4 * Copyright 2016 - 2018 Red Hat, Inc.
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"
20 #include "qemu/error-report.h"
21 #include "standard-headers/linux/pci_regs.h"
22 #include "qemu/event_notifier.h"
23 #include "qemu/vfio-helpers.h"
26 #define QEMU_VFIO_DEBUG 0
28 #define QEMU_VFIO_IOVA_MIN 0x10000ULL
29 /* XXX: Once VFIO exposes the iova bit width in the IOMMU capability interface,
30 * we can use a runtime limit; alternatively it's also possible to do platform
31 * specific detection by reading sysfs entries. Until then, 39 is a safe bet.
33 #define QEMU_VFIO_IOVA_MAX (1ULL << 39)
36 /* Page aligned addr. */
42 struct QEMUVFIOState {
45 /* These fields are protected by BQL */
49 RAMBlockNotifier ram_notifier;
50 struct vfio_region_info config_region_info, bar_region_info[6];
52 /* These fields are protected by @lock */
53 /* VFIO's IO virtual address space is managed by splitting into a few
56 * --------------- <= 0
58 * |-------------| <= QEMU_VFIO_IOVA_MIN
62 * |-------------| <= low_water_mark
66 * |-------------| <= high_water_mark
70 * |-------------| <= QEMU_VFIO_IOVA_MAX
75 * - Addresses lower than QEMU_VFIO_IOVA_MIN are reserved as invalid;
77 * - Fixed mappings of HVAs are assigned "low" IOVAs in the range of
78 * [QEMU_VFIO_IOVA_MIN, low_water_mark). Once allocated they will not be
79 * reclaimed - low_water_mark never shrinks;
81 * - IOVAs in range [low_water_mark, high_water_mark) are free;
83 * - IOVAs in range [high_water_mark, QEMU_VFIO_IOVA_MAX) are volatile
84 * mappings. At each qemu_vfio_dma_reset_temporary() call, the whole area
85 * is recycled. The caller should make sure I/O's depending on these
86 * mappings are completed before calling.
88 uint64_t low_water_mark;
89 uint64_t high_water_mark;
90 IOVAMapping *mappings;
95 * Find group file by PCI device address as specified @device, and return the
96 * path. The returned string is owned by caller and should be g_free'ed later.
98 static char *sysfs_find_group_file(const char *device, Error **errp)
105 sysfs_link = g_strdup_printf("/sys/bus/pci/devices/%s/iommu_group", device);
106 sysfs_group = g_malloc0(PATH_MAX);
107 if (readlink(sysfs_link, sysfs_group, PATH_MAX - 1) == -1) {
108 error_setg_errno(errp, errno, "Failed to find iommu group sysfs path");
111 p = strrchr(sysfs_group, '/');
113 error_setg(errp, "Failed to find iommu group number");
117 path = g_strdup_printf("/dev/vfio/%s", p + 1);
124 static inline void assert_bar_index_valid(QEMUVFIOState *s, int index)
126 assert(index >= 0 && index < ARRAY_SIZE(s->bar_region_info));
129 static int qemu_vfio_pci_init_bar(QEMUVFIOState *s, int index, Error **errp)
131 assert_bar_index_valid(s, index);
132 s->bar_region_info[index] = (struct vfio_region_info) {
133 .index = VFIO_PCI_BAR0_REGION_INDEX + index,
134 .argsz = sizeof(struct vfio_region_info),
136 if (ioctl(s->device, VFIO_DEVICE_GET_REGION_INFO, &s->bar_region_info[index])) {
137 error_setg_errno(errp, errno, "Failed to get BAR region info");
145 * Map a PCI bar area.
147 void *qemu_vfio_pci_map_bar(QEMUVFIOState *s, int index,
148 uint64_t offset, uint64_t size,
152 assert_bar_index_valid(s, index);
153 p = mmap(NULL, MIN(size, s->bar_region_info[index].size - offset),
154 PROT_READ | PROT_WRITE, MAP_SHARED,
155 s->device, s->bar_region_info[index].offset + offset);
156 if (p == MAP_FAILED) {
157 error_setg_errno(errp, errno, "Failed to map BAR region");
164 * Unmap a PCI bar area.
166 void qemu_vfio_pci_unmap_bar(QEMUVFIOState *s, int index, void *bar,
167 uint64_t offset, uint64_t size)
170 munmap(bar, MIN(size, s->bar_region_info[index].size - offset));
175 * Initialize device IRQ with @irq_type and and register an event notifier.
177 int qemu_vfio_pci_init_irq(QEMUVFIOState *s, EventNotifier *e,
178 int irq_type, Error **errp)
181 struct vfio_irq_set *irq_set;
183 struct vfio_irq_info irq_info = { .argsz = sizeof(irq_info) };
185 irq_info.index = irq_type;
186 if (ioctl(s->device, VFIO_DEVICE_GET_IRQ_INFO, &irq_info)) {
187 error_setg_errno(errp, errno, "Failed to get device interrupt info");
190 if (!(irq_info.flags & VFIO_IRQ_INFO_EVENTFD)) {
191 error_setg(errp, "Device interrupt doesn't support eventfd");
195 irq_set_size = sizeof(*irq_set) + sizeof(int);
196 irq_set = g_malloc0(irq_set_size);
198 /* Get to a known IRQ state */
199 *irq_set = (struct vfio_irq_set) {
200 .argsz = irq_set_size,
201 .flags = VFIO_IRQ_SET_DATA_EVENTFD | VFIO_IRQ_SET_ACTION_TRIGGER,
202 .index = irq_info.index,
207 *(int *)&irq_set->data = event_notifier_get_fd(e);
208 r = ioctl(s->device, VFIO_DEVICE_SET_IRQS, irq_set);
211 error_setg_errno(errp, errno, "Failed to setup device interrupt");
217 static int qemu_vfio_pci_read_config(QEMUVFIOState *s, void *buf,
223 ret = pread(s->device, buf, size, s->config_region_info.offset + ofs);
224 } while (ret == -1 && errno == EINTR);
225 return ret == size ? 0 : -errno;
228 static int qemu_vfio_pci_write_config(QEMUVFIOState *s, void *buf, int size, int ofs)
233 ret = pwrite(s->device, buf, size, s->config_region_info.offset + ofs);
234 } while (ret == -1 && errno == EINTR);
235 return ret == size ? 0 : -errno;
238 static int qemu_vfio_init_pci(QEMUVFIOState *s, const char *device,
244 struct vfio_group_status group_status = { .argsz = sizeof(group_status) };
245 struct vfio_iommu_type1_info iommu_info = { .argsz = sizeof(iommu_info) };
246 struct vfio_device_info device_info = { .argsz = sizeof(device_info) };
247 char *group_file = NULL;
249 /* Create a new container */
250 s->container = open("/dev/vfio/vfio", O_RDWR);
252 if (s->container == -1) {
253 error_setg_errno(errp, errno, "Failed to open /dev/vfio/vfio");
256 if (ioctl(s->container, VFIO_GET_API_VERSION) != VFIO_API_VERSION) {
257 error_setg(errp, "Invalid VFIO version");
262 if (!ioctl(s->container, VFIO_CHECK_EXTENSION, VFIO_TYPE1_IOMMU)) {
263 error_setg_errno(errp, errno, "VFIO IOMMU check failed");
269 group_file = sysfs_find_group_file(device, errp);
275 s->group = open(group_file, O_RDWR);
276 if (s->group == -1) {
277 error_setg_errno(errp, errno, "Failed to open VFIO group file: %s",
285 /* Test the group is viable and available */
286 if (ioctl(s->group, VFIO_GROUP_GET_STATUS, &group_status)) {
287 error_setg_errno(errp, errno, "Failed to get VFIO group status");
292 if (!(group_status.flags & VFIO_GROUP_FLAGS_VIABLE)) {
293 error_setg(errp, "VFIO group is not viable");
298 /* Add the group to the container */
299 if (ioctl(s->group, VFIO_GROUP_SET_CONTAINER, &s->container)) {
300 error_setg_errno(errp, errno, "Failed to add group to VFIO container");
305 /* Enable the IOMMU model we want */
306 if (ioctl(s->container, VFIO_SET_IOMMU, VFIO_TYPE1_IOMMU)) {
307 error_setg_errno(errp, errno, "Failed to set VFIO IOMMU type");
312 /* Get additional IOMMU info */
313 if (ioctl(s->container, VFIO_IOMMU_GET_INFO, &iommu_info)) {
314 error_setg_errno(errp, errno, "Failed to get IOMMU info");
319 s->device = ioctl(s->group, VFIO_GROUP_GET_DEVICE_FD, device);
322 error_setg_errno(errp, errno, "Failed to get device fd");
327 /* Test and setup the device */
328 if (ioctl(s->device, VFIO_DEVICE_GET_INFO, &device_info)) {
329 error_setg_errno(errp, errno, "Failed to get device info");
334 if (device_info.num_regions < VFIO_PCI_CONFIG_REGION_INDEX) {
335 error_setg(errp, "Invalid device regions");
340 s->config_region_info = (struct vfio_region_info) {
341 .index = VFIO_PCI_CONFIG_REGION_INDEX,
342 .argsz = sizeof(struct vfio_region_info),
344 if (ioctl(s->device, VFIO_DEVICE_GET_REGION_INFO, &s->config_region_info)) {
345 error_setg_errno(errp, errno, "Failed to get config region info");
350 for (i = 0; i < ARRAY_SIZE(s->bar_region_info); i++) {
351 ret = qemu_vfio_pci_init_bar(s, i, errp);
357 /* Enable bus master */
358 ret = qemu_vfio_pci_read_config(s, &pci_cmd, sizeof(pci_cmd), PCI_COMMAND);
362 pci_cmd |= PCI_COMMAND_MASTER;
363 ret = qemu_vfio_pci_write_config(s, &pci_cmd, sizeof(pci_cmd), PCI_COMMAND);
375 static void qemu_vfio_ram_block_added(RAMBlockNotifier *n,
376 void *host, size_t size)
378 QEMUVFIOState *s = container_of(n, QEMUVFIOState, ram_notifier);
379 trace_qemu_vfio_ram_block_added(s, host, size);
380 qemu_vfio_dma_map(s, host, size, false, NULL);
383 static void qemu_vfio_ram_block_removed(RAMBlockNotifier *n,
384 void *host, size_t size)
386 QEMUVFIOState *s = container_of(n, QEMUVFIOState, ram_notifier);
388 trace_qemu_vfio_ram_block_removed(s, host, size);
389 qemu_vfio_dma_unmap(s, host);
393 static int qemu_vfio_init_ramblock(RAMBlock *rb, void *opaque)
395 void *host_addr = qemu_ram_get_host_addr(rb);
396 ram_addr_t length = qemu_ram_get_used_length(rb);
398 QEMUVFIOState *s = opaque;
403 ret = qemu_vfio_dma_map(s, host_addr, length, false, NULL);
405 fprintf(stderr, "qemu_vfio_init_ramblock: failed %p %" PRId64 "\n",
406 host_addr, (uint64_t)length);
411 static void qemu_vfio_open_common(QEMUVFIOState *s)
413 qemu_mutex_init(&s->lock);
414 s->ram_notifier.ram_block_added = qemu_vfio_ram_block_added;
415 s->ram_notifier.ram_block_removed = qemu_vfio_ram_block_removed;
416 ram_block_notifier_add(&s->ram_notifier);
417 s->low_water_mark = QEMU_VFIO_IOVA_MIN;
418 s->high_water_mark = QEMU_VFIO_IOVA_MAX;
419 qemu_ram_foreach_block(qemu_vfio_init_ramblock, s);
423 * Open a PCI device, e.g. "0000:00:01.0".
425 QEMUVFIOState *qemu_vfio_open_pci(const char *device, Error **errp)
428 QEMUVFIOState *s = g_new0(QEMUVFIOState, 1);
430 r = qemu_vfio_init_pci(s, device, errp);
435 qemu_vfio_open_common(s);
439 static void qemu_vfio_dump_mapping(IOVAMapping *m)
441 if (QEMU_VFIO_DEBUG) {
442 printf(" vfio mapping %p %" PRIx64 " to %" PRIx64 "\n", m->host,
443 (uint64_t)m->size, (uint64_t)m->iova);
447 static void qemu_vfio_dump_mappings(QEMUVFIOState *s)
451 if (QEMU_VFIO_DEBUG) {
452 printf("vfio mappings\n");
453 for (i = 0; i < s->nr_mappings; ++i) {
454 qemu_vfio_dump_mapping(&s->mappings[i]);
460 * Find the mapping entry that contains [host, host + size) and set @index to
461 * the position. If no entry contains it, @index is the position _after_ which
462 * to insert the new mapping. IOW, it is the index of the largest element that
463 * is smaller than @host, or -1 if no entry is.
465 static IOVAMapping *qemu_vfio_find_mapping(QEMUVFIOState *s, void *host,
468 IOVAMapping *p = s->mappings;
469 IOVAMapping *q = p ? p + s->nr_mappings - 1 : NULL;
471 trace_qemu_vfio_find_mapping(s, host);
477 mid = p + (q - p) / 2;
481 if (mid->host > host) {
483 } else if (mid->host < host) {
489 if (mid->host > host) {
491 } else if (mid < &s->mappings[s->nr_mappings - 1]
492 && (mid + 1)->host <= host) {
495 *index = mid - &s->mappings[0];
496 if (mid >= &s->mappings[0] &&
497 mid->host <= host && mid->host + mid->size > host) {
498 assert(mid < &s->mappings[s->nr_mappings]);
501 /* At this point *index + 1 is the right position to insert the new
507 * Allocate IOVA and and create a new mapping record and insert it in @s.
509 static IOVAMapping *qemu_vfio_add_mapping(QEMUVFIOState *s,
510 void *host, size_t size,
511 int index, uint64_t iova)
514 IOVAMapping m = {.host = host, .size = size, .iova = iova};
517 assert(QEMU_IS_ALIGNED(size, qemu_real_host_page_size));
518 assert(QEMU_IS_ALIGNED(s->low_water_mark, qemu_real_host_page_size));
519 assert(QEMU_IS_ALIGNED(s->high_water_mark, qemu_real_host_page_size));
520 trace_qemu_vfio_new_mapping(s, host, size, index, iova);
524 s->mappings = g_renew(IOVAMapping, s->mappings, s->nr_mappings);
525 insert = &s->mappings[index];
526 shift = s->nr_mappings - index - 1;
528 memmove(insert + 1, insert, shift * sizeof(s->mappings[0]));
534 /* Do the DMA mapping with VFIO. */
535 static int qemu_vfio_do_mapping(QEMUVFIOState *s, void *host, size_t size,
538 struct vfio_iommu_type1_dma_map dma_map = {
539 .argsz = sizeof(dma_map),
540 .flags = VFIO_DMA_MAP_FLAG_READ | VFIO_DMA_MAP_FLAG_WRITE,
542 .vaddr = (uintptr_t)host,
545 trace_qemu_vfio_do_mapping(s, host, size, iova);
547 if (ioctl(s->container, VFIO_IOMMU_MAP_DMA, &dma_map)) {
548 error_report("VFIO_MAP_DMA failed: %s", strerror(errno));
555 * Undo the DMA mapping from @s with VFIO, and remove from mapping list.
557 static void qemu_vfio_undo_mapping(QEMUVFIOState *s, IOVAMapping *mapping,
561 struct vfio_iommu_type1_dma_unmap unmap = {
562 .argsz = sizeof(unmap),
564 .iova = mapping->iova,
565 .size = mapping->size,
568 index = mapping - s->mappings;
569 assert(mapping->size > 0);
570 assert(QEMU_IS_ALIGNED(mapping->size, qemu_real_host_page_size));
571 assert(index >= 0 && index < s->nr_mappings);
572 if (ioctl(s->container, VFIO_IOMMU_UNMAP_DMA, &unmap)) {
573 error_setg_errno(errp, errno, "VFIO_UNMAP_DMA failed");
575 memmove(mapping, &s->mappings[index + 1],
576 sizeof(s->mappings[0]) * (s->nr_mappings - index - 1));
578 s->mappings = g_renew(IOVAMapping, s->mappings, s->nr_mappings);
581 /* Check if the mapping list is (ascending) ordered. */
582 static bool qemu_vfio_verify_mappings(QEMUVFIOState *s)
585 if (QEMU_VFIO_DEBUG) {
586 for (i = 0; i < s->nr_mappings - 1; ++i) {
587 if (!(s->mappings[i].host < s->mappings[i + 1].host)) {
588 fprintf(stderr, "item %d not sorted!\n", i);
589 qemu_vfio_dump_mappings(s);
592 if (!(s->mappings[i].host + s->mappings[i].size <=
593 s->mappings[i + 1].host)) {
594 fprintf(stderr, "item %d overlap with next!\n", i);
595 qemu_vfio_dump_mappings(s);
603 /* Map [host, host + size) area into a contiguous IOVA address space, and store
604 * the result in @iova if not NULL. The caller need to make sure the area is
605 * aligned to page size, and mustn't overlap with existing mapping areas (split
606 * mapping status within this area is not allowed).
608 int qemu_vfio_dma_map(QEMUVFIOState *s, void *host, size_t size,
609 bool temporary, uint64_t *iova)
613 IOVAMapping *mapping;
616 assert(QEMU_PTR_IS_ALIGNED(host, qemu_real_host_page_size));
617 assert(QEMU_IS_ALIGNED(size, qemu_real_host_page_size));
618 trace_qemu_vfio_dma_map(s, host, size, temporary, iova);
619 qemu_mutex_lock(&s->lock);
620 mapping = qemu_vfio_find_mapping(s, host, &index);
622 iova0 = mapping->iova + ((uint8_t *)host - (uint8_t *)mapping->host);
624 if (s->high_water_mark - s->low_water_mark + 1 < size) {
629 iova0 = s->low_water_mark;
630 mapping = qemu_vfio_add_mapping(s, host, size, index + 1, iova0);
635 assert(qemu_vfio_verify_mappings(s));
636 ret = qemu_vfio_do_mapping(s, host, size, iova0);
638 qemu_vfio_undo_mapping(s, mapping, NULL);
641 s->low_water_mark += size;
642 qemu_vfio_dump_mappings(s);
644 iova0 = s->high_water_mark - size;
645 ret = qemu_vfio_do_mapping(s, host, size, iova0);
649 s->high_water_mark -= size;
656 qemu_mutex_unlock(&s->lock);
660 /* Reset the high watermark and free all "temporary" mappings. */
661 int qemu_vfio_dma_reset_temporary(QEMUVFIOState *s)
663 struct vfio_iommu_type1_dma_unmap unmap = {
664 .argsz = sizeof(unmap),
666 .iova = s->high_water_mark,
667 .size = QEMU_VFIO_IOVA_MAX - s->high_water_mark,
669 trace_qemu_vfio_dma_reset_temporary(s);
670 qemu_mutex_lock(&s->lock);
671 if (ioctl(s->container, VFIO_IOMMU_UNMAP_DMA, &unmap)) {
672 error_report("VFIO_UNMAP_DMA failed: %s", strerror(errno));
673 qemu_mutex_unlock(&s->lock);
676 s->high_water_mark = QEMU_VFIO_IOVA_MAX;
677 qemu_mutex_unlock(&s->lock);
681 /* Unmapping the whole area that was previously mapped with
682 * qemu_vfio_dma_map(). */
683 void qemu_vfio_dma_unmap(QEMUVFIOState *s, void *host)
692 trace_qemu_vfio_dma_unmap(s, host);
693 qemu_mutex_lock(&s->lock);
694 m = qemu_vfio_find_mapping(s, host, &index);
698 qemu_vfio_undo_mapping(s, m, NULL);
700 qemu_mutex_unlock(&s->lock);
703 static void qemu_vfio_reset(QEMUVFIOState *s)
705 ioctl(s->device, VFIO_DEVICE_RESET);
708 /* Close and free the VFIO resources. */
709 void qemu_vfio_close(QEMUVFIOState *s)
716 for (i = 0; i < s->nr_mappings; ++i) {
717 qemu_vfio_undo_mapping(s, &s->mappings[i], NULL);
719 ram_block_notifier_remove(&s->ram_notifier);