2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
4 * Copyright (c) 2011 NetApp, Inc.
7 * Redistribution and use in source and binary forms, with or without
8 * 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 NETAPP, INC ``AS 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 NETAPP, INC 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
31 #include <sys/cdefs.h>
32 __FBSDID("$FreeBSD$");
34 #include <sys/param.h>
35 #include <sys/linker_set.h>
36 #include <sys/select.h>
38 #include <sys/ioctl.h>
39 #include <net/ethernet.h>
40 #include <net/if.h> /* IFNAMSIZ */
53 #include <pthread_np.h>
60 #include "net_utils.h"
61 #include "net_backends.h"
64 #define VTNET_RINGSZ 1024
66 #define VTNET_MAXSEGS 256
68 #define VTNET_MAX_PKT_LEN (65536 + 64)
70 #define VTNET_MIN_MTU ETHERMIN
71 #define VTNET_MAX_MTU 65535
73 #define VTNET_S_HOSTCAPS \
74 ( VIRTIO_NET_F_MAC | VIRTIO_NET_F_STATUS | \
75 VIRTIO_F_NOTIFY_ON_EMPTY | VIRTIO_RING_F_INDIRECT_DESC)
78 * PCI config-space "registers"
80 struct virtio_net_config {
83 uint16_t max_virtqueue_pairs;
92 #define VTNET_CTLQ 2 /* NB: not yet supported */
99 static int pci_vtnet_debug;
100 #define DPRINTF(params) if (pci_vtnet_debug) PRINTLN params
101 #define WPRINTF(params) PRINTLN params
106 struct pci_vtnet_softc {
107 struct virtio_softc vsc_vs;
108 struct vqueue_info vsc_queues[VTNET_MAXQ - 1];
109 pthread_mutex_t vsc_mtx;
111 net_backend_t *vsc_be;
113 int resetting; /* protected by tx_mtx */
115 uint64_t vsc_features; /* negotiated features */
117 pthread_mutex_t rx_mtx;
118 int rx_merge; /* merged rx bufs in use */
121 pthread_mutex_t tx_mtx;
122 pthread_cond_t tx_cond;
128 struct virtio_net_config vsc_config;
129 struct virtio_consts vsc_consts;
132 static void pci_vtnet_reset(void *);
133 /* static void pci_vtnet_notify(void *, struct vqueue_info *); */
134 static int pci_vtnet_cfgread(void *, int, int, uint32_t *);
135 static int pci_vtnet_cfgwrite(void *, int, int, uint32_t);
136 static void pci_vtnet_neg_features(void *, uint64_t);
138 static struct virtio_consts vtnet_vi_consts = {
139 "vtnet", /* our name */
140 VTNET_MAXQ - 1, /* we currently support 2 virtqueues */
141 sizeof(struct virtio_net_config), /* config reg size */
142 pci_vtnet_reset, /* reset */
143 NULL, /* device-wide qnotify -- not used */
144 pci_vtnet_cfgread, /* read PCI config */
145 pci_vtnet_cfgwrite, /* write PCI config */
146 pci_vtnet_neg_features, /* apply negotiated features */
147 VTNET_S_HOSTCAPS, /* our capabilities */
151 pci_vtnet_reset(void *vsc)
153 struct pci_vtnet_softc *sc = vsc;
155 DPRINTF(("vtnet: device reset requested !"));
157 /* Acquire the RX lock to block RX processing. */
158 pthread_mutex_lock(&sc->rx_mtx);
161 * Make sure receive operation is disabled at least until we
162 * re-negotiate the features, since receive operation depends
163 * on the value of sc->rx_merge and the header length, which
164 * are both set in pci_vtnet_neg_features().
165 * Receive operation will be enabled again once the guest adds
166 * the first receive buffers and kicks us.
168 netbe_rx_disable(sc->vsc_be);
170 /* Set sc->resetting and give a chance to the TX thread to stop. */
171 pthread_mutex_lock(&sc->tx_mtx);
173 while (sc->tx_in_progress) {
174 pthread_mutex_unlock(&sc->tx_mtx);
176 pthread_mutex_lock(&sc->tx_mtx);
180 * Now reset rings, MSI-X vectors, and negotiated capabilities.
181 * Do that with the TX lock held, since we need to reset
184 vi_reset_dev(&sc->vsc_vs);
187 pthread_mutex_unlock(&sc->tx_mtx);
188 pthread_mutex_unlock(&sc->rx_mtx);
191 static __inline struct iovec *
192 iov_trim_hdr(struct iovec *iov, int *iovcnt, unsigned int hlen)
196 if (iov[0].iov_len < hlen) {
198 * Not enough header space in the first fragment.
199 * That's not ok for us.
204 iov[0].iov_len -= hlen;
205 if (iov[0].iov_len == 0) {
209 * Only space for the header. That's not
216 iov[0].iov_base = (void *)((uintptr_t)iov[0].iov_base + hlen);
223 struct virtio_mrg_rxbuf_info {
230 pci_vtnet_rx(struct pci_vtnet_softc *sc)
232 int prepend_hdr_len = sc->vhdrlen - sc->be_vhdrlen;
233 struct virtio_mrg_rxbuf_info info[VTNET_MAXSEGS];
234 struct iovec iov[VTNET_MAXSEGS + 1];
235 struct vqueue_info *vq;
237 vq = &sc->vsc_queues[VTNET_RXQ];
239 struct virtio_net_rxhdr *hdr;
248 plen = netbe_peek_recvlen(sc->vsc_be);
251 * No more packets (plen == 0), or backend errored
252 * (plen < 0). Interrupt if needed and stop.
254 vq_endchains(vq, /*used_all_avail=*/0);
257 plen += prepend_hdr_len;
260 * Get a descriptor chain to store the next ingress
261 * packet. In case of mergeable rx buffers, get as
262 * many chains as necessary in order to make room
270 int n = vq_getchain(vq, &info[n_chains].idx, riov,
271 VTNET_MAXSEGS - riov_len, NULL);
275 * No rx buffers. Enable RX kicks and double
279 if (!vq_has_descs(vq)) {
281 * Still no buffers. Return the unused
282 * chains (if any), interrupt if needed
283 * (including for NOTIFY_ON_EMPTY), and
284 * disable the backend until the next
287 vq_retchains(vq, n_chains);
288 vq_endchains(vq, /*used_all_avail=*/1);
289 netbe_rx_disable(sc->vsc_be);
293 /* More rx buffers found, so keep going. */
297 assert(n >= 1 && riov_len + n <= VTNET_MAXSEGS);
303 info[n_chains].len = (uint32_t)count_iov(riov, n);
304 riov_bytes += info[n_chains].len;
307 } while (riov_bytes < plen && riov_len < VTNET_MAXSEGS);
310 hdr = riov[0].iov_base;
311 if (prepend_hdr_len > 0) {
313 * The frontend uses a virtio-net header, but the
314 * backend does not. We need to prepend a zeroed
317 riov = iov_trim_hdr(riov, &riov_len, prepend_hdr_len);
320 * The first collected chain is nonsensical,
321 * as it is not even enough to store the
322 * virtio-net header. Just drop it.
324 vq_relchain(vq, info[0].idx, 0);
325 vq_retchains(vq, n_chains - 1);
328 memset(hdr, 0, prepend_hdr_len);
331 rlen = netbe_recv(sc->vsc_be, riov, riov_len);
332 if (rlen != plen - prepend_hdr_len) {
334 * If this happens it means there is something
335 * wrong with the backend (e.g., some other
336 * process is stealing our packets).
338 WPRINTF(("netbe_recv: expected %zd bytes, "
339 "got %zd", plen - prepend_hdr_len, rlen));
340 vq_retchains(vq, n_chains);
344 ulen = (uint32_t)plen;
347 * Publish the used buffers to the guest, reporting the
348 * number of bytes that we wrote.
351 vq_relchain(vq, info[0].idx, ulen);
361 vq_relchain_prepare(vq, info[i].idx, iolen);
367 vq_relchain_publish(vq);
368 assert(i == n_chains);
375 * Called when there is read activity on the backend file descriptor.
376 * Each buffer posted by the guest is assumed to be able to contain
377 * an entire ethernet frame + rx header.
380 pci_vtnet_rx_callback(int fd, enum ev_type type, void *param)
382 struct pci_vtnet_softc *sc = param;
384 pthread_mutex_lock(&sc->rx_mtx);
386 pthread_mutex_unlock(&sc->rx_mtx);
390 /* Called on RX kick. */
392 pci_vtnet_ping_rxq(void *vsc, struct vqueue_info *vq)
394 struct pci_vtnet_softc *sc = vsc;
397 * A qnotify means that the rx process can now begin.
399 pthread_mutex_lock(&sc->rx_mtx);
401 netbe_rx_enable(sc->vsc_be);
402 pthread_mutex_unlock(&sc->rx_mtx);
405 /* TX virtqueue processing, called by the TX thread. */
407 pci_vtnet_proctx(struct pci_vtnet_softc *sc, struct vqueue_info *vq)
409 struct iovec iov[VTNET_MAXSEGS + 1];
410 struct iovec *siov = iov;
416 * Obtain chain of descriptors. The first descriptor also
417 * contains the virtio-net header.
419 n = vq_getchain(vq, &idx, iov, VTNET_MAXSEGS, NULL);
420 assert(n >= 1 && n <= VTNET_MAXSEGS);
422 if (sc->vhdrlen != sc->be_vhdrlen) {
424 * The frontend uses a virtio-net header, but the backend
425 * does not. We simply strip the header and ignore it, as
426 * it should be zero-filled.
428 siov = iov_trim_hdr(siov, &n, sc->vhdrlen);
432 /* The chain is nonsensical. Just drop it. */
435 len = netbe_send(sc->vsc_be, siov, n);
438 * If send failed, report that 0 bytes
446 * Return the processed chain to the guest, reporting
447 * the number of bytes that we read.
449 vq_relchain(vq, idx, len);
452 /* Called on TX kick. */
454 pci_vtnet_ping_txq(void *vsc, struct vqueue_info *vq)
456 struct pci_vtnet_softc *sc = vsc;
459 * Any ring entries to process?
461 if (!vq_has_descs(vq))
464 /* Signal the tx thread for processing */
465 pthread_mutex_lock(&sc->tx_mtx);
467 if (sc->tx_in_progress == 0)
468 pthread_cond_signal(&sc->tx_cond);
469 pthread_mutex_unlock(&sc->tx_mtx);
473 * Thread which will handle processing of TX desc
476 pci_vtnet_tx_thread(void *param)
478 struct pci_vtnet_softc *sc = param;
479 struct vqueue_info *vq;
482 vq = &sc->vsc_queues[VTNET_TXQ];
485 * Let us wait till the tx queue pointers get initialised &
488 pthread_mutex_lock(&sc->tx_mtx);
489 error = pthread_cond_wait(&sc->tx_cond, &sc->tx_mtx);
493 /* note - tx mutex is locked here */
494 while (sc->resetting || !vq_has_descs(vq)) {
496 if (!sc->resetting && vq_has_descs(vq))
499 sc->tx_in_progress = 0;
500 error = pthread_cond_wait(&sc->tx_cond, &sc->tx_mtx);
504 sc->tx_in_progress = 1;
505 pthread_mutex_unlock(&sc->tx_mtx);
509 * Run through entries, placing them into
510 * iovecs and sending when an end-of-packet
513 pci_vtnet_proctx(sc, vq);
514 } while (vq_has_descs(vq));
517 * Generate an interrupt if needed.
519 vq_endchains(vq, /*used_all_avail=*/1);
521 pthread_mutex_lock(&sc->tx_mtx);
527 pci_vtnet_ping_ctlq(void *vsc, struct vqueue_info *vq)
530 DPRINTF(("vtnet: control qnotify!"));
535 pci_vtnet_init(struct vmctx *ctx, struct pci_devinst *pi, char *opts)
537 struct pci_vtnet_softc *sc;
538 char tname[MAXCOMLEN + 1];
541 unsigned long mtu = ETHERMTU;
544 * Allocate data structures for further virtio initializations.
545 * sc also contains a copy of vtnet_vi_consts, since capabilities
546 * change depending on the backend.
548 sc = calloc(1, sizeof(struct pci_vtnet_softc));
550 sc->vsc_consts = vtnet_vi_consts;
551 pthread_mutex_init(&sc->vsc_mtx, NULL);
553 sc->vsc_queues[VTNET_RXQ].vq_qsize = VTNET_RINGSZ;
554 sc->vsc_queues[VTNET_RXQ].vq_notify = pci_vtnet_ping_rxq;
555 sc->vsc_queues[VTNET_TXQ].vq_qsize = VTNET_RINGSZ;
556 sc->vsc_queues[VTNET_TXQ].vq_notify = pci_vtnet_ping_txq;
558 sc->vsc_queues[VTNET_CTLQ].vq_qsize = VTNET_RINGSZ;
559 sc->vsc_queues[VTNET_CTLQ].vq_notify = pci_vtnet_ping_ctlq;
563 * Attempt to open the backend device and read the MAC address
573 /* Get the device name. */
574 devname = vtopts = strdup(opts);
575 (void) strsep(&vtopts, ",");
578 * Parse the list of options in the form
579 * key1=value1,...,keyN=valueN.
581 while (vtopts != NULL) {
582 char *value = vtopts;
585 key = strsep(&value, "=");
589 (void) strsep(&vtopts, ",");
591 if (strcmp(key, "mac") == 0) {
592 err = net_parsemac(value, sc->vsc_config.mac);
596 } else if (strcmp(key, "mtu") == 0) {
597 err = net_parsemtu(value, &mtu);
601 if (mtu < VTNET_MIN_MTU || mtu > VTNET_MAX_MTU) {
616 err = netbe_init(&sc->vsc_be, devname, pci_vtnet_rx_callback,
623 sc->vsc_consts.vc_hv_caps |= VIRTIO_NET_F_MRG_RXBUF |
624 netbe_get_cap(sc->vsc_be);
628 net_genmac(pi, sc->vsc_config.mac);
631 sc->vsc_config.mtu = mtu;
633 sc->vsc_consts.vc_hv_caps |= VIRTIO_NET_F_MTU;
637 * Since we do not actually support multiqueue,
638 * set the maximum virtqueue pairs to 1.
640 sc->vsc_config.max_virtqueue_pairs = 1;
642 /* initialize config space */
643 pci_set_cfgdata16(pi, PCIR_DEVICE, VIRTIO_DEV_NET);
644 pci_set_cfgdata16(pi, PCIR_VENDOR, VIRTIO_VENDOR);
645 pci_set_cfgdata8(pi, PCIR_CLASS, PCIC_NETWORK);
646 pci_set_cfgdata16(pi, PCIR_SUBDEV_0, VIRTIO_TYPE_NET);
647 pci_set_cfgdata16(pi, PCIR_SUBVEND_0, VIRTIO_VENDOR);
649 /* Link is up if we managed to open backend device. */
650 sc->vsc_config.status = (opts == NULL || sc->vsc_be);
652 vi_softc_linkup(&sc->vsc_vs, &sc->vsc_consts, sc, pi, sc->vsc_queues);
653 sc->vsc_vs.vs_mtx = &sc->vsc_mtx;
655 /* use BAR 1 to map MSI-X table and PBA, if we're using MSI-X */
656 if (vi_intr_init(&sc->vsc_vs, 1, fbsdrun_virtio_msix())) {
661 /* use BAR 0 to map config regs in IO space */
662 vi_set_io_bar(&sc->vsc_vs, 0);
667 sc->vhdrlen = sizeof(struct virtio_net_rxhdr) - 2;
668 pthread_mutex_init(&sc->rx_mtx, NULL);
671 * Initialize tx semaphore & spawn TX processing thread.
672 * As of now, only one thread for TX desc processing is
675 sc->tx_in_progress = 0;
676 pthread_mutex_init(&sc->tx_mtx, NULL);
677 pthread_cond_init(&sc->tx_cond, NULL);
678 pthread_create(&sc->tx_tid, NULL, pci_vtnet_tx_thread, (void *)sc);
679 snprintf(tname, sizeof(tname), "vtnet-%d:%d tx", pi->pi_slot,
681 pthread_set_name_np(sc->tx_tid, tname);
687 pci_vtnet_cfgwrite(void *vsc, int offset, int size, uint32_t value)
689 struct pci_vtnet_softc *sc = vsc;
692 if (offset < (int)sizeof(sc->vsc_config.mac)) {
693 assert(offset + size <= (int)sizeof(sc->vsc_config.mac));
695 * The driver is allowed to change the MAC address
697 ptr = &sc->vsc_config.mac[offset];
698 memcpy(ptr, &value, size);
700 /* silently ignore other writes */
701 DPRINTF(("vtnet: write to readonly reg %d", offset));
708 pci_vtnet_cfgread(void *vsc, int offset, int size, uint32_t *retval)
710 struct pci_vtnet_softc *sc = vsc;
713 ptr = (uint8_t *)&sc->vsc_config + offset;
714 memcpy(retval, ptr, size);
719 pci_vtnet_neg_features(void *vsc, uint64_t negotiated_features)
721 struct pci_vtnet_softc *sc = vsc;
723 sc->vsc_features = negotiated_features;
725 if (negotiated_features & VIRTIO_NET_F_MRG_RXBUF) {
726 sc->vhdrlen = sizeof(struct virtio_net_rxhdr);
730 * Without mergeable rx buffers, virtio-net header is 2
731 * bytes shorter than sizeof(struct virtio_net_rxhdr).
733 sc->vhdrlen = sizeof(struct virtio_net_rxhdr) - 2;
737 /* Tell the backend to enable some capabilities it has advertised. */
738 netbe_set_cap(sc->vsc_be, negotiated_features, sc->vhdrlen);
739 sc->be_vhdrlen = netbe_get_vnet_hdr_len(sc->vsc_be);
740 assert(sc->be_vhdrlen == 0 || sc->be_vhdrlen == sc->vhdrlen);
743 static struct pci_devemu pci_de_vnet = {
744 .pe_emu = "virtio-net",
745 .pe_init = pci_vtnet_init,
746 .pe_barwrite = vi_pci_write,
747 .pe_barread = vi_pci_read
749 PCI_EMUL_SET(pci_de_vnet);