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;
270 const struct internal_config *internal_conf =
271 eal_get_internal_configuration();
273 /* if primary, try to open the group */
274 if (internal_conf->process_type == RTE_PROC_PRIMARY) {
275 /* try regular group format */
276 snprintf(filename, sizeof(filename),
277 VFIO_GROUP_FMT, iommu_group_num);
278 vfio_group_fd = open(filename, O_RDWR);
279 if (vfio_group_fd < 0) {
280 /* if file not found, it's not an error */
281 if (errno != ENOENT) {
282 RTE_LOG(ERR, EAL, "Cannot open %s: %s\n", filename,
287 /* special case: try no-IOMMU path as well */
288 snprintf(filename, sizeof(filename),
289 VFIO_NOIOMMU_GROUP_FMT,
291 vfio_group_fd = open(filename, O_RDWR);
292 if (vfio_group_fd < 0) {
293 if (errno != ENOENT) {
294 RTE_LOG(ERR, EAL, "Cannot open %s: %s\n", filename,
300 /* noiommu group found */
303 return vfio_group_fd;
305 /* if we're in a secondary process, request group fd from the primary
306 * process via mp channel.
308 p->req = SOCKET_REQ_GROUP;
309 p->group_num = iommu_group_num;
310 strcpy(mp_req.name, EAL_VFIO_MP);
311 mp_req.len_param = sizeof(*p);
315 if (rte_mp_request_sync(&mp_req, &mp_reply, &ts) == 0 &&
316 mp_reply.nb_received == 1) {
317 mp_rep = &mp_reply.msgs[0];
318 p = (struct vfio_mp_param *)mp_rep->param;
319 if (p->result == SOCKET_OK && mp_rep->num_fds == 1) {
320 vfio_group_fd = mp_rep->fds[0];
321 } else if (p->result == SOCKET_NO_FD) {
322 RTE_LOG(ERR, EAL, " bad VFIO group fd\n");
328 if (vfio_group_fd < 0)
329 RTE_LOG(ERR, EAL, " cannot request group fd\n");
330 return vfio_group_fd;
333 static struct vfio_config *
334 get_vfio_cfg_by_group_num(int iommu_group_num)
336 struct vfio_config *vfio_cfg;
339 for (i = 0; i < VFIO_MAX_CONTAINERS; i++) {
340 vfio_cfg = &vfio_cfgs[i];
341 for (j = 0; j < VFIO_MAX_GROUPS; j++) {
342 if (vfio_cfg->vfio_groups[j].group_num ==
352 vfio_get_group_fd(struct vfio_config *vfio_cfg,
357 struct vfio_group *cur_grp;
359 /* check if we already have the group descriptor open */
360 for (i = 0; i < VFIO_MAX_GROUPS; i++)
361 if (vfio_cfg->vfio_groups[i].group_num == iommu_group_num)
362 return vfio_cfg->vfio_groups[i].fd;
364 /* Lets see first if there is room for a new group */
365 if (vfio_cfg->vfio_active_groups == VFIO_MAX_GROUPS) {
366 RTE_LOG(ERR, EAL, "Maximum number of VFIO groups reached!\n");
370 /* Now lets get an index for the new group */
371 for (i = 0; i < VFIO_MAX_GROUPS; i++)
372 if (vfio_cfg->vfio_groups[i].group_num == -1) {
373 cur_grp = &vfio_cfg->vfio_groups[i];
377 /* This should not happen */
378 if (i == VFIO_MAX_GROUPS) {
379 RTE_LOG(ERR, EAL, "No VFIO group free slot found\n");
383 vfio_group_fd = vfio_open_group_fd(iommu_group_num);
384 if (vfio_group_fd <= 0) {
385 RTE_LOG(ERR, EAL, "Failed to open group %d\n", iommu_group_num);
389 cur_grp->group_num = iommu_group_num;
390 cur_grp->fd = vfio_group_fd;
391 vfio_cfg->vfio_active_groups++;
393 return vfio_group_fd;
396 static struct vfio_config *
397 get_vfio_cfg_by_group_fd(int vfio_group_fd)
399 struct vfio_config *vfio_cfg;
402 for (i = 0; i < VFIO_MAX_CONTAINERS; i++) {
403 vfio_cfg = &vfio_cfgs[i];
404 for (j = 0; j < VFIO_MAX_GROUPS; j++)
405 if (vfio_cfg->vfio_groups[j].fd == vfio_group_fd)
412 static struct vfio_config *
413 get_vfio_cfg_by_container_fd(int container_fd)
417 if (container_fd == RTE_VFIO_DEFAULT_CONTAINER_FD)
418 return default_vfio_cfg;
420 for (i = 0; i < VFIO_MAX_CONTAINERS; i++) {
421 if (vfio_cfgs[i].vfio_container_fd == container_fd)
422 return &vfio_cfgs[i];
429 rte_vfio_get_group_fd(int iommu_group_num)
431 struct vfio_config *vfio_cfg;
433 /* get the vfio_config it belongs to */
434 vfio_cfg = get_vfio_cfg_by_group_num(iommu_group_num);
435 vfio_cfg = vfio_cfg ? vfio_cfg : default_vfio_cfg;
437 return vfio_get_group_fd(vfio_cfg, iommu_group_num);
441 get_vfio_group_idx(int vfio_group_fd)
443 struct vfio_config *vfio_cfg;
446 for (i = 0; i < VFIO_MAX_CONTAINERS; i++) {
447 vfio_cfg = &vfio_cfgs[i];
448 for (j = 0; j < VFIO_MAX_GROUPS; j++)
449 if (vfio_cfg->vfio_groups[j].fd == vfio_group_fd)
457 vfio_group_device_get(int vfio_group_fd)
459 struct vfio_config *vfio_cfg;
462 vfio_cfg = get_vfio_cfg_by_group_fd(vfio_group_fd);
463 if (vfio_cfg == NULL) {
464 RTE_LOG(ERR, EAL, " invalid group fd!\n");
468 i = get_vfio_group_idx(vfio_group_fd);
469 if (i < 0 || i > (VFIO_MAX_GROUPS - 1))
470 RTE_LOG(ERR, EAL, " wrong vfio_group index (%d)\n", i);
472 vfio_cfg->vfio_groups[i].devices++;
476 vfio_group_device_put(int vfio_group_fd)
478 struct vfio_config *vfio_cfg;
481 vfio_cfg = get_vfio_cfg_by_group_fd(vfio_group_fd);
482 if (vfio_cfg == NULL) {
483 RTE_LOG(ERR, EAL, " invalid group fd!\n");
487 i = get_vfio_group_idx(vfio_group_fd);
488 if (i < 0 || i > (VFIO_MAX_GROUPS - 1))
489 RTE_LOG(ERR, EAL, " wrong vfio_group index (%d)\n", i);
491 vfio_cfg->vfio_groups[i].devices--;
495 vfio_group_device_count(int vfio_group_fd)
497 struct vfio_config *vfio_cfg;
500 vfio_cfg = get_vfio_cfg_by_group_fd(vfio_group_fd);
501 if (vfio_cfg == NULL) {
502 RTE_LOG(ERR, EAL, " invalid group fd!\n");
506 i = get_vfio_group_idx(vfio_group_fd);
507 if (i < 0 || i > (VFIO_MAX_GROUPS - 1)) {
508 RTE_LOG(ERR, EAL, " wrong vfio_group index (%d)\n", i);
512 return vfio_cfg->vfio_groups[i].devices;
516 vfio_mem_event_callback(enum rte_mem_event type, const void *addr, size_t len,
517 void *arg __rte_unused)
519 rte_iova_t iova_start, iova_expected;
520 struct rte_memseg_list *msl;
521 struct rte_memseg *ms;
525 msl = rte_mem_virt2memseg_list(addr);
527 /* for IOVA as VA mode, no need to care for IOVA addresses */
528 if (rte_eal_iova_mode() == RTE_IOVA_VA && msl->external == 0) {
529 uint64_t vfio_va = (uint64_t)(uintptr_t)addr;
530 if (type == RTE_MEM_EVENT_ALLOC)
531 vfio_dma_mem_map(default_vfio_cfg, vfio_va, vfio_va,
534 vfio_dma_mem_map(default_vfio_cfg, vfio_va, vfio_va,
539 #ifdef RTE_ARCH_PPC_64
540 ms = rte_mem_virt2memseg(addr, msl);
541 while (cur_len < len) {
542 int idx = rte_fbarray_find_idx(&msl->memseg_arr, ms);
544 rte_fbarray_set_free(&msl->memseg_arr, idx);
550 /* memsegs are contiguous in memory */
551 ms = rte_mem_virt2memseg(addr, msl);
554 * This memory is not guaranteed to be contiguous, but it still could
555 * be, or it could have some small contiguous chunks. Since the number
556 * of VFIO mappings is limited, and VFIO appears to not concatenate
557 * adjacent mappings, we have to do this ourselves.
559 * So, find contiguous chunks, then map them.
561 va_start = ms->addr_64;
562 iova_start = iova_expected = ms->iova;
563 while (cur_len < len) {
564 bool new_contig_area = ms->iova != iova_expected;
565 bool last_seg = (len - cur_len) == ms->len;
566 bool skip_last = false;
568 /* only do mappings when current contiguous area ends */
569 if (new_contig_area) {
570 if (type == RTE_MEM_EVENT_ALLOC)
571 vfio_dma_mem_map(default_vfio_cfg, va_start,
573 iova_expected - iova_start, 1);
575 vfio_dma_mem_map(default_vfio_cfg, va_start,
577 iova_expected - iova_start, 0);
578 va_start = ms->addr_64;
579 iova_start = ms->iova;
581 /* some memory segments may have invalid IOVA */
582 if (ms->iova == RTE_BAD_IOVA) {
583 RTE_LOG(DEBUG, EAL, "Memory segment at %p has bad IOVA, skipping\n",
587 iova_expected = ms->iova + ms->len;
592 * don't count previous segment, and don't attempt to
593 * dereference a potentially invalid pointer.
595 if (skip_last && !last_seg) {
596 iova_expected = iova_start = ms->iova;
597 va_start = ms->addr_64;
598 } else if (!skip_last && last_seg) {
599 /* this is the last segment and we're not skipping */
600 if (type == RTE_MEM_EVENT_ALLOC)
601 vfio_dma_mem_map(default_vfio_cfg, va_start,
603 iova_expected - iova_start, 1);
605 vfio_dma_mem_map(default_vfio_cfg, va_start,
607 iova_expected - iova_start, 0);
610 #ifdef RTE_ARCH_PPC_64
612 ms = rte_mem_virt2memseg(addr, msl);
613 while (cur_len < len) {
614 int idx = rte_fbarray_find_idx(&msl->memseg_arr, ms);
616 rte_fbarray_set_used(&msl->memseg_arr, idx);
624 vfio_sync_default_container(void)
626 struct rte_mp_msg mp_req, *mp_rep;
627 struct rte_mp_reply mp_reply = {0};
628 struct timespec ts = {.tv_sec = 5, .tv_nsec = 0};
629 struct vfio_mp_param *p = (struct vfio_mp_param *)mp_req.param;
633 /* cannot be called from primary */
634 if (rte_eal_process_type() != RTE_PROC_SECONDARY)
637 /* default container fd should have been opened in rte_vfio_enable() */
638 if (!default_vfio_cfg->vfio_enabled ||
639 default_vfio_cfg->vfio_container_fd < 0) {
640 RTE_LOG(ERR, EAL, "VFIO support is not initialized\n");
644 /* find default container's IOMMU type */
645 p->req = SOCKET_REQ_IOMMU_TYPE;
646 strcpy(mp_req.name, EAL_VFIO_MP);
647 mp_req.len_param = sizeof(*p);
651 if (rte_mp_request_sync(&mp_req, &mp_reply, &ts) == 0 &&
652 mp_reply.nb_received == 1) {
653 mp_rep = &mp_reply.msgs[0];
654 p = (struct vfio_mp_param *)mp_rep->param;
655 if (p->result == SOCKET_OK)
656 iommu_type_id = p->iommu_type_id;
659 if (iommu_type_id < 0) {
660 RTE_LOG(ERR, EAL, "Could not get IOMMU type for default container\n");
664 /* we now have an fd for default container, as well as its IOMMU type.
665 * now, set up default VFIO container config to match.
667 for (i = 0; i < RTE_DIM(iommu_types); i++) {
668 const struct vfio_iommu_type *t = &iommu_types[i];
669 if (t->type_id != iommu_type_id)
672 /* we found our IOMMU type */
673 default_vfio_cfg->vfio_iommu_type = t;
677 RTE_LOG(ERR, EAL, "Could not find IOMMU type id (%i)\n",
683 rte_vfio_clear_group(int vfio_group_fd)
686 struct vfio_config *vfio_cfg;
688 vfio_cfg = get_vfio_cfg_by_group_fd(vfio_group_fd);
689 if (vfio_cfg == NULL) {
690 RTE_LOG(ERR, EAL, " invalid group fd!\n");
694 i = get_vfio_group_idx(vfio_group_fd);
697 vfio_cfg->vfio_groups[i].group_num = -1;
698 vfio_cfg->vfio_groups[i].fd = -1;
699 vfio_cfg->vfio_groups[i].devices = 0;
700 vfio_cfg->vfio_active_groups--;
706 rte_vfio_setup_device(const char *sysfs_base, const char *dev_addr,
707 int *vfio_dev_fd, struct vfio_device_info *device_info)
709 struct vfio_group_status group_status = {
710 .argsz = sizeof(group_status)
712 struct vfio_config *vfio_cfg;
713 struct user_mem_maps *user_mem_maps;
714 int vfio_container_fd;
719 const struct internal_config *internal_conf =
720 eal_get_internal_configuration();
722 /* get group number */
723 ret = rte_vfio_get_group_num(sysfs_base, dev_addr, &iommu_group_num);
725 RTE_LOG(WARNING, EAL, " %s not managed by VFIO driver, skipping\n",
730 /* if negative, something failed */
734 /* get the actual group fd */
735 vfio_group_fd = rte_vfio_get_group_fd(iommu_group_num);
736 if (vfio_group_fd < 0)
739 /* if group_fd == 0, that means the device isn't managed by VFIO */
740 if (vfio_group_fd == 0) {
741 RTE_LOG(WARNING, EAL, " %s not managed by VFIO driver, skipping\n",
747 * at this point, we know that this group is viable (meaning, all devices
748 * are either bound to VFIO or not bound to anything)
751 /* check if the group is viable */
752 ret = ioctl(vfio_group_fd, VFIO_GROUP_GET_STATUS, &group_status);
754 RTE_LOG(ERR, EAL, " %s cannot get group status, "
755 "error %i (%s)\n", dev_addr, errno, strerror(errno));
756 close(vfio_group_fd);
757 rte_vfio_clear_group(vfio_group_fd);
759 } else if (!(group_status.flags & VFIO_GROUP_FLAGS_VIABLE)) {
760 RTE_LOG(ERR, EAL, " %s VFIO group is not viable! "
761 "Not all devices in IOMMU group bound to VFIO or unbound\n",
763 close(vfio_group_fd);
764 rte_vfio_clear_group(vfio_group_fd);
768 /* get the vfio_config it belongs to */
769 vfio_cfg = get_vfio_cfg_by_group_num(iommu_group_num);
770 vfio_cfg = vfio_cfg ? vfio_cfg : default_vfio_cfg;
771 vfio_container_fd = vfio_cfg->vfio_container_fd;
772 user_mem_maps = &vfio_cfg->mem_maps;
774 /* check if group does not have a container yet */
775 if (!(group_status.flags & VFIO_GROUP_FLAGS_CONTAINER_SET)) {
777 /* add group to a container */
778 ret = ioctl(vfio_group_fd, VFIO_GROUP_SET_CONTAINER,
781 RTE_LOG(ERR, EAL, " %s cannot add VFIO group to container, "
782 "error %i (%s)\n", dev_addr, errno, strerror(errno));
783 close(vfio_group_fd);
784 rte_vfio_clear_group(vfio_group_fd);
789 * pick an IOMMU type and set up DMA mappings for container
791 * needs to be done only once, only when first group is
792 * assigned to a container and only in primary process.
793 * Note this can happen several times with the hotplug
796 if (internal_conf->process_type == RTE_PROC_PRIMARY &&
797 vfio_cfg->vfio_active_groups == 1 &&
798 vfio_group_device_count(vfio_group_fd) == 0) {
799 const struct vfio_iommu_type *t;
801 /* select an IOMMU type which we will be using */
802 t = vfio_set_iommu_type(vfio_container_fd);
805 " %s failed to select IOMMU type\n",
807 close(vfio_group_fd);
808 rte_vfio_clear_group(vfio_group_fd);
811 /* lock memory hotplug before mapping and release it
812 * after registering callback, to prevent races
814 rte_mcfg_mem_read_lock();
815 if (vfio_cfg == default_vfio_cfg)
816 ret = t->dma_map_func(vfio_container_fd);
821 " %s DMA remapping failed, error %i (%s)\n",
822 dev_addr, errno, strerror(errno));
823 close(vfio_group_fd);
824 rte_vfio_clear_group(vfio_group_fd);
825 rte_mcfg_mem_read_unlock();
829 vfio_cfg->vfio_iommu_type = t;
831 /* re-map all user-mapped segments */
832 rte_spinlock_recursive_lock(&user_mem_maps->lock);
834 /* this IOMMU type may not support DMA mapping, but
835 * if we have mappings in the list - that means we have
836 * previously mapped something successfully, so we can
837 * be sure that DMA mapping is supported.
839 for (i = 0; i < user_mem_maps->n_maps; i++) {
840 struct user_mem_map *map;
841 map = &user_mem_maps->maps[i];
843 ret = t->dma_user_map_func(
845 map->addr, map->iova, map->len,
848 RTE_LOG(ERR, EAL, "Couldn't map user memory for DMA: "
850 "iova: 0x%" PRIx64 " "
851 "len: 0x%" PRIu64 "\n",
852 map->addr, map->iova,
854 rte_spinlock_recursive_unlock(
855 &user_mem_maps->lock);
856 rte_mcfg_mem_read_unlock();
860 rte_spinlock_recursive_unlock(&user_mem_maps->lock);
862 /* register callback for mem events */
863 if (vfio_cfg == default_vfio_cfg)
864 ret = rte_mem_event_callback_register(
865 VFIO_MEM_EVENT_CLB_NAME,
866 vfio_mem_event_callback, NULL);
869 /* unlock memory hotplug */
870 rte_mcfg_mem_read_unlock();
872 if (ret && rte_errno != ENOTSUP) {
873 RTE_LOG(ERR, EAL, "Could not install memory event callback for VFIO\n");
877 RTE_LOG(DEBUG, EAL, "Memory event callbacks not supported\n");
879 RTE_LOG(DEBUG, EAL, "Installed memory event callback for VFIO\n");
881 } else if (rte_eal_process_type() != RTE_PROC_PRIMARY &&
882 vfio_cfg == default_vfio_cfg &&
883 vfio_cfg->vfio_iommu_type == NULL) {
884 /* if we're not a primary process, we do not set up the VFIO
885 * container because it's already been set up by the primary
886 * process. instead, we simply ask the primary about VFIO type
887 * we are using, and set the VFIO config up appropriately.
889 ret = vfio_sync_default_container();
891 RTE_LOG(ERR, EAL, "Could not sync default VFIO container\n");
892 close(vfio_group_fd);
893 rte_vfio_clear_group(vfio_group_fd);
896 /* we have successfully initialized VFIO, notify user */
897 const struct vfio_iommu_type *t =
898 default_vfio_cfg->vfio_iommu_type;
899 RTE_LOG(INFO, EAL, " using IOMMU type %d (%s)\n",
900 t->type_id, t->name);
903 rte_eal_vfio_get_vf_token(vf_token);
905 /* get a file descriptor for the device with VF token firstly */
906 if (!rte_uuid_is_null(vf_token)) {
907 char vf_token_str[RTE_UUID_STRLEN];
910 rte_uuid_unparse(vf_token, vf_token_str, sizeof(vf_token_str));
911 snprintf(dev, sizeof(dev),
912 "%s vf_token=%s", dev_addr, vf_token_str);
914 *vfio_dev_fd = ioctl(vfio_group_fd, VFIO_GROUP_GET_DEVICE_FD,
916 if (*vfio_dev_fd >= 0)
920 /* get a file descriptor for the device */
921 *vfio_dev_fd = ioctl(vfio_group_fd, VFIO_GROUP_GET_DEVICE_FD, dev_addr);
922 if (*vfio_dev_fd < 0) {
923 /* if we cannot get a device fd, this implies a problem with
924 * the VFIO group or the container not having IOMMU configured.
927 RTE_LOG(WARNING, EAL, "Getting a vfio_dev_fd for %s failed\n",
929 close(vfio_group_fd);
930 rte_vfio_clear_group(vfio_group_fd);
934 /* test and setup the device */
936 ret = ioctl(*vfio_dev_fd, VFIO_DEVICE_GET_INFO, device_info);
938 RTE_LOG(ERR, EAL, " %s cannot get device info, "
939 "error %i (%s)\n", dev_addr, errno,
942 close(vfio_group_fd);
943 rte_vfio_clear_group(vfio_group_fd);
946 vfio_group_device_get(vfio_group_fd);
952 rte_vfio_release_device(const char *sysfs_base, const char *dev_addr,
955 struct vfio_config *vfio_cfg;
960 /* we don't want any DMA mapping messages to come while we're detaching
961 * VFIO device, because this might be the last device and we might need
962 * to unregister the callback.
964 rte_mcfg_mem_read_lock();
966 /* get group number */
967 ret = rte_vfio_get_group_num(sysfs_base, dev_addr, &iommu_group_num);
969 RTE_LOG(WARNING, EAL, " %s not managed by VFIO driver\n",
971 /* This is an error at this point. */
976 /* get the actual group fd */
977 vfio_group_fd = rte_vfio_get_group_fd(iommu_group_num);
978 if (vfio_group_fd <= 0) {
979 RTE_LOG(INFO, EAL, "rte_vfio_get_group_fd failed for %s\n",
985 /* get the vfio_config it belongs to */
986 vfio_cfg = get_vfio_cfg_by_group_num(iommu_group_num);
987 vfio_cfg = vfio_cfg ? vfio_cfg : default_vfio_cfg;
989 /* At this point we got an active group. Closing it will make the
990 * container detachment. If this is the last active group, VFIO kernel
991 * code will unset the container and the IOMMU mappings.
994 /* Closing a device */
995 if (close(vfio_dev_fd) < 0) {
996 RTE_LOG(INFO, EAL, "Error when closing vfio_dev_fd for %s\n",
1002 /* An VFIO group can have several devices attached. Just when there is
1003 * no devices remaining should the group be closed.
1005 vfio_group_device_put(vfio_group_fd);
1006 if (!vfio_group_device_count(vfio_group_fd)) {
1008 if (close(vfio_group_fd) < 0) {
1009 RTE_LOG(INFO, EAL, "Error when closing vfio_group_fd for %s\n",
1015 if (rte_vfio_clear_group(vfio_group_fd) < 0) {
1016 RTE_LOG(INFO, EAL, "Error when clearing group for %s\n",
1023 /* if there are no active device groups, unregister the callback to
1024 * avoid spurious attempts to map/unmap memory from VFIO.
1026 if (vfio_cfg == default_vfio_cfg && vfio_cfg->vfio_active_groups == 0 &&
1027 rte_eal_process_type() != RTE_PROC_SECONDARY)
1028 rte_mem_event_callback_unregister(VFIO_MEM_EVENT_CLB_NAME,
1035 rte_mcfg_mem_read_unlock();
1040 rte_vfio_enable(const char *modname)
1042 /* initialize group list */
1045 const struct internal_config *internal_conf =
1046 eal_get_internal_configuration();
1048 rte_spinlock_recursive_t lock = RTE_SPINLOCK_RECURSIVE_INITIALIZER;
1050 for (i = 0; i < VFIO_MAX_CONTAINERS; i++) {
1051 vfio_cfgs[i].vfio_container_fd = -1;
1052 vfio_cfgs[i].vfio_active_groups = 0;
1053 vfio_cfgs[i].vfio_iommu_type = NULL;
1054 vfio_cfgs[i].mem_maps.lock = lock;
1056 for (j = 0; j < VFIO_MAX_GROUPS; j++) {
1057 vfio_cfgs[i].vfio_groups[j].fd = -1;
1058 vfio_cfgs[i].vfio_groups[j].group_num = -1;
1059 vfio_cfgs[i].vfio_groups[j].devices = 0;
1063 /* inform the user that we are probing for VFIO */
1064 RTE_LOG(INFO, EAL, "Probing VFIO support...\n");
1066 /* check if vfio module is loaded */
1067 vfio_available = rte_eal_check_module(modname);
1069 /* return error directly */
1070 if (vfio_available == -1) {
1071 RTE_LOG(INFO, EAL, "Could not get loaded module details!\n");
1075 /* return 0 if VFIO modules not loaded */
1076 if (vfio_available == 0) {
1077 RTE_LOG(DEBUG, EAL, "VFIO modules not loaded, "
1078 "skipping VFIO support...\n");
1082 if (internal_conf->process_type == RTE_PROC_PRIMARY) {
1083 /* open a new container */
1084 default_vfio_cfg->vfio_container_fd =
1085 rte_vfio_get_container_fd();
1087 /* get the default container from the primary process */
1088 default_vfio_cfg->vfio_container_fd =
1089 vfio_get_default_container_fd();
1092 /* check if we have VFIO driver enabled */
1093 if (default_vfio_cfg->vfio_container_fd != -1) {
1094 RTE_LOG(INFO, EAL, "VFIO support initialized\n");
1095 default_vfio_cfg->vfio_enabled = 1;
1097 RTE_LOG(NOTICE, EAL, "VFIO support could not be initialized\n");
1104 rte_vfio_is_enabled(const char *modname)
1106 const int mod_available = rte_eal_check_module(modname) > 0;
1107 return default_vfio_cfg->vfio_enabled && mod_available;
1111 vfio_get_default_container_fd(void)
1113 struct rte_mp_msg mp_req, *mp_rep;
1114 struct rte_mp_reply mp_reply = {0};
1115 struct timespec ts = {.tv_sec = 5, .tv_nsec = 0};
1116 struct vfio_mp_param *p = (struct vfio_mp_param *)mp_req.param;
1118 const struct internal_config *internal_conf =
1119 eal_get_internal_configuration();
1121 if (default_vfio_cfg->vfio_enabled)
1122 return default_vfio_cfg->vfio_container_fd;
1124 if (internal_conf->process_type == RTE_PROC_PRIMARY) {
1125 /* if we were secondary process we would try requesting
1126 * container fd from the primary, but we're the primary
1127 * process so just exit here
1132 p->req = SOCKET_REQ_DEFAULT_CONTAINER;
1133 strcpy(mp_req.name, EAL_VFIO_MP);
1134 mp_req.len_param = sizeof(*p);
1137 if (rte_mp_request_sync(&mp_req, &mp_reply, &ts) == 0 &&
1138 mp_reply.nb_received == 1) {
1139 mp_rep = &mp_reply.msgs[0];
1140 p = (struct vfio_mp_param *)mp_rep->param;
1141 if (p->result == SOCKET_OK && mp_rep->num_fds == 1) {
1142 container_fd = mp_rep->fds[0];
1143 free(mp_reply.msgs);
1144 return container_fd;
1148 free(mp_reply.msgs);
1149 RTE_LOG(ERR, EAL, " cannot request default container fd\n");
1154 vfio_get_iommu_type(void)
1156 if (default_vfio_cfg->vfio_iommu_type == NULL)
1159 return default_vfio_cfg->vfio_iommu_type->type_id;
1162 const struct vfio_iommu_type *
1163 vfio_set_iommu_type(int vfio_container_fd)
1166 for (idx = 0; idx < RTE_DIM(iommu_types); idx++) {
1167 const struct vfio_iommu_type *t = &iommu_types[idx];
1169 int ret = ioctl(vfio_container_fd, VFIO_SET_IOMMU,
1172 RTE_LOG(INFO, EAL, " using IOMMU type %d (%s)\n",
1173 t->type_id, t->name);
1176 /* not an error, there may be more supported IOMMU types */
1177 RTE_LOG(DEBUG, EAL, " set IOMMU type %d (%s) failed, "
1178 "error %i (%s)\n", t->type_id, t->name, errno,
1181 /* if we didn't find a suitable IOMMU type, fail */
1186 vfio_has_supported_extensions(int vfio_container_fd)
1189 unsigned idx, n_extensions = 0;
1190 for (idx = 0; idx < RTE_DIM(iommu_types); idx++) {
1191 const struct vfio_iommu_type *t = &iommu_types[idx];
1193 ret = ioctl(vfio_container_fd, VFIO_CHECK_EXTENSION,
1196 RTE_LOG(ERR, EAL, " could not get IOMMU type, "
1197 "error %i (%s)\n", errno,
1199 close(vfio_container_fd);
1201 } else if (ret == 1) {
1202 /* we found a supported extension */
1205 RTE_LOG(DEBUG, EAL, " IOMMU type %d (%s) is %s\n",
1206 t->type_id, t->name,
1207 ret ? "supported" : "not supported");
1210 /* if we didn't find any supported IOMMU types, fail */
1211 if (!n_extensions) {
1212 close(vfio_container_fd);
1220 rte_vfio_get_container_fd(void)
1222 int ret, vfio_container_fd;
1223 struct rte_mp_msg mp_req, *mp_rep;
1224 struct rte_mp_reply mp_reply = {0};
1225 struct timespec ts = {.tv_sec = 5, .tv_nsec = 0};
1226 struct vfio_mp_param *p = (struct vfio_mp_param *)mp_req.param;
1227 const struct internal_config *internal_conf =
1228 eal_get_internal_configuration();
1231 /* if we're in a primary process, try to open the container */
1232 if (internal_conf->process_type == RTE_PROC_PRIMARY) {
1233 vfio_container_fd = open(VFIO_CONTAINER_PATH, O_RDWR);
1234 if (vfio_container_fd < 0) {
1235 RTE_LOG(ERR, EAL, " cannot open VFIO container, "
1236 "error %i (%s)\n", errno, strerror(errno));
1240 /* check VFIO API version */
1241 ret = ioctl(vfio_container_fd, VFIO_GET_API_VERSION);
1242 if (ret != VFIO_API_VERSION) {
1244 RTE_LOG(ERR, EAL, " could not get VFIO API version, "
1245 "error %i (%s)\n", errno, strerror(errno));
1247 RTE_LOG(ERR, EAL, " unsupported VFIO API version!\n");
1248 close(vfio_container_fd);
1252 ret = vfio_has_supported_extensions(vfio_container_fd);
1254 RTE_LOG(ERR, EAL, " no supported IOMMU "
1255 "extensions found!\n");
1259 return vfio_container_fd;
1262 * if we're in a secondary process, request container fd from the
1263 * primary process via mp channel
1265 p->req = SOCKET_REQ_CONTAINER;
1266 strcpy(mp_req.name, EAL_VFIO_MP);
1267 mp_req.len_param = sizeof(*p);
1270 vfio_container_fd = -1;
1271 if (rte_mp_request_sync(&mp_req, &mp_reply, &ts) == 0 &&
1272 mp_reply.nb_received == 1) {
1273 mp_rep = &mp_reply.msgs[0];
1274 p = (struct vfio_mp_param *)mp_rep->param;
1275 if (p->result == SOCKET_OK && mp_rep->num_fds == 1) {
1276 vfio_container_fd = mp_rep->fds[0];
1277 free(mp_reply.msgs);
1278 return vfio_container_fd;
1282 free(mp_reply.msgs);
1283 RTE_LOG(ERR, EAL, " cannot request container fd\n");
1288 rte_vfio_get_group_num(const char *sysfs_base,
1289 const char *dev_addr, int *iommu_group_num)
1291 char linkname[PATH_MAX];
1292 char filename[PATH_MAX];
1293 char *tok[16], *group_tok, *end;
1296 memset(linkname, 0, sizeof(linkname));
1297 memset(filename, 0, sizeof(filename));
1299 /* try to find out IOMMU group for this device */
1300 snprintf(linkname, sizeof(linkname),
1301 "%s/%s/iommu_group", sysfs_base, dev_addr);
1303 ret = readlink(linkname, filename, sizeof(filename));
1305 /* if the link doesn't exist, no VFIO for us */
1309 ret = rte_strsplit(filename, sizeof(filename),
1310 tok, RTE_DIM(tok), '/');
1313 RTE_LOG(ERR, EAL, " %s cannot get IOMMU group\n", dev_addr);
1317 /* IOMMU group is always the last token */
1319 group_tok = tok[ret - 1];
1321 *iommu_group_num = strtol(group_tok, &end, 10);
1322 if ((end != group_tok && *end != '\0') || errno != 0) {
1323 RTE_LOG(ERR, EAL, " %s error parsing IOMMU number!\n", dev_addr);
1331 type1_map_contig(const struct rte_memseg_list *msl, const struct rte_memseg *ms,
1332 size_t len, void *arg)
1334 int *vfio_container_fd = arg;
1339 return vfio_type1_dma_mem_map(*vfio_container_fd, ms->addr_64, ms->iova,
1344 type1_map(const struct rte_memseg_list *msl, const struct rte_memseg *ms,
1347 int *vfio_container_fd = arg;
1349 /* skip external memory that isn't a heap */
1350 if (msl->external && !msl->heap)
1353 /* skip any segments with invalid IOVA addresses */
1354 if (ms->iova == RTE_BAD_IOVA)
1357 /* if IOVA mode is VA, we've already mapped the internal segments */
1358 if (!msl->external && rte_eal_iova_mode() == RTE_IOVA_VA)
1361 return vfio_type1_dma_mem_map(*vfio_container_fd, ms->addr_64, ms->iova,
1366 vfio_type1_dma_mem_map(int vfio_container_fd, uint64_t vaddr, uint64_t iova,
1367 uint64_t len, int do_map)
1369 struct vfio_iommu_type1_dma_map dma_map;
1370 struct vfio_iommu_type1_dma_unmap dma_unmap;
1374 memset(&dma_map, 0, sizeof(dma_map));
1375 dma_map.argsz = sizeof(struct vfio_iommu_type1_dma_map);
1376 dma_map.vaddr = vaddr;
1378 dma_map.iova = iova;
1379 dma_map.flags = VFIO_DMA_MAP_FLAG_READ |
1380 VFIO_DMA_MAP_FLAG_WRITE;
1382 ret = ioctl(vfio_container_fd, VFIO_IOMMU_MAP_DMA, &dma_map);
1385 * In case the mapping was already done EEXIST will be
1386 * returned from kernel.
1388 if (errno == EEXIST) {
1390 " Memory segment is already mapped,"
1394 " cannot set up DMA remapping,"
1396 errno, strerror(errno));
1401 memset(&dma_unmap, 0, sizeof(dma_unmap));
1402 dma_unmap.argsz = sizeof(struct vfio_iommu_type1_dma_unmap);
1403 dma_unmap.size = len;
1404 dma_unmap.iova = iova;
1406 ret = ioctl(vfio_container_fd, VFIO_IOMMU_UNMAP_DMA,
1409 RTE_LOG(ERR, EAL, " cannot clear DMA remapping, error %i (%s)\n",
1410 errno, strerror(errno));
1419 vfio_type1_dma_map(int vfio_container_fd)
1421 if (rte_eal_iova_mode() == RTE_IOVA_VA) {
1422 /* with IOVA as VA mode, we can get away with mapping contiguous
1423 * chunks rather than going page-by-page.
1425 int ret = rte_memseg_contig_walk(type1_map_contig,
1426 &vfio_container_fd);
1429 /* we have to continue the walk because we've skipped the
1430 * external segments during the config walk.
1433 return rte_memseg_walk(type1_map, &vfio_container_fd);
1437 vfio_spapr_dma_do_map(int vfio_container_fd, uint64_t vaddr, uint64_t iova,
1438 uint64_t len, int do_map)
1440 struct vfio_iommu_type1_dma_map dma_map;
1441 struct vfio_iommu_type1_dma_unmap dma_unmap;
1443 struct vfio_iommu_spapr_register_memory reg = {
1444 .argsz = sizeof(reg),
1447 reg.vaddr = (uintptr_t) vaddr;
1451 ret = ioctl(vfio_container_fd,
1452 VFIO_IOMMU_SPAPR_REGISTER_MEMORY, ®);
1454 RTE_LOG(ERR, EAL, " cannot register vaddr for IOMMU, "
1455 "error %i (%s)\n", errno, strerror(errno));
1459 memset(&dma_map, 0, sizeof(dma_map));
1460 dma_map.argsz = sizeof(struct vfio_iommu_type1_dma_map);
1461 dma_map.vaddr = vaddr;
1463 dma_map.iova = iova;
1464 dma_map.flags = VFIO_DMA_MAP_FLAG_READ |
1465 VFIO_DMA_MAP_FLAG_WRITE;
1467 ret = ioctl(vfio_container_fd, VFIO_IOMMU_MAP_DMA, &dma_map);
1470 * In case the mapping was already done EBUSY will be
1471 * returned from kernel.
1473 if (errno == EBUSY) {
1475 " Memory segment is already mapped,"
1479 " cannot set up DMA remapping,"
1480 " error %i (%s)\n", errno,
1487 memset(&dma_unmap, 0, sizeof(dma_unmap));
1488 dma_unmap.argsz = sizeof(struct vfio_iommu_type1_dma_unmap);
1489 dma_unmap.size = len;
1490 dma_unmap.iova = iova;
1492 ret = ioctl(vfio_container_fd, VFIO_IOMMU_UNMAP_DMA,
1495 RTE_LOG(ERR, EAL, " cannot clear DMA remapping, error %i (%s)\n",
1496 errno, strerror(errno));
1500 ret = ioctl(vfio_container_fd,
1501 VFIO_IOMMU_SPAPR_UNREGISTER_MEMORY, ®);
1503 RTE_LOG(ERR, EAL, " cannot unregister vaddr for IOMMU, error %i (%s)\n",
1504 errno, strerror(errno));
1513 vfio_spapr_map_walk(const struct rte_memseg_list *msl,
1514 const struct rte_memseg *ms, void *arg)
1516 int *vfio_container_fd = arg;
1518 /* skip external memory that isn't a heap */
1519 if (msl->external && !msl->heap)
1522 /* skip any segments with invalid IOVA addresses */
1523 if (ms->iova == RTE_BAD_IOVA)
1526 return vfio_spapr_dma_do_map(*vfio_container_fd, ms->addr_64, ms->iova,
1531 vfio_spapr_unmap_walk(const struct rte_memseg_list *msl,
1532 const struct rte_memseg *ms, void *arg)
1534 int *vfio_container_fd = arg;
1536 /* skip external memory that isn't a heap */
1537 if (msl->external && !msl->heap)
1540 /* skip any segments with invalid IOVA addresses */
1541 if (ms->iova == RTE_BAD_IOVA)
1544 return vfio_spapr_dma_do_map(*vfio_container_fd, ms->addr_64, ms->iova,
1548 struct spapr_walk_param {
1549 uint64_t window_size;
1550 uint64_t hugepage_sz;
1554 vfio_spapr_window_size_walk(const struct rte_memseg_list *msl,
1555 const struct rte_memseg *ms, void *arg)
1557 struct spapr_walk_param *param = arg;
1558 uint64_t max = ms->iova + ms->len;
1560 /* skip external memory that isn't a heap */
1561 if (msl->external && !msl->heap)
1564 /* skip any segments with invalid IOVA addresses */
1565 if (ms->iova == RTE_BAD_IOVA)
1568 if (max > param->window_size) {
1569 param->hugepage_sz = ms->hugepage_sz;
1570 param->window_size = max;
1577 vfio_spapr_create_new_dma_window(int vfio_container_fd,
1578 struct vfio_iommu_spapr_tce_create *create) {
1579 struct vfio_iommu_spapr_tce_remove remove = {
1580 .argsz = sizeof(remove),
1582 struct vfio_iommu_spapr_tce_info info = {
1583 .argsz = sizeof(info),
1587 /* query spapr iommu info */
1588 ret = ioctl(vfio_container_fd, VFIO_IOMMU_SPAPR_TCE_GET_INFO, &info);
1590 RTE_LOG(ERR, EAL, " cannot get iommu info, "
1591 "error %i (%s)\n", errno, strerror(errno));
1595 /* remove default DMA of 32 bit window */
1596 remove.start_addr = info.dma32_window_start;
1597 ret = ioctl(vfio_container_fd, VFIO_IOMMU_SPAPR_TCE_REMOVE, &remove);
1599 RTE_LOG(ERR, EAL, " cannot remove default DMA window, "
1600 "error %i (%s)\n", errno, strerror(errno));
1604 /* create new DMA window */
1605 ret = ioctl(vfio_container_fd, VFIO_IOMMU_SPAPR_TCE_CREATE, create);
1607 #ifdef VFIO_IOMMU_SPAPR_INFO_DDW
1608 /* try possible page_shift and levels for workaround */
1611 for (levels = create->levels + 1;
1612 ret && levels <= info.ddw.levels; levels++) {
1613 create->levels = levels;
1614 ret = ioctl(vfio_container_fd,
1615 VFIO_IOMMU_SPAPR_TCE_CREATE, create);
1619 RTE_LOG(ERR, EAL, " cannot create new DMA window, "
1620 "error %i (%s)\n", errno, strerror(errno));
1625 if (create->start_addr != 0) {
1626 RTE_LOG(ERR, EAL, " DMA window start address != 0\n");
1634 vfio_spapr_dma_mem_map(int vfio_container_fd, uint64_t vaddr, uint64_t iova,
1635 uint64_t len, int do_map)
1637 struct spapr_walk_param param;
1638 struct vfio_iommu_spapr_tce_create create = {
1639 .argsz = sizeof(create),
1641 struct vfio_config *vfio_cfg;
1642 struct user_mem_maps *user_mem_maps;
1645 vfio_cfg = get_vfio_cfg_by_container_fd(vfio_container_fd);
1646 if (vfio_cfg == NULL) {
1647 RTE_LOG(ERR, EAL, " invalid container fd!\n");
1651 user_mem_maps = &vfio_cfg->mem_maps;
1652 rte_spinlock_recursive_lock(&user_mem_maps->lock);
1654 /* check if window size needs to be adjusted */
1655 memset(¶m, 0, sizeof(param));
1657 /* we're inside a callback so use thread-unsafe version */
1658 if (rte_memseg_walk_thread_unsafe(vfio_spapr_window_size_walk,
1660 RTE_LOG(ERR, EAL, "Could not get window size\n");
1665 /* also check user maps */
1666 for (i = 0; i < user_mem_maps->n_maps; i++) {
1667 uint64_t max = user_mem_maps->maps[i].iova +
1668 user_mem_maps->maps[i].len;
1669 param.window_size = RTE_MAX(param.window_size, max);
1672 /* sPAPR requires window size to be a power of 2 */
1673 create.window_size = rte_align64pow2(param.window_size);
1674 create.page_shift = __builtin_ctzll(param.hugepage_sz);
1678 /* re-create window and remap the entire memory */
1679 if (iova + len > create.window_size) {
1680 /* release all maps before recreating the window */
1681 if (rte_memseg_walk_thread_unsafe(vfio_spapr_unmap_walk,
1682 &vfio_container_fd) < 0) {
1683 RTE_LOG(ERR, EAL, "Could not release DMA maps\n");
1687 /* release all user maps */
1688 for (i = 0; i < user_mem_maps->n_maps; i++) {
1689 struct user_mem_map *map =
1690 &user_mem_maps->maps[i];
1691 if (vfio_spapr_dma_do_map(vfio_container_fd,
1692 map->addr, map->iova, map->len,
1694 RTE_LOG(ERR, EAL, "Could not release user DMA maps\n");
1699 create.window_size = rte_align64pow2(iova + len);
1700 if (vfio_spapr_create_new_dma_window(vfio_container_fd,
1702 RTE_LOG(ERR, EAL, "Could not create new DMA window\n");
1706 /* we're inside a callback, so use thread-unsafe version
1708 if (rte_memseg_walk_thread_unsafe(vfio_spapr_map_walk,
1709 &vfio_container_fd) < 0) {
1710 RTE_LOG(ERR, EAL, "Could not recreate DMA maps\n");
1714 /* remap all user maps */
1715 for (i = 0; i < user_mem_maps->n_maps; i++) {
1716 struct user_mem_map *map =
1717 &user_mem_maps->maps[i];
1718 if (vfio_spapr_dma_do_map(vfio_container_fd,
1719 map->addr, map->iova, map->len,
1721 RTE_LOG(ERR, EAL, "Could not recreate user DMA maps\n");
1727 if (vfio_spapr_dma_do_map(vfio_container_fd, vaddr, iova, len, 1)) {
1728 RTE_LOG(ERR, EAL, "Failed to map DMA\n");
1733 /* for unmap, check if iova within DMA window */
1734 if (iova > create.window_size) {
1735 RTE_LOG(ERR, EAL, "iova beyond DMA window for unmap");
1740 vfio_spapr_dma_do_map(vfio_container_fd, vaddr, iova, len, 0);
1743 rte_spinlock_recursive_unlock(&user_mem_maps->lock);
1748 vfio_spapr_dma_map(int vfio_container_fd)
1750 struct vfio_iommu_spapr_tce_create create = {
1751 .argsz = sizeof(create),
1753 struct spapr_walk_param param;
1755 memset(¶m, 0, sizeof(param));
1757 /* create DMA window from 0 to max(phys_addr + len) */
1758 rte_memseg_walk(vfio_spapr_window_size_walk, ¶m);
1760 /* sPAPR requires window size to be a power of 2 */
1761 create.window_size = rte_align64pow2(param.window_size);
1762 create.page_shift = __builtin_ctzll(param.hugepage_sz);
1765 if (vfio_spapr_create_new_dma_window(vfio_container_fd, &create) < 0) {
1766 RTE_LOG(ERR, EAL, "Could not create new DMA window\n");
1770 /* map all DPDK segments for DMA. use 1:1 PA to IOVA mapping */
1771 if (rte_memseg_walk(vfio_spapr_map_walk, &vfio_container_fd) < 0)
1778 vfio_noiommu_dma_map(int __rte_unused vfio_container_fd)
1780 /* No-IOMMU mode does not need DMA mapping */
1785 vfio_noiommu_dma_mem_map(int __rte_unused vfio_container_fd,
1786 uint64_t __rte_unused vaddr,
1787 uint64_t __rte_unused iova, uint64_t __rte_unused len,
1788 int __rte_unused do_map)
1790 /* No-IOMMU mode does not need DMA mapping */
1795 vfio_dma_mem_map(struct vfio_config *vfio_cfg, uint64_t vaddr, uint64_t iova,
1796 uint64_t len, int do_map)
1798 const struct vfio_iommu_type *t = vfio_cfg->vfio_iommu_type;
1801 RTE_LOG(ERR, EAL, " VFIO support not initialized\n");
1806 if (!t->dma_user_map_func) {
1808 " VFIO custom DMA region maping not supported by IOMMU %s\n",
1810 rte_errno = ENOTSUP;
1814 return t->dma_user_map_func(vfio_cfg->vfio_container_fd, vaddr, iova,
1819 container_dma_map(struct vfio_config *vfio_cfg, uint64_t vaddr, uint64_t iova,
1822 struct user_mem_map *new_map;
1823 struct user_mem_maps *user_mem_maps;
1826 user_mem_maps = &vfio_cfg->mem_maps;
1827 rte_spinlock_recursive_lock(&user_mem_maps->lock);
1828 if (user_mem_maps->n_maps == VFIO_MAX_USER_MEM_MAPS) {
1829 RTE_LOG(ERR, EAL, "No more space for user mem maps\n");
1835 if (vfio_dma_mem_map(vfio_cfg, vaddr, iova, len, 1)) {
1836 /* technically, this will fail if there are currently no devices
1837 * plugged in, even if a device were added later, this mapping
1838 * might have succeeded. however, since we cannot verify if this
1839 * is a valid mapping without having a device attached, consider
1840 * this to be unsupported, because we can't just store any old
1841 * mapping and pollute list of active mappings willy-nilly.
1843 RTE_LOG(ERR, EAL, "Couldn't map new region for DMA\n");
1847 /* create new user mem map entry */
1848 new_map = &user_mem_maps->maps[user_mem_maps->n_maps++];
1849 new_map->addr = vaddr;
1850 new_map->iova = iova;
1853 compact_user_maps(user_mem_maps);
1855 rte_spinlock_recursive_unlock(&user_mem_maps->lock);
1860 container_dma_unmap(struct vfio_config *vfio_cfg, uint64_t vaddr, uint64_t iova,
1863 struct user_mem_map *map, *new_map = NULL;
1864 struct user_mem_maps *user_mem_maps;
1867 user_mem_maps = &vfio_cfg->mem_maps;
1868 rte_spinlock_recursive_lock(&user_mem_maps->lock);
1870 /* find our mapping */
1871 map = find_user_mem_map(user_mem_maps, vaddr, iova, len);
1873 RTE_LOG(ERR, EAL, "Couldn't find previously mapped region\n");
1878 if (map->addr != vaddr || map->iova != iova || map->len != len) {
1879 /* we're partially unmapping a previously mapped region, so we
1880 * need to split entry into two.
1882 if (user_mem_maps->n_maps == VFIO_MAX_USER_MEM_MAPS) {
1883 RTE_LOG(ERR, EAL, "Not enough space to store partial mapping\n");
1888 new_map = &user_mem_maps->maps[user_mem_maps->n_maps++];
1891 /* unmap the entry */
1892 if (vfio_dma_mem_map(vfio_cfg, vaddr, iova, len, 0)) {
1893 /* there may not be any devices plugged in, so unmapping will
1894 * fail with ENODEV/ENOTSUP rte_errno values, but that doesn't
1895 * stop us from removing the mapping, as the assumption is we
1896 * won't be needing this memory any more and thus will want to
1897 * prevent it from being remapped again on hotplug. so, only
1898 * fail if we indeed failed to unmap (e.g. if the mapping was
1899 * within our mapped range but had invalid alignment).
1901 if (rte_errno != ENODEV && rte_errno != ENOTSUP) {
1902 RTE_LOG(ERR, EAL, "Couldn't unmap region for DMA\n");
1906 RTE_LOG(DEBUG, EAL, "DMA unmapping failed, but removing mappings anyway\n");
1909 /* remove map from the list of active mappings */
1910 if (new_map != NULL) {
1911 adjust_map(map, new_map, vaddr, len);
1913 /* if we've created a new map by splitting, sort everything */
1914 if (!is_null_map(new_map)) {
1915 compact_user_maps(user_mem_maps);
1917 /* we've created a new mapping, but it was unused */
1918 user_mem_maps->n_maps--;
1921 memset(map, 0, sizeof(*map));
1922 compact_user_maps(user_mem_maps);
1923 user_mem_maps->n_maps--;
1927 rte_spinlock_recursive_unlock(&user_mem_maps->lock);
1932 rte_vfio_noiommu_is_enabled(void)
1938 fd = open(VFIO_NOIOMMU_MODE, O_RDONLY);
1940 if (errno != ENOENT) {
1941 RTE_LOG(ERR, EAL, " cannot open vfio noiommu file %i (%s)\n",
1942 errno, strerror(errno));
1946 * else the file does not exists
1947 * i.e. noiommu is not enabled
1952 cnt = read(fd, &c, 1);
1955 RTE_LOG(ERR, EAL, " unable to read from vfio noiommu "
1956 "file %i (%s)\n", errno, strerror(errno));
1964 rte_vfio_container_create(void)
1968 /* Find an empty slot to store new vfio config */
1969 for (i = 1; i < VFIO_MAX_CONTAINERS; i++) {
1970 if (vfio_cfgs[i].vfio_container_fd == -1)
1974 if (i == VFIO_MAX_CONTAINERS) {
1975 RTE_LOG(ERR, EAL, "exceed max vfio container limit\n");
1979 vfio_cfgs[i].vfio_container_fd = rte_vfio_get_container_fd();
1980 if (vfio_cfgs[i].vfio_container_fd < 0) {
1981 RTE_LOG(NOTICE, EAL, "fail to create a new container\n");
1985 return vfio_cfgs[i].vfio_container_fd;
1989 rte_vfio_container_destroy(int container_fd)
1991 struct vfio_config *vfio_cfg;
1994 vfio_cfg = get_vfio_cfg_by_container_fd(container_fd);
1995 if (vfio_cfg == NULL) {
1996 RTE_LOG(ERR, EAL, "Invalid container fd\n");
2000 for (i = 0; i < VFIO_MAX_GROUPS; i++)
2001 if (vfio_cfg->vfio_groups[i].group_num != -1)
2002 rte_vfio_container_group_unbind(container_fd,
2003 vfio_cfg->vfio_groups[i].group_num);
2005 close(container_fd);
2006 vfio_cfg->vfio_container_fd = -1;
2007 vfio_cfg->vfio_active_groups = 0;
2008 vfio_cfg->vfio_iommu_type = NULL;
2014 rte_vfio_container_group_bind(int container_fd, int iommu_group_num)
2016 struct vfio_config *vfio_cfg;
2018 vfio_cfg = get_vfio_cfg_by_container_fd(container_fd);
2019 if (vfio_cfg == NULL) {
2020 RTE_LOG(ERR, EAL, "Invalid container fd\n");
2024 return vfio_get_group_fd(vfio_cfg, iommu_group_num);
2028 rte_vfio_container_group_unbind(int container_fd, int iommu_group_num)
2030 struct vfio_config *vfio_cfg;
2031 struct vfio_group *cur_grp = NULL;
2034 vfio_cfg = get_vfio_cfg_by_container_fd(container_fd);
2035 if (vfio_cfg == NULL) {
2036 RTE_LOG(ERR, EAL, "Invalid container fd\n");
2040 for (i = 0; i < VFIO_MAX_GROUPS; i++) {
2041 if (vfio_cfg->vfio_groups[i].group_num == iommu_group_num) {
2042 cur_grp = &vfio_cfg->vfio_groups[i];
2047 /* This should not happen */
2048 if (i == VFIO_MAX_GROUPS || cur_grp == NULL) {
2049 RTE_LOG(ERR, EAL, "Specified group number not found\n");
2053 if (cur_grp->fd >= 0 && close(cur_grp->fd) < 0) {
2054 RTE_LOG(ERR, EAL, "Error when closing vfio_group_fd for"
2055 " iommu_group_num %d\n", iommu_group_num);
2058 cur_grp->group_num = -1;
2060 cur_grp->devices = 0;
2061 vfio_cfg->vfio_active_groups--;
2067 rte_vfio_container_dma_map(int container_fd, uint64_t vaddr, uint64_t iova,
2070 struct vfio_config *vfio_cfg;
2077 vfio_cfg = get_vfio_cfg_by_container_fd(container_fd);
2078 if (vfio_cfg == NULL) {
2079 RTE_LOG(ERR, EAL, "Invalid container fd\n");
2083 return container_dma_map(vfio_cfg, vaddr, iova, len);
2087 rte_vfio_container_dma_unmap(int container_fd, uint64_t vaddr, uint64_t iova,
2090 struct vfio_config *vfio_cfg;
2097 vfio_cfg = get_vfio_cfg_by_container_fd(container_fd);
2098 if (vfio_cfg == NULL) {
2099 RTE_LOG(ERR, EAL, "Invalid container fd\n");
2103 return container_dma_unmap(vfio_cfg, vaddr, iova, len);
2109 rte_vfio_setup_device(__rte_unused const char *sysfs_base,
2110 __rte_unused const char *dev_addr,
2111 __rte_unused int *vfio_dev_fd,
2112 __rte_unused struct vfio_device_info *device_info)
2118 rte_vfio_release_device(__rte_unused const char *sysfs_base,
2119 __rte_unused const char *dev_addr, __rte_unused int fd)
2125 rte_vfio_enable(__rte_unused const char *modname)
2131 rte_vfio_is_enabled(__rte_unused const char *modname)
2137 rte_vfio_noiommu_is_enabled(void)
2143 rte_vfio_clear_group(__rte_unused int vfio_group_fd)
2149 rte_vfio_get_group_num(__rte_unused const char *sysfs_base,
2150 __rte_unused const char *dev_addr,
2151 __rte_unused int *iommu_group_num)
2157 rte_vfio_get_container_fd(void)
2163 rte_vfio_get_group_fd(__rte_unused int iommu_group_num)
2169 rte_vfio_container_create(void)
2175 rte_vfio_container_destroy(__rte_unused int container_fd)
2181 rte_vfio_container_group_bind(__rte_unused int container_fd,
2182 __rte_unused int iommu_group_num)
2188 rte_vfio_container_group_unbind(__rte_unused int container_fd,
2189 __rte_unused int iommu_group_num)
2195 rte_vfio_container_dma_map(__rte_unused int container_fd,
2196 __rte_unused uint64_t vaddr,
2197 __rte_unused uint64_t iova,
2198 __rte_unused uint64_t len)
2204 rte_vfio_container_dma_unmap(__rte_unused int container_fd,
2205 __rte_unused uint64_t vaddr,
2206 __rte_unused uint64_t iova,
2207 __rte_unused uint64_t len)
2212 #endif /* VFIO_PRESENT */