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