xref: /qemu/hw/vfio/listener.c (revision 06b40d250ecfa1633209c2e431a7a38acfd03a98)
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-device.h"
29 #include "hw/vfio/pci.h"
30 #include "system/address-spaces.h"
31 #include "system/memory.h"
32 #include "system/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 "system/kvm.h"
38 #include "system/reset.h"
39 #include "system/runstate.h"
40 #include "trace.h"
41 #include "qapi/error.h"
42 #include "migration/misc.h"
43 #include "migration/qemu-file.h"
44 #include "system/tcg.h"
45 #include "system/tpm.h"
46 #include "vfio-migration-internal.h"
47 #include "vfio-helpers.h"
48 #include "vfio-listener.h"
49 
50 /*
51  * Device state interfaces
52  */
53 
54 
55 static bool vfio_log_sync_needed(const VFIOContainerBase *bcontainer)
56 {
57     VFIODevice *vbasedev;
58 
59     if (!vfio_container_dirty_tracking_is_started(bcontainer)) {
60         return false;
61     }
62 
63     QLIST_FOREACH(vbasedev, &bcontainer->device_list, container_next) {
64         VFIOMigration *migration = vbasedev->migration;
65 
66         if (!migration) {
67             return false;
68         }
69 
70         if (vbasedev->pre_copy_dirty_page_tracking == ON_OFF_AUTO_OFF &&
71             (vfio_device_state_is_running(vbasedev) ||
72              vfio_device_state_is_precopy(vbasedev))) {
73             return false;
74         }
75     }
76     return true;
77 }
78 
79 static bool vfio_listener_skipped_section(MemoryRegionSection *section)
80 {
81     return (!memory_region_is_ram(section->mr) &&
82             !memory_region_is_iommu(section->mr)) ||
83            memory_region_is_protected(section->mr) ||
84            /*
85             * Sizing an enabled 64-bit BAR can cause spurious mappings to
86             * addresses in the upper part of the 64-bit address space.  These
87             * are never accessed by the CPU and beyond the address width of
88             * some IOMMU hardware.  TODO: VFIO should tell us the IOMMU width.
89             */
90            section->offset_within_address_space & (1ULL << 63);
91 }
92 
93 /* Called with rcu_read_lock held.  */
94 static bool vfio_get_xlat_addr(IOMMUTLBEntry *iotlb, void **vaddr,
95                                ram_addr_t *ram_addr, bool *read_only,
96                                Error **errp)
97 {
98     bool ret, mr_has_discard_manager;
99 
100     ret = memory_get_xlat_addr(iotlb, vaddr, ram_addr, read_only,
101                                &mr_has_discard_manager, errp);
102     if (ret && mr_has_discard_manager) {
103         /*
104          * Malicious VMs might trigger discarding of IOMMU-mapped memory. The
105          * pages will remain pinned inside vfio until unmapped, resulting in a
106          * higher memory consumption than expected. If memory would get
107          * populated again later, there would be an inconsistency between pages
108          * pinned by vfio and pages seen by QEMU. This is the case until
109          * unmapped from the IOMMU (e.g., during device reset).
110          *
111          * With malicious guests, we really only care about pinning more memory
112          * than expected. RLIMIT_MEMLOCK set for the user/process can never be
113          * exceeded and can be used to mitigate this problem.
114          */
115         warn_report_once("Using vfio with vIOMMUs and coordinated discarding of"
116                          " RAM (e.g., virtio-mem) works, however, malicious"
117                          " guests can trigger pinning of more memory than"
118                          " intended via an IOMMU. It's possible to mitigate "
119                          " by setting/adjusting RLIMIT_MEMLOCK.");
120     }
121     return ret;
122 }
123 
124 static void vfio_iommu_map_notify(IOMMUNotifier *n, IOMMUTLBEntry *iotlb)
125 {
126     VFIOGuestIOMMU *giommu = container_of(n, VFIOGuestIOMMU, n);
127     VFIOContainerBase *bcontainer = giommu->bcontainer;
128     hwaddr iova = iotlb->iova + giommu->iommu_offset;
129     void *vaddr;
130     int ret;
131     Error *local_err = NULL;
132 
133     trace_vfio_iommu_map_notify(iotlb->perm == IOMMU_NONE ? "UNMAP" : "MAP",
134                                 iova, iova + iotlb->addr_mask);
135 
136     if (iotlb->target_as != &address_space_memory) {
137         error_setg(&local_err,
138                    "Wrong target AS \"%s\", only system memory is allowed",
139                    iotlb->target_as->name ? iotlb->target_as->name : "none");
140         if (migration_is_running()) {
141             migration_file_set_error(-EINVAL, local_err);
142         } else {
143             error_report_err(local_err);
144         }
145         return;
146     }
147 
148     rcu_read_lock();
149 
150     if ((iotlb->perm & IOMMU_RW) != IOMMU_NONE) {
151         bool read_only;
152 
153         if (!vfio_get_xlat_addr(iotlb, &vaddr, NULL, &read_only, &local_err)) {
154             error_report_err(local_err);
155             goto out;
156         }
157         /*
158          * vaddr is only valid until rcu_read_unlock(). But after
159          * vfio_dma_map has set up the mapping the pages will be
160          * pinned by the kernel. This makes sure that the RAM backend
161          * of vaddr will always be there, even if the memory object is
162          * destroyed and its backing memory munmap-ed.
163          */
164         ret = vfio_container_dma_map(bcontainer, iova,
165                                      iotlb->addr_mask + 1, vaddr,
166                                      read_only);
167         if (ret) {
168             error_report("vfio_container_dma_map(%p, 0x%"HWADDR_PRIx", "
169                          "0x%"HWADDR_PRIx", %p) = %d (%s)",
170                          bcontainer, iova,
171                          iotlb->addr_mask + 1, vaddr, ret, strerror(-ret));
172         }
173     } else {
174         ret = vfio_container_dma_unmap(bcontainer, iova,
175                                        iotlb->addr_mask + 1, iotlb);
176         if (ret) {
177             error_setg(&local_err,
178                        "vfio_container_dma_unmap(%p, 0x%"HWADDR_PRIx", "
179                        "0x%"HWADDR_PRIx") = %d (%s)",
180                        bcontainer, iova,
181                        iotlb->addr_mask + 1, ret, strerror(-ret));
182             if (migration_is_running()) {
183                 migration_file_set_error(ret, local_err);
184             } else {
185                 error_report_err(local_err);
186             }
187         }
188     }
189 out:
190     rcu_read_unlock();
191 }
192 
193 static void vfio_ram_discard_notify_discard(RamDiscardListener *rdl,
194                                             MemoryRegionSection *section)
195 {
196     VFIORamDiscardListener *vrdl = container_of(rdl, VFIORamDiscardListener,
197                                                 listener);
198     VFIOContainerBase *bcontainer = vrdl->bcontainer;
199     const hwaddr size = int128_get64(section->size);
200     const hwaddr iova = section->offset_within_address_space;
201     int ret;
202 
203     /* Unmap with a single call. */
204     ret = vfio_container_dma_unmap(bcontainer, iova, size , NULL);
205     if (ret) {
206         error_report("%s: vfio_container_dma_unmap() failed: %s", __func__,
207                      strerror(-ret));
208     }
209 }
210 
211 static int vfio_ram_discard_notify_populate(RamDiscardListener *rdl,
212                                             MemoryRegionSection *section)
213 {
214     VFIORamDiscardListener *vrdl = container_of(rdl, VFIORamDiscardListener,
215                                                 listener);
216     VFIOContainerBase *bcontainer = vrdl->bcontainer;
217     const hwaddr end = section->offset_within_region +
218                        int128_get64(section->size);
219     hwaddr start, next, iova;
220     void *vaddr;
221     int ret;
222 
223     /*
224      * Map in (aligned within memory region) minimum granularity, so we can
225      * unmap in minimum granularity later.
226      */
227     for (start = section->offset_within_region; start < end; start = next) {
228         next = ROUND_UP(start + 1, vrdl->granularity);
229         next = MIN(next, end);
230 
231         iova = start - section->offset_within_region +
232                section->offset_within_address_space;
233         vaddr = memory_region_get_ram_ptr(section->mr) + start;
234 
235         ret = vfio_container_dma_map(bcontainer, iova, next - start,
236                                      vaddr, section->readonly);
237         if (ret) {
238             /* Rollback */
239             vfio_ram_discard_notify_discard(rdl, section);
240             return ret;
241         }
242     }
243     return 0;
244 }
245 
246 static void vfio_ram_discard_register_listener(VFIOContainerBase *bcontainer,
247                                                MemoryRegionSection *section)
248 {
249     RamDiscardManager *rdm = memory_region_get_ram_discard_manager(section->mr);
250     int target_page_size = qemu_target_page_size();
251     VFIORamDiscardListener *vrdl;
252 
253     /* Ignore some corner cases not relevant in practice. */
254     g_assert(QEMU_IS_ALIGNED(section->offset_within_region, target_page_size));
255     g_assert(QEMU_IS_ALIGNED(section->offset_within_address_space,
256                              target_page_size));
257     g_assert(QEMU_IS_ALIGNED(int128_get64(section->size), target_page_size));
258 
259     vrdl = g_new0(VFIORamDiscardListener, 1);
260     vrdl->bcontainer = bcontainer;
261     vrdl->mr = section->mr;
262     vrdl->offset_within_address_space = section->offset_within_address_space;
263     vrdl->size = int128_get64(section->size);
264     vrdl->granularity = ram_discard_manager_get_min_granularity(rdm,
265                                                                 section->mr);
266 
267     g_assert(vrdl->granularity && is_power_of_2(vrdl->granularity));
268     g_assert(bcontainer->pgsizes &&
269              vrdl->granularity >= 1ULL << ctz64(bcontainer->pgsizes));
270 
271     ram_discard_listener_init(&vrdl->listener,
272                               vfio_ram_discard_notify_populate,
273                               vfio_ram_discard_notify_discard, true);
274     ram_discard_manager_register_listener(rdm, &vrdl->listener, section);
275     QLIST_INSERT_HEAD(&bcontainer->vrdl_list, vrdl, next);
276 
277     /*
278      * Sanity-check if we have a theoretically problematic setup where we could
279      * exceed the maximum number of possible DMA mappings over time. We assume
280      * that each mapped section in the same address space as a RamDiscardManager
281      * section consumes exactly one DMA mapping, with the exception of
282      * RamDiscardManager sections; i.e., we don't expect to have gIOMMU sections
283      * in the same address space as RamDiscardManager sections.
284      *
285      * We assume that each section in the address space consumes one memslot.
286      * We take the number of KVM memory slots as a best guess for the maximum
287      * number of sections in the address space we could have over time,
288      * also consuming DMA mappings.
289      */
290     if (bcontainer->dma_max_mappings) {
291         unsigned int vrdl_count = 0, vrdl_mappings = 0, max_memslots = 512;
292 
293 #ifdef CONFIG_KVM
294         if (kvm_enabled()) {
295             max_memslots = kvm_get_max_memslots();
296         }
297 #endif
298 
299         QLIST_FOREACH(vrdl, &bcontainer->vrdl_list, next) {
300             hwaddr start, end;
301 
302             start = QEMU_ALIGN_DOWN(vrdl->offset_within_address_space,
303                                     vrdl->granularity);
304             end = ROUND_UP(vrdl->offset_within_address_space + vrdl->size,
305                            vrdl->granularity);
306             vrdl_mappings += (end - start) / vrdl->granularity;
307             vrdl_count++;
308         }
309 
310         if (vrdl_mappings + max_memslots - vrdl_count >
311             bcontainer->dma_max_mappings) {
312             warn_report("%s: possibly running out of DMA mappings. E.g., try"
313                         " increasing the 'block-size' of virtio-mem devies."
314                         " Maximum possible DMA mappings: %d, Maximum possible"
315                         " memslots: %d", __func__, bcontainer->dma_max_mappings,
316                         max_memslots);
317         }
318     }
319 }
320 
321 static void vfio_ram_discard_unregister_listener(VFIOContainerBase *bcontainer,
322                                                  MemoryRegionSection *section)
323 {
324     RamDiscardManager *rdm = memory_region_get_ram_discard_manager(section->mr);
325     VFIORamDiscardListener *vrdl = NULL;
326 
327     QLIST_FOREACH(vrdl, &bcontainer->vrdl_list, next) {
328         if (vrdl->mr == section->mr &&
329             vrdl->offset_within_address_space ==
330             section->offset_within_address_space) {
331             break;
332         }
333     }
334 
335     if (!vrdl) {
336         hw_error("vfio: Trying to unregister missing RAM discard listener");
337     }
338 
339     ram_discard_manager_unregister_listener(rdm, &vrdl->listener);
340     QLIST_REMOVE(vrdl, next);
341     g_free(vrdl);
342 }
343 
344 static bool vfio_known_safe_misalignment(MemoryRegionSection *section)
345 {
346     MemoryRegion *mr = section->mr;
347 
348     if (!TPM_IS_CRB(mr->owner)) {
349         return false;
350     }
351 
352     /* this is a known safe misaligned region, just trace for debug purpose */
353     trace_vfio_known_safe_misalignment(memory_region_name(mr),
354                                        section->offset_within_address_space,
355                                        section->offset_within_region,
356                                        qemu_real_host_page_size());
357     return true;
358 }
359 
360 static bool vfio_listener_valid_section(MemoryRegionSection *section,
361                                         const char *name)
362 {
363     if (vfio_listener_skipped_section(section)) {
364         trace_vfio_listener_region_skip(name,
365                 section->offset_within_address_space,
366                 section->offset_within_address_space +
367                 int128_get64(int128_sub(section->size, int128_one())));
368         return false;
369     }
370 
371     if (unlikely((section->offset_within_address_space &
372                   ~qemu_real_host_page_mask()) !=
373                  (section->offset_within_region & ~qemu_real_host_page_mask()))) {
374         if (!vfio_known_safe_misalignment(section)) {
375             error_report("%s received unaligned region %s iova=0x%"PRIx64
376                          " offset_within_region=0x%"PRIx64
377                          " qemu_real_host_page_size=0x%"PRIxPTR,
378                          __func__, memory_region_name(section->mr),
379                          section->offset_within_address_space,
380                          section->offset_within_region,
381                          qemu_real_host_page_size());
382         }
383         return false;
384     }
385 
386     return true;
387 }
388 
389 static bool vfio_get_section_iova_range(VFIOContainerBase *bcontainer,
390                                         MemoryRegionSection *section,
391                                         hwaddr *out_iova, hwaddr *out_end,
392                                         Int128 *out_llend)
393 {
394     Int128 llend;
395     hwaddr iova;
396 
397     iova = REAL_HOST_PAGE_ALIGN(section->offset_within_address_space);
398     llend = int128_make64(section->offset_within_address_space);
399     llend = int128_add(llend, section->size);
400     llend = int128_and(llend, int128_exts64(qemu_real_host_page_mask()));
401 
402     if (int128_ge(int128_make64(iova), llend)) {
403         return false;
404     }
405 
406     *out_iova = iova;
407     *out_end = int128_get64(int128_sub(llend, int128_one()));
408     if (out_llend) {
409         *out_llend = llend;
410     }
411     return true;
412 }
413 
414 static void vfio_device_error_append(VFIODevice *vbasedev, Error **errp)
415 {
416     /*
417      * MMIO region mapping failures are not fatal but in this case PCI
418      * peer-to-peer transactions are broken.
419      */
420     if (vbasedev && vbasedev->type == VFIO_DEVICE_TYPE_PCI) {
421         error_append_hint(errp, "%s: PCI peer-to-peer transactions "
422                           "on BARs are not supported.\n", vbasedev->name);
423     }
424 }
425 
426 static void vfio_listener_region_add(MemoryListener *listener,
427                                      MemoryRegionSection *section)
428 {
429     VFIOContainerBase *bcontainer = container_of(listener, VFIOContainerBase,
430                                                  listener);
431     hwaddr iova, end;
432     Int128 llend, llsize;
433     void *vaddr;
434     int ret;
435     Error *err = NULL;
436 
437     if (!vfio_listener_valid_section(section, "region_add")) {
438         return;
439     }
440 
441     if (!vfio_get_section_iova_range(bcontainer, section, &iova, &end,
442                                      &llend)) {
443         if (memory_region_is_ram_device(section->mr)) {
444             trace_vfio_listener_region_add_no_dma_map(
445                 memory_region_name(section->mr),
446                 section->offset_within_address_space,
447                 int128_getlo(section->size),
448                 qemu_real_host_page_size());
449         }
450         return;
451     }
452 
453     /* PPC64/pseries machine only */
454     if (!vfio_container_add_section_window(bcontainer, section, &err)) {
455         goto mmio_dma_error;
456     }
457 
458     memory_region_ref(section->mr);
459 
460     if (memory_region_is_iommu(section->mr)) {
461         VFIOGuestIOMMU *giommu;
462         IOMMUMemoryRegion *iommu_mr = IOMMU_MEMORY_REGION(section->mr);
463         int iommu_idx;
464 
465         trace_vfio_listener_region_add_iommu(section->mr->name, iova, end);
466         /*
467          * FIXME: For VFIO iommu types which have KVM acceleration to
468          * avoid bouncing all map/unmaps through qemu this way, this
469          * would be the right place to wire that up (tell the KVM
470          * device emulation the VFIO iommu handles to use).
471          */
472         giommu = g_malloc0(sizeof(*giommu));
473         giommu->iommu_mr = iommu_mr;
474         giommu->iommu_offset = section->offset_within_address_space -
475                                section->offset_within_region;
476         giommu->bcontainer = bcontainer;
477         llend = int128_add(int128_make64(section->offset_within_region),
478                            section->size);
479         llend = int128_sub(llend, int128_one());
480         iommu_idx = memory_region_iommu_attrs_to_index(iommu_mr,
481                                                        MEMTXATTRS_UNSPECIFIED);
482         iommu_notifier_init(&giommu->n, vfio_iommu_map_notify,
483                             IOMMU_NOTIFIER_IOTLB_EVENTS,
484                             section->offset_within_region,
485                             int128_get64(llend),
486                             iommu_idx);
487 
488         ret = memory_region_register_iommu_notifier(section->mr, &giommu->n,
489                                                     &err);
490         if (ret) {
491             g_free(giommu);
492             goto fail;
493         }
494         QLIST_INSERT_HEAD(&bcontainer->giommu_list, giommu, giommu_next);
495         memory_region_iommu_replay(giommu->iommu_mr, &giommu->n);
496 
497         return;
498     }
499 
500     /* Here we assume that memory_region_is_ram(section->mr)==true */
501 
502     /*
503      * For RAM memory regions with a RamDiscardManager, we only want to map the
504      * actually populated parts - and update the mapping whenever we're notified
505      * about changes.
506      */
507     if (memory_region_has_ram_discard_manager(section->mr)) {
508         vfio_ram_discard_register_listener(bcontainer, section);
509         return;
510     }
511 
512     vaddr = memory_region_get_ram_ptr(section->mr) +
513             section->offset_within_region +
514             (iova - section->offset_within_address_space);
515 
516     trace_vfio_listener_region_add_ram(iova, end, vaddr);
517 
518     llsize = int128_sub(llend, int128_make64(iova));
519 
520     if (memory_region_is_ram_device(section->mr)) {
521         hwaddr pgmask = (1ULL << ctz64(bcontainer->pgsizes)) - 1;
522 
523         if ((iova & pgmask) || (int128_get64(llsize) & pgmask)) {
524             trace_vfio_listener_region_add_no_dma_map(
525                 memory_region_name(section->mr),
526                 section->offset_within_address_space,
527                 int128_getlo(section->size),
528                 pgmask + 1);
529             return;
530         }
531     }
532 
533     ret = vfio_container_dma_map(bcontainer, iova, int128_get64(llsize),
534                                  vaddr, section->readonly);
535     if (ret) {
536         error_setg(&err, "vfio_container_dma_map(%p, 0x%"HWADDR_PRIx", "
537                    "0x%"HWADDR_PRIx", %p) = %d (%s)",
538                    bcontainer, iova, int128_get64(llsize), vaddr, ret,
539                    strerror(-ret));
540     mmio_dma_error:
541         if (memory_region_is_ram_device(section->mr)) {
542             /* Allow unexpected mappings not to be fatal for RAM devices */
543             VFIODevice *vbasedev =
544                 vfio_get_vfio_device(memory_region_owner(section->mr));
545             vfio_device_error_append(vbasedev, &err);
546             warn_report_err_once(err);
547             return;
548         }
549         goto fail;
550     }
551 
552     return;
553 
554 fail:
555     if (!bcontainer->initialized) {
556         /*
557          * At machine init time or when the device is attached to the
558          * VM, store the first error in the container so we can
559          * gracefully fail the device realize routine.
560          */
561         if (!bcontainer->error) {
562             error_propagate_prepend(&bcontainer->error, err,
563                                     "Region %s: ",
564                                     memory_region_name(section->mr));
565         } else {
566             error_free(err);
567         }
568     } else {
569         /*
570          * At runtime, there's not much we can do other than throw a
571          * hardware error.
572          */
573         error_report_err(err);
574         hw_error("vfio: DMA mapping failed, unable to continue");
575     }
576 }
577 
578 static void vfio_listener_region_del(MemoryListener *listener,
579                                      MemoryRegionSection *section)
580 {
581     VFIOContainerBase *bcontainer = container_of(listener, VFIOContainerBase,
582                                                  listener);
583     hwaddr iova, end;
584     Int128 llend, llsize;
585     int ret;
586     bool try_unmap = true;
587 
588     if (!vfio_listener_valid_section(section, "region_del")) {
589         return;
590     }
591 
592     if (memory_region_is_iommu(section->mr)) {
593         VFIOGuestIOMMU *giommu;
594 
595         trace_vfio_listener_region_del_iommu(section->mr->name);
596         QLIST_FOREACH(giommu, &bcontainer->giommu_list, giommu_next) {
597             if (MEMORY_REGION(giommu->iommu_mr) == section->mr &&
598                 giommu->n.start == section->offset_within_region) {
599                 memory_region_unregister_iommu_notifier(section->mr,
600                                                         &giommu->n);
601                 QLIST_REMOVE(giommu, giommu_next);
602                 g_free(giommu);
603                 break;
604             }
605         }
606 
607         /*
608          * FIXME: We assume the one big unmap below is adequate to
609          * remove any individual page mappings in the IOMMU which
610          * might have been copied into VFIO. This works for a page table
611          * based IOMMU where a big unmap flattens a large range of IO-PTEs.
612          * That may not be true for all IOMMU types.
613          */
614     }
615 
616     if (!vfio_get_section_iova_range(bcontainer, section, &iova, &end,
617                                      &llend)) {
618         return;
619     }
620 
621     llsize = int128_sub(llend, int128_make64(iova));
622 
623     trace_vfio_listener_region_del(iova, end);
624 
625     if (memory_region_is_ram_device(section->mr)) {
626         hwaddr pgmask;
627 
628         pgmask = (1ULL << ctz64(bcontainer->pgsizes)) - 1;
629         try_unmap = !((iova & pgmask) || (int128_get64(llsize) & pgmask));
630     } else if (memory_region_has_ram_discard_manager(section->mr)) {
631         vfio_ram_discard_unregister_listener(bcontainer, section);
632         /* Unregistering will trigger an unmap. */
633         try_unmap = false;
634     }
635 
636     if (try_unmap) {
637         if (int128_eq(llsize, int128_2_64())) {
638             /* The unmap ioctl doesn't accept a full 64-bit span. */
639             llsize = int128_rshift(llsize, 1);
640             ret = vfio_container_dma_unmap(bcontainer, iova,
641                                            int128_get64(llsize), NULL);
642             if (ret) {
643                 error_report("vfio_container_dma_unmap(%p, 0x%"HWADDR_PRIx", "
644                              "0x%"HWADDR_PRIx") = %d (%s)",
645                              bcontainer, iova, int128_get64(llsize), ret,
646                              strerror(-ret));
647             }
648             iova += int128_get64(llsize);
649         }
650         ret = vfio_container_dma_unmap(bcontainer, iova,
651                                        int128_get64(llsize), NULL);
652         if (ret) {
653             error_report("vfio_container_dma_unmap(%p, 0x%"HWADDR_PRIx", "
654                          "0x%"HWADDR_PRIx") = %d (%s)",
655                          bcontainer, iova, int128_get64(llsize), ret,
656                          strerror(-ret));
657         }
658     }
659 
660     memory_region_unref(section->mr);
661 
662     /* PPC64/pseries machine only */
663     vfio_container_del_section_window(bcontainer, section);
664 }
665 
666 typedef struct VFIODirtyRanges {
667     hwaddr min32;
668     hwaddr max32;
669     hwaddr min64;
670     hwaddr max64;
671     hwaddr minpci64;
672     hwaddr maxpci64;
673 } VFIODirtyRanges;
674 
675 typedef struct VFIODirtyRangesListener {
676     VFIOContainerBase *bcontainer;
677     VFIODirtyRanges ranges;
678     MemoryListener listener;
679 } VFIODirtyRangesListener;
680 
681 static bool vfio_section_is_vfio_pci(MemoryRegionSection *section,
682                                      VFIOContainerBase *bcontainer)
683 {
684     VFIOPCIDevice *pcidev;
685     VFIODevice *vbasedev;
686     Object *owner;
687 
688     owner = memory_region_owner(section->mr);
689 
690     QLIST_FOREACH(vbasedev, &bcontainer->device_list, container_next) {
691         if (vbasedev->type != VFIO_DEVICE_TYPE_PCI) {
692             continue;
693         }
694         pcidev = container_of(vbasedev, VFIOPCIDevice, vbasedev);
695         if (OBJECT(pcidev) == owner) {
696             return true;
697         }
698     }
699 
700     return false;
701 }
702 
703 static void vfio_dirty_tracking_update_range(VFIODirtyRanges *range,
704                                              hwaddr iova, hwaddr end,
705                                              bool update_pci)
706 {
707     hwaddr *min, *max;
708 
709     /*
710      * The address space passed to the dirty tracker is reduced to three ranges:
711      * one for 32-bit DMA ranges, one for 64-bit DMA ranges and one for the
712      * PCI 64-bit hole.
713      *
714      * The underlying reports of dirty will query a sub-interval of each of
715      * these ranges.
716      *
717      * The purpose of the three range handling is to handle known cases of big
718      * holes in the address space, like the x86 AMD 1T hole, and firmware (like
719      * OVMF) which may relocate the pci-hole64 to the end of the address space.
720      * The latter would otherwise generate large ranges for tracking, stressing
721      * the limits of supported hardware. The pci-hole32 will always be below 4G
722      * (overlapping or not) so it doesn't need special handling and is part of
723      * the 32-bit range.
724      *
725      * The alternative would be an IOVATree but that has a much bigger runtime
726      * overhead and unnecessary complexity.
727      */
728     if (update_pci && iova >= UINT32_MAX) {
729         min = &range->minpci64;
730         max = &range->maxpci64;
731     } else {
732         min = (end <= UINT32_MAX) ? &range->min32 : &range->min64;
733         max = (end <= UINT32_MAX) ? &range->max32 : &range->max64;
734     }
735     if (*min > iova) {
736         *min = iova;
737     }
738     if (*max < end) {
739         *max = end;
740     }
741 
742     trace_vfio_device_dirty_tracking_update(iova, end, *min, *max);
743 }
744 
745 static void vfio_dirty_tracking_update(MemoryListener *listener,
746                                        MemoryRegionSection *section)
747 {
748     VFIODirtyRangesListener *dirty =
749         container_of(listener, VFIODirtyRangesListener, listener);
750     hwaddr iova, end;
751 
752     if (!vfio_listener_valid_section(section, "tracking_update") ||
753         !vfio_get_section_iova_range(dirty->bcontainer, section,
754                                      &iova, &end, NULL)) {
755         return;
756     }
757 
758     vfio_dirty_tracking_update_range(&dirty->ranges, iova, end,
759                       vfio_section_is_vfio_pci(section, dirty->bcontainer));
760 }
761 
762 static const MemoryListener vfio_dirty_tracking_listener = {
763     .name = "vfio-tracking",
764     .region_add = vfio_dirty_tracking_update,
765 };
766 
767 static void vfio_dirty_tracking_init(VFIOContainerBase *bcontainer,
768                                      VFIODirtyRanges *ranges)
769 {
770     VFIODirtyRangesListener dirty;
771 
772     memset(&dirty, 0, sizeof(dirty));
773     dirty.ranges.min32 = UINT32_MAX;
774     dirty.ranges.min64 = UINT64_MAX;
775     dirty.ranges.minpci64 = UINT64_MAX;
776     dirty.listener = vfio_dirty_tracking_listener;
777     dirty.bcontainer = bcontainer;
778 
779     memory_listener_register(&dirty.listener,
780                              bcontainer->space->as);
781 
782     *ranges = dirty.ranges;
783 
784     /*
785      * The memory listener is synchronous, and used to calculate the range
786      * to dirty tracking. Unregister it after we are done as we are not
787      * interested in any follow-up updates.
788      */
789     memory_listener_unregister(&dirty.listener);
790 }
791 
792 static void vfio_devices_dma_logging_stop(VFIOContainerBase *bcontainer)
793 {
794     uint64_t buf[DIV_ROUND_UP(sizeof(struct vfio_device_feature),
795                               sizeof(uint64_t))] = {};
796     struct vfio_device_feature *feature = (struct vfio_device_feature *)buf;
797     VFIODevice *vbasedev;
798 
799     feature->argsz = sizeof(buf);
800     feature->flags = VFIO_DEVICE_FEATURE_SET |
801                      VFIO_DEVICE_FEATURE_DMA_LOGGING_STOP;
802 
803     QLIST_FOREACH(vbasedev, &bcontainer->device_list, container_next) {
804         if (!vbasedev->dirty_tracking) {
805             continue;
806         }
807 
808         if (ioctl(vbasedev->fd, VFIO_DEVICE_FEATURE, feature)) {
809             warn_report("%s: Failed to stop DMA logging, err %d (%s)",
810                         vbasedev->name, -errno, strerror(errno));
811         }
812         vbasedev->dirty_tracking = false;
813     }
814 }
815 
816 static struct vfio_device_feature *
817 vfio_device_feature_dma_logging_start_create(VFIOContainerBase *bcontainer,
818                                              VFIODirtyRanges *tracking)
819 {
820     struct vfio_device_feature *feature;
821     size_t feature_size;
822     struct vfio_device_feature_dma_logging_control *control;
823     struct vfio_device_feature_dma_logging_range *ranges;
824 
825     feature_size = sizeof(struct vfio_device_feature) +
826                    sizeof(struct vfio_device_feature_dma_logging_control);
827     feature = g_try_malloc0(feature_size);
828     if (!feature) {
829         errno = ENOMEM;
830         return NULL;
831     }
832     feature->argsz = feature_size;
833     feature->flags = VFIO_DEVICE_FEATURE_SET |
834                      VFIO_DEVICE_FEATURE_DMA_LOGGING_START;
835 
836     control = (struct vfio_device_feature_dma_logging_control *)feature->data;
837     control->page_size = qemu_real_host_page_size();
838 
839     /*
840      * DMA logging uAPI guarantees to support at least a number of ranges that
841      * fits into a single host kernel base page.
842      */
843     control->num_ranges = !!tracking->max32 + !!tracking->max64 +
844         !!tracking->maxpci64;
845     ranges = g_try_new0(struct vfio_device_feature_dma_logging_range,
846                         control->num_ranges);
847     if (!ranges) {
848         g_free(feature);
849         errno = ENOMEM;
850 
851         return NULL;
852     }
853 
854     control->ranges = (uintptr_t)ranges;
855     if (tracking->max32) {
856         ranges->iova = tracking->min32;
857         ranges->length = (tracking->max32 - tracking->min32) + 1;
858         ranges++;
859     }
860     if (tracking->max64) {
861         ranges->iova = tracking->min64;
862         ranges->length = (tracking->max64 - tracking->min64) + 1;
863         ranges++;
864     }
865     if (tracking->maxpci64) {
866         ranges->iova = tracking->minpci64;
867         ranges->length = (tracking->maxpci64 - tracking->minpci64) + 1;
868     }
869 
870     trace_vfio_device_dirty_tracking_start(control->num_ranges,
871                                            tracking->min32, tracking->max32,
872                                            tracking->min64, tracking->max64,
873                                            tracking->minpci64, tracking->maxpci64);
874 
875     return feature;
876 }
877 
878 static void vfio_device_feature_dma_logging_start_destroy(
879     struct vfio_device_feature *feature)
880 {
881     struct vfio_device_feature_dma_logging_control *control =
882         (struct vfio_device_feature_dma_logging_control *)feature->data;
883     struct vfio_device_feature_dma_logging_range *ranges =
884         (struct vfio_device_feature_dma_logging_range *)(uintptr_t)control->ranges;
885 
886     g_free(ranges);
887     g_free(feature);
888 }
889 
890 static bool vfio_devices_dma_logging_start(VFIOContainerBase *bcontainer,
891                                           Error **errp)
892 {
893     struct vfio_device_feature *feature;
894     VFIODirtyRanges ranges;
895     VFIODevice *vbasedev;
896     int ret = 0;
897 
898     vfio_dirty_tracking_init(bcontainer, &ranges);
899     feature = vfio_device_feature_dma_logging_start_create(bcontainer,
900                                                            &ranges);
901     if (!feature) {
902         error_setg_errno(errp, errno, "Failed to prepare DMA logging");
903         return false;
904     }
905 
906     QLIST_FOREACH(vbasedev, &bcontainer->device_list, container_next) {
907         if (vbasedev->dirty_tracking) {
908             continue;
909         }
910 
911         ret = ioctl(vbasedev->fd, VFIO_DEVICE_FEATURE, feature);
912         if (ret) {
913             ret = -errno;
914             error_setg_errno(errp, errno, "%s: Failed to start DMA logging",
915                              vbasedev->name);
916             goto out;
917         }
918         vbasedev->dirty_tracking = true;
919     }
920 
921 out:
922     if (ret) {
923         vfio_devices_dma_logging_stop(bcontainer);
924     }
925 
926     vfio_device_feature_dma_logging_start_destroy(feature);
927 
928     return ret == 0;
929 }
930 
931 static bool vfio_listener_log_global_start(MemoryListener *listener,
932                                            Error **errp)
933 {
934     ERRP_GUARD();
935     VFIOContainerBase *bcontainer = container_of(listener, VFIOContainerBase,
936                                                  listener);
937     bool ret;
938 
939     if (vfio_container_devices_dirty_tracking_is_supported(bcontainer)) {
940         ret = vfio_devices_dma_logging_start(bcontainer, errp);
941     } else {
942         ret = vfio_container_set_dirty_page_tracking(bcontainer, true, errp) == 0;
943     }
944 
945     if (!ret) {
946         error_prepend(errp, "vfio: Could not start dirty page tracking - ");
947     }
948     return ret;
949 }
950 
951 static void vfio_listener_log_global_stop(MemoryListener *listener)
952 {
953     VFIOContainerBase *bcontainer = container_of(listener, VFIOContainerBase,
954                                                  listener);
955     Error *local_err = NULL;
956     int ret = 0;
957 
958     if (vfio_container_devices_dirty_tracking_is_supported(bcontainer)) {
959         vfio_devices_dma_logging_stop(bcontainer);
960     } else {
961         ret = vfio_container_set_dirty_page_tracking(bcontainer, false,
962                                                      &local_err);
963     }
964 
965     if (ret) {
966         error_prepend(&local_err,
967                       "vfio: Could not stop dirty page tracking - ");
968         if (migration_is_running()) {
969             migration_file_set_error(ret, local_err);
970         } else {
971             error_report_err(local_err);
972         }
973     }
974 }
975 
976 typedef struct {
977     IOMMUNotifier n;
978     VFIOGuestIOMMU *giommu;
979 } vfio_giommu_dirty_notifier;
980 
981 static void vfio_iommu_map_dirty_notify(IOMMUNotifier *n, IOMMUTLBEntry *iotlb)
982 {
983     vfio_giommu_dirty_notifier *gdn = container_of(n,
984                                                 vfio_giommu_dirty_notifier, n);
985     VFIOGuestIOMMU *giommu = gdn->giommu;
986     VFIOContainerBase *bcontainer = giommu->bcontainer;
987     hwaddr iova = iotlb->iova + giommu->iommu_offset;
988     ram_addr_t translated_addr;
989     Error *local_err = NULL;
990     int ret = -EINVAL;
991 
992     trace_vfio_iommu_map_dirty_notify(iova, iova + iotlb->addr_mask);
993 
994     if (iotlb->target_as != &address_space_memory) {
995         error_setg(&local_err,
996                    "Wrong target AS \"%s\", only system memory is allowed",
997                    iotlb->target_as->name ? iotlb->target_as->name : "none");
998         goto out;
999     }
1000 
1001     rcu_read_lock();
1002     if (!vfio_get_xlat_addr(iotlb, NULL, &translated_addr, NULL, &local_err)) {
1003         goto out_unlock;
1004     }
1005 
1006     ret = vfio_container_query_dirty_bitmap(bcontainer, iova, iotlb->addr_mask + 1,
1007                                 translated_addr, &local_err);
1008     if (ret) {
1009         error_prepend(&local_err,
1010                       "vfio_iommu_map_dirty_notify(%p, 0x%"HWADDR_PRIx", "
1011                       "0x%"HWADDR_PRIx") failed - ", bcontainer, iova,
1012                       iotlb->addr_mask + 1);
1013     }
1014 
1015 out_unlock:
1016     rcu_read_unlock();
1017 
1018 out:
1019     if (ret) {
1020         if (migration_is_running()) {
1021             migration_file_set_error(ret, local_err);
1022         } else {
1023             error_report_err(local_err);
1024         }
1025     }
1026 }
1027 
1028 static int vfio_ram_discard_query_dirty_bitmap(MemoryRegionSection *section,
1029                                              void *opaque)
1030 {
1031     const hwaddr size = int128_get64(section->size);
1032     const hwaddr iova = section->offset_within_address_space;
1033     const ram_addr_t ram_addr = memory_region_get_ram_addr(section->mr) +
1034                                 section->offset_within_region;
1035     VFIORamDiscardListener *vrdl = opaque;
1036     Error *local_err = NULL;
1037     int ret;
1038 
1039     /*
1040      * Sync the whole mapped region (spanning multiple individual mappings)
1041      * in one go.
1042      */
1043     ret = vfio_container_query_dirty_bitmap(vrdl->bcontainer, iova, size, ram_addr,
1044                                 &local_err);
1045     if (ret) {
1046         error_report_err(local_err);
1047     }
1048     return ret;
1049 }
1050 
1051 static int
1052 vfio_sync_ram_discard_listener_dirty_bitmap(VFIOContainerBase *bcontainer,
1053                                             MemoryRegionSection *section)
1054 {
1055     RamDiscardManager *rdm = memory_region_get_ram_discard_manager(section->mr);
1056     VFIORamDiscardListener *vrdl = NULL;
1057 
1058     QLIST_FOREACH(vrdl, &bcontainer->vrdl_list, next) {
1059         if (vrdl->mr == section->mr &&
1060             vrdl->offset_within_address_space ==
1061             section->offset_within_address_space) {
1062             break;
1063         }
1064     }
1065 
1066     if (!vrdl) {
1067         hw_error("vfio: Trying to sync missing RAM discard listener");
1068     }
1069 
1070     /*
1071      * We only want/can synchronize the bitmap for actually mapped parts -
1072      * which correspond to populated parts. Replay all populated parts.
1073      */
1074     return ram_discard_manager_replay_populated(rdm, section,
1075                                                 vfio_ram_discard_query_dirty_bitmap,
1076                                                 &vrdl);
1077 }
1078 
1079 static int vfio_sync_iommu_dirty_bitmap(VFIOContainerBase *bcontainer,
1080                                         MemoryRegionSection *section)
1081 {
1082     VFIOGuestIOMMU *giommu;
1083     bool found = false;
1084     Int128 llend;
1085     vfio_giommu_dirty_notifier gdn;
1086     int idx;
1087 
1088     QLIST_FOREACH(giommu, &bcontainer->giommu_list, giommu_next) {
1089         if (MEMORY_REGION(giommu->iommu_mr) == section->mr &&
1090             giommu->n.start == section->offset_within_region) {
1091             found = true;
1092             break;
1093         }
1094     }
1095 
1096     if (!found) {
1097         return 0;
1098     }
1099 
1100     gdn.giommu = giommu;
1101     idx = memory_region_iommu_attrs_to_index(giommu->iommu_mr,
1102                                              MEMTXATTRS_UNSPECIFIED);
1103 
1104     llend = int128_add(int128_make64(section->offset_within_region),
1105                        section->size);
1106     llend = int128_sub(llend, int128_one());
1107 
1108     iommu_notifier_init(&gdn.n, vfio_iommu_map_dirty_notify, IOMMU_NOTIFIER_MAP,
1109                         section->offset_within_region, int128_get64(llend),
1110                         idx);
1111     memory_region_iommu_replay(giommu->iommu_mr, &gdn.n);
1112 
1113     return 0;
1114 }
1115 
1116 static int vfio_sync_dirty_bitmap(VFIOContainerBase *bcontainer,
1117                                   MemoryRegionSection *section, Error **errp)
1118 {
1119     ram_addr_t ram_addr;
1120 
1121     if (memory_region_is_iommu(section->mr)) {
1122         return vfio_sync_iommu_dirty_bitmap(bcontainer, section);
1123     } else if (memory_region_has_ram_discard_manager(section->mr)) {
1124         int ret;
1125 
1126         ret = vfio_sync_ram_discard_listener_dirty_bitmap(bcontainer, section);
1127         if (ret) {
1128             error_setg(errp,
1129                        "Failed to sync dirty bitmap with RAM discard listener");
1130         }
1131         return ret;
1132     }
1133 
1134     ram_addr = memory_region_get_ram_addr(section->mr) +
1135                section->offset_within_region;
1136 
1137     return vfio_container_query_dirty_bitmap(bcontainer,
1138                    REAL_HOST_PAGE_ALIGN(section->offset_within_address_space),
1139                                  int128_get64(section->size), ram_addr, errp);
1140 }
1141 
1142 static void vfio_listener_log_sync(MemoryListener *listener,
1143         MemoryRegionSection *section)
1144 {
1145     VFIOContainerBase *bcontainer = container_of(listener, VFIOContainerBase,
1146                                                  listener);
1147     int ret;
1148     Error *local_err = NULL;
1149 
1150     if (vfio_listener_skipped_section(section)) {
1151         return;
1152     }
1153 
1154     if (vfio_log_sync_needed(bcontainer)) {
1155         ret = vfio_sync_dirty_bitmap(bcontainer, section, &local_err);
1156         if (ret) {
1157             if (migration_is_running()) {
1158                 migration_file_set_error(ret, local_err);
1159             } else {
1160                 error_report_err(local_err);
1161             }
1162         }
1163     }
1164 }
1165 
1166 static const MemoryListener vfio_memory_listener = {
1167     .name = "vfio",
1168     .region_add = vfio_listener_region_add,
1169     .region_del = vfio_listener_region_del,
1170     .log_global_start = vfio_listener_log_global_start,
1171     .log_global_stop = vfio_listener_log_global_stop,
1172     .log_sync = vfio_listener_log_sync,
1173 };
1174 
1175 bool vfio_listener_register(VFIOContainerBase *bcontainer, Error **errp)
1176 {
1177     bcontainer->listener = vfio_memory_listener;
1178     memory_listener_register(&bcontainer->listener, bcontainer->space->as);
1179 
1180     if (bcontainer->error) {
1181         error_propagate_prepend(errp, bcontainer->error,
1182                                 "memory listener initialization failed: ");
1183         return false;
1184     }
1185 
1186     return true;
1187 }
1188 
1189 void vfio_listener_unregister(VFIOContainerBase *bcontainer)
1190 {
1191     memory_listener_unregister(&bcontainer->listener);
1192 }
1193