master
c 296 lines 9.35 KB
Raw
1 /*
2 * Copyright (c) 2021-2025 Oracle and/or its affiliates.
3 *
4 * SPDX-License-Identifier: GPL-2.0-or-later
5 */
6
7 #include "qemu/osdep.h"
8 #include <sys/ioctl.h>
9 #include <linux/vfio.h>
10 #include "hw/vfio/vfio-container-legacy.h"
11 #include "hw/vfio/vfio-device.h"
12 #include "hw/vfio/vfio-listener.h"
13 #include "migration/blocker.h"
14 #include "migration/cpr.h"
15 #include "migration/migration.h"
16 #include "migration/vmstate.h"
17 #include "qapi/error.h"
18 #include "qemu/error-report.h"
19
20 static bool vfio_dma_unmap_vaddr_all(VFIOLegacyContainer *container,
21 Error **errp)
22 {
23 struct vfio_iommu_type1_dma_unmap unmap = {
24 .argsz = sizeof(unmap),
25 .flags = VFIO_DMA_UNMAP_FLAG_VADDR | VFIO_DMA_UNMAP_FLAG_ALL,
26 .iova = 0,
27 .size = 0,
28 };
29 if (ioctl(container->fd, VFIO_IOMMU_UNMAP_DMA, &unmap)) {
30 error_setg_errno(errp, errno, "vfio_dma_unmap_vaddr_all");
31 return false;
32 }
33 container->cpr.vaddr_unmapped = true;
34 return true;
35 }
36
37 /*
38 * Set the new @vaddr for any mappings registered during cpr load.
39 * The incoming state is cleared thereafter.
40 */
41 static int vfio_legacy_cpr_dma_map(const VFIOContainer *bcontainer,
42 hwaddr iova, uint64_t size, void *vaddr,
43 bool readonly, MemoryRegion *mr)
44 {
45 const VFIOLegacyContainer *container = VFIO_IOMMU_LEGACY(bcontainer);
46
47 struct vfio_iommu_type1_dma_map map = {
48 .argsz = sizeof(map),
49 .flags = VFIO_DMA_MAP_FLAG_VADDR,
50 .vaddr = (__u64)(uintptr_t)vaddr,
51 .iova = iova,
52 .size = size,
53 };
54
55 if (ioctl(container->fd, VFIO_IOMMU_MAP_DMA, &map)) {
56 return -errno;
57 }
58
59 return 0;
60 }
61
62 static void vfio_region_remap(MemoryListener *listener,
63 MemoryRegionSection *section)
64 {
65 VFIOLegacyContainer *container = container_of(listener,
66 VFIOLegacyContainer,
67 cpr.remap_listener);
68 vfio_container_region_add(VFIO_IOMMU(container), section, true);
69 }
70
71 static bool vfio_cpr_supported(VFIOLegacyContainer *container, Error **errp)
72 {
73 if (!ioctl(container->fd, VFIO_CHECK_EXTENSION, VFIO_UPDATE_VADDR)) {
74 error_setg(errp, "VFIO container does not support VFIO_UPDATE_VADDR");
75 return false;
76
77 } else if (!ioctl(container->fd, VFIO_CHECK_EXTENSION, VFIO_UNMAP_ALL)) {
78 error_setg(errp, "VFIO container does not support VFIO_UNMAP_ALL");
79 return false;
80
81 } else {
82 return true;
83 }
84 }
85
86 static int vfio_container_pre_save(void *opaque)
87 {
88 VFIOLegacyContainer *container = opaque;
89 Error *local_err = NULL;
90
91 if (!vfio_dma_unmap_vaddr_all(container, &local_err)) {
92 error_report_err(local_err);
93 return -1;
94 }
95 return 0;
96 }
97
98 static int vfio_container_post_load(void *opaque, int version_id)
99 {
100 VFIOLegacyContainer *container = opaque;
101 VFIOContainer *bcontainer = VFIO_IOMMU(container);
102 VFIOIOMMUClass *vioc = VFIO_IOMMU_GET_CLASS(bcontainer);
103 dma_map_fn saved_dma_map = vioc->dma_map;
104 Error *local_err = NULL;
105
106 /* During incoming CPR, divert calls to dma_map. */
107 vioc->dma_map = vfio_legacy_cpr_dma_map;
108
109 if (!vfio_listener_register(bcontainer, &local_err)) {
110 error_report_err(local_err);
111 return -1;
112 }
113
114 /* Restore original dma_map function */
115 vioc->dma_map = saved_dma_map;
116
117 return 0;
118 }
119
120 static const VMStateDescription vfio_container_vmstate = {
121 .name = "vfio-container",
122 .version_id = 0,
123 .minimum_version_id = 0,
124 .priority = MIG_PRI_LOW, /* Must happen after devices and groups */
125 .pre_save = vfio_container_pre_save,
126 .post_load = vfio_container_post_load,
127 .needed = cpr_incoming_needed,
128 .fields = (VMStateField[]) {
129 VMSTATE_END_OF_LIST()
130 }
131 };
132
133 static int vfio_cpr_fail_notifier(NotifierWithReturn *notifier,
134 MigrationEvent *e, Error **errp)
135 {
136 VFIOLegacyContainer *container =
137 container_of(notifier, VFIOLegacyContainer, cpr.transfer_notifier);
138 VFIOContainer *bcontainer = VFIO_IOMMU(container);
139
140 if (e->type != MIG_EVENT_FAILED) {
141 return 0;
142 }
143
144 if (container->cpr.vaddr_unmapped) {
145 /*
146 * Force a call to vfio_region_remap for each mapped section by
147 * temporarily registering a listener, and temporarily diverting
148 * dma_map to vfio_legacy_cpr_dma_map. The latter restores vaddr.
149 */
150
151 VFIOIOMMUClass *vioc = VFIO_IOMMU_GET_CLASS(bcontainer);
152 dma_map_fn saved_dma_map = vioc->dma_map;
153 vioc->dma_map = vfio_legacy_cpr_dma_map;
154
155 container->cpr.remap_listener = (MemoryListener) {
156 .name = "vfio cpr recover",
157 .region_add = vfio_region_remap
158 };
159 memory_listener_register(&container->cpr.remap_listener,
160 bcontainer->space->as);
161 memory_listener_unregister(&container->cpr.remap_listener);
162 container->cpr.vaddr_unmapped = false;
163 vioc->dma_map = saved_dma_map;
164 }
165 return 0;
166 }
167
168 bool vfio_legacy_cpr_register_container(VFIOLegacyContainer *container,
169 Error **errp)
170 {
171 VFIOContainer *bcontainer = VFIO_IOMMU(container);
172 Error **cpr_blocker = &container->cpr.blocker;
173
174 migration_add_notifier_mode(&bcontainer->cpr_reboot_notifier,
175 vfio_cpr_reboot_notifier,
176 MIG_MODE_CPR_REBOOT);
177
178 if (!vfio_cpr_supported(container, cpr_blocker)) {
179 return migrate_add_blocker_modes(cpr_blocker,
180 BIT(MIG_MODE_CPR_TRANSFER) | BIT(MIG_MODE_CPR_EXEC),
181 errp) == 0;
182 }
183
184 vfio_cpr_add_kvm_notifier();
185
186 vmstate_register(NULL, -1, &vfio_container_vmstate, container);
187
188 migration_add_notifier_modes(&container->cpr.transfer_notifier,
189 vfio_cpr_fail_notifier,
190 BIT(MIG_MODE_CPR_TRANSFER) | BIT(MIG_MODE_CPR_EXEC));
191 return true;
192 }
193
194 void vfio_legacy_cpr_unregister_container(VFIOLegacyContainer *container)
195 {
196 VFIOContainer *bcontainer = VFIO_IOMMU(container);
197
198 migration_remove_notifier(&bcontainer->cpr_reboot_notifier);
199 migrate_del_blocker(&container->cpr.blocker);
200 vmstate_unregister(NULL, &vfio_container_vmstate, container);
201 migration_remove_notifier(&container->cpr.transfer_notifier);
202 }
203
204 /*
205 * In old QEMU, VFIO_DMA_UNMAP_FLAG_VADDR may fail on some mapping after
206 * succeeding for others, so the latter have lost their vaddr. Call this
207 * to restore vaddr for a section with a giommu.
208 *
209 * The giommu already exists. Find it and replay it, which calls
210 * vfio_legacy_cpr_dma_map further down the stack.
211 */
212 void vfio_cpr_giommu_remap(VFIOContainer *bcontainer,
213 MemoryRegionSection *section)
214 {
215 VFIOGuestIOMMU *giommu;
216 hwaddr as_offset = section->offset_within_address_space;
217 hwaddr iommu_offset = as_offset - section->offset_within_region;
218
219 QLIST_FOREACH(giommu, &bcontainer->giommu_list, giommu_next) {
220 if (giommu->iommu_mr == IOMMU_MEMORY_REGION(section->mr) &&
221 giommu->iommu_offset == iommu_offset) {
222 break;
223 }
224 }
225 g_assert(giommu);
226 memory_region_iommu_replay(giommu->iommu_mr, &giommu->n);
227 }
228
229 static int vfio_cpr_rdm_remap(const MemoryRegionSection *section, void *opaque)
230 {
231 RamDiscardListener *rdl = opaque;
232
233 return rdl->notify_populate(rdl, section);
234 }
235
236 /*
237 * In old QEMU, VFIO_DMA_UNMAP_FLAG_VADDR may fail on some mapping after
238 * succeeding for others, so the latter have lost their vaddr. Call this
239 * to restore vaddr for populated parts in a section with a RamDiscardManager.
240 *
241 * The ram discard listener already exists. Call its replay_populated function
242 * directly, which calls vfio_legacy_cpr_dma_map.
243 */
244 bool vfio_cpr_ram_discard_replay_populated(VFIOContainer *bcontainer,
245 const MemoryRegionSection *section)
246 {
247 RamDiscardManager *rdm = memory_region_get_ram_discard_manager(section->mr);
248 VFIORamDiscardListener *vrdl =
249 vfio_find_ram_discard_listener(bcontainer, section);
250
251 g_assert(vrdl);
252 return ram_discard_manager_replay_populated(rdm, section,
253 vfio_cpr_rdm_remap,
254 &vrdl->listener) == 0;
255 }
256
257 int vfio_cpr_group_get_device_fd(int d, const char *name)
258 {
259 const int id = 0;
260 int fd = cpr_find_fd(name, id);
261
262 if (fd < 0) {
263 fd = ioctl(d, VFIO_GROUP_GET_DEVICE_FD, name);
264 if (fd >= 0) {
265 cpr_save_fd(name, id, fd);
266 }
267 }
268 return fd;
269 }
270
271 static bool same_device(int fd1, int fd2)
272 {
273 struct stat st1, st2;
274
275 return !fstat(fd1, &st1) && !fstat(fd2, &st2) && st1.st_dev == st2.st_dev;
276 }
277
278 bool vfio_cpr_container_match(VFIOLegacyContainer *container, VFIOGroup *group,
279 int fd)
280 {
281 if (container->fd == fd) {
282 return true;
283 }
284 if (!same_device(container->fd, fd)) {
285 return false;
286 }
287 /*
288 * Same device, different fd. This occurs when the container fd is
289 * cpr_save'd multiple times, once for each groupid, so SCM_RIGHTS
290 * produces duplicates. De-dup it.
291 */
292 cpr_delete_fd("vfio_container_for_group", group->groupid);
293 close(fd);
294 cpr_save_fd("vfio_container_for_group", group->groupid, container->fd);
295 return true;
296 }