xref: /openbmc/qemu/hw/vfio/common.c (revision ef5cc166)
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/migration.h"
43 #include "migration/misc.h"
44 #include "migration/blocker.h"
45 #include "migration/qemu-file.h"
46 #include "sysemu/tpm.h"
47 
48 VFIODeviceList vfio_device_list =
49     QLIST_HEAD_INITIALIZER(vfio_device_list);
50 static QLIST_HEAD(, VFIOAddressSpace) vfio_address_spaces =
51     QLIST_HEAD_INITIALIZER(vfio_address_spaces);
52 
53 #ifdef CONFIG_KVM
54 /*
55  * We have a single VFIO pseudo device per KVM VM.  Once created it lives
56  * for the life of the VM.  Closing the file descriptor only drops our
57  * reference to it and the device's reference to kvm.  Therefore once
58  * initialized, this file descriptor is only released on QEMU exit and
59  * we'll re-use it should another vfio device be attached before then.
60  */
61 int vfio_kvm_device_fd = -1;
62 #endif
63 
64 /*
65  * Device state interfaces
66  */
67 
68 bool vfio_mig_active(void)
69 {
70     VFIODevice *vbasedev;
71 
72     if (QLIST_EMPTY(&vfio_device_list)) {
73         return false;
74     }
75 
76     QLIST_FOREACH(vbasedev, &vfio_device_list, global_next) {
77         if (vbasedev->migration_blocker) {
78             return false;
79         }
80     }
81     return true;
82 }
83 
84 static Error *multiple_devices_migration_blocker;
85 
86 /*
87  * Multiple devices migration is allowed only if all devices support P2P
88  * migration. Single device migration is allowed regardless of P2P migration
89  * support.
90  */
91 static bool vfio_multiple_devices_migration_is_supported(void)
92 {
93     VFIODevice *vbasedev;
94     unsigned int device_num = 0;
95     bool all_support_p2p = true;
96 
97     QLIST_FOREACH(vbasedev, &vfio_device_list, global_next) {
98         if (vbasedev->migration) {
99             device_num++;
100 
101             if (!(vbasedev->migration->mig_flags & VFIO_MIGRATION_P2P)) {
102                 all_support_p2p = false;
103             }
104         }
105     }
106 
107     return all_support_p2p || device_num <= 1;
108 }
109 
110 int vfio_block_multiple_devices_migration(VFIODevice *vbasedev, Error **errp)
111 {
112     int ret;
113 
114     if (vfio_multiple_devices_migration_is_supported()) {
115         return 0;
116     }
117 
118     if (vbasedev->enable_migration == ON_OFF_AUTO_ON) {
119         error_setg(errp, "Multiple VFIO devices migration is supported only if "
120                          "all of them support P2P migration");
121         return -EINVAL;
122     }
123 
124     if (multiple_devices_migration_blocker) {
125         return 0;
126     }
127 
128     error_setg(&multiple_devices_migration_blocker,
129                "Multiple VFIO devices migration is supported only if all of "
130                "them support P2P migration");
131     ret = migrate_add_blocker(&multiple_devices_migration_blocker, errp);
132 
133     return ret;
134 }
135 
136 void vfio_unblock_multiple_devices_migration(void)
137 {
138     if (!multiple_devices_migration_blocker ||
139         !vfio_multiple_devices_migration_is_supported()) {
140         return;
141     }
142 
143     migrate_del_blocker(&multiple_devices_migration_blocker);
144 }
145 
146 bool vfio_viommu_preset(VFIODevice *vbasedev)
147 {
148     return vbasedev->bcontainer->space->as != &address_space_memory;
149 }
150 
151 static void vfio_set_migration_error(int err)
152 {
153     MigrationState *ms = migrate_get_current();
154 
155     if (migration_is_setup_or_active(ms->state)) {
156         WITH_QEMU_LOCK_GUARD(&ms->qemu_file_lock) {
157             if (ms->to_dst_file) {
158                 qemu_file_set_error(ms->to_dst_file, err);
159             }
160         }
161     }
162 }
163 
164 bool vfio_device_state_is_running(VFIODevice *vbasedev)
165 {
166     VFIOMigration *migration = vbasedev->migration;
167 
168     return migration->device_state == VFIO_DEVICE_STATE_RUNNING ||
169            migration->device_state == VFIO_DEVICE_STATE_RUNNING_P2P;
170 }
171 
172 bool vfio_device_state_is_precopy(VFIODevice *vbasedev)
173 {
174     VFIOMigration *migration = vbasedev->migration;
175 
176     return migration->device_state == VFIO_DEVICE_STATE_PRE_COPY ||
177            migration->device_state == VFIO_DEVICE_STATE_PRE_COPY_P2P;
178 }
179 
180 static bool vfio_devices_all_dirty_tracking(VFIOContainerBase *bcontainer)
181 {
182     VFIODevice *vbasedev;
183     MigrationState *ms = migrate_get_current();
184 
185     if (ms->state != MIGRATION_STATUS_ACTIVE &&
186         ms->state != MIGRATION_STATUS_DEVICE) {
187         return false;
188     }
189 
190     QLIST_FOREACH(vbasedev, &bcontainer->device_list, container_next) {
191         VFIOMigration *migration = vbasedev->migration;
192 
193         if (!migration) {
194             return false;
195         }
196 
197         if (vbasedev->pre_copy_dirty_page_tracking == ON_OFF_AUTO_OFF &&
198             (vfio_device_state_is_running(vbasedev) ||
199              vfio_device_state_is_precopy(vbasedev))) {
200             return false;
201         }
202     }
203     return true;
204 }
205 
206 bool vfio_devices_all_device_dirty_tracking(const VFIOContainerBase *bcontainer)
207 {
208     VFIODevice *vbasedev;
209 
210     QLIST_FOREACH(vbasedev, &bcontainer->device_list, container_next) {
211         if (!vbasedev->dirty_pages_supported) {
212             return false;
213         }
214     }
215 
216     return true;
217 }
218 
219 /*
220  * Check if all VFIO devices are running and migration is active, which is
221  * essentially equivalent to the migration being in pre-copy phase.
222  */
223 bool
224 vfio_devices_all_running_and_mig_active(const VFIOContainerBase *bcontainer)
225 {
226     VFIODevice *vbasedev;
227 
228     if (!migration_is_active(migrate_get_current())) {
229         return false;
230     }
231 
232     QLIST_FOREACH(vbasedev, &bcontainer->device_list, container_next) {
233         VFIOMigration *migration = vbasedev->migration;
234 
235         if (!migration) {
236             return false;
237         }
238 
239         if (vfio_device_state_is_running(vbasedev) ||
240             vfio_device_state_is_precopy(vbasedev)) {
241             continue;
242         } else {
243             return false;
244         }
245     }
246     return true;
247 }
248 
249 static bool vfio_listener_skipped_section(MemoryRegionSection *section)
250 {
251     return (!memory_region_is_ram(section->mr) &&
252             !memory_region_is_iommu(section->mr)) ||
253            memory_region_is_protected(section->mr) ||
254            /*
255             * Sizing an enabled 64-bit BAR can cause spurious mappings to
256             * addresses in the upper part of the 64-bit address space.  These
257             * are never accessed by the CPU and beyond the address width of
258             * some IOMMU hardware.  TODO: VFIO should tell us the IOMMU width.
259             */
260            section->offset_within_address_space & (1ULL << 63);
261 }
262 
263 /* Called with rcu_read_lock held.  */
264 static bool vfio_get_xlat_addr(IOMMUTLBEntry *iotlb, void **vaddr,
265                                ram_addr_t *ram_addr, bool *read_only)
266 {
267     bool ret, mr_has_discard_manager;
268 
269     ret = memory_get_xlat_addr(iotlb, vaddr, ram_addr, read_only,
270                                &mr_has_discard_manager);
271     if (ret && mr_has_discard_manager) {
272         /*
273          * Malicious VMs might trigger discarding of IOMMU-mapped memory. The
274          * pages will remain pinned inside vfio until unmapped, resulting in a
275          * higher memory consumption than expected. If memory would get
276          * populated again later, there would be an inconsistency between pages
277          * pinned by vfio and pages seen by QEMU. This is the case until
278          * unmapped from the IOMMU (e.g., during device reset).
279          *
280          * With malicious guests, we really only care about pinning more memory
281          * than expected. RLIMIT_MEMLOCK set for the user/process can never be
282          * exceeded and can be used to mitigate this problem.
283          */
284         warn_report_once("Using vfio with vIOMMUs and coordinated discarding of"
285                          " RAM (e.g., virtio-mem) works, however, malicious"
286                          " guests can trigger pinning of more memory than"
287                          " intended via an IOMMU. It's possible to mitigate "
288                          " by setting/adjusting RLIMIT_MEMLOCK.");
289     }
290     return ret;
291 }
292 
293 static void vfio_iommu_map_notify(IOMMUNotifier *n, IOMMUTLBEntry *iotlb)
294 {
295     VFIOGuestIOMMU *giommu = container_of(n, VFIOGuestIOMMU, n);
296     VFIOContainerBase *bcontainer = giommu->bcontainer;
297     hwaddr iova = iotlb->iova + giommu->iommu_offset;
298     void *vaddr;
299     int ret;
300 
301     trace_vfio_iommu_map_notify(iotlb->perm == IOMMU_NONE ? "UNMAP" : "MAP",
302                                 iova, iova + iotlb->addr_mask);
303 
304     if (iotlb->target_as != &address_space_memory) {
305         error_report("Wrong target AS \"%s\", only system memory is allowed",
306                      iotlb->target_as->name ? iotlb->target_as->name : "none");
307         vfio_set_migration_error(-EINVAL);
308         return;
309     }
310 
311     rcu_read_lock();
312 
313     if ((iotlb->perm & IOMMU_RW) != IOMMU_NONE) {
314         bool read_only;
315 
316         if (!vfio_get_xlat_addr(iotlb, &vaddr, NULL, &read_only)) {
317             goto out;
318         }
319         /*
320          * vaddr is only valid until rcu_read_unlock(). But after
321          * vfio_dma_map has set up the mapping the pages will be
322          * pinned by the kernel. This makes sure that the RAM backend
323          * of vaddr will always be there, even if the memory object is
324          * destroyed and its backing memory munmap-ed.
325          */
326         ret = vfio_container_dma_map(bcontainer, iova,
327                                      iotlb->addr_mask + 1, vaddr,
328                                      read_only);
329         if (ret) {
330             error_report("vfio_container_dma_map(%p, 0x%"HWADDR_PRIx", "
331                          "0x%"HWADDR_PRIx", %p) = %d (%s)",
332                          bcontainer, iova,
333                          iotlb->addr_mask + 1, vaddr, ret, strerror(-ret));
334         }
335     } else {
336         ret = vfio_container_dma_unmap(bcontainer, iova,
337                                        iotlb->addr_mask + 1, iotlb);
338         if (ret) {
339             error_report("vfio_container_dma_unmap(%p, 0x%"HWADDR_PRIx", "
340                          "0x%"HWADDR_PRIx") = %d (%s)",
341                          bcontainer, iova,
342                          iotlb->addr_mask + 1, ret, strerror(-ret));
343             vfio_set_migration_error(ret);
344         }
345     }
346 out:
347     rcu_read_unlock();
348 }
349 
350 static void vfio_ram_discard_notify_discard(RamDiscardListener *rdl,
351                                             MemoryRegionSection *section)
352 {
353     VFIORamDiscardListener *vrdl = container_of(rdl, VFIORamDiscardListener,
354                                                 listener);
355     VFIOContainerBase *bcontainer = vrdl->bcontainer;
356     const hwaddr size = int128_get64(section->size);
357     const hwaddr iova = section->offset_within_address_space;
358     int ret;
359 
360     /* Unmap with a single call. */
361     ret = vfio_container_dma_unmap(bcontainer, iova, size , NULL);
362     if (ret) {
363         error_report("%s: vfio_container_dma_unmap() failed: %s", __func__,
364                      strerror(-ret));
365     }
366 }
367 
368 static int vfio_ram_discard_notify_populate(RamDiscardListener *rdl,
369                                             MemoryRegionSection *section)
370 {
371     VFIORamDiscardListener *vrdl = container_of(rdl, VFIORamDiscardListener,
372                                                 listener);
373     VFIOContainerBase *bcontainer = vrdl->bcontainer;
374     const hwaddr end = section->offset_within_region +
375                        int128_get64(section->size);
376     hwaddr start, next, iova;
377     void *vaddr;
378     int ret;
379 
380     /*
381      * Map in (aligned within memory region) minimum granularity, so we can
382      * unmap in minimum granularity later.
383      */
384     for (start = section->offset_within_region; start < end; start = next) {
385         next = ROUND_UP(start + 1, vrdl->granularity);
386         next = MIN(next, end);
387 
388         iova = start - section->offset_within_region +
389                section->offset_within_address_space;
390         vaddr = memory_region_get_ram_ptr(section->mr) + start;
391 
392         ret = vfio_container_dma_map(bcontainer, iova, next - start,
393                                      vaddr, section->readonly);
394         if (ret) {
395             /* Rollback */
396             vfio_ram_discard_notify_discard(rdl, section);
397             return ret;
398         }
399     }
400     return 0;
401 }
402 
403 static void vfio_register_ram_discard_listener(VFIOContainerBase *bcontainer,
404                                                MemoryRegionSection *section)
405 {
406     RamDiscardManager *rdm = memory_region_get_ram_discard_manager(section->mr);
407     VFIORamDiscardListener *vrdl;
408 
409     /* Ignore some corner cases not relevant in practice. */
410     g_assert(QEMU_IS_ALIGNED(section->offset_within_region, TARGET_PAGE_SIZE));
411     g_assert(QEMU_IS_ALIGNED(section->offset_within_address_space,
412                              TARGET_PAGE_SIZE));
413     g_assert(QEMU_IS_ALIGNED(int128_get64(section->size), TARGET_PAGE_SIZE));
414 
415     vrdl = g_new0(VFIORamDiscardListener, 1);
416     vrdl->bcontainer = bcontainer;
417     vrdl->mr = section->mr;
418     vrdl->offset_within_address_space = section->offset_within_address_space;
419     vrdl->size = int128_get64(section->size);
420     vrdl->granularity = ram_discard_manager_get_min_granularity(rdm,
421                                                                 section->mr);
422 
423     g_assert(vrdl->granularity && is_power_of_2(vrdl->granularity));
424     g_assert(bcontainer->pgsizes &&
425              vrdl->granularity >= 1ULL << ctz64(bcontainer->pgsizes));
426 
427     ram_discard_listener_init(&vrdl->listener,
428                               vfio_ram_discard_notify_populate,
429                               vfio_ram_discard_notify_discard, true);
430     ram_discard_manager_register_listener(rdm, &vrdl->listener, section);
431     QLIST_INSERT_HEAD(&bcontainer->vrdl_list, vrdl, next);
432 
433     /*
434      * Sanity-check if we have a theoretically problematic setup where we could
435      * exceed the maximum number of possible DMA mappings over time. We assume
436      * that each mapped section in the same address space as a RamDiscardManager
437      * section consumes exactly one DMA mapping, with the exception of
438      * RamDiscardManager sections; i.e., we don't expect to have gIOMMU sections
439      * in the same address space as RamDiscardManager sections.
440      *
441      * We assume that each section in the address space consumes one memslot.
442      * We take the number of KVM memory slots as a best guess for the maximum
443      * number of sections in the address space we could have over time,
444      * also consuming DMA mappings.
445      */
446     if (bcontainer->dma_max_mappings) {
447         unsigned int vrdl_count = 0, vrdl_mappings = 0, max_memslots = 512;
448 
449 #ifdef CONFIG_KVM
450         if (kvm_enabled()) {
451             max_memslots = kvm_get_max_memslots();
452         }
453 #endif
454 
455         QLIST_FOREACH(vrdl, &bcontainer->vrdl_list, next) {
456             hwaddr start, end;
457 
458             start = QEMU_ALIGN_DOWN(vrdl->offset_within_address_space,
459                                     vrdl->granularity);
460             end = ROUND_UP(vrdl->offset_within_address_space + vrdl->size,
461                            vrdl->granularity);
462             vrdl_mappings += (end - start) / vrdl->granularity;
463             vrdl_count++;
464         }
465 
466         if (vrdl_mappings + max_memslots - vrdl_count >
467             bcontainer->dma_max_mappings) {
468             warn_report("%s: possibly running out of DMA mappings. E.g., try"
469                         " increasing the 'block-size' of virtio-mem devies."
470                         " Maximum possible DMA mappings: %d, Maximum possible"
471                         " memslots: %d", __func__, bcontainer->dma_max_mappings,
472                         max_memslots);
473         }
474     }
475 }
476 
477 static void vfio_unregister_ram_discard_listener(VFIOContainerBase *bcontainer,
478                                                  MemoryRegionSection *section)
479 {
480     RamDiscardManager *rdm = memory_region_get_ram_discard_manager(section->mr);
481     VFIORamDiscardListener *vrdl = NULL;
482 
483     QLIST_FOREACH(vrdl, &bcontainer->vrdl_list, next) {
484         if (vrdl->mr == section->mr &&
485             vrdl->offset_within_address_space ==
486             section->offset_within_address_space) {
487             break;
488         }
489     }
490 
491     if (!vrdl) {
492         hw_error("vfio: Trying to unregister missing RAM discard listener");
493     }
494 
495     ram_discard_manager_unregister_listener(rdm, &vrdl->listener);
496     QLIST_REMOVE(vrdl, next);
497     g_free(vrdl);
498 }
499 
500 static bool vfio_known_safe_misalignment(MemoryRegionSection *section)
501 {
502     MemoryRegion *mr = section->mr;
503 
504     if (!TPM_IS_CRB(mr->owner)) {
505         return false;
506     }
507 
508     /* this is a known safe misaligned region, just trace for debug purpose */
509     trace_vfio_known_safe_misalignment(memory_region_name(mr),
510                                        section->offset_within_address_space,
511                                        section->offset_within_region,
512                                        qemu_real_host_page_size());
513     return true;
514 }
515 
516 static bool vfio_listener_valid_section(MemoryRegionSection *section,
517                                         const char *name)
518 {
519     if (vfio_listener_skipped_section(section)) {
520         trace_vfio_listener_region_skip(name,
521                 section->offset_within_address_space,
522                 section->offset_within_address_space +
523                 int128_get64(int128_sub(section->size, int128_one())));
524         return false;
525     }
526 
527     if (unlikely((section->offset_within_address_space &
528                   ~qemu_real_host_page_mask()) !=
529                  (section->offset_within_region & ~qemu_real_host_page_mask()))) {
530         if (!vfio_known_safe_misalignment(section)) {
531             error_report("%s received unaligned region %s iova=0x%"PRIx64
532                          " offset_within_region=0x%"PRIx64
533                          " qemu_real_host_page_size=0x%"PRIxPTR,
534                          __func__, memory_region_name(section->mr),
535                          section->offset_within_address_space,
536                          section->offset_within_region,
537                          qemu_real_host_page_size());
538         }
539         return false;
540     }
541 
542     return true;
543 }
544 
545 static bool vfio_get_section_iova_range(VFIOContainerBase *bcontainer,
546                                         MemoryRegionSection *section,
547                                         hwaddr *out_iova, hwaddr *out_end,
548                                         Int128 *out_llend)
549 {
550     Int128 llend;
551     hwaddr iova;
552 
553     iova = REAL_HOST_PAGE_ALIGN(section->offset_within_address_space);
554     llend = int128_make64(section->offset_within_address_space);
555     llend = int128_add(llend, section->size);
556     llend = int128_and(llend, int128_exts64(qemu_real_host_page_mask()));
557 
558     if (int128_ge(int128_make64(iova), llend)) {
559         return false;
560     }
561 
562     *out_iova = iova;
563     *out_end = int128_get64(int128_sub(llend, int128_one()));
564     if (out_llend) {
565         *out_llend = llend;
566     }
567     return true;
568 }
569 
570 static void vfio_listener_region_add(MemoryListener *listener,
571                                      MemoryRegionSection *section)
572 {
573     VFIOContainerBase *bcontainer = container_of(listener, VFIOContainerBase,
574                                                  listener);
575     hwaddr iova, end;
576     Int128 llend, llsize;
577     void *vaddr;
578     int ret;
579     Error *err = NULL;
580 
581     if (!vfio_listener_valid_section(section, "region_add")) {
582         return;
583     }
584 
585     if (!vfio_get_section_iova_range(bcontainer, section, &iova, &end,
586                                      &llend)) {
587         if (memory_region_is_ram_device(section->mr)) {
588             trace_vfio_listener_region_add_no_dma_map(
589                 memory_region_name(section->mr),
590                 section->offset_within_address_space,
591                 int128_getlo(section->size),
592                 qemu_real_host_page_size());
593         }
594         return;
595     }
596 
597     if (vfio_container_add_section_window(bcontainer, section, &err)) {
598         goto fail;
599     }
600 
601     memory_region_ref(section->mr);
602 
603     if (memory_region_is_iommu(section->mr)) {
604         VFIOGuestIOMMU *giommu;
605         IOMMUMemoryRegion *iommu_mr = IOMMU_MEMORY_REGION(section->mr);
606         int iommu_idx;
607 
608         trace_vfio_listener_region_add_iommu(iova, end);
609         /*
610          * FIXME: For VFIO iommu types which have KVM acceleration to
611          * avoid bouncing all map/unmaps through qemu this way, this
612          * would be the right place to wire that up (tell the KVM
613          * device emulation the VFIO iommu handles to use).
614          */
615         giommu = g_malloc0(sizeof(*giommu));
616         giommu->iommu_mr = iommu_mr;
617         giommu->iommu_offset = section->offset_within_address_space -
618                                section->offset_within_region;
619         giommu->bcontainer = bcontainer;
620         llend = int128_add(int128_make64(section->offset_within_region),
621                            section->size);
622         llend = int128_sub(llend, int128_one());
623         iommu_idx = memory_region_iommu_attrs_to_index(iommu_mr,
624                                                        MEMTXATTRS_UNSPECIFIED);
625         iommu_notifier_init(&giommu->n, vfio_iommu_map_notify,
626                             IOMMU_NOTIFIER_IOTLB_EVENTS,
627                             section->offset_within_region,
628                             int128_get64(llend),
629                             iommu_idx);
630 
631         ret = memory_region_iommu_set_page_size_mask(giommu->iommu_mr,
632                                                      bcontainer->pgsizes,
633                                                      &err);
634         if (ret) {
635             g_free(giommu);
636             goto fail;
637         }
638 
639         if (bcontainer->iova_ranges) {
640             ret = memory_region_iommu_set_iova_ranges(giommu->iommu_mr,
641                                                       bcontainer->iova_ranges,
642                                                       &err);
643             if (ret) {
644                 g_free(giommu);
645                 goto fail;
646             }
647         }
648 
649         ret = memory_region_register_iommu_notifier(section->mr, &giommu->n,
650                                                     &err);
651         if (ret) {
652             g_free(giommu);
653             goto fail;
654         }
655         QLIST_INSERT_HEAD(&bcontainer->giommu_list, giommu, giommu_next);
656         memory_region_iommu_replay(giommu->iommu_mr, &giommu->n);
657 
658         return;
659     }
660 
661     /* Here we assume that memory_region_is_ram(section->mr)==true */
662 
663     /*
664      * For RAM memory regions with a RamDiscardManager, we only want to map the
665      * actually populated parts - and update the mapping whenever we're notified
666      * about changes.
667      */
668     if (memory_region_has_ram_discard_manager(section->mr)) {
669         vfio_register_ram_discard_listener(bcontainer, section);
670         return;
671     }
672 
673     vaddr = memory_region_get_ram_ptr(section->mr) +
674             section->offset_within_region +
675             (iova - section->offset_within_address_space);
676 
677     trace_vfio_listener_region_add_ram(iova, end, vaddr);
678 
679     llsize = int128_sub(llend, int128_make64(iova));
680 
681     if (memory_region_is_ram_device(section->mr)) {
682         hwaddr pgmask = (1ULL << ctz64(bcontainer->pgsizes)) - 1;
683 
684         if ((iova & pgmask) || (int128_get64(llsize) & pgmask)) {
685             trace_vfio_listener_region_add_no_dma_map(
686                 memory_region_name(section->mr),
687                 section->offset_within_address_space,
688                 int128_getlo(section->size),
689                 pgmask + 1);
690             return;
691         }
692     }
693 
694     ret = vfio_container_dma_map(bcontainer, iova, int128_get64(llsize),
695                                  vaddr, section->readonly);
696     if (ret) {
697         error_setg(&err, "vfio_container_dma_map(%p, 0x%"HWADDR_PRIx", "
698                    "0x%"HWADDR_PRIx", %p) = %d (%s)",
699                    bcontainer, iova, int128_get64(llsize), vaddr, ret,
700                    strerror(-ret));
701         if (memory_region_is_ram_device(section->mr)) {
702             /* Allow unexpected mappings not to be fatal for RAM devices */
703             error_report_err(err);
704             return;
705         }
706         goto fail;
707     }
708 
709     return;
710 
711 fail:
712     if (memory_region_is_ram_device(section->mr)) {
713         error_reportf_err(err, "PCI p2p may not work: ");
714         return;
715     }
716     /*
717      * On the initfn path, store the first error in the container so we
718      * can gracefully fail.  Runtime, there's not much we can do other
719      * than throw a hardware error.
720      */
721     if (!bcontainer->initialized) {
722         if (!bcontainer->error) {
723             error_propagate_prepend(&bcontainer->error, err,
724                                     "Region %s: ",
725                                     memory_region_name(section->mr));
726         } else {
727             error_free(err);
728         }
729     } else {
730         error_report_err(err);
731         hw_error("vfio: DMA mapping failed, unable to continue");
732     }
733 }
734 
735 static void vfio_listener_region_del(MemoryListener *listener,
736                                      MemoryRegionSection *section)
737 {
738     VFIOContainerBase *bcontainer = container_of(listener, VFIOContainerBase,
739                                                  listener);
740     hwaddr iova, end;
741     Int128 llend, llsize;
742     int ret;
743     bool try_unmap = true;
744 
745     if (!vfio_listener_valid_section(section, "region_del")) {
746         return;
747     }
748 
749     if (memory_region_is_iommu(section->mr)) {
750         VFIOGuestIOMMU *giommu;
751 
752         QLIST_FOREACH(giommu, &bcontainer->giommu_list, giommu_next) {
753             if (MEMORY_REGION(giommu->iommu_mr) == section->mr &&
754                 giommu->n.start == section->offset_within_region) {
755                 memory_region_unregister_iommu_notifier(section->mr,
756                                                         &giommu->n);
757                 QLIST_REMOVE(giommu, giommu_next);
758                 g_free(giommu);
759                 break;
760             }
761         }
762 
763         /*
764          * FIXME: We assume the one big unmap below is adequate to
765          * remove any individual page mappings in the IOMMU which
766          * might have been copied into VFIO. This works for a page table
767          * based IOMMU where a big unmap flattens a large range of IO-PTEs.
768          * That may not be true for all IOMMU types.
769          */
770     }
771 
772     if (!vfio_get_section_iova_range(bcontainer, section, &iova, &end,
773                                      &llend)) {
774         return;
775     }
776 
777     llsize = int128_sub(llend, int128_make64(iova));
778 
779     trace_vfio_listener_region_del(iova, end);
780 
781     if (memory_region_is_ram_device(section->mr)) {
782         hwaddr pgmask;
783 
784         pgmask = (1ULL << ctz64(bcontainer->pgsizes)) - 1;
785         try_unmap = !((iova & pgmask) || (int128_get64(llsize) & pgmask));
786     } else if (memory_region_has_ram_discard_manager(section->mr)) {
787         vfio_unregister_ram_discard_listener(bcontainer, section);
788         /* Unregistering will trigger an unmap. */
789         try_unmap = false;
790     }
791 
792     if (try_unmap) {
793         if (int128_eq(llsize, int128_2_64())) {
794             /* The unmap ioctl doesn't accept a full 64-bit span. */
795             llsize = int128_rshift(llsize, 1);
796             ret = vfio_container_dma_unmap(bcontainer, iova,
797                                            int128_get64(llsize), NULL);
798             if (ret) {
799                 error_report("vfio_container_dma_unmap(%p, 0x%"HWADDR_PRIx", "
800                              "0x%"HWADDR_PRIx") = %d (%s)",
801                              bcontainer, iova, int128_get64(llsize), ret,
802                              strerror(-ret));
803             }
804             iova += int128_get64(llsize);
805         }
806         ret = vfio_container_dma_unmap(bcontainer, iova,
807                                        int128_get64(llsize), NULL);
808         if (ret) {
809             error_report("vfio_container_dma_unmap(%p, 0x%"HWADDR_PRIx", "
810                          "0x%"HWADDR_PRIx") = %d (%s)",
811                          bcontainer, iova, int128_get64(llsize), ret,
812                          strerror(-ret));
813         }
814     }
815 
816     memory_region_unref(section->mr);
817 
818     vfio_container_del_section_window(bcontainer, section);
819 }
820 
821 typedef struct VFIODirtyRanges {
822     hwaddr min32;
823     hwaddr max32;
824     hwaddr min64;
825     hwaddr max64;
826     hwaddr minpci64;
827     hwaddr maxpci64;
828 } VFIODirtyRanges;
829 
830 typedef struct VFIODirtyRangesListener {
831     VFIOContainerBase *bcontainer;
832     VFIODirtyRanges ranges;
833     MemoryListener listener;
834 } VFIODirtyRangesListener;
835 
836 static bool vfio_section_is_vfio_pci(MemoryRegionSection *section,
837                                      VFIOContainerBase *bcontainer)
838 {
839     VFIOPCIDevice *pcidev;
840     VFIODevice *vbasedev;
841     Object *owner;
842 
843     owner = memory_region_owner(section->mr);
844 
845     QLIST_FOREACH(vbasedev, &bcontainer->device_list, container_next) {
846         if (vbasedev->type != VFIO_DEVICE_TYPE_PCI) {
847             continue;
848         }
849         pcidev = container_of(vbasedev, VFIOPCIDevice, vbasedev);
850         if (OBJECT(pcidev) == owner) {
851             return true;
852         }
853     }
854 
855     return false;
856 }
857 
858 static void vfio_dirty_tracking_update(MemoryListener *listener,
859                                        MemoryRegionSection *section)
860 {
861     VFIODirtyRangesListener *dirty = container_of(listener,
862                                                   VFIODirtyRangesListener,
863                                                   listener);
864     VFIODirtyRanges *range = &dirty->ranges;
865     hwaddr iova, end, *min, *max;
866 
867     if (!vfio_listener_valid_section(section, "tracking_update") ||
868         !vfio_get_section_iova_range(dirty->bcontainer, section,
869                                      &iova, &end, NULL)) {
870         return;
871     }
872 
873     /*
874      * The address space passed to the dirty tracker is reduced to three ranges:
875      * one for 32-bit DMA ranges, one for 64-bit DMA ranges and one for the
876      * PCI 64-bit hole.
877      *
878      * The underlying reports of dirty will query a sub-interval of each of
879      * these ranges.
880      *
881      * The purpose of the three range handling is to handle known cases of big
882      * holes in the address space, like the x86 AMD 1T hole, and firmware (like
883      * OVMF) which may relocate the pci-hole64 to the end of the address space.
884      * The latter would otherwise generate large ranges for tracking, stressing
885      * the limits of supported hardware. The pci-hole32 will always be below 4G
886      * (overlapping or not) so it doesn't need special handling and is part of
887      * the 32-bit range.
888      *
889      * The alternative would be an IOVATree but that has a much bigger runtime
890      * overhead and unnecessary complexity.
891      */
892     if (vfio_section_is_vfio_pci(section, dirty->bcontainer) &&
893         iova >= UINT32_MAX) {
894         min = &range->minpci64;
895         max = &range->maxpci64;
896     } else {
897         min = (end <= UINT32_MAX) ? &range->min32 : &range->min64;
898         max = (end <= UINT32_MAX) ? &range->max32 : &range->max64;
899     }
900     if (*min > iova) {
901         *min = iova;
902     }
903     if (*max < end) {
904         *max = end;
905     }
906 
907     trace_vfio_device_dirty_tracking_update(iova, end, *min, *max);
908     return;
909 }
910 
911 static const MemoryListener vfio_dirty_tracking_listener = {
912     .name = "vfio-tracking",
913     .region_add = vfio_dirty_tracking_update,
914 };
915 
916 static void vfio_dirty_tracking_init(VFIOContainerBase *bcontainer,
917                                      VFIODirtyRanges *ranges)
918 {
919     VFIODirtyRangesListener dirty;
920 
921     memset(&dirty, 0, sizeof(dirty));
922     dirty.ranges.min32 = UINT32_MAX;
923     dirty.ranges.min64 = UINT64_MAX;
924     dirty.ranges.minpci64 = UINT64_MAX;
925     dirty.listener = vfio_dirty_tracking_listener;
926     dirty.bcontainer = bcontainer;
927 
928     memory_listener_register(&dirty.listener,
929                              bcontainer->space->as);
930 
931     *ranges = dirty.ranges;
932 
933     /*
934      * The memory listener is synchronous, and used to calculate the range
935      * to dirty tracking. Unregister it after we are done as we are not
936      * interested in any follow-up updates.
937      */
938     memory_listener_unregister(&dirty.listener);
939 }
940 
941 static void vfio_devices_dma_logging_stop(VFIOContainerBase *bcontainer)
942 {
943     uint64_t buf[DIV_ROUND_UP(sizeof(struct vfio_device_feature),
944                               sizeof(uint64_t))] = {};
945     struct vfio_device_feature *feature = (struct vfio_device_feature *)buf;
946     VFIODevice *vbasedev;
947 
948     feature->argsz = sizeof(buf);
949     feature->flags = VFIO_DEVICE_FEATURE_SET |
950                      VFIO_DEVICE_FEATURE_DMA_LOGGING_STOP;
951 
952     QLIST_FOREACH(vbasedev, &bcontainer->device_list, container_next) {
953         if (!vbasedev->dirty_tracking) {
954             continue;
955         }
956 
957         if (ioctl(vbasedev->fd, VFIO_DEVICE_FEATURE, feature)) {
958             warn_report("%s: Failed to stop DMA logging, err %d (%s)",
959                         vbasedev->name, -errno, strerror(errno));
960         }
961         vbasedev->dirty_tracking = false;
962     }
963 }
964 
965 static struct vfio_device_feature *
966 vfio_device_feature_dma_logging_start_create(VFIOContainerBase *bcontainer,
967                                              VFIODirtyRanges *tracking)
968 {
969     struct vfio_device_feature *feature;
970     size_t feature_size;
971     struct vfio_device_feature_dma_logging_control *control;
972     struct vfio_device_feature_dma_logging_range *ranges;
973 
974     feature_size = sizeof(struct vfio_device_feature) +
975                    sizeof(struct vfio_device_feature_dma_logging_control);
976     feature = g_try_malloc0(feature_size);
977     if (!feature) {
978         errno = ENOMEM;
979         return NULL;
980     }
981     feature->argsz = feature_size;
982     feature->flags = VFIO_DEVICE_FEATURE_SET |
983                      VFIO_DEVICE_FEATURE_DMA_LOGGING_START;
984 
985     control = (struct vfio_device_feature_dma_logging_control *)feature->data;
986     control->page_size = qemu_real_host_page_size();
987 
988     /*
989      * DMA logging uAPI guarantees to support at least a number of ranges that
990      * fits into a single host kernel base page.
991      */
992     control->num_ranges = !!tracking->max32 + !!tracking->max64 +
993         !!tracking->maxpci64;
994     ranges = g_try_new0(struct vfio_device_feature_dma_logging_range,
995                         control->num_ranges);
996     if (!ranges) {
997         g_free(feature);
998         errno = ENOMEM;
999 
1000         return NULL;
1001     }
1002 
1003     control->ranges = (uintptr_t)ranges;
1004     if (tracking->max32) {
1005         ranges->iova = tracking->min32;
1006         ranges->length = (tracking->max32 - tracking->min32) + 1;
1007         ranges++;
1008     }
1009     if (tracking->max64) {
1010         ranges->iova = tracking->min64;
1011         ranges->length = (tracking->max64 - tracking->min64) + 1;
1012         ranges++;
1013     }
1014     if (tracking->maxpci64) {
1015         ranges->iova = tracking->minpci64;
1016         ranges->length = (tracking->maxpci64 - tracking->minpci64) + 1;
1017     }
1018 
1019     trace_vfio_device_dirty_tracking_start(control->num_ranges,
1020                                            tracking->min32, tracking->max32,
1021                                            tracking->min64, tracking->max64,
1022                                            tracking->minpci64, tracking->maxpci64);
1023 
1024     return feature;
1025 }
1026 
1027 static void vfio_device_feature_dma_logging_start_destroy(
1028     struct vfio_device_feature *feature)
1029 {
1030     struct vfio_device_feature_dma_logging_control *control =
1031         (struct vfio_device_feature_dma_logging_control *)feature->data;
1032     struct vfio_device_feature_dma_logging_range *ranges =
1033         (struct vfio_device_feature_dma_logging_range *)(uintptr_t)control->ranges;
1034 
1035     g_free(ranges);
1036     g_free(feature);
1037 }
1038 
1039 static int vfio_devices_dma_logging_start(VFIOContainerBase *bcontainer)
1040 {
1041     struct vfio_device_feature *feature;
1042     VFIODirtyRanges ranges;
1043     VFIODevice *vbasedev;
1044     int ret = 0;
1045 
1046     vfio_dirty_tracking_init(bcontainer, &ranges);
1047     feature = vfio_device_feature_dma_logging_start_create(bcontainer,
1048                                                            &ranges);
1049     if (!feature) {
1050         return -errno;
1051     }
1052 
1053     QLIST_FOREACH(vbasedev, &bcontainer->device_list, container_next) {
1054         if (vbasedev->dirty_tracking) {
1055             continue;
1056         }
1057 
1058         ret = ioctl(vbasedev->fd, VFIO_DEVICE_FEATURE, feature);
1059         if (ret) {
1060             ret = -errno;
1061             error_report("%s: Failed to start DMA logging, err %d (%s)",
1062                          vbasedev->name, ret, strerror(errno));
1063             goto out;
1064         }
1065         vbasedev->dirty_tracking = true;
1066     }
1067 
1068 out:
1069     if (ret) {
1070         vfio_devices_dma_logging_stop(bcontainer);
1071     }
1072 
1073     vfio_device_feature_dma_logging_start_destroy(feature);
1074 
1075     return ret;
1076 }
1077 
1078 static void vfio_listener_log_global_start(MemoryListener *listener)
1079 {
1080     VFIOContainerBase *bcontainer = container_of(listener, VFIOContainerBase,
1081                                                  listener);
1082     int ret;
1083 
1084     if (vfio_devices_all_device_dirty_tracking(bcontainer)) {
1085         ret = vfio_devices_dma_logging_start(bcontainer);
1086     } else {
1087         ret = vfio_container_set_dirty_page_tracking(bcontainer, true);
1088     }
1089 
1090     if (ret) {
1091         error_report("vfio: Could not start dirty page tracking, err: %d (%s)",
1092                      ret, strerror(-ret));
1093         vfio_set_migration_error(ret);
1094     }
1095 }
1096 
1097 static void vfio_listener_log_global_stop(MemoryListener *listener)
1098 {
1099     VFIOContainerBase *bcontainer = container_of(listener, VFIOContainerBase,
1100                                                  listener);
1101     int ret = 0;
1102 
1103     if (vfio_devices_all_device_dirty_tracking(bcontainer)) {
1104         vfio_devices_dma_logging_stop(bcontainer);
1105     } else {
1106         ret = vfio_container_set_dirty_page_tracking(bcontainer, false);
1107     }
1108 
1109     if (ret) {
1110         error_report("vfio: Could not stop dirty page tracking, err: %d (%s)",
1111                      ret, strerror(-ret));
1112         vfio_set_migration_error(ret);
1113     }
1114 }
1115 
1116 static int vfio_device_dma_logging_report(VFIODevice *vbasedev, hwaddr iova,
1117                                           hwaddr size, void *bitmap)
1118 {
1119     uint64_t buf[DIV_ROUND_UP(sizeof(struct vfio_device_feature) +
1120                         sizeof(struct vfio_device_feature_dma_logging_report),
1121                         sizeof(uint64_t))] = {};
1122     struct vfio_device_feature *feature = (struct vfio_device_feature *)buf;
1123     struct vfio_device_feature_dma_logging_report *report =
1124         (struct vfio_device_feature_dma_logging_report *)feature->data;
1125 
1126     report->iova = iova;
1127     report->length = size;
1128     report->page_size = qemu_real_host_page_size();
1129     report->bitmap = (uintptr_t)bitmap;
1130 
1131     feature->argsz = sizeof(buf);
1132     feature->flags = VFIO_DEVICE_FEATURE_GET |
1133                      VFIO_DEVICE_FEATURE_DMA_LOGGING_REPORT;
1134 
1135     if (ioctl(vbasedev->fd, VFIO_DEVICE_FEATURE, feature)) {
1136         return -errno;
1137     }
1138 
1139     return 0;
1140 }
1141 
1142 int vfio_devices_query_dirty_bitmap(const VFIOContainerBase *bcontainer,
1143                                     VFIOBitmap *vbmap, hwaddr iova,
1144                                     hwaddr size)
1145 {
1146     VFIODevice *vbasedev;
1147     int ret;
1148 
1149     QLIST_FOREACH(vbasedev, &bcontainer->device_list, container_next) {
1150         ret = vfio_device_dma_logging_report(vbasedev, iova, size,
1151                                              vbmap->bitmap);
1152         if (ret) {
1153             error_report("%s: Failed to get DMA logging report, iova: "
1154                          "0x%" HWADDR_PRIx ", size: 0x%" HWADDR_PRIx
1155                          ", err: %d (%s)",
1156                          vbasedev->name, iova, size, ret, strerror(-ret));
1157 
1158             return ret;
1159         }
1160     }
1161 
1162     return 0;
1163 }
1164 
1165 int vfio_get_dirty_bitmap(const VFIOContainerBase *bcontainer, uint64_t iova,
1166                           uint64_t size, ram_addr_t ram_addr)
1167 {
1168     bool all_device_dirty_tracking =
1169         vfio_devices_all_device_dirty_tracking(bcontainer);
1170     uint64_t dirty_pages;
1171     VFIOBitmap vbmap;
1172     int ret;
1173 
1174     if (!bcontainer->dirty_pages_supported && !all_device_dirty_tracking) {
1175         cpu_physical_memory_set_dirty_range(ram_addr, size,
1176                                             tcg_enabled() ? DIRTY_CLIENTS_ALL :
1177                                             DIRTY_CLIENTS_NOCODE);
1178         return 0;
1179     }
1180 
1181     ret = vfio_bitmap_alloc(&vbmap, size);
1182     if (ret) {
1183         return ret;
1184     }
1185 
1186     if (all_device_dirty_tracking) {
1187         ret = vfio_devices_query_dirty_bitmap(bcontainer, &vbmap, iova, size);
1188     } else {
1189         ret = vfio_container_query_dirty_bitmap(bcontainer, &vbmap, iova, size);
1190     }
1191 
1192     if (ret) {
1193         goto out;
1194     }
1195 
1196     dirty_pages = cpu_physical_memory_set_dirty_lebitmap(vbmap.bitmap, ram_addr,
1197                                                          vbmap.pages);
1198 
1199     trace_vfio_get_dirty_bitmap(iova, size, vbmap.size, ram_addr, dirty_pages);
1200 out:
1201     g_free(vbmap.bitmap);
1202 
1203     return ret;
1204 }
1205 
1206 typedef struct {
1207     IOMMUNotifier n;
1208     VFIOGuestIOMMU *giommu;
1209 } vfio_giommu_dirty_notifier;
1210 
1211 static void vfio_iommu_map_dirty_notify(IOMMUNotifier *n, IOMMUTLBEntry *iotlb)
1212 {
1213     vfio_giommu_dirty_notifier *gdn = container_of(n,
1214                                                 vfio_giommu_dirty_notifier, n);
1215     VFIOGuestIOMMU *giommu = gdn->giommu;
1216     VFIOContainerBase *bcontainer = giommu->bcontainer;
1217     hwaddr iova = iotlb->iova + giommu->iommu_offset;
1218     ram_addr_t translated_addr;
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)) {
1231         ret = vfio_get_dirty_bitmap(bcontainer, iova, iotlb->addr_mask + 1,
1232                                     translated_addr);
1233         if (ret) {
1234             error_report("vfio_iommu_map_dirty_notify(%p, 0x%"HWADDR_PRIx", "
1235                          "0x%"HWADDR_PRIx") = %d (%s)",
1236                          bcontainer, iova, iotlb->addr_mask + 1, ret,
1237                          strerror(-ret));
1238         }
1239     }
1240     rcu_read_unlock();
1241 
1242 out:
1243     if (ret) {
1244         vfio_set_migration_error(ret);
1245     }
1246 }
1247 
1248 static int vfio_ram_discard_get_dirty_bitmap(MemoryRegionSection *section,
1249                                              void *opaque)
1250 {
1251     const hwaddr size = int128_get64(section->size);
1252     const hwaddr iova = section->offset_within_address_space;
1253     const ram_addr_t ram_addr = memory_region_get_ram_addr(section->mr) +
1254                                 section->offset_within_region;
1255     VFIORamDiscardListener *vrdl = opaque;
1256 
1257     /*
1258      * Sync the whole mapped region (spanning multiple individual mappings)
1259      * in one go.
1260      */
1261     return vfio_get_dirty_bitmap(vrdl->bcontainer, iova, size, ram_addr);
1262 }
1263 
1264 static int
1265 vfio_sync_ram_discard_listener_dirty_bitmap(VFIOContainerBase *bcontainer,
1266                                             MemoryRegionSection *section)
1267 {
1268     RamDiscardManager *rdm = memory_region_get_ram_discard_manager(section->mr);
1269     VFIORamDiscardListener *vrdl = NULL;
1270 
1271     QLIST_FOREACH(vrdl, &bcontainer->vrdl_list, next) {
1272         if (vrdl->mr == section->mr &&
1273             vrdl->offset_within_address_space ==
1274             section->offset_within_address_space) {
1275             break;
1276         }
1277     }
1278 
1279     if (!vrdl) {
1280         hw_error("vfio: Trying to sync missing RAM discard listener");
1281     }
1282 
1283     /*
1284      * We only want/can synchronize the bitmap for actually mapped parts -
1285      * which correspond to populated parts. Replay all populated parts.
1286      */
1287     return ram_discard_manager_replay_populated(rdm, section,
1288                                               vfio_ram_discard_get_dirty_bitmap,
1289                                                 &vrdl);
1290 }
1291 
1292 static int vfio_sync_dirty_bitmap(VFIOContainerBase *bcontainer,
1293                                   MemoryRegionSection *section)
1294 {
1295     ram_addr_t ram_addr;
1296 
1297     if (memory_region_is_iommu(section->mr)) {
1298         VFIOGuestIOMMU *giommu;
1299 
1300         QLIST_FOREACH(giommu, &bcontainer->giommu_list, giommu_next) {
1301             if (MEMORY_REGION(giommu->iommu_mr) == section->mr &&
1302                 giommu->n.start == section->offset_within_region) {
1303                 Int128 llend;
1304                 vfio_giommu_dirty_notifier gdn = { .giommu = giommu };
1305                 int idx = memory_region_iommu_attrs_to_index(giommu->iommu_mr,
1306                                                        MEMTXATTRS_UNSPECIFIED);
1307 
1308                 llend = int128_add(int128_make64(section->offset_within_region),
1309                                    section->size);
1310                 llend = int128_sub(llend, int128_one());
1311 
1312                 iommu_notifier_init(&gdn.n,
1313                                     vfio_iommu_map_dirty_notify,
1314                                     IOMMU_NOTIFIER_MAP,
1315                                     section->offset_within_region,
1316                                     int128_get64(llend),
1317                                     idx);
1318                 memory_region_iommu_replay(giommu->iommu_mr, &gdn.n);
1319                 break;
1320             }
1321         }
1322         return 0;
1323     } else if (memory_region_has_ram_discard_manager(section->mr)) {
1324         return vfio_sync_ram_discard_listener_dirty_bitmap(bcontainer, section);
1325     }
1326 
1327     ram_addr = memory_region_get_ram_addr(section->mr) +
1328                section->offset_within_region;
1329 
1330     return vfio_get_dirty_bitmap(bcontainer,
1331                    REAL_HOST_PAGE_ALIGN(section->offset_within_address_space),
1332                    int128_get64(section->size), ram_addr);
1333 }
1334 
1335 static void vfio_listener_log_sync(MemoryListener *listener,
1336         MemoryRegionSection *section)
1337 {
1338     VFIOContainerBase *bcontainer = container_of(listener, VFIOContainerBase,
1339                                                  listener);
1340     int ret;
1341 
1342     if (vfio_listener_skipped_section(section)) {
1343         return;
1344     }
1345 
1346     if (vfio_devices_all_dirty_tracking(bcontainer)) {
1347         ret = vfio_sync_dirty_bitmap(bcontainer, section);
1348         if (ret) {
1349             error_report("vfio: Failed to sync dirty bitmap, err: %d (%s)", ret,
1350                          strerror(-ret));
1351             vfio_set_migration_error(ret);
1352         }
1353     }
1354 }
1355 
1356 const MemoryListener vfio_memory_listener = {
1357     .name = "vfio",
1358     .region_add = vfio_listener_region_add,
1359     .region_del = vfio_listener_region_del,
1360     .log_global_start = vfio_listener_log_global_start,
1361     .log_global_stop = vfio_listener_log_global_stop,
1362     .log_sync = vfio_listener_log_sync,
1363 };
1364 
1365 void vfio_reset_handler(void *opaque)
1366 {
1367     VFIODevice *vbasedev;
1368 
1369     QLIST_FOREACH(vbasedev, &vfio_device_list, global_next) {
1370         if (vbasedev->dev->realized) {
1371             vbasedev->ops->vfio_compute_needs_reset(vbasedev);
1372         }
1373     }
1374 
1375     QLIST_FOREACH(vbasedev, &vfio_device_list, global_next) {
1376         if (vbasedev->dev->realized && vbasedev->needs_reset) {
1377             vbasedev->ops->vfio_hot_reset_multi(vbasedev);
1378         }
1379     }
1380 }
1381 
1382 int vfio_kvm_device_add_fd(int fd, Error **errp)
1383 {
1384 #ifdef CONFIG_KVM
1385     struct kvm_device_attr attr = {
1386         .group = KVM_DEV_VFIO_FILE,
1387         .attr = KVM_DEV_VFIO_FILE_ADD,
1388         .addr = (uint64_t)(unsigned long)&fd,
1389     };
1390 
1391     if (!kvm_enabled()) {
1392         return 0;
1393     }
1394 
1395     if (vfio_kvm_device_fd < 0) {
1396         struct kvm_create_device cd = {
1397             .type = KVM_DEV_TYPE_VFIO,
1398         };
1399 
1400         if (kvm_vm_ioctl(kvm_state, KVM_CREATE_DEVICE, &cd)) {
1401             error_setg_errno(errp, errno, "Failed to create KVM VFIO device");
1402             return -errno;
1403         }
1404 
1405         vfio_kvm_device_fd = cd.fd;
1406     }
1407 
1408     if (ioctl(vfio_kvm_device_fd, KVM_SET_DEVICE_ATTR, &attr)) {
1409         error_setg_errno(errp, errno, "Failed to add fd %d to KVM VFIO device",
1410                          fd);
1411         return -errno;
1412     }
1413 #endif
1414     return 0;
1415 }
1416 
1417 int vfio_kvm_device_del_fd(int fd, Error **errp)
1418 {
1419 #ifdef CONFIG_KVM
1420     struct kvm_device_attr attr = {
1421         .group = KVM_DEV_VFIO_FILE,
1422         .attr = KVM_DEV_VFIO_FILE_DEL,
1423         .addr = (uint64_t)(unsigned long)&fd,
1424     };
1425 
1426     if (vfio_kvm_device_fd < 0) {
1427         error_setg(errp, "KVM VFIO device isn't created yet");
1428         return -EINVAL;
1429     }
1430 
1431     if (ioctl(vfio_kvm_device_fd, KVM_SET_DEVICE_ATTR, &attr)) {
1432         error_setg_errno(errp, errno,
1433                          "Failed to remove fd %d from KVM VFIO device", fd);
1434         return -errno;
1435     }
1436 #endif
1437     return 0;
1438 }
1439 
1440 VFIOAddressSpace *vfio_get_address_space(AddressSpace *as)
1441 {
1442     VFIOAddressSpace *space;
1443 
1444     QLIST_FOREACH(space, &vfio_address_spaces, list) {
1445         if (space->as == as) {
1446             return space;
1447         }
1448     }
1449 
1450     /* No suitable VFIOAddressSpace, create a new one */
1451     space = g_malloc0(sizeof(*space));
1452     space->as = as;
1453     QLIST_INIT(&space->containers);
1454 
1455     if (QLIST_EMPTY(&vfio_address_spaces)) {
1456         qemu_register_reset(vfio_reset_handler, NULL);
1457     }
1458 
1459     QLIST_INSERT_HEAD(&vfio_address_spaces, space, list);
1460 
1461     return space;
1462 }
1463 
1464 void vfio_put_address_space(VFIOAddressSpace *space)
1465 {
1466     if (!QLIST_EMPTY(&space->containers)) {
1467         return;
1468     }
1469 
1470     QLIST_REMOVE(space, list);
1471     g_free(space);
1472 
1473     if (QLIST_EMPTY(&vfio_address_spaces)) {
1474         qemu_unregister_reset(vfio_reset_handler, NULL);
1475     }
1476 }
1477 
1478 struct vfio_device_info *vfio_get_device_info(int fd)
1479 {
1480     struct vfio_device_info *info;
1481     uint32_t argsz = sizeof(*info);
1482 
1483     info = g_malloc0(argsz);
1484 
1485 retry:
1486     info->argsz = argsz;
1487 
1488     if (ioctl(fd, VFIO_DEVICE_GET_INFO, info)) {
1489         g_free(info);
1490         return NULL;
1491     }
1492 
1493     if (info->argsz > argsz) {
1494         argsz = info->argsz;
1495         info = g_realloc(info, argsz);
1496         goto retry;
1497     }
1498 
1499     return info;
1500 }
1501 
1502 int vfio_attach_device(char *name, VFIODevice *vbasedev,
1503                        AddressSpace *as, Error **errp)
1504 {
1505     const VFIOIOMMUClass *ops =
1506         VFIO_IOMMU_CLASS(object_class_by_name(TYPE_VFIO_IOMMU_LEGACY));
1507 
1508     if (vbasedev->iommufd) {
1509         ops = VFIO_IOMMU_CLASS(object_class_by_name(TYPE_VFIO_IOMMU_IOMMUFD));
1510     }
1511 
1512     assert(ops);
1513 
1514     return ops->attach_device(name, vbasedev, as, errp);
1515 }
1516 
1517 void vfio_detach_device(VFIODevice *vbasedev)
1518 {
1519     if (!vbasedev->bcontainer) {
1520         return;
1521     }
1522     vbasedev->bcontainer->ops->detach_device(vbasedev);
1523 }
1524