]>
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; | |
ea85cf35 | 341 | int ret; |
73fa0d10 | 342 | |
ea85cf35 KW |
343 | if (mm == current->mm) { |
344 | ret = get_user_pages_fast(vaddr, 1, !!(prot & IOMMU_WRITE), | |
345 | page); | |
346 | } else { | |
347 | unsigned int flags = 0; | |
348 | ||
349 | if (prot & IOMMU_WRITE) | |
350 | flags |= FOLL_WRITE; | |
351 | ||
352 | down_read(&mm->mmap_sem); | |
353 | ret = get_user_pages_remote(NULL, mm, vaddr, 1, flags, page, | |
5b56d49f | 354 | NULL, NULL); |
ea85cf35 KW |
355 | up_read(&mm->mmap_sem); |
356 | } | |
357 | ||
358 | if (ret == 1) { | |
73fa0d10 AW |
359 | *pfn = page_to_pfn(page[0]); |
360 | return 0; | |
361 | } | |
362 | ||
ea85cf35 | 363 | down_read(&mm->mmap_sem); |
73fa0d10 | 364 | |
ea85cf35 | 365 | vma = find_vma_intersection(mm, vaddr, vaddr + 1); |
73fa0d10 AW |
366 | |
367 | if (vma && vma->vm_flags & VM_PFNMAP) { | |
368 | *pfn = ((vaddr - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff; | |
369 | if (is_invalid_reserved_pfn(*pfn)) | |
370 | ret = 0; | |
371 | } | |
372 | ||
ea85cf35 | 373 | up_read(&mm->mmap_sem); |
73fa0d10 AW |
374 | return ret; |
375 | } | |
376 | ||
166fd7d9 AW |
377 | /* |
378 | * Attempt to pin pages. We really don't want to track all the pfns and | |
379 | * the iommu can only map chunks of consecutive pfns anyway, so get the | |
380 | * first page and all consecutive pages with the same locking. | |
381 | */ | |
8f0d5bb9 | 382 | static long vfio_pin_pages_remote(struct vfio_dma *dma, unsigned long vaddr, |
7cb671e7 AW |
383 | long npage, unsigned long *pfn_base, |
384 | bool lock_cap, unsigned long limit) | |
73fa0d10 | 385 | { |
7cb671e7 | 386 | unsigned long pfn = 0; |
6c38c055 | 387 | long ret, pinned = 0, lock_acct = 0; |
babbf176 | 388 | bool rsvd; |
a54eb550 | 389 | dma_addr_t iova = vaddr - dma->vaddr + dma->iova; |
73fa0d10 | 390 | |
6c38c055 AW |
391 | /* This code path is only user initiated */ |
392 | if (!current->mm) | |
166fd7d9 | 393 | return -ENODEV; |
73fa0d10 | 394 | |
6c38c055 | 395 | ret = vaddr_get_pfn(current->mm, vaddr, dma->prot, pfn_base); |
166fd7d9 | 396 | if (ret) |
6c38c055 | 397 | return ret; |
73fa0d10 | 398 | |
6c38c055 | 399 | pinned++; |
babbf176 | 400 | rsvd = is_invalid_reserved_pfn(*pfn_base); |
73fa0d10 | 401 | |
a54eb550 KW |
402 | /* |
403 | * Reserved pages aren't counted against the user, externally pinned | |
404 | * pages are already counted against the user. | |
405 | */ | |
406 | if (!rsvd && !vfio_find_vpfn(dma, iova)) { | |
6c38c055 | 407 | if (!lock_cap && current->mm->locked_vm + 1 > limit) { |
a54eb550 KW |
408 | put_pfn(*pfn_base, dma->prot); |
409 | pr_warn("%s: RLIMIT_MEMLOCK (%ld) exceeded\n", __func__, | |
410 | limit << PAGE_SHIFT); | |
6c38c055 | 411 | return -ENOMEM; |
a54eb550 KW |
412 | } |
413 | lock_acct++; | |
5c6c2b21 AW |
414 | } |
415 | ||
6c38c055 AW |
416 | if (unlikely(disable_hugepages)) |
417 | goto out; | |
73fa0d10 | 418 | |
6c38c055 AW |
419 | /* Lock all the consecutive pages from pfn_base */ |
420 | for (vaddr += PAGE_SIZE, iova += PAGE_SIZE; pinned < npage; | |
421 | pinned++, vaddr += PAGE_SIZE, iova += PAGE_SIZE) { | |
6c38c055 AW |
422 | ret = vaddr_get_pfn(current->mm, vaddr, dma->prot, &pfn); |
423 | if (ret) | |
424 | break; | |
425 | ||
426 | if (pfn != *pfn_base + pinned || | |
427 | rsvd != is_invalid_reserved_pfn(pfn)) { | |
428 | put_pfn(pfn, dma->prot); | |
429 | break; | |
430 | } | |
166fd7d9 | 431 | |
6c38c055 AW |
432 | if (!rsvd && !vfio_find_vpfn(dma, iova)) { |
433 | if (!lock_cap && | |
434 | current->mm->locked_vm + lock_acct + 1 > limit) { | |
a54eb550 | 435 | put_pfn(pfn, dma->prot); |
6c38c055 AW |
436 | pr_warn("%s: RLIMIT_MEMLOCK (%ld) exceeded\n", |
437 | __func__, limit << PAGE_SHIFT); | |
0cfef2b7 AW |
438 | ret = -ENOMEM; |
439 | goto unpin_out; | |
a54eb550 | 440 | } |
6c38c055 | 441 | lock_acct++; |
166fd7d9 AW |
442 | } |
443 | } | |
444 | ||
6c38c055 | 445 | out: |
0cfef2b7 AW |
446 | ret = vfio_lock_acct(current, lock_acct, &lock_cap); |
447 | ||
448 | unpin_out: | |
449 | if (ret) { | |
450 | if (!rsvd) { | |
451 | for (pfn = *pfn_base ; pinned ; pfn++, pinned--) | |
452 | put_pfn(pfn, dma->prot); | |
453 | } | |
454 | ||
455 | return ret; | |
456 | } | |
166fd7d9 | 457 | |
6c38c055 | 458 | return pinned; |
166fd7d9 AW |
459 | } |
460 | ||
a54eb550 KW |
461 | static long vfio_unpin_pages_remote(struct vfio_dma *dma, dma_addr_t iova, |
462 | unsigned long pfn, long npage, | |
463 | bool do_accounting) | |
166fd7d9 | 464 | { |
a54eb550 | 465 | long unlocked = 0, locked = 0; |
166fd7d9 AW |
466 | long i; |
467 | ||
6c38c055 | 468 | for (i = 0; i < npage; i++, iova += PAGE_SIZE) { |
a54eb550 KW |
469 | if (put_pfn(pfn++, dma->prot)) { |
470 | unlocked++; | |
6c38c055 | 471 | if (vfio_find_vpfn(dma, iova)) |
a54eb550 KW |
472 | locked++; |
473 | } | |
474 | } | |
475 | ||
476 | if (do_accounting) | |
0cfef2b7 | 477 | vfio_lock_acct(dma->task, locked - unlocked, NULL); |
a54eb550 KW |
478 | |
479 | return unlocked; | |
480 | } | |
481 | ||
482 | static int vfio_pin_page_external(struct vfio_dma *dma, unsigned long vaddr, | |
483 | unsigned long *pfn_base, bool do_accounting) | |
484 | { | |
a54eb550 KW |
485 | struct mm_struct *mm; |
486 | int ret; | |
a54eb550 KW |
487 | |
488 | mm = get_task_mm(dma->task); | |
489 | if (!mm) | |
490 | return -ENODEV; | |
491 | ||
492 | ret = vaddr_get_pfn(mm, vaddr, dma->prot, pfn_base); | |
80dbe1fb AW |
493 | if (!ret && do_accounting && !is_invalid_reserved_pfn(*pfn_base)) { |
494 | ret = vfio_lock_acct(dma->task, 1, NULL); | |
0cfef2b7 AW |
495 | if (ret) { |
496 | put_pfn(*pfn_base, dma->prot); | |
80dbe1fb AW |
497 | if (ret == -ENOMEM) |
498 | pr_warn("%s: Task %s (%d) RLIMIT_MEMLOCK " | |
499 | "(%ld) exceeded\n", __func__, | |
500 | dma->task->comm, task_pid_nr(dma->task), | |
501 | task_rlimit(dma->task, RLIMIT_MEMLOCK)); | |
0cfef2b7 AW |
502 | } |
503 | } | |
504 | ||
a54eb550 KW |
505 | mmput(mm); |
506 | return ret; | |
507 | } | |
508 | ||
509 | static int vfio_unpin_page_external(struct vfio_dma *dma, dma_addr_t iova, | |
510 | bool do_accounting) | |
511 | { | |
512 | int unlocked; | |
513 | struct vfio_pfn *vpfn = vfio_find_vpfn(dma, iova); | |
514 | ||
515 | if (!vpfn) | |
516 | return 0; | |
517 | ||
518 | unlocked = vfio_iova_put_vfio_pfn(dma, vpfn); | |
166fd7d9 AW |
519 | |
520 | if (do_accounting) | |
0cfef2b7 | 521 | vfio_lock_acct(dma->task, -unlocked, NULL); |
166fd7d9 AW |
522 | |
523 | return unlocked; | |
524 | } | |
525 | ||
a54eb550 KW |
526 | static int vfio_iommu_type1_pin_pages(void *iommu_data, |
527 | unsigned long *user_pfn, | |
528 | int npage, int prot, | |
529 | unsigned long *phys_pfn) | |
530 | { | |
531 | struct vfio_iommu *iommu = iommu_data; | |
532 | int i, j, ret; | |
533 | unsigned long remote_vaddr; | |
534 | struct vfio_dma *dma; | |
535 | bool do_accounting; | |
536 | ||
537 | if (!iommu || !user_pfn || !phys_pfn) | |
538 | return -EINVAL; | |
539 | ||
540 | /* Supported for v2 version only */ | |
541 | if (!iommu->v2) | |
542 | return -EACCES; | |
543 | ||
544 | mutex_lock(&iommu->lock); | |
545 | ||
c086de81 KW |
546 | /* Fail if notifier list is empty */ |
547 | if ((!iommu->external_domain) || (!iommu->notifier.head)) { | |
a54eb550 KW |
548 | ret = -EINVAL; |
549 | goto pin_done; | |
550 | } | |
551 | ||
552 | /* | |
553 | * If iommu capable domain exist in the container then all pages are | |
554 | * already pinned and accounted. Accouting should be done if there is no | |
555 | * iommu capable domain in the container. | |
556 | */ | |
557 | do_accounting = !IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu); | |
558 | ||
559 | for (i = 0; i < npage; i++) { | |
560 | dma_addr_t iova; | |
561 | struct vfio_pfn *vpfn; | |
562 | ||
563 | iova = user_pfn[i] << PAGE_SHIFT; | |
2b8bb1d7 | 564 | dma = vfio_find_dma(iommu, iova, PAGE_SIZE); |
a54eb550 KW |
565 | if (!dma) { |
566 | ret = -EINVAL; | |
567 | goto pin_unwind; | |
568 | } | |
569 | ||
570 | if ((dma->prot & prot) != prot) { | |
571 | ret = -EPERM; | |
572 | goto pin_unwind; | |
573 | } | |
574 | ||
575 | vpfn = vfio_iova_get_vfio_pfn(dma, iova); | |
576 | if (vpfn) { | |
577 | phys_pfn[i] = vpfn->pfn; | |
578 | continue; | |
579 | } | |
580 | ||
581 | remote_vaddr = dma->vaddr + iova - dma->iova; | |
582 | ret = vfio_pin_page_external(dma, remote_vaddr, &phys_pfn[i], | |
583 | do_accounting); | |
80dbe1fb | 584 | if (ret) |
a54eb550 | 585 | goto pin_unwind; |
a54eb550 KW |
586 | |
587 | ret = vfio_add_to_pfn_list(dma, iova, phys_pfn[i]); | |
588 | if (ret) { | |
589 | vfio_unpin_page_external(dma, iova, do_accounting); | |
590 | goto pin_unwind; | |
591 | } | |
592 | } | |
593 | ||
594 | ret = i; | |
595 | goto pin_done; | |
596 | ||
597 | pin_unwind: | |
598 | phys_pfn[i] = 0; | |
599 | for (j = 0; j < i; j++) { | |
600 | dma_addr_t iova; | |
601 | ||
602 | iova = user_pfn[j] << PAGE_SHIFT; | |
2b8bb1d7 | 603 | dma = vfio_find_dma(iommu, iova, PAGE_SIZE); |
a54eb550 KW |
604 | vfio_unpin_page_external(dma, iova, do_accounting); |
605 | phys_pfn[j] = 0; | |
606 | } | |
607 | pin_done: | |
608 | mutex_unlock(&iommu->lock); | |
609 | return ret; | |
610 | } | |
611 | ||
612 | static int vfio_iommu_type1_unpin_pages(void *iommu_data, | |
613 | unsigned long *user_pfn, | |
614 | int npage) | |
615 | { | |
616 | struct vfio_iommu *iommu = iommu_data; | |
617 | bool do_accounting; | |
618 | int i; | |
619 | ||
620 | if (!iommu || !user_pfn) | |
621 | return -EINVAL; | |
622 | ||
623 | /* Supported for v2 version only */ | |
624 | if (!iommu->v2) | |
625 | return -EACCES; | |
626 | ||
627 | mutex_lock(&iommu->lock); | |
628 | ||
629 | if (!iommu->external_domain) { | |
630 | mutex_unlock(&iommu->lock); | |
631 | return -EINVAL; | |
632 | } | |
633 | ||
634 | do_accounting = !IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu); | |
635 | for (i = 0; i < npage; i++) { | |
636 | struct vfio_dma *dma; | |
637 | dma_addr_t iova; | |
638 | ||
639 | iova = user_pfn[i] << PAGE_SHIFT; | |
2b8bb1d7 | 640 | dma = vfio_find_dma(iommu, iova, PAGE_SIZE); |
a54eb550 KW |
641 | if (!dma) |
642 | goto unpin_exit; | |
643 | vfio_unpin_page_external(dma, iova, do_accounting); | |
644 | } | |
645 | ||
646 | unpin_exit: | |
647 | mutex_unlock(&iommu->lock); | |
648 | return i > npage ? npage : (i > 0 ? i : -EINVAL); | |
649 | } | |
650 | ||
651 | static long vfio_unmap_unpin(struct vfio_iommu *iommu, struct vfio_dma *dma, | |
652 | bool do_accounting) | |
166fd7d9 | 653 | { |
1ef3e2bc AW |
654 | dma_addr_t iova = dma->iova, end = dma->iova + dma->size; |
655 | struct vfio_domain *domain, *d; | |
166fd7d9 AW |
656 | long unlocked = 0; |
657 | ||
1ef3e2bc | 658 | if (!dma->size) |
a54eb550 KW |
659 | return 0; |
660 | ||
661 | if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu)) | |
662 | return 0; | |
663 | ||
1ef3e2bc AW |
664 | /* |
665 | * We use the IOMMU to track the physical addresses, otherwise we'd | |
666 | * need a much more complicated tracking system. Unfortunately that | |
667 | * means we need to use one of the iommu domains to figure out the | |
668 | * pfns to unpin. The rest need to be unmapped in advance so we have | |
669 | * no iommu translations remaining when the pages are unpinned. | |
670 | */ | |
671 | domain = d = list_first_entry(&iommu->domain_list, | |
672 | struct vfio_domain, next); | |
673 | ||
c5e66887 | 674 | list_for_each_entry_continue(d, &iommu->domain_list, next) { |
1ef3e2bc | 675 | iommu_unmap(d->domain, dma->iova, dma->size); |
c5e66887 AW |
676 | cond_resched(); |
677 | } | |
1ef3e2bc | 678 | |
166fd7d9 | 679 | while (iova < end) { |
6fe1010d AW |
680 | size_t unmapped, len; |
681 | phys_addr_t phys, next; | |
166fd7d9 | 682 | |
1ef3e2bc | 683 | phys = iommu_iova_to_phys(domain->domain, iova); |
166fd7d9 AW |
684 | if (WARN_ON(!phys)) { |
685 | iova += PAGE_SIZE; | |
686 | continue; | |
73fa0d10 | 687 | } |
166fd7d9 | 688 | |
6fe1010d AW |
689 | /* |
690 | * To optimize for fewer iommu_unmap() calls, each of which | |
691 | * may require hardware cache flushing, try to find the | |
692 | * largest contiguous physical memory chunk to unmap. | |
693 | */ | |
694 | for (len = PAGE_SIZE; | |
695 | !domain->fgsp && iova + len < end; len += PAGE_SIZE) { | |
696 | next = iommu_iova_to_phys(domain->domain, iova + len); | |
697 | if (next != phys + len) | |
698 | break; | |
699 | } | |
700 | ||
701 | unmapped = iommu_unmap(domain->domain, iova, len); | |
1ef3e2bc | 702 | if (WARN_ON(!unmapped)) |
166fd7d9 AW |
703 | break; |
704 | ||
a54eb550 KW |
705 | unlocked += vfio_unpin_pages_remote(dma, iova, |
706 | phys >> PAGE_SHIFT, | |
2169037d | 707 | unmapped >> PAGE_SHIFT, |
a54eb550 | 708 | false); |
166fd7d9 | 709 | iova += unmapped; |
c5e66887 AW |
710 | |
711 | cond_resched(); | |
73fa0d10 | 712 | } |
166fd7d9 | 713 | |
a54eb550 KW |
714 | dma->iommu_mapped = false; |
715 | if (do_accounting) { | |
0cfef2b7 | 716 | vfio_lock_acct(dma->task, -unlocked, NULL); |
a54eb550 KW |
717 | return 0; |
718 | } | |
719 | return unlocked; | |
73fa0d10 AW |
720 | } |
721 | ||
1ef3e2bc | 722 | static void vfio_remove_dma(struct vfio_iommu *iommu, struct vfio_dma *dma) |
73fa0d10 | 723 | { |
a54eb550 | 724 | vfio_unmap_unpin(iommu, dma, true); |
1ef3e2bc | 725 | vfio_unlink_dma(iommu, dma); |
8f0d5bb9 | 726 | put_task_struct(dma->task); |
1ef3e2bc AW |
727 | kfree(dma); |
728 | } | |
73fa0d10 | 729 | |
1ef3e2bc AW |
730 | static unsigned long vfio_pgsize_bitmap(struct vfio_iommu *iommu) |
731 | { | |
732 | struct vfio_domain *domain; | |
4644321f | 733 | unsigned long bitmap = ULONG_MAX; |
166fd7d9 | 734 | |
1ef3e2bc AW |
735 | mutex_lock(&iommu->lock); |
736 | list_for_each_entry(domain, &iommu->domain_list, next) | |
d16e0faa | 737 | bitmap &= domain->domain->pgsize_bitmap; |
1ef3e2bc | 738 | mutex_unlock(&iommu->lock); |
73fa0d10 | 739 | |
4644321f EA |
740 | /* |
741 | * In case the IOMMU supports page sizes smaller than PAGE_SIZE | |
742 | * we pretend PAGE_SIZE is supported and hide sub-PAGE_SIZE sizes. | |
743 | * That way the user will be able to map/unmap buffers whose size/ | |
744 | * start address is aligned with PAGE_SIZE. Pinning code uses that | |
745 | * granularity while iommu driver can use the sub-PAGE_SIZE size | |
746 | * to map the buffer. | |
747 | */ | |
748 | if (bitmap & ~PAGE_MASK) { | |
749 | bitmap &= PAGE_MASK; | |
750 | bitmap |= PAGE_SIZE; | |
751 | } | |
752 | ||
1ef3e2bc | 753 | return bitmap; |
73fa0d10 AW |
754 | } |
755 | ||
756 | static int vfio_dma_do_unmap(struct vfio_iommu *iommu, | |
757 | struct vfio_iommu_type1_dma_unmap *unmap) | |
758 | { | |
73fa0d10 | 759 | uint64_t mask; |
c086de81 | 760 | struct vfio_dma *dma, *dma_last = NULL; |
1ef3e2bc | 761 | size_t unmapped = 0; |
c086de81 | 762 | int ret = 0, retries = 0; |
73fa0d10 | 763 | |
1ef3e2bc | 764 | mask = ((uint64_t)1 << __ffs(vfio_pgsize_bitmap(iommu))) - 1; |
73fa0d10 AW |
765 | |
766 | if (unmap->iova & mask) | |
767 | return -EINVAL; | |
f5bfdbf2 | 768 | if (!unmap->size || unmap->size & mask) |
73fa0d10 AW |
769 | return -EINVAL; |
770 | ||
73fa0d10 | 771 | WARN_ON(mask & PAGE_MASK); |
c086de81 | 772 | again: |
73fa0d10 AW |
773 | mutex_lock(&iommu->lock); |
774 | ||
1ef3e2bc AW |
775 | /* |
776 | * vfio-iommu-type1 (v1) - User mappings were coalesced together to | |
777 | * avoid tracking individual mappings. This means that the granularity | |
778 | * of the original mapping was lost and the user was allowed to attempt | |
779 | * to unmap any range. Depending on the contiguousness of physical | |
780 | * memory and page sizes supported by the IOMMU, arbitrary unmaps may | |
781 | * or may not have worked. We only guaranteed unmap granularity | |
782 | * matching the original mapping; even though it was untracked here, | |
783 | * the original mappings are reflected in IOMMU mappings. This | |
784 | * resulted in a couple unusual behaviors. First, if a range is not | |
785 | * able to be unmapped, ex. a set of 4k pages that was mapped as a | |
786 | * 2M hugepage into the IOMMU, the unmap ioctl returns success but with | |
787 | * a zero sized unmap. Also, if an unmap request overlaps the first | |
788 | * address of a hugepage, the IOMMU will unmap the entire hugepage. | |
789 | * This also returns success and the returned unmap size reflects the | |
790 | * actual size unmapped. | |
791 | * | |
792 | * We attempt to maintain compatibility with this "v1" interface, but | |
793 | * we take control out of the hands of the IOMMU. Therefore, an unmap | |
794 | * request offset from the beginning of the original mapping will | |
795 | * return success with zero sized unmap. And an unmap request covering | |
796 | * the first iova of mapping will unmap the entire range. | |
797 | * | |
798 | * The v2 version of this interface intends to be more deterministic. | |
799 | * Unmap requests must fully cover previous mappings. Multiple | |
800 | * mappings may still be unmaped by specifying large ranges, but there | |
801 | * must not be any previous mappings bisected by the range. An error | |
802 | * will be returned if these conditions are not met. The v2 interface | |
803 | * will only return success and a size of zero if there were no | |
804 | * mappings within the range. | |
805 | */ | |
806 | if (iommu->v2) { | |
7c03f428 | 807 | dma = vfio_find_dma(iommu, unmap->iova, 1); |
1ef3e2bc AW |
808 | if (dma && dma->iova != unmap->iova) { |
809 | ret = -EINVAL; | |
810 | goto unlock; | |
811 | } | |
812 | dma = vfio_find_dma(iommu, unmap->iova + unmap->size - 1, 0); | |
813 | if (dma && dma->iova + dma->size != unmap->iova + unmap->size) { | |
814 | ret = -EINVAL; | |
815 | goto unlock; | |
816 | } | |
817 | } | |
818 | ||
166fd7d9 | 819 | while ((dma = vfio_find_dma(iommu, unmap->iova, unmap->size))) { |
1ef3e2bc | 820 | if (!iommu->v2 && unmap->iova > dma->iova) |
166fd7d9 | 821 | break; |
8f0d5bb9 KW |
822 | /* |
823 | * Task with same address space who mapped this iova range is | |
824 | * allowed to unmap the iova range. | |
825 | */ | |
826 | if (dma->task->mm != current->mm) | |
827 | break; | |
c086de81 KW |
828 | |
829 | if (!RB_EMPTY_ROOT(&dma->pfn_list)) { | |
830 | struct vfio_iommu_type1_dma_unmap nb_unmap; | |
831 | ||
832 | if (dma_last == dma) { | |
833 | BUG_ON(++retries > 10); | |
834 | } else { | |
835 | dma_last = dma; | |
836 | retries = 0; | |
837 | } | |
838 | ||
839 | nb_unmap.iova = dma->iova; | |
840 | nb_unmap.size = dma->size; | |
841 | ||
842 | /* | |
843 | * Notify anyone (mdev vendor drivers) to invalidate and | |
844 | * unmap iovas within the range we're about to unmap. | |
845 | * Vendor drivers MUST unpin pages in response to an | |
846 | * invalidation. | |
847 | */ | |
848 | mutex_unlock(&iommu->lock); | |
849 | blocking_notifier_call_chain(&iommu->notifier, | |
850 | VFIO_IOMMU_NOTIFY_DMA_UNMAP, | |
851 | &nb_unmap); | |
852 | goto again; | |
853 | } | |
1ef3e2bc AW |
854 | unmapped += dma->size; |
855 | vfio_remove_dma(iommu, dma); | |
166fd7d9 | 856 | } |
cd9b2268 | 857 | |
1ef3e2bc | 858 | unlock: |
73fa0d10 | 859 | mutex_unlock(&iommu->lock); |
166fd7d9 | 860 | |
1ef3e2bc | 861 | /* Report how much was unmapped */ |
166fd7d9 AW |
862 | unmap->size = unmapped; |
863 | ||
864 | return ret; | |
865 | } | |
866 | ||
867 | /* | |
868 | * Turns out AMD IOMMU has a page table bug where it won't map large pages | |
869 | * to a region that previously mapped smaller pages. This should be fixed | |
870 | * soon, so this is just a temporary workaround to break mappings down into | |
871 | * PAGE_SIZE. Better to map smaller pages than nothing. | |
872 | */ | |
1ef3e2bc | 873 | static int map_try_harder(struct vfio_domain *domain, dma_addr_t iova, |
166fd7d9 AW |
874 | unsigned long pfn, long npage, int prot) |
875 | { | |
876 | long i; | |
089f1c6b | 877 | int ret = 0; |
166fd7d9 AW |
878 | |
879 | for (i = 0; i < npage; i++, pfn++, iova += PAGE_SIZE) { | |
1ef3e2bc | 880 | ret = iommu_map(domain->domain, iova, |
166fd7d9 | 881 | (phys_addr_t)pfn << PAGE_SHIFT, |
1ef3e2bc | 882 | PAGE_SIZE, prot | domain->prot); |
166fd7d9 AW |
883 | if (ret) |
884 | break; | |
885 | } | |
886 | ||
887 | for (; i < npage && i > 0; i--, iova -= PAGE_SIZE) | |
1ef3e2bc AW |
888 | iommu_unmap(domain->domain, iova, PAGE_SIZE); |
889 | ||
890 | return ret; | |
891 | } | |
892 | ||
893 | static int vfio_iommu_map(struct vfio_iommu *iommu, dma_addr_t iova, | |
894 | unsigned long pfn, long npage, int prot) | |
895 | { | |
896 | struct vfio_domain *d; | |
897 | int ret; | |
898 | ||
899 | list_for_each_entry(d, &iommu->domain_list, next) { | |
900 | ret = iommu_map(d->domain, iova, (phys_addr_t)pfn << PAGE_SHIFT, | |
901 | npage << PAGE_SHIFT, prot | d->prot); | |
902 | if (ret) { | |
903 | if (ret != -EBUSY || | |
904 | map_try_harder(d, iova, pfn, npage, prot)) | |
905 | goto unwind; | |
906 | } | |
c5e66887 AW |
907 | |
908 | cond_resched(); | |
1ef3e2bc AW |
909 | } |
910 | ||
911 | return 0; | |
912 | ||
913 | unwind: | |
914 | list_for_each_entry_continue_reverse(d, &iommu->domain_list, next) | |
915 | iommu_unmap(d->domain, iova, npage << PAGE_SHIFT); | |
166fd7d9 | 916 | |
cd9b2268 | 917 | return ret; |
73fa0d10 AW |
918 | } |
919 | ||
8f0d5bb9 KW |
920 | static int vfio_pin_map_dma(struct vfio_iommu *iommu, struct vfio_dma *dma, |
921 | size_t map_size) | |
922 | { | |
923 | dma_addr_t iova = dma->iova; | |
924 | unsigned long vaddr = dma->vaddr; | |
925 | size_t size = map_size; | |
926 | long npage; | |
7cb671e7 AW |
927 | unsigned long pfn, limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT; |
928 | bool lock_cap = capable(CAP_IPC_LOCK); | |
8f0d5bb9 KW |
929 | int ret = 0; |
930 | ||
931 | while (size) { | |
932 | /* Pin a contiguous chunk of memory */ | |
933 | npage = vfio_pin_pages_remote(dma, vaddr + dma->size, | |
7cb671e7 AW |
934 | size >> PAGE_SHIFT, &pfn, |
935 | lock_cap, limit); | |
8f0d5bb9 KW |
936 | if (npage <= 0) { |
937 | WARN_ON(!npage); | |
938 | ret = (int)npage; | |
939 | break; | |
940 | } | |
941 | ||
942 | /* Map it! */ | |
943 | ret = vfio_iommu_map(iommu, iova + dma->size, pfn, npage, | |
944 | dma->prot); | |
945 | if (ret) { | |
a54eb550 KW |
946 | vfio_unpin_pages_remote(dma, iova + dma->size, pfn, |
947 | npage, true); | |
8f0d5bb9 KW |
948 | break; |
949 | } | |
950 | ||
951 | size -= npage << PAGE_SHIFT; | |
952 | dma->size += npage << PAGE_SHIFT; | |
953 | } | |
954 | ||
a54eb550 KW |
955 | dma->iommu_mapped = true; |
956 | ||
8f0d5bb9 KW |
957 | if (ret) |
958 | vfio_remove_dma(iommu, dma); | |
959 | ||
960 | return ret; | |
961 | } | |
962 | ||
73fa0d10 AW |
963 | static int vfio_dma_do_map(struct vfio_iommu *iommu, |
964 | struct vfio_iommu_type1_dma_map *map) | |
965 | { | |
c8dbca16 | 966 | dma_addr_t iova = map->iova; |
166fd7d9 | 967 | unsigned long vaddr = map->vaddr; |
73fa0d10 AW |
968 | size_t size = map->size; |
969 | int ret = 0, prot = 0; | |
970 | uint64_t mask; | |
1ef3e2bc | 971 | struct vfio_dma *dma; |
166fd7d9 | 972 | |
c8dbca16 AW |
973 | /* Verify that none of our __u64 fields overflow */ |
974 | if (map->size != size || map->vaddr != vaddr || map->iova != iova) | |
975 | return -EINVAL; | |
73fa0d10 | 976 | |
1ef3e2bc | 977 | mask = ((uint64_t)1 << __ffs(vfio_pgsize_bitmap(iommu))) - 1; |
73fa0d10 | 978 | |
c8dbca16 AW |
979 | WARN_ON(mask & PAGE_MASK); |
980 | ||
73fa0d10 AW |
981 | /* READ/WRITE from device perspective */ |
982 | if (map->flags & VFIO_DMA_MAP_FLAG_WRITE) | |
983 | prot |= IOMMU_WRITE; | |
984 | if (map->flags & VFIO_DMA_MAP_FLAG_READ) | |
985 | prot |= IOMMU_READ; | |
986 | ||
c8dbca16 | 987 | if (!prot || !size || (size | iova | vaddr) & mask) |
73fa0d10 AW |
988 | return -EINVAL; |
989 | ||
c8dbca16 AW |
990 | /* Don't allow IOVA or virtual address wrap */ |
991 | if (iova + size - 1 < iova || vaddr + size - 1 < vaddr) | |
73fa0d10 AW |
992 | return -EINVAL; |
993 | ||
994 | mutex_lock(&iommu->lock); | |
995 | ||
c8dbca16 | 996 | if (vfio_find_dma(iommu, iova, size)) { |
8f0d5bb9 KW |
997 | ret = -EEXIST; |
998 | goto out_unlock; | |
73fa0d10 AW |
999 | } |
1000 | ||
1ef3e2bc AW |
1001 | dma = kzalloc(sizeof(*dma), GFP_KERNEL); |
1002 | if (!dma) { | |
8f0d5bb9 KW |
1003 | ret = -ENOMEM; |
1004 | goto out_unlock; | |
1ef3e2bc AW |
1005 | } |
1006 | ||
c8dbca16 AW |
1007 | dma->iova = iova; |
1008 | dma->vaddr = vaddr; | |
1ef3e2bc | 1009 | dma->prot = prot; |
8f0d5bb9 KW |
1010 | get_task_struct(current); |
1011 | dma->task = current; | |
a54eb550 | 1012 | dma->pfn_list = RB_ROOT; |
166fd7d9 | 1013 | |
1ef3e2bc AW |
1014 | /* Insert zero-sized and grow as we map chunks of it */ |
1015 | vfio_link_dma(iommu, dma); | |
166fd7d9 | 1016 | |
a54eb550 KW |
1017 | /* Don't pin and map if container doesn't contain IOMMU capable domain*/ |
1018 | if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu)) | |
1019 | dma->size = size; | |
1020 | else | |
1021 | ret = vfio_pin_map_dma(iommu, dma, size); | |
1022 | ||
8f0d5bb9 | 1023 | out_unlock: |
1ef3e2bc AW |
1024 | mutex_unlock(&iommu->lock); |
1025 | return ret; | |
1026 | } | |
1027 | ||
1028 | static int vfio_bus_type(struct device *dev, void *data) | |
1029 | { | |
1030 | struct bus_type **bus = data; | |
1031 | ||
1032 | if (*bus && *bus != dev->bus) | |
1033 | return -EINVAL; | |
1034 | ||
1035 | *bus = dev->bus; | |
1036 | ||
1037 | return 0; | |
1038 | } | |
1039 | ||
1040 | static int vfio_iommu_replay(struct vfio_iommu *iommu, | |
1041 | struct vfio_domain *domain) | |
1042 | { | |
1043 | struct vfio_domain *d; | |
1044 | struct rb_node *n; | |
7cb671e7 AW |
1045 | unsigned long limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT; |
1046 | bool lock_cap = capable(CAP_IPC_LOCK); | |
1ef3e2bc AW |
1047 | int ret; |
1048 | ||
1049 | /* Arbitrarily pick the first domain in the list for lookups */ | |
1050 | d = list_first_entry(&iommu->domain_list, struct vfio_domain, next); | |
1051 | n = rb_first(&iommu->dma_list); | |
1052 | ||
1ef3e2bc AW |
1053 | for (; n; n = rb_next(n)) { |
1054 | struct vfio_dma *dma; | |
1055 | dma_addr_t iova; | |
1056 | ||
1057 | dma = rb_entry(n, struct vfio_dma, node); | |
1058 | iova = dma->iova; | |
1059 | ||
1060 | while (iova < dma->iova + dma->size) { | |
a54eb550 | 1061 | phys_addr_t phys; |
1ef3e2bc | 1062 | size_t size; |
73fa0d10 | 1063 | |
a54eb550 KW |
1064 | if (dma->iommu_mapped) { |
1065 | phys_addr_t p; | |
1066 | dma_addr_t i; | |
1067 | ||
1068 | phys = iommu_iova_to_phys(d->domain, iova); | |
1069 | ||
1070 | if (WARN_ON(!phys)) { | |
1071 | iova += PAGE_SIZE; | |
1072 | continue; | |
1073 | } | |
1074 | ||
1075 | size = PAGE_SIZE; | |
1076 | p = phys + size; | |
1077 | i = iova + size; | |
1078 | while (i < dma->iova + dma->size && | |
1079 | p == iommu_iova_to_phys(d->domain, i)) { | |
1080 | size += PAGE_SIZE; | |
1081 | p += PAGE_SIZE; | |
1082 | i += PAGE_SIZE; | |
1083 | } | |
1084 | } else { | |
1085 | unsigned long pfn; | |
1086 | unsigned long vaddr = dma->vaddr + | |
1087 | (iova - dma->iova); | |
1088 | size_t n = dma->iova + dma->size - iova; | |
1089 | long npage; | |
1090 | ||
1091 | npage = vfio_pin_pages_remote(dma, vaddr, | |
1092 | n >> PAGE_SHIFT, | |
7cb671e7 AW |
1093 | &pfn, lock_cap, |
1094 | limit); | |
a54eb550 KW |
1095 | if (npage <= 0) { |
1096 | WARN_ON(!npage); | |
1097 | ret = (int)npage; | |
1098 | return ret; | |
1099 | } | |
1100 | ||
1101 | phys = pfn << PAGE_SHIFT; | |
1102 | size = npage << PAGE_SHIFT; | |
166fd7d9 AW |
1103 | } |
1104 | ||
1ef3e2bc AW |
1105 | ret = iommu_map(domain->domain, iova, phys, |
1106 | size, dma->prot | domain->prot); | |
1107 | if (ret) | |
1108 | return ret; | |
d93b3ac0 | 1109 | |
1ef3e2bc AW |
1110 | iova += size; |
1111 | } | |
a54eb550 | 1112 | dma->iommu_mapped = true; |
166fd7d9 | 1113 | } |
1ef3e2bc | 1114 | return 0; |
73fa0d10 AW |
1115 | } |
1116 | ||
6fe1010d AW |
1117 | /* |
1118 | * We change our unmap behavior slightly depending on whether the IOMMU | |
1119 | * supports fine-grained superpages. IOMMUs like AMD-Vi will use a superpage | |
1120 | * for practically any contiguous power-of-two mapping we give it. This means | |
1121 | * we don't need to look for contiguous chunks ourselves to make unmapping | |
1122 | * more efficient. On IOMMUs with coarse-grained super pages, like Intel VT-d | |
1123 | * with discrete 2M/1G/512G/1T superpages, identifying contiguous chunks | |
1124 | * significantly boosts non-hugetlbfs mappings and doesn't seem to hurt when | |
1125 | * hugetlbfs is in use. | |
1126 | */ | |
1127 | static void vfio_test_domain_fgsp(struct vfio_domain *domain) | |
1128 | { | |
1129 | struct page *pages; | |
1130 | int ret, order = get_order(PAGE_SIZE * 2); | |
1131 | ||
1132 | pages = alloc_pages(GFP_KERNEL | __GFP_ZERO, order); | |
1133 | if (!pages) | |
1134 | return; | |
1135 | ||
1136 | ret = iommu_map(domain->domain, 0, page_to_phys(pages), PAGE_SIZE * 2, | |
1137 | IOMMU_READ | IOMMU_WRITE | domain->prot); | |
1138 | if (!ret) { | |
1139 | size_t unmapped = iommu_unmap(domain->domain, 0, PAGE_SIZE); | |
1140 | ||
1141 | if (unmapped == PAGE_SIZE) | |
1142 | iommu_unmap(domain->domain, PAGE_SIZE, PAGE_SIZE); | |
1143 | else | |
1144 | domain->fgsp = true; | |
1145 | } | |
1146 | ||
1147 | __free_pages(pages, order); | |
1148 | } | |
1149 | ||
7896c998 KW |
1150 | static struct vfio_group *find_iommu_group(struct vfio_domain *domain, |
1151 | struct iommu_group *iommu_group) | |
1152 | { | |
1153 | struct vfio_group *g; | |
1154 | ||
1155 | list_for_each_entry(g, &domain->group_list, next) { | |
1156 | if (g->iommu_group == iommu_group) | |
1157 | return g; | |
1158 | } | |
1159 | ||
1160 | return NULL; | |
1161 | } | |
1162 | ||
9d3a4de4 | 1163 | static bool vfio_iommu_has_sw_msi(struct iommu_group *group, phys_addr_t *base) |
5d704992 EA |
1164 | { |
1165 | struct list_head group_resv_regions; | |
1166 | struct iommu_resv_region *region, *next; | |
1167 | bool ret = false; | |
1168 | ||
1169 | INIT_LIST_HEAD(&group_resv_regions); | |
1170 | iommu_get_group_resv_regions(group, &group_resv_regions); | |
1171 | list_for_each_entry(region, &group_resv_regions, list) { | |
9d3a4de4 | 1172 | if (region->type == IOMMU_RESV_SW_MSI) { |
5d704992 EA |
1173 | *base = region->start; |
1174 | ret = true; | |
1175 | goto out; | |
1176 | } | |
1177 | } | |
1178 | out: | |
1179 | list_for_each_entry_safe(region, next, &group_resv_regions, list) | |
1180 | kfree(region); | |
1181 | return ret; | |
1182 | } | |
1183 | ||
73fa0d10 AW |
1184 | static int vfio_iommu_type1_attach_group(void *iommu_data, |
1185 | struct iommu_group *iommu_group) | |
1186 | { | |
1187 | struct vfio_iommu *iommu = iommu_data; | |
7896c998 | 1188 | struct vfio_group *group; |
1ef3e2bc | 1189 | struct vfio_domain *domain, *d; |
a54eb550 | 1190 | struct bus_type *bus = NULL, *mdev_bus; |
73fa0d10 | 1191 | int ret; |
9d72f87b | 1192 | bool resv_msi, msi_remap; |
5d704992 | 1193 | phys_addr_t resv_msi_base; |
73fa0d10 | 1194 | |
73fa0d10 AW |
1195 | mutex_lock(&iommu->lock); |
1196 | ||
1ef3e2bc | 1197 | list_for_each_entry(d, &iommu->domain_list, next) { |
7896c998 | 1198 | if (find_iommu_group(d, iommu_group)) { |
73fa0d10 | 1199 | mutex_unlock(&iommu->lock); |
73fa0d10 AW |
1200 | return -EINVAL; |
1201 | } | |
1202 | } | |
1203 | ||
a54eb550 KW |
1204 | if (iommu->external_domain) { |
1205 | if (find_iommu_group(iommu->external_domain, iommu_group)) { | |
1206 | mutex_unlock(&iommu->lock); | |
1207 | return -EINVAL; | |
1208 | } | |
1209 | } | |
1210 | ||
1ef3e2bc AW |
1211 | group = kzalloc(sizeof(*group), GFP_KERNEL); |
1212 | domain = kzalloc(sizeof(*domain), GFP_KERNEL); | |
1213 | if (!group || !domain) { | |
1214 | ret = -ENOMEM; | |
1215 | goto out_free; | |
1216 | } | |
1217 | ||
1218 | group->iommu_group = iommu_group; | |
1219 | ||
1220 | /* Determine bus_type in order to allocate a domain */ | |
1221 | ret = iommu_group_for_each_dev(iommu_group, &bus, vfio_bus_type); | |
1222 | if (ret) | |
1223 | goto out_free; | |
1224 | ||
a54eb550 KW |
1225 | mdev_bus = symbol_get(mdev_bus_type); |
1226 | ||
1227 | if (mdev_bus) { | |
1228 | if ((bus == mdev_bus) && !iommu_present(bus)) { | |
1229 | symbol_put(mdev_bus_type); | |
1230 | if (!iommu->external_domain) { | |
1231 | INIT_LIST_HEAD(&domain->group_list); | |
1232 | iommu->external_domain = domain; | |
1233 | } else | |
1234 | kfree(domain); | |
1235 | ||
1236 | list_add(&group->next, | |
1237 | &iommu->external_domain->group_list); | |
1238 | mutex_unlock(&iommu->lock); | |
1239 | return 0; | |
1240 | } | |
1241 | symbol_put(mdev_bus_type); | |
1242 | } | |
1243 | ||
1ef3e2bc AW |
1244 | domain->domain = iommu_domain_alloc(bus); |
1245 | if (!domain->domain) { | |
1246 | ret = -EIO; | |
1247 | goto out_free; | |
1248 | } | |
1249 | ||
f5c9eceb WD |
1250 | if (iommu->nesting) { |
1251 | int attr = 1; | |
1252 | ||
1253 | ret = iommu_domain_set_attr(domain->domain, DOMAIN_ATTR_NESTING, | |
1254 | &attr); | |
1255 | if (ret) | |
1256 | goto out_domain; | |
1257 | } | |
1258 | ||
1ef3e2bc AW |
1259 | ret = iommu_attach_group(domain->domain, iommu_group); |
1260 | if (ret) | |
1261 | goto out_domain; | |
1262 | ||
9d3a4de4 | 1263 | resv_msi = vfio_iommu_has_sw_msi(iommu_group, &resv_msi_base); |
5d704992 | 1264 | |
1ef3e2bc AW |
1265 | INIT_LIST_HEAD(&domain->group_list); |
1266 | list_add(&group->next, &domain->group_list); | |
1267 | ||
9d72f87b EA |
1268 | msi_remap = resv_msi ? irq_domain_check_msi_remap() : |
1269 | iommu_capable(bus, IOMMU_CAP_INTR_REMAP); | |
1270 | ||
1271 | if (!allow_unsafe_interrupts && !msi_remap) { | |
1ef3e2bc AW |
1272 | pr_warn("%s: No interrupt remapping support. Use the module param \"allow_unsafe_interrupts\" to enable VFIO IOMMU support on this platform\n", |
1273 | __func__); | |
1274 | ret = -EPERM; | |
1275 | goto out_detach; | |
1276 | } | |
1277 | ||
eb165f05 | 1278 | if (iommu_capable(bus, IOMMU_CAP_CACHE_COHERENCY)) |
1ef3e2bc AW |
1279 | domain->prot |= IOMMU_CACHE; |
1280 | ||
73fa0d10 | 1281 | /* |
1ef3e2bc AW |
1282 | * Try to match an existing compatible domain. We don't want to |
1283 | * preclude an IOMMU driver supporting multiple bus_types and being | |
1284 | * able to include different bus_types in the same IOMMU domain, so | |
1285 | * we test whether the domains use the same iommu_ops rather than | |
1286 | * testing if they're on the same bus_type. | |
73fa0d10 | 1287 | */ |
1ef3e2bc AW |
1288 | list_for_each_entry(d, &iommu->domain_list, next) { |
1289 | if (d->domain->ops == domain->domain->ops && | |
1290 | d->prot == domain->prot) { | |
1291 | iommu_detach_group(domain->domain, iommu_group); | |
1292 | if (!iommu_attach_group(d->domain, iommu_group)) { | |
1293 | list_add(&group->next, &d->group_list); | |
1294 | iommu_domain_free(domain->domain); | |
1295 | kfree(domain); | |
1296 | mutex_unlock(&iommu->lock); | |
1297 | return 0; | |
1298 | } | |
1299 | ||
1300 | ret = iommu_attach_group(domain->domain, iommu_group); | |
1301 | if (ret) | |
1302 | goto out_domain; | |
1303 | } | |
73fa0d10 AW |
1304 | } |
1305 | ||
6fe1010d AW |
1306 | vfio_test_domain_fgsp(domain); |
1307 | ||
1ef3e2bc AW |
1308 | /* replay mappings on new domains */ |
1309 | ret = vfio_iommu_replay(iommu, domain); | |
1310 | if (ret) | |
1311 | goto out_detach; | |
1312 | ||
2c9f1af5 WY |
1313 | if (resv_msi) { |
1314 | ret = iommu_get_msi_cookie(domain->domain, resv_msi_base); | |
1315 | if (ret) | |
1316 | goto out_detach; | |
1317 | } | |
5d704992 | 1318 | |
1ef3e2bc | 1319 | list_add(&domain->next, &iommu->domain_list); |
73fa0d10 AW |
1320 | |
1321 | mutex_unlock(&iommu->lock); | |
1322 | ||
1323 | return 0; | |
1ef3e2bc AW |
1324 | |
1325 | out_detach: | |
1326 | iommu_detach_group(domain->domain, iommu_group); | |
1327 | out_domain: | |
1328 | iommu_domain_free(domain->domain); | |
1329 | out_free: | |
1330 | kfree(domain); | |
1331 | kfree(group); | |
1332 | mutex_unlock(&iommu->lock); | |
1333 | return ret; | |
1334 | } | |
1335 | ||
1336 | static void vfio_iommu_unmap_unpin_all(struct vfio_iommu *iommu) | |
1337 | { | |
1338 | struct rb_node *node; | |
1339 | ||
1340 | while ((node = rb_first(&iommu->dma_list))) | |
1341 | vfio_remove_dma(iommu, rb_entry(node, struct vfio_dma, node)); | |
73fa0d10 AW |
1342 | } |
1343 | ||
a54eb550 KW |
1344 | static void vfio_iommu_unmap_unpin_reaccount(struct vfio_iommu *iommu) |
1345 | { | |
1346 | struct rb_node *n, *p; | |
1347 | ||
1348 | n = rb_first(&iommu->dma_list); | |
1349 | for (; n; n = rb_next(n)) { | |
1350 | struct vfio_dma *dma; | |
1351 | long locked = 0, unlocked = 0; | |
1352 | ||
1353 | dma = rb_entry(n, struct vfio_dma, node); | |
1354 | unlocked += vfio_unmap_unpin(iommu, dma, false); | |
1355 | p = rb_first(&dma->pfn_list); | |
1356 | for (; p; p = rb_next(p)) { | |
1357 | struct vfio_pfn *vpfn = rb_entry(p, struct vfio_pfn, | |
1358 | node); | |
1359 | ||
1360 | if (!is_invalid_reserved_pfn(vpfn->pfn)) | |
1361 | locked++; | |
1362 | } | |
0cfef2b7 | 1363 | vfio_lock_acct(dma->task, locked - unlocked, NULL); |
a54eb550 KW |
1364 | } |
1365 | } | |
1366 | ||
1367 | static void vfio_sanity_check_pfn_list(struct vfio_iommu *iommu) | |
1368 | { | |
1369 | struct rb_node *n; | |
1370 | ||
1371 | n = rb_first(&iommu->dma_list); | |
1372 | for (; n; n = rb_next(n)) { | |
1373 | struct vfio_dma *dma; | |
1374 | ||
1375 | dma = rb_entry(n, struct vfio_dma, node); | |
1376 | ||
1377 | if (WARN_ON(!RB_EMPTY_ROOT(&dma->pfn_list))) | |
1378 | break; | |
1379 | } | |
3cedd7d7 KW |
1380 | /* mdev vendor driver must unregister notifier */ |
1381 | WARN_ON(iommu->notifier.head); | |
a54eb550 KW |
1382 | } |
1383 | ||
73fa0d10 AW |
1384 | static void vfio_iommu_type1_detach_group(void *iommu_data, |
1385 | struct iommu_group *iommu_group) | |
1386 | { | |
1387 | struct vfio_iommu *iommu = iommu_data; | |
1ef3e2bc | 1388 | struct vfio_domain *domain; |
73fa0d10 AW |
1389 | struct vfio_group *group; |
1390 | ||
1391 | mutex_lock(&iommu->lock); | |
1392 | ||
a54eb550 KW |
1393 | if (iommu->external_domain) { |
1394 | group = find_iommu_group(iommu->external_domain, iommu_group); | |
1395 | if (group) { | |
1396 | list_del(&group->next); | |
1397 | kfree(group); | |
1398 | ||
1399 | if (list_empty(&iommu->external_domain->group_list)) { | |
1400 | vfio_sanity_check_pfn_list(iommu); | |
1401 | ||
1402 | if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu)) | |
1403 | vfio_iommu_unmap_unpin_all(iommu); | |
1404 | ||
1405 | kfree(iommu->external_domain); | |
1406 | iommu->external_domain = NULL; | |
1407 | } | |
1408 | goto detach_group_done; | |
1409 | } | |
1410 | } | |
1411 | ||
1ef3e2bc | 1412 | list_for_each_entry(domain, &iommu->domain_list, next) { |
7896c998 KW |
1413 | group = find_iommu_group(domain, iommu_group); |
1414 | if (!group) | |
1415 | continue; | |
1ef3e2bc | 1416 | |
7896c998 KW |
1417 | iommu_detach_group(domain->domain, iommu_group); |
1418 | list_del(&group->next); | |
1419 | kfree(group); | |
1420 | /* | |
a54eb550 KW |
1421 | * Group ownership provides privilege, if the group list is |
1422 | * empty, the domain goes away. If it's the last domain with | |
1423 | * iommu and external domain doesn't exist, then all the | |
1424 | * mappings go away too. If it's the last domain with iommu and | |
1425 | * external domain exist, update accounting | |
7896c998 KW |
1426 | */ |
1427 | if (list_empty(&domain->group_list)) { | |
a54eb550 KW |
1428 | if (list_is_singular(&iommu->domain_list)) { |
1429 | if (!iommu->external_domain) | |
1430 | vfio_iommu_unmap_unpin_all(iommu); | |
1431 | else | |
1432 | vfio_iommu_unmap_unpin_reaccount(iommu); | |
1433 | } | |
7896c998 KW |
1434 | iommu_domain_free(domain->domain); |
1435 | list_del(&domain->next); | |
1436 | kfree(domain); | |
73fa0d10 | 1437 | } |
a54eb550 | 1438 | break; |
73fa0d10 AW |
1439 | } |
1440 | ||
a54eb550 | 1441 | detach_group_done: |
73fa0d10 AW |
1442 | mutex_unlock(&iommu->lock); |
1443 | } | |
1444 | ||
1445 | static void *vfio_iommu_type1_open(unsigned long arg) | |
1446 | { | |
1447 | struct vfio_iommu *iommu; | |
1448 | ||
73fa0d10 AW |
1449 | iommu = kzalloc(sizeof(*iommu), GFP_KERNEL); |
1450 | if (!iommu) | |
1451 | return ERR_PTR(-ENOMEM); | |
1452 | ||
f5c9eceb WD |
1453 | switch (arg) { |
1454 | case VFIO_TYPE1_IOMMU: | |
1455 | break; | |
1456 | case VFIO_TYPE1_NESTING_IOMMU: | |
1457 | iommu->nesting = true; | |
1458 | case VFIO_TYPE1v2_IOMMU: | |
1459 | iommu->v2 = true; | |
1460 | break; | |
1461 | default: | |
1462 | kfree(iommu); | |
1463 | return ERR_PTR(-EINVAL); | |
1464 | } | |
1465 | ||
1ef3e2bc | 1466 | INIT_LIST_HEAD(&iommu->domain_list); |
cd9b2268 | 1467 | iommu->dma_list = RB_ROOT; |
73fa0d10 | 1468 | mutex_init(&iommu->lock); |
c086de81 | 1469 | BLOCKING_INIT_NOTIFIER_HEAD(&iommu->notifier); |
73fa0d10 AW |
1470 | |
1471 | return iommu; | |
1472 | } | |
1473 | ||
a54eb550 KW |
1474 | static void vfio_release_domain(struct vfio_domain *domain, bool external) |
1475 | { | |
1476 | struct vfio_group *group, *group_tmp; | |
1477 | ||
1478 | list_for_each_entry_safe(group, group_tmp, | |
1479 | &domain->group_list, next) { | |
1480 | if (!external) | |
1481 | iommu_detach_group(domain->domain, group->iommu_group); | |
1482 | list_del(&group->next); | |
1483 | kfree(group); | |
1484 | } | |
1485 | ||
1486 | if (!external) | |
1487 | iommu_domain_free(domain->domain); | |
1488 | } | |
1489 | ||
73fa0d10 AW |
1490 | static void vfio_iommu_type1_release(void *iommu_data) |
1491 | { | |
1492 | struct vfio_iommu *iommu = iommu_data; | |
1ef3e2bc | 1493 | struct vfio_domain *domain, *domain_tmp; |
a54eb550 KW |
1494 | |
1495 | if (iommu->external_domain) { | |
1496 | vfio_release_domain(iommu->external_domain, true); | |
1497 | vfio_sanity_check_pfn_list(iommu); | |
1498 | kfree(iommu->external_domain); | |
1499 | } | |
73fa0d10 | 1500 | |
1ef3e2bc | 1501 | vfio_iommu_unmap_unpin_all(iommu); |
73fa0d10 | 1502 | |
1ef3e2bc AW |
1503 | list_for_each_entry_safe(domain, domain_tmp, |
1504 | &iommu->domain_list, next) { | |
a54eb550 | 1505 | vfio_release_domain(domain, false); |
1ef3e2bc AW |
1506 | list_del(&domain->next); |
1507 | kfree(domain); | |
73fa0d10 | 1508 | } |
73fa0d10 AW |
1509 | kfree(iommu); |
1510 | } | |
1511 | ||
aa429318 AW |
1512 | static int vfio_domains_have_iommu_cache(struct vfio_iommu *iommu) |
1513 | { | |
1514 | struct vfio_domain *domain; | |
1515 | int ret = 1; | |
1516 | ||
1517 | mutex_lock(&iommu->lock); | |
1518 | list_for_each_entry(domain, &iommu->domain_list, next) { | |
1519 | if (!(domain->prot & IOMMU_CACHE)) { | |
1520 | ret = 0; | |
f5bfdbf2 | 1521 | break; |
aa429318 | 1522 | } |
73fa0d10 | 1523 | } |
aa429318 | 1524 | mutex_unlock(&iommu->lock); |
73fa0d10 | 1525 | |
aa429318 | 1526 | return ret; |
73fa0d10 AW |
1527 | } |
1528 | ||
1529 | static long vfio_iommu_type1_ioctl(void *iommu_data, | |
1530 | unsigned int cmd, unsigned long arg) | |
1531 | { | |
1532 | struct vfio_iommu *iommu = iommu_data; | |
1533 | unsigned long minsz; | |
1534 | ||
1535 | if (cmd == VFIO_CHECK_EXTENSION) { | |
1536 | switch (arg) { | |
1537 | case VFIO_TYPE1_IOMMU: | |
1ef3e2bc | 1538 | case VFIO_TYPE1v2_IOMMU: |
f5c9eceb | 1539 | case VFIO_TYPE1_NESTING_IOMMU: |
73fa0d10 | 1540 | return 1; |
aa429318 AW |
1541 | case VFIO_DMA_CC_IOMMU: |
1542 | if (!iommu) | |
1543 | return 0; | |
1544 | return vfio_domains_have_iommu_cache(iommu); | |
73fa0d10 AW |
1545 | default: |
1546 | return 0; | |
1547 | } | |
1548 | } else if (cmd == VFIO_IOMMU_GET_INFO) { | |
1549 | struct vfio_iommu_type1_info info; | |
1550 | ||
1551 | minsz = offsetofend(struct vfio_iommu_type1_info, iova_pgsizes); | |
1552 | ||
1553 | if (copy_from_user(&info, (void __user *)arg, minsz)) | |
1554 | return -EFAULT; | |
1555 | ||
1556 | if (info.argsz < minsz) | |
1557 | return -EINVAL; | |
1558 | ||
d4f50ee2 | 1559 | info.flags = VFIO_IOMMU_INFO_PGSIZES; |
73fa0d10 | 1560 | |
1ef3e2bc | 1561 | info.iova_pgsizes = vfio_pgsize_bitmap(iommu); |
73fa0d10 | 1562 | |
8160c4e4 MT |
1563 | return copy_to_user((void __user *)arg, &info, minsz) ? |
1564 | -EFAULT : 0; | |
73fa0d10 AW |
1565 | |
1566 | } else if (cmd == VFIO_IOMMU_MAP_DMA) { | |
1567 | struct vfio_iommu_type1_dma_map map; | |
1568 | uint32_t mask = VFIO_DMA_MAP_FLAG_READ | | |
1569 | VFIO_DMA_MAP_FLAG_WRITE; | |
1570 | ||
1571 | minsz = offsetofend(struct vfio_iommu_type1_dma_map, size); | |
1572 | ||
1573 | if (copy_from_user(&map, (void __user *)arg, minsz)) | |
1574 | return -EFAULT; | |
1575 | ||
1576 | if (map.argsz < minsz || map.flags & ~mask) | |
1577 | return -EINVAL; | |
1578 | ||
1579 | return vfio_dma_do_map(iommu, &map); | |
1580 | ||
1581 | } else if (cmd == VFIO_IOMMU_UNMAP_DMA) { | |
1582 | struct vfio_iommu_type1_dma_unmap unmap; | |
166fd7d9 | 1583 | long ret; |
73fa0d10 AW |
1584 | |
1585 | minsz = offsetofend(struct vfio_iommu_type1_dma_unmap, size); | |
1586 | ||
1587 | if (copy_from_user(&unmap, (void __user *)arg, minsz)) | |
1588 | return -EFAULT; | |
1589 | ||
1590 | if (unmap.argsz < minsz || unmap.flags) | |
1591 | return -EINVAL; | |
1592 | ||
166fd7d9 AW |
1593 | ret = vfio_dma_do_unmap(iommu, &unmap); |
1594 | if (ret) | |
1595 | return ret; | |
1596 | ||
8160c4e4 MT |
1597 | return copy_to_user((void __user *)arg, &unmap, minsz) ? |
1598 | -EFAULT : 0; | |
73fa0d10 AW |
1599 | } |
1600 | ||
1601 | return -ENOTTY; | |
1602 | } | |
1603 | ||
c086de81 | 1604 | static int vfio_iommu_type1_register_notifier(void *iommu_data, |
22195cbd | 1605 | unsigned long *events, |
c086de81 KW |
1606 | struct notifier_block *nb) |
1607 | { | |
1608 | struct vfio_iommu *iommu = iommu_data; | |
1609 | ||
22195cbd JS |
1610 | /* clear known events */ |
1611 | *events &= ~VFIO_IOMMU_NOTIFY_DMA_UNMAP; | |
1612 | ||
1613 | /* refuse to register if still events remaining */ | |
1614 | if (*events) | |
1615 | return -EINVAL; | |
1616 | ||
c086de81 KW |
1617 | return blocking_notifier_chain_register(&iommu->notifier, nb); |
1618 | } | |
1619 | ||
1620 | static int vfio_iommu_type1_unregister_notifier(void *iommu_data, | |
1621 | struct notifier_block *nb) | |
1622 | { | |
1623 | struct vfio_iommu *iommu = iommu_data; | |
1624 | ||
1625 | return blocking_notifier_chain_unregister(&iommu->notifier, nb); | |
1626 | } | |
1627 | ||
73fa0d10 | 1628 | static const struct vfio_iommu_driver_ops vfio_iommu_driver_ops_type1 = { |
c086de81 KW |
1629 | .name = "vfio-iommu-type1", |
1630 | .owner = THIS_MODULE, | |
1631 | .open = vfio_iommu_type1_open, | |
1632 | .release = vfio_iommu_type1_release, | |
1633 | .ioctl = vfio_iommu_type1_ioctl, | |
1634 | .attach_group = vfio_iommu_type1_attach_group, | |
1635 | .detach_group = vfio_iommu_type1_detach_group, | |
1636 | .pin_pages = vfio_iommu_type1_pin_pages, | |
1637 | .unpin_pages = vfio_iommu_type1_unpin_pages, | |
1638 | .register_notifier = vfio_iommu_type1_register_notifier, | |
1639 | .unregister_notifier = vfio_iommu_type1_unregister_notifier, | |
73fa0d10 AW |
1640 | }; |
1641 | ||
1642 | static int __init vfio_iommu_type1_init(void) | |
1643 | { | |
73fa0d10 AW |
1644 | return vfio_register_iommu_driver(&vfio_iommu_driver_ops_type1); |
1645 | } | |
1646 | ||
1647 | static void __exit vfio_iommu_type1_cleanup(void) | |
1648 | { | |
1649 | vfio_unregister_iommu_driver(&vfio_iommu_driver_ops_type1); | |
1650 | } | |
1651 | ||
1652 | module_init(vfio_iommu_type1_init); | |
1653 | module_exit(vfio_iommu_type1_cleanup); | |
1654 | ||
1655 | MODULE_VERSION(DRIVER_VERSION); | |
1656 | MODULE_LICENSE("GPL v2"); | |
1657 | MODULE_AUTHOR(DRIVER_AUTHOR); | |
1658 | MODULE_DESCRIPTION(DRIVER_DESC); |