| 1 | /* |
| 2 | * VFIO: IOMMU DMA mapping support for Type1 IOMMU |
| 3 | * |
| 4 | * Copyright (C) 2012 Red Hat, Inc. All rights reserved. |
| 5 | * Author: Alex Williamson <alex.williamson@redhat.com> |
| 6 | * |
| 7 | * This program is free software; you can redistribute it and/or modify |
| 8 | * it under the terms of the GNU General Public License version 2 as |
| 9 | * published by the Free Software Foundation. |
| 10 | * |
| 11 | * Derived from original vfio: |
| 12 | * Copyright 2010 Cisco Systems, Inc. All rights reserved. |
| 13 | * Author: Tom Lyon, pugs@cisco.com |
| 14 | * |
| 15 | * We arbitrarily define a Type1 IOMMU as one matching the below code. |
| 16 | * It could be called the x86 IOMMU as it's designed for AMD-Vi & Intel |
| 17 | * VT-d, but that makes it harder to re-use as theoretically anyone |
| 18 | * implementing a similar IOMMU could make use of this. We expect the |
| 19 | * IOMMU to support the IOMMU API and have few to no restrictions around |
| 20 | * the IOVA range that can be mapped. The Type1 IOMMU is currently |
| 21 | * optimized for relatively static mappings of a userspace process with |
| 22 | * userpsace pages pinned into memory. We also assume devices and IOMMU |
| 23 | * domains are PCI based as the IOMMU API is still centered around a |
| 24 | * device/bus interface rather than a group interface. |
| 25 | */ |
| 26 | |
| 27 | #include <linux/compat.h> |
| 28 | #include <linux/device.h> |
| 29 | #include <linux/fs.h> |
| 30 | #include <linux/iommu.h> |
| 31 | #include <linux/module.h> |
| 32 | #include <linux/mm.h> |
| 33 | #include <linux/pci.h> /* pci_bus_type */ |
| 34 | #include <linux/sched.h> |
| 35 | #include <linux/slab.h> |
| 36 | #include <linux/uaccess.h> |
| 37 | #include <linux/vfio.h> |
| 38 | #include <linux/workqueue.h> |
| 39 | |
| 40 | #define DRIVER_VERSION "0.2" |
| 41 | #define DRIVER_AUTHOR "Alex Williamson <alex.williamson@redhat.com>" |
| 42 | #define DRIVER_DESC "Type1 IOMMU driver for VFIO" |
| 43 | |
| 44 | static bool allow_unsafe_interrupts; |
| 45 | module_param_named(allow_unsafe_interrupts, |
| 46 | allow_unsafe_interrupts, bool, S_IRUGO | S_IWUSR); |
| 47 | MODULE_PARM_DESC(allow_unsafe_interrupts, |
| 48 | "Enable VFIO IOMMU support for on platforms without interrupt remapping support."); |
| 49 | |
| 50 | struct vfio_iommu { |
| 51 | struct iommu_domain *domain; |
| 52 | struct mutex lock; |
| 53 | struct list_head dma_list; |
| 54 | struct list_head group_list; |
| 55 | bool cache; |
| 56 | }; |
| 57 | |
| 58 | struct vfio_dma { |
| 59 | struct list_head next; |
| 60 | dma_addr_t iova; /* Device address */ |
| 61 | unsigned long vaddr; /* Process virtual addr */ |
| 62 | long npage; /* Number of pages */ |
| 63 | int prot; /* IOMMU_READ/WRITE */ |
| 64 | }; |
| 65 | |
| 66 | struct vfio_group { |
| 67 | struct iommu_group *iommu_group; |
| 68 | struct list_head next; |
| 69 | }; |
| 70 | |
| 71 | /* |
| 72 | * This code handles mapping and unmapping of user data buffers |
| 73 | * into DMA'ble space using the IOMMU |
| 74 | */ |
| 75 | |
| 76 | #define NPAGE_TO_SIZE(npage) ((size_t)(npage) << PAGE_SHIFT) |
| 77 | |
| 78 | struct vwork { |
| 79 | struct mm_struct *mm; |
| 80 | long npage; |
| 81 | struct work_struct work; |
| 82 | }; |
| 83 | |
| 84 | /* delayed decrement/increment for locked_vm */ |
| 85 | static void vfio_lock_acct_bg(struct work_struct *work) |
| 86 | { |
| 87 | struct vwork *vwork = container_of(work, struct vwork, work); |
| 88 | struct mm_struct *mm; |
| 89 | |
| 90 | mm = vwork->mm; |
| 91 | down_write(&mm->mmap_sem); |
| 92 | mm->locked_vm += vwork->npage; |
| 93 | up_write(&mm->mmap_sem); |
| 94 | mmput(mm); |
| 95 | kfree(vwork); |
| 96 | } |
| 97 | |
| 98 | static void vfio_lock_acct(long npage) |
| 99 | { |
| 100 | struct vwork *vwork; |
| 101 | struct mm_struct *mm; |
| 102 | |
| 103 | if (!current->mm) |
| 104 | return; /* process exited */ |
| 105 | |
| 106 | if (down_write_trylock(¤t->mm->mmap_sem)) { |
| 107 | current->mm->locked_vm += npage; |
| 108 | up_write(¤t->mm->mmap_sem); |
| 109 | return; |
| 110 | } |
| 111 | |
| 112 | /* |
| 113 | * Couldn't get mmap_sem lock, so must setup to update |
| 114 | * mm->locked_vm later. If locked_vm were atomic, we |
| 115 | * wouldn't need this silliness |
| 116 | */ |
| 117 | vwork = kmalloc(sizeof(struct vwork), GFP_KERNEL); |
| 118 | if (!vwork) |
| 119 | return; |
| 120 | mm = get_task_mm(current); |
| 121 | if (!mm) { |
| 122 | kfree(vwork); |
| 123 | return; |
| 124 | } |
| 125 | INIT_WORK(&vwork->work, vfio_lock_acct_bg); |
| 126 | vwork->mm = mm; |
| 127 | vwork->npage = npage; |
| 128 | schedule_work(&vwork->work); |
| 129 | } |
| 130 | |
| 131 | /* |
| 132 | * Some mappings aren't backed by a struct page, for example an mmap'd |
| 133 | * MMIO range for our own or another device. These use a different |
| 134 | * pfn conversion and shouldn't be tracked as locked pages. |
| 135 | */ |
| 136 | static bool is_invalid_reserved_pfn(unsigned long pfn) |
| 137 | { |
| 138 | if (pfn_valid(pfn)) { |
| 139 | bool reserved; |
| 140 | struct page *tail = pfn_to_page(pfn); |
| 141 | struct page *head = compound_head(tail); |
| 142 | reserved = !!(PageReserved(head)); |
| 143 | if (head != tail) { |
| 144 | /* |
| 145 | * "head" is not a dangling pointer |
| 146 | * (compound_head takes care of that) |
| 147 | * but the hugepage may have been split |
| 148 | * from under us (and we may not hold a |
| 149 | * reference count on the head page so it can |
| 150 | * be reused before we run PageReferenced), so |
| 151 | * we've to check PageTail before returning |
| 152 | * what we just read. |
| 153 | */ |
| 154 | smp_rmb(); |
| 155 | if (PageTail(tail)) |
| 156 | return reserved; |
| 157 | } |
| 158 | return PageReserved(tail); |
| 159 | } |
| 160 | |
| 161 | return true; |
| 162 | } |
| 163 | |
| 164 | static int put_pfn(unsigned long pfn, int prot) |
| 165 | { |
| 166 | if (!is_invalid_reserved_pfn(pfn)) { |
| 167 | struct page *page = pfn_to_page(pfn); |
| 168 | if (prot & IOMMU_WRITE) |
| 169 | SetPageDirty(page); |
| 170 | put_page(page); |
| 171 | return 1; |
| 172 | } |
| 173 | return 0; |
| 174 | } |
| 175 | |
| 176 | /* Unmap DMA region */ |
| 177 | static long __vfio_dma_do_unmap(struct vfio_iommu *iommu, dma_addr_t iova, |
| 178 | long npage, int prot) |
| 179 | { |
| 180 | long i, unlocked = 0; |
| 181 | |
| 182 | for (i = 0; i < npage; i++, iova += PAGE_SIZE) { |
| 183 | unsigned long pfn; |
| 184 | |
| 185 | pfn = iommu_iova_to_phys(iommu->domain, iova) >> PAGE_SHIFT; |
| 186 | if (pfn) { |
| 187 | iommu_unmap(iommu->domain, iova, PAGE_SIZE); |
| 188 | unlocked += put_pfn(pfn, prot); |
| 189 | } |
| 190 | } |
| 191 | return unlocked; |
| 192 | } |
| 193 | |
| 194 | static void vfio_dma_unmap(struct vfio_iommu *iommu, dma_addr_t iova, |
| 195 | long npage, int prot) |
| 196 | { |
| 197 | long unlocked; |
| 198 | |
| 199 | unlocked = __vfio_dma_do_unmap(iommu, iova, npage, prot); |
| 200 | vfio_lock_acct(-unlocked); |
| 201 | } |
| 202 | |
| 203 | static int vaddr_get_pfn(unsigned long vaddr, int prot, unsigned long *pfn) |
| 204 | { |
| 205 | struct page *page[1]; |
| 206 | struct vm_area_struct *vma; |
| 207 | int ret = -EFAULT; |
| 208 | |
| 209 | if (get_user_pages_fast(vaddr, 1, !!(prot & IOMMU_WRITE), page) == 1) { |
| 210 | *pfn = page_to_pfn(page[0]); |
| 211 | return 0; |
| 212 | } |
| 213 | |
| 214 | down_read(¤t->mm->mmap_sem); |
| 215 | |
| 216 | vma = find_vma_intersection(current->mm, vaddr, vaddr + 1); |
| 217 | |
| 218 | if (vma && vma->vm_flags & VM_PFNMAP) { |
| 219 | *pfn = ((vaddr - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff; |
| 220 | if (is_invalid_reserved_pfn(*pfn)) |
| 221 | ret = 0; |
| 222 | } |
| 223 | |
| 224 | up_read(¤t->mm->mmap_sem); |
| 225 | |
| 226 | return ret; |
| 227 | } |
| 228 | |
| 229 | /* Map DMA region */ |
| 230 | static int __vfio_dma_map(struct vfio_iommu *iommu, dma_addr_t iova, |
| 231 | unsigned long vaddr, long npage, int prot) |
| 232 | { |
| 233 | dma_addr_t start = iova; |
| 234 | long i, locked = 0; |
| 235 | int ret; |
| 236 | |
| 237 | /* Verify that pages are not already mapped */ |
| 238 | for (i = 0; i < npage; i++, iova += PAGE_SIZE) |
| 239 | if (iommu_iova_to_phys(iommu->domain, iova)) |
| 240 | return -EBUSY; |
| 241 | |
| 242 | iova = start; |
| 243 | |
| 244 | if (iommu->cache) |
| 245 | prot |= IOMMU_CACHE; |
| 246 | |
| 247 | /* |
| 248 | * XXX We break mappings into pages and use get_user_pages_fast to |
| 249 | * pin the pages in memory. It's been suggested that mlock might |
| 250 | * provide a more efficient mechanism, but nothing prevents the |
| 251 | * user from munlocking the pages, which could then allow the user |
| 252 | * access to random host memory. We also have no guarantee from the |
| 253 | * IOMMU API that the iommu driver can unmap sub-pages of previous |
| 254 | * mappings. This means we might lose an entire range if a single |
| 255 | * page within it is unmapped. Single page mappings are inefficient, |
| 256 | * but provide the most flexibility for now. |
| 257 | */ |
| 258 | for (i = 0; i < npage; i++, iova += PAGE_SIZE, vaddr += PAGE_SIZE) { |
| 259 | unsigned long pfn = 0; |
| 260 | |
| 261 | ret = vaddr_get_pfn(vaddr, prot, &pfn); |
| 262 | if (ret) { |
| 263 | __vfio_dma_do_unmap(iommu, start, i, prot); |
| 264 | return ret; |
| 265 | } |
| 266 | |
| 267 | /* |
| 268 | * Only add actual locked pages to accounting |
| 269 | * XXX We're effectively marking a page locked for every |
| 270 | * IOVA page even though it's possible the user could be |
| 271 | * backing multiple IOVAs with the same vaddr. This over- |
| 272 | * penalizes the user process, but we currently have no |
| 273 | * easy way to do this properly. |
| 274 | */ |
| 275 | if (!is_invalid_reserved_pfn(pfn)) |
| 276 | locked++; |
| 277 | |
| 278 | ret = iommu_map(iommu->domain, iova, |
| 279 | (phys_addr_t)pfn << PAGE_SHIFT, |
| 280 | PAGE_SIZE, prot); |
| 281 | if (ret) { |
| 282 | /* Back out mappings on error */ |
| 283 | put_pfn(pfn, prot); |
| 284 | __vfio_dma_do_unmap(iommu, start, i, prot); |
| 285 | return ret; |
| 286 | } |
| 287 | } |
| 288 | vfio_lock_acct(locked); |
| 289 | return 0; |
| 290 | } |
| 291 | |
| 292 | static inline bool ranges_overlap(dma_addr_t start1, size_t size1, |
| 293 | dma_addr_t start2, size_t size2) |
| 294 | { |
| 295 | if (start1 < start2) |
| 296 | return (start2 - start1 < size1); |
| 297 | else if (start2 < start1) |
| 298 | return (start1 - start2 < size2); |
| 299 | return (size1 > 0 && size2 > 0); |
| 300 | } |
| 301 | |
| 302 | static struct vfio_dma *vfio_find_dma(struct vfio_iommu *iommu, |
| 303 | dma_addr_t start, size_t size) |
| 304 | { |
| 305 | struct vfio_dma *dma; |
| 306 | |
| 307 | list_for_each_entry(dma, &iommu->dma_list, next) { |
| 308 | if (ranges_overlap(dma->iova, NPAGE_TO_SIZE(dma->npage), |
| 309 | start, size)) |
| 310 | return dma; |
| 311 | } |
| 312 | return NULL; |
| 313 | } |
| 314 | |
| 315 | static long vfio_remove_dma_overlap(struct vfio_iommu *iommu, dma_addr_t start, |
| 316 | size_t size, struct vfio_dma *dma) |
| 317 | { |
| 318 | struct vfio_dma *split; |
| 319 | long npage_lo, npage_hi; |
| 320 | |
| 321 | /* Existing dma region is completely covered, unmap all */ |
| 322 | if (start <= dma->iova && |
| 323 | start + size >= dma->iova + NPAGE_TO_SIZE(dma->npage)) { |
| 324 | vfio_dma_unmap(iommu, dma->iova, dma->npage, dma->prot); |
| 325 | list_del(&dma->next); |
| 326 | npage_lo = dma->npage; |
| 327 | kfree(dma); |
| 328 | return npage_lo; |
| 329 | } |
| 330 | |
| 331 | /* Overlap low address of existing range */ |
| 332 | if (start <= dma->iova) { |
| 333 | size_t overlap; |
| 334 | |
| 335 | overlap = start + size - dma->iova; |
| 336 | npage_lo = overlap >> PAGE_SHIFT; |
| 337 | |
| 338 | vfio_dma_unmap(iommu, dma->iova, npage_lo, dma->prot); |
| 339 | dma->iova += overlap; |
| 340 | dma->vaddr += overlap; |
| 341 | dma->npage -= npage_lo; |
| 342 | return npage_lo; |
| 343 | } |
| 344 | |
| 345 | /* Overlap high address of existing range */ |
| 346 | if (start + size >= dma->iova + NPAGE_TO_SIZE(dma->npage)) { |
| 347 | size_t overlap; |
| 348 | |
| 349 | overlap = dma->iova + NPAGE_TO_SIZE(dma->npage) - start; |
| 350 | npage_hi = overlap >> PAGE_SHIFT; |
| 351 | |
| 352 | vfio_dma_unmap(iommu, start, npage_hi, dma->prot); |
| 353 | dma->npage -= npage_hi; |
| 354 | return npage_hi; |
| 355 | } |
| 356 | |
| 357 | /* Split existing */ |
| 358 | npage_lo = (start - dma->iova) >> PAGE_SHIFT; |
| 359 | npage_hi = dma->npage - (size >> PAGE_SHIFT) - npage_lo; |
| 360 | |
| 361 | split = kzalloc(sizeof *split, GFP_KERNEL); |
| 362 | if (!split) |
| 363 | return -ENOMEM; |
| 364 | |
| 365 | vfio_dma_unmap(iommu, start, size >> PAGE_SHIFT, dma->prot); |
| 366 | |
| 367 | dma->npage = npage_lo; |
| 368 | |
| 369 | split->npage = npage_hi; |
| 370 | split->iova = start + size; |
| 371 | split->vaddr = dma->vaddr + NPAGE_TO_SIZE(npage_lo) + size; |
| 372 | split->prot = dma->prot; |
| 373 | list_add(&split->next, &iommu->dma_list); |
| 374 | return size >> PAGE_SHIFT; |
| 375 | } |
| 376 | |
| 377 | static int vfio_dma_do_unmap(struct vfio_iommu *iommu, |
| 378 | struct vfio_iommu_type1_dma_unmap *unmap) |
| 379 | { |
| 380 | long ret = 0, npage = unmap->size >> PAGE_SHIFT; |
| 381 | struct vfio_dma *dma, *tmp; |
| 382 | uint64_t mask; |
| 383 | |
| 384 | mask = ((uint64_t)1 << __ffs(iommu->domain->ops->pgsize_bitmap)) - 1; |
| 385 | |
| 386 | if (unmap->iova & mask) |
| 387 | return -EINVAL; |
| 388 | if (unmap->size & mask) |
| 389 | return -EINVAL; |
| 390 | |
| 391 | /* XXX We still break these down into PAGE_SIZE */ |
| 392 | WARN_ON(mask & PAGE_MASK); |
| 393 | |
| 394 | mutex_lock(&iommu->lock); |
| 395 | |
| 396 | list_for_each_entry_safe(dma, tmp, &iommu->dma_list, next) { |
| 397 | if (ranges_overlap(dma->iova, NPAGE_TO_SIZE(dma->npage), |
| 398 | unmap->iova, unmap->size)) { |
| 399 | ret = vfio_remove_dma_overlap(iommu, unmap->iova, |
| 400 | unmap->size, dma); |
| 401 | if (ret > 0) |
| 402 | npage -= ret; |
| 403 | if (ret < 0 || npage == 0) |
| 404 | break; |
| 405 | } |
| 406 | } |
| 407 | mutex_unlock(&iommu->lock); |
| 408 | return ret > 0 ? 0 : (int)ret; |
| 409 | } |
| 410 | |
| 411 | static int vfio_dma_do_map(struct vfio_iommu *iommu, |
| 412 | struct vfio_iommu_type1_dma_map *map) |
| 413 | { |
| 414 | struct vfio_dma *dma, *pdma = NULL; |
| 415 | dma_addr_t iova = map->iova; |
| 416 | unsigned long locked, lock_limit, vaddr = map->vaddr; |
| 417 | size_t size = map->size; |
| 418 | int ret = 0, prot = 0; |
| 419 | uint64_t mask; |
| 420 | long npage; |
| 421 | |
| 422 | mask = ((uint64_t)1 << __ffs(iommu->domain->ops->pgsize_bitmap)) - 1; |
| 423 | |
| 424 | /* READ/WRITE from device perspective */ |
| 425 | if (map->flags & VFIO_DMA_MAP_FLAG_WRITE) |
| 426 | prot |= IOMMU_WRITE; |
| 427 | if (map->flags & VFIO_DMA_MAP_FLAG_READ) |
| 428 | prot |= IOMMU_READ; |
| 429 | |
| 430 | if (!prot) |
| 431 | return -EINVAL; /* No READ/WRITE? */ |
| 432 | |
| 433 | if (vaddr & mask) |
| 434 | return -EINVAL; |
| 435 | if (iova & mask) |
| 436 | return -EINVAL; |
| 437 | if (size & mask) |
| 438 | return -EINVAL; |
| 439 | |
| 440 | /* XXX We still break these down into PAGE_SIZE */ |
| 441 | WARN_ON(mask & PAGE_MASK); |
| 442 | |
| 443 | /* Don't allow IOVA wrap */ |
| 444 | if (iova + size && iova + size < iova) |
| 445 | return -EINVAL; |
| 446 | |
| 447 | /* Don't allow virtual address wrap */ |
| 448 | if (vaddr + size && vaddr + size < vaddr) |
| 449 | return -EINVAL; |
| 450 | |
| 451 | npage = size >> PAGE_SHIFT; |
| 452 | if (!npage) |
| 453 | return -EINVAL; |
| 454 | |
| 455 | mutex_lock(&iommu->lock); |
| 456 | |
| 457 | if (vfio_find_dma(iommu, iova, size)) { |
| 458 | ret = -EBUSY; |
| 459 | goto out_lock; |
| 460 | } |
| 461 | |
| 462 | /* account for locked pages */ |
| 463 | locked = current->mm->locked_vm + npage; |
| 464 | lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT; |
| 465 | if (locked > lock_limit && !capable(CAP_IPC_LOCK)) { |
| 466 | pr_warn("%s: RLIMIT_MEMLOCK (%ld) exceeded\n", |
| 467 | __func__, rlimit(RLIMIT_MEMLOCK)); |
| 468 | ret = -ENOMEM; |
| 469 | goto out_lock; |
| 470 | } |
| 471 | |
| 472 | ret = __vfio_dma_map(iommu, iova, vaddr, npage, prot); |
| 473 | if (ret) |
| 474 | goto out_lock; |
| 475 | |
| 476 | /* Check if we abut a region below - nothing below 0 */ |
| 477 | if (iova) { |
| 478 | dma = vfio_find_dma(iommu, iova - 1, 1); |
| 479 | if (dma && dma->prot == prot && |
| 480 | dma->vaddr + NPAGE_TO_SIZE(dma->npage) == vaddr) { |
| 481 | |
| 482 | dma->npage += npage; |
| 483 | iova = dma->iova; |
| 484 | vaddr = dma->vaddr; |
| 485 | npage = dma->npage; |
| 486 | size = NPAGE_TO_SIZE(npage); |
| 487 | |
| 488 | pdma = dma; |
| 489 | } |
| 490 | } |
| 491 | |
| 492 | /* Check if we abut a region above - nothing above ~0 + 1 */ |
| 493 | if (iova + size) { |
| 494 | dma = vfio_find_dma(iommu, iova + size, 1); |
| 495 | if (dma && dma->prot == prot && |
| 496 | dma->vaddr == vaddr + size) { |
| 497 | |
| 498 | dma->npage += npage; |
| 499 | dma->iova = iova; |
| 500 | dma->vaddr = vaddr; |
| 501 | |
| 502 | /* |
| 503 | * If merged above and below, remove previously |
| 504 | * merged entry. New entry covers it. |
| 505 | */ |
| 506 | if (pdma) { |
| 507 | list_del(&pdma->next); |
| 508 | kfree(pdma); |
| 509 | } |
| 510 | pdma = dma; |
| 511 | } |
| 512 | } |
| 513 | |
| 514 | /* Isolated, new region */ |
| 515 | if (!pdma) { |
| 516 | dma = kzalloc(sizeof *dma, GFP_KERNEL); |
| 517 | if (!dma) { |
| 518 | ret = -ENOMEM; |
| 519 | vfio_dma_unmap(iommu, iova, npage, prot); |
| 520 | goto out_lock; |
| 521 | } |
| 522 | |
| 523 | dma->npage = npage; |
| 524 | dma->iova = iova; |
| 525 | dma->vaddr = vaddr; |
| 526 | dma->prot = prot; |
| 527 | list_add(&dma->next, &iommu->dma_list); |
| 528 | } |
| 529 | |
| 530 | out_lock: |
| 531 | mutex_unlock(&iommu->lock); |
| 532 | return ret; |
| 533 | } |
| 534 | |
| 535 | static int vfio_iommu_type1_attach_group(void *iommu_data, |
| 536 | struct iommu_group *iommu_group) |
| 537 | { |
| 538 | struct vfio_iommu *iommu = iommu_data; |
| 539 | struct vfio_group *group, *tmp; |
| 540 | int ret; |
| 541 | |
| 542 | group = kzalloc(sizeof(*group), GFP_KERNEL); |
| 543 | if (!group) |
| 544 | return -ENOMEM; |
| 545 | |
| 546 | mutex_lock(&iommu->lock); |
| 547 | |
| 548 | list_for_each_entry(tmp, &iommu->group_list, next) { |
| 549 | if (tmp->iommu_group == iommu_group) { |
| 550 | mutex_unlock(&iommu->lock); |
| 551 | kfree(group); |
| 552 | return -EINVAL; |
| 553 | } |
| 554 | } |
| 555 | |
| 556 | /* |
| 557 | * TODO: Domain have capabilities that might change as we add |
| 558 | * groups (see iommu->cache, currently never set). Check for |
| 559 | * them and potentially disallow groups to be attached when it |
| 560 | * would change capabilities (ugh). |
| 561 | */ |
| 562 | ret = iommu_attach_group(iommu->domain, iommu_group); |
| 563 | if (ret) { |
| 564 | mutex_unlock(&iommu->lock); |
| 565 | kfree(group); |
| 566 | return ret; |
| 567 | } |
| 568 | |
| 569 | group->iommu_group = iommu_group; |
| 570 | list_add(&group->next, &iommu->group_list); |
| 571 | |
| 572 | mutex_unlock(&iommu->lock); |
| 573 | |
| 574 | return 0; |
| 575 | } |
| 576 | |
| 577 | static void vfio_iommu_type1_detach_group(void *iommu_data, |
| 578 | struct iommu_group *iommu_group) |
| 579 | { |
| 580 | struct vfio_iommu *iommu = iommu_data; |
| 581 | struct vfio_group *group; |
| 582 | |
| 583 | mutex_lock(&iommu->lock); |
| 584 | |
| 585 | list_for_each_entry(group, &iommu->group_list, next) { |
| 586 | if (group->iommu_group == iommu_group) { |
| 587 | iommu_detach_group(iommu->domain, iommu_group); |
| 588 | list_del(&group->next); |
| 589 | kfree(group); |
| 590 | break; |
| 591 | } |
| 592 | } |
| 593 | |
| 594 | mutex_unlock(&iommu->lock); |
| 595 | } |
| 596 | |
| 597 | static void *vfio_iommu_type1_open(unsigned long arg) |
| 598 | { |
| 599 | struct vfio_iommu *iommu; |
| 600 | |
| 601 | if (arg != VFIO_TYPE1_IOMMU) |
| 602 | return ERR_PTR(-EINVAL); |
| 603 | |
| 604 | iommu = kzalloc(sizeof(*iommu), GFP_KERNEL); |
| 605 | if (!iommu) |
| 606 | return ERR_PTR(-ENOMEM); |
| 607 | |
| 608 | INIT_LIST_HEAD(&iommu->group_list); |
| 609 | INIT_LIST_HEAD(&iommu->dma_list); |
| 610 | mutex_init(&iommu->lock); |
| 611 | |
| 612 | /* |
| 613 | * Wish we didn't have to know about bus_type here. |
| 614 | */ |
| 615 | iommu->domain = iommu_domain_alloc(&pci_bus_type); |
| 616 | if (!iommu->domain) { |
| 617 | kfree(iommu); |
| 618 | return ERR_PTR(-EIO); |
| 619 | } |
| 620 | |
| 621 | /* |
| 622 | * Wish we could specify required capabilities rather than create |
| 623 | * a domain, see what comes out and hope it doesn't change along |
| 624 | * the way. Fortunately we know interrupt remapping is global for |
| 625 | * our iommus. |
| 626 | */ |
| 627 | if (!allow_unsafe_interrupts && |
| 628 | !iommu_domain_has_cap(iommu->domain, IOMMU_CAP_INTR_REMAP)) { |
| 629 | pr_warn("%s: No interrupt remapping support. Use the module param \"allow_unsafe_interrupts\" to enable VFIO IOMMU support on this platform\n", |
| 630 | __func__); |
| 631 | iommu_domain_free(iommu->domain); |
| 632 | kfree(iommu); |
| 633 | return ERR_PTR(-EPERM); |
| 634 | } |
| 635 | |
| 636 | return iommu; |
| 637 | } |
| 638 | |
| 639 | static void vfio_iommu_type1_release(void *iommu_data) |
| 640 | { |
| 641 | struct vfio_iommu *iommu = iommu_data; |
| 642 | struct vfio_group *group, *group_tmp; |
| 643 | struct vfio_dma *dma, *dma_tmp; |
| 644 | |
| 645 | list_for_each_entry_safe(group, group_tmp, &iommu->group_list, next) { |
| 646 | iommu_detach_group(iommu->domain, group->iommu_group); |
| 647 | list_del(&group->next); |
| 648 | kfree(group); |
| 649 | } |
| 650 | |
| 651 | list_for_each_entry_safe(dma, dma_tmp, &iommu->dma_list, next) { |
| 652 | vfio_dma_unmap(iommu, dma->iova, dma->npage, dma->prot); |
| 653 | list_del(&dma->next); |
| 654 | kfree(dma); |
| 655 | } |
| 656 | |
| 657 | iommu_domain_free(iommu->domain); |
| 658 | iommu->domain = NULL; |
| 659 | kfree(iommu); |
| 660 | } |
| 661 | |
| 662 | static long vfio_iommu_type1_ioctl(void *iommu_data, |
| 663 | unsigned int cmd, unsigned long arg) |
| 664 | { |
| 665 | struct vfio_iommu *iommu = iommu_data; |
| 666 | unsigned long minsz; |
| 667 | |
| 668 | if (cmd == VFIO_CHECK_EXTENSION) { |
| 669 | switch (arg) { |
| 670 | case VFIO_TYPE1_IOMMU: |
| 671 | return 1; |
| 672 | default: |
| 673 | return 0; |
| 674 | } |
| 675 | } else if (cmd == VFIO_IOMMU_GET_INFO) { |
| 676 | struct vfio_iommu_type1_info info; |
| 677 | |
| 678 | minsz = offsetofend(struct vfio_iommu_type1_info, iova_pgsizes); |
| 679 | |
| 680 | if (copy_from_user(&info, (void __user *)arg, minsz)) |
| 681 | return -EFAULT; |
| 682 | |
| 683 | if (info.argsz < minsz) |
| 684 | return -EINVAL; |
| 685 | |
| 686 | info.flags = 0; |
| 687 | |
| 688 | info.iova_pgsizes = iommu->domain->ops->pgsize_bitmap; |
| 689 | |
| 690 | return copy_to_user((void __user *)arg, &info, minsz); |
| 691 | |
| 692 | } else if (cmd == VFIO_IOMMU_MAP_DMA) { |
| 693 | struct vfio_iommu_type1_dma_map map; |
| 694 | uint32_t mask = VFIO_DMA_MAP_FLAG_READ | |
| 695 | VFIO_DMA_MAP_FLAG_WRITE; |
| 696 | |
| 697 | minsz = offsetofend(struct vfio_iommu_type1_dma_map, size); |
| 698 | |
| 699 | if (copy_from_user(&map, (void __user *)arg, minsz)) |
| 700 | return -EFAULT; |
| 701 | |
| 702 | if (map.argsz < minsz || map.flags & ~mask) |
| 703 | return -EINVAL; |
| 704 | |
| 705 | return vfio_dma_do_map(iommu, &map); |
| 706 | |
| 707 | } else if (cmd == VFIO_IOMMU_UNMAP_DMA) { |
| 708 | struct vfio_iommu_type1_dma_unmap unmap; |
| 709 | |
| 710 | minsz = offsetofend(struct vfio_iommu_type1_dma_unmap, size); |
| 711 | |
| 712 | if (copy_from_user(&unmap, (void __user *)arg, minsz)) |
| 713 | return -EFAULT; |
| 714 | |
| 715 | if (unmap.argsz < minsz || unmap.flags) |
| 716 | return -EINVAL; |
| 717 | |
| 718 | return vfio_dma_do_unmap(iommu, &unmap); |
| 719 | } |
| 720 | |
| 721 | return -ENOTTY; |
| 722 | } |
| 723 | |
| 724 | static const struct vfio_iommu_driver_ops vfio_iommu_driver_ops_type1 = { |
| 725 | .name = "vfio-iommu-type1", |
| 726 | .owner = THIS_MODULE, |
| 727 | .open = vfio_iommu_type1_open, |
| 728 | .release = vfio_iommu_type1_release, |
| 729 | .ioctl = vfio_iommu_type1_ioctl, |
| 730 | .attach_group = vfio_iommu_type1_attach_group, |
| 731 | .detach_group = vfio_iommu_type1_detach_group, |
| 732 | }; |
| 733 | |
| 734 | static int __init vfio_iommu_type1_init(void) |
| 735 | { |
| 736 | if (!iommu_present(&pci_bus_type)) |
| 737 | return -ENODEV; |
| 738 | |
| 739 | return vfio_register_iommu_driver(&vfio_iommu_driver_ops_type1); |
| 740 | } |
| 741 | |
| 742 | static void __exit vfio_iommu_type1_cleanup(void) |
| 743 | { |
| 744 | vfio_unregister_iommu_driver(&vfio_iommu_driver_ops_type1); |
| 745 | } |
| 746 | |
| 747 | module_init(vfio_iommu_type1_init); |
| 748 | module_exit(vfio_iommu_type1_cleanup); |
| 749 | |
| 750 | MODULE_VERSION(DRIVER_VERSION); |
| 751 | MODULE_LICENSE("GPL v2"); |
| 752 | MODULE_AUTHOR(DRIVER_AUTHOR); |
| 753 | MODULE_DESCRIPTION(DRIVER_DESC); |