xref: /openbmc/qemu/hw/vfio/common.c (revision 7c08eefc)
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 "hw/hw.h"
34 #include "qemu/error-report.h"
35 #include "qemu/main-loop.h"
36 #include "qemu/range.h"
37 #include "sysemu/kvm.h"
38 #include "sysemu/reset.h"
39 #include "sysemu/runstate.h"
40 #include "trace.h"
41 #include "qapi/error.h"
42 #include "migration/misc.h"
43 #include "migration/blocker.h"
44 #include "migration/qemu-file.h"
45 #include "sysemu/tpm.h"
46 
47 VFIODeviceList vfio_device_list =
48     QLIST_HEAD_INITIALIZER(vfio_device_list);
49 static QLIST_HEAD(, VFIOAddressSpace) vfio_address_spaces =
50     QLIST_HEAD_INITIALIZER(vfio_address_spaces);
51 
52 #ifdef CONFIG_KVM
53 /*
54  * We have a single VFIO pseudo device per KVM VM.  Once created it lives
55  * for the life of the VM.  Closing the file descriptor only drops our
56  * reference to it and the device's reference to kvm.  Therefore once
57  * initialized, this file descriptor is only released on QEMU exit and
58  * we'll re-use it should another vfio device be attached before then.
59  */
60 int vfio_kvm_device_fd = -1;
61 #endif
62 
63 /*
64  * Device state interfaces
65  */
66 
67 bool vfio_mig_active(void)
68 {
69     VFIODevice *vbasedev;
70 
71     if (QLIST_EMPTY(&vfio_device_list)) {
72         return false;
73     }
74 
75     QLIST_FOREACH(vbasedev, &vfio_device_list, global_next) {
76         if (vbasedev->migration_blocker) {
77             return false;
78         }
79     }
80     return true;
81 }
82 
83 static Error *multiple_devices_migration_blocker;
84 
85 /*
86  * Multiple devices migration is allowed only if all devices support P2P
87  * migration. Single device migration is allowed regardless of P2P migration
88  * support.
89  */
90 static bool vfio_multiple_devices_migration_is_supported(void)
91 {
92     VFIODevice *vbasedev;
93     unsigned int device_num = 0;
94     bool all_support_p2p = true;
95 
96     QLIST_FOREACH(vbasedev, &vfio_device_list, global_next) {
97         if (vbasedev->migration) {
98             device_num++;
99 
100             if (!(vbasedev->migration->mig_flags & VFIO_MIGRATION_P2P)) {
101                 all_support_p2p = false;
102             }
103         }
104     }
105 
106     return all_support_p2p || device_num <= 1;
107 }
108 
109 int vfio_block_multiple_devices_migration(VFIODevice *vbasedev, Error **errp)
110 {
111     int ret;
112 
113     if (vfio_multiple_devices_migration_is_supported()) {
114         return 0;
115     }
116 
117     if (vbasedev->enable_migration == ON_OFF_AUTO_ON) {
118         error_setg(errp, "Multiple VFIO devices migration is supported only if "
119                          "all of them support P2P migration");
120         return -EINVAL;
121     }
122 
123     if (multiple_devices_migration_blocker) {
124         return 0;
125     }
126 
127     error_setg(&multiple_devices_migration_blocker,
128                "Multiple VFIO devices migration is supported only if all of "
129                "them support P2P migration");
130     ret = migrate_add_blocker_normal(&multiple_devices_migration_blocker, errp);
131 
132     return ret;
133 }
134 
135 void vfio_unblock_multiple_devices_migration(void)
136 {
137     if (!multiple_devices_migration_blocker ||
138         !vfio_multiple_devices_migration_is_supported()) {
139         return;
140     }
141 
142     migrate_del_blocker(&multiple_devices_migration_blocker);
143 }
144 
145 bool vfio_viommu_preset(VFIODevice *vbasedev)
146 {
147     return vbasedev->bcontainer->space->as != &address_space_memory;
148 }
149 
150 static void vfio_set_migration_error(int ret)
151 {
152     if (migration_is_setup_or_active()) {
153         migration_file_set_error(ret, NULL);
154     }
155 }
156 
157 bool vfio_device_state_is_running(VFIODevice *vbasedev)
158 {
159     VFIOMigration *migration = vbasedev->migration;
160 
161     return migration->device_state == VFIO_DEVICE_STATE_RUNNING ||
162            migration->device_state == VFIO_DEVICE_STATE_RUNNING_P2P;
163 }
164 
165 bool vfio_device_state_is_precopy(VFIODevice *vbasedev)
166 {
167     VFIOMigration *migration = vbasedev->migration;
168 
169     return migration->device_state == VFIO_DEVICE_STATE_PRE_COPY ||
170            migration->device_state == VFIO_DEVICE_STATE_PRE_COPY_P2P;
171 }
172 
173 static bool vfio_devices_all_dirty_tracking(VFIOContainerBase *bcontainer)
174 {
175     VFIODevice *vbasedev;
176 
177     if (!migration_is_active() && !migration_is_device()) {
178         return false;
179     }
180 
181     QLIST_FOREACH(vbasedev, &bcontainer->device_list, container_next) {
182         VFIOMigration *migration = vbasedev->migration;
183 
184         if (!migration) {
185             return false;
186         }
187 
188         if (vbasedev->pre_copy_dirty_page_tracking == ON_OFF_AUTO_OFF &&
189             (vfio_device_state_is_running(vbasedev) ||
190              vfio_device_state_is_precopy(vbasedev))) {
191             return false;
192         }
193     }
194     return true;
195 }
196 
197 bool vfio_devices_all_device_dirty_tracking(const VFIOContainerBase *bcontainer)
198 {
199     VFIODevice *vbasedev;
200 
201     QLIST_FOREACH(vbasedev, &bcontainer->device_list, container_next) {
202         if (!vbasedev->dirty_pages_supported) {
203             return false;
204         }
205     }
206 
207     return true;
208 }
209 
210 /*
211  * Check if all VFIO devices are running and migration is active, which is
212  * essentially equivalent to the migration being in pre-copy phase.
213  */
214 bool
215 vfio_devices_all_running_and_mig_active(const VFIOContainerBase *bcontainer)
216 {
217     VFIODevice *vbasedev;
218 
219     if (!migration_is_active()) {
220         return false;
221     }
222 
223     QLIST_FOREACH(vbasedev, &bcontainer->device_list, container_next) {
224         VFIOMigration *migration = vbasedev->migration;
225 
226         if (!migration) {
227             return false;
228         }
229 
230         if (vfio_device_state_is_running(vbasedev) ||
231             vfio_device_state_is_precopy(vbasedev)) {
232             continue;
233         } else {
234             return false;
235         }
236     }
237     return true;
238 }
239 
240 static bool vfio_listener_skipped_section(MemoryRegionSection *section)
241 {
242     return (!memory_region_is_ram(section->mr) &&
243             !memory_region_is_iommu(section->mr)) ||
244            memory_region_is_protected(section->mr) ||
245            /*
246             * Sizing an enabled 64-bit BAR can cause spurious mappings to
247             * addresses in the upper part of the 64-bit address space.  These
248             * are never accessed by the CPU and beyond the address width of
249             * some IOMMU hardware.  TODO: VFIO should tell us the IOMMU width.
250             */
251            section->offset_within_address_space & (1ULL << 63);
252 }
253 
254 /* Called with rcu_read_lock held.  */
255 static bool vfio_get_xlat_addr(IOMMUTLBEntry *iotlb, void **vaddr,
256                                ram_addr_t *ram_addr, bool *read_only,
257                                Error **errp)
258 {
259     bool ret, mr_has_discard_manager;
260 
261     ret = memory_get_xlat_addr(iotlb, vaddr, ram_addr, read_only,
262                                &mr_has_discard_manager, errp);
263     if (ret && mr_has_discard_manager) {
264         /*
265          * Malicious VMs might trigger discarding of IOMMU-mapped memory. The
266          * pages will remain pinned inside vfio until unmapped, resulting in a
267          * higher memory consumption than expected. If memory would get
268          * populated again later, there would be an inconsistency between pages
269          * pinned by vfio and pages seen by QEMU. This is the case until
270          * unmapped from the IOMMU (e.g., during device reset).
271          *
272          * With malicious guests, we really only care about pinning more memory
273          * than expected. RLIMIT_MEMLOCK set for the user/process can never be
274          * exceeded and can be used to mitigate this problem.
275          */
276         warn_report_once("Using vfio with vIOMMUs and coordinated discarding of"
277                          " RAM (e.g., virtio-mem) works, however, malicious"
278                          " guests can trigger pinning of more memory than"
279                          " intended via an IOMMU. It's possible to mitigate "
280                          " by setting/adjusting RLIMIT_MEMLOCK.");
281     }
282     return ret;
283 }
284 
285 static void vfio_iommu_map_notify(IOMMUNotifier *n, IOMMUTLBEntry *iotlb)
286 {
287     VFIOGuestIOMMU *giommu = container_of(n, VFIOGuestIOMMU, n);
288     VFIOContainerBase *bcontainer = giommu->bcontainer;
289     hwaddr iova = iotlb->iova + giommu->iommu_offset;
290     void *vaddr;
291     int ret;
292     Error *local_err = NULL;
293 
294     trace_vfio_iommu_map_notify(iotlb->perm == IOMMU_NONE ? "UNMAP" : "MAP",
295                                 iova, iova + iotlb->addr_mask);
296 
297     if (iotlb->target_as != &address_space_memory) {
298         error_report("Wrong target AS \"%s\", only system memory is allowed",
299                      iotlb->target_as->name ? iotlb->target_as->name : "none");
300         vfio_set_migration_error(-EINVAL);
301         return;
302     }
303 
304     rcu_read_lock();
305 
306     if ((iotlb->perm & IOMMU_RW) != IOMMU_NONE) {
307         bool read_only;
308 
309         if (!vfio_get_xlat_addr(iotlb, &vaddr, NULL, &read_only, &local_err)) {
310             error_report_err(local_err);
311             goto out;
312         }
313         /*
314          * vaddr is only valid until rcu_read_unlock(). But after
315          * vfio_dma_map has set up the mapping the pages will be
316          * pinned by the kernel. This makes sure that the RAM backend
317          * of vaddr will always be there, even if the memory object is
318          * destroyed and its backing memory munmap-ed.
319          */
320         ret = vfio_container_dma_map(bcontainer, iova,
321                                      iotlb->addr_mask + 1, vaddr,
322                                      read_only);
323         if (ret) {
324             error_report("vfio_container_dma_map(%p, 0x%"HWADDR_PRIx", "
325                          "0x%"HWADDR_PRIx", %p) = %d (%s)",
326                          bcontainer, iova,
327                          iotlb->addr_mask + 1, vaddr, ret, strerror(-ret));
328         }
329     } else {
330         ret = vfio_container_dma_unmap(bcontainer, iova,
331                                        iotlb->addr_mask + 1, iotlb);
332         if (ret) {
333             error_report("vfio_container_dma_unmap(%p, 0x%"HWADDR_PRIx", "
334                          "0x%"HWADDR_PRIx") = %d (%s)",
335                          bcontainer, iova,
336                          iotlb->addr_mask + 1, ret, strerror(-ret));
337             vfio_set_migration_error(ret);
338         }
339     }
340 out:
341     rcu_read_unlock();
342 }
343 
344 static void vfio_ram_discard_notify_discard(RamDiscardListener *rdl,
345                                             MemoryRegionSection *section)
346 {
347     VFIORamDiscardListener *vrdl = container_of(rdl, VFIORamDiscardListener,
348                                                 listener);
349     VFIOContainerBase *bcontainer = vrdl->bcontainer;
350     const hwaddr size = int128_get64(section->size);
351     const hwaddr iova = section->offset_within_address_space;
352     int ret;
353 
354     /* Unmap with a single call. */
355     ret = vfio_container_dma_unmap(bcontainer, iova, size , NULL);
356     if (ret) {
357         error_report("%s: vfio_container_dma_unmap() failed: %s", __func__,
358                      strerror(-ret));
359     }
360 }
361 
362 static int vfio_ram_discard_notify_populate(RamDiscardListener *rdl,
363                                             MemoryRegionSection *section)
364 {
365     VFIORamDiscardListener *vrdl = container_of(rdl, VFIORamDiscardListener,
366                                                 listener);
367     VFIOContainerBase *bcontainer = vrdl->bcontainer;
368     const hwaddr end = section->offset_within_region +
369                        int128_get64(section->size);
370     hwaddr start, next, iova;
371     void *vaddr;
372     int ret;
373 
374     /*
375      * Map in (aligned within memory region) minimum granularity, so we can
376      * unmap in minimum granularity later.
377      */
378     for (start = section->offset_within_region; start < end; start = next) {
379         next = ROUND_UP(start + 1, vrdl->granularity);
380         next = MIN(next, end);
381 
382         iova = start - section->offset_within_region +
383                section->offset_within_address_space;
384         vaddr = memory_region_get_ram_ptr(section->mr) + start;
385 
386         ret = vfio_container_dma_map(bcontainer, iova, next - start,
387                                      vaddr, section->readonly);
388         if (ret) {
389             /* Rollback */
390             vfio_ram_discard_notify_discard(rdl, section);
391             return ret;
392         }
393     }
394     return 0;
395 }
396 
397 static void vfio_register_ram_discard_listener(VFIOContainerBase *bcontainer,
398                                                MemoryRegionSection *section)
399 {
400     RamDiscardManager *rdm = memory_region_get_ram_discard_manager(section->mr);
401     VFIORamDiscardListener *vrdl;
402 
403     /* Ignore some corner cases not relevant in practice. */
404     g_assert(QEMU_IS_ALIGNED(section->offset_within_region, TARGET_PAGE_SIZE));
405     g_assert(QEMU_IS_ALIGNED(section->offset_within_address_space,
406                              TARGET_PAGE_SIZE));
407     g_assert(QEMU_IS_ALIGNED(int128_get64(section->size), TARGET_PAGE_SIZE));
408 
409     vrdl = g_new0(VFIORamDiscardListener, 1);
410     vrdl->bcontainer = bcontainer;
411     vrdl->mr = section->mr;
412     vrdl->offset_within_address_space = section->offset_within_address_space;
413     vrdl->size = int128_get64(section->size);
414     vrdl->granularity = ram_discard_manager_get_min_granularity(rdm,
415                                                                 section->mr);
416 
417     g_assert(vrdl->granularity && is_power_of_2(vrdl->granularity));
418     g_assert(bcontainer->pgsizes &&
419              vrdl->granularity >= 1ULL << ctz64(bcontainer->pgsizes));
420 
421     ram_discard_listener_init(&vrdl->listener,
422                               vfio_ram_discard_notify_populate,
423                               vfio_ram_discard_notify_discard, true);
424     ram_discard_manager_register_listener(rdm, &vrdl->listener, section);
425     QLIST_INSERT_HEAD(&bcontainer->vrdl_list, vrdl, next);
426 
427     /*
428      * Sanity-check if we have a theoretically problematic setup where we could
429      * exceed the maximum number of possible DMA mappings over time. We assume
430      * that each mapped section in the same address space as a RamDiscardManager
431      * section consumes exactly one DMA mapping, with the exception of
432      * RamDiscardManager sections; i.e., we don't expect to have gIOMMU sections
433      * in the same address space as RamDiscardManager sections.
434      *
435      * We assume that each section in the address space consumes one memslot.
436      * We take the number of KVM memory slots as a best guess for the maximum
437      * number of sections in the address space we could have over time,
438      * also consuming DMA mappings.
439      */
440     if (bcontainer->dma_max_mappings) {
441         unsigned int vrdl_count = 0, vrdl_mappings = 0, max_memslots = 512;
442 
443 #ifdef CONFIG_KVM
444         if (kvm_enabled()) {
445             max_memslots = kvm_get_max_memslots();
446         }
447 #endif
448 
449         QLIST_FOREACH(vrdl, &bcontainer->vrdl_list, next) {
450             hwaddr start, end;
451 
452             start = QEMU_ALIGN_DOWN(vrdl->offset_within_address_space,
453                                     vrdl->granularity);
454             end = ROUND_UP(vrdl->offset_within_address_space + vrdl->size,
455                            vrdl->granularity);
456             vrdl_mappings += (end - start) / vrdl->granularity;
457             vrdl_count++;
458         }
459 
460         if (vrdl_mappings + max_memslots - vrdl_count >
461             bcontainer->dma_max_mappings) {
462             warn_report("%s: possibly running out of DMA mappings. E.g., try"
463                         " increasing the 'block-size' of virtio-mem devies."
464                         " Maximum possible DMA mappings: %d, Maximum possible"
465                         " memslots: %d", __func__, bcontainer->dma_max_mappings,
466                         max_memslots);
467         }
468     }
469 }
470 
471 static void vfio_unregister_ram_discard_listener(VFIOContainerBase *bcontainer,
472                                                  MemoryRegionSection *section)
473 {
474     RamDiscardManager *rdm = memory_region_get_ram_discard_manager(section->mr);
475     VFIORamDiscardListener *vrdl = NULL;
476 
477     QLIST_FOREACH(vrdl, &bcontainer->vrdl_list, next) {
478         if (vrdl->mr == section->mr &&
479             vrdl->offset_within_address_space ==
480             section->offset_within_address_space) {
481             break;
482         }
483     }
484 
485     if (!vrdl) {
486         hw_error("vfio: Trying to unregister missing RAM discard listener");
487     }
488 
489     ram_discard_manager_unregister_listener(rdm, &vrdl->listener);
490     QLIST_REMOVE(vrdl, next);
491     g_free(vrdl);
492 }
493 
494 static bool vfio_known_safe_misalignment(MemoryRegionSection *section)
495 {
496     MemoryRegion *mr = section->mr;
497 
498     if (!TPM_IS_CRB(mr->owner)) {
499         return false;
500     }
501 
502     /* this is a known safe misaligned region, just trace for debug purpose */
503     trace_vfio_known_safe_misalignment(memory_region_name(mr),
504                                        section->offset_within_address_space,
505                                        section->offset_within_region,
506                                        qemu_real_host_page_size());
507     return true;
508 }
509 
510 static bool vfio_listener_valid_section(MemoryRegionSection *section,
511                                         const char *name)
512 {
513     if (vfio_listener_skipped_section(section)) {
514         trace_vfio_listener_region_skip(name,
515                 section->offset_within_address_space,
516                 section->offset_within_address_space +
517                 int128_get64(int128_sub(section->size, int128_one())));
518         return false;
519     }
520 
521     if (unlikely((section->offset_within_address_space &
522                   ~qemu_real_host_page_mask()) !=
523                  (section->offset_within_region & ~qemu_real_host_page_mask()))) {
524         if (!vfio_known_safe_misalignment(section)) {
525             error_report("%s received unaligned region %s iova=0x%"PRIx64
526                          " offset_within_region=0x%"PRIx64
527                          " qemu_real_host_page_size=0x%"PRIxPTR,
528                          __func__, memory_region_name(section->mr),
529                          section->offset_within_address_space,
530                          section->offset_within_region,
531                          qemu_real_host_page_size());
532         }
533         return false;
534     }
535 
536     return true;
537 }
538 
539 static bool vfio_get_section_iova_range(VFIOContainerBase *bcontainer,
540                                         MemoryRegionSection *section,
541                                         hwaddr *out_iova, hwaddr *out_end,
542                                         Int128 *out_llend)
543 {
544     Int128 llend;
545     hwaddr iova;
546 
547     iova = REAL_HOST_PAGE_ALIGN(section->offset_within_address_space);
548     llend = int128_make64(section->offset_within_address_space);
549     llend = int128_add(llend, section->size);
550     llend = int128_and(llend, int128_exts64(qemu_real_host_page_mask()));
551 
552     if (int128_ge(int128_make64(iova), llend)) {
553         return false;
554     }
555 
556     *out_iova = iova;
557     *out_end = int128_get64(int128_sub(llend, int128_one()));
558     if (out_llend) {
559         *out_llend = llend;
560     }
561     return true;
562 }
563 
564 static void vfio_listener_region_add(MemoryListener *listener,
565                                      MemoryRegionSection *section)
566 {
567     VFIOContainerBase *bcontainer = container_of(listener, VFIOContainerBase,
568                                                  listener);
569     hwaddr iova, end;
570     Int128 llend, llsize;
571     void *vaddr;
572     int ret;
573     Error *err = NULL;
574 
575     if (!vfio_listener_valid_section(section, "region_add")) {
576         return;
577     }
578 
579     if (!vfio_get_section_iova_range(bcontainer, section, &iova, &end,
580                                      &llend)) {
581         if (memory_region_is_ram_device(section->mr)) {
582             trace_vfio_listener_region_add_no_dma_map(
583                 memory_region_name(section->mr),
584                 section->offset_within_address_space,
585                 int128_getlo(section->size),
586                 qemu_real_host_page_size());
587         }
588         return;
589     }
590 
591     if (!vfio_container_add_section_window(bcontainer, section, &err)) {
592         goto fail;
593     }
594 
595     memory_region_ref(section->mr);
596 
597     if (memory_region_is_iommu(section->mr)) {
598         VFIOGuestIOMMU *giommu;
599         IOMMUMemoryRegion *iommu_mr = IOMMU_MEMORY_REGION(section->mr);
600         int iommu_idx;
601 
602         trace_vfio_listener_region_add_iommu(iova, end);
603         /*
604          * FIXME: For VFIO iommu types which have KVM acceleration to
605          * avoid bouncing all map/unmaps through qemu this way, this
606          * would be the right place to wire that up (tell the KVM
607          * device emulation the VFIO iommu handles to use).
608          */
609         giommu = g_malloc0(sizeof(*giommu));
610         giommu->iommu_mr = iommu_mr;
611         giommu->iommu_offset = section->offset_within_address_space -
612                                section->offset_within_region;
613         giommu->bcontainer = bcontainer;
614         llend = int128_add(int128_make64(section->offset_within_region),
615                            section->size);
616         llend = int128_sub(llend, int128_one());
617         iommu_idx = memory_region_iommu_attrs_to_index(iommu_mr,
618                                                        MEMTXATTRS_UNSPECIFIED);
619         iommu_notifier_init(&giommu->n, vfio_iommu_map_notify,
620                             IOMMU_NOTIFIER_IOTLB_EVENTS,
621                             section->offset_within_region,
622                             int128_get64(llend),
623                             iommu_idx);
624 
625         ret = memory_region_iommu_set_page_size_mask(giommu->iommu_mr,
626                                                      bcontainer->pgsizes,
627                                                      &err);
628         if (ret) {
629             g_free(giommu);
630             goto fail;
631         }
632 
633         ret = memory_region_register_iommu_notifier(section->mr, &giommu->n,
634                                                     &err);
635         if (ret) {
636             g_free(giommu);
637             goto fail;
638         }
639         QLIST_INSERT_HEAD(&bcontainer->giommu_list, giommu, giommu_next);
640         memory_region_iommu_replay(giommu->iommu_mr, &giommu->n);
641 
642         return;
643     }
644 
645     /* Here we assume that memory_region_is_ram(section->mr)==true */
646 
647     /*
648      * For RAM memory regions with a RamDiscardManager, we only want to map the
649      * actually populated parts - and update the mapping whenever we're notified
650      * about changes.
651      */
652     if (memory_region_has_ram_discard_manager(section->mr)) {
653         vfio_register_ram_discard_listener(bcontainer, section);
654         return;
655     }
656 
657     vaddr = memory_region_get_ram_ptr(section->mr) +
658             section->offset_within_region +
659             (iova - section->offset_within_address_space);
660 
661     trace_vfio_listener_region_add_ram(iova, end, vaddr);
662 
663     llsize = int128_sub(llend, int128_make64(iova));
664 
665     if (memory_region_is_ram_device(section->mr)) {
666         hwaddr pgmask = (1ULL << ctz64(bcontainer->pgsizes)) - 1;
667 
668         if ((iova & pgmask) || (int128_get64(llsize) & pgmask)) {
669             trace_vfio_listener_region_add_no_dma_map(
670                 memory_region_name(section->mr),
671                 section->offset_within_address_space,
672                 int128_getlo(section->size),
673                 pgmask + 1);
674             return;
675         }
676     }
677 
678     ret = vfio_container_dma_map(bcontainer, iova, int128_get64(llsize),
679                                  vaddr, section->readonly);
680     if (ret) {
681         error_setg(&err, "vfio_container_dma_map(%p, 0x%"HWADDR_PRIx", "
682                    "0x%"HWADDR_PRIx", %p) = %d (%s)",
683                    bcontainer, iova, int128_get64(llsize), vaddr, ret,
684                    strerror(-ret));
685         if (memory_region_is_ram_device(section->mr)) {
686             /* Allow unexpected mappings not to be fatal for RAM devices */
687             error_report_err(err);
688             return;
689         }
690         goto fail;
691     }
692 
693     return;
694 
695 fail:
696     if (memory_region_is_ram_device(section->mr)) {
697         error_reportf_err(err, "PCI p2p may not work: ");
698         return;
699     }
700     /*
701      * On the initfn path, store the first error in the container so we
702      * can gracefully fail.  Runtime, there's not much we can do other
703      * than throw a hardware error.
704      */
705     if (!bcontainer->initialized) {
706         if (!bcontainer->error) {
707             error_propagate_prepend(&bcontainer->error, err,
708                                     "Region %s: ",
709                                     memory_region_name(section->mr));
710         } else {
711             error_free(err);
712         }
713     } else {
714         error_report_err(err);
715         hw_error("vfio: DMA mapping failed, unable to continue");
716     }
717 }
718 
719 static void vfio_listener_region_del(MemoryListener *listener,
720                                      MemoryRegionSection *section)
721 {
722     VFIOContainerBase *bcontainer = container_of(listener, VFIOContainerBase,
723                                                  listener);
724     hwaddr iova, end;
725     Int128 llend, llsize;
726     int ret;
727     bool try_unmap = true;
728 
729     if (!vfio_listener_valid_section(section, "region_del")) {
730         return;
731     }
732 
733     if (memory_region_is_iommu(section->mr)) {
734         VFIOGuestIOMMU *giommu;
735 
736         QLIST_FOREACH(giommu, &bcontainer->giommu_list, giommu_next) {
737             if (MEMORY_REGION(giommu->iommu_mr) == section->mr &&
738                 giommu->n.start == section->offset_within_region) {
739                 memory_region_unregister_iommu_notifier(section->mr,
740                                                         &giommu->n);
741                 QLIST_REMOVE(giommu, giommu_next);
742                 g_free(giommu);
743                 break;
744             }
745         }
746 
747         /*
748          * FIXME: We assume the one big unmap below is adequate to
749          * remove any individual page mappings in the IOMMU which
750          * might have been copied into VFIO. This works for a page table
751          * based IOMMU where a big unmap flattens a large range of IO-PTEs.
752          * That may not be true for all IOMMU types.
753          */
754     }
755 
756     if (!vfio_get_section_iova_range(bcontainer, section, &iova, &end,
757                                      &llend)) {
758         return;
759     }
760 
761     llsize = int128_sub(llend, int128_make64(iova));
762 
763     trace_vfio_listener_region_del(iova, end);
764 
765     if (memory_region_is_ram_device(section->mr)) {
766         hwaddr pgmask;
767 
768         pgmask = (1ULL << ctz64(bcontainer->pgsizes)) - 1;
769         try_unmap = !((iova & pgmask) || (int128_get64(llsize) & pgmask));
770     } else if (memory_region_has_ram_discard_manager(section->mr)) {
771         vfio_unregister_ram_discard_listener(bcontainer, section);
772         /* Unregistering will trigger an unmap. */
773         try_unmap = false;
774     }
775 
776     if (try_unmap) {
777         if (int128_eq(llsize, int128_2_64())) {
778             /* The unmap ioctl doesn't accept a full 64-bit span. */
779             llsize = int128_rshift(llsize, 1);
780             ret = vfio_container_dma_unmap(bcontainer, iova,
781                                            int128_get64(llsize), NULL);
782             if (ret) {
783                 error_report("vfio_container_dma_unmap(%p, 0x%"HWADDR_PRIx", "
784                              "0x%"HWADDR_PRIx") = %d (%s)",
785                              bcontainer, iova, int128_get64(llsize), ret,
786                              strerror(-ret));
787             }
788             iova += int128_get64(llsize);
789         }
790         ret = vfio_container_dma_unmap(bcontainer, iova,
791                                        int128_get64(llsize), NULL);
792         if (ret) {
793             error_report("vfio_container_dma_unmap(%p, 0x%"HWADDR_PRIx", "
794                          "0x%"HWADDR_PRIx") = %d (%s)",
795                          bcontainer, iova, int128_get64(llsize), ret,
796                          strerror(-ret));
797         }
798     }
799 
800     memory_region_unref(section->mr);
801 
802     vfio_container_del_section_window(bcontainer, section);
803 }
804 
805 typedef struct VFIODirtyRanges {
806     hwaddr min32;
807     hwaddr max32;
808     hwaddr min64;
809     hwaddr max64;
810     hwaddr minpci64;
811     hwaddr maxpci64;
812 } VFIODirtyRanges;
813 
814 typedef struct VFIODirtyRangesListener {
815     VFIOContainerBase *bcontainer;
816     VFIODirtyRanges ranges;
817     MemoryListener listener;
818 } VFIODirtyRangesListener;
819 
820 static bool vfio_section_is_vfio_pci(MemoryRegionSection *section,
821                                      VFIOContainerBase *bcontainer)
822 {
823     VFIOPCIDevice *pcidev;
824     VFIODevice *vbasedev;
825     Object *owner;
826 
827     owner = memory_region_owner(section->mr);
828 
829     QLIST_FOREACH(vbasedev, &bcontainer->device_list, container_next) {
830         if (vbasedev->type != VFIO_DEVICE_TYPE_PCI) {
831             continue;
832         }
833         pcidev = container_of(vbasedev, VFIOPCIDevice, vbasedev);
834         if (OBJECT(pcidev) == owner) {
835             return true;
836         }
837     }
838 
839     return false;
840 }
841 
842 static void vfio_dirty_tracking_update_range(VFIODirtyRanges *range,
843                                              hwaddr iova, hwaddr end,
844                                              bool update_pci)
845 {
846     hwaddr *min, *max;
847 
848     /*
849      * The address space passed to the dirty tracker is reduced to three ranges:
850      * one for 32-bit DMA ranges, one for 64-bit DMA ranges and one for the
851      * PCI 64-bit hole.
852      *
853      * The underlying reports of dirty will query a sub-interval of each of
854      * these ranges.
855      *
856      * The purpose of the three range handling is to handle known cases of big
857      * holes in the address space, like the x86 AMD 1T hole, and firmware (like
858      * OVMF) which may relocate the pci-hole64 to the end of the address space.
859      * The latter would otherwise generate large ranges for tracking, stressing
860      * the limits of supported hardware. The pci-hole32 will always be below 4G
861      * (overlapping or not) so it doesn't need special handling and is part of
862      * the 32-bit range.
863      *
864      * The alternative would be an IOVATree but that has a much bigger runtime
865      * overhead and unnecessary complexity.
866      */
867     if (update_pci && iova >= UINT32_MAX) {
868         min = &range->minpci64;
869         max = &range->maxpci64;
870     } else {
871         min = (end <= UINT32_MAX) ? &range->min32 : &range->min64;
872         max = (end <= UINT32_MAX) ? &range->max32 : &range->max64;
873     }
874     if (*min > iova) {
875         *min = iova;
876     }
877     if (*max < end) {
878         *max = end;
879     }
880 
881     trace_vfio_device_dirty_tracking_update(iova, end, *min, *max);
882 }
883 
884 static void vfio_dirty_tracking_update(MemoryListener *listener,
885                                        MemoryRegionSection *section)
886 {
887     VFIODirtyRangesListener *dirty =
888         container_of(listener, VFIODirtyRangesListener, listener);
889     hwaddr iova, end;
890 
891     if (!vfio_listener_valid_section(section, "tracking_update") ||
892         !vfio_get_section_iova_range(dirty->bcontainer, section,
893                                      &iova, &end, NULL)) {
894         return;
895     }
896 
897     vfio_dirty_tracking_update_range(&dirty->ranges, iova, end,
898                       vfio_section_is_vfio_pci(section, dirty->bcontainer));
899 }
900 
901 static const MemoryListener vfio_dirty_tracking_listener = {
902     .name = "vfio-tracking",
903     .region_add = vfio_dirty_tracking_update,
904 };
905 
906 static void vfio_dirty_tracking_init(VFIOContainerBase *bcontainer,
907                                      VFIODirtyRanges *ranges)
908 {
909     VFIODirtyRangesListener dirty;
910 
911     memset(&dirty, 0, sizeof(dirty));
912     dirty.ranges.min32 = UINT32_MAX;
913     dirty.ranges.min64 = UINT64_MAX;
914     dirty.ranges.minpci64 = UINT64_MAX;
915     dirty.listener = vfio_dirty_tracking_listener;
916     dirty.bcontainer = bcontainer;
917 
918     memory_listener_register(&dirty.listener,
919                              bcontainer->space->as);
920 
921     *ranges = dirty.ranges;
922 
923     /*
924      * The memory listener is synchronous, and used to calculate the range
925      * to dirty tracking. Unregister it after we are done as we are not
926      * interested in any follow-up updates.
927      */
928     memory_listener_unregister(&dirty.listener);
929 }
930 
931 static void vfio_devices_dma_logging_stop(VFIOContainerBase *bcontainer)
932 {
933     uint64_t buf[DIV_ROUND_UP(sizeof(struct vfio_device_feature),
934                               sizeof(uint64_t))] = {};
935     struct vfio_device_feature *feature = (struct vfio_device_feature *)buf;
936     VFIODevice *vbasedev;
937 
938     feature->argsz = sizeof(buf);
939     feature->flags = VFIO_DEVICE_FEATURE_SET |
940                      VFIO_DEVICE_FEATURE_DMA_LOGGING_STOP;
941 
942     QLIST_FOREACH(vbasedev, &bcontainer->device_list, container_next) {
943         if (!vbasedev->dirty_tracking) {
944             continue;
945         }
946 
947         if (ioctl(vbasedev->fd, VFIO_DEVICE_FEATURE, feature)) {
948             warn_report("%s: Failed to stop DMA logging, err %d (%s)",
949                         vbasedev->name, -errno, strerror(errno));
950         }
951         vbasedev->dirty_tracking = false;
952     }
953 }
954 
955 static struct vfio_device_feature *
956 vfio_device_feature_dma_logging_start_create(VFIOContainerBase *bcontainer,
957                                              VFIODirtyRanges *tracking)
958 {
959     struct vfio_device_feature *feature;
960     size_t feature_size;
961     struct vfio_device_feature_dma_logging_control *control;
962     struct vfio_device_feature_dma_logging_range *ranges;
963 
964     feature_size = sizeof(struct vfio_device_feature) +
965                    sizeof(struct vfio_device_feature_dma_logging_control);
966     feature = g_try_malloc0(feature_size);
967     if (!feature) {
968         errno = ENOMEM;
969         return NULL;
970     }
971     feature->argsz = feature_size;
972     feature->flags = VFIO_DEVICE_FEATURE_SET |
973                      VFIO_DEVICE_FEATURE_DMA_LOGGING_START;
974 
975     control = (struct vfio_device_feature_dma_logging_control *)feature->data;
976     control->page_size = qemu_real_host_page_size();
977 
978     /*
979      * DMA logging uAPI guarantees to support at least a number of ranges that
980      * fits into a single host kernel base page.
981      */
982     control->num_ranges = !!tracking->max32 + !!tracking->max64 +
983         !!tracking->maxpci64;
984     ranges = g_try_new0(struct vfio_device_feature_dma_logging_range,
985                         control->num_ranges);
986     if (!ranges) {
987         g_free(feature);
988         errno = ENOMEM;
989 
990         return NULL;
991     }
992 
993     control->ranges = (uintptr_t)ranges;
994     if (tracking->max32) {
995         ranges->iova = tracking->min32;
996         ranges->length = (tracking->max32 - tracking->min32) + 1;
997         ranges++;
998     }
999     if (tracking->max64) {
1000         ranges->iova = tracking->min64;
1001         ranges->length = (tracking->max64 - tracking->min64) + 1;
1002         ranges++;
1003     }
1004     if (tracking->maxpci64) {
1005         ranges->iova = tracking->minpci64;
1006         ranges->length = (tracking->maxpci64 - tracking->minpci64) + 1;
1007     }
1008 
1009     trace_vfio_device_dirty_tracking_start(control->num_ranges,
1010                                            tracking->min32, tracking->max32,
1011                                            tracking->min64, tracking->max64,
1012                                            tracking->minpci64, tracking->maxpci64);
1013 
1014     return feature;
1015 }
1016 
1017 static void vfio_device_feature_dma_logging_start_destroy(
1018     struct vfio_device_feature *feature)
1019 {
1020     struct vfio_device_feature_dma_logging_control *control =
1021         (struct vfio_device_feature_dma_logging_control *)feature->data;
1022     struct vfio_device_feature_dma_logging_range *ranges =
1023         (struct vfio_device_feature_dma_logging_range *)(uintptr_t)control->ranges;
1024 
1025     g_free(ranges);
1026     g_free(feature);
1027 }
1028 
1029 static bool vfio_devices_dma_logging_start(VFIOContainerBase *bcontainer,
1030                                           Error **errp)
1031 {
1032     struct vfio_device_feature *feature;
1033     VFIODirtyRanges ranges;
1034     VFIODevice *vbasedev;
1035     int ret = 0;
1036 
1037     vfio_dirty_tracking_init(bcontainer, &ranges);
1038     feature = vfio_device_feature_dma_logging_start_create(bcontainer,
1039                                                            &ranges);
1040     if (!feature) {
1041         error_setg_errno(errp, errno, "Failed to prepare DMA logging");
1042         return false;
1043     }
1044 
1045     QLIST_FOREACH(vbasedev, &bcontainer->device_list, container_next) {
1046         if (vbasedev->dirty_tracking) {
1047             continue;
1048         }
1049 
1050         ret = ioctl(vbasedev->fd, VFIO_DEVICE_FEATURE, feature);
1051         if (ret) {
1052             ret = -errno;
1053             error_setg_errno(errp, errno, "%s: Failed to start DMA logging",
1054                              vbasedev->name);
1055             goto out;
1056         }
1057         vbasedev->dirty_tracking = true;
1058     }
1059 
1060 out:
1061     if (ret) {
1062         vfio_devices_dma_logging_stop(bcontainer);
1063     }
1064 
1065     vfio_device_feature_dma_logging_start_destroy(feature);
1066 
1067     return ret == 0;
1068 }
1069 
1070 static bool vfio_listener_log_global_start(MemoryListener *listener,
1071                                            Error **errp)
1072 {
1073     ERRP_GUARD();
1074     VFIOContainerBase *bcontainer = container_of(listener, VFIOContainerBase,
1075                                                  listener);
1076     bool ret;
1077 
1078     if (vfio_devices_all_device_dirty_tracking(bcontainer)) {
1079         ret = vfio_devices_dma_logging_start(bcontainer, errp);
1080     } else {
1081         ret = vfio_container_set_dirty_page_tracking(bcontainer, true, errp) == 0;
1082     }
1083 
1084     if (!ret) {
1085         error_prepend(errp, "vfio: Could not start dirty page tracking - ");
1086     }
1087     return ret;
1088 }
1089 
1090 static void vfio_listener_log_global_stop(MemoryListener *listener)
1091 {
1092     VFIOContainerBase *bcontainer = container_of(listener, VFIOContainerBase,
1093                                                  listener);
1094     Error *local_err = NULL;
1095     int ret = 0;
1096 
1097     if (vfio_devices_all_device_dirty_tracking(bcontainer)) {
1098         vfio_devices_dma_logging_stop(bcontainer);
1099     } else {
1100         ret = vfio_container_set_dirty_page_tracking(bcontainer, false,
1101                                                      &local_err);
1102     }
1103 
1104     if (ret) {
1105         error_prepend(&local_err,
1106                       "vfio: Could not stop dirty page tracking - ");
1107         error_report_err(local_err);
1108         vfio_set_migration_error(ret);
1109     }
1110 }
1111 
1112 static int vfio_device_dma_logging_report(VFIODevice *vbasedev, hwaddr iova,
1113                                           hwaddr size, void *bitmap)
1114 {
1115     uint64_t buf[DIV_ROUND_UP(sizeof(struct vfio_device_feature) +
1116                         sizeof(struct vfio_device_feature_dma_logging_report),
1117                         sizeof(uint64_t))] = {};
1118     struct vfio_device_feature *feature = (struct vfio_device_feature *)buf;
1119     struct vfio_device_feature_dma_logging_report *report =
1120         (struct vfio_device_feature_dma_logging_report *)feature->data;
1121 
1122     report->iova = iova;
1123     report->length = size;
1124     report->page_size = qemu_real_host_page_size();
1125     report->bitmap = (uintptr_t)bitmap;
1126 
1127     feature->argsz = sizeof(buf);
1128     feature->flags = VFIO_DEVICE_FEATURE_GET |
1129                      VFIO_DEVICE_FEATURE_DMA_LOGGING_REPORT;
1130 
1131     if (ioctl(vbasedev->fd, VFIO_DEVICE_FEATURE, feature)) {
1132         return -errno;
1133     }
1134 
1135     return 0;
1136 }
1137 
1138 int vfio_devices_query_dirty_bitmap(const VFIOContainerBase *bcontainer,
1139                  VFIOBitmap *vbmap, hwaddr iova, hwaddr size, Error **errp)
1140 {
1141     VFIODevice *vbasedev;
1142     int ret;
1143 
1144     QLIST_FOREACH(vbasedev, &bcontainer->device_list, container_next) {
1145         ret = vfio_device_dma_logging_report(vbasedev, iova, size,
1146                                              vbmap->bitmap);
1147         if (ret) {
1148             error_setg_errno(errp, -ret,
1149                              "%s: Failed to get DMA logging report, iova: "
1150                              "0x%" HWADDR_PRIx ", size: 0x%" HWADDR_PRIx,
1151                              vbasedev->name, iova, size);
1152 
1153             return ret;
1154         }
1155     }
1156 
1157     return 0;
1158 }
1159 
1160 int vfio_get_dirty_bitmap(const VFIOContainerBase *bcontainer, uint64_t iova,
1161                           uint64_t size, ram_addr_t ram_addr, Error **errp)
1162 {
1163     bool all_device_dirty_tracking =
1164         vfio_devices_all_device_dirty_tracking(bcontainer);
1165     uint64_t dirty_pages;
1166     VFIOBitmap vbmap;
1167     int ret;
1168 
1169     if (!bcontainer->dirty_pages_supported && !all_device_dirty_tracking) {
1170         cpu_physical_memory_set_dirty_range(ram_addr, size,
1171                                             tcg_enabled() ? DIRTY_CLIENTS_ALL :
1172                                             DIRTY_CLIENTS_NOCODE);
1173         return 0;
1174     }
1175 
1176     ret = vfio_bitmap_alloc(&vbmap, size);
1177     if (ret) {
1178         error_setg_errno(errp, -ret,
1179                          "Failed to allocate dirty tracking bitmap");
1180         return ret;
1181     }
1182 
1183     if (all_device_dirty_tracking) {
1184         ret = vfio_devices_query_dirty_bitmap(bcontainer, &vbmap, iova, size,
1185                                               errp);
1186     } else {
1187         ret = vfio_container_query_dirty_bitmap(bcontainer, &vbmap, iova, size,
1188                                                 errp);
1189     }
1190 
1191     if (ret) {
1192         goto out;
1193     }
1194 
1195     dirty_pages = cpu_physical_memory_set_dirty_lebitmap(vbmap.bitmap, ram_addr,
1196                                                          vbmap.pages);
1197 
1198     trace_vfio_get_dirty_bitmap(iova, size, vbmap.size, ram_addr, dirty_pages);
1199 out:
1200     g_free(vbmap.bitmap);
1201 
1202     return ret;
1203 }
1204 
1205 typedef struct {
1206     IOMMUNotifier n;
1207     VFIOGuestIOMMU *giommu;
1208 } vfio_giommu_dirty_notifier;
1209 
1210 static void vfio_iommu_map_dirty_notify(IOMMUNotifier *n, IOMMUTLBEntry *iotlb)
1211 {
1212     vfio_giommu_dirty_notifier *gdn = container_of(n,
1213                                                 vfio_giommu_dirty_notifier, n);
1214     VFIOGuestIOMMU *giommu = gdn->giommu;
1215     VFIOContainerBase *bcontainer = giommu->bcontainer;
1216     hwaddr iova = iotlb->iova + giommu->iommu_offset;
1217     ram_addr_t translated_addr;
1218     Error *local_err = NULL;
1219     int ret = -EINVAL;
1220 
1221     trace_vfio_iommu_map_dirty_notify(iova, iova + iotlb->addr_mask);
1222 
1223     if (iotlb->target_as != &address_space_memory) {
1224         error_report("Wrong target AS \"%s\", only system memory is allowed",
1225                      iotlb->target_as->name ? iotlb->target_as->name : "none");
1226         goto out;
1227     }
1228 
1229     rcu_read_lock();
1230     if (!vfio_get_xlat_addr(iotlb, NULL, &translated_addr, NULL, &local_err)) {
1231         error_report_err(local_err);
1232         goto out_unlock;
1233     }
1234 
1235     ret = vfio_get_dirty_bitmap(bcontainer, iova, iotlb->addr_mask + 1,
1236                                 translated_addr, &local_err);
1237     if (ret) {
1238         error_prepend(&local_err,
1239                       "vfio_iommu_map_dirty_notify(%p, 0x%"HWADDR_PRIx", "
1240                       "0x%"HWADDR_PRIx") failed - ", bcontainer, iova,
1241                       iotlb->addr_mask + 1);
1242         error_report_err(local_err);
1243     }
1244 
1245 out_unlock:
1246     rcu_read_unlock();
1247 
1248 out:
1249     if (ret) {
1250         vfio_set_migration_error(ret);
1251     }
1252 }
1253 
1254 static int vfio_ram_discard_get_dirty_bitmap(MemoryRegionSection *section,
1255                                              void *opaque)
1256 {
1257     const hwaddr size = int128_get64(section->size);
1258     const hwaddr iova = section->offset_within_address_space;
1259     const ram_addr_t ram_addr = memory_region_get_ram_addr(section->mr) +
1260                                 section->offset_within_region;
1261     VFIORamDiscardListener *vrdl = opaque;
1262     Error *local_err = NULL;
1263     int ret;
1264 
1265     /*
1266      * Sync the whole mapped region (spanning multiple individual mappings)
1267      * in one go.
1268      */
1269     ret = vfio_get_dirty_bitmap(vrdl->bcontainer, iova, size, ram_addr,
1270                                 &local_err);
1271     if (ret) {
1272         error_report_err(local_err);
1273     }
1274     return ret;
1275 }
1276 
1277 static int
1278 vfio_sync_ram_discard_listener_dirty_bitmap(VFIOContainerBase *bcontainer,
1279                                             MemoryRegionSection *section)
1280 {
1281     RamDiscardManager *rdm = memory_region_get_ram_discard_manager(section->mr);
1282     VFIORamDiscardListener *vrdl = NULL;
1283 
1284     QLIST_FOREACH(vrdl, &bcontainer->vrdl_list, next) {
1285         if (vrdl->mr == section->mr &&
1286             vrdl->offset_within_address_space ==
1287             section->offset_within_address_space) {
1288             break;
1289         }
1290     }
1291 
1292     if (!vrdl) {
1293         hw_error("vfio: Trying to sync missing RAM discard listener");
1294     }
1295 
1296     /*
1297      * We only want/can synchronize the bitmap for actually mapped parts -
1298      * which correspond to populated parts. Replay all populated parts.
1299      */
1300     return ram_discard_manager_replay_populated(rdm, section,
1301                                               vfio_ram_discard_get_dirty_bitmap,
1302                                                 &vrdl);
1303 }
1304 
1305 static int vfio_sync_iommu_dirty_bitmap(VFIOContainerBase *bcontainer,
1306                                         MemoryRegionSection *section)
1307 {
1308     VFIOGuestIOMMU *giommu;
1309     bool found = false;
1310     Int128 llend;
1311     vfio_giommu_dirty_notifier gdn;
1312     int idx;
1313 
1314     QLIST_FOREACH(giommu, &bcontainer->giommu_list, giommu_next) {
1315         if (MEMORY_REGION(giommu->iommu_mr) == section->mr &&
1316             giommu->n.start == section->offset_within_region) {
1317             found = true;
1318             break;
1319         }
1320     }
1321 
1322     if (!found) {
1323         return 0;
1324     }
1325 
1326     gdn.giommu = giommu;
1327     idx = memory_region_iommu_attrs_to_index(giommu->iommu_mr,
1328                                              MEMTXATTRS_UNSPECIFIED);
1329 
1330     llend = int128_add(int128_make64(section->offset_within_region),
1331                        section->size);
1332     llend = int128_sub(llend, int128_one());
1333 
1334     iommu_notifier_init(&gdn.n, vfio_iommu_map_dirty_notify, IOMMU_NOTIFIER_MAP,
1335                         section->offset_within_region, int128_get64(llend),
1336                         idx);
1337     memory_region_iommu_replay(giommu->iommu_mr, &gdn.n);
1338 
1339     return 0;
1340 }
1341 
1342 static int vfio_sync_dirty_bitmap(VFIOContainerBase *bcontainer,
1343                                   MemoryRegionSection *section, Error **errp)
1344 {
1345     ram_addr_t ram_addr;
1346 
1347     if (memory_region_is_iommu(section->mr)) {
1348         return vfio_sync_iommu_dirty_bitmap(bcontainer, section);
1349     } else if (memory_region_has_ram_discard_manager(section->mr)) {
1350         int ret;
1351 
1352         ret = vfio_sync_ram_discard_listener_dirty_bitmap(bcontainer, section);
1353         if (ret) {
1354             error_setg(errp,
1355                        "Failed to sync dirty bitmap with RAM discard listener");
1356         }
1357         return ret;
1358     }
1359 
1360     ram_addr = memory_region_get_ram_addr(section->mr) +
1361                section->offset_within_region;
1362 
1363     return vfio_get_dirty_bitmap(bcontainer,
1364                    REAL_HOST_PAGE_ALIGN(section->offset_within_address_space),
1365                                  int128_get64(section->size), ram_addr, errp);
1366 }
1367 
1368 static void vfio_listener_log_sync(MemoryListener *listener,
1369         MemoryRegionSection *section)
1370 {
1371     VFIOContainerBase *bcontainer = container_of(listener, VFIOContainerBase,
1372                                                  listener);
1373     int ret;
1374     Error *local_err = NULL;
1375 
1376     if (vfio_listener_skipped_section(section)) {
1377         return;
1378     }
1379 
1380     if (vfio_devices_all_dirty_tracking(bcontainer)) {
1381         ret = vfio_sync_dirty_bitmap(bcontainer, section, &local_err);
1382         if (ret) {
1383             error_report_err(local_err);
1384             vfio_set_migration_error(ret);
1385         }
1386     }
1387 }
1388 
1389 const MemoryListener vfio_memory_listener = {
1390     .name = "vfio",
1391     .region_add = vfio_listener_region_add,
1392     .region_del = vfio_listener_region_del,
1393     .log_global_start = vfio_listener_log_global_start,
1394     .log_global_stop = vfio_listener_log_global_stop,
1395     .log_sync = vfio_listener_log_sync,
1396 };
1397 
1398 void vfio_reset_handler(void *opaque)
1399 {
1400     VFIODevice *vbasedev;
1401 
1402     QLIST_FOREACH(vbasedev, &vfio_device_list, global_next) {
1403         if (vbasedev->dev->realized) {
1404             vbasedev->ops->vfio_compute_needs_reset(vbasedev);
1405         }
1406     }
1407 
1408     QLIST_FOREACH(vbasedev, &vfio_device_list, global_next) {
1409         if (vbasedev->dev->realized && vbasedev->needs_reset) {
1410             vbasedev->ops->vfio_hot_reset_multi(vbasedev);
1411         }
1412     }
1413 }
1414 
1415 int vfio_kvm_device_add_fd(int fd, Error **errp)
1416 {
1417 #ifdef CONFIG_KVM
1418     struct kvm_device_attr attr = {
1419         .group = KVM_DEV_VFIO_FILE,
1420         .attr = KVM_DEV_VFIO_FILE_ADD,
1421         .addr = (uint64_t)(unsigned long)&fd,
1422     };
1423 
1424     if (!kvm_enabled()) {
1425         return 0;
1426     }
1427 
1428     if (vfio_kvm_device_fd < 0) {
1429         struct kvm_create_device cd = {
1430             .type = KVM_DEV_TYPE_VFIO,
1431         };
1432 
1433         if (kvm_vm_ioctl(kvm_state, KVM_CREATE_DEVICE, &cd)) {
1434             error_setg_errno(errp, errno, "Failed to create KVM VFIO device");
1435             return -errno;
1436         }
1437 
1438         vfio_kvm_device_fd = cd.fd;
1439     }
1440 
1441     if (ioctl(vfio_kvm_device_fd, KVM_SET_DEVICE_ATTR, &attr)) {
1442         error_setg_errno(errp, errno, "Failed to add fd %d to KVM VFIO device",
1443                          fd);
1444         return -errno;
1445     }
1446 #endif
1447     return 0;
1448 }
1449 
1450 int vfio_kvm_device_del_fd(int fd, Error **errp)
1451 {
1452 #ifdef CONFIG_KVM
1453     struct kvm_device_attr attr = {
1454         .group = KVM_DEV_VFIO_FILE,
1455         .attr = KVM_DEV_VFIO_FILE_DEL,
1456         .addr = (uint64_t)(unsigned long)&fd,
1457     };
1458 
1459     if (vfio_kvm_device_fd < 0) {
1460         error_setg(errp, "KVM VFIO device isn't created yet");
1461         return -EINVAL;
1462     }
1463 
1464     if (ioctl(vfio_kvm_device_fd, KVM_SET_DEVICE_ATTR, &attr)) {
1465         error_setg_errno(errp, errno,
1466                          "Failed to remove fd %d from KVM VFIO device", fd);
1467         return -errno;
1468     }
1469 #endif
1470     return 0;
1471 }
1472 
1473 VFIOAddressSpace *vfio_get_address_space(AddressSpace *as)
1474 {
1475     VFIOAddressSpace *space;
1476 
1477     QLIST_FOREACH(space, &vfio_address_spaces, list) {
1478         if (space->as == as) {
1479             return space;
1480         }
1481     }
1482 
1483     /* No suitable VFIOAddressSpace, create a new one */
1484     space = g_malloc0(sizeof(*space));
1485     space->as = as;
1486     QLIST_INIT(&space->containers);
1487 
1488     if (QLIST_EMPTY(&vfio_address_spaces)) {
1489         qemu_register_reset(vfio_reset_handler, NULL);
1490     }
1491 
1492     QLIST_INSERT_HEAD(&vfio_address_spaces, space, list);
1493 
1494     return space;
1495 }
1496 
1497 void vfio_put_address_space(VFIOAddressSpace *space)
1498 {
1499     if (!QLIST_EMPTY(&space->containers)) {
1500         return;
1501     }
1502 
1503     QLIST_REMOVE(space, list);
1504     g_free(space);
1505 
1506     if (QLIST_EMPTY(&vfio_address_spaces)) {
1507         qemu_unregister_reset(vfio_reset_handler, NULL);
1508     }
1509 }
1510 
1511 void vfio_address_space_insert(VFIOAddressSpace *space,
1512                                VFIOContainerBase *bcontainer)
1513 {
1514     QLIST_INSERT_HEAD(&space->containers, bcontainer, next);
1515     bcontainer->space = space;
1516 }
1517 
1518 struct vfio_device_info *vfio_get_device_info(int fd)
1519 {
1520     struct vfio_device_info *info;
1521     uint32_t argsz = sizeof(*info);
1522 
1523     info = g_malloc0(argsz);
1524 
1525 retry:
1526     info->argsz = argsz;
1527 
1528     if (ioctl(fd, VFIO_DEVICE_GET_INFO, info)) {
1529         g_free(info);
1530         return NULL;
1531     }
1532 
1533     if (info->argsz > argsz) {
1534         argsz = info->argsz;
1535         info = g_realloc(info, argsz);
1536         goto retry;
1537     }
1538 
1539     return info;
1540 }
1541 
1542 bool vfio_attach_device(char *name, VFIODevice *vbasedev,
1543                         AddressSpace *as, Error **errp)
1544 {
1545     const VFIOIOMMUClass *ops =
1546         VFIO_IOMMU_CLASS(object_class_by_name(TYPE_VFIO_IOMMU_LEGACY));
1547     HostIOMMUDevice *hiod;
1548 
1549     if (vbasedev->iommufd) {
1550         ops = VFIO_IOMMU_CLASS(object_class_by_name(TYPE_VFIO_IOMMU_IOMMUFD));
1551     }
1552 
1553     assert(ops);
1554 
1555     if (!ops->attach_device(name, vbasedev, as, errp)) {
1556         return false;
1557     }
1558 
1559     hiod = HOST_IOMMU_DEVICE(object_new(ops->hiod_typename));
1560     if (!HOST_IOMMU_DEVICE_GET_CLASS(hiod)->realize(hiod, vbasedev, errp)) {
1561         object_unref(hiod);
1562         ops->detach_device(vbasedev);
1563         return false;
1564     }
1565     vbasedev->hiod = hiod;
1566 
1567     return true;
1568 }
1569 
1570 void vfio_detach_device(VFIODevice *vbasedev)
1571 {
1572     if (!vbasedev->bcontainer) {
1573         return;
1574     }
1575     object_unref(vbasedev->hiod);
1576     VFIO_IOMMU_GET_CLASS(vbasedev->bcontainer)->detach_device(vbasedev);
1577 }
1578