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