Commit | Line | Data |
---|---|---|
73fa0d10 AW |
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> | |
cd9b2268 | 33 | #include <linux/rbtree.h> |
73fa0d10 AW |
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 | ||
5c6c2b21 AW |
50 | static bool disable_hugepages; |
51 | module_param_named(disable_hugepages, | |
52 | disable_hugepages, bool, S_IRUGO | S_IWUSR); | |
53 | MODULE_PARM_DESC(disable_hugepages, | |
54 | "Disable VFIO IOMMU support for IOMMU hugepages."); | |
55 | ||
73fa0d10 | 56 | struct vfio_iommu { |
1ef3e2bc | 57 | struct list_head domain_list; |
73fa0d10 | 58 | struct mutex lock; |
cd9b2268 | 59 | struct rb_root dma_list; |
1ef3e2bc AW |
60 | bool v2; |
61 | }; | |
62 | ||
63 | struct vfio_domain { | |
64 | struct iommu_domain *domain; | |
65 | struct list_head next; | |
73fa0d10 | 66 | struct list_head group_list; |
1ef3e2bc | 67 | int prot; /* IOMMU_CACHE */ |
73fa0d10 AW |
68 | }; |
69 | ||
70 | struct vfio_dma { | |
cd9b2268 | 71 | struct rb_node node; |
73fa0d10 AW |
72 | dma_addr_t iova; /* Device address */ |
73 | unsigned long vaddr; /* Process virtual addr */ | |
166fd7d9 | 74 | size_t size; /* Map size (bytes) */ |
73fa0d10 AW |
75 | int prot; /* IOMMU_READ/WRITE */ |
76 | }; | |
77 | ||
78 | struct vfio_group { | |
79 | struct iommu_group *iommu_group; | |
80 | struct list_head next; | |
81 | }; | |
82 | ||
83 | /* | |
84 | * This code handles mapping and unmapping of user data buffers | |
85 | * into DMA'ble space using the IOMMU | |
86 | */ | |
87 | ||
cd9b2268 AW |
88 | static struct vfio_dma *vfio_find_dma(struct vfio_iommu *iommu, |
89 | dma_addr_t start, size_t size) | |
90 | { | |
91 | struct rb_node *node = iommu->dma_list.rb_node; | |
92 | ||
93 | while (node) { | |
94 | struct vfio_dma *dma = rb_entry(node, struct vfio_dma, node); | |
95 | ||
96 | if (start + size <= dma->iova) | |
97 | node = node->rb_left; | |
166fd7d9 | 98 | else if (start >= dma->iova + dma->size) |
cd9b2268 AW |
99 | node = node->rb_right; |
100 | else | |
101 | return dma; | |
102 | } | |
103 | ||
104 | return NULL; | |
105 | } | |
106 | ||
1ef3e2bc | 107 | static void vfio_link_dma(struct vfio_iommu *iommu, struct vfio_dma *new) |
cd9b2268 AW |
108 | { |
109 | struct rb_node **link = &iommu->dma_list.rb_node, *parent = NULL; | |
110 | struct vfio_dma *dma; | |
111 | ||
112 | while (*link) { | |
113 | parent = *link; | |
114 | dma = rb_entry(parent, struct vfio_dma, node); | |
115 | ||
166fd7d9 | 116 | if (new->iova + new->size <= dma->iova) |
cd9b2268 AW |
117 | link = &(*link)->rb_left; |
118 | else | |
119 | link = &(*link)->rb_right; | |
120 | } | |
121 | ||
122 | rb_link_node(&new->node, parent, link); | |
123 | rb_insert_color(&new->node, &iommu->dma_list); | |
124 | } | |
125 | ||
1ef3e2bc | 126 | static void vfio_unlink_dma(struct vfio_iommu *iommu, struct vfio_dma *old) |
cd9b2268 AW |
127 | { |
128 | rb_erase(&old->node, &iommu->dma_list); | |
129 | } | |
130 | ||
73fa0d10 AW |
131 | struct vwork { |
132 | struct mm_struct *mm; | |
133 | long npage; | |
134 | struct work_struct work; | |
135 | }; | |
136 | ||
137 | /* delayed decrement/increment for locked_vm */ | |
138 | static void vfio_lock_acct_bg(struct work_struct *work) | |
139 | { | |
140 | struct vwork *vwork = container_of(work, struct vwork, work); | |
141 | struct mm_struct *mm; | |
142 | ||
143 | mm = vwork->mm; | |
144 | down_write(&mm->mmap_sem); | |
145 | mm->locked_vm += vwork->npage; | |
146 | up_write(&mm->mmap_sem); | |
147 | mmput(mm); | |
148 | kfree(vwork); | |
149 | } | |
150 | ||
151 | static void vfio_lock_acct(long npage) | |
152 | { | |
153 | struct vwork *vwork; | |
154 | struct mm_struct *mm; | |
155 | ||
166fd7d9 AW |
156 | if (!current->mm || !npage) |
157 | return; /* process exited or nothing to do */ | |
73fa0d10 AW |
158 | |
159 | if (down_write_trylock(¤t->mm->mmap_sem)) { | |
160 | current->mm->locked_vm += npage; | |
161 | up_write(¤t->mm->mmap_sem); | |
162 | return; | |
163 | } | |
164 | ||
165 | /* | |
166 | * Couldn't get mmap_sem lock, so must setup to update | |
167 | * mm->locked_vm later. If locked_vm were atomic, we | |
168 | * wouldn't need this silliness | |
169 | */ | |
170 | vwork = kmalloc(sizeof(struct vwork), GFP_KERNEL); | |
171 | if (!vwork) | |
172 | return; | |
173 | mm = get_task_mm(current); | |
174 | if (!mm) { | |
175 | kfree(vwork); | |
176 | return; | |
177 | } | |
178 | INIT_WORK(&vwork->work, vfio_lock_acct_bg); | |
179 | vwork->mm = mm; | |
180 | vwork->npage = npage; | |
181 | schedule_work(&vwork->work); | |
182 | } | |
183 | ||
184 | /* | |
185 | * Some mappings aren't backed by a struct page, for example an mmap'd | |
186 | * MMIO range for our own or another device. These use a different | |
187 | * pfn conversion and shouldn't be tracked as locked pages. | |
188 | */ | |
189 | static bool is_invalid_reserved_pfn(unsigned long pfn) | |
190 | { | |
191 | if (pfn_valid(pfn)) { | |
192 | bool reserved; | |
193 | struct page *tail = pfn_to_page(pfn); | |
668f9abb | 194 | struct page *head = compound_head(tail); |
73fa0d10 AW |
195 | reserved = !!(PageReserved(head)); |
196 | if (head != tail) { | |
197 | /* | |
198 | * "head" is not a dangling pointer | |
668f9abb | 199 | * (compound_head takes care of that) |
73fa0d10 AW |
200 | * but the hugepage may have been split |
201 | * from under us (and we may not hold a | |
202 | * reference count on the head page so it can | |
203 | * be reused before we run PageReferenced), so | |
204 | * we've to check PageTail before returning | |
205 | * what we just read. | |
206 | */ | |
207 | smp_rmb(); | |
208 | if (PageTail(tail)) | |
209 | return reserved; | |
210 | } | |
211 | return PageReserved(tail); | |
212 | } | |
213 | ||
214 | return true; | |
215 | } | |
216 | ||
217 | static int put_pfn(unsigned long pfn, int prot) | |
218 | { | |
219 | if (!is_invalid_reserved_pfn(pfn)) { | |
220 | struct page *page = pfn_to_page(pfn); | |
221 | if (prot & IOMMU_WRITE) | |
222 | SetPageDirty(page); | |
223 | put_page(page); | |
224 | return 1; | |
225 | } | |
226 | return 0; | |
227 | } | |
228 | ||
73fa0d10 AW |
229 | static int vaddr_get_pfn(unsigned long vaddr, int prot, unsigned long *pfn) |
230 | { | |
231 | struct page *page[1]; | |
232 | struct vm_area_struct *vma; | |
233 | int ret = -EFAULT; | |
234 | ||
235 | if (get_user_pages_fast(vaddr, 1, !!(prot & IOMMU_WRITE), page) == 1) { | |
236 | *pfn = page_to_pfn(page[0]); | |
237 | return 0; | |
238 | } | |
239 | ||
240 | down_read(¤t->mm->mmap_sem); | |
241 | ||
242 | vma = find_vma_intersection(current->mm, vaddr, vaddr + 1); | |
243 | ||
244 | if (vma && vma->vm_flags & VM_PFNMAP) { | |
245 | *pfn = ((vaddr - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff; | |
246 | if (is_invalid_reserved_pfn(*pfn)) | |
247 | ret = 0; | |
248 | } | |
249 | ||
250 | up_read(¤t->mm->mmap_sem); | |
251 | ||
252 | return ret; | |
253 | } | |
254 | ||
166fd7d9 AW |
255 | /* |
256 | * Attempt to pin pages. We really don't want to track all the pfns and | |
257 | * the iommu can only map chunks of consecutive pfns anyway, so get the | |
258 | * first page and all consecutive pages with the same locking. | |
259 | */ | |
260 | static long vfio_pin_pages(unsigned long vaddr, long npage, | |
261 | int prot, unsigned long *pfn_base) | |
73fa0d10 | 262 | { |
166fd7d9 AW |
263 | unsigned long limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT; |
264 | bool lock_cap = capable(CAP_IPC_LOCK); | |
265 | long ret, i; | |
73fa0d10 | 266 | |
166fd7d9 AW |
267 | if (!current->mm) |
268 | return -ENODEV; | |
73fa0d10 | 269 | |
166fd7d9 AW |
270 | ret = vaddr_get_pfn(vaddr, prot, pfn_base); |
271 | if (ret) | |
272 | return ret; | |
73fa0d10 | 273 | |
166fd7d9 AW |
274 | if (is_invalid_reserved_pfn(*pfn_base)) |
275 | return 1; | |
73fa0d10 | 276 | |
166fd7d9 AW |
277 | if (!lock_cap && current->mm->locked_vm + 1 > limit) { |
278 | put_pfn(*pfn_base, prot); | |
279 | pr_warn("%s: RLIMIT_MEMLOCK (%ld) exceeded\n", __func__, | |
280 | limit << PAGE_SHIFT); | |
281 | return -ENOMEM; | |
282 | } | |
283 | ||
5c6c2b21 AW |
284 | if (unlikely(disable_hugepages)) { |
285 | vfio_lock_acct(1); | |
286 | return 1; | |
287 | } | |
288 | ||
166fd7d9 AW |
289 | /* Lock all the consecutive pages from pfn_base */ |
290 | for (i = 1, vaddr += PAGE_SIZE; i < npage; i++, vaddr += PAGE_SIZE) { | |
73fa0d10 AW |
291 | unsigned long pfn = 0; |
292 | ||
293 | ret = vaddr_get_pfn(vaddr, prot, &pfn); | |
166fd7d9 AW |
294 | if (ret) |
295 | break; | |
296 | ||
297 | if (pfn != *pfn_base + i || is_invalid_reserved_pfn(pfn)) { | |
298 | put_pfn(pfn, prot); | |
299 | break; | |
73fa0d10 AW |
300 | } |
301 | ||
166fd7d9 AW |
302 | if (!lock_cap && current->mm->locked_vm + i + 1 > limit) { |
303 | put_pfn(pfn, prot); | |
304 | pr_warn("%s: RLIMIT_MEMLOCK (%ld) exceeded\n", | |
305 | __func__, limit << PAGE_SHIFT); | |
306 | break; | |
307 | } | |
308 | } | |
309 | ||
310 | vfio_lock_acct(i); | |
311 | ||
312 | return i; | |
313 | } | |
314 | ||
315 | static long vfio_unpin_pages(unsigned long pfn, long npage, | |
316 | int prot, bool do_accounting) | |
317 | { | |
318 | unsigned long unlocked = 0; | |
319 | long i; | |
320 | ||
321 | for (i = 0; i < npage; i++) | |
322 | unlocked += put_pfn(pfn++, prot); | |
323 | ||
324 | if (do_accounting) | |
325 | vfio_lock_acct(-unlocked); | |
326 | ||
327 | return unlocked; | |
328 | } | |
329 | ||
1ef3e2bc | 330 | static void vfio_unmap_unpin(struct vfio_iommu *iommu, struct vfio_dma *dma) |
166fd7d9 | 331 | { |
1ef3e2bc AW |
332 | dma_addr_t iova = dma->iova, end = dma->iova + dma->size; |
333 | struct vfio_domain *domain, *d; | |
166fd7d9 AW |
334 | long unlocked = 0; |
335 | ||
1ef3e2bc AW |
336 | if (!dma->size) |
337 | return; | |
338 | /* | |
339 | * We use the IOMMU to track the physical addresses, otherwise we'd | |
340 | * need a much more complicated tracking system. Unfortunately that | |
341 | * means we need to use one of the iommu domains to figure out the | |
342 | * pfns to unpin. The rest need to be unmapped in advance so we have | |
343 | * no iommu translations remaining when the pages are unpinned. | |
344 | */ | |
345 | domain = d = list_first_entry(&iommu->domain_list, | |
346 | struct vfio_domain, next); | |
347 | ||
348 | list_for_each_entry_continue(d, &iommu->domain_list, next) | |
349 | iommu_unmap(d->domain, dma->iova, dma->size); | |
350 | ||
166fd7d9 AW |
351 | while (iova < end) { |
352 | size_t unmapped; | |
353 | phys_addr_t phys; | |
354 | ||
1ef3e2bc | 355 | phys = iommu_iova_to_phys(domain->domain, iova); |
166fd7d9 AW |
356 | if (WARN_ON(!phys)) { |
357 | iova += PAGE_SIZE; | |
358 | continue; | |
73fa0d10 | 359 | } |
166fd7d9 | 360 | |
1ef3e2bc AW |
361 | unmapped = iommu_unmap(domain->domain, iova, PAGE_SIZE); |
362 | if (WARN_ON(!unmapped)) | |
166fd7d9 AW |
363 | break; |
364 | ||
365 | unlocked += vfio_unpin_pages(phys >> PAGE_SHIFT, | |
366 | unmapped >> PAGE_SHIFT, | |
367 | dma->prot, false); | |
368 | iova += unmapped; | |
73fa0d10 | 369 | } |
166fd7d9 AW |
370 | |
371 | vfio_lock_acct(-unlocked); | |
73fa0d10 AW |
372 | } |
373 | ||
1ef3e2bc | 374 | static void vfio_remove_dma(struct vfio_iommu *iommu, struct vfio_dma *dma) |
73fa0d10 | 375 | { |
1ef3e2bc AW |
376 | vfio_unmap_unpin(iommu, dma); |
377 | vfio_unlink_dma(iommu, dma); | |
378 | kfree(dma); | |
379 | } | |
73fa0d10 | 380 | |
1ef3e2bc AW |
381 | static unsigned long vfio_pgsize_bitmap(struct vfio_iommu *iommu) |
382 | { | |
383 | struct vfio_domain *domain; | |
384 | unsigned long bitmap = PAGE_MASK; | |
166fd7d9 | 385 | |
1ef3e2bc AW |
386 | mutex_lock(&iommu->lock); |
387 | list_for_each_entry(domain, &iommu->domain_list, next) | |
388 | bitmap &= domain->domain->ops->pgsize_bitmap; | |
389 | mutex_unlock(&iommu->lock); | |
73fa0d10 | 390 | |
1ef3e2bc | 391 | return bitmap; |
73fa0d10 AW |
392 | } |
393 | ||
394 | static int vfio_dma_do_unmap(struct vfio_iommu *iommu, | |
395 | struct vfio_iommu_type1_dma_unmap *unmap) | |
396 | { | |
73fa0d10 | 397 | uint64_t mask; |
cd9b2268 | 398 | struct vfio_dma *dma; |
1ef3e2bc | 399 | size_t unmapped = 0; |
cd9b2268 | 400 | int ret = 0; |
73fa0d10 | 401 | |
1ef3e2bc | 402 | mask = ((uint64_t)1 << __ffs(vfio_pgsize_bitmap(iommu))) - 1; |
73fa0d10 AW |
403 | |
404 | if (unmap->iova & mask) | |
405 | return -EINVAL; | |
f5bfdbf2 | 406 | if (!unmap->size || unmap->size & mask) |
73fa0d10 AW |
407 | return -EINVAL; |
408 | ||
73fa0d10 AW |
409 | WARN_ON(mask & PAGE_MASK); |
410 | ||
411 | mutex_lock(&iommu->lock); | |
412 | ||
1ef3e2bc AW |
413 | /* |
414 | * vfio-iommu-type1 (v1) - User mappings were coalesced together to | |
415 | * avoid tracking individual mappings. This means that the granularity | |
416 | * of the original mapping was lost and the user was allowed to attempt | |
417 | * to unmap any range. Depending on the contiguousness of physical | |
418 | * memory and page sizes supported by the IOMMU, arbitrary unmaps may | |
419 | * or may not have worked. We only guaranteed unmap granularity | |
420 | * matching the original mapping; even though it was untracked here, | |
421 | * the original mappings are reflected in IOMMU mappings. This | |
422 | * resulted in a couple unusual behaviors. First, if a range is not | |
423 | * able to be unmapped, ex. a set of 4k pages that was mapped as a | |
424 | * 2M hugepage into the IOMMU, the unmap ioctl returns success but with | |
425 | * a zero sized unmap. Also, if an unmap request overlaps the first | |
426 | * address of a hugepage, the IOMMU will unmap the entire hugepage. | |
427 | * This also returns success and the returned unmap size reflects the | |
428 | * actual size unmapped. | |
429 | * | |
430 | * We attempt to maintain compatibility with this "v1" interface, but | |
431 | * we take control out of the hands of the IOMMU. Therefore, an unmap | |
432 | * request offset from the beginning of the original mapping will | |
433 | * return success with zero sized unmap. And an unmap request covering | |
434 | * the first iova of mapping will unmap the entire range. | |
435 | * | |
436 | * The v2 version of this interface intends to be more deterministic. | |
437 | * Unmap requests must fully cover previous mappings. Multiple | |
438 | * mappings may still be unmaped by specifying large ranges, but there | |
439 | * must not be any previous mappings bisected by the range. An error | |
440 | * will be returned if these conditions are not met. The v2 interface | |
441 | * will only return success and a size of zero if there were no | |
442 | * mappings within the range. | |
443 | */ | |
444 | if (iommu->v2) { | |
445 | dma = vfio_find_dma(iommu, unmap->iova, 0); | |
446 | if (dma && dma->iova != unmap->iova) { | |
447 | ret = -EINVAL; | |
448 | goto unlock; | |
449 | } | |
450 | dma = vfio_find_dma(iommu, unmap->iova + unmap->size - 1, 0); | |
451 | if (dma && dma->iova + dma->size != unmap->iova + unmap->size) { | |
452 | ret = -EINVAL; | |
453 | goto unlock; | |
454 | } | |
455 | } | |
456 | ||
166fd7d9 | 457 | while ((dma = vfio_find_dma(iommu, unmap->iova, unmap->size))) { |
1ef3e2bc | 458 | if (!iommu->v2 && unmap->iova > dma->iova) |
166fd7d9 | 459 | break; |
1ef3e2bc AW |
460 | unmapped += dma->size; |
461 | vfio_remove_dma(iommu, dma); | |
166fd7d9 | 462 | } |
cd9b2268 | 463 | |
1ef3e2bc | 464 | unlock: |
73fa0d10 | 465 | mutex_unlock(&iommu->lock); |
166fd7d9 | 466 | |
1ef3e2bc | 467 | /* Report how much was unmapped */ |
166fd7d9 AW |
468 | unmap->size = unmapped; |
469 | ||
470 | return ret; | |
471 | } | |
472 | ||
473 | /* | |
474 | * Turns out AMD IOMMU has a page table bug where it won't map large pages | |
475 | * to a region that previously mapped smaller pages. This should be fixed | |
476 | * soon, so this is just a temporary workaround to break mappings down into | |
477 | * PAGE_SIZE. Better to map smaller pages than nothing. | |
478 | */ | |
1ef3e2bc | 479 | static int map_try_harder(struct vfio_domain *domain, dma_addr_t iova, |
166fd7d9 AW |
480 | unsigned long pfn, long npage, int prot) |
481 | { | |
482 | long i; | |
483 | int ret; | |
484 | ||
485 | for (i = 0; i < npage; i++, pfn++, iova += PAGE_SIZE) { | |
1ef3e2bc | 486 | ret = iommu_map(domain->domain, iova, |
166fd7d9 | 487 | (phys_addr_t)pfn << PAGE_SHIFT, |
1ef3e2bc | 488 | PAGE_SIZE, prot | domain->prot); |
166fd7d9 AW |
489 | if (ret) |
490 | break; | |
491 | } | |
492 | ||
493 | for (; i < npage && i > 0; i--, iova -= PAGE_SIZE) | |
1ef3e2bc AW |
494 | iommu_unmap(domain->domain, iova, PAGE_SIZE); |
495 | ||
496 | return ret; | |
497 | } | |
498 | ||
499 | static int vfio_iommu_map(struct vfio_iommu *iommu, dma_addr_t iova, | |
500 | unsigned long pfn, long npage, int prot) | |
501 | { | |
502 | struct vfio_domain *d; | |
503 | int ret; | |
504 | ||
505 | list_for_each_entry(d, &iommu->domain_list, next) { | |
506 | ret = iommu_map(d->domain, iova, (phys_addr_t)pfn << PAGE_SHIFT, | |
507 | npage << PAGE_SHIFT, prot | d->prot); | |
508 | if (ret) { | |
509 | if (ret != -EBUSY || | |
510 | map_try_harder(d, iova, pfn, npage, prot)) | |
511 | goto unwind; | |
512 | } | |
513 | } | |
514 | ||
515 | return 0; | |
516 | ||
517 | unwind: | |
518 | list_for_each_entry_continue_reverse(d, &iommu->domain_list, next) | |
519 | iommu_unmap(d->domain, iova, npage << PAGE_SHIFT); | |
166fd7d9 | 520 | |
cd9b2268 | 521 | return ret; |
73fa0d10 AW |
522 | } |
523 | ||
524 | static int vfio_dma_do_map(struct vfio_iommu *iommu, | |
525 | struct vfio_iommu_type1_dma_map *map) | |
526 | { | |
c8dbca16 | 527 | dma_addr_t iova = map->iova; |
166fd7d9 | 528 | unsigned long vaddr = map->vaddr; |
73fa0d10 | 529 | size_t size = map->size; |
166fd7d9 | 530 | long npage; |
73fa0d10 AW |
531 | int ret = 0, prot = 0; |
532 | uint64_t mask; | |
1ef3e2bc | 533 | struct vfio_dma *dma; |
d93b3ac0 | 534 | unsigned long pfn; |
166fd7d9 | 535 | |
c8dbca16 AW |
536 | /* Verify that none of our __u64 fields overflow */ |
537 | if (map->size != size || map->vaddr != vaddr || map->iova != iova) | |
538 | return -EINVAL; | |
73fa0d10 | 539 | |
1ef3e2bc | 540 | mask = ((uint64_t)1 << __ffs(vfio_pgsize_bitmap(iommu))) - 1; |
73fa0d10 | 541 | |
c8dbca16 AW |
542 | WARN_ON(mask & PAGE_MASK); |
543 | ||
73fa0d10 AW |
544 | /* READ/WRITE from device perspective */ |
545 | if (map->flags & VFIO_DMA_MAP_FLAG_WRITE) | |
546 | prot |= IOMMU_WRITE; | |
547 | if (map->flags & VFIO_DMA_MAP_FLAG_READ) | |
548 | prot |= IOMMU_READ; | |
549 | ||
c8dbca16 | 550 | if (!prot || !size || (size | iova | vaddr) & mask) |
73fa0d10 AW |
551 | return -EINVAL; |
552 | ||
c8dbca16 AW |
553 | /* Don't allow IOVA or virtual address wrap */ |
554 | if (iova + size - 1 < iova || vaddr + size - 1 < vaddr) | |
73fa0d10 AW |
555 | return -EINVAL; |
556 | ||
557 | mutex_lock(&iommu->lock); | |
558 | ||
c8dbca16 | 559 | if (vfio_find_dma(iommu, iova, size)) { |
166fd7d9 AW |
560 | mutex_unlock(&iommu->lock); |
561 | return -EEXIST; | |
73fa0d10 AW |
562 | } |
563 | ||
1ef3e2bc AW |
564 | dma = kzalloc(sizeof(*dma), GFP_KERNEL); |
565 | if (!dma) { | |
566 | mutex_unlock(&iommu->lock); | |
567 | return -ENOMEM; | |
568 | } | |
569 | ||
c8dbca16 AW |
570 | dma->iova = iova; |
571 | dma->vaddr = vaddr; | |
1ef3e2bc | 572 | dma->prot = prot; |
166fd7d9 | 573 | |
1ef3e2bc AW |
574 | /* Insert zero-sized and grow as we map chunks of it */ |
575 | vfio_link_dma(iommu, dma); | |
166fd7d9 | 576 | |
c8dbca16 | 577 | while (size) { |
166fd7d9 | 578 | /* Pin a contiguous chunk of memory */ |
c8dbca16 AW |
579 | npage = vfio_pin_pages(vaddr + dma->size, |
580 | size >> PAGE_SHIFT, prot, &pfn); | |
166fd7d9 AW |
581 | if (npage <= 0) { |
582 | WARN_ON(!npage); | |
583 | ret = (int)npage; | |
1ef3e2bc | 584 | break; |
166fd7d9 AW |
585 | } |
586 | ||
1ef3e2bc | 587 | /* Map it! */ |
c8dbca16 | 588 | ret = vfio_iommu_map(iommu, iova + dma->size, pfn, npage, prot); |
166fd7d9 | 589 | if (ret) { |
1ef3e2bc AW |
590 | vfio_unpin_pages(pfn, npage, prot, true); |
591 | break; | |
166fd7d9 AW |
592 | } |
593 | ||
c8dbca16 AW |
594 | size -= npage << PAGE_SHIFT; |
595 | dma->size += npage << PAGE_SHIFT; | |
1ef3e2bc | 596 | } |
166fd7d9 | 597 | |
1ef3e2bc AW |
598 | if (ret) |
599 | vfio_remove_dma(iommu, dma); | |
166fd7d9 | 600 | |
1ef3e2bc AW |
601 | mutex_unlock(&iommu->lock); |
602 | return ret; | |
603 | } | |
604 | ||
605 | static int vfio_bus_type(struct device *dev, void *data) | |
606 | { | |
607 | struct bus_type **bus = data; | |
608 | ||
609 | if (*bus && *bus != dev->bus) | |
610 | return -EINVAL; | |
611 | ||
612 | *bus = dev->bus; | |
613 | ||
614 | return 0; | |
615 | } | |
616 | ||
617 | static int vfio_iommu_replay(struct vfio_iommu *iommu, | |
618 | struct vfio_domain *domain) | |
619 | { | |
620 | struct vfio_domain *d; | |
621 | struct rb_node *n; | |
622 | int ret; | |
623 | ||
624 | /* Arbitrarily pick the first domain in the list for lookups */ | |
625 | d = list_first_entry(&iommu->domain_list, struct vfio_domain, next); | |
626 | n = rb_first(&iommu->dma_list); | |
627 | ||
628 | /* If there's not a domain, there better not be any mappings */ | |
629 | if (WARN_ON(n && !d)) | |
630 | return -EINVAL; | |
631 | ||
632 | for (; n; n = rb_next(n)) { | |
633 | struct vfio_dma *dma; | |
634 | dma_addr_t iova; | |
635 | ||
636 | dma = rb_entry(n, struct vfio_dma, node); | |
637 | iova = dma->iova; | |
638 | ||
639 | while (iova < dma->iova + dma->size) { | |
640 | phys_addr_t phys = iommu_iova_to_phys(d->domain, iova); | |
641 | size_t size; | |
73fa0d10 | 642 | |
1ef3e2bc AW |
643 | if (WARN_ON(!phys)) { |
644 | iova += PAGE_SIZE; | |
645 | continue; | |
166fd7d9 AW |
646 | } |
647 | ||
1ef3e2bc | 648 | size = PAGE_SIZE; |
73fa0d10 | 649 | |
1ef3e2bc AW |
650 | while (iova + size < dma->iova + dma->size && |
651 | phys + size == iommu_iova_to_phys(d->domain, | |
652 | iova + size)) | |
653 | size += PAGE_SIZE; | |
d93b3ac0 | 654 | |
1ef3e2bc AW |
655 | ret = iommu_map(domain->domain, iova, phys, |
656 | size, dma->prot | domain->prot); | |
657 | if (ret) | |
658 | return ret; | |
d93b3ac0 | 659 | |
1ef3e2bc AW |
660 | iova += size; |
661 | } | |
166fd7d9 | 662 | } |
73fa0d10 | 663 | |
1ef3e2bc | 664 | return 0; |
73fa0d10 AW |
665 | } |
666 | ||
667 | static int vfio_iommu_type1_attach_group(void *iommu_data, | |
668 | struct iommu_group *iommu_group) | |
669 | { | |
670 | struct vfio_iommu *iommu = iommu_data; | |
1ef3e2bc AW |
671 | struct vfio_group *group, *g; |
672 | struct vfio_domain *domain, *d; | |
673 | struct bus_type *bus = NULL; | |
73fa0d10 AW |
674 | int ret; |
675 | ||
73fa0d10 AW |
676 | mutex_lock(&iommu->lock); |
677 | ||
1ef3e2bc AW |
678 | list_for_each_entry(d, &iommu->domain_list, next) { |
679 | list_for_each_entry(g, &d->group_list, next) { | |
680 | if (g->iommu_group != iommu_group) | |
681 | continue; | |
682 | ||
73fa0d10 | 683 | mutex_unlock(&iommu->lock); |
73fa0d10 AW |
684 | return -EINVAL; |
685 | } | |
686 | } | |
687 | ||
1ef3e2bc AW |
688 | group = kzalloc(sizeof(*group), GFP_KERNEL); |
689 | domain = kzalloc(sizeof(*domain), GFP_KERNEL); | |
690 | if (!group || !domain) { | |
691 | ret = -ENOMEM; | |
692 | goto out_free; | |
693 | } | |
694 | ||
695 | group->iommu_group = iommu_group; | |
696 | ||
697 | /* Determine bus_type in order to allocate a domain */ | |
698 | ret = iommu_group_for_each_dev(iommu_group, &bus, vfio_bus_type); | |
699 | if (ret) | |
700 | goto out_free; | |
701 | ||
702 | domain->domain = iommu_domain_alloc(bus); | |
703 | if (!domain->domain) { | |
704 | ret = -EIO; | |
705 | goto out_free; | |
706 | } | |
707 | ||
708 | ret = iommu_attach_group(domain->domain, iommu_group); | |
709 | if (ret) | |
710 | goto out_domain; | |
711 | ||
712 | INIT_LIST_HEAD(&domain->group_list); | |
713 | list_add(&group->next, &domain->group_list); | |
714 | ||
715 | if (!allow_unsafe_interrupts && | |
716 | !iommu_domain_has_cap(domain->domain, IOMMU_CAP_INTR_REMAP)) { | |
717 | pr_warn("%s: No interrupt remapping support. Use the module param \"allow_unsafe_interrupts\" to enable VFIO IOMMU support on this platform\n", | |
718 | __func__); | |
719 | ret = -EPERM; | |
720 | goto out_detach; | |
721 | } | |
722 | ||
723 | if (iommu_domain_has_cap(domain->domain, IOMMU_CAP_CACHE_COHERENCY)) | |
724 | domain->prot |= IOMMU_CACHE; | |
725 | ||
73fa0d10 | 726 | /* |
1ef3e2bc AW |
727 | * Try to match an existing compatible domain. We don't want to |
728 | * preclude an IOMMU driver supporting multiple bus_types and being | |
729 | * able to include different bus_types in the same IOMMU domain, so | |
730 | * we test whether the domains use the same iommu_ops rather than | |
731 | * testing if they're on the same bus_type. | |
73fa0d10 | 732 | */ |
1ef3e2bc AW |
733 | list_for_each_entry(d, &iommu->domain_list, next) { |
734 | if (d->domain->ops == domain->domain->ops && | |
735 | d->prot == domain->prot) { | |
736 | iommu_detach_group(domain->domain, iommu_group); | |
737 | if (!iommu_attach_group(d->domain, iommu_group)) { | |
738 | list_add(&group->next, &d->group_list); | |
739 | iommu_domain_free(domain->domain); | |
740 | kfree(domain); | |
741 | mutex_unlock(&iommu->lock); | |
742 | return 0; | |
743 | } | |
744 | ||
745 | ret = iommu_attach_group(domain->domain, iommu_group); | |
746 | if (ret) | |
747 | goto out_domain; | |
748 | } | |
73fa0d10 AW |
749 | } |
750 | ||
1ef3e2bc AW |
751 | /* replay mappings on new domains */ |
752 | ret = vfio_iommu_replay(iommu, domain); | |
753 | if (ret) | |
754 | goto out_detach; | |
755 | ||
756 | list_add(&domain->next, &iommu->domain_list); | |
73fa0d10 AW |
757 | |
758 | mutex_unlock(&iommu->lock); | |
759 | ||
760 | return 0; | |
1ef3e2bc AW |
761 | |
762 | out_detach: | |
763 | iommu_detach_group(domain->domain, iommu_group); | |
764 | out_domain: | |
765 | iommu_domain_free(domain->domain); | |
766 | out_free: | |
767 | kfree(domain); | |
768 | kfree(group); | |
769 | mutex_unlock(&iommu->lock); | |
770 | return ret; | |
771 | } | |
772 | ||
773 | static void vfio_iommu_unmap_unpin_all(struct vfio_iommu *iommu) | |
774 | { | |
775 | struct rb_node *node; | |
776 | ||
777 | while ((node = rb_first(&iommu->dma_list))) | |
778 | vfio_remove_dma(iommu, rb_entry(node, struct vfio_dma, node)); | |
73fa0d10 AW |
779 | } |
780 | ||
781 | static void vfio_iommu_type1_detach_group(void *iommu_data, | |
782 | struct iommu_group *iommu_group) | |
783 | { | |
784 | struct vfio_iommu *iommu = iommu_data; | |
1ef3e2bc | 785 | struct vfio_domain *domain; |
73fa0d10 AW |
786 | struct vfio_group *group; |
787 | ||
788 | mutex_lock(&iommu->lock); | |
789 | ||
1ef3e2bc AW |
790 | list_for_each_entry(domain, &iommu->domain_list, next) { |
791 | list_for_each_entry(group, &domain->group_list, next) { | |
792 | if (group->iommu_group != iommu_group) | |
793 | continue; | |
794 | ||
795 | iommu_detach_group(domain->domain, iommu_group); | |
73fa0d10 AW |
796 | list_del(&group->next); |
797 | kfree(group); | |
1ef3e2bc AW |
798 | /* |
799 | * Group ownership provides privilege, if the group | |
800 | * list is empty, the domain goes away. If it's the | |
801 | * last domain, then all the mappings go away too. | |
802 | */ | |
803 | if (list_empty(&domain->group_list)) { | |
804 | if (list_is_singular(&iommu->domain_list)) | |
805 | vfio_iommu_unmap_unpin_all(iommu); | |
806 | iommu_domain_free(domain->domain); | |
807 | list_del(&domain->next); | |
808 | kfree(domain); | |
809 | } | |
810 | goto done; | |
73fa0d10 AW |
811 | } |
812 | } | |
813 | ||
1ef3e2bc | 814 | done: |
73fa0d10 AW |
815 | mutex_unlock(&iommu->lock); |
816 | } | |
817 | ||
818 | static void *vfio_iommu_type1_open(unsigned long arg) | |
819 | { | |
820 | struct vfio_iommu *iommu; | |
821 | ||
1ef3e2bc | 822 | if (arg != VFIO_TYPE1_IOMMU && arg != VFIO_TYPE1v2_IOMMU) |
73fa0d10 AW |
823 | return ERR_PTR(-EINVAL); |
824 | ||
825 | iommu = kzalloc(sizeof(*iommu), GFP_KERNEL); | |
826 | if (!iommu) | |
827 | return ERR_PTR(-ENOMEM); | |
828 | ||
1ef3e2bc | 829 | INIT_LIST_HEAD(&iommu->domain_list); |
cd9b2268 | 830 | iommu->dma_list = RB_ROOT; |
73fa0d10 | 831 | mutex_init(&iommu->lock); |
1ef3e2bc | 832 | iommu->v2 = (arg == VFIO_TYPE1v2_IOMMU); |
73fa0d10 AW |
833 | |
834 | return iommu; | |
835 | } | |
836 | ||
837 | static void vfio_iommu_type1_release(void *iommu_data) | |
838 | { | |
839 | struct vfio_iommu *iommu = iommu_data; | |
1ef3e2bc | 840 | struct vfio_domain *domain, *domain_tmp; |
73fa0d10 | 841 | struct vfio_group *group, *group_tmp; |
73fa0d10 | 842 | |
1ef3e2bc | 843 | vfio_iommu_unmap_unpin_all(iommu); |
73fa0d10 | 844 | |
1ef3e2bc AW |
845 | list_for_each_entry_safe(domain, domain_tmp, |
846 | &iommu->domain_list, next) { | |
847 | list_for_each_entry_safe(group, group_tmp, | |
848 | &domain->group_list, next) { | |
849 | iommu_detach_group(domain->domain, group->iommu_group); | |
850 | list_del(&group->next); | |
851 | kfree(group); | |
852 | } | |
853 | iommu_domain_free(domain->domain); | |
854 | list_del(&domain->next); | |
855 | kfree(domain); | |
73fa0d10 AW |
856 | } |
857 | ||
73fa0d10 AW |
858 | kfree(iommu); |
859 | } | |
860 | ||
aa429318 AW |
861 | static int vfio_domains_have_iommu_cache(struct vfio_iommu *iommu) |
862 | { | |
863 | struct vfio_domain *domain; | |
864 | int ret = 1; | |
865 | ||
866 | mutex_lock(&iommu->lock); | |
867 | list_for_each_entry(domain, &iommu->domain_list, next) { | |
868 | if (!(domain->prot & IOMMU_CACHE)) { | |
869 | ret = 0; | |
f5bfdbf2 | 870 | break; |
aa429318 | 871 | } |
73fa0d10 | 872 | } |
aa429318 | 873 | mutex_unlock(&iommu->lock); |
73fa0d10 | 874 | |
aa429318 | 875 | return ret; |
73fa0d10 AW |
876 | } |
877 | ||
878 | static long vfio_iommu_type1_ioctl(void *iommu_data, | |
879 | unsigned int cmd, unsigned long arg) | |
880 | { | |
881 | struct vfio_iommu *iommu = iommu_data; | |
882 | unsigned long minsz; | |
883 | ||
884 | if (cmd == VFIO_CHECK_EXTENSION) { | |
885 | switch (arg) { | |
886 | case VFIO_TYPE1_IOMMU: | |
1ef3e2bc | 887 | case VFIO_TYPE1v2_IOMMU: |
73fa0d10 | 888 | return 1; |
aa429318 AW |
889 | case VFIO_DMA_CC_IOMMU: |
890 | if (!iommu) | |
891 | return 0; | |
892 | return vfio_domains_have_iommu_cache(iommu); | |
73fa0d10 AW |
893 | default: |
894 | return 0; | |
895 | } | |
896 | } else if (cmd == VFIO_IOMMU_GET_INFO) { | |
897 | struct vfio_iommu_type1_info info; | |
898 | ||
899 | minsz = offsetofend(struct vfio_iommu_type1_info, iova_pgsizes); | |
900 | ||
901 | if (copy_from_user(&info, (void __user *)arg, minsz)) | |
902 | return -EFAULT; | |
903 | ||
904 | if (info.argsz < minsz) | |
905 | return -EINVAL; | |
906 | ||
907 | info.flags = 0; | |
908 | ||
1ef3e2bc | 909 | info.iova_pgsizes = vfio_pgsize_bitmap(iommu); |
73fa0d10 AW |
910 | |
911 | return copy_to_user((void __user *)arg, &info, minsz); | |
912 | ||
913 | } else if (cmd == VFIO_IOMMU_MAP_DMA) { | |
914 | struct vfio_iommu_type1_dma_map map; | |
915 | uint32_t mask = VFIO_DMA_MAP_FLAG_READ | | |
916 | VFIO_DMA_MAP_FLAG_WRITE; | |
917 | ||
918 | minsz = offsetofend(struct vfio_iommu_type1_dma_map, size); | |
919 | ||
920 | if (copy_from_user(&map, (void __user *)arg, minsz)) | |
921 | return -EFAULT; | |
922 | ||
923 | if (map.argsz < minsz || map.flags & ~mask) | |
924 | return -EINVAL; | |
925 | ||
926 | return vfio_dma_do_map(iommu, &map); | |
927 | ||
928 | } else if (cmd == VFIO_IOMMU_UNMAP_DMA) { | |
929 | struct vfio_iommu_type1_dma_unmap unmap; | |
166fd7d9 | 930 | long ret; |
73fa0d10 AW |
931 | |
932 | minsz = offsetofend(struct vfio_iommu_type1_dma_unmap, size); | |
933 | ||
934 | if (copy_from_user(&unmap, (void __user *)arg, minsz)) | |
935 | return -EFAULT; | |
936 | ||
937 | if (unmap.argsz < minsz || unmap.flags) | |
938 | return -EINVAL; | |
939 | ||
166fd7d9 AW |
940 | ret = vfio_dma_do_unmap(iommu, &unmap); |
941 | if (ret) | |
942 | return ret; | |
943 | ||
944 | return copy_to_user((void __user *)arg, &unmap, minsz); | |
73fa0d10 AW |
945 | } |
946 | ||
947 | return -ENOTTY; | |
948 | } | |
949 | ||
950 | static const struct vfio_iommu_driver_ops vfio_iommu_driver_ops_type1 = { | |
951 | .name = "vfio-iommu-type1", | |
952 | .owner = THIS_MODULE, | |
953 | .open = vfio_iommu_type1_open, | |
954 | .release = vfio_iommu_type1_release, | |
955 | .ioctl = vfio_iommu_type1_ioctl, | |
956 | .attach_group = vfio_iommu_type1_attach_group, | |
957 | .detach_group = vfio_iommu_type1_detach_group, | |
958 | }; | |
959 | ||
960 | static int __init vfio_iommu_type1_init(void) | |
961 | { | |
73fa0d10 AW |
962 | return vfio_register_iommu_driver(&vfio_iommu_driver_ops_type1); |
963 | } | |
964 | ||
965 | static void __exit vfio_iommu_type1_cleanup(void) | |
966 | { | |
967 | vfio_unregister_iommu_driver(&vfio_iommu_driver_ops_type1); | |
968 | } | |
969 | ||
970 | module_init(vfio_iommu_type1_init); | |
971 | module_exit(vfio_iommu_type1_cleanup); | |
972 | ||
973 | MODULE_VERSION(DRIVER_VERSION); | |
974 | MODULE_LICENSE("GPL v2"); | |
975 | MODULE_AUTHOR(DRIVER_AUTHOR); | |
976 | MODULE_DESCRIPTION(DRIVER_DESC); |