xref: /freebsd-12.1/sys/dev/netmap/netmap_kern.h (revision c8953e12)
1 /*
2  * Copyright (C) 2011-2012 Matteo Landi, Luigi Rizzo. All rights reserved.
3  *
4  * Redistribution and use in source and binary forms, with or without
5  * modification, are permitted provided that the following conditions
6  * are met:
7  *   1. Redistributions of source code must retain the above copyright
8  *      notice, this list of conditions and the following disclaimer.
9  *   2. Redistributions in binary form must reproduce the above copyright
10  *      notice, this list of conditions and the following disclaimer in the
11  *    documentation and/or other materials provided with the distribution.
12  *
13  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
14  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
15  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
16  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
17  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
18  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
19  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
20  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
21  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
22  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
23  * SUCH DAMAGE.
24  */
25 
26 /*
27  * $FreeBSD$
28  * $Id: netmap_kern.h 11343 2012-07-03 09:08:38Z luigi $
29  *
30  * The header contains the definitions of constants and function
31  * prototypes used only in kernelspace.
32  */
33 
34 #ifndef _NET_NETMAP_KERN_H_
35 #define _NET_NETMAP_KERN_H_
36 
37 #define NETMAP_MEM2    // use the new memory allocator
38 
39 #if defined(__FreeBSD__)
40 #define likely(x)	__builtin_expect(!!(x), 1)
41 #define unlikely(x)	__builtin_expect(!!(x), 0)
42 
43 #define	NM_LOCK_T	struct mtx
44 #define	NM_SELINFO_T	struct selinfo
45 #define	MBUF_LEN(m)	((m)->m_pkthdr.len)
46 #define	NM_SEND_UP(ifp, m)	((ifp)->if_input)(ifp, m)
47 #elif defined (linux)
48 #define	NM_LOCK_T	spinlock_t
49 #define	NM_SELINFO_T	wait_queue_head_t
50 #define	MBUF_LEN(m)	((m)->len)
51 #define	NM_SEND_UP(ifp, m)	netif_rx(m)
52 
53 #ifndef DEV_NETMAP
54 #define DEV_NETMAP
55 #endif
56 
57 /*
58  * IFCAP_NETMAP goes into net_device's flags (if_capabilities)
59  * and priv_flags (if_capenable). The latter used to be 16 bits
60  * up to linux 2.6.36, so we need to use a 16 bit value on older
61  * platforms and tolerate the clash with IFF_DYNAMIC and IFF_BRIDGE_PORT.
62  * For the 32-bit value, 0x100000 (bit 20) has no clashes up to 3.3.1
63  */
64 #if LINUX_VERSION_CODE < KERNEL_VERSION(2,6,37)
65 #define IFCAP_NETMAP	0x8000
66 #else
67 #define IFCAP_NETMAP	0x100000
68 #endif
69 
70 #elif defined (__APPLE__)
71 #warning apple support is experimental
72 #define likely(x)	__builtin_expect(!!(x), 1)
73 #define unlikely(x)	__builtin_expect(!!(x), 0)
74 #define	NM_LOCK_T	IOLock *
75 #define	NM_SELINFO_T	struct selinfo
76 #define	MBUF_LEN(m)	((m)->m_pkthdr.len)
77 #define	NM_SEND_UP(ifp, m)	((ifp)->if_input)(ifp, m)
78 
79 #else
80 #error unsupported platform
81 #endif
82 
83 #ifdef MALLOC_DECLARE
84 MALLOC_DECLARE(M_NETMAP);
85 #endif
86 
87 #define ND(format, ...)
88 #define D(format, ...)						\
89 	do {							\
90 		struct timeval __xxts;				\
91 		microtime(&__xxts);				\
92 		printf("%03d.%06d %s [%d] " format "\n",	\
93 		(int)__xxts.tv_sec % 1000, (int)__xxts.tv_usec,	\
94 		__FUNCTION__, __LINE__, ##__VA_ARGS__);		\
95 	} while (0)
96 
97 struct netmap_adapter;
98 
99 /*
100  * private, kernel view of a ring. Keeps track of the status of
101  * a ring across system calls.
102  *
103  *	nr_hwcur	index of the next buffer to refill.
104  *			It corresponds to ring->cur - ring->reserved
105  *
106  *	nr_hwavail	the number of slots "owned" by userspace.
107  *			nr_hwavail =:= ring->avail + ring->reserved
108  *
109  * The indexes in the NIC and netmap rings are offset by nkr_hwofs slots.
110  * This is so that, on a reset, buffers owned by userspace are not
111  * modified by the kernel. In particular:
112  * RX rings: the next empty buffer (hwcur + hwavail + hwofs) coincides with
113  * 	the next empty buffer as known by the hardware (next_to_check or so).
114  * TX rings: hwcur + hwofs coincides with next_to_send
115  */
116 struct netmap_kring {
117 	struct netmap_ring *ring;
118 	u_int nr_hwcur;
119 	int nr_hwavail;
120 	u_int nr_kflags;	/* private driver flags */
121 #define NKR_PENDINTR	0x1	// Pending interrupt.
122 	u_int nkr_num_slots;
123 
124 	int	nkr_hwofs;	/* offset between NIC and netmap ring */
125 	struct netmap_adapter *na;
126 	NM_SELINFO_T si;	/* poll/select wait queue */
127 	NM_LOCK_T q_lock;	/* used if no device lock available */
128 } __attribute__((__aligned__(64)));
129 
130 /*
131  * This struct extends the 'struct adapter' (or
132  * equivalent) device descriptor. It contains all fields needed to
133  * support netmap operation.
134  */
135 struct netmap_adapter {
136 	int refcount; /* number of user-space descriptors using this
137 			 interface, which is equal to the number of
138 			 struct netmap_if objs in the mapped region. */
139 	/*
140 	 * The selwakeup in the interrupt thread can use per-ring
141 	 * and/or global wait queues. We track how many clients
142 	 * of each type we have so we can optimize the drivers,
143 	 * and especially avoid huge contention on the locks.
144 	 */
145 	int na_single;	/* threads attached to a single hw queue */
146 	int na_multi;	/* threads attached to multiple hw queues */
147 
148 	int separate_locks; /* set if the interface suports different
149 			       locks for rx, tx and core. */
150 
151 	u_int num_rx_rings; /* number of tx/rx ring pairs */
152 	u_int num_tx_rings; // if nonzero, overrides num_rx_rings
153 
154 	u_int num_tx_desc; /* number of descriptor in each queue */
155 	u_int num_rx_desc;
156 	//u_int buff_size;	// XXX deprecate, use NETMAP_BUF_SIZE
157 
158 	/* tx_rings and rx_rings are private but allocated
159 	 * as a contiguous chunk of memory. Each array has
160 	 * N+1 entries, for the adapter queues and for the host queue.
161 	 */
162 	struct netmap_kring *tx_rings; /* array of TX rings. */
163 	struct netmap_kring *rx_rings; /* array of RX rings. */
164 
165 	NM_SELINFO_T tx_si, rx_si;	/* global wait queues */
166 
167 	/* copy of if_qflush and if_transmit pointers, to intercept
168 	 * packets from the network stack when netmap is active.
169 	 */
170 	int     (*if_transmit)(struct ifnet *, struct mbuf *);
171 
172 	/* references to the ifnet and device routines, used by
173 	 * the generic netmap functions.
174 	 */
175 	struct ifnet *ifp; /* adapter is ifp->if_softc */
176 
177 	NM_LOCK_T core_lock;	/* used if no device lock available */
178 
179 	int (*nm_register)(struct ifnet *, int onoff);
180 	void (*nm_lock)(struct ifnet *, int what, u_int ringid);
181 	int (*nm_txsync)(struct ifnet *, u_int ring, int lock);
182 	int (*nm_rxsync)(struct ifnet *, u_int ring, int lock);
183 
184 	int bdg_port;
185 #ifdef linux
186 	struct net_device_ops nm_ndo;
187 	int if_refcount;	// XXX additions for bridge
188 #endif /* linux */
189 };
190 
191 /*
192  * The combination of "enable" (ifp->if_capabilities &IFCAP_NETMAP)
193  * and refcount gives the status of the interface, namely:
194  *
195  *	enable	refcount	Status
196  *
197  *	FALSE	0		normal operation
198  *	FALSE	!= 0		-- (impossible)
199  *	TRUE	1		netmap mode
200  *	TRUE	0		being deleted.
201  */
202 
203 #define NETMAP_DELETING(_na)  (  ((_na)->refcount == 0) &&	\
204 	( (_na)->ifp->if_capenable & IFCAP_NETMAP) )
205 
206 /*
207  * parameters for (*nm_lock)(adapter, what, index)
208  */
209 enum {
210 	NETMAP_NO_LOCK = 0,
211 	NETMAP_CORE_LOCK, NETMAP_CORE_UNLOCK,
212 	NETMAP_TX_LOCK, NETMAP_TX_UNLOCK,
213 	NETMAP_RX_LOCK, NETMAP_RX_UNLOCK,
214 #ifdef __FreeBSD__
215 #define	NETMAP_REG_LOCK		NETMAP_CORE_LOCK
216 #define	NETMAP_REG_UNLOCK	NETMAP_CORE_UNLOCK
217 #else
218 	NETMAP_REG_LOCK, NETMAP_REG_UNLOCK
219 #endif
220 };
221 
222 /*
223  * The following are support routines used by individual drivers to
224  * support netmap operation.
225  *
226  * netmap_attach() initializes a struct netmap_adapter, allocating the
227  * 	struct netmap_ring's and the struct selinfo.
228  *
229  * netmap_detach() frees the memory allocated by netmap_attach().
230  *
231  * netmap_start() replaces the if_transmit routine of the interface,
232  *	and is used to intercept packets coming from the stack.
233  *
234  * netmap_load_map/netmap_reload_map are helper routines to set/reset
235  *	the dmamap for a packet buffer
236  *
237  * netmap_reset() is a helper routine to be called in the driver
238  *	when reinitializing a ring.
239  */
240 int netmap_attach(struct netmap_adapter *, int);
241 void netmap_detach(struct ifnet *);
242 int netmap_start(struct ifnet *, struct mbuf *);
243 enum txrx { NR_RX = 0, NR_TX = 1 };
244 struct netmap_slot *netmap_reset(struct netmap_adapter *na,
245 	enum txrx tx, int n, u_int new_cur);
246 int netmap_ring_reinit(struct netmap_kring *);
247 
248 extern int netmap_buf_size;
249 #define NETMAP_BUF_SIZE	netmap_buf_size
250 extern int netmap_mitigate;
251 extern int netmap_no_pendintr;
252 extern u_int netmap_total_buffers;
253 extern char *netmap_buffer_base;
254 extern int netmap_verbose;	// XXX debugging
255 enum {                                  /* verbose flags */
256 	NM_VERB_ON = 1,                 /* generic verbose */
257 	NM_VERB_HOST = 0x2,             /* verbose host stack */
258 	NM_VERB_RXSYNC = 0x10,          /* verbose on rxsync/txsync */
259 	NM_VERB_TXSYNC = 0x20,
260 	NM_VERB_RXINTR = 0x100,         /* verbose on rx/tx intr (driver) */
261 	NM_VERB_TXINTR = 0x200,
262 	NM_VERB_NIC_RXSYNC = 0x1000,    /* verbose on rx/tx intr (driver) */
263 	NM_VERB_NIC_TXSYNC = 0x2000,
264 };
265 
266 /*
267  * NA returns a pointer to the struct netmap adapter from the ifp,
268  * WNA is used to write it.
269  */
270 #ifndef WNA
271 #define	WNA(_ifp)	(_ifp)->if_pspare[0]
272 #endif
273 #define	NA(_ifp)	((struct netmap_adapter *)WNA(_ifp))
274 
275 
276 #ifdef __FreeBSD__
277 /* Callback invoked by the dma machinery after a successfull dmamap_load */
278 static void netmap_dmamap_cb(__unused void *arg,
279     __unused bus_dma_segment_t * segs, __unused int nseg, __unused int error)
280 {
281 }
282 
283 /* bus_dmamap_load wrapper: call aforementioned function if map != NULL.
284  * XXX can we do it without a callback ?
285  */
286 static inline void
287 netmap_load_map(bus_dma_tag_t tag, bus_dmamap_t map, void *buf)
288 {
289 	if (map)
290 		bus_dmamap_load(tag, map, buf, NETMAP_BUF_SIZE,
291 		    netmap_dmamap_cb, NULL, BUS_DMA_NOWAIT);
292 }
293 
294 /* update the map when a buffer changes. */
295 static inline void
296 netmap_reload_map(bus_dma_tag_t tag, bus_dmamap_t map, void *buf)
297 {
298 	if (map) {
299 		bus_dmamap_unload(tag, map);
300 		bus_dmamap_load(tag, map, buf, NETMAP_BUF_SIZE,
301 		    netmap_dmamap_cb, NULL, BUS_DMA_NOWAIT);
302 	}
303 }
304 #else /* linux */
305 
306 /*
307  * XXX How do we redefine these functions:
308  *
309  * on linux we need
310  *	dma_map_single(&pdev->dev, virt_addr, len, direction)
311  *	dma_unmap_single(&adapter->pdev->dev, phys_addr, len, direction
312  * The len can be implicit (on netmap it is NETMAP_BUF_SIZE)
313  * unfortunately the direction is not, so we need to change
314  * something to have a cross API
315  */
316 #define netmap_load_map(_t, _m, _b)
317 #define netmap_reload_map(_t, _m, _b)
318 #if 0
319 	struct e1000_buffer *buffer_info =  &tx_ring->buffer_info[l];
320 	/* set time_stamp *before* dma to help avoid a possible race */
321 	buffer_info->time_stamp = jiffies;
322 	buffer_info->mapped_as_page = false;
323 	buffer_info->length = len;
324 	//buffer_info->next_to_watch = l;
325 	/* reload dma map */
326 	dma_unmap_single(&adapter->pdev->dev, buffer_info->dma,
327 			NETMAP_BUF_SIZE, DMA_TO_DEVICE);
328 	buffer_info->dma = dma_map_single(&adapter->pdev->dev,
329 			addr, NETMAP_BUF_SIZE, DMA_TO_DEVICE);
330 
331 	if (dma_mapping_error(&adapter->pdev->dev, buffer_info->dma)) {
332 		D("dma mapping error");
333 		/* goto dma_error; See e1000_put_txbuf() */
334 		/* XXX reset */
335 	}
336 	tx_desc->buffer_addr = htole64(buffer_info->dma); //XXX
337 
338 #endif
339 
340 /*
341  * The bus_dmamap_sync() can be one of wmb() or rmb() depending on direction.
342  */
343 #define bus_dmamap_sync(_a, _b, _c)
344 
345 #endif /* linux */
346 
347 /*
348  * functions to map NIC to KRING indexes (n2k) and vice versa (k2n)
349  */
350 static inline int
351 netmap_idx_n2k(struct netmap_kring *kr, int idx)
352 {
353 	int n = kr->nkr_num_slots;
354 	idx += kr->nkr_hwofs;
355 	if (idx < 0)
356 		return idx + n;
357 	else if (idx < n)
358 		return idx;
359 	else
360 		return idx - n;
361 }
362 
363 
364 static inline int
365 netmap_idx_k2n(struct netmap_kring *kr, int idx)
366 {
367 	int n = kr->nkr_num_slots;
368 	idx -= kr->nkr_hwofs;
369 	if (idx < 0)
370 		return idx + n;
371 	else if (idx < n)
372 		return idx;
373 	else
374 		return idx - n;
375 }
376 
377 
378 #ifdef NETMAP_MEM2
379 /* Entries of the look-up table. */
380 struct lut_entry {
381 	void *vaddr;		/* virtual address. */
382 	vm_paddr_t paddr;	/* phisical address. */
383 };
384 
385 struct netmap_obj_pool;
386 extern struct lut_entry *netmap_buffer_lut;
387 #define NMB_VA(i)	(netmap_buffer_lut[i].vaddr)
388 #define NMB_PA(i)	(netmap_buffer_lut[i].paddr)
389 #else /* NETMAP_MEM1 */
390 #define NMB_VA(i)	(netmap_buffer_base + (i * NETMAP_BUF_SIZE) )
391 #endif /* NETMAP_MEM2 */
392 
393 /*
394  * NMB return the virtual address of a buffer (buffer 0 on bad index)
395  * PNMB also fills the physical address
396  */
397 static inline void *
398 NMB(struct netmap_slot *slot)
399 {
400 	uint32_t i = slot->buf_idx;
401 	return (unlikely(i >= netmap_total_buffers)) ?  NMB_VA(0) : NMB_VA(i);
402 }
403 
404 static inline void *
405 PNMB(struct netmap_slot *slot, uint64_t *pp)
406 {
407 	uint32_t i = slot->buf_idx;
408 	void *ret = (i >= netmap_total_buffers) ? NMB_VA(0) : NMB_VA(i);
409 #ifdef NETMAP_MEM2
410 	*pp = (i >= netmap_total_buffers) ? NMB_PA(0) : NMB_PA(i);
411 #else
412 	*pp = vtophys(ret);
413 #endif
414 	return ret;
415 }
416 
417 /* default functions to handle rx/tx interrupts */
418 int netmap_rx_irq(struct ifnet *, int, int *);
419 #define netmap_tx_irq(_n, _q) netmap_rx_irq(_n, _q, NULL)
420 
421 extern int netmap_copy;
422 #endif /* _NET_NETMAP_KERN_H_ */
423