xref: /openbmc/qemu/hw/vfio/common.c (revision b2580720)
1 /*
2  * generic functions used by VFIO devices
3  *
4  * Copyright Red Hat, Inc. 2012
5  *
6  * Authors:
7  *  Alex Williamson <alex.williamson@redhat.com>
8  *
9  * This work is licensed under the terms of the GNU GPL, version 2.  See
10  * the COPYING file in the top-level directory.
11  *
12  * Based on qemu-kvm device-assignment:
13  *  Adapted for KVM by Qumranet.
14  *  Copyright (c) 2007, Neocleus, Alex Novik (alex@neocleus.com)
15  *  Copyright (c) 2007, Neocleus, Guy Zana (guy@neocleus.com)
16  *  Copyright (C) 2008, Qumranet, Amit Shah (amit.shah@qumranet.com)
17  *  Copyright (C) 2008, Red Hat, Amit Shah (amit.shah@redhat.com)
18  *  Copyright (C) 2008, IBM, Muli Ben-Yehuda (muli@il.ibm.com)
19  */
20 
21 #include "qemu/osdep.h"
22 #include <sys/ioctl.h>
23 #ifdef CONFIG_KVM
24 #include <linux/kvm.h>
25 #endif
26 #include <linux/vfio.h>
27 
28 #include "hw/vfio/vfio-common.h"
29 #include "hw/vfio/pci.h"
30 #include "exec/address-spaces.h"
31 #include "exec/memory.h"
32 #include "exec/ram_addr.h"
33 #include "hw/hw.h"
34 #include "qemu/error-report.h"
35 #include "qemu/main-loop.h"
36 #include "qemu/range.h"
37 #include "sysemu/kvm.h"
38 #include "sysemu/reset.h"
39 #include "sysemu/runstate.h"
40 #include "trace.h"
41 #include "qapi/error.h"
42 #include "migration/misc.h"
43 #include "migration/blocker.h"
44 #include "migration/qemu-file.h"
45 #include "sysemu/tpm.h"
46 
47 VFIODeviceList vfio_device_list =
48     QLIST_HEAD_INITIALIZER(vfio_device_list);
49 static QLIST_HEAD(, VFIOAddressSpace) vfio_address_spaces =
50     QLIST_HEAD_INITIALIZER(vfio_address_spaces);
51 
52 #ifdef CONFIG_KVM
53 /*
54  * We have a single VFIO pseudo device per KVM VM.  Once created it lives
55  * for the life of the VM.  Closing the file descriptor only drops our
56  * reference to it and the device's reference to kvm.  Therefore once
57  * initialized, this file descriptor is only released on QEMU exit and
58  * we'll re-use it should another vfio device be attached before then.
59  */
60 int vfio_kvm_device_fd = -1;
61 #endif
62 
63 /*
64  * Device state interfaces
65  */
66 
67 bool vfio_mig_active(void)
68 {
69     VFIODevice *vbasedev;
70 
71     if (QLIST_EMPTY(&vfio_device_list)) {
72         return false;
73     }
74 
75     QLIST_FOREACH(vbasedev, &vfio_device_list, global_next) {
76         if (vbasedev->migration_blocker) {
77             return false;
78         }
79     }
80     return true;
81 }
82 
83 static Error *multiple_devices_migration_blocker;
84 
85 /*
86  * Multiple devices migration is allowed only if all devices support P2P
87  * migration. Single device migration is allowed regardless of P2P migration
88  * support.
89  */
90 static bool vfio_multiple_devices_migration_is_supported(void)
91 {
92     VFIODevice *vbasedev;
93     unsigned int device_num = 0;
94     bool all_support_p2p = true;
95 
96     QLIST_FOREACH(vbasedev, &vfio_device_list, global_next) {
97         if (vbasedev->migration) {
98             device_num++;
99 
100             if (!(vbasedev->migration->mig_flags & VFIO_MIGRATION_P2P)) {
101                 all_support_p2p = false;
102             }
103         }
104     }
105 
106     return all_support_p2p || device_num <= 1;
107 }
108 
109 int vfio_block_multiple_devices_migration(VFIODevice *vbasedev, Error **errp)
110 {
111     int ret;
112 
113     if (vfio_multiple_devices_migration_is_supported()) {
114         return 0;
115     }
116 
117     if (vbasedev->enable_migration == ON_OFF_AUTO_ON) {
118         error_setg(errp, "Multiple VFIO devices migration is supported only if "
119                          "all of them support P2P migration");
120         return -EINVAL;
121     }
122 
123     if (multiple_devices_migration_blocker) {
124         return 0;
125     }
126 
127     error_setg(&multiple_devices_migration_blocker,
128                "Multiple VFIO devices migration is supported only if all of "
129                "them support P2P migration");
130     ret = migrate_add_blocker_normal(&multiple_devices_migration_blocker, errp);
131 
132     return ret;
133 }
134 
135 void vfio_unblock_multiple_devices_migration(void)
136 {
137     if (!multiple_devices_migration_blocker ||
138         !vfio_multiple_devices_migration_is_supported()) {
139         return;
140     }
141 
142     migrate_del_blocker(&multiple_devices_migration_blocker);
143 }
144 
145 bool vfio_viommu_preset(VFIODevice *vbasedev)
146 {
147     return vbasedev->bcontainer->space->as != &address_space_memory;
148 }
149 
150 static void vfio_set_migration_error(int ret)
151 {
152     if (migration_is_setup_or_active()) {
153         migration_file_set_error(ret, NULL);
154     }
155 }
156 
157 bool vfio_device_state_is_running(VFIODevice *vbasedev)
158 {
159     VFIOMigration *migration = vbasedev->migration;
160 
161     return migration->device_state == VFIO_DEVICE_STATE_RUNNING ||
162            migration->device_state == VFIO_DEVICE_STATE_RUNNING_P2P;
163 }
164 
165 bool vfio_device_state_is_precopy(VFIODevice *vbasedev)
166 {
167     VFIOMigration *migration = vbasedev->migration;
168 
169     return migration->device_state == VFIO_DEVICE_STATE_PRE_COPY ||
170            migration->device_state == VFIO_DEVICE_STATE_PRE_COPY_P2P;
171 }
172 
173 static bool vfio_devices_all_dirty_tracking(VFIOContainerBase *bcontainer)
174 {
175     VFIODevice *vbasedev;
176 
177     if (!migration_is_active() && !migration_is_device()) {
178         return false;
179     }
180 
181     QLIST_FOREACH(vbasedev, &bcontainer->device_list, container_next) {
182         VFIOMigration *migration = vbasedev->migration;
183 
184         if (!migration) {
185             return false;
186         }
187 
188         if (vbasedev->pre_copy_dirty_page_tracking == ON_OFF_AUTO_OFF &&
189             (vfio_device_state_is_running(vbasedev) ||
190              vfio_device_state_is_precopy(vbasedev))) {
191             return false;
192         }
193     }
194     return true;
195 }
196 
197 bool vfio_devices_all_device_dirty_tracking(const VFIOContainerBase *bcontainer)
198 {
199     VFIODevice *vbasedev;
200 
201     QLIST_FOREACH(vbasedev, &bcontainer->device_list, container_next) {
202         if (!vbasedev->dirty_pages_supported) {
203             return false;
204         }
205     }
206 
207     return true;
208 }
209 
210 /*
211  * Check if all VFIO devices are running and migration is active, which is
212  * essentially equivalent to the migration being in pre-copy phase.
213  */
214 bool
215 vfio_devices_all_running_and_mig_active(const VFIOContainerBase *bcontainer)
216 {
217     VFIODevice *vbasedev;
218 
219     if (!migration_is_active()) {
220         return false;
221     }
222 
223     QLIST_FOREACH(vbasedev, &bcontainer->device_list, container_next) {
224         VFIOMigration *migration = vbasedev->migration;
225 
226         if (!migration) {
227             return false;
228         }
229 
230         if (vfio_device_state_is_running(vbasedev) ||
231             vfio_device_state_is_precopy(vbasedev)) {
232             continue;
233         } else {
234             return false;
235         }
236     }
237     return true;
238 }
239 
240 static bool vfio_listener_skipped_section(MemoryRegionSection *section)
241 {
242     return (!memory_region_is_ram(section->mr) &&
243             !memory_region_is_iommu(section->mr)) ||
244            memory_region_is_protected(section->mr) ||
245            /*
246             * Sizing an enabled 64-bit BAR can cause spurious mappings to
247             * addresses in the upper part of the 64-bit address space.  These
248             * are never accessed by the CPU and beyond the address width of
249             * some IOMMU hardware.  TODO: VFIO should tell us the IOMMU width.
250             */
251            section->offset_within_address_space & (1ULL << 63);
252 }
253 
254 /* Called with rcu_read_lock held.  */
255 static bool vfio_get_xlat_addr(IOMMUTLBEntry *iotlb, void **vaddr,
256                                ram_addr_t *ram_addr, bool *read_only,
257                                Error **errp)
258 {
259     bool ret, mr_has_discard_manager;
260 
261     ret = memory_get_xlat_addr(iotlb, vaddr, ram_addr, read_only,
262                                &mr_has_discard_manager, errp);
263     if (ret && mr_has_discard_manager) {
264         /*
265          * Malicious VMs might trigger discarding of IOMMU-mapped memory. The
266          * pages will remain pinned inside vfio until unmapped, resulting in a
267          * higher memory consumption than expected. If memory would get
268          * populated again later, there would be an inconsistency between pages
269          * pinned by vfio and pages seen by QEMU. This is the case until
270          * unmapped from the IOMMU (e.g., during device reset).
271          *
272          * With malicious guests, we really only care about pinning more memory
273          * than expected. RLIMIT_MEMLOCK set for the user/process can never be
274          * exceeded and can be used to mitigate this problem.
275          */
276         warn_report_once("Using vfio with vIOMMUs and coordinated discarding of"
277                          " RAM (e.g., virtio-mem) works, however, malicious"
278                          " guests can trigger pinning of more memory than"
279                          " intended via an IOMMU. It's possible to mitigate "
280                          " by setting/adjusting RLIMIT_MEMLOCK.");
281     }
282     return ret;
283 }
284 
285 static void vfio_iommu_map_notify(IOMMUNotifier *n, IOMMUTLBEntry *iotlb)
286 {
287     VFIOGuestIOMMU *giommu = container_of(n, VFIOGuestIOMMU, n);
288     VFIOContainerBase *bcontainer = giommu->bcontainer;
289     hwaddr iova = iotlb->iova + giommu->iommu_offset;
290     void *vaddr;
291     int ret;
292     Error *local_err = NULL;
293 
294     trace_vfio_iommu_map_notify(iotlb->perm == IOMMU_NONE ? "UNMAP" : "MAP",
295                                 iova, iova + iotlb->addr_mask);
296 
297     if (iotlb->target_as != &address_space_memory) {
298         error_report("Wrong target AS \"%s\", only system memory is allowed",
299                      iotlb->target_as->name ? iotlb->target_as->name : "none");
300         vfio_set_migration_error(-EINVAL);
301         return;
302     }
303 
304     rcu_read_lock();
305 
306     if ((iotlb->perm & IOMMU_RW) != IOMMU_NONE) {
307         bool read_only;
308 
309         if (!vfio_get_xlat_addr(iotlb, &vaddr, NULL, &read_only, &local_err)) {
310             error_report_err(local_err);
311             goto out;
312         }
313         /*
314          * vaddr is only valid until rcu_read_unlock(). But after
315          * vfio_dma_map has set up the mapping the pages will be
316          * pinned by the kernel. This makes sure that the RAM backend
317          * of vaddr will always be there, even if the memory object is
318          * destroyed and its backing memory munmap-ed.
319          */
320         ret = vfio_container_dma_map(bcontainer, iova,
321                                      iotlb->addr_mask + 1, vaddr,
322                                      read_only);
323         if (ret) {
324             error_report("vfio_container_dma_map(%p, 0x%"HWADDR_PRIx", "
325                          "0x%"HWADDR_PRIx", %p) = %d (%s)",
326                          bcontainer, iova,
327                          iotlb->addr_mask + 1, vaddr, ret, strerror(-ret));
328         }
329     } else {
330         ret = vfio_container_dma_unmap(bcontainer, iova,
331                                        iotlb->addr_mask + 1, iotlb);
332         if (ret) {
333             error_report("vfio_container_dma_unmap(%p, 0x%"HWADDR_PRIx", "
334                          "0x%"HWADDR_PRIx") = %d (%s)",
335                          bcontainer, iova,
336                          iotlb->addr_mask + 1, ret, strerror(-ret));
337             vfio_set_migration_error(ret);
338         }
339     }
340 out:
341     rcu_read_unlock();
342 }
343 
344 static void vfio_ram_discard_notify_discard(RamDiscardListener *rdl,
345                                             MemoryRegionSection *section)
346 {
347     VFIORamDiscardListener *vrdl = container_of(rdl, VFIORamDiscardListener,
348                                                 listener);
349     VFIOContainerBase *bcontainer = vrdl->bcontainer;
350     const hwaddr size = int128_get64(section->size);
351     const hwaddr iova = section->offset_within_address_space;
352     int ret;
353 
354     /* Unmap with a single call. */
355     ret = vfio_container_dma_unmap(bcontainer, iova, size , NULL);
356     if (ret) {
357         error_report("%s: vfio_container_dma_unmap() failed: %s", __func__,
358                      strerror(-ret));
359     }
360 }
361 
362 static int vfio_ram_discard_notify_populate(RamDiscardListener *rdl,
363                                             MemoryRegionSection *section)
364 {
365     VFIORamDiscardListener *vrdl = container_of(rdl, VFIORamDiscardListener,
366                                                 listener);
367     VFIOContainerBase *bcontainer = vrdl->bcontainer;
368     const hwaddr end = section->offset_within_region +
369                        int128_get64(section->size);
370     hwaddr start, next, iova;
371     void *vaddr;
372     int ret;
373 
374     /*
375      * Map in (aligned within memory region) minimum granularity, so we can
376      * unmap in minimum granularity later.
377      */
378     for (start = section->offset_within_region; start < end; start = next) {
379         next = ROUND_UP(start + 1, vrdl->granularity);
380         next = MIN(next, end);
381 
382         iova = start - section->offset_within_region +
383                section->offset_within_address_space;
384         vaddr = memory_region_get_ram_ptr(section->mr) + start;
385 
386         ret = vfio_container_dma_map(bcontainer, iova, next - start,
387                                      vaddr, section->readonly);
388         if (ret) {
389             /* Rollback */
390             vfio_ram_discard_notify_discard(rdl, section);
391             return ret;
392         }
393     }
394     return 0;
395 }
396 
397 static void vfio_register_ram_discard_listener(VFIOContainerBase *bcontainer,
398                                                MemoryRegionSection *section)
399 {
400     RamDiscardManager *rdm = memory_region_get_ram_discard_manager(section->mr);
401     VFIORamDiscardListener *vrdl;
402 
403     /* Ignore some corner cases not relevant in practice. */
404     g_assert(QEMU_IS_ALIGNED(section->offset_within_region, TARGET_PAGE_SIZE));
405     g_assert(QEMU_IS_ALIGNED(section->offset_within_address_space,
406                              TARGET_PAGE_SIZE));
407     g_assert(QEMU_IS_ALIGNED(int128_get64(section->size), TARGET_PAGE_SIZE));
408 
409     vrdl = g_new0(VFIORamDiscardListener, 1);
410     vrdl->bcontainer = bcontainer;
411     vrdl->mr = section->mr;
412     vrdl->offset_within_address_space = section->offset_within_address_space;
413     vrdl->size = int128_get64(section->size);
414     vrdl->granularity = ram_discard_manager_get_min_granularity(rdm,
415                                                                 section->mr);
416 
417     g_assert(vrdl->granularity && is_power_of_2(vrdl->granularity));
418     g_assert(bcontainer->pgsizes &&
419              vrdl->granularity >= 1ULL << ctz64(bcontainer->pgsizes));
420 
421     ram_discard_listener_init(&vrdl->listener,
422                               vfio_ram_discard_notify_populate,
423                               vfio_ram_discard_notify_discard, true);
424     ram_discard_manager_register_listener(rdm, &vrdl->listener, section);
425     QLIST_INSERT_HEAD(&bcontainer->vrdl_list, vrdl, next);
426 
427     /*
428      * Sanity-check if we have a theoretically problematic setup where we could
429      * exceed the maximum number of possible DMA mappings over time. We assume
430      * that each mapped section in the same address space as a RamDiscardManager
431      * section consumes exactly one DMA mapping, with the exception of
432      * RamDiscardManager sections; i.e., we don't expect to have gIOMMU sections
433      * in the same address space as RamDiscardManager sections.
434      *
435      * We assume that each section in the address space consumes one memslot.
436      * We take the number of KVM memory slots as a best guess for the maximum
437      * number of sections in the address space we could have over time,
438      * also consuming DMA mappings.
439      */
440     if (bcontainer->dma_max_mappings) {
441         unsigned int vrdl_count = 0, vrdl_mappings = 0, max_memslots = 512;
442 
443 #ifdef CONFIG_KVM
444         if (kvm_enabled()) {
445             max_memslots = kvm_get_max_memslots();
446         }
447 #endif
448 
449         QLIST_FOREACH(vrdl, &bcontainer->vrdl_list, next) {
450             hwaddr start, end;
451 
452             start = QEMU_ALIGN_DOWN(vrdl->offset_within_address_space,
453                                     vrdl->granularity);
454             end = ROUND_UP(vrdl->offset_within_address_space + vrdl->size,
455                            vrdl->granularity);
456             vrdl_mappings += (end - start) / vrdl->granularity;
457             vrdl_count++;
458         }
459 
460         if (vrdl_mappings + max_memslots - vrdl_count >
461             bcontainer->dma_max_mappings) {
462             warn_report("%s: possibly running out of DMA mappings. E.g., try"
463                         " increasing the 'block-size' of virtio-mem devies."
464                         " Maximum possible DMA mappings: %d, Maximum possible"
465                         " memslots: %d", __func__, bcontainer->dma_max_mappings,
466                         max_memslots);
467         }
468     }
469 }
470 
471 static void vfio_unregister_ram_discard_listener(VFIOContainerBase *bcontainer,
472                                                  MemoryRegionSection *section)
473 {
474     RamDiscardManager *rdm = memory_region_get_ram_discard_manager(section->mr);
475     VFIORamDiscardListener *vrdl = NULL;
476 
477     QLIST_FOREACH(vrdl, &bcontainer->vrdl_list, next) {
478         if (vrdl->mr == section->mr &&
479             vrdl->offset_within_address_space ==
480             section->offset_within_address_space) {
481             break;
482         }
483     }
484 
485     if (!vrdl) {
486         hw_error("vfio: Trying to unregister missing RAM discard listener");
487     }
488 
489     ram_discard_manager_unregister_listener(rdm, &vrdl->listener);
490     QLIST_REMOVE(vrdl, next);
491     g_free(vrdl);
492 }
493 
494 static bool vfio_known_safe_misalignment(MemoryRegionSection *section)
495 {
496     MemoryRegion *mr = section->mr;
497 
498     if (!TPM_IS_CRB(mr->owner)) {
499         return false;
500     }
501 
502     /* this is a known safe misaligned region, just trace for debug purpose */
503     trace_vfio_known_safe_misalignment(memory_region_name(mr),
504                                        section->offset_within_address_space,
505                                        section->offset_within_region,
506                                        qemu_real_host_page_size());
507     return true;
508 }
509 
510 static bool vfio_listener_valid_section(MemoryRegionSection *section,
511                                         const char *name)
512 {
513     if (vfio_listener_skipped_section(section)) {
514         trace_vfio_listener_region_skip(name,
515                 section->offset_within_address_space,
516                 section->offset_within_address_space +
517                 int128_get64(int128_sub(section->size, int128_one())));
518         return false;
519     }
520 
521     if (unlikely((section->offset_within_address_space &
522                   ~qemu_real_host_page_mask()) !=
523                  (section->offset_within_region & ~qemu_real_host_page_mask()))) {
524         if (!vfio_known_safe_misalignment(section)) {
525             error_report("%s received unaligned region %s iova=0x%"PRIx64
526                          " offset_within_region=0x%"PRIx64
527                          " qemu_real_host_page_size=0x%"PRIxPTR,
528                          __func__, memory_region_name(section->mr),
529                          section->offset_within_address_space,
530                          section->offset_within_region,
531                          qemu_real_host_page_size());
532         }
533         return false;
534     }
535 
536     return true;
537 }
538 
539 static bool vfio_get_section_iova_range(VFIOContainerBase *bcontainer,
540                                         MemoryRegionSection *section,
541                                         hwaddr *out_iova, hwaddr *out_end,
542                                         Int128 *out_llend)
543 {
544     Int128 llend;
545     hwaddr iova;
546 
547     iova = REAL_HOST_PAGE_ALIGN(section->offset_within_address_space);
548     llend = int128_make64(section->offset_within_address_space);
549     llend = int128_add(llend, section->size);
550     llend = int128_and(llend, int128_exts64(qemu_real_host_page_mask()));
551 
552     if (int128_ge(int128_make64(iova), llend)) {
553         return false;
554     }
555 
556     *out_iova = iova;
557     *out_end = int128_get64(int128_sub(llend, int128_one()));
558     if (out_llend) {
559         *out_llend = llend;
560     }
561     return true;
562 }
563 
564 static void vfio_listener_region_add(MemoryListener *listener,
565                                      MemoryRegionSection *section)
566 {
567     VFIOContainerBase *bcontainer = container_of(listener, VFIOContainerBase,
568                                                  listener);
569     hwaddr iova, end;
570     Int128 llend, llsize;
571     void *vaddr;
572     int ret;
573     Error *err = NULL;
574 
575     if (!vfio_listener_valid_section(section, "region_add")) {
576         return;
577     }
578 
579     if (!vfio_get_section_iova_range(bcontainer, section, &iova, &end,
580                                      &llend)) {
581         if (memory_region_is_ram_device(section->mr)) {
582             trace_vfio_listener_region_add_no_dma_map(
583                 memory_region_name(section->mr),
584                 section->offset_within_address_space,
585                 int128_getlo(section->size),
586                 qemu_real_host_page_size());
587         }
588         return;
589     }
590 
591     if (!vfio_container_add_section_window(bcontainer, section, &err)) {
592         goto fail;
593     }
594 
595     memory_region_ref(section->mr);
596 
597     if (memory_region_is_iommu(section->mr)) {
598         VFIOGuestIOMMU *giommu;
599         IOMMUMemoryRegion *iommu_mr = IOMMU_MEMORY_REGION(section->mr);
600         int iommu_idx;
601 
602         trace_vfio_listener_region_add_iommu(iova, end);
603         /*
604          * FIXME: For VFIO iommu types which have KVM acceleration to
605          * avoid bouncing all map/unmaps through qemu this way, this
606          * would be the right place to wire that up (tell the KVM
607          * device emulation the VFIO iommu handles to use).
608          */
609         giommu = g_malloc0(sizeof(*giommu));
610         giommu->iommu_mr = iommu_mr;
611         giommu->iommu_offset = section->offset_within_address_space -
612                                section->offset_within_region;
613         giommu->bcontainer = bcontainer;
614         llend = int128_add(int128_make64(section->offset_within_region),
615                            section->size);
616         llend = int128_sub(llend, int128_one());
617         iommu_idx = memory_region_iommu_attrs_to_index(iommu_mr,
618                                                        MEMTXATTRS_UNSPECIFIED);
619         iommu_notifier_init(&giommu->n, vfio_iommu_map_notify,
620                             IOMMU_NOTIFIER_IOTLB_EVENTS,
621                             section->offset_within_region,
622                             int128_get64(llend),
623                             iommu_idx);
624 
625         ret = memory_region_iommu_set_page_size_mask(giommu->iommu_mr,
626                                                      bcontainer->pgsizes,
627                                                      &err);
628         if (ret) {
629             g_free(giommu);
630             goto fail;
631         }
632 
633         if (bcontainer->iova_ranges) {
634             ret = memory_region_iommu_set_iova_ranges(giommu->iommu_mr,
635                                                       bcontainer->iova_ranges,
636                                                       &err);
637             if (ret) {
638                 g_free(giommu);
639                 goto fail;
640             }
641         }
642 
643         ret = memory_region_register_iommu_notifier(section->mr, &giommu->n,
644                                                     &err);
645         if (ret) {
646             g_free(giommu);
647             goto fail;
648         }
649         QLIST_INSERT_HEAD(&bcontainer->giommu_list, giommu, giommu_next);
650         memory_region_iommu_replay(giommu->iommu_mr, &giommu->n);
651 
652         return;
653     }
654 
655     /* Here we assume that memory_region_is_ram(section->mr)==true */
656 
657     /*
658      * For RAM memory regions with a RamDiscardManager, we only want to map the
659      * actually populated parts - and update the mapping whenever we're notified
660      * about changes.
661      */
662     if (memory_region_has_ram_discard_manager(section->mr)) {
663         vfio_register_ram_discard_listener(bcontainer, section);
664         return;
665     }
666 
667     vaddr = memory_region_get_ram_ptr(section->mr) +
668             section->offset_within_region +
669             (iova - section->offset_within_address_space);
670 
671     trace_vfio_listener_region_add_ram(iova, end, vaddr);
672 
673     llsize = int128_sub(llend, int128_make64(iova));
674 
675     if (memory_region_is_ram_device(section->mr)) {
676         hwaddr pgmask = (1ULL << ctz64(bcontainer->pgsizes)) - 1;
677 
678         if ((iova & pgmask) || (int128_get64(llsize) & pgmask)) {
679             trace_vfio_listener_region_add_no_dma_map(
680                 memory_region_name(section->mr),
681                 section->offset_within_address_space,
682                 int128_getlo(section->size),
683                 pgmask + 1);
684             return;
685         }
686     }
687 
688     ret = vfio_container_dma_map(bcontainer, iova, int128_get64(llsize),
689                                  vaddr, section->readonly);
690     if (ret) {
691         error_setg(&err, "vfio_container_dma_map(%p, 0x%"HWADDR_PRIx", "
692                    "0x%"HWADDR_PRIx", %p) = %d (%s)",
693                    bcontainer, iova, int128_get64(llsize), vaddr, ret,
694                    strerror(-ret));
695         if (memory_region_is_ram_device(section->mr)) {
696             /* Allow unexpected mappings not to be fatal for RAM devices */
697             error_report_err(err);
698             return;
699         }
700         goto fail;
701     }
702 
703     return;
704 
705 fail:
706     if (memory_region_is_ram_device(section->mr)) {
707         error_reportf_err(err, "PCI p2p may not work: ");
708         return;
709     }
710     /*
711      * On the initfn path, store the first error in the container so we
712      * can gracefully fail.  Runtime, there's not much we can do other
713      * than throw a hardware error.
714      */
715     if (!bcontainer->initialized) {
716         if (!bcontainer->error) {
717             error_propagate_prepend(&bcontainer->error, err,
718                                     "Region %s: ",
719                                     memory_region_name(section->mr));
720         } else {
721             error_free(err);
722         }
723     } else {
724         error_report_err(err);
725         hw_error("vfio: DMA mapping failed, unable to continue");
726     }
727 }
728 
729 static void vfio_listener_region_del(MemoryListener *listener,
730                                      MemoryRegionSection *section)
731 {
732     VFIOContainerBase *bcontainer = container_of(listener, VFIOContainerBase,
733                                                  listener);
734     hwaddr iova, end;
735     Int128 llend, llsize;
736     int ret;
737     bool try_unmap = true;
738 
739     if (!vfio_listener_valid_section(section, "region_del")) {
740         return;
741     }
742 
743     if (memory_region_is_iommu(section->mr)) {
744         VFIOGuestIOMMU *giommu;
745 
746         QLIST_FOREACH(giommu, &bcontainer->giommu_list, giommu_next) {
747             if (MEMORY_REGION(giommu->iommu_mr) == section->mr &&
748                 giommu->n.start == section->offset_within_region) {
749                 memory_region_unregister_iommu_notifier(section->mr,
750                                                         &giommu->n);
751                 QLIST_REMOVE(giommu, giommu_next);
752                 g_free(giommu);
753                 break;
754             }
755         }
756 
757         /*
758          * FIXME: We assume the one big unmap below is adequate to
759          * remove any individual page mappings in the IOMMU which
760          * might have been copied into VFIO. This works for a page table
761          * based IOMMU where a big unmap flattens a large range of IO-PTEs.
762          * That may not be true for all IOMMU types.
763          */
764     }
765 
766     if (!vfio_get_section_iova_range(bcontainer, section, &iova, &end,
767                                      &llend)) {
768         return;
769     }
770 
771     llsize = int128_sub(llend, int128_make64(iova));
772 
773     trace_vfio_listener_region_del(iova, end);
774 
775     if (memory_region_is_ram_device(section->mr)) {
776         hwaddr pgmask;
777 
778         pgmask = (1ULL << ctz64(bcontainer->pgsizes)) - 1;
779         try_unmap = !((iova & pgmask) || (int128_get64(llsize) & pgmask));
780     } else if (memory_region_has_ram_discard_manager(section->mr)) {
781         vfio_unregister_ram_discard_listener(bcontainer, section);
782         /* Unregistering will trigger an unmap. */
783         try_unmap = false;
784     }
785 
786     if (try_unmap) {
787         if (int128_eq(llsize, int128_2_64())) {
788             /* The unmap ioctl doesn't accept a full 64-bit span. */
789             llsize = int128_rshift(llsize, 1);
790             ret = vfio_container_dma_unmap(bcontainer, iova,
791                                            int128_get64(llsize), NULL);
792             if (ret) {
793                 error_report("vfio_container_dma_unmap(%p, 0x%"HWADDR_PRIx", "
794                              "0x%"HWADDR_PRIx") = %d (%s)",
795                              bcontainer, iova, int128_get64(llsize), ret,
796                              strerror(-ret));
797             }
798             iova += int128_get64(llsize);
799         }
800         ret = vfio_container_dma_unmap(bcontainer, iova,
801                                        int128_get64(llsize), NULL);
802         if (ret) {
803             error_report("vfio_container_dma_unmap(%p, 0x%"HWADDR_PRIx", "
804                          "0x%"HWADDR_PRIx") = %d (%s)",
805                          bcontainer, iova, int128_get64(llsize), ret,
806                          strerror(-ret));
807         }
808     }
809 
810     memory_region_unref(section->mr);
811 
812     vfio_container_del_section_window(bcontainer, section);
813 }
814 
815 typedef struct VFIODirtyRanges {
816     hwaddr min32;
817     hwaddr max32;
818     hwaddr min64;
819     hwaddr max64;
820     hwaddr minpci64;
821     hwaddr maxpci64;
822 } VFIODirtyRanges;
823 
824 typedef struct VFIODirtyRangesListener {
825     VFIOContainerBase *bcontainer;
826     VFIODirtyRanges ranges;
827     MemoryListener listener;
828 } VFIODirtyRangesListener;
829 
830 static bool vfio_section_is_vfio_pci(MemoryRegionSection *section,
831                                      VFIOContainerBase *bcontainer)
832 {
833     VFIOPCIDevice *pcidev;
834     VFIODevice *vbasedev;
835     Object *owner;
836 
837     owner = memory_region_owner(section->mr);
838 
839     QLIST_FOREACH(vbasedev, &bcontainer->device_list, container_next) {
840         if (vbasedev->type != VFIO_DEVICE_TYPE_PCI) {
841             continue;
842         }
843         pcidev = container_of(vbasedev, VFIOPCIDevice, vbasedev);
844         if (OBJECT(pcidev) == owner) {
845             return true;
846         }
847     }
848 
849     return false;
850 }
851 
852 static void vfio_dirty_tracking_update(MemoryListener *listener,
853                                        MemoryRegionSection *section)
854 {
855     VFIODirtyRangesListener *dirty = container_of(listener,
856                                                   VFIODirtyRangesListener,
857                                                   listener);
858     VFIODirtyRanges *range = &dirty->ranges;
859     hwaddr iova, end, *min, *max;
860 
861     if (!vfio_listener_valid_section(section, "tracking_update") ||
862         !vfio_get_section_iova_range(dirty->bcontainer, section,
863                                      &iova, &end, NULL)) {
864         return;
865     }
866 
867     /*
868      * The address space passed to the dirty tracker is reduced to three ranges:
869      * one for 32-bit DMA ranges, one for 64-bit DMA ranges and one for the
870      * PCI 64-bit hole.
871      *
872      * The underlying reports of dirty will query a sub-interval of each of
873      * these ranges.
874      *
875      * The purpose of the three range handling is to handle known cases of big
876      * holes in the address space, like the x86 AMD 1T hole, and firmware (like
877      * OVMF) which may relocate the pci-hole64 to the end of the address space.
878      * The latter would otherwise generate large ranges for tracking, stressing
879      * the limits of supported hardware. The pci-hole32 will always be below 4G
880      * (overlapping or not) so it doesn't need special handling and is part of
881      * the 32-bit range.
882      *
883      * The alternative would be an IOVATree but that has a much bigger runtime
884      * overhead and unnecessary complexity.
885      */
886     if (vfio_section_is_vfio_pci(section, dirty->bcontainer) &&
887         iova >= UINT32_MAX) {
888         min = &range->minpci64;
889         max = &range->maxpci64;
890     } else {
891         min = (end <= UINT32_MAX) ? &range->min32 : &range->min64;
892         max = (end <= UINT32_MAX) ? &range->max32 : &range->max64;
893     }
894     if (*min > iova) {
895         *min = iova;
896     }
897     if (*max < end) {
898         *max = end;
899     }
900 
901     trace_vfio_device_dirty_tracking_update(iova, end, *min, *max);
902     return;
903 }
904 
905 static const MemoryListener vfio_dirty_tracking_listener = {
906     .name = "vfio-tracking",
907     .region_add = vfio_dirty_tracking_update,
908 };
909 
910 static void vfio_dirty_tracking_init(VFIOContainerBase *bcontainer,
911                                      VFIODirtyRanges *ranges)
912 {
913     VFIODirtyRangesListener dirty;
914 
915     memset(&dirty, 0, sizeof(dirty));
916     dirty.ranges.min32 = UINT32_MAX;
917     dirty.ranges.min64 = UINT64_MAX;
918     dirty.ranges.minpci64 = UINT64_MAX;
919     dirty.listener = vfio_dirty_tracking_listener;
920     dirty.bcontainer = bcontainer;
921 
922     memory_listener_register(&dirty.listener,
923                              bcontainer->space->as);
924 
925     *ranges = dirty.ranges;
926 
927     /*
928      * The memory listener is synchronous, and used to calculate the range
929      * to dirty tracking. Unregister it after we are done as we are not
930      * interested in any follow-up updates.
931      */
932     memory_listener_unregister(&dirty.listener);
933 }
934 
935 static void vfio_devices_dma_logging_stop(VFIOContainerBase *bcontainer)
936 {
937     uint64_t buf[DIV_ROUND_UP(sizeof(struct vfio_device_feature),
938                               sizeof(uint64_t))] = {};
939     struct vfio_device_feature *feature = (struct vfio_device_feature *)buf;
940     VFIODevice *vbasedev;
941 
942     feature->argsz = sizeof(buf);
943     feature->flags = VFIO_DEVICE_FEATURE_SET |
944                      VFIO_DEVICE_FEATURE_DMA_LOGGING_STOP;
945 
946     QLIST_FOREACH(vbasedev, &bcontainer->device_list, container_next) {
947         if (!vbasedev->dirty_tracking) {
948             continue;
949         }
950 
951         if (ioctl(vbasedev->fd, VFIO_DEVICE_FEATURE, feature)) {
952             warn_report("%s: Failed to stop DMA logging, err %d (%s)",
953                         vbasedev->name, -errno, strerror(errno));
954         }
955         vbasedev->dirty_tracking = false;
956     }
957 }
958 
959 static struct vfio_device_feature *
960 vfio_device_feature_dma_logging_start_create(VFIOContainerBase *bcontainer,
961                                              VFIODirtyRanges *tracking)
962 {
963     struct vfio_device_feature *feature;
964     size_t feature_size;
965     struct vfio_device_feature_dma_logging_control *control;
966     struct vfio_device_feature_dma_logging_range *ranges;
967 
968     feature_size = sizeof(struct vfio_device_feature) +
969                    sizeof(struct vfio_device_feature_dma_logging_control);
970     feature = g_try_malloc0(feature_size);
971     if (!feature) {
972         errno = ENOMEM;
973         return NULL;
974     }
975     feature->argsz = feature_size;
976     feature->flags = VFIO_DEVICE_FEATURE_SET |
977                      VFIO_DEVICE_FEATURE_DMA_LOGGING_START;
978 
979     control = (struct vfio_device_feature_dma_logging_control *)feature->data;
980     control->page_size = qemu_real_host_page_size();
981 
982     /*
983      * DMA logging uAPI guarantees to support at least a number of ranges that
984      * fits into a single host kernel base page.
985      */
986     control->num_ranges = !!tracking->max32 + !!tracking->max64 +
987         !!tracking->maxpci64;
988     ranges = g_try_new0(struct vfio_device_feature_dma_logging_range,
989                         control->num_ranges);
990     if (!ranges) {
991         g_free(feature);
992         errno = ENOMEM;
993 
994         return NULL;
995     }
996 
997     control->ranges = (uintptr_t)ranges;
998     if (tracking->max32) {
999         ranges->iova = tracking->min32;
1000         ranges->length = (tracking->max32 - tracking->min32) + 1;
1001         ranges++;
1002     }
1003     if (tracking->max64) {
1004         ranges->iova = tracking->min64;
1005         ranges->length = (tracking->max64 - tracking->min64) + 1;
1006         ranges++;
1007     }
1008     if (tracking->maxpci64) {
1009         ranges->iova = tracking->minpci64;
1010         ranges->length = (tracking->maxpci64 - tracking->minpci64) + 1;
1011     }
1012 
1013     trace_vfio_device_dirty_tracking_start(control->num_ranges,
1014                                            tracking->min32, tracking->max32,
1015                                            tracking->min64, tracking->max64,
1016                                            tracking->minpci64, tracking->maxpci64);
1017 
1018     return feature;
1019 }
1020 
1021 static void vfio_device_feature_dma_logging_start_destroy(
1022     struct vfio_device_feature *feature)
1023 {
1024     struct vfio_device_feature_dma_logging_control *control =
1025         (struct vfio_device_feature_dma_logging_control *)feature->data;
1026     struct vfio_device_feature_dma_logging_range *ranges =
1027         (struct vfio_device_feature_dma_logging_range *)(uintptr_t)control->ranges;
1028 
1029     g_free(ranges);
1030     g_free(feature);
1031 }
1032 
1033 static int vfio_devices_dma_logging_start(VFIOContainerBase *bcontainer,
1034                                           Error **errp)
1035 {
1036     struct vfio_device_feature *feature;
1037     VFIODirtyRanges ranges;
1038     VFIODevice *vbasedev;
1039     int ret = 0;
1040 
1041     vfio_dirty_tracking_init(bcontainer, &ranges);
1042     feature = vfio_device_feature_dma_logging_start_create(bcontainer,
1043                                                            &ranges);
1044     if (!feature) {
1045         error_setg_errno(errp, errno, "Failed to prepare DMA logging");
1046         return -errno;
1047     }
1048 
1049     QLIST_FOREACH(vbasedev, &bcontainer->device_list, container_next) {
1050         if (vbasedev->dirty_tracking) {
1051             continue;
1052         }
1053 
1054         ret = ioctl(vbasedev->fd, VFIO_DEVICE_FEATURE, feature);
1055         if (ret) {
1056             ret = -errno;
1057             error_setg_errno(errp, errno, "%s: Failed to start DMA logging",
1058                              vbasedev->name);
1059             goto out;
1060         }
1061         vbasedev->dirty_tracking = true;
1062     }
1063 
1064 out:
1065     if (ret) {
1066         vfio_devices_dma_logging_stop(bcontainer);
1067     }
1068 
1069     vfio_device_feature_dma_logging_start_destroy(feature);
1070 
1071     return ret;
1072 }
1073 
1074 static bool vfio_listener_log_global_start(MemoryListener *listener,
1075                                            Error **errp)
1076 {
1077     ERRP_GUARD();
1078     VFIOContainerBase *bcontainer = container_of(listener, VFIOContainerBase,
1079                                                  listener);
1080     int ret;
1081 
1082     if (vfio_devices_all_device_dirty_tracking(bcontainer)) {
1083         ret = vfio_devices_dma_logging_start(bcontainer, errp);
1084     } else {
1085         ret = vfio_container_set_dirty_page_tracking(bcontainer, true, errp);
1086     }
1087 
1088     if (ret) {
1089         error_prepend(errp, "vfio: Could not start dirty page tracking - ");
1090     }
1091     return !ret;
1092 }
1093 
1094 static void vfio_listener_log_global_stop(MemoryListener *listener)
1095 {
1096     VFIOContainerBase *bcontainer = container_of(listener, VFIOContainerBase,
1097                                                  listener);
1098     Error *local_err = NULL;
1099     int ret = 0;
1100 
1101     if (vfio_devices_all_device_dirty_tracking(bcontainer)) {
1102         vfio_devices_dma_logging_stop(bcontainer);
1103     } else {
1104         ret = vfio_container_set_dirty_page_tracking(bcontainer, false,
1105                                                      &local_err);
1106     }
1107 
1108     if (ret) {
1109         error_prepend(&local_err,
1110                       "vfio: Could not stop dirty page tracking - ");
1111         error_report_err(local_err);
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, hwaddr size, Error **errp)
1144 {
1145     VFIODevice *vbasedev;
1146     int ret;
1147 
1148     QLIST_FOREACH(vbasedev, &bcontainer->device_list, container_next) {
1149         ret = vfio_device_dma_logging_report(vbasedev, iova, size,
1150                                              vbmap->bitmap);
1151         if (ret) {
1152             error_setg_errno(errp, -ret,
1153                              "%s: Failed to get DMA logging report, iova: "
1154                              "0x%" HWADDR_PRIx ", size: 0x%" HWADDR_PRIx,
1155                              vbasedev->name, iova, size);
1156 
1157             return ret;
1158         }
1159     }
1160 
1161     return 0;
1162 }
1163 
1164 int vfio_get_dirty_bitmap(const VFIOContainerBase *bcontainer, uint64_t iova,
1165                           uint64_t size, ram_addr_t ram_addr, Error **errp)
1166 {
1167     bool all_device_dirty_tracking =
1168         vfio_devices_all_device_dirty_tracking(bcontainer);
1169     uint64_t dirty_pages;
1170     VFIOBitmap vbmap;
1171     int ret;
1172 
1173     if (!bcontainer->dirty_pages_supported && !all_device_dirty_tracking) {
1174         cpu_physical_memory_set_dirty_range(ram_addr, size,
1175                                             tcg_enabled() ? DIRTY_CLIENTS_ALL :
1176                                             DIRTY_CLIENTS_NOCODE);
1177         return 0;
1178     }
1179 
1180     ret = vfio_bitmap_alloc(&vbmap, size);
1181     if (ret) {
1182         error_setg_errno(errp, -ret,
1183                          "Failed to allocate dirty tracking bitmap");
1184         return ret;
1185     }
1186 
1187     if (all_device_dirty_tracking) {
1188         ret = vfio_devices_query_dirty_bitmap(bcontainer, &vbmap, iova, size,
1189                                               errp);
1190     } else {
1191         ret = vfio_container_query_dirty_bitmap(bcontainer, &vbmap, iova, size,
1192                                                 errp);
1193     }
1194 
1195     if (ret) {
1196         goto out;
1197     }
1198 
1199     dirty_pages = cpu_physical_memory_set_dirty_lebitmap(vbmap.bitmap, ram_addr,
1200                                                          vbmap.pages);
1201 
1202     trace_vfio_get_dirty_bitmap(iova, size, vbmap.size, ram_addr, dirty_pages);
1203 out:
1204     g_free(vbmap.bitmap);
1205 
1206     return ret;
1207 }
1208 
1209 typedef struct {
1210     IOMMUNotifier n;
1211     VFIOGuestIOMMU *giommu;
1212 } vfio_giommu_dirty_notifier;
1213 
1214 static void vfio_iommu_map_dirty_notify(IOMMUNotifier *n, IOMMUTLBEntry *iotlb)
1215 {
1216     vfio_giommu_dirty_notifier *gdn = container_of(n,
1217                                                 vfio_giommu_dirty_notifier, n);
1218     VFIOGuestIOMMU *giommu = gdn->giommu;
1219     VFIOContainerBase *bcontainer = giommu->bcontainer;
1220     hwaddr iova = iotlb->iova + giommu->iommu_offset;
1221     ram_addr_t translated_addr;
1222     Error *local_err = NULL;
1223     int ret = -EINVAL;
1224 
1225     trace_vfio_iommu_map_dirty_notify(iova, iova + iotlb->addr_mask);
1226 
1227     if (iotlb->target_as != &address_space_memory) {
1228         error_report("Wrong target AS \"%s\", only system memory is allowed",
1229                      iotlb->target_as->name ? iotlb->target_as->name : "none");
1230         goto out;
1231     }
1232 
1233     rcu_read_lock();
1234     if (!vfio_get_xlat_addr(iotlb, NULL, &translated_addr, NULL, &local_err)) {
1235         error_report_err(local_err);
1236         goto out_unlock;
1237     }
1238 
1239     ret = vfio_get_dirty_bitmap(bcontainer, iova, iotlb->addr_mask + 1,
1240                                 translated_addr, &local_err);
1241     if (ret) {
1242         error_prepend(&local_err,
1243                       "vfio_iommu_map_dirty_notify(%p, 0x%"HWADDR_PRIx", "
1244                       "0x%"HWADDR_PRIx") failed - ", bcontainer, iova,
1245                       iotlb->addr_mask + 1);
1246         error_report_err(local_err);
1247     }
1248 
1249 out_unlock:
1250     rcu_read_unlock();
1251 
1252 out:
1253     if (ret) {
1254         vfio_set_migration_error(ret);
1255     }
1256 }
1257 
1258 static int vfio_ram_discard_get_dirty_bitmap(MemoryRegionSection *section,
1259                                              void *opaque)
1260 {
1261     const hwaddr size = int128_get64(section->size);
1262     const hwaddr iova = section->offset_within_address_space;
1263     const ram_addr_t ram_addr = memory_region_get_ram_addr(section->mr) +
1264                                 section->offset_within_region;
1265     VFIORamDiscardListener *vrdl = opaque;
1266     Error *local_err = NULL;
1267     int ret;
1268 
1269     /*
1270      * Sync the whole mapped region (spanning multiple individual mappings)
1271      * in one go.
1272      */
1273     ret = vfio_get_dirty_bitmap(vrdl->bcontainer, iova, size, ram_addr,
1274                                 &local_err);
1275     if (ret) {
1276         error_report_err(local_err);
1277     }
1278     return ret;
1279 }
1280 
1281 static int
1282 vfio_sync_ram_discard_listener_dirty_bitmap(VFIOContainerBase *bcontainer,
1283                                             MemoryRegionSection *section)
1284 {
1285     RamDiscardManager *rdm = memory_region_get_ram_discard_manager(section->mr);
1286     VFIORamDiscardListener *vrdl = NULL;
1287 
1288     QLIST_FOREACH(vrdl, &bcontainer->vrdl_list, next) {
1289         if (vrdl->mr == section->mr &&
1290             vrdl->offset_within_address_space ==
1291             section->offset_within_address_space) {
1292             break;
1293         }
1294     }
1295 
1296     if (!vrdl) {
1297         hw_error("vfio: Trying to sync missing RAM discard listener");
1298     }
1299 
1300     /*
1301      * We only want/can synchronize the bitmap for actually mapped parts -
1302      * which correspond to populated parts. Replay all populated parts.
1303      */
1304     return ram_discard_manager_replay_populated(rdm, section,
1305                                               vfio_ram_discard_get_dirty_bitmap,
1306                                                 &vrdl);
1307 }
1308 
1309 static int vfio_sync_dirty_bitmap(VFIOContainerBase *bcontainer,
1310                                   MemoryRegionSection *section, Error **errp)
1311 {
1312     ram_addr_t ram_addr;
1313 
1314     if (memory_region_is_iommu(section->mr)) {
1315         VFIOGuestIOMMU *giommu;
1316 
1317         QLIST_FOREACH(giommu, &bcontainer->giommu_list, giommu_next) {
1318             if (MEMORY_REGION(giommu->iommu_mr) == section->mr &&
1319                 giommu->n.start == section->offset_within_region) {
1320                 Int128 llend;
1321                 vfio_giommu_dirty_notifier gdn = { .giommu = giommu };
1322                 int idx = memory_region_iommu_attrs_to_index(giommu->iommu_mr,
1323                                                        MEMTXATTRS_UNSPECIFIED);
1324 
1325                 llend = int128_add(int128_make64(section->offset_within_region),
1326                                    section->size);
1327                 llend = int128_sub(llend, int128_one());
1328 
1329                 iommu_notifier_init(&gdn.n,
1330                                     vfio_iommu_map_dirty_notify,
1331                                     IOMMU_NOTIFIER_MAP,
1332                                     section->offset_within_region,
1333                                     int128_get64(llend),
1334                                     idx);
1335                 memory_region_iommu_replay(giommu->iommu_mr, &gdn.n);
1336                 break;
1337             }
1338         }
1339         return 0;
1340     } else if (memory_region_has_ram_discard_manager(section->mr)) {
1341         int ret;
1342 
1343         ret = vfio_sync_ram_discard_listener_dirty_bitmap(bcontainer, section);
1344         if (ret) {
1345             error_setg(errp,
1346                        "Failed to sync dirty bitmap with RAM discard listener");
1347         }
1348         return ret;
1349     }
1350 
1351     ram_addr = memory_region_get_ram_addr(section->mr) +
1352                section->offset_within_region;
1353 
1354     return vfio_get_dirty_bitmap(bcontainer,
1355                    REAL_HOST_PAGE_ALIGN(section->offset_within_address_space),
1356                                  int128_get64(section->size), ram_addr, errp);
1357 }
1358 
1359 static void vfio_listener_log_sync(MemoryListener *listener,
1360         MemoryRegionSection *section)
1361 {
1362     VFIOContainerBase *bcontainer = container_of(listener, VFIOContainerBase,
1363                                                  listener);
1364     int ret;
1365     Error *local_err = NULL;
1366 
1367     if (vfio_listener_skipped_section(section)) {
1368         return;
1369     }
1370 
1371     if (vfio_devices_all_dirty_tracking(bcontainer)) {
1372         ret = vfio_sync_dirty_bitmap(bcontainer, section, &local_err);
1373         if (ret) {
1374             error_report_err(local_err);
1375             vfio_set_migration_error(ret);
1376         }
1377     }
1378 }
1379 
1380 const MemoryListener vfio_memory_listener = {
1381     .name = "vfio",
1382     .region_add = vfio_listener_region_add,
1383     .region_del = vfio_listener_region_del,
1384     .log_global_start = vfio_listener_log_global_start,
1385     .log_global_stop = vfio_listener_log_global_stop,
1386     .log_sync = vfio_listener_log_sync,
1387 };
1388 
1389 void vfio_reset_handler(void *opaque)
1390 {
1391     VFIODevice *vbasedev;
1392 
1393     QLIST_FOREACH(vbasedev, &vfio_device_list, global_next) {
1394         if (vbasedev->dev->realized) {
1395             vbasedev->ops->vfio_compute_needs_reset(vbasedev);
1396         }
1397     }
1398 
1399     QLIST_FOREACH(vbasedev, &vfio_device_list, global_next) {
1400         if (vbasedev->dev->realized && vbasedev->needs_reset) {
1401             vbasedev->ops->vfio_hot_reset_multi(vbasedev);
1402         }
1403     }
1404 }
1405 
1406 int vfio_kvm_device_add_fd(int fd, Error **errp)
1407 {
1408 #ifdef CONFIG_KVM
1409     struct kvm_device_attr attr = {
1410         .group = KVM_DEV_VFIO_FILE,
1411         .attr = KVM_DEV_VFIO_FILE_ADD,
1412         .addr = (uint64_t)(unsigned long)&fd,
1413     };
1414 
1415     if (!kvm_enabled()) {
1416         return 0;
1417     }
1418 
1419     if (vfio_kvm_device_fd < 0) {
1420         struct kvm_create_device cd = {
1421             .type = KVM_DEV_TYPE_VFIO,
1422         };
1423 
1424         if (kvm_vm_ioctl(kvm_state, KVM_CREATE_DEVICE, &cd)) {
1425             error_setg_errno(errp, errno, "Failed to create KVM VFIO device");
1426             return -errno;
1427         }
1428 
1429         vfio_kvm_device_fd = cd.fd;
1430     }
1431 
1432     if (ioctl(vfio_kvm_device_fd, KVM_SET_DEVICE_ATTR, &attr)) {
1433         error_setg_errno(errp, errno, "Failed to add fd %d to KVM VFIO device",
1434                          fd);
1435         return -errno;
1436     }
1437 #endif
1438     return 0;
1439 }
1440 
1441 int vfio_kvm_device_del_fd(int fd, Error **errp)
1442 {
1443 #ifdef CONFIG_KVM
1444     struct kvm_device_attr attr = {
1445         .group = KVM_DEV_VFIO_FILE,
1446         .attr = KVM_DEV_VFIO_FILE_DEL,
1447         .addr = (uint64_t)(unsigned long)&fd,
1448     };
1449 
1450     if (vfio_kvm_device_fd < 0) {
1451         error_setg(errp, "KVM VFIO device isn't created yet");
1452         return -EINVAL;
1453     }
1454 
1455     if (ioctl(vfio_kvm_device_fd, KVM_SET_DEVICE_ATTR, &attr)) {
1456         error_setg_errno(errp, errno,
1457                          "Failed to remove fd %d from KVM VFIO device", fd);
1458         return -errno;
1459     }
1460 #endif
1461     return 0;
1462 }
1463 
1464 VFIOAddressSpace *vfio_get_address_space(AddressSpace *as)
1465 {
1466     VFIOAddressSpace *space;
1467 
1468     QLIST_FOREACH(space, &vfio_address_spaces, list) {
1469         if (space->as == as) {
1470             return space;
1471         }
1472     }
1473 
1474     /* No suitable VFIOAddressSpace, create a new one */
1475     space = g_malloc0(sizeof(*space));
1476     space->as = as;
1477     QLIST_INIT(&space->containers);
1478 
1479     if (QLIST_EMPTY(&vfio_address_spaces)) {
1480         qemu_register_reset(vfio_reset_handler, NULL);
1481     }
1482 
1483     QLIST_INSERT_HEAD(&vfio_address_spaces, space, list);
1484 
1485     return space;
1486 }
1487 
1488 void vfio_put_address_space(VFIOAddressSpace *space)
1489 {
1490     if (!QLIST_EMPTY(&space->containers)) {
1491         return;
1492     }
1493 
1494     QLIST_REMOVE(space, list);
1495     g_free(space);
1496 
1497     if (QLIST_EMPTY(&vfio_address_spaces)) {
1498         qemu_unregister_reset(vfio_reset_handler, NULL);
1499     }
1500 }
1501 
1502 struct vfio_device_info *vfio_get_device_info(int fd)
1503 {
1504     struct vfio_device_info *info;
1505     uint32_t argsz = sizeof(*info);
1506 
1507     info = g_malloc0(argsz);
1508 
1509 retry:
1510     info->argsz = argsz;
1511 
1512     if (ioctl(fd, VFIO_DEVICE_GET_INFO, info)) {
1513         g_free(info);
1514         return NULL;
1515     }
1516 
1517     if (info->argsz > argsz) {
1518         argsz = info->argsz;
1519         info = g_realloc(info, argsz);
1520         goto retry;
1521     }
1522 
1523     return info;
1524 }
1525 
1526 bool vfio_attach_device(char *name, VFIODevice *vbasedev,
1527                         AddressSpace *as, Error **errp)
1528 {
1529     const VFIOIOMMUClass *ops =
1530         VFIO_IOMMU_CLASS(object_class_by_name(TYPE_VFIO_IOMMU_LEGACY));
1531 
1532     if (vbasedev->iommufd) {
1533         ops = VFIO_IOMMU_CLASS(object_class_by_name(TYPE_VFIO_IOMMU_IOMMUFD));
1534     }
1535 
1536     assert(ops);
1537 
1538     return ops->attach_device(name, vbasedev, as, errp);
1539 }
1540 
1541 void vfio_detach_device(VFIODevice *vbasedev)
1542 {
1543     if (!vbasedev->bcontainer) {
1544         return;
1545     }
1546     vbasedev->bcontainer->ops->detach_device(vbasedev);
1547 }
1548