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