xref: /openbmc/qemu/hw/vfio/common.c (revision 75bbe5e5)
1 /*
2  * generic functions used by VFIO devices
3  *
4  * Copyright Red Hat, Inc. 2012
5  *
6  * Authors:
7  *  Alex Williamson <alex.williamson@redhat.com>
8  *
9  * This work is licensed under the terms of the GNU GPL, version 2.  See
10  * the COPYING file in the top-level directory.
11  *
12  * Based on qemu-kvm device-assignment:
13  *  Adapted for KVM by Qumranet.
14  *  Copyright (c) 2007, Neocleus, Alex Novik (alex@neocleus.com)
15  *  Copyright (c) 2007, Neocleus, Guy Zana (guy@neocleus.com)
16  *  Copyright (C) 2008, Qumranet, Amit Shah (amit.shah@qumranet.com)
17  *  Copyright (C) 2008, Red Hat, Amit Shah (amit.shah@redhat.com)
18  *  Copyright (C) 2008, IBM, Muli Ben-Yehuda (muli@il.ibm.com)
19  */
20 
21 #include "qemu/osdep.h"
22 #include <sys/ioctl.h>
23 #ifdef CONFIG_KVM
24 #include <linux/kvm.h>
25 #endif
26 #include <linux/vfio.h>
27 
28 #include "hw/vfio/vfio-common.h"
29 #include "hw/vfio/vfio.h"
30 #include "exec/address-spaces.h"
31 #include "exec/memory.h"
32 #include "exec/ram_addr.h"
33 #include "hw/hw.h"
34 #include "qemu/error-report.h"
35 #include "qemu/main-loop.h"
36 #include "qemu/range.h"
37 #include "sysemu/kvm.h"
38 #include "sysemu/reset.h"
39 #include "sysemu/runstate.h"
40 #include "trace.h"
41 #include "qapi/error.h"
42 #include "migration/migration.h"
43 #include "sysemu/tpm.h"
44 
45 VFIOGroupList vfio_group_list =
46     QLIST_HEAD_INITIALIZER(vfio_group_list);
47 static QLIST_HEAD(, VFIOAddressSpace) vfio_address_spaces =
48     QLIST_HEAD_INITIALIZER(vfio_address_spaces);
49 
50 #ifdef CONFIG_KVM
51 /*
52  * We have a single VFIO pseudo device per KVM VM.  Once created it lives
53  * for the life of the VM.  Closing the file descriptor only drops our
54  * reference to it and the device's reference to kvm.  Therefore once
55  * initialized, this file descriptor is only released on QEMU exit and
56  * we'll re-use it should another vfio device be attached before then.
57  */
58 static int vfio_kvm_device_fd = -1;
59 #endif
60 
61 /*
62  * Common VFIO interrupt disable
63  */
64 void vfio_disable_irqindex(VFIODevice *vbasedev, int index)
65 {
66     struct vfio_irq_set irq_set = {
67         .argsz = sizeof(irq_set),
68         .flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_TRIGGER,
69         .index = index,
70         .start = 0,
71         .count = 0,
72     };
73 
74     ioctl(vbasedev->fd, VFIO_DEVICE_SET_IRQS, &irq_set);
75 }
76 
77 void vfio_unmask_single_irqindex(VFIODevice *vbasedev, int index)
78 {
79     struct vfio_irq_set irq_set = {
80         .argsz = sizeof(irq_set),
81         .flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_UNMASK,
82         .index = index,
83         .start = 0,
84         .count = 1,
85     };
86 
87     ioctl(vbasedev->fd, VFIO_DEVICE_SET_IRQS, &irq_set);
88 }
89 
90 void vfio_mask_single_irqindex(VFIODevice *vbasedev, int index)
91 {
92     struct vfio_irq_set irq_set = {
93         .argsz = sizeof(irq_set),
94         .flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_MASK,
95         .index = index,
96         .start = 0,
97         .count = 1,
98     };
99 
100     ioctl(vbasedev->fd, VFIO_DEVICE_SET_IRQS, &irq_set);
101 }
102 
103 static inline const char *action_to_str(int action)
104 {
105     switch (action) {
106     case VFIO_IRQ_SET_ACTION_MASK:
107         return "MASK";
108     case VFIO_IRQ_SET_ACTION_UNMASK:
109         return "UNMASK";
110     case VFIO_IRQ_SET_ACTION_TRIGGER:
111         return "TRIGGER";
112     default:
113         return "UNKNOWN ACTION";
114     }
115 }
116 
117 static const char *index_to_str(VFIODevice *vbasedev, int index)
118 {
119     if (vbasedev->type != VFIO_DEVICE_TYPE_PCI) {
120         return NULL;
121     }
122 
123     switch (index) {
124     case VFIO_PCI_INTX_IRQ_INDEX:
125         return "INTX";
126     case VFIO_PCI_MSI_IRQ_INDEX:
127         return "MSI";
128     case VFIO_PCI_MSIX_IRQ_INDEX:
129         return "MSIX";
130     case VFIO_PCI_ERR_IRQ_INDEX:
131         return "ERR";
132     case VFIO_PCI_REQ_IRQ_INDEX:
133         return "REQ";
134     default:
135         return NULL;
136     }
137 }
138 
139 static int vfio_ram_block_discard_disable(VFIOContainer *container, bool state)
140 {
141     switch (container->iommu_type) {
142     case VFIO_TYPE1v2_IOMMU:
143     case VFIO_TYPE1_IOMMU:
144         /*
145          * We support coordinated discarding of RAM via the RamDiscardManager.
146          */
147         return ram_block_uncoordinated_discard_disable(state);
148     default:
149         /*
150          * VFIO_SPAPR_TCE_IOMMU most probably works just fine with
151          * RamDiscardManager, however, it is completely untested.
152          *
153          * VFIO_SPAPR_TCE_v2_IOMMU with "DMA memory preregistering" does
154          * completely the opposite of managing mapping/pinning dynamically as
155          * required by RamDiscardManager. We would have to special-case sections
156          * with a RamDiscardManager.
157          */
158         return ram_block_discard_disable(state);
159     }
160 }
161 
162 int vfio_set_irq_signaling(VFIODevice *vbasedev, int index, int subindex,
163                            int action, int fd, Error **errp)
164 {
165     struct vfio_irq_set *irq_set;
166     int argsz, ret = 0;
167     const char *name;
168     int32_t *pfd;
169 
170     argsz = sizeof(*irq_set) + sizeof(*pfd);
171 
172     irq_set = g_malloc0(argsz);
173     irq_set->argsz = argsz;
174     irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD | action;
175     irq_set->index = index;
176     irq_set->start = subindex;
177     irq_set->count = 1;
178     pfd = (int32_t *)&irq_set->data;
179     *pfd = fd;
180 
181     if (ioctl(vbasedev->fd, VFIO_DEVICE_SET_IRQS, irq_set)) {
182         ret = -errno;
183     }
184     g_free(irq_set);
185 
186     if (!ret) {
187         return 0;
188     }
189 
190     error_setg_errno(errp, -ret, "VFIO_DEVICE_SET_IRQS failure");
191 
192     name = index_to_str(vbasedev, index);
193     if (name) {
194         error_prepend(errp, "%s-%d: ", name, subindex);
195     } else {
196         error_prepend(errp, "index %d-%d: ", index, subindex);
197     }
198     error_prepend(errp,
199                   "Failed to %s %s eventfd signaling for interrupt ",
200                   fd < 0 ? "tear down" : "set up", action_to_str(action));
201     return ret;
202 }
203 
204 /*
205  * IO Port/MMIO - Beware of the endians, VFIO is always little endian
206  */
207 void vfio_region_write(void *opaque, hwaddr addr,
208                        uint64_t data, unsigned size)
209 {
210     VFIORegion *region = opaque;
211     VFIODevice *vbasedev = region->vbasedev;
212     union {
213         uint8_t byte;
214         uint16_t word;
215         uint32_t dword;
216         uint64_t qword;
217     } buf;
218 
219     switch (size) {
220     case 1:
221         buf.byte = data;
222         break;
223     case 2:
224         buf.word = cpu_to_le16(data);
225         break;
226     case 4:
227         buf.dword = cpu_to_le32(data);
228         break;
229     case 8:
230         buf.qword = cpu_to_le64(data);
231         break;
232     default:
233         hw_error("vfio: unsupported write size, %u bytes", size);
234         break;
235     }
236 
237     if (pwrite(vbasedev->fd, &buf, size, region->fd_offset + addr) != size) {
238         error_report("%s(%s:region%d+0x%"HWADDR_PRIx", 0x%"PRIx64
239                      ",%d) failed: %m",
240                      __func__, vbasedev->name, region->nr,
241                      addr, data, size);
242     }
243 
244     trace_vfio_region_write(vbasedev->name, region->nr, addr, data, size);
245 
246     /*
247      * A read or write to a BAR always signals an INTx EOI.  This will
248      * do nothing if not pending (including not in INTx mode).  We assume
249      * that a BAR access is in response to an interrupt and that BAR
250      * accesses will service the interrupt.  Unfortunately, we don't know
251      * which access will service the interrupt, so we're potentially
252      * getting quite a few host interrupts per guest interrupt.
253      */
254     vbasedev->ops->vfio_eoi(vbasedev);
255 }
256 
257 uint64_t vfio_region_read(void *opaque,
258                           hwaddr addr, unsigned size)
259 {
260     VFIORegion *region = opaque;
261     VFIODevice *vbasedev = region->vbasedev;
262     union {
263         uint8_t byte;
264         uint16_t word;
265         uint32_t dword;
266         uint64_t qword;
267     } buf;
268     uint64_t data = 0;
269 
270     if (pread(vbasedev->fd, &buf, size, region->fd_offset + addr) != size) {
271         error_report("%s(%s:region%d+0x%"HWADDR_PRIx", %d) failed: %m",
272                      __func__, vbasedev->name, region->nr,
273                      addr, size);
274         return (uint64_t)-1;
275     }
276     switch (size) {
277     case 1:
278         data = buf.byte;
279         break;
280     case 2:
281         data = le16_to_cpu(buf.word);
282         break;
283     case 4:
284         data = le32_to_cpu(buf.dword);
285         break;
286     case 8:
287         data = le64_to_cpu(buf.qword);
288         break;
289     default:
290         hw_error("vfio: unsupported read size, %u bytes", size);
291         break;
292     }
293 
294     trace_vfio_region_read(vbasedev->name, region->nr, addr, size, data);
295 
296     /* Same as write above */
297     vbasedev->ops->vfio_eoi(vbasedev);
298 
299     return data;
300 }
301 
302 const MemoryRegionOps vfio_region_ops = {
303     .read = vfio_region_read,
304     .write = vfio_region_write,
305     .endianness = DEVICE_LITTLE_ENDIAN,
306     .valid = {
307         .min_access_size = 1,
308         .max_access_size = 8,
309     },
310     .impl = {
311         .min_access_size = 1,
312         .max_access_size = 8,
313     },
314 };
315 
316 /*
317  * Device state interfaces
318  */
319 
320 bool vfio_mig_active(void)
321 {
322     VFIOGroup *group;
323     VFIODevice *vbasedev;
324 
325     if (QLIST_EMPTY(&vfio_group_list)) {
326         return false;
327     }
328 
329     QLIST_FOREACH(group, &vfio_group_list, next) {
330         QLIST_FOREACH(vbasedev, &group->device_list, next) {
331             if (vbasedev->migration_blocker) {
332                 return false;
333             }
334         }
335     }
336     return true;
337 }
338 
339 static bool vfio_devices_all_dirty_tracking(VFIOContainer *container)
340 {
341     VFIOGroup *group;
342     VFIODevice *vbasedev;
343     MigrationState *ms = migrate_get_current();
344 
345     if (!migration_is_setup_or_active(ms->state)) {
346         return false;
347     }
348 
349     QLIST_FOREACH(group, &container->group_list, container_next) {
350         QLIST_FOREACH(vbasedev, &group->device_list, next) {
351             VFIOMigration *migration = vbasedev->migration;
352 
353             if (!migration) {
354                 return false;
355             }
356 
357             if ((vbasedev->pre_copy_dirty_page_tracking == ON_OFF_AUTO_OFF)
358                 && (migration->device_state & VFIO_DEVICE_STATE_V1_RUNNING)) {
359                 return false;
360             }
361         }
362     }
363     return true;
364 }
365 
366 static bool vfio_devices_all_running_and_saving(VFIOContainer *container)
367 {
368     VFIOGroup *group;
369     VFIODevice *vbasedev;
370     MigrationState *ms = migrate_get_current();
371 
372     if (!migration_is_setup_or_active(ms->state)) {
373         return false;
374     }
375 
376     QLIST_FOREACH(group, &container->group_list, container_next) {
377         QLIST_FOREACH(vbasedev, &group->device_list, next) {
378             VFIOMigration *migration = vbasedev->migration;
379 
380             if (!migration) {
381                 return false;
382             }
383 
384             if ((migration->device_state & VFIO_DEVICE_STATE_V1_SAVING) &&
385                 (migration->device_state & VFIO_DEVICE_STATE_V1_RUNNING)) {
386                 continue;
387             } else {
388                 return false;
389             }
390         }
391     }
392     return true;
393 }
394 
395 static int vfio_dma_unmap_bitmap(VFIOContainer *container,
396                                  hwaddr iova, ram_addr_t size,
397                                  IOMMUTLBEntry *iotlb)
398 {
399     struct vfio_iommu_type1_dma_unmap *unmap;
400     struct vfio_bitmap *bitmap;
401     uint64_t pages = REAL_HOST_PAGE_ALIGN(size) / qemu_real_host_page_size();
402     int ret;
403 
404     unmap = g_malloc0(sizeof(*unmap) + sizeof(*bitmap));
405 
406     unmap->argsz = sizeof(*unmap) + sizeof(*bitmap);
407     unmap->iova = iova;
408     unmap->size = size;
409     unmap->flags |= VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP;
410     bitmap = (struct vfio_bitmap *)&unmap->data;
411 
412     /*
413      * cpu_physical_memory_set_dirty_lebitmap() supports pages in bitmap of
414      * qemu_real_host_page_size to mark those dirty. Hence set bitmap_pgsize
415      * to qemu_real_host_page_size.
416      */
417 
418     bitmap->pgsize = qemu_real_host_page_size();
419     bitmap->size = ROUND_UP(pages, sizeof(__u64) * BITS_PER_BYTE) /
420                    BITS_PER_BYTE;
421 
422     if (bitmap->size > container->max_dirty_bitmap_size) {
423         error_report("UNMAP: Size of bitmap too big 0x%"PRIx64,
424                      (uint64_t)bitmap->size);
425         ret = -E2BIG;
426         goto unmap_exit;
427     }
428 
429     bitmap->data = g_try_malloc0(bitmap->size);
430     if (!bitmap->data) {
431         ret = -ENOMEM;
432         goto unmap_exit;
433     }
434 
435     ret = ioctl(container->fd, VFIO_IOMMU_UNMAP_DMA, unmap);
436     if (!ret) {
437         cpu_physical_memory_set_dirty_lebitmap((unsigned long *)bitmap->data,
438                 iotlb->translated_addr, pages);
439     } else {
440         error_report("VFIO_UNMAP_DMA with DIRTY_BITMAP : %m");
441     }
442 
443     g_free(bitmap->data);
444 unmap_exit:
445     g_free(unmap);
446     return ret;
447 }
448 
449 /*
450  * DMA - Mapping and unmapping for the "type1" IOMMU interface used on x86
451  */
452 static int vfio_dma_unmap(VFIOContainer *container,
453                           hwaddr iova, ram_addr_t size,
454                           IOMMUTLBEntry *iotlb)
455 {
456     struct vfio_iommu_type1_dma_unmap unmap = {
457         .argsz = sizeof(unmap),
458         .flags = 0,
459         .iova = iova,
460         .size = size,
461     };
462 
463     if (iotlb && container->dirty_pages_supported &&
464         vfio_devices_all_running_and_saving(container)) {
465         return vfio_dma_unmap_bitmap(container, iova, size, iotlb);
466     }
467 
468     while (ioctl(container->fd, VFIO_IOMMU_UNMAP_DMA, &unmap)) {
469         /*
470          * The type1 backend has an off-by-one bug in the kernel (71a7d3d78e3c
471          * v4.15) where an overflow in its wrap-around check prevents us from
472          * unmapping the last page of the address space.  Test for the error
473          * condition and re-try the unmap excluding the last page.  The
474          * expectation is that we've never mapped the last page anyway and this
475          * unmap request comes via vIOMMU support which also makes it unlikely
476          * that this page is used.  This bug was introduced well after type1 v2
477          * support was introduced, so we shouldn't need to test for v1.  A fix
478          * is queued for kernel v5.0 so this workaround can be removed once
479          * affected kernels are sufficiently deprecated.
480          */
481         if (errno == EINVAL && unmap.size && !(unmap.iova + unmap.size) &&
482             container->iommu_type == VFIO_TYPE1v2_IOMMU) {
483             trace_vfio_dma_unmap_overflow_workaround();
484             unmap.size -= 1ULL << ctz64(container->pgsizes);
485             continue;
486         }
487         error_report("VFIO_UNMAP_DMA failed: %s", strerror(errno));
488         return -errno;
489     }
490 
491     return 0;
492 }
493 
494 static int vfio_dma_map(VFIOContainer *container, hwaddr iova,
495                         ram_addr_t size, void *vaddr, bool readonly)
496 {
497     struct vfio_iommu_type1_dma_map map = {
498         .argsz = sizeof(map),
499         .flags = VFIO_DMA_MAP_FLAG_READ,
500         .vaddr = (__u64)(uintptr_t)vaddr,
501         .iova = iova,
502         .size = size,
503     };
504 
505     if (!readonly) {
506         map.flags |= VFIO_DMA_MAP_FLAG_WRITE;
507     }
508 
509     /*
510      * Try the mapping, if it fails with EBUSY, unmap the region and try
511      * again.  This shouldn't be necessary, but we sometimes see it in
512      * the VGA ROM space.
513      */
514     if (ioctl(container->fd, VFIO_IOMMU_MAP_DMA, &map) == 0 ||
515         (errno == EBUSY && vfio_dma_unmap(container, iova, size, NULL) == 0 &&
516          ioctl(container->fd, VFIO_IOMMU_MAP_DMA, &map) == 0)) {
517         return 0;
518     }
519 
520     error_report("VFIO_MAP_DMA failed: %s", strerror(errno));
521     return -errno;
522 }
523 
524 static void vfio_host_win_add(VFIOContainer *container,
525                               hwaddr min_iova, hwaddr max_iova,
526                               uint64_t iova_pgsizes)
527 {
528     VFIOHostDMAWindow *hostwin;
529 
530     QLIST_FOREACH(hostwin, &container->hostwin_list, hostwin_next) {
531         if (ranges_overlap(hostwin->min_iova,
532                            hostwin->max_iova - hostwin->min_iova + 1,
533                            min_iova,
534                            max_iova - min_iova + 1)) {
535             hw_error("%s: Overlapped IOMMU are not enabled", __func__);
536         }
537     }
538 
539     hostwin = g_malloc0(sizeof(*hostwin));
540 
541     hostwin->min_iova = min_iova;
542     hostwin->max_iova = max_iova;
543     hostwin->iova_pgsizes = iova_pgsizes;
544     QLIST_INSERT_HEAD(&container->hostwin_list, hostwin, hostwin_next);
545 }
546 
547 static int vfio_host_win_del(VFIOContainer *container, hwaddr min_iova,
548                              hwaddr max_iova)
549 {
550     VFIOHostDMAWindow *hostwin;
551 
552     QLIST_FOREACH(hostwin, &container->hostwin_list, hostwin_next) {
553         if (hostwin->min_iova == min_iova && hostwin->max_iova == max_iova) {
554             QLIST_REMOVE(hostwin, hostwin_next);
555             g_free(hostwin);
556             return 0;
557         }
558     }
559 
560     return -1;
561 }
562 
563 static bool vfio_listener_skipped_section(MemoryRegionSection *section)
564 {
565     return (!memory_region_is_ram(section->mr) &&
566             !memory_region_is_iommu(section->mr)) ||
567            memory_region_is_protected(section->mr) ||
568            /*
569             * Sizing an enabled 64-bit BAR can cause spurious mappings to
570             * addresses in the upper part of the 64-bit address space.  These
571             * are never accessed by the CPU and beyond the address width of
572             * some IOMMU hardware.  TODO: VFIO should tell us the IOMMU width.
573             */
574            section->offset_within_address_space & (1ULL << 63);
575 }
576 
577 /* Called with rcu_read_lock held.  */
578 static bool vfio_get_xlat_addr(IOMMUTLBEntry *iotlb, void **vaddr,
579                                ram_addr_t *ram_addr, bool *read_only)
580 {
581     MemoryRegion *mr;
582     hwaddr xlat;
583     hwaddr len = iotlb->addr_mask + 1;
584     bool writable = iotlb->perm & IOMMU_WO;
585 
586     /*
587      * The IOMMU TLB entry we have just covers translation through
588      * this IOMMU to its immediate target.  We need to translate
589      * it the rest of the way through to memory.
590      */
591     mr = address_space_translate(&address_space_memory,
592                                  iotlb->translated_addr,
593                                  &xlat, &len, writable,
594                                  MEMTXATTRS_UNSPECIFIED);
595     if (!memory_region_is_ram(mr)) {
596         error_report("iommu map to non memory area %"HWADDR_PRIx"",
597                      xlat);
598         return false;
599     } else if (memory_region_has_ram_discard_manager(mr)) {
600         RamDiscardManager *rdm = memory_region_get_ram_discard_manager(mr);
601         MemoryRegionSection tmp = {
602             .mr = mr,
603             .offset_within_region = xlat,
604             .size = int128_make64(len),
605         };
606 
607         /*
608          * Malicious VMs can map memory into the IOMMU, which is expected
609          * to remain discarded. vfio will pin all pages, populating memory.
610          * Disallow that. vmstate priorities make sure any RamDiscardManager
611          * were already restored before IOMMUs are restored.
612          */
613         if (!ram_discard_manager_is_populated(rdm, &tmp)) {
614             error_report("iommu map to discarded memory (e.g., unplugged via"
615                          " virtio-mem): %"HWADDR_PRIx"",
616                          iotlb->translated_addr);
617             return false;
618         }
619 
620         /*
621          * Malicious VMs might trigger discarding of IOMMU-mapped memory. The
622          * pages will remain pinned inside vfio until unmapped, resulting in a
623          * higher memory consumption than expected. If memory would get
624          * populated again later, there would be an inconsistency between pages
625          * pinned by vfio and pages seen by QEMU. This is the case until
626          * unmapped from the IOMMU (e.g., during device reset).
627          *
628          * With malicious guests, we really only care about pinning more memory
629          * than expected. RLIMIT_MEMLOCK set for the user/process can never be
630          * exceeded and can be used to mitigate this problem.
631          */
632         warn_report_once("Using vfio with vIOMMUs and coordinated discarding of"
633                          " RAM (e.g., virtio-mem) works, however, malicious"
634                          " guests can trigger pinning of more memory than"
635                          " intended via an IOMMU. It's possible to mitigate "
636                          " by setting/adjusting RLIMIT_MEMLOCK.");
637     }
638 
639     /*
640      * Translation truncates length to the IOMMU page size,
641      * check that it did not truncate too much.
642      */
643     if (len & iotlb->addr_mask) {
644         error_report("iommu has granularity incompatible with target AS");
645         return false;
646     }
647 
648     if (vaddr) {
649         *vaddr = memory_region_get_ram_ptr(mr) + xlat;
650     }
651 
652     if (ram_addr) {
653         *ram_addr = memory_region_get_ram_addr(mr) + xlat;
654     }
655 
656     if (read_only) {
657         *read_only = !writable || mr->readonly;
658     }
659 
660     return true;
661 }
662 
663 static void vfio_iommu_map_notify(IOMMUNotifier *n, IOMMUTLBEntry *iotlb)
664 {
665     VFIOGuestIOMMU *giommu = container_of(n, VFIOGuestIOMMU, n);
666     VFIOContainer *container = giommu->container;
667     hwaddr iova = iotlb->iova + giommu->iommu_offset;
668     void *vaddr;
669     int ret;
670 
671     trace_vfio_iommu_map_notify(iotlb->perm == IOMMU_NONE ? "UNMAP" : "MAP",
672                                 iova, iova + iotlb->addr_mask);
673 
674     if (iotlb->target_as != &address_space_memory) {
675         error_report("Wrong target AS \"%s\", only system memory is allowed",
676                      iotlb->target_as->name ? iotlb->target_as->name : "none");
677         return;
678     }
679 
680     rcu_read_lock();
681 
682     if ((iotlb->perm & IOMMU_RW) != IOMMU_NONE) {
683         bool read_only;
684 
685         if (!vfio_get_xlat_addr(iotlb, &vaddr, NULL, &read_only)) {
686             goto out;
687         }
688         /*
689          * vaddr is only valid until rcu_read_unlock(). But after
690          * vfio_dma_map has set up the mapping the pages will be
691          * pinned by the kernel. This makes sure that the RAM backend
692          * of vaddr will always be there, even if the memory object is
693          * destroyed and its backing memory munmap-ed.
694          */
695         ret = vfio_dma_map(container, iova,
696                            iotlb->addr_mask + 1, vaddr,
697                            read_only);
698         if (ret) {
699             error_report("vfio_dma_map(%p, 0x%"HWADDR_PRIx", "
700                          "0x%"HWADDR_PRIx", %p) = %d (%m)",
701                          container, iova,
702                          iotlb->addr_mask + 1, vaddr, ret);
703         }
704     } else {
705         ret = vfio_dma_unmap(container, iova, iotlb->addr_mask + 1, iotlb);
706         if (ret) {
707             error_report("vfio_dma_unmap(%p, 0x%"HWADDR_PRIx", "
708                          "0x%"HWADDR_PRIx") = %d (%m)",
709                          container, iova,
710                          iotlb->addr_mask + 1, ret);
711         }
712     }
713 out:
714     rcu_read_unlock();
715 }
716 
717 static void vfio_ram_discard_notify_discard(RamDiscardListener *rdl,
718                                             MemoryRegionSection *section)
719 {
720     VFIORamDiscardListener *vrdl = container_of(rdl, VFIORamDiscardListener,
721                                                 listener);
722     const hwaddr size = int128_get64(section->size);
723     const hwaddr iova = section->offset_within_address_space;
724     int ret;
725 
726     /* Unmap with a single call. */
727     ret = vfio_dma_unmap(vrdl->container, iova, size , NULL);
728     if (ret) {
729         error_report("%s: vfio_dma_unmap() failed: %s", __func__,
730                      strerror(-ret));
731     }
732 }
733 
734 static int vfio_ram_discard_notify_populate(RamDiscardListener *rdl,
735                                             MemoryRegionSection *section)
736 {
737     VFIORamDiscardListener *vrdl = container_of(rdl, VFIORamDiscardListener,
738                                                 listener);
739     const hwaddr end = section->offset_within_region +
740                        int128_get64(section->size);
741     hwaddr start, next, iova;
742     void *vaddr;
743     int ret;
744 
745     /*
746      * Map in (aligned within memory region) minimum granularity, so we can
747      * unmap in minimum granularity later.
748      */
749     for (start = section->offset_within_region; start < end; start = next) {
750         next = ROUND_UP(start + 1, vrdl->granularity);
751         next = MIN(next, end);
752 
753         iova = start - section->offset_within_region +
754                section->offset_within_address_space;
755         vaddr = memory_region_get_ram_ptr(section->mr) + start;
756 
757         ret = vfio_dma_map(vrdl->container, iova, next - start,
758                            vaddr, section->readonly);
759         if (ret) {
760             /* Rollback */
761             vfio_ram_discard_notify_discard(rdl, section);
762             return ret;
763         }
764     }
765     return 0;
766 }
767 
768 static void vfio_register_ram_discard_listener(VFIOContainer *container,
769                                                MemoryRegionSection *section)
770 {
771     RamDiscardManager *rdm = memory_region_get_ram_discard_manager(section->mr);
772     VFIORamDiscardListener *vrdl;
773 
774     /* Ignore some corner cases not relevant in practice. */
775     g_assert(QEMU_IS_ALIGNED(section->offset_within_region, TARGET_PAGE_SIZE));
776     g_assert(QEMU_IS_ALIGNED(section->offset_within_address_space,
777                              TARGET_PAGE_SIZE));
778     g_assert(QEMU_IS_ALIGNED(int128_get64(section->size), TARGET_PAGE_SIZE));
779 
780     vrdl = g_new0(VFIORamDiscardListener, 1);
781     vrdl->container = container;
782     vrdl->mr = section->mr;
783     vrdl->offset_within_address_space = section->offset_within_address_space;
784     vrdl->size = int128_get64(section->size);
785     vrdl->granularity = ram_discard_manager_get_min_granularity(rdm,
786                                                                 section->mr);
787 
788     g_assert(vrdl->granularity && is_power_of_2(vrdl->granularity));
789     g_assert(container->pgsizes &&
790              vrdl->granularity >= 1ULL << ctz64(container->pgsizes));
791 
792     ram_discard_listener_init(&vrdl->listener,
793                               vfio_ram_discard_notify_populate,
794                               vfio_ram_discard_notify_discard, true);
795     ram_discard_manager_register_listener(rdm, &vrdl->listener, section);
796     QLIST_INSERT_HEAD(&container->vrdl_list, vrdl, next);
797 
798     /*
799      * Sanity-check if we have a theoretically problematic setup where we could
800      * exceed the maximum number of possible DMA mappings over time. We assume
801      * that each mapped section in the same address space as a RamDiscardManager
802      * section consumes exactly one DMA mapping, with the exception of
803      * RamDiscardManager sections; i.e., we don't expect to have gIOMMU sections
804      * in the same address space as RamDiscardManager sections.
805      *
806      * We assume that each section in the address space consumes one memslot.
807      * We take the number of KVM memory slots as a best guess for the maximum
808      * number of sections in the address space we could have over time,
809      * also consuming DMA mappings.
810      */
811     if (container->dma_max_mappings) {
812         unsigned int vrdl_count = 0, vrdl_mappings = 0, max_memslots = 512;
813 
814 #ifdef CONFIG_KVM
815         if (kvm_enabled()) {
816             max_memslots = kvm_get_max_memslots();
817         }
818 #endif
819 
820         QLIST_FOREACH(vrdl, &container->vrdl_list, next) {
821             hwaddr start, end;
822 
823             start = QEMU_ALIGN_DOWN(vrdl->offset_within_address_space,
824                                     vrdl->granularity);
825             end = ROUND_UP(vrdl->offset_within_address_space + vrdl->size,
826                            vrdl->granularity);
827             vrdl_mappings += (end - start) / vrdl->granularity;
828             vrdl_count++;
829         }
830 
831         if (vrdl_mappings + max_memslots - vrdl_count >
832             container->dma_max_mappings) {
833             warn_report("%s: possibly running out of DMA mappings. E.g., try"
834                         " increasing the 'block-size' of virtio-mem devies."
835                         " Maximum possible DMA mappings: %d, Maximum possible"
836                         " memslots: %d", __func__, container->dma_max_mappings,
837                         max_memslots);
838         }
839     }
840 }
841 
842 static void vfio_unregister_ram_discard_listener(VFIOContainer *container,
843                                                  MemoryRegionSection *section)
844 {
845     RamDiscardManager *rdm = memory_region_get_ram_discard_manager(section->mr);
846     VFIORamDiscardListener *vrdl = NULL;
847 
848     QLIST_FOREACH(vrdl, &container->vrdl_list, next) {
849         if (vrdl->mr == section->mr &&
850             vrdl->offset_within_address_space ==
851             section->offset_within_address_space) {
852             break;
853         }
854     }
855 
856     if (!vrdl) {
857         hw_error("vfio: Trying to unregister missing RAM discard listener");
858     }
859 
860     ram_discard_manager_unregister_listener(rdm, &vrdl->listener);
861     QLIST_REMOVE(vrdl, next);
862     g_free(vrdl);
863 }
864 
865 static bool vfio_known_safe_misalignment(MemoryRegionSection *section)
866 {
867     MemoryRegion *mr = section->mr;
868 
869     if (!TPM_IS_CRB(mr->owner)) {
870         return false;
871     }
872 
873     /* this is a known safe misaligned region, just trace for debug purpose */
874     trace_vfio_known_safe_misalignment(memory_region_name(mr),
875                                        section->offset_within_address_space,
876                                        section->offset_within_region,
877                                        qemu_real_host_page_size());
878     return true;
879 }
880 
881 static void vfio_listener_region_add(MemoryListener *listener,
882                                      MemoryRegionSection *section)
883 {
884     VFIOContainer *container = container_of(listener, VFIOContainer, listener);
885     hwaddr iova, end;
886     Int128 llend, llsize;
887     void *vaddr;
888     int ret;
889     VFIOHostDMAWindow *hostwin;
890     bool hostwin_found;
891     Error *err = NULL;
892 
893     if (vfio_listener_skipped_section(section)) {
894         trace_vfio_listener_region_add_skip(
895                 section->offset_within_address_space,
896                 section->offset_within_address_space +
897                 int128_get64(int128_sub(section->size, int128_one())));
898         return;
899     }
900 
901     if (unlikely((section->offset_within_address_space &
902                   ~qemu_real_host_page_mask()) !=
903                  (section->offset_within_region & ~qemu_real_host_page_mask()))) {
904         if (!vfio_known_safe_misalignment(section)) {
905             error_report("%s received unaligned region %s iova=0x%"PRIx64
906                          " offset_within_region=0x%"PRIx64
907                          " qemu_real_host_page_size=0x%"PRIxPTR,
908                          __func__, memory_region_name(section->mr),
909                          section->offset_within_address_space,
910                          section->offset_within_region,
911                          qemu_real_host_page_size());
912         }
913         return;
914     }
915 
916     iova = REAL_HOST_PAGE_ALIGN(section->offset_within_address_space);
917     llend = int128_make64(section->offset_within_address_space);
918     llend = int128_add(llend, section->size);
919     llend = int128_and(llend, int128_exts64(qemu_real_host_page_mask()));
920 
921     if (int128_ge(int128_make64(iova), llend)) {
922         if (memory_region_is_ram_device(section->mr)) {
923             trace_vfio_listener_region_add_no_dma_map(
924                 memory_region_name(section->mr),
925                 section->offset_within_address_space,
926                 int128_getlo(section->size),
927                 qemu_real_host_page_size());
928         }
929         return;
930     }
931     end = int128_get64(int128_sub(llend, int128_one()));
932 
933     if (container->iommu_type == VFIO_SPAPR_TCE_v2_IOMMU) {
934         hwaddr pgsize = 0;
935 
936         /* For now intersections are not allowed, we may relax this later */
937         QLIST_FOREACH(hostwin, &container->hostwin_list, hostwin_next) {
938             if (ranges_overlap(hostwin->min_iova,
939                                hostwin->max_iova - hostwin->min_iova + 1,
940                                section->offset_within_address_space,
941                                int128_get64(section->size))) {
942                 error_setg(&err,
943                     "region [0x%"PRIx64",0x%"PRIx64"] overlaps with existing"
944                     "host DMA window [0x%"PRIx64",0x%"PRIx64"]",
945                     section->offset_within_address_space,
946                     section->offset_within_address_space +
947                         int128_get64(section->size) - 1,
948                     hostwin->min_iova, hostwin->max_iova);
949                 goto fail;
950             }
951         }
952 
953         ret = vfio_spapr_create_window(container, section, &pgsize);
954         if (ret) {
955             error_setg_errno(&err, -ret, "Failed to create SPAPR window");
956             goto fail;
957         }
958 
959         vfio_host_win_add(container, section->offset_within_address_space,
960                           section->offset_within_address_space +
961                           int128_get64(section->size) - 1, pgsize);
962 #ifdef CONFIG_KVM
963         if (kvm_enabled()) {
964             VFIOGroup *group;
965             IOMMUMemoryRegion *iommu_mr = IOMMU_MEMORY_REGION(section->mr);
966             struct kvm_vfio_spapr_tce param;
967             struct kvm_device_attr attr = {
968                 .group = KVM_DEV_VFIO_GROUP,
969                 .attr = KVM_DEV_VFIO_GROUP_SET_SPAPR_TCE,
970                 .addr = (uint64_t)(unsigned long)&param,
971             };
972 
973             if (!memory_region_iommu_get_attr(iommu_mr, IOMMU_ATTR_SPAPR_TCE_FD,
974                                               &param.tablefd)) {
975                 QLIST_FOREACH(group, &container->group_list, container_next) {
976                     param.groupfd = group->fd;
977                     if (ioctl(vfio_kvm_device_fd, KVM_SET_DEVICE_ATTR, &attr)) {
978                         error_report("vfio: failed to setup fd %d "
979                                      "for a group with fd %d: %s",
980                                      param.tablefd, param.groupfd,
981                                      strerror(errno));
982                         return;
983                     }
984                     trace_vfio_spapr_group_attach(param.groupfd, param.tablefd);
985                 }
986             }
987         }
988 #endif
989     }
990 
991     hostwin_found = false;
992     QLIST_FOREACH(hostwin, &container->hostwin_list, hostwin_next) {
993         if (hostwin->min_iova <= iova && end <= hostwin->max_iova) {
994             hostwin_found = true;
995             break;
996         }
997     }
998 
999     if (!hostwin_found) {
1000         error_setg(&err, "Container %p can't map guest IOVA region"
1001                    " 0x%"HWADDR_PRIx"..0x%"HWADDR_PRIx, container, iova, end);
1002         goto fail;
1003     }
1004 
1005     memory_region_ref(section->mr);
1006 
1007     if (memory_region_is_iommu(section->mr)) {
1008         VFIOGuestIOMMU *giommu;
1009         IOMMUMemoryRegion *iommu_mr = IOMMU_MEMORY_REGION(section->mr);
1010         int iommu_idx;
1011 
1012         trace_vfio_listener_region_add_iommu(iova, end);
1013         /*
1014          * FIXME: For VFIO iommu types which have KVM acceleration to
1015          * avoid bouncing all map/unmaps through qemu this way, this
1016          * would be the right place to wire that up (tell the KVM
1017          * device emulation the VFIO iommu handles to use).
1018          */
1019         giommu = g_malloc0(sizeof(*giommu));
1020         giommu->iommu_mr = iommu_mr;
1021         giommu->iommu_offset = section->offset_within_address_space -
1022                                section->offset_within_region;
1023         giommu->container = container;
1024         llend = int128_add(int128_make64(section->offset_within_region),
1025                            section->size);
1026         llend = int128_sub(llend, int128_one());
1027         iommu_idx = memory_region_iommu_attrs_to_index(iommu_mr,
1028                                                        MEMTXATTRS_UNSPECIFIED);
1029         iommu_notifier_init(&giommu->n, vfio_iommu_map_notify,
1030                             IOMMU_NOTIFIER_IOTLB_EVENTS,
1031                             section->offset_within_region,
1032                             int128_get64(llend),
1033                             iommu_idx);
1034 
1035         ret = memory_region_iommu_set_page_size_mask(giommu->iommu_mr,
1036                                                      container->pgsizes,
1037                                                      &err);
1038         if (ret) {
1039             g_free(giommu);
1040             goto fail;
1041         }
1042 
1043         ret = memory_region_register_iommu_notifier(section->mr, &giommu->n,
1044                                                     &err);
1045         if (ret) {
1046             g_free(giommu);
1047             goto fail;
1048         }
1049         QLIST_INSERT_HEAD(&container->giommu_list, giommu, giommu_next);
1050         memory_region_iommu_replay(giommu->iommu_mr, &giommu->n);
1051 
1052         return;
1053     }
1054 
1055     /* Here we assume that memory_region_is_ram(section->mr)==true */
1056 
1057     /*
1058      * For RAM memory regions with a RamDiscardManager, we only want to map the
1059      * actually populated parts - and update the mapping whenever we're notified
1060      * about changes.
1061      */
1062     if (memory_region_has_ram_discard_manager(section->mr)) {
1063         vfio_register_ram_discard_listener(container, section);
1064         return;
1065     }
1066 
1067     vaddr = memory_region_get_ram_ptr(section->mr) +
1068             section->offset_within_region +
1069             (iova - section->offset_within_address_space);
1070 
1071     trace_vfio_listener_region_add_ram(iova, end, vaddr);
1072 
1073     llsize = int128_sub(llend, int128_make64(iova));
1074 
1075     if (memory_region_is_ram_device(section->mr)) {
1076         hwaddr pgmask = (1ULL << ctz64(hostwin->iova_pgsizes)) - 1;
1077 
1078         if ((iova & pgmask) || (int128_get64(llsize) & pgmask)) {
1079             trace_vfio_listener_region_add_no_dma_map(
1080                 memory_region_name(section->mr),
1081                 section->offset_within_address_space,
1082                 int128_getlo(section->size),
1083                 pgmask + 1);
1084             return;
1085         }
1086     }
1087 
1088     ret = vfio_dma_map(container, iova, int128_get64(llsize),
1089                        vaddr, section->readonly);
1090     if (ret) {
1091         error_setg(&err, "vfio_dma_map(%p, 0x%"HWADDR_PRIx", "
1092                    "0x%"HWADDR_PRIx", %p) = %d (%m)",
1093                    container, iova, int128_get64(llsize), vaddr, ret);
1094         if (memory_region_is_ram_device(section->mr)) {
1095             /* Allow unexpected mappings not to be fatal for RAM devices */
1096             error_report_err(err);
1097             return;
1098         }
1099         goto fail;
1100     }
1101 
1102     return;
1103 
1104 fail:
1105     if (memory_region_is_ram_device(section->mr)) {
1106         error_report("failed to vfio_dma_map. pci p2p may not work");
1107         return;
1108     }
1109     /*
1110      * On the initfn path, store the first error in the container so we
1111      * can gracefully fail.  Runtime, there's not much we can do other
1112      * than throw a hardware error.
1113      */
1114     if (!container->initialized) {
1115         if (!container->error) {
1116             error_propagate_prepend(&container->error, err,
1117                                     "Region %s: ",
1118                                     memory_region_name(section->mr));
1119         } else {
1120             error_free(err);
1121         }
1122     } else {
1123         error_report_err(err);
1124         hw_error("vfio: DMA mapping failed, unable to continue");
1125     }
1126 }
1127 
1128 static void vfio_listener_region_del(MemoryListener *listener,
1129                                      MemoryRegionSection *section)
1130 {
1131     VFIOContainer *container = container_of(listener, VFIOContainer, listener);
1132     hwaddr iova, end;
1133     Int128 llend, llsize;
1134     int ret;
1135     bool try_unmap = true;
1136 
1137     if (vfio_listener_skipped_section(section)) {
1138         trace_vfio_listener_region_del_skip(
1139                 section->offset_within_address_space,
1140                 section->offset_within_address_space +
1141                 int128_get64(int128_sub(section->size, int128_one())));
1142         return;
1143     }
1144 
1145     if (unlikely((section->offset_within_address_space &
1146                   ~qemu_real_host_page_mask()) !=
1147                  (section->offset_within_region & ~qemu_real_host_page_mask()))) {
1148         error_report("%s received unaligned region", __func__);
1149         return;
1150     }
1151 
1152     if (memory_region_is_iommu(section->mr)) {
1153         VFIOGuestIOMMU *giommu;
1154 
1155         QLIST_FOREACH(giommu, &container->giommu_list, giommu_next) {
1156             if (MEMORY_REGION(giommu->iommu_mr) == section->mr &&
1157                 giommu->n.start == section->offset_within_region) {
1158                 memory_region_unregister_iommu_notifier(section->mr,
1159                                                         &giommu->n);
1160                 QLIST_REMOVE(giommu, giommu_next);
1161                 g_free(giommu);
1162                 break;
1163             }
1164         }
1165 
1166         /*
1167          * FIXME: We assume the one big unmap below is adequate to
1168          * remove any individual page mappings in the IOMMU which
1169          * might have been copied into VFIO. This works for a page table
1170          * based IOMMU where a big unmap flattens a large range of IO-PTEs.
1171          * That may not be true for all IOMMU types.
1172          */
1173     }
1174 
1175     iova = REAL_HOST_PAGE_ALIGN(section->offset_within_address_space);
1176     llend = int128_make64(section->offset_within_address_space);
1177     llend = int128_add(llend, section->size);
1178     llend = int128_and(llend, int128_exts64(qemu_real_host_page_mask()));
1179 
1180     if (int128_ge(int128_make64(iova), llend)) {
1181         return;
1182     }
1183     end = int128_get64(int128_sub(llend, int128_one()));
1184 
1185     llsize = int128_sub(llend, int128_make64(iova));
1186 
1187     trace_vfio_listener_region_del(iova, end);
1188 
1189     if (memory_region_is_ram_device(section->mr)) {
1190         hwaddr pgmask;
1191         VFIOHostDMAWindow *hostwin;
1192         bool hostwin_found = false;
1193 
1194         QLIST_FOREACH(hostwin, &container->hostwin_list, hostwin_next) {
1195             if (hostwin->min_iova <= iova && end <= hostwin->max_iova) {
1196                 hostwin_found = true;
1197                 break;
1198             }
1199         }
1200         assert(hostwin_found); /* or region_add() would have failed */
1201 
1202         pgmask = (1ULL << ctz64(hostwin->iova_pgsizes)) - 1;
1203         try_unmap = !((iova & pgmask) || (int128_get64(llsize) & pgmask));
1204     } else if (memory_region_has_ram_discard_manager(section->mr)) {
1205         vfio_unregister_ram_discard_listener(container, section);
1206         /* Unregistering will trigger an unmap. */
1207         try_unmap = false;
1208     }
1209 
1210     if (try_unmap) {
1211         if (int128_eq(llsize, int128_2_64())) {
1212             /* The unmap ioctl doesn't accept a full 64-bit span. */
1213             llsize = int128_rshift(llsize, 1);
1214             ret = vfio_dma_unmap(container, iova, int128_get64(llsize), NULL);
1215             if (ret) {
1216                 error_report("vfio_dma_unmap(%p, 0x%"HWADDR_PRIx", "
1217                              "0x%"HWADDR_PRIx") = %d (%m)",
1218                              container, iova, int128_get64(llsize), ret);
1219             }
1220             iova += int128_get64(llsize);
1221         }
1222         ret = vfio_dma_unmap(container, iova, int128_get64(llsize), NULL);
1223         if (ret) {
1224             error_report("vfio_dma_unmap(%p, 0x%"HWADDR_PRIx", "
1225                          "0x%"HWADDR_PRIx") = %d (%m)",
1226                          container, iova, int128_get64(llsize), ret);
1227         }
1228     }
1229 
1230     memory_region_unref(section->mr);
1231 
1232     if (container->iommu_type == VFIO_SPAPR_TCE_v2_IOMMU) {
1233         vfio_spapr_remove_window(container,
1234                                  section->offset_within_address_space);
1235         if (vfio_host_win_del(container,
1236                               section->offset_within_address_space,
1237                               section->offset_within_address_space +
1238                               int128_get64(section->size) - 1) < 0) {
1239             hw_error("%s: Cannot delete missing window at %"HWADDR_PRIx,
1240                      __func__, section->offset_within_address_space);
1241         }
1242     }
1243 }
1244 
1245 static void vfio_set_dirty_page_tracking(VFIOContainer *container, bool start)
1246 {
1247     int ret;
1248     struct vfio_iommu_type1_dirty_bitmap dirty = {
1249         .argsz = sizeof(dirty),
1250     };
1251 
1252     if (start) {
1253         dirty.flags = VFIO_IOMMU_DIRTY_PAGES_FLAG_START;
1254     } else {
1255         dirty.flags = VFIO_IOMMU_DIRTY_PAGES_FLAG_STOP;
1256     }
1257 
1258     ret = ioctl(container->fd, VFIO_IOMMU_DIRTY_PAGES, &dirty);
1259     if (ret) {
1260         error_report("Failed to set dirty tracking flag 0x%x errno: %d",
1261                      dirty.flags, errno);
1262     }
1263 }
1264 
1265 static void vfio_listener_log_global_start(MemoryListener *listener)
1266 {
1267     VFIOContainer *container = container_of(listener, VFIOContainer, listener);
1268 
1269     vfio_set_dirty_page_tracking(container, true);
1270 }
1271 
1272 static void vfio_listener_log_global_stop(MemoryListener *listener)
1273 {
1274     VFIOContainer *container = container_of(listener, VFIOContainer, listener);
1275 
1276     vfio_set_dirty_page_tracking(container, false);
1277 }
1278 
1279 static int vfio_get_dirty_bitmap(VFIOContainer *container, uint64_t iova,
1280                                  uint64_t size, ram_addr_t ram_addr)
1281 {
1282     struct vfio_iommu_type1_dirty_bitmap *dbitmap;
1283     struct vfio_iommu_type1_dirty_bitmap_get *range;
1284     uint64_t pages;
1285     int ret;
1286 
1287     dbitmap = g_malloc0(sizeof(*dbitmap) + sizeof(*range));
1288 
1289     dbitmap->argsz = sizeof(*dbitmap) + sizeof(*range);
1290     dbitmap->flags = VFIO_IOMMU_DIRTY_PAGES_FLAG_GET_BITMAP;
1291     range = (struct vfio_iommu_type1_dirty_bitmap_get *)&dbitmap->data;
1292     range->iova = iova;
1293     range->size = size;
1294 
1295     /*
1296      * cpu_physical_memory_set_dirty_lebitmap() supports pages in bitmap of
1297      * qemu_real_host_page_size to mark those dirty. Hence set bitmap's pgsize
1298      * to qemu_real_host_page_size.
1299      */
1300     range->bitmap.pgsize = qemu_real_host_page_size();
1301 
1302     pages = REAL_HOST_PAGE_ALIGN(range->size) / qemu_real_host_page_size();
1303     range->bitmap.size = ROUND_UP(pages, sizeof(__u64) * BITS_PER_BYTE) /
1304                                          BITS_PER_BYTE;
1305     range->bitmap.data = g_try_malloc0(range->bitmap.size);
1306     if (!range->bitmap.data) {
1307         ret = -ENOMEM;
1308         goto err_out;
1309     }
1310 
1311     ret = ioctl(container->fd, VFIO_IOMMU_DIRTY_PAGES, dbitmap);
1312     if (ret) {
1313         error_report("Failed to get dirty bitmap for iova: 0x%"PRIx64
1314                 " size: 0x%"PRIx64" err: %d", (uint64_t)range->iova,
1315                 (uint64_t)range->size, errno);
1316         goto err_out;
1317     }
1318 
1319     cpu_physical_memory_set_dirty_lebitmap((unsigned long *)range->bitmap.data,
1320                                             ram_addr, pages);
1321 
1322     trace_vfio_get_dirty_bitmap(container->fd, range->iova, range->size,
1323                                 range->bitmap.size, ram_addr);
1324 err_out:
1325     g_free(range->bitmap.data);
1326     g_free(dbitmap);
1327 
1328     return ret;
1329 }
1330 
1331 typedef struct {
1332     IOMMUNotifier n;
1333     VFIOGuestIOMMU *giommu;
1334 } vfio_giommu_dirty_notifier;
1335 
1336 static void vfio_iommu_map_dirty_notify(IOMMUNotifier *n, IOMMUTLBEntry *iotlb)
1337 {
1338     vfio_giommu_dirty_notifier *gdn = container_of(n,
1339                                                 vfio_giommu_dirty_notifier, n);
1340     VFIOGuestIOMMU *giommu = gdn->giommu;
1341     VFIOContainer *container = giommu->container;
1342     hwaddr iova = iotlb->iova + giommu->iommu_offset;
1343     ram_addr_t translated_addr;
1344 
1345     trace_vfio_iommu_map_dirty_notify(iova, iova + iotlb->addr_mask);
1346 
1347     if (iotlb->target_as != &address_space_memory) {
1348         error_report("Wrong target AS \"%s\", only system memory is allowed",
1349                      iotlb->target_as->name ? iotlb->target_as->name : "none");
1350         return;
1351     }
1352 
1353     rcu_read_lock();
1354     if (vfio_get_xlat_addr(iotlb, NULL, &translated_addr, NULL)) {
1355         int ret;
1356 
1357         ret = vfio_get_dirty_bitmap(container, iova, iotlb->addr_mask + 1,
1358                                     translated_addr);
1359         if (ret) {
1360             error_report("vfio_iommu_map_dirty_notify(%p, 0x%"HWADDR_PRIx", "
1361                          "0x%"HWADDR_PRIx") = %d (%m)",
1362                          container, iova,
1363                          iotlb->addr_mask + 1, ret);
1364         }
1365     }
1366     rcu_read_unlock();
1367 }
1368 
1369 static int vfio_ram_discard_get_dirty_bitmap(MemoryRegionSection *section,
1370                                              void *opaque)
1371 {
1372     const hwaddr size = int128_get64(section->size);
1373     const hwaddr iova = section->offset_within_address_space;
1374     const ram_addr_t ram_addr = memory_region_get_ram_addr(section->mr) +
1375                                 section->offset_within_region;
1376     VFIORamDiscardListener *vrdl = opaque;
1377 
1378     /*
1379      * Sync the whole mapped region (spanning multiple individual mappings)
1380      * in one go.
1381      */
1382     return vfio_get_dirty_bitmap(vrdl->container, iova, size, ram_addr);
1383 }
1384 
1385 static int vfio_sync_ram_discard_listener_dirty_bitmap(VFIOContainer *container,
1386                                                    MemoryRegionSection *section)
1387 {
1388     RamDiscardManager *rdm = memory_region_get_ram_discard_manager(section->mr);
1389     VFIORamDiscardListener *vrdl = NULL;
1390 
1391     QLIST_FOREACH(vrdl, &container->vrdl_list, next) {
1392         if (vrdl->mr == section->mr &&
1393             vrdl->offset_within_address_space ==
1394             section->offset_within_address_space) {
1395             break;
1396         }
1397     }
1398 
1399     if (!vrdl) {
1400         hw_error("vfio: Trying to sync missing RAM discard listener");
1401     }
1402 
1403     /*
1404      * We only want/can synchronize the bitmap for actually mapped parts -
1405      * which correspond to populated parts. Replay all populated parts.
1406      */
1407     return ram_discard_manager_replay_populated(rdm, section,
1408                                               vfio_ram_discard_get_dirty_bitmap,
1409                                                 &vrdl);
1410 }
1411 
1412 static int vfio_sync_dirty_bitmap(VFIOContainer *container,
1413                                   MemoryRegionSection *section)
1414 {
1415     ram_addr_t ram_addr;
1416 
1417     if (memory_region_is_iommu(section->mr)) {
1418         VFIOGuestIOMMU *giommu;
1419 
1420         QLIST_FOREACH(giommu, &container->giommu_list, giommu_next) {
1421             if (MEMORY_REGION(giommu->iommu_mr) == section->mr &&
1422                 giommu->n.start == section->offset_within_region) {
1423                 Int128 llend;
1424                 vfio_giommu_dirty_notifier gdn = { .giommu = giommu };
1425                 int idx = memory_region_iommu_attrs_to_index(giommu->iommu_mr,
1426                                                        MEMTXATTRS_UNSPECIFIED);
1427 
1428                 llend = int128_add(int128_make64(section->offset_within_region),
1429                                    section->size);
1430                 llend = int128_sub(llend, int128_one());
1431 
1432                 iommu_notifier_init(&gdn.n,
1433                                     vfio_iommu_map_dirty_notify,
1434                                     IOMMU_NOTIFIER_MAP,
1435                                     section->offset_within_region,
1436                                     int128_get64(llend),
1437                                     idx);
1438                 memory_region_iommu_replay(giommu->iommu_mr, &gdn.n);
1439                 break;
1440             }
1441         }
1442         return 0;
1443     } else if (memory_region_has_ram_discard_manager(section->mr)) {
1444         return vfio_sync_ram_discard_listener_dirty_bitmap(container, section);
1445     }
1446 
1447     ram_addr = memory_region_get_ram_addr(section->mr) +
1448                section->offset_within_region;
1449 
1450     return vfio_get_dirty_bitmap(container,
1451                    REAL_HOST_PAGE_ALIGN(section->offset_within_address_space),
1452                    int128_get64(section->size), ram_addr);
1453 }
1454 
1455 static void vfio_listener_log_sync(MemoryListener *listener,
1456         MemoryRegionSection *section)
1457 {
1458     VFIOContainer *container = container_of(listener, VFIOContainer, listener);
1459 
1460     if (vfio_listener_skipped_section(section) ||
1461         !container->dirty_pages_supported) {
1462         return;
1463     }
1464 
1465     if (vfio_devices_all_dirty_tracking(container)) {
1466         vfio_sync_dirty_bitmap(container, section);
1467     }
1468 }
1469 
1470 static const MemoryListener vfio_memory_listener = {
1471     .name = "vfio",
1472     .region_add = vfio_listener_region_add,
1473     .region_del = vfio_listener_region_del,
1474     .log_global_start = vfio_listener_log_global_start,
1475     .log_global_stop = vfio_listener_log_global_stop,
1476     .log_sync = vfio_listener_log_sync,
1477 };
1478 
1479 static void vfio_listener_release(VFIOContainer *container)
1480 {
1481     memory_listener_unregister(&container->listener);
1482     if (container->iommu_type == VFIO_SPAPR_TCE_v2_IOMMU) {
1483         memory_listener_unregister(&container->prereg_listener);
1484     }
1485 }
1486 
1487 static struct vfio_info_cap_header *
1488 vfio_get_cap(void *ptr, uint32_t cap_offset, uint16_t id)
1489 {
1490     struct vfio_info_cap_header *hdr;
1491 
1492     for (hdr = ptr + cap_offset; hdr != ptr; hdr = ptr + hdr->next) {
1493         if (hdr->id == id) {
1494             return hdr;
1495         }
1496     }
1497 
1498     return NULL;
1499 }
1500 
1501 struct vfio_info_cap_header *
1502 vfio_get_region_info_cap(struct vfio_region_info *info, uint16_t id)
1503 {
1504     if (!(info->flags & VFIO_REGION_INFO_FLAG_CAPS)) {
1505         return NULL;
1506     }
1507 
1508     return vfio_get_cap((void *)info, info->cap_offset, id);
1509 }
1510 
1511 static struct vfio_info_cap_header *
1512 vfio_get_iommu_type1_info_cap(struct vfio_iommu_type1_info *info, uint16_t id)
1513 {
1514     if (!(info->flags & VFIO_IOMMU_INFO_CAPS)) {
1515         return NULL;
1516     }
1517 
1518     return vfio_get_cap((void *)info, info->cap_offset, id);
1519 }
1520 
1521 struct vfio_info_cap_header *
1522 vfio_get_device_info_cap(struct vfio_device_info *info, uint16_t id)
1523 {
1524     if (!(info->flags & VFIO_DEVICE_FLAGS_CAPS)) {
1525         return NULL;
1526     }
1527 
1528     return vfio_get_cap((void *)info, info->cap_offset, id);
1529 }
1530 
1531 bool vfio_get_info_dma_avail(struct vfio_iommu_type1_info *info,
1532                              unsigned int *avail)
1533 {
1534     struct vfio_info_cap_header *hdr;
1535     struct vfio_iommu_type1_info_dma_avail *cap;
1536 
1537     /* If the capability cannot be found, assume no DMA limiting */
1538     hdr = vfio_get_iommu_type1_info_cap(info,
1539                                         VFIO_IOMMU_TYPE1_INFO_DMA_AVAIL);
1540     if (hdr == NULL) {
1541         return false;
1542     }
1543 
1544     if (avail != NULL) {
1545         cap = (void *) hdr;
1546         *avail = cap->avail;
1547     }
1548 
1549     return true;
1550 }
1551 
1552 static int vfio_setup_region_sparse_mmaps(VFIORegion *region,
1553                                           struct vfio_region_info *info)
1554 {
1555     struct vfio_info_cap_header *hdr;
1556     struct vfio_region_info_cap_sparse_mmap *sparse;
1557     int i, j;
1558 
1559     hdr = vfio_get_region_info_cap(info, VFIO_REGION_INFO_CAP_SPARSE_MMAP);
1560     if (!hdr) {
1561         return -ENODEV;
1562     }
1563 
1564     sparse = container_of(hdr, struct vfio_region_info_cap_sparse_mmap, header);
1565 
1566     trace_vfio_region_sparse_mmap_header(region->vbasedev->name,
1567                                          region->nr, sparse->nr_areas);
1568 
1569     region->mmaps = g_new0(VFIOMmap, sparse->nr_areas);
1570 
1571     for (i = 0, j = 0; i < sparse->nr_areas; i++) {
1572         if (sparse->areas[i].size) {
1573             trace_vfio_region_sparse_mmap_entry(i, sparse->areas[i].offset,
1574                                             sparse->areas[i].offset +
1575                                             sparse->areas[i].size - 1);
1576             region->mmaps[j].offset = sparse->areas[i].offset;
1577             region->mmaps[j].size = sparse->areas[i].size;
1578             j++;
1579         }
1580     }
1581 
1582     region->nr_mmaps = j;
1583     region->mmaps = g_realloc(region->mmaps, j * sizeof(VFIOMmap));
1584 
1585     return 0;
1586 }
1587 
1588 int vfio_region_setup(Object *obj, VFIODevice *vbasedev, VFIORegion *region,
1589                       int index, const char *name)
1590 {
1591     struct vfio_region_info *info;
1592     int ret;
1593 
1594     ret = vfio_get_region_info(vbasedev, index, &info);
1595     if (ret) {
1596         return ret;
1597     }
1598 
1599     region->vbasedev = vbasedev;
1600     region->flags = info->flags;
1601     region->size = info->size;
1602     region->fd_offset = info->offset;
1603     region->nr = index;
1604 
1605     if (region->size) {
1606         region->mem = g_new0(MemoryRegion, 1);
1607         memory_region_init_io(region->mem, obj, &vfio_region_ops,
1608                               region, name, region->size);
1609 
1610         if (!vbasedev->no_mmap &&
1611             region->flags & VFIO_REGION_INFO_FLAG_MMAP) {
1612 
1613             ret = vfio_setup_region_sparse_mmaps(region, info);
1614 
1615             if (ret) {
1616                 region->nr_mmaps = 1;
1617                 region->mmaps = g_new0(VFIOMmap, region->nr_mmaps);
1618                 region->mmaps[0].offset = 0;
1619                 region->mmaps[0].size = region->size;
1620             }
1621         }
1622     }
1623 
1624     g_free(info);
1625 
1626     trace_vfio_region_setup(vbasedev->name, index, name,
1627                             region->flags, region->fd_offset, region->size);
1628     return 0;
1629 }
1630 
1631 static void vfio_subregion_unmap(VFIORegion *region, int index)
1632 {
1633     trace_vfio_region_unmap(memory_region_name(&region->mmaps[index].mem),
1634                             region->mmaps[index].offset,
1635                             region->mmaps[index].offset +
1636                             region->mmaps[index].size - 1);
1637     memory_region_del_subregion(region->mem, &region->mmaps[index].mem);
1638     munmap(region->mmaps[index].mmap, region->mmaps[index].size);
1639     object_unparent(OBJECT(&region->mmaps[index].mem));
1640     region->mmaps[index].mmap = NULL;
1641 }
1642 
1643 int vfio_region_mmap(VFIORegion *region)
1644 {
1645     int i, prot = 0;
1646     char *name;
1647 
1648     if (!region->mem) {
1649         return 0;
1650     }
1651 
1652     prot |= region->flags & VFIO_REGION_INFO_FLAG_READ ? PROT_READ : 0;
1653     prot |= region->flags & VFIO_REGION_INFO_FLAG_WRITE ? PROT_WRITE : 0;
1654 
1655     for (i = 0; i < region->nr_mmaps; i++) {
1656         region->mmaps[i].mmap = mmap(NULL, region->mmaps[i].size, prot,
1657                                      MAP_SHARED, region->vbasedev->fd,
1658                                      region->fd_offset +
1659                                      region->mmaps[i].offset);
1660         if (region->mmaps[i].mmap == MAP_FAILED) {
1661             int ret = -errno;
1662 
1663             trace_vfio_region_mmap_fault(memory_region_name(region->mem), i,
1664                                          region->fd_offset +
1665                                          region->mmaps[i].offset,
1666                                          region->fd_offset +
1667                                          region->mmaps[i].offset +
1668                                          region->mmaps[i].size - 1, ret);
1669 
1670             region->mmaps[i].mmap = NULL;
1671 
1672             for (i--; i >= 0; i--) {
1673                 vfio_subregion_unmap(region, i);
1674             }
1675 
1676             return ret;
1677         }
1678 
1679         name = g_strdup_printf("%s mmaps[%d]",
1680                                memory_region_name(region->mem), i);
1681         memory_region_init_ram_device_ptr(&region->mmaps[i].mem,
1682                                           memory_region_owner(region->mem),
1683                                           name, region->mmaps[i].size,
1684                                           region->mmaps[i].mmap);
1685         g_free(name);
1686         memory_region_add_subregion(region->mem, region->mmaps[i].offset,
1687                                     &region->mmaps[i].mem);
1688 
1689         trace_vfio_region_mmap(memory_region_name(&region->mmaps[i].mem),
1690                                region->mmaps[i].offset,
1691                                region->mmaps[i].offset +
1692                                region->mmaps[i].size - 1);
1693     }
1694 
1695     return 0;
1696 }
1697 
1698 void vfio_region_unmap(VFIORegion *region)
1699 {
1700     int i;
1701 
1702     if (!region->mem) {
1703         return;
1704     }
1705 
1706     for (i = 0; i < region->nr_mmaps; i++) {
1707         if (region->mmaps[i].mmap) {
1708             vfio_subregion_unmap(region, i);
1709         }
1710     }
1711 }
1712 
1713 void vfio_region_exit(VFIORegion *region)
1714 {
1715     int i;
1716 
1717     if (!region->mem) {
1718         return;
1719     }
1720 
1721     for (i = 0; i < region->nr_mmaps; i++) {
1722         if (region->mmaps[i].mmap) {
1723             memory_region_del_subregion(region->mem, &region->mmaps[i].mem);
1724         }
1725     }
1726 
1727     trace_vfio_region_exit(region->vbasedev->name, region->nr);
1728 }
1729 
1730 void vfio_region_finalize(VFIORegion *region)
1731 {
1732     int i;
1733 
1734     if (!region->mem) {
1735         return;
1736     }
1737 
1738     for (i = 0; i < region->nr_mmaps; i++) {
1739         if (region->mmaps[i].mmap) {
1740             munmap(region->mmaps[i].mmap, region->mmaps[i].size);
1741             object_unparent(OBJECT(&region->mmaps[i].mem));
1742         }
1743     }
1744 
1745     object_unparent(OBJECT(region->mem));
1746 
1747     g_free(region->mem);
1748     g_free(region->mmaps);
1749 
1750     trace_vfio_region_finalize(region->vbasedev->name, region->nr);
1751 
1752     region->mem = NULL;
1753     region->mmaps = NULL;
1754     region->nr_mmaps = 0;
1755     region->size = 0;
1756     region->flags = 0;
1757     region->nr = 0;
1758 }
1759 
1760 void vfio_region_mmaps_set_enabled(VFIORegion *region, bool enabled)
1761 {
1762     int i;
1763 
1764     if (!region->mem) {
1765         return;
1766     }
1767 
1768     for (i = 0; i < region->nr_mmaps; i++) {
1769         if (region->mmaps[i].mmap) {
1770             memory_region_set_enabled(&region->mmaps[i].mem, enabled);
1771         }
1772     }
1773 
1774     trace_vfio_region_mmaps_set_enabled(memory_region_name(region->mem),
1775                                         enabled);
1776 }
1777 
1778 void vfio_reset_handler(void *opaque)
1779 {
1780     VFIOGroup *group;
1781     VFIODevice *vbasedev;
1782 
1783     QLIST_FOREACH(group, &vfio_group_list, next) {
1784         QLIST_FOREACH(vbasedev, &group->device_list, next) {
1785             if (vbasedev->dev->realized) {
1786                 vbasedev->ops->vfio_compute_needs_reset(vbasedev);
1787             }
1788         }
1789     }
1790 
1791     QLIST_FOREACH(group, &vfio_group_list, next) {
1792         QLIST_FOREACH(vbasedev, &group->device_list, next) {
1793             if (vbasedev->dev->realized && vbasedev->needs_reset) {
1794                 vbasedev->ops->vfio_hot_reset_multi(vbasedev);
1795             }
1796         }
1797     }
1798 }
1799 
1800 static void vfio_kvm_device_add_group(VFIOGroup *group)
1801 {
1802 #ifdef CONFIG_KVM
1803     struct kvm_device_attr attr = {
1804         .group = KVM_DEV_VFIO_GROUP,
1805         .attr = KVM_DEV_VFIO_GROUP_ADD,
1806         .addr = (uint64_t)(unsigned long)&group->fd,
1807     };
1808 
1809     if (!kvm_enabled()) {
1810         return;
1811     }
1812 
1813     if (vfio_kvm_device_fd < 0) {
1814         struct kvm_create_device cd = {
1815             .type = KVM_DEV_TYPE_VFIO,
1816         };
1817 
1818         if (kvm_vm_ioctl(kvm_state, KVM_CREATE_DEVICE, &cd)) {
1819             error_report("Failed to create KVM VFIO device: %m");
1820             return;
1821         }
1822 
1823         vfio_kvm_device_fd = cd.fd;
1824     }
1825 
1826     if (ioctl(vfio_kvm_device_fd, KVM_SET_DEVICE_ATTR, &attr)) {
1827         error_report("Failed to add group %d to KVM VFIO device: %m",
1828                      group->groupid);
1829     }
1830 #endif
1831 }
1832 
1833 static void vfio_kvm_device_del_group(VFIOGroup *group)
1834 {
1835 #ifdef CONFIG_KVM
1836     struct kvm_device_attr attr = {
1837         .group = KVM_DEV_VFIO_GROUP,
1838         .attr = KVM_DEV_VFIO_GROUP_DEL,
1839         .addr = (uint64_t)(unsigned long)&group->fd,
1840     };
1841 
1842     if (vfio_kvm_device_fd < 0) {
1843         return;
1844     }
1845 
1846     if (ioctl(vfio_kvm_device_fd, KVM_SET_DEVICE_ATTR, &attr)) {
1847         error_report("Failed to remove group %d from KVM VFIO device: %m",
1848                      group->groupid);
1849     }
1850 #endif
1851 }
1852 
1853 static VFIOAddressSpace *vfio_get_address_space(AddressSpace *as)
1854 {
1855     VFIOAddressSpace *space;
1856 
1857     QLIST_FOREACH(space, &vfio_address_spaces, list) {
1858         if (space->as == as) {
1859             return space;
1860         }
1861     }
1862 
1863     /* No suitable VFIOAddressSpace, create a new one */
1864     space = g_malloc0(sizeof(*space));
1865     space->as = as;
1866     QLIST_INIT(&space->containers);
1867 
1868     QLIST_INSERT_HEAD(&vfio_address_spaces, space, list);
1869 
1870     return space;
1871 }
1872 
1873 static void vfio_put_address_space(VFIOAddressSpace *space)
1874 {
1875     if (QLIST_EMPTY(&space->containers)) {
1876         QLIST_REMOVE(space, list);
1877         g_free(space);
1878     }
1879 }
1880 
1881 /*
1882  * vfio_get_iommu_type - selects the richest iommu_type (v2 first)
1883  */
1884 static int vfio_get_iommu_type(VFIOContainer *container,
1885                                Error **errp)
1886 {
1887     int iommu_types[] = { VFIO_TYPE1v2_IOMMU, VFIO_TYPE1_IOMMU,
1888                           VFIO_SPAPR_TCE_v2_IOMMU, VFIO_SPAPR_TCE_IOMMU };
1889     int i;
1890 
1891     for (i = 0; i < ARRAY_SIZE(iommu_types); i++) {
1892         if (ioctl(container->fd, VFIO_CHECK_EXTENSION, iommu_types[i])) {
1893             return iommu_types[i];
1894         }
1895     }
1896     error_setg(errp, "No available IOMMU models");
1897     return -EINVAL;
1898 }
1899 
1900 static int vfio_init_container(VFIOContainer *container, int group_fd,
1901                                Error **errp)
1902 {
1903     int iommu_type, ret;
1904 
1905     iommu_type = vfio_get_iommu_type(container, errp);
1906     if (iommu_type < 0) {
1907         return iommu_type;
1908     }
1909 
1910     ret = ioctl(group_fd, VFIO_GROUP_SET_CONTAINER, &container->fd);
1911     if (ret) {
1912         error_setg_errno(errp, errno, "Failed to set group container");
1913         return -errno;
1914     }
1915 
1916     while (ioctl(container->fd, VFIO_SET_IOMMU, iommu_type)) {
1917         if (iommu_type == VFIO_SPAPR_TCE_v2_IOMMU) {
1918             /*
1919              * On sPAPR, despite the IOMMU subdriver always advertises v1 and
1920              * v2, the running platform may not support v2 and there is no
1921              * way to guess it until an IOMMU group gets added to the container.
1922              * So in case it fails with v2, try v1 as a fallback.
1923              */
1924             iommu_type = VFIO_SPAPR_TCE_IOMMU;
1925             continue;
1926         }
1927         error_setg_errno(errp, errno, "Failed to set iommu for container");
1928         return -errno;
1929     }
1930 
1931     container->iommu_type = iommu_type;
1932     return 0;
1933 }
1934 
1935 static int vfio_get_iommu_info(VFIOContainer *container,
1936                                struct vfio_iommu_type1_info **info)
1937 {
1938 
1939     size_t argsz = sizeof(struct vfio_iommu_type1_info);
1940 
1941     *info = g_new0(struct vfio_iommu_type1_info, 1);
1942 again:
1943     (*info)->argsz = argsz;
1944 
1945     if (ioctl(container->fd, VFIO_IOMMU_GET_INFO, *info)) {
1946         g_free(*info);
1947         *info = NULL;
1948         return -errno;
1949     }
1950 
1951     if (((*info)->argsz > argsz)) {
1952         argsz = (*info)->argsz;
1953         *info = g_realloc(*info, argsz);
1954         goto again;
1955     }
1956 
1957     return 0;
1958 }
1959 
1960 static struct vfio_info_cap_header *
1961 vfio_get_iommu_info_cap(struct vfio_iommu_type1_info *info, uint16_t id)
1962 {
1963     struct vfio_info_cap_header *hdr;
1964     void *ptr = info;
1965 
1966     if (!(info->flags & VFIO_IOMMU_INFO_CAPS)) {
1967         return NULL;
1968     }
1969 
1970     for (hdr = ptr + info->cap_offset; hdr != ptr; hdr = ptr + hdr->next) {
1971         if (hdr->id == id) {
1972             return hdr;
1973         }
1974     }
1975 
1976     return NULL;
1977 }
1978 
1979 static void vfio_get_iommu_info_migration(VFIOContainer *container,
1980                                          struct vfio_iommu_type1_info *info)
1981 {
1982     struct vfio_info_cap_header *hdr;
1983     struct vfio_iommu_type1_info_cap_migration *cap_mig;
1984 
1985     hdr = vfio_get_iommu_info_cap(info, VFIO_IOMMU_TYPE1_INFO_CAP_MIGRATION);
1986     if (!hdr) {
1987         return;
1988     }
1989 
1990     cap_mig = container_of(hdr, struct vfio_iommu_type1_info_cap_migration,
1991                             header);
1992 
1993     /*
1994      * cpu_physical_memory_set_dirty_lebitmap() supports pages in bitmap of
1995      * qemu_real_host_page_size to mark those dirty.
1996      */
1997     if (cap_mig->pgsize_bitmap & qemu_real_host_page_size()) {
1998         container->dirty_pages_supported = true;
1999         container->max_dirty_bitmap_size = cap_mig->max_dirty_bitmap_size;
2000         container->dirty_pgsizes = cap_mig->pgsize_bitmap;
2001     }
2002 }
2003 
2004 static int vfio_connect_container(VFIOGroup *group, AddressSpace *as,
2005                                   Error **errp)
2006 {
2007     VFIOContainer *container;
2008     int ret, fd;
2009     VFIOAddressSpace *space;
2010 
2011     space = vfio_get_address_space(as);
2012 
2013     /*
2014      * VFIO is currently incompatible with discarding of RAM insofar as the
2015      * madvise to purge (zap) the page from QEMU's address space does not
2016      * interact with the memory API and therefore leaves stale virtual to
2017      * physical mappings in the IOMMU if the page was previously pinned.  We
2018      * therefore set discarding broken for each group added to a container,
2019      * whether the container is used individually or shared.  This provides
2020      * us with options to allow devices within a group to opt-in and allow
2021      * discarding, so long as it is done consistently for a group (for instance
2022      * if the device is an mdev device where it is known that the host vendor
2023      * driver will never pin pages outside of the working set of the guest
2024      * driver, which would thus not be discarding candidates).
2025      *
2026      * The first opportunity to induce pinning occurs here where we attempt to
2027      * attach the group to existing containers within the AddressSpace.  If any
2028      * pages are already zapped from the virtual address space, such as from
2029      * previous discards, new pinning will cause valid mappings to be
2030      * re-established.  Likewise, when the overall MemoryListener for a new
2031      * container is registered, a replay of mappings within the AddressSpace
2032      * will occur, re-establishing any previously zapped pages as well.
2033      *
2034      * Especially virtio-balloon is currently only prevented from discarding
2035      * new memory, it will not yet set ram_block_discard_set_required() and
2036      * therefore, neither stops us here or deals with the sudden memory
2037      * consumption of inflated memory.
2038      *
2039      * We do support discarding of memory coordinated via the RamDiscardManager
2040      * with some IOMMU types. vfio_ram_block_discard_disable() handles the
2041      * details once we know which type of IOMMU we are using.
2042      */
2043 
2044     QLIST_FOREACH(container, &space->containers, next) {
2045         if (!ioctl(group->fd, VFIO_GROUP_SET_CONTAINER, &container->fd)) {
2046             ret = vfio_ram_block_discard_disable(container, true);
2047             if (ret) {
2048                 error_setg_errno(errp, -ret,
2049                                  "Cannot set discarding of RAM broken");
2050                 if (ioctl(group->fd, VFIO_GROUP_UNSET_CONTAINER,
2051                           &container->fd)) {
2052                     error_report("vfio: error disconnecting group %d from"
2053                                  " container", group->groupid);
2054                 }
2055                 return ret;
2056             }
2057             group->container = container;
2058             QLIST_INSERT_HEAD(&container->group_list, group, container_next);
2059             vfio_kvm_device_add_group(group);
2060             return 0;
2061         }
2062     }
2063 
2064     fd = qemu_open_old("/dev/vfio/vfio", O_RDWR);
2065     if (fd < 0) {
2066         error_setg_errno(errp, errno, "failed to open /dev/vfio/vfio");
2067         ret = -errno;
2068         goto put_space_exit;
2069     }
2070 
2071     ret = ioctl(fd, VFIO_GET_API_VERSION);
2072     if (ret != VFIO_API_VERSION) {
2073         error_setg(errp, "supported vfio version: %d, "
2074                    "reported version: %d", VFIO_API_VERSION, ret);
2075         ret = -EINVAL;
2076         goto close_fd_exit;
2077     }
2078 
2079     container = g_malloc0(sizeof(*container));
2080     container->space = space;
2081     container->fd = fd;
2082     container->error = NULL;
2083     container->dirty_pages_supported = false;
2084     container->dma_max_mappings = 0;
2085     QLIST_INIT(&container->giommu_list);
2086     QLIST_INIT(&container->hostwin_list);
2087     QLIST_INIT(&container->vrdl_list);
2088 
2089     ret = vfio_init_container(container, group->fd, errp);
2090     if (ret) {
2091         goto free_container_exit;
2092     }
2093 
2094     ret = vfio_ram_block_discard_disable(container, true);
2095     if (ret) {
2096         error_setg_errno(errp, -ret, "Cannot set discarding of RAM broken");
2097         goto free_container_exit;
2098     }
2099 
2100     switch (container->iommu_type) {
2101     case VFIO_TYPE1v2_IOMMU:
2102     case VFIO_TYPE1_IOMMU:
2103     {
2104         struct vfio_iommu_type1_info *info;
2105 
2106         /*
2107          * FIXME: This assumes that a Type1 IOMMU can map any 64-bit
2108          * IOVA whatsoever.  That's not actually true, but the current
2109          * kernel interface doesn't tell us what it can map, and the
2110          * existing Type1 IOMMUs generally support any IOVA we're
2111          * going to actually try in practice.
2112          */
2113         ret = vfio_get_iommu_info(container, &info);
2114 
2115         if (ret || !(info->flags & VFIO_IOMMU_INFO_PGSIZES)) {
2116             /* Assume 4k IOVA page size */
2117             info->iova_pgsizes = 4096;
2118         }
2119         vfio_host_win_add(container, 0, (hwaddr)-1, info->iova_pgsizes);
2120         container->pgsizes = info->iova_pgsizes;
2121 
2122         /* The default in the kernel ("dma_entry_limit") is 65535. */
2123         container->dma_max_mappings = 65535;
2124         if (!ret) {
2125             vfio_get_info_dma_avail(info, &container->dma_max_mappings);
2126             vfio_get_iommu_info_migration(container, info);
2127         }
2128         g_free(info);
2129         break;
2130     }
2131     case VFIO_SPAPR_TCE_v2_IOMMU:
2132     case VFIO_SPAPR_TCE_IOMMU:
2133     {
2134         struct vfio_iommu_spapr_tce_info info;
2135         bool v2 = container->iommu_type == VFIO_SPAPR_TCE_v2_IOMMU;
2136 
2137         /*
2138          * The host kernel code implementing VFIO_IOMMU_DISABLE is called
2139          * when container fd is closed so we do not call it explicitly
2140          * in this file.
2141          */
2142         if (!v2) {
2143             ret = ioctl(fd, VFIO_IOMMU_ENABLE);
2144             if (ret) {
2145                 error_setg_errno(errp, errno, "failed to enable container");
2146                 ret = -errno;
2147                 goto enable_discards_exit;
2148             }
2149         } else {
2150             container->prereg_listener = vfio_prereg_listener;
2151 
2152             memory_listener_register(&container->prereg_listener,
2153                                      &address_space_memory);
2154             if (container->error) {
2155                 memory_listener_unregister(&container->prereg_listener);
2156                 ret = -1;
2157                 error_propagate_prepend(errp, container->error,
2158                     "RAM memory listener initialization failed: ");
2159                 goto enable_discards_exit;
2160             }
2161         }
2162 
2163         info.argsz = sizeof(info);
2164         ret = ioctl(fd, VFIO_IOMMU_SPAPR_TCE_GET_INFO, &info);
2165         if (ret) {
2166             error_setg_errno(errp, errno,
2167                              "VFIO_IOMMU_SPAPR_TCE_GET_INFO failed");
2168             ret = -errno;
2169             if (v2) {
2170                 memory_listener_unregister(&container->prereg_listener);
2171             }
2172             goto enable_discards_exit;
2173         }
2174 
2175         if (v2) {
2176             container->pgsizes = info.ddw.pgsizes;
2177             /*
2178              * There is a default window in just created container.
2179              * To make region_add/del simpler, we better remove this
2180              * window now and let those iommu_listener callbacks
2181              * create/remove them when needed.
2182              */
2183             ret = vfio_spapr_remove_window(container, info.dma32_window_start);
2184             if (ret) {
2185                 error_setg_errno(errp, -ret,
2186                                  "failed to remove existing window");
2187                 goto enable_discards_exit;
2188             }
2189         } else {
2190             /* The default table uses 4K pages */
2191             container->pgsizes = 0x1000;
2192             vfio_host_win_add(container, info.dma32_window_start,
2193                               info.dma32_window_start +
2194                               info.dma32_window_size - 1,
2195                               0x1000);
2196         }
2197     }
2198     }
2199 
2200     vfio_kvm_device_add_group(group);
2201 
2202     QLIST_INIT(&container->group_list);
2203     QLIST_INSERT_HEAD(&space->containers, container, next);
2204 
2205     group->container = container;
2206     QLIST_INSERT_HEAD(&container->group_list, group, container_next);
2207 
2208     container->listener = vfio_memory_listener;
2209 
2210     memory_listener_register(&container->listener, container->space->as);
2211 
2212     if (container->error) {
2213         ret = -1;
2214         error_propagate_prepend(errp, container->error,
2215             "memory listener initialization failed: ");
2216         goto listener_release_exit;
2217     }
2218 
2219     container->initialized = true;
2220 
2221     return 0;
2222 listener_release_exit:
2223     QLIST_REMOVE(group, container_next);
2224     QLIST_REMOVE(container, next);
2225     vfio_kvm_device_del_group(group);
2226     vfio_listener_release(container);
2227 
2228 enable_discards_exit:
2229     vfio_ram_block_discard_disable(container, false);
2230 
2231 free_container_exit:
2232     g_free(container);
2233 
2234 close_fd_exit:
2235     close(fd);
2236 
2237 put_space_exit:
2238     vfio_put_address_space(space);
2239 
2240     return ret;
2241 }
2242 
2243 static void vfio_disconnect_container(VFIOGroup *group)
2244 {
2245     VFIOContainer *container = group->container;
2246 
2247     QLIST_REMOVE(group, container_next);
2248     group->container = NULL;
2249 
2250     /*
2251      * Explicitly release the listener first before unset container,
2252      * since unset may destroy the backend container if it's the last
2253      * group.
2254      */
2255     if (QLIST_EMPTY(&container->group_list)) {
2256         vfio_listener_release(container);
2257     }
2258 
2259     if (ioctl(group->fd, VFIO_GROUP_UNSET_CONTAINER, &container->fd)) {
2260         error_report("vfio: error disconnecting group %d from container",
2261                      group->groupid);
2262     }
2263 
2264     if (QLIST_EMPTY(&container->group_list)) {
2265         VFIOAddressSpace *space = container->space;
2266         VFIOGuestIOMMU *giommu, *tmp;
2267         VFIOHostDMAWindow *hostwin, *next;
2268 
2269         QLIST_REMOVE(container, next);
2270 
2271         QLIST_FOREACH_SAFE(giommu, &container->giommu_list, giommu_next, tmp) {
2272             memory_region_unregister_iommu_notifier(
2273                     MEMORY_REGION(giommu->iommu_mr), &giommu->n);
2274             QLIST_REMOVE(giommu, giommu_next);
2275             g_free(giommu);
2276         }
2277 
2278         QLIST_FOREACH_SAFE(hostwin, &container->hostwin_list, hostwin_next,
2279                            next) {
2280             QLIST_REMOVE(hostwin, hostwin_next);
2281             g_free(hostwin);
2282         }
2283 
2284         trace_vfio_disconnect_container(container->fd);
2285         close(container->fd);
2286         g_free(container);
2287 
2288         vfio_put_address_space(space);
2289     }
2290 }
2291 
2292 VFIOGroup *vfio_get_group(int groupid, AddressSpace *as, Error **errp)
2293 {
2294     VFIOGroup *group;
2295     char path[32];
2296     struct vfio_group_status status = { .argsz = sizeof(status) };
2297 
2298     QLIST_FOREACH(group, &vfio_group_list, next) {
2299         if (group->groupid == groupid) {
2300             /* Found it.  Now is it already in the right context? */
2301             if (group->container->space->as == as) {
2302                 return group;
2303             } else {
2304                 error_setg(errp, "group %d used in multiple address spaces",
2305                            group->groupid);
2306                 return NULL;
2307             }
2308         }
2309     }
2310 
2311     group = g_malloc0(sizeof(*group));
2312 
2313     snprintf(path, sizeof(path), "/dev/vfio/%d", groupid);
2314     group->fd = qemu_open_old(path, O_RDWR);
2315     if (group->fd < 0) {
2316         error_setg_errno(errp, errno, "failed to open %s", path);
2317         goto free_group_exit;
2318     }
2319 
2320     if (ioctl(group->fd, VFIO_GROUP_GET_STATUS, &status)) {
2321         error_setg_errno(errp, errno, "failed to get group %d status", groupid);
2322         goto close_fd_exit;
2323     }
2324 
2325     if (!(status.flags & VFIO_GROUP_FLAGS_VIABLE)) {
2326         error_setg(errp, "group %d is not viable", groupid);
2327         error_append_hint(errp,
2328                           "Please ensure all devices within the iommu_group "
2329                           "are bound to their vfio bus driver.\n");
2330         goto close_fd_exit;
2331     }
2332 
2333     group->groupid = groupid;
2334     QLIST_INIT(&group->device_list);
2335 
2336     if (vfio_connect_container(group, as, errp)) {
2337         error_prepend(errp, "failed to setup container for group %d: ",
2338                       groupid);
2339         goto close_fd_exit;
2340     }
2341 
2342     if (QLIST_EMPTY(&vfio_group_list)) {
2343         qemu_register_reset(vfio_reset_handler, NULL);
2344     }
2345 
2346     QLIST_INSERT_HEAD(&vfio_group_list, group, next);
2347 
2348     return group;
2349 
2350 close_fd_exit:
2351     close(group->fd);
2352 
2353 free_group_exit:
2354     g_free(group);
2355 
2356     return NULL;
2357 }
2358 
2359 void vfio_put_group(VFIOGroup *group)
2360 {
2361     if (!group || !QLIST_EMPTY(&group->device_list)) {
2362         return;
2363     }
2364 
2365     if (!group->ram_block_discard_allowed) {
2366         vfio_ram_block_discard_disable(group->container, false);
2367     }
2368     vfio_kvm_device_del_group(group);
2369     vfio_disconnect_container(group);
2370     QLIST_REMOVE(group, next);
2371     trace_vfio_put_group(group->fd);
2372     close(group->fd);
2373     g_free(group);
2374 
2375     if (QLIST_EMPTY(&vfio_group_list)) {
2376         qemu_unregister_reset(vfio_reset_handler, NULL);
2377     }
2378 }
2379 
2380 int vfio_get_device(VFIOGroup *group, const char *name,
2381                     VFIODevice *vbasedev, Error **errp)
2382 {
2383     struct vfio_device_info dev_info = { .argsz = sizeof(dev_info) };
2384     int ret, fd;
2385 
2386     fd = ioctl(group->fd, VFIO_GROUP_GET_DEVICE_FD, name);
2387     if (fd < 0) {
2388         error_setg_errno(errp, errno, "error getting device from group %d",
2389                          group->groupid);
2390         error_append_hint(errp,
2391                       "Verify all devices in group %d are bound to vfio-<bus> "
2392                       "or pci-stub and not already in use\n", group->groupid);
2393         return fd;
2394     }
2395 
2396     ret = ioctl(fd, VFIO_DEVICE_GET_INFO, &dev_info);
2397     if (ret) {
2398         error_setg_errno(errp, errno, "error getting device info");
2399         close(fd);
2400         return ret;
2401     }
2402 
2403     /*
2404      * Set discarding of RAM as not broken for this group if the driver knows
2405      * the device operates compatibly with discarding.  Setting must be
2406      * consistent per group, but since compatibility is really only possible
2407      * with mdev currently, we expect singleton groups.
2408      */
2409     if (vbasedev->ram_block_discard_allowed !=
2410         group->ram_block_discard_allowed) {
2411         if (!QLIST_EMPTY(&group->device_list)) {
2412             error_setg(errp, "Inconsistent setting of support for discarding "
2413                        "RAM (e.g., balloon) within group");
2414             close(fd);
2415             return -1;
2416         }
2417 
2418         if (!group->ram_block_discard_allowed) {
2419             group->ram_block_discard_allowed = true;
2420             vfio_ram_block_discard_disable(group->container, false);
2421         }
2422     }
2423 
2424     vbasedev->fd = fd;
2425     vbasedev->group = group;
2426     QLIST_INSERT_HEAD(&group->device_list, vbasedev, next);
2427 
2428     vbasedev->num_irqs = dev_info.num_irqs;
2429     vbasedev->num_regions = dev_info.num_regions;
2430     vbasedev->flags = dev_info.flags;
2431 
2432     trace_vfio_get_device(name, dev_info.flags, dev_info.num_regions,
2433                           dev_info.num_irqs);
2434 
2435     vbasedev->reset_works = !!(dev_info.flags & VFIO_DEVICE_FLAGS_RESET);
2436     return 0;
2437 }
2438 
2439 void vfio_put_base_device(VFIODevice *vbasedev)
2440 {
2441     if (!vbasedev->group) {
2442         return;
2443     }
2444     QLIST_REMOVE(vbasedev, next);
2445     vbasedev->group = NULL;
2446     trace_vfio_put_base_device(vbasedev->fd);
2447     close(vbasedev->fd);
2448 }
2449 
2450 int vfio_get_region_info(VFIODevice *vbasedev, int index,
2451                          struct vfio_region_info **info)
2452 {
2453     size_t argsz = sizeof(struct vfio_region_info);
2454 
2455     *info = g_malloc0(argsz);
2456 
2457     (*info)->index = index;
2458 retry:
2459     (*info)->argsz = argsz;
2460 
2461     if (ioctl(vbasedev->fd, VFIO_DEVICE_GET_REGION_INFO, *info)) {
2462         g_free(*info);
2463         *info = NULL;
2464         return -errno;
2465     }
2466 
2467     if ((*info)->argsz > argsz) {
2468         argsz = (*info)->argsz;
2469         *info = g_realloc(*info, argsz);
2470 
2471         goto retry;
2472     }
2473 
2474     return 0;
2475 }
2476 
2477 int vfio_get_dev_region_info(VFIODevice *vbasedev, uint32_t type,
2478                              uint32_t subtype, struct vfio_region_info **info)
2479 {
2480     int i;
2481 
2482     for (i = 0; i < vbasedev->num_regions; i++) {
2483         struct vfio_info_cap_header *hdr;
2484         struct vfio_region_info_cap_type *cap_type;
2485 
2486         if (vfio_get_region_info(vbasedev, i, info)) {
2487             continue;
2488         }
2489 
2490         hdr = vfio_get_region_info_cap(*info, VFIO_REGION_INFO_CAP_TYPE);
2491         if (!hdr) {
2492             g_free(*info);
2493             continue;
2494         }
2495 
2496         cap_type = container_of(hdr, struct vfio_region_info_cap_type, header);
2497 
2498         trace_vfio_get_dev_region(vbasedev->name, i,
2499                                   cap_type->type, cap_type->subtype);
2500 
2501         if (cap_type->type == type && cap_type->subtype == subtype) {
2502             return 0;
2503         }
2504 
2505         g_free(*info);
2506     }
2507 
2508     *info = NULL;
2509     return -ENODEV;
2510 }
2511 
2512 bool vfio_has_region_cap(VFIODevice *vbasedev, int region, uint16_t cap_type)
2513 {
2514     struct vfio_region_info *info = NULL;
2515     bool ret = false;
2516 
2517     if (!vfio_get_region_info(vbasedev, region, &info)) {
2518         if (vfio_get_region_info_cap(info, cap_type)) {
2519             ret = true;
2520         }
2521         g_free(info);
2522     }
2523 
2524     return ret;
2525 }
2526 
2527 /*
2528  * Interfaces for IBM EEH (Enhanced Error Handling)
2529  */
2530 static bool vfio_eeh_container_ok(VFIOContainer *container)
2531 {
2532     /*
2533      * As of 2016-03-04 (linux-4.5) the host kernel EEH/VFIO
2534      * implementation is broken if there are multiple groups in a
2535      * container.  The hardware works in units of Partitionable
2536      * Endpoints (== IOMMU groups) and the EEH operations naively
2537      * iterate across all groups in the container, without any logic
2538      * to make sure the groups have their state synchronized.  For
2539      * certain operations (ENABLE) that might be ok, until an error
2540      * occurs, but for others (GET_STATE) it's clearly broken.
2541      */
2542 
2543     /*
2544      * XXX Once fixed kernels exist, test for them here
2545      */
2546 
2547     if (QLIST_EMPTY(&container->group_list)) {
2548         return false;
2549     }
2550 
2551     if (QLIST_NEXT(QLIST_FIRST(&container->group_list), container_next)) {
2552         return false;
2553     }
2554 
2555     return true;
2556 }
2557 
2558 static int vfio_eeh_container_op(VFIOContainer *container, uint32_t op)
2559 {
2560     struct vfio_eeh_pe_op pe_op = {
2561         .argsz = sizeof(pe_op),
2562         .op = op,
2563     };
2564     int ret;
2565 
2566     if (!vfio_eeh_container_ok(container)) {
2567         error_report("vfio/eeh: EEH_PE_OP 0x%x: "
2568                      "kernel requires a container with exactly one group", op);
2569         return -EPERM;
2570     }
2571 
2572     ret = ioctl(container->fd, VFIO_EEH_PE_OP, &pe_op);
2573     if (ret < 0) {
2574         error_report("vfio/eeh: EEH_PE_OP 0x%x failed: %m", op);
2575         return -errno;
2576     }
2577 
2578     return ret;
2579 }
2580 
2581 static VFIOContainer *vfio_eeh_as_container(AddressSpace *as)
2582 {
2583     VFIOAddressSpace *space = vfio_get_address_space(as);
2584     VFIOContainer *container = NULL;
2585 
2586     if (QLIST_EMPTY(&space->containers)) {
2587         /* No containers to act on */
2588         goto out;
2589     }
2590 
2591     container = QLIST_FIRST(&space->containers);
2592 
2593     if (QLIST_NEXT(container, next)) {
2594         /* We don't yet have logic to synchronize EEH state across
2595          * multiple containers */
2596         container = NULL;
2597         goto out;
2598     }
2599 
2600 out:
2601     vfio_put_address_space(space);
2602     return container;
2603 }
2604 
2605 bool vfio_eeh_as_ok(AddressSpace *as)
2606 {
2607     VFIOContainer *container = vfio_eeh_as_container(as);
2608 
2609     return (container != NULL) && vfio_eeh_container_ok(container);
2610 }
2611 
2612 int vfio_eeh_as_op(AddressSpace *as, uint32_t op)
2613 {
2614     VFIOContainer *container = vfio_eeh_as_container(as);
2615 
2616     if (!container) {
2617         return -ENODEV;
2618     }
2619     return vfio_eeh_container_op(container, op);
2620 }
2621