FreeBSD/Linux Kernel Cross Reference
sys/net/netmap.h
1 /*-
2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3 *
4 * Copyright (C) 2011-2014 Matteo Landi, Luigi Rizzo. All rights reserved.
5 *
6 * Redistribution and use in source and binary forms, with or without
7 * modification, are permitted provided that the following conditions
8 * are met:
9 *
10 * 1. Redistributions of source code must retain the above copyright
11 * notice, this list of conditions and the following disclaimer.
12 * 2. Redistributions in binary form must reproduce the above copyright
13 * notice, this list of conditions and the following disclaimer in the
14 * documentation and/or other materials provided with the distribution.
15 *
16 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``S IS''AND
17 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
19 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
20 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
21 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
22 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
23 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
24 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
25 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
26 * SUCH DAMAGE.
27 */
28
29 /*
30 * $FreeBSD$
31 *
32 * Definitions of constants and the structures used by the netmap
33 * framework, for the part visible to both kernel and userspace.
34 * Detailed info on netmap is available with "man netmap" or at
35 *
36 * http://info.iet.unipi.it/~luigi/netmap/
37 *
38 * This API is also used to communicate with the VALE software switch
39 */
40
41 #ifndef _NET_NETMAP_H_
42 #define _NET_NETMAP_H_
43
44 #define NETMAP_API 14 /* current API version */
45
46 #define NETMAP_MIN_API 14 /* min and max versions accepted */
47 #define NETMAP_MAX_API 15
48 /*
49 * Some fields should be cache-aligned to reduce contention.
50 * The alignment is architecture and OS dependent, but rather than
51 * digging into OS headers to find the exact value we use an estimate
52 * that should cover most architectures.
53 */
54 #define NM_CACHE_ALIGN 128
55
56 /*
57 * --- Netmap data structures ---
58 *
59 * The userspace data structures used by netmap are shown below.
60 * They are allocated by the kernel and mmap()ed by userspace threads.
61 * Pointers are implemented as memory offsets or indexes,
62 * so that they can be easily dereferenced in kernel and userspace.
63
64 KERNEL (opaque, obviously)
65
66 ====================================================================
67 |
68 USERSPACE | struct netmap_ring
69 +---->+---------------+
70 / | head,cur,tail |
71 struct netmap_if (nifp, 1 per fd) / | buf_ofs |
72 +----------------+ / | other fields |
73 | ni_tx_rings | / +===============+
74 | ni_rx_rings | / | buf_idx, len | slot[0]
75 | | / | flags, ptr |
76 | | / +---------------+
77 +================+ / | buf_idx, len | slot[1]
78 | txring_ofs[0] | (rel.to nifp)--' | flags, ptr |
79 | txring_ofs[1] | +---------------+
80 (tx+htx entries) (num_slots entries)
81 | txring_ofs[t] | | buf_idx, len | slot[n-1]
82 +----------------+ | flags, ptr |
83 | rxring_ofs[0] | +---------------+
84 | rxring_ofs[1] |
85 (rx+hrx entries)
86 | rxring_ofs[r] |
87 +----------------+
88
89 * For each "interface" (NIC, host stack, PIPE, VALE switch port) bound to
90 * a file descriptor, the mmap()ed region contains a (logically readonly)
91 * struct netmap_if pointing to struct netmap_ring's.
92 *
93 * There is one netmap_ring per physical NIC ring, plus at least one tx/rx ring
94 * pair attached to the host stack (these pairs are unused for non-NIC ports).
95 *
96 * All physical/host stack ports share the same memory region,
97 * so that zero-copy can be implemented between them.
98 * VALE switch ports instead have separate memory regions.
99 *
100 * The netmap_ring is the userspace-visible replica of the NIC ring.
101 * Each slot has the index of a buffer (MTU-sized and residing in the
102 * mmapped region), its length and some flags. An extra 64-bit pointer
103 * is provided for user-supplied buffers in the tx path.
104 *
105 * In user space, the buffer address is computed as
106 * (char *)ring + buf_ofs + index * NETMAP_BUF_SIZE
107 *
108 * Added in NETMAP_API 11:
109 *
110 * + NIOCREGIF can request the allocation of extra spare buffers from
111 * the same memory pool. The desired number of buffers must be in
112 * nr_arg3. The ioctl may return fewer buffers, depending on memory
113 * availability. nr_arg3 will return the actual value, and, once
114 * mapped, nifp->ni_bufs_head will be the index of the first buffer.
115 *
116 * The buffers are linked to each other using the first uint32_t
117 * as the index. On close, ni_bufs_head must point to the list of
118 * buffers to be released.
119 *
120 * + NIOCREGIF can attach to PIPE rings sharing the same memory
121 * space with a parent device. The ifname indicates the parent device,
122 * which must already exist. Flags in nr_flags indicate if we want to
123 * bind the master or slave side, the index (from nr_ringid)
124 * is just a cookie and does not need to be sequential.
125 *
126 * + NIOCREGIF can also attach to 'monitor' rings that replicate
127 * the content of specific rings, also from the same memory space.
128 *
129 * Extra flags in nr_flags support the above functions.
130 * Application libraries may use the following naming scheme:
131 * netmap:foo all NIC rings pairs
132 * netmap:foo^ only host rings pairs
133 * netmap:foo^k the k-th host rings pair
134 * netmap:foo+ all NIC rings + host rings pairs
135 * netmap:foo-k the k-th NIC rings pair
136 * netmap:foo{k PIPE rings pair k, master side
137 * netmap:foo}k PIPE rings pair k, slave side
138 *
139 * Some notes about host rings:
140 *
141 * + The RX host rings are used to store those packets that the host network
142 * stack is trying to transmit through a NIC queue, but only if that queue
143 * is currently in netmap mode. Netmap will not intercept host stack mbufs
144 * designated to NIC queues that are not in netmap mode. As a consequence,
145 * registering a netmap port with netmap:foo^ is not enough to intercept
146 * mbufs in the RX host rings; the netmap port should be registered with
147 * netmap:foo*, or another registration should be done to open at least a
148 * NIC TX queue in netmap mode.
149 *
150 * + Netmap is not currently able to deal with intercepted trasmit mbufs which
151 * require offloadings like TSO, UFO, checksumming offloadings, etc. It is
152 * responsibility of the user to disable those offloadings (e.g. using
153 * ifconfig on FreeBSD or ethtool -K on Linux) for an interface that is being
154 * used in netmap mode. If the offloadings are not disabled, GSO and/or
155 * unchecksummed packets may be dropped immediately or end up in the host RX
156 * rings, and will be dropped as soon as the packet reaches another netmap
157 * adapter.
158 */
159
160 /*
161 * struct netmap_slot is a buffer descriptor
162 */
163 struct netmap_slot {
164 uint32_t buf_idx; /* buffer index */
165 uint16_t len; /* length for this slot */
166 uint16_t flags; /* buf changed, etc. */
167 uint64_t ptr; /* pointer for indirect buffers */
168 };
169
170 /*
171 * The following flags control how the slot is used
172 */
173
174 #define NS_BUF_CHANGED 0x0001 /* buf_idx changed */
175 /*
176 * must be set whenever buf_idx is changed (as it might be
177 * necessary to recompute the physical address and mapping)
178 *
179 * It is also set by the kernel whenever the buf_idx is
180 * changed internally (e.g., by pipes). Applications may
181 * use this information to know when they can reuse the
182 * contents of previously prepared buffers.
183 */
184
185 #define NS_REPORT 0x0002 /* ask the hardware to report results */
186 /*
187 * Request notification when slot is used by the hardware.
188 * Normally transmit completions are handled lazily and
189 * may be unreported. This flag lets us know when a slot
190 * has been sent (e.g. to terminate the sender).
191 */
192
193 #define NS_FORWARD 0x0004 /* pass packet 'forward' */
194 /*
195 * (Only for physical ports, rx rings with NR_FORWARD set).
196 * Slot released to the kernel (i.e. before ring->head) with
197 * this flag set are passed to the peer ring (host/NIC),
198 * thus restoring the host-NIC connection for these slots.
199 * This supports efficient traffic monitoring or firewalling.
200 */
201
202 #define NS_NO_LEARN 0x0008 /* disable bridge learning */
203 /*
204 * On a VALE switch, do not 'learn' the source port for
205 * this buffer.
206 */
207
208 #define NS_INDIRECT 0x0010 /* userspace buffer */
209 /*
210 * (VALE tx rings only) data is in a userspace buffer,
211 * whose address is in the 'ptr' field in the slot.
212 */
213
214 #define NS_MOREFRAG 0x0020 /* packet has more fragments */
215 /*
216 * (VALE ports, ptnetmap ports and some NIC ports, e.g.
217 * ixgbe and i40e on Linux)
218 * Set on all but the last slot of a multi-segment packet.
219 * The 'len' field refers to the individual fragment.
220 */
221
222 #define NS_TXMON 0x0040
223 /* (monitor ports only) the packet comes from the TX
224 * ring of the monitored port
225 */
226
227 #define NS_PORT_SHIFT 8
228 #define NS_PORT_MASK (0xff << NS_PORT_SHIFT)
229 /*
230 * The high 8 bits of the flag, if not zero, indicate the
231 * destination port for the VALE switch, overriding
232 * the lookup table.
233 */
234
235 #define NS_RFRAGS(_slot) ( ((_slot)->flags >> 8) & 0xff)
236 /*
237 * (VALE rx rings only) the high 8 bits
238 * are the number of fragments.
239 */
240
241 #define NETMAP_MAX_FRAGS 64 /* max number of fragments */
242
243
244 /*
245 * struct netmap_ring
246 *
247 * Netmap representation of a TX or RX ring (also known as "queue").
248 * This is a queue implemented as a fixed-size circular array.
249 * At the software level the important fields are: head, cur, tail.
250 *
251 * In TX rings:
252 *
253 * head first slot available for transmission.
254 * cur wakeup point. select() and poll() will unblock
255 * when 'tail' moves past 'cur'
256 * tail (readonly) first slot reserved to the kernel
257 *
258 * [head .. tail-1] can be used for new packets to send;
259 * 'head' and 'cur' must be incremented as slots are filled
260 * with new packets to be sent;
261 * 'cur' can be moved further ahead if we need more space
262 * for new transmissions. XXX todo (2014-03-12)
263 *
264 * In RX rings:
265 *
266 * head first valid received packet
267 * cur wakeup point. select() and poll() will unblock
268 * when 'tail' moves past 'cur'
269 * tail (readonly) first slot reserved to the kernel
270 *
271 * [head .. tail-1] contain received packets;
272 * 'head' and 'cur' must be incremented as slots are consumed
273 * and can be returned to the kernel;
274 * 'cur' can be moved further ahead if we want to wait for
275 * new packets without returning the previous ones.
276 *
277 * DATA OWNERSHIP/LOCKING:
278 * The netmap_ring, and all slots and buffers in the range
279 * [head .. tail-1] are owned by the user program;
280 * the kernel only accesses them during a netmap system call
281 * and in the user thread context.
282 *
283 * Other slots and buffers are reserved for use by the kernel
284 */
285 struct netmap_ring {
286 /*
287 * buf_ofs is meant to be used through macros.
288 * It contains the offset of the buffer region from this
289 * descriptor.
290 */
291 const int64_t buf_ofs;
292 const uint32_t num_slots; /* number of slots in the ring. */
293 const uint32_t nr_buf_size;
294 const uint16_t ringid;
295 const uint16_t dir; /* 0: tx, 1: rx */
296
297 uint32_t head; /* (u) first user slot */
298 uint32_t cur; /* (u) wakeup point */
299 uint32_t tail; /* (k) first kernel slot */
300
301 uint32_t flags;
302
303 struct timeval ts; /* (k) time of last *sync() */
304
305 /* opaque room for a mutex or similar object */
306 #if !defined(_WIN32) || defined(__CYGWIN__)
307 uint8_t __attribute__((__aligned__(NM_CACHE_ALIGN))) sem[128];
308 #else
309 uint8_t __declspec(align(NM_CACHE_ALIGN)) sem[128];
310 #endif
311
312 /* the slots follow. This struct has variable size */
313 struct netmap_slot slot[0]; /* array of slots. */
314 };
315
316
317 /*
318 * RING FLAGS
319 */
320 #define NR_TIMESTAMP 0x0002 /* set timestamp on *sync() */
321 /*
322 * updates the 'ts' field on each netmap syscall. This saves
323 * saves a separate gettimeofday(), and is not much worse than
324 * software timestamps generated in the interrupt handler.
325 */
326
327 #define NR_FORWARD 0x0004 /* enable NS_FORWARD for ring */
328 /*
329 * Enables the NS_FORWARD slot flag for the ring.
330 */
331
332 /*
333 * Helper functions for kernel and userspace
334 */
335
336 /*
337 * Check if space is available in the ring. We use ring->head, which
338 * points to the next netmap slot to be published to netmap. It is
339 * possible that the applications moves ring->cur ahead of ring->tail
340 * (e.g., by setting ring->cur <== ring->tail), if it wants more slots
341 * than the ones currently available, and it wants to be notified when
342 * more arrive. See netmap(4) for more details and examples.
343 */
344 static inline int
345 nm_ring_empty(struct netmap_ring *ring)
346 {
347 return (ring->head == ring->tail);
348 }
349
350 /*
351 * Netmap representation of an interface and its queue(s).
352 * This is initialized by the kernel when binding a file
353 * descriptor to a port, and should be considered as readonly
354 * by user programs. The kernel never uses it.
355 *
356 * There is one netmap_if for each file descriptor on which we want
357 * to select/poll.
358 * select/poll operates on one or all pairs depending on the value of
359 * nmr_queueid passed on the ioctl.
360 */
361 struct netmap_if {
362 char ni_name[IFNAMSIZ]; /* name of the interface. */
363 const uint32_t ni_version; /* API version, currently unused */
364 const uint32_t ni_flags; /* properties */
365 #define NI_PRIV_MEM 0x1 /* private memory region */
366
367 /*
368 * The number of packet rings available in netmap mode.
369 * Physical NICs can have different numbers of tx and rx rings.
370 * Physical NICs also have at least a 'host' rings pair.
371 * Additionally, clients can request additional ring pairs to
372 * be used for internal communication.
373 */
374 const uint32_t ni_tx_rings; /* number of HW tx rings */
375 const uint32_t ni_rx_rings; /* number of HW rx rings */
376
377 uint32_t ni_bufs_head; /* head index for extra bufs */
378 const uint32_t ni_host_tx_rings; /* number of SW tx rings */
379 const uint32_t ni_host_rx_rings; /* number of SW rx rings */
380 uint32_t ni_spare1[3];
381 /*
382 * The following array contains the offset of each netmap ring
383 * from this structure, in the following order:
384 * - NIC tx rings (ni_tx_rings);
385 * - host tx rings (ni_host_tx_rings);
386 * - NIC rx rings (ni_rx_rings);
387 * - host rx ring (ni_host_rx_rings);
388 *
389 * The area is filled up by the kernel on NETMAP_REQ_REGISTER,
390 * and then only read by userspace code.
391 */
392 const ssize_t ring_ofs[0];
393 };
394
395 /* Legacy interface to interact with a netmap control device.
396 * Included for backward compatibility. The user should not include this
397 * file directly. */
398 #include "netmap_legacy.h"
399
400 /*
401 * New API to control netmap control devices. New applications should only use
402 * nmreq_xyz structs with the NIOCCTRL ioctl() command.
403 *
404 * NIOCCTRL takes a nmreq_header struct, which contains the required
405 * API version, the name of a netmap port, a command type, and pointers
406 * to request body and options.
407 *
408 * nr_name (in)
409 * The name of the port (em0, valeXXX:YYY, eth0{pn1 etc.)
410 *
411 * nr_version (in/out)
412 * Must match NETMAP_API as used in the kernel, error otherwise.
413 * Always returns the desired value on output.
414 *
415 * nr_reqtype (in)
416 * One of the NETMAP_REQ_* command types below
417 *
418 * nr_body (in)
419 * Pointer to a command-specific struct, described by one
420 * of the struct nmreq_xyz below.
421 *
422 * nr_options (in)
423 * Command specific options, if any.
424 *
425 * A NETMAP_REQ_REGISTER command activates netmap mode on the netmap
426 * port (e.g. physical interface) specified by nmreq_header.nr_name.
427 * The request body (struct nmreq_register) has several arguments to
428 * specify how the port is to be registered.
429 *
430 * nr_tx_slots, nr_tx_slots, nr_tx_rings, nr_rx_rings,
431 * nr_host_tx_rings, nr_host_rx_rings (in/out)
432 * On input, non-zero values may be used to reconfigure the port
433 * according to the requested values, but this is not guaranteed.
434 * On output the actual values in use are reported.
435 *
436 * nr_mode (in)
437 * Indicate what set of rings must be bound to the netmap
438 * device (e.g. all NIC rings, host rings only, NIC and
439 * host rings, ...). Values are in NR_REG_*.
440 *
441 * nr_ringid (in)
442 * If nr_mode == NR_REG_ONE_NIC (only a single couple of TX/RX
443 * rings), indicate which NIC TX and/or RX ring is to be bound
444 * (0..nr_*x_rings-1).
445 *
446 * nr_flags (in)
447 * Indicate special options for how to open the port.
448 *
449 * NR_NO_TX_POLL can be OR-ed to make select()/poll() push
450 * packets on tx rings only if POLLOUT is set.
451 * The default is to push any pending packet.
452 *
453 * NR_DO_RX_POLL can be OR-ed to make select()/poll() release
454 * packets on rx rings also when POLLIN is NOT set.
455 * The default is to touch the rx ring only with POLLIN.
456 * Note that this is the opposite of TX because it
457 * reflects the common usage.
458 *
459 * Other options are NR_MONITOR_TX, NR_MONITOR_RX, NR_ZCOPY_MON,
460 * NR_EXCLUSIVE, NR_RX_RINGS_ONLY, NR_TX_RINGS_ONLY and
461 * NR_ACCEPT_VNET_HDR.
462 *
463 * nr_mem_id (in/out)
464 * The identity of the memory region used.
465 * On input, 0 means the system decides autonomously,
466 * other values may try to select a specific region.
467 * On return the actual value is reported.
468 * Region '1' is the global allocator, normally shared
469 * by all interfaces. Other values are private regions.
470 * If two ports the same region zero-copy is possible.
471 *
472 * nr_extra_bufs (in/out)
473 * Number of extra buffers to be allocated.
474 *
475 * The other NETMAP_REQ_* commands are described below.
476 *
477 */
478
479 /* maximum size of a request, including all options */
480 #define NETMAP_REQ_MAXSIZE 4096
481
482 /* Header common to all request options. */
483 struct nmreq_option {
484 /* Pointer ot the next option. */
485 uint64_t nro_next;
486 /* Option type. */
487 uint32_t nro_reqtype;
488 /* (out) status of the option:
489 * 0: recognized and processed
490 * !=0: errno value
491 */
492 uint32_t nro_status;
493 /* Option size, used only for options that can have variable size
494 * (e.g. because they contain arrays). For fixed-size options this
495 * field should be set to zero. */
496 uint64_t nro_size;
497 };
498
499 /* Header common to all requests. Do not reorder these fields, as we need
500 * the second one (nr_reqtype) to know how much to copy from/to userspace. */
501 struct nmreq_header {
502 uint16_t nr_version; /* API version */
503 uint16_t nr_reqtype; /* nmreq type (NETMAP_REQ_*) */
504 uint32_t nr_reserved; /* must be zero */
505 #define NETMAP_REQ_IFNAMSIZ 64
506 char nr_name[NETMAP_REQ_IFNAMSIZ]; /* port name */
507 uint64_t nr_options; /* command-specific options */
508 uint64_t nr_body; /* ptr to nmreq_xyz struct */
509 };
510
511 enum {
512 /* Register a netmap port with the device. */
513 NETMAP_REQ_REGISTER = 1,
514 /* Get information from a netmap port. */
515 NETMAP_REQ_PORT_INFO_GET,
516 /* Attach a netmap port to a VALE switch. */
517 NETMAP_REQ_VALE_ATTACH,
518 /* Detach a netmap port from a VALE switch. */
519 NETMAP_REQ_VALE_DETACH,
520 /* List the ports attached to a VALE switch. */
521 NETMAP_REQ_VALE_LIST,
522 /* Set the port header length (was virtio-net header length). */
523 NETMAP_REQ_PORT_HDR_SET,
524 /* Get the port header length (was virtio-net header length). */
525 NETMAP_REQ_PORT_HDR_GET,
526 /* Create a new persistent VALE port. */
527 NETMAP_REQ_VALE_NEWIF,
528 /* Delete a persistent VALE port. */
529 NETMAP_REQ_VALE_DELIF,
530 /* Enable polling kernel thread(s) on an attached VALE port. */
531 NETMAP_REQ_VALE_POLLING_ENABLE,
532 /* Disable polling kernel thread(s) on an attached VALE port. */
533 NETMAP_REQ_VALE_POLLING_DISABLE,
534 /* Get info about the pools of a memory allocator. */
535 NETMAP_REQ_POOLS_INFO_GET,
536 /* Start an in-kernel loop that syncs the rings periodically or
537 * on notifications. The loop runs in the context of the ioctl
538 * syscall, and only stops on NETMAP_REQ_SYNC_KLOOP_STOP. */
539 NETMAP_REQ_SYNC_KLOOP_START,
540 /* Stops the thread executing the in-kernel loop. The thread
541 * returns from the ioctl syscall. */
542 NETMAP_REQ_SYNC_KLOOP_STOP,
543 /* Enable CSB mode on a registered netmap control device. */
544 NETMAP_REQ_CSB_ENABLE,
545 };
546
547 enum {
548 /* On NETMAP_REQ_REGISTER, ask netmap to use memory allocated
549 * from user-space allocated memory pools (e.g. hugepages).
550 */
551 NETMAP_REQ_OPT_EXTMEM = 1,
552
553 /* ON NETMAP_REQ_SYNC_KLOOP_START, ask netmap to use eventfd-based
554 * notifications to synchronize the kernel loop with the application.
555 */
556 NETMAP_REQ_OPT_SYNC_KLOOP_EVENTFDS,
557
558 /* On NETMAP_REQ_REGISTER, ask netmap to work in CSB mode, where
559 * head, cur and tail pointers are not exchanged through the
560 * struct netmap_ring header, but rather using an user-provided
561 * memory area (see struct nm_csb_atok and struct nm_csb_ktoa).
562 */
563 NETMAP_REQ_OPT_CSB,
564
565 /* An extension to NETMAP_REQ_OPT_SYNC_KLOOP_EVENTFDS, which specifies
566 * if the TX and/or RX rings are synced in the context of the VM exit.
567 * This requires the 'ioeventfd' fields to be valid (cannot be < 0).
568 */
569 NETMAP_REQ_OPT_SYNC_KLOOP_MODE,
570
571 /* This is a marker to count the number of available options.
572 * New options must be added above it. */
573 NETMAP_REQ_OPT_MAX,
574 };
575
576 /*
577 * nr_reqtype: NETMAP_REQ_REGISTER
578 * Bind (register) a netmap port to this control device.
579 */
580 struct nmreq_register {
581 uint64_t nr_offset; /* nifp offset in the shared region */
582 uint64_t nr_memsize; /* size of the shared region */
583 uint32_t nr_tx_slots; /* slots in tx rings */
584 uint32_t nr_rx_slots; /* slots in rx rings */
585 uint16_t nr_tx_rings; /* number of tx rings */
586 uint16_t nr_rx_rings; /* number of rx rings */
587 uint16_t nr_host_tx_rings; /* number of host tx rings */
588 uint16_t nr_host_rx_rings; /* number of host rx rings */
589
590 uint16_t nr_mem_id; /* id of the memory allocator */
591 uint16_t nr_ringid; /* ring(s) we care about */
592 uint32_t nr_mode; /* specify NR_REG_* modes */
593 uint32_t nr_extra_bufs; /* number of requested extra buffers */
594
595 uint64_t nr_flags; /* additional flags (see below) */
596 /* monitors use nr_ringid and nr_mode to select the rings to monitor */
597 #define NR_MONITOR_TX 0x100
598 #define NR_MONITOR_RX 0x200
599 #define NR_ZCOPY_MON 0x400
600 /* request exclusive access to the selected rings */
601 #define NR_EXCLUSIVE 0x800
602 /* 0x1000 unused */
603 #define NR_RX_RINGS_ONLY 0x2000
604 #define NR_TX_RINGS_ONLY 0x4000
605 /* Applications set this flag if they are able to deal with virtio-net headers,
606 * that is send/receive frames that start with a virtio-net header.
607 * If not set, NETMAP_REQ_REGISTER will fail with netmap ports that require
608 * applications to use those headers. If the flag is set, the application can
609 * use the NETMAP_VNET_HDR_GET command to figure out the header length. */
610 #define NR_ACCEPT_VNET_HDR 0x8000
611 /* The following two have the same meaning of NETMAP_NO_TX_POLL and
612 * NETMAP_DO_RX_POLL. */
613 #define NR_DO_RX_POLL 0x10000
614 #define NR_NO_TX_POLL 0x20000
615 };
616
617 /* Valid values for nmreq_register.nr_mode (see above). */
618 enum { NR_REG_DEFAULT = 0, /* backward compat, should not be used. */
619 NR_REG_ALL_NIC = 1,
620 NR_REG_SW = 2,
621 NR_REG_NIC_SW = 3,
622 NR_REG_ONE_NIC = 4,
623 NR_REG_PIPE_MASTER = 5, /* deprecated, use "x{y" port name syntax */
624 NR_REG_PIPE_SLAVE = 6, /* deprecated, use "x}y" port name syntax */
625 NR_REG_NULL = 7,
626 NR_REG_ONE_SW = 8,
627 };
628
629 /* A single ioctl number is shared by all the new API command.
630 * Demultiplexing is done using the hdr.nr_reqtype field.
631 * FreeBSD uses the size value embedded in the _IOWR to determine
632 * how much to copy in/out, so we define the ioctl() command
633 * specifying only nmreq_header, and copyin/copyout the rest. */
634 #define NIOCCTRL _IOWR('i', 151, struct nmreq_header)
635
636 /* The ioctl commands to sync TX/RX netmap rings.
637 * NIOCTXSYNC, NIOCRXSYNC synchronize tx or rx queues,
638 * whose identity is set in NETMAP_REQ_REGISTER through nr_ringid.
639 * These are non blocking and take no argument. */
640 #define NIOCTXSYNC _IO('i', 148) /* sync tx queues */
641 #define NIOCRXSYNC _IO('i', 149) /* sync rx queues */
642
643 /*
644 * nr_reqtype: NETMAP_REQ_PORT_INFO_GET
645 * Get information about a netmap port, including number of rings.
646 * slots per ring, id of the memory allocator, etc. The netmap
647 * control device used for this operation does not need to be bound
648 * to a netmap port.
649 */
650 struct nmreq_port_info_get {
651 uint64_t nr_memsize; /* size of the shared region */
652 uint32_t nr_tx_slots; /* slots in tx rings */
653 uint32_t nr_rx_slots; /* slots in rx rings */
654 uint16_t nr_tx_rings; /* number of tx rings */
655 uint16_t nr_rx_rings; /* number of rx rings */
656 uint16_t nr_host_tx_rings; /* number of host tx rings */
657 uint16_t nr_host_rx_rings; /* number of host rx rings */
658 uint16_t nr_mem_id; /* memory allocator id (in/out) */
659 uint16_t pad[3];
660 };
661
662 #define NM_BDG_NAME "vale" /* prefix for bridge port name */
663
664 /*
665 * nr_reqtype: NETMAP_REQ_VALE_ATTACH
666 * Attach a netmap port to a VALE switch. Both the name of the netmap
667 * port and the VALE switch are specified through the nr_name argument.
668 * The attach operation could need to register a port, so at least
669 * the same arguments are available.
670 * port_index will contain the index where the port has been attached.
671 */
672 struct nmreq_vale_attach {
673 struct nmreq_register reg;
674 uint32_t port_index;
675 uint32_t pad1;
676 };
677
678 /*
679 * nr_reqtype: NETMAP_REQ_VALE_DETACH
680 * Detach a netmap port from a VALE switch. Both the name of the netmap
681 * port and the VALE switch are specified through the nr_name argument.
682 * port_index will contain the index where the port was attached.
683 */
684 struct nmreq_vale_detach {
685 uint32_t port_index;
686 uint32_t pad1;
687 };
688
689 /*
690 * nr_reqtype: NETMAP_REQ_VALE_LIST
691 * List the ports of a VALE switch.
692 */
693 struct nmreq_vale_list {
694 /* Name of the VALE port (valeXXX:YYY) or empty. */
695 uint16_t nr_bridge_idx;
696 uint16_t pad1;
697 uint32_t nr_port_idx;
698 };
699
700 /*
701 * nr_reqtype: NETMAP_REQ_PORT_HDR_SET or NETMAP_REQ_PORT_HDR_GET
702 * Set or get the port header length of the port identified by hdr.nr_name.
703 * The control device does not need to be bound to a netmap port.
704 */
705 struct nmreq_port_hdr {
706 uint32_t nr_hdr_len;
707 uint32_t pad1;
708 };
709
710 /*
711 * nr_reqtype: NETMAP_REQ_VALE_NEWIF
712 * Create a new persistent VALE port.
713 */
714 struct nmreq_vale_newif {
715 uint32_t nr_tx_slots; /* slots in tx rings */
716 uint32_t nr_rx_slots; /* slots in rx rings */
717 uint16_t nr_tx_rings; /* number of tx rings */
718 uint16_t nr_rx_rings; /* number of rx rings */
719 uint16_t nr_mem_id; /* id of the memory allocator */
720 uint16_t pad1;
721 };
722
723 /*
724 * nr_reqtype: NETMAP_REQ_VALE_POLLING_ENABLE or NETMAP_REQ_VALE_POLLING_DISABLE
725 * Enable or disable polling kthreads on a VALE port.
726 */
727 struct nmreq_vale_polling {
728 uint32_t nr_mode;
729 #define NETMAP_POLLING_MODE_SINGLE_CPU 1
730 #define NETMAP_POLLING_MODE_MULTI_CPU 2
731 uint32_t nr_first_cpu_id;
732 uint32_t nr_num_polling_cpus;
733 uint32_t pad1;
734 };
735
736 /*
737 * nr_reqtype: NETMAP_REQ_POOLS_INFO_GET
738 * Get info about the pools of the memory allocator of the netmap
739 * port specified by hdr.nr_name and nr_mem_id. The netmap control
740 * device used for this operation does not need to be bound to a netmap
741 * port.
742 */
743 struct nmreq_pools_info {
744 uint64_t nr_memsize;
745 uint16_t nr_mem_id; /* in/out argument */
746 uint16_t pad1[3];
747 uint64_t nr_if_pool_offset;
748 uint32_t nr_if_pool_objtotal;
749 uint32_t nr_if_pool_objsize;
750 uint64_t nr_ring_pool_offset;
751 uint32_t nr_ring_pool_objtotal;
752 uint32_t nr_ring_pool_objsize;
753 uint64_t nr_buf_pool_offset;
754 uint32_t nr_buf_pool_objtotal;
755 uint32_t nr_buf_pool_objsize;
756 };
757
758 /*
759 * nr_reqtype: NETMAP_REQ_SYNC_KLOOP_START
760 * Start an in-kernel loop that syncs the rings periodically or on
761 * notifications. The loop runs in the context of the ioctl syscall,
762 * and only stops on NETMAP_REQ_SYNC_KLOOP_STOP.
763 * The registered netmap port must be open in CSB mode.
764 */
765 struct nmreq_sync_kloop_start {
766 /* Sleeping is the default synchronization method for the kloop.
767 * The 'sleep_us' field specifies how many microseconds to sleep for
768 * when there is no work to do, before doing another kloop iteration.
769 */
770 uint32_t sleep_us;
771 uint32_t pad1;
772 };
773
774 /* A CSB entry for the application --> kernel direction. */
775 struct nm_csb_atok {
776 uint32_t head; /* AW+ KR+ the head of the appl netmap_ring */
777 uint32_t cur; /* AW+ KR+ the cur of the appl netmap_ring */
778 uint32_t appl_need_kick; /* AW+ KR+ kern --> appl notification enable */
779 uint32_t sync_flags; /* AW+ KR+ the flags of the appl [tx|rx]sync() */
780 uint32_t pad[12]; /* pad to a 64 bytes cacheline */
781 };
782
783 /* A CSB entry for the application <-- kernel direction. */
784 struct nm_csb_ktoa {
785 uint32_t hwcur; /* AR+ KW+ the hwcur of the kern netmap_kring */
786 uint32_t hwtail; /* AR+ KW+ the hwtail of the kern netmap_kring */
787 uint32_t kern_need_kick; /* AR+ KW+ appl-->kern notification enable */
788 uint32_t pad[13];
789 };
790
791 #ifdef __linux__
792
793 #ifdef __KERNEL__
794 #define nm_stst_barrier smp_wmb
795 #define nm_ldld_barrier smp_rmb
796 #define nm_stld_barrier smp_mb
797 #else /* !__KERNEL__ */
798 static inline void nm_stst_barrier(void)
799 {
800 /* A memory barrier with release semantic has the combined
801 * effect of a store-store barrier and a load-store barrier,
802 * which is fine for us. */
803 __atomic_thread_fence(__ATOMIC_RELEASE);
804 }
805 static inline void nm_ldld_barrier(void)
806 {
807 /* A memory barrier with acquire semantic has the combined
808 * effect of a load-load barrier and a store-load barrier,
809 * which is fine for us. */
810 __atomic_thread_fence(__ATOMIC_ACQUIRE);
811 }
812 #endif /* !__KERNEL__ */
813
814 #elif defined(__FreeBSD__)
815
816 #ifdef _KERNEL
817 #define nm_stst_barrier atomic_thread_fence_rel
818 #define nm_ldld_barrier atomic_thread_fence_acq
819 #define nm_stld_barrier atomic_thread_fence_seq_cst
820 #else /* !_KERNEL */
821 #include <stdatomic.h>
822 static inline void nm_stst_barrier(void)
823 {
824 atomic_thread_fence(memory_order_release);
825 }
826 static inline void nm_ldld_barrier(void)
827 {
828 atomic_thread_fence(memory_order_acquire);
829 }
830 #endif /* !_KERNEL */
831
832 #else /* !__linux__ && !__FreeBSD__ */
833 #error "OS not supported"
834 #endif /* !__linux__ && !__FreeBSD__ */
835
836 /* Application side of sync-kloop: Write ring pointers (cur, head) to the CSB.
837 * This routine is coupled with sync_kloop_kernel_read(). */
838 static inline void
839 nm_sync_kloop_appl_write(struct nm_csb_atok *atok, uint32_t cur,
840 uint32_t head)
841 {
842 /* Issue a first store-store barrier to make sure writes to the
843 * netmap ring do not overcome updates on atok->cur and atok->head. */
844 nm_stst_barrier();
845
846 /*
847 * We need to write cur and head to the CSB but we cannot do it atomically.
848 * There is no way we can prevent the host from reading the updated value
849 * of one of the two and the old value of the other. However, if we make
850 * sure that the host never reads a value of head more recent than the
851 * value of cur we are safe. We can allow the host to read a value of cur
852 * more recent than the value of head, since in the netmap ring cur can be
853 * ahead of head and cur cannot wrap around head because it must be behind
854 * tail. Inverting the order of writes below could instead result into the
855 * host to think head went ahead of cur, which would cause the sync
856 * prologue to fail.
857 *
858 * The following memory barrier scheme is used to make this happen:
859 *
860 * Guest Host
861 *
862 * STORE(cur) LOAD(head)
863 * wmb() <-----------> rmb()
864 * STORE(head) LOAD(cur)
865 *
866 */
867 atok->cur = cur;
868 nm_stst_barrier();
869 atok->head = head;
870 }
871
872 /* Application side of sync-kloop: Read kring pointers (hwcur, hwtail) from
873 * the CSB. This routine is coupled with sync_kloop_kernel_write(). */
874 static inline void
875 nm_sync_kloop_appl_read(struct nm_csb_ktoa *ktoa, uint32_t *hwtail,
876 uint32_t *hwcur)
877 {
878 /*
879 * We place a memory barrier to make sure that the update of hwtail never
880 * overtakes the update of hwcur.
881 * (see explanation in sync_kloop_kernel_write).
882 */
883 *hwtail = ktoa->hwtail;
884 nm_ldld_barrier();
885 *hwcur = ktoa->hwcur;
886
887 /* Make sure that loads from ktoa->hwtail and ktoa->hwcur are not delayed
888 * after the loads from the netmap ring. */
889 nm_ldld_barrier();
890 }
891
892 /*
893 * data for NETMAP_REQ_OPT_* options
894 */
895
896 struct nmreq_opt_sync_kloop_eventfds {
897 struct nmreq_option nro_opt; /* common header */
898 /* An array of N entries for bidirectional notifications between
899 * the kernel loop and the application. The number of entries and
900 * their order must agree with the CSB arrays passed in the
901 * NETMAP_REQ_OPT_CSB option. Each entry contains a file descriptor
902 * backed by an eventfd.
903 *
904 * If any of the 'ioeventfd' entries is < 0, the event loop uses
905 * the sleeping synchronization strategy (according to sleep_us),
906 * and keeps kern_need_kick always disabled.
907 * Each 'irqfd' can be < 0, and in that case the corresponding queue
908 * is never notified.
909 */
910 struct {
911 /* Notifier for the application --> kernel loop direction. */
912 int32_t ioeventfd;
913 /* Notifier for the kernel loop --> application direction. */
914 int32_t irqfd;
915 } eventfds[0];
916 };
917
918 struct nmreq_opt_sync_kloop_mode {
919 struct nmreq_option nro_opt; /* common header */
920 #define NM_OPT_SYNC_KLOOP_DIRECT_TX (1 << 0)
921 #define NM_OPT_SYNC_KLOOP_DIRECT_RX (1 << 1)
922 uint32_t mode;
923 };
924
925 struct nmreq_opt_extmem {
926 struct nmreq_option nro_opt; /* common header */
927 uint64_t nro_usrptr; /* (in) ptr to usr memory */
928 struct nmreq_pools_info nro_info; /* (in/out) */
929 };
930
931 struct nmreq_opt_csb {
932 struct nmreq_option nro_opt;
933
934 /* Array of CSB entries for application --> kernel communication
935 * (N entries). */
936 uint64_t csb_atok;
937
938 /* Array of CSB entries for kernel --> application communication
939 * (N entries). */
940 uint64_t csb_ktoa;
941 };
942
943 #endif /* _NET_NETMAP_H_ */
Cache object: b695e5ec66b5cdee6c50c05aa3570a47
|