xref: /openbmc/qemu/hw/vfio/common.c (revision 88cd34ee)
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 "hw/hw.h"
33 #include "qemu/error-report.h"
34 #include "qemu/main-loop.h"
35 #include "qemu/range.h"
36 #include "sysemu/balloon.h"
37 #include "sysemu/kvm.h"
38 #include "sysemu/reset.h"
39 #include "trace.h"
40 #include "qapi/error.h"
41 
42 VFIOGroupList vfio_group_list =
43     QLIST_HEAD_INITIALIZER(vfio_group_list);
44 static QLIST_HEAD(, VFIOAddressSpace) vfio_address_spaces =
45     QLIST_HEAD_INITIALIZER(vfio_address_spaces);
46 
47 #ifdef CONFIG_KVM
48 /*
49  * We have a single VFIO pseudo device per KVM VM.  Once created it lives
50  * for the life of the VM.  Closing the file descriptor only drops our
51  * reference to it and the device's reference to kvm.  Therefore once
52  * initialized, this file descriptor is only released on QEMU exit and
53  * we'll re-use it should another vfio device be attached before then.
54  */
55 static int vfio_kvm_device_fd = -1;
56 #endif
57 
58 /*
59  * Common VFIO interrupt disable
60  */
61 void vfio_disable_irqindex(VFIODevice *vbasedev, int index)
62 {
63     struct vfio_irq_set irq_set = {
64         .argsz = sizeof(irq_set),
65         .flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_TRIGGER,
66         .index = index,
67         .start = 0,
68         .count = 0,
69     };
70 
71     ioctl(vbasedev->fd, VFIO_DEVICE_SET_IRQS, &irq_set);
72 }
73 
74 void vfio_unmask_single_irqindex(VFIODevice *vbasedev, int index)
75 {
76     struct vfio_irq_set irq_set = {
77         .argsz = sizeof(irq_set),
78         .flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_UNMASK,
79         .index = index,
80         .start = 0,
81         .count = 1,
82     };
83 
84     ioctl(vbasedev->fd, VFIO_DEVICE_SET_IRQS, &irq_set);
85 }
86 
87 void vfio_mask_single_irqindex(VFIODevice *vbasedev, int index)
88 {
89     struct vfio_irq_set irq_set = {
90         .argsz = sizeof(irq_set),
91         .flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_MASK,
92         .index = index,
93         .start = 0,
94         .count = 1,
95     };
96 
97     ioctl(vbasedev->fd, VFIO_DEVICE_SET_IRQS, &irq_set);
98 }
99 
100 static inline const char *action_to_str(int action)
101 {
102     switch (action) {
103     case VFIO_IRQ_SET_ACTION_MASK:
104         return "MASK";
105     case VFIO_IRQ_SET_ACTION_UNMASK:
106         return "UNMASK";
107     case VFIO_IRQ_SET_ACTION_TRIGGER:
108         return "TRIGGER";
109     default:
110         return "UNKNOWN ACTION";
111     }
112 }
113 
114 static const char *index_to_str(VFIODevice *vbasedev, int index)
115 {
116     if (vbasedev->type != VFIO_DEVICE_TYPE_PCI) {
117         return NULL;
118     }
119 
120     switch (index) {
121     case VFIO_PCI_INTX_IRQ_INDEX:
122         return "INTX";
123     case VFIO_PCI_MSI_IRQ_INDEX:
124         return "MSI";
125     case VFIO_PCI_MSIX_IRQ_INDEX:
126         return "MSIX";
127     case VFIO_PCI_ERR_IRQ_INDEX:
128         return "ERR";
129     case VFIO_PCI_REQ_IRQ_INDEX:
130         return "REQ";
131     default:
132         return NULL;
133     }
134 }
135 
136 int vfio_set_irq_signaling(VFIODevice *vbasedev, int index, int subindex,
137                            int action, int fd, Error **errp)
138 {
139     struct vfio_irq_set *irq_set;
140     int argsz, ret = 0;
141     const char *name;
142     int32_t *pfd;
143 
144     argsz = sizeof(*irq_set) + sizeof(*pfd);
145 
146     irq_set = g_malloc0(argsz);
147     irq_set->argsz = argsz;
148     irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD | action;
149     irq_set->index = index;
150     irq_set->start = subindex;
151     irq_set->count = 1;
152     pfd = (int32_t *)&irq_set->data;
153     *pfd = fd;
154 
155     if (ioctl(vbasedev->fd, VFIO_DEVICE_SET_IRQS, irq_set)) {
156         ret = -errno;
157     }
158     g_free(irq_set);
159 
160     if (!ret) {
161         return 0;
162     }
163 
164     error_setg_errno(errp, -ret, "VFIO_DEVICE_SET_IRQS failure");
165 
166     name = index_to_str(vbasedev, index);
167     if (name) {
168         error_prepend(errp, "%s-%d: ", name, subindex);
169     } else {
170         error_prepend(errp, "index %d-%d: ", index, subindex);
171     }
172     error_prepend(errp,
173                   "Failed to %s %s eventfd signaling for interrupt ",
174                   fd < 0 ? "tear down" : "set up", action_to_str(action));
175     return ret;
176 }
177 
178 /*
179  * IO Port/MMIO - Beware of the endians, VFIO is always little endian
180  */
181 void vfio_region_write(void *opaque, hwaddr addr,
182                        uint64_t data, unsigned size)
183 {
184     VFIORegion *region = opaque;
185     VFIODevice *vbasedev = region->vbasedev;
186     union {
187         uint8_t byte;
188         uint16_t word;
189         uint32_t dword;
190         uint64_t qword;
191     } buf;
192 
193     switch (size) {
194     case 1:
195         buf.byte = data;
196         break;
197     case 2:
198         buf.word = cpu_to_le16(data);
199         break;
200     case 4:
201         buf.dword = cpu_to_le32(data);
202         break;
203     case 8:
204         buf.qword = cpu_to_le64(data);
205         break;
206     default:
207         hw_error("vfio: unsupported write size, %d bytes", size);
208         break;
209     }
210 
211     if (pwrite(vbasedev->fd, &buf, size, region->fd_offset + addr) != size) {
212         error_report("%s(%s:region%d+0x%"HWADDR_PRIx", 0x%"PRIx64
213                      ",%d) failed: %m",
214                      __func__, vbasedev->name, region->nr,
215                      addr, data, size);
216     }
217 
218     trace_vfio_region_write(vbasedev->name, region->nr, addr, data, size);
219 
220     /*
221      * A read or write to a BAR always signals an INTx EOI.  This will
222      * do nothing if not pending (including not in INTx mode).  We assume
223      * that a BAR access is in response to an interrupt and that BAR
224      * accesses will service the interrupt.  Unfortunately, we don't know
225      * which access will service the interrupt, so we're potentially
226      * getting quite a few host interrupts per guest interrupt.
227      */
228     vbasedev->ops->vfio_eoi(vbasedev);
229 }
230 
231 uint64_t vfio_region_read(void *opaque,
232                           hwaddr addr, unsigned size)
233 {
234     VFIORegion *region = opaque;
235     VFIODevice *vbasedev = region->vbasedev;
236     union {
237         uint8_t byte;
238         uint16_t word;
239         uint32_t dword;
240         uint64_t qword;
241     } buf;
242     uint64_t data = 0;
243 
244     if (pread(vbasedev->fd, &buf, size, region->fd_offset + addr) != size) {
245         error_report("%s(%s:region%d+0x%"HWADDR_PRIx", %d) failed: %m",
246                      __func__, vbasedev->name, region->nr,
247                      addr, size);
248         return (uint64_t)-1;
249     }
250     switch (size) {
251     case 1:
252         data = buf.byte;
253         break;
254     case 2:
255         data = le16_to_cpu(buf.word);
256         break;
257     case 4:
258         data = le32_to_cpu(buf.dword);
259         break;
260     case 8:
261         data = le64_to_cpu(buf.qword);
262         break;
263     default:
264         hw_error("vfio: unsupported read size, %d bytes", size);
265         break;
266     }
267 
268     trace_vfio_region_read(vbasedev->name, region->nr, addr, size, data);
269 
270     /* Same as write above */
271     vbasedev->ops->vfio_eoi(vbasedev);
272 
273     return data;
274 }
275 
276 const MemoryRegionOps vfio_region_ops = {
277     .read = vfio_region_read,
278     .write = vfio_region_write,
279     .endianness = DEVICE_LITTLE_ENDIAN,
280     .valid = {
281         .min_access_size = 1,
282         .max_access_size = 8,
283     },
284     .impl = {
285         .min_access_size = 1,
286         .max_access_size = 8,
287     },
288 };
289 
290 /*
291  * DMA - Mapping and unmapping for the "type1" IOMMU interface used on x86
292  */
293 static int vfio_dma_unmap(VFIOContainer *container,
294                           hwaddr iova, ram_addr_t size)
295 {
296     struct vfio_iommu_type1_dma_unmap unmap = {
297         .argsz = sizeof(unmap),
298         .flags = 0,
299         .iova = iova,
300         .size = size,
301     };
302 
303     while (ioctl(container->fd, VFIO_IOMMU_UNMAP_DMA, &unmap)) {
304         /*
305          * The type1 backend has an off-by-one bug in the kernel (71a7d3d78e3c
306          * v4.15) where an overflow in its wrap-around check prevents us from
307          * unmapping the last page of the address space.  Test for the error
308          * condition and re-try the unmap excluding the last page.  The
309          * expectation is that we've never mapped the last page anyway and this
310          * unmap request comes via vIOMMU support which also makes it unlikely
311          * that this page is used.  This bug was introduced well after type1 v2
312          * support was introduced, so we shouldn't need to test for v1.  A fix
313          * is queued for kernel v5.0 so this workaround can be removed once
314          * affected kernels are sufficiently deprecated.
315          */
316         if (errno == EINVAL && unmap.size && !(unmap.iova + unmap.size) &&
317             container->iommu_type == VFIO_TYPE1v2_IOMMU) {
318             trace_vfio_dma_unmap_overflow_workaround();
319             unmap.size -= 1ULL << ctz64(container->pgsizes);
320             continue;
321         }
322         error_report("VFIO_UNMAP_DMA failed: %s", strerror(errno));
323         return -errno;
324     }
325 
326     return 0;
327 }
328 
329 static int vfio_dma_map(VFIOContainer *container, hwaddr iova,
330                         ram_addr_t size, void *vaddr, bool readonly)
331 {
332     struct vfio_iommu_type1_dma_map map = {
333         .argsz = sizeof(map),
334         .flags = VFIO_DMA_MAP_FLAG_READ,
335         .vaddr = (__u64)(uintptr_t)vaddr,
336         .iova = iova,
337         .size = size,
338     };
339 
340     if (!readonly) {
341         map.flags |= VFIO_DMA_MAP_FLAG_WRITE;
342     }
343 
344     /*
345      * Try the mapping, if it fails with EBUSY, unmap the region and try
346      * again.  This shouldn't be necessary, but we sometimes see it in
347      * the VGA ROM space.
348      */
349     if (ioctl(container->fd, VFIO_IOMMU_MAP_DMA, &map) == 0 ||
350         (errno == EBUSY && vfio_dma_unmap(container, iova, size) == 0 &&
351          ioctl(container->fd, VFIO_IOMMU_MAP_DMA, &map) == 0)) {
352         return 0;
353     }
354 
355     error_report("VFIO_MAP_DMA failed: %s", strerror(errno));
356     return -errno;
357 }
358 
359 static void vfio_host_win_add(VFIOContainer *container,
360                               hwaddr min_iova, hwaddr max_iova,
361                               uint64_t iova_pgsizes)
362 {
363     VFIOHostDMAWindow *hostwin;
364 
365     QLIST_FOREACH(hostwin, &container->hostwin_list, hostwin_next) {
366         if (ranges_overlap(hostwin->min_iova,
367                            hostwin->max_iova - hostwin->min_iova + 1,
368                            min_iova,
369                            max_iova - min_iova + 1)) {
370             hw_error("%s: Overlapped IOMMU are not enabled", __func__);
371         }
372     }
373 
374     hostwin = g_malloc0(sizeof(*hostwin));
375 
376     hostwin->min_iova = min_iova;
377     hostwin->max_iova = max_iova;
378     hostwin->iova_pgsizes = iova_pgsizes;
379     QLIST_INSERT_HEAD(&container->hostwin_list, hostwin, hostwin_next);
380 }
381 
382 static int vfio_host_win_del(VFIOContainer *container, hwaddr min_iova,
383                              hwaddr max_iova)
384 {
385     VFIOHostDMAWindow *hostwin;
386 
387     QLIST_FOREACH(hostwin, &container->hostwin_list, hostwin_next) {
388         if (hostwin->min_iova == min_iova && hostwin->max_iova == max_iova) {
389             QLIST_REMOVE(hostwin, hostwin_next);
390             return 0;
391         }
392     }
393 
394     return -1;
395 }
396 
397 static bool vfio_listener_skipped_section(MemoryRegionSection *section)
398 {
399     return (!memory_region_is_ram(section->mr) &&
400             !memory_region_is_iommu(section->mr)) ||
401            /*
402             * Sizing an enabled 64-bit BAR can cause spurious mappings to
403             * addresses in the upper part of the 64-bit address space.  These
404             * are never accessed by the CPU and beyond the address width of
405             * some IOMMU hardware.  TODO: VFIO should tell us the IOMMU width.
406             */
407            section->offset_within_address_space & (1ULL << 63);
408 }
409 
410 /* Called with rcu_read_lock held.  */
411 static bool vfio_get_vaddr(IOMMUTLBEntry *iotlb, void **vaddr,
412                            bool *read_only)
413 {
414     MemoryRegion *mr;
415     hwaddr xlat;
416     hwaddr len = iotlb->addr_mask + 1;
417     bool writable = iotlb->perm & IOMMU_WO;
418 
419     /*
420      * The IOMMU TLB entry we have just covers translation through
421      * this IOMMU to its immediate target.  We need to translate
422      * it the rest of the way through to memory.
423      */
424     mr = address_space_translate(&address_space_memory,
425                                  iotlb->translated_addr,
426                                  &xlat, &len, writable,
427                                  MEMTXATTRS_UNSPECIFIED);
428     if (!memory_region_is_ram(mr)) {
429         error_report("iommu map to non memory area %"HWADDR_PRIx"",
430                      xlat);
431         return false;
432     }
433 
434     /*
435      * Translation truncates length to the IOMMU page size,
436      * check that it did not truncate too much.
437      */
438     if (len & iotlb->addr_mask) {
439         error_report("iommu has granularity incompatible with target AS");
440         return false;
441     }
442 
443     *vaddr = memory_region_get_ram_ptr(mr) + xlat;
444     *read_only = !writable || mr->readonly;
445 
446     return true;
447 }
448 
449 static void vfio_iommu_map_notify(IOMMUNotifier *n, IOMMUTLBEntry *iotlb)
450 {
451     VFIOGuestIOMMU *giommu = container_of(n, VFIOGuestIOMMU, n);
452     VFIOContainer *container = giommu->container;
453     hwaddr iova = iotlb->iova + giommu->iommu_offset;
454     bool read_only;
455     void *vaddr;
456     int ret;
457 
458     trace_vfio_iommu_map_notify(iotlb->perm == IOMMU_NONE ? "UNMAP" : "MAP",
459                                 iova, iova + iotlb->addr_mask);
460 
461     if (iotlb->target_as != &address_space_memory) {
462         error_report("Wrong target AS \"%s\", only system memory is allowed",
463                      iotlb->target_as->name ? iotlb->target_as->name : "none");
464         return;
465     }
466 
467     rcu_read_lock();
468 
469     if ((iotlb->perm & IOMMU_RW) != IOMMU_NONE) {
470         if (!vfio_get_vaddr(iotlb, &vaddr, &read_only)) {
471             goto out;
472         }
473         /*
474          * vaddr is only valid until rcu_read_unlock(). But after
475          * vfio_dma_map has set up the mapping the pages will be
476          * pinned by the kernel. This makes sure that the RAM backend
477          * of vaddr will always be there, even if the memory object is
478          * destroyed and its backing memory munmap-ed.
479          */
480         ret = vfio_dma_map(container, iova,
481                            iotlb->addr_mask + 1, vaddr,
482                            read_only);
483         if (ret) {
484             error_report("vfio_dma_map(%p, 0x%"HWADDR_PRIx", "
485                          "0x%"HWADDR_PRIx", %p) = %d (%m)",
486                          container, iova,
487                          iotlb->addr_mask + 1, vaddr, ret);
488         }
489     } else {
490         ret = vfio_dma_unmap(container, iova, iotlb->addr_mask + 1);
491         if (ret) {
492             error_report("vfio_dma_unmap(%p, 0x%"HWADDR_PRIx", "
493                          "0x%"HWADDR_PRIx") = %d (%m)",
494                          container, iova,
495                          iotlb->addr_mask + 1, ret);
496         }
497     }
498 out:
499     rcu_read_unlock();
500 }
501 
502 static void vfio_listener_region_add(MemoryListener *listener,
503                                      MemoryRegionSection *section)
504 {
505     VFIOContainer *container = container_of(listener, VFIOContainer, listener);
506     hwaddr iova, end;
507     Int128 llend, llsize;
508     void *vaddr;
509     int ret;
510     VFIOHostDMAWindow *hostwin;
511     bool hostwin_found;
512     Error *err = NULL;
513 
514     if (vfio_listener_skipped_section(section)) {
515         trace_vfio_listener_region_add_skip(
516                 section->offset_within_address_space,
517                 section->offset_within_address_space +
518                 int128_get64(int128_sub(section->size, int128_one())));
519         return;
520     }
521 
522     if (unlikely((section->offset_within_address_space & ~TARGET_PAGE_MASK) !=
523                  (section->offset_within_region & ~TARGET_PAGE_MASK))) {
524         error_report("%s received unaligned region", __func__);
525         return;
526     }
527 
528     iova = TARGET_PAGE_ALIGN(section->offset_within_address_space);
529     llend = int128_make64(section->offset_within_address_space);
530     llend = int128_add(llend, section->size);
531     llend = int128_and(llend, int128_exts64(TARGET_PAGE_MASK));
532 
533     if (int128_ge(int128_make64(iova), llend)) {
534         return;
535     }
536     end = int128_get64(int128_sub(llend, int128_one()));
537 
538     if (container->iommu_type == VFIO_SPAPR_TCE_v2_IOMMU) {
539         hwaddr pgsize = 0;
540 
541         /* For now intersections are not allowed, we may relax this later */
542         QLIST_FOREACH(hostwin, &container->hostwin_list, hostwin_next) {
543             if (ranges_overlap(hostwin->min_iova,
544                                hostwin->max_iova - hostwin->min_iova + 1,
545                                section->offset_within_address_space,
546                                int128_get64(section->size))) {
547                 error_setg(&err,
548                     "region [0x%"PRIx64",0x%"PRIx64"] overlaps with existing"
549                     "host DMA window [0x%"PRIx64",0x%"PRIx64"]",
550                     section->offset_within_address_space,
551                     section->offset_within_address_space +
552                         int128_get64(section->size) - 1,
553                     hostwin->min_iova, hostwin->max_iova);
554                 goto fail;
555             }
556         }
557 
558         ret = vfio_spapr_create_window(container, section, &pgsize);
559         if (ret) {
560             error_setg_errno(&err, -ret, "Failed to create SPAPR window");
561             goto fail;
562         }
563 
564         vfio_host_win_add(container, section->offset_within_address_space,
565                           section->offset_within_address_space +
566                           int128_get64(section->size) - 1, pgsize);
567 #ifdef CONFIG_KVM
568         if (kvm_enabled()) {
569             VFIOGroup *group;
570             IOMMUMemoryRegion *iommu_mr = IOMMU_MEMORY_REGION(section->mr);
571             struct kvm_vfio_spapr_tce param;
572             struct kvm_device_attr attr = {
573                 .group = KVM_DEV_VFIO_GROUP,
574                 .attr = KVM_DEV_VFIO_GROUP_SET_SPAPR_TCE,
575                 .addr = (uint64_t)(unsigned long)&param,
576             };
577 
578             if (!memory_region_iommu_get_attr(iommu_mr, IOMMU_ATTR_SPAPR_TCE_FD,
579                                               &param.tablefd)) {
580                 QLIST_FOREACH(group, &container->group_list, container_next) {
581                     param.groupfd = group->fd;
582                     if (ioctl(vfio_kvm_device_fd, KVM_SET_DEVICE_ATTR, &attr)) {
583                         error_report("vfio: failed to setup fd %d "
584                                      "for a group with fd %d: %s",
585                                      param.tablefd, param.groupfd,
586                                      strerror(errno));
587                         return;
588                     }
589                     trace_vfio_spapr_group_attach(param.groupfd, param.tablefd);
590                 }
591             }
592         }
593 #endif
594     }
595 
596     hostwin_found = false;
597     QLIST_FOREACH(hostwin, &container->hostwin_list, hostwin_next) {
598         if (hostwin->min_iova <= iova && end <= hostwin->max_iova) {
599             hostwin_found = true;
600             break;
601         }
602     }
603 
604     if (!hostwin_found) {
605         error_setg(&err, "Container %p can't map guest IOVA region"
606                    " 0x%"HWADDR_PRIx"..0x%"HWADDR_PRIx, container, iova, end);
607         goto fail;
608     }
609 
610     memory_region_ref(section->mr);
611 
612     if (memory_region_is_iommu(section->mr)) {
613         VFIOGuestIOMMU *giommu;
614         IOMMUMemoryRegion *iommu_mr = IOMMU_MEMORY_REGION(section->mr);
615         int iommu_idx;
616 
617         trace_vfio_listener_region_add_iommu(iova, end);
618         /*
619          * FIXME: For VFIO iommu types which have KVM acceleration to
620          * avoid bouncing all map/unmaps through qemu this way, this
621          * would be the right place to wire that up (tell the KVM
622          * device emulation the VFIO iommu handles to use).
623          */
624         giommu = g_malloc0(sizeof(*giommu));
625         giommu->iommu = iommu_mr;
626         giommu->iommu_offset = section->offset_within_address_space -
627                                section->offset_within_region;
628         giommu->container = container;
629         llend = int128_add(int128_make64(section->offset_within_region),
630                            section->size);
631         llend = int128_sub(llend, int128_one());
632         iommu_idx = memory_region_iommu_attrs_to_index(iommu_mr,
633                                                        MEMTXATTRS_UNSPECIFIED);
634         iommu_notifier_init(&giommu->n, vfio_iommu_map_notify,
635                             IOMMU_NOTIFIER_ALL,
636                             section->offset_within_region,
637                             int128_get64(llend),
638                             iommu_idx);
639 
640         ret = memory_region_register_iommu_notifier(section->mr, &giommu->n,
641                                                     &err);
642         if (ret) {
643             g_free(giommu);
644             goto fail;
645         }
646         QLIST_INSERT_HEAD(&container->giommu_list, giommu, giommu_next);
647         memory_region_iommu_replay(giommu->iommu, &giommu->n);
648 
649         return;
650     }
651 
652     /* Here we assume that memory_region_is_ram(section->mr)==true */
653 
654     vaddr = memory_region_get_ram_ptr(section->mr) +
655             section->offset_within_region +
656             (iova - section->offset_within_address_space);
657 
658     trace_vfio_listener_region_add_ram(iova, end, vaddr);
659 
660     llsize = int128_sub(llend, int128_make64(iova));
661 
662     if (memory_region_is_ram_device(section->mr)) {
663         hwaddr pgmask = (1ULL << ctz64(hostwin->iova_pgsizes)) - 1;
664 
665         if ((iova & pgmask) || (int128_get64(llsize) & pgmask)) {
666             trace_vfio_listener_region_add_no_dma_map(
667                 memory_region_name(section->mr),
668                 section->offset_within_address_space,
669                 int128_getlo(section->size),
670                 pgmask + 1);
671             return;
672         }
673     }
674 
675     ret = vfio_dma_map(container, iova, int128_get64(llsize),
676                        vaddr, section->readonly);
677     if (ret) {
678         error_setg(&err, "vfio_dma_map(%p, 0x%"HWADDR_PRIx", "
679                    "0x%"HWADDR_PRIx", %p) = %d (%m)",
680                    container, iova, int128_get64(llsize), vaddr, ret);
681         if (memory_region_is_ram_device(section->mr)) {
682             /* Allow unexpected mappings not to be fatal for RAM devices */
683             error_report_err(err);
684             return;
685         }
686         goto fail;
687     }
688 
689     return;
690 
691 fail:
692     if (memory_region_is_ram_device(section->mr)) {
693         error_report("failed to vfio_dma_map. pci p2p may not work");
694         return;
695     }
696     /*
697      * On the initfn path, store the first error in the container so we
698      * can gracefully fail.  Runtime, there's not much we can do other
699      * than throw a hardware error.
700      */
701     if (!container->initialized) {
702         if (!container->error) {
703             error_propagate_prepend(&container->error, err,
704                                     "Region %s: ",
705                                     memory_region_name(section->mr));
706         } else {
707             error_free(err);
708         }
709     } else {
710         error_report_err(err);
711         hw_error("vfio: DMA mapping failed, unable to continue");
712     }
713 }
714 
715 static void vfio_listener_region_del(MemoryListener *listener,
716                                      MemoryRegionSection *section)
717 {
718     VFIOContainer *container = container_of(listener, VFIOContainer, listener);
719     hwaddr iova, end;
720     Int128 llend, llsize;
721     int ret;
722     bool try_unmap = true;
723 
724     if (vfio_listener_skipped_section(section)) {
725         trace_vfio_listener_region_del_skip(
726                 section->offset_within_address_space,
727                 section->offset_within_address_space +
728                 int128_get64(int128_sub(section->size, int128_one())));
729         return;
730     }
731 
732     if (unlikely((section->offset_within_address_space & ~TARGET_PAGE_MASK) !=
733                  (section->offset_within_region & ~TARGET_PAGE_MASK))) {
734         error_report("%s received unaligned region", __func__);
735         return;
736     }
737 
738     if (memory_region_is_iommu(section->mr)) {
739         VFIOGuestIOMMU *giommu;
740 
741         QLIST_FOREACH(giommu, &container->giommu_list, giommu_next) {
742             if (MEMORY_REGION(giommu->iommu) == section->mr &&
743                 giommu->n.start == section->offset_within_region) {
744                 memory_region_unregister_iommu_notifier(section->mr,
745                                                         &giommu->n);
746                 QLIST_REMOVE(giommu, giommu_next);
747                 g_free(giommu);
748                 break;
749             }
750         }
751 
752         /*
753          * FIXME: We assume the one big unmap below is adequate to
754          * remove any individual page mappings in the IOMMU which
755          * might have been copied into VFIO. This works for a page table
756          * based IOMMU where a big unmap flattens a large range of IO-PTEs.
757          * That may not be true for all IOMMU types.
758          */
759     }
760 
761     iova = TARGET_PAGE_ALIGN(section->offset_within_address_space);
762     llend = int128_make64(section->offset_within_address_space);
763     llend = int128_add(llend, section->size);
764     llend = int128_and(llend, int128_exts64(TARGET_PAGE_MASK));
765 
766     if (int128_ge(int128_make64(iova), llend)) {
767         return;
768     }
769     end = int128_get64(int128_sub(llend, int128_one()));
770 
771     llsize = int128_sub(llend, int128_make64(iova));
772 
773     trace_vfio_listener_region_del(iova, end);
774 
775     if (memory_region_is_ram_device(section->mr)) {
776         hwaddr pgmask;
777         VFIOHostDMAWindow *hostwin;
778         bool hostwin_found = false;
779 
780         QLIST_FOREACH(hostwin, &container->hostwin_list, hostwin_next) {
781             if (hostwin->min_iova <= iova && end <= hostwin->max_iova) {
782                 hostwin_found = true;
783                 break;
784             }
785         }
786         assert(hostwin_found); /* or region_add() would have failed */
787 
788         pgmask = (1ULL << ctz64(hostwin->iova_pgsizes)) - 1;
789         try_unmap = !((iova & pgmask) || (int128_get64(llsize) & pgmask));
790     }
791 
792     if (try_unmap) {
793         ret = vfio_dma_unmap(container, iova, int128_get64(llsize));
794         if (ret) {
795             error_report("vfio_dma_unmap(%p, 0x%"HWADDR_PRIx", "
796                          "0x%"HWADDR_PRIx") = %d (%m)",
797                          container, iova, int128_get64(llsize), ret);
798         }
799     }
800 
801     memory_region_unref(section->mr);
802 
803     if (container->iommu_type == VFIO_SPAPR_TCE_v2_IOMMU) {
804         vfio_spapr_remove_window(container,
805                                  section->offset_within_address_space);
806         if (vfio_host_win_del(container,
807                               section->offset_within_address_space,
808                               section->offset_within_address_space +
809                               int128_get64(section->size) - 1) < 0) {
810             hw_error("%s: Cannot delete missing window at %"HWADDR_PRIx,
811                      __func__, section->offset_within_address_space);
812         }
813     }
814 }
815 
816 static const MemoryListener vfio_memory_listener = {
817     .region_add = vfio_listener_region_add,
818     .region_del = vfio_listener_region_del,
819 };
820 
821 static void vfio_listener_release(VFIOContainer *container)
822 {
823     memory_listener_unregister(&container->listener);
824     if (container->iommu_type == VFIO_SPAPR_TCE_v2_IOMMU) {
825         memory_listener_unregister(&container->prereg_listener);
826     }
827 }
828 
829 struct vfio_info_cap_header *
830 vfio_get_region_info_cap(struct vfio_region_info *info, uint16_t id)
831 {
832     struct vfio_info_cap_header *hdr;
833     void *ptr = info;
834 
835     if (!(info->flags & VFIO_REGION_INFO_FLAG_CAPS)) {
836         return NULL;
837     }
838 
839     for (hdr = ptr + info->cap_offset; hdr != ptr; hdr = ptr + hdr->next) {
840         if (hdr->id == id) {
841             return hdr;
842         }
843     }
844 
845     return NULL;
846 }
847 
848 static int vfio_setup_region_sparse_mmaps(VFIORegion *region,
849                                           struct vfio_region_info *info)
850 {
851     struct vfio_info_cap_header *hdr;
852     struct vfio_region_info_cap_sparse_mmap *sparse;
853     int i, j;
854 
855     hdr = vfio_get_region_info_cap(info, VFIO_REGION_INFO_CAP_SPARSE_MMAP);
856     if (!hdr) {
857         return -ENODEV;
858     }
859 
860     sparse = container_of(hdr, struct vfio_region_info_cap_sparse_mmap, header);
861 
862     trace_vfio_region_sparse_mmap_header(region->vbasedev->name,
863                                          region->nr, sparse->nr_areas);
864 
865     region->mmaps = g_new0(VFIOMmap, sparse->nr_areas);
866 
867     for (i = 0, j = 0; i < sparse->nr_areas; i++) {
868         trace_vfio_region_sparse_mmap_entry(i, sparse->areas[i].offset,
869                                             sparse->areas[i].offset +
870                                             sparse->areas[i].size);
871 
872         if (sparse->areas[i].size) {
873             region->mmaps[j].offset = sparse->areas[i].offset;
874             region->mmaps[j].size = sparse->areas[i].size;
875             j++;
876         }
877     }
878 
879     region->nr_mmaps = j;
880     region->mmaps = g_realloc(region->mmaps, j * sizeof(VFIOMmap));
881 
882     return 0;
883 }
884 
885 int vfio_region_setup(Object *obj, VFIODevice *vbasedev, VFIORegion *region,
886                       int index, const char *name)
887 {
888     struct vfio_region_info *info;
889     int ret;
890 
891     ret = vfio_get_region_info(vbasedev, index, &info);
892     if (ret) {
893         return ret;
894     }
895 
896     region->vbasedev = vbasedev;
897     region->flags = info->flags;
898     region->size = info->size;
899     region->fd_offset = info->offset;
900     region->nr = index;
901 
902     if (region->size) {
903         region->mem = g_new0(MemoryRegion, 1);
904         memory_region_init_io(region->mem, obj, &vfio_region_ops,
905                               region, name, region->size);
906 
907         if (!vbasedev->no_mmap &&
908             region->flags & VFIO_REGION_INFO_FLAG_MMAP) {
909 
910             ret = vfio_setup_region_sparse_mmaps(region, info);
911 
912             if (ret) {
913                 region->nr_mmaps = 1;
914                 region->mmaps = g_new0(VFIOMmap, region->nr_mmaps);
915                 region->mmaps[0].offset = 0;
916                 region->mmaps[0].size = region->size;
917             }
918         }
919     }
920 
921     g_free(info);
922 
923     trace_vfio_region_setup(vbasedev->name, index, name,
924                             region->flags, region->fd_offset, region->size);
925     return 0;
926 }
927 
928 int vfio_region_mmap(VFIORegion *region)
929 {
930     int i, prot = 0;
931     char *name;
932 
933     if (!region->mem) {
934         return 0;
935     }
936 
937     prot |= region->flags & VFIO_REGION_INFO_FLAG_READ ? PROT_READ : 0;
938     prot |= region->flags & VFIO_REGION_INFO_FLAG_WRITE ? PROT_WRITE : 0;
939 
940     for (i = 0; i < region->nr_mmaps; i++) {
941         region->mmaps[i].mmap = mmap(NULL, region->mmaps[i].size, prot,
942                                      MAP_SHARED, region->vbasedev->fd,
943                                      region->fd_offset +
944                                      region->mmaps[i].offset);
945         if (region->mmaps[i].mmap == MAP_FAILED) {
946             int ret = -errno;
947 
948             trace_vfio_region_mmap_fault(memory_region_name(region->mem), i,
949                                          region->fd_offset +
950                                          region->mmaps[i].offset,
951                                          region->fd_offset +
952                                          region->mmaps[i].offset +
953                                          region->mmaps[i].size - 1, ret);
954 
955             region->mmaps[i].mmap = NULL;
956 
957             for (i--; i >= 0; i--) {
958                 memory_region_del_subregion(region->mem, &region->mmaps[i].mem);
959                 munmap(region->mmaps[i].mmap, region->mmaps[i].size);
960                 object_unparent(OBJECT(&region->mmaps[i].mem));
961                 region->mmaps[i].mmap = NULL;
962             }
963 
964             return ret;
965         }
966 
967         name = g_strdup_printf("%s mmaps[%d]",
968                                memory_region_name(region->mem), i);
969         memory_region_init_ram_device_ptr(&region->mmaps[i].mem,
970                                           memory_region_owner(region->mem),
971                                           name, region->mmaps[i].size,
972                                           region->mmaps[i].mmap);
973         g_free(name);
974         memory_region_add_subregion(region->mem, region->mmaps[i].offset,
975                                     &region->mmaps[i].mem);
976 
977         trace_vfio_region_mmap(memory_region_name(&region->mmaps[i].mem),
978                                region->mmaps[i].offset,
979                                region->mmaps[i].offset +
980                                region->mmaps[i].size - 1);
981     }
982 
983     return 0;
984 }
985 
986 void vfio_region_exit(VFIORegion *region)
987 {
988     int i;
989 
990     if (!region->mem) {
991         return;
992     }
993 
994     for (i = 0; i < region->nr_mmaps; i++) {
995         if (region->mmaps[i].mmap) {
996             memory_region_del_subregion(region->mem, &region->mmaps[i].mem);
997         }
998     }
999 
1000     trace_vfio_region_exit(region->vbasedev->name, region->nr);
1001 }
1002 
1003 void vfio_region_finalize(VFIORegion *region)
1004 {
1005     int i;
1006 
1007     if (!region->mem) {
1008         return;
1009     }
1010 
1011     for (i = 0; i < region->nr_mmaps; i++) {
1012         if (region->mmaps[i].mmap) {
1013             munmap(region->mmaps[i].mmap, region->mmaps[i].size);
1014             object_unparent(OBJECT(&region->mmaps[i].mem));
1015         }
1016     }
1017 
1018     object_unparent(OBJECT(region->mem));
1019 
1020     g_free(region->mem);
1021     g_free(region->mmaps);
1022 
1023     trace_vfio_region_finalize(region->vbasedev->name, region->nr);
1024 
1025     region->mem = NULL;
1026     region->mmaps = NULL;
1027     region->nr_mmaps = 0;
1028     region->size = 0;
1029     region->flags = 0;
1030     region->nr = 0;
1031 }
1032 
1033 void vfio_region_mmaps_set_enabled(VFIORegion *region, bool enabled)
1034 {
1035     int i;
1036 
1037     if (!region->mem) {
1038         return;
1039     }
1040 
1041     for (i = 0; i < region->nr_mmaps; i++) {
1042         if (region->mmaps[i].mmap) {
1043             memory_region_set_enabled(&region->mmaps[i].mem, enabled);
1044         }
1045     }
1046 
1047     trace_vfio_region_mmaps_set_enabled(memory_region_name(region->mem),
1048                                         enabled);
1049 }
1050 
1051 void vfio_reset_handler(void *opaque)
1052 {
1053     VFIOGroup *group;
1054     VFIODevice *vbasedev;
1055 
1056     QLIST_FOREACH(group, &vfio_group_list, next) {
1057         QLIST_FOREACH(vbasedev, &group->device_list, next) {
1058             if (vbasedev->dev->realized) {
1059                 vbasedev->ops->vfio_compute_needs_reset(vbasedev);
1060             }
1061         }
1062     }
1063 
1064     QLIST_FOREACH(group, &vfio_group_list, next) {
1065         QLIST_FOREACH(vbasedev, &group->device_list, next) {
1066             if (vbasedev->dev->realized && vbasedev->needs_reset) {
1067                 vbasedev->ops->vfio_hot_reset_multi(vbasedev);
1068             }
1069         }
1070     }
1071 }
1072 
1073 static void vfio_kvm_device_add_group(VFIOGroup *group)
1074 {
1075 #ifdef CONFIG_KVM
1076     struct kvm_device_attr attr = {
1077         .group = KVM_DEV_VFIO_GROUP,
1078         .attr = KVM_DEV_VFIO_GROUP_ADD,
1079         .addr = (uint64_t)(unsigned long)&group->fd,
1080     };
1081 
1082     if (!kvm_enabled()) {
1083         return;
1084     }
1085 
1086     if (vfio_kvm_device_fd < 0) {
1087         struct kvm_create_device cd = {
1088             .type = KVM_DEV_TYPE_VFIO,
1089         };
1090 
1091         if (kvm_vm_ioctl(kvm_state, KVM_CREATE_DEVICE, &cd)) {
1092             error_report("Failed to create KVM VFIO device: %m");
1093             return;
1094         }
1095 
1096         vfio_kvm_device_fd = cd.fd;
1097     }
1098 
1099     if (ioctl(vfio_kvm_device_fd, KVM_SET_DEVICE_ATTR, &attr)) {
1100         error_report("Failed to add group %d to KVM VFIO device: %m",
1101                      group->groupid);
1102     }
1103 #endif
1104 }
1105 
1106 static void vfio_kvm_device_del_group(VFIOGroup *group)
1107 {
1108 #ifdef CONFIG_KVM
1109     struct kvm_device_attr attr = {
1110         .group = KVM_DEV_VFIO_GROUP,
1111         .attr = KVM_DEV_VFIO_GROUP_DEL,
1112         .addr = (uint64_t)(unsigned long)&group->fd,
1113     };
1114 
1115     if (vfio_kvm_device_fd < 0) {
1116         return;
1117     }
1118 
1119     if (ioctl(vfio_kvm_device_fd, KVM_SET_DEVICE_ATTR, &attr)) {
1120         error_report("Failed to remove group %d from KVM VFIO device: %m",
1121                      group->groupid);
1122     }
1123 #endif
1124 }
1125 
1126 static VFIOAddressSpace *vfio_get_address_space(AddressSpace *as)
1127 {
1128     VFIOAddressSpace *space;
1129 
1130     QLIST_FOREACH(space, &vfio_address_spaces, list) {
1131         if (space->as == as) {
1132             return space;
1133         }
1134     }
1135 
1136     /* No suitable VFIOAddressSpace, create a new one */
1137     space = g_malloc0(sizeof(*space));
1138     space->as = as;
1139     QLIST_INIT(&space->containers);
1140 
1141     QLIST_INSERT_HEAD(&vfio_address_spaces, space, list);
1142 
1143     return space;
1144 }
1145 
1146 static void vfio_put_address_space(VFIOAddressSpace *space)
1147 {
1148     if (QLIST_EMPTY(&space->containers)) {
1149         QLIST_REMOVE(space, list);
1150         g_free(space);
1151     }
1152 }
1153 
1154 /*
1155  * vfio_get_iommu_type - selects the richest iommu_type (v2 first)
1156  */
1157 static int vfio_get_iommu_type(VFIOContainer *container,
1158                                Error **errp)
1159 {
1160     int iommu_types[] = { VFIO_TYPE1v2_IOMMU, VFIO_TYPE1_IOMMU,
1161                           VFIO_SPAPR_TCE_v2_IOMMU, VFIO_SPAPR_TCE_IOMMU };
1162     int i;
1163 
1164     for (i = 0; i < ARRAY_SIZE(iommu_types); i++) {
1165         if (ioctl(container->fd, VFIO_CHECK_EXTENSION, iommu_types[i])) {
1166             return iommu_types[i];
1167         }
1168     }
1169     error_setg(errp, "No available IOMMU models");
1170     return -EINVAL;
1171 }
1172 
1173 static int vfio_init_container(VFIOContainer *container, int group_fd,
1174                                Error **errp)
1175 {
1176     int iommu_type, ret;
1177 
1178     iommu_type = vfio_get_iommu_type(container, errp);
1179     if (iommu_type < 0) {
1180         return iommu_type;
1181     }
1182 
1183     ret = ioctl(group_fd, VFIO_GROUP_SET_CONTAINER, &container->fd);
1184     if (ret) {
1185         error_setg_errno(errp, errno, "Failed to set group container");
1186         return -errno;
1187     }
1188 
1189     while (ioctl(container->fd, VFIO_SET_IOMMU, iommu_type)) {
1190         if (iommu_type == VFIO_SPAPR_TCE_v2_IOMMU) {
1191             /*
1192              * On sPAPR, despite the IOMMU subdriver always advertises v1 and
1193              * v2, the running platform may not support v2 and there is no
1194              * way to guess it until an IOMMU group gets added to the container.
1195              * So in case it fails with v2, try v1 as a fallback.
1196              */
1197             iommu_type = VFIO_SPAPR_TCE_IOMMU;
1198             continue;
1199         }
1200         error_setg_errno(errp, errno, "Failed to set iommu for container");
1201         return -errno;
1202     }
1203 
1204     container->iommu_type = iommu_type;
1205     return 0;
1206 }
1207 
1208 static int vfio_connect_container(VFIOGroup *group, AddressSpace *as,
1209                                   Error **errp)
1210 {
1211     VFIOContainer *container;
1212     int ret, fd;
1213     VFIOAddressSpace *space;
1214 
1215     space = vfio_get_address_space(as);
1216 
1217     /*
1218      * VFIO is currently incompatible with memory ballooning insofar as the
1219      * madvise to purge (zap) the page from QEMU's address space does not
1220      * interact with the memory API and therefore leaves stale virtual to
1221      * physical mappings in the IOMMU if the page was previously pinned.  We
1222      * therefore add a balloon inhibit for each group added to a container,
1223      * whether the container is used individually or shared.  This provides
1224      * us with options to allow devices within a group to opt-in and allow
1225      * ballooning, so long as it is done consistently for a group (for instance
1226      * if the device is an mdev device where it is known that the host vendor
1227      * driver will never pin pages outside of the working set of the guest
1228      * driver, which would thus not be ballooning candidates).
1229      *
1230      * The first opportunity to induce pinning occurs here where we attempt to
1231      * attach the group to existing containers within the AddressSpace.  If any
1232      * pages are already zapped from the virtual address space, such as from a
1233      * previous ballooning opt-in, new pinning will cause valid mappings to be
1234      * re-established.  Likewise, when the overall MemoryListener for a new
1235      * container is registered, a replay of mappings within the AddressSpace
1236      * will occur, re-establishing any previously zapped pages as well.
1237      *
1238      * NB. Balloon inhibiting does not currently block operation of the
1239      * balloon driver or revoke previously pinned pages, it only prevents
1240      * calling madvise to modify the virtual mapping of ballooned pages.
1241      */
1242     qemu_balloon_inhibit(true);
1243 
1244     QLIST_FOREACH(container, &space->containers, next) {
1245         if (!ioctl(group->fd, VFIO_GROUP_SET_CONTAINER, &container->fd)) {
1246             group->container = container;
1247             QLIST_INSERT_HEAD(&container->group_list, group, container_next);
1248             vfio_kvm_device_add_group(group);
1249             return 0;
1250         }
1251     }
1252 
1253     fd = qemu_open("/dev/vfio/vfio", O_RDWR);
1254     if (fd < 0) {
1255         error_setg_errno(errp, errno, "failed to open /dev/vfio/vfio");
1256         ret = -errno;
1257         goto put_space_exit;
1258     }
1259 
1260     ret = ioctl(fd, VFIO_GET_API_VERSION);
1261     if (ret != VFIO_API_VERSION) {
1262         error_setg(errp, "supported vfio version: %d, "
1263                    "reported version: %d", VFIO_API_VERSION, ret);
1264         ret = -EINVAL;
1265         goto close_fd_exit;
1266     }
1267 
1268     container = g_malloc0(sizeof(*container));
1269     container->space = space;
1270     container->fd = fd;
1271     container->error = NULL;
1272     QLIST_INIT(&container->giommu_list);
1273     QLIST_INIT(&container->hostwin_list);
1274 
1275     ret = vfio_init_container(container, group->fd, errp);
1276     if (ret) {
1277         goto free_container_exit;
1278     }
1279 
1280     switch (container->iommu_type) {
1281     case VFIO_TYPE1v2_IOMMU:
1282     case VFIO_TYPE1_IOMMU:
1283     {
1284         struct vfio_iommu_type1_info info;
1285 
1286         /*
1287          * FIXME: This assumes that a Type1 IOMMU can map any 64-bit
1288          * IOVA whatsoever.  That's not actually true, but the current
1289          * kernel interface doesn't tell us what it can map, and the
1290          * existing Type1 IOMMUs generally support any IOVA we're
1291          * going to actually try in practice.
1292          */
1293         info.argsz = sizeof(info);
1294         ret = ioctl(fd, VFIO_IOMMU_GET_INFO, &info);
1295         /* Ignore errors */
1296         if (ret || !(info.flags & VFIO_IOMMU_INFO_PGSIZES)) {
1297             /* Assume 4k IOVA page size */
1298             info.iova_pgsizes = 4096;
1299         }
1300         vfio_host_win_add(container, 0, (hwaddr)-1, info.iova_pgsizes);
1301         container->pgsizes = info.iova_pgsizes;
1302         break;
1303     }
1304     case VFIO_SPAPR_TCE_v2_IOMMU:
1305     case VFIO_SPAPR_TCE_IOMMU:
1306     {
1307         struct vfio_iommu_spapr_tce_info info;
1308         bool v2 = container->iommu_type == VFIO_SPAPR_TCE_v2_IOMMU;
1309 
1310         /*
1311          * The host kernel code implementing VFIO_IOMMU_DISABLE is called
1312          * when container fd is closed so we do not call it explicitly
1313          * in this file.
1314          */
1315         if (!v2) {
1316             ret = ioctl(fd, VFIO_IOMMU_ENABLE);
1317             if (ret) {
1318                 error_setg_errno(errp, errno, "failed to enable container");
1319                 ret = -errno;
1320                 goto free_container_exit;
1321             }
1322         } else {
1323             container->prereg_listener = vfio_prereg_listener;
1324 
1325             memory_listener_register(&container->prereg_listener,
1326                                      &address_space_memory);
1327             if (container->error) {
1328                 memory_listener_unregister(&container->prereg_listener);
1329                 ret = -1;
1330                 error_propagate_prepend(errp, container->error,
1331                     "RAM memory listener initialization failed: ");
1332                 goto free_container_exit;
1333             }
1334         }
1335 
1336         info.argsz = sizeof(info);
1337         ret = ioctl(fd, VFIO_IOMMU_SPAPR_TCE_GET_INFO, &info);
1338         if (ret) {
1339             error_setg_errno(errp, errno,
1340                              "VFIO_IOMMU_SPAPR_TCE_GET_INFO failed");
1341             ret = -errno;
1342             if (v2) {
1343                 memory_listener_unregister(&container->prereg_listener);
1344             }
1345             goto free_container_exit;
1346         }
1347 
1348         if (v2) {
1349             container->pgsizes = info.ddw.pgsizes;
1350             /*
1351              * There is a default window in just created container.
1352              * To make region_add/del simpler, we better remove this
1353              * window now and let those iommu_listener callbacks
1354              * create/remove them when needed.
1355              */
1356             ret = vfio_spapr_remove_window(container, info.dma32_window_start);
1357             if (ret) {
1358                 error_setg_errno(errp, -ret,
1359                                  "failed to remove existing window");
1360                 goto free_container_exit;
1361             }
1362         } else {
1363             /* The default table uses 4K pages */
1364             container->pgsizes = 0x1000;
1365             vfio_host_win_add(container, info.dma32_window_start,
1366                               info.dma32_window_start +
1367                               info.dma32_window_size - 1,
1368                               0x1000);
1369         }
1370     }
1371     }
1372 
1373     vfio_kvm_device_add_group(group);
1374 
1375     QLIST_INIT(&container->group_list);
1376     QLIST_INSERT_HEAD(&space->containers, container, next);
1377 
1378     group->container = container;
1379     QLIST_INSERT_HEAD(&container->group_list, group, container_next);
1380 
1381     container->listener = vfio_memory_listener;
1382 
1383     memory_listener_register(&container->listener, container->space->as);
1384 
1385     if (container->error) {
1386         ret = -1;
1387         error_propagate_prepend(errp, container->error,
1388             "memory listener initialization failed: ");
1389         goto listener_release_exit;
1390     }
1391 
1392     container->initialized = true;
1393 
1394     return 0;
1395 listener_release_exit:
1396     QLIST_REMOVE(group, container_next);
1397     QLIST_REMOVE(container, next);
1398     vfio_kvm_device_del_group(group);
1399     vfio_listener_release(container);
1400 
1401 free_container_exit:
1402     g_free(container);
1403 
1404 close_fd_exit:
1405     close(fd);
1406 
1407 put_space_exit:
1408     qemu_balloon_inhibit(false);
1409     vfio_put_address_space(space);
1410 
1411     return ret;
1412 }
1413 
1414 static void vfio_disconnect_container(VFIOGroup *group)
1415 {
1416     VFIOContainer *container = group->container;
1417 
1418     QLIST_REMOVE(group, container_next);
1419     group->container = NULL;
1420 
1421     /*
1422      * Explicitly release the listener first before unset container,
1423      * since unset may destroy the backend container if it's the last
1424      * group.
1425      */
1426     if (QLIST_EMPTY(&container->group_list)) {
1427         vfio_listener_release(container);
1428     }
1429 
1430     if (ioctl(group->fd, VFIO_GROUP_UNSET_CONTAINER, &container->fd)) {
1431         error_report("vfio: error disconnecting group %d from container",
1432                      group->groupid);
1433     }
1434 
1435     if (QLIST_EMPTY(&container->group_list)) {
1436         VFIOAddressSpace *space = container->space;
1437         VFIOGuestIOMMU *giommu, *tmp;
1438 
1439         QLIST_REMOVE(container, next);
1440 
1441         QLIST_FOREACH_SAFE(giommu, &container->giommu_list, giommu_next, tmp) {
1442             memory_region_unregister_iommu_notifier(
1443                     MEMORY_REGION(giommu->iommu), &giommu->n);
1444             QLIST_REMOVE(giommu, giommu_next);
1445             g_free(giommu);
1446         }
1447 
1448         trace_vfio_disconnect_container(container->fd);
1449         close(container->fd);
1450         g_free(container);
1451 
1452         vfio_put_address_space(space);
1453     }
1454 }
1455 
1456 VFIOGroup *vfio_get_group(int groupid, AddressSpace *as, Error **errp)
1457 {
1458     VFIOGroup *group;
1459     char path[32];
1460     struct vfio_group_status status = { .argsz = sizeof(status) };
1461 
1462     QLIST_FOREACH(group, &vfio_group_list, next) {
1463         if (group->groupid == groupid) {
1464             /* Found it.  Now is it already in the right context? */
1465             if (group->container->space->as == as) {
1466                 return group;
1467             } else {
1468                 error_setg(errp, "group %d used in multiple address spaces",
1469                            group->groupid);
1470                 return NULL;
1471             }
1472         }
1473     }
1474 
1475     group = g_malloc0(sizeof(*group));
1476 
1477     snprintf(path, sizeof(path), "/dev/vfio/%d", groupid);
1478     group->fd = qemu_open(path, O_RDWR);
1479     if (group->fd < 0) {
1480         error_setg_errno(errp, errno, "failed to open %s", path);
1481         goto free_group_exit;
1482     }
1483 
1484     if (ioctl(group->fd, VFIO_GROUP_GET_STATUS, &status)) {
1485         error_setg_errno(errp, errno, "failed to get group %d status", groupid);
1486         goto close_fd_exit;
1487     }
1488 
1489     if (!(status.flags & VFIO_GROUP_FLAGS_VIABLE)) {
1490         error_setg(errp, "group %d is not viable", groupid);
1491         error_append_hint(errp,
1492                           "Please ensure all devices within the iommu_group "
1493                           "are bound to their vfio bus driver.\n");
1494         goto close_fd_exit;
1495     }
1496 
1497     group->groupid = groupid;
1498     QLIST_INIT(&group->device_list);
1499 
1500     if (vfio_connect_container(group, as, errp)) {
1501         error_prepend(errp, "failed to setup container for group %d: ",
1502                       groupid);
1503         goto close_fd_exit;
1504     }
1505 
1506     if (QLIST_EMPTY(&vfio_group_list)) {
1507         qemu_register_reset(vfio_reset_handler, NULL);
1508     }
1509 
1510     QLIST_INSERT_HEAD(&vfio_group_list, group, next);
1511 
1512     return group;
1513 
1514 close_fd_exit:
1515     close(group->fd);
1516 
1517 free_group_exit:
1518     g_free(group);
1519 
1520     return NULL;
1521 }
1522 
1523 void vfio_put_group(VFIOGroup *group)
1524 {
1525     if (!group || !QLIST_EMPTY(&group->device_list)) {
1526         return;
1527     }
1528 
1529     if (!group->balloon_allowed) {
1530         qemu_balloon_inhibit(false);
1531     }
1532     vfio_kvm_device_del_group(group);
1533     vfio_disconnect_container(group);
1534     QLIST_REMOVE(group, next);
1535     trace_vfio_put_group(group->fd);
1536     close(group->fd);
1537     g_free(group);
1538 
1539     if (QLIST_EMPTY(&vfio_group_list)) {
1540         qemu_unregister_reset(vfio_reset_handler, NULL);
1541     }
1542 }
1543 
1544 int vfio_get_device(VFIOGroup *group, const char *name,
1545                     VFIODevice *vbasedev, Error **errp)
1546 {
1547     struct vfio_device_info dev_info = { .argsz = sizeof(dev_info) };
1548     int ret, fd;
1549 
1550     fd = ioctl(group->fd, VFIO_GROUP_GET_DEVICE_FD, name);
1551     if (fd < 0) {
1552         error_setg_errno(errp, errno, "error getting device from group %d",
1553                          group->groupid);
1554         error_append_hint(errp,
1555                       "Verify all devices in group %d are bound to vfio-<bus> "
1556                       "or pci-stub and not already in use\n", group->groupid);
1557         return fd;
1558     }
1559 
1560     ret = ioctl(fd, VFIO_DEVICE_GET_INFO, &dev_info);
1561     if (ret) {
1562         error_setg_errno(errp, errno, "error getting device info");
1563         close(fd);
1564         return ret;
1565     }
1566 
1567     /*
1568      * Clear the balloon inhibitor for this group if the driver knows the
1569      * device operates compatibly with ballooning.  Setting must be consistent
1570      * per group, but since compatibility is really only possible with mdev
1571      * currently, we expect singleton groups.
1572      */
1573     if (vbasedev->balloon_allowed != group->balloon_allowed) {
1574         if (!QLIST_EMPTY(&group->device_list)) {
1575             error_setg(errp,
1576                        "Inconsistent device balloon setting within group");
1577             close(fd);
1578             return -1;
1579         }
1580 
1581         if (!group->balloon_allowed) {
1582             group->balloon_allowed = true;
1583             qemu_balloon_inhibit(false);
1584         }
1585     }
1586 
1587     vbasedev->fd = fd;
1588     vbasedev->group = group;
1589     QLIST_INSERT_HEAD(&group->device_list, vbasedev, next);
1590 
1591     vbasedev->num_irqs = dev_info.num_irqs;
1592     vbasedev->num_regions = dev_info.num_regions;
1593     vbasedev->flags = dev_info.flags;
1594 
1595     trace_vfio_get_device(name, dev_info.flags, dev_info.num_regions,
1596                           dev_info.num_irqs);
1597 
1598     vbasedev->reset_works = !!(dev_info.flags & VFIO_DEVICE_FLAGS_RESET);
1599     return 0;
1600 }
1601 
1602 void vfio_put_base_device(VFIODevice *vbasedev)
1603 {
1604     if (!vbasedev->group) {
1605         return;
1606     }
1607     QLIST_REMOVE(vbasedev, next);
1608     vbasedev->group = NULL;
1609     trace_vfio_put_base_device(vbasedev->fd);
1610     close(vbasedev->fd);
1611 }
1612 
1613 int vfio_get_region_info(VFIODevice *vbasedev, int index,
1614                          struct vfio_region_info **info)
1615 {
1616     size_t argsz = sizeof(struct vfio_region_info);
1617 
1618     *info = g_malloc0(argsz);
1619 
1620     (*info)->index = index;
1621 retry:
1622     (*info)->argsz = argsz;
1623 
1624     if (ioctl(vbasedev->fd, VFIO_DEVICE_GET_REGION_INFO, *info)) {
1625         g_free(*info);
1626         *info = NULL;
1627         return -errno;
1628     }
1629 
1630     if ((*info)->argsz > argsz) {
1631         argsz = (*info)->argsz;
1632         *info = g_realloc(*info, argsz);
1633 
1634         goto retry;
1635     }
1636 
1637     return 0;
1638 }
1639 
1640 int vfio_get_dev_region_info(VFIODevice *vbasedev, uint32_t type,
1641                              uint32_t subtype, struct vfio_region_info **info)
1642 {
1643     int i;
1644 
1645     for (i = 0; i < vbasedev->num_regions; i++) {
1646         struct vfio_info_cap_header *hdr;
1647         struct vfio_region_info_cap_type *cap_type;
1648 
1649         if (vfio_get_region_info(vbasedev, i, info)) {
1650             continue;
1651         }
1652 
1653         hdr = vfio_get_region_info_cap(*info, VFIO_REGION_INFO_CAP_TYPE);
1654         if (!hdr) {
1655             g_free(*info);
1656             continue;
1657         }
1658 
1659         cap_type = container_of(hdr, struct vfio_region_info_cap_type, header);
1660 
1661         trace_vfio_get_dev_region(vbasedev->name, i,
1662                                   cap_type->type, cap_type->subtype);
1663 
1664         if (cap_type->type == type && cap_type->subtype == subtype) {
1665             return 0;
1666         }
1667 
1668         g_free(*info);
1669     }
1670 
1671     *info = NULL;
1672     return -ENODEV;
1673 }
1674 
1675 bool vfio_has_region_cap(VFIODevice *vbasedev, int region, uint16_t cap_type)
1676 {
1677     struct vfio_region_info *info = NULL;
1678     bool ret = false;
1679 
1680     if (!vfio_get_region_info(vbasedev, region, &info)) {
1681         if (vfio_get_region_info_cap(info, cap_type)) {
1682             ret = true;
1683         }
1684         g_free(info);
1685     }
1686 
1687     return ret;
1688 }
1689 
1690 /*
1691  * Interfaces for IBM EEH (Enhanced Error Handling)
1692  */
1693 static bool vfio_eeh_container_ok(VFIOContainer *container)
1694 {
1695     /*
1696      * As of 2016-03-04 (linux-4.5) the host kernel EEH/VFIO
1697      * implementation is broken if there are multiple groups in a
1698      * container.  The hardware works in units of Partitionable
1699      * Endpoints (== IOMMU groups) and the EEH operations naively
1700      * iterate across all groups in the container, without any logic
1701      * to make sure the groups have their state synchronized.  For
1702      * certain operations (ENABLE) that might be ok, until an error
1703      * occurs, but for others (GET_STATE) it's clearly broken.
1704      */
1705 
1706     /*
1707      * XXX Once fixed kernels exist, test for them here
1708      */
1709 
1710     if (QLIST_EMPTY(&container->group_list)) {
1711         return false;
1712     }
1713 
1714     if (QLIST_NEXT(QLIST_FIRST(&container->group_list), container_next)) {
1715         return false;
1716     }
1717 
1718     return true;
1719 }
1720 
1721 static int vfio_eeh_container_op(VFIOContainer *container, uint32_t op)
1722 {
1723     struct vfio_eeh_pe_op pe_op = {
1724         .argsz = sizeof(pe_op),
1725         .op = op,
1726     };
1727     int ret;
1728 
1729     if (!vfio_eeh_container_ok(container)) {
1730         error_report("vfio/eeh: EEH_PE_OP 0x%x: "
1731                      "kernel requires a container with exactly one group", op);
1732         return -EPERM;
1733     }
1734 
1735     ret = ioctl(container->fd, VFIO_EEH_PE_OP, &pe_op);
1736     if (ret < 0) {
1737         error_report("vfio/eeh: EEH_PE_OP 0x%x failed: %m", op);
1738         return -errno;
1739     }
1740 
1741     return ret;
1742 }
1743 
1744 static VFIOContainer *vfio_eeh_as_container(AddressSpace *as)
1745 {
1746     VFIOAddressSpace *space = vfio_get_address_space(as);
1747     VFIOContainer *container = NULL;
1748 
1749     if (QLIST_EMPTY(&space->containers)) {
1750         /* No containers to act on */
1751         goto out;
1752     }
1753 
1754     container = QLIST_FIRST(&space->containers);
1755 
1756     if (QLIST_NEXT(container, next)) {
1757         /* We don't yet have logic to synchronize EEH state across
1758          * multiple containers */
1759         container = NULL;
1760         goto out;
1761     }
1762 
1763 out:
1764     vfio_put_address_space(space);
1765     return container;
1766 }
1767 
1768 bool vfio_eeh_as_ok(AddressSpace *as)
1769 {
1770     VFIOContainer *container = vfio_eeh_as_container(as);
1771 
1772     return (container != NULL) && vfio_eeh_container_ok(container);
1773 }
1774 
1775 int vfio_eeh_as_op(AddressSpace *as, uint32_t op)
1776 {
1777     VFIOContainer *container = vfio_eeh_as_container(as);
1778 
1779     if (!container) {
1780         return -ENODEV;
1781     }
1782     return vfio_eeh_container_op(container, op);
1783 }
1784