]>
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> |
3f07c014 | 34 | #include <linux/sched/signal.h> |
6e84f315 | 35 | #include <linux/sched/mm.h> |
73fa0d10 AW |
36 | #include <linux/slab.h> |
37 | #include <linux/uaccess.h> | |
38 | #include <linux/vfio.h> | |
39 | #include <linux/workqueue.h> | |
a54eb550 | 40 | #include <linux/mdev.h> |
c086de81 | 41 | #include <linux/notifier.h> |
5d704992 | 42 | #include <linux/dma-iommu.h> |
9d72f87b | 43 | #include <linux/irqdomain.h> |
73fa0d10 AW |
44 | |
45 | #define DRIVER_VERSION "0.2" | |
46 | #define DRIVER_AUTHOR "Alex Williamson <alex.williamson@redhat.com>" | |
47 | #define DRIVER_DESC "Type1 IOMMU driver for VFIO" | |
48 | ||
49 | static bool allow_unsafe_interrupts; | |
50 | module_param_named(allow_unsafe_interrupts, | |
51 | allow_unsafe_interrupts, bool, S_IRUGO | S_IWUSR); | |
52 | MODULE_PARM_DESC(allow_unsafe_interrupts, | |
53 | "Enable VFIO IOMMU support for on platforms without interrupt remapping support."); | |
54 | ||
5c6c2b21 AW |
55 | static bool disable_hugepages; |
56 | module_param_named(disable_hugepages, | |
57 | disable_hugepages, bool, S_IRUGO | S_IWUSR); | |
58 | MODULE_PARM_DESC(disable_hugepages, | |
59 | "Disable VFIO IOMMU support for IOMMU hugepages."); | |
60 | ||
73fa0d10 | 61 | struct vfio_iommu { |
1ef3e2bc | 62 | struct list_head domain_list; |
a54eb550 | 63 | struct vfio_domain *external_domain; /* domain for external user */ |
73fa0d10 | 64 | struct mutex lock; |
cd9b2268 | 65 | struct rb_root dma_list; |
c086de81 | 66 | struct blocking_notifier_head notifier; |
f5c9eceb WD |
67 | bool v2; |
68 | bool nesting; | |
1ef3e2bc AW |
69 | }; |
70 | ||
71 | struct vfio_domain { | |
72 | struct iommu_domain *domain; | |
73 | struct list_head next; | |
73fa0d10 | 74 | struct list_head group_list; |
1ef3e2bc | 75 | int prot; /* IOMMU_CACHE */ |
6fe1010d | 76 | bool fgsp; /* Fine-grained super pages */ |
73fa0d10 AW |
77 | }; |
78 | ||
79 | struct vfio_dma { | |
cd9b2268 | 80 | struct rb_node node; |
73fa0d10 AW |
81 | dma_addr_t iova; /* Device address */ |
82 | unsigned long vaddr; /* Process virtual addr */ | |
166fd7d9 | 83 | size_t size; /* Map size (bytes) */ |
73fa0d10 | 84 | int prot; /* IOMMU_READ/WRITE */ |
a54eb550 | 85 | bool iommu_mapped; |
8f0d5bb9 | 86 | struct task_struct *task; |
a54eb550 | 87 | struct rb_root pfn_list; /* Ex-user pinned pfn list */ |
73fa0d10 AW |
88 | }; |
89 | ||
90 | struct vfio_group { | |
91 | struct iommu_group *iommu_group; | |
92 | struct list_head next; | |
93 | }; | |
94 | ||
a54eb550 KW |
95 | /* |
96 | * Guest RAM pinning working set or DMA target | |
97 | */ | |
98 | struct vfio_pfn { | |
99 | struct rb_node node; | |
100 | dma_addr_t iova; /* Device address */ | |
101 | unsigned long pfn; /* Host pfn */ | |
102 | atomic_t ref_count; | |
103 | }; | |
104 | ||
105 | #define IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu) \ | |
106 | (!list_empty(&iommu->domain_list)) | |
107 | ||
108 | static int put_pfn(unsigned long pfn, int prot); | |
109 | ||
73fa0d10 AW |
110 | /* |
111 | * This code handles mapping and unmapping of user data buffers | |
112 | * into DMA'ble space using the IOMMU | |
113 | */ | |
114 | ||
cd9b2268 AW |
115 | static struct vfio_dma *vfio_find_dma(struct vfio_iommu *iommu, |
116 | dma_addr_t start, size_t size) | |
117 | { | |
118 | struct rb_node *node = iommu->dma_list.rb_node; | |
119 | ||
120 | while (node) { | |
121 | struct vfio_dma *dma = rb_entry(node, struct vfio_dma, node); | |
122 | ||
123 | if (start + size <= dma->iova) | |
124 | node = node->rb_left; | |
166fd7d9 | 125 | else if (start >= dma->iova + dma->size) |
cd9b2268 AW |
126 | node = node->rb_right; |
127 | else | |
128 | return dma; | |
129 | } | |
130 | ||
131 | return NULL; | |
132 | } | |
133 | ||
1ef3e2bc | 134 | static void vfio_link_dma(struct vfio_iommu *iommu, struct vfio_dma *new) |
cd9b2268 AW |
135 | { |
136 | struct rb_node **link = &iommu->dma_list.rb_node, *parent = NULL; | |
137 | struct vfio_dma *dma; | |
138 | ||
139 | while (*link) { | |
140 | parent = *link; | |
141 | dma = rb_entry(parent, struct vfio_dma, node); | |
142 | ||
166fd7d9 | 143 | if (new->iova + new->size <= dma->iova) |
cd9b2268 AW |
144 | link = &(*link)->rb_left; |
145 | else | |
146 | link = &(*link)->rb_right; | |
147 | } | |
148 | ||
149 | rb_link_node(&new->node, parent, link); | |
150 | rb_insert_color(&new->node, &iommu->dma_list); | |
151 | } | |
152 | ||
1ef3e2bc | 153 | static void vfio_unlink_dma(struct vfio_iommu *iommu, struct vfio_dma *old) |
cd9b2268 AW |
154 | { |
155 | rb_erase(&old->node, &iommu->dma_list); | |
156 | } | |
157 | ||
a54eb550 KW |
158 | /* |
159 | * Helper Functions for host iova-pfn list | |
160 | */ | |
161 | static struct vfio_pfn *vfio_find_vpfn(struct vfio_dma *dma, dma_addr_t iova) | |
162 | { | |
163 | struct vfio_pfn *vpfn; | |
164 | struct rb_node *node = dma->pfn_list.rb_node; | |
165 | ||
166 | while (node) { | |
167 | vpfn = rb_entry(node, struct vfio_pfn, node); | |
168 | ||
169 | if (iova < vpfn->iova) | |
170 | node = node->rb_left; | |
171 | else if (iova > vpfn->iova) | |
172 | node = node->rb_right; | |
173 | else | |
174 | return vpfn; | |
175 | } | |
176 | return NULL; | |
177 | } | |
178 | ||
179 | static void vfio_link_pfn(struct vfio_dma *dma, | |
180 | struct vfio_pfn *new) | |
181 | { | |
182 | struct rb_node **link, *parent = NULL; | |
183 | struct vfio_pfn *vpfn; | |
184 | ||
185 | link = &dma->pfn_list.rb_node; | |
186 | while (*link) { | |
187 | parent = *link; | |
188 | vpfn = rb_entry(parent, struct vfio_pfn, node); | |
189 | ||
190 | if (new->iova < vpfn->iova) | |
191 | link = &(*link)->rb_left; | |
192 | else | |
193 | link = &(*link)->rb_right; | |
194 | } | |
195 | ||
196 | rb_link_node(&new->node, parent, link); | |
197 | rb_insert_color(&new->node, &dma->pfn_list); | |
198 | } | |
199 | ||
200 | static void vfio_unlink_pfn(struct vfio_dma *dma, struct vfio_pfn *old) | |
201 | { | |
202 | rb_erase(&old->node, &dma->pfn_list); | |
203 | } | |
204 | ||
205 | static int vfio_add_to_pfn_list(struct vfio_dma *dma, dma_addr_t iova, | |
206 | unsigned long pfn) | |
207 | { | |
208 | struct vfio_pfn *vpfn; | |
209 | ||
210 | vpfn = kzalloc(sizeof(*vpfn), GFP_KERNEL); | |
211 | if (!vpfn) | |
212 | return -ENOMEM; | |
213 | ||
214 | vpfn->iova = iova; | |
215 | vpfn->pfn = pfn; | |
216 | atomic_set(&vpfn->ref_count, 1); | |
217 | vfio_link_pfn(dma, vpfn); | |
218 | return 0; | |
219 | } | |
220 | ||
221 | static void vfio_remove_from_pfn_list(struct vfio_dma *dma, | |
222 | struct vfio_pfn *vpfn) | |
223 | { | |
224 | vfio_unlink_pfn(dma, vpfn); | |
225 | kfree(vpfn); | |
226 | } | |
227 | ||
228 | static struct vfio_pfn *vfio_iova_get_vfio_pfn(struct vfio_dma *dma, | |
229 | unsigned long iova) | |
230 | { | |
231 | struct vfio_pfn *vpfn = vfio_find_vpfn(dma, iova); | |
232 | ||
233 | if (vpfn) | |
234 | atomic_inc(&vpfn->ref_count); | |
235 | return vpfn; | |
236 | } | |
237 | ||
238 | static int vfio_iova_put_vfio_pfn(struct vfio_dma *dma, struct vfio_pfn *vpfn) | |
239 | { | |
240 | int ret = 0; | |
241 | ||
242 | if (atomic_dec_and_test(&vpfn->ref_count)) { | |
243 | ret = put_pfn(vpfn->pfn, dma->prot); | |
244 | vfio_remove_from_pfn_list(dma, vpfn); | |
245 | } | |
246 | return ret; | |
247 | } | |
248 | ||
0cfef2b7 | 249 | static int vfio_lock_acct(struct task_struct *task, long npage, bool *lock_cap) |
73fa0d10 | 250 | { |
73fa0d10 | 251 | struct mm_struct *mm; |
6c38c055 | 252 | bool is_current; |
0cfef2b7 | 253 | int ret; |
73fa0d10 | 254 | |
3624a248 | 255 | if (!npage) |
0cfef2b7 | 256 | return 0; |
3624a248 | 257 | |
6c38c055 AW |
258 | is_current = (task->mm == current->mm); |
259 | ||
260 | mm = is_current ? task->mm : get_task_mm(task); | |
3624a248 | 261 | if (!mm) |
0cfef2b7 | 262 | return -ESRCH; /* process exited */ |
73fa0d10 | 263 | |
0cfef2b7 AW |
264 | ret = down_write_killable(&mm->mmap_sem); |
265 | if (!ret) { | |
266 | if (npage > 0) { | |
267 | if (lock_cap ? !*lock_cap : | |
268 | !has_capability(task, CAP_IPC_LOCK)) { | |
269 | unsigned long limit; | |
270 | ||
271 | limit = task_rlimit(task, | |
272 | RLIMIT_MEMLOCK) >> PAGE_SHIFT; | |
273 | ||
274 | if (mm->locked_vm + npage > limit) | |
275 | ret = -ENOMEM; | |
276 | } | |
277 | } | |
278 | ||
279 | if (!ret) | |
280 | mm->locked_vm += npage; | |
73fa0d10 | 281 | |
0cfef2b7 | 282 | up_write(&mm->mmap_sem); |
6c38c055 AW |
283 | } |
284 | ||
0cfef2b7 | 285 | if (!is_current) |
3624a248 | 286 | mmput(mm); |
0cfef2b7 AW |
287 | |
288 | return ret; | |
73fa0d10 AW |
289 | } |
290 | ||
291 | /* | |
292 | * Some mappings aren't backed by a struct page, for example an mmap'd | |
293 | * MMIO range for our own or another device. These use a different | |
294 | * pfn conversion and shouldn't be tracked as locked pages. | |
295 | */ | |
296 | static bool is_invalid_reserved_pfn(unsigned long pfn) | |
297 | { | |
298 | if (pfn_valid(pfn)) { | |
299 | bool reserved; | |
300 | struct page *tail = pfn_to_page(pfn); | |
668f9abb | 301 | struct page *head = compound_head(tail); |
73fa0d10 AW |
302 | reserved = !!(PageReserved(head)); |
303 | if (head != tail) { | |
304 | /* | |
305 | * "head" is not a dangling pointer | |
668f9abb | 306 | * (compound_head takes care of that) |
73fa0d10 AW |
307 | * but the hugepage may have been split |
308 | * from under us (and we may not hold a | |
309 | * reference count on the head page so it can | |
310 | * be reused before we run PageReferenced), so | |
311 | * we've to check PageTail before returning | |
312 | * what we just read. | |
313 | */ | |
314 | smp_rmb(); | |
315 | if (PageTail(tail)) | |
316 | return reserved; | |
317 | } | |
318 | return PageReserved(tail); | |
319 | } | |
320 | ||
321 | return true; | |
322 | } | |
323 | ||
324 | static int put_pfn(unsigned long pfn, int prot) | |
325 | { | |
326 | if (!is_invalid_reserved_pfn(pfn)) { | |
327 | struct page *page = pfn_to_page(pfn); | |
328 | if (prot & IOMMU_WRITE) | |
329 | SetPageDirty(page); | |
330 | put_page(page); | |
331 | return 1; | |
332 | } | |
333 | return 0; | |
334 | } | |
335 | ||
ea85cf35 KW |
336 | static int vaddr_get_pfn(struct mm_struct *mm, unsigned long vaddr, |
337 | int prot, unsigned long *pfn) | |
73fa0d10 AW |
338 | { |
339 | struct page *page[1]; | |
340 | struct vm_area_struct *vma; | |
acf4a054 | 341 | struct vm_area_struct *vmas[1]; |
ea85cf35 | 342 | int ret; |
73fa0d10 | 343 | |
ea85cf35 | 344 | if (mm == current->mm) { |
acf4a054 DW |
345 | ret = get_user_pages_longterm(vaddr, 1, !!(prot & IOMMU_WRITE), |
346 | page, vmas); | |
ea85cf35 KW |
347 | } else { |
348 | unsigned int flags = 0; | |
349 | ||
350 | if (prot & IOMMU_WRITE) | |
351 | flags |= FOLL_WRITE; | |
352 | ||
353 | down_read(&mm->mmap_sem); | |
354 | ret = get_user_pages_remote(NULL, mm, vaddr, 1, flags, page, | |
acf4a054 DW |
355 | vmas, NULL); |
356 | /* | |
357 | * The lifetime of a vaddr_get_pfn() page pin is | |
358 | * userspace-controlled. In the fs-dax case this could | |
359 | * lead to indefinite stalls in filesystem operations. | |
360 | * Disallow attempts to pin fs-dax pages via this | |
361 | * interface. | |
362 | */ | |
363 | if (ret > 0 && vma_is_fsdax(vmas[0])) { | |
364 | ret = -EOPNOTSUPP; | |
365 | put_page(page[0]); | |
366 | } | |
ea85cf35 KW |
367 | up_read(&mm->mmap_sem); |
368 | } | |
369 | ||
370 | if (ret == 1) { | |
73fa0d10 AW |
371 | *pfn = page_to_pfn(page[0]); |
372 | return 0; | |
373 | } | |
374 | ||
ea85cf35 | 375 | down_read(&mm->mmap_sem); |
73fa0d10 | 376 | |
ea85cf35 | 377 | vma = find_vma_intersection(mm, vaddr, vaddr + 1); |
73fa0d10 AW |
378 | |
379 | if (vma && vma->vm_flags & VM_PFNMAP) { | |
380 | *pfn = ((vaddr - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff; | |
381 | if (is_invalid_reserved_pfn(*pfn)) | |
382 | ret = 0; | |
383 | } | |
384 | ||
ea85cf35 | 385 | up_read(&mm->mmap_sem); |
73fa0d10 AW |
386 | return ret; |
387 | } | |
388 | ||
166fd7d9 AW |
389 | /* |
390 | * Attempt to pin pages. We really don't want to track all the pfns and | |
391 | * the iommu can only map chunks of consecutive pfns anyway, so get the | |
392 | * first page and all consecutive pages with the same locking. | |
393 | */ | |
8f0d5bb9 | 394 | static long vfio_pin_pages_remote(struct vfio_dma *dma, unsigned long vaddr, |
7cb671e7 AW |
395 | long npage, unsigned long *pfn_base, |
396 | bool lock_cap, unsigned long limit) | |
73fa0d10 | 397 | { |
7cb671e7 | 398 | unsigned long pfn = 0; |
6c38c055 | 399 | long ret, pinned = 0, lock_acct = 0; |
babbf176 | 400 | bool rsvd; |
a54eb550 | 401 | dma_addr_t iova = vaddr - dma->vaddr + dma->iova; |
73fa0d10 | 402 | |
6c38c055 AW |
403 | /* This code path is only user initiated */ |
404 | if (!current->mm) | |
166fd7d9 | 405 | return -ENODEV; |
73fa0d10 | 406 | |
6c38c055 | 407 | ret = vaddr_get_pfn(current->mm, vaddr, dma->prot, pfn_base); |
166fd7d9 | 408 | if (ret) |
6c38c055 | 409 | return ret; |
73fa0d10 | 410 | |
6c38c055 | 411 | pinned++; |
babbf176 | 412 | rsvd = is_invalid_reserved_pfn(*pfn_base); |
73fa0d10 | 413 | |
a54eb550 KW |
414 | /* |
415 | * Reserved pages aren't counted against the user, externally pinned | |
416 | * pages are already counted against the user. | |
417 | */ | |
418 | if (!rsvd && !vfio_find_vpfn(dma, iova)) { | |
6c38c055 | 419 | if (!lock_cap && current->mm->locked_vm + 1 > limit) { |
a54eb550 KW |
420 | put_pfn(*pfn_base, dma->prot); |
421 | pr_warn("%s: RLIMIT_MEMLOCK (%ld) exceeded\n", __func__, | |
422 | limit << PAGE_SHIFT); | |
6c38c055 | 423 | return -ENOMEM; |
a54eb550 KW |
424 | } |
425 | lock_acct++; | |
5c6c2b21 AW |
426 | } |
427 | ||
6c38c055 AW |
428 | if (unlikely(disable_hugepages)) |
429 | goto out; | |
73fa0d10 | 430 | |
6c38c055 AW |
431 | /* Lock all the consecutive pages from pfn_base */ |
432 | for (vaddr += PAGE_SIZE, iova += PAGE_SIZE; pinned < npage; | |
433 | pinned++, vaddr += PAGE_SIZE, iova += PAGE_SIZE) { | |
6c38c055 AW |
434 | ret = vaddr_get_pfn(current->mm, vaddr, dma->prot, &pfn); |
435 | if (ret) | |
436 | break; | |
437 | ||
438 | if (pfn != *pfn_base + pinned || | |
439 | rsvd != is_invalid_reserved_pfn(pfn)) { | |
440 | put_pfn(pfn, dma->prot); | |
441 | break; | |
442 | } | |
166fd7d9 | 443 | |
6c38c055 AW |
444 | if (!rsvd && !vfio_find_vpfn(dma, iova)) { |
445 | if (!lock_cap && | |
446 | current->mm->locked_vm + lock_acct + 1 > limit) { | |
a54eb550 | 447 | put_pfn(pfn, dma->prot); |
6c38c055 AW |
448 | pr_warn("%s: RLIMIT_MEMLOCK (%ld) exceeded\n", |
449 | __func__, limit << PAGE_SHIFT); | |
0cfef2b7 AW |
450 | ret = -ENOMEM; |
451 | goto unpin_out; | |
a54eb550 | 452 | } |
6c38c055 | 453 | lock_acct++; |
166fd7d9 AW |
454 | } |
455 | } | |
456 | ||
6c38c055 | 457 | out: |
0cfef2b7 AW |
458 | ret = vfio_lock_acct(current, lock_acct, &lock_cap); |
459 | ||
460 | unpin_out: | |
461 | if (ret) { | |
462 | if (!rsvd) { | |
463 | for (pfn = *pfn_base ; pinned ; pfn++, pinned--) | |
464 | put_pfn(pfn, dma->prot); | |
465 | } | |
466 | ||
467 | return ret; | |
468 | } | |
166fd7d9 | 469 | |
6c38c055 | 470 | return pinned; |
166fd7d9 AW |
471 | } |
472 | ||
a54eb550 KW |
473 | static long vfio_unpin_pages_remote(struct vfio_dma *dma, dma_addr_t iova, |
474 | unsigned long pfn, long npage, | |
475 | bool do_accounting) | |
166fd7d9 | 476 | { |
a54eb550 | 477 | long unlocked = 0, locked = 0; |
166fd7d9 AW |
478 | long i; |
479 | ||
6c38c055 | 480 | for (i = 0; i < npage; i++, iova += PAGE_SIZE) { |
a54eb550 KW |
481 | if (put_pfn(pfn++, dma->prot)) { |
482 | unlocked++; | |
6c38c055 | 483 | if (vfio_find_vpfn(dma, iova)) |
a54eb550 KW |
484 | locked++; |
485 | } | |
486 | } | |
487 | ||
488 | if (do_accounting) | |
0cfef2b7 | 489 | vfio_lock_acct(dma->task, locked - unlocked, NULL); |
a54eb550 KW |
490 | |
491 | return unlocked; | |
492 | } | |
493 | ||
494 | static int vfio_pin_page_external(struct vfio_dma *dma, unsigned long vaddr, | |
495 | unsigned long *pfn_base, bool do_accounting) | |
496 | { | |
a54eb550 KW |
497 | struct mm_struct *mm; |
498 | int ret; | |
a54eb550 KW |
499 | |
500 | mm = get_task_mm(dma->task); | |
501 | if (!mm) | |
502 | return -ENODEV; | |
503 | ||
504 | ret = vaddr_get_pfn(mm, vaddr, dma->prot, pfn_base); | |
80dbe1fb AW |
505 | if (!ret && do_accounting && !is_invalid_reserved_pfn(*pfn_base)) { |
506 | ret = vfio_lock_acct(dma->task, 1, NULL); | |
0cfef2b7 AW |
507 | if (ret) { |
508 | put_pfn(*pfn_base, dma->prot); | |
80dbe1fb AW |
509 | if (ret == -ENOMEM) |
510 | pr_warn("%s: Task %s (%d) RLIMIT_MEMLOCK " | |
511 | "(%ld) exceeded\n", __func__, | |
512 | dma->task->comm, task_pid_nr(dma->task), | |
513 | task_rlimit(dma->task, RLIMIT_MEMLOCK)); | |
0cfef2b7 AW |
514 | } |
515 | } | |
516 | ||
a54eb550 KW |
517 | mmput(mm); |
518 | return ret; | |
519 | } | |
520 | ||
521 | static int vfio_unpin_page_external(struct vfio_dma *dma, dma_addr_t iova, | |
522 | bool do_accounting) | |
523 | { | |
524 | int unlocked; | |
525 | struct vfio_pfn *vpfn = vfio_find_vpfn(dma, iova); | |
526 | ||
527 | if (!vpfn) | |
528 | return 0; | |
529 | ||
530 | unlocked = vfio_iova_put_vfio_pfn(dma, vpfn); | |
166fd7d9 AW |
531 | |
532 | if (do_accounting) | |
0cfef2b7 | 533 | vfio_lock_acct(dma->task, -unlocked, NULL); |
166fd7d9 AW |
534 | |
535 | return unlocked; | |
536 | } | |
537 | ||
a54eb550 KW |
538 | static int vfio_iommu_type1_pin_pages(void *iommu_data, |
539 | unsigned long *user_pfn, | |
540 | int npage, int prot, | |
541 | unsigned long *phys_pfn) | |
542 | { | |
543 | struct vfio_iommu *iommu = iommu_data; | |
544 | int i, j, ret; | |
545 | unsigned long remote_vaddr; | |
546 | struct vfio_dma *dma; | |
547 | bool do_accounting; | |
548 | ||
549 | if (!iommu || !user_pfn || !phys_pfn) | |
550 | return -EINVAL; | |
551 | ||
552 | /* Supported for v2 version only */ | |
553 | if (!iommu->v2) | |
554 | return -EACCES; | |
555 | ||
556 | mutex_lock(&iommu->lock); | |
557 | ||
c086de81 KW |
558 | /* Fail if notifier list is empty */ |
559 | if ((!iommu->external_domain) || (!iommu->notifier.head)) { | |
a54eb550 KW |
560 | ret = -EINVAL; |
561 | goto pin_done; | |
562 | } | |
563 | ||
564 | /* | |
565 | * If iommu capable domain exist in the container then all pages are | |
566 | * already pinned and accounted. Accouting should be done if there is no | |
567 | * iommu capable domain in the container. | |
568 | */ | |
569 | do_accounting = !IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu); | |
570 | ||
571 | for (i = 0; i < npage; i++) { | |
572 | dma_addr_t iova; | |
573 | struct vfio_pfn *vpfn; | |
574 | ||
575 | iova = user_pfn[i] << PAGE_SHIFT; | |
2b8bb1d7 | 576 | dma = vfio_find_dma(iommu, iova, PAGE_SIZE); |
a54eb550 KW |
577 | if (!dma) { |
578 | ret = -EINVAL; | |
579 | goto pin_unwind; | |
580 | } | |
581 | ||
582 | if ((dma->prot & prot) != prot) { | |
583 | ret = -EPERM; | |
584 | goto pin_unwind; | |
585 | } | |
586 | ||
587 | vpfn = vfio_iova_get_vfio_pfn(dma, iova); | |
588 | if (vpfn) { | |
589 | phys_pfn[i] = vpfn->pfn; | |
590 | continue; | |
591 | } | |
592 | ||
593 | remote_vaddr = dma->vaddr + iova - dma->iova; | |
594 | ret = vfio_pin_page_external(dma, remote_vaddr, &phys_pfn[i], | |
595 | do_accounting); | |
80dbe1fb | 596 | if (ret) |
a54eb550 | 597 | goto pin_unwind; |
a54eb550 KW |
598 | |
599 | ret = vfio_add_to_pfn_list(dma, iova, phys_pfn[i]); | |
600 | if (ret) { | |
601 | vfio_unpin_page_external(dma, iova, do_accounting); | |
602 | goto pin_unwind; | |
603 | } | |
604 | } | |
605 | ||
606 | ret = i; | |
607 | goto pin_done; | |
608 | ||
609 | pin_unwind: | |
610 | phys_pfn[i] = 0; | |
611 | for (j = 0; j < i; j++) { | |
612 | dma_addr_t iova; | |
613 | ||
614 | iova = user_pfn[j] << PAGE_SHIFT; | |
2b8bb1d7 | 615 | dma = vfio_find_dma(iommu, iova, PAGE_SIZE); |
a54eb550 KW |
616 | vfio_unpin_page_external(dma, iova, do_accounting); |
617 | phys_pfn[j] = 0; | |
618 | } | |
619 | pin_done: | |
620 | mutex_unlock(&iommu->lock); | |
621 | return ret; | |
622 | } | |
623 | ||
624 | static int vfio_iommu_type1_unpin_pages(void *iommu_data, | |
625 | unsigned long *user_pfn, | |
626 | int npage) | |
627 | { | |
628 | struct vfio_iommu *iommu = iommu_data; | |
629 | bool do_accounting; | |
630 | int i; | |
631 | ||
632 | if (!iommu || !user_pfn) | |
633 | return -EINVAL; | |
634 | ||
635 | /* Supported for v2 version only */ | |
636 | if (!iommu->v2) | |
637 | return -EACCES; | |
638 | ||
639 | mutex_lock(&iommu->lock); | |
640 | ||
641 | if (!iommu->external_domain) { | |
642 | mutex_unlock(&iommu->lock); | |
643 | return -EINVAL; | |
644 | } | |
645 | ||
646 | do_accounting = !IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu); | |
647 | for (i = 0; i < npage; i++) { | |
648 | struct vfio_dma *dma; | |
649 | dma_addr_t iova; | |
650 | ||
651 | iova = user_pfn[i] << PAGE_SHIFT; | |
2b8bb1d7 | 652 | dma = vfio_find_dma(iommu, iova, PAGE_SIZE); |
a54eb550 KW |
653 | if (!dma) |
654 | goto unpin_exit; | |
655 | vfio_unpin_page_external(dma, iova, do_accounting); | |
656 | } | |
657 | ||
658 | unpin_exit: | |
659 | mutex_unlock(&iommu->lock); | |
660 | return i > npage ? npage : (i > 0 ? i : -EINVAL); | |
661 | } | |
662 | ||
663 | static long vfio_unmap_unpin(struct vfio_iommu *iommu, struct vfio_dma *dma, | |
664 | bool do_accounting) | |
166fd7d9 | 665 | { |
1ef3e2bc AW |
666 | dma_addr_t iova = dma->iova, end = dma->iova + dma->size; |
667 | struct vfio_domain *domain, *d; | |
166fd7d9 AW |
668 | long unlocked = 0; |
669 | ||
1ef3e2bc | 670 | if (!dma->size) |
a54eb550 KW |
671 | return 0; |
672 | ||
673 | if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu)) | |
674 | return 0; | |
675 | ||
1ef3e2bc AW |
676 | /* |
677 | * We use the IOMMU to track the physical addresses, otherwise we'd | |
678 | * need a much more complicated tracking system. Unfortunately that | |
679 | * means we need to use one of the iommu domains to figure out the | |
680 | * pfns to unpin. The rest need to be unmapped in advance so we have | |
681 | * no iommu translations remaining when the pages are unpinned. | |
682 | */ | |
683 | domain = d = list_first_entry(&iommu->domain_list, | |
684 | struct vfio_domain, next); | |
685 | ||
c5e66887 | 686 | list_for_each_entry_continue(d, &iommu->domain_list, next) { |
1ef3e2bc | 687 | iommu_unmap(d->domain, dma->iova, dma->size); |
c5e66887 AW |
688 | cond_resched(); |
689 | } | |
1ef3e2bc | 690 | |
166fd7d9 | 691 | while (iova < end) { |
6fe1010d AW |
692 | size_t unmapped, len; |
693 | phys_addr_t phys, next; | |
166fd7d9 | 694 | |
1ef3e2bc | 695 | phys = iommu_iova_to_phys(domain->domain, iova); |
166fd7d9 AW |
696 | if (WARN_ON(!phys)) { |
697 | iova += PAGE_SIZE; | |
698 | continue; | |
73fa0d10 | 699 | } |
166fd7d9 | 700 | |
6fe1010d AW |
701 | /* |
702 | * To optimize for fewer iommu_unmap() calls, each of which | |
703 | * may require hardware cache flushing, try to find the | |
704 | * largest contiguous physical memory chunk to unmap. | |
705 | */ | |
706 | for (len = PAGE_SIZE; | |
707 | !domain->fgsp && iova + len < end; len += PAGE_SIZE) { | |
708 | next = iommu_iova_to_phys(domain->domain, iova + len); | |
709 | if (next != phys + len) | |
710 | break; | |
711 | } | |
712 | ||
713 | unmapped = iommu_unmap(domain->domain, iova, len); | |
1ef3e2bc | 714 | if (WARN_ON(!unmapped)) |
166fd7d9 AW |
715 | break; |
716 | ||
a54eb550 KW |
717 | unlocked += vfio_unpin_pages_remote(dma, iova, |
718 | phys >> PAGE_SHIFT, | |
2169037d | 719 | unmapped >> PAGE_SHIFT, |
a54eb550 | 720 | false); |
166fd7d9 | 721 | iova += unmapped; |
c5e66887 AW |
722 | |
723 | cond_resched(); | |
73fa0d10 | 724 | } |
166fd7d9 | 725 | |
a54eb550 KW |
726 | dma->iommu_mapped = false; |
727 | if (do_accounting) { | |
0cfef2b7 | 728 | vfio_lock_acct(dma->task, -unlocked, NULL); |
a54eb550 KW |
729 | return 0; |
730 | } | |
731 | return unlocked; | |
73fa0d10 AW |
732 | } |
733 | ||
1ef3e2bc | 734 | static void vfio_remove_dma(struct vfio_iommu *iommu, struct vfio_dma *dma) |
73fa0d10 | 735 | { |
a54eb550 | 736 | vfio_unmap_unpin(iommu, dma, true); |
1ef3e2bc | 737 | vfio_unlink_dma(iommu, dma); |
8f0d5bb9 | 738 | put_task_struct(dma->task); |
1ef3e2bc AW |
739 | kfree(dma); |
740 | } | |
73fa0d10 | 741 | |
1ef3e2bc AW |
742 | static unsigned long vfio_pgsize_bitmap(struct vfio_iommu *iommu) |
743 | { | |
744 | struct vfio_domain *domain; | |
4644321f | 745 | unsigned long bitmap = ULONG_MAX; |
166fd7d9 | 746 | |
1ef3e2bc AW |
747 | mutex_lock(&iommu->lock); |
748 | list_for_each_entry(domain, &iommu->domain_list, next) | |
d16e0faa | 749 | bitmap &= domain->domain->pgsize_bitmap; |
1ef3e2bc | 750 | mutex_unlock(&iommu->lock); |
73fa0d10 | 751 | |
4644321f EA |
752 | /* |
753 | * In case the IOMMU supports page sizes smaller than PAGE_SIZE | |
754 | * we pretend PAGE_SIZE is supported and hide sub-PAGE_SIZE sizes. | |
755 | * That way the user will be able to map/unmap buffers whose size/ | |
756 | * start address is aligned with PAGE_SIZE. Pinning code uses that | |
757 | * granularity while iommu driver can use the sub-PAGE_SIZE size | |
758 | * to map the buffer. | |
759 | */ | |
760 | if (bitmap & ~PAGE_MASK) { | |
761 | bitmap &= PAGE_MASK; | |
762 | bitmap |= PAGE_SIZE; | |
763 | } | |
764 | ||
1ef3e2bc | 765 | return bitmap; |
73fa0d10 AW |
766 | } |
767 | ||
768 | static int vfio_dma_do_unmap(struct vfio_iommu *iommu, | |
769 | struct vfio_iommu_type1_dma_unmap *unmap) | |
770 | { | |
73fa0d10 | 771 | uint64_t mask; |
c086de81 | 772 | struct vfio_dma *dma, *dma_last = NULL; |
1ef3e2bc | 773 | size_t unmapped = 0; |
c086de81 | 774 | int ret = 0, retries = 0; |
73fa0d10 | 775 | |
1ef3e2bc | 776 | mask = ((uint64_t)1 << __ffs(vfio_pgsize_bitmap(iommu))) - 1; |
73fa0d10 AW |
777 | |
778 | if (unmap->iova & mask) | |
779 | return -EINVAL; | |
f5bfdbf2 | 780 | if (!unmap->size || unmap->size & mask) |
73fa0d10 | 781 | return -EINVAL; |
71a7d3d7 DC |
782 | if (unmap->iova + unmap->size < unmap->iova || |
783 | unmap->size > SIZE_MAX) | |
784 | return -EINVAL; | |
73fa0d10 | 785 | |
73fa0d10 | 786 | WARN_ON(mask & PAGE_MASK); |
c086de81 | 787 | again: |
73fa0d10 AW |
788 | mutex_lock(&iommu->lock); |
789 | ||
1ef3e2bc AW |
790 | /* |
791 | * vfio-iommu-type1 (v1) - User mappings were coalesced together to | |
792 | * avoid tracking individual mappings. This means that the granularity | |
793 | * of the original mapping was lost and the user was allowed to attempt | |
794 | * to unmap any range. Depending on the contiguousness of physical | |
795 | * memory and page sizes supported by the IOMMU, arbitrary unmaps may | |
796 | * or may not have worked. We only guaranteed unmap granularity | |
797 | * matching the original mapping; even though it was untracked here, | |
798 | * the original mappings are reflected in IOMMU mappings. This | |
799 | * resulted in a couple unusual behaviors. First, if a range is not | |
800 | * able to be unmapped, ex. a set of 4k pages that was mapped as a | |
801 | * 2M hugepage into the IOMMU, the unmap ioctl returns success but with | |
802 | * a zero sized unmap. Also, if an unmap request overlaps the first | |
803 | * address of a hugepage, the IOMMU will unmap the entire hugepage. | |
804 | * This also returns success and the returned unmap size reflects the | |
805 | * actual size unmapped. | |
806 | * | |
807 | * We attempt to maintain compatibility with this "v1" interface, but | |
808 | * we take control out of the hands of the IOMMU. Therefore, an unmap | |
809 | * request offset from the beginning of the original mapping will | |
810 | * return success with zero sized unmap. And an unmap request covering | |
811 | * the first iova of mapping will unmap the entire range. | |
812 | * | |
813 | * The v2 version of this interface intends to be more deterministic. | |
814 | * Unmap requests must fully cover previous mappings. Multiple | |
815 | * mappings may still be unmaped by specifying large ranges, but there | |
816 | * must not be any previous mappings bisected by the range. An error | |
817 | * will be returned if these conditions are not met. The v2 interface | |
818 | * will only return success and a size of zero if there were no | |
819 | * mappings within the range. | |
820 | */ | |
821 | if (iommu->v2) { | |
7c03f428 | 822 | dma = vfio_find_dma(iommu, unmap->iova, 1); |
1ef3e2bc AW |
823 | if (dma && dma->iova != unmap->iova) { |
824 | ret = -EINVAL; | |
825 | goto unlock; | |
826 | } | |
827 | dma = vfio_find_dma(iommu, unmap->iova + unmap->size - 1, 0); | |
828 | if (dma && dma->iova + dma->size != unmap->iova + unmap->size) { | |
829 | ret = -EINVAL; | |
830 | goto unlock; | |
831 | } | |
832 | } | |
833 | ||
166fd7d9 | 834 | while ((dma = vfio_find_dma(iommu, unmap->iova, unmap->size))) { |
1ef3e2bc | 835 | if (!iommu->v2 && unmap->iova > dma->iova) |
166fd7d9 | 836 | break; |
8f0d5bb9 KW |
837 | /* |
838 | * Task with same address space who mapped this iova range is | |
839 | * allowed to unmap the iova range. | |
840 | */ | |
841 | if (dma->task->mm != current->mm) | |
842 | break; | |
c086de81 KW |
843 | |
844 | if (!RB_EMPTY_ROOT(&dma->pfn_list)) { | |
845 | struct vfio_iommu_type1_dma_unmap nb_unmap; | |
846 | ||
847 | if (dma_last == dma) { | |
848 | BUG_ON(++retries > 10); | |
849 | } else { | |
850 | dma_last = dma; | |
851 | retries = 0; | |
852 | } | |
853 | ||
854 | nb_unmap.iova = dma->iova; | |
855 | nb_unmap.size = dma->size; | |
856 | ||
857 | /* | |
858 | * Notify anyone (mdev vendor drivers) to invalidate and | |
859 | * unmap iovas within the range we're about to unmap. | |
860 | * Vendor drivers MUST unpin pages in response to an | |
861 | * invalidation. | |
862 | */ | |
863 | mutex_unlock(&iommu->lock); | |
864 | blocking_notifier_call_chain(&iommu->notifier, | |
865 | VFIO_IOMMU_NOTIFY_DMA_UNMAP, | |
866 | &nb_unmap); | |
867 | goto again; | |
868 | } | |
1ef3e2bc AW |
869 | unmapped += dma->size; |
870 | vfio_remove_dma(iommu, dma); | |
166fd7d9 | 871 | } |
cd9b2268 | 872 | |
1ef3e2bc | 873 | unlock: |
73fa0d10 | 874 | mutex_unlock(&iommu->lock); |
166fd7d9 | 875 | |
1ef3e2bc | 876 | /* Report how much was unmapped */ |
166fd7d9 AW |
877 | unmap->size = unmapped; |
878 | ||
879 | return ret; | |
880 | } | |
881 | ||
882 | /* | |
883 | * Turns out AMD IOMMU has a page table bug where it won't map large pages | |
884 | * to a region that previously mapped smaller pages. This should be fixed | |
885 | * soon, so this is just a temporary workaround to break mappings down into | |
886 | * PAGE_SIZE. Better to map smaller pages than nothing. | |
887 | */ | |
1ef3e2bc | 888 | static int map_try_harder(struct vfio_domain *domain, dma_addr_t iova, |
166fd7d9 AW |
889 | unsigned long pfn, long npage, int prot) |
890 | { | |
891 | long i; | |
089f1c6b | 892 | int ret = 0; |
166fd7d9 AW |
893 | |
894 | for (i = 0; i < npage; i++, pfn++, iova += PAGE_SIZE) { | |
1ef3e2bc | 895 | ret = iommu_map(domain->domain, iova, |
166fd7d9 | 896 | (phys_addr_t)pfn << PAGE_SHIFT, |
1ef3e2bc | 897 | PAGE_SIZE, prot | domain->prot); |
166fd7d9 AW |
898 | if (ret) |
899 | break; | |
900 | } | |
901 | ||
902 | for (; i < npage && i > 0; i--, iova -= PAGE_SIZE) | |
1ef3e2bc AW |
903 | iommu_unmap(domain->domain, iova, PAGE_SIZE); |
904 | ||
905 | return ret; | |
906 | } | |
907 | ||
908 | static int vfio_iommu_map(struct vfio_iommu *iommu, dma_addr_t iova, | |
909 | unsigned long pfn, long npage, int prot) | |
910 | { | |
911 | struct vfio_domain *d; | |
912 | int ret; | |
913 | ||
914 | list_for_each_entry(d, &iommu->domain_list, next) { | |
915 | ret = iommu_map(d->domain, iova, (phys_addr_t)pfn << PAGE_SHIFT, | |
916 | npage << PAGE_SHIFT, prot | d->prot); | |
917 | if (ret) { | |
918 | if (ret != -EBUSY || | |
919 | map_try_harder(d, iova, pfn, npage, prot)) | |
920 | goto unwind; | |
921 | } | |
c5e66887 AW |
922 | |
923 | cond_resched(); | |
1ef3e2bc AW |
924 | } |
925 | ||
926 | return 0; | |
927 | ||
928 | unwind: | |
929 | list_for_each_entry_continue_reverse(d, &iommu->domain_list, next) | |
930 | iommu_unmap(d->domain, iova, npage << PAGE_SHIFT); | |
166fd7d9 | 931 | |
cd9b2268 | 932 | return ret; |
73fa0d10 AW |
933 | } |
934 | ||
8f0d5bb9 KW |
935 | static int vfio_pin_map_dma(struct vfio_iommu *iommu, struct vfio_dma *dma, |
936 | size_t map_size) | |
937 | { | |
938 | dma_addr_t iova = dma->iova; | |
939 | unsigned long vaddr = dma->vaddr; | |
940 | size_t size = map_size; | |
941 | long npage; | |
7cb671e7 AW |
942 | unsigned long pfn, limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT; |
943 | bool lock_cap = capable(CAP_IPC_LOCK); | |
8f0d5bb9 KW |
944 | int ret = 0; |
945 | ||
946 | while (size) { | |
947 | /* Pin a contiguous chunk of memory */ | |
948 | npage = vfio_pin_pages_remote(dma, vaddr + dma->size, | |
7cb671e7 AW |
949 | size >> PAGE_SHIFT, &pfn, |
950 | lock_cap, limit); | |
8f0d5bb9 KW |
951 | if (npage <= 0) { |
952 | WARN_ON(!npage); | |
953 | ret = (int)npage; | |
954 | break; | |
955 | } | |
956 | ||
957 | /* Map it! */ | |
958 | ret = vfio_iommu_map(iommu, iova + dma->size, pfn, npage, | |
959 | dma->prot); | |
960 | if (ret) { | |
a54eb550 KW |
961 | vfio_unpin_pages_remote(dma, iova + dma->size, pfn, |
962 | npage, true); | |
8f0d5bb9 KW |
963 | break; |
964 | } | |
965 | ||
966 | size -= npage << PAGE_SHIFT; | |
967 | dma->size += npage << PAGE_SHIFT; | |
968 | } | |
969 | ||
a54eb550 KW |
970 | dma->iommu_mapped = true; |
971 | ||
8f0d5bb9 KW |
972 | if (ret) |
973 | vfio_remove_dma(iommu, dma); | |
974 | ||
975 | return ret; | |
976 | } | |
977 | ||
73fa0d10 AW |
978 | static int vfio_dma_do_map(struct vfio_iommu *iommu, |
979 | struct vfio_iommu_type1_dma_map *map) | |
980 | { | |
c8dbca16 | 981 | dma_addr_t iova = map->iova; |
166fd7d9 | 982 | unsigned long vaddr = map->vaddr; |
73fa0d10 AW |
983 | size_t size = map->size; |
984 | int ret = 0, prot = 0; | |
985 | uint64_t mask; | |
1ef3e2bc | 986 | struct vfio_dma *dma; |
166fd7d9 | 987 | |
c8dbca16 AW |
988 | /* Verify that none of our __u64 fields overflow */ |
989 | if (map->size != size || map->vaddr != vaddr || map->iova != iova) | |
990 | return -EINVAL; | |
73fa0d10 | 991 | |
1ef3e2bc | 992 | mask = ((uint64_t)1 << __ffs(vfio_pgsize_bitmap(iommu))) - 1; |
73fa0d10 | 993 | |
c8dbca16 AW |
994 | WARN_ON(mask & PAGE_MASK); |
995 | ||
73fa0d10 AW |
996 | /* READ/WRITE from device perspective */ |
997 | if (map->flags & VFIO_DMA_MAP_FLAG_WRITE) | |
998 | prot |= IOMMU_WRITE; | |
999 | if (map->flags & VFIO_DMA_MAP_FLAG_READ) | |
1000 | prot |= IOMMU_READ; | |
1001 | ||
c8dbca16 | 1002 | if (!prot || !size || (size | iova | vaddr) & mask) |
73fa0d10 AW |
1003 | return -EINVAL; |
1004 | ||
c8dbca16 AW |
1005 | /* Don't allow IOVA or virtual address wrap */ |
1006 | if (iova + size - 1 < iova || vaddr + size - 1 < vaddr) | |
73fa0d10 AW |
1007 | return -EINVAL; |
1008 | ||
1009 | mutex_lock(&iommu->lock); | |
1010 | ||
c8dbca16 | 1011 | if (vfio_find_dma(iommu, iova, size)) { |
8f0d5bb9 KW |
1012 | ret = -EEXIST; |
1013 | goto out_unlock; | |
73fa0d10 AW |
1014 | } |
1015 | ||
1ef3e2bc AW |
1016 | dma = kzalloc(sizeof(*dma), GFP_KERNEL); |
1017 | if (!dma) { | |
8f0d5bb9 KW |
1018 | ret = -ENOMEM; |
1019 | goto out_unlock; | |
1ef3e2bc AW |
1020 | } |
1021 | ||
c8dbca16 AW |
1022 | dma->iova = iova; |
1023 | dma->vaddr = vaddr; | |
1ef3e2bc | 1024 | dma->prot = prot; |
8f0d5bb9 KW |
1025 | get_task_struct(current); |
1026 | dma->task = current; | |
a54eb550 | 1027 | dma->pfn_list = RB_ROOT; |
166fd7d9 | 1028 | |
1ef3e2bc AW |
1029 | /* Insert zero-sized and grow as we map chunks of it */ |
1030 | vfio_link_dma(iommu, dma); | |
166fd7d9 | 1031 | |
a54eb550 KW |
1032 | /* Don't pin and map if container doesn't contain IOMMU capable domain*/ |
1033 | if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu)) | |
1034 | dma->size = size; | |
1035 | else | |
1036 | ret = vfio_pin_map_dma(iommu, dma, size); | |
1037 | ||
8f0d5bb9 | 1038 | out_unlock: |
1ef3e2bc AW |
1039 | mutex_unlock(&iommu->lock); |
1040 | return ret; | |
1041 | } | |
1042 | ||
1043 | static int vfio_bus_type(struct device *dev, void *data) | |
1044 | { | |
1045 | struct bus_type **bus = data; | |
1046 | ||
1047 | if (*bus && *bus != dev->bus) | |
1048 | return -EINVAL; | |
1049 | ||
1050 | *bus = dev->bus; | |
1051 | ||
1052 | return 0; | |
1053 | } | |
1054 | ||
1055 | static int vfio_iommu_replay(struct vfio_iommu *iommu, | |
1056 | struct vfio_domain *domain) | |
1057 | { | |
1058 | struct vfio_domain *d; | |
1059 | struct rb_node *n; | |
7cb671e7 AW |
1060 | unsigned long limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT; |
1061 | bool lock_cap = capable(CAP_IPC_LOCK); | |
1ef3e2bc AW |
1062 | int ret; |
1063 | ||
1064 | /* Arbitrarily pick the first domain in the list for lookups */ | |
1065 | d = list_first_entry(&iommu->domain_list, struct vfio_domain, next); | |
1066 | n = rb_first(&iommu->dma_list); | |
1067 | ||
1ef3e2bc AW |
1068 | for (; n; n = rb_next(n)) { |
1069 | struct vfio_dma *dma; | |
1070 | dma_addr_t iova; | |
1071 | ||
1072 | dma = rb_entry(n, struct vfio_dma, node); | |
1073 | iova = dma->iova; | |
1074 | ||
1075 | while (iova < dma->iova + dma->size) { | |
a54eb550 | 1076 | phys_addr_t phys; |
1ef3e2bc | 1077 | size_t size; |
73fa0d10 | 1078 | |
a54eb550 KW |
1079 | if (dma->iommu_mapped) { |
1080 | phys_addr_t p; | |
1081 | dma_addr_t i; | |
1082 | ||
1083 | phys = iommu_iova_to_phys(d->domain, iova); | |
1084 | ||
1085 | if (WARN_ON(!phys)) { | |
1086 | iova += PAGE_SIZE; | |
1087 | continue; | |
1088 | } | |
1089 | ||
1090 | size = PAGE_SIZE; | |
1091 | p = phys + size; | |
1092 | i = iova + size; | |
1093 | while (i < dma->iova + dma->size && | |
1094 | p == iommu_iova_to_phys(d->domain, i)) { | |
1095 | size += PAGE_SIZE; | |
1096 | p += PAGE_SIZE; | |
1097 | i += PAGE_SIZE; | |
1098 | } | |
1099 | } else { | |
1100 | unsigned long pfn; | |
1101 | unsigned long vaddr = dma->vaddr + | |
1102 | (iova - dma->iova); | |
1103 | size_t n = dma->iova + dma->size - iova; | |
1104 | long npage; | |
1105 | ||
1106 | npage = vfio_pin_pages_remote(dma, vaddr, | |
1107 | n >> PAGE_SHIFT, | |
7cb671e7 AW |
1108 | &pfn, lock_cap, |
1109 | limit); | |
a54eb550 KW |
1110 | if (npage <= 0) { |
1111 | WARN_ON(!npage); | |
1112 | ret = (int)npage; | |
1113 | return ret; | |
1114 | } | |
1115 | ||
1116 | phys = pfn << PAGE_SHIFT; | |
1117 | size = npage << PAGE_SHIFT; | |
166fd7d9 AW |
1118 | } |
1119 | ||
1ef3e2bc AW |
1120 | ret = iommu_map(domain->domain, iova, phys, |
1121 | size, dma->prot | domain->prot); | |
1122 | if (ret) | |
1123 | return ret; | |
d93b3ac0 | 1124 | |
1ef3e2bc AW |
1125 | iova += size; |
1126 | } | |
a54eb550 | 1127 | dma->iommu_mapped = true; |
166fd7d9 | 1128 | } |
1ef3e2bc | 1129 | return 0; |
73fa0d10 AW |
1130 | } |
1131 | ||
6fe1010d AW |
1132 | /* |
1133 | * We change our unmap behavior slightly depending on whether the IOMMU | |
1134 | * supports fine-grained superpages. IOMMUs like AMD-Vi will use a superpage | |
1135 | * for practically any contiguous power-of-two mapping we give it. This means | |
1136 | * we don't need to look for contiguous chunks ourselves to make unmapping | |
1137 | * more efficient. On IOMMUs with coarse-grained super pages, like Intel VT-d | |
1138 | * with discrete 2M/1G/512G/1T superpages, identifying contiguous chunks | |
1139 | * significantly boosts non-hugetlbfs mappings and doesn't seem to hurt when | |
1140 | * hugetlbfs is in use. | |
1141 | */ | |
1142 | static void vfio_test_domain_fgsp(struct vfio_domain *domain) | |
1143 | { | |
1144 | struct page *pages; | |
1145 | int ret, order = get_order(PAGE_SIZE * 2); | |
1146 | ||
1147 | pages = alloc_pages(GFP_KERNEL | __GFP_ZERO, order); | |
1148 | if (!pages) | |
1149 | return; | |
1150 | ||
1151 | ret = iommu_map(domain->domain, 0, page_to_phys(pages), PAGE_SIZE * 2, | |
1152 | IOMMU_READ | IOMMU_WRITE | domain->prot); | |
1153 | if (!ret) { | |
1154 | size_t unmapped = iommu_unmap(domain->domain, 0, PAGE_SIZE); | |
1155 | ||
1156 | if (unmapped == PAGE_SIZE) | |
1157 | iommu_unmap(domain->domain, PAGE_SIZE, PAGE_SIZE); | |
1158 | else | |
1159 | domain->fgsp = true; | |
1160 | } | |
1161 | ||
1162 | __free_pages(pages, order); | |
1163 | } | |
1164 | ||
7896c998 KW |
1165 | static struct vfio_group *find_iommu_group(struct vfio_domain *domain, |
1166 | struct iommu_group *iommu_group) | |
1167 | { | |
1168 | struct vfio_group *g; | |
1169 | ||
1170 | list_for_each_entry(g, &domain->group_list, next) { | |
1171 | if (g->iommu_group == iommu_group) | |
1172 | return g; | |
1173 | } | |
1174 | ||
1175 | return NULL; | |
1176 | } | |
1177 | ||
9d3a4de4 | 1178 | static bool vfio_iommu_has_sw_msi(struct iommu_group *group, phys_addr_t *base) |
5d704992 EA |
1179 | { |
1180 | struct list_head group_resv_regions; | |
1181 | struct iommu_resv_region *region, *next; | |
1182 | bool ret = false; | |
1183 | ||
1184 | INIT_LIST_HEAD(&group_resv_regions); | |
1185 | iommu_get_group_resv_regions(group, &group_resv_regions); | |
1186 | list_for_each_entry(region, &group_resv_regions, list) { | |
f203f7f1 RM |
1187 | /* |
1188 | * The presence of any 'real' MSI regions should take | |
1189 | * precedence over the software-managed one if the | |
1190 | * IOMMU driver happens to advertise both types. | |
1191 | */ | |
1192 | if (region->type == IOMMU_RESV_MSI) { | |
1193 | ret = false; | |
1194 | break; | |
1195 | } | |
1196 | ||
9d3a4de4 | 1197 | if (region->type == IOMMU_RESV_SW_MSI) { |
5d704992 EA |
1198 | *base = region->start; |
1199 | ret = true; | |
5d704992 EA |
1200 | } |
1201 | } | |
5d704992 EA |
1202 | list_for_each_entry_safe(region, next, &group_resv_regions, list) |
1203 | kfree(region); | |
1204 | return ret; | |
1205 | } | |
1206 | ||
73fa0d10 AW |
1207 | static int vfio_iommu_type1_attach_group(void *iommu_data, |
1208 | struct iommu_group *iommu_group) | |
1209 | { | |
1210 | struct vfio_iommu *iommu = iommu_data; | |
7896c998 | 1211 | struct vfio_group *group; |
1ef3e2bc | 1212 | struct vfio_domain *domain, *d; |
a54eb550 | 1213 | struct bus_type *bus = NULL, *mdev_bus; |
73fa0d10 | 1214 | int ret; |
9d72f87b | 1215 | bool resv_msi, msi_remap; |
5d704992 | 1216 | phys_addr_t resv_msi_base; |
73fa0d10 | 1217 | |
73fa0d10 AW |
1218 | mutex_lock(&iommu->lock); |
1219 | ||
1ef3e2bc | 1220 | list_for_each_entry(d, &iommu->domain_list, next) { |
7896c998 | 1221 | if (find_iommu_group(d, iommu_group)) { |
73fa0d10 | 1222 | mutex_unlock(&iommu->lock); |
73fa0d10 AW |
1223 | return -EINVAL; |
1224 | } | |
1225 | } | |
1226 | ||
a54eb550 KW |
1227 | if (iommu->external_domain) { |
1228 | if (find_iommu_group(iommu->external_domain, iommu_group)) { | |
1229 | mutex_unlock(&iommu->lock); | |
1230 | return -EINVAL; | |
1231 | } | |
1232 | } | |
1233 | ||
1ef3e2bc AW |
1234 | group = kzalloc(sizeof(*group), GFP_KERNEL); |
1235 | domain = kzalloc(sizeof(*domain), GFP_KERNEL); | |
1236 | if (!group || !domain) { | |
1237 | ret = -ENOMEM; | |
1238 | goto out_free; | |
1239 | } | |
1240 | ||
1241 | group->iommu_group = iommu_group; | |
1242 | ||
1243 | /* Determine bus_type in order to allocate a domain */ | |
1244 | ret = iommu_group_for_each_dev(iommu_group, &bus, vfio_bus_type); | |
1245 | if (ret) | |
1246 | goto out_free; | |
1247 | ||
a54eb550 KW |
1248 | mdev_bus = symbol_get(mdev_bus_type); |
1249 | ||
1250 | if (mdev_bus) { | |
1251 | if ((bus == mdev_bus) && !iommu_present(bus)) { | |
1252 | symbol_put(mdev_bus_type); | |
1253 | if (!iommu->external_domain) { | |
1254 | INIT_LIST_HEAD(&domain->group_list); | |
1255 | iommu->external_domain = domain; | |
1256 | } else | |
1257 | kfree(domain); | |
1258 | ||
1259 | list_add(&group->next, | |
1260 | &iommu->external_domain->group_list); | |
1261 | mutex_unlock(&iommu->lock); | |
1262 | return 0; | |
1263 | } | |
1264 | symbol_put(mdev_bus_type); | |
1265 | } | |
1266 | ||
1ef3e2bc AW |
1267 | domain->domain = iommu_domain_alloc(bus); |
1268 | if (!domain->domain) { | |
1269 | ret = -EIO; | |
1270 | goto out_free; | |
1271 | } | |
1272 | ||
f5c9eceb WD |
1273 | if (iommu->nesting) { |
1274 | int attr = 1; | |
1275 | ||
1276 | ret = iommu_domain_set_attr(domain->domain, DOMAIN_ATTR_NESTING, | |
1277 | &attr); | |
1278 | if (ret) | |
1279 | goto out_domain; | |
1280 | } | |
1281 | ||
1ef3e2bc AW |
1282 | ret = iommu_attach_group(domain->domain, iommu_group); |
1283 | if (ret) | |
1284 | goto out_domain; | |
1285 | ||
9d3a4de4 | 1286 | resv_msi = vfio_iommu_has_sw_msi(iommu_group, &resv_msi_base); |
5d704992 | 1287 | |
1ef3e2bc AW |
1288 | INIT_LIST_HEAD(&domain->group_list); |
1289 | list_add(&group->next, &domain->group_list); | |
1290 | ||
db406cc0 RM |
1291 | msi_remap = irq_domain_check_msi_remap() || |
1292 | iommu_capable(bus, IOMMU_CAP_INTR_REMAP); | |
9d72f87b EA |
1293 | |
1294 | if (!allow_unsafe_interrupts && !msi_remap) { | |
1ef3e2bc AW |
1295 | pr_warn("%s: No interrupt remapping support. Use the module param \"allow_unsafe_interrupts\" to enable VFIO IOMMU support on this platform\n", |
1296 | __func__); | |
1297 | ret = -EPERM; | |
1298 | goto out_detach; | |
1299 | } | |
1300 | ||
eb165f05 | 1301 | if (iommu_capable(bus, IOMMU_CAP_CACHE_COHERENCY)) |
1ef3e2bc AW |
1302 | domain->prot |= IOMMU_CACHE; |
1303 | ||
73fa0d10 | 1304 | /* |
1ef3e2bc AW |
1305 | * Try to match an existing compatible domain. We don't want to |
1306 | * preclude an IOMMU driver supporting multiple bus_types and being | |
1307 | * able to include different bus_types in the same IOMMU domain, so | |
1308 | * we test whether the domains use the same iommu_ops rather than | |
1309 | * testing if they're on the same bus_type. | |
73fa0d10 | 1310 | */ |
1ef3e2bc AW |
1311 | list_for_each_entry(d, &iommu->domain_list, next) { |
1312 | if (d->domain->ops == domain->domain->ops && | |
1313 | d->prot == domain->prot) { | |
1314 | iommu_detach_group(domain->domain, iommu_group); | |
1315 | if (!iommu_attach_group(d->domain, iommu_group)) { | |
1316 | list_add(&group->next, &d->group_list); | |
1317 | iommu_domain_free(domain->domain); | |
1318 | kfree(domain); | |
1319 | mutex_unlock(&iommu->lock); | |
1320 | return 0; | |
1321 | } | |
1322 | ||
1323 | ret = iommu_attach_group(domain->domain, iommu_group); | |
1324 | if (ret) | |
1325 | goto out_domain; | |
1326 | } | |
73fa0d10 AW |
1327 | } |
1328 | ||
6fe1010d AW |
1329 | vfio_test_domain_fgsp(domain); |
1330 | ||
1ef3e2bc AW |
1331 | /* replay mappings on new domains */ |
1332 | ret = vfio_iommu_replay(iommu, domain); | |
1333 | if (ret) | |
1334 | goto out_detach; | |
1335 | ||
2c9f1af5 WY |
1336 | if (resv_msi) { |
1337 | ret = iommu_get_msi_cookie(domain->domain, resv_msi_base); | |
1338 | if (ret) | |
1339 | goto out_detach; | |
1340 | } | |
5d704992 | 1341 | |
1ef3e2bc | 1342 | list_add(&domain->next, &iommu->domain_list); |
73fa0d10 AW |
1343 | |
1344 | mutex_unlock(&iommu->lock); | |
1345 | ||
1346 | return 0; | |
1ef3e2bc AW |
1347 | |
1348 | out_detach: | |
1349 | iommu_detach_group(domain->domain, iommu_group); | |
1350 | out_domain: | |
1351 | iommu_domain_free(domain->domain); | |
1352 | out_free: | |
1353 | kfree(domain); | |
1354 | kfree(group); | |
1355 | mutex_unlock(&iommu->lock); | |
1356 | return ret; | |
1357 | } | |
1358 | ||
1359 | static void vfio_iommu_unmap_unpin_all(struct vfio_iommu *iommu) | |
1360 | { | |
1361 | struct rb_node *node; | |
1362 | ||
1363 | while ((node = rb_first(&iommu->dma_list))) | |
1364 | vfio_remove_dma(iommu, rb_entry(node, struct vfio_dma, node)); | |
73fa0d10 AW |
1365 | } |
1366 | ||
a54eb550 KW |
1367 | static void vfio_iommu_unmap_unpin_reaccount(struct vfio_iommu *iommu) |
1368 | { | |
1369 | struct rb_node *n, *p; | |
1370 | ||
1371 | n = rb_first(&iommu->dma_list); | |
1372 | for (; n; n = rb_next(n)) { | |
1373 | struct vfio_dma *dma; | |
1374 | long locked = 0, unlocked = 0; | |
1375 | ||
1376 | dma = rb_entry(n, struct vfio_dma, node); | |
1377 | unlocked += vfio_unmap_unpin(iommu, dma, false); | |
1378 | p = rb_first(&dma->pfn_list); | |
1379 | for (; p; p = rb_next(p)) { | |
1380 | struct vfio_pfn *vpfn = rb_entry(p, struct vfio_pfn, | |
1381 | node); | |
1382 | ||
1383 | if (!is_invalid_reserved_pfn(vpfn->pfn)) | |
1384 | locked++; | |
1385 | } | |
0cfef2b7 | 1386 | vfio_lock_acct(dma->task, locked - unlocked, NULL); |
a54eb550 KW |
1387 | } |
1388 | } | |
1389 | ||
1390 | static void vfio_sanity_check_pfn_list(struct vfio_iommu *iommu) | |
1391 | { | |
1392 | struct rb_node *n; | |
1393 | ||
1394 | n = rb_first(&iommu->dma_list); | |
1395 | for (; n; n = rb_next(n)) { | |
1396 | struct vfio_dma *dma; | |
1397 | ||
1398 | dma = rb_entry(n, struct vfio_dma, node); | |
1399 | ||
1400 | if (WARN_ON(!RB_EMPTY_ROOT(&dma->pfn_list))) | |
1401 | break; | |
1402 | } | |
3cedd7d7 KW |
1403 | /* mdev vendor driver must unregister notifier */ |
1404 | WARN_ON(iommu->notifier.head); | |
a54eb550 KW |
1405 | } |
1406 | ||
73fa0d10 AW |
1407 | static void vfio_iommu_type1_detach_group(void *iommu_data, |
1408 | struct iommu_group *iommu_group) | |
1409 | { | |
1410 | struct vfio_iommu *iommu = iommu_data; | |
1ef3e2bc | 1411 | struct vfio_domain *domain; |
73fa0d10 AW |
1412 | struct vfio_group *group; |
1413 | ||
1414 | mutex_lock(&iommu->lock); | |
1415 | ||
a54eb550 KW |
1416 | if (iommu->external_domain) { |
1417 | group = find_iommu_group(iommu->external_domain, iommu_group); | |
1418 | if (group) { | |
1419 | list_del(&group->next); | |
1420 | kfree(group); | |
1421 | ||
1422 | if (list_empty(&iommu->external_domain->group_list)) { | |
1423 | vfio_sanity_check_pfn_list(iommu); | |
1424 | ||
1425 | if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu)) | |
1426 | vfio_iommu_unmap_unpin_all(iommu); | |
1427 | ||
1428 | kfree(iommu->external_domain); | |
1429 | iommu->external_domain = NULL; | |
1430 | } | |
1431 | goto detach_group_done; | |
1432 | } | |
1433 | } | |
1434 | ||
1ef3e2bc | 1435 | list_for_each_entry(domain, &iommu->domain_list, next) { |
7896c998 KW |
1436 | group = find_iommu_group(domain, iommu_group); |
1437 | if (!group) | |
1438 | continue; | |
1ef3e2bc | 1439 | |
7896c998 KW |
1440 | iommu_detach_group(domain->domain, iommu_group); |
1441 | list_del(&group->next); | |
1442 | kfree(group); | |
1443 | /* | |
a54eb550 KW |
1444 | * Group ownership provides privilege, if the group list is |
1445 | * empty, the domain goes away. If it's the last domain with | |
1446 | * iommu and external domain doesn't exist, then all the | |
1447 | * mappings go away too. If it's the last domain with iommu and | |
1448 | * external domain exist, update accounting | |
7896c998 KW |
1449 | */ |
1450 | if (list_empty(&domain->group_list)) { | |
a54eb550 KW |
1451 | if (list_is_singular(&iommu->domain_list)) { |
1452 | if (!iommu->external_domain) | |
1453 | vfio_iommu_unmap_unpin_all(iommu); | |
1454 | else | |
1455 | vfio_iommu_unmap_unpin_reaccount(iommu); | |
1456 | } | |
7896c998 KW |
1457 | iommu_domain_free(domain->domain); |
1458 | list_del(&domain->next); | |
1459 | kfree(domain); | |
73fa0d10 | 1460 | } |
a54eb550 | 1461 | break; |
73fa0d10 AW |
1462 | } |
1463 | ||
a54eb550 | 1464 | detach_group_done: |
73fa0d10 AW |
1465 | mutex_unlock(&iommu->lock); |
1466 | } | |
1467 | ||
1468 | static void *vfio_iommu_type1_open(unsigned long arg) | |
1469 | { | |
1470 | struct vfio_iommu *iommu; | |
1471 | ||
73fa0d10 AW |
1472 | iommu = kzalloc(sizeof(*iommu), GFP_KERNEL); |
1473 | if (!iommu) | |
1474 | return ERR_PTR(-ENOMEM); | |
1475 | ||
f5c9eceb WD |
1476 | switch (arg) { |
1477 | case VFIO_TYPE1_IOMMU: | |
1478 | break; | |
1479 | case VFIO_TYPE1_NESTING_IOMMU: | |
1480 | iommu->nesting = true; | |
1481 | case VFIO_TYPE1v2_IOMMU: | |
1482 | iommu->v2 = true; | |
1483 | break; | |
1484 | default: | |
1485 | kfree(iommu); | |
1486 | return ERR_PTR(-EINVAL); | |
1487 | } | |
1488 | ||
1ef3e2bc | 1489 | INIT_LIST_HEAD(&iommu->domain_list); |
cd9b2268 | 1490 | iommu->dma_list = RB_ROOT; |
73fa0d10 | 1491 | mutex_init(&iommu->lock); |
c086de81 | 1492 | BLOCKING_INIT_NOTIFIER_HEAD(&iommu->notifier); |
73fa0d10 AW |
1493 | |
1494 | return iommu; | |
1495 | } | |
1496 | ||
a54eb550 KW |
1497 | static void vfio_release_domain(struct vfio_domain *domain, bool external) |
1498 | { | |
1499 | struct vfio_group *group, *group_tmp; | |
1500 | ||
1501 | list_for_each_entry_safe(group, group_tmp, | |
1502 | &domain->group_list, next) { | |
1503 | if (!external) | |
1504 | iommu_detach_group(domain->domain, group->iommu_group); | |
1505 | list_del(&group->next); | |
1506 | kfree(group); | |
1507 | } | |
1508 | ||
1509 | if (!external) | |
1510 | iommu_domain_free(domain->domain); | |
1511 | } | |
1512 | ||
73fa0d10 AW |
1513 | static void vfio_iommu_type1_release(void *iommu_data) |
1514 | { | |
1515 | struct vfio_iommu *iommu = iommu_data; | |
1ef3e2bc | 1516 | struct vfio_domain *domain, *domain_tmp; |
a54eb550 KW |
1517 | |
1518 | if (iommu->external_domain) { | |
1519 | vfio_release_domain(iommu->external_domain, true); | |
1520 | vfio_sanity_check_pfn_list(iommu); | |
1521 | kfree(iommu->external_domain); | |
1522 | } | |
73fa0d10 | 1523 | |
1ef3e2bc | 1524 | vfio_iommu_unmap_unpin_all(iommu); |
73fa0d10 | 1525 | |
1ef3e2bc AW |
1526 | list_for_each_entry_safe(domain, domain_tmp, |
1527 | &iommu->domain_list, next) { | |
a54eb550 | 1528 | vfio_release_domain(domain, false); |
1ef3e2bc AW |
1529 | list_del(&domain->next); |
1530 | kfree(domain); | |
73fa0d10 | 1531 | } |
73fa0d10 AW |
1532 | kfree(iommu); |
1533 | } | |
1534 | ||
aa429318 AW |
1535 | static int vfio_domains_have_iommu_cache(struct vfio_iommu *iommu) |
1536 | { | |
1537 | struct vfio_domain *domain; | |
1538 | int ret = 1; | |
1539 | ||
1540 | mutex_lock(&iommu->lock); | |
1541 | list_for_each_entry(domain, &iommu->domain_list, next) { | |
1542 | if (!(domain->prot & IOMMU_CACHE)) { | |
1543 | ret = 0; | |
f5bfdbf2 | 1544 | break; |
aa429318 | 1545 | } |
73fa0d10 | 1546 | } |
aa429318 | 1547 | mutex_unlock(&iommu->lock); |
73fa0d10 | 1548 | |
aa429318 | 1549 | return ret; |
73fa0d10 AW |
1550 | } |
1551 | ||
1552 | static long vfio_iommu_type1_ioctl(void *iommu_data, | |
1553 | unsigned int cmd, unsigned long arg) | |
1554 | { | |
1555 | struct vfio_iommu *iommu = iommu_data; | |
1556 | unsigned long minsz; | |
1557 | ||
1558 | if (cmd == VFIO_CHECK_EXTENSION) { | |
1559 | switch (arg) { | |
1560 | case VFIO_TYPE1_IOMMU: | |
1ef3e2bc | 1561 | case VFIO_TYPE1v2_IOMMU: |
f5c9eceb | 1562 | case VFIO_TYPE1_NESTING_IOMMU: |
73fa0d10 | 1563 | return 1; |
aa429318 AW |
1564 | case VFIO_DMA_CC_IOMMU: |
1565 | if (!iommu) | |
1566 | return 0; | |
1567 | return vfio_domains_have_iommu_cache(iommu); | |
73fa0d10 AW |
1568 | default: |
1569 | return 0; | |
1570 | } | |
1571 | } else if (cmd == VFIO_IOMMU_GET_INFO) { | |
1572 | struct vfio_iommu_type1_info info; | |
1573 | ||
1574 | minsz = offsetofend(struct vfio_iommu_type1_info, iova_pgsizes); | |
1575 | ||
1576 | if (copy_from_user(&info, (void __user *)arg, minsz)) | |
1577 | return -EFAULT; | |
1578 | ||
1579 | if (info.argsz < minsz) | |
1580 | return -EINVAL; | |
1581 | ||
d4f50ee2 | 1582 | info.flags = VFIO_IOMMU_INFO_PGSIZES; |
73fa0d10 | 1583 | |
1ef3e2bc | 1584 | info.iova_pgsizes = vfio_pgsize_bitmap(iommu); |
73fa0d10 | 1585 | |
8160c4e4 MT |
1586 | return copy_to_user((void __user *)arg, &info, minsz) ? |
1587 | -EFAULT : 0; | |
73fa0d10 AW |
1588 | |
1589 | } else if (cmd == VFIO_IOMMU_MAP_DMA) { | |
1590 | struct vfio_iommu_type1_dma_map map; | |
1591 | uint32_t mask = VFIO_DMA_MAP_FLAG_READ | | |
1592 | VFIO_DMA_MAP_FLAG_WRITE; | |
1593 | ||
1594 | minsz = offsetofend(struct vfio_iommu_type1_dma_map, size); | |
1595 | ||
1596 | if (copy_from_user(&map, (void __user *)arg, minsz)) | |
1597 | return -EFAULT; | |
1598 | ||
1599 | if (map.argsz < minsz || map.flags & ~mask) | |
1600 | return -EINVAL; | |
1601 | ||
1602 | return vfio_dma_do_map(iommu, &map); | |
1603 | ||
1604 | } else if (cmd == VFIO_IOMMU_UNMAP_DMA) { | |
1605 | struct vfio_iommu_type1_dma_unmap unmap; | |
166fd7d9 | 1606 | long ret; |
73fa0d10 AW |
1607 | |
1608 | minsz = offsetofend(struct vfio_iommu_type1_dma_unmap, size); | |
1609 | ||
1610 | if (copy_from_user(&unmap, (void __user *)arg, minsz)) | |
1611 | return -EFAULT; | |
1612 | ||
1613 | if (unmap.argsz < minsz || unmap.flags) | |
1614 | return -EINVAL; | |
1615 | ||
166fd7d9 AW |
1616 | ret = vfio_dma_do_unmap(iommu, &unmap); |
1617 | if (ret) | |
1618 | return ret; | |
1619 | ||
8160c4e4 MT |
1620 | return copy_to_user((void __user *)arg, &unmap, minsz) ? |
1621 | -EFAULT : 0; | |
73fa0d10 AW |
1622 | } |
1623 | ||
1624 | return -ENOTTY; | |
1625 | } | |
1626 | ||
c086de81 | 1627 | static int vfio_iommu_type1_register_notifier(void *iommu_data, |
22195cbd | 1628 | unsigned long *events, |
c086de81 KW |
1629 | struct notifier_block *nb) |
1630 | { | |
1631 | struct vfio_iommu *iommu = iommu_data; | |
1632 | ||
22195cbd JS |
1633 | /* clear known events */ |
1634 | *events &= ~VFIO_IOMMU_NOTIFY_DMA_UNMAP; | |
1635 | ||
1636 | /* refuse to register if still events remaining */ | |
1637 | if (*events) | |
1638 | return -EINVAL; | |
1639 | ||
c086de81 KW |
1640 | return blocking_notifier_chain_register(&iommu->notifier, nb); |
1641 | } | |
1642 | ||
1643 | static int vfio_iommu_type1_unregister_notifier(void *iommu_data, | |
1644 | struct notifier_block *nb) | |
1645 | { | |
1646 | struct vfio_iommu *iommu = iommu_data; | |
1647 | ||
1648 | return blocking_notifier_chain_unregister(&iommu->notifier, nb); | |
1649 | } | |
1650 | ||
73fa0d10 | 1651 | static const struct vfio_iommu_driver_ops vfio_iommu_driver_ops_type1 = { |
c086de81 KW |
1652 | .name = "vfio-iommu-type1", |
1653 | .owner = THIS_MODULE, | |
1654 | .open = vfio_iommu_type1_open, | |
1655 | .release = vfio_iommu_type1_release, | |
1656 | .ioctl = vfio_iommu_type1_ioctl, | |
1657 | .attach_group = vfio_iommu_type1_attach_group, | |
1658 | .detach_group = vfio_iommu_type1_detach_group, | |
1659 | .pin_pages = vfio_iommu_type1_pin_pages, | |
1660 | .unpin_pages = vfio_iommu_type1_unpin_pages, | |
1661 | .register_notifier = vfio_iommu_type1_register_notifier, | |
1662 | .unregister_notifier = vfio_iommu_type1_unregister_notifier, | |
73fa0d10 AW |
1663 | }; |
1664 | ||
1665 | static int __init vfio_iommu_type1_init(void) | |
1666 | { | |
73fa0d10 AW |
1667 | return vfio_register_iommu_driver(&vfio_iommu_driver_ops_type1); |
1668 | } | |
1669 | ||
1670 | static void __exit vfio_iommu_type1_cleanup(void) | |
1671 | { | |
1672 | vfio_unregister_iommu_driver(&vfio_iommu_driver_ops_type1); | |
1673 | } | |
1674 | ||
1675 | module_init(vfio_iommu_type1_init); | |
1676 | module_exit(vfio_iommu_type1_cleanup); | |
1677 | ||
1678 | MODULE_VERSION(DRIVER_VERSION); | |
1679 | MODULE_LICENSE("GPL v2"); | |
1680 | MODULE_AUTHOR(DRIVER_AUTHOR); | |
1681 | MODULE_DESCRIPTION(DRIVER_DESC); |