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 <machine/vmm_snapshot.h>
40 #include <net/ethernet.h>
41 #include <net/if.h> /* IFNAMSIZ */
54 #include <pthread_np.h>
61 #include "net_utils.h"
62 #include "net_backends.h"
65 #define VTNET_RINGSZ 1024
67 #define VTNET_MAXSEGS 256
69 #define VTNET_MAX_PKT_LEN (65536 + 64)
71 #define VTNET_MIN_MTU ETHERMIN
72 #define VTNET_MAX_MTU 65535
74 #define VTNET_S_HOSTCAPS \
75 ( VIRTIO_NET_F_MAC | VIRTIO_NET_F_STATUS | \
76 VIRTIO_F_NOTIFY_ON_EMPTY | VIRTIO_RING_F_INDIRECT_DESC)
79 * PCI config-space "registers"
81 struct virtio_net_config {
84 uint16_t max_virtqueue_pairs;
93 #define VTNET_CTLQ 2 /* NB: not yet supported */
100 static int pci_vtnet_debug;
101 #define DPRINTF(params) if (pci_vtnet_debug) PRINTLN params
102 #define WPRINTF(params) PRINTLN params
107 struct pci_vtnet_softc {
108 struct virtio_softc vsc_vs;
109 struct vqueue_info vsc_queues[VTNET_MAXQ - 1];
110 pthread_mutex_t vsc_mtx;
112 net_backend_t *vsc_be;
114 bool features_negotiated; /* protected by rx_mtx */
116 int resetting; /* protected by tx_mtx */
118 uint64_t vsc_features; /* negotiated features */
120 pthread_mutex_t rx_mtx;
121 int rx_merge; /* merged rx bufs in use */
124 pthread_mutex_t tx_mtx;
125 pthread_cond_t tx_cond;
131 struct virtio_net_config vsc_config;
132 struct virtio_consts vsc_consts;
135 static void pci_vtnet_reset(void *);
136 /* static void pci_vtnet_notify(void *, struct vqueue_info *); */
137 static int pci_vtnet_cfgread(void *, int, int, uint32_t *);
138 static int pci_vtnet_cfgwrite(void *, int, int, uint32_t);
139 static void pci_vtnet_neg_features(void *, uint64_t);
140 #ifdef BHYVE_SNAPSHOT
141 static void pci_vtnet_pause(void *);
142 static void pci_vtnet_resume(void *);
143 static int pci_vtnet_snapshot(void *, struct vm_snapshot_meta *);
146 static struct virtio_consts vtnet_vi_consts = {
147 "vtnet", /* our name */
148 VTNET_MAXQ - 1, /* we currently support 2 virtqueues */
149 sizeof(struct virtio_net_config), /* config reg size */
150 pci_vtnet_reset, /* reset */
151 NULL, /* device-wide qnotify -- not used */
152 pci_vtnet_cfgread, /* read PCI config */
153 pci_vtnet_cfgwrite, /* write PCI config */
154 pci_vtnet_neg_features, /* apply negotiated features */
155 VTNET_S_HOSTCAPS, /* our capabilities */
156 #ifdef BHYVE_SNAPSHOT
157 pci_vtnet_pause, /* pause rx/tx threads */
158 pci_vtnet_resume, /* resume rx/tx threads */
159 pci_vtnet_snapshot, /* save / restore device state */
164 pci_vtnet_reset(void *vsc)
166 struct pci_vtnet_softc *sc = vsc;
168 DPRINTF(("vtnet: device reset requested !"));
170 /* Acquire the RX lock to block RX processing. */
171 pthread_mutex_lock(&sc->rx_mtx);
174 * Make sure receive operation is disabled at least until we
175 * re-negotiate the features, since receive operation depends
176 * on the value of sc->rx_merge and the header length, which
177 * are both set in pci_vtnet_neg_features().
178 * Receive operation will be enabled again once the guest adds
179 * the first receive buffers and kicks us.
181 sc->features_negotiated = false;
182 netbe_rx_disable(sc->vsc_be);
184 /* Set sc->resetting and give a chance to the TX thread to stop. */
185 pthread_mutex_lock(&sc->tx_mtx);
187 while (sc->tx_in_progress) {
188 pthread_mutex_unlock(&sc->tx_mtx);
190 pthread_mutex_lock(&sc->tx_mtx);
194 * Now reset rings, MSI-X vectors, and negotiated capabilities.
195 * Do that with the TX lock held, since we need to reset
198 vi_reset_dev(&sc->vsc_vs);
201 pthread_mutex_unlock(&sc->tx_mtx);
202 pthread_mutex_unlock(&sc->rx_mtx);
205 static __inline struct iovec *
206 iov_trim_hdr(struct iovec *iov, int *iovcnt, unsigned int hlen)
210 if (iov[0].iov_len < hlen) {
212 * Not enough header space in the first fragment.
213 * That's not ok for us.
218 iov[0].iov_len -= hlen;
219 if (iov[0].iov_len == 0) {
223 * Only space for the header. That's not
230 iov[0].iov_base = (void *)((uintptr_t)iov[0].iov_base + hlen);
237 struct virtio_mrg_rxbuf_info {
244 pci_vtnet_rx(struct pci_vtnet_softc *sc)
246 int prepend_hdr_len = sc->vhdrlen - sc->be_vhdrlen;
247 struct virtio_mrg_rxbuf_info info[VTNET_MAXSEGS];
248 struct iovec iov[VTNET_MAXSEGS + 1];
249 struct vqueue_info *vq;
251 vq = &sc->vsc_queues[VTNET_RXQ];
253 /* Features must be negotiated */
254 if (!sc->features_negotiated) {
259 struct virtio_net_rxhdr *hdr;
268 plen = netbe_peek_recvlen(sc->vsc_be);
271 * No more packets (plen == 0), or backend errored
272 * (plen < 0). Interrupt if needed and stop.
274 vq_endchains(vq, /*used_all_avail=*/0);
277 plen += prepend_hdr_len;
280 * Get a descriptor chain to store the next ingress
281 * packet. In case of mergeable rx buffers, get as
282 * many chains as necessary in order to make room
290 int n = vq_getchain(vq, &info[n_chains].idx, riov,
291 VTNET_MAXSEGS - riov_len, NULL);
295 * No rx buffers. Enable RX kicks and double
299 if (!vq_has_descs(vq)) {
301 * Still no buffers. Return the unused
302 * chains (if any), interrupt if needed
303 * (including for NOTIFY_ON_EMPTY), and
304 * disable the backend until the next
307 vq_retchains(vq, n_chains);
308 vq_endchains(vq, /*used_all_avail=*/1);
309 netbe_rx_disable(sc->vsc_be);
313 /* More rx buffers found, so keep going. */
317 assert(n >= 1 && riov_len + n <= VTNET_MAXSEGS);
323 info[n_chains].len = (uint32_t)count_iov(riov, n);
324 riov_bytes += info[n_chains].len;
327 } while (riov_bytes < plen && riov_len < VTNET_MAXSEGS);
330 hdr = riov[0].iov_base;
331 if (prepend_hdr_len > 0) {
333 * The frontend uses a virtio-net header, but the
334 * backend does not. We need to prepend a zeroed
337 riov = iov_trim_hdr(riov, &riov_len, prepend_hdr_len);
340 * The first collected chain is nonsensical,
341 * as it is not even enough to store the
342 * virtio-net header. Just drop it.
344 vq_relchain(vq, info[0].idx, 0);
345 vq_retchains(vq, n_chains - 1);
348 memset(hdr, 0, prepend_hdr_len);
351 rlen = netbe_recv(sc->vsc_be, riov, riov_len);
352 if (rlen != plen - prepend_hdr_len) {
354 * If this happens it means there is something
355 * wrong with the backend (e.g., some other
356 * process is stealing our packets).
358 WPRINTF(("netbe_recv: expected %zd bytes, "
359 "got %zd", plen - prepend_hdr_len, rlen));
360 vq_retchains(vq, n_chains);
364 ulen = (uint32_t)plen;
367 * Publish the used buffers to the guest, reporting the
368 * number of bytes that we wrote.
371 vq_relchain(vq, info[0].idx, ulen);
381 vq_relchain_prepare(vq, info[i].idx, iolen);
387 vq_relchain_publish(vq);
388 assert(i == n_chains);
395 * Called when there is read activity on the backend file descriptor.
396 * Each buffer posted by the guest is assumed to be able to contain
397 * an entire ethernet frame + rx header.
400 pci_vtnet_rx_callback(int fd, enum ev_type type, void *param)
402 struct pci_vtnet_softc *sc = param;
404 pthread_mutex_lock(&sc->rx_mtx);
406 pthread_mutex_unlock(&sc->rx_mtx);
410 /* Called on RX kick. */
412 pci_vtnet_ping_rxq(void *vsc, struct vqueue_info *vq)
414 struct pci_vtnet_softc *sc = vsc;
417 * A qnotify means that the rx process can now begin.
418 * Enable RX only if features are negotiated.
420 pthread_mutex_lock(&sc->rx_mtx);
421 if (!sc->features_negotiated) {
422 pthread_mutex_unlock(&sc->rx_mtx);
427 netbe_rx_enable(sc->vsc_be);
428 pthread_mutex_unlock(&sc->rx_mtx);
431 /* TX virtqueue processing, called by the TX thread. */
433 pci_vtnet_proctx(struct pci_vtnet_softc *sc, struct vqueue_info *vq)
435 struct iovec iov[VTNET_MAXSEGS + 1];
436 struct iovec *siov = iov;
442 * Obtain chain of descriptors. The first descriptor also
443 * contains the virtio-net header.
445 n = vq_getchain(vq, &idx, iov, VTNET_MAXSEGS, NULL);
446 assert(n >= 1 && n <= VTNET_MAXSEGS);
448 if (sc->vhdrlen != sc->be_vhdrlen) {
450 * The frontend uses a virtio-net header, but the backend
451 * does not. We simply strip the header and ignore it, as
452 * it should be zero-filled.
454 siov = iov_trim_hdr(siov, &n, sc->vhdrlen);
458 /* The chain is nonsensical. Just drop it. */
461 len = netbe_send(sc->vsc_be, siov, n);
464 * If send failed, report that 0 bytes
472 * Return the processed chain to the guest, reporting
473 * the number of bytes that we read.
475 vq_relchain(vq, idx, len);
478 /* Called on TX kick. */
480 pci_vtnet_ping_txq(void *vsc, struct vqueue_info *vq)
482 struct pci_vtnet_softc *sc = vsc;
485 * Any ring entries to process?
487 if (!vq_has_descs(vq))
490 /* Signal the tx thread for processing */
491 pthread_mutex_lock(&sc->tx_mtx);
493 if (sc->tx_in_progress == 0)
494 pthread_cond_signal(&sc->tx_cond);
495 pthread_mutex_unlock(&sc->tx_mtx);
499 * Thread which will handle processing of TX desc
502 pci_vtnet_tx_thread(void *param)
504 struct pci_vtnet_softc *sc = param;
505 struct vqueue_info *vq;
508 vq = &sc->vsc_queues[VTNET_TXQ];
511 * Let us wait till the tx queue pointers get initialised &
514 pthread_mutex_lock(&sc->tx_mtx);
515 error = pthread_cond_wait(&sc->tx_cond, &sc->tx_mtx);
519 /* note - tx mutex is locked here */
520 while (sc->resetting || !vq_has_descs(vq)) {
522 if (!sc->resetting && vq_has_descs(vq))
525 sc->tx_in_progress = 0;
526 error = pthread_cond_wait(&sc->tx_cond, &sc->tx_mtx);
530 sc->tx_in_progress = 1;
531 pthread_mutex_unlock(&sc->tx_mtx);
535 * Run through entries, placing them into
536 * iovecs and sending when an end-of-packet
539 pci_vtnet_proctx(sc, vq);
540 } while (vq_has_descs(vq));
543 * Generate an interrupt if needed.
545 vq_endchains(vq, /*used_all_avail=*/1);
547 pthread_mutex_lock(&sc->tx_mtx);
553 pci_vtnet_ping_ctlq(void *vsc, struct vqueue_info *vq)
556 DPRINTF(("vtnet: control qnotify!"));
561 pci_vtnet_init(struct vmctx *ctx, struct pci_devinst *pi, char *opts)
563 struct pci_vtnet_softc *sc;
564 char tname[MAXCOMLEN + 1];
567 unsigned long mtu = ETHERMTU;
570 * Allocate data structures for further virtio initializations.
571 * sc also contains a copy of vtnet_vi_consts, since capabilities
572 * change depending on the backend.
574 sc = calloc(1, sizeof(struct pci_vtnet_softc));
576 sc->vsc_consts = vtnet_vi_consts;
577 pthread_mutex_init(&sc->vsc_mtx, NULL);
579 sc->vsc_queues[VTNET_RXQ].vq_qsize = VTNET_RINGSZ;
580 sc->vsc_queues[VTNET_RXQ].vq_notify = pci_vtnet_ping_rxq;
581 sc->vsc_queues[VTNET_TXQ].vq_qsize = VTNET_RINGSZ;
582 sc->vsc_queues[VTNET_TXQ].vq_notify = pci_vtnet_ping_txq;
584 sc->vsc_queues[VTNET_CTLQ].vq_qsize = VTNET_RINGSZ;
585 sc->vsc_queues[VTNET_CTLQ].vq_notify = pci_vtnet_ping_ctlq;
589 * Attempt to open the backend device and read the MAC address
599 /* Get the device name. */
600 optscopy = vtopts = strdup(opts);
601 (void) strsep(&vtopts, ",");
604 * Parse the list of options in the form
605 * key1=value1,...,keyN=valueN.
607 while (vtopts != NULL) {
608 char *value = vtopts;
611 key = strsep(&value, "=");
615 (void) strsep(&vtopts, ",");
617 if (strcmp(key, "mac") == 0) {
618 err = net_parsemac(value, sc->vsc_config.mac);
622 } else if (strcmp(key, "mtu") == 0) {
623 err = net_parsemtu(value, &mtu);
627 if (mtu < VTNET_MIN_MTU || mtu > VTNET_MAX_MTU) {
643 err = netbe_init(&sc->vsc_be, opts, pci_vtnet_rx_callback,
650 sc->vsc_consts.vc_hv_caps |= VIRTIO_NET_F_MRG_RXBUF |
651 netbe_get_cap(sc->vsc_be);
655 net_genmac(pi, sc->vsc_config.mac);
658 sc->vsc_config.mtu = mtu;
660 sc->vsc_consts.vc_hv_caps |= VIRTIO_NET_F_MTU;
664 * Since we do not actually support multiqueue,
665 * set the maximum virtqueue pairs to 1.
667 sc->vsc_config.max_virtqueue_pairs = 1;
669 /* initialize config space */
670 pci_set_cfgdata16(pi, PCIR_DEVICE, VIRTIO_DEV_NET);
671 pci_set_cfgdata16(pi, PCIR_VENDOR, VIRTIO_VENDOR);
672 pci_set_cfgdata8(pi, PCIR_CLASS, PCIC_NETWORK);
673 pci_set_cfgdata16(pi, PCIR_SUBDEV_0, VIRTIO_TYPE_NET);
674 pci_set_cfgdata16(pi, PCIR_SUBVEND_0, VIRTIO_VENDOR);
676 /* Link is up if we managed to open backend device. */
677 sc->vsc_config.status = (opts == NULL || sc->vsc_be);
679 vi_softc_linkup(&sc->vsc_vs, &sc->vsc_consts, sc, pi, sc->vsc_queues);
680 sc->vsc_vs.vs_mtx = &sc->vsc_mtx;
682 /* use BAR 1 to map MSI-X table and PBA, if we're using MSI-X */
683 if (vi_intr_init(&sc->vsc_vs, 1, fbsdrun_virtio_msix())) {
688 /* use BAR 0 to map config regs in IO space */
689 vi_set_io_bar(&sc->vsc_vs, 0);
694 sc->vhdrlen = sizeof(struct virtio_net_rxhdr) - 2;
695 pthread_mutex_init(&sc->rx_mtx, NULL);
698 * Initialize tx semaphore & spawn TX processing thread.
699 * As of now, only one thread for TX desc processing is
702 sc->tx_in_progress = 0;
703 pthread_mutex_init(&sc->tx_mtx, NULL);
704 pthread_cond_init(&sc->tx_cond, NULL);
705 pthread_create(&sc->tx_tid, NULL, pci_vtnet_tx_thread, (void *)sc);
706 snprintf(tname, sizeof(tname), "vtnet-%d:%d tx", pi->pi_slot,
708 pthread_set_name_np(sc->tx_tid, tname);
714 pci_vtnet_cfgwrite(void *vsc, int offset, int size, uint32_t value)
716 struct pci_vtnet_softc *sc = vsc;
719 if (offset < (int)sizeof(sc->vsc_config.mac)) {
720 assert(offset + size <= (int)sizeof(sc->vsc_config.mac));
722 * The driver is allowed to change the MAC address
724 ptr = &sc->vsc_config.mac[offset];
725 memcpy(ptr, &value, size);
727 /* silently ignore other writes */
728 DPRINTF(("vtnet: write to readonly reg %d", offset));
735 pci_vtnet_cfgread(void *vsc, int offset, int size, uint32_t *retval)
737 struct pci_vtnet_softc *sc = vsc;
740 ptr = (uint8_t *)&sc->vsc_config + offset;
741 memcpy(retval, ptr, size);
746 pci_vtnet_neg_features(void *vsc, uint64_t negotiated_features)
748 struct pci_vtnet_softc *sc = vsc;
750 sc->vsc_features = negotiated_features;
752 if (negotiated_features & VIRTIO_NET_F_MRG_RXBUF) {
753 sc->vhdrlen = sizeof(struct virtio_net_rxhdr);
757 * Without mergeable rx buffers, virtio-net header is 2
758 * bytes shorter than sizeof(struct virtio_net_rxhdr).
760 sc->vhdrlen = sizeof(struct virtio_net_rxhdr) - 2;
764 /* Tell the backend to enable some capabilities it has advertised. */
765 netbe_set_cap(sc->vsc_be, negotiated_features, sc->vhdrlen);
766 sc->be_vhdrlen = netbe_get_vnet_hdr_len(sc->vsc_be);
767 assert(sc->be_vhdrlen == 0 || sc->be_vhdrlen == sc->vhdrlen);
769 pthread_mutex_lock(&sc->rx_mtx);
770 sc->features_negotiated = true;
771 pthread_mutex_unlock(&sc->rx_mtx);
774 #ifdef BHYVE_SNAPSHOT
776 pci_vtnet_pause(void *vsc)
778 struct pci_vtnet_softc *sc = vsc;
780 DPRINTF(("vtnet: device pause requested !\n"));
782 /* Acquire the RX lock to block RX processing. */
783 pthread_mutex_lock(&sc->rx_mtx);
785 /* Wait for the transmit thread to finish its processing. */
786 pthread_mutex_lock(&sc->tx_mtx);
787 while (sc->tx_in_progress) {
788 pthread_mutex_unlock(&sc->tx_mtx);
790 pthread_mutex_lock(&sc->tx_mtx);
795 pci_vtnet_resume(void *vsc)
797 struct pci_vtnet_softc *sc = vsc;
799 DPRINTF(("vtnet: device resume requested !\n"));
801 pthread_mutex_unlock(&sc->tx_mtx);
802 /* The RX lock should have been acquired in vtnet_pause. */
803 pthread_mutex_unlock(&sc->rx_mtx);
807 pci_vtnet_snapshot(void *vsc, struct vm_snapshot_meta *meta)
810 struct pci_vtnet_softc *sc = vsc;
812 DPRINTF(("vtnet: device snapshot requested !\n"));
815 * Queues and consts should have been saved by the more generic
816 * vi_pci_snapshot function. We need to save only our features and
820 SNAPSHOT_VAR_OR_LEAVE(sc->vsc_features, meta, ret, done);
822 /* Force reapply negociated features at restore time */
823 if (meta->op == VM_SNAPSHOT_RESTORE) {
824 pci_vtnet_neg_features(sc, sc->vsc_features);
825 netbe_rx_enable(sc->vsc_be);
828 SNAPSHOT_VAR_OR_LEAVE(sc->vsc_config, meta, ret, done);
829 SNAPSHOT_VAR_OR_LEAVE(sc->rx_merge, meta, ret, done);
831 SNAPSHOT_VAR_OR_LEAVE(sc->vhdrlen, meta, ret, done);
832 SNAPSHOT_VAR_OR_LEAVE(sc->be_vhdrlen, meta, ret, done);
839 static struct pci_devemu pci_de_vnet = {
840 .pe_emu = "virtio-net",
841 .pe_init = pci_vtnet_init,
842 .pe_barwrite = vi_pci_write,
843 .pe_barread = vi_pci_read,
844 #ifdef BHYVE_SNAPSHOT
845 .pe_snapshot = vi_pci_snapshot,
846 .pe_pause = vi_pci_pause,
847 .pe_resume = vi_pci_resume,
850 PCI_EMUL_SET(pci_de_vnet);