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 | ||
6bd06f5a SS |
105 | struct vfio_regions { |
106 | struct list_head list; | |
107 | dma_addr_t iova; | |
108 | phys_addr_t phys; | |
109 | size_t len; | |
110 | }; | |
111 | ||
a54eb550 KW |
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 | ||
0cfef2b7 | 256 | static int vfio_lock_acct(struct task_struct *task, long npage, bool *lock_cap) |
73fa0d10 | 257 | { |
73fa0d10 | 258 | struct mm_struct *mm; |
6c38c055 | 259 | bool is_current; |
0cfef2b7 | 260 | int ret; |
73fa0d10 | 261 | |
3624a248 | 262 | if (!npage) |
0cfef2b7 | 263 | return 0; |
3624a248 | 264 | |
6c38c055 AW |
265 | is_current = (task->mm == current->mm); |
266 | ||
267 | mm = is_current ? task->mm : get_task_mm(task); | |
3624a248 | 268 | if (!mm) |
0cfef2b7 | 269 | return -ESRCH; /* process exited */ |
73fa0d10 | 270 | |
0cfef2b7 AW |
271 | ret = down_write_killable(&mm->mmap_sem); |
272 | if (!ret) { | |
273 | if (npage > 0) { | |
274 | if (lock_cap ? !*lock_cap : | |
275 | !has_capability(task, CAP_IPC_LOCK)) { | |
276 | unsigned long limit; | |
277 | ||
278 | limit = task_rlimit(task, | |
279 | RLIMIT_MEMLOCK) >> PAGE_SHIFT; | |
280 | ||
281 | if (mm->locked_vm + npage > limit) | |
282 | ret = -ENOMEM; | |
283 | } | |
284 | } | |
285 | ||
286 | if (!ret) | |
287 | mm->locked_vm += npage; | |
73fa0d10 | 288 | |
0cfef2b7 | 289 | up_write(&mm->mmap_sem); |
6c38c055 AW |
290 | } |
291 | ||
0cfef2b7 | 292 | if (!is_current) |
3624a248 | 293 | mmput(mm); |
0cfef2b7 AW |
294 | |
295 | return ret; | |
73fa0d10 AW |
296 | } |
297 | ||
298 | /* | |
299 | * Some mappings aren't backed by a struct page, for example an mmap'd | |
300 | * MMIO range for our own or another device. These use a different | |
301 | * pfn conversion and shouldn't be tracked as locked pages. | |
302 | */ | |
303 | static bool is_invalid_reserved_pfn(unsigned long pfn) | |
304 | { | |
305 | if (pfn_valid(pfn)) { | |
306 | bool reserved; | |
307 | struct page *tail = pfn_to_page(pfn); | |
668f9abb | 308 | struct page *head = compound_head(tail); |
73fa0d10 AW |
309 | reserved = !!(PageReserved(head)); |
310 | if (head != tail) { | |
311 | /* | |
312 | * "head" is not a dangling pointer | |
668f9abb | 313 | * (compound_head takes care of that) |
73fa0d10 AW |
314 | * but the hugepage may have been split |
315 | * from under us (and we may not hold a | |
316 | * reference count on the head page so it can | |
317 | * be reused before we run PageReferenced), so | |
318 | * we've to check PageTail before returning | |
319 | * what we just read. | |
320 | */ | |
321 | smp_rmb(); | |
322 | if (PageTail(tail)) | |
323 | return reserved; | |
324 | } | |
325 | return PageReserved(tail); | |
326 | } | |
327 | ||
328 | return true; | |
329 | } | |
330 | ||
331 | static int put_pfn(unsigned long pfn, int prot) | |
332 | { | |
333 | if (!is_invalid_reserved_pfn(pfn)) { | |
334 | struct page *page = pfn_to_page(pfn); | |
335 | if (prot & IOMMU_WRITE) | |
336 | SetPageDirty(page); | |
337 | put_page(page); | |
338 | return 1; | |
339 | } | |
340 | return 0; | |
341 | } | |
342 | ||
ea85cf35 KW |
343 | static int vaddr_get_pfn(struct mm_struct *mm, unsigned long vaddr, |
344 | int prot, unsigned long *pfn) | |
73fa0d10 AW |
345 | { |
346 | struct page *page[1]; | |
347 | struct vm_area_struct *vma; | |
94db151d | 348 | struct vm_area_struct *vmas[1]; |
ea85cf35 | 349 | int ret; |
73fa0d10 | 350 | |
ea85cf35 | 351 | if (mm == current->mm) { |
94db151d DW |
352 | ret = get_user_pages_longterm(vaddr, 1, !!(prot & IOMMU_WRITE), |
353 | page, vmas); | |
ea85cf35 KW |
354 | } else { |
355 | unsigned int flags = 0; | |
356 | ||
357 | if (prot & IOMMU_WRITE) | |
358 | flags |= FOLL_WRITE; | |
359 | ||
360 | down_read(&mm->mmap_sem); | |
361 | ret = get_user_pages_remote(NULL, mm, vaddr, 1, flags, page, | |
94db151d DW |
362 | vmas, NULL); |
363 | /* | |
364 | * The lifetime of a vaddr_get_pfn() page pin is | |
365 | * userspace-controlled. In the fs-dax case this could | |
366 | * lead to indefinite stalls in filesystem operations. | |
367 | * Disallow attempts to pin fs-dax pages via this | |
368 | * interface. | |
369 | */ | |
370 | if (ret > 0 && vma_is_fsdax(vmas[0])) { | |
371 | ret = -EOPNOTSUPP; | |
372 | put_page(page[0]); | |
373 | } | |
ea85cf35 KW |
374 | up_read(&mm->mmap_sem); |
375 | } | |
376 | ||
377 | if (ret == 1) { | |
73fa0d10 AW |
378 | *pfn = page_to_pfn(page[0]); |
379 | return 0; | |
380 | } | |
381 | ||
ea85cf35 | 382 | down_read(&mm->mmap_sem); |
73fa0d10 | 383 | |
ea85cf35 | 384 | vma = find_vma_intersection(mm, vaddr, vaddr + 1); |
73fa0d10 AW |
385 | |
386 | if (vma && vma->vm_flags & VM_PFNMAP) { | |
387 | *pfn = ((vaddr - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff; | |
388 | if (is_invalid_reserved_pfn(*pfn)) | |
389 | ret = 0; | |
390 | } | |
391 | ||
ea85cf35 | 392 | up_read(&mm->mmap_sem); |
73fa0d10 AW |
393 | return ret; |
394 | } | |
395 | ||
166fd7d9 AW |
396 | /* |
397 | * Attempt to pin pages. We really don't want to track all the pfns and | |
398 | * the iommu can only map chunks of consecutive pfns anyway, so get the | |
399 | * first page and all consecutive pages with the same locking. | |
400 | */ | |
8f0d5bb9 | 401 | static long vfio_pin_pages_remote(struct vfio_dma *dma, unsigned long vaddr, |
7cb671e7 AW |
402 | long npage, unsigned long *pfn_base, |
403 | bool lock_cap, unsigned long limit) | |
73fa0d10 | 404 | { |
7cb671e7 | 405 | unsigned long pfn = 0; |
6c38c055 | 406 | long ret, pinned = 0, lock_acct = 0; |
a54eb550 | 407 | dma_addr_t iova = vaddr - dma->vaddr + dma->iova; |
73fa0d10 | 408 | |
6c38c055 AW |
409 | /* This code path is only user initiated */ |
410 | if (!current->mm) | |
166fd7d9 | 411 | return -ENODEV; |
73fa0d10 | 412 | |
6c38c055 | 413 | ret = vaddr_get_pfn(current->mm, vaddr, dma->prot, pfn_base); |
166fd7d9 | 414 | if (ret) |
6c38c055 | 415 | return ret; |
73fa0d10 | 416 | |
356e88eb JCXF |
417 | if (is_invalid_reserved_pfn(*pfn_base)) { |
418 | struct vm_area_struct *vma; | |
419 | ||
420 | down_read(¤t->mm->mmap_sem); | |
421 | vma = find_vma_intersection(current->mm, vaddr, vaddr + 1); | |
422 | pinned = min_t(long, npage, vma_pages(vma)); | |
423 | up_read(¤t->mm->mmap_sem); | |
424 | return pinned; | |
425 | } | |
426 | ||
6c38c055 | 427 | pinned++; |
73fa0d10 | 428 | |
a54eb550 KW |
429 | /* |
430 | * Reserved pages aren't counted against the user, externally pinned | |
431 | * pages are already counted against the user. | |
432 | */ | |
356e88eb | 433 | if (!vfio_find_vpfn(dma, iova)) { |
6c38c055 | 434 | if (!lock_cap && current->mm->locked_vm + 1 > limit) { |
a54eb550 KW |
435 | put_pfn(*pfn_base, dma->prot); |
436 | pr_warn("%s: RLIMIT_MEMLOCK (%ld) exceeded\n", __func__, | |
437 | limit << PAGE_SHIFT); | |
6c38c055 | 438 | return -ENOMEM; |
a54eb550 KW |
439 | } |
440 | lock_acct++; | |
5c6c2b21 AW |
441 | } |
442 | ||
6c38c055 AW |
443 | if (unlikely(disable_hugepages)) |
444 | goto out; | |
73fa0d10 | 445 | |
6c38c055 AW |
446 | /* Lock all the consecutive pages from pfn_base */ |
447 | for (vaddr += PAGE_SIZE, iova += PAGE_SIZE; pinned < npage; | |
448 | pinned++, vaddr += PAGE_SIZE, iova += PAGE_SIZE) { | |
6c38c055 AW |
449 | ret = vaddr_get_pfn(current->mm, vaddr, dma->prot, &pfn); |
450 | if (ret) | |
451 | break; | |
452 | ||
356e88eb | 453 | if (pfn != *pfn_base + pinned) { |
6c38c055 AW |
454 | put_pfn(pfn, dma->prot); |
455 | break; | |
456 | } | |
166fd7d9 | 457 | |
356e88eb | 458 | if (!vfio_find_vpfn(dma, iova)) { |
6c38c055 AW |
459 | if (!lock_cap && |
460 | current->mm->locked_vm + lock_acct + 1 > limit) { | |
a54eb550 | 461 | put_pfn(pfn, dma->prot); |
6c38c055 AW |
462 | pr_warn("%s: RLIMIT_MEMLOCK (%ld) exceeded\n", |
463 | __func__, limit << PAGE_SHIFT); | |
0cfef2b7 AW |
464 | ret = -ENOMEM; |
465 | goto unpin_out; | |
a54eb550 | 466 | } |
6c38c055 | 467 | lock_acct++; |
166fd7d9 AW |
468 | } |
469 | } | |
470 | ||
6c38c055 | 471 | out: |
0cfef2b7 AW |
472 | ret = vfio_lock_acct(current, lock_acct, &lock_cap); |
473 | ||
474 | unpin_out: | |
475 | if (ret) { | |
356e88eb JCXF |
476 | for (pfn = *pfn_base ; pinned ; pfn++, pinned--) |
477 | put_pfn(pfn, dma->prot); | |
0cfef2b7 AW |
478 | |
479 | return ret; | |
480 | } | |
166fd7d9 | 481 | |
6c38c055 | 482 | return pinned; |
166fd7d9 AW |
483 | } |
484 | ||
a54eb550 KW |
485 | static long vfio_unpin_pages_remote(struct vfio_dma *dma, dma_addr_t iova, |
486 | unsigned long pfn, long npage, | |
487 | bool do_accounting) | |
166fd7d9 | 488 | { |
a54eb550 | 489 | long unlocked = 0, locked = 0; |
166fd7d9 AW |
490 | long i; |
491 | ||
6c38c055 | 492 | for (i = 0; i < npage; i++, iova += PAGE_SIZE) { |
a54eb550 KW |
493 | if (put_pfn(pfn++, dma->prot)) { |
494 | unlocked++; | |
6c38c055 | 495 | if (vfio_find_vpfn(dma, iova)) |
a54eb550 KW |
496 | locked++; |
497 | } | |
498 | } | |
499 | ||
500 | if (do_accounting) | |
0cfef2b7 | 501 | vfio_lock_acct(dma->task, locked - unlocked, NULL); |
a54eb550 KW |
502 | |
503 | return unlocked; | |
504 | } | |
505 | ||
506 | static int vfio_pin_page_external(struct vfio_dma *dma, unsigned long vaddr, | |
507 | unsigned long *pfn_base, bool do_accounting) | |
508 | { | |
a54eb550 KW |
509 | struct mm_struct *mm; |
510 | int ret; | |
a54eb550 KW |
511 | |
512 | mm = get_task_mm(dma->task); | |
513 | if (!mm) | |
514 | return -ENODEV; | |
515 | ||
516 | ret = vaddr_get_pfn(mm, vaddr, dma->prot, pfn_base); | |
80dbe1fb AW |
517 | if (!ret && do_accounting && !is_invalid_reserved_pfn(*pfn_base)) { |
518 | ret = vfio_lock_acct(dma->task, 1, NULL); | |
0cfef2b7 AW |
519 | if (ret) { |
520 | put_pfn(*pfn_base, dma->prot); | |
80dbe1fb AW |
521 | if (ret == -ENOMEM) |
522 | pr_warn("%s: Task %s (%d) RLIMIT_MEMLOCK " | |
523 | "(%ld) exceeded\n", __func__, | |
524 | dma->task->comm, task_pid_nr(dma->task), | |
525 | task_rlimit(dma->task, RLIMIT_MEMLOCK)); | |
0cfef2b7 AW |
526 | } |
527 | } | |
528 | ||
a54eb550 KW |
529 | mmput(mm); |
530 | return ret; | |
531 | } | |
532 | ||
533 | static int vfio_unpin_page_external(struct vfio_dma *dma, dma_addr_t iova, | |
534 | bool do_accounting) | |
535 | { | |
536 | int unlocked; | |
537 | struct vfio_pfn *vpfn = vfio_find_vpfn(dma, iova); | |
538 | ||
539 | if (!vpfn) | |
540 | return 0; | |
541 | ||
542 | unlocked = vfio_iova_put_vfio_pfn(dma, vpfn); | |
166fd7d9 AW |
543 | |
544 | if (do_accounting) | |
0cfef2b7 | 545 | vfio_lock_acct(dma->task, -unlocked, NULL); |
166fd7d9 AW |
546 | |
547 | return unlocked; | |
548 | } | |
549 | ||
a54eb550 KW |
550 | static int vfio_iommu_type1_pin_pages(void *iommu_data, |
551 | unsigned long *user_pfn, | |
552 | int npage, int prot, | |
553 | unsigned long *phys_pfn) | |
554 | { | |
555 | struct vfio_iommu *iommu = iommu_data; | |
556 | int i, j, ret; | |
557 | unsigned long remote_vaddr; | |
558 | struct vfio_dma *dma; | |
559 | bool do_accounting; | |
560 | ||
561 | if (!iommu || !user_pfn || !phys_pfn) | |
562 | return -EINVAL; | |
563 | ||
564 | /* Supported for v2 version only */ | |
565 | if (!iommu->v2) | |
566 | return -EACCES; | |
567 | ||
568 | mutex_lock(&iommu->lock); | |
569 | ||
c086de81 KW |
570 | /* Fail if notifier list is empty */ |
571 | if ((!iommu->external_domain) || (!iommu->notifier.head)) { | |
a54eb550 KW |
572 | ret = -EINVAL; |
573 | goto pin_done; | |
574 | } | |
575 | ||
576 | /* | |
577 | * If iommu capable domain exist in the container then all pages are | |
578 | * already pinned and accounted. Accouting should be done if there is no | |
579 | * iommu capable domain in the container. | |
580 | */ | |
581 | do_accounting = !IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu); | |
582 | ||
583 | for (i = 0; i < npage; i++) { | |
584 | dma_addr_t iova; | |
585 | struct vfio_pfn *vpfn; | |
586 | ||
587 | iova = user_pfn[i] << PAGE_SHIFT; | |
2b8bb1d7 | 588 | dma = vfio_find_dma(iommu, iova, PAGE_SIZE); |
a54eb550 KW |
589 | if (!dma) { |
590 | ret = -EINVAL; | |
591 | goto pin_unwind; | |
592 | } | |
593 | ||
594 | if ((dma->prot & prot) != prot) { | |
595 | ret = -EPERM; | |
596 | goto pin_unwind; | |
597 | } | |
598 | ||
599 | vpfn = vfio_iova_get_vfio_pfn(dma, iova); | |
600 | if (vpfn) { | |
601 | phys_pfn[i] = vpfn->pfn; | |
602 | continue; | |
603 | } | |
604 | ||
605 | remote_vaddr = dma->vaddr + iova - dma->iova; | |
606 | ret = vfio_pin_page_external(dma, remote_vaddr, &phys_pfn[i], | |
607 | do_accounting); | |
80dbe1fb | 608 | if (ret) |
a54eb550 | 609 | goto pin_unwind; |
a54eb550 KW |
610 | |
611 | ret = vfio_add_to_pfn_list(dma, iova, phys_pfn[i]); | |
612 | if (ret) { | |
613 | vfio_unpin_page_external(dma, iova, do_accounting); | |
614 | goto pin_unwind; | |
615 | } | |
616 | } | |
617 | ||
618 | ret = i; | |
619 | goto pin_done; | |
620 | ||
621 | pin_unwind: | |
622 | phys_pfn[i] = 0; | |
623 | for (j = 0; j < i; j++) { | |
624 | dma_addr_t iova; | |
625 | ||
626 | iova = user_pfn[j] << PAGE_SHIFT; | |
2b8bb1d7 | 627 | dma = vfio_find_dma(iommu, iova, PAGE_SIZE); |
a54eb550 KW |
628 | vfio_unpin_page_external(dma, iova, do_accounting); |
629 | phys_pfn[j] = 0; | |
630 | } | |
631 | pin_done: | |
632 | mutex_unlock(&iommu->lock); | |
633 | return ret; | |
634 | } | |
635 | ||
636 | static int vfio_iommu_type1_unpin_pages(void *iommu_data, | |
637 | unsigned long *user_pfn, | |
638 | int npage) | |
639 | { | |
640 | struct vfio_iommu *iommu = iommu_data; | |
641 | bool do_accounting; | |
642 | int i; | |
643 | ||
644 | if (!iommu || !user_pfn) | |
645 | return -EINVAL; | |
646 | ||
647 | /* Supported for v2 version only */ | |
648 | if (!iommu->v2) | |
649 | return -EACCES; | |
650 | ||
651 | mutex_lock(&iommu->lock); | |
652 | ||
653 | if (!iommu->external_domain) { | |
654 | mutex_unlock(&iommu->lock); | |
655 | return -EINVAL; | |
656 | } | |
657 | ||
658 | do_accounting = !IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu); | |
659 | for (i = 0; i < npage; i++) { | |
660 | struct vfio_dma *dma; | |
661 | dma_addr_t iova; | |
662 | ||
663 | iova = user_pfn[i] << PAGE_SHIFT; | |
2b8bb1d7 | 664 | dma = vfio_find_dma(iommu, iova, PAGE_SIZE); |
a54eb550 KW |
665 | if (!dma) |
666 | goto unpin_exit; | |
667 | vfio_unpin_page_external(dma, iova, do_accounting); | |
668 | } | |
669 | ||
670 | unpin_exit: | |
671 | mutex_unlock(&iommu->lock); | |
672 | return i > npage ? npage : (i > 0 ? i : -EINVAL); | |
673 | } | |
674 | ||
6bd06f5a SS |
675 | static long vfio_sync_unpin(struct vfio_dma *dma, struct vfio_domain *domain, |
676 | struct list_head *regions) | |
677 | { | |
678 | long unlocked = 0; | |
679 | struct vfio_regions *entry, *next; | |
680 | ||
681 | iommu_tlb_sync(domain->domain); | |
682 | ||
683 | list_for_each_entry_safe(entry, next, regions, list) { | |
684 | unlocked += vfio_unpin_pages_remote(dma, | |
685 | entry->iova, | |
686 | entry->phys >> PAGE_SHIFT, | |
687 | entry->len >> PAGE_SHIFT, | |
688 | false); | |
689 | list_del(&entry->list); | |
690 | kfree(entry); | |
691 | } | |
692 | ||
693 | cond_resched(); | |
694 | ||
695 | return unlocked; | |
696 | } | |
697 | ||
698 | /* | |
699 | * Generally, VFIO needs to unpin remote pages after each IOTLB flush. | |
700 | * Therefore, when using IOTLB flush sync interface, VFIO need to keep track | |
701 | * of these regions (currently using a list). | |
702 | * | |
703 | * This value specifies maximum number of regions for each IOTLB flush sync. | |
704 | */ | |
705 | #define VFIO_IOMMU_TLB_SYNC_MAX 512 | |
706 | ||
707 | static size_t unmap_unpin_fast(struct vfio_domain *domain, | |
708 | struct vfio_dma *dma, dma_addr_t *iova, | |
709 | size_t len, phys_addr_t phys, long *unlocked, | |
710 | struct list_head *unmapped_list, | |
711 | int *unmapped_cnt) | |
712 | { | |
713 | size_t unmapped = 0; | |
714 | struct vfio_regions *entry = kzalloc(sizeof(*entry), GFP_KERNEL); | |
715 | ||
716 | if (entry) { | |
717 | unmapped = iommu_unmap_fast(domain->domain, *iova, len); | |
718 | ||
719 | if (!unmapped) { | |
720 | kfree(entry); | |
721 | } else { | |
722 | iommu_tlb_range_add(domain->domain, *iova, unmapped); | |
723 | entry->iova = *iova; | |
724 | entry->phys = phys; | |
725 | entry->len = unmapped; | |
726 | list_add_tail(&entry->list, unmapped_list); | |
727 | ||
728 | *iova += unmapped; | |
729 | (*unmapped_cnt)++; | |
730 | } | |
731 | } | |
732 | ||
733 | /* | |
734 | * Sync if the number of fast-unmap regions hits the limit | |
735 | * or in case of errors. | |
736 | */ | |
737 | if (*unmapped_cnt >= VFIO_IOMMU_TLB_SYNC_MAX || !unmapped) { | |
738 | *unlocked += vfio_sync_unpin(dma, domain, | |
739 | unmapped_list); | |
740 | *unmapped_cnt = 0; | |
741 | } | |
742 | ||
743 | return unmapped; | |
744 | } | |
745 | ||
746 | static size_t unmap_unpin_slow(struct vfio_domain *domain, | |
747 | struct vfio_dma *dma, dma_addr_t *iova, | |
748 | size_t len, phys_addr_t phys, | |
749 | long *unlocked) | |
750 | { | |
751 | size_t unmapped = iommu_unmap(domain->domain, *iova, len); | |
752 | ||
753 | if (unmapped) { | |
754 | *unlocked += vfio_unpin_pages_remote(dma, *iova, | |
755 | phys >> PAGE_SHIFT, | |
756 | unmapped >> PAGE_SHIFT, | |
757 | false); | |
758 | *iova += unmapped; | |
759 | cond_resched(); | |
760 | } | |
761 | return unmapped; | |
762 | } | |
763 | ||
a54eb550 KW |
764 | static long vfio_unmap_unpin(struct vfio_iommu *iommu, struct vfio_dma *dma, |
765 | bool do_accounting) | |
166fd7d9 | 766 | { |
1ef3e2bc AW |
767 | dma_addr_t iova = dma->iova, end = dma->iova + dma->size; |
768 | struct vfio_domain *domain, *d; | |
6bd06f5a SS |
769 | LIST_HEAD(unmapped_region_list); |
770 | int unmapped_region_cnt = 0; | |
166fd7d9 AW |
771 | long unlocked = 0; |
772 | ||
1ef3e2bc | 773 | if (!dma->size) |
a54eb550 KW |
774 | return 0; |
775 | ||
776 | if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu)) | |
777 | return 0; | |
778 | ||
1ef3e2bc AW |
779 | /* |
780 | * We use the IOMMU to track the physical addresses, otherwise we'd | |
781 | * need a much more complicated tracking system. Unfortunately that | |
782 | * means we need to use one of the iommu domains to figure out the | |
783 | * pfns to unpin. The rest need to be unmapped in advance so we have | |
784 | * no iommu translations remaining when the pages are unpinned. | |
785 | */ | |
786 | domain = d = list_first_entry(&iommu->domain_list, | |
787 | struct vfio_domain, next); | |
788 | ||
c5e66887 | 789 | list_for_each_entry_continue(d, &iommu->domain_list, next) { |
1ef3e2bc | 790 | iommu_unmap(d->domain, dma->iova, dma->size); |
c5e66887 AW |
791 | cond_resched(); |
792 | } | |
1ef3e2bc | 793 | |
166fd7d9 | 794 | while (iova < end) { |
6fe1010d AW |
795 | size_t unmapped, len; |
796 | phys_addr_t phys, next; | |
166fd7d9 | 797 | |
1ef3e2bc | 798 | phys = iommu_iova_to_phys(domain->domain, iova); |
166fd7d9 AW |
799 | if (WARN_ON(!phys)) { |
800 | iova += PAGE_SIZE; | |
801 | continue; | |
73fa0d10 | 802 | } |
166fd7d9 | 803 | |
6fe1010d AW |
804 | /* |
805 | * To optimize for fewer iommu_unmap() calls, each of which | |
806 | * may require hardware cache flushing, try to find the | |
807 | * largest contiguous physical memory chunk to unmap. | |
808 | */ | |
809 | for (len = PAGE_SIZE; | |
810 | !domain->fgsp && iova + len < end; len += PAGE_SIZE) { | |
811 | next = iommu_iova_to_phys(domain->domain, iova + len); | |
812 | if (next != phys + len) | |
813 | break; | |
814 | } | |
815 | ||
6bd06f5a SS |
816 | /* |
817 | * First, try to use fast unmap/unpin. In case of failure, | |
818 | * switch to slow unmap/unpin path. | |
819 | */ | |
820 | unmapped = unmap_unpin_fast(domain, dma, &iova, len, phys, | |
821 | &unlocked, &unmapped_region_list, | |
822 | &unmapped_region_cnt); | |
823 | if (!unmapped) { | |
824 | unmapped = unmap_unpin_slow(domain, dma, &iova, len, | |
825 | phys, &unlocked); | |
826 | if (WARN_ON(!unmapped)) | |
827 | break; | |
828 | } | |
73fa0d10 | 829 | } |
166fd7d9 | 830 | |
a54eb550 | 831 | dma->iommu_mapped = false; |
6bd06f5a SS |
832 | |
833 | if (unmapped_region_cnt) | |
834 | unlocked += vfio_sync_unpin(dma, domain, &unmapped_region_list); | |
835 | ||
a54eb550 | 836 | if (do_accounting) { |
0cfef2b7 | 837 | vfio_lock_acct(dma->task, -unlocked, NULL); |
a54eb550 KW |
838 | return 0; |
839 | } | |
840 | return unlocked; | |
73fa0d10 AW |
841 | } |
842 | ||
1ef3e2bc | 843 | static void vfio_remove_dma(struct vfio_iommu *iommu, struct vfio_dma *dma) |
73fa0d10 | 844 | { |
a54eb550 | 845 | vfio_unmap_unpin(iommu, dma, true); |
1ef3e2bc | 846 | vfio_unlink_dma(iommu, dma); |
8f0d5bb9 | 847 | put_task_struct(dma->task); |
1ef3e2bc AW |
848 | kfree(dma); |
849 | } | |
73fa0d10 | 850 | |
1ef3e2bc AW |
851 | static unsigned long vfio_pgsize_bitmap(struct vfio_iommu *iommu) |
852 | { | |
853 | struct vfio_domain *domain; | |
4644321f | 854 | unsigned long bitmap = ULONG_MAX; |
166fd7d9 | 855 | |
1ef3e2bc AW |
856 | mutex_lock(&iommu->lock); |
857 | list_for_each_entry(domain, &iommu->domain_list, next) | |
d16e0faa | 858 | bitmap &= domain->domain->pgsize_bitmap; |
1ef3e2bc | 859 | mutex_unlock(&iommu->lock); |
73fa0d10 | 860 | |
4644321f EA |
861 | /* |
862 | * In case the IOMMU supports page sizes smaller than PAGE_SIZE | |
863 | * we pretend PAGE_SIZE is supported and hide sub-PAGE_SIZE sizes. | |
864 | * That way the user will be able to map/unmap buffers whose size/ | |
865 | * start address is aligned with PAGE_SIZE. Pinning code uses that | |
866 | * granularity while iommu driver can use the sub-PAGE_SIZE size | |
867 | * to map the buffer. | |
868 | */ | |
869 | if (bitmap & ~PAGE_MASK) { | |
870 | bitmap &= PAGE_MASK; | |
871 | bitmap |= PAGE_SIZE; | |
872 | } | |
873 | ||
1ef3e2bc | 874 | return bitmap; |
73fa0d10 AW |
875 | } |
876 | ||
877 | static int vfio_dma_do_unmap(struct vfio_iommu *iommu, | |
878 | struct vfio_iommu_type1_dma_unmap *unmap) | |
879 | { | |
73fa0d10 | 880 | uint64_t mask; |
c086de81 | 881 | struct vfio_dma *dma, *dma_last = NULL; |
1ef3e2bc | 882 | size_t unmapped = 0; |
c086de81 | 883 | int ret = 0, retries = 0; |
73fa0d10 | 884 | |
1ef3e2bc | 885 | mask = ((uint64_t)1 << __ffs(vfio_pgsize_bitmap(iommu))) - 1; |
73fa0d10 AW |
886 | |
887 | if (unmap->iova & mask) | |
888 | return -EINVAL; | |
f5bfdbf2 | 889 | if (!unmap->size || unmap->size & mask) |
73fa0d10 | 890 | return -EINVAL; |
71a7d3d7 DC |
891 | if (unmap->iova + unmap->size < unmap->iova || |
892 | unmap->size > SIZE_MAX) | |
893 | return -EINVAL; | |
73fa0d10 | 894 | |
73fa0d10 | 895 | WARN_ON(mask & PAGE_MASK); |
c086de81 | 896 | again: |
73fa0d10 AW |
897 | mutex_lock(&iommu->lock); |
898 | ||
1ef3e2bc AW |
899 | /* |
900 | * vfio-iommu-type1 (v1) - User mappings were coalesced together to | |
901 | * avoid tracking individual mappings. This means that the granularity | |
902 | * of the original mapping was lost and the user was allowed to attempt | |
903 | * to unmap any range. Depending on the contiguousness of physical | |
904 | * memory and page sizes supported by the IOMMU, arbitrary unmaps may | |
905 | * or may not have worked. We only guaranteed unmap granularity | |
906 | * matching the original mapping; even though it was untracked here, | |
907 | * the original mappings are reflected in IOMMU mappings. This | |
908 | * resulted in a couple unusual behaviors. First, if a range is not | |
909 | * able to be unmapped, ex. a set of 4k pages that was mapped as a | |
910 | * 2M hugepage into the IOMMU, the unmap ioctl returns success but with | |
911 | * a zero sized unmap. Also, if an unmap request overlaps the first | |
912 | * address of a hugepage, the IOMMU will unmap the entire hugepage. | |
913 | * This also returns success and the returned unmap size reflects the | |
914 | * actual size unmapped. | |
915 | * | |
916 | * We attempt to maintain compatibility with this "v1" interface, but | |
917 | * we take control out of the hands of the IOMMU. Therefore, an unmap | |
918 | * request offset from the beginning of the original mapping will | |
919 | * return success with zero sized unmap. And an unmap request covering | |
920 | * the first iova of mapping will unmap the entire range. | |
921 | * | |
922 | * The v2 version of this interface intends to be more deterministic. | |
923 | * Unmap requests must fully cover previous mappings. Multiple | |
924 | * mappings may still be unmaped by specifying large ranges, but there | |
925 | * must not be any previous mappings bisected by the range. An error | |
926 | * will be returned if these conditions are not met. The v2 interface | |
927 | * will only return success and a size of zero if there were no | |
928 | * mappings within the range. | |
929 | */ | |
930 | if (iommu->v2) { | |
7c03f428 | 931 | dma = vfio_find_dma(iommu, unmap->iova, 1); |
1ef3e2bc AW |
932 | if (dma && dma->iova != unmap->iova) { |
933 | ret = -EINVAL; | |
934 | goto unlock; | |
935 | } | |
936 | dma = vfio_find_dma(iommu, unmap->iova + unmap->size - 1, 0); | |
937 | if (dma && dma->iova + dma->size != unmap->iova + unmap->size) { | |
938 | ret = -EINVAL; | |
939 | goto unlock; | |
940 | } | |
941 | } | |
942 | ||
166fd7d9 | 943 | while ((dma = vfio_find_dma(iommu, unmap->iova, unmap->size))) { |
1ef3e2bc | 944 | if (!iommu->v2 && unmap->iova > dma->iova) |
166fd7d9 | 945 | break; |
8f0d5bb9 KW |
946 | /* |
947 | * Task with same address space who mapped this iova range is | |
948 | * allowed to unmap the iova range. | |
949 | */ | |
950 | if (dma->task->mm != current->mm) | |
951 | break; | |
c086de81 KW |
952 | |
953 | if (!RB_EMPTY_ROOT(&dma->pfn_list)) { | |
954 | struct vfio_iommu_type1_dma_unmap nb_unmap; | |
955 | ||
956 | if (dma_last == dma) { | |
957 | BUG_ON(++retries > 10); | |
958 | } else { | |
959 | dma_last = dma; | |
960 | retries = 0; | |
961 | } | |
962 | ||
963 | nb_unmap.iova = dma->iova; | |
964 | nb_unmap.size = dma->size; | |
965 | ||
966 | /* | |
967 | * Notify anyone (mdev vendor drivers) to invalidate and | |
968 | * unmap iovas within the range we're about to unmap. | |
969 | * Vendor drivers MUST unpin pages in response to an | |
970 | * invalidation. | |
971 | */ | |
972 | mutex_unlock(&iommu->lock); | |
973 | blocking_notifier_call_chain(&iommu->notifier, | |
974 | VFIO_IOMMU_NOTIFY_DMA_UNMAP, | |
975 | &nb_unmap); | |
976 | goto again; | |
977 | } | |
1ef3e2bc AW |
978 | unmapped += dma->size; |
979 | vfio_remove_dma(iommu, dma); | |
166fd7d9 | 980 | } |
cd9b2268 | 981 | |
1ef3e2bc | 982 | unlock: |
73fa0d10 | 983 | mutex_unlock(&iommu->lock); |
166fd7d9 | 984 | |
1ef3e2bc | 985 | /* Report how much was unmapped */ |
166fd7d9 AW |
986 | unmap->size = unmapped; |
987 | ||
988 | return ret; | |
989 | } | |
990 | ||
991 | /* | |
992 | * Turns out AMD IOMMU has a page table bug where it won't map large pages | |
993 | * to a region that previously mapped smaller pages. This should be fixed | |
994 | * soon, so this is just a temporary workaround to break mappings down into | |
995 | * PAGE_SIZE. Better to map smaller pages than nothing. | |
996 | */ | |
1ef3e2bc | 997 | static int map_try_harder(struct vfio_domain *domain, dma_addr_t iova, |
166fd7d9 AW |
998 | unsigned long pfn, long npage, int prot) |
999 | { | |
1000 | long i; | |
089f1c6b | 1001 | int ret = 0; |
166fd7d9 AW |
1002 | |
1003 | for (i = 0; i < npage; i++, pfn++, iova += PAGE_SIZE) { | |
1ef3e2bc | 1004 | ret = iommu_map(domain->domain, iova, |
166fd7d9 | 1005 | (phys_addr_t)pfn << PAGE_SHIFT, |
1ef3e2bc | 1006 | PAGE_SIZE, prot | domain->prot); |
166fd7d9 AW |
1007 | if (ret) |
1008 | break; | |
1009 | } | |
1010 | ||
1011 | for (; i < npage && i > 0; i--, iova -= PAGE_SIZE) | |
1ef3e2bc AW |
1012 | iommu_unmap(domain->domain, iova, PAGE_SIZE); |
1013 | ||
1014 | return ret; | |
1015 | } | |
1016 | ||
1017 | static int vfio_iommu_map(struct vfio_iommu *iommu, dma_addr_t iova, | |
1018 | unsigned long pfn, long npage, int prot) | |
1019 | { | |
1020 | struct vfio_domain *d; | |
1021 | int ret; | |
1022 | ||
1023 | list_for_each_entry(d, &iommu->domain_list, next) { | |
1024 | ret = iommu_map(d->domain, iova, (phys_addr_t)pfn << PAGE_SHIFT, | |
1025 | npage << PAGE_SHIFT, prot | d->prot); | |
1026 | if (ret) { | |
1027 | if (ret != -EBUSY || | |
1028 | map_try_harder(d, iova, pfn, npage, prot)) | |
1029 | goto unwind; | |
1030 | } | |
c5e66887 AW |
1031 | |
1032 | cond_resched(); | |
1ef3e2bc AW |
1033 | } |
1034 | ||
1035 | return 0; | |
1036 | ||
1037 | unwind: | |
1038 | list_for_each_entry_continue_reverse(d, &iommu->domain_list, next) | |
1039 | iommu_unmap(d->domain, iova, npage << PAGE_SHIFT); | |
166fd7d9 | 1040 | |
cd9b2268 | 1041 | return ret; |
73fa0d10 AW |
1042 | } |
1043 | ||
8f0d5bb9 KW |
1044 | static int vfio_pin_map_dma(struct vfio_iommu *iommu, struct vfio_dma *dma, |
1045 | size_t map_size) | |
1046 | { | |
1047 | dma_addr_t iova = dma->iova; | |
1048 | unsigned long vaddr = dma->vaddr; | |
1049 | size_t size = map_size; | |
1050 | long npage; | |
7cb671e7 AW |
1051 | unsigned long pfn, limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT; |
1052 | bool lock_cap = capable(CAP_IPC_LOCK); | |
8f0d5bb9 KW |
1053 | int ret = 0; |
1054 | ||
1055 | while (size) { | |
1056 | /* Pin a contiguous chunk of memory */ | |
1057 | npage = vfio_pin_pages_remote(dma, vaddr + dma->size, | |
7cb671e7 AW |
1058 | size >> PAGE_SHIFT, &pfn, |
1059 | lock_cap, limit); | |
8f0d5bb9 KW |
1060 | if (npage <= 0) { |
1061 | WARN_ON(!npage); | |
1062 | ret = (int)npage; | |
1063 | break; | |
1064 | } | |
1065 | ||
1066 | /* Map it! */ | |
1067 | ret = vfio_iommu_map(iommu, iova + dma->size, pfn, npage, | |
1068 | dma->prot); | |
1069 | if (ret) { | |
a54eb550 KW |
1070 | vfio_unpin_pages_remote(dma, iova + dma->size, pfn, |
1071 | npage, true); | |
8f0d5bb9 KW |
1072 | break; |
1073 | } | |
1074 | ||
1075 | size -= npage << PAGE_SHIFT; | |
1076 | dma->size += npage << PAGE_SHIFT; | |
1077 | } | |
1078 | ||
a54eb550 KW |
1079 | dma->iommu_mapped = true; |
1080 | ||
8f0d5bb9 KW |
1081 | if (ret) |
1082 | vfio_remove_dma(iommu, dma); | |
1083 | ||
1084 | return ret; | |
1085 | } | |
1086 | ||
73fa0d10 AW |
1087 | static int vfio_dma_do_map(struct vfio_iommu *iommu, |
1088 | struct vfio_iommu_type1_dma_map *map) | |
1089 | { | |
c8dbca16 | 1090 | dma_addr_t iova = map->iova; |
166fd7d9 | 1091 | unsigned long vaddr = map->vaddr; |
73fa0d10 AW |
1092 | size_t size = map->size; |
1093 | int ret = 0, prot = 0; | |
1094 | uint64_t mask; | |
1ef3e2bc | 1095 | struct vfio_dma *dma; |
166fd7d9 | 1096 | |
c8dbca16 AW |
1097 | /* Verify that none of our __u64 fields overflow */ |
1098 | if (map->size != size || map->vaddr != vaddr || map->iova != iova) | |
1099 | return -EINVAL; | |
73fa0d10 | 1100 | |
1ef3e2bc | 1101 | mask = ((uint64_t)1 << __ffs(vfio_pgsize_bitmap(iommu))) - 1; |
73fa0d10 | 1102 | |
c8dbca16 AW |
1103 | WARN_ON(mask & PAGE_MASK); |
1104 | ||
73fa0d10 AW |
1105 | /* READ/WRITE from device perspective */ |
1106 | if (map->flags & VFIO_DMA_MAP_FLAG_WRITE) | |
1107 | prot |= IOMMU_WRITE; | |
1108 | if (map->flags & VFIO_DMA_MAP_FLAG_READ) | |
1109 | prot |= IOMMU_READ; | |
1110 | ||
c8dbca16 | 1111 | if (!prot || !size || (size | iova | vaddr) & mask) |
73fa0d10 AW |
1112 | return -EINVAL; |
1113 | ||
c8dbca16 AW |
1114 | /* Don't allow IOVA or virtual address wrap */ |
1115 | if (iova + size - 1 < iova || vaddr + size - 1 < vaddr) | |
73fa0d10 AW |
1116 | return -EINVAL; |
1117 | ||
1118 | mutex_lock(&iommu->lock); | |
1119 | ||
c8dbca16 | 1120 | if (vfio_find_dma(iommu, iova, size)) { |
8f0d5bb9 KW |
1121 | ret = -EEXIST; |
1122 | goto out_unlock; | |
73fa0d10 AW |
1123 | } |
1124 | ||
1ef3e2bc AW |
1125 | dma = kzalloc(sizeof(*dma), GFP_KERNEL); |
1126 | if (!dma) { | |
8f0d5bb9 KW |
1127 | ret = -ENOMEM; |
1128 | goto out_unlock; | |
1ef3e2bc AW |
1129 | } |
1130 | ||
c8dbca16 AW |
1131 | dma->iova = iova; |
1132 | dma->vaddr = vaddr; | |
1ef3e2bc | 1133 | dma->prot = prot; |
8f0d5bb9 KW |
1134 | get_task_struct(current); |
1135 | dma->task = current; | |
a54eb550 | 1136 | dma->pfn_list = RB_ROOT; |
166fd7d9 | 1137 | |
1ef3e2bc AW |
1138 | /* Insert zero-sized and grow as we map chunks of it */ |
1139 | vfio_link_dma(iommu, dma); | |
166fd7d9 | 1140 | |
a54eb550 KW |
1141 | /* Don't pin and map if container doesn't contain IOMMU capable domain*/ |
1142 | if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu)) | |
1143 | dma->size = size; | |
1144 | else | |
1145 | ret = vfio_pin_map_dma(iommu, dma, size); | |
1146 | ||
8f0d5bb9 | 1147 | out_unlock: |
1ef3e2bc AW |
1148 | mutex_unlock(&iommu->lock); |
1149 | return ret; | |
1150 | } | |
1151 | ||
1152 | static int vfio_bus_type(struct device *dev, void *data) | |
1153 | { | |
1154 | struct bus_type **bus = data; | |
1155 | ||
1156 | if (*bus && *bus != dev->bus) | |
1157 | return -EINVAL; | |
1158 | ||
1159 | *bus = dev->bus; | |
1160 | ||
1161 | return 0; | |
1162 | } | |
1163 | ||
1164 | static int vfio_iommu_replay(struct vfio_iommu *iommu, | |
1165 | struct vfio_domain *domain) | |
1166 | { | |
1167 | struct vfio_domain *d; | |
1168 | struct rb_node *n; | |
7cb671e7 AW |
1169 | unsigned long limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT; |
1170 | bool lock_cap = capable(CAP_IPC_LOCK); | |
1ef3e2bc AW |
1171 | int ret; |
1172 | ||
1173 | /* Arbitrarily pick the first domain in the list for lookups */ | |
1174 | d = list_first_entry(&iommu->domain_list, struct vfio_domain, next); | |
1175 | n = rb_first(&iommu->dma_list); | |
1176 | ||
1ef3e2bc AW |
1177 | for (; n; n = rb_next(n)) { |
1178 | struct vfio_dma *dma; | |
1179 | dma_addr_t iova; | |
1180 | ||
1181 | dma = rb_entry(n, struct vfio_dma, node); | |
1182 | iova = dma->iova; | |
1183 | ||
1184 | while (iova < dma->iova + dma->size) { | |
a54eb550 | 1185 | phys_addr_t phys; |
1ef3e2bc | 1186 | size_t size; |
73fa0d10 | 1187 | |
a54eb550 KW |
1188 | if (dma->iommu_mapped) { |
1189 | phys_addr_t p; | |
1190 | dma_addr_t i; | |
1191 | ||
1192 | phys = iommu_iova_to_phys(d->domain, iova); | |
1193 | ||
1194 | if (WARN_ON(!phys)) { | |
1195 | iova += PAGE_SIZE; | |
1196 | continue; | |
1197 | } | |
1198 | ||
1199 | size = PAGE_SIZE; | |
1200 | p = phys + size; | |
1201 | i = iova + size; | |
1202 | while (i < dma->iova + dma->size && | |
1203 | p == iommu_iova_to_phys(d->domain, i)) { | |
1204 | size += PAGE_SIZE; | |
1205 | p += PAGE_SIZE; | |
1206 | i += PAGE_SIZE; | |
1207 | } | |
1208 | } else { | |
1209 | unsigned long pfn; | |
1210 | unsigned long vaddr = dma->vaddr + | |
1211 | (iova - dma->iova); | |
1212 | size_t n = dma->iova + dma->size - iova; | |
1213 | long npage; | |
1214 | ||
1215 | npage = vfio_pin_pages_remote(dma, vaddr, | |
1216 | n >> PAGE_SHIFT, | |
7cb671e7 AW |
1217 | &pfn, lock_cap, |
1218 | limit); | |
a54eb550 KW |
1219 | if (npage <= 0) { |
1220 | WARN_ON(!npage); | |
1221 | ret = (int)npage; | |
1222 | return ret; | |
1223 | } | |
1224 | ||
1225 | phys = pfn << PAGE_SHIFT; | |
1226 | size = npage << PAGE_SHIFT; | |
166fd7d9 AW |
1227 | } |
1228 | ||
1ef3e2bc AW |
1229 | ret = iommu_map(domain->domain, iova, phys, |
1230 | size, dma->prot | domain->prot); | |
1231 | if (ret) | |
1232 | return ret; | |
d93b3ac0 | 1233 | |
1ef3e2bc AW |
1234 | iova += size; |
1235 | } | |
a54eb550 | 1236 | dma->iommu_mapped = true; |
166fd7d9 | 1237 | } |
1ef3e2bc | 1238 | return 0; |
73fa0d10 AW |
1239 | } |
1240 | ||
6fe1010d AW |
1241 | /* |
1242 | * We change our unmap behavior slightly depending on whether the IOMMU | |
1243 | * supports fine-grained superpages. IOMMUs like AMD-Vi will use a superpage | |
1244 | * for practically any contiguous power-of-two mapping we give it. This means | |
1245 | * we don't need to look for contiguous chunks ourselves to make unmapping | |
1246 | * more efficient. On IOMMUs with coarse-grained super pages, like Intel VT-d | |
1247 | * with discrete 2M/1G/512G/1T superpages, identifying contiguous chunks | |
1248 | * significantly boosts non-hugetlbfs mappings and doesn't seem to hurt when | |
1249 | * hugetlbfs is in use. | |
1250 | */ | |
1251 | static void vfio_test_domain_fgsp(struct vfio_domain *domain) | |
1252 | { | |
1253 | struct page *pages; | |
1254 | int ret, order = get_order(PAGE_SIZE * 2); | |
1255 | ||
1256 | pages = alloc_pages(GFP_KERNEL | __GFP_ZERO, order); | |
1257 | if (!pages) | |
1258 | return; | |
1259 | ||
1260 | ret = iommu_map(domain->domain, 0, page_to_phys(pages), PAGE_SIZE * 2, | |
1261 | IOMMU_READ | IOMMU_WRITE | domain->prot); | |
1262 | if (!ret) { | |
1263 | size_t unmapped = iommu_unmap(domain->domain, 0, PAGE_SIZE); | |
1264 | ||
1265 | if (unmapped == PAGE_SIZE) | |
1266 | iommu_unmap(domain->domain, PAGE_SIZE, PAGE_SIZE); | |
1267 | else | |
1268 | domain->fgsp = true; | |
1269 | } | |
1270 | ||
1271 | __free_pages(pages, order); | |
1272 | } | |
1273 | ||
7896c998 KW |
1274 | static struct vfio_group *find_iommu_group(struct vfio_domain *domain, |
1275 | struct iommu_group *iommu_group) | |
1276 | { | |
1277 | struct vfio_group *g; | |
1278 | ||
1279 | list_for_each_entry(g, &domain->group_list, next) { | |
1280 | if (g->iommu_group == iommu_group) | |
1281 | return g; | |
1282 | } | |
1283 | ||
1284 | return NULL; | |
1285 | } | |
1286 | ||
9d3a4de4 | 1287 | static bool vfio_iommu_has_sw_msi(struct iommu_group *group, phys_addr_t *base) |
5d704992 EA |
1288 | { |
1289 | struct list_head group_resv_regions; | |
1290 | struct iommu_resv_region *region, *next; | |
1291 | bool ret = false; | |
1292 | ||
1293 | INIT_LIST_HEAD(&group_resv_regions); | |
1294 | iommu_get_group_resv_regions(group, &group_resv_regions); | |
1295 | list_for_each_entry(region, &group_resv_regions, list) { | |
f203f7f1 RM |
1296 | /* |
1297 | * The presence of any 'real' MSI regions should take | |
1298 | * precedence over the software-managed one if the | |
1299 | * IOMMU driver happens to advertise both types. | |
1300 | */ | |
1301 | if (region->type == IOMMU_RESV_MSI) { | |
1302 | ret = false; | |
1303 | break; | |
1304 | } | |
1305 | ||
9d3a4de4 | 1306 | if (region->type == IOMMU_RESV_SW_MSI) { |
5d704992 EA |
1307 | *base = region->start; |
1308 | ret = true; | |
5d704992 EA |
1309 | } |
1310 | } | |
5d704992 EA |
1311 | list_for_each_entry_safe(region, next, &group_resv_regions, list) |
1312 | kfree(region); | |
1313 | return ret; | |
1314 | } | |
1315 | ||
73fa0d10 AW |
1316 | static int vfio_iommu_type1_attach_group(void *iommu_data, |
1317 | struct iommu_group *iommu_group) | |
1318 | { | |
1319 | struct vfio_iommu *iommu = iommu_data; | |
7896c998 | 1320 | struct vfio_group *group; |
1ef3e2bc | 1321 | struct vfio_domain *domain, *d; |
a54eb550 | 1322 | struct bus_type *bus = NULL, *mdev_bus; |
73fa0d10 | 1323 | int ret; |
9d72f87b | 1324 | bool resv_msi, msi_remap; |
5d704992 | 1325 | phys_addr_t resv_msi_base; |
73fa0d10 | 1326 | |
73fa0d10 AW |
1327 | mutex_lock(&iommu->lock); |
1328 | ||
1ef3e2bc | 1329 | list_for_each_entry(d, &iommu->domain_list, next) { |
7896c998 | 1330 | if (find_iommu_group(d, iommu_group)) { |
73fa0d10 | 1331 | mutex_unlock(&iommu->lock); |
73fa0d10 AW |
1332 | return -EINVAL; |
1333 | } | |
1334 | } | |
1335 | ||
a54eb550 KW |
1336 | if (iommu->external_domain) { |
1337 | if (find_iommu_group(iommu->external_domain, iommu_group)) { | |
1338 | mutex_unlock(&iommu->lock); | |
1339 | return -EINVAL; | |
1340 | } | |
1341 | } | |
1342 | ||
1ef3e2bc AW |
1343 | group = kzalloc(sizeof(*group), GFP_KERNEL); |
1344 | domain = kzalloc(sizeof(*domain), GFP_KERNEL); | |
1345 | if (!group || !domain) { | |
1346 | ret = -ENOMEM; | |
1347 | goto out_free; | |
1348 | } | |
1349 | ||
1350 | group->iommu_group = iommu_group; | |
1351 | ||
1352 | /* Determine bus_type in order to allocate a domain */ | |
1353 | ret = iommu_group_for_each_dev(iommu_group, &bus, vfio_bus_type); | |
1354 | if (ret) | |
1355 | goto out_free; | |
1356 | ||
a54eb550 KW |
1357 | mdev_bus = symbol_get(mdev_bus_type); |
1358 | ||
1359 | if (mdev_bus) { | |
1360 | if ((bus == mdev_bus) && !iommu_present(bus)) { | |
1361 | symbol_put(mdev_bus_type); | |
1362 | if (!iommu->external_domain) { | |
1363 | INIT_LIST_HEAD(&domain->group_list); | |
1364 | iommu->external_domain = domain; | |
1365 | } else | |
1366 | kfree(domain); | |
1367 | ||
1368 | list_add(&group->next, | |
1369 | &iommu->external_domain->group_list); | |
1370 | mutex_unlock(&iommu->lock); | |
1371 | return 0; | |
1372 | } | |
1373 | symbol_put(mdev_bus_type); | |
1374 | } | |
1375 | ||
1ef3e2bc AW |
1376 | domain->domain = iommu_domain_alloc(bus); |
1377 | if (!domain->domain) { | |
1378 | ret = -EIO; | |
1379 | goto out_free; | |
1380 | } | |
1381 | ||
f5c9eceb WD |
1382 | if (iommu->nesting) { |
1383 | int attr = 1; | |
1384 | ||
1385 | ret = iommu_domain_set_attr(domain->domain, DOMAIN_ATTR_NESTING, | |
1386 | &attr); | |
1387 | if (ret) | |
1388 | goto out_domain; | |
1389 | } | |
1390 | ||
1ef3e2bc AW |
1391 | ret = iommu_attach_group(domain->domain, iommu_group); |
1392 | if (ret) | |
1393 | goto out_domain; | |
1394 | ||
9d3a4de4 | 1395 | resv_msi = vfio_iommu_has_sw_msi(iommu_group, &resv_msi_base); |
5d704992 | 1396 | |
1ef3e2bc AW |
1397 | INIT_LIST_HEAD(&domain->group_list); |
1398 | list_add(&group->next, &domain->group_list); | |
1399 | ||
db406cc0 RM |
1400 | msi_remap = irq_domain_check_msi_remap() || |
1401 | iommu_capable(bus, IOMMU_CAP_INTR_REMAP); | |
9d72f87b EA |
1402 | |
1403 | if (!allow_unsafe_interrupts && !msi_remap) { | |
1ef3e2bc AW |
1404 | pr_warn("%s: No interrupt remapping support. Use the module param \"allow_unsafe_interrupts\" to enable VFIO IOMMU support on this platform\n", |
1405 | __func__); | |
1406 | ret = -EPERM; | |
1407 | goto out_detach; | |
1408 | } | |
1409 | ||
eb165f05 | 1410 | if (iommu_capable(bus, IOMMU_CAP_CACHE_COHERENCY)) |
1ef3e2bc AW |
1411 | domain->prot |= IOMMU_CACHE; |
1412 | ||
73fa0d10 | 1413 | /* |
1ef3e2bc AW |
1414 | * Try to match an existing compatible domain. We don't want to |
1415 | * preclude an IOMMU driver supporting multiple bus_types and being | |
1416 | * able to include different bus_types in the same IOMMU domain, so | |
1417 | * we test whether the domains use the same iommu_ops rather than | |
1418 | * testing if they're on the same bus_type. | |
73fa0d10 | 1419 | */ |
1ef3e2bc AW |
1420 | list_for_each_entry(d, &iommu->domain_list, next) { |
1421 | if (d->domain->ops == domain->domain->ops && | |
1422 | d->prot == domain->prot) { | |
1423 | iommu_detach_group(domain->domain, iommu_group); | |
1424 | if (!iommu_attach_group(d->domain, iommu_group)) { | |
1425 | list_add(&group->next, &d->group_list); | |
1426 | iommu_domain_free(domain->domain); | |
1427 | kfree(domain); | |
1428 | mutex_unlock(&iommu->lock); | |
1429 | return 0; | |
1430 | } | |
1431 | ||
1432 | ret = iommu_attach_group(domain->domain, iommu_group); | |
1433 | if (ret) | |
1434 | goto out_domain; | |
1435 | } | |
73fa0d10 AW |
1436 | } |
1437 | ||
6fe1010d AW |
1438 | vfio_test_domain_fgsp(domain); |
1439 | ||
1ef3e2bc AW |
1440 | /* replay mappings on new domains */ |
1441 | ret = vfio_iommu_replay(iommu, domain); | |
1442 | if (ret) | |
1443 | goto out_detach; | |
1444 | ||
2c9f1af5 WY |
1445 | if (resv_msi) { |
1446 | ret = iommu_get_msi_cookie(domain->domain, resv_msi_base); | |
1447 | if (ret) | |
1448 | goto out_detach; | |
1449 | } | |
5d704992 | 1450 | |
1ef3e2bc | 1451 | list_add(&domain->next, &iommu->domain_list); |
73fa0d10 AW |
1452 | |
1453 | mutex_unlock(&iommu->lock); | |
1454 | ||
1455 | return 0; | |
1ef3e2bc AW |
1456 | |
1457 | out_detach: | |
1458 | iommu_detach_group(domain->domain, iommu_group); | |
1459 | out_domain: | |
1460 | iommu_domain_free(domain->domain); | |
1461 | out_free: | |
1462 | kfree(domain); | |
1463 | kfree(group); | |
1464 | mutex_unlock(&iommu->lock); | |
1465 | return ret; | |
1466 | } | |
1467 | ||
1468 | static void vfio_iommu_unmap_unpin_all(struct vfio_iommu *iommu) | |
1469 | { | |
1470 | struct rb_node *node; | |
1471 | ||
1472 | while ((node = rb_first(&iommu->dma_list))) | |
1473 | vfio_remove_dma(iommu, rb_entry(node, struct vfio_dma, node)); | |
73fa0d10 AW |
1474 | } |
1475 | ||
a54eb550 KW |
1476 | static void vfio_iommu_unmap_unpin_reaccount(struct vfio_iommu *iommu) |
1477 | { | |
1478 | struct rb_node *n, *p; | |
1479 | ||
1480 | n = rb_first(&iommu->dma_list); | |
1481 | for (; n; n = rb_next(n)) { | |
1482 | struct vfio_dma *dma; | |
1483 | long locked = 0, unlocked = 0; | |
1484 | ||
1485 | dma = rb_entry(n, struct vfio_dma, node); | |
1486 | unlocked += vfio_unmap_unpin(iommu, dma, false); | |
1487 | p = rb_first(&dma->pfn_list); | |
1488 | for (; p; p = rb_next(p)) { | |
1489 | struct vfio_pfn *vpfn = rb_entry(p, struct vfio_pfn, | |
1490 | node); | |
1491 | ||
1492 | if (!is_invalid_reserved_pfn(vpfn->pfn)) | |
1493 | locked++; | |
1494 | } | |
0cfef2b7 | 1495 | vfio_lock_acct(dma->task, locked - unlocked, NULL); |
a54eb550 KW |
1496 | } |
1497 | } | |
1498 | ||
1499 | static void vfio_sanity_check_pfn_list(struct vfio_iommu *iommu) | |
1500 | { | |
1501 | struct rb_node *n; | |
1502 | ||
1503 | n = rb_first(&iommu->dma_list); | |
1504 | for (; n; n = rb_next(n)) { | |
1505 | struct vfio_dma *dma; | |
1506 | ||
1507 | dma = rb_entry(n, struct vfio_dma, node); | |
1508 | ||
1509 | if (WARN_ON(!RB_EMPTY_ROOT(&dma->pfn_list))) | |
1510 | break; | |
1511 | } | |
3cedd7d7 KW |
1512 | /* mdev vendor driver must unregister notifier */ |
1513 | WARN_ON(iommu->notifier.head); | |
a54eb550 KW |
1514 | } |
1515 | ||
73fa0d10 AW |
1516 | static void vfio_iommu_type1_detach_group(void *iommu_data, |
1517 | struct iommu_group *iommu_group) | |
1518 | { | |
1519 | struct vfio_iommu *iommu = iommu_data; | |
1ef3e2bc | 1520 | struct vfio_domain *domain; |
73fa0d10 AW |
1521 | struct vfio_group *group; |
1522 | ||
1523 | mutex_lock(&iommu->lock); | |
1524 | ||
a54eb550 KW |
1525 | if (iommu->external_domain) { |
1526 | group = find_iommu_group(iommu->external_domain, iommu_group); | |
1527 | if (group) { | |
1528 | list_del(&group->next); | |
1529 | kfree(group); | |
1530 | ||
1531 | if (list_empty(&iommu->external_domain->group_list)) { | |
1532 | vfio_sanity_check_pfn_list(iommu); | |
1533 | ||
1534 | if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu)) | |
1535 | vfio_iommu_unmap_unpin_all(iommu); | |
1536 | ||
1537 | kfree(iommu->external_domain); | |
1538 | iommu->external_domain = NULL; | |
1539 | } | |
1540 | goto detach_group_done; | |
1541 | } | |
1542 | } | |
1543 | ||
1ef3e2bc | 1544 | list_for_each_entry(domain, &iommu->domain_list, next) { |
7896c998 KW |
1545 | group = find_iommu_group(domain, iommu_group); |
1546 | if (!group) | |
1547 | continue; | |
1ef3e2bc | 1548 | |
7896c998 KW |
1549 | iommu_detach_group(domain->domain, iommu_group); |
1550 | list_del(&group->next); | |
1551 | kfree(group); | |
1552 | /* | |
a54eb550 KW |
1553 | * Group ownership provides privilege, if the group list is |
1554 | * empty, the domain goes away. If it's the last domain with | |
1555 | * iommu and external domain doesn't exist, then all the | |
1556 | * mappings go away too. If it's the last domain with iommu and | |
1557 | * external domain exist, update accounting | |
7896c998 KW |
1558 | */ |
1559 | if (list_empty(&domain->group_list)) { | |
a54eb550 KW |
1560 | if (list_is_singular(&iommu->domain_list)) { |
1561 | if (!iommu->external_domain) | |
1562 | vfio_iommu_unmap_unpin_all(iommu); | |
1563 | else | |
1564 | vfio_iommu_unmap_unpin_reaccount(iommu); | |
1565 | } | |
7896c998 KW |
1566 | iommu_domain_free(domain->domain); |
1567 | list_del(&domain->next); | |
1568 | kfree(domain); | |
73fa0d10 | 1569 | } |
a54eb550 | 1570 | break; |
73fa0d10 AW |
1571 | } |
1572 | ||
a54eb550 | 1573 | detach_group_done: |
73fa0d10 AW |
1574 | mutex_unlock(&iommu->lock); |
1575 | } | |
1576 | ||
1577 | static void *vfio_iommu_type1_open(unsigned long arg) | |
1578 | { | |
1579 | struct vfio_iommu *iommu; | |
1580 | ||
73fa0d10 AW |
1581 | iommu = kzalloc(sizeof(*iommu), GFP_KERNEL); |
1582 | if (!iommu) | |
1583 | return ERR_PTR(-ENOMEM); | |
1584 | ||
f5c9eceb WD |
1585 | switch (arg) { |
1586 | case VFIO_TYPE1_IOMMU: | |
1587 | break; | |
1588 | case VFIO_TYPE1_NESTING_IOMMU: | |
1589 | iommu->nesting = true; | |
1590 | case VFIO_TYPE1v2_IOMMU: | |
1591 | iommu->v2 = true; | |
1592 | break; | |
1593 | default: | |
1594 | kfree(iommu); | |
1595 | return ERR_PTR(-EINVAL); | |
1596 | } | |
1597 | ||
1ef3e2bc | 1598 | INIT_LIST_HEAD(&iommu->domain_list); |
cd9b2268 | 1599 | iommu->dma_list = RB_ROOT; |
73fa0d10 | 1600 | mutex_init(&iommu->lock); |
c086de81 | 1601 | BLOCKING_INIT_NOTIFIER_HEAD(&iommu->notifier); |
73fa0d10 AW |
1602 | |
1603 | return iommu; | |
1604 | } | |
1605 | ||
a54eb550 KW |
1606 | static void vfio_release_domain(struct vfio_domain *domain, bool external) |
1607 | { | |
1608 | struct vfio_group *group, *group_tmp; | |
1609 | ||
1610 | list_for_each_entry_safe(group, group_tmp, | |
1611 | &domain->group_list, next) { | |
1612 | if (!external) | |
1613 | iommu_detach_group(domain->domain, group->iommu_group); | |
1614 | list_del(&group->next); | |
1615 | kfree(group); | |
1616 | } | |
1617 | ||
1618 | if (!external) | |
1619 | iommu_domain_free(domain->domain); | |
1620 | } | |
1621 | ||
73fa0d10 AW |
1622 | static void vfio_iommu_type1_release(void *iommu_data) |
1623 | { | |
1624 | struct vfio_iommu *iommu = iommu_data; | |
1ef3e2bc | 1625 | struct vfio_domain *domain, *domain_tmp; |
a54eb550 KW |
1626 | |
1627 | if (iommu->external_domain) { | |
1628 | vfio_release_domain(iommu->external_domain, true); | |
1629 | vfio_sanity_check_pfn_list(iommu); | |
1630 | kfree(iommu->external_domain); | |
1631 | } | |
73fa0d10 | 1632 | |
1ef3e2bc | 1633 | vfio_iommu_unmap_unpin_all(iommu); |
73fa0d10 | 1634 | |
1ef3e2bc AW |
1635 | list_for_each_entry_safe(domain, domain_tmp, |
1636 | &iommu->domain_list, next) { | |
a54eb550 | 1637 | vfio_release_domain(domain, false); |
1ef3e2bc AW |
1638 | list_del(&domain->next); |
1639 | kfree(domain); | |
73fa0d10 | 1640 | } |
73fa0d10 AW |
1641 | kfree(iommu); |
1642 | } | |
1643 | ||
aa429318 AW |
1644 | static int vfio_domains_have_iommu_cache(struct vfio_iommu *iommu) |
1645 | { | |
1646 | struct vfio_domain *domain; | |
1647 | int ret = 1; | |
1648 | ||
1649 | mutex_lock(&iommu->lock); | |
1650 | list_for_each_entry(domain, &iommu->domain_list, next) { | |
1651 | if (!(domain->prot & IOMMU_CACHE)) { | |
1652 | ret = 0; | |
f5bfdbf2 | 1653 | break; |
aa429318 | 1654 | } |
73fa0d10 | 1655 | } |
aa429318 | 1656 | mutex_unlock(&iommu->lock); |
73fa0d10 | 1657 | |
aa429318 | 1658 | return ret; |
73fa0d10 AW |
1659 | } |
1660 | ||
1661 | static long vfio_iommu_type1_ioctl(void *iommu_data, | |
1662 | unsigned int cmd, unsigned long arg) | |
1663 | { | |
1664 | struct vfio_iommu *iommu = iommu_data; | |
1665 | unsigned long minsz; | |
1666 | ||
1667 | if (cmd == VFIO_CHECK_EXTENSION) { | |
1668 | switch (arg) { | |
1669 | case VFIO_TYPE1_IOMMU: | |
1ef3e2bc | 1670 | case VFIO_TYPE1v2_IOMMU: |
f5c9eceb | 1671 | case VFIO_TYPE1_NESTING_IOMMU: |
73fa0d10 | 1672 | return 1; |
aa429318 AW |
1673 | case VFIO_DMA_CC_IOMMU: |
1674 | if (!iommu) | |
1675 | return 0; | |
1676 | return vfio_domains_have_iommu_cache(iommu); | |
73fa0d10 AW |
1677 | default: |
1678 | return 0; | |
1679 | } | |
1680 | } else if (cmd == VFIO_IOMMU_GET_INFO) { | |
1681 | struct vfio_iommu_type1_info info; | |
1682 | ||
1683 | minsz = offsetofend(struct vfio_iommu_type1_info, iova_pgsizes); | |
1684 | ||
1685 | if (copy_from_user(&info, (void __user *)arg, minsz)) | |
1686 | return -EFAULT; | |
1687 | ||
1688 | if (info.argsz < minsz) | |
1689 | return -EINVAL; | |
1690 | ||
d4f50ee2 | 1691 | info.flags = VFIO_IOMMU_INFO_PGSIZES; |
73fa0d10 | 1692 | |
1ef3e2bc | 1693 | info.iova_pgsizes = vfio_pgsize_bitmap(iommu); |
73fa0d10 | 1694 | |
8160c4e4 MT |
1695 | return copy_to_user((void __user *)arg, &info, minsz) ? |
1696 | -EFAULT : 0; | |
73fa0d10 AW |
1697 | |
1698 | } else if (cmd == VFIO_IOMMU_MAP_DMA) { | |
1699 | struct vfio_iommu_type1_dma_map map; | |
1700 | uint32_t mask = VFIO_DMA_MAP_FLAG_READ | | |
1701 | VFIO_DMA_MAP_FLAG_WRITE; | |
1702 | ||
1703 | minsz = offsetofend(struct vfio_iommu_type1_dma_map, size); | |
1704 | ||
1705 | if (copy_from_user(&map, (void __user *)arg, minsz)) | |
1706 | return -EFAULT; | |
1707 | ||
1708 | if (map.argsz < minsz || map.flags & ~mask) | |
1709 | return -EINVAL; | |
1710 | ||
1711 | return vfio_dma_do_map(iommu, &map); | |
1712 | ||
1713 | } else if (cmd == VFIO_IOMMU_UNMAP_DMA) { | |
1714 | struct vfio_iommu_type1_dma_unmap unmap; | |
166fd7d9 | 1715 | long ret; |
73fa0d10 AW |
1716 | |
1717 | minsz = offsetofend(struct vfio_iommu_type1_dma_unmap, size); | |
1718 | ||
1719 | if (copy_from_user(&unmap, (void __user *)arg, minsz)) | |
1720 | return -EFAULT; | |
1721 | ||
1722 | if (unmap.argsz < minsz || unmap.flags) | |
1723 | return -EINVAL; | |
1724 | ||
166fd7d9 AW |
1725 | ret = vfio_dma_do_unmap(iommu, &unmap); |
1726 | if (ret) | |
1727 | return ret; | |
1728 | ||
8160c4e4 MT |
1729 | return copy_to_user((void __user *)arg, &unmap, minsz) ? |
1730 | -EFAULT : 0; | |
73fa0d10 AW |
1731 | } |
1732 | ||
1733 | return -ENOTTY; | |
1734 | } | |
1735 | ||
c086de81 | 1736 | static int vfio_iommu_type1_register_notifier(void *iommu_data, |
22195cbd | 1737 | unsigned long *events, |
c086de81 KW |
1738 | struct notifier_block *nb) |
1739 | { | |
1740 | struct vfio_iommu *iommu = iommu_data; | |
1741 | ||
22195cbd JS |
1742 | /* clear known events */ |
1743 | *events &= ~VFIO_IOMMU_NOTIFY_DMA_UNMAP; | |
1744 | ||
1745 | /* refuse to register if still events remaining */ | |
1746 | if (*events) | |
1747 | return -EINVAL; | |
1748 | ||
c086de81 KW |
1749 | return blocking_notifier_chain_register(&iommu->notifier, nb); |
1750 | } | |
1751 | ||
1752 | static int vfio_iommu_type1_unregister_notifier(void *iommu_data, | |
1753 | struct notifier_block *nb) | |
1754 | { | |
1755 | struct vfio_iommu *iommu = iommu_data; | |
1756 | ||
1757 | return blocking_notifier_chain_unregister(&iommu->notifier, nb); | |
1758 | } | |
1759 | ||
73fa0d10 | 1760 | static const struct vfio_iommu_driver_ops vfio_iommu_driver_ops_type1 = { |
c086de81 KW |
1761 | .name = "vfio-iommu-type1", |
1762 | .owner = THIS_MODULE, | |
1763 | .open = vfio_iommu_type1_open, | |
1764 | .release = vfio_iommu_type1_release, | |
1765 | .ioctl = vfio_iommu_type1_ioctl, | |
1766 | .attach_group = vfio_iommu_type1_attach_group, | |
1767 | .detach_group = vfio_iommu_type1_detach_group, | |
1768 | .pin_pages = vfio_iommu_type1_pin_pages, | |
1769 | .unpin_pages = vfio_iommu_type1_unpin_pages, | |
1770 | .register_notifier = vfio_iommu_type1_register_notifier, | |
1771 | .unregister_notifier = vfio_iommu_type1_unregister_notifier, | |
73fa0d10 AW |
1772 | }; |
1773 | ||
1774 | static int __init vfio_iommu_type1_init(void) | |
1775 | { | |
73fa0d10 AW |
1776 | return vfio_register_iommu_driver(&vfio_iommu_driver_ops_type1); |
1777 | } | |
1778 | ||
1779 | static void __exit vfio_iommu_type1_cleanup(void) | |
1780 | { | |
1781 | vfio_unregister_iommu_driver(&vfio_iommu_driver_ops_type1); | |
1782 | } | |
1783 | ||
1784 | module_init(vfio_iommu_type1_init); | |
1785 | module_exit(vfio_iommu_type1_cleanup); | |
1786 | ||
1787 | MODULE_VERSION(DRIVER_VERSION); | |
1788 | MODULE_LICENSE("GPL v2"); | |
1789 | MODULE_AUTHOR(DRIVER_AUTHOR); | |
1790 | MODULE_DESCRIPTION(DRIVER_DESC); |