xref: /freebsd-13.1/sys/net/iflib.c (revision eaeff159)
1 /*-
2  * Copyright (c) 2014-2018, Matthew Macy <[email protected]>
3  * All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions are met:
7  *
8  *  1. Redistributions of source code must retain the above copyright notice,
9  *     this list of conditions and the following disclaimer.
10  *
11  *  2. Neither the name of Matthew Macy nor the names of its
12  *     contributors may be used to endorse or promote products derived from
13  *     this software without specific prior written permission.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
16  * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18  * ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
19  * LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
20  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
21  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
22  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
23  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
24  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
25  * POSSIBILITY OF SUCH DAMAGE.
26  */
27 
28 #include <sys/cdefs.h>
29 __FBSDID("$FreeBSD$");
30 
31 #include "opt_inet.h"
32 #include "opt_inet6.h"
33 #include "opt_acpi.h"
34 #include "opt_sched.h"
35 
36 #include <sys/param.h>
37 #include <sys/types.h>
38 #include <sys/bus.h>
39 #include <sys/eventhandler.h>
40 #include <sys/kernel.h>
41 #include <sys/lock.h>
42 #include <sys/mutex.h>
43 #include <sys/module.h>
44 #include <sys/kobj.h>
45 #include <sys/rman.h>
46 #include <sys/sbuf.h>
47 #include <sys/smp.h>
48 #include <sys/socket.h>
49 #include <sys/sockio.h>
50 #include <sys/sysctl.h>
51 #include <sys/syslog.h>
52 #include <sys/taskqueue.h>
53 #include <sys/limits.h>
54 
55 #include <net/if.h>
56 #include <net/if_var.h>
57 #include <net/if_types.h>
58 #include <net/if_media.h>
59 #include <net/bpf.h>
60 #include <net/ethernet.h>
61 #include <net/mp_ring.h>
62 #include <net/debugnet.h>
63 #include <net/pfil.h>
64 #include <net/vnet.h>
65 
66 #include <netinet/in.h>
67 #include <netinet/in_pcb.h>
68 #include <netinet/tcp_lro.h>
69 #include <netinet/in_systm.h>
70 #include <netinet/if_ether.h>
71 #include <netinet/ip.h>
72 #include <netinet/ip6.h>
73 #include <netinet/tcp.h>
74 #include <netinet/ip_var.h>
75 #include <netinet6/ip6_var.h>
76 
77 #include <machine/bus.h>
78 #include <machine/in_cksum.h>
79 
80 #include <vm/vm.h>
81 #include <vm/pmap.h>
82 
83 #include <dev/led/led.h>
84 #include <dev/pci/pcireg.h>
85 #include <dev/pci/pcivar.h>
86 #include <dev/pci/pci_private.h>
87 
88 #include <net/iflib.h>
89 #include <net/iflib_private.h>
90 
91 #include "ifdi_if.h"
92 
93 #ifdef PCI_IOV
94 #include <dev/pci/pci_iov.h>
95 #endif
96 
97 #include <sys/bitstring.h>
98 /*
99  * enable accounting of every mbuf as it comes in to and goes out of
100  * iflib's software descriptor references
101  */
102 #define MEMORY_LOGGING 0
103 /*
104  * Enable mbuf vectors for compressing long mbuf chains
105  */
106 
107 /*
108  * NB:
109  * - Prefetching in tx cleaning should perhaps be a tunable. The distance ahead
110  *   we prefetch needs to be determined by the time spent in m_free vis a vis
111  *   the cost of a prefetch. This will of course vary based on the workload:
112  *      - NFLX's m_free path is dominated by vm-based M_EXT manipulation which
113  *        is quite expensive, thus suggesting very little prefetch.
114  *      - small packet forwarding which is just returning a single mbuf to
115  *        UMA will typically be very fast vis a vis the cost of a memory
116  *        access.
117  */
118 
119 /*
120  * File organization:
121  *  - private structures
122  *  - iflib private utility functions
123  *  - ifnet functions
124  *  - vlan registry and other exported functions
125  *  - iflib public core functions
126  *
127  *
128  */
129 MALLOC_DEFINE(M_IFLIB, "iflib", "ifnet library");
130 
131 #define	IFLIB_RXEOF_MORE (1U << 0)
132 #define	IFLIB_RXEOF_EMPTY (2U << 0)
133 
134 struct iflib_txq;
135 typedef struct iflib_txq *iflib_txq_t;
136 struct iflib_rxq;
137 typedef struct iflib_rxq *iflib_rxq_t;
138 struct iflib_fl;
139 typedef struct iflib_fl *iflib_fl_t;
140 
141 struct iflib_ctx;
142 
143 static void iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid);
144 static void iflib_timer(void *arg);
145 static void iflib_tqg_detach(if_ctx_t ctx);
146 
147 typedef struct iflib_filter_info {
148 	driver_filter_t *ifi_filter;
149 	void *ifi_filter_arg;
150 	struct grouptask *ifi_task;
151 	void *ifi_ctx;
152 } *iflib_filter_info_t;
153 
154 struct iflib_ctx {
155 	KOBJ_FIELDS;
156 	/*
157 	 * Pointer to hardware driver's softc
158 	 */
159 	void *ifc_softc;
160 	device_t ifc_dev;
161 	if_t ifc_ifp;
162 
163 	cpuset_t ifc_cpus;
164 	if_shared_ctx_t ifc_sctx;
165 	struct if_softc_ctx ifc_softc_ctx;
166 
167 	struct sx ifc_ctx_sx;
168 	struct mtx ifc_state_mtx;
169 
170 	iflib_txq_t ifc_txqs;
171 	iflib_rxq_t ifc_rxqs;
172 	uint32_t ifc_if_flags;
173 	uint32_t ifc_flags;
174 	uint32_t ifc_max_fl_buf_size;
175 	uint32_t ifc_rx_mbuf_sz;
176 
177 	int ifc_link_state;
178 	int ifc_watchdog_events;
179 	struct cdev *ifc_led_dev;
180 	struct resource *ifc_msix_mem;
181 
182 	struct if_irq ifc_legacy_irq;
183 	struct grouptask ifc_admin_task;
184 	struct grouptask ifc_vflr_task;
185 	struct iflib_filter_info ifc_filter_info;
186 	struct ifmedia	ifc_media;
187 	struct ifmedia	*ifc_mediap;
188 
189 	struct sysctl_oid *ifc_sysctl_node;
190 	uint16_t ifc_sysctl_ntxqs;
191 	uint16_t ifc_sysctl_nrxqs;
192 	uint16_t ifc_sysctl_qs_eq_override;
193 	uint16_t ifc_sysctl_rx_budget;
194 	uint16_t ifc_sysctl_tx_abdicate;
195 	uint16_t ifc_sysctl_core_offset;
196 #define	CORE_OFFSET_UNSPECIFIED	0xffff
197 	uint8_t  ifc_sysctl_separate_txrx;
198 	uint8_t  ifc_sysctl_use_logical_cores;
199 	bool	 ifc_cpus_are_physical_cores;
200 
201 	qidx_t ifc_sysctl_ntxds[8];
202 	qidx_t ifc_sysctl_nrxds[8];
203 	struct if_txrx ifc_txrx;
204 #define isc_txd_encap  ifc_txrx.ift_txd_encap
205 #define isc_txd_flush  ifc_txrx.ift_txd_flush
206 #define isc_txd_credits_update  ifc_txrx.ift_txd_credits_update
207 #define isc_rxd_available ifc_txrx.ift_rxd_available
208 #define isc_rxd_pkt_get ifc_txrx.ift_rxd_pkt_get
209 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
210 #define isc_rxd_flush ifc_txrx.ift_rxd_flush
211 #define isc_legacy_intr ifc_txrx.ift_legacy_intr
212 	eventhandler_tag ifc_vlan_attach_event;
213 	eventhandler_tag ifc_vlan_detach_event;
214 	struct ether_addr ifc_mac;
215 };
216 
217 void *
218 iflib_get_softc(if_ctx_t ctx)
219 {
220 
221 	return (ctx->ifc_softc);
222 }
223 
224 device_t
225 iflib_get_dev(if_ctx_t ctx)
226 {
227 
228 	return (ctx->ifc_dev);
229 }
230 
231 if_t
232 iflib_get_ifp(if_ctx_t ctx)
233 {
234 
235 	return (ctx->ifc_ifp);
236 }
237 
238 struct ifmedia *
239 iflib_get_media(if_ctx_t ctx)
240 {
241 
242 	return (ctx->ifc_mediap);
243 }
244 
245 uint32_t
246 iflib_get_flags(if_ctx_t ctx)
247 {
248 	return (ctx->ifc_flags);
249 }
250 
251 void
252 iflib_set_mac(if_ctx_t ctx, uint8_t mac[ETHER_ADDR_LEN])
253 {
254 
255 	bcopy(mac, ctx->ifc_mac.octet, ETHER_ADDR_LEN);
256 }
257 
258 if_softc_ctx_t
259 iflib_get_softc_ctx(if_ctx_t ctx)
260 {
261 
262 	return (&ctx->ifc_softc_ctx);
263 }
264 
265 if_shared_ctx_t
266 iflib_get_sctx(if_ctx_t ctx)
267 {
268 
269 	return (ctx->ifc_sctx);
270 }
271 
272 #define IP_ALIGNED(m) ((((uintptr_t)(m)->m_data) & 0x3) == 0x2)
273 #define CACHE_PTR_INCREMENT (CACHE_LINE_SIZE/sizeof(void*))
274 #define CACHE_PTR_NEXT(ptr) ((void *)(((uintptr_t)(ptr)+CACHE_LINE_SIZE-1) & (CACHE_LINE_SIZE-1)))
275 
276 #define LINK_ACTIVE(ctx) ((ctx)->ifc_link_state == LINK_STATE_UP)
277 #define CTX_IS_VF(ctx) ((ctx)->ifc_sctx->isc_flags & IFLIB_IS_VF)
278 
279 typedef struct iflib_sw_rx_desc_array {
280 	bus_dmamap_t	*ifsd_map;         /* bus_dma maps for packet */
281 	struct mbuf	**ifsd_m;           /* pkthdr mbufs */
282 	caddr_t		*ifsd_cl;          /* direct cluster pointer for rx */
283 	bus_addr_t	*ifsd_ba;          /* bus addr of cluster for rx */
284 } iflib_rxsd_array_t;
285 
286 typedef struct iflib_sw_tx_desc_array {
287 	bus_dmamap_t    *ifsd_map;         /* bus_dma maps for packet */
288 	bus_dmamap_t	*ifsd_tso_map;     /* bus_dma maps for TSO packet */
289 	struct mbuf    **ifsd_m;           /* pkthdr mbufs */
290 } if_txsd_vec_t;
291 
292 /* magic number that should be high enough for any hardware */
293 #define IFLIB_MAX_TX_SEGS		128
294 #define IFLIB_RX_COPY_THRESH		128
295 #define IFLIB_MAX_RX_REFRESH		32
296 /* The minimum descriptors per second before we start coalescing */
297 #define IFLIB_MIN_DESC_SEC		16384
298 #define IFLIB_DEFAULT_TX_UPDATE_FREQ	16
299 #define IFLIB_QUEUE_IDLE		0
300 #define IFLIB_QUEUE_HUNG		1
301 #define IFLIB_QUEUE_WORKING		2
302 /* maximum number of txqs that can share an rx interrupt */
303 #define IFLIB_MAX_TX_SHARED_INTR	4
304 
305 /* this should really scale with ring size - this is a fairly arbitrary value */
306 #define TX_BATCH_SIZE			32
307 
308 #define IFLIB_RESTART_BUDGET		8
309 
310 #define CSUM_OFFLOAD		(CSUM_IP_TSO|CSUM_IP6_TSO|CSUM_IP| \
311 				 CSUM_IP_UDP|CSUM_IP_TCP|CSUM_IP_SCTP| \
312 				 CSUM_IP6_UDP|CSUM_IP6_TCP|CSUM_IP6_SCTP)
313 
314 struct iflib_txq {
315 	qidx_t		ift_in_use;
316 	qidx_t		ift_cidx;
317 	qidx_t		ift_cidx_processed;
318 	qidx_t		ift_pidx;
319 	uint8_t		ift_gen;
320 	uint8_t		ift_br_offset;
321 	uint16_t	ift_npending;
322 	uint16_t	ift_db_pending;
323 	uint16_t	ift_rs_pending;
324 	/* implicit pad */
325 	uint8_t		ift_txd_size[8];
326 	uint64_t	ift_processed;
327 	uint64_t	ift_cleaned;
328 	uint64_t	ift_cleaned_prev;
329 #if MEMORY_LOGGING
330 	uint64_t	ift_enqueued;
331 	uint64_t	ift_dequeued;
332 #endif
333 	uint64_t	ift_no_tx_dma_setup;
334 	uint64_t	ift_no_desc_avail;
335 	uint64_t	ift_mbuf_defrag_failed;
336 	uint64_t	ift_mbuf_defrag;
337 	uint64_t	ift_map_failed;
338 	uint64_t	ift_txd_encap_efbig;
339 	uint64_t	ift_pullups;
340 	uint64_t	ift_last_timer_tick;
341 
342 	struct mtx	ift_mtx;
343 	struct mtx	ift_db_mtx;
344 
345 	/* constant values */
346 	if_ctx_t	ift_ctx;
347 	struct ifmp_ring        *ift_br;
348 	struct grouptask	ift_task;
349 	qidx_t		ift_size;
350 	uint16_t	ift_id;
351 	struct callout	ift_timer;
352 #ifdef DEV_NETMAP
353 	struct callout	ift_netmap_timer;
354 #endif /* DEV_NETMAP */
355 
356 	if_txsd_vec_t	ift_sds;
357 	uint8_t		ift_qstatus;
358 	uint8_t		ift_closed;
359 	uint8_t		ift_update_freq;
360 	struct iflib_filter_info ift_filter_info;
361 	bus_dma_tag_t	ift_buf_tag;
362 	bus_dma_tag_t	ift_tso_buf_tag;
363 	iflib_dma_info_t	ift_ifdi;
364 #define	MTX_NAME_LEN	32
365 	char                    ift_mtx_name[MTX_NAME_LEN];
366 	bus_dma_segment_t	ift_segs[IFLIB_MAX_TX_SEGS]  __aligned(CACHE_LINE_SIZE);
367 #ifdef IFLIB_DIAGNOSTICS
368 	uint64_t ift_cpu_exec_count[256];
369 #endif
370 } __aligned(CACHE_LINE_SIZE);
371 
372 struct iflib_fl {
373 	qidx_t		ifl_cidx;
374 	qidx_t		ifl_pidx;
375 	qidx_t		ifl_credits;
376 	uint8_t		ifl_gen;
377 	uint8_t		ifl_rxd_size;
378 #if MEMORY_LOGGING
379 	uint64_t	ifl_m_enqueued;
380 	uint64_t	ifl_m_dequeued;
381 	uint64_t	ifl_cl_enqueued;
382 	uint64_t	ifl_cl_dequeued;
383 #endif
384 	/* implicit pad */
385 	bitstr_t 	*ifl_rx_bitmap;
386 	qidx_t		ifl_fragidx;
387 	/* constant */
388 	qidx_t		ifl_size;
389 	uint16_t	ifl_buf_size;
390 	uint16_t	ifl_cltype;
391 	uma_zone_t	ifl_zone;
392 	iflib_rxsd_array_t	ifl_sds;
393 	iflib_rxq_t	ifl_rxq;
394 	uint8_t		ifl_id;
395 	bus_dma_tag_t	ifl_buf_tag;
396 	iflib_dma_info_t	ifl_ifdi;
397 	uint64_t	ifl_bus_addrs[IFLIB_MAX_RX_REFRESH] __aligned(CACHE_LINE_SIZE);
398 	qidx_t		ifl_rxd_idxs[IFLIB_MAX_RX_REFRESH];
399 }  __aligned(CACHE_LINE_SIZE);
400 
401 static inline qidx_t
402 get_inuse(int size, qidx_t cidx, qidx_t pidx, uint8_t gen)
403 {
404 	qidx_t used;
405 
406 	if (pidx > cidx)
407 		used = pidx - cidx;
408 	else if (pidx < cidx)
409 		used = size - cidx + pidx;
410 	else if (gen == 0 && pidx == cidx)
411 		used = 0;
412 	else if (gen == 1 && pidx == cidx)
413 		used = size;
414 	else
415 		panic("bad state");
416 
417 	return (used);
418 }
419 
420 #define TXQ_AVAIL(txq) (txq->ift_size - get_inuse(txq->ift_size, txq->ift_cidx, txq->ift_pidx, txq->ift_gen))
421 
422 #define IDXDIFF(head, tail, wrap) \
423 	((head) >= (tail) ? (head) - (tail) : (wrap) - (tail) + (head))
424 
425 struct iflib_rxq {
426 	if_ctx_t	ifr_ctx;
427 	iflib_fl_t	ifr_fl;
428 	uint64_t	ifr_rx_irq;
429 	struct pfil_head	*pfil;
430 	/*
431 	 * If there is a separate completion queue (IFLIB_HAS_RXCQ), this is
432 	 * the completion queue consumer index.  Otherwise it's unused.
433 	 */
434 	qidx_t		ifr_cq_cidx;
435 	uint16_t	ifr_id;
436 	uint8_t		ifr_nfl;
437 	uint8_t		ifr_ntxqirq;
438 	uint8_t		ifr_txqid[IFLIB_MAX_TX_SHARED_INTR];
439 	uint8_t		ifr_fl_offset;
440 	struct lro_ctrl			ifr_lc;
441 	struct grouptask        ifr_task;
442 	struct callout		ifr_watchdog;
443 	struct iflib_filter_info ifr_filter_info;
444 	iflib_dma_info_t		ifr_ifdi;
445 
446 	/* dynamically allocate if any drivers need a value substantially larger than this */
447 	struct if_rxd_frag	ifr_frags[IFLIB_MAX_RX_SEGS] __aligned(CACHE_LINE_SIZE);
448 #ifdef IFLIB_DIAGNOSTICS
449 	uint64_t ifr_cpu_exec_count[256];
450 #endif
451 }  __aligned(CACHE_LINE_SIZE);
452 
453 typedef struct if_rxsd {
454 	caddr_t *ifsd_cl;
455 	iflib_fl_t ifsd_fl;
456 } *if_rxsd_t;
457 
458 /* multiple of word size */
459 #ifdef __LP64__
460 #define PKT_INFO_SIZE	6
461 #define RXD_INFO_SIZE	5
462 #define PKT_TYPE uint64_t
463 #else
464 #define PKT_INFO_SIZE	11
465 #define RXD_INFO_SIZE	8
466 #define PKT_TYPE uint32_t
467 #endif
468 #define PKT_LOOP_BOUND  ((PKT_INFO_SIZE/3)*3)
469 #define RXD_LOOP_BOUND  ((RXD_INFO_SIZE/4)*4)
470 
471 typedef struct if_pkt_info_pad {
472 	PKT_TYPE pkt_val[PKT_INFO_SIZE];
473 } *if_pkt_info_pad_t;
474 typedef struct if_rxd_info_pad {
475 	PKT_TYPE rxd_val[RXD_INFO_SIZE];
476 } *if_rxd_info_pad_t;
477 
478 CTASSERT(sizeof(struct if_pkt_info_pad) == sizeof(struct if_pkt_info));
479 CTASSERT(sizeof(struct if_rxd_info_pad) == sizeof(struct if_rxd_info));
480 
481 static inline void
482 pkt_info_zero(if_pkt_info_t pi)
483 {
484 	if_pkt_info_pad_t pi_pad;
485 
486 	pi_pad = (if_pkt_info_pad_t)pi;
487 	pi_pad->pkt_val[0] = 0; pi_pad->pkt_val[1] = 0; pi_pad->pkt_val[2] = 0;
488 	pi_pad->pkt_val[3] = 0; pi_pad->pkt_val[4] = 0; pi_pad->pkt_val[5] = 0;
489 #ifndef __LP64__
490 	pi_pad->pkt_val[6] = 0; pi_pad->pkt_val[7] = 0; pi_pad->pkt_val[8] = 0;
491 	pi_pad->pkt_val[9] = 0; pi_pad->pkt_val[10] = 0;
492 #endif
493 }
494 
495 static device_method_t iflib_pseudo_methods[] = {
496 	DEVMETHOD(device_attach, noop_attach),
497 	DEVMETHOD(device_detach, iflib_pseudo_detach),
498 	DEVMETHOD_END
499 };
500 
501 driver_t iflib_pseudodriver = {
502 	"iflib_pseudo", iflib_pseudo_methods, sizeof(struct iflib_ctx),
503 };
504 
505 static inline void
506 rxd_info_zero(if_rxd_info_t ri)
507 {
508 	if_rxd_info_pad_t ri_pad;
509 	int i;
510 
511 	ri_pad = (if_rxd_info_pad_t)ri;
512 	for (i = 0; i < RXD_LOOP_BOUND; i += 4) {
513 		ri_pad->rxd_val[i] = 0;
514 		ri_pad->rxd_val[i+1] = 0;
515 		ri_pad->rxd_val[i+2] = 0;
516 		ri_pad->rxd_val[i+3] = 0;
517 	}
518 #ifdef __LP64__
519 	ri_pad->rxd_val[RXD_INFO_SIZE-1] = 0;
520 #endif
521 }
522 
523 /*
524  * Only allow a single packet to take up most 1/nth of the tx ring
525  */
526 #define MAX_SINGLE_PACKET_FRACTION 12
527 #define IF_BAD_DMA (bus_addr_t)-1
528 
529 #define CTX_ACTIVE(ctx) ((if_getdrvflags((ctx)->ifc_ifp) & IFF_DRV_RUNNING))
530 
531 #define CTX_LOCK_INIT(_sc)  sx_init(&(_sc)->ifc_ctx_sx, "iflib ctx lock")
532 #define CTX_LOCK(ctx) sx_xlock(&(ctx)->ifc_ctx_sx)
533 #define CTX_UNLOCK(ctx) sx_xunlock(&(ctx)->ifc_ctx_sx)
534 #define CTX_LOCK_DESTROY(ctx) sx_destroy(&(ctx)->ifc_ctx_sx)
535 
536 #define STATE_LOCK_INIT(_sc, _name)  mtx_init(&(_sc)->ifc_state_mtx, _name, "iflib state lock", MTX_DEF)
537 #define STATE_LOCK(ctx) mtx_lock(&(ctx)->ifc_state_mtx)
538 #define STATE_UNLOCK(ctx) mtx_unlock(&(ctx)->ifc_state_mtx)
539 #define STATE_LOCK_DESTROY(ctx) mtx_destroy(&(ctx)->ifc_state_mtx)
540 
541 #define CALLOUT_LOCK(txq)	mtx_lock(&txq->ift_mtx)
542 #define CALLOUT_UNLOCK(txq) 	mtx_unlock(&txq->ift_mtx)
543 
544 void
545 iflib_set_detach(if_ctx_t ctx)
546 {
547 	STATE_LOCK(ctx);
548 	ctx->ifc_flags |= IFC_IN_DETACH;
549 	STATE_UNLOCK(ctx);
550 }
551 
552 /* Our boot-time initialization hook */
553 static int	iflib_module_event_handler(module_t, int, void *);
554 
555 static moduledata_t iflib_moduledata = {
556 	"iflib",
557 	iflib_module_event_handler,
558 	NULL
559 };
560 
561 DECLARE_MODULE(iflib, iflib_moduledata, SI_SUB_INIT_IF, SI_ORDER_ANY);
562 MODULE_VERSION(iflib, 1);
563 
564 MODULE_DEPEND(iflib, pci, 1, 1, 1);
565 MODULE_DEPEND(iflib, ether, 1, 1, 1);
566 
567 TASKQGROUP_DEFINE(if_io_tqg, mp_ncpus, 1);
568 TASKQGROUP_DEFINE(if_config_tqg, 1, 1);
569 
570 #ifndef IFLIB_DEBUG_COUNTERS
571 #ifdef INVARIANTS
572 #define IFLIB_DEBUG_COUNTERS 1
573 #else
574 #define IFLIB_DEBUG_COUNTERS 0
575 #endif /* !INVARIANTS */
576 #endif
577 
578 static SYSCTL_NODE(_net, OID_AUTO, iflib, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
579     "iflib driver parameters");
580 
581 /*
582  * XXX need to ensure that this can't accidentally cause the head to be moved backwards
583  */
584 static int iflib_min_tx_latency = 0;
585 SYSCTL_INT(_net_iflib, OID_AUTO, min_tx_latency, CTLFLAG_RW,
586 		   &iflib_min_tx_latency, 0, "minimize transmit latency at the possible expense of throughput");
587 static int iflib_no_tx_batch = 0;
588 SYSCTL_INT(_net_iflib, OID_AUTO, no_tx_batch, CTLFLAG_RW,
589 		   &iflib_no_tx_batch, 0, "minimize transmit latency at the possible expense of throughput");
590 static int iflib_timer_default = 1000;
591 SYSCTL_INT(_net_iflib, OID_AUTO, timer_default, CTLFLAG_RW,
592 		   &iflib_timer_default, 0, "number of ticks between iflib_timer calls");
593 
594 
595 #if IFLIB_DEBUG_COUNTERS
596 
597 static int iflib_tx_seen;
598 static int iflib_tx_sent;
599 static int iflib_tx_encap;
600 static int iflib_rx_allocs;
601 static int iflib_fl_refills;
602 static int iflib_fl_refills_large;
603 static int iflib_tx_frees;
604 
605 SYSCTL_INT(_net_iflib, OID_AUTO, tx_seen, CTLFLAG_RD,
606 		   &iflib_tx_seen, 0, "# TX mbufs seen");
607 SYSCTL_INT(_net_iflib, OID_AUTO, tx_sent, CTLFLAG_RD,
608 		   &iflib_tx_sent, 0, "# TX mbufs sent");
609 SYSCTL_INT(_net_iflib, OID_AUTO, tx_encap, CTLFLAG_RD,
610 		   &iflib_tx_encap, 0, "# TX mbufs encapped");
611 SYSCTL_INT(_net_iflib, OID_AUTO, tx_frees, CTLFLAG_RD,
612 		   &iflib_tx_frees, 0, "# TX frees");
613 SYSCTL_INT(_net_iflib, OID_AUTO, rx_allocs, CTLFLAG_RD,
614 		   &iflib_rx_allocs, 0, "# RX allocations");
615 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills, CTLFLAG_RD,
616 		   &iflib_fl_refills, 0, "# refills");
617 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills_large, CTLFLAG_RD,
618 		   &iflib_fl_refills_large, 0, "# large refills");
619 
620 static int iflib_txq_drain_flushing;
621 static int iflib_txq_drain_oactive;
622 static int iflib_txq_drain_notready;
623 
624 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_flushing, CTLFLAG_RD,
625 		   &iflib_txq_drain_flushing, 0, "# drain flushes");
626 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_oactive, CTLFLAG_RD,
627 		   &iflib_txq_drain_oactive, 0, "# drain oactives");
628 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_notready, CTLFLAG_RD,
629 		   &iflib_txq_drain_notready, 0, "# drain notready");
630 
631 static int iflib_encap_load_mbuf_fail;
632 static int iflib_encap_pad_mbuf_fail;
633 static int iflib_encap_txq_avail_fail;
634 static int iflib_encap_txd_encap_fail;
635 
636 SYSCTL_INT(_net_iflib, OID_AUTO, encap_load_mbuf_fail, CTLFLAG_RD,
637 		   &iflib_encap_load_mbuf_fail, 0, "# busdma load failures");
638 SYSCTL_INT(_net_iflib, OID_AUTO, encap_pad_mbuf_fail, CTLFLAG_RD,
639 		   &iflib_encap_pad_mbuf_fail, 0, "# runt frame pad failures");
640 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txq_avail_fail, CTLFLAG_RD,
641 		   &iflib_encap_txq_avail_fail, 0, "# txq avail failures");
642 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txd_encap_fail, CTLFLAG_RD,
643 		   &iflib_encap_txd_encap_fail, 0, "# driver encap failures");
644 
645 static int iflib_task_fn_rxs;
646 static int iflib_rx_intr_enables;
647 static int iflib_fast_intrs;
648 static int iflib_rx_unavail;
649 static int iflib_rx_ctx_inactive;
650 static int iflib_rx_if_input;
651 static int iflib_rxd_flush;
652 
653 static int iflib_verbose_debug;
654 
655 SYSCTL_INT(_net_iflib, OID_AUTO, task_fn_rx, CTLFLAG_RD,
656 		   &iflib_task_fn_rxs, 0, "# task_fn_rx calls");
657 SYSCTL_INT(_net_iflib, OID_AUTO, rx_intr_enables, CTLFLAG_RD,
658 		   &iflib_rx_intr_enables, 0, "# RX intr enables");
659 SYSCTL_INT(_net_iflib, OID_AUTO, fast_intrs, CTLFLAG_RD,
660 		   &iflib_fast_intrs, 0, "# fast_intr calls");
661 SYSCTL_INT(_net_iflib, OID_AUTO, rx_unavail, CTLFLAG_RD,
662 		   &iflib_rx_unavail, 0, "# times rxeof called with no available data");
663 SYSCTL_INT(_net_iflib, OID_AUTO, rx_ctx_inactive, CTLFLAG_RD,
664 		   &iflib_rx_ctx_inactive, 0, "# times rxeof called with inactive context");
665 SYSCTL_INT(_net_iflib, OID_AUTO, rx_if_input, CTLFLAG_RD,
666 		   &iflib_rx_if_input, 0, "# times rxeof called if_input");
667 SYSCTL_INT(_net_iflib, OID_AUTO, rxd_flush, CTLFLAG_RD,
668 	         &iflib_rxd_flush, 0, "# times rxd_flush called");
669 SYSCTL_INT(_net_iflib, OID_AUTO, verbose_debug, CTLFLAG_RW,
670 		   &iflib_verbose_debug, 0, "enable verbose debugging");
671 
672 #define DBG_COUNTER_INC(name) atomic_add_int(&(iflib_ ## name), 1)
673 static void
674 iflib_debug_reset(void)
675 {
676 	iflib_tx_seen = iflib_tx_sent = iflib_tx_encap = iflib_rx_allocs =
677 		iflib_fl_refills = iflib_fl_refills_large = iflib_tx_frees =
678 		iflib_txq_drain_flushing = iflib_txq_drain_oactive =
679 		iflib_txq_drain_notready =
680 		iflib_encap_load_mbuf_fail = iflib_encap_pad_mbuf_fail =
681 		iflib_encap_txq_avail_fail = iflib_encap_txd_encap_fail =
682 		iflib_task_fn_rxs = iflib_rx_intr_enables = iflib_fast_intrs =
683 		iflib_rx_unavail =
684 		iflib_rx_ctx_inactive = iflib_rx_if_input =
685 		iflib_rxd_flush = 0;
686 }
687 
688 #else
689 #define DBG_COUNTER_INC(name)
690 static void iflib_debug_reset(void) {}
691 #endif
692 
693 #define IFLIB_DEBUG 0
694 
695 static void iflib_tx_structures_free(if_ctx_t ctx);
696 static void iflib_rx_structures_free(if_ctx_t ctx);
697 static int iflib_queues_alloc(if_ctx_t ctx);
698 static int iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq);
699 static int iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget);
700 static int iflib_qset_structures_setup(if_ctx_t ctx);
701 static int iflib_msix_init(if_ctx_t ctx);
702 static int iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filterarg, int *rid, const char *str);
703 static void iflib_txq_check_drain(iflib_txq_t txq, int budget);
704 static uint32_t iflib_txq_can_drain(struct ifmp_ring *);
705 #ifdef ALTQ
706 static void iflib_altq_if_start(if_t ifp);
707 static int iflib_altq_if_transmit(if_t ifp, struct mbuf *m);
708 #endif
709 static int iflib_register(if_ctx_t);
710 static void iflib_deregister(if_ctx_t);
711 static void iflib_unregister_vlan_handlers(if_ctx_t ctx);
712 static uint16_t iflib_get_mbuf_size_for(unsigned int size);
713 static void iflib_init_locked(if_ctx_t ctx);
714 static void iflib_add_device_sysctl_pre(if_ctx_t ctx);
715 static void iflib_add_device_sysctl_post(if_ctx_t ctx);
716 static void iflib_ifmp_purge(iflib_txq_t txq);
717 static void _iflib_pre_assert(if_softc_ctx_t scctx);
718 static void iflib_if_init_locked(if_ctx_t ctx);
719 static void iflib_free_intr_mem(if_ctx_t ctx);
720 #ifndef __NO_STRICT_ALIGNMENT
721 static struct mbuf * iflib_fixup_rx(struct mbuf *m);
722 #endif
723 
724 static SLIST_HEAD(cpu_offset_list, cpu_offset) cpu_offsets =
725     SLIST_HEAD_INITIALIZER(cpu_offsets);
726 struct cpu_offset {
727 	SLIST_ENTRY(cpu_offset) entries;
728 	cpuset_t	set;
729 	unsigned int	refcount;
730 	uint16_t	next_cpuid;
731 };
732 static struct mtx cpu_offset_mtx;
733 MTX_SYSINIT(iflib_cpu_offset, &cpu_offset_mtx, "iflib_cpu_offset lock",
734     MTX_DEF);
735 
736 DEBUGNET_DEFINE(iflib);
737 
738 static int
739 iflib_num_rx_descs(if_ctx_t ctx)
740 {
741 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
742 	if_shared_ctx_t sctx = ctx->ifc_sctx;
743 	uint16_t first_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
744 
745 	return scctx->isc_nrxd[first_rxq];
746 }
747 
748 static int
749 iflib_num_tx_descs(if_ctx_t ctx)
750 {
751 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
752 	if_shared_ctx_t sctx = ctx->ifc_sctx;
753 	uint16_t first_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
754 
755 	return scctx->isc_ntxd[first_txq];
756 }
757 
758 #ifdef DEV_NETMAP
759 #include <sys/selinfo.h>
760 #include <net/netmap.h>
761 #include <dev/netmap/netmap_kern.h>
762 
763 MODULE_DEPEND(iflib, netmap, 1, 1, 1);
764 
765 static int netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, bool init);
766 static void iflib_netmap_timer(void *arg);
767 
768 /*
769  * device-specific sysctl variables:
770  *
771  * iflib_crcstrip: 0: keep CRC in rx frames (default), 1: strip it.
772  *	During regular operations the CRC is stripped, but on some
773  *	hardware reception of frames not multiple of 64 is slower,
774  *	so using crcstrip=0 helps in benchmarks.
775  *
776  * iflib_rx_miss, iflib_rx_miss_bufs:
777  *	count packets that might be missed due to lost interrupts.
778  */
779 SYSCTL_DECL(_dev_netmap);
780 /*
781  * The xl driver by default strips CRCs and we do not override it.
782  */
783 
784 int iflib_crcstrip = 1;
785 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_crcstrip,
786     CTLFLAG_RW, &iflib_crcstrip, 1, "strip CRC on RX frames");
787 
788 int iflib_rx_miss, iflib_rx_miss_bufs;
789 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss,
790     CTLFLAG_RW, &iflib_rx_miss, 0, "potentially missed RX intr");
791 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss_bufs,
792     CTLFLAG_RW, &iflib_rx_miss_bufs, 0, "potentially missed RX intr bufs");
793 
794 /*
795  * Register/unregister. We are already under netmap lock.
796  * Only called on the first register or the last unregister.
797  */
798 static int
799 iflib_netmap_register(struct netmap_adapter *na, int onoff)
800 {
801 	if_t ifp = na->ifp;
802 	if_ctx_t ctx = ifp->if_softc;
803 	int status;
804 
805 	CTX_LOCK(ctx);
806 	if (!CTX_IS_VF(ctx))
807 		IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip);
808 
809 	iflib_stop(ctx);
810 
811 	/*
812 	 * Enable (or disable) netmap flags, and intercept (or restore)
813 	 * ifp->if_transmit. This is done once the device has been stopped
814 	 * to prevent race conditions. Also, this must be done after
815 	 * calling netmap_disable_all_rings() and before calling
816 	 * netmap_enable_all_rings(), so that these two functions see the
817 	 * updated state of the NAF_NETMAP_ON bit.
818 	 */
819 	if (onoff) {
820 		nm_set_native_flags(na);
821 	} else {
822 		nm_clear_native_flags(na);
823 	}
824 
825 	iflib_init_locked(ctx);
826 	IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip); // XXX why twice ?
827 	status = ifp->if_drv_flags & IFF_DRV_RUNNING ? 0 : 1;
828 	if (status)
829 		nm_clear_native_flags(na);
830 	CTX_UNLOCK(ctx);
831 	return (status);
832 }
833 
834 static int
835 iflib_netmap_config(struct netmap_adapter *na, struct nm_config_info *info)
836 {
837 	if_t ifp = na->ifp;
838 	if_ctx_t ctx = ifp->if_softc;
839 	iflib_rxq_t rxq = &ctx->ifc_rxqs[0];
840 	iflib_fl_t fl = &rxq->ifr_fl[0];
841 
842 	info->num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets;
843 	info->num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets;
844 	info->num_tx_descs = iflib_num_tx_descs(ctx);
845 	info->num_rx_descs = iflib_num_rx_descs(ctx);
846 	info->rx_buf_maxsize = fl->ifl_buf_size;
847 	nm_prinf("txr %u rxr %u txd %u rxd %u rbufsz %u",
848 		info->num_tx_rings, info->num_rx_rings, info->num_tx_descs,
849 		info->num_rx_descs, info->rx_buf_maxsize);
850 
851 	return 0;
852 }
853 
854 static int
855 netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, bool init)
856 {
857 	struct netmap_adapter *na = kring->na;
858 	u_int const lim = kring->nkr_num_slots - 1;
859 	struct netmap_ring *ring = kring->ring;
860 	bus_dmamap_t *map;
861 	struct if_rxd_update iru;
862 	if_ctx_t ctx = rxq->ifr_ctx;
863 	iflib_fl_t fl = &rxq->ifr_fl[0];
864 	u_int nic_i_first, nic_i;
865 	u_int nm_i;
866 	int i, n;
867 #if IFLIB_DEBUG_COUNTERS
868 	int rf_count = 0;
869 #endif
870 
871 	/*
872 	 * This function is used both at initialization and in rxsync.
873 	 * At initialization we need to prepare (with isc_rxd_refill())
874 	 * all the netmap buffers currently owned by the kernel, in
875 	 * such a way to keep fl->ifl_pidx and kring->nr_hwcur in sync
876 	 * (except for kring->nkr_hwofs). These may be less than
877 	 * kring->nkr_num_slots if netmap_reset() was called while
878 	 * an application using the kring that still owned some
879 	 * buffers.
880 	 * At rxsync time, both indexes point to the next buffer to be
881 	 * refilled.
882 	 * In any case we publish (with isc_rxd_flush()) up to
883 	 * (fl->ifl_pidx - 1) % N (included), to avoid the NIC tail/prod
884 	 * pointer to overrun the head/cons pointer, although this is
885 	 * not necessary for some NICs (e.g. vmx).
886 	 */
887 	if (__predict_false(init)) {
888 		n = kring->nkr_num_slots - nm_kr_rxspace(kring);
889 	} else {
890 		n = kring->rhead - kring->nr_hwcur;
891 		if (n == 0)
892 			return (0); /* Nothing to do. */
893 		if (n < 0)
894 			n += kring->nkr_num_slots;
895 	}
896 
897 	iru_init(&iru, rxq, 0 /* flid */);
898 	map = fl->ifl_sds.ifsd_map;
899 	nic_i = fl->ifl_pidx;
900 	nm_i = netmap_idx_n2k(kring, nic_i);
901 	if (__predict_false(init)) {
902 		/*
903 		 * On init/reset, nic_i must be 0, and we must
904 		 * start to refill from hwtail (see netmap_reset()).
905 		 */
906 		MPASS(nic_i == 0);
907 		MPASS(nm_i == kring->nr_hwtail);
908 	} else
909 		MPASS(nm_i == kring->nr_hwcur);
910 	DBG_COUNTER_INC(fl_refills);
911 	while (n > 0) {
912 #if IFLIB_DEBUG_COUNTERS
913 		if (++rf_count == 9)
914 			DBG_COUNTER_INC(fl_refills_large);
915 #endif
916 		nic_i_first = nic_i;
917 		for (i = 0; n > 0 && i < IFLIB_MAX_RX_REFRESH; n--, i++) {
918 			struct netmap_slot *slot = &ring->slot[nm_i];
919 			void *addr = PNMB(na, slot, &fl->ifl_bus_addrs[i]);
920 
921 			MPASS(i < IFLIB_MAX_RX_REFRESH);
922 
923 			if (addr == NETMAP_BUF_BASE(na)) /* bad buf */
924 			        return netmap_ring_reinit(kring);
925 
926 			fl->ifl_rxd_idxs[i] = nic_i;
927 
928 			if (__predict_false(init)) {
929 				netmap_load_map(na, fl->ifl_buf_tag,
930 				    map[nic_i], addr);
931 			} else if (slot->flags & NS_BUF_CHANGED) {
932 				/* buffer has changed, reload map */
933 				netmap_reload_map(na, fl->ifl_buf_tag,
934 				    map[nic_i], addr);
935 			}
936 			bus_dmamap_sync(fl->ifl_buf_tag, map[nic_i],
937 			    BUS_DMASYNC_PREREAD);
938 			slot->flags &= ~NS_BUF_CHANGED;
939 
940 			nm_i = nm_next(nm_i, lim);
941 			nic_i = nm_next(nic_i, lim);
942 		}
943 
944 		iru.iru_pidx = nic_i_first;
945 		iru.iru_count = i;
946 		ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
947 	}
948 	fl->ifl_pidx = nic_i;
949 	/*
950 	 * At the end of the loop we must have refilled everything
951 	 * we could possibly refill.
952 	 */
953 	MPASS(nm_i == kring->rhead);
954 	kring->nr_hwcur = nm_i;
955 
956 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
957 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
958 	ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, fl->ifl_id,
959 	    nm_prev(nic_i, lim));
960 	DBG_COUNTER_INC(rxd_flush);
961 
962 	return (0);
963 }
964 
965 #define NETMAP_TX_TIMER_US	90
966 
967 /*
968  * Reconcile kernel and user view of the transmit ring.
969  *
970  * All information is in the kring.
971  * Userspace wants to send packets up to the one before kring->rhead,
972  * kernel knows kring->nr_hwcur is the first unsent packet.
973  *
974  * Here we push packets out (as many as possible), and possibly
975  * reclaim buffers from previously completed transmission.
976  *
977  * The caller (netmap) guarantees that there is only one instance
978  * running at any time. Any interference with other driver
979  * methods should be handled by the individual drivers.
980  */
981 static int
982 iflib_netmap_txsync(struct netmap_kring *kring, int flags)
983 {
984 	struct netmap_adapter *na = kring->na;
985 	if_t ifp = na->ifp;
986 	struct netmap_ring *ring = kring->ring;
987 	u_int nm_i;	/* index into the netmap kring */
988 	u_int nic_i;	/* index into the NIC ring */
989 	u_int n;
990 	u_int const lim = kring->nkr_num_slots - 1;
991 	u_int const head = kring->rhead;
992 	struct if_pkt_info pi;
993 	int tx_pkts = 0, tx_bytes = 0;
994 
995 	/*
996 	 * interrupts on every tx packet are expensive so request
997 	 * them every half ring, or where NS_REPORT is set
998 	 */
999 	u_int report_frequency = kring->nkr_num_slots >> 1;
1000 	/* device-specific */
1001 	if_ctx_t ctx = ifp->if_softc;
1002 	iflib_txq_t txq = &ctx->ifc_txqs[kring->ring_id];
1003 
1004 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1005 	    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1006 
1007 	/*
1008 	 * First part: process new packets to send.
1009 	 * nm_i is the current index in the netmap kring,
1010 	 * nic_i is the corresponding index in the NIC ring.
1011 	 *
1012 	 * If we have packets to send (nm_i != head)
1013 	 * iterate over the netmap ring, fetch length and update
1014 	 * the corresponding slot in the NIC ring. Some drivers also
1015 	 * need to update the buffer's physical address in the NIC slot
1016 	 * even NS_BUF_CHANGED is not set (PNMB computes the addresses).
1017 	 *
1018 	 * The netmap_reload_map() calls is especially expensive,
1019 	 * even when (as in this case) the tag is 0, so do only
1020 	 * when the buffer has actually changed.
1021 	 *
1022 	 * If possible do not set the report/intr bit on all slots,
1023 	 * but only a few times per ring or when NS_REPORT is set.
1024 	 *
1025 	 * Finally, on 10G and faster drivers, it might be useful
1026 	 * to prefetch the next slot and txr entry.
1027 	 */
1028 
1029 	nm_i = kring->nr_hwcur;
1030 	if (nm_i != head) {	/* we have new packets to send */
1031 		uint32_t pkt_len = 0, seg_idx = 0;
1032 		int nic_i_start = -1, flags = 0;
1033 		pkt_info_zero(&pi);
1034 		pi.ipi_segs = txq->ift_segs;
1035 		pi.ipi_qsidx = kring->ring_id;
1036 		nic_i = netmap_idx_k2n(kring, nm_i);
1037 
1038 		__builtin_prefetch(&ring->slot[nm_i]);
1039 		__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i]);
1040 		__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i]);
1041 
1042 		for (n = 0; nm_i != head; n++) {
1043 			struct netmap_slot *slot = &ring->slot[nm_i];
1044 			u_int len = slot->len;
1045 			uint64_t paddr;
1046 			void *addr = PNMB(na, slot, &paddr);
1047 
1048 			flags |= (slot->flags & NS_REPORT ||
1049 				nic_i == 0 || nic_i == report_frequency) ?
1050 				IPI_TX_INTR : 0;
1051 
1052 			/*
1053 			 * If this is the first packet fragment, save the
1054 			 * index of the first NIC slot for later.
1055 			 */
1056 			if (nic_i_start < 0)
1057 				nic_i_start = nic_i;
1058 
1059 			pi.ipi_segs[seg_idx].ds_addr = paddr;
1060 			pi.ipi_segs[seg_idx].ds_len = len;
1061 			if (len) {
1062 				pkt_len += len;
1063 				seg_idx++;
1064 			}
1065 
1066 			if (!(slot->flags & NS_MOREFRAG)) {
1067 				pi.ipi_len = pkt_len;
1068 				pi.ipi_nsegs = seg_idx;
1069 				pi.ipi_pidx = nic_i_start;
1070 				pi.ipi_ndescs = 0;
1071 				pi.ipi_flags = flags;
1072 
1073 				/* Prepare the NIC TX ring. */
1074 				ctx->isc_txd_encap(ctx->ifc_softc, &pi);
1075 				DBG_COUNTER_INC(tx_encap);
1076 
1077 				/* Update transmit counters */
1078 				tx_bytes += pi.ipi_len;
1079 				tx_pkts++;
1080 
1081 				/* Reinit per-packet info for the next one. */
1082 				flags = seg_idx = pkt_len = 0;
1083 				nic_i_start = -1;
1084 			}
1085 
1086 			/* prefetch for next round */
1087 			__builtin_prefetch(&ring->slot[nm_i + 1]);
1088 			__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i + 1]);
1089 			__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i + 1]);
1090 
1091 			NM_CHECK_ADDR_LEN(na, addr, len);
1092 
1093 			if (slot->flags & NS_BUF_CHANGED) {
1094 				/* buffer has changed, reload map */
1095 				netmap_reload_map(na, txq->ift_buf_tag,
1096 				    txq->ift_sds.ifsd_map[nic_i], addr);
1097 			}
1098 			/* make sure changes to the buffer are synced */
1099 			bus_dmamap_sync(txq->ift_buf_tag,
1100 			    txq->ift_sds.ifsd_map[nic_i],
1101 			    BUS_DMASYNC_PREWRITE);
1102 
1103 			slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED | NS_MOREFRAG);
1104 			nm_i = nm_next(nm_i, lim);
1105 			nic_i = nm_next(nic_i, lim);
1106 		}
1107 		kring->nr_hwcur = nm_i;
1108 
1109 		/* synchronize the NIC ring */
1110 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1111 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1112 
1113 		/* (re)start the tx unit up to slot nic_i (excluded) */
1114 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, nic_i);
1115 	}
1116 
1117 	/*
1118 	 * Second part: reclaim buffers for completed transmissions.
1119 	 *
1120 	 * If there are unclaimed buffers, attempt to reclaim them.
1121 	 * If we don't manage to reclaim them all, and TX IRQs are not in use,
1122 	 * trigger a per-tx-queue timer to try again later.
1123 	 */
1124 	if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1125 		if (iflib_tx_credits_update(ctx, txq)) {
1126 			/* some tx completed, increment avail */
1127 			nic_i = txq->ift_cidx_processed;
1128 			kring->nr_hwtail = nm_prev(netmap_idx_n2k(kring, nic_i), lim);
1129 		}
1130 	}
1131 
1132 	if (!(ctx->ifc_flags & IFC_NETMAP_TX_IRQ))
1133 		if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) {
1134 			callout_reset_sbt_on(&txq->ift_netmap_timer,
1135 			    NETMAP_TX_TIMER_US * SBT_1US, SBT_1US,
1136 			    iflib_netmap_timer, txq,
1137 			    txq->ift_netmap_timer.c_cpu, 0);
1138 		}
1139 
1140 	if_inc_counter(ifp, IFCOUNTER_OBYTES, tx_bytes);
1141 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, tx_pkts);
1142 
1143 	return (0);
1144 }
1145 
1146 /*
1147  * Reconcile kernel and user view of the receive ring.
1148  * Same as for the txsync, this routine must be efficient.
1149  * The caller guarantees a single invocations, but races against
1150  * the rest of the driver should be handled here.
1151  *
1152  * On call, kring->rhead is the first packet that userspace wants
1153  * to keep, and kring->rcur is the wakeup point.
1154  * The kernel has previously reported packets up to kring->rtail.
1155  *
1156  * If (flags & NAF_FORCE_READ) also check for incoming packets irrespective
1157  * of whether or not we received an interrupt.
1158  */
1159 static int
1160 iflib_netmap_rxsync(struct netmap_kring *kring, int flags)
1161 {
1162 	struct netmap_adapter *na = kring->na;
1163 	struct netmap_ring *ring = kring->ring;
1164 	if_t ifp = na->ifp;
1165 	uint32_t nm_i;	/* index into the netmap ring */
1166 	uint32_t nic_i;	/* index into the NIC ring */
1167 	u_int n;
1168 	u_int const lim = kring->nkr_num_slots - 1;
1169 	int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
1170 	int i = 0, rx_bytes = 0, rx_pkts = 0;
1171 
1172 	if_ctx_t ctx = ifp->if_softc;
1173 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1174 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1175 	iflib_rxq_t rxq = &ctx->ifc_rxqs[kring->ring_id];
1176 	iflib_fl_t fl = &rxq->ifr_fl[0];
1177 	struct if_rxd_info ri;
1178 	qidx_t *cidxp;
1179 
1180 	/*
1181 	 * netmap only uses free list 0, to avoid out of order consumption
1182 	 * of receive buffers
1183 	 */
1184 
1185 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
1186 	    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1187 
1188 	/*
1189 	 * First part: import newly received packets.
1190 	 *
1191 	 * nm_i is the index of the next free slot in the netmap ring,
1192 	 * nic_i is the index of the next received packet in the NIC ring
1193 	 * (or in the free list 0 if IFLIB_HAS_RXCQ is set), and they may
1194 	 * differ in case if_init() has been called while
1195 	 * in netmap mode. For the receive ring we have
1196 	 *
1197 	 *	nic_i = fl->ifl_cidx;
1198 	 *	nm_i = kring->nr_hwtail (previous)
1199 	 * and
1200 	 *	nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1201 	 *
1202 	 * fl->ifl_cidx is set to 0 on a ring reinit
1203 	 */
1204 	if (netmap_no_pendintr || force_update) {
1205 		uint32_t hwtail_lim = nm_prev(kring->nr_hwcur, lim);
1206 		bool have_rxcq = sctx->isc_flags & IFLIB_HAS_RXCQ;
1207 		int crclen = iflib_crcstrip ? 0 : 4;
1208 		int error, avail;
1209 
1210 		/*
1211 		 * For the free list consumer index, we use the same
1212 		 * logic as in iflib_rxeof().
1213 		 */
1214 		if (have_rxcq)
1215 			cidxp = &rxq->ifr_cq_cidx;
1216 		else
1217 			cidxp = &fl->ifl_cidx;
1218 		avail = ctx->isc_rxd_available(ctx->ifc_softc,
1219 		    rxq->ifr_id, *cidxp, USHRT_MAX);
1220 
1221 		nic_i = fl->ifl_cidx;
1222 		nm_i = netmap_idx_n2k(kring, nic_i);
1223 		MPASS(nm_i == kring->nr_hwtail);
1224 		for (n = 0; avail > 0 && nm_i != hwtail_lim; n++, avail--) {
1225 			rxd_info_zero(&ri);
1226 			ri.iri_frags = rxq->ifr_frags;
1227 			ri.iri_qsidx = kring->ring_id;
1228 			ri.iri_ifp = ctx->ifc_ifp;
1229 			ri.iri_cidx = *cidxp;
1230 
1231 			error = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
1232 			for (i = 0; i < ri.iri_nfrags; i++) {
1233 				if (error) {
1234 					ring->slot[nm_i].len = 0;
1235 					ring->slot[nm_i].flags = 0;
1236 				} else {
1237 					ring->slot[nm_i].len = ri.iri_frags[i].irf_len;
1238 					if (i == (ri.iri_nfrags - 1)) {
1239 						ring->slot[nm_i].len -= crclen;
1240 						ring->slot[nm_i].flags = 0;
1241 
1242 						/* Update receive counters */
1243 						rx_bytes += ri.iri_len;
1244 						rx_pkts++;
1245 					} else
1246 						ring->slot[nm_i].flags = NS_MOREFRAG;
1247 				}
1248 
1249 				bus_dmamap_sync(fl->ifl_buf_tag,
1250 				    fl->ifl_sds.ifsd_map[nic_i], BUS_DMASYNC_POSTREAD);
1251 				nm_i = nm_next(nm_i, lim);
1252 				fl->ifl_cidx = nic_i = nm_next(nic_i, lim);
1253 			}
1254 
1255 			if (have_rxcq) {
1256 				*cidxp = ri.iri_cidx;
1257 				while (*cidxp >= scctx->isc_nrxd[0])
1258 					*cidxp -= scctx->isc_nrxd[0];
1259 			}
1260 
1261 		}
1262 		if (n) { /* update the state variables */
1263 			if (netmap_no_pendintr && !force_update) {
1264 				/* diagnostics */
1265 				iflib_rx_miss ++;
1266 				iflib_rx_miss_bufs += n;
1267 			}
1268 			kring->nr_hwtail = nm_i;
1269 		}
1270 		kring->nr_kflags &= ~NKR_PENDINTR;
1271 	}
1272 	/*
1273 	 * Second part: skip past packets that userspace has released.
1274 	 * (kring->nr_hwcur to head excluded),
1275 	 * and make the buffers available for reception.
1276 	 * As usual nm_i is the index in the netmap ring,
1277 	 * nic_i is the index in the NIC ring, and
1278 	 * nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1279 	 */
1280 	netmap_fl_refill(rxq, kring, false);
1281 
1282 	if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes);
1283 	if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts);
1284 
1285 	return (0);
1286 }
1287 
1288 static void
1289 iflib_netmap_intr(struct netmap_adapter *na, int onoff)
1290 {
1291 	if_ctx_t ctx = na->ifp->if_softc;
1292 
1293 	CTX_LOCK(ctx);
1294 	if (onoff) {
1295 		IFDI_INTR_ENABLE(ctx);
1296 	} else {
1297 		IFDI_INTR_DISABLE(ctx);
1298 	}
1299 	CTX_UNLOCK(ctx);
1300 }
1301 
1302 static int
1303 iflib_netmap_attach(if_ctx_t ctx)
1304 {
1305 	struct netmap_adapter na;
1306 
1307 	bzero(&na, sizeof(na));
1308 
1309 	na.ifp = ctx->ifc_ifp;
1310 	na.na_flags = NAF_BDG_MAYSLEEP | NAF_MOREFRAG;
1311 	MPASS(ctx->ifc_softc_ctx.isc_ntxqsets);
1312 	MPASS(ctx->ifc_softc_ctx.isc_nrxqsets);
1313 
1314 	na.num_tx_desc = iflib_num_tx_descs(ctx);
1315 	na.num_rx_desc = iflib_num_rx_descs(ctx);
1316 	na.nm_txsync = iflib_netmap_txsync;
1317 	na.nm_rxsync = iflib_netmap_rxsync;
1318 	na.nm_register = iflib_netmap_register;
1319 	na.nm_intr = iflib_netmap_intr;
1320 	na.nm_config = iflib_netmap_config;
1321 	na.num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets;
1322 	na.num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets;
1323 	return (netmap_attach(&na));
1324 }
1325 
1326 static int
1327 iflib_netmap_txq_init(if_ctx_t ctx, iflib_txq_t txq)
1328 {
1329 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1330 	struct netmap_slot *slot;
1331 
1332 	slot = netmap_reset(na, NR_TX, txq->ift_id, 0);
1333 	if (slot == NULL)
1334 		return (0);
1335 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxd[0]; i++) {
1336 		/*
1337 		 * In netmap mode, set the map for the packet buffer.
1338 		 * NOTE: Some drivers (not this one) also need to set
1339 		 * the physical buffer address in the NIC ring.
1340 		 * netmap_idx_n2k() maps a nic index, i, into the corresponding
1341 		 * netmap slot index, si
1342 		 */
1343 		int si = netmap_idx_n2k(na->tx_rings[txq->ift_id], i);
1344 		netmap_load_map(na, txq->ift_buf_tag, txq->ift_sds.ifsd_map[i],
1345 		    NMB(na, slot + si));
1346 	}
1347 	return (1);
1348 }
1349 
1350 static int
1351 iflib_netmap_rxq_init(if_ctx_t ctx, iflib_rxq_t rxq)
1352 {
1353 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1354 	struct netmap_kring *kring;
1355 	struct netmap_slot *slot;
1356 
1357 	slot = netmap_reset(na, NR_RX, rxq->ifr_id, 0);
1358 	if (slot == NULL)
1359 		return (0);
1360 	kring = na->rx_rings[rxq->ifr_id];
1361 	netmap_fl_refill(rxq, kring, true);
1362 	return (1);
1363 }
1364 
1365 static void
1366 iflib_netmap_timer(void *arg)
1367 {
1368 	iflib_txq_t txq = arg;
1369 	if_ctx_t ctx = txq->ift_ctx;
1370 
1371 	/*
1372 	 * Wake up the netmap application, to give it a chance to
1373 	 * call txsync and reclaim more completed TX buffers.
1374 	 */
1375 	netmap_tx_irq(ctx->ifc_ifp, txq->ift_id);
1376 }
1377 
1378 #define iflib_netmap_detach(ifp) netmap_detach(ifp)
1379 
1380 #else
1381 #define iflib_netmap_txq_init(ctx, txq) (0)
1382 #define iflib_netmap_rxq_init(ctx, rxq) (0)
1383 #define iflib_netmap_detach(ifp)
1384 #define netmap_enable_all_rings(ifp)
1385 #define netmap_disable_all_rings(ifp)
1386 
1387 #define iflib_netmap_attach(ctx) (0)
1388 #define netmap_rx_irq(ifp, qid, budget) (0)
1389 #endif
1390 
1391 #if defined(__i386__) || defined(__amd64__)
1392 static __inline void
1393 prefetch(void *x)
1394 {
1395 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1396 }
1397 static __inline void
1398 prefetch2cachelines(void *x)
1399 {
1400 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1401 #if (CACHE_LINE_SIZE < 128)
1402 	__asm volatile("prefetcht0 %0" :: "m" (*(((unsigned long *)x)+CACHE_LINE_SIZE/(sizeof(unsigned long)))));
1403 #endif
1404 }
1405 #else
1406 #define prefetch(x)
1407 #define prefetch2cachelines(x)
1408 #endif
1409 
1410 static void
1411 iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid)
1412 {
1413 	iflib_fl_t fl;
1414 
1415 	fl = &rxq->ifr_fl[flid];
1416 	iru->iru_paddrs = fl->ifl_bus_addrs;
1417 	iru->iru_idxs = fl->ifl_rxd_idxs;
1418 	iru->iru_qsidx = rxq->ifr_id;
1419 	iru->iru_buf_size = fl->ifl_buf_size;
1420 	iru->iru_flidx = fl->ifl_id;
1421 }
1422 
1423 static void
1424 _iflib_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg, int err)
1425 {
1426 	if (err)
1427 		return;
1428 	*(bus_addr_t *) arg = segs[0].ds_addr;
1429 }
1430 
1431 int
1432 iflib_dma_alloc_align(if_ctx_t ctx, int size, int align, iflib_dma_info_t dma, int mapflags)
1433 {
1434 	int err;
1435 	device_t dev = ctx->ifc_dev;
1436 
1437 	err = bus_dma_tag_create(bus_get_dma_tag(dev),	/* parent */
1438 				align, 0,		/* alignment, bounds */
1439 				BUS_SPACE_MAXADDR,	/* lowaddr */
1440 				BUS_SPACE_MAXADDR,	/* highaddr */
1441 				NULL, NULL,		/* filter, filterarg */
1442 				size,			/* maxsize */
1443 				1,			/* nsegments */
1444 				size,			/* maxsegsize */
1445 				BUS_DMA_ALLOCNOW,	/* flags */
1446 				NULL,			/* lockfunc */
1447 				NULL,			/* lockarg */
1448 				&dma->idi_tag);
1449 	if (err) {
1450 		device_printf(dev,
1451 		    "%s: bus_dma_tag_create failed: %d\n",
1452 		    __func__, err);
1453 		goto fail_0;
1454 	}
1455 
1456 	err = bus_dmamem_alloc(dma->idi_tag, (void**) &dma->idi_vaddr,
1457 	    BUS_DMA_NOWAIT | BUS_DMA_COHERENT | BUS_DMA_ZERO, &dma->idi_map);
1458 	if (err) {
1459 		device_printf(dev,
1460 		    "%s: bus_dmamem_alloc(%ju) failed: %d\n",
1461 		    __func__, (uintmax_t)size, err);
1462 		goto fail_1;
1463 	}
1464 
1465 	dma->idi_paddr = IF_BAD_DMA;
1466 	err = bus_dmamap_load(dma->idi_tag, dma->idi_map, dma->idi_vaddr,
1467 	    size, _iflib_dmamap_cb, &dma->idi_paddr, mapflags | BUS_DMA_NOWAIT);
1468 	if (err || dma->idi_paddr == IF_BAD_DMA) {
1469 		device_printf(dev,
1470 		    "%s: bus_dmamap_load failed: %d\n",
1471 		    __func__, err);
1472 		goto fail_2;
1473 	}
1474 
1475 	dma->idi_size = size;
1476 	return (0);
1477 
1478 fail_2:
1479 	bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1480 fail_1:
1481 	bus_dma_tag_destroy(dma->idi_tag);
1482 fail_0:
1483 	dma->idi_tag = NULL;
1484 
1485 	return (err);
1486 }
1487 
1488 int
1489 iflib_dma_alloc(if_ctx_t ctx, int size, iflib_dma_info_t dma, int mapflags)
1490 {
1491 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1492 
1493 	KASSERT(sctx->isc_q_align != 0, ("alignment value not initialized"));
1494 
1495 	return (iflib_dma_alloc_align(ctx, size, sctx->isc_q_align, dma, mapflags));
1496 }
1497 
1498 int
1499 iflib_dma_alloc_multi(if_ctx_t ctx, int *sizes, iflib_dma_info_t *dmalist, int mapflags, int count)
1500 {
1501 	int i, err;
1502 	iflib_dma_info_t *dmaiter;
1503 
1504 	dmaiter = dmalist;
1505 	for (i = 0; i < count; i++, dmaiter++) {
1506 		if ((err = iflib_dma_alloc(ctx, sizes[i], *dmaiter, mapflags)) != 0)
1507 			break;
1508 	}
1509 	if (err)
1510 		iflib_dma_free_multi(dmalist, i);
1511 	return (err);
1512 }
1513 
1514 void
1515 iflib_dma_free(iflib_dma_info_t dma)
1516 {
1517 	if (dma->idi_tag == NULL)
1518 		return;
1519 	if (dma->idi_paddr != IF_BAD_DMA) {
1520 		bus_dmamap_sync(dma->idi_tag, dma->idi_map,
1521 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1522 		bus_dmamap_unload(dma->idi_tag, dma->idi_map);
1523 		dma->idi_paddr = IF_BAD_DMA;
1524 	}
1525 	if (dma->idi_vaddr != NULL) {
1526 		bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1527 		dma->idi_vaddr = NULL;
1528 	}
1529 	bus_dma_tag_destroy(dma->idi_tag);
1530 	dma->idi_tag = NULL;
1531 }
1532 
1533 void
1534 iflib_dma_free_multi(iflib_dma_info_t *dmalist, int count)
1535 {
1536 	int i;
1537 	iflib_dma_info_t *dmaiter = dmalist;
1538 
1539 	for (i = 0; i < count; i++, dmaiter++)
1540 		iflib_dma_free(*dmaiter);
1541 }
1542 
1543 static int
1544 iflib_fast_intr(void *arg)
1545 {
1546 	iflib_filter_info_t info = arg;
1547 	struct grouptask *gtask = info->ifi_task;
1548 	int result;
1549 
1550 	DBG_COUNTER_INC(fast_intrs);
1551 	if (info->ifi_filter != NULL) {
1552 		result = info->ifi_filter(info->ifi_filter_arg);
1553 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1554 			return (result);
1555 	}
1556 
1557 	GROUPTASK_ENQUEUE(gtask);
1558 	return (FILTER_HANDLED);
1559 }
1560 
1561 static int
1562 iflib_fast_intr_rxtx(void *arg)
1563 {
1564 	iflib_filter_info_t info = arg;
1565 	struct grouptask *gtask = info->ifi_task;
1566 	if_ctx_t ctx;
1567 	iflib_rxq_t rxq = (iflib_rxq_t)info->ifi_ctx;
1568 	iflib_txq_t txq;
1569 	void *sc;
1570 	int i, cidx, result;
1571 	qidx_t txqid;
1572 	bool intr_enable, intr_legacy;
1573 
1574 	DBG_COUNTER_INC(fast_intrs);
1575 	if (info->ifi_filter != NULL) {
1576 		result = info->ifi_filter(info->ifi_filter_arg);
1577 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1578 			return (result);
1579 	}
1580 
1581 	ctx = rxq->ifr_ctx;
1582 	sc = ctx->ifc_softc;
1583 	intr_enable = false;
1584 	intr_legacy = !!(ctx->ifc_flags & IFC_LEGACY);
1585 	MPASS(rxq->ifr_ntxqirq);
1586 	for (i = 0; i < rxq->ifr_ntxqirq; i++) {
1587 		txqid = rxq->ifr_txqid[i];
1588 		txq = &ctx->ifc_txqs[txqid];
1589 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
1590 		    BUS_DMASYNC_POSTREAD);
1591 		if (!ctx->isc_txd_credits_update(sc, txqid, false)) {
1592 			if (intr_legacy)
1593 				intr_enable = true;
1594 			else
1595 				IFDI_TX_QUEUE_INTR_ENABLE(ctx, txqid);
1596 			continue;
1597 		}
1598 		GROUPTASK_ENQUEUE(&txq->ift_task);
1599 	}
1600 	if (ctx->ifc_sctx->isc_flags & IFLIB_HAS_RXCQ)
1601 		cidx = rxq->ifr_cq_cidx;
1602 	else
1603 		cidx = rxq->ifr_fl[0].ifl_cidx;
1604 	if (iflib_rxd_avail(ctx, rxq, cidx, 1))
1605 		GROUPTASK_ENQUEUE(gtask);
1606 	else {
1607 		if (intr_legacy)
1608 			intr_enable = true;
1609 		else
1610 			IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
1611 		DBG_COUNTER_INC(rx_intr_enables);
1612 	}
1613 	if (intr_enable)
1614 		IFDI_INTR_ENABLE(ctx);
1615 	return (FILTER_HANDLED);
1616 }
1617 
1618 static int
1619 iflib_fast_intr_ctx(void *arg)
1620 {
1621 	iflib_filter_info_t info = arg;
1622 	struct grouptask *gtask = info->ifi_task;
1623 	int result;
1624 
1625 	DBG_COUNTER_INC(fast_intrs);
1626 	if (info->ifi_filter != NULL) {
1627 		result = info->ifi_filter(info->ifi_filter_arg);
1628 		if ((result & FILTER_SCHEDULE_THREAD) == 0)
1629 			return (result);
1630 	}
1631 
1632 	GROUPTASK_ENQUEUE(gtask);
1633 	return (FILTER_HANDLED);
1634 }
1635 
1636 static int
1637 _iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
1638 		 driver_filter_t filter, driver_intr_t handler, void *arg,
1639 		 const char *name)
1640 {
1641 	struct resource *res;
1642 	void *tag = NULL;
1643 	device_t dev = ctx->ifc_dev;
1644 	int flags, i, rc;
1645 
1646 	flags = RF_ACTIVE;
1647 	if (ctx->ifc_flags & IFC_LEGACY)
1648 		flags |= RF_SHAREABLE;
1649 	MPASS(rid < 512);
1650 	i = rid;
1651 	res = bus_alloc_resource_any(dev, SYS_RES_IRQ, &i, flags);
1652 	if (res == NULL) {
1653 		device_printf(dev,
1654 		    "failed to allocate IRQ for rid %d, name %s.\n", rid, name);
1655 		return (ENOMEM);
1656 	}
1657 	irq->ii_res = res;
1658 	KASSERT(filter == NULL || handler == NULL, ("filter and handler can't both be non-NULL"));
1659 	rc = bus_setup_intr(dev, res, INTR_MPSAFE | INTR_TYPE_NET,
1660 						filter, handler, arg, &tag);
1661 	if (rc != 0) {
1662 		device_printf(dev,
1663 		    "failed to setup interrupt for rid %d, name %s: %d\n",
1664 					  rid, name ? name : "unknown", rc);
1665 		return (rc);
1666 	} else if (name)
1667 		bus_describe_intr(dev, res, tag, "%s", name);
1668 
1669 	irq->ii_tag = tag;
1670 	return (0);
1671 }
1672 
1673 /*********************************************************************
1674  *
1675  *  Allocate DMA resources for TX buffers as well as memory for the TX
1676  *  mbuf map.  TX DMA maps (non-TSO/TSO) and TX mbuf map are kept in a
1677  *  iflib_sw_tx_desc_array structure, storing all the information that
1678  *  is needed to transmit a packet on the wire.  This is called only
1679  *  once at attach, setup is done every reset.
1680  *
1681  **********************************************************************/
1682 static int
1683 iflib_txsd_alloc(iflib_txq_t txq)
1684 {
1685 	if_ctx_t ctx = txq->ift_ctx;
1686 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1687 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1688 	device_t dev = ctx->ifc_dev;
1689 	bus_size_t tsomaxsize;
1690 	int err, nsegments, ntsosegments;
1691 	bool tso;
1692 
1693 	nsegments = scctx->isc_tx_nsegments;
1694 	ntsosegments = scctx->isc_tx_tso_segments_max;
1695 	tsomaxsize = scctx->isc_tx_tso_size_max;
1696 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_VLAN_MTU)
1697 		tsomaxsize += sizeof(struct ether_vlan_header);
1698 	MPASS(scctx->isc_ntxd[0] > 0);
1699 	MPASS(scctx->isc_ntxd[txq->ift_br_offset] > 0);
1700 	MPASS(nsegments > 0);
1701 	if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) {
1702 		MPASS(ntsosegments > 0);
1703 		MPASS(sctx->isc_tso_maxsize >= tsomaxsize);
1704 	}
1705 
1706 	/*
1707 	 * Set up DMA tags for TX buffers.
1708 	 */
1709 	if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1710 			       1, 0,			/* alignment, bounds */
1711 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1712 			       BUS_SPACE_MAXADDR,	/* highaddr */
1713 			       NULL, NULL,		/* filter, filterarg */
1714 			       sctx->isc_tx_maxsize,		/* maxsize */
1715 			       nsegments,	/* nsegments */
1716 			       sctx->isc_tx_maxsegsize,	/* maxsegsize */
1717 			       0,			/* flags */
1718 			       NULL,			/* lockfunc */
1719 			       NULL,			/* lockfuncarg */
1720 			       &txq->ift_buf_tag))) {
1721 		device_printf(dev,"Unable to allocate TX DMA tag: %d\n", err);
1722 		device_printf(dev,"maxsize: %ju nsegments: %d maxsegsize: %ju\n",
1723 		    (uintmax_t)sctx->isc_tx_maxsize, nsegments, (uintmax_t)sctx->isc_tx_maxsegsize);
1724 		goto fail;
1725 	}
1726 	tso = (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) != 0;
1727 	if (tso && (err = bus_dma_tag_create(bus_get_dma_tag(dev),
1728 			       1, 0,			/* alignment, bounds */
1729 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1730 			       BUS_SPACE_MAXADDR,	/* highaddr */
1731 			       NULL, NULL,		/* filter, filterarg */
1732 			       tsomaxsize,		/* maxsize */
1733 			       ntsosegments,	/* nsegments */
1734 			       sctx->isc_tso_maxsegsize,/* maxsegsize */
1735 			       0,			/* flags */
1736 			       NULL,			/* lockfunc */
1737 			       NULL,			/* lockfuncarg */
1738 			       &txq->ift_tso_buf_tag))) {
1739 		device_printf(dev, "Unable to allocate TSO TX DMA tag: %d\n",
1740 		    err);
1741 		goto fail;
1742 	}
1743 
1744 	/* Allocate memory for the TX mbuf map. */
1745 	if (!(txq->ift_sds.ifsd_m =
1746 	    (struct mbuf **) malloc(sizeof(struct mbuf *) *
1747 	    scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1748 		device_printf(dev, "Unable to allocate TX mbuf map memory\n");
1749 		err = ENOMEM;
1750 		goto fail;
1751 	}
1752 
1753 	/*
1754 	 * Create the DMA maps for TX buffers.
1755 	 */
1756 	if ((txq->ift_sds.ifsd_map = (bus_dmamap_t *)malloc(
1757 	    sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset],
1758 	    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
1759 		device_printf(dev,
1760 		    "Unable to allocate TX buffer DMA map memory\n");
1761 		err = ENOMEM;
1762 		goto fail;
1763 	}
1764 	if (tso && (txq->ift_sds.ifsd_tso_map = (bus_dmamap_t *)malloc(
1765 	    sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset],
1766 	    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
1767 		device_printf(dev,
1768 		    "Unable to allocate TSO TX buffer map memory\n");
1769 		err = ENOMEM;
1770 		goto fail;
1771 	}
1772 	for (int i = 0; i < scctx->isc_ntxd[txq->ift_br_offset]; i++) {
1773 		err = bus_dmamap_create(txq->ift_buf_tag, 0,
1774 		    &txq->ift_sds.ifsd_map[i]);
1775 		if (err != 0) {
1776 			device_printf(dev, "Unable to create TX DMA map\n");
1777 			goto fail;
1778 		}
1779 		if (!tso)
1780 			continue;
1781 		err = bus_dmamap_create(txq->ift_tso_buf_tag, 0,
1782 		    &txq->ift_sds.ifsd_tso_map[i]);
1783 		if (err != 0) {
1784 			device_printf(dev, "Unable to create TSO TX DMA map\n");
1785 			goto fail;
1786 		}
1787 	}
1788 	return (0);
1789 fail:
1790 	/* We free all, it handles case where we are in the middle */
1791 	iflib_tx_structures_free(ctx);
1792 	return (err);
1793 }
1794 
1795 static void
1796 iflib_txsd_destroy(if_ctx_t ctx, iflib_txq_t txq, int i)
1797 {
1798 	bus_dmamap_t map;
1799 
1800 	if (txq->ift_sds.ifsd_map != NULL) {
1801 		map = txq->ift_sds.ifsd_map[i];
1802 		bus_dmamap_sync(txq->ift_buf_tag, map, BUS_DMASYNC_POSTWRITE);
1803 		bus_dmamap_unload(txq->ift_buf_tag, map);
1804 		bus_dmamap_destroy(txq->ift_buf_tag, map);
1805 		txq->ift_sds.ifsd_map[i] = NULL;
1806 	}
1807 
1808 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1809 		map = txq->ift_sds.ifsd_tso_map[i];
1810 		bus_dmamap_sync(txq->ift_tso_buf_tag, map,
1811 		    BUS_DMASYNC_POSTWRITE);
1812 		bus_dmamap_unload(txq->ift_tso_buf_tag, map);
1813 		bus_dmamap_destroy(txq->ift_tso_buf_tag, map);
1814 		txq->ift_sds.ifsd_tso_map[i] = NULL;
1815 	}
1816 }
1817 
1818 static void
1819 iflib_txq_destroy(iflib_txq_t txq)
1820 {
1821 	if_ctx_t ctx = txq->ift_ctx;
1822 
1823 	for (int i = 0; i < txq->ift_size; i++)
1824 		iflib_txsd_destroy(ctx, txq, i);
1825 
1826 	if (txq->ift_br != NULL) {
1827 		ifmp_ring_free(txq->ift_br);
1828 		txq->ift_br = NULL;
1829 	}
1830 
1831 	mtx_destroy(&txq->ift_mtx);
1832 
1833 	if (txq->ift_sds.ifsd_map != NULL) {
1834 		free(txq->ift_sds.ifsd_map, M_IFLIB);
1835 		txq->ift_sds.ifsd_map = NULL;
1836 	}
1837 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1838 		free(txq->ift_sds.ifsd_tso_map, M_IFLIB);
1839 		txq->ift_sds.ifsd_tso_map = NULL;
1840 	}
1841 	if (txq->ift_sds.ifsd_m != NULL) {
1842 		free(txq->ift_sds.ifsd_m, M_IFLIB);
1843 		txq->ift_sds.ifsd_m = NULL;
1844 	}
1845 	if (txq->ift_buf_tag != NULL) {
1846 		bus_dma_tag_destroy(txq->ift_buf_tag);
1847 		txq->ift_buf_tag = NULL;
1848 	}
1849 	if (txq->ift_tso_buf_tag != NULL) {
1850 		bus_dma_tag_destroy(txq->ift_tso_buf_tag);
1851 		txq->ift_tso_buf_tag = NULL;
1852 	}
1853 	if (txq->ift_ifdi != NULL) {
1854 		free(txq->ift_ifdi, M_IFLIB);
1855 	}
1856 }
1857 
1858 static void
1859 iflib_txsd_free(if_ctx_t ctx, iflib_txq_t txq, int i)
1860 {
1861 	struct mbuf **mp;
1862 
1863 	mp = &txq->ift_sds.ifsd_m[i];
1864 	if (*mp == NULL)
1865 		return;
1866 
1867 	if (txq->ift_sds.ifsd_map != NULL) {
1868 		bus_dmamap_sync(txq->ift_buf_tag,
1869 		    txq->ift_sds.ifsd_map[i], BUS_DMASYNC_POSTWRITE);
1870 		bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[i]);
1871 	}
1872 	if (txq->ift_sds.ifsd_tso_map != NULL) {
1873 		bus_dmamap_sync(txq->ift_tso_buf_tag,
1874 		    txq->ift_sds.ifsd_tso_map[i], BUS_DMASYNC_POSTWRITE);
1875 		bus_dmamap_unload(txq->ift_tso_buf_tag,
1876 		    txq->ift_sds.ifsd_tso_map[i]);
1877 	}
1878 	m_freem(*mp);
1879 	DBG_COUNTER_INC(tx_frees);
1880 	*mp = NULL;
1881 }
1882 
1883 static int
1884 iflib_txq_setup(iflib_txq_t txq)
1885 {
1886 	if_ctx_t ctx = txq->ift_ctx;
1887 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1888 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1889 	iflib_dma_info_t di;
1890 	int i;
1891 
1892 	/* Set number of descriptors available */
1893 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
1894 	/* XXX make configurable */
1895 	txq->ift_update_freq = IFLIB_DEFAULT_TX_UPDATE_FREQ;
1896 
1897 	/* Reset indices */
1898 	txq->ift_cidx_processed = 0;
1899 	txq->ift_pidx = txq->ift_cidx = txq->ift_npending = 0;
1900 	txq->ift_size = scctx->isc_ntxd[txq->ift_br_offset];
1901 
1902 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1903 		bzero((void *)di->idi_vaddr, di->idi_size);
1904 
1905 	IFDI_TXQ_SETUP(ctx, txq->ift_id);
1906 	for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++)
1907 		bus_dmamap_sync(di->idi_tag, di->idi_map,
1908 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1909 	return (0);
1910 }
1911 
1912 /*********************************************************************
1913  *
1914  *  Allocate DMA resources for RX buffers as well as memory for the RX
1915  *  mbuf map, direct RX cluster pointer map and RX cluster bus address
1916  *  map.  RX DMA map, RX mbuf map, direct RX cluster pointer map and
1917  *  RX cluster map are kept in a iflib_sw_rx_desc_array structure.
1918  *  Since we use use one entry in iflib_sw_rx_desc_array per received
1919  *  packet, the maximum number of entries we'll need is equal to the
1920  *  number of hardware receive descriptors that we've allocated.
1921  *
1922  **********************************************************************/
1923 static int
1924 iflib_rxsd_alloc(iflib_rxq_t rxq)
1925 {
1926 	if_ctx_t ctx = rxq->ifr_ctx;
1927 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1928 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1929 	device_t dev = ctx->ifc_dev;
1930 	iflib_fl_t fl;
1931 	int			err;
1932 
1933 	MPASS(scctx->isc_nrxd[0] > 0);
1934 	MPASS(scctx->isc_nrxd[rxq->ifr_fl_offset] > 0);
1935 
1936 	fl = rxq->ifr_fl;
1937 	for (int i = 0; i <  rxq->ifr_nfl; i++, fl++) {
1938 		fl->ifl_size = scctx->isc_nrxd[rxq->ifr_fl_offset]; /* this isn't necessarily the same */
1939 		/* Set up DMA tag for RX buffers. */
1940 		err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1941 					 1, 0,			/* alignment, bounds */
1942 					 BUS_SPACE_MAXADDR,	/* lowaddr */
1943 					 BUS_SPACE_MAXADDR,	/* highaddr */
1944 					 NULL, NULL,		/* filter, filterarg */
1945 					 sctx->isc_rx_maxsize,	/* maxsize */
1946 					 sctx->isc_rx_nsegments,	/* nsegments */
1947 					 sctx->isc_rx_maxsegsize,	/* maxsegsize */
1948 					 0,			/* flags */
1949 					 NULL,			/* lockfunc */
1950 					 NULL,			/* lockarg */
1951 					 &fl->ifl_buf_tag);
1952 		if (err) {
1953 			device_printf(dev,
1954 			    "Unable to allocate RX DMA tag: %d\n", err);
1955 			goto fail;
1956 		}
1957 
1958 		/* Allocate memory for the RX mbuf map. */
1959 		if (!(fl->ifl_sds.ifsd_m =
1960 		      (struct mbuf **) malloc(sizeof(struct mbuf *) *
1961 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1962 			device_printf(dev,
1963 			    "Unable to allocate RX mbuf map memory\n");
1964 			err = ENOMEM;
1965 			goto fail;
1966 		}
1967 
1968 		/* Allocate memory for the direct RX cluster pointer map. */
1969 		if (!(fl->ifl_sds.ifsd_cl =
1970 		      (caddr_t *) malloc(sizeof(caddr_t) *
1971 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1972 			device_printf(dev,
1973 			    "Unable to allocate RX cluster map memory\n");
1974 			err = ENOMEM;
1975 			goto fail;
1976 		}
1977 
1978 		/* Allocate memory for the RX cluster bus address map. */
1979 		if (!(fl->ifl_sds.ifsd_ba =
1980 		      (bus_addr_t *) malloc(sizeof(bus_addr_t) *
1981 					      scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1982 			device_printf(dev,
1983 			    "Unable to allocate RX bus address map memory\n");
1984 			err = ENOMEM;
1985 			goto fail;
1986 		}
1987 
1988 		/*
1989 		 * Create the DMA maps for RX buffers.
1990 		 */
1991 		if (!(fl->ifl_sds.ifsd_map =
1992 		      (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) {
1993 			device_printf(dev,
1994 			    "Unable to allocate RX buffer DMA map memory\n");
1995 			err = ENOMEM;
1996 			goto fail;
1997 		}
1998 		for (int i = 0; i < scctx->isc_nrxd[rxq->ifr_fl_offset]; i++) {
1999 			err = bus_dmamap_create(fl->ifl_buf_tag, 0,
2000 			    &fl->ifl_sds.ifsd_map[i]);
2001 			if (err != 0) {
2002 				device_printf(dev, "Unable to create RX buffer DMA map\n");
2003 				goto fail;
2004 			}
2005 		}
2006 	}
2007 	return (0);
2008 
2009 fail:
2010 	iflib_rx_structures_free(ctx);
2011 	return (err);
2012 }
2013 
2014 /*
2015  * Internal service routines
2016  */
2017 
2018 struct rxq_refill_cb_arg {
2019 	int               error;
2020 	bus_dma_segment_t seg;
2021 	int               nseg;
2022 };
2023 
2024 static void
2025 _rxq_refill_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error)
2026 {
2027 	struct rxq_refill_cb_arg *cb_arg = arg;
2028 
2029 	cb_arg->error = error;
2030 	cb_arg->seg = segs[0];
2031 	cb_arg->nseg = nseg;
2032 }
2033 
2034 /**
2035  * iflib_fl_refill - refill an rxq free-buffer list
2036  * @ctx: the iflib context
2037  * @fl: the free list to refill
2038  * @count: the number of new buffers to allocate
2039  *
2040  * (Re)populate an rxq free-buffer list with up to @count new packet buffers.
2041  * The caller must assure that @count does not exceed the queue's capacity
2042  * minus one (since we always leave a descriptor unavailable).
2043  */
2044 static uint8_t
2045 iflib_fl_refill(if_ctx_t ctx, iflib_fl_t fl, int count)
2046 {
2047 	struct if_rxd_update iru;
2048 	struct rxq_refill_cb_arg cb_arg;
2049 	struct mbuf *m;
2050 	caddr_t cl, *sd_cl;
2051 	struct mbuf **sd_m;
2052 	bus_dmamap_t *sd_map;
2053 	bus_addr_t bus_addr, *sd_ba;
2054 	int err, frag_idx, i, idx, n, pidx;
2055 	qidx_t credits;
2056 
2057 	MPASS(count <= fl->ifl_size - fl->ifl_credits - 1);
2058 
2059 	sd_m = fl->ifl_sds.ifsd_m;
2060 	sd_map = fl->ifl_sds.ifsd_map;
2061 	sd_cl = fl->ifl_sds.ifsd_cl;
2062 	sd_ba = fl->ifl_sds.ifsd_ba;
2063 	pidx = fl->ifl_pidx;
2064 	idx = pidx;
2065 	frag_idx = fl->ifl_fragidx;
2066 	credits = fl->ifl_credits;
2067 
2068 	i = 0;
2069 	n = count;
2070 	MPASS(n > 0);
2071 	MPASS(credits + n <= fl->ifl_size);
2072 
2073 	if (pidx < fl->ifl_cidx)
2074 		MPASS(pidx + n <= fl->ifl_cidx);
2075 	if (pidx == fl->ifl_cidx && (credits < fl->ifl_size))
2076 		MPASS(fl->ifl_gen == 0);
2077 	if (pidx > fl->ifl_cidx)
2078 		MPASS(n <= fl->ifl_size - pidx + fl->ifl_cidx);
2079 
2080 	DBG_COUNTER_INC(fl_refills);
2081 	if (n > 8)
2082 		DBG_COUNTER_INC(fl_refills_large);
2083 	iru_init(&iru, fl->ifl_rxq, fl->ifl_id);
2084 	while (n-- > 0) {
2085 		/*
2086 		 * We allocate an uninitialized mbuf + cluster, mbuf is
2087 		 * initialized after rx.
2088 		 *
2089 		 * If the cluster is still set then we know a minimum sized
2090 		 * packet was received
2091 		 */
2092 		bit_ffc_at(fl->ifl_rx_bitmap, frag_idx, fl->ifl_size,
2093 		    &frag_idx);
2094 		if (frag_idx < 0)
2095 			bit_ffc(fl->ifl_rx_bitmap, fl->ifl_size, &frag_idx);
2096 		MPASS(frag_idx >= 0);
2097 		if ((cl = sd_cl[frag_idx]) == NULL) {
2098 			cl = uma_zalloc(fl->ifl_zone, M_NOWAIT);
2099 			if (__predict_false(cl == NULL))
2100 				break;
2101 
2102 			cb_arg.error = 0;
2103 			MPASS(sd_map != NULL);
2104 			err = bus_dmamap_load(fl->ifl_buf_tag, sd_map[frag_idx],
2105 			    cl, fl->ifl_buf_size, _rxq_refill_cb, &cb_arg,
2106 			    BUS_DMA_NOWAIT);
2107 			if (__predict_false(err != 0 || cb_arg.error)) {
2108 				uma_zfree(fl->ifl_zone, cl);
2109 				break;
2110 			}
2111 
2112 			sd_ba[frag_idx] = bus_addr = cb_arg.seg.ds_addr;
2113 			sd_cl[frag_idx] = cl;
2114 #if MEMORY_LOGGING
2115 			fl->ifl_cl_enqueued++;
2116 #endif
2117 		} else {
2118 			bus_addr = sd_ba[frag_idx];
2119 		}
2120 		bus_dmamap_sync(fl->ifl_buf_tag, sd_map[frag_idx],
2121 		    BUS_DMASYNC_PREREAD);
2122 
2123 		if (sd_m[frag_idx] == NULL) {
2124 			m = m_gethdr(M_NOWAIT, MT_NOINIT);
2125 			if (__predict_false(m == NULL))
2126 				break;
2127 			sd_m[frag_idx] = m;
2128 		}
2129 		bit_set(fl->ifl_rx_bitmap, frag_idx);
2130 #if MEMORY_LOGGING
2131 		fl->ifl_m_enqueued++;
2132 #endif
2133 
2134 		DBG_COUNTER_INC(rx_allocs);
2135 		fl->ifl_rxd_idxs[i] = frag_idx;
2136 		fl->ifl_bus_addrs[i] = bus_addr;
2137 		credits++;
2138 		i++;
2139 		MPASS(credits <= fl->ifl_size);
2140 		if (++idx == fl->ifl_size) {
2141 #ifdef INVARIANTS
2142 			fl->ifl_gen = 1;
2143 #endif
2144 			idx = 0;
2145 		}
2146 		if (n == 0 || i == IFLIB_MAX_RX_REFRESH) {
2147 			iru.iru_pidx = pidx;
2148 			iru.iru_count = i;
2149 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2150 			fl->ifl_pidx = idx;
2151 			fl->ifl_credits = credits;
2152 			pidx = idx;
2153 			i = 0;
2154 		}
2155 	}
2156 
2157 	if (n < count - 1) {
2158 		if (i != 0) {
2159 			iru.iru_pidx = pidx;
2160 			iru.iru_count = i;
2161 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
2162 			fl->ifl_pidx = idx;
2163 			fl->ifl_credits = credits;
2164 		}
2165 		DBG_COUNTER_INC(rxd_flush);
2166 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2167 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2168 		ctx->isc_rxd_flush(ctx->ifc_softc, fl->ifl_rxq->ifr_id,
2169 		    fl->ifl_id, fl->ifl_pidx);
2170 		if (__predict_true(bit_test(fl->ifl_rx_bitmap, frag_idx))) {
2171 			fl->ifl_fragidx = frag_idx + 1;
2172 			if (fl->ifl_fragidx == fl->ifl_size)
2173 				fl->ifl_fragidx = 0;
2174 		} else {
2175 			fl->ifl_fragidx = frag_idx;
2176 		}
2177 	}
2178 
2179 	return (n == -1 ? 0 : IFLIB_RXEOF_EMPTY);
2180 }
2181 
2182 static inline uint8_t
2183 iflib_fl_refill_all(if_ctx_t ctx, iflib_fl_t fl)
2184 {
2185 	/*
2186 	 * We leave an unused descriptor to avoid pidx to catch up with cidx.
2187 	 * This is important as it confuses most NICs. For instance,
2188 	 * Intel NICs have (per receive ring) RDH and RDT registers, where
2189 	 * RDH points to the next receive descriptor to be used by the NIC,
2190 	 * and RDT for the next receive descriptor to be published by the
2191 	 * driver to the NIC (RDT - 1 is thus the last valid one).
2192 	 * The condition RDH == RDT means no descriptors are available to
2193 	 * the NIC, and thus it would be ambiguous if it also meant that
2194 	 * all the descriptors are available to the NIC.
2195 	 */
2196 	int32_t reclaimable = fl->ifl_size - fl->ifl_credits - 1;
2197 #ifdef INVARIANTS
2198 	int32_t delta = fl->ifl_size - get_inuse(fl->ifl_size, fl->ifl_cidx, fl->ifl_pidx, fl->ifl_gen) - 1;
2199 #endif
2200 
2201 	MPASS(fl->ifl_credits <= fl->ifl_size);
2202 	MPASS(reclaimable == delta);
2203 
2204 	if (reclaimable > 0)
2205 		return (iflib_fl_refill(ctx, fl, reclaimable));
2206 	return (0);
2207 }
2208 
2209 uint8_t
2210 iflib_in_detach(if_ctx_t ctx)
2211 {
2212 	bool in_detach;
2213 
2214 	STATE_LOCK(ctx);
2215 	in_detach = !!(ctx->ifc_flags & IFC_IN_DETACH);
2216 	STATE_UNLOCK(ctx);
2217 	return (in_detach);
2218 }
2219 
2220 static void
2221 iflib_fl_bufs_free(iflib_fl_t fl)
2222 {
2223 	iflib_dma_info_t idi = fl->ifl_ifdi;
2224 	bus_dmamap_t sd_map;
2225 	uint32_t i;
2226 
2227 	for (i = 0; i < fl->ifl_size; i++) {
2228 		struct mbuf **sd_m = &fl->ifl_sds.ifsd_m[i];
2229 		caddr_t *sd_cl = &fl->ifl_sds.ifsd_cl[i];
2230 
2231 		if (*sd_cl != NULL) {
2232 			sd_map = fl->ifl_sds.ifsd_map[i];
2233 			bus_dmamap_sync(fl->ifl_buf_tag, sd_map,
2234 			    BUS_DMASYNC_POSTREAD);
2235 			bus_dmamap_unload(fl->ifl_buf_tag, sd_map);
2236 			uma_zfree(fl->ifl_zone, *sd_cl);
2237 			*sd_cl = NULL;
2238 			if (*sd_m != NULL) {
2239 				m_init(*sd_m, M_NOWAIT, MT_DATA, 0);
2240 				m_free_raw(*sd_m);
2241 				*sd_m = NULL;
2242 			}
2243 		} else {
2244 			MPASS(*sd_m == NULL);
2245 		}
2246 #if MEMORY_LOGGING
2247 		fl->ifl_m_dequeued++;
2248 		fl->ifl_cl_dequeued++;
2249 #endif
2250 	}
2251 #ifdef INVARIANTS
2252 	for (i = 0; i < fl->ifl_size; i++) {
2253 		MPASS(fl->ifl_sds.ifsd_cl[i] == NULL);
2254 		MPASS(fl->ifl_sds.ifsd_m[i] == NULL);
2255 	}
2256 #endif
2257 	/*
2258 	 * Reset free list values
2259 	 */
2260 	fl->ifl_credits = fl->ifl_cidx = fl->ifl_pidx = fl->ifl_gen = fl->ifl_fragidx = 0;
2261 	bzero(idi->idi_vaddr, idi->idi_size);
2262 }
2263 
2264 /*********************************************************************
2265  *
2266  *  Initialize a free list and its buffers.
2267  *
2268  **********************************************************************/
2269 static int
2270 iflib_fl_setup(iflib_fl_t fl)
2271 {
2272 	iflib_rxq_t rxq = fl->ifl_rxq;
2273 	if_ctx_t ctx = rxq->ifr_ctx;
2274 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2275 	int qidx;
2276 
2277 	bit_nclear(fl->ifl_rx_bitmap, 0, fl->ifl_size - 1);
2278 	/*
2279 	** Free current RX buffer structs and their mbufs
2280 	*/
2281 	iflib_fl_bufs_free(fl);
2282 	/* Now replenish the mbufs */
2283 	MPASS(fl->ifl_credits == 0);
2284 	qidx = rxq->ifr_fl_offset + fl->ifl_id;
2285 	if (scctx->isc_rxd_buf_size[qidx] != 0)
2286 		fl->ifl_buf_size = scctx->isc_rxd_buf_size[qidx];
2287 	else
2288 		fl->ifl_buf_size = ctx->ifc_rx_mbuf_sz;
2289 	/*
2290 	 * ifl_buf_size may be a driver-supplied value, so pull it up
2291 	 * to the selected mbuf size.
2292 	 */
2293 	fl->ifl_buf_size = iflib_get_mbuf_size_for(fl->ifl_buf_size);
2294 	if (fl->ifl_buf_size > ctx->ifc_max_fl_buf_size)
2295 		ctx->ifc_max_fl_buf_size = fl->ifl_buf_size;
2296 	fl->ifl_cltype = m_gettype(fl->ifl_buf_size);
2297 	fl->ifl_zone = m_getzone(fl->ifl_buf_size);
2298 
2299 	/*
2300 	 * Avoid pre-allocating zillions of clusters to an idle card
2301 	 * potentially speeding up attach. In any case make sure
2302 	 * to leave a descriptor unavailable. See the comment in
2303 	 * iflib_fl_refill_all().
2304 	 */
2305 	MPASS(fl->ifl_size > 0);
2306 	(void)iflib_fl_refill(ctx, fl, min(128, fl->ifl_size - 1));
2307 	if (min(128, fl->ifl_size - 1) != fl->ifl_credits)
2308 		return (ENOBUFS);
2309 	/*
2310 	 * handle failure
2311 	 */
2312 	MPASS(rxq != NULL);
2313 	MPASS(fl->ifl_ifdi != NULL);
2314 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2315 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2316 	return (0);
2317 }
2318 
2319 /*********************************************************************
2320  *
2321  *  Free receive ring data structures
2322  *
2323  **********************************************************************/
2324 static void
2325 iflib_rx_sds_free(iflib_rxq_t rxq)
2326 {
2327 	iflib_fl_t fl;
2328 	int i, j;
2329 
2330 	if (rxq->ifr_fl != NULL) {
2331 		for (i = 0; i < rxq->ifr_nfl; i++) {
2332 			fl = &rxq->ifr_fl[i];
2333 			if (fl->ifl_buf_tag != NULL) {
2334 				if (fl->ifl_sds.ifsd_map != NULL) {
2335 					for (j = 0; j < fl->ifl_size; j++) {
2336 						bus_dmamap_sync(
2337 						    fl->ifl_buf_tag,
2338 						    fl->ifl_sds.ifsd_map[j],
2339 						    BUS_DMASYNC_POSTREAD);
2340 						bus_dmamap_unload(
2341 						    fl->ifl_buf_tag,
2342 						    fl->ifl_sds.ifsd_map[j]);
2343 						bus_dmamap_destroy(
2344 						    fl->ifl_buf_tag,
2345 						    fl->ifl_sds.ifsd_map[j]);
2346 					}
2347 				}
2348 				bus_dma_tag_destroy(fl->ifl_buf_tag);
2349 				fl->ifl_buf_tag = NULL;
2350 			}
2351 			free(fl->ifl_sds.ifsd_m, M_IFLIB);
2352 			free(fl->ifl_sds.ifsd_cl, M_IFLIB);
2353 			free(fl->ifl_sds.ifsd_ba, M_IFLIB);
2354 			free(fl->ifl_sds.ifsd_map, M_IFLIB);
2355 			free(fl->ifl_rx_bitmap, M_IFLIB);
2356 			fl->ifl_sds.ifsd_m = NULL;
2357 			fl->ifl_sds.ifsd_cl = NULL;
2358 			fl->ifl_sds.ifsd_ba = NULL;
2359 			fl->ifl_sds.ifsd_map = NULL;
2360 			fl->ifl_rx_bitmap = NULL;
2361 		}
2362 		free(rxq->ifr_fl, M_IFLIB);
2363 		rxq->ifr_fl = NULL;
2364 		free(rxq->ifr_ifdi, M_IFLIB);
2365 		rxq->ifr_ifdi = NULL;
2366 		rxq->ifr_cq_cidx = 0;
2367 	}
2368 }
2369 
2370 /*
2371  * Timer routine
2372  */
2373 static void
2374 iflib_timer(void *arg)
2375 {
2376 	iflib_txq_t txq = arg;
2377 	if_ctx_t ctx = txq->ift_ctx;
2378 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2379 	uint64_t this_tick = ticks;
2380 
2381 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
2382 		return;
2383 
2384 	/*
2385 	** Check on the state of the TX queue(s), this
2386 	** can be done without the lock because its RO
2387 	** and the HUNG state will be static if set.
2388 	*/
2389 	if (this_tick - txq->ift_last_timer_tick >= iflib_timer_default) {
2390 		txq->ift_last_timer_tick = this_tick;
2391 		IFDI_TIMER(ctx, txq->ift_id);
2392 		if ((txq->ift_qstatus == IFLIB_QUEUE_HUNG) &&
2393 		    ((txq->ift_cleaned_prev == txq->ift_cleaned) ||
2394 		     (sctx->isc_pause_frames == 0)))
2395 			goto hung;
2396 
2397 		if (txq->ift_qstatus != IFLIB_QUEUE_IDLE &&
2398 		    ifmp_ring_is_stalled(txq->ift_br)) {
2399 			KASSERT(ctx->ifc_link_state == LINK_STATE_UP,
2400 			    ("queue can't be marked as hung if interface is down"));
2401 			txq->ift_qstatus = IFLIB_QUEUE_HUNG;
2402 		}
2403 		txq->ift_cleaned_prev = txq->ift_cleaned;
2404 	}
2405 	/* handle any laggards */
2406 	if (txq->ift_db_pending)
2407 		GROUPTASK_ENQUEUE(&txq->ift_task);
2408 
2409 	sctx->isc_pause_frames = 0;
2410 	if (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)
2411 		callout_reset_on(&txq->ift_timer, iflib_timer_default, iflib_timer,
2412 		    txq, txq->ift_timer.c_cpu);
2413 	return;
2414 
2415  hung:
2416 	device_printf(ctx->ifc_dev,
2417 	    "Watchdog timeout (TX: %d desc avail: %d pidx: %d) -- resetting\n",
2418 	    txq->ift_id, TXQ_AVAIL(txq), txq->ift_pidx);
2419 	STATE_LOCK(ctx);
2420 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2421 	ctx->ifc_flags |= (IFC_DO_WATCHDOG|IFC_DO_RESET);
2422 	iflib_admin_intr_deferred(ctx);
2423 	STATE_UNLOCK(ctx);
2424 }
2425 
2426 static uint16_t
2427 iflib_get_mbuf_size_for(unsigned int size)
2428 {
2429 
2430 	if (size <= MCLBYTES)
2431 		return (MCLBYTES);
2432 	else
2433 		return (MJUMPAGESIZE);
2434 }
2435 
2436 static void
2437 iflib_calc_rx_mbuf_sz(if_ctx_t ctx)
2438 {
2439 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2440 
2441 	/*
2442 	 * XXX don't set the max_frame_size to larger
2443 	 * than the hardware can handle
2444 	 */
2445 	ctx->ifc_rx_mbuf_sz =
2446 	    iflib_get_mbuf_size_for(sctx->isc_max_frame_size);
2447 }
2448 
2449 uint32_t
2450 iflib_get_rx_mbuf_sz(if_ctx_t ctx)
2451 {
2452 
2453 	return (ctx->ifc_rx_mbuf_sz);
2454 }
2455 
2456 static void
2457 iflib_init_locked(if_ctx_t ctx)
2458 {
2459 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2460 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2461 	if_t ifp = ctx->ifc_ifp;
2462 	iflib_fl_t fl;
2463 	iflib_txq_t txq;
2464 	iflib_rxq_t rxq;
2465 	int i, j, tx_ip_csum_flags, tx_ip6_csum_flags;
2466 
2467 	if_setdrvflagbits(ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2468 	IFDI_INTR_DISABLE(ctx);
2469 
2470 	/*
2471 	 * See iflib_stop(). Useful in case iflib_init_locked() is
2472 	 * called without first calling iflib_stop().
2473 	 */
2474 	netmap_disable_all_rings(ifp);
2475 
2476 	tx_ip_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP | CSUM_SCTP);
2477 	tx_ip6_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_UDP | CSUM_IP6_SCTP);
2478 	/* Set hardware offload abilities */
2479 	if_clearhwassist(ifp);
2480 	if (if_getcapenable(ifp) & IFCAP_TXCSUM)
2481 		if_sethwassistbits(ifp, tx_ip_csum_flags, 0);
2482 	if (if_getcapenable(ifp) & IFCAP_TXCSUM_IPV6)
2483 		if_sethwassistbits(ifp,  tx_ip6_csum_flags, 0);
2484 	if (if_getcapenable(ifp) & IFCAP_TSO4)
2485 		if_sethwassistbits(ifp, CSUM_IP_TSO, 0);
2486 	if (if_getcapenable(ifp) & IFCAP_TSO6)
2487 		if_sethwassistbits(ifp, CSUM_IP6_TSO, 0);
2488 
2489 	for (i = 0, txq = ctx->ifc_txqs; i < sctx->isc_ntxqsets; i++, txq++) {
2490 		CALLOUT_LOCK(txq);
2491 		callout_stop(&txq->ift_timer);
2492 #ifdef DEV_NETMAP
2493 		callout_stop(&txq->ift_netmap_timer);
2494 #endif /* DEV_NETMAP */
2495 		CALLOUT_UNLOCK(txq);
2496 		iflib_netmap_txq_init(ctx, txq);
2497 	}
2498 
2499 	/*
2500 	 * Calculate a suitable Rx mbuf size prior to calling IFDI_INIT, so
2501 	 * that drivers can use the value when setting up the hardware receive
2502 	 * buffers.
2503 	 */
2504 	iflib_calc_rx_mbuf_sz(ctx);
2505 
2506 #ifdef INVARIANTS
2507 	i = if_getdrvflags(ifp);
2508 #endif
2509 	IFDI_INIT(ctx);
2510 	MPASS(if_getdrvflags(ifp) == i);
2511 	for (i = 0, rxq = ctx->ifc_rxqs; i < sctx->isc_nrxqsets; i++, rxq++) {
2512 		if (iflib_netmap_rxq_init(ctx, rxq) > 0) {
2513 			/* This rxq is in netmap mode. Skip normal init. */
2514 			continue;
2515 		}
2516 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
2517 			if (iflib_fl_setup(fl)) {
2518 				device_printf(ctx->ifc_dev,
2519 				    "setting up free list %d failed - "
2520 				    "check cluster settings\n", j);
2521 				goto done;
2522 			}
2523 		}
2524 	}
2525 done:
2526 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_RUNNING, IFF_DRV_OACTIVE);
2527 	IFDI_INTR_ENABLE(ctx);
2528 	txq = ctx->ifc_txqs;
2529 	for (i = 0; i < sctx->isc_ntxqsets; i++, txq++)
2530 		callout_reset_on(&txq->ift_timer, iflib_timer_default, iflib_timer, txq,
2531 			txq->ift_timer.c_cpu);
2532 
2533         /* Re-enable txsync/rxsync. */
2534 	netmap_enable_all_rings(ifp);
2535 }
2536 
2537 static int
2538 iflib_media_change(if_t ifp)
2539 {
2540 	if_ctx_t ctx = if_getsoftc(ifp);
2541 	int err;
2542 
2543 	CTX_LOCK(ctx);
2544 	if ((err = IFDI_MEDIA_CHANGE(ctx)) == 0)
2545 		iflib_if_init_locked(ctx);
2546 	CTX_UNLOCK(ctx);
2547 	return (err);
2548 }
2549 
2550 static void
2551 iflib_media_status(if_t ifp, struct ifmediareq *ifmr)
2552 {
2553 	if_ctx_t ctx = if_getsoftc(ifp);
2554 
2555 	CTX_LOCK(ctx);
2556 	IFDI_UPDATE_ADMIN_STATUS(ctx);
2557 	IFDI_MEDIA_STATUS(ctx, ifmr);
2558 	CTX_UNLOCK(ctx);
2559 }
2560 
2561 void
2562 iflib_stop(if_ctx_t ctx)
2563 {
2564 	iflib_txq_t txq = ctx->ifc_txqs;
2565 	iflib_rxq_t rxq = ctx->ifc_rxqs;
2566 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2567 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2568 	iflib_dma_info_t di;
2569 	iflib_fl_t fl;
2570 	int i, j;
2571 
2572 	/* Tell the stack that the interface is no longer active */
2573 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2574 
2575 	IFDI_INTR_DISABLE(ctx);
2576 	DELAY(1000);
2577 	IFDI_STOP(ctx);
2578 	DELAY(1000);
2579 
2580 	/*
2581 	 * Stop any pending txsync/rxsync and prevent new ones
2582 	 * form starting. Processes blocked in poll() will get
2583 	 * POLLERR.
2584 	 */
2585 	netmap_disable_all_rings(ctx->ifc_ifp);
2586 
2587 	iflib_debug_reset();
2588 	/* Wait for current tx queue users to exit to disarm watchdog timer. */
2589 	for (i = 0; i < scctx->isc_ntxqsets; i++, txq++) {
2590 		/* make sure all transmitters have completed before proceeding XXX */
2591 
2592 		CALLOUT_LOCK(txq);
2593 		callout_stop(&txq->ift_timer);
2594 #ifdef DEV_NETMAP
2595 		callout_stop(&txq->ift_netmap_timer);
2596 #endif /* DEV_NETMAP */
2597 		CALLOUT_UNLOCK(txq);
2598 
2599 		/* clean any enqueued buffers */
2600 		iflib_ifmp_purge(txq);
2601 		/* Free any existing tx buffers. */
2602 		for (j = 0; j < txq->ift_size; j++) {
2603 			iflib_txsd_free(ctx, txq, j);
2604 		}
2605 		txq->ift_processed = txq->ift_cleaned = txq->ift_cidx_processed = 0;
2606 		txq->ift_in_use = txq->ift_gen = txq->ift_cidx = txq->ift_pidx = txq->ift_no_desc_avail = 0;
2607 		txq->ift_closed = txq->ift_mbuf_defrag = txq->ift_mbuf_defrag_failed = 0;
2608 		txq->ift_no_tx_dma_setup = txq->ift_txd_encap_efbig = txq->ift_map_failed = 0;
2609 		txq->ift_pullups = 0;
2610 		ifmp_ring_reset_stats(txq->ift_br);
2611 		for (j = 0, di = txq->ift_ifdi; j < sctx->isc_ntxqs; j++, di++)
2612 			bzero((void *)di->idi_vaddr, di->idi_size);
2613 	}
2614 	for (i = 0; i < scctx->isc_nrxqsets; i++, rxq++) {
2615 		gtaskqueue_drain(rxq->ifr_task.gt_taskqueue,
2616 		    &rxq->ifr_task.gt_task);
2617 
2618 		rxq->ifr_cq_cidx = 0;
2619 		for (j = 0, di = rxq->ifr_ifdi; j < sctx->isc_nrxqs; j++, di++)
2620 			bzero((void *)di->idi_vaddr, di->idi_size);
2621 		/* also resets the free lists pidx/cidx */
2622 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
2623 			iflib_fl_bufs_free(fl);
2624 	}
2625 }
2626 
2627 static inline caddr_t
2628 calc_next_rxd(iflib_fl_t fl, int cidx)
2629 {
2630 	qidx_t size;
2631 	int nrxd;
2632 	caddr_t start, end, cur, next;
2633 
2634 	nrxd = fl->ifl_size;
2635 	size = fl->ifl_rxd_size;
2636 	start = fl->ifl_ifdi->idi_vaddr;
2637 
2638 	if (__predict_false(size == 0))
2639 		return (start);
2640 	cur = start + size*cidx;
2641 	end = start + size*nrxd;
2642 	next = CACHE_PTR_NEXT(cur);
2643 	return (next < end ? next : start);
2644 }
2645 
2646 static inline void
2647 prefetch_pkts(iflib_fl_t fl, int cidx)
2648 {
2649 	int nextptr;
2650 	int nrxd = fl->ifl_size;
2651 	caddr_t next_rxd;
2652 
2653 	nextptr = (cidx + CACHE_PTR_INCREMENT) & (nrxd-1);
2654 	prefetch(&fl->ifl_sds.ifsd_m[nextptr]);
2655 	prefetch(&fl->ifl_sds.ifsd_cl[nextptr]);
2656 	next_rxd = calc_next_rxd(fl, cidx);
2657 	prefetch(next_rxd);
2658 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 1) & (nrxd-1)]);
2659 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 2) & (nrxd-1)]);
2660 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 3) & (nrxd-1)]);
2661 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 4) & (nrxd-1)]);
2662 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 1) & (nrxd-1)]);
2663 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 2) & (nrxd-1)]);
2664 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 3) & (nrxd-1)]);
2665 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 4) & (nrxd-1)]);
2666 }
2667 
2668 static struct mbuf *
2669 rxd_frag_to_sd(iflib_rxq_t rxq, if_rxd_frag_t irf, bool unload, if_rxsd_t sd,
2670     int *pf_rv, if_rxd_info_t ri)
2671 {
2672 	bus_dmamap_t map;
2673 	iflib_fl_t fl;
2674 	caddr_t payload;
2675 	struct mbuf *m;
2676 	int flid, cidx, len, next;
2677 
2678 	map = NULL;
2679 	flid = irf->irf_flid;
2680 	cidx = irf->irf_idx;
2681 	fl = &rxq->ifr_fl[flid];
2682 	sd->ifsd_fl = fl;
2683 	m = fl->ifl_sds.ifsd_m[cidx];
2684 	sd->ifsd_cl = &fl->ifl_sds.ifsd_cl[cidx];
2685 	fl->ifl_credits--;
2686 #if MEMORY_LOGGING
2687 	fl->ifl_m_dequeued++;
2688 #endif
2689 	if (rxq->ifr_ctx->ifc_flags & IFC_PREFETCH)
2690 		prefetch_pkts(fl, cidx);
2691 	next = (cidx + CACHE_PTR_INCREMENT) & (fl->ifl_size-1);
2692 	prefetch(&fl->ifl_sds.ifsd_map[next]);
2693 	map = fl->ifl_sds.ifsd_map[cidx];
2694 
2695 	bus_dmamap_sync(fl->ifl_buf_tag, map, BUS_DMASYNC_POSTREAD);
2696 
2697 	if (rxq->pfil != NULL && PFIL_HOOKED_IN(rxq->pfil) && pf_rv != NULL &&
2698 	    irf->irf_len != 0) {
2699 		payload  = *sd->ifsd_cl;
2700 		payload +=  ri->iri_pad;
2701 		len = ri->iri_len - ri->iri_pad;
2702 		*pf_rv = pfil_run_hooks(rxq->pfil, payload, ri->iri_ifp,
2703 		    len | PFIL_MEMPTR | PFIL_IN, NULL);
2704 		switch (*pf_rv) {
2705 		case PFIL_DROPPED:
2706 		case PFIL_CONSUMED:
2707 			/*
2708 			 * The filter ate it.  Everything is recycled.
2709 			 */
2710 			m = NULL;
2711 			unload = 0;
2712 			break;
2713 		case PFIL_REALLOCED:
2714 			/*
2715 			 * The filter copied it.  Everything is recycled.
2716 			 */
2717 			m = pfil_mem2mbuf(payload);
2718 			unload = 0;
2719 			break;
2720 		case PFIL_PASS:
2721 			/*
2722 			 * Filter said it was OK, so receive like
2723 			 * normal
2724 			 */
2725 			fl->ifl_sds.ifsd_m[cidx] = NULL;
2726 			break;
2727 		default:
2728 			MPASS(0);
2729 		}
2730 	} else {
2731 		fl->ifl_sds.ifsd_m[cidx] = NULL;
2732 		if (pf_rv != NULL)
2733 			*pf_rv = PFIL_PASS;
2734 	}
2735 
2736 	if (unload && irf->irf_len != 0)
2737 		bus_dmamap_unload(fl->ifl_buf_tag, map);
2738 	fl->ifl_cidx = (fl->ifl_cidx + 1) & (fl->ifl_size-1);
2739 	if (__predict_false(fl->ifl_cidx == 0))
2740 		fl->ifl_gen = 0;
2741 	bit_clear(fl->ifl_rx_bitmap, cidx);
2742 	return (m);
2743 }
2744 
2745 static struct mbuf *
2746 assemble_segments(iflib_rxq_t rxq, if_rxd_info_t ri, if_rxsd_t sd, int *pf_rv)
2747 {
2748 	struct mbuf *m, *mh, *mt;
2749 	caddr_t cl;
2750 	int  *pf_rv_ptr, flags, i, padlen;
2751 	bool consumed;
2752 
2753 	i = 0;
2754 	mh = NULL;
2755 	consumed = false;
2756 	*pf_rv = PFIL_PASS;
2757 	pf_rv_ptr = pf_rv;
2758 	do {
2759 		m = rxd_frag_to_sd(rxq, &ri->iri_frags[i], !consumed, sd,
2760 		    pf_rv_ptr, ri);
2761 
2762 		MPASS(*sd->ifsd_cl != NULL);
2763 
2764 		/*
2765 		 * Exclude zero-length frags & frags from
2766 		 * packets the filter has consumed or dropped
2767 		 */
2768 		if (ri->iri_frags[i].irf_len == 0 || consumed ||
2769 		    *pf_rv == PFIL_CONSUMED || *pf_rv == PFIL_DROPPED) {
2770 			if (mh == NULL) {
2771 				/* everything saved here */
2772 				consumed = true;
2773 				pf_rv_ptr = NULL;
2774 				continue;
2775 			}
2776 			/* XXX we can save the cluster here, but not the mbuf */
2777 			m_init(m, M_NOWAIT, MT_DATA, 0);
2778 			m_free(m);
2779 			continue;
2780 		}
2781 		if (mh == NULL) {
2782 			flags = M_PKTHDR|M_EXT;
2783 			mh = mt = m;
2784 			padlen = ri->iri_pad;
2785 		} else {
2786 			flags = M_EXT;
2787 			mt->m_next = m;
2788 			mt = m;
2789 			/* assuming padding is only on the first fragment */
2790 			padlen = 0;
2791 		}
2792 		cl = *sd->ifsd_cl;
2793 		*sd->ifsd_cl = NULL;
2794 
2795 		/* Can these two be made one ? */
2796 		m_init(m, M_NOWAIT, MT_DATA, flags);
2797 		m_cljset(m, cl, sd->ifsd_fl->ifl_cltype);
2798 		/*
2799 		 * These must follow m_init and m_cljset
2800 		 */
2801 		m->m_data += padlen;
2802 		ri->iri_len -= padlen;
2803 		m->m_len = ri->iri_frags[i].irf_len;
2804 	} while (++i < ri->iri_nfrags);
2805 
2806 	return (mh);
2807 }
2808 
2809 /*
2810  * Process one software descriptor
2811  */
2812 static struct mbuf *
2813 iflib_rxd_pkt_get(iflib_rxq_t rxq, if_rxd_info_t ri)
2814 {
2815 	struct if_rxsd sd;
2816 	struct mbuf *m;
2817 	int pf_rv;
2818 
2819 	/* should I merge this back in now that the two paths are basically duplicated? */
2820 	if (ri->iri_nfrags == 1 &&
2821 	    ri->iri_frags[0].irf_len != 0 &&
2822 	    ri->iri_frags[0].irf_len <= MIN(IFLIB_RX_COPY_THRESH, MHLEN)) {
2823 		m = rxd_frag_to_sd(rxq, &ri->iri_frags[0], false, &sd,
2824 		    &pf_rv, ri);
2825 		if (pf_rv != PFIL_PASS && pf_rv != PFIL_REALLOCED)
2826 			return (m);
2827 		if (pf_rv == PFIL_PASS) {
2828 			m_init(m, M_NOWAIT, MT_DATA, M_PKTHDR);
2829 #ifndef __NO_STRICT_ALIGNMENT
2830 			if (!IP_ALIGNED(m))
2831 				m->m_data += 2;
2832 #endif
2833 			memcpy(m->m_data, *sd.ifsd_cl, ri->iri_len);
2834 			m->m_len = ri->iri_frags[0].irf_len;
2835 		}
2836 	} else {
2837 		m = assemble_segments(rxq, ri, &sd, &pf_rv);
2838 		if (m == NULL)
2839 			return (NULL);
2840 		if (pf_rv != PFIL_PASS && pf_rv != PFIL_REALLOCED)
2841 			return (m);
2842 	}
2843 	m->m_pkthdr.len = ri->iri_len;
2844 	m->m_pkthdr.rcvif = ri->iri_ifp;
2845 	m->m_flags |= ri->iri_flags;
2846 	m->m_pkthdr.ether_vtag = ri->iri_vtag;
2847 	m->m_pkthdr.flowid = ri->iri_flowid;
2848 	M_HASHTYPE_SET(m, ri->iri_rsstype);
2849 	m->m_pkthdr.csum_flags = ri->iri_csum_flags;
2850 	m->m_pkthdr.csum_data = ri->iri_csum_data;
2851 	return (m);
2852 }
2853 
2854 #if defined(INET6) || defined(INET)
2855 static void
2856 iflib_get_ip_forwarding(struct lro_ctrl *lc, bool *v4, bool *v6)
2857 {
2858 	CURVNET_SET(lc->ifp->if_vnet);
2859 #if defined(INET6)
2860 	*v6 = V_ip6_forwarding;
2861 #endif
2862 #if defined(INET)
2863 	*v4 = V_ipforwarding;
2864 #endif
2865 	CURVNET_RESTORE();
2866 }
2867 
2868 /*
2869  * Returns true if it's possible this packet could be LROed.
2870  * if it returns false, it is guaranteed that tcp_lro_rx()
2871  * would not return zero.
2872  */
2873 static bool
2874 iflib_check_lro_possible(struct mbuf *m, bool v4_forwarding, bool v6_forwarding)
2875 {
2876 	struct ether_header *eh;
2877 
2878 	eh = mtod(m, struct ether_header *);
2879 	switch (eh->ether_type) {
2880 #if defined(INET6)
2881 		case htons(ETHERTYPE_IPV6):
2882 			return (!v6_forwarding);
2883 #endif
2884 #if defined (INET)
2885 		case htons(ETHERTYPE_IP):
2886 			return (!v4_forwarding);
2887 #endif
2888 	}
2889 
2890 	return false;
2891 }
2892 #else
2893 static void
2894 iflib_get_ip_forwarding(struct lro_ctrl *lc __unused, bool *v4 __unused, bool *v6 __unused)
2895 {
2896 }
2897 #endif
2898 
2899 static void
2900 _task_fn_rx_watchdog(void *context)
2901 {
2902 	iflib_rxq_t rxq = context;
2903 
2904 	GROUPTASK_ENQUEUE(&rxq->ifr_task);
2905 }
2906 
2907 static uint8_t
2908 iflib_rxeof(iflib_rxq_t rxq, qidx_t budget)
2909 {
2910 	if_t ifp;
2911 	if_ctx_t ctx = rxq->ifr_ctx;
2912 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2913 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2914 	int avail, i;
2915 	qidx_t *cidxp;
2916 	struct if_rxd_info ri;
2917 	int err, budget_left, rx_bytes, rx_pkts;
2918 	iflib_fl_t fl;
2919 	int lro_enabled;
2920 	bool v4_forwarding, v6_forwarding, lro_possible;
2921 	uint8_t retval = 0;
2922 
2923 	/*
2924 	 * XXX early demux data packets so that if_input processing only handles
2925 	 * acks in interrupt context
2926 	 */
2927 	struct mbuf *m, *mh, *mt, *mf;
2928 
2929 	NET_EPOCH_ASSERT();
2930 
2931 	lro_possible = v4_forwarding = v6_forwarding = false;
2932 	ifp = ctx->ifc_ifp;
2933 	mh = mt = NULL;
2934 	MPASS(budget > 0);
2935 	rx_pkts	= rx_bytes = 0;
2936 	if (sctx->isc_flags & IFLIB_HAS_RXCQ)
2937 		cidxp = &rxq->ifr_cq_cidx;
2938 	else
2939 		cidxp = &rxq->ifr_fl[0].ifl_cidx;
2940 	if ((avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget)) == 0) {
2941 		for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2942 			retval |= iflib_fl_refill_all(ctx, fl);
2943 		DBG_COUNTER_INC(rx_unavail);
2944 		return (retval);
2945 	}
2946 
2947 	/* pfil needs the vnet to be set */
2948 	CURVNET_SET_QUIET(ifp->if_vnet);
2949 	for (budget_left = budget; budget_left > 0 && avail > 0;) {
2950 		if (__predict_false(!CTX_ACTIVE(ctx))) {
2951 			DBG_COUNTER_INC(rx_ctx_inactive);
2952 			break;
2953 		}
2954 		/*
2955 		 * Reset client set fields to their default values
2956 		 */
2957 		rxd_info_zero(&ri);
2958 		ri.iri_qsidx = rxq->ifr_id;
2959 		ri.iri_cidx = *cidxp;
2960 		ri.iri_ifp = ifp;
2961 		ri.iri_frags = rxq->ifr_frags;
2962 		err = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
2963 
2964 		if (err)
2965 			goto err;
2966 		rx_pkts += 1;
2967 		rx_bytes += ri.iri_len;
2968 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
2969 			*cidxp = ri.iri_cidx;
2970 			/* Update our consumer index */
2971 			/* XXX NB: shurd - check if this is still safe */
2972 			while (rxq->ifr_cq_cidx >= scctx->isc_nrxd[0])
2973 				rxq->ifr_cq_cidx -= scctx->isc_nrxd[0];
2974 			/* was this only a completion queue message? */
2975 			if (__predict_false(ri.iri_nfrags == 0))
2976 				continue;
2977 		}
2978 		MPASS(ri.iri_nfrags != 0);
2979 		MPASS(ri.iri_len != 0);
2980 
2981 		/* will advance the cidx on the corresponding free lists */
2982 		m = iflib_rxd_pkt_get(rxq, &ri);
2983 		avail--;
2984 		budget_left--;
2985 		if (avail == 0 && budget_left)
2986 			avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget_left);
2987 
2988 		if (__predict_false(m == NULL))
2989 			continue;
2990 
2991 		/* imm_pkt: -- cxgb */
2992 		if (mh == NULL)
2993 			mh = mt = m;
2994 		else {
2995 			mt->m_nextpkt = m;
2996 			mt = m;
2997 		}
2998 	}
2999 	CURVNET_RESTORE();
3000 	/* make sure that we can refill faster than drain */
3001 	for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
3002 		retval |= iflib_fl_refill_all(ctx, fl);
3003 
3004 	lro_enabled = (if_getcapenable(ifp) & IFCAP_LRO);
3005 	if (lro_enabled)
3006 		iflib_get_ip_forwarding(&rxq->ifr_lc, &v4_forwarding, &v6_forwarding);
3007 	mt = mf = NULL;
3008 	while (mh != NULL) {
3009 		m = mh;
3010 		mh = mh->m_nextpkt;
3011 		m->m_nextpkt = NULL;
3012 #ifndef __NO_STRICT_ALIGNMENT
3013 		if (!IP_ALIGNED(m) && (m = iflib_fixup_rx(m)) == NULL)
3014 			continue;
3015 #endif
3016 #if defined(INET6) || defined(INET)
3017 		if (lro_enabled) {
3018 			if (!lro_possible) {
3019 				lro_possible = iflib_check_lro_possible(m, v4_forwarding, v6_forwarding);
3020 				if (lro_possible && mf != NULL) {
3021 					ifp->if_input(ifp, mf);
3022 					DBG_COUNTER_INC(rx_if_input);
3023 					mt = mf = NULL;
3024 				}
3025 			}
3026 			if ((m->m_pkthdr.csum_flags & (CSUM_L4_CALC|CSUM_L4_VALID)) ==
3027 			    (CSUM_L4_CALC|CSUM_L4_VALID)) {
3028 				if (lro_possible && tcp_lro_rx(&rxq->ifr_lc, m, 0) == 0)
3029 					continue;
3030 			}
3031 		}
3032 #endif
3033 		if (lro_possible) {
3034 			ifp->if_input(ifp, m);
3035 			DBG_COUNTER_INC(rx_if_input);
3036 			continue;
3037 		}
3038 
3039 		if (mf == NULL)
3040 			mf = m;
3041 		if (mt != NULL)
3042 			mt->m_nextpkt = m;
3043 		mt = m;
3044 	}
3045 	if (mf != NULL) {
3046 		ifp->if_input(ifp, mf);
3047 		DBG_COUNTER_INC(rx_if_input);
3048 	}
3049 
3050 	if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes);
3051 	if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts);
3052 
3053 	/*
3054 	 * Flush any outstanding LRO work
3055 	 */
3056 #if defined(INET6) || defined(INET)
3057 	tcp_lro_flush_all(&rxq->ifr_lc);
3058 #endif
3059 	if (avail != 0 || iflib_rxd_avail(ctx, rxq, *cidxp, 1) != 0)
3060 		retval |= IFLIB_RXEOF_MORE;
3061 	return (retval);
3062 err:
3063 	STATE_LOCK(ctx);
3064 	ctx->ifc_flags |= IFC_DO_RESET;
3065 	iflib_admin_intr_deferred(ctx);
3066 	STATE_UNLOCK(ctx);
3067 	return (0);
3068 }
3069 
3070 #define TXD_NOTIFY_COUNT(txq) (((txq)->ift_size / (txq)->ift_update_freq)-1)
3071 static inline qidx_t
3072 txq_max_db_deferred(iflib_txq_t txq, qidx_t in_use)
3073 {
3074 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
3075 	qidx_t minthresh = txq->ift_size / 8;
3076 	if (in_use > 4*minthresh)
3077 		return (notify_count);
3078 	if (in_use > 2*minthresh)
3079 		return (notify_count >> 1);
3080 	if (in_use > minthresh)
3081 		return (notify_count >> 3);
3082 	return (0);
3083 }
3084 
3085 static inline qidx_t
3086 txq_max_rs_deferred(iflib_txq_t txq)
3087 {
3088 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
3089 	qidx_t minthresh = txq->ift_size / 8;
3090 	if (txq->ift_in_use > 4*minthresh)
3091 		return (notify_count);
3092 	if (txq->ift_in_use > 2*minthresh)
3093 		return (notify_count >> 1);
3094 	if (txq->ift_in_use > minthresh)
3095 		return (notify_count >> 2);
3096 	return (2);
3097 }
3098 
3099 #define M_CSUM_FLAGS(m) ((m)->m_pkthdr.csum_flags)
3100 #define M_HAS_VLANTAG(m) (m->m_flags & M_VLANTAG)
3101 
3102 #define TXQ_MAX_DB_DEFERRED(txq, in_use) txq_max_db_deferred((txq), (in_use))
3103 #define TXQ_MAX_RS_DEFERRED(txq) txq_max_rs_deferred(txq)
3104 #define TXQ_MAX_DB_CONSUMED(size) (size >> 4)
3105 
3106 /* forward compatibility for cxgb */
3107 #define FIRST_QSET(ctx) 0
3108 #define NTXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_ntxqsets)
3109 #define NRXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_nrxqsets)
3110 #define QIDX(ctx, m) ((((m)->m_pkthdr.flowid & ctx->ifc_softc_ctx.isc_rss_table_mask) % NTXQSETS(ctx)) + FIRST_QSET(ctx))
3111 #define DESC_RECLAIMABLE(q) ((int)((q)->ift_processed - (q)->ift_cleaned - (q)->ift_ctx->ifc_softc_ctx.isc_tx_nsegments))
3112 
3113 /* XXX we should be setting this to something other than zero */
3114 #define RECLAIM_THRESH(ctx) ((ctx)->ifc_sctx->isc_tx_reclaim_thresh)
3115 #define	MAX_TX_DESC(ctx) MAX((ctx)->ifc_softc_ctx.isc_tx_tso_segments_max, \
3116     (ctx)->ifc_softc_ctx.isc_tx_nsegments)
3117 
3118 static inline bool
3119 iflib_txd_db_check(iflib_txq_t txq, int ring)
3120 {
3121 	if_ctx_t ctx = txq->ift_ctx;
3122 	qidx_t dbval, max;
3123 
3124 	max = TXQ_MAX_DB_DEFERRED(txq, txq->ift_in_use);
3125 
3126 	/* force || threshold exceeded || at the edge of the ring */
3127 	if (ring || (txq->ift_db_pending >= max) || (TXQ_AVAIL(txq) <= MAX_TX_DESC(ctx) + 2)) {
3128 
3129 		/*
3130 		 * 'npending' is used if the card's doorbell is in terms of the number of descriptors
3131 		 * pending flush (BRCM). 'pidx' is used in cases where the card's doorbeel uses the
3132 		 * producer index explicitly (INTC).
3133 		 */
3134 		dbval = txq->ift_npending ? txq->ift_npending : txq->ift_pidx;
3135 		bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3136 		    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
3137 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, dbval);
3138 
3139 		/*
3140 		 * Absent bugs there are zero packets pending so reset pending counts to zero.
3141 		 */
3142 		txq->ift_db_pending = txq->ift_npending = 0;
3143 		return (true);
3144 	}
3145 	return (false);
3146 }
3147 
3148 #ifdef PKT_DEBUG
3149 static void
3150 print_pkt(if_pkt_info_t pi)
3151 {
3152 	printf("pi len:  %d qsidx: %d nsegs: %d ndescs: %d flags: %x pidx: %d\n",
3153 	       pi->ipi_len, pi->ipi_qsidx, pi->ipi_nsegs, pi->ipi_ndescs, pi->ipi_flags, pi->ipi_pidx);
3154 	printf("pi new_pidx: %d csum_flags: %lx tso_segsz: %d mflags: %x vtag: %d\n",
3155 	       pi->ipi_new_pidx, pi->ipi_csum_flags, pi->ipi_tso_segsz, pi->ipi_mflags, pi->ipi_vtag);
3156 	printf("pi etype: %d ehdrlen: %d ip_hlen: %d ipproto: %d\n",
3157 	       pi->ipi_etype, pi->ipi_ehdrlen, pi->ipi_ip_hlen, pi->ipi_ipproto);
3158 }
3159 #endif
3160 
3161 #define IS_TSO4(pi) ((pi)->ipi_csum_flags & CSUM_IP_TSO)
3162 #define IS_TX_OFFLOAD4(pi) ((pi)->ipi_csum_flags & (CSUM_IP_TCP | CSUM_IP_TSO))
3163 #define IS_TSO6(pi) ((pi)->ipi_csum_flags & CSUM_IP6_TSO)
3164 #define IS_TX_OFFLOAD6(pi) ((pi)->ipi_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_TSO))
3165 
3166 static int
3167 iflib_parse_header(iflib_txq_t txq, if_pkt_info_t pi, struct mbuf **mp)
3168 {
3169 	if_shared_ctx_t sctx = txq->ift_ctx->ifc_sctx;
3170 	struct ether_vlan_header *eh;
3171 	struct mbuf *m;
3172 
3173 	m = *mp;
3174 	if ((sctx->isc_flags & IFLIB_NEED_SCRATCH) &&
3175 	    M_WRITABLE(m) == 0) {
3176 		if ((m = m_dup(m, M_NOWAIT)) == NULL) {
3177 			return (ENOMEM);
3178 		} else {
3179 			m_freem(*mp);
3180 			DBG_COUNTER_INC(tx_frees);
3181 			*mp = m;
3182 		}
3183 	}
3184 
3185 	/*
3186 	 * Determine where frame payload starts.
3187 	 * Jump over vlan headers if already present,
3188 	 * helpful for QinQ too.
3189 	 */
3190 	if (__predict_false(m->m_len < sizeof(*eh))) {
3191 		txq->ift_pullups++;
3192 		if (__predict_false((m = m_pullup(m, sizeof(*eh))) == NULL))
3193 			return (ENOMEM);
3194 	}
3195 	eh = mtod(m, struct ether_vlan_header *);
3196 	if (eh->evl_encap_proto == htons(ETHERTYPE_VLAN)) {
3197 		pi->ipi_etype = ntohs(eh->evl_proto);
3198 		pi->ipi_ehdrlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
3199 	} else {
3200 		pi->ipi_etype = ntohs(eh->evl_encap_proto);
3201 		pi->ipi_ehdrlen = ETHER_HDR_LEN;
3202 	}
3203 
3204 	switch (pi->ipi_etype) {
3205 #ifdef INET
3206 	case ETHERTYPE_IP:
3207 	{
3208 		struct mbuf *n;
3209 		struct ip *ip = NULL;
3210 		struct tcphdr *th = NULL;
3211 		int minthlen;
3212 
3213 		minthlen = min(m->m_pkthdr.len, pi->ipi_ehdrlen + sizeof(*ip) + sizeof(*th));
3214 		if (__predict_false(m->m_len < minthlen)) {
3215 			/*
3216 			 * if this code bloat is causing too much of a hit
3217 			 * move it to a separate function and mark it noinline
3218 			 */
3219 			if (m->m_len == pi->ipi_ehdrlen) {
3220 				n = m->m_next;
3221 				MPASS(n);
3222 				if (n->m_len >= sizeof(*ip))  {
3223 					ip = (struct ip *)n->m_data;
3224 					if (n->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3225 						th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3226 				} else {
3227 					txq->ift_pullups++;
3228 					if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
3229 						return (ENOMEM);
3230 					ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3231 				}
3232 			} else {
3233 				txq->ift_pullups++;
3234 				if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
3235 					return (ENOMEM);
3236 				ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3237 				if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3238 					th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3239 			}
3240 		} else {
3241 			ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
3242 			if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
3243 				th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
3244 		}
3245 		pi->ipi_ip_hlen = ip->ip_hl << 2;
3246 		pi->ipi_ipproto = ip->ip_p;
3247 		pi->ipi_flags |= IPI_TX_IPV4;
3248 
3249 		/* TCP checksum offload may require TCP header length */
3250 		if (IS_TX_OFFLOAD4(pi)) {
3251 			if (__predict_true(pi->ipi_ipproto == IPPROTO_TCP)) {
3252 				if (__predict_false(th == NULL)) {
3253 					txq->ift_pullups++;
3254 					if (__predict_false((m = m_pullup(m, (ip->ip_hl << 2) + sizeof(*th))) == NULL))
3255 						return (ENOMEM);
3256 					th = (struct tcphdr *)((caddr_t)ip + pi->ipi_ip_hlen);
3257 				}
3258 				pi->ipi_tcp_hflags = th->th_flags;
3259 				pi->ipi_tcp_hlen = th->th_off << 2;
3260 				pi->ipi_tcp_seq = th->th_seq;
3261 			}
3262 			if (IS_TSO4(pi)) {
3263 				if (__predict_false(ip->ip_p != IPPROTO_TCP))
3264 					return (ENXIO);
3265 				/*
3266 				 * TSO always requires hardware checksum offload.
3267 				 */
3268 				pi->ipi_csum_flags |= (CSUM_IP_TCP | CSUM_IP);
3269 				th->th_sum = in_pseudo(ip->ip_src.s_addr,
3270 						       ip->ip_dst.s_addr, htons(IPPROTO_TCP));
3271 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3272 				if (sctx->isc_flags & IFLIB_TSO_INIT_IP) {
3273 					ip->ip_sum = 0;
3274 					ip->ip_len = htons(pi->ipi_ip_hlen + pi->ipi_tcp_hlen + pi->ipi_tso_segsz);
3275 				}
3276 			}
3277 		}
3278 		if ((sctx->isc_flags & IFLIB_NEED_ZERO_CSUM) && (pi->ipi_csum_flags & CSUM_IP))
3279                        ip->ip_sum = 0;
3280 
3281 		break;
3282 	}
3283 #endif
3284 #ifdef INET6
3285 	case ETHERTYPE_IPV6:
3286 	{
3287 		struct ip6_hdr *ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen);
3288 		struct tcphdr *th;
3289 		pi->ipi_ip_hlen = sizeof(struct ip6_hdr);
3290 
3291 		if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) {
3292 			txq->ift_pullups++;
3293 			if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL))
3294 				return (ENOMEM);
3295 		}
3296 		th = (struct tcphdr *)((caddr_t)ip6 + pi->ipi_ip_hlen);
3297 
3298 		/* XXX-BZ this will go badly in case of ext hdrs. */
3299 		pi->ipi_ipproto = ip6->ip6_nxt;
3300 		pi->ipi_flags |= IPI_TX_IPV6;
3301 
3302 		/* TCP checksum offload may require TCP header length */
3303 		if (IS_TX_OFFLOAD6(pi)) {
3304 			if (pi->ipi_ipproto == IPPROTO_TCP) {
3305 				if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) {
3306 					txq->ift_pullups++;
3307 					if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) == NULL))
3308 						return (ENOMEM);
3309 				}
3310 				pi->ipi_tcp_hflags = th->th_flags;
3311 				pi->ipi_tcp_hlen = th->th_off << 2;
3312 				pi->ipi_tcp_seq = th->th_seq;
3313 			}
3314 			if (IS_TSO6(pi)) {
3315 				if (__predict_false(ip6->ip6_nxt != IPPROTO_TCP))
3316 					return (ENXIO);
3317 				/*
3318 				 * TSO always requires hardware checksum offload.
3319 				 */
3320 				pi->ipi_csum_flags |= CSUM_IP6_TCP;
3321 				th->th_sum = in6_cksum_pseudo(ip6, 0, IPPROTO_TCP, 0);
3322 				pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
3323 			}
3324 		}
3325 		break;
3326 	}
3327 #endif
3328 	default:
3329 		pi->ipi_csum_flags &= ~CSUM_OFFLOAD;
3330 		pi->ipi_ip_hlen = 0;
3331 		break;
3332 	}
3333 	*mp = m;
3334 
3335 	return (0);
3336 }
3337 
3338 /*
3339  * If dodgy hardware rejects the scatter gather chain we've handed it
3340  * we'll need to remove the mbuf chain from ifsg_m[] before we can add the
3341  * m_defrag'd mbufs
3342  */
3343 static __noinline struct mbuf *
3344 iflib_remove_mbuf(iflib_txq_t txq)
3345 {
3346 	int ntxd, pidx;
3347 	struct mbuf *m, **ifsd_m;
3348 
3349 	ifsd_m = txq->ift_sds.ifsd_m;
3350 	ntxd = txq->ift_size;
3351 	pidx = txq->ift_pidx & (ntxd - 1);
3352 	ifsd_m = txq->ift_sds.ifsd_m;
3353 	m = ifsd_m[pidx];
3354 	ifsd_m[pidx] = NULL;
3355 	bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[pidx]);
3356 	if (txq->ift_sds.ifsd_tso_map != NULL)
3357 		bus_dmamap_unload(txq->ift_tso_buf_tag,
3358 		    txq->ift_sds.ifsd_tso_map[pidx]);
3359 #if MEMORY_LOGGING
3360 	txq->ift_dequeued++;
3361 #endif
3362 	return (m);
3363 }
3364 
3365 static inline caddr_t
3366 calc_next_txd(iflib_txq_t txq, int cidx, uint8_t qid)
3367 {
3368 	qidx_t size;
3369 	int ntxd;
3370 	caddr_t start, end, cur, next;
3371 
3372 	ntxd = txq->ift_size;
3373 	size = txq->ift_txd_size[qid];
3374 	start = txq->ift_ifdi[qid].idi_vaddr;
3375 
3376 	if (__predict_false(size == 0))
3377 		return (start);
3378 	cur = start + size*cidx;
3379 	end = start + size*ntxd;
3380 	next = CACHE_PTR_NEXT(cur);
3381 	return (next < end ? next : start);
3382 }
3383 
3384 /*
3385  * Pad an mbuf to ensure a minimum ethernet frame size.
3386  * min_frame_size is the frame size (less CRC) to pad the mbuf to
3387  */
3388 static __noinline int
3389 iflib_ether_pad(device_t dev, struct mbuf **m_head, uint16_t min_frame_size)
3390 {
3391 	/*
3392 	 * 18 is enough bytes to pad an ARP packet to 46 bytes, and
3393 	 * and ARP message is the smallest common payload I can think of
3394 	 */
3395 	static char pad[18];	/* just zeros */
3396 	int n;
3397 	struct mbuf *new_head;
3398 
3399 	if (!M_WRITABLE(*m_head)) {
3400 		new_head = m_dup(*m_head, M_NOWAIT);
3401 		if (new_head == NULL) {
3402 			m_freem(*m_head);
3403 			device_printf(dev, "cannot pad short frame, m_dup() failed");
3404 			DBG_COUNTER_INC(encap_pad_mbuf_fail);
3405 			DBG_COUNTER_INC(tx_frees);
3406 			return ENOMEM;
3407 		}
3408 		m_freem(*m_head);
3409 		*m_head = new_head;
3410 	}
3411 
3412 	for (n = min_frame_size - (*m_head)->m_pkthdr.len;
3413 	     n > 0; n -= sizeof(pad))
3414 		if (!m_append(*m_head, min(n, sizeof(pad)), pad))
3415 			break;
3416 
3417 	if (n > 0) {
3418 		m_freem(*m_head);
3419 		device_printf(dev, "cannot pad short frame\n");
3420 		DBG_COUNTER_INC(encap_pad_mbuf_fail);
3421 		DBG_COUNTER_INC(tx_frees);
3422 		return (ENOBUFS);
3423 	}
3424 
3425 	return 0;
3426 }
3427 
3428 static int
3429 iflib_encap(iflib_txq_t txq, struct mbuf **m_headp)
3430 {
3431 	if_ctx_t		ctx;
3432 	if_shared_ctx_t		sctx;
3433 	if_softc_ctx_t		scctx;
3434 	bus_dma_tag_t		buf_tag;
3435 	bus_dma_segment_t	*segs;
3436 	struct mbuf		*m_head, **ifsd_m;
3437 	void			*next_txd;
3438 	bus_dmamap_t		map;
3439 	struct if_pkt_info	pi;
3440 	int remap = 0;
3441 	int err, nsegs, ndesc, max_segs, pidx, cidx, next, ntxd;
3442 
3443 	ctx = txq->ift_ctx;
3444 	sctx = ctx->ifc_sctx;
3445 	scctx = &ctx->ifc_softc_ctx;
3446 	segs = txq->ift_segs;
3447 	ntxd = txq->ift_size;
3448 	m_head = *m_headp;
3449 	map = NULL;
3450 
3451 	/*
3452 	 * If we're doing TSO the next descriptor to clean may be quite far ahead
3453 	 */
3454 	cidx = txq->ift_cidx;
3455 	pidx = txq->ift_pidx;
3456 	if (ctx->ifc_flags & IFC_PREFETCH) {
3457 		next = (cidx + CACHE_PTR_INCREMENT) & (ntxd-1);
3458 		if (!(ctx->ifc_flags & IFLIB_HAS_TXCQ)) {
3459 			next_txd = calc_next_txd(txq, cidx, 0);
3460 			prefetch(next_txd);
3461 		}
3462 
3463 		/* prefetch the next cache line of mbuf pointers and flags */
3464 		prefetch(&txq->ift_sds.ifsd_m[next]);
3465 		prefetch(&txq->ift_sds.ifsd_map[next]);
3466 		next = (cidx + CACHE_LINE_SIZE) & (ntxd-1);
3467 	}
3468 	map = txq->ift_sds.ifsd_map[pidx];
3469 	ifsd_m = txq->ift_sds.ifsd_m;
3470 
3471 	if (m_head->m_pkthdr.csum_flags & CSUM_TSO) {
3472 		buf_tag = txq->ift_tso_buf_tag;
3473 		max_segs = scctx->isc_tx_tso_segments_max;
3474 		map = txq->ift_sds.ifsd_tso_map[pidx];
3475 		MPASS(buf_tag != NULL);
3476 		MPASS(max_segs > 0);
3477 	} else {
3478 		buf_tag = txq->ift_buf_tag;
3479 		max_segs = scctx->isc_tx_nsegments;
3480 		map = txq->ift_sds.ifsd_map[pidx];
3481 	}
3482 	if ((sctx->isc_flags & IFLIB_NEED_ETHER_PAD) &&
3483 	    __predict_false(m_head->m_pkthdr.len < scctx->isc_min_frame_size)) {
3484 		err = iflib_ether_pad(ctx->ifc_dev, m_headp, scctx->isc_min_frame_size);
3485 		if (err) {
3486 			DBG_COUNTER_INC(encap_txd_encap_fail);
3487 			return err;
3488 		}
3489 	}
3490 	m_head = *m_headp;
3491 
3492 	pkt_info_zero(&pi);
3493 	pi.ipi_mflags = (m_head->m_flags & (M_VLANTAG|M_BCAST|M_MCAST));
3494 	pi.ipi_pidx = pidx;
3495 	pi.ipi_qsidx = txq->ift_id;
3496 	pi.ipi_len = m_head->m_pkthdr.len;
3497 	pi.ipi_csum_flags = m_head->m_pkthdr.csum_flags;
3498 	pi.ipi_vtag = M_HAS_VLANTAG(m_head) ? m_head->m_pkthdr.ether_vtag : 0;
3499 
3500 	/* deliberate bitwise OR to make one condition */
3501 	if (__predict_true((pi.ipi_csum_flags | pi.ipi_vtag))) {
3502 		if (__predict_false((err = iflib_parse_header(txq, &pi, m_headp)) != 0)) {
3503 			DBG_COUNTER_INC(encap_txd_encap_fail);
3504 			return (err);
3505 		}
3506 		m_head = *m_headp;
3507 	}
3508 
3509 retry:
3510 	err = bus_dmamap_load_mbuf_sg(buf_tag, map, m_head, segs, &nsegs,
3511 	    BUS_DMA_NOWAIT);
3512 defrag:
3513 	if (__predict_false(err)) {
3514 		switch (err) {
3515 		case EFBIG:
3516 			/* try collapse once and defrag once */
3517 			if (remap == 0) {
3518 				m_head = m_collapse(*m_headp, M_NOWAIT, max_segs);
3519 				/* try defrag if collapsing fails */
3520 				if (m_head == NULL)
3521 					remap++;
3522 			}
3523 			if (remap == 1) {
3524 				txq->ift_mbuf_defrag++;
3525 				m_head = m_defrag(*m_headp, M_NOWAIT);
3526 			}
3527 			/*
3528 			 * remap should never be >1 unless bus_dmamap_load_mbuf_sg
3529 			 * failed to map an mbuf that was run through m_defrag
3530 			 */
3531 			MPASS(remap <= 1);
3532 			if (__predict_false(m_head == NULL || remap > 1))
3533 				goto defrag_failed;
3534 			remap++;
3535 			*m_headp = m_head;
3536 			goto retry;
3537 			break;
3538 		case ENOMEM:
3539 			txq->ift_no_tx_dma_setup++;
3540 			break;
3541 		default:
3542 			txq->ift_no_tx_dma_setup++;
3543 			m_freem(*m_headp);
3544 			DBG_COUNTER_INC(tx_frees);
3545 			*m_headp = NULL;
3546 			break;
3547 		}
3548 		txq->ift_map_failed++;
3549 		DBG_COUNTER_INC(encap_load_mbuf_fail);
3550 		DBG_COUNTER_INC(encap_txd_encap_fail);
3551 		return (err);
3552 	}
3553 	ifsd_m[pidx] = m_head;
3554 	/*
3555 	 * XXX assumes a 1 to 1 relationship between segments and
3556 	 *        descriptors - this does not hold true on all drivers, e.g.
3557 	 *        cxgb
3558 	 */
3559 	if (__predict_false(nsegs + 2 > TXQ_AVAIL(txq))) {
3560 		txq->ift_no_desc_avail++;
3561 		bus_dmamap_unload(buf_tag, map);
3562 		DBG_COUNTER_INC(encap_txq_avail_fail);
3563 		DBG_COUNTER_INC(encap_txd_encap_fail);
3564 		if ((txq->ift_task.gt_task.ta_flags & TASK_ENQUEUED) == 0)
3565 			GROUPTASK_ENQUEUE(&txq->ift_task);
3566 		return (ENOBUFS);
3567 	}
3568 	/*
3569 	 * On Intel cards we can greatly reduce the number of TX interrupts
3570 	 * we see by only setting report status on every Nth descriptor.
3571 	 * However, this also means that the driver will need to keep track
3572 	 * of the descriptors that RS was set on to check them for the DD bit.
3573 	 */
3574 	txq->ift_rs_pending += nsegs + 1;
3575 	if (txq->ift_rs_pending > TXQ_MAX_RS_DEFERRED(txq) ||
3576 	     iflib_no_tx_batch || (TXQ_AVAIL(txq) - nsegs) <= MAX_TX_DESC(ctx) + 2) {
3577 		pi.ipi_flags |= IPI_TX_INTR;
3578 		txq->ift_rs_pending = 0;
3579 	}
3580 
3581 	pi.ipi_segs = segs;
3582 	pi.ipi_nsegs = nsegs;
3583 
3584 	MPASS(pidx >= 0 && pidx < txq->ift_size);
3585 #ifdef PKT_DEBUG
3586 	print_pkt(&pi);
3587 #endif
3588 	if ((err = ctx->isc_txd_encap(ctx->ifc_softc, &pi)) == 0) {
3589 		bus_dmamap_sync(buf_tag, map, BUS_DMASYNC_PREWRITE);
3590 		DBG_COUNTER_INC(tx_encap);
3591 		MPASS(pi.ipi_new_pidx < txq->ift_size);
3592 
3593 		ndesc = pi.ipi_new_pidx - pi.ipi_pidx;
3594 		if (pi.ipi_new_pidx < pi.ipi_pidx) {
3595 			ndesc += txq->ift_size;
3596 			txq->ift_gen = 1;
3597 		}
3598 		/*
3599 		 * drivers can need as many as
3600 		 * two sentinels
3601 		 */
3602 		MPASS(ndesc <= pi.ipi_nsegs + 2);
3603 		MPASS(pi.ipi_new_pidx != pidx);
3604 		MPASS(ndesc > 0);
3605 		txq->ift_in_use += ndesc;
3606 		txq->ift_db_pending += ndesc;
3607 
3608 		/*
3609 		 * We update the last software descriptor again here because there may
3610 		 * be a sentinel and/or there may be more mbufs than segments
3611 		 */
3612 		txq->ift_pidx = pi.ipi_new_pidx;
3613 		txq->ift_npending += pi.ipi_ndescs;
3614 	} else {
3615 		*m_headp = m_head = iflib_remove_mbuf(txq);
3616 		if (err == EFBIG) {
3617 			txq->ift_txd_encap_efbig++;
3618 			if (remap < 2) {
3619 				remap = 1;
3620 				goto defrag;
3621 			}
3622 		}
3623 		goto defrag_failed;
3624 	}
3625 	/*
3626 	 * err can't possibly be non-zero here, so we don't neet to test it
3627 	 * to see if we need to DBG_COUNTER_INC(encap_txd_encap_fail).
3628 	 */
3629 	return (err);
3630 
3631 defrag_failed:
3632 	txq->ift_mbuf_defrag_failed++;
3633 	txq->ift_map_failed++;
3634 	m_freem(*m_headp);
3635 	DBG_COUNTER_INC(tx_frees);
3636 	*m_headp = NULL;
3637 	DBG_COUNTER_INC(encap_txd_encap_fail);
3638 	return (ENOMEM);
3639 }
3640 
3641 static void
3642 iflib_tx_desc_free(iflib_txq_t txq, int n)
3643 {
3644 	uint32_t qsize, cidx, mask, gen;
3645 	struct mbuf *m, **ifsd_m;
3646 	bool do_prefetch;
3647 
3648 	cidx = txq->ift_cidx;
3649 	gen = txq->ift_gen;
3650 	qsize = txq->ift_size;
3651 	mask = qsize-1;
3652 	ifsd_m = txq->ift_sds.ifsd_m;
3653 	do_prefetch = (txq->ift_ctx->ifc_flags & IFC_PREFETCH);
3654 
3655 	while (n-- > 0) {
3656 		if (do_prefetch) {
3657 			prefetch(ifsd_m[(cidx + 3) & mask]);
3658 			prefetch(ifsd_m[(cidx + 4) & mask]);
3659 		}
3660 		if ((m = ifsd_m[cidx]) != NULL) {
3661 			prefetch(&ifsd_m[(cidx + CACHE_PTR_INCREMENT) & mask]);
3662 			if (m->m_pkthdr.csum_flags & CSUM_TSO) {
3663 				bus_dmamap_sync(txq->ift_tso_buf_tag,
3664 				    txq->ift_sds.ifsd_tso_map[cidx],
3665 				    BUS_DMASYNC_POSTWRITE);
3666 				bus_dmamap_unload(txq->ift_tso_buf_tag,
3667 				    txq->ift_sds.ifsd_tso_map[cidx]);
3668 			} else {
3669 				bus_dmamap_sync(txq->ift_buf_tag,
3670 				    txq->ift_sds.ifsd_map[cidx],
3671 				    BUS_DMASYNC_POSTWRITE);
3672 				bus_dmamap_unload(txq->ift_buf_tag,
3673 				    txq->ift_sds.ifsd_map[cidx]);
3674 			}
3675 			/* XXX we don't support any drivers that batch packets yet */
3676 			MPASS(m->m_nextpkt == NULL);
3677 			m_freem(m);
3678 			ifsd_m[cidx] = NULL;
3679 #if MEMORY_LOGGING
3680 			txq->ift_dequeued++;
3681 #endif
3682 			DBG_COUNTER_INC(tx_frees);
3683 		}
3684 		if (__predict_false(++cidx == qsize)) {
3685 			cidx = 0;
3686 			gen = 0;
3687 		}
3688 	}
3689 	txq->ift_cidx = cidx;
3690 	txq->ift_gen = gen;
3691 }
3692 
3693 static __inline int
3694 iflib_completed_tx_reclaim(iflib_txq_t txq, int thresh)
3695 {
3696 	int reclaim;
3697 	if_ctx_t ctx = txq->ift_ctx;
3698 
3699 	KASSERT(thresh >= 0, ("invalid threshold to reclaim"));
3700 	MPASS(thresh /*+ MAX_TX_DESC(txq->ift_ctx) */ < txq->ift_size);
3701 
3702 	/*
3703 	 * Need a rate-limiting check so that this isn't called every time
3704 	 */
3705 	iflib_tx_credits_update(ctx, txq);
3706 	reclaim = DESC_RECLAIMABLE(txq);
3707 
3708 	if (reclaim <= thresh /* + MAX_TX_DESC(txq->ift_ctx) */) {
3709 #ifdef INVARIANTS
3710 		if (iflib_verbose_debug) {
3711 			printf("%s processed=%ju cleaned=%ju tx_nsegments=%d reclaim=%d thresh=%d\n", __FUNCTION__,
3712 			       txq->ift_processed, txq->ift_cleaned, txq->ift_ctx->ifc_softc_ctx.isc_tx_nsegments,
3713 			       reclaim, thresh);
3714 		}
3715 #endif
3716 		return (0);
3717 	}
3718 	iflib_tx_desc_free(txq, reclaim);
3719 	txq->ift_cleaned += reclaim;
3720 	txq->ift_in_use -= reclaim;
3721 
3722 	return (reclaim);
3723 }
3724 
3725 static struct mbuf **
3726 _ring_peek_one(struct ifmp_ring *r, int cidx, int offset, int remaining)
3727 {
3728 	int next, size;
3729 	struct mbuf **items;
3730 
3731 	size = r->size;
3732 	next = (cidx + CACHE_PTR_INCREMENT) & (size-1);
3733 	items = __DEVOLATILE(struct mbuf **, &r->items[0]);
3734 
3735 	prefetch(items[(cidx + offset) & (size-1)]);
3736 	if (remaining > 1) {
3737 		prefetch2cachelines(&items[next]);
3738 		prefetch2cachelines(items[(cidx + offset + 1) & (size-1)]);
3739 		prefetch2cachelines(items[(cidx + offset + 2) & (size-1)]);
3740 		prefetch2cachelines(items[(cidx + offset + 3) & (size-1)]);
3741 	}
3742 	return (__DEVOLATILE(struct mbuf **, &r->items[(cidx + offset) & (size-1)]));
3743 }
3744 
3745 static void
3746 iflib_txq_check_drain(iflib_txq_t txq, int budget)
3747 {
3748 
3749 	ifmp_ring_check_drainage(txq->ift_br, budget);
3750 }
3751 
3752 static uint32_t
3753 iflib_txq_can_drain(struct ifmp_ring *r)
3754 {
3755 	iflib_txq_t txq = r->cookie;
3756 	if_ctx_t ctx = txq->ift_ctx;
3757 
3758 	if (TXQ_AVAIL(txq) > MAX_TX_DESC(ctx) + 2)
3759 		return (1);
3760 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3761 	    BUS_DMASYNC_POSTREAD);
3762 	return (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id,
3763 	    false));
3764 }
3765 
3766 static uint32_t
3767 iflib_txq_drain(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3768 {
3769 	iflib_txq_t txq = r->cookie;
3770 	if_ctx_t ctx = txq->ift_ctx;
3771 	if_t ifp = ctx->ifc_ifp;
3772 	struct mbuf *m, **mp;
3773 	int avail, bytes_sent, skipped, count, err, i;
3774 	int mcast_sent, pkt_sent, reclaimed;
3775 	bool do_prefetch, rang, ring;
3776 
3777 	if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING) ||
3778 			    !LINK_ACTIVE(ctx))) {
3779 		DBG_COUNTER_INC(txq_drain_notready);
3780 		return (0);
3781 	}
3782 	reclaimed = iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
3783 	rang = iflib_txd_db_check(txq, reclaimed && txq->ift_db_pending);
3784 	avail = IDXDIFF(pidx, cidx, r->size);
3785 
3786 	if (__predict_false(ctx->ifc_flags & IFC_QFLUSH)) {
3787 		/*
3788 		 * The driver is unloading so we need to free all pending packets.
3789 		 */
3790 		DBG_COUNTER_INC(txq_drain_flushing);
3791 		for (i = 0; i < avail; i++) {
3792 			if (__predict_true(r->items[(cidx + i) & (r->size-1)] != (void *)txq))
3793 				m_freem(r->items[(cidx + i) & (r->size-1)]);
3794 			r->items[(cidx + i) & (r->size-1)] = NULL;
3795 		}
3796 		return (avail);
3797 	}
3798 
3799 	if (__predict_false(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) {
3800 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3801 		CALLOUT_LOCK(txq);
3802 		callout_stop(&txq->ift_timer);
3803 		CALLOUT_UNLOCK(txq);
3804 		DBG_COUNTER_INC(txq_drain_oactive);
3805 		return (0);
3806 	}
3807 
3808 	/*
3809 	 * If we've reclaimed any packets this queue cannot be hung.
3810 	 */
3811 	if (reclaimed)
3812 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3813 	skipped = mcast_sent = bytes_sent = pkt_sent = 0;
3814 	count = MIN(avail, TX_BATCH_SIZE);
3815 #ifdef INVARIANTS
3816 	if (iflib_verbose_debug)
3817 		printf("%s avail=%d ifc_flags=%x txq_avail=%d ", __FUNCTION__,
3818 		       avail, ctx->ifc_flags, TXQ_AVAIL(txq));
3819 #endif
3820 	do_prefetch = (ctx->ifc_flags & IFC_PREFETCH);
3821 	err = 0;
3822 	for (i = 0; i < count && TXQ_AVAIL(txq) >= MAX_TX_DESC(ctx) + 2; i++) {
3823 		int rem = do_prefetch ? count - i : 0;
3824 
3825 		mp = _ring_peek_one(r, cidx, i, rem);
3826 		MPASS(mp != NULL && *mp != NULL);
3827 
3828 		/*
3829 		 * Completion interrupts will use the address of the txq
3830 		 * as a sentinel to enqueue _something_ in order to acquire
3831 		 * the lock on the mp_ring (there's no direct lock call).
3832 		 * We obviously whave to check for these sentinel cases
3833 		 * and skip them.
3834 		 */
3835 		if (__predict_false(*mp == (struct mbuf *)txq)) {
3836 			skipped++;
3837 			continue;
3838 		}
3839 		err = iflib_encap(txq, mp);
3840 		if (__predict_false(err)) {
3841 			/* no room - bail out */
3842 			if (err == ENOBUFS)
3843 				break;
3844 			skipped++;
3845 			/* we can't send this packet - skip it */
3846 			continue;
3847 		}
3848 		pkt_sent++;
3849 		m = *mp;
3850 		DBG_COUNTER_INC(tx_sent);
3851 		bytes_sent += m->m_pkthdr.len;
3852 		mcast_sent += !!(m->m_flags & M_MCAST);
3853 
3854 		if (__predict_false(!(ifp->if_drv_flags & IFF_DRV_RUNNING)))
3855 			break;
3856 		ETHER_BPF_MTAP(ifp, m);
3857 		rang = iflib_txd_db_check(txq, false);
3858 	}
3859 
3860 	/* deliberate use of bitwise or to avoid gratuitous short-circuit */
3861 	ring = rang ? false  : (iflib_min_tx_latency | err);
3862 	iflib_txd_db_check(txq, ring);
3863 	if_inc_counter(ifp, IFCOUNTER_OBYTES, bytes_sent);
3864 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, pkt_sent);
3865 	if (mcast_sent)
3866 		if_inc_counter(ifp, IFCOUNTER_OMCASTS, mcast_sent);
3867 #ifdef INVARIANTS
3868 	if (iflib_verbose_debug)
3869 		printf("consumed=%d\n", skipped + pkt_sent);
3870 #endif
3871 	return (skipped + pkt_sent);
3872 }
3873 
3874 static uint32_t
3875 iflib_txq_drain_always(struct ifmp_ring *r)
3876 {
3877 	return (1);
3878 }
3879 
3880 static uint32_t
3881 iflib_txq_drain_free(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3882 {
3883 	int i, avail;
3884 	struct mbuf **mp;
3885 	iflib_txq_t txq;
3886 
3887 	txq = r->cookie;
3888 
3889 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3890 	CALLOUT_LOCK(txq);
3891 	callout_stop(&txq->ift_timer);
3892 	CALLOUT_UNLOCK(txq);
3893 
3894 	avail = IDXDIFF(pidx, cidx, r->size);
3895 	for (i = 0; i < avail; i++) {
3896 		mp = _ring_peek_one(r, cidx, i, avail - i);
3897 		if (__predict_false(*mp == (struct mbuf *)txq))
3898 			continue;
3899 		m_freem(*mp);
3900 		DBG_COUNTER_INC(tx_frees);
3901 	}
3902 	MPASS(ifmp_ring_is_stalled(r) == 0);
3903 	return (avail);
3904 }
3905 
3906 static void
3907 iflib_ifmp_purge(iflib_txq_t txq)
3908 {
3909 	struct ifmp_ring *r;
3910 
3911 	r = txq->ift_br;
3912 	r->drain = iflib_txq_drain_free;
3913 	r->can_drain = iflib_txq_drain_always;
3914 
3915 	ifmp_ring_check_drainage(r, r->size);
3916 
3917 	r->drain = iflib_txq_drain;
3918 	r->can_drain = iflib_txq_can_drain;
3919 }
3920 
3921 static void
3922 _task_fn_tx(void *context)
3923 {
3924 	iflib_txq_t txq = context;
3925 	if_ctx_t ctx = txq->ift_ctx;
3926 	if_t ifp = ctx->ifc_ifp;
3927 	int abdicate = ctx->ifc_sysctl_tx_abdicate;
3928 
3929 #ifdef IFLIB_DIAGNOSTICS
3930 	txq->ift_cpu_exec_count[curcpu]++;
3931 #endif
3932 	if (!(if_getdrvflags(ifp) & IFF_DRV_RUNNING))
3933 		return;
3934 #ifdef DEV_NETMAP
3935 	if ((if_getcapenable(ifp) & IFCAP_NETMAP) &&
3936 	    netmap_tx_irq(ifp, txq->ift_id))
3937 		goto skip_ifmp;
3938 #endif
3939 #ifdef ALTQ
3940 	if (ALTQ_IS_ENABLED(&ifp->if_snd))
3941 		iflib_altq_if_start(ifp);
3942 #endif
3943 	if (txq->ift_db_pending)
3944 		ifmp_ring_enqueue(txq->ift_br, (void **)&txq, 1, TX_BATCH_SIZE, abdicate);
3945 	else if (!abdicate)
3946 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3947 	/*
3948 	 * When abdicating, we always need to check drainage, not just when we don't enqueue
3949 	 */
3950 	if (abdicate)
3951 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3952 #ifdef DEV_NETMAP
3953 skip_ifmp:
3954 #endif
3955 	if (ctx->ifc_flags & IFC_LEGACY)
3956 		IFDI_INTR_ENABLE(ctx);
3957 	else
3958 		IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3959 }
3960 
3961 static void
3962 _task_fn_rx(void *context)
3963 {
3964 	iflib_rxq_t rxq = context;
3965 	if_ctx_t ctx = rxq->ifr_ctx;
3966 	uint8_t more;
3967 	uint16_t budget;
3968 #ifdef DEV_NETMAP
3969 	u_int work = 0;
3970 	int nmirq;
3971 #endif
3972 
3973 #ifdef IFLIB_DIAGNOSTICS
3974 	rxq->ifr_cpu_exec_count[curcpu]++;
3975 #endif
3976 	DBG_COUNTER_INC(task_fn_rxs);
3977 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3978 		return;
3979 #ifdef DEV_NETMAP
3980 	nmirq = netmap_rx_irq(ctx->ifc_ifp, rxq->ifr_id, &work);
3981 	if (nmirq != NM_IRQ_PASS) {
3982 		more = (nmirq == NM_IRQ_RESCHED) ? IFLIB_RXEOF_MORE : 0;
3983 		goto skip_rxeof;
3984 	}
3985 #endif
3986 	budget = ctx->ifc_sysctl_rx_budget;
3987 	if (budget == 0)
3988 		budget = 16;	/* XXX */
3989 	more = iflib_rxeof(rxq, budget);
3990 #ifdef DEV_NETMAP
3991 skip_rxeof:
3992 #endif
3993 	if ((more & IFLIB_RXEOF_MORE) == 0) {
3994 		if (ctx->ifc_flags & IFC_LEGACY)
3995 			IFDI_INTR_ENABLE(ctx);
3996 		else
3997 			IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
3998 		DBG_COUNTER_INC(rx_intr_enables);
3999 	}
4000 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
4001 		return;
4002 
4003 	if (more & IFLIB_RXEOF_MORE)
4004 		GROUPTASK_ENQUEUE(&rxq->ifr_task);
4005 	else if (more & IFLIB_RXEOF_EMPTY)
4006 		callout_reset_curcpu(&rxq->ifr_watchdog, 1, &_task_fn_rx_watchdog, rxq);
4007 }
4008 
4009 static void
4010 _task_fn_admin(void *context)
4011 {
4012 	if_ctx_t ctx = context;
4013 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
4014 	iflib_txq_t txq;
4015 	int i;
4016 	bool oactive, running, do_reset, do_watchdog, in_detach;
4017 
4018 	STATE_LOCK(ctx);
4019 	running = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING);
4020 	oactive = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE);
4021 	do_reset = (ctx->ifc_flags & IFC_DO_RESET);
4022 	do_watchdog = (ctx->ifc_flags & IFC_DO_WATCHDOG);
4023 	in_detach = (ctx->ifc_flags & IFC_IN_DETACH);
4024 	ctx->ifc_flags &= ~(IFC_DO_RESET|IFC_DO_WATCHDOG);
4025 	STATE_UNLOCK(ctx);
4026 
4027 	if ((!running && !oactive) && !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
4028 		return;
4029 	if (in_detach)
4030 		return;
4031 
4032 	CTX_LOCK(ctx);
4033 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
4034 		CALLOUT_LOCK(txq);
4035 		callout_stop(&txq->ift_timer);
4036 		CALLOUT_UNLOCK(txq);
4037 	}
4038 	if (do_watchdog) {
4039 		ctx->ifc_watchdog_events++;
4040 		IFDI_WATCHDOG_RESET(ctx);
4041 	}
4042 	IFDI_UPDATE_ADMIN_STATUS(ctx);
4043 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
4044 		callout_reset_on(&txq->ift_timer, iflib_timer_default, iflib_timer, txq,
4045 		    txq->ift_timer.c_cpu);
4046 	}
4047 	IFDI_LINK_INTR_ENABLE(ctx);
4048 	if (do_reset)
4049 		iflib_if_init_locked(ctx);
4050 	CTX_UNLOCK(ctx);
4051 
4052 	if (LINK_ACTIVE(ctx) == 0)
4053 		return;
4054 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
4055 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
4056 }
4057 
4058 static void
4059 _task_fn_iov(void *context)
4060 {
4061 	if_ctx_t ctx = context;
4062 
4063 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING) &&
4064 	    !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN))
4065 		return;
4066 
4067 	CTX_LOCK(ctx);
4068 	IFDI_VFLR_HANDLE(ctx);
4069 	CTX_UNLOCK(ctx);
4070 }
4071 
4072 static int
4073 iflib_sysctl_int_delay(SYSCTL_HANDLER_ARGS)
4074 {
4075 	int err;
4076 	if_int_delay_info_t info;
4077 	if_ctx_t ctx;
4078 
4079 	info = (if_int_delay_info_t)arg1;
4080 	ctx = info->iidi_ctx;
4081 	info->iidi_req = req;
4082 	info->iidi_oidp = oidp;
4083 	CTX_LOCK(ctx);
4084 	err = IFDI_SYSCTL_INT_DELAY(ctx, info);
4085 	CTX_UNLOCK(ctx);
4086 	return (err);
4087 }
4088 
4089 /*********************************************************************
4090  *
4091  *  IFNET FUNCTIONS
4092  *
4093  **********************************************************************/
4094 
4095 static void
4096 iflib_if_init_locked(if_ctx_t ctx)
4097 {
4098 	iflib_stop(ctx);
4099 	iflib_init_locked(ctx);
4100 }
4101 
4102 static void
4103 iflib_if_init(void *arg)
4104 {
4105 	if_ctx_t ctx = arg;
4106 
4107 	CTX_LOCK(ctx);
4108 	iflib_if_init_locked(ctx);
4109 	CTX_UNLOCK(ctx);
4110 }
4111 
4112 static int
4113 iflib_if_transmit(if_t ifp, struct mbuf *m)
4114 {
4115 	if_ctx_t	ctx = if_getsoftc(ifp);
4116 
4117 	iflib_txq_t txq;
4118 	int err, qidx;
4119 	int abdicate = ctx->ifc_sysctl_tx_abdicate;
4120 
4121 	if (__predict_false((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0 || !LINK_ACTIVE(ctx))) {
4122 		DBG_COUNTER_INC(tx_frees);
4123 		m_freem(m);
4124 		return (ENETDOWN);
4125 	}
4126 
4127 	MPASS(m->m_nextpkt == NULL);
4128 	/* ALTQ-enabled interfaces always use queue 0. */
4129 	qidx = 0;
4130 	if ((NTXQSETS(ctx) > 1) && M_HASHTYPE_GET(m) && !ALTQ_IS_ENABLED(&ifp->if_snd))
4131 		qidx = QIDX(ctx, m);
4132 	/*
4133 	 * XXX calculate buf_ring based on flowid (divvy up bits?)
4134 	 */
4135 	txq = &ctx->ifc_txqs[qidx];
4136 
4137 #ifdef DRIVER_BACKPRESSURE
4138 	if (txq->ift_closed) {
4139 		while (m != NULL) {
4140 			next = m->m_nextpkt;
4141 			m->m_nextpkt = NULL;
4142 			m_freem(m);
4143 			DBG_COUNTER_INC(tx_frees);
4144 			m = next;
4145 		}
4146 		return (ENOBUFS);
4147 	}
4148 #endif
4149 #ifdef notyet
4150 	qidx = count = 0;
4151 	mp = marr;
4152 	next = m;
4153 	do {
4154 		count++;
4155 		next = next->m_nextpkt;
4156 	} while (next != NULL);
4157 
4158 	if (count > nitems(marr))
4159 		if ((mp = malloc(count*sizeof(struct mbuf *), M_IFLIB, M_NOWAIT)) == NULL) {
4160 			/* XXX check nextpkt */
4161 			m_freem(m);
4162 			/* XXX simplify for now */
4163 			DBG_COUNTER_INC(tx_frees);
4164 			return (ENOBUFS);
4165 		}
4166 	for (next = m, i = 0; next != NULL; i++) {
4167 		mp[i] = next;
4168 		next = next->m_nextpkt;
4169 		mp[i]->m_nextpkt = NULL;
4170 	}
4171 #endif
4172 	DBG_COUNTER_INC(tx_seen);
4173 	err = ifmp_ring_enqueue(txq->ift_br, (void **)&m, 1, TX_BATCH_SIZE, abdicate);
4174 
4175 	if (abdicate)
4176 		GROUPTASK_ENQUEUE(&txq->ift_task);
4177  	if (err) {
4178 		if (!abdicate)
4179 			GROUPTASK_ENQUEUE(&txq->ift_task);
4180 		/* support forthcoming later */
4181 #ifdef DRIVER_BACKPRESSURE
4182 		txq->ift_closed = TRUE;
4183 #endif
4184 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
4185 		m_freem(m);
4186 		DBG_COUNTER_INC(tx_frees);
4187 	}
4188 
4189 	return (err);
4190 }
4191 
4192 #ifdef ALTQ
4193 /*
4194  * The overall approach to integrating iflib with ALTQ is to continue to use
4195  * the iflib mp_ring machinery between the ALTQ queue(s) and the hardware
4196  * ring.  Technically, when using ALTQ, queueing to an intermediate mp_ring
4197  * is redundant/unnecessary, but doing so minimizes the amount of
4198  * ALTQ-specific code required in iflib.  It is assumed that the overhead of
4199  * redundantly queueing to an intermediate mp_ring is swamped by the
4200  * performance limitations inherent in using ALTQ.
4201  *
4202  * When ALTQ support is compiled in, all iflib drivers will use a transmit
4203  * routine, iflib_altq_if_transmit(), that checks if ALTQ is enabled for the
4204  * given interface.  If ALTQ is enabled for an interface, then all
4205  * transmitted packets for that interface will be submitted to the ALTQ
4206  * subsystem via IFQ_ENQUEUE().  We don't use the legacy if_transmit()
4207  * implementation because it uses IFQ_HANDOFF(), which will duplicatively
4208  * update stats that the iflib machinery handles, and which is sensitve to
4209  * the disused IFF_DRV_OACTIVE flag.  Additionally, iflib_altq_if_start()
4210  * will be installed as the start routine for use by ALTQ facilities that
4211  * need to trigger queue drains on a scheduled basis.
4212  *
4213  */
4214 static void
4215 iflib_altq_if_start(if_t ifp)
4216 {
4217 	struct ifaltq *ifq = &ifp->if_snd;
4218 	struct mbuf *m;
4219 
4220 	IFQ_LOCK(ifq);
4221 	IFQ_DEQUEUE_NOLOCK(ifq, m);
4222 	while (m != NULL) {
4223 		iflib_if_transmit(ifp, m);
4224 		IFQ_DEQUEUE_NOLOCK(ifq, m);
4225 	}
4226 	IFQ_UNLOCK(ifq);
4227 }
4228 
4229 static int
4230 iflib_altq_if_transmit(if_t ifp, struct mbuf *m)
4231 {
4232 	int err;
4233 
4234 	if (ALTQ_IS_ENABLED(&ifp->if_snd)) {
4235 		IFQ_ENQUEUE(&ifp->if_snd, m, err);
4236 		if (err == 0)
4237 			iflib_altq_if_start(ifp);
4238 	} else
4239 		err = iflib_if_transmit(ifp, m);
4240 
4241 	return (err);
4242 }
4243 #endif /* ALTQ */
4244 
4245 static void
4246 iflib_if_qflush(if_t ifp)
4247 {
4248 	if_ctx_t ctx = if_getsoftc(ifp);
4249 	iflib_txq_t txq = ctx->ifc_txqs;
4250 	int i;
4251 
4252 	STATE_LOCK(ctx);
4253 	ctx->ifc_flags |= IFC_QFLUSH;
4254 	STATE_UNLOCK(ctx);
4255 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
4256 		while (!(ifmp_ring_is_idle(txq->ift_br) || ifmp_ring_is_stalled(txq->ift_br)))
4257 			iflib_txq_check_drain(txq, 0);
4258 	STATE_LOCK(ctx);
4259 	ctx->ifc_flags &= ~IFC_QFLUSH;
4260 	STATE_UNLOCK(ctx);
4261 
4262 	/*
4263 	 * When ALTQ is enabled, this will also take care of purging the
4264 	 * ALTQ queue(s).
4265 	 */
4266 	if_qflush(ifp);
4267 }
4268 
4269 #define IFCAP_FLAGS (IFCAP_HWCSUM_IPV6 | IFCAP_HWCSUM | IFCAP_LRO | \
4270 		     IFCAP_TSO | IFCAP_VLAN_HWTAGGING | IFCAP_HWSTATS | \
4271 		     IFCAP_VLAN_MTU | IFCAP_VLAN_HWFILTER | \
4272 		     IFCAP_VLAN_HWTSO | IFCAP_VLAN_HWCSUM | IFCAP_MEXTPG)
4273 
4274 static int
4275 iflib_if_ioctl(if_t ifp, u_long command, caddr_t data)
4276 {
4277 	if_ctx_t ctx = if_getsoftc(ifp);
4278 	struct ifreq	*ifr = (struct ifreq *)data;
4279 #if defined(INET) || defined(INET6)
4280 	struct ifaddr	*ifa = (struct ifaddr *)data;
4281 #endif
4282 	bool		avoid_reset = false;
4283 	int		err = 0, reinit = 0, bits;
4284 
4285 	switch (command) {
4286 	case SIOCSIFADDR:
4287 #ifdef INET
4288 		if (ifa->ifa_addr->sa_family == AF_INET)
4289 			avoid_reset = true;
4290 #endif
4291 #ifdef INET6
4292 		if (ifa->ifa_addr->sa_family == AF_INET6)
4293 			avoid_reset = true;
4294 #endif
4295 		/*
4296 		** Calling init results in link renegotiation,
4297 		** so we avoid doing it when possible.
4298 		*/
4299 		if (avoid_reset) {
4300 			if_setflagbits(ifp, IFF_UP,0);
4301 			if (!(if_getdrvflags(ifp) & IFF_DRV_RUNNING))
4302 				reinit = 1;
4303 #ifdef INET
4304 			if (!(if_getflags(ifp) & IFF_NOARP))
4305 				arp_ifinit(ifp, ifa);
4306 #endif
4307 		} else
4308 			err = ether_ioctl(ifp, command, data);
4309 		break;
4310 	case SIOCSIFMTU:
4311 		CTX_LOCK(ctx);
4312 		if (ifr->ifr_mtu == if_getmtu(ifp)) {
4313 			CTX_UNLOCK(ctx);
4314 			break;
4315 		}
4316 		bits = if_getdrvflags(ifp);
4317 		/* stop the driver and free any clusters before proceeding */
4318 		iflib_stop(ctx);
4319 
4320 		if ((err = IFDI_MTU_SET(ctx, ifr->ifr_mtu)) == 0) {
4321 			STATE_LOCK(ctx);
4322 			if (ifr->ifr_mtu > ctx->ifc_max_fl_buf_size)
4323 				ctx->ifc_flags |= IFC_MULTISEG;
4324 			else
4325 				ctx->ifc_flags &= ~IFC_MULTISEG;
4326 			STATE_UNLOCK(ctx);
4327 			err = if_setmtu(ifp, ifr->ifr_mtu);
4328 		}
4329 		iflib_init_locked(ctx);
4330 		STATE_LOCK(ctx);
4331 		if_setdrvflags(ifp, bits);
4332 		STATE_UNLOCK(ctx);
4333 		CTX_UNLOCK(ctx);
4334 		break;
4335 	case SIOCSIFFLAGS:
4336 		CTX_LOCK(ctx);
4337 		if (if_getflags(ifp) & IFF_UP) {
4338 			if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4339 				if ((if_getflags(ifp) ^ ctx->ifc_if_flags) &
4340 				    (IFF_PROMISC | IFF_ALLMULTI)) {
4341 					CTX_UNLOCK(ctx);
4342 					err = IFDI_PROMISC_SET(ctx, if_getflags(ifp));
4343 					CTX_LOCK(ctx);
4344 				}
4345 			} else
4346 				reinit = 1;
4347 		} else if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4348 			iflib_stop(ctx);
4349 		}
4350 		ctx->ifc_if_flags = if_getflags(ifp);
4351 		CTX_UNLOCK(ctx);
4352 		break;
4353 	case SIOCADDMULTI:
4354 	case SIOCDELMULTI:
4355 		if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
4356 			CTX_LOCK(ctx);
4357 			IFDI_INTR_DISABLE(ctx);
4358 			IFDI_MULTI_SET(ctx);
4359 			IFDI_INTR_ENABLE(ctx);
4360 			CTX_UNLOCK(ctx);
4361 		}
4362 		break;
4363 	case SIOCSIFMEDIA:
4364 		CTX_LOCK(ctx);
4365 		IFDI_MEDIA_SET(ctx);
4366 		CTX_UNLOCK(ctx);
4367 		/* FALLTHROUGH */
4368 	case SIOCGIFMEDIA:
4369 	case SIOCGIFXMEDIA:
4370 		err = ifmedia_ioctl(ifp, ifr, ctx->ifc_mediap, command);
4371 		break;
4372 	case SIOCGI2C:
4373 	{
4374 		struct ifi2creq i2c;
4375 
4376 		err = copyin(ifr_data_get_ptr(ifr), &i2c, sizeof(i2c));
4377 		if (err != 0)
4378 			break;
4379 		if (i2c.dev_addr != 0xA0 && i2c.dev_addr != 0xA2) {
4380 			err = EINVAL;
4381 			break;
4382 		}
4383 		if (i2c.len > sizeof(i2c.data)) {
4384 			err = EINVAL;
4385 			break;
4386 		}
4387 
4388 		if ((err = IFDI_I2C_REQ(ctx, &i2c)) == 0)
4389 			err = copyout(&i2c, ifr_data_get_ptr(ifr),
4390 			    sizeof(i2c));
4391 		break;
4392 	}
4393 	case SIOCSIFCAP:
4394 	{
4395 		int mask, setmask, oldmask;
4396 
4397 		oldmask = if_getcapenable(ifp);
4398 		mask = ifr->ifr_reqcap ^ oldmask;
4399 		mask &= ctx->ifc_softc_ctx.isc_capabilities | IFCAP_MEXTPG;
4400 		setmask = 0;
4401 #ifdef TCP_OFFLOAD
4402 		setmask |= mask & (IFCAP_TOE4|IFCAP_TOE6);
4403 #endif
4404 		setmask |= (mask & IFCAP_FLAGS);
4405 		setmask |= (mask & IFCAP_WOL);
4406 
4407 		/*
4408 		 * If any RX csum has changed, change all the ones that
4409 		 * are supported by the driver.
4410 		 */
4411 		if (setmask & (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6)) {
4412 			setmask |= ctx->ifc_softc_ctx.isc_capabilities &
4413 			    (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6);
4414 		}
4415 
4416 		/*
4417 		 * want to ensure that traffic has stopped before we change any of the flags
4418 		 */
4419 		if (setmask) {
4420 			CTX_LOCK(ctx);
4421 			bits = if_getdrvflags(ifp);
4422 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4423 				iflib_stop(ctx);
4424 			STATE_LOCK(ctx);
4425 			if_togglecapenable(ifp, setmask);
4426 			ctx->ifc_softc_ctx.isc_capenable ^= setmask;
4427 			STATE_UNLOCK(ctx);
4428 			if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL)
4429 				iflib_init_locked(ctx);
4430 			STATE_LOCK(ctx);
4431 			if_setdrvflags(ifp, bits);
4432 			STATE_UNLOCK(ctx);
4433 			CTX_UNLOCK(ctx);
4434 		}
4435 		if_vlancap(ifp);
4436 		break;
4437 	}
4438 	case SIOCGPRIVATE_0:
4439 	case SIOCSDRVSPEC:
4440 	case SIOCGDRVSPEC:
4441 		CTX_LOCK(ctx);
4442 		err = IFDI_PRIV_IOCTL(ctx, command, data);
4443 		CTX_UNLOCK(ctx);
4444 		break;
4445 	default:
4446 		err = ether_ioctl(ifp, command, data);
4447 		break;
4448 	}
4449 	if (reinit)
4450 		iflib_if_init(ctx);
4451 	return (err);
4452 }
4453 
4454 static uint64_t
4455 iflib_if_get_counter(if_t ifp, ift_counter cnt)
4456 {
4457 	if_ctx_t ctx = if_getsoftc(ifp);
4458 
4459 	return (IFDI_GET_COUNTER(ctx, cnt));
4460 }
4461 
4462 /*********************************************************************
4463  *
4464  *  OTHER FUNCTIONS EXPORTED TO THE STACK
4465  *
4466  **********************************************************************/
4467 
4468 static void
4469 iflib_vlan_register(void *arg, if_t ifp, uint16_t vtag)
4470 {
4471 	if_ctx_t ctx = if_getsoftc(ifp);
4472 
4473 	if ((void *)ctx != arg)
4474 		return;
4475 
4476 	if ((vtag == 0) || (vtag > 4095))
4477 		return;
4478 
4479 	if (iflib_in_detach(ctx))
4480 		return;
4481 
4482 	CTX_LOCK(ctx);
4483 	/* Driver may need all untagged packets to be flushed */
4484 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4485 		iflib_stop(ctx);
4486 	IFDI_VLAN_REGISTER(ctx, vtag);
4487 	/* Re-init to load the changes, if required */
4488 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4489 		iflib_init_locked(ctx);
4490 	CTX_UNLOCK(ctx);
4491 }
4492 
4493 static void
4494 iflib_vlan_unregister(void *arg, if_t ifp, uint16_t vtag)
4495 {
4496 	if_ctx_t ctx = if_getsoftc(ifp);
4497 
4498 	if ((void *)ctx != arg)
4499 		return;
4500 
4501 	if ((vtag == 0) || (vtag > 4095))
4502 		return;
4503 
4504 	CTX_LOCK(ctx);
4505 	/* Driver may need all tagged packets to be flushed */
4506 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4507 		iflib_stop(ctx);
4508 	IFDI_VLAN_UNREGISTER(ctx, vtag);
4509 	/* Re-init to load the changes, if required */
4510 	if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG))
4511 		iflib_init_locked(ctx);
4512 	CTX_UNLOCK(ctx);
4513 }
4514 
4515 static void
4516 iflib_led_func(void *arg, int onoff)
4517 {
4518 	if_ctx_t ctx = arg;
4519 
4520 	CTX_LOCK(ctx);
4521 	IFDI_LED_FUNC(ctx, onoff);
4522 	CTX_UNLOCK(ctx);
4523 }
4524 
4525 /*********************************************************************
4526  *
4527  *  BUS FUNCTION DEFINITIONS
4528  *
4529  **********************************************************************/
4530 
4531 int
4532 iflib_device_probe(device_t dev)
4533 {
4534 	const pci_vendor_info_t *ent;
4535 	if_shared_ctx_t sctx;
4536 	uint16_t pci_device_id, pci_rev_id, pci_subdevice_id, pci_subvendor_id;
4537 	uint16_t pci_vendor_id;
4538 
4539 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4540 		return (ENOTSUP);
4541 
4542 	pci_vendor_id = pci_get_vendor(dev);
4543 	pci_device_id = pci_get_device(dev);
4544 	pci_subvendor_id = pci_get_subvendor(dev);
4545 	pci_subdevice_id = pci_get_subdevice(dev);
4546 	pci_rev_id = pci_get_revid(dev);
4547 	if (sctx->isc_parse_devinfo != NULL)
4548 		sctx->isc_parse_devinfo(&pci_device_id, &pci_subvendor_id, &pci_subdevice_id, &pci_rev_id);
4549 
4550 	ent = sctx->isc_vendor_info;
4551 	while (ent->pvi_vendor_id != 0) {
4552 		if (pci_vendor_id != ent->pvi_vendor_id) {
4553 			ent++;
4554 			continue;
4555 		}
4556 		if ((pci_device_id == ent->pvi_device_id) &&
4557 		    ((pci_subvendor_id == ent->pvi_subvendor_id) ||
4558 		     (ent->pvi_subvendor_id == 0)) &&
4559 		    ((pci_subdevice_id == ent->pvi_subdevice_id) ||
4560 		     (ent->pvi_subdevice_id == 0)) &&
4561 		    ((pci_rev_id == ent->pvi_rev_id) ||
4562 		     (ent->pvi_rev_id == 0))) {
4563 			device_set_desc_copy(dev, ent->pvi_name);
4564 			/* this needs to be changed to zero if the bus probing code
4565 			 * ever stops re-probing on best match because the sctx
4566 			 * may have its values over written by register calls
4567 			 * in subsequent probes
4568 			 */
4569 			return (BUS_PROBE_DEFAULT);
4570 		}
4571 		ent++;
4572 	}
4573 	return (ENXIO);
4574 }
4575 
4576 int
4577 iflib_device_probe_vendor(device_t dev)
4578 {
4579 	int probe;
4580 
4581 	probe = iflib_device_probe(dev);
4582 	if (probe == BUS_PROBE_DEFAULT)
4583 		return (BUS_PROBE_VENDOR);
4584 	else
4585 		return (probe);
4586 }
4587 
4588 static void
4589 iflib_reset_qvalues(if_ctx_t ctx)
4590 {
4591 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4592 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4593 	device_t dev = ctx->ifc_dev;
4594 	int i;
4595 
4596 	if (ctx->ifc_sysctl_ntxqs != 0)
4597 		scctx->isc_ntxqsets = ctx->ifc_sysctl_ntxqs;
4598 	if (ctx->ifc_sysctl_nrxqs != 0)
4599 		scctx->isc_nrxqsets = ctx->ifc_sysctl_nrxqs;
4600 
4601 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4602 		if (ctx->ifc_sysctl_ntxds[i] != 0)
4603 			scctx->isc_ntxd[i] = ctx->ifc_sysctl_ntxds[i];
4604 		else
4605 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4606 	}
4607 
4608 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4609 		if (ctx->ifc_sysctl_nrxds[i] != 0)
4610 			scctx->isc_nrxd[i] = ctx->ifc_sysctl_nrxds[i];
4611 		else
4612 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4613 	}
4614 
4615 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4616 		if (scctx->isc_nrxd[i] < sctx->isc_nrxd_min[i]) {
4617 			device_printf(dev, "nrxd%d: %d less than nrxd_min %d - resetting to min\n",
4618 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_min[i]);
4619 			scctx->isc_nrxd[i] = sctx->isc_nrxd_min[i];
4620 		}
4621 		if (scctx->isc_nrxd[i] > sctx->isc_nrxd_max[i]) {
4622 			device_printf(dev, "nrxd%d: %d greater than nrxd_max %d - resetting to max\n",
4623 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_max[i]);
4624 			scctx->isc_nrxd[i] = sctx->isc_nrxd_max[i];
4625 		}
4626 		if (!powerof2(scctx->isc_nrxd[i])) {
4627 			device_printf(dev, "nrxd%d: %d is not a power of 2 - using default value of %d\n",
4628 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_default[i]);
4629 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4630 		}
4631 	}
4632 
4633 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4634 		if (scctx->isc_ntxd[i] < sctx->isc_ntxd_min[i]) {
4635 			device_printf(dev, "ntxd%d: %d less than ntxd_min %d - resetting to min\n",
4636 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_min[i]);
4637 			scctx->isc_ntxd[i] = sctx->isc_ntxd_min[i];
4638 		}
4639 		if (scctx->isc_ntxd[i] > sctx->isc_ntxd_max[i]) {
4640 			device_printf(dev, "ntxd%d: %d greater than ntxd_max %d - resetting to max\n",
4641 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_max[i]);
4642 			scctx->isc_ntxd[i] = sctx->isc_ntxd_max[i];
4643 		}
4644 		if (!powerof2(scctx->isc_ntxd[i])) {
4645 			device_printf(dev, "ntxd%d: %d is not a power of 2 - using default value of %d\n",
4646 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_default[i]);
4647 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4648 		}
4649 	}
4650 }
4651 
4652 static void
4653 iflib_add_pfil(if_ctx_t ctx)
4654 {
4655 	struct pfil_head *pfil;
4656 	struct pfil_head_args pa;
4657 	iflib_rxq_t rxq;
4658 	int i;
4659 
4660 	pa.pa_version = PFIL_VERSION;
4661 	pa.pa_flags = PFIL_IN;
4662 	pa.pa_type = PFIL_TYPE_ETHERNET;
4663 	pa.pa_headname = ctx->ifc_ifp->if_xname;
4664 	pfil = pfil_head_register(&pa);
4665 
4666 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
4667 		rxq->pfil = pfil;
4668 	}
4669 }
4670 
4671 static void
4672 iflib_rem_pfil(if_ctx_t ctx)
4673 {
4674 	struct pfil_head *pfil;
4675 	iflib_rxq_t rxq;
4676 	int i;
4677 
4678 	rxq = ctx->ifc_rxqs;
4679 	pfil = rxq->pfil;
4680 	for (i = 0; i < NRXQSETS(ctx); i++, rxq++) {
4681 		rxq->pfil = NULL;
4682 	}
4683 	pfil_head_unregister(pfil);
4684 }
4685 
4686 
4687 /*
4688  * Advance forward by n members of the cpuset ctx->ifc_cpus starting from
4689  * cpuid and wrapping as necessary.
4690  */
4691 static unsigned int
4692 cpuid_advance(if_ctx_t ctx, unsigned int cpuid, unsigned int n)
4693 {
4694 	unsigned int first_valid;
4695 	unsigned int last_valid;
4696 
4697 	/* cpuid should always be in the valid set */
4698 	MPASS(CPU_ISSET(cpuid, &ctx->ifc_cpus));
4699 
4700 	/* valid set should never be empty */
4701 	MPASS(!CPU_EMPTY(&ctx->ifc_cpus));
4702 
4703 	first_valid = CPU_FFS(&ctx->ifc_cpus) - 1;
4704 	last_valid = CPU_FLS(&ctx->ifc_cpus) - 1;
4705 	n = n % CPU_COUNT(&ctx->ifc_cpus);
4706 	while (n > 0) {
4707 		do {
4708 			cpuid++;
4709 			if (cpuid > last_valid)
4710 				cpuid = first_valid;
4711 		} while (!CPU_ISSET(cpuid, &ctx->ifc_cpus));
4712 		n--;
4713 	}
4714 
4715 	return (cpuid);
4716 }
4717 
4718 #if defined(SMP) && defined(SCHED_ULE)
4719 extern struct cpu_group *cpu_top;              /* CPU topology */
4720 
4721 static int
4722 find_child_with_core(int cpu, struct cpu_group *grp)
4723 {
4724 	int i;
4725 
4726 	if (grp->cg_children == 0)
4727 		return -1;
4728 
4729 	MPASS(grp->cg_child);
4730 	for (i = 0; i < grp->cg_children; i++) {
4731 		if (CPU_ISSET(cpu, &grp->cg_child[i].cg_mask))
4732 			return i;
4733 	}
4734 
4735 	return -1;
4736 }
4737 
4738 
4739 /*
4740  * Find an L2 neighbor of the given CPU or return -1 if none found.  This
4741  * does not distinguish among multiple L2 neighbors if the given CPU has
4742  * more than one (it will always return the same result in that case).
4743  */
4744 static int
4745 find_l2_neighbor(int cpu)
4746 {
4747 	struct cpu_group *grp;
4748 	int i;
4749 
4750 	grp = cpu_top;
4751 	if (grp == NULL)
4752 		return -1;
4753 
4754 	/*
4755 	 * Find the smallest CPU group that contains the given core.
4756 	 */
4757 	i = 0;
4758 	while ((i = find_child_with_core(cpu, grp)) != -1) {
4759 		/*
4760 		 * If the smallest group containing the given CPU has less
4761 		 * than two members, we conclude the given CPU has no
4762 		 * L2 neighbor.
4763 		 */
4764 		if (grp->cg_child[i].cg_count <= 1)
4765 			return (-1);
4766 		grp = &grp->cg_child[i];
4767 	}
4768 
4769 	/* Must share L2. */
4770 	if (grp->cg_level > CG_SHARE_L2 || grp->cg_level == CG_SHARE_NONE)
4771 		return -1;
4772 
4773 	/*
4774 	 * Select the first member of the set that isn't the reference
4775 	 * CPU, which at this point is guaranteed to exist.
4776 	 */
4777 	for (i = 0; i < CPU_SETSIZE; i++) {
4778 		if (CPU_ISSET(i, &grp->cg_mask) && i != cpu)
4779 			return (i);
4780 	}
4781 
4782 	/* Should never be reached */
4783 	return (-1);
4784 }
4785 
4786 #else
4787 static int
4788 find_l2_neighbor(int cpu)
4789 {
4790 
4791 	return (-1);
4792 }
4793 #endif
4794 
4795 /*
4796  * CPU mapping behaviors
4797  * ---------------------
4798  * 'separate txrx' refers to the separate_txrx sysctl
4799  * 'use logical' refers to the use_logical_cores sysctl
4800  * 'INTR CPUS' indicates whether bus_get_cpus(INTR_CPUS) succeeded
4801  *
4802  *  separate     use     INTR
4803  *    txrx     logical   CPUS   result
4804  * ---------- --------- ------ ------------------------------------------------
4805  *     -          -       X     RX and TX queues mapped to consecutive physical
4806  *                              cores with RX/TX pairs on same core and excess
4807  *                              of either following
4808  *     -          X       X     RX and TX queues mapped to consecutive cores
4809  *                              of any type with RX/TX pairs on same core and
4810  *                              excess of either following
4811  *     X          -       X     RX and TX queues mapped to consecutive physical
4812  *                              cores; all RX then all TX
4813  *     X          X       X     RX queues mapped to consecutive physical cores
4814  *                              first, then TX queues mapped to L2 neighbor of
4815  *                              the corresponding RX queue if one exists,
4816  *                              otherwise to consecutive physical cores
4817  *     -         n/a      -     RX and TX queues mapped to consecutive cores of
4818  *                              any type with RX/TX pairs on same core and excess
4819  *                              of either following
4820  *     X         n/a      -     RX and TX queues mapped to consecutive cores of
4821  *                              any type; all RX then all TX
4822  */
4823 static unsigned int
4824 get_cpuid_for_queue(if_ctx_t ctx, unsigned int base_cpuid, unsigned int qid,
4825     bool is_tx)
4826 {
4827 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4828 	unsigned int core_index;
4829 
4830 	if (ctx->ifc_sysctl_separate_txrx) {
4831 		/*
4832 		 * When using separate CPUs for TX and RX, the assignment
4833 		 * will always be of a consecutive CPU out of the set of
4834 		 * context CPUs, except for the specific case where the
4835 		 * context CPUs are phsyical cores, the use of logical cores
4836 		 * has been enabled, the assignment is for TX, the TX qid
4837 		 * corresponds to an RX qid, and the CPU assigned to the
4838 		 * corresponding RX queue has an L2 neighbor.
4839 		 */
4840 		if (ctx->ifc_sysctl_use_logical_cores &&
4841 		    ctx->ifc_cpus_are_physical_cores &&
4842 		    is_tx && qid < scctx->isc_nrxqsets) {
4843 			int l2_neighbor;
4844 			unsigned int rx_cpuid;
4845 
4846 			rx_cpuid = cpuid_advance(ctx, base_cpuid, qid);
4847 			l2_neighbor = find_l2_neighbor(rx_cpuid);
4848 			if (l2_neighbor != -1) {
4849 				return (l2_neighbor);
4850 			}
4851 			/*
4852 			 * ... else fall through to the normal
4853 			 * consecutive-after-RX assignment scheme.
4854 			 *
4855 			 * Note that we are assuming that all RX queue CPUs
4856 			 * have an L2 neighbor, or all do not.  If a mixed
4857 			 * scenario is possible, we will have to keep track
4858 			 * separately of how many queues prior to this one
4859 			 * were not able to be assigned to an L2 neighbor.
4860 			 */
4861 		}
4862 		if (is_tx)
4863 			core_index = scctx->isc_nrxqsets + qid;
4864 		else
4865 			core_index = qid;
4866 	} else {
4867 		core_index = qid;
4868 	}
4869 
4870 	return (cpuid_advance(ctx, base_cpuid, core_index));
4871 }
4872 
4873 static uint16_t
4874 get_ctx_core_offset(if_ctx_t ctx)
4875 {
4876 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4877 	struct cpu_offset *op;
4878 	cpuset_t assigned_cpus;
4879 	unsigned int cores_consumed;
4880 	unsigned int base_cpuid = ctx->ifc_sysctl_core_offset;
4881 	unsigned int first_valid;
4882 	unsigned int last_valid;
4883 	unsigned int i;
4884 
4885 	first_valid = CPU_FFS(&ctx->ifc_cpus) - 1;
4886 	last_valid = CPU_FLS(&ctx->ifc_cpus) - 1;
4887 
4888 	if (base_cpuid != CORE_OFFSET_UNSPECIFIED) {
4889 		/*
4890 		 * Align the user-chosen base CPU ID to the next valid CPU
4891 		 * for this device.  If the chosen base CPU ID is smaller
4892 		 * than the first valid CPU or larger than the last valid
4893 		 * CPU, we assume the user does not know what the valid
4894 		 * range is for this device and is thinking in terms of a
4895 		 * zero-based reference frame, and so we shift the given
4896 		 * value into the valid range (and wrap accordingly) so the
4897 		 * intent is translated to the proper frame of reference.
4898 		 * If the base CPU ID is within the valid first/last, but
4899 		 * does not correspond to a valid CPU, it is advanced to the
4900 		 * next valid CPU (wrapping if necessary).
4901 		 */
4902 		if (base_cpuid < first_valid || base_cpuid > last_valid) {
4903 			/* shift from zero-based to first_valid-based */
4904 			base_cpuid += first_valid;
4905 			/* wrap to range [first_valid, last_valid] */
4906 			base_cpuid = (base_cpuid - first_valid) %
4907 			    (last_valid - first_valid + 1);
4908 		}
4909 		if (!CPU_ISSET(base_cpuid, &ctx->ifc_cpus)) {
4910 			/*
4911 			 * base_cpuid is in [first_valid, last_valid], but
4912 			 * not a member of the valid set.  In this case,
4913 			 * there will always be a member of the valid set
4914 			 * with a CPU ID that is greater than base_cpuid,
4915 			 * and we simply advance to it.
4916 			 */
4917 			while (!CPU_ISSET(base_cpuid, &ctx->ifc_cpus))
4918 				base_cpuid++;
4919 		}
4920 		return (base_cpuid);
4921 	}
4922 
4923 	/*
4924 	 * Determine how many cores will be consumed by performing the CPU
4925 	 * assignments and counting how many of the assigned CPUs correspond
4926 	 * to CPUs in the set of context CPUs.  This is done using the CPU
4927 	 * ID first_valid as the base CPU ID, as the base CPU must be within
4928 	 * the set of context CPUs.
4929 	 *
4930 	 * Note not all assigned CPUs will be in the set of context CPUs
4931 	 * when separate CPUs are being allocated to TX and RX queues,
4932 	 * assignment to logical cores has been enabled, the set of context
4933 	 * CPUs contains only physical CPUs, and TX queues are mapped to L2
4934 	 * neighbors of CPUs that RX queues have been mapped to - in this
4935 	 * case we do only want to count how many CPUs in the set of context
4936 	 * CPUs have been consumed, as that determines the next CPU in that
4937 	 * set to start allocating at for the next device for which
4938 	 * core_offset is not set.
4939 	 */
4940 	CPU_ZERO(&assigned_cpus);
4941 	for (i = 0; i < scctx->isc_ntxqsets; i++)
4942 		CPU_SET(get_cpuid_for_queue(ctx, first_valid, i, true),
4943 		    &assigned_cpus);
4944 	for (i = 0; i < scctx->isc_nrxqsets; i++)
4945 		CPU_SET(get_cpuid_for_queue(ctx, first_valid, i, false),
4946 		    &assigned_cpus);
4947 	CPU_AND(&assigned_cpus, &assigned_cpus, &ctx->ifc_cpus);
4948 	cores_consumed = CPU_COUNT(&assigned_cpus);
4949 
4950 	mtx_lock(&cpu_offset_mtx);
4951 	SLIST_FOREACH(op, &cpu_offsets, entries) {
4952 		if (CPU_CMP(&ctx->ifc_cpus, &op->set) == 0) {
4953 			base_cpuid = op->next_cpuid;
4954 			op->next_cpuid = cpuid_advance(ctx, op->next_cpuid,
4955 			    cores_consumed);
4956 			MPASS(op->refcount < UINT_MAX);
4957 			op->refcount++;
4958 			break;
4959 		}
4960 	}
4961 	if (base_cpuid == CORE_OFFSET_UNSPECIFIED) {
4962 		base_cpuid = first_valid;
4963 		op = malloc(sizeof(struct cpu_offset), M_IFLIB,
4964 		    M_NOWAIT | M_ZERO);
4965 		if (op == NULL) {
4966 			device_printf(ctx->ifc_dev,
4967 			    "allocation for cpu offset failed.\n");
4968 		} else {
4969 			op->next_cpuid = cpuid_advance(ctx, base_cpuid,
4970 			    cores_consumed);
4971 			op->refcount = 1;
4972 			CPU_COPY(&ctx->ifc_cpus, &op->set);
4973 			SLIST_INSERT_HEAD(&cpu_offsets, op, entries);
4974 		}
4975 	}
4976 	mtx_unlock(&cpu_offset_mtx);
4977 
4978 	return (base_cpuid);
4979 }
4980 
4981 static void
4982 unref_ctx_core_offset(if_ctx_t ctx)
4983 {
4984 	struct cpu_offset *op, *top;
4985 
4986 	mtx_lock(&cpu_offset_mtx);
4987 	SLIST_FOREACH_SAFE(op, &cpu_offsets, entries, top) {
4988 		if (CPU_CMP(&ctx->ifc_cpus, &op->set) == 0) {
4989 			MPASS(op->refcount > 0);
4990 			op->refcount--;
4991 			if (op->refcount == 0) {
4992 				SLIST_REMOVE(&cpu_offsets, op, cpu_offset, entries);
4993 				free(op, M_IFLIB);
4994 			}
4995 			break;
4996 		}
4997 	}
4998 	mtx_unlock(&cpu_offset_mtx);
4999 }
5000 
5001 int
5002 iflib_device_register(device_t dev, void *sc, if_shared_ctx_t sctx, if_ctx_t *ctxp)
5003 {
5004 	if_ctx_t ctx;
5005 	if_t ifp;
5006 	if_softc_ctx_t scctx;
5007 	kobjop_desc_t kobj_desc;
5008 	kobj_method_t *kobj_method;
5009 	int err, msix, rid;
5010 	int num_txd, num_rxd;
5011 
5012 	ctx = malloc(sizeof(* ctx), M_IFLIB, M_WAITOK|M_ZERO);
5013 
5014 	if (sc == NULL) {
5015 		sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
5016 		device_set_softc(dev, ctx);
5017 		ctx->ifc_flags |= IFC_SC_ALLOCATED;
5018 	}
5019 
5020 	ctx->ifc_sctx = sctx;
5021 	ctx->ifc_dev = dev;
5022 	ctx->ifc_softc = sc;
5023 
5024 	if ((err = iflib_register(ctx)) != 0) {
5025 		device_printf(dev, "iflib_register failed %d\n", err);
5026 		goto fail_ctx_free;
5027 	}
5028 	iflib_add_device_sysctl_pre(ctx);
5029 
5030 	scctx = &ctx->ifc_softc_ctx;
5031 	ifp = ctx->ifc_ifp;
5032 
5033 	iflib_reset_qvalues(ctx);
5034 	CTX_LOCK(ctx);
5035 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
5036 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
5037 		goto fail_unlock;
5038 	}
5039 	_iflib_pre_assert(scctx);
5040 	ctx->ifc_txrx = *scctx->isc_txrx;
5041 
5042 	if (sctx->isc_flags & IFLIB_DRIVER_MEDIA)
5043 		ctx->ifc_mediap = scctx->isc_media;
5044 
5045 #ifdef INVARIANTS
5046 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
5047 		MPASS(scctx->isc_tx_csum_flags);
5048 #endif
5049 
5050 	if_setcapabilities(ifp,
5051 	    scctx->isc_capabilities | IFCAP_HWSTATS | IFCAP_MEXTPG);
5052 	if_setcapenable(ifp,
5053 	    scctx->isc_capenable | IFCAP_HWSTATS | IFCAP_MEXTPG);
5054 
5055 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
5056 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
5057 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
5058 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
5059 
5060 	num_txd = iflib_num_tx_descs(ctx);
5061 	num_rxd = iflib_num_rx_descs(ctx);
5062 
5063 	/* XXX change for per-queue sizes */
5064 	device_printf(dev, "Using %d TX descriptors and %d RX descriptors\n",
5065 	    num_txd, num_rxd);
5066 
5067 	if (scctx->isc_tx_nsegments > num_txd / MAX_SINGLE_PACKET_FRACTION)
5068 		scctx->isc_tx_nsegments = max(1, num_txd /
5069 		    MAX_SINGLE_PACKET_FRACTION);
5070 	if (scctx->isc_tx_tso_segments_max > num_txd /
5071 	    MAX_SINGLE_PACKET_FRACTION)
5072 		scctx->isc_tx_tso_segments_max = max(1,
5073 		    num_txd / MAX_SINGLE_PACKET_FRACTION);
5074 
5075 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
5076 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
5077 		/*
5078 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
5079 		 * but some MACs do.
5080 		 */
5081 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
5082 		    IP_MAXPACKET));
5083 		/*
5084 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
5085 		 * into account.  In the worst case, each of these calls will
5086 		 * add another mbuf and, thus, the requirement for another DMA
5087 		 * segment.  So for best performance, it doesn't make sense to
5088 		 * advertize a maximum of TSO segments that typically will
5089 		 * require defragmentation in iflib_encap().
5090 		 */
5091 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
5092 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
5093 	}
5094 	if (scctx->isc_rss_table_size == 0)
5095 		scctx->isc_rss_table_size = 64;
5096 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
5097 
5098 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
5099 	/* XXX format name */
5100 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx,
5101 	    NULL, NULL, "admin");
5102 
5103 	/* Set up cpu set.  If it fails, use the set of all CPUs. */
5104 	if (bus_get_cpus(dev, INTR_CPUS, sizeof(ctx->ifc_cpus), &ctx->ifc_cpus) != 0) {
5105 		device_printf(dev, "Unable to fetch CPU list\n");
5106 		CPU_COPY(&all_cpus, &ctx->ifc_cpus);
5107 		ctx->ifc_cpus_are_physical_cores = false;
5108 	} else
5109 		ctx->ifc_cpus_are_physical_cores = true;
5110 	MPASS(CPU_COUNT(&ctx->ifc_cpus) > 0);
5111 
5112 	/*
5113 	** Now set up MSI or MSI-X, should return us the number of supported
5114 	** vectors (will be 1 for a legacy interrupt and MSI).
5115 	*/
5116 	if (sctx->isc_flags & IFLIB_SKIP_MSIX) {
5117 		msix = scctx->isc_vectors;
5118 	} else if (scctx->isc_msix_bar != 0)
5119 	       /*
5120 		* The simple fact that isc_msix_bar is not 0 does not mean we
5121 		* we have a good value there that is known to work.
5122 		*/
5123 		msix = iflib_msix_init(ctx);
5124 	else {
5125 		scctx->isc_vectors = 1;
5126 		scctx->isc_ntxqsets = 1;
5127 		scctx->isc_nrxqsets = 1;
5128 		scctx->isc_intr = IFLIB_INTR_LEGACY;
5129 		msix = 0;
5130 	}
5131 	/* Get memory for the station queues */
5132 	if ((err = iflib_queues_alloc(ctx))) {
5133 		device_printf(dev, "Unable to allocate queue memory\n");
5134 		goto fail_intr_free;
5135 	}
5136 
5137 	if ((err = iflib_qset_structures_setup(ctx)))
5138 		goto fail_queues;
5139 
5140 	/*
5141 	 * Now that we know how many queues there are, get the core offset.
5142 	 */
5143 	ctx->ifc_sysctl_core_offset = get_ctx_core_offset(ctx);
5144 
5145 	if (msix > 1) {
5146 		/*
5147 		 * When using MSI-X, ensure that ifdi_{r,t}x_queue_intr_enable
5148 		 * aren't the default NULL implementation.
5149 		 */
5150 		kobj_desc = &ifdi_rx_queue_intr_enable_desc;
5151 		kobj_method = kobj_lookup_method(((kobj_t)ctx)->ops->cls, NULL,
5152 		    kobj_desc);
5153 		if (kobj_method == &kobj_desc->deflt) {
5154 			device_printf(dev,
5155 			    "MSI-X requires ifdi_rx_queue_intr_enable method");
5156 			err = EOPNOTSUPP;
5157 			goto fail_queues;
5158 		}
5159 		kobj_desc = &ifdi_tx_queue_intr_enable_desc;
5160 		kobj_method = kobj_lookup_method(((kobj_t)ctx)->ops->cls, NULL,
5161 		    kobj_desc);
5162 		if (kobj_method == &kobj_desc->deflt) {
5163 			device_printf(dev,
5164 			    "MSI-X requires ifdi_tx_queue_intr_enable method");
5165 			err = EOPNOTSUPP;
5166 			goto fail_queues;
5167 		}
5168 
5169 		/*
5170 		 * Assign the MSI-X vectors.
5171 		 * Note that the default NULL ifdi_msix_intr_assign method will
5172 		 * fail here, too.
5173 		 */
5174 		err = IFDI_MSIX_INTR_ASSIGN(ctx, msix);
5175 		if (err != 0) {
5176 			device_printf(dev, "IFDI_MSIX_INTR_ASSIGN failed %d\n",
5177 			    err);
5178 			goto fail_queues;
5179 		}
5180 	} else if (scctx->isc_intr != IFLIB_INTR_MSIX) {
5181 		rid = 0;
5182 		if (scctx->isc_intr == IFLIB_INTR_MSI) {
5183 			MPASS(msix == 1);
5184 			rid = 1;
5185 		}
5186 		if ((err = iflib_legacy_setup(ctx, ctx->isc_legacy_intr, ctx->ifc_softc, &rid, "irq0")) != 0) {
5187 			device_printf(dev, "iflib_legacy_setup failed %d\n", err);
5188 			goto fail_queues;
5189 		}
5190 	} else {
5191 		device_printf(dev,
5192 		    "Cannot use iflib with only 1 MSI-X interrupt!\n");
5193 		err = ENODEV;
5194 		goto fail_queues;
5195 	}
5196 
5197 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
5198 
5199 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
5200 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
5201 		goto fail_detach;
5202 	}
5203 
5204 	/*
5205 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
5206 	 * This must appear after the call to ether_ifattach() because
5207 	 * ether_ifattach() sets if_hdrlen to the default value.
5208 	 */
5209 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
5210 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
5211 
5212 	if ((err = iflib_netmap_attach(ctx))) {
5213 		device_printf(ctx->ifc_dev, "netmap attach failed: %d\n", err);
5214 		goto fail_detach;
5215 	}
5216 	*ctxp = ctx;
5217 
5218 	DEBUGNET_SET(ctx->ifc_ifp, iflib);
5219 
5220 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
5221 	iflib_add_device_sysctl_post(ctx);
5222 	iflib_add_pfil(ctx);
5223 	ctx->ifc_flags |= IFC_INIT_DONE;
5224 	CTX_UNLOCK(ctx);
5225 
5226 	return (0);
5227 
5228 fail_detach:
5229 	ether_ifdetach(ctx->ifc_ifp);
5230 fail_queues:
5231 	iflib_tqg_detach(ctx);
5232 	iflib_tx_structures_free(ctx);
5233 	iflib_rx_structures_free(ctx);
5234 	IFDI_DETACH(ctx);
5235 	IFDI_QUEUES_FREE(ctx);
5236 fail_intr_free:
5237 	iflib_free_intr_mem(ctx);
5238 fail_unlock:
5239 	CTX_UNLOCK(ctx);
5240 	iflib_deregister(ctx);
5241 fail_ctx_free:
5242 	device_set_softc(ctx->ifc_dev, NULL);
5243         if (ctx->ifc_flags & IFC_SC_ALLOCATED)
5244                 free(ctx->ifc_softc, M_IFLIB);
5245         free(ctx, M_IFLIB);
5246 	return (err);
5247 }
5248 
5249 int
5250 iflib_pseudo_register(device_t dev, if_shared_ctx_t sctx, if_ctx_t *ctxp,
5251 					  struct iflib_cloneattach_ctx *clctx)
5252 {
5253 	int num_txd, num_rxd;
5254 	int err;
5255 	if_ctx_t ctx;
5256 	if_t ifp;
5257 	if_softc_ctx_t scctx;
5258 	int i;
5259 	void *sc;
5260 
5261 	ctx = malloc(sizeof(*ctx), M_IFLIB, M_WAITOK|M_ZERO);
5262 	sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
5263 	ctx->ifc_flags |= IFC_SC_ALLOCATED;
5264 	if (sctx->isc_flags & (IFLIB_PSEUDO|IFLIB_VIRTUAL))
5265 		ctx->ifc_flags |= IFC_PSEUDO;
5266 
5267 	ctx->ifc_sctx = sctx;
5268 	ctx->ifc_softc = sc;
5269 	ctx->ifc_dev = dev;
5270 
5271 	if ((err = iflib_register(ctx)) != 0) {
5272 		device_printf(dev, "%s: iflib_register failed %d\n", __func__, err);
5273 		goto fail_ctx_free;
5274 	}
5275 	iflib_add_device_sysctl_pre(ctx);
5276 
5277 	scctx = &ctx->ifc_softc_ctx;
5278 	ifp = ctx->ifc_ifp;
5279 
5280 	iflib_reset_qvalues(ctx);
5281 	CTX_LOCK(ctx);
5282 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
5283 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
5284 		goto fail_unlock;
5285 	}
5286 	if (sctx->isc_flags & IFLIB_GEN_MAC)
5287 		ether_gen_addr(ifp, &ctx->ifc_mac);
5288 	if ((err = IFDI_CLONEATTACH(ctx, clctx->cc_ifc, clctx->cc_name,
5289 								clctx->cc_params)) != 0) {
5290 		device_printf(dev, "IFDI_CLONEATTACH failed %d\n", err);
5291 		goto fail_unlock;
5292 	}
5293 #ifdef INVARIANTS
5294 	if (scctx->isc_capabilities & IFCAP_TXCSUM)
5295 		MPASS(scctx->isc_tx_csum_flags);
5296 #endif
5297 
5298 	if_setcapabilities(ifp, scctx->isc_capabilities | IFCAP_HWSTATS | IFCAP_LINKSTATE);
5299 	if_setcapenable(ifp, scctx->isc_capenable | IFCAP_HWSTATS | IFCAP_LINKSTATE);
5300 
5301 	ifp->if_flags |= IFF_NOGROUP;
5302 	if (sctx->isc_flags & IFLIB_PSEUDO) {
5303 		ifmedia_add(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO, 0, NULL);
5304 		ifmedia_set(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO);
5305 		if (sctx->isc_flags & IFLIB_PSEUDO_ETHER) {
5306 			ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
5307 		} else {
5308 			if_attach(ctx->ifc_ifp);
5309 			bpfattach(ctx->ifc_ifp, DLT_NULL, sizeof(u_int32_t));
5310 		}
5311 
5312 		if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
5313 			device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
5314 			goto fail_detach;
5315 		}
5316 		*ctxp = ctx;
5317 
5318 		/*
5319 		 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
5320 		 * This must appear after the call to ether_ifattach() because
5321 		 * ether_ifattach() sets if_hdrlen to the default value.
5322 		 */
5323 		if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
5324 			if_setifheaderlen(ifp,
5325 			    sizeof(struct ether_vlan_header));
5326 
5327 		if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
5328 		iflib_add_device_sysctl_post(ctx);
5329 		ctx->ifc_flags |= IFC_INIT_DONE;
5330 		CTX_UNLOCK(ctx);
5331 		return (0);
5332 	}
5333 	ifmedia_add(ctx->ifc_mediap, IFM_ETHER | IFM_1000_T | IFM_FDX, 0, NULL);
5334 	ifmedia_add(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO, 0, NULL);
5335 	ifmedia_set(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO);
5336 
5337 	_iflib_pre_assert(scctx);
5338 	ctx->ifc_txrx = *scctx->isc_txrx;
5339 
5340 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
5341 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
5342 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
5343 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
5344 
5345 	num_txd = iflib_num_tx_descs(ctx);
5346 	num_rxd = iflib_num_rx_descs(ctx);
5347 
5348 	/* XXX change for per-queue sizes */
5349 	device_printf(dev, "Using %d TX descriptors and %d RX descriptors\n",
5350 	    num_txd, num_rxd);
5351 
5352 	if (scctx->isc_tx_nsegments > num_txd / MAX_SINGLE_PACKET_FRACTION)
5353 		scctx->isc_tx_nsegments = max(1, num_txd /
5354 		    MAX_SINGLE_PACKET_FRACTION);
5355 	if (scctx->isc_tx_tso_segments_max > num_txd /
5356 	    MAX_SINGLE_PACKET_FRACTION)
5357 		scctx->isc_tx_tso_segments_max = max(1,
5358 		    num_txd / MAX_SINGLE_PACKET_FRACTION);
5359 
5360 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
5361 	if (if_getcapabilities(ifp) & IFCAP_TSO) {
5362 		/*
5363 		 * The stack can't handle a TSO size larger than IP_MAXPACKET,
5364 		 * but some MACs do.
5365 		 */
5366 		if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max,
5367 		    IP_MAXPACKET));
5368 		/*
5369 		 * Take maximum number of m_pullup(9)'s in iflib_parse_header()
5370 		 * into account.  In the worst case, each of these calls will
5371 		 * add another mbuf and, thus, the requirement for another DMA
5372 		 * segment.  So for best performance, it doesn't make sense to
5373 		 * advertize a maximum of TSO segments that typically will
5374 		 * require defragmentation in iflib_encap().
5375 		 */
5376 		if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3);
5377 		if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max);
5378 	}
5379 	if (scctx->isc_rss_table_size == 0)
5380 		scctx->isc_rss_table_size = 64;
5381 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
5382 
5383 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
5384 	/* XXX format name */
5385 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx,
5386 	    NULL, NULL, "admin");
5387 
5388 	/* XXX --- can support > 1 -- but keep it simple for now */
5389 	scctx->isc_intr = IFLIB_INTR_LEGACY;
5390 
5391 	/* Get memory for the station queues */
5392 	if ((err = iflib_queues_alloc(ctx))) {
5393 		device_printf(dev, "Unable to allocate queue memory\n");
5394 		goto fail_iflib_detach;
5395 	}
5396 
5397 	if ((err = iflib_qset_structures_setup(ctx))) {
5398 		device_printf(dev, "qset structure setup failed %d\n", err);
5399 		goto fail_queues;
5400 	}
5401 
5402 	/*
5403 	 * XXX What if anything do we want to do about interrupts?
5404 	 */
5405 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet);
5406 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
5407 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
5408 		goto fail_detach;
5409 	}
5410 
5411 	/*
5412 	 * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported.
5413 	 * This must appear after the call to ether_ifattach() because
5414 	 * ether_ifattach() sets if_hdrlen to the default value.
5415 	 */
5416 	if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU)
5417 		if_setifheaderlen(ifp, sizeof(struct ether_vlan_header));
5418 
5419 	/* XXX handle more than one queue */
5420 	for (i = 0; i < scctx->isc_nrxqsets; i++)
5421 		IFDI_RX_CLSET(ctx, 0, i, ctx->ifc_rxqs[i].ifr_fl[0].ifl_sds.ifsd_cl);
5422 
5423 	*ctxp = ctx;
5424 
5425 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
5426 	iflib_add_device_sysctl_post(ctx);
5427 	ctx->ifc_flags |= IFC_INIT_DONE;
5428 	CTX_UNLOCK(ctx);
5429 
5430 	return (0);
5431 fail_detach:
5432 	ether_ifdetach(ctx->ifc_ifp);
5433 fail_queues:
5434 	iflib_tqg_detach(ctx);
5435 	iflib_tx_structures_free(ctx);
5436 	iflib_rx_structures_free(ctx);
5437 fail_iflib_detach:
5438 	IFDI_DETACH(ctx);
5439 	IFDI_QUEUES_FREE(ctx);
5440 fail_unlock:
5441 	CTX_UNLOCK(ctx);
5442 	iflib_deregister(ctx);
5443 fail_ctx_free:
5444 	free(ctx->ifc_softc, M_IFLIB);
5445 	free(ctx, M_IFLIB);
5446 	return (err);
5447 }
5448 
5449 int
5450 iflib_pseudo_deregister(if_ctx_t ctx)
5451 {
5452 	if_t ifp = ctx->ifc_ifp;
5453 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5454 
5455 	/* Unregister VLAN event handlers early */
5456 	iflib_unregister_vlan_handlers(ctx);
5457 
5458 	if ((sctx->isc_flags & IFLIB_PSEUDO)  &&
5459 		(sctx->isc_flags & IFLIB_PSEUDO_ETHER) == 0) {
5460 		bpfdetach(ifp);
5461 		if_detach(ifp);
5462 	} else {
5463 		ether_ifdetach(ifp);
5464 	}
5465 
5466 	iflib_tqg_detach(ctx);
5467 	iflib_tx_structures_free(ctx);
5468 	iflib_rx_structures_free(ctx);
5469 	IFDI_DETACH(ctx);
5470 	IFDI_QUEUES_FREE(ctx);
5471 
5472 	iflib_deregister(ctx);
5473 
5474 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
5475 		free(ctx->ifc_softc, M_IFLIB);
5476 	free(ctx, M_IFLIB);
5477 	return (0);
5478 }
5479 
5480 int
5481 iflib_device_attach(device_t dev)
5482 {
5483 	if_ctx_t ctx;
5484 	if_shared_ctx_t sctx;
5485 
5486 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
5487 		return (ENOTSUP);
5488 
5489 	pci_enable_busmaster(dev);
5490 
5491 	return (iflib_device_register(dev, NULL, sctx, &ctx));
5492 }
5493 
5494 int
5495 iflib_device_deregister(if_ctx_t ctx)
5496 {
5497 	if_t ifp = ctx->ifc_ifp;
5498 	device_t dev = ctx->ifc_dev;
5499 
5500 	/* Make sure VLANS are not using driver */
5501 	if (if_vlantrunkinuse(ifp)) {
5502 		device_printf(dev, "Vlan in use, detach first\n");
5503 		return (EBUSY);
5504 	}
5505 #ifdef PCI_IOV
5506 	if (!CTX_IS_VF(ctx) && pci_iov_detach(dev) != 0) {
5507 		device_printf(dev, "SR-IOV in use; detach first.\n");
5508 		return (EBUSY);
5509 	}
5510 #endif
5511 
5512 	STATE_LOCK(ctx);
5513 	ctx->ifc_flags |= IFC_IN_DETACH;
5514 	STATE_UNLOCK(ctx);
5515 
5516 	/* Unregister VLAN handlers before calling iflib_stop() */
5517 	iflib_unregister_vlan_handlers(ctx);
5518 
5519 	iflib_netmap_detach(ifp);
5520 	ether_ifdetach(ifp);
5521 
5522 	CTX_LOCK(ctx);
5523 	iflib_stop(ctx);
5524 	CTX_UNLOCK(ctx);
5525 
5526 	iflib_rem_pfil(ctx);
5527 	if (ctx->ifc_led_dev != NULL)
5528 		led_destroy(ctx->ifc_led_dev);
5529 
5530 	iflib_tqg_detach(ctx);
5531 	iflib_tx_structures_free(ctx);
5532 	iflib_rx_structures_free(ctx);
5533 
5534 	CTX_LOCK(ctx);
5535 	IFDI_DETACH(ctx);
5536 	IFDI_QUEUES_FREE(ctx);
5537 	CTX_UNLOCK(ctx);
5538 
5539 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
5540 	iflib_free_intr_mem(ctx);
5541 
5542 	bus_generic_detach(dev);
5543 
5544 	iflib_deregister(ctx);
5545 
5546 	device_set_softc(ctx->ifc_dev, NULL);
5547 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
5548 		free(ctx->ifc_softc, M_IFLIB);
5549 	unref_ctx_core_offset(ctx);
5550 	free(ctx, M_IFLIB);
5551 	return (0);
5552 }
5553 
5554 static void
5555 iflib_tqg_detach(if_ctx_t ctx)
5556 {
5557 	iflib_txq_t txq;
5558 	iflib_rxq_t rxq;
5559 	int i;
5560 	struct taskqgroup *tqg;
5561 
5562 	/* XXX drain any dependent tasks */
5563 	tqg = qgroup_if_io_tqg;
5564 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
5565 		callout_drain(&txq->ift_timer);
5566 #ifdef DEV_NETMAP
5567 		callout_drain(&txq->ift_netmap_timer);
5568 #endif /* DEV_NETMAP */
5569 		if (txq->ift_task.gt_uniq != NULL)
5570 			taskqgroup_detach(tqg, &txq->ift_task);
5571 	}
5572 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
5573 		if (rxq->ifr_task.gt_uniq != NULL)
5574 			taskqgroup_detach(tqg, &rxq->ifr_task);
5575 	}
5576 	tqg = qgroup_if_config_tqg;
5577 	if (ctx->ifc_admin_task.gt_uniq != NULL)
5578 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
5579 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
5580 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
5581 }
5582 
5583 static void
5584 iflib_free_intr_mem(if_ctx_t ctx)
5585 {
5586 
5587 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_MSIX) {
5588 		iflib_irq_free(ctx, &ctx->ifc_legacy_irq);
5589 	}
5590 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_LEGACY) {
5591 		pci_release_msi(ctx->ifc_dev);
5592 	}
5593 	if (ctx->ifc_msix_mem != NULL) {
5594 		bus_release_resource(ctx->ifc_dev, SYS_RES_MEMORY,
5595 		    rman_get_rid(ctx->ifc_msix_mem), ctx->ifc_msix_mem);
5596 		ctx->ifc_msix_mem = NULL;
5597 	}
5598 }
5599 
5600 int
5601 iflib_device_detach(device_t dev)
5602 {
5603 	if_ctx_t ctx = device_get_softc(dev);
5604 
5605 	return (iflib_device_deregister(ctx));
5606 }
5607 
5608 int
5609 iflib_device_suspend(device_t dev)
5610 {
5611 	if_ctx_t ctx = device_get_softc(dev);
5612 
5613 	CTX_LOCK(ctx);
5614 	IFDI_SUSPEND(ctx);
5615 	CTX_UNLOCK(ctx);
5616 
5617 	return bus_generic_suspend(dev);
5618 }
5619 int
5620 iflib_device_shutdown(device_t dev)
5621 {
5622 	if_ctx_t ctx = device_get_softc(dev);
5623 
5624 	CTX_LOCK(ctx);
5625 	IFDI_SHUTDOWN(ctx);
5626 	CTX_UNLOCK(ctx);
5627 
5628 	return bus_generic_suspend(dev);
5629 }
5630 
5631 int
5632 iflib_device_resume(device_t dev)
5633 {
5634 	if_ctx_t ctx = device_get_softc(dev);
5635 	iflib_txq_t txq = ctx->ifc_txqs;
5636 
5637 	CTX_LOCK(ctx);
5638 	IFDI_RESUME(ctx);
5639 	iflib_if_init_locked(ctx);
5640 	CTX_UNLOCK(ctx);
5641 	for (int i = 0; i < NTXQSETS(ctx); i++, txq++)
5642 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
5643 
5644 	return (bus_generic_resume(dev));
5645 }
5646 
5647 int
5648 iflib_device_iov_init(device_t dev, uint16_t num_vfs, const nvlist_t *params)
5649 {
5650 	int error;
5651 	if_ctx_t ctx = device_get_softc(dev);
5652 
5653 	CTX_LOCK(ctx);
5654 	error = IFDI_IOV_INIT(ctx, num_vfs, params);
5655 	CTX_UNLOCK(ctx);
5656 
5657 	return (error);
5658 }
5659 
5660 void
5661 iflib_device_iov_uninit(device_t dev)
5662 {
5663 	if_ctx_t ctx = device_get_softc(dev);
5664 
5665 	CTX_LOCK(ctx);
5666 	IFDI_IOV_UNINIT(ctx);
5667 	CTX_UNLOCK(ctx);
5668 }
5669 
5670 int
5671 iflib_device_iov_add_vf(device_t dev, uint16_t vfnum, const nvlist_t *params)
5672 {
5673 	int error;
5674 	if_ctx_t ctx = device_get_softc(dev);
5675 
5676 	CTX_LOCK(ctx);
5677 	error = IFDI_IOV_VF_ADD(ctx, vfnum, params);
5678 	CTX_UNLOCK(ctx);
5679 
5680 	return (error);
5681 }
5682 
5683 /*********************************************************************
5684  *
5685  *  MODULE FUNCTION DEFINITIONS
5686  *
5687  **********************************************************************/
5688 
5689 /*
5690  * - Start a fast taskqueue thread for each core
5691  * - Start a taskqueue for control operations
5692  */
5693 static int
5694 iflib_module_init(void)
5695 {
5696 	iflib_timer_default = hz / 2;
5697 	return (0);
5698 }
5699 
5700 static int
5701 iflib_module_event_handler(module_t mod, int what, void *arg)
5702 {
5703 	int err;
5704 
5705 	switch (what) {
5706 	case MOD_LOAD:
5707 		if ((err = iflib_module_init()) != 0)
5708 			return (err);
5709 		break;
5710 	case MOD_UNLOAD:
5711 		return (EBUSY);
5712 	default:
5713 		return (EOPNOTSUPP);
5714 	}
5715 
5716 	return (0);
5717 }
5718 
5719 /*********************************************************************
5720  *
5721  *  PUBLIC FUNCTION DEFINITIONS
5722  *     ordered as in iflib.h
5723  *
5724  **********************************************************************/
5725 
5726 static void
5727 _iflib_assert(if_shared_ctx_t sctx)
5728 {
5729 	int i;
5730 
5731 	MPASS(sctx->isc_tx_maxsize);
5732 	MPASS(sctx->isc_tx_maxsegsize);
5733 
5734 	MPASS(sctx->isc_rx_maxsize);
5735 	MPASS(sctx->isc_rx_nsegments);
5736 	MPASS(sctx->isc_rx_maxsegsize);
5737 
5738 	MPASS(sctx->isc_nrxqs >= 1 && sctx->isc_nrxqs <= 8);
5739 	for (i = 0; i < sctx->isc_nrxqs; i++) {
5740 		MPASS(sctx->isc_nrxd_min[i]);
5741 		MPASS(powerof2(sctx->isc_nrxd_min[i]));
5742 		MPASS(sctx->isc_nrxd_max[i]);
5743 		MPASS(powerof2(sctx->isc_nrxd_max[i]));
5744 		MPASS(sctx->isc_nrxd_default[i]);
5745 		MPASS(powerof2(sctx->isc_nrxd_default[i]));
5746 	}
5747 
5748 	MPASS(sctx->isc_ntxqs >= 1 && sctx->isc_ntxqs <= 8);
5749 	for (i = 0; i < sctx->isc_ntxqs; i++) {
5750 		MPASS(sctx->isc_ntxd_min[i]);
5751 		MPASS(powerof2(sctx->isc_ntxd_min[i]));
5752 		MPASS(sctx->isc_ntxd_max[i]);
5753 		MPASS(powerof2(sctx->isc_ntxd_max[i]));
5754 		MPASS(sctx->isc_ntxd_default[i]);
5755 		MPASS(powerof2(sctx->isc_ntxd_default[i]));
5756 	}
5757 }
5758 
5759 static void
5760 _iflib_pre_assert(if_softc_ctx_t scctx)
5761 {
5762 
5763 	MPASS(scctx->isc_txrx->ift_txd_encap);
5764 	MPASS(scctx->isc_txrx->ift_txd_flush);
5765 	MPASS(scctx->isc_txrx->ift_txd_credits_update);
5766 	MPASS(scctx->isc_txrx->ift_rxd_available);
5767 	MPASS(scctx->isc_txrx->ift_rxd_pkt_get);
5768 	MPASS(scctx->isc_txrx->ift_rxd_refill);
5769 	MPASS(scctx->isc_txrx->ift_rxd_flush);
5770 }
5771 
5772 static int
5773 iflib_register(if_ctx_t ctx)
5774 {
5775 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5776 	driver_t *driver = sctx->isc_driver;
5777 	device_t dev = ctx->ifc_dev;
5778 	if_t ifp;
5779 	u_char type;
5780 	int iflags;
5781 
5782 	if ((sctx->isc_flags & IFLIB_PSEUDO) == 0)
5783 		_iflib_assert(sctx);
5784 
5785 	CTX_LOCK_INIT(ctx);
5786 	STATE_LOCK_INIT(ctx, device_get_nameunit(ctx->ifc_dev));
5787 	if (sctx->isc_flags & IFLIB_PSEUDO) {
5788 		if (sctx->isc_flags & IFLIB_PSEUDO_ETHER)
5789 			type = IFT_ETHER;
5790 		else
5791 			type = IFT_PPP;
5792 	} else
5793 		type = IFT_ETHER;
5794 	ifp = ctx->ifc_ifp = if_alloc(type);
5795 	if (ifp == NULL) {
5796 		device_printf(dev, "can not allocate ifnet structure\n");
5797 		return (ENOMEM);
5798 	}
5799 
5800 	/*
5801 	 * Initialize our context's device specific methods
5802 	 */
5803 	kobj_init((kobj_t) ctx, (kobj_class_t) driver);
5804 	kobj_class_compile((kobj_class_t) driver);
5805 
5806 	if_initname(ifp, device_get_name(dev), device_get_unit(dev));
5807 	if_setsoftc(ifp, ctx);
5808 	if_setdev(ifp, dev);
5809 	if_setinitfn(ifp, iflib_if_init);
5810 	if_setioctlfn(ifp, iflib_if_ioctl);
5811 #ifdef ALTQ
5812 	if_setstartfn(ifp, iflib_altq_if_start);
5813 	if_settransmitfn(ifp, iflib_altq_if_transmit);
5814 	if_setsendqready(ifp);
5815 #else
5816 	if_settransmitfn(ifp, iflib_if_transmit);
5817 #endif
5818 	if_setqflushfn(ifp, iflib_if_qflush);
5819 	iflags = IFF_MULTICAST | IFF_KNOWSEPOCH;
5820 
5821 	if ((sctx->isc_flags & IFLIB_PSEUDO) &&
5822 		(sctx->isc_flags & IFLIB_PSEUDO_ETHER) == 0)
5823 		iflags |= IFF_POINTOPOINT;
5824 	else
5825 		iflags |= IFF_BROADCAST | IFF_SIMPLEX;
5826 	if_setflags(ifp, iflags);
5827 	ctx->ifc_vlan_attach_event =
5828 		EVENTHANDLER_REGISTER(vlan_config, iflib_vlan_register, ctx,
5829 							  EVENTHANDLER_PRI_FIRST);
5830 	ctx->ifc_vlan_detach_event =
5831 		EVENTHANDLER_REGISTER(vlan_unconfig, iflib_vlan_unregister, ctx,
5832 							  EVENTHANDLER_PRI_FIRST);
5833 
5834 	if ((sctx->isc_flags & IFLIB_DRIVER_MEDIA) == 0) {
5835 		ctx->ifc_mediap = &ctx->ifc_media;
5836 		ifmedia_init(ctx->ifc_mediap, IFM_IMASK,
5837 		    iflib_media_change, iflib_media_status);
5838 	}
5839 	return (0);
5840 }
5841 
5842 static void
5843 iflib_unregister_vlan_handlers(if_ctx_t ctx)
5844 {
5845 	/* Unregister VLAN events */
5846 	if (ctx->ifc_vlan_attach_event != NULL) {
5847 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
5848 		ctx->ifc_vlan_attach_event = NULL;
5849 	}
5850 	if (ctx->ifc_vlan_detach_event != NULL) {
5851 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
5852 		ctx->ifc_vlan_detach_event = NULL;
5853 	}
5854 
5855 }
5856 
5857 static void
5858 iflib_deregister(if_ctx_t ctx)
5859 {
5860 	if_t ifp = ctx->ifc_ifp;
5861 
5862 	/* Remove all media */
5863 	ifmedia_removeall(&ctx->ifc_media);
5864 
5865 	/* Ensure that VLAN event handlers are unregistered */
5866 	iflib_unregister_vlan_handlers(ctx);
5867 
5868 	/* Release kobject reference */
5869 	kobj_delete((kobj_t) ctx, NULL);
5870 
5871 	/* Free the ifnet structure */
5872 	if_free(ifp);
5873 
5874 	STATE_LOCK_DESTROY(ctx);
5875 
5876 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
5877 	CTX_LOCK_DESTROY(ctx);
5878 }
5879 
5880 static int
5881 iflib_queues_alloc(if_ctx_t ctx)
5882 {
5883 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5884 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5885 	device_t dev = ctx->ifc_dev;
5886 	int nrxqsets = scctx->isc_nrxqsets;
5887 	int ntxqsets = scctx->isc_ntxqsets;
5888 	iflib_txq_t txq;
5889 	iflib_rxq_t rxq;
5890 	iflib_fl_t fl = NULL;
5891 	int i, j, cpu, err, txconf, rxconf;
5892 	iflib_dma_info_t ifdip;
5893 	uint32_t *rxqsizes = scctx->isc_rxqsizes;
5894 	uint32_t *txqsizes = scctx->isc_txqsizes;
5895 	uint8_t nrxqs = sctx->isc_nrxqs;
5896 	uint8_t ntxqs = sctx->isc_ntxqs;
5897 	int nfree_lists = sctx->isc_nfl ? sctx->isc_nfl : 1;
5898 	int fl_offset = (sctx->isc_flags & IFLIB_HAS_RXCQ ? 1 : 0);
5899 	caddr_t *vaddrs;
5900 	uint64_t *paddrs;
5901 
5902 	KASSERT(ntxqs > 0, ("number of queues per qset must be at least 1"));
5903 	KASSERT(nrxqs > 0, ("number of queues per qset must be at least 1"));
5904 	KASSERT(nrxqs >= fl_offset + nfree_lists,
5905            ("there must be at least a rxq for each free list"));
5906 
5907 	/* Allocate the TX ring struct memory */
5908 	if (!(ctx->ifc_txqs =
5909 	    (iflib_txq_t) malloc(sizeof(struct iflib_txq) *
5910 	    ntxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5911 		device_printf(dev, "Unable to allocate TX ring memory\n");
5912 		err = ENOMEM;
5913 		goto fail;
5914 	}
5915 
5916 	/* Now allocate the RX */
5917 	if (!(ctx->ifc_rxqs =
5918 	    (iflib_rxq_t) malloc(sizeof(struct iflib_rxq) *
5919 	    nrxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) {
5920 		device_printf(dev, "Unable to allocate RX ring memory\n");
5921 		err = ENOMEM;
5922 		goto rx_fail;
5923 	}
5924 
5925 	txq = ctx->ifc_txqs;
5926 	rxq = ctx->ifc_rxqs;
5927 
5928 	/*
5929 	 * XXX handle allocation failure
5930 	 */
5931 	for (txconf = i = 0, cpu = CPU_FIRST(); i < ntxqsets; i++, txconf++, txq++, cpu = CPU_NEXT(cpu)) {
5932 		/* Set up some basics */
5933 
5934 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * ntxqs,
5935 		    M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
5936 			device_printf(dev,
5937 			    "Unable to allocate TX DMA info memory\n");
5938 			err = ENOMEM;
5939 			goto err_tx_desc;
5940 		}
5941 		txq->ift_ifdi = ifdip;
5942 		for (j = 0; j < ntxqs; j++, ifdip++) {
5943 			if (iflib_dma_alloc(ctx, txqsizes[j], ifdip, 0)) {
5944 				device_printf(dev,
5945 				    "Unable to allocate TX descriptors\n");
5946 				err = ENOMEM;
5947 				goto err_tx_desc;
5948 			}
5949 			txq->ift_txd_size[j] = scctx->isc_txd_size[j];
5950 			bzero((void *)ifdip->idi_vaddr, txqsizes[j]);
5951 		}
5952 		txq->ift_ctx = ctx;
5953 		txq->ift_id = i;
5954 		if (sctx->isc_flags & IFLIB_HAS_TXCQ) {
5955 			txq->ift_br_offset = 1;
5956 		} else {
5957 			txq->ift_br_offset = 0;
5958 		}
5959 
5960 		if (iflib_txsd_alloc(txq)) {
5961 			device_printf(dev, "Critical Failure setting up TX buffers\n");
5962 			err = ENOMEM;
5963 			goto err_tx_desc;
5964 		}
5965 
5966 		/* Initialize the TX lock */
5967 		snprintf(txq->ift_mtx_name, MTX_NAME_LEN, "%s:TX(%d):callout",
5968 		    device_get_nameunit(dev), txq->ift_id);
5969 		mtx_init(&txq->ift_mtx, txq->ift_mtx_name, NULL, MTX_DEF);
5970 		callout_init_mtx(&txq->ift_timer, &txq->ift_mtx, 0);
5971 		txq->ift_timer.c_cpu = cpu;
5972 #ifdef DEV_NETMAP
5973 		callout_init_mtx(&txq->ift_netmap_timer, &txq->ift_mtx, 0);
5974 		txq->ift_netmap_timer.c_cpu = cpu;
5975 #endif /* DEV_NETMAP */
5976 
5977 		err = ifmp_ring_alloc(&txq->ift_br, 2048, txq, iflib_txq_drain,
5978 				      iflib_txq_can_drain, M_IFLIB, M_WAITOK);
5979 		if (err) {
5980 			/* XXX free any allocated rings */
5981 			device_printf(dev, "Unable to allocate buf_ring\n");
5982 			goto err_tx_desc;
5983 		}
5984 	}
5985 
5986 	for (rxconf = i = 0; i < nrxqsets; i++, rxconf++, rxq++) {
5987 		/* Set up some basics */
5988 		callout_init(&rxq->ifr_watchdog, 1);
5989 
5990 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * nrxqs,
5991 		   M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) {
5992 			device_printf(dev,
5993 			    "Unable to allocate RX DMA info memory\n");
5994 			err = ENOMEM;
5995 			goto err_tx_desc;
5996 		}
5997 
5998 		rxq->ifr_ifdi = ifdip;
5999 		/* XXX this needs to be changed if #rx queues != #tx queues */
6000 		rxq->ifr_ntxqirq = 1;
6001 		rxq->ifr_txqid[0] = i;
6002 		for (j = 0; j < nrxqs; j++, ifdip++) {
6003 			if (iflib_dma_alloc(ctx, rxqsizes[j], ifdip, 0)) {
6004 				device_printf(dev,
6005 				    "Unable to allocate RX descriptors\n");
6006 				err = ENOMEM;
6007 				goto err_tx_desc;
6008 			}
6009 			bzero((void *)ifdip->idi_vaddr, rxqsizes[j]);
6010 		}
6011 		rxq->ifr_ctx = ctx;
6012 		rxq->ifr_id = i;
6013 		rxq->ifr_fl_offset = fl_offset;
6014 		rxq->ifr_nfl = nfree_lists;
6015 		if (!(fl =
6016 			  (iflib_fl_t) malloc(sizeof(struct iflib_fl) * nfree_lists, M_IFLIB, M_NOWAIT | M_ZERO))) {
6017 			device_printf(dev, "Unable to allocate free list memory\n");
6018 			err = ENOMEM;
6019 			goto err_tx_desc;
6020 		}
6021 		rxq->ifr_fl = fl;
6022 		for (j = 0; j < nfree_lists; j++) {
6023 			fl[j].ifl_rxq = rxq;
6024 			fl[j].ifl_id = j;
6025 			fl[j].ifl_ifdi = &rxq->ifr_ifdi[j + rxq->ifr_fl_offset];
6026 			fl[j].ifl_rxd_size = scctx->isc_rxd_size[j];
6027 		}
6028 		/* Allocate receive buffers for the ring */
6029 		if (iflib_rxsd_alloc(rxq)) {
6030 			device_printf(dev,
6031 			    "Critical Failure setting up receive buffers\n");
6032 			err = ENOMEM;
6033 			goto err_rx_desc;
6034 		}
6035 
6036 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
6037 			fl->ifl_rx_bitmap = bit_alloc(fl->ifl_size, M_IFLIB,
6038 			    M_WAITOK);
6039 	}
6040 
6041 	/* TXQs */
6042 	vaddrs = malloc(sizeof(caddr_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
6043 	paddrs = malloc(sizeof(uint64_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
6044 	for (i = 0; i < ntxqsets; i++) {
6045 		iflib_dma_info_t di = ctx->ifc_txqs[i].ift_ifdi;
6046 
6047 		for (j = 0; j < ntxqs; j++, di++) {
6048 			vaddrs[i*ntxqs + j] = di->idi_vaddr;
6049 			paddrs[i*ntxqs + j] = di->idi_paddr;
6050 		}
6051 	}
6052 	if ((err = IFDI_TX_QUEUES_ALLOC(ctx, vaddrs, paddrs, ntxqs, ntxqsets)) != 0) {
6053 		device_printf(ctx->ifc_dev,
6054 		    "Unable to allocate device TX queue\n");
6055 		iflib_tx_structures_free(ctx);
6056 		free(vaddrs, M_IFLIB);
6057 		free(paddrs, M_IFLIB);
6058 		goto err_rx_desc;
6059 	}
6060 	free(vaddrs, M_IFLIB);
6061 	free(paddrs, M_IFLIB);
6062 
6063 	/* RXQs */
6064 	vaddrs = malloc(sizeof(caddr_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
6065 	paddrs = malloc(sizeof(uint64_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
6066 	for (i = 0; i < nrxqsets; i++) {
6067 		iflib_dma_info_t di = ctx->ifc_rxqs[i].ifr_ifdi;
6068 
6069 		for (j = 0; j < nrxqs; j++, di++) {
6070 			vaddrs[i*nrxqs + j] = di->idi_vaddr;
6071 			paddrs[i*nrxqs + j] = di->idi_paddr;
6072 		}
6073 	}
6074 	if ((err = IFDI_RX_QUEUES_ALLOC(ctx, vaddrs, paddrs, nrxqs, nrxqsets)) != 0) {
6075 		device_printf(ctx->ifc_dev,
6076 		    "Unable to allocate device RX queue\n");
6077 		iflib_tx_structures_free(ctx);
6078 		free(vaddrs, M_IFLIB);
6079 		free(paddrs, M_IFLIB);
6080 		goto err_rx_desc;
6081 	}
6082 	free(vaddrs, M_IFLIB);
6083 	free(paddrs, M_IFLIB);
6084 
6085 	return (0);
6086 
6087 /* XXX handle allocation failure changes */
6088 err_rx_desc:
6089 err_tx_desc:
6090 rx_fail:
6091 	if (ctx->ifc_rxqs != NULL)
6092 		free(ctx->ifc_rxqs, M_IFLIB);
6093 	ctx->ifc_rxqs = NULL;
6094 	if (ctx->ifc_txqs != NULL)
6095 		free(ctx->ifc_txqs, M_IFLIB);
6096 	ctx->ifc_txqs = NULL;
6097 fail:
6098 	return (err);
6099 }
6100 
6101 static int
6102 iflib_tx_structures_setup(if_ctx_t ctx)
6103 {
6104 	iflib_txq_t txq = ctx->ifc_txqs;
6105 	int i;
6106 
6107 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
6108 		iflib_txq_setup(txq);
6109 
6110 	return (0);
6111 }
6112 
6113 static void
6114 iflib_tx_structures_free(if_ctx_t ctx)
6115 {
6116 	iflib_txq_t txq = ctx->ifc_txqs;
6117 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6118 	int i, j;
6119 
6120 	for (i = 0; i < NTXQSETS(ctx); i++, txq++) {
6121 		for (j = 0; j < sctx->isc_ntxqs; j++)
6122 			iflib_dma_free(&txq->ift_ifdi[j]);
6123 		iflib_txq_destroy(txq);
6124 	}
6125 	free(ctx->ifc_txqs, M_IFLIB);
6126 	ctx->ifc_txqs = NULL;
6127 }
6128 
6129 /*********************************************************************
6130  *
6131  *  Initialize all receive rings.
6132  *
6133  **********************************************************************/
6134 static int
6135 iflib_rx_structures_setup(if_ctx_t ctx)
6136 {
6137 	iflib_rxq_t rxq = ctx->ifc_rxqs;
6138 	int q;
6139 #if defined(INET6) || defined(INET)
6140 	int err, i;
6141 #endif
6142 
6143 	for (q = 0; q < ctx->ifc_softc_ctx.isc_nrxqsets; q++, rxq++) {
6144 #if defined(INET6) || defined(INET)
6145 		if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_LRO) {
6146 			err = tcp_lro_init_args(&rxq->ifr_lc, ctx->ifc_ifp,
6147 			    TCP_LRO_ENTRIES, min(1024,
6148 			    ctx->ifc_softc_ctx.isc_nrxd[rxq->ifr_fl_offset]));
6149 			if (err != 0) {
6150 				device_printf(ctx->ifc_dev,
6151 				    "LRO Initialization failed!\n");
6152 				goto fail;
6153 			}
6154 		}
6155 #endif
6156 		IFDI_RXQ_SETUP(ctx, rxq->ifr_id);
6157 	}
6158 	return (0);
6159 #if defined(INET6) || defined(INET)
6160 fail:
6161 	/*
6162 	 * Free LRO resources allocated so far, we will only handle
6163 	 * the rings that completed, the failing case will have
6164 	 * cleaned up for itself.  'q' failed, so its the terminus.
6165 	 */
6166 	rxq = ctx->ifc_rxqs;
6167 	for (i = 0; i < q; ++i, rxq++) {
6168 		if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_LRO)
6169 			tcp_lro_free(&rxq->ifr_lc);
6170 	}
6171 	return (err);
6172 #endif
6173 }
6174 
6175 /*********************************************************************
6176  *
6177  *  Free all receive rings.
6178  *
6179  **********************************************************************/
6180 static void
6181 iflib_rx_structures_free(if_ctx_t ctx)
6182 {
6183 	iflib_rxq_t rxq = ctx->ifc_rxqs;
6184 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6185 	int i, j;
6186 
6187 	for (i = 0; i < ctx->ifc_softc_ctx.isc_nrxqsets; i++, rxq++) {
6188 		for (j = 0; j < sctx->isc_nrxqs; j++)
6189 			iflib_dma_free(&rxq->ifr_ifdi[j]);
6190 		iflib_rx_sds_free(rxq);
6191 #if defined(INET6) || defined(INET)
6192 		if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_LRO)
6193 			tcp_lro_free(&rxq->ifr_lc);
6194 #endif
6195 	}
6196 	free(ctx->ifc_rxqs, M_IFLIB);
6197 	ctx->ifc_rxqs = NULL;
6198 }
6199 
6200 static int
6201 iflib_qset_structures_setup(if_ctx_t ctx)
6202 {
6203 	int err;
6204 
6205 	/*
6206 	 * It is expected that the caller takes care of freeing queues if this
6207 	 * fails.
6208 	 */
6209 	if ((err = iflib_tx_structures_setup(ctx)) != 0) {
6210 		device_printf(ctx->ifc_dev, "iflib_tx_structures_setup failed: %d\n", err);
6211 		return (err);
6212 	}
6213 
6214 	if ((err = iflib_rx_structures_setup(ctx)) != 0)
6215 		device_printf(ctx->ifc_dev, "iflib_rx_structures_setup failed: %d\n", err);
6216 
6217 	return (err);
6218 }
6219 
6220 int
6221 iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
6222 		driver_filter_t filter, void *filter_arg, driver_intr_t handler, void *arg, const char *name)
6223 {
6224 
6225 	return (_iflib_irq_alloc(ctx, irq, rid, filter, handler, arg, name));
6226 }
6227 
6228 /* Just to avoid copy/paste */
6229 static inline int
6230 iflib_irq_set_affinity(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type,
6231     int qid, struct grouptask *gtask, struct taskqgroup *tqg, void *uniq,
6232     const char *name)
6233 {
6234 	device_t dev;
6235 	unsigned int base_cpuid, cpuid;
6236 	int err;
6237 
6238 	dev = ctx->ifc_dev;
6239 	base_cpuid = ctx->ifc_sysctl_core_offset;
6240 	cpuid = get_cpuid_for_queue(ctx, base_cpuid, qid, type == IFLIB_INTR_TX);
6241 	err = taskqgroup_attach_cpu(tqg, gtask, uniq, cpuid, dev,
6242 	    irq ? irq->ii_res : NULL, name);
6243 	if (err) {
6244 		device_printf(dev, "taskqgroup_attach_cpu failed %d\n", err);
6245 		return (err);
6246 	}
6247 #ifdef notyet
6248 	if (cpuid > ctx->ifc_cpuid_highest)
6249 		ctx->ifc_cpuid_highest = cpuid;
6250 #endif
6251 	return (0);
6252 }
6253 
6254 int
6255 iflib_irq_alloc_generic(if_ctx_t ctx, if_irq_t irq, int rid,
6256 			iflib_intr_type_t type, driver_filter_t *filter,
6257 			void *filter_arg, int qid, const char *name)
6258 {
6259 	device_t dev;
6260 	struct grouptask *gtask;
6261 	struct taskqgroup *tqg;
6262 	iflib_filter_info_t info;
6263 	gtask_fn_t *fn;
6264 	int tqrid, err;
6265 	driver_filter_t *intr_fast;
6266 	void *q;
6267 
6268 	info = &ctx->ifc_filter_info;
6269 	tqrid = rid;
6270 
6271 	switch (type) {
6272 	/* XXX merge tx/rx for netmap? */
6273 	case IFLIB_INTR_TX:
6274 		q = &ctx->ifc_txqs[qid];
6275 		info = &ctx->ifc_txqs[qid].ift_filter_info;
6276 		gtask = &ctx->ifc_txqs[qid].ift_task;
6277 		tqg = qgroup_if_io_tqg;
6278 		fn = _task_fn_tx;
6279 		intr_fast = iflib_fast_intr;
6280 		GROUPTASK_INIT(gtask, 0, fn, q);
6281 		ctx->ifc_flags |= IFC_NETMAP_TX_IRQ;
6282 		break;
6283 	case IFLIB_INTR_RX:
6284 		q = &ctx->ifc_rxqs[qid];
6285 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
6286 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
6287 		tqg = qgroup_if_io_tqg;
6288 		fn = _task_fn_rx;
6289 		intr_fast = iflib_fast_intr;
6290 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
6291 		break;
6292 	case IFLIB_INTR_RXTX:
6293 		q = &ctx->ifc_rxqs[qid];
6294 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
6295 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
6296 		tqg = qgroup_if_io_tqg;
6297 		fn = _task_fn_rx;
6298 		intr_fast = iflib_fast_intr_rxtx;
6299 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
6300 		break;
6301 	case IFLIB_INTR_ADMIN:
6302 		q = ctx;
6303 		tqrid = -1;
6304 		info = &ctx->ifc_filter_info;
6305 		gtask = &ctx->ifc_admin_task;
6306 		tqg = qgroup_if_config_tqg;
6307 		fn = _task_fn_admin;
6308 		intr_fast = iflib_fast_intr_ctx;
6309 		break;
6310 	default:
6311 		device_printf(ctx->ifc_dev, "%s: unknown net intr type\n",
6312 		    __func__);
6313 		return (EINVAL);
6314 	}
6315 
6316 	info->ifi_filter = filter;
6317 	info->ifi_filter_arg = filter_arg;
6318 	info->ifi_task = gtask;
6319 	info->ifi_ctx = q;
6320 
6321 	dev = ctx->ifc_dev;
6322 	err = _iflib_irq_alloc(ctx, irq, rid, intr_fast, NULL, info,  name);
6323 	if (err != 0) {
6324 		device_printf(dev, "_iflib_irq_alloc failed %d\n", err);
6325 		return (err);
6326 	}
6327 	if (type == IFLIB_INTR_ADMIN)
6328 		return (0);
6329 
6330 	if (tqrid != -1) {
6331 		err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg, q,
6332 		    name);
6333 		if (err)
6334 			return (err);
6335 	} else {
6336 		taskqgroup_attach(tqg, gtask, q, dev, irq->ii_res, name);
6337 	}
6338 
6339 	return (0);
6340 }
6341 
6342 void
6343 iflib_softirq_alloc_generic(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type, void *arg, int qid, const char *name)
6344 {
6345 	device_t dev;
6346 	struct grouptask *gtask;
6347 	struct taskqgroup *tqg;
6348 	gtask_fn_t *fn;
6349 	void *q;
6350 	int err;
6351 
6352 	switch (type) {
6353 	case IFLIB_INTR_TX:
6354 		q = &ctx->ifc_txqs[qid];
6355 		gtask = &ctx->ifc_txqs[qid].ift_task;
6356 		tqg = qgroup_if_io_tqg;
6357 		fn = _task_fn_tx;
6358 		GROUPTASK_INIT(gtask, 0, fn, q);
6359 		break;
6360 	case IFLIB_INTR_RX:
6361 		q = &ctx->ifc_rxqs[qid];
6362 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
6363 		tqg = qgroup_if_io_tqg;
6364 		fn = _task_fn_rx;
6365 		NET_GROUPTASK_INIT(gtask, 0, fn, q);
6366 		break;
6367 	case IFLIB_INTR_IOV:
6368 		q = ctx;
6369 		gtask = &ctx->ifc_vflr_task;
6370 		tqg = qgroup_if_config_tqg;
6371 		fn = _task_fn_iov;
6372 		GROUPTASK_INIT(gtask, 0, fn, q);
6373 		break;
6374 	default:
6375 		panic("unknown net intr type");
6376 	}
6377 	err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg, q, name);
6378 	if (err) {
6379 		dev = ctx->ifc_dev;
6380 		taskqgroup_attach(tqg, gtask, q, dev, irq ? irq->ii_res : NULL,
6381 		    name);
6382 	}
6383 }
6384 
6385 void
6386 iflib_irq_free(if_ctx_t ctx, if_irq_t irq)
6387 {
6388 
6389 	if (irq->ii_tag)
6390 		bus_teardown_intr(ctx->ifc_dev, irq->ii_res, irq->ii_tag);
6391 
6392 	if (irq->ii_res)
6393 		bus_release_resource(ctx->ifc_dev, SYS_RES_IRQ,
6394 		    rman_get_rid(irq->ii_res), irq->ii_res);
6395 }
6396 
6397 static int
6398 iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filter_arg, int *rid, const char *name)
6399 {
6400 	iflib_txq_t txq = ctx->ifc_txqs;
6401 	iflib_rxq_t rxq = ctx->ifc_rxqs;
6402 	if_irq_t irq = &ctx->ifc_legacy_irq;
6403 	iflib_filter_info_t info;
6404 	device_t dev;
6405 	struct grouptask *gtask;
6406 	struct resource *res;
6407 	struct taskqgroup *tqg;
6408 	void *q;
6409 	int err, tqrid;
6410 	bool rx_only;
6411 
6412 	q = &ctx->ifc_rxqs[0];
6413 	info = &rxq[0].ifr_filter_info;
6414 	gtask = &rxq[0].ifr_task;
6415 	tqg = qgroup_if_io_tqg;
6416 	tqrid = *rid;
6417 	rx_only = (ctx->ifc_sctx->isc_flags & IFLIB_SINGLE_IRQ_RX_ONLY) != 0;
6418 
6419 	ctx->ifc_flags |= IFC_LEGACY;
6420 	info->ifi_filter = filter;
6421 	info->ifi_filter_arg = filter_arg;
6422 	info->ifi_task = gtask;
6423 	info->ifi_ctx = rx_only ? ctx : q;
6424 
6425 	dev = ctx->ifc_dev;
6426 	/* We allocate a single interrupt resource */
6427 	err = _iflib_irq_alloc(ctx, irq, tqrid, rx_only ? iflib_fast_intr_ctx :
6428 	    iflib_fast_intr_rxtx, NULL, info, name);
6429 	if (err != 0)
6430 		return (err);
6431 	NET_GROUPTASK_INIT(gtask, 0, _task_fn_rx, q);
6432 	res = irq->ii_res;
6433 	taskqgroup_attach(tqg, gtask, q, dev, res, name);
6434 
6435 	GROUPTASK_INIT(&txq->ift_task, 0, _task_fn_tx, txq);
6436 	taskqgroup_attach(qgroup_if_io_tqg, &txq->ift_task, txq, dev, res,
6437 	    "tx");
6438 	return (0);
6439 }
6440 
6441 void
6442 iflib_led_create(if_ctx_t ctx)
6443 {
6444 
6445 	ctx->ifc_led_dev = led_create(iflib_led_func, ctx,
6446 	    device_get_nameunit(ctx->ifc_dev));
6447 }
6448 
6449 void
6450 iflib_tx_intr_deferred(if_ctx_t ctx, int txqid)
6451 {
6452 
6453 	GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
6454 }
6455 
6456 void
6457 iflib_rx_intr_deferred(if_ctx_t ctx, int rxqid)
6458 {
6459 
6460 	GROUPTASK_ENQUEUE(&ctx->ifc_rxqs[rxqid].ifr_task);
6461 }
6462 
6463 void
6464 iflib_admin_intr_deferred(if_ctx_t ctx)
6465 {
6466 
6467 	MPASS(ctx->ifc_admin_task.gt_taskqueue != NULL);
6468 	GROUPTASK_ENQUEUE(&ctx->ifc_admin_task);
6469 }
6470 
6471 void
6472 iflib_iov_intr_deferred(if_ctx_t ctx)
6473 {
6474 
6475 	GROUPTASK_ENQUEUE(&ctx->ifc_vflr_task);
6476 }
6477 
6478 void
6479 iflib_io_tqg_attach(struct grouptask *gt, void *uniq, int cpu, const char *name)
6480 {
6481 
6482 	taskqgroup_attach_cpu(qgroup_if_io_tqg, gt, uniq, cpu, NULL, NULL,
6483 	    name);
6484 }
6485 
6486 void
6487 iflib_config_gtask_init(void *ctx, struct grouptask *gtask, gtask_fn_t *fn,
6488 	const char *name)
6489 {
6490 
6491 	GROUPTASK_INIT(gtask, 0, fn, ctx);
6492 	taskqgroup_attach(qgroup_if_config_tqg, gtask, gtask, NULL, NULL,
6493 	    name);
6494 }
6495 
6496 void
6497 iflib_config_gtask_deinit(struct grouptask *gtask)
6498 {
6499 
6500 	taskqgroup_detach(qgroup_if_config_tqg, gtask);
6501 }
6502 
6503 void
6504 iflib_link_state_change(if_ctx_t ctx, int link_state, uint64_t baudrate)
6505 {
6506 	if_t ifp = ctx->ifc_ifp;
6507 	iflib_txq_t txq = ctx->ifc_txqs;
6508 
6509 	if_setbaudrate(ifp, baudrate);
6510 	if (baudrate >= IF_Gbps(10)) {
6511 		STATE_LOCK(ctx);
6512 		ctx->ifc_flags |= IFC_PREFETCH;
6513 		STATE_UNLOCK(ctx);
6514 	}
6515 	/* If link down, disable watchdog */
6516 	if ((ctx->ifc_link_state == LINK_STATE_UP) && (link_state == LINK_STATE_DOWN)) {
6517 		for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxqsets; i++, txq++)
6518 			txq->ift_qstatus = IFLIB_QUEUE_IDLE;
6519 	}
6520 	ctx->ifc_link_state = link_state;
6521 	if_link_state_change(ifp, link_state);
6522 }
6523 
6524 static int
6525 iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq)
6526 {
6527 	int credits;
6528 #ifdef INVARIANTS
6529 	int credits_pre = txq->ift_cidx_processed;
6530 #endif
6531 
6532 	bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
6533 	    BUS_DMASYNC_POSTREAD);
6534 	if ((credits = ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, true)) == 0)
6535 		return (0);
6536 
6537 	txq->ift_processed += credits;
6538 	txq->ift_cidx_processed += credits;
6539 
6540 	MPASS(credits_pre + credits == txq->ift_cidx_processed);
6541 	if (txq->ift_cidx_processed >= txq->ift_size)
6542 		txq->ift_cidx_processed -= txq->ift_size;
6543 	return (credits);
6544 }
6545 
6546 static int
6547 iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget)
6548 {
6549 	iflib_fl_t fl;
6550 	u_int i;
6551 
6552 	for (i = 0, fl = &rxq->ifr_fl[0]; i < rxq->ifr_nfl; i++, fl++)
6553 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
6554 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
6555 	return (ctx->isc_rxd_available(ctx->ifc_softc, rxq->ifr_id, cidx,
6556 	    budget));
6557 }
6558 
6559 void
6560 iflib_add_int_delay_sysctl(if_ctx_t ctx, const char *name,
6561 	const char *description, if_int_delay_info_t info,
6562 	int offset, int value)
6563 {
6564 	info->iidi_ctx = ctx;
6565 	info->iidi_offset = offset;
6566 	info->iidi_value = value;
6567 	SYSCTL_ADD_PROC(device_get_sysctl_ctx(ctx->ifc_dev),
6568 	    SYSCTL_CHILDREN(device_get_sysctl_tree(ctx->ifc_dev)),
6569 	    OID_AUTO, name, CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE,
6570 	    info, 0, iflib_sysctl_int_delay, "I", description);
6571 }
6572 
6573 struct sx *
6574 iflib_ctx_lock_get(if_ctx_t ctx)
6575 {
6576 
6577 	return (&ctx->ifc_ctx_sx);
6578 }
6579 
6580 static int
6581 iflib_msix_init(if_ctx_t ctx)
6582 {
6583 	device_t dev = ctx->ifc_dev;
6584 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6585 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6586 	int admincnt, bar, err, iflib_num_rx_queues, iflib_num_tx_queues;
6587 	int msgs, queuemsgs, queues, rx_queues, tx_queues, vectors;
6588 
6589 	iflib_num_tx_queues = ctx->ifc_sysctl_ntxqs;
6590 	iflib_num_rx_queues = ctx->ifc_sysctl_nrxqs;
6591 
6592 	if (bootverbose)
6593 		device_printf(dev, "msix_init qsets capped at %d\n",
6594 		    imax(scctx->isc_ntxqsets, scctx->isc_nrxqsets));
6595 
6596 	/* Override by tuneable */
6597 	if (scctx->isc_disable_msix)
6598 		goto msi;
6599 
6600 	/* First try MSI-X */
6601 	if ((msgs = pci_msix_count(dev)) == 0) {
6602 		if (bootverbose)
6603 			device_printf(dev, "MSI-X not supported or disabled\n");
6604 		goto msi;
6605 	}
6606 
6607 	bar = ctx->ifc_softc_ctx.isc_msix_bar;
6608 	/*
6609 	 * bar == -1 => "trust me I know what I'm doing"
6610 	 * Some drivers are for hardware that is so shoddily
6611 	 * documented that no one knows which bars are which
6612 	 * so the developer has to map all bars. This hack
6613 	 * allows shoddy garbage to use MSI-X in this framework.
6614 	 */
6615 	if (bar != -1) {
6616 		ctx->ifc_msix_mem = bus_alloc_resource_any(dev,
6617 	            SYS_RES_MEMORY, &bar, RF_ACTIVE);
6618 		if (ctx->ifc_msix_mem == NULL) {
6619 			device_printf(dev, "Unable to map MSI-X table\n");
6620 			goto msi;
6621 		}
6622 	}
6623 
6624 	admincnt = sctx->isc_admin_intrcnt;
6625 #if IFLIB_DEBUG
6626 	/* use only 1 qset in debug mode */
6627 	queuemsgs = min(msgs - admincnt, 1);
6628 #else
6629 	queuemsgs = msgs - admincnt;
6630 #endif
6631 #ifdef RSS
6632 	queues = imin(queuemsgs, rss_getnumbuckets());
6633 #else
6634 	queues = queuemsgs;
6635 #endif
6636 	queues = imin(CPU_COUNT(&ctx->ifc_cpus), queues);
6637 	if (bootverbose)
6638 		device_printf(dev,
6639 		    "intr CPUs: %d queue msgs: %d admincnt: %d\n",
6640 		    CPU_COUNT(&ctx->ifc_cpus), queuemsgs, admincnt);
6641 #ifdef  RSS
6642 	/* If we're doing RSS, clamp at the number of RSS buckets */
6643 	if (queues > rss_getnumbuckets())
6644 		queues = rss_getnumbuckets();
6645 #endif
6646 	if (iflib_num_rx_queues > 0 && iflib_num_rx_queues < queuemsgs - admincnt)
6647 		rx_queues = iflib_num_rx_queues;
6648 	else
6649 		rx_queues = queues;
6650 
6651 	if (rx_queues > scctx->isc_nrxqsets)
6652 		rx_queues = scctx->isc_nrxqsets;
6653 
6654 	/*
6655 	 * We want this to be all logical CPUs by default
6656 	 */
6657 	if (iflib_num_tx_queues > 0 && iflib_num_tx_queues < queues)
6658 		tx_queues = iflib_num_tx_queues;
6659 	else
6660 		tx_queues = mp_ncpus;
6661 
6662 	if (tx_queues > scctx->isc_ntxqsets)
6663 		tx_queues = scctx->isc_ntxqsets;
6664 
6665 	if (ctx->ifc_sysctl_qs_eq_override == 0) {
6666 #ifdef INVARIANTS
6667 		if (tx_queues != rx_queues)
6668 			device_printf(dev,
6669 			    "queue equality override not set, capping rx_queues at %d and tx_queues at %d\n",
6670 			    min(rx_queues, tx_queues), min(rx_queues, tx_queues));
6671 #endif
6672 		tx_queues = min(rx_queues, tx_queues);
6673 		rx_queues = min(rx_queues, tx_queues);
6674 	}
6675 
6676 	vectors = rx_queues + admincnt;
6677 	if (msgs < vectors) {
6678 		device_printf(dev,
6679 		    "insufficient number of MSI-X vectors "
6680 		    "(supported %d, need %d)\n", msgs, vectors);
6681 		goto msi;
6682 	}
6683 
6684 	device_printf(dev, "Using %d RX queues %d TX queues\n", rx_queues,
6685 	    tx_queues);
6686 	msgs = vectors;
6687 	if ((err = pci_alloc_msix(dev, &vectors)) == 0) {
6688 		if (vectors != msgs) {
6689 			device_printf(dev,
6690 			    "Unable to allocate sufficient MSI-X vectors "
6691 			    "(got %d, need %d)\n", vectors, msgs);
6692 			pci_release_msi(dev);
6693 			if (bar != -1) {
6694 				bus_release_resource(dev, SYS_RES_MEMORY, bar,
6695 				    ctx->ifc_msix_mem);
6696 				ctx->ifc_msix_mem = NULL;
6697 			}
6698 			goto msi;
6699 		}
6700 		device_printf(dev, "Using MSI-X interrupts with %d vectors\n",
6701 		    vectors);
6702 		scctx->isc_vectors = vectors;
6703 		scctx->isc_nrxqsets = rx_queues;
6704 		scctx->isc_ntxqsets = tx_queues;
6705 		scctx->isc_intr = IFLIB_INTR_MSIX;
6706 
6707 		return (vectors);
6708 	} else {
6709 		device_printf(dev,
6710 		    "failed to allocate %d MSI-X vectors, err: %d\n", vectors,
6711 		    err);
6712 		if (bar != -1) {
6713 			bus_release_resource(dev, SYS_RES_MEMORY, bar,
6714 			    ctx->ifc_msix_mem);
6715 			ctx->ifc_msix_mem = NULL;
6716 		}
6717 	}
6718 
6719 msi:
6720 	vectors = pci_msi_count(dev);
6721 	scctx->isc_nrxqsets = 1;
6722 	scctx->isc_ntxqsets = 1;
6723 	scctx->isc_vectors = vectors;
6724 	if (vectors == 1 && pci_alloc_msi(dev, &vectors) == 0) {
6725 		device_printf(dev,"Using an MSI interrupt\n");
6726 		scctx->isc_intr = IFLIB_INTR_MSI;
6727 	} else {
6728 		scctx->isc_vectors = 1;
6729 		device_printf(dev,"Using a Legacy interrupt\n");
6730 		scctx->isc_intr = IFLIB_INTR_LEGACY;
6731 	}
6732 
6733 	return (vectors);
6734 }
6735 
6736 static const char *ring_states[] = { "IDLE", "BUSY", "STALLED", "ABDICATED" };
6737 
6738 static int
6739 mp_ring_state_handler(SYSCTL_HANDLER_ARGS)
6740 {
6741 	int rc;
6742 	uint16_t *state = ((uint16_t *)oidp->oid_arg1);
6743 	struct sbuf *sb;
6744 	const char *ring_state = "UNKNOWN";
6745 
6746 	/* XXX needed ? */
6747 	rc = sysctl_wire_old_buffer(req, 0);
6748 	MPASS(rc == 0);
6749 	if (rc != 0)
6750 		return (rc);
6751 	sb = sbuf_new_for_sysctl(NULL, NULL, 80, req);
6752 	MPASS(sb != NULL);
6753 	if (sb == NULL)
6754 		return (ENOMEM);
6755 	if (state[3] <= 3)
6756 		ring_state = ring_states[state[3]];
6757 
6758 	sbuf_printf(sb, "pidx_head: %04hd pidx_tail: %04hd cidx: %04hd state: %s",
6759 		    state[0], state[1], state[2], ring_state);
6760 	rc = sbuf_finish(sb);
6761 	sbuf_delete(sb);
6762         return(rc);
6763 }
6764 
6765 enum iflib_ndesc_handler {
6766 	IFLIB_NTXD_HANDLER,
6767 	IFLIB_NRXD_HANDLER,
6768 };
6769 
6770 static int
6771 mp_ndesc_handler(SYSCTL_HANDLER_ARGS)
6772 {
6773 	if_ctx_t ctx = (void *)arg1;
6774 	enum iflib_ndesc_handler type = arg2;
6775 	char buf[256] = {0};
6776 	qidx_t *ndesc;
6777 	char *p, *next;
6778 	int nqs, rc, i;
6779 
6780 	nqs = 8;
6781 	switch(type) {
6782 	case IFLIB_NTXD_HANDLER:
6783 		ndesc = ctx->ifc_sysctl_ntxds;
6784 		if (ctx->ifc_sctx)
6785 			nqs = ctx->ifc_sctx->isc_ntxqs;
6786 		break;
6787 	case IFLIB_NRXD_HANDLER:
6788 		ndesc = ctx->ifc_sysctl_nrxds;
6789 		if (ctx->ifc_sctx)
6790 			nqs = ctx->ifc_sctx->isc_nrxqs;
6791 		break;
6792 	default:
6793 		printf("%s: unhandled type\n", __func__);
6794 		return (EINVAL);
6795 	}
6796 	if (nqs == 0)
6797 		nqs = 8;
6798 
6799 	for (i=0; i<8; i++) {
6800 		if (i >= nqs)
6801 			break;
6802 		if (i)
6803 			strcat(buf, ",");
6804 		sprintf(strchr(buf, 0), "%d", ndesc[i]);
6805 	}
6806 
6807 	rc = sysctl_handle_string(oidp, buf, sizeof(buf), req);
6808 	if (rc || req->newptr == NULL)
6809 		return rc;
6810 
6811 	for (i = 0, next = buf, p = strsep(&next, " ,"); i < 8 && p;
6812 	    i++, p = strsep(&next, " ,")) {
6813 		ndesc[i] = strtoul(p, NULL, 10);
6814 	}
6815 
6816 	return(rc);
6817 }
6818 
6819 #define NAME_BUFLEN 32
6820 static void
6821 iflib_add_device_sysctl_pre(if_ctx_t ctx)
6822 {
6823         device_t dev = iflib_get_dev(ctx);
6824 	struct sysctl_oid_list *child, *oid_list;
6825 	struct sysctl_ctx_list *ctx_list;
6826 	struct sysctl_oid *node;
6827 
6828 	ctx_list = device_get_sysctl_ctx(dev);
6829 	child = SYSCTL_CHILDREN(device_get_sysctl_tree(dev));
6830 	ctx->ifc_sysctl_node = node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, "iflib",
6831 	    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "IFLIB fields");
6832 	oid_list = SYSCTL_CHILDREN(node);
6833 
6834 	SYSCTL_ADD_CONST_STRING(ctx_list, oid_list, OID_AUTO, "driver_version",
6835 		       CTLFLAG_RD, ctx->ifc_sctx->isc_driver_version,
6836 		       "driver version");
6837 
6838 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_ntxqs",
6839 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_ntxqs, 0,
6840 			"# of txqs to use, 0 => use default #");
6841 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_nrxqs",
6842 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_nrxqs, 0,
6843 			"# of rxqs to use, 0 => use default #");
6844 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_qs_enable",
6845 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_qs_eq_override, 0,
6846                        "permit #txq != #rxq");
6847 	SYSCTL_ADD_INT(ctx_list, oid_list, OID_AUTO, "disable_msix",
6848                       CTLFLAG_RWTUN, &ctx->ifc_softc_ctx.isc_disable_msix, 0,
6849                       "disable MSI-X (default 0)");
6850 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "rx_budget",
6851 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_rx_budget, 0,
6852 		       "set the RX budget");
6853 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "tx_abdicate",
6854 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_tx_abdicate, 0,
6855 		       "cause TX to abdicate instead of running to completion");
6856 	ctx->ifc_sysctl_core_offset = CORE_OFFSET_UNSPECIFIED;
6857 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "core_offset",
6858 		       CTLFLAG_RDTUN, &ctx->ifc_sysctl_core_offset, 0,
6859 		       "offset to start using cores at");
6860 	SYSCTL_ADD_U8(ctx_list, oid_list, OID_AUTO, "separate_txrx",
6861 		       CTLFLAG_RDTUN, &ctx->ifc_sysctl_separate_txrx, 0,
6862 		       "use separate cores for TX and RX");
6863 	SYSCTL_ADD_U8(ctx_list, oid_list, OID_AUTO, "use_logical_cores",
6864 		      CTLFLAG_RDTUN, &ctx->ifc_sysctl_use_logical_cores, 0,
6865 		      "try to make use of logical cores for TX and RX");
6866 
6867 	/* XXX change for per-queue sizes */
6868 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_ntxds",
6869 	    CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_NEEDGIANT, ctx,
6870 	    IFLIB_NTXD_HANDLER, mp_ndesc_handler, "A",
6871 	    "list of # of TX descriptors to use, 0 = use default #");
6872 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_nrxds",
6873 	    CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_NEEDGIANT, ctx,
6874 	    IFLIB_NRXD_HANDLER, mp_ndesc_handler, "A",
6875 	    "list of # of RX descriptors to use, 0 = use default #");
6876 }
6877 
6878 static void
6879 iflib_add_device_sysctl_post(if_ctx_t ctx)
6880 {
6881 	if_shared_ctx_t sctx = ctx->ifc_sctx;
6882 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
6883         device_t dev = iflib_get_dev(ctx);
6884 	struct sysctl_oid_list *child;
6885 	struct sysctl_ctx_list *ctx_list;
6886 	iflib_fl_t fl;
6887 	iflib_txq_t txq;
6888 	iflib_rxq_t rxq;
6889 	int i, j;
6890 	char namebuf[NAME_BUFLEN];
6891 	char *qfmt;
6892 	struct sysctl_oid *queue_node, *fl_node, *node;
6893 	struct sysctl_oid_list *queue_list, *fl_list;
6894 	ctx_list = device_get_sysctl_ctx(dev);
6895 
6896 	node = ctx->ifc_sysctl_node;
6897 	child = SYSCTL_CHILDREN(node);
6898 
6899 	if (scctx->isc_ntxqsets > 100)
6900 		qfmt = "txq%03d";
6901 	else if (scctx->isc_ntxqsets > 10)
6902 		qfmt = "txq%02d";
6903 	else
6904 		qfmt = "txq%d";
6905 	for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) {
6906 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6907 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6908 		    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "Queue Name");
6909 		queue_list = SYSCTL_CHILDREN(queue_node);
6910 		SYSCTL_ADD_INT(ctx_list, queue_list, OID_AUTO, "cpu",
6911 			       CTLFLAG_RD,
6912 			       &txq->ift_task.gt_cpu, 0, "cpu this queue is bound to");
6913 #if MEMORY_LOGGING
6914 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_dequeued",
6915 				CTLFLAG_RD,
6916 				&txq->ift_dequeued, "total mbufs freed");
6917 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_enqueued",
6918 				CTLFLAG_RD,
6919 				&txq->ift_enqueued, "total mbufs enqueued");
6920 #endif
6921 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag",
6922 				   CTLFLAG_RD,
6923 				   &txq->ift_mbuf_defrag, "# of times m_defrag was called");
6924 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "m_pullups",
6925 				   CTLFLAG_RD,
6926 				   &txq->ift_pullups, "# of times m_pullup was called");
6927 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag_failed",
6928 				   CTLFLAG_RD,
6929 				   &txq->ift_mbuf_defrag_failed, "# of times m_defrag failed");
6930 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_desc_avail",
6931 				   CTLFLAG_RD,
6932 				   &txq->ift_no_desc_avail, "# of times no descriptors were available");
6933 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "tx_map_failed",
6934 				   CTLFLAG_RD,
6935 				   &txq->ift_map_failed, "# of times DMA map failed");
6936 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txd_encap_efbig",
6937 				   CTLFLAG_RD,
6938 				   &txq->ift_txd_encap_efbig, "# of times txd_encap returned EFBIG");
6939 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_tx_dma_setup",
6940 				   CTLFLAG_RD,
6941 				   &txq->ift_no_tx_dma_setup, "# of times map failed for other than EFBIG");
6942 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_pidx",
6943 				   CTLFLAG_RD,
6944 				   &txq->ift_pidx, 1, "Producer Index");
6945 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx",
6946 				   CTLFLAG_RD,
6947 				   &txq->ift_cidx, 1, "Consumer Index");
6948 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx_processed",
6949 				   CTLFLAG_RD,
6950 				   &txq->ift_cidx_processed, 1, "Consumer Index seen by credit update");
6951 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_in_use",
6952 				   CTLFLAG_RD,
6953 				   &txq->ift_in_use, 1, "descriptors in use");
6954 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_processed",
6955 				   CTLFLAG_RD,
6956 				   &txq->ift_processed, "descriptors procesed for clean");
6957 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_cleaned",
6958 				   CTLFLAG_RD,
6959 				   &txq->ift_cleaned, "total cleaned");
6960 		SYSCTL_ADD_PROC(ctx_list, queue_list, OID_AUTO, "ring_state",
6961 		    CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_NEEDGIANT,
6962 		    __DEVOLATILE(uint64_t *, &txq->ift_br->state), 0,
6963 		    mp_ring_state_handler, "A", "soft ring state");
6964 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_enqueues",
6965 				       CTLFLAG_RD, &txq->ift_br->enqueues,
6966 				       "# of enqueues to the mp_ring for this queue");
6967 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_drops",
6968 				       CTLFLAG_RD, &txq->ift_br->drops,
6969 				       "# of drops in the mp_ring for this queue");
6970 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_starts",
6971 				       CTLFLAG_RD, &txq->ift_br->starts,
6972 				       "# of normal consumer starts in the mp_ring for this queue");
6973 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_stalls",
6974 				       CTLFLAG_RD, &txq->ift_br->stalls,
6975 					       "# of consumer stalls in the mp_ring for this queue");
6976 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_restarts",
6977 			       CTLFLAG_RD, &txq->ift_br->restarts,
6978 				       "# of consumer restarts in the mp_ring for this queue");
6979 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_abdications",
6980 				       CTLFLAG_RD, &txq->ift_br->abdications,
6981 				       "# of consumer abdications in the mp_ring for this queue");
6982 	}
6983 
6984 	if (scctx->isc_nrxqsets > 100)
6985 		qfmt = "rxq%03d";
6986 	else if (scctx->isc_nrxqsets > 10)
6987 		qfmt = "rxq%02d";
6988 	else
6989 		qfmt = "rxq%d";
6990 	for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) {
6991 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
6992 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
6993 		    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "Queue Name");
6994 		queue_list = SYSCTL_CHILDREN(queue_node);
6995 		SYSCTL_ADD_INT(ctx_list, queue_list, OID_AUTO, "cpu",
6996 			       CTLFLAG_RD,
6997 			       &rxq->ifr_task.gt_cpu, 0, "cpu this queue is bound to");
6998 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
6999 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_cidx",
7000 				       CTLFLAG_RD,
7001 				       &rxq->ifr_cq_cidx, 1, "Consumer Index");
7002 		}
7003 
7004 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
7005 			snprintf(namebuf, NAME_BUFLEN, "rxq_fl%d", j);
7006 			fl_node = SYSCTL_ADD_NODE(ctx_list, queue_list, OID_AUTO, namebuf,
7007 			    CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "freelist Name");
7008 			fl_list = SYSCTL_CHILDREN(fl_node);
7009 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "pidx",
7010 				       CTLFLAG_RD,
7011 				       &fl->ifl_pidx, 1, "Producer Index");
7012 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "cidx",
7013 				       CTLFLAG_RD,
7014 				       &fl->ifl_cidx, 1, "Consumer Index");
7015 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "credits",
7016 				       CTLFLAG_RD,
7017 				       &fl->ifl_credits, 1, "credits available");
7018 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "buf_size",
7019 				       CTLFLAG_RD,
7020 				       &fl->ifl_buf_size, 1, "buffer size");
7021 #if MEMORY_LOGGING
7022 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_enqueued",
7023 					CTLFLAG_RD,
7024 					&fl->ifl_m_enqueued, "mbufs allocated");
7025 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_dequeued",
7026 					CTLFLAG_RD,
7027 					&fl->ifl_m_dequeued, "mbufs freed");
7028 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_enqueued",
7029 					CTLFLAG_RD,
7030 					&fl->ifl_cl_enqueued, "clusters allocated");
7031 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_dequeued",
7032 					CTLFLAG_RD,
7033 					&fl->ifl_cl_dequeued, "clusters freed");
7034 #endif
7035 		}
7036 	}
7037 
7038 }
7039 
7040 void
7041 iflib_request_reset(if_ctx_t ctx)
7042 {
7043 
7044 	STATE_LOCK(ctx);
7045 	ctx->ifc_flags |= IFC_DO_RESET;
7046 	STATE_UNLOCK(ctx);
7047 }
7048 
7049 #ifndef __NO_STRICT_ALIGNMENT
7050 static struct mbuf *
7051 iflib_fixup_rx(struct mbuf *m)
7052 {
7053 	struct mbuf *n;
7054 
7055 	if (m->m_len <= (MCLBYTES - ETHER_HDR_LEN)) {
7056 		bcopy(m->m_data, m->m_data + ETHER_HDR_LEN, m->m_len);
7057 		m->m_data += ETHER_HDR_LEN;
7058 		n = m;
7059 	} else {
7060 		MGETHDR(n, M_NOWAIT, MT_DATA);
7061 		if (n == NULL) {
7062 			m_freem(m);
7063 			return (NULL);
7064 		}
7065 		bcopy(m->m_data, n->m_data, ETHER_HDR_LEN);
7066 		m->m_data += ETHER_HDR_LEN;
7067 		m->m_len -= ETHER_HDR_LEN;
7068 		n->m_len = ETHER_HDR_LEN;
7069 		M_MOVE_PKTHDR(n, m);
7070 		n->m_next = m;
7071 	}
7072 	return (n);
7073 }
7074 #endif
7075 
7076 #ifdef DEBUGNET
7077 static void
7078 iflib_debugnet_init(if_t ifp, int *nrxr, int *ncl, int *clsize)
7079 {
7080 	if_ctx_t ctx;
7081 
7082 	ctx = if_getsoftc(ifp);
7083 	CTX_LOCK(ctx);
7084 	*nrxr = NRXQSETS(ctx);
7085 	*ncl = ctx->ifc_rxqs[0].ifr_fl->ifl_size;
7086 	*clsize = ctx->ifc_rxqs[0].ifr_fl->ifl_buf_size;
7087 	CTX_UNLOCK(ctx);
7088 }
7089 
7090 static void
7091 iflib_debugnet_event(if_t ifp, enum debugnet_ev event)
7092 {
7093 	if_ctx_t ctx;
7094 	if_softc_ctx_t scctx;
7095 	iflib_fl_t fl;
7096 	iflib_rxq_t rxq;
7097 	int i, j;
7098 
7099 	ctx = if_getsoftc(ifp);
7100 	scctx = &ctx->ifc_softc_ctx;
7101 
7102 	switch (event) {
7103 	case DEBUGNET_START:
7104 		for (i = 0; i < scctx->isc_nrxqsets; i++) {
7105 			rxq = &ctx->ifc_rxqs[i];
7106 			for (j = 0; j < rxq->ifr_nfl; j++) {
7107 				fl = rxq->ifr_fl;
7108 				fl->ifl_zone = m_getzone(fl->ifl_buf_size);
7109 			}
7110 		}
7111 		iflib_no_tx_batch = 1;
7112 		break;
7113 	default:
7114 		break;
7115 	}
7116 }
7117 
7118 static int
7119 iflib_debugnet_transmit(if_t ifp, struct mbuf *m)
7120 {
7121 	if_ctx_t ctx;
7122 	iflib_txq_t txq;
7123 	int error;
7124 
7125 	ctx = if_getsoftc(ifp);
7126 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
7127 	    IFF_DRV_RUNNING)
7128 		return (EBUSY);
7129 
7130 	txq = &ctx->ifc_txqs[0];
7131 	error = iflib_encap(txq, &m);
7132 	if (error == 0)
7133 		(void)iflib_txd_db_check(txq, true);
7134 	return (error);
7135 }
7136 
7137 static int
7138 iflib_debugnet_poll(if_t ifp, int count)
7139 {
7140 	struct epoch_tracker et;
7141 	if_ctx_t ctx;
7142 	if_softc_ctx_t scctx;
7143 	iflib_txq_t txq;
7144 	int i;
7145 
7146 	ctx = if_getsoftc(ifp);
7147 	scctx = &ctx->ifc_softc_ctx;
7148 
7149 	if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) !=
7150 	    IFF_DRV_RUNNING)
7151 		return (EBUSY);
7152 
7153 	txq = &ctx->ifc_txqs[0];
7154 	(void)iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
7155 
7156 	NET_EPOCH_ENTER(et);
7157 	for (i = 0; i < scctx->isc_nrxqsets; i++)
7158 		(void)iflib_rxeof(&ctx->ifc_rxqs[i], 16 /* XXX */);
7159 	NET_EPOCH_EXIT(et);
7160 	return (0);
7161 }
7162 #endif /* DEBUGNET */
7163