xref: /openbmc/qemu/hw/vfio/common.c (revision 87417811)
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/vfio.h"
30 #include "hw/vfio/pci.h"
31 #include "exec/address-spaces.h"
32 #include "exec/memory.h"
33 #include "exec/ram_addr.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 "sysemu/kvm.h"
39 #include "sysemu/reset.h"
40 #include "sysemu/runstate.h"
41 #include "trace.h"
42 #include "qapi/error.h"
43 #include "migration/migration.h"
44 #include "migration/misc.h"
45 #include "migration/blocker.h"
46 #include "migration/qemu-file.h"
47 #include "sysemu/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 
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, 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  */
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, 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 
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(multiple_devices_migration_blocker, errp);
133     if (ret < 0) {
134         error_free(multiple_devices_migration_blocker);
135         multiple_devices_migration_blocker = NULL;
136     }
137 
138     return ret;
139 }
140 
141 void vfio_unblock_multiple_devices_migration(void)
142 {
143     if (!multiple_devices_migration_blocker ||
144         !vfio_multiple_devices_migration_is_supported()) {
145         return;
146     }
147 
148     migrate_del_blocker(multiple_devices_migration_blocker);
149     error_free(multiple_devices_migration_blocker);
150     multiple_devices_migration_blocker = NULL;
151 }
152 
153 bool vfio_viommu_preset(VFIODevice *vbasedev)
154 {
155     return vbasedev->container->space->as != &address_space_memory;
156 }
157 
158 static void vfio_set_migration_error(int err)
159 {
160     MigrationState *ms = migrate_get_current();
161 
162     if (migration_is_setup_or_active(ms->state)) {
163         WITH_QEMU_LOCK_GUARD(&ms->qemu_file_lock) {
164             if (ms->to_dst_file) {
165                 qemu_file_set_error(ms->to_dst_file, err);
166             }
167         }
168     }
169 }
170 
171 bool vfio_device_state_is_running(VFIODevice *vbasedev)
172 {
173     VFIOMigration *migration = vbasedev->migration;
174 
175     return migration->device_state == VFIO_DEVICE_STATE_RUNNING ||
176            migration->device_state == VFIO_DEVICE_STATE_RUNNING_P2P;
177 }
178 
179 bool vfio_device_state_is_precopy(VFIODevice *vbasedev)
180 {
181     VFIOMigration *migration = vbasedev->migration;
182 
183     return migration->device_state == VFIO_DEVICE_STATE_PRE_COPY ||
184            migration->device_state == VFIO_DEVICE_STATE_PRE_COPY_P2P;
185 }
186 
187 static bool vfio_devices_all_dirty_tracking(VFIOContainer *container)
188 {
189     VFIODevice *vbasedev;
190     MigrationState *ms = migrate_get_current();
191 
192     if (ms->state != MIGRATION_STATUS_ACTIVE &&
193         ms->state != MIGRATION_STATUS_DEVICE) {
194         return false;
195     }
196 
197     QLIST_FOREACH(vbasedev, &container->device_list, container_next) {
198         VFIOMigration *migration = vbasedev->migration;
199 
200         if (!migration) {
201             return false;
202         }
203 
204         if (vbasedev->pre_copy_dirty_page_tracking == ON_OFF_AUTO_OFF &&
205             (vfio_device_state_is_running(vbasedev) ||
206              vfio_device_state_is_precopy(vbasedev))) {
207             return false;
208         }
209     }
210     return true;
211 }
212 
213 bool vfio_devices_all_device_dirty_tracking(VFIOContainer *container)
214 {
215     VFIODevice *vbasedev;
216 
217     QLIST_FOREACH(vbasedev, &container->device_list, container_next) {
218         if (!vbasedev->dirty_pages_supported) {
219             return false;
220         }
221     }
222 
223     return true;
224 }
225 
226 /*
227  * Check if all VFIO devices are running and migration is active, which is
228  * essentially equivalent to the migration being in pre-copy phase.
229  */
230 bool vfio_devices_all_running_and_mig_active(VFIOContainer *container)
231 {
232     VFIODevice *vbasedev;
233 
234     if (!migration_is_active(migrate_get_current())) {
235         return false;
236     }
237 
238     QLIST_FOREACH(vbasedev, &container->device_list, container_next) {
239         VFIOMigration *migration = vbasedev->migration;
240 
241         if (!migration) {
242             return false;
243         }
244 
245         if (vfio_device_state_is_running(vbasedev) ||
246             vfio_device_state_is_precopy(vbasedev)) {
247             continue;
248         } else {
249             return false;
250         }
251     }
252     return true;
253 }
254 
255 void vfio_host_win_add(VFIOContainer *container, hwaddr min_iova,
256                        hwaddr max_iova, uint64_t iova_pgsizes)
257 {
258     VFIOHostDMAWindow *hostwin;
259 
260     QLIST_FOREACH(hostwin, &container->hostwin_list, hostwin_next) {
261         if (ranges_overlap(hostwin->min_iova,
262                            hostwin->max_iova - hostwin->min_iova + 1,
263                            min_iova,
264                            max_iova - min_iova + 1)) {
265             hw_error("%s: Overlapped IOMMU are not enabled", __func__);
266         }
267     }
268 
269     hostwin = g_malloc0(sizeof(*hostwin));
270 
271     hostwin->min_iova = min_iova;
272     hostwin->max_iova = max_iova;
273     hostwin->iova_pgsizes = iova_pgsizes;
274     QLIST_INSERT_HEAD(&container->hostwin_list, hostwin, hostwin_next);
275 }
276 
277 int vfio_host_win_del(VFIOContainer *container,
278                       hwaddr min_iova, hwaddr max_iova)
279 {
280     VFIOHostDMAWindow *hostwin;
281 
282     QLIST_FOREACH(hostwin, &container->hostwin_list, hostwin_next) {
283         if (hostwin->min_iova == min_iova && hostwin->max_iova == max_iova) {
284             QLIST_REMOVE(hostwin, hostwin_next);
285             g_free(hostwin);
286             return 0;
287         }
288     }
289 
290     return -1;
291 }
292 
293 static bool vfio_listener_skipped_section(MemoryRegionSection *section)
294 {
295     return (!memory_region_is_ram(section->mr) &&
296             !memory_region_is_iommu(section->mr)) ||
297            memory_region_is_protected(section->mr) ||
298            /*
299             * Sizing an enabled 64-bit BAR can cause spurious mappings to
300             * addresses in the upper part of the 64-bit address space.  These
301             * are never accessed by the CPU and beyond the address width of
302             * some IOMMU hardware.  TODO: VFIO should tell us the IOMMU width.
303             */
304            section->offset_within_address_space & (1ULL << 63);
305 }
306 
307 /* Called with rcu_read_lock held.  */
308 static bool vfio_get_xlat_addr(IOMMUTLBEntry *iotlb, void **vaddr,
309                                ram_addr_t *ram_addr, bool *read_only)
310 {
311     bool ret, mr_has_discard_manager;
312 
313     ret = memory_get_xlat_addr(iotlb, vaddr, ram_addr, read_only,
314                                &mr_has_discard_manager);
315     if (ret && mr_has_discard_manager) {
316         /*
317          * Malicious VMs might trigger discarding of IOMMU-mapped memory. The
318          * pages will remain pinned inside vfio until unmapped, resulting in a
319          * higher memory consumption than expected. If memory would get
320          * populated again later, there would be an inconsistency between pages
321          * pinned by vfio and pages seen by QEMU. This is the case until
322          * unmapped from the IOMMU (e.g., during device reset).
323          *
324          * With malicious guests, we really only care about pinning more memory
325          * than expected. RLIMIT_MEMLOCK set for the user/process can never be
326          * exceeded and can be used to mitigate this problem.
327          */
328         warn_report_once("Using vfio with vIOMMUs and coordinated discarding of"
329                          " RAM (e.g., virtio-mem) works, however, malicious"
330                          " guests can trigger pinning of more memory than"
331                          " intended via an IOMMU. It's possible to mitigate "
332                          " by setting/adjusting RLIMIT_MEMLOCK.");
333     }
334     return ret;
335 }
336 
337 static void vfio_iommu_map_notify(IOMMUNotifier *n, IOMMUTLBEntry *iotlb)
338 {
339     VFIOGuestIOMMU *giommu = container_of(n, VFIOGuestIOMMU, n);
340     VFIOContainer *container = giommu->container;
341     hwaddr iova = iotlb->iova + giommu->iommu_offset;
342     void *vaddr;
343     int ret;
344 
345     trace_vfio_iommu_map_notify(iotlb->perm == IOMMU_NONE ? "UNMAP" : "MAP",
346                                 iova, iova + iotlb->addr_mask);
347 
348     if (iotlb->target_as != &address_space_memory) {
349         error_report("Wrong target AS \"%s\", only system memory is allowed",
350                      iotlb->target_as->name ? iotlb->target_as->name : "none");
351         vfio_set_migration_error(-EINVAL);
352         return;
353     }
354 
355     rcu_read_lock();
356 
357     if ((iotlb->perm & IOMMU_RW) != IOMMU_NONE) {
358         bool read_only;
359 
360         if (!vfio_get_xlat_addr(iotlb, &vaddr, NULL, &read_only)) {
361             goto out;
362         }
363         /*
364          * vaddr is only valid until rcu_read_unlock(). But after
365          * vfio_dma_map has set up the mapping the pages will be
366          * pinned by the kernel. This makes sure that the RAM backend
367          * of vaddr will always be there, even if the memory object is
368          * destroyed and its backing memory munmap-ed.
369          */
370         ret = vfio_dma_map(container, iova,
371                            iotlb->addr_mask + 1, vaddr,
372                            read_only);
373         if (ret) {
374             error_report("vfio_dma_map(%p, 0x%"HWADDR_PRIx", "
375                          "0x%"HWADDR_PRIx", %p) = %d (%s)",
376                          container, iova,
377                          iotlb->addr_mask + 1, vaddr, ret, strerror(-ret));
378         }
379     } else {
380         ret = vfio_dma_unmap(container, iova, iotlb->addr_mask + 1, iotlb);
381         if (ret) {
382             error_report("vfio_dma_unmap(%p, 0x%"HWADDR_PRIx", "
383                          "0x%"HWADDR_PRIx") = %d (%s)",
384                          container, iova,
385                          iotlb->addr_mask + 1, ret, strerror(-ret));
386             vfio_set_migration_error(ret);
387         }
388     }
389 out:
390     rcu_read_unlock();
391 }
392 
393 static void vfio_ram_discard_notify_discard(RamDiscardListener *rdl,
394                                             MemoryRegionSection *section)
395 {
396     VFIORamDiscardListener *vrdl = container_of(rdl, VFIORamDiscardListener,
397                                                 listener);
398     const hwaddr size = int128_get64(section->size);
399     const hwaddr iova = section->offset_within_address_space;
400     int ret;
401 
402     /* Unmap with a single call. */
403     ret = vfio_dma_unmap(vrdl->container, iova, size , NULL);
404     if (ret) {
405         error_report("%s: vfio_dma_unmap() failed: %s", __func__,
406                      strerror(-ret));
407     }
408 }
409 
410 static int vfio_ram_discard_notify_populate(RamDiscardListener *rdl,
411                                             MemoryRegionSection *section)
412 {
413     VFIORamDiscardListener *vrdl = container_of(rdl, VFIORamDiscardListener,
414                                                 listener);
415     const hwaddr end = section->offset_within_region +
416                        int128_get64(section->size);
417     hwaddr start, next, iova;
418     void *vaddr;
419     int ret;
420 
421     /*
422      * Map in (aligned within memory region) minimum granularity, so we can
423      * unmap in minimum granularity later.
424      */
425     for (start = section->offset_within_region; start < end; start = next) {
426         next = ROUND_UP(start + 1, vrdl->granularity);
427         next = MIN(next, end);
428 
429         iova = start - section->offset_within_region +
430                section->offset_within_address_space;
431         vaddr = memory_region_get_ram_ptr(section->mr) + start;
432 
433         ret = vfio_dma_map(vrdl->container, iova, next - start,
434                            vaddr, section->readonly);
435         if (ret) {
436             /* Rollback */
437             vfio_ram_discard_notify_discard(rdl, section);
438             return ret;
439         }
440     }
441     return 0;
442 }
443 
444 static void vfio_register_ram_discard_listener(VFIOContainer *container,
445                                                MemoryRegionSection *section)
446 {
447     RamDiscardManager *rdm = memory_region_get_ram_discard_manager(section->mr);
448     VFIORamDiscardListener *vrdl;
449 
450     /* Ignore some corner cases not relevant in practice. */
451     g_assert(QEMU_IS_ALIGNED(section->offset_within_region, TARGET_PAGE_SIZE));
452     g_assert(QEMU_IS_ALIGNED(section->offset_within_address_space,
453                              TARGET_PAGE_SIZE));
454     g_assert(QEMU_IS_ALIGNED(int128_get64(section->size), TARGET_PAGE_SIZE));
455 
456     vrdl = g_new0(VFIORamDiscardListener, 1);
457     vrdl->container = container;
458     vrdl->mr = section->mr;
459     vrdl->offset_within_address_space = section->offset_within_address_space;
460     vrdl->size = int128_get64(section->size);
461     vrdl->granularity = ram_discard_manager_get_min_granularity(rdm,
462                                                                 section->mr);
463 
464     g_assert(vrdl->granularity && is_power_of_2(vrdl->granularity));
465     g_assert(container->pgsizes &&
466              vrdl->granularity >= 1ULL << ctz64(container->pgsizes));
467 
468     ram_discard_listener_init(&vrdl->listener,
469                               vfio_ram_discard_notify_populate,
470                               vfio_ram_discard_notify_discard, true);
471     ram_discard_manager_register_listener(rdm, &vrdl->listener, section);
472     QLIST_INSERT_HEAD(&container->vrdl_list, vrdl, next);
473 
474     /*
475      * Sanity-check if we have a theoretically problematic setup where we could
476      * exceed the maximum number of possible DMA mappings over time. We assume
477      * that each mapped section in the same address space as a RamDiscardManager
478      * section consumes exactly one DMA mapping, with the exception of
479      * RamDiscardManager sections; i.e., we don't expect to have gIOMMU sections
480      * in the same address space as RamDiscardManager sections.
481      *
482      * We assume that each section in the address space consumes one memslot.
483      * We take the number of KVM memory slots as a best guess for the maximum
484      * number of sections in the address space we could have over time,
485      * also consuming DMA mappings.
486      */
487     if (container->dma_max_mappings) {
488         unsigned int vrdl_count = 0, vrdl_mappings = 0, max_memslots = 512;
489 
490 #ifdef CONFIG_KVM
491         if (kvm_enabled()) {
492             max_memslots = kvm_get_max_memslots();
493         }
494 #endif
495 
496         QLIST_FOREACH(vrdl, &container->vrdl_list, next) {
497             hwaddr start, end;
498 
499             start = QEMU_ALIGN_DOWN(vrdl->offset_within_address_space,
500                                     vrdl->granularity);
501             end = ROUND_UP(vrdl->offset_within_address_space + vrdl->size,
502                            vrdl->granularity);
503             vrdl_mappings += (end - start) / vrdl->granularity;
504             vrdl_count++;
505         }
506 
507         if (vrdl_mappings + max_memslots - vrdl_count >
508             container->dma_max_mappings) {
509             warn_report("%s: possibly running out of DMA mappings. E.g., try"
510                         " increasing the 'block-size' of virtio-mem devies."
511                         " Maximum possible DMA mappings: %d, Maximum possible"
512                         " memslots: %d", __func__, container->dma_max_mappings,
513                         max_memslots);
514         }
515     }
516 }
517 
518 static void vfio_unregister_ram_discard_listener(VFIOContainer *container,
519                                                  MemoryRegionSection *section)
520 {
521     RamDiscardManager *rdm = memory_region_get_ram_discard_manager(section->mr);
522     VFIORamDiscardListener *vrdl = NULL;
523 
524     QLIST_FOREACH(vrdl, &container->vrdl_list, next) {
525         if (vrdl->mr == section->mr &&
526             vrdl->offset_within_address_space ==
527             section->offset_within_address_space) {
528             break;
529         }
530     }
531 
532     if (!vrdl) {
533         hw_error("vfio: Trying to unregister missing RAM discard listener");
534     }
535 
536     ram_discard_manager_unregister_listener(rdm, &vrdl->listener);
537     QLIST_REMOVE(vrdl, next);
538     g_free(vrdl);
539 }
540 
541 static VFIOHostDMAWindow *vfio_find_hostwin(VFIOContainer *container,
542                                             hwaddr iova, hwaddr end)
543 {
544     VFIOHostDMAWindow *hostwin;
545     bool hostwin_found = false;
546 
547     QLIST_FOREACH(hostwin, &container->hostwin_list, hostwin_next) {
548         if (hostwin->min_iova <= iova && end <= hostwin->max_iova) {
549             hostwin_found = true;
550             break;
551         }
552     }
553 
554     return hostwin_found ? hostwin : NULL;
555 }
556 
557 static bool vfio_known_safe_misalignment(MemoryRegionSection *section)
558 {
559     MemoryRegion *mr = section->mr;
560 
561     if (!TPM_IS_CRB(mr->owner)) {
562         return false;
563     }
564 
565     /* this is a known safe misaligned region, just trace for debug purpose */
566     trace_vfio_known_safe_misalignment(memory_region_name(mr),
567                                        section->offset_within_address_space,
568                                        section->offset_within_region,
569                                        qemu_real_host_page_size());
570     return true;
571 }
572 
573 static bool vfio_listener_valid_section(MemoryRegionSection *section,
574                                         const char *name)
575 {
576     if (vfio_listener_skipped_section(section)) {
577         trace_vfio_listener_region_skip(name,
578                 section->offset_within_address_space,
579                 section->offset_within_address_space +
580                 int128_get64(int128_sub(section->size, int128_one())));
581         return false;
582     }
583 
584     if (unlikely((section->offset_within_address_space &
585                   ~qemu_real_host_page_mask()) !=
586                  (section->offset_within_region & ~qemu_real_host_page_mask()))) {
587         if (!vfio_known_safe_misalignment(section)) {
588             error_report("%s received unaligned region %s iova=0x%"PRIx64
589                          " offset_within_region=0x%"PRIx64
590                          " qemu_real_host_page_size=0x%"PRIxPTR,
591                          __func__, memory_region_name(section->mr),
592                          section->offset_within_address_space,
593                          section->offset_within_region,
594                          qemu_real_host_page_size());
595         }
596         return false;
597     }
598 
599     return true;
600 }
601 
602 static bool vfio_get_section_iova_range(VFIOContainer *container,
603                                         MemoryRegionSection *section,
604                                         hwaddr *out_iova, hwaddr *out_end,
605                                         Int128 *out_llend)
606 {
607     Int128 llend;
608     hwaddr iova;
609 
610     iova = REAL_HOST_PAGE_ALIGN(section->offset_within_address_space);
611     llend = int128_make64(section->offset_within_address_space);
612     llend = int128_add(llend, section->size);
613     llend = int128_and(llend, int128_exts64(qemu_real_host_page_mask()));
614 
615     if (int128_ge(int128_make64(iova), llend)) {
616         return false;
617     }
618 
619     *out_iova = iova;
620     *out_end = int128_get64(int128_sub(llend, int128_one()));
621     if (out_llend) {
622         *out_llend = llend;
623     }
624     return true;
625 }
626 
627 static void vfio_listener_region_add(MemoryListener *listener,
628                                      MemoryRegionSection *section)
629 {
630     VFIOContainer *container = container_of(listener, VFIOContainer, listener);
631     hwaddr iova, end;
632     Int128 llend, llsize;
633     void *vaddr;
634     int ret;
635     VFIOHostDMAWindow *hostwin;
636     Error *err = NULL;
637 
638     if (!vfio_listener_valid_section(section, "region_add")) {
639         return;
640     }
641 
642     if (!vfio_get_section_iova_range(container, section, &iova, &end, &llend)) {
643         if (memory_region_is_ram_device(section->mr)) {
644             trace_vfio_listener_region_add_no_dma_map(
645                 memory_region_name(section->mr),
646                 section->offset_within_address_space,
647                 int128_getlo(section->size),
648                 qemu_real_host_page_size());
649         }
650         return;
651     }
652 
653     if (vfio_container_add_section_window(container, section, &err)) {
654         goto fail;
655     }
656 
657     hostwin = vfio_find_hostwin(container, iova, end);
658     if (!hostwin) {
659         error_setg(&err, "Container %p can't map guest IOVA region"
660                    " 0x%"HWADDR_PRIx"..0x%"HWADDR_PRIx, container, iova, end);
661         goto fail;
662     }
663 
664     memory_region_ref(section->mr);
665 
666     if (memory_region_is_iommu(section->mr)) {
667         VFIOGuestIOMMU *giommu;
668         IOMMUMemoryRegion *iommu_mr = IOMMU_MEMORY_REGION(section->mr);
669         int iommu_idx;
670 
671         trace_vfio_listener_region_add_iommu(iova, end);
672         /*
673          * FIXME: For VFIO iommu types which have KVM acceleration to
674          * avoid bouncing all map/unmaps through qemu this way, this
675          * would be the right place to wire that up (tell the KVM
676          * device emulation the VFIO iommu handles to use).
677          */
678         giommu = g_malloc0(sizeof(*giommu));
679         giommu->iommu_mr = iommu_mr;
680         giommu->iommu_offset = section->offset_within_address_space -
681                                section->offset_within_region;
682         giommu->container = container;
683         llend = int128_add(int128_make64(section->offset_within_region),
684                            section->size);
685         llend = int128_sub(llend, int128_one());
686         iommu_idx = memory_region_iommu_attrs_to_index(iommu_mr,
687                                                        MEMTXATTRS_UNSPECIFIED);
688         iommu_notifier_init(&giommu->n, vfio_iommu_map_notify,
689                             IOMMU_NOTIFIER_IOTLB_EVENTS,
690                             section->offset_within_region,
691                             int128_get64(llend),
692                             iommu_idx);
693 
694         ret = memory_region_iommu_set_page_size_mask(giommu->iommu_mr,
695                                                      container->pgsizes,
696                                                      &err);
697         if (ret) {
698             g_free(giommu);
699             goto fail;
700         }
701 
702         ret = memory_region_register_iommu_notifier(section->mr, &giommu->n,
703                                                     &err);
704         if (ret) {
705             g_free(giommu);
706             goto fail;
707         }
708         QLIST_INSERT_HEAD(&container->giommu_list, giommu, giommu_next);
709         memory_region_iommu_replay(giommu->iommu_mr, &giommu->n);
710 
711         return;
712     }
713 
714     /* Here we assume that memory_region_is_ram(section->mr)==true */
715 
716     /*
717      * For RAM memory regions with a RamDiscardManager, we only want to map the
718      * actually populated parts - and update the mapping whenever we're notified
719      * about changes.
720      */
721     if (memory_region_has_ram_discard_manager(section->mr)) {
722         vfio_register_ram_discard_listener(container, section);
723         return;
724     }
725 
726     vaddr = memory_region_get_ram_ptr(section->mr) +
727             section->offset_within_region +
728             (iova - section->offset_within_address_space);
729 
730     trace_vfio_listener_region_add_ram(iova, end, vaddr);
731 
732     llsize = int128_sub(llend, int128_make64(iova));
733 
734     if (memory_region_is_ram_device(section->mr)) {
735         hwaddr pgmask = (1ULL << ctz64(hostwin->iova_pgsizes)) - 1;
736 
737         if ((iova & pgmask) || (int128_get64(llsize) & pgmask)) {
738             trace_vfio_listener_region_add_no_dma_map(
739                 memory_region_name(section->mr),
740                 section->offset_within_address_space,
741                 int128_getlo(section->size),
742                 pgmask + 1);
743             return;
744         }
745     }
746 
747     ret = vfio_dma_map(container, iova, int128_get64(llsize),
748                        vaddr, section->readonly);
749     if (ret) {
750         error_setg(&err, "vfio_dma_map(%p, 0x%"HWADDR_PRIx", "
751                    "0x%"HWADDR_PRIx", %p) = %d (%s)",
752                    container, iova, int128_get64(llsize), vaddr, ret,
753                    strerror(-ret));
754         if (memory_region_is_ram_device(section->mr)) {
755             /* Allow unexpected mappings not to be fatal for RAM devices */
756             error_report_err(err);
757             return;
758         }
759         goto fail;
760     }
761 
762     return;
763 
764 fail:
765     if (memory_region_is_ram_device(section->mr)) {
766         error_reportf_err(err, "PCI p2p may not work: ");
767         return;
768     }
769     /*
770      * On the initfn path, store the first error in the container so we
771      * can gracefully fail.  Runtime, there's not much we can do other
772      * than throw a hardware error.
773      */
774     if (!container->initialized) {
775         if (!container->error) {
776             error_propagate_prepend(&container->error, err,
777                                     "Region %s: ",
778                                     memory_region_name(section->mr));
779         } else {
780             error_free(err);
781         }
782     } else {
783         error_report_err(err);
784         hw_error("vfio: DMA mapping failed, unable to continue");
785     }
786 }
787 
788 static void vfio_listener_region_del(MemoryListener *listener,
789                                      MemoryRegionSection *section)
790 {
791     VFIOContainer *container = container_of(listener, VFIOContainer, listener);
792     hwaddr iova, end;
793     Int128 llend, llsize;
794     int ret;
795     bool try_unmap = true;
796 
797     if (!vfio_listener_valid_section(section, "region_del")) {
798         return;
799     }
800 
801     if (memory_region_is_iommu(section->mr)) {
802         VFIOGuestIOMMU *giommu;
803 
804         QLIST_FOREACH(giommu, &container->giommu_list, giommu_next) {
805             if (MEMORY_REGION(giommu->iommu_mr) == section->mr &&
806                 giommu->n.start == section->offset_within_region) {
807                 memory_region_unregister_iommu_notifier(section->mr,
808                                                         &giommu->n);
809                 QLIST_REMOVE(giommu, giommu_next);
810                 g_free(giommu);
811                 break;
812             }
813         }
814 
815         /*
816          * FIXME: We assume the one big unmap below is adequate to
817          * remove any individual page mappings in the IOMMU which
818          * might have been copied into VFIO. This works for a page table
819          * based IOMMU where a big unmap flattens a large range of IO-PTEs.
820          * That may not be true for all IOMMU types.
821          */
822     }
823 
824     if (!vfio_get_section_iova_range(container, section, &iova, &end, &llend)) {
825         return;
826     }
827 
828     llsize = int128_sub(llend, int128_make64(iova));
829 
830     trace_vfio_listener_region_del(iova, end);
831 
832     if (memory_region_is_ram_device(section->mr)) {
833         hwaddr pgmask;
834         VFIOHostDMAWindow *hostwin;
835 
836         hostwin = vfio_find_hostwin(container, iova, end);
837         assert(hostwin); /* or region_add() would have failed */
838 
839         pgmask = (1ULL << ctz64(hostwin->iova_pgsizes)) - 1;
840         try_unmap = !((iova & pgmask) || (int128_get64(llsize) & pgmask));
841     } else if (memory_region_has_ram_discard_manager(section->mr)) {
842         vfio_unregister_ram_discard_listener(container, section);
843         /* Unregistering will trigger an unmap. */
844         try_unmap = false;
845     }
846 
847     if (try_unmap) {
848         if (int128_eq(llsize, int128_2_64())) {
849             /* The unmap ioctl doesn't accept a full 64-bit span. */
850             llsize = int128_rshift(llsize, 1);
851             ret = vfio_dma_unmap(container, iova, int128_get64(llsize), NULL);
852             if (ret) {
853                 error_report("vfio_dma_unmap(%p, 0x%"HWADDR_PRIx", "
854                              "0x%"HWADDR_PRIx") = %d (%s)",
855                              container, iova, int128_get64(llsize), ret,
856                              strerror(-ret));
857             }
858             iova += int128_get64(llsize);
859         }
860         ret = vfio_dma_unmap(container, iova, int128_get64(llsize), NULL);
861         if (ret) {
862             error_report("vfio_dma_unmap(%p, 0x%"HWADDR_PRIx", "
863                          "0x%"HWADDR_PRIx") = %d (%s)",
864                          container, iova, int128_get64(llsize), ret,
865                          strerror(-ret));
866         }
867     }
868 
869     memory_region_unref(section->mr);
870 
871     vfio_container_del_section_window(container, section);
872 }
873 
874 typedef struct VFIODirtyRanges {
875     hwaddr min32;
876     hwaddr max32;
877     hwaddr min64;
878     hwaddr max64;
879     hwaddr minpci64;
880     hwaddr maxpci64;
881 } VFIODirtyRanges;
882 
883 typedef struct VFIODirtyRangesListener {
884     VFIOContainer *container;
885     VFIODirtyRanges ranges;
886     MemoryListener listener;
887 } VFIODirtyRangesListener;
888 
889 static bool vfio_section_is_vfio_pci(MemoryRegionSection *section,
890                                      VFIOContainer *container)
891 {
892     VFIOPCIDevice *pcidev;
893     VFIODevice *vbasedev;
894     Object *owner;
895 
896     owner = memory_region_owner(section->mr);
897 
898     QLIST_FOREACH(vbasedev, &container->device_list, container_next) {
899         if (vbasedev->type != VFIO_DEVICE_TYPE_PCI) {
900             continue;
901         }
902         pcidev = container_of(vbasedev, VFIOPCIDevice, vbasedev);
903         if (OBJECT(pcidev) == owner) {
904             return true;
905         }
906     }
907 
908     return false;
909 }
910 
911 static void vfio_dirty_tracking_update(MemoryListener *listener,
912                                        MemoryRegionSection *section)
913 {
914     VFIODirtyRangesListener *dirty = container_of(listener,
915                                                   VFIODirtyRangesListener,
916                                                   listener);
917     VFIODirtyRanges *range = &dirty->ranges;
918     hwaddr iova, end, *min, *max;
919 
920     if (!vfio_listener_valid_section(section, "tracking_update") ||
921         !vfio_get_section_iova_range(dirty->container, section,
922                                      &iova, &end, NULL)) {
923         return;
924     }
925 
926     /*
927      * The address space passed to the dirty tracker is reduced to three ranges:
928      * one for 32-bit DMA ranges, one for 64-bit DMA ranges and one for the
929      * PCI 64-bit hole.
930      *
931      * The underlying reports of dirty will query a sub-interval of each of
932      * these ranges.
933      *
934      * The purpose of the three range handling is to handle known cases of big
935      * holes in the address space, like the x86 AMD 1T hole, and firmware (like
936      * OVMF) which may relocate the pci-hole64 to the end of the address space.
937      * The latter would otherwise generate large ranges for tracking, stressing
938      * the limits of supported hardware. The pci-hole32 will always be below 4G
939      * (overlapping or not) so it doesn't need special handling and is part of
940      * the 32-bit range.
941      *
942      * The alternative would be an IOVATree but that has a much bigger runtime
943      * overhead and unnecessary complexity.
944      */
945     if (vfio_section_is_vfio_pci(section, dirty->container) &&
946         iova >= UINT32_MAX) {
947         min = &range->minpci64;
948         max = &range->maxpci64;
949     } else {
950         min = (end <= UINT32_MAX) ? &range->min32 : &range->min64;
951         max = (end <= UINT32_MAX) ? &range->max32 : &range->max64;
952     }
953     if (*min > iova) {
954         *min = iova;
955     }
956     if (*max < end) {
957         *max = end;
958     }
959 
960     trace_vfio_device_dirty_tracking_update(iova, end, *min, *max);
961     return;
962 }
963 
964 static const MemoryListener vfio_dirty_tracking_listener = {
965     .name = "vfio-tracking",
966     .region_add = vfio_dirty_tracking_update,
967 };
968 
969 static void vfio_dirty_tracking_init(VFIOContainer *container,
970                                      VFIODirtyRanges *ranges)
971 {
972     VFIODirtyRangesListener dirty;
973 
974     memset(&dirty, 0, sizeof(dirty));
975     dirty.ranges.min32 = UINT32_MAX;
976     dirty.ranges.min64 = UINT64_MAX;
977     dirty.ranges.minpci64 = UINT64_MAX;
978     dirty.listener = vfio_dirty_tracking_listener;
979     dirty.container = container;
980 
981     memory_listener_register(&dirty.listener,
982                              container->space->as);
983 
984     *ranges = dirty.ranges;
985 
986     /*
987      * The memory listener is synchronous, and used to calculate the range
988      * to dirty tracking. Unregister it after we are done as we are not
989      * interested in any follow-up updates.
990      */
991     memory_listener_unregister(&dirty.listener);
992 }
993 
994 static void vfio_devices_dma_logging_stop(VFIOContainer *container)
995 {
996     uint64_t buf[DIV_ROUND_UP(sizeof(struct vfio_device_feature),
997                               sizeof(uint64_t))] = {};
998     struct vfio_device_feature *feature = (struct vfio_device_feature *)buf;
999     VFIODevice *vbasedev;
1000 
1001     feature->argsz = sizeof(buf);
1002     feature->flags = VFIO_DEVICE_FEATURE_SET |
1003                      VFIO_DEVICE_FEATURE_DMA_LOGGING_STOP;
1004 
1005     QLIST_FOREACH(vbasedev, &container->device_list, container_next) {
1006         if (!vbasedev->dirty_tracking) {
1007             continue;
1008         }
1009 
1010         if (ioctl(vbasedev->fd, VFIO_DEVICE_FEATURE, feature)) {
1011             warn_report("%s: Failed to stop DMA logging, err %d (%s)",
1012                         vbasedev->name, -errno, strerror(errno));
1013         }
1014         vbasedev->dirty_tracking = false;
1015     }
1016 }
1017 
1018 static struct vfio_device_feature *
1019 vfio_device_feature_dma_logging_start_create(VFIOContainer *container,
1020                                              VFIODirtyRanges *tracking)
1021 {
1022     struct vfio_device_feature *feature;
1023     size_t feature_size;
1024     struct vfio_device_feature_dma_logging_control *control;
1025     struct vfio_device_feature_dma_logging_range *ranges;
1026 
1027     feature_size = sizeof(struct vfio_device_feature) +
1028                    sizeof(struct vfio_device_feature_dma_logging_control);
1029     feature = g_try_malloc0(feature_size);
1030     if (!feature) {
1031         errno = ENOMEM;
1032         return NULL;
1033     }
1034     feature->argsz = feature_size;
1035     feature->flags = VFIO_DEVICE_FEATURE_SET |
1036                      VFIO_DEVICE_FEATURE_DMA_LOGGING_START;
1037 
1038     control = (struct vfio_device_feature_dma_logging_control *)feature->data;
1039     control->page_size = qemu_real_host_page_size();
1040 
1041     /*
1042      * DMA logging uAPI guarantees to support at least a number of ranges that
1043      * fits into a single host kernel base page.
1044      */
1045     control->num_ranges = !!tracking->max32 + !!tracking->max64 +
1046         !!tracking->maxpci64;
1047     ranges = g_try_new0(struct vfio_device_feature_dma_logging_range,
1048                         control->num_ranges);
1049     if (!ranges) {
1050         g_free(feature);
1051         errno = ENOMEM;
1052 
1053         return NULL;
1054     }
1055 
1056     control->ranges = (__u64)(uintptr_t)ranges;
1057     if (tracking->max32) {
1058         ranges->iova = tracking->min32;
1059         ranges->length = (tracking->max32 - tracking->min32) + 1;
1060         ranges++;
1061     }
1062     if (tracking->max64) {
1063         ranges->iova = tracking->min64;
1064         ranges->length = (tracking->max64 - tracking->min64) + 1;
1065         ranges++;
1066     }
1067     if (tracking->maxpci64) {
1068         ranges->iova = tracking->minpci64;
1069         ranges->length = (tracking->maxpci64 - tracking->minpci64) + 1;
1070     }
1071 
1072     trace_vfio_device_dirty_tracking_start(control->num_ranges,
1073                                            tracking->min32, tracking->max32,
1074                                            tracking->min64, tracking->max64,
1075                                            tracking->minpci64, tracking->maxpci64);
1076 
1077     return feature;
1078 }
1079 
1080 static void vfio_device_feature_dma_logging_start_destroy(
1081     struct vfio_device_feature *feature)
1082 {
1083     struct vfio_device_feature_dma_logging_control *control =
1084         (struct vfio_device_feature_dma_logging_control *)feature->data;
1085     struct vfio_device_feature_dma_logging_range *ranges =
1086         (struct vfio_device_feature_dma_logging_range *)(uintptr_t)control->ranges;
1087 
1088     g_free(ranges);
1089     g_free(feature);
1090 }
1091 
1092 static int vfio_devices_dma_logging_start(VFIOContainer *container)
1093 {
1094     struct vfio_device_feature *feature;
1095     VFIODirtyRanges ranges;
1096     VFIODevice *vbasedev;
1097     int ret = 0;
1098 
1099     vfio_dirty_tracking_init(container, &ranges);
1100     feature = vfio_device_feature_dma_logging_start_create(container,
1101                                                            &ranges);
1102     if (!feature) {
1103         return -errno;
1104     }
1105 
1106     QLIST_FOREACH(vbasedev, &container->device_list, container_next) {
1107         if (vbasedev->dirty_tracking) {
1108             continue;
1109         }
1110 
1111         ret = ioctl(vbasedev->fd, VFIO_DEVICE_FEATURE, feature);
1112         if (ret) {
1113             ret = -errno;
1114             error_report("%s: Failed to start DMA logging, err %d (%s)",
1115                          vbasedev->name, ret, strerror(errno));
1116             goto out;
1117         }
1118         vbasedev->dirty_tracking = true;
1119     }
1120 
1121 out:
1122     if (ret) {
1123         vfio_devices_dma_logging_stop(container);
1124     }
1125 
1126     vfio_device_feature_dma_logging_start_destroy(feature);
1127 
1128     return ret;
1129 }
1130 
1131 static void vfio_listener_log_global_start(MemoryListener *listener)
1132 {
1133     VFIOContainer *container = container_of(listener, VFIOContainer, listener);
1134     int ret;
1135 
1136     if (vfio_devices_all_device_dirty_tracking(container)) {
1137         ret = vfio_devices_dma_logging_start(container);
1138     } else {
1139         ret = vfio_set_dirty_page_tracking(container, true);
1140     }
1141 
1142     if (ret) {
1143         error_report("vfio: Could not start dirty page tracking, err: %d (%s)",
1144                      ret, strerror(-ret));
1145         vfio_set_migration_error(ret);
1146     }
1147 }
1148 
1149 static void vfio_listener_log_global_stop(MemoryListener *listener)
1150 {
1151     VFIOContainer *container = container_of(listener, VFIOContainer, listener);
1152     int ret = 0;
1153 
1154     if (vfio_devices_all_device_dirty_tracking(container)) {
1155         vfio_devices_dma_logging_stop(container);
1156     } else {
1157         ret = vfio_set_dirty_page_tracking(container, false);
1158     }
1159 
1160     if (ret) {
1161         error_report("vfio: Could not stop dirty page tracking, err: %d (%s)",
1162                      ret, strerror(-ret));
1163         vfio_set_migration_error(ret);
1164     }
1165 }
1166 
1167 static int vfio_device_dma_logging_report(VFIODevice *vbasedev, hwaddr iova,
1168                                           hwaddr size, void *bitmap)
1169 {
1170     uint64_t buf[DIV_ROUND_UP(sizeof(struct vfio_device_feature) +
1171                         sizeof(struct vfio_device_feature_dma_logging_report),
1172                         sizeof(__u64))] = {};
1173     struct vfio_device_feature *feature = (struct vfio_device_feature *)buf;
1174     struct vfio_device_feature_dma_logging_report *report =
1175         (struct vfio_device_feature_dma_logging_report *)feature->data;
1176 
1177     report->iova = iova;
1178     report->length = size;
1179     report->page_size = qemu_real_host_page_size();
1180     report->bitmap = (__u64)(uintptr_t)bitmap;
1181 
1182     feature->argsz = sizeof(buf);
1183     feature->flags = VFIO_DEVICE_FEATURE_GET |
1184                      VFIO_DEVICE_FEATURE_DMA_LOGGING_REPORT;
1185 
1186     if (ioctl(vbasedev->fd, VFIO_DEVICE_FEATURE, feature)) {
1187         return -errno;
1188     }
1189 
1190     return 0;
1191 }
1192 
1193 int vfio_devices_query_dirty_bitmap(VFIOContainer *container,
1194                                     VFIOBitmap *vbmap, hwaddr iova,
1195                                     hwaddr size)
1196 {
1197     VFIODevice *vbasedev;
1198     int ret;
1199 
1200     QLIST_FOREACH(vbasedev, &container->device_list, container_next) {
1201         ret = vfio_device_dma_logging_report(vbasedev, iova, size,
1202                                              vbmap->bitmap);
1203         if (ret) {
1204             error_report("%s: Failed to get DMA logging report, iova: "
1205                          "0x%" HWADDR_PRIx ", size: 0x%" HWADDR_PRIx
1206                          ", err: %d (%s)",
1207                          vbasedev->name, iova, size, ret, strerror(-ret));
1208 
1209             return ret;
1210         }
1211     }
1212 
1213     return 0;
1214 }
1215 
1216 int vfio_get_dirty_bitmap(VFIOContainer *container, uint64_t iova,
1217                           uint64_t size, ram_addr_t ram_addr)
1218 {
1219     bool all_device_dirty_tracking =
1220         vfio_devices_all_device_dirty_tracking(container);
1221     uint64_t dirty_pages;
1222     VFIOBitmap vbmap;
1223     int ret;
1224 
1225     if (!container->dirty_pages_supported && !all_device_dirty_tracking) {
1226         cpu_physical_memory_set_dirty_range(ram_addr, size,
1227                                             tcg_enabled() ? DIRTY_CLIENTS_ALL :
1228                                             DIRTY_CLIENTS_NOCODE);
1229         return 0;
1230     }
1231 
1232     ret = vfio_bitmap_alloc(&vbmap, size);
1233     if (ret) {
1234         return ret;
1235     }
1236 
1237     if (all_device_dirty_tracking) {
1238         ret = vfio_devices_query_dirty_bitmap(container, &vbmap, iova, size);
1239     } else {
1240         ret = vfio_query_dirty_bitmap(container, &vbmap, iova, size);
1241     }
1242 
1243     if (ret) {
1244         goto out;
1245     }
1246 
1247     dirty_pages = cpu_physical_memory_set_dirty_lebitmap(vbmap.bitmap, ram_addr,
1248                                                          vbmap.pages);
1249 
1250     trace_vfio_get_dirty_bitmap(container->fd, iova, size, vbmap.size,
1251                                 ram_addr, dirty_pages);
1252 out:
1253     g_free(vbmap.bitmap);
1254 
1255     return ret;
1256 }
1257 
1258 typedef struct {
1259     IOMMUNotifier n;
1260     VFIOGuestIOMMU *giommu;
1261 } vfio_giommu_dirty_notifier;
1262 
1263 static void vfio_iommu_map_dirty_notify(IOMMUNotifier *n, IOMMUTLBEntry *iotlb)
1264 {
1265     vfio_giommu_dirty_notifier *gdn = container_of(n,
1266                                                 vfio_giommu_dirty_notifier, n);
1267     VFIOGuestIOMMU *giommu = gdn->giommu;
1268     VFIOContainer *container = giommu->container;
1269     hwaddr iova = iotlb->iova + giommu->iommu_offset;
1270     ram_addr_t translated_addr;
1271     int ret = -EINVAL;
1272 
1273     trace_vfio_iommu_map_dirty_notify(iova, iova + iotlb->addr_mask);
1274 
1275     if (iotlb->target_as != &address_space_memory) {
1276         error_report("Wrong target AS \"%s\", only system memory is allowed",
1277                      iotlb->target_as->name ? iotlb->target_as->name : "none");
1278         goto out;
1279     }
1280 
1281     rcu_read_lock();
1282     if (vfio_get_xlat_addr(iotlb, NULL, &translated_addr, NULL)) {
1283         ret = vfio_get_dirty_bitmap(container, iova, iotlb->addr_mask + 1,
1284                                     translated_addr);
1285         if (ret) {
1286             error_report("vfio_iommu_map_dirty_notify(%p, 0x%"HWADDR_PRIx", "
1287                          "0x%"HWADDR_PRIx") = %d (%s)",
1288                          container, iova, iotlb->addr_mask + 1, ret,
1289                          strerror(-ret));
1290         }
1291     }
1292     rcu_read_unlock();
1293 
1294 out:
1295     if (ret) {
1296         vfio_set_migration_error(ret);
1297     }
1298 }
1299 
1300 static int vfio_ram_discard_get_dirty_bitmap(MemoryRegionSection *section,
1301                                              void *opaque)
1302 {
1303     const hwaddr size = int128_get64(section->size);
1304     const hwaddr iova = section->offset_within_address_space;
1305     const ram_addr_t ram_addr = memory_region_get_ram_addr(section->mr) +
1306                                 section->offset_within_region;
1307     VFIORamDiscardListener *vrdl = opaque;
1308 
1309     /*
1310      * Sync the whole mapped region (spanning multiple individual mappings)
1311      * in one go.
1312      */
1313     return vfio_get_dirty_bitmap(vrdl->container, iova, size, ram_addr);
1314 }
1315 
1316 static int vfio_sync_ram_discard_listener_dirty_bitmap(VFIOContainer *container,
1317                                                    MemoryRegionSection *section)
1318 {
1319     RamDiscardManager *rdm = memory_region_get_ram_discard_manager(section->mr);
1320     VFIORamDiscardListener *vrdl = NULL;
1321 
1322     QLIST_FOREACH(vrdl, &container->vrdl_list, next) {
1323         if (vrdl->mr == section->mr &&
1324             vrdl->offset_within_address_space ==
1325             section->offset_within_address_space) {
1326             break;
1327         }
1328     }
1329 
1330     if (!vrdl) {
1331         hw_error("vfio: Trying to sync missing RAM discard listener");
1332     }
1333 
1334     /*
1335      * We only want/can synchronize the bitmap for actually mapped parts -
1336      * which correspond to populated parts. Replay all populated parts.
1337      */
1338     return ram_discard_manager_replay_populated(rdm, section,
1339                                               vfio_ram_discard_get_dirty_bitmap,
1340                                                 &vrdl);
1341 }
1342 
1343 static int vfio_sync_dirty_bitmap(VFIOContainer *container,
1344                                   MemoryRegionSection *section)
1345 {
1346     ram_addr_t ram_addr;
1347 
1348     if (memory_region_is_iommu(section->mr)) {
1349         VFIOGuestIOMMU *giommu;
1350 
1351         QLIST_FOREACH(giommu, &container->giommu_list, giommu_next) {
1352             if (MEMORY_REGION(giommu->iommu_mr) == section->mr &&
1353                 giommu->n.start == section->offset_within_region) {
1354                 Int128 llend;
1355                 vfio_giommu_dirty_notifier gdn = { .giommu = giommu };
1356                 int idx = memory_region_iommu_attrs_to_index(giommu->iommu_mr,
1357                                                        MEMTXATTRS_UNSPECIFIED);
1358 
1359                 llend = int128_add(int128_make64(section->offset_within_region),
1360                                    section->size);
1361                 llend = int128_sub(llend, int128_one());
1362 
1363                 iommu_notifier_init(&gdn.n,
1364                                     vfio_iommu_map_dirty_notify,
1365                                     IOMMU_NOTIFIER_MAP,
1366                                     section->offset_within_region,
1367                                     int128_get64(llend),
1368                                     idx);
1369                 memory_region_iommu_replay(giommu->iommu_mr, &gdn.n);
1370                 break;
1371             }
1372         }
1373         return 0;
1374     } else if (memory_region_has_ram_discard_manager(section->mr)) {
1375         return vfio_sync_ram_discard_listener_dirty_bitmap(container, section);
1376     }
1377 
1378     ram_addr = memory_region_get_ram_addr(section->mr) +
1379                section->offset_within_region;
1380 
1381     return vfio_get_dirty_bitmap(container,
1382                    REAL_HOST_PAGE_ALIGN(section->offset_within_address_space),
1383                    int128_get64(section->size), ram_addr);
1384 }
1385 
1386 static void vfio_listener_log_sync(MemoryListener *listener,
1387         MemoryRegionSection *section)
1388 {
1389     VFIOContainer *container = container_of(listener, VFIOContainer, listener);
1390     int ret;
1391 
1392     if (vfio_listener_skipped_section(section)) {
1393         return;
1394     }
1395 
1396     if (vfio_devices_all_dirty_tracking(container)) {
1397         ret = vfio_sync_dirty_bitmap(container, section);
1398         if (ret) {
1399             error_report("vfio: Failed to sync dirty bitmap, err: %d (%s)", ret,
1400                          strerror(-ret));
1401             vfio_set_migration_error(ret);
1402         }
1403     }
1404 }
1405 
1406 const MemoryListener vfio_memory_listener = {
1407     .name = "vfio",
1408     .region_add = vfio_listener_region_add,
1409     .region_del = vfio_listener_region_del,
1410     .log_global_start = vfio_listener_log_global_start,
1411     .log_global_stop = vfio_listener_log_global_stop,
1412     .log_sync = vfio_listener_log_sync,
1413 };
1414 
1415 void vfio_reset_handler(void *opaque)
1416 {
1417     VFIODevice *vbasedev;
1418 
1419     QLIST_FOREACH(vbasedev, &vfio_device_list, next) {
1420         if (vbasedev->dev->realized) {
1421             vbasedev->ops->vfio_compute_needs_reset(vbasedev);
1422         }
1423     }
1424 
1425     QLIST_FOREACH(vbasedev, &vfio_device_list, next) {
1426         if (vbasedev->dev->realized && vbasedev->needs_reset) {
1427             vbasedev->ops->vfio_hot_reset_multi(vbasedev);
1428         }
1429     }
1430 }
1431 
1432 int vfio_kvm_device_add_fd(int fd, Error **errp)
1433 {
1434 #ifdef CONFIG_KVM
1435     struct kvm_device_attr attr = {
1436         .group = KVM_DEV_VFIO_FILE,
1437         .attr = KVM_DEV_VFIO_FILE_ADD,
1438         .addr = (uint64_t)(unsigned long)&fd,
1439     };
1440 
1441     if (!kvm_enabled()) {
1442         return 0;
1443     }
1444 
1445     if (vfio_kvm_device_fd < 0) {
1446         struct kvm_create_device cd = {
1447             .type = KVM_DEV_TYPE_VFIO,
1448         };
1449 
1450         if (kvm_vm_ioctl(kvm_state, KVM_CREATE_DEVICE, &cd)) {
1451             error_setg_errno(errp, errno, "Failed to create KVM VFIO device");
1452             return -errno;
1453         }
1454 
1455         vfio_kvm_device_fd = cd.fd;
1456     }
1457 
1458     if (ioctl(vfio_kvm_device_fd, KVM_SET_DEVICE_ATTR, &attr)) {
1459         error_setg_errno(errp, errno, "Failed to add fd %d to KVM VFIO device",
1460                          fd);
1461         return -errno;
1462     }
1463 #endif
1464     return 0;
1465 }
1466 
1467 int vfio_kvm_device_del_fd(int fd, Error **errp)
1468 {
1469 #ifdef CONFIG_KVM
1470     struct kvm_device_attr attr = {
1471         .group = KVM_DEV_VFIO_FILE,
1472         .attr = KVM_DEV_VFIO_FILE_DEL,
1473         .addr = (uint64_t)(unsigned long)&fd,
1474     };
1475 
1476     if (vfio_kvm_device_fd < 0) {
1477         error_setg(errp, "KVM VFIO device isn't created yet");
1478         return -EINVAL;
1479     }
1480 
1481     if (ioctl(vfio_kvm_device_fd, KVM_SET_DEVICE_ATTR, &attr)) {
1482         error_setg_errno(errp, errno,
1483                          "Failed to remove fd %d from KVM VFIO device", fd);
1484         return -errno;
1485     }
1486 #endif
1487     return 0;
1488 }
1489 
1490 VFIOAddressSpace *vfio_get_address_space(AddressSpace *as)
1491 {
1492     VFIOAddressSpace *space;
1493 
1494     QLIST_FOREACH(space, &vfio_address_spaces, list) {
1495         if (space->as == as) {
1496             return space;
1497         }
1498     }
1499 
1500     /* No suitable VFIOAddressSpace, create a new one */
1501     space = g_malloc0(sizeof(*space));
1502     space->as = as;
1503     QLIST_INIT(&space->containers);
1504 
1505     if (QLIST_EMPTY(&vfio_address_spaces)) {
1506         qemu_register_reset(vfio_reset_handler, NULL);
1507     }
1508 
1509     QLIST_INSERT_HEAD(&vfio_address_spaces, space, list);
1510 
1511     return space;
1512 }
1513 
1514 void vfio_put_address_space(VFIOAddressSpace *space)
1515 {
1516     if (QLIST_EMPTY(&space->containers)) {
1517         QLIST_REMOVE(space, list);
1518         g_free(space);
1519     }
1520     if (QLIST_EMPTY(&vfio_address_spaces)) {
1521         qemu_unregister_reset(vfio_reset_handler, NULL);
1522     }
1523 }
1524 
1525 struct vfio_device_info *vfio_get_device_info(int fd)
1526 {
1527     struct vfio_device_info *info;
1528     uint32_t argsz = sizeof(*info);
1529 
1530     info = g_malloc0(argsz);
1531 
1532 retry:
1533     info->argsz = argsz;
1534 
1535     if (ioctl(fd, VFIO_DEVICE_GET_INFO, info)) {
1536         g_free(info);
1537         return NULL;
1538     }
1539 
1540     if (info->argsz > argsz) {
1541         argsz = info->argsz;
1542         info = g_realloc(info, argsz);
1543         goto retry;
1544     }
1545 
1546     return info;
1547 }
1548