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 vfio_sync_default_container(void)
555 struct rte_mp_msg mp_req, *mp_rep;
556 struct rte_mp_reply mp_reply;
557 struct timespec ts = {.tv_sec = 5, .tv_nsec = 0};
558 struct vfio_mp_param *p = (struct vfio_mp_param *)mp_req.param;
562 /* cannot be called from primary */
563 if (rte_eal_process_type() != RTE_PROC_SECONDARY)
566 /* default container fd should have been opened in rte_vfio_enable() */
567 if (!default_vfio_cfg->vfio_enabled ||
568 default_vfio_cfg->vfio_container_fd < 0) {
569 RTE_LOG(ERR, EAL, "VFIO support is not initialized\n");
573 /* find default container's IOMMU type */
574 p->req = SOCKET_REQ_IOMMU_TYPE;
575 strcpy(mp_req.name, EAL_VFIO_MP);
576 mp_req.len_param = sizeof(*p);
580 if (rte_mp_request_sync(&mp_req, &mp_reply, &ts) == 0 &&
581 mp_reply.nb_received == 1) {
582 mp_rep = &mp_reply.msgs[0];
583 p = (struct vfio_mp_param *)mp_rep->param;
584 if (p->result == SOCKET_OK)
585 iommu_type_id = p->iommu_type_id;
588 if (iommu_type_id < 0) {
589 RTE_LOG(ERR, EAL, "Could not get IOMMU type for default container\n");
593 /* we now have an fd for default container, as well as its IOMMU type.
594 * now, set up default VFIO container config to match.
596 for (i = 0; i < RTE_DIM(iommu_types); i++) {
597 const struct vfio_iommu_type *t = &iommu_types[i];
598 if (t->type_id != iommu_type_id)
601 /* we found our IOMMU type */
602 default_vfio_cfg->vfio_iommu_type = t;
606 RTE_LOG(ERR, EAL, "Could not find IOMMU type id (%i)\n",
612 rte_vfio_clear_group(int vfio_group_fd)
615 struct vfio_config *vfio_cfg;
617 vfio_cfg = get_vfio_cfg_by_group_fd(vfio_group_fd);
618 if (vfio_cfg == NULL) {
619 RTE_LOG(ERR, EAL, " invalid group fd!\n");
623 i = get_vfio_group_idx(vfio_group_fd);
626 vfio_cfg->vfio_groups[i].group_num = -1;
627 vfio_cfg->vfio_groups[i].fd = -1;
628 vfio_cfg->vfio_groups[i].devices = 0;
629 vfio_cfg->vfio_active_groups--;
635 rte_vfio_setup_device(const char *sysfs_base, const char *dev_addr,
636 int *vfio_dev_fd, struct vfio_device_info *device_info)
638 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
639 rte_rwlock_t *mem_lock = &mcfg->memory_hotplug_lock;
640 struct vfio_group_status group_status = {
641 .argsz = sizeof(group_status)
643 struct vfio_config *vfio_cfg;
644 struct user_mem_maps *user_mem_maps;
645 int vfio_container_fd;
650 /* get group number */
651 ret = rte_vfio_get_group_num(sysfs_base, dev_addr, &iommu_group_num);
653 RTE_LOG(WARNING, EAL, " %s not managed by VFIO driver, skipping\n",
658 /* if negative, something failed */
662 /* get the actual group fd */
663 vfio_group_fd = rte_vfio_get_group_fd(iommu_group_num);
664 if (vfio_group_fd < 0)
667 /* if group_fd == 0, that means the device isn't managed by VFIO */
668 if (vfio_group_fd == 0) {
669 RTE_LOG(WARNING, EAL, " %s not managed by VFIO driver, skipping\n",
675 * at this point, we know that this group is viable (meaning, all devices
676 * are either bound to VFIO or not bound to anything)
679 /* check if the group is viable */
680 ret = ioctl(vfio_group_fd, VFIO_GROUP_GET_STATUS, &group_status);
682 RTE_LOG(ERR, EAL, " %s cannot get group status, "
683 "error %i (%s)\n", dev_addr, errno, strerror(errno));
684 close(vfio_group_fd);
685 rte_vfio_clear_group(vfio_group_fd);
687 } else if (!(group_status.flags & VFIO_GROUP_FLAGS_VIABLE)) {
688 RTE_LOG(ERR, EAL, " %s VFIO group is not viable! "
689 "Not all devices in IOMMU group bound to VFIO or unbound\n",
691 close(vfio_group_fd);
692 rte_vfio_clear_group(vfio_group_fd);
696 /* get the vfio_config it belongs to */
697 vfio_cfg = get_vfio_cfg_by_group_num(iommu_group_num);
698 vfio_cfg = vfio_cfg ? vfio_cfg : default_vfio_cfg;
699 vfio_container_fd = vfio_cfg->vfio_container_fd;
700 user_mem_maps = &vfio_cfg->mem_maps;
702 /* check if group does not have a container yet */
703 if (!(group_status.flags & VFIO_GROUP_FLAGS_CONTAINER_SET)) {
705 /* add group to a container */
706 ret = ioctl(vfio_group_fd, VFIO_GROUP_SET_CONTAINER,
709 RTE_LOG(ERR, EAL, " %s cannot add VFIO group to container, "
710 "error %i (%s)\n", dev_addr, errno, strerror(errno));
711 close(vfio_group_fd);
712 rte_vfio_clear_group(vfio_group_fd);
717 * pick an IOMMU type and set up DMA mappings for container
719 * needs to be done only once, only when first group is
720 * assigned to a container and only in primary process.
721 * Note this can happen several times with the hotplug
724 if (internal_config.process_type == RTE_PROC_PRIMARY &&
725 vfio_cfg->vfio_active_groups == 1 &&
726 vfio_group_device_count(vfio_group_fd) == 0) {
727 const struct vfio_iommu_type *t;
729 /* select an IOMMU type which we will be using */
730 t = vfio_set_iommu_type(vfio_container_fd);
733 " %s failed to select IOMMU type\n",
735 close(vfio_group_fd);
736 rte_vfio_clear_group(vfio_group_fd);
739 /* lock memory hotplug before mapping and release it
740 * after registering callback, to prevent races
742 rte_rwlock_read_lock(mem_lock);
743 if (vfio_cfg == default_vfio_cfg)
744 ret = t->dma_map_func(vfio_container_fd);
749 " %s DMA remapping failed, error %i (%s)\n",
750 dev_addr, errno, strerror(errno));
751 close(vfio_group_fd);
752 rte_vfio_clear_group(vfio_group_fd);
753 rte_rwlock_read_unlock(mem_lock);
757 vfio_cfg->vfio_iommu_type = t;
759 /* re-map all user-mapped segments */
760 rte_spinlock_recursive_lock(&user_mem_maps->lock);
762 /* this IOMMU type may not support DMA mapping, but
763 * if we have mappings in the list - that means we have
764 * previously mapped something successfully, so we can
765 * be sure that DMA mapping is supported.
767 for (i = 0; i < user_mem_maps->n_maps; i++) {
768 struct user_mem_map *map;
769 map = &user_mem_maps->maps[i];
771 ret = t->dma_user_map_func(
773 map->addr, map->iova, map->len,
776 RTE_LOG(ERR, EAL, "Couldn't map user memory for DMA: "
778 "iova: 0x%" PRIx64 " "
779 "len: 0x%" PRIu64 "\n",
780 map->addr, map->iova,
782 rte_spinlock_recursive_unlock(
783 &user_mem_maps->lock);
784 rte_rwlock_read_unlock(mem_lock);
788 rte_spinlock_recursive_unlock(&user_mem_maps->lock);
790 /* register callback for mem events */
791 if (vfio_cfg == default_vfio_cfg)
792 ret = rte_mem_event_callback_register(
793 VFIO_MEM_EVENT_CLB_NAME,
794 vfio_mem_event_callback, NULL);
797 /* unlock memory hotplug */
798 rte_rwlock_read_unlock(mem_lock);
800 if (ret && rte_errno != ENOTSUP) {
801 RTE_LOG(ERR, EAL, "Could not install memory event callback for VFIO\n");
805 RTE_LOG(DEBUG, EAL, "Memory event callbacks not supported\n");
807 RTE_LOG(DEBUG, EAL, "Installed memory event callback for VFIO\n");
809 } else if (rte_eal_process_type() != RTE_PROC_PRIMARY &&
810 vfio_cfg == default_vfio_cfg &&
811 vfio_cfg->vfio_iommu_type == NULL) {
812 /* if we're not a primary process, we do not set up the VFIO
813 * container because it's already been set up by the primary
814 * process. instead, we simply ask the primary about VFIO type
815 * we are using, and set the VFIO config up appropriately.
817 ret = vfio_sync_default_container();
819 RTE_LOG(ERR, EAL, "Could not sync default VFIO container\n");
820 close(vfio_group_fd);
821 rte_vfio_clear_group(vfio_group_fd);
824 /* we have successfully initialized VFIO, notify user */
825 const struct vfio_iommu_type *t =
826 default_vfio_cfg->vfio_iommu_type;
827 RTE_LOG(NOTICE, EAL, " using IOMMU type %d (%s)\n",
828 t->type_id, t->name);
831 /* get a file descriptor for the device */
832 *vfio_dev_fd = ioctl(vfio_group_fd, VFIO_GROUP_GET_DEVICE_FD, dev_addr);
833 if (*vfio_dev_fd < 0) {
834 /* if we cannot get a device fd, this implies a problem with
835 * the VFIO group or the container not having IOMMU configured.
838 RTE_LOG(WARNING, EAL, "Getting a vfio_dev_fd for %s failed\n",
840 close(vfio_group_fd);
841 rte_vfio_clear_group(vfio_group_fd);
845 /* test and setup the device */
846 ret = ioctl(*vfio_dev_fd, VFIO_DEVICE_GET_INFO, device_info);
848 RTE_LOG(ERR, EAL, " %s cannot get device info, "
849 "error %i (%s)\n", dev_addr, errno,
852 close(vfio_group_fd);
853 rte_vfio_clear_group(vfio_group_fd);
856 vfio_group_device_get(vfio_group_fd);
862 rte_vfio_release_device(const char *sysfs_base, const char *dev_addr,
865 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
866 rte_rwlock_t *mem_lock = &mcfg->memory_hotplug_lock;
867 struct vfio_group_status group_status = {
868 .argsz = sizeof(group_status)
870 struct vfio_config *vfio_cfg;
875 /* we don't want any DMA mapping messages to come while we're detaching
876 * VFIO device, because this might be the last device and we might need
877 * to unregister the callback.
879 rte_rwlock_read_lock(mem_lock);
881 /* get group number */
882 ret = rte_vfio_get_group_num(sysfs_base, dev_addr, &iommu_group_num);
884 RTE_LOG(WARNING, EAL, " %s not managed by VFIO driver\n",
886 /* This is an error at this point. */
891 /* get the actual group fd */
892 vfio_group_fd = rte_vfio_get_group_fd(iommu_group_num);
893 if (vfio_group_fd <= 0) {
894 RTE_LOG(INFO, EAL, "rte_vfio_get_group_fd failed for %s\n",
900 /* get the vfio_config it belongs to */
901 vfio_cfg = get_vfio_cfg_by_group_num(iommu_group_num);
902 vfio_cfg = vfio_cfg ? vfio_cfg : default_vfio_cfg;
904 /* At this point we got an active group. Closing it will make the
905 * container detachment. If this is the last active group, VFIO kernel
906 * code will unset the container and the IOMMU mappings.
909 /* Closing a device */
910 if (close(vfio_dev_fd) < 0) {
911 RTE_LOG(INFO, EAL, "Error when closing vfio_dev_fd for %s\n",
917 /* An VFIO group can have several devices attached. Just when there is
918 * no devices remaining should the group be closed.
920 vfio_group_device_put(vfio_group_fd);
921 if (!vfio_group_device_count(vfio_group_fd)) {
923 if (close(vfio_group_fd) < 0) {
924 RTE_LOG(INFO, EAL, "Error when closing vfio_group_fd for %s\n",
930 if (rte_vfio_clear_group(vfio_group_fd) < 0) {
931 RTE_LOG(INFO, EAL, "Error when clearing group for %s\n",
938 /* if there are no active device groups, unregister the callback to
939 * avoid spurious attempts to map/unmap memory from VFIO.
941 if (vfio_cfg == default_vfio_cfg && vfio_cfg->vfio_active_groups == 0 &&
942 rte_eal_process_type() != RTE_PROC_SECONDARY)
943 rte_mem_event_callback_unregister(VFIO_MEM_EVENT_CLB_NAME,
950 rte_rwlock_read_unlock(mem_lock);
955 rte_vfio_enable(const char *modname)
957 /* initialize group list */
961 rte_spinlock_recursive_t lock = RTE_SPINLOCK_RECURSIVE_INITIALIZER;
963 for (i = 0; i < VFIO_MAX_CONTAINERS; i++) {
964 vfio_cfgs[i].vfio_container_fd = -1;
965 vfio_cfgs[i].vfio_active_groups = 0;
966 vfio_cfgs[i].vfio_iommu_type = NULL;
967 vfio_cfgs[i].mem_maps.lock = lock;
969 for (j = 0; j < VFIO_MAX_GROUPS; j++) {
970 vfio_cfgs[i].vfio_groups[j].fd = -1;
971 vfio_cfgs[i].vfio_groups[j].group_num = -1;
972 vfio_cfgs[i].vfio_groups[j].devices = 0;
976 /* inform the user that we are probing for VFIO */
977 RTE_LOG(INFO, EAL, "Probing VFIO support...\n");
979 /* check if vfio module is loaded */
980 vfio_available = rte_eal_check_module(modname);
982 /* return error directly */
983 if (vfio_available == -1) {
984 RTE_LOG(INFO, EAL, "Could not get loaded module details!\n");
988 /* return 0 if VFIO modules not loaded */
989 if (vfio_available == 0) {
990 RTE_LOG(DEBUG, EAL, "VFIO modules not loaded, "
991 "skipping VFIO support...\n");
995 if (internal_config.process_type == RTE_PROC_PRIMARY) {
996 /* open a new container */
997 default_vfio_cfg->vfio_container_fd =
998 rte_vfio_get_container_fd();
1000 /* get the default container from the primary process */
1001 default_vfio_cfg->vfio_container_fd =
1002 vfio_get_default_container_fd();
1005 /* check if we have VFIO driver enabled */
1006 if (default_vfio_cfg->vfio_container_fd != -1) {
1007 RTE_LOG(NOTICE, EAL, "VFIO support initialized\n");
1008 default_vfio_cfg->vfio_enabled = 1;
1010 RTE_LOG(NOTICE, EAL, "VFIO support could not be initialized\n");
1017 rte_vfio_is_enabled(const char *modname)
1019 const int mod_available = rte_eal_check_module(modname) > 0;
1020 return default_vfio_cfg->vfio_enabled && mod_available;
1024 vfio_get_default_container_fd(void)
1026 struct rte_mp_msg mp_req, *mp_rep;
1027 struct rte_mp_reply mp_reply;
1028 struct timespec ts = {.tv_sec = 5, .tv_nsec = 0};
1029 struct vfio_mp_param *p = (struct vfio_mp_param *)mp_req.param;
1031 if (default_vfio_cfg->vfio_enabled)
1032 return default_vfio_cfg->vfio_container_fd;
1034 if (internal_config.process_type == RTE_PROC_PRIMARY) {
1035 /* if we were secondary process we would try requesting
1036 * container fd from the primary, but we're the primary
1037 * process so just exit here
1042 p->req = SOCKET_REQ_DEFAULT_CONTAINER;
1043 strcpy(mp_req.name, EAL_VFIO_MP);
1044 mp_req.len_param = sizeof(*p);
1047 if (rte_mp_request_sync(&mp_req, &mp_reply, &ts) == 0 &&
1048 mp_reply.nb_received == 1) {
1049 mp_rep = &mp_reply.msgs[0];
1050 p = (struct vfio_mp_param *)mp_rep->param;
1051 if (p->result == SOCKET_OK && mp_rep->num_fds == 1) {
1052 free(mp_reply.msgs);
1053 return mp_rep->fds[0];
1055 free(mp_reply.msgs);
1058 RTE_LOG(ERR, EAL, " cannot request default container fd\n");
1063 vfio_get_iommu_type(void)
1065 if (default_vfio_cfg->vfio_iommu_type == NULL)
1068 return default_vfio_cfg->vfio_iommu_type->type_id;
1071 const struct vfio_iommu_type *
1072 vfio_set_iommu_type(int vfio_container_fd)
1075 for (idx = 0; idx < RTE_DIM(iommu_types); idx++) {
1076 const struct vfio_iommu_type *t = &iommu_types[idx];
1078 int ret = ioctl(vfio_container_fd, VFIO_SET_IOMMU,
1081 RTE_LOG(NOTICE, EAL, " using IOMMU type %d (%s)\n",
1082 t->type_id, t->name);
1085 /* not an error, there may be more supported IOMMU types */
1086 RTE_LOG(DEBUG, EAL, " set IOMMU type %d (%s) failed, "
1087 "error %i (%s)\n", t->type_id, t->name, errno,
1090 /* if we didn't find a suitable IOMMU type, fail */
1095 vfio_has_supported_extensions(int vfio_container_fd)
1098 unsigned idx, n_extensions = 0;
1099 for (idx = 0; idx < RTE_DIM(iommu_types); idx++) {
1100 const struct vfio_iommu_type *t = &iommu_types[idx];
1102 ret = ioctl(vfio_container_fd, VFIO_CHECK_EXTENSION,
1105 RTE_LOG(ERR, EAL, " could not get IOMMU type, "
1106 "error %i (%s)\n", errno,
1108 close(vfio_container_fd);
1110 } else if (ret == 1) {
1111 /* we found a supported extension */
1114 RTE_LOG(DEBUG, EAL, " IOMMU type %d (%s) is %s\n",
1115 t->type_id, t->name,
1116 ret ? "supported" : "not supported");
1119 /* if we didn't find any supported IOMMU types, fail */
1120 if (!n_extensions) {
1121 close(vfio_container_fd);
1129 rte_vfio_get_container_fd(void)
1131 int ret, vfio_container_fd;
1132 struct rte_mp_msg mp_req, *mp_rep;
1133 struct rte_mp_reply mp_reply;
1134 struct timespec ts = {.tv_sec = 5, .tv_nsec = 0};
1135 struct vfio_mp_param *p = (struct vfio_mp_param *)mp_req.param;
1138 /* if we're in a primary process, try to open the container */
1139 if (internal_config.process_type == RTE_PROC_PRIMARY) {
1140 vfio_container_fd = open(VFIO_CONTAINER_PATH, O_RDWR);
1141 if (vfio_container_fd < 0) {
1142 RTE_LOG(ERR, EAL, " cannot open VFIO container, "
1143 "error %i (%s)\n", errno, strerror(errno));
1147 /* check VFIO API version */
1148 ret = ioctl(vfio_container_fd, VFIO_GET_API_VERSION);
1149 if (ret != VFIO_API_VERSION) {
1151 RTE_LOG(ERR, EAL, " could not get VFIO API version, "
1152 "error %i (%s)\n", errno, strerror(errno));
1154 RTE_LOG(ERR, EAL, " unsupported VFIO API version!\n");
1155 close(vfio_container_fd);
1159 ret = vfio_has_supported_extensions(vfio_container_fd);
1161 RTE_LOG(ERR, EAL, " no supported IOMMU "
1162 "extensions found!\n");
1166 return vfio_container_fd;
1169 * if we're in a secondary process, request container fd from the
1170 * primary process via mp channel
1172 p->req = SOCKET_REQ_CONTAINER;
1173 strcpy(mp_req.name, EAL_VFIO_MP);
1174 mp_req.len_param = sizeof(*p);
1177 vfio_container_fd = -1;
1178 if (rte_mp_request_sync(&mp_req, &mp_reply, &ts) == 0 &&
1179 mp_reply.nb_received == 1) {
1180 mp_rep = &mp_reply.msgs[0];
1181 p = (struct vfio_mp_param *)mp_rep->param;
1182 if (p->result == SOCKET_OK && mp_rep->num_fds == 1) {
1183 vfio_container_fd = mp_rep->fds[0];
1184 free(mp_reply.msgs);
1185 return vfio_container_fd;
1187 free(mp_reply.msgs);
1190 RTE_LOG(ERR, EAL, " cannot request container fd\n");
1195 rte_vfio_get_group_num(const char *sysfs_base,
1196 const char *dev_addr, int *iommu_group_num)
1198 char linkname[PATH_MAX];
1199 char filename[PATH_MAX];
1200 char *tok[16], *group_tok, *end;
1203 memset(linkname, 0, sizeof(linkname));
1204 memset(filename, 0, sizeof(filename));
1206 /* try to find out IOMMU group for this device */
1207 snprintf(linkname, sizeof(linkname),
1208 "%s/%s/iommu_group", sysfs_base, dev_addr);
1210 ret = readlink(linkname, filename, sizeof(filename));
1212 /* if the link doesn't exist, no VFIO for us */
1216 ret = rte_strsplit(filename, sizeof(filename),
1217 tok, RTE_DIM(tok), '/');
1220 RTE_LOG(ERR, EAL, " %s cannot get IOMMU group\n", dev_addr);
1224 /* IOMMU group is always the last token */
1226 group_tok = tok[ret - 1];
1228 *iommu_group_num = strtol(group_tok, &end, 10);
1229 if ((end != group_tok && *end != '\0') || errno != 0) {
1230 RTE_LOG(ERR, EAL, " %s error parsing IOMMU number!\n", dev_addr);
1238 type1_map(const struct rte_memseg_list *msl, const struct rte_memseg *ms,
1241 int *vfio_container_fd = arg;
1246 return vfio_type1_dma_mem_map(*vfio_container_fd, ms->addr_64, ms->iova,
1251 vfio_type1_dma_mem_map(int vfio_container_fd, uint64_t vaddr, uint64_t iova,
1252 uint64_t len, int do_map)
1254 struct vfio_iommu_type1_dma_map dma_map;
1255 struct vfio_iommu_type1_dma_unmap dma_unmap;
1259 memset(&dma_map, 0, sizeof(dma_map));
1260 dma_map.argsz = sizeof(struct vfio_iommu_type1_dma_map);
1261 dma_map.vaddr = vaddr;
1263 dma_map.iova = iova;
1264 dma_map.flags = VFIO_DMA_MAP_FLAG_READ |
1265 VFIO_DMA_MAP_FLAG_WRITE;
1267 ret = ioctl(vfio_container_fd, VFIO_IOMMU_MAP_DMA, &dma_map);
1270 * In case the mapping was already done EEXIST will be
1271 * returned from kernel.
1273 if (errno == EEXIST) {
1275 " Memory segment is allready mapped,"
1279 " cannot set up DMA remapping,"
1281 errno, strerror(errno));
1286 memset(&dma_unmap, 0, sizeof(dma_unmap));
1287 dma_unmap.argsz = sizeof(struct vfio_iommu_type1_dma_unmap);
1288 dma_unmap.size = len;
1289 dma_unmap.iova = iova;
1291 ret = ioctl(vfio_container_fd, VFIO_IOMMU_UNMAP_DMA,
1294 RTE_LOG(ERR, EAL, " cannot clear DMA remapping, error %i (%s)\n",
1295 errno, strerror(errno));
1304 vfio_type1_dma_map(int vfio_container_fd)
1306 return rte_memseg_walk(type1_map, &vfio_container_fd);
1310 vfio_spapr_dma_do_map(int vfio_container_fd, uint64_t vaddr, uint64_t iova,
1311 uint64_t len, int do_map)
1313 struct vfio_iommu_type1_dma_map dma_map;
1314 struct vfio_iommu_type1_dma_unmap dma_unmap;
1316 struct vfio_iommu_spapr_register_memory reg = {
1317 .argsz = sizeof(reg),
1320 reg.vaddr = (uintptr_t) vaddr;
1324 ret = ioctl(vfio_container_fd,
1325 VFIO_IOMMU_SPAPR_REGISTER_MEMORY, ®);
1327 RTE_LOG(ERR, EAL, " cannot register vaddr for IOMMU, "
1328 "error %i (%s)\n", errno, strerror(errno));
1332 memset(&dma_map, 0, sizeof(dma_map));
1333 dma_map.argsz = sizeof(struct vfio_iommu_type1_dma_map);
1334 dma_map.vaddr = vaddr;
1336 dma_map.iova = iova;
1337 dma_map.flags = VFIO_DMA_MAP_FLAG_READ |
1338 VFIO_DMA_MAP_FLAG_WRITE;
1340 ret = ioctl(vfio_container_fd, VFIO_IOMMU_MAP_DMA, &dma_map);
1343 * In case the mapping was already done EBUSY will be
1344 * returned from kernel.
1346 if (errno == EBUSY) {
1348 " Memory segment is allready mapped,"
1352 " cannot set up DMA remapping,"
1353 " error %i (%s)\n", errno,
1360 ret = ioctl(vfio_container_fd,
1361 VFIO_IOMMU_SPAPR_UNREGISTER_MEMORY, ®);
1363 RTE_LOG(ERR, EAL, " cannot unregister vaddr for IOMMU, error %i (%s)\n",
1364 errno, strerror(errno));
1368 memset(&dma_unmap, 0, sizeof(dma_unmap));
1369 dma_unmap.argsz = sizeof(struct vfio_iommu_type1_dma_unmap);
1370 dma_unmap.size = len;
1371 dma_unmap.iova = iova;
1373 ret = ioctl(vfio_container_fd, VFIO_IOMMU_UNMAP_DMA,
1376 RTE_LOG(ERR, EAL, " cannot clear DMA remapping, error %i (%s)\n",
1377 errno, strerror(errno));
1386 vfio_spapr_map_walk(const struct rte_memseg_list *msl,
1387 const struct rte_memseg *ms, void *arg)
1389 int *vfio_container_fd = arg;
1394 return vfio_spapr_dma_do_map(*vfio_container_fd, ms->addr_64, ms->iova,
1398 struct spapr_walk_param {
1399 uint64_t window_size;
1400 uint64_t hugepage_sz;
1403 vfio_spapr_window_size_walk(const struct rte_memseg_list *msl,
1404 const struct rte_memseg *ms, void *arg)
1406 struct spapr_walk_param *param = arg;
1407 uint64_t max = ms->iova + ms->len;
1412 if (max > param->window_size) {
1413 param->hugepage_sz = ms->hugepage_sz;
1414 param->window_size = max;
1421 vfio_spapr_create_new_dma_window(int vfio_container_fd,
1422 struct vfio_iommu_spapr_tce_create *create) {
1423 struct vfio_iommu_spapr_tce_remove remove = {
1424 .argsz = sizeof(remove),
1426 struct vfio_iommu_spapr_tce_info info = {
1427 .argsz = sizeof(info),
1431 /* query spapr iommu info */
1432 ret = ioctl(vfio_container_fd, VFIO_IOMMU_SPAPR_TCE_GET_INFO, &info);
1434 RTE_LOG(ERR, EAL, " cannot get iommu info, "
1435 "error %i (%s)\n", errno, strerror(errno));
1439 /* remove default DMA of 32 bit window */
1440 remove.start_addr = info.dma32_window_start;
1441 ret = ioctl(vfio_container_fd, VFIO_IOMMU_SPAPR_TCE_REMOVE, &remove);
1443 RTE_LOG(ERR, EAL, " cannot remove default DMA window, "
1444 "error %i (%s)\n", errno, strerror(errno));
1448 /* create new DMA window */
1449 ret = ioctl(vfio_container_fd, VFIO_IOMMU_SPAPR_TCE_CREATE, create);
1451 RTE_LOG(ERR, EAL, " cannot create new DMA window, "
1452 "error %i (%s)\n", errno, strerror(errno));
1456 if (create->start_addr != 0) {
1457 RTE_LOG(ERR, EAL, " DMA window start address != 0\n");
1465 vfio_spapr_dma_mem_map(int vfio_container_fd, uint64_t vaddr, uint64_t iova,
1466 uint64_t len, int do_map)
1468 struct spapr_walk_param param;
1469 struct vfio_iommu_spapr_tce_create create = {
1470 .argsz = sizeof(create),
1472 struct vfio_config *vfio_cfg;
1473 struct user_mem_maps *user_mem_maps;
1476 vfio_cfg = get_vfio_cfg_by_container_fd(vfio_container_fd);
1477 if (vfio_cfg == NULL) {
1478 RTE_LOG(ERR, EAL, " invalid container fd!\n");
1482 user_mem_maps = &vfio_cfg->mem_maps;
1483 rte_spinlock_recursive_lock(&user_mem_maps->lock);
1485 /* check if window size needs to be adjusted */
1486 memset(¶m, 0, sizeof(param));
1488 /* we're inside a callback so use thread-unsafe version */
1489 if (rte_memseg_walk_thread_unsafe(vfio_spapr_window_size_walk,
1491 RTE_LOG(ERR, EAL, "Could not get window size\n");
1496 /* also check user maps */
1497 for (i = 0; i < user_mem_maps->n_maps; i++) {
1498 uint64_t max = user_mem_maps->maps[i].iova +
1499 user_mem_maps->maps[i].len;
1500 create.window_size = RTE_MAX(create.window_size, max);
1503 /* sPAPR requires window size to be a power of 2 */
1504 create.window_size = rte_align64pow2(param.window_size);
1505 create.page_shift = __builtin_ctzll(param.hugepage_sz);
1510 /* re-create window and remap the entire memory */
1511 if (iova > create.window_size) {
1512 if (vfio_spapr_create_new_dma_window(vfio_container_fd,
1514 RTE_LOG(ERR, EAL, "Could not create new DMA window\n");
1518 /* we're inside a callback, so use thread-unsafe version
1520 if (rte_memseg_walk_thread_unsafe(vfio_spapr_map_walk,
1521 &vfio_container_fd) < 0) {
1522 RTE_LOG(ERR, EAL, "Could not recreate DMA maps\n");
1526 /* remap all user maps */
1527 for (i = 0; i < user_mem_maps->n_maps; i++) {
1528 struct user_mem_map *map =
1529 &user_mem_maps->maps[i];
1530 if (vfio_spapr_dma_do_map(vfio_container_fd,
1531 map->addr, map->iova, map->len,
1533 RTE_LOG(ERR, EAL, "Could not recreate user DMA maps\n");
1540 /* now that we've remapped all of the memory that was present
1541 * before, map the segment that we were requested to map.
1543 * however, if we were called by the callback, the memory we
1544 * were called with was already in the memseg list, so previous
1545 * mapping should've mapped that segment already.
1547 * virt2memseg_list is a relatively cheap check, so use that. if
1548 * memory is within any memseg list, it's a memseg, so it's
1551 addr = (void *)(uintptr_t)vaddr;
1552 if (rte_mem_virt2memseg_list(addr) == NULL &&
1553 vfio_spapr_dma_do_map(vfio_container_fd,
1554 vaddr, iova, len, 1) < 0) {
1555 RTE_LOG(ERR, EAL, "Could not map segment\n");
1560 /* for unmap, check if iova within DMA window */
1561 if (iova > create.window_size) {
1562 RTE_LOG(ERR, EAL, "iova beyond DMA window for unmap");
1567 vfio_spapr_dma_do_map(vfio_container_fd, vaddr, iova, len, 0);
1570 rte_spinlock_recursive_unlock(&user_mem_maps->lock);
1575 vfio_spapr_dma_map(int vfio_container_fd)
1577 struct vfio_iommu_spapr_tce_create create = {
1578 .argsz = sizeof(create),
1580 struct spapr_walk_param param;
1582 memset(¶m, 0, sizeof(param));
1584 /* create DMA window from 0 to max(phys_addr + len) */
1585 rte_memseg_walk(vfio_spapr_window_size_walk, ¶m);
1587 /* sPAPR requires window size to be a power of 2 */
1588 create.window_size = rte_align64pow2(param.window_size);
1589 create.page_shift = __builtin_ctzll(param.hugepage_sz);
1592 if (vfio_spapr_create_new_dma_window(vfio_container_fd, &create) < 0) {
1593 RTE_LOG(ERR, EAL, "Could not create new DMA window\n");
1597 /* map all DPDK segments for DMA. use 1:1 PA to IOVA mapping */
1598 if (rte_memseg_walk(vfio_spapr_map_walk, &vfio_container_fd) < 0)
1605 vfio_noiommu_dma_map(int __rte_unused vfio_container_fd)
1607 /* No-IOMMU mode does not need DMA mapping */
1612 vfio_noiommu_dma_mem_map(int __rte_unused vfio_container_fd,
1613 uint64_t __rte_unused vaddr,
1614 uint64_t __rte_unused iova, uint64_t __rte_unused len,
1615 int __rte_unused do_map)
1617 /* No-IOMMU mode does not need DMA mapping */
1622 vfio_dma_mem_map(struct vfio_config *vfio_cfg, uint64_t vaddr, uint64_t iova,
1623 uint64_t len, int do_map)
1625 const struct vfio_iommu_type *t = vfio_cfg->vfio_iommu_type;
1628 RTE_LOG(ERR, EAL, " VFIO support not initialized\n");
1633 if (!t->dma_user_map_func) {
1635 " VFIO custom DMA region maping not supported by IOMMU %s\n",
1637 rte_errno = ENOTSUP;
1641 return t->dma_user_map_func(vfio_cfg->vfio_container_fd, vaddr, iova,
1646 container_dma_map(struct vfio_config *vfio_cfg, uint64_t vaddr, uint64_t iova,
1649 struct user_mem_map *new_map;
1650 struct user_mem_maps *user_mem_maps;
1653 user_mem_maps = &vfio_cfg->mem_maps;
1654 rte_spinlock_recursive_lock(&user_mem_maps->lock);
1655 if (user_mem_maps->n_maps == VFIO_MAX_USER_MEM_MAPS) {
1656 RTE_LOG(ERR, EAL, "No more space for user mem maps\n");
1662 if (vfio_dma_mem_map(vfio_cfg, vaddr, iova, len, 1)) {
1663 /* technically, this will fail if there are currently no devices
1664 * plugged in, even if a device were added later, this mapping
1665 * might have succeeded. however, since we cannot verify if this
1666 * is a valid mapping without having a device attached, consider
1667 * this to be unsupported, because we can't just store any old
1668 * mapping and pollute list of active mappings willy-nilly.
1670 RTE_LOG(ERR, EAL, "Couldn't map new region for DMA\n");
1674 /* create new user mem map entry */
1675 new_map = &user_mem_maps->maps[user_mem_maps->n_maps++];
1676 new_map->addr = vaddr;
1677 new_map->iova = iova;
1680 compact_user_maps(user_mem_maps);
1682 rte_spinlock_recursive_unlock(&user_mem_maps->lock);
1687 container_dma_unmap(struct vfio_config *vfio_cfg, uint64_t vaddr, uint64_t iova,
1690 struct user_mem_map *map, *new_map = NULL;
1691 struct user_mem_maps *user_mem_maps;
1694 user_mem_maps = &vfio_cfg->mem_maps;
1695 rte_spinlock_recursive_lock(&user_mem_maps->lock);
1697 /* find our mapping */
1698 map = find_user_mem_map(user_mem_maps, vaddr, iova, len);
1700 RTE_LOG(ERR, EAL, "Couldn't find previously mapped region\n");
1705 if (map->addr != vaddr || map->iova != iova || map->len != len) {
1706 /* we're partially unmapping a previously mapped region, so we
1707 * need to split entry into two.
1709 if (user_mem_maps->n_maps == VFIO_MAX_USER_MEM_MAPS) {
1710 RTE_LOG(ERR, EAL, "Not enough space to store partial mapping\n");
1715 new_map = &user_mem_maps->maps[user_mem_maps->n_maps++];
1718 /* unmap the entry */
1719 if (vfio_dma_mem_map(vfio_cfg, vaddr, iova, len, 0)) {
1720 /* there may not be any devices plugged in, so unmapping will
1721 * fail with ENODEV/ENOTSUP rte_errno values, but that doesn't
1722 * stop us from removing the mapping, as the assumption is we
1723 * won't be needing this memory any more and thus will want to
1724 * prevent it from being remapped again on hotplug. so, only
1725 * fail if we indeed failed to unmap (e.g. if the mapping was
1726 * within our mapped range but had invalid alignment).
1728 if (rte_errno != ENODEV && rte_errno != ENOTSUP) {
1729 RTE_LOG(ERR, EAL, "Couldn't unmap region for DMA\n");
1733 RTE_LOG(DEBUG, EAL, "DMA unmapping failed, but removing mappings anyway\n");
1736 /* remove map from the list of active mappings */
1737 if (new_map != NULL) {
1738 adjust_map(map, new_map, vaddr, len);
1740 /* if we've created a new map by splitting, sort everything */
1741 if (!is_null_map(new_map)) {
1742 compact_user_maps(user_mem_maps);
1744 /* we've created a new mapping, but it was unused */
1745 user_mem_maps->n_maps--;
1748 memset(map, 0, sizeof(*map));
1749 compact_user_maps(user_mem_maps);
1750 user_mem_maps->n_maps--;
1754 rte_spinlock_recursive_unlock(&user_mem_maps->lock);
1759 rte_vfio_dma_map(uint64_t vaddr, uint64_t iova, uint64_t len)
1766 return container_dma_map(default_vfio_cfg, vaddr, iova, len);
1770 rte_vfio_dma_unmap(uint64_t vaddr, uint64_t iova, uint64_t len)
1777 return container_dma_unmap(default_vfio_cfg, vaddr, iova, len);
1781 rte_vfio_noiommu_is_enabled(void)
1787 fd = open(VFIO_NOIOMMU_MODE, O_RDONLY);
1789 if (errno != ENOENT) {
1790 RTE_LOG(ERR, EAL, " cannot open vfio noiommu file %i (%s)\n",
1791 errno, strerror(errno));
1795 * else the file does not exists
1796 * i.e. noiommu is not enabled
1801 cnt = read(fd, &c, 1);
1804 RTE_LOG(ERR, EAL, " unable to read from vfio noiommu "
1805 "file %i (%s)\n", errno, strerror(errno));
1813 rte_vfio_container_create(void)
1817 /* Find an empty slot to store new vfio config */
1818 for (i = 1; i < VFIO_MAX_CONTAINERS; i++) {
1819 if (vfio_cfgs[i].vfio_container_fd == -1)
1823 if (i == VFIO_MAX_CONTAINERS) {
1824 RTE_LOG(ERR, EAL, "exceed max vfio container limit\n");
1828 vfio_cfgs[i].vfio_container_fd = rte_vfio_get_container_fd();
1829 if (vfio_cfgs[i].vfio_container_fd < 0) {
1830 RTE_LOG(NOTICE, EAL, "fail to create a new container\n");
1834 return vfio_cfgs[i].vfio_container_fd;
1837 int __rte_experimental
1838 rte_vfio_container_destroy(int container_fd)
1840 struct vfio_config *vfio_cfg;
1843 vfio_cfg = get_vfio_cfg_by_container_fd(container_fd);
1844 if (vfio_cfg == NULL) {
1845 RTE_LOG(ERR, EAL, "Invalid container fd\n");
1849 for (i = 0; i < VFIO_MAX_GROUPS; i++)
1850 if (vfio_cfg->vfio_groups[i].group_num != -1)
1851 rte_vfio_container_group_unbind(container_fd,
1852 vfio_cfg->vfio_groups[i].group_num);
1854 close(container_fd);
1855 vfio_cfg->vfio_container_fd = -1;
1856 vfio_cfg->vfio_active_groups = 0;
1857 vfio_cfg->vfio_iommu_type = NULL;
1863 rte_vfio_container_group_bind(int container_fd, int iommu_group_num)
1865 struct vfio_config *vfio_cfg;
1867 vfio_cfg = get_vfio_cfg_by_container_fd(container_fd);
1868 if (vfio_cfg == NULL) {
1869 RTE_LOG(ERR, EAL, "Invalid container fd\n");
1873 return vfio_get_group_fd(vfio_cfg, iommu_group_num);
1877 rte_vfio_container_group_unbind(int container_fd, int iommu_group_num)
1879 struct vfio_config *vfio_cfg;
1880 struct vfio_group *cur_grp = NULL;
1883 vfio_cfg = get_vfio_cfg_by_container_fd(container_fd);
1884 if (vfio_cfg == NULL) {
1885 RTE_LOG(ERR, EAL, "Invalid container fd\n");
1889 for (i = 0; i < VFIO_MAX_GROUPS; i++) {
1890 if (vfio_cfg->vfio_groups[i].group_num == iommu_group_num) {
1891 cur_grp = &vfio_cfg->vfio_groups[i];
1896 /* This should not happen */
1897 if (i == VFIO_MAX_GROUPS || cur_grp == NULL) {
1898 RTE_LOG(ERR, EAL, "Specified group number not found\n");
1902 if (cur_grp->fd >= 0 && close(cur_grp->fd) < 0) {
1903 RTE_LOG(ERR, EAL, "Error when closing vfio_group_fd for"
1904 " iommu_group_num %d\n", iommu_group_num);
1907 cur_grp->group_num = -1;
1909 cur_grp->devices = 0;
1910 vfio_cfg->vfio_active_groups--;
1916 rte_vfio_container_dma_map(int container_fd, uint64_t vaddr, uint64_t iova,
1919 struct vfio_config *vfio_cfg;
1926 vfio_cfg = get_vfio_cfg_by_container_fd(container_fd);
1927 if (vfio_cfg == NULL) {
1928 RTE_LOG(ERR, EAL, "Invalid container fd\n");
1932 return container_dma_map(vfio_cfg, vaddr, iova, len);
1936 rte_vfio_container_dma_unmap(int container_fd, uint64_t vaddr, uint64_t iova,
1939 struct vfio_config *vfio_cfg;
1946 vfio_cfg = get_vfio_cfg_by_container_fd(container_fd);
1947 if (vfio_cfg == NULL) {
1948 RTE_LOG(ERR, EAL, "Invalid container fd\n");
1952 return container_dma_unmap(vfio_cfg, vaddr, iova, len);
1958 rte_vfio_dma_map(uint64_t __rte_unused vaddr, __rte_unused uint64_t iova,
1959 __rte_unused uint64_t len)
1965 rte_vfio_dma_unmap(uint64_t __rte_unused vaddr, uint64_t __rte_unused iova,
1966 __rte_unused uint64_t len)
1972 rte_vfio_setup_device(__rte_unused const char *sysfs_base,
1973 __rte_unused const char *dev_addr,
1974 __rte_unused int *vfio_dev_fd,
1975 __rte_unused struct vfio_device_info *device_info)
1981 rte_vfio_release_device(__rte_unused const char *sysfs_base,
1982 __rte_unused const char *dev_addr, __rte_unused int fd)
1988 rte_vfio_enable(__rte_unused const char *modname)
1994 rte_vfio_is_enabled(__rte_unused const char *modname)
2000 rte_vfio_noiommu_is_enabled(void)
2006 rte_vfio_clear_group(__rte_unused int vfio_group_fd)
2012 rte_vfio_get_group_num(__rte_unused const char *sysfs_base,
2013 __rte_unused const char *dev_addr,
2014 __rte_unused int *iommu_group_num)
2020 rte_vfio_get_container_fd(void)
2026 rte_vfio_get_group_fd(__rte_unused int iommu_group_num)
2032 rte_vfio_container_create(void)
2038 rte_vfio_container_destroy(__rte_unused int container_fd)
2044 rte_vfio_container_group_bind(__rte_unused int container_fd,
2045 __rte_unused int iommu_group_num)
2051 rte_vfio_container_group_unbind(__rte_unused int container_fd,
2052 __rte_unused int iommu_group_num)
2058 rte_vfio_container_dma_map(__rte_unused int container_fd,
2059 __rte_unused uint64_t vaddr,
2060 __rte_unused uint64_t iova,
2061 __rte_unused uint64_t len)
2067 rte_vfio_container_dma_unmap(__rte_unused int container_fd,
2068 __rte_unused uint64_t vaddr,
2069 __rte_unused uint64_t iova,
2070 __rte_unused uint64_t len)
2075 #endif /* VFIO_PRESENT */