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 13 /* current API version */
46 #define NETMAP_MIN_API 13 /* 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+1 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 one tx/rx ring
94 * pair attached to the host stack (this pair is 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 request space for extra rings (and buffers)
121 * allocated in the same memory space. The number of extra rings
122 * is in nr_arg1, and is advisory. This is a no-op on NICs where
123 * the size of the memory space is fixed.
125 * + NIOCREGIF can attach to PIPE rings sharing the same memory
126 * space with a parent device. The ifname indicates the parent device,
127 * which must already exist. Flags in nr_flags indicate if we want to
128 * bind the master or slave side, the index (from nr_ringid)
129 * is just a cookie and does not need to be sequential.
131 * + NIOCREGIF can also attach to 'monitor' rings that replicate
132 * the content of specific rings, also from the same memory space.
134 * Extra flags in nr_flags support the above functions.
135 * Application libraries may use the following naming scheme:
136 * netmap:foo all NIC ring pairs
137 * netmap:foo^ only host ring pair
138 * netmap:foo+ all NIC ring + host ring pairs
139 * netmap:foo-k the k-th NIC ring pair
140 * netmap:foo{k PIPE ring pair k, master side
141 * netmap:foo}k PIPE ring pair k, slave side
143 * Some notes about host rings:
145 * + The RX host ring is used to store those packets that the host network
146 * stack is trying to transmit through a NIC queue, but only if that queue
147 * is currently in netmap mode. Netmap will not intercept host stack mbufs
148 * designated to NIC queues that are not in netmap mode. As a consequence,
149 * registering a netmap port with netmap:foo^ is not enough to intercept
150 * mbufs in the RX host ring; the netmap port should be registered with
151 * netmap:foo*, or another registration should be done to open at least a
152 * NIC TX queue in netmap mode.
154 * + Netmap is not currently able to deal with intercepted trasmit mbufs which
155 * require offloadings like TSO, UFO, checksumming offloadings, etc. It is
156 * responsibility of the user to disable those offloadings (e.g. using
157 * ifconfig on FreeBSD or ethtool -K on Linux) for an interface that is being
158 * used in netmap mode. If the offloadings are not disabled, GSO and/or
159 * unchecksummed packets may be dropped immediately or end up in the host RX
160 * ring, and will be dropped as soon as the packet reaches another netmap
165 * struct netmap_slot is a buffer descriptor
168 uint32_t buf_idx; /* buffer index */
169 uint16_t len; /* length for this slot */
170 uint16_t flags; /* buf changed, etc. */
171 uint64_t ptr; /* pointer for indirect buffers */
175 * The following flags control how the slot is used
178 #define NS_BUF_CHANGED 0x0001 /* buf_idx changed */
180 * must be set whenever buf_idx is changed (as it might be
181 * necessary to recompute the physical address and mapping)
183 * It is also set by the kernel whenever the buf_idx is
184 * changed internally (e.g., by pipes). Applications may
185 * use this information to know when they can reuse the
186 * contents of previously prepared buffers.
189 #define NS_REPORT 0x0002 /* ask the hardware to report results */
191 * Request notification when slot is used by the hardware.
192 * Normally transmit completions are handled lazily and
193 * may be unreported. This flag lets us know when a slot
194 * has been sent (e.g. to terminate the sender).
197 #define NS_FORWARD 0x0004 /* pass packet 'forward' */
199 * (Only for physical ports, rx rings with NR_FORWARD set).
200 * Slot released to the kernel (i.e. before ring->head) with
201 * this flag set are passed to the peer ring (host/NIC),
202 * thus restoring the host-NIC connection for these slots.
203 * This supports efficient traffic monitoring or firewalling.
206 #define NS_NO_LEARN 0x0008 /* disable bridge learning */
208 * On a VALE switch, do not 'learn' the source port for
212 #define NS_INDIRECT 0x0010 /* userspace buffer */
214 * (VALE tx rings only) data is in a userspace buffer,
215 * whose address is in the 'ptr' field in the slot.
218 #define NS_MOREFRAG 0x0020 /* packet has more fragments */
220 * (VALE ports, ptnetmap ports and some NIC ports, e.g.
221 * ixgbe and i40e on Linux)
222 * Set on all but the last slot of a multi-segment packet.
223 * The 'len' field refers to the individual fragment.
226 #define NS_PORT_SHIFT 8
227 #define NS_PORT_MASK (0xff << NS_PORT_SHIFT)
229 * The high 8 bits of the flag, if not zero, indicate the
230 * destination port for the VALE switch, overriding
234 #define NS_RFRAGS(_slot) ( ((_slot)->flags >> 8) & 0xff)
236 * (VALE rx rings only) the high 8 bits
237 * are the number of fragments.
240 #define NETMAP_MAX_FRAGS 64 /* max number of fragments */
246 * Netmap representation of a TX or RX ring (also known as "queue").
247 * This is a queue implemented as a fixed-size circular array.
248 * At the software level the important fields are: head, cur, tail.
252 * head first slot available for transmission.
253 * cur wakeup point. select() and poll() will unblock
254 * when 'tail' moves past 'cur'
255 * tail (readonly) first slot reserved to the kernel
257 * [head .. tail-1] can be used for new packets to send;
258 * 'head' and 'cur' must be incremented as slots are filled
259 * with new packets to be sent;
260 * 'cur' can be moved further ahead if we need more space
261 * for new transmissions. XXX todo (2014-03-12)
265 * head first valid received packet
266 * cur wakeup point. select() and poll() will unblock
267 * when 'tail' moves past 'cur'
268 * tail (readonly) first slot reserved to the kernel
270 * [head .. tail-1] contain received packets;
271 * 'head' and 'cur' must be incremented as slots are consumed
272 * and can be returned to the kernel;
273 * 'cur' can be moved further ahead if we want to wait for
274 * new packets without returning the previous ones.
276 * DATA OWNERSHIP/LOCKING:
277 * The netmap_ring, and all slots and buffers in the range
278 * [head .. tail-1] are owned by the user program;
279 * the kernel only accesses them during a netmap system call
280 * and in the user thread context.
282 * Other slots and buffers are reserved for use by the kernel
286 * buf_ofs is meant to be used through macros.
287 * It contains the offset of the buffer region from this
290 const int64_t buf_ofs;
291 const uint32_t num_slots; /* number of slots in the ring. */
292 const uint32_t nr_buf_size;
293 const uint16_t ringid;
294 const uint16_t dir; /* 0: tx, 1: rx */
296 uint32_t head; /* (u) first user slot */
297 uint32_t cur; /* (u) wakeup point */
298 uint32_t tail; /* (k) first kernel slot */
302 struct timeval ts; /* (k) time of last *sync() */
304 /* opaque room for a mutex or similar object */
305 #if !defined(_WIN32) || defined(__CYGWIN__)
306 uint8_t __attribute__((__aligned__(NM_CACHE_ALIGN))) sem[128];
308 uint8_t __declspec(align(NM_CACHE_ALIGN)) sem[128];
311 /* the slots follow. This struct has variable size */
312 struct netmap_slot slot[0]; /* array of slots. */
319 #define NR_TIMESTAMP 0x0002 /* set timestamp on *sync() */
321 * updates the 'ts' field on each netmap syscall. This saves
322 * saves a separate gettimeofday(), and is not much worse than
323 * software timestamps generated in the interrupt handler.
326 #define NR_FORWARD 0x0004 /* enable NS_FORWARD for ring */
328 * Enables the NS_FORWARD slot flag for the ring.
332 * Helper functions for kernel and userspace
336 * Check if space is available in the ring. We use ring->head, which
337 * points to the next netmap slot to be published to netmap. It is
338 * possible that the applications moves ring->cur ahead of ring->tail
339 * (e.g., by setting ring->cur <== ring->tail), if it wants more slots
340 * than the ones currently available, and it wants to be notified when
341 * more arrive. See netmap(4) for more details and examples.
344 nm_ring_empty(struct netmap_ring *ring)
346 return (ring->head == ring->tail);
350 * Netmap representation of an interface and its queue(s).
351 * This is initialized by the kernel when binding a file
352 * descriptor to a port, and should be considered as readonly
353 * by user programs. The kernel never uses it.
355 * There is one netmap_if for each file descriptor on which we want
357 * select/poll operates on one or all pairs depending on the value of
358 * nmr_queueid passed on the ioctl.
361 char ni_name[IFNAMSIZ]; /* name of the interface. */
362 const uint32_t ni_version; /* API version, currently unused */
363 const uint32_t ni_flags; /* properties */
364 #define NI_PRIV_MEM 0x1 /* private memory region */
367 * The number of packet rings available in netmap mode.
368 * Physical NICs can have different numbers of tx and rx rings.
369 * Physical NICs also have a 'host' ring pair.
370 * Additionally, clients can request additional ring pairs to
371 * be used for internal communication.
373 const uint32_t ni_tx_rings; /* number of HW tx rings */
374 const uint32_t ni_rx_rings; /* number of HW rx rings */
376 uint32_t ni_bufs_head; /* head index for extra bufs */
377 uint32_t ni_spare1[5];
379 * The following array contains the offset of each netmap ring
380 * from this structure, in the following order:
381 * NIC tx rings (ni_tx_rings); host tx ring (1); extra tx rings;
382 * NIC rx rings (ni_rx_rings); host tx ring (1); extra rx rings.
384 * The area is filled up by the kernel on NIOCREGIF,
385 * and then only read by userspace code.
387 const ssize_t ring_ofs[0];
390 /* Legacy interface to interact with a netmap control device.
391 * Included for backward compatibility. The user should not include this
393 #include "netmap_legacy.h"
396 * New API to control netmap control devices. New applications should only use
397 * nmreq_xyz structs with the NIOCCTRL ioctl() command.
399 * NIOCCTRL takes a nmreq_header struct, which contains the required
400 * API version, the name of a netmap port, a command type, and pointers
401 * to request body and options.
404 * The name of the port (em0, valeXXX:YYY, eth0{pn1 etc.)
406 * nr_version (in/out)
407 * Must match NETMAP_API as used in the kernel, error otherwise.
408 * Always returns the desired value on output.
411 * One of the NETMAP_REQ_* command types below
414 * Pointer to a command-specific struct, described by one
415 * of the struct nmreq_xyz below.
418 * Command specific options, if any.
420 * A NETMAP_REQ_REGISTER command activates netmap mode on the netmap
421 * port (e.g. physical interface) specified by nmreq_header.nr_name.
422 * The request body (struct nmreq_register) has several arguments to
423 * specify how the port is to be registered.
425 * nr_tx_slots, nr_tx_slots, nr_tx_rings, nr_rx_rings (in/out)
426 * On input, non-zero values may be used to reconfigure the port
427 * according to the requested values, but this is not guaranteed.
428 * On output the actual values in use are reported.
431 * Indicate what set of rings must be bound to the netmap
432 * device (e.g. all NIC rings, host rings only, NIC and
433 * host rings, ...). Values are in NR_REG_*.
436 * If nr_mode == NR_REG_ONE_NIC (only a single couple of TX/RX
437 * rings), indicate which NIC TX and/or RX ring is to be bound
438 * (0..nr_*x_rings-1).
441 * Indicate special options for how to open the port.
443 * NR_NO_TX_POLL can be OR-ed to make select()/poll() push
444 * packets on tx rings only if POLLOUT is set.
445 * The default is to push any pending packet.
447 * NR_DO_RX_POLL can be OR-ed to make select()/poll() release
448 * packets on rx rings also when POLLIN is NOT set.
449 * The default is to touch the rx ring only with POLLIN.
450 * Note that this is the opposite of TX because it
451 * reflects the common usage.
453 * Other options are NR_MONITOR_TX, NR_MONITOR_RX, NR_ZCOPY_MON,
454 * NR_EXCLUSIVE, NR_RX_RINGS_ONLY, NR_TX_RINGS_ONLY and
455 * NR_ACCEPT_VNET_HDR.
458 * The identity of the memory region used.
459 * On input, 0 means the system decides autonomously,
460 * other values may try to select a specific region.
461 * On return the actual value is reported.
462 * Region '1' is the global allocator, normally shared
463 * by all interfaces. Other values are private regions.
464 * If two ports the same region zero-copy is possible.
466 * nr_extra_bufs (in/out)
467 * Number of extra buffers to be allocated.
469 * The other NETMAP_REQ_* commands are described below.
473 /* maximum size of a request, including all options */
474 #define NETMAP_REQ_MAXSIZE 4096
476 /* Header common to all request options. */
477 struct nmreq_option {
478 /* Pointer ot the next option. */
481 uint32_t nro_reqtype;
482 /* (out) status of the option:
483 * 0: recognized and processed
487 /* Option size, used only for options that can have variable size
488 * (e.g. because they contain arrays). For fixed-size options this
489 * field should be set to zero. */
493 /* Header common to all requests. Do not reorder these fields, as we need
494 * the second one (nr_reqtype) to know how much to copy from/to userspace. */
495 struct nmreq_header {
496 uint16_t nr_version; /* API version */
497 uint16_t nr_reqtype; /* nmreq type (NETMAP_REQ_*) */
498 uint32_t nr_reserved; /* must be zero */
499 #define NETMAP_REQ_IFNAMSIZ 64
500 char nr_name[NETMAP_REQ_IFNAMSIZ]; /* port name */
501 uint64_t nr_options; /* command-specific options */
502 uint64_t nr_body; /* ptr to nmreq_xyz struct */
506 /* Register a netmap port with the device. */
507 NETMAP_REQ_REGISTER = 1,
508 /* Get information from a netmap port. */
509 NETMAP_REQ_PORT_INFO_GET,
510 /* Attach a netmap port to a VALE switch. */
511 NETMAP_REQ_VALE_ATTACH,
512 /* Detach a netmap port from a VALE switch. */
513 NETMAP_REQ_VALE_DETACH,
514 /* List the ports attached to a VALE switch. */
515 NETMAP_REQ_VALE_LIST,
516 /* Set the port header length (was virtio-net header length). */
517 NETMAP_REQ_PORT_HDR_SET,
518 /* Get the port header length (was virtio-net header length). */
519 NETMAP_REQ_PORT_HDR_GET,
520 /* Create a new persistent VALE port. */
521 NETMAP_REQ_VALE_NEWIF,
522 /* Delete a persistent VALE port. */
523 NETMAP_REQ_VALE_DELIF,
524 /* Enable polling kernel thread(s) on an attached VALE port. */
525 NETMAP_REQ_VALE_POLLING_ENABLE,
526 /* Disable polling kernel thread(s) on an attached VALE port. */
527 NETMAP_REQ_VALE_POLLING_DISABLE,
528 /* Get info about the pools of a memory allocator. */
529 NETMAP_REQ_POOLS_INFO_GET,
530 /* Start an in-kernel loop that syncs the rings periodically or
531 * on notifications. The loop runs in the context of the ioctl
532 * syscall, and only stops on NETMAP_REQ_SYNC_KLOOP_STOP. */
533 NETMAP_REQ_SYNC_KLOOP_START,
534 /* Stops the thread executing the in-kernel loop. The thread
535 * returns from the ioctl syscall. */
536 NETMAP_REQ_SYNC_KLOOP_STOP,
537 /* Enable CSB mode on a registered netmap control device. */
538 NETMAP_REQ_CSB_ENABLE,
542 /* On NETMAP_REQ_REGISTER, ask netmap to use memory allocated
543 * from user-space allocated memory pools (e.g. hugepages).
545 NETMAP_REQ_OPT_EXTMEM = 1,
547 /* ON NETMAP_REQ_SYNC_KLOOP_START, ask netmap to use eventfd-based
548 * notifications to synchronize the kernel loop with the application.
550 NETMAP_REQ_OPT_SYNC_KLOOP_EVENTFDS,
552 /* On NETMAP_REQ_REGISTER, ask netmap to work in CSB mode, where
553 * head, cur and tail pointers are not exchanged through the
554 * struct netmap_ring header, but rather using an user-provided
555 * memory area (see struct nm_csb_atok and struct nm_csb_ktoa).
559 /* An extension to NETMAP_REQ_OPT_SYNC_KLOOP_EVENTFDS, which specifies
560 * if the TX and/or RX rings are synced in the context of the VM exit.
561 * This requires the 'ioeventfd' fields to be valid (cannot be < 0).
563 NETMAP_REQ_OPT_SYNC_KLOOP_MODE,
567 * nr_reqtype: NETMAP_REQ_REGISTER
568 * Bind (register) a netmap port to this control device.
570 struct nmreq_register {
571 uint64_t nr_offset; /* nifp offset in the shared region */
572 uint64_t nr_memsize; /* size of the shared region */
573 uint32_t nr_tx_slots; /* slots in tx rings */
574 uint32_t nr_rx_slots; /* slots in rx rings */
575 uint16_t nr_tx_rings; /* number of tx rings */
576 uint16_t nr_rx_rings; /* number of rx rings */
578 uint16_t nr_mem_id; /* id of the memory allocator */
579 uint16_t nr_ringid; /* ring(s) we care about */
580 uint32_t nr_mode; /* specify NR_REG_* modes */
581 uint32_t nr_extra_bufs; /* number of requested extra buffers */
583 uint64_t nr_flags; /* additional flags (see below) */
584 /* monitors use nr_ringid and nr_mode to select the rings to monitor */
585 #define NR_MONITOR_TX 0x100
586 #define NR_MONITOR_RX 0x200
587 #define NR_ZCOPY_MON 0x400
588 /* request exclusive access to the selected rings */
589 #define NR_EXCLUSIVE 0x800
591 #define NR_RX_RINGS_ONLY 0x2000
592 #define NR_TX_RINGS_ONLY 0x4000
593 /* Applications set this flag if they are able to deal with virtio-net headers,
594 * that is send/receive frames that start with a virtio-net header.
595 * If not set, NIOCREGIF will fail with netmap ports that require applications
596 * to use those headers. If the flag is set, the application can use the
597 * NETMAP_VNET_HDR_GET command to figure out the header length. */
598 #define NR_ACCEPT_VNET_HDR 0x8000
599 /* The following two have the same meaning of NETMAP_NO_TX_POLL and
600 * NETMAP_DO_RX_POLL. */
601 #define NR_DO_RX_POLL 0x10000
602 #define NR_NO_TX_POLL 0x20000
605 /* Valid values for nmreq_register.nr_mode (see above). */
606 enum { NR_REG_DEFAULT = 0, /* backward compat, should not be used. */
611 NR_REG_PIPE_MASTER = 5, /* deprecated, use "x{y" port name syntax */
612 NR_REG_PIPE_SLAVE = 6, /* deprecated, use "x}y" port name syntax */
616 /* A single ioctl number is shared by all the new API command.
617 * Demultiplexing is done using the hdr.nr_reqtype field.
618 * FreeBSD uses the size value embedded in the _IOWR to determine
619 * how much to copy in/out, so we define the ioctl() command
620 * specifying only nmreq_header, and copyin/copyout the rest. */
621 #define NIOCCTRL _IOWR('i', 151, struct nmreq_header)
623 /* The ioctl commands to sync TX/RX netmap rings.
624 * NIOCTXSYNC, NIOCRXSYNC synchronize tx or rx queues,
625 * whose identity is set in NIOCREGIF through nr_ringid.
626 * These are non blocking and take no argument. */
627 #define NIOCTXSYNC _IO('i', 148) /* sync tx queues */
628 #define NIOCRXSYNC _IO('i', 149) /* sync rx queues */
631 * nr_reqtype: NETMAP_REQ_PORT_INFO_GET
632 * Get information about a netmap port, including number of rings.
633 * slots per ring, id of the memory allocator, etc. The netmap
634 * control device used for this operation does not need to be bound
637 struct nmreq_port_info_get {
638 uint64_t nr_memsize; /* size of the shared region */
639 uint32_t nr_tx_slots; /* slots in tx rings */
640 uint32_t nr_rx_slots; /* slots in rx rings */
641 uint16_t nr_tx_rings; /* number of tx rings */
642 uint16_t nr_rx_rings; /* number of rx rings */
643 uint16_t nr_mem_id; /* memory allocator id (in/out) */
647 #define NM_BDG_NAME "vale" /* prefix for bridge port name */
650 * nr_reqtype: NETMAP_REQ_VALE_ATTACH
651 * Attach a netmap port to a VALE switch. Both the name of the netmap
652 * port and the VALE switch are specified through the nr_name argument.
653 * The attach operation could need to register a port, so at least
654 * the same arguments are available.
655 * port_index will contain the index where the port has been attached.
657 struct nmreq_vale_attach {
658 struct nmreq_register reg;
664 * nr_reqtype: NETMAP_REQ_VALE_DETACH
665 * Detach a netmap port from a VALE switch. Both the name of the netmap
666 * port and the VALE switch are specified through the nr_name argument.
667 * port_index will contain the index where the port was attached.
669 struct nmreq_vale_detach {
675 * nr_reqtype: NETMAP_REQ_VALE_LIST
676 * List the ports of a VALE switch.
678 struct nmreq_vale_list {
679 /* Name of the VALE port (valeXXX:YYY) or empty. */
680 uint16_t nr_bridge_idx;
682 uint32_t nr_port_idx;
686 * nr_reqtype: NETMAP_REQ_PORT_HDR_SET or NETMAP_REQ_PORT_HDR_GET
687 * Set or get the port header length of the port identified by hdr.nr_name.
688 * The control device does not need to be bound to a netmap port.
690 struct nmreq_port_hdr {
696 * nr_reqtype: NETMAP_REQ_VALE_NEWIF
697 * Create a new persistent VALE port.
699 struct nmreq_vale_newif {
700 uint32_t nr_tx_slots; /* slots in tx rings */
701 uint32_t nr_rx_slots; /* slots in rx rings */
702 uint16_t nr_tx_rings; /* number of tx rings */
703 uint16_t nr_rx_rings; /* number of rx rings */
704 uint16_t nr_mem_id; /* id of the memory allocator */
709 * nr_reqtype: NETMAP_REQ_VALE_POLLING_ENABLE or NETMAP_REQ_VALE_POLLING_DISABLE
710 * Enable or disable polling kthreads on a VALE port.
712 struct nmreq_vale_polling {
714 #define NETMAP_POLLING_MODE_SINGLE_CPU 1
715 #define NETMAP_POLLING_MODE_MULTI_CPU 2
716 uint32_t nr_first_cpu_id;
717 uint32_t nr_num_polling_cpus;
722 * nr_reqtype: NETMAP_REQ_POOLS_INFO_GET
723 * Get info about the pools of the memory allocator of the netmap
724 * port specified by hdr.nr_name and nr_mem_id. The netmap control
725 * device used for this operation does not need to be bound to a netmap
728 struct nmreq_pools_info {
730 uint16_t nr_mem_id; /* in/out argument */
732 uint64_t nr_if_pool_offset;
733 uint32_t nr_if_pool_objtotal;
734 uint32_t nr_if_pool_objsize;
735 uint64_t nr_ring_pool_offset;
736 uint32_t nr_ring_pool_objtotal;
737 uint32_t nr_ring_pool_objsize;
738 uint64_t nr_buf_pool_offset;
739 uint32_t nr_buf_pool_objtotal;
740 uint32_t nr_buf_pool_objsize;
744 * nr_reqtype: NETMAP_REQ_SYNC_KLOOP_START
745 * Start an in-kernel loop that syncs the rings periodically or on
746 * notifications. The loop runs in the context of the ioctl syscall,
747 * and only stops on NETMAP_REQ_SYNC_KLOOP_STOP.
748 * The registered netmap port must be open in CSB mode.
750 struct nmreq_sync_kloop_start {
751 /* Sleeping is the default synchronization method for the kloop.
752 * The 'sleep_us' field specifies how many microsconds to sleep for
753 * when there is no work to do, before doing another kloop iteration.
759 /* A CSB entry for the application --> kernel direction. */
761 uint32_t head; /* AW+ KR+ the head of the appl netmap_ring */
762 uint32_t cur; /* AW+ KR+ the cur of the appl netmap_ring */
763 uint32_t appl_need_kick; /* AW+ KR+ kern --> appl notification enable */
764 uint32_t sync_flags; /* AW+ KR+ the flags of the appl [tx|rx]sync() */
765 uint32_t pad[12]; /* pad to a 64 bytes cacheline */
768 /* A CSB entry for the application <-- kernel direction. */
770 uint32_t hwcur; /* AR+ KW+ the hwcur of the kern netmap_kring */
771 uint32_t hwtail; /* AR+ KW+ the hwtail of the kern netmap_kring */
772 uint32_t kern_need_kick; /* AR+ KW+ appl-->kern notification enable */
779 #define nm_stst_barrier smp_wmb
780 #define nm_ldld_barrier smp_rmb
781 #define nm_stld_barrier smp_mb
782 #else /* !__KERNEL__ */
783 static inline void nm_stst_barrier(void)
785 /* A memory barrier with release semantic has the combined
786 * effect of a store-store barrier and a load-store barrier,
787 * which is fine for us. */
788 __atomic_thread_fence(__ATOMIC_RELEASE);
790 static inline void nm_ldld_barrier(void)
792 /* A memory barrier with acquire semantic has the combined
793 * effect of a load-load barrier and a store-load barrier,
794 * which is fine for us. */
795 __atomic_thread_fence(__ATOMIC_ACQUIRE);
797 #endif /* !__KERNEL__ */
799 #elif defined(__FreeBSD__)
802 #define nm_stst_barrier atomic_thread_fence_rel
803 #define nm_ldld_barrier atomic_thread_fence_acq
804 #define nm_stld_barrier atomic_thread_fence_seq_cst
806 #include <stdatomic.h>
807 static inline void nm_stst_barrier(void)
809 atomic_thread_fence(memory_order_release);
811 static inline void nm_ldld_barrier(void)
813 atomic_thread_fence(memory_order_acquire);
815 #endif /* !_KERNEL */
817 #else /* !__linux__ && !__FreeBSD__ */
818 #error "OS not supported"
819 #endif /* !__linux__ && !__FreeBSD__ */
821 /* Application side of sync-kloop: Write ring pointers (cur, head) to the CSB.
822 * This routine is coupled with sync_kloop_kernel_read(). */
824 nm_sync_kloop_appl_write(struct nm_csb_atok *atok, uint32_t cur,
827 /* Issue a first store-store barrier to make sure writes to the
828 * netmap ring do not overcome updates on atok->cur and atok->head. */
832 * We need to write cur and head to the CSB but we cannot do it atomically.
833 * There is no way we can prevent the host from reading the updated value
834 * of one of the two and the old value of the other. However, if we make
835 * sure that the host never reads a value of head more recent than the
836 * value of cur we are safe. We can allow the host to read a value of cur
837 * more recent than the value of head, since in the netmap ring cur can be
838 * ahead of head and cur cannot wrap around head because it must be behind
839 * tail. Inverting the order of writes below could instead result into the
840 * host to think head went ahead of cur, which would cause the sync
843 * The following memory barrier scheme is used to make this happen:
847 * STORE(cur) LOAD(head)
848 * wmb() <-----------> rmb()
849 * STORE(head) LOAD(cur)
857 /* Application side of sync-kloop: Read kring pointers (hwcur, hwtail) from
858 * the CSB. This routine is coupled with sync_kloop_kernel_write(). */
860 nm_sync_kloop_appl_read(struct nm_csb_ktoa *ktoa, uint32_t *hwtail,
864 * We place a memory barrier to make sure that the update of hwtail never
865 * overtakes the update of hwcur.
866 * (see explanation in sync_kloop_kernel_write).
868 *hwtail = ktoa->hwtail;
870 *hwcur = ktoa->hwcur;
872 /* Make sure that loads from ktoa->hwtail and ktoa->hwcur are not delayed
873 * after the loads from the netmap ring. */
878 * data for NETMAP_REQ_OPT_* options
881 struct nmreq_opt_sync_kloop_eventfds {
882 struct nmreq_option nro_opt; /* common header */
883 /* An array of N entries for bidirectional notifications between
884 * the kernel loop and the application. The number of entries and
885 * their order must agree with the CSB arrays passed in the
886 * NETMAP_REQ_OPT_CSB option. Each entry contains a file descriptor
887 * backed by an eventfd.
889 * If any of the 'ioeventfd' entries is < 0, the event loop uses
890 * the sleeping synchronization strategy (according to sleep_us),
891 * and keeps kern_need_kick always disabled.
892 * Each 'irqfd' can be < 0, and in that case the corresponding queue
896 /* Notifier for the application --> kernel loop direction. */
898 /* Notifier for the kernel loop --> application direction. */
903 struct nmreq_opt_sync_kloop_mode {
904 struct nmreq_option nro_opt; /* common header */
905 #define NM_OPT_SYNC_KLOOP_DIRECT_TX (1 << 0)
906 #define NM_OPT_SYNC_KLOOP_DIRECT_RX (1 << 1)
910 struct nmreq_opt_extmem {
911 struct nmreq_option nro_opt; /* common header */
912 uint64_t nro_usrptr; /* (in) ptr to usr memory */
913 struct nmreq_pools_info nro_info; /* (in/out) */
916 struct nmreq_opt_csb {
917 struct nmreq_option nro_opt;
919 /* Array of CSB entries for application --> kernel communication
923 /* Array of CSB entries for kernel --> application communication
928 #endif /* _NET_NETMAP_H_ */