2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
4 * Copyright (C) 2011-2014 Matteo Landi, Luigi Rizzo. All rights reserved.
6 * Redistribution and use in source and binary forms, with or without
7 * modification, are permitted provided that the following conditions
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.
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
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
36 * http://info.iet.unipi.it/~luigi/netmap/
38 * This API is also used to communicate with the VALE software switch
41 #ifndef _NET_NETMAP_H_
42 #define _NET_NETMAP_H_
44 #define NETMAP_API 14 /* current API version */
46 #define NETMAP_MIN_API 14 /* min and max versions accepted */
47 #define NETMAP_MAX_API 15
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.
54 #define NM_CACHE_ALIGN 128
57 * --- Netmap data structures ---
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.
64 KERNEL (opaque, obviously)
66 ====================================================================
68 USERSPACE | struct netmap_ring
69 +---->+---------------+
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]
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] | +---------------+
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.
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).
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.
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.
105 * In user space, the buffer address is computed as
106 * (char *)ring + buf_ofs + index * NETMAP_BUF_SIZE
108 * Added in NETMAP_API 11:
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.
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.
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.
126 * + NIOCREGIF can also attach to 'monitor' rings that replicate
127 * the content of specific rings, also from the same memory space.
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
139 * Some notes about host rings:
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.
150 * + Netmap is not currently able to deal with intercepted transmit 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
161 * struct netmap_slot is a buffer descriptor
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 */
171 * The following flags control how the slot is used
174 #define NS_BUF_CHANGED 0x0001 /* buf_idx changed */
176 * must be set whenever buf_idx is changed (as it might be
177 * necessary to recompute the physical address and mapping)
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.
185 #define NS_REPORT 0x0002 /* ask the hardware to report results */
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).
193 #define NS_FORWARD 0x0004 /* pass packet 'forward' */
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.
202 #define NS_NO_LEARN 0x0008 /* disable bridge learning */
204 * On a VALE switch, do not 'learn' the source port for
208 #define NS_INDIRECT 0x0010 /* userspace buffer */
210 * (VALE tx rings only) data is in a userspace buffer,
211 * whose address is in the 'ptr' field in the slot.
214 #define NS_MOREFRAG 0x0020 /* packet has more fragments */
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.
222 #define NS_TXMON 0x0040
223 /* (monitor ports only) the packet comes from the TX
224 * ring of the monitored port
227 #define NS_PORT_SHIFT 8
228 #define NS_PORT_MASK (0xff << NS_PORT_SHIFT)
230 * The high 8 bits of the flag, if not zero, indicate the
231 * destination port for the VALE switch, overriding
235 #define NS_RFRAGS(_slot) ( ((_slot)->flags >> 8) & 0xff)
237 * (VALE rx rings only) the high 8 bits
238 * are the number of fragments.
241 #define NETMAP_MAX_FRAGS 64 /* max number of fragments */
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.
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
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)
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
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.
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.
283 * Other slots and buffers are reserved for use by the kernel
287 * buf_ofs is meant to be used through macros.
288 * It contains the offset of the buffer region from this
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 */
297 uint32_t head; /* (u) first user slot */
298 uint32_t cur; /* (u) wakeup point */
299 uint32_t tail; /* (k) first kernel slot */
303 struct timeval ts; /* (k) time of last *sync() */
305 /* offset_mask is used to isolate the part of the ptr field
306 * in the slots used to contain an offset in the buffer.
307 * It is zero if the ring has not be opened using the
308 * NETMAP_REQ_OPT_OFFSETS option.
310 const uint64_t offset_mask;
311 /* the alignment requirement, in bytes, for the start
312 * of the packets inside the buffers.
313 * User programs should take this alignment into
314 * account when specifying buffer-offsets in TX slots.
316 const uint64_t buf_align;
318 /* opaque room for a mutex or similar object */
319 #if !defined(_WIN32) || defined(__CYGWIN__)
320 uint8_t __attribute__((__aligned__(NM_CACHE_ALIGN))) sem[128];
322 uint8_t __declspec(align(NM_CACHE_ALIGN)) sem[128];
325 /* the slots follow. This struct has variable size */
326 struct netmap_slot slot[0]; /* array of slots. */
333 #define NR_TIMESTAMP 0x0002 /* set timestamp on *sync() */
335 * updates the 'ts' field on each netmap syscall. This saves
336 * saves a separate gettimeofday(), and is not much worse than
337 * software timestamps generated in the interrupt handler.
340 #define NR_FORWARD 0x0004 /* enable NS_FORWARD for ring */
342 * Enables the NS_FORWARD slot flag for the ring.
346 * Helper functions for kernel and userspace
350 * Check if space is available in the ring. We use ring->head, which
351 * points to the next netmap slot to be published to netmap. It is
352 * possible that the applications moves ring->cur ahead of ring->tail
353 * (e.g., by setting ring->cur <== ring->tail), if it wants more slots
354 * than the ones currently available, and it wants to be notified when
355 * more arrive. See netmap(4) for more details and examples.
358 nm_ring_empty(struct netmap_ring *ring)
360 return (ring->head == ring->tail);
364 * Netmap representation of an interface and its queue(s).
365 * This is initialized by the kernel when binding a file
366 * descriptor to a port, and should be considered as readonly
367 * by user programs. The kernel never uses it.
369 * There is one netmap_if for each file descriptor on which we want
371 * select/poll operates on one or all pairs depending on the value of
372 * nmr_queueid passed on the ioctl.
375 char ni_name[IFNAMSIZ]; /* name of the interface. */
376 const uint32_t ni_version; /* API version, currently unused */
377 const uint32_t ni_flags; /* properties */
378 #define NI_PRIV_MEM 0x1 /* private memory region */
381 * The number of packet rings available in netmap mode.
382 * Physical NICs can have different numbers of tx and rx rings.
383 * Physical NICs also have at least a 'host' rings pair.
384 * Additionally, clients can request additional ring pairs to
385 * be used for internal communication.
387 const uint32_t ni_tx_rings; /* number of HW tx rings */
388 const uint32_t ni_rx_rings; /* number of HW rx rings */
390 uint32_t ni_bufs_head; /* head index for extra bufs */
391 const uint32_t ni_host_tx_rings; /* number of SW tx rings */
392 const uint32_t ni_host_rx_rings; /* number of SW rx rings */
393 uint32_t ni_spare1[3];
395 * The following array contains the offset of each netmap ring
396 * from this structure, in the following order:
397 * - NIC tx rings (ni_tx_rings);
398 * - host tx rings (ni_host_tx_rings);
399 * - NIC rx rings (ni_rx_rings);
400 * - host rx ring (ni_host_rx_rings);
402 * The area is filled up by the kernel on NETMAP_REQ_REGISTER,
403 * and then only read by userspace code.
405 const ssize_t ring_ofs[0];
408 /* Legacy interface to interact with a netmap control device.
409 * Included for backward compatibility. The user should not include this
411 #include "netmap_legacy.h"
414 * New API to control netmap control devices. New applications should only use
415 * nmreq_xyz structs with the NIOCCTRL ioctl() command.
417 * NIOCCTRL takes a nmreq_header struct, which contains the required
418 * API version, the name of a netmap port, a command type, and pointers
419 * to request body and options.
422 * The name of the port (em0, valeXXX:YYY, eth0{pn1 etc.)
424 * nr_version (in/out)
425 * Must match NETMAP_API as used in the kernel, error otherwise.
426 * Always returns the desired value on output.
429 * One of the NETMAP_REQ_* command types below
432 * Pointer to a command-specific struct, described by one
433 * of the struct nmreq_xyz below.
436 * Command specific options, if any.
438 * A NETMAP_REQ_REGISTER command activates netmap mode on the netmap
439 * port (e.g. physical interface) specified by nmreq_header.nr_name.
440 * The request body (struct nmreq_register) has several arguments to
441 * specify how the port is to be registered.
443 * nr_tx_slots, nr_tx_slots, nr_tx_rings, nr_rx_rings,
444 * nr_host_tx_rings, nr_host_rx_rings (in/out)
445 * On input, non-zero values may be used to reconfigure the port
446 * according to the requested values, but this is not guaranteed.
447 * On output the actual values in use are reported.
450 * Indicate what set of rings must be bound to the netmap
451 * device (e.g. all NIC rings, host rings only, NIC and
452 * host rings, ...). Values are in NR_REG_*.
455 * If nr_mode == NR_REG_ONE_NIC (only a single couple of TX/RX
456 * rings), indicate which NIC TX and/or RX ring is to be bound
457 * (0..nr_*x_rings-1).
460 * Indicate special options for how to open the port.
462 * NR_NO_TX_POLL can be OR-ed to make select()/poll() push
463 * packets on tx rings only if POLLOUT is set.
464 * The default is to push any pending packet.
466 * NR_DO_RX_POLL can be OR-ed to make select()/poll() release
467 * packets on rx rings also when POLLIN is NOT set.
468 * The default is to touch the rx ring only with POLLIN.
469 * Note that this is the opposite of TX because it
470 * reflects the common usage.
472 * Other options are NR_MONITOR_TX, NR_MONITOR_RX, NR_ZCOPY_MON,
473 * NR_EXCLUSIVE, NR_RX_RINGS_ONLY, NR_TX_RINGS_ONLY and
474 * NR_ACCEPT_VNET_HDR.
477 * The identity of the memory region used.
478 * On input, 0 means the system decides autonomously,
479 * other values may try to select a specific region.
480 * On return the actual value is reported.
481 * Region '1' is the global allocator, normally shared
482 * by all interfaces. Other values are private regions.
483 * If two ports the same region zero-copy is possible.
485 * nr_extra_bufs (in/out)
486 * Number of extra buffers to be allocated.
488 * The other NETMAP_REQ_* commands are described below.
492 /* maximum size of a request, including all options */
493 #define NETMAP_REQ_MAXSIZE 4096
495 /* Header common to all request options. */
496 struct nmreq_option {
497 /* Pointer to the next option. */
500 uint32_t nro_reqtype;
501 /* (out) status of the option:
502 * 0: recognized and processed
506 /* Option size, used only for options that can have variable size
507 * (e.g. because they contain arrays). For fixed-size options this
508 * field should be set to zero. */
512 /* Header common to all requests. Do not reorder these fields, as we need
513 * the second one (nr_reqtype) to know how much to copy from/to userspace. */
514 struct nmreq_header {
515 uint16_t nr_version; /* API version */
516 uint16_t nr_reqtype; /* nmreq type (NETMAP_REQ_*) */
517 uint32_t nr_reserved; /* must be zero */
518 #define NETMAP_REQ_IFNAMSIZ 64
519 char nr_name[NETMAP_REQ_IFNAMSIZ]; /* port name */
520 uint64_t nr_options; /* command-specific options */
521 uint64_t nr_body; /* ptr to nmreq_xyz struct */
525 /* Register a netmap port with the device. */
526 NETMAP_REQ_REGISTER = 1,
527 /* Get information from a netmap port. */
528 NETMAP_REQ_PORT_INFO_GET,
529 /* Attach a netmap port to a VALE switch. */
530 NETMAP_REQ_VALE_ATTACH,
531 /* Detach a netmap port from a VALE switch. */
532 NETMAP_REQ_VALE_DETACH,
533 /* List the ports attached to a VALE switch. */
534 NETMAP_REQ_VALE_LIST,
535 /* Set the port header length (was virtio-net header length). */
536 NETMAP_REQ_PORT_HDR_SET,
537 /* Get the port header length (was virtio-net header length). */
538 NETMAP_REQ_PORT_HDR_GET,
539 /* Create a new persistent VALE port. */
540 NETMAP_REQ_VALE_NEWIF,
541 /* Delete a persistent VALE port. */
542 NETMAP_REQ_VALE_DELIF,
543 /* Enable polling kernel thread(s) on an attached VALE port. */
544 NETMAP_REQ_VALE_POLLING_ENABLE,
545 /* Disable polling kernel thread(s) on an attached VALE port. */
546 NETMAP_REQ_VALE_POLLING_DISABLE,
547 /* Get info about the pools of a memory allocator. */
548 NETMAP_REQ_POOLS_INFO_GET,
549 /* Start an in-kernel loop that syncs the rings periodically or
550 * on notifications. The loop runs in the context of the ioctl
551 * syscall, and only stops on NETMAP_REQ_SYNC_KLOOP_STOP. */
552 NETMAP_REQ_SYNC_KLOOP_START,
553 /* Stops the thread executing the in-kernel loop. The thread
554 * returns from the ioctl syscall. */
555 NETMAP_REQ_SYNC_KLOOP_STOP,
556 /* Enable CSB mode on a registered netmap control device. */
557 NETMAP_REQ_CSB_ENABLE,
561 /* On NETMAP_REQ_REGISTER, ask netmap to use memory allocated
562 * from user-space allocated memory pools (e.g. hugepages).
564 NETMAP_REQ_OPT_EXTMEM = 1,
566 /* ON NETMAP_REQ_SYNC_KLOOP_START, ask netmap to use eventfd-based
567 * notifications to synchronize the kernel loop with the application.
569 NETMAP_REQ_OPT_SYNC_KLOOP_EVENTFDS,
571 /* On NETMAP_REQ_REGISTER, ask netmap to work in CSB mode, where
572 * head, cur and tail pointers are not exchanged through the
573 * struct netmap_ring header, but rather using an user-provided
574 * memory area (see struct nm_csb_atok and struct nm_csb_ktoa).
578 /* An extension to NETMAP_REQ_OPT_SYNC_KLOOP_EVENTFDS, which specifies
579 * if the TX and/or RX rings are synced in the context of the VM exit.
580 * This requires the 'ioeventfd' fields to be valid (cannot be < 0).
582 NETMAP_REQ_OPT_SYNC_KLOOP_MODE,
584 /* On NETMAP_REQ_REGISTER, ask for (part of) the ptr field in the
585 * slots of the registered rings to be used as an offset field
586 * for the start of the packets inside the netmap buffer.
588 NETMAP_REQ_OPT_OFFSETS,
590 /* This is a marker to count the number of available options.
591 * New options must be added above it. */
596 * nr_reqtype: NETMAP_REQ_REGISTER
597 * Bind (register) a netmap port to this control device.
599 struct nmreq_register {
600 uint64_t nr_offset; /* nifp offset in the shared region */
601 uint64_t nr_memsize; /* size of the shared region */
602 uint32_t nr_tx_slots; /* slots in tx rings */
603 uint32_t nr_rx_slots; /* slots in rx rings */
604 uint16_t nr_tx_rings; /* number of tx rings */
605 uint16_t nr_rx_rings; /* number of rx rings */
606 uint16_t nr_host_tx_rings; /* number of host tx rings */
607 uint16_t nr_host_rx_rings; /* number of host rx rings */
609 uint16_t nr_mem_id; /* id of the memory allocator */
610 uint16_t nr_ringid; /* ring(s) we care about */
611 uint32_t nr_mode; /* specify NR_REG_* modes */
612 uint32_t nr_extra_bufs; /* number of requested extra buffers */
614 uint64_t nr_flags; /* additional flags (see below) */
615 /* monitors use nr_ringid and nr_mode to select the rings to monitor */
616 #define NR_MONITOR_TX 0x100
617 #define NR_MONITOR_RX 0x200
618 #define NR_ZCOPY_MON 0x400
619 /* request exclusive access to the selected rings */
620 #define NR_EXCLUSIVE 0x800
622 #define NR_RX_RINGS_ONLY 0x2000
623 #define NR_TX_RINGS_ONLY 0x4000
624 /* Applications set this flag if they are able to deal with virtio-net headers,
625 * that is send/receive frames that start with a virtio-net header.
626 * If not set, NETMAP_REQ_REGISTER will fail with netmap ports that require
627 * applications to use those headers. If the flag is set, the application can
628 * use the NETMAP_VNET_HDR_GET command to figure out the header length. */
629 #define NR_ACCEPT_VNET_HDR 0x8000
630 /* The following two have the same meaning of NETMAP_NO_TX_POLL and
631 * NETMAP_DO_RX_POLL. */
632 #define NR_DO_RX_POLL 0x10000
633 #define NR_NO_TX_POLL 0x20000
636 /* Valid values for nmreq_register.nr_mode (see above). */
637 enum { NR_REG_DEFAULT = 0, /* backward compat, should not be used. */
642 NR_REG_PIPE_MASTER = 5, /* deprecated, use "x{y" port name syntax */
643 NR_REG_PIPE_SLAVE = 6, /* deprecated, use "x}y" port name syntax */
648 /* A single ioctl number is shared by all the new API command.
649 * Demultiplexing is done using the hdr.nr_reqtype field.
650 * FreeBSD uses the size value embedded in the _IOWR to determine
651 * how much to copy in/out, so we define the ioctl() command
652 * specifying only nmreq_header, and copyin/copyout the rest. */
653 #define NIOCCTRL _IOWR('i', 151, struct nmreq_header)
655 /* The ioctl commands to sync TX/RX netmap rings.
656 * NIOCTXSYNC, NIOCRXSYNC synchronize tx or rx queues,
657 * whose identity is set in NETMAP_REQ_REGISTER through nr_ringid.
658 * These are non blocking and take no argument. */
659 #define NIOCTXSYNC _IO('i', 148) /* sync tx queues */
660 #define NIOCRXSYNC _IO('i', 149) /* sync rx queues */
663 * nr_reqtype: NETMAP_REQ_PORT_INFO_GET
664 * Get information about a netmap port, including number of rings.
665 * slots per ring, id of the memory allocator, etc. The netmap
666 * control device used for this operation does not need to be bound
669 struct nmreq_port_info_get {
670 uint64_t nr_memsize; /* size of the shared region */
671 uint32_t nr_tx_slots; /* slots in tx rings */
672 uint32_t nr_rx_slots; /* slots in rx rings */
673 uint16_t nr_tx_rings; /* number of tx rings */
674 uint16_t nr_rx_rings; /* number of rx rings */
675 uint16_t nr_host_tx_rings; /* number of host tx rings */
676 uint16_t nr_host_rx_rings; /* number of host rx rings */
677 uint16_t nr_mem_id; /* memory allocator id (in/out) */
681 #define NM_BDG_NAME "vale" /* prefix for bridge port name */
684 * nr_reqtype: NETMAP_REQ_VALE_ATTACH
685 * Attach a netmap port to a VALE switch. Both the name of the netmap
686 * port and the VALE switch are specified through the nr_name argument.
687 * The attach operation could need to register a port, so at least
688 * the same arguments are available.
689 * port_index will contain the index where the port has been attached.
691 struct nmreq_vale_attach {
692 struct nmreq_register reg;
698 * nr_reqtype: NETMAP_REQ_VALE_DETACH
699 * Detach a netmap port from a VALE switch. Both the name of the netmap
700 * port and the VALE switch are specified through the nr_name argument.
701 * port_index will contain the index where the port was attached.
703 struct nmreq_vale_detach {
709 * nr_reqtype: NETMAP_REQ_VALE_LIST
710 * List the ports of a VALE switch.
712 struct nmreq_vale_list {
713 /* Name of the VALE port (valeXXX:YYY) or empty. */
714 uint16_t nr_bridge_idx;
716 uint32_t nr_port_idx;
720 * nr_reqtype: NETMAP_REQ_PORT_HDR_SET or NETMAP_REQ_PORT_HDR_GET
721 * Set or get the port header length of the port identified by hdr.nr_name.
722 * The control device does not need to be bound to a netmap port.
724 struct nmreq_port_hdr {
730 * nr_reqtype: NETMAP_REQ_VALE_NEWIF
731 * Create a new persistent VALE port.
733 struct nmreq_vale_newif {
734 uint32_t nr_tx_slots; /* slots in tx rings */
735 uint32_t nr_rx_slots; /* slots in rx rings */
736 uint16_t nr_tx_rings; /* number of tx rings */
737 uint16_t nr_rx_rings; /* number of rx rings */
738 uint16_t nr_mem_id; /* id of the memory allocator */
743 * nr_reqtype: NETMAP_REQ_VALE_POLLING_ENABLE or NETMAP_REQ_VALE_POLLING_DISABLE
744 * Enable or disable polling kthreads on a VALE port.
746 struct nmreq_vale_polling {
748 #define NETMAP_POLLING_MODE_SINGLE_CPU 1
749 #define NETMAP_POLLING_MODE_MULTI_CPU 2
750 uint32_t nr_first_cpu_id;
751 uint32_t nr_num_polling_cpus;
756 * nr_reqtype: NETMAP_REQ_POOLS_INFO_GET
757 * Get info about the pools of the memory allocator of the netmap
758 * port specified by hdr.nr_name and nr_mem_id. The netmap control
759 * device used for this operation does not need to be bound to a netmap
762 struct nmreq_pools_info {
764 uint16_t nr_mem_id; /* in/out argument */
766 uint64_t nr_if_pool_offset;
767 uint32_t nr_if_pool_objtotal;
768 uint32_t nr_if_pool_objsize;
769 uint64_t nr_ring_pool_offset;
770 uint32_t nr_ring_pool_objtotal;
771 uint32_t nr_ring_pool_objsize;
772 uint64_t nr_buf_pool_offset;
773 uint32_t nr_buf_pool_objtotal;
774 uint32_t nr_buf_pool_objsize;
778 * nr_reqtype: NETMAP_REQ_SYNC_KLOOP_START
779 * Start an in-kernel loop that syncs the rings periodically or on
780 * notifications. The loop runs in the context of the ioctl syscall,
781 * and only stops on NETMAP_REQ_SYNC_KLOOP_STOP.
782 * The registered netmap port must be open in CSB mode.
784 struct nmreq_sync_kloop_start {
785 /* Sleeping is the default synchronization method for the kloop.
786 * The 'sleep_us' field specifies how many microsconds to sleep for
787 * when there is no work to do, before doing another kloop iteration.
793 /* A CSB entry for the application --> kernel direction. */
795 uint32_t head; /* AW+ KR+ the head of the appl netmap_ring */
796 uint32_t cur; /* AW+ KR+ the cur of the appl netmap_ring */
797 uint32_t appl_need_kick; /* AW+ KR+ kern --> appl notification enable */
798 uint32_t sync_flags; /* AW+ KR+ the flags of the appl [tx|rx]sync() */
799 uint32_t pad[12]; /* pad to a 64 bytes cacheline */
802 /* A CSB entry for the application <-- kernel direction. */
804 uint32_t hwcur; /* AR+ KW+ the hwcur of the kern netmap_kring */
805 uint32_t hwtail; /* AR+ KW+ the hwtail of the kern netmap_kring */
806 uint32_t kern_need_kick; /* AR+ KW+ appl-->kern notification enable */
813 #define nm_stst_barrier smp_wmb
814 #define nm_ldld_barrier smp_rmb
815 #define nm_stld_barrier smp_mb
816 #else /* !__KERNEL__ */
817 static inline void nm_stst_barrier(void)
819 /* A memory barrier with release semantic has the combined
820 * effect of a store-store barrier and a load-store barrier,
821 * which is fine for us. */
822 __atomic_thread_fence(__ATOMIC_RELEASE);
824 static inline void nm_ldld_barrier(void)
826 /* A memory barrier with acquire semantic has the combined
827 * effect of a load-load barrier and a store-load barrier,
828 * which is fine for us. */
829 __atomic_thread_fence(__ATOMIC_ACQUIRE);
831 #endif /* !__KERNEL__ */
833 #elif defined(__FreeBSD__)
836 #define nm_stst_barrier atomic_thread_fence_rel
837 #define nm_ldld_barrier atomic_thread_fence_acq
838 #define nm_stld_barrier atomic_thread_fence_seq_cst
843 using std::memory_order_release;
844 using std::memory_order_acquire;
846 #else /* __cplusplus */
847 #include <stdatomic.h>
848 #endif /* __cplusplus */
850 static inline void nm_stst_barrier(void)
852 atomic_thread_fence(memory_order_release);
854 static inline void nm_ldld_barrier(void)
856 atomic_thread_fence(memory_order_acquire);
858 #endif /* !_KERNEL */
860 #else /* !__linux__ && !__FreeBSD__ */
861 #error "OS not supported"
862 #endif /* !__linux__ && !__FreeBSD__ */
864 /* Application side of sync-kloop: Write ring pointers (cur, head) to the CSB.
865 * This routine is coupled with sync_kloop_kernel_read(). */
867 nm_sync_kloop_appl_write(struct nm_csb_atok *atok, uint32_t cur,
870 /* Issue a first store-store barrier to make sure writes to the
871 * netmap ring do not overcome updates on atok->cur and atok->head. */
875 * We need to write cur and head to the CSB but we cannot do it atomically.
876 * There is no way we can prevent the host from reading the updated value
877 * of one of the two and the old value of the other. However, if we make
878 * sure that the host never reads a value of head more recent than the
879 * value of cur we are safe. We can allow the host to read a value of cur
880 * more recent than the value of head, since in the netmap ring cur can be
881 * ahead of head and cur cannot wrap around head because it must be behind
882 * tail. Inverting the order of writes below could instead result into the
883 * host to think head went ahead of cur, which would cause the sync
886 * The following memory barrier scheme is used to make this happen:
890 * STORE(cur) LOAD(head)
891 * wmb() <-----------> rmb()
892 * STORE(head) LOAD(cur)
900 /* Application side of sync-kloop: Read kring pointers (hwcur, hwtail) from
901 * the CSB. This routine is coupled with sync_kloop_kernel_write(). */
903 nm_sync_kloop_appl_read(struct nm_csb_ktoa *ktoa, uint32_t *hwtail,
907 * We place a memory barrier to make sure that the update of hwtail never
908 * overtakes the update of hwcur.
909 * (see explanation in sync_kloop_kernel_write).
911 *hwtail = ktoa->hwtail;
913 *hwcur = ktoa->hwcur;
915 /* Make sure that loads from ktoa->hwtail and ktoa->hwcur are not delayed
916 * after the loads from the netmap ring. */
921 * data for NETMAP_REQ_OPT_* options
924 struct nmreq_opt_sync_kloop_eventfds {
925 struct nmreq_option nro_opt; /* common header */
926 /* An array of N entries for bidirectional notifications between
927 * the kernel loop and the application. The number of entries and
928 * their order must agree with the CSB arrays passed in the
929 * NETMAP_REQ_OPT_CSB option. Each entry contains a file descriptor
930 * backed by an eventfd.
932 * If any of the 'ioeventfd' entries is < 0, the event loop uses
933 * the sleeping synchronization strategy (according to sleep_us),
934 * and keeps kern_need_kick always disabled.
935 * Each 'irqfd' can be < 0, and in that case the corresponding queue
939 /* Notifier for the application --> kernel loop direction. */
941 /* Notifier for the kernel loop --> application direction. */
946 struct nmreq_opt_sync_kloop_mode {
947 struct nmreq_option nro_opt; /* common header */
948 #define NM_OPT_SYNC_KLOOP_DIRECT_TX (1 << 0)
949 #define NM_OPT_SYNC_KLOOP_DIRECT_RX (1 << 1)
953 struct nmreq_opt_extmem {
954 struct nmreq_option nro_opt; /* common header */
955 uint64_t nro_usrptr; /* (in) ptr to usr memory */
956 struct nmreq_pools_info nro_info; /* (in/out) */
959 struct nmreq_opt_csb {
960 struct nmreq_option nro_opt;
962 /* Array of CSB entries for application --> kernel communication
966 /* Array of CSB entries for kernel --> application communication
971 /* option NETMAP_REQ_OPT_OFFSETS */
972 struct nmreq_opt_offsets {
973 struct nmreq_option nro_opt;
974 /* the user must declare the maximum offset value that she is
975 * going to put into the offset slot-fields. Any larger value
976 * found at runtime will be cropped. On output the (possibly
977 * higher) effective max value is returned.
979 uint64_t nro_max_offset;
980 /* optional initial offset value, to be set in all slots. */
981 uint64_t nro_initial_offset;
982 /* number of bits in the lower part of the 'ptr' field to be
983 * used as the offset field. On output the (possibly larger)
984 * effective number of bits is returned.
985 * 0 means: use the whole ptr field.
987 uint32_t nro_offset_bits;
988 /* required alignment for the beginning of the packets
989 * (base of the buffer plus offset) in the TX slots.
991 uint32_t nro_tx_align;
992 /* Reserved: set to zero. */
993 uint64_t nro_min_gap;
996 #endif /* _NET_NETMAP_H_ */