1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright(c) 2010-2018 Intel Corporation
11 #include <rte_errno.h>
13 #include <rte_memory.h>
14 #include <rte_eal_memconfig.h>
17 #include "eal_filesystem.h"
19 #include "eal_private.h"
23 #define VFIO_MEM_EVENT_CLB_NAME "vfio_mem_event_clb"
25 /* hot plug/unplug of VFIO groups may cause all DMA maps to be dropped. we can
26 * recreate the mappings for DPDK segments, but we cannot do so for memory that
27 * was registered by the user themselves, so we need to store the user mappings
28 * somewhere, to recreate them later.
30 #define VFIO_MAX_USER_MEM_MAPS 256
37 struct user_mem_maps {
38 rte_spinlock_recursive_t lock;
40 struct user_mem_map maps[VFIO_MAX_USER_MEM_MAPS];
45 int vfio_container_fd;
46 int vfio_active_groups;
47 const struct vfio_iommu_type *vfio_iommu_type;
48 struct vfio_group vfio_groups[VFIO_MAX_GROUPS];
49 struct user_mem_maps mem_maps;
52 /* per-process VFIO config */
53 static struct vfio_config vfio_cfgs[VFIO_MAX_CONTAINERS];
54 static struct vfio_config *default_vfio_cfg = &vfio_cfgs[0];
56 static int vfio_type1_dma_map(int);
57 static int vfio_type1_dma_mem_map(int, uint64_t, uint64_t, uint64_t, int);
58 static int vfio_spapr_dma_map(int);
59 static int vfio_spapr_dma_mem_map(int, uint64_t, uint64_t, uint64_t, int);
60 static int vfio_noiommu_dma_map(int);
61 static int vfio_noiommu_dma_mem_map(int, uint64_t, uint64_t, uint64_t, int);
62 static int vfio_dma_mem_map(struct vfio_config *vfio_cfg, uint64_t vaddr,
63 uint64_t iova, uint64_t len, int do_map);
65 /* IOMMU types we support */
66 static const struct vfio_iommu_type iommu_types[] = {
67 /* x86 IOMMU, otherwise known as type 1 */
69 .type_id = RTE_VFIO_TYPE1,
71 .dma_map_func = &vfio_type1_dma_map,
72 .dma_user_map_func = &vfio_type1_dma_mem_map
74 /* ppc64 IOMMU, otherwise known as spapr */
76 .type_id = RTE_VFIO_SPAPR,
78 .dma_map_func = &vfio_spapr_dma_map,
79 .dma_user_map_func = &vfio_spapr_dma_mem_map
83 .type_id = RTE_VFIO_NOIOMMU,
85 .dma_map_func = &vfio_noiommu_dma_map,
86 .dma_user_map_func = &vfio_noiommu_dma_mem_map
91 is_null_map(const struct user_mem_map *map)
93 return map->addr == 0 && map->iova == 0 && map->len == 0;
96 /* we may need to merge user mem maps together in case of user mapping/unmapping
97 * chunks of memory, so we'll need a comparator function to sort segments.
100 user_mem_map_cmp(const void *a, const void *b)
102 const struct user_mem_map *umm_a = a;
103 const struct user_mem_map *umm_b = b;
105 /* move null entries to end */
106 if (is_null_map(umm_a))
108 if (is_null_map(umm_b))
111 /* sort by iova first */
112 if (umm_a->iova < umm_b->iova)
114 if (umm_a->iova > umm_b->iova)
117 if (umm_a->addr < umm_b->addr)
119 if (umm_a->addr > umm_b->addr)
122 if (umm_a->len < umm_b->len)
124 if (umm_a->len > umm_b->len)
130 /* adjust user map entry. this may result in shortening of existing map, or in
131 * splitting existing map in two pieces.
134 adjust_map(struct user_mem_map *src, struct user_mem_map *end,
135 uint64_t remove_va_start, uint64_t remove_len)
137 /* if va start is same as start address, we're simply moving start */
138 if (remove_va_start == src->addr) {
139 src->addr += remove_len;
140 src->iova += remove_len;
141 src->len -= remove_len;
142 } else if (remove_va_start + remove_len == src->addr + src->len) {
143 /* we're shrinking mapping from the end */
144 src->len -= remove_len;
146 /* we're blowing a hole in the middle */
147 struct user_mem_map tmp;
148 uint64_t total_len = src->len;
150 /* adjust source segment length */
151 src->len = remove_va_start - src->addr;
153 /* create temporary segment in the middle */
154 tmp.addr = src->addr + src->len;
155 tmp.iova = src->iova + src->len;
156 tmp.len = remove_len;
158 /* populate end segment - this one we will be keeping */
159 end->addr = tmp.addr + tmp.len;
160 end->iova = tmp.iova + tmp.len;
161 end->len = total_len - src->len - tmp.len;
165 /* try merging two maps into one, return 1 if succeeded */
167 merge_map(struct user_mem_map *left, struct user_mem_map *right)
169 if (left->addr + left->len != right->addr)
171 if (left->iova + left->len != right->iova)
174 left->len += right->len;
176 memset(right, 0, sizeof(*right));
181 static struct user_mem_map *
182 find_user_mem_map(struct user_mem_maps *user_mem_maps, uint64_t addr,
183 uint64_t iova, uint64_t len)
185 uint64_t va_end = addr + len;
186 uint64_t iova_end = iova + len;
189 for (i = 0; i < user_mem_maps->n_maps; i++) {
190 struct user_mem_map *map = &user_mem_maps->maps[i];
191 uint64_t map_va_end = map->addr + map->len;
192 uint64_t map_iova_end = map->iova + map->len;
195 if (addr < map->addr || addr >= map_va_end)
197 /* check if VA end is within boundaries */
198 if (va_end <= map->addr || va_end > map_va_end)
201 /* check start IOVA */
202 if (iova < map->iova || iova >= map_iova_end)
204 /* check if IOVA end is within boundaries */
205 if (iova_end <= map->iova || iova_end > map_iova_end)
208 /* we've found our map */
214 /* this will sort all user maps, and merge/compact any adjacent maps */
216 compact_user_maps(struct user_mem_maps *user_mem_maps)
218 int i, n_merged, cur_idx;
220 qsort(user_mem_maps->maps, user_mem_maps->n_maps,
221 sizeof(user_mem_maps->maps[0]), user_mem_map_cmp);
223 /* we'll go over the list backwards when merging */
225 for (i = user_mem_maps->n_maps - 2; i >= 0; i--) {
226 struct user_mem_map *l, *r;
228 l = &user_mem_maps->maps[i];
229 r = &user_mem_maps->maps[i + 1];
231 if (is_null_map(l) || is_null_map(r))
238 /* the entries are still sorted, but now they have holes in them, so
239 * walk through the list and remove the holes
243 for (i = 0; i < user_mem_maps->n_maps; i++) {
244 if (!is_null_map(&user_mem_maps->maps[i])) {
245 struct user_mem_map *src, *dst;
247 src = &user_mem_maps->maps[i];
248 dst = &user_mem_maps->maps[cur_idx++];
251 memcpy(dst, src, sizeof(*src));
252 memset(src, 0, sizeof(*src));
256 user_mem_maps->n_maps = cur_idx;
261 vfio_open_group_fd(int iommu_group_num)
264 char filename[PATH_MAX];
265 struct rte_mp_msg mp_req, *mp_rep;
266 struct rte_mp_reply mp_reply;
267 struct timespec ts = {.tv_sec = 5, .tv_nsec = 0};
268 struct vfio_mp_param *p = (struct vfio_mp_param *)mp_req.param;
270 /* if primary, try to open the group */
271 if (internal_config.process_type == RTE_PROC_PRIMARY) {
272 /* try regular group format */
273 snprintf(filename, sizeof(filename),
274 VFIO_GROUP_FMT, iommu_group_num);
275 vfio_group_fd = open(filename, O_RDWR);
276 if (vfio_group_fd < 0) {
277 /* if file not found, it's not an error */
278 if (errno != ENOENT) {
279 RTE_LOG(ERR, EAL, "Cannot open %s: %s\n", filename,
284 /* special case: try no-IOMMU path as well */
285 snprintf(filename, sizeof(filename),
286 VFIO_NOIOMMU_GROUP_FMT,
288 vfio_group_fd = open(filename, O_RDWR);
289 if (vfio_group_fd < 0) {
290 if (errno != ENOENT) {
291 RTE_LOG(ERR, EAL, "Cannot open %s: %s\n", filename,
297 /* noiommu group found */
300 return vfio_group_fd;
302 /* if we're in a secondary process, request group fd from the primary
303 * process via mp channel.
305 p->req = SOCKET_REQ_GROUP;
306 p->group_num = iommu_group_num;
307 strcpy(mp_req.name, EAL_VFIO_MP);
308 mp_req.len_param = sizeof(*p);
312 if (rte_mp_request_sync(&mp_req, &mp_reply, &ts) == 0 &&
313 mp_reply.nb_received == 1) {
314 mp_rep = &mp_reply.msgs[0];
315 p = (struct vfio_mp_param *)mp_rep->param;
316 if (p->result == SOCKET_OK && mp_rep->num_fds == 1) {
317 vfio_group_fd = mp_rep->fds[0];
318 } else if (p->result == SOCKET_NO_FD) {
319 RTE_LOG(ERR, EAL, " bad VFIO group fd\n");
325 if (vfio_group_fd < 0)
326 RTE_LOG(ERR, EAL, " cannot request group fd\n");
327 return vfio_group_fd;
330 static struct vfio_config *
331 get_vfio_cfg_by_group_num(int iommu_group_num)
333 struct vfio_config *vfio_cfg;
336 for (i = 0; i < VFIO_MAX_CONTAINERS; i++) {
337 vfio_cfg = &vfio_cfgs[i];
338 for (j = 0; j < VFIO_MAX_GROUPS; j++) {
339 if (vfio_cfg->vfio_groups[j].group_num ==
349 vfio_get_group_fd(struct vfio_config *vfio_cfg,
354 struct vfio_group *cur_grp;
356 /* check if we already have the group descriptor open */
357 for (i = 0; i < VFIO_MAX_GROUPS; i++)
358 if (vfio_cfg->vfio_groups[i].group_num == iommu_group_num)
359 return vfio_cfg->vfio_groups[i].fd;
361 /* Lets see first if there is room for a new group */
362 if (vfio_cfg->vfio_active_groups == VFIO_MAX_GROUPS) {
363 RTE_LOG(ERR, EAL, "Maximum number of VFIO groups reached!\n");
367 /* Now lets get an index for the new group */
368 for (i = 0; i < VFIO_MAX_GROUPS; i++)
369 if (vfio_cfg->vfio_groups[i].group_num == -1) {
370 cur_grp = &vfio_cfg->vfio_groups[i];
374 /* This should not happen */
375 if (i == VFIO_MAX_GROUPS) {
376 RTE_LOG(ERR, EAL, "No VFIO group free slot found\n");
380 vfio_group_fd = vfio_open_group_fd(iommu_group_num);
381 if (vfio_group_fd < 0) {
382 RTE_LOG(ERR, EAL, "Failed to open group %d\n", iommu_group_num);
386 cur_grp->group_num = iommu_group_num;
387 cur_grp->fd = vfio_group_fd;
388 vfio_cfg->vfio_active_groups++;
390 return vfio_group_fd;
393 static struct vfio_config *
394 get_vfio_cfg_by_group_fd(int vfio_group_fd)
396 struct vfio_config *vfio_cfg;
399 for (i = 0; i < VFIO_MAX_CONTAINERS; i++) {
400 vfio_cfg = &vfio_cfgs[i];
401 for (j = 0; j < VFIO_MAX_GROUPS; j++)
402 if (vfio_cfg->vfio_groups[j].fd == vfio_group_fd)
409 static struct vfio_config *
410 get_vfio_cfg_by_container_fd(int container_fd)
414 for (i = 0; i < VFIO_MAX_CONTAINERS; i++) {
415 if (vfio_cfgs[i].vfio_container_fd == container_fd)
416 return &vfio_cfgs[i];
423 rte_vfio_get_group_fd(int iommu_group_num)
425 struct vfio_config *vfio_cfg;
427 /* get the vfio_config it belongs to */
428 vfio_cfg = get_vfio_cfg_by_group_num(iommu_group_num);
429 vfio_cfg = vfio_cfg ? vfio_cfg : default_vfio_cfg;
431 return vfio_get_group_fd(vfio_cfg, iommu_group_num);
435 get_vfio_group_idx(int vfio_group_fd)
437 struct vfio_config *vfio_cfg;
440 for (i = 0; i < VFIO_MAX_CONTAINERS; i++) {
441 vfio_cfg = &vfio_cfgs[i];
442 for (j = 0; j < VFIO_MAX_GROUPS; j++)
443 if (vfio_cfg->vfio_groups[j].fd == vfio_group_fd)
451 vfio_group_device_get(int vfio_group_fd)
453 struct vfio_config *vfio_cfg;
456 vfio_cfg = get_vfio_cfg_by_group_fd(vfio_group_fd);
457 if (vfio_cfg == NULL) {
458 RTE_LOG(ERR, EAL, " invalid group fd!\n");
462 i = get_vfio_group_idx(vfio_group_fd);
463 if (i < 0 || i > (VFIO_MAX_GROUPS - 1))
464 RTE_LOG(ERR, EAL, " wrong vfio_group index (%d)\n", i);
466 vfio_cfg->vfio_groups[i].devices++;
470 vfio_group_device_put(int vfio_group_fd)
472 struct vfio_config *vfio_cfg;
475 vfio_cfg = get_vfio_cfg_by_group_fd(vfio_group_fd);
476 if (vfio_cfg == NULL) {
477 RTE_LOG(ERR, EAL, " invalid group fd!\n");
481 i = get_vfio_group_idx(vfio_group_fd);
482 if (i < 0 || i > (VFIO_MAX_GROUPS - 1))
483 RTE_LOG(ERR, EAL, " wrong vfio_group index (%d)\n", i);
485 vfio_cfg->vfio_groups[i].devices--;
489 vfio_group_device_count(int vfio_group_fd)
491 struct vfio_config *vfio_cfg;
494 vfio_cfg = get_vfio_cfg_by_group_fd(vfio_group_fd);
495 if (vfio_cfg == NULL) {
496 RTE_LOG(ERR, EAL, " invalid group fd!\n");
500 i = get_vfio_group_idx(vfio_group_fd);
501 if (i < 0 || i > (VFIO_MAX_GROUPS - 1)) {
502 RTE_LOG(ERR, EAL, " wrong vfio_group index (%d)\n", i);
506 return vfio_cfg->vfio_groups[i].devices;
510 vfio_mem_event_callback(enum rte_mem_event type, const void *addr, size_t len,
511 void *arg __rte_unused)
513 struct rte_memseg_list *msl;
514 struct rte_memseg *ms;
517 msl = rte_mem_virt2memseg_list(addr);
519 /* for IOVA as VA mode, no need to care for IOVA addresses */
520 if (rte_eal_iova_mode() == RTE_IOVA_VA && msl->external == 0) {
521 uint64_t vfio_va = (uint64_t)(uintptr_t)addr;
522 if (type == RTE_MEM_EVENT_ALLOC)
523 vfio_dma_mem_map(default_vfio_cfg, vfio_va, vfio_va,
526 vfio_dma_mem_map(default_vfio_cfg, vfio_va, vfio_va,
531 /* memsegs are contiguous in memory */
532 ms = rte_mem_virt2memseg(addr, msl);
533 while (cur_len < len) {
534 /* some memory segments may have invalid IOVA */
535 if (ms->iova == RTE_BAD_IOVA) {
536 RTE_LOG(DEBUG, EAL, "Memory segment at %p has bad IOVA, skipping\n",
540 if (type == RTE_MEM_EVENT_ALLOC)
541 vfio_dma_mem_map(default_vfio_cfg, ms->addr_64,
542 ms->iova, ms->len, 1);
544 vfio_dma_mem_map(default_vfio_cfg, ms->addr_64,
545 ms->iova, ms->len, 0);
553 rte_vfio_clear_group(int vfio_group_fd)
556 struct vfio_config *vfio_cfg;
558 vfio_cfg = get_vfio_cfg_by_group_fd(vfio_group_fd);
559 if (vfio_cfg == NULL) {
560 RTE_LOG(ERR, EAL, " invalid group fd!\n");
564 i = get_vfio_group_idx(vfio_group_fd);
567 vfio_cfg->vfio_groups[i].group_num = -1;
568 vfio_cfg->vfio_groups[i].fd = -1;
569 vfio_cfg->vfio_groups[i].devices = 0;
570 vfio_cfg->vfio_active_groups--;
576 rte_vfio_setup_device(const char *sysfs_base, const char *dev_addr,
577 int *vfio_dev_fd, struct vfio_device_info *device_info)
579 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
580 rte_rwlock_t *mem_lock = &mcfg->memory_hotplug_lock;
581 struct vfio_group_status group_status = {
582 .argsz = sizeof(group_status)
584 struct vfio_config *vfio_cfg;
585 struct user_mem_maps *user_mem_maps;
586 int vfio_container_fd;
591 /* get group number */
592 ret = rte_vfio_get_group_num(sysfs_base, dev_addr, &iommu_group_num);
594 RTE_LOG(WARNING, EAL, " %s not managed by VFIO driver, skipping\n",
599 /* if negative, something failed */
603 /* get the actual group fd */
604 vfio_group_fd = rte_vfio_get_group_fd(iommu_group_num);
605 if (vfio_group_fd < 0)
608 /* if group_fd == 0, that means the device isn't managed by VFIO */
609 if (vfio_group_fd == 0) {
610 RTE_LOG(WARNING, EAL, " %s not managed by VFIO driver, skipping\n",
616 * at this point, we know that this group is viable (meaning, all devices
617 * are either bound to VFIO or not bound to anything)
620 /* check if the group is viable */
621 ret = ioctl(vfio_group_fd, VFIO_GROUP_GET_STATUS, &group_status);
623 RTE_LOG(ERR, EAL, " %s cannot get group status, "
624 "error %i (%s)\n", dev_addr, errno, strerror(errno));
625 close(vfio_group_fd);
626 rte_vfio_clear_group(vfio_group_fd);
628 } else if (!(group_status.flags & VFIO_GROUP_FLAGS_VIABLE)) {
629 RTE_LOG(ERR, EAL, " %s VFIO group is not viable!\n", dev_addr);
630 close(vfio_group_fd);
631 rte_vfio_clear_group(vfio_group_fd);
635 /* get the vfio_config it belongs to */
636 vfio_cfg = get_vfio_cfg_by_group_num(iommu_group_num);
637 vfio_cfg = vfio_cfg ? vfio_cfg : default_vfio_cfg;
638 vfio_container_fd = vfio_cfg->vfio_container_fd;
639 user_mem_maps = &vfio_cfg->mem_maps;
641 /* check if group does not have a container yet */
642 if (!(group_status.flags & VFIO_GROUP_FLAGS_CONTAINER_SET)) {
644 /* add group to a container */
645 ret = ioctl(vfio_group_fd, VFIO_GROUP_SET_CONTAINER,
648 RTE_LOG(ERR, EAL, " %s cannot add VFIO group to container, "
649 "error %i (%s)\n", dev_addr, errno, strerror(errno));
650 close(vfio_group_fd);
651 rte_vfio_clear_group(vfio_group_fd);
656 * pick an IOMMU type and set up DMA mappings for container
658 * needs to be done only once, only when first group is
659 * assigned to a container and only in primary process.
660 * Note this can happen several times with the hotplug
663 if (internal_config.process_type == RTE_PROC_PRIMARY &&
664 vfio_cfg->vfio_active_groups == 1 &&
665 vfio_group_device_count(vfio_group_fd) == 0) {
666 const struct vfio_iommu_type *t;
668 /* select an IOMMU type which we will be using */
669 t = vfio_set_iommu_type(vfio_container_fd);
672 " %s failed to select IOMMU type\n",
674 close(vfio_group_fd);
675 rte_vfio_clear_group(vfio_group_fd);
678 /* lock memory hotplug before mapping and release it
679 * after registering callback, to prevent races
681 rte_rwlock_read_lock(mem_lock);
682 if (vfio_cfg == default_vfio_cfg)
683 ret = t->dma_map_func(vfio_container_fd);
688 " %s DMA remapping failed, error %i (%s)\n",
689 dev_addr, errno, strerror(errno));
690 close(vfio_group_fd);
691 rte_vfio_clear_group(vfio_group_fd);
692 rte_rwlock_read_unlock(mem_lock);
696 vfio_cfg->vfio_iommu_type = t;
698 /* re-map all user-mapped segments */
699 rte_spinlock_recursive_lock(&user_mem_maps->lock);
701 /* this IOMMU type may not support DMA mapping, but
702 * if we have mappings in the list - that means we have
703 * previously mapped something successfully, so we can
704 * be sure that DMA mapping is supported.
706 for (i = 0; i < user_mem_maps->n_maps; i++) {
707 struct user_mem_map *map;
708 map = &user_mem_maps->maps[i];
710 ret = t->dma_user_map_func(
712 map->addr, map->iova, map->len,
715 RTE_LOG(ERR, EAL, "Couldn't map user memory for DMA: "
717 "iova: 0x%" PRIx64 " "
718 "len: 0x%" PRIu64 "\n",
719 map->addr, map->iova,
721 rte_spinlock_recursive_unlock(
722 &user_mem_maps->lock);
723 rte_rwlock_read_unlock(mem_lock);
727 rte_spinlock_recursive_unlock(&user_mem_maps->lock);
729 /* register callback for mem events */
730 if (vfio_cfg == default_vfio_cfg)
731 ret = rte_mem_event_callback_register(
732 VFIO_MEM_EVENT_CLB_NAME,
733 vfio_mem_event_callback, NULL);
736 /* unlock memory hotplug */
737 rte_rwlock_read_unlock(mem_lock);
739 if (ret && rte_errno != ENOTSUP) {
740 RTE_LOG(ERR, EAL, "Could not install memory event callback for VFIO\n");
744 RTE_LOG(DEBUG, EAL, "Memory event callbacks not supported\n");
746 RTE_LOG(DEBUG, EAL, "Installed memory event callback for VFIO\n");
750 /* get a file descriptor for the device */
751 *vfio_dev_fd = ioctl(vfio_group_fd, VFIO_GROUP_GET_DEVICE_FD, dev_addr);
752 if (*vfio_dev_fd < 0) {
753 /* if we cannot get a device fd, this implies a problem with
754 * the VFIO group or the container not having IOMMU configured.
757 RTE_LOG(WARNING, EAL, "Getting a vfio_dev_fd for %s failed\n",
759 close(vfio_group_fd);
760 rte_vfio_clear_group(vfio_group_fd);
764 /* test and setup the device */
765 ret = ioctl(*vfio_dev_fd, VFIO_DEVICE_GET_INFO, device_info);
767 RTE_LOG(ERR, EAL, " %s cannot get device info, "
768 "error %i (%s)\n", dev_addr, errno,
771 close(vfio_group_fd);
772 rte_vfio_clear_group(vfio_group_fd);
775 vfio_group_device_get(vfio_group_fd);
781 rte_vfio_release_device(const char *sysfs_base, const char *dev_addr,
784 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
785 rte_rwlock_t *mem_lock = &mcfg->memory_hotplug_lock;
786 struct vfio_group_status group_status = {
787 .argsz = sizeof(group_status)
789 struct vfio_config *vfio_cfg;
794 /* we don't want any DMA mapping messages to come while we're detaching
795 * VFIO device, because this might be the last device and we might need
796 * to unregister the callback.
798 rte_rwlock_read_lock(mem_lock);
800 /* get group number */
801 ret = rte_vfio_get_group_num(sysfs_base, dev_addr, &iommu_group_num);
803 RTE_LOG(WARNING, EAL, " %s not managed by VFIO driver\n",
805 /* This is an error at this point. */
810 /* get the actual group fd */
811 vfio_group_fd = rte_vfio_get_group_fd(iommu_group_num);
812 if (vfio_group_fd <= 0) {
813 RTE_LOG(INFO, EAL, "rte_vfio_get_group_fd failed for %s\n",
819 /* get the vfio_config it belongs to */
820 vfio_cfg = get_vfio_cfg_by_group_num(iommu_group_num);
821 vfio_cfg = vfio_cfg ? vfio_cfg : default_vfio_cfg;
823 /* At this point we got an active group. Closing it will make the
824 * container detachment. If this is the last active group, VFIO kernel
825 * code will unset the container and the IOMMU mappings.
828 /* Closing a device */
829 if (close(vfio_dev_fd) < 0) {
830 RTE_LOG(INFO, EAL, "Error when closing vfio_dev_fd for %s\n",
836 /* An VFIO group can have several devices attached. Just when there is
837 * no devices remaining should the group be closed.
839 vfio_group_device_put(vfio_group_fd);
840 if (!vfio_group_device_count(vfio_group_fd)) {
842 if (close(vfio_group_fd) < 0) {
843 RTE_LOG(INFO, EAL, "Error when closing vfio_group_fd for %s\n",
849 if (rte_vfio_clear_group(vfio_group_fd) < 0) {
850 RTE_LOG(INFO, EAL, "Error when clearing group for %s\n",
857 /* if there are no active device groups, unregister the callback to
858 * avoid spurious attempts to map/unmap memory from VFIO.
860 if (vfio_cfg == default_vfio_cfg && vfio_cfg->vfio_active_groups == 0 &&
861 rte_eal_process_type() != RTE_PROC_SECONDARY)
862 rte_mem_event_callback_unregister(VFIO_MEM_EVENT_CLB_NAME,
869 rte_rwlock_read_unlock(mem_lock);
874 rte_vfio_enable(const char *modname)
876 /* initialize group list */
880 rte_spinlock_recursive_t lock = RTE_SPINLOCK_RECURSIVE_INITIALIZER;
882 for (i = 0; i < VFIO_MAX_CONTAINERS; i++) {
883 vfio_cfgs[i].vfio_container_fd = -1;
884 vfio_cfgs[i].vfio_active_groups = 0;
885 vfio_cfgs[i].vfio_iommu_type = NULL;
886 vfio_cfgs[i].mem_maps.lock = lock;
888 for (j = 0; j < VFIO_MAX_GROUPS; j++) {
889 vfio_cfgs[i].vfio_groups[j].fd = -1;
890 vfio_cfgs[i].vfio_groups[j].group_num = -1;
891 vfio_cfgs[i].vfio_groups[j].devices = 0;
895 /* inform the user that we are probing for VFIO */
896 RTE_LOG(INFO, EAL, "Probing VFIO support...\n");
898 /* check if vfio module is loaded */
899 vfio_available = rte_eal_check_module(modname);
901 /* return error directly */
902 if (vfio_available == -1) {
903 RTE_LOG(INFO, EAL, "Could not get loaded module details!\n");
907 /* return 0 if VFIO modules not loaded */
908 if (vfio_available == 0) {
909 RTE_LOG(DEBUG, EAL, "VFIO modules not loaded, "
910 "skipping VFIO support...\n");
914 if (internal_config.process_type == RTE_PROC_PRIMARY) {
915 /* open a new container */
916 default_vfio_cfg->vfio_container_fd =
917 rte_vfio_get_container_fd();
919 /* get the default container from the primary process */
920 default_vfio_cfg->vfio_container_fd =
921 vfio_get_default_container_fd();
924 /* check if we have VFIO driver enabled */
925 if (default_vfio_cfg->vfio_container_fd != -1) {
926 RTE_LOG(NOTICE, EAL, "VFIO support initialized\n");
927 default_vfio_cfg->vfio_enabled = 1;
929 RTE_LOG(NOTICE, EAL, "VFIO support could not be initialized\n");
936 rte_vfio_is_enabled(const char *modname)
938 const int mod_available = rte_eal_check_module(modname) > 0;
939 return default_vfio_cfg->vfio_enabled && mod_available;
943 vfio_get_default_container_fd(void)
945 struct rte_mp_msg mp_req, *mp_rep;
946 struct rte_mp_reply mp_reply;
947 struct timespec ts = {.tv_sec = 5, .tv_nsec = 0};
948 struct vfio_mp_param *p = (struct vfio_mp_param *)mp_req.param;
950 if (default_vfio_cfg->vfio_enabled)
951 return default_vfio_cfg->vfio_container_fd;
953 if (internal_config.process_type == RTE_PROC_PRIMARY) {
954 /* if we were secondary process we would try requesting
955 * container fd from the primary, but we're the primary
956 * process so just exit here
961 p->req = SOCKET_REQ_DEFAULT_CONTAINER;
962 strcpy(mp_req.name, EAL_VFIO_MP);
963 mp_req.len_param = sizeof(*p);
966 if (rte_mp_request_sync(&mp_req, &mp_reply, &ts) == 0 &&
967 mp_reply.nb_received == 1) {
968 mp_rep = &mp_reply.msgs[0];
969 p = (struct vfio_mp_param *)mp_rep->param;
970 if (p->result == SOCKET_OK && mp_rep->num_fds == 1) {
972 return mp_rep->fds[0];
977 RTE_LOG(ERR, EAL, " cannot request default container fd\n");
981 const struct vfio_iommu_type *
982 vfio_set_iommu_type(int vfio_container_fd)
985 for (idx = 0; idx < RTE_DIM(iommu_types); idx++) {
986 const struct vfio_iommu_type *t = &iommu_types[idx];
988 int ret = ioctl(vfio_container_fd, VFIO_SET_IOMMU,
991 RTE_LOG(NOTICE, EAL, " using IOMMU type %d (%s)\n",
992 t->type_id, t->name);
995 /* not an error, there may be more supported IOMMU types */
996 RTE_LOG(DEBUG, EAL, " set IOMMU type %d (%s) failed, "
997 "error %i (%s)\n", t->type_id, t->name, errno,
1000 /* if we didn't find a suitable IOMMU type, fail */
1005 vfio_has_supported_extensions(int vfio_container_fd)
1008 unsigned idx, n_extensions = 0;
1009 for (idx = 0; idx < RTE_DIM(iommu_types); idx++) {
1010 const struct vfio_iommu_type *t = &iommu_types[idx];
1012 ret = ioctl(vfio_container_fd, VFIO_CHECK_EXTENSION,
1015 RTE_LOG(ERR, EAL, " could not get IOMMU type, "
1016 "error %i (%s)\n", errno,
1018 close(vfio_container_fd);
1020 } else if (ret == 1) {
1021 /* we found a supported extension */
1024 RTE_LOG(DEBUG, EAL, " IOMMU type %d (%s) is %s\n",
1025 t->type_id, t->name,
1026 ret ? "supported" : "not supported");
1029 /* if we didn't find any supported IOMMU types, fail */
1030 if (!n_extensions) {
1031 close(vfio_container_fd);
1039 rte_vfio_get_container_fd(void)
1041 int ret, vfio_container_fd;
1042 struct rte_mp_msg mp_req, *mp_rep;
1043 struct rte_mp_reply mp_reply;
1044 struct timespec ts = {.tv_sec = 5, .tv_nsec = 0};
1045 struct vfio_mp_param *p = (struct vfio_mp_param *)mp_req.param;
1048 /* if we're in a primary process, try to open the container */
1049 if (internal_config.process_type == RTE_PROC_PRIMARY) {
1050 vfio_container_fd = open(VFIO_CONTAINER_PATH, O_RDWR);
1051 if (vfio_container_fd < 0) {
1052 RTE_LOG(ERR, EAL, " cannot open VFIO container, "
1053 "error %i (%s)\n", errno, strerror(errno));
1057 /* check VFIO API version */
1058 ret = ioctl(vfio_container_fd, VFIO_GET_API_VERSION);
1059 if (ret != VFIO_API_VERSION) {
1061 RTE_LOG(ERR, EAL, " could not get VFIO API version, "
1062 "error %i (%s)\n", errno, strerror(errno));
1064 RTE_LOG(ERR, EAL, " unsupported VFIO API version!\n");
1065 close(vfio_container_fd);
1069 ret = vfio_has_supported_extensions(vfio_container_fd);
1071 RTE_LOG(ERR, EAL, " no supported IOMMU "
1072 "extensions found!\n");
1076 return vfio_container_fd;
1079 * if we're in a secondary process, request container fd from the
1080 * primary process via mp channel
1082 p->req = SOCKET_REQ_CONTAINER;
1083 strcpy(mp_req.name, EAL_VFIO_MP);
1084 mp_req.len_param = sizeof(*p);
1087 vfio_container_fd = -1;
1088 if (rte_mp_request_sync(&mp_req, &mp_reply, &ts) == 0 &&
1089 mp_reply.nb_received == 1) {
1090 mp_rep = &mp_reply.msgs[0];
1091 p = (struct vfio_mp_param *)mp_rep->param;
1092 if (p->result == SOCKET_OK && mp_rep->num_fds == 1) {
1093 vfio_container_fd = mp_rep->fds[0];
1094 free(mp_reply.msgs);
1095 return vfio_container_fd;
1097 free(mp_reply.msgs);
1100 RTE_LOG(ERR, EAL, " cannot request container fd\n");
1105 rte_vfio_get_group_num(const char *sysfs_base,
1106 const char *dev_addr, int *iommu_group_num)
1108 char linkname[PATH_MAX];
1109 char filename[PATH_MAX];
1110 char *tok[16], *group_tok, *end;
1113 memset(linkname, 0, sizeof(linkname));
1114 memset(filename, 0, sizeof(filename));
1116 /* try to find out IOMMU group for this device */
1117 snprintf(linkname, sizeof(linkname),
1118 "%s/%s/iommu_group", sysfs_base, dev_addr);
1120 ret = readlink(linkname, filename, sizeof(filename));
1122 /* if the link doesn't exist, no VFIO for us */
1126 ret = rte_strsplit(filename, sizeof(filename),
1127 tok, RTE_DIM(tok), '/');
1130 RTE_LOG(ERR, EAL, " %s cannot get IOMMU group\n", dev_addr);
1134 /* IOMMU group is always the last token */
1136 group_tok = tok[ret - 1];
1138 *iommu_group_num = strtol(group_tok, &end, 10);
1139 if ((end != group_tok && *end != '\0') || errno != 0) {
1140 RTE_LOG(ERR, EAL, " %s error parsing IOMMU number!\n", dev_addr);
1148 type1_map(const struct rte_memseg_list *msl, const struct rte_memseg *ms,
1151 int *vfio_container_fd = arg;
1156 return vfio_type1_dma_mem_map(*vfio_container_fd, ms->addr_64, ms->iova,
1161 vfio_type1_dma_mem_map(int vfio_container_fd, uint64_t vaddr, uint64_t iova,
1162 uint64_t len, int do_map)
1164 struct vfio_iommu_type1_dma_map dma_map;
1165 struct vfio_iommu_type1_dma_unmap dma_unmap;
1169 memset(&dma_map, 0, sizeof(dma_map));
1170 dma_map.argsz = sizeof(struct vfio_iommu_type1_dma_map);
1171 dma_map.vaddr = vaddr;
1173 dma_map.iova = iova;
1174 dma_map.flags = VFIO_DMA_MAP_FLAG_READ |
1175 VFIO_DMA_MAP_FLAG_WRITE;
1177 ret = ioctl(vfio_container_fd, VFIO_IOMMU_MAP_DMA, &dma_map);
1179 RTE_LOG(ERR, EAL, " cannot set up DMA remapping, error %i (%s)\n",
1180 errno, strerror(errno));
1184 memset(&dma_unmap, 0, sizeof(dma_unmap));
1185 dma_unmap.argsz = sizeof(struct vfio_iommu_type1_dma_unmap);
1186 dma_unmap.size = len;
1187 dma_unmap.iova = iova;
1189 ret = ioctl(vfio_container_fd, VFIO_IOMMU_UNMAP_DMA,
1192 RTE_LOG(ERR, EAL, " cannot clear DMA remapping, error %i (%s)\n",
1193 errno, strerror(errno));
1202 vfio_type1_dma_map(int vfio_container_fd)
1204 return rte_memseg_walk(type1_map, &vfio_container_fd);
1208 vfio_spapr_dma_do_map(int vfio_container_fd, uint64_t vaddr, uint64_t iova,
1209 uint64_t len, int do_map)
1211 struct vfio_iommu_type1_dma_map dma_map;
1212 struct vfio_iommu_type1_dma_unmap dma_unmap;
1214 struct vfio_iommu_spapr_register_memory reg = {
1215 .argsz = sizeof(reg),
1218 reg.vaddr = (uintptr_t) vaddr;
1222 ret = ioctl(vfio_container_fd,
1223 VFIO_IOMMU_SPAPR_REGISTER_MEMORY, ®);
1225 RTE_LOG(ERR, EAL, " cannot register vaddr for IOMMU, "
1226 "error %i (%s)\n", errno, strerror(errno));
1230 memset(&dma_map, 0, sizeof(dma_map));
1231 dma_map.argsz = sizeof(struct vfio_iommu_type1_dma_map);
1232 dma_map.vaddr = vaddr;
1234 dma_map.iova = iova;
1235 dma_map.flags = VFIO_DMA_MAP_FLAG_READ |
1236 VFIO_DMA_MAP_FLAG_WRITE;
1238 ret = ioctl(vfio_container_fd, VFIO_IOMMU_MAP_DMA, &dma_map);
1240 RTE_LOG(ERR, EAL, " cannot set up DMA remapping, error %i (%s)\n",
1241 errno, strerror(errno));
1246 ret = ioctl(vfio_container_fd,
1247 VFIO_IOMMU_SPAPR_UNREGISTER_MEMORY, ®);
1249 RTE_LOG(ERR, EAL, " cannot unregister vaddr for IOMMU, error %i (%s)\n",
1250 errno, strerror(errno));
1254 memset(&dma_unmap, 0, sizeof(dma_unmap));
1255 dma_unmap.argsz = sizeof(struct vfio_iommu_type1_dma_unmap);
1256 dma_unmap.size = len;
1257 dma_unmap.iova = iova;
1259 ret = ioctl(vfio_container_fd, VFIO_IOMMU_UNMAP_DMA,
1262 RTE_LOG(ERR, EAL, " cannot clear DMA remapping, error %i (%s)\n",
1263 errno, strerror(errno));
1272 vfio_spapr_map_walk(const struct rte_memseg_list *msl,
1273 const struct rte_memseg *ms, void *arg)
1275 int *vfio_container_fd = arg;
1280 return vfio_spapr_dma_do_map(*vfio_container_fd, ms->addr_64, ms->iova,
1284 struct spapr_walk_param {
1285 uint64_t window_size;
1286 uint64_t hugepage_sz;
1289 vfio_spapr_window_size_walk(const struct rte_memseg_list *msl,
1290 const struct rte_memseg *ms, void *arg)
1292 struct spapr_walk_param *param = arg;
1293 uint64_t max = ms->iova + ms->len;
1298 if (max > param->window_size) {
1299 param->hugepage_sz = ms->hugepage_sz;
1300 param->window_size = max;
1307 vfio_spapr_create_new_dma_window(int vfio_container_fd,
1308 struct vfio_iommu_spapr_tce_create *create) {
1309 struct vfio_iommu_spapr_tce_remove remove = {
1310 .argsz = sizeof(remove),
1312 struct vfio_iommu_spapr_tce_info info = {
1313 .argsz = sizeof(info),
1317 /* query spapr iommu info */
1318 ret = ioctl(vfio_container_fd, VFIO_IOMMU_SPAPR_TCE_GET_INFO, &info);
1320 RTE_LOG(ERR, EAL, " cannot get iommu info, "
1321 "error %i (%s)\n", errno, strerror(errno));
1325 /* remove default DMA of 32 bit window */
1326 remove.start_addr = info.dma32_window_start;
1327 ret = ioctl(vfio_container_fd, VFIO_IOMMU_SPAPR_TCE_REMOVE, &remove);
1329 RTE_LOG(ERR, EAL, " cannot remove default DMA window, "
1330 "error %i (%s)\n", errno, strerror(errno));
1334 /* create new DMA window */
1335 ret = ioctl(vfio_container_fd, VFIO_IOMMU_SPAPR_TCE_CREATE, create);
1337 RTE_LOG(ERR, EAL, " cannot create new DMA window, "
1338 "error %i (%s)\n", errno, strerror(errno));
1342 if (create->start_addr != 0) {
1343 RTE_LOG(ERR, EAL, " DMA window start address != 0\n");
1351 vfio_spapr_dma_mem_map(int vfio_container_fd, uint64_t vaddr, uint64_t iova,
1352 uint64_t len, int do_map)
1354 struct spapr_walk_param param;
1355 struct vfio_iommu_spapr_tce_create create = {
1356 .argsz = sizeof(create),
1358 struct vfio_config *vfio_cfg;
1359 struct user_mem_maps *user_mem_maps;
1362 vfio_cfg = get_vfio_cfg_by_container_fd(vfio_container_fd);
1363 if (vfio_cfg == NULL) {
1364 RTE_LOG(ERR, EAL, " invalid container fd!\n");
1368 user_mem_maps = &vfio_cfg->mem_maps;
1369 rte_spinlock_recursive_lock(&user_mem_maps->lock);
1371 /* check if window size needs to be adjusted */
1372 memset(¶m, 0, sizeof(param));
1374 /* we're inside a callback so use thread-unsafe version */
1375 if (rte_memseg_walk_thread_unsafe(vfio_spapr_window_size_walk,
1377 RTE_LOG(ERR, EAL, "Could not get window size\n");
1382 /* also check user maps */
1383 for (i = 0; i < user_mem_maps->n_maps; i++) {
1384 uint64_t max = user_mem_maps->maps[i].iova +
1385 user_mem_maps->maps[i].len;
1386 create.window_size = RTE_MAX(create.window_size, max);
1389 /* sPAPR requires window size to be a power of 2 */
1390 create.window_size = rte_align64pow2(param.window_size);
1391 create.page_shift = __builtin_ctzll(param.hugepage_sz);
1396 /* re-create window and remap the entire memory */
1397 if (iova > create.window_size) {
1398 if (vfio_spapr_create_new_dma_window(vfio_container_fd,
1400 RTE_LOG(ERR, EAL, "Could not create new DMA window\n");
1404 /* we're inside a callback, so use thread-unsafe version
1406 if (rte_memseg_walk_thread_unsafe(vfio_spapr_map_walk,
1407 &vfio_container_fd) < 0) {
1408 RTE_LOG(ERR, EAL, "Could not recreate DMA maps\n");
1412 /* remap all user maps */
1413 for (i = 0; i < user_mem_maps->n_maps; i++) {
1414 struct user_mem_map *map =
1415 &user_mem_maps->maps[i];
1416 if (vfio_spapr_dma_do_map(vfio_container_fd,
1417 map->addr, map->iova, map->len,
1419 RTE_LOG(ERR, EAL, "Could not recreate user DMA maps\n");
1426 /* now that we've remapped all of the memory that was present
1427 * before, map the segment that we were requested to map.
1429 * however, if we were called by the callback, the memory we
1430 * were called with was already in the memseg list, so previous
1431 * mapping should've mapped that segment already.
1433 * virt2memseg_list is a relatively cheap check, so use that. if
1434 * memory is within any memseg list, it's a memseg, so it's
1437 addr = (void *)(uintptr_t)vaddr;
1438 if (rte_mem_virt2memseg_list(addr) == NULL &&
1439 vfio_spapr_dma_do_map(vfio_container_fd,
1440 vaddr, iova, len, 1) < 0) {
1441 RTE_LOG(ERR, EAL, "Could not map segment\n");
1446 /* for unmap, check if iova within DMA window */
1447 if (iova > create.window_size) {
1448 RTE_LOG(ERR, EAL, "iova beyond DMA window for unmap");
1453 vfio_spapr_dma_do_map(vfio_container_fd, vaddr, iova, len, 0);
1456 rte_spinlock_recursive_unlock(&user_mem_maps->lock);
1461 vfio_spapr_dma_map(int vfio_container_fd)
1463 struct vfio_iommu_spapr_tce_create create = {
1464 .argsz = sizeof(create),
1466 struct spapr_walk_param param;
1468 memset(¶m, 0, sizeof(param));
1470 /* create DMA window from 0 to max(phys_addr + len) */
1471 rte_memseg_walk(vfio_spapr_window_size_walk, ¶m);
1473 /* sPAPR requires window size to be a power of 2 */
1474 create.window_size = rte_align64pow2(param.window_size);
1475 create.page_shift = __builtin_ctzll(param.hugepage_sz);
1478 if (vfio_spapr_create_new_dma_window(vfio_container_fd, &create) < 0) {
1479 RTE_LOG(ERR, EAL, "Could not create new DMA window\n");
1483 /* map all DPDK segments for DMA. use 1:1 PA to IOVA mapping */
1484 if (rte_memseg_walk(vfio_spapr_map_walk, &vfio_container_fd) < 0)
1491 vfio_noiommu_dma_map(int __rte_unused vfio_container_fd)
1493 /* No-IOMMU mode does not need DMA mapping */
1498 vfio_noiommu_dma_mem_map(int __rte_unused vfio_container_fd,
1499 uint64_t __rte_unused vaddr,
1500 uint64_t __rte_unused iova, uint64_t __rte_unused len,
1501 int __rte_unused do_map)
1503 /* No-IOMMU mode does not need DMA mapping */
1508 vfio_dma_mem_map(struct vfio_config *vfio_cfg, uint64_t vaddr, uint64_t iova,
1509 uint64_t len, int do_map)
1511 const struct vfio_iommu_type *t = vfio_cfg->vfio_iommu_type;
1514 RTE_LOG(ERR, EAL, " VFIO support not initialized\n");
1519 if (!t->dma_user_map_func) {
1521 " VFIO custom DMA region maping not supported by IOMMU %s\n",
1523 rte_errno = ENOTSUP;
1527 return t->dma_user_map_func(vfio_cfg->vfio_container_fd, vaddr, iova,
1532 container_dma_map(struct vfio_config *vfio_cfg, uint64_t vaddr, uint64_t iova,
1535 struct user_mem_map *new_map;
1536 struct user_mem_maps *user_mem_maps;
1539 user_mem_maps = &vfio_cfg->mem_maps;
1540 rte_spinlock_recursive_lock(&user_mem_maps->lock);
1541 if (user_mem_maps->n_maps == VFIO_MAX_USER_MEM_MAPS) {
1542 RTE_LOG(ERR, EAL, "No more space for user mem maps\n");
1548 if (vfio_dma_mem_map(vfio_cfg, vaddr, iova, len, 1)) {
1549 /* technically, this will fail if there are currently no devices
1550 * plugged in, even if a device were added later, this mapping
1551 * might have succeeded. however, since we cannot verify if this
1552 * is a valid mapping without having a device attached, consider
1553 * this to be unsupported, because we can't just store any old
1554 * mapping and pollute list of active mappings willy-nilly.
1556 RTE_LOG(ERR, EAL, "Couldn't map new region for DMA\n");
1560 /* create new user mem map entry */
1561 new_map = &user_mem_maps->maps[user_mem_maps->n_maps++];
1562 new_map->addr = vaddr;
1563 new_map->iova = iova;
1566 compact_user_maps(user_mem_maps);
1568 rte_spinlock_recursive_unlock(&user_mem_maps->lock);
1573 container_dma_unmap(struct vfio_config *vfio_cfg, uint64_t vaddr, uint64_t iova,
1576 struct user_mem_map *map, *new_map = NULL;
1577 struct user_mem_maps *user_mem_maps;
1580 user_mem_maps = &vfio_cfg->mem_maps;
1581 rte_spinlock_recursive_lock(&user_mem_maps->lock);
1583 /* find our mapping */
1584 map = find_user_mem_map(user_mem_maps, vaddr, iova, len);
1586 RTE_LOG(ERR, EAL, "Couldn't find previously mapped region\n");
1591 if (map->addr != vaddr || map->iova != iova || map->len != len) {
1592 /* we're partially unmapping a previously mapped region, so we
1593 * need to split entry into two.
1595 if (user_mem_maps->n_maps == VFIO_MAX_USER_MEM_MAPS) {
1596 RTE_LOG(ERR, EAL, "Not enough space to store partial mapping\n");
1601 new_map = &user_mem_maps->maps[user_mem_maps->n_maps++];
1604 /* unmap the entry */
1605 if (vfio_dma_mem_map(vfio_cfg, vaddr, iova, len, 0)) {
1606 /* there may not be any devices plugged in, so unmapping will
1607 * fail with ENODEV/ENOTSUP rte_errno values, but that doesn't
1608 * stop us from removing the mapping, as the assumption is we
1609 * won't be needing this memory any more and thus will want to
1610 * prevent it from being remapped again on hotplug. so, only
1611 * fail if we indeed failed to unmap (e.g. if the mapping was
1612 * within our mapped range but had invalid alignment).
1614 if (rte_errno != ENODEV && rte_errno != ENOTSUP) {
1615 RTE_LOG(ERR, EAL, "Couldn't unmap region for DMA\n");
1619 RTE_LOG(DEBUG, EAL, "DMA unmapping failed, but removing mappings anyway\n");
1622 /* remove map from the list of active mappings */
1623 if (new_map != NULL) {
1624 adjust_map(map, new_map, vaddr, len);
1626 /* if we've created a new map by splitting, sort everything */
1627 if (!is_null_map(new_map)) {
1628 compact_user_maps(user_mem_maps);
1630 /* we've created a new mapping, but it was unused */
1631 user_mem_maps->n_maps--;
1634 memset(map, 0, sizeof(*map));
1635 compact_user_maps(user_mem_maps);
1636 user_mem_maps->n_maps--;
1640 rte_spinlock_recursive_unlock(&user_mem_maps->lock);
1645 rte_vfio_dma_map(uint64_t vaddr, uint64_t iova, uint64_t len)
1652 return container_dma_map(default_vfio_cfg, vaddr, iova, len);
1656 rte_vfio_dma_unmap(uint64_t vaddr, uint64_t iova, uint64_t len)
1663 return container_dma_unmap(default_vfio_cfg, vaddr, iova, len);
1667 rte_vfio_noiommu_is_enabled(void)
1673 fd = open(VFIO_NOIOMMU_MODE, O_RDONLY);
1675 if (errno != ENOENT) {
1676 RTE_LOG(ERR, EAL, " cannot open vfio noiommu file %i (%s)\n",
1677 errno, strerror(errno));
1681 * else the file does not exists
1682 * i.e. noiommu is not enabled
1687 cnt = read(fd, &c, 1);
1690 RTE_LOG(ERR, EAL, " unable to read from vfio noiommu "
1691 "file %i (%s)\n", errno, strerror(errno));
1699 rte_vfio_container_create(void)
1703 /* Find an empty slot to store new vfio config */
1704 for (i = 1; i < VFIO_MAX_CONTAINERS; i++) {
1705 if (vfio_cfgs[i].vfio_container_fd == -1)
1709 if (i == VFIO_MAX_CONTAINERS) {
1710 RTE_LOG(ERR, EAL, "exceed max vfio container limit\n");
1714 vfio_cfgs[i].vfio_container_fd = rte_vfio_get_container_fd();
1715 if (vfio_cfgs[i].vfio_container_fd < 0) {
1716 RTE_LOG(NOTICE, EAL, "fail to create a new container\n");
1720 return vfio_cfgs[i].vfio_container_fd;
1723 int __rte_experimental
1724 rte_vfio_container_destroy(int container_fd)
1726 struct vfio_config *vfio_cfg;
1729 vfio_cfg = get_vfio_cfg_by_container_fd(container_fd);
1730 if (vfio_cfg == NULL) {
1731 RTE_LOG(ERR, EAL, "Invalid container fd\n");
1735 for (i = 0; i < VFIO_MAX_GROUPS; i++)
1736 if (vfio_cfg->vfio_groups[i].group_num != -1)
1737 rte_vfio_container_group_unbind(container_fd,
1738 vfio_cfg->vfio_groups[i].group_num);
1740 close(container_fd);
1741 vfio_cfg->vfio_container_fd = -1;
1742 vfio_cfg->vfio_active_groups = 0;
1743 vfio_cfg->vfio_iommu_type = NULL;
1749 rte_vfio_container_group_bind(int container_fd, int iommu_group_num)
1751 struct vfio_config *vfio_cfg;
1753 vfio_cfg = get_vfio_cfg_by_container_fd(container_fd);
1754 if (vfio_cfg == NULL) {
1755 RTE_LOG(ERR, EAL, "Invalid container fd\n");
1759 return vfio_get_group_fd(vfio_cfg, iommu_group_num);
1763 rte_vfio_container_group_unbind(int container_fd, int iommu_group_num)
1765 struct vfio_config *vfio_cfg;
1766 struct vfio_group *cur_grp = NULL;
1769 vfio_cfg = get_vfio_cfg_by_container_fd(container_fd);
1770 if (vfio_cfg == NULL) {
1771 RTE_LOG(ERR, EAL, "Invalid container fd\n");
1775 for (i = 0; i < VFIO_MAX_GROUPS; i++) {
1776 if (vfio_cfg->vfio_groups[i].group_num == iommu_group_num) {
1777 cur_grp = &vfio_cfg->vfio_groups[i];
1782 /* This should not happen */
1783 if (i == VFIO_MAX_GROUPS || cur_grp == NULL) {
1784 RTE_LOG(ERR, EAL, "Specified group number not found\n");
1788 if (cur_grp->fd >= 0 && close(cur_grp->fd) < 0) {
1789 RTE_LOG(ERR, EAL, "Error when closing vfio_group_fd for"
1790 " iommu_group_num %d\n", iommu_group_num);
1793 cur_grp->group_num = -1;
1795 cur_grp->devices = 0;
1796 vfio_cfg->vfio_active_groups--;
1802 rte_vfio_container_dma_map(int container_fd, uint64_t vaddr, uint64_t iova,
1805 struct vfio_config *vfio_cfg;
1812 vfio_cfg = get_vfio_cfg_by_container_fd(container_fd);
1813 if (vfio_cfg == NULL) {
1814 RTE_LOG(ERR, EAL, "Invalid container fd\n");
1818 return container_dma_map(vfio_cfg, vaddr, iova, len);
1822 rte_vfio_container_dma_unmap(int container_fd, uint64_t vaddr, uint64_t iova,
1825 struct vfio_config *vfio_cfg;
1832 vfio_cfg = get_vfio_cfg_by_container_fd(container_fd);
1833 if (vfio_cfg == NULL) {
1834 RTE_LOG(ERR, EAL, "Invalid container fd\n");
1838 return container_dma_unmap(vfio_cfg, vaddr, iova, len);
1844 rte_vfio_dma_map(uint64_t __rte_unused vaddr, __rte_unused uint64_t iova,
1845 __rte_unused uint64_t len)
1851 rte_vfio_dma_unmap(uint64_t __rte_unused vaddr, uint64_t __rte_unused iova,
1852 __rte_unused uint64_t len)
1858 rte_vfio_setup_device(__rte_unused const char *sysfs_base,
1859 __rte_unused const char *dev_addr,
1860 __rte_unused int *vfio_dev_fd,
1861 __rte_unused struct vfio_device_info *device_info)
1867 rte_vfio_release_device(__rte_unused const char *sysfs_base,
1868 __rte_unused const char *dev_addr, __rte_unused int fd)
1874 rte_vfio_enable(__rte_unused const char *modname)
1880 rte_vfio_is_enabled(__rte_unused const char *modname)
1886 rte_vfio_noiommu_is_enabled(void)
1892 rte_vfio_clear_group(__rte_unused int vfio_group_fd)
1898 rte_vfio_get_group_num(__rte_unused const char *sysfs_base,
1899 __rte_unused const char *dev_addr,
1900 __rte_unused int *iommu_group_num)
1906 rte_vfio_get_container_fd(void)
1912 rte_vfio_get_group_fd(__rte_unused int iommu_group_num)
1918 rte_vfio_container_create(void)
1924 rte_vfio_container_destroy(__rte_unused int container_fd)
1930 rte_vfio_container_group_bind(__rte_unused int container_fd,
1931 __rte_unused int iommu_group_num)
1937 rte_vfio_container_group_unbind(__rte_unused int container_fd,
1938 __rte_unused int iommu_group_num)
1944 rte_vfio_container_dma_map(__rte_unused int container_fd,
1945 __rte_unused uint64_t vaddr,
1946 __rte_unused uint64_t iova,
1947 __rte_unused uint64_t len)
1953 rte_vfio_container_dma_unmap(__rte_unused int container_fd,
1954 __rte_unused uint64_t vaddr,
1955 __rte_unused uint64_t iova,
1956 __rte_unused uint64_t len)
1961 #endif /* VFIO_PRESENT */