xref: /openbmc/qemu/hw/vfio/common.c (revision fe7f9b8e)
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/range.h"
35 #include "sysemu/kvm.h"
36 #include "trace.h"
37 #include "qapi/error.h"
38 
39 struct vfio_group_head vfio_group_list =
40     QLIST_HEAD_INITIALIZER(vfio_group_list);
41 struct vfio_as_head vfio_address_spaces =
42     QLIST_HEAD_INITIALIZER(vfio_address_spaces);
43 
44 #ifdef CONFIG_KVM
45 /*
46  * We have a single VFIO pseudo device per KVM VM.  Once created it lives
47  * for the life of the VM.  Closing the file descriptor only drops our
48  * reference to it and the device's reference to kvm.  Therefore once
49  * initialized, this file descriptor is only released on QEMU exit and
50  * we'll re-use it should another vfio device be attached before then.
51  */
52 static int vfio_kvm_device_fd = -1;
53 #endif
54 
55 /*
56  * Common VFIO interrupt disable
57  */
58 void vfio_disable_irqindex(VFIODevice *vbasedev, int index)
59 {
60     struct vfio_irq_set irq_set = {
61         .argsz = sizeof(irq_set),
62         .flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_TRIGGER,
63         .index = index,
64         .start = 0,
65         .count = 0,
66     };
67 
68     ioctl(vbasedev->fd, VFIO_DEVICE_SET_IRQS, &irq_set);
69 }
70 
71 void vfio_unmask_single_irqindex(VFIODevice *vbasedev, int index)
72 {
73     struct vfio_irq_set irq_set = {
74         .argsz = sizeof(irq_set),
75         .flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_UNMASK,
76         .index = index,
77         .start = 0,
78         .count = 1,
79     };
80 
81     ioctl(vbasedev->fd, VFIO_DEVICE_SET_IRQS, &irq_set);
82 }
83 
84 void vfio_mask_single_irqindex(VFIODevice *vbasedev, int index)
85 {
86     struct vfio_irq_set irq_set = {
87         .argsz = sizeof(irq_set),
88         .flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_MASK,
89         .index = index,
90         .start = 0,
91         .count = 1,
92     };
93 
94     ioctl(vbasedev->fd, VFIO_DEVICE_SET_IRQS, &irq_set);
95 }
96 
97 /*
98  * IO Port/MMIO - Beware of the endians, VFIO is always little endian
99  */
100 void vfio_region_write(void *opaque, hwaddr addr,
101                        uint64_t data, unsigned size)
102 {
103     VFIORegion *region = opaque;
104     VFIODevice *vbasedev = region->vbasedev;
105     union {
106         uint8_t byte;
107         uint16_t word;
108         uint32_t dword;
109         uint64_t qword;
110     } buf;
111 
112     switch (size) {
113     case 1:
114         buf.byte = data;
115         break;
116     case 2:
117         buf.word = cpu_to_le16(data);
118         break;
119     case 4:
120         buf.dword = cpu_to_le32(data);
121         break;
122     case 8:
123         buf.qword = cpu_to_le64(data);
124         break;
125     default:
126         hw_error("vfio: unsupported write size, %d bytes", size);
127         break;
128     }
129 
130     if (pwrite(vbasedev->fd, &buf, size, region->fd_offset + addr) != size) {
131         error_report("%s(%s:region%d+0x%"HWADDR_PRIx", 0x%"PRIx64
132                      ",%d) failed: %m",
133                      __func__, vbasedev->name, region->nr,
134                      addr, data, size);
135     }
136 
137     trace_vfio_region_write(vbasedev->name, region->nr, addr, data, size);
138 
139     /*
140      * A read or write to a BAR always signals an INTx EOI.  This will
141      * do nothing if not pending (including not in INTx mode).  We assume
142      * that a BAR access is in response to an interrupt and that BAR
143      * accesses will service the interrupt.  Unfortunately, we don't know
144      * which access will service the interrupt, so we're potentially
145      * getting quite a few host interrupts per guest interrupt.
146      */
147     vbasedev->ops->vfio_eoi(vbasedev);
148 }
149 
150 uint64_t vfio_region_read(void *opaque,
151                           hwaddr addr, unsigned size)
152 {
153     VFIORegion *region = opaque;
154     VFIODevice *vbasedev = region->vbasedev;
155     union {
156         uint8_t byte;
157         uint16_t word;
158         uint32_t dword;
159         uint64_t qword;
160     } buf;
161     uint64_t data = 0;
162 
163     if (pread(vbasedev->fd, &buf, size, region->fd_offset + addr) != size) {
164         error_report("%s(%s:region%d+0x%"HWADDR_PRIx", %d) failed: %m",
165                      __func__, vbasedev->name, region->nr,
166                      addr, size);
167         return (uint64_t)-1;
168     }
169     switch (size) {
170     case 1:
171         data = buf.byte;
172         break;
173     case 2:
174         data = le16_to_cpu(buf.word);
175         break;
176     case 4:
177         data = le32_to_cpu(buf.dword);
178         break;
179     case 8:
180         data = le64_to_cpu(buf.qword);
181         break;
182     default:
183         hw_error("vfio: unsupported read size, %d bytes", size);
184         break;
185     }
186 
187     trace_vfio_region_read(vbasedev->name, region->nr, addr, size, data);
188 
189     /* Same as write above */
190     vbasedev->ops->vfio_eoi(vbasedev);
191 
192     return data;
193 }
194 
195 const MemoryRegionOps vfio_region_ops = {
196     .read = vfio_region_read,
197     .write = vfio_region_write,
198     .endianness = DEVICE_LITTLE_ENDIAN,
199     .valid = {
200         .min_access_size = 1,
201         .max_access_size = 8,
202     },
203     .impl = {
204         .min_access_size = 1,
205         .max_access_size = 8,
206     },
207 };
208 
209 /*
210  * DMA - Mapping and unmapping for the "type1" IOMMU interface used on x86
211  */
212 static int vfio_dma_unmap(VFIOContainer *container,
213                           hwaddr iova, ram_addr_t size)
214 {
215     struct vfio_iommu_type1_dma_unmap unmap = {
216         .argsz = sizeof(unmap),
217         .flags = 0,
218         .iova = iova,
219         .size = size,
220     };
221 
222     if (ioctl(container->fd, VFIO_IOMMU_UNMAP_DMA, &unmap)) {
223         error_report("VFIO_UNMAP_DMA: %d", -errno);
224         return -errno;
225     }
226 
227     return 0;
228 }
229 
230 static int vfio_dma_map(VFIOContainer *container, hwaddr iova,
231                         ram_addr_t size, void *vaddr, bool readonly)
232 {
233     struct vfio_iommu_type1_dma_map map = {
234         .argsz = sizeof(map),
235         .flags = VFIO_DMA_MAP_FLAG_READ,
236         .vaddr = (__u64)(uintptr_t)vaddr,
237         .iova = iova,
238         .size = size,
239     };
240 
241     if (!readonly) {
242         map.flags |= VFIO_DMA_MAP_FLAG_WRITE;
243     }
244 
245     /*
246      * Try the mapping, if it fails with EBUSY, unmap the region and try
247      * again.  This shouldn't be necessary, but we sometimes see it in
248      * the VGA ROM space.
249      */
250     if (ioctl(container->fd, VFIO_IOMMU_MAP_DMA, &map) == 0 ||
251         (errno == EBUSY && vfio_dma_unmap(container, iova, size) == 0 &&
252          ioctl(container->fd, VFIO_IOMMU_MAP_DMA, &map) == 0)) {
253         return 0;
254     }
255 
256     error_report("VFIO_MAP_DMA: %d", -errno);
257     return -errno;
258 }
259 
260 static void vfio_host_win_add(VFIOContainer *container,
261                               hwaddr min_iova, hwaddr max_iova,
262                               uint64_t iova_pgsizes)
263 {
264     VFIOHostDMAWindow *hostwin;
265 
266     QLIST_FOREACH(hostwin, &container->hostwin_list, hostwin_next) {
267         if (ranges_overlap(hostwin->min_iova,
268                            hostwin->max_iova - hostwin->min_iova + 1,
269                            min_iova,
270                            max_iova - min_iova + 1)) {
271             hw_error("%s: Overlapped IOMMU are not enabled", __func__);
272         }
273     }
274 
275     hostwin = g_malloc0(sizeof(*hostwin));
276 
277     hostwin->min_iova = min_iova;
278     hostwin->max_iova = max_iova;
279     hostwin->iova_pgsizes = iova_pgsizes;
280     QLIST_INSERT_HEAD(&container->hostwin_list, hostwin, hostwin_next);
281 }
282 
283 static int vfio_host_win_del(VFIOContainer *container, hwaddr min_iova,
284                              hwaddr max_iova)
285 {
286     VFIOHostDMAWindow *hostwin;
287 
288     QLIST_FOREACH(hostwin, &container->hostwin_list, hostwin_next) {
289         if (hostwin->min_iova == min_iova && hostwin->max_iova == max_iova) {
290             QLIST_REMOVE(hostwin, hostwin_next);
291             return 0;
292         }
293     }
294 
295     return -1;
296 }
297 
298 static bool vfio_listener_skipped_section(MemoryRegionSection *section)
299 {
300     return (!memory_region_is_ram(section->mr) &&
301             !memory_region_is_iommu(section->mr)) ||
302            /*
303             * Sizing an enabled 64-bit BAR can cause spurious mappings to
304             * addresses in the upper part of the 64-bit address space.  These
305             * are never accessed by the CPU and beyond the address width of
306             * some IOMMU hardware.  TODO: VFIO should tell us the IOMMU width.
307             */
308            section->offset_within_address_space & (1ULL << 63);
309 }
310 
311 /* Called with rcu_read_lock held.  */
312 static bool vfio_get_vaddr(IOMMUTLBEntry *iotlb, void **vaddr,
313                            bool *read_only)
314 {
315     MemoryRegion *mr;
316     hwaddr xlat;
317     hwaddr len = iotlb->addr_mask + 1;
318     bool writable = iotlb->perm & IOMMU_WO;
319 
320     /*
321      * The IOMMU TLB entry we have just covers translation through
322      * this IOMMU to its immediate target.  We need to translate
323      * it the rest of the way through to memory.
324      */
325     mr = address_space_translate(&address_space_memory,
326                                  iotlb->translated_addr,
327                                  &xlat, &len, writable);
328     if (!memory_region_is_ram(mr)) {
329         error_report("iommu map to non memory area %"HWADDR_PRIx"",
330                      xlat);
331         return false;
332     }
333 
334     /*
335      * Translation truncates length to the IOMMU page size,
336      * check that it did not truncate too much.
337      */
338     if (len & iotlb->addr_mask) {
339         error_report("iommu has granularity incompatible with target AS");
340         return false;
341     }
342 
343     *vaddr = memory_region_get_ram_ptr(mr) + xlat;
344     *read_only = !writable || mr->readonly;
345 
346     return true;
347 }
348 
349 static void vfio_iommu_map_notify(IOMMUNotifier *n, IOMMUTLBEntry *iotlb)
350 {
351     VFIOGuestIOMMU *giommu = container_of(n, VFIOGuestIOMMU, n);
352     VFIOContainer *container = giommu->container;
353     hwaddr iova = iotlb->iova + giommu->iommu_offset;
354     bool read_only;
355     void *vaddr;
356     int ret;
357 
358     trace_vfio_iommu_map_notify(iotlb->perm == IOMMU_NONE ? "UNMAP" : "MAP",
359                                 iova, iova + iotlb->addr_mask);
360 
361     if (iotlb->target_as != &address_space_memory) {
362         error_report("Wrong target AS \"%s\", only system memory is allowed",
363                      iotlb->target_as->name ? iotlb->target_as->name : "none");
364         return;
365     }
366 
367     rcu_read_lock();
368 
369     if ((iotlb->perm & IOMMU_RW) != IOMMU_NONE) {
370         if (!vfio_get_vaddr(iotlb, &vaddr, &read_only)) {
371             goto out;
372         }
373         /*
374          * vaddr is only valid until rcu_read_unlock(). But after
375          * vfio_dma_map has set up the mapping the pages will be
376          * pinned by the kernel. This makes sure that the RAM backend
377          * of vaddr will always be there, even if the memory object is
378          * destroyed and its backing memory munmap-ed.
379          */
380         ret = vfio_dma_map(container, iova,
381                            iotlb->addr_mask + 1, vaddr,
382                            read_only);
383         if (ret) {
384             error_report("vfio_dma_map(%p, 0x%"HWADDR_PRIx", "
385                          "0x%"HWADDR_PRIx", %p) = %d (%m)",
386                          container, iova,
387                          iotlb->addr_mask + 1, vaddr, ret);
388         }
389     } else {
390         ret = vfio_dma_unmap(container, iova, iotlb->addr_mask + 1);
391         if (ret) {
392             error_report("vfio_dma_unmap(%p, 0x%"HWADDR_PRIx", "
393                          "0x%"HWADDR_PRIx") = %d (%m)",
394                          container, iova,
395                          iotlb->addr_mask + 1, ret);
396         }
397     }
398 out:
399     rcu_read_unlock();
400 }
401 
402 static void vfio_listener_region_add(MemoryListener *listener,
403                                      MemoryRegionSection *section)
404 {
405     VFIOContainer *container = container_of(listener, VFIOContainer, listener);
406     hwaddr iova, end;
407     Int128 llend, llsize;
408     void *vaddr;
409     int ret;
410     VFIOHostDMAWindow *hostwin;
411     bool hostwin_found;
412 
413     if (vfio_listener_skipped_section(section)) {
414         trace_vfio_listener_region_add_skip(
415                 section->offset_within_address_space,
416                 section->offset_within_address_space +
417                 int128_get64(int128_sub(section->size, int128_one())));
418         return;
419     }
420 
421     if (unlikely((section->offset_within_address_space & ~TARGET_PAGE_MASK) !=
422                  (section->offset_within_region & ~TARGET_PAGE_MASK))) {
423         error_report("%s received unaligned region", __func__);
424         return;
425     }
426 
427     iova = TARGET_PAGE_ALIGN(section->offset_within_address_space);
428     llend = int128_make64(section->offset_within_address_space);
429     llend = int128_add(llend, section->size);
430     llend = int128_and(llend, int128_exts64(TARGET_PAGE_MASK));
431 
432     if (int128_ge(int128_make64(iova), llend)) {
433         return;
434     }
435     end = int128_get64(int128_sub(llend, int128_one()));
436 
437     if (container->iommu_type == VFIO_SPAPR_TCE_v2_IOMMU) {
438         hwaddr pgsize = 0;
439 
440         /* For now intersections are not allowed, we may relax this later */
441         QLIST_FOREACH(hostwin, &container->hostwin_list, hostwin_next) {
442             if (ranges_overlap(hostwin->min_iova,
443                                hostwin->max_iova - hostwin->min_iova + 1,
444                                section->offset_within_address_space,
445                                int128_get64(section->size))) {
446                 ret = -1;
447                 goto fail;
448             }
449         }
450 
451         ret = vfio_spapr_create_window(container, section, &pgsize);
452         if (ret) {
453             goto fail;
454         }
455 
456         vfio_host_win_add(container, section->offset_within_address_space,
457                           section->offset_within_address_space +
458                           int128_get64(section->size) - 1, pgsize);
459 #ifdef CONFIG_KVM
460         if (kvm_enabled()) {
461             VFIOGroup *group;
462             IOMMUMemoryRegion *iommu_mr = IOMMU_MEMORY_REGION(section->mr);
463             struct kvm_vfio_spapr_tce param;
464             struct kvm_device_attr attr = {
465                 .group = KVM_DEV_VFIO_GROUP,
466                 .attr = KVM_DEV_VFIO_GROUP_SET_SPAPR_TCE,
467                 .addr = (uint64_t)(unsigned long)&param,
468             };
469 
470             if (!memory_region_iommu_get_attr(iommu_mr, IOMMU_ATTR_SPAPR_TCE_FD,
471                                               &param.tablefd)) {
472                 QLIST_FOREACH(group, &container->group_list, container_next) {
473                     param.groupfd = group->fd;
474                     if (ioctl(vfio_kvm_device_fd, KVM_SET_DEVICE_ATTR, &attr)) {
475                         error_report("vfio: failed to setup fd %d "
476                                      "for a group with fd %d: %s",
477                                      param.tablefd, param.groupfd,
478                                      strerror(errno));
479                         return;
480                     }
481                     trace_vfio_spapr_group_attach(param.groupfd, param.tablefd);
482                 }
483             }
484         }
485 #endif
486     }
487 
488     hostwin_found = false;
489     QLIST_FOREACH(hostwin, &container->hostwin_list, hostwin_next) {
490         if (hostwin->min_iova <= iova && end <= hostwin->max_iova) {
491             hostwin_found = true;
492             break;
493         }
494     }
495 
496     if (!hostwin_found) {
497         error_report("vfio: IOMMU container %p can't map guest IOVA region"
498                      " 0x%"HWADDR_PRIx"..0x%"HWADDR_PRIx,
499                      container, iova, end);
500         ret = -EFAULT;
501         goto fail;
502     }
503 
504     memory_region_ref(section->mr);
505 
506     if (memory_region_is_iommu(section->mr)) {
507         VFIOGuestIOMMU *giommu;
508         IOMMUMemoryRegion *iommu_mr = IOMMU_MEMORY_REGION(section->mr);
509 
510         trace_vfio_listener_region_add_iommu(iova, end);
511         /*
512          * FIXME: For VFIO iommu types which have KVM acceleration to
513          * avoid bouncing all map/unmaps through qemu this way, this
514          * would be the right place to wire that up (tell the KVM
515          * device emulation the VFIO iommu handles to use).
516          */
517         giommu = g_malloc0(sizeof(*giommu));
518         giommu->iommu = iommu_mr;
519         giommu->iommu_offset = section->offset_within_address_space -
520                                section->offset_within_region;
521         giommu->container = container;
522         llend = int128_add(int128_make64(section->offset_within_region),
523                            section->size);
524         llend = int128_sub(llend, int128_one());
525         iommu_notifier_init(&giommu->n, vfio_iommu_map_notify,
526                             IOMMU_NOTIFIER_ALL,
527                             section->offset_within_region,
528                             int128_get64(llend));
529         QLIST_INSERT_HEAD(&container->giommu_list, giommu, giommu_next);
530 
531         memory_region_register_iommu_notifier(section->mr, &giommu->n);
532         memory_region_iommu_replay(giommu->iommu, &giommu->n);
533 
534         return;
535     }
536 
537     /* Here we assume that memory_region_is_ram(section->mr)==true */
538 
539     vaddr = memory_region_get_ram_ptr(section->mr) +
540             section->offset_within_region +
541             (iova - section->offset_within_address_space);
542 
543     trace_vfio_listener_region_add_ram(iova, end, vaddr);
544 
545     llsize = int128_sub(llend, int128_make64(iova));
546 
547     if (memory_region_is_ram_device(section->mr)) {
548         hwaddr pgmask = (1ULL << ctz64(hostwin->iova_pgsizes)) - 1;
549 
550         if ((iova & pgmask) || (int128_get64(llsize) & pgmask)) {
551             trace_vfio_listener_region_add_no_dma_map(
552                 memory_region_name(section->mr),
553                 section->offset_within_address_space,
554                 int128_getlo(section->size),
555                 pgmask + 1);
556             return;
557         }
558     }
559 
560     ret = vfio_dma_map(container, iova, int128_get64(llsize),
561                        vaddr, section->readonly);
562     if (ret) {
563         error_report("vfio_dma_map(%p, 0x%"HWADDR_PRIx", "
564                      "0x%"HWADDR_PRIx", %p) = %d (%m)",
565                      container, iova, int128_get64(llsize), vaddr, ret);
566         if (memory_region_is_ram_device(section->mr)) {
567             /* Allow unexpected mappings not to be fatal for RAM devices */
568             return;
569         }
570         goto fail;
571     }
572 
573     return;
574 
575 fail:
576     if (memory_region_is_ram_device(section->mr)) {
577         error_report("failed to vfio_dma_map. pci p2p may not work");
578         return;
579     }
580     /*
581      * On the initfn path, store the first error in the container so we
582      * can gracefully fail.  Runtime, there's not much we can do other
583      * than throw a hardware error.
584      */
585     if (!container->initialized) {
586         if (!container->error) {
587             container->error = ret;
588         }
589     } else {
590         hw_error("vfio: DMA mapping failed, unable to continue");
591     }
592 }
593 
594 static void vfio_listener_region_del(MemoryListener *listener,
595                                      MemoryRegionSection *section)
596 {
597     VFIOContainer *container = container_of(listener, VFIOContainer, listener);
598     hwaddr iova, end;
599     Int128 llend, llsize;
600     int ret;
601     bool try_unmap = true;
602 
603     if (vfio_listener_skipped_section(section)) {
604         trace_vfio_listener_region_del_skip(
605                 section->offset_within_address_space,
606                 section->offset_within_address_space +
607                 int128_get64(int128_sub(section->size, int128_one())));
608         return;
609     }
610 
611     if (unlikely((section->offset_within_address_space & ~TARGET_PAGE_MASK) !=
612                  (section->offset_within_region & ~TARGET_PAGE_MASK))) {
613         error_report("%s received unaligned region", __func__);
614         return;
615     }
616 
617     if (memory_region_is_iommu(section->mr)) {
618         VFIOGuestIOMMU *giommu;
619 
620         QLIST_FOREACH(giommu, &container->giommu_list, giommu_next) {
621             if (MEMORY_REGION(giommu->iommu) == section->mr &&
622                 giommu->n.start == section->offset_within_region) {
623                 memory_region_unregister_iommu_notifier(section->mr,
624                                                         &giommu->n);
625                 QLIST_REMOVE(giommu, giommu_next);
626                 g_free(giommu);
627                 break;
628             }
629         }
630 
631         /*
632          * FIXME: We assume the one big unmap below is adequate to
633          * remove any individual page mappings in the IOMMU which
634          * might have been copied into VFIO. This works for a page table
635          * based IOMMU where a big unmap flattens a large range of IO-PTEs.
636          * That may not be true for all IOMMU types.
637          */
638     }
639 
640     iova = TARGET_PAGE_ALIGN(section->offset_within_address_space);
641     llend = int128_make64(section->offset_within_address_space);
642     llend = int128_add(llend, section->size);
643     llend = int128_and(llend, int128_exts64(TARGET_PAGE_MASK));
644 
645     if (int128_ge(int128_make64(iova), llend)) {
646         return;
647     }
648     end = int128_get64(int128_sub(llend, int128_one()));
649 
650     llsize = int128_sub(llend, int128_make64(iova));
651 
652     trace_vfio_listener_region_del(iova, end);
653 
654     if (memory_region_is_ram_device(section->mr)) {
655         hwaddr pgmask;
656         VFIOHostDMAWindow *hostwin;
657         bool hostwin_found = false;
658 
659         QLIST_FOREACH(hostwin, &container->hostwin_list, hostwin_next) {
660             if (hostwin->min_iova <= iova && end <= hostwin->max_iova) {
661                 hostwin_found = true;
662                 break;
663             }
664         }
665         assert(hostwin_found); /* or region_add() would have failed */
666 
667         pgmask = (1ULL << ctz64(hostwin->iova_pgsizes)) - 1;
668         try_unmap = !((iova & pgmask) || (int128_get64(llsize) & pgmask));
669     }
670 
671     if (try_unmap) {
672         ret = vfio_dma_unmap(container, iova, int128_get64(llsize));
673         if (ret) {
674             error_report("vfio_dma_unmap(%p, 0x%"HWADDR_PRIx", "
675                          "0x%"HWADDR_PRIx") = %d (%m)",
676                          container, iova, int128_get64(llsize), ret);
677         }
678     }
679 
680     memory_region_unref(section->mr);
681 
682     if (container->iommu_type == VFIO_SPAPR_TCE_v2_IOMMU) {
683         vfio_spapr_remove_window(container,
684                                  section->offset_within_address_space);
685         if (vfio_host_win_del(container,
686                               section->offset_within_address_space,
687                               section->offset_within_address_space +
688                               int128_get64(section->size) - 1) < 0) {
689             hw_error("%s: Cannot delete missing window at %"HWADDR_PRIx,
690                      __func__, section->offset_within_address_space);
691         }
692     }
693 }
694 
695 static const MemoryListener vfio_memory_listener = {
696     .region_add = vfio_listener_region_add,
697     .region_del = vfio_listener_region_del,
698 };
699 
700 static void vfio_listener_release(VFIOContainer *container)
701 {
702     memory_listener_unregister(&container->listener);
703     if (container->iommu_type == VFIO_SPAPR_TCE_v2_IOMMU) {
704         memory_listener_unregister(&container->prereg_listener);
705     }
706 }
707 
708 static struct vfio_info_cap_header *
709 vfio_get_region_info_cap(struct vfio_region_info *info, uint16_t id)
710 {
711     struct vfio_info_cap_header *hdr;
712     void *ptr = info;
713 
714     if (!(info->flags & VFIO_REGION_INFO_FLAG_CAPS)) {
715         return NULL;
716     }
717 
718     for (hdr = ptr + info->cap_offset; hdr != ptr; hdr = ptr + hdr->next) {
719         if (hdr->id == id) {
720             return hdr;
721         }
722     }
723 
724     return NULL;
725 }
726 
727 static int vfio_setup_region_sparse_mmaps(VFIORegion *region,
728                                           struct vfio_region_info *info)
729 {
730     struct vfio_info_cap_header *hdr;
731     struct vfio_region_info_cap_sparse_mmap *sparse;
732     int i, j;
733 
734     hdr = vfio_get_region_info_cap(info, VFIO_REGION_INFO_CAP_SPARSE_MMAP);
735     if (!hdr) {
736         return -ENODEV;
737     }
738 
739     sparse = container_of(hdr, struct vfio_region_info_cap_sparse_mmap, header);
740 
741     trace_vfio_region_sparse_mmap_header(region->vbasedev->name,
742                                          region->nr, sparse->nr_areas);
743 
744     region->mmaps = g_new0(VFIOMmap, sparse->nr_areas);
745 
746     for (i = 0, j = 0; i < sparse->nr_areas; i++) {
747         trace_vfio_region_sparse_mmap_entry(i, sparse->areas[i].offset,
748                                             sparse->areas[i].offset +
749                                             sparse->areas[i].size);
750 
751         if (sparse->areas[i].size) {
752             region->mmaps[j].offset = sparse->areas[i].offset;
753             region->mmaps[j].size = sparse->areas[i].size;
754             j++;
755         }
756     }
757 
758     region->nr_mmaps = j;
759     region->mmaps = g_realloc(region->mmaps, j * sizeof(VFIOMmap));
760 
761     return 0;
762 }
763 
764 int vfio_region_setup(Object *obj, VFIODevice *vbasedev, VFIORegion *region,
765                       int index, const char *name)
766 {
767     struct vfio_region_info *info;
768     int ret;
769 
770     ret = vfio_get_region_info(vbasedev, index, &info);
771     if (ret) {
772         return ret;
773     }
774 
775     region->vbasedev = vbasedev;
776     region->flags = info->flags;
777     region->size = info->size;
778     region->fd_offset = info->offset;
779     region->nr = index;
780 
781     if (region->size) {
782         region->mem = g_new0(MemoryRegion, 1);
783         memory_region_init_io(region->mem, obj, &vfio_region_ops,
784                               region, name, region->size);
785 
786         if (!vbasedev->no_mmap &&
787             region->flags & VFIO_REGION_INFO_FLAG_MMAP) {
788 
789             ret = vfio_setup_region_sparse_mmaps(region, info);
790 
791             if (ret) {
792                 region->nr_mmaps = 1;
793                 region->mmaps = g_new0(VFIOMmap, region->nr_mmaps);
794                 region->mmaps[0].offset = 0;
795                 region->mmaps[0].size = region->size;
796             }
797         }
798     }
799 
800     g_free(info);
801 
802     trace_vfio_region_setup(vbasedev->name, index, name,
803                             region->flags, region->fd_offset, region->size);
804     return 0;
805 }
806 
807 int vfio_region_mmap(VFIORegion *region)
808 {
809     int i, prot = 0;
810     char *name;
811 
812     if (!region->mem) {
813         return 0;
814     }
815 
816     prot |= region->flags & VFIO_REGION_INFO_FLAG_READ ? PROT_READ : 0;
817     prot |= region->flags & VFIO_REGION_INFO_FLAG_WRITE ? PROT_WRITE : 0;
818 
819     for (i = 0; i < region->nr_mmaps; i++) {
820         region->mmaps[i].mmap = mmap(NULL, region->mmaps[i].size, prot,
821                                      MAP_SHARED, region->vbasedev->fd,
822                                      region->fd_offset +
823                                      region->mmaps[i].offset);
824         if (region->mmaps[i].mmap == MAP_FAILED) {
825             int ret = -errno;
826 
827             trace_vfio_region_mmap_fault(memory_region_name(region->mem), i,
828                                          region->fd_offset +
829                                          region->mmaps[i].offset,
830                                          region->fd_offset +
831                                          region->mmaps[i].offset +
832                                          region->mmaps[i].size - 1, ret);
833 
834             region->mmaps[i].mmap = NULL;
835 
836             for (i--; i >= 0; i--) {
837                 memory_region_del_subregion(region->mem, &region->mmaps[i].mem);
838                 munmap(region->mmaps[i].mmap, region->mmaps[i].size);
839                 object_unparent(OBJECT(&region->mmaps[i].mem));
840                 region->mmaps[i].mmap = NULL;
841             }
842 
843             return ret;
844         }
845 
846         name = g_strdup_printf("%s mmaps[%d]",
847                                memory_region_name(region->mem), i);
848         memory_region_init_ram_device_ptr(&region->mmaps[i].mem,
849                                           memory_region_owner(region->mem),
850                                           name, region->mmaps[i].size,
851                                           region->mmaps[i].mmap);
852         g_free(name);
853         memory_region_add_subregion(region->mem, region->mmaps[i].offset,
854                                     &region->mmaps[i].mem);
855 
856         trace_vfio_region_mmap(memory_region_name(&region->mmaps[i].mem),
857                                region->mmaps[i].offset,
858                                region->mmaps[i].offset +
859                                region->mmaps[i].size - 1);
860     }
861 
862     return 0;
863 }
864 
865 void vfio_region_exit(VFIORegion *region)
866 {
867     int i;
868 
869     if (!region->mem) {
870         return;
871     }
872 
873     for (i = 0; i < region->nr_mmaps; i++) {
874         if (region->mmaps[i].mmap) {
875             memory_region_del_subregion(region->mem, &region->mmaps[i].mem);
876         }
877     }
878 
879     trace_vfio_region_exit(region->vbasedev->name, region->nr);
880 }
881 
882 void vfio_region_finalize(VFIORegion *region)
883 {
884     int i;
885 
886     if (!region->mem) {
887         return;
888     }
889 
890     for (i = 0; i < region->nr_mmaps; i++) {
891         if (region->mmaps[i].mmap) {
892             munmap(region->mmaps[i].mmap, region->mmaps[i].size);
893             object_unparent(OBJECT(&region->mmaps[i].mem));
894         }
895     }
896 
897     object_unparent(OBJECT(region->mem));
898 
899     g_free(region->mem);
900     g_free(region->mmaps);
901 
902     trace_vfio_region_finalize(region->vbasedev->name, region->nr);
903 
904     region->mem = NULL;
905     region->mmaps = NULL;
906     region->nr_mmaps = 0;
907     region->size = 0;
908     region->flags = 0;
909     region->nr = 0;
910 }
911 
912 void vfio_region_mmaps_set_enabled(VFIORegion *region, bool enabled)
913 {
914     int i;
915 
916     if (!region->mem) {
917         return;
918     }
919 
920     for (i = 0; i < region->nr_mmaps; i++) {
921         if (region->mmaps[i].mmap) {
922             memory_region_set_enabled(&region->mmaps[i].mem, enabled);
923         }
924     }
925 
926     trace_vfio_region_mmaps_set_enabled(memory_region_name(region->mem),
927                                         enabled);
928 }
929 
930 void vfio_reset_handler(void *opaque)
931 {
932     VFIOGroup *group;
933     VFIODevice *vbasedev;
934 
935     QLIST_FOREACH(group, &vfio_group_list, next) {
936         QLIST_FOREACH(vbasedev, &group->device_list, next) {
937             if (vbasedev->dev->realized) {
938                 vbasedev->ops->vfio_compute_needs_reset(vbasedev);
939             }
940         }
941     }
942 
943     QLIST_FOREACH(group, &vfio_group_list, next) {
944         QLIST_FOREACH(vbasedev, &group->device_list, next) {
945             if (vbasedev->dev->realized && vbasedev->needs_reset) {
946                 vbasedev->ops->vfio_hot_reset_multi(vbasedev);
947             }
948         }
949     }
950 }
951 
952 static void vfio_kvm_device_add_group(VFIOGroup *group)
953 {
954 #ifdef CONFIG_KVM
955     struct kvm_device_attr attr = {
956         .group = KVM_DEV_VFIO_GROUP,
957         .attr = KVM_DEV_VFIO_GROUP_ADD,
958         .addr = (uint64_t)(unsigned long)&group->fd,
959     };
960 
961     if (!kvm_enabled()) {
962         return;
963     }
964 
965     if (vfio_kvm_device_fd < 0) {
966         struct kvm_create_device cd = {
967             .type = KVM_DEV_TYPE_VFIO,
968         };
969 
970         if (kvm_vm_ioctl(kvm_state, KVM_CREATE_DEVICE, &cd)) {
971             error_report("Failed to create KVM VFIO device: %m");
972             return;
973         }
974 
975         vfio_kvm_device_fd = cd.fd;
976     }
977 
978     if (ioctl(vfio_kvm_device_fd, KVM_SET_DEVICE_ATTR, &attr)) {
979         error_report("Failed to add group %d to KVM VFIO device: %m",
980                      group->groupid);
981     }
982 #endif
983 }
984 
985 static void vfio_kvm_device_del_group(VFIOGroup *group)
986 {
987 #ifdef CONFIG_KVM
988     struct kvm_device_attr attr = {
989         .group = KVM_DEV_VFIO_GROUP,
990         .attr = KVM_DEV_VFIO_GROUP_DEL,
991         .addr = (uint64_t)(unsigned long)&group->fd,
992     };
993 
994     if (vfio_kvm_device_fd < 0) {
995         return;
996     }
997 
998     if (ioctl(vfio_kvm_device_fd, KVM_SET_DEVICE_ATTR, &attr)) {
999         error_report("Failed to remove group %d from KVM VFIO device: %m",
1000                      group->groupid);
1001     }
1002 #endif
1003 }
1004 
1005 static VFIOAddressSpace *vfio_get_address_space(AddressSpace *as)
1006 {
1007     VFIOAddressSpace *space;
1008 
1009     QLIST_FOREACH(space, &vfio_address_spaces, list) {
1010         if (space->as == as) {
1011             return space;
1012         }
1013     }
1014 
1015     /* No suitable VFIOAddressSpace, create a new one */
1016     space = g_malloc0(sizeof(*space));
1017     space->as = as;
1018     QLIST_INIT(&space->containers);
1019 
1020     QLIST_INSERT_HEAD(&vfio_address_spaces, space, list);
1021 
1022     return space;
1023 }
1024 
1025 static void vfio_put_address_space(VFIOAddressSpace *space)
1026 {
1027     if (QLIST_EMPTY(&space->containers)) {
1028         QLIST_REMOVE(space, list);
1029         g_free(space);
1030     }
1031 }
1032 
1033 static int vfio_connect_container(VFIOGroup *group, AddressSpace *as,
1034                                   Error **errp)
1035 {
1036     VFIOContainer *container;
1037     int ret, fd;
1038     VFIOAddressSpace *space;
1039 
1040     space = vfio_get_address_space(as);
1041 
1042     QLIST_FOREACH(container, &space->containers, next) {
1043         if (!ioctl(group->fd, VFIO_GROUP_SET_CONTAINER, &container->fd)) {
1044             group->container = container;
1045             QLIST_INSERT_HEAD(&container->group_list, group, container_next);
1046             vfio_kvm_device_add_group(group);
1047             return 0;
1048         }
1049     }
1050 
1051     fd = qemu_open("/dev/vfio/vfio", O_RDWR);
1052     if (fd < 0) {
1053         error_setg_errno(errp, errno, "failed to open /dev/vfio/vfio");
1054         ret = -errno;
1055         goto put_space_exit;
1056     }
1057 
1058     ret = ioctl(fd, VFIO_GET_API_VERSION);
1059     if (ret != VFIO_API_VERSION) {
1060         error_setg(errp, "supported vfio version: %d, "
1061                    "reported version: %d", VFIO_API_VERSION, ret);
1062         ret = -EINVAL;
1063         goto close_fd_exit;
1064     }
1065 
1066     container = g_malloc0(sizeof(*container));
1067     container->space = space;
1068     container->fd = fd;
1069     QLIST_INIT(&container->giommu_list);
1070     QLIST_INIT(&container->hostwin_list);
1071     if (ioctl(fd, VFIO_CHECK_EXTENSION, VFIO_TYPE1_IOMMU) ||
1072         ioctl(fd, VFIO_CHECK_EXTENSION, VFIO_TYPE1v2_IOMMU)) {
1073         bool v2 = !!ioctl(fd, VFIO_CHECK_EXTENSION, VFIO_TYPE1v2_IOMMU);
1074         struct vfio_iommu_type1_info info;
1075 
1076         ret = ioctl(group->fd, VFIO_GROUP_SET_CONTAINER, &fd);
1077         if (ret) {
1078             error_setg_errno(errp, errno, "failed to set group container");
1079             ret = -errno;
1080             goto free_container_exit;
1081         }
1082 
1083         container->iommu_type = v2 ? VFIO_TYPE1v2_IOMMU : VFIO_TYPE1_IOMMU;
1084         ret = ioctl(fd, VFIO_SET_IOMMU, container->iommu_type);
1085         if (ret) {
1086             error_setg_errno(errp, errno, "failed to set iommu for container");
1087             ret = -errno;
1088             goto free_container_exit;
1089         }
1090 
1091         /*
1092          * FIXME: This assumes that a Type1 IOMMU can map any 64-bit
1093          * IOVA whatsoever.  That's not actually true, but the current
1094          * kernel interface doesn't tell us what it can map, and the
1095          * existing Type1 IOMMUs generally support any IOVA we're
1096          * going to actually try in practice.
1097          */
1098         info.argsz = sizeof(info);
1099         ret = ioctl(fd, VFIO_IOMMU_GET_INFO, &info);
1100         /* Ignore errors */
1101         if (ret || !(info.flags & VFIO_IOMMU_INFO_PGSIZES)) {
1102             /* Assume 4k IOVA page size */
1103             info.iova_pgsizes = 4096;
1104         }
1105         vfio_host_win_add(container, 0, (hwaddr)-1, info.iova_pgsizes);
1106     } else if (ioctl(fd, VFIO_CHECK_EXTENSION, VFIO_SPAPR_TCE_IOMMU) ||
1107                ioctl(fd, VFIO_CHECK_EXTENSION, VFIO_SPAPR_TCE_v2_IOMMU)) {
1108         struct vfio_iommu_spapr_tce_info info;
1109         bool v2 = !!ioctl(fd, VFIO_CHECK_EXTENSION, VFIO_SPAPR_TCE_v2_IOMMU);
1110 
1111         ret = ioctl(group->fd, VFIO_GROUP_SET_CONTAINER, &fd);
1112         if (ret) {
1113             error_setg_errno(errp, errno, "failed to set group container");
1114             ret = -errno;
1115             goto free_container_exit;
1116         }
1117         container->iommu_type =
1118             v2 ? VFIO_SPAPR_TCE_v2_IOMMU : VFIO_SPAPR_TCE_IOMMU;
1119         ret = ioctl(fd, VFIO_SET_IOMMU, container->iommu_type);
1120         if (ret) {
1121             container->iommu_type = VFIO_SPAPR_TCE_IOMMU;
1122             v2 = false;
1123             ret = ioctl(fd, VFIO_SET_IOMMU, container->iommu_type);
1124         }
1125         if (ret) {
1126             error_setg_errno(errp, errno, "failed to set iommu for container");
1127             ret = -errno;
1128             goto free_container_exit;
1129         }
1130 
1131         /*
1132          * The host kernel code implementing VFIO_IOMMU_DISABLE is called
1133          * when container fd is closed so we do not call it explicitly
1134          * in this file.
1135          */
1136         if (!v2) {
1137             ret = ioctl(fd, VFIO_IOMMU_ENABLE);
1138             if (ret) {
1139                 error_setg_errno(errp, errno, "failed to enable container");
1140                 ret = -errno;
1141                 goto free_container_exit;
1142             }
1143         } else {
1144             container->prereg_listener = vfio_prereg_listener;
1145 
1146             memory_listener_register(&container->prereg_listener,
1147                                      &address_space_memory);
1148             if (container->error) {
1149                 memory_listener_unregister(&container->prereg_listener);
1150                 ret = container->error;
1151                 error_setg(errp,
1152                     "RAM memory listener initialization failed for container");
1153                 goto free_container_exit;
1154             }
1155         }
1156 
1157         info.argsz = sizeof(info);
1158         ret = ioctl(fd, VFIO_IOMMU_SPAPR_TCE_GET_INFO, &info);
1159         if (ret) {
1160             error_setg_errno(errp, errno,
1161                              "VFIO_IOMMU_SPAPR_TCE_GET_INFO failed");
1162             ret = -errno;
1163             if (v2) {
1164                 memory_listener_unregister(&container->prereg_listener);
1165             }
1166             goto free_container_exit;
1167         }
1168 
1169         if (v2) {
1170             /*
1171              * There is a default window in just created container.
1172              * To make region_add/del simpler, we better remove this
1173              * window now and let those iommu_listener callbacks
1174              * create/remove them when needed.
1175              */
1176             ret = vfio_spapr_remove_window(container, info.dma32_window_start);
1177             if (ret) {
1178                 error_setg_errno(errp, -ret,
1179                                  "failed to remove existing window");
1180                 goto free_container_exit;
1181             }
1182         } else {
1183             /* The default table uses 4K pages */
1184             vfio_host_win_add(container, info.dma32_window_start,
1185                               info.dma32_window_start +
1186                               info.dma32_window_size - 1,
1187                               0x1000);
1188         }
1189     } else {
1190         error_setg(errp, "No available IOMMU models");
1191         ret = -EINVAL;
1192         goto free_container_exit;
1193     }
1194 
1195     vfio_kvm_device_add_group(group);
1196 
1197     QLIST_INIT(&container->group_list);
1198     QLIST_INSERT_HEAD(&space->containers, container, next);
1199 
1200     group->container = container;
1201     QLIST_INSERT_HEAD(&container->group_list, group, container_next);
1202 
1203     container->listener = vfio_memory_listener;
1204 
1205     memory_listener_register(&container->listener, container->space->as);
1206 
1207     if (container->error) {
1208         ret = container->error;
1209         error_setg_errno(errp, -ret,
1210                          "memory listener initialization failed for container");
1211         goto listener_release_exit;
1212     }
1213 
1214     container->initialized = true;
1215 
1216     return 0;
1217 listener_release_exit:
1218     QLIST_REMOVE(group, container_next);
1219     QLIST_REMOVE(container, next);
1220     vfio_kvm_device_del_group(group);
1221     vfio_listener_release(container);
1222 
1223 free_container_exit:
1224     g_free(container);
1225 
1226 close_fd_exit:
1227     close(fd);
1228 
1229 put_space_exit:
1230     vfio_put_address_space(space);
1231 
1232     return ret;
1233 }
1234 
1235 static void vfio_disconnect_container(VFIOGroup *group)
1236 {
1237     VFIOContainer *container = group->container;
1238 
1239     QLIST_REMOVE(group, container_next);
1240     group->container = NULL;
1241 
1242     /*
1243      * Explicitly release the listener first before unset container,
1244      * since unset may destroy the backend container if it's the last
1245      * group.
1246      */
1247     if (QLIST_EMPTY(&container->group_list)) {
1248         vfio_listener_release(container);
1249     }
1250 
1251     if (ioctl(group->fd, VFIO_GROUP_UNSET_CONTAINER, &container->fd)) {
1252         error_report("vfio: error disconnecting group %d from container",
1253                      group->groupid);
1254     }
1255 
1256     if (QLIST_EMPTY(&container->group_list)) {
1257         VFIOAddressSpace *space = container->space;
1258         VFIOGuestIOMMU *giommu, *tmp;
1259 
1260         QLIST_REMOVE(container, next);
1261 
1262         QLIST_FOREACH_SAFE(giommu, &container->giommu_list, giommu_next, tmp) {
1263             memory_region_unregister_iommu_notifier(
1264                     MEMORY_REGION(giommu->iommu), &giommu->n);
1265             QLIST_REMOVE(giommu, giommu_next);
1266             g_free(giommu);
1267         }
1268 
1269         trace_vfio_disconnect_container(container->fd);
1270         close(container->fd);
1271         g_free(container);
1272 
1273         vfio_put_address_space(space);
1274     }
1275 }
1276 
1277 VFIOGroup *vfio_get_group(int groupid, AddressSpace *as, Error **errp)
1278 {
1279     VFIOGroup *group;
1280     char path[32];
1281     struct vfio_group_status status = { .argsz = sizeof(status) };
1282 
1283     QLIST_FOREACH(group, &vfio_group_list, next) {
1284         if (group->groupid == groupid) {
1285             /* Found it.  Now is it already in the right context? */
1286             if (group->container->space->as == as) {
1287                 return group;
1288             } else {
1289                 error_setg(errp, "group %d used in multiple address spaces",
1290                            group->groupid);
1291                 return NULL;
1292             }
1293         }
1294     }
1295 
1296     group = g_malloc0(sizeof(*group));
1297 
1298     snprintf(path, sizeof(path), "/dev/vfio/%d", groupid);
1299     group->fd = qemu_open(path, O_RDWR);
1300     if (group->fd < 0) {
1301         error_setg_errno(errp, errno, "failed to open %s", path);
1302         goto free_group_exit;
1303     }
1304 
1305     if (ioctl(group->fd, VFIO_GROUP_GET_STATUS, &status)) {
1306         error_setg_errno(errp, errno, "failed to get group %d status", groupid);
1307         goto close_fd_exit;
1308     }
1309 
1310     if (!(status.flags & VFIO_GROUP_FLAGS_VIABLE)) {
1311         error_setg(errp, "group %d is not viable", groupid);
1312         error_append_hint(errp,
1313                           "Please ensure all devices within the iommu_group "
1314                           "are bound to their vfio bus driver.\n");
1315         goto close_fd_exit;
1316     }
1317 
1318     group->groupid = groupid;
1319     QLIST_INIT(&group->device_list);
1320 
1321     if (vfio_connect_container(group, as, errp)) {
1322         error_prepend(errp, "failed to setup container for group %d: ",
1323                       groupid);
1324         goto close_fd_exit;
1325     }
1326 
1327     if (QLIST_EMPTY(&vfio_group_list)) {
1328         qemu_register_reset(vfio_reset_handler, NULL);
1329     }
1330 
1331     QLIST_INSERT_HEAD(&vfio_group_list, group, next);
1332 
1333     return group;
1334 
1335 close_fd_exit:
1336     close(group->fd);
1337 
1338 free_group_exit:
1339     g_free(group);
1340 
1341     return NULL;
1342 }
1343 
1344 void vfio_put_group(VFIOGroup *group)
1345 {
1346     if (!group || !QLIST_EMPTY(&group->device_list)) {
1347         return;
1348     }
1349 
1350     vfio_kvm_device_del_group(group);
1351     vfio_disconnect_container(group);
1352     QLIST_REMOVE(group, next);
1353     trace_vfio_put_group(group->fd);
1354     close(group->fd);
1355     g_free(group);
1356 
1357     if (QLIST_EMPTY(&vfio_group_list)) {
1358         qemu_unregister_reset(vfio_reset_handler, NULL);
1359     }
1360 }
1361 
1362 int vfio_get_device(VFIOGroup *group, const char *name,
1363                     VFIODevice *vbasedev, Error **errp)
1364 {
1365     struct vfio_device_info dev_info = { .argsz = sizeof(dev_info) };
1366     int ret, fd;
1367 
1368     fd = ioctl(group->fd, VFIO_GROUP_GET_DEVICE_FD, name);
1369     if (fd < 0) {
1370         error_setg_errno(errp, errno, "error getting device from group %d",
1371                          group->groupid);
1372         error_append_hint(errp,
1373                       "Verify all devices in group %d are bound to vfio-<bus> "
1374                       "or pci-stub and not already in use\n", group->groupid);
1375         return fd;
1376     }
1377 
1378     ret = ioctl(fd, VFIO_DEVICE_GET_INFO, &dev_info);
1379     if (ret) {
1380         error_setg_errno(errp, errno, "error getting device info");
1381         close(fd);
1382         return ret;
1383     }
1384 
1385     vbasedev->fd = fd;
1386     vbasedev->group = group;
1387     QLIST_INSERT_HEAD(&group->device_list, vbasedev, next);
1388 
1389     vbasedev->num_irqs = dev_info.num_irqs;
1390     vbasedev->num_regions = dev_info.num_regions;
1391     vbasedev->flags = dev_info.flags;
1392 
1393     trace_vfio_get_device(name, dev_info.flags, dev_info.num_regions,
1394                           dev_info.num_irqs);
1395 
1396     vbasedev->reset_works = !!(dev_info.flags & VFIO_DEVICE_FLAGS_RESET);
1397     return 0;
1398 }
1399 
1400 void vfio_put_base_device(VFIODevice *vbasedev)
1401 {
1402     if (!vbasedev->group) {
1403         return;
1404     }
1405     QLIST_REMOVE(vbasedev, next);
1406     vbasedev->group = NULL;
1407     trace_vfio_put_base_device(vbasedev->fd);
1408     close(vbasedev->fd);
1409 }
1410 
1411 int vfio_get_region_info(VFIODevice *vbasedev, int index,
1412                          struct vfio_region_info **info)
1413 {
1414     size_t argsz = sizeof(struct vfio_region_info);
1415 
1416     *info = g_malloc0(argsz);
1417 
1418     (*info)->index = index;
1419 retry:
1420     (*info)->argsz = argsz;
1421 
1422     if (ioctl(vbasedev->fd, VFIO_DEVICE_GET_REGION_INFO, *info)) {
1423         g_free(*info);
1424         *info = NULL;
1425         return -errno;
1426     }
1427 
1428     if ((*info)->argsz > argsz) {
1429         argsz = (*info)->argsz;
1430         *info = g_realloc(*info, argsz);
1431 
1432         goto retry;
1433     }
1434 
1435     return 0;
1436 }
1437 
1438 int vfio_get_dev_region_info(VFIODevice *vbasedev, uint32_t type,
1439                              uint32_t subtype, struct vfio_region_info **info)
1440 {
1441     int i;
1442 
1443     for (i = 0; i < vbasedev->num_regions; i++) {
1444         struct vfio_info_cap_header *hdr;
1445         struct vfio_region_info_cap_type *cap_type;
1446 
1447         if (vfio_get_region_info(vbasedev, i, info)) {
1448             continue;
1449         }
1450 
1451         hdr = vfio_get_region_info_cap(*info, VFIO_REGION_INFO_CAP_TYPE);
1452         if (!hdr) {
1453             g_free(*info);
1454             continue;
1455         }
1456 
1457         cap_type = container_of(hdr, struct vfio_region_info_cap_type, header);
1458 
1459         trace_vfio_get_dev_region(vbasedev->name, i,
1460                                   cap_type->type, cap_type->subtype);
1461 
1462         if (cap_type->type == type && cap_type->subtype == subtype) {
1463             return 0;
1464         }
1465 
1466         g_free(*info);
1467     }
1468 
1469     *info = NULL;
1470     return -ENODEV;
1471 }
1472 
1473 bool vfio_has_region_cap(VFIODevice *vbasedev, int region, uint16_t cap_type)
1474 {
1475     struct vfio_region_info *info = NULL;
1476     bool ret = false;
1477 
1478     if (!vfio_get_region_info(vbasedev, region, &info)) {
1479         if (vfio_get_region_info_cap(info, cap_type)) {
1480             ret = true;
1481         }
1482         g_free(info);
1483     }
1484 
1485     return ret;
1486 }
1487 
1488 /*
1489  * Interfaces for IBM EEH (Enhanced Error Handling)
1490  */
1491 static bool vfio_eeh_container_ok(VFIOContainer *container)
1492 {
1493     /*
1494      * As of 2016-03-04 (linux-4.5) the host kernel EEH/VFIO
1495      * implementation is broken if there are multiple groups in a
1496      * container.  The hardware works in units of Partitionable
1497      * Endpoints (== IOMMU groups) and the EEH operations naively
1498      * iterate across all groups in the container, without any logic
1499      * to make sure the groups have their state synchronized.  For
1500      * certain operations (ENABLE) that might be ok, until an error
1501      * occurs, but for others (GET_STATE) it's clearly broken.
1502      */
1503 
1504     /*
1505      * XXX Once fixed kernels exist, test for them here
1506      */
1507 
1508     if (QLIST_EMPTY(&container->group_list)) {
1509         return false;
1510     }
1511 
1512     if (QLIST_NEXT(QLIST_FIRST(&container->group_list), container_next)) {
1513         return false;
1514     }
1515 
1516     return true;
1517 }
1518 
1519 static int vfio_eeh_container_op(VFIOContainer *container, uint32_t op)
1520 {
1521     struct vfio_eeh_pe_op pe_op = {
1522         .argsz = sizeof(pe_op),
1523         .op = op,
1524     };
1525     int ret;
1526 
1527     if (!vfio_eeh_container_ok(container)) {
1528         error_report("vfio/eeh: EEH_PE_OP 0x%x: "
1529                      "kernel requires a container with exactly one group", op);
1530         return -EPERM;
1531     }
1532 
1533     ret = ioctl(container->fd, VFIO_EEH_PE_OP, &pe_op);
1534     if (ret < 0) {
1535         error_report("vfio/eeh: EEH_PE_OP 0x%x failed: %m", op);
1536         return -errno;
1537     }
1538 
1539     return ret;
1540 }
1541 
1542 static VFIOContainer *vfio_eeh_as_container(AddressSpace *as)
1543 {
1544     VFIOAddressSpace *space = vfio_get_address_space(as);
1545     VFIOContainer *container = NULL;
1546 
1547     if (QLIST_EMPTY(&space->containers)) {
1548         /* No containers to act on */
1549         goto out;
1550     }
1551 
1552     container = QLIST_FIRST(&space->containers);
1553 
1554     if (QLIST_NEXT(container, next)) {
1555         /* We don't yet have logic to synchronize EEH state across
1556          * multiple containers */
1557         container = NULL;
1558         goto out;
1559     }
1560 
1561 out:
1562     vfio_put_address_space(space);
1563     return container;
1564 }
1565 
1566 bool vfio_eeh_as_ok(AddressSpace *as)
1567 {
1568     VFIOContainer *container = vfio_eeh_as_container(as);
1569 
1570     return (container != NULL) && vfio_eeh_container_ok(container);
1571 }
1572 
1573 int vfio_eeh_as_op(AddressSpace *as, uint32_t op)
1574 {
1575     VFIOContainer *container = vfio_eeh_as_container(as);
1576 
1577     if (!container) {
1578         return -ENODEV;
1579     }
1580     return vfio_eeh_container_op(container, op);
1581 }
1582