xref: /freebsd-12.1/sys/net/iflib.c (revision 80098bb1)
1 /*-
2  * Copyright (c) 2014-2017, 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/sockio.h>
41 #include <sys/kernel.h>
42 #include <sys/lock.h>
43 #include <sys/mutex.h>
44 #include <sys/module.h>
45 #include <sys/kobj.h>
46 #include <sys/rman.h>
47 #include <sys/sbuf.h>
48 #include <sys/smp.h>
49 #include <sys/socket.h>
50 #include <sys/sysctl.h>
51 #include <sys/syslog.h>
52 #include <sys/taskqueue.h>
53 #include <sys/limits.h>
54 
55 
56 #include <net/if.h>
57 #include <net/if_var.h>
58 #include <net/if_types.h>
59 #include <net/if_media.h>
60 #include <net/bpf.h>
61 #include <net/ethernet.h>
62 #include <net/mp_ring.h>
63 #include <net/vnet.h>
64 
65 #include <netinet/in.h>
66 #include <netinet/in_pcb.h>
67 #include <netinet/tcp_lro.h>
68 #include <netinet/in_systm.h>
69 #include <netinet/if_ether.h>
70 #include <netinet/ip.h>
71 #include <netinet/ip6.h>
72 #include <netinet/tcp.h>
73 #include <netinet/ip_var.h>
74 #include <netinet6/ip6_var.h>
75 
76 #include <machine/bus.h>
77 #include <machine/in_cksum.h>
78 
79 #include <vm/vm.h>
80 #include <vm/pmap.h>
81 
82 #include <dev/led/led.h>
83 #include <dev/pci/pcireg.h>
84 #include <dev/pci/pcivar.h>
85 #include <dev/pci/pci_private.h>
86 
87 #include <net/iflib.h>
88 
89 #include "ifdi_if.h"
90 
91 #if defined(__i386__) || defined(__amd64__)
92 #include <sys/memdesc.h>
93 #include <machine/bus.h>
94 #include <machine/md_var.h>
95 #include <machine/specialreg.h>
96 #include <x86/include/busdma_impl.h>
97 #include <x86/iommu/busdma_dmar.h>
98 #endif
99 
100 #include <sys/bitstring.h>
101 /*
102  * enable accounting of every mbuf as it comes in to and goes out of
103  * iflib's software descriptor references
104  */
105 #define MEMORY_LOGGING 0
106 /*
107  * Enable mbuf vectors for compressing long mbuf chains
108  */
109 
110 /*
111  * NB:
112  * - Prefetching in tx cleaning should perhaps be a tunable. The distance ahead
113  *   we prefetch needs to be determined by the time spent in m_free vis a vis
114  *   the cost of a prefetch. This will of course vary based on the workload:
115  *      - NFLX's m_free path is dominated by vm-based M_EXT manipulation which
116  *        is quite expensive, thus suggesting very little prefetch.
117  *      - small packet forwarding which is just returning a single mbuf to
118  *        UMA will typically be very fast vis a vis the cost of a memory
119  *        access.
120  */
121 
122 
123 /*
124  * File organization:
125  *  - private structures
126  *  - iflib private utility functions
127  *  - ifnet functions
128  *  - vlan registry and other exported functions
129  *  - iflib public core functions
130  *
131  *
132  */
133 static MALLOC_DEFINE(M_IFLIB, "iflib", "ifnet library");
134 
135 struct iflib_txq;
136 typedef struct iflib_txq *iflib_txq_t;
137 struct iflib_rxq;
138 typedef struct iflib_rxq *iflib_rxq_t;
139 struct iflib_fl;
140 typedef struct iflib_fl *iflib_fl_t;
141 
142 struct iflib_ctx;
143 
144 static void iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid);
145 
146 typedef struct iflib_filter_info {
147 	driver_filter_t *ifi_filter;
148 	void *ifi_filter_arg;
149 	struct grouptask *ifi_task;
150 	void *ifi_ctx;
151 } *iflib_filter_info_t;
152 
153 struct iflib_ctx {
154 	KOBJ_FIELDS;
155    /*
156    * Pointer to hardware driver's softc
157    */
158 	void *ifc_softc;
159 	device_t ifc_dev;
160 	if_t ifc_ifp;
161 
162 	cpuset_t ifc_cpus;
163 	if_shared_ctx_t ifc_sctx;
164 	struct if_softc_ctx ifc_softc_ctx;
165 
166 	struct mtx ifc_mtx;
167 
168 	uint16_t ifc_nhwtxqs;
169 	uint16_t ifc_nhwrxqs;
170 
171 	iflib_txq_t ifc_txqs;
172 	iflib_rxq_t ifc_rxqs;
173 	uint32_t ifc_if_flags;
174 	uint32_t ifc_flags;
175 	uint32_t ifc_max_fl_buf_size;
176 	int ifc_in_detach;
177 
178 	int ifc_link_state;
179 	int ifc_link_irq;
180 	int ifc_watchdog_events;
181 	struct cdev *ifc_led_dev;
182 	struct resource *ifc_msix_mem;
183 
184 	struct if_irq ifc_legacy_irq;
185 	struct grouptask ifc_admin_task;
186 	struct grouptask ifc_vflr_task;
187 	struct iflib_filter_info ifc_filter_info;
188 	struct ifmedia	ifc_media;
189 
190 	struct sysctl_oid *ifc_sysctl_node;
191 	uint16_t ifc_sysctl_ntxqs;
192 	uint16_t ifc_sysctl_nrxqs;
193 	uint16_t ifc_sysctl_qs_eq_override;
194 	uint16_t ifc_sysctl_rx_budget;
195 
196 	qidx_t ifc_sysctl_ntxds[8];
197 	qidx_t ifc_sysctl_nrxds[8];
198 	struct if_txrx ifc_txrx;
199 #define isc_txd_encap  ifc_txrx.ift_txd_encap
200 #define isc_txd_flush  ifc_txrx.ift_txd_flush
201 #define isc_txd_credits_update  ifc_txrx.ift_txd_credits_update
202 #define isc_rxd_available ifc_txrx.ift_rxd_available
203 #define isc_rxd_pkt_get ifc_txrx.ift_rxd_pkt_get
204 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
205 #define isc_rxd_flush ifc_txrx.ift_rxd_flush
206 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
207 #define isc_rxd_refill ifc_txrx.ift_rxd_refill
208 #define isc_legacy_intr ifc_txrx.ift_legacy_intr
209 	eventhandler_tag ifc_vlan_attach_event;
210 	eventhandler_tag ifc_vlan_detach_event;
211 	uint8_t ifc_mac[ETHER_ADDR_LEN];
212 	char ifc_mtx_name[16];
213 };
214 
215 
216 void *
217 iflib_get_softc(if_ctx_t ctx)
218 {
219 
220 	return (ctx->ifc_softc);
221 }
222 
223 device_t
224 iflib_get_dev(if_ctx_t ctx)
225 {
226 
227 	return (ctx->ifc_dev);
228 }
229 
230 if_t
231 iflib_get_ifp(if_ctx_t ctx)
232 {
233 
234 	return (ctx->ifc_ifp);
235 }
236 
237 struct ifmedia *
238 iflib_get_media(if_ctx_t ctx)
239 {
240 
241 	return (&ctx->ifc_media);
242 }
243 
244 void
245 iflib_set_mac(if_ctx_t ctx, uint8_t mac[ETHER_ADDR_LEN])
246 {
247 
248 	bcopy(mac, ctx->ifc_mac, ETHER_ADDR_LEN);
249 }
250 
251 if_softc_ctx_t
252 iflib_get_softc_ctx(if_ctx_t ctx)
253 {
254 
255 	return (&ctx->ifc_softc_ctx);
256 }
257 
258 if_shared_ctx_t
259 iflib_get_sctx(if_ctx_t ctx)
260 {
261 
262 	return (ctx->ifc_sctx);
263 }
264 
265 #define IP_ALIGNED(m) ((((uintptr_t)(m)->m_data) & 0x3) == 0x2)
266 #define CACHE_PTR_INCREMENT (CACHE_LINE_SIZE/sizeof(void*))
267 #define CACHE_PTR_NEXT(ptr) ((void *)(((uintptr_t)(ptr)+CACHE_LINE_SIZE-1) & (CACHE_LINE_SIZE-1)))
268 
269 #define LINK_ACTIVE(ctx) ((ctx)->ifc_link_state == LINK_STATE_UP)
270 #define CTX_IS_VF(ctx) ((ctx)->ifc_sctx->isc_flags & IFLIB_IS_VF)
271 
272 #define RX_SW_DESC_MAP_CREATED	(1 << 0)
273 #define TX_SW_DESC_MAP_CREATED	(1 << 1)
274 #define RX_SW_DESC_INUSE        (1 << 3)
275 #define TX_SW_DESC_MAPPED       (1 << 4)
276 
277 #define	M_TOOBIG		M_PROTO1
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 	uint8_t		*ifsd_flags;
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 	struct mbuf    **ifsd_m;           /* pkthdr mbufs */
289 	uint8_t		*ifsd_flags;
290 } if_txsd_vec_t;
291 
292 
293 /* magic number that should be high enough for any hardware */
294 #define IFLIB_MAX_TX_SEGS		128
295 /* bnxt supports 64 with hardware LRO enabled */
296 #define IFLIB_MAX_RX_SEGS		64
297 #define IFLIB_RX_COPY_THRESH		128
298 #define IFLIB_MAX_RX_REFRESH		32
299 /* The minimum descriptors per second before we start coalescing */
300 #define IFLIB_MIN_DESC_SEC		16384
301 #define IFLIB_DEFAULT_TX_UPDATE_FREQ	16
302 #define IFLIB_QUEUE_IDLE		0
303 #define IFLIB_QUEUE_HUNG		1
304 #define IFLIB_QUEUE_WORKING		2
305 /* maximum number of txqs that can share an rx interrupt */
306 #define IFLIB_MAX_TX_SHARED_INTR	4
307 
308 /* this should really scale with ring size - this is a fairly arbitrary value */
309 #define TX_BATCH_SIZE			32
310 
311 #define IFLIB_RESTART_BUDGET		8
312 
313 #define	IFC_LEGACY		0x001
314 #define	IFC_QFLUSH		0x002
315 #define	IFC_MULTISEG		0x004
316 #define	IFC_DMAR		0x008
317 #define	IFC_SC_ALLOCATED	0x010
318 #define	IFC_INIT_DONE		0x020
319 #define	IFC_PREFETCH		0x040
320 #define	IFC_DO_RESET		0x080
321 #define	IFC_CHECK_HUNG		0x100
322 
323 #define CSUM_OFFLOAD		(CSUM_IP_TSO|CSUM_IP6_TSO|CSUM_IP| \
324 				 CSUM_IP_UDP|CSUM_IP_TCP|CSUM_IP_SCTP| \
325 				 CSUM_IP6_UDP|CSUM_IP6_TCP|CSUM_IP6_SCTP)
326 struct iflib_txq {
327 	qidx_t		ift_in_use;
328 	qidx_t		ift_cidx;
329 	qidx_t		ift_cidx_processed;
330 	qidx_t		ift_pidx;
331 	uint8_t		ift_gen;
332 	uint8_t		ift_br_offset;
333 	uint16_t	ift_npending;
334 	uint16_t	ift_db_pending;
335 	uint16_t	ift_rs_pending;
336 	/* implicit pad */
337 	uint8_t		ift_txd_size[8];
338 	uint64_t	ift_processed;
339 	uint64_t	ift_cleaned;
340 	uint64_t	ift_cleaned_prev;
341 #if MEMORY_LOGGING
342 	uint64_t	ift_enqueued;
343 	uint64_t	ift_dequeued;
344 #endif
345 	uint64_t	ift_no_tx_dma_setup;
346 	uint64_t	ift_no_desc_avail;
347 	uint64_t	ift_mbuf_defrag_failed;
348 	uint64_t	ift_mbuf_defrag;
349 	uint64_t	ift_map_failed;
350 	uint64_t	ift_txd_encap_efbig;
351 	uint64_t	ift_pullups;
352 
353 	struct mtx	ift_mtx;
354 	struct mtx	ift_db_mtx;
355 
356 	/* constant values */
357 	if_ctx_t	ift_ctx;
358 	struct ifmp_ring        *ift_br;
359 	struct grouptask	ift_task;
360 	qidx_t		ift_size;
361 	uint16_t	ift_id;
362 	struct callout	ift_timer;
363 
364 	if_txsd_vec_t	ift_sds;
365 	uint8_t		ift_qstatus;
366 	uint8_t		ift_closed;
367 	uint8_t		ift_update_freq;
368 	struct iflib_filter_info ift_filter_info;
369 	bus_dma_tag_t		ift_desc_tag;
370 	bus_dma_tag_t		ift_tso_desc_tag;
371 	iflib_dma_info_t	ift_ifdi;
372 #define MTX_NAME_LEN 16
373 	char                    ift_mtx_name[MTX_NAME_LEN];
374 	char                    ift_db_mtx_name[MTX_NAME_LEN];
375 	bus_dma_segment_t	ift_segs[IFLIB_MAX_TX_SEGS]  __aligned(CACHE_LINE_SIZE);
376 #ifdef IFLIB_DIAGNOSTICS
377 	uint64_t ift_cpu_exec_count[256];
378 #endif
379 } __aligned(CACHE_LINE_SIZE);
380 
381 struct iflib_fl {
382 	qidx_t		ifl_cidx;
383 	qidx_t		ifl_pidx;
384 	qidx_t		ifl_credits;
385 	uint8_t		ifl_gen;
386 	uint8_t		ifl_rxd_size;
387 #if MEMORY_LOGGING
388 	uint64_t	ifl_m_enqueued;
389 	uint64_t	ifl_m_dequeued;
390 	uint64_t	ifl_cl_enqueued;
391 	uint64_t	ifl_cl_dequeued;
392 #endif
393 	/* implicit pad */
394 
395 	bitstr_t 	*ifl_rx_bitmap;
396 	qidx_t		ifl_fragidx;
397 	/* constant */
398 	qidx_t		ifl_size;
399 	uint16_t	ifl_buf_size;
400 	uint16_t	ifl_cltype;
401 	uma_zone_t	ifl_zone;
402 	iflib_rxsd_array_t	ifl_sds;
403 	iflib_rxq_t	ifl_rxq;
404 	uint8_t		ifl_id;
405 	bus_dma_tag_t           ifl_desc_tag;
406 	iflib_dma_info_t	ifl_ifdi;
407 	uint64_t	ifl_bus_addrs[IFLIB_MAX_RX_REFRESH] __aligned(CACHE_LINE_SIZE);
408 	caddr_t		ifl_vm_addrs[IFLIB_MAX_RX_REFRESH];
409 	qidx_t	ifl_rxd_idxs[IFLIB_MAX_RX_REFRESH];
410 }  __aligned(CACHE_LINE_SIZE);
411 
412 static inline qidx_t
413 get_inuse(int size, qidx_t cidx, qidx_t pidx, uint8_t gen)
414 {
415 	qidx_t used;
416 
417 	if (pidx > cidx)
418 		used = pidx - cidx;
419 	else if (pidx < cidx)
420 		used = size - cidx + pidx;
421 	else if (gen == 0 && pidx == cidx)
422 		used = 0;
423 	else if (gen == 1 && pidx == cidx)
424 		used = size;
425 	else
426 		panic("bad state");
427 
428 	return (used);
429 }
430 
431 #define TXQ_AVAIL(txq) (txq->ift_size - get_inuse(txq->ift_size, txq->ift_cidx, txq->ift_pidx, txq->ift_gen))
432 
433 #define IDXDIFF(head, tail, wrap) \
434 	((head) >= (tail) ? (head) - (tail) : (wrap) - (tail) + (head))
435 
436 struct iflib_rxq {
437 	/* If there is a separate completion queue -
438 	 * these are the cq cidx and pidx. Otherwise
439 	 * these are unused.
440 	 */
441 	qidx_t		ifr_size;
442 	qidx_t		ifr_cq_cidx;
443 	qidx_t		ifr_cq_pidx;
444 	uint8_t		ifr_cq_gen;
445 	uint8_t		ifr_fl_offset;
446 
447 	if_ctx_t	ifr_ctx;
448 	iflib_fl_t	ifr_fl;
449 	uint64_t	ifr_rx_irq;
450 	uint16_t	ifr_id;
451 	uint8_t		ifr_lro_enabled;
452 	uint8_t		ifr_nfl;
453 	uint8_t		ifr_ntxqirq;
454 	uint8_t		ifr_txqid[IFLIB_MAX_TX_SHARED_INTR];
455 	struct lro_ctrl			ifr_lc;
456 	struct grouptask        ifr_task;
457 	struct iflib_filter_info ifr_filter_info;
458 	iflib_dma_info_t		ifr_ifdi;
459 
460 	/* dynamically allocate if any drivers need a value substantially larger than this */
461 	struct if_rxd_frag	ifr_frags[IFLIB_MAX_RX_SEGS] __aligned(CACHE_LINE_SIZE);
462 #ifdef IFLIB_DIAGNOSTICS
463 	uint64_t ifr_cpu_exec_count[256];
464 #endif
465 }  __aligned(CACHE_LINE_SIZE);
466 
467 typedef struct if_rxsd {
468 	caddr_t *ifsd_cl;
469 	struct mbuf **ifsd_m;
470 	iflib_fl_t ifsd_fl;
471 	qidx_t ifsd_cidx;
472 } *if_rxsd_t;
473 
474 /* multiple of word size */
475 #ifdef __LP64__
476 #define PKT_INFO_SIZE	6
477 #define RXD_INFO_SIZE	5
478 #define PKT_TYPE uint64_t
479 #else
480 #define PKT_INFO_SIZE	11
481 #define RXD_INFO_SIZE	8
482 #define PKT_TYPE uint32_t
483 #endif
484 #define PKT_LOOP_BOUND  ((PKT_INFO_SIZE/3)*3)
485 #define RXD_LOOP_BOUND  ((RXD_INFO_SIZE/4)*4)
486 
487 typedef struct if_pkt_info_pad {
488 	PKT_TYPE pkt_val[PKT_INFO_SIZE];
489 } *if_pkt_info_pad_t;
490 typedef struct if_rxd_info_pad {
491 	PKT_TYPE rxd_val[RXD_INFO_SIZE];
492 } *if_rxd_info_pad_t;
493 
494 CTASSERT(sizeof(struct if_pkt_info_pad) == sizeof(struct if_pkt_info));
495 CTASSERT(sizeof(struct if_rxd_info_pad) == sizeof(struct if_rxd_info));
496 
497 
498 static inline void
499 pkt_info_zero(if_pkt_info_t pi)
500 {
501 	if_pkt_info_pad_t pi_pad;
502 
503 	pi_pad = (if_pkt_info_pad_t)pi;
504 	pi_pad->pkt_val[0] = 0; pi_pad->pkt_val[1] = 0; pi_pad->pkt_val[2] = 0;
505 	pi_pad->pkt_val[3] = 0; pi_pad->pkt_val[4] = 0; pi_pad->pkt_val[5] = 0;
506 #ifndef __LP64__
507 	pi_pad->pkt_val[6] = 0; pi_pad->pkt_val[7] = 0; pi_pad->pkt_val[8] = 0;
508 	pi_pad->pkt_val[9] = 0; pi_pad->pkt_val[10] = 0;
509 #endif
510 }
511 
512 static inline void
513 rxd_info_zero(if_rxd_info_t ri)
514 {
515 	if_rxd_info_pad_t ri_pad;
516 	int i;
517 
518 	ri_pad = (if_rxd_info_pad_t)ri;
519 	for (i = 0; i < RXD_LOOP_BOUND; i += 4) {
520 		ri_pad->rxd_val[i] = 0;
521 		ri_pad->rxd_val[i+1] = 0;
522 		ri_pad->rxd_val[i+2] = 0;
523 		ri_pad->rxd_val[i+3] = 0;
524 	}
525 #ifdef __LP64__
526 	ri_pad->rxd_val[RXD_INFO_SIZE-1] = 0;
527 #endif
528 }
529 
530 /*
531  * Only allow a single packet to take up most 1/nth of the tx ring
532  */
533 #define MAX_SINGLE_PACKET_FRACTION 12
534 #define IF_BAD_DMA (bus_addr_t)-1
535 
536 #define CTX_ACTIVE(ctx) ((if_getdrvflags((ctx)->ifc_ifp) & IFF_DRV_RUNNING))
537 
538 #define CTX_LOCK_INIT(_sc, _name)  mtx_init(&(_sc)->ifc_mtx, _name, "iflib ctx lock", MTX_DEF)
539 
540 #define CTX_LOCK(ctx) mtx_lock(&(ctx)->ifc_mtx)
541 #define CTX_UNLOCK(ctx) mtx_unlock(&(ctx)->ifc_mtx)
542 #define CTX_LOCK_DESTROY(ctx) mtx_destroy(&(ctx)->ifc_mtx)
543 
544 
545 #define CALLOUT_LOCK(txq)	mtx_lock(&txq->ift_mtx)
546 #define CALLOUT_UNLOCK(txq) 	mtx_unlock(&txq->ift_mtx)
547 
548 
549 /* Our boot-time initialization hook */
550 static int	iflib_module_event_handler(module_t, int, void *);
551 
552 static moduledata_t iflib_moduledata = {
553 	"iflib",
554 	iflib_module_event_handler,
555 	NULL
556 };
557 
558 DECLARE_MODULE(iflib, iflib_moduledata, SI_SUB_INIT_IF, SI_ORDER_ANY);
559 MODULE_VERSION(iflib, 1);
560 
561 MODULE_DEPEND(iflib, pci, 1, 1, 1);
562 MODULE_DEPEND(iflib, ether, 1, 1, 1);
563 
564 TASKQGROUP_DEFINE(if_io_tqg, mp_ncpus, 1);
565 TASKQGROUP_DEFINE(if_config_tqg, 1, 1);
566 
567 #ifndef IFLIB_DEBUG_COUNTERS
568 #ifdef INVARIANTS
569 #define IFLIB_DEBUG_COUNTERS 1
570 #else
571 #define IFLIB_DEBUG_COUNTERS 0
572 #endif /* !INVARIANTS */
573 #endif
574 
575 static SYSCTL_NODE(_net, OID_AUTO, iflib, CTLFLAG_RD, 0,
576                    "iflib driver parameters");
577 
578 /*
579  * XXX need to ensure that this can't accidentally cause the head to be moved backwards
580  */
581 static int iflib_min_tx_latency = 0;
582 SYSCTL_INT(_net_iflib, OID_AUTO, min_tx_latency, CTLFLAG_RW,
583 		   &iflib_min_tx_latency, 0, "minimize transmit latency at the possible expense of throughput");
584 static int iflib_no_tx_batch = 0;
585 SYSCTL_INT(_net_iflib, OID_AUTO, no_tx_batch, CTLFLAG_RW,
586 		   &iflib_no_tx_batch, 0, "minimize transmit latency at the possible expense of throughput");
587 
588 
589 #if IFLIB_DEBUG_COUNTERS
590 
591 static int iflib_tx_seen;
592 static int iflib_tx_sent;
593 static int iflib_tx_encap;
594 static int iflib_rx_allocs;
595 static int iflib_fl_refills;
596 static int iflib_fl_refills_large;
597 static int iflib_tx_frees;
598 
599 SYSCTL_INT(_net_iflib, OID_AUTO, tx_seen, CTLFLAG_RD,
600 		   &iflib_tx_seen, 0, "# tx mbufs seen");
601 SYSCTL_INT(_net_iflib, OID_AUTO, tx_sent, CTLFLAG_RD,
602 		   &iflib_tx_sent, 0, "# tx mbufs sent");
603 SYSCTL_INT(_net_iflib, OID_AUTO, tx_encap, CTLFLAG_RD,
604 		   &iflib_tx_encap, 0, "# tx mbufs encapped");
605 SYSCTL_INT(_net_iflib, OID_AUTO, tx_frees, CTLFLAG_RD,
606 		   &iflib_tx_frees, 0, "# tx frees");
607 SYSCTL_INT(_net_iflib, OID_AUTO, rx_allocs, CTLFLAG_RD,
608 		   &iflib_rx_allocs, 0, "# rx allocations");
609 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills, CTLFLAG_RD,
610 		   &iflib_fl_refills, 0, "# refills");
611 SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills_large, CTLFLAG_RD,
612 		   &iflib_fl_refills_large, 0, "# large refills");
613 
614 
615 static int iflib_txq_drain_flushing;
616 static int iflib_txq_drain_oactive;
617 static int iflib_txq_drain_notready;
618 static int iflib_txq_drain_encapfail;
619 
620 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_flushing, CTLFLAG_RD,
621 		   &iflib_txq_drain_flushing, 0, "# drain flushes");
622 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_oactive, CTLFLAG_RD,
623 		   &iflib_txq_drain_oactive, 0, "# drain oactives");
624 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_notready, CTLFLAG_RD,
625 		   &iflib_txq_drain_notready, 0, "# drain notready");
626 SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_encapfail, CTLFLAG_RD,
627 		   &iflib_txq_drain_encapfail, 0, "# drain encap fails");
628 
629 
630 static int iflib_encap_load_mbuf_fail;
631 static int iflib_encap_pad_mbuf_fail;
632 static int iflib_encap_txq_avail_fail;
633 static int iflib_encap_txd_encap_fail;
634 
635 SYSCTL_INT(_net_iflib, OID_AUTO, encap_load_mbuf_fail, CTLFLAG_RD,
636 		   &iflib_encap_load_mbuf_fail, 0, "# busdma load failures");
637 SYSCTL_INT(_net_iflib, OID_AUTO, encap_pad_mbuf_fail, CTLFLAG_RD,
638 		   &iflib_encap_pad_mbuf_fail, 0, "# runt frame pad failures");
639 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txq_avail_fail, CTLFLAG_RD,
640 		   &iflib_encap_txq_avail_fail, 0, "# txq avail failures");
641 SYSCTL_INT(_net_iflib, OID_AUTO, encap_txd_encap_fail, CTLFLAG_RD,
642 		   &iflib_encap_txd_encap_fail, 0, "# driver encap failures");
643 
644 static int iflib_task_fn_rxs;
645 static int iflib_rx_intr_enables;
646 static int iflib_fast_intrs;
647 static int iflib_intr_link;
648 static int iflib_intr_msix;
649 static int iflib_rx_unavail;
650 static int iflib_rx_ctx_inactive;
651 static int iflib_rx_zero_len;
652 static int iflib_rx_if_input;
653 static int iflib_rx_mbuf_null;
654 static int iflib_rxd_flush;
655 
656 static int iflib_verbose_debug;
657 
658 SYSCTL_INT(_net_iflib, OID_AUTO, intr_link, CTLFLAG_RD,
659 		   &iflib_intr_link, 0, "# intr link calls");
660 SYSCTL_INT(_net_iflib, OID_AUTO, intr_msix, CTLFLAG_RD,
661 		   &iflib_intr_msix, 0, "# intr msix calls");
662 SYSCTL_INT(_net_iflib, OID_AUTO, task_fn_rx, CTLFLAG_RD,
663 		   &iflib_task_fn_rxs, 0, "# task_fn_rx calls");
664 SYSCTL_INT(_net_iflib, OID_AUTO, rx_intr_enables, CTLFLAG_RD,
665 		   &iflib_rx_intr_enables, 0, "# rx intr enables");
666 SYSCTL_INT(_net_iflib, OID_AUTO, fast_intrs, CTLFLAG_RD,
667 		   &iflib_fast_intrs, 0, "# fast_intr calls");
668 SYSCTL_INT(_net_iflib, OID_AUTO, rx_unavail, CTLFLAG_RD,
669 		   &iflib_rx_unavail, 0, "# times rxeof called with no available data");
670 SYSCTL_INT(_net_iflib, OID_AUTO, rx_ctx_inactive, CTLFLAG_RD,
671 		   &iflib_rx_ctx_inactive, 0, "# times rxeof called with inactive context");
672 SYSCTL_INT(_net_iflib, OID_AUTO, rx_zero_len, CTLFLAG_RD,
673 		   &iflib_rx_zero_len, 0, "# times rxeof saw zero len mbuf");
674 SYSCTL_INT(_net_iflib, OID_AUTO, rx_if_input, CTLFLAG_RD,
675 		   &iflib_rx_if_input, 0, "# times rxeof called if_input");
676 SYSCTL_INT(_net_iflib, OID_AUTO, rx_mbuf_null, CTLFLAG_RD,
677 		   &iflib_rx_mbuf_null, 0, "# times rxeof got null mbuf");
678 SYSCTL_INT(_net_iflib, OID_AUTO, rxd_flush, CTLFLAG_RD,
679 	         &iflib_rxd_flush, 0, "# times rxd_flush called");
680 SYSCTL_INT(_net_iflib, OID_AUTO, verbose_debug, CTLFLAG_RW,
681 		   &iflib_verbose_debug, 0, "enable verbose debugging");
682 
683 #define DBG_COUNTER_INC(name) atomic_add_int(&(iflib_ ## name), 1)
684 static void
685 iflib_debug_reset(void)
686 {
687 	iflib_tx_seen = iflib_tx_sent = iflib_tx_encap = iflib_rx_allocs =
688 		iflib_fl_refills = iflib_fl_refills_large = iflib_tx_frees =
689 		iflib_txq_drain_flushing = iflib_txq_drain_oactive =
690 		iflib_txq_drain_notready = iflib_txq_drain_encapfail =
691 		iflib_encap_load_mbuf_fail = iflib_encap_pad_mbuf_fail =
692 		iflib_encap_txq_avail_fail = iflib_encap_txd_encap_fail =
693 		iflib_task_fn_rxs = iflib_rx_intr_enables = iflib_fast_intrs =
694 		iflib_intr_link = iflib_intr_msix = iflib_rx_unavail =
695 		iflib_rx_ctx_inactive = iflib_rx_zero_len = iflib_rx_if_input =
696 		iflib_rx_mbuf_null = iflib_rxd_flush = 0;
697 }
698 
699 #else
700 #define DBG_COUNTER_INC(name)
701 static void iflib_debug_reset(void) {}
702 #endif
703 
704 
705 
706 #define IFLIB_DEBUG 0
707 
708 static void iflib_tx_structures_free(if_ctx_t ctx);
709 static void iflib_rx_structures_free(if_ctx_t ctx);
710 static int iflib_queues_alloc(if_ctx_t ctx);
711 static int iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq);
712 static int iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget);
713 static int iflib_qset_structures_setup(if_ctx_t ctx);
714 static int iflib_msix_init(if_ctx_t ctx);
715 static int iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filterarg, int *rid, char *str);
716 static void iflib_txq_check_drain(iflib_txq_t txq, int budget);
717 static uint32_t iflib_txq_can_drain(struct ifmp_ring *);
718 static int iflib_register(if_ctx_t);
719 static void iflib_init_locked(if_ctx_t ctx);
720 static void iflib_add_device_sysctl_pre(if_ctx_t ctx);
721 static void iflib_add_device_sysctl_post(if_ctx_t ctx);
722 static void iflib_ifmp_purge(iflib_txq_t txq);
723 static void _iflib_pre_assert(if_softc_ctx_t scctx);
724 static void iflib_stop(if_ctx_t ctx);
725 static void iflib_if_init_locked(if_ctx_t ctx);
726 #ifndef __NO_STRICT_ALIGNMENT
727 static struct mbuf * iflib_fixup_rx(struct mbuf *m);
728 #endif
729 
730 #ifdef DEV_NETMAP
731 #include <sys/selinfo.h>
732 #include <net/netmap.h>
733 #include <dev/netmap/netmap_kern.h>
734 
735 MODULE_DEPEND(iflib, netmap, 1, 1, 1);
736 
737 static int netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, uint32_t nm_i, bool init);
738 
739 /*
740  * device-specific sysctl variables:
741  *
742  * iflib_crcstrip: 0: keep CRC in rx frames (default), 1: strip it.
743  *	During regular operations the CRC is stripped, but on some
744  *	hardware reception of frames not multiple of 64 is slower,
745  *	so using crcstrip=0 helps in benchmarks.
746  *
747  * iflib_rx_miss, iflib_rx_miss_bufs:
748  *	count packets that might be missed due to lost interrupts.
749  */
750 SYSCTL_DECL(_dev_netmap);
751 /*
752  * The xl driver by default strips CRCs and we do not override it.
753  */
754 
755 int iflib_crcstrip = 1;
756 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_crcstrip,
757     CTLFLAG_RW, &iflib_crcstrip, 1, "strip CRC on rx frames");
758 
759 int iflib_rx_miss, iflib_rx_miss_bufs;
760 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss,
761     CTLFLAG_RW, &iflib_rx_miss, 0, "potentially missed rx intr");
762 SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss_bufs,
763     CTLFLAG_RW, &iflib_rx_miss_bufs, 0, "potentially missed rx intr bufs");
764 
765 /*
766  * Register/unregister. We are already under netmap lock.
767  * Only called on the first register or the last unregister.
768  */
769 static int
770 iflib_netmap_register(struct netmap_adapter *na, int onoff)
771 {
772 	struct ifnet *ifp = na->ifp;
773 	if_ctx_t ctx = ifp->if_softc;
774 	int status;
775 
776 	CTX_LOCK(ctx);
777 	IFDI_INTR_DISABLE(ctx);
778 
779 	/* Tell the stack that the interface is no longer active */
780 	ifp->if_drv_flags &= ~(IFF_DRV_RUNNING | IFF_DRV_OACTIVE);
781 
782 	if (!CTX_IS_VF(ctx))
783 		IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip);
784 
785 	/* enable or disable flags and callbacks in na and ifp */
786 	if (onoff) {
787 		nm_set_native_flags(na);
788 	} else {
789 		nm_clear_native_flags(na);
790 	}
791 	iflib_stop(ctx);
792 	iflib_init_locked(ctx);
793 	IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip); // XXX why twice ?
794 	status = ifp->if_drv_flags & IFF_DRV_RUNNING ? 0 : 1;
795 	if (status)
796 		nm_clear_native_flags(na);
797 	CTX_UNLOCK(ctx);
798 	return (status);
799 }
800 
801 static int
802 netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, uint32_t nm_i, bool init)
803 {
804 	struct netmap_adapter *na = kring->na;
805 	u_int const lim = kring->nkr_num_slots - 1;
806 	u_int head = kring->rhead;
807 	struct netmap_ring *ring = kring->ring;
808 	bus_dmamap_t *map;
809 	struct if_rxd_update iru;
810 	if_ctx_t ctx = rxq->ifr_ctx;
811 	iflib_fl_t fl = &rxq->ifr_fl[0];
812 	uint32_t refill_pidx, nic_i;
813 
814 	if (nm_i == head && __predict_true(!init))
815 		return 0;
816 	iru_init(&iru, rxq, 0 /* flid */);
817 	map = fl->ifl_sds.ifsd_map;
818 	refill_pidx = netmap_idx_k2n(kring, nm_i);
819 	/*
820 	 * IMPORTANT: we must leave one free slot in the ring,
821 	 * so move head back by one unit
822 	 */
823 	head = nm_prev(head, lim);
824 	while (nm_i != head) {
825 		for (int tmp_pidx = 0; tmp_pidx < IFLIB_MAX_RX_REFRESH && nm_i != head; tmp_pidx++) {
826 			struct netmap_slot *slot = &ring->slot[nm_i];
827 			void *addr = PNMB(na, slot, &fl->ifl_bus_addrs[tmp_pidx]);
828 			uint32_t nic_i_dma = refill_pidx;
829 			nic_i = netmap_idx_k2n(kring, nm_i);
830 
831 			MPASS(tmp_pidx < IFLIB_MAX_RX_REFRESH);
832 
833 			if (addr == NETMAP_BUF_BASE(na)) /* bad buf */
834 			        return netmap_ring_reinit(kring);
835 
836 			fl->ifl_vm_addrs[tmp_pidx] = addr;
837 			if (__predict_false(init) && map) {
838 				netmap_load_map(na, fl->ifl_ifdi->idi_tag, map[nic_i], addr);
839 			} else if (map && (slot->flags & NS_BUF_CHANGED)) {
840 				/* buffer has changed, reload map */
841 				netmap_reload_map(na, fl->ifl_ifdi->idi_tag, map[nic_i], addr);
842 			}
843 			slot->flags &= ~NS_BUF_CHANGED;
844 
845 			nm_i = nm_next(nm_i, lim);
846 			fl->ifl_rxd_idxs[tmp_pidx] = nic_i = nm_next(nic_i, lim);
847 			if (nm_i != head && tmp_pidx < IFLIB_MAX_RX_REFRESH-1)
848 				continue;
849 
850 			iru.iru_pidx = refill_pidx;
851 			iru.iru_count = tmp_pidx+1;
852 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
853 
854 			refill_pidx = nic_i;
855 			if (map == NULL)
856 				continue;
857 
858 			for (int n = 0; n < iru.iru_count; n++) {
859 				bus_dmamap_sync(fl->ifl_ifdi->idi_tag, map[nic_i_dma],
860 						BUS_DMASYNC_PREREAD);
861 				/* XXX - change this to not use the netmap func*/
862 				nic_i_dma = nm_next(nic_i_dma, lim);
863 			}
864 		}
865 	}
866 	kring->nr_hwcur = head;
867 
868 	if (map)
869 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
870 				BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
871 	ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, fl->ifl_id, nic_i);
872 	return (0);
873 }
874 
875 /*
876  * Reconcile kernel and user view of the transmit ring.
877  *
878  * All information is in the kring.
879  * Userspace wants to send packets up to the one before kring->rhead,
880  * kernel knows kring->nr_hwcur is the first unsent packet.
881  *
882  * Here we push packets out (as many as possible), and possibly
883  * reclaim buffers from previously completed transmission.
884  *
885  * The caller (netmap) guarantees that there is only one instance
886  * running at any time. Any interference with other driver
887  * methods should be handled by the individual drivers.
888  */
889 static int
890 iflib_netmap_txsync(struct netmap_kring *kring, int flags)
891 {
892 	struct netmap_adapter *na = kring->na;
893 	struct ifnet *ifp = na->ifp;
894 	struct netmap_ring *ring = kring->ring;
895 	u_int nm_i;	/* index into the netmap ring */
896 	u_int nic_i;	/* index into the NIC ring */
897 	u_int n;
898 	u_int const lim = kring->nkr_num_slots - 1;
899 	u_int const head = kring->rhead;
900 	struct if_pkt_info pi;
901 
902 	/*
903 	 * interrupts on every tx packet are expensive so request
904 	 * them every half ring, or where NS_REPORT is set
905 	 */
906 	u_int report_frequency = kring->nkr_num_slots >> 1;
907 	/* device-specific */
908 	if_ctx_t ctx = ifp->if_softc;
909 	iflib_txq_t txq = &ctx->ifc_txqs[kring->ring_id];
910 
911 	if (txq->ift_sds.ifsd_map)
912 		bus_dmamap_sync(txq->ift_desc_tag, txq->ift_ifdi->idi_map,
913 				BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
914 
915 
916 	/*
917 	 * First part: process new packets to send.
918 	 * nm_i is the current index in the netmap ring,
919 	 * nic_i is the corresponding index in the NIC ring.
920 	 *
921 	 * If we have packets to send (nm_i != head)
922 	 * iterate over the netmap ring, fetch length and update
923 	 * the corresponding slot in the NIC ring. Some drivers also
924 	 * need to update the buffer's physical address in the NIC slot
925 	 * even NS_BUF_CHANGED is not set (PNMB computes the addresses).
926 	 *
927 	 * The netmap_reload_map() calls is especially expensive,
928 	 * even when (as in this case) the tag is 0, so do only
929 	 * when the buffer has actually changed.
930 	 *
931 	 * If possible do not set the report/intr bit on all slots,
932 	 * but only a few times per ring or when NS_REPORT is set.
933 	 *
934 	 * Finally, on 10G and faster drivers, it might be useful
935 	 * to prefetch the next slot and txr entry.
936 	 */
937 
938 	nm_i = netmap_idx_n2k(kring, kring->nr_hwcur);
939 	pkt_info_zero(&pi);
940 	pi.ipi_segs = txq->ift_segs;
941 	pi.ipi_qsidx = kring->ring_id;
942 	if (nm_i != head) {	/* we have new packets to send */
943 		nic_i = netmap_idx_k2n(kring, nm_i);
944 
945 		__builtin_prefetch(&ring->slot[nm_i]);
946 		__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i]);
947 		if (txq->ift_sds.ifsd_map)
948 			__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i]);
949 
950 		for (n = 0; nm_i != head; n++) {
951 			struct netmap_slot *slot = &ring->slot[nm_i];
952 			u_int len = slot->len;
953 			uint64_t paddr;
954 			void *addr = PNMB(na, slot, &paddr);
955 			int flags = (slot->flags & NS_REPORT ||
956 				nic_i == 0 || nic_i == report_frequency) ?
957 				IPI_TX_INTR : 0;
958 
959 			/* device-specific */
960 			pi.ipi_len = len;
961 			pi.ipi_segs[0].ds_addr = paddr;
962 			pi.ipi_segs[0].ds_len = len;
963 			pi.ipi_nsegs = 1;
964 			pi.ipi_ndescs = 0;
965 			pi.ipi_pidx = nic_i;
966 			pi.ipi_flags = flags;
967 
968 			/* Fill the slot in the NIC ring. */
969 			ctx->isc_txd_encap(ctx->ifc_softc, &pi);
970 
971 			/* prefetch for next round */
972 			__builtin_prefetch(&ring->slot[nm_i + 1]);
973 			__builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i + 1]);
974 			if (txq->ift_sds.ifsd_map) {
975 				__builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i + 1]);
976 
977 				NM_CHECK_ADDR_LEN(na, addr, len);
978 
979 				if (slot->flags & NS_BUF_CHANGED) {
980 					/* buffer has changed, reload map */
981 					netmap_reload_map(na, txq->ift_desc_tag, txq->ift_sds.ifsd_map[nic_i], addr);
982 				}
983 				/* make sure changes to the buffer are synced */
984 				bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_sds.ifsd_map[nic_i],
985 						BUS_DMASYNC_PREWRITE);
986 			}
987 			slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED);
988 			nm_i = nm_next(nm_i, lim);
989 			nic_i = nm_next(nic_i, lim);
990 		}
991 		kring->nr_hwcur = head;
992 
993 		/* synchronize the NIC ring */
994 		if (txq->ift_sds.ifsd_map)
995 			bus_dmamap_sync(txq->ift_desc_tag, txq->ift_ifdi->idi_map,
996 						BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
997 
998 		/* (re)start the tx unit up to slot nic_i (excluded) */
999 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, nic_i);
1000 	}
1001 
1002 	/*
1003 	 * Second part: reclaim buffers for completed transmissions.
1004 	 */
1005 	if (iflib_tx_credits_update(ctx, txq)) {
1006 		/* some tx completed, increment avail */
1007 		nic_i = txq->ift_cidx_processed;
1008 		kring->nr_hwtail = nm_prev(netmap_idx_n2k(kring, nic_i), lim);
1009 	}
1010 	return (0);
1011 }
1012 
1013 /*
1014  * Reconcile kernel and user view of the receive ring.
1015  * Same as for the txsync, this routine must be efficient.
1016  * The caller guarantees a single invocations, but races against
1017  * the rest of the driver should be handled here.
1018  *
1019  * On call, kring->rhead is the first packet that userspace wants
1020  * to keep, and kring->rcur is the wakeup point.
1021  * The kernel has previously reported packets up to kring->rtail.
1022  *
1023  * If (flags & NAF_FORCE_READ) also check for incoming packets irrespective
1024  * of whether or not we received an interrupt.
1025  */
1026 static int
1027 iflib_netmap_rxsync(struct netmap_kring *kring, int flags)
1028 {
1029 	struct netmap_adapter *na = kring->na;
1030 	struct netmap_ring *ring = kring->ring;
1031 	uint32_t nm_i;	/* index into the netmap ring */
1032 	uint32_t nic_i;	/* index into the NIC ring */
1033 	u_int i, n;
1034 	u_int const lim = kring->nkr_num_slots - 1;
1035 	u_int const head = netmap_idx_n2k(kring, kring->rhead);
1036 	int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR;
1037 	struct if_rxd_info ri;
1038 
1039 	struct ifnet *ifp = na->ifp;
1040 	if_ctx_t ctx = ifp->if_softc;
1041 	iflib_rxq_t rxq = &ctx->ifc_rxqs[kring->ring_id];
1042 	iflib_fl_t fl = rxq->ifr_fl;
1043 	if (head > lim)
1044 		return netmap_ring_reinit(kring);
1045 
1046 	/* XXX check sync modes */
1047 	for (i = 0, fl = rxq->ifr_fl; i < rxq->ifr_nfl; i++, fl++) {
1048 		if (fl->ifl_sds.ifsd_map == NULL)
1049 			continue;
1050 		bus_dmamap_sync(rxq->ifr_fl[i].ifl_desc_tag, fl->ifl_ifdi->idi_map,
1051 				BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1052 	}
1053 	/*
1054 	 * First part: import newly received packets.
1055 	 *
1056 	 * nm_i is the index of the next free slot in the netmap ring,
1057 	 * nic_i is the index of the next received packet in the NIC ring,
1058 	 * and they may differ in case if_init() has been called while
1059 	 * in netmap mode. For the receive ring we have
1060 	 *
1061 	 *	nic_i = rxr->next_check;
1062 	 *	nm_i = kring->nr_hwtail (previous)
1063 	 * and
1064 	 *	nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1065 	 *
1066 	 * rxr->next_check is set to 0 on a ring reinit
1067 	 */
1068 	if (netmap_no_pendintr || force_update) {
1069 		int crclen = iflib_crcstrip ? 0 : 4;
1070 		int error, avail;
1071 		uint16_t slot_flags = kring->nkr_slot_flags;
1072 
1073 		for (i = 0; i < rxq->ifr_nfl; i++) {
1074 			fl = &rxq->ifr_fl[i];
1075 			nic_i = fl->ifl_cidx;
1076 			nm_i = netmap_idx_n2k(kring, nic_i);
1077 			avail = iflib_rxd_avail(ctx, rxq, nic_i, USHRT_MAX);
1078 			for (n = 0; avail > 0; n++, avail--) {
1079 				rxd_info_zero(&ri);
1080 				ri.iri_frags = rxq->ifr_frags;
1081 				ri.iri_qsidx = kring->ring_id;
1082 				ri.iri_ifp = ctx->ifc_ifp;
1083 				ri.iri_cidx = nic_i;
1084 
1085 				error = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
1086 				ring->slot[nm_i].len = error ? 0 : ri.iri_len - crclen;
1087 				ring->slot[nm_i].flags = slot_flags;
1088 				if (fl->ifl_sds.ifsd_map)
1089 					bus_dmamap_sync(fl->ifl_ifdi->idi_tag,
1090 							fl->ifl_sds.ifsd_map[nic_i], BUS_DMASYNC_POSTREAD);
1091 				nm_i = nm_next(nm_i, lim);
1092 				nic_i = nm_next(nic_i, lim);
1093 			}
1094 			if (n) { /* update the state variables */
1095 				if (netmap_no_pendintr && !force_update) {
1096 					/* diagnostics */
1097 					iflib_rx_miss ++;
1098 					iflib_rx_miss_bufs += n;
1099 				}
1100 				fl->ifl_cidx = nic_i;
1101 				kring->nr_hwtail = netmap_idx_k2n(kring, nm_i);
1102 			}
1103 			kring->nr_kflags &= ~NKR_PENDINTR;
1104 		}
1105 	}
1106 	/*
1107 	 * Second part: skip past packets that userspace has released.
1108 	 * (kring->nr_hwcur to head excluded),
1109 	 * and make the buffers available for reception.
1110 	 * As usual nm_i is the index in the netmap ring,
1111 	 * nic_i is the index in the NIC ring, and
1112 	 * nm_i == (nic_i + kring->nkr_hwofs) % ring_size
1113 	 */
1114 	/* XXX not sure how this will work with multiple free lists */
1115 	nm_i = netmap_idx_n2k(kring, kring->nr_hwcur);
1116 
1117 	return (netmap_fl_refill(rxq, kring, nm_i, false));
1118 }
1119 
1120 static void
1121 iflib_netmap_intr(struct netmap_adapter *na, int onoff)
1122 {
1123 	struct ifnet *ifp = na->ifp;
1124 	if_ctx_t ctx = ifp->if_softc;
1125 
1126 	CTX_LOCK(ctx);
1127 	if (onoff) {
1128 		IFDI_INTR_ENABLE(ctx);
1129 	} else {
1130 		IFDI_INTR_DISABLE(ctx);
1131 	}
1132 	CTX_UNLOCK(ctx);
1133 }
1134 
1135 
1136 static int
1137 iflib_netmap_attach(if_ctx_t ctx)
1138 {
1139 	struct netmap_adapter na;
1140 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1141 
1142 	bzero(&na, sizeof(na));
1143 
1144 	na.ifp = ctx->ifc_ifp;
1145 	na.na_flags = NAF_BDG_MAYSLEEP;
1146 	MPASS(ctx->ifc_softc_ctx.isc_ntxqsets);
1147 	MPASS(ctx->ifc_softc_ctx.isc_nrxqsets);
1148 
1149 	na.num_tx_desc = scctx->isc_ntxd[0];
1150 	na.num_rx_desc = scctx->isc_nrxd[0];
1151 	na.nm_txsync = iflib_netmap_txsync;
1152 	na.nm_rxsync = iflib_netmap_rxsync;
1153 	na.nm_register = iflib_netmap_register;
1154 	na.nm_intr = iflib_netmap_intr;
1155 	na.num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets;
1156 	na.num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets;
1157 	return (netmap_attach(&na));
1158 }
1159 
1160 static void
1161 iflib_netmap_txq_init(if_ctx_t ctx, iflib_txq_t txq)
1162 {
1163 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1164 	struct netmap_slot *slot;
1165 
1166 	slot = netmap_reset(na, NR_TX, txq->ift_id, 0);
1167 	if (slot == NULL)
1168 		return;
1169 	if (txq->ift_sds.ifsd_map == NULL)
1170 		return;
1171 
1172 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxd[0]; i++) {
1173 
1174 		/*
1175 		 * In netmap mode, set the map for the packet buffer.
1176 		 * NOTE: Some drivers (not this one) also need to set
1177 		 * the physical buffer address in the NIC ring.
1178 		 * netmap_idx_n2k() maps a nic index, i, into the corresponding
1179 		 * netmap slot index, si
1180 		 */
1181 		int si = netmap_idx_n2k(&na->tx_rings[txq->ift_id], i);
1182 		netmap_load_map(na, txq->ift_desc_tag, txq->ift_sds.ifsd_map[i], NMB(na, slot + si));
1183 	}
1184 }
1185 
1186 static void
1187 iflib_netmap_rxq_init(if_ctx_t ctx, iflib_rxq_t rxq)
1188 {
1189 	struct netmap_adapter *na = NA(ctx->ifc_ifp);
1190 	struct netmap_kring *kring = &na->rx_rings[rxq->ifr_id];
1191 	struct netmap_slot *slot;
1192 	uint32_t nm_i;
1193 
1194 	slot = netmap_reset(na, NR_RX, rxq->ifr_id, 0);
1195 	if (slot == NULL)
1196 		return;
1197 	nm_i = netmap_idx_n2k(kring, 0);
1198 	netmap_fl_refill(rxq, kring, nm_i, true);
1199 }
1200 
1201 #define iflib_netmap_detach(ifp) netmap_detach(ifp)
1202 
1203 #else
1204 #define iflib_netmap_txq_init(ctx, txq)
1205 #define iflib_netmap_rxq_init(ctx, rxq)
1206 #define iflib_netmap_detach(ifp)
1207 
1208 #define iflib_netmap_attach(ctx) (0)
1209 #define netmap_rx_irq(ifp, qid, budget) (0)
1210 #define netmap_tx_irq(ifp, qid) do {} while (0)
1211 
1212 #endif
1213 
1214 #if defined(__i386__) || defined(__amd64__)
1215 static __inline void
1216 prefetch(void *x)
1217 {
1218 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1219 }
1220 static __inline void
1221 prefetch2cachelines(void *x)
1222 {
1223 	__asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x));
1224 #if (CACHE_LINE_SIZE < 128)
1225 	__asm volatile("prefetcht0 %0" :: "m" (*(((unsigned long *)x)+CACHE_LINE_SIZE/(sizeof(unsigned long)))));
1226 #endif
1227 }
1228 #else
1229 #define prefetch(x)
1230 #define prefetch2cachelines(x)
1231 #endif
1232 
1233 static void
1234 iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid)
1235 {
1236 	iflib_fl_t fl;
1237 
1238 	fl = &rxq->ifr_fl[flid];
1239 	iru->iru_paddrs = fl->ifl_bus_addrs;
1240 	iru->iru_vaddrs = &fl->ifl_vm_addrs[0];
1241 	iru->iru_idxs = fl->ifl_rxd_idxs;
1242 	iru->iru_qsidx = rxq->ifr_id;
1243 	iru->iru_buf_size = fl->ifl_buf_size;
1244 	iru->iru_flidx = fl->ifl_id;
1245 }
1246 
1247 static void
1248 _iflib_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg, int err)
1249 {
1250 	if (err)
1251 		return;
1252 	*(bus_addr_t *) arg = segs[0].ds_addr;
1253 }
1254 
1255 int
1256 iflib_dma_alloc(if_ctx_t ctx, int size, iflib_dma_info_t dma, int mapflags)
1257 {
1258 	int err;
1259 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1260 	device_t dev = ctx->ifc_dev;
1261 
1262 	KASSERT(sctx->isc_q_align != 0, ("alignment value not initialized"));
1263 
1264 	err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1265 				sctx->isc_q_align, 0,	/* alignment, bounds */
1266 				BUS_SPACE_MAXADDR,	/* lowaddr */
1267 				BUS_SPACE_MAXADDR,	/* highaddr */
1268 				NULL, NULL,		/* filter, filterarg */
1269 				size,			/* maxsize */
1270 				1,			/* nsegments */
1271 				size,			/* maxsegsize */
1272 				BUS_DMA_ALLOCNOW,	/* flags */
1273 				NULL,			/* lockfunc */
1274 				NULL,			/* lockarg */
1275 				&dma->idi_tag);
1276 	if (err) {
1277 		device_printf(dev,
1278 		    "%s: bus_dma_tag_create failed: %d\n",
1279 		    __func__, err);
1280 		goto fail_0;
1281 	}
1282 
1283 	err = bus_dmamem_alloc(dma->idi_tag, (void**) &dma->idi_vaddr,
1284 	    BUS_DMA_NOWAIT | BUS_DMA_COHERENT | BUS_DMA_ZERO, &dma->idi_map);
1285 	if (err) {
1286 		device_printf(dev,
1287 		    "%s: bus_dmamem_alloc(%ju) failed: %d\n",
1288 		    __func__, (uintmax_t)size, err);
1289 		goto fail_1;
1290 	}
1291 
1292 	dma->idi_paddr = IF_BAD_DMA;
1293 	err = bus_dmamap_load(dma->idi_tag, dma->idi_map, dma->idi_vaddr,
1294 	    size, _iflib_dmamap_cb, &dma->idi_paddr, mapflags | BUS_DMA_NOWAIT);
1295 	if (err || dma->idi_paddr == IF_BAD_DMA) {
1296 		device_printf(dev,
1297 		    "%s: bus_dmamap_load failed: %d\n",
1298 		    __func__, err);
1299 		goto fail_2;
1300 	}
1301 
1302 	dma->idi_size = size;
1303 	return (0);
1304 
1305 fail_2:
1306 	bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1307 fail_1:
1308 	bus_dma_tag_destroy(dma->idi_tag);
1309 fail_0:
1310 	dma->idi_tag = NULL;
1311 
1312 	return (err);
1313 }
1314 
1315 int
1316 iflib_dma_alloc_multi(if_ctx_t ctx, int *sizes, iflib_dma_info_t *dmalist, int mapflags, int count)
1317 {
1318 	int i, err;
1319 	iflib_dma_info_t *dmaiter;
1320 
1321 	dmaiter = dmalist;
1322 	for (i = 0; i < count; i++, dmaiter++) {
1323 		if ((err = iflib_dma_alloc(ctx, sizes[i], *dmaiter, mapflags)) != 0)
1324 			break;
1325 	}
1326 	if (err)
1327 		iflib_dma_free_multi(dmalist, i);
1328 	return (err);
1329 }
1330 
1331 void
1332 iflib_dma_free(iflib_dma_info_t dma)
1333 {
1334 	if (dma->idi_tag == NULL)
1335 		return;
1336 	if (dma->idi_paddr != IF_BAD_DMA) {
1337 		bus_dmamap_sync(dma->idi_tag, dma->idi_map,
1338 		    BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
1339 		bus_dmamap_unload(dma->idi_tag, dma->idi_map);
1340 		dma->idi_paddr = IF_BAD_DMA;
1341 	}
1342 	if (dma->idi_vaddr != NULL) {
1343 		bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map);
1344 		dma->idi_vaddr = NULL;
1345 	}
1346 	bus_dma_tag_destroy(dma->idi_tag);
1347 	dma->idi_tag = NULL;
1348 }
1349 
1350 void
1351 iflib_dma_free_multi(iflib_dma_info_t *dmalist, int count)
1352 {
1353 	int i;
1354 	iflib_dma_info_t *dmaiter = dmalist;
1355 
1356 	for (i = 0; i < count; i++, dmaiter++)
1357 		iflib_dma_free(*dmaiter);
1358 }
1359 
1360 #ifdef EARLY_AP_STARTUP
1361 static const int iflib_started = 1;
1362 #else
1363 /*
1364  * We used to abuse the smp_started flag to decide if the queues have been
1365  * fully initialized (by late taskqgroup_adjust() calls in a SYSINIT()).
1366  * That gave bad races, since the SYSINIT() runs strictly after smp_started
1367  * is set.  Run a SYSINIT() strictly after that to just set a usable
1368  * completion flag.
1369  */
1370 
1371 static int iflib_started;
1372 
1373 static void
1374 iflib_record_started(void *arg)
1375 {
1376 	iflib_started = 1;
1377 }
1378 
1379 SYSINIT(iflib_record_started, SI_SUB_SMP + 1, SI_ORDER_FIRST,
1380 	iflib_record_started, NULL);
1381 #endif
1382 
1383 static int
1384 iflib_fast_intr(void *arg)
1385 {
1386 	iflib_filter_info_t info = arg;
1387 	struct grouptask *gtask = info->ifi_task;
1388 	if (!iflib_started)
1389 		return (FILTER_HANDLED);
1390 
1391 	DBG_COUNTER_INC(fast_intrs);
1392 	if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED)
1393 		return (FILTER_HANDLED);
1394 
1395 	GROUPTASK_ENQUEUE(gtask);
1396 	return (FILTER_HANDLED);
1397 }
1398 
1399 static int
1400 iflib_fast_intr_rxtx(void *arg)
1401 {
1402 	iflib_filter_info_t info = arg;
1403 	struct grouptask *gtask = info->ifi_task;
1404 	iflib_rxq_t rxq = (iflib_rxq_t)info->ifi_ctx;
1405 	if_ctx_t ctx;
1406 	int i, cidx;
1407 
1408 	if (!iflib_started)
1409 		return (FILTER_HANDLED);
1410 
1411 	DBG_COUNTER_INC(fast_intrs);
1412 	if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED)
1413 		return (FILTER_HANDLED);
1414 
1415 	for (i = 0; i < rxq->ifr_ntxqirq; i++) {
1416 		qidx_t txqid = rxq->ifr_txqid[i];
1417 
1418 		ctx = rxq->ifr_ctx;
1419 
1420 		if (!ctx->isc_txd_credits_update(ctx->ifc_softc, txqid, false)) {
1421 			IFDI_TX_QUEUE_INTR_ENABLE(ctx, txqid);
1422 			continue;
1423 		}
1424 		GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
1425 	}
1426 	if (ctx->ifc_sctx->isc_flags & IFLIB_HAS_RXCQ)
1427 		cidx = rxq->ifr_cq_cidx;
1428 	else
1429 		cidx = rxq->ifr_fl[0].ifl_cidx;
1430 	if (iflib_rxd_avail(ctx, rxq, cidx, 1))
1431 		GROUPTASK_ENQUEUE(gtask);
1432 	else
1433 		IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
1434 	return (FILTER_HANDLED);
1435 }
1436 
1437 
1438 static int
1439 iflib_fast_intr_ctx(void *arg)
1440 {
1441 	iflib_filter_info_t info = arg;
1442 	struct grouptask *gtask = info->ifi_task;
1443 
1444 	if (!iflib_started)
1445 		return (FILTER_HANDLED);
1446 
1447 	DBG_COUNTER_INC(fast_intrs);
1448 	if (info->ifi_filter != NULL && info->ifi_filter(info->ifi_filter_arg) == FILTER_HANDLED)
1449 		return (FILTER_HANDLED);
1450 
1451 	GROUPTASK_ENQUEUE(gtask);
1452 	return (FILTER_HANDLED);
1453 }
1454 
1455 static int
1456 _iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
1457 	driver_filter_t filter, driver_intr_t handler, void *arg,
1458 				 char *name)
1459 {
1460 	int rc, flags;
1461 	struct resource *res;
1462 	void *tag = NULL;
1463 	device_t dev = ctx->ifc_dev;
1464 
1465 	flags = RF_ACTIVE;
1466 	if (ctx->ifc_flags & IFC_LEGACY)
1467 		flags |= RF_SHAREABLE;
1468 	MPASS(rid < 512);
1469 	irq->ii_rid = rid;
1470 	res = bus_alloc_resource_any(dev, SYS_RES_IRQ, &irq->ii_rid, flags);
1471 	if (res == NULL) {
1472 		device_printf(dev,
1473 		    "failed to allocate IRQ for rid %d, name %s.\n", rid, name);
1474 		return (ENOMEM);
1475 	}
1476 	irq->ii_res = res;
1477 	KASSERT(filter == NULL || handler == NULL, ("filter and handler can't both be non-NULL"));
1478 	rc = bus_setup_intr(dev, res, INTR_MPSAFE | INTR_TYPE_NET,
1479 						filter, handler, arg, &tag);
1480 	if (rc != 0) {
1481 		device_printf(dev,
1482 		    "failed to setup interrupt for rid %d, name %s: %d\n",
1483 					  rid, name ? name : "unknown", rc);
1484 		return (rc);
1485 	} else if (name)
1486 		bus_describe_intr(dev, res, tag, "%s", name);
1487 
1488 	irq->ii_tag = tag;
1489 	return (0);
1490 }
1491 
1492 
1493 /*********************************************************************
1494  *
1495  *  Allocate memory for tx_buffer structures. The tx_buffer stores all
1496  *  the information needed to transmit a packet on the wire. This is
1497  *  called only once at attach, setup is done every reset.
1498  *
1499  **********************************************************************/
1500 
1501 static int
1502 iflib_txsd_alloc(iflib_txq_t txq)
1503 {
1504 	if_ctx_t ctx = txq->ift_ctx;
1505 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1506 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1507 	device_t dev = ctx->ifc_dev;
1508 	int err, nsegments, ntsosegments;
1509 
1510 	nsegments = scctx->isc_tx_nsegments;
1511 	ntsosegments = scctx->isc_tx_tso_segments_max;
1512 	MPASS(scctx->isc_ntxd[0] > 0);
1513 	MPASS(scctx->isc_ntxd[txq->ift_br_offset] > 0);
1514 	MPASS(nsegments > 0);
1515 	MPASS(ntsosegments > 0);
1516 	/*
1517 	 * Setup DMA descriptor areas.
1518 	 */
1519 	if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1520 			       1, 0,			/* alignment, bounds */
1521 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1522 			       BUS_SPACE_MAXADDR,	/* highaddr */
1523 			       NULL, NULL,		/* filter, filterarg */
1524 			       sctx->isc_tx_maxsize,		/* maxsize */
1525 			       nsegments,	/* nsegments */
1526 			       sctx->isc_tx_maxsegsize,	/* maxsegsize */
1527 			       0,			/* flags */
1528 			       NULL,			/* lockfunc */
1529 			       NULL,			/* lockfuncarg */
1530 			       &txq->ift_desc_tag))) {
1531 		device_printf(dev,"Unable to allocate TX DMA tag: %d\n", err);
1532 		device_printf(dev,"maxsize: %ju nsegments: %d maxsegsize: %ju\n",
1533 		    (uintmax_t)sctx->isc_tx_maxsize, nsegments, (uintmax_t)sctx->isc_tx_maxsegsize);
1534 		goto fail;
1535 	}
1536 	if ((err = bus_dma_tag_create(bus_get_dma_tag(dev),
1537 			       1, 0,			/* alignment, bounds */
1538 			       BUS_SPACE_MAXADDR,	/* lowaddr */
1539 			       BUS_SPACE_MAXADDR,	/* highaddr */
1540 			       NULL, NULL,		/* filter, filterarg */
1541 			       scctx->isc_tx_tso_size_max,		/* maxsize */
1542 			       ntsosegments,	/* nsegments */
1543 			       scctx->isc_tx_tso_segsize_max,	/* maxsegsize */
1544 			       0,			/* flags */
1545 			       NULL,			/* lockfunc */
1546 			       NULL,			/* lockfuncarg */
1547 			       &txq->ift_tso_desc_tag))) {
1548 		device_printf(dev,"Unable to allocate TX TSO DMA tag: %d\n", err);
1549 
1550 		goto fail;
1551 	}
1552 	if (!(txq->ift_sds.ifsd_flags =
1553 	    (uint8_t *) mallocarray(scctx->isc_ntxd[txq->ift_br_offset],
1554 	    sizeof(uint8_t), M_IFLIB, M_NOWAIT | M_ZERO))) {
1555 		device_printf(dev, "Unable to allocate tx_buffer memory\n");
1556 		err = ENOMEM;
1557 		goto fail;
1558 	}
1559 	if (!(txq->ift_sds.ifsd_m =
1560 	    (struct mbuf **) mallocarray(scctx->isc_ntxd[txq->ift_br_offset],
1561 	    sizeof(struct mbuf *), M_IFLIB, M_NOWAIT | M_ZERO))) {
1562 		device_printf(dev, "Unable to allocate tx_buffer memory\n");
1563 		err = ENOMEM;
1564 		goto fail;
1565 	}
1566 
1567         /* Create the descriptor buffer dma maps */
1568 #if defined(ACPI_DMAR) || (! (defined(__i386__) || defined(__amd64__)))
1569 	if ((ctx->ifc_flags & IFC_DMAR) == 0)
1570 		return (0);
1571 
1572 	if (!(txq->ift_sds.ifsd_map =
1573 	    (bus_dmamap_t *) mallocarray(scctx->isc_ntxd[txq->ift_br_offset],
1574 	    sizeof(bus_dmamap_t), M_IFLIB, M_NOWAIT | M_ZERO))) {
1575 		device_printf(dev, "Unable to allocate tx_buffer map memory\n");
1576 		err = ENOMEM;
1577 		goto fail;
1578 	}
1579 
1580 	for (int i = 0; i < scctx->isc_ntxd[txq->ift_br_offset]; i++) {
1581 		err = bus_dmamap_create(txq->ift_desc_tag, 0, &txq->ift_sds.ifsd_map[i]);
1582 		if (err != 0) {
1583 			device_printf(dev, "Unable to create TX DMA map\n");
1584 			goto fail;
1585 		}
1586 	}
1587 #endif
1588 	return (0);
1589 fail:
1590 	/* We free all, it handles case where we are in the middle */
1591 	iflib_tx_structures_free(ctx);
1592 	return (err);
1593 }
1594 
1595 static void
1596 iflib_txsd_destroy(if_ctx_t ctx, iflib_txq_t txq, int i)
1597 {
1598 	bus_dmamap_t map;
1599 
1600 	map = NULL;
1601 	if (txq->ift_sds.ifsd_map != NULL)
1602 		map = txq->ift_sds.ifsd_map[i];
1603 	if (map != NULL) {
1604 		bus_dmamap_unload(txq->ift_desc_tag, map);
1605 		bus_dmamap_destroy(txq->ift_desc_tag, map);
1606 		txq->ift_sds.ifsd_map[i] = NULL;
1607 	}
1608 }
1609 
1610 static void
1611 iflib_txq_destroy(iflib_txq_t txq)
1612 {
1613 	if_ctx_t ctx = txq->ift_ctx;
1614 
1615 	for (int i = 0; i < txq->ift_size; i++)
1616 		iflib_txsd_destroy(ctx, txq, i);
1617 	if (txq->ift_sds.ifsd_map != NULL) {
1618 		free(txq->ift_sds.ifsd_map, M_IFLIB);
1619 		txq->ift_sds.ifsd_map = NULL;
1620 	}
1621 	if (txq->ift_sds.ifsd_m != NULL) {
1622 		free(txq->ift_sds.ifsd_m, M_IFLIB);
1623 		txq->ift_sds.ifsd_m = NULL;
1624 	}
1625 	if (txq->ift_sds.ifsd_flags != NULL) {
1626 		free(txq->ift_sds.ifsd_flags, M_IFLIB);
1627 		txq->ift_sds.ifsd_flags = NULL;
1628 	}
1629 	if (txq->ift_desc_tag != NULL) {
1630 		bus_dma_tag_destroy(txq->ift_desc_tag);
1631 		txq->ift_desc_tag = NULL;
1632 	}
1633 	if (txq->ift_tso_desc_tag != NULL) {
1634 		bus_dma_tag_destroy(txq->ift_tso_desc_tag);
1635 		txq->ift_tso_desc_tag = NULL;
1636 	}
1637 }
1638 
1639 static void
1640 iflib_txsd_free(if_ctx_t ctx, iflib_txq_t txq, int i)
1641 {
1642 	struct mbuf **mp;
1643 
1644 	mp = &txq->ift_sds.ifsd_m[i];
1645 	if (*mp == NULL)
1646 		return;
1647 
1648 	if (txq->ift_sds.ifsd_map != NULL) {
1649 		bus_dmamap_sync(txq->ift_desc_tag,
1650 				txq->ift_sds.ifsd_map[i],
1651 				BUS_DMASYNC_POSTWRITE);
1652 		bus_dmamap_unload(txq->ift_desc_tag,
1653 				  txq->ift_sds.ifsd_map[i]);
1654 	}
1655 	m_free(*mp);
1656 	DBG_COUNTER_INC(tx_frees);
1657 	*mp = NULL;
1658 }
1659 
1660 static int
1661 iflib_txq_setup(iflib_txq_t txq)
1662 {
1663 	if_ctx_t ctx = txq->ift_ctx;
1664 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1665 	iflib_dma_info_t di;
1666 	int i;
1667 
1668 	/* Set number of descriptors available */
1669 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
1670 	/* XXX make configurable */
1671 	txq->ift_update_freq = IFLIB_DEFAULT_TX_UPDATE_FREQ;
1672 
1673 	/* Reset indices */
1674 	txq->ift_cidx_processed = 0;
1675 	txq->ift_pidx = txq->ift_cidx = txq->ift_npending = 0;
1676 	txq->ift_size = scctx->isc_ntxd[txq->ift_br_offset];
1677 
1678 	for (i = 0, di = txq->ift_ifdi; i < ctx->ifc_nhwtxqs; i++, di++)
1679 		bzero((void *)di->idi_vaddr, di->idi_size);
1680 
1681 	IFDI_TXQ_SETUP(ctx, txq->ift_id);
1682 	for (i = 0, di = txq->ift_ifdi; i < ctx->ifc_nhwtxqs; i++, di++)
1683 		bus_dmamap_sync(di->idi_tag, di->idi_map,
1684 						BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1685 	return (0);
1686 }
1687 
1688 /*********************************************************************
1689  *
1690  *  Allocate memory for rx_buffer structures. Since we use one
1691  *  rx_buffer per received packet, the maximum number of rx_buffer's
1692  *  that we'll need is equal to the number of receive descriptors
1693  *  that we've allocated.
1694  *
1695  **********************************************************************/
1696 static int
1697 iflib_rxsd_alloc(iflib_rxq_t rxq)
1698 {
1699 	if_ctx_t ctx = rxq->ifr_ctx;
1700 	if_shared_ctx_t sctx = ctx->ifc_sctx;
1701 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
1702 	device_t dev = ctx->ifc_dev;
1703 	iflib_fl_t fl;
1704 	int			err;
1705 
1706 	MPASS(scctx->isc_nrxd[0] > 0);
1707 	MPASS(scctx->isc_nrxd[rxq->ifr_fl_offset] > 0);
1708 
1709 	fl = rxq->ifr_fl;
1710 	for (int i = 0; i <  rxq->ifr_nfl; i++, fl++) {
1711 		fl->ifl_size = scctx->isc_nrxd[rxq->ifr_fl_offset]; /* this isn't necessarily the same */
1712 		err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */
1713 					 1, 0,			/* alignment, bounds */
1714 					 BUS_SPACE_MAXADDR,	/* lowaddr */
1715 					 BUS_SPACE_MAXADDR,	/* highaddr */
1716 					 NULL, NULL,		/* filter, filterarg */
1717 					 sctx->isc_rx_maxsize,	/* maxsize */
1718 					 sctx->isc_rx_nsegments,	/* nsegments */
1719 					 sctx->isc_rx_maxsegsize,	/* maxsegsize */
1720 					 0,			/* flags */
1721 					 NULL,			/* lockfunc */
1722 					 NULL,			/* lockarg */
1723 					 &fl->ifl_desc_tag);
1724 		if (err) {
1725 			device_printf(dev, "%s: bus_dma_tag_create failed %d\n",
1726 				__func__, err);
1727 			goto fail;
1728 		}
1729 		if (!(fl->ifl_sds.ifsd_flags =
1730 		      (uint8_t *) mallocarray(scctx->isc_nrxd[rxq->ifr_fl_offset],
1731 		          sizeof(uint8_t), M_IFLIB, M_NOWAIT | M_ZERO))) {
1732 			device_printf(dev, "Unable to allocate tx_buffer memory\n");
1733 			err = ENOMEM;
1734 			goto fail;
1735 		}
1736 		if (!(fl->ifl_sds.ifsd_m =
1737 		      (struct mbuf **) mallocarray(scctx->isc_nrxd[rxq->ifr_fl_offset],
1738 		          sizeof(struct mbuf *), M_IFLIB, M_NOWAIT | M_ZERO))) {
1739 			device_printf(dev, "Unable to allocate tx_buffer memory\n");
1740 			err = ENOMEM;
1741 			goto fail;
1742 		}
1743 		if (!(fl->ifl_sds.ifsd_cl =
1744 		      (caddr_t *) mallocarray(scctx->isc_nrxd[rxq->ifr_fl_offset],
1745 		          sizeof(caddr_t), M_IFLIB, M_NOWAIT | M_ZERO))) {
1746 			device_printf(dev, "Unable to allocate tx_buffer memory\n");
1747 			err = ENOMEM;
1748 			goto fail;
1749 		}
1750 
1751 		/* Create the descriptor buffer dma maps */
1752 #if defined(ACPI_DMAR) || (! (defined(__i386__) || defined(__amd64__)))
1753 		if ((ctx->ifc_flags & IFC_DMAR) == 0)
1754 			continue;
1755 
1756 		if (!(fl->ifl_sds.ifsd_map =
1757 		      (bus_dmamap_t *) mallocarray(scctx->isc_nrxd[rxq->ifr_fl_offset],
1758 		          sizeof(bus_dmamap_t), M_IFLIB, M_NOWAIT | M_ZERO))) {
1759 			device_printf(dev, "Unable to allocate tx_buffer map memory\n");
1760 			err = ENOMEM;
1761 			goto fail;
1762 		}
1763 
1764 		for (int i = 0; i < scctx->isc_nrxd[rxq->ifr_fl_offset]; i++) {
1765 			err = bus_dmamap_create(fl->ifl_desc_tag, 0, &fl->ifl_sds.ifsd_map[i]);
1766 			if (err != 0) {
1767 				device_printf(dev, "Unable to create RX buffer DMA map\n");
1768 				goto fail;
1769 			}
1770 		}
1771 #endif
1772 	}
1773 	return (0);
1774 
1775 fail:
1776 	iflib_rx_structures_free(ctx);
1777 	return (err);
1778 }
1779 
1780 
1781 /*
1782  * Internal service routines
1783  */
1784 
1785 struct rxq_refill_cb_arg {
1786 	int               error;
1787 	bus_dma_segment_t seg;
1788 	int               nseg;
1789 };
1790 
1791 static void
1792 _rxq_refill_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error)
1793 {
1794 	struct rxq_refill_cb_arg *cb_arg = arg;
1795 
1796 	cb_arg->error = error;
1797 	cb_arg->seg = segs[0];
1798 	cb_arg->nseg = nseg;
1799 }
1800 
1801 
1802 #ifdef ACPI_DMAR
1803 #define IS_DMAR(ctx) (ctx->ifc_flags & IFC_DMAR)
1804 #else
1805 #define IS_DMAR(ctx) (0)
1806 #endif
1807 
1808 /**
1809  *	rxq_refill - refill an rxq  free-buffer list
1810  *	@ctx: the iflib context
1811  *	@rxq: the free-list to refill
1812  *	@n: the number of new buffers to allocate
1813  *
1814  *	(Re)populate an rxq free-buffer list with up to @n new packet buffers.
1815  *	The caller must assure that @n does not exceed the queue's capacity.
1816  */
1817 static void
1818 _iflib_fl_refill(if_ctx_t ctx, iflib_fl_t fl, int count)
1819 {
1820 	struct mbuf *m;
1821 	int idx, frag_idx = fl->ifl_fragidx;
1822         int pidx = fl->ifl_pidx;
1823 	caddr_t cl, *sd_cl;
1824 	struct mbuf **sd_m;
1825 	uint8_t *sd_flags;
1826 	struct if_rxd_update iru;
1827 	bus_dmamap_t *sd_map;
1828 	int n, i = 0;
1829 	uint64_t bus_addr;
1830 	int err;
1831 	qidx_t credits;
1832 
1833 	sd_m = fl->ifl_sds.ifsd_m;
1834 	sd_map = fl->ifl_sds.ifsd_map;
1835 	sd_cl = fl->ifl_sds.ifsd_cl;
1836 	sd_flags = fl->ifl_sds.ifsd_flags;
1837 	idx = pidx;
1838 	credits = fl->ifl_credits;
1839 
1840 	n  = count;
1841 	MPASS(n > 0);
1842 	MPASS(credits + n <= fl->ifl_size);
1843 
1844 	if (pidx < fl->ifl_cidx)
1845 		MPASS(pidx + n <= fl->ifl_cidx);
1846 	if (pidx == fl->ifl_cidx && (credits < fl->ifl_size))
1847 		MPASS(fl->ifl_gen == 0);
1848 	if (pidx > fl->ifl_cidx)
1849 		MPASS(n <= fl->ifl_size - pidx + fl->ifl_cidx);
1850 
1851 	DBG_COUNTER_INC(fl_refills);
1852 	if (n > 8)
1853 		DBG_COUNTER_INC(fl_refills_large);
1854 	iru_init(&iru, fl->ifl_rxq, fl->ifl_id);
1855 	while (n--) {
1856 		/*
1857 		 * We allocate an uninitialized mbuf + cluster, mbuf is
1858 		 * initialized after rx.
1859 		 *
1860 		 * If the cluster is still set then we know a minimum sized packet was received
1861 		 */
1862 		bit_ffc_at(fl->ifl_rx_bitmap, frag_idx, fl->ifl_size,  &frag_idx);
1863 		if ((frag_idx < 0) || (frag_idx >= fl->ifl_size))
1864                 	bit_ffc(fl->ifl_rx_bitmap, fl->ifl_size, &frag_idx);
1865 		if ((cl = sd_cl[frag_idx]) == NULL) {
1866                        if ((cl = sd_cl[frag_idx] = m_cljget(NULL, M_NOWAIT, fl->ifl_buf_size)) == NULL)
1867 				break;
1868 #if MEMORY_LOGGING
1869 			fl->ifl_cl_enqueued++;
1870 #endif
1871 		}
1872 		if ((m = m_gethdr(M_NOWAIT, MT_NOINIT)) == NULL) {
1873 			break;
1874 		}
1875 #if MEMORY_LOGGING
1876 		fl->ifl_m_enqueued++;
1877 #endif
1878 
1879 		DBG_COUNTER_INC(rx_allocs);
1880 #if defined(__i386__) || defined(__amd64__)
1881 		if (!IS_DMAR(ctx)) {
1882 			bus_addr = pmap_kextract((vm_offset_t)cl);
1883 		} else
1884 #endif
1885 		{
1886 			struct rxq_refill_cb_arg cb_arg;
1887 			iflib_rxq_t q;
1888 
1889 			cb_arg.error = 0;
1890 			q = fl->ifl_rxq;
1891 			MPASS(sd_map != NULL);
1892 			MPASS(sd_map[frag_idx] != NULL);
1893 			err = bus_dmamap_load(fl->ifl_desc_tag, sd_map[frag_idx],
1894 		         cl, fl->ifl_buf_size, _rxq_refill_cb, &cb_arg, 0);
1895 			bus_dmamap_sync(fl->ifl_desc_tag, sd_map[frag_idx],
1896 					BUS_DMASYNC_PREREAD);
1897 
1898 			if (err != 0 || cb_arg.error) {
1899 				/*
1900 				 * !zone_pack ?
1901 				 */
1902 				if (fl->ifl_zone == zone_pack)
1903 					uma_zfree(fl->ifl_zone, cl);
1904 				m_free(m);
1905 				n = 0;
1906 				goto done;
1907 			}
1908 			bus_addr = cb_arg.seg.ds_addr;
1909 		}
1910                 bit_set(fl->ifl_rx_bitmap, frag_idx);
1911 		sd_flags[frag_idx] |= RX_SW_DESC_INUSE;
1912 
1913 		MPASS(sd_m[frag_idx] == NULL);
1914 		sd_cl[frag_idx] = cl;
1915 		sd_m[frag_idx] = m;
1916 		fl->ifl_rxd_idxs[i] = frag_idx;
1917 		fl->ifl_bus_addrs[i] = bus_addr;
1918 		fl->ifl_vm_addrs[i] = cl;
1919 		credits++;
1920 		i++;
1921 		MPASS(credits <= fl->ifl_size);
1922 		if (++idx == fl->ifl_size) {
1923 			fl->ifl_gen = 1;
1924 			idx = 0;
1925 		}
1926 		if (n == 0 || i == IFLIB_MAX_RX_REFRESH) {
1927 			iru.iru_pidx = pidx;
1928 			iru.iru_count = i;
1929 			ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
1930 			i = 0;
1931 			pidx = idx;
1932 			fl->ifl_pidx = idx;
1933 			fl->ifl_credits = credits;
1934 		}
1935 
1936 	}
1937 done:
1938 	if (i) {
1939 		iru.iru_pidx = pidx;
1940 		iru.iru_count = i;
1941 		ctx->isc_rxd_refill(ctx->ifc_softc, &iru);
1942 		fl->ifl_pidx = idx;
1943 		fl->ifl_credits = credits;
1944 	}
1945 	DBG_COUNTER_INC(rxd_flush);
1946 	if (fl->ifl_pidx == 0)
1947 		pidx = fl->ifl_size - 1;
1948 	else
1949 		pidx = fl->ifl_pidx - 1;
1950 
1951 	if (sd_map)
1952 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
1953 				BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
1954 	ctx->isc_rxd_flush(ctx->ifc_softc, fl->ifl_rxq->ifr_id, fl->ifl_id, pidx);
1955 	fl->ifl_fragidx = frag_idx;
1956 }
1957 
1958 static __inline void
1959 __iflib_fl_refill_lt(if_ctx_t ctx, iflib_fl_t fl, int max)
1960 {
1961 	/* we avoid allowing pidx to catch up with cidx as it confuses ixl */
1962 	int32_t reclaimable = fl->ifl_size - fl->ifl_credits - 1;
1963 #ifdef INVARIANTS
1964 	int32_t delta = fl->ifl_size - get_inuse(fl->ifl_size, fl->ifl_cidx, fl->ifl_pidx, fl->ifl_gen) - 1;
1965 #endif
1966 
1967 	MPASS(fl->ifl_credits <= fl->ifl_size);
1968 	MPASS(reclaimable == delta);
1969 
1970 	if (reclaimable > 0)
1971 		_iflib_fl_refill(ctx, fl, min(max, reclaimable));
1972 }
1973 
1974 static void
1975 iflib_fl_bufs_free(iflib_fl_t fl)
1976 {
1977 	iflib_dma_info_t idi = fl->ifl_ifdi;
1978 	uint32_t i;
1979 
1980 	for (i = 0; i < fl->ifl_size; i++) {
1981 		struct mbuf **sd_m = &fl->ifl_sds.ifsd_m[i];
1982 		uint8_t *sd_flags = &fl->ifl_sds.ifsd_flags[i];
1983 		caddr_t *sd_cl = &fl->ifl_sds.ifsd_cl[i];
1984 
1985 		if (*sd_flags & RX_SW_DESC_INUSE) {
1986 			if (fl->ifl_sds.ifsd_map != NULL) {
1987 				bus_dmamap_t sd_map = fl->ifl_sds.ifsd_map[i];
1988 				bus_dmamap_unload(fl->ifl_desc_tag, sd_map);
1989 				bus_dmamap_destroy(fl->ifl_desc_tag, sd_map);
1990 			}
1991 			if (*sd_m != NULL) {
1992 				m_init(*sd_m, M_NOWAIT, MT_DATA, 0);
1993 				uma_zfree(zone_mbuf, *sd_m);
1994 			}
1995 			if (*sd_cl != NULL)
1996 				uma_zfree(fl->ifl_zone, *sd_cl);
1997 			*sd_flags = 0;
1998 		} else {
1999 			MPASS(*sd_cl == NULL);
2000 			MPASS(*sd_m == NULL);
2001 		}
2002 #if MEMORY_LOGGING
2003 		fl->ifl_m_dequeued++;
2004 		fl->ifl_cl_dequeued++;
2005 #endif
2006 		*sd_cl = NULL;
2007 		*sd_m = NULL;
2008 	}
2009 #ifdef INVARIANTS
2010 	for (i = 0; i < fl->ifl_size; i++) {
2011 		MPASS(fl->ifl_sds.ifsd_flags[i] == 0);
2012 		MPASS(fl->ifl_sds.ifsd_cl[i] == NULL);
2013 		MPASS(fl->ifl_sds.ifsd_m[i] == NULL);
2014 	}
2015 #endif
2016 	/*
2017 	 * Reset free list values
2018 	 */
2019 	fl->ifl_credits = fl->ifl_cidx = fl->ifl_pidx = fl->ifl_gen = fl->ifl_fragidx = 0;
2020 	bzero(idi->idi_vaddr, idi->idi_size);
2021 }
2022 
2023 /*********************************************************************
2024  *
2025  *  Initialize a receive ring and its buffers.
2026  *
2027  **********************************************************************/
2028 static int
2029 iflib_fl_setup(iflib_fl_t fl)
2030 {
2031 	iflib_rxq_t rxq = fl->ifl_rxq;
2032 	if_ctx_t ctx = rxq->ifr_ctx;
2033 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2034 
2035 	bit_nclear(fl->ifl_rx_bitmap, 0, fl->ifl_size - 1);
2036 	/*
2037 	** Free current RX buffer structs and their mbufs
2038 	*/
2039 	iflib_fl_bufs_free(fl);
2040 	/* Now replenish the mbufs */
2041 	MPASS(fl->ifl_credits == 0);
2042 	/*
2043 	 * XXX don't set the max_frame_size to larger
2044 	 * than the hardware can handle
2045 	 */
2046 	if (sctx->isc_max_frame_size <= 2048)
2047 		fl->ifl_buf_size = MCLBYTES;
2048 #ifndef CONTIGMALLOC_WORKS
2049 	else
2050 		fl->ifl_buf_size = MJUMPAGESIZE;
2051 #else
2052 	else if (sctx->isc_max_frame_size <= 4096)
2053 		fl->ifl_buf_size = MJUMPAGESIZE;
2054 	else if (sctx->isc_max_frame_size <= 9216)
2055 		fl->ifl_buf_size = MJUM9BYTES;
2056 	else
2057 		fl->ifl_buf_size = MJUM16BYTES;
2058 #endif
2059 	if (fl->ifl_buf_size > ctx->ifc_max_fl_buf_size)
2060 		ctx->ifc_max_fl_buf_size = fl->ifl_buf_size;
2061 	fl->ifl_cltype = m_gettype(fl->ifl_buf_size);
2062 	fl->ifl_zone = m_getzone(fl->ifl_buf_size);
2063 
2064 
2065 	/* avoid pre-allocating zillions of clusters to an idle card
2066 	 * potentially speeding up attach
2067 	 */
2068 	_iflib_fl_refill(ctx, fl, min(128, fl->ifl_size));
2069 	MPASS(min(128, fl->ifl_size) == fl->ifl_credits);
2070 	if (min(128, fl->ifl_size) != fl->ifl_credits)
2071 		return (ENOBUFS);
2072 	/*
2073 	 * handle failure
2074 	 */
2075 	MPASS(rxq != NULL);
2076 	MPASS(fl->ifl_ifdi != NULL);
2077 	bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2078 	    BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2079 	return (0);
2080 }
2081 
2082 /*********************************************************************
2083  *
2084  *  Free receive ring data structures
2085  *
2086  **********************************************************************/
2087 static void
2088 iflib_rx_sds_free(iflib_rxq_t rxq)
2089 {
2090 	iflib_fl_t fl;
2091 	int i;
2092 
2093 	if (rxq->ifr_fl != NULL) {
2094 		for (i = 0; i < rxq->ifr_nfl; i++) {
2095 			fl = &rxq->ifr_fl[i];
2096 			if (fl->ifl_desc_tag != NULL) {
2097 				bus_dma_tag_destroy(fl->ifl_desc_tag);
2098 				fl->ifl_desc_tag = NULL;
2099 			}
2100 			free(fl->ifl_sds.ifsd_m, M_IFLIB);
2101 			free(fl->ifl_sds.ifsd_cl, M_IFLIB);
2102 			/* XXX destroy maps first */
2103 			free(fl->ifl_sds.ifsd_map, M_IFLIB);
2104 			fl->ifl_sds.ifsd_m = NULL;
2105 			fl->ifl_sds.ifsd_cl = NULL;
2106 			fl->ifl_sds.ifsd_map = NULL;
2107 		}
2108 		free(rxq->ifr_fl, M_IFLIB);
2109 		rxq->ifr_fl = NULL;
2110 		rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
2111 	}
2112 }
2113 
2114 /*
2115  * MI independent logic
2116  *
2117  */
2118 static void
2119 iflib_timer(void *arg)
2120 {
2121 	iflib_txq_t txq = arg;
2122 	if_ctx_t ctx = txq->ift_ctx;
2123 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2124 
2125 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
2126 		return;
2127 	/*
2128 	** Check on the state of the TX queue(s), this
2129 	** can be done without the lock because its RO
2130 	** and the HUNG state will be static if set.
2131 	*/
2132 	IFDI_TIMER(ctx, txq->ift_id);
2133 	if ((txq->ift_qstatus == IFLIB_QUEUE_HUNG) &&
2134 	    ((txq->ift_cleaned_prev == txq->ift_cleaned) ||
2135 	     (sctx->isc_pause_frames == 0)))
2136 		goto hung;
2137 
2138 	if (ifmp_ring_is_stalled(txq->ift_br))
2139 		txq->ift_qstatus = IFLIB_QUEUE_HUNG;
2140 	txq->ift_cleaned_prev = txq->ift_cleaned;
2141 	/* handle any laggards */
2142 	if (txq->ift_db_pending)
2143 		GROUPTASK_ENQUEUE(&txq->ift_task);
2144 
2145 	sctx->isc_pause_frames = 0;
2146 	if (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)
2147 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq, txq->ift_timer.c_cpu);
2148 	return;
2149 hung:
2150 	CTX_LOCK(ctx);
2151 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2152 	device_printf(ctx->ifc_dev,  "TX(%d) desc avail = %d, pidx = %d\n",
2153 				  txq->ift_id, TXQ_AVAIL(txq), txq->ift_pidx);
2154 
2155 	IFDI_WATCHDOG_RESET(ctx);
2156 	ctx->ifc_watchdog_events++;
2157 
2158 	ctx->ifc_flags |= IFC_DO_RESET;
2159 	iflib_admin_intr_deferred(ctx);
2160 	CTX_UNLOCK(ctx);
2161 }
2162 
2163 static void
2164 iflib_init_locked(if_ctx_t ctx)
2165 {
2166 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
2167 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2168 	if_t ifp = ctx->ifc_ifp;
2169 	iflib_fl_t fl;
2170 	iflib_txq_t txq;
2171 	iflib_rxq_t rxq;
2172 	int i, j, tx_ip_csum_flags, tx_ip6_csum_flags;
2173 
2174 
2175 	if_setdrvflagbits(ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2176 	IFDI_INTR_DISABLE(ctx);
2177 
2178 	tx_ip_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP | CSUM_SCTP);
2179 	tx_ip6_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_UDP | CSUM_IP6_SCTP);
2180 	/* Set hardware offload abilities */
2181 	if_clearhwassist(ifp);
2182 	if (if_getcapenable(ifp) & IFCAP_TXCSUM)
2183 		if_sethwassistbits(ifp, tx_ip_csum_flags, 0);
2184 	if (if_getcapenable(ifp) & IFCAP_TXCSUM_IPV6)
2185 		if_sethwassistbits(ifp,  tx_ip6_csum_flags, 0);
2186 	if (if_getcapenable(ifp) & IFCAP_TSO4)
2187 		if_sethwassistbits(ifp, CSUM_IP_TSO, 0);
2188 	if (if_getcapenable(ifp) & IFCAP_TSO6)
2189 		if_sethwassistbits(ifp, CSUM_IP6_TSO, 0);
2190 
2191 	for (i = 0, txq = ctx->ifc_txqs; i < sctx->isc_ntxqsets; i++, txq++) {
2192 		CALLOUT_LOCK(txq);
2193 		callout_stop(&txq->ift_timer);
2194 		CALLOUT_UNLOCK(txq);
2195 		iflib_netmap_txq_init(ctx, txq);
2196 	}
2197 #ifdef INVARIANTS
2198 	i = if_getdrvflags(ifp);
2199 #endif
2200 	IFDI_INIT(ctx);
2201 	MPASS(if_getdrvflags(ifp) == i);
2202 	for (i = 0, rxq = ctx->ifc_rxqs; i < sctx->isc_nrxqsets; i++, rxq++) {
2203 		/* XXX this should really be done on a per-queue basis */
2204 		if (if_getcapenable(ifp) & IFCAP_NETMAP) {
2205 			MPASS(rxq->ifr_id == i);
2206 			iflib_netmap_rxq_init(ctx, rxq);
2207 			continue;
2208 		}
2209 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
2210 			if (iflib_fl_setup(fl)) {
2211 				device_printf(ctx->ifc_dev, "freelist setup failed - check cluster settings\n");
2212 				goto done;
2213 			}
2214 		}
2215 	}
2216 	done:
2217 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_RUNNING, IFF_DRV_OACTIVE);
2218 	IFDI_INTR_ENABLE(ctx);
2219 	txq = ctx->ifc_txqs;
2220 	for (i = 0; i < sctx->isc_ntxqsets; i++, txq++)
2221 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq,
2222 			txq->ift_timer.c_cpu);
2223 }
2224 
2225 static int
2226 iflib_media_change(if_t ifp)
2227 {
2228 	if_ctx_t ctx = if_getsoftc(ifp);
2229 	int err;
2230 
2231 	CTX_LOCK(ctx);
2232 	if ((err = IFDI_MEDIA_CHANGE(ctx)) == 0)
2233 		iflib_init_locked(ctx);
2234 	CTX_UNLOCK(ctx);
2235 	return (err);
2236 }
2237 
2238 static void
2239 iflib_media_status(if_t ifp, struct ifmediareq *ifmr)
2240 {
2241 	if_ctx_t ctx = if_getsoftc(ifp);
2242 
2243 	CTX_LOCK(ctx);
2244 	IFDI_UPDATE_ADMIN_STATUS(ctx);
2245 	IFDI_MEDIA_STATUS(ctx, ifmr);
2246 	CTX_UNLOCK(ctx);
2247 }
2248 
2249 static void
2250 iflib_stop(if_ctx_t ctx)
2251 {
2252 	iflib_txq_t txq = ctx->ifc_txqs;
2253 	iflib_rxq_t rxq = ctx->ifc_rxqs;
2254 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2255 	iflib_dma_info_t di;
2256 	iflib_fl_t fl;
2257 	int i, j;
2258 
2259 	/* Tell the stack that the interface is no longer active */
2260 	if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING);
2261 
2262 	IFDI_INTR_DISABLE(ctx);
2263 	DELAY(1000);
2264 	IFDI_STOP(ctx);
2265 	DELAY(1000);
2266 
2267 	iflib_debug_reset();
2268 	/* Wait for current tx queue users to exit to disarm watchdog timer. */
2269 	for (i = 0; i < scctx->isc_ntxqsets; i++, txq++) {
2270 		/* make sure all transmitters have completed before proceeding XXX */
2271 
2272 		/* clean any enqueued buffers */
2273 		iflib_ifmp_purge(txq);
2274 		/* Free any existing tx buffers. */
2275 		for (j = 0; j < txq->ift_size; j++) {
2276 			iflib_txsd_free(ctx, txq, j);
2277 		}
2278 		txq->ift_processed = txq->ift_cleaned = txq->ift_cidx_processed = 0;
2279 		txq->ift_in_use = txq->ift_gen = txq->ift_cidx = txq->ift_pidx = txq->ift_no_desc_avail = 0;
2280 		txq->ift_closed = txq->ift_mbuf_defrag = txq->ift_mbuf_defrag_failed = 0;
2281 		txq->ift_no_tx_dma_setup = txq->ift_txd_encap_efbig = txq->ift_map_failed = 0;
2282 		txq->ift_pullups = 0;
2283 		ifmp_ring_reset_stats(txq->ift_br);
2284 		for (j = 0, di = txq->ift_ifdi; j < ctx->ifc_nhwtxqs; j++, di++)
2285 			bzero((void *)di->idi_vaddr, di->idi_size);
2286 	}
2287 	for (i = 0; i < scctx->isc_nrxqsets; i++, rxq++) {
2288 		/* make sure all transmitters have completed before proceeding XXX */
2289 
2290 		for (j = 0, di = txq->ift_ifdi; j < ctx->ifc_nhwrxqs; j++, di++)
2291 			bzero((void *)di->idi_vaddr, di->idi_size);
2292 		/* also resets the free lists pidx/cidx */
2293 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
2294 			iflib_fl_bufs_free(fl);
2295 	}
2296 }
2297 
2298 static inline caddr_t
2299 calc_next_rxd(iflib_fl_t fl, int cidx)
2300 {
2301 	qidx_t size;
2302 	int nrxd;
2303 	caddr_t start, end, cur, next;
2304 
2305 	nrxd = fl->ifl_size;
2306 	size = fl->ifl_rxd_size;
2307 	start = fl->ifl_ifdi->idi_vaddr;
2308 
2309 	if (__predict_false(size == 0))
2310 		return (start);
2311 	cur = start + size*cidx;
2312 	end = start + size*nrxd;
2313 	next = CACHE_PTR_NEXT(cur);
2314 	return (next < end ? next : start);
2315 }
2316 
2317 static inline void
2318 prefetch_pkts(iflib_fl_t fl, int cidx)
2319 {
2320 	int nextptr;
2321 	int nrxd = fl->ifl_size;
2322 	caddr_t next_rxd;
2323 
2324 
2325 	nextptr = (cidx + CACHE_PTR_INCREMENT) & (nrxd-1);
2326 	prefetch(&fl->ifl_sds.ifsd_m[nextptr]);
2327 	prefetch(&fl->ifl_sds.ifsd_cl[nextptr]);
2328 	next_rxd = calc_next_rxd(fl, cidx);
2329 	prefetch(next_rxd);
2330 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 1) & (nrxd-1)]);
2331 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 2) & (nrxd-1)]);
2332 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 3) & (nrxd-1)]);
2333 	prefetch(fl->ifl_sds.ifsd_m[(cidx + 4) & (nrxd-1)]);
2334 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 1) & (nrxd-1)]);
2335 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 2) & (nrxd-1)]);
2336 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 3) & (nrxd-1)]);
2337 	prefetch(fl->ifl_sds.ifsd_cl[(cidx + 4) & (nrxd-1)]);
2338 }
2339 
2340 static void
2341 rxd_frag_to_sd(iflib_rxq_t rxq, if_rxd_frag_t irf, int unload, if_rxsd_t sd)
2342 {
2343 	int flid, cidx;
2344 	bus_dmamap_t map;
2345 	iflib_fl_t fl;
2346 	iflib_dma_info_t di;
2347 	int next;
2348 
2349 	map = NULL;
2350 	flid = irf->irf_flid;
2351 	cidx = irf->irf_idx;
2352 	fl = &rxq->ifr_fl[flid];
2353 	sd->ifsd_fl = fl;
2354 	sd->ifsd_cidx = cidx;
2355 	sd->ifsd_m = &fl->ifl_sds.ifsd_m[cidx];
2356 	sd->ifsd_cl = &fl->ifl_sds.ifsd_cl[cidx];
2357 	fl->ifl_credits--;
2358 #if MEMORY_LOGGING
2359 	fl->ifl_m_dequeued++;
2360 #endif
2361 	if (rxq->ifr_ctx->ifc_flags & IFC_PREFETCH)
2362 		prefetch_pkts(fl, cidx);
2363 	if (fl->ifl_sds.ifsd_map != NULL) {
2364 		next = (cidx + CACHE_PTR_INCREMENT) & (fl->ifl_size-1);
2365 		prefetch(&fl->ifl_sds.ifsd_map[next]);
2366 		map = fl->ifl_sds.ifsd_map[cidx];
2367 		di = fl->ifl_ifdi;
2368 		next = (cidx + CACHE_LINE_SIZE) & (fl->ifl_size-1);
2369 		prefetch(&fl->ifl_sds.ifsd_flags[next]);
2370 		bus_dmamap_sync(di->idi_tag, di->idi_map,
2371 				BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE);
2372 
2373 	/* not valid assert if bxe really does SGE from non-contiguous elements */
2374 		MPASS(fl->ifl_cidx == cidx);
2375 		if (unload)
2376 			bus_dmamap_unload(fl->ifl_desc_tag, map);
2377 	}
2378 	fl->ifl_cidx = (fl->ifl_cidx + 1) & (fl->ifl_size-1);
2379 	if (__predict_false(fl->ifl_cidx == 0))
2380 		fl->ifl_gen = 0;
2381 	if (map != NULL)
2382 		bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map,
2383 			BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
2384         bit_clear(fl->ifl_rx_bitmap, cidx);
2385 }
2386 
2387 static struct mbuf *
2388 assemble_segments(iflib_rxq_t rxq, if_rxd_info_t ri, if_rxsd_t sd)
2389 {
2390 	int i, padlen , flags;
2391 	struct mbuf *m, *mh, *mt;
2392 	caddr_t cl;
2393 
2394 	i = 0;
2395 	mh = NULL;
2396 	do {
2397 		rxd_frag_to_sd(rxq, &ri->iri_frags[i], TRUE, sd);
2398 
2399 		MPASS(*sd->ifsd_cl != NULL);
2400 		MPASS(*sd->ifsd_m != NULL);
2401 
2402 		/* Don't include zero-length frags */
2403 		if (ri->iri_frags[i].irf_len == 0) {
2404 			/* XXX we can save the cluster here, but not the mbuf */
2405 			m_init(*sd->ifsd_m, M_NOWAIT, MT_DATA, 0);
2406 			m_free(*sd->ifsd_m);
2407 			*sd->ifsd_m = NULL;
2408 			continue;
2409 		}
2410 		m = *sd->ifsd_m;
2411 		*sd->ifsd_m = NULL;
2412 		if (mh == NULL) {
2413 			flags = M_PKTHDR|M_EXT;
2414 			mh = mt = m;
2415 			padlen = ri->iri_pad;
2416 		} else {
2417 			flags = M_EXT;
2418 			mt->m_next = m;
2419 			mt = m;
2420 			/* assuming padding is only on the first fragment */
2421 			padlen = 0;
2422 		}
2423 		cl = *sd->ifsd_cl;
2424 		*sd->ifsd_cl = NULL;
2425 
2426 		/* Can these two be made one ? */
2427 		m_init(m, M_NOWAIT, MT_DATA, flags);
2428 		m_cljset(m, cl, sd->ifsd_fl->ifl_cltype);
2429 		/*
2430 		 * These must follow m_init and m_cljset
2431 		 */
2432 		m->m_data += padlen;
2433 		ri->iri_len -= padlen;
2434 		m->m_len = ri->iri_frags[i].irf_len;
2435 	} while (++i < ri->iri_nfrags);
2436 
2437 	return (mh);
2438 }
2439 
2440 /*
2441  * Process one software descriptor
2442  */
2443 static struct mbuf *
2444 iflib_rxd_pkt_get(iflib_rxq_t rxq, if_rxd_info_t ri)
2445 {
2446 	struct if_rxsd sd;
2447 	struct mbuf *m;
2448 
2449 	/* should I merge this back in now that the two paths are basically duplicated? */
2450 	if (ri->iri_nfrags == 1 &&
2451 	    ri->iri_frags[0].irf_len <= IFLIB_RX_COPY_THRESH) {
2452 		rxd_frag_to_sd(rxq, &ri->iri_frags[0], FALSE, &sd);
2453 		m = *sd.ifsd_m;
2454 		*sd.ifsd_m = NULL;
2455 		m_init(m, M_NOWAIT, MT_DATA, M_PKTHDR);
2456 #ifndef __NO_STRICT_ALIGNMENT
2457 		if (!IP_ALIGNED(m))
2458 			m->m_data += 2;
2459 #endif
2460 		memcpy(m->m_data, *sd.ifsd_cl, ri->iri_len);
2461 		m->m_len = ri->iri_frags[0].irf_len;
2462        } else {
2463 		m = assemble_segments(rxq, ri, &sd);
2464 	}
2465 	m->m_pkthdr.len = ri->iri_len;
2466 	m->m_pkthdr.rcvif = ri->iri_ifp;
2467 	m->m_flags |= ri->iri_flags;
2468 	m->m_pkthdr.ether_vtag = ri->iri_vtag;
2469 	m->m_pkthdr.flowid = ri->iri_flowid;
2470 	M_HASHTYPE_SET(m, ri->iri_rsstype);
2471 	m->m_pkthdr.csum_flags = ri->iri_csum_flags;
2472 	m->m_pkthdr.csum_data = ri->iri_csum_data;
2473 	return (m);
2474 }
2475 
2476 #if defined(INET6) || defined(INET)
2477 static void
2478 iflib_get_ip_forwarding(struct lro_ctrl *lc, bool *v4, bool *v6)
2479 {
2480 	CURVNET_SET(lc->ifp->if_vnet);
2481 #if defined(INET6)
2482 	*v6 = VNET(ip6_forwarding);
2483 #endif
2484 #if defined(INET)
2485 	*v4 = VNET(ipforwarding);
2486 #endif
2487 	CURVNET_RESTORE();
2488 }
2489 
2490 /*
2491  * Returns true if it's possible this packet could be LROed.
2492  * if it returns false, it is guaranteed that tcp_lro_rx()
2493  * would not return zero.
2494  */
2495 static bool
2496 iflib_check_lro_possible(struct mbuf *m, bool v4_forwarding, bool v6_forwarding)
2497 {
2498 	struct ether_header *eh;
2499 	uint16_t eh_type;
2500 
2501 	eh = mtod(m, struct ether_header *);
2502 	eh_type = ntohs(eh->ether_type);
2503 	switch (eh_type) {
2504 #if defined(INET6)
2505 		case ETHERTYPE_IPV6:
2506 			return !v6_forwarding;
2507 #endif
2508 #if defined (INET)
2509 		case ETHERTYPE_IP:
2510 			return !v4_forwarding;
2511 #endif
2512 	}
2513 
2514 	return false;
2515 }
2516 #else
2517 static void
2518 iflib_get_ip_forwarding(struct lro_ctrl *lc __unused, bool *v4 __unused, bool *v6 __unused)
2519 {
2520 }
2521 #endif
2522 
2523 static bool
2524 iflib_rxeof(iflib_rxq_t rxq, qidx_t budget)
2525 {
2526 	if_ctx_t ctx = rxq->ifr_ctx;
2527 	if_shared_ctx_t sctx = ctx->ifc_sctx;
2528 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
2529 	int avail, i;
2530 	qidx_t *cidxp;
2531 	struct if_rxd_info ri;
2532 	int err, budget_left, rx_bytes, rx_pkts;
2533 	iflib_fl_t fl;
2534 	struct ifnet *ifp;
2535 	int lro_enabled;
2536 	bool lro_possible = false;
2537 	bool v4_forwarding, v6_forwarding;
2538 
2539 	/*
2540 	 * XXX early demux data packets so that if_input processing only handles
2541 	 * acks in interrupt context
2542 	 */
2543 	struct mbuf *m, *mh, *mt, *mf;
2544 
2545 	ifp = ctx->ifc_ifp;
2546 	mh = mt = NULL;
2547 	MPASS(budget > 0);
2548 	rx_pkts	= rx_bytes = 0;
2549 	if (sctx->isc_flags & IFLIB_HAS_RXCQ)
2550 		cidxp = &rxq->ifr_cq_cidx;
2551 	else
2552 		cidxp = &rxq->ifr_fl[0].ifl_cidx;
2553 	if ((avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget)) == 0) {
2554 		for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2555 			__iflib_fl_refill_lt(ctx, fl, budget + 8);
2556 		DBG_COUNTER_INC(rx_unavail);
2557 		return (false);
2558 	}
2559 
2560 	for (budget_left = budget; (budget_left > 0) && (avail > 0); budget_left--, avail--) {
2561 		if (__predict_false(!CTX_ACTIVE(ctx))) {
2562 			DBG_COUNTER_INC(rx_ctx_inactive);
2563 			break;
2564 		}
2565 		/*
2566 		 * Reset client set fields to their default values
2567 		 */
2568 		rxd_info_zero(&ri);
2569 		ri.iri_qsidx = rxq->ifr_id;
2570 		ri.iri_cidx = *cidxp;
2571 		ri.iri_ifp = ifp;
2572 		ri.iri_frags = rxq->ifr_frags;
2573 		err = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri);
2574 
2575 		if (err)
2576 			goto err;
2577 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
2578 			*cidxp = ri.iri_cidx;
2579 			/* Update our consumer index */
2580 			/* XXX NB: shurd - check if this is still safe */
2581 			while (rxq->ifr_cq_cidx >= scctx->isc_nrxd[0]) {
2582 				rxq->ifr_cq_cidx -= scctx->isc_nrxd[0];
2583 				rxq->ifr_cq_gen = 0;
2584 			}
2585 			/* was this only a completion queue message? */
2586 			if (__predict_false(ri.iri_nfrags == 0))
2587 				continue;
2588 		}
2589 		MPASS(ri.iri_nfrags != 0);
2590 		MPASS(ri.iri_len != 0);
2591 
2592 		/* will advance the cidx on the corresponding free lists */
2593 		m = iflib_rxd_pkt_get(rxq, &ri);
2594 		if (avail == 0 && budget_left)
2595 			avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget_left);
2596 
2597 		if (__predict_false(m == NULL)) {
2598 			DBG_COUNTER_INC(rx_mbuf_null);
2599 			continue;
2600 		}
2601 		/* imm_pkt: -- cxgb */
2602 		if (mh == NULL)
2603 			mh = mt = m;
2604 		else {
2605 			mt->m_nextpkt = m;
2606 			mt = m;
2607 		}
2608 	}
2609 	/* make sure that we can refill faster than drain */
2610 	for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++)
2611 		__iflib_fl_refill_lt(ctx, fl, budget + 8);
2612 
2613 	lro_enabled = (if_getcapenable(ifp) & IFCAP_LRO);
2614 	if (lro_enabled)
2615 		iflib_get_ip_forwarding(&rxq->ifr_lc, &v4_forwarding, &v6_forwarding);
2616 	mt = mf = NULL;
2617 	while (mh != NULL) {
2618 		m = mh;
2619 		mh = mh->m_nextpkt;
2620 		m->m_nextpkt = NULL;
2621 #ifndef __NO_STRICT_ALIGNMENT
2622 		if (!IP_ALIGNED(m) && (m = iflib_fixup_rx(m)) == NULL)
2623 			continue;
2624 #endif
2625 		rx_bytes += m->m_pkthdr.len;
2626 		rx_pkts++;
2627 #if defined(INET6) || defined(INET)
2628 		if (lro_enabled) {
2629 			if (!lro_possible) {
2630 				lro_possible = iflib_check_lro_possible(m, v4_forwarding, v6_forwarding);
2631 				if (lro_possible && mf != NULL) {
2632 					ifp->if_input(ifp, mf);
2633 					DBG_COUNTER_INC(rx_if_input);
2634 					mt = mf = NULL;
2635 				}
2636 			}
2637 			if ((m->m_pkthdr.csum_flags & (CSUM_L4_CALC|CSUM_L4_VALID)) ==
2638 			    (CSUM_L4_CALC|CSUM_L4_VALID)) {
2639 				if (lro_possible && tcp_lro_rx(&rxq->ifr_lc, m, 0) == 0)
2640 					continue;
2641 			}
2642 		}
2643 #endif
2644 		if (lro_possible) {
2645 			ifp->if_input(ifp, m);
2646 			DBG_COUNTER_INC(rx_if_input);
2647 			continue;
2648 		}
2649 
2650 		if (mf == NULL)
2651 			mf = m;
2652 		if (mt != NULL)
2653 			mt->m_nextpkt = m;
2654 		mt = m;
2655 	}
2656 	if (mf != NULL) {
2657 		ifp->if_input(ifp, mf);
2658 		DBG_COUNTER_INC(rx_if_input);
2659 	}
2660 
2661 	if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes);
2662 	if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts);
2663 
2664 	/*
2665 	 * Flush any outstanding LRO work
2666 	 */
2667 #if defined(INET6) || defined(INET)
2668 	tcp_lro_flush_all(&rxq->ifr_lc);
2669 #endif
2670 	if (avail)
2671 		return true;
2672 	return (iflib_rxd_avail(ctx, rxq, *cidxp, 1));
2673 err:
2674 	CTX_LOCK(ctx);
2675 	ctx->ifc_flags |= IFC_DO_RESET;
2676 	iflib_admin_intr_deferred(ctx);
2677 	CTX_UNLOCK(ctx);
2678 	return (false);
2679 }
2680 
2681 #define TXD_NOTIFY_COUNT(txq) (((txq)->ift_size / (txq)->ift_update_freq)-1)
2682 static inline qidx_t
2683 txq_max_db_deferred(iflib_txq_t txq, qidx_t in_use)
2684 {
2685 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2686 	qidx_t minthresh = txq->ift_size / 8;
2687 	if (in_use > 4*minthresh)
2688 		return (notify_count);
2689 	if (in_use > 2*minthresh)
2690 		return (notify_count >> 1);
2691 	if (in_use > minthresh)
2692 		return (notify_count >> 3);
2693 	return (0);
2694 }
2695 
2696 static inline qidx_t
2697 txq_max_rs_deferred(iflib_txq_t txq)
2698 {
2699 	qidx_t notify_count = TXD_NOTIFY_COUNT(txq);
2700 	qidx_t minthresh = txq->ift_size / 8;
2701 	if (txq->ift_in_use > 4*minthresh)
2702 		return (notify_count);
2703 	if (txq->ift_in_use > 2*minthresh)
2704 		return (notify_count >> 1);
2705 	if (txq->ift_in_use > minthresh)
2706 		return (notify_count >> 2);
2707 	return (2);
2708 }
2709 
2710 #define M_CSUM_FLAGS(m) ((m)->m_pkthdr.csum_flags)
2711 #define M_HAS_VLANTAG(m) (m->m_flags & M_VLANTAG)
2712 
2713 #define TXQ_MAX_DB_DEFERRED(txq, in_use) txq_max_db_deferred((txq), (in_use))
2714 #define TXQ_MAX_RS_DEFERRED(txq) txq_max_rs_deferred(txq)
2715 #define TXQ_MAX_DB_CONSUMED(size) (size >> 4)
2716 
2717 /* forward compatibility for cxgb */
2718 #define FIRST_QSET(ctx) 0
2719 #define NTXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_ntxqsets)
2720 #define NRXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_nrxqsets)
2721 #define QIDX(ctx, m) ((((m)->m_pkthdr.flowid & ctx->ifc_softc_ctx.isc_rss_table_mask) % NTXQSETS(ctx)) + FIRST_QSET(ctx))
2722 #define DESC_RECLAIMABLE(q) ((int)((q)->ift_processed - (q)->ift_cleaned - (q)->ift_ctx->ifc_softc_ctx.isc_tx_nsegments))
2723 
2724 /* XXX we should be setting this to something other than zero */
2725 #define RECLAIM_THRESH(ctx) ((ctx)->ifc_sctx->isc_tx_reclaim_thresh)
2726 #define MAX_TX_DESC(ctx) ((ctx)->ifc_softc_ctx.isc_tx_tso_segments_max)
2727 
2728 static inline bool
2729 iflib_txd_db_check(if_ctx_t ctx, iflib_txq_t txq, int ring, qidx_t in_use)
2730 {
2731 	qidx_t dbval, max;
2732 	bool rang;
2733 
2734 	rang = false;
2735 	max = TXQ_MAX_DB_DEFERRED(txq, in_use);
2736 	if (ring || txq->ift_db_pending >= max) {
2737 		dbval = txq->ift_npending ? txq->ift_npending : txq->ift_pidx;
2738 		ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, dbval);
2739 		txq->ift_db_pending = txq->ift_npending = 0;
2740 		rang = true;
2741 	}
2742 	return (rang);
2743 }
2744 
2745 #ifdef PKT_DEBUG
2746 static void
2747 print_pkt(if_pkt_info_t pi)
2748 {
2749 	printf("pi len:  %d qsidx: %d nsegs: %d ndescs: %d flags: %x pidx: %d\n",
2750 	       pi->ipi_len, pi->ipi_qsidx, pi->ipi_nsegs, pi->ipi_ndescs, pi->ipi_flags, pi->ipi_pidx);
2751 	printf("pi new_pidx: %d csum_flags: %lx tso_segsz: %d mflags: %x vtag: %d\n",
2752 	       pi->ipi_new_pidx, pi->ipi_csum_flags, pi->ipi_tso_segsz, pi->ipi_mflags, pi->ipi_vtag);
2753 	printf("pi etype: %d ehdrlen: %d ip_hlen: %d ipproto: %d\n",
2754 	       pi->ipi_etype, pi->ipi_ehdrlen, pi->ipi_ip_hlen, pi->ipi_ipproto);
2755 }
2756 #endif
2757 
2758 #define IS_TSO4(pi) ((pi)->ipi_csum_flags & CSUM_IP_TSO)
2759 #define IS_TSO6(pi) ((pi)->ipi_csum_flags & CSUM_IP6_TSO)
2760 
2761 static int
2762 iflib_parse_header(iflib_txq_t txq, if_pkt_info_t pi, struct mbuf **mp)
2763 {
2764 	if_shared_ctx_t sctx = txq->ift_ctx->ifc_sctx;
2765 	struct ether_vlan_header *eh;
2766 	struct mbuf *m, *n;
2767 
2768 	n = m = *mp;
2769 	if ((sctx->isc_flags & IFLIB_NEED_SCRATCH) &&
2770 	    M_WRITABLE(m) == 0) {
2771 		if ((m = m_dup(m, M_NOWAIT)) == NULL) {
2772 			return (ENOMEM);
2773 		} else {
2774 			m_freem(*mp);
2775 			n = *mp = m;
2776 		}
2777 	}
2778 
2779 	/*
2780 	 * Determine where frame payload starts.
2781 	 * Jump over vlan headers if already present,
2782 	 * helpful for QinQ too.
2783 	 */
2784 	if (__predict_false(m->m_len < sizeof(*eh))) {
2785 		txq->ift_pullups++;
2786 		if (__predict_false((m = m_pullup(m, sizeof(*eh))) == NULL))
2787 			return (ENOMEM);
2788 	}
2789 	eh = mtod(m, struct ether_vlan_header *);
2790 	if (eh->evl_encap_proto == htons(ETHERTYPE_VLAN)) {
2791 		pi->ipi_etype = ntohs(eh->evl_proto);
2792 		pi->ipi_ehdrlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN;
2793 	} else {
2794 		pi->ipi_etype = ntohs(eh->evl_encap_proto);
2795 		pi->ipi_ehdrlen = ETHER_HDR_LEN;
2796 	}
2797 
2798 	switch (pi->ipi_etype) {
2799 #ifdef INET
2800 	case ETHERTYPE_IP:
2801 	{
2802 		struct ip *ip = NULL;
2803 		struct tcphdr *th = NULL;
2804 		int minthlen;
2805 
2806 		minthlen = min(m->m_pkthdr.len, pi->ipi_ehdrlen + sizeof(*ip) + sizeof(*th));
2807 		if (__predict_false(m->m_len < minthlen)) {
2808 			/*
2809 			 * if this code bloat is causing too much of a hit
2810 			 * move it to a separate function and mark it noinline
2811 			 */
2812 			if (m->m_len == pi->ipi_ehdrlen) {
2813 				n = m->m_next;
2814 				MPASS(n);
2815 				if (n->m_len >= sizeof(*ip))  {
2816 					ip = (struct ip *)n->m_data;
2817 					if (n->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2818 						th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2819 				} else {
2820 					txq->ift_pullups++;
2821 					if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
2822 						return (ENOMEM);
2823 					ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2824 				}
2825 			} else {
2826 				txq->ift_pullups++;
2827 				if (__predict_false((m = m_pullup(m, minthlen)) == NULL))
2828 					return (ENOMEM);
2829 				ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2830 				if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2831 					th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2832 			}
2833 		} else {
2834 			ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen);
2835 			if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th))
2836 				th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2));
2837 		}
2838 		pi->ipi_ip_hlen = ip->ip_hl << 2;
2839 		pi->ipi_ipproto = ip->ip_p;
2840 		pi->ipi_flags |= IPI_TX_IPV4;
2841 
2842 		if ((sctx->isc_flags & IFLIB_NEED_ZERO_CSUM) && (pi->ipi_csum_flags & CSUM_IP))
2843                        ip->ip_sum = 0;
2844 
2845 		if (IS_TSO4(pi)) {
2846 			if (pi->ipi_ipproto == IPPROTO_TCP) {
2847 				if (__predict_false(th == NULL)) {
2848 					txq->ift_pullups++;
2849 					if (__predict_false((m = m_pullup(m, (ip->ip_hl << 2) + sizeof(*th))) == NULL))
2850 						return (ENOMEM);
2851 					th = (struct tcphdr *)((caddr_t)ip + pi->ipi_ip_hlen);
2852 				}
2853 				pi->ipi_tcp_hflags = th->th_flags;
2854 				pi->ipi_tcp_hlen = th->th_off << 2;
2855 				pi->ipi_tcp_seq = th->th_seq;
2856 			}
2857 			if (__predict_false(ip->ip_p != IPPROTO_TCP))
2858 				return (ENXIO);
2859 			th->th_sum = in_pseudo(ip->ip_src.s_addr,
2860 					       ip->ip_dst.s_addr, htons(IPPROTO_TCP));
2861 			pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
2862 			if (sctx->isc_flags & IFLIB_TSO_INIT_IP) {
2863 				ip->ip_sum = 0;
2864 				ip->ip_len = htons(pi->ipi_ip_hlen + pi->ipi_tcp_hlen + pi->ipi_tso_segsz);
2865 			}
2866 		}
2867 		break;
2868 	}
2869 #endif
2870 #ifdef INET6
2871 	case ETHERTYPE_IPV6:
2872 	{
2873 		struct ip6_hdr *ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen);
2874 		struct tcphdr *th;
2875 		pi->ipi_ip_hlen = sizeof(struct ip6_hdr);
2876 
2877 		if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) {
2878 			if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL))
2879 				return (ENOMEM);
2880 		}
2881 		th = (struct tcphdr *)((caddr_t)ip6 + pi->ipi_ip_hlen);
2882 
2883 		/* XXX-BZ this will go badly in case of ext hdrs. */
2884 		pi->ipi_ipproto = ip6->ip6_nxt;
2885 		pi->ipi_flags |= IPI_TX_IPV6;
2886 
2887 		if (IS_TSO6(pi)) {
2888 			if (pi->ipi_ipproto == IPPROTO_TCP) {
2889 				if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) {
2890 					if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) == NULL))
2891 						return (ENOMEM);
2892 				}
2893 				pi->ipi_tcp_hflags = th->th_flags;
2894 				pi->ipi_tcp_hlen = th->th_off << 2;
2895 			}
2896 
2897 			if (__predict_false(ip6->ip6_nxt != IPPROTO_TCP))
2898 				return (ENXIO);
2899 			/*
2900 			 * The corresponding flag is set by the stack in the IPv4
2901 			 * TSO case, but not in IPv6 (at least in FreeBSD 10.2).
2902 			 * So, set it here because the rest of the flow requires it.
2903 			 */
2904 			pi->ipi_csum_flags |= CSUM_TCP_IPV6;
2905 			th->th_sum = in6_cksum_pseudo(ip6, 0, IPPROTO_TCP, 0);
2906 			pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz;
2907 		}
2908 		break;
2909 	}
2910 #endif
2911 	default:
2912 		pi->ipi_csum_flags &= ~CSUM_OFFLOAD;
2913 		pi->ipi_ip_hlen = 0;
2914 		break;
2915 	}
2916 	*mp = m;
2917 
2918 	return (0);
2919 }
2920 
2921 static  __noinline  struct mbuf *
2922 collapse_pkthdr(struct mbuf *m0)
2923 {
2924 	struct mbuf *m, *m_next, *tmp;
2925 
2926 	m = m0;
2927 	m_next = m->m_next;
2928 	while (m_next != NULL && m_next->m_len == 0) {
2929 		m = m_next;
2930 		m->m_next = NULL;
2931 		m_free(m);
2932 		m_next = m_next->m_next;
2933 	}
2934 	m = m0;
2935 	m->m_next = m_next;
2936 	if ((m_next->m_flags & M_EXT) == 0) {
2937 		m = m_defrag(m, M_NOWAIT);
2938 	} else {
2939 		tmp = m_next->m_next;
2940 		memcpy(m_next, m, MPKTHSIZE);
2941 		m = m_next;
2942 		m->m_next = tmp;
2943 	}
2944 	return (m);
2945 }
2946 
2947 /*
2948  * If dodgy hardware rejects the scatter gather chain we've handed it
2949  * we'll need to remove the mbuf chain from ifsg_m[] before we can add the
2950  * m_defrag'd mbufs
2951  */
2952 static __noinline struct mbuf *
2953 iflib_remove_mbuf(iflib_txq_t txq)
2954 {
2955 	int ntxd, i, pidx;
2956 	struct mbuf *m, *mh, **ifsd_m;
2957 
2958 	pidx = txq->ift_pidx;
2959 	ifsd_m = txq->ift_sds.ifsd_m;
2960 	ntxd = txq->ift_size;
2961 	mh = m = ifsd_m[pidx];
2962 	ifsd_m[pidx] = NULL;
2963 #if MEMORY_LOGGING
2964 	txq->ift_dequeued++;
2965 #endif
2966 	i = 1;
2967 
2968 	while (m) {
2969 		ifsd_m[(pidx + i) & (ntxd -1)] = NULL;
2970 #if MEMORY_LOGGING
2971 		txq->ift_dequeued++;
2972 #endif
2973 		m = m->m_next;
2974 		i++;
2975 	}
2976 	return (mh);
2977 }
2978 
2979 static int
2980 iflib_busdma_load_mbuf_sg(iflib_txq_t txq, bus_dma_tag_t tag, bus_dmamap_t map,
2981 			  struct mbuf **m0, bus_dma_segment_t *segs, int *nsegs,
2982 			  int max_segs, int flags)
2983 {
2984 	if_ctx_t ctx;
2985 	if_shared_ctx_t		sctx;
2986 	if_softc_ctx_t		scctx;
2987 	int i, next, pidx, err, ntxd, count;
2988 	struct mbuf *m, *tmp, **ifsd_m;
2989 
2990 	m = *m0;
2991 
2992 	/*
2993 	 * Please don't ever do this
2994 	 */
2995 	if (__predict_false(m->m_len == 0))
2996 		*m0 = m = collapse_pkthdr(m);
2997 
2998 	ctx = txq->ift_ctx;
2999 	sctx = ctx->ifc_sctx;
3000 	scctx = &ctx->ifc_softc_ctx;
3001 	ifsd_m = txq->ift_sds.ifsd_m;
3002 	ntxd = txq->ift_size;
3003 	pidx = txq->ift_pidx;
3004 	if (map != NULL) {
3005 		uint8_t *ifsd_flags = txq->ift_sds.ifsd_flags;
3006 
3007 		err = bus_dmamap_load_mbuf_sg(tag, map,
3008 					      *m0, segs, nsegs, BUS_DMA_NOWAIT);
3009 		if (err)
3010 			return (err);
3011 		ifsd_flags[pidx] |= TX_SW_DESC_MAPPED;
3012 		count = 0;
3013 		m = *m0;
3014 		do {
3015 			if (__predict_false(m->m_len <= 0)) {
3016 				tmp = m;
3017 				m = m->m_next;
3018 				tmp->m_next = NULL;
3019 				m_free(tmp);
3020 				continue;
3021 			}
3022 			m = m->m_next;
3023 			count++;
3024 		} while (m != NULL);
3025 		if (count > *nsegs) {
3026 			ifsd_m[pidx] = *m0;
3027 			ifsd_m[pidx]->m_flags |= M_TOOBIG;
3028 			return (0);
3029 		}
3030 		m = *m0;
3031 		count = 0;
3032 		do {
3033 			next = (pidx + count) & (ntxd-1);
3034 			MPASS(ifsd_m[next] == NULL);
3035 			ifsd_m[next] = m;
3036 			count++;
3037 			tmp = m;
3038 			m = m->m_next;
3039 		} while (m != NULL);
3040 	} else {
3041 		int buflen, sgsize, maxsegsz, max_sgsize;
3042 		vm_offset_t vaddr;
3043 		vm_paddr_t curaddr;
3044 
3045 		count = i = 0;
3046 		m = *m0;
3047 		if (m->m_pkthdr.csum_flags & CSUM_TSO)
3048 			maxsegsz = scctx->isc_tx_tso_segsize_max;
3049 		else
3050 			maxsegsz = sctx->isc_tx_maxsegsize;
3051 
3052 		do {
3053 			if (__predict_false(m->m_len <= 0)) {
3054 				tmp = m;
3055 				m = m->m_next;
3056 				tmp->m_next = NULL;
3057 				m_free(tmp);
3058 				continue;
3059 			}
3060 			buflen = m->m_len;
3061 			vaddr = (vm_offset_t)m->m_data;
3062 			/*
3063 			 * see if we can't be smarter about physically
3064 			 * contiguous mappings
3065 			 */
3066 			next = (pidx + count) & (ntxd-1);
3067 			MPASS(ifsd_m[next] == NULL);
3068 #if MEMORY_LOGGING
3069 			txq->ift_enqueued++;
3070 #endif
3071 			ifsd_m[next] = m;
3072 			while (buflen > 0) {
3073 				if (i >= max_segs)
3074 					goto err;
3075 				max_sgsize = MIN(buflen, maxsegsz);
3076 				curaddr = pmap_kextract(vaddr);
3077 				sgsize = PAGE_SIZE - (curaddr & PAGE_MASK);
3078 				sgsize = MIN(sgsize, max_sgsize);
3079 				segs[i].ds_addr = curaddr;
3080 				segs[i].ds_len = sgsize;
3081 				vaddr += sgsize;
3082 				buflen -= sgsize;
3083 				i++;
3084 			}
3085 			count++;
3086 			tmp = m;
3087 			m = m->m_next;
3088 		} while (m != NULL);
3089 		*nsegs = i;
3090 	}
3091 	return (0);
3092 err:
3093 	*m0 = iflib_remove_mbuf(txq);
3094 	return (EFBIG);
3095 }
3096 
3097 static inline caddr_t
3098 calc_next_txd(iflib_txq_t txq, int cidx, uint8_t qid)
3099 {
3100 	qidx_t size;
3101 	int ntxd;
3102 	caddr_t start, end, cur, next;
3103 
3104 	ntxd = txq->ift_size;
3105 	size = txq->ift_txd_size[qid];
3106 	start = txq->ift_ifdi[qid].idi_vaddr;
3107 
3108 	if (__predict_false(size == 0))
3109 		return (start);
3110 	cur = start + size*cidx;
3111 	end = start + size*ntxd;
3112 	next = CACHE_PTR_NEXT(cur);
3113 	return (next < end ? next : start);
3114 }
3115 
3116 /*
3117  * Pad an mbuf to ensure a minimum ethernet frame size.
3118  * min_frame_size is the frame size (less CRC) to pad the mbuf to
3119  */
3120 static __noinline int
3121 iflib_ether_pad(device_t dev, struct mbuf **m_head, uint16_t min_frame_size)
3122 {
3123 	/*
3124 	 * 18 is enough bytes to pad an ARP packet to 46 bytes, and
3125 	 * and ARP message is the smallest common payload I can think of
3126 	 */
3127 	static char pad[18];	/* just zeros */
3128 	int n;
3129 	struct mbuf *new_head;
3130 
3131 	if (!M_WRITABLE(*m_head)) {
3132 		new_head = m_dup(*m_head, M_NOWAIT);
3133 		if (new_head == NULL) {
3134 			m_freem(*m_head);
3135 			device_printf(dev, "cannot pad short frame, m_dup() failed");
3136 			DBG_COUNTER_INC(encap_pad_mbuf_fail);
3137 			return ENOMEM;
3138 		}
3139 		m_freem(*m_head);
3140 		*m_head = new_head;
3141 	}
3142 
3143 	for (n = min_frame_size - (*m_head)->m_pkthdr.len;
3144 	     n > 0; n -= sizeof(pad))
3145 		if (!m_append(*m_head, min(n, sizeof(pad)), pad))
3146 			break;
3147 
3148 	if (n > 0) {
3149 		m_freem(*m_head);
3150 		device_printf(dev, "cannot pad short frame\n");
3151 		DBG_COUNTER_INC(encap_pad_mbuf_fail);
3152 		return (ENOBUFS);
3153 	}
3154 
3155 	return 0;
3156 }
3157 
3158 static int
3159 iflib_encap(iflib_txq_t txq, struct mbuf **m_headp)
3160 {
3161 	if_ctx_t		ctx;
3162 	if_shared_ctx_t		sctx;
3163 	if_softc_ctx_t		scctx;
3164 	bus_dma_segment_t	*segs;
3165 	struct mbuf		*m_head;
3166 	void			*next_txd;
3167 	bus_dmamap_t		map;
3168 	struct if_pkt_info	pi;
3169 	int remap = 0;
3170 	int err, nsegs, ndesc, max_segs, pidx, cidx, next, ntxd;
3171 	bus_dma_tag_t desc_tag;
3172 
3173 	segs = txq->ift_segs;
3174 	ctx = txq->ift_ctx;
3175 	sctx = ctx->ifc_sctx;
3176 	scctx = &ctx->ifc_softc_ctx;
3177 	segs = txq->ift_segs;
3178 	ntxd = txq->ift_size;
3179 	m_head = *m_headp;
3180 	map = NULL;
3181 
3182 	/*
3183 	 * If we're doing TSO the next descriptor to clean may be quite far ahead
3184 	 */
3185 	cidx = txq->ift_cidx;
3186 	pidx = txq->ift_pidx;
3187 	if (ctx->ifc_flags & IFC_PREFETCH) {
3188 		next = (cidx + CACHE_PTR_INCREMENT) & (ntxd-1);
3189 		if (!(ctx->ifc_flags & IFLIB_HAS_TXCQ)) {
3190 			next_txd = calc_next_txd(txq, cidx, 0);
3191 			prefetch(next_txd);
3192 		}
3193 
3194 		/* prefetch the next cache line of mbuf pointers and flags */
3195 		prefetch(&txq->ift_sds.ifsd_m[next]);
3196 		if (txq->ift_sds.ifsd_map != NULL) {
3197 			prefetch(&txq->ift_sds.ifsd_map[next]);
3198 			next = (cidx + CACHE_LINE_SIZE) & (ntxd-1);
3199 			prefetch(&txq->ift_sds.ifsd_flags[next]);
3200 		}
3201 	} else if (txq->ift_sds.ifsd_map != NULL)
3202 		map = txq->ift_sds.ifsd_map[pidx];
3203 
3204 	if (m_head->m_pkthdr.csum_flags & CSUM_TSO) {
3205 		desc_tag = txq->ift_tso_desc_tag;
3206 		max_segs = scctx->isc_tx_tso_segments_max;
3207 	} else {
3208 		desc_tag = txq->ift_desc_tag;
3209 		max_segs = scctx->isc_tx_nsegments;
3210 	}
3211 	if ((sctx->isc_flags & IFLIB_NEED_ETHER_PAD) &&
3212 	    __predict_false(m_head->m_pkthdr.len < scctx->isc_min_frame_size)) {
3213 		err = iflib_ether_pad(ctx->ifc_dev, m_headp, scctx->isc_min_frame_size);
3214 		if (err)
3215 			return err;
3216 	}
3217 	m_head = *m_headp;
3218 
3219 	pkt_info_zero(&pi);
3220 	pi.ipi_mflags = (m_head->m_flags & (M_VLANTAG|M_BCAST|M_MCAST));
3221 	pi.ipi_pidx = pidx;
3222 	pi.ipi_qsidx = txq->ift_id;
3223 	pi.ipi_len = m_head->m_pkthdr.len;
3224 	pi.ipi_csum_flags = m_head->m_pkthdr.csum_flags;
3225 	pi.ipi_vtag = (m_head->m_flags & M_VLANTAG) ? m_head->m_pkthdr.ether_vtag : 0;
3226 
3227 	/* deliberate bitwise OR to make one condition */
3228 	if (__predict_true((pi.ipi_csum_flags | pi.ipi_vtag))) {
3229 		if (__predict_false((err = iflib_parse_header(txq, &pi, m_headp)) != 0))
3230 			return (err);
3231 		m_head = *m_headp;
3232 	}
3233 
3234 retry:
3235 	err = iflib_busdma_load_mbuf_sg(txq, desc_tag, map, m_headp, segs, &nsegs, max_segs, BUS_DMA_NOWAIT);
3236 defrag:
3237 	if (__predict_false(err)) {
3238 		switch (err) {
3239 		case EFBIG:
3240 			/* try collapse once and defrag once */
3241 			if (remap == 0)
3242 				m_head = m_collapse(*m_headp, M_NOWAIT, max_segs);
3243 			if (remap == 1)
3244 				m_head = m_defrag(*m_headp, M_NOWAIT);
3245 			remap++;
3246 			if (__predict_false(m_head == NULL))
3247 				goto defrag_failed;
3248 			txq->ift_mbuf_defrag++;
3249 			*m_headp = m_head;
3250 			goto retry;
3251 			break;
3252 		case ENOMEM:
3253 			txq->ift_no_tx_dma_setup++;
3254 			break;
3255 		default:
3256 			txq->ift_no_tx_dma_setup++;
3257 			m_freem(*m_headp);
3258 			DBG_COUNTER_INC(tx_frees);
3259 			*m_headp = NULL;
3260 			break;
3261 		}
3262 		txq->ift_map_failed++;
3263 		DBG_COUNTER_INC(encap_load_mbuf_fail);
3264 		return (err);
3265 	}
3266 
3267 	/*
3268 	 * XXX assumes a 1 to 1 relationship between segments and
3269 	 *        descriptors - this does not hold true on all drivers, e.g.
3270 	 *        cxgb
3271 	 */
3272 	if (__predict_false(nsegs + 2 > TXQ_AVAIL(txq))) {
3273 		txq->ift_no_desc_avail++;
3274 		if (map != NULL)
3275 			bus_dmamap_unload(desc_tag, map);
3276 		DBG_COUNTER_INC(encap_txq_avail_fail);
3277 		if ((txq->ift_task.gt_task.ta_flags & TASK_ENQUEUED) == 0)
3278 			GROUPTASK_ENQUEUE(&txq->ift_task);
3279 		return (ENOBUFS);
3280 	}
3281 	/*
3282 	 * On Intel cards we can greatly reduce the number of TX interrupts
3283 	 * we see by only setting report status on every Nth descriptor.
3284 	 * However, this also means that the driver will need to keep track
3285 	 * of the descriptors that RS was set on to check them for the DD bit.
3286 	 */
3287 	txq->ift_rs_pending += nsegs + 1;
3288 	if (txq->ift_rs_pending > TXQ_MAX_RS_DEFERRED(txq) ||
3289 	     iflib_no_tx_batch || (TXQ_AVAIL(txq) - nsegs - 1) <= MAX_TX_DESC(ctx)) {
3290 		pi.ipi_flags |= IPI_TX_INTR;
3291 		txq->ift_rs_pending = 0;
3292 	}
3293 
3294 	pi.ipi_segs = segs;
3295 	pi.ipi_nsegs = nsegs;
3296 
3297 	MPASS(pidx >= 0 && pidx < txq->ift_size);
3298 #ifdef PKT_DEBUG
3299 	print_pkt(&pi);
3300 #endif
3301 	if (map != NULL)
3302 		bus_dmamap_sync(desc_tag, map, BUS_DMASYNC_PREWRITE);
3303 	if ((err = ctx->isc_txd_encap(ctx->ifc_softc, &pi)) == 0) {
3304 		if (map != NULL)
3305 			bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map,
3306 					BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE);
3307 		DBG_COUNTER_INC(tx_encap);
3308 		MPASS(pi.ipi_new_pidx < txq->ift_size);
3309 
3310 		ndesc = pi.ipi_new_pidx - pi.ipi_pidx;
3311 		if (pi.ipi_new_pidx < pi.ipi_pidx) {
3312 			ndesc += txq->ift_size;
3313 			txq->ift_gen = 1;
3314 		}
3315 		/*
3316 		 * drivers can need as many as
3317 		 * two sentinels
3318 		 */
3319 		MPASS(ndesc <= pi.ipi_nsegs + 2);
3320 		MPASS(pi.ipi_new_pidx != pidx);
3321 		MPASS(ndesc > 0);
3322 		txq->ift_in_use += ndesc;
3323 
3324 		/*
3325 		 * We update the last software descriptor again here because there may
3326 		 * be a sentinel and/or there may be more mbufs than segments
3327 		 */
3328 		txq->ift_pidx = pi.ipi_new_pidx;
3329 		txq->ift_npending += pi.ipi_ndescs;
3330 	} else if (__predict_false(err == EFBIG && remap < 2)) {
3331 		*m_headp = m_head = iflib_remove_mbuf(txq);
3332 		remap = 1;
3333 		txq->ift_txd_encap_efbig++;
3334 		goto defrag;
3335 	} else
3336 		DBG_COUNTER_INC(encap_txd_encap_fail);
3337 	return (err);
3338 
3339 defrag_failed:
3340 	txq->ift_mbuf_defrag_failed++;
3341 	txq->ift_map_failed++;
3342 	m_freem(*m_headp);
3343 	DBG_COUNTER_INC(tx_frees);
3344 	*m_headp = NULL;
3345 	return (ENOMEM);
3346 }
3347 
3348 static void
3349 iflib_tx_desc_free(iflib_txq_t txq, int n)
3350 {
3351 	int hasmap;
3352 	uint32_t qsize, cidx, mask, gen;
3353 	struct mbuf *m, **ifsd_m;
3354 	uint8_t *ifsd_flags;
3355 	bus_dmamap_t *ifsd_map;
3356 	bool do_prefetch;
3357 
3358 	cidx = txq->ift_cidx;
3359 	gen = txq->ift_gen;
3360 	qsize = txq->ift_size;
3361 	mask = qsize-1;
3362 	hasmap = txq->ift_sds.ifsd_map != NULL;
3363 	ifsd_flags = txq->ift_sds.ifsd_flags;
3364 	ifsd_m = txq->ift_sds.ifsd_m;
3365 	ifsd_map = txq->ift_sds.ifsd_map;
3366 	do_prefetch = (txq->ift_ctx->ifc_flags & IFC_PREFETCH);
3367 
3368 	while (n--) {
3369 		if (do_prefetch) {
3370 			prefetch(ifsd_m[(cidx + 3) & mask]);
3371 			prefetch(ifsd_m[(cidx + 4) & mask]);
3372 		}
3373 		if (ifsd_m[cidx] != NULL) {
3374 			prefetch(&ifsd_m[(cidx + CACHE_PTR_INCREMENT) & mask]);
3375 			prefetch(&ifsd_flags[(cidx + CACHE_PTR_INCREMENT) & mask]);
3376 			if (hasmap && (ifsd_flags[cidx] & TX_SW_DESC_MAPPED)) {
3377 				/*
3378 				 * does it matter if it's not the TSO tag? If so we'll
3379 				 * have to add the type to flags
3380 				 */
3381 				bus_dmamap_unload(txq->ift_desc_tag, ifsd_map[cidx]);
3382 				ifsd_flags[cidx] &= ~TX_SW_DESC_MAPPED;
3383 			}
3384 			if ((m = ifsd_m[cidx]) != NULL) {
3385 				/* XXX we don't support any drivers that batch packets yet */
3386 				MPASS(m->m_nextpkt == NULL);
3387 				/* if the number of clusters exceeds the number of segments
3388 				 * there won't be space on the ring to save a pointer to each
3389 				 * cluster so we simply free the list here
3390 				 */
3391 				if (m->m_flags & M_TOOBIG) {
3392 					m_freem(m);
3393 				} else {
3394 					m_free(m);
3395 				}
3396 				ifsd_m[cidx] = NULL;
3397 #if MEMORY_LOGGING
3398 				txq->ift_dequeued++;
3399 #endif
3400 				DBG_COUNTER_INC(tx_frees);
3401 			}
3402 		}
3403 		if (__predict_false(++cidx == qsize)) {
3404 			cidx = 0;
3405 			gen = 0;
3406 		}
3407 	}
3408 	txq->ift_cidx = cidx;
3409 	txq->ift_gen = gen;
3410 }
3411 
3412 static __inline int
3413 iflib_completed_tx_reclaim(iflib_txq_t txq, int thresh)
3414 {
3415 	int reclaim;
3416 	if_ctx_t ctx = txq->ift_ctx;
3417 
3418 	KASSERT(thresh >= 0, ("invalid threshold to reclaim"));
3419 	MPASS(thresh /*+ MAX_TX_DESC(txq->ift_ctx) */ < txq->ift_size);
3420 
3421 	/*
3422 	 * Need a rate-limiting check so that this isn't called every time
3423 	 */
3424 	iflib_tx_credits_update(ctx, txq);
3425 	reclaim = DESC_RECLAIMABLE(txq);
3426 
3427 	if (reclaim <= thresh /* + MAX_TX_DESC(txq->ift_ctx) */) {
3428 #ifdef INVARIANTS
3429 		if (iflib_verbose_debug) {
3430 			printf("%s processed=%ju cleaned=%ju tx_nsegments=%d reclaim=%d thresh=%d\n", __FUNCTION__,
3431 			       txq->ift_processed, txq->ift_cleaned, txq->ift_ctx->ifc_softc_ctx.isc_tx_nsegments,
3432 			       reclaim, thresh);
3433 
3434 		}
3435 #endif
3436 		return (0);
3437 	}
3438 	iflib_tx_desc_free(txq, reclaim);
3439 	txq->ift_cleaned += reclaim;
3440 	txq->ift_in_use -= reclaim;
3441 
3442 	return (reclaim);
3443 }
3444 
3445 static struct mbuf **
3446 _ring_peek_one(struct ifmp_ring *r, int cidx, int offset, int remaining)
3447 {
3448 	int next, size;
3449 	struct mbuf **items;
3450 
3451 	size = r->size;
3452 	next = (cidx + CACHE_PTR_INCREMENT) & (size-1);
3453 	items = __DEVOLATILE(struct mbuf **, &r->items[0]);
3454 
3455 	prefetch(items[(cidx + offset) & (size-1)]);
3456 	if (remaining > 1) {
3457 		prefetch2cachelines(&items[next]);
3458 		prefetch2cachelines(items[(cidx + offset + 1) & (size-1)]);
3459 		prefetch2cachelines(items[(cidx + offset + 2) & (size-1)]);
3460 		prefetch2cachelines(items[(cidx + offset + 3) & (size-1)]);
3461 	}
3462 	return (__DEVOLATILE(struct mbuf **, &r->items[(cidx + offset) & (size-1)]));
3463 }
3464 
3465 static void
3466 iflib_txq_check_drain(iflib_txq_t txq, int budget)
3467 {
3468 
3469 	ifmp_ring_check_drainage(txq->ift_br, budget);
3470 }
3471 
3472 static uint32_t
3473 iflib_txq_can_drain(struct ifmp_ring *r)
3474 {
3475 	iflib_txq_t txq = r->cookie;
3476 	if_ctx_t ctx = txq->ift_ctx;
3477 
3478 	return ((TXQ_AVAIL(txq) > MAX_TX_DESC(ctx) + 2) ||
3479 		ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, false));
3480 }
3481 
3482 static uint32_t
3483 iflib_txq_drain(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3484 {
3485 	iflib_txq_t txq = r->cookie;
3486 	if_ctx_t ctx = txq->ift_ctx;
3487 	struct ifnet *ifp = ctx->ifc_ifp;
3488 	struct mbuf **mp, *m;
3489 	int i, count, consumed, pkt_sent, bytes_sent, mcast_sent, avail;
3490 	int reclaimed, err, in_use_prev, desc_used;
3491 	bool do_prefetch, ring, rang;
3492 
3493 	if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING) ||
3494 			    !LINK_ACTIVE(ctx))) {
3495 		DBG_COUNTER_INC(txq_drain_notready);
3496 		return (0);
3497 	}
3498 	reclaimed = iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx));
3499 	rang = iflib_txd_db_check(ctx, txq, reclaimed, txq->ift_in_use);
3500 	avail = IDXDIFF(pidx, cidx, r->size);
3501 	if (__predict_false(ctx->ifc_flags & IFC_QFLUSH)) {
3502 		DBG_COUNTER_INC(txq_drain_flushing);
3503 		for (i = 0; i < avail; i++) {
3504 			m_free(r->items[(cidx + i) & (r->size-1)]);
3505 			r->items[(cidx + i) & (r->size-1)] = NULL;
3506 		}
3507 		return (avail);
3508 	}
3509 
3510 	if (__predict_false(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) {
3511 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3512 		CALLOUT_LOCK(txq);
3513 		callout_stop(&txq->ift_timer);
3514 		CALLOUT_UNLOCK(txq);
3515 		DBG_COUNTER_INC(txq_drain_oactive);
3516 		return (0);
3517 	}
3518 	if (reclaimed)
3519 		txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3520 	consumed = mcast_sent = bytes_sent = pkt_sent = 0;
3521 	count = MIN(avail, TX_BATCH_SIZE);
3522 #ifdef INVARIANTS
3523 	if (iflib_verbose_debug)
3524 		printf("%s avail=%d ifc_flags=%x txq_avail=%d ", __FUNCTION__,
3525 		       avail, ctx->ifc_flags, TXQ_AVAIL(txq));
3526 #endif
3527 	do_prefetch = (ctx->ifc_flags & IFC_PREFETCH);
3528 	avail = TXQ_AVAIL(txq);
3529 	for (desc_used = i = 0; i < count && avail > MAX_TX_DESC(ctx) + 2; i++) {
3530 		int pidx_prev, rem = do_prefetch ? count - i : 0;
3531 
3532 		mp = _ring_peek_one(r, cidx, i, rem);
3533 		MPASS(mp != NULL && *mp != NULL);
3534 		if (__predict_false(*mp == (struct mbuf *)txq)) {
3535 			consumed++;
3536 			reclaimed++;
3537 			continue;
3538 		}
3539 		in_use_prev = txq->ift_in_use;
3540 		pidx_prev = txq->ift_pidx;
3541 		err = iflib_encap(txq, mp);
3542 		if (__predict_false(err)) {
3543 			DBG_COUNTER_INC(txq_drain_encapfail);
3544 			/* no room - bail out */
3545 			if (err == ENOBUFS)
3546 				break;
3547 			consumed++;
3548 			DBG_COUNTER_INC(txq_drain_encapfail);
3549 			/* we can't send this packet - skip it */
3550 			continue;
3551 		}
3552 		consumed++;
3553 		pkt_sent++;
3554 		m = *mp;
3555 		DBG_COUNTER_INC(tx_sent);
3556 		bytes_sent += m->m_pkthdr.len;
3557 		mcast_sent += !!(m->m_flags & M_MCAST);
3558 		avail = TXQ_AVAIL(txq);
3559 
3560 		txq->ift_db_pending += (txq->ift_in_use - in_use_prev);
3561 		desc_used += (txq->ift_in_use - in_use_prev);
3562 		ETHER_BPF_MTAP(ifp, m);
3563 		if (__predict_false(!(ifp->if_drv_flags & IFF_DRV_RUNNING)))
3564 			break;
3565 		rang = iflib_txd_db_check(ctx, txq, false, in_use_prev);
3566 	}
3567 
3568 	/* deliberate use of bitwise or to avoid gratuitous short-circuit */
3569 	ring = rang ? false  : (iflib_min_tx_latency | err) || (TXQ_AVAIL(txq) < MAX_TX_DESC(ctx));
3570 	iflib_txd_db_check(ctx, txq, ring, txq->ift_in_use);
3571 	if_inc_counter(ifp, IFCOUNTER_OBYTES, bytes_sent);
3572 	if_inc_counter(ifp, IFCOUNTER_OPACKETS, pkt_sent);
3573 	if (mcast_sent)
3574 		if_inc_counter(ifp, IFCOUNTER_OMCASTS, mcast_sent);
3575 #ifdef INVARIANTS
3576 	if (iflib_verbose_debug)
3577 		printf("consumed=%d\n", consumed);
3578 #endif
3579 	return (consumed);
3580 }
3581 
3582 static uint32_t
3583 iflib_txq_drain_always(struct ifmp_ring *r)
3584 {
3585 	return (1);
3586 }
3587 
3588 static uint32_t
3589 iflib_txq_drain_free(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx)
3590 {
3591 	int i, avail;
3592 	struct mbuf **mp;
3593 	iflib_txq_t txq;
3594 
3595 	txq = r->cookie;
3596 
3597 	txq->ift_qstatus = IFLIB_QUEUE_IDLE;
3598 	CALLOUT_LOCK(txq);
3599 	callout_stop(&txq->ift_timer);
3600 	CALLOUT_UNLOCK(txq);
3601 
3602 	avail = IDXDIFF(pidx, cidx, r->size);
3603 	for (i = 0; i < avail; i++) {
3604 		mp = _ring_peek_one(r, cidx, i, avail - i);
3605 		if (__predict_false(*mp == (struct mbuf *)txq))
3606 			continue;
3607 		m_freem(*mp);
3608 	}
3609 	MPASS(ifmp_ring_is_stalled(r) == 0);
3610 	return (avail);
3611 }
3612 
3613 static void
3614 iflib_ifmp_purge(iflib_txq_t txq)
3615 {
3616 	struct ifmp_ring *r;
3617 
3618 	r = txq->ift_br;
3619 	r->drain = iflib_txq_drain_free;
3620 	r->can_drain = iflib_txq_drain_always;
3621 
3622 	ifmp_ring_check_drainage(r, r->size);
3623 
3624 	r->drain = iflib_txq_drain;
3625 	r->can_drain = iflib_txq_can_drain;
3626 }
3627 
3628 static void
3629 _task_fn_tx(void *context)
3630 {
3631 	iflib_txq_t txq = context;
3632 	if_ctx_t ctx = txq->ift_ctx;
3633 	struct ifnet *ifp = ctx->ifc_ifp;
3634 	int rc;
3635 
3636 #ifdef IFLIB_DIAGNOSTICS
3637 	txq->ift_cpu_exec_count[curcpu]++;
3638 #endif
3639 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
3640 		return;
3641 	if (if_getcapenable(ifp) & IFCAP_NETMAP) {
3642 		if (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, false))
3643 			netmap_tx_irq(ifp, txq->ift_id);
3644 		IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3645 		return;
3646 	}
3647 	if (txq->ift_db_pending)
3648 		ifmp_ring_enqueue(txq->ift_br, (void **)&txq, 1, TX_BATCH_SIZE);
3649 	ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3650 	if (ctx->ifc_flags & IFC_LEGACY)
3651 		IFDI_INTR_ENABLE(ctx);
3652 	else {
3653 		rc = IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id);
3654 		KASSERT(rc != ENOTSUP, ("MSI-X support requires queue_intr_enable, but not implemented in driver"));
3655 	}
3656 }
3657 
3658 static void
3659 _task_fn_rx(void *context)
3660 {
3661 	iflib_rxq_t rxq = context;
3662 	if_ctx_t ctx = rxq->ifr_ctx;
3663 	bool more;
3664 	int rc;
3665 	uint16_t budget;
3666 
3667 #ifdef IFLIB_DIAGNOSTICS
3668 	rxq->ifr_cpu_exec_count[curcpu]++;
3669 #endif
3670 	DBG_COUNTER_INC(task_fn_rxs);
3671 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3672 		return;
3673 	more = true;
3674 #ifdef DEV_NETMAP
3675 	if (if_getcapenable(ctx->ifc_ifp) & IFCAP_NETMAP) {
3676 		u_int work = 0;
3677 		if (netmap_rx_irq(ctx->ifc_ifp, rxq->ifr_id, &work)) {
3678 			more = false;
3679 		}
3680 	}
3681 #endif
3682 	budget = ctx->ifc_sysctl_rx_budget;
3683 	if (budget == 0)
3684 		budget = 16;	/* XXX */
3685 	if (more == false || (more = iflib_rxeof(rxq, budget)) == false) {
3686 		if (ctx->ifc_flags & IFC_LEGACY)
3687 			IFDI_INTR_ENABLE(ctx);
3688 		else {
3689 			DBG_COUNTER_INC(rx_intr_enables);
3690 			rc = IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id);
3691 			KASSERT(rc != ENOTSUP, ("MSI-X support requires queue_intr_enable, but not implemented in driver"));
3692 		}
3693 	}
3694 	if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)))
3695 		return;
3696 	if (more)
3697 		GROUPTASK_ENQUEUE(&rxq->ifr_task);
3698 }
3699 
3700 static void
3701 _task_fn_admin(void *context)
3702 {
3703 	if_ctx_t ctx = context;
3704 	if_softc_ctx_t sctx = &ctx->ifc_softc_ctx;
3705 	iflib_txq_t txq;
3706 	int i;
3707 
3708 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)) {
3709 		if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) {
3710 			return;
3711 		}
3712 	}
3713 
3714 	CTX_LOCK(ctx);
3715 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) {
3716 		CALLOUT_LOCK(txq);
3717 		callout_stop(&txq->ift_timer);
3718 		CALLOUT_UNLOCK(txq);
3719 	}
3720 	IFDI_UPDATE_ADMIN_STATUS(ctx);
3721 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
3722 		callout_reset_on(&txq->ift_timer, hz/2, iflib_timer, txq, txq->ift_timer.c_cpu);
3723 	IFDI_LINK_INTR_ENABLE(ctx);
3724 	if (ctx->ifc_flags & IFC_DO_RESET) {
3725 		ctx->ifc_flags &= ~IFC_DO_RESET;
3726 		iflib_if_init_locked(ctx);
3727 	}
3728 	CTX_UNLOCK(ctx);
3729 
3730 	if (LINK_ACTIVE(ctx) == 0)
3731 		return;
3732 	for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++)
3733 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
3734 }
3735 
3736 
3737 static void
3738 _task_fn_iov(void *context)
3739 {
3740 	if_ctx_t ctx = context;
3741 
3742 	if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))
3743 		return;
3744 
3745 	CTX_LOCK(ctx);
3746 	IFDI_VFLR_HANDLE(ctx);
3747 	CTX_UNLOCK(ctx);
3748 }
3749 
3750 static int
3751 iflib_sysctl_int_delay(SYSCTL_HANDLER_ARGS)
3752 {
3753 	int err;
3754 	if_int_delay_info_t info;
3755 	if_ctx_t ctx;
3756 
3757 	info = (if_int_delay_info_t)arg1;
3758 	ctx = info->iidi_ctx;
3759 	info->iidi_req = req;
3760 	info->iidi_oidp = oidp;
3761 	CTX_LOCK(ctx);
3762 	err = IFDI_SYSCTL_INT_DELAY(ctx, info);
3763 	CTX_UNLOCK(ctx);
3764 	return (err);
3765 }
3766 
3767 /*********************************************************************
3768  *
3769  *  IFNET FUNCTIONS
3770  *
3771  **********************************************************************/
3772 
3773 static void
3774 iflib_if_init_locked(if_ctx_t ctx)
3775 {
3776 	iflib_stop(ctx);
3777 	iflib_init_locked(ctx);
3778 }
3779 
3780 
3781 static void
3782 iflib_if_init(void *arg)
3783 {
3784 	if_ctx_t ctx = arg;
3785 
3786 	CTX_LOCK(ctx);
3787 	iflib_if_init_locked(ctx);
3788 	CTX_UNLOCK(ctx);
3789 }
3790 
3791 static int
3792 iflib_if_transmit(if_t ifp, struct mbuf *m)
3793 {
3794 	if_ctx_t	ctx = if_getsoftc(ifp);
3795 
3796 	iflib_txq_t txq;
3797 	int err, qidx;
3798 
3799 	if (__predict_false((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0 || !LINK_ACTIVE(ctx))) {
3800 		DBG_COUNTER_INC(tx_frees);
3801 		m_freem(m);
3802 		return (ENOBUFS);
3803 	}
3804 
3805 	MPASS(m->m_nextpkt == NULL);
3806 	qidx = 0;
3807 	if ((NTXQSETS(ctx) > 1) && M_HASHTYPE_GET(m))
3808 		qidx = QIDX(ctx, m);
3809 	/*
3810 	 * XXX calculate buf_ring based on flowid (divvy up bits?)
3811 	 */
3812 	txq = &ctx->ifc_txqs[qidx];
3813 
3814 #ifdef DRIVER_BACKPRESSURE
3815 	if (txq->ift_closed) {
3816 		while (m != NULL) {
3817 			next = m->m_nextpkt;
3818 			m->m_nextpkt = NULL;
3819 			m_freem(m);
3820 			m = next;
3821 		}
3822 		return (ENOBUFS);
3823 	}
3824 #endif
3825 #ifdef notyet
3826 	qidx = count = 0;
3827 	mp = marr;
3828 	next = m;
3829 	do {
3830 		count++;
3831 		next = next->m_nextpkt;
3832 	} while (next != NULL);
3833 
3834 	if (count > nitems(marr))
3835 		if ((mp = malloc(count*sizeof(struct mbuf *), M_IFLIB, M_NOWAIT)) == NULL) {
3836 			/* XXX check nextpkt */
3837 			m_freem(m);
3838 			/* XXX simplify for now */
3839 			DBG_COUNTER_INC(tx_frees);
3840 			return (ENOBUFS);
3841 		}
3842 	for (next = m, i = 0; next != NULL; i++) {
3843 		mp[i] = next;
3844 		next = next->m_nextpkt;
3845 		mp[i]->m_nextpkt = NULL;
3846 	}
3847 #endif
3848 	DBG_COUNTER_INC(tx_seen);
3849 	err = ifmp_ring_enqueue(txq->ift_br, (void **)&m, 1, TX_BATCH_SIZE);
3850 
3851 	GROUPTASK_ENQUEUE(&txq->ift_task);
3852 	if (err) {
3853 		/* support forthcoming later */
3854 #ifdef DRIVER_BACKPRESSURE
3855 		txq->ift_closed = TRUE;
3856 #endif
3857 		ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE);
3858 		m_freem(m);
3859 	}
3860 
3861 	return (err);
3862 }
3863 
3864 static void
3865 iflib_if_qflush(if_t ifp)
3866 {
3867 	if_ctx_t ctx = if_getsoftc(ifp);
3868 	iflib_txq_t txq = ctx->ifc_txqs;
3869 	int i;
3870 
3871 	CTX_LOCK(ctx);
3872 	ctx->ifc_flags |= IFC_QFLUSH;
3873 	CTX_UNLOCK(ctx);
3874 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
3875 		while (!(ifmp_ring_is_idle(txq->ift_br) || ifmp_ring_is_stalled(txq->ift_br)))
3876 			iflib_txq_check_drain(txq, 0);
3877 	CTX_LOCK(ctx);
3878 	ctx->ifc_flags &= ~IFC_QFLUSH;
3879 	CTX_UNLOCK(ctx);
3880 
3881 	if_qflush(ifp);
3882 }
3883 
3884 
3885 #define IFCAP_FLAGS (IFCAP_TXCSUM_IPV6 | IFCAP_RXCSUM_IPV6 | IFCAP_HWCSUM | IFCAP_LRO | \
3886 		     IFCAP_TSO4 | IFCAP_TSO6 | IFCAP_VLAN_HWTAGGING | IFCAP_HWSTATS | \
3887 		     IFCAP_VLAN_MTU | IFCAP_VLAN_HWFILTER | IFCAP_VLAN_HWTSO)
3888 
3889 static int
3890 iflib_if_ioctl(if_t ifp, u_long command, caddr_t data)
3891 {
3892 	if_ctx_t ctx = if_getsoftc(ifp);
3893 	struct ifreq	*ifr = (struct ifreq *)data;
3894 #if defined(INET) || defined(INET6)
3895 	struct ifaddr	*ifa = (struct ifaddr *)data;
3896 #endif
3897 	bool		avoid_reset = FALSE;
3898 	int		err = 0, reinit = 0, bits;
3899 
3900 	switch (command) {
3901 	case SIOCSIFADDR:
3902 #ifdef INET
3903 		if (ifa->ifa_addr->sa_family == AF_INET)
3904 			avoid_reset = TRUE;
3905 #endif
3906 #ifdef INET6
3907 		if (ifa->ifa_addr->sa_family == AF_INET6)
3908 			avoid_reset = TRUE;
3909 #endif
3910 		/*
3911 		** Calling init results in link renegotiation,
3912 		** so we avoid doing it when possible.
3913 		*/
3914 		if (avoid_reset) {
3915 			if_setflagbits(ifp, IFF_UP,0);
3916 			if (!(if_getdrvflags(ifp)& IFF_DRV_RUNNING))
3917 				reinit = 1;
3918 #ifdef INET
3919 			if (!(if_getflags(ifp) & IFF_NOARP))
3920 				arp_ifinit(ifp, ifa);
3921 #endif
3922 		} else
3923 			err = ether_ioctl(ifp, command, data);
3924 		break;
3925 	case SIOCSIFMTU:
3926 		CTX_LOCK(ctx);
3927 		if (ifr->ifr_mtu == if_getmtu(ifp)) {
3928 			CTX_UNLOCK(ctx);
3929 			break;
3930 		}
3931 		bits = if_getdrvflags(ifp);
3932 		/* stop the driver and free any clusters before proceeding */
3933 		iflib_stop(ctx);
3934 
3935 		if ((err = IFDI_MTU_SET(ctx, ifr->ifr_mtu)) == 0) {
3936 			if (ifr->ifr_mtu > ctx->ifc_max_fl_buf_size)
3937 				ctx->ifc_flags |= IFC_MULTISEG;
3938 			else
3939 				ctx->ifc_flags &= ~IFC_MULTISEG;
3940 			err = if_setmtu(ifp, ifr->ifr_mtu);
3941 		}
3942 		iflib_init_locked(ctx);
3943 		if_setdrvflags(ifp, bits);
3944 		CTX_UNLOCK(ctx);
3945 		break;
3946 	case SIOCSIFFLAGS:
3947 		CTX_LOCK(ctx);
3948 		if (if_getflags(ifp) & IFF_UP) {
3949 			if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
3950 				if ((if_getflags(ifp) ^ ctx->ifc_if_flags) &
3951 				    (IFF_PROMISC | IFF_ALLMULTI)) {
3952 					err = IFDI_PROMISC_SET(ctx, if_getflags(ifp));
3953 				}
3954 			} else
3955 				reinit = 1;
3956 		} else if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
3957 			iflib_stop(ctx);
3958 		}
3959 		ctx->ifc_if_flags = if_getflags(ifp);
3960 		CTX_UNLOCK(ctx);
3961 		break;
3962 	case SIOCADDMULTI:
3963 	case SIOCDELMULTI:
3964 		if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) {
3965 			CTX_LOCK(ctx);
3966 			IFDI_INTR_DISABLE(ctx);
3967 			IFDI_MULTI_SET(ctx);
3968 			IFDI_INTR_ENABLE(ctx);
3969 			CTX_UNLOCK(ctx);
3970 		}
3971 		break;
3972 	case SIOCSIFMEDIA:
3973 		CTX_LOCK(ctx);
3974 		IFDI_MEDIA_SET(ctx);
3975 		CTX_UNLOCK(ctx);
3976 		/* falls thru */
3977 	case SIOCGIFMEDIA:
3978 	case SIOCGIFXMEDIA:
3979 		err = ifmedia_ioctl(ifp, ifr, &ctx->ifc_media, command);
3980 		break;
3981 	case SIOCGI2C:
3982 	{
3983 		struct ifi2creq i2c;
3984 
3985 		err = copyin(ifr->ifr_data, &i2c, sizeof(i2c));
3986 		if (err != 0)
3987 			break;
3988 		if (i2c.dev_addr != 0xA0 && i2c.dev_addr != 0xA2) {
3989 			err = EINVAL;
3990 			break;
3991 		}
3992 		if (i2c.len > sizeof(i2c.data)) {
3993 			err = EINVAL;
3994 			break;
3995 		}
3996 
3997 		if ((err = IFDI_I2C_REQ(ctx, &i2c)) == 0)
3998 			err = copyout(&i2c, ifr->ifr_data, sizeof(i2c));
3999 		break;
4000 	}
4001 	case SIOCSIFCAP:
4002 	{
4003 		int mask, setmask;
4004 
4005 		mask = ifr->ifr_reqcap ^ if_getcapenable(ifp);
4006 		setmask = 0;
4007 #ifdef TCP_OFFLOAD
4008 		setmask |= mask & (IFCAP_TOE4|IFCAP_TOE6);
4009 #endif
4010 		setmask |= (mask & IFCAP_FLAGS);
4011 
4012 		if (setmask  & (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6))
4013 			setmask |= (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6);
4014 		if ((mask & IFCAP_WOL) &&
4015 		    (if_getcapabilities(ifp) & IFCAP_WOL) != 0)
4016 			setmask |= (mask & (IFCAP_WOL_MCAST|IFCAP_WOL_MAGIC));
4017 		if_vlancap(ifp);
4018 		/*
4019 		 * want to ensure that traffic has stopped before we change any of the flags
4020 		 */
4021 		if (setmask) {
4022 			CTX_LOCK(ctx);
4023 			bits = if_getdrvflags(ifp);
4024 			if (bits & IFF_DRV_RUNNING)
4025 				iflib_stop(ctx);
4026 			if_togglecapenable(ifp, setmask);
4027 			if (bits & IFF_DRV_RUNNING)
4028 				iflib_init_locked(ctx);
4029 			if_setdrvflags(ifp, bits);
4030 			CTX_UNLOCK(ctx);
4031 		}
4032 		break;
4033 	    }
4034 	case SIOCGPRIVATE_0:
4035 	case SIOCSDRVSPEC:
4036 	case SIOCGDRVSPEC:
4037 		CTX_LOCK(ctx);
4038 		err = IFDI_PRIV_IOCTL(ctx, command, data);
4039 		CTX_UNLOCK(ctx);
4040 		break;
4041 	default:
4042 		err = ether_ioctl(ifp, command, data);
4043 		break;
4044 	}
4045 	if (reinit)
4046 		iflib_if_init(ctx);
4047 	return (err);
4048 }
4049 
4050 static uint64_t
4051 iflib_if_get_counter(if_t ifp, ift_counter cnt)
4052 {
4053 	if_ctx_t ctx = if_getsoftc(ifp);
4054 
4055 	return (IFDI_GET_COUNTER(ctx, cnt));
4056 }
4057 
4058 /*********************************************************************
4059  *
4060  *  OTHER FUNCTIONS EXPORTED TO THE STACK
4061  *
4062  **********************************************************************/
4063 
4064 static void
4065 iflib_vlan_register(void *arg, if_t ifp, uint16_t vtag)
4066 {
4067 	if_ctx_t ctx = if_getsoftc(ifp);
4068 
4069 	if ((void *)ctx != arg)
4070 		return;
4071 
4072 	if ((vtag == 0) || (vtag > 4095))
4073 		return;
4074 
4075 	CTX_LOCK(ctx);
4076 	IFDI_VLAN_REGISTER(ctx, vtag);
4077 	/* Re-init to load the changes */
4078 	if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
4079 		iflib_if_init_locked(ctx);
4080 	CTX_UNLOCK(ctx);
4081 }
4082 
4083 static void
4084 iflib_vlan_unregister(void *arg, if_t ifp, uint16_t vtag)
4085 {
4086 	if_ctx_t ctx = if_getsoftc(ifp);
4087 
4088 	if ((void *)ctx != arg)
4089 		return;
4090 
4091 	if ((vtag == 0) || (vtag > 4095))
4092 		return;
4093 
4094 	CTX_LOCK(ctx);
4095 	IFDI_VLAN_UNREGISTER(ctx, vtag);
4096 	/* Re-init to load the changes */
4097 	if (if_getcapenable(ifp) & IFCAP_VLAN_HWFILTER)
4098 		iflib_if_init_locked(ctx);
4099 	CTX_UNLOCK(ctx);
4100 }
4101 
4102 static void
4103 iflib_led_func(void *arg, int onoff)
4104 {
4105 	if_ctx_t ctx = arg;
4106 
4107 	CTX_LOCK(ctx);
4108 	IFDI_LED_FUNC(ctx, onoff);
4109 	CTX_UNLOCK(ctx);
4110 }
4111 
4112 /*********************************************************************
4113  *
4114  *  BUS FUNCTION DEFINITIONS
4115  *
4116  **********************************************************************/
4117 
4118 int
4119 iflib_device_probe(device_t dev)
4120 {
4121 	pci_vendor_info_t *ent;
4122 
4123 	uint16_t	pci_vendor_id, pci_device_id;
4124 	uint16_t	pci_subvendor_id, pci_subdevice_id;
4125 	uint16_t	pci_rev_id;
4126 	if_shared_ctx_t sctx;
4127 
4128 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4129 		return (ENOTSUP);
4130 
4131 	pci_vendor_id = pci_get_vendor(dev);
4132 	pci_device_id = pci_get_device(dev);
4133 	pci_subvendor_id = pci_get_subvendor(dev);
4134 	pci_subdevice_id = pci_get_subdevice(dev);
4135 	pci_rev_id = pci_get_revid(dev);
4136 	if (sctx->isc_parse_devinfo != NULL)
4137 		sctx->isc_parse_devinfo(&pci_device_id, &pci_subvendor_id, &pci_subdevice_id, &pci_rev_id);
4138 
4139 	ent = sctx->isc_vendor_info;
4140 	while (ent->pvi_vendor_id != 0) {
4141 		if (pci_vendor_id != ent->pvi_vendor_id) {
4142 			ent++;
4143 			continue;
4144 		}
4145 		if ((pci_device_id == ent->pvi_device_id) &&
4146 		    ((pci_subvendor_id == ent->pvi_subvendor_id) ||
4147 		     (ent->pvi_subvendor_id == 0)) &&
4148 		    ((pci_subdevice_id == ent->pvi_subdevice_id) ||
4149 		     (ent->pvi_subdevice_id == 0)) &&
4150 		    ((pci_rev_id == ent->pvi_rev_id) ||
4151 		     (ent->pvi_rev_id == 0))) {
4152 
4153 			device_set_desc_copy(dev, ent->pvi_name);
4154 			/* this needs to be changed to zero if the bus probing code
4155 			 * ever stops re-probing on best match because the sctx
4156 			 * may have its values over written by register calls
4157 			 * in subsequent probes
4158 			 */
4159 			return (BUS_PROBE_DEFAULT);
4160 		}
4161 		ent++;
4162 	}
4163 	return (ENXIO);
4164 }
4165 
4166 int
4167 iflib_device_register(device_t dev, void *sc, if_shared_ctx_t sctx, if_ctx_t *ctxp)
4168 {
4169 	int err, rid, msix, msix_bar;
4170 	if_ctx_t ctx;
4171 	if_t ifp;
4172 	if_softc_ctx_t scctx;
4173 	int i;
4174 	uint16_t main_txq;
4175 	uint16_t main_rxq;
4176 
4177 
4178 	ctx = malloc(sizeof(* ctx), M_IFLIB, M_WAITOK|M_ZERO);
4179 
4180 	if (sc == NULL) {
4181 		sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO);
4182 		device_set_softc(dev, ctx);
4183 		ctx->ifc_flags |= IFC_SC_ALLOCATED;
4184 	}
4185 
4186 	ctx->ifc_sctx = sctx;
4187 	ctx->ifc_dev = dev;
4188 	ctx->ifc_softc = sc;
4189 
4190 	if ((err = iflib_register(ctx)) != 0) {
4191 		device_printf(dev, "iflib_register failed %d\n", err);
4192 		return (err);
4193 	}
4194 	iflib_add_device_sysctl_pre(ctx);
4195 
4196 	scctx = &ctx->ifc_softc_ctx;
4197 	ifp = ctx->ifc_ifp;
4198 
4199 	/*
4200 	 * XXX sanity check that ntxd & nrxd are a power of 2
4201 	 */
4202 	if (ctx->ifc_sysctl_ntxqs != 0)
4203 		scctx->isc_ntxqsets = ctx->ifc_sysctl_ntxqs;
4204 	if (ctx->ifc_sysctl_nrxqs != 0)
4205 		scctx->isc_nrxqsets = ctx->ifc_sysctl_nrxqs;
4206 
4207 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4208 		if (ctx->ifc_sysctl_ntxds[i] != 0)
4209 			scctx->isc_ntxd[i] = ctx->ifc_sysctl_ntxds[i];
4210 		else
4211 			scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i];
4212 	}
4213 
4214 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4215 		if (ctx->ifc_sysctl_nrxds[i] != 0)
4216 			scctx->isc_nrxd[i] = ctx->ifc_sysctl_nrxds[i];
4217 		else
4218 			scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i];
4219 	}
4220 
4221 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4222 		if (scctx->isc_nrxd[i] < sctx->isc_nrxd_min[i]) {
4223 			device_printf(dev, "nrxd%d: %d less than nrxd_min %d - resetting to min\n",
4224 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_min[i]);
4225 			scctx->isc_nrxd[i] = sctx->isc_nrxd_min[i];
4226 		}
4227 		if (scctx->isc_nrxd[i] > sctx->isc_nrxd_max[i]) {
4228 			device_printf(dev, "nrxd%d: %d greater than nrxd_max %d - resetting to max\n",
4229 				      i, scctx->isc_nrxd[i], sctx->isc_nrxd_max[i]);
4230 			scctx->isc_nrxd[i] = sctx->isc_nrxd_max[i];
4231 		}
4232 	}
4233 
4234 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4235 		if (scctx->isc_ntxd[i] < sctx->isc_ntxd_min[i]) {
4236 			device_printf(dev, "ntxd%d: %d less than ntxd_min %d - resetting to min\n",
4237 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_min[i]);
4238 			scctx->isc_ntxd[i] = sctx->isc_ntxd_min[i];
4239 		}
4240 		if (scctx->isc_ntxd[i] > sctx->isc_ntxd_max[i]) {
4241 			device_printf(dev, "ntxd%d: %d greater than ntxd_max %d - resetting to max\n",
4242 				      i, scctx->isc_ntxd[i], sctx->isc_ntxd_max[i]);
4243 			scctx->isc_ntxd[i] = sctx->isc_ntxd_max[i];
4244 		}
4245 	}
4246 
4247 	if ((err = IFDI_ATTACH_PRE(ctx)) != 0) {
4248 		device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err);
4249 		return (err);
4250 	}
4251 	_iflib_pre_assert(scctx);
4252 	ctx->ifc_txrx = *scctx->isc_txrx;
4253 
4254 #ifdef INVARIANTS
4255 	MPASS(scctx->isc_capenable);
4256 	if (scctx->isc_capenable & IFCAP_TXCSUM)
4257 		MPASS(scctx->isc_tx_csum_flags);
4258 #endif
4259 
4260 	if_setcapabilities(ifp, scctx->isc_capenable | IFCAP_HWSTATS);
4261 	if_setcapenable(ifp, scctx->isc_capenable | IFCAP_HWSTATS);
4262 
4263 	if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets))
4264 		scctx->isc_ntxqsets = scctx->isc_ntxqsets_max;
4265 	if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets))
4266 		scctx->isc_nrxqsets = scctx->isc_nrxqsets_max;
4267 
4268 #ifdef ACPI_DMAR
4269 	if (dmar_get_dma_tag(device_get_parent(dev), dev) != NULL)
4270 		ctx->ifc_flags |= IFC_DMAR;
4271 #elif !(defined(__i386__) || defined(__amd64__))
4272 	/* set unconditionally for !x86 */
4273 	ctx->ifc_flags |= IFC_DMAR;
4274 #endif
4275 
4276 	msix_bar = scctx->isc_msix_bar;
4277 	main_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0;
4278 	main_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0;
4279 
4280 	/* XXX change for per-queue sizes */
4281 	device_printf(dev, "using %d tx descriptors and %d rx descriptors\n",
4282 		      scctx->isc_ntxd[main_txq], scctx->isc_nrxd[main_rxq]);
4283 	for (i = 0; i < sctx->isc_nrxqs; i++) {
4284 		if (!powerof2(scctx->isc_nrxd[i])) {
4285 			/* round down instead? */
4286 			device_printf(dev, "# rx descriptors must be a power of 2\n");
4287 			err = EINVAL;
4288 			goto fail;
4289 		}
4290 	}
4291 	for (i = 0; i < sctx->isc_ntxqs; i++) {
4292 		if (!powerof2(scctx->isc_ntxd[i])) {
4293 			device_printf(dev,
4294 			    "# tx descriptors must be a power of 2");
4295 			err = EINVAL;
4296 			goto fail;
4297 		}
4298 	}
4299 
4300 	if (scctx->isc_tx_nsegments > scctx->isc_ntxd[main_txq] /
4301 	    MAX_SINGLE_PACKET_FRACTION)
4302 		scctx->isc_tx_nsegments = max(1, scctx->isc_ntxd[main_txq] /
4303 		    MAX_SINGLE_PACKET_FRACTION);
4304 	if (scctx->isc_tx_tso_segments_max > scctx->isc_ntxd[main_txq] /
4305 	    MAX_SINGLE_PACKET_FRACTION)
4306 		scctx->isc_tx_tso_segments_max = max(1,
4307 		    scctx->isc_ntxd[main_txq] / MAX_SINGLE_PACKET_FRACTION);
4308 
4309 	/*
4310 	 * Protect the stack against modern hardware
4311 	 */
4312 	if (scctx->isc_tx_tso_size_max > FREEBSD_TSO_SIZE_MAX)
4313 		scctx->isc_tx_tso_size_max = FREEBSD_TSO_SIZE_MAX;
4314 
4315 	/* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */
4316 	ifp->if_hw_tsomaxsegcount = scctx->isc_tx_tso_segments_max;
4317 	ifp->if_hw_tsomax = scctx->isc_tx_tso_size_max;
4318 	ifp->if_hw_tsomaxsegsize = scctx->isc_tx_tso_segsize_max;
4319 	if (scctx->isc_rss_table_size == 0)
4320 		scctx->isc_rss_table_size = 64;
4321 	scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1;
4322 
4323 	GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx);
4324 	/* XXX format name */
4325 	taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx, -1, "admin");
4326 
4327 	/* Set up cpu set.  If it fails, use the set of all CPUs. */
4328 	if (bus_get_cpus(dev, INTR_CPUS, sizeof(ctx->ifc_cpus), &ctx->ifc_cpus) != 0) {
4329 		device_printf(dev, "Unable to fetch CPU list\n");
4330 		CPU_COPY(&all_cpus, &ctx->ifc_cpus);
4331 	}
4332 	MPASS(CPU_COUNT(&ctx->ifc_cpus) > 0);
4333 
4334 	/*
4335 	** Now setup MSI or MSI/X, should
4336 	** return us the number of supported
4337 	** vectors. (Will be 1 for MSI)
4338 	*/
4339 	if (sctx->isc_flags & IFLIB_SKIP_MSIX) {
4340 		msix = scctx->isc_vectors;
4341 	} else if (scctx->isc_msix_bar != 0)
4342 	       /*
4343 		* The simple fact that isc_msix_bar is not 0 does not mean we
4344 		* we have a good value there that is known to work.
4345 		*/
4346 		msix = iflib_msix_init(ctx);
4347 	else {
4348 		scctx->isc_vectors = 1;
4349 		scctx->isc_ntxqsets = 1;
4350 		scctx->isc_nrxqsets = 1;
4351 		scctx->isc_intr = IFLIB_INTR_LEGACY;
4352 		msix = 0;
4353 	}
4354 	/* Get memory for the station queues */
4355 	if ((err = iflib_queues_alloc(ctx))) {
4356 		device_printf(dev, "Unable to allocate queue memory\n");
4357 		goto fail;
4358 	}
4359 
4360 	if ((err = iflib_qset_structures_setup(ctx))) {
4361 		device_printf(dev, "qset structure setup failed %d\n", err);
4362 		goto fail_queues;
4363 	}
4364 
4365 	/*
4366 	 * Group taskqueues aren't properly set up until SMP is started,
4367 	 * so we disable interrupts until we can handle them post
4368 	 * SI_SUB_SMP.
4369 	 *
4370 	 * XXX: disabling interrupts doesn't actually work, at least for
4371 	 * the non-MSI case.  When they occur before SI_SUB_SMP completes,
4372 	 * we do null handling and depend on this not causing too large an
4373 	 * interrupt storm.
4374 	 */
4375 	IFDI_INTR_DISABLE(ctx);
4376 	if (msix > 1 && (err = IFDI_MSIX_INTR_ASSIGN(ctx, msix)) != 0) {
4377 		device_printf(dev, "IFDI_MSIX_INTR_ASSIGN failed %d\n", err);
4378 		goto fail_intr_free;
4379 	}
4380 	if (msix <= 1) {
4381 		rid = 0;
4382 		if (scctx->isc_intr == IFLIB_INTR_MSI) {
4383 			MPASS(msix == 1);
4384 			rid = 1;
4385 		}
4386 		if ((err = iflib_legacy_setup(ctx, ctx->isc_legacy_intr, ctx->ifc_softc, &rid, "irq0")) != 0) {
4387 			device_printf(dev, "iflib_legacy_setup failed %d\n", err);
4388 			goto fail_intr_free;
4389 		}
4390 	}
4391 	ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac);
4392 	if ((err = IFDI_ATTACH_POST(ctx)) != 0) {
4393 		device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err);
4394 		goto fail_detach;
4395 	}
4396 	if ((err = iflib_netmap_attach(ctx))) {
4397 		device_printf(ctx->ifc_dev, "netmap attach failed: %d\n", err);
4398 		goto fail_detach;
4399 	}
4400 	*ctxp = ctx;
4401 
4402 	if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter);
4403 	iflib_add_device_sysctl_post(ctx);
4404 	ctx->ifc_flags |= IFC_INIT_DONE;
4405 	return (0);
4406 fail_detach:
4407 	ether_ifdetach(ctx->ifc_ifp);
4408 fail_intr_free:
4409 	if (scctx->isc_intr == IFLIB_INTR_MSIX || scctx->isc_intr == IFLIB_INTR_MSI)
4410 		pci_release_msi(ctx->ifc_dev);
4411 fail_queues:
4412 	/* XXX free queues */
4413 fail:
4414 	IFDI_DETACH(ctx);
4415 	return (err);
4416 }
4417 
4418 int
4419 iflib_device_attach(device_t dev)
4420 {
4421 	if_ctx_t ctx;
4422 	if_shared_ctx_t sctx;
4423 
4424 	if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC)
4425 		return (ENOTSUP);
4426 
4427 	pci_enable_busmaster(dev);
4428 
4429 	return (iflib_device_register(dev, NULL, sctx, &ctx));
4430 }
4431 
4432 int
4433 iflib_device_deregister(if_ctx_t ctx)
4434 {
4435 	if_t ifp = ctx->ifc_ifp;
4436 	iflib_txq_t txq;
4437 	iflib_rxq_t rxq;
4438 	device_t dev = ctx->ifc_dev;
4439 	int i, j;
4440 	struct taskqgroup *tqg;
4441 	iflib_fl_t fl;
4442 
4443 	/* Make sure VLANS are not using driver */
4444 	if (if_vlantrunkinuse(ifp)) {
4445 		device_printf(dev,"Vlan in use, detach first\n");
4446 		return (EBUSY);
4447 	}
4448 
4449 	CTX_LOCK(ctx);
4450 	ctx->ifc_in_detach = 1;
4451 	iflib_stop(ctx);
4452 	CTX_UNLOCK(ctx);
4453 
4454 	/* Unregister VLAN events */
4455 	if (ctx->ifc_vlan_attach_event != NULL)
4456 		EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event);
4457 	if (ctx->ifc_vlan_detach_event != NULL)
4458 		EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event);
4459 
4460 	iflib_netmap_detach(ifp);
4461 	ether_ifdetach(ifp);
4462 	/* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/
4463 	CTX_LOCK_DESTROY(ctx);
4464 	if (ctx->ifc_led_dev != NULL)
4465 		led_destroy(ctx->ifc_led_dev);
4466 	/* XXX drain any dependent tasks */
4467 	tqg = qgroup_if_io_tqg;
4468 	for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) {
4469 		callout_drain(&txq->ift_timer);
4470 		if (txq->ift_task.gt_uniq != NULL)
4471 			taskqgroup_detach(tqg, &txq->ift_task);
4472 	}
4473 	for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) {
4474 		if (rxq->ifr_task.gt_uniq != NULL)
4475 			taskqgroup_detach(tqg, &rxq->ifr_task);
4476 
4477 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
4478 			free(fl->ifl_rx_bitmap, M_IFLIB);
4479 
4480 	}
4481 	tqg = qgroup_if_config_tqg;
4482 	if (ctx->ifc_admin_task.gt_uniq != NULL)
4483 		taskqgroup_detach(tqg, &ctx->ifc_admin_task);
4484 	if (ctx->ifc_vflr_task.gt_uniq != NULL)
4485 		taskqgroup_detach(tqg, &ctx->ifc_vflr_task);
4486 
4487 	IFDI_DETACH(ctx);
4488 	device_set_softc(ctx->ifc_dev, NULL);
4489 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_LEGACY) {
4490 		pci_release_msi(dev);
4491 	}
4492 	if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_MSIX) {
4493 		iflib_irq_free(ctx, &ctx->ifc_legacy_irq);
4494 	}
4495 	if (ctx->ifc_msix_mem != NULL) {
4496 		bus_release_resource(ctx->ifc_dev, SYS_RES_MEMORY,
4497 			ctx->ifc_softc_ctx.isc_msix_bar, ctx->ifc_msix_mem);
4498 		ctx->ifc_msix_mem = NULL;
4499 	}
4500 
4501 	bus_generic_detach(dev);
4502 	if_free(ifp);
4503 
4504 	iflib_tx_structures_free(ctx);
4505 	iflib_rx_structures_free(ctx);
4506 	if (ctx->ifc_flags & IFC_SC_ALLOCATED)
4507 		free(ctx->ifc_softc, M_IFLIB);
4508 	free(ctx, M_IFLIB);
4509 	return (0);
4510 }
4511 
4512 
4513 int
4514 iflib_device_detach(device_t dev)
4515 {
4516 	if_ctx_t ctx = device_get_softc(dev);
4517 
4518 	return (iflib_device_deregister(ctx));
4519 }
4520 
4521 int
4522 iflib_device_suspend(device_t dev)
4523 {
4524 	if_ctx_t ctx = device_get_softc(dev);
4525 
4526 	CTX_LOCK(ctx);
4527 	IFDI_SUSPEND(ctx);
4528 	CTX_UNLOCK(ctx);
4529 
4530 	return bus_generic_suspend(dev);
4531 }
4532 int
4533 iflib_device_shutdown(device_t dev)
4534 {
4535 	if_ctx_t ctx = device_get_softc(dev);
4536 
4537 	CTX_LOCK(ctx);
4538 	IFDI_SHUTDOWN(ctx);
4539 	CTX_UNLOCK(ctx);
4540 
4541 	return bus_generic_suspend(dev);
4542 }
4543 
4544 
4545 int
4546 iflib_device_resume(device_t dev)
4547 {
4548 	if_ctx_t ctx = device_get_softc(dev);
4549 	iflib_txq_t txq = ctx->ifc_txqs;
4550 
4551 	CTX_LOCK(ctx);
4552 	IFDI_RESUME(ctx);
4553 	iflib_init_locked(ctx);
4554 	CTX_UNLOCK(ctx);
4555 	for (int i = 0; i < NTXQSETS(ctx); i++, txq++)
4556 		iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET);
4557 
4558 	return (bus_generic_resume(dev));
4559 }
4560 
4561 int
4562 iflib_device_iov_init(device_t dev, uint16_t num_vfs, const nvlist_t *params)
4563 {
4564 	int error;
4565 	if_ctx_t ctx = device_get_softc(dev);
4566 
4567 	CTX_LOCK(ctx);
4568 	error = IFDI_IOV_INIT(ctx, num_vfs, params);
4569 	CTX_UNLOCK(ctx);
4570 
4571 	return (error);
4572 }
4573 
4574 void
4575 iflib_device_iov_uninit(device_t dev)
4576 {
4577 	if_ctx_t ctx = device_get_softc(dev);
4578 
4579 	CTX_LOCK(ctx);
4580 	IFDI_IOV_UNINIT(ctx);
4581 	CTX_UNLOCK(ctx);
4582 }
4583 
4584 int
4585 iflib_device_iov_add_vf(device_t dev, uint16_t vfnum, const nvlist_t *params)
4586 {
4587 	int error;
4588 	if_ctx_t ctx = device_get_softc(dev);
4589 
4590 	CTX_LOCK(ctx);
4591 	error = IFDI_IOV_VF_ADD(ctx, vfnum, params);
4592 	CTX_UNLOCK(ctx);
4593 
4594 	return (error);
4595 }
4596 
4597 /*********************************************************************
4598  *
4599  *  MODULE FUNCTION DEFINITIONS
4600  *
4601  **********************************************************************/
4602 
4603 /*
4604  * - Start a fast taskqueue thread for each core
4605  * - Start a taskqueue for control operations
4606  */
4607 static int
4608 iflib_module_init(void)
4609 {
4610 	return (0);
4611 }
4612 
4613 static int
4614 iflib_module_event_handler(module_t mod, int what, void *arg)
4615 {
4616 	int err;
4617 
4618 	switch (what) {
4619 	case MOD_LOAD:
4620 		if ((err = iflib_module_init()) != 0)
4621 			return (err);
4622 		break;
4623 	case MOD_UNLOAD:
4624 		return (EBUSY);
4625 	default:
4626 		return (EOPNOTSUPP);
4627 	}
4628 
4629 	return (0);
4630 }
4631 
4632 /*********************************************************************
4633  *
4634  *  PUBLIC FUNCTION DEFINITIONS
4635  *     ordered as in iflib.h
4636  *
4637  **********************************************************************/
4638 
4639 
4640 static void
4641 _iflib_assert(if_shared_ctx_t sctx)
4642 {
4643 	MPASS(sctx->isc_tx_maxsize);
4644 	MPASS(sctx->isc_tx_maxsegsize);
4645 
4646 	MPASS(sctx->isc_rx_maxsize);
4647 	MPASS(sctx->isc_rx_nsegments);
4648 	MPASS(sctx->isc_rx_maxsegsize);
4649 
4650 	MPASS(sctx->isc_nrxd_min[0]);
4651 	MPASS(sctx->isc_nrxd_max[0]);
4652 	MPASS(sctx->isc_nrxd_default[0]);
4653 	MPASS(sctx->isc_ntxd_min[0]);
4654 	MPASS(sctx->isc_ntxd_max[0]);
4655 	MPASS(sctx->isc_ntxd_default[0]);
4656 }
4657 
4658 static void
4659 _iflib_pre_assert(if_softc_ctx_t scctx)
4660 {
4661 
4662 	MPASS(scctx->isc_txrx->ift_txd_encap);
4663 	MPASS(scctx->isc_txrx->ift_txd_flush);
4664 	MPASS(scctx->isc_txrx->ift_txd_credits_update);
4665 	MPASS(scctx->isc_txrx->ift_rxd_available);
4666 	MPASS(scctx->isc_txrx->ift_rxd_pkt_get);
4667 	MPASS(scctx->isc_txrx->ift_rxd_refill);
4668 	MPASS(scctx->isc_txrx->ift_rxd_flush);
4669 }
4670 
4671 static int
4672 iflib_register(if_ctx_t ctx)
4673 {
4674 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4675 	driver_t *driver = sctx->isc_driver;
4676 	device_t dev = ctx->ifc_dev;
4677 	if_t ifp;
4678 
4679 	_iflib_assert(sctx);
4680 
4681 	CTX_LOCK_INIT(ctx, device_get_nameunit(ctx->ifc_dev));
4682 
4683 	ifp = ctx->ifc_ifp = if_gethandle(IFT_ETHER);
4684 	if (ifp == NULL) {
4685 		device_printf(dev, "can not allocate ifnet structure\n");
4686 		return (ENOMEM);
4687 	}
4688 
4689 	/*
4690 	 * Initialize our context's device specific methods
4691 	 */
4692 	kobj_init((kobj_t) ctx, (kobj_class_t) driver);
4693 	kobj_class_compile((kobj_class_t) driver);
4694 	driver->refs++;
4695 
4696 	if_initname(ifp, device_get_name(dev), device_get_unit(dev));
4697 	if_setsoftc(ifp, ctx);
4698 	if_setdev(ifp, dev);
4699 	if_setinitfn(ifp, iflib_if_init);
4700 	if_setioctlfn(ifp, iflib_if_ioctl);
4701 	if_settransmitfn(ifp, iflib_if_transmit);
4702 	if_setqflushfn(ifp, iflib_if_qflush);
4703 	if_setflags(ifp, IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST);
4704 
4705 	ctx->ifc_vlan_attach_event =
4706 		EVENTHANDLER_REGISTER(vlan_config, iflib_vlan_register, ctx,
4707 							  EVENTHANDLER_PRI_FIRST);
4708 	ctx->ifc_vlan_detach_event =
4709 		EVENTHANDLER_REGISTER(vlan_unconfig, iflib_vlan_unregister, ctx,
4710 							  EVENTHANDLER_PRI_FIRST);
4711 
4712 	ifmedia_init(&ctx->ifc_media, IFM_IMASK,
4713 					 iflib_media_change, iflib_media_status);
4714 
4715 	return (0);
4716 }
4717 
4718 
4719 static int
4720 iflib_queues_alloc(if_ctx_t ctx)
4721 {
4722 	if_shared_ctx_t sctx = ctx->ifc_sctx;
4723 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
4724 	device_t dev = ctx->ifc_dev;
4725 	int nrxqsets = scctx->isc_nrxqsets;
4726 	int ntxqsets = scctx->isc_ntxqsets;
4727 	iflib_txq_t txq;
4728 	iflib_rxq_t rxq;
4729 	iflib_fl_t fl = NULL;
4730 	int i, j, cpu, err, txconf, rxconf;
4731 	iflib_dma_info_t ifdip;
4732 	uint32_t *rxqsizes = scctx->isc_rxqsizes;
4733 	uint32_t *txqsizes = scctx->isc_txqsizes;
4734 	uint8_t nrxqs = sctx->isc_nrxqs;
4735 	uint8_t ntxqs = sctx->isc_ntxqs;
4736 	int nfree_lists = sctx->isc_nfl ? sctx->isc_nfl : 1;
4737 	caddr_t *vaddrs;
4738 	uint64_t *paddrs;
4739 	struct ifmp_ring **brscp;
4740 
4741 	KASSERT(ntxqs > 0, ("number of queues per qset must be at least 1"));
4742 	KASSERT(nrxqs > 0, ("number of queues per qset must be at least 1"));
4743 
4744 	brscp = NULL;
4745 	txq = NULL;
4746 	rxq = NULL;
4747 
4748 /* Allocate the TX ring struct memory */
4749 	if (!(txq =
4750 	    (iflib_txq_t) mallocarray(ntxqsets, sizeof(struct iflib_txq),
4751 	        M_IFLIB, M_NOWAIT | M_ZERO))) {
4752 		device_printf(dev, "Unable to allocate TX ring memory\n");
4753 		err = ENOMEM;
4754 		goto fail;
4755 	}
4756 
4757 	/* Now allocate the RX */
4758 	if (!(rxq =
4759 	    (iflib_rxq_t) mallocarray(nrxqsets, sizeof(struct iflib_rxq),
4760 	        M_IFLIB, M_NOWAIT | M_ZERO))) {
4761 		device_printf(dev, "Unable to allocate RX ring memory\n");
4762 		err = ENOMEM;
4763 		goto rx_fail;
4764 	}
4765 
4766 	ctx->ifc_txqs = txq;
4767 	ctx->ifc_rxqs = rxq;
4768 
4769 	/*
4770 	 * XXX handle allocation failure
4771 	 */
4772 	for (txconf = i = 0, cpu = CPU_FIRST(); i < ntxqsets; i++, txconf++, txq++, cpu = CPU_NEXT(cpu)) {
4773 		/* Set up some basics */
4774 
4775 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * ntxqs, M_IFLIB, M_WAITOK|M_ZERO)) == NULL) {
4776 			device_printf(dev, "failed to allocate iflib_dma_info\n");
4777 			err = ENOMEM;
4778 			goto err_tx_desc;
4779 		}
4780 		txq->ift_ifdi = ifdip;
4781 		for (j = 0; j < ntxqs; j++, ifdip++) {
4782 			if (iflib_dma_alloc(ctx, txqsizes[j], ifdip, BUS_DMA_NOWAIT)) {
4783 				device_printf(dev, "Unable to allocate Descriptor memory\n");
4784 				err = ENOMEM;
4785 				goto err_tx_desc;
4786 			}
4787 			txq->ift_txd_size[j] = scctx->isc_txd_size[j];
4788 			bzero((void *)ifdip->idi_vaddr, txqsizes[j]);
4789 		}
4790 		txq->ift_ctx = ctx;
4791 		txq->ift_id = i;
4792 		if (sctx->isc_flags & IFLIB_HAS_TXCQ) {
4793 			txq->ift_br_offset = 1;
4794 		} else {
4795 			txq->ift_br_offset = 0;
4796 		}
4797 		/* XXX fix this */
4798 		txq->ift_timer.c_cpu = cpu;
4799 
4800 		if (iflib_txsd_alloc(txq)) {
4801 			device_printf(dev, "Critical Failure setting up TX buffers\n");
4802 			err = ENOMEM;
4803 			goto err_tx_desc;
4804 		}
4805 
4806 		/* Initialize the TX lock */
4807 		snprintf(txq->ift_mtx_name, MTX_NAME_LEN, "%s:tx(%d):callout",
4808 		    device_get_nameunit(dev), txq->ift_id);
4809 		mtx_init(&txq->ift_mtx, txq->ift_mtx_name, NULL, MTX_DEF);
4810 		callout_init_mtx(&txq->ift_timer, &txq->ift_mtx, 0);
4811 
4812 		snprintf(txq->ift_db_mtx_name, MTX_NAME_LEN, "%s:tx(%d):db",
4813 			 device_get_nameunit(dev), txq->ift_id);
4814 
4815 		err = ifmp_ring_alloc(&txq->ift_br, 2048, txq, iflib_txq_drain,
4816 				      iflib_txq_can_drain, M_IFLIB, M_WAITOK);
4817 		if (err) {
4818 			/* XXX free any allocated rings */
4819 			device_printf(dev, "Unable to allocate buf_ring\n");
4820 			goto err_tx_desc;
4821 		}
4822 	}
4823 
4824 	for (rxconf = i = 0; i < nrxqsets; i++, rxconf++, rxq++) {
4825 		/* Set up some basics */
4826 
4827 		if ((ifdip = malloc(sizeof(struct iflib_dma_info) * nrxqs, M_IFLIB, M_WAITOK|M_ZERO)) == NULL) {
4828 			device_printf(dev, "failed to allocate iflib_dma_info\n");
4829 			err = ENOMEM;
4830 			goto err_tx_desc;
4831 		}
4832 
4833 		rxq->ifr_ifdi = ifdip;
4834 		/* XXX this needs to be changed if #rx queues != #tx queues */
4835 		rxq->ifr_ntxqirq = 1;
4836 		rxq->ifr_txqid[0] = i;
4837 		for (j = 0; j < nrxqs; j++, ifdip++) {
4838 			if (iflib_dma_alloc(ctx, rxqsizes[j], ifdip, BUS_DMA_NOWAIT)) {
4839 				device_printf(dev, "Unable to allocate Descriptor memory\n");
4840 				err = ENOMEM;
4841 				goto err_tx_desc;
4842 			}
4843 			bzero((void *)ifdip->idi_vaddr, rxqsizes[j]);
4844 		}
4845 		rxq->ifr_ctx = ctx;
4846 		rxq->ifr_id = i;
4847 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
4848 			rxq->ifr_fl_offset = 1;
4849 		} else {
4850 			rxq->ifr_fl_offset = 0;
4851 		}
4852 		rxq->ifr_nfl = nfree_lists;
4853 		if (!(fl =
4854 			  (iflib_fl_t) mallocarray(nfree_lists, sizeof(struct iflib_fl),
4855 			      M_IFLIB, M_NOWAIT | M_ZERO))) {
4856 			device_printf(dev, "Unable to allocate free list memory\n");
4857 			err = ENOMEM;
4858 			goto err_tx_desc;
4859 		}
4860 		rxq->ifr_fl = fl;
4861 		for (j = 0; j < nfree_lists; j++) {
4862 			fl[j].ifl_rxq = rxq;
4863 			fl[j].ifl_id = j;
4864 			fl[j].ifl_ifdi = &rxq->ifr_ifdi[j + rxq->ifr_fl_offset];
4865 			fl[j].ifl_rxd_size = scctx->isc_rxd_size[j];
4866 		}
4867         /* Allocate receive buffers for the ring*/
4868 		if (iflib_rxsd_alloc(rxq)) {
4869 			device_printf(dev,
4870 			    "Critical Failure setting up receive buffers\n");
4871 			err = ENOMEM;
4872 			goto err_rx_desc;
4873 		}
4874 
4875 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++)
4876 			fl->ifl_rx_bitmap = bit_alloc(fl->ifl_size, M_IFLIB, M_WAITOK|M_ZERO);
4877 	}
4878 
4879 	/* TXQs */
4880 	vaddrs = malloc(sizeof(caddr_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
4881 	paddrs = malloc(sizeof(uint64_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK);
4882 	for (i = 0; i < ntxqsets; i++) {
4883 		iflib_dma_info_t di = ctx->ifc_txqs[i].ift_ifdi;
4884 
4885 		for (j = 0; j < ntxqs; j++, di++) {
4886 			vaddrs[i*ntxqs + j] = di->idi_vaddr;
4887 			paddrs[i*ntxqs + j] = di->idi_paddr;
4888 		}
4889 	}
4890 	if ((err = IFDI_TX_QUEUES_ALLOC(ctx, vaddrs, paddrs, ntxqs, ntxqsets)) != 0) {
4891 		device_printf(ctx->ifc_dev, "device queue allocation failed\n");
4892 		iflib_tx_structures_free(ctx);
4893 		free(vaddrs, M_IFLIB);
4894 		free(paddrs, M_IFLIB);
4895 		goto err_rx_desc;
4896 	}
4897 	free(vaddrs, M_IFLIB);
4898 	free(paddrs, M_IFLIB);
4899 
4900 	/* RXQs */
4901 	vaddrs = malloc(sizeof(caddr_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
4902 	paddrs = malloc(sizeof(uint64_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK);
4903 	for (i = 0; i < nrxqsets; i++) {
4904 		iflib_dma_info_t di = ctx->ifc_rxqs[i].ifr_ifdi;
4905 
4906 		for (j = 0; j < nrxqs; j++, di++) {
4907 			vaddrs[i*nrxqs + j] = di->idi_vaddr;
4908 			paddrs[i*nrxqs + j] = di->idi_paddr;
4909 		}
4910 	}
4911 	if ((err = IFDI_RX_QUEUES_ALLOC(ctx, vaddrs, paddrs, nrxqs, nrxqsets)) != 0) {
4912 		device_printf(ctx->ifc_dev, "device queue allocation failed\n");
4913 		iflib_tx_structures_free(ctx);
4914 		free(vaddrs, M_IFLIB);
4915 		free(paddrs, M_IFLIB);
4916 		goto err_rx_desc;
4917 	}
4918 	free(vaddrs, M_IFLIB);
4919 	free(paddrs, M_IFLIB);
4920 
4921 	return (0);
4922 
4923 /* XXX handle allocation failure changes */
4924 err_rx_desc:
4925 err_tx_desc:
4926 	if (ctx->ifc_rxqs != NULL)
4927 		free(ctx->ifc_rxqs, M_IFLIB);
4928 	ctx->ifc_rxqs = NULL;
4929 	if (ctx->ifc_txqs != NULL)
4930 		free(ctx->ifc_txqs, M_IFLIB);
4931 	ctx->ifc_txqs = NULL;
4932 rx_fail:
4933 	if (brscp != NULL)
4934 		free(brscp, M_IFLIB);
4935 	if (rxq != NULL)
4936 		free(rxq, M_IFLIB);
4937 	if (txq != NULL)
4938 		free(txq, M_IFLIB);
4939 fail:
4940 	return (err);
4941 }
4942 
4943 static int
4944 iflib_tx_structures_setup(if_ctx_t ctx)
4945 {
4946 	iflib_txq_t txq = ctx->ifc_txqs;
4947 	int i;
4948 
4949 	for (i = 0; i < NTXQSETS(ctx); i++, txq++)
4950 		iflib_txq_setup(txq);
4951 
4952 	return (0);
4953 }
4954 
4955 static void
4956 iflib_tx_structures_free(if_ctx_t ctx)
4957 {
4958 	iflib_txq_t txq = ctx->ifc_txqs;
4959 	int i, j;
4960 
4961 	for (i = 0; i < NTXQSETS(ctx); i++, txq++) {
4962 		iflib_txq_destroy(txq);
4963 		for (j = 0; j < ctx->ifc_nhwtxqs; j++)
4964 			iflib_dma_free(&txq->ift_ifdi[j]);
4965 	}
4966 	free(ctx->ifc_txqs, M_IFLIB);
4967 	ctx->ifc_txqs = NULL;
4968 	IFDI_QUEUES_FREE(ctx);
4969 }
4970 
4971 /*********************************************************************
4972  *
4973  *  Initialize all receive rings.
4974  *
4975  **********************************************************************/
4976 static int
4977 iflib_rx_structures_setup(if_ctx_t ctx)
4978 {
4979 	iflib_rxq_t rxq = ctx->ifc_rxqs;
4980 	int q;
4981 #if defined(INET6) || defined(INET)
4982 	int i, err;
4983 #endif
4984 
4985 	for (q = 0; q < ctx->ifc_softc_ctx.isc_nrxqsets; q++, rxq++) {
4986 #if defined(INET6) || defined(INET)
4987 		tcp_lro_free(&rxq->ifr_lc);
4988 		if ((err = tcp_lro_init_args(&rxq->ifr_lc, ctx->ifc_ifp,
4989 		    TCP_LRO_ENTRIES, min(1024,
4990 		    ctx->ifc_softc_ctx.isc_nrxd[rxq->ifr_fl_offset]))) != 0) {
4991 			device_printf(ctx->ifc_dev, "LRO Initialization failed!\n");
4992 			goto fail;
4993 		}
4994 		rxq->ifr_lro_enabled = TRUE;
4995 #endif
4996 		IFDI_RXQ_SETUP(ctx, rxq->ifr_id);
4997 	}
4998 	return (0);
4999 #if defined(INET6) || defined(INET)
5000 fail:
5001 	/*
5002 	 * Free RX software descriptors allocated so far, we will only handle
5003 	 * the rings that completed, the failing case will have
5004 	 * cleaned up for itself. 'q' failed, so its the terminus.
5005 	 */
5006 	rxq = ctx->ifc_rxqs;
5007 	for (i = 0; i < q; ++i, rxq++) {
5008 		iflib_rx_sds_free(rxq);
5009 		rxq->ifr_cq_gen = rxq->ifr_cq_cidx = rxq->ifr_cq_pidx = 0;
5010 	}
5011 	return (err);
5012 #endif
5013 }
5014 
5015 /*********************************************************************
5016  *
5017  *  Free all receive rings.
5018  *
5019  **********************************************************************/
5020 static void
5021 iflib_rx_structures_free(if_ctx_t ctx)
5022 {
5023 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5024 
5025 	for (int i = 0; i < ctx->ifc_softc_ctx.isc_nrxqsets; i++, rxq++) {
5026 		iflib_rx_sds_free(rxq);
5027 	}
5028 }
5029 
5030 static int
5031 iflib_qset_structures_setup(if_ctx_t ctx)
5032 {
5033 	int err;
5034 
5035 	if ((err = iflib_tx_structures_setup(ctx)) != 0)
5036 		return (err);
5037 
5038 	if ((err = iflib_rx_structures_setup(ctx)) != 0) {
5039 		device_printf(ctx->ifc_dev, "iflib_rx_structures_setup failed: %d\n", err);
5040 		iflib_tx_structures_free(ctx);
5041 		iflib_rx_structures_free(ctx);
5042 	}
5043 	return (err);
5044 }
5045 
5046 int
5047 iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid,
5048 				driver_filter_t filter, void *filter_arg, driver_intr_t handler, void *arg, char *name)
5049 {
5050 
5051 	return (_iflib_irq_alloc(ctx, irq, rid, filter, handler, arg, name));
5052 }
5053 
5054 #ifdef SMP
5055 static int
5056 find_nth(if_ctx_t ctx, int qid)
5057 {
5058 	cpuset_t cpus;
5059 	int i, cpuid, eqid, count;
5060 
5061 	CPU_COPY(&ctx->ifc_cpus, &cpus);
5062 	count = CPU_COUNT(&cpus);
5063 	eqid = qid % count;
5064 	/* clear up to the qid'th bit */
5065 	for (i = 0; i < eqid; i++) {
5066 		cpuid = CPU_FFS(&cpus);
5067 		MPASS(cpuid != 0);
5068 		CPU_CLR(cpuid-1, &cpus);
5069 	}
5070 	cpuid = CPU_FFS(&cpus);
5071 	MPASS(cpuid != 0);
5072 	return (cpuid-1);
5073 }
5074 
5075 #ifdef SCHED_ULE
5076 extern struct cpu_group *cpu_top;              /* CPU topology */
5077 
5078 static int
5079 find_child_with_core(int cpu, struct cpu_group *grp)
5080 {
5081 	int i;
5082 
5083 	if (grp->cg_children == 0)
5084 		return -1;
5085 
5086 	MPASS(grp->cg_child);
5087 	for (i = 0; i < grp->cg_children; i++) {
5088 		if (CPU_ISSET(cpu, &grp->cg_child[i].cg_mask))
5089 			return i;
5090 	}
5091 
5092 	return -1;
5093 }
5094 
5095 /*
5096  * Find the nth thread on the specified core
5097  */
5098 static int
5099 find_thread(int cpu, int thread_num)
5100 {
5101 	struct cpu_group *grp;
5102 	int i;
5103 	cpuset_t cs;
5104 
5105 	grp = cpu_top;
5106 	if (grp == NULL)
5107 		return cpu;
5108 	i = 0;
5109 	while ((i = find_child_with_core(cpu, grp)) != -1) {
5110 		/* If the child only has one cpu, don't descend */
5111 		if (grp->cg_child[i].cg_count <= 1)
5112 			break;
5113 		grp = &grp->cg_child[i];
5114 	}
5115 
5116 	/* If they don't share at least an L2 cache, use the same CPU */
5117 	if (grp->cg_level > CG_SHARE_L2 || grp->cg_level == CG_SHARE_NONE)
5118 		return cpu;
5119 
5120 	/* Now pick one */
5121 	CPU_COPY(&grp->cg_mask, &cs);
5122 	for (i = thread_num % grp->cg_count; i > 0; i--) {
5123 		MPASS(CPU_FFS(&cs));
5124 		CPU_CLR(CPU_FFS(&cs) - 1, &cs);
5125 	}
5126 	MPASS(CPU_FFS(&cs));
5127 	return CPU_FFS(&cs) - 1;
5128 }
5129 #else
5130 static int
5131 find_thread(int cpu, int thread_num __unused)
5132 {
5133 	return cpu;
5134 }
5135 #endif
5136 
5137 static int
5138 get_thread_num(if_ctx_t ctx, iflib_intr_type_t type, int qid)
5139 {
5140 	switch (type) {
5141 	case IFLIB_INTR_TX:
5142 		/* TX queues get threads on the same core as the corresponding RX queue */
5143 		/* XXX handle multiple RX threads per core and more than two threads per core */
5144 		return qid / CPU_COUNT(&ctx->ifc_cpus) + 1;
5145 	case IFLIB_INTR_RX:
5146 	case IFLIB_INTR_RXTX:
5147 		/* RX queues get the first thread on their core */
5148 		return qid / CPU_COUNT(&ctx->ifc_cpus);
5149 	default:
5150 		return -1;
5151 	}
5152 }
5153 #else
5154 #define get_thread_num(ctx, type, qid)	CPU_FIRST()
5155 #define find_thread(cpuid, tid)		CPU_FIRST()
5156 #define find_nth(ctx, gid)		CPU_FIRST()
5157 #endif
5158 
5159 /* Just to avoid copy/paste */
5160 static inline int
5161 iflib_irq_set_affinity(if_ctx_t ctx, int irq, iflib_intr_type_t type, int qid,
5162     struct grouptask *gtask, struct taskqgroup *tqg, void *uniq, char *name)
5163 {
5164 	int cpuid;
5165 	int err, tid;
5166 
5167 	cpuid = find_nth(ctx, qid);
5168 	tid = get_thread_num(ctx, type, qid);
5169 	MPASS(tid >= 0);
5170 	cpuid = find_thread(cpuid, tid);
5171 	err = taskqgroup_attach_cpu(tqg, gtask, uniq, cpuid, irq, name);
5172 	if (err) {
5173 		device_printf(ctx->ifc_dev, "taskqgroup_attach_cpu failed %d\n", err);
5174 		return (err);
5175 	}
5176 #ifdef notyet
5177 	if (cpuid > ctx->ifc_cpuid_highest)
5178 		ctx->ifc_cpuid_highest = cpuid;
5179 #endif
5180 	return 0;
5181 }
5182 
5183 int
5184 iflib_irq_alloc_generic(if_ctx_t ctx, if_irq_t irq, int rid,
5185 						iflib_intr_type_t type, driver_filter_t *filter,
5186 						void *filter_arg, int qid, char *name)
5187 {
5188 	struct grouptask *gtask;
5189 	struct taskqgroup *tqg;
5190 	iflib_filter_info_t info;
5191 	gtask_fn_t *fn;
5192 	int tqrid, err;
5193 	driver_filter_t *intr_fast;
5194 	void *q;
5195 
5196 	info = &ctx->ifc_filter_info;
5197 	tqrid = rid;
5198 
5199 	switch (type) {
5200 	/* XXX merge tx/rx for netmap? */
5201 	case IFLIB_INTR_TX:
5202 		q = &ctx->ifc_txqs[qid];
5203 		info = &ctx->ifc_txqs[qid].ift_filter_info;
5204 		gtask = &ctx->ifc_txqs[qid].ift_task;
5205 		tqg = qgroup_if_io_tqg;
5206 		fn = _task_fn_tx;
5207 		intr_fast = iflib_fast_intr;
5208 		GROUPTASK_INIT(gtask, 0, fn, q);
5209 		break;
5210 	case IFLIB_INTR_RX:
5211 		q = &ctx->ifc_rxqs[qid];
5212 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
5213 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5214 		tqg = qgroup_if_io_tqg;
5215 		fn = _task_fn_rx;
5216 		intr_fast = iflib_fast_intr;
5217 		GROUPTASK_INIT(gtask, 0, fn, q);
5218 		break;
5219 	case IFLIB_INTR_RXTX:
5220 		q = &ctx->ifc_rxqs[qid];
5221 		info = &ctx->ifc_rxqs[qid].ifr_filter_info;
5222 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5223 		tqg = qgroup_if_io_tqg;
5224 		fn = _task_fn_rx;
5225 		intr_fast = iflib_fast_intr_rxtx;
5226 		GROUPTASK_INIT(gtask, 0, fn, q);
5227 		break;
5228 	case IFLIB_INTR_ADMIN:
5229 		q = ctx;
5230 		tqrid = -1;
5231 		info = &ctx->ifc_filter_info;
5232 		gtask = &ctx->ifc_admin_task;
5233 		tqg = qgroup_if_config_tqg;
5234 		fn = _task_fn_admin;
5235 		intr_fast = iflib_fast_intr_ctx;
5236 		break;
5237 	default:
5238 		panic("unknown net intr type");
5239 	}
5240 
5241 	info->ifi_filter = filter;
5242 	info->ifi_filter_arg = filter_arg;
5243 	info->ifi_task = gtask;
5244 	info->ifi_ctx = q;
5245 
5246 	err = _iflib_irq_alloc(ctx, irq, rid, intr_fast, NULL, info,  name);
5247 	if (err != 0) {
5248 		device_printf(ctx->ifc_dev, "_iflib_irq_alloc failed %d\n", err);
5249 		return (err);
5250 	}
5251 	if (type == IFLIB_INTR_ADMIN)
5252 		return (0);
5253 
5254 	if (tqrid != -1) {
5255 		err = iflib_irq_set_affinity(ctx, rman_get_start(irq->ii_res), type, qid, gtask, tqg, q, name);
5256 		if (err)
5257 			return (err);
5258 	} else {
5259 		taskqgroup_attach(tqg, gtask, q, rman_get_start(irq->ii_res), name);
5260 	}
5261 
5262 	return (0);
5263 }
5264 
5265 void
5266 iflib_softirq_alloc_generic(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type,  void *arg, int qid, char *name)
5267 {
5268 	struct grouptask *gtask;
5269 	struct taskqgroup *tqg;
5270 	gtask_fn_t *fn;
5271 	void *q;
5272 	int irq_num = -1;
5273 	int err;
5274 
5275 	switch (type) {
5276 	case IFLIB_INTR_TX:
5277 		q = &ctx->ifc_txqs[qid];
5278 		gtask = &ctx->ifc_txqs[qid].ift_task;
5279 		tqg = qgroup_if_io_tqg;
5280 		fn = _task_fn_tx;
5281 		if (irq != NULL)
5282 			irq_num = rman_get_start(irq->ii_res);
5283 		break;
5284 	case IFLIB_INTR_RX:
5285 		q = &ctx->ifc_rxqs[qid];
5286 		gtask = &ctx->ifc_rxqs[qid].ifr_task;
5287 		tqg = qgroup_if_io_tqg;
5288 		fn = _task_fn_rx;
5289 		if (irq != NULL)
5290 			irq_num = rman_get_start(irq->ii_res);
5291 		break;
5292 	case IFLIB_INTR_IOV:
5293 		q = ctx;
5294 		gtask = &ctx->ifc_vflr_task;
5295 		tqg = qgroup_if_config_tqg;
5296 		fn = _task_fn_iov;
5297 		break;
5298 	default:
5299 		panic("unknown net intr type");
5300 	}
5301 	GROUPTASK_INIT(gtask, 0, fn, q);
5302 	if (irq_num != -1) {
5303 		err = iflib_irq_set_affinity(ctx, irq_num, type, qid, gtask, tqg, q, name);
5304 		if (err)
5305 			taskqgroup_attach(tqg, gtask, q, irq_num, name);
5306 	}
5307 	else {
5308 		taskqgroup_attach(tqg, gtask, q, irq_num, name);
5309 	}
5310 }
5311 
5312 void
5313 iflib_irq_free(if_ctx_t ctx, if_irq_t irq)
5314 {
5315 	if (irq->ii_tag)
5316 		bus_teardown_intr(ctx->ifc_dev, irq->ii_res, irq->ii_tag);
5317 
5318 	if (irq->ii_res)
5319 		bus_release_resource(ctx->ifc_dev, SYS_RES_IRQ, irq->ii_rid, irq->ii_res);
5320 }
5321 
5322 static int
5323 iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filter_arg, int *rid, char *name)
5324 {
5325 	iflib_txq_t txq = ctx->ifc_txqs;
5326 	iflib_rxq_t rxq = ctx->ifc_rxqs;
5327 	if_irq_t irq = &ctx->ifc_legacy_irq;
5328 	iflib_filter_info_t info;
5329 	struct grouptask *gtask;
5330 	struct taskqgroup *tqg;
5331 	gtask_fn_t *fn;
5332 	int tqrid;
5333 	void *q;
5334 	int err;
5335 
5336 	q = &ctx->ifc_rxqs[0];
5337 	info = &rxq[0].ifr_filter_info;
5338 	gtask = &rxq[0].ifr_task;
5339 	tqg = qgroup_if_io_tqg;
5340 	tqrid = irq->ii_rid = *rid;
5341 	fn = _task_fn_rx;
5342 
5343 	ctx->ifc_flags |= IFC_LEGACY;
5344 	info->ifi_filter = filter;
5345 	info->ifi_filter_arg = filter_arg;
5346 	info->ifi_task = gtask;
5347 	info->ifi_ctx = ctx;
5348 
5349 	/* We allocate a single interrupt resource */
5350 	if ((err = _iflib_irq_alloc(ctx, irq, tqrid, iflib_fast_intr_ctx, NULL, info, name)) != 0)
5351 		return (err);
5352 	GROUPTASK_INIT(gtask, 0, fn, q);
5353 	taskqgroup_attach(tqg, gtask, q, rman_get_start(irq->ii_res), name);
5354 
5355 	GROUPTASK_INIT(&txq->ift_task, 0, _task_fn_tx, txq);
5356 	taskqgroup_attach(qgroup_if_io_tqg, &txq->ift_task, txq, rman_get_start(irq->ii_res), "tx");
5357 	return (0);
5358 }
5359 
5360 void
5361 iflib_led_create(if_ctx_t ctx)
5362 {
5363 
5364 	ctx->ifc_led_dev = led_create(iflib_led_func, ctx,
5365 	    device_get_nameunit(ctx->ifc_dev));
5366 }
5367 
5368 void
5369 iflib_tx_intr_deferred(if_ctx_t ctx, int txqid)
5370 {
5371 
5372 	GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task);
5373 }
5374 
5375 void
5376 iflib_rx_intr_deferred(if_ctx_t ctx, int rxqid)
5377 {
5378 
5379 	GROUPTASK_ENQUEUE(&ctx->ifc_rxqs[rxqid].ifr_task);
5380 }
5381 
5382 void
5383 iflib_admin_intr_deferred(if_ctx_t ctx)
5384 {
5385 #ifdef INVARIANTS
5386 	struct grouptask *gtask;
5387 
5388 	gtask = &ctx->ifc_admin_task;
5389 	MPASS(gtask != NULL && gtask->gt_taskqueue != NULL);
5390 #endif
5391 
5392 	GROUPTASK_ENQUEUE(&ctx->ifc_admin_task);
5393 }
5394 
5395 void
5396 iflib_iov_intr_deferred(if_ctx_t ctx)
5397 {
5398 
5399 	GROUPTASK_ENQUEUE(&ctx->ifc_vflr_task);
5400 }
5401 
5402 void
5403 iflib_io_tqg_attach(struct grouptask *gt, void *uniq, int cpu, char *name)
5404 {
5405 
5406 	taskqgroup_attach_cpu(qgroup_if_io_tqg, gt, uniq, cpu, -1, name);
5407 }
5408 
5409 void
5410 iflib_config_gtask_init(if_ctx_t ctx, struct grouptask *gtask, gtask_fn_t *fn,
5411 	char *name)
5412 {
5413 
5414 	GROUPTASK_INIT(gtask, 0, fn, ctx);
5415 	taskqgroup_attach(qgroup_if_config_tqg, gtask, gtask, -1, name);
5416 }
5417 
5418 void
5419 iflib_config_gtask_deinit(struct grouptask *gtask)
5420 {
5421 
5422 	taskqgroup_detach(qgroup_if_config_tqg, gtask);
5423 }
5424 
5425 void
5426 iflib_link_state_change(if_ctx_t ctx, int link_state, uint64_t baudrate)
5427 {
5428 	if_t ifp = ctx->ifc_ifp;
5429 	iflib_txq_t txq = ctx->ifc_txqs;
5430 
5431 	if_setbaudrate(ifp, baudrate);
5432 	if (baudrate >= IF_Gbps(10))
5433 		ctx->ifc_flags |= IFC_PREFETCH;
5434 
5435 	/* If link down, disable watchdog */
5436 	if ((ctx->ifc_link_state == LINK_STATE_UP) && (link_state == LINK_STATE_DOWN)) {
5437 		for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxqsets; i++, txq++)
5438 			txq->ift_qstatus = IFLIB_QUEUE_IDLE;
5439 	}
5440 	ctx->ifc_link_state = link_state;
5441 	if_link_state_change(ifp, link_state);
5442 }
5443 
5444 static int
5445 iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq)
5446 {
5447 	int credits;
5448 #ifdef INVARIANTS
5449 	int credits_pre = txq->ift_cidx_processed;
5450 #endif
5451 
5452 	if (ctx->isc_txd_credits_update == NULL)
5453 		return (0);
5454 
5455 	if ((credits = ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, true)) == 0)
5456 		return (0);
5457 
5458 	txq->ift_processed += credits;
5459 	txq->ift_cidx_processed += credits;
5460 
5461 	MPASS(credits_pre + credits == txq->ift_cidx_processed);
5462 	if (txq->ift_cidx_processed >= txq->ift_size)
5463 		txq->ift_cidx_processed -= txq->ift_size;
5464 	return (credits);
5465 }
5466 
5467 static int
5468 iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget)
5469 {
5470 
5471 	return (ctx->isc_rxd_available(ctx->ifc_softc, rxq->ifr_id, cidx,
5472 	    budget));
5473 }
5474 
5475 void
5476 iflib_add_int_delay_sysctl(if_ctx_t ctx, const char *name,
5477 	const char *description, if_int_delay_info_t info,
5478 	int offset, int value)
5479 {
5480 	info->iidi_ctx = ctx;
5481 	info->iidi_offset = offset;
5482 	info->iidi_value = value;
5483 	SYSCTL_ADD_PROC(device_get_sysctl_ctx(ctx->ifc_dev),
5484 	    SYSCTL_CHILDREN(device_get_sysctl_tree(ctx->ifc_dev)),
5485 	    OID_AUTO, name, CTLTYPE_INT|CTLFLAG_RW,
5486 	    info, 0, iflib_sysctl_int_delay, "I", description);
5487 }
5488 
5489 struct mtx *
5490 iflib_ctx_lock_get(if_ctx_t ctx)
5491 {
5492 
5493 	return (&ctx->ifc_mtx);
5494 }
5495 
5496 static int
5497 iflib_msix_init(if_ctx_t ctx)
5498 {
5499 	device_t dev = ctx->ifc_dev;
5500 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5501 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5502 	int vectors, queues, rx_queues, tx_queues, queuemsgs, msgs;
5503 	int iflib_num_tx_queues, iflib_num_rx_queues;
5504 	int err, admincnt, bar;
5505 
5506 	iflib_num_tx_queues = ctx->ifc_sysctl_ntxqs;
5507 	iflib_num_rx_queues = ctx->ifc_sysctl_nrxqs;
5508 
5509 	device_printf(dev, "msix_init qsets capped at %d\n", imax(scctx->isc_ntxqsets, scctx->isc_nrxqsets));
5510 
5511 	bar = ctx->ifc_softc_ctx.isc_msix_bar;
5512 	admincnt = sctx->isc_admin_intrcnt;
5513 	/* Override by global tuneable */
5514 	{
5515 		int i;
5516 		size_t len = sizeof(i);
5517 		err = kernel_sysctlbyname(curthread, "hw.pci.enable_msix", &i, &len, NULL, 0, NULL, 0);
5518 		if (err == 0) {
5519 			if (i == 0)
5520 				goto msi;
5521 		}
5522 		else {
5523 			device_printf(dev, "unable to read hw.pci.enable_msix.");
5524 		}
5525 	}
5526 	/* Override by tuneable */
5527 	if (scctx->isc_disable_msix)
5528 		goto msi;
5529 
5530 	/*
5531 	** When used in a virtualized environment
5532 	** PCI BUSMASTER capability may not be set
5533 	** so explicity set it here and rewrite
5534 	** the ENABLE in the MSIX control register
5535 	** at this point to cause the host to
5536 	** successfully initialize us.
5537 	*/
5538 	{
5539 		int msix_ctrl, rid;
5540 
5541  		pci_enable_busmaster(dev);
5542 		rid = 0;
5543 		if (pci_find_cap(dev, PCIY_MSIX, &rid) == 0 && rid != 0) {
5544 			rid += PCIR_MSIX_CTRL;
5545 			msix_ctrl = pci_read_config(dev, rid, 2);
5546 			msix_ctrl |= PCIM_MSIXCTRL_MSIX_ENABLE;
5547 			pci_write_config(dev, rid, msix_ctrl, 2);
5548 		} else {
5549 			device_printf(dev, "PCIY_MSIX capability not found; "
5550 			                   "or rid %d == 0.\n", rid);
5551 			goto msi;
5552 		}
5553 	}
5554 
5555 	/*
5556 	 * bar == -1 => "trust me I know what I'm doing"
5557 	 * Some drivers are for hardware that is so shoddily
5558 	 * documented that no one knows which bars are which
5559 	 * so the developer has to map all bars. This hack
5560 	 * allows shoddy garbage to use msix in this framework.
5561 	 */
5562 	if (bar != -1) {
5563 		ctx->ifc_msix_mem = bus_alloc_resource_any(dev,
5564 	            SYS_RES_MEMORY, &bar, RF_ACTIVE);
5565 		if (ctx->ifc_msix_mem == NULL) {
5566 			/* May not be enabled */
5567 			device_printf(dev, "Unable to map MSIX table \n");
5568 			goto msi;
5569 		}
5570 	}
5571 	/* First try MSI/X */
5572 	if ((msgs = pci_msix_count(dev)) == 0) { /* system has msix disabled */
5573 		device_printf(dev, "System has MSIX disabled \n");
5574 		bus_release_resource(dev, SYS_RES_MEMORY,
5575 		    bar, ctx->ifc_msix_mem);
5576 		ctx->ifc_msix_mem = NULL;
5577 		goto msi;
5578 	}
5579 #if IFLIB_DEBUG
5580 	/* use only 1 qset in debug mode */
5581 	queuemsgs = min(msgs - admincnt, 1);
5582 #else
5583 	queuemsgs = msgs - admincnt;
5584 #endif
5585 #ifdef RSS
5586 	queues = imin(queuemsgs, rss_getnumbuckets());
5587 #else
5588 	queues = queuemsgs;
5589 #endif
5590 	queues = imin(CPU_COUNT(&ctx->ifc_cpus), queues);
5591 	device_printf(dev, "pxm cpus: %d queue msgs: %d admincnt: %d\n",
5592 				  CPU_COUNT(&ctx->ifc_cpus), queuemsgs, admincnt);
5593 #ifdef  RSS
5594 	/* If we're doing RSS, clamp at the number of RSS buckets */
5595 	if (queues > rss_getnumbuckets())
5596 		queues = rss_getnumbuckets();
5597 #endif
5598 	if (iflib_num_rx_queues > 0 && iflib_num_rx_queues < queuemsgs - admincnt)
5599 		rx_queues = iflib_num_rx_queues;
5600 	else
5601 		rx_queues = queues;
5602 
5603 	if (rx_queues > scctx->isc_nrxqsets)
5604 		rx_queues = scctx->isc_nrxqsets;
5605 
5606 	/*
5607 	 * We want this to be all logical CPUs by default
5608 	 */
5609 	if (iflib_num_tx_queues > 0 && iflib_num_tx_queues < queues)
5610 		tx_queues = iflib_num_tx_queues;
5611 	else
5612 		tx_queues = mp_ncpus;
5613 
5614 	if (tx_queues > scctx->isc_ntxqsets)
5615 		tx_queues = scctx->isc_ntxqsets;
5616 
5617 	if (ctx->ifc_sysctl_qs_eq_override == 0) {
5618 #ifdef INVARIANTS
5619 		if (tx_queues != rx_queues)
5620 			device_printf(dev, "queue equality override not set, capping rx_queues at %d and tx_queues at %d\n",
5621 				      min(rx_queues, tx_queues), min(rx_queues, tx_queues));
5622 #endif
5623 		tx_queues = min(rx_queues, tx_queues);
5624 		rx_queues = min(rx_queues, tx_queues);
5625 	}
5626 
5627 	device_printf(dev, "using %d rx queues %d tx queues \n", rx_queues, tx_queues);
5628 
5629 	vectors = rx_queues + admincnt;
5630 	if ((err = pci_alloc_msix(dev, &vectors)) == 0) {
5631 		device_printf(dev,
5632 					  "Using MSIX interrupts with %d vectors\n", vectors);
5633 		scctx->isc_vectors = vectors;
5634 		scctx->isc_nrxqsets = rx_queues;
5635 		scctx->isc_ntxqsets = tx_queues;
5636 		scctx->isc_intr = IFLIB_INTR_MSIX;
5637 
5638 		return (vectors);
5639 	} else {
5640 		device_printf(dev, "failed to allocate %d msix vectors, err: %d - using MSI\n", vectors, err);
5641 	}
5642 msi:
5643 	vectors = pci_msi_count(dev);
5644 	scctx->isc_nrxqsets = 1;
5645 	scctx->isc_ntxqsets = 1;
5646 	scctx->isc_vectors = vectors;
5647 	if (vectors == 1 && pci_alloc_msi(dev, &vectors) == 0) {
5648 		device_printf(dev,"Using an MSI interrupt\n");
5649 		scctx->isc_intr = IFLIB_INTR_MSI;
5650 	} else {
5651 		device_printf(dev,"Using a Legacy interrupt\n");
5652 		scctx->isc_intr = IFLIB_INTR_LEGACY;
5653 	}
5654 
5655 	return (vectors);
5656 }
5657 
5658 char * ring_states[] = { "IDLE", "BUSY", "STALLED", "ABDICATED" };
5659 
5660 static int
5661 mp_ring_state_handler(SYSCTL_HANDLER_ARGS)
5662 {
5663 	int rc;
5664 	uint16_t *state = ((uint16_t *)oidp->oid_arg1);
5665 	struct sbuf *sb;
5666 	char *ring_state = "UNKNOWN";
5667 
5668 	/* XXX needed ? */
5669 	rc = sysctl_wire_old_buffer(req, 0);
5670 	MPASS(rc == 0);
5671 	if (rc != 0)
5672 		return (rc);
5673 	sb = sbuf_new_for_sysctl(NULL, NULL, 80, req);
5674 	MPASS(sb != NULL);
5675 	if (sb == NULL)
5676 		return (ENOMEM);
5677 	if (state[3] <= 3)
5678 		ring_state = ring_states[state[3]];
5679 
5680 	sbuf_printf(sb, "pidx_head: %04hd pidx_tail: %04hd cidx: %04hd state: %s",
5681 		    state[0], state[1], state[2], ring_state);
5682 	rc = sbuf_finish(sb);
5683 	sbuf_delete(sb);
5684         return(rc);
5685 }
5686 
5687 enum iflib_ndesc_handler {
5688 	IFLIB_NTXD_HANDLER,
5689 	IFLIB_NRXD_HANDLER,
5690 };
5691 
5692 static int
5693 mp_ndesc_handler(SYSCTL_HANDLER_ARGS)
5694 {
5695 	if_ctx_t ctx = (void *)arg1;
5696 	enum iflib_ndesc_handler type = arg2;
5697 	char buf[256] = {0};
5698 	qidx_t *ndesc;
5699 	char *p, *next;
5700 	int nqs, rc, i;
5701 
5702 	MPASS(type == IFLIB_NTXD_HANDLER || type == IFLIB_NRXD_HANDLER);
5703 
5704 	nqs = 8;
5705 	switch(type) {
5706 	case IFLIB_NTXD_HANDLER:
5707 		ndesc = ctx->ifc_sysctl_ntxds;
5708 		if (ctx->ifc_sctx)
5709 			nqs = ctx->ifc_sctx->isc_ntxqs;
5710 		break;
5711 	case IFLIB_NRXD_HANDLER:
5712 		ndesc = ctx->ifc_sysctl_nrxds;
5713 		if (ctx->ifc_sctx)
5714 			nqs = ctx->ifc_sctx->isc_nrxqs;
5715 		break;
5716 	}
5717 	if (nqs == 0)
5718 		nqs = 8;
5719 
5720 	for (i=0; i<8; i++) {
5721 		if (i >= nqs)
5722 			break;
5723 		if (i)
5724 			strcat(buf, ",");
5725 		sprintf(strchr(buf, 0), "%d", ndesc[i]);
5726 	}
5727 
5728 	rc = sysctl_handle_string(oidp, buf, sizeof(buf), req);
5729 	if (rc || req->newptr == NULL)
5730 		return rc;
5731 
5732 	for (i = 0, next = buf, p = strsep(&next, " ,"); i < 8 && p;
5733 	    i++, p = strsep(&next, " ,")) {
5734 		ndesc[i] = strtoul(p, NULL, 10);
5735 	}
5736 
5737 	return(rc);
5738 }
5739 
5740 #define NAME_BUFLEN 32
5741 static void
5742 iflib_add_device_sysctl_pre(if_ctx_t ctx)
5743 {
5744         device_t dev = iflib_get_dev(ctx);
5745 	struct sysctl_oid_list *child, *oid_list;
5746 	struct sysctl_ctx_list *ctx_list;
5747 	struct sysctl_oid *node;
5748 
5749 	ctx_list = device_get_sysctl_ctx(dev);
5750 	child = SYSCTL_CHILDREN(device_get_sysctl_tree(dev));
5751 	ctx->ifc_sysctl_node = node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, "iflib",
5752 						      CTLFLAG_RD, NULL, "IFLIB fields");
5753 	oid_list = SYSCTL_CHILDREN(node);
5754 
5755 	SYSCTL_ADD_STRING(ctx_list, oid_list, OID_AUTO, "driver_version",
5756 		       CTLFLAG_RD, ctx->ifc_sctx->isc_driver_version, 0,
5757 		       "driver version");
5758 
5759 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_ntxqs",
5760 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_ntxqs, 0,
5761 			"# of txqs to use, 0 => use default #");
5762 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_nrxqs",
5763 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_nrxqs, 0,
5764 			"# of rxqs to use, 0 => use default #");
5765 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_qs_enable",
5766 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_qs_eq_override, 0,
5767                        "permit #txq != #rxq");
5768 	SYSCTL_ADD_INT(ctx_list, oid_list, OID_AUTO, "disable_msix",
5769                       CTLFLAG_RWTUN, &ctx->ifc_softc_ctx.isc_disable_msix, 0,
5770                       "disable MSIX (default 0)");
5771 	SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "rx_budget",
5772 		       CTLFLAG_RWTUN, &ctx->ifc_sysctl_rx_budget, 0,
5773                        "set the rx budget");
5774 
5775 	/* XXX change for per-queue sizes */
5776 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_ntxds",
5777 		       CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NTXD_HANDLER,
5778                        mp_ndesc_handler, "A",
5779                        "list of # of tx descriptors to use, 0 = use default #");
5780 	SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_nrxds",
5781 		       CTLTYPE_STRING|CTLFLAG_RWTUN, ctx, IFLIB_NRXD_HANDLER,
5782                        mp_ndesc_handler, "A",
5783                        "list of # of rx descriptors to use, 0 = use default #");
5784 }
5785 
5786 static void
5787 iflib_add_device_sysctl_post(if_ctx_t ctx)
5788 {
5789 	if_shared_ctx_t sctx = ctx->ifc_sctx;
5790 	if_softc_ctx_t scctx = &ctx->ifc_softc_ctx;
5791         device_t dev = iflib_get_dev(ctx);
5792 	struct sysctl_oid_list *child;
5793 	struct sysctl_ctx_list *ctx_list;
5794 	iflib_fl_t fl;
5795 	iflib_txq_t txq;
5796 	iflib_rxq_t rxq;
5797 	int i, j;
5798 	char namebuf[NAME_BUFLEN];
5799 	char *qfmt;
5800 	struct sysctl_oid *queue_node, *fl_node, *node;
5801 	struct sysctl_oid_list *queue_list, *fl_list;
5802 	ctx_list = device_get_sysctl_ctx(dev);
5803 
5804 	node = ctx->ifc_sysctl_node;
5805 	child = SYSCTL_CHILDREN(node);
5806 
5807 	if (scctx->isc_ntxqsets > 100)
5808 		qfmt = "txq%03d";
5809 	else if (scctx->isc_ntxqsets > 10)
5810 		qfmt = "txq%02d";
5811 	else
5812 		qfmt = "txq%d";
5813 	for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) {
5814 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
5815 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
5816 					     CTLFLAG_RD, NULL, "Queue Name");
5817 		queue_list = SYSCTL_CHILDREN(queue_node);
5818 #if MEMORY_LOGGING
5819 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_dequeued",
5820 				CTLFLAG_RD,
5821 				&txq->ift_dequeued, "total mbufs freed");
5822 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_enqueued",
5823 				CTLFLAG_RD,
5824 				&txq->ift_enqueued, "total mbufs enqueued");
5825 #endif
5826 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag",
5827 				   CTLFLAG_RD,
5828 				   &txq->ift_mbuf_defrag, "# of times m_defrag was called");
5829 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "m_pullups",
5830 				   CTLFLAG_RD,
5831 				   &txq->ift_pullups, "# of times m_pullup was called");
5832 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag_failed",
5833 				   CTLFLAG_RD,
5834 				   &txq->ift_mbuf_defrag_failed, "# of times m_defrag failed");
5835 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_desc_avail",
5836 				   CTLFLAG_RD,
5837 				   &txq->ift_no_desc_avail, "# of times no descriptors were available");
5838 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "tx_map_failed",
5839 				   CTLFLAG_RD,
5840 				   &txq->ift_map_failed, "# of times dma map failed");
5841 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txd_encap_efbig",
5842 				   CTLFLAG_RD,
5843 				   &txq->ift_txd_encap_efbig, "# of times txd_encap returned EFBIG");
5844 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_tx_dma_setup",
5845 				   CTLFLAG_RD,
5846 				   &txq->ift_no_tx_dma_setup, "# of times map failed for other than EFBIG");
5847 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_pidx",
5848 				   CTLFLAG_RD,
5849 				   &txq->ift_pidx, 1, "Producer Index");
5850 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx",
5851 				   CTLFLAG_RD,
5852 				   &txq->ift_cidx, 1, "Consumer Index");
5853 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx_processed",
5854 				   CTLFLAG_RD,
5855 				   &txq->ift_cidx_processed, 1, "Consumer Index seen by credit update");
5856 		SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_in_use",
5857 				   CTLFLAG_RD,
5858 				   &txq->ift_in_use, 1, "descriptors in use");
5859 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_processed",
5860 				   CTLFLAG_RD,
5861 				   &txq->ift_processed, "descriptors procesed for clean");
5862 		SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_cleaned",
5863 				   CTLFLAG_RD,
5864 				   &txq->ift_cleaned, "total cleaned");
5865 		SYSCTL_ADD_PROC(ctx_list, queue_list, OID_AUTO, "ring_state",
5866 				CTLTYPE_STRING | CTLFLAG_RD, __DEVOLATILE(uint64_t *, &txq->ift_br->state),
5867 				0, mp_ring_state_handler, "A", "soft ring state");
5868 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_enqueues",
5869 				       CTLFLAG_RD, &txq->ift_br->enqueues,
5870 				       "# of enqueues to the mp_ring for this queue");
5871 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_drops",
5872 				       CTLFLAG_RD, &txq->ift_br->drops,
5873 				       "# of drops in the mp_ring for this queue");
5874 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_starts",
5875 				       CTLFLAG_RD, &txq->ift_br->starts,
5876 				       "# of normal consumer starts in the mp_ring for this queue");
5877 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_stalls",
5878 				       CTLFLAG_RD, &txq->ift_br->stalls,
5879 					       "# of consumer stalls in the mp_ring for this queue");
5880 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_restarts",
5881 			       CTLFLAG_RD, &txq->ift_br->restarts,
5882 				       "# of consumer restarts in the mp_ring for this queue");
5883 		SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_abdications",
5884 				       CTLFLAG_RD, &txq->ift_br->abdications,
5885 				       "# of consumer abdications in the mp_ring for this queue");
5886 	}
5887 
5888 	if (scctx->isc_nrxqsets > 100)
5889 		qfmt = "rxq%03d";
5890 	else if (scctx->isc_nrxqsets > 10)
5891 		qfmt = "rxq%02d";
5892 	else
5893 		qfmt = "rxq%d";
5894 	for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) {
5895 		snprintf(namebuf, NAME_BUFLEN, qfmt, i);
5896 		queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf,
5897 					     CTLFLAG_RD, NULL, "Queue Name");
5898 		queue_list = SYSCTL_CHILDREN(queue_node);
5899 		if (sctx->isc_flags & IFLIB_HAS_RXCQ) {
5900 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_pidx",
5901 				       CTLFLAG_RD,
5902 				       &rxq->ifr_cq_pidx, 1, "Producer Index");
5903 			SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_cidx",
5904 				       CTLFLAG_RD,
5905 				       &rxq->ifr_cq_cidx, 1, "Consumer Index");
5906 		}
5907 
5908 		for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) {
5909 			snprintf(namebuf, NAME_BUFLEN, "rxq_fl%d", j);
5910 			fl_node = SYSCTL_ADD_NODE(ctx_list, queue_list, OID_AUTO, namebuf,
5911 						     CTLFLAG_RD, NULL, "freelist Name");
5912 			fl_list = SYSCTL_CHILDREN(fl_node);
5913 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "pidx",
5914 				       CTLFLAG_RD,
5915 				       &fl->ifl_pidx, 1, "Producer Index");
5916 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "cidx",
5917 				       CTLFLAG_RD,
5918 				       &fl->ifl_cidx, 1, "Consumer Index");
5919 			SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "credits",
5920 				       CTLFLAG_RD,
5921 				       &fl->ifl_credits, 1, "credits available");
5922 #if MEMORY_LOGGING
5923 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_enqueued",
5924 					CTLFLAG_RD,
5925 					&fl->ifl_m_enqueued, "mbufs allocated");
5926 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_dequeued",
5927 					CTLFLAG_RD,
5928 					&fl->ifl_m_dequeued, "mbufs freed");
5929 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_enqueued",
5930 					CTLFLAG_RD,
5931 					&fl->ifl_cl_enqueued, "clusters allocated");
5932 			SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_dequeued",
5933 					CTLFLAG_RD,
5934 					&fl->ifl_cl_dequeued, "clusters freed");
5935 #endif
5936 
5937 		}
5938 	}
5939 
5940 }
5941 
5942 #ifndef __NO_STRICT_ALIGNMENT
5943 static struct mbuf *
5944 iflib_fixup_rx(struct mbuf *m)
5945 {
5946 	struct mbuf *n;
5947 
5948 	if (m->m_len <= (MCLBYTES - ETHER_HDR_LEN)) {
5949 		bcopy(m->m_data, m->m_data + ETHER_HDR_LEN, m->m_len);
5950 		m->m_data += ETHER_HDR_LEN;
5951 		n = m;
5952 	} else {
5953 		MGETHDR(n, M_NOWAIT, MT_DATA);
5954 		if (n == NULL) {
5955 			m_freem(m);
5956 			return (NULL);
5957 		}
5958 		bcopy(m->m_data, n->m_data, ETHER_HDR_LEN);
5959 		m->m_data += ETHER_HDR_LEN;
5960 		m->m_len -= ETHER_HDR_LEN;
5961 		n->m_len = ETHER_HDR_LEN;
5962 		M_MOVE_PKTHDR(n, m);
5963 		n->m_next = m;
5964 	}
5965 	return (n);
5966 }
5967 #endif
5968