FreeBSD/Linux Kernel Cross Reference
sys/net/netmap.h
1 /*
2 * Copyright (C) 2011-2014 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 *
8 * 1. Redistributions of source code must retain the above copyright
9 * notice, this list of conditions and the following disclaimer.
10 * 2. Redistributions in binary form must reproduce the above copyright
11 * notice, this list of conditions and the following disclaimer in the
12 * documentation and/or other materials provided with the distribution.
13 *
14 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``S IS''AND
15 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
16 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
17 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
18 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
19 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
20 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
21 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
22 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
23 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
24 * SUCH DAMAGE.
25 */
26
27 /*
28 * $FreeBSD$
29 *
30 * Definitions of constants and the structures used by the netmap
31 * framework, for the part visible to both kernel and userspace.
32 * Detailed info on netmap is available with "man netmap" or at
33 *
34 * http://info.iet.unipi.it/~luigi/netmap/
35 *
36 * This API is also used to communicate with the VALE software switch
37 */
38
39 #ifndef _NET_NETMAP_H_
40 #define _NET_NETMAP_H_
41
42 #define NETMAP_API 13 /* current API version */
43
44 #define NETMAP_MIN_API 13 /* min and max versions accepted */
45 #define NETMAP_MAX_API 15
46 /*
47 * Some fields should be cache-aligned to reduce contention.
48 * The alignment is architecture and OS dependent, but rather than
49 * digging into OS headers to find the exact value we use an estimate
50 * that should cover most architectures.
51 */
52 #define NM_CACHE_ALIGN 128
53
54 /*
55 * --- Netmap data structures ---
56 *
57 * The userspace data structures used by netmap are shown below.
58 * They are allocated by the kernel and mmap()ed by userspace threads.
59 * Pointers are implemented as memory offsets or indexes,
60 * so that they can be easily dereferenced in kernel and userspace.
61
62 KERNEL (opaque, obviously)
63
64 ====================================================================
65 |
66 USERSPACE | struct netmap_ring
67 +---->+---------------+
68 / | head,cur,tail |
69 struct netmap_if (nifp, 1 per fd) / | buf_ofs |
70 +---------------+ / | other fields |
71 | ni_tx_rings | / +===============+
72 | ni_rx_rings | / | buf_idx, len | slot[0]
73 | | / | flags, ptr |
74 | | / +---------------+
75 +===============+ / | buf_idx, len | slot[1]
76 | txring_ofs[0] | (rel.to nifp)--' | flags, ptr |
77 | txring_ofs[1] | +---------------+
78 (tx+1 entries) (num_slots entries)
79 | txring_ofs[t] | | buf_idx, len | slot[n-1]
80 +---------------+ | flags, ptr |
81 | rxring_ofs[0] | +---------------+
82 | rxring_ofs[1] |
83 (rx+1 entries)
84 | rxring_ofs[r] |
85 +---------------+
86
87 * For each "interface" (NIC, host stack, PIPE, VALE switch port) bound to
88 * a file descriptor, the mmap()ed region contains a (logically readonly)
89 * struct netmap_if pointing to struct netmap_ring's.
90 *
91 * There is one netmap_ring per physical NIC ring, plus one tx/rx ring
92 * pair attached to the host stack (this pair is unused for non-NIC ports).
93 *
94 * All physical/host stack ports share the same memory region,
95 * so that zero-copy can be implemented between them.
96 * VALE switch ports instead have separate memory regions.
97 *
98 * The netmap_ring is the userspace-visible replica of the NIC ring.
99 * Each slot has the index of a buffer (MTU-sized and residing in the
100 * mmapped region), its length and some flags. An extra 64-bit pointer
101 * is provided for user-supplied buffers in the tx path.
102 *
103 * In user space, the buffer address is computed as
104 * (char *)ring + buf_ofs + index * NETMAP_BUF_SIZE
105 *
106 * Added in NETMAP_API 11:
107 *
108 * + NIOCREGIF can request the allocation of extra spare buffers from
109 * the same memory pool. The desired number of buffers must be in
110 * nr_arg3. The ioctl may return fewer buffers, depending on memory
111 * availability. nr_arg3 will return the actual value, and, once
112 * mapped, nifp->ni_bufs_head will be the index of the first buffer.
113 *
114 * The buffers are linked to each other using the first uint32_t
115 * as the index. On close, ni_bufs_head must point to the list of
116 * buffers to be released.
117 *
118 * + NIOCREGIF can request space for extra rings (and buffers)
119 * allocated in the same memory space. The number of extra rings
120 * is in nr_arg1, and is advisory. This is a no-op on NICs where
121 * the size of the memory space is fixed.
122 *
123 * + NIOCREGIF can attach to PIPE rings sharing the same memory
124 * space with a parent device. The ifname indicates the parent device,
125 * which must already exist. Flags in nr_flags indicate if we want to
126 * bind the master or slave side, the index (from nr_ringid)
127 * is just a cookie and does not need to be sequential.
128 *
129 * + NIOCREGIF can also attach to 'monitor' rings that replicate
130 * the content of specific rings, also from the same memory space.
131 *
132 * Extra flags in nr_flags support the above functions.
133 * Application libraries may use the following naming scheme:
134 * netmap:foo all NIC ring pairs
135 * netmap:foo^ only host ring pair
136 * netmap:foo+ all NIC ring + host ring pairs
137 * netmap:foo-k the k-th NIC ring pair
138 * netmap:foo{k PIPE ring pair k, master side
139 * netmap:foo}k PIPE ring pair k, slave side
140 *
141 * Some notes about host rings:
142 *
143 * + The RX host ring is used to store those packets that the host network
144 * stack is trying to transmit through a NIC queue, but only if that queue
145 * is currently in netmap mode. Netmap will not intercept host stack mbufs
146 * designated to NIC queues that are not in netmap mode. As a consequence,
147 * registering a netmap port with netmap:foo^ is not enough to intercept
148 * mbufs in the RX host ring; the netmap port should be registered with
149 * netmap:foo*, or another registration should be done to open at least a
150 * NIC TX queue in netmap mode.
151 *
152 * + Netmap is not currently able to deal with intercepted trasmit mbufs which
153 * require offloadings like TSO, UFO, checksumming offloadings, etc. It is
154 * responsibility of the user to disable those offloadings (e.g. using
155 * ifconfig on FreeBSD or ethtool -K on Linux) for an interface that is being
156 * used in netmap mode. If the offloadings are not disabled, GSO and/or
157 * unchecksummed packets may be dropped immediately or end up in the host RX
158 * ring, and will be dropped as soon as the packet reaches another netmap
159 * adapter.
160 */
161
162 /*
163 * struct netmap_slot is a buffer descriptor
164 */
165 struct netmap_slot {
166 uint32_t buf_idx; /* buffer index */
167 uint16_t len; /* length for this slot */
168 uint16_t flags; /* buf changed, etc. */
169 uint64_t ptr; /* pointer for indirect buffers */
170 };
171
172 /*
173 * The following flags control how the slot is used
174 */
175
176 #define NS_BUF_CHANGED 0x0001 /* buf_idx changed */
177 /*
178 * must be set whenever buf_idx is changed (as it might be
179 * necessary to recompute the physical address and mapping)
180 *
181 * It is also set by the kernel whenever the buf_idx is
182 * changed internally (e.g., by pipes). Applications may
183 * use this information to know when they can reuse the
184 * contents of previously prepared buffers.
185 */
186
187 #define NS_REPORT 0x0002 /* ask the hardware to report results */
188 /*
189 * Request notification when slot is used by the hardware.
190 * Normally transmit completions are handled lazily and
191 * may be unreported. This flag lets us know when a slot
192 * has been sent (e.g. to terminate the sender).
193 */
194
195 #define NS_FORWARD 0x0004 /* pass packet 'forward' */
196 /*
197 * (Only for physical ports, rx rings with NR_FORWARD set).
198 * Slot released to the kernel (i.e. before ring->head) with
199 * this flag set are passed to the peer ring (host/NIC),
200 * thus restoring the host-NIC connection for these slots.
201 * This supports efficient traffic monitoring or firewalling.
202 */
203
204 #define NS_NO_LEARN 0x0008 /* disable bridge learning */
205 /*
206 * On a VALE switch, do not 'learn' the source port for
207 * this buffer.
208 */
209
210 #define NS_INDIRECT 0x0010 /* userspace buffer */
211 /*
212 * (VALE tx rings only) data is in a userspace buffer,
213 * whose address is in the 'ptr' field in the slot.
214 */
215
216 #define NS_MOREFRAG 0x0020 /* packet has more fragments */
217 /*
218 * (VALE ports, ptnetmap ports and some NIC ports, e.g.
219 * ixgbe and i40e on Linux)
220 * Set on all but the last slot of a multi-segment packet.
221 * The 'len' field refers to the individual fragment.
222 */
223
224 #define NS_PORT_SHIFT 8
225 #define NS_PORT_MASK (0xff << NS_PORT_SHIFT)
226 /*
227 * The high 8 bits of the flag, if not zero, indicate the
228 * destination port for the VALE switch, overriding
229 * the lookup table.
230 */
231
232 #define NS_RFRAGS(_slot) ( ((_slot)->flags >> 8) & 0xff)
233 /*
234 * (VALE rx rings only) the high 8 bits
235 * are the number of fragments.
236 */
237
238 #define NETMAP_MAX_FRAGS 64 /* max number of fragments */
239
240
241 /*
242 * struct netmap_ring
243 *
244 * Netmap representation of a TX or RX ring (also known as "queue").
245 * This is a queue implemented as a fixed-size circular array.
246 * At the software level the important fields are: head, cur, tail.
247 *
248 * In TX rings:
249 *
250 * head first slot available for transmission.
251 * cur wakeup point. select() and poll() will unblock
252 * when 'tail' moves past 'cur'
253 * tail (readonly) first slot reserved to the kernel
254 *
255 * [head .. tail-1] can be used for new packets to send;
256 * 'head' and 'cur' must be incremented as slots are filled
257 * with new packets to be sent;
258 * 'cur' can be moved further ahead if we need more space
259 * for new transmissions. XXX todo (2014-03-12)
260 *
261 * In RX rings:
262 *
263 * head first valid received packet
264 * cur wakeup point. select() and poll() will unblock
265 * when 'tail' moves past 'cur'
266 * tail (readonly) first slot reserved to the kernel
267 *
268 * [head .. tail-1] contain received packets;
269 * 'head' and 'cur' must be incremented as slots are consumed
270 * and can be returned to the kernel;
271 * 'cur' can be moved further ahead if we want to wait for
272 * new packets without returning the previous ones.
273 *
274 * DATA OWNERSHIP/LOCKING:
275 * The netmap_ring, and all slots and buffers in the range
276 * [head .. tail-1] are owned by the user program;
277 * the kernel only accesses them during a netmap system call
278 * and in the user thread context.
279 *
280 * Other slots and buffers are reserved for use by the kernel
281 */
282 struct netmap_ring {
283 /*
284 * buf_ofs is meant to be used through macros.
285 * It contains the offset of the buffer region from this
286 * descriptor.
287 */
288 const int64_t buf_ofs;
289 const uint32_t num_slots; /* number of slots in the ring. */
290 const uint32_t nr_buf_size;
291 const uint16_t ringid;
292 const uint16_t dir; /* 0: tx, 1: rx */
293
294 uint32_t head; /* (u) first user slot */
295 uint32_t cur; /* (u) wakeup point */
296 uint32_t tail; /* (k) first kernel slot */
297
298 uint32_t flags;
299
300 struct timeval ts; /* (k) time of last *sync() */
301
302 /* opaque room for a mutex or similar object */
303 #if !defined(_WIN32) || defined(__CYGWIN__)
304 uint8_t __attribute__((__aligned__(NM_CACHE_ALIGN))) sem[128];
305 #else
306 uint8_t __declspec(align(NM_CACHE_ALIGN)) sem[128];
307 #endif
308
309 /* the slots follow. This struct has variable size */
310 struct netmap_slot slot[0]; /* array of slots. */
311 };
312
313
314 /*
315 * RING FLAGS
316 */
317 #define NR_TIMESTAMP 0x0002 /* set timestamp on *sync() */
318 /*
319 * updates the 'ts' field on each netmap syscall. This saves
320 * saves a separate gettimeofday(), and is not much worse than
321 * software timestamps generated in the interrupt handler.
322 */
323
324 #define NR_FORWARD 0x0004 /* enable NS_FORWARD for ring */
325 /*
326 * Enables the NS_FORWARD slot flag for the ring.
327 */
328
329 /*
330 * Helper functions for kernel and userspace
331 */
332
333 /*
334 * Check if space is available in the ring. We use ring->head, which
335 * points to the next netmap slot to be published to netmap. It is
336 * possible that the applications moves ring->cur ahead of ring->tail
337 * (e.g., by setting ring->cur <== ring->tail), if it wants more slots
338 * than the ones currently available, and it wants to be notified when
339 * more arrive. See netmap(4) for more details and examples.
340 */
341 static inline int
342 nm_ring_empty(struct netmap_ring *ring)
343 {
344 return (ring->head == ring->tail);
345 }
346
347 /*
348 * Netmap representation of an interface and its queue(s).
349 * This is initialized by the kernel when binding a file
350 * descriptor to a port, and should be considered as readonly
351 * by user programs. The kernel never uses it.
352 *
353 * There is one netmap_if for each file descriptor on which we want
354 * to select/poll.
355 * select/poll operates on one or all pairs depending on the value of
356 * nmr_queueid passed on the ioctl.
357 */
358 struct netmap_if {
359 char ni_name[IFNAMSIZ]; /* name of the interface. */
360 const uint32_t ni_version; /* API version, currently unused */
361 const uint32_t ni_flags; /* properties */
362 #define NI_PRIV_MEM 0x1 /* private memory region */
363
364 /*
365 * The number of packet rings available in netmap mode.
366 * Physical NICs can have different numbers of tx and rx rings.
367 * Physical NICs also have a 'host' ring pair.
368 * Additionally, clients can request additional ring pairs to
369 * be used for internal communication.
370 */
371 const uint32_t ni_tx_rings; /* number of HW tx rings */
372 const uint32_t ni_rx_rings; /* number of HW rx rings */
373
374 uint32_t ni_bufs_head; /* head index for extra bufs */
375 uint32_t ni_spare1[5];
376 /*
377 * The following array contains the offset of each netmap ring
378 * from this structure, in the following order:
379 * NIC tx rings (ni_tx_rings); host tx ring (1); extra tx rings;
380 * NIC rx rings (ni_rx_rings); host tx ring (1); extra rx rings.
381 *
382 * The area is filled up by the kernel on NIOCREGIF,
383 * and then only read by userspace code.
384 */
385 const ssize_t ring_ofs[0];
386 };
387
388 /* Legacy interface to interact with a netmap control device.
389 * Included for backward compatibility. The user should not include this
390 * file directly. */
391 #include "netmap_legacy.h"
392
393 /*
394 * New API to control netmap control devices. New applications should only use
395 * nmreq_xyz structs with the NIOCCTRL ioctl() command.
396 *
397 * NIOCCTRL takes a nmreq_header struct, which contains the required
398 * API version, the name of a netmap port, a command type, and pointers
399 * to request body and options.
400 *
401 * nr_name (in)
402 * The name of the port (em0, valeXXX:YYY, eth0{pn1 etc.)
403 *
404 * nr_version (in/out)
405 * Must match NETMAP_API as used in the kernel, error otherwise.
406 * Always returns the desired value on output.
407 *
408 * nr_reqtype (in)
409 * One of the NETMAP_REQ_* command types below
410 *
411 * nr_body (in)
412 * Pointer to a command-specific struct, described by one
413 * of the struct nmreq_xyz below.
414 *
415 * nr_options (in)
416 * Command specific options, if any.
417 *
418 * A NETMAP_REQ_REGISTER command activates netmap mode on the netmap
419 * port (e.g. physical interface) specified by nmreq_header.nr_name.
420 * The request body (struct nmreq_register) has several arguments to
421 * specify how the port is to be registered.
422 *
423 * nr_tx_slots, nr_tx_slots, nr_tx_rings, nr_rx_rings (in/out)
424 * On input, non-zero values may be used to reconfigure the port
425 * according to the requested values, but this is not guaranteed.
426 * On output the actual values in use are reported.
427 *
428 * nr_mode (in)
429 * Indicate what set of rings must be bound to the netmap
430 * device (e.g. all NIC rings, host rings only, NIC and
431 * host rings, ...). Values are in NR_REG_*.
432 *
433 * nr_ringid (in)
434 * If nr_mode == NR_REG_ONE_NIC (only a single couple of TX/RX
435 * rings), indicate which NIC TX and/or RX ring is to be bound
436 * (0..nr_*x_rings-1).
437 *
438 * nr_flags (in)
439 * Indicate special options for how to open the port.
440 *
441 * NR_NO_TX_POLL can be OR-ed to make select()/poll() push
442 * packets on tx rings only if POLLOUT is set.
443 * The default is to push any pending packet.
444 *
445 * NR_DO_RX_POLL can be OR-ed to make select()/poll() release
446 * packets on rx rings also when POLLIN is NOT set.
447 * The default is to touch the rx ring only with POLLIN.
448 * Note that this is the opposite of TX because it
449 * reflects the common usage.
450 *
451 * Other options are NR_MONITOR_TX, NR_MONITOR_RX, NR_ZCOPY_MON,
452 * NR_EXCLUSIVE, NR_RX_RINGS_ONLY, NR_TX_RINGS_ONLY and
453 * NR_ACCEPT_VNET_HDR.
454 *
455 * nr_mem_id (in/out)
456 * The identity of the memory region used.
457 * On input, 0 means the system decides autonomously,
458 * other values may try to select a specific region.
459 * On return the actual value is reported.
460 * Region '1' is the global allocator, normally shared
461 * by all interfaces. Other values are private regions.
462 * If two ports the same region zero-copy is possible.
463 *
464 * nr_extra_bufs (in/out)
465 * Number of extra buffers to be allocated.
466 *
467 * The other NETMAP_REQ_* commands are described below.
468 *
469 */
470
471 /* maximum size of a request, including all options */
472 #define NETMAP_REQ_MAXSIZE 4096
473
474 /* Header common to all request options. */
475 struct nmreq_option {
476 /* Pointer ot the next option. */
477 uint64_t nro_next;
478 /* Option type. */
479 uint32_t nro_reqtype;
480 /* (out) status of the option:
481 * 0: recognized and processed
482 * !=0: errno value
483 */
484 uint32_t nro_status;
485 /* Option size, used only for options that can have variable size
486 * (e.g. because they contain arrays). For fixed-size options this
487 * field should be set to zero. */
488 uint64_t nro_size;
489 };
490
491 /* Header common to all requests. Do not reorder these fields, as we need
492 * the second one (nr_reqtype) to know how much to copy from/to userspace. */
493 struct nmreq_header {
494 uint16_t nr_version; /* API version */
495 uint16_t nr_reqtype; /* nmreq type (NETMAP_REQ_*) */
496 uint32_t nr_reserved; /* must be zero */
497 #define NETMAP_REQ_IFNAMSIZ 64
498 char nr_name[NETMAP_REQ_IFNAMSIZ]; /* port name */
499 uint64_t nr_options; /* command-specific options */
500 uint64_t nr_body; /* ptr to nmreq_xyz struct */
501 };
502
503 enum {
504 /* Register a netmap port with the device. */
505 NETMAP_REQ_REGISTER = 1,
506 /* Get information from a netmap port. */
507 NETMAP_REQ_PORT_INFO_GET,
508 /* Attach a netmap port to a VALE switch. */
509 NETMAP_REQ_VALE_ATTACH,
510 /* Detach a netmap port from a VALE switch. */
511 NETMAP_REQ_VALE_DETACH,
512 /* List the ports attached to a VALE switch. */
513 NETMAP_REQ_VALE_LIST,
514 /* Set the port header length (was virtio-net header length). */
515 NETMAP_REQ_PORT_HDR_SET,
516 /* Get the port header length (was virtio-net header length). */
517 NETMAP_REQ_PORT_HDR_GET,
518 /* Create a new persistent VALE port. */
519 NETMAP_REQ_VALE_NEWIF,
520 /* Delete a persistent VALE port. */
521 NETMAP_REQ_VALE_DELIF,
522 /* Enable polling kernel thread(s) on an attached VALE port. */
523 NETMAP_REQ_VALE_POLLING_ENABLE,
524 /* Disable polling kernel thread(s) on an attached VALE port. */
525 NETMAP_REQ_VALE_POLLING_DISABLE,
526 /* Get info about the pools of a memory allocator. */
527 NETMAP_REQ_POOLS_INFO_GET,
528 /* Start an in-kernel loop that syncs the rings periodically or
529 * on notifications. The loop runs in the context of the ioctl
530 * syscall, and only stops on NETMAP_REQ_SYNC_KLOOP_STOP. */
531 NETMAP_REQ_SYNC_KLOOP_START,
532 /* Stops the thread executing the in-kernel loop. The thread
533 * returns from the ioctl syscall. */
534 NETMAP_REQ_SYNC_KLOOP_STOP,
535 /* Enable CSB mode on a registered netmap control device. */
536 NETMAP_REQ_CSB_ENABLE,
537 };
538
539 enum {
540 /* On NETMAP_REQ_REGISTER, ask netmap to use memory allocated
541 * from user-space allocated memory pools (e.g. hugepages).
542 */
543 NETMAP_REQ_OPT_EXTMEM = 1,
544
545 /* ON NETMAP_REQ_SYNC_KLOOP_START, ask netmap to use eventfd-based
546 * notifications to synchronize the kernel loop with the application.
547 */
548 NETMAP_REQ_OPT_SYNC_KLOOP_EVENTFDS,
549
550 /* On NETMAP_REQ_REGISTER, ask netmap to work in CSB mode, where
551 * head, cur and tail pointers are not exchanged through the
552 * struct netmap_ring header, but rather using an user-provided
553 * memory area (see struct nm_csb_atok and struct nm_csb_ktoa).
554 */
555 NETMAP_REQ_OPT_CSB,
556
557 /* An extension to NETMAP_REQ_OPT_SYNC_KLOOP_EVENTFDS, which specifies
558 * if the TX and/or RX rings are synced in the context of the VM exit.
559 * This requires the 'ioeventfd' fields to be valid (cannot be < 0).
560 */
561 NETMAP_REQ_OPT_SYNC_KLOOP_MODE,
562 };
563
564 /*
565 * nr_reqtype: NETMAP_REQ_REGISTER
566 * Bind (register) a netmap port to this control device.
567 */
568 struct nmreq_register {
569 uint64_t nr_offset; /* nifp offset in the shared region */
570 uint64_t nr_memsize; /* size of the shared region */
571 uint32_t nr_tx_slots; /* slots in tx rings */
572 uint32_t nr_rx_slots; /* slots in rx rings */
573 uint16_t nr_tx_rings; /* number of tx rings */
574 uint16_t nr_rx_rings; /* number of rx rings */
575
576 uint16_t nr_mem_id; /* id of the memory allocator */
577 uint16_t nr_ringid; /* ring(s) we care about */
578 uint32_t nr_mode; /* specify NR_REG_* modes */
579 uint32_t nr_extra_bufs; /* number of requested extra buffers */
580
581 uint64_t nr_flags; /* additional flags (see below) */
582 /* monitors use nr_ringid and nr_mode to select the rings to monitor */
583 #define NR_MONITOR_TX 0x100
584 #define NR_MONITOR_RX 0x200
585 #define NR_ZCOPY_MON 0x400
586 /* request exclusive access to the selected rings */
587 #define NR_EXCLUSIVE 0x800
588 /* 0x1000 unused */
589 #define NR_RX_RINGS_ONLY 0x2000
590 #define NR_TX_RINGS_ONLY 0x4000
591 /* Applications set this flag if they are able to deal with virtio-net headers,
592 * that is send/receive frames that start with a virtio-net header.
593 * If not set, NIOCREGIF will fail with netmap ports that require applications
594 * to use those headers. If the flag is set, the application can use the
595 * NETMAP_VNET_HDR_GET command to figure out the header length. */
596 #define NR_ACCEPT_VNET_HDR 0x8000
597 /* The following two have the same meaning of NETMAP_NO_TX_POLL and
598 * NETMAP_DO_RX_POLL. */
599 #define NR_DO_RX_POLL 0x10000
600 #define NR_NO_TX_POLL 0x20000
601 };
602
603 /* Valid values for nmreq_register.nr_mode (see above). */
604 enum { NR_REG_DEFAULT = 0, /* backward compat, should not be used. */
605 NR_REG_ALL_NIC = 1,
606 NR_REG_SW = 2,
607 NR_REG_NIC_SW = 3,
608 NR_REG_ONE_NIC = 4,
609 NR_REG_PIPE_MASTER = 5, /* deprecated, use "x{y" port name syntax */
610 NR_REG_PIPE_SLAVE = 6, /* deprecated, use "x}y" port name syntax */
611 NR_REG_NULL = 7,
612 };
613
614 /* A single ioctl number is shared by all the new API command.
615 * Demultiplexing is done using the hdr.nr_reqtype field.
616 * FreeBSD uses the size value embedded in the _IOWR to determine
617 * how much to copy in/out, so we define the ioctl() command
618 * specifying only nmreq_header, and copyin/copyout the rest. */
619 #define NIOCCTRL _IOWR('i', 151, struct nmreq_header)
620
621 /* The ioctl commands to sync TX/RX netmap rings.
622 * NIOCTXSYNC, NIOCRXSYNC synchronize tx or rx queues,
623 * whose identity is set in NIOCREGIF through nr_ringid.
624 * These are non blocking and take no argument. */
625 #define NIOCTXSYNC _IO('i', 148) /* sync tx queues */
626 #define NIOCRXSYNC _IO('i', 149) /* sync rx queues */
627
628 /*
629 * nr_reqtype: NETMAP_REQ_PORT_INFO_GET
630 * Get information about a netmap port, including number of rings.
631 * slots per ring, id of the memory allocator, etc. The netmap
632 * control device used for this operation does not need to be bound
633 * to a netmap port.
634 */
635 struct nmreq_port_info_get {
636 uint64_t nr_memsize; /* size of the shared region */
637 uint32_t nr_tx_slots; /* slots in tx rings */
638 uint32_t nr_rx_slots; /* slots in rx rings */
639 uint16_t nr_tx_rings; /* number of tx rings */
640 uint16_t nr_rx_rings; /* number of rx rings */
641 uint16_t nr_mem_id; /* memory allocator id (in/out) */
642 uint16_t pad1;
643 };
644
645 #define NM_BDG_NAME "vale" /* prefix for bridge port name */
646
647 /*
648 * nr_reqtype: NETMAP_REQ_VALE_ATTACH
649 * Attach a netmap port to a VALE switch. Both the name of the netmap
650 * port and the VALE switch are specified through the nr_name argument.
651 * The attach operation could need to register a port, so at least
652 * the same arguments are available.
653 * port_index will contain the index where the port has been attached.
654 */
655 struct nmreq_vale_attach {
656 struct nmreq_register reg;
657 uint32_t port_index;
658 uint32_t pad1;
659 };
660
661 /*
662 * nr_reqtype: NETMAP_REQ_VALE_DETACH
663 * Detach a netmap port from a VALE switch. Both the name of the netmap
664 * port and the VALE switch are specified through the nr_name argument.
665 * port_index will contain the index where the port was attached.
666 */
667 struct nmreq_vale_detach {
668 uint32_t port_index;
669 uint32_t pad1;
670 };
671
672 /*
673 * nr_reqtype: NETMAP_REQ_VALE_LIST
674 * List the ports of a VALE switch.
675 */
676 struct nmreq_vale_list {
677 /* Name of the VALE port (valeXXX:YYY) or empty. */
678 uint16_t nr_bridge_idx;
679 uint16_t pad1;
680 uint32_t nr_port_idx;
681 };
682
683 /*
684 * nr_reqtype: NETMAP_REQ_PORT_HDR_SET or NETMAP_REQ_PORT_HDR_GET
685 * Set or get the port header length of the port identified by hdr.nr_name.
686 * The control device does not need to be bound to a netmap port.
687 */
688 struct nmreq_port_hdr {
689 uint32_t nr_hdr_len;
690 uint32_t pad1;
691 };
692
693 /*
694 * nr_reqtype: NETMAP_REQ_VALE_NEWIF
695 * Create a new persistent VALE port.
696 */
697 struct nmreq_vale_newif {
698 uint32_t nr_tx_slots; /* slots in tx rings */
699 uint32_t nr_rx_slots; /* slots in rx rings */
700 uint16_t nr_tx_rings; /* number of tx rings */
701 uint16_t nr_rx_rings; /* number of rx rings */
702 uint16_t nr_mem_id; /* id of the memory allocator */
703 uint16_t pad1;
704 };
705
706 /*
707 * nr_reqtype: NETMAP_REQ_VALE_POLLING_ENABLE or NETMAP_REQ_VALE_POLLING_DISABLE
708 * Enable or disable polling kthreads on a VALE port.
709 */
710 struct nmreq_vale_polling {
711 uint32_t nr_mode;
712 #define NETMAP_POLLING_MODE_SINGLE_CPU 1
713 #define NETMAP_POLLING_MODE_MULTI_CPU 2
714 uint32_t nr_first_cpu_id;
715 uint32_t nr_num_polling_cpus;
716 uint32_t pad1;
717 };
718
719 /*
720 * nr_reqtype: NETMAP_REQ_POOLS_INFO_GET
721 * Get info about the pools of the memory allocator of the netmap
722 * port specified by hdr.nr_name and nr_mem_id. The netmap control
723 * device used for this operation does not need to be bound to a netmap
724 * port.
725 */
726 struct nmreq_pools_info {
727 uint64_t nr_memsize;
728 uint16_t nr_mem_id; /* in/out argument */
729 uint16_t pad1[3];
730 uint64_t nr_if_pool_offset;
731 uint32_t nr_if_pool_objtotal;
732 uint32_t nr_if_pool_objsize;
733 uint64_t nr_ring_pool_offset;
734 uint32_t nr_ring_pool_objtotal;
735 uint32_t nr_ring_pool_objsize;
736 uint64_t nr_buf_pool_offset;
737 uint32_t nr_buf_pool_objtotal;
738 uint32_t nr_buf_pool_objsize;
739 };
740
741 /*
742 * nr_reqtype: NETMAP_REQ_SYNC_KLOOP_START
743 * Start an in-kernel loop that syncs the rings periodically or on
744 * notifications. The loop runs in the context of the ioctl syscall,
745 * and only stops on NETMAP_REQ_SYNC_KLOOP_STOP.
746 * The registered netmap port must be open in CSB mode.
747 */
748 struct nmreq_sync_kloop_start {
749 /* Sleeping is the default synchronization method for the kloop.
750 * The 'sleep_us' field specifies how many microsconds to sleep for
751 * when there is no work to do, before doing another kloop iteration.
752 */
753 uint32_t sleep_us;
754 uint32_t pad1;
755 };
756
757 /* A CSB entry for the application --> kernel direction. */
758 struct nm_csb_atok {
759 uint32_t head; /* AW+ KR+ the head of the appl netmap_ring */
760 uint32_t cur; /* AW+ KR+ the cur of the appl netmap_ring */
761 uint32_t appl_need_kick; /* AW+ KR+ kern --> appl notification enable */
762 uint32_t sync_flags; /* AW+ KR+ the flags of the appl [tx|rx]sync() */
763 uint32_t pad[12]; /* pad to a 64 bytes cacheline */
764 };
765
766 /* A CSB entry for the application <-- kernel direction. */
767 struct nm_csb_ktoa {
768 uint32_t hwcur; /* AR+ KW+ the hwcur of the kern netmap_kring */
769 uint32_t hwtail; /* AR+ KW+ the hwtail of the kern netmap_kring */
770 uint32_t kern_need_kick; /* AR+ KW+ appl-->kern notification enable */
771 uint32_t pad[13];
772 };
773
774 #ifdef __linux__
775
776 #ifdef __KERNEL__
777 #define nm_stst_barrier smp_wmb
778 #define nm_ldld_barrier smp_rmb
779 #define nm_stld_barrier smp_mb
780 #else /* !__KERNEL__ */
781 static inline void nm_stst_barrier(void)
782 {
783 /* A memory barrier with release semantic has the combined
784 * effect of a store-store barrier and a load-store barrier,
785 * which is fine for us. */
786 __atomic_thread_fence(__ATOMIC_RELEASE);
787 }
788 static inline void nm_ldld_barrier(void)
789 {
790 /* A memory barrier with acquire semantic has the combined
791 * effect of a load-load barrier and a store-load barrier,
792 * which is fine for us. */
793 __atomic_thread_fence(__ATOMIC_ACQUIRE);
794 }
795 #endif /* !__KERNEL__ */
796
797 #elif defined(__FreeBSD__)
798
799 #ifdef _KERNEL
800 #define nm_stst_barrier atomic_thread_fence_rel
801 #define nm_ldld_barrier atomic_thread_fence_acq
802 #define nm_stld_barrier atomic_thread_fence_seq_cst
803 #else /* !_KERNEL */
804 #include <stdatomic.h>
805 static inline void nm_stst_barrier(void)
806 {
807 atomic_thread_fence(memory_order_release);
808 }
809 static inline void nm_ldld_barrier(void)
810 {
811 atomic_thread_fence(memory_order_acquire);
812 }
813 #endif /* !_KERNEL */
814
815 #else /* !__linux__ && !__FreeBSD__ */
816 #error "OS not supported"
817 #endif /* !__linux__ && !__FreeBSD__ */
818
819 /* Application side of sync-kloop: Write ring pointers (cur, head) to the CSB.
820 * This routine is coupled with sync_kloop_kernel_read(). */
821 static inline void
822 nm_sync_kloop_appl_write(struct nm_csb_atok *atok, uint32_t cur,
823 uint32_t head)
824 {
825 /* Issue a first store-store barrier to make sure writes to the
826 * netmap ring do not overcome updates on atok->cur and atok->head. */
827 nm_stst_barrier();
828
829 /*
830 * We need to write cur and head to the CSB but we cannot do it atomically.
831 * There is no way we can prevent the host from reading the updated value
832 * of one of the two and the old value of the other. However, if we make
833 * sure that the host never reads a value of head more recent than the
834 * value of cur we are safe. We can allow the host to read a value of cur
835 * more recent than the value of head, since in the netmap ring cur can be
836 * ahead of head and cur cannot wrap around head because it must be behind
837 * tail. Inverting the order of writes below could instead result into the
838 * host to think head went ahead of cur, which would cause the sync
839 * prologue to fail.
840 *
841 * The following memory barrier scheme is used to make this happen:
842 *
843 * Guest Host
844 *
845 * STORE(cur) LOAD(head)
846 * wmb() <-----------> rmb()
847 * STORE(head) LOAD(cur)
848 *
849 */
850 atok->cur = cur;
851 nm_stst_barrier();
852 atok->head = head;
853 }
854
855 /* Application side of sync-kloop: Read kring pointers (hwcur, hwtail) from
856 * the CSB. This routine is coupled with sync_kloop_kernel_write(). */
857 static inline void
858 nm_sync_kloop_appl_read(struct nm_csb_ktoa *ktoa, uint32_t *hwtail,
859 uint32_t *hwcur)
860 {
861 /*
862 * We place a memory barrier to make sure that the update of hwtail never
863 * overtakes the update of hwcur.
864 * (see explanation in sync_kloop_kernel_write).
865 */
866 *hwtail = ktoa->hwtail;
867 nm_ldld_barrier();
868 *hwcur = ktoa->hwcur;
869
870 /* Make sure that loads from ktoa->hwtail and ktoa->hwcur are not delayed
871 * after the loads from the netmap ring. */
872 nm_ldld_barrier();
873 }
874
875 /*
876 * data for NETMAP_REQ_OPT_* options
877 */
878
879 struct nmreq_opt_sync_kloop_eventfds {
880 struct nmreq_option nro_opt; /* common header */
881 /* An array of N entries for bidirectional notifications between
882 * the kernel loop and the application. The number of entries and
883 * their order must agree with the CSB arrays passed in the
884 * NETMAP_REQ_OPT_CSB option. Each entry contains a file descriptor
885 * backed by an eventfd.
886 *
887 * If any of the 'ioeventfd' entries is < 0, the event loop uses
888 * the sleeping synchronization strategy (according to sleep_us),
889 * and keeps kern_need_kick always disabled.
890 * Each 'irqfd' can be < 0, and in that case the corresponding queue
891 * is never notified.
892 */
893 struct {
894 /* Notifier for the application --> kernel loop direction. */
895 int32_t ioeventfd;
896 /* Notifier for the kernel loop --> application direction. */
897 int32_t irqfd;
898 } eventfds[0];
899 };
900
901 struct nmreq_opt_sync_kloop_mode {
902 struct nmreq_option nro_opt; /* common header */
903 #define NM_OPT_SYNC_KLOOP_DIRECT_TX (1 << 0)
904 #define NM_OPT_SYNC_KLOOP_DIRECT_RX (1 << 1)
905 uint32_t mode;
906 };
907
908 struct nmreq_opt_extmem {
909 struct nmreq_option nro_opt; /* common header */
910 uint64_t nro_usrptr; /* (in) ptr to usr memory */
911 struct nmreq_pools_info nro_info; /* (in/out) */
912 };
913
914 struct nmreq_opt_csb {
915 struct nmreq_option nro_opt;
916
917 /* Array of CSB entries for application --> kernel communication
918 * (N entries). */
919 uint64_t csb_atok;
920
921 /* Array of CSB entries for kernel --> application communication
922 * (N entries). */
923 uint64_t csb_ktoa;
924 };
925
926 #endif /* _NET_NETMAP_H_ */
Cache object: bc1607c9ff6395efcd70d7b5e0f510a4
|