1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright 2017 6WIND S.A.
3 * Copyright 2017 Mellanox Technologies, Ltd
9 #include <linux/sockios.h>
10 #include <linux/netlink.h>
11 #include <linux/rtnetlink.h>
13 #include <net/if_arp.h>
14 #include <netinet/ip.h>
21 #include <sys/ioctl.h>
22 #include <sys/queue.h>
23 #include <sys/socket.h>
26 #include <rte_alarm.h>
28 #include <rte_bus_vdev.h>
29 #include <rte_common.h>
31 #include <rte_errno.h>
32 #include <rte_ethdev.h>
33 #include <rte_ether.h>
34 #include <rte_hypervisor.h>
35 #include <rte_kvargs.h>
37 #include <rte_string_fns.h>
39 #define VDEV_NETVSC_DRIVER net_vdev_netvsc
40 #define VDEV_NETVSC_DRIVER_NAME RTE_STR(VDEV_NETVSC_DRIVER)
41 #define VDEV_NETVSC_DRIVER_NAME_LEN 15
42 #define VDEV_NETVSC_ARG_IFACE "iface"
43 #define VDEV_NETVSC_ARG_MAC "mac"
44 #define VDEV_NETVSC_ARG_FORCE "force"
45 #define VDEV_NETVSC_ARG_IGNORE "ignore"
46 #define VDEV_NETVSC_PROBE_MS 1000
48 #define NETVSC_CLASS_ID "{f8615163-df3e-46c5-913f-f2d2f965ed0e}"
49 #define NETVSC_MAX_ROUTE_LINE_SIZE 300
51 #define DRV_LOG(level, ...) \
52 rte_log(RTE_LOG_ ## level, \
53 vdev_netvsc_logtype, \
54 RTE_FMT(VDEV_NETVSC_DRIVER_NAME ": " \
55 RTE_FMT_HEAD(__VA_ARGS__,) "\n", \
56 RTE_FMT_TAIL(__VA_ARGS__,)))
58 /** Driver-specific log messages type. */
59 static int vdev_netvsc_logtype;
61 /** Context structure for a vdev_netvsc instance. */
62 struct vdev_netvsc_ctx {
63 LIST_ENTRY(vdev_netvsc_ctx) entry; /**< Next entry in list. */
64 unsigned int id; /**< Unique ID. */
65 char name[64]; /**< Unique name. */
66 char devname[64]; /**< Fail-safe instance name. */
67 char devargs[256]; /**< Fail-safe device arguments. */
68 char if_name[IF_NAMESIZE]; /**< NetVSC netdevice name. */
69 unsigned int if_index; /**< NetVSC netdevice index. */
70 struct rte_ether_addr if_addr; /**< NetVSC MAC address. */
71 int pipe[2]; /**< Fail-safe communication pipe. */
72 char yield[256]; /**< PCI sub-device arguments. */
75 /** Context list is common to all driver instances. */
76 static LIST_HEAD(, vdev_netvsc_ctx) vdev_netvsc_ctx_list =
77 LIST_HEAD_INITIALIZER(vdev_netvsc_ctx_list);
79 /** Number of entries in context list. */
80 static unsigned int vdev_netvsc_ctx_count;
82 /** Number of driver instances relying on context list. */
83 static unsigned int vdev_netvsc_ctx_inst;
86 * Destroy a vdev_netvsc context instance.
92 vdev_netvsc_ctx_destroy(struct vdev_netvsc_ctx *ctx)
94 if (ctx->pipe[0] != -1)
96 if (ctx->pipe[1] != -1)
102 * Determine if a network interface is NetVSC.
105 * Pointer to netdevice description structure (name and index).
108 * A nonzero value when interface is detected as NetVSC. In case of error,
109 * rte_errno is updated and 0 returned.
112 vdev_netvsc_iface_is_netvsc(const struct if_nameindex *iface)
114 static const char temp[] = "/sys/class/net/%s/device/class_id";
115 char path[sizeof(temp) + IF_NAMESIZE];
120 ret = snprintf(path, sizeof(path), temp, iface->if_name);
121 if (ret == -1 || (size_t)ret >= sizeof(path)) {
125 f = fopen(path, "r");
130 ret = fscanf(f, NETVSC_CLASS_ID "%n", &len);
133 ret = len == (int)strlen(NETVSC_CLASS_ID);
139 * Iterate over system network interfaces.
141 * This function runs a given callback function for each netdevice found on
145 * Callback function pointer. List traversal is aborted when this function
146 * returns a nonzero value.
148 * Indicates the device type to iterate - netvsc or non-netvsc.
150 * Variable parameter list passed as @p va_list to @p func.
153 * 0 when the entire list is traversed successfully, a negative error code
154 * in case or failure, or the nonzero value returned by @p func when list
155 * traversal is aborted.
158 vdev_netvsc_foreach_iface(int (*func)(const struct if_nameindex *iface,
159 const struct rte_ether_addr *eth_addr,
160 va_list ap), int is_netvsc, ...)
162 struct if_nameindex *iface = if_nameindex();
163 int s = socket(PF_INET, SOCK_DGRAM, IPPROTO_IP);
169 DRV_LOG(ERR, "cannot retrieve system network interfaces");
174 DRV_LOG(ERR, "cannot open socket: %s", rte_strerror(errno));
177 for (i = 0; iface[i].if_name; ++i) {
180 struct rte_ether_addr eth_addr;
183 is_netvsc_ret = vdev_netvsc_iface_is_netvsc(&iface[i]) ? 1 : 0;
184 if (is_netvsc ^ is_netvsc_ret)
186 strlcpy(req.ifr_name, iface[i].if_name, sizeof(req.ifr_name));
187 if (ioctl(s, SIOCGIFHWADDR, &req) == -1) {
188 DRV_LOG(WARNING, "cannot retrieve information about"
189 " interface \"%s\": %s",
190 req.ifr_name, rte_strerror(errno));
193 if (req.ifr_hwaddr.sa_family != ARPHRD_ETHER)
195 memcpy(eth_addr.addr_bytes, req.ifr_hwaddr.sa_data,
196 RTE_DIM(eth_addr.addr_bytes));
197 va_start(ap, is_netvsc);
198 ret = func(&iface[i], ð_addr, ap);
207 if_freenameindex(iface);
212 * Determine if a network interface has a route.
215 * Network device name.
217 * Address family: AF_INET for IPv4 or AF_INET6 for IPv6.
220 * 1 when interface has a route, negative errno value in case of error and
224 vdev_netvsc_has_route(const struct if_nameindex *iface,
225 const unsigned char family)
228 * The implementation can be simpler by getifaddrs() function usage but
229 * it works for IPv6 only starting from glibc 2.3.3.
236 struct nlmsghdr *retmsg = (struct nlmsghdr *)buf;
237 struct sockaddr_nl sa;
239 struct nlmsghdr nlhdr;
240 struct ifaddrmsg addrmsg;
243 if (!iface || (family != AF_INET && family != AF_INET6)) {
244 DRV_LOG(ERR, "%s", rte_strerror(EINVAL));
247 sock = socket(AF_NETLINK, SOCK_RAW, NETLINK_ROUTE);
249 DRV_LOG(ERR, "cannot open socket: %s", rte_strerror(errno));
252 memset(&sa, 0, sizeof(sa));
253 sa.nl_family = AF_NETLINK;
254 sa.nl_groups = RTMGRP_LINK | RTMGRP_IPV4_IFADDR;
255 res = bind(sock, (struct sockaddr *)&sa, sizeof(sa));
258 DRV_LOG(ERR, "cannot bind socket: %s", rte_strerror(errno));
261 memset(&msg, 0, sizeof(msg));
262 msg.nlhdr.nlmsg_len = NLMSG_LENGTH(sizeof(struct ifaddrmsg));
263 msg.nlhdr.nlmsg_flags = NLM_F_REQUEST | NLM_F_DUMP;
264 msg.nlhdr.nlmsg_type = RTM_GETADDR;
265 msg.nlhdr.nlmsg_pid = getpid();
266 msg.addrmsg.ifa_family = family;
267 msg.addrmsg.ifa_index = iface->if_index;
268 res = send(sock, &msg, msg.nlhdr.nlmsg_len, 0);
271 DRV_LOG(ERR, "cannot send socket message: %s",
272 rte_strerror(errno));
275 memset(buf, 0, sizeof(buf));
276 len = recv(sock, buf, sizeof(buf), 0);
279 DRV_LOG(ERR, "cannot receive socket message: %s",
280 rte_strerror(errno));
283 while (NLMSG_OK(retmsg, (unsigned int)len)) {
284 struct ifaddrmsg *retaddr =
285 (struct ifaddrmsg *)NLMSG_DATA(retmsg);
287 if (retaddr->ifa_family == family &&
288 retaddr->ifa_index == iface->if_index) {
289 struct rtattr *retrta = IFA_RTA(retaddr);
290 int attlen = IFA_PAYLOAD(retmsg);
292 while (RTA_OK(retrta, attlen)) {
293 if (retrta->rta_type == IFA_ADDRESS) {
295 DRV_LOG(DEBUG, "interface %s has IP",
299 retrta = RTA_NEXT(retrta, attlen);
302 retmsg = NLMSG_NEXT(retmsg, len);
310 * Retrieve network interface data from sysfs symbolic link.
313 * Output data buffer.
315 * Output buffer size.
319 * Symbolic link path relative to netdevice sysfs entry.
322 * 0 on success, a negative error code otherwise.
325 vdev_netvsc_sysfs_readlink(char *buf, size_t size, const char *if_name,
328 struct vdev_netvsc_ctx *ctx;
329 char in[RTE_MAX(sizeof(ctx->yield), 256u)];
332 ret = snprintf(in, sizeof(in), "/sys/class/net/%s/%s",
334 if (ret == -1 || (size_t)ret >= sizeof(in))
336 ret = readlink(in, buf, size);
339 if ((size_t)ret >= size - 1)
346 * Probe a network interface to associate with vdev_netvsc context.
348 * This function determines if the network device matches the properties of
349 * the NetVSC interface associated with the vdev_netvsc context and
350 * communicates its bus address to the fail-safe PMD instance if so.
352 * It is normally used with vdev_netvsc_foreach_iface().
355 * Pointer to netdevice description structure (name and index).
356 * @param[in] eth_addr
357 * MAC address associated with @p iface.
359 * Variable arguments list comprising:
361 * - struct vdev_netvsc_ctx *ctx:
362 * Context to associate network interface with.
365 * A nonzero value when interface matches, 0 otherwise or in case of
369 vdev_netvsc_device_probe(const struct if_nameindex *iface,
370 const struct rte_ether_addr *eth_addr,
373 struct vdev_netvsc_ctx *ctx = va_arg(ap, struct vdev_netvsc_ctx *);
374 char buf[RTE_MAX(sizeof(ctx->yield), 256u)];
379 /* Skip non-matching or unwanted NetVSC interfaces. */
380 if (ctx->if_index == iface->if_index) {
381 if (!strcmp(ctx->if_name, iface->if_name))
384 "NetVSC interface \"%s\" (index %u) renamed \"%s\"",
385 ctx->if_name, ctx->if_index, iface->if_name);
386 strlcpy(ctx->if_name, iface->if_name, sizeof(ctx->if_name));
389 if (!rte_is_same_ether_addr(eth_addr, &ctx->if_addr))
391 /* Look for associated PCI device. */
392 ret = vdev_netvsc_sysfs_readlink(buf, sizeof(buf), iface->if_name,
396 addr = strrchr(buf, '/');
397 addr = addr ? addr + 1 : buf;
398 if (strcmp(addr, "pci"))
400 ret = vdev_netvsc_sysfs_readlink(buf, sizeof(buf), iface->if_name,
404 addr = strrchr(buf, '/');
405 addr = addr ? addr + 1 : buf;
409 /* Send PCI device argument to fail-safe PMD instance. */
410 if (strcmp(addr, ctx->yield))
411 DRV_LOG(DEBUG, "associating PCI device \"%s\" with NetVSC"
412 " interface \"%s\" (index %u)", addr, ctx->if_name,
414 memmove(buf, addr, len + 1);
417 ret = write(ctx->pipe[1], addr, len + 1);
420 if (errno == EINTR || errno == EAGAIN)
422 DRV_LOG(WARNING, "cannot associate PCI device name \"%s\" with"
423 " interface \"%s\": %s", addr, ctx->if_name,
424 rte_strerror(errno));
427 if ((size_t)ret != len + 1) {
429 * Attempt to override previous partial write, no need to
430 * recover if that fails.
432 ret = write(ctx->pipe[1], "\n", 1);
437 memcpy(ctx->yield, addr, len + 1);
442 * Alarm callback that regularly probes system network interfaces.
444 * This callback runs at a frequency determined by VDEV_NETVSC_PROBE_MS as
445 * long as an vdev_netvsc context instance exists.
451 vdev_netvsc_alarm(__rte_unused void *arg)
453 struct vdev_netvsc_ctx *ctx;
456 LIST_FOREACH(ctx, &vdev_netvsc_ctx_list, entry) {
457 ret = vdev_netvsc_foreach_iface(vdev_netvsc_device_probe, 0,
462 if (!vdev_netvsc_ctx_count)
464 ret = rte_eal_alarm_set(VDEV_NETVSC_PROBE_MS * 1000,
465 vdev_netvsc_alarm, NULL);
467 DRV_LOG(ERR, "unable to reschedule alarm callback: %s",
473 * Probe a NetVSC interface to generate a vdev_netvsc context from.
475 * This function instantiates vdev_netvsc contexts either for all NetVSC
476 * devices found on the system or only a subset provided as device
479 * It is normally used with vdev_netvsc_foreach_iface().
482 * Pointer to netdevice description structure (name and index).
483 * @param[in] eth_addr
484 * MAC address associated with @p iface.
486 * Variable arguments list comprising:
488 * - const char *name:
489 * Name associated with current driver instance.
491 * - struct rte_kvargs *kvargs:
492 * Device arguments provided to current driver instance.
495 * Accept specified interface even if not detected as NetVSC.
497 * - unsigned int specified:
498 * Number of specific netdevices provided as device arguments.
500 * - unsigned int *matched:
501 * The number of specified netdevices matched by this function.
504 * A nonzero value when interface matches, 0 otherwise or in case of
508 vdev_netvsc_netvsc_probe(const struct if_nameindex *iface,
509 const struct rte_ether_addr *eth_addr,
512 const char *name = va_arg(ap, const char *);
513 struct rte_kvargs *kvargs = va_arg(ap, struct rte_kvargs *);
514 unsigned int specified = va_arg(ap, unsigned int);
515 unsigned int *matched = va_arg(ap, unsigned int *);
517 struct vdev_netvsc_ctx *ctx;
520 /* Probe all interfaces when none are specified. */
522 for (i = 0; i != kvargs->count; ++i) {
523 const struct rte_kvargs_pair *pair = &kvargs->pairs[i];
525 if (!strcmp(pair->key, VDEV_NETVSC_ARG_IFACE)) {
526 if (!strcmp(pair->value, iface->if_name))
528 } else if (!strcmp(pair->key, VDEV_NETVSC_ARG_MAC)) {
529 struct rte_ether_addr tmp;
531 if (rte_ether_unformat_addr(pair->value, &tmp) != 0) {
533 "invalid MAC address format"
538 if (rte_is_same_ether_addr(eth_addr, &tmp))
542 if (i == kvargs->count)
546 /* Weed out interfaces already handled. */
547 LIST_FOREACH(ctx, &vdev_netvsc_ctx_list, entry)
548 if (ctx->if_index == iface->if_index)
554 "interface \"%s\" (index %u) is already handled,"
556 iface->if_name, iface->if_index);
559 /* Routed NetVSC should not be probed. */
560 if (vdev_netvsc_has_route(iface, AF_INET) ||
561 vdev_netvsc_has_route(iface, AF_INET6)) {
564 DRV_LOG(WARNING, "probably using routed NetVSC interface \"%s\""
565 " (index %u)", iface->if_name, iface->if_index);
567 /* Create interface context. */
568 ctx = calloc(1, sizeof(*ctx));
571 DRV_LOG(ERR, "cannot allocate context for interface \"%s\": %s",
572 iface->if_name, rte_strerror(errno));
575 ctx->id = vdev_netvsc_ctx_count;
576 strlcpy(ctx->if_name, iface->if_name, sizeof(ctx->if_name));
577 ctx->if_index = iface->if_index;
578 ctx->if_addr = *eth_addr;
581 ctx->yield[0] = '\0';
582 if (pipe(ctx->pipe) == -1) {
585 "cannot allocate control pipe for interface \"%s\": %s",
586 ctx->if_name, rte_strerror(errno));
589 for (i = 0; i != RTE_DIM(ctx->pipe); ++i) {
590 int flf = fcntl(ctx->pipe[i], F_GETFL);
593 fcntl(ctx->pipe[i], F_SETFL, flf | O_NONBLOCK) != -1)
596 DRV_LOG(ERR, "cannot toggle non-blocking flag on control file"
597 " descriptor #%u (%d): %s", i, ctx->pipe[i],
598 rte_strerror(errno));
601 /* Generate virtual device name and arguments. */
603 ret = snprintf(ctx->name, sizeof(ctx->name), "%s_id%u",
605 if (ret == -1 || (size_t)ret >= sizeof(ctx->name))
607 ret = snprintf(ctx->devname, sizeof(ctx->devname), "net_failsafe_vsc%u",
609 if (ret == -1 || (size_t)ret >= sizeof(ctx->devname))
611 ret = snprintf(ctx->devargs, sizeof(ctx->devargs),
612 "fd(%d),dev(net_tap_vsc%u,remote=%s)",
613 ctx->pipe[0], ctx->id, ctx->if_name);
614 if (ret == -1 || (size_t)ret >= sizeof(ctx->devargs))
618 DRV_LOG(ERR, "generated virtual device name or argument list"
619 " too long for interface \"%s\"", ctx->if_name);
622 /* Request virtual device generation. */
623 DRV_LOG(DEBUG, "generating virtual device \"%s\" with arguments \"%s\"",
624 ctx->devname, ctx->devargs);
625 vdev_netvsc_foreach_iface(vdev_netvsc_device_probe, 0, ctx);
626 ret = rte_eal_hotplug_add("vdev", ctx->devname, ctx->devargs);
629 LIST_INSERT_HEAD(&vdev_netvsc_ctx_list, ctx, entry);
630 ++vdev_netvsc_ctx_count;
631 DRV_LOG(DEBUG, "added NetVSC interface \"%s\" to context list",
636 vdev_netvsc_ctx_destroy(ctx);
641 * Probe NetVSC interfaces.
643 * This function probes system netdevices according to the specified device
644 * arguments and starts a periodic alarm callback to notify the resulting
645 * fail-safe PMD instances of their sub-devices whereabouts.
648 * Virtual device context for driver instance.
651 * Always 0, even in case of errors.
654 vdev_netvsc_vdev_probe(struct rte_vdev_device *dev)
656 static const char *const vdev_netvsc_arg[] = {
657 VDEV_NETVSC_ARG_IFACE,
659 VDEV_NETVSC_ARG_FORCE,
660 VDEV_NETVSC_ARG_IGNORE,
663 const char *name = rte_vdev_device_name(dev);
664 const char *args = rte_vdev_device_args(dev);
665 struct rte_kvargs *kvargs = rte_kvargs_parse(args ? args : "",
667 unsigned int specified = 0;
668 unsigned int matched = 0;
674 DRV_LOG(DEBUG, "invoked as \"%s\", using arguments \"%s\"", name, args);
676 DRV_LOG(ERR, "cannot parse arguments list");
679 for (i = 0; i != kvargs->count; ++i) {
680 const struct rte_kvargs_pair *pair = &kvargs->pairs[i];
682 if (!strcmp(pair->key, VDEV_NETVSC_ARG_FORCE))
683 force = !!atoi(pair->value);
684 else if (!strcmp(pair->key, VDEV_NETVSC_ARG_IGNORE))
685 ignore = !!atoi(pair->value);
686 else if (!strcmp(pair->key, VDEV_NETVSC_ARG_IFACE) ||
687 !strcmp(pair->key, VDEV_NETVSC_ARG_MAC))
692 rte_kvargs_free(kvargs);
696 DRV_LOG(ERR, "More than one way used to specify the netvsc"
700 rte_eal_alarm_cancel(vdev_netvsc_alarm, NULL);
701 /* Gather interfaces. */
702 ret = vdev_netvsc_foreach_iface(vdev_netvsc_netvsc_probe, 1, name,
703 kvargs, specified, &matched);
706 if (specified && matched < specified) {
708 DRV_LOG(ERR, "Cannot find the specified netvsc device");
711 /* Try to force probing on non-netvsc specified device. */
712 if (vdev_netvsc_foreach_iface(vdev_netvsc_netvsc_probe, 0, name,
713 kvargs, specified, &matched) < 0)
715 if (matched < specified) {
716 DRV_LOG(ERR, "Cannot find the specified device");
719 DRV_LOG(WARNING, "non-netvsc device was probed as netvsc");
721 ret = rte_eal_alarm_set(VDEV_NETVSC_PROBE_MS * 1000,
722 vdev_netvsc_alarm, NULL);
724 DRV_LOG(ERR, "unable to schedule alarm callback: %s",
730 rte_kvargs_free(kvargs);
731 ++vdev_netvsc_ctx_inst;
736 * Remove driver instance.
738 * The alarm callback and underlying vdev_netvsc context instances are only
739 * destroyed after the last PMD instance is removed.
742 * Virtual device context for driver instance.
748 vdev_netvsc_vdev_remove(__rte_unused struct rte_vdev_device *dev)
750 if (--vdev_netvsc_ctx_inst)
752 rte_eal_alarm_cancel(vdev_netvsc_alarm, NULL);
753 while (!LIST_EMPTY(&vdev_netvsc_ctx_list)) {
754 struct vdev_netvsc_ctx *ctx = LIST_FIRST(&vdev_netvsc_ctx_list);
756 LIST_REMOVE(ctx, entry);
757 --vdev_netvsc_ctx_count;
758 vdev_netvsc_ctx_destroy(ctx);
763 /** Virtual device descriptor. */
764 static struct rte_vdev_driver vdev_netvsc_vdev = {
765 .probe = vdev_netvsc_vdev_probe,
766 .remove = vdev_netvsc_vdev_remove,
769 RTE_PMD_REGISTER_VDEV(VDEV_NETVSC_DRIVER, vdev_netvsc_vdev);
770 RTE_PMD_REGISTER_ALIAS(VDEV_NETVSC_DRIVER, eth_vdev_netvsc);
771 RTE_PMD_REGISTER_PARAM_STRING(net_vdev_netvsc,
772 VDEV_NETVSC_ARG_IFACE "=<string> "
773 VDEV_NETVSC_ARG_MAC "=<string> "
774 VDEV_NETVSC_ARG_FORCE "=<int> "
775 VDEV_NETVSC_ARG_IGNORE "=<int>");
777 /** Initialize driver log type. */
778 RTE_INIT(vdev_netvsc_init_log)
780 vdev_netvsc_logtype = rte_log_register("pmd.net.vdev_netvsc");
781 if (vdev_netvsc_logtype >= 0)
782 rte_log_set_level(vdev_netvsc_logtype, RTE_LOG_NOTICE);
785 /** Compare function for vdev find device operation. */
787 vdev_netvsc_cmp_rte_device(const struct rte_device *dev1,
788 __rte_unused const void *_dev2)
790 return strncmp(dev1->devargs->name, VDEV_NETVSC_DRIVER_NAME,
791 VDEV_NETVSC_DRIVER_NAME_LEN);
795 * A callback called by vdev bus scan function to ensure this driver probing
796 * automatically in Hyper-V VM system unless it already exists in the
800 vdev_netvsc_scan_callback(__rte_unused void *arg)
802 struct rte_device *dev;
803 struct rte_devargs *devargs;
804 struct rte_bus *vbus = rte_bus_find_by_name("vdev");
806 RTE_EAL_DEVARGS_FOREACH("vdev", devargs)
807 if (!strncmp(devargs->name, VDEV_NETVSC_DRIVER_NAME,
808 VDEV_NETVSC_DRIVER_NAME_LEN))
811 dev = vbus->find_device(NULL, vdev_netvsc_cmp_rte_device,
812 VDEV_NETVSC_DRIVER_NAME);
815 if (rte_devargs_add(RTE_DEVTYPE_VIRTUAL, VDEV_NETVSC_DRIVER_NAME))
816 DRV_LOG(ERR, "unable to add netvsc devargs.");
819 /** Initialize the custom scan. */
820 RTE_INIT(vdev_netvsc_custom_scan_add)
822 if (rte_hypervisor_get() == RTE_HYPERVISOR_HYPERV)
823 rte_vdev_add_custom_scan(vdev_netvsc_scan_callback, NULL);