]>
Commit | Line | Data |
---|---|---|
73fa0d10 AW |
1 | /* |
2 | * VFIO: IOMMU DMA mapping support for Type1 IOMMU | |
3 | * | |
4 | * Copyright (C) 2012 Red Hat, Inc. All rights reserved. | |
5 | * Author: Alex Williamson <alex.williamson@redhat.com> | |
6 | * | |
7 | * This program is free software; you can redistribute it and/or modify | |
8 | * it under the terms of the GNU General Public License version 2 as | |
9 | * published by the Free Software Foundation. | |
10 | * | |
11 | * Derived from original vfio: | |
12 | * Copyright 2010 Cisco Systems, Inc. All rights reserved. | |
13 | * Author: Tom Lyon, pugs@cisco.com | |
14 | * | |
15 | * We arbitrarily define a Type1 IOMMU as one matching the below code. | |
16 | * It could be called the x86 IOMMU as it's designed for AMD-Vi & Intel | |
17 | * VT-d, but that makes it harder to re-use as theoretically anyone | |
18 | * implementing a similar IOMMU could make use of this. We expect the | |
19 | * IOMMU to support the IOMMU API and have few to no restrictions around | |
20 | * the IOVA range that can be mapped. The Type1 IOMMU is currently | |
21 | * optimized for relatively static mappings of a userspace process with | |
22 | * userpsace pages pinned into memory. We also assume devices and IOMMU | |
23 | * domains are PCI based as the IOMMU API is still centered around a | |
24 | * device/bus interface rather than a group interface. | |
25 | */ | |
26 | ||
27 | #include <linux/compat.h> | |
28 | #include <linux/device.h> | |
29 | #include <linux/fs.h> | |
30 | #include <linux/iommu.h> | |
31 | #include <linux/module.h> | |
32 | #include <linux/mm.h> | |
cd9b2268 | 33 | #include <linux/rbtree.h> |
3f07c014 | 34 | #include <linux/sched/signal.h> |
6e84f315 | 35 | #include <linux/sched/mm.h> |
73fa0d10 AW |
36 | #include <linux/slab.h> |
37 | #include <linux/uaccess.h> | |
38 | #include <linux/vfio.h> | |
39 | #include <linux/workqueue.h> | |
a54eb550 | 40 | #include <linux/mdev.h> |
c086de81 | 41 | #include <linux/notifier.h> |
5d704992 | 42 | #include <linux/dma-iommu.h> |
9d72f87b | 43 | #include <linux/irqdomain.h> |
73fa0d10 AW |
44 | |
45 | #define DRIVER_VERSION "0.2" | |
46 | #define DRIVER_AUTHOR "Alex Williamson <alex.williamson@redhat.com>" | |
47 | #define DRIVER_DESC "Type1 IOMMU driver for VFIO" | |
48 | ||
49 | static bool allow_unsafe_interrupts; | |
50 | module_param_named(allow_unsafe_interrupts, | |
51 | allow_unsafe_interrupts, bool, S_IRUGO | S_IWUSR); | |
52 | MODULE_PARM_DESC(allow_unsafe_interrupts, | |
53 | "Enable VFIO IOMMU support for on platforms without interrupt remapping support."); | |
54 | ||
5c6c2b21 AW |
55 | static bool disable_hugepages; |
56 | module_param_named(disable_hugepages, | |
57 | disable_hugepages, bool, S_IRUGO | S_IWUSR); | |
58 | MODULE_PARM_DESC(disable_hugepages, | |
59 | "Disable VFIO IOMMU support for IOMMU hugepages."); | |
60 | ||
73fa0d10 | 61 | struct vfio_iommu { |
1ef3e2bc | 62 | struct list_head domain_list; |
a54eb550 | 63 | struct vfio_domain *external_domain; /* domain for external user */ |
73fa0d10 | 64 | struct mutex lock; |
cd9b2268 | 65 | struct rb_root dma_list; |
c086de81 | 66 | struct blocking_notifier_head notifier; |
f5c9eceb WD |
67 | bool v2; |
68 | bool nesting; | |
1ef3e2bc AW |
69 | }; |
70 | ||
71 | struct vfio_domain { | |
72 | struct iommu_domain *domain; | |
73 | struct list_head next; | |
73fa0d10 | 74 | struct list_head group_list; |
1ef3e2bc | 75 | int prot; /* IOMMU_CACHE */ |
6fe1010d | 76 | bool fgsp; /* Fine-grained super pages */ |
73fa0d10 AW |
77 | }; |
78 | ||
79 | struct vfio_dma { | |
cd9b2268 | 80 | struct rb_node node; |
73fa0d10 AW |
81 | dma_addr_t iova; /* Device address */ |
82 | unsigned long vaddr; /* Process virtual addr */ | |
166fd7d9 | 83 | size_t size; /* Map size (bytes) */ |
73fa0d10 | 84 | int prot; /* IOMMU_READ/WRITE */ |
a54eb550 | 85 | bool iommu_mapped; |
8f0d5bb9 | 86 | struct task_struct *task; |
a54eb550 | 87 | struct rb_root pfn_list; /* Ex-user pinned pfn list */ |
73fa0d10 AW |
88 | }; |
89 | ||
90 | struct vfio_group { | |
91 | struct iommu_group *iommu_group; | |
92 | struct list_head next; | |
93 | }; | |
94 | ||
a54eb550 KW |
95 | /* |
96 | * Guest RAM pinning working set or DMA target | |
97 | */ | |
98 | struct vfio_pfn { | |
99 | struct rb_node node; | |
100 | dma_addr_t iova; /* Device address */ | |
101 | unsigned long pfn; /* Host pfn */ | |
102 | atomic_t ref_count; | |
103 | }; | |
104 | ||
105 | #define IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu) \ | |
106 | (!list_empty(&iommu->domain_list)) | |
107 | ||
108 | static int put_pfn(unsigned long pfn, int prot); | |
109 | ||
73fa0d10 AW |
110 | /* |
111 | * This code handles mapping and unmapping of user data buffers | |
112 | * into DMA'ble space using the IOMMU | |
113 | */ | |
114 | ||
cd9b2268 AW |
115 | static struct vfio_dma *vfio_find_dma(struct vfio_iommu *iommu, |
116 | dma_addr_t start, size_t size) | |
117 | { | |
118 | struct rb_node *node = iommu->dma_list.rb_node; | |
119 | ||
120 | while (node) { | |
121 | struct vfio_dma *dma = rb_entry(node, struct vfio_dma, node); | |
122 | ||
123 | if (start + size <= dma->iova) | |
124 | node = node->rb_left; | |
166fd7d9 | 125 | else if (start >= dma->iova + dma->size) |
cd9b2268 AW |
126 | node = node->rb_right; |
127 | else | |
128 | return dma; | |
129 | } | |
130 | ||
131 | return NULL; | |
132 | } | |
133 | ||
1ef3e2bc | 134 | static void vfio_link_dma(struct vfio_iommu *iommu, struct vfio_dma *new) |
cd9b2268 AW |
135 | { |
136 | struct rb_node **link = &iommu->dma_list.rb_node, *parent = NULL; | |
137 | struct vfio_dma *dma; | |
138 | ||
139 | while (*link) { | |
140 | parent = *link; | |
141 | dma = rb_entry(parent, struct vfio_dma, node); | |
142 | ||
166fd7d9 | 143 | if (new->iova + new->size <= dma->iova) |
cd9b2268 AW |
144 | link = &(*link)->rb_left; |
145 | else | |
146 | link = &(*link)->rb_right; | |
147 | } | |
148 | ||
149 | rb_link_node(&new->node, parent, link); | |
150 | rb_insert_color(&new->node, &iommu->dma_list); | |
151 | } | |
152 | ||
1ef3e2bc | 153 | static void vfio_unlink_dma(struct vfio_iommu *iommu, struct vfio_dma *old) |
cd9b2268 AW |
154 | { |
155 | rb_erase(&old->node, &iommu->dma_list); | |
156 | } | |
157 | ||
a54eb550 KW |
158 | /* |
159 | * Helper Functions for host iova-pfn list | |
160 | */ | |
161 | static struct vfio_pfn *vfio_find_vpfn(struct vfio_dma *dma, dma_addr_t iova) | |
162 | { | |
163 | struct vfio_pfn *vpfn; | |
164 | struct rb_node *node = dma->pfn_list.rb_node; | |
165 | ||
166 | while (node) { | |
167 | vpfn = rb_entry(node, struct vfio_pfn, node); | |
168 | ||
169 | if (iova < vpfn->iova) | |
170 | node = node->rb_left; | |
171 | else if (iova > vpfn->iova) | |
172 | node = node->rb_right; | |
173 | else | |
174 | return vpfn; | |
175 | } | |
176 | return NULL; | |
177 | } | |
178 | ||
179 | static void vfio_link_pfn(struct vfio_dma *dma, | |
180 | struct vfio_pfn *new) | |
181 | { | |
182 | struct rb_node **link, *parent = NULL; | |
183 | struct vfio_pfn *vpfn; | |
184 | ||
185 | link = &dma->pfn_list.rb_node; | |
186 | while (*link) { | |
187 | parent = *link; | |
188 | vpfn = rb_entry(parent, struct vfio_pfn, node); | |
189 | ||
190 | if (new->iova < vpfn->iova) | |
191 | link = &(*link)->rb_left; | |
192 | else | |
193 | link = &(*link)->rb_right; | |
194 | } | |
195 | ||
196 | rb_link_node(&new->node, parent, link); | |
197 | rb_insert_color(&new->node, &dma->pfn_list); | |
198 | } | |
199 | ||
200 | static void vfio_unlink_pfn(struct vfio_dma *dma, struct vfio_pfn *old) | |
201 | { | |
202 | rb_erase(&old->node, &dma->pfn_list); | |
203 | } | |
204 | ||
205 | static int vfio_add_to_pfn_list(struct vfio_dma *dma, dma_addr_t iova, | |
206 | unsigned long pfn) | |
207 | { | |
208 | struct vfio_pfn *vpfn; | |
209 | ||
210 | vpfn = kzalloc(sizeof(*vpfn), GFP_KERNEL); | |
211 | if (!vpfn) | |
212 | return -ENOMEM; | |
213 | ||
214 | vpfn->iova = iova; | |
215 | vpfn->pfn = pfn; | |
216 | atomic_set(&vpfn->ref_count, 1); | |
217 | vfio_link_pfn(dma, vpfn); | |
218 | return 0; | |
219 | } | |
220 | ||
221 | static void vfio_remove_from_pfn_list(struct vfio_dma *dma, | |
222 | struct vfio_pfn *vpfn) | |
223 | { | |
224 | vfio_unlink_pfn(dma, vpfn); | |
225 | kfree(vpfn); | |
226 | } | |
227 | ||
228 | static struct vfio_pfn *vfio_iova_get_vfio_pfn(struct vfio_dma *dma, | |
229 | unsigned long iova) | |
230 | { | |
231 | struct vfio_pfn *vpfn = vfio_find_vpfn(dma, iova); | |
232 | ||
233 | if (vpfn) | |
234 | atomic_inc(&vpfn->ref_count); | |
235 | return vpfn; | |
236 | } | |
237 | ||
238 | static int vfio_iova_put_vfio_pfn(struct vfio_dma *dma, struct vfio_pfn *vpfn) | |
239 | { | |
240 | int ret = 0; | |
241 | ||
242 | if (atomic_dec_and_test(&vpfn->ref_count)) { | |
243 | ret = put_pfn(vpfn->pfn, dma->prot); | |
244 | vfio_remove_from_pfn_list(dma, vpfn); | |
245 | } | |
246 | return ret; | |
247 | } | |
248 | ||
73fa0d10 AW |
249 | struct vwork { |
250 | struct mm_struct *mm; | |
251 | long npage; | |
252 | struct work_struct work; | |
253 | }; | |
254 | ||
255 | /* delayed decrement/increment for locked_vm */ | |
256 | static void vfio_lock_acct_bg(struct work_struct *work) | |
257 | { | |
258 | struct vwork *vwork = container_of(work, struct vwork, work); | |
259 | struct mm_struct *mm; | |
260 | ||
261 | mm = vwork->mm; | |
262 | down_write(&mm->mmap_sem); | |
263 | mm->locked_vm += vwork->npage; | |
264 | up_write(&mm->mmap_sem); | |
265 | mmput(mm); | |
266 | kfree(vwork); | |
267 | } | |
268 | ||
3624a248 | 269 | static void vfio_lock_acct(struct task_struct *task, long npage) |
73fa0d10 AW |
270 | { |
271 | struct vwork *vwork; | |
272 | struct mm_struct *mm; | |
6c38c055 | 273 | bool is_current; |
73fa0d10 | 274 | |
3624a248 KW |
275 | if (!npage) |
276 | return; | |
277 | ||
6c38c055 AW |
278 | is_current = (task->mm == current->mm); |
279 | ||
280 | mm = is_current ? task->mm : get_task_mm(task); | |
3624a248 | 281 | if (!mm) |
6c38c055 | 282 | return; /* process exited */ |
73fa0d10 | 283 | |
3624a248 KW |
284 | if (down_write_trylock(&mm->mmap_sem)) { |
285 | mm->locked_vm += npage; | |
286 | up_write(&mm->mmap_sem); | |
6c38c055 AW |
287 | if (!is_current) |
288 | mmput(mm); | |
73fa0d10 AW |
289 | return; |
290 | } | |
291 | ||
6c38c055 AW |
292 | if (is_current) { |
293 | mm = get_task_mm(task); | |
294 | if (!mm) | |
295 | return; | |
296 | } | |
297 | ||
73fa0d10 AW |
298 | /* |
299 | * Couldn't get mmap_sem lock, so must setup to update | |
300 | * mm->locked_vm later. If locked_vm were atomic, we | |
301 | * wouldn't need this silliness | |
302 | */ | |
303 | vwork = kmalloc(sizeof(struct vwork), GFP_KERNEL); | |
6c38c055 | 304 | if (WARN_ON(!vwork)) { |
3624a248 | 305 | mmput(mm); |
73fa0d10 AW |
306 | return; |
307 | } | |
308 | INIT_WORK(&vwork->work, vfio_lock_acct_bg); | |
309 | vwork->mm = mm; | |
310 | vwork->npage = npage; | |
311 | schedule_work(&vwork->work); | |
312 | } | |
313 | ||
314 | /* | |
315 | * Some mappings aren't backed by a struct page, for example an mmap'd | |
316 | * MMIO range for our own or another device. These use a different | |
317 | * pfn conversion and shouldn't be tracked as locked pages. | |
318 | */ | |
319 | static bool is_invalid_reserved_pfn(unsigned long pfn) | |
320 | { | |
321 | if (pfn_valid(pfn)) { | |
322 | bool reserved; | |
323 | struct page *tail = pfn_to_page(pfn); | |
668f9abb | 324 | struct page *head = compound_head(tail); |
73fa0d10 AW |
325 | reserved = !!(PageReserved(head)); |
326 | if (head != tail) { | |
327 | /* | |
328 | * "head" is not a dangling pointer | |
668f9abb | 329 | * (compound_head takes care of that) |
73fa0d10 AW |
330 | * but the hugepage may have been split |
331 | * from under us (and we may not hold a | |
332 | * reference count on the head page so it can | |
333 | * be reused before we run PageReferenced), so | |
334 | * we've to check PageTail before returning | |
335 | * what we just read. | |
336 | */ | |
337 | smp_rmb(); | |
338 | if (PageTail(tail)) | |
339 | return reserved; | |
340 | } | |
341 | return PageReserved(tail); | |
342 | } | |
343 | ||
344 | return true; | |
345 | } | |
346 | ||
347 | static int put_pfn(unsigned long pfn, int prot) | |
348 | { | |
349 | if (!is_invalid_reserved_pfn(pfn)) { | |
350 | struct page *page = pfn_to_page(pfn); | |
351 | if (prot & IOMMU_WRITE) | |
352 | SetPageDirty(page); | |
353 | put_page(page); | |
354 | return 1; | |
355 | } | |
356 | return 0; | |
357 | } | |
358 | ||
ea85cf35 KW |
359 | static int vaddr_get_pfn(struct mm_struct *mm, unsigned long vaddr, |
360 | int prot, unsigned long *pfn) | |
73fa0d10 AW |
361 | { |
362 | struct page *page[1]; | |
363 | struct vm_area_struct *vma; | |
ea85cf35 | 364 | int ret; |
73fa0d10 | 365 | |
ea85cf35 KW |
366 | if (mm == current->mm) { |
367 | ret = get_user_pages_fast(vaddr, 1, !!(prot & IOMMU_WRITE), | |
368 | page); | |
369 | } else { | |
370 | unsigned int flags = 0; | |
371 | ||
372 | if (prot & IOMMU_WRITE) | |
373 | flags |= FOLL_WRITE; | |
374 | ||
375 | down_read(&mm->mmap_sem); | |
376 | ret = get_user_pages_remote(NULL, mm, vaddr, 1, flags, page, | |
5b56d49f | 377 | NULL, NULL); |
ea85cf35 KW |
378 | up_read(&mm->mmap_sem); |
379 | } | |
380 | ||
381 | if (ret == 1) { | |
73fa0d10 AW |
382 | *pfn = page_to_pfn(page[0]); |
383 | return 0; | |
384 | } | |
385 | ||
ea85cf35 | 386 | down_read(&mm->mmap_sem); |
73fa0d10 | 387 | |
ea85cf35 | 388 | vma = find_vma_intersection(mm, vaddr, vaddr + 1); |
73fa0d10 AW |
389 | |
390 | if (vma && vma->vm_flags & VM_PFNMAP) { | |
391 | *pfn = ((vaddr - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff; | |
392 | if (is_invalid_reserved_pfn(*pfn)) | |
393 | ret = 0; | |
394 | } | |
395 | ||
ea85cf35 | 396 | up_read(&mm->mmap_sem); |
73fa0d10 AW |
397 | return ret; |
398 | } | |
399 | ||
166fd7d9 AW |
400 | /* |
401 | * Attempt to pin pages. We really don't want to track all the pfns and | |
402 | * the iommu can only map chunks of consecutive pfns anyway, so get the | |
403 | * first page and all consecutive pages with the same locking. | |
404 | */ | |
8f0d5bb9 | 405 | static long vfio_pin_pages_remote(struct vfio_dma *dma, unsigned long vaddr, |
a54eb550 | 406 | long npage, unsigned long *pfn_base) |
73fa0d10 | 407 | { |
6c38c055 AW |
408 | unsigned long limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT; |
409 | bool lock_cap = capable(CAP_IPC_LOCK); | |
410 | long ret, pinned = 0, lock_acct = 0; | |
babbf176 | 411 | bool rsvd; |
a54eb550 | 412 | dma_addr_t iova = vaddr - dma->vaddr + dma->iova; |
73fa0d10 | 413 | |
6c38c055 AW |
414 | /* This code path is only user initiated */ |
415 | if (!current->mm) | |
166fd7d9 | 416 | return -ENODEV; |
73fa0d10 | 417 | |
6c38c055 | 418 | ret = vaddr_get_pfn(current->mm, vaddr, dma->prot, pfn_base); |
166fd7d9 | 419 | if (ret) |
6c38c055 | 420 | return ret; |
73fa0d10 | 421 | |
6c38c055 | 422 | pinned++; |
babbf176 | 423 | rsvd = is_invalid_reserved_pfn(*pfn_base); |
73fa0d10 | 424 | |
a54eb550 KW |
425 | /* |
426 | * Reserved pages aren't counted against the user, externally pinned | |
427 | * pages are already counted against the user. | |
428 | */ | |
429 | if (!rsvd && !vfio_find_vpfn(dma, iova)) { | |
6c38c055 | 430 | if (!lock_cap && current->mm->locked_vm + 1 > limit) { |
a54eb550 KW |
431 | put_pfn(*pfn_base, dma->prot); |
432 | pr_warn("%s: RLIMIT_MEMLOCK (%ld) exceeded\n", __func__, | |
433 | limit << PAGE_SHIFT); | |
6c38c055 | 434 | return -ENOMEM; |
a54eb550 KW |
435 | } |
436 | lock_acct++; | |
5c6c2b21 AW |
437 | } |
438 | ||
6c38c055 AW |
439 | if (unlikely(disable_hugepages)) |
440 | goto out; | |
73fa0d10 | 441 | |
6c38c055 AW |
442 | /* Lock all the consecutive pages from pfn_base */ |
443 | for (vaddr += PAGE_SIZE, iova += PAGE_SIZE; pinned < npage; | |
444 | pinned++, vaddr += PAGE_SIZE, iova += PAGE_SIZE) { | |
445 | unsigned long pfn = 0; | |
446 | ||
447 | ret = vaddr_get_pfn(current->mm, vaddr, dma->prot, &pfn); | |
448 | if (ret) | |
449 | break; | |
450 | ||
451 | if (pfn != *pfn_base + pinned || | |
452 | rsvd != is_invalid_reserved_pfn(pfn)) { | |
453 | put_pfn(pfn, dma->prot); | |
454 | break; | |
455 | } | |
166fd7d9 | 456 | |
6c38c055 AW |
457 | if (!rsvd && !vfio_find_vpfn(dma, iova)) { |
458 | if (!lock_cap && | |
459 | current->mm->locked_vm + lock_acct + 1 > limit) { | |
a54eb550 | 460 | put_pfn(pfn, dma->prot); |
6c38c055 AW |
461 | pr_warn("%s: RLIMIT_MEMLOCK (%ld) exceeded\n", |
462 | __func__, limit << PAGE_SHIFT); | |
a54eb550 KW |
463 | break; |
464 | } | |
6c38c055 | 465 | lock_acct++; |
166fd7d9 AW |
466 | } |
467 | } | |
468 | ||
6c38c055 AW |
469 | out: |
470 | vfio_lock_acct(current, lock_acct); | |
166fd7d9 | 471 | |
6c38c055 | 472 | return pinned; |
166fd7d9 AW |
473 | } |
474 | ||
a54eb550 KW |
475 | static long vfio_unpin_pages_remote(struct vfio_dma *dma, dma_addr_t iova, |
476 | unsigned long pfn, long npage, | |
477 | bool do_accounting) | |
166fd7d9 | 478 | { |
a54eb550 | 479 | long unlocked = 0, locked = 0; |
166fd7d9 AW |
480 | long i; |
481 | ||
6c38c055 | 482 | for (i = 0; i < npage; i++, iova += PAGE_SIZE) { |
a54eb550 KW |
483 | if (put_pfn(pfn++, dma->prot)) { |
484 | unlocked++; | |
6c38c055 | 485 | if (vfio_find_vpfn(dma, iova)) |
a54eb550 KW |
486 | locked++; |
487 | } | |
488 | } | |
489 | ||
490 | if (do_accounting) | |
491 | vfio_lock_acct(dma->task, locked - unlocked); | |
492 | ||
493 | return unlocked; | |
494 | } | |
495 | ||
496 | static int vfio_pin_page_external(struct vfio_dma *dma, unsigned long vaddr, | |
497 | unsigned long *pfn_base, bool do_accounting) | |
498 | { | |
499 | unsigned long limit; | |
d1b333d1 | 500 | bool lock_cap = has_capability(dma->task, CAP_IPC_LOCK); |
a54eb550 KW |
501 | struct mm_struct *mm; |
502 | int ret; | |
503 | bool rsvd; | |
504 | ||
505 | mm = get_task_mm(dma->task); | |
506 | if (!mm) | |
507 | return -ENODEV; | |
508 | ||
509 | ret = vaddr_get_pfn(mm, vaddr, dma->prot, pfn_base); | |
510 | if (ret) | |
511 | goto pin_page_exit; | |
512 | ||
513 | rsvd = is_invalid_reserved_pfn(*pfn_base); | |
514 | limit = task_rlimit(dma->task, RLIMIT_MEMLOCK) >> PAGE_SHIFT; | |
515 | ||
516 | if (!rsvd && !lock_cap && mm->locked_vm + 1 > limit) { | |
517 | put_pfn(*pfn_base, dma->prot); | |
518 | pr_warn("%s: Task %s (%d) RLIMIT_MEMLOCK (%ld) exceeded\n", | |
519 | __func__, dma->task->comm, task_pid_nr(dma->task), | |
520 | limit << PAGE_SHIFT); | |
521 | ret = -ENOMEM; | |
522 | goto pin_page_exit; | |
523 | } | |
524 | ||
525 | if (!rsvd && do_accounting) | |
526 | vfio_lock_acct(dma->task, 1); | |
527 | ret = 1; | |
528 | ||
529 | pin_page_exit: | |
530 | mmput(mm); | |
531 | return ret; | |
532 | } | |
533 | ||
534 | static int vfio_unpin_page_external(struct vfio_dma *dma, dma_addr_t iova, | |
535 | bool do_accounting) | |
536 | { | |
537 | int unlocked; | |
538 | struct vfio_pfn *vpfn = vfio_find_vpfn(dma, iova); | |
539 | ||
540 | if (!vpfn) | |
541 | return 0; | |
542 | ||
543 | unlocked = vfio_iova_put_vfio_pfn(dma, vpfn); | |
166fd7d9 AW |
544 | |
545 | if (do_accounting) | |
8f0d5bb9 | 546 | vfio_lock_acct(dma->task, -unlocked); |
166fd7d9 AW |
547 | |
548 | return unlocked; | |
549 | } | |
550 | ||
a54eb550 KW |
551 | static int vfio_iommu_type1_pin_pages(void *iommu_data, |
552 | unsigned long *user_pfn, | |
553 | int npage, int prot, | |
554 | unsigned long *phys_pfn) | |
555 | { | |
556 | struct vfio_iommu *iommu = iommu_data; | |
557 | int i, j, ret; | |
558 | unsigned long remote_vaddr; | |
559 | struct vfio_dma *dma; | |
560 | bool do_accounting; | |
561 | ||
562 | if (!iommu || !user_pfn || !phys_pfn) | |
563 | return -EINVAL; | |
564 | ||
565 | /* Supported for v2 version only */ | |
566 | if (!iommu->v2) | |
567 | return -EACCES; | |
568 | ||
569 | mutex_lock(&iommu->lock); | |
570 | ||
c086de81 KW |
571 | /* Fail if notifier list is empty */ |
572 | if ((!iommu->external_domain) || (!iommu->notifier.head)) { | |
a54eb550 KW |
573 | ret = -EINVAL; |
574 | goto pin_done; | |
575 | } | |
576 | ||
577 | /* | |
578 | * If iommu capable domain exist in the container then all pages are | |
579 | * already pinned and accounted. Accouting should be done if there is no | |
580 | * iommu capable domain in the container. | |
581 | */ | |
582 | do_accounting = !IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu); | |
583 | ||
584 | for (i = 0; i < npage; i++) { | |
585 | dma_addr_t iova; | |
586 | struct vfio_pfn *vpfn; | |
587 | ||
588 | iova = user_pfn[i] << PAGE_SHIFT; | |
2b8bb1d7 | 589 | dma = vfio_find_dma(iommu, iova, PAGE_SIZE); |
a54eb550 KW |
590 | if (!dma) { |
591 | ret = -EINVAL; | |
592 | goto pin_unwind; | |
593 | } | |
594 | ||
595 | if ((dma->prot & prot) != prot) { | |
596 | ret = -EPERM; | |
597 | goto pin_unwind; | |
598 | } | |
599 | ||
600 | vpfn = vfio_iova_get_vfio_pfn(dma, iova); | |
601 | if (vpfn) { | |
602 | phys_pfn[i] = vpfn->pfn; | |
603 | continue; | |
604 | } | |
605 | ||
606 | remote_vaddr = dma->vaddr + iova - dma->iova; | |
607 | ret = vfio_pin_page_external(dma, remote_vaddr, &phys_pfn[i], | |
608 | do_accounting); | |
609 | if (ret <= 0) { | |
610 | WARN_ON(!ret); | |
611 | goto pin_unwind; | |
612 | } | |
613 | ||
614 | ret = vfio_add_to_pfn_list(dma, iova, phys_pfn[i]); | |
615 | if (ret) { | |
616 | vfio_unpin_page_external(dma, iova, do_accounting); | |
617 | goto pin_unwind; | |
618 | } | |
619 | } | |
620 | ||
621 | ret = i; | |
622 | goto pin_done; | |
623 | ||
624 | pin_unwind: | |
625 | phys_pfn[i] = 0; | |
626 | for (j = 0; j < i; j++) { | |
627 | dma_addr_t iova; | |
628 | ||
629 | iova = user_pfn[j] << PAGE_SHIFT; | |
2b8bb1d7 | 630 | dma = vfio_find_dma(iommu, iova, PAGE_SIZE); |
a54eb550 KW |
631 | vfio_unpin_page_external(dma, iova, do_accounting); |
632 | phys_pfn[j] = 0; | |
633 | } | |
634 | pin_done: | |
635 | mutex_unlock(&iommu->lock); | |
636 | return ret; | |
637 | } | |
638 | ||
639 | static int vfio_iommu_type1_unpin_pages(void *iommu_data, | |
640 | unsigned long *user_pfn, | |
641 | int npage) | |
642 | { | |
643 | struct vfio_iommu *iommu = iommu_data; | |
644 | bool do_accounting; | |
645 | int i; | |
646 | ||
647 | if (!iommu || !user_pfn) | |
648 | return -EINVAL; | |
649 | ||
650 | /* Supported for v2 version only */ | |
651 | if (!iommu->v2) | |
652 | return -EACCES; | |
653 | ||
654 | mutex_lock(&iommu->lock); | |
655 | ||
656 | if (!iommu->external_domain) { | |
657 | mutex_unlock(&iommu->lock); | |
658 | return -EINVAL; | |
659 | } | |
660 | ||
661 | do_accounting = !IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu); | |
662 | for (i = 0; i < npage; i++) { | |
663 | struct vfio_dma *dma; | |
664 | dma_addr_t iova; | |
665 | ||
666 | iova = user_pfn[i] << PAGE_SHIFT; | |
2b8bb1d7 | 667 | dma = vfio_find_dma(iommu, iova, PAGE_SIZE); |
a54eb550 KW |
668 | if (!dma) |
669 | goto unpin_exit; | |
670 | vfio_unpin_page_external(dma, iova, do_accounting); | |
671 | } | |
672 | ||
673 | unpin_exit: | |
674 | mutex_unlock(&iommu->lock); | |
675 | return i > npage ? npage : (i > 0 ? i : -EINVAL); | |
676 | } | |
677 | ||
678 | static long vfio_unmap_unpin(struct vfio_iommu *iommu, struct vfio_dma *dma, | |
679 | bool do_accounting) | |
166fd7d9 | 680 | { |
1ef3e2bc AW |
681 | dma_addr_t iova = dma->iova, end = dma->iova + dma->size; |
682 | struct vfio_domain *domain, *d; | |
166fd7d9 AW |
683 | long unlocked = 0; |
684 | ||
1ef3e2bc | 685 | if (!dma->size) |
a54eb550 KW |
686 | return 0; |
687 | ||
688 | if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu)) | |
689 | return 0; | |
690 | ||
1ef3e2bc AW |
691 | /* |
692 | * We use the IOMMU to track the physical addresses, otherwise we'd | |
693 | * need a much more complicated tracking system. Unfortunately that | |
694 | * means we need to use one of the iommu domains to figure out the | |
695 | * pfns to unpin. The rest need to be unmapped in advance so we have | |
696 | * no iommu translations remaining when the pages are unpinned. | |
697 | */ | |
698 | domain = d = list_first_entry(&iommu->domain_list, | |
699 | struct vfio_domain, next); | |
700 | ||
c5e66887 | 701 | list_for_each_entry_continue(d, &iommu->domain_list, next) { |
1ef3e2bc | 702 | iommu_unmap(d->domain, dma->iova, dma->size); |
c5e66887 AW |
703 | cond_resched(); |
704 | } | |
1ef3e2bc | 705 | |
166fd7d9 | 706 | while (iova < end) { |
6fe1010d AW |
707 | size_t unmapped, len; |
708 | phys_addr_t phys, next; | |
166fd7d9 | 709 | |
1ef3e2bc | 710 | phys = iommu_iova_to_phys(domain->domain, iova); |
166fd7d9 AW |
711 | if (WARN_ON(!phys)) { |
712 | iova += PAGE_SIZE; | |
713 | continue; | |
73fa0d10 | 714 | } |
166fd7d9 | 715 | |
6fe1010d AW |
716 | /* |
717 | * To optimize for fewer iommu_unmap() calls, each of which | |
718 | * may require hardware cache flushing, try to find the | |
719 | * largest contiguous physical memory chunk to unmap. | |
720 | */ | |
721 | for (len = PAGE_SIZE; | |
722 | !domain->fgsp && iova + len < end; len += PAGE_SIZE) { | |
723 | next = iommu_iova_to_phys(domain->domain, iova + len); | |
724 | if (next != phys + len) | |
725 | break; | |
726 | } | |
727 | ||
728 | unmapped = iommu_unmap(domain->domain, iova, len); | |
1ef3e2bc | 729 | if (WARN_ON(!unmapped)) |
166fd7d9 AW |
730 | break; |
731 | ||
a54eb550 KW |
732 | unlocked += vfio_unpin_pages_remote(dma, iova, |
733 | phys >> PAGE_SHIFT, | |
2169037d | 734 | unmapped >> PAGE_SHIFT, |
a54eb550 | 735 | false); |
166fd7d9 | 736 | iova += unmapped; |
c5e66887 AW |
737 | |
738 | cond_resched(); | |
73fa0d10 | 739 | } |
166fd7d9 | 740 | |
a54eb550 KW |
741 | dma->iommu_mapped = false; |
742 | if (do_accounting) { | |
743 | vfio_lock_acct(dma->task, -unlocked); | |
744 | return 0; | |
745 | } | |
746 | return unlocked; | |
73fa0d10 AW |
747 | } |
748 | ||
1ef3e2bc | 749 | static void vfio_remove_dma(struct vfio_iommu *iommu, struct vfio_dma *dma) |
73fa0d10 | 750 | { |
a54eb550 | 751 | vfio_unmap_unpin(iommu, dma, true); |
1ef3e2bc | 752 | vfio_unlink_dma(iommu, dma); |
8f0d5bb9 | 753 | put_task_struct(dma->task); |
1ef3e2bc AW |
754 | kfree(dma); |
755 | } | |
73fa0d10 | 756 | |
1ef3e2bc AW |
757 | static unsigned long vfio_pgsize_bitmap(struct vfio_iommu *iommu) |
758 | { | |
759 | struct vfio_domain *domain; | |
4644321f | 760 | unsigned long bitmap = ULONG_MAX; |
166fd7d9 | 761 | |
1ef3e2bc AW |
762 | mutex_lock(&iommu->lock); |
763 | list_for_each_entry(domain, &iommu->domain_list, next) | |
d16e0faa | 764 | bitmap &= domain->domain->pgsize_bitmap; |
1ef3e2bc | 765 | mutex_unlock(&iommu->lock); |
73fa0d10 | 766 | |
4644321f EA |
767 | /* |
768 | * In case the IOMMU supports page sizes smaller than PAGE_SIZE | |
769 | * we pretend PAGE_SIZE is supported and hide sub-PAGE_SIZE sizes. | |
770 | * That way the user will be able to map/unmap buffers whose size/ | |
771 | * start address is aligned with PAGE_SIZE. Pinning code uses that | |
772 | * granularity while iommu driver can use the sub-PAGE_SIZE size | |
773 | * to map the buffer. | |
774 | */ | |
775 | if (bitmap & ~PAGE_MASK) { | |
776 | bitmap &= PAGE_MASK; | |
777 | bitmap |= PAGE_SIZE; | |
778 | } | |
779 | ||
1ef3e2bc | 780 | return bitmap; |
73fa0d10 AW |
781 | } |
782 | ||
783 | static int vfio_dma_do_unmap(struct vfio_iommu *iommu, | |
784 | struct vfio_iommu_type1_dma_unmap *unmap) | |
785 | { | |
73fa0d10 | 786 | uint64_t mask; |
c086de81 | 787 | struct vfio_dma *dma, *dma_last = NULL; |
1ef3e2bc | 788 | size_t unmapped = 0; |
c086de81 | 789 | int ret = 0, retries = 0; |
73fa0d10 | 790 | |
1ef3e2bc | 791 | mask = ((uint64_t)1 << __ffs(vfio_pgsize_bitmap(iommu))) - 1; |
73fa0d10 AW |
792 | |
793 | if (unmap->iova & mask) | |
794 | return -EINVAL; | |
f5bfdbf2 | 795 | if (!unmap->size || unmap->size & mask) |
73fa0d10 AW |
796 | return -EINVAL; |
797 | ||
73fa0d10 | 798 | WARN_ON(mask & PAGE_MASK); |
c086de81 | 799 | again: |
73fa0d10 AW |
800 | mutex_lock(&iommu->lock); |
801 | ||
1ef3e2bc AW |
802 | /* |
803 | * vfio-iommu-type1 (v1) - User mappings were coalesced together to | |
804 | * avoid tracking individual mappings. This means that the granularity | |
805 | * of the original mapping was lost and the user was allowed to attempt | |
806 | * to unmap any range. Depending on the contiguousness of physical | |
807 | * memory and page sizes supported by the IOMMU, arbitrary unmaps may | |
808 | * or may not have worked. We only guaranteed unmap granularity | |
809 | * matching the original mapping; even though it was untracked here, | |
810 | * the original mappings are reflected in IOMMU mappings. This | |
811 | * resulted in a couple unusual behaviors. First, if a range is not | |
812 | * able to be unmapped, ex. a set of 4k pages that was mapped as a | |
813 | * 2M hugepage into the IOMMU, the unmap ioctl returns success but with | |
814 | * a zero sized unmap. Also, if an unmap request overlaps the first | |
815 | * address of a hugepage, the IOMMU will unmap the entire hugepage. | |
816 | * This also returns success and the returned unmap size reflects the | |
817 | * actual size unmapped. | |
818 | * | |
819 | * We attempt to maintain compatibility with this "v1" interface, but | |
820 | * we take control out of the hands of the IOMMU. Therefore, an unmap | |
821 | * request offset from the beginning of the original mapping will | |
822 | * return success with zero sized unmap. And an unmap request covering | |
823 | * the first iova of mapping will unmap the entire range. | |
824 | * | |
825 | * The v2 version of this interface intends to be more deterministic. | |
826 | * Unmap requests must fully cover previous mappings. Multiple | |
827 | * mappings may still be unmaped by specifying large ranges, but there | |
828 | * must not be any previous mappings bisected by the range. An error | |
829 | * will be returned if these conditions are not met. The v2 interface | |
830 | * will only return success and a size of zero if there were no | |
831 | * mappings within the range. | |
832 | */ | |
833 | if (iommu->v2) { | |
7c03f428 | 834 | dma = vfio_find_dma(iommu, unmap->iova, 1); |
1ef3e2bc AW |
835 | if (dma && dma->iova != unmap->iova) { |
836 | ret = -EINVAL; | |
837 | goto unlock; | |
838 | } | |
839 | dma = vfio_find_dma(iommu, unmap->iova + unmap->size - 1, 0); | |
840 | if (dma && dma->iova + dma->size != unmap->iova + unmap->size) { | |
841 | ret = -EINVAL; | |
842 | goto unlock; | |
843 | } | |
844 | } | |
845 | ||
166fd7d9 | 846 | while ((dma = vfio_find_dma(iommu, unmap->iova, unmap->size))) { |
1ef3e2bc | 847 | if (!iommu->v2 && unmap->iova > dma->iova) |
166fd7d9 | 848 | break; |
8f0d5bb9 KW |
849 | /* |
850 | * Task with same address space who mapped this iova range is | |
851 | * allowed to unmap the iova range. | |
852 | */ | |
853 | if (dma->task->mm != current->mm) | |
854 | break; | |
c086de81 KW |
855 | |
856 | if (!RB_EMPTY_ROOT(&dma->pfn_list)) { | |
857 | struct vfio_iommu_type1_dma_unmap nb_unmap; | |
858 | ||
859 | if (dma_last == dma) { | |
860 | BUG_ON(++retries > 10); | |
861 | } else { | |
862 | dma_last = dma; | |
863 | retries = 0; | |
864 | } | |
865 | ||
866 | nb_unmap.iova = dma->iova; | |
867 | nb_unmap.size = dma->size; | |
868 | ||
869 | /* | |
870 | * Notify anyone (mdev vendor drivers) to invalidate and | |
871 | * unmap iovas within the range we're about to unmap. | |
872 | * Vendor drivers MUST unpin pages in response to an | |
873 | * invalidation. | |
874 | */ | |
875 | mutex_unlock(&iommu->lock); | |
876 | blocking_notifier_call_chain(&iommu->notifier, | |
877 | VFIO_IOMMU_NOTIFY_DMA_UNMAP, | |
878 | &nb_unmap); | |
879 | goto again; | |
880 | } | |
1ef3e2bc AW |
881 | unmapped += dma->size; |
882 | vfio_remove_dma(iommu, dma); | |
166fd7d9 | 883 | } |
cd9b2268 | 884 | |
1ef3e2bc | 885 | unlock: |
73fa0d10 | 886 | mutex_unlock(&iommu->lock); |
166fd7d9 | 887 | |
1ef3e2bc | 888 | /* Report how much was unmapped */ |
166fd7d9 AW |
889 | unmap->size = unmapped; |
890 | ||
891 | return ret; | |
892 | } | |
893 | ||
894 | /* | |
895 | * Turns out AMD IOMMU has a page table bug where it won't map large pages | |
896 | * to a region that previously mapped smaller pages. This should be fixed | |
897 | * soon, so this is just a temporary workaround to break mappings down into | |
898 | * PAGE_SIZE. Better to map smaller pages than nothing. | |
899 | */ | |
1ef3e2bc | 900 | static int map_try_harder(struct vfio_domain *domain, dma_addr_t iova, |
166fd7d9 AW |
901 | unsigned long pfn, long npage, int prot) |
902 | { | |
903 | long i; | |
089f1c6b | 904 | int ret = 0; |
166fd7d9 AW |
905 | |
906 | for (i = 0; i < npage; i++, pfn++, iova += PAGE_SIZE) { | |
1ef3e2bc | 907 | ret = iommu_map(domain->domain, iova, |
166fd7d9 | 908 | (phys_addr_t)pfn << PAGE_SHIFT, |
1ef3e2bc | 909 | PAGE_SIZE, prot | domain->prot); |
166fd7d9 AW |
910 | if (ret) |
911 | break; | |
912 | } | |
913 | ||
914 | for (; i < npage && i > 0; i--, iova -= PAGE_SIZE) | |
1ef3e2bc AW |
915 | iommu_unmap(domain->domain, iova, PAGE_SIZE); |
916 | ||
917 | return ret; | |
918 | } | |
919 | ||
920 | static int vfio_iommu_map(struct vfio_iommu *iommu, dma_addr_t iova, | |
921 | unsigned long pfn, long npage, int prot) | |
922 | { | |
923 | struct vfio_domain *d; | |
924 | int ret; | |
925 | ||
926 | list_for_each_entry(d, &iommu->domain_list, next) { | |
927 | ret = iommu_map(d->domain, iova, (phys_addr_t)pfn << PAGE_SHIFT, | |
928 | npage << PAGE_SHIFT, prot | d->prot); | |
929 | if (ret) { | |
930 | if (ret != -EBUSY || | |
931 | map_try_harder(d, iova, pfn, npage, prot)) | |
932 | goto unwind; | |
933 | } | |
c5e66887 AW |
934 | |
935 | cond_resched(); | |
1ef3e2bc AW |
936 | } |
937 | ||
938 | return 0; | |
939 | ||
940 | unwind: | |
941 | list_for_each_entry_continue_reverse(d, &iommu->domain_list, next) | |
942 | iommu_unmap(d->domain, iova, npage << PAGE_SHIFT); | |
166fd7d9 | 943 | |
cd9b2268 | 944 | return ret; |
73fa0d10 AW |
945 | } |
946 | ||
8f0d5bb9 KW |
947 | static int vfio_pin_map_dma(struct vfio_iommu *iommu, struct vfio_dma *dma, |
948 | size_t map_size) | |
949 | { | |
950 | dma_addr_t iova = dma->iova; | |
951 | unsigned long vaddr = dma->vaddr; | |
952 | size_t size = map_size; | |
953 | long npage; | |
954 | unsigned long pfn; | |
955 | int ret = 0; | |
956 | ||
957 | while (size) { | |
958 | /* Pin a contiguous chunk of memory */ | |
959 | npage = vfio_pin_pages_remote(dma, vaddr + dma->size, | |
a54eb550 | 960 | size >> PAGE_SHIFT, &pfn); |
8f0d5bb9 KW |
961 | if (npage <= 0) { |
962 | WARN_ON(!npage); | |
963 | ret = (int)npage; | |
964 | break; | |
965 | } | |
966 | ||
967 | /* Map it! */ | |
968 | ret = vfio_iommu_map(iommu, iova + dma->size, pfn, npage, | |
969 | dma->prot); | |
970 | if (ret) { | |
a54eb550 KW |
971 | vfio_unpin_pages_remote(dma, iova + dma->size, pfn, |
972 | npage, true); | |
8f0d5bb9 KW |
973 | break; |
974 | } | |
975 | ||
976 | size -= npage << PAGE_SHIFT; | |
977 | dma->size += npage << PAGE_SHIFT; | |
978 | } | |
979 | ||
a54eb550 KW |
980 | dma->iommu_mapped = true; |
981 | ||
8f0d5bb9 KW |
982 | if (ret) |
983 | vfio_remove_dma(iommu, dma); | |
984 | ||
985 | return ret; | |
986 | } | |
987 | ||
73fa0d10 AW |
988 | static int vfio_dma_do_map(struct vfio_iommu *iommu, |
989 | struct vfio_iommu_type1_dma_map *map) | |
990 | { | |
c8dbca16 | 991 | dma_addr_t iova = map->iova; |
166fd7d9 | 992 | unsigned long vaddr = map->vaddr; |
73fa0d10 AW |
993 | size_t size = map->size; |
994 | int ret = 0, prot = 0; | |
995 | uint64_t mask; | |
1ef3e2bc | 996 | struct vfio_dma *dma; |
166fd7d9 | 997 | |
c8dbca16 AW |
998 | /* Verify that none of our __u64 fields overflow */ |
999 | if (map->size != size || map->vaddr != vaddr || map->iova != iova) | |
1000 | return -EINVAL; | |
73fa0d10 | 1001 | |
1ef3e2bc | 1002 | mask = ((uint64_t)1 << __ffs(vfio_pgsize_bitmap(iommu))) - 1; |
73fa0d10 | 1003 | |
c8dbca16 AW |
1004 | WARN_ON(mask & PAGE_MASK); |
1005 | ||
73fa0d10 AW |
1006 | /* READ/WRITE from device perspective */ |
1007 | if (map->flags & VFIO_DMA_MAP_FLAG_WRITE) | |
1008 | prot |= IOMMU_WRITE; | |
1009 | if (map->flags & VFIO_DMA_MAP_FLAG_READ) | |
1010 | prot |= IOMMU_READ; | |
1011 | ||
c8dbca16 | 1012 | if (!prot || !size || (size | iova | vaddr) & mask) |
73fa0d10 AW |
1013 | return -EINVAL; |
1014 | ||
c8dbca16 AW |
1015 | /* Don't allow IOVA or virtual address wrap */ |
1016 | if (iova + size - 1 < iova || vaddr + size - 1 < vaddr) | |
73fa0d10 AW |
1017 | return -EINVAL; |
1018 | ||
1019 | mutex_lock(&iommu->lock); | |
1020 | ||
c8dbca16 | 1021 | if (vfio_find_dma(iommu, iova, size)) { |
8f0d5bb9 KW |
1022 | ret = -EEXIST; |
1023 | goto out_unlock; | |
73fa0d10 AW |
1024 | } |
1025 | ||
1ef3e2bc AW |
1026 | dma = kzalloc(sizeof(*dma), GFP_KERNEL); |
1027 | if (!dma) { | |
8f0d5bb9 KW |
1028 | ret = -ENOMEM; |
1029 | goto out_unlock; | |
1ef3e2bc AW |
1030 | } |
1031 | ||
c8dbca16 AW |
1032 | dma->iova = iova; |
1033 | dma->vaddr = vaddr; | |
1ef3e2bc | 1034 | dma->prot = prot; |
8f0d5bb9 KW |
1035 | get_task_struct(current); |
1036 | dma->task = current; | |
a54eb550 | 1037 | dma->pfn_list = RB_ROOT; |
166fd7d9 | 1038 | |
1ef3e2bc AW |
1039 | /* Insert zero-sized and grow as we map chunks of it */ |
1040 | vfio_link_dma(iommu, dma); | |
166fd7d9 | 1041 | |
a54eb550 KW |
1042 | /* Don't pin and map if container doesn't contain IOMMU capable domain*/ |
1043 | if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu)) | |
1044 | dma->size = size; | |
1045 | else | |
1046 | ret = vfio_pin_map_dma(iommu, dma, size); | |
1047 | ||
8f0d5bb9 | 1048 | out_unlock: |
1ef3e2bc AW |
1049 | mutex_unlock(&iommu->lock); |
1050 | return ret; | |
1051 | } | |
1052 | ||
1053 | static int vfio_bus_type(struct device *dev, void *data) | |
1054 | { | |
1055 | struct bus_type **bus = data; | |
1056 | ||
1057 | if (*bus && *bus != dev->bus) | |
1058 | return -EINVAL; | |
1059 | ||
1060 | *bus = dev->bus; | |
1061 | ||
1062 | return 0; | |
1063 | } | |
1064 | ||
1065 | static int vfio_iommu_replay(struct vfio_iommu *iommu, | |
1066 | struct vfio_domain *domain) | |
1067 | { | |
1068 | struct vfio_domain *d; | |
1069 | struct rb_node *n; | |
1070 | int ret; | |
1071 | ||
1072 | /* Arbitrarily pick the first domain in the list for lookups */ | |
1073 | d = list_first_entry(&iommu->domain_list, struct vfio_domain, next); | |
1074 | n = rb_first(&iommu->dma_list); | |
1075 | ||
1ef3e2bc AW |
1076 | for (; n; n = rb_next(n)) { |
1077 | struct vfio_dma *dma; | |
1078 | dma_addr_t iova; | |
1079 | ||
1080 | dma = rb_entry(n, struct vfio_dma, node); | |
1081 | iova = dma->iova; | |
1082 | ||
1083 | while (iova < dma->iova + dma->size) { | |
a54eb550 | 1084 | phys_addr_t phys; |
1ef3e2bc | 1085 | size_t size; |
73fa0d10 | 1086 | |
a54eb550 KW |
1087 | if (dma->iommu_mapped) { |
1088 | phys_addr_t p; | |
1089 | dma_addr_t i; | |
1090 | ||
1091 | phys = iommu_iova_to_phys(d->domain, iova); | |
1092 | ||
1093 | if (WARN_ON(!phys)) { | |
1094 | iova += PAGE_SIZE; | |
1095 | continue; | |
1096 | } | |
1097 | ||
1098 | size = PAGE_SIZE; | |
1099 | p = phys + size; | |
1100 | i = iova + size; | |
1101 | while (i < dma->iova + dma->size && | |
1102 | p == iommu_iova_to_phys(d->domain, i)) { | |
1103 | size += PAGE_SIZE; | |
1104 | p += PAGE_SIZE; | |
1105 | i += PAGE_SIZE; | |
1106 | } | |
1107 | } else { | |
1108 | unsigned long pfn; | |
1109 | unsigned long vaddr = dma->vaddr + | |
1110 | (iova - dma->iova); | |
1111 | size_t n = dma->iova + dma->size - iova; | |
1112 | long npage; | |
1113 | ||
1114 | npage = vfio_pin_pages_remote(dma, vaddr, | |
1115 | n >> PAGE_SHIFT, | |
1116 | &pfn); | |
1117 | if (npage <= 0) { | |
1118 | WARN_ON(!npage); | |
1119 | ret = (int)npage; | |
1120 | return ret; | |
1121 | } | |
1122 | ||
1123 | phys = pfn << PAGE_SHIFT; | |
1124 | size = npage << PAGE_SHIFT; | |
166fd7d9 AW |
1125 | } |
1126 | ||
1ef3e2bc AW |
1127 | ret = iommu_map(domain->domain, iova, phys, |
1128 | size, dma->prot | domain->prot); | |
1129 | if (ret) | |
1130 | return ret; | |
d93b3ac0 | 1131 | |
1ef3e2bc AW |
1132 | iova += size; |
1133 | } | |
a54eb550 | 1134 | dma->iommu_mapped = true; |
166fd7d9 | 1135 | } |
1ef3e2bc | 1136 | return 0; |
73fa0d10 AW |
1137 | } |
1138 | ||
6fe1010d AW |
1139 | /* |
1140 | * We change our unmap behavior slightly depending on whether the IOMMU | |
1141 | * supports fine-grained superpages. IOMMUs like AMD-Vi will use a superpage | |
1142 | * for practically any contiguous power-of-two mapping we give it. This means | |
1143 | * we don't need to look for contiguous chunks ourselves to make unmapping | |
1144 | * more efficient. On IOMMUs with coarse-grained super pages, like Intel VT-d | |
1145 | * with discrete 2M/1G/512G/1T superpages, identifying contiguous chunks | |
1146 | * significantly boosts non-hugetlbfs mappings and doesn't seem to hurt when | |
1147 | * hugetlbfs is in use. | |
1148 | */ | |
1149 | static void vfio_test_domain_fgsp(struct vfio_domain *domain) | |
1150 | { | |
1151 | struct page *pages; | |
1152 | int ret, order = get_order(PAGE_SIZE * 2); | |
1153 | ||
1154 | pages = alloc_pages(GFP_KERNEL | __GFP_ZERO, order); | |
1155 | if (!pages) | |
1156 | return; | |
1157 | ||
1158 | ret = iommu_map(domain->domain, 0, page_to_phys(pages), PAGE_SIZE * 2, | |
1159 | IOMMU_READ | IOMMU_WRITE | domain->prot); | |
1160 | if (!ret) { | |
1161 | size_t unmapped = iommu_unmap(domain->domain, 0, PAGE_SIZE); | |
1162 | ||
1163 | if (unmapped == PAGE_SIZE) | |
1164 | iommu_unmap(domain->domain, PAGE_SIZE, PAGE_SIZE); | |
1165 | else | |
1166 | domain->fgsp = true; | |
1167 | } | |
1168 | ||
1169 | __free_pages(pages, order); | |
1170 | } | |
1171 | ||
7896c998 KW |
1172 | static struct vfio_group *find_iommu_group(struct vfio_domain *domain, |
1173 | struct iommu_group *iommu_group) | |
1174 | { | |
1175 | struct vfio_group *g; | |
1176 | ||
1177 | list_for_each_entry(g, &domain->group_list, next) { | |
1178 | if (g->iommu_group == iommu_group) | |
1179 | return g; | |
1180 | } | |
1181 | ||
1182 | return NULL; | |
1183 | } | |
1184 | ||
9d3a4de4 | 1185 | static bool vfio_iommu_has_sw_msi(struct iommu_group *group, phys_addr_t *base) |
5d704992 EA |
1186 | { |
1187 | struct list_head group_resv_regions; | |
1188 | struct iommu_resv_region *region, *next; | |
1189 | bool ret = false; | |
1190 | ||
1191 | INIT_LIST_HEAD(&group_resv_regions); | |
1192 | iommu_get_group_resv_regions(group, &group_resv_regions); | |
1193 | list_for_each_entry(region, &group_resv_regions, list) { | |
9d3a4de4 | 1194 | if (region->type == IOMMU_RESV_SW_MSI) { |
5d704992 EA |
1195 | *base = region->start; |
1196 | ret = true; | |
1197 | goto out; | |
1198 | } | |
1199 | } | |
1200 | out: | |
1201 | list_for_each_entry_safe(region, next, &group_resv_regions, list) | |
1202 | kfree(region); | |
1203 | return ret; | |
1204 | } | |
1205 | ||
73fa0d10 AW |
1206 | static int vfio_iommu_type1_attach_group(void *iommu_data, |
1207 | struct iommu_group *iommu_group) | |
1208 | { | |
1209 | struct vfio_iommu *iommu = iommu_data; | |
7896c998 | 1210 | struct vfio_group *group; |
1ef3e2bc | 1211 | struct vfio_domain *domain, *d; |
a54eb550 | 1212 | struct bus_type *bus = NULL, *mdev_bus; |
73fa0d10 | 1213 | int ret; |
9d72f87b | 1214 | bool resv_msi, msi_remap; |
5d704992 | 1215 | phys_addr_t resv_msi_base; |
73fa0d10 | 1216 | |
73fa0d10 AW |
1217 | mutex_lock(&iommu->lock); |
1218 | ||
1ef3e2bc | 1219 | list_for_each_entry(d, &iommu->domain_list, next) { |
7896c998 | 1220 | if (find_iommu_group(d, iommu_group)) { |
73fa0d10 | 1221 | mutex_unlock(&iommu->lock); |
73fa0d10 AW |
1222 | return -EINVAL; |
1223 | } | |
1224 | } | |
1225 | ||
a54eb550 KW |
1226 | if (iommu->external_domain) { |
1227 | if (find_iommu_group(iommu->external_domain, iommu_group)) { | |
1228 | mutex_unlock(&iommu->lock); | |
1229 | return -EINVAL; | |
1230 | } | |
1231 | } | |
1232 | ||
1ef3e2bc AW |
1233 | group = kzalloc(sizeof(*group), GFP_KERNEL); |
1234 | domain = kzalloc(sizeof(*domain), GFP_KERNEL); | |
1235 | if (!group || !domain) { | |
1236 | ret = -ENOMEM; | |
1237 | goto out_free; | |
1238 | } | |
1239 | ||
1240 | group->iommu_group = iommu_group; | |
1241 | ||
1242 | /* Determine bus_type in order to allocate a domain */ | |
1243 | ret = iommu_group_for_each_dev(iommu_group, &bus, vfio_bus_type); | |
1244 | if (ret) | |
1245 | goto out_free; | |
1246 | ||
a54eb550 KW |
1247 | mdev_bus = symbol_get(mdev_bus_type); |
1248 | ||
1249 | if (mdev_bus) { | |
1250 | if ((bus == mdev_bus) && !iommu_present(bus)) { | |
1251 | symbol_put(mdev_bus_type); | |
1252 | if (!iommu->external_domain) { | |
1253 | INIT_LIST_HEAD(&domain->group_list); | |
1254 | iommu->external_domain = domain; | |
1255 | } else | |
1256 | kfree(domain); | |
1257 | ||
1258 | list_add(&group->next, | |
1259 | &iommu->external_domain->group_list); | |
1260 | mutex_unlock(&iommu->lock); | |
1261 | return 0; | |
1262 | } | |
1263 | symbol_put(mdev_bus_type); | |
1264 | } | |
1265 | ||
1ef3e2bc AW |
1266 | domain->domain = iommu_domain_alloc(bus); |
1267 | if (!domain->domain) { | |
1268 | ret = -EIO; | |
1269 | goto out_free; | |
1270 | } | |
1271 | ||
f5c9eceb WD |
1272 | if (iommu->nesting) { |
1273 | int attr = 1; | |
1274 | ||
1275 | ret = iommu_domain_set_attr(domain->domain, DOMAIN_ATTR_NESTING, | |
1276 | &attr); | |
1277 | if (ret) | |
1278 | goto out_domain; | |
1279 | } | |
1280 | ||
1ef3e2bc AW |
1281 | ret = iommu_attach_group(domain->domain, iommu_group); |
1282 | if (ret) | |
1283 | goto out_domain; | |
1284 | ||
9d3a4de4 | 1285 | resv_msi = vfio_iommu_has_sw_msi(iommu_group, &resv_msi_base); |
5d704992 | 1286 | |
1ef3e2bc AW |
1287 | INIT_LIST_HEAD(&domain->group_list); |
1288 | list_add(&group->next, &domain->group_list); | |
1289 | ||
9d72f87b EA |
1290 | msi_remap = resv_msi ? irq_domain_check_msi_remap() : |
1291 | iommu_capable(bus, IOMMU_CAP_INTR_REMAP); | |
1292 | ||
1293 | if (!allow_unsafe_interrupts && !msi_remap) { | |
1ef3e2bc AW |
1294 | pr_warn("%s: No interrupt remapping support. Use the module param \"allow_unsafe_interrupts\" to enable VFIO IOMMU support on this platform\n", |
1295 | __func__); | |
1296 | ret = -EPERM; | |
1297 | goto out_detach; | |
1298 | } | |
1299 | ||
eb165f05 | 1300 | if (iommu_capable(bus, IOMMU_CAP_CACHE_COHERENCY)) |
1ef3e2bc AW |
1301 | domain->prot |= IOMMU_CACHE; |
1302 | ||
73fa0d10 | 1303 | /* |
1ef3e2bc AW |
1304 | * Try to match an existing compatible domain. We don't want to |
1305 | * preclude an IOMMU driver supporting multiple bus_types and being | |
1306 | * able to include different bus_types in the same IOMMU domain, so | |
1307 | * we test whether the domains use the same iommu_ops rather than | |
1308 | * testing if they're on the same bus_type. | |
73fa0d10 | 1309 | */ |
1ef3e2bc AW |
1310 | list_for_each_entry(d, &iommu->domain_list, next) { |
1311 | if (d->domain->ops == domain->domain->ops && | |
1312 | d->prot == domain->prot) { | |
1313 | iommu_detach_group(domain->domain, iommu_group); | |
1314 | if (!iommu_attach_group(d->domain, iommu_group)) { | |
1315 | list_add(&group->next, &d->group_list); | |
1316 | iommu_domain_free(domain->domain); | |
1317 | kfree(domain); | |
1318 | mutex_unlock(&iommu->lock); | |
1319 | return 0; | |
1320 | } | |
1321 | ||
1322 | ret = iommu_attach_group(domain->domain, iommu_group); | |
1323 | if (ret) | |
1324 | goto out_domain; | |
1325 | } | |
73fa0d10 AW |
1326 | } |
1327 | ||
6fe1010d AW |
1328 | vfio_test_domain_fgsp(domain); |
1329 | ||
1ef3e2bc AW |
1330 | /* replay mappings on new domains */ |
1331 | ret = vfio_iommu_replay(iommu, domain); | |
1332 | if (ret) | |
1333 | goto out_detach; | |
1334 | ||
2c9f1af5 WY |
1335 | if (resv_msi) { |
1336 | ret = iommu_get_msi_cookie(domain->domain, resv_msi_base); | |
1337 | if (ret) | |
1338 | goto out_detach; | |
1339 | } | |
5d704992 | 1340 | |
1ef3e2bc | 1341 | list_add(&domain->next, &iommu->domain_list); |
73fa0d10 AW |
1342 | |
1343 | mutex_unlock(&iommu->lock); | |
1344 | ||
1345 | return 0; | |
1ef3e2bc AW |
1346 | |
1347 | out_detach: | |
1348 | iommu_detach_group(domain->domain, iommu_group); | |
1349 | out_domain: | |
1350 | iommu_domain_free(domain->domain); | |
1351 | out_free: | |
1352 | kfree(domain); | |
1353 | kfree(group); | |
1354 | mutex_unlock(&iommu->lock); | |
1355 | return ret; | |
1356 | } | |
1357 | ||
1358 | static void vfio_iommu_unmap_unpin_all(struct vfio_iommu *iommu) | |
1359 | { | |
1360 | struct rb_node *node; | |
1361 | ||
1362 | while ((node = rb_first(&iommu->dma_list))) | |
1363 | vfio_remove_dma(iommu, rb_entry(node, struct vfio_dma, node)); | |
73fa0d10 AW |
1364 | } |
1365 | ||
a54eb550 KW |
1366 | static void vfio_iommu_unmap_unpin_reaccount(struct vfio_iommu *iommu) |
1367 | { | |
1368 | struct rb_node *n, *p; | |
1369 | ||
1370 | n = rb_first(&iommu->dma_list); | |
1371 | for (; n; n = rb_next(n)) { | |
1372 | struct vfio_dma *dma; | |
1373 | long locked = 0, unlocked = 0; | |
1374 | ||
1375 | dma = rb_entry(n, struct vfio_dma, node); | |
1376 | unlocked += vfio_unmap_unpin(iommu, dma, false); | |
1377 | p = rb_first(&dma->pfn_list); | |
1378 | for (; p; p = rb_next(p)) { | |
1379 | struct vfio_pfn *vpfn = rb_entry(p, struct vfio_pfn, | |
1380 | node); | |
1381 | ||
1382 | if (!is_invalid_reserved_pfn(vpfn->pfn)) | |
1383 | locked++; | |
1384 | } | |
1385 | vfio_lock_acct(dma->task, locked - unlocked); | |
1386 | } | |
1387 | } | |
1388 | ||
1389 | static void vfio_sanity_check_pfn_list(struct vfio_iommu *iommu) | |
1390 | { | |
1391 | struct rb_node *n; | |
1392 | ||
1393 | n = rb_first(&iommu->dma_list); | |
1394 | for (; n; n = rb_next(n)) { | |
1395 | struct vfio_dma *dma; | |
1396 | ||
1397 | dma = rb_entry(n, struct vfio_dma, node); | |
1398 | ||
1399 | if (WARN_ON(!RB_EMPTY_ROOT(&dma->pfn_list))) | |
1400 | break; | |
1401 | } | |
3cedd7d7 KW |
1402 | /* mdev vendor driver must unregister notifier */ |
1403 | WARN_ON(iommu->notifier.head); | |
a54eb550 KW |
1404 | } |
1405 | ||
73fa0d10 AW |
1406 | static void vfio_iommu_type1_detach_group(void *iommu_data, |
1407 | struct iommu_group *iommu_group) | |
1408 | { | |
1409 | struct vfio_iommu *iommu = iommu_data; | |
1ef3e2bc | 1410 | struct vfio_domain *domain; |
73fa0d10 AW |
1411 | struct vfio_group *group; |
1412 | ||
1413 | mutex_lock(&iommu->lock); | |
1414 | ||
a54eb550 KW |
1415 | if (iommu->external_domain) { |
1416 | group = find_iommu_group(iommu->external_domain, iommu_group); | |
1417 | if (group) { | |
1418 | list_del(&group->next); | |
1419 | kfree(group); | |
1420 | ||
1421 | if (list_empty(&iommu->external_domain->group_list)) { | |
1422 | vfio_sanity_check_pfn_list(iommu); | |
1423 | ||
1424 | if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu)) | |
1425 | vfio_iommu_unmap_unpin_all(iommu); | |
1426 | ||
1427 | kfree(iommu->external_domain); | |
1428 | iommu->external_domain = NULL; | |
1429 | } | |
1430 | goto detach_group_done; | |
1431 | } | |
1432 | } | |
1433 | ||
1ef3e2bc | 1434 | list_for_each_entry(domain, &iommu->domain_list, next) { |
7896c998 KW |
1435 | group = find_iommu_group(domain, iommu_group); |
1436 | if (!group) | |
1437 | continue; | |
1ef3e2bc | 1438 | |
7896c998 KW |
1439 | iommu_detach_group(domain->domain, iommu_group); |
1440 | list_del(&group->next); | |
1441 | kfree(group); | |
1442 | /* | |
a54eb550 KW |
1443 | * Group ownership provides privilege, if the group list is |
1444 | * empty, the domain goes away. If it's the last domain with | |
1445 | * iommu and external domain doesn't exist, then all the | |
1446 | * mappings go away too. If it's the last domain with iommu and | |
1447 | * external domain exist, update accounting | |
7896c998 KW |
1448 | */ |
1449 | if (list_empty(&domain->group_list)) { | |
a54eb550 KW |
1450 | if (list_is_singular(&iommu->domain_list)) { |
1451 | if (!iommu->external_domain) | |
1452 | vfio_iommu_unmap_unpin_all(iommu); | |
1453 | else | |
1454 | vfio_iommu_unmap_unpin_reaccount(iommu); | |
1455 | } | |
7896c998 KW |
1456 | iommu_domain_free(domain->domain); |
1457 | list_del(&domain->next); | |
1458 | kfree(domain); | |
73fa0d10 | 1459 | } |
a54eb550 | 1460 | break; |
73fa0d10 AW |
1461 | } |
1462 | ||
a54eb550 | 1463 | detach_group_done: |
73fa0d10 AW |
1464 | mutex_unlock(&iommu->lock); |
1465 | } | |
1466 | ||
1467 | static void *vfio_iommu_type1_open(unsigned long arg) | |
1468 | { | |
1469 | struct vfio_iommu *iommu; | |
1470 | ||
73fa0d10 AW |
1471 | iommu = kzalloc(sizeof(*iommu), GFP_KERNEL); |
1472 | if (!iommu) | |
1473 | return ERR_PTR(-ENOMEM); | |
1474 | ||
f5c9eceb WD |
1475 | switch (arg) { |
1476 | case VFIO_TYPE1_IOMMU: | |
1477 | break; | |
1478 | case VFIO_TYPE1_NESTING_IOMMU: | |
1479 | iommu->nesting = true; | |
1480 | case VFIO_TYPE1v2_IOMMU: | |
1481 | iommu->v2 = true; | |
1482 | break; | |
1483 | default: | |
1484 | kfree(iommu); | |
1485 | return ERR_PTR(-EINVAL); | |
1486 | } | |
1487 | ||
1ef3e2bc | 1488 | INIT_LIST_HEAD(&iommu->domain_list); |
cd9b2268 | 1489 | iommu->dma_list = RB_ROOT; |
73fa0d10 | 1490 | mutex_init(&iommu->lock); |
c086de81 | 1491 | BLOCKING_INIT_NOTIFIER_HEAD(&iommu->notifier); |
73fa0d10 AW |
1492 | |
1493 | return iommu; | |
1494 | } | |
1495 | ||
a54eb550 KW |
1496 | static void vfio_release_domain(struct vfio_domain *domain, bool external) |
1497 | { | |
1498 | struct vfio_group *group, *group_tmp; | |
1499 | ||
1500 | list_for_each_entry_safe(group, group_tmp, | |
1501 | &domain->group_list, next) { | |
1502 | if (!external) | |
1503 | iommu_detach_group(domain->domain, group->iommu_group); | |
1504 | list_del(&group->next); | |
1505 | kfree(group); | |
1506 | } | |
1507 | ||
1508 | if (!external) | |
1509 | iommu_domain_free(domain->domain); | |
1510 | } | |
1511 | ||
73fa0d10 AW |
1512 | static void vfio_iommu_type1_release(void *iommu_data) |
1513 | { | |
1514 | struct vfio_iommu *iommu = iommu_data; | |
1ef3e2bc | 1515 | struct vfio_domain *domain, *domain_tmp; |
a54eb550 KW |
1516 | |
1517 | if (iommu->external_domain) { | |
1518 | vfio_release_domain(iommu->external_domain, true); | |
1519 | vfio_sanity_check_pfn_list(iommu); | |
1520 | kfree(iommu->external_domain); | |
1521 | } | |
73fa0d10 | 1522 | |
1ef3e2bc | 1523 | vfio_iommu_unmap_unpin_all(iommu); |
73fa0d10 | 1524 | |
1ef3e2bc AW |
1525 | list_for_each_entry_safe(domain, domain_tmp, |
1526 | &iommu->domain_list, next) { | |
a54eb550 | 1527 | vfio_release_domain(domain, false); |
1ef3e2bc AW |
1528 | list_del(&domain->next); |
1529 | kfree(domain); | |
73fa0d10 | 1530 | } |
73fa0d10 AW |
1531 | kfree(iommu); |
1532 | } | |
1533 | ||
aa429318 AW |
1534 | static int vfio_domains_have_iommu_cache(struct vfio_iommu *iommu) |
1535 | { | |
1536 | struct vfio_domain *domain; | |
1537 | int ret = 1; | |
1538 | ||
1539 | mutex_lock(&iommu->lock); | |
1540 | list_for_each_entry(domain, &iommu->domain_list, next) { | |
1541 | if (!(domain->prot & IOMMU_CACHE)) { | |
1542 | ret = 0; | |
f5bfdbf2 | 1543 | break; |
aa429318 | 1544 | } |
73fa0d10 | 1545 | } |
aa429318 | 1546 | mutex_unlock(&iommu->lock); |
73fa0d10 | 1547 | |
aa429318 | 1548 | return ret; |
73fa0d10 AW |
1549 | } |
1550 | ||
1551 | static long vfio_iommu_type1_ioctl(void *iommu_data, | |
1552 | unsigned int cmd, unsigned long arg) | |
1553 | { | |
1554 | struct vfio_iommu *iommu = iommu_data; | |
1555 | unsigned long minsz; | |
1556 | ||
1557 | if (cmd == VFIO_CHECK_EXTENSION) { | |
1558 | switch (arg) { | |
1559 | case VFIO_TYPE1_IOMMU: | |
1ef3e2bc | 1560 | case VFIO_TYPE1v2_IOMMU: |
f5c9eceb | 1561 | case VFIO_TYPE1_NESTING_IOMMU: |
73fa0d10 | 1562 | return 1; |
aa429318 AW |
1563 | case VFIO_DMA_CC_IOMMU: |
1564 | if (!iommu) | |
1565 | return 0; | |
1566 | return vfio_domains_have_iommu_cache(iommu); | |
73fa0d10 AW |
1567 | default: |
1568 | return 0; | |
1569 | } | |
1570 | } else if (cmd == VFIO_IOMMU_GET_INFO) { | |
1571 | struct vfio_iommu_type1_info info; | |
1572 | ||
1573 | minsz = offsetofend(struct vfio_iommu_type1_info, iova_pgsizes); | |
1574 | ||
1575 | if (copy_from_user(&info, (void __user *)arg, minsz)) | |
1576 | return -EFAULT; | |
1577 | ||
1578 | if (info.argsz < minsz) | |
1579 | return -EINVAL; | |
1580 | ||
d4f50ee2 | 1581 | info.flags = VFIO_IOMMU_INFO_PGSIZES; |
73fa0d10 | 1582 | |
1ef3e2bc | 1583 | info.iova_pgsizes = vfio_pgsize_bitmap(iommu); |
73fa0d10 | 1584 | |
8160c4e4 MT |
1585 | return copy_to_user((void __user *)arg, &info, minsz) ? |
1586 | -EFAULT : 0; | |
73fa0d10 AW |
1587 | |
1588 | } else if (cmd == VFIO_IOMMU_MAP_DMA) { | |
1589 | struct vfio_iommu_type1_dma_map map; | |
1590 | uint32_t mask = VFIO_DMA_MAP_FLAG_READ | | |
1591 | VFIO_DMA_MAP_FLAG_WRITE; | |
1592 | ||
1593 | minsz = offsetofend(struct vfio_iommu_type1_dma_map, size); | |
1594 | ||
1595 | if (copy_from_user(&map, (void __user *)arg, minsz)) | |
1596 | return -EFAULT; | |
1597 | ||
1598 | if (map.argsz < minsz || map.flags & ~mask) | |
1599 | return -EINVAL; | |
1600 | ||
1601 | return vfio_dma_do_map(iommu, &map); | |
1602 | ||
1603 | } else if (cmd == VFIO_IOMMU_UNMAP_DMA) { | |
1604 | struct vfio_iommu_type1_dma_unmap unmap; | |
166fd7d9 | 1605 | long ret; |
73fa0d10 AW |
1606 | |
1607 | minsz = offsetofend(struct vfio_iommu_type1_dma_unmap, size); | |
1608 | ||
1609 | if (copy_from_user(&unmap, (void __user *)arg, minsz)) | |
1610 | return -EFAULT; | |
1611 | ||
1612 | if (unmap.argsz < minsz || unmap.flags) | |
1613 | return -EINVAL; | |
1614 | ||
166fd7d9 AW |
1615 | ret = vfio_dma_do_unmap(iommu, &unmap); |
1616 | if (ret) | |
1617 | return ret; | |
1618 | ||
8160c4e4 MT |
1619 | return copy_to_user((void __user *)arg, &unmap, minsz) ? |
1620 | -EFAULT : 0; | |
73fa0d10 AW |
1621 | } |
1622 | ||
1623 | return -ENOTTY; | |
1624 | } | |
1625 | ||
c086de81 | 1626 | static int vfio_iommu_type1_register_notifier(void *iommu_data, |
22195cbd | 1627 | unsigned long *events, |
c086de81 KW |
1628 | struct notifier_block *nb) |
1629 | { | |
1630 | struct vfio_iommu *iommu = iommu_data; | |
1631 | ||
22195cbd JS |
1632 | /* clear known events */ |
1633 | *events &= ~VFIO_IOMMU_NOTIFY_DMA_UNMAP; | |
1634 | ||
1635 | /* refuse to register if still events remaining */ | |
1636 | if (*events) | |
1637 | return -EINVAL; | |
1638 | ||
c086de81 KW |
1639 | return blocking_notifier_chain_register(&iommu->notifier, nb); |
1640 | } | |
1641 | ||
1642 | static int vfio_iommu_type1_unregister_notifier(void *iommu_data, | |
1643 | struct notifier_block *nb) | |
1644 | { | |
1645 | struct vfio_iommu *iommu = iommu_data; | |
1646 | ||
1647 | return blocking_notifier_chain_unregister(&iommu->notifier, nb); | |
1648 | } | |
1649 | ||
73fa0d10 | 1650 | static const struct vfio_iommu_driver_ops vfio_iommu_driver_ops_type1 = { |
c086de81 KW |
1651 | .name = "vfio-iommu-type1", |
1652 | .owner = THIS_MODULE, | |
1653 | .open = vfio_iommu_type1_open, | |
1654 | .release = vfio_iommu_type1_release, | |
1655 | .ioctl = vfio_iommu_type1_ioctl, | |
1656 | .attach_group = vfio_iommu_type1_attach_group, | |
1657 | .detach_group = vfio_iommu_type1_detach_group, | |
1658 | .pin_pages = vfio_iommu_type1_pin_pages, | |
1659 | .unpin_pages = vfio_iommu_type1_unpin_pages, | |
1660 | .register_notifier = vfio_iommu_type1_register_notifier, | |
1661 | .unregister_notifier = vfio_iommu_type1_unregister_notifier, | |
73fa0d10 AW |
1662 | }; |
1663 | ||
1664 | static int __init vfio_iommu_type1_init(void) | |
1665 | { | |
73fa0d10 AW |
1666 | return vfio_register_iommu_driver(&vfio_iommu_driver_ops_type1); |
1667 | } | |
1668 | ||
1669 | static void __exit vfio_iommu_type1_cleanup(void) | |
1670 | { | |
1671 | vfio_unregister_iommu_driver(&vfio_iommu_driver_ops_type1); | |
1672 | } | |
1673 | ||
1674 | module_init(vfio_iommu_type1_init); | |
1675 | module_exit(vfio_iommu_type1_cleanup); | |
1676 | ||
1677 | MODULE_VERSION(DRIVER_VERSION); | |
1678 | MODULE_LICENSE("GPL v2"); | |
1679 | MODULE_AUTHOR(DRIVER_AUTHOR); | |
1680 | MODULE_DESCRIPTION(DRIVER_DESC); |