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 MEMTXATTRS_UNSPECIFIED); 329 if (!memory_region_is_ram(mr)) { 330 error_report("iommu map to non memory area %"HWADDR_PRIx"", 331 xlat); 332 return false; 333 } 334 335 /* 336 * Translation truncates length to the IOMMU page size, 337 * check that it did not truncate too much. 338 */ 339 if (len & iotlb->addr_mask) { 340 error_report("iommu has granularity incompatible with target AS"); 341 return false; 342 } 343 344 *vaddr = memory_region_get_ram_ptr(mr) + xlat; 345 *read_only = !writable || mr->readonly; 346 347 return true; 348 } 349 350 static void vfio_iommu_map_notify(IOMMUNotifier *n, IOMMUTLBEntry *iotlb) 351 { 352 VFIOGuestIOMMU *giommu = container_of(n, VFIOGuestIOMMU, n); 353 VFIOContainer *container = giommu->container; 354 hwaddr iova = iotlb->iova + giommu->iommu_offset; 355 bool read_only; 356 void *vaddr; 357 int ret; 358 359 trace_vfio_iommu_map_notify(iotlb->perm == IOMMU_NONE ? "UNMAP" : "MAP", 360 iova, iova + iotlb->addr_mask); 361 362 if (iotlb->target_as != &address_space_memory) { 363 error_report("Wrong target AS \"%s\", only system memory is allowed", 364 iotlb->target_as->name ? iotlb->target_as->name : "none"); 365 return; 366 } 367 368 rcu_read_lock(); 369 370 if ((iotlb->perm & IOMMU_RW) != IOMMU_NONE) { 371 if (!vfio_get_vaddr(iotlb, &vaddr, &read_only)) { 372 goto out; 373 } 374 /* 375 * vaddr is only valid until rcu_read_unlock(). But after 376 * vfio_dma_map has set up the mapping the pages will be 377 * pinned by the kernel. This makes sure that the RAM backend 378 * of vaddr will always be there, even if the memory object is 379 * destroyed and its backing memory munmap-ed. 380 */ 381 ret = vfio_dma_map(container, iova, 382 iotlb->addr_mask + 1, vaddr, 383 read_only); 384 if (ret) { 385 error_report("vfio_dma_map(%p, 0x%"HWADDR_PRIx", " 386 "0x%"HWADDR_PRIx", %p) = %d (%m)", 387 container, iova, 388 iotlb->addr_mask + 1, vaddr, ret); 389 } 390 } else { 391 ret = vfio_dma_unmap(container, iova, iotlb->addr_mask + 1); 392 if (ret) { 393 error_report("vfio_dma_unmap(%p, 0x%"HWADDR_PRIx", " 394 "0x%"HWADDR_PRIx") = %d (%m)", 395 container, iova, 396 iotlb->addr_mask + 1, ret); 397 } 398 } 399 out: 400 rcu_read_unlock(); 401 } 402 403 static void vfio_listener_region_add(MemoryListener *listener, 404 MemoryRegionSection *section) 405 { 406 VFIOContainer *container = container_of(listener, VFIOContainer, listener); 407 hwaddr iova, end; 408 Int128 llend, llsize; 409 void *vaddr; 410 int ret; 411 VFIOHostDMAWindow *hostwin; 412 bool hostwin_found; 413 414 if (vfio_listener_skipped_section(section)) { 415 trace_vfio_listener_region_add_skip( 416 section->offset_within_address_space, 417 section->offset_within_address_space + 418 int128_get64(int128_sub(section->size, int128_one()))); 419 return; 420 } 421 422 if (unlikely((section->offset_within_address_space & ~TARGET_PAGE_MASK) != 423 (section->offset_within_region & ~TARGET_PAGE_MASK))) { 424 error_report("%s received unaligned region", __func__); 425 return; 426 } 427 428 iova = TARGET_PAGE_ALIGN(section->offset_within_address_space); 429 llend = int128_make64(section->offset_within_address_space); 430 llend = int128_add(llend, section->size); 431 llend = int128_and(llend, int128_exts64(TARGET_PAGE_MASK)); 432 433 if (int128_ge(int128_make64(iova), llend)) { 434 return; 435 } 436 end = int128_get64(int128_sub(llend, int128_one())); 437 438 if (container->iommu_type == VFIO_SPAPR_TCE_v2_IOMMU) { 439 hwaddr pgsize = 0; 440 441 /* For now intersections are not allowed, we may relax this later */ 442 QLIST_FOREACH(hostwin, &container->hostwin_list, hostwin_next) { 443 if (ranges_overlap(hostwin->min_iova, 444 hostwin->max_iova - hostwin->min_iova + 1, 445 section->offset_within_address_space, 446 int128_get64(section->size))) { 447 ret = -1; 448 goto fail; 449 } 450 } 451 452 ret = vfio_spapr_create_window(container, section, &pgsize); 453 if (ret) { 454 goto fail; 455 } 456 457 vfio_host_win_add(container, section->offset_within_address_space, 458 section->offset_within_address_space + 459 int128_get64(section->size) - 1, pgsize); 460 #ifdef CONFIG_KVM 461 if (kvm_enabled()) { 462 VFIOGroup *group; 463 IOMMUMemoryRegion *iommu_mr = IOMMU_MEMORY_REGION(section->mr); 464 struct kvm_vfio_spapr_tce param; 465 struct kvm_device_attr attr = { 466 .group = KVM_DEV_VFIO_GROUP, 467 .attr = KVM_DEV_VFIO_GROUP_SET_SPAPR_TCE, 468 .addr = (uint64_t)(unsigned long)¶m, 469 }; 470 471 if (!memory_region_iommu_get_attr(iommu_mr, IOMMU_ATTR_SPAPR_TCE_FD, 472 ¶m.tablefd)) { 473 QLIST_FOREACH(group, &container->group_list, container_next) { 474 param.groupfd = group->fd; 475 if (ioctl(vfio_kvm_device_fd, KVM_SET_DEVICE_ATTR, &attr)) { 476 error_report("vfio: failed to setup fd %d " 477 "for a group with fd %d: %s", 478 param.tablefd, param.groupfd, 479 strerror(errno)); 480 return; 481 } 482 trace_vfio_spapr_group_attach(param.groupfd, param.tablefd); 483 } 484 } 485 } 486 #endif 487 } 488 489 hostwin_found = false; 490 QLIST_FOREACH(hostwin, &container->hostwin_list, hostwin_next) { 491 if (hostwin->min_iova <= iova && end <= hostwin->max_iova) { 492 hostwin_found = true; 493 break; 494 } 495 } 496 497 if (!hostwin_found) { 498 error_report("vfio: IOMMU container %p can't map guest IOVA region" 499 " 0x%"HWADDR_PRIx"..0x%"HWADDR_PRIx, 500 container, iova, end); 501 ret = -EFAULT; 502 goto fail; 503 } 504 505 memory_region_ref(section->mr); 506 507 if (memory_region_is_iommu(section->mr)) { 508 VFIOGuestIOMMU *giommu; 509 IOMMUMemoryRegion *iommu_mr = IOMMU_MEMORY_REGION(section->mr); 510 int iommu_idx; 511 512 trace_vfio_listener_region_add_iommu(iova, end); 513 /* 514 * FIXME: For VFIO iommu types which have KVM acceleration to 515 * avoid bouncing all map/unmaps through qemu this way, this 516 * would be the right place to wire that up (tell the KVM 517 * device emulation the VFIO iommu handles to use). 518 */ 519 giommu = g_malloc0(sizeof(*giommu)); 520 giommu->iommu = iommu_mr; 521 giommu->iommu_offset = section->offset_within_address_space - 522 section->offset_within_region; 523 giommu->container = container; 524 llend = int128_add(int128_make64(section->offset_within_region), 525 section->size); 526 llend = int128_sub(llend, int128_one()); 527 iommu_idx = memory_region_iommu_attrs_to_index(iommu_mr, 528 MEMTXATTRS_UNSPECIFIED); 529 iommu_notifier_init(&giommu->n, vfio_iommu_map_notify, 530 IOMMU_NOTIFIER_ALL, 531 section->offset_within_region, 532 int128_get64(llend), 533 iommu_idx); 534 QLIST_INSERT_HEAD(&container->giommu_list, giommu, giommu_next); 535 536 memory_region_register_iommu_notifier(section->mr, &giommu->n); 537 memory_region_iommu_replay(giommu->iommu, &giommu->n); 538 539 return; 540 } 541 542 /* Here we assume that memory_region_is_ram(section->mr)==true */ 543 544 vaddr = memory_region_get_ram_ptr(section->mr) + 545 section->offset_within_region + 546 (iova - section->offset_within_address_space); 547 548 trace_vfio_listener_region_add_ram(iova, end, vaddr); 549 550 llsize = int128_sub(llend, int128_make64(iova)); 551 552 if (memory_region_is_ram_device(section->mr)) { 553 hwaddr pgmask = (1ULL << ctz64(hostwin->iova_pgsizes)) - 1; 554 555 if ((iova & pgmask) || (int128_get64(llsize) & pgmask)) { 556 trace_vfio_listener_region_add_no_dma_map( 557 memory_region_name(section->mr), 558 section->offset_within_address_space, 559 int128_getlo(section->size), 560 pgmask + 1); 561 return; 562 } 563 } 564 565 ret = vfio_dma_map(container, iova, int128_get64(llsize), 566 vaddr, section->readonly); 567 if (ret) { 568 error_report("vfio_dma_map(%p, 0x%"HWADDR_PRIx", " 569 "0x%"HWADDR_PRIx", %p) = %d (%m)", 570 container, iova, int128_get64(llsize), vaddr, ret); 571 if (memory_region_is_ram_device(section->mr)) { 572 /* Allow unexpected mappings not to be fatal for RAM devices */ 573 return; 574 } 575 goto fail; 576 } 577 578 return; 579 580 fail: 581 if (memory_region_is_ram_device(section->mr)) { 582 error_report("failed to vfio_dma_map. pci p2p may not work"); 583 return; 584 } 585 /* 586 * On the initfn path, store the first error in the container so we 587 * can gracefully fail. Runtime, there's not much we can do other 588 * than throw a hardware error. 589 */ 590 if (!container->initialized) { 591 if (!container->error) { 592 container->error = ret; 593 } 594 } else { 595 hw_error("vfio: DMA mapping failed, unable to continue"); 596 } 597 } 598 599 static void vfio_listener_region_del(MemoryListener *listener, 600 MemoryRegionSection *section) 601 { 602 VFIOContainer *container = container_of(listener, VFIOContainer, listener); 603 hwaddr iova, end; 604 Int128 llend, llsize; 605 int ret; 606 bool try_unmap = true; 607 608 if (vfio_listener_skipped_section(section)) { 609 trace_vfio_listener_region_del_skip( 610 section->offset_within_address_space, 611 section->offset_within_address_space + 612 int128_get64(int128_sub(section->size, int128_one()))); 613 return; 614 } 615 616 if (unlikely((section->offset_within_address_space & ~TARGET_PAGE_MASK) != 617 (section->offset_within_region & ~TARGET_PAGE_MASK))) { 618 error_report("%s received unaligned region", __func__); 619 return; 620 } 621 622 if (memory_region_is_iommu(section->mr)) { 623 VFIOGuestIOMMU *giommu; 624 625 QLIST_FOREACH(giommu, &container->giommu_list, giommu_next) { 626 if (MEMORY_REGION(giommu->iommu) == section->mr && 627 giommu->n.start == section->offset_within_region) { 628 memory_region_unregister_iommu_notifier(section->mr, 629 &giommu->n); 630 QLIST_REMOVE(giommu, giommu_next); 631 g_free(giommu); 632 break; 633 } 634 } 635 636 /* 637 * FIXME: We assume the one big unmap below is adequate to 638 * remove any individual page mappings in the IOMMU which 639 * might have been copied into VFIO. This works for a page table 640 * based IOMMU where a big unmap flattens a large range of IO-PTEs. 641 * That may not be true for all IOMMU types. 642 */ 643 } 644 645 iova = TARGET_PAGE_ALIGN(section->offset_within_address_space); 646 llend = int128_make64(section->offset_within_address_space); 647 llend = int128_add(llend, section->size); 648 llend = int128_and(llend, int128_exts64(TARGET_PAGE_MASK)); 649 650 if (int128_ge(int128_make64(iova), llend)) { 651 return; 652 } 653 end = int128_get64(int128_sub(llend, int128_one())); 654 655 llsize = int128_sub(llend, int128_make64(iova)); 656 657 trace_vfio_listener_region_del(iova, end); 658 659 if (memory_region_is_ram_device(section->mr)) { 660 hwaddr pgmask; 661 VFIOHostDMAWindow *hostwin; 662 bool hostwin_found = false; 663 664 QLIST_FOREACH(hostwin, &container->hostwin_list, hostwin_next) { 665 if (hostwin->min_iova <= iova && end <= hostwin->max_iova) { 666 hostwin_found = true; 667 break; 668 } 669 } 670 assert(hostwin_found); /* or region_add() would have failed */ 671 672 pgmask = (1ULL << ctz64(hostwin->iova_pgsizes)) - 1; 673 try_unmap = !((iova & pgmask) || (int128_get64(llsize) & pgmask)); 674 } 675 676 if (try_unmap) { 677 ret = vfio_dma_unmap(container, iova, int128_get64(llsize)); 678 if (ret) { 679 error_report("vfio_dma_unmap(%p, 0x%"HWADDR_PRIx", " 680 "0x%"HWADDR_PRIx") = %d (%m)", 681 container, iova, int128_get64(llsize), ret); 682 } 683 } 684 685 memory_region_unref(section->mr); 686 687 if (container->iommu_type == VFIO_SPAPR_TCE_v2_IOMMU) { 688 vfio_spapr_remove_window(container, 689 section->offset_within_address_space); 690 if (vfio_host_win_del(container, 691 section->offset_within_address_space, 692 section->offset_within_address_space + 693 int128_get64(section->size) - 1) < 0) { 694 hw_error("%s: Cannot delete missing window at %"HWADDR_PRIx, 695 __func__, section->offset_within_address_space); 696 } 697 } 698 } 699 700 static const MemoryListener vfio_memory_listener = { 701 .region_add = vfio_listener_region_add, 702 .region_del = vfio_listener_region_del, 703 }; 704 705 static void vfio_listener_release(VFIOContainer *container) 706 { 707 memory_listener_unregister(&container->listener); 708 if (container->iommu_type == VFIO_SPAPR_TCE_v2_IOMMU) { 709 memory_listener_unregister(&container->prereg_listener); 710 } 711 } 712 713 static struct vfio_info_cap_header * 714 vfio_get_region_info_cap(struct vfio_region_info *info, uint16_t id) 715 { 716 struct vfio_info_cap_header *hdr; 717 void *ptr = info; 718 719 if (!(info->flags & VFIO_REGION_INFO_FLAG_CAPS)) { 720 return NULL; 721 } 722 723 for (hdr = ptr + info->cap_offset; hdr != ptr; hdr = ptr + hdr->next) { 724 if (hdr->id == id) { 725 return hdr; 726 } 727 } 728 729 return NULL; 730 } 731 732 static int vfio_setup_region_sparse_mmaps(VFIORegion *region, 733 struct vfio_region_info *info) 734 { 735 struct vfio_info_cap_header *hdr; 736 struct vfio_region_info_cap_sparse_mmap *sparse; 737 int i, j; 738 739 hdr = vfio_get_region_info_cap(info, VFIO_REGION_INFO_CAP_SPARSE_MMAP); 740 if (!hdr) { 741 return -ENODEV; 742 } 743 744 sparse = container_of(hdr, struct vfio_region_info_cap_sparse_mmap, header); 745 746 trace_vfio_region_sparse_mmap_header(region->vbasedev->name, 747 region->nr, sparse->nr_areas); 748 749 region->mmaps = g_new0(VFIOMmap, sparse->nr_areas); 750 751 for (i = 0, j = 0; i < sparse->nr_areas; i++) { 752 trace_vfio_region_sparse_mmap_entry(i, sparse->areas[i].offset, 753 sparse->areas[i].offset + 754 sparse->areas[i].size); 755 756 if (sparse->areas[i].size) { 757 region->mmaps[j].offset = sparse->areas[i].offset; 758 region->mmaps[j].size = sparse->areas[i].size; 759 j++; 760 } 761 } 762 763 region->nr_mmaps = j; 764 region->mmaps = g_realloc(region->mmaps, j * sizeof(VFIOMmap)); 765 766 return 0; 767 } 768 769 int vfio_region_setup(Object *obj, VFIODevice *vbasedev, VFIORegion *region, 770 int index, const char *name) 771 { 772 struct vfio_region_info *info; 773 int ret; 774 775 ret = vfio_get_region_info(vbasedev, index, &info); 776 if (ret) { 777 return ret; 778 } 779 780 region->vbasedev = vbasedev; 781 region->flags = info->flags; 782 region->size = info->size; 783 region->fd_offset = info->offset; 784 region->nr = index; 785 786 if (region->size) { 787 region->mem = g_new0(MemoryRegion, 1); 788 memory_region_init_io(region->mem, obj, &vfio_region_ops, 789 region, name, region->size); 790 791 if (!vbasedev->no_mmap && 792 region->flags & VFIO_REGION_INFO_FLAG_MMAP) { 793 794 ret = vfio_setup_region_sparse_mmaps(region, info); 795 796 if (ret) { 797 region->nr_mmaps = 1; 798 region->mmaps = g_new0(VFIOMmap, region->nr_mmaps); 799 region->mmaps[0].offset = 0; 800 region->mmaps[0].size = region->size; 801 } 802 } 803 } 804 805 g_free(info); 806 807 trace_vfio_region_setup(vbasedev->name, index, name, 808 region->flags, region->fd_offset, region->size); 809 return 0; 810 } 811 812 int vfio_region_mmap(VFIORegion *region) 813 { 814 int i, prot = 0; 815 char *name; 816 817 if (!region->mem) { 818 return 0; 819 } 820 821 prot |= region->flags & VFIO_REGION_INFO_FLAG_READ ? PROT_READ : 0; 822 prot |= region->flags & VFIO_REGION_INFO_FLAG_WRITE ? PROT_WRITE : 0; 823 824 for (i = 0; i < region->nr_mmaps; i++) { 825 region->mmaps[i].mmap = mmap(NULL, region->mmaps[i].size, prot, 826 MAP_SHARED, region->vbasedev->fd, 827 region->fd_offset + 828 region->mmaps[i].offset); 829 if (region->mmaps[i].mmap == MAP_FAILED) { 830 int ret = -errno; 831 832 trace_vfio_region_mmap_fault(memory_region_name(region->mem), i, 833 region->fd_offset + 834 region->mmaps[i].offset, 835 region->fd_offset + 836 region->mmaps[i].offset + 837 region->mmaps[i].size - 1, ret); 838 839 region->mmaps[i].mmap = NULL; 840 841 for (i--; i >= 0; i--) { 842 memory_region_del_subregion(region->mem, ®ion->mmaps[i].mem); 843 munmap(region->mmaps[i].mmap, region->mmaps[i].size); 844 object_unparent(OBJECT(®ion->mmaps[i].mem)); 845 region->mmaps[i].mmap = NULL; 846 } 847 848 return ret; 849 } 850 851 name = g_strdup_printf("%s mmaps[%d]", 852 memory_region_name(region->mem), i); 853 memory_region_init_ram_device_ptr(®ion->mmaps[i].mem, 854 memory_region_owner(region->mem), 855 name, region->mmaps[i].size, 856 region->mmaps[i].mmap); 857 g_free(name); 858 memory_region_add_subregion(region->mem, region->mmaps[i].offset, 859 ®ion->mmaps[i].mem); 860 861 trace_vfio_region_mmap(memory_region_name(®ion->mmaps[i].mem), 862 region->mmaps[i].offset, 863 region->mmaps[i].offset + 864 region->mmaps[i].size - 1); 865 } 866 867 return 0; 868 } 869 870 void vfio_region_exit(VFIORegion *region) 871 { 872 int i; 873 874 if (!region->mem) { 875 return; 876 } 877 878 for (i = 0; i < region->nr_mmaps; i++) { 879 if (region->mmaps[i].mmap) { 880 memory_region_del_subregion(region->mem, ®ion->mmaps[i].mem); 881 } 882 } 883 884 trace_vfio_region_exit(region->vbasedev->name, region->nr); 885 } 886 887 void vfio_region_finalize(VFIORegion *region) 888 { 889 int i; 890 891 if (!region->mem) { 892 return; 893 } 894 895 for (i = 0; i < region->nr_mmaps; i++) { 896 if (region->mmaps[i].mmap) { 897 munmap(region->mmaps[i].mmap, region->mmaps[i].size); 898 object_unparent(OBJECT(®ion->mmaps[i].mem)); 899 } 900 } 901 902 object_unparent(OBJECT(region->mem)); 903 904 g_free(region->mem); 905 g_free(region->mmaps); 906 907 trace_vfio_region_finalize(region->vbasedev->name, region->nr); 908 909 region->mem = NULL; 910 region->mmaps = NULL; 911 region->nr_mmaps = 0; 912 region->size = 0; 913 region->flags = 0; 914 region->nr = 0; 915 } 916 917 void vfio_region_mmaps_set_enabled(VFIORegion *region, bool enabled) 918 { 919 int i; 920 921 if (!region->mem) { 922 return; 923 } 924 925 for (i = 0; i < region->nr_mmaps; i++) { 926 if (region->mmaps[i].mmap) { 927 memory_region_set_enabled(®ion->mmaps[i].mem, enabled); 928 } 929 } 930 931 trace_vfio_region_mmaps_set_enabled(memory_region_name(region->mem), 932 enabled); 933 } 934 935 void vfio_reset_handler(void *opaque) 936 { 937 VFIOGroup *group; 938 VFIODevice *vbasedev; 939 940 QLIST_FOREACH(group, &vfio_group_list, next) { 941 QLIST_FOREACH(vbasedev, &group->device_list, next) { 942 if (vbasedev->dev->realized) { 943 vbasedev->ops->vfio_compute_needs_reset(vbasedev); 944 } 945 } 946 } 947 948 QLIST_FOREACH(group, &vfio_group_list, next) { 949 QLIST_FOREACH(vbasedev, &group->device_list, next) { 950 if (vbasedev->dev->realized && vbasedev->needs_reset) { 951 vbasedev->ops->vfio_hot_reset_multi(vbasedev); 952 } 953 } 954 } 955 } 956 957 static void vfio_kvm_device_add_group(VFIOGroup *group) 958 { 959 #ifdef CONFIG_KVM 960 struct kvm_device_attr attr = { 961 .group = KVM_DEV_VFIO_GROUP, 962 .attr = KVM_DEV_VFIO_GROUP_ADD, 963 .addr = (uint64_t)(unsigned long)&group->fd, 964 }; 965 966 if (!kvm_enabled()) { 967 return; 968 } 969 970 if (vfio_kvm_device_fd < 0) { 971 struct kvm_create_device cd = { 972 .type = KVM_DEV_TYPE_VFIO, 973 }; 974 975 if (kvm_vm_ioctl(kvm_state, KVM_CREATE_DEVICE, &cd)) { 976 error_report("Failed to create KVM VFIO device: %m"); 977 return; 978 } 979 980 vfio_kvm_device_fd = cd.fd; 981 } 982 983 if (ioctl(vfio_kvm_device_fd, KVM_SET_DEVICE_ATTR, &attr)) { 984 error_report("Failed to add group %d to KVM VFIO device: %m", 985 group->groupid); 986 } 987 #endif 988 } 989 990 static void vfio_kvm_device_del_group(VFIOGroup *group) 991 { 992 #ifdef CONFIG_KVM 993 struct kvm_device_attr attr = { 994 .group = KVM_DEV_VFIO_GROUP, 995 .attr = KVM_DEV_VFIO_GROUP_DEL, 996 .addr = (uint64_t)(unsigned long)&group->fd, 997 }; 998 999 if (vfio_kvm_device_fd < 0) { 1000 return; 1001 } 1002 1003 if (ioctl(vfio_kvm_device_fd, KVM_SET_DEVICE_ATTR, &attr)) { 1004 error_report("Failed to remove group %d from KVM VFIO device: %m", 1005 group->groupid); 1006 } 1007 #endif 1008 } 1009 1010 static VFIOAddressSpace *vfio_get_address_space(AddressSpace *as) 1011 { 1012 VFIOAddressSpace *space; 1013 1014 QLIST_FOREACH(space, &vfio_address_spaces, list) { 1015 if (space->as == as) { 1016 return space; 1017 } 1018 } 1019 1020 /* No suitable VFIOAddressSpace, create a new one */ 1021 space = g_malloc0(sizeof(*space)); 1022 space->as = as; 1023 QLIST_INIT(&space->containers); 1024 1025 QLIST_INSERT_HEAD(&vfio_address_spaces, space, list); 1026 1027 return space; 1028 } 1029 1030 static void vfio_put_address_space(VFIOAddressSpace *space) 1031 { 1032 if (QLIST_EMPTY(&space->containers)) { 1033 QLIST_REMOVE(space, list); 1034 g_free(space); 1035 } 1036 } 1037 1038 static int vfio_connect_container(VFIOGroup *group, AddressSpace *as, 1039 Error **errp) 1040 { 1041 VFIOContainer *container; 1042 int ret, fd; 1043 VFIOAddressSpace *space; 1044 1045 space = vfio_get_address_space(as); 1046 1047 QLIST_FOREACH(container, &space->containers, next) { 1048 if (!ioctl(group->fd, VFIO_GROUP_SET_CONTAINER, &container->fd)) { 1049 group->container = container; 1050 QLIST_INSERT_HEAD(&container->group_list, group, container_next); 1051 vfio_kvm_device_add_group(group); 1052 return 0; 1053 } 1054 } 1055 1056 fd = qemu_open("/dev/vfio/vfio", O_RDWR); 1057 if (fd < 0) { 1058 error_setg_errno(errp, errno, "failed to open /dev/vfio/vfio"); 1059 ret = -errno; 1060 goto put_space_exit; 1061 } 1062 1063 ret = ioctl(fd, VFIO_GET_API_VERSION); 1064 if (ret != VFIO_API_VERSION) { 1065 error_setg(errp, "supported vfio version: %d, " 1066 "reported version: %d", VFIO_API_VERSION, ret); 1067 ret = -EINVAL; 1068 goto close_fd_exit; 1069 } 1070 1071 container = g_malloc0(sizeof(*container)); 1072 container->space = space; 1073 container->fd = fd; 1074 QLIST_INIT(&container->giommu_list); 1075 QLIST_INIT(&container->hostwin_list); 1076 if (ioctl(fd, VFIO_CHECK_EXTENSION, VFIO_TYPE1_IOMMU) || 1077 ioctl(fd, VFIO_CHECK_EXTENSION, VFIO_TYPE1v2_IOMMU)) { 1078 bool v2 = !!ioctl(fd, VFIO_CHECK_EXTENSION, VFIO_TYPE1v2_IOMMU); 1079 struct vfio_iommu_type1_info info; 1080 1081 ret = ioctl(group->fd, VFIO_GROUP_SET_CONTAINER, &fd); 1082 if (ret) { 1083 error_setg_errno(errp, errno, "failed to set group container"); 1084 ret = -errno; 1085 goto free_container_exit; 1086 } 1087 1088 container->iommu_type = v2 ? VFIO_TYPE1v2_IOMMU : VFIO_TYPE1_IOMMU; 1089 ret = ioctl(fd, VFIO_SET_IOMMU, container->iommu_type); 1090 if (ret) { 1091 error_setg_errno(errp, errno, "failed to set iommu for container"); 1092 ret = -errno; 1093 goto free_container_exit; 1094 } 1095 1096 /* 1097 * FIXME: This assumes that a Type1 IOMMU can map any 64-bit 1098 * IOVA whatsoever. That's not actually true, but the current 1099 * kernel interface doesn't tell us what it can map, and the 1100 * existing Type1 IOMMUs generally support any IOVA we're 1101 * going to actually try in practice. 1102 */ 1103 info.argsz = sizeof(info); 1104 ret = ioctl(fd, VFIO_IOMMU_GET_INFO, &info); 1105 /* Ignore errors */ 1106 if (ret || !(info.flags & VFIO_IOMMU_INFO_PGSIZES)) { 1107 /* Assume 4k IOVA page size */ 1108 info.iova_pgsizes = 4096; 1109 } 1110 vfio_host_win_add(container, 0, (hwaddr)-1, info.iova_pgsizes); 1111 } else if (ioctl(fd, VFIO_CHECK_EXTENSION, VFIO_SPAPR_TCE_IOMMU) || 1112 ioctl(fd, VFIO_CHECK_EXTENSION, VFIO_SPAPR_TCE_v2_IOMMU)) { 1113 struct vfio_iommu_spapr_tce_info info; 1114 bool v2 = !!ioctl(fd, VFIO_CHECK_EXTENSION, VFIO_SPAPR_TCE_v2_IOMMU); 1115 1116 ret = ioctl(group->fd, VFIO_GROUP_SET_CONTAINER, &fd); 1117 if (ret) { 1118 error_setg_errno(errp, errno, "failed to set group container"); 1119 ret = -errno; 1120 goto free_container_exit; 1121 } 1122 container->iommu_type = 1123 v2 ? VFIO_SPAPR_TCE_v2_IOMMU : VFIO_SPAPR_TCE_IOMMU; 1124 ret = ioctl(fd, VFIO_SET_IOMMU, container->iommu_type); 1125 if (ret) { 1126 container->iommu_type = VFIO_SPAPR_TCE_IOMMU; 1127 v2 = false; 1128 ret = ioctl(fd, VFIO_SET_IOMMU, container->iommu_type); 1129 } 1130 if (ret) { 1131 error_setg_errno(errp, errno, "failed to set iommu for container"); 1132 ret = -errno; 1133 goto free_container_exit; 1134 } 1135 1136 /* 1137 * The host kernel code implementing VFIO_IOMMU_DISABLE is called 1138 * when container fd is closed so we do not call it explicitly 1139 * in this file. 1140 */ 1141 if (!v2) { 1142 ret = ioctl(fd, VFIO_IOMMU_ENABLE); 1143 if (ret) { 1144 error_setg_errno(errp, errno, "failed to enable container"); 1145 ret = -errno; 1146 goto free_container_exit; 1147 } 1148 } else { 1149 container->prereg_listener = vfio_prereg_listener; 1150 1151 memory_listener_register(&container->prereg_listener, 1152 &address_space_memory); 1153 if (container->error) { 1154 memory_listener_unregister(&container->prereg_listener); 1155 ret = container->error; 1156 error_setg(errp, 1157 "RAM memory listener initialization failed for container"); 1158 goto free_container_exit; 1159 } 1160 } 1161 1162 info.argsz = sizeof(info); 1163 ret = ioctl(fd, VFIO_IOMMU_SPAPR_TCE_GET_INFO, &info); 1164 if (ret) { 1165 error_setg_errno(errp, errno, 1166 "VFIO_IOMMU_SPAPR_TCE_GET_INFO failed"); 1167 ret = -errno; 1168 if (v2) { 1169 memory_listener_unregister(&container->prereg_listener); 1170 } 1171 goto free_container_exit; 1172 } 1173 1174 if (v2) { 1175 /* 1176 * There is a default window in just created container. 1177 * To make region_add/del simpler, we better remove this 1178 * window now and let those iommu_listener callbacks 1179 * create/remove them when needed. 1180 */ 1181 ret = vfio_spapr_remove_window(container, info.dma32_window_start); 1182 if (ret) { 1183 error_setg_errno(errp, -ret, 1184 "failed to remove existing window"); 1185 goto free_container_exit; 1186 } 1187 } else { 1188 /* The default table uses 4K pages */ 1189 vfio_host_win_add(container, info.dma32_window_start, 1190 info.dma32_window_start + 1191 info.dma32_window_size - 1, 1192 0x1000); 1193 } 1194 } else { 1195 error_setg(errp, "No available IOMMU models"); 1196 ret = -EINVAL; 1197 goto free_container_exit; 1198 } 1199 1200 vfio_kvm_device_add_group(group); 1201 1202 QLIST_INIT(&container->group_list); 1203 QLIST_INSERT_HEAD(&space->containers, container, next); 1204 1205 group->container = container; 1206 QLIST_INSERT_HEAD(&container->group_list, group, container_next); 1207 1208 container->listener = vfio_memory_listener; 1209 1210 memory_listener_register(&container->listener, container->space->as); 1211 1212 if (container->error) { 1213 ret = container->error; 1214 error_setg_errno(errp, -ret, 1215 "memory listener initialization failed for container"); 1216 goto listener_release_exit; 1217 } 1218 1219 container->initialized = true; 1220 1221 return 0; 1222 listener_release_exit: 1223 QLIST_REMOVE(group, container_next); 1224 QLIST_REMOVE(container, next); 1225 vfio_kvm_device_del_group(group); 1226 vfio_listener_release(container); 1227 1228 free_container_exit: 1229 g_free(container); 1230 1231 close_fd_exit: 1232 close(fd); 1233 1234 put_space_exit: 1235 vfio_put_address_space(space); 1236 1237 return ret; 1238 } 1239 1240 static void vfio_disconnect_container(VFIOGroup *group) 1241 { 1242 VFIOContainer *container = group->container; 1243 1244 QLIST_REMOVE(group, container_next); 1245 group->container = NULL; 1246 1247 /* 1248 * Explicitly release the listener first before unset container, 1249 * since unset may destroy the backend container if it's the last 1250 * group. 1251 */ 1252 if (QLIST_EMPTY(&container->group_list)) { 1253 vfio_listener_release(container); 1254 } 1255 1256 if (ioctl(group->fd, VFIO_GROUP_UNSET_CONTAINER, &container->fd)) { 1257 error_report("vfio: error disconnecting group %d from container", 1258 group->groupid); 1259 } 1260 1261 if (QLIST_EMPTY(&container->group_list)) { 1262 VFIOAddressSpace *space = container->space; 1263 VFIOGuestIOMMU *giommu, *tmp; 1264 1265 QLIST_REMOVE(container, next); 1266 1267 QLIST_FOREACH_SAFE(giommu, &container->giommu_list, giommu_next, tmp) { 1268 memory_region_unregister_iommu_notifier( 1269 MEMORY_REGION(giommu->iommu), &giommu->n); 1270 QLIST_REMOVE(giommu, giommu_next); 1271 g_free(giommu); 1272 } 1273 1274 trace_vfio_disconnect_container(container->fd); 1275 close(container->fd); 1276 g_free(container); 1277 1278 vfio_put_address_space(space); 1279 } 1280 } 1281 1282 VFIOGroup *vfio_get_group(int groupid, AddressSpace *as, Error **errp) 1283 { 1284 VFIOGroup *group; 1285 char path[32]; 1286 struct vfio_group_status status = { .argsz = sizeof(status) }; 1287 1288 QLIST_FOREACH(group, &vfio_group_list, next) { 1289 if (group->groupid == groupid) { 1290 /* Found it. Now is it already in the right context? */ 1291 if (group->container->space->as == as) { 1292 return group; 1293 } else { 1294 error_setg(errp, "group %d used in multiple address spaces", 1295 group->groupid); 1296 return NULL; 1297 } 1298 } 1299 } 1300 1301 group = g_malloc0(sizeof(*group)); 1302 1303 snprintf(path, sizeof(path), "/dev/vfio/%d", groupid); 1304 group->fd = qemu_open(path, O_RDWR); 1305 if (group->fd < 0) { 1306 error_setg_errno(errp, errno, "failed to open %s", path); 1307 goto free_group_exit; 1308 } 1309 1310 if (ioctl(group->fd, VFIO_GROUP_GET_STATUS, &status)) { 1311 error_setg_errno(errp, errno, "failed to get group %d status", groupid); 1312 goto close_fd_exit; 1313 } 1314 1315 if (!(status.flags & VFIO_GROUP_FLAGS_VIABLE)) { 1316 error_setg(errp, "group %d is not viable", groupid); 1317 error_append_hint(errp, 1318 "Please ensure all devices within the iommu_group " 1319 "are bound to their vfio bus driver.\n"); 1320 goto close_fd_exit; 1321 } 1322 1323 group->groupid = groupid; 1324 QLIST_INIT(&group->device_list); 1325 1326 if (vfio_connect_container(group, as, errp)) { 1327 error_prepend(errp, "failed to setup container for group %d: ", 1328 groupid); 1329 goto close_fd_exit; 1330 } 1331 1332 if (QLIST_EMPTY(&vfio_group_list)) { 1333 qemu_register_reset(vfio_reset_handler, NULL); 1334 } 1335 1336 QLIST_INSERT_HEAD(&vfio_group_list, group, next); 1337 1338 return group; 1339 1340 close_fd_exit: 1341 close(group->fd); 1342 1343 free_group_exit: 1344 g_free(group); 1345 1346 return NULL; 1347 } 1348 1349 void vfio_put_group(VFIOGroup *group) 1350 { 1351 if (!group || !QLIST_EMPTY(&group->device_list)) { 1352 return; 1353 } 1354 1355 vfio_kvm_device_del_group(group); 1356 vfio_disconnect_container(group); 1357 QLIST_REMOVE(group, next); 1358 trace_vfio_put_group(group->fd); 1359 close(group->fd); 1360 g_free(group); 1361 1362 if (QLIST_EMPTY(&vfio_group_list)) { 1363 qemu_unregister_reset(vfio_reset_handler, NULL); 1364 } 1365 } 1366 1367 int vfio_get_device(VFIOGroup *group, const char *name, 1368 VFIODevice *vbasedev, Error **errp) 1369 { 1370 struct vfio_device_info dev_info = { .argsz = sizeof(dev_info) }; 1371 int ret, fd; 1372 1373 fd = ioctl(group->fd, VFIO_GROUP_GET_DEVICE_FD, name); 1374 if (fd < 0) { 1375 error_setg_errno(errp, errno, "error getting device from group %d", 1376 group->groupid); 1377 error_append_hint(errp, 1378 "Verify all devices in group %d are bound to vfio-<bus> " 1379 "or pci-stub and not already in use\n", group->groupid); 1380 return fd; 1381 } 1382 1383 ret = ioctl(fd, VFIO_DEVICE_GET_INFO, &dev_info); 1384 if (ret) { 1385 error_setg_errno(errp, errno, "error getting device info"); 1386 close(fd); 1387 return ret; 1388 } 1389 1390 vbasedev->fd = fd; 1391 vbasedev->group = group; 1392 QLIST_INSERT_HEAD(&group->device_list, vbasedev, next); 1393 1394 vbasedev->num_irqs = dev_info.num_irqs; 1395 vbasedev->num_regions = dev_info.num_regions; 1396 vbasedev->flags = dev_info.flags; 1397 1398 trace_vfio_get_device(name, dev_info.flags, dev_info.num_regions, 1399 dev_info.num_irqs); 1400 1401 vbasedev->reset_works = !!(dev_info.flags & VFIO_DEVICE_FLAGS_RESET); 1402 return 0; 1403 } 1404 1405 void vfio_put_base_device(VFIODevice *vbasedev) 1406 { 1407 if (!vbasedev->group) { 1408 return; 1409 } 1410 QLIST_REMOVE(vbasedev, next); 1411 vbasedev->group = NULL; 1412 trace_vfio_put_base_device(vbasedev->fd); 1413 close(vbasedev->fd); 1414 } 1415 1416 int vfio_get_region_info(VFIODevice *vbasedev, int index, 1417 struct vfio_region_info **info) 1418 { 1419 size_t argsz = sizeof(struct vfio_region_info); 1420 1421 *info = g_malloc0(argsz); 1422 1423 (*info)->index = index; 1424 retry: 1425 (*info)->argsz = argsz; 1426 1427 if (ioctl(vbasedev->fd, VFIO_DEVICE_GET_REGION_INFO, *info)) { 1428 g_free(*info); 1429 *info = NULL; 1430 return -errno; 1431 } 1432 1433 if ((*info)->argsz > argsz) { 1434 argsz = (*info)->argsz; 1435 *info = g_realloc(*info, argsz); 1436 1437 goto retry; 1438 } 1439 1440 return 0; 1441 } 1442 1443 int vfio_get_dev_region_info(VFIODevice *vbasedev, uint32_t type, 1444 uint32_t subtype, struct vfio_region_info **info) 1445 { 1446 int i; 1447 1448 for (i = 0; i < vbasedev->num_regions; i++) { 1449 struct vfio_info_cap_header *hdr; 1450 struct vfio_region_info_cap_type *cap_type; 1451 1452 if (vfio_get_region_info(vbasedev, i, info)) { 1453 continue; 1454 } 1455 1456 hdr = vfio_get_region_info_cap(*info, VFIO_REGION_INFO_CAP_TYPE); 1457 if (!hdr) { 1458 g_free(*info); 1459 continue; 1460 } 1461 1462 cap_type = container_of(hdr, struct vfio_region_info_cap_type, header); 1463 1464 trace_vfio_get_dev_region(vbasedev->name, i, 1465 cap_type->type, cap_type->subtype); 1466 1467 if (cap_type->type == type && cap_type->subtype == subtype) { 1468 return 0; 1469 } 1470 1471 g_free(*info); 1472 } 1473 1474 *info = NULL; 1475 return -ENODEV; 1476 } 1477 1478 bool vfio_has_region_cap(VFIODevice *vbasedev, int region, uint16_t cap_type) 1479 { 1480 struct vfio_region_info *info = NULL; 1481 bool ret = false; 1482 1483 if (!vfio_get_region_info(vbasedev, region, &info)) { 1484 if (vfio_get_region_info_cap(info, cap_type)) { 1485 ret = true; 1486 } 1487 g_free(info); 1488 } 1489 1490 return ret; 1491 } 1492 1493 /* 1494 * Interfaces for IBM EEH (Enhanced Error Handling) 1495 */ 1496 static bool vfio_eeh_container_ok(VFIOContainer *container) 1497 { 1498 /* 1499 * As of 2016-03-04 (linux-4.5) the host kernel EEH/VFIO 1500 * implementation is broken if there are multiple groups in a 1501 * container. The hardware works in units of Partitionable 1502 * Endpoints (== IOMMU groups) and the EEH operations naively 1503 * iterate across all groups in the container, without any logic 1504 * to make sure the groups have their state synchronized. For 1505 * certain operations (ENABLE) that might be ok, until an error 1506 * occurs, but for others (GET_STATE) it's clearly broken. 1507 */ 1508 1509 /* 1510 * XXX Once fixed kernels exist, test for them here 1511 */ 1512 1513 if (QLIST_EMPTY(&container->group_list)) { 1514 return false; 1515 } 1516 1517 if (QLIST_NEXT(QLIST_FIRST(&container->group_list), container_next)) { 1518 return false; 1519 } 1520 1521 return true; 1522 } 1523 1524 static int vfio_eeh_container_op(VFIOContainer *container, uint32_t op) 1525 { 1526 struct vfio_eeh_pe_op pe_op = { 1527 .argsz = sizeof(pe_op), 1528 .op = op, 1529 }; 1530 int ret; 1531 1532 if (!vfio_eeh_container_ok(container)) { 1533 error_report("vfio/eeh: EEH_PE_OP 0x%x: " 1534 "kernel requires a container with exactly one group", op); 1535 return -EPERM; 1536 } 1537 1538 ret = ioctl(container->fd, VFIO_EEH_PE_OP, &pe_op); 1539 if (ret < 0) { 1540 error_report("vfio/eeh: EEH_PE_OP 0x%x failed: %m", op); 1541 return -errno; 1542 } 1543 1544 return ret; 1545 } 1546 1547 static VFIOContainer *vfio_eeh_as_container(AddressSpace *as) 1548 { 1549 VFIOAddressSpace *space = vfio_get_address_space(as); 1550 VFIOContainer *container = NULL; 1551 1552 if (QLIST_EMPTY(&space->containers)) { 1553 /* No containers to act on */ 1554 goto out; 1555 } 1556 1557 container = QLIST_FIRST(&space->containers); 1558 1559 if (QLIST_NEXT(container, next)) { 1560 /* We don't yet have logic to synchronize EEH state across 1561 * multiple containers */ 1562 container = NULL; 1563 goto out; 1564 } 1565 1566 out: 1567 vfio_put_address_space(space); 1568 return container; 1569 } 1570 1571 bool vfio_eeh_as_ok(AddressSpace *as) 1572 { 1573 VFIOContainer *container = vfio_eeh_as_container(as); 1574 1575 return (container != NULL) && vfio_eeh_container_ok(container); 1576 } 1577 1578 int vfio_eeh_as_op(AddressSpace *as, uint32_t op) 1579 { 1580 VFIOContainer *container = vfio_eeh_as_container(as); 1581 1582 if (!container) { 1583 return -ENODEV; 1584 } 1585 return vfio_eeh_container_op(container, op); 1586 } 1587