]>
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 | ||
a5354ea8 AW |
61 | static unsigned int dma_entry_limit __read_mostly = U16_MAX; |
62 | module_param_named(dma_entry_limit, dma_entry_limit, uint, 0644); | |
63 | MODULE_PARM_DESC(dma_entry_limit, | |
64 | "Maximum number of user DMA mappings per container (65535)."); | |
65 | ||
73fa0d10 | 66 | struct vfio_iommu { |
1ef3e2bc | 67 | struct list_head domain_list; |
a54eb550 | 68 | struct vfio_domain *external_domain; /* domain for external user */ |
73fa0d10 | 69 | struct mutex lock; |
cd9b2268 | 70 | struct rb_root dma_list; |
c086de81 | 71 | struct blocking_notifier_head notifier; |
a5354ea8 | 72 | unsigned int dma_avail; |
f5c9eceb WD |
73 | bool v2; |
74 | bool nesting; | |
1ef3e2bc AW |
75 | }; |
76 | ||
77 | struct vfio_domain { | |
78 | struct iommu_domain *domain; | |
79 | struct list_head next; | |
73fa0d10 | 80 | struct list_head group_list; |
1ef3e2bc | 81 | int prot; /* IOMMU_CACHE */ |
6fe1010d | 82 | bool fgsp; /* Fine-grained super pages */ |
73fa0d10 AW |
83 | }; |
84 | ||
85 | struct vfio_dma { | |
cd9b2268 | 86 | struct rb_node node; |
73fa0d10 AW |
87 | dma_addr_t iova; /* Device address */ |
88 | unsigned long vaddr; /* Process virtual addr */ | |
166fd7d9 | 89 | size_t size; /* Map size (bytes) */ |
73fa0d10 | 90 | int prot; /* IOMMU_READ/WRITE */ |
a54eb550 | 91 | bool iommu_mapped; |
ddfd451d | 92 | bool lock_cap; /* capable(CAP_IPC_LOCK) */ |
8f0d5bb9 | 93 | struct task_struct *task; |
a54eb550 | 94 | struct rb_root pfn_list; /* Ex-user pinned pfn list */ |
73fa0d10 AW |
95 | }; |
96 | ||
97 | struct vfio_group { | |
98 | struct iommu_group *iommu_group; | |
99 | struct list_head next; | |
100 | }; | |
101 | ||
a54eb550 KW |
102 | /* |
103 | * Guest RAM pinning working set or DMA target | |
104 | */ | |
105 | struct vfio_pfn { | |
106 | struct rb_node node; | |
107 | dma_addr_t iova; /* Device address */ | |
108 | unsigned long pfn; /* Host pfn */ | |
109 | atomic_t ref_count; | |
110 | }; | |
111 | ||
112 | #define IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu) \ | |
113 | (!list_empty(&iommu->domain_list)) | |
114 | ||
115 | static int put_pfn(unsigned long pfn, int prot); | |
116 | ||
73fa0d10 AW |
117 | /* |
118 | * This code handles mapping and unmapping of user data buffers | |
119 | * into DMA'ble space using the IOMMU | |
120 | */ | |
121 | ||
cd9b2268 AW |
122 | static struct vfio_dma *vfio_find_dma(struct vfio_iommu *iommu, |
123 | dma_addr_t start, size_t size) | |
124 | { | |
125 | struct rb_node *node = iommu->dma_list.rb_node; | |
126 | ||
127 | while (node) { | |
128 | struct vfio_dma *dma = rb_entry(node, struct vfio_dma, node); | |
129 | ||
130 | if (start + size <= dma->iova) | |
131 | node = node->rb_left; | |
166fd7d9 | 132 | else if (start >= dma->iova + dma->size) |
cd9b2268 AW |
133 | node = node->rb_right; |
134 | else | |
135 | return dma; | |
136 | } | |
137 | ||
138 | return NULL; | |
139 | } | |
140 | ||
1ef3e2bc | 141 | static void vfio_link_dma(struct vfio_iommu *iommu, struct vfio_dma *new) |
cd9b2268 AW |
142 | { |
143 | struct rb_node **link = &iommu->dma_list.rb_node, *parent = NULL; | |
144 | struct vfio_dma *dma; | |
145 | ||
146 | while (*link) { | |
147 | parent = *link; | |
148 | dma = rb_entry(parent, struct vfio_dma, node); | |
149 | ||
166fd7d9 | 150 | if (new->iova + new->size <= dma->iova) |
cd9b2268 AW |
151 | link = &(*link)->rb_left; |
152 | else | |
153 | link = &(*link)->rb_right; | |
154 | } | |
155 | ||
156 | rb_link_node(&new->node, parent, link); | |
157 | rb_insert_color(&new->node, &iommu->dma_list); | |
158 | } | |
159 | ||
1ef3e2bc | 160 | static void vfio_unlink_dma(struct vfio_iommu *iommu, struct vfio_dma *old) |
cd9b2268 AW |
161 | { |
162 | rb_erase(&old->node, &iommu->dma_list); | |
163 | } | |
164 | ||
a54eb550 KW |
165 | /* |
166 | * Helper Functions for host iova-pfn list | |
167 | */ | |
168 | static struct vfio_pfn *vfio_find_vpfn(struct vfio_dma *dma, dma_addr_t iova) | |
169 | { | |
170 | struct vfio_pfn *vpfn; | |
171 | struct rb_node *node = dma->pfn_list.rb_node; | |
172 | ||
173 | while (node) { | |
174 | vpfn = rb_entry(node, struct vfio_pfn, node); | |
175 | ||
176 | if (iova < vpfn->iova) | |
177 | node = node->rb_left; | |
178 | else if (iova > vpfn->iova) | |
179 | node = node->rb_right; | |
180 | else | |
181 | return vpfn; | |
182 | } | |
183 | return NULL; | |
184 | } | |
185 | ||
186 | static void vfio_link_pfn(struct vfio_dma *dma, | |
187 | struct vfio_pfn *new) | |
188 | { | |
189 | struct rb_node **link, *parent = NULL; | |
190 | struct vfio_pfn *vpfn; | |
191 | ||
192 | link = &dma->pfn_list.rb_node; | |
193 | while (*link) { | |
194 | parent = *link; | |
195 | vpfn = rb_entry(parent, struct vfio_pfn, node); | |
196 | ||
197 | if (new->iova < vpfn->iova) | |
198 | link = &(*link)->rb_left; | |
199 | else | |
200 | link = &(*link)->rb_right; | |
201 | } | |
202 | ||
203 | rb_link_node(&new->node, parent, link); | |
204 | rb_insert_color(&new->node, &dma->pfn_list); | |
205 | } | |
206 | ||
207 | static void vfio_unlink_pfn(struct vfio_dma *dma, struct vfio_pfn *old) | |
208 | { | |
209 | rb_erase(&old->node, &dma->pfn_list); | |
210 | } | |
211 | ||
212 | static int vfio_add_to_pfn_list(struct vfio_dma *dma, dma_addr_t iova, | |
213 | unsigned long pfn) | |
214 | { | |
215 | struct vfio_pfn *vpfn; | |
216 | ||
217 | vpfn = kzalloc(sizeof(*vpfn), GFP_KERNEL); | |
218 | if (!vpfn) | |
219 | return -ENOMEM; | |
220 | ||
221 | vpfn->iova = iova; | |
222 | vpfn->pfn = pfn; | |
223 | atomic_set(&vpfn->ref_count, 1); | |
224 | vfio_link_pfn(dma, vpfn); | |
225 | return 0; | |
226 | } | |
227 | ||
228 | static void vfio_remove_from_pfn_list(struct vfio_dma *dma, | |
229 | struct vfio_pfn *vpfn) | |
230 | { | |
231 | vfio_unlink_pfn(dma, vpfn); | |
232 | kfree(vpfn); | |
233 | } | |
234 | ||
235 | static struct vfio_pfn *vfio_iova_get_vfio_pfn(struct vfio_dma *dma, | |
236 | unsigned long iova) | |
237 | { | |
238 | struct vfio_pfn *vpfn = vfio_find_vpfn(dma, iova); | |
239 | ||
240 | if (vpfn) | |
241 | atomic_inc(&vpfn->ref_count); | |
242 | return vpfn; | |
243 | } | |
244 | ||
245 | static int vfio_iova_put_vfio_pfn(struct vfio_dma *dma, struct vfio_pfn *vpfn) | |
246 | { | |
247 | int ret = 0; | |
248 | ||
249 | if (atomic_dec_and_test(&vpfn->ref_count)) { | |
250 | ret = put_pfn(vpfn->pfn, dma->prot); | |
251 | vfio_remove_from_pfn_list(dma, vpfn); | |
252 | } | |
253 | return ret; | |
254 | } | |
255 | ||
ddfd451d | 256 | static int vfio_lock_acct(struct vfio_dma *dma, long npage, bool async) |
73fa0d10 | 257 | { |
73fa0d10 | 258 | struct mm_struct *mm; |
0cfef2b7 | 259 | int ret; |
73fa0d10 | 260 | |
3624a248 | 261 | if (!npage) |
0cfef2b7 | 262 | return 0; |
3624a248 | 263 | |
ddfd451d | 264 | mm = async ? get_task_mm(dma->task) : dma->task->mm; |
3624a248 | 265 | if (!mm) |
0cfef2b7 | 266 | return -ESRCH; /* process exited */ |
73fa0d10 | 267 | |
0cfef2b7 AW |
268 | ret = down_write_killable(&mm->mmap_sem); |
269 | if (!ret) { | |
270 | if (npage > 0) { | |
ddfd451d | 271 | if (!dma->lock_cap) { |
0cfef2b7 AW |
272 | unsigned long limit; |
273 | ||
ddfd451d | 274 | limit = task_rlimit(dma->task, |
0cfef2b7 AW |
275 | RLIMIT_MEMLOCK) >> PAGE_SHIFT; |
276 | ||
277 | if (mm->locked_vm + npage > limit) | |
278 | ret = -ENOMEM; | |
279 | } | |
280 | } | |
281 | ||
282 | if (!ret) | |
283 | mm->locked_vm += npage; | |
73fa0d10 | 284 | |
0cfef2b7 | 285 | up_write(&mm->mmap_sem); |
6c38c055 AW |
286 | } |
287 | ||
ddfd451d | 288 | if (async) |
3624a248 | 289 | mmput(mm); |
0cfef2b7 AW |
290 | |
291 | return ret; | |
73fa0d10 AW |
292 | } |
293 | ||
294 | /* | |
295 | * Some mappings aren't backed by a struct page, for example an mmap'd | |
296 | * MMIO range for our own or another device. These use a different | |
297 | * pfn conversion and shouldn't be tracked as locked pages. | |
298 | */ | |
299 | static bool is_invalid_reserved_pfn(unsigned long pfn) | |
300 | { | |
301 | if (pfn_valid(pfn)) { | |
302 | bool reserved; | |
303 | struct page *tail = pfn_to_page(pfn); | |
668f9abb | 304 | struct page *head = compound_head(tail); |
73fa0d10 AW |
305 | reserved = !!(PageReserved(head)); |
306 | if (head != tail) { | |
307 | /* | |
308 | * "head" is not a dangling pointer | |
668f9abb | 309 | * (compound_head takes care of that) |
73fa0d10 AW |
310 | * but the hugepage may have been split |
311 | * from under us (and we may not hold a | |
312 | * reference count on the head page so it can | |
313 | * be reused before we run PageReferenced), so | |
314 | * we've to check PageTail before returning | |
315 | * what we just read. | |
316 | */ | |
317 | smp_rmb(); | |
318 | if (PageTail(tail)) | |
319 | return reserved; | |
320 | } | |
321 | return PageReserved(tail); | |
322 | } | |
323 | ||
324 | return true; | |
325 | } | |
326 | ||
327 | static int put_pfn(unsigned long pfn, int prot) | |
328 | { | |
329 | if (!is_invalid_reserved_pfn(pfn)) { | |
330 | struct page *page = pfn_to_page(pfn); | |
331 | if (prot & IOMMU_WRITE) | |
332 | SetPageDirty(page); | |
333 | put_page(page); | |
334 | return 1; | |
335 | } | |
336 | return 0; | |
337 | } | |
338 | ||
ea85cf35 KW |
339 | static int vaddr_get_pfn(struct mm_struct *mm, unsigned long vaddr, |
340 | int prot, unsigned long *pfn) | |
73fa0d10 AW |
341 | { |
342 | struct page *page[1]; | |
343 | struct vm_area_struct *vma; | |
acf4a054 | 344 | struct vm_area_struct *vmas[1]; |
d432bec2 | 345 | unsigned int flags = 0; |
ea85cf35 | 346 | int ret; |
73fa0d10 | 347 | |
d432bec2 JG |
348 | if (prot & IOMMU_WRITE) |
349 | flags |= FOLL_WRITE; | |
350 | ||
351 | down_read(&mm->mmap_sem); | |
ea85cf35 | 352 | if (mm == current->mm) { |
d432bec2 | 353 | ret = get_user_pages_longterm(vaddr, 1, flags, page, vmas); |
ea85cf35 | 354 | } else { |
ea85cf35 | 355 | ret = get_user_pages_remote(NULL, mm, vaddr, 1, flags, page, |
acf4a054 DW |
356 | vmas, NULL); |
357 | /* | |
358 | * The lifetime of a vaddr_get_pfn() page pin is | |
359 | * userspace-controlled. In the fs-dax case this could | |
360 | * lead to indefinite stalls in filesystem operations. | |
361 | * Disallow attempts to pin fs-dax pages via this | |
362 | * interface. | |
363 | */ | |
364 | if (ret > 0 && vma_is_fsdax(vmas[0])) { | |
365 | ret = -EOPNOTSUPP; | |
366 | put_page(page[0]); | |
367 | } | |
ea85cf35 | 368 | } |
d432bec2 | 369 | up_read(&mm->mmap_sem); |
ea85cf35 KW |
370 | |
371 | if (ret == 1) { | |
73fa0d10 AW |
372 | *pfn = page_to_pfn(page[0]); |
373 | return 0; | |
374 | } | |
375 | ||
ea85cf35 | 376 | down_read(&mm->mmap_sem); |
73fa0d10 | 377 | |
ea85cf35 | 378 | vma = find_vma_intersection(mm, vaddr, vaddr + 1); |
73fa0d10 AW |
379 | |
380 | if (vma && vma->vm_flags & VM_PFNMAP) { | |
381 | *pfn = ((vaddr - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff; | |
382 | if (is_invalid_reserved_pfn(*pfn)) | |
383 | ret = 0; | |
384 | } | |
385 | ||
ea85cf35 | 386 | up_read(&mm->mmap_sem); |
73fa0d10 AW |
387 | return ret; |
388 | } | |
389 | ||
166fd7d9 AW |
390 | /* |
391 | * Attempt to pin pages. We really don't want to track all the pfns and | |
392 | * the iommu can only map chunks of consecutive pfns anyway, so get the | |
393 | * first page and all consecutive pages with the same locking. | |
394 | */ | |
8f0d5bb9 | 395 | static long vfio_pin_pages_remote(struct vfio_dma *dma, unsigned long vaddr, |
7cb671e7 | 396 | long npage, unsigned long *pfn_base, |
ddfd451d | 397 | unsigned long limit) |
73fa0d10 | 398 | { |
7cb671e7 | 399 | unsigned long pfn = 0; |
6c38c055 | 400 | long ret, pinned = 0, lock_acct = 0; |
babbf176 | 401 | bool rsvd; |
a54eb550 | 402 | dma_addr_t iova = vaddr - dma->vaddr + dma->iova; |
73fa0d10 | 403 | |
6c38c055 AW |
404 | /* This code path is only user initiated */ |
405 | if (!current->mm) | |
166fd7d9 | 406 | return -ENODEV; |
73fa0d10 | 407 | |
6c38c055 | 408 | ret = vaddr_get_pfn(current->mm, vaddr, dma->prot, pfn_base); |
166fd7d9 | 409 | if (ret) |
6c38c055 | 410 | return ret; |
73fa0d10 | 411 | |
6c38c055 | 412 | pinned++; |
babbf176 | 413 | rsvd = is_invalid_reserved_pfn(*pfn_base); |
73fa0d10 | 414 | |
a54eb550 KW |
415 | /* |
416 | * Reserved pages aren't counted against the user, externally pinned | |
417 | * pages are already counted against the user. | |
418 | */ | |
419 | if (!rsvd && !vfio_find_vpfn(dma, iova)) { | |
ddfd451d | 420 | if (!dma->lock_cap && current->mm->locked_vm + 1 > limit) { |
a54eb550 KW |
421 | put_pfn(*pfn_base, dma->prot); |
422 | pr_warn("%s: RLIMIT_MEMLOCK (%ld) exceeded\n", __func__, | |
423 | limit << PAGE_SHIFT); | |
6c38c055 | 424 | return -ENOMEM; |
a54eb550 KW |
425 | } |
426 | lock_acct++; | |
5c6c2b21 AW |
427 | } |
428 | ||
6c38c055 AW |
429 | if (unlikely(disable_hugepages)) |
430 | goto out; | |
73fa0d10 | 431 | |
6c38c055 AW |
432 | /* Lock all the consecutive pages from pfn_base */ |
433 | for (vaddr += PAGE_SIZE, iova += PAGE_SIZE; pinned < npage; | |
434 | pinned++, vaddr += PAGE_SIZE, iova += PAGE_SIZE) { | |
6c38c055 AW |
435 | ret = vaddr_get_pfn(current->mm, vaddr, dma->prot, &pfn); |
436 | if (ret) | |
437 | break; | |
438 | ||
439 | if (pfn != *pfn_base + pinned || | |
440 | rsvd != is_invalid_reserved_pfn(pfn)) { | |
441 | put_pfn(pfn, dma->prot); | |
442 | break; | |
443 | } | |
166fd7d9 | 444 | |
6c38c055 | 445 | if (!rsvd && !vfio_find_vpfn(dma, iova)) { |
ddfd451d | 446 | if (!dma->lock_cap && |
6c38c055 | 447 | current->mm->locked_vm + lock_acct + 1 > limit) { |
a54eb550 | 448 | put_pfn(pfn, dma->prot); |
6c38c055 AW |
449 | pr_warn("%s: RLIMIT_MEMLOCK (%ld) exceeded\n", |
450 | __func__, limit << PAGE_SHIFT); | |
0cfef2b7 AW |
451 | ret = -ENOMEM; |
452 | goto unpin_out; | |
a54eb550 | 453 | } |
6c38c055 | 454 | lock_acct++; |
166fd7d9 AW |
455 | } |
456 | } | |
457 | ||
6c38c055 | 458 | out: |
ddfd451d | 459 | ret = vfio_lock_acct(dma, lock_acct, false); |
0cfef2b7 AW |
460 | |
461 | unpin_out: | |
462 | if (ret) { | |
463 | if (!rsvd) { | |
464 | for (pfn = *pfn_base ; pinned ; pfn++, pinned--) | |
465 | put_pfn(pfn, dma->prot); | |
466 | } | |
467 | ||
468 | return ret; | |
469 | } | |
166fd7d9 | 470 | |
6c38c055 | 471 | return pinned; |
166fd7d9 AW |
472 | } |
473 | ||
a54eb550 KW |
474 | static long vfio_unpin_pages_remote(struct vfio_dma *dma, dma_addr_t iova, |
475 | unsigned long pfn, long npage, | |
476 | bool do_accounting) | |
166fd7d9 | 477 | { |
a54eb550 | 478 | long unlocked = 0, locked = 0; |
166fd7d9 AW |
479 | long i; |
480 | ||
6c38c055 | 481 | for (i = 0; i < npage; i++, iova += PAGE_SIZE) { |
a54eb550 KW |
482 | if (put_pfn(pfn++, dma->prot)) { |
483 | unlocked++; | |
6c38c055 | 484 | if (vfio_find_vpfn(dma, iova)) |
a54eb550 KW |
485 | locked++; |
486 | } | |
487 | } | |
488 | ||
489 | if (do_accounting) | |
ddfd451d | 490 | vfio_lock_acct(dma, locked - unlocked, true); |
a54eb550 KW |
491 | |
492 | return unlocked; | |
493 | } | |
494 | ||
495 | static int vfio_pin_page_external(struct vfio_dma *dma, unsigned long vaddr, | |
496 | unsigned long *pfn_base, bool do_accounting) | |
497 | { | |
a54eb550 KW |
498 | struct mm_struct *mm; |
499 | int ret; | |
a54eb550 KW |
500 | |
501 | mm = get_task_mm(dma->task); | |
502 | if (!mm) | |
503 | return -ENODEV; | |
504 | ||
505 | ret = vaddr_get_pfn(mm, vaddr, dma->prot, pfn_base); | |
80dbe1fb | 506 | if (!ret && do_accounting && !is_invalid_reserved_pfn(*pfn_base)) { |
ddfd451d | 507 | ret = vfio_lock_acct(dma, 1, true); |
0cfef2b7 AW |
508 | if (ret) { |
509 | put_pfn(*pfn_base, dma->prot); | |
80dbe1fb AW |
510 | if (ret == -ENOMEM) |
511 | pr_warn("%s: Task %s (%d) RLIMIT_MEMLOCK " | |
512 | "(%ld) exceeded\n", __func__, | |
513 | dma->task->comm, task_pid_nr(dma->task), | |
514 | task_rlimit(dma->task, RLIMIT_MEMLOCK)); | |
0cfef2b7 AW |
515 | } |
516 | } | |
517 | ||
a54eb550 KW |
518 | mmput(mm); |
519 | return ret; | |
520 | } | |
521 | ||
522 | static int vfio_unpin_page_external(struct vfio_dma *dma, dma_addr_t iova, | |
523 | bool do_accounting) | |
524 | { | |
525 | int unlocked; | |
526 | struct vfio_pfn *vpfn = vfio_find_vpfn(dma, iova); | |
527 | ||
528 | if (!vpfn) | |
529 | return 0; | |
530 | ||
531 | unlocked = vfio_iova_put_vfio_pfn(dma, vpfn); | |
166fd7d9 AW |
532 | |
533 | if (do_accounting) | |
ddfd451d | 534 | vfio_lock_acct(dma, -unlocked, true); |
166fd7d9 AW |
535 | |
536 | return unlocked; | |
537 | } | |
538 | ||
a54eb550 KW |
539 | static int vfio_iommu_type1_pin_pages(void *iommu_data, |
540 | unsigned long *user_pfn, | |
541 | int npage, int prot, | |
542 | unsigned long *phys_pfn) | |
543 | { | |
544 | struct vfio_iommu *iommu = iommu_data; | |
545 | int i, j, ret; | |
546 | unsigned long remote_vaddr; | |
547 | struct vfio_dma *dma; | |
548 | bool do_accounting; | |
549 | ||
550 | if (!iommu || !user_pfn || !phys_pfn) | |
551 | return -EINVAL; | |
552 | ||
553 | /* Supported for v2 version only */ | |
554 | if (!iommu->v2) | |
555 | return -EACCES; | |
556 | ||
557 | mutex_lock(&iommu->lock); | |
558 | ||
c086de81 KW |
559 | /* Fail if notifier list is empty */ |
560 | if ((!iommu->external_domain) || (!iommu->notifier.head)) { | |
a54eb550 KW |
561 | ret = -EINVAL; |
562 | goto pin_done; | |
563 | } | |
564 | ||
565 | /* | |
566 | * If iommu capable domain exist in the container then all pages are | |
567 | * already pinned and accounted. Accouting should be done if there is no | |
568 | * iommu capable domain in the container. | |
569 | */ | |
570 | do_accounting = !IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu); | |
571 | ||
572 | for (i = 0; i < npage; i++) { | |
573 | dma_addr_t iova; | |
574 | struct vfio_pfn *vpfn; | |
575 | ||
576 | iova = user_pfn[i] << PAGE_SHIFT; | |
2b8bb1d7 | 577 | dma = vfio_find_dma(iommu, iova, PAGE_SIZE); |
a54eb550 KW |
578 | if (!dma) { |
579 | ret = -EINVAL; | |
580 | goto pin_unwind; | |
581 | } | |
582 | ||
583 | if ((dma->prot & prot) != prot) { | |
584 | ret = -EPERM; | |
585 | goto pin_unwind; | |
586 | } | |
587 | ||
588 | vpfn = vfio_iova_get_vfio_pfn(dma, iova); | |
589 | if (vpfn) { | |
590 | phys_pfn[i] = vpfn->pfn; | |
591 | continue; | |
592 | } | |
593 | ||
594 | remote_vaddr = dma->vaddr + iova - dma->iova; | |
595 | ret = vfio_pin_page_external(dma, remote_vaddr, &phys_pfn[i], | |
596 | do_accounting); | |
80dbe1fb | 597 | if (ret) |
a54eb550 | 598 | goto pin_unwind; |
a54eb550 KW |
599 | |
600 | ret = vfio_add_to_pfn_list(dma, iova, phys_pfn[i]); | |
601 | if (ret) { | |
602 | vfio_unpin_page_external(dma, iova, do_accounting); | |
603 | goto pin_unwind; | |
604 | } | |
605 | } | |
606 | ||
607 | ret = i; | |
608 | goto pin_done; | |
609 | ||
610 | pin_unwind: | |
611 | phys_pfn[i] = 0; | |
612 | for (j = 0; j < i; j++) { | |
613 | dma_addr_t iova; | |
614 | ||
615 | iova = user_pfn[j] << PAGE_SHIFT; | |
2b8bb1d7 | 616 | dma = vfio_find_dma(iommu, iova, PAGE_SIZE); |
a54eb550 KW |
617 | vfio_unpin_page_external(dma, iova, do_accounting); |
618 | phys_pfn[j] = 0; | |
619 | } | |
620 | pin_done: | |
621 | mutex_unlock(&iommu->lock); | |
622 | return ret; | |
623 | } | |
624 | ||
625 | static int vfio_iommu_type1_unpin_pages(void *iommu_data, | |
626 | unsigned long *user_pfn, | |
627 | int npage) | |
628 | { | |
629 | struct vfio_iommu *iommu = iommu_data; | |
630 | bool do_accounting; | |
631 | int i; | |
632 | ||
633 | if (!iommu || !user_pfn) | |
634 | return -EINVAL; | |
635 | ||
636 | /* Supported for v2 version only */ | |
637 | if (!iommu->v2) | |
638 | return -EACCES; | |
639 | ||
640 | mutex_lock(&iommu->lock); | |
641 | ||
642 | if (!iommu->external_domain) { | |
643 | mutex_unlock(&iommu->lock); | |
644 | return -EINVAL; | |
645 | } | |
646 | ||
647 | do_accounting = !IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu); | |
648 | for (i = 0; i < npage; i++) { | |
649 | struct vfio_dma *dma; | |
650 | dma_addr_t iova; | |
651 | ||
652 | iova = user_pfn[i] << PAGE_SHIFT; | |
2b8bb1d7 | 653 | dma = vfio_find_dma(iommu, iova, PAGE_SIZE); |
a54eb550 KW |
654 | if (!dma) |
655 | goto unpin_exit; | |
656 | vfio_unpin_page_external(dma, iova, do_accounting); | |
657 | } | |
658 | ||
659 | unpin_exit: | |
660 | mutex_unlock(&iommu->lock); | |
661 | return i > npage ? npage : (i > 0 ? i : -EINVAL); | |
662 | } | |
663 | ||
664 | static long vfio_unmap_unpin(struct vfio_iommu *iommu, struct vfio_dma *dma, | |
665 | bool do_accounting) | |
166fd7d9 | 666 | { |
1ef3e2bc AW |
667 | dma_addr_t iova = dma->iova, end = dma->iova + dma->size; |
668 | struct vfio_domain *domain, *d; | |
166fd7d9 AW |
669 | long unlocked = 0; |
670 | ||
1ef3e2bc | 671 | if (!dma->size) |
a54eb550 KW |
672 | return 0; |
673 | ||
674 | if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu)) | |
675 | return 0; | |
676 | ||
1ef3e2bc AW |
677 | /* |
678 | * We use the IOMMU to track the physical addresses, otherwise we'd | |
679 | * need a much more complicated tracking system. Unfortunately that | |
680 | * means we need to use one of the iommu domains to figure out the | |
681 | * pfns to unpin. The rest need to be unmapped in advance so we have | |
682 | * no iommu translations remaining when the pages are unpinned. | |
683 | */ | |
684 | domain = d = list_first_entry(&iommu->domain_list, | |
685 | struct vfio_domain, next); | |
686 | ||
c5e66887 | 687 | list_for_each_entry_continue(d, &iommu->domain_list, next) { |
1ef3e2bc | 688 | iommu_unmap(d->domain, dma->iova, dma->size); |
c5e66887 AW |
689 | cond_resched(); |
690 | } | |
1ef3e2bc | 691 | |
166fd7d9 | 692 | while (iova < end) { |
6fe1010d AW |
693 | size_t unmapped, len; |
694 | phys_addr_t phys, next; | |
166fd7d9 | 695 | |
1ef3e2bc | 696 | phys = iommu_iova_to_phys(domain->domain, iova); |
166fd7d9 AW |
697 | if (WARN_ON(!phys)) { |
698 | iova += PAGE_SIZE; | |
699 | continue; | |
73fa0d10 | 700 | } |
166fd7d9 | 701 | |
6fe1010d AW |
702 | /* |
703 | * To optimize for fewer iommu_unmap() calls, each of which | |
704 | * may require hardware cache flushing, try to find the | |
705 | * largest contiguous physical memory chunk to unmap. | |
706 | */ | |
707 | for (len = PAGE_SIZE; | |
708 | !domain->fgsp && iova + len < end; len += PAGE_SIZE) { | |
709 | next = iommu_iova_to_phys(domain->domain, iova + len); | |
710 | if (next != phys + len) | |
711 | break; | |
712 | } | |
713 | ||
714 | unmapped = iommu_unmap(domain->domain, iova, len); | |
1ef3e2bc | 715 | if (WARN_ON(!unmapped)) |
166fd7d9 AW |
716 | break; |
717 | ||
a54eb550 KW |
718 | unlocked += vfio_unpin_pages_remote(dma, iova, |
719 | phys >> PAGE_SHIFT, | |
2169037d | 720 | unmapped >> PAGE_SHIFT, |
a54eb550 | 721 | false); |
166fd7d9 | 722 | iova += unmapped; |
c5e66887 AW |
723 | |
724 | cond_resched(); | |
73fa0d10 | 725 | } |
166fd7d9 | 726 | |
a54eb550 KW |
727 | dma->iommu_mapped = false; |
728 | if (do_accounting) { | |
ddfd451d | 729 | vfio_lock_acct(dma, -unlocked, true); |
a54eb550 KW |
730 | return 0; |
731 | } | |
732 | return unlocked; | |
73fa0d10 AW |
733 | } |
734 | ||
1ef3e2bc | 735 | static void vfio_remove_dma(struct vfio_iommu *iommu, struct vfio_dma *dma) |
73fa0d10 | 736 | { |
a54eb550 | 737 | vfio_unmap_unpin(iommu, dma, true); |
1ef3e2bc | 738 | vfio_unlink_dma(iommu, dma); |
8f0d5bb9 | 739 | put_task_struct(dma->task); |
1ef3e2bc | 740 | kfree(dma); |
a5354ea8 | 741 | iommu->dma_avail++; |
1ef3e2bc | 742 | } |
73fa0d10 | 743 | |
1ef3e2bc AW |
744 | static unsigned long vfio_pgsize_bitmap(struct vfio_iommu *iommu) |
745 | { | |
746 | struct vfio_domain *domain; | |
4644321f | 747 | unsigned long bitmap = ULONG_MAX; |
166fd7d9 | 748 | |
1ef3e2bc AW |
749 | mutex_lock(&iommu->lock); |
750 | list_for_each_entry(domain, &iommu->domain_list, next) | |
d16e0faa | 751 | bitmap &= domain->domain->pgsize_bitmap; |
1ef3e2bc | 752 | mutex_unlock(&iommu->lock); |
73fa0d10 | 753 | |
4644321f EA |
754 | /* |
755 | * In case the IOMMU supports page sizes smaller than PAGE_SIZE | |
756 | * we pretend PAGE_SIZE is supported and hide sub-PAGE_SIZE sizes. | |
757 | * That way the user will be able to map/unmap buffers whose size/ | |
758 | * start address is aligned with PAGE_SIZE. Pinning code uses that | |
759 | * granularity while iommu driver can use the sub-PAGE_SIZE size | |
760 | * to map the buffer. | |
761 | */ | |
762 | if (bitmap & ~PAGE_MASK) { | |
763 | bitmap &= PAGE_MASK; | |
764 | bitmap |= PAGE_SIZE; | |
765 | } | |
766 | ||
1ef3e2bc | 767 | return bitmap; |
73fa0d10 AW |
768 | } |
769 | ||
770 | static int vfio_dma_do_unmap(struct vfio_iommu *iommu, | |
771 | struct vfio_iommu_type1_dma_unmap *unmap) | |
772 | { | |
73fa0d10 | 773 | uint64_t mask; |
c086de81 | 774 | struct vfio_dma *dma, *dma_last = NULL; |
1ef3e2bc | 775 | size_t unmapped = 0; |
c086de81 | 776 | int ret = 0, retries = 0; |
73fa0d10 | 777 | |
1ef3e2bc | 778 | mask = ((uint64_t)1 << __ffs(vfio_pgsize_bitmap(iommu))) - 1; |
73fa0d10 AW |
779 | |
780 | if (unmap->iova & mask) | |
781 | return -EINVAL; | |
f5bfdbf2 | 782 | if (!unmap->size || unmap->size & mask) |
73fa0d10 | 783 | return -EINVAL; |
71a7d3d7 DC |
784 | if (unmap->iova + unmap->size < unmap->iova || |
785 | unmap->size > SIZE_MAX) | |
786 | return -EINVAL; | |
73fa0d10 | 787 | |
73fa0d10 | 788 | WARN_ON(mask & PAGE_MASK); |
c086de81 | 789 | again: |
73fa0d10 AW |
790 | mutex_lock(&iommu->lock); |
791 | ||
1ef3e2bc AW |
792 | /* |
793 | * vfio-iommu-type1 (v1) - User mappings were coalesced together to | |
794 | * avoid tracking individual mappings. This means that the granularity | |
795 | * of the original mapping was lost and the user was allowed to attempt | |
796 | * to unmap any range. Depending on the contiguousness of physical | |
797 | * memory and page sizes supported by the IOMMU, arbitrary unmaps may | |
798 | * or may not have worked. We only guaranteed unmap granularity | |
799 | * matching the original mapping; even though it was untracked here, | |
800 | * the original mappings are reflected in IOMMU mappings. This | |
801 | * resulted in a couple unusual behaviors. First, if a range is not | |
802 | * able to be unmapped, ex. a set of 4k pages that was mapped as a | |
803 | * 2M hugepage into the IOMMU, the unmap ioctl returns success but with | |
804 | * a zero sized unmap. Also, if an unmap request overlaps the first | |
805 | * address of a hugepage, the IOMMU will unmap the entire hugepage. | |
806 | * This also returns success and the returned unmap size reflects the | |
807 | * actual size unmapped. | |
808 | * | |
809 | * We attempt to maintain compatibility with this "v1" interface, but | |
810 | * we take control out of the hands of the IOMMU. Therefore, an unmap | |
811 | * request offset from the beginning of the original mapping will | |
812 | * return success with zero sized unmap. And an unmap request covering | |
813 | * the first iova of mapping will unmap the entire range. | |
814 | * | |
815 | * The v2 version of this interface intends to be more deterministic. | |
816 | * Unmap requests must fully cover previous mappings. Multiple | |
817 | * mappings may still be unmaped by specifying large ranges, but there | |
818 | * must not be any previous mappings bisected by the range. An error | |
819 | * will be returned if these conditions are not met. The v2 interface | |
820 | * will only return success and a size of zero if there were no | |
821 | * mappings within the range. | |
822 | */ | |
823 | if (iommu->v2) { | |
7c03f428 | 824 | dma = vfio_find_dma(iommu, unmap->iova, 1); |
1ef3e2bc AW |
825 | if (dma && dma->iova != unmap->iova) { |
826 | ret = -EINVAL; | |
827 | goto unlock; | |
828 | } | |
829 | dma = vfio_find_dma(iommu, unmap->iova + unmap->size - 1, 0); | |
830 | if (dma && dma->iova + dma->size != unmap->iova + unmap->size) { | |
831 | ret = -EINVAL; | |
832 | goto unlock; | |
833 | } | |
834 | } | |
835 | ||
166fd7d9 | 836 | while ((dma = vfio_find_dma(iommu, unmap->iova, unmap->size))) { |
1ef3e2bc | 837 | if (!iommu->v2 && unmap->iova > dma->iova) |
166fd7d9 | 838 | break; |
8f0d5bb9 KW |
839 | /* |
840 | * Task with same address space who mapped this iova range is | |
841 | * allowed to unmap the iova range. | |
842 | */ | |
843 | if (dma->task->mm != current->mm) | |
844 | break; | |
c086de81 KW |
845 | |
846 | if (!RB_EMPTY_ROOT(&dma->pfn_list)) { | |
847 | struct vfio_iommu_type1_dma_unmap nb_unmap; | |
848 | ||
849 | if (dma_last == dma) { | |
850 | BUG_ON(++retries > 10); | |
851 | } else { | |
852 | dma_last = dma; | |
853 | retries = 0; | |
854 | } | |
855 | ||
856 | nb_unmap.iova = dma->iova; | |
857 | nb_unmap.size = dma->size; | |
858 | ||
859 | /* | |
860 | * Notify anyone (mdev vendor drivers) to invalidate and | |
861 | * unmap iovas within the range we're about to unmap. | |
862 | * Vendor drivers MUST unpin pages in response to an | |
863 | * invalidation. | |
864 | */ | |
865 | mutex_unlock(&iommu->lock); | |
866 | blocking_notifier_call_chain(&iommu->notifier, | |
867 | VFIO_IOMMU_NOTIFY_DMA_UNMAP, | |
868 | &nb_unmap); | |
869 | goto again; | |
870 | } | |
1ef3e2bc AW |
871 | unmapped += dma->size; |
872 | vfio_remove_dma(iommu, dma); | |
166fd7d9 | 873 | } |
cd9b2268 | 874 | |
1ef3e2bc | 875 | unlock: |
73fa0d10 | 876 | mutex_unlock(&iommu->lock); |
166fd7d9 | 877 | |
1ef3e2bc | 878 | /* Report how much was unmapped */ |
166fd7d9 AW |
879 | unmap->size = unmapped; |
880 | ||
881 | return ret; | |
882 | } | |
883 | ||
884 | /* | |
885 | * Turns out AMD IOMMU has a page table bug where it won't map large pages | |
886 | * to a region that previously mapped smaller pages. This should be fixed | |
887 | * soon, so this is just a temporary workaround to break mappings down into | |
888 | * PAGE_SIZE. Better to map smaller pages than nothing. | |
889 | */ | |
1ef3e2bc | 890 | static int map_try_harder(struct vfio_domain *domain, dma_addr_t iova, |
166fd7d9 AW |
891 | unsigned long pfn, long npage, int prot) |
892 | { | |
893 | long i; | |
089f1c6b | 894 | int ret = 0; |
166fd7d9 AW |
895 | |
896 | for (i = 0; i < npage; i++, pfn++, iova += PAGE_SIZE) { | |
1ef3e2bc | 897 | ret = iommu_map(domain->domain, iova, |
166fd7d9 | 898 | (phys_addr_t)pfn << PAGE_SHIFT, |
1ef3e2bc | 899 | PAGE_SIZE, prot | domain->prot); |
166fd7d9 AW |
900 | if (ret) |
901 | break; | |
902 | } | |
903 | ||
904 | for (; i < npage && i > 0; i--, iova -= PAGE_SIZE) | |
1ef3e2bc AW |
905 | iommu_unmap(domain->domain, iova, PAGE_SIZE); |
906 | ||
907 | return ret; | |
908 | } | |
909 | ||
910 | static int vfio_iommu_map(struct vfio_iommu *iommu, dma_addr_t iova, | |
911 | unsigned long pfn, long npage, int prot) | |
912 | { | |
913 | struct vfio_domain *d; | |
914 | int ret; | |
915 | ||
916 | list_for_each_entry(d, &iommu->domain_list, next) { | |
917 | ret = iommu_map(d->domain, iova, (phys_addr_t)pfn << PAGE_SHIFT, | |
918 | npage << PAGE_SHIFT, prot | d->prot); | |
919 | if (ret) { | |
920 | if (ret != -EBUSY || | |
921 | map_try_harder(d, iova, pfn, npage, prot)) | |
922 | goto unwind; | |
923 | } | |
c5e66887 AW |
924 | |
925 | cond_resched(); | |
1ef3e2bc AW |
926 | } |
927 | ||
928 | return 0; | |
929 | ||
930 | unwind: | |
931 | list_for_each_entry_continue_reverse(d, &iommu->domain_list, next) | |
932 | iommu_unmap(d->domain, iova, npage << PAGE_SHIFT); | |
166fd7d9 | 933 | |
cd9b2268 | 934 | return ret; |
73fa0d10 AW |
935 | } |
936 | ||
8f0d5bb9 KW |
937 | static int vfio_pin_map_dma(struct vfio_iommu *iommu, struct vfio_dma *dma, |
938 | size_t map_size) | |
939 | { | |
940 | dma_addr_t iova = dma->iova; | |
941 | unsigned long vaddr = dma->vaddr; | |
942 | size_t size = map_size; | |
943 | long npage; | |
7cb671e7 | 944 | unsigned long pfn, limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT; |
8f0d5bb9 KW |
945 | int ret = 0; |
946 | ||
947 | while (size) { | |
948 | /* Pin a contiguous chunk of memory */ | |
949 | npage = vfio_pin_pages_remote(dma, vaddr + dma->size, | |
ddfd451d | 950 | size >> PAGE_SHIFT, &pfn, 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 | ||
a5354ea8 AW |
1016 | if (!iommu->dma_avail) { |
1017 | ret = -ENOSPC; | |
1018 | goto out_unlock; | |
1019 | } | |
1020 | ||
1ef3e2bc AW |
1021 | dma = kzalloc(sizeof(*dma), GFP_KERNEL); |
1022 | if (!dma) { | |
8f0d5bb9 KW |
1023 | ret = -ENOMEM; |
1024 | goto out_unlock; | |
1ef3e2bc AW |
1025 | } |
1026 | ||
a5354ea8 | 1027 | iommu->dma_avail--; |
c8dbca16 AW |
1028 | dma->iova = iova; |
1029 | dma->vaddr = vaddr; | |
1ef3e2bc | 1030 | dma->prot = prot; |
ddfd451d AW |
1031 | |
1032 | /* | |
1033 | * We need to be able to both add to a task's locked memory and test | |
1034 | * against the locked memory limit and we need to be able to do both | |
1035 | * outside of this call path as pinning can be asynchronous via the | |
1036 | * external interfaces for mdev devices. RLIMIT_MEMLOCK requires a | |
1037 | * task_struct and VM locked pages requires an mm_struct, however | |
1038 | * holding an indefinite mm reference is not recommended, therefore we | |
1039 | * only hold a reference to a task. We could hold a reference to | |
1040 | * current, however QEMU uses this call path through vCPU threads, | |
1041 | * which can be killed resulting in a NULL mm and failure in the unmap | |
1042 | * path when called via a different thread. Avoid this problem by | |
1043 | * using the group_leader as threads within the same group require | |
1044 | * both CLONE_THREAD and CLONE_VM and will therefore use the same | |
1045 | * mm_struct. | |
1046 | * | |
1047 | * Previously we also used the task for testing CAP_IPC_LOCK at the | |
1048 | * time of pinning and accounting, however has_capability() makes use | |
1049 | * of real_cred, a copy-on-write field, so we can't guarantee that it | |
1050 | * matches group_leader, or in fact that it might not change by the | |
1051 | * time it's evaluated. If a process were to call MAP_DMA with | |
1052 | * CAP_IPC_LOCK but later drop it, it doesn't make sense that they | |
1053 | * possibly see different results for an iommu_mapped vfio_dma vs | |
1054 | * externally mapped. Therefore track CAP_IPC_LOCK in vfio_dma at the | |
1055 | * time of calling MAP_DMA. | |
1056 | */ | |
1057 | get_task_struct(current->group_leader); | |
1058 | dma->task = current->group_leader; | |
1059 | dma->lock_cap = capable(CAP_IPC_LOCK); | |
1060 | ||
a54eb550 | 1061 | dma->pfn_list = RB_ROOT; |
166fd7d9 | 1062 | |
1ef3e2bc AW |
1063 | /* Insert zero-sized and grow as we map chunks of it */ |
1064 | vfio_link_dma(iommu, dma); | |
166fd7d9 | 1065 | |
a54eb550 KW |
1066 | /* Don't pin and map if container doesn't contain IOMMU capable domain*/ |
1067 | if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu)) | |
1068 | dma->size = size; | |
1069 | else | |
1070 | ret = vfio_pin_map_dma(iommu, dma, size); | |
1071 | ||
8f0d5bb9 | 1072 | out_unlock: |
1ef3e2bc AW |
1073 | mutex_unlock(&iommu->lock); |
1074 | return ret; | |
1075 | } | |
1076 | ||
1077 | static int vfio_bus_type(struct device *dev, void *data) | |
1078 | { | |
1079 | struct bus_type **bus = data; | |
1080 | ||
1081 | if (*bus && *bus != dev->bus) | |
1082 | return -EINVAL; | |
1083 | ||
1084 | *bus = dev->bus; | |
1085 | ||
1086 | return 0; | |
1087 | } | |
1088 | ||
1089 | static int vfio_iommu_replay(struct vfio_iommu *iommu, | |
1090 | struct vfio_domain *domain) | |
1091 | { | |
1092 | struct vfio_domain *d; | |
1093 | struct rb_node *n; | |
7cb671e7 | 1094 | unsigned long limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT; |
1ef3e2bc AW |
1095 | int ret; |
1096 | ||
1097 | /* Arbitrarily pick the first domain in the list for lookups */ | |
1098 | d = list_first_entry(&iommu->domain_list, struct vfio_domain, next); | |
1099 | n = rb_first(&iommu->dma_list); | |
1100 | ||
1ef3e2bc AW |
1101 | for (; n; n = rb_next(n)) { |
1102 | struct vfio_dma *dma; | |
1103 | dma_addr_t iova; | |
1104 | ||
1105 | dma = rb_entry(n, struct vfio_dma, node); | |
1106 | iova = dma->iova; | |
1107 | ||
1108 | while (iova < dma->iova + dma->size) { | |
a54eb550 | 1109 | phys_addr_t phys; |
1ef3e2bc | 1110 | size_t size; |
73fa0d10 | 1111 | |
a54eb550 KW |
1112 | if (dma->iommu_mapped) { |
1113 | phys_addr_t p; | |
1114 | dma_addr_t i; | |
1115 | ||
1116 | phys = iommu_iova_to_phys(d->domain, iova); | |
1117 | ||
1118 | if (WARN_ON(!phys)) { | |
1119 | iova += PAGE_SIZE; | |
1120 | continue; | |
1121 | } | |
1122 | ||
1123 | size = PAGE_SIZE; | |
1124 | p = phys + size; | |
1125 | i = iova + size; | |
1126 | while (i < dma->iova + dma->size && | |
1127 | p == iommu_iova_to_phys(d->domain, i)) { | |
1128 | size += PAGE_SIZE; | |
1129 | p += PAGE_SIZE; | |
1130 | i += PAGE_SIZE; | |
1131 | } | |
1132 | } else { | |
1133 | unsigned long pfn; | |
1134 | unsigned long vaddr = dma->vaddr + | |
1135 | (iova - dma->iova); | |
1136 | size_t n = dma->iova + dma->size - iova; | |
1137 | long npage; | |
1138 | ||
1139 | npage = vfio_pin_pages_remote(dma, vaddr, | |
1140 | n >> PAGE_SHIFT, | |
ddfd451d | 1141 | &pfn, limit); |
a54eb550 KW |
1142 | if (npage <= 0) { |
1143 | WARN_ON(!npage); | |
1144 | ret = (int)npage; | |
1145 | return ret; | |
1146 | } | |
1147 | ||
1148 | phys = pfn << PAGE_SHIFT; | |
1149 | size = npage << PAGE_SHIFT; | |
166fd7d9 AW |
1150 | } |
1151 | ||
1ef3e2bc AW |
1152 | ret = iommu_map(domain->domain, iova, phys, |
1153 | size, dma->prot | domain->prot); | |
1154 | if (ret) | |
1155 | return ret; | |
d93b3ac0 | 1156 | |
1ef3e2bc AW |
1157 | iova += size; |
1158 | } | |
a54eb550 | 1159 | dma->iommu_mapped = true; |
166fd7d9 | 1160 | } |
1ef3e2bc | 1161 | return 0; |
73fa0d10 AW |
1162 | } |
1163 | ||
6fe1010d AW |
1164 | /* |
1165 | * We change our unmap behavior slightly depending on whether the IOMMU | |
1166 | * supports fine-grained superpages. IOMMUs like AMD-Vi will use a superpage | |
1167 | * for practically any contiguous power-of-two mapping we give it. This means | |
1168 | * we don't need to look for contiguous chunks ourselves to make unmapping | |
1169 | * more efficient. On IOMMUs with coarse-grained super pages, like Intel VT-d | |
1170 | * with discrete 2M/1G/512G/1T superpages, identifying contiguous chunks | |
1171 | * significantly boosts non-hugetlbfs mappings and doesn't seem to hurt when | |
1172 | * hugetlbfs is in use. | |
1173 | */ | |
1174 | static void vfio_test_domain_fgsp(struct vfio_domain *domain) | |
1175 | { | |
1176 | struct page *pages; | |
1177 | int ret, order = get_order(PAGE_SIZE * 2); | |
1178 | ||
1179 | pages = alloc_pages(GFP_KERNEL | __GFP_ZERO, order); | |
1180 | if (!pages) | |
1181 | return; | |
1182 | ||
1183 | ret = iommu_map(domain->domain, 0, page_to_phys(pages), PAGE_SIZE * 2, | |
1184 | IOMMU_READ | IOMMU_WRITE | domain->prot); | |
1185 | if (!ret) { | |
1186 | size_t unmapped = iommu_unmap(domain->domain, 0, PAGE_SIZE); | |
1187 | ||
1188 | if (unmapped == PAGE_SIZE) | |
1189 | iommu_unmap(domain->domain, PAGE_SIZE, PAGE_SIZE); | |
1190 | else | |
1191 | domain->fgsp = true; | |
1192 | } | |
1193 | ||
1194 | __free_pages(pages, order); | |
1195 | } | |
1196 | ||
7896c998 KW |
1197 | static struct vfio_group *find_iommu_group(struct vfio_domain *domain, |
1198 | struct iommu_group *iommu_group) | |
1199 | { | |
1200 | struct vfio_group *g; | |
1201 | ||
1202 | list_for_each_entry(g, &domain->group_list, next) { | |
1203 | if (g->iommu_group == iommu_group) | |
1204 | return g; | |
1205 | } | |
1206 | ||
1207 | return NULL; | |
1208 | } | |
1209 | ||
9d3a4de4 | 1210 | static bool vfio_iommu_has_sw_msi(struct iommu_group *group, phys_addr_t *base) |
5d704992 EA |
1211 | { |
1212 | struct list_head group_resv_regions; | |
1213 | struct iommu_resv_region *region, *next; | |
1214 | bool ret = false; | |
1215 | ||
1216 | INIT_LIST_HEAD(&group_resv_regions); | |
1217 | iommu_get_group_resv_regions(group, &group_resv_regions); | |
1218 | list_for_each_entry(region, &group_resv_regions, list) { | |
f203f7f1 RM |
1219 | /* |
1220 | * The presence of any 'real' MSI regions should take | |
1221 | * precedence over the software-managed one if the | |
1222 | * IOMMU driver happens to advertise both types. | |
1223 | */ | |
1224 | if (region->type == IOMMU_RESV_MSI) { | |
1225 | ret = false; | |
1226 | break; | |
1227 | } | |
1228 | ||
9d3a4de4 | 1229 | if (region->type == IOMMU_RESV_SW_MSI) { |
5d704992 EA |
1230 | *base = region->start; |
1231 | ret = true; | |
5d704992 EA |
1232 | } |
1233 | } | |
5d704992 EA |
1234 | list_for_each_entry_safe(region, next, &group_resv_regions, list) |
1235 | kfree(region); | |
1236 | return ret; | |
1237 | } | |
1238 | ||
73fa0d10 AW |
1239 | static int vfio_iommu_type1_attach_group(void *iommu_data, |
1240 | struct iommu_group *iommu_group) | |
1241 | { | |
1242 | struct vfio_iommu *iommu = iommu_data; | |
7896c998 | 1243 | struct vfio_group *group; |
1ef3e2bc | 1244 | struct vfio_domain *domain, *d; |
a54eb550 | 1245 | struct bus_type *bus = NULL, *mdev_bus; |
73fa0d10 | 1246 | int ret; |
9d72f87b | 1247 | bool resv_msi, msi_remap; |
5d704992 | 1248 | phys_addr_t resv_msi_base; |
73fa0d10 | 1249 | |
73fa0d10 AW |
1250 | mutex_lock(&iommu->lock); |
1251 | ||
1ef3e2bc | 1252 | list_for_each_entry(d, &iommu->domain_list, next) { |
7896c998 | 1253 | if (find_iommu_group(d, iommu_group)) { |
73fa0d10 | 1254 | mutex_unlock(&iommu->lock); |
73fa0d10 AW |
1255 | return -EINVAL; |
1256 | } | |
1257 | } | |
1258 | ||
a54eb550 KW |
1259 | if (iommu->external_domain) { |
1260 | if (find_iommu_group(iommu->external_domain, iommu_group)) { | |
1261 | mutex_unlock(&iommu->lock); | |
1262 | return -EINVAL; | |
1263 | } | |
1264 | } | |
1265 | ||
1ef3e2bc AW |
1266 | group = kzalloc(sizeof(*group), GFP_KERNEL); |
1267 | domain = kzalloc(sizeof(*domain), GFP_KERNEL); | |
1268 | if (!group || !domain) { | |
1269 | ret = -ENOMEM; | |
1270 | goto out_free; | |
1271 | } | |
1272 | ||
1273 | group->iommu_group = iommu_group; | |
1274 | ||
1275 | /* Determine bus_type in order to allocate a domain */ | |
1276 | ret = iommu_group_for_each_dev(iommu_group, &bus, vfio_bus_type); | |
1277 | if (ret) | |
1278 | goto out_free; | |
1279 | ||
a54eb550 KW |
1280 | mdev_bus = symbol_get(mdev_bus_type); |
1281 | ||
1282 | if (mdev_bus) { | |
1283 | if ((bus == mdev_bus) && !iommu_present(bus)) { | |
1284 | symbol_put(mdev_bus_type); | |
1285 | if (!iommu->external_domain) { | |
1286 | INIT_LIST_HEAD(&domain->group_list); | |
1287 | iommu->external_domain = domain; | |
1288 | } else | |
1289 | kfree(domain); | |
1290 | ||
1291 | list_add(&group->next, | |
1292 | &iommu->external_domain->group_list); | |
1293 | mutex_unlock(&iommu->lock); | |
1294 | return 0; | |
1295 | } | |
1296 | symbol_put(mdev_bus_type); | |
1297 | } | |
1298 | ||
1ef3e2bc AW |
1299 | domain->domain = iommu_domain_alloc(bus); |
1300 | if (!domain->domain) { | |
1301 | ret = -EIO; | |
1302 | goto out_free; | |
1303 | } | |
1304 | ||
f5c9eceb WD |
1305 | if (iommu->nesting) { |
1306 | int attr = 1; | |
1307 | ||
1308 | ret = iommu_domain_set_attr(domain->domain, DOMAIN_ATTR_NESTING, | |
1309 | &attr); | |
1310 | if (ret) | |
1311 | goto out_domain; | |
1312 | } | |
1313 | ||
1ef3e2bc AW |
1314 | ret = iommu_attach_group(domain->domain, iommu_group); |
1315 | if (ret) | |
1316 | goto out_domain; | |
1317 | ||
9d3a4de4 | 1318 | resv_msi = vfio_iommu_has_sw_msi(iommu_group, &resv_msi_base); |
5d704992 | 1319 | |
1ef3e2bc AW |
1320 | INIT_LIST_HEAD(&domain->group_list); |
1321 | list_add(&group->next, &domain->group_list); | |
1322 | ||
db406cc0 RM |
1323 | msi_remap = irq_domain_check_msi_remap() || |
1324 | iommu_capable(bus, IOMMU_CAP_INTR_REMAP); | |
9d72f87b EA |
1325 | |
1326 | if (!allow_unsafe_interrupts && !msi_remap) { | |
1ef3e2bc AW |
1327 | pr_warn("%s: No interrupt remapping support. Use the module param \"allow_unsafe_interrupts\" to enable VFIO IOMMU support on this platform\n", |
1328 | __func__); | |
1329 | ret = -EPERM; | |
1330 | goto out_detach; | |
1331 | } | |
1332 | ||
eb165f05 | 1333 | if (iommu_capable(bus, IOMMU_CAP_CACHE_COHERENCY)) |
1ef3e2bc AW |
1334 | domain->prot |= IOMMU_CACHE; |
1335 | ||
73fa0d10 | 1336 | /* |
1ef3e2bc AW |
1337 | * Try to match an existing compatible domain. We don't want to |
1338 | * preclude an IOMMU driver supporting multiple bus_types and being | |
1339 | * able to include different bus_types in the same IOMMU domain, so | |
1340 | * we test whether the domains use the same iommu_ops rather than | |
1341 | * testing if they're on the same bus_type. | |
73fa0d10 | 1342 | */ |
1ef3e2bc AW |
1343 | list_for_each_entry(d, &iommu->domain_list, next) { |
1344 | if (d->domain->ops == domain->domain->ops && | |
1345 | d->prot == domain->prot) { | |
1346 | iommu_detach_group(domain->domain, iommu_group); | |
1347 | if (!iommu_attach_group(d->domain, iommu_group)) { | |
1348 | list_add(&group->next, &d->group_list); | |
1349 | iommu_domain_free(domain->domain); | |
1350 | kfree(domain); | |
1351 | mutex_unlock(&iommu->lock); | |
1352 | return 0; | |
1353 | } | |
1354 | ||
1355 | ret = iommu_attach_group(domain->domain, iommu_group); | |
1356 | if (ret) | |
1357 | goto out_domain; | |
1358 | } | |
73fa0d10 AW |
1359 | } |
1360 | ||
6fe1010d AW |
1361 | vfio_test_domain_fgsp(domain); |
1362 | ||
1ef3e2bc AW |
1363 | /* replay mappings on new domains */ |
1364 | ret = vfio_iommu_replay(iommu, domain); | |
1365 | if (ret) | |
1366 | goto out_detach; | |
1367 | ||
2c9f1af5 WY |
1368 | if (resv_msi) { |
1369 | ret = iommu_get_msi_cookie(domain->domain, resv_msi_base); | |
1370 | if (ret) | |
1371 | goto out_detach; | |
1372 | } | |
5d704992 | 1373 | |
1ef3e2bc | 1374 | list_add(&domain->next, &iommu->domain_list); |
73fa0d10 AW |
1375 | |
1376 | mutex_unlock(&iommu->lock); | |
1377 | ||
1378 | return 0; | |
1ef3e2bc AW |
1379 | |
1380 | out_detach: | |
1381 | iommu_detach_group(domain->domain, iommu_group); | |
1382 | out_domain: | |
1383 | iommu_domain_free(domain->domain); | |
1384 | out_free: | |
1385 | kfree(domain); | |
1386 | kfree(group); | |
1387 | mutex_unlock(&iommu->lock); | |
1388 | return ret; | |
1389 | } | |
1390 | ||
1391 | static void vfio_iommu_unmap_unpin_all(struct vfio_iommu *iommu) | |
1392 | { | |
1393 | struct rb_node *node; | |
1394 | ||
1395 | while ((node = rb_first(&iommu->dma_list))) | |
1396 | vfio_remove_dma(iommu, rb_entry(node, struct vfio_dma, node)); | |
73fa0d10 AW |
1397 | } |
1398 | ||
a54eb550 KW |
1399 | static void vfio_iommu_unmap_unpin_reaccount(struct vfio_iommu *iommu) |
1400 | { | |
1401 | struct rb_node *n, *p; | |
1402 | ||
1403 | n = rb_first(&iommu->dma_list); | |
1404 | for (; n; n = rb_next(n)) { | |
1405 | struct vfio_dma *dma; | |
1406 | long locked = 0, unlocked = 0; | |
1407 | ||
1408 | dma = rb_entry(n, struct vfio_dma, node); | |
1409 | unlocked += vfio_unmap_unpin(iommu, dma, false); | |
1410 | p = rb_first(&dma->pfn_list); | |
1411 | for (; p; p = rb_next(p)) { | |
1412 | struct vfio_pfn *vpfn = rb_entry(p, struct vfio_pfn, | |
1413 | node); | |
1414 | ||
1415 | if (!is_invalid_reserved_pfn(vpfn->pfn)) | |
1416 | locked++; | |
1417 | } | |
ddfd451d | 1418 | vfio_lock_acct(dma, locked - unlocked, true); |
a54eb550 KW |
1419 | } |
1420 | } | |
1421 | ||
1422 | static void vfio_sanity_check_pfn_list(struct vfio_iommu *iommu) | |
1423 | { | |
1424 | struct rb_node *n; | |
1425 | ||
1426 | n = rb_first(&iommu->dma_list); | |
1427 | for (; n; n = rb_next(n)) { | |
1428 | struct vfio_dma *dma; | |
1429 | ||
1430 | dma = rb_entry(n, struct vfio_dma, node); | |
1431 | ||
1432 | if (WARN_ON(!RB_EMPTY_ROOT(&dma->pfn_list))) | |
1433 | break; | |
1434 | } | |
3cedd7d7 KW |
1435 | /* mdev vendor driver must unregister notifier */ |
1436 | WARN_ON(iommu->notifier.head); | |
a54eb550 KW |
1437 | } |
1438 | ||
73fa0d10 AW |
1439 | static void vfio_iommu_type1_detach_group(void *iommu_data, |
1440 | struct iommu_group *iommu_group) | |
1441 | { | |
1442 | struct vfio_iommu *iommu = iommu_data; | |
1ef3e2bc | 1443 | struct vfio_domain *domain; |
73fa0d10 AW |
1444 | struct vfio_group *group; |
1445 | ||
1446 | mutex_lock(&iommu->lock); | |
1447 | ||
a54eb550 KW |
1448 | if (iommu->external_domain) { |
1449 | group = find_iommu_group(iommu->external_domain, iommu_group); | |
1450 | if (group) { | |
1451 | list_del(&group->next); | |
1452 | kfree(group); | |
1453 | ||
1454 | if (list_empty(&iommu->external_domain->group_list)) { | |
1455 | vfio_sanity_check_pfn_list(iommu); | |
1456 | ||
1457 | if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu)) | |
1458 | vfio_iommu_unmap_unpin_all(iommu); | |
1459 | ||
1460 | kfree(iommu->external_domain); | |
1461 | iommu->external_domain = NULL; | |
1462 | } | |
1463 | goto detach_group_done; | |
1464 | } | |
1465 | } | |
1466 | ||
1ef3e2bc | 1467 | list_for_each_entry(domain, &iommu->domain_list, next) { |
7896c998 KW |
1468 | group = find_iommu_group(domain, iommu_group); |
1469 | if (!group) | |
1470 | continue; | |
1ef3e2bc | 1471 | |
7896c998 KW |
1472 | iommu_detach_group(domain->domain, iommu_group); |
1473 | list_del(&group->next); | |
1474 | kfree(group); | |
1475 | /* | |
a54eb550 KW |
1476 | * Group ownership provides privilege, if the group list is |
1477 | * empty, the domain goes away. If it's the last domain with | |
1478 | * iommu and external domain doesn't exist, then all the | |
1479 | * mappings go away too. If it's the last domain with iommu and | |
1480 | * external domain exist, update accounting | |
7896c998 KW |
1481 | */ |
1482 | if (list_empty(&domain->group_list)) { | |
a54eb550 KW |
1483 | if (list_is_singular(&iommu->domain_list)) { |
1484 | if (!iommu->external_domain) | |
1485 | vfio_iommu_unmap_unpin_all(iommu); | |
1486 | else | |
1487 | vfio_iommu_unmap_unpin_reaccount(iommu); | |
1488 | } | |
7896c998 KW |
1489 | iommu_domain_free(domain->domain); |
1490 | list_del(&domain->next); | |
1491 | kfree(domain); | |
73fa0d10 | 1492 | } |
a54eb550 | 1493 | break; |
73fa0d10 AW |
1494 | } |
1495 | ||
a54eb550 | 1496 | detach_group_done: |
73fa0d10 AW |
1497 | mutex_unlock(&iommu->lock); |
1498 | } | |
1499 | ||
1500 | static void *vfio_iommu_type1_open(unsigned long arg) | |
1501 | { | |
1502 | struct vfio_iommu *iommu; | |
1503 | ||
73fa0d10 AW |
1504 | iommu = kzalloc(sizeof(*iommu), GFP_KERNEL); |
1505 | if (!iommu) | |
1506 | return ERR_PTR(-ENOMEM); | |
1507 | ||
f5c9eceb WD |
1508 | switch (arg) { |
1509 | case VFIO_TYPE1_IOMMU: | |
1510 | break; | |
1511 | case VFIO_TYPE1_NESTING_IOMMU: | |
1512 | iommu->nesting = true; | |
1513 | case VFIO_TYPE1v2_IOMMU: | |
1514 | iommu->v2 = true; | |
1515 | break; | |
1516 | default: | |
1517 | kfree(iommu); | |
1518 | return ERR_PTR(-EINVAL); | |
1519 | } | |
1520 | ||
1ef3e2bc | 1521 | INIT_LIST_HEAD(&iommu->domain_list); |
cd9b2268 | 1522 | iommu->dma_list = RB_ROOT; |
a5354ea8 | 1523 | iommu->dma_avail = dma_entry_limit; |
73fa0d10 | 1524 | mutex_init(&iommu->lock); |
c086de81 | 1525 | BLOCKING_INIT_NOTIFIER_HEAD(&iommu->notifier); |
73fa0d10 AW |
1526 | |
1527 | return iommu; | |
1528 | } | |
1529 | ||
a54eb550 KW |
1530 | static void vfio_release_domain(struct vfio_domain *domain, bool external) |
1531 | { | |
1532 | struct vfio_group *group, *group_tmp; | |
1533 | ||
1534 | list_for_each_entry_safe(group, group_tmp, | |
1535 | &domain->group_list, next) { | |
1536 | if (!external) | |
1537 | iommu_detach_group(domain->domain, group->iommu_group); | |
1538 | list_del(&group->next); | |
1539 | kfree(group); | |
1540 | } | |
1541 | ||
1542 | if (!external) | |
1543 | iommu_domain_free(domain->domain); | |
1544 | } | |
1545 | ||
73fa0d10 AW |
1546 | static void vfio_iommu_type1_release(void *iommu_data) |
1547 | { | |
1548 | struct vfio_iommu *iommu = iommu_data; | |
1ef3e2bc | 1549 | struct vfio_domain *domain, *domain_tmp; |
a54eb550 KW |
1550 | |
1551 | if (iommu->external_domain) { | |
1552 | vfio_release_domain(iommu->external_domain, true); | |
1553 | vfio_sanity_check_pfn_list(iommu); | |
1554 | kfree(iommu->external_domain); | |
1555 | } | |
73fa0d10 | 1556 | |
1ef3e2bc | 1557 | vfio_iommu_unmap_unpin_all(iommu); |
73fa0d10 | 1558 | |
1ef3e2bc AW |
1559 | list_for_each_entry_safe(domain, domain_tmp, |
1560 | &iommu->domain_list, next) { | |
a54eb550 | 1561 | vfio_release_domain(domain, false); |
1ef3e2bc AW |
1562 | list_del(&domain->next); |
1563 | kfree(domain); | |
73fa0d10 | 1564 | } |
73fa0d10 AW |
1565 | kfree(iommu); |
1566 | } | |
1567 | ||
aa429318 AW |
1568 | static int vfio_domains_have_iommu_cache(struct vfio_iommu *iommu) |
1569 | { | |
1570 | struct vfio_domain *domain; | |
1571 | int ret = 1; | |
1572 | ||
1573 | mutex_lock(&iommu->lock); | |
1574 | list_for_each_entry(domain, &iommu->domain_list, next) { | |
1575 | if (!(domain->prot & IOMMU_CACHE)) { | |
1576 | ret = 0; | |
f5bfdbf2 | 1577 | break; |
aa429318 | 1578 | } |
73fa0d10 | 1579 | } |
aa429318 | 1580 | mutex_unlock(&iommu->lock); |
73fa0d10 | 1581 | |
aa429318 | 1582 | return ret; |
73fa0d10 AW |
1583 | } |
1584 | ||
1585 | static long vfio_iommu_type1_ioctl(void *iommu_data, | |
1586 | unsigned int cmd, unsigned long arg) | |
1587 | { | |
1588 | struct vfio_iommu *iommu = iommu_data; | |
1589 | unsigned long minsz; | |
1590 | ||
1591 | if (cmd == VFIO_CHECK_EXTENSION) { | |
1592 | switch (arg) { | |
1593 | case VFIO_TYPE1_IOMMU: | |
1ef3e2bc | 1594 | case VFIO_TYPE1v2_IOMMU: |
f5c9eceb | 1595 | case VFIO_TYPE1_NESTING_IOMMU: |
73fa0d10 | 1596 | return 1; |
aa429318 AW |
1597 | case VFIO_DMA_CC_IOMMU: |
1598 | if (!iommu) | |
1599 | return 0; | |
1600 | return vfio_domains_have_iommu_cache(iommu); | |
73fa0d10 AW |
1601 | default: |
1602 | return 0; | |
1603 | } | |
1604 | } else if (cmd == VFIO_IOMMU_GET_INFO) { | |
1605 | struct vfio_iommu_type1_info info; | |
1606 | ||
1607 | minsz = offsetofend(struct vfio_iommu_type1_info, iova_pgsizes); | |
1608 | ||
1609 | if (copy_from_user(&info, (void __user *)arg, minsz)) | |
1610 | return -EFAULT; | |
1611 | ||
1612 | if (info.argsz < minsz) | |
1613 | return -EINVAL; | |
1614 | ||
d4f50ee2 | 1615 | info.flags = VFIO_IOMMU_INFO_PGSIZES; |
73fa0d10 | 1616 | |
1ef3e2bc | 1617 | info.iova_pgsizes = vfio_pgsize_bitmap(iommu); |
73fa0d10 | 1618 | |
8160c4e4 MT |
1619 | return copy_to_user((void __user *)arg, &info, minsz) ? |
1620 | -EFAULT : 0; | |
73fa0d10 AW |
1621 | |
1622 | } else if (cmd == VFIO_IOMMU_MAP_DMA) { | |
1623 | struct vfio_iommu_type1_dma_map map; | |
1624 | uint32_t mask = VFIO_DMA_MAP_FLAG_READ | | |
1625 | VFIO_DMA_MAP_FLAG_WRITE; | |
1626 | ||
1627 | minsz = offsetofend(struct vfio_iommu_type1_dma_map, size); | |
1628 | ||
1629 | if (copy_from_user(&map, (void __user *)arg, minsz)) | |
1630 | return -EFAULT; | |
1631 | ||
1632 | if (map.argsz < minsz || map.flags & ~mask) | |
1633 | return -EINVAL; | |
1634 | ||
1635 | return vfio_dma_do_map(iommu, &map); | |
1636 | ||
1637 | } else if (cmd == VFIO_IOMMU_UNMAP_DMA) { | |
1638 | struct vfio_iommu_type1_dma_unmap unmap; | |
166fd7d9 | 1639 | long ret; |
73fa0d10 AW |
1640 | |
1641 | minsz = offsetofend(struct vfio_iommu_type1_dma_unmap, size); | |
1642 | ||
1643 | if (copy_from_user(&unmap, (void __user *)arg, minsz)) | |
1644 | return -EFAULT; | |
1645 | ||
1646 | if (unmap.argsz < minsz || unmap.flags) | |
1647 | return -EINVAL; | |
1648 | ||
166fd7d9 AW |
1649 | ret = vfio_dma_do_unmap(iommu, &unmap); |
1650 | if (ret) | |
1651 | return ret; | |
1652 | ||
8160c4e4 MT |
1653 | return copy_to_user((void __user *)arg, &unmap, minsz) ? |
1654 | -EFAULT : 0; | |
73fa0d10 AW |
1655 | } |
1656 | ||
1657 | return -ENOTTY; | |
1658 | } | |
1659 | ||
c086de81 | 1660 | static int vfio_iommu_type1_register_notifier(void *iommu_data, |
22195cbd | 1661 | unsigned long *events, |
c086de81 KW |
1662 | struct notifier_block *nb) |
1663 | { | |
1664 | struct vfio_iommu *iommu = iommu_data; | |
1665 | ||
22195cbd JS |
1666 | /* clear known events */ |
1667 | *events &= ~VFIO_IOMMU_NOTIFY_DMA_UNMAP; | |
1668 | ||
1669 | /* refuse to register if still events remaining */ | |
1670 | if (*events) | |
1671 | return -EINVAL; | |
1672 | ||
c086de81 KW |
1673 | return blocking_notifier_chain_register(&iommu->notifier, nb); |
1674 | } | |
1675 | ||
1676 | static int vfio_iommu_type1_unregister_notifier(void *iommu_data, | |
1677 | struct notifier_block *nb) | |
1678 | { | |
1679 | struct vfio_iommu *iommu = iommu_data; | |
1680 | ||
1681 | return blocking_notifier_chain_unregister(&iommu->notifier, nb); | |
1682 | } | |
1683 | ||
73fa0d10 | 1684 | static const struct vfio_iommu_driver_ops vfio_iommu_driver_ops_type1 = { |
c086de81 KW |
1685 | .name = "vfio-iommu-type1", |
1686 | .owner = THIS_MODULE, | |
1687 | .open = vfio_iommu_type1_open, | |
1688 | .release = vfio_iommu_type1_release, | |
1689 | .ioctl = vfio_iommu_type1_ioctl, | |
1690 | .attach_group = vfio_iommu_type1_attach_group, | |
1691 | .detach_group = vfio_iommu_type1_detach_group, | |
1692 | .pin_pages = vfio_iommu_type1_pin_pages, | |
1693 | .unpin_pages = vfio_iommu_type1_unpin_pages, | |
1694 | .register_notifier = vfio_iommu_type1_register_notifier, | |
1695 | .unregister_notifier = vfio_iommu_type1_unregister_notifier, | |
73fa0d10 AW |
1696 | }; |
1697 | ||
1698 | static int __init vfio_iommu_type1_init(void) | |
1699 | { | |
73fa0d10 AW |
1700 | return vfio_register_iommu_driver(&vfio_iommu_driver_ops_type1); |
1701 | } | |
1702 | ||
1703 | static void __exit vfio_iommu_type1_cleanup(void) | |
1704 | { | |
1705 | vfio_unregister_iommu_driver(&vfio_iommu_driver_ops_type1); | |
1706 | } | |
1707 | ||
1708 | module_init(vfio_iommu_type1_init); | |
1709 | module_exit(vfio_iommu_type1_cleanup); | |
1710 | ||
1711 | MODULE_VERSION(DRIVER_VERSION); | |
1712 | MODULE_LICENSE("GPL v2"); | |
1713 | MODULE_AUTHOR(DRIVER_AUTHOR); | |
1714 | MODULE_DESCRIPTION(DRIVER_DESC); |