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 /* memsegs are contiguous in memory */
536 ms = rte_mem_virt2memseg(addr, msl);
537 while (cur_len < len) {
538 /* some memory segments may have invalid IOVA */
539 if (ms->iova == RTE_BAD_IOVA) {
540 RTE_LOG(DEBUG, EAL, "Memory segment at %p has bad IOVA, skipping\n",
544 if (type == RTE_MEM_EVENT_ALLOC)
545 vfio_dma_mem_map(default_vfio_cfg, ms->addr_64,
546 ms->iova, ms->len, 1);
548 vfio_dma_mem_map(default_vfio_cfg, ms->addr_64,
549 ms->iova, ms->len, 0);
557 vfio_sync_default_container(void)
559 struct rte_mp_msg mp_req, *mp_rep;
560 struct rte_mp_reply mp_reply = {0};
561 struct timespec ts = {.tv_sec = 5, .tv_nsec = 0};
562 struct vfio_mp_param *p = (struct vfio_mp_param *)mp_req.param;
566 /* cannot be called from primary */
567 if (rte_eal_process_type() != RTE_PROC_SECONDARY)
570 /* default container fd should have been opened in rte_vfio_enable() */
571 if (!default_vfio_cfg->vfio_enabled ||
572 default_vfio_cfg->vfio_container_fd < 0) {
573 RTE_LOG(ERR, EAL, "VFIO support is not initialized\n");
577 /* find default container's IOMMU type */
578 p->req = SOCKET_REQ_IOMMU_TYPE;
579 strcpy(mp_req.name, EAL_VFIO_MP);
580 mp_req.len_param = sizeof(*p);
584 if (rte_mp_request_sync(&mp_req, &mp_reply, &ts) == 0 &&
585 mp_reply.nb_received == 1) {
586 mp_rep = &mp_reply.msgs[0];
587 p = (struct vfio_mp_param *)mp_rep->param;
588 if (p->result == SOCKET_OK)
589 iommu_type_id = p->iommu_type_id;
592 if (iommu_type_id < 0) {
593 RTE_LOG(ERR, EAL, "Could not get IOMMU type for default container\n");
597 /* we now have an fd for default container, as well as its IOMMU type.
598 * now, set up default VFIO container config to match.
600 for (i = 0; i < RTE_DIM(iommu_types); i++) {
601 const struct vfio_iommu_type *t = &iommu_types[i];
602 if (t->type_id != iommu_type_id)
605 /* we found our IOMMU type */
606 default_vfio_cfg->vfio_iommu_type = t;
610 RTE_LOG(ERR, EAL, "Could not find IOMMU type id (%i)\n",
616 rte_vfio_clear_group(int vfio_group_fd)
619 struct vfio_config *vfio_cfg;
621 vfio_cfg = get_vfio_cfg_by_group_fd(vfio_group_fd);
622 if (vfio_cfg == NULL) {
623 RTE_LOG(ERR, EAL, " invalid group fd!\n");
627 i = get_vfio_group_idx(vfio_group_fd);
630 vfio_cfg->vfio_groups[i].group_num = -1;
631 vfio_cfg->vfio_groups[i].fd = -1;
632 vfio_cfg->vfio_groups[i].devices = 0;
633 vfio_cfg->vfio_active_groups--;
639 rte_vfio_setup_device(const char *sysfs_base, const char *dev_addr,
640 int *vfio_dev_fd, struct vfio_device_info *device_info)
642 struct vfio_group_status group_status = {
643 .argsz = sizeof(group_status)
645 struct vfio_config *vfio_cfg;
646 struct user_mem_maps *user_mem_maps;
647 int vfio_container_fd;
652 /* get group number */
653 ret = rte_vfio_get_group_num(sysfs_base, dev_addr, &iommu_group_num);
655 RTE_LOG(WARNING, EAL, " %s not managed by VFIO driver, skipping\n",
660 /* if negative, something failed */
664 /* get the actual group fd */
665 vfio_group_fd = rte_vfio_get_group_fd(iommu_group_num);
666 if (vfio_group_fd < 0)
669 /* if group_fd == 0, that means the device isn't managed by VFIO */
670 if (vfio_group_fd == 0) {
671 RTE_LOG(WARNING, EAL, " %s not managed by VFIO driver, skipping\n",
677 * at this point, we know that this group is viable (meaning, all devices
678 * are either bound to VFIO or not bound to anything)
681 /* check if the group is viable */
682 ret = ioctl(vfio_group_fd, VFIO_GROUP_GET_STATUS, &group_status);
684 RTE_LOG(ERR, EAL, " %s cannot get group status, "
685 "error %i (%s)\n", dev_addr, errno, strerror(errno));
686 close(vfio_group_fd);
687 rte_vfio_clear_group(vfio_group_fd);
689 } else if (!(group_status.flags & VFIO_GROUP_FLAGS_VIABLE)) {
690 RTE_LOG(ERR, EAL, " %s VFIO group is not viable! "
691 "Not all devices in IOMMU group bound to VFIO or unbound\n",
693 close(vfio_group_fd);
694 rte_vfio_clear_group(vfio_group_fd);
698 /* get the vfio_config it belongs to */
699 vfio_cfg = get_vfio_cfg_by_group_num(iommu_group_num);
700 vfio_cfg = vfio_cfg ? vfio_cfg : default_vfio_cfg;
701 vfio_container_fd = vfio_cfg->vfio_container_fd;
702 user_mem_maps = &vfio_cfg->mem_maps;
704 /* check if group does not have a container yet */
705 if (!(group_status.flags & VFIO_GROUP_FLAGS_CONTAINER_SET)) {
707 /* add group to a container */
708 ret = ioctl(vfio_group_fd, VFIO_GROUP_SET_CONTAINER,
711 RTE_LOG(ERR, EAL, " %s cannot add VFIO group to container, "
712 "error %i (%s)\n", dev_addr, errno, strerror(errno));
713 close(vfio_group_fd);
714 rte_vfio_clear_group(vfio_group_fd);
719 * pick an IOMMU type and set up DMA mappings for container
721 * needs to be done only once, only when first group is
722 * assigned to a container and only in primary process.
723 * Note this can happen several times with the hotplug
726 if (internal_config.process_type == RTE_PROC_PRIMARY &&
727 vfio_cfg->vfio_active_groups == 1 &&
728 vfio_group_device_count(vfio_group_fd) == 0) {
729 const struct vfio_iommu_type *t;
731 /* select an IOMMU type which we will be using */
732 t = vfio_set_iommu_type(vfio_container_fd);
735 " %s failed to select IOMMU type\n",
737 close(vfio_group_fd);
738 rte_vfio_clear_group(vfio_group_fd);
741 /* lock memory hotplug before mapping and release it
742 * after registering callback, to prevent races
744 rte_mcfg_mem_read_lock();
745 if (vfio_cfg == default_vfio_cfg)
746 ret = t->dma_map_func(vfio_container_fd);
751 " %s DMA remapping failed, error %i (%s)\n",
752 dev_addr, errno, strerror(errno));
753 close(vfio_group_fd);
754 rte_vfio_clear_group(vfio_group_fd);
755 rte_mcfg_mem_read_unlock();
759 vfio_cfg->vfio_iommu_type = t;
761 /* re-map all user-mapped segments */
762 rte_spinlock_recursive_lock(&user_mem_maps->lock);
764 /* this IOMMU type may not support DMA mapping, but
765 * if we have mappings in the list - that means we have
766 * previously mapped something successfully, so we can
767 * be sure that DMA mapping is supported.
769 for (i = 0; i < user_mem_maps->n_maps; i++) {
770 struct user_mem_map *map;
771 map = &user_mem_maps->maps[i];
773 ret = t->dma_user_map_func(
775 map->addr, map->iova, map->len,
778 RTE_LOG(ERR, EAL, "Couldn't map user memory for DMA: "
780 "iova: 0x%" PRIx64 " "
781 "len: 0x%" PRIu64 "\n",
782 map->addr, map->iova,
784 rte_spinlock_recursive_unlock(
785 &user_mem_maps->lock);
786 rte_mcfg_mem_read_unlock();
790 rte_spinlock_recursive_unlock(&user_mem_maps->lock);
792 /* register callback for mem events */
793 if (vfio_cfg == default_vfio_cfg)
794 ret = rte_mem_event_callback_register(
795 VFIO_MEM_EVENT_CLB_NAME,
796 vfio_mem_event_callback, NULL);
799 /* unlock memory hotplug */
800 rte_mcfg_mem_read_unlock();
802 if (ret && rte_errno != ENOTSUP) {
803 RTE_LOG(ERR, EAL, "Could not install memory event callback for VFIO\n");
807 RTE_LOG(DEBUG, EAL, "Memory event callbacks not supported\n");
809 RTE_LOG(DEBUG, EAL, "Installed memory event callback for VFIO\n");
811 } else if (rte_eal_process_type() != RTE_PROC_PRIMARY &&
812 vfio_cfg == default_vfio_cfg &&
813 vfio_cfg->vfio_iommu_type == NULL) {
814 /* if we're not a primary process, we do not set up the VFIO
815 * container because it's already been set up by the primary
816 * process. instead, we simply ask the primary about VFIO type
817 * we are using, and set the VFIO config up appropriately.
819 ret = vfio_sync_default_container();
821 RTE_LOG(ERR, EAL, "Could not sync default VFIO container\n");
822 close(vfio_group_fd);
823 rte_vfio_clear_group(vfio_group_fd);
826 /* we have successfully initialized VFIO, notify user */
827 const struct vfio_iommu_type *t =
828 default_vfio_cfg->vfio_iommu_type;
829 RTE_LOG(NOTICE, EAL, " using IOMMU type %d (%s)\n",
830 t->type_id, t->name);
833 /* get a file descriptor for the device */
834 *vfio_dev_fd = ioctl(vfio_group_fd, VFIO_GROUP_GET_DEVICE_FD, dev_addr);
835 if (*vfio_dev_fd < 0) {
836 /* if we cannot get a device fd, this implies a problem with
837 * the VFIO group or the container not having IOMMU configured.
840 RTE_LOG(WARNING, EAL, "Getting a vfio_dev_fd for %s failed\n",
842 close(vfio_group_fd);
843 rte_vfio_clear_group(vfio_group_fd);
847 /* test and setup the device */
848 ret = ioctl(*vfio_dev_fd, VFIO_DEVICE_GET_INFO, device_info);
850 RTE_LOG(ERR, EAL, " %s cannot get device info, "
851 "error %i (%s)\n", dev_addr, errno,
854 close(vfio_group_fd);
855 rte_vfio_clear_group(vfio_group_fd);
858 vfio_group_device_get(vfio_group_fd);
864 rte_vfio_release_device(const char *sysfs_base, const char *dev_addr,
867 struct vfio_group_status group_status = {
868 .argsz = sizeof(group_status)
870 struct vfio_config *vfio_cfg;
875 /* we don't want any DMA mapping messages to come while we're detaching
876 * VFIO device, because this might be the last device and we might need
877 * to unregister the callback.
879 rte_mcfg_mem_read_lock();
881 /* get group number */
882 ret = rte_vfio_get_group_num(sysfs_base, dev_addr, &iommu_group_num);
884 RTE_LOG(WARNING, EAL, " %s not managed by VFIO driver\n",
886 /* This is an error at this point. */
891 /* get the actual group fd */
892 vfio_group_fd = rte_vfio_get_group_fd(iommu_group_num);
893 if (vfio_group_fd <= 0) {
894 RTE_LOG(INFO, EAL, "rte_vfio_get_group_fd failed for %s\n",
900 /* get the vfio_config it belongs to */
901 vfio_cfg = get_vfio_cfg_by_group_num(iommu_group_num);
902 vfio_cfg = vfio_cfg ? vfio_cfg : default_vfio_cfg;
904 /* At this point we got an active group. Closing it will make the
905 * container detachment. If this is the last active group, VFIO kernel
906 * code will unset the container and the IOMMU mappings.
909 /* Closing a device */
910 if (close(vfio_dev_fd) < 0) {
911 RTE_LOG(INFO, EAL, "Error when closing vfio_dev_fd for %s\n",
917 /* An VFIO group can have several devices attached. Just when there is
918 * no devices remaining should the group be closed.
920 vfio_group_device_put(vfio_group_fd);
921 if (!vfio_group_device_count(vfio_group_fd)) {
923 if (close(vfio_group_fd) < 0) {
924 RTE_LOG(INFO, EAL, "Error when closing vfio_group_fd for %s\n",
930 if (rte_vfio_clear_group(vfio_group_fd) < 0) {
931 RTE_LOG(INFO, EAL, "Error when clearing group for %s\n",
938 /* if there are no active device groups, unregister the callback to
939 * avoid spurious attempts to map/unmap memory from VFIO.
941 if (vfio_cfg == default_vfio_cfg && vfio_cfg->vfio_active_groups == 0 &&
942 rte_eal_process_type() != RTE_PROC_SECONDARY)
943 rte_mem_event_callback_unregister(VFIO_MEM_EVENT_CLB_NAME,
950 rte_mcfg_mem_read_unlock();
955 rte_vfio_enable(const char *modname)
957 /* initialize group list */
961 rte_spinlock_recursive_t lock = RTE_SPINLOCK_RECURSIVE_INITIALIZER;
963 for (i = 0; i < VFIO_MAX_CONTAINERS; i++) {
964 vfio_cfgs[i].vfio_container_fd = -1;
965 vfio_cfgs[i].vfio_active_groups = 0;
966 vfio_cfgs[i].vfio_iommu_type = NULL;
967 vfio_cfgs[i].mem_maps.lock = lock;
969 for (j = 0; j < VFIO_MAX_GROUPS; j++) {
970 vfio_cfgs[i].vfio_groups[j].fd = -1;
971 vfio_cfgs[i].vfio_groups[j].group_num = -1;
972 vfio_cfgs[i].vfio_groups[j].devices = 0;
976 /* inform the user that we are probing for VFIO */
977 RTE_LOG(INFO, EAL, "Probing VFIO support...\n");
979 /* check if vfio module is loaded */
980 vfio_available = rte_eal_check_module(modname);
982 /* return error directly */
983 if (vfio_available == -1) {
984 RTE_LOG(INFO, EAL, "Could not get loaded module details!\n");
988 /* return 0 if VFIO modules not loaded */
989 if (vfio_available == 0) {
990 RTE_LOG(DEBUG, EAL, "VFIO modules not loaded, "
991 "skipping VFIO support...\n");
995 if (internal_config.process_type == RTE_PROC_PRIMARY) {
996 /* open a new container */
997 default_vfio_cfg->vfio_container_fd =
998 rte_vfio_get_container_fd();
1000 /* get the default container from the primary process */
1001 default_vfio_cfg->vfio_container_fd =
1002 vfio_get_default_container_fd();
1005 /* check if we have VFIO driver enabled */
1006 if (default_vfio_cfg->vfio_container_fd != -1) {
1007 RTE_LOG(NOTICE, EAL, "VFIO support initialized\n");
1008 default_vfio_cfg->vfio_enabled = 1;
1010 RTE_LOG(NOTICE, EAL, "VFIO support could not be initialized\n");
1017 rte_vfio_is_enabled(const char *modname)
1019 const int mod_available = rte_eal_check_module(modname) > 0;
1020 return default_vfio_cfg->vfio_enabled && mod_available;
1024 vfio_get_default_container_fd(void)
1026 struct rte_mp_msg mp_req, *mp_rep;
1027 struct rte_mp_reply mp_reply = {0};
1028 struct timespec ts = {.tv_sec = 5, .tv_nsec = 0};
1029 struct vfio_mp_param *p = (struct vfio_mp_param *)mp_req.param;
1031 if (default_vfio_cfg->vfio_enabled)
1032 return default_vfio_cfg->vfio_container_fd;
1034 if (internal_config.process_type == RTE_PROC_PRIMARY) {
1035 /* if we were secondary process we would try requesting
1036 * container fd from the primary, but we're the primary
1037 * process so just exit here
1042 p->req = SOCKET_REQ_DEFAULT_CONTAINER;
1043 strcpy(mp_req.name, EAL_VFIO_MP);
1044 mp_req.len_param = sizeof(*p);
1047 if (rte_mp_request_sync(&mp_req, &mp_reply, &ts) == 0 &&
1048 mp_reply.nb_received == 1) {
1049 mp_rep = &mp_reply.msgs[0];
1050 p = (struct vfio_mp_param *)mp_rep->param;
1051 if (p->result == SOCKET_OK && mp_rep->num_fds == 1) {
1052 free(mp_reply.msgs);
1053 return mp_rep->fds[0];
1057 free(mp_reply.msgs);
1058 RTE_LOG(ERR, EAL, " cannot request default container fd\n");
1063 vfio_get_iommu_type(void)
1065 if (default_vfio_cfg->vfio_iommu_type == NULL)
1068 return default_vfio_cfg->vfio_iommu_type->type_id;
1071 const struct vfio_iommu_type *
1072 vfio_set_iommu_type(int vfio_container_fd)
1075 for (idx = 0; idx < RTE_DIM(iommu_types); idx++) {
1076 const struct vfio_iommu_type *t = &iommu_types[idx];
1078 int ret = ioctl(vfio_container_fd, VFIO_SET_IOMMU,
1081 RTE_LOG(NOTICE, EAL, " using IOMMU type %d (%s)\n",
1082 t->type_id, t->name);
1085 /* not an error, there may be more supported IOMMU types */
1086 RTE_LOG(DEBUG, EAL, " set IOMMU type %d (%s) failed, "
1087 "error %i (%s)\n", t->type_id, t->name, errno,
1090 /* if we didn't find a suitable IOMMU type, fail */
1095 vfio_has_supported_extensions(int vfio_container_fd)
1098 unsigned idx, n_extensions = 0;
1099 for (idx = 0; idx < RTE_DIM(iommu_types); idx++) {
1100 const struct vfio_iommu_type *t = &iommu_types[idx];
1102 ret = ioctl(vfio_container_fd, VFIO_CHECK_EXTENSION,
1105 RTE_LOG(ERR, EAL, " could not get IOMMU type, "
1106 "error %i (%s)\n", errno,
1108 close(vfio_container_fd);
1110 } else if (ret == 1) {
1111 /* we found a supported extension */
1114 RTE_LOG(DEBUG, EAL, " IOMMU type %d (%s) is %s\n",
1115 t->type_id, t->name,
1116 ret ? "supported" : "not supported");
1119 /* if we didn't find any supported IOMMU types, fail */
1120 if (!n_extensions) {
1121 close(vfio_container_fd);
1129 rte_vfio_get_container_fd(void)
1131 int ret, vfio_container_fd;
1132 struct rte_mp_msg mp_req, *mp_rep;
1133 struct rte_mp_reply mp_reply = {0};
1134 struct timespec ts = {.tv_sec = 5, .tv_nsec = 0};
1135 struct vfio_mp_param *p = (struct vfio_mp_param *)mp_req.param;
1138 /* if we're in a primary process, try to open the container */
1139 if (internal_config.process_type == RTE_PROC_PRIMARY) {
1140 vfio_container_fd = open(VFIO_CONTAINER_PATH, O_RDWR);
1141 if (vfio_container_fd < 0) {
1142 RTE_LOG(ERR, EAL, " cannot open VFIO container, "
1143 "error %i (%s)\n", errno, strerror(errno));
1147 /* check VFIO API version */
1148 ret = ioctl(vfio_container_fd, VFIO_GET_API_VERSION);
1149 if (ret != VFIO_API_VERSION) {
1151 RTE_LOG(ERR, EAL, " could not get VFIO API version, "
1152 "error %i (%s)\n", errno, strerror(errno));
1154 RTE_LOG(ERR, EAL, " unsupported VFIO API version!\n");
1155 close(vfio_container_fd);
1159 ret = vfio_has_supported_extensions(vfio_container_fd);
1161 RTE_LOG(ERR, EAL, " no supported IOMMU "
1162 "extensions found!\n");
1166 return vfio_container_fd;
1169 * if we're in a secondary process, request container fd from the
1170 * primary process via mp channel
1172 p->req = SOCKET_REQ_CONTAINER;
1173 strcpy(mp_req.name, EAL_VFIO_MP);
1174 mp_req.len_param = sizeof(*p);
1177 vfio_container_fd = -1;
1178 if (rte_mp_request_sync(&mp_req, &mp_reply, &ts) == 0 &&
1179 mp_reply.nb_received == 1) {
1180 mp_rep = &mp_reply.msgs[0];
1181 p = (struct vfio_mp_param *)mp_rep->param;
1182 if (p->result == SOCKET_OK && mp_rep->num_fds == 1) {
1183 vfio_container_fd = mp_rep->fds[0];
1184 free(mp_reply.msgs);
1185 return vfio_container_fd;
1189 free(mp_reply.msgs);
1190 RTE_LOG(ERR, EAL, " cannot request container fd\n");
1195 rte_vfio_get_group_num(const char *sysfs_base,
1196 const char *dev_addr, int *iommu_group_num)
1198 char linkname[PATH_MAX];
1199 char filename[PATH_MAX];
1200 char *tok[16], *group_tok, *end;
1203 memset(linkname, 0, sizeof(linkname));
1204 memset(filename, 0, sizeof(filename));
1206 /* try to find out IOMMU group for this device */
1207 snprintf(linkname, sizeof(linkname),
1208 "%s/%s/iommu_group", sysfs_base, dev_addr);
1210 ret = readlink(linkname, filename, sizeof(filename));
1212 /* if the link doesn't exist, no VFIO for us */
1216 ret = rte_strsplit(filename, sizeof(filename),
1217 tok, RTE_DIM(tok), '/');
1220 RTE_LOG(ERR, EAL, " %s cannot get IOMMU group\n", dev_addr);
1224 /* IOMMU group is always the last token */
1226 group_tok = tok[ret - 1];
1228 *iommu_group_num = strtol(group_tok, &end, 10);
1229 if ((end != group_tok && *end != '\0') || errno != 0) {
1230 RTE_LOG(ERR, EAL, " %s error parsing IOMMU number!\n", dev_addr);
1238 type1_map_contig(const struct rte_memseg_list *msl, const struct rte_memseg *ms,
1239 size_t len, void *arg)
1241 int *vfio_container_fd = arg;
1246 return vfio_type1_dma_mem_map(*vfio_container_fd, ms->addr_64, ms->iova,
1251 type1_map(const struct rte_memseg_list *msl, const struct rte_memseg *ms,
1254 int *vfio_container_fd = arg;
1256 /* skip external memory that isn't a heap */
1257 if (msl->external && !msl->heap)
1260 /* skip any segments with invalid IOVA addresses */
1261 if (ms->iova == RTE_BAD_IOVA)
1264 /* if IOVA mode is VA, we've already mapped the internal segments */
1265 if (!msl->external && rte_eal_iova_mode() == RTE_IOVA_VA)
1268 return vfio_type1_dma_mem_map(*vfio_container_fd, ms->addr_64, ms->iova,
1273 vfio_type1_dma_mem_map(int vfio_container_fd, uint64_t vaddr, uint64_t iova,
1274 uint64_t len, int do_map)
1276 struct vfio_iommu_type1_dma_map dma_map;
1277 struct vfio_iommu_type1_dma_unmap dma_unmap;
1281 memset(&dma_map, 0, sizeof(dma_map));
1282 dma_map.argsz = sizeof(struct vfio_iommu_type1_dma_map);
1283 dma_map.vaddr = vaddr;
1285 dma_map.iova = iova;
1286 dma_map.flags = VFIO_DMA_MAP_FLAG_READ |
1287 VFIO_DMA_MAP_FLAG_WRITE;
1289 ret = ioctl(vfio_container_fd, VFIO_IOMMU_MAP_DMA, &dma_map);
1292 * In case the mapping was already done EEXIST will be
1293 * returned from kernel.
1295 if (errno == EEXIST) {
1297 " Memory segment is already mapped,"
1301 " cannot set up DMA remapping,"
1303 errno, strerror(errno));
1308 memset(&dma_unmap, 0, sizeof(dma_unmap));
1309 dma_unmap.argsz = sizeof(struct vfio_iommu_type1_dma_unmap);
1310 dma_unmap.size = len;
1311 dma_unmap.iova = iova;
1313 ret = ioctl(vfio_container_fd, VFIO_IOMMU_UNMAP_DMA,
1316 RTE_LOG(ERR, EAL, " cannot clear DMA remapping, error %i (%s)\n",
1317 errno, strerror(errno));
1326 vfio_type1_dma_map(int vfio_container_fd)
1328 if (rte_eal_iova_mode() == RTE_IOVA_VA) {
1329 /* with IOVA as VA mode, we can get away with mapping contiguous
1330 * chunks rather than going page-by-page.
1332 int ret = rte_memseg_contig_walk(type1_map_contig,
1333 &vfio_container_fd);
1336 /* we have to continue the walk because we've skipped the
1337 * external segments during the config walk.
1340 return rte_memseg_walk(type1_map, &vfio_container_fd);
1344 vfio_spapr_dma_do_map(int vfio_container_fd, uint64_t vaddr, uint64_t iova,
1345 uint64_t len, int do_map)
1347 struct vfio_iommu_type1_dma_map dma_map;
1348 struct vfio_iommu_type1_dma_unmap dma_unmap;
1350 struct vfio_iommu_spapr_register_memory reg = {
1351 .argsz = sizeof(reg),
1354 reg.vaddr = (uintptr_t) vaddr;
1358 ret = ioctl(vfio_container_fd,
1359 VFIO_IOMMU_SPAPR_REGISTER_MEMORY, ®);
1361 RTE_LOG(ERR, EAL, " cannot register vaddr for IOMMU, "
1362 "error %i (%s)\n", errno, strerror(errno));
1366 memset(&dma_map, 0, sizeof(dma_map));
1367 dma_map.argsz = sizeof(struct vfio_iommu_type1_dma_map);
1368 dma_map.vaddr = vaddr;
1370 dma_map.iova = iova;
1371 dma_map.flags = VFIO_DMA_MAP_FLAG_READ |
1372 VFIO_DMA_MAP_FLAG_WRITE;
1374 ret = ioctl(vfio_container_fd, VFIO_IOMMU_MAP_DMA, &dma_map);
1377 * In case the mapping was already done EBUSY will be
1378 * returned from kernel.
1380 if (errno == EBUSY) {
1382 " Memory segment is already mapped,"
1386 " cannot set up DMA remapping,"
1387 " error %i (%s)\n", errno,
1394 memset(&dma_unmap, 0, sizeof(dma_unmap));
1395 dma_unmap.argsz = sizeof(struct vfio_iommu_type1_dma_unmap);
1396 dma_unmap.size = len;
1397 dma_unmap.iova = iova;
1399 ret = ioctl(vfio_container_fd, VFIO_IOMMU_UNMAP_DMA,
1402 RTE_LOG(ERR, EAL, " cannot clear DMA remapping, error %i (%s)\n",
1403 errno, strerror(errno));
1407 ret = ioctl(vfio_container_fd,
1408 VFIO_IOMMU_SPAPR_UNREGISTER_MEMORY, ®);
1410 RTE_LOG(ERR, EAL, " cannot unregister vaddr for IOMMU, error %i (%s)\n",
1411 errno, strerror(errno));
1419 struct spapr_remap_walk_param {
1420 int vfio_container_fd;
1425 vfio_spapr_map_walk(const struct rte_memseg_list *msl,
1426 const struct rte_memseg *ms, void *arg)
1428 struct spapr_remap_walk_param *param = arg;
1430 /* skip external memory that isn't a heap */
1431 if (msl->external && !msl->heap)
1434 /* skip any segments with invalid IOVA addresses */
1435 if (ms->iova == RTE_BAD_IOVA)
1438 if (ms->addr_64 == param->addr_64)
1441 return vfio_spapr_dma_do_map(param->vfio_container_fd, ms->addr_64, ms->iova,
1446 vfio_spapr_unmap_walk(const struct rte_memseg_list *msl,
1447 const struct rte_memseg *ms, void *arg)
1449 struct spapr_remap_walk_param *param = arg;
1451 /* skip external memory that isn't a heap */
1452 if (msl->external && !msl->heap)
1455 /* skip any segments with invalid IOVA addresses */
1456 if (ms->iova == RTE_BAD_IOVA)
1459 if (ms->addr_64 == param->addr_64)
1462 return vfio_spapr_dma_do_map(param->vfio_container_fd, ms->addr_64, ms->iova,
1466 struct spapr_walk_param {
1467 uint64_t window_size;
1468 uint64_t hugepage_sz;
1473 vfio_spapr_window_size_walk(const struct rte_memseg_list *msl,
1474 const struct rte_memseg *ms, void *arg)
1476 struct spapr_walk_param *param = arg;
1477 uint64_t max = ms->iova + ms->len;
1479 /* skip external memory that isn't a heap */
1480 if (msl->external && !msl->heap)
1483 /* skip any segments with invalid IOVA addresses */
1484 if (ms->iova == RTE_BAD_IOVA)
1487 /* do not iterate ms we haven't mapped yet */
1488 if (param->addr_64 && ms->addr_64 == param->addr_64)
1491 if (max > param->window_size) {
1492 param->hugepage_sz = ms->hugepage_sz;
1493 param->window_size = max;
1500 vfio_spapr_create_new_dma_window(int vfio_container_fd,
1501 struct vfio_iommu_spapr_tce_create *create) {
1502 struct vfio_iommu_spapr_tce_remove remove = {
1503 .argsz = sizeof(remove),
1505 struct vfio_iommu_spapr_tce_info info = {
1506 .argsz = sizeof(info),
1510 /* query spapr iommu info */
1511 ret = ioctl(vfio_container_fd, VFIO_IOMMU_SPAPR_TCE_GET_INFO, &info);
1513 RTE_LOG(ERR, EAL, " cannot get iommu info, "
1514 "error %i (%s)\n", errno, strerror(errno));
1518 /* remove default DMA of 32 bit window */
1519 remove.start_addr = info.dma32_window_start;
1520 ret = ioctl(vfio_container_fd, VFIO_IOMMU_SPAPR_TCE_REMOVE, &remove);
1522 RTE_LOG(ERR, EAL, " cannot remove default DMA window, "
1523 "error %i (%s)\n", errno, strerror(errno));
1527 /* create new DMA window */
1528 ret = ioctl(vfio_container_fd, VFIO_IOMMU_SPAPR_TCE_CREATE, create);
1530 #ifdef VFIO_IOMMU_SPAPR_INFO_DDW
1531 /* try possible page_shift and levels for workaround */
1534 for (levels = 1; levels <= info.ddw.levels; levels++) {
1535 uint32_t pgsizes = info.ddw.pgsizes;
1537 while (pgsizes != 0) {
1538 create->page_shift = 31 - __builtin_clz(pgsizes);
1539 create->levels = levels;
1540 ret = ioctl(vfio_container_fd,
1541 VFIO_IOMMU_SPAPR_TCE_CREATE, create);
1544 pgsizes &= ~(1 << create->page_shift);
1551 RTE_LOG(ERR, EAL, " cannot create new DMA window, "
1552 "error %i (%s)\n", errno, strerror(errno));
1557 if (create->start_addr != 0) {
1558 RTE_LOG(ERR, EAL, " DMA window start address != 0\n");
1566 vfio_spapr_dma_mem_map(int vfio_container_fd, uint64_t vaddr, uint64_t iova,
1567 uint64_t len, int do_map)
1569 struct spapr_walk_param param;
1570 struct vfio_iommu_spapr_tce_create create = {
1571 .argsz = sizeof(create),
1573 struct vfio_config *vfio_cfg;
1574 struct user_mem_maps *user_mem_maps;
1577 vfio_cfg = get_vfio_cfg_by_container_fd(vfio_container_fd);
1578 if (vfio_cfg == NULL) {
1579 RTE_LOG(ERR, EAL, " invalid container fd!\n");
1583 user_mem_maps = &vfio_cfg->mem_maps;
1584 rte_spinlock_recursive_lock(&user_mem_maps->lock);
1586 /* check if window size needs to be adjusted */
1587 memset(¶m, 0, sizeof(param));
1588 param.addr_64 = vaddr;
1590 /* we're inside a callback so use thread-unsafe version */
1591 if (rte_memseg_walk_thread_unsafe(vfio_spapr_window_size_walk,
1593 RTE_LOG(ERR, EAL, "Could not get window size\n");
1598 /* also check user maps */
1599 for (i = 0; i < user_mem_maps->n_maps; i++) {
1600 uint64_t max = user_mem_maps->maps[i].iova +
1601 user_mem_maps->maps[i].len;
1602 param.window_size = RTE_MAX(param.window_size, max);
1605 /* sPAPR requires window size to be a power of 2 */
1606 create.window_size = rte_align64pow2(param.window_size);
1607 create.page_shift = __builtin_ctzll(param.hugepage_sz);
1611 /* re-create window and remap the entire memory */
1612 if (iova + len > create.window_size) {
1613 struct spapr_remap_walk_param remap_param = {
1614 .vfio_container_fd = vfio_container_fd,
1618 /* release all maps before recreating the window */
1619 if (rte_memseg_walk_thread_unsafe(vfio_spapr_unmap_walk,
1620 &remap_param) < 0) {
1621 RTE_LOG(ERR, EAL, "Could not release DMA maps\n");
1625 /* release all user maps */
1626 for (i = 0; i < user_mem_maps->n_maps; i++) {
1627 struct user_mem_map *map =
1628 &user_mem_maps->maps[i];
1629 if (vfio_spapr_dma_do_map(vfio_container_fd,
1630 map->addr, map->iova, map->len,
1632 RTE_LOG(ERR, EAL, "Could not release user DMA maps\n");
1637 create.window_size = rte_align64pow2(iova + len);
1638 if (vfio_spapr_create_new_dma_window(vfio_container_fd,
1640 RTE_LOG(ERR, EAL, "Could not create new DMA window\n");
1644 /* we're inside a callback, so use thread-unsafe version
1646 if (rte_memseg_walk_thread_unsafe(vfio_spapr_map_walk,
1647 &remap_param) < 0) {
1648 RTE_LOG(ERR, EAL, "Could not recreate DMA maps\n");
1652 /* remap all user maps */
1653 for (i = 0; i < user_mem_maps->n_maps; i++) {
1654 struct user_mem_map *map =
1655 &user_mem_maps->maps[i];
1656 if (vfio_spapr_dma_do_map(vfio_container_fd,
1657 map->addr, map->iova, map->len,
1659 RTE_LOG(ERR, EAL, "Could not recreate user DMA maps\n");
1665 if (vfio_spapr_dma_do_map(vfio_container_fd, vaddr, iova, len, 1)) {
1666 RTE_LOG(ERR, EAL, "Failed to map DMA\n");
1671 /* for unmap, check if iova within DMA window */
1672 if (iova > create.window_size) {
1673 RTE_LOG(ERR, EAL, "iova beyond DMA window for unmap");
1678 vfio_spapr_dma_do_map(vfio_container_fd, vaddr, iova, len, 0);
1681 rte_spinlock_recursive_unlock(&user_mem_maps->lock);
1686 vfio_spapr_dma_map(int vfio_container_fd)
1688 struct vfio_iommu_spapr_tce_create create = {
1689 .argsz = sizeof(create),
1691 struct spapr_walk_param param;
1693 memset(¶m, 0, sizeof(param));
1694 param.addr_64 = 0UL;
1696 /* create DMA window from 0 to max(phys_addr + len) */
1697 rte_memseg_walk(vfio_spapr_window_size_walk, ¶m);
1699 /* sPAPR requires window size to be a power of 2 */
1700 create.window_size = rte_align64pow2(param.window_size);
1701 create.page_shift = __builtin_ctzll(param.hugepage_sz);
1704 if (vfio_spapr_create_new_dma_window(vfio_container_fd, &create) < 0) {
1705 RTE_LOG(ERR, EAL, "Could not create new DMA window\n");
1709 /* map all DPDK segments for DMA. use 1:1 PA to IOVA mapping */
1710 if (rte_memseg_walk(vfio_spapr_map_walk, &vfio_container_fd) < 0)
1717 vfio_noiommu_dma_map(int __rte_unused vfio_container_fd)
1719 /* No-IOMMU mode does not need DMA mapping */
1724 vfio_noiommu_dma_mem_map(int __rte_unused vfio_container_fd,
1725 uint64_t __rte_unused vaddr,
1726 uint64_t __rte_unused iova, uint64_t __rte_unused len,
1727 int __rte_unused do_map)
1729 /* No-IOMMU mode does not need DMA mapping */
1734 vfio_dma_mem_map(struct vfio_config *vfio_cfg, uint64_t vaddr, uint64_t iova,
1735 uint64_t len, int do_map)
1737 const struct vfio_iommu_type *t = vfio_cfg->vfio_iommu_type;
1740 RTE_LOG(ERR, EAL, " VFIO support not initialized\n");
1745 if (!t->dma_user_map_func) {
1747 " VFIO custom DMA region maping not supported by IOMMU %s\n",
1749 rte_errno = ENOTSUP;
1753 return t->dma_user_map_func(vfio_cfg->vfio_container_fd, vaddr, iova,
1758 container_dma_map(struct vfio_config *vfio_cfg, uint64_t vaddr, uint64_t iova,
1761 struct user_mem_map *new_map;
1762 struct user_mem_maps *user_mem_maps;
1765 user_mem_maps = &vfio_cfg->mem_maps;
1766 rte_spinlock_recursive_lock(&user_mem_maps->lock);
1767 if (user_mem_maps->n_maps == VFIO_MAX_USER_MEM_MAPS) {
1768 RTE_LOG(ERR, EAL, "No more space for user mem maps\n");
1774 if (vfio_dma_mem_map(vfio_cfg, vaddr, iova, len, 1)) {
1775 /* technically, this will fail if there are currently no devices
1776 * plugged in, even if a device were added later, this mapping
1777 * might have succeeded. however, since we cannot verify if this
1778 * is a valid mapping without having a device attached, consider
1779 * this to be unsupported, because we can't just store any old
1780 * mapping and pollute list of active mappings willy-nilly.
1782 RTE_LOG(ERR, EAL, "Couldn't map new region for DMA\n");
1786 /* create new user mem map entry */
1787 new_map = &user_mem_maps->maps[user_mem_maps->n_maps++];
1788 new_map->addr = vaddr;
1789 new_map->iova = iova;
1792 compact_user_maps(user_mem_maps);
1794 rte_spinlock_recursive_unlock(&user_mem_maps->lock);
1799 container_dma_unmap(struct vfio_config *vfio_cfg, uint64_t vaddr, uint64_t iova,
1802 struct user_mem_map *map, *new_map = NULL;
1803 struct user_mem_maps *user_mem_maps;
1806 user_mem_maps = &vfio_cfg->mem_maps;
1807 rte_spinlock_recursive_lock(&user_mem_maps->lock);
1809 /* find our mapping */
1810 map = find_user_mem_map(user_mem_maps, vaddr, iova, len);
1812 RTE_LOG(ERR, EAL, "Couldn't find previously mapped region\n");
1817 if (map->addr != vaddr || map->iova != iova || map->len != len) {
1818 /* we're partially unmapping a previously mapped region, so we
1819 * need to split entry into two.
1821 if (user_mem_maps->n_maps == VFIO_MAX_USER_MEM_MAPS) {
1822 RTE_LOG(ERR, EAL, "Not enough space to store partial mapping\n");
1827 new_map = &user_mem_maps->maps[user_mem_maps->n_maps++];
1830 /* unmap the entry */
1831 if (vfio_dma_mem_map(vfio_cfg, vaddr, iova, len, 0)) {
1832 /* there may not be any devices plugged in, so unmapping will
1833 * fail with ENODEV/ENOTSUP rte_errno values, but that doesn't
1834 * stop us from removing the mapping, as the assumption is we
1835 * won't be needing this memory any more and thus will want to
1836 * prevent it from being remapped again on hotplug. so, only
1837 * fail if we indeed failed to unmap (e.g. if the mapping was
1838 * within our mapped range but had invalid alignment).
1840 if (rte_errno != ENODEV && rte_errno != ENOTSUP) {
1841 RTE_LOG(ERR, EAL, "Couldn't unmap region for DMA\n");
1845 RTE_LOG(DEBUG, EAL, "DMA unmapping failed, but removing mappings anyway\n");
1848 /* remove map from the list of active mappings */
1849 if (new_map != NULL) {
1850 adjust_map(map, new_map, vaddr, len);
1852 /* if we've created a new map by splitting, sort everything */
1853 if (!is_null_map(new_map)) {
1854 compact_user_maps(user_mem_maps);
1856 /* we've created a new mapping, but it was unused */
1857 user_mem_maps->n_maps--;
1860 memset(map, 0, sizeof(*map));
1861 compact_user_maps(user_mem_maps);
1862 user_mem_maps->n_maps--;
1866 rte_spinlock_recursive_unlock(&user_mem_maps->lock);
1871 rte_vfio_noiommu_is_enabled(void)
1877 fd = open(VFIO_NOIOMMU_MODE, O_RDONLY);
1879 if (errno != ENOENT) {
1880 RTE_LOG(ERR, EAL, " cannot open vfio noiommu file %i (%s)\n",
1881 errno, strerror(errno));
1885 * else the file does not exists
1886 * i.e. noiommu is not enabled
1891 cnt = read(fd, &c, 1);
1894 RTE_LOG(ERR, EAL, " unable to read from vfio noiommu "
1895 "file %i (%s)\n", errno, strerror(errno));
1903 rte_vfio_container_create(void)
1907 /* Find an empty slot to store new vfio config */
1908 for (i = 1; i < VFIO_MAX_CONTAINERS; i++) {
1909 if (vfio_cfgs[i].vfio_container_fd == -1)
1913 if (i == VFIO_MAX_CONTAINERS) {
1914 RTE_LOG(ERR, EAL, "exceed max vfio container limit\n");
1918 vfio_cfgs[i].vfio_container_fd = rte_vfio_get_container_fd();
1919 if (vfio_cfgs[i].vfio_container_fd < 0) {
1920 RTE_LOG(NOTICE, EAL, "fail to create a new container\n");
1924 return vfio_cfgs[i].vfio_container_fd;
1928 rte_vfio_container_destroy(int container_fd)
1930 struct vfio_config *vfio_cfg;
1933 vfio_cfg = get_vfio_cfg_by_container_fd(container_fd);
1934 if (vfio_cfg == NULL) {
1935 RTE_LOG(ERR, EAL, "Invalid container fd\n");
1939 for (i = 0; i < VFIO_MAX_GROUPS; i++)
1940 if (vfio_cfg->vfio_groups[i].group_num != -1)
1941 rte_vfio_container_group_unbind(container_fd,
1942 vfio_cfg->vfio_groups[i].group_num);
1944 close(container_fd);
1945 vfio_cfg->vfio_container_fd = -1;
1946 vfio_cfg->vfio_active_groups = 0;
1947 vfio_cfg->vfio_iommu_type = NULL;
1953 rte_vfio_container_group_bind(int container_fd, int iommu_group_num)
1955 struct vfio_config *vfio_cfg;
1957 vfio_cfg = get_vfio_cfg_by_container_fd(container_fd);
1958 if (vfio_cfg == NULL) {
1959 RTE_LOG(ERR, EAL, "Invalid container fd\n");
1963 return vfio_get_group_fd(vfio_cfg, iommu_group_num);
1967 rte_vfio_container_group_unbind(int container_fd, int iommu_group_num)
1969 struct vfio_config *vfio_cfg;
1970 struct vfio_group *cur_grp = NULL;
1973 vfio_cfg = get_vfio_cfg_by_container_fd(container_fd);
1974 if (vfio_cfg == NULL) {
1975 RTE_LOG(ERR, EAL, "Invalid container fd\n");
1979 for (i = 0; i < VFIO_MAX_GROUPS; i++) {
1980 if (vfio_cfg->vfio_groups[i].group_num == iommu_group_num) {
1981 cur_grp = &vfio_cfg->vfio_groups[i];
1986 /* This should not happen */
1987 if (i == VFIO_MAX_GROUPS || cur_grp == NULL) {
1988 RTE_LOG(ERR, EAL, "Specified group number not found\n");
1992 if (cur_grp->fd >= 0 && close(cur_grp->fd) < 0) {
1993 RTE_LOG(ERR, EAL, "Error when closing vfio_group_fd for"
1994 " iommu_group_num %d\n", iommu_group_num);
1997 cur_grp->group_num = -1;
1999 cur_grp->devices = 0;
2000 vfio_cfg->vfio_active_groups--;
2006 rte_vfio_container_dma_map(int container_fd, uint64_t vaddr, uint64_t iova,
2009 struct vfio_config *vfio_cfg;
2016 vfio_cfg = get_vfio_cfg_by_container_fd(container_fd);
2017 if (vfio_cfg == NULL) {
2018 RTE_LOG(ERR, EAL, "Invalid container fd\n");
2022 return container_dma_map(vfio_cfg, vaddr, iova, len);
2026 rte_vfio_container_dma_unmap(int container_fd, uint64_t vaddr, uint64_t iova,
2029 struct vfio_config *vfio_cfg;
2036 vfio_cfg = get_vfio_cfg_by_container_fd(container_fd);
2037 if (vfio_cfg == NULL) {
2038 RTE_LOG(ERR, EAL, "Invalid container fd\n");
2042 return container_dma_unmap(vfio_cfg, vaddr, iova, len);
2048 rte_vfio_setup_device(__rte_unused const char *sysfs_base,
2049 __rte_unused const char *dev_addr,
2050 __rte_unused int *vfio_dev_fd,
2051 __rte_unused struct vfio_device_info *device_info)
2057 rte_vfio_release_device(__rte_unused const char *sysfs_base,
2058 __rte_unused const char *dev_addr, __rte_unused int fd)
2064 rte_vfio_enable(__rte_unused const char *modname)
2070 rte_vfio_is_enabled(__rte_unused const char *modname)
2076 rte_vfio_noiommu_is_enabled(void)
2082 rte_vfio_clear_group(__rte_unused int vfio_group_fd)
2088 rte_vfio_get_group_num(__rte_unused const char *sysfs_base,
2089 __rte_unused const char *dev_addr,
2090 __rte_unused int *iommu_group_num)
2096 rte_vfio_get_container_fd(void)
2102 rte_vfio_get_group_fd(__rte_unused int iommu_group_num)
2108 rte_vfio_container_create(void)
2114 rte_vfio_container_destroy(__rte_unused int container_fd)
2120 rte_vfio_container_group_bind(__rte_unused int container_fd,
2121 __rte_unused int iommu_group_num)
2127 rte_vfio_container_group_unbind(__rte_unused int container_fd,
2128 __rte_unused int iommu_group_num)
2134 rte_vfio_container_dma_map(__rte_unused int container_fd,
2135 __rte_unused uint64_t vaddr,
2136 __rte_unused uint64_t iova,
2137 __rte_unused uint64_t len)
2143 rte_vfio_container_dma_unmap(__rte_unused int container_fd,
2144 __rte_unused uint64_t vaddr,
2145 __rte_unused uint64_t iova,
2146 __rte_unused uint64_t len)
2151 #endif /* VFIO_PRESENT */