4 * Copyright 2016 - 2018 Red Hat, Inc.
7 * Fam Zheng <famz@redhat.com>
9 * This work is licensed under the terms of the GNU GPL, version 2 or later.
10 * See the COPYING file in the top-level directory.
13 #include "qemu/osdep.h"
14 #include <sys/ioctl.h>
15 #include <linux/vfio.h>
16 #include "qapi/error.h"
17 #include "exec/ramlist.h"
18 #include "exec/cpu-common.h"
19 #include "exec/memory.h"
21 #include "qemu/error-report.h"
22 #include "standard-headers/linux/pci_regs.h"
23 #include "qemu/event_notifier.h"
24 #include "qemu/vfio-helpers.h"
25 #include "qemu/lockable.h"
28 #define QEMU_VFIO_DEBUG 0
30 #define QEMU_VFIO_IOVA_MIN 0x10000ULL
31 /* XXX: Once VFIO exposes the iova bit width in the IOMMU capability interface,
32 * we can use a runtime limit; alternatively it's also possible to do platform
33 * specific detection by reading sysfs entries. Until then, 39 is a safe bet.
35 #define QEMU_VFIO_IOVA_MAX (1ULL << 39)
38 /* Page aligned addr. */
49 struct QEMUVFIOState {
52 /* These fields are protected by BQL */
56 RAMBlockNotifier ram_notifier;
57 struct vfio_region_info config_region_info, bar_region_info[6];
58 struct IOVARange *usable_iova_ranges;
59 uint8_t nb_iova_ranges;
61 /* These fields are protected by @lock */
62 /* VFIO's IO virtual address space is managed by splitting into a few
65 * --------------- <= 0
67 * |-------------| <= QEMU_VFIO_IOVA_MIN
71 * |-------------| <= low_water_mark
75 * |-------------| <= high_water_mark
79 * |-------------| <= QEMU_VFIO_IOVA_MAX
84 * - Addresses lower than QEMU_VFIO_IOVA_MIN are reserved as invalid;
86 * - Fixed mappings of HVAs are assigned "low" IOVAs in the range of
87 * [QEMU_VFIO_IOVA_MIN, low_water_mark). Once allocated they will not be
88 * reclaimed - low_water_mark never shrinks;
90 * - IOVAs in range [low_water_mark, high_water_mark) are free;
92 * - IOVAs in range [high_water_mark, QEMU_VFIO_IOVA_MAX) are volatile
93 * mappings. At each qemu_vfio_dma_reset_temporary() call, the whole area
94 * is recycled. The caller should make sure I/O's depending on these
95 * mappings are completed before calling.
97 uint64_t low_water_mark;
98 uint64_t high_water_mark;
99 IOVAMapping *mappings;
104 * Find group file by PCI device address as specified @device, and return the
105 * path. The returned string is owned by caller and should be g_free'ed later.
107 static char *sysfs_find_group_file(const char *device, Error **errp)
114 sysfs_link = g_strdup_printf("/sys/bus/pci/devices/%s/iommu_group", device);
115 sysfs_group = g_malloc0(PATH_MAX);
116 if (readlink(sysfs_link, sysfs_group, PATH_MAX - 1) == -1) {
117 error_setg_errno(errp, errno, "Failed to find iommu group sysfs path");
120 p = strrchr(sysfs_group, '/');
122 error_setg(errp, "Failed to find iommu group number");
126 path = g_strdup_printf("/dev/vfio/%s", p + 1);
133 static inline void assert_bar_index_valid(QEMUVFIOState *s, int index)
135 assert(index >= 0 && index < ARRAY_SIZE(s->bar_region_info));
138 static int qemu_vfio_pci_init_bar(QEMUVFIOState *s, int index, Error **errp)
140 g_autofree char *barname = NULL;
141 assert_bar_index_valid(s, index);
142 s->bar_region_info[index] = (struct vfio_region_info) {
143 .index = VFIO_PCI_BAR0_REGION_INDEX + index,
144 .argsz = sizeof(struct vfio_region_info),
146 if (ioctl(s->device, VFIO_DEVICE_GET_REGION_INFO, &s->bar_region_info[index])) {
147 error_setg_errno(errp, errno, "Failed to get BAR region info");
150 barname = g_strdup_printf("bar[%d]", index);
151 trace_qemu_vfio_region_info(barname, s->bar_region_info[index].offset,
152 s->bar_region_info[index].size,
153 s->bar_region_info[index].cap_offset);
159 * Map a PCI bar area.
161 void *qemu_vfio_pci_map_bar(QEMUVFIOState *s, int index,
162 uint64_t offset, uint64_t size, int prot,
166 assert(QEMU_IS_ALIGNED(offset, qemu_real_host_page_size));
167 assert_bar_index_valid(s, index);
168 p = mmap(NULL, MIN(size, s->bar_region_info[index].size - offset),
170 s->device, s->bar_region_info[index].offset + offset);
171 trace_qemu_vfio_pci_map_bar(index, s->bar_region_info[index].offset ,
173 if (p == MAP_FAILED) {
174 error_setg_errno(errp, errno, "Failed to map BAR region");
181 * Unmap a PCI bar area.
183 void qemu_vfio_pci_unmap_bar(QEMUVFIOState *s, int index, void *bar,
184 uint64_t offset, uint64_t size)
187 munmap(bar, MIN(size, s->bar_region_info[index].size - offset));
192 * Initialize device IRQ with @irq_type and register an event notifier.
194 int qemu_vfio_pci_init_irq(QEMUVFIOState *s, EventNotifier *e,
195 int irq_type, Error **errp)
198 struct vfio_irq_set *irq_set;
200 struct vfio_irq_info irq_info = { .argsz = sizeof(irq_info) };
202 irq_info.index = irq_type;
203 if (ioctl(s->device, VFIO_DEVICE_GET_IRQ_INFO, &irq_info)) {
204 error_setg_errno(errp, errno, "Failed to get device interrupt info");
207 if (!(irq_info.flags & VFIO_IRQ_INFO_EVENTFD)) {
208 error_setg(errp, "Device interrupt doesn't support eventfd");
212 irq_set_size = sizeof(*irq_set) + sizeof(int);
213 irq_set = g_malloc0(irq_set_size);
215 /* Get to a known IRQ state */
216 *irq_set = (struct vfio_irq_set) {
217 .argsz = irq_set_size,
218 .flags = VFIO_IRQ_SET_DATA_EVENTFD | VFIO_IRQ_SET_ACTION_TRIGGER,
219 .index = irq_info.index,
224 *(int *)&irq_set->data = event_notifier_get_fd(e);
225 r = ioctl(s->device, VFIO_DEVICE_SET_IRQS, irq_set);
228 error_setg_errno(errp, errno, "Failed to setup device interrupt");
234 static int qemu_vfio_pci_read_config(QEMUVFIOState *s, void *buf,
239 trace_qemu_vfio_pci_read_config(buf, ofs, size,
240 s->config_region_info.offset,
241 s->config_region_info.size);
242 assert(QEMU_IS_ALIGNED(s->config_region_info.offset + ofs, size));
244 ret = pread(s->device, buf, size, s->config_region_info.offset + ofs);
245 } while (ret == -1 && errno == EINTR);
246 return ret == size ? 0 : -errno;
249 static int qemu_vfio_pci_write_config(QEMUVFIOState *s, void *buf, int size, int ofs)
253 trace_qemu_vfio_pci_write_config(buf, ofs, size,
254 s->config_region_info.offset,
255 s->config_region_info.size);
256 assert(QEMU_IS_ALIGNED(s->config_region_info.offset + ofs, size));
258 ret = pwrite(s->device, buf, size, s->config_region_info.offset + ofs);
259 } while (ret == -1 && errno == EINTR);
260 return ret == size ? 0 : -errno;
263 static void collect_usable_iova_ranges(QEMUVFIOState *s, void *buf)
265 struct vfio_iommu_type1_info *info = (struct vfio_iommu_type1_info *)buf;
266 struct vfio_info_cap_header *cap = (void *)buf + info->cap_offset;
267 struct vfio_iommu_type1_info_cap_iova_range *cap_iova_range;
270 while (cap->id != VFIO_IOMMU_TYPE1_INFO_CAP_IOVA_RANGE) {
274 cap = (struct vfio_info_cap_header *)(buf + cap->next);
277 cap_iova_range = (struct vfio_iommu_type1_info_cap_iova_range *)cap;
279 s->nb_iova_ranges = cap_iova_range->nr_iovas;
280 if (s->nb_iova_ranges > 1) {
281 s->usable_iova_ranges =
282 g_realloc(s->usable_iova_ranges,
283 s->nb_iova_ranges * sizeof(struct IOVARange));
286 for (i = 0; i < s->nb_iova_ranges; i++) {
287 s->usable_iova_ranges[i].start = cap_iova_range->iova_ranges[i].start;
288 s->usable_iova_ranges[i].end = cap_iova_range->iova_ranges[i].end;
292 static int qemu_vfio_init_pci(QEMUVFIOState *s, const char *device,
298 struct vfio_group_status group_status = { .argsz = sizeof(group_status) };
299 struct vfio_iommu_type1_info *iommu_info = NULL;
300 size_t iommu_info_size = sizeof(*iommu_info);
301 struct vfio_device_info device_info = { .argsz = sizeof(device_info) };
302 char *group_file = NULL;
304 s->usable_iova_ranges = NULL;
306 /* Create a new container */
307 s->container = open("/dev/vfio/vfio", O_RDWR);
309 if (s->container == -1) {
310 error_setg_errno(errp, errno, "Failed to open /dev/vfio/vfio");
313 if (ioctl(s->container, VFIO_GET_API_VERSION) != VFIO_API_VERSION) {
314 error_setg(errp, "Invalid VFIO version");
319 if (!ioctl(s->container, VFIO_CHECK_EXTENSION, VFIO_TYPE1_IOMMU)) {
320 error_setg_errno(errp, errno, "VFIO IOMMU Type1 is not supported");
326 group_file = sysfs_find_group_file(device, errp);
332 s->group = open(group_file, O_RDWR);
333 if (s->group == -1) {
334 error_setg_errno(errp, errno, "Failed to open VFIO group file: %s",
342 /* Test the group is viable and available */
343 if (ioctl(s->group, VFIO_GROUP_GET_STATUS, &group_status)) {
344 error_setg_errno(errp, errno, "Failed to get VFIO group status");
349 if (!(group_status.flags & VFIO_GROUP_FLAGS_VIABLE)) {
350 error_setg(errp, "VFIO group is not viable");
355 /* Add the group to the container */
356 if (ioctl(s->group, VFIO_GROUP_SET_CONTAINER, &s->container)) {
357 error_setg_errno(errp, errno, "Failed to add group to VFIO container");
362 /* Enable the IOMMU model we want */
363 if (ioctl(s->container, VFIO_SET_IOMMU, VFIO_TYPE1_IOMMU)) {
364 error_setg_errno(errp, errno, "Failed to set VFIO IOMMU type");
369 iommu_info = g_malloc0(iommu_info_size);
370 iommu_info->argsz = iommu_info_size;
372 /* Get additional IOMMU info */
373 if (ioctl(s->container, VFIO_IOMMU_GET_INFO, iommu_info)) {
374 error_setg_errno(errp, errno, "Failed to get IOMMU info");
380 * if the kernel does not report usable IOVA regions, choose
381 * the legacy [QEMU_VFIO_IOVA_MIN, QEMU_VFIO_IOVA_MAX -1] region
383 s->nb_iova_ranges = 1;
384 s->usable_iova_ranges = g_new0(struct IOVARange, 1);
385 s->usable_iova_ranges[0].start = QEMU_VFIO_IOVA_MIN;
386 s->usable_iova_ranges[0].end = QEMU_VFIO_IOVA_MAX - 1;
388 if (iommu_info->argsz > iommu_info_size) {
389 iommu_info_size = iommu_info->argsz;
390 iommu_info = g_realloc(iommu_info, iommu_info_size);
391 if (ioctl(s->container, VFIO_IOMMU_GET_INFO, iommu_info)) {
395 collect_usable_iova_ranges(s, iommu_info);
398 s->device = ioctl(s->group, VFIO_GROUP_GET_DEVICE_FD, device);
401 error_setg_errno(errp, errno, "Failed to get device fd");
406 /* Test and setup the device */
407 if (ioctl(s->device, VFIO_DEVICE_GET_INFO, &device_info)) {
408 error_setg_errno(errp, errno, "Failed to get device info");
413 if (device_info.num_regions < VFIO_PCI_CONFIG_REGION_INDEX) {
414 error_setg(errp, "Invalid device regions");
419 s->config_region_info = (struct vfio_region_info) {
420 .index = VFIO_PCI_CONFIG_REGION_INDEX,
421 .argsz = sizeof(struct vfio_region_info),
423 if (ioctl(s->device, VFIO_DEVICE_GET_REGION_INFO, &s->config_region_info)) {
424 error_setg_errno(errp, errno, "Failed to get config region info");
428 trace_qemu_vfio_region_info("config", s->config_region_info.offset,
429 s->config_region_info.size,
430 s->config_region_info.cap_offset);
432 for (i = 0; i < ARRAY_SIZE(s->bar_region_info); i++) {
433 ret = qemu_vfio_pci_init_bar(s, i, errp);
439 /* Enable bus master */
440 ret = qemu_vfio_pci_read_config(s, &pci_cmd, sizeof(pci_cmd), PCI_COMMAND);
444 pci_cmd |= PCI_COMMAND_MASTER;
445 ret = qemu_vfio_pci_write_config(s, &pci_cmd, sizeof(pci_cmd), PCI_COMMAND);
452 g_free(s->usable_iova_ranges);
453 s->usable_iova_ranges = NULL;
454 s->nb_iova_ranges = 0;
462 static void qemu_vfio_ram_block_added(RAMBlockNotifier *n, void *host,
463 size_t size, size_t max_size)
465 QEMUVFIOState *s = container_of(n, QEMUVFIOState, ram_notifier);
468 trace_qemu_vfio_ram_block_added(s, host, max_size);
469 ret = qemu_vfio_dma_map(s, host, max_size, false, NULL);
471 error_report("qemu_vfio_dma_map(%p, %zu) failed: %s", host, max_size,
476 static void qemu_vfio_ram_block_removed(RAMBlockNotifier *n, void *host,
477 size_t size, size_t max_size)
479 QEMUVFIOState *s = container_of(n, QEMUVFIOState, ram_notifier);
481 trace_qemu_vfio_ram_block_removed(s, host, max_size);
482 qemu_vfio_dma_unmap(s, host);
486 static void qemu_vfio_open_common(QEMUVFIOState *s)
488 qemu_mutex_init(&s->lock);
489 s->ram_notifier.ram_block_added = qemu_vfio_ram_block_added;
490 s->ram_notifier.ram_block_removed = qemu_vfio_ram_block_removed;
491 s->low_water_mark = QEMU_VFIO_IOVA_MIN;
492 s->high_water_mark = QEMU_VFIO_IOVA_MAX;
493 ram_block_notifier_add(&s->ram_notifier);
497 * Open a PCI device, e.g. "0000:00:01.0".
499 QEMUVFIOState *qemu_vfio_open_pci(const char *device, Error **errp)
502 QEMUVFIOState *s = g_new0(QEMUVFIOState, 1);
505 * VFIO may pin all memory inside mappings, resulting it in pinning
506 * all memory inside RAM blocks unconditionally.
508 r = ram_block_discard_disable(true);
510 error_setg_errno(errp, -r, "Cannot set discarding of RAM broken");
515 r = qemu_vfio_init_pci(s, device, errp);
517 ram_block_discard_disable(false);
521 qemu_vfio_open_common(s);
525 static void qemu_vfio_dump_mappings(QEMUVFIOState *s)
527 for (int i = 0; i < s->nr_mappings; ++i) {
528 trace_qemu_vfio_dump_mapping(s->mappings[i].host,
530 s->mappings[i].size);
535 * Find the mapping entry that contains [host, host + size) and set @index to
536 * the position. If no entry contains it, @index is the position _after_ which
537 * to insert the new mapping. IOW, it is the index of the largest element that
538 * is smaller than @host, or -1 if no entry is.
540 static IOVAMapping *qemu_vfio_find_mapping(QEMUVFIOState *s, void *host,
543 IOVAMapping *p = s->mappings;
544 IOVAMapping *q = p ? p + s->nr_mappings - 1 : NULL;
546 trace_qemu_vfio_find_mapping(s, host);
552 mid = p + (q - p) / 2;
556 if (mid->host > host) {
558 } else if (mid->host < host) {
564 if (mid->host > host) {
566 } else if (mid < &s->mappings[s->nr_mappings - 1]
567 && (mid + 1)->host <= host) {
570 *index = mid - &s->mappings[0];
571 if (mid >= &s->mappings[0] &&
572 mid->host <= host && mid->host + mid->size > host) {
573 assert(mid < &s->mappings[s->nr_mappings]);
576 /* At this point *index + 1 is the right position to insert the new
582 * Allocate IOVA and create a new mapping record and insert it in @s.
584 static IOVAMapping *qemu_vfio_add_mapping(QEMUVFIOState *s,
585 void *host, size_t size,
586 int index, uint64_t iova)
589 IOVAMapping m = {.host = host, .size = size, .iova = iova};
592 assert(QEMU_IS_ALIGNED(size, qemu_real_host_page_size));
593 assert(QEMU_IS_ALIGNED(s->low_water_mark, qemu_real_host_page_size));
594 assert(QEMU_IS_ALIGNED(s->high_water_mark, qemu_real_host_page_size));
595 trace_qemu_vfio_new_mapping(s, host, size, index, iova);
599 s->mappings = g_renew(IOVAMapping, s->mappings, s->nr_mappings);
600 insert = &s->mappings[index];
601 shift = s->nr_mappings - index - 1;
603 memmove(insert + 1, insert, shift * sizeof(s->mappings[0]));
609 /* Do the DMA mapping with VFIO. */
610 static int qemu_vfio_do_mapping(QEMUVFIOState *s, void *host, size_t size,
613 struct vfio_iommu_type1_dma_map dma_map = {
614 .argsz = sizeof(dma_map),
615 .flags = VFIO_DMA_MAP_FLAG_READ | VFIO_DMA_MAP_FLAG_WRITE,
617 .vaddr = (uintptr_t)host,
620 trace_qemu_vfio_do_mapping(s, host, iova, size);
622 if (ioctl(s->container, VFIO_IOMMU_MAP_DMA, &dma_map)) {
623 error_report("VFIO_MAP_DMA failed: %s", strerror(errno));
630 * Undo the DMA mapping from @s with VFIO, and remove from mapping list.
632 static void qemu_vfio_undo_mapping(QEMUVFIOState *s, IOVAMapping *mapping,
636 struct vfio_iommu_type1_dma_unmap unmap = {
637 .argsz = sizeof(unmap),
639 .iova = mapping->iova,
640 .size = mapping->size,
643 index = mapping - s->mappings;
644 assert(mapping->size > 0);
645 assert(QEMU_IS_ALIGNED(mapping->size, qemu_real_host_page_size));
646 assert(index >= 0 && index < s->nr_mappings);
647 if (ioctl(s->container, VFIO_IOMMU_UNMAP_DMA, &unmap)) {
648 error_setg_errno(errp, errno, "VFIO_UNMAP_DMA failed");
650 memmove(mapping, &s->mappings[index + 1],
651 sizeof(s->mappings[0]) * (s->nr_mappings - index - 1));
653 s->mappings = g_renew(IOVAMapping, s->mappings, s->nr_mappings);
656 /* Check if the mapping list is (ascending) ordered. */
657 static bool qemu_vfio_verify_mappings(QEMUVFIOState *s)
660 if (QEMU_VFIO_DEBUG) {
661 for (i = 0; i < s->nr_mappings - 1; ++i) {
662 if (!(s->mappings[i].host < s->mappings[i + 1].host)) {
663 fprintf(stderr, "item %d not sorted!\n", i);
664 qemu_vfio_dump_mappings(s);
667 if (!(s->mappings[i].host + s->mappings[i].size <=
668 s->mappings[i + 1].host)) {
669 fprintf(stderr, "item %d overlap with next!\n", i);
670 qemu_vfio_dump_mappings(s);
679 qemu_vfio_find_fixed_iova(QEMUVFIOState *s, size_t size, uint64_t *iova)
683 for (i = 0; i < s->nb_iova_ranges; i++) {
684 if (s->usable_iova_ranges[i].end < s->low_water_mark) {
688 MAX(s->low_water_mark, s->usable_iova_ranges[i].start);
690 if (s->usable_iova_ranges[i].end - s->low_water_mark + 1 >= size ||
691 s->usable_iova_ranges[i].end - s->low_water_mark + 1 == 0) {
692 *iova = s->low_water_mark;
693 s->low_water_mark += size;
701 qemu_vfio_find_temp_iova(QEMUVFIOState *s, size_t size, uint64_t *iova)
705 for (i = s->nb_iova_ranges - 1; i >= 0; i--) {
706 if (s->usable_iova_ranges[i].start > s->high_water_mark) {
710 MIN(s->high_water_mark, s->usable_iova_ranges[i].end + 1);
712 if (s->high_water_mark - s->usable_iova_ranges[i].start + 1 >= size ||
713 s->high_water_mark - s->usable_iova_ranges[i].start + 1 == 0) {
714 *iova = s->high_water_mark - size;
715 s->high_water_mark = *iova;
722 /* Map [host, host + size) area into a contiguous IOVA address space, and store
723 * the result in @iova if not NULL. The caller need to make sure the area is
724 * aligned to page size, and mustn't overlap with existing mapping areas (split
725 * mapping status within this area is not allowed).
727 int qemu_vfio_dma_map(QEMUVFIOState *s, void *host, size_t size,
728 bool temporary, uint64_t *iova)
732 IOVAMapping *mapping;
735 assert(QEMU_PTR_IS_ALIGNED(host, qemu_real_host_page_size));
736 assert(QEMU_IS_ALIGNED(size, qemu_real_host_page_size));
737 trace_qemu_vfio_dma_map(s, host, size, temporary, iova);
738 qemu_mutex_lock(&s->lock);
739 mapping = qemu_vfio_find_mapping(s, host, &index);
741 iova0 = mapping->iova + ((uint8_t *)host - (uint8_t *)mapping->host);
743 if (s->high_water_mark - s->low_water_mark + 1 < size) {
748 if (qemu_vfio_find_fixed_iova(s, size, &iova0)) {
753 mapping = qemu_vfio_add_mapping(s, host, size, index + 1, iova0);
758 assert(qemu_vfio_verify_mappings(s));
759 ret = qemu_vfio_do_mapping(s, host, size, iova0);
761 qemu_vfio_undo_mapping(s, mapping, NULL);
764 qemu_vfio_dump_mappings(s);
766 if (qemu_vfio_find_temp_iova(s, size, &iova0)) {
770 ret = qemu_vfio_do_mapping(s, host, size, iova0);
776 trace_qemu_vfio_dma_mapped(s, host, iova0, size);
781 qemu_mutex_unlock(&s->lock);
785 /* Reset the high watermark and free all "temporary" mappings. */
786 int qemu_vfio_dma_reset_temporary(QEMUVFIOState *s)
788 struct vfio_iommu_type1_dma_unmap unmap = {
789 .argsz = sizeof(unmap),
791 .iova = s->high_water_mark,
792 .size = QEMU_VFIO_IOVA_MAX - s->high_water_mark,
794 trace_qemu_vfio_dma_reset_temporary(s);
795 QEMU_LOCK_GUARD(&s->lock);
796 if (ioctl(s->container, VFIO_IOMMU_UNMAP_DMA, &unmap)) {
797 error_report("VFIO_UNMAP_DMA failed: %s", strerror(errno));
800 s->high_water_mark = QEMU_VFIO_IOVA_MAX;
804 /* Unmapping the whole area that was previously mapped with
805 * qemu_vfio_dma_map(). */
806 void qemu_vfio_dma_unmap(QEMUVFIOState *s, void *host)
815 trace_qemu_vfio_dma_unmap(s, host);
816 qemu_mutex_lock(&s->lock);
817 m = qemu_vfio_find_mapping(s, host, &index);
821 qemu_vfio_undo_mapping(s, m, NULL);
823 qemu_mutex_unlock(&s->lock);
826 static void qemu_vfio_reset(QEMUVFIOState *s)
828 ioctl(s->device, VFIO_DEVICE_RESET);
831 /* Close and free the VFIO resources. */
832 void qemu_vfio_close(QEMUVFIOState *s)
839 for (i = 0; i < s->nr_mappings; ++i) {
840 qemu_vfio_undo_mapping(s, &s->mappings[i], NULL);
842 ram_block_notifier_remove(&s->ram_notifier);
843 g_free(s->usable_iova_ranges);
844 s->nb_iova_ranges = 0;
849 ram_block_discard_disable(false);