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
21 #include "qemu/osdep.h"
22 #include <sys/ioctl.h>
23 #ifdef CONFIG_KVM
24 #include <linux/kvm.h>
25 #endif
26 #include <linux/vfio.h>
27
28 #include "hw/vfio/vfio-common.h"
29 #include "hw/vfio/pci.h"
30 #include "exec/address-spaces.h"
31 #include "exec/memory.h"
32 #include "exec/ram_addr.h"
33 #include "exec/target_page.h"
34 #include "hw/hw.h"
35 #include "qemu/error-report.h"
36 #include "qemu/main-loop.h"
37 #include "qemu/range.h"
38 #include "system/kvm.h"
39 #include "system/reset.h"
40 #include "system/runstate.h"
41 #include "trace.h"
42 #include "qapi/error.h"
43 #include "migration/misc.h"
44 #include "migration/blocker.h"
45 #include "migration/qemu-file.h"
46 #include "system/tcg.h"
47 #include "system/tpm.h"
48
49 VFIODeviceList vfio_device_list =
50 QLIST_HEAD_INITIALIZER(vfio_device_list);
51 static QLIST_HEAD(, VFIOAddressSpace) vfio_address_spaces =
52 QLIST_HEAD_INITIALIZER(vfio_address_spaces);
53
54 #ifdef CONFIG_KVM
55 /*
56 * We have a single VFIO pseudo device per KVM VM. Once created it lives
57 * for the life of the VM. Closing the file descriptor only drops our
58 * reference to it and the device's reference to kvm. Therefore once
59 * initialized, this file descriptor is only released on QEMU exit and
60 * we'll re-use it should another vfio device be attached before then.
61 */
62 int vfio_kvm_device_fd = -1;
63 #endif
64
65 /*
66 * Device state interfaces
67 */
68
vfio_mig_active(void)69 bool vfio_mig_active(void)
70 {
71 VFIODevice *vbasedev;
72
73 if (QLIST_EMPTY(&vfio_device_list)) {
74 return false;
75 }
76
77 QLIST_FOREACH(vbasedev, &vfio_device_list, global_next) {
78 if (vbasedev->migration_blocker) {
79 return false;
80 }
81 }
82 return true;
83 }
84
85 static Error *multiple_devices_migration_blocker;
86
87 /*
88 * Multiple devices migration is allowed only if all devices support P2P
89 * migration. Single device migration is allowed regardless of P2P migration
90 * support.
91 */
vfio_multiple_devices_migration_is_supported(void)92 static bool vfio_multiple_devices_migration_is_supported(void)
93 {
94 VFIODevice *vbasedev;
95 unsigned int device_num = 0;
96 bool all_support_p2p = true;
97
98 QLIST_FOREACH(vbasedev, &vfio_device_list, global_next) {
99 if (vbasedev->migration) {
100 device_num++;
101
102 if (!(vbasedev->migration->mig_flags & VFIO_MIGRATION_P2P)) {
103 all_support_p2p = false;
104 }
105 }
106 }
107
108 return all_support_p2p || device_num <= 1;
109 }
110
vfio_block_multiple_devices_migration(VFIODevice * vbasedev,Error ** errp)111 int vfio_block_multiple_devices_migration(VFIODevice *vbasedev, Error **errp)
112 {
113 int ret;
114
115 if (vfio_multiple_devices_migration_is_supported()) {
116 return 0;
117 }
118
119 if (vbasedev->enable_migration == ON_OFF_AUTO_ON) {
120 error_setg(errp, "Multiple VFIO devices migration is supported only if "
121 "all of them support P2P migration");
122 return -EINVAL;
123 }
124
125 if (multiple_devices_migration_blocker) {
126 return 0;
127 }
128
129 error_setg(&multiple_devices_migration_blocker,
130 "Multiple VFIO devices migration is supported only if all of "
131 "them support P2P migration");
132 ret = migrate_add_blocker_normal(&multiple_devices_migration_blocker, errp);
133
134 return ret;
135 }
136
vfio_unblock_multiple_devices_migration(void)137 void vfio_unblock_multiple_devices_migration(void)
138 {
139 if (!multiple_devices_migration_blocker ||
140 !vfio_multiple_devices_migration_is_supported()) {
141 return;
142 }
143
144 migrate_del_blocker(&multiple_devices_migration_blocker);
145 }
146
vfio_viommu_preset(VFIODevice * vbasedev)147 bool vfio_viommu_preset(VFIODevice *vbasedev)
148 {
149 return vbasedev->bcontainer->space->as != &address_space_memory;
150 }
151
vfio_set_migration_error(int ret)152 static void vfio_set_migration_error(int ret)
153 {
154 if (migration_is_running()) {
155 migration_file_set_error(ret, NULL);
156 }
157 }
158
vfio_device_state_is_running(VFIODevice * vbasedev)159 bool vfio_device_state_is_running(VFIODevice *vbasedev)
160 {
161 VFIOMigration *migration = vbasedev->migration;
162
163 return migration->device_state == VFIO_DEVICE_STATE_RUNNING ||
164 migration->device_state == VFIO_DEVICE_STATE_RUNNING_P2P;
165 }
166
vfio_device_state_is_precopy(VFIODevice * vbasedev)167 bool vfio_device_state_is_precopy(VFIODevice *vbasedev)
168 {
169 VFIOMigration *migration = vbasedev->migration;
170
171 return migration->device_state == VFIO_DEVICE_STATE_PRE_COPY ||
172 migration->device_state == VFIO_DEVICE_STATE_PRE_COPY_P2P;
173 }
174
vfio_devices_all_device_dirty_tracking_started(const VFIOContainerBase * bcontainer)175 static bool vfio_devices_all_device_dirty_tracking_started(
176 const VFIOContainerBase *bcontainer)
177 {
178 VFIODevice *vbasedev;
179
180 QLIST_FOREACH(vbasedev, &bcontainer->device_list, container_next) {
181 if (!vbasedev->dirty_tracking) {
182 return false;
183 }
184 }
185
186 return true;
187 }
188
vfio_devices_all_dirty_tracking_started(const VFIOContainerBase * bcontainer)189 bool vfio_devices_all_dirty_tracking_started(
190 const VFIOContainerBase *bcontainer)
191 {
192 return vfio_devices_all_device_dirty_tracking_started(bcontainer) ||
193 bcontainer->dirty_pages_started;
194 }
195
vfio_log_sync_needed(const VFIOContainerBase * bcontainer)196 static bool vfio_log_sync_needed(const VFIOContainerBase *bcontainer)
197 {
198 VFIODevice *vbasedev;
199
200 if (!vfio_devices_all_dirty_tracking_started(bcontainer)) {
201 return false;
202 }
203
204 QLIST_FOREACH(vbasedev, &bcontainer->device_list, container_next) {
205 VFIOMigration *migration = vbasedev->migration;
206
207 if (!migration) {
208 return false;
209 }
210
211 if (vbasedev->pre_copy_dirty_page_tracking == ON_OFF_AUTO_OFF &&
212 (vfio_device_state_is_running(vbasedev) ||
213 vfio_device_state_is_precopy(vbasedev))) {
214 return false;
215 }
216 }
217 return true;
218 }
219
vfio_devices_all_device_dirty_tracking(const VFIOContainerBase * bcontainer)220 bool vfio_devices_all_device_dirty_tracking(const VFIOContainerBase *bcontainer)
221 {
222 VFIODevice *vbasedev;
223
224 QLIST_FOREACH(vbasedev, &bcontainer->device_list, container_next) {
225 if (vbasedev->device_dirty_page_tracking == ON_OFF_AUTO_OFF) {
226 return false;
227 }
228 if (!vbasedev->dirty_pages_supported) {
229 return false;
230 }
231 }
232
233 return true;
234 }
235
vfio_listener_skipped_section(MemoryRegionSection * section)236 static bool vfio_listener_skipped_section(MemoryRegionSection *section)
237 {
238 return (!memory_region_is_ram(section->mr) &&
239 !memory_region_is_iommu(section->mr)) ||
240 memory_region_is_protected(section->mr) ||
241 /*
242 * Sizing an enabled 64-bit BAR can cause spurious mappings to
243 * addresses in the upper part of the 64-bit address space. These
244 * are never accessed by the CPU and beyond the address width of
245 * some IOMMU hardware. TODO: VFIO should tell us the IOMMU width.
246 */
247 section->offset_within_address_space & (1ULL << 63);
248 }
249
250 /* Called with rcu_read_lock held. */
vfio_get_xlat_addr(IOMMUTLBEntry * iotlb,void ** vaddr,ram_addr_t * ram_addr,bool * read_only,Error ** errp)251 static bool vfio_get_xlat_addr(IOMMUTLBEntry *iotlb, void **vaddr,
252 ram_addr_t *ram_addr, bool *read_only,
253 Error **errp)
254 {
255 bool ret, mr_has_discard_manager;
256
257 ret = memory_get_xlat_addr(iotlb, vaddr, ram_addr, read_only,
258 &mr_has_discard_manager, errp);
259 if (ret && mr_has_discard_manager) {
260 /*
261 * Malicious VMs might trigger discarding of IOMMU-mapped memory. The
262 * pages will remain pinned inside vfio until unmapped, resulting in a
263 * higher memory consumption than expected. If memory would get
264 * populated again later, there would be an inconsistency between pages
265 * pinned by vfio and pages seen by QEMU. This is the case until
266 * unmapped from the IOMMU (e.g., during device reset).
267 *
268 * With malicious guests, we really only care about pinning more memory
269 * than expected. RLIMIT_MEMLOCK set for the user/process can never be
270 * exceeded and can be used to mitigate this problem.
271 */
272 warn_report_once("Using vfio with vIOMMUs and coordinated discarding of"
273 " RAM (e.g., virtio-mem) works, however, malicious"
274 " guests can trigger pinning of more memory than"
275 " intended via an IOMMU. It's possible to mitigate "
276 " by setting/adjusting RLIMIT_MEMLOCK.");
277 }
278 return ret;
279 }
280
vfio_iommu_map_notify(IOMMUNotifier * n,IOMMUTLBEntry * iotlb)281 static void vfio_iommu_map_notify(IOMMUNotifier *n, IOMMUTLBEntry *iotlb)
282 {
283 VFIOGuestIOMMU *giommu = container_of(n, VFIOGuestIOMMU, n);
284 VFIOContainerBase *bcontainer = giommu->bcontainer;
285 hwaddr iova = iotlb->iova + giommu->iommu_offset;
286 void *vaddr;
287 int ret;
288 Error *local_err = NULL;
289
290 trace_vfio_iommu_map_notify(iotlb->perm == IOMMU_NONE ? "UNMAP" : "MAP",
291 iova, iova + iotlb->addr_mask);
292
293 if (iotlb->target_as != &address_space_memory) {
294 error_report("Wrong target AS \"%s\", only system memory is allowed",
295 iotlb->target_as->name ? iotlb->target_as->name : "none");
296 vfio_set_migration_error(-EINVAL);
297 return;
298 }
299
300 rcu_read_lock();
301
302 if ((iotlb->perm & IOMMU_RW) != IOMMU_NONE) {
303 bool read_only;
304
305 if (!vfio_get_xlat_addr(iotlb, &vaddr, NULL, &read_only, &local_err)) {
306 error_report_err(local_err);
307 goto out;
308 }
309 /*
310 * vaddr is only valid until rcu_read_unlock(). But after
311 * vfio_dma_map has set up the mapping the pages will be
312 * pinned by the kernel. This makes sure that the RAM backend
313 * of vaddr will always be there, even if the memory object is
314 * destroyed and its backing memory munmap-ed.
315 */
316 ret = vfio_container_dma_map(bcontainer, iova,
317 iotlb->addr_mask + 1, vaddr,
318 read_only);
319 if (ret) {
320 error_report("vfio_container_dma_map(%p, 0x%"HWADDR_PRIx", "
321 "0x%"HWADDR_PRIx", %p) = %d (%s)",
322 bcontainer, iova,
323 iotlb->addr_mask + 1, vaddr, ret, strerror(-ret));
324 }
325 } else {
326 ret = vfio_container_dma_unmap(bcontainer, iova,
327 iotlb->addr_mask + 1, iotlb);
328 if (ret) {
329 error_report("vfio_container_dma_unmap(%p, 0x%"HWADDR_PRIx", "
330 "0x%"HWADDR_PRIx") = %d (%s)",
331 bcontainer, iova,
332 iotlb->addr_mask + 1, ret, strerror(-ret));
333 vfio_set_migration_error(ret);
334 }
335 }
336 out:
337 rcu_read_unlock();
338 }
339
vfio_ram_discard_notify_discard(RamDiscardListener * rdl,MemoryRegionSection * section)340 static void vfio_ram_discard_notify_discard(RamDiscardListener *rdl,
341 MemoryRegionSection *section)
342 {
343 VFIORamDiscardListener *vrdl = container_of(rdl, VFIORamDiscardListener,
344 listener);
345 VFIOContainerBase *bcontainer = vrdl->bcontainer;
346 const hwaddr size = int128_get64(section->size);
347 const hwaddr iova = section->offset_within_address_space;
348 int ret;
349
350 /* Unmap with a single call. */
351 ret = vfio_container_dma_unmap(bcontainer, iova, size , NULL);
352 if (ret) {
353 error_report("%s: vfio_container_dma_unmap() failed: %s", __func__,
354 strerror(-ret));
355 }
356 }
357
vfio_ram_discard_notify_populate(RamDiscardListener * rdl,MemoryRegionSection * section)358 static int vfio_ram_discard_notify_populate(RamDiscardListener *rdl,
359 MemoryRegionSection *section)
360 {
361 VFIORamDiscardListener *vrdl = container_of(rdl, VFIORamDiscardListener,
362 listener);
363 VFIOContainerBase *bcontainer = vrdl->bcontainer;
364 const hwaddr end = section->offset_within_region +
365 int128_get64(section->size);
366 hwaddr start, next, iova;
367 void *vaddr;
368 int ret;
369
370 /*
371 * Map in (aligned within memory region) minimum granularity, so we can
372 * unmap in minimum granularity later.
373 */
374 for (start = section->offset_within_region; start < end; start = next) {
375 next = ROUND_UP(start + 1, vrdl->granularity);
376 next = MIN(next, end);
377
378 iova = start - section->offset_within_region +
379 section->offset_within_address_space;
380 vaddr = memory_region_get_ram_ptr(section->mr) + start;
381
382 ret = vfio_container_dma_map(bcontainer, iova, next - start,
383 vaddr, section->readonly);
384 if (ret) {
385 /* Rollback */
386 vfio_ram_discard_notify_discard(rdl, section);
387 return ret;
388 }
389 }
390 return 0;
391 }
392
vfio_register_ram_discard_listener(VFIOContainerBase * bcontainer,MemoryRegionSection * section)393 static void vfio_register_ram_discard_listener(VFIOContainerBase *bcontainer,
394 MemoryRegionSection *section)
395 {
396 RamDiscardManager *rdm = memory_region_get_ram_discard_manager(section->mr);
397 int target_page_size = qemu_target_page_size();
398 VFIORamDiscardListener *vrdl;
399
400 /* Ignore some corner cases not relevant in practice. */
401 g_assert(QEMU_IS_ALIGNED(section->offset_within_region, target_page_size));
402 g_assert(QEMU_IS_ALIGNED(section->offset_within_address_space,
403 target_page_size));
404 g_assert(QEMU_IS_ALIGNED(int128_get64(section->size), target_page_size));
405
406 vrdl = g_new0(VFIORamDiscardListener, 1);
407 vrdl->bcontainer = bcontainer;
408 vrdl->mr = section->mr;
409 vrdl->offset_within_address_space = section->offset_within_address_space;
410 vrdl->size = int128_get64(section->size);
411 vrdl->granularity = ram_discard_manager_get_min_granularity(rdm,
412 section->mr);
413
414 g_assert(vrdl->granularity && is_power_of_2(vrdl->granularity));
415 g_assert(bcontainer->pgsizes &&
416 vrdl->granularity >= 1ULL << ctz64(bcontainer->pgsizes));
417
418 ram_discard_listener_init(&vrdl->listener,
419 vfio_ram_discard_notify_populate,
420 vfio_ram_discard_notify_discard, true);
421 ram_discard_manager_register_listener(rdm, &vrdl->listener, section);
422 QLIST_INSERT_HEAD(&bcontainer->vrdl_list, vrdl, next);
423
424 /*
425 * Sanity-check if we have a theoretically problematic setup where we could
426 * exceed the maximum number of possible DMA mappings over time. We assume
427 * that each mapped section in the same address space as a RamDiscardManager
428 * section consumes exactly one DMA mapping, with the exception of
429 * RamDiscardManager sections; i.e., we don't expect to have gIOMMU sections
430 * in the same address space as RamDiscardManager sections.
431 *
432 * We assume that each section in the address space consumes one memslot.
433 * We take the number of KVM memory slots as a best guess for the maximum
434 * number of sections in the address space we could have over time,
435 * also consuming DMA mappings.
436 */
437 if (bcontainer->dma_max_mappings) {
438 unsigned int vrdl_count = 0, vrdl_mappings = 0, max_memslots = 512;
439
440 #ifdef CONFIG_KVM
441 if (kvm_enabled()) {
442 max_memslots = kvm_get_max_memslots();
443 }
444 #endif
445
446 QLIST_FOREACH(vrdl, &bcontainer->vrdl_list, next) {
447 hwaddr start, end;
448
449 start = QEMU_ALIGN_DOWN(vrdl->offset_within_address_space,
450 vrdl->granularity);
451 end = ROUND_UP(vrdl->offset_within_address_space + vrdl->size,
452 vrdl->granularity);
453 vrdl_mappings += (end - start) / vrdl->granularity;
454 vrdl_count++;
455 }
456
457 if (vrdl_mappings + max_memslots - vrdl_count >
458 bcontainer->dma_max_mappings) {
459 warn_report("%s: possibly running out of DMA mappings. E.g., try"
460 " increasing the 'block-size' of virtio-mem devies."
461 " Maximum possible DMA mappings: %d, Maximum possible"
462 " memslots: %d", __func__, bcontainer->dma_max_mappings,
463 max_memslots);
464 }
465 }
466 }
467
vfio_unregister_ram_discard_listener(VFIOContainerBase * bcontainer,MemoryRegionSection * section)468 static void vfio_unregister_ram_discard_listener(VFIOContainerBase *bcontainer,
469 MemoryRegionSection *section)
470 {
471 RamDiscardManager *rdm = memory_region_get_ram_discard_manager(section->mr);
472 VFIORamDiscardListener *vrdl = NULL;
473
474 QLIST_FOREACH(vrdl, &bcontainer->vrdl_list, next) {
475 if (vrdl->mr == section->mr &&
476 vrdl->offset_within_address_space ==
477 section->offset_within_address_space) {
478 break;
479 }
480 }
481
482 if (!vrdl) {
483 hw_error("vfio: Trying to unregister missing RAM discard listener");
484 }
485
486 ram_discard_manager_unregister_listener(rdm, &vrdl->listener);
487 QLIST_REMOVE(vrdl, next);
488 g_free(vrdl);
489 }
490
vfio_known_safe_misalignment(MemoryRegionSection * section)491 static bool vfio_known_safe_misalignment(MemoryRegionSection *section)
492 {
493 MemoryRegion *mr = section->mr;
494
495 if (!TPM_IS_CRB(mr->owner)) {
496 return false;
497 }
498
499 /* this is a known safe misaligned region, just trace for debug purpose */
500 trace_vfio_known_safe_misalignment(memory_region_name(mr),
501 section->offset_within_address_space,
502 section->offset_within_region,
503 qemu_real_host_page_size());
504 return true;
505 }
506
vfio_listener_valid_section(MemoryRegionSection * section,const char * name)507 static bool vfio_listener_valid_section(MemoryRegionSection *section,
508 const char *name)
509 {
510 if (vfio_listener_skipped_section(section)) {
511 trace_vfio_listener_region_skip(name,
512 section->offset_within_address_space,
513 section->offset_within_address_space +
514 int128_get64(int128_sub(section->size, int128_one())));
515 return false;
516 }
517
518 if (unlikely((section->offset_within_address_space &
519 ~qemu_real_host_page_mask()) !=
520 (section->offset_within_region & ~qemu_real_host_page_mask()))) {
521 if (!vfio_known_safe_misalignment(section)) {
522 error_report("%s received unaligned region %s iova=0x%"PRIx64
523 " offset_within_region=0x%"PRIx64
524 " qemu_real_host_page_size=0x%"PRIxPTR,
525 __func__, memory_region_name(section->mr),
526 section->offset_within_address_space,
527 section->offset_within_region,
528 qemu_real_host_page_size());
529 }
530 return false;
531 }
532
533 return true;
534 }
535
vfio_get_section_iova_range(VFIOContainerBase * bcontainer,MemoryRegionSection * section,hwaddr * out_iova,hwaddr * out_end,Int128 * out_llend)536 static bool vfio_get_section_iova_range(VFIOContainerBase *bcontainer,
537 MemoryRegionSection *section,
538 hwaddr *out_iova, hwaddr *out_end,
539 Int128 *out_llend)
540 {
541 Int128 llend;
542 hwaddr iova;
543
544 iova = REAL_HOST_PAGE_ALIGN(section->offset_within_address_space);
545 llend = int128_make64(section->offset_within_address_space);
546 llend = int128_add(llend, section->size);
547 llend = int128_and(llend, int128_exts64(qemu_real_host_page_mask()));
548
549 if (int128_ge(int128_make64(iova), llend)) {
550 return false;
551 }
552
553 *out_iova = iova;
554 *out_end = int128_get64(int128_sub(llend, int128_one()));
555 if (out_llend) {
556 *out_llend = llend;
557 }
558 return true;
559 }
560
vfio_device_error_append(VFIODevice * vbasedev,Error ** errp)561 static void vfio_device_error_append(VFIODevice *vbasedev, Error **errp)
562 {
563 /*
564 * MMIO region mapping failures are not fatal but in this case PCI
565 * peer-to-peer transactions are broken.
566 */
567 if (vbasedev && vbasedev->type == VFIO_DEVICE_TYPE_PCI) {
568 error_append_hint(errp, "%s: PCI peer-to-peer transactions "
569 "on BARs are not supported.\n", vbasedev->name);
570 }
571 }
572
vfio_listener_region_add(MemoryListener * listener,MemoryRegionSection * section)573 static void vfio_listener_region_add(MemoryListener *listener,
574 MemoryRegionSection *section)
575 {
576 VFIOContainerBase *bcontainer = container_of(listener, VFIOContainerBase,
577 listener);
578 hwaddr iova, end;
579 Int128 llend, llsize;
580 void *vaddr;
581 int ret;
582 Error *err = NULL;
583
584 if (!vfio_listener_valid_section(section, "region_add")) {
585 return;
586 }
587
588 if (!vfio_get_section_iova_range(bcontainer, section, &iova, &end,
589 &llend)) {
590 if (memory_region_is_ram_device(section->mr)) {
591 trace_vfio_listener_region_add_no_dma_map(
592 memory_region_name(section->mr),
593 section->offset_within_address_space,
594 int128_getlo(section->size),
595 qemu_real_host_page_size());
596 }
597 return;
598 }
599
600 /* PPC64/pseries machine only */
601 if (!vfio_container_add_section_window(bcontainer, section, &err)) {
602 goto mmio_dma_error;
603 }
604
605 memory_region_ref(section->mr);
606
607 if (memory_region_is_iommu(section->mr)) {
608 VFIOGuestIOMMU *giommu;
609 IOMMUMemoryRegion *iommu_mr = IOMMU_MEMORY_REGION(section->mr);
610 int iommu_idx;
611
612 trace_vfio_listener_region_add_iommu(section->mr->name, iova, end);
613 /*
614 * FIXME: For VFIO iommu types which have KVM acceleration to
615 * avoid bouncing all map/unmaps through qemu this way, this
616 * would be the right place to wire that up (tell the KVM
617 * device emulation the VFIO iommu handles to use).
618 */
619 giommu = g_malloc0(sizeof(*giommu));
620 giommu->iommu_mr = iommu_mr;
621 giommu->iommu_offset = section->offset_within_address_space -
622 section->offset_within_region;
623 giommu->bcontainer = bcontainer;
624 llend = int128_add(int128_make64(section->offset_within_region),
625 section->size);
626 llend = int128_sub(llend, int128_one());
627 iommu_idx = memory_region_iommu_attrs_to_index(iommu_mr,
628 MEMTXATTRS_UNSPECIFIED);
629 iommu_notifier_init(&giommu->n, vfio_iommu_map_notify,
630 IOMMU_NOTIFIER_IOTLB_EVENTS,
631 section->offset_within_region,
632 int128_get64(llend),
633 iommu_idx);
634
635 ret = memory_region_register_iommu_notifier(section->mr, &giommu->n,
636 &err);
637 if (ret) {
638 g_free(giommu);
639 goto fail;
640 }
641 QLIST_INSERT_HEAD(&bcontainer->giommu_list, giommu, giommu_next);
642 memory_region_iommu_replay(giommu->iommu_mr, &giommu->n);
643
644 return;
645 }
646
647 /* Here we assume that memory_region_is_ram(section->mr)==true */
648
649 /*
650 * For RAM memory regions with a RamDiscardManager, we only want to map the
651 * actually populated parts - and update the mapping whenever we're notified
652 * about changes.
653 */
654 if (memory_region_has_ram_discard_manager(section->mr)) {
655 vfio_register_ram_discard_listener(bcontainer, section);
656 return;
657 }
658
659 vaddr = memory_region_get_ram_ptr(section->mr) +
660 section->offset_within_region +
661 (iova - section->offset_within_address_space);
662
663 trace_vfio_listener_region_add_ram(iova, end, vaddr);
664
665 llsize = int128_sub(llend, int128_make64(iova));
666
667 if (memory_region_is_ram_device(section->mr)) {
668 hwaddr pgmask = (1ULL << ctz64(bcontainer->pgsizes)) - 1;
669
670 if ((iova & pgmask) || (int128_get64(llsize) & pgmask)) {
671 trace_vfio_listener_region_add_no_dma_map(
672 memory_region_name(section->mr),
673 section->offset_within_address_space,
674 int128_getlo(section->size),
675 pgmask + 1);
676 return;
677 }
678 }
679
680 ret = vfio_container_dma_map(bcontainer, iova, int128_get64(llsize),
681 vaddr, section->readonly);
682 if (ret) {
683 error_setg(&err, "vfio_container_dma_map(%p, 0x%"HWADDR_PRIx", "
684 "0x%"HWADDR_PRIx", %p) = %d (%s)",
685 bcontainer, iova, int128_get64(llsize), vaddr, ret,
686 strerror(-ret));
687 mmio_dma_error:
688 if (memory_region_is_ram_device(section->mr)) {
689 /* Allow unexpected mappings not to be fatal for RAM devices */
690 VFIODevice *vbasedev =
691 vfio_get_vfio_device(memory_region_owner(section->mr));
692 vfio_device_error_append(vbasedev, &err);
693 warn_report_err_once(err);
694 return;
695 }
696 goto fail;
697 }
698
699 return;
700
701 fail:
702 if (!bcontainer->initialized) {
703 /*
704 * At machine init time or when the device is attached to the
705 * VM, store the first error in the container so we can
706 * gracefully fail the device realize routine.
707 */
708 if (!bcontainer->error) {
709 error_propagate_prepend(&bcontainer->error, err,
710 "Region %s: ",
711 memory_region_name(section->mr));
712 } else {
713 error_free(err);
714 }
715 } else {
716 /*
717 * At runtime, there's not much we can do other than throw a
718 * hardware error.
719 */
720 error_report_err(err);
721 hw_error("vfio: DMA mapping failed, unable to continue");
722 }
723 }
724
vfio_listener_region_del(MemoryListener * listener,MemoryRegionSection * section)725 static void vfio_listener_region_del(MemoryListener *listener,
726 MemoryRegionSection *section)
727 {
728 VFIOContainerBase *bcontainer = container_of(listener, VFIOContainerBase,
729 listener);
730 hwaddr iova, end;
731 Int128 llend, llsize;
732 int ret;
733 bool try_unmap = true;
734
735 if (!vfio_listener_valid_section(section, "region_del")) {
736 return;
737 }
738
739 if (memory_region_is_iommu(section->mr)) {
740 VFIOGuestIOMMU *giommu;
741
742 trace_vfio_listener_region_del_iommu(section->mr->name);
743 QLIST_FOREACH(giommu, &bcontainer->giommu_list, giommu_next) {
744 if (MEMORY_REGION(giommu->iommu_mr) == section->mr &&
745 giommu->n.start == section->offset_within_region) {
746 memory_region_unregister_iommu_notifier(section->mr,
747 &giommu->n);
748 QLIST_REMOVE(giommu, giommu_next);
749 g_free(giommu);
750 break;
751 }
752 }
753
754 /*
755 * FIXME: We assume the one big unmap below is adequate to
756 * remove any individual page mappings in the IOMMU which
757 * might have been copied into VFIO. This works for a page table
758 * based IOMMU where a big unmap flattens a large range of IO-PTEs.
759 * That may not be true for all IOMMU types.
760 */
761 }
762
763 if (!vfio_get_section_iova_range(bcontainer, section, &iova, &end,
764 &llend)) {
765 return;
766 }
767
768 llsize = int128_sub(llend, int128_make64(iova));
769
770 trace_vfio_listener_region_del(iova, end);
771
772 if (memory_region_is_ram_device(section->mr)) {
773 hwaddr pgmask;
774
775 pgmask = (1ULL << ctz64(bcontainer->pgsizes)) - 1;
776 try_unmap = !((iova & pgmask) || (int128_get64(llsize) & pgmask));
777 } else if (memory_region_has_ram_discard_manager(section->mr)) {
778 vfio_unregister_ram_discard_listener(bcontainer, section);
779 /* Unregistering will trigger an unmap. */
780 try_unmap = false;
781 }
782
783 if (try_unmap) {
784 if (int128_eq(llsize, int128_2_64())) {
785 /* The unmap ioctl doesn't accept a full 64-bit span. */
786 llsize = int128_rshift(llsize, 1);
787 ret = vfio_container_dma_unmap(bcontainer, iova,
788 int128_get64(llsize), NULL);
789 if (ret) {
790 error_report("vfio_container_dma_unmap(%p, 0x%"HWADDR_PRIx", "
791 "0x%"HWADDR_PRIx") = %d (%s)",
792 bcontainer, iova, int128_get64(llsize), ret,
793 strerror(-ret));
794 }
795 iova += int128_get64(llsize);
796 }
797 ret = vfio_container_dma_unmap(bcontainer, iova,
798 int128_get64(llsize), NULL);
799 if (ret) {
800 error_report("vfio_container_dma_unmap(%p, 0x%"HWADDR_PRIx", "
801 "0x%"HWADDR_PRIx") = %d (%s)",
802 bcontainer, iova, int128_get64(llsize), ret,
803 strerror(-ret));
804 }
805 }
806
807 memory_region_unref(section->mr);
808
809 /* PPC64/pseries machine only */
810 vfio_container_del_section_window(bcontainer, section);
811 }
812
813 typedef struct VFIODirtyRanges {
814 hwaddr min32;
815 hwaddr max32;
816 hwaddr min64;
817 hwaddr max64;
818 hwaddr minpci64;
819 hwaddr maxpci64;
820 } VFIODirtyRanges;
821
822 typedef struct VFIODirtyRangesListener {
823 VFIOContainerBase *bcontainer;
824 VFIODirtyRanges ranges;
825 MemoryListener listener;
826 } VFIODirtyRangesListener;
827
vfio_section_is_vfio_pci(MemoryRegionSection * section,VFIOContainerBase * bcontainer)828 static bool vfio_section_is_vfio_pci(MemoryRegionSection *section,
829 VFIOContainerBase *bcontainer)
830 {
831 VFIOPCIDevice *pcidev;
832 VFIODevice *vbasedev;
833 Object *owner;
834
835 owner = memory_region_owner(section->mr);
836
837 QLIST_FOREACH(vbasedev, &bcontainer->device_list, container_next) {
838 if (vbasedev->type != VFIO_DEVICE_TYPE_PCI) {
839 continue;
840 }
841 pcidev = container_of(vbasedev, VFIOPCIDevice, vbasedev);
842 if (OBJECT(pcidev) == owner) {
843 return true;
844 }
845 }
846
847 return false;
848 }
849
vfio_dirty_tracking_update_range(VFIODirtyRanges * range,hwaddr iova,hwaddr end,bool update_pci)850 static void vfio_dirty_tracking_update_range(VFIODirtyRanges *range,
851 hwaddr iova, hwaddr end,
852 bool update_pci)
853 {
854 hwaddr *min, *max;
855
856 /*
857 * The address space passed to the dirty tracker is reduced to three ranges:
858 * one for 32-bit DMA ranges, one for 64-bit DMA ranges and one for the
859 * PCI 64-bit hole.
860 *
861 * The underlying reports of dirty will query a sub-interval of each of
862 * these ranges.
863 *
864 * The purpose of the three range handling is to handle known cases of big
865 * holes in the address space, like the x86 AMD 1T hole, and firmware (like
866 * OVMF) which may relocate the pci-hole64 to the end of the address space.
867 * The latter would otherwise generate large ranges for tracking, stressing
868 * the limits of supported hardware. The pci-hole32 will always be below 4G
869 * (overlapping or not) so it doesn't need special handling and is part of
870 * the 32-bit range.
871 *
872 * The alternative would be an IOVATree but that has a much bigger runtime
873 * overhead and unnecessary complexity.
874 */
875 if (update_pci && iova >= UINT32_MAX) {
876 min = &range->minpci64;
877 max = &range->maxpci64;
878 } else {
879 min = (end <= UINT32_MAX) ? &range->min32 : &range->min64;
880 max = (end <= UINT32_MAX) ? &range->max32 : &range->max64;
881 }
882 if (*min > iova) {
883 *min = iova;
884 }
885 if (*max < end) {
886 *max = end;
887 }
888
889 trace_vfio_device_dirty_tracking_update(iova, end, *min, *max);
890 }
891
vfio_dirty_tracking_update(MemoryListener * listener,MemoryRegionSection * section)892 static void vfio_dirty_tracking_update(MemoryListener *listener,
893 MemoryRegionSection *section)
894 {
895 VFIODirtyRangesListener *dirty =
896 container_of(listener, VFIODirtyRangesListener, listener);
897 hwaddr iova, end;
898
899 if (!vfio_listener_valid_section(section, "tracking_update") ||
900 !vfio_get_section_iova_range(dirty->bcontainer, section,
901 &iova, &end, NULL)) {
902 return;
903 }
904
905 vfio_dirty_tracking_update_range(&dirty->ranges, iova, end,
906 vfio_section_is_vfio_pci(section, dirty->bcontainer));
907 }
908
909 static const MemoryListener vfio_dirty_tracking_listener = {
910 .name = "vfio-tracking",
911 .region_add = vfio_dirty_tracking_update,
912 };
913
vfio_dirty_tracking_init(VFIOContainerBase * bcontainer,VFIODirtyRanges * ranges)914 static void vfio_dirty_tracking_init(VFIOContainerBase *bcontainer,
915 VFIODirtyRanges *ranges)
916 {
917 VFIODirtyRangesListener dirty;
918
919 memset(&dirty, 0, sizeof(dirty));
920 dirty.ranges.min32 = UINT32_MAX;
921 dirty.ranges.min64 = UINT64_MAX;
922 dirty.ranges.minpci64 = UINT64_MAX;
923 dirty.listener = vfio_dirty_tracking_listener;
924 dirty.bcontainer = bcontainer;
925
926 memory_listener_register(&dirty.listener,
927 bcontainer->space->as);
928
929 *ranges = dirty.ranges;
930
931 /*
932 * The memory listener is synchronous, and used to calculate the range
933 * to dirty tracking. Unregister it after we are done as we are not
934 * interested in any follow-up updates.
935 */
936 memory_listener_unregister(&dirty.listener);
937 }
938
vfio_devices_dma_logging_stop(VFIOContainerBase * bcontainer)939 static void vfio_devices_dma_logging_stop(VFIOContainerBase *bcontainer)
940 {
941 uint64_t buf[DIV_ROUND_UP(sizeof(struct vfio_device_feature),
942 sizeof(uint64_t))] = {};
943 struct vfio_device_feature *feature = (struct vfio_device_feature *)buf;
944 VFIODevice *vbasedev;
945
946 feature->argsz = sizeof(buf);
947 feature->flags = VFIO_DEVICE_FEATURE_SET |
948 VFIO_DEVICE_FEATURE_DMA_LOGGING_STOP;
949
950 QLIST_FOREACH(vbasedev, &bcontainer->device_list, container_next) {
951 if (!vbasedev->dirty_tracking) {
952 continue;
953 }
954
955 if (ioctl(vbasedev->fd, VFIO_DEVICE_FEATURE, feature)) {
956 warn_report("%s: Failed to stop DMA logging, err %d (%s)",
957 vbasedev->name, -errno, strerror(errno));
958 }
959 vbasedev->dirty_tracking = false;
960 }
961 }
962
963 static struct vfio_device_feature *
vfio_device_feature_dma_logging_start_create(VFIOContainerBase * bcontainer,VFIODirtyRanges * tracking)964 vfio_device_feature_dma_logging_start_create(VFIOContainerBase *bcontainer,
965 VFIODirtyRanges *tracking)
966 {
967 struct vfio_device_feature *feature;
968 size_t feature_size;
969 struct vfio_device_feature_dma_logging_control *control;
970 struct vfio_device_feature_dma_logging_range *ranges;
971
972 feature_size = sizeof(struct vfio_device_feature) +
973 sizeof(struct vfio_device_feature_dma_logging_control);
974 feature = g_try_malloc0(feature_size);
975 if (!feature) {
976 errno = ENOMEM;
977 return NULL;
978 }
979 feature->argsz = feature_size;
980 feature->flags = VFIO_DEVICE_FEATURE_SET |
981 VFIO_DEVICE_FEATURE_DMA_LOGGING_START;
982
983 control = (struct vfio_device_feature_dma_logging_control *)feature->data;
984 control->page_size = qemu_real_host_page_size();
985
986 /*
987 * DMA logging uAPI guarantees to support at least a number of ranges that
988 * fits into a single host kernel base page.
989 */
990 control->num_ranges = !!tracking->max32 + !!tracking->max64 +
991 !!tracking->maxpci64;
992 ranges = g_try_new0(struct vfio_device_feature_dma_logging_range,
993 control->num_ranges);
994 if (!ranges) {
995 g_free(feature);
996 errno = ENOMEM;
997
998 return NULL;
999 }
1000
1001 control->ranges = (uintptr_t)ranges;
1002 if (tracking->max32) {
1003 ranges->iova = tracking->min32;
1004 ranges->length = (tracking->max32 - tracking->min32) + 1;
1005 ranges++;
1006 }
1007 if (tracking->max64) {
1008 ranges->iova = tracking->min64;
1009 ranges->length = (tracking->max64 - tracking->min64) + 1;
1010 ranges++;
1011 }
1012 if (tracking->maxpci64) {
1013 ranges->iova = tracking->minpci64;
1014 ranges->length = (tracking->maxpci64 - tracking->minpci64) + 1;
1015 }
1016
1017 trace_vfio_device_dirty_tracking_start(control->num_ranges,
1018 tracking->min32, tracking->max32,
1019 tracking->min64, tracking->max64,
1020 tracking->minpci64, tracking->maxpci64);
1021
1022 return feature;
1023 }
1024
vfio_device_feature_dma_logging_start_destroy(struct vfio_device_feature * feature)1025 static void vfio_device_feature_dma_logging_start_destroy(
1026 struct vfio_device_feature *feature)
1027 {
1028 struct vfio_device_feature_dma_logging_control *control =
1029 (struct vfio_device_feature_dma_logging_control *)feature->data;
1030 struct vfio_device_feature_dma_logging_range *ranges =
1031 (struct vfio_device_feature_dma_logging_range *)(uintptr_t)control->ranges;
1032
1033 g_free(ranges);
1034 g_free(feature);
1035 }
1036
vfio_devices_dma_logging_start(VFIOContainerBase * bcontainer,Error ** errp)1037 static bool vfio_devices_dma_logging_start(VFIOContainerBase *bcontainer,
1038 Error **errp)
1039 {
1040 struct vfio_device_feature *feature;
1041 VFIODirtyRanges ranges;
1042 VFIODevice *vbasedev;
1043 int ret = 0;
1044
1045 vfio_dirty_tracking_init(bcontainer, &ranges);
1046 feature = vfio_device_feature_dma_logging_start_create(bcontainer,
1047 &ranges);
1048 if (!feature) {
1049 error_setg_errno(errp, errno, "Failed to prepare DMA logging");
1050 return false;
1051 }
1052
1053 QLIST_FOREACH(vbasedev, &bcontainer->device_list, container_next) {
1054 if (vbasedev->dirty_tracking) {
1055 continue;
1056 }
1057
1058 ret = ioctl(vbasedev->fd, VFIO_DEVICE_FEATURE, feature);
1059 if (ret) {
1060 ret = -errno;
1061 error_setg_errno(errp, errno, "%s: Failed to start DMA logging",
1062 vbasedev->name);
1063 goto out;
1064 }
1065 vbasedev->dirty_tracking = true;
1066 }
1067
1068 out:
1069 if (ret) {
1070 vfio_devices_dma_logging_stop(bcontainer);
1071 }
1072
1073 vfio_device_feature_dma_logging_start_destroy(feature);
1074
1075 return ret == 0;
1076 }
1077
vfio_listener_log_global_start(MemoryListener * listener,Error ** errp)1078 static bool vfio_listener_log_global_start(MemoryListener *listener,
1079 Error **errp)
1080 {
1081 ERRP_GUARD();
1082 VFIOContainerBase *bcontainer = container_of(listener, VFIOContainerBase,
1083 listener);
1084 bool ret;
1085
1086 if (vfio_devices_all_device_dirty_tracking(bcontainer)) {
1087 ret = vfio_devices_dma_logging_start(bcontainer, errp);
1088 } else {
1089 ret = vfio_container_set_dirty_page_tracking(bcontainer, true, errp) == 0;
1090 }
1091
1092 if (!ret) {
1093 error_prepend(errp, "vfio: Could not start dirty page tracking - ");
1094 }
1095 return ret;
1096 }
1097
vfio_listener_log_global_stop(MemoryListener * listener)1098 static void vfio_listener_log_global_stop(MemoryListener *listener)
1099 {
1100 VFIOContainerBase *bcontainer = container_of(listener, VFIOContainerBase,
1101 listener);
1102 Error *local_err = NULL;
1103 int ret = 0;
1104
1105 if (vfio_devices_all_device_dirty_tracking(bcontainer)) {
1106 vfio_devices_dma_logging_stop(bcontainer);
1107 } else {
1108 ret = vfio_container_set_dirty_page_tracking(bcontainer, false,
1109 &local_err);
1110 }
1111
1112 if (ret) {
1113 error_prepend(&local_err,
1114 "vfio: Could not stop dirty page tracking - ");
1115 error_report_err(local_err);
1116 vfio_set_migration_error(ret);
1117 }
1118 }
1119
vfio_device_dma_logging_report(VFIODevice * vbasedev,hwaddr iova,hwaddr size,void * bitmap)1120 static int vfio_device_dma_logging_report(VFIODevice *vbasedev, hwaddr iova,
1121 hwaddr size, void *bitmap)
1122 {
1123 uint64_t buf[DIV_ROUND_UP(sizeof(struct vfio_device_feature) +
1124 sizeof(struct vfio_device_feature_dma_logging_report),
1125 sizeof(uint64_t))] = {};
1126 struct vfio_device_feature *feature = (struct vfio_device_feature *)buf;
1127 struct vfio_device_feature_dma_logging_report *report =
1128 (struct vfio_device_feature_dma_logging_report *)feature->data;
1129
1130 report->iova = iova;
1131 report->length = size;
1132 report->page_size = qemu_real_host_page_size();
1133 report->bitmap = (uintptr_t)bitmap;
1134
1135 feature->argsz = sizeof(buf);
1136 feature->flags = VFIO_DEVICE_FEATURE_GET |
1137 VFIO_DEVICE_FEATURE_DMA_LOGGING_REPORT;
1138
1139 if (ioctl(vbasedev->fd, VFIO_DEVICE_FEATURE, feature)) {
1140 return -errno;
1141 }
1142
1143 return 0;
1144 }
1145
vfio_devices_query_dirty_bitmap(const VFIOContainerBase * bcontainer,VFIOBitmap * vbmap,hwaddr iova,hwaddr size,Error ** errp)1146 int vfio_devices_query_dirty_bitmap(const VFIOContainerBase *bcontainer,
1147 VFIOBitmap *vbmap, hwaddr iova, hwaddr size, Error **errp)
1148 {
1149 VFIODevice *vbasedev;
1150 int ret;
1151
1152 QLIST_FOREACH(vbasedev, &bcontainer->device_list, container_next) {
1153 ret = vfio_device_dma_logging_report(vbasedev, iova, size,
1154 vbmap->bitmap);
1155 if (ret) {
1156 error_setg_errno(errp, -ret,
1157 "%s: Failed to get DMA logging report, iova: "
1158 "0x%" HWADDR_PRIx ", size: 0x%" HWADDR_PRIx,
1159 vbasedev->name, iova, size);
1160
1161 return ret;
1162 }
1163 }
1164
1165 return 0;
1166 }
1167
vfio_get_dirty_bitmap(const VFIOContainerBase * bcontainer,uint64_t iova,uint64_t size,ram_addr_t ram_addr,Error ** errp)1168 int vfio_get_dirty_bitmap(const VFIOContainerBase *bcontainer, uint64_t iova,
1169 uint64_t size, ram_addr_t ram_addr, Error **errp)
1170 {
1171 bool all_device_dirty_tracking =
1172 vfio_devices_all_device_dirty_tracking(bcontainer);
1173 uint64_t dirty_pages;
1174 VFIOBitmap vbmap;
1175 int ret;
1176
1177 if (!bcontainer->dirty_pages_supported && !all_device_dirty_tracking) {
1178 cpu_physical_memory_set_dirty_range(ram_addr, size,
1179 tcg_enabled() ? DIRTY_CLIENTS_ALL :
1180 DIRTY_CLIENTS_NOCODE);
1181 return 0;
1182 }
1183
1184 ret = vfio_bitmap_alloc(&vbmap, size);
1185 if (ret) {
1186 error_setg_errno(errp, -ret,
1187 "Failed to allocate dirty tracking bitmap");
1188 return ret;
1189 }
1190
1191 if (all_device_dirty_tracking) {
1192 ret = vfio_devices_query_dirty_bitmap(bcontainer, &vbmap, iova, size,
1193 errp);
1194 } else {
1195 ret = vfio_container_query_dirty_bitmap(bcontainer, &vbmap, iova, size,
1196 errp);
1197 }
1198
1199 if (ret) {
1200 goto out;
1201 }
1202
1203 dirty_pages = cpu_physical_memory_set_dirty_lebitmap(vbmap.bitmap, ram_addr,
1204 vbmap.pages);
1205
1206 trace_vfio_get_dirty_bitmap(iova, size, vbmap.size, ram_addr, dirty_pages);
1207 out:
1208 g_free(vbmap.bitmap);
1209
1210 return ret;
1211 }
1212
1213 typedef struct {
1214 IOMMUNotifier n;
1215 VFIOGuestIOMMU *giommu;
1216 } vfio_giommu_dirty_notifier;
1217
vfio_iommu_map_dirty_notify(IOMMUNotifier * n,IOMMUTLBEntry * iotlb)1218 static void vfio_iommu_map_dirty_notify(IOMMUNotifier *n, IOMMUTLBEntry *iotlb)
1219 {
1220 vfio_giommu_dirty_notifier *gdn = container_of(n,
1221 vfio_giommu_dirty_notifier, n);
1222 VFIOGuestIOMMU *giommu = gdn->giommu;
1223 VFIOContainerBase *bcontainer = giommu->bcontainer;
1224 hwaddr iova = iotlb->iova + giommu->iommu_offset;
1225 ram_addr_t translated_addr;
1226 Error *local_err = NULL;
1227 int ret = -EINVAL;
1228
1229 trace_vfio_iommu_map_dirty_notify(iova, iova + iotlb->addr_mask);
1230
1231 if (iotlb->target_as != &address_space_memory) {
1232 error_report("Wrong target AS \"%s\", only system memory is allowed",
1233 iotlb->target_as->name ? iotlb->target_as->name : "none");
1234 goto out;
1235 }
1236
1237 rcu_read_lock();
1238 if (!vfio_get_xlat_addr(iotlb, NULL, &translated_addr, NULL, &local_err)) {
1239 error_report_err(local_err);
1240 goto out_unlock;
1241 }
1242
1243 ret = vfio_get_dirty_bitmap(bcontainer, iova, iotlb->addr_mask + 1,
1244 translated_addr, &local_err);
1245 if (ret) {
1246 error_prepend(&local_err,
1247 "vfio_iommu_map_dirty_notify(%p, 0x%"HWADDR_PRIx", "
1248 "0x%"HWADDR_PRIx") failed - ", bcontainer, iova,
1249 iotlb->addr_mask + 1);
1250 error_report_err(local_err);
1251 }
1252
1253 out_unlock:
1254 rcu_read_unlock();
1255
1256 out:
1257 if (ret) {
1258 vfio_set_migration_error(ret);
1259 }
1260 }
1261
vfio_ram_discard_get_dirty_bitmap(MemoryRegionSection * section,void * opaque)1262 static int vfio_ram_discard_get_dirty_bitmap(MemoryRegionSection *section,
1263 void *opaque)
1264 {
1265 const hwaddr size = int128_get64(section->size);
1266 const hwaddr iova = section->offset_within_address_space;
1267 const ram_addr_t ram_addr = memory_region_get_ram_addr(section->mr) +
1268 section->offset_within_region;
1269 VFIORamDiscardListener *vrdl = opaque;
1270 Error *local_err = NULL;
1271 int ret;
1272
1273 /*
1274 * Sync the whole mapped region (spanning multiple individual mappings)
1275 * in one go.
1276 */
1277 ret = vfio_get_dirty_bitmap(vrdl->bcontainer, iova, size, ram_addr,
1278 &local_err);
1279 if (ret) {
1280 error_report_err(local_err);
1281 }
1282 return ret;
1283 }
1284
1285 static int
vfio_sync_ram_discard_listener_dirty_bitmap(VFIOContainerBase * bcontainer,MemoryRegionSection * section)1286 vfio_sync_ram_discard_listener_dirty_bitmap(VFIOContainerBase *bcontainer,
1287 MemoryRegionSection *section)
1288 {
1289 RamDiscardManager *rdm = memory_region_get_ram_discard_manager(section->mr);
1290 VFIORamDiscardListener *vrdl = NULL;
1291
1292 QLIST_FOREACH(vrdl, &bcontainer->vrdl_list, next) {
1293 if (vrdl->mr == section->mr &&
1294 vrdl->offset_within_address_space ==
1295 section->offset_within_address_space) {
1296 break;
1297 }
1298 }
1299
1300 if (!vrdl) {
1301 hw_error("vfio: Trying to sync missing RAM discard listener");
1302 }
1303
1304 /*
1305 * We only want/can synchronize the bitmap for actually mapped parts -
1306 * which correspond to populated parts. Replay all populated parts.
1307 */
1308 return ram_discard_manager_replay_populated(rdm, section,
1309 vfio_ram_discard_get_dirty_bitmap,
1310 &vrdl);
1311 }
1312
vfio_sync_iommu_dirty_bitmap(VFIOContainerBase * bcontainer,MemoryRegionSection * section)1313 static int vfio_sync_iommu_dirty_bitmap(VFIOContainerBase *bcontainer,
1314 MemoryRegionSection *section)
1315 {
1316 VFIOGuestIOMMU *giommu;
1317 bool found = false;
1318 Int128 llend;
1319 vfio_giommu_dirty_notifier gdn;
1320 int idx;
1321
1322 QLIST_FOREACH(giommu, &bcontainer->giommu_list, giommu_next) {
1323 if (MEMORY_REGION(giommu->iommu_mr) == section->mr &&
1324 giommu->n.start == section->offset_within_region) {
1325 found = true;
1326 break;
1327 }
1328 }
1329
1330 if (!found) {
1331 return 0;
1332 }
1333
1334 gdn.giommu = giommu;
1335 idx = memory_region_iommu_attrs_to_index(giommu->iommu_mr,
1336 MEMTXATTRS_UNSPECIFIED);
1337
1338 llend = int128_add(int128_make64(section->offset_within_region),
1339 section->size);
1340 llend = int128_sub(llend, int128_one());
1341
1342 iommu_notifier_init(&gdn.n, vfio_iommu_map_dirty_notify, IOMMU_NOTIFIER_MAP,
1343 section->offset_within_region, int128_get64(llend),
1344 idx);
1345 memory_region_iommu_replay(giommu->iommu_mr, &gdn.n);
1346
1347 return 0;
1348 }
1349
vfio_sync_dirty_bitmap(VFIOContainerBase * bcontainer,MemoryRegionSection * section,Error ** errp)1350 static int vfio_sync_dirty_bitmap(VFIOContainerBase *bcontainer,
1351 MemoryRegionSection *section, Error **errp)
1352 {
1353 ram_addr_t ram_addr;
1354
1355 if (memory_region_is_iommu(section->mr)) {
1356 return vfio_sync_iommu_dirty_bitmap(bcontainer, section);
1357 } else if (memory_region_has_ram_discard_manager(section->mr)) {
1358 int ret;
1359
1360 ret = vfio_sync_ram_discard_listener_dirty_bitmap(bcontainer, section);
1361 if (ret) {
1362 error_setg(errp,
1363 "Failed to sync dirty bitmap with RAM discard listener");
1364 }
1365 return ret;
1366 }
1367
1368 ram_addr = memory_region_get_ram_addr(section->mr) +
1369 section->offset_within_region;
1370
1371 return vfio_get_dirty_bitmap(bcontainer,
1372 REAL_HOST_PAGE_ALIGN(section->offset_within_address_space),
1373 int128_get64(section->size), ram_addr, errp);
1374 }
1375
vfio_listener_log_sync(MemoryListener * listener,MemoryRegionSection * section)1376 static void vfio_listener_log_sync(MemoryListener *listener,
1377 MemoryRegionSection *section)
1378 {
1379 VFIOContainerBase *bcontainer = container_of(listener, VFIOContainerBase,
1380 listener);
1381 int ret;
1382 Error *local_err = NULL;
1383
1384 if (vfio_listener_skipped_section(section)) {
1385 return;
1386 }
1387
1388 if (vfio_log_sync_needed(bcontainer)) {
1389 ret = vfio_sync_dirty_bitmap(bcontainer, section, &local_err);
1390 if (ret) {
1391 error_report_err(local_err);
1392 vfio_set_migration_error(ret);
1393 }
1394 }
1395 }
1396
1397 const MemoryListener vfio_memory_listener = {
1398 .name = "vfio",
1399 .region_add = vfio_listener_region_add,
1400 .region_del = vfio_listener_region_del,
1401 .log_global_start = vfio_listener_log_global_start,
1402 .log_global_stop = vfio_listener_log_global_stop,
1403 .log_sync = vfio_listener_log_sync,
1404 };
1405
vfio_reset_handler(void * opaque)1406 void vfio_reset_handler(void *opaque)
1407 {
1408 VFIODevice *vbasedev;
1409
1410 trace_vfio_reset_handler();
1411 QLIST_FOREACH(vbasedev, &vfio_device_list, global_next) {
1412 if (vbasedev->dev->realized) {
1413 vbasedev->ops->vfio_compute_needs_reset(vbasedev);
1414 }
1415 }
1416
1417 QLIST_FOREACH(vbasedev, &vfio_device_list, global_next) {
1418 if (vbasedev->dev->realized && vbasedev->needs_reset) {
1419 vbasedev->ops->vfio_hot_reset_multi(vbasedev);
1420 }
1421 }
1422 }
1423
vfio_kvm_device_add_fd(int fd,Error ** errp)1424 int vfio_kvm_device_add_fd(int fd, Error **errp)
1425 {
1426 #ifdef CONFIG_KVM
1427 struct kvm_device_attr attr = {
1428 .group = KVM_DEV_VFIO_FILE,
1429 .attr = KVM_DEV_VFIO_FILE_ADD,
1430 .addr = (uint64_t)(unsigned long)&fd,
1431 };
1432
1433 if (!kvm_enabled()) {
1434 return 0;
1435 }
1436
1437 if (vfio_kvm_device_fd < 0) {
1438 struct kvm_create_device cd = {
1439 .type = KVM_DEV_TYPE_VFIO,
1440 };
1441
1442 if (kvm_vm_ioctl(kvm_state, KVM_CREATE_DEVICE, &cd)) {
1443 error_setg_errno(errp, errno, "Failed to create KVM VFIO device");
1444 return -errno;
1445 }
1446
1447 vfio_kvm_device_fd = cd.fd;
1448 }
1449
1450 if (ioctl(vfio_kvm_device_fd, KVM_SET_DEVICE_ATTR, &attr)) {
1451 error_setg_errno(errp, errno, "Failed to add fd %d to KVM VFIO device",
1452 fd);
1453 return -errno;
1454 }
1455 #endif
1456 return 0;
1457 }
1458
vfio_kvm_device_del_fd(int fd,Error ** errp)1459 int vfio_kvm_device_del_fd(int fd, Error **errp)
1460 {
1461 #ifdef CONFIG_KVM
1462 struct kvm_device_attr attr = {
1463 .group = KVM_DEV_VFIO_FILE,
1464 .attr = KVM_DEV_VFIO_FILE_DEL,
1465 .addr = (uint64_t)(unsigned long)&fd,
1466 };
1467
1468 if (vfio_kvm_device_fd < 0) {
1469 error_setg(errp, "KVM VFIO device isn't created yet");
1470 return -EINVAL;
1471 }
1472
1473 if (ioctl(vfio_kvm_device_fd, KVM_SET_DEVICE_ATTR, &attr)) {
1474 error_setg_errno(errp, errno,
1475 "Failed to remove fd %d from KVM VFIO device", fd);
1476 return -errno;
1477 }
1478 #endif
1479 return 0;
1480 }
1481
vfio_get_address_space(AddressSpace * as)1482 VFIOAddressSpace *vfio_get_address_space(AddressSpace *as)
1483 {
1484 VFIOAddressSpace *space;
1485
1486 QLIST_FOREACH(space, &vfio_address_spaces, list) {
1487 if (space->as == as) {
1488 return space;
1489 }
1490 }
1491
1492 /* No suitable VFIOAddressSpace, create a new one */
1493 space = g_malloc0(sizeof(*space));
1494 space->as = as;
1495 QLIST_INIT(&space->containers);
1496
1497 if (QLIST_EMPTY(&vfio_address_spaces)) {
1498 qemu_register_reset(vfio_reset_handler, NULL);
1499 }
1500
1501 QLIST_INSERT_HEAD(&vfio_address_spaces, space, list);
1502
1503 return space;
1504 }
1505
vfio_put_address_space(VFIOAddressSpace * space)1506 void vfio_put_address_space(VFIOAddressSpace *space)
1507 {
1508 if (!QLIST_EMPTY(&space->containers)) {
1509 return;
1510 }
1511
1512 QLIST_REMOVE(space, list);
1513 g_free(space);
1514
1515 if (QLIST_EMPTY(&vfio_address_spaces)) {
1516 qemu_unregister_reset(vfio_reset_handler, NULL);
1517 }
1518 }
1519
vfio_address_space_insert(VFIOAddressSpace * space,VFIOContainerBase * bcontainer)1520 void vfio_address_space_insert(VFIOAddressSpace *space,
1521 VFIOContainerBase *bcontainer)
1522 {
1523 QLIST_INSERT_HEAD(&space->containers, bcontainer, next);
1524 bcontainer->space = space;
1525 }
1526
vfio_get_device_info(int fd)1527 struct vfio_device_info *vfio_get_device_info(int fd)
1528 {
1529 struct vfio_device_info *info;
1530 uint32_t argsz = sizeof(*info);
1531
1532 info = g_malloc0(argsz);
1533
1534 retry:
1535 info->argsz = argsz;
1536
1537 if (ioctl(fd, VFIO_DEVICE_GET_INFO, info)) {
1538 g_free(info);
1539 return NULL;
1540 }
1541
1542 if (info->argsz > argsz) {
1543 argsz = info->argsz;
1544 info = g_realloc(info, argsz);
1545 goto retry;
1546 }
1547
1548 return info;
1549 }
1550
vfio_attach_device(char * name,VFIODevice * vbasedev,AddressSpace * as,Error ** errp)1551 bool vfio_attach_device(char *name, VFIODevice *vbasedev,
1552 AddressSpace *as, Error **errp)
1553 {
1554 const VFIOIOMMUClass *ops =
1555 VFIO_IOMMU_CLASS(object_class_by_name(TYPE_VFIO_IOMMU_LEGACY));
1556 HostIOMMUDevice *hiod = NULL;
1557
1558 if (vbasedev->iommufd) {
1559 ops = VFIO_IOMMU_CLASS(object_class_by_name(TYPE_VFIO_IOMMU_IOMMUFD));
1560 }
1561
1562 assert(ops);
1563
1564
1565 if (!vbasedev->mdev) {
1566 hiod = HOST_IOMMU_DEVICE(object_new(ops->hiod_typename));
1567 vbasedev->hiod = hiod;
1568 }
1569
1570 if (!ops->attach_device(name, vbasedev, as, errp)) {
1571 object_unref(hiod);
1572 vbasedev->hiod = NULL;
1573 return false;
1574 }
1575
1576 return true;
1577 }
1578
vfio_detach_device(VFIODevice * vbasedev)1579 void vfio_detach_device(VFIODevice *vbasedev)
1580 {
1581 if (!vbasedev->bcontainer) {
1582 return;
1583 }
1584 object_unref(vbasedev->hiod);
1585 VFIO_IOMMU_GET_CLASS(vbasedev->bcontainer)->detach_device(vbasedev);
1586 }
1587