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