]>
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; |
f5c9eceb WD |
60 | bool v2; |
61 | bool nesting; | |
1ef3e2bc AW |
62 | }; |
63 | ||
64 | struct vfio_domain { | |
65 | struct iommu_domain *domain; | |
66 | struct list_head next; | |
73fa0d10 | 67 | struct list_head group_list; |
1ef3e2bc | 68 | int prot; /* IOMMU_CACHE */ |
73fa0d10 AW |
69 | }; |
70 | ||
71 | struct vfio_dma { | |
cd9b2268 | 72 | struct rb_node node; |
73fa0d10 AW |
73 | dma_addr_t iova; /* Device address */ |
74 | unsigned long vaddr; /* Process virtual addr */ | |
166fd7d9 | 75 | size_t size; /* Map size (bytes) */ |
73fa0d10 AW |
76 | int prot; /* IOMMU_READ/WRITE */ |
77 | }; | |
78 | ||
79 | struct vfio_group { | |
80 | struct iommu_group *iommu_group; | |
81 | struct list_head next; | |
82 | }; | |
83 | ||
84 | /* | |
85 | * This code handles mapping and unmapping of user data buffers | |
86 | * into DMA'ble space using the IOMMU | |
87 | */ | |
88 | ||
cd9b2268 AW |
89 | static struct vfio_dma *vfio_find_dma(struct vfio_iommu *iommu, |
90 | dma_addr_t start, size_t size) | |
91 | { | |
92 | struct rb_node *node = iommu->dma_list.rb_node; | |
93 | ||
94 | while (node) { | |
95 | struct vfio_dma *dma = rb_entry(node, struct vfio_dma, node); | |
96 | ||
97 | if (start + size <= dma->iova) | |
98 | node = node->rb_left; | |
166fd7d9 | 99 | else if (start >= dma->iova + dma->size) |
cd9b2268 AW |
100 | node = node->rb_right; |
101 | else | |
102 | return dma; | |
103 | } | |
104 | ||
105 | return NULL; | |
106 | } | |
107 | ||
1ef3e2bc | 108 | static void vfio_link_dma(struct vfio_iommu *iommu, struct vfio_dma *new) |
cd9b2268 AW |
109 | { |
110 | struct rb_node **link = &iommu->dma_list.rb_node, *parent = NULL; | |
111 | struct vfio_dma *dma; | |
112 | ||
113 | while (*link) { | |
114 | parent = *link; | |
115 | dma = rb_entry(parent, struct vfio_dma, node); | |
116 | ||
166fd7d9 | 117 | if (new->iova + new->size <= dma->iova) |
cd9b2268 AW |
118 | link = &(*link)->rb_left; |
119 | else | |
120 | link = &(*link)->rb_right; | |
121 | } | |
122 | ||
123 | rb_link_node(&new->node, parent, link); | |
124 | rb_insert_color(&new->node, &iommu->dma_list); | |
125 | } | |
126 | ||
1ef3e2bc | 127 | static void vfio_unlink_dma(struct vfio_iommu *iommu, struct vfio_dma *old) |
cd9b2268 AW |
128 | { |
129 | rb_erase(&old->node, &iommu->dma_list); | |
130 | } | |
131 | ||
73fa0d10 AW |
132 | struct vwork { |
133 | struct mm_struct *mm; | |
134 | long npage; | |
135 | struct work_struct work; | |
136 | }; | |
137 | ||
138 | /* delayed decrement/increment for locked_vm */ | |
139 | static void vfio_lock_acct_bg(struct work_struct *work) | |
140 | { | |
141 | struct vwork *vwork = container_of(work, struct vwork, work); | |
142 | struct mm_struct *mm; | |
143 | ||
144 | mm = vwork->mm; | |
145 | down_write(&mm->mmap_sem); | |
146 | mm->locked_vm += vwork->npage; | |
147 | up_write(&mm->mmap_sem); | |
148 | mmput(mm); | |
149 | kfree(vwork); | |
150 | } | |
151 | ||
152 | static void vfio_lock_acct(long npage) | |
153 | { | |
154 | struct vwork *vwork; | |
155 | struct mm_struct *mm; | |
156 | ||
166fd7d9 AW |
157 | if (!current->mm || !npage) |
158 | return; /* process exited or nothing to do */ | |
73fa0d10 AW |
159 | |
160 | if (down_write_trylock(¤t->mm->mmap_sem)) { | |
161 | current->mm->locked_vm += npage; | |
162 | up_write(¤t->mm->mmap_sem); | |
163 | return; | |
164 | } | |
165 | ||
166 | /* | |
167 | * Couldn't get mmap_sem lock, so must setup to update | |
168 | * mm->locked_vm later. If locked_vm were atomic, we | |
169 | * wouldn't need this silliness | |
170 | */ | |
171 | vwork = kmalloc(sizeof(struct vwork), GFP_KERNEL); | |
172 | if (!vwork) | |
173 | return; | |
174 | mm = get_task_mm(current); | |
175 | if (!mm) { | |
176 | kfree(vwork); | |
177 | return; | |
178 | } | |
179 | INIT_WORK(&vwork->work, vfio_lock_acct_bg); | |
180 | vwork->mm = mm; | |
181 | vwork->npage = npage; | |
182 | schedule_work(&vwork->work); | |
183 | } | |
184 | ||
185 | /* | |
186 | * Some mappings aren't backed by a struct page, for example an mmap'd | |
187 | * MMIO range for our own or another device. These use a different | |
188 | * pfn conversion and shouldn't be tracked as locked pages. | |
189 | */ | |
190 | static bool is_invalid_reserved_pfn(unsigned long pfn) | |
191 | { | |
192 | if (pfn_valid(pfn)) { | |
193 | bool reserved; | |
194 | struct page *tail = pfn_to_page(pfn); | |
668f9abb | 195 | struct page *head = compound_head(tail); |
73fa0d10 AW |
196 | reserved = !!(PageReserved(head)); |
197 | if (head != tail) { | |
198 | /* | |
199 | * "head" is not a dangling pointer | |
668f9abb | 200 | * (compound_head takes care of that) |
73fa0d10 AW |
201 | * but the hugepage may have been split |
202 | * from under us (and we may not hold a | |
203 | * reference count on the head page so it can | |
204 | * be reused before we run PageReferenced), so | |
205 | * we've to check PageTail before returning | |
206 | * what we just read. | |
207 | */ | |
208 | smp_rmb(); | |
209 | if (PageTail(tail)) | |
210 | return reserved; | |
211 | } | |
212 | return PageReserved(tail); | |
213 | } | |
214 | ||
215 | return true; | |
216 | } | |
217 | ||
218 | static int put_pfn(unsigned long pfn, int prot) | |
219 | { | |
220 | if (!is_invalid_reserved_pfn(pfn)) { | |
221 | struct page *page = pfn_to_page(pfn); | |
222 | if (prot & IOMMU_WRITE) | |
223 | SetPageDirty(page); | |
224 | put_page(page); | |
225 | return 1; | |
226 | } | |
227 | return 0; | |
228 | } | |
229 | ||
73fa0d10 AW |
230 | static int vaddr_get_pfn(unsigned long vaddr, int prot, unsigned long *pfn) |
231 | { | |
232 | struct page *page[1]; | |
233 | struct vm_area_struct *vma; | |
234 | int ret = -EFAULT; | |
235 | ||
236 | if (get_user_pages_fast(vaddr, 1, !!(prot & IOMMU_WRITE), page) == 1) { | |
237 | *pfn = page_to_pfn(page[0]); | |
238 | return 0; | |
239 | } | |
240 | ||
241 | down_read(¤t->mm->mmap_sem); | |
242 | ||
243 | vma = find_vma_intersection(current->mm, vaddr, vaddr + 1); | |
244 | ||
245 | if (vma && vma->vm_flags & VM_PFNMAP) { | |
246 | *pfn = ((vaddr - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff; | |
247 | if (is_invalid_reserved_pfn(*pfn)) | |
248 | ret = 0; | |
249 | } | |
250 | ||
251 | up_read(¤t->mm->mmap_sem); | |
252 | ||
253 | return ret; | |
254 | } | |
255 | ||
166fd7d9 AW |
256 | /* |
257 | * Attempt to pin pages. We really don't want to track all the pfns and | |
258 | * the iommu can only map chunks of consecutive pfns anyway, so get the | |
259 | * first page and all consecutive pages with the same locking. | |
260 | */ | |
261 | static long vfio_pin_pages(unsigned long vaddr, long npage, | |
262 | int prot, unsigned long *pfn_base) | |
73fa0d10 | 263 | { |
166fd7d9 AW |
264 | unsigned long limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT; |
265 | bool lock_cap = capable(CAP_IPC_LOCK); | |
266 | long ret, i; | |
73fa0d10 | 267 | |
166fd7d9 AW |
268 | if (!current->mm) |
269 | return -ENODEV; | |
73fa0d10 | 270 | |
166fd7d9 AW |
271 | ret = vaddr_get_pfn(vaddr, prot, pfn_base); |
272 | if (ret) | |
273 | return ret; | |
73fa0d10 | 274 | |
166fd7d9 AW |
275 | if (is_invalid_reserved_pfn(*pfn_base)) |
276 | return 1; | |
73fa0d10 | 277 | |
166fd7d9 AW |
278 | if (!lock_cap && current->mm->locked_vm + 1 > limit) { |
279 | put_pfn(*pfn_base, prot); | |
280 | pr_warn("%s: RLIMIT_MEMLOCK (%ld) exceeded\n", __func__, | |
281 | limit << PAGE_SHIFT); | |
282 | return -ENOMEM; | |
283 | } | |
284 | ||
5c6c2b21 AW |
285 | if (unlikely(disable_hugepages)) { |
286 | vfio_lock_acct(1); | |
287 | return 1; | |
288 | } | |
289 | ||
166fd7d9 AW |
290 | /* Lock all the consecutive pages from pfn_base */ |
291 | for (i = 1, vaddr += PAGE_SIZE; i < npage; i++, vaddr += PAGE_SIZE) { | |
73fa0d10 AW |
292 | unsigned long pfn = 0; |
293 | ||
294 | ret = vaddr_get_pfn(vaddr, prot, &pfn); | |
166fd7d9 AW |
295 | if (ret) |
296 | break; | |
297 | ||
298 | if (pfn != *pfn_base + i || is_invalid_reserved_pfn(pfn)) { | |
299 | put_pfn(pfn, prot); | |
300 | break; | |
73fa0d10 AW |
301 | } |
302 | ||
166fd7d9 AW |
303 | if (!lock_cap && current->mm->locked_vm + i + 1 > limit) { |
304 | put_pfn(pfn, prot); | |
305 | pr_warn("%s: RLIMIT_MEMLOCK (%ld) exceeded\n", | |
306 | __func__, limit << PAGE_SHIFT); | |
307 | break; | |
308 | } | |
309 | } | |
310 | ||
311 | vfio_lock_acct(i); | |
312 | ||
313 | return i; | |
314 | } | |
315 | ||
316 | static long vfio_unpin_pages(unsigned long pfn, long npage, | |
317 | int prot, bool do_accounting) | |
318 | { | |
319 | unsigned long unlocked = 0; | |
320 | long i; | |
321 | ||
322 | for (i = 0; i < npage; i++) | |
323 | unlocked += put_pfn(pfn++, prot); | |
324 | ||
325 | if (do_accounting) | |
326 | vfio_lock_acct(-unlocked); | |
327 | ||
328 | return unlocked; | |
329 | } | |
330 | ||
1ef3e2bc | 331 | static void vfio_unmap_unpin(struct vfio_iommu *iommu, struct vfio_dma *dma) |
166fd7d9 | 332 | { |
1ef3e2bc AW |
333 | dma_addr_t iova = dma->iova, end = dma->iova + dma->size; |
334 | struct vfio_domain *domain, *d; | |
166fd7d9 AW |
335 | long unlocked = 0; |
336 | ||
1ef3e2bc AW |
337 | if (!dma->size) |
338 | return; | |
339 | /* | |
340 | * We use the IOMMU to track the physical addresses, otherwise we'd | |
341 | * need a much more complicated tracking system. Unfortunately that | |
342 | * means we need to use one of the iommu domains to figure out the | |
343 | * pfns to unpin. The rest need to be unmapped in advance so we have | |
344 | * no iommu translations remaining when the pages are unpinned. | |
345 | */ | |
346 | domain = d = list_first_entry(&iommu->domain_list, | |
347 | struct vfio_domain, next); | |
348 | ||
349 | list_for_each_entry_continue(d, &iommu->domain_list, next) | |
350 | iommu_unmap(d->domain, dma->iova, dma->size); | |
351 | ||
166fd7d9 AW |
352 | while (iova < end) { |
353 | size_t unmapped; | |
354 | phys_addr_t phys; | |
355 | ||
1ef3e2bc | 356 | phys = iommu_iova_to_phys(domain->domain, iova); |
166fd7d9 AW |
357 | if (WARN_ON(!phys)) { |
358 | iova += PAGE_SIZE; | |
359 | continue; | |
73fa0d10 | 360 | } |
166fd7d9 | 361 | |
1ef3e2bc AW |
362 | unmapped = iommu_unmap(domain->domain, iova, PAGE_SIZE); |
363 | if (WARN_ON(!unmapped)) | |
166fd7d9 AW |
364 | break; |
365 | ||
366 | unlocked += vfio_unpin_pages(phys >> PAGE_SHIFT, | |
367 | unmapped >> PAGE_SHIFT, | |
368 | dma->prot, false); | |
369 | iova += unmapped; | |
73fa0d10 | 370 | } |
166fd7d9 AW |
371 | |
372 | vfio_lock_acct(-unlocked); | |
73fa0d10 AW |
373 | } |
374 | ||
1ef3e2bc | 375 | static void vfio_remove_dma(struct vfio_iommu *iommu, struct vfio_dma *dma) |
73fa0d10 | 376 | { |
1ef3e2bc AW |
377 | vfio_unmap_unpin(iommu, dma); |
378 | vfio_unlink_dma(iommu, dma); | |
379 | kfree(dma); | |
380 | } | |
73fa0d10 | 381 | |
1ef3e2bc AW |
382 | static unsigned long vfio_pgsize_bitmap(struct vfio_iommu *iommu) |
383 | { | |
384 | struct vfio_domain *domain; | |
385 | unsigned long bitmap = PAGE_MASK; | |
166fd7d9 | 386 | |
1ef3e2bc AW |
387 | mutex_lock(&iommu->lock); |
388 | list_for_each_entry(domain, &iommu->domain_list, next) | |
389 | bitmap &= domain->domain->ops->pgsize_bitmap; | |
390 | mutex_unlock(&iommu->lock); | |
73fa0d10 | 391 | |
1ef3e2bc | 392 | return bitmap; |
73fa0d10 AW |
393 | } |
394 | ||
395 | static int vfio_dma_do_unmap(struct vfio_iommu *iommu, | |
396 | struct vfio_iommu_type1_dma_unmap *unmap) | |
397 | { | |
73fa0d10 | 398 | uint64_t mask; |
cd9b2268 | 399 | struct vfio_dma *dma; |
1ef3e2bc | 400 | size_t unmapped = 0; |
cd9b2268 | 401 | int ret = 0; |
73fa0d10 | 402 | |
1ef3e2bc | 403 | mask = ((uint64_t)1 << __ffs(vfio_pgsize_bitmap(iommu))) - 1; |
73fa0d10 AW |
404 | |
405 | if (unmap->iova & mask) | |
406 | return -EINVAL; | |
f5bfdbf2 | 407 | if (!unmap->size || unmap->size & mask) |
73fa0d10 AW |
408 | return -EINVAL; |
409 | ||
73fa0d10 AW |
410 | WARN_ON(mask & PAGE_MASK); |
411 | ||
412 | mutex_lock(&iommu->lock); | |
413 | ||
1ef3e2bc AW |
414 | /* |
415 | * vfio-iommu-type1 (v1) - User mappings were coalesced together to | |
416 | * avoid tracking individual mappings. This means that the granularity | |
417 | * of the original mapping was lost and the user was allowed to attempt | |
418 | * to unmap any range. Depending on the contiguousness of physical | |
419 | * memory and page sizes supported by the IOMMU, arbitrary unmaps may | |
420 | * or may not have worked. We only guaranteed unmap granularity | |
421 | * matching the original mapping; even though it was untracked here, | |
422 | * the original mappings are reflected in IOMMU mappings. This | |
423 | * resulted in a couple unusual behaviors. First, if a range is not | |
424 | * able to be unmapped, ex. a set of 4k pages that was mapped as a | |
425 | * 2M hugepage into the IOMMU, the unmap ioctl returns success but with | |
426 | * a zero sized unmap. Also, if an unmap request overlaps the first | |
427 | * address of a hugepage, the IOMMU will unmap the entire hugepage. | |
428 | * This also returns success and the returned unmap size reflects the | |
429 | * actual size unmapped. | |
430 | * | |
431 | * We attempt to maintain compatibility with this "v1" interface, but | |
432 | * we take control out of the hands of the IOMMU. Therefore, an unmap | |
433 | * request offset from the beginning of the original mapping will | |
434 | * return success with zero sized unmap. And an unmap request covering | |
435 | * the first iova of mapping will unmap the entire range. | |
436 | * | |
437 | * The v2 version of this interface intends to be more deterministic. | |
438 | * Unmap requests must fully cover previous mappings. Multiple | |
439 | * mappings may still be unmaped by specifying large ranges, but there | |
440 | * must not be any previous mappings bisected by the range. An error | |
441 | * will be returned if these conditions are not met. The v2 interface | |
442 | * will only return success and a size of zero if there were no | |
443 | * mappings within the range. | |
444 | */ | |
445 | if (iommu->v2) { | |
446 | dma = vfio_find_dma(iommu, unmap->iova, 0); | |
447 | if (dma && dma->iova != unmap->iova) { | |
448 | ret = -EINVAL; | |
449 | goto unlock; | |
450 | } | |
451 | dma = vfio_find_dma(iommu, unmap->iova + unmap->size - 1, 0); | |
452 | if (dma && dma->iova + dma->size != unmap->iova + unmap->size) { | |
453 | ret = -EINVAL; | |
454 | goto unlock; | |
455 | } | |
456 | } | |
457 | ||
166fd7d9 | 458 | while ((dma = vfio_find_dma(iommu, unmap->iova, unmap->size))) { |
1ef3e2bc | 459 | if (!iommu->v2 && unmap->iova > dma->iova) |
166fd7d9 | 460 | break; |
1ef3e2bc AW |
461 | unmapped += dma->size; |
462 | vfio_remove_dma(iommu, dma); | |
166fd7d9 | 463 | } |
cd9b2268 | 464 | |
1ef3e2bc | 465 | unlock: |
73fa0d10 | 466 | mutex_unlock(&iommu->lock); |
166fd7d9 | 467 | |
1ef3e2bc | 468 | /* Report how much was unmapped */ |
166fd7d9 AW |
469 | unmap->size = unmapped; |
470 | ||
471 | return ret; | |
472 | } | |
473 | ||
474 | /* | |
475 | * Turns out AMD IOMMU has a page table bug where it won't map large pages | |
476 | * to a region that previously mapped smaller pages. This should be fixed | |
477 | * soon, so this is just a temporary workaround to break mappings down into | |
478 | * PAGE_SIZE. Better to map smaller pages than nothing. | |
479 | */ | |
1ef3e2bc | 480 | static int map_try_harder(struct vfio_domain *domain, dma_addr_t iova, |
166fd7d9 AW |
481 | unsigned long pfn, long npage, int prot) |
482 | { | |
483 | long i; | |
484 | int ret; | |
485 | ||
486 | for (i = 0; i < npage; i++, pfn++, iova += PAGE_SIZE) { | |
1ef3e2bc | 487 | ret = iommu_map(domain->domain, iova, |
166fd7d9 | 488 | (phys_addr_t)pfn << PAGE_SHIFT, |
1ef3e2bc | 489 | PAGE_SIZE, prot | domain->prot); |
166fd7d9 AW |
490 | if (ret) |
491 | break; | |
492 | } | |
493 | ||
494 | for (; i < npage && i > 0; i--, iova -= PAGE_SIZE) | |
1ef3e2bc AW |
495 | iommu_unmap(domain->domain, iova, PAGE_SIZE); |
496 | ||
497 | return ret; | |
498 | } | |
499 | ||
500 | static int vfio_iommu_map(struct vfio_iommu *iommu, dma_addr_t iova, | |
501 | unsigned long pfn, long npage, int prot) | |
502 | { | |
503 | struct vfio_domain *d; | |
504 | int ret; | |
505 | ||
506 | list_for_each_entry(d, &iommu->domain_list, next) { | |
507 | ret = iommu_map(d->domain, iova, (phys_addr_t)pfn << PAGE_SHIFT, | |
508 | npage << PAGE_SHIFT, prot | d->prot); | |
509 | if (ret) { | |
510 | if (ret != -EBUSY || | |
511 | map_try_harder(d, iova, pfn, npage, prot)) | |
512 | goto unwind; | |
513 | } | |
514 | } | |
515 | ||
516 | return 0; | |
517 | ||
518 | unwind: | |
519 | list_for_each_entry_continue_reverse(d, &iommu->domain_list, next) | |
520 | iommu_unmap(d->domain, iova, npage << PAGE_SHIFT); | |
166fd7d9 | 521 | |
cd9b2268 | 522 | return ret; |
73fa0d10 AW |
523 | } |
524 | ||
525 | static int vfio_dma_do_map(struct vfio_iommu *iommu, | |
526 | struct vfio_iommu_type1_dma_map *map) | |
527 | { | |
c8dbca16 | 528 | dma_addr_t iova = map->iova; |
166fd7d9 | 529 | unsigned long vaddr = map->vaddr; |
73fa0d10 | 530 | size_t size = map->size; |
166fd7d9 | 531 | long npage; |
73fa0d10 AW |
532 | int ret = 0, prot = 0; |
533 | uint64_t mask; | |
1ef3e2bc | 534 | struct vfio_dma *dma; |
d93b3ac0 | 535 | unsigned long pfn; |
166fd7d9 | 536 | |
c8dbca16 AW |
537 | /* Verify that none of our __u64 fields overflow */ |
538 | if (map->size != size || map->vaddr != vaddr || map->iova != iova) | |
539 | return -EINVAL; | |
73fa0d10 | 540 | |
1ef3e2bc | 541 | mask = ((uint64_t)1 << __ffs(vfio_pgsize_bitmap(iommu))) - 1; |
73fa0d10 | 542 | |
c8dbca16 AW |
543 | WARN_ON(mask & PAGE_MASK); |
544 | ||
73fa0d10 AW |
545 | /* READ/WRITE from device perspective */ |
546 | if (map->flags & VFIO_DMA_MAP_FLAG_WRITE) | |
547 | prot |= IOMMU_WRITE; | |
548 | if (map->flags & VFIO_DMA_MAP_FLAG_READ) | |
549 | prot |= IOMMU_READ; | |
550 | ||
c8dbca16 | 551 | if (!prot || !size || (size | iova | vaddr) & mask) |
73fa0d10 AW |
552 | return -EINVAL; |
553 | ||
c8dbca16 AW |
554 | /* Don't allow IOVA or virtual address wrap */ |
555 | if (iova + size - 1 < iova || vaddr + size - 1 < vaddr) | |
73fa0d10 AW |
556 | return -EINVAL; |
557 | ||
558 | mutex_lock(&iommu->lock); | |
559 | ||
c8dbca16 | 560 | if (vfio_find_dma(iommu, iova, size)) { |
166fd7d9 AW |
561 | mutex_unlock(&iommu->lock); |
562 | return -EEXIST; | |
73fa0d10 AW |
563 | } |
564 | ||
1ef3e2bc AW |
565 | dma = kzalloc(sizeof(*dma), GFP_KERNEL); |
566 | if (!dma) { | |
567 | mutex_unlock(&iommu->lock); | |
568 | return -ENOMEM; | |
569 | } | |
570 | ||
c8dbca16 AW |
571 | dma->iova = iova; |
572 | dma->vaddr = vaddr; | |
1ef3e2bc | 573 | dma->prot = prot; |
166fd7d9 | 574 | |
1ef3e2bc AW |
575 | /* Insert zero-sized and grow as we map chunks of it */ |
576 | vfio_link_dma(iommu, dma); | |
166fd7d9 | 577 | |
c8dbca16 | 578 | while (size) { |
166fd7d9 | 579 | /* Pin a contiguous chunk of memory */ |
c8dbca16 AW |
580 | npage = vfio_pin_pages(vaddr + dma->size, |
581 | size >> PAGE_SHIFT, prot, &pfn); | |
166fd7d9 AW |
582 | if (npage <= 0) { |
583 | WARN_ON(!npage); | |
584 | ret = (int)npage; | |
1ef3e2bc | 585 | break; |
166fd7d9 AW |
586 | } |
587 | ||
1ef3e2bc | 588 | /* Map it! */ |
c8dbca16 | 589 | ret = vfio_iommu_map(iommu, iova + dma->size, pfn, npage, prot); |
166fd7d9 | 590 | if (ret) { |
1ef3e2bc AW |
591 | vfio_unpin_pages(pfn, npage, prot, true); |
592 | break; | |
166fd7d9 AW |
593 | } |
594 | ||
c8dbca16 AW |
595 | size -= npage << PAGE_SHIFT; |
596 | dma->size += npage << PAGE_SHIFT; | |
1ef3e2bc | 597 | } |
166fd7d9 | 598 | |
1ef3e2bc AW |
599 | if (ret) |
600 | vfio_remove_dma(iommu, dma); | |
166fd7d9 | 601 | |
1ef3e2bc AW |
602 | mutex_unlock(&iommu->lock); |
603 | return ret; | |
604 | } | |
605 | ||
606 | static int vfio_bus_type(struct device *dev, void *data) | |
607 | { | |
608 | struct bus_type **bus = data; | |
609 | ||
610 | if (*bus && *bus != dev->bus) | |
611 | return -EINVAL; | |
612 | ||
613 | *bus = dev->bus; | |
614 | ||
615 | return 0; | |
616 | } | |
617 | ||
618 | static int vfio_iommu_replay(struct vfio_iommu *iommu, | |
619 | struct vfio_domain *domain) | |
620 | { | |
621 | struct vfio_domain *d; | |
622 | struct rb_node *n; | |
623 | int ret; | |
624 | ||
625 | /* Arbitrarily pick the first domain in the list for lookups */ | |
626 | d = list_first_entry(&iommu->domain_list, struct vfio_domain, next); | |
627 | n = rb_first(&iommu->dma_list); | |
628 | ||
629 | /* If there's not a domain, there better not be any mappings */ | |
630 | if (WARN_ON(n && !d)) | |
631 | return -EINVAL; | |
632 | ||
633 | for (; n; n = rb_next(n)) { | |
634 | struct vfio_dma *dma; | |
635 | dma_addr_t iova; | |
636 | ||
637 | dma = rb_entry(n, struct vfio_dma, node); | |
638 | iova = dma->iova; | |
639 | ||
640 | while (iova < dma->iova + dma->size) { | |
641 | phys_addr_t phys = iommu_iova_to_phys(d->domain, iova); | |
642 | size_t size; | |
73fa0d10 | 643 | |
1ef3e2bc AW |
644 | if (WARN_ON(!phys)) { |
645 | iova += PAGE_SIZE; | |
646 | continue; | |
166fd7d9 AW |
647 | } |
648 | ||
1ef3e2bc | 649 | size = PAGE_SIZE; |
73fa0d10 | 650 | |
1ef3e2bc AW |
651 | while (iova + size < dma->iova + dma->size && |
652 | phys + size == iommu_iova_to_phys(d->domain, | |
653 | iova + size)) | |
654 | size += PAGE_SIZE; | |
d93b3ac0 | 655 | |
1ef3e2bc AW |
656 | ret = iommu_map(domain->domain, iova, phys, |
657 | size, dma->prot | domain->prot); | |
658 | if (ret) | |
659 | return ret; | |
d93b3ac0 | 660 | |
1ef3e2bc AW |
661 | iova += size; |
662 | } | |
166fd7d9 | 663 | } |
73fa0d10 | 664 | |
1ef3e2bc | 665 | return 0; |
73fa0d10 AW |
666 | } |
667 | ||
668 | static int vfio_iommu_type1_attach_group(void *iommu_data, | |
669 | struct iommu_group *iommu_group) | |
670 | { | |
671 | struct vfio_iommu *iommu = iommu_data; | |
1ef3e2bc AW |
672 | struct vfio_group *group, *g; |
673 | struct vfio_domain *domain, *d; | |
674 | struct bus_type *bus = NULL; | |
73fa0d10 AW |
675 | int ret; |
676 | ||
73fa0d10 AW |
677 | mutex_lock(&iommu->lock); |
678 | ||
1ef3e2bc AW |
679 | list_for_each_entry(d, &iommu->domain_list, next) { |
680 | list_for_each_entry(g, &d->group_list, next) { | |
681 | if (g->iommu_group != iommu_group) | |
682 | continue; | |
683 | ||
73fa0d10 | 684 | mutex_unlock(&iommu->lock); |
73fa0d10 AW |
685 | return -EINVAL; |
686 | } | |
687 | } | |
688 | ||
1ef3e2bc AW |
689 | group = kzalloc(sizeof(*group), GFP_KERNEL); |
690 | domain = kzalloc(sizeof(*domain), GFP_KERNEL); | |
691 | if (!group || !domain) { | |
692 | ret = -ENOMEM; | |
693 | goto out_free; | |
694 | } | |
695 | ||
696 | group->iommu_group = iommu_group; | |
697 | ||
698 | /* Determine bus_type in order to allocate a domain */ | |
699 | ret = iommu_group_for_each_dev(iommu_group, &bus, vfio_bus_type); | |
700 | if (ret) | |
701 | goto out_free; | |
702 | ||
703 | domain->domain = iommu_domain_alloc(bus); | |
704 | if (!domain->domain) { | |
705 | ret = -EIO; | |
706 | goto out_free; | |
707 | } | |
708 | ||
f5c9eceb WD |
709 | if (iommu->nesting) { |
710 | int attr = 1; | |
711 | ||
712 | ret = iommu_domain_set_attr(domain->domain, DOMAIN_ATTR_NESTING, | |
713 | &attr); | |
714 | if (ret) | |
715 | goto out_domain; | |
716 | } | |
717 | ||
1ef3e2bc AW |
718 | ret = iommu_attach_group(domain->domain, iommu_group); |
719 | if (ret) | |
720 | goto out_domain; | |
721 | ||
722 | INIT_LIST_HEAD(&domain->group_list); | |
723 | list_add(&group->next, &domain->group_list); | |
724 | ||
725 | if (!allow_unsafe_interrupts && | |
eb165f05 | 726 | !iommu_capable(bus, IOMMU_CAP_INTR_REMAP)) { |
1ef3e2bc AW |
727 | pr_warn("%s: No interrupt remapping support. Use the module param \"allow_unsafe_interrupts\" to enable VFIO IOMMU support on this platform\n", |
728 | __func__); | |
729 | ret = -EPERM; | |
730 | goto out_detach; | |
731 | } | |
732 | ||
eb165f05 | 733 | if (iommu_capable(bus, IOMMU_CAP_CACHE_COHERENCY)) |
1ef3e2bc AW |
734 | domain->prot |= IOMMU_CACHE; |
735 | ||
73fa0d10 | 736 | /* |
1ef3e2bc AW |
737 | * Try to match an existing compatible domain. We don't want to |
738 | * preclude an IOMMU driver supporting multiple bus_types and being | |
739 | * able to include different bus_types in the same IOMMU domain, so | |
740 | * we test whether the domains use the same iommu_ops rather than | |
741 | * testing if they're on the same bus_type. | |
73fa0d10 | 742 | */ |
1ef3e2bc AW |
743 | list_for_each_entry(d, &iommu->domain_list, next) { |
744 | if (d->domain->ops == domain->domain->ops && | |
745 | d->prot == domain->prot) { | |
746 | iommu_detach_group(domain->domain, iommu_group); | |
747 | if (!iommu_attach_group(d->domain, iommu_group)) { | |
748 | list_add(&group->next, &d->group_list); | |
749 | iommu_domain_free(domain->domain); | |
750 | kfree(domain); | |
751 | mutex_unlock(&iommu->lock); | |
752 | return 0; | |
753 | } | |
754 | ||
755 | ret = iommu_attach_group(domain->domain, iommu_group); | |
756 | if (ret) | |
757 | goto out_domain; | |
758 | } | |
73fa0d10 AW |
759 | } |
760 | ||
1ef3e2bc AW |
761 | /* replay mappings on new domains */ |
762 | ret = vfio_iommu_replay(iommu, domain); | |
763 | if (ret) | |
764 | goto out_detach; | |
765 | ||
766 | list_add(&domain->next, &iommu->domain_list); | |
73fa0d10 AW |
767 | |
768 | mutex_unlock(&iommu->lock); | |
769 | ||
770 | return 0; | |
1ef3e2bc AW |
771 | |
772 | out_detach: | |
773 | iommu_detach_group(domain->domain, iommu_group); | |
774 | out_domain: | |
775 | iommu_domain_free(domain->domain); | |
776 | out_free: | |
777 | kfree(domain); | |
778 | kfree(group); | |
779 | mutex_unlock(&iommu->lock); | |
780 | return ret; | |
781 | } | |
782 | ||
783 | static void vfio_iommu_unmap_unpin_all(struct vfio_iommu *iommu) | |
784 | { | |
785 | struct rb_node *node; | |
786 | ||
787 | while ((node = rb_first(&iommu->dma_list))) | |
788 | vfio_remove_dma(iommu, rb_entry(node, struct vfio_dma, node)); | |
73fa0d10 AW |
789 | } |
790 | ||
791 | static void vfio_iommu_type1_detach_group(void *iommu_data, | |
792 | struct iommu_group *iommu_group) | |
793 | { | |
794 | struct vfio_iommu *iommu = iommu_data; | |
1ef3e2bc | 795 | struct vfio_domain *domain; |
73fa0d10 AW |
796 | struct vfio_group *group; |
797 | ||
798 | mutex_lock(&iommu->lock); | |
799 | ||
1ef3e2bc AW |
800 | list_for_each_entry(domain, &iommu->domain_list, next) { |
801 | list_for_each_entry(group, &domain->group_list, next) { | |
802 | if (group->iommu_group != iommu_group) | |
803 | continue; | |
804 | ||
805 | iommu_detach_group(domain->domain, iommu_group); | |
73fa0d10 AW |
806 | list_del(&group->next); |
807 | kfree(group); | |
1ef3e2bc AW |
808 | /* |
809 | * Group ownership provides privilege, if the group | |
810 | * list is empty, the domain goes away. If it's the | |
811 | * last domain, then all the mappings go away too. | |
812 | */ | |
813 | if (list_empty(&domain->group_list)) { | |
814 | if (list_is_singular(&iommu->domain_list)) | |
815 | vfio_iommu_unmap_unpin_all(iommu); | |
816 | iommu_domain_free(domain->domain); | |
817 | list_del(&domain->next); | |
818 | kfree(domain); | |
819 | } | |
820 | goto done; | |
73fa0d10 AW |
821 | } |
822 | } | |
823 | ||
1ef3e2bc | 824 | done: |
73fa0d10 AW |
825 | mutex_unlock(&iommu->lock); |
826 | } | |
827 | ||
828 | static void *vfio_iommu_type1_open(unsigned long arg) | |
829 | { | |
830 | struct vfio_iommu *iommu; | |
831 | ||
73fa0d10 AW |
832 | iommu = kzalloc(sizeof(*iommu), GFP_KERNEL); |
833 | if (!iommu) | |
834 | return ERR_PTR(-ENOMEM); | |
835 | ||
f5c9eceb WD |
836 | switch (arg) { |
837 | case VFIO_TYPE1_IOMMU: | |
838 | break; | |
839 | case VFIO_TYPE1_NESTING_IOMMU: | |
840 | iommu->nesting = true; | |
841 | case VFIO_TYPE1v2_IOMMU: | |
842 | iommu->v2 = true; | |
843 | break; | |
844 | default: | |
845 | kfree(iommu); | |
846 | return ERR_PTR(-EINVAL); | |
847 | } | |
848 | ||
1ef3e2bc | 849 | INIT_LIST_HEAD(&iommu->domain_list); |
cd9b2268 | 850 | iommu->dma_list = RB_ROOT; |
73fa0d10 | 851 | mutex_init(&iommu->lock); |
73fa0d10 AW |
852 | |
853 | return iommu; | |
854 | } | |
855 | ||
856 | static void vfio_iommu_type1_release(void *iommu_data) | |
857 | { | |
858 | struct vfio_iommu *iommu = iommu_data; | |
1ef3e2bc | 859 | struct vfio_domain *domain, *domain_tmp; |
73fa0d10 | 860 | struct vfio_group *group, *group_tmp; |
73fa0d10 | 861 | |
1ef3e2bc | 862 | vfio_iommu_unmap_unpin_all(iommu); |
73fa0d10 | 863 | |
1ef3e2bc AW |
864 | list_for_each_entry_safe(domain, domain_tmp, |
865 | &iommu->domain_list, next) { | |
866 | list_for_each_entry_safe(group, group_tmp, | |
867 | &domain->group_list, next) { | |
868 | iommu_detach_group(domain->domain, group->iommu_group); | |
869 | list_del(&group->next); | |
870 | kfree(group); | |
871 | } | |
872 | iommu_domain_free(domain->domain); | |
873 | list_del(&domain->next); | |
874 | kfree(domain); | |
73fa0d10 AW |
875 | } |
876 | ||
73fa0d10 AW |
877 | kfree(iommu); |
878 | } | |
879 | ||
aa429318 AW |
880 | static int vfio_domains_have_iommu_cache(struct vfio_iommu *iommu) |
881 | { | |
882 | struct vfio_domain *domain; | |
883 | int ret = 1; | |
884 | ||
885 | mutex_lock(&iommu->lock); | |
886 | list_for_each_entry(domain, &iommu->domain_list, next) { | |
887 | if (!(domain->prot & IOMMU_CACHE)) { | |
888 | ret = 0; | |
f5bfdbf2 | 889 | break; |
aa429318 | 890 | } |
73fa0d10 | 891 | } |
aa429318 | 892 | mutex_unlock(&iommu->lock); |
73fa0d10 | 893 | |
aa429318 | 894 | return ret; |
73fa0d10 AW |
895 | } |
896 | ||
897 | static long vfio_iommu_type1_ioctl(void *iommu_data, | |
898 | unsigned int cmd, unsigned long arg) | |
899 | { | |
900 | struct vfio_iommu *iommu = iommu_data; | |
901 | unsigned long minsz; | |
902 | ||
903 | if (cmd == VFIO_CHECK_EXTENSION) { | |
904 | switch (arg) { | |
905 | case VFIO_TYPE1_IOMMU: | |
1ef3e2bc | 906 | case VFIO_TYPE1v2_IOMMU: |
f5c9eceb | 907 | case VFIO_TYPE1_NESTING_IOMMU: |
73fa0d10 | 908 | return 1; |
aa429318 AW |
909 | case VFIO_DMA_CC_IOMMU: |
910 | if (!iommu) | |
911 | return 0; | |
912 | return vfio_domains_have_iommu_cache(iommu); | |
73fa0d10 AW |
913 | default: |
914 | return 0; | |
915 | } | |
916 | } else if (cmd == VFIO_IOMMU_GET_INFO) { | |
917 | struct vfio_iommu_type1_info info; | |
918 | ||
919 | minsz = offsetofend(struct vfio_iommu_type1_info, iova_pgsizes); | |
920 | ||
921 | if (copy_from_user(&info, (void __user *)arg, minsz)) | |
922 | return -EFAULT; | |
923 | ||
924 | if (info.argsz < minsz) | |
925 | return -EINVAL; | |
926 | ||
927 | info.flags = 0; | |
928 | ||
1ef3e2bc | 929 | info.iova_pgsizes = vfio_pgsize_bitmap(iommu); |
73fa0d10 AW |
930 | |
931 | return copy_to_user((void __user *)arg, &info, minsz); | |
932 | ||
933 | } else if (cmd == VFIO_IOMMU_MAP_DMA) { | |
934 | struct vfio_iommu_type1_dma_map map; | |
935 | uint32_t mask = VFIO_DMA_MAP_FLAG_READ | | |
936 | VFIO_DMA_MAP_FLAG_WRITE; | |
937 | ||
938 | minsz = offsetofend(struct vfio_iommu_type1_dma_map, size); | |
939 | ||
940 | if (copy_from_user(&map, (void __user *)arg, minsz)) | |
941 | return -EFAULT; | |
942 | ||
943 | if (map.argsz < minsz || map.flags & ~mask) | |
944 | return -EINVAL; | |
945 | ||
946 | return vfio_dma_do_map(iommu, &map); | |
947 | ||
948 | } else if (cmd == VFIO_IOMMU_UNMAP_DMA) { | |
949 | struct vfio_iommu_type1_dma_unmap unmap; | |
166fd7d9 | 950 | long ret; |
73fa0d10 AW |
951 | |
952 | minsz = offsetofend(struct vfio_iommu_type1_dma_unmap, size); | |
953 | ||
954 | if (copy_from_user(&unmap, (void __user *)arg, minsz)) | |
955 | return -EFAULT; | |
956 | ||
957 | if (unmap.argsz < minsz || unmap.flags) | |
958 | return -EINVAL; | |
959 | ||
166fd7d9 AW |
960 | ret = vfio_dma_do_unmap(iommu, &unmap); |
961 | if (ret) | |
962 | return ret; | |
963 | ||
964 | return copy_to_user((void __user *)arg, &unmap, minsz); | |
73fa0d10 AW |
965 | } |
966 | ||
967 | return -ENOTTY; | |
968 | } | |
969 | ||
970 | static const struct vfio_iommu_driver_ops vfio_iommu_driver_ops_type1 = { | |
971 | .name = "vfio-iommu-type1", | |
972 | .owner = THIS_MODULE, | |
973 | .open = vfio_iommu_type1_open, | |
974 | .release = vfio_iommu_type1_release, | |
975 | .ioctl = vfio_iommu_type1_ioctl, | |
976 | .attach_group = vfio_iommu_type1_attach_group, | |
977 | .detach_group = vfio_iommu_type1_detach_group, | |
978 | }; | |
979 | ||
980 | static int __init vfio_iommu_type1_init(void) | |
981 | { | |
73fa0d10 AW |
982 | return vfio_register_iommu_driver(&vfio_iommu_driver_ops_type1); |
983 | } | |
984 | ||
985 | static void __exit vfio_iommu_type1_cleanup(void) | |
986 | { | |
987 | vfio_unregister_iommu_driver(&vfio_iommu_driver_ops_type1); | |
988 | } | |
989 | ||
990 | module_init(vfio_iommu_type1_init); | |
991 | module_exit(vfio_iommu_type1_cleanup); | |
992 | ||
993 | MODULE_VERSION(DRIVER_VERSION); | |
994 | MODULE_LICENSE("GPL v2"); | |
995 | MODULE_AUTHOR(DRIVER_AUTHOR); | |
996 | MODULE_DESCRIPTION(DRIVER_DESC); |