]>
Commit | Line | Data |
---|---|---|
e2c7d025 EA |
1 | /* |
2 | * generic functions used by VFIO devices | |
3 | * | |
4 | * Copyright Red Hat, Inc. 2012 | |
5 | * | |
6 | * Authors: | |
7 | * Alex Williamson <alex.williamson@redhat.com> | |
8 | * | |
9 | * This work is licensed under the terms of the GNU GPL, version 2. See | |
10 | * the COPYING file in the top-level directory. | |
11 | * | |
12 | * Based on qemu-kvm device-assignment: | |
13 | * Adapted for KVM by Qumranet. | |
14 | * Copyright (c) 2007, Neocleus, Alex Novik (alex@neocleus.com) | |
15 | * Copyright (c) 2007, Neocleus, Guy Zana (guy@neocleus.com) | |
16 | * Copyright (C) 2008, Qumranet, Amit Shah (amit.shah@qumranet.com) | |
17 | * Copyright (C) 2008, Red Hat, Amit Shah (amit.shah@redhat.com) | |
18 | * Copyright (C) 2008, IBM, Muli Ben-Yehuda (muli@il.ibm.com) | |
19 | */ | |
20 | ||
c6eacb1a | 21 | #include "qemu/osdep.h" |
e2c7d025 | 22 | #include <sys/ioctl.h> |
a9c94277 MA |
23 | #ifdef CONFIG_KVM |
24 | #include <linux/kvm.h> | |
25 | #endif | |
e2c7d025 EA |
26 | #include <linux/vfio.h> |
27 | ||
28 | #include "hw/vfio/vfio-common.h" | |
29 | #include "hw/vfio/vfio.h" | |
30 | #include "exec/address-spaces.h" | |
31 | #include "exec/memory.h" | |
b6dd6504 | 32 | #include "exec/ram_addr.h" |
e2c7d025 EA |
33 | #include "hw/hw.h" |
34 | #include "qemu/error-report.h" | |
db725815 | 35 | #include "qemu/main-loop.h" |
f4ec5e26 | 36 | #include "qemu/range.h" |
e2c7d025 | 37 | #include "sysemu/kvm.h" |
71e8a915 | 38 | #include "sysemu/reset.h" |
0fd7616e | 39 | #include "sysemu/runstate.h" |
e2c7d025 | 40 | #include "trace.h" |
01905f58 | 41 | #include "qapi/error.h" |
b6dd6504 | 42 | #include "migration/migration.h" |
8b942af3 | 43 | #include "migration/misc.h" |
29d81b71 | 44 | #include "migration/blocker.h" |
851d6d1a | 45 | #include "sysemu/tpm.h" |
e2c7d025 | 46 | |
f481ee2d | 47 | VFIOGroupList vfio_group_list = |
39cb514f | 48 | QLIST_HEAD_INITIALIZER(vfio_group_list); |
10ca76b4 | 49 | static QLIST_HEAD(, VFIOAddressSpace) vfio_address_spaces = |
e2c7d025 EA |
50 | QLIST_HEAD_INITIALIZER(vfio_address_spaces); |
51 | ||
52 | #ifdef CONFIG_KVM | |
53 | /* | |
54 | * We have a single VFIO pseudo device per KVM VM. Once created it lives | |
55 | * for the life of the VM. Closing the file descriptor only drops our | |
56 | * reference to it and the device's reference to kvm. Therefore once | |
57 | * initialized, this file descriptor is only released on QEMU exit and | |
58 | * we'll re-use it should another vfio device be attached before then. | |
59 | */ | |
60 | static int vfio_kvm_device_fd = -1; | |
61 | #endif | |
62 | ||
63 | /* | |
64 | * Common VFIO interrupt disable | |
65 | */ | |
66 | void vfio_disable_irqindex(VFIODevice *vbasedev, int index) | |
67 | { | |
68 | struct vfio_irq_set irq_set = { | |
69 | .argsz = sizeof(irq_set), | |
70 | .flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_TRIGGER, | |
71 | .index = index, | |
72 | .start = 0, | |
73 | .count = 0, | |
74 | }; | |
75 | ||
76 | ioctl(vbasedev->fd, VFIO_DEVICE_SET_IRQS, &irq_set); | |
77 | } | |
78 | ||
79 | void vfio_unmask_single_irqindex(VFIODevice *vbasedev, int index) | |
80 | { | |
81 | struct vfio_irq_set irq_set = { | |
82 | .argsz = sizeof(irq_set), | |
83 | .flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_UNMASK, | |
84 | .index = index, | |
85 | .start = 0, | |
86 | .count = 1, | |
87 | }; | |
88 | ||
89 | ioctl(vbasedev->fd, VFIO_DEVICE_SET_IRQS, &irq_set); | |
90 | } | |
91 | ||
92 | void vfio_mask_single_irqindex(VFIODevice *vbasedev, int index) | |
93 | { | |
94 | struct vfio_irq_set irq_set = { | |
95 | .argsz = sizeof(irq_set), | |
96 | .flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_MASK, | |
97 | .index = index, | |
98 | .start = 0, | |
99 | .count = 1, | |
100 | }; | |
101 | ||
102 | ioctl(vbasedev->fd, VFIO_DEVICE_SET_IRQS, &irq_set); | |
103 | } | |
104 | ||
201a7331 EA |
105 | static inline const char *action_to_str(int action) |
106 | { | |
107 | switch (action) { | |
108 | case VFIO_IRQ_SET_ACTION_MASK: | |
109 | return "MASK"; | |
110 | case VFIO_IRQ_SET_ACTION_UNMASK: | |
111 | return "UNMASK"; | |
112 | case VFIO_IRQ_SET_ACTION_TRIGGER: | |
113 | return "TRIGGER"; | |
114 | default: | |
115 | return "UNKNOWN ACTION"; | |
116 | } | |
117 | } | |
118 | ||
119 | static const char *index_to_str(VFIODevice *vbasedev, int index) | |
120 | { | |
121 | if (vbasedev->type != VFIO_DEVICE_TYPE_PCI) { | |
122 | return NULL; | |
123 | } | |
124 | ||
125 | switch (index) { | |
126 | case VFIO_PCI_INTX_IRQ_INDEX: | |
127 | return "INTX"; | |
128 | case VFIO_PCI_MSI_IRQ_INDEX: | |
129 | return "MSI"; | |
130 | case VFIO_PCI_MSIX_IRQ_INDEX: | |
131 | return "MSIX"; | |
132 | case VFIO_PCI_ERR_IRQ_INDEX: | |
133 | return "ERR"; | |
134 | case VFIO_PCI_REQ_IRQ_INDEX: | |
135 | return "REQ"; | |
136 | default: | |
137 | return NULL; | |
138 | } | |
139 | } | |
140 | ||
53d1b5fc DH |
141 | static int vfio_ram_block_discard_disable(VFIOContainer *container, bool state) |
142 | { | |
143 | switch (container->iommu_type) { | |
144 | case VFIO_TYPE1v2_IOMMU: | |
145 | case VFIO_TYPE1_IOMMU: | |
146 | /* | |
147 | * We support coordinated discarding of RAM via the RamDiscardManager. | |
148 | */ | |
149 | return ram_block_uncoordinated_discard_disable(state); | |
150 | default: | |
151 | /* | |
152 | * VFIO_SPAPR_TCE_IOMMU most probably works just fine with | |
153 | * RamDiscardManager, however, it is completely untested. | |
154 | * | |
155 | * VFIO_SPAPR_TCE_v2_IOMMU with "DMA memory preregistering" does | |
156 | * completely the opposite of managing mapping/pinning dynamically as | |
157 | * required by RamDiscardManager. We would have to special-case sections | |
158 | * with a RamDiscardManager. | |
159 | */ | |
160 | return ram_block_discard_disable(state); | |
161 | } | |
162 | } | |
163 | ||
201a7331 EA |
164 | int vfio_set_irq_signaling(VFIODevice *vbasedev, int index, int subindex, |
165 | int action, int fd, Error **errp) | |
166 | { | |
167 | struct vfio_irq_set *irq_set; | |
168 | int argsz, ret = 0; | |
169 | const char *name; | |
170 | int32_t *pfd; | |
171 | ||
172 | argsz = sizeof(*irq_set) + sizeof(*pfd); | |
173 | ||
174 | irq_set = g_malloc0(argsz); | |
175 | irq_set->argsz = argsz; | |
176 | irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD | action; | |
177 | irq_set->index = index; | |
178 | irq_set->start = subindex; | |
179 | irq_set->count = 1; | |
180 | pfd = (int32_t *)&irq_set->data; | |
181 | *pfd = fd; | |
182 | ||
183 | if (ioctl(vbasedev->fd, VFIO_DEVICE_SET_IRQS, irq_set)) { | |
184 | ret = -errno; | |
185 | } | |
186 | g_free(irq_set); | |
187 | ||
188 | if (!ret) { | |
189 | return 0; | |
190 | } | |
191 | ||
192 | error_setg_errno(errp, -ret, "VFIO_DEVICE_SET_IRQS failure"); | |
193 | ||
194 | name = index_to_str(vbasedev, index); | |
195 | if (name) { | |
196 | error_prepend(errp, "%s-%d: ", name, subindex); | |
197 | } else { | |
198 | error_prepend(errp, "index %d-%d: ", index, subindex); | |
199 | } | |
200 | error_prepend(errp, | |
201 | "Failed to %s %s eventfd signaling for interrupt ", | |
202 | fd < 0 ? "tear down" : "set up", action_to_str(action)); | |
203 | return ret; | |
204 | } | |
205 | ||
e2c7d025 EA |
206 | /* |
207 | * IO Port/MMIO - Beware of the endians, VFIO is always little endian | |
208 | */ | |
209 | void vfio_region_write(void *opaque, hwaddr addr, | |
210 | uint64_t data, unsigned size) | |
211 | { | |
212 | VFIORegion *region = opaque; | |
213 | VFIODevice *vbasedev = region->vbasedev; | |
214 | union { | |
215 | uint8_t byte; | |
216 | uint16_t word; | |
217 | uint32_t dword; | |
218 | uint64_t qword; | |
219 | } buf; | |
220 | ||
221 | switch (size) { | |
222 | case 1: | |
223 | buf.byte = data; | |
224 | break; | |
225 | case 2: | |
226 | buf.word = cpu_to_le16(data); | |
227 | break; | |
228 | case 4: | |
229 | buf.dword = cpu_to_le32(data); | |
230 | break; | |
38d49e8c JRZ |
231 | case 8: |
232 | buf.qword = cpu_to_le64(data); | |
233 | break; | |
e2c7d025 | 234 | default: |
c624b6b3 | 235 | hw_error("vfio: unsupported write size, %u bytes", size); |
e2c7d025 EA |
236 | break; |
237 | } | |
238 | ||
239 | if (pwrite(vbasedev->fd, &buf, size, region->fd_offset + addr) != size) { | |
240 | error_report("%s(%s:region%d+0x%"HWADDR_PRIx", 0x%"PRIx64 | |
241 | ",%d) failed: %m", | |
242 | __func__, vbasedev->name, region->nr, | |
243 | addr, data, size); | |
244 | } | |
245 | ||
246 | trace_vfio_region_write(vbasedev->name, region->nr, addr, data, size); | |
247 | ||
248 | /* | |
249 | * A read or write to a BAR always signals an INTx EOI. This will | |
250 | * do nothing if not pending (including not in INTx mode). We assume | |
251 | * that a BAR access is in response to an interrupt and that BAR | |
252 | * accesses will service the interrupt. Unfortunately, we don't know | |
253 | * which access will service the interrupt, so we're potentially | |
254 | * getting quite a few host interrupts per guest interrupt. | |
255 | */ | |
256 | vbasedev->ops->vfio_eoi(vbasedev); | |
257 | } | |
258 | ||
259 | uint64_t vfio_region_read(void *opaque, | |
260 | hwaddr addr, unsigned size) | |
261 | { | |
262 | VFIORegion *region = opaque; | |
263 | VFIODevice *vbasedev = region->vbasedev; | |
264 | union { | |
265 | uint8_t byte; | |
266 | uint16_t word; | |
267 | uint32_t dword; | |
268 | uint64_t qword; | |
269 | } buf; | |
270 | uint64_t data = 0; | |
271 | ||
272 | if (pread(vbasedev->fd, &buf, size, region->fd_offset + addr) != size) { | |
273 | error_report("%s(%s:region%d+0x%"HWADDR_PRIx", %d) failed: %m", | |
274 | __func__, vbasedev->name, region->nr, | |
275 | addr, size); | |
276 | return (uint64_t)-1; | |
277 | } | |
278 | switch (size) { | |
279 | case 1: | |
280 | data = buf.byte; | |
281 | break; | |
282 | case 2: | |
283 | data = le16_to_cpu(buf.word); | |
284 | break; | |
285 | case 4: | |
286 | data = le32_to_cpu(buf.dword); | |
287 | break; | |
38d49e8c JRZ |
288 | case 8: |
289 | data = le64_to_cpu(buf.qword); | |
290 | break; | |
e2c7d025 | 291 | default: |
c624b6b3 | 292 | hw_error("vfio: unsupported read size, %u bytes", size); |
e2c7d025 EA |
293 | break; |
294 | } | |
295 | ||
296 | trace_vfio_region_read(vbasedev->name, region->nr, addr, size, data); | |
297 | ||
298 | /* Same as write above */ | |
299 | vbasedev->ops->vfio_eoi(vbasedev); | |
300 | ||
301 | return data; | |
302 | } | |
303 | ||
304 | const MemoryRegionOps vfio_region_ops = { | |
305 | .read = vfio_region_read, | |
306 | .write = vfio_region_write, | |
307 | .endianness = DEVICE_LITTLE_ENDIAN, | |
15126cba JRZ |
308 | .valid = { |
309 | .min_access_size = 1, | |
310 | .max_access_size = 8, | |
311 | }, | |
38d49e8c JRZ |
312 | .impl = { |
313 | .min_access_size = 1, | |
314 | .max_access_size = 8, | |
315 | }, | |
e2c7d025 EA |
316 | }; |
317 | ||
b6dd6504 KW |
318 | /* |
319 | * Device state interfaces | |
320 | */ | |
321 | ||
3710586c KW |
322 | bool vfio_mig_active(void) |
323 | { | |
324 | VFIOGroup *group; | |
325 | VFIODevice *vbasedev; | |
326 | ||
327 | if (QLIST_EMPTY(&vfio_group_list)) { | |
328 | return false; | |
329 | } | |
330 | ||
331 | QLIST_FOREACH(group, &vfio_group_list, next) { | |
332 | QLIST_FOREACH(vbasedev, &group->device_list, next) { | |
333 | if (vbasedev->migration_blocker) { | |
334 | return false; | |
335 | } | |
336 | } | |
337 | } | |
338 | return true; | |
339 | } | |
340 | ||
29d81b71 AH |
341 | static Error *multiple_devices_migration_blocker; |
342 | ||
343 | static unsigned int vfio_migratable_device_num(void) | |
344 | { | |
345 | VFIOGroup *group; | |
346 | VFIODevice *vbasedev; | |
347 | unsigned int device_num = 0; | |
348 | ||
349 | QLIST_FOREACH(group, &vfio_group_list, next) { | |
350 | QLIST_FOREACH(vbasedev, &group->device_list, next) { | |
351 | if (vbasedev->migration) { | |
352 | device_num++; | |
353 | } | |
354 | } | |
355 | } | |
356 | ||
357 | return device_num; | |
358 | } | |
359 | ||
360 | int vfio_block_multiple_devices_migration(Error **errp) | |
361 | { | |
362 | int ret; | |
363 | ||
364 | if (multiple_devices_migration_blocker || | |
365 | vfio_migratable_device_num() <= 1) { | |
366 | return 0; | |
367 | } | |
368 | ||
369 | error_setg(&multiple_devices_migration_blocker, | |
370 | "Migration is currently not supported with multiple " | |
371 | "VFIO devices"); | |
372 | ret = migrate_add_blocker(multiple_devices_migration_blocker, errp); | |
373 | if (ret < 0) { | |
374 | error_free(multiple_devices_migration_blocker); | |
375 | multiple_devices_migration_blocker = NULL; | |
376 | } | |
377 | ||
378 | return ret; | |
379 | } | |
380 | ||
381 | void vfio_unblock_multiple_devices_migration(void) | |
382 | { | |
383 | if (!multiple_devices_migration_blocker || | |
384 | vfio_migratable_device_num() > 1) { | |
385 | return; | |
386 | } | |
387 | ||
388 | migrate_del_blocker(multiple_devices_migration_blocker); | |
389 | error_free(multiple_devices_migration_blocker); | |
390 | multiple_devices_migration_blocker = NULL; | |
391 | } | |
392 | ||
758b96b6 | 393 | static bool vfio_devices_all_dirty_tracking(VFIOContainer *container) |
b6dd6504 KW |
394 | { |
395 | VFIOGroup *group; | |
396 | VFIODevice *vbasedev; | |
397 | MigrationState *ms = migrate_get_current(); | |
398 | ||
399 | if (!migration_is_setup_or_active(ms->state)) { | |
400 | return false; | |
401 | } | |
402 | ||
403 | QLIST_FOREACH(group, &container->group_list, container_next) { | |
404 | QLIST_FOREACH(vbasedev, &group->device_list, next) { | |
405 | VFIOMigration *migration = vbasedev->migration; | |
406 | ||
407 | if (!migration) { | |
408 | return false; | |
409 | } | |
410 | ||
31bcbbb5 AH |
411 | if (!migration->v2 && |
412 | (vbasedev->pre_copy_dirty_page_tracking == ON_OFF_AUTO_OFF) && | |
6eeb2909 | 413 | (migration->device_state_v1 & VFIO_DEVICE_STATE_V1_RUNNING)) { |
b6dd6504 KW |
414 | return false; |
415 | } | |
31bcbbb5 AH |
416 | |
417 | if (migration->v2 && | |
418 | vbasedev->pre_copy_dirty_page_tracking == ON_OFF_AUTO_OFF && | |
419 | migration->device_state == VFIO_DEVICE_STATE_RUNNING) { | |
420 | return false; | |
421 | } | |
b6dd6504 KW |
422 | } |
423 | } | |
424 | return true; | |
425 | } | |
426 | ||
8b942af3 AH |
427 | /* |
428 | * Check if all VFIO devices are running and migration is active, which is | |
429 | * essentially equivalent to the migration being in pre-copy phase. | |
430 | */ | |
431 | static bool vfio_devices_all_running_and_mig_active(VFIOContainer *container) | |
9e7b0442 KW |
432 | { |
433 | VFIOGroup *group; | |
434 | VFIODevice *vbasedev; | |
9e7b0442 | 435 | |
8b942af3 | 436 | if (!migration_is_active(migrate_get_current())) { |
9e7b0442 KW |
437 | return false; |
438 | } | |
439 | ||
440 | QLIST_FOREACH(group, &container->group_list, container_next) { | |
441 | QLIST_FOREACH(vbasedev, &group->device_list, next) { | |
442 | VFIOMigration *migration = vbasedev->migration; | |
443 | ||
444 | if (!migration) { | |
445 | return false; | |
446 | } | |
447 | ||
31bcbbb5 AH |
448 | if (!migration->v2 && |
449 | migration->device_state_v1 & VFIO_DEVICE_STATE_V1_RUNNING) { | |
450 | continue; | |
451 | } | |
452 | ||
453 | if (migration->v2 && | |
454 | migration->device_state == VFIO_DEVICE_STATE_RUNNING) { | |
9e7b0442 KW |
455 | continue; |
456 | } else { | |
457 | return false; | |
458 | } | |
459 | } | |
460 | } | |
461 | return true; | |
462 | } | |
463 | ||
464 | static int vfio_dma_unmap_bitmap(VFIOContainer *container, | |
465 | hwaddr iova, ram_addr_t size, | |
466 | IOMMUTLBEntry *iotlb) | |
467 | { | |
468 | struct vfio_iommu_type1_dma_unmap *unmap; | |
469 | struct vfio_bitmap *bitmap; | |
8e3b0cbb | 470 | uint64_t pages = REAL_HOST_PAGE_ALIGN(size) / qemu_real_host_page_size(); |
9e7b0442 KW |
471 | int ret; |
472 | ||
473 | unmap = g_malloc0(sizeof(*unmap) + sizeof(*bitmap)); | |
474 | ||
475 | unmap->argsz = sizeof(*unmap) + sizeof(*bitmap); | |
476 | unmap->iova = iova; | |
477 | unmap->size = size; | |
478 | unmap->flags |= VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP; | |
479 | bitmap = (struct vfio_bitmap *)&unmap->data; | |
480 | ||
481 | /* | |
1eb7f642 KJ |
482 | * cpu_physical_memory_set_dirty_lebitmap() supports pages in bitmap of |
483 | * qemu_real_host_page_size to mark those dirty. Hence set bitmap_pgsize | |
484 | * to qemu_real_host_page_size. | |
9e7b0442 KW |
485 | */ |
486 | ||
8e3b0cbb | 487 | bitmap->pgsize = qemu_real_host_page_size(); |
9e7b0442 KW |
488 | bitmap->size = ROUND_UP(pages, sizeof(__u64) * BITS_PER_BYTE) / |
489 | BITS_PER_BYTE; | |
490 | ||
491 | if (bitmap->size > container->max_dirty_bitmap_size) { | |
492 | error_report("UNMAP: Size of bitmap too big 0x%"PRIx64, | |
493 | (uint64_t)bitmap->size); | |
494 | ret = -E2BIG; | |
495 | goto unmap_exit; | |
496 | } | |
497 | ||
498 | bitmap->data = g_try_malloc0(bitmap->size); | |
499 | if (!bitmap->data) { | |
500 | ret = -ENOMEM; | |
501 | goto unmap_exit; | |
502 | } | |
503 | ||
504 | ret = ioctl(container->fd, VFIO_IOMMU_UNMAP_DMA, unmap); | |
505 | if (!ret) { | |
506 | cpu_physical_memory_set_dirty_lebitmap((unsigned long *)bitmap->data, | |
507 | iotlb->translated_addr, pages); | |
508 | } else { | |
509 | error_report("VFIO_UNMAP_DMA with DIRTY_BITMAP : %m"); | |
510 | } | |
511 | ||
512 | g_free(bitmap->data); | |
513 | unmap_exit: | |
514 | g_free(unmap); | |
515 | return ret; | |
516 | } | |
517 | ||
e2c7d025 EA |
518 | /* |
519 | * DMA - Mapping and unmapping for the "type1" IOMMU interface used on x86 | |
520 | */ | |
521 | static int vfio_dma_unmap(VFIOContainer *container, | |
9e7b0442 KW |
522 | hwaddr iova, ram_addr_t size, |
523 | IOMMUTLBEntry *iotlb) | |
e2c7d025 EA |
524 | { |
525 | struct vfio_iommu_type1_dma_unmap unmap = { | |
526 | .argsz = sizeof(unmap), | |
527 | .flags = 0, | |
528 | .iova = iova, | |
529 | .size = size, | |
530 | }; | |
531 | ||
9e7b0442 | 532 | if (iotlb && container->dirty_pages_supported && |
8b942af3 | 533 | vfio_devices_all_running_and_mig_active(container)) { |
9e7b0442 KW |
534 | return vfio_dma_unmap_bitmap(container, iova, size, iotlb); |
535 | } | |
536 | ||
567d7d3e AW |
537 | while (ioctl(container->fd, VFIO_IOMMU_UNMAP_DMA, &unmap)) { |
538 | /* | |
539 | * The type1 backend has an off-by-one bug in the kernel (71a7d3d78e3c | |
540 | * v4.15) where an overflow in its wrap-around check prevents us from | |
541 | * unmapping the last page of the address space. Test for the error | |
542 | * condition and re-try the unmap excluding the last page. The | |
543 | * expectation is that we've never mapped the last page anyway and this | |
544 | * unmap request comes via vIOMMU support which also makes it unlikely | |
545 | * that this page is used. This bug was introduced well after type1 v2 | |
546 | * support was introduced, so we shouldn't need to test for v1. A fix | |
547 | * is queued for kernel v5.0 so this workaround can be removed once | |
548 | * affected kernels are sufficiently deprecated. | |
549 | */ | |
550 | if (errno == EINVAL && unmap.size && !(unmap.iova + unmap.size) && | |
551 | container->iommu_type == VFIO_TYPE1v2_IOMMU) { | |
552 | trace_vfio_dma_unmap_overflow_workaround(); | |
553 | unmap.size -= 1ULL << ctz64(container->pgsizes); | |
554 | continue; | |
555 | } | |
b09d51c9 | 556 | error_report("VFIO_UNMAP_DMA failed: %s", strerror(errno)); |
e2c7d025 EA |
557 | return -errno; |
558 | } | |
559 | ||
8b942af3 | 560 | if (iotlb && vfio_devices_all_running_and_mig_active(container)) { |
b051a3f6 AH |
561 | cpu_physical_memory_set_dirty_range(iotlb->translated_addr, size, |
562 | tcg_enabled() ? DIRTY_CLIENTS_ALL : | |
563 | DIRTY_CLIENTS_NOCODE); | |
564 | } | |
565 | ||
e2c7d025 EA |
566 | return 0; |
567 | } | |
568 | ||
569 | static int vfio_dma_map(VFIOContainer *container, hwaddr iova, | |
570 | ram_addr_t size, void *vaddr, bool readonly) | |
571 | { | |
572 | struct vfio_iommu_type1_dma_map map = { | |
573 | .argsz = sizeof(map), | |
574 | .flags = VFIO_DMA_MAP_FLAG_READ, | |
575 | .vaddr = (__u64)(uintptr_t)vaddr, | |
576 | .iova = iova, | |
577 | .size = size, | |
578 | }; | |
579 | ||
580 | if (!readonly) { | |
581 | map.flags |= VFIO_DMA_MAP_FLAG_WRITE; | |
582 | } | |
583 | ||
584 | /* | |
585 | * Try the mapping, if it fails with EBUSY, unmap the region and try | |
586 | * again. This shouldn't be necessary, but we sometimes see it in | |
b6af0975 | 587 | * the VGA ROM space. |
e2c7d025 EA |
588 | */ |
589 | if (ioctl(container->fd, VFIO_IOMMU_MAP_DMA, &map) == 0 || | |
9e7b0442 | 590 | (errno == EBUSY && vfio_dma_unmap(container, iova, size, NULL) == 0 && |
e2c7d025 EA |
591 | ioctl(container->fd, VFIO_IOMMU_MAP_DMA, &map) == 0)) { |
592 | return 0; | |
593 | } | |
594 | ||
b09d51c9 | 595 | error_report("VFIO_MAP_DMA failed: %s", strerror(errno)); |
e2c7d025 EA |
596 | return -errno; |
597 | } | |
598 | ||
f4ec5e26 AK |
599 | static void vfio_host_win_add(VFIOContainer *container, |
600 | hwaddr min_iova, hwaddr max_iova, | |
601 | uint64_t iova_pgsizes) | |
602 | { | |
603 | VFIOHostDMAWindow *hostwin; | |
604 | ||
605 | QLIST_FOREACH(hostwin, &container->hostwin_list, hostwin_next) { | |
606 | if (ranges_overlap(hostwin->min_iova, | |
607 | hostwin->max_iova - hostwin->min_iova + 1, | |
608 | min_iova, | |
609 | max_iova - min_iova + 1)) { | |
610 | hw_error("%s: Overlapped IOMMU are not enabled", __func__); | |
611 | } | |
612 | } | |
613 | ||
614 | hostwin = g_malloc0(sizeof(*hostwin)); | |
615 | ||
616 | hostwin->min_iova = min_iova; | |
617 | hostwin->max_iova = max_iova; | |
618 | hostwin->iova_pgsizes = iova_pgsizes; | |
619 | QLIST_INSERT_HEAD(&container->hostwin_list, hostwin, hostwin_next); | |
620 | } | |
621 | ||
2e4109de AK |
622 | static int vfio_host_win_del(VFIOContainer *container, hwaddr min_iova, |
623 | hwaddr max_iova) | |
624 | { | |
625 | VFIOHostDMAWindow *hostwin; | |
626 | ||
627 | QLIST_FOREACH(hostwin, &container->hostwin_list, hostwin_next) { | |
628 | if (hostwin->min_iova == min_iova && hostwin->max_iova == max_iova) { | |
629 | QLIST_REMOVE(hostwin, hostwin_next); | |
f3bc3a73 | 630 | g_free(hostwin); |
2e4109de AK |
631 | return 0; |
632 | } | |
633 | } | |
634 | ||
635 | return -1; | |
636 | } | |
637 | ||
e2c7d025 EA |
638 | static bool vfio_listener_skipped_section(MemoryRegionSection *section) |
639 | { | |
640 | return (!memory_region_is_ram(section->mr) && | |
641 | !memory_region_is_iommu(section->mr)) || | |
56918a12 | 642 | memory_region_is_protected(section->mr) || |
e2c7d025 EA |
643 | /* |
644 | * Sizing an enabled 64-bit BAR can cause spurious mappings to | |
645 | * addresses in the upper part of the 64-bit address space. These | |
646 | * are never accessed by the CPU and beyond the address width of | |
647 | * some IOMMU hardware. TODO: VFIO should tell us the IOMMU width. | |
648 | */ | |
649 | section->offset_within_address_space & (1ULL << 63); | |
650 | } | |
651 | ||
4a4b88fb | 652 | /* Called with rcu_read_lock held. */ |
9a04fe09 KW |
653 | static bool vfio_get_xlat_addr(IOMMUTLBEntry *iotlb, void **vaddr, |
654 | ram_addr_t *ram_addr, bool *read_only) | |
e2c7d025 | 655 | { |
baa44bce | 656 | bool ret, mr_has_discard_manager; |
0fd7616e | 657 | |
baa44bce CL |
658 | ret = memory_get_xlat_addr(iotlb, vaddr, ram_addr, read_only, |
659 | &mr_has_discard_manager); | |
660 | if (ret && mr_has_discard_manager) { | |
0fd7616e DH |
661 | /* |
662 | * Malicious VMs might trigger discarding of IOMMU-mapped memory. The | |
663 | * pages will remain pinned inside vfio until unmapped, resulting in a | |
664 | * higher memory consumption than expected. If memory would get | |
665 | * populated again later, there would be an inconsistency between pages | |
666 | * pinned by vfio and pages seen by QEMU. This is the case until | |
667 | * unmapped from the IOMMU (e.g., during device reset). | |
668 | * | |
669 | * With malicious guests, we really only care about pinning more memory | |
670 | * than expected. RLIMIT_MEMLOCK set for the user/process can never be | |
671 | * exceeded and can be used to mitigate this problem. | |
672 | */ | |
673 | warn_report_once("Using vfio with vIOMMUs and coordinated discarding of" | |
674 | " RAM (e.g., virtio-mem) works, however, malicious" | |
675 | " guests can trigger pinning of more memory than" | |
676 | " intended via an IOMMU. It's possible to mitigate " | |
677 | " by setting/adjusting RLIMIT_MEMLOCK."); | |
e2c7d025 | 678 | } |
baa44bce | 679 | return ret; |
4a4b88fb PX |
680 | } |
681 | ||
682 | static void vfio_iommu_map_notify(IOMMUNotifier *n, IOMMUTLBEntry *iotlb) | |
683 | { | |
684 | VFIOGuestIOMMU *giommu = container_of(n, VFIOGuestIOMMU, n); | |
685 | VFIOContainer *container = giommu->container; | |
686 | hwaddr iova = iotlb->iova + giommu->iommu_offset; | |
4a4b88fb PX |
687 | void *vaddr; |
688 | int ret; | |
689 | ||
690 | trace_vfio_iommu_map_notify(iotlb->perm == IOMMU_NONE ? "UNMAP" : "MAP", | |
691 | iova, iova + iotlb->addr_mask); | |
692 | ||
693 | if (iotlb->target_as != &address_space_memory) { | |
694 | error_report("Wrong target AS \"%s\", only system memory is allowed", | |
695 | iotlb->target_as->name ? iotlb->target_as->name : "none"); | |
696 | return; | |
697 | } | |
698 | ||
699 | rcu_read_lock(); | |
700 | ||
e2c7d025 | 701 | if ((iotlb->perm & IOMMU_RW) != IOMMU_NONE) { |
9a04fe09 KW |
702 | bool read_only; |
703 | ||
704 | if (!vfio_get_xlat_addr(iotlb, &vaddr, NULL, &read_only)) { | |
dfbd90e5 PX |
705 | goto out; |
706 | } | |
4a4b88fb PX |
707 | /* |
708 | * vaddr is only valid until rcu_read_unlock(). But after | |
709 | * vfio_dma_map has set up the mapping the pages will be | |
710 | * pinned by the kernel. This makes sure that the RAM backend | |
711 | * of vaddr will always be there, even if the memory object is | |
712 | * destroyed and its backing memory munmap-ed. | |
713 | */ | |
d78c19b5 | 714 | ret = vfio_dma_map(container, iova, |
e2c7d025 | 715 | iotlb->addr_mask + 1, vaddr, |
4a4b88fb | 716 | read_only); |
e2c7d025 EA |
717 | if (ret) { |
718 | error_report("vfio_dma_map(%p, 0x%"HWADDR_PRIx", " | |
719 | "0x%"HWADDR_PRIx", %p) = %d (%m)", | |
d78c19b5 | 720 | container, iova, |
e2c7d025 EA |
721 | iotlb->addr_mask + 1, vaddr, ret); |
722 | } | |
723 | } else { | |
9e7b0442 | 724 | ret = vfio_dma_unmap(container, iova, iotlb->addr_mask + 1, iotlb); |
e2c7d025 EA |
725 | if (ret) { |
726 | error_report("vfio_dma_unmap(%p, 0x%"HWADDR_PRIx", " | |
727 | "0x%"HWADDR_PRIx") = %d (%m)", | |
d78c19b5 | 728 | container, iova, |
e2c7d025 EA |
729 | iotlb->addr_mask + 1, ret); |
730 | } | |
731 | } | |
41063e1e PB |
732 | out: |
733 | rcu_read_unlock(); | |
e2c7d025 EA |
734 | } |
735 | ||
5e3b981c DH |
736 | static void vfio_ram_discard_notify_discard(RamDiscardListener *rdl, |
737 | MemoryRegionSection *section) | |
738 | { | |
739 | VFIORamDiscardListener *vrdl = container_of(rdl, VFIORamDiscardListener, | |
740 | listener); | |
741 | const hwaddr size = int128_get64(section->size); | |
742 | const hwaddr iova = section->offset_within_address_space; | |
743 | int ret; | |
744 | ||
745 | /* Unmap with a single call. */ | |
746 | ret = vfio_dma_unmap(vrdl->container, iova, size , NULL); | |
747 | if (ret) { | |
748 | error_report("%s: vfio_dma_unmap() failed: %s", __func__, | |
749 | strerror(-ret)); | |
750 | } | |
751 | } | |
752 | ||
753 | static int vfio_ram_discard_notify_populate(RamDiscardListener *rdl, | |
754 | MemoryRegionSection *section) | |
755 | { | |
756 | VFIORamDiscardListener *vrdl = container_of(rdl, VFIORamDiscardListener, | |
757 | listener); | |
758 | const hwaddr end = section->offset_within_region + | |
759 | int128_get64(section->size); | |
760 | hwaddr start, next, iova; | |
761 | void *vaddr; | |
762 | int ret; | |
763 | ||
764 | /* | |
765 | * Map in (aligned within memory region) minimum granularity, so we can | |
766 | * unmap in minimum granularity later. | |
767 | */ | |
768 | for (start = section->offset_within_region; start < end; start = next) { | |
769 | next = ROUND_UP(start + 1, vrdl->granularity); | |
770 | next = MIN(next, end); | |
771 | ||
772 | iova = start - section->offset_within_region + | |
773 | section->offset_within_address_space; | |
774 | vaddr = memory_region_get_ram_ptr(section->mr) + start; | |
775 | ||
776 | ret = vfio_dma_map(vrdl->container, iova, next - start, | |
777 | vaddr, section->readonly); | |
778 | if (ret) { | |
779 | /* Rollback */ | |
780 | vfio_ram_discard_notify_discard(rdl, section); | |
781 | return ret; | |
782 | } | |
783 | } | |
784 | return 0; | |
785 | } | |
786 | ||
787 | static void vfio_register_ram_discard_listener(VFIOContainer *container, | |
788 | MemoryRegionSection *section) | |
789 | { | |
790 | RamDiscardManager *rdm = memory_region_get_ram_discard_manager(section->mr); | |
791 | VFIORamDiscardListener *vrdl; | |
792 | ||
793 | /* Ignore some corner cases not relevant in practice. */ | |
794 | g_assert(QEMU_IS_ALIGNED(section->offset_within_region, TARGET_PAGE_SIZE)); | |
795 | g_assert(QEMU_IS_ALIGNED(section->offset_within_address_space, | |
796 | TARGET_PAGE_SIZE)); | |
797 | g_assert(QEMU_IS_ALIGNED(int128_get64(section->size), TARGET_PAGE_SIZE)); | |
798 | ||
799 | vrdl = g_new0(VFIORamDiscardListener, 1); | |
800 | vrdl->container = container; | |
801 | vrdl->mr = section->mr; | |
802 | vrdl->offset_within_address_space = section->offset_within_address_space; | |
803 | vrdl->size = int128_get64(section->size); | |
804 | vrdl->granularity = ram_discard_manager_get_min_granularity(rdm, | |
805 | section->mr); | |
806 | ||
807 | g_assert(vrdl->granularity && is_power_of_2(vrdl->granularity)); | |
a5dba9bc DH |
808 | g_assert(container->pgsizes && |
809 | vrdl->granularity >= 1ULL << ctz64(container->pgsizes)); | |
5e3b981c DH |
810 | |
811 | ram_discard_listener_init(&vrdl->listener, | |
812 | vfio_ram_discard_notify_populate, | |
813 | vfio_ram_discard_notify_discard, true); | |
814 | ram_discard_manager_register_listener(rdm, &vrdl->listener, section); | |
815 | QLIST_INSERT_HEAD(&container->vrdl_list, vrdl, next); | |
a74317f6 DH |
816 | |
817 | /* | |
818 | * Sanity-check if we have a theoretically problematic setup where we could | |
819 | * exceed the maximum number of possible DMA mappings over time. We assume | |
820 | * that each mapped section in the same address space as a RamDiscardManager | |
821 | * section consumes exactly one DMA mapping, with the exception of | |
822 | * RamDiscardManager sections; i.e., we don't expect to have gIOMMU sections | |
823 | * in the same address space as RamDiscardManager sections. | |
824 | * | |
825 | * We assume that each section in the address space consumes one memslot. | |
826 | * We take the number of KVM memory slots as a best guess for the maximum | |
827 | * number of sections in the address space we could have over time, | |
828 | * also consuming DMA mappings. | |
829 | */ | |
830 | if (container->dma_max_mappings) { | |
831 | unsigned int vrdl_count = 0, vrdl_mappings = 0, max_memslots = 512; | |
832 | ||
833 | #ifdef CONFIG_KVM | |
834 | if (kvm_enabled()) { | |
835 | max_memslots = kvm_get_max_memslots(); | |
836 | } | |
837 | #endif | |
838 | ||
839 | QLIST_FOREACH(vrdl, &container->vrdl_list, next) { | |
840 | hwaddr start, end; | |
841 | ||
842 | start = QEMU_ALIGN_DOWN(vrdl->offset_within_address_space, | |
843 | vrdl->granularity); | |
844 | end = ROUND_UP(vrdl->offset_within_address_space + vrdl->size, | |
845 | vrdl->granularity); | |
846 | vrdl_mappings += (end - start) / vrdl->granularity; | |
847 | vrdl_count++; | |
848 | } | |
849 | ||
850 | if (vrdl_mappings + max_memslots - vrdl_count > | |
851 | container->dma_max_mappings) { | |
852 | warn_report("%s: possibly running out of DMA mappings. E.g., try" | |
853 | " increasing the 'block-size' of virtio-mem devies." | |
854 | " Maximum possible DMA mappings: %d, Maximum possible" | |
855 | " memslots: %d", __func__, container->dma_max_mappings, | |
856 | max_memslots); | |
857 | } | |
858 | } | |
5e3b981c DH |
859 | } |
860 | ||
861 | static void vfio_unregister_ram_discard_listener(VFIOContainer *container, | |
862 | MemoryRegionSection *section) | |
863 | { | |
864 | RamDiscardManager *rdm = memory_region_get_ram_discard_manager(section->mr); | |
865 | VFIORamDiscardListener *vrdl = NULL; | |
866 | ||
867 | QLIST_FOREACH(vrdl, &container->vrdl_list, next) { | |
868 | if (vrdl->mr == section->mr && | |
869 | vrdl->offset_within_address_space == | |
870 | section->offset_within_address_space) { | |
871 | break; | |
872 | } | |
873 | } | |
874 | ||
875 | if (!vrdl) { | |
876 | hw_error("vfio: Trying to unregister missing RAM discard listener"); | |
877 | } | |
878 | ||
879 | ram_discard_manager_unregister_listener(rdm, &vrdl->listener); | |
880 | QLIST_REMOVE(vrdl, next); | |
881 | g_free(vrdl); | |
882 | } | |
883 | ||
851d6d1a EA |
884 | static bool vfio_known_safe_misalignment(MemoryRegionSection *section) |
885 | { | |
886 | MemoryRegion *mr = section->mr; | |
887 | ||
888 | if (!TPM_IS_CRB(mr->owner)) { | |
889 | return false; | |
890 | } | |
891 | ||
892 | /* this is a known safe misaligned region, just trace for debug purpose */ | |
893 | trace_vfio_known_safe_misalignment(memory_region_name(mr), | |
894 | section->offset_within_address_space, | |
895 | section->offset_within_region, | |
896 | qemu_real_host_page_size()); | |
897 | return true; | |
898 | } | |
899 | ||
e2c7d025 EA |
900 | static void vfio_listener_region_add(MemoryListener *listener, |
901 | MemoryRegionSection *section) | |
902 | { | |
ee0bf0e5 | 903 | VFIOContainer *container = container_of(listener, VFIOContainer, listener); |
e2c7d025 | 904 | hwaddr iova, end; |
55efcc53 | 905 | Int128 llend, llsize; |
e2c7d025 EA |
906 | void *vaddr; |
907 | int ret; | |
f4ec5e26 AK |
908 | VFIOHostDMAWindow *hostwin; |
909 | bool hostwin_found; | |
d7d87836 | 910 | Error *err = NULL; |
e2c7d025 EA |
911 | |
912 | if (vfio_listener_skipped_section(section)) { | |
913 | trace_vfio_listener_region_add_skip( | |
914 | section->offset_within_address_space, | |
915 | section->offset_within_address_space + | |
916 | int128_get64(int128_sub(section->size, int128_one()))); | |
917 | return; | |
918 | } | |
919 | ||
1eb7f642 | 920 | if (unlikely((section->offset_within_address_space & |
8e3b0cbb MAL |
921 | ~qemu_real_host_page_mask()) != |
922 | (section->offset_within_region & ~qemu_real_host_page_mask()))) { | |
851d6d1a EA |
923 | if (!vfio_known_safe_misalignment(section)) { |
924 | error_report("%s received unaligned region %s iova=0x%"PRIx64 | |
925 | " offset_within_region=0x%"PRIx64 | |
926 | " qemu_real_host_page_size=0x%"PRIxPTR, | |
927 | __func__, memory_region_name(section->mr), | |
928 | section->offset_within_address_space, | |
929 | section->offset_within_region, | |
930 | qemu_real_host_page_size()); | |
931 | } | |
e2c7d025 EA |
932 | return; |
933 | } | |
934 | ||
1eb7f642 | 935 | iova = REAL_HOST_PAGE_ALIGN(section->offset_within_address_space); |
e2c7d025 EA |
936 | llend = int128_make64(section->offset_within_address_space); |
937 | llend = int128_add(llend, section->size); | |
8e3b0cbb | 938 | llend = int128_and(llend, int128_exts64(qemu_real_host_page_mask())); |
e2c7d025 EA |
939 | |
940 | if (int128_ge(int128_make64(iova), llend)) { | |
e4b34708 KJ |
941 | if (memory_region_is_ram_device(section->mr)) { |
942 | trace_vfio_listener_region_add_no_dma_map( | |
943 | memory_region_name(section->mr), | |
944 | section->offset_within_address_space, | |
945 | int128_getlo(section->size), | |
8e3b0cbb | 946 | qemu_real_host_page_size()); |
e4b34708 | 947 | } |
e2c7d025 EA |
948 | return; |
949 | } | |
55efcc53 | 950 | end = int128_get64(int128_sub(llend, int128_one())); |
3898aad3 | 951 | |
2e4109de | 952 | if (container->iommu_type == VFIO_SPAPR_TCE_v2_IOMMU) { |
2e4109de AK |
953 | hwaddr pgsize = 0; |
954 | ||
955 | /* For now intersections are not allowed, we may relax this later */ | |
956 | QLIST_FOREACH(hostwin, &container->hostwin_list, hostwin_next) { | |
957 | if (ranges_overlap(hostwin->min_iova, | |
958 | hostwin->max_iova - hostwin->min_iova + 1, | |
959 | section->offset_within_address_space, | |
960 | int128_get64(section->size))) { | |
d7d87836 EA |
961 | error_setg(&err, |
962 | "region [0x%"PRIx64",0x%"PRIx64"] overlaps with existing" | |
963 | "host DMA window [0x%"PRIx64",0x%"PRIx64"]", | |
964 | section->offset_within_address_space, | |
965 | section->offset_within_address_space + | |
966 | int128_get64(section->size) - 1, | |
967 | hostwin->min_iova, hostwin->max_iova); | |
2e4109de AK |
968 | goto fail; |
969 | } | |
970 | } | |
971 | ||
972 | ret = vfio_spapr_create_window(container, section, &pgsize); | |
973 | if (ret) { | |
d7d87836 | 974 | error_setg_errno(&err, -ret, "Failed to create SPAPR window"); |
2e4109de AK |
975 | goto fail; |
976 | } | |
977 | ||
978 | vfio_host_win_add(container, section->offset_within_address_space, | |
979 | section->offset_within_address_space + | |
980 | int128_get64(section->size) - 1, pgsize); | |
07bc681a AK |
981 | #ifdef CONFIG_KVM |
982 | if (kvm_enabled()) { | |
983 | VFIOGroup *group; | |
984 | IOMMUMemoryRegion *iommu_mr = IOMMU_MEMORY_REGION(section->mr); | |
985 | struct kvm_vfio_spapr_tce param; | |
986 | struct kvm_device_attr attr = { | |
987 | .group = KVM_DEV_VFIO_GROUP, | |
988 | .attr = KVM_DEV_VFIO_GROUP_SET_SPAPR_TCE, | |
989 | .addr = (uint64_t)(unsigned long)¶m, | |
990 | }; | |
991 | ||
992 | if (!memory_region_iommu_get_attr(iommu_mr, IOMMU_ATTR_SPAPR_TCE_FD, | |
993 | ¶m.tablefd)) { | |
994 | QLIST_FOREACH(group, &container->group_list, container_next) { | |
995 | param.groupfd = group->fd; | |
996 | if (ioctl(vfio_kvm_device_fd, KVM_SET_DEVICE_ATTR, &attr)) { | |
997 | error_report("vfio: failed to setup fd %d " | |
998 | "for a group with fd %d: %s", | |
999 | param.tablefd, param.groupfd, | |
1000 | strerror(errno)); | |
1001 | return; | |
1002 | } | |
1003 | trace_vfio_spapr_group_attach(param.groupfd, param.tablefd); | |
1004 | } | |
1005 | } | |
1006 | } | |
1007 | #endif | |
2e4109de AK |
1008 | } |
1009 | ||
f4ec5e26 AK |
1010 | hostwin_found = false; |
1011 | QLIST_FOREACH(hostwin, &container->hostwin_list, hostwin_next) { | |
1012 | if (hostwin->min_iova <= iova && end <= hostwin->max_iova) { | |
1013 | hostwin_found = true; | |
1014 | break; | |
1015 | } | |
1016 | } | |
1017 | ||
1018 | if (!hostwin_found) { | |
d7d87836 EA |
1019 | error_setg(&err, "Container %p can't map guest IOVA region" |
1020 | " 0x%"HWADDR_PRIx"..0x%"HWADDR_PRIx, container, iova, end); | |
3898aad3 DG |
1021 | goto fail; |
1022 | } | |
e2c7d025 EA |
1023 | |
1024 | memory_region_ref(section->mr); | |
1025 | ||
1026 | if (memory_region_is_iommu(section->mr)) { | |
1027 | VFIOGuestIOMMU *giommu; | |
3df9d748 | 1028 | IOMMUMemoryRegion *iommu_mr = IOMMU_MEMORY_REGION(section->mr); |
cb1efcf4 | 1029 | int iommu_idx; |
e2c7d025 | 1030 | |
55efcc53 | 1031 | trace_vfio_listener_region_add_iommu(iova, end); |
e2c7d025 | 1032 | /* |
e2c7d025 EA |
1033 | * FIXME: For VFIO iommu types which have KVM acceleration to |
1034 | * avoid bouncing all map/unmaps through qemu this way, this | |
1035 | * would be the right place to wire that up (tell the KVM | |
1036 | * device emulation the VFIO iommu handles to use). | |
1037 | */ | |
e2c7d025 | 1038 | giommu = g_malloc0(sizeof(*giommu)); |
44ee6aaa | 1039 | giommu->iommu_mr = iommu_mr; |
d78c19b5 AK |
1040 | giommu->iommu_offset = section->offset_within_address_space - |
1041 | section->offset_within_region; | |
e2c7d025 | 1042 | giommu->container = container; |
698feb5e PX |
1043 | llend = int128_add(int128_make64(section->offset_within_region), |
1044 | section->size); | |
1045 | llend = int128_sub(llend, int128_one()); | |
cb1efcf4 PM |
1046 | iommu_idx = memory_region_iommu_attrs_to_index(iommu_mr, |
1047 | MEMTXATTRS_UNSPECIFIED); | |
698feb5e | 1048 | iommu_notifier_init(&giommu->n, vfio_iommu_map_notify, |
8dca037b | 1049 | IOMMU_NOTIFIER_IOTLB_EVENTS, |
698feb5e | 1050 | section->offset_within_region, |
cb1efcf4 PM |
1051 | int128_get64(llend), |
1052 | iommu_idx); | |
508ce5eb | 1053 | |
44ee6aaa | 1054 | ret = memory_region_iommu_set_page_size_mask(giommu->iommu_mr, |
b9177498 BB |
1055 | container->pgsizes, |
1056 | &err); | |
1057 | if (ret) { | |
1058 | g_free(giommu); | |
1059 | goto fail; | |
1060 | } | |
1061 | ||
549d4005 EA |
1062 | ret = memory_region_register_iommu_notifier(section->mr, &giommu->n, |
1063 | &err); | |
1064 | if (ret) { | |
1065 | g_free(giommu); | |
1066 | goto fail; | |
1067 | } | |
1068 | QLIST_INSERT_HEAD(&container->giommu_list, giommu, giommu_next); | |
44ee6aaa | 1069 | memory_region_iommu_replay(giommu->iommu_mr, &giommu->n); |
e2c7d025 EA |
1070 | |
1071 | return; | |
1072 | } | |
1073 | ||
1074 | /* Here we assume that memory_region_is_ram(section->mr)==true */ | |
1075 | ||
5e3b981c DH |
1076 | /* |
1077 | * For RAM memory regions with a RamDiscardManager, we only want to map the | |
1078 | * actually populated parts - and update the mapping whenever we're notified | |
1079 | * about changes. | |
1080 | */ | |
1081 | if (memory_region_has_ram_discard_manager(section->mr)) { | |
1082 | vfio_register_ram_discard_listener(container, section); | |
1083 | return; | |
1084 | } | |
1085 | ||
e2c7d025 EA |
1086 | vaddr = memory_region_get_ram_ptr(section->mr) + |
1087 | section->offset_within_region + | |
1088 | (iova - section->offset_within_address_space); | |
1089 | ||
55efcc53 | 1090 | trace_vfio_listener_region_add_ram(iova, end, vaddr); |
e2c7d025 | 1091 | |
55efcc53 BD |
1092 | llsize = int128_sub(llend, int128_make64(iova)); |
1093 | ||
567b5b30 AK |
1094 | if (memory_region_is_ram_device(section->mr)) { |
1095 | hwaddr pgmask = (1ULL << ctz64(hostwin->iova_pgsizes)) - 1; | |
1096 | ||
1097 | if ((iova & pgmask) || (int128_get64(llsize) & pgmask)) { | |
5c086005 EA |
1098 | trace_vfio_listener_region_add_no_dma_map( |
1099 | memory_region_name(section->mr), | |
1100 | section->offset_within_address_space, | |
1101 | int128_getlo(section->size), | |
1102 | pgmask + 1); | |
567b5b30 AK |
1103 | return; |
1104 | } | |
1105 | } | |
1106 | ||
55efcc53 BD |
1107 | ret = vfio_dma_map(container, iova, int128_get64(llsize), |
1108 | vaddr, section->readonly); | |
e2c7d025 | 1109 | if (ret) { |
d7d87836 EA |
1110 | error_setg(&err, "vfio_dma_map(%p, 0x%"HWADDR_PRIx", " |
1111 | "0x%"HWADDR_PRIx", %p) = %d (%m)", | |
1112 | container, iova, int128_get64(llsize), vaddr, ret); | |
567b5b30 AK |
1113 | if (memory_region_is_ram_device(section->mr)) { |
1114 | /* Allow unexpected mappings not to be fatal for RAM devices */ | |
d7d87836 | 1115 | error_report_err(err); |
567b5b30 AK |
1116 | return; |
1117 | } | |
ac6dc389 DG |
1118 | goto fail; |
1119 | } | |
e2c7d025 | 1120 | |
ac6dc389 DG |
1121 | return; |
1122 | ||
1123 | fail: | |
567b5b30 AK |
1124 | if (memory_region_is_ram_device(section->mr)) { |
1125 | error_report("failed to vfio_dma_map. pci p2p may not work"); | |
1126 | return; | |
1127 | } | |
ac6dc389 DG |
1128 | /* |
1129 | * On the initfn path, store the first error in the container so we | |
1130 | * can gracefully fail. Runtime, there's not much we can do other | |
1131 | * than throw a hardware error. | |
1132 | */ | |
1133 | if (!container->initialized) { | |
1134 | if (!container->error) { | |
d7d87836 EA |
1135 | error_propagate_prepend(&container->error, err, |
1136 | "Region %s: ", | |
1137 | memory_region_name(section->mr)); | |
1138 | } else { | |
1139 | error_free(err); | |
e2c7d025 | 1140 | } |
ac6dc389 | 1141 | } else { |
d7d87836 | 1142 | error_report_err(err); |
ac6dc389 | 1143 | hw_error("vfio: DMA mapping failed, unable to continue"); |
e2c7d025 EA |
1144 | } |
1145 | } | |
1146 | ||
1147 | static void vfio_listener_region_del(MemoryListener *listener, | |
1148 | MemoryRegionSection *section) | |
1149 | { | |
ee0bf0e5 | 1150 | VFIOContainer *container = container_of(listener, VFIOContainer, listener); |
e2c7d025 | 1151 | hwaddr iova, end; |
7a057b4f | 1152 | Int128 llend, llsize; |
e2c7d025 | 1153 | int ret; |
567b5b30 | 1154 | bool try_unmap = true; |
e2c7d025 EA |
1155 | |
1156 | if (vfio_listener_skipped_section(section)) { | |
1157 | trace_vfio_listener_region_del_skip( | |
1158 | section->offset_within_address_space, | |
1159 | section->offset_within_address_space + | |
1160 | int128_get64(int128_sub(section->size, int128_one()))); | |
1161 | return; | |
1162 | } | |
1163 | ||
1eb7f642 | 1164 | if (unlikely((section->offset_within_address_space & |
8e3b0cbb MAL |
1165 | ~qemu_real_host_page_mask()) != |
1166 | (section->offset_within_region & ~qemu_real_host_page_mask()))) { | |
ec6600be EA |
1167 | if (!vfio_known_safe_misalignment(section)) { |
1168 | error_report("%s received unaligned region %s iova=0x%"PRIx64 | |
1169 | " offset_within_region=0x%"PRIx64 | |
1170 | " qemu_real_host_page_size=0x%"PRIxPTR, | |
1171 | __func__, memory_region_name(section->mr), | |
1172 | section->offset_within_address_space, | |
1173 | section->offset_within_region, | |
1174 | qemu_real_host_page_size()); | |
1175 | } | |
e2c7d025 EA |
1176 | return; |
1177 | } | |
1178 | ||
1179 | if (memory_region_is_iommu(section->mr)) { | |
1180 | VFIOGuestIOMMU *giommu; | |
1181 | ||
1182 | QLIST_FOREACH(giommu, &container->giommu_list, giommu_next) { | |
44ee6aaa | 1183 | if (MEMORY_REGION(giommu->iommu_mr) == section->mr && |
698feb5e | 1184 | giommu->n.start == section->offset_within_region) { |
3df9d748 | 1185 | memory_region_unregister_iommu_notifier(section->mr, |
d22d8956 | 1186 | &giommu->n); |
e2c7d025 EA |
1187 | QLIST_REMOVE(giommu, giommu_next); |
1188 | g_free(giommu); | |
1189 | break; | |
1190 | } | |
1191 | } | |
1192 | ||
1193 | /* | |
1194 | * FIXME: We assume the one big unmap below is adequate to | |
1195 | * remove any individual page mappings in the IOMMU which | |
1196 | * might have been copied into VFIO. This works for a page table | |
1197 | * based IOMMU where a big unmap flattens a large range of IO-PTEs. | |
1198 | * That may not be true for all IOMMU types. | |
1199 | */ | |
1200 | } | |
1201 | ||
1eb7f642 | 1202 | iova = REAL_HOST_PAGE_ALIGN(section->offset_within_address_space); |
7a057b4f AK |
1203 | llend = int128_make64(section->offset_within_address_space); |
1204 | llend = int128_add(llend, section->size); | |
8e3b0cbb | 1205 | llend = int128_and(llend, int128_exts64(qemu_real_host_page_mask())); |
e2c7d025 | 1206 | |
7a057b4f | 1207 | if (int128_ge(int128_make64(iova), llend)) { |
e2c7d025 EA |
1208 | return; |
1209 | } | |
7a057b4f AK |
1210 | end = int128_get64(int128_sub(llend, int128_one())); |
1211 | ||
1212 | llsize = int128_sub(llend, int128_make64(iova)); | |
e2c7d025 | 1213 | |
7a057b4f | 1214 | trace_vfio_listener_region_del(iova, end); |
e2c7d025 | 1215 | |
567b5b30 AK |
1216 | if (memory_region_is_ram_device(section->mr)) { |
1217 | hwaddr pgmask; | |
1218 | VFIOHostDMAWindow *hostwin; | |
1219 | bool hostwin_found = false; | |
1220 | ||
1221 | QLIST_FOREACH(hostwin, &container->hostwin_list, hostwin_next) { | |
1222 | if (hostwin->min_iova <= iova && end <= hostwin->max_iova) { | |
1223 | hostwin_found = true; | |
1224 | break; | |
1225 | } | |
1226 | } | |
1227 | assert(hostwin_found); /* or region_add() would have failed */ | |
1228 | ||
1229 | pgmask = (1ULL << ctz64(hostwin->iova_pgsizes)) - 1; | |
1230 | try_unmap = !((iova & pgmask) || (int128_get64(llsize) & pgmask)); | |
5e3b981c DH |
1231 | } else if (memory_region_has_ram_discard_manager(section->mr)) { |
1232 | vfio_unregister_ram_discard_listener(container, section); | |
1233 | /* Unregistering will trigger an unmap. */ | |
1234 | try_unmap = false; | |
e2c7d025 | 1235 | } |
2e4109de | 1236 | |
567b5b30 | 1237 | if (try_unmap) { |
1b296c3d JPB |
1238 | if (int128_eq(llsize, int128_2_64())) { |
1239 | /* The unmap ioctl doesn't accept a full 64-bit span. */ | |
1240 | llsize = int128_rshift(llsize, 1); | |
1241 | ret = vfio_dma_unmap(container, iova, int128_get64(llsize), NULL); | |
1242 | if (ret) { | |
1243 | error_report("vfio_dma_unmap(%p, 0x%"HWADDR_PRIx", " | |
1244 | "0x%"HWADDR_PRIx") = %d (%m)", | |
1245 | container, iova, int128_get64(llsize), ret); | |
1246 | } | |
1247 | iova += int128_get64(llsize); | |
1248 | } | |
9e7b0442 | 1249 | ret = vfio_dma_unmap(container, iova, int128_get64(llsize), NULL); |
567b5b30 AK |
1250 | if (ret) { |
1251 | error_report("vfio_dma_unmap(%p, 0x%"HWADDR_PRIx", " | |
1252 | "0x%"HWADDR_PRIx") = %d (%m)", | |
1253 | container, iova, int128_get64(llsize), ret); | |
1254 | } | |
1255 | } | |
1256 | ||
1257 | memory_region_unref(section->mr); | |
1258 | ||
2e4109de AK |
1259 | if (container->iommu_type == VFIO_SPAPR_TCE_v2_IOMMU) { |
1260 | vfio_spapr_remove_window(container, | |
1261 | section->offset_within_address_space); | |
1262 | if (vfio_host_win_del(container, | |
1263 | section->offset_within_address_space, | |
1264 | section->offset_within_address_space + | |
1265 | int128_get64(section->size) - 1) < 0) { | |
1266 | hw_error("%s: Cannot delete missing window at %"HWADDR_PRIx, | |
1267 | __func__, section->offset_within_address_space); | |
1268 | } | |
1269 | } | |
e2c7d025 EA |
1270 | } |
1271 | ||
758b96b6 KZ |
1272 | static void vfio_set_dirty_page_tracking(VFIOContainer *container, bool start) |
1273 | { | |
1274 | int ret; | |
1275 | struct vfio_iommu_type1_dirty_bitmap dirty = { | |
1276 | .argsz = sizeof(dirty), | |
1277 | }; | |
1278 | ||
b051a3f6 AH |
1279 | if (!container->dirty_pages_supported) { |
1280 | return; | |
1281 | } | |
1282 | ||
758b96b6 KZ |
1283 | if (start) { |
1284 | dirty.flags = VFIO_IOMMU_DIRTY_PAGES_FLAG_START; | |
1285 | } else { | |
1286 | dirty.flags = VFIO_IOMMU_DIRTY_PAGES_FLAG_STOP; | |
1287 | } | |
1288 | ||
1289 | ret = ioctl(container->fd, VFIO_IOMMU_DIRTY_PAGES, &dirty); | |
1290 | if (ret) { | |
1291 | error_report("Failed to set dirty tracking flag 0x%x errno: %d", | |
1292 | dirty.flags, errno); | |
1293 | } | |
1294 | } | |
1295 | ||
1296 | static void vfio_listener_log_global_start(MemoryListener *listener) | |
1297 | { | |
1298 | VFIOContainer *container = container_of(listener, VFIOContainer, listener); | |
1299 | ||
1300 | vfio_set_dirty_page_tracking(container, true); | |
1301 | } | |
1302 | ||
1303 | static void vfio_listener_log_global_stop(MemoryListener *listener) | |
1304 | { | |
1305 | VFIOContainer *container = container_of(listener, VFIOContainer, listener); | |
1306 | ||
1307 | vfio_set_dirty_page_tracking(container, false); | |
1308 | } | |
1309 | ||
b6dd6504 KW |
1310 | static int vfio_get_dirty_bitmap(VFIOContainer *container, uint64_t iova, |
1311 | uint64_t size, ram_addr_t ram_addr) | |
1312 | { | |
1313 | struct vfio_iommu_type1_dirty_bitmap *dbitmap; | |
1314 | struct vfio_iommu_type1_dirty_bitmap_get *range; | |
1315 | uint64_t pages; | |
1316 | int ret; | |
1317 | ||
b051a3f6 AH |
1318 | if (!container->dirty_pages_supported) { |
1319 | cpu_physical_memory_set_dirty_range(ram_addr, size, | |
1320 | tcg_enabled() ? DIRTY_CLIENTS_ALL : | |
1321 | DIRTY_CLIENTS_NOCODE); | |
1322 | return 0; | |
1323 | } | |
1324 | ||
b6dd6504 KW |
1325 | dbitmap = g_malloc0(sizeof(*dbitmap) + sizeof(*range)); |
1326 | ||
1327 | dbitmap->argsz = sizeof(*dbitmap) + sizeof(*range); | |
1328 | dbitmap->flags = VFIO_IOMMU_DIRTY_PAGES_FLAG_GET_BITMAP; | |
1329 | range = (struct vfio_iommu_type1_dirty_bitmap_get *)&dbitmap->data; | |
1330 | range->iova = iova; | |
1331 | range->size = size; | |
1332 | ||
1333 | /* | |
1eb7f642 KJ |
1334 | * cpu_physical_memory_set_dirty_lebitmap() supports pages in bitmap of |
1335 | * qemu_real_host_page_size to mark those dirty. Hence set bitmap's pgsize | |
1336 | * to qemu_real_host_page_size. | |
b6dd6504 | 1337 | */ |
8e3b0cbb | 1338 | range->bitmap.pgsize = qemu_real_host_page_size(); |
b6dd6504 | 1339 | |
8e3b0cbb | 1340 | pages = REAL_HOST_PAGE_ALIGN(range->size) / qemu_real_host_page_size(); |
b6dd6504 KW |
1341 | range->bitmap.size = ROUND_UP(pages, sizeof(__u64) * BITS_PER_BYTE) / |
1342 | BITS_PER_BYTE; | |
1343 | range->bitmap.data = g_try_malloc0(range->bitmap.size); | |
1344 | if (!range->bitmap.data) { | |
1345 | ret = -ENOMEM; | |
1346 | goto err_out; | |
1347 | } | |
1348 | ||
1349 | ret = ioctl(container->fd, VFIO_IOMMU_DIRTY_PAGES, dbitmap); | |
1350 | if (ret) { | |
1351 | error_report("Failed to get dirty bitmap for iova: 0x%"PRIx64 | |
1352 | " size: 0x%"PRIx64" err: %d", (uint64_t)range->iova, | |
1353 | (uint64_t)range->size, errno); | |
1354 | goto err_out; | |
1355 | } | |
1356 | ||
1357 | cpu_physical_memory_set_dirty_lebitmap((unsigned long *)range->bitmap.data, | |
1358 | ram_addr, pages); | |
1359 | ||
1360 | trace_vfio_get_dirty_bitmap(container->fd, range->iova, range->size, | |
1361 | range->bitmap.size, ram_addr); | |
1362 | err_out: | |
1363 | g_free(range->bitmap.data); | |
1364 | g_free(dbitmap); | |
1365 | ||
1366 | return ret; | |
1367 | } | |
1368 | ||
9a04fe09 KW |
1369 | typedef struct { |
1370 | IOMMUNotifier n; | |
1371 | VFIOGuestIOMMU *giommu; | |
1372 | } vfio_giommu_dirty_notifier; | |
1373 | ||
1374 | static void vfio_iommu_map_dirty_notify(IOMMUNotifier *n, IOMMUTLBEntry *iotlb) | |
1375 | { | |
1376 | vfio_giommu_dirty_notifier *gdn = container_of(n, | |
1377 | vfio_giommu_dirty_notifier, n); | |
1378 | VFIOGuestIOMMU *giommu = gdn->giommu; | |
1379 | VFIOContainer *container = giommu->container; | |
1380 | hwaddr iova = iotlb->iova + giommu->iommu_offset; | |
1381 | ram_addr_t translated_addr; | |
1382 | ||
1383 | trace_vfio_iommu_map_dirty_notify(iova, iova + iotlb->addr_mask); | |
1384 | ||
1385 | if (iotlb->target_as != &address_space_memory) { | |
1386 | error_report("Wrong target AS \"%s\", only system memory is allowed", | |
1387 | iotlb->target_as->name ? iotlb->target_as->name : "none"); | |
1388 | return; | |
1389 | } | |
1390 | ||
1391 | rcu_read_lock(); | |
1392 | if (vfio_get_xlat_addr(iotlb, NULL, &translated_addr, NULL)) { | |
1393 | int ret; | |
1394 | ||
1395 | ret = vfio_get_dirty_bitmap(container, iova, iotlb->addr_mask + 1, | |
1396 | translated_addr); | |
1397 | if (ret) { | |
1398 | error_report("vfio_iommu_map_dirty_notify(%p, 0x%"HWADDR_PRIx", " | |
1399 | "0x%"HWADDR_PRIx") = %d (%m)", | |
1400 | container, iova, | |
1401 | iotlb->addr_mask + 1, ret); | |
1402 | } | |
1403 | } | |
1404 | rcu_read_unlock(); | |
1405 | } | |
1406 | ||
5e3b981c DH |
1407 | static int vfio_ram_discard_get_dirty_bitmap(MemoryRegionSection *section, |
1408 | void *opaque) | |
1409 | { | |
1410 | const hwaddr size = int128_get64(section->size); | |
1411 | const hwaddr iova = section->offset_within_address_space; | |
1412 | const ram_addr_t ram_addr = memory_region_get_ram_addr(section->mr) + | |
1413 | section->offset_within_region; | |
1414 | VFIORamDiscardListener *vrdl = opaque; | |
1415 | ||
1416 | /* | |
1417 | * Sync the whole mapped region (spanning multiple individual mappings) | |
1418 | * in one go. | |
1419 | */ | |
1420 | return vfio_get_dirty_bitmap(vrdl->container, iova, size, ram_addr); | |
1421 | } | |
1422 | ||
1423 | static int vfio_sync_ram_discard_listener_dirty_bitmap(VFIOContainer *container, | |
1424 | MemoryRegionSection *section) | |
1425 | { | |
1426 | RamDiscardManager *rdm = memory_region_get_ram_discard_manager(section->mr); | |
1427 | VFIORamDiscardListener *vrdl = NULL; | |
1428 | ||
1429 | QLIST_FOREACH(vrdl, &container->vrdl_list, next) { | |
1430 | if (vrdl->mr == section->mr && | |
1431 | vrdl->offset_within_address_space == | |
1432 | section->offset_within_address_space) { | |
1433 | break; | |
1434 | } | |
1435 | } | |
1436 | ||
1437 | if (!vrdl) { | |
1438 | hw_error("vfio: Trying to sync missing RAM discard listener"); | |
1439 | } | |
1440 | ||
1441 | /* | |
1442 | * We only want/can synchronize the bitmap for actually mapped parts - | |
1443 | * which correspond to populated parts. Replay all populated parts. | |
1444 | */ | |
1445 | return ram_discard_manager_replay_populated(rdm, section, | |
1446 | vfio_ram_discard_get_dirty_bitmap, | |
1447 | &vrdl); | |
1448 | } | |
1449 | ||
b6dd6504 KW |
1450 | static int vfio_sync_dirty_bitmap(VFIOContainer *container, |
1451 | MemoryRegionSection *section) | |
1452 | { | |
1453 | ram_addr_t ram_addr; | |
1454 | ||
9a04fe09 KW |
1455 | if (memory_region_is_iommu(section->mr)) { |
1456 | VFIOGuestIOMMU *giommu; | |
1457 | ||
1458 | QLIST_FOREACH(giommu, &container->giommu_list, giommu_next) { | |
44ee6aaa | 1459 | if (MEMORY_REGION(giommu->iommu_mr) == section->mr && |
9a04fe09 KW |
1460 | giommu->n.start == section->offset_within_region) { |
1461 | Int128 llend; | |
1462 | vfio_giommu_dirty_notifier gdn = { .giommu = giommu }; | |
44ee6aaa | 1463 | int idx = memory_region_iommu_attrs_to_index(giommu->iommu_mr, |
9a04fe09 KW |
1464 | MEMTXATTRS_UNSPECIFIED); |
1465 | ||
1466 | llend = int128_add(int128_make64(section->offset_within_region), | |
1467 | section->size); | |
1468 | llend = int128_sub(llend, int128_one()); | |
1469 | ||
1470 | iommu_notifier_init(&gdn.n, | |
1471 | vfio_iommu_map_dirty_notify, | |
1472 | IOMMU_NOTIFIER_MAP, | |
1473 | section->offset_within_region, | |
1474 | int128_get64(llend), | |
1475 | idx); | |
44ee6aaa | 1476 | memory_region_iommu_replay(giommu->iommu_mr, &gdn.n); |
9a04fe09 KW |
1477 | break; |
1478 | } | |
1479 | } | |
1480 | return 0; | |
5e3b981c DH |
1481 | } else if (memory_region_has_ram_discard_manager(section->mr)) { |
1482 | return vfio_sync_ram_discard_listener_dirty_bitmap(container, section); | |
9a04fe09 KW |
1483 | } |
1484 | ||
b6dd6504 KW |
1485 | ram_addr = memory_region_get_ram_addr(section->mr) + |
1486 | section->offset_within_region; | |
1487 | ||
1488 | return vfio_get_dirty_bitmap(container, | |
1eb7f642 KJ |
1489 | REAL_HOST_PAGE_ALIGN(section->offset_within_address_space), |
1490 | int128_get64(section->size), ram_addr); | |
b6dd6504 KW |
1491 | } |
1492 | ||
4292d501 | 1493 | static void vfio_listener_log_sync(MemoryListener *listener, |
b6dd6504 KW |
1494 | MemoryRegionSection *section) |
1495 | { | |
1496 | VFIOContainer *container = container_of(listener, VFIOContainer, listener); | |
1497 | ||
b051a3f6 | 1498 | if (vfio_listener_skipped_section(section)) { |
b6dd6504 KW |
1499 | return; |
1500 | } | |
1501 | ||
758b96b6 | 1502 | if (vfio_devices_all_dirty_tracking(container)) { |
b6dd6504 KW |
1503 | vfio_sync_dirty_bitmap(container, section); |
1504 | } | |
1505 | } | |
1506 | ||
51b833f4 | 1507 | static const MemoryListener vfio_memory_listener = { |
142518bd | 1508 | .name = "vfio", |
e2c7d025 EA |
1509 | .region_add = vfio_listener_region_add, |
1510 | .region_del = vfio_listener_region_del, | |
758b96b6 KZ |
1511 | .log_global_start = vfio_listener_log_global_start, |
1512 | .log_global_stop = vfio_listener_log_global_stop, | |
4292d501 | 1513 | .log_sync = vfio_listener_log_sync, |
e2c7d025 EA |
1514 | }; |
1515 | ||
51b833f4 | 1516 | static void vfio_listener_release(VFIOContainer *container) |
e2c7d025 | 1517 | { |
ee0bf0e5 | 1518 | memory_listener_unregister(&container->listener); |
318f67ce AK |
1519 | if (container->iommu_type == VFIO_SPAPR_TCE_v2_IOMMU) { |
1520 | memory_listener_unregister(&container->prereg_listener); | |
1521 | } | |
e2c7d025 EA |
1522 | } |
1523 | ||
3ab7a0b4 MR |
1524 | static struct vfio_info_cap_header * |
1525 | vfio_get_cap(void *ptr, uint32_t cap_offset, uint16_t id) | |
b53b0f69 AW |
1526 | { |
1527 | struct vfio_info_cap_header *hdr; | |
b53b0f69 | 1528 | |
3ab7a0b4 | 1529 | for (hdr = ptr + cap_offset; hdr != ptr; hdr = ptr + hdr->next) { |
b53b0f69 AW |
1530 | if (hdr->id == id) { |
1531 | return hdr; | |
1532 | } | |
1533 | } | |
1534 | ||
1535 | return NULL; | |
1536 | } | |
1537 | ||
3ab7a0b4 MR |
1538 | struct vfio_info_cap_header * |
1539 | vfio_get_region_info_cap(struct vfio_region_info *info, uint16_t id) | |
1540 | { | |
1541 | if (!(info->flags & VFIO_REGION_INFO_FLAG_CAPS)) { | |
1542 | return NULL; | |
1543 | } | |
1544 | ||
1545 | return vfio_get_cap((void *)info, info->cap_offset, id); | |
1546 | } | |
1547 | ||
7486a628 MR |
1548 | static struct vfio_info_cap_header * |
1549 | vfio_get_iommu_type1_info_cap(struct vfio_iommu_type1_info *info, uint16_t id) | |
1550 | { | |
1551 | if (!(info->flags & VFIO_IOMMU_INFO_CAPS)) { | |
1552 | return NULL; | |
1553 | } | |
1554 | ||
1555 | return vfio_get_cap((void *)info, info->cap_offset, id); | |
1556 | } | |
1557 | ||
92fe289a MR |
1558 | struct vfio_info_cap_header * |
1559 | vfio_get_device_info_cap(struct vfio_device_info *info, uint16_t id) | |
1560 | { | |
1561 | if (!(info->flags & VFIO_DEVICE_FLAGS_CAPS)) { | |
1562 | return NULL; | |
1563 | } | |
1564 | ||
1565 | return vfio_get_cap((void *)info, info->cap_offset, id); | |
1566 | } | |
1567 | ||
7486a628 MR |
1568 | bool vfio_get_info_dma_avail(struct vfio_iommu_type1_info *info, |
1569 | unsigned int *avail) | |
1570 | { | |
1571 | struct vfio_info_cap_header *hdr; | |
1572 | struct vfio_iommu_type1_info_dma_avail *cap; | |
1573 | ||
1574 | /* If the capability cannot be found, assume no DMA limiting */ | |
1575 | hdr = vfio_get_iommu_type1_info_cap(info, | |
1576 | VFIO_IOMMU_TYPE1_INFO_DMA_AVAIL); | |
1577 | if (hdr == NULL) { | |
1578 | return false; | |
1579 | } | |
1580 | ||
1581 | if (avail != NULL) { | |
1582 | cap = (void *) hdr; | |
1583 | *avail = cap->avail; | |
1584 | } | |
1585 | ||
1586 | return true; | |
1587 | } | |
1588 | ||
24acf72b AW |
1589 | static int vfio_setup_region_sparse_mmaps(VFIORegion *region, |
1590 | struct vfio_region_info *info) | |
b53b0f69 AW |
1591 | { |
1592 | struct vfio_info_cap_header *hdr; | |
1593 | struct vfio_region_info_cap_sparse_mmap *sparse; | |
24acf72b | 1594 | int i, j; |
b53b0f69 AW |
1595 | |
1596 | hdr = vfio_get_region_info_cap(info, VFIO_REGION_INFO_CAP_SPARSE_MMAP); | |
1597 | if (!hdr) { | |
24acf72b | 1598 | return -ENODEV; |
b53b0f69 AW |
1599 | } |
1600 | ||
1601 | sparse = container_of(hdr, struct vfio_region_info_cap_sparse_mmap, header); | |
1602 | ||
1603 | trace_vfio_region_sparse_mmap_header(region->vbasedev->name, | |
1604 | region->nr, sparse->nr_areas); | |
1605 | ||
24acf72b AW |
1606 | region->mmaps = g_new0(VFIOMmap, sparse->nr_areas); |
1607 | ||
1608 | for (i = 0, j = 0; i < sparse->nr_areas; i++) { | |
24acf72b | 1609 | if (sparse->areas[i].size) { |
99510d27 XC |
1610 | trace_vfio_region_sparse_mmap_entry(i, sparse->areas[i].offset, |
1611 | sparse->areas[i].offset + | |
1612 | sparse->areas[i].size - 1); | |
24acf72b AW |
1613 | region->mmaps[j].offset = sparse->areas[i].offset; |
1614 | region->mmaps[j].size = sparse->areas[i].size; | |
1615 | j++; | |
1616 | } | |
b53b0f69 | 1617 | } |
24acf72b AW |
1618 | |
1619 | region->nr_mmaps = j; | |
1620 | region->mmaps = g_realloc(region->mmaps, j * sizeof(VFIOMmap)); | |
1621 | ||
1622 | return 0; | |
b53b0f69 AW |
1623 | } |
1624 | ||
db0da029 AW |
1625 | int vfio_region_setup(Object *obj, VFIODevice *vbasedev, VFIORegion *region, |
1626 | int index, const char *name) | |
e2c7d025 | 1627 | { |
db0da029 AW |
1628 | struct vfio_region_info *info; |
1629 | int ret; | |
1630 | ||
1631 | ret = vfio_get_region_info(vbasedev, index, &info); | |
1632 | if (ret) { | |
1633 | return ret; | |
1634 | } | |
1635 | ||
1636 | region->vbasedev = vbasedev; | |
1637 | region->flags = info->flags; | |
1638 | region->size = info->size; | |
1639 | region->fd_offset = info->offset; | |
1640 | region->nr = index; | |
1641 | ||
1642 | if (region->size) { | |
1643 | region->mem = g_new0(MemoryRegion, 1); | |
1644 | memory_region_init_io(region->mem, obj, &vfio_region_ops, | |
1645 | region, name, region->size); | |
e2c7d025 | 1646 | |
db0da029 | 1647 | if (!vbasedev->no_mmap && |
95251725 | 1648 | region->flags & VFIO_REGION_INFO_FLAG_MMAP) { |
e2c7d025 | 1649 | |
24acf72b | 1650 | ret = vfio_setup_region_sparse_mmaps(region, info); |
db0da029 | 1651 | |
24acf72b | 1652 | if (ret) { |
b53b0f69 AW |
1653 | region->nr_mmaps = 1; |
1654 | region->mmaps = g_new0(VFIOMmap, region->nr_mmaps); | |
1655 | region->mmaps[0].offset = 0; | |
1656 | region->mmaps[0].size = region->size; | |
1657 | } | |
e2c7d025 | 1658 | } |
db0da029 AW |
1659 | } |
1660 | ||
1661 | g_free(info); | |
1662 | ||
1663 | trace_vfio_region_setup(vbasedev->name, index, name, | |
1664 | region->flags, region->fd_offset, region->size); | |
1665 | return 0; | |
1666 | } | |
e2c7d025 | 1667 | |
0f7a903b KW |
1668 | static void vfio_subregion_unmap(VFIORegion *region, int index) |
1669 | { | |
1670 | trace_vfio_region_unmap(memory_region_name(®ion->mmaps[index].mem), | |
1671 | region->mmaps[index].offset, | |
1672 | region->mmaps[index].offset + | |
1673 | region->mmaps[index].size - 1); | |
1674 | memory_region_del_subregion(region->mem, ®ion->mmaps[index].mem); | |
1675 | munmap(region->mmaps[index].mmap, region->mmaps[index].size); | |
1676 | object_unparent(OBJECT(®ion->mmaps[index].mem)); | |
1677 | region->mmaps[index].mmap = NULL; | |
1678 | } | |
1679 | ||
db0da029 AW |
1680 | int vfio_region_mmap(VFIORegion *region) |
1681 | { | |
1682 | int i, prot = 0; | |
1683 | char *name; | |
1684 | ||
1685 | if (!region->mem) { | |
1686 | return 0; | |
1687 | } | |
1688 | ||
1689 | prot |= region->flags & VFIO_REGION_INFO_FLAG_READ ? PROT_READ : 0; | |
1690 | prot |= region->flags & VFIO_REGION_INFO_FLAG_WRITE ? PROT_WRITE : 0; | |
1691 | ||
1692 | for (i = 0; i < region->nr_mmaps; i++) { | |
1693 | region->mmaps[i].mmap = mmap(NULL, region->mmaps[i].size, prot, | |
1694 | MAP_SHARED, region->vbasedev->fd, | |
1695 | region->fd_offset + | |
1696 | region->mmaps[i].offset); | |
1697 | if (region->mmaps[i].mmap == MAP_FAILED) { | |
1698 | int ret = -errno; | |
1699 | ||
1700 | trace_vfio_region_mmap_fault(memory_region_name(region->mem), i, | |
1701 | region->fd_offset + | |
1702 | region->mmaps[i].offset, | |
1703 | region->fd_offset + | |
1704 | region->mmaps[i].offset + | |
1705 | region->mmaps[i].size - 1, ret); | |
1706 | ||
1707 | region->mmaps[i].mmap = NULL; | |
1708 | ||
1709 | for (i--; i >= 0; i--) { | |
0f7a903b | 1710 | vfio_subregion_unmap(region, i); |
db0da029 AW |
1711 | } |
1712 | ||
1713 | return ret; | |
e2c7d025 EA |
1714 | } |
1715 | ||
db0da029 AW |
1716 | name = g_strdup_printf("%s mmaps[%d]", |
1717 | memory_region_name(region->mem), i); | |
21e00fa5 AW |
1718 | memory_region_init_ram_device_ptr(®ion->mmaps[i].mem, |
1719 | memory_region_owner(region->mem), | |
1720 | name, region->mmaps[i].size, | |
1721 | region->mmaps[i].mmap); | |
db0da029 | 1722 | g_free(name); |
db0da029 AW |
1723 | memory_region_add_subregion(region->mem, region->mmaps[i].offset, |
1724 | ®ion->mmaps[i].mem); | |
1725 | ||
1726 | trace_vfio_region_mmap(memory_region_name(®ion->mmaps[i].mem), | |
1727 | region->mmaps[i].offset, | |
1728 | region->mmaps[i].offset + | |
1729 | region->mmaps[i].size - 1); | |
1730 | } | |
1731 | ||
1732 | return 0; | |
1733 | } | |
1734 | ||
0f7a903b KW |
1735 | void vfio_region_unmap(VFIORegion *region) |
1736 | { | |
1737 | int i; | |
1738 | ||
1739 | if (!region->mem) { | |
1740 | return; | |
1741 | } | |
1742 | ||
1743 | for (i = 0; i < region->nr_mmaps; i++) { | |
1744 | if (region->mmaps[i].mmap) { | |
1745 | vfio_subregion_unmap(region, i); | |
1746 | } | |
1747 | } | |
1748 | } | |
1749 | ||
db0da029 AW |
1750 | void vfio_region_exit(VFIORegion *region) |
1751 | { | |
1752 | int i; | |
1753 | ||
1754 | if (!region->mem) { | |
1755 | return; | |
1756 | } | |
1757 | ||
1758 | for (i = 0; i < region->nr_mmaps; i++) { | |
1759 | if (region->mmaps[i].mmap) { | |
1760 | memory_region_del_subregion(region->mem, ®ion->mmaps[i].mem); | |
e2c7d025 | 1761 | } |
db0da029 | 1762 | } |
e2c7d025 | 1763 | |
db0da029 AW |
1764 | trace_vfio_region_exit(region->vbasedev->name, region->nr); |
1765 | } | |
1766 | ||
1767 | void vfio_region_finalize(VFIORegion *region) | |
1768 | { | |
1769 | int i; | |
1770 | ||
1771 | if (!region->mem) { | |
1772 | return; | |
e2c7d025 EA |
1773 | } |
1774 | ||
db0da029 AW |
1775 | for (i = 0; i < region->nr_mmaps; i++) { |
1776 | if (region->mmaps[i].mmap) { | |
1777 | munmap(region->mmaps[i].mmap, region->mmaps[i].size); | |
1778 | object_unparent(OBJECT(®ion->mmaps[i].mem)); | |
1779 | } | |
1780 | } | |
1781 | ||
1782 | object_unparent(OBJECT(region->mem)); | |
1783 | ||
1784 | g_free(region->mem); | |
1785 | g_free(region->mmaps); | |
1786 | ||
1787 | trace_vfio_region_finalize(region->vbasedev->name, region->nr); | |
92f86bff GH |
1788 | |
1789 | region->mem = NULL; | |
1790 | region->mmaps = NULL; | |
1791 | region->nr_mmaps = 0; | |
1792 | region->size = 0; | |
1793 | region->flags = 0; | |
1794 | region->nr = 0; | |
db0da029 AW |
1795 | } |
1796 | ||
1797 | void vfio_region_mmaps_set_enabled(VFIORegion *region, bool enabled) | |
1798 | { | |
1799 | int i; | |
1800 | ||
1801 | if (!region->mem) { | |
1802 | return; | |
1803 | } | |
1804 | ||
1805 | for (i = 0; i < region->nr_mmaps; i++) { | |
1806 | if (region->mmaps[i].mmap) { | |
1807 | memory_region_set_enabled(®ion->mmaps[i].mem, enabled); | |
1808 | } | |
1809 | } | |
e2c7d025 | 1810 | |
db0da029 AW |
1811 | trace_vfio_region_mmaps_set_enabled(memory_region_name(region->mem), |
1812 | enabled); | |
e2c7d025 EA |
1813 | } |
1814 | ||
1815 | void vfio_reset_handler(void *opaque) | |
1816 | { | |
1817 | VFIOGroup *group; | |
1818 | VFIODevice *vbasedev; | |
1819 | ||
1820 | QLIST_FOREACH(group, &vfio_group_list, next) { | |
1821 | QLIST_FOREACH(vbasedev, &group->device_list, next) { | |
7da624e2 AW |
1822 | if (vbasedev->dev->realized) { |
1823 | vbasedev->ops->vfio_compute_needs_reset(vbasedev); | |
1824 | } | |
e2c7d025 EA |
1825 | } |
1826 | } | |
1827 | ||
1828 | QLIST_FOREACH(group, &vfio_group_list, next) { | |
1829 | QLIST_FOREACH(vbasedev, &group->device_list, next) { | |
7da624e2 | 1830 | if (vbasedev->dev->realized && vbasedev->needs_reset) { |
e2c7d025 EA |
1831 | vbasedev->ops->vfio_hot_reset_multi(vbasedev); |
1832 | } | |
1833 | } | |
1834 | } | |
1835 | } | |
1836 | ||
1837 | static void vfio_kvm_device_add_group(VFIOGroup *group) | |
1838 | { | |
1839 | #ifdef CONFIG_KVM | |
1840 | struct kvm_device_attr attr = { | |
1841 | .group = KVM_DEV_VFIO_GROUP, | |
1842 | .attr = KVM_DEV_VFIO_GROUP_ADD, | |
1843 | .addr = (uint64_t)(unsigned long)&group->fd, | |
1844 | }; | |
1845 | ||
1846 | if (!kvm_enabled()) { | |
1847 | return; | |
1848 | } | |
1849 | ||
1850 | if (vfio_kvm_device_fd < 0) { | |
1851 | struct kvm_create_device cd = { | |
1852 | .type = KVM_DEV_TYPE_VFIO, | |
1853 | }; | |
1854 | ||
1855 | if (kvm_vm_ioctl(kvm_state, KVM_CREATE_DEVICE, &cd)) { | |
78e5b17f | 1856 | error_report("Failed to create KVM VFIO device: %m"); |
e2c7d025 EA |
1857 | return; |
1858 | } | |
1859 | ||
1860 | vfio_kvm_device_fd = cd.fd; | |
1861 | } | |
1862 | ||
1863 | if (ioctl(vfio_kvm_device_fd, KVM_SET_DEVICE_ATTR, &attr)) { | |
1864 | error_report("Failed to add group %d to KVM VFIO device: %m", | |
1865 | group->groupid); | |
1866 | } | |
1867 | #endif | |
1868 | } | |
1869 | ||
1870 | static void vfio_kvm_device_del_group(VFIOGroup *group) | |
1871 | { | |
1872 | #ifdef CONFIG_KVM | |
1873 | struct kvm_device_attr attr = { | |
1874 | .group = KVM_DEV_VFIO_GROUP, | |
1875 | .attr = KVM_DEV_VFIO_GROUP_DEL, | |
1876 | .addr = (uint64_t)(unsigned long)&group->fd, | |
1877 | }; | |
1878 | ||
1879 | if (vfio_kvm_device_fd < 0) { | |
1880 | return; | |
1881 | } | |
1882 | ||
1883 | if (ioctl(vfio_kvm_device_fd, KVM_SET_DEVICE_ATTR, &attr)) { | |
1884 | error_report("Failed to remove group %d from KVM VFIO device: %m", | |
1885 | group->groupid); | |
1886 | } | |
1887 | #endif | |
1888 | } | |
1889 | ||
1890 | static VFIOAddressSpace *vfio_get_address_space(AddressSpace *as) | |
1891 | { | |
1892 | VFIOAddressSpace *space; | |
1893 | ||
1894 | QLIST_FOREACH(space, &vfio_address_spaces, list) { | |
1895 | if (space->as == as) { | |
1896 | return space; | |
1897 | } | |
1898 | } | |
1899 | ||
1900 | /* No suitable VFIOAddressSpace, create a new one */ | |
1901 | space = g_malloc0(sizeof(*space)); | |
1902 | space->as = as; | |
1903 | QLIST_INIT(&space->containers); | |
1904 | ||
1905 | QLIST_INSERT_HEAD(&vfio_address_spaces, space, list); | |
1906 | ||
1907 | return space; | |
1908 | } | |
1909 | ||
1910 | static void vfio_put_address_space(VFIOAddressSpace *space) | |
1911 | { | |
1912 | if (QLIST_EMPTY(&space->containers)) { | |
1913 | QLIST_REMOVE(space, list); | |
1914 | g_free(space); | |
1915 | } | |
1916 | } | |
1917 | ||
2b6326c0 EA |
1918 | /* |
1919 | * vfio_get_iommu_type - selects the richest iommu_type (v2 first) | |
1920 | */ | |
1921 | static int vfio_get_iommu_type(VFIOContainer *container, | |
1922 | Error **errp) | |
1923 | { | |
1924 | int iommu_types[] = { VFIO_TYPE1v2_IOMMU, VFIO_TYPE1_IOMMU, | |
1925 | VFIO_SPAPR_TCE_v2_IOMMU, VFIO_SPAPR_TCE_IOMMU }; | |
1926 | int i; | |
1927 | ||
1928 | for (i = 0; i < ARRAY_SIZE(iommu_types); i++) { | |
1929 | if (ioctl(container->fd, VFIO_CHECK_EXTENSION, iommu_types[i])) { | |
1930 | return iommu_types[i]; | |
1931 | } | |
1932 | } | |
1933 | error_setg(errp, "No available IOMMU models"); | |
1934 | return -EINVAL; | |
1935 | } | |
1936 | ||
1937 | static int vfio_init_container(VFIOContainer *container, int group_fd, | |
1938 | Error **errp) | |
1939 | { | |
1940 | int iommu_type, ret; | |
1941 | ||
1942 | iommu_type = vfio_get_iommu_type(container, errp); | |
1943 | if (iommu_type < 0) { | |
1944 | return iommu_type; | |
1945 | } | |
1946 | ||
1947 | ret = ioctl(group_fd, VFIO_GROUP_SET_CONTAINER, &container->fd); | |
1948 | if (ret) { | |
1949 | error_setg_errno(errp, errno, "Failed to set group container"); | |
1950 | return -errno; | |
1951 | } | |
1952 | ||
1953 | while (ioctl(container->fd, VFIO_SET_IOMMU, iommu_type)) { | |
1954 | if (iommu_type == VFIO_SPAPR_TCE_v2_IOMMU) { | |
1955 | /* | |
1956 | * On sPAPR, despite the IOMMU subdriver always advertises v1 and | |
1957 | * v2, the running platform may not support v2 and there is no | |
1958 | * way to guess it until an IOMMU group gets added to the container. | |
1959 | * So in case it fails with v2, try v1 as a fallback. | |
1960 | */ | |
1961 | iommu_type = VFIO_SPAPR_TCE_IOMMU; | |
1962 | continue; | |
1963 | } | |
1964 | error_setg_errno(errp, errno, "Failed to set iommu for container"); | |
1965 | return -errno; | |
1966 | } | |
1967 | ||
1968 | container->iommu_type = iommu_type; | |
1969 | return 0; | |
1970 | } | |
1971 | ||
87ea529c KW |
1972 | static int vfio_get_iommu_info(VFIOContainer *container, |
1973 | struct vfio_iommu_type1_info **info) | |
1974 | { | |
1975 | ||
1976 | size_t argsz = sizeof(struct vfio_iommu_type1_info); | |
1977 | ||
1978 | *info = g_new0(struct vfio_iommu_type1_info, 1); | |
1979 | again: | |
1980 | (*info)->argsz = argsz; | |
1981 | ||
1982 | if (ioctl(container->fd, VFIO_IOMMU_GET_INFO, *info)) { | |
1983 | g_free(*info); | |
1984 | *info = NULL; | |
1985 | return -errno; | |
1986 | } | |
1987 | ||
1988 | if (((*info)->argsz > argsz)) { | |
1989 | argsz = (*info)->argsz; | |
1990 | *info = g_realloc(*info, argsz); | |
1991 | goto again; | |
1992 | } | |
1993 | ||
1994 | return 0; | |
1995 | } | |
1996 | ||
1997 | static struct vfio_info_cap_header * | |
1998 | vfio_get_iommu_info_cap(struct vfio_iommu_type1_info *info, uint16_t id) | |
1999 | { | |
2000 | struct vfio_info_cap_header *hdr; | |
2001 | void *ptr = info; | |
2002 | ||
2003 | if (!(info->flags & VFIO_IOMMU_INFO_CAPS)) { | |
2004 | return NULL; | |
2005 | } | |
2006 | ||
2007 | for (hdr = ptr + info->cap_offset; hdr != ptr; hdr = ptr + hdr->next) { | |
2008 | if (hdr->id == id) { | |
2009 | return hdr; | |
2010 | } | |
2011 | } | |
2012 | ||
2013 | return NULL; | |
2014 | } | |
2015 | ||
2016 | static void vfio_get_iommu_info_migration(VFIOContainer *container, | |
2017 | struct vfio_iommu_type1_info *info) | |
2018 | { | |
2019 | struct vfio_info_cap_header *hdr; | |
2020 | struct vfio_iommu_type1_info_cap_migration *cap_mig; | |
2021 | ||
2022 | hdr = vfio_get_iommu_info_cap(info, VFIO_IOMMU_TYPE1_INFO_CAP_MIGRATION); | |
2023 | if (!hdr) { | |
2024 | return; | |
2025 | } | |
2026 | ||
2027 | cap_mig = container_of(hdr, struct vfio_iommu_type1_info_cap_migration, | |
2028 | header); | |
2029 | ||
2030 | /* | |
1eb7f642 KJ |
2031 | * cpu_physical_memory_set_dirty_lebitmap() supports pages in bitmap of |
2032 | * qemu_real_host_page_size to mark those dirty. | |
87ea529c | 2033 | */ |
8e3b0cbb | 2034 | if (cap_mig->pgsize_bitmap & qemu_real_host_page_size()) { |
87ea529c KW |
2035 | container->dirty_pages_supported = true; |
2036 | container->max_dirty_bitmap_size = cap_mig->max_dirty_bitmap_size; | |
2037 | container->dirty_pgsizes = cap_mig->pgsize_bitmap; | |
2038 | } | |
2039 | } | |
2040 | ||
01905f58 EA |
2041 | static int vfio_connect_container(VFIOGroup *group, AddressSpace *as, |
2042 | Error **errp) | |
e2c7d025 EA |
2043 | { |
2044 | VFIOContainer *container; | |
2045 | int ret, fd; | |
2046 | VFIOAddressSpace *space; | |
2047 | ||
2048 | space = vfio_get_address_space(as); | |
2049 | ||
c65ee433 | 2050 | /* |
aff92b82 | 2051 | * VFIO is currently incompatible with discarding of RAM insofar as the |
c65ee433 AW |
2052 | * madvise to purge (zap) the page from QEMU's address space does not |
2053 | * interact with the memory API and therefore leaves stale virtual to | |
2054 | * physical mappings in the IOMMU if the page was previously pinned. We | |
aff92b82 | 2055 | * therefore set discarding broken for each group added to a container, |
c65ee433 AW |
2056 | * whether the container is used individually or shared. This provides |
2057 | * us with options to allow devices within a group to opt-in and allow | |
aff92b82 | 2058 | * discarding, so long as it is done consistently for a group (for instance |
c65ee433 AW |
2059 | * if the device is an mdev device where it is known that the host vendor |
2060 | * driver will never pin pages outside of the working set of the guest | |
aff92b82 | 2061 | * driver, which would thus not be discarding candidates). |
c65ee433 AW |
2062 | * |
2063 | * The first opportunity to induce pinning occurs here where we attempt to | |
2064 | * attach the group to existing containers within the AddressSpace. If any | |
aff92b82 DH |
2065 | * pages are already zapped from the virtual address space, such as from |
2066 | * previous discards, new pinning will cause valid mappings to be | |
c65ee433 AW |
2067 | * re-established. Likewise, when the overall MemoryListener for a new |
2068 | * container is registered, a replay of mappings within the AddressSpace | |
2069 | * will occur, re-establishing any previously zapped pages as well. | |
2070 | * | |
aff92b82 DH |
2071 | * Especially virtio-balloon is currently only prevented from discarding |
2072 | * new memory, it will not yet set ram_block_discard_set_required() and | |
2073 | * therefore, neither stops us here or deals with the sudden memory | |
2074 | * consumption of inflated memory. | |
53d1b5fc DH |
2075 | * |
2076 | * We do support discarding of memory coordinated via the RamDiscardManager | |
2077 | * with some IOMMU types. vfio_ram_block_discard_disable() handles the | |
2078 | * details once we know which type of IOMMU we are using. | |
c65ee433 | 2079 | */ |
c65ee433 | 2080 | |
e2c7d025 EA |
2081 | QLIST_FOREACH(container, &space->containers, next) { |
2082 | if (!ioctl(group->fd, VFIO_GROUP_SET_CONTAINER, &container->fd)) { | |
53d1b5fc DH |
2083 | ret = vfio_ram_block_discard_disable(container, true); |
2084 | if (ret) { | |
2085 | error_setg_errno(errp, -ret, | |
2086 | "Cannot set discarding of RAM broken"); | |
2087 | if (ioctl(group->fd, VFIO_GROUP_UNSET_CONTAINER, | |
2088 | &container->fd)) { | |
2089 | error_report("vfio: error disconnecting group %d from" | |
2090 | " container", group->groupid); | |
2091 | } | |
2092 | return ret; | |
2093 | } | |
e2c7d025 EA |
2094 | group->container = container; |
2095 | QLIST_INSERT_HEAD(&container->group_list, group, container_next); | |
2016986a | 2096 | vfio_kvm_device_add_group(group); |
e2c7d025 EA |
2097 | return 0; |
2098 | } | |
2099 | } | |
2100 | ||
448058aa | 2101 | fd = qemu_open_old("/dev/vfio/vfio", O_RDWR); |
e2c7d025 | 2102 | if (fd < 0) { |
01905f58 | 2103 | error_setg_errno(errp, errno, "failed to open /dev/vfio/vfio"); |
e2c7d025 EA |
2104 | ret = -errno; |
2105 | goto put_space_exit; | |
2106 | } | |
2107 | ||
2108 | ret = ioctl(fd, VFIO_GET_API_VERSION); | |
2109 | if (ret != VFIO_API_VERSION) { | |
01905f58 EA |
2110 | error_setg(errp, "supported vfio version: %d, " |
2111 | "reported version: %d", VFIO_API_VERSION, ret); | |
e2c7d025 EA |
2112 | ret = -EINVAL; |
2113 | goto close_fd_exit; | |
2114 | } | |
2115 | ||
2116 | container = g_malloc0(sizeof(*container)); | |
2117 | container->space = space; | |
2118 | container->fd = fd; | |
d7d87836 | 2119 | container->error = NULL; |
87ea529c | 2120 | container->dirty_pages_supported = false; |
3eed155c | 2121 | container->dma_max_mappings = 0; |
f7f9c7b2 LY |
2122 | QLIST_INIT(&container->giommu_list); |
2123 | QLIST_INIT(&container->hostwin_list); | |
5e3b981c | 2124 | QLIST_INIT(&container->vrdl_list); |
2e6e697e | 2125 | |
2b6326c0 EA |
2126 | ret = vfio_init_container(container, group->fd, errp); |
2127 | if (ret) { | |
2128 | goto free_container_exit; | |
2129 | } | |
e2c7d025 | 2130 | |
53d1b5fc DH |
2131 | ret = vfio_ram_block_discard_disable(container, true); |
2132 | if (ret) { | |
2133 | error_setg_errno(errp, -ret, "Cannot set discarding of RAM broken"); | |
2134 | goto free_container_exit; | |
2135 | } | |
2136 | ||
2b6326c0 EA |
2137 | switch (container->iommu_type) { |
2138 | case VFIO_TYPE1v2_IOMMU: | |
2139 | case VFIO_TYPE1_IOMMU: | |
2140 | { | |
87ea529c | 2141 | struct vfio_iommu_type1_info *info; |
3898aad3 | 2142 | |
87ea529c | 2143 | ret = vfio_get_iommu_info(container, &info); |
85b6d2b5 AW |
2144 | if (ret) { |
2145 | error_setg_errno(errp, -ret, "Failed to get VFIO IOMMU info"); | |
2146 | goto enable_discards_exit; | |
2147 | } | |
87ea529c | 2148 | |
85b6d2b5 AW |
2149 | if (info->flags & VFIO_IOMMU_INFO_PGSIZES) { |
2150 | container->pgsizes = info->iova_pgsizes; | |
2151 | } else { | |
2152 | container->pgsizes = qemu_real_host_page_size(); | |
87ea529c | 2153 | } |
85b6d2b5 AW |
2154 | |
2155 | if (!vfio_get_info_dma_avail(info, &container->dma_max_mappings)) { | |
2156 | container->dma_max_mappings = 65535; | |
7a140a57 | 2157 | } |
85b6d2b5 | 2158 | vfio_get_iommu_info_migration(container, info); |
87ea529c | 2159 | g_free(info); |
85b6d2b5 AW |
2160 | |
2161 | /* | |
2162 | * FIXME: We should parse VFIO_IOMMU_TYPE1_INFO_CAP_IOVA_RANGE | |
2163 | * information to get the actual window extent rather than assume | |
2164 | * a 64-bit IOVA address space. | |
2165 | */ | |
2166 | vfio_host_win_add(container, 0, (hwaddr)-1, container->pgsizes); | |
2167 | ||
2b6326c0 EA |
2168 | break; |
2169 | } | |
2170 | case VFIO_SPAPR_TCE_v2_IOMMU: | |
2171 | case VFIO_SPAPR_TCE_IOMMU: | |
2172 | { | |
3898aad3 | 2173 | struct vfio_iommu_spapr_tce_info info; |
2b6326c0 | 2174 | bool v2 = container->iommu_type == VFIO_SPAPR_TCE_v2_IOMMU; |
e2c7d025 EA |
2175 | |
2176 | /* | |
2177 | * The host kernel code implementing VFIO_IOMMU_DISABLE is called | |
2178 | * when container fd is closed so we do not call it explicitly | |
2179 | * in this file. | |
2180 | */ | |
318f67ce AK |
2181 | if (!v2) { |
2182 | ret = ioctl(fd, VFIO_IOMMU_ENABLE); | |
2183 | if (ret) { | |
01905f58 | 2184 | error_setg_errno(errp, errno, "failed to enable container"); |
318f67ce | 2185 | ret = -errno; |
53d1b5fc | 2186 | goto enable_discards_exit; |
318f67ce AK |
2187 | } |
2188 | } else { | |
2189 | container->prereg_listener = vfio_prereg_listener; | |
2190 | ||
2191 | memory_listener_register(&container->prereg_listener, | |
2192 | &address_space_memory); | |
2193 | if (container->error) { | |
2194 | memory_listener_unregister(&container->prereg_listener); | |
d7d87836 EA |
2195 | ret = -1; |
2196 | error_propagate_prepend(errp, container->error, | |
2197 | "RAM memory listener initialization failed: "); | |
53d1b5fc | 2198 | goto enable_discards_exit; |
318f67ce | 2199 | } |
e2c7d025 | 2200 | } |
3898aad3 | 2201 | |
3898aad3 DG |
2202 | info.argsz = sizeof(info); |
2203 | ret = ioctl(fd, VFIO_IOMMU_SPAPR_TCE_GET_INFO, &info); | |
2204 | if (ret) { | |
01905f58 EA |
2205 | error_setg_errno(errp, errno, |
2206 | "VFIO_IOMMU_SPAPR_TCE_GET_INFO failed"); | |
3898aad3 | 2207 | ret = -errno; |
318f67ce AK |
2208 | if (v2) { |
2209 | memory_listener_unregister(&container->prereg_listener); | |
2210 | } | |
53d1b5fc | 2211 | goto enable_discards_exit; |
3898aad3 | 2212 | } |
7a140a57 | 2213 | |
2e4109de | 2214 | if (v2) { |
c26bc185 | 2215 | container->pgsizes = info.ddw.pgsizes; |
2e4109de AK |
2216 | /* |
2217 | * There is a default window in just created container. | |
2218 | * To make region_add/del simpler, we better remove this | |
2219 | * window now and let those iommu_listener callbacks | |
2220 | * create/remove them when needed. | |
2221 | */ | |
2222 | ret = vfio_spapr_remove_window(container, info.dma32_window_start); | |
2223 | if (ret) { | |
01905f58 EA |
2224 | error_setg_errno(errp, -ret, |
2225 | "failed to remove existing window"); | |
53d1b5fc | 2226 | goto enable_discards_exit; |
2e4109de AK |
2227 | } |
2228 | } else { | |
2229 | /* The default table uses 4K pages */ | |
c26bc185 | 2230 | container->pgsizes = 0x1000; |
2e4109de AK |
2231 | vfio_host_win_add(container, info.dma32_window_start, |
2232 | info.dma32_window_start + | |
2233 | info.dma32_window_size - 1, | |
2234 | 0x1000); | |
2235 | } | |
2b6326c0 | 2236 | } |
e2c7d025 EA |
2237 | } |
2238 | ||
8c37faa4 AK |
2239 | vfio_kvm_device_add_group(group); |
2240 | ||
2241 | QLIST_INIT(&container->group_list); | |
2242 | QLIST_INSERT_HEAD(&space->containers, container, next); | |
2243 | ||
2244 | group->container = container; | |
2245 | QLIST_INSERT_HEAD(&container->group_list, group, container_next); | |
2246 | ||
ee0bf0e5 DG |
2247 | container->listener = vfio_memory_listener; |
2248 | ||
2249 | memory_listener_register(&container->listener, container->space->as); | |
2250 | ||
2251 | if (container->error) { | |
d7d87836 EA |
2252 | ret = -1; |
2253 | error_propagate_prepend(errp, container->error, | |
2254 | "memory listener initialization failed: "); | |
ee0bf0e5 DG |
2255 | goto listener_release_exit; |
2256 | } | |
2257 | ||
2258 | container->initialized = true; | |
2259 | ||
e2c7d025 EA |
2260 | return 0; |
2261 | listener_release_exit: | |
8c37faa4 AK |
2262 | QLIST_REMOVE(group, container_next); |
2263 | QLIST_REMOVE(container, next); | |
2264 | vfio_kvm_device_del_group(group); | |
e2c7d025 EA |
2265 | vfio_listener_release(container); |
2266 | ||
53d1b5fc DH |
2267 | enable_discards_exit: |
2268 | vfio_ram_block_discard_disable(container, false); | |
2269 | ||
e2c7d025 EA |
2270 | free_container_exit: |
2271 | g_free(container); | |
2272 | ||
2273 | close_fd_exit: | |
2274 | close(fd); | |
2275 | ||
2276 | put_space_exit: | |
2277 | vfio_put_address_space(space); | |
2278 | ||
2279 | return ret; | |
2280 | } | |
2281 | ||
2282 | static void vfio_disconnect_container(VFIOGroup *group) | |
2283 | { | |
2284 | VFIOContainer *container = group->container; | |
2285 | ||
36968626 PX |
2286 | QLIST_REMOVE(group, container_next); |
2287 | group->container = NULL; | |
2288 | ||
2289 | /* | |
2290 | * Explicitly release the listener first before unset container, | |
2291 | * since unset may destroy the backend container if it's the last | |
2292 | * group. | |
2293 | */ | |
2294 | if (QLIST_EMPTY(&container->group_list)) { | |
2295 | vfio_listener_release(container); | |
2296 | } | |
2297 | ||
e2c7d025 EA |
2298 | if (ioctl(group->fd, VFIO_GROUP_UNSET_CONTAINER, &container->fd)) { |
2299 | error_report("vfio: error disconnecting group %d from container", | |
2300 | group->groupid); | |
2301 | } | |
2302 | ||
e2c7d025 EA |
2303 | if (QLIST_EMPTY(&container->group_list)) { |
2304 | VFIOAddressSpace *space = container->space; | |
f8d8a944 | 2305 | VFIOGuestIOMMU *giommu, *tmp; |
f3bc3a73 | 2306 | VFIOHostDMAWindow *hostwin, *next; |
e2c7d025 | 2307 | |
e2c7d025 | 2308 | QLIST_REMOVE(container, next); |
f8d8a944 AK |
2309 | |
2310 | QLIST_FOREACH_SAFE(giommu, &container->giommu_list, giommu_next, tmp) { | |
3df9d748 | 2311 | memory_region_unregister_iommu_notifier( |
44ee6aaa | 2312 | MEMORY_REGION(giommu->iommu_mr), &giommu->n); |
f8d8a944 AK |
2313 | QLIST_REMOVE(giommu, giommu_next); |
2314 | g_free(giommu); | |
2315 | } | |
2316 | ||
f3bc3a73 PL |
2317 | QLIST_FOREACH_SAFE(hostwin, &container->hostwin_list, hostwin_next, |
2318 | next) { | |
2319 | QLIST_REMOVE(hostwin, hostwin_next); | |
2320 | g_free(hostwin); | |
2321 | } | |
2322 | ||
e2c7d025 EA |
2323 | trace_vfio_disconnect_container(container->fd); |
2324 | close(container->fd); | |
2325 | g_free(container); | |
2326 | ||
2327 | vfio_put_address_space(space); | |
2328 | } | |
2329 | } | |
2330 | ||
1b808d5b | 2331 | VFIOGroup *vfio_get_group(int groupid, AddressSpace *as, Error **errp) |
e2c7d025 EA |
2332 | { |
2333 | VFIOGroup *group; | |
2334 | char path[32]; | |
2335 | struct vfio_group_status status = { .argsz = sizeof(status) }; | |
2336 | ||
2337 | QLIST_FOREACH(group, &vfio_group_list, next) { | |
2338 | if (group->groupid == groupid) { | |
2339 | /* Found it. Now is it already in the right context? */ | |
2340 | if (group->container->space->as == as) { | |
2341 | return group; | |
2342 | } else { | |
1b808d5b EA |
2343 | error_setg(errp, "group %d used in multiple address spaces", |
2344 | group->groupid); | |
e2c7d025 EA |
2345 | return NULL; |
2346 | } | |
2347 | } | |
2348 | } | |
2349 | ||
2350 | group = g_malloc0(sizeof(*group)); | |
2351 | ||
2352 | snprintf(path, sizeof(path), "/dev/vfio/%d", groupid); | |
448058aa | 2353 | group->fd = qemu_open_old(path, O_RDWR); |
e2c7d025 | 2354 | if (group->fd < 0) { |
1b808d5b | 2355 | error_setg_errno(errp, errno, "failed to open %s", path); |
e2c7d025 EA |
2356 | goto free_group_exit; |
2357 | } | |
2358 | ||
2359 | if (ioctl(group->fd, VFIO_GROUP_GET_STATUS, &status)) { | |
1b808d5b | 2360 | error_setg_errno(errp, errno, "failed to get group %d status", groupid); |
e2c7d025 EA |
2361 | goto close_fd_exit; |
2362 | } | |
2363 | ||
2364 | if (!(status.flags & VFIO_GROUP_FLAGS_VIABLE)) { | |
1b808d5b EA |
2365 | error_setg(errp, "group %d is not viable", groupid); |
2366 | error_append_hint(errp, | |
2367 | "Please ensure all devices within the iommu_group " | |
2368 | "are bound to their vfio bus driver.\n"); | |
e2c7d025 EA |
2369 | goto close_fd_exit; |
2370 | } | |
2371 | ||
2372 | group->groupid = groupid; | |
2373 | QLIST_INIT(&group->device_list); | |
2374 | ||
1b808d5b EA |
2375 | if (vfio_connect_container(group, as, errp)) { |
2376 | error_prepend(errp, "failed to setup container for group %d: ", | |
2377 | groupid); | |
e2c7d025 EA |
2378 | goto close_fd_exit; |
2379 | } | |
2380 | ||
2381 | if (QLIST_EMPTY(&vfio_group_list)) { | |
2382 | qemu_register_reset(vfio_reset_handler, NULL); | |
2383 | } | |
2384 | ||
2385 | QLIST_INSERT_HEAD(&vfio_group_list, group, next); | |
2386 | ||
e2c7d025 EA |
2387 | return group; |
2388 | ||
2389 | close_fd_exit: | |
2390 | close(group->fd); | |
2391 | ||
2392 | free_group_exit: | |
2393 | g_free(group); | |
2394 | ||
2395 | return NULL; | |
2396 | } | |
2397 | ||
2398 | void vfio_put_group(VFIOGroup *group) | |
2399 | { | |
77a10d04 | 2400 | if (!group || !QLIST_EMPTY(&group->device_list)) { |
e2c7d025 EA |
2401 | return; |
2402 | } | |
2403 | ||
aff92b82 | 2404 | if (!group->ram_block_discard_allowed) { |
53d1b5fc | 2405 | vfio_ram_block_discard_disable(group->container, false); |
238e9172 | 2406 | } |
e2c7d025 EA |
2407 | vfio_kvm_device_del_group(group); |
2408 | vfio_disconnect_container(group); | |
2409 | QLIST_REMOVE(group, next); | |
2410 | trace_vfio_put_group(group->fd); | |
2411 | close(group->fd); | |
2412 | g_free(group); | |
2413 | ||
2414 | if (QLIST_EMPTY(&vfio_group_list)) { | |
2415 | qemu_unregister_reset(vfio_reset_handler, NULL); | |
2416 | } | |
2417 | } | |
2418 | ||
2419 | int vfio_get_device(VFIOGroup *group, const char *name, | |
59f7d674 | 2420 | VFIODevice *vbasedev, Error **errp) |
e2c7d025 EA |
2421 | { |
2422 | struct vfio_device_info dev_info = { .argsz = sizeof(dev_info) }; | |
217e9fdc | 2423 | int ret, fd; |
e2c7d025 | 2424 | |
217e9fdc PB |
2425 | fd = ioctl(group->fd, VFIO_GROUP_GET_DEVICE_FD, name); |
2426 | if (fd < 0) { | |
59f7d674 EA |
2427 | error_setg_errno(errp, errno, "error getting device from group %d", |
2428 | group->groupid); | |
2429 | error_append_hint(errp, | |
2430 | "Verify all devices in group %d are bound to vfio-<bus> " | |
2431 | "or pci-stub and not already in use\n", group->groupid); | |
217e9fdc | 2432 | return fd; |
e2c7d025 EA |
2433 | } |
2434 | ||
217e9fdc | 2435 | ret = ioctl(fd, VFIO_DEVICE_GET_INFO, &dev_info); |
e2c7d025 | 2436 | if (ret) { |
59f7d674 | 2437 | error_setg_errno(errp, errno, "error getting device info"); |
217e9fdc PB |
2438 | close(fd); |
2439 | return ret; | |
e2c7d025 EA |
2440 | } |
2441 | ||
238e9172 | 2442 | /* |
aff92b82 DH |
2443 | * Set discarding of RAM as not broken for this group if the driver knows |
2444 | * the device operates compatibly with discarding. Setting must be | |
2445 | * consistent per group, but since compatibility is really only possible | |
2446 | * with mdev currently, we expect singleton groups. | |
238e9172 | 2447 | */ |
aff92b82 DH |
2448 | if (vbasedev->ram_block_discard_allowed != |
2449 | group->ram_block_discard_allowed) { | |
238e9172 | 2450 | if (!QLIST_EMPTY(&group->device_list)) { |
aff92b82 DH |
2451 | error_setg(errp, "Inconsistent setting of support for discarding " |
2452 | "RAM (e.g., balloon) within group"); | |
8709b395 | 2453 | close(fd); |
238e9172 AW |
2454 | return -1; |
2455 | } | |
2456 | ||
aff92b82 DH |
2457 | if (!group->ram_block_discard_allowed) { |
2458 | group->ram_block_discard_allowed = true; | |
53d1b5fc | 2459 | vfio_ram_block_discard_disable(group->container, false); |
238e9172 AW |
2460 | } |
2461 | } | |
2462 | ||
217e9fdc PB |
2463 | vbasedev->fd = fd; |
2464 | vbasedev->group = group; | |
2465 | QLIST_INSERT_HEAD(&group->device_list, vbasedev, next); | |
2466 | ||
e2c7d025 EA |
2467 | vbasedev->num_irqs = dev_info.num_irqs; |
2468 | vbasedev->num_regions = dev_info.num_regions; | |
2469 | vbasedev->flags = dev_info.flags; | |
2470 | ||
2471 | trace_vfio_get_device(name, dev_info.flags, dev_info.num_regions, | |
2472 | dev_info.num_irqs); | |
2473 | ||
2474 | vbasedev->reset_works = !!(dev_info.flags & VFIO_DEVICE_FLAGS_RESET); | |
217e9fdc | 2475 | return 0; |
e2c7d025 EA |
2476 | } |
2477 | ||
2478 | void vfio_put_base_device(VFIODevice *vbasedev) | |
2479 | { | |
77a10d04 PB |
2480 | if (!vbasedev->group) { |
2481 | return; | |
2482 | } | |
e2c7d025 EA |
2483 | QLIST_REMOVE(vbasedev, next); |
2484 | vbasedev->group = NULL; | |
2485 | trace_vfio_put_base_device(vbasedev->fd); | |
2486 | close(vbasedev->fd); | |
2487 | } | |
2488 | ||
46900226 AW |
2489 | int vfio_get_region_info(VFIODevice *vbasedev, int index, |
2490 | struct vfio_region_info **info) | |
2491 | { | |
2492 | size_t argsz = sizeof(struct vfio_region_info); | |
2493 | ||
2494 | *info = g_malloc0(argsz); | |
2495 | ||
2496 | (*info)->index = index; | |
b53b0f69 | 2497 | retry: |
46900226 AW |
2498 | (*info)->argsz = argsz; |
2499 | ||
2500 | if (ioctl(vbasedev->fd, VFIO_DEVICE_GET_REGION_INFO, *info)) { | |
2501 | g_free(*info); | |
e61a424f | 2502 | *info = NULL; |
46900226 AW |
2503 | return -errno; |
2504 | } | |
2505 | ||
b53b0f69 AW |
2506 | if ((*info)->argsz > argsz) { |
2507 | argsz = (*info)->argsz; | |
2508 | *info = g_realloc(*info, argsz); | |
2509 | ||
2510 | goto retry; | |
2511 | } | |
2512 | ||
46900226 AW |
2513 | return 0; |
2514 | } | |
2515 | ||
e61a424f AW |
2516 | int vfio_get_dev_region_info(VFIODevice *vbasedev, uint32_t type, |
2517 | uint32_t subtype, struct vfio_region_info **info) | |
2518 | { | |
2519 | int i; | |
2520 | ||
2521 | for (i = 0; i < vbasedev->num_regions; i++) { | |
2522 | struct vfio_info_cap_header *hdr; | |
2523 | struct vfio_region_info_cap_type *cap_type; | |
2524 | ||
2525 | if (vfio_get_region_info(vbasedev, i, info)) { | |
2526 | continue; | |
2527 | } | |
2528 | ||
2529 | hdr = vfio_get_region_info_cap(*info, VFIO_REGION_INFO_CAP_TYPE); | |
2530 | if (!hdr) { | |
2531 | g_free(*info); | |
2532 | continue; | |
2533 | } | |
2534 | ||
2535 | cap_type = container_of(hdr, struct vfio_region_info_cap_type, header); | |
2536 | ||
2537 | trace_vfio_get_dev_region(vbasedev->name, i, | |
2538 | cap_type->type, cap_type->subtype); | |
2539 | ||
2540 | if (cap_type->type == type && cap_type->subtype == subtype) { | |
2541 | return 0; | |
2542 | } | |
2543 | ||
2544 | g_free(*info); | |
2545 | } | |
2546 | ||
2547 | *info = NULL; | |
2548 | return -ENODEV; | |
2549 | } | |
2550 | ||
ae0215b2 AK |
2551 | bool vfio_has_region_cap(VFIODevice *vbasedev, int region, uint16_t cap_type) |
2552 | { | |
2553 | struct vfio_region_info *info = NULL; | |
2554 | bool ret = false; | |
2555 | ||
2556 | if (!vfio_get_region_info(vbasedev, region, &info)) { | |
2557 | if (vfio_get_region_info_cap(info, cap_type)) { | |
2558 | ret = true; | |
2559 | } | |
2560 | g_free(info); | |
2561 | } | |
2562 | ||
2563 | return ret; | |
2564 | } | |
2565 | ||
3153119e DG |
2566 | /* |
2567 | * Interfaces for IBM EEH (Enhanced Error Handling) | |
2568 | */ | |
2569 | static bool vfio_eeh_container_ok(VFIOContainer *container) | |
2570 | { | |
2571 | /* | |
2572 | * As of 2016-03-04 (linux-4.5) the host kernel EEH/VFIO | |
2573 | * implementation is broken if there are multiple groups in a | |
2574 | * container. The hardware works in units of Partitionable | |
2575 | * Endpoints (== IOMMU groups) and the EEH operations naively | |
2576 | * iterate across all groups in the container, without any logic | |
2577 | * to make sure the groups have their state synchronized. For | |
2578 | * certain operations (ENABLE) that might be ok, until an error | |
2579 | * occurs, but for others (GET_STATE) it's clearly broken. | |
2580 | */ | |
2581 | ||
2582 | /* | |
2583 | * XXX Once fixed kernels exist, test for them here | |
2584 | */ | |
2585 | ||
2586 | if (QLIST_EMPTY(&container->group_list)) { | |
2587 | return false; | |
2588 | } | |
2589 | ||
2590 | if (QLIST_NEXT(QLIST_FIRST(&container->group_list), container_next)) { | |
2591 | return false; | |
2592 | } | |
2593 | ||
2594 | return true; | |
2595 | } | |
2596 | ||
2597 | static int vfio_eeh_container_op(VFIOContainer *container, uint32_t op) | |
2598 | { | |
2599 | struct vfio_eeh_pe_op pe_op = { | |
2600 | .argsz = sizeof(pe_op), | |
2601 | .op = op, | |
2602 | }; | |
2603 | int ret; | |
2604 | ||
2605 | if (!vfio_eeh_container_ok(container)) { | |
2606 | error_report("vfio/eeh: EEH_PE_OP 0x%x: " | |
2607 | "kernel requires a container with exactly one group", op); | |
2608 | return -EPERM; | |
2609 | } | |
2610 | ||
2611 | ret = ioctl(container->fd, VFIO_EEH_PE_OP, &pe_op); | |
2612 | if (ret < 0) { | |
2613 | error_report("vfio/eeh: EEH_PE_OP 0x%x failed: %m", op); | |
2614 | return -errno; | |
2615 | } | |
2616 | ||
d917e88d | 2617 | return ret; |
3153119e DG |
2618 | } |
2619 | ||
2620 | static VFIOContainer *vfio_eeh_as_container(AddressSpace *as) | |
2621 | { | |
2622 | VFIOAddressSpace *space = vfio_get_address_space(as); | |
2623 | VFIOContainer *container = NULL; | |
2624 | ||
2625 | if (QLIST_EMPTY(&space->containers)) { | |
2626 | /* No containers to act on */ | |
2627 | goto out; | |
2628 | } | |
2629 | ||
2630 | container = QLIST_FIRST(&space->containers); | |
2631 | ||
2632 | if (QLIST_NEXT(container, next)) { | |
2633 | /* We don't yet have logic to synchronize EEH state across | |
2634 | * multiple containers */ | |
2635 | container = NULL; | |
2636 | goto out; | |
2637 | } | |
2638 | ||
2639 | out: | |
2640 | vfio_put_address_space(space); | |
2641 | return container; | |
2642 | } | |
2643 | ||
2644 | bool vfio_eeh_as_ok(AddressSpace *as) | |
2645 | { | |
2646 | VFIOContainer *container = vfio_eeh_as_container(as); | |
2647 | ||
2648 | return (container != NULL) && vfio_eeh_container_ok(container); | |
2649 | } | |
2650 | ||
2651 | int vfio_eeh_as_op(AddressSpace *as, uint32_t op) | |
2652 | { | |
2653 | VFIOContainer *container = vfio_eeh_as_container(as); | |
2654 | ||
2655 | if (!container) { | |
2656 | return -ENODEV; | |
2657 | } | |
2658 | return vfio_eeh_container_op(container, op); | |
2659 | } |