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"
18 #include "eal_memcfg.h"
20 #include "eal_private.h"
24 #define VFIO_MEM_EVENT_CLB_NAME "vfio_mem_event_clb"
26 /* hot plug/unplug of VFIO groups may cause all DMA maps to be dropped. we can
27 * recreate the mappings for DPDK segments, but we cannot do so for memory that
28 * was registered by the user themselves, so we need to store the user mappings
29 * somewhere, to recreate them later.
31 #define VFIO_MAX_USER_MEM_MAPS 256
38 struct user_mem_maps {
39 rte_spinlock_recursive_t lock;
41 struct user_mem_map maps[VFIO_MAX_USER_MEM_MAPS];
46 int vfio_container_fd;
47 int vfio_active_groups;
48 const struct vfio_iommu_type *vfio_iommu_type;
49 struct vfio_group vfio_groups[VFIO_MAX_GROUPS];
50 struct user_mem_maps mem_maps;
53 /* per-process VFIO config */
54 static struct vfio_config vfio_cfgs[VFIO_MAX_CONTAINERS];
55 static struct vfio_config *default_vfio_cfg = &vfio_cfgs[0];
57 static int vfio_type1_dma_map(int);
58 static int vfio_type1_dma_mem_map(int, uint64_t, uint64_t, uint64_t, int);
59 static int vfio_spapr_dma_map(int);
60 static int vfio_spapr_dma_mem_map(int, uint64_t, uint64_t, uint64_t, int);
61 static int vfio_noiommu_dma_map(int);
62 static int vfio_noiommu_dma_mem_map(int, uint64_t, uint64_t, uint64_t, int);
63 static int vfio_dma_mem_map(struct vfio_config *vfio_cfg, uint64_t vaddr,
64 uint64_t iova, uint64_t len, int do_map);
66 /* IOMMU types we support */
67 static const struct vfio_iommu_type iommu_types[] = {
68 /* x86 IOMMU, otherwise known as type 1 */
70 .type_id = RTE_VFIO_TYPE1,
72 .dma_map_func = &vfio_type1_dma_map,
73 .dma_user_map_func = &vfio_type1_dma_mem_map
75 /* ppc64 IOMMU, otherwise known as spapr */
77 .type_id = RTE_VFIO_SPAPR,
79 .dma_map_func = &vfio_spapr_dma_map,
80 .dma_user_map_func = &vfio_spapr_dma_mem_map
84 .type_id = RTE_VFIO_NOIOMMU,
86 .dma_map_func = &vfio_noiommu_dma_map,
87 .dma_user_map_func = &vfio_noiommu_dma_mem_map
92 is_null_map(const struct user_mem_map *map)
94 return map->addr == 0 && map->iova == 0 && map->len == 0;
97 /* we may need to merge user mem maps together in case of user mapping/unmapping
98 * chunks of memory, so we'll need a comparator function to sort segments.
101 user_mem_map_cmp(const void *a, const void *b)
103 const struct user_mem_map *umm_a = a;
104 const struct user_mem_map *umm_b = b;
106 /* move null entries to end */
107 if (is_null_map(umm_a))
109 if (is_null_map(umm_b))
112 /* sort by iova first */
113 if (umm_a->iova < umm_b->iova)
115 if (umm_a->iova > umm_b->iova)
118 if (umm_a->addr < umm_b->addr)
120 if (umm_a->addr > umm_b->addr)
123 if (umm_a->len < umm_b->len)
125 if (umm_a->len > umm_b->len)
131 /* adjust user map entry. this may result in shortening of existing map, or in
132 * splitting existing map in two pieces.
135 adjust_map(struct user_mem_map *src, struct user_mem_map *end,
136 uint64_t remove_va_start, uint64_t remove_len)
138 /* if va start is same as start address, we're simply moving start */
139 if (remove_va_start == src->addr) {
140 src->addr += remove_len;
141 src->iova += remove_len;
142 src->len -= remove_len;
143 } else if (remove_va_start + remove_len == src->addr + src->len) {
144 /* we're shrinking mapping from the end */
145 src->len -= remove_len;
147 /* we're blowing a hole in the middle */
148 struct user_mem_map tmp;
149 uint64_t total_len = src->len;
151 /* adjust source segment length */
152 src->len = remove_va_start - src->addr;
154 /* create temporary segment in the middle */
155 tmp.addr = src->addr + src->len;
156 tmp.iova = src->iova + src->len;
157 tmp.len = remove_len;
159 /* populate end segment - this one we will be keeping */
160 end->addr = tmp.addr + tmp.len;
161 end->iova = tmp.iova + tmp.len;
162 end->len = total_len - src->len - tmp.len;
166 /* try merging two maps into one, return 1 if succeeded */
168 merge_map(struct user_mem_map *left, struct user_mem_map *right)
170 if (left->addr + left->len != right->addr)
172 if (left->iova + left->len != right->iova)
175 left->len += right->len;
177 memset(right, 0, sizeof(*right));
182 static struct user_mem_map *
183 find_user_mem_map(struct user_mem_maps *user_mem_maps, uint64_t addr,
184 uint64_t iova, uint64_t len)
186 uint64_t va_end = addr + len;
187 uint64_t iova_end = iova + len;
190 for (i = 0; i < user_mem_maps->n_maps; i++) {
191 struct user_mem_map *map = &user_mem_maps->maps[i];
192 uint64_t map_va_end = map->addr + map->len;
193 uint64_t map_iova_end = map->iova + map->len;
196 if (addr < map->addr || addr >= map_va_end)
198 /* check if VA end is within boundaries */
199 if (va_end <= map->addr || va_end > map_va_end)
202 /* check start IOVA */
203 if (iova < map->iova || iova >= map_iova_end)
205 /* check if IOVA end is within boundaries */
206 if (iova_end <= map->iova || iova_end > map_iova_end)
209 /* we've found our map */
215 /* this will sort all user maps, and merge/compact any adjacent maps */
217 compact_user_maps(struct user_mem_maps *user_mem_maps)
219 int i, n_merged, cur_idx;
221 qsort(user_mem_maps->maps, user_mem_maps->n_maps,
222 sizeof(user_mem_maps->maps[0]), user_mem_map_cmp);
224 /* we'll go over the list backwards when merging */
226 for (i = user_mem_maps->n_maps - 2; i >= 0; i--) {
227 struct user_mem_map *l, *r;
229 l = &user_mem_maps->maps[i];
230 r = &user_mem_maps->maps[i + 1];
232 if (is_null_map(l) || is_null_map(r))
239 /* the entries are still sorted, but now they have holes in them, so
240 * walk through the list and remove the holes
244 for (i = 0; i < user_mem_maps->n_maps; i++) {
245 if (!is_null_map(&user_mem_maps->maps[i])) {
246 struct user_mem_map *src, *dst;
248 src = &user_mem_maps->maps[i];
249 dst = &user_mem_maps->maps[cur_idx++];
252 memcpy(dst, src, sizeof(*src));
253 memset(src, 0, sizeof(*src));
257 user_mem_maps->n_maps = cur_idx;
262 vfio_open_group_fd(int iommu_group_num)
265 char filename[PATH_MAX];
266 struct rte_mp_msg mp_req, *mp_rep;
267 struct rte_mp_reply mp_reply = {0};
268 struct timespec ts = {.tv_sec = 5, .tv_nsec = 0};
269 struct vfio_mp_param *p = (struct vfio_mp_param *)mp_req.param;
271 /* if primary, try to open the group */
272 if (internal_config.process_type == RTE_PROC_PRIMARY) {
273 /* try regular group format */
274 snprintf(filename, sizeof(filename),
275 VFIO_GROUP_FMT, iommu_group_num);
276 vfio_group_fd = open(filename, O_RDWR);
277 if (vfio_group_fd < 0) {
278 /* if file not found, it's not an error */
279 if (errno != ENOENT) {
280 RTE_LOG(ERR, EAL, "Cannot open %s: %s\n", filename,
285 /* special case: try no-IOMMU path as well */
286 snprintf(filename, sizeof(filename),
287 VFIO_NOIOMMU_GROUP_FMT,
289 vfio_group_fd = open(filename, O_RDWR);
290 if (vfio_group_fd < 0) {
291 if (errno != ENOENT) {
292 RTE_LOG(ERR, EAL, "Cannot open %s: %s\n", filename,
298 /* noiommu group found */
301 return vfio_group_fd;
303 /* if we're in a secondary process, request group fd from the primary
304 * process via mp channel.
306 p->req = SOCKET_REQ_GROUP;
307 p->group_num = iommu_group_num;
308 strcpy(mp_req.name, EAL_VFIO_MP);
309 mp_req.len_param = sizeof(*p);
313 if (rte_mp_request_sync(&mp_req, &mp_reply, &ts) == 0 &&
314 mp_reply.nb_received == 1) {
315 mp_rep = &mp_reply.msgs[0];
316 p = (struct vfio_mp_param *)mp_rep->param;
317 if (p->result == SOCKET_OK && mp_rep->num_fds == 1) {
318 vfio_group_fd = mp_rep->fds[0];
319 } else if (p->result == SOCKET_NO_FD) {
320 RTE_LOG(ERR, EAL, " bad VFIO group fd\n");
326 if (vfio_group_fd < 0)
327 RTE_LOG(ERR, EAL, " cannot request group fd\n");
328 return vfio_group_fd;
331 static struct vfio_config *
332 get_vfio_cfg_by_group_num(int iommu_group_num)
334 struct vfio_config *vfio_cfg;
337 for (i = 0; i < VFIO_MAX_CONTAINERS; i++) {
338 vfio_cfg = &vfio_cfgs[i];
339 for (j = 0; j < VFIO_MAX_GROUPS; j++) {
340 if (vfio_cfg->vfio_groups[j].group_num ==
350 vfio_get_group_fd(struct vfio_config *vfio_cfg,
355 struct vfio_group *cur_grp;
357 /* check if we already have the group descriptor open */
358 for (i = 0; i < VFIO_MAX_GROUPS; i++)
359 if (vfio_cfg->vfio_groups[i].group_num == iommu_group_num)
360 return vfio_cfg->vfio_groups[i].fd;
362 /* Lets see first if there is room for a new group */
363 if (vfio_cfg->vfio_active_groups == VFIO_MAX_GROUPS) {
364 RTE_LOG(ERR, EAL, "Maximum number of VFIO groups reached!\n");
368 /* Now lets get an index for the new group */
369 for (i = 0; i < VFIO_MAX_GROUPS; i++)
370 if (vfio_cfg->vfio_groups[i].group_num == -1) {
371 cur_grp = &vfio_cfg->vfio_groups[i];
375 /* This should not happen */
376 if (i == VFIO_MAX_GROUPS) {
377 RTE_LOG(ERR, EAL, "No VFIO group free slot found\n");
381 vfio_group_fd = vfio_open_group_fd(iommu_group_num);
382 if (vfio_group_fd < 0) {
383 RTE_LOG(ERR, EAL, "Failed to open group %d\n", iommu_group_num);
387 cur_grp->group_num = iommu_group_num;
388 cur_grp->fd = vfio_group_fd;
389 vfio_cfg->vfio_active_groups++;
391 return vfio_group_fd;
394 static struct vfio_config *
395 get_vfio_cfg_by_group_fd(int vfio_group_fd)
397 struct vfio_config *vfio_cfg;
400 for (i = 0; i < VFIO_MAX_CONTAINERS; i++) {
401 vfio_cfg = &vfio_cfgs[i];
402 for (j = 0; j < VFIO_MAX_GROUPS; j++)
403 if (vfio_cfg->vfio_groups[j].fd == vfio_group_fd)
410 static struct vfio_config *
411 get_vfio_cfg_by_container_fd(int container_fd)
415 if (container_fd == RTE_VFIO_DEFAULT_CONTAINER_FD)
416 return default_vfio_cfg;
418 for (i = 0; i < VFIO_MAX_CONTAINERS; i++) {
419 if (vfio_cfgs[i].vfio_container_fd == container_fd)
420 return &vfio_cfgs[i];
427 rte_vfio_get_group_fd(int iommu_group_num)
429 struct vfio_config *vfio_cfg;
431 /* get the vfio_config it belongs to */
432 vfio_cfg = get_vfio_cfg_by_group_num(iommu_group_num);
433 vfio_cfg = vfio_cfg ? vfio_cfg : default_vfio_cfg;
435 return vfio_get_group_fd(vfio_cfg, iommu_group_num);
439 get_vfio_group_idx(int vfio_group_fd)
441 struct vfio_config *vfio_cfg;
444 for (i = 0; i < VFIO_MAX_CONTAINERS; i++) {
445 vfio_cfg = &vfio_cfgs[i];
446 for (j = 0; j < VFIO_MAX_GROUPS; j++)
447 if (vfio_cfg->vfio_groups[j].fd == vfio_group_fd)
455 vfio_group_device_get(int vfio_group_fd)
457 struct vfio_config *vfio_cfg;
460 vfio_cfg = get_vfio_cfg_by_group_fd(vfio_group_fd);
461 if (vfio_cfg == NULL) {
462 RTE_LOG(ERR, EAL, " invalid group fd!\n");
466 i = get_vfio_group_idx(vfio_group_fd);
467 if (i < 0 || i > (VFIO_MAX_GROUPS - 1))
468 RTE_LOG(ERR, EAL, " wrong vfio_group index (%d)\n", i);
470 vfio_cfg->vfio_groups[i].devices++;
474 vfio_group_device_put(int vfio_group_fd)
476 struct vfio_config *vfio_cfg;
479 vfio_cfg = get_vfio_cfg_by_group_fd(vfio_group_fd);
480 if (vfio_cfg == NULL) {
481 RTE_LOG(ERR, EAL, " invalid group fd!\n");
485 i = get_vfio_group_idx(vfio_group_fd);
486 if (i < 0 || i > (VFIO_MAX_GROUPS - 1))
487 RTE_LOG(ERR, EAL, " wrong vfio_group index (%d)\n", i);
489 vfio_cfg->vfio_groups[i].devices--;
493 vfio_group_device_count(int vfio_group_fd)
495 struct vfio_config *vfio_cfg;
498 vfio_cfg = get_vfio_cfg_by_group_fd(vfio_group_fd);
499 if (vfio_cfg == NULL) {
500 RTE_LOG(ERR, EAL, " invalid group fd!\n");
504 i = get_vfio_group_idx(vfio_group_fd);
505 if (i < 0 || i > (VFIO_MAX_GROUPS - 1)) {
506 RTE_LOG(ERR, EAL, " wrong vfio_group index (%d)\n", i);
510 return vfio_cfg->vfio_groups[i].devices;
514 vfio_mem_event_callback(enum rte_mem_event type, const void *addr, size_t len,
515 void *arg __rte_unused)
517 struct rte_memseg_list *msl;
518 struct rte_memseg *ms;
521 msl = rte_mem_virt2memseg_list(addr);
523 /* for IOVA as VA mode, no need to care for IOVA addresses */
524 if (rte_eal_iova_mode() == RTE_IOVA_VA && msl->external == 0) {
525 uint64_t vfio_va = (uint64_t)(uintptr_t)addr;
526 if (type == RTE_MEM_EVENT_ALLOC)
527 vfio_dma_mem_map(default_vfio_cfg, vfio_va, vfio_va,
530 vfio_dma_mem_map(default_vfio_cfg, vfio_va, vfio_va,
535 #ifdef RTE_ARCH_PPC_64
536 ms = rte_mem_virt2memseg(addr, msl);
537 while (cur_len < len) {
538 int idx = rte_fbarray_find_idx(&msl->memseg_arr, ms);
540 rte_fbarray_set_free(&msl->memseg_arr, idx);
546 /* memsegs are contiguous in memory */
547 ms = rte_mem_virt2memseg(addr, msl);
548 while (cur_len < len) {
549 /* some memory segments may have invalid IOVA */
550 if (ms->iova == RTE_BAD_IOVA) {
551 RTE_LOG(DEBUG, EAL, "Memory segment at %p has bad IOVA, skipping\n",
555 if (type == RTE_MEM_EVENT_ALLOC)
556 vfio_dma_mem_map(default_vfio_cfg, ms->addr_64,
557 ms->iova, ms->len, 1);
559 vfio_dma_mem_map(default_vfio_cfg, ms->addr_64,
560 ms->iova, ms->len, 0);
565 #ifdef RTE_ARCH_PPC_64
567 ms = rte_mem_virt2memseg(addr, msl);
568 while (cur_len < len) {
569 int idx = rte_fbarray_find_idx(&msl->memseg_arr, ms);
571 rte_fbarray_set_used(&msl->memseg_arr, idx);
579 vfio_sync_default_container(void)
581 struct rte_mp_msg mp_req, *mp_rep;
582 struct rte_mp_reply mp_reply = {0};
583 struct timespec ts = {.tv_sec = 5, .tv_nsec = 0};
584 struct vfio_mp_param *p = (struct vfio_mp_param *)mp_req.param;
588 /* cannot be called from primary */
589 if (rte_eal_process_type() != RTE_PROC_SECONDARY)
592 /* default container fd should have been opened in rte_vfio_enable() */
593 if (!default_vfio_cfg->vfio_enabled ||
594 default_vfio_cfg->vfio_container_fd < 0) {
595 RTE_LOG(ERR, EAL, "VFIO support is not initialized\n");
599 /* find default container's IOMMU type */
600 p->req = SOCKET_REQ_IOMMU_TYPE;
601 strcpy(mp_req.name, EAL_VFIO_MP);
602 mp_req.len_param = sizeof(*p);
606 if (rte_mp_request_sync(&mp_req, &mp_reply, &ts) == 0 &&
607 mp_reply.nb_received == 1) {
608 mp_rep = &mp_reply.msgs[0];
609 p = (struct vfio_mp_param *)mp_rep->param;
610 if (p->result == SOCKET_OK)
611 iommu_type_id = p->iommu_type_id;
614 if (iommu_type_id < 0) {
615 RTE_LOG(ERR, EAL, "Could not get IOMMU type for default container\n");
619 /* we now have an fd for default container, as well as its IOMMU type.
620 * now, set up default VFIO container config to match.
622 for (i = 0; i < RTE_DIM(iommu_types); i++) {
623 const struct vfio_iommu_type *t = &iommu_types[i];
624 if (t->type_id != iommu_type_id)
627 /* we found our IOMMU type */
628 default_vfio_cfg->vfio_iommu_type = t;
632 RTE_LOG(ERR, EAL, "Could not find IOMMU type id (%i)\n",
638 rte_vfio_clear_group(int vfio_group_fd)
641 struct vfio_config *vfio_cfg;
643 vfio_cfg = get_vfio_cfg_by_group_fd(vfio_group_fd);
644 if (vfio_cfg == NULL) {
645 RTE_LOG(ERR, EAL, " invalid group fd!\n");
649 i = get_vfio_group_idx(vfio_group_fd);
652 vfio_cfg->vfio_groups[i].group_num = -1;
653 vfio_cfg->vfio_groups[i].fd = -1;
654 vfio_cfg->vfio_groups[i].devices = 0;
655 vfio_cfg->vfio_active_groups--;
661 rte_vfio_setup_device(const char *sysfs_base, const char *dev_addr,
662 int *vfio_dev_fd, struct vfio_device_info *device_info)
664 struct vfio_group_status group_status = {
665 .argsz = sizeof(group_status)
667 struct vfio_config *vfio_cfg;
668 struct user_mem_maps *user_mem_maps;
669 int vfio_container_fd;
674 /* get group number */
675 ret = rte_vfio_get_group_num(sysfs_base, dev_addr, &iommu_group_num);
677 RTE_LOG(WARNING, EAL, " %s not managed by VFIO driver, skipping\n",
682 /* if negative, something failed */
686 /* get the actual group fd */
687 vfio_group_fd = rte_vfio_get_group_fd(iommu_group_num);
688 if (vfio_group_fd < 0)
691 /* if group_fd == 0, that means the device isn't managed by VFIO */
692 if (vfio_group_fd == 0) {
693 RTE_LOG(WARNING, EAL, " %s not managed by VFIO driver, skipping\n",
699 * at this point, we know that this group is viable (meaning, all devices
700 * are either bound to VFIO or not bound to anything)
703 /* check if the group is viable */
704 ret = ioctl(vfio_group_fd, VFIO_GROUP_GET_STATUS, &group_status);
706 RTE_LOG(ERR, EAL, " %s cannot get group status, "
707 "error %i (%s)\n", dev_addr, errno, strerror(errno));
708 close(vfio_group_fd);
709 rte_vfio_clear_group(vfio_group_fd);
711 } else if (!(group_status.flags & VFIO_GROUP_FLAGS_VIABLE)) {
712 RTE_LOG(ERR, EAL, " %s VFIO group is not viable! "
713 "Not all devices in IOMMU group bound to VFIO or unbound\n",
715 close(vfio_group_fd);
716 rte_vfio_clear_group(vfio_group_fd);
720 /* get the vfio_config it belongs to */
721 vfio_cfg = get_vfio_cfg_by_group_num(iommu_group_num);
722 vfio_cfg = vfio_cfg ? vfio_cfg : default_vfio_cfg;
723 vfio_container_fd = vfio_cfg->vfio_container_fd;
724 user_mem_maps = &vfio_cfg->mem_maps;
726 /* check if group does not have a container yet */
727 if (!(group_status.flags & VFIO_GROUP_FLAGS_CONTAINER_SET)) {
729 /* add group to a container */
730 ret = ioctl(vfio_group_fd, VFIO_GROUP_SET_CONTAINER,
733 RTE_LOG(ERR, EAL, " %s cannot add VFIO group to container, "
734 "error %i (%s)\n", dev_addr, errno, strerror(errno));
735 close(vfio_group_fd);
736 rte_vfio_clear_group(vfio_group_fd);
741 * pick an IOMMU type and set up DMA mappings for container
743 * needs to be done only once, only when first group is
744 * assigned to a container and only in primary process.
745 * Note this can happen several times with the hotplug
748 if (internal_config.process_type == RTE_PROC_PRIMARY &&
749 vfio_cfg->vfio_active_groups == 1 &&
750 vfio_group_device_count(vfio_group_fd) == 0) {
751 const struct vfio_iommu_type *t;
753 /* select an IOMMU type which we will be using */
754 t = vfio_set_iommu_type(vfio_container_fd);
757 " %s failed to select IOMMU type\n",
759 close(vfio_group_fd);
760 rte_vfio_clear_group(vfio_group_fd);
763 /* lock memory hotplug before mapping and release it
764 * after registering callback, to prevent races
766 rte_mcfg_mem_read_lock();
767 if (vfio_cfg == default_vfio_cfg)
768 ret = t->dma_map_func(vfio_container_fd);
773 " %s DMA remapping failed, error %i (%s)\n",
774 dev_addr, errno, strerror(errno));
775 close(vfio_group_fd);
776 rte_vfio_clear_group(vfio_group_fd);
777 rte_mcfg_mem_read_unlock();
781 vfio_cfg->vfio_iommu_type = t;
783 /* re-map all user-mapped segments */
784 rte_spinlock_recursive_lock(&user_mem_maps->lock);
786 /* this IOMMU type may not support DMA mapping, but
787 * if we have mappings in the list - that means we have
788 * previously mapped something successfully, so we can
789 * be sure that DMA mapping is supported.
791 for (i = 0; i < user_mem_maps->n_maps; i++) {
792 struct user_mem_map *map;
793 map = &user_mem_maps->maps[i];
795 ret = t->dma_user_map_func(
797 map->addr, map->iova, map->len,
800 RTE_LOG(ERR, EAL, "Couldn't map user memory for DMA: "
802 "iova: 0x%" PRIx64 " "
803 "len: 0x%" PRIu64 "\n",
804 map->addr, map->iova,
806 rte_spinlock_recursive_unlock(
807 &user_mem_maps->lock);
808 rte_mcfg_mem_read_unlock();
812 rte_spinlock_recursive_unlock(&user_mem_maps->lock);
814 /* register callback for mem events */
815 if (vfio_cfg == default_vfio_cfg)
816 ret = rte_mem_event_callback_register(
817 VFIO_MEM_EVENT_CLB_NAME,
818 vfio_mem_event_callback, NULL);
821 /* unlock memory hotplug */
822 rte_mcfg_mem_read_unlock();
824 if (ret && rte_errno != ENOTSUP) {
825 RTE_LOG(ERR, EAL, "Could not install memory event callback for VFIO\n");
829 RTE_LOG(DEBUG, EAL, "Memory event callbacks not supported\n");
831 RTE_LOG(DEBUG, EAL, "Installed memory event callback for VFIO\n");
833 } else if (rte_eal_process_type() != RTE_PROC_PRIMARY &&
834 vfio_cfg == default_vfio_cfg &&
835 vfio_cfg->vfio_iommu_type == NULL) {
836 /* if we're not a primary process, we do not set up the VFIO
837 * container because it's already been set up by the primary
838 * process. instead, we simply ask the primary about VFIO type
839 * we are using, and set the VFIO config up appropriately.
841 ret = vfio_sync_default_container();
843 RTE_LOG(ERR, EAL, "Could not sync default VFIO container\n");
844 close(vfio_group_fd);
845 rte_vfio_clear_group(vfio_group_fd);
848 /* we have successfully initialized VFIO, notify user */
849 const struct vfio_iommu_type *t =
850 default_vfio_cfg->vfio_iommu_type;
851 RTE_LOG(NOTICE, EAL, " using IOMMU type %d (%s)\n",
852 t->type_id, t->name);
855 /* get a file descriptor for the device */
856 *vfio_dev_fd = ioctl(vfio_group_fd, VFIO_GROUP_GET_DEVICE_FD, dev_addr);
857 if (*vfio_dev_fd < 0) {
858 /* if we cannot get a device fd, this implies a problem with
859 * the VFIO group or the container not having IOMMU configured.
862 RTE_LOG(WARNING, EAL, "Getting a vfio_dev_fd for %s failed\n",
864 close(vfio_group_fd);
865 rte_vfio_clear_group(vfio_group_fd);
869 /* test and setup the device */
870 ret = ioctl(*vfio_dev_fd, VFIO_DEVICE_GET_INFO, device_info);
872 RTE_LOG(ERR, EAL, " %s cannot get device info, "
873 "error %i (%s)\n", dev_addr, errno,
876 close(vfio_group_fd);
877 rte_vfio_clear_group(vfio_group_fd);
880 vfio_group_device_get(vfio_group_fd);
886 rte_vfio_release_device(const char *sysfs_base, const char *dev_addr,
889 struct vfio_group_status group_status = {
890 .argsz = sizeof(group_status)
892 struct vfio_config *vfio_cfg;
897 /* we don't want any DMA mapping messages to come while we're detaching
898 * VFIO device, because this might be the last device and we might need
899 * to unregister the callback.
901 rte_mcfg_mem_read_lock();
903 /* get group number */
904 ret = rte_vfio_get_group_num(sysfs_base, dev_addr, &iommu_group_num);
906 RTE_LOG(WARNING, EAL, " %s not managed by VFIO driver\n",
908 /* This is an error at this point. */
913 /* get the actual group fd */
914 vfio_group_fd = rte_vfio_get_group_fd(iommu_group_num);
915 if (vfio_group_fd <= 0) {
916 RTE_LOG(INFO, EAL, "rte_vfio_get_group_fd failed for %s\n",
922 /* get the vfio_config it belongs to */
923 vfio_cfg = get_vfio_cfg_by_group_num(iommu_group_num);
924 vfio_cfg = vfio_cfg ? vfio_cfg : default_vfio_cfg;
926 /* At this point we got an active group. Closing it will make the
927 * container detachment. If this is the last active group, VFIO kernel
928 * code will unset the container and the IOMMU mappings.
931 /* Closing a device */
932 if (close(vfio_dev_fd) < 0) {
933 RTE_LOG(INFO, EAL, "Error when closing vfio_dev_fd for %s\n",
939 /* An VFIO group can have several devices attached. Just when there is
940 * no devices remaining should the group be closed.
942 vfio_group_device_put(vfio_group_fd);
943 if (!vfio_group_device_count(vfio_group_fd)) {
945 if (close(vfio_group_fd) < 0) {
946 RTE_LOG(INFO, EAL, "Error when closing vfio_group_fd for %s\n",
952 if (rte_vfio_clear_group(vfio_group_fd) < 0) {
953 RTE_LOG(INFO, EAL, "Error when clearing group for %s\n",
960 /* if there are no active device groups, unregister the callback to
961 * avoid spurious attempts to map/unmap memory from VFIO.
963 if (vfio_cfg == default_vfio_cfg && vfio_cfg->vfio_active_groups == 0 &&
964 rte_eal_process_type() != RTE_PROC_SECONDARY)
965 rte_mem_event_callback_unregister(VFIO_MEM_EVENT_CLB_NAME,
972 rte_mcfg_mem_read_unlock();
977 rte_vfio_enable(const char *modname)
979 /* initialize group list */
983 rte_spinlock_recursive_t lock = RTE_SPINLOCK_RECURSIVE_INITIALIZER;
985 for (i = 0; i < VFIO_MAX_CONTAINERS; i++) {
986 vfio_cfgs[i].vfio_container_fd = -1;
987 vfio_cfgs[i].vfio_active_groups = 0;
988 vfio_cfgs[i].vfio_iommu_type = NULL;
989 vfio_cfgs[i].mem_maps.lock = lock;
991 for (j = 0; j < VFIO_MAX_GROUPS; j++) {
992 vfio_cfgs[i].vfio_groups[j].fd = -1;
993 vfio_cfgs[i].vfio_groups[j].group_num = -1;
994 vfio_cfgs[i].vfio_groups[j].devices = 0;
998 /* inform the user that we are probing for VFIO */
999 RTE_LOG(INFO, EAL, "Probing VFIO support...\n");
1001 /* check if vfio module is loaded */
1002 vfio_available = rte_eal_check_module(modname);
1004 /* return error directly */
1005 if (vfio_available == -1) {
1006 RTE_LOG(INFO, EAL, "Could not get loaded module details!\n");
1010 /* return 0 if VFIO modules not loaded */
1011 if (vfio_available == 0) {
1012 RTE_LOG(DEBUG, EAL, "VFIO modules not loaded, "
1013 "skipping VFIO support...\n");
1017 if (internal_config.process_type == RTE_PROC_PRIMARY) {
1018 /* open a new container */
1019 default_vfio_cfg->vfio_container_fd =
1020 rte_vfio_get_container_fd();
1022 /* get the default container from the primary process */
1023 default_vfio_cfg->vfio_container_fd =
1024 vfio_get_default_container_fd();
1027 /* check if we have VFIO driver enabled */
1028 if (default_vfio_cfg->vfio_container_fd != -1) {
1029 RTE_LOG(NOTICE, EAL, "VFIO support initialized\n");
1030 default_vfio_cfg->vfio_enabled = 1;
1032 RTE_LOG(NOTICE, EAL, "VFIO support could not be initialized\n");
1039 rte_vfio_is_enabled(const char *modname)
1041 const int mod_available = rte_eal_check_module(modname) > 0;
1042 return default_vfio_cfg->vfio_enabled && mod_available;
1046 vfio_get_default_container_fd(void)
1048 struct rte_mp_msg mp_req, *mp_rep;
1049 struct rte_mp_reply mp_reply = {0};
1050 struct timespec ts = {.tv_sec = 5, .tv_nsec = 0};
1051 struct vfio_mp_param *p = (struct vfio_mp_param *)mp_req.param;
1053 if (default_vfio_cfg->vfio_enabled)
1054 return default_vfio_cfg->vfio_container_fd;
1056 if (internal_config.process_type == RTE_PROC_PRIMARY) {
1057 /* if we were secondary process we would try requesting
1058 * container fd from the primary, but we're the primary
1059 * process so just exit here
1064 p->req = SOCKET_REQ_DEFAULT_CONTAINER;
1065 strcpy(mp_req.name, EAL_VFIO_MP);
1066 mp_req.len_param = sizeof(*p);
1069 if (rte_mp_request_sync(&mp_req, &mp_reply, &ts) == 0 &&
1070 mp_reply.nb_received == 1) {
1071 mp_rep = &mp_reply.msgs[0];
1072 p = (struct vfio_mp_param *)mp_rep->param;
1073 if (p->result == SOCKET_OK && mp_rep->num_fds == 1) {
1074 free(mp_reply.msgs);
1075 return mp_rep->fds[0];
1079 free(mp_reply.msgs);
1080 RTE_LOG(ERR, EAL, " cannot request default container fd\n");
1085 vfio_get_iommu_type(void)
1087 if (default_vfio_cfg->vfio_iommu_type == NULL)
1090 return default_vfio_cfg->vfio_iommu_type->type_id;
1093 const struct vfio_iommu_type *
1094 vfio_set_iommu_type(int vfio_container_fd)
1097 for (idx = 0; idx < RTE_DIM(iommu_types); idx++) {
1098 const struct vfio_iommu_type *t = &iommu_types[idx];
1100 int ret = ioctl(vfio_container_fd, VFIO_SET_IOMMU,
1103 RTE_LOG(NOTICE, EAL, " using IOMMU type %d (%s)\n",
1104 t->type_id, t->name);
1107 /* not an error, there may be more supported IOMMU types */
1108 RTE_LOG(DEBUG, EAL, " set IOMMU type %d (%s) failed, "
1109 "error %i (%s)\n", t->type_id, t->name, errno,
1112 /* if we didn't find a suitable IOMMU type, fail */
1117 vfio_has_supported_extensions(int vfio_container_fd)
1120 unsigned idx, n_extensions = 0;
1121 for (idx = 0; idx < RTE_DIM(iommu_types); idx++) {
1122 const struct vfio_iommu_type *t = &iommu_types[idx];
1124 ret = ioctl(vfio_container_fd, VFIO_CHECK_EXTENSION,
1127 RTE_LOG(ERR, EAL, " could not get IOMMU type, "
1128 "error %i (%s)\n", errno,
1130 close(vfio_container_fd);
1132 } else if (ret == 1) {
1133 /* we found a supported extension */
1136 RTE_LOG(DEBUG, EAL, " IOMMU type %d (%s) is %s\n",
1137 t->type_id, t->name,
1138 ret ? "supported" : "not supported");
1141 /* if we didn't find any supported IOMMU types, fail */
1142 if (!n_extensions) {
1143 close(vfio_container_fd);
1151 rte_vfio_get_container_fd(void)
1153 int ret, vfio_container_fd;
1154 struct rte_mp_msg mp_req, *mp_rep;
1155 struct rte_mp_reply mp_reply = {0};
1156 struct timespec ts = {.tv_sec = 5, .tv_nsec = 0};
1157 struct vfio_mp_param *p = (struct vfio_mp_param *)mp_req.param;
1160 /* if we're in a primary process, try to open the container */
1161 if (internal_config.process_type == RTE_PROC_PRIMARY) {
1162 vfio_container_fd = open(VFIO_CONTAINER_PATH, O_RDWR);
1163 if (vfio_container_fd < 0) {
1164 RTE_LOG(ERR, EAL, " cannot open VFIO container, "
1165 "error %i (%s)\n", errno, strerror(errno));
1169 /* check VFIO API version */
1170 ret = ioctl(vfio_container_fd, VFIO_GET_API_VERSION);
1171 if (ret != VFIO_API_VERSION) {
1173 RTE_LOG(ERR, EAL, " could not get VFIO API version, "
1174 "error %i (%s)\n", errno, strerror(errno));
1176 RTE_LOG(ERR, EAL, " unsupported VFIO API version!\n");
1177 close(vfio_container_fd);
1181 ret = vfio_has_supported_extensions(vfio_container_fd);
1183 RTE_LOG(ERR, EAL, " no supported IOMMU "
1184 "extensions found!\n");
1188 return vfio_container_fd;
1191 * if we're in a secondary process, request container fd from the
1192 * primary process via mp channel
1194 p->req = SOCKET_REQ_CONTAINER;
1195 strcpy(mp_req.name, EAL_VFIO_MP);
1196 mp_req.len_param = sizeof(*p);
1199 vfio_container_fd = -1;
1200 if (rte_mp_request_sync(&mp_req, &mp_reply, &ts) == 0 &&
1201 mp_reply.nb_received == 1) {
1202 mp_rep = &mp_reply.msgs[0];
1203 p = (struct vfio_mp_param *)mp_rep->param;
1204 if (p->result == SOCKET_OK && mp_rep->num_fds == 1) {
1205 vfio_container_fd = mp_rep->fds[0];
1206 free(mp_reply.msgs);
1207 return vfio_container_fd;
1211 free(mp_reply.msgs);
1212 RTE_LOG(ERR, EAL, " cannot request container fd\n");
1217 rte_vfio_get_group_num(const char *sysfs_base,
1218 const char *dev_addr, int *iommu_group_num)
1220 char linkname[PATH_MAX];
1221 char filename[PATH_MAX];
1222 char *tok[16], *group_tok, *end;
1225 memset(linkname, 0, sizeof(linkname));
1226 memset(filename, 0, sizeof(filename));
1228 /* try to find out IOMMU group for this device */
1229 snprintf(linkname, sizeof(linkname),
1230 "%s/%s/iommu_group", sysfs_base, dev_addr);
1232 ret = readlink(linkname, filename, sizeof(filename));
1234 /* if the link doesn't exist, no VFIO for us */
1238 ret = rte_strsplit(filename, sizeof(filename),
1239 tok, RTE_DIM(tok), '/');
1242 RTE_LOG(ERR, EAL, " %s cannot get IOMMU group\n", dev_addr);
1246 /* IOMMU group is always the last token */
1248 group_tok = tok[ret - 1];
1250 *iommu_group_num = strtol(group_tok, &end, 10);
1251 if ((end != group_tok && *end != '\0') || errno != 0) {
1252 RTE_LOG(ERR, EAL, " %s error parsing IOMMU number!\n", dev_addr);
1260 type1_map_contig(const struct rte_memseg_list *msl, const struct rte_memseg *ms,
1261 size_t len, void *arg)
1263 int *vfio_container_fd = arg;
1268 return vfio_type1_dma_mem_map(*vfio_container_fd, ms->addr_64, ms->iova,
1273 type1_map(const struct rte_memseg_list *msl, const struct rte_memseg *ms,
1276 int *vfio_container_fd = arg;
1278 /* skip external memory that isn't a heap */
1279 if (msl->external && !msl->heap)
1282 /* skip any segments with invalid IOVA addresses */
1283 if (ms->iova == RTE_BAD_IOVA)
1286 /* if IOVA mode is VA, we've already mapped the internal segments */
1287 if (!msl->external && rte_eal_iova_mode() == RTE_IOVA_VA)
1290 return vfio_type1_dma_mem_map(*vfio_container_fd, ms->addr_64, ms->iova,
1295 vfio_type1_dma_mem_map(int vfio_container_fd, uint64_t vaddr, uint64_t iova,
1296 uint64_t len, int do_map)
1298 struct vfio_iommu_type1_dma_map dma_map;
1299 struct vfio_iommu_type1_dma_unmap dma_unmap;
1303 memset(&dma_map, 0, sizeof(dma_map));
1304 dma_map.argsz = sizeof(struct vfio_iommu_type1_dma_map);
1305 dma_map.vaddr = vaddr;
1307 dma_map.iova = iova;
1308 dma_map.flags = VFIO_DMA_MAP_FLAG_READ |
1309 VFIO_DMA_MAP_FLAG_WRITE;
1311 ret = ioctl(vfio_container_fd, VFIO_IOMMU_MAP_DMA, &dma_map);
1314 * In case the mapping was already done EEXIST will be
1315 * returned from kernel.
1317 if (errno == EEXIST) {
1319 " Memory segment is already mapped,"
1323 " cannot set up DMA remapping,"
1325 errno, strerror(errno));
1330 memset(&dma_unmap, 0, sizeof(dma_unmap));
1331 dma_unmap.argsz = sizeof(struct vfio_iommu_type1_dma_unmap);
1332 dma_unmap.size = len;
1333 dma_unmap.iova = iova;
1335 ret = ioctl(vfio_container_fd, VFIO_IOMMU_UNMAP_DMA,
1338 RTE_LOG(ERR, EAL, " cannot clear DMA remapping, error %i (%s)\n",
1339 errno, strerror(errno));
1348 vfio_type1_dma_map(int vfio_container_fd)
1350 if (rte_eal_iova_mode() == RTE_IOVA_VA) {
1351 /* with IOVA as VA mode, we can get away with mapping contiguous
1352 * chunks rather than going page-by-page.
1354 int ret = rte_memseg_contig_walk(type1_map_contig,
1355 &vfio_container_fd);
1358 /* we have to continue the walk because we've skipped the
1359 * external segments during the config walk.
1362 return rte_memseg_walk(type1_map, &vfio_container_fd);
1366 vfio_spapr_dma_do_map(int vfio_container_fd, uint64_t vaddr, uint64_t iova,
1367 uint64_t len, int do_map)
1369 struct vfio_iommu_type1_dma_map dma_map;
1370 struct vfio_iommu_type1_dma_unmap dma_unmap;
1372 struct vfio_iommu_spapr_register_memory reg = {
1373 .argsz = sizeof(reg),
1376 reg.vaddr = (uintptr_t) vaddr;
1380 ret = ioctl(vfio_container_fd,
1381 VFIO_IOMMU_SPAPR_REGISTER_MEMORY, ®);
1383 RTE_LOG(ERR, EAL, " cannot register vaddr for IOMMU, "
1384 "error %i (%s)\n", errno, strerror(errno));
1388 memset(&dma_map, 0, sizeof(dma_map));
1389 dma_map.argsz = sizeof(struct vfio_iommu_type1_dma_map);
1390 dma_map.vaddr = vaddr;
1392 dma_map.iova = iova;
1393 dma_map.flags = VFIO_DMA_MAP_FLAG_READ |
1394 VFIO_DMA_MAP_FLAG_WRITE;
1396 ret = ioctl(vfio_container_fd, VFIO_IOMMU_MAP_DMA, &dma_map);
1399 * In case the mapping was already done EBUSY will be
1400 * returned from kernel.
1402 if (errno == EBUSY) {
1404 " Memory segment is already mapped,"
1408 " cannot set up DMA remapping,"
1409 " error %i (%s)\n", errno,
1416 memset(&dma_unmap, 0, sizeof(dma_unmap));
1417 dma_unmap.argsz = sizeof(struct vfio_iommu_type1_dma_unmap);
1418 dma_unmap.size = len;
1419 dma_unmap.iova = iova;
1421 ret = ioctl(vfio_container_fd, VFIO_IOMMU_UNMAP_DMA,
1424 RTE_LOG(ERR, EAL, " cannot clear DMA remapping, error %i (%s)\n",
1425 errno, strerror(errno));
1429 ret = ioctl(vfio_container_fd,
1430 VFIO_IOMMU_SPAPR_UNREGISTER_MEMORY, ®);
1432 RTE_LOG(ERR, EAL, " cannot unregister vaddr for IOMMU, error %i (%s)\n",
1433 errno, strerror(errno));
1442 vfio_spapr_map_walk(const struct rte_memseg_list *msl,
1443 const struct rte_memseg *ms, void *arg)
1445 int *vfio_container_fd = arg;
1447 /* skip external memory that isn't a heap */
1448 if (msl->external && !msl->heap)
1451 /* skip any segments with invalid IOVA addresses */
1452 if (ms->iova == RTE_BAD_IOVA)
1455 return vfio_spapr_dma_do_map(*vfio_container_fd, ms->addr_64, ms->iova,
1460 vfio_spapr_unmap_walk(const struct rte_memseg_list *msl,
1461 const struct rte_memseg *ms, void *arg)
1463 int *vfio_container_fd = arg;
1465 /* skip external memory that isn't a heap */
1466 if (msl->external && !msl->heap)
1469 /* skip any segments with invalid IOVA addresses */
1470 if (ms->iova == RTE_BAD_IOVA)
1473 return vfio_spapr_dma_do_map(*vfio_container_fd, ms->addr_64, ms->iova,
1477 struct spapr_walk_param {
1478 uint64_t window_size;
1479 uint64_t hugepage_sz;
1483 vfio_spapr_window_size_walk(const struct rte_memseg_list *msl,
1484 const struct rte_memseg *ms, void *arg)
1486 struct spapr_walk_param *param = arg;
1487 uint64_t max = ms->iova + ms->len;
1489 /* skip external memory that isn't a heap */
1490 if (msl->external && !msl->heap)
1493 /* skip any segments with invalid IOVA addresses */
1494 if (ms->iova == RTE_BAD_IOVA)
1497 if (max > param->window_size) {
1498 param->hugepage_sz = ms->hugepage_sz;
1499 param->window_size = max;
1506 vfio_spapr_create_new_dma_window(int vfio_container_fd,
1507 struct vfio_iommu_spapr_tce_create *create) {
1508 struct vfio_iommu_spapr_tce_remove remove = {
1509 .argsz = sizeof(remove),
1511 struct vfio_iommu_spapr_tce_info info = {
1512 .argsz = sizeof(info),
1516 /* query spapr iommu info */
1517 ret = ioctl(vfio_container_fd, VFIO_IOMMU_SPAPR_TCE_GET_INFO, &info);
1519 RTE_LOG(ERR, EAL, " cannot get iommu info, "
1520 "error %i (%s)\n", errno, strerror(errno));
1524 /* remove default DMA of 32 bit window */
1525 remove.start_addr = info.dma32_window_start;
1526 ret = ioctl(vfio_container_fd, VFIO_IOMMU_SPAPR_TCE_REMOVE, &remove);
1528 RTE_LOG(ERR, EAL, " cannot remove default DMA window, "
1529 "error %i (%s)\n", errno, strerror(errno));
1533 /* create new DMA window */
1534 ret = ioctl(vfio_container_fd, VFIO_IOMMU_SPAPR_TCE_CREATE, create);
1536 #ifdef VFIO_IOMMU_SPAPR_INFO_DDW
1537 /* try possible page_shift and levels for workaround */
1540 for (levels = create->levels + 1;
1541 ret && levels <= info.ddw.levels; levels++) {
1542 create->levels = levels;
1543 ret = ioctl(vfio_container_fd,
1544 VFIO_IOMMU_SPAPR_TCE_CREATE, create);
1548 RTE_LOG(ERR, EAL, " cannot create new DMA window, "
1549 "error %i (%s)\n", errno, strerror(errno));
1554 if (create->start_addr != 0) {
1555 RTE_LOG(ERR, EAL, " DMA window start address != 0\n");
1563 vfio_spapr_dma_mem_map(int vfio_container_fd, uint64_t vaddr, uint64_t iova,
1564 uint64_t len, int do_map)
1566 struct spapr_walk_param param;
1567 struct vfio_iommu_spapr_tce_create create = {
1568 .argsz = sizeof(create),
1570 struct vfio_config *vfio_cfg;
1571 struct user_mem_maps *user_mem_maps;
1574 vfio_cfg = get_vfio_cfg_by_container_fd(vfio_container_fd);
1575 if (vfio_cfg == NULL) {
1576 RTE_LOG(ERR, EAL, " invalid container fd!\n");
1580 user_mem_maps = &vfio_cfg->mem_maps;
1581 rte_spinlock_recursive_lock(&user_mem_maps->lock);
1583 /* check if window size needs to be adjusted */
1584 memset(¶m, 0, sizeof(param));
1586 /* we're inside a callback so use thread-unsafe version */
1587 if (rte_memseg_walk_thread_unsafe(vfio_spapr_window_size_walk,
1589 RTE_LOG(ERR, EAL, "Could not get window size\n");
1594 /* also check user maps */
1595 for (i = 0; i < user_mem_maps->n_maps; i++) {
1596 uint64_t max = user_mem_maps->maps[i].iova +
1597 user_mem_maps->maps[i].len;
1598 param.window_size = RTE_MAX(param.window_size, max);
1601 /* sPAPR requires window size to be a power of 2 */
1602 create.window_size = rte_align64pow2(param.window_size);
1603 create.page_shift = __builtin_ctzll(param.hugepage_sz);
1607 /* re-create window and remap the entire memory */
1608 if (iova + len > create.window_size) {
1609 /* release all maps before recreating the window */
1610 if (rte_memseg_walk_thread_unsafe(vfio_spapr_unmap_walk,
1611 &vfio_container_fd) < 0) {
1612 RTE_LOG(ERR, EAL, "Could not release DMA maps\n");
1616 /* release all user maps */
1617 for (i = 0; i < user_mem_maps->n_maps; i++) {
1618 struct user_mem_map *map =
1619 &user_mem_maps->maps[i];
1620 if (vfio_spapr_dma_do_map(vfio_container_fd,
1621 map->addr, map->iova, map->len,
1623 RTE_LOG(ERR, EAL, "Could not release user DMA maps\n");
1628 create.window_size = rte_align64pow2(iova + len);
1629 if (vfio_spapr_create_new_dma_window(vfio_container_fd,
1631 RTE_LOG(ERR, EAL, "Could not create new DMA window\n");
1635 /* we're inside a callback, so use thread-unsafe version
1637 if (rte_memseg_walk_thread_unsafe(vfio_spapr_map_walk,
1638 &vfio_container_fd) < 0) {
1639 RTE_LOG(ERR, EAL, "Could not recreate DMA maps\n");
1643 /* remap all user maps */
1644 for (i = 0; i < user_mem_maps->n_maps; i++) {
1645 struct user_mem_map *map =
1646 &user_mem_maps->maps[i];
1647 if (vfio_spapr_dma_do_map(vfio_container_fd,
1648 map->addr, map->iova, map->len,
1650 RTE_LOG(ERR, EAL, "Could not recreate user DMA maps\n");
1656 if (vfio_spapr_dma_do_map(vfio_container_fd, vaddr, iova, len, 1)) {
1657 RTE_LOG(ERR, EAL, "Failed to map DMA\n");
1662 /* for unmap, check if iova within DMA window */
1663 if (iova > create.window_size) {
1664 RTE_LOG(ERR, EAL, "iova beyond DMA window for unmap");
1669 vfio_spapr_dma_do_map(vfio_container_fd, vaddr, iova, len, 0);
1672 rte_spinlock_recursive_unlock(&user_mem_maps->lock);
1677 vfio_spapr_dma_map(int vfio_container_fd)
1679 struct vfio_iommu_spapr_tce_create create = {
1680 .argsz = sizeof(create),
1682 struct spapr_walk_param param;
1684 memset(¶m, 0, sizeof(param));
1686 /* create DMA window from 0 to max(phys_addr + len) */
1687 rte_memseg_walk(vfio_spapr_window_size_walk, ¶m);
1689 /* sPAPR requires window size to be a power of 2 */
1690 create.window_size = rte_align64pow2(param.window_size);
1691 create.page_shift = __builtin_ctzll(param.hugepage_sz);
1694 if (vfio_spapr_create_new_dma_window(vfio_container_fd, &create) < 0) {
1695 RTE_LOG(ERR, EAL, "Could not create new DMA window\n");
1699 /* map all DPDK segments for DMA. use 1:1 PA to IOVA mapping */
1700 if (rte_memseg_walk(vfio_spapr_map_walk, &vfio_container_fd) < 0)
1707 vfio_noiommu_dma_map(int __rte_unused vfio_container_fd)
1709 /* No-IOMMU mode does not need DMA mapping */
1714 vfio_noiommu_dma_mem_map(int __rte_unused vfio_container_fd,
1715 uint64_t __rte_unused vaddr,
1716 uint64_t __rte_unused iova, uint64_t __rte_unused len,
1717 int __rte_unused do_map)
1719 /* No-IOMMU mode does not need DMA mapping */
1724 vfio_dma_mem_map(struct vfio_config *vfio_cfg, uint64_t vaddr, uint64_t iova,
1725 uint64_t len, int do_map)
1727 const struct vfio_iommu_type *t = vfio_cfg->vfio_iommu_type;
1730 RTE_LOG(ERR, EAL, " VFIO support not initialized\n");
1735 if (!t->dma_user_map_func) {
1737 " VFIO custom DMA region maping not supported by IOMMU %s\n",
1739 rte_errno = ENOTSUP;
1743 return t->dma_user_map_func(vfio_cfg->vfio_container_fd, vaddr, iova,
1748 container_dma_map(struct vfio_config *vfio_cfg, uint64_t vaddr, uint64_t iova,
1751 struct user_mem_map *new_map;
1752 struct user_mem_maps *user_mem_maps;
1755 user_mem_maps = &vfio_cfg->mem_maps;
1756 rte_spinlock_recursive_lock(&user_mem_maps->lock);
1757 if (user_mem_maps->n_maps == VFIO_MAX_USER_MEM_MAPS) {
1758 RTE_LOG(ERR, EAL, "No more space for user mem maps\n");
1764 if (vfio_dma_mem_map(vfio_cfg, vaddr, iova, len, 1)) {
1765 /* technically, this will fail if there are currently no devices
1766 * plugged in, even if a device were added later, this mapping
1767 * might have succeeded. however, since we cannot verify if this
1768 * is a valid mapping without having a device attached, consider
1769 * this to be unsupported, because we can't just store any old
1770 * mapping and pollute list of active mappings willy-nilly.
1772 RTE_LOG(ERR, EAL, "Couldn't map new region for DMA\n");
1776 /* create new user mem map entry */
1777 new_map = &user_mem_maps->maps[user_mem_maps->n_maps++];
1778 new_map->addr = vaddr;
1779 new_map->iova = iova;
1782 compact_user_maps(user_mem_maps);
1784 rte_spinlock_recursive_unlock(&user_mem_maps->lock);
1789 container_dma_unmap(struct vfio_config *vfio_cfg, uint64_t vaddr, uint64_t iova,
1792 struct user_mem_map *map, *new_map = NULL;
1793 struct user_mem_maps *user_mem_maps;
1796 user_mem_maps = &vfio_cfg->mem_maps;
1797 rte_spinlock_recursive_lock(&user_mem_maps->lock);
1799 /* find our mapping */
1800 map = find_user_mem_map(user_mem_maps, vaddr, iova, len);
1802 RTE_LOG(ERR, EAL, "Couldn't find previously mapped region\n");
1807 if (map->addr != vaddr || map->iova != iova || map->len != len) {
1808 /* we're partially unmapping a previously mapped region, so we
1809 * need to split entry into two.
1811 if (user_mem_maps->n_maps == VFIO_MAX_USER_MEM_MAPS) {
1812 RTE_LOG(ERR, EAL, "Not enough space to store partial mapping\n");
1817 new_map = &user_mem_maps->maps[user_mem_maps->n_maps++];
1820 /* unmap the entry */
1821 if (vfio_dma_mem_map(vfio_cfg, vaddr, iova, len, 0)) {
1822 /* there may not be any devices plugged in, so unmapping will
1823 * fail with ENODEV/ENOTSUP rte_errno values, but that doesn't
1824 * stop us from removing the mapping, as the assumption is we
1825 * won't be needing this memory any more and thus will want to
1826 * prevent it from being remapped again on hotplug. so, only
1827 * fail if we indeed failed to unmap (e.g. if the mapping was
1828 * within our mapped range but had invalid alignment).
1830 if (rte_errno != ENODEV && rte_errno != ENOTSUP) {
1831 RTE_LOG(ERR, EAL, "Couldn't unmap region for DMA\n");
1835 RTE_LOG(DEBUG, EAL, "DMA unmapping failed, but removing mappings anyway\n");
1838 /* remove map from the list of active mappings */
1839 if (new_map != NULL) {
1840 adjust_map(map, new_map, vaddr, len);
1842 /* if we've created a new map by splitting, sort everything */
1843 if (!is_null_map(new_map)) {
1844 compact_user_maps(user_mem_maps);
1846 /* we've created a new mapping, but it was unused */
1847 user_mem_maps->n_maps--;
1850 memset(map, 0, sizeof(*map));
1851 compact_user_maps(user_mem_maps);
1852 user_mem_maps->n_maps--;
1856 rte_spinlock_recursive_unlock(&user_mem_maps->lock);
1861 rte_vfio_noiommu_is_enabled(void)
1867 fd = open(VFIO_NOIOMMU_MODE, O_RDONLY);
1869 if (errno != ENOENT) {
1870 RTE_LOG(ERR, EAL, " cannot open vfio noiommu file %i (%s)\n",
1871 errno, strerror(errno));
1875 * else the file does not exists
1876 * i.e. noiommu is not enabled
1881 cnt = read(fd, &c, 1);
1884 RTE_LOG(ERR, EAL, " unable to read from vfio noiommu "
1885 "file %i (%s)\n", errno, strerror(errno));
1893 rte_vfio_container_create(void)
1897 /* Find an empty slot to store new vfio config */
1898 for (i = 1; i < VFIO_MAX_CONTAINERS; i++) {
1899 if (vfio_cfgs[i].vfio_container_fd == -1)
1903 if (i == VFIO_MAX_CONTAINERS) {
1904 RTE_LOG(ERR, EAL, "exceed max vfio container limit\n");
1908 vfio_cfgs[i].vfio_container_fd = rte_vfio_get_container_fd();
1909 if (vfio_cfgs[i].vfio_container_fd < 0) {
1910 RTE_LOG(NOTICE, EAL, "fail to create a new container\n");
1914 return vfio_cfgs[i].vfio_container_fd;
1918 rte_vfio_container_destroy(int container_fd)
1920 struct vfio_config *vfio_cfg;
1923 vfio_cfg = get_vfio_cfg_by_container_fd(container_fd);
1924 if (vfio_cfg == NULL) {
1925 RTE_LOG(ERR, EAL, "Invalid container fd\n");
1929 for (i = 0; i < VFIO_MAX_GROUPS; i++)
1930 if (vfio_cfg->vfio_groups[i].group_num != -1)
1931 rte_vfio_container_group_unbind(container_fd,
1932 vfio_cfg->vfio_groups[i].group_num);
1934 close(container_fd);
1935 vfio_cfg->vfio_container_fd = -1;
1936 vfio_cfg->vfio_active_groups = 0;
1937 vfio_cfg->vfio_iommu_type = NULL;
1943 rte_vfio_container_group_bind(int container_fd, int iommu_group_num)
1945 struct vfio_config *vfio_cfg;
1947 vfio_cfg = get_vfio_cfg_by_container_fd(container_fd);
1948 if (vfio_cfg == NULL) {
1949 RTE_LOG(ERR, EAL, "Invalid container fd\n");
1953 return vfio_get_group_fd(vfio_cfg, iommu_group_num);
1957 rte_vfio_container_group_unbind(int container_fd, int iommu_group_num)
1959 struct vfio_config *vfio_cfg;
1960 struct vfio_group *cur_grp = NULL;
1963 vfio_cfg = get_vfio_cfg_by_container_fd(container_fd);
1964 if (vfio_cfg == NULL) {
1965 RTE_LOG(ERR, EAL, "Invalid container fd\n");
1969 for (i = 0; i < VFIO_MAX_GROUPS; i++) {
1970 if (vfio_cfg->vfio_groups[i].group_num == iommu_group_num) {
1971 cur_grp = &vfio_cfg->vfio_groups[i];
1976 /* This should not happen */
1977 if (i == VFIO_MAX_GROUPS || cur_grp == NULL) {
1978 RTE_LOG(ERR, EAL, "Specified group number not found\n");
1982 if (cur_grp->fd >= 0 && close(cur_grp->fd) < 0) {
1983 RTE_LOG(ERR, EAL, "Error when closing vfio_group_fd for"
1984 " iommu_group_num %d\n", iommu_group_num);
1987 cur_grp->group_num = -1;
1989 cur_grp->devices = 0;
1990 vfio_cfg->vfio_active_groups--;
1996 rte_vfio_container_dma_map(int container_fd, uint64_t vaddr, uint64_t iova,
1999 struct vfio_config *vfio_cfg;
2006 vfio_cfg = get_vfio_cfg_by_container_fd(container_fd);
2007 if (vfio_cfg == NULL) {
2008 RTE_LOG(ERR, EAL, "Invalid container fd\n");
2012 return container_dma_map(vfio_cfg, vaddr, iova, len);
2016 rte_vfio_container_dma_unmap(int container_fd, uint64_t vaddr, uint64_t iova,
2019 struct vfio_config *vfio_cfg;
2026 vfio_cfg = get_vfio_cfg_by_container_fd(container_fd);
2027 if (vfio_cfg == NULL) {
2028 RTE_LOG(ERR, EAL, "Invalid container fd\n");
2032 return container_dma_unmap(vfio_cfg, vaddr, iova, len);
2038 rte_vfio_setup_device(__rte_unused const char *sysfs_base,
2039 __rte_unused const char *dev_addr,
2040 __rte_unused int *vfio_dev_fd,
2041 __rte_unused struct vfio_device_info *device_info)
2047 rte_vfio_release_device(__rte_unused const char *sysfs_base,
2048 __rte_unused const char *dev_addr, __rte_unused int fd)
2054 rte_vfio_enable(__rte_unused const char *modname)
2060 rte_vfio_is_enabled(__rte_unused const char *modname)
2066 rte_vfio_noiommu_is_enabled(void)
2072 rte_vfio_clear_group(__rte_unused int vfio_group_fd)
2078 rte_vfio_get_group_num(__rte_unused const char *sysfs_base,
2079 __rte_unused const char *dev_addr,
2080 __rte_unused int *iommu_group_num)
2086 rte_vfio_get_container_fd(void)
2092 rte_vfio_get_group_fd(__rte_unused int iommu_group_num)
2098 rte_vfio_container_create(void)
2104 rte_vfio_container_destroy(__rte_unused int container_fd)
2110 rte_vfio_container_group_bind(__rte_unused int container_fd,
2111 __rte_unused int iommu_group_num)
2117 rte_vfio_container_group_unbind(__rte_unused int container_fd,
2118 __rte_unused int iommu_group_num)
2124 rte_vfio_container_dma_map(__rte_unused int container_fd,
2125 __rte_unused uint64_t vaddr,
2126 __rte_unused uint64_t iova,
2127 __rte_unused uint64_t len)
2133 rte_vfio_container_dma_unmap(__rte_unused int container_fd,
2134 __rte_unused uint64_t vaddr,
2135 __rte_unused uint64_t iova,
2136 __rte_unused uint64_t len)
2141 #endif /* VFIO_PRESENT */