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;
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 for (i = 0; i < VFIO_MAX_CONTAINERS; i++) {
416 if (vfio_cfgs[i].vfio_container_fd == container_fd)
417 return &vfio_cfgs[i];
424 rte_vfio_get_group_fd(int iommu_group_num)
426 struct vfio_config *vfio_cfg;
428 /* get the vfio_config it belongs to */
429 vfio_cfg = get_vfio_cfg_by_group_num(iommu_group_num);
430 vfio_cfg = vfio_cfg ? vfio_cfg : default_vfio_cfg;
432 return vfio_get_group_fd(vfio_cfg, iommu_group_num);
436 get_vfio_group_idx(int vfio_group_fd)
438 struct vfio_config *vfio_cfg;
441 for (i = 0; i < VFIO_MAX_CONTAINERS; i++) {
442 vfio_cfg = &vfio_cfgs[i];
443 for (j = 0; j < VFIO_MAX_GROUPS; j++)
444 if (vfio_cfg->vfio_groups[j].fd == vfio_group_fd)
452 vfio_group_device_get(int vfio_group_fd)
454 struct vfio_config *vfio_cfg;
457 vfio_cfg = get_vfio_cfg_by_group_fd(vfio_group_fd);
458 if (vfio_cfg == NULL) {
459 RTE_LOG(ERR, EAL, " invalid group fd!\n");
463 i = get_vfio_group_idx(vfio_group_fd);
464 if (i < 0 || i > (VFIO_MAX_GROUPS - 1))
465 RTE_LOG(ERR, EAL, " wrong vfio_group index (%d)\n", i);
467 vfio_cfg->vfio_groups[i].devices++;
471 vfio_group_device_put(int vfio_group_fd)
473 struct vfio_config *vfio_cfg;
476 vfio_cfg = get_vfio_cfg_by_group_fd(vfio_group_fd);
477 if (vfio_cfg == NULL) {
478 RTE_LOG(ERR, EAL, " invalid group fd!\n");
482 i = get_vfio_group_idx(vfio_group_fd);
483 if (i < 0 || i > (VFIO_MAX_GROUPS - 1))
484 RTE_LOG(ERR, EAL, " wrong vfio_group index (%d)\n", i);
486 vfio_cfg->vfio_groups[i].devices--;
490 vfio_group_device_count(int vfio_group_fd)
492 struct vfio_config *vfio_cfg;
495 vfio_cfg = get_vfio_cfg_by_group_fd(vfio_group_fd);
496 if (vfio_cfg == NULL) {
497 RTE_LOG(ERR, EAL, " invalid group fd!\n");
501 i = get_vfio_group_idx(vfio_group_fd);
502 if (i < 0 || i > (VFIO_MAX_GROUPS - 1)) {
503 RTE_LOG(ERR, EAL, " wrong vfio_group index (%d)\n", i);
507 return vfio_cfg->vfio_groups[i].devices;
511 vfio_mem_event_callback(enum rte_mem_event type, const void *addr, size_t len,
512 void *arg __rte_unused)
514 struct rte_memseg_list *msl;
515 struct rte_memseg *ms;
518 msl = rte_mem_virt2memseg_list(addr);
520 /* for IOVA as VA mode, no need to care for IOVA addresses */
521 if (rte_eal_iova_mode() == RTE_IOVA_VA && msl->external == 0) {
522 uint64_t vfio_va = (uint64_t)(uintptr_t)addr;
523 if (type == RTE_MEM_EVENT_ALLOC)
524 vfio_dma_mem_map(default_vfio_cfg, vfio_va, vfio_va,
527 vfio_dma_mem_map(default_vfio_cfg, vfio_va, vfio_va,
532 /* memsegs are contiguous in memory */
533 ms = rte_mem_virt2memseg(addr, msl);
534 while (cur_len < len) {
535 /* some memory segments may have invalid IOVA */
536 if (ms->iova == RTE_BAD_IOVA) {
537 RTE_LOG(DEBUG, EAL, "Memory segment at %p has bad IOVA, skipping\n",
541 if (type == RTE_MEM_EVENT_ALLOC)
542 vfio_dma_mem_map(default_vfio_cfg, ms->addr_64,
543 ms->iova, ms->len, 1);
545 vfio_dma_mem_map(default_vfio_cfg, ms->addr_64,
546 ms->iova, ms->len, 0);
554 vfio_sync_default_container(void)
556 struct rte_mp_msg mp_req, *mp_rep;
557 struct rte_mp_reply mp_reply;
558 struct timespec ts = {.tv_sec = 5, .tv_nsec = 0};
559 struct vfio_mp_param *p = (struct vfio_mp_param *)mp_req.param;
563 /* cannot be called from primary */
564 if (rte_eal_process_type() != RTE_PROC_SECONDARY)
567 /* default container fd should have been opened in rte_vfio_enable() */
568 if (!default_vfio_cfg->vfio_enabled ||
569 default_vfio_cfg->vfio_container_fd < 0) {
570 RTE_LOG(ERR, EAL, "VFIO support is not initialized\n");
574 /* find default container's IOMMU type */
575 p->req = SOCKET_REQ_IOMMU_TYPE;
576 strcpy(mp_req.name, EAL_VFIO_MP);
577 mp_req.len_param = sizeof(*p);
581 if (rte_mp_request_sync(&mp_req, &mp_reply, &ts) == 0 &&
582 mp_reply.nb_received == 1) {
583 mp_rep = &mp_reply.msgs[0];
584 p = (struct vfio_mp_param *)mp_rep->param;
585 if (p->result == SOCKET_OK)
586 iommu_type_id = p->iommu_type_id;
589 if (iommu_type_id < 0) {
590 RTE_LOG(ERR, EAL, "Could not get IOMMU type for default container\n");
594 /* we now have an fd for default container, as well as its IOMMU type.
595 * now, set up default VFIO container config to match.
597 for (i = 0; i < RTE_DIM(iommu_types); i++) {
598 const struct vfio_iommu_type *t = &iommu_types[i];
599 if (t->type_id != iommu_type_id)
602 /* we found our IOMMU type */
603 default_vfio_cfg->vfio_iommu_type = t;
607 RTE_LOG(ERR, EAL, "Could not find IOMMU type id (%i)\n",
613 rte_vfio_clear_group(int vfio_group_fd)
616 struct vfio_config *vfio_cfg;
618 vfio_cfg = get_vfio_cfg_by_group_fd(vfio_group_fd);
619 if (vfio_cfg == NULL) {
620 RTE_LOG(ERR, EAL, " invalid group fd!\n");
624 i = get_vfio_group_idx(vfio_group_fd);
627 vfio_cfg->vfio_groups[i].group_num = -1;
628 vfio_cfg->vfio_groups[i].fd = -1;
629 vfio_cfg->vfio_groups[i].devices = 0;
630 vfio_cfg->vfio_active_groups--;
636 rte_vfio_setup_device(const char *sysfs_base, const char *dev_addr,
637 int *vfio_dev_fd, struct vfio_device_info *device_info)
639 struct vfio_group_status group_status = {
640 .argsz = sizeof(group_status)
642 struct vfio_config *vfio_cfg;
643 struct user_mem_maps *user_mem_maps;
644 int vfio_container_fd;
649 /* get group number */
650 ret = rte_vfio_get_group_num(sysfs_base, dev_addr, &iommu_group_num);
652 RTE_LOG(WARNING, EAL, " %s not managed by VFIO driver, skipping\n",
657 /* if negative, something failed */
661 /* get the actual group fd */
662 vfio_group_fd = rte_vfio_get_group_fd(iommu_group_num);
663 if (vfio_group_fd < 0)
666 /* if group_fd == 0, that means the device isn't managed by VFIO */
667 if (vfio_group_fd == 0) {
668 RTE_LOG(WARNING, EAL, " %s not managed by VFIO driver, skipping\n",
674 * at this point, we know that this group is viable (meaning, all devices
675 * are either bound to VFIO or not bound to anything)
678 /* check if the group is viable */
679 ret = ioctl(vfio_group_fd, VFIO_GROUP_GET_STATUS, &group_status);
681 RTE_LOG(ERR, EAL, " %s cannot get group status, "
682 "error %i (%s)\n", dev_addr, errno, strerror(errno));
683 close(vfio_group_fd);
684 rte_vfio_clear_group(vfio_group_fd);
686 } else if (!(group_status.flags & VFIO_GROUP_FLAGS_VIABLE)) {
687 RTE_LOG(ERR, EAL, " %s VFIO group is not viable! "
688 "Not all devices in IOMMU group bound to VFIO or unbound\n",
690 close(vfio_group_fd);
691 rte_vfio_clear_group(vfio_group_fd);
695 /* get the vfio_config it belongs to */
696 vfio_cfg = get_vfio_cfg_by_group_num(iommu_group_num);
697 vfio_cfg = vfio_cfg ? vfio_cfg : default_vfio_cfg;
698 vfio_container_fd = vfio_cfg->vfio_container_fd;
699 user_mem_maps = &vfio_cfg->mem_maps;
701 /* check if group does not have a container yet */
702 if (!(group_status.flags & VFIO_GROUP_FLAGS_CONTAINER_SET)) {
704 /* add group to a container */
705 ret = ioctl(vfio_group_fd, VFIO_GROUP_SET_CONTAINER,
708 RTE_LOG(ERR, EAL, " %s cannot add VFIO group to container, "
709 "error %i (%s)\n", dev_addr, errno, strerror(errno));
710 close(vfio_group_fd);
711 rte_vfio_clear_group(vfio_group_fd);
716 * pick an IOMMU type and set up DMA mappings for container
718 * needs to be done only once, only when first group is
719 * assigned to a container and only in primary process.
720 * Note this can happen several times with the hotplug
723 if (internal_config.process_type == RTE_PROC_PRIMARY &&
724 vfio_cfg->vfio_active_groups == 1 &&
725 vfio_group_device_count(vfio_group_fd) == 0) {
726 const struct vfio_iommu_type *t;
728 /* select an IOMMU type which we will be using */
729 t = vfio_set_iommu_type(vfio_container_fd);
732 " %s failed to select IOMMU type\n",
734 close(vfio_group_fd);
735 rte_vfio_clear_group(vfio_group_fd);
738 /* lock memory hotplug before mapping and release it
739 * after registering callback, to prevent races
741 rte_mcfg_mem_read_lock();
742 if (vfio_cfg == default_vfio_cfg)
743 ret = t->dma_map_func(vfio_container_fd);
748 " %s DMA remapping failed, error %i (%s)\n",
749 dev_addr, errno, strerror(errno));
750 close(vfio_group_fd);
751 rte_vfio_clear_group(vfio_group_fd);
752 rte_mcfg_mem_read_unlock();
756 vfio_cfg->vfio_iommu_type = t;
758 /* re-map all user-mapped segments */
759 rte_spinlock_recursive_lock(&user_mem_maps->lock);
761 /* this IOMMU type may not support DMA mapping, but
762 * if we have mappings in the list - that means we have
763 * previously mapped something successfully, so we can
764 * be sure that DMA mapping is supported.
766 for (i = 0; i < user_mem_maps->n_maps; i++) {
767 struct user_mem_map *map;
768 map = &user_mem_maps->maps[i];
770 ret = t->dma_user_map_func(
772 map->addr, map->iova, map->len,
775 RTE_LOG(ERR, EAL, "Couldn't map user memory for DMA: "
777 "iova: 0x%" PRIx64 " "
778 "len: 0x%" PRIu64 "\n",
779 map->addr, map->iova,
781 rte_spinlock_recursive_unlock(
782 &user_mem_maps->lock);
783 rte_mcfg_mem_read_unlock();
787 rte_spinlock_recursive_unlock(&user_mem_maps->lock);
789 /* register callback for mem events */
790 if (vfio_cfg == default_vfio_cfg)
791 ret = rte_mem_event_callback_register(
792 VFIO_MEM_EVENT_CLB_NAME,
793 vfio_mem_event_callback, NULL);
796 /* unlock memory hotplug */
797 rte_mcfg_mem_read_unlock();
799 if (ret && rte_errno != ENOTSUP) {
800 RTE_LOG(ERR, EAL, "Could not install memory event callback for VFIO\n");
804 RTE_LOG(DEBUG, EAL, "Memory event callbacks not supported\n");
806 RTE_LOG(DEBUG, EAL, "Installed memory event callback for VFIO\n");
808 } else if (rte_eal_process_type() != RTE_PROC_PRIMARY &&
809 vfio_cfg == default_vfio_cfg &&
810 vfio_cfg->vfio_iommu_type == NULL) {
811 /* if we're not a primary process, we do not set up the VFIO
812 * container because it's already been set up by the primary
813 * process. instead, we simply ask the primary about VFIO type
814 * we are using, and set the VFIO config up appropriately.
816 ret = vfio_sync_default_container();
818 RTE_LOG(ERR, EAL, "Could not sync default VFIO container\n");
819 close(vfio_group_fd);
820 rte_vfio_clear_group(vfio_group_fd);
823 /* we have successfully initialized VFIO, notify user */
824 const struct vfio_iommu_type *t =
825 default_vfio_cfg->vfio_iommu_type;
826 RTE_LOG(NOTICE, EAL, " using IOMMU type %d (%s)\n",
827 t->type_id, t->name);
830 /* get a file descriptor for the device */
831 *vfio_dev_fd = ioctl(vfio_group_fd, VFIO_GROUP_GET_DEVICE_FD, dev_addr);
832 if (*vfio_dev_fd < 0) {
833 /* if we cannot get a device fd, this implies a problem with
834 * the VFIO group or the container not having IOMMU configured.
837 RTE_LOG(WARNING, EAL, "Getting a vfio_dev_fd for %s failed\n",
839 close(vfio_group_fd);
840 rte_vfio_clear_group(vfio_group_fd);
844 /* test and setup the device */
845 ret = ioctl(*vfio_dev_fd, VFIO_DEVICE_GET_INFO, device_info);
847 RTE_LOG(ERR, EAL, " %s cannot get device info, "
848 "error %i (%s)\n", dev_addr, errno,
851 close(vfio_group_fd);
852 rte_vfio_clear_group(vfio_group_fd);
855 vfio_group_device_get(vfio_group_fd);
861 rte_vfio_release_device(const char *sysfs_base, const char *dev_addr,
864 struct vfio_group_status group_status = {
865 .argsz = sizeof(group_status)
867 struct vfio_config *vfio_cfg;
872 /* we don't want any DMA mapping messages to come while we're detaching
873 * VFIO device, because this might be the last device and we might need
874 * to unregister the callback.
876 rte_mcfg_mem_read_lock();
878 /* get group number */
879 ret = rte_vfio_get_group_num(sysfs_base, dev_addr, &iommu_group_num);
881 RTE_LOG(WARNING, EAL, " %s not managed by VFIO driver\n",
883 /* This is an error at this point. */
888 /* get the actual group fd */
889 vfio_group_fd = rte_vfio_get_group_fd(iommu_group_num);
890 if (vfio_group_fd <= 0) {
891 RTE_LOG(INFO, EAL, "rte_vfio_get_group_fd failed for %s\n",
897 /* get the vfio_config it belongs to */
898 vfio_cfg = get_vfio_cfg_by_group_num(iommu_group_num);
899 vfio_cfg = vfio_cfg ? vfio_cfg : default_vfio_cfg;
901 /* At this point we got an active group. Closing it will make the
902 * container detachment. If this is the last active group, VFIO kernel
903 * code will unset the container and the IOMMU mappings.
906 /* Closing a device */
907 if (close(vfio_dev_fd) < 0) {
908 RTE_LOG(INFO, EAL, "Error when closing vfio_dev_fd for %s\n",
914 /* An VFIO group can have several devices attached. Just when there is
915 * no devices remaining should the group be closed.
917 vfio_group_device_put(vfio_group_fd);
918 if (!vfio_group_device_count(vfio_group_fd)) {
920 if (close(vfio_group_fd) < 0) {
921 RTE_LOG(INFO, EAL, "Error when closing vfio_group_fd for %s\n",
927 if (rte_vfio_clear_group(vfio_group_fd) < 0) {
928 RTE_LOG(INFO, EAL, "Error when clearing group for %s\n",
935 /* if there are no active device groups, unregister the callback to
936 * avoid spurious attempts to map/unmap memory from VFIO.
938 if (vfio_cfg == default_vfio_cfg && vfio_cfg->vfio_active_groups == 0 &&
939 rte_eal_process_type() != RTE_PROC_SECONDARY)
940 rte_mem_event_callback_unregister(VFIO_MEM_EVENT_CLB_NAME,
947 rte_mcfg_mem_read_unlock();
952 rte_vfio_enable(const char *modname)
954 /* initialize group list */
958 rte_spinlock_recursive_t lock = RTE_SPINLOCK_RECURSIVE_INITIALIZER;
960 for (i = 0; i < VFIO_MAX_CONTAINERS; i++) {
961 vfio_cfgs[i].vfio_container_fd = -1;
962 vfio_cfgs[i].vfio_active_groups = 0;
963 vfio_cfgs[i].vfio_iommu_type = NULL;
964 vfio_cfgs[i].mem_maps.lock = lock;
966 for (j = 0; j < VFIO_MAX_GROUPS; j++) {
967 vfio_cfgs[i].vfio_groups[j].fd = -1;
968 vfio_cfgs[i].vfio_groups[j].group_num = -1;
969 vfio_cfgs[i].vfio_groups[j].devices = 0;
973 /* inform the user that we are probing for VFIO */
974 RTE_LOG(INFO, EAL, "Probing VFIO support...\n");
976 /* check if vfio module is loaded */
977 vfio_available = rte_eal_check_module(modname);
979 /* return error directly */
980 if (vfio_available == -1) {
981 RTE_LOG(INFO, EAL, "Could not get loaded module details!\n");
985 /* return 0 if VFIO modules not loaded */
986 if (vfio_available == 0) {
987 RTE_LOG(DEBUG, EAL, "VFIO modules not loaded, "
988 "skipping VFIO support...\n");
992 if (internal_config.process_type == RTE_PROC_PRIMARY) {
993 /* open a new container */
994 default_vfio_cfg->vfio_container_fd =
995 rte_vfio_get_container_fd();
997 /* get the default container from the primary process */
998 default_vfio_cfg->vfio_container_fd =
999 vfio_get_default_container_fd();
1002 /* check if we have VFIO driver enabled */
1003 if (default_vfio_cfg->vfio_container_fd != -1) {
1004 RTE_LOG(NOTICE, EAL, "VFIO support initialized\n");
1005 default_vfio_cfg->vfio_enabled = 1;
1007 RTE_LOG(NOTICE, EAL, "VFIO support could not be initialized\n");
1014 rte_vfio_is_enabled(const char *modname)
1016 const int mod_available = rte_eal_check_module(modname) > 0;
1017 return default_vfio_cfg->vfio_enabled && mod_available;
1021 vfio_get_default_container_fd(void)
1023 struct rte_mp_msg mp_req, *mp_rep;
1024 struct rte_mp_reply mp_reply;
1025 struct timespec ts = {.tv_sec = 5, .tv_nsec = 0};
1026 struct vfio_mp_param *p = (struct vfio_mp_param *)mp_req.param;
1028 if (default_vfio_cfg->vfio_enabled)
1029 return default_vfio_cfg->vfio_container_fd;
1031 if (internal_config.process_type == RTE_PROC_PRIMARY) {
1032 /* if we were secondary process we would try requesting
1033 * container fd from the primary, but we're the primary
1034 * process so just exit here
1039 p->req = SOCKET_REQ_DEFAULT_CONTAINER;
1040 strcpy(mp_req.name, EAL_VFIO_MP);
1041 mp_req.len_param = sizeof(*p);
1044 if (rte_mp_request_sync(&mp_req, &mp_reply, &ts) == 0 &&
1045 mp_reply.nb_received == 1) {
1046 mp_rep = &mp_reply.msgs[0];
1047 p = (struct vfio_mp_param *)mp_rep->param;
1048 if (p->result == SOCKET_OK && mp_rep->num_fds == 1) {
1049 free(mp_reply.msgs);
1050 return mp_rep->fds[0];
1052 free(mp_reply.msgs);
1055 RTE_LOG(ERR, EAL, " cannot request default container fd\n");
1060 vfio_get_iommu_type(void)
1062 if (default_vfio_cfg->vfio_iommu_type == NULL)
1065 return default_vfio_cfg->vfio_iommu_type->type_id;
1068 const struct vfio_iommu_type *
1069 vfio_set_iommu_type(int vfio_container_fd)
1072 for (idx = 0; idx < RTE_DIM(iommu_types); idx++) {
1073 const struct vfio_iommu_type *t = &iommu_types[idx];
1075 int ret = ioctl(vfio_container_fd, VFIO_SET_IOMMU,
1078 RTE_LOG(NOTICE, EAL, " using IOMMU type %d (%s)\n",
1079 t->type_id, t->name);
1082 /* not an error, there may be more supported IOMMU types */
1083 RTE_LOG(DEBUG, EAL, " set IOMMU type %d (%s) failed, "
1084 "error %i (%s)\n", t->type_id, t->name, errno,
1087 /* if we didn't find a suitable IOMMU type, fail */
1092 vfio_has_supported_extensions(int vfio_container_fd)
1095 unsigned idx, n_extensions = 0;
1096 for (idx = 0; idx < RTE_DIM(iommu_types); idx++) {
1097 const struct vfio_iommu_type *t = &iommu_types[idx];
1099 ret = ioctl(vfio_container_fd, VFIO_CHECK_EXTENSION,
1102 RTE_LOG(ERR, EAL, " could not get IOMMU type, "
1103 "error %i (%s)\n", errno,
1105 close(vfio_container_fd);
1107 } else if (ret == 1) {
1108 /* we found a supported extension */
1111 RTE_LOG(DEBUG, EAL, " IOMMU type %d (%s) is %s\n",
1112 t->type_id, t->name,
1113 ret ? "supported" : "not supported");
1116 /* if we didn't find any supported IOMMU types, fail */
1117 if (!n_extensions) {
1118 close(vfio_container_fd);
1126 rte_vfio_get_container_fd(void)
1128 int ret, vfio_container_fd;
1129 struct rte_mp_msg mp_req, *mp_rep;
1130 struct rte_mp_reply mp_reply;
1131 struct timespec ts = {.tv_sec = 5, .tv_nsec = 0};
1132 struct vfio_mp_param *p = (struct vfio_mp_param *)mp_req.param;
1135 /* if we're in a primary process, try to open the container */
1136 if (internal_config.process_type == RTE_PROC_PRIMARY) {
1137 vfio_container_fd = open(VFIO_CONTAINER_PATH, O_RDWR);
1138 if (vfio_container_fd < 0) {
1139 RTE_LOG(ERR, EAL, " cannot open VFIO container, "
1140 "error %i (%s)\n", errno, strerror(errno));
1144 /* check VFIO API version */
1145 ret = ioctl(vfio_container_fd, VFIO_GET_API_VERSION);
1146 if (ret != VFIO_API_VERSION) {
1148 RTE_LOG(ERR, EAL, " could not get VFIO API version, "
1149 "error %i (%s)\n", errno, strerror(errno));
1151 RTE_LOG(ERR, EAL, " unsupported VFIO API version!\n");
1152 close(vfio_container_fd);
1156 ret = vfio_has_supported_extensions(vfio_container_fd);
1158 RTE_LOG(ERR, EAL, " no supported IOMMU "
1159 "extensions found!\n");
1163 return vfio_container_fd;
1166 * if we're in a secondary process, request container fd from the
1167 * primary process via mp channel
1169 p->req = SOCKET_REQ_CONTAINER;
1170 strcpy(mp_req.name, EAL_VFIO_MP);
1171 mp_req.len_param = sizeof(*p);
1174 vfio_container_fd = -1;
1175 if (rte_mp_request_sync(&mp_req, &mp_reply, &ts) == 0 &&
1176 mp_reply.nb_received == 1) {
1177 mp_rep = &mp_reply.msgs[0];
1178 p = (struct vfio_mp_param *)mp_rep->param;
1179 if (p->result == SOCKET_OK && mp_rep->num_fds == 1) {
1180 vfio_container_fd = mp_rep->fds[0];
1181 free(mp_reply.msgs);
1182 return vfio_container_fd;
1184 free(mp_reply.msgs);
1187 RTE_LOG(ERR, EAL, " cannot request container fd\n");
1192 rte_vfio_get_group_num(const char *sysfs_base,
1193 const char *dev_addr, int *iommu_group_num)
1195 char linkname[PATH_MAX];
1196 char filename[PATH_MAX];
1197 char *tok[16], *group_tok, *end;
1200 memset(linkname, 0, sizeof(linkname));
1201 memset(filename, 0, sizeof(filename));
1203 /* try to find out IOMMU group for this device */
1204 snprintf(linkname, sizeof(linkname),
1205 "%s/%s/iommu_group", sysfs_base, dev_addr);
1207 ret = readlink(linkname, filename, sizeof(filename));
1209 /* if the link doesn't exist, no VFIO for us */
1213 ret = rte_strsplit(filename, sizeof(filename),
1214 tok, RTE_DIM(tok), '/');
1217 RTE_LOG(ERR, EAL, " %s cannot get IOMMU group\n", dev_addr);
1221 /* IOMMU group is always the last token */
1223 group_tok = tok[ret - 1];
1225 *iommu_group_num = strtol(group_tok, &end, 10);
1226 if ((end != group_tok && *end != '\0') || errno != 0) {
1227 RTE_LOG(ERR, EAL, " %s error parsing IOMMU number!\n", dev_addr);
1235 type1_map_contig(const struct rte_memseg_list *msl, const struct rte_memseg *ms,
1236 size_t len, void *arg)
1238 int *vfio_container_fd = arg;
1243 return vfio_type1_dma_mem_map(*vfio_container_fd, ms->addr_64, ms->iova,
1248 type1_map(const struct rte_memseg_list *msl, const struct rte_memseg *ms,
1251 int *vfio_container_fd = arg;
1256 return vfio_type1_dma_mem_map(*vfio_container_fd, ms->addr_64, ms->iova,
1261 vfio_type1_dma_mem_map(int vfio_container_fd, uint64_t vaddr, uint64_t iova,
1262 uint64_t len, int do_map)
1264 struct vfio_iommu_type1_dma_map dma_map;
1265 struct vfio_iommu_type1_dma_unmap dma_unmap;
1269 memset(&dma_map, 0, sizeof(dma_map));
1270 dma_map.argsz = sizeof(struct vfio_iommu_type1_dma_map);
1271 dma_map.vaddr = vaddr;
1273 dma_map.iova = iova;
1274 dma_map.flags = VFIO_DMA_MAP_FLAG_READ |
1275 VFIO_DMA_MAP_FLAG_WRITE;
1277 ret = ioctl(vfio_container_fd, VFIO_IOMMU_MAP_DMA, &dma_map);
1280 * In case the mapping was already done EEXIST will be
1281 * returned from kernel.
1283 if (errno == EEXIST) {
1285 " Memory segment is allready mapped,"
1289 " cannot set up DMA remapping,"
1291 errno, strerror(errno));
1296 memset(&dma_unmap, 0, sizeof(dma_unmap));
1297 dma_unmap.argsz = sizeof(struct vfio_iommu_type1_dma_unmap);
1298 dma_unmap.size = len;
1299 dma_unmap.iova = iova;
1301 ret = ioctl(vfio_container_fd, VFIO_IOMMU_UNMAP_DMA,
1304 RTE_LOG(ERR, EAL, " cannot clear DMA remapping, error %i (%s)\n",
1305 errno, strerror(errno));
1314 vfio_type1_dma_map(int vfio_container_fd)
1316 if (rte_eal_iova_mode() == RTE_IOVA_VA) {
1317 /* with IOVA as VA mode, we can get away with mapping contiguous
1318 * chunks rather than going page-by-page.
1320 return rte_memseg_contig_walk(type1_map_contig,
1321 &vfio_container_fd);
1323 return rte_memseg_walk(type1_map, &vfio_container_fd);
1327 vfio_spapr_dma_do_map(int vfio_container_fd, uint64_t vaddr, uint64_t iova,
1328 uint64_t len, int do_map)
1330 struct vfio_iommu_type1_dma_map dma_map;
1331 struct vfio_iommu_type1_dma_unmap dma_unmap;
1333 struct vfio_iommu_spapr_register_memory reg = {
1334 .argsz = sizeof(reg),
1337 reg.vaddr = (uintptr_t) vaddr;
1341 ret = ioctl(vfio_container_fd,
1342 VFIO_IOMMU_SPAPR_REGISTER_MEMORY, ®);
1344 RTE_LOG(ERR, EAL, " cannot register vaddr for IOMMU, "
1345 "error %i (%s)\n", errno, strerror(errno));
1349 memset(&dma_map, 0, sizeof(dma_map));
1350 dma_map.argsz = sizeof(struct vfio_iommu_type1_dma_map);
1351 dma_map.vaddr = vaddr;
1353 dma_map.iova = iova;
1354 dma_map.flags = VFIO_DMA_MAP_FLAG_READ |
1355 VFIO_DMA_MAP_FLAG_WRITE;
1357 ret = ioctl(vfio_container_fd, VFIO_IOMMU_MAP_DMA, &dma_map);
1360 * In case the mapping was already done EBUSY will be
1361 * returned from kernel.
1363 if (errno == EBUSY) {
1365 " Memory segment is allready mapped,"
1369 " cannot set up DMA remapping,"
1370 " error %i (%s)\n", errno,
1377 memset(&dma_unmap, 0, sizeof(dma_unmap));
1378 dma_unmap.argsz = sizeof(struct vfio_iommu_type1_dma_unmap);
1379 dma_unmap.size = len;
1380 dma_unmap.iova = iova;
1382 ret = ioctl(vfio_container_fd, VFIO_IOMMU_UNMAP_DMA,
1385 RTE_LOG(ERR, EAL, " cannot clear DMA remapping, error %i (%s)\n",
1386 errno, strerror(errno));
1390 ret = ioctl(vfio_container_fd,
1391 VFIO_IOMMU_SPAPR_UNREGISTER_MEMORY, ®);
1393 RTE_LOG(ERR, EAL, " cannot unregister vaddr for IOMMU, error %i (%s)\n",
1394 errno, strerror(errno));
1402 struct spapr_remap_walk_param {
1403 int vfio_container_fd;
1408 vfio_spapr_map_walk(const struct rte_memseg_list *msl,
1409 const struct rte_memseg *ms, void *arg)
1411 struct spapr_remap_walk_param *param = arg;
1413 if (msl->external || ms->addr_64 == param->addr_64)
1416 return vfio_spapr_dma_do_map(param->vfio_container_fd, ms->addr_64, ms->iova,
1421 vfio_spapr_unmap_walk(const struct rte_memseg_list *msl,
1422 const struct rte_memseg *ms, void *arg)
1424 struct spapr_remap_walk_param *param = arg;
1426 if (msl->external || ms->addr_64 == param->addr_64)
1429 return vfio_spapr_dma_do_map(param->vfio_container_fd, ms->addr_64, ms->iova,
1433 struct spapr_walk_param {
1434 uint64_t window_size;
1435 uint64_t hugepage_sz;
1440 vfio_spapr_window_size_walk(const struct rte_memseg_list *msl,
1441 const struct rte_memseg *ms, void *arg)
1443 struct spapr_walk_param *param = arg;
1444 uint64_t max = ms->iova + ms->len;
1449 /* do not iterate ms we haven't mapped yet */
1450 if (param->addr_64 && ms->addr_64 == param->addr_64)
1453 if (max > param->window_size) {
1454 param->hugepage_sz = ms->hugepage_sz;
1455 param->window_size = max;
1462 vfio_spapr_create_new_dma_window(int vfio_container_fd,
1463 struct vfio_iommu_spapr_tce_create *create) {
1464 struct vfio_iommu_spapr_tce_remove remove = {
1465 .argsz = sizeof(remove),
1467 struct vfio_iommu_spapr_tce_info info = {
1468 .argsz = sizeof(info),
1472 /* query spapr iommu info */
1473 ret = ioctl(vfio_container_fd, VFIO_IOMMU_SPAPR_TCE_GET_INFO, &info);
1475 RTE_LOG(ERR, EAL, " cannot get iommu info, "
1476 "error %i (%s)\n", errno, strerror(errno));
1480 /* remove default DMA of 32 bit window */
1481 remove.start_addr = info.dma32_window_start;
1482 ret = ioctl(vfio_container_fd, VFIO_IOMMU_SPAPR_TCE_REMOVE, &remove);
1484 RTE_LOG(ERR, EAL, " cannot remove default DMA window, "
1485 "error %i (%s)\n", errno, strerror(errno));
1489 /* create new DMA window */
1490 ret = ioctl(vfio_container_fd, VFIO_IOMMU_SPAPR_TCE_CREATE, create);
1492 #ifdef VFIO_IOMMU_SPAPR_INFO_DDW
1493 /* try possible page_shift and levels for workaround */
1496 for (levels = 1; levels <= info.ddw.levels; levels++) {
1497 uint32_t pgsizes = info.ddw.pgsizes;
1499 while (pgsizes != 0) {
1500 create->page_shift = 31 - __builtin_clz(pgsizes);
1501 create->levels = levels;
1502 ret = ioctl(vfio_container_fd,
1503 VFIO_IOMMU_SPAPR_TCE_CREATE, create);
1506 pgsizes &= ~(1 << create->page_shift);
1513 RTE_LOG(ERR, EAL, " cannot create new DMA window, "
1514 "error %i (%s)\n", errno, strerror(errno));
1519 if (create->start_addr != 0) {
1520 RTE_LOG(ERR, EAL, " DMA window start address != 0\n");
1528 vfio_spapr_dma_mem_map(int vfio_container_fd, uint64_t vaddr, uint64_t iova,
1529 uint64_t len, int do_map)
1531 struct spapr_walk_param param;
1532 struct vfio_iommu_spapr_tce_create create = {
1533 .argsz = sizeof(create),
1535 struct vfio_config *vfio_cfg;
1536 struct user_mem_maps *user_mem_maps;
1539 vfio_cfg = get_vfio_cfg_by_container_fd(vfio_container_fd);
1540 if (vfio_cfg == NULL) {
1541 RTE_LOG(ERR, EAL, " invalid container fd!\n");
1545 user_mem_maps = &vfio_cfg->mem_maps;
1546 rte_spinlock_recursive_lock(&user_mem_maps->lock);
1548 /* check if window size needs to be adjusted */
1549 memset(¶m, 0, sizeof(param));
1550 param.addr_64 = vaddr;
1552 /* we're inside a callback so use thread-unsafe version */
1553 if (rte_memseg_walk_thread_unsafe(vfio_spapr_window_size_walk,
1555 RTE_LOG(ERR, EAL, "Could not get window size\n");
1560 /* also check user maps */
1561 for (i = 0; i < user_mem_maps->n_maps; i++) {
1562 uint64_t max = user_mem_maps->maps[i].iova +
1563 user_mem_maps->maps[i].len;
1564 param.window_size = RTE_MAX(param.window_size, max);
1567 /* sPAPR requires window size to be a power of 2 */
1568 create.window_size = rte_align64pow2(param.window_size);
1569 create.page_shift = __builtin_ctzll(param.hugepage_sz);
1573 /* re-create window and remap the entire memory */
1574 if (iova + len > create.window_size) {
1575 struct spapr_remap_walk_param remap_param = {
1576 .vfio_container_fd = vfio_container_fd,
1580 /* release all maps before recreating the window */
1581 if (rte_memseg_walk_thread_unsafe(vfio_spapr_unmap_walk,
1582 &remap_param) < 0) {
1583 RTE_LOG(ERR, EAL, "Could not release DMA maps\n");
1587 /* release all user maps */
1588 for (i = 0; i < user_mem_maps->n_maps; i++) {
1589 struct user_mem_map *map =
1590 &user_mem_maps->maps[i];
1591 if (vfio_spapr_dma_do_map(vfio_container_fd,
1592 map->addr, map->iova, map->len,
1594 RTE_LOG(ERR, EAL, "Could not release user DMA maps\n");
1599 create.window_size = rte_align64pow2(iova + len);
1600 if (vfio_spapr_create_new_dma_window(vfio_container_fd,
1602 RTE_LOG(ERR, EAL, "Could not create new DMA window\n");
1606 /* we're inside a callback, so use thread-unsafe version
1608 if (rte_memseg_walk_thread_unsafe(vfio_spapr_map_walk,
1609 &remap_param) < 0) {
1610 RTE_LOG(ERR, EAL, "Could not recreate DMA maps\n");
1614 /* remap all user maps */
1615 for (i = 0; i < user_mem_maps->n_maps; i++) {
1616 struct user_mem_map *map =
1617 &user_mem_maps->maps[i];
1618 if (vfio_spapr_dma_do_map(vfio_container_fd,
1619 map->addr, map->iova, map->len,
1621 RTE_LOG(ERR, EAL, "Could not recreate user DMA maps\n");
1627 if (vfio_spapr_dma_do_map(vfio_container_fd, vaddr, iova, len, 1)) {
1628 RTE_LOG(ERR, EAL, "Failed to map DMA\n");
1633 /* for unmap, check if iova within DMA window */
1634 if (iova > create.window_size) {
1635 RTE_LOG(ERR, EAL, "iova beyond DMA window for unmap");
1640 vfio_spapr_dma_do_map(vfio_container_fd, vaddr, iova, len, 0);
1643 rte_spinlock_recursive_unlock(&user_mem_maps->lock);
1648 vfio_spapr_dma_map(int vfio_container_fd)
1650 struct vfio_iommu_spapr_tce_create create = {
1651 .argsz = sizeof(create),
1653 struct spapr_walk_param param;
1655 memset(¶m, 0, sizeof(param));
1656 param.addr_64 = 0UL;
1658 /* create DMA window from 0 to max(phys_addr + len) */
1659 rte_memseg_walk(vfio_spapr_window_size_walk, ¶m);
1661 /* sPAPR requires window size to be a power of 2 */
1662 create.window_size = rte_align64pow2(param.window_size);
1663 create.page_shift = __builtin_ctzll(param.hugepage_sz);
1666 if (vfio_spapr_create_new_dma_window(vfio_container_fd, &create) < 0) {
1667 RTE_LOG(ERR, EAL, "Could not create new DMA window\n");
1671 /* map all DPDK segments for DMA. use 1:1 PA to IOVA mapping */
1672 if (rte_memseg_walk(vfio_spapr_map_walk, &vfio_container_fd) < 0)
1679 vfio_noiommu_dma_map(int __rte_unused vfio_container_fd)
1681 /* No-IOMMU mode does not need DMA mapping */
1686 vfio_noiommu_dma_mem_map(int __rte_unused vfio_container_fd,
1687 uint64_t __rte_unused vaddr,
1688 uint64_t __rte_unused iova, uint64_t __rte_unused len,
1689 int __rte_unused do_map)
1691 /* No-IOMMU mode does not need DMA mapping */
1696 vfio_dma_mem_map(struct vfio_config *vfio_cfg, uint64_t vaddr, uint64_t iova,
1697 uint64_t len, int do_map)
1699 const struct vfio_iommu_type *t = vfio_cfg->vfio_iommu_type;
1702 RTE_LOG(ERR, EAL, " VFIO support not initialized\n");
1707 if (!t->dma_user_map_func) {
1709 " VFIO custom DMA region maping not supported by IOMMU %s\n",
1711 rte_errno = ENOTSUP;
1715 return t->dma_user_map_func(vfio_cfg->vfio_container_fd, vaddr, iova,
1720 container_dma_map(struct vfio_config *vfio_cfg, uint64_t vaddr, uint64_t iova,
1723 struct user_mem_map *new_map;
1724 struct user_mem_maps *user_mem_maps;
1727 user_mem_maps = &vfio_cfg->mem_maps;
1728 rte_spinlock_recursive_lock(&user_mem_maps->lock);
1729 if (user_mem_maps->n_maps == VFIO_MAX_USER_MEM_MAPS) {
1730 RTE_LOG(ERR, EAL, "No more space for user mem maps\n");
1736 if (vfio_dma_mem_map(vfio_cfg, vaddr, iova, len, 1)) {
1737 /* technically, this will fail if there are currently no devices
1738 * plugged in, even if a device were added later, this mapping
1739 * might have succeeded. however, since we cannot verify if this
1740 * is a valid mapping without having a device attached, consider
1741 * this to be unsupported, because we can't just store any old
1742 * mapping and pollute list of active mappings willy-nilly.
1744 RTE_LOG(ERR, EAL, "Couldn't map new region for DMA\n");
1748 /* create new user mem map entry */
1749 new_map = &user_mem_maps->maps[user_mem_maps->n_maps++];
1750 new_map->addr = vaddr;
1751 new_map->iova = iova;
1754 compact_user_maps(user_mem_maps);
1756 rte_spinlock_recursive_unlock(&user_mem_maps->lock);
1761 container_dma_unmap(struct vfio_config *vfio_cfg, uint64_t vaddr, uint64_t iova,
1764 struct user_mem_map *map, *new_map = NULL;
1765 struct user_mem_maps *user_mem_maps;
1768 user_mem_maps = &vfio_cfg->mem_maps;
1769 rte_spinlock_recursive_lock(&user_mem_maps->lock);
1771 /* find our mapping */
1772 map = find_user_mem_map(user_mem_maps, vaddr, iova, len);
1774 RTE_LOG(ERR, EAL, "Couldn't find previously mapped region\n");
1779 if (map->addr != vaddr || map->iova != iova || map->len != len) {
1780 /* we're partially unmapping a previously mapped region, so we
1781 * need to split entry into two.
1783 if (user_mem_maps->n_maps == VFIO_MAX_USER_MEM_MAPS) {
1784 RTE_LOG(ERR, EAL, "Not enough space to store partial mapping\n");
1789 new_map = &user_mem_maps->maps[user_mem_maps->n_maps++];
1792 /* unmap the entry */
1793 if (vfio_dma_mem_map(vfio_cfg, vaddr, iova, len, 0)) {
1794 /* there may not be any devices plugged in, so unmapping will
1795 * fail with ENODEV/ENOTSUP rte_errno values, but that doesn't
1796 * stop us from removing the mapping, as the assumption is we
1797 * won't be needing this memory any more and thus will want to
1798 * prevent it from being remapped again on hotplug. so, only
1799 * fail if we indeed failed to unmap (e.g. if the mapping was
1800 * within our mapped range but had invalid alignment).
1802 if (rte_errno != ENODEV && rte_errno != ENOTSUP) {
1803 RTE_LOG(ERR, EAL, "Couldn't unmap region for DMA\n");
1807 RTE_LOG(DEBUG, EAL, "DMA unmapping failed, but removing mappings anyway\n");
1810 /* remove map from the list of active mappings */
1811 if (new_map != NULL) {
1812 adjust_map(map, new_map, vaddr, len);
1814 /* if we've created a new map by splitting, sort everything */
1815 if (!is_null_map(new_map)) {
1816 compact_user_maps(user_mem_maps);
1818 /* we've created a new mapping, but it was unused */
1819 user_mem_maps->n_maps--;
1822 memset(map, 0, sizeof(*map));
1823 compact_user_maps(user_mem_maps);
1824 user_mem_maps->n_maps--;
1828 rte_spinlock_recursive_unlock(&user_mem_maps->lock);
1833 rte_vfio_dma_map(uint64_t vaddr, uint64_t iova, uint64_t len)
1840 return container_dma_map(default_vfio_cfg, vaddr, iova, len);
1844 rte_vfio_dma_unmap(uint64_t vaddr, uint64_t iova, uint64_t len)
1851 return container_dma_unmap(default_vfio_cfg, vaddr, iova, len);
1855 rte_vfio_noiommu_is_enabled(void)
1861 fd = open(VFIO_NOIOMMU_MODE, O_RDONLY);
1863 if (errno != ENOENT) {
1864 RTE_LOG(ERR, EAL, " cannot open vfio noiommu file %i (%s)\n",
1865 errno, strerror(errno));
1869 * else the file does not exists
1870 * i.e. noiommu is not enabled
1875 cnt = read(fd, &c, 1);
1878 RTE_LOG(ERR, EAL, " unable to read from vfio noiommu "
1879 "file %i (%s)\n", errno, strerror(errno));
1887 rte_vfio_container_create(void)
1891 /* Find an empty slot to store new vfio config */
1892 for (i = 1; i < VFIO_MAX_CONTAINERS; i++) {
1893 if (vfio_cfgs[i].vfio_container_fd == -1)
1897 if (i == VFIO_MAX_CONTAINERS) {
1898 RTE_LOG(ERR, EAL, "exceed max vfio container limit\n");
1902 vfio_cfgs[i].vfio_container_fd = rte_vfio_get_container_fd();
1903 if (vfio_cfgs[i].vfio_container_fd < 0) {
1904 RTE_LOG(NOTICE, EAL, "fail to create a new container\n");
1908 return vfio_cfgs[i].vfio_container_fd;
1912 rte_vfio_container_destroy(int container_fd)
1914 struct vfio_config *vfio_cfg;
1917 vfio_cfg = get_vfio_cfg_by_container_fd(container_fd);
1918 if (vfio_cfg == NULL) {
1919 RTE_LOG(ERR, EAL, "Invalid container fd\n");
1923 for (i = 0; i < VFIO_MAX_GROUPS; i++)
1924 if (vfio_cfg->vfio_groups[i].group_num != -1)
1925 rte_vfio_container_group_unbind(container_fd,
1926 vfio_cfg->vfio_groups[i].group_num);
1928 close(container_fd);
1929 vfio_cfg->vfio_container_fd = -1;
1930 vfio_cfg->vfio_active_groups = 0;
1931 vfio_cfg->vfio_iommu_type = NULL;
1937 rte_vfio_container_group_bind(int container_fd, int iommu_group_num)
1939 struct vfio_config *vfio_cfg;
1941 vfio_cfg = get_vfio_cfg_by_container_fd(container_fd);
1942 if (vfio_cfg == NULL) {
1943 RTE_LOG(ERR, EAL, "Invalid container fd\n");
1947 return vfio_get_group_fd(vfio_cfg, iommu_group_num);
1951 rte_vfio_container_group_unbind(int container_fd, int iommu_group_num)
1953 struct vfio_config *vfio_cfg;
1954 struct vfio_group *cur_grp = NULL;
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 for (i = 0; i < VFIO_MAX_GROUPS; i++) {
1964 if (vfio_cfg->vfio_groups[i].group_num == iommu_group_num) {
1965 cur_grp = &vfio_cfg->vfio_groups[i];
1970 /* This should not happen */
1971 if (i == VFIO_MAX_GROUPS || cur_grp == NULL) {
1972 RTE_LOG(ERR, EAL, "Specified group number not found\n");
1976 if (cur_grp->fd >= 0 && close(cur_grp->fd) < 0) {
1977 RTE_LOG(ERR, EAL, "Error when closing vfio_group_fd for"
1978 " iommu_group_num %d\n", iommu_group_num);
1981 cur_grp->group_num = -1;
1983 cur_grp->devices = 0;
1984 vfio_cfg->vfio_active_groups--;
1990 rte_vfio_container_dma_map(int container_fd, uint64_t vaddr, uint64_t iova,
1993 struct vfio_config *vfio_cfg;
2000 vfio_cfg = get_vfio_cfg_by_container_fd(container_fd);
2001 if (vfio_cfg == NULL) {
2002 RTE_LOG(ERR, EAL, "Invalid container fd\n");
2006 return container_dma_map(vfio_cfg, vaddr, iova, len);
2010 rte_vfio_container_dma_unmap(int container_fd, uint64_t vaddr, uint64_t iova,
2013 struct vfio_config *vfio_cfg;
2020 vfio_cfg = get_vfio_cfg_by_container_fd(container_fd);
2021 if (vfio_cfg == NULL) {
2022 RTE_LOG(ERR, EAL, "Invalid container fd\n");
2026 return container_dma_unmap(vfio_cfg, vaddr, iova, len);
2032 rte_vfio_dma_map(uint64_t __rte_unused vaddr, __rte_unused uint64_t iova,
2033 __rte_unused uint64_t len)
2039 rte_vfio_dma_unmap(uint64_t __rte_unused vaddr, uint64_t __rte_unused iova,
2040 __rte_unused uint64_t len)
2046 rte_vfio_setup_device(__rte_unused const char *sysfs_base,
2047 __rte_unused const char *dev_addr,
2048 __rte_unused int *vfio_dev_fd,
2049 __rte_unused struct vfio_device_info *device_info)
2055 rte_vfio_release_device(__rte_unused const char *sysfs_base,
2056 __rte_unused const char *dev_addr, __rte_unused int fd)
2062 rte_vfio_enable(__rte_unused const char *modname)
2068 rte_vfio_is_enabled(__rte_unused const char *modname)
2074 rte_vfio_noiommu_is_enabled(void)
2080 rte_vfio_clear_group(__rte_unused int vfio_group_fd)
2086 rte_vfio_get_group_num(__rte_unused const char *sysfs_base,
2087 __rte_unused const char *dev_addr,
2088 __rte_unused int *iommu_group_num)
2094 rte_vfio_get_container_fd(void)
2100 rte_vfio_get_group_fd(__rte_unused int iommu_group_num)
2106 rte_vfio_container_create(void)
2112 rte_vfio_container_destroy(__rte_unused int container_fd)
2118 rte_vfio_container_group_bind(__rte_unused int container_fd,
2119 __rte_unused int iommu_group_num)
2125 rte_vfio_container_group_unbind(__rte_unused int container_fd,
2126 __rte_unused int iommu_group_num)
2132 rte_vfio_container_dma_map(__rte_unused int container_fd,
2133 __rte_unused uint64_t vaddr,
2134 __rte_unused uint64_t iova,
2135 __rte_unused uint64_t len)
2141 rte_vfio_container_dma_unmap(__rte_unused int container_fd,
2142 __rte_unused uint64_t vaddr,
2143 __rte_unused uint64_t iova,
2144 __rte_unused uint64_t len)
2149 #endif /* VFIO_PRESENT */