]>
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" |
236e0a45 | 45 | #include "migration/qemu-file.h" |
851d6d1a | 46 | #include "sysemu/tpm.h" |
e2c7d025 | 47 | |
f481ee2d | 48 | VFIOGroupList vfio_group_list = |
39cb514f | 49 | QLIST_HEAD_INITIALIZER(vfio_group_list); |
10ca76b4 | 50 | static QLIST_HEAD(, VFIOAddressSpace) vfio_address_spaces = |
e2c7d025 EA |
51 | QLIST_HEAD_INITIALIZER(vfio_address_spaces); |
52 | ||
53 | #ifdef CONFIG_KVM | |
54 | /* | |
55 | * We have a single VFIO pseudo device per KVM VM. Once created it lives | |
56 | * for the life of the VM. Closing the file descriptor only drops our | |
57 | * reference to it and the device's reference to kvm. Therefore once | |
58 | * initialized, this file descriptor is only released on QEMU exit and | |
59 | * we'll re-use it should another vfio device be attached before then. | |
60 | */ | |
61 | static int vfio_kvm_device_fd = -1; | |
62 | #endif | |
63 | ||
64 | /* | |
65 | * Common VFIO interrupt disable | |
66 | */ | |
67 | void vfio_disable_irqindex(VFIODevice *vbasedev, int index) | |
68 | { | |
69 | struct vfio_irq_set irq_set = { | |
70 | .argsz = sizeof(irq_set), | |
71 | .flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_TRIGGER, | |
72 | .index = index, | |
73 | .start = 0, | |
74 | .count = 0, | |
75 | }; | |
76 | ||
77 | ioctl(vbasedev->fd, VFIO_DEVICE_SET_IRQS, &irq_set); | |
78 | } | |
79 | ||
80 | void vfio_unmask_single_irqindex(VFIODevice *vbasedev, int index) | |
81 | { | |
82 | struct vfio_irq_set irq_set = { | |
83 | .argsz = sizeof(irq_set), | |
84 | .flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_UNMASK, | |
85 | .index = index, | |
86 | .start = 0, | |
87 | .count = 1, | |
88 | }; | |
89 | ||
90 | ioctl(vbasedev->fd, VFIO_DEVICE_SET_IRQS, &irq_set); | |
91 | } | |
92 | ||
93 | void vfio_mask_single_irqindex(VFIODevice *vbasedev, int index) | |
94 | { | |
95 | struct vfio_irq_set irq_set = { | |
96 | .argsz = sizeof(irq_set), | |
97 | .flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_MASK, | |
98 | .index = index, | |
99 | .start = 0, | |
100 | .count = 1, | |
101 | }; | |
102 | ||
103 | ioctl(vbasedev->fd, VFIO_DEVICE_SET_IRQS, &irq_set); | |
104 | } | |
105 | ||
201a7331 EA |
106 | static inline const char *action_to_str(int action) |
107 | { | |
108 | switch (action) { | |
109 | case VFIO_IRQ_SET_ACTION_MASK: | |
110 | return "MASK"; | |
111 | case VFIO_IRQ_SET_ACTION_UNMASK: | |
112 | return "UNMASK"; | |
113 | case VFIO_IRQ_SET_ACTION_TRIGGER: | |
114 | return "TRIGGER"; | |
115 | default: | |
116 | return "UNKNOWN ACTION"; | |
117 | } | |
118 | } | |
119 | ||
120 | static const char *index_to_str(VFIODevice *vbasedev, int index) | |
121 | { | |
122 | if (vbasedev->type != VFIO_DEVICE_TYPE_PCI) { | |
123 | return NULL; | |
124 | } | |
125 | ||
126 | switch (index) { | |
127 | case VFIO_PCI_INTX_IRQ_INDEX: | |
128 | return "INTX"; | |
129 | case VFIO_PCI_MSI_IRQ_INDEX: | |
130 | return "MSI"; | |
131 | case VFIO_PCI_MSIX_IRQ_INDEX: | |
132 | return "MSIX"; | |
133 | case VFIO_PCI_ERR_IRQ_INDEX: | |
134 | return "ERR"; | |
135 | case VFIO_PCI_REQ_IRQ_INDEX: | |
136 | return "REQ"; | |
137 | default: | |
138 | return NULL; | |
139 | } | |
140 | } | |
141 | ||
53d1b5fc DH |
142 | static int vfio_ram_block_discard_disable(VFIOContainer *container, bool state) |
143 | { | |
144 | switch (container->iommu_type) { | |
145 | case VFIO_TYPE1v2_IOMMU: | |
146 | case VFIO_TYPE1_IOMMU: | |
147 | /* | |
148 | * We support coordinated discarding of RAM via the RamDiscardManager. | |
149 | */ | |
150 | return ram_block_uncoordinated_discard_disable(state); | |
151 | default: | |
152 | /* | |
153 | * VFIO_SPAPR_TCE_IOMMU most probably works just fine with | |
154 | * RamDiscardManager, however, it is completely untested. | |
155 | * | |
156 | * VFIO_SPAPR_TCE_v2_IOMMU with "DMA memory preregistering" does | |
157 | * completely the opposite of managing mapping/pinning dynamically as | |
158 | * required by RamDiscardManager. We would have to special-case sections | |
159 | * with a RamDiscardManager. | |
160 | */ | |
161 | return ram_block_discard_disable(state); | |
162 | } | |
163 | } | |
164 | ||
201a7331 EA |
165 | int vfio_set_irq_signaling(VFIODevice *vbasedev, int index, int subindex, |
166 | int action, int fd, Error **errp) | |
167 | { | |
168 | struct vfio_irq_set *irq_set; | |
169 | int argsz, ret = 0; | |
170 | const char *name; | |
171 | int32_t *pfd; | |
172 | ||
173 | argsz = sizeof(*irq_set) + sizeof(*pfd); | |
174 | ||
175 | irq_set = g_malloc0(argsz); | |
176 | irq_set->argsz = argsz; | |
177 | irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD | action; | |
178 | irq_set->index = index; | |
179 | irq_set->start = subindex; | |
180 | irq_set->count = 1; | |
181 | pfd = (int32_t *)&irq_set->data; | |
182 | *pfd = fd; | |
183 | ||
184 | if (ioctl(vbasedev->fd, VFIO_DEVICE_SET_IRQS, irq_set)) { | |
185 | ret = -errno; | |
186 | } | |
187 | g_free(irq_set); | |
188 | ||
189 | if (!ret) { | |
190 | return 0; | |
191 | } | |
192 | ||
193 | error_setg_errno(errp, -ret, "VFIO_DEVICE_SET_IRQS failure"); | |
194 | ||
195 | name = index_to_str(vbasedev, index); | |
196 | if (name) { | |
197 | error_prepend(errp, "%s-%d: ", name, subindex); | |
198 | } else { | |
199 | error_prepend(errp, "index %d-%d: ", index, subindex); | |
200 | } | |
201 | error_prepend(errp, | |
202 | "Failed to %s %s eventfd signaling for interrupt ", | |
203 | fd < 0 ? "tear down" : "set up", action_to_str(action)); | |
204 | return ret; | |
205 | } | |
206 | ||
e2c7d025 EA |
207 | /* |
208 | * IO Port/MMIO - Beware of the endians, VFIO is always little endian | |
209 | */ | |
210 | void vfio_region_write(void *opaque, hwaddr addr, | |
211 | uint64_t data, unsigned size) | |
212 | { | |
213 | VFIORegion *region = opaque; | |
214 | VFIODevice *vbasedev = region->vbasedev; | |
215 | union { | |
216 | uint8_t byte; | |
217 | uint16_t word; | |
218 | uint32_t dword; | |
219 | uint64_t qword; | |
220 | } buf; | |
221 | ||
222 | switch (size) { | |
223 | case 1: | |
224 | buf.byte = data; | |
225 | break; | |
226 | case 2: | |
227 | buf.word = cpu_to_le16(data); | |
228 | break; | |
229 | case 4: | |
230 | buf.dword = cpu_to_le32(data); | |
231 | break; | |
38d49e8c JRZ |
232 | case 8: |
233 | buf.qword = cpu_to_le64(data); | |
234 | break; | |
e2c7d025 | 235 | default: |
c624b6b3 | 236 | hw_error("vfio: unsupported write size, %u bytes", size); |
e2c7d025 EA |
237 | break; |
238 | } | |
239 | ||
240 | if (pwrite(vbasedev->fd, &buf, size, region->fd_offset + addr) != size) { | |
241 | error_report("%s(%s:region%d+0x%"HWADDR_PRIx", 0x%"PRIx64 | |
242 | ",%d) failed: %m", | |
243 | __func__, vbasedev->name, region->nr, | |
244 | addr, data, size); | |
245 | } | |
246 | ||
247 | trace_vfio_region_write(vbasedev->name, region->nr, addr, data, size); | |
248 | ||
249 | /* | |
250 | * A read or write to a BAR always signals an INTx EOI. This will | |
251 | * do nothing if not pending (including not in INTx mode). We assume | |
252 | * that a BAR access is in response to an interrupt and that BAR | |
253 | * accesses will service the interrupt. Unfortunately, we don't know | |
254 | * which access will service the interrupt, so we're potentially | |
255 | * getting quite a few host interrupts per guest interrupt. | |
256 | */ | |
257 | vbasedev->ops->vfio_eoi(vbasedev); | |
258 | } | |
259 | ||
260 | uint64_t vfio_region_read(void *opaque, | |
261 | hwaddr addr, unsigned size) | |
262 | { | |
263 | VFIORegion *region = opaque; | |
264 | VFIODevice *vbasedev = region->vbasedev; | |
265 | union { | |
266 | uint8_t byte; | |
267 | uint16_t word; | |
268 | uint32_t dword; | |
269 | uint64_t qword; | |
270 | } buf; | |
271 | uint64_t data = 0; | |
272 | ||
273 | if (pread(vbasedev->fd, &buf, size, region->fd_offset + addr) != size) { | |
274 | error_report("%s(%s:region%d+0x%"HWADDR_PRIx", %d) failed: %m", | |
275 | __func__, vbasedev->name, region->nr, | |
276 | addr, size); | |
277 | return (uint64_t)-1; | |
278 | } | |
279 | switch (size) { | |
280 | case 1: | |
281 | data = buf.byte; | |
282 | break; | |
283 | case 2: | |
284 | data = le16_to_cpu(buf.word); | |
285 | break; | |
286 | case 4: | |
287 | data = le32_to_cpu(buf.dword); | |
288 | break; | |
38d49e8c JRZ |
289 | case 8: |
290 | data = le64_to_cpu(buf.qword); | |
291 | break; | |
e2c7d025 | 292 | default: |
c624b6b3 | 293 | hw_error("vfio: unsupported read size, %u bytes", size); |
e2c7d025 EA |
294 | break; |
295 | } | |
296 | ||
297 | trace_vfio_region_read(vbasedev->name, region->nr, addr, size, data); | |
298 | ||
299 | /* Same as write above */ | |
300 | vbasedev->ops->vfio_eoi(vbasedev); | |
301 | ||
302 | return data; | |
303 | } | |
304 | ||
305 | const MemoryRegionOps vfio_region_ops = { | |
306 | .read = vfio_region_read, | |
307 | .write = vfio_region_write, | |
308 | .endianness = DEVICE_LITTLE_ENDIAN, | |
15126cba JRZ |
309 | .valid = { |
310 | .min_access_size = 1, | |
311 | .max_access_size = 8, | |
312 | }, | |
38d49e8c JRZ |
313 | .impl = { |
314 | .min_access_size = 1, | |
315 | .max_access_size = 8, | |
316 | }, | |
e2c7d025 EA |
317 | }; |
318 | ||
b6dd6504 KW |
319 | /* |
320 | * Device state interfaces | |
321 | */ | |
322 | ||
725ccd7e AH |
323 | typedef struct { |
324 | unsigned long *bitmap; | |
325 | hwaddr size; | |
326 | hwaddr pages; | |
327 | } VFIOBitmap; | |
328 | ||
329 | static int vfio_bitmap_alloc(VFIOBitmap *vbmap, hwaddr size) | |
330 | { | |
331 | vbmap->pages = REAL_HOST_PAGE_ALIGN(size) / qemu_real_host_page_size(); | |
332 | vbmap->size = ROUND_UP(vbmap->pages, sizeof(__u64) * BITS_PER_BYTE) / | |
333 | BITS_PER_BYTE; | |
334 | vbmap->bitmap = g_try_malloc0(vbmap->size); | |
335 | if (!vbmap->bitmap) { | |
336 | return -ENOMEM; | |
337 | } | |
338 | ||
339 | return 0; | |
340 | } | |
341 | ||
3710586c KW |
342 | bool vfio_mig_active(void) |
343 | { | |
344 | VFIOGroup *group; | |
345 | VFIODevice *vbasedev; | |
346 | ||
347 | if (QLIST_EMPTY(&vfio_group_list)) { | |
348 | return false; | |
349 | } | |
350 | ||
351 | QLIST_FOREACH(group, &vfio_group_list, next) { | |
352 | QLIST_FOREACH(vbasedev, &group->device_list, next) { | |
353 | if (vbasedev->migration_blocker) { | |
354 | return false; | |
355 | } | |
356 | } | |
357 | } | |
358 | return true; | |
359 | } | |
360 | ||
29d81b71 AH |
361 | static Error *multiple_devices_migration_blocker; |
362 | ||
363 | static unsigned int vfio_migratable_device_num(void) | |
364 | { | |
365 | VFIOGroup *group; | |
366 | VFIODevice *vbasedev; | |
367 | unsigned int device_num = 0; | |
368 | ||
369 | QLIST_FOREACH(group, &vfio_group_list, next) { | |
370 | QLIST_FOREACH(vbasedev, &group->device_list, next) { | |
371 | if (vbasedev->migration) { | |
372 | device_num++; | |
373 | } | |
374 | } | |
375 | } | |
376 | ||
377 | return device_num; | |
378 | } | |
379 | ||
380 | int vfio_block_multiple_devices_migration(Error **errp) | |
381 | { | |
382 | int ret; | |
383 | ||
384 | if (multiple_devices_migration_blocker || | |
385 | vfio_migratable_device_num() <= 1) { | |
386 | return 0; | |
387 | } | |
388 | ||
389 | error_setg(&multiple_devices_migration_blocker, | |
390 | "Migration is currently not supported with multiple " | |
391 | "VFIO devices"); | |
392 | ret = migrate_add_blocker(multiple_devices_migration_blocker, errp); | |
393 | if (ret < 0) { | |
394 | error_free(multiple_devices_migration_blocker); | |
395 | multiple_devices_migration_blocker = NULL; | |
396 | } | |
397 | ||
398 | return ret; | |
399 | } | |
400 | ||
401 | void vfio_unblock_multiple_devices_migration(void) | |
402 | { | |
403 | if (!multiple_devices_migration_blocker || | |
404 | vfio_migratable_device_num() > 1) { | |
405 | return; | |
406 | } | |
407 | ||
408 | migrate_del_blocker(multiple_devices_migration_blocker); | |
409 | error_free(multiple_devices_migration_blocker); | |
410 | multiple_devices_migration_blocker = NULL; | |
411 | } | |
412 | ||
236e0a45 AH |
413 | static void vfio_set_migration_error(int err) |
414 | { | |
415 | MigrationState *ms = migrate_get_current(); | |
416 | ||
417 | if (migration_is_setup_or_active(ms->state)) { | |
418 | WITH_QEMU_LOCK_GUARD(&ms->qemu_file_lock) { | |
419 | if (ms->to_dst_file) { | |
420 | qemu_file_set_error(ms->to_dst_file, err); | |
421 | } | |
422 | } | |
423 | } | |
424 | } | |
425 | ||
758b96b6 | 426 | static bool vfio_devices_all_dirty_tracking(VFIOContainer *container) |
b6dd6504 KW |
427 | { |
428 | VFIOGroup *group; | |
429 | VFIODevice *vbasedev; | |
430 | MigrationState *ms = migrate_get_current(); | |
431 | ||
432 | if (!migration_is_setup_or_active(ms->state)) { | |
433 | return false; | |
434 | } | |
435 | ||
436 | QLIST_FOREACH(group, &container->group_list, container_next) { | |
437 | QLIST_FOREACH(vbasedev, &group->device_list, next) { | |
438 | VFIOMigration *migration = vbasedev->migration; | |
439 | ||
440 | if (!migration) { | |
441 | return false; | |
442 | } | |
443 | ||
7429aebe | 444 | if (vbasedev->pre_copy_dirty_page_tracking == ON_OFF_AUTO_OFF && |
31bcbbb5 AH |
445 | migration->device_state == VFIO_DEVICE_STATE_RUNNING) { |
446 | return false; | |
447 | } | |
b6dd6504 KW |
448 | } |
449 | } | |
450 | return true; | |
451 | } | |
452 | ||
8b942af3 AH |
453 | /* |
454 | * Check if all VFIO devices are running and migration is active, which is | |
455 | * essentially equivalent to the migration being in pre-copy phase. | |
456 | */ | |
457 | static bool vfio_devices_all_running_and_mig_active(VFIOContainer *container) | |
9e7b0442 KW |
458 | { |
459 | VFIOGroup *group; | |
460 | VFIODevice *vbasedev; | |
9e7b0442 | 461 | |
8b942af3 | 462 | if (!migration_is_active(migrate_get_current())) { |
9e7b0442 KW |
463 | return false; |
464 | } | |
465 | ||
466 | QLIST_FOREACH(group, &container->group_list, container_next) { | |
467 | QLIST_FOREACH(vbasedev, &group->device_list, next) { | |
468 | VFIOMigration *migration = vbasedev->migration; | |
469 | ||
470 | if (!migration) { | |
471 | return false; | |
472 | } | |
473 | ||
7429aebe | 474 | if (migration->device_state == VFIO_DEVICE_STATE_RUNNING) { |
9e7b0442 KW |
475 | continue; |
476 | } else { | |
477 | return false; | |
478 | } | |
479 | } | |
480 | } | |
481 | return true; | |
482 | } | |
483 | ||
484 | static int vfio_dma_unmap_bitmap(VFIOContainer *container, | |
485 | hwaddr iova, ram_addr_t size, | |
486 | IOMMUTLBEntry *iotlb) | |
487 | { | |
488 | struct vfio_iommu_type1_dma_unmap *unmap; | |
489 | struct vfio_bitmap *bitmap; | |
725ccd7e | 490 | VFIOBitmap vbmap; |
9e7b0442 KW |
491 | int ret; |
492 | ||
725ccd7e AH |
493 | ret = vfio_bitmap_alloc(&vbmap, size); |
494 | if (ret) { | |
495 | return ret; | |
496 | } | |
497 | ||
9e7b0442 KW |
498 | unmap = g_malloc0(sizeof(*unmap) + sizeof(*bitmap)); |
499 | ||
500 | unmap->argsz = sizeof(*unmap) + sizeof(*bitmap); | |
501 | unmap->iova = iova; | |
502 | unmap->size = size; | |
503 | unmap->flags |= VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP; | |
504 | bitmap = (struct vfio_bitmap *)&unmap->data; | |
505 | ||
506 | /* | |
1eb7f642 KJ |
507 | * cpu_physical_memory_set_dirty_lebitmap() supports pages in bitmap of |
508 | * qemu_real_host_page_size to mark those dirty. Hence set bitmap_pgsize | |
509 | * to qemu_real_host_page_size. | |
9e7b0442 | 510 | */ |
8e3b0cbb | 511 | bitmap->pgsize = qemu_real_host_page_size(); |
725ccd7e AH |
512 | bitmap->size = vbmap.size; |
513 | bitmap->data = (__u64 *)vbmap.bitmap; | |
9e7b0442 | 514 | |
725ccd7e AH |
515 | if (vbmap.size > container->max_dirty_bitmap_size) { |
516 | error_report("UNMAP: Size of bitmap too big 0x%"PRIx64, vbmap.size); | |
9e7b0442 KW |
517 | ret = -E2BIG; |
518 | goto unmap_exit; | |
519 | } | |
520 | ||
9e7b0442 KW |
521 | ret = ioctl(container->fd, VFIO_IOMMU_UNMAP_DMA, unmap); |
522 | if (!ret) { | |
725ccd7e AH |
523 | cpu_physical_memory_set_dirty_lebitmap(vbmap.bitmap, |
524 | iotlb->translated_addr, vbmap.pages); | |
9e7b0442 KW |
525 | } else { |
526 | error_report("VFIO_UNMAP_DMA with DIRTY_BITMAP : %m"); | |
527 | } | |
528 | ||
9e7b0442 KW |
529 | unmap_exit: |
530 | g_free(unmap); | |
725ccd7e AH |
531 | g_free(vbmap.bitmap); |
532 | ||
9e7b0442 KW |
533 | return ret; |
534 | } | |
535 | ||
e2c7d025 EA |
536 | /* |
537 | * DMA - Mapping and unmapping for the "type1" IOMMU interface used on x86 | |
538 | */ | |
539 | static int vfio_dma_unmap(VFIOContainer *container, | |
9e7b0442 KW |
540 | hwaddr iova, ram_addr_t size, |
541 | IOMMUTLBEntry *iotlb) | |
e2c7d025 EA |
542 | { |
543 | struct vfio_iommu_type1_dma_unmap unmap = { | |
544 | .argsz = sizeof(unmap), | |
545 | .flags = 0, | |
546 | .iova = iova, | |
547 | .size = size, | |
548 | }; | |
549 | ||
9e7b0442 | 550 | if (iotlb && container->dirty_pages_supported && |
8b942af3 | 551 | vfio_devices_all_running_and_mig_active(container)) { |
9e7b0442 KW |
552 | return vfio_dma_unmap_bitmap(container, iova, size, iotlb); |
553 | } | |
554 | ||
567d7d3e AW |
555 | while (ioctl(container->fd, VFIO_IOMMU_UNMAP_DMA, &unmap)) { |
556 | /* | |
557 | * The type1 backend has an off-by-one bug in the kernel (71a7d3d78e3c | |
558 | * v4.15) where an overflow in its wrap-around check prevents us from | |
559 | * unmapping the last page of the address space. Test for the error | |
560 | * condition and re-try the unmap excluding the last page. The | |
561 | * expectation is that we've never mapped the last page anyway and this | |
562 | * unmap request comes via vIOMMU support which also makes it unlikely | |
563 | * that this page is used. This bug was introduced well after type1 v2 | |
564 | * support was introduced, so we shouldn't need to test for v1. A fix | |
565 | * is queued for kernel v5.0 so this workaround can be removed once | |
566 | * affected kernels are sufficiently deprecated. | |
567 | */ | |
568 | if (errno == EINVAL && unmap.size && !(unmap.iova + unmap.size) && | |
569 | container->iommu_type == VFIO_TYPE1v2_IOMMU) { | |
570 | trace_vfio_dma_unmap_overflow_workaround(); | |
571 | unmap.size -= 1ULL << ctz64(container->pgsizes); | |
572 | continue; | |
573 | } | |
b09d51c9 | 574 | error_report("VFIO_UNMAP_DMA failed: %s", strerror(errno)); |
e2c7d025 EA |
575 | return -errno; |
576 | } | |
577 | ||
8b942af3 | 578 | if (iotlb && vfio_devices_all_running_and_mig_active(container)) { |
b051a3f6 AH |
579 | cpu_physical_memory_set_dirty_range(iotlb->translated_addr, size, |
580 | tcg_enabled() ? DIRTY_CLIENTS_ALL : | |
581 | DIRTY_CLIENTS_NOCODE); | |
582 | } | |
583 | ||
e2c7d025 EA |
584 | return 0; |
585 | } | |
586 | ||
587 | static int vfio_dma_map(VFIOContainer *container, hwaddr iova, | |
588 | ram_addr_t size, void *vaddr, bool readonly) | |
589 | { | |
590 | struct vfio_iommu_type1_dma_map map = { | |
591 | .argsz = sizeof(map), | |
592 | .flags = VFIO_DMA_MAP_FLAG_READ, | |
593 | .vaddr = (__u64)(uintptr_t)vaddr, | |
594 | .iova = iova, | |
595 | .size = size, | |
596 | }; | |
597 | ||
598 | if (!readonly) { | |
599 | map.flags |= VFIO_DMA_MAP_FLAG_WRITE; | |
600 | } | |
601 | ||
602 | /* | |
603 | * Try the mapping, if it fails with EBUSY, unmap the region and try | |
604 | * again. This shouldn't be necessary, but we sometimes see it in | |
b6af0975 | 605 | * the VGA ROM space. |
e2c7d025 EA |
606 | */ |
607 | if (ioctl(container->fd, VFIO_IOMMU_MAP_DMA, &map) == 0 || | |
9e7b0442 | 608 | (errno == EBUSY && vfio_dma_unmap(container, iova, size, NULL) == 0 && |
e2c7d025 EA |
609 | ioctl(container->fd, VFIO_IOMMU_MAP_DMA, &map) == 0)) { |
610 | return 0; | |
611 | } | |
612 | ||
b09d51c9 | 613 | error_report("VFIO_MAP_DMA failed: %s", strerror(errno)); |
e2c7d025 EA |
614 | return -errno; |
615 | } | |
616 | ||
f4ec5e26 AK |
617 | static void vfio_host_win_add(VFIOContainer *container, |
618 | hwaddr min_iova, hwaddr max_iova, | |
619 | uint64_t iova_pgsizes) | |
620 | { | |
621 | VFIOHostDMAWindow *hostwin; | |
622 | ||
623 | QLIST_FOREACH(hostwin, &container->hostwin_list, hostwin_next) { | |
624 | if (ranges_overlap(hostwin->min_iova, | |
625 | hostwin->max_iova - hostwin->min_iova + 1, | |
626 | min_iova, | |
627 | max_iova - min_iova + 1)) { | |
628 | hw_error("%s: Overlapped IOMMU are not enabled", __func__); | |
629 | } | |
630 | } | |
631 | ||
632 | hostwin = g_malloc0(sizeof(*hostwin)); | |
633 | ||
634 | hostwin->min_iova = min_iova; | |
635 | hostwin->max_iova = max_iova; | |
636 | hostwin->iova_pgsizes = iova_pgsizes; | |
637 | QLIST_INSERT_HEAD(&container->hostwin_list, hostwin, hostwin_next); | |
638 | } | |
639 | ||
2e4109de AK |
640 | static int vfio_host_win_del(VFIOContainer *container, hwaddr min_iova, |
641 | hwaddr max_iova) | |
642 | { | |
643 | VFIOHostDMAWindow *hostwin; | |
644 | ||
645 | QLIST_FOREACH(hostwin, &container->hostwin_list, hostwin_next) { | |
646 | if (hostwin->min_iova == min_iova && hostwin->max_iova == max_iova) { | |
647 | QLIST_REMOVE(hostwin, hostwin_next); | |
f3bc3a73 | 648 | g_free(hostwin); |
2e4109de AK |
649 | return 0; |
650 | } | |
651 | } | |
652 | ||
653 | return -1; | |
654 | } | |
655 | ||
e2c7d025 EA |
656 | static bool vfio_listener_skipped_section(MemoryRegionSection *section) |
657 | { | |
658 | return (!memory_region_is_ram(section->mr) && | |
659 | !memory_region_is_iommu(section->mr)) || | |
56918a12 | 660 | memory_region_is_protected(section->mr) || |
e2c7d025 EA |
661 | /* |
662 | * Sizing an enabled 64-bit BAR can cause spurious mappings to | |
663 | * addresses in the upper part of the 64-bit address space. These | |
664 | * are never accessed by the CPU and beyond the address width of | |
665 | * some IOMMU hardware. TODO: VFIO should tell us the IOMMU width. | |
666 | */ | |
667 | section->offset_within_address_space & (1ULL << 63); | |
668 | } | |
669 | ||
4a4b88fb | 670 | /* Called with rcu_read_lock held. */ |
9a04fe09 KW |
671 | static bool vfio_get_xlat_addr(IOMMUTLBEntry *iotlb, void **vaddr, |
672 | ram_addr_t *ram_addr, bool *read_only) | |
e2c7d025 | 673 | { |
baa44bce | 674 | bool ret, mr_has_discard_manager; |
0fd7616e | 675 | |
baa44bce CL |
676 | ret = memory_get_xlat_addr(iotlb, vaddr, ram_addr, read_only, |
677 | &mr_has_discard_manager); | |
678 | if (ret && mr_has_discard_manager) { | |
0fd7616e DH |
679 | /* |
680 | * Malicious VMs might trigger discarding of IOMMU-mapped memory. The | |
681 | * pages will remain pinned inside vfio until unmapped, resulting in a | |
682 | * higher memory consumption than expected. If memory would get | |
683 | * populated again later, there would be an inconsistency between pages | |
684 | * pinned by vfio and pages seen by QEMU. This is the case until | |
685 | * unmapped from the IOMMU (e.g., during device reset). | |
686 | * | |
687 | * With malicious guests, we really only care about pinning more memory | |
688 | * than expected. RLIMIT_MEMLOCK set for the user/process can never be | |
689 | * exceeded and can be used to mitigate this problem. | |
690 | */ | |
691 | warn_report_once("Using vfio with vIOMMUs and coordinated discarding of" | |
692 | " RAM (e.g., virtio-mem) works, however, malicious" | |
693 | " guests can trigger pinning of more memory than" | |
694 | " intended via an IOMMU. It's possible to mitigate " | |
695 | " by setting/adjusting RLIMIT_MEMLOCK."); | |
e2c7d025 | 696 | } |
baa44bce | 697 | return ret; |
4a4b88fb PX |
698 | } |
699 | ||
700 | static void vfio_iommu_map_notify(IOMMUNotifier *n, IOMMUTLBEntry *iotlb) | |
701 | { | |
702 | VFIOGuestIOMMU *giommu = container_of(n, VFIOGuestIOMMU, n); | |
703 | VFIOContainer *container = giommu->container; | |
704 | hwaddr iova = iotlb->iova + giommu->iommu_offset; | |
4a4b88fb PX |
705 | void *vaddr; |
706 | int ret; | |
707 | ||
708 | trace_vfio_iommu_map_notify(iotlb->perm == IOMMU_NONE ? "UNMAP" : "MAP", | |
709 | iova, iova + iotlb->addr_mask); | |
710 | ||
711 | if (iotlb->target_as != &address_space_memory) { | |
712 | error_report("Wrong target AS \"%s\", only system memory is allowed", | |
713 | iotlb->target_as->name ? iotlb->target_as->name : "none"); | |
236e0a45 | 714 | vfio_set_migration_error(-EINVAL); |
4a4b88fb PX |
715 | return; |
716 | } | |
717 | ||
718 | rcu_read_lock(); | |
719 | ||
e2c7d025 | 720 | if ((iotlb->perm & IOMMU_RW) != IOMMU_NONE) { |
9a04fe09 KW |
721 | bool read_only; |
722 | ||
723 | if (!vfio_get_xlat_addr(iotlb, &vaddr, NULL, &read_only)) { | |
dfbd90e5 PX |
724 | goto out; |
725 | } | |
4a4b88fb PX |
726 | /* |
727 | * vaddr is only valid until rcu_read_unlock(). But after | |
728 | * vfio_dma_map has set up the mapping the pages will be | |
729 | * pinned by the kernel. This makes sure that the RAM backend | |
730 | * of vaddr will always be there, even if the memory object is | |
731 | * destroyed and its backing memory munmap-ed. | |
732 | */ | |
d78c19b5 | 733 | ret = vfio_dma_map(container, iova, |
e2c7d025 | 734 | iotlb->addr_mask + 1, vaddr, |
4a4b88fb | 735 | read_only); |
e2c7d025 EA |
736 | if (ret) { |
737 | error_report("vfio_dma_map(%p, 0x%"HWADDR_PRIx", " | |
db9b829b | 738 | "0x%"HWADDR_PRIx", %p) = %d (%s)", |
d78c19b5 | 739 | container, iova, |
db9b829b | 740 | iotlb->addr_mask + 1, vaddr, ret, strerror(-ret)); |
e2c7d025 EA |
741 | } |
742 | } else { | |
9e7b0442 | 743 | ret = vfio_dma_unmap(container, iova, iotlb->addr_mask + 1, iotlb); |
e2c7d025 EA |
744 | if (ret) { |
745 | error_report("vfio_dma_unmap(%p, 0x%"HWADDR_PRIx", " | |
db9b829b | 746 | "0x%"HWADDR_PRIx") = %d (%s)", |
d78c19b5 | 747 | container, iova, |
db9b829b | 748 | iotlb->addr_mask + 1, ret, strerror(-ret)); |
236e0a45 | 749 | vfio_set_migration_error(ret); |
e2c7d025 EA |
750 | } |
751 | } | |
41063e1e PB |
752 | out: |
753 | rcu_read_unlock(); | |
e2c7d025 EA |
754 | } |
755 | ||
5e3b981c DH |
756 | static void vfio_ram_discard_notify_discard(RamDiscardListener *rdl, |
757 | MemoryRegionSection *section) | |
758 | { | |
759 | VFIORamDiscardListener *vrdl = container_of(rdl, VFIORamDiscardListener, | |
760 | listener); | |
761 | const hwaddr size = int128_get64(section->size); | |
762 | const hwaddr iova = section->offset_within_address_space; | |
763 | int ret; | |
764 | ||
765 | /* Unmap with a single call. */ | |
766 | ret = vfio_dma_unmap(vrdl->container, iova, size , NULL); | |
767 | if (ret) { | |
768 | error_report("%s: vfio_dma_unmap() failed: %s", __func__, | |
769 | strerror(-ret)); | |
770 | } | |
771 | } | |
772 | ||
773 | static int vfio_ram_discard_notify_populate(RamDiscardListener *rdl, | |
774 | MemoryRegionSection *section) | |
775 | { | |
776 | VFIORamDiscardListener *vrdl = container_of(rdl, VFIORamDiscardListener, | |
777 | listener); | |
778 | const hwaddr end = section->offset_within_region + | |
779 | int128_get64(section->size); | |
780 | hwaddr start, next, iova; | |
781 | void *vaddr; | |
782 | int ret; | |
783 | ||
784 | /* | |
785 | * Map in (aligned within memory region) minimum granularity, so we can | |
786 | * unmap in minimum granularity later. | |
787 | */ | |
788 | for (start = section->offset_within_region; start < end; start = next) { | |
789 | next = ROUND_UP(start + 1, vrdl->granularity); | |
790 | next = MIN(next, end); | |
791 | ||
792 | iova = start - section->offset_within_region + | |
793 | section->offset_within_address_space; | |
794 | vaddr = memory_region_get_ram_ptr(section->mr) + start; | |
795 | ||
796 | ret = vfio_dma_map(vrdl->container, iova, next - start, | |
797 | vaddr, section->readonly); | |
798 | if (ret) { | |
799 | /* Rollback */ | |
800 | vfio_ram_discard_notify_discard(rdl, section); | |
801 | return ret; | |
802 | } | |
803 | } | |
804 | return 0; | |
805 | } | |
806 | ||
807 | static void vfio_register_ram_discard_listener(VFIOContainer *container, | |
808 | MemoryRegionSection *section) | |
809 | { | |
810 | RamDiscardManager *rdm = memory_region_get_ram_discard_manager(section->mr); | |
811 | VFIORamDiscardListener *vrdl; | |
812 | ||
813 | /* Ignore some corner cases not relevant in practice. */ | |
814 | g_assert(QEMU_IS_ALIGNED(section->offset_within_region, TARGET_PAGE_SIZE)); | |
815 | g_assert(QEMU_IS_ALIGNED(section->offset_within_address_space, | |
816 | TARGET_PAGE_SIZE)); | |
817 | g_assert(QEMU_IS_ALIGNED(int128_get64(section->size), TARGET_PAGE_SIZE)); | |
818 | ||
819 | vrdl = g_new0(VFIORamDiscardListener, 1); | |
820 | vrdl->container = container; | |
821 | vrdl->mr = section->mr; | |
822 | vrdl->offset_within_address_space = section->offset_within_address_space; | |
823 | vrdl->size = int128_get64(section->size); | |
824 | vrdl->granularity = ram_discard_manager_get_min_granularity(rdm, | |
825 | section->mr); | |
826 | ||
827 | g_assert(vrdl->granularity && is_power_of_2(vrdl->granularity)); | |
a5dba9bc DH |
828 | g_assert(container->pgsizes && |
829 | vrdl->granularity >= 1ULL << ctz64(container->pgsizes)); | |
5e3b981c DH |
830 | |
831 | ram_discard_listener_init(&vrdl->listener, | |
832 | vfio_ram_discard_notify_populate, | |
833 | vfio_ram_discard_notify_discard, true); | |
834 | ram_discard_manager_register_listener(rdm, &vrdl->listener, section); | |
835 | QLIST_INSERT_HEAD(&container->vrdl_list, vrdl, next); | |
a74317f6 DH |
836 | |
837 | /* | |
838 | * Sanity-check if we have a theoretically problematic setup where we could | |
839 | * exceed the maximum number of possible DMA mappings over time. We assume | |
840 | * that each mapped section in the same address space as a RamDiscardManager | |
841 | * section consumes exactly one DMA mapping, with the exception of | |
842 | * RamDiscardManager sections; i.e., we don't expect to have gIOMMU sections | |
843 | * in the same address space as RamDiscardManager sections. | |
844 | * | |
845 | * We assume that each section in the address space consumes one memslot. | |
846 | * We take the number of KVM memory slots as a best guess for the maximum | |
847 | * number of sections in the address space we could have over time, | |
848 | * also consuming DMA mappings. | |
849 | */ | |
850 | if (container->dma_max_mappings) { | |
851 | unsigned int vrdl_count = 0, vrdl_mappings = 0, max_memslots = 512; | |
852 | ||
853 | #ifdef CONFIG_KVM | |
854 | if (kvm_enabled()) { | |
855 | max_memslots = kvm_get_max_memslots(); | |
856 | } | |
857 | #endif | |
858 | ||
859 | QLIST_FOREACH(vrdl, &container->vrdl_list, next) { | |
860 | hwaddr start, end; | |
861 | ||
862 | start = QEMU_ALIGN_DOWN(vrdl->offset_within_address_space, | |
863 | vrdl->granularity); | |
864 | end = ROUND_UP(vrdl->offset_within_address_space + vrdl->size, | |
865 | vrdl->granularity); | |
866 | vrdl_mappings += (end - start) / vrdl->granularity; | |
867 | vrdl_count++; | |
868 | } | |
869 | ||
870 | if (vrdl_mappings + max_memslots - vrdl_count > | |
871 | container->dma_max_mappings) { | |
872 | warn_report("%s: possibly running out of DMA mappings. E.g., try" | |
873 | " increasing the 'block-size' of virtio-mem devies." | |
874 | " Maximum possible DMA mappings: %d, Maximum possible" | |
875 | " memslots: %d", __func__, container->dma_max_mappings, | |
876 | max_memslots); | |
877 | } | |
878 | } | |
5e3b981c DH |
879 | } |
880 | ||
881 | static void vfio_unregister_ram_discard_listener(VFIOContainer *container, | |
882 | MemoryRegionSection *section) | |
883 | { | |
884 | RamDiscardManager *rdm = memory_region_get_ram_discard_manager(section->mr); | |
885 | VFIORamDiscardListener *vrdl = NULL; | |
886 | ||
887 | QLIST_FOREACH(vrdl, &container->vrdl_list, next) { | |
888 | if (vrdl->mr == section->mr && | |
889 | vrdl->offset_within_address_space == | |
890 | section->offset_within_address_space) { | |
891 | break; | |
892 | } | |
893 | } | |
894 | ||
895 | if (!vrdl) { | |
896 | hw_error("vfio: Trying to unregister missing RAM discard listener"); | |
897 | } | |
898 | ||
899 | ram_discard_manager_unregister_listener(rdm, &vrdl->listener); | |
900 | QLIST_REMOVE(vrdl, next); | |
901 | g_free(vrdl); | |
902 | } | |
903 | ||
fbc6c921 JM |
904 | static VFIOHostDMAWindow *vfio_find_hostwin(VFIOContainer *container, |
905 | hwaddr iova, hwaddr end) | |
906 | { | |
907 | VFIOHostDMAWindow *hostwin; | |
908 | bool hostwin_found = false; | |
909 | ||
910 | QLIST_FOREACH(hostwin, &container->hostwin_list, hostwin_next) { | |
911 | if (hostwin->min_iova <= iova && end <= hostwin->max_iova) { | |
912 | hostwin_found = true; | |
913 | break; | |
914 | } | |
915 | } | |
916 | ||
917 | return hostwin_found ? hostwin : NULL; | |
918 | } | |
919 | ||
851d6d1a EA |
920 | static bool vfio_known_safe_misalignment(MemoryRegionSection *section) |
921 | { | |
922 | MemoryRegion *mr = section->mr; | |
923 | ||
924 | if (!TPM_IS_CRB(mr->owner)) { | |
925 | return false; | |
926 | } | |
927 | ||
928 | /* this is a known safe misaligned region, just trace for debug purpose */ | |
929 | trace_vfio_known_safe_misalignment(memory_region_name(mr), | |
930 | section->offset_within_address_space, | |
931 | section->offset_within_region, | |
932 | qemu_real_host_page_size()); | |
933 | return true; | |
934 | } | |
935 | ||
b92f2376 JM |
936 | static bool vfio_listener_valid_section(MemoryRegionSection *section, |
937 | const char *name) | |
e2c7d025 | 938 | { |
e2c7d025 | 939 | if (vfio_listener_skipped_section(section)) { |
b92f2376 | 940 | trace_vfio_listener_region_skip(name, |
e2c7d025 EA |
941 | section->offset_within_address_space, |
942 | section->offset_within_address_space + | |
943 | int128_get64(int128_sub(section->size, int128_one()))); | |
b92f2376 | 944 | return false; |
e2c7d025 EA |
945 | } |
946 | ||
1eb7f642 | 947 | if (unlikely((section->offset_within_address_space & |
8e3b0cbb MAL |
948 | ~qemu_real_host_page_mask()) != |
949 | (section->offset_within_region & ~qemu_real_host_page_mask()))) { | |
851d6d1a EA |
950 | if (!vfio_known_safe_misalignment(section)) { |
951 | error_report("%s received unaligned region %s iova=0x%"PRIx64 | |
952 | " offset_within_region=0x%"PRIx64 | |
953 | " qemu_real_host_page_size=0x%"PRIxPTR, | |
954 | __func__, memory_region_name(section->mr), | |
955 | section->offset_within_address_space, | |
956 | section->offset_within_region, | |
957 | qemu_real_host_page_size()); | |
958 | } | |
b92f2376 JM |
959 | return false; |
960 | } | |
961 | ||
962 | return true; | |
963 | } | |
964 | ||
965 | static void vfio_listener_region_add(MemoryListener *listener, | |
966 | MemoryRegionSection *section) | |
967 | { | |
968 | VFIOContainer *container = container_of(listener, VFIOContainer, listener); | |
969 | hwaddr iova, end; | |
970 | Int128 llend, llsize; | |
971 | void *vaddr; | |
972 | int ret; | |
973 | VFIOHostDMAWindow *hostwin; | |
974 | Error *err = NULL; | |
975 | ||
976 | if (!vfio_listener_valid_section(section, "region_add")) { | |
e2c7d025 EA |
977 | return; |
978 | } | |
979 | ||
1eb7f642 | 980 | iova = REAL_HOST_PAGE_ALIGN(section->offset_within_address_space); |
e2c7d025 EA |
981 | llend = int128_make64(section->offset_within_address_space); |
982 | llend = int128_add(llend, section->size); | |
8e3b0cbb | 983 | llend = int128_and(llend, int128_exts64(qemu_real_host_page_mask())); |
e2c7d025 EA |
984 | |
985 | if (int128_ge(int128_make64(iova), llend)) { | |
e4b34708 KJ |
986 | if (memory_region_is_ram_device(section->mr)) { |
987 | trace_vfio_listener_region_add_no_dma_map( | |
988 | memory_region_name(section->mr), | |
989 | section->offset_within_address_space, | |
990 | int128_getlo(section->size), | |
8e3b0cbb | 991 | qemu_real_host_page_size()); |
e4b34708 | 992 | } |
e2c7d025 EA |
993 | return; |
994 | } | |
55efcc53 | 995 | end = int128_get64(int128_sub(llend, int128_one())); |
3898aad3 | 996 | |
2e4109de | 997 | if (container->iommu_type == VFIO_SPAPR_TCE_v2_IOMMU) { |
2e4109de AK |
998 | hwaddr pgsize = 0; |
999 | ||
1000 | /* For now intersections are not allowed, we may relax this later */ | |
1001 | QLIST_FOREACH(hostwin, &container->hostwin_list, hostwin_next) { | |
1002 | if (ranges_overlap(hostwin->min_iova, | |
1003 | hostwin->max_iova - hostwin->min_iova + 1, | |
1004 | section->offset_within_address_space, | |
1005 | int128_get64(section->size))) { | |
d7d87836 EA |
1006 | error_setg(&err, |
1007 | "region [0x%"PRIx64",0x%"PRIx64"] overlaps with existing" | |
1008 | "host DMA window [0x%"PRIx64",0x%"PRIx64"]", | |
1009 | section->offset_within_address_space, | |
1010 | section->offset_within_address_space + | |
1011 | int128_get64(section->size) - 1, | |
1012 | hostwin->min_iova, hostwin->max_iova); | |
2e4109de AK |
1013 | goto fail; |
1014 | } | |
1015 | } | |
1016 | ||
1017 | ret = vfio_spapr_create_window(container, section, &pgsize); | |
1018 | if (ret) { | |
d7d87836 | 1019 | error_setg_errno(&err, -ret, "Failed to create SPAPR window"); |
2e4109de AK |
1020 | goto fail; |
1021 | } | |
1022 | ||
1023 | vfio_host_win_add(container, section->offset_within_address_space, | |
1024 | section->offset_within_address_space + | |
1025 | int128_get64(section->size) - 1, pgsize); | |
07bc681a AK |
1026 | #ifdef CONFIG_KVM |
1027 | if (kvm_enabled()) { | |
1028 | VFIOGroup *group; | |
1029 | IOMMUMemoryRegion *iommu_mr = IOMMU_MEMORY_REGION(section->mr); | |
1030 | struct kvm_vfio_spapr_tce param; | |
1031 | struct kvm_device_attr attr = { | |
1032 | .group = KVM_DEV_VFIO_GROUP, | |
1033 | .attr = KVM_DEV_VFIO_GROUP_SET_SPAPR_TCE, | |
1034 | .addr = (uint64_t)(unsigned long)¶m, | |
1035 | }; | |
1036 | ||
1037 | if (!memory_region_iommu_get_attr(iommu_mr, IOMMU_ATTR_SPAPR_TCE_FD, | |
1038 | ¶m.tablefd)) { | |
1039 | QLIST_FOREACH(group, &container->group_list, container_next) { | |
1040 | param.groupfd = group->fd; | |
1041 | if (ioctl(vfio_kvm_device_fd, KVM_SET_DEVICE_ATTR, &attr)) { | |
1042 | error_report("vfio: failed to setup fd %d " | |
1043 | "for a group with fd %d: %s", | |
1044 | param.tablefd, param.groupfd, | |
1045 | strerror(errno)); | |
1046 | return; | |
1047 | } | |
1048 | trace_vfio_spapr_group_attach(param.groupfd, param.tablefd); | |
1049 | } | |
1050 | } | |
1051 | } | |
1052 | #endif | |
2e4109de AK |
1053 | } |
1054 | ||
fbc6c921 JM |
1055 | hostwin = vfio_find_hostwin(container, iova, end); |
1056 | if (!hostwin) { | |
d7d87836 EA |
1057 | error_setg(&err, "Container %p can't map guest IOVA region" |
1058 | " 0x%"HWADDR_PRIx"..0x%"HWADDR_PRIx, container, iova, end); | |
3898aad3 DG |
1059 | goto fail; |
1060 | } | |
e2c7d025 EA |
1061 | |
1062 | memory_region_ref(section->mr); | |
1063 | ||
1064 | if (memory_region_is_iommu(section->mr)) { | |
1065 | VFIOGuestIOMMU *giommu; | |
3df9d748 | 1066 | IOMMUMemoryRegion *iommu_mr = IOMMU_MEMORY_REGION(section->mr); |
cb1efcf4 | 1067 | int iommu_idx; |
e2c7d025 | 1068 | |
55efcc53 | 1069 | trace_vfio_listener_region_add_iommu(iova, end); |
e2c7d025 | 1070 | /* |
e2c7d025 EA |
1071 | * FIXME: For VFIO iommu types which have KVM acceleration to |
1072 | * avoid bouncing all map/unmaps through qemu this way, this | |
1073 | * would be the right place to wire that up (tell the KVM | |
1074 | * device emulation the VFIO iommu handles to use). | |
1075 | */ | |
e2c7d025 | 1076 | giommu = g_malloc0(sizeof(*giommu)); |
44ee6aaa | 1077 | giommu->iommu_mr = iommu_mr; |
d78c19b5 AK |
1078 | giommu->iommu_offset = section->offset_within_address_space - |
1079 | section->offset_within_region; | |
e2c7d025 | 1080 | giommu->container = container; |
698feb5e PX |
1081 | llend = int128_add(int128_make64(section->offset_within_region), |
1082 | section->size); | |
1083 | llend = int128_sub(llend, int128_one()); | |
cb1efcf4 PM |
1084 | iommu_idx = memory_region_iommu_attrs_to_index(iommu_mr, |
1085 | MEMTXATTRS_UNSPECIFIED); | |
698feb5e | 1086 | iommu_notifier_init(&giommu->n, vfio_iommu_map_notify, |
8dca037b | 1087 | IOMMU_NOTIFIER_IOTLB_EVENTS, |
698feb5e | 1088 | section->offset_within_region, |
cb1efcf4 PM |
1089 | int128_get64(llend), |
1090 | iommu_idx); | |
508ce5eb | 1091 | |
44ee6aaa | 1092 | ret = memory_region_iommu_set_page_size_mask(giommu->iommu_mr, |
b9177498 BB |
1093 | container->pgsizes, |
1094 | &err); | |
1095 | if (ret) { | |
1096 | g_free(giommu); | |
1097 | goto fail; | |
1098 | } | |
1099 | ||
549d4005 EA |
1100 | ret = memory_region_register_iommu_notifier(section->mr, &giommu->n, |
1101 | &err); | |
1102 | if (ret) { | |
1103 | g_free(giommu); | |
1104 | goto fail; | |
1105 | } | |
1106 | QLIST_INSERT_HEAD(&container->giommu_list, giommu, giommu_next); | |
44ee6aaa | 1107 | memory_region_iommu_replay(giommu->iommu_mr, &giommu->n); |
e2c7d025 EA |
1108 | |
1109 | return; | |
1110 | } | |
1111 | ||
1112 | /* Here we assume that memory_region_is_ram(section->mr)==true */ | |
1113 | ||
5e3b981c DH |
1114 | /* |
1115 | * For RAM memory regions with a RamDiscardManager, we only want to map the | |
1116 | * actually populated parts - and update the mapping whenever we're notified | |
1117 | * about changes. | |
1118 | */ | |
1119 | if (memory_region_has_ram_discard_manager(section->mr)) { | |
1120 | vfio_register_ram_discard_listener(container, section); | |
1121 | return; | |
1122 | } | |
1123 | ||
e2c7d025 EA |
1124 | vaddr = memory_region_get_ram_ptr(section->mr) + |
1125 | section->offset_within_region + | |
1126 | (iova - section->offset_within_address_space); | |
1127 | ||
55efcc53 | 1128 | trace_vfio_listener_region_add_ram(iova, end, vaddr); |
e2c7d025 | 1129 | |
55efcc53 BD |
1130 | llsize = int128_sub(llend, int128_make64(iova)); |
1131 | ||
567b5b30 AK |
1132 | if (memory_region_is_ram_device(section->mr)) { |
1133 | hwaddr pgmask = (1ULL << ctz64(hostwin->iova_pgsizes)) - 1; | |
1134 | ||
1135 | if ((iova & pgmask) || (int128_get64(llsize) & pgmask)) { | |
5c086005 EA |
1136 | trace_vfio_listener_region_add_no_dma_map( |
1137 | memory_region_name(section->mr), | |
1138 | section->offset_within_address_space, | |
1139 | int128_getlo(section->size), | |
1140 | pgmask + 1); | |
567b5b30 AK |
1141 | return; |
1142 | } | |
1143 | } | |
1144 | ||
55efcc53 BD |
1145 | ret = vfio_dma_map(container, iova, int128_get64(llsize), |
1146 | vaddr, section->readonly); | |
e2c7d025 | 1147 | if (ret) { |
d7d87836 | 1148 | error_setg(&err, "vfio_dma_map(%p, 0x%"HWADDR_PRIx", " |
db9b829b AH |
1149 | "0x%"HWADDR_PRIx", %p) = %d (%s)", |
1150 | container, iova, int128_get64(llsize), vaddr, ret, | |
1151 | strerror(-ret)); | |
567b5b30 AK |
1152 | if (memory_region_is_ram_device(section->mr)) { |
1153 | /* Allow unexpected mappings not to be fatal for RAM devices */ | |
d7d87836 | 1154 | error_report_err(err); |
567b5b30 AK |
1155 | return; |
1156 | } | |
ac6dc389 DG |
1157 | goto fail; |
1158 | } | |
e2c7d025 | 1159 | |
ac6dc389 DG |
1160 | return; |
1161 | ||
1162 | fail: | |
567b5b30 AK |
1163 | if (memory_region_is_ram_device(section->mr)) { |
1164 | error_report("failed to vfio_dma_map. pci p2p may not work"); | |
1165 | return; | |
1166 | } | |
ac6dc389 DG |
1167 | /* |
1168 | * On the initfn path, store the first error in the container so we | |
1169 | * can gracefully fail. Runtime, there's not much we can do other | |
1170 | * than throw a hardware error. | |
1171 | */ | |
1172 | if (!container->initialized) { | |
1173 | if (!container->error) { | |
d7d87836 EA |
1174 | error_propagate_prepend(&container->error, err, |
1175 | "Region %s: ", | |
1176 | memory_region_name(section->mr)); | |
1177 | } else { | |
1178 | error_free(err); | |
e2c7d025 | 1179 | } |
ac6dc389 | 1180 | } else { |
d7d87836 | 1181 | error_report_err(err); |
ac6dc389 | 1182 | hw_error("vfio: DMA mapping failed, unable to continue"); |
e2c7d025 EA |
1183 | } |
1184 | } | |
1185 | ||
1186 | static void vfio_listener_region_del(MemoryListener *listener, | |
1187 | MemoryRegionSection *section) | |
1188 | { | |
ee0bf0e5 | 1189 | VFIOContainer *container = container_of(listener, VFIOContainer, listener); |
e2c7d025 | 1190 | hwaddr iova, end; |
7a057b4f | 1191 | Int128 llend, llsize; |
e2c7d025 | 1192 | int ret; |
567b5b30 | 1193 | bool try_unmap = true; |
e2c7d025 | 1194 | |
b92f2376 | 1195 | if (!vfio_listener_valid_section(section, "region_del")) { |
e2c7d025 EA |
1196 | return; |
1197 | } | |
1198 | ||
1199 | if (memory_region_is_iommu(section->mr)) { | |
1200 | VFIOGuestIOMMU *giommu; | |
1201 | ||
1202 | QLIST_FOREACH(giommu, &container->giommu_list, giommu_next) { | |
44ee6aaa | 1203 | if (MEMORY_REGION(giommu->iommu_mr) == section->mr && |
698feb5e | 1204 | giommu->n.start == section->offset_within_region) { |
3df9d748 | 1205 | memory_region_unregister_iommu_notifier(section->mr, |
d22d8956 | 1206 | &giommu->n); |
e2c7d025 EA |
1207 | QLIST_REMOVE(giommu, giommu_next); |
1208 | g_free(giommu); | |
1209 | break; | |
1210 | } | |
1211 | } | |
1212 | ||
1213 | /* | |
1214 | * FIXME: We assume the one big unmap below is adequate to | |
1215 | * remove any individual page mappings in the IOMMU which | |
1216 | * might have been copied into VFIO. This works for a page table | |
1217 | * based IOMMU where a big unmap flattens a large range of IO-PTEs. | |
1218 | * That may not be true for all IOMMU types. | |
1219 | */ | |
1220 | } | |
1221 | ||
1eb7f642 | 1222 | iova = REAL_HOST_PAGE_ALIGN(section->offset_within_address_space); |
7a057b4f AK |
1223 | llend = int128_make64(section->offset_within_address_space); |
1224 | llend = int128_add(llend, section->size); | |
8e3b0cbb | 1225 | llend = int128_and(llend, int128_exts64(qemu_real_host_page_mask())); |
e2c7d025 | 1226 | |
7a057b4f | 1227 | if (int128_ge(int128_make64(iova), llend)) { |
e2c7d025 EA |
1228 | return; |
1229 | } | |
7a057b4f AK |
1230 | end = int128_get64(int128_sub(llend, int128_one())); |
1231 | ||
1232 | llsize = int128_sub(llend, int128_make64(iova)); | |
e2c7d025 | 1233 | |
7a057b4f | 1234 | trace_vfio_listener_region_del(iova, end); |
e2c7d025 | 1235 | |
567b5b30 AK |
1236 | if (memory_region_is_ram_device(section->mr)) { |
1237 | hwaddr pgmask; | |
1238 | VFIOHostDMAWindow *hostwin; | |
567b5b30 | 1239 | |
fbc6c921 JM |
1240 | hostwin = vfio_find_hostwin(container, iova, end); |
1241 | assert(hostwin); /* or region_add() would have failed */ | |
567b5b30 AK |
1242 | |
1243 | pgmask = (1ULL << ctz64(hostwin->iova_pgsizes)) - 1; | |
1244 | try_unmap = !((iova & pgmask) || (int128_get64(llsize) & pgmask)); | |
5e3b981c DH |
1245 | } else if (memory_region_has_ram_discard_manager(section->mr)) { |
1246 | vfio_unregister_ram_discard_listener(container, section); | |
1247 | /* Unregistering will trigger an unmap. */ | |
1248 | try_unmap = false; | |
e2c7d025 | 1249 | } |
2e4109de | 1250 | |
567b5b30 | 1251 | if (try_unmap) { |
1b296c3d JPB |
1252 | if (int128_eq(llsize, int128_2_64())) { |
1253 | /* The unmap ioctl doesn't accept a full 64-bit span. */ | |
1254 | llsize = int128_rshift(llsize, 1); | |
1255 | ret = vfio_dma_unmap(container, iova, int128_get64(llsize), NULL); | |
1256 | if (ret) { | |
1257 | error_report("vfio_dma_unmap(%p, 0x%"HWADDR_PRIx", " | |
db9b829b AH |
1258 | "0x%"HWADDR_PRIx") = %d (%s)", |
1259 | container, iova, int128_get64(llsize), ret, | |
1260 | strerror(-ret)); | |
1b296c3d JPB |
1261 | } |
1262 | iova += int128_get64(llsize); | |
1263 | } | |
9e7b0442 | 1264 | ret = vfio_dma_unmap(container, iova, int128_get64(llsize), NULL); |
567b5b30 AK |
1265 | if (ret) { |
1266 | error_report("vfio_dma_unmap(%p, 0x%"HWADDR_PRIx", " | |
db9b829b AH |
1267 | "0x%"HWADDR_PRIx") = %d (%s)", |
1268 | container, iova, int128_get64(llsize), ret, | |
1269 | strerror(-ret)); | |
567b5b30 AK |
1270 | } |
1271 | } | |
1272 | ||
1273 | memory_region_unref(section->mr); | |
1274 | ||
2e4109de AK |
1275 | if (container->iommu_type == VFIO_SPAPR_TCE_v2_IOMMU) { |
1276 | vfio_spapr_remove_window(container, | |
1277 | section->offset_within_address_space); | |
1278 | if (vfio_host_win_del(container, | |
1279 | section->offset_within_address_space, | |
1280 | section->offset_within_address_space + | |
1281 | int128_get64(section->size) - 1) < 0) { | |
1282 | hw_error("%s: Cannot delete missing window at %"HWADDR_PRIx, | |
1283 | __func__, section->offset_within_address_space); | |
1284 | } | |
1285 | } | |
e2c7d025 EA |
1286 | } |
1287 | ||
236e0a45 | 1288 | static int vfio_set_dirty_page_tracking(VFIOContainer *container, bool start) |
758b96b6 KZ |
1289 | { |
1290 | int ret; | |
1291 | struct vfio_iommu_type1_dirty_bitmap dirty = { | |
1292 | .argsz = sizeof(dirty), | |
1293 | }; | |
1294 | ||
b051a3f6 | 1295 | if (!container->dirty_pages_supported) { |
236e0a45 | 1296 | return 0; |
b051a3f6 AH |
1297 | } |
1298 | ||
758b96b6 KZ |
1299 | if (start) { |
1300 | dirty.flags = VFIO_IOMMU_DIRTY_PAGES_FLAG_START; | |
1301 | } else { | |
1302 | dirty.flags = VFIO_IOMMU_DIRTY_PAGES_FLAG_STOP; | |
1303 | } | |
1304 | ||
1305 | ret = ioctl(container->fd, VFIO_IOMMU_DIRTY_PAGES, &dirty); | |
1306 | if (ret) { | |
236e0a45 | 1307 | ret = -errno; |
758b96b6 KZ |
1308 | error_report("Failed to set dirty tracking flag 0x%x errno: %d", |
1309 | dirty.flags, errno); | |
1310 | } | |
236e0a45 AH |
1311 | |
1312 | return ret; | |
758b96b6 KZ |
1313 | } |
1314 | ||
1315 | static void vfio_listener_log_global_start(MemoryListener *listener) | |
1316 | { | |
1317 | VFIOContainer *container = container_of(listener, VFIOContainer, listener); | |
236e0a45 | 1318 | int ret; |
758b96b6 | 1319 | |
236e0a45 AH |
1320 | ret = vfio_set_dirty_page_tracking(container, true); |
1321 | if (ret) { | |
1322 | vfio_set_migration_error(ret); | |
1323 | } | |
758b96b6 KZ |
1324 | } |
1325 | ||
1326 | static void vfio_listener_log_global_stop(MemoryListener *listener) | |
1327 | { | |
1328 | VFIOContainer *container = container_of(listener, VFIOContainer, listener); | |
236e0a45 | 1329 | int ret; |
758b96b6 | 1330 | |
236e0a45 AH |
1331 | ret = vfio_set_dirty_page_tracking(container, false); |
1332 | if (ret) { | |
1333 | vfio_set_migration_error(ret); | |
1334 | } | |
758b96b6 KZ |
1335 | } |
1336 | ||
b6dd6504 KW |
1337 | static int vfio_get_dirty_bitmap(VFIOContainer *container, uint64_t iova, |
1338 | uint64_t size, ram_addr_t ram_addr) | |
1339 | { | |
1340 | struct vfio_iommu_type1_dirty_bitmap *dbitmap; | |
1341 | struct vfio_iommu_type1_dirty_bitmap_get *range; | |
725ccd7e | 1342 | VFIOBitmap vbmap; |
b6dd6504 KW |
1343 | int ret; |
1344 | ||
b051a3f6 AH |
1345 | if (!container->dirty_pages_supported) { |
1346 | cpu_physical_memory_set_dirty_range(ram_addr, size, | |
1347 | tcg_enabled() ? DIRTY_CLIENTS_ALL : | |
1348 | DIRTY_CLIENTS_NOCODE); | |
1349 | return 0; | |
1350 | } | |
1351 | ||
725ccd7e AH |
1352 | ret = vfio_bitmap_alloc(&vbmap, size); |
1353 | if (ret) { | |
1354 | return ret; | |
1355 | } | |
1356 | ||
b6dd6504 KW |
1357 | dbitmap = g_malloc0(sizeof(*dbitmap) + sizeof(*range)); |
1358 | ||
1359 | dbitmap->argsz = sizeof(*dbitmap) + sizeof(*range); | |
1360 | dbitmap->flags = VFIO_IOMMU_DIRTY_PAGES_FLAG_GET_BITMAP; | |
1361 | range = (struct vfio_iommu_type1_dirty_bitmap_get *)&dbitmap->data; | |
1362 | range->iova = iova; | |
1363 | range->size = size; | |
1364 | ||
1365 | /* | |
1eb7f642 KJ |
1366 | * cpu_physical_memory_set_dirty_lebitmap() supports pages in bitmap of |
1367 | * qemu_real_host_page_size to mark those dirty. Hence set bitmap's pgsize | |
1368 | * to qemu_real_host_page_size. | |
b6dd6504 | 1369 | */ |
8e3b0cbb | 1370 | range->bitmap.pgsize = qemu_real_host_page_size(); |
725ccd7e AH |
1371 | range->bitmap.size = vbmap.size; |
1372 | range->bitmap.data = (__u64 *)vbmap.bitmap; | |
b6dd6504 KW |
1373 | |
1374 | ret = ioctl(container->fd, VFIO_IOMMU_DIRTY_PAGES, dbitmap); | |
1375 | if (ret) { | |
3e2413a6 | 1376 | ret = -errno; |
b6dd6504 KW |
1377 | error_report("Failed to get dirty bitmap for iova: 0x%"PRIx64 |
1378 | " size: 0x%"PRIx64" err: %d", (uint64_t)range->iova, | |
1379 | (uint64_t)range->size, errno); | |
1380 | goto err_out; | |
1381 | } | |
1382 | ||
725ccd7e AH |
1383 | cpu_physical_memory_set_dirty_lebitmap(vbmap.bitmap, ram_addr, |
1384 | vbmap.pages); | |
b6dd6504 KW |
1385 | |
1386 | trace_vfio_get_dirty_bitmap(container->fd, range->iova, range->size, | |
1387 | range->bitmap.size, ram_addr); | |
1388 | err_out: | |
b6dd6504 | 1389 | g_free(dbitmap); |
725ccd7e | 1390 | g_free(vbmap.bitmap); |
b6dd6504 KW |
1391 | |
1392 | return ret; | |
1393 | } | |
1394 | ||
9a04fe09 KW |
1395 | typedef struct { |
1396 | IOMMUNotifier n; | |
1397 | VFIOGuestIOMMU *giommu; | |
1398 | } vfio_giommu_dirty_notifier; | |
1399 | ||
1400 | static void vfio_iommu_map_dirty_notify(IOMMUNotifier *n, IOMMUTLBEntry *iotlb) | |
1401 | { | |
1402 | vfio_giommu_dirty_notifier *gdn = container_of(n, | |
1403 | vfio_giommu_dirty_notifier, n); | |
1404 | VFIOGuestIOMMU *giommu = gdn->giommu; | |
1405 | VFIOContainer *container = giommu->container; | |
1406 | hwaddr iova = iotlb->iova + giommu->iommu_offset; | |
1407 | ram_addr_t translated_addr; | |
236e0a45 | 1408 | int ret = -EINVAL; |
9a04fe09 KW |
1409 | |
1410 | trace_vfio_iommu_map_dirty_notify(iova, iova + iotlb->addr_mask); | |
1411 | ||
1412 | if (iotlb->target_as != &address_space_memory) { | |
1413 | error_report("Wrong target AS \"%s\", only system memory is allowed", | |
1414 | iotlb->target_as->name ? iotlb->target_as->name : "none"); | |
236e0a45 | 1415 | goto out; |
9a04fe09 KW |
1416 | } |
1417 | ||
1418 | rcu_read_lock(); | |
1419 | if (vfio_get_xlat_addr(iotlb, NULL, &translated_addr, NULL)) { | |
9a04fe09 KW |
1420 | ret = vfio_get_dirty_bitmap(container, iova, iotlb->addr_mask + 1, |
1421 | translated_addr); | |
1422 | if (ret) { | |
1423 | error_report("vfio_iommu_map_dirty_notify(%p, 0x%"HWADDR_PRIx", " | |
db9b829b AH |
1424 | "0x%"HWADDR_PRIx") = %d (%s)", |
1425 | container, iova, iotlb->addr_mask + 1, ret, | |
1426 | strerror(-ret)); | |
9a04fe09 KW |
1427 | } |
1428 | } | |
1429 | rcu_read_unlock(); | |
236e0a45 AH |
1430 | |
1431 | out: | |
1432 | if (ret) { | |
1433 | vfio_set_migration_error(ret); | |
1434 | } | |
9a04fe09 KW |
1435 | } |
1436 | ||
5e3b981c DH |
1437 | static int vfio_ram_discard_get_dirty_bitmap(MemoryRegionSection *section, |
1438 | void *opaque) | |
1439 | { | |
1440 | const hwaddr size = int128_get64(section->size); | |
1441 | const hwaddr iova = section->offset_within_address_space; | |
1442 | const ram_addr_t ram_addr = memory_region_get_ram_addr(section->mr) + | |
1443 | section->offset_within_region; | |
1444 | VFIORamDiscardListener *vrdl = opaque; | |
1445 | ||
1446 | /* | |
1447 | * Sync the whole mapped region (spanning multiple individual mappings) | |
1448 | * in one go. | |
1449 | */ | |
1450 | return vfio_get_dirty_bitmap(vrdl->container, iova, size, ram_addr); | |
1451 | } | |
1452 | ||
1453 | static int vfio_sync_ram_discard_listener_dirty_bitmap(VFIOContainer *container, | |
1454 | MemoryRegionSection *section) | |
1455 | { | |
1456 | RamDiscardManager *rdm = memory_region_get_ram_discard_manager(section->mr); | |
1457 | VFIORamDiscardListener *vrdl = NULL; | |
1458 | ||
1459 | QLIST_FOREACH(vrdl, &container->vrdl_list, next) { | |
1460 | if (vrdl->mr == section->mr && | |
1461 | vrdl->offset_within_address_space == | |
1462 | section->offset_within_address_space) { | |
1463 | break; | |
1464 | } | |
1465 | } | |
1466 | ||
1467 | if (!vrdl) { | |
1468 | hw_error("vfio: Trying to sync missing RAM discard listener"); | |
1469 | } | |
1470 | ||
1471 | /* | |
1472 | * We only want/can synchronize the bitmap for actually mapped parts - | |
1473 | * which correspond to populated parts. Replay all populated parts. | |
1474 | */ | |
1475 | return ram_discard_manager_replay_populated(rdm, section, | |
1476 | vfio_ram_discard_get_dirty_bitmap, | |
1477 | &vrdl); | |
1478 | } | |
1479 | ||
b6dd6504 KW |
1480 | static int vfio_sync_dirty_bitmap(VFIOContainer *container, |
1481 | MemoryRegionSection *section) | |
1482 | { | |
1483 | ram_addr_t ram_addr; | |
1484 | ||
9a04fe09 KW |
1485 | if (memory_region_is_iommu(section->mr)) { |
1486 | VFIOGuestIOMMU *giommu; | |
1487 | ||
1488 | QLIST_FOREACH(giommu, &container->giommu_list, giommu_next) { | |
44ee6aaa | 1489 | if (MEMORY_REGION(giommu->iommu_mr) == section->mr && |
9a04fe09 KW |
1490 | giommu->n.start == section->offset_within_region) { |
1491 | Int128 llend; | |
1492 | vfio_giommu_dirty_notifier gdn = { .giommu = giommu }; | |
44ee6aaa | 1493 | int idx = memory_region_iommu_attrs_to_index(giommu->iommu_mr, |
9a04fe09 KW |
1494 | MEMTXATTRS_UNSPECIFIED); |
1495 | ||
1496 | llend = int128_add(int128_make64(section->offset_within_region), | |
1497 | section->size); | |
1498 | llend = int128_sub(llend, int128_one()); | |
1499 | ||
1500 | iommu_notifier_init(&gdn.n, | |
1501 | vfio_iommu_map_dirty_notify, | |
1502 | IOMMU_NOTIFIER_MAP, | |
1503 | section->offset_within_region, | |
1504 | int128_get64(llend), | |
1505 | idx); | |
44ee6aaa | 1506 | memory_region_iommu_replay(giommu->iommu_mr, &gdn.n); |
9a04fe09 KW |
1507 | break; |
1508 | } | |
1509 | } | |
1510 | return 0; | |
5e3b981c DH |
1511 | } else if (memory_region_has_ram_discard_manager(section->mr)) { |
1512 | return vfio_sync_ram_discard_listener_dirty_bitmap(container, section); | |
9a04fe09 KW |
1513 | } |
1514 | ||
b6dd6504 KW |
1515 | ram_addr = memory_region_get_ram_addr(section->mr) + |
1516 | section->offset_within_region; | |
1517 | ||
1518 | return vfio_get_dirty_bitmap(container, | |
1eb7f642 KJ |
1519 | REAL_HOST_PAGE_ALIGN(section->offset_within_address_space), |
1520 | int128_get64(section->size), ram_addr); | |
b6dd6504 KW |
1521 | } |
1522 | ||
4292d501 | 1523 | static void vfio_listener_log_sync(MemoryListener *listener, |
b6dd6504 KW |
1524 | MemoryRegionSection *section) |
1525 | { | |
1526 | VFIOContainer *container = container_of(listener, VFIOContainer, listener); | |
236e0a45 | 1527 | int ret; |
b6dd6504 | 1528 | |
b051a3f6 | 1529 | if (vfio_listener_skipped_section(section)) { |
b6dd6504 KW |
1530 | return; |
1531 | } | |
1532 | ||
758b96b6 | 1533 | if (vfio_devices_all_dirty_tracking(container)) { |
236e0a45 AH |
1534 | ret = vfio_sync_dirty_bitmap(container, section); |
1535 | if (ret) { | |
1536 | error_report("vfio: Failed to sync dirty bitmap, err: %d (%s)", ret, | |
1537 | strerror(-ret)); | |
1538 | vfio_set_migration_error(ret); | |
1539 | } | |
b6dd6504 KW |
1540 | } |
1541 | } | |
1542 | ||
51b833f4 | 1543 | static const MemoryListener vfio_memory_listener = { |
142518bd | 1544 | .name = "vfio", |
e2c7d025 EA |
1545 | .region_add = vfio_listener_region_add, |
1546 | .region_del = vfio_listener_region_del, | |
758b96b6 KZ |
1547 | .log_global_start = vfio_listener_log_global_start, |
1548 | .log_global_stop = vfio_listener_log_global_stop, | |
4292d501 | 1549 | .log_sync = vfio_listener_log_sync, |
e2c7d025 EA |
1550 | }; |
1551 | ||
51b833f4 | 1552 | static void vfio_listener_release(VFIOContainer *container) |
e2c7d025 | 1553 | { |
ee0bf0e5 | 1554 | memory_listener_unregister(&container->listener); |
318f67ce AK |
1555 | if (container->iommu_type == VFIO_SPAPR_TCE_v2_IOMMU) { |
1556 | memory_listener_unregister(&container->prereg_listener); | |
1557 | } | |
e2c7d025 EA |
1558 | } |
1559 | ||
3ab7a0b4 MR |
1560 | static struct vfio_info_cap_header * |
1561 | vfio_get_cap(void *ptr, uint32_t cap_offset, uint16_t id) | |
b53b0f69 AW |
1562 | { |
1563 | struct vfio_info_cap_header *hdr; | |
b53b0f69 | 1564 | |
3ab7a0b4 | 1565 | for (hdr = ptr + cap_offset; hdr != ptr; hdr = ptr + hdr->next) { |
b53b0f69 AW |
1566 | if (hdr->id == id) { |
1567 | return hdr; | |
1568 | } | |
1569 | } | |
1570 | ||
1571 | return NULL; | |
1572 | } | |
1573 | ||
3ab7a0b4 MR |
1574 | struct vfio_info_cap_header * |
1575 | vfio_get_region_info_cap(struct vfio_region_info *info, uint16_t id) | |
1576 | { | |
1577 | if (!(info->flags & VFIO_REGION_INFO_FLAG_CAPS)) { | |
1578 | return NULL; | |
1579 | } | |
1580 | ||
1581 | return vfio_get_cap((void *)info, info->cap_offset, id); | |
1582 | } | |
1583 | ||
7486a628 MR |
1584 | static struct vfio_info_cap_header * |
1585 | vfio_get_iommu_type1_info_cap(struct vfio_iommu_type1_info *info, uint16_t id) | |
1586 | { | |
1587 | if (!(info->flags & VFIO_IOMMU_INFO_CAPS)) { | |
1588 | return NULL; | |
1589 | } | |
1590 | ||
1591 | return vfio_get_cap((void *)info, info->cap_offset, id); | |
1592 | } | |
1593 | ||
92fe289a MR |
1594 | struct vfio_info_cap_header * |
1595 | vfio_get_device_info_cap(struct vfio_device_info *info, uint16_t id) | |
1596 | { | |
1597 | if (!(info->flags & VFIO_DEVICE_FLAGS_CAPS)) { | |
1598 | return NULL; | |
1599 | } | |
1600 | ||
1601 | return vfio_get_cap((void *)info, info->cap_offset, id); | |
1602 | } | |
1603 | ||
7486a628 MR |
1604 | bool vfio_get_info_dma_avail(struct vfio_iommu_type1_info *info, |
1605 | unsigned int *avail) | |
1606 | { | |
1607 | struct vfio_info_cap_header *hdr; | |
1608 | struct vfio_iommu_type1_info_dma_avail *cap; | |
1609 | ||
1610 | /* If the capability cannot be found, assume no DMA limiting */ | |
1611 | hdr = vfio_get_iommu_type1_info_cap(info, | |
1612 | VFIO_IOMMU_TYPE1_INFO_DMA_AVAIL); | |
1613 | if (hdr == NULL) { | |
1614 | return false; | |
1615 | } | |
1616 | ||
1617 | if (avail != NULL) { | |
1618 | cap = (void *) hdr; | |
1619 | *avail = cap->avail; | |
1620 | } | |
1621 | ||
1622 | return true; | |
1623 | } | |
1624 | ||
24acf72b AW |
1625 | static int vfio_setup_region_sparse_mmaps(VFIORegion *region, |
1626 | struct vfio_region_info *info) | |
b53b0f69 AW |
1627 | { |
1628 | struct vfio_info_cap_header *hdr; | |
1629 | struct vfio_region_info_cap_sparse_mmap *sparse; | |
24acf72b | 1630 | int i, j; |
b53b0f69 AW |
1631 | |
1632 | hdr = vfio_get_region_info_cap(info, VFIO_REGION_INFO_CAP_SPARSE_MMAP); | |
1633 | if (!hdr) { | |
24acf72b | 1634 | return -ENODEV; |
b53b0f69 AW |
1635 | } |
1636 | ||
1637 | sparse = container_of(hdr, struct vfio_region_info_cap_sparse_mmap, header); | |
1638 | ||
1639 | trace_vfio_region_sparse_mmap_header(region->vbasedev->name, | |
1640 | region->nr, sparse->nr_areas); | |
1641 | ||
24acf72b AW |
1642 | region->mmaps = g_new0(VFIOMmap, sparse->nr_areas); |
1643 | ||
1644 | for (i = 0, j = 0; i < sparse->nr_areas; i++) { | |
24acf72b | 1645 | if (sparse->areas[i].size) { |
99510d27 XC |
1646 | trace_vfio_region_sparse_mmap_entry(i, sparse->areas[i].offset, |
1647 | sparse->areas[i].offset + | |
1648 | sparse->areas[i].size - 1); | |
24acf72b AW |
1649 | region->mmaps[j].offset = sparse->areas[i].offset; |
1650 | region->mmaps[j].size = sparse->areas[i].size; | |
1651 | j++; | |
1652 | } | |
b53b0f69 | 1653 | } |
24acf72b AW |
1654 | |
1655 | region->nr_mmaps = j; | |
1656 | region->mmaps = g_realloc(region->mmaps, j * sizeof(VFIOMmap)); | |
1657 | ||
1658 | return 0; | |
b53b0f69 AW |
1659 | } |
1660 | ||
db0da029 AW |
1661 | int vfio_region_setup(Object *obj, VFIODevice *vbasedev, VFIORegion *region, |
1662 | int index, const char *name) | |
e2c7d025 | 1663 | { |
db0da029 AW |
1664 | struct vfio_region_info *info; |
1665 | int ret; | |
1666 | ||
1667 | ret = vfio_get_region_info(vbasedev, index, &info); | |
1668 | if (ret) { | |
1669 | return ret; | |
1670 | } | |
1671 | ||
1672 | region->vbasedev = vbasedev; | |
1673 | region->flags = info->flags; | |
1674 | region->size = info->size; | |
1675 | region->fd_offset = info->offset; | |
1676 | region->nr = index; | |
1677 | ||
1678 | if (region->size) { | |
1679 | region->mem = g_new0(MemoryRegion, 1); | |
1680 | memory_region_init_io(region->mem, obj, &vfio_region_ops, | |
1681 | region, name, region->size); | |
e2c7d025 | 1682 | |
db0da029 | 1683 | if (!vbasedev->no_mmap && |
95251725 | 1684 | region->flags & VFIO_REGION_INFO_FLAG_MMAP) { |
e2c7d025 | 1685 | |
24acf72b | 1686 | ret = vfio_setup_region_sparse_mmaps(region, info); |
db0da029 | 1687 | |
24acf72b | 1688 | if (ret) { |
b53b0f69 AW |
1689 | region->nr_mmaps = 1; |
1690 | region->mmaps = g_new0(VFIOMmap, region->nr_mmaps); | |
1691 | region->mmaps[0].offset = 0; | |
1692 | region->mmaps[0].size = region->size; | |
1693 | } | |
e2c7d025 | 1694 | } |
db0da029 AW |
1695 | } |
1696 | ||
1697 | g_free(info); | |
1698 | ||
1699 | trace_vfio_region_setup(vbasedev->name, index, name, | |
1700 | region->flags, region->fd_offset, region->size); | |
1701 | return 0; | |
1702 | } | |
e2c7d025 | 1703 | |
0f7a903b KW |
1704 | static void vfio_subregion_unmap(VFIORegion *region, int index) |
1705 | { | |
1706 | trace_vfio_region_unmap(memory_region_name(®ion->mmaps[index].mem), | |
1707 | region->mmaps[index].offset, | |
1708 | region->mmaps[index].offset + | |
1709 | region->mmaps[index].size - 1); | |
1710 | memory_region_del_subregion(region->mem, ®ion->mmaps[index].mem); | |
1711 | munmap(region->mmaps[index].mmap, region->mmaps[index].size); | |
1712 | object_unparent(OBJECT(®ion->mmaps[index].mem)); | |
1713 | region->mmaps[index].mmap = NULL; | |
1714 | } | |
1715 | ||
db0da029 AW |
1716 | int vfio_region_mmap(VFIORegion *region) |
1717 | { | |
1718 | int i, prot = 0; | |
1719 | char *name; | |
1720 | ||
1721 | if (!region->mem) { | |
1722 | return 0; | |
1723 | } | |
1724 | ||
1725 | prot |= region->flags & VFIO_REGION_INFO_FLAG_READ ? PROT_READ : 0; | |
1726 | prot |= region->flags & VFIO_REGION_INFO_FLAG_WRITE ? PROT_WRITE : 0; | |
1727 | ||
1728 | for (i = 0; i < region->nr_mmaps; i++) { | |
1729 | region->mmaps[i].mmap = mmap(NULL, region->mmaps[i].size, prot, | |
1730 | MAP_SHARED, region->vbasedev->fd, | |
1731 | region->fd_offset + | |
1732 | region->mmaps[i].offset); | |
1733 | if (region->mmaps[i].mmap == MAP_FAILED) { | |
1734 | int ret = -errno; | |
1735 | ||
1736 | trace_vfio_region_mmap_fault(memory_region_name(region->mem), i, | |
1737 | region->fd_offset + | |
1738 | region->mmaps[i].offset, | |
1739 | region->fd_offset + | |
1740 | region->mmaps[i].offset + | |
1741 | region->mmaps[i].size - 1, ret); | |
1742 | ||
1743 | region->mmaps[i].mmap = NULL; | |
1744 | ||
1745 | for (i--; i >= 0; i--) { | |
0f7a903b | 1746 | vfio_subregion_unmap(region, i); |
db0da029 AW |
1747 | } |
1748 | ||
1749 | return ret; | |
e2c7d025 EA |
1750 | } |
1751 | ||
db0da029 AW |
1752 | name = g_strdup_printf("%s mmaps[%d]", |
1753 | memory_region_name(region->mem), i); | |
21e00fa5 AW |
1754 | memory_region_init_ram_device_ptr(®ion->mmaps[i].mem, |
1755 | memory_region_owner(region->mem), | |
1756 | name, region->mmaps[i].size, | |
1757 | region->mmaps[i].mmap); | |
db0da029 | 1758 | g_free(name); |
db0da029 AW |
1759 | memory_region_add_subregion(region->mem, region->mmaps[i].offset, |
1760 | ®ion->mmaps[i].mem); | |
1761 | ||
1762 | trace_vfio_region_mmap(memory_region_name(®ion->mmaps[i].mem), | |
1763 | region->mmaps[i].offset, | |
1764 | region->mmaps[i].offset + | |
1765 | region->mmaps[i].size - 1); | |
1766 | } | |
1767 | ||
1768 | return 0; | |
1769 | } | |
1770 | ||
0f7a903b KW |
1771 | void vfio_region_unmap(VFIORegion *region) |
1772 | { | |
1773 | int i; | |
1774 | ||
1775 | if (!region->mem) { | |
1776 | return; | |
1777 | } | |
1778 | ||
1779 | for (i = 0; i < region->nr_mmaps; i++) { | |
1780 | if (region->mmaps[i].mmap) { | |
1781 | vfio_subregion_unmap(region, i); | |
1782 | } | |
1783 | } | |
1784 | } | |
1785 | ||
db0da029 AW |
1786 | void vfio_region_exit(VFIORegion *region) |
1787 | { | |
1788 | int i; | |
1789 | ||
1790 | if (!region->mem) { | |
1791 | return; | |
1792 | } | |
1793 | ||
1794 | for (i = 0; i < region->nr_mmaps; i++) { | |
1795 | if (region->mmaps[i].mmap) { | |
1796 | memory_region_del_subregion(region->mem, ®ion->mmaps[i].mem); | |
e2c7d025 | 1797 | } |
db0da029 | 1798 | } |
e2c7d025 | 1799 | |
db0da029 AW |
1800 | trace_vfio_region_exit(region->vbasedev->name, region->nr); |
1801 | } | |
1802 | ||
1803 | void vfio_region_finalize(VFIORegion *region) | |
1804 | { | |
1805 | int i; | |
1806 | ||
1807 | if (!region->mem) { | |
1808 | return; | |
e2c7d025 EA |
1809 | } |
1810 | ||
db0da029 AW |
1811 | for (i = 0; i < region->nr_mmaps; i++) { |
1812 | if (region->mmaps[i].mmap) { | |
1813 | munmap(region->mmaps[i].mmap, region->mmaps[i].size); | |
1814 | object_unparent(OBJECT(®ion->mmaps[i].mem)); | |
1815 | } | |
1816 | } | |
1817 | ||
1818 | object_unparent(OBJECT(region->mem)); | |
1819 | ||
1820 | g_free(region->mem); | |
1821 | g_free(region->mmaps); | |
1822 | ||
1823 | trace_vfio_region_finalize(region->vbasedev->name, region->nr); | |
92f86bff GH |
1824 | |
1825 | region->mem = NULL; | |
1826 | region->mmaps = NULL; | |
1827 | region->nr_mmaps = 0; | |
1828 | region->size = 0; | |
1829 | region->flags = 0; | |
1830 | region->nr = 0; | |
db0da029 AW |
1831 | } |
1832 | ||
1833 | void vfio_region_mmaps_set_enabled(VFIORegion *region, bool enabled) | |
1834 | { | |
1835 | int i; | |
1836 | ||
1837 | if (!region->mem) { | |
1838 | return; | |
1839 | } | |
1840 | ||
1841 | for (i = 0; i < region->nr_mmaps; i++) { | |
1842 | if (region->mmaps[i].mmap) { | |
1843 | memory_region_set_enabled(®ion->mmaps[i].mem, enabled); | |
1844 | } | |
1845 | } | |
e2c7d025 | 1846 | |
db0da029 AW |
1847 | trace_vfio_region_mmaps_set_enabled(memory_region_name(region->mem), |
1848 | enabled); | |
e2c7d025 EA |
1849 | } |
1850 | ||
1851 | void vfio_reset_handler(void *opaque) | |
1852 | { | |
1853 | VFIOGroup *group; | |
1854 | VFIODevice *vbasedev; | |
1855 | ||
1856 | QLIST_FOREACH(group, &vfio_group_list, next) { | |
1857 | QLIST_FOREACH(vbasedev, &group->device_list, next) { | |
7da624e2 AW |
1858 | if (vbasedev->dev->realized) { |
1859 | vbasedev->ops->vfio_compute_needs_reset(vbasedev); | |
1860 | } | |
e2c7d025 EA |
1861 | } |
1862 | } | |
1863 | ||
1864 | QLIST_FOREACH(group, &vfio_group_list, next) { | |
1865 | QLIST_FOREACH(vbasedev, &group->device_list, next) { | |
7da624e2 | 1866 | if (vbasedev->dev->realized && vbasedev->needs_reset) { |
e2c7d025 EA |
1867 | vbasedev->ops->vfio_hot_reset_multi(vbasedev); |
1868 | } | |
1869 | } | |
1870 | } | |
1871 | } | |
1872 | ||
1873 | static void vfio_kvm_device_add_group(VFIOGroup *group) | |
1874 | { | |
1875 | #ifdef CONFIG_KVM | |
1876 | struct kvm_device_attr attr = { | |
1877 | .group = KVM_DEV_VFIO_GROUP, | |
1878 | .attr = KVM_DEV_VFIO_GROUP_ADD, | |
1879 | .addr = (uint64_t)(unsigned long)&group->fd, | |
1880 | }; | |
1881 | ||
1882 | if (!kvm_enabled()) { | |
1883 | return; | |
1884 | } | |
1885 | ||
1886 | if (vfio_kvm_device_fd < 0) { | |
1887 | struct kvm_create_device cd = { | |
1888 | .type = KVM_DEV_TYPE_VFIO, | |
1889 | }; | |
1890 | ||
1891 | if (kvm_vm_ioctl(kvm_state, KVM_CREATE_DEVICE, &cd)) { | |
78e5b17f | 1892 | error_report("Failed to create KVM VFIO device: %m"); |
e2c7d025 EA |
1893 | return; |
1894 | } | |
1895 | ||
1896 | vfio_kvm_device_fd = cd.fd; | |
1897 | } | |
1898 | ||
1899 | if (ioctl(vfio_kvm_device_fd, KVM_SET_DEVICE_ATTR, &attr)) { | |
1900 | error_report("Failed to add group %d to KVM VFIO device: %m", | |
1901 | group->groupid); | |
1902 | } | |
1903 | #endif | |
1904 | } | |
1905 | ||
1906 | static void vfio_kvm_device_del_group(VFIOGroup *group) | |
1907 | { | |
1908 | #ifdef CONFIG_KVM | |
1909 | struct kvm_device_attr attr = { | |
1910 | .group = KVM_DEV_VFIO_GROUP, | |
1911 | .attr = KVM_DEV_VFIO_GROUP_DEL, | |
1912 | .addr = (uint64_t)(unsigned long)&group->fd, | |
1913 | }; | |
1914 | ||
1915 | if (vfio_kvm_device_fd < 0) { | |
1916 | return; | |
1917 | } | |
1918 | ||
1919 | if (ioctl(vfio_kvm_device_fd, KVM_SET_DEVICE_ATTR, &attr)) { | |
1920 | error_report("Failed to remove group %d from KVM VFIO device: %m", | |
1921 | group->groupid); | |
1922 | } | |
1923 | #endif | |
1924 | } | |
1925 | ||
1926 | static VFIOAddressSpace *vfio_get_address_space(AddressSpace *as) | |
1927 | { | |
1928 | VFIOAddressSpace *space; | |
1929 | ||
1930 | QLIST_FOREACH(space, &vfio_address_spaces, list) { | |
1931 | if (space->as == as) { | |
1932 | return space; | |
1933 | } | |
1934 | } | |
1935 | ||
1936 | /* No suitable VFIOAddressSpace, create a new one */ | |
1937 | space = g_malloc0(sizeof(*space)); | |
1938 | space->as = as; | |
1939 | QLIST_INIT(&space->containers); | |
1940 | ||
1941 | QLIST_INSERT_HEAD(&vfio_address_spaces, space, list); | |
1942 | ||
1943 | return space; | |
1944 | } | |
1945 | ||
1946 | static void vfio_put_address_space(VFIOAddressSpace *space) | |
1947 | { | |
1948 | if (QLIST_EMPTY(&space->containers)) { | |
1949 | QLIST_REMOVE(space, list); | |
1950 | g_free(space); | |
1951 | } | |
1952 | } | |
1953 | ||
2b6326c0 EA |
1954 | /* |
1955 | * vfio_get_iommu_type - selects the richest iommu_type (v2 first) | |
1956 | */ | |
1957 | static int vfio_get_iommu_type(VFIOContainer *container, | |
1958 | Error **errp) | |
1959 | { | |
1960 | int iommu_types[] = { VFIO_TYPE1v2_IOMMU, VFIO_TYPE1_IOMMU, | |
1961 | VFIO_SPAPR_TCE_v2_IOMMU, VFIO_SPAPR_TCE_IOMMU }; | |
1962 | int i; | |
1963 | ||
1964 | for (i = 0; i < ARRAY_SIZE(iommu_types); i++) { | |
1965 | if (ioctl(container->fd, VFIO_CHECK_EXTENSION, iommu_types[i])) { | |
1966 | return iommu_types[i]; | |
1967 | } | |
1968 | } | |
1969 | error_setg(errp, "No available IOMMU models"); | |
1970 | return -EINVAL; | |
1971 | } | |
1972 | ||
1973 | static int vfio_init_container(VFIOContainer *container, int group_fd, | |
1974 | Error **errp) | |
1975 | { | |
1976 | int iommu_type, ret; | |
1977 | ||
1978 | iommu_type = vfio_get_iommu_type(container, errp); | |
1979 | if (iommu_type < 0) { | |
1980 | return iommu_type; | |
1981 | } | |
1982 | ||
1983 | ret = ioctl(group_fd, VFIO_GROUP_SET_CONTAINER, &container->fd); | |
1984 | if (ret) { | |
1985 | error_setg_errno(errp, errno, "Failed to set group container"); | |
1986 | return -errno; | |
1987 | } | |
1988 | ||
1989 | while (ioctl(container->fd, VFIO_SET_IOMMU, iommu_type)) { | |
1990 | if (iommu_type == VFIO_SPAPR_TCE_v2_IOMMU) { | |
1991 | /* | |
1992 | * On sPAPR, despite the IOMMU subdriver always advertises v1 and | |
1993 | * v2, the running platform may not support v2 and there is no | |
1994 | * way to guess it until an IOMMU group gets added to the container. | |
1995 | * So in case it fails with v2, try v1 as a fallback. | |
1996 | */ | |
1997 | iommu_type = VFIO_SPAPR_TCE_IOMMU; | |
1998 | continue; | |
1999 | } | |
2000 | error_setg_errno(errp, errno, "Failed to set iommu for container"); | |
2001 | return -errno; | |
2002 | } | |
2003 | ||
2004 | container->iommu_type = iommu_type; | |
2005 | return 0; | |
2006 | } | |
2007 | ||
87ea529c KW |
2008 | static int vfio_get_iommu_info(VFIOContainer *container, |
2009 | struct vfio_iommu_type1_info **info) | |
2010 | { | |
2011 | ||
2012 | size_t argsz = sizeof(struct vfio_iommu_type1_info); | |
2013 | ||
2014 | *info = g_new0(struct vfio_iommu_type1_info, 1); | |
2015 | again: | |
2016 | (*info)->argsz = argsz; | |
2017 | ||
2018 | if (ioctl(container->fd, VFIO_IOMMU_GET_INFO, *info)) { | |
2019 | g_free(*info); | |
2020 | *info = NULL; | |
2021 | return -errno; | |
2022 | } | |
2023 | ||
2024 | if (((*info)->argsz > argsz)) { | |
2025 | argsz = (*info)->argsz; | |
2026 | *info = g_realloc(*info, argsz); | |
2027 | goto again; | |
2028 | } | |
2029 | ||
2030 | return 0; | |
2031 | } | |
2032 | ||
2033 | static struct vfio_info_cap_header * | |
2034 | vfio_get_iommu_info_cap(struct vfio_iommu_type1_info *info, uint16_t id) | |
2035 | { | |
2036 | struct vfio_info_cap_header *hdr; | |
2037 | void *ptr = info; | |
2038 | ||
2039 | if (!(info->flags & VFIO_IOMMU_INFO_CAPS)) { | |
2040 | return NULL; | |
2041 | } | |
2042 | ||
2043 | for (hdr = ptr + info->cap_offset; hdr != ptr; hdr = ptr + hdr->next) { | |
2044 | if (hdr->id == id) { | |
2045 | return hdr; | |
2046 | } | |
2047 | } | |
2048 | ||
2049 | return NULL; | |
2050 | } | |
2051 | ||
2052 | static void vfio_get_iommu_info_migration(VFIOContainer *container, | |
2053 | struct vfio_iommu_type1_info *info) | |
2054 | { | |
2055 | struct vfio_info_cap_header *hdr; | |
2056 | struct vfio_iommu_type1_info_cap_migration *cap_mig; | |
2057 | ||
2058 | hdr = vfio_get_iommu_info_cap(info, VFIO_IOMMU_TYPE1_INFO_CAP_MIGRATION); | |
2059 | if (!hdr) { | |
2060 | return; | |
2061 | } | |
2062 | ||
2063 | cap_mig = container_of(hdr, struct vfio_iommu_type1_info_cap_migration, | |
2064 | header); | |
2065 | ||
2066 | /* | |
1eb7f642 KJ |
2067 | * cpu_physical_memory_set_dirty_lebitmap() supports pages in bitmap of |
2068 | * qemu_real_host_page_size to mark those dirty. | |
87ea529c | 2069 | */ |
8e3b0cbb | 2070 | if (cap_mig->pgsize_bitmap & qemu_real_host_page_size()) { |
87ea529c KW |
2071 | container->dirty_pages_supported = true; |
2072 | container->max_dirty_bitmap_size = cap_mig->max_dirty_bitmap_size; | |
2073 | container->dirty_pgsizes = cap_mig->pgsize_bitmap; | |
2074 | } | |
2075 | } | |
2076 | ||
01905f58 EA |
2077 | static int vfio_connect_container(VFIOGroup *group, AddressSpace *as, |
2078 | Error **errp) | |
e2c7d025 EA |
2079 | { |
2080 | VFIOContainer *container; | |
2081 | int ret, fd; | |
2082 | VFIOAddressSpace *space; | |
2083 | ||
2084 | space = vfio_get_address_space(as); | |
2085 | ||
c65ee433 | 2086 | /* |
aff92b82 | 2087 | * VFIO is currently incompatible with discarding of RAM insofar as the |
c65ee433 AW |
2088 | * madvise to purge (zap) the page from QEMU's address space does not |
2089 | * interact with the memory API and therefore leaves stale virtual to | |
2090 | * physical mappings in the IOMMU if the page was previously pinned. We | |
aff92b82 | 2091 | * therefore set discarding broken for each group added to a container, |
c65ee433 AW |
2092 | * whether the container is used individually or shared. This provides |
2093 | * us with options to allow devices within a group to opt-in and allow | |
aff92b82 | 2094 | * discarding, so long as it is done consistently for a group (for instance |
c65ee433 AW |
2095 | * if the device is an mdev device where it is known that the host vendor |
2096 | * driver will never pin pages outside of the working set of the guest | |
aff92b82 | 2097 | * driver, which would thus not be discarding candidates). |
c65ee433 AW |
2098 | * |
2099 | * The first opportunity to induce pinning occurs here where we attempt to | |
2100 | * attach the group to existing containers within the AddressSpace. If any | |
aff92b82 DH |
2101 | * pages are already zapped from the virtual address space, such as from |
2102 | * previous discards, new pinning will cause valid mappings to be | |
c65ee433 AW |
2103 | * re-established. Likewise, when the overall MemoryListener for a new |
2104 | * container is registered, a replay of mappings within the AddressSpace | |
2105 | * will occur, re-establishing any previously zapped pages as well. | |
2106 | * | |
aff92b82 DH |
2107 | * Especially virtio-balloon is currently only prevented from discarding |
2108 | * new memory, it will not yet set ram_block_discard_set_required() and | |
2109 | * therefore, neither stops us here or deals with the sudden memory | |
2110 | * consumption of inflated memory. | |
53d1b5fc DH |
2111 | * |
2112 | * We do support discarding of memory coordinated via the RamDiscardManager | |
2113 | * with some IOMMU types. vfio_ram_block_discard_disable() handles the | |
2114 | * details once we know which type of IOMMU we are using. | |
c65ee433 | 2115 | */ |
c65ee433 | 2116 | |
e2c7d025 EA |
2117 | QLIST_FOREACH(container, &space->containers, next) { |
2118 | if (!ioctl(group->fd, VFIO_GROUP_SET_CONTAINER, &container->fd)) { | |
53d1b5fc DH |
2119 | ret = vfio_ram_block_discard_disable(container, true); |
2120 | if (ret) { | |
2121 | error_setg_errno(errp, -ret, | |
2122 | "Cannot set discarding of RAM broken"); | |
2123 | if (ioctl(group->fd, VFIO_GROUP_UNSET_CONTAINER, | |
2124 | &container->fd)) { | |
2125 | error_report("vfio: error disconnecting group %d from" | |
2126 | " container", group->groupid); | |
2127 | } | |
2128 | return ret; | |
2129 | } | |
e2c7d025 EA |
2130 | group->container = container; |
2131 | QLIST_INSERT_HEAD(&container->group_list, group, container_next); | |
2016986a | 2132 | vfio_kvm_device_add_group(group); |
e2c7d025 EA |
2133 | return 0; |
2134 | } | |
2135 | } | |
2136 | ||
448058aa | 2137 | fd = qemu_open_old("/dev/vfio/vfio", O_RDWR); |
e2c7d025 | 2138 | if (fd < 0) { |
01905f58 | 2139 | error_setg_errno(errp, errno, "failed to open /dev/vfio/vfio"); |
e2c7d025 EA |
2140 | ret = -errno; |
2141 | goto put_space_exit; | |
2142 | } | |
2143 | ||
2144 | ret = ioctl(fd, VFIO_GET_API_VERSION); | |
2145 | if (ret != VFIO_API_VERSION) { | |
01905f58 EA |
2146 | error_setg(errp, "supported vfio version: %d, " |
2147 | "reported version: %d", VFIO_API_VERSION, ret); | |
e2c7d025 EA |
2148 | ret = -EINVAL; |
2149 | goto close_fd_exit; | |
2150 | } | |
2151 | ||
2152 | container = g_malloc0(sizeof(*container)); | |
2153 | container->space = space; | |
2154 | container->fd = fd; | |
d7d87836 | 2155 | container->error = NULL; |
87ea529c | 2156 | container->dirty_pages_supported = false; |
3eed155c | 2157 | container->dma_max_mappings = 0; |
f7f9c7b2 LY |
2158 | QLIST_INIT(&container->giommu_list); |
2159 | QLIST_INIT(&container->hostwin_list); | |
5e3b981c | 2160 | QLIST_INIT(&container->vrdl_list); |
2e6e697e | 2161 | |
2b6326c0 EA |
2162 | ret = vfio_init_container(container, group->fd, errp); |
2163 | if (ret) { | |
2164 | goto free_container_exit; | |
2165 | } | |
e2c7d025 | 2166 | |
53d1b5fc DH |
2167 | ret = vfio_ram_block_discard_disable(container, true); |
2168 | if (ret) { | |
2169 | error_setg_errno(errp, -ret, "Cannot set discarding of RAM broken"); | |
2170 | goto free_container_exit; | |
2171 | } | |
2172 | ||
2b6326c0 EA |
2173 | switch (container->iommu_type) { |
2174 | case VFIO_TYPE1v2_IOMMU: | |
2175 | case VFIO_TYPE1_IOMMU: | |
2176 | { | |
87ea529c | 2177 | struct vfio_iommu_type1_info *info; |
3898aad3 | 2178 | |
87ea529c | 2179 | ret = vfio_get_iommu_info(container, &info); |
85b6d2b5 AW |
2180 | if (ret) { |
2181 | error_setg_errno(errp, -ret, "Failed to get VFIO IOMMU info"); | |
2182 | goto enable_discards_exit; | |
2183 | } | |
87ea529c | 2184 | |
85b6d2b5 AW |
2185 | if (info->flags & VFIO_IOMMU_INFO_PGSIZES) { |
2186 | container->pgsizes = info->iova_pgsizes; | |
2187 | } else { | |
2188 | container->pgsizes = qemu_real_host_page_size(); | |
87ea529c | 2189 | } |
85b6d2b5 AW |
2190 | |
2191 | if (!vfio_get_info_dma_avail(info, &container->dma_max_mappings)) { | |
2192 | container->dma_max_mappings = 65535; | |
7a140a57 | 2193 | } |
85b6d2b5 | 2194 | vfio_get_iommu_info_migration(container, info); |
87ea529c | 2195 | g_free(info); |
85b6d2b5 AW |
2196 | |
2197 | /* | |
2198 | * FIXME: We should parse VFIO_IOMMU_TYPE1_INFO_CAP_IOVA_RANGE | |
2199 | * information to get the actual window extent rather than assume | |
2200 | * a 64-bit IOVA address space. | |
2201 | */ | |
2202 | vfio_host_win_add(container, 0, (hwaddr)-1, container->pgsizes); | |
2203 | ||
2b6326c0 EA |
2204 | break; |
2205 | } | |
2206 | case VFIO_SPAPR_TCE_v2_IOMMU: | |
2207 | case VFIO_SPAPR_TCE_IOMMU: | |
2208 | { | |
3898aad3 | 2209 | struct vfio_iommu_spapr_tce_info info; |
2b6326c0 | 2210 | bool v2 = container->iommu_type == VFIO_SPAPR_TCE_v2_IOMMU; |
e2c7d025 EA |
2211 | |
2212 | /* | |
2213 | * The host kernel code implementing VFIO_IOMMU_DISABLE is called | |
2214 | * when container fd is closed so we do not call it explicitly | |
2215 | * in this file. | |
2216 | */ | |
318f67ce AK |
2217 | if (!v2) { |
2218 | ret = ioctl(fd, VFIO_IOMMU_ENABLE); | |
2219 | if (ret) { | |
01905f58 | 2220 | error_setg_errno(errp, errno, "failed to enable container"); |
318f67ce | 2221 | ret = -errno; |
53d1b5fc | 2222 | goto enable_discards_exit; |
318f67ce AK |
2223 | } |
2224 | } else { | |
2225 | container->prereg_listener = vfio_prereg_listener; | |
2226 | ||
2227 | memory_listener_register(&container->prereg_listener, | |
2228 | &address_space_memory); | |
2229 | if (container->error) { | |
2230 | memory_listener_unregister(&container->prereg_listener); | |
d7d87836 EA |
2231 | ret = -1; |
2232 | error_propagate_prepend(errp, container->error, | |
2233 | "RAM memory listener initialization failed: "); | |
53d1b5fc | 2234 | goto enable_discards_exit; |
318f67ce | 2235 | } |
e2c7d025 | 2236 | } |
3898aad3 | 2237 | |
3898aad3 DG |
2238 | info.argsz = sizeof(info); |
2239 | ret = ioctl(fd, VFIO_IOMMU_SPAPR_TCE_GET_INFO, &info); | |
2240 | if (ret) { | |
01905f58 EA |
2241 | error_setg_errno(errp, errno, |
2242 | "VFIO_IOMMU_SPAPR_TCE_GET_INFO failed"); | |
3898aad3 | 2243 | ret = -errno; |
318f67ce AK |
2244 | if (v2) { |
2245 | memory_listener_unregister(&container->prereg_listener); | |
2246 | } | |
53d1b5fc | 2247 | goto enable_discards_exit; |
3898aad3 | 2248 | } |
7a140a57 | 2249 | |
2e4109de | 2250 | if (v2) { |
c26bc185 | 2251 | container->pgsizes = info.ddw.pgsizes; |
2e4109de AK |
2252 | /* |
2253 | * There is a default window in just created container. | |
2254 | * To make region_add/del simpler, we better remove this | |
2255 | * window now and let those iommu_listener callbacks | |
2256 | * create/remove them when needed. | |
2257 | */ | |
2258 | ret = vfio_spapr_remove_window(container, info.dma32_window_start); | |
2259 | if (ret) { | |
01905f58 EA |
2260 | error_setg_errno(errp, -ret, |
2261 | "failed to remove existing window"); | |
53d1b5fc | 2262 | goto enable_discards_exit; |
2e4109de AK |
2263 | } |
2264 | } else { | |
2265 | /* The default table uses 4K pages */ | |
c26bc185 | 2266 | container->pgsizes = 0x1000; |
2e4109de AK |
2267 | vfio_host_win_add(container, info.dma32_window_start, |
2268 | info.dma32_window_start + | |
2269 | info.dma32_window_size - 1, | |
2270 | 0x1000); | |
2271 | } | |
2b6326c0 | 2272 | } |
e2c7d025 EA |
2273 | } |
2274 | ||
8c37faa4 AK |
2275 | vfio_kvm_device_add_group(group); |
2276 | ||
2277 | QLIST_INIT(&container->group_list); | |
2278 | QLIST_INSERT_HEAD(&space->containers, container, next); | |
2279 | ||
2280 | group->container = container; | |
2281 | QLIST_INSERT_HEAD(&container->group_list, group, container_next); | |
2282 | ||
ee0bf0e5 DG |
2283 | container->listener = vfio_memory_listener; |
2284 | ||
2285 | memory_listener_register(&container->listener, container->space->as); | |
2286 | ||
2287 | if (container->error) { | |
d7d87836 EA |
2288 | ret = -1; |
2289 | error_propagate_prepend(errp, container->error, | |
2290 | "memory listener initialization failed: "); | |
ee0bf0e5 DG |
2291 | goto listener_release_exit; |
2292 | } | |
2293 | ||
2294 | container->initialized = true; | |
2295 | ||
e2c7d025 EA |
2296 | return 0; |
2297 | listener_release_exit: | |
8c37faa4 AK |
2298 | QLIST_REMOVE(group, container_next); |
2299 | QLIST_REMOVE(container, next); | |
2300 | vfio_kvm_device_del_group(group); | |
e2c7d025 EA |
2301 | vfio_listener_release(container); |
2302 | ||
53d1b5fc DH |
2303 | enable_discards_exit: |
2304 | vfio_ram_block_discard_disable(container, false); | |
2305 | ||
e2c7d025 EA |
2306 | free_container_exit: |
2307 | g_free(container); | |
2308 | ||
2309 | close_fd_exit: | |
2310 | close(fd); | |
2311 | ||
2312 | put_space_exit: | |
2313 | vfio_put_address_space(space); | |
2314 | ||
2315 | return ret; | |
2316 | } | |
2317 | ||
2318 | static void vfio_disconnect_container(VFIOGroup *group) | |
2319 | { | |
2320 | VFIOContainer *container = group->container; | |
2321 | ||
36968626 PX |
2322 | QLIST_REMOVE(group, container_next); |
2323 | group->container = NULL; | |
2324 | ||
2325 | /* | |
2326 | * Explicitly release the listener first before unset container, | |
2327 | * since unset may destroy the backend container if it's the last | |
2328 | * group. | |
2329 | */ | |
2330 | if (QLIST_EMPTY(&container->group_list)) { | |
2331 | vfio_listener_release(container); | |
2332 | } | |
2333 | ||
e2c7d025 EA |
2334 | if (ioctl(group->fd, VFIO_GROUP_UNSET_CONTAINER, &container->fd)) { |
2335 | error_report("vfio: error disconnecting group %d from container", | |
2336 | group->groupid); | |
2337 | } | |
2338 | ||
e2c7d025 EA |
2339 | if (QLIST_EMPTY(&container->group_list)) { |
2340 | VFIOAddressSpace *space = container->space; | |
f8d8a944 | 2341 | VFIOGuestIOMMU *giommu, *tmp; |
f3bc3a73 | 2342 | VFIOHostDMAWindow *hostwin, *next; |
e2c7d025 | 2343 | |
e2c7d025 | 2344 | QLIST_REMOVE(container, next); |
f8d8a944 AK |
2345 | |
2346 | QLIST_FOREACH_SAFE(giommu, &container->giommu_list, giommu_next, tmp) { | |
3df9d748 | 2347 | memory_region_unregister_iommu_notifier( |
44ee6aaa | 2348 | MEMORY_REGION(giommu->iommu_mr), &giommu->n); |
f8d8a944 AK |
2349 | QLIST_REMOVE(giommu, giommu_next); |
2350 | g_free(giommu); | |
2351 | } | |
2352 | ||
f3bc3a73 PL |
2353 | QLIST_FOREACH_SAFE(hostwin, &container->hostwin_list, hostwin_next, |
2354 | next) { | |
2355 | QLIST_REMOVE(hostwin, hostwin_next); | |
2356 | g_free(hostwin); | |
2357 | } | |
2358 | ||
e2c7d025 EA |
2359 | trace_vfio_disconnect_container(container->fd); |
2360 | close(container->fd); | |
2361 | g_free(container); | |
2362 | ||
2363 | vfio_put_address_space(space); | |
2364 | } | |
2365 | } | |
2366 | ||
1b808d5b | 2367 | VFIOGroup *vfio_get_group(int groupid, AddressSpace *as, Error **errp) |
e2c7d025 EA |
2368 | { |
2369 | VFIOGroup *group; | |
2370 | char path[32]; | |
2371 | struct vfio_group_status status = { .argsz = sizeof(status) }; | |
2372 | ||
2373 | QLIST_FOREACH(group, &vfio_group_list, next) { | |
2374 | if (group->groupid == groupid) { | |
2375 | /* Found it. Now is it already in the right context? */ | |
2376 | if (group->container->space->as == as) { | |
2377 | return group; | |
2378 | } else { | |
1b808d5b EA |
2379 | error_setg(errp, "group %d used in multiple address spaces", |
2380 | group->groupid); | |
e2c7d025 EA |
2381 | return NULL; |
2382 | } | |
2383 | } | |
2384 | } | |
2385 | ||
2386 | group = g_malloc0(sizeof(*group)); | |
2387 | ||
2388 | snprintf(path, sizeof(path), "/dev/vfio/%d", groupid); | |
448058aa | 2389 | group->fd = qemu_open_old(path, O_RDWR); |
e2c7d025 | 2390 | if (group->fd < 0) { |
1b808d5b | 2391 | error_setg_errno(errp, errno, "failed to open %s", path); |
e2c7d025 EA |
2392 | goto free_group_exit; |
2393 | } | |
2394 | ||
2395 | if (ioctl(group->fd, VFIO_GROUP_GET_STATUS, &status)) { | |
1b808d5b | 2396 | error_setg_errno(errp, errno, "failed to get group %d status", groupid); |
e2c7d025 EA |
2397 | goto close_fd_exit; |
2398 | } | |
2399 | ||
2400 | if (!(status.flags & VFIO_GROUP_FLAGS_VIABLE)) { | |
1b808d5b EA |
2401 | error_setg(errp, "group %d is not viable", groupid); |
2402 | error_append_hint(errp, | |
2403 | "Please ensure all devices within the iommu_group " | |
2404 | "are bound to their vfio bus driver.\n"); | |
e2c7d025 EA |
2405 | goto close_fd_exit; |
2406 | } | |
2407 | ||
2408 | group->groupid = groupid; | |
2409 | QLIST_INIT(&group->device_list); | |
2410 | ||
1b808d5b EA |
2411 | if (vfio_connect_container(group, as, errp)) { |
2412 | error_prepend(errp, "failed to setup container for group %d: ", | |
2413 | groupid); | |
e2c7d025 EA |
2414 | goto close_fd_exit; |
2415 | } | |
2416 | ||
2417 | if (QLIST_EMPTY(&vfio_group_list)) { | |
2418 | qemu_register_reset(vfio_reset_handler, NULL); | |
2419 | } | |
2420 | ||
2421 | QLIST_INSERT_HEAD(&vfio_group_list, group, next); | |
2422 | ||
e2c7d025 EA |
2423 | return group; |
2424 | ||
2425 | close_fd_exit: | |
2426 | close(group->fd); | |
2427 | ||
2428 | free_group_exit: | |
2429 | g_free(group); | |
2430 | ||
2431 | return NULL; | |
2432 | } | |
2433 | ||
2434 | void vfio_put_group(VFIOGroup *group) | |
2435 | { | |
77a10d04 | 2436 | if (!group || !QLIST_EMPTY(&group->device_list)) { |
e2c7d025 EA |
2437 | return; |
2438 | } | |
2439 | ||
aff92b82 | 2440 | if (!group->ram_block_discard_allowed) { |
53d1b5fc | 2441 | vfio_ram_block_discard_disable(group->container, false); |
238e9172 | 2442 | } |
e2c7d025 EA |
2443 | vfio_kvm_device_del_group(group); |
2444 | vfio_disconnect_container(group); | |
2445 | QLIST_REMOVE(group, next); | |
2446 | trace_vfio_put_group(group->fd); | |
2447 | close(group->fd); | |
2448 | g_free(group); | |
2449 | ||
2450 | if (QLIST_EMPTY(&vfio_group_list)) { | |
2451 | qemu_unregister_reset(vfio_reset_handler, NULL); | |
2452 | } | |
2453 | } | |
2454 | ||
2455 | int vfio_get_device(VFIOGroup *group, const char *name, | |
59f7d674 | 2456 | VFIODevice *vbasedev, Error **errp) |
e2c7d025 EA |
2457 | { |
2458 | struct vfio_device_info dev_info = { .argsz = sizeof(dev_info) }; | |
217e9fdc | 2459 | int ret, fd; |
e2c7d025 | 2460 | |
217e9fdc PB |
2461 | fd = ioctl(group->fd, VFIO_GROUP_GET_DEVICE_FD, name); |
2462 | if (fd < 0) { | |
59f7d674 EA |
2463 | error_setg_errno(errp, errno, "error getting device from group %d", |
2464 | group->groupid); | |
2465 | error_append_hint(errp, | |
2466 | "Verify all devices in group %d are bound to vfio-<bus> " | |
2467 | "or pci-stub and not already in use\n", group->groupid); | |
217e9fdc | 2468 | return fd; |
e2c7d025 EA |
2469 | } |
2470 | ||
217e9fdc | 2471 | ret = ioctl(fd, VFIO_DEVICE_GET_INFO, &dev_info); |
e2c7d025 | 2472 | if (ret) { |
59f7d674 | 2473 | error_setg_errno(errp, errno, "error getting device info"); |
217e9fdc PB |
2474 | close(fd); |
2475 | return ret; | |
e2c7d025 EA |
2476 | } |
2477 | ||
238e9172 | 2478 | /* |
aff92b82 DH |
2479 | * Set discarding of RAM as not broken for this group if the driver knows |
2480 | * the device operates compatibly with discarding. Setting must be | |
2481 | * consistent per group, but since compatibility is really only possible | |
2482 | * with mdev currently, we expect singleton groups. | |
238e9172 | 2483 | */ |
aff92b82 DH |
2484 | if (vbasedev->ram_block_discard_allowed != |
2485 | group->ram_block_discard_allowed) { | |
238e9172 | 2486 | if (!QLIST_EMPTY(&group->device_list)) { |
aff92b82 DH |
2487 | error_setg(errp, "Inconsistent setting of support for discarding " |
2488 | "RAM (e.g., balloon) within group"); | |
8709b395 | 2489 | close(fd); |
238e9172 AW |
2490 | return -1; |
2491 | } | |
2492 | ||
aff92b82 DH |
2493 | if (!group->ram_block_discard_allowed) { |
2494 | group->ram_block_discard_allowed = true; | |
53d1b5fc | 2495 | vfio_ram_block_discard_disable(group->container, false); |
238e9172 AW |
2496 | } |
2497 | } | |
2498 | ||
217e9fdc PB |
2499 | vbasedev->fd = fd; |
2500 | vbasedev->group = group; | |
2501 | QLIST_INSERT_HEAD(&group->device_list, vbasedev, next); | |
2502 | ||
e2c7d025 EA |
2503 | vbasedev->num_irqs = dev_info.num_irqs; |
2504 | vbasedev->num_regions = dev_info.num_regions; | |
2505 | vbasedev->flags = dev_info.flags; | |
2506 | ||
2507 | trace_vfio_get_device(name, dev_info.flags, dev_info.num_regions, | |
2508 | dev_info.num_irqs); | |
2509 | ||
2510 | vbasedev->reset_works = !!(dev_info.flags & VFIO_DEVICE_FLAGS_RESET); | |
217e9fdc | 2511 | return 0; |
e2c7d025 EA |
2512 | } |
2513 | ||
2514 | void vfio_put_base_device(VFIODevice *vbasedev) | |
2515 | { | |
77a10d04 PB |
2516 | if (!vbasedev->group) { |
2517 | return; | |
2518 | } | |
e2c7d025 EA |
2519 | QLIST_REMOVE(vbasedev, next); |
2520 | vbasedev->group = NULL; | |
2521 | trace_vfio_put_base_device(vbasedev->fd); | |
2522 | close(vbasedev->fd); | |
2523 | } | |
2524 | ||
46900226 AW |
2525 | int vfio_get_region_info(VFIODevice *vbasedev, int index, |
2526 | struct vfio_region_info **info) | |
2527 | { | |
2528 | size_t argsz = sizeof(struct vfio_region_info); | |
2529 | ||
2530 | *info = g_malloc0(argsz); | |
2531 | ||
2532 | (*info)->index = index; | |
b53b0f69 | 2533 | retry: |
46900226 AW |
2534 | (*info)->argsz = argsz; |
2535 | ||
2536 | if (ioctl(vbasedev->fd, VFIO_DEVICE_GET_REGION_INFO, *info)) { | |
2537 | g_free(*info); | |
e61a424f | 2538 | *info = NULL; |
46900226 AW |
2539 | return -errno; |
2540 | } | |
2541 | ||
b53b0f69 AW |
2542 | if ((*info)->argsz > argsz) { |
2543 | argsz = (*info)->argsz; | |
2544 | *info = g_realloc(*info, argsz); | |
2545 | ||
2546 | goto retry; | |
2547 | } | |
2548 | ||
46900226 AW |
2549 | return 0; |
2550 | } | |
2551 | ||
e61a424f AW |
2552 | int vfio_get_dev_region_info(VFIODevice *vbasedev, uint32_t type, |
2553 | uint32_t subtype, struct vfio_region_info **info) | |
2554 | { | |
2555 | int i; | |
2556 | ||
2557 | for (i = 0; i < vbasedev->num_regions; i++) { | |
2558 | struct vfio_info_cap_header *hdr; | |
2559 | struct vfio_region_info_cap_type *cap_type; | |
2560 | ||
2561 | if (vfio_get_region_info(vbasedev, i, info)) { | |
2562 | continue; | |
2563 | } | |
2564 | ||
2565 | hdr = vfio_get_region_info_cap(*info, VFIO_REGION_INFO_CAP_TYPE); | |
2566 | if (!hdr) { | |
2567 | g_free(*info); | |
2568 | continue; | |
2569 | } | |
2570 | ||
2571 | cap_type = container_of(hdr, struct vfio_region_info_cap_type, header); | |
2572 | ||
2573 | trace_vfio_get_dev_region(vbasedev->name, i, | |
2574 | cap_type->type, cap_type->subtype); | |
2575 | ||
2576 | if (cap_type->type == type && cap_type->subtype == subtype) { | |
2577 | return 0; | |
2578 | } | |
2579 | ||
2580 | g_free(*info); | |
2581 | } | |
2582 | ||
2583 | *info = NULL; | |
2584 | return -ENODEV; | |
2585 | } | |
2586 | ||
ae0215b2 AK |
2587 | bool vfio_has_region_cap(VFIODevice *vbasedev, int region, uint16_t cap_type) |
2588 | { | |
2589 | struct vfio_region_info *info = NULL; | |
2590 | bool ret = false; | |
2591 | ||
2592 | if (!vfio_get_region_info(vbasedev, region, &info)) { | |
2593 | if (vfio_get_region_info_cap(info, cap_type)) { | |
2594 | ret = true; | |
2595 | } | |
2596 | g_free(info); | |
2597 | } | |
2598 | ||
2599 | return ret; | |
2600 | } | |
2601 | ||
3153119e DG |
2602 | /* |
2603 | * Interfaces for IBM EEH (Enhanced Error Handling) | |
2604 | */ | |
2605 | static bool vfio_eeh_container_ok(VFIOContainer *container) | |
2606 | { | |
2607 | /* | |
2608 | * As of 2016-03-04 (linux-4.5) the host kernel EEH/VFIO | |
2609 | * implementation is broken if there are multiple groups in a | |
2610 | * container. The hardware works in units of Partitionable | |
2611 | * Endpoints (== IOMMU groups) and the EEH operations naively | |
2612 | * iterate across all groups in the container, without any logic | |
2613 | * to make sure the groups have their state synchronized. For | |
2614 | * certain operations (ENABLE) that might be ok, until an error | |
2615 | * occurs, but for others (GET_STATE) it's clearly broken. | |
2616 | */ | |
2617 | ||
2618 | /* | |
2619 | * XXX Once fixed kernels exist, test for them here | |
2620 | */ | |
2621 | ||
2622 | if (QLIST_EMPTY(&container->group_list)) { | |
2623 | return false; | |
2624 | } | |
2625 | ||
2626 | if (QLIST_NEXT(QLIST_FIRST(&container->group_list), container_next)) { | |
2627 | return false; | |
2628 | } | |
2629 | ||
2630 | return true; | |
2631 | } | |
2632 | ||
2633 | static int vfio_eeh_container_op(VFIOContainer *container, uint32_t op) | |
2634 | { | |
2635 | struct vfio_eeh_pe_op pe_op = { | |
2636 | .argsz = sizeof(pe_op), | |
2637 | .op = op, | |
2638 | }; | |
2639 | int ret; | |
2640 | ||
2641 | if (!vfio_eeh_container_ok(container)) { | |
2642 | error_report("vfio/eeh: EEH_PE_OP 0x%x: " | |
2643 | "kernel requires a container with exactly one group", op); | |
2644 | return -EPERM; | |
2645 | } | |
2646 | ||
2647 | ret = ioctl(container->fd, VFIO_EEH_PE_OP, &pe_op); | |
2648 | if (ret < 0) { | |
2649 | error_report("vfio/eeh: EEH_PE_OP 0x%x failed: %m", op); | |
2650 | return -errno; | |
2651 | } | |
2652 | ||
d917e88d | 2653 | return ret; |
3153119e DG |
2654 | } |
2655 | ||
2656 | static VFIOContainer *vfio_eeh_as_container(AddressSpace *as) | |
2657 | { | |
2658 | VFIOAddressSpace *space = vfio_get_address_space(as); | |
2659 | VFIOContainer *container = NULL; | |
2660 | ||
2661 | if (QLIST_EMPTY(&space->containers)) { | |
2662 | /* No containers to act on */ | |
2663 | goto out; | |
2664 | } | |
2665 | ||
2666 | container = QLIST_FIRST(&space->containers); | |
2667 | ||
2668 | if (QLIST_NEXT(container, next)) { | |
2669 | /* We don't yet have logic to synchronize EEH state across | |
2670 | * multiple containers */ | |
2671 | container = NULL; | |
2672 | goto out; | |
2673 | } | |
2674 | ||
2675 | out: | |
2676 | vfio_put_address_space(space); | |
2677 | return container; | |
2678 | } | |
2679 | ||
2680 | bool vfio_eeh_as_ok(AddressSpace *as) | |
2681 | { | |
2682 | VFIOContainer *container = vfio_eeh_as_container(as); | |
2683 | ||
2684 | return (container != NULL) && vfio_eeh_container_ok(container); | |
2685 | } | |
2686 | ||
2687 | int vfio_eeh_as_op(AddressSpace *as, uint32_t op) | |
2688 | { | |
2689 | VFIOContainer *container = vfio_eeh_as_container(as); | |
2690 | ||
2691 | if (!container) { | |
2692 | return -ENODEV; | |
2693 | } | |
2694 | return vfio_eeh_container_op(container, op); | |
2695 | } |