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 ==
348 static struct vfio_config *
349 get_vfio_cfg_by_group_fd(int vfio_group_fd)
351 struct vfio_config *vfio_cfg;
354 for (i = 0; i < VFIO_MAX_CONTAINERS; i++) {
355 vfio_cfg = &vfio_cfgs[i];
356 for (j = 0; j < VFIO_MAX_GROUPS; j++)
357 if (vfio_cfg->vfio_groups[j].fd == vfio_group_fd)
364 static struct vfio_config *
365 get_vfio_cfg_by_container_fd(int container_fd)
369 for (i = 0; i < VFIO_MAX_CONTAINERS; i++) {
370 if (vfio_cfgs[i].vfio_container_fd == container_fd)
371 return &vfio_cfgs[i];
378 rte_vfio_get_group_fd(int iommu_group_num)
382 struct vfio_group *cur_grp;
383 struct vfio_config *vfio_cfg;
385 /* get the vfio_config it belongs to */
386 vfio_cfg = get_vfio_cfg_by_group_num(iommu_group_num);
387 vfio_cfg = vfio_cfg ? vfio_cfg : default_vfio_cfg;
389 /* check if we already have the group descriptor open */
390 for (i = 0; i < VFIO_MAX_GROUPS; i++)
391 if (vfio_cfg->vfio_groups[i].group_num == iommu_group_num)
392 return vfio_cfg->vfio_groups[i].fd;
394 /* Lets see first if there is room for a new group */
395 if (vfio_cfg->vfio_active_groups == VFIO_MAX_GROUPS) {
396 RTE_LOG(ERR, EAL, "Maximum number of VFIO groups reached!\n");
400 /* Now lets get an index for the new group */
401 for (i = 0; i < VFIO_MAX_GROUPS; i++)
402 if (vfio_cfg->vfio_groups[i].group_num == -1) {
403 cur_grp = &vfio_cfg->vfio_groups[i];
407 /* This should not happen */
408 if (i == VFIO_MAX_GROUPS) {
409 RTE_LOG(ERR, EAL, "No VFIO group free slot found\n");
413 vfio_group_fd = vfio_open_group_fd(iommu_group_num);
414 if (vfio_group_fd < 0) {
415 RTE_LOG(ERR, EAL, "Failed to open group %d\n", iommu_group_num);
419 cur_grp->group_num = iommu_group_num;
420 cur_grp->fd = vfio_group_fd;
421 vfio_cfg->vfio_active_groups++;
423 return vfio_group_fd;
427 get_vfio_group_idx(int vfio_group_fd)
429 struct vfio_config *vfio_cfg;
432 for (i = 0; i < VFIO_MAX_CONTAINERS; i++) {
433 vfio_cfg = &vfio_cfgs[i];
434 for (j = 0; j < VFIO_MAX_GROUPS; j++)
435 if (vfio_cfg->vfio_groups[j].fd == vfio_group_fd)
443 vfio_group_device_get(int vfio_group_fd)
445 struct vfio_config *vfio_cfg;
448 vfio_cfg = get_vfio_cfg_by_group_fd(vfio_group_fd);
449 if (vfio_cfg == NULL) {
450 RTE_LOG(ERR, EAL, " invalid group fd!\n");
454 i = get_vfio_group_idx(vfio_group_fd);
455 if (i < 0 || i > (VFIO_MAX_GROUPS - 1))
456 RTE_LOG(ERR, EAL, " wrong vfio_group index (%d)\n", i);
458 vfio_cfg->vfio_groups[i].devices++;
462 vfio_group_device_put(int vfio_group_fd)
464 struct vfio_config *vfio_cfg;
467 vfio_cfg = get_vfio_cfg_by_group_fd(vfio_group_fd);
468 if (vfio_cfg == NULL) {
469 RTE_LOG(ERR, EAL, " invalid group fd!\n");
473 i = get_vfio_group_idx(vfio_group_fd);
474 if (i < 0 || i > (VFIO_MAX_GROUPS - 1))
475 RTE_LOG(ERR, EAL, " wrong vfio_group index (%d)\n", i);
477 vfio_cfg->vfio_groups[i].devices--;
481 vfio_group_device_count(int vfio_group_fd)
483 struct vfio_config *vfio_cfg;
486 vfio_cfg = get_vfio_cfg_by_group_fd(vfio_group_fd);
487 if (vfio_cfg == NULL) {
488 RTE_LOG(ERR, EAL, " invalid group fd!\n");
492 i = get_vfio_group_idx(vfio_group_fd);
493 if (i < 0 || i > (VFIO_MAX_GROUPS - 1)) {
494 RTE_LOG(ERR, EAL, " wrong vfio_group index (%d)\n", i);
498 return vfio_cfg->vfio_groups[i].devices;
502 vfio_mem_event_callback(enum rte_mem_event type, const void *addr, size_t len,
503 void *arg __rte_unused)
505 struct rte_memseg_list *msl;
506 struct rte_memseg *ms;
509 msl = rte_mem_virt2memseg_list(addr);
511 /* for IOVA as VA mode, no need to care for IOVA addresses */
512 if (rte_eal_iova_mode() == RTE_IOVA_VA && msl->external == 0) {
513 uint64_t vfio_va = (uint64_t)(uintptr_t)addr;
514 if (type == RTE_MEM_EVENT_ALLOC)
515 vfio_dma_mem_map(default_vfio_cfg, vfio_va, vfio_va,
518 vfio_dma_mem_map(default_vfio_cfg, vfio_va, vfio_va,
523 /* memsegs are contiguous in memory */
524 ms = rte_mem_virt2memseg(addr, msl);
525 while (cur_len < len) {
526 /* some memory segments may have invalid IOVA */
527 if (ms->iova == RTE_BAD_IOVA) {
528 RTE_LOG(DEBUG, EAL, "Memory segment at %p has bad IOVA, skipping\n",
532 if (type == RTE_MEM_EVENT_ALLOC)
533 vfio_dma_mem_map(default_vfio_cfg, ms->addr_64,
534 ms->iova, ms->len, 1);
536 vfio_dma_mem_map(default_vfio_cfg, ms->addr_64,
537 ms->iova, ms->len, 0);
545 rte_vfio_clear_group(int vfio_group_fd)
548 struct vfio_config *vfio_cfg;
550 vfio_cfg = get_vfio_cfg_by_group_fd(vfio_group_fd);
551 if (vfio_cfg == NULL) {
552 RTE_LOG(ERR, EAL, " invalid group fd!\n");
556 i = get_vfio_group_idx(vfio_group_fd);
559 vfio_cfg->vfio_groups[i].group_num = -1;
560 vfio_cfg->vfio_groups[i].fd = -1;
561 vfio_cfg->vfio_groups[i].devices = 0;
562 vfio_cfg->vfio_active_groups--;
568 rte_vfio_setup_device(const char *sysfs_base, const char *dev_addr,
569 int *vfio_dev_fd, struct vfio_device_info *device_info)
571 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
572 rte_rwlock_t *mem_lock = &mcfg->memory_hotplug_lock;
573 struct vfio_group_status group_status = {
574 .argsz = sizeof(group_status)
576 struct vfio_config *vfio_cfg;
577 struct user_mem_maps *user_mem_maps;
578 int vfio_container_fd;
583 /* get group number */
584 ret = rte_vfio_get_group_num(sysfs_base, dev_addr, &iommu_group_num);
586 RTE_LOG(WARNING, EAL, " %s not managed by VFIO driver, skipping\n",
591 /* if negative, something failed */
595 /* get the actual group fd */
596 vfio_group_fd = rte_vfio_get_group_fd(iommu_group_num);
597 if (vfio_group_fd < 0)
600 /* if group_fd == 0, that means the device isn't managed by VFIO */
601 if (vfio_group_fd == 0) {
602 RTE_LOG(WARNING, EAL, " %s not managed by VFIO driver, skipping\n",
608 * at this point, we know that this group is viable (meaning, all devices
609 * are either bound to VFIO or not bound to anything)
612 /* check if the group is viable */
613 ret = ioctl(vfio_group_fd, VFIO_GROUP_GET_STATUS, &group_status);
615 RTE_LOG(ERR, EAL, " %s cannot get group status, "
616 "error %i (%s)\n", dev_addr, errno, strerror(errno));
617 close(vfio_group_fd);
618 rte_vfio_clear_group(vfio_group_fd);
620 } else if (!(group_status.flags & VFIO_GROUP_FLAGS_VIABLE)) {
621 RTE_LOG(ERR, EAL, " %s VFIO group is not viable!\n", dev_addr);
622 close(vfio_group_fd);
623 rte_vfio_clear_group(vfio_group_fd);
627 /* get the vfio_config it belongs to */
628 vfio_cfg = get_vfio_cfg_by_group_num(iommu_group_num);
629 vfio_cfg = vfio_cfg ? vfio_cfg : default_vfio_cfg;
630 vfio_container_fd = vfio_cfg->vfio_container_fd;
631 user_mem_maps = &vfio_cfg->mem_maps;
633 /* check if group does not have a container yet */
634 if (!(group_status.flags & VFIO_GROUP_FLAGS_CONTAINER_SET)) {
636 /* add group to a container */
637 ret = ioctl(vfio_group_fd, VFIO_GROUP_SET_CONTAINER,
640 RTE_LOG(ERR, EAL, " %s cannot add VFIO group to container, "
641 "error %i (%s)\n", dev_addr, errno, strerror(errno));
642 close(vfio_group_fd);
643 rte_vfio_clear_group(vfio_group_fd);
648 * pick an IOMMU type and set up DMA mappings for container
650 * needs to be done only once, only when first group is
651 * assigned to a container and only in primary process.
652 * Note this can happen several times with the hotplug
655 if (internal_config.process_type == RTE_PROC_PRIMARY &&
656 vfio_cfg->vfio_active_groups == 1 &&
657 vfio_group_device_count(vfio_group_fd) == 0) {
658 const struct vfio_iommu_type *t;
660 /* select an IOMMU type which we will be using */
661 t = vfio_set_iommu_type(vfio_container_fd);
664 " %s failed to select IOMMU type\n",
666 close(vfio_group_fd);
667 rte_vfio_clear_group(vfio_group_fd);
670 /* lock memory hotplug before mapping and release it
671 * after registering callback, to prevent races
673 rte_rwlock_read_lock(mem_lock);
674 if (vfio_cfg == default_vfio_cfg)
675 ret = t->dma_map_func(vfio_container_fd);
680 " %s DMA remapping failed, error %i (%s)\n",
681 dev_addr, errno, strerror(errno));
682 close(vfio_group_fd);
683 rte_vfio_clear_group(vfio_group_fd);
684 rte_rwlock_read_unlock(mem_lock);
688 vfio_cfg->vfio_iommu_type = t;
690 /* re-map all user-mapped segments */
691 rte_spinlock_recursive_lock(&user_mem_maps->lock);
693 /* this IOMMU type may not support DMA mapping, but
694 * if we have mappings in the list - that means we have
695 * previously mapped something successfully, so we can
696 * be sure that DMA mapping is supported.
698 for (i = 0; i < user_mem_maps->n_maps; i++) {
699 struct user_mem_map *map;
700 map = &user_mem_maps->maps[i];
702 ret = t->dma_user_map_func(
704 map->addr, map->iova, map->len,
707 RTE_LOG(ERR, EAL, "Couldn't map user memory for DMA: "
709 "iova: 0x%" PRIx64 " "
710 "len: 0x%" PRIu64 "\n",
711 map->addr, map->iova,
713 rte_spinlock_recursive_unlock(
714 &user_mem_maps->lock);
715 rte_rwlock_read_unlock(mem_lock);
719 rte_spinlock_recursive_unlock(&user_mem_maps->lock);
721 /* register callback for mem events */
722 if (vfio_cfg == default_vfio_cfg)
723 ret = rte_mem_event_callback_register(
724 VFIO_MEM_EVENT_CLB_NAME,
725 vfio_mem_event_callback, NULL);
728 /* unlock memory hotplug */
729 rte_rwlock_read_unlock(mem_lock);
731 if (ret && rte_errno != ENOTSUP) {
732 RTE_LOG(ERR, EAL, "Could not install memory event callback for VFIO\n");
736 RTE_LOG(DEBUG, EAL, "Memory event callbacks not supported\n");
738 RTE_LOG(DEBUG, EAL, "Installed memory event callback for VFIO\n");
742 /* get a file descriptor for the device */
743 *vfio_dev_fd = ioctl(vfio_group_fd, VFIO_GROUP_GET_DEVICE_FD, dev_addr);
744 if (*vfio_dev_fd < 0) {
745 /* if we cannot get a device fd, this implies a problem with
746 * the VFIO group or the container not having IOMMU configured.
749 RTE_LOG(WARNING, EAL, "Getting a vfio_dev_fd for %s failed\n",
751 close(vfio_group_fd);
752 rte_vfio_clear_group(vfio_group_fd);
756 /* test and setup the device */
757 ret = ioctl(*vfio_dev_fd, VFIO_DEVICE_GET_INFO, device_info);
759 RTE_LOG(ERR, EAL, " %s cannot get device info, "
760 "error %i (%s)\n", dev_addr, errno,
763 close(vfio_group_fd);
764 rte_vfio_clear_group(vfio_group_fd);
767 vfio_group_device_get(vfio_group_fd);
773 rte_vfio_release_device(const char *sysfs_base, const char *dev_addr,
776 struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
777 rte_rwlock_t *mem_lock = &mcfg->memory_hotplug_lock;
778 struct vfio_group_status group_status = {
779 .argsz = sizeof(group_status)
781 struct vfio_config *vfio_cfg;
786 /* we don't want any DMA mapping messages to come while we're detaching
787 * VFIO device, because this might be the last device and we might need
788 * to unregister the callback.
790 rte_rwlock_read_lock(mem_lock);
792 /* get group number */
793 ret = rte_vfio_get_group_num(sysfs_base, dev_addr, &iommu_group_num);
795 RTE_LOG(WARNING, EAL, " %s not managed by VFIO driver\n",
797 /* This is an error at this point. */
802 /* get the actual group fd */
803 vfio_group_fd = rte_vfio_get_group_fd(iommu_group_num);
804 if (vfio_group_fd <= 0) {
805 RTE_LOG(INFO, EAL, "rte_vfio_get_group_fd failed for %s\n",
811 /* get the vfio_config it belongs to */
812 vfio_cfg = get_vfio_cfg_by_group_num(iommu_group_num);
813 vfio_cfg = vfio_cfg ? vfio_cfg : default_vfio_cfg;
815 /* At this point we got an active group. Closing it will make the
816 * container detachment. If this is the last active group, VFIO kernel
817 * code will unset the container and the IOMMU mappings.
820 /* Closing a device */
821 if (close(vfio_dev_fd) < 0) {
822 RTE_LOG(INFO, EAL, "Error when closing vfio_dev_fd for %s\n",
828 /* An VFIO group can have several devices attached. Just when there is
829 * no devices remaining should the group be closed.
831 vfio_group_device_put(vfio_group_fd);
832 if (!vfio_group_device_count(vfio_group_fd)) {
834 if (close(vfio_group_fd) < 0) {
835 RTE_LOG(INFO, EAL, "Error when closing vfio_group_fd for %s\n",
841 if (rte_vfio_clear_group(vfio_group_fd) < 0) {
842 RTE_LOG(INFO, EAL, "Error when clearing group for %s\n",
849 /* if there are no active device groups, unregister the callback to
850 * avoid spurious attempts to map/unmap memory from VFIO.
852 if (vfio_cfg == default_vfio_cfg && vfio_cfg->vfio_active_groups == 0)
853 rte_mem_event_callback_unregister(VFIO_MEM_EVENT_CLB_NAME,
860 rte_rwlock_read_unlock(mem_lock);
865 rte_vfio_enable(const char *modname)
867 /* initialize group list */
871 rte_spinlock_recursive_t lock = RTE_SPINLOCK_RECURSIVE_INITIALIZER;
873 for (i = 0; i < VFIO_MAX_CONTAINERS; i++) {
874 vfio_cfgs[i].vfio_container_fd = -1;
875 vfio_cfgs[i].vfio_active_groups = 0;
876 vfio_cfgs[i].vfio_iommu_type = NULL;
877 vfio_cfgs[i].mem_maps.lock = lock;
879 for (j = 0; j < VFIO_MAX_GROUPS; j++) {
880 vfio_cfgs[i].vfio_groups[j].fd = -1;
881 vfio_cfgs[i].vfio_groups[j].group_num = -1;
882 vfio_cfgs[i].vfio_groups[j].devices = 0;
886 /* inform the user that we are probing for VFIO */
887 RTE_LOG(INFO, EAL, "Probing VFIO support...\n");
889 /* check if vfio module is loaded */
890 vfio_available = rte_eal_check_module(modname);
892 /* return error directly */
893 if (vfio_available == -1) {
894 RTE_LOG(INFO, EAL, "Could not get loaded module details!\n");
898 /* return 0 if VFIO modules not loaded */
899 if (vfio_available == 0) {
900 RTE_LOG(DEBUG, EAL, "VFIO modules not loaded, "
901 "skipping VFIO support...\n");
905 default_vfio_cfg->vfio_container_fd = rte_vfio_get_container_fd();
907 /* check if we have VFIO driver enabled */
908 if (default_vfio_cfg->vfio_container_fd != -1) {
909 RTE_LOG(NOTICE, EAL, "VFIO support initialized\n");
910 default_vfio_cfg->vfio_enabled = 1;
912 RTE_LOG(NOTICE, EAL, "VFIO support could not be initialized\n");
919 rte_vfio_is_enabled(const char *modname)
921 const int mod_available = rte_eal_check_module(modname) > 0;
922 return default_vfio_cfg->vfio_enabled && mod_available;
925 const struct vfio_iommu_type *
926 vfio_set_iommu_type(int vfio_container_fd)
929 for (idx = 0; idx < RTE_DIM(iommu_types); idx++) {
930 const struct vfio_iommu_type *t = &iommu_types[idx];
932 int ret = ioctl(vfio_container_fd, VFIO_SET_IOMMU,
935 RTE_LOG(NOTICE, EAL, " using IOMMU type %d (%s)\n",
936 t->type_id, t->name);
939 /* not an error, there may be more supported IOMMU types */
940 RTE_LOG(DEBUG, EAL, " set IOMMU type %d (%s) failed, "
941 "error %i (%s)\n", t->type_id, t->name, errno,
944 /* if we didn't find a suitable IOMMU type, fail */
949 vfio_has_supported_extensions(int vfio_container_fd)
952 unsigned idx, n_extensions = 0;
953 for (idx = 0; idx < RTE_DIM(iommu_types); idx++) {
954 const struct vfio_iommu_type *t = &iommu_types[idx];
956 ret = ioctl(vfio_container_fd, VFIO_CHECK_EXTENSION,
959 RTE_LOG(ERR, EAL, " could not get IOMMU type, "
960 "error %i (%s)\n", errno,
962 close(vfio_container_fd);
964 } else if (ret == 1) {
965 /* we found a supported extension */
968 RTE_LOG(DEBUG, EAL, " IOMMU type %d (%s) is %s\n",
970 ret ? "supported" : "not supported");
973 /* if we didn't find any supported IOMMU types, fail */
975 close(vfio_container_fd);
983 rte_vfio_get_container_fd(void)
985 int ret, vfio_container_fd;
986 struct rte_mp_msg mp_req, *mp_rep;
987 struct rte_mp_reply mp_reply;
988 struct timespec ts = {.tv_sec = 5, .tv_nsec = 0};
989 struct vfio_mp_param *p = (struct vfio_mp_param *)mp_req.param;
992 /* if we're in a primary process, try to open the container */
993 if (internal_config.process_type == RTE_PROC_PRIMARY) {
994 vfio_container_fd = open(VFIO_CONTAINER_PATH, O_RDWR);
995 if (vfio_container_fd < 0) {
996 RTE_LOG(ERR, EAL, " cannot open VFIO container, "
997 "error %i (%s)\n", errno, strerror(errno));
1001 /* check VFIO API version */
1002 ret = ioctl(vfio_container_fd, VFIO_GET_API_VERSION);
1003 if (ret != VFIO_API_VERSION) {
1005 RTE_LOG(ERR, EAL, " could not get VFIO API version, "
1006 "error %i (%s)\n", errno, strerror(errno));
1008 RTE_LOG(ERR, EAL, " unsupported VFIO API version!\n");
1009 close(vfio_container_fd);
1013 ret = vfio_has_supported_extensions(vfio_container_fd);
1015 RTE_LOG(ERR, EAL, " no supported IOMMU "
1016 "extensions found!\n");
1020 return vfio_container_fd;
1023 * if we're in a secondary process, request container fd from the
1024 * primary process via mp channel
1026 p->req = SOCKET_REQ_CONTAINER;
1027 strcpy(mp_req.name, EAL_VFIO_MP);
1028 mp_req.len_param = sizeof(*p);
1031 vfio_container_fd = -1;
1032 if (rte_mp_request_sync(&mp_req, &mp_reply, &ts) == 0 &&
1033 mp_reply.nb_received == 1) {
1034 mp_rep = &mp_reply.msgs[0];
1035 p = (struct vfio_mp_param *)mp_rep->param;
1036 if (p->result == SOCKET_OK && mp_rep->num_fds == 1) {
1037 free(mp_reply.msgs);
1038 return mp_rep->fds[0];
1040 free(mp_reply.msgs);
1043 RTE_LOG(ERR, EAL, " cannot request container fd\n");
1048 rte_vfio_get_group_num(const char *sysfs_base,
1049 const char *dev_addr, int *iommu_group_num)
1051 char linkname[PATH_MAX];
1052 char filename[PATH_MAX];
1053 char *tok[16], *group_tok, *end;
1056 memset(linkname, 0, sizeof(linkname));
1057 memset(filename, 0, sizeof(filename));
1059 /* try to find out IOMMU group for this device */
1060 snprintf(linkname, sizeof(linkname),
1061 "%s/%s/iommu_group", sysfs_base, dev_addr);
1063 ret = readlink(linkname, filename, sizeof(filename));
1065 /* if the link doesn't exist, no VFIO for us */
1069 ret = rte_strsplit(filename, sizeof(filename),
1070 tok, RTE_DIM(tok), '/');
1073 RTE_LOG(ERR, EAL, " %s cannot get IOMMU group\n", dev_addr);
1077 /* IOMMU group is always the last token */
1079 group_tok = tok[ret - 1];
1081 *iommu_group_num = strtol(group_tok, &end, 10);
1082 if ((end != group_tok && *end != '\0') || errno != 0) {
1083 RTE_LOG(ERR, EAL, " %s error parsing IOMMU number!\n", dev_addr);
1091 type1_map(const struct rte_memseg_list *msl, const struct rte_memseg *ms,
1094 int *vfio_container_fd = arg;
1099 return vfio_type1_dma_mem_map(*vfio_container_fd, ms->addr_64, ms->iova,
1104 vfio_type1_dma_mem_map(int vfio_container_fd, uint64_t vaddr, uint64_t iova,
1105 uint64_t len, int do_map)
1107 struct vfio_iommu_type1_dma_map dma_map;
1108 struct vfio_iommu_type1_dma_unmap dma_unmap;
1112 memset(&dma_map, 0, sizeof(dma_map));
1113 dma_map.argsz = sizeof(struct vfio_iommu_type1_dma_map);
1114 dma_map.vaddr = vaddr;
1116 dma_map.iova = iova;
1117 dma_map.flags = VFIO_DMA_MAP_FLAG_READ |
1118 VFIO_DMA_MAP_FLAG_WRITE;
1120 ret = ioctl(vfio_container_fd, VFIO_IOMMU_MAP_DMA, &dma_map);
1122 RTE_LOG(ERR, EAL, " cannot set up DMA remapping, error %i (%s)\n",
1123 errno, strerror(errno));
1127 memset(&dma_unmap, 0, sizeof(dma_unmap));
1128 dma_unmap.argsz = sizeof(struct vfio_iommu_type1_dma_unmap);
1129 dma_unmap.size = len;
1130 dma_unmap.iova = iova;
1132 ret = ioctl(vfio_container_fd, VFIO_IOMMU_UNMAP_DMA,
1135 RTE_LOG(ERR, EAL, " cannot clear DMA remapping, error %i (%s)\n",
1136 errno, strerror(errno));
1145 vfio_type1_dma_map(int vfio_container_fd)
1147 return rte_memseg_walk(type1_map, &vfio_container_fd);
1151 vfio_spapr_dma_do_map(int vfio_container_fd, uint64_t vaddr, uint64_t iova,
1152 uint64_t len, int do_map)
1154 struct vfio_iommu_type1_dma_map dma_map;
1155 struct vfio_iommu_type1_dma_unmap dma_unmap;
1157 struct vfio_iommu_spapr_register_memory reg = {
1158 .argsz = sizeof(reg),
1161 reg.vaddr = (uintptr_t) vaddr;
1165 ret = ioctl(vfio_container_fd,
1166 VFIO_IOMMU_SPAPR_REGISTER_MEMORY, ®);
1168 RTE_LOG(ERR, EAL, " cannot register vaddr for IOMMU, "
1169 "error %i (%s)\n", errno, strerror(errno));
1173 memset(&dma_map, 0, sizeof(dma_map));
1174 dma_map.argsz = sizeof(struct vfio_iommu_type1_dma_map);
1175 dma_map.vaddr = vaddr;
1177 dma_map.iova = iova;
1178 dma_map.flags = VFIO_DMA_MAP_FLAG_READ |
1179 VFIO_DMA_MAP_FLAG_WRITE;
1181 ret = ioctl(vfio_container_fd, VFIO_IOMMU_MAP_DMA, &dma_map);
1183 RTE_LOG(ERR, EAL, " cannot set up DMA remapping, error %i (%s)\n",
1184 errno, strerror(errno));
1189 ret = ioctl(vfio_container_fd,
1190 VFIO_IOMMU_SPAPR_UNREGISTER_MEMORY, ®);
1192 RTE_LOG(ERR, EAL, " cannot unregister vaddr for IOMMU, error %i (%s)\n",
1193 errno, strerror(errno));
1197 memset(&dma_unmap, 0, sizeof(dma_unmap));
1198 dma_unmap.argsz = sizeof(struct vfio_iommu_type1_dma_unmap);
1199 dma_unmap.size = len;
1200 dma_unmap.iova = iova;
1202 ret = ioctl(vfio_container_fd, VFIO_IOMMU_UNMAP_DMA,
1205 RTE_LOG(ERR, EAL, " cannot clear DMA remapping, error %i (%s)\n",
1206 errno, strerror(errno));
1215 vfio_spapr_map_walk(const struct rte_memseg_list *msl,
1216 const struct rte_memseg *ms, void *arg)
1218 int *vfio_container_fd = arg;
1223 return vfio_spapr_dma_do_map(*vfio_container_fd, ms->addr_64, ms->iova,
1227 struct spapr_walk_param {
1228 uint64_t window_size;
1229 uint64_t hugepage_sz;
1232 vfio_spapr_window_size_walk(const struct rte_memseg_list *msl,
1233 const struct rte_memseg *ms, void *arg)
1235 struct spapr_walk_param *param = arg;
1236 uint64_t max = ms->iova + ms->len;
1241 if (max > param->window_size) {
1242 param->hugepage_sz = ms->hugepage_sz;
1243 param->window_size = max;
1250 vfio_spapr_create_new_dma_window(int vfio_container_fd,
1251 struct vfio_iommu_spapr_tce_create *create) {
1252 struct vfio_iommu_spapr_tce_remove remove = {
1253 .argsz = sizeof(remove),
1255 struct vfio_iommu_spapr_tce_info info = {
1256 .argsz = sizeof(info),
1260 /* query spapr iommu info */
1261 ret = ioctl(vfio_container_fd, VFIO_IOMMU_SPAPR_TCE_GET_INFO, &info);
1263 RTE_LOG(ERR, EAL, " cannot get iommu info, "
1264 "error %i (%s)\n", errno, strerror(errno));
1268 /* remove default DMA of 32 bit window */
1269 remove.start_addr = info.dma32_window_start;
1270 ret = ioctl(vfio_container_fd, VFIO_IOMMU_SPAPR_TCE_REMOVE, &remove);
1272 RTE_LOG(ERR, EAL, " cannot remove default DMA window, "
1273 "error %i (%s)\n", errno, strerror(errno));
1277 /* create new DMA window */
1278 ret = ioctl(vfio_container_fd, VFIO_IOMMU_SPAPR_TCE_CREATE, create);
1280 RTE_LOG(ERR, EAL, " cannot create new DMA window, "
1281 "error %i (%s)\n", errno, strerror(errno));
1285 if (create->start_addr != 0) {
1286 RTE_LOG(ERR, EAL, " DMA window start address != 0\n");
1294 vfio_spapr_dma_mem_map(int vfio_container_fd, uint64_t vaddr, uint64_t iova,
1295 uint64_t len, int do_map)
1297 struct spapr_walk_param param;
1298 struct vfio_iommu_spapr_tce_create create = {
1299 .argsz = sizeof(create),
1301 struct vfio_config *vfio_cfg;
1302 struct user_mem_maps *user_mem_maps;
1305 vfio_cfg = get_vfio_cfg_by_container_fd(vfio_container_fd);
1306 if (vfio_cfg == NULL) {
1307 RTE_LOG(ERR, EAL, " invalid container fd!\n");
1311 user_mem_maps = &vfio_cfg->mem_maps;
1312 rte_spinlock_recursive_lock(&user_mem_maps->lock);
1314 /* check if window size needs to be adjusted */
1315 memset(¶m, 0, sizeof(param));
1317 /* we're inside a callback so use thread-unsafe version */
1318 if (rte_memseg_walk_thread_unsafe(vfio_spapr_window_size_walk,
1320 RTE_LOG(ERR, EAL, "Could not get window size\n");
1325 /* also check user maps */
1326 for (i = 0; i < user_mem_maps->n_maps; i++) {
1327 uint64_t max = user_mem_maps->maps[i].iova +
1328 user_mem_maps->maps[i].len;
1329 create.window_size = RTE_MAX(create.window_size, max);
1332 /* sPAPR requires window size to be a power of 2 */
1333 create.window_size = rte_align64pow2(param.window_size);
1334 create.page_shift = __builtin_ctzll(param.hugepage_sz);
1339 /* re-create window and remap the entire memory */
1340 if (iova > create.window_size) {
1341 if (vfio_spapr_create_new_dma_window(vfio_container_fd,
1343 RTE_LOG(ERR, EAL, "Could not create new DMA window\n");
1347 /* we're inside a callback, so use thread-unsafe version
1349 if (rte_memseg_walk_thread_unsafe(vfio_spapr_map_walk,
1350 &vfio_container_fd) < 0) {
1351 RTE_LOG(ERR, EAL, "Could not recreate DMA maps\n");
1355 /* remap all user maps */
1356 for (i = 0; i < user_mem_maps->n_maps; i++) {
1357 struct user_mem_map *map =
1358 &user_mem_maps->maps[i];
1359 if (vfio_spapr_dma_do_map(vfio_container_fd,
1360 map->addr, map->iova, map->len,
1362 RTE_LOG(ERR, EAL, "Could not recreate user DMA maps\n");
1369 /* now that we've remapped all of the memory that was present
1370 * before, map the segment that we were requested to map.
1372 * however, if we were called by the callback, the memory we
1373 * were called with was already in the memseg list, so previous
1374 * mapping should've mapped that segment already.
1376 * virt2memseg_list is a relatively cheap check, so use that. if
1377 * memory is within any memseg list, it's a memseg, so it's
1380 addr = (void *)(uintptr_t)vaddr;
1381 if (rte_mem_virt2memseg_list(addr) == NULL &&
1382 vfio_spapr_dma_do_map(vfio_container_fd,
1383 vaddr, iova, len, 1) < 0) {
1384 RTE_LOG(ERR, EAL, "Could not map segment\n");
1389 /* for unmap, check if iova within DMA window */
1390 if (iova > create.window_size) {
1391 RTE_LOG(ERR, EAL, "iova beyond DMA window for unmap");
1396 vfio_spapr_dma_do_map(vfio_container_fd, vaddr, iova, len, 0);
1399 rte_spinlock_recursive_unlock(&user_mem_maps->lock);
1404 vfio_spapr_dma_map(int vfio_container_fd)
1406 struct vfio_iommu_spapr_tce_create create = {
1407 .argsz = sizeof(create),
1409 struct spapr_walk_param param;
1411 memset(¶m, 0, sizeof(param));
1413 /* create DMA window from 0 to max(phys_addr + len) */
1414 rte_memseg_walk(vfio_spapr_window_size_walk, ¶m);
1416 /* sPAPR requires window size to be a power of 2 */
1417 create.window_size = rte_align64pow2(param.window_size);
1418 create.page_shift = __builtin_ctzll(param.hugepage_sz);
1421 if (vfio_spapr_create_new_dma_window(vfio_container_fd, &create) < 0) {
1422 RTE_LOG(ERR, EAL, "Could not create new DMA window\n");
1426 /* map all DPDK segments for DMA. use 1:1 PA to IOVA mapping */
1427 if (rte_memseg_walk(vfio_spapr_map_walk, &vfio_container_fd) < 0)
1434 vfio_noiommu_dma_map(int __rte_unused vfio_container_fd)
1436 /* No-IOMMU mode does not need DMA mapping */
1441 vfio_noiommu_dma_mem_map(int __rte_unused vfio_container_fd,
1442 uint64_t __rte_unused vaddr,
1443 uint64_t __rte_unused iova, uint64_t __rte_unused len,
1444 int __rte_unused do_map)
1446 /* No-IOMMU mode does not need DMA mapping */
1451 vfio_dma_mem_map(struct vfio_config *vfio_cfg, uint64_t vaddr, uint64_t iova,
1452 uint64_t len, int do_map)
1454 const struct vfio_iommu_type *t = vfio_cfg->vfio_iommu_type;
1457 RTE_LOG(ERR, EAL, " VFIO support not initialized\n");
1462 if (!t->dma_user_map_func) {
1464 " VFIO custom DMA region maping not supported by IOMMU %s\n",
1466 rte_errno = ENOTSUP;
1470 return t->dma_user_map_func(vfio_cfg->vfio_container_fd, vaddr, iova,
1475 container_dma_map(struct vfio_config *vfio_cfg, uint64_t vaddr, uint64_t iova,
1478 struct user_mem_map *new_map;
1479 struct user_mem_maps *user_mem_maps;
1482 user_mem_maps = &vfio_cfg->mem_maps;
1483 rte_spinlock_recursive_lock(&user_mem_maps->lock);
1484 if (user_mem_maps->n_maps == VFIO_MAX_USER_MEM_MAPS) {
1485 RTE_LOG(ERR, EAL, "No more space for user mem maps\n");
1491 if (vfio_dma_mem_map(vfio_cfg, vaddr, iova, len, 1)) {
1492 /* technically, this will fail if there are currently no devices
1493 * plugged in, even if a device were added later, this mapping
1494 * might have succeeded. however, since we cannot verify if this
1495 * is a valid mapping without having a device attached, consider
1496 * this to be unsupported, because we can't just store any old
1497 * mapping and pollute list of active mappings willy-nilly.
1499 RTE_LOG(ERR, EAL, "Couldn't map new region for DMA\n");
1503 /* create new user mem map entry */
1504 new_map = &user_mem_maps->maps[user_mem_maps->n_maps++];
1505 new_map->addr = vaddr;
1506 new_map->iova = iova;
1509 compact_user_maps(user_mem_maps);
1511 rte_spinlock_recursive_unlock(&user_mem_maps->lock);
1516 container_dma_unmap(struct vfio_config *vfio_cfg, uint64_t vaddr, uint64_t iova,
1519 struct user_mem_map *map, *new_map = NULL;
1520 struct user_mem_maps *user_mem_maps;
1523 user_mem_maps = &vfio_cfg->mem_maps;
1524 rte_spinlock_recursive_lock(&user_mem_maps->lock);
1526 /* find our mapping */
1527 map = find_user_mem_map(user_mem_maps, vaddr, iova, len);
1529 RTE_LOG(ERR, EAL, "Couldn't find previously mapped region\n");
1534 if (map->addr != vaddr || map->iova != iova || map->len != len) {
1535 /* we're partially unmapping a previously mapped region, so we
1536 * need to split entry into two.
1538 if (user_mem_maps->n_maps == VFIO_MAX_USER_MEM_MAPS) {
1539 RTE_LOG(ERR, EAL, "Not enough space to store partial mapping\n");
1544 new_map = &user_mem_maps->maps[user_mem_maps->n_maps++];
1547 /* unmap the entry */
1548 if (vfio_dma_mem_map(vfio_cfg, vaddr, iova, len, 0)) {
1549 /* there may not be any devices plugged in, so unmapping will
1550 * fail with ENODEV/ENOTSUP rte_errno values, but that doesn't
1551 * stop us from removing the mapping, as the assumption is we
1552 * won't be needing this memory any more and thus will want to
1553 * prevent it from being remapped again on hotplug. so, only
1554 * fail if we indeed failed to unmap (e.g. if the mapping was
1555 * within our mapped range but had invalid alignment).
1557 if (rte_errno != ENODEV && rte_errno != ENOTSUP) {
1558 RTE_LOG(ERR, EAL, "Couldn't unmap region for DMA\n");
1562 RTE_LOG(DEBUG, EAL, "DMA unmapping failed, but removing mappings anyway\n");
1565 /* remove map from the list of active mappings */
1566 if (new_map != NULL) {
1567 adjust_map(map, new_map, vaddr, len);
1569 /* if we've created a new map by splitting, sort everything */
1570 if (!is_null_map(new_map)) {
1571 compact_user_maps(user_mem_maps);
1573 /* we've created a new mapping, but it was unused */
1574 user_mem_maps->n_maps--;
1577 memset(map, 0, sizeof(*map));
1578 compact_user_maps(user_mem_maps);
1579 user_mem_maps->n_maps--;
1583 rte_spinlock_recursive_unlock(&user_mem_maps->lock);
1588 rte_vfio_dma_map(uint64_t vaddr, uint64_t iova, uint64_t len)
1595 return container_dma_map(default_vfio_cfg, vaddr, iova, len);
1599 rte_vfio_dma_unmap(uint64_t vaddr, uint64_t iova, uint64_t len)
1606 return container_dma_unmap(default_vfio_cfg, vaddr, iova, len);
1610 rte_vfio_noiommu_is_enabled(void)
1616 fd = open(VFIO_NOIOMMU_MODE, O_RDONLY);
1618 if (errno != ENOENT) {
1619 RTE_LOG(ERR, EAL, " cannot open vfio noiommu file %i (%s)\n",
1620 errno, strerror(errno));
1624 * else the file does not exists
1625 * i.e. noiommu is not enabled
1630 cnt = read(fd, &c, 1);
1633 RTE_LOG(ERR, EAL, " unable to read from vfio noiommu "
1634 "file %i (%s)\n", errno, strerror(errno));
1642 rte_vfio_container_create(void)
1646 /* Find an empty slot to store new vfio config */
1647 for (i = 1; i < VFIO_MAX_CONTAINERS; i++) {
1648 if (vfio_cfgs[i].vfio_container_fd == -1)
1652 if (i == VFIO_MAX_CONTAINERS) {
1653 RTE_LOG(ERR, EAL, "exceed max vfio container limit\n");
1657 vfio_cfgs[i].vfio_container_fd = rte_vfio_get_container_fd();
1658 if (vfio_cfgs[i].vfio_container_fd < 0) {
1659 RTE_LOG(NOTICE, EAL, "fail to create a new container\n");
1663 return vfio_cfgs[i].vfio_container_fd;
1666 int __rte_experimental
1667 rte_vfio_container_destroy(int container_fd)
1669 struct vfio_config *vfio_cfg;
1672 vfio_cfg = get_vfio_cfg_by_container_fd(container_fd);
1673 if (vfio_cfg == NULL) {
1674 RTE_LOG(ERR, EAL, "Invalid container fd\n");
1678 for (i = 0; i < VFIO_MAX_GROUPS; i++)
1679 if (vfio_cfg->vfio_groups[i].group_num != -1)
1680 rte_vfio_container_group_unbind(container_fd,
1681 vfio_cfg->vfio_groups[i].group_num);
1683 close(container_fd);
1684 vfio_cfg->vfio_container_fd = -1;
1685 vfio_cfg->vfio_active_groups = 0;
1686 vfio_cfg->vfio_iommu_type = NULL;
1692 rte_vfio_container_group_bind(int container_fd, int iommu_group_num)
1694 struct vfio_config *vfio_cfg;
1695 struct vfio_group *cur_grp;
1699 vfio_cfg = get_vfio_cfg_by_container_fd(container_fd);
1700 if (vfio_cfg == NULL) {
1701 RTE_LOG(ERR, EAL, "Invalid container fd\n");
1705 /* Check room for new group */
1706 if (vfio_cfg->vfio_active_groups == VFIO_MAX_GROUPS) {
1707 RTE_LOG(ERR, EAL, "Maximum number of VFIO groups reached!\n");
1711 /* Get an index for the new group */
1712 for (i = 0; i < VFIO_MAX_GROUPS; i++)
1713 if (vfio_cfg->vfio_groups[i].group_num == -1) {
1714 cur_grp = &vfio_cfg->vfio_groups[i];
1718 /* This should not happen */
1719 if (i == VFIO_MAX_GROUPS) {
1720 RTE_LOG(ERR, EAL, "No VFIO group free slot found\n");
1724 vfio_group_fd = vfio_open_group_fd(iommu_group_num);
1725 if (vfio_group_fd < 0) {
1726 RTE_LOG(ERR, EAL, "Failed to open group %d\n", iommu_group_num);
1729 cur_grp->group_num = iommu_group_num;
1730 cur_grp->fd = vfio_group_fd;
1731 cur_grp->devices = 0;
1732 vfio_cfg->vfio_active_groups++;
1734 return vfio_group_fd;
1738 rte_vfio_container_group_unbind(int container_fd, int iommu_group_num)
1740 struct vfio_config *vfio_cfg;
1741 struct vfio_group *cur_grp = NULL;
1744 vfio_cfg = get_vfio_cfg_by_container_fd(container_fd);
1745 if (vfio_cfg == NULL) {
1746 RTE_LOG(ERR, EAL, "Invalid container fd\n");
1750 for (i = 0; i < VFIO_MAX_GROUPS; i++) {
1751 if (vfio_cfg->vfio_groups[i].group_num == iommu_group_num) {
1752 cur_grp = &vfio_cfg->vfio_groups[i];
1757 /* This should not happen */
1758 if (i == VFIO_MAX_GROUPS || cur_grp == NULL) {
1759 RTE_LOG(ERR, EAL, "Specified group number not found\n");
1763 if (cur_grp->fd >= 0 && close(cur_grp->fd) < 0) {
1764 RTE_LOG(ERR, EAL, "Error when closing vfio_group_fd for"
1765 " iommu_group_num %d\n", iommu_group_num);
1768 cur_grp->group_num = -1;
1770 cur_grp->devices = 0;
1771 vfio_cfg->vfio_active_groups--;
1777 rte_vfio_container_dma_map(int container_fd, uint64_t vaddr, uint64_t iova,
1780 struct vfio_config *vfio_cfg;
1787 vfio_cfg = get_vfio_cfg_by_container_fd(container_fd);
1788 if (vfio_cfg == NULL) {
1789 RTE_LOG(ERR, EAL, "Invalid container fd\n");
1793 return container_dma_map(vfio_cfg, vaddr, iova, len);
1797 rte_vfio_container_dma_unmap(int container_fd, uint64_t vaddr, uint64_t iova,
1800 struct vfio_config *vfio_cfg;
1807 vfio_cfg = get_vfio_cfg_by_container_fd(container_fd);
1808 if (vfio_cfg == NULL) {
1809 RTE_LOG(ERR, EAL, "Invalid container fd\n");
1813 return container_dma_unmap(vfio_cfg, vaddr, iova, len);
1819 rte_vfio_dma_map(uint64_t __rte_unused vaddr, __rte_unused uint64_t iova,
1820 __rte_unused uint64_t len)
1826 rte_vfio_dma_unmap(uint64_t __rte_unused vaddr, uint64_t __rte_unused iova,
1827 __rte_unused uint64_t len)
1833 rte_vfio_setup_device(__rte_unused const char *sysfs_base,
1834 __rte_unused const char *dev_addr,
1835 __rte_unused int *vfio_dev_fd,
1836 __rte_unused struct vfio_device_info *device_info)
1842 rte_vfio_release_device(__rte_unused const char *sysfs_base,
1843 __rte_unused const char *dev_addr, __rte_unused int fd)
1849 rte_vfio_enable(__rte_unused const char *modname)
1855 rte_vfio_is_enabled(__rte_unused const char *modname)
1861 rte_vfio_noiommu_is_enabled(void)
1867 rte_vfio_clear_group(__rte_unused int vfio_group_fd)
1873 rte_vfio_get_group_num(__rte_unused const char *sysfs_base,
1874 __rte_unused const char *dev_addr,
1875 __rte_unused int *iommu_group_num)
1881 rte_vfio_get_container_fd(void)
1887 rte_vfio_get_group_fd(__rte_unused int iommu_group_num)
1893 rte_vfio_container_create(void)
1899 rte_vfio_container_destroy(__rte_unused int container_fd)
1905 rte_vfio_container_group_bind(__rte_unused int container_fd,
1906 __rte_unused int iommu_group_num)
1912 rte_vfio_container_group_unbind(__rte_unused int container_fd,
1913 __rte_unused int iommu_group_num)
1919 rte_vfio_container_dma_map(__rte_unused int container_fd,
1920 __rte_unused uint64_t vaddr,
1921 __rte_unused uint64_t iova,
1922 __rte_unused uint64_t len)
1928 rte_vfio_container_dma_unmap(__rte_unused int container_fd,
1929 __rte_unused uint64_t vaddr,
1930 __rte_unused uint64_t iova,
1931 __rte_unused uint64_t len)
1936 #endif /* VFIO_PRESENT */