]>
Commit | Line | Data |
---|---|---|
73fa0d10 AW |
1 | /* |
2 | * VFIO: IOMMU DMA mapping support for Type1 IOMMU | |
3 | * | |
4 | * Copyright (C) 2012 Red Hat, Inc. All rights reserved. | |
5 | * Author: Alex Williamson <alex.williamson@redhat.com> | |
6 | * | |
7 | * This program is free software; you can redistribute it and/or modify | |
8 | * it under the terms of the GNU General Public License version 2 as | |
9 | * published by the Free Software Foundation. | |
10 | * | |
11 | * Derived from original vfio: | |
12 | * Copyright 2010 Cisco Systems, Inc. All rights reserved. | |
13 | * Author: Tom Lyon, pugs@cisco.com | |
14 | * | |
15 | * We arbitrarily define a Type1 IOMMU as one matching the below code. | |
16 | * It could be called the x86 IOMMU as it's designed for AMD-Vi & Intel | |
17 | * VT-d, but that makes it harder to re-use as theoretically anyone | |
18 | * implementing a similar IOMMU could make use of this. We expect the | |
19 | * IOMMU to support the IOMMU API and have few to no restrictions around | |
20 | * the IOVA range that can be mapped. The Type1 IOMMU is currently | |
21 | * optimized for relatively static mappings of a userspace process with | |
22 | * userpsace pages pinned into memory. We also assume devices and IOMMU | |
23 | * domains are PCI based as the IOMMU API is still centered around a | |
24 | * device/bus interface rather than a group interface. | |
25 | */ | |
26 | ||
27 | #include <linux/compat.h> | |
28 | #include <linux/device.h> | |
29 | #include <linux/fs.h> | |
30 | #include <linux/iommu.h> | |
31 | #include <linux/module.h> | |
32 | #include <linux/mm.h> | |
cd9b2268 | 33 | #include <linux/rbtree.h> |
3f07c014 | 34 | #include <linux/sched/signal.h> |
6e84f315 | 35 | #include <linux/sched/mm.h> |
73fa0d10 AW |
36 | #include <linux/slab.h> |
37 | #include <linux/uaccess.h> | |
38 | #include <linux/vfio.h> | |
39 | #include <linux/workqueue.h> | |
a54eb550 | 40 | #include <linux/mdev.h> |
c086de81 | 41 | #include <linux/notifier.h> |
5d704992 | 42 | #include <linux/dma-iommu.h> |
9d72f87b | 43 | #include <linux/irqdomain.h> |
73fa0d10 AW |
44 | |
45 | #define DRIVER_VERSION "0.2" | |
46 | #define DRIVER_AUTHOR "Alex Williamson <alex.williamson@redhat.com>" | |
47 | #define DRIVER_DESC "Type1 IOMMU driver for VFIO" | |
48 | ||
49 | static bool allow_unsafe_interrupts; | |
50 | module_param_named(allow_unsafe_interrupts, | |
51 | allow_unsafe_interrupts, bool, S_IRUGO | S_IWUSR); | |
52 | MODULE_PARM_DESC(allow_unsafe_interrupts, | |
53 | "Enable VFIO IOMMU support for on platforms without interrupt remapping support."); | |
54 | ||
5c6c2b21 AW |
55 | static bool disable_hugepages; |
56 | module_param_named(disable_hugepages, | |
57 | disable_hugepages, bool, S_IRUGO | S_IWUSR); | |
58 | MODULE_PARM_DESC(disable_hugepages, | |
59 | "Disable VFIO IOMMU support for IOMMU hugepages."); | |
60 | ||
49285593 AW |
61 | static unsigned int dma_entry_limit __read_mostly = U16_MAX; |
62 | module_param_named(dma_entry_limit, dma_entry_limit, uint, 0644); | |
63 | MODULE_PARM_DESC(dma_entry_limit, | |
64 | "Maximum number of user DMA mappings per container (65535)."); | |
65 | ||
73fa0d10 | 66 | struct vfio_iommu { |
1ef3e2bc | 67 | struct list_head domain_list; |
a54eb550 | 68 | struct vfio_domain *external_domain; /* domain for external user */ |
73fa0d10 | 69 | struct mutex lock; |
cd9b2268 | 70 | struct rb_root dma_list; |
c086de81 | 71 | struct blocking_notifier_head notifier; |
49285593 | 72 | unsigned int dma_avail; |
f5c9eceb WD |
73 | bool v2; |
74 | bool nesting; | |
1ef3e2bc AW |
75 | }; |
76 | ||
77 | struct vfio_domain { | |
78 | struct iommu_domain *domain; | |
79 | struct list_head next; | |
73fa0d10 | 80 | struct list_head group_list; |
1ef3e2bc | 81 | int prot; /* IOMMU_CACHE */ |
6fe1010d | 82 | bool fgsp; /* Fine-grained super pages */ |
73fa0d10 AW |
83 | }; |
84 | ||
85 | struct vfio_dma { | |
cd9b2268 | 86 | struct rb_node node; |
73fa0d10 AW |
87 | dma_addr_t iova; /* Device address */ |
88 | unsigned long vaddr; /* Process virtual addr */ | |
166fd7d9 | 89 | size_t size; /* Map size (bytes) */ |
73fa0d10 | 90 | int prot; /* IOMMU_READ/WRITE */ |
a54eb550 | 91 | bool iommu_mapped; |
48d8476b | 92 | bool lock_cap; /* capable(CAP_IPC_LOCK) */ |
8f0d5bb9 | 93 | struct task_struct *task; |
a54eb550 | 94 | struct rb_root pfn_list; /* Ex-user pinned pfn list */ |
73fa0d10 AW |
95 | }; |
96 | ||
97 | struct vfio_group { | |
98 | struct iommu_group *iommu_group; | |
99 | struct list_head next; | |
100 | }; | |
101 | ||
a54eb550 KW |
102 | /* |
103 | * Guest RAM pinning working set or DMA target | |
104 | */ | |
105 | struct vfio_pfn { | |
106 | struct rb_node node; | |
107 | dma_addr_t iova; /* Device address */ | |
108 | unsigned long pfn; /* Host pfn */ | |
109 | atomic_t ref_count; | |
110 | }; | |
111 | ||
6bd06f5a SS |
112 | struct vfio_regions { |
113 | struct list_head list; | |
114 | dma_addr_t iova; | |
115 | phys_addr_t phys; | |
116 | size_t len; | |
117 | }; | |
118 | ||
a54eb550 KW |
119 | #define IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu) \ |
120 | (!list_empty(&iommu->domain_list)) | |
121 | ||
122 | static int put_pfn(unsigned long pfn, int prot); | |
123 | ||
73fa0d10 AW |
124 | /* |
125 | * This code handles mapping and unmapping of user data buffers | |
126 | * into DMA'ble space using the IOMMU | |
127 | */ | |
128 | ||
cd9b2268 AW |
129 | static struct vfio_dma *vfio_find_dma(struct vfio_iommu *iommu, |
130 | dma_addr_t start, size_t size) | |
131 | { | |
132 | struct rb_node *node = iommu->dma_list.rb_node; | |
133 | ||
134 | while (node) { | |
135 | struct vfio_dma *dma = rb_entry(node, struct vfio_dma, node); | |
136 | ||
137 | if (start + size <= dma->iova) | |
138 | node = node->rb_left; | |
166fd7d9 | 139 | else if (start >= dma->iova + dma->size) |
cd9b2268 AW |
140 | node = node->rb_right; |
141 | else | |
142 | return dma; | |
143 | } | |
144 | ||
145 | return NULL; | |
146 | } | |
147 | ||
1ef3e2bc | 148 | static void vfio_link_dma(struct vfio_iommu *iommu, struct vfio_dma *new) |
cd9b2268 AW |
149 | { |
150 | struct rb_node **link = &iommu->dma_list.rb_node, *parent = NULL; | |
151 | struct vfio_dma *dma; | |
152 | ||
153 | while (*link) { | |
154 | parent = *link; | |
155 | dma = rb_entry(parent, struct vfio_dma, node); | |
156 | ||
166fd7d9 | 157 | if (new->iova + new->size <= dma->iova) |
cd9b2268 AW |
158 | link = &(*link)->rb_left; |
159 | else | |
160 | link = &(*link)->rb_right; | |
161 | } | |
162 | ||
163 | rb_link_node(&new->node, parent, link); | |
164 | rb_insert_color(&new->node, &iommu->dma_list); | |
165 | } | |
166 | ||
1ef3e2bc | 167 | static void vfio_unlink_dma(struct vfio_iommu *iommu, struct vfio_dma *old) |
cd9b2268 AW |
168 | { |
169 | rb_erase(&old->node, &iommu->dma_list); | |
170 | } | |
171 | ||
a54eb550 KW |
172 | /* |
173 | * Helper Functions for host iova-pfn list | |
174 | */ | |
175 | static struct vfio_pfn *vfio_find_vpfn(struct vfio_dma *dma, dma_addr_t iova) | |
176 | { | |
177 | struct vfio_pfn *vpfn; | |
178 | struct rb_node *node = dma->pfn_list.rb_node; | |
179 | ||
180 | while (node) { | |
181 | vpfn = rb_entry(node, struct vfio_pfn, node); | |
182 | ||
183 | if (iova < vpfn->iova) | |
184 | node = node->rb_left; | |
185 | else if (iova > vpfn->iova) | |
186 | node = node->rb_right; | |
187 | else | |
188 | return vpfn; | |
189 | } | |
190 | return NULL; | |
191 | } | |
192 | ||
193 | static void vfio_link_pfn(struct vfio_dma *dma, | |
194 | struct vfio_pfn *new) | |
195 | { | |
196 | struct rb_node **link, *parent = NULL; | |
197 | struct vfio_pfn *vpfn; | |
198 | ||
199 | link = &dma->pfn_list.rb_node; | |
200 | while (*link) { | |
201 | parent = *link; | |
202 | vpfn = rb_entry(parent, struct vfio_pfn, node); | |
203 | ||
204 | if (new->iova < vpfn->iova) | |
205 | link = &(*link)->rb_left; | |
206 | else | |
207 | link = &(*link)->rb_right; | |
208 | } | |
209 | ||
210 | rb_link_node(&new->node, parent, link); | |
211 | rb_insert_color(&new->node, &dma->pfn_list); | |
212 | } | |
213 | ||
214 | static void vfio_unlink_pfn(struct vfio_dma *dma, struct vfio_pfn *old) | |
215 | { | |
216 | rb_erase(&old->node, &dma->pfn_list); | |
217 | } | |
218 | ||
219 | static int vfio_add_to_pfn_list(struct vfio_dma *dma, dma_addr_t iova, | |
220 | unsigned long pfn) | |
221 | { | |
222 | struct vfio_pfn *vpfn; | |
223 | ||
224 | vpfn = kzalloc(sizeof(*vpfn), GFP_KERNEL); | |
225 | if (!vpfn) | |
226 | return -ENOMEM; | |
227 | ||
228 | vpfn->iova = iova; | |
229 | vpfn->pfn = pfn; | |
230 | atomic_set(&vpfn->ref_count, 1); | |
231 | vfio_link_pfn(dma, vpfn); | |
232 | return 0; | |
233 | } | |
234 | ||
235 | static void vfio_remove_from_pfn_list(struct vfio_dma *dma, | |
236 | struct vfio_pfn *vpfn) | |
237 | { | |
238 | vfio_unlink_pfn(dma, vpfn); | |
239 | kfree(vpfn); | |
240 | } | |
241 | ||
242 | static struct vfio_pfn *vfio_iova_get_vfio_pfn(struct vfio_dma *dma, | |
243 | unsigned long iova) | |
244 | { | |
245 | struct vfio_pfn *vpfn = vfio_find_vpfn(dma, iova); | |
246 | ||
247 | if (vpfn) | |
248 | atomic_inc(&vpfn->ref_count); | |
249 | return vpfn; | |
250 | } | |
251 | ||
252 | static int vfio_iova_put_vfio_pfn(struct vfio_dma *dma, struct vfio_pfn *vpfn) | |
253 | { | |
254 | int ret = 0; | |
255 | ||
256 | if (atomic_dec_and_test(&vpfn->ref_count)) { | |
257 | ret = put_pfn(vpfn->pfn, dma->prot); | |
258 | vfio_remove_from_pfn_list(dma, vpfn); | |
259 | } | |
260 | return ret; | |
261 | } | |
262 | ||
48d8476b | 263 | static int vfio_lock_acct(struct vfio_dma *dma, long npage, bool async) |
73fa0d10 | 264 | { |
73fa0d10 | 265 | struct mm_struct *mm; |
0cfef2b7 | 266 | int ret; |
73fa0d10 | 267 | |
3624a248 | 268 | if (!npage) |
0cfef2b7 | 269 | return 0; |
3624a248 | 270 | |
48d8476b | 271 | mm = async ? get_task_mm(dma->task) : dma->task->mm; |
3624a248 | 272 | if (!mm) |
0cfef2b7 | 273 | return -ESRCH; /* process exited */ |
73fa0d10 | 274 | |
0cfef2b7 AW |
275 | ret = down_write_killable(&mm->mmap_sem); |
276 | if (!ret) { | |
277 | if (npage > 0) { | |
48d8476b | 278 | if (!dma->lock_cap) { |
0cfef2b7 AW |
279 | unsigned long limit; |
280 | ||
48d8476b | 281 | limit = task_rlimit(dma->task, |
0cfef2b7 AW |
282 | RLIMIT_MEMLOCK) >> PAGE_SHIFT; |
283 | ||
284 | if (mm->locked_vm + npage > limit) | |
285 | ret = -ENOMEM; | |
286 | } | |
287 | } | |
288 | ||
289 | if (!ret) | |
290 | mm->locked_vm += npage; | |
73fa0d10 | 291 | |
0cfef2b7 | 292 | up_write(&mm->mmap_sem); |
6c38c055 AW |
293 | } |
294 | ||
48d8476b | 295 | if (async) |
3624a248 | 296 | mmput(mm); |
0cfef2b7 AW |
297 | |
298 | return ret; | |
73fa0d10 AW |
299 | } |
300 | ||
301 | /* | |
302 | * Some mappings aren't backed by a struct page, for example an mmap'd | |
303 | * MMIO range for our own or another device. These use a different | |
304 | * pfn conversion and shouldn't be tracked as locked pages. | |
305 | */ | |
306 | static bool is_invalid_reserved_pfn(unsigned long pfn) | |
307 | { | |
308 | if (pfn_valid(pfn)) { | |
309 | bool reserved; | |
310 | struct page *tail = pfn_to_page(pfn); | |
668f9abb | 311 | struct page *head = compound_head(tail); |
73fa0d10 AW |
312 | reserved = !!(PageReserved(head)); |
313 | if (head != tail) { | |
314 | /* | |
315 | * "head" is not a dangling pointer | |
668f9abb | 316 | * (compound_head takes care of that) |
73fa0d10 AW |
317 | * but the hugepage may have been split |
318 | * from under us (and we may not hold a | |
319 | * reference count on the head page so it can | |
320 | * be reused before we run PageReferenced), so | |
321 | * we've to check PageTail before returning | |
322 | * what we just read. | |
323 | */ | |
324 | smp_rmb(); | |
325 | if (PageTail(tail)) | |
326 | return reserved; | |
327 | } | |
328 | return PageReserved(tail); | |
329 | } | |
330 | ||
331 | return true; | |
332 | } | |
333 | ||
334 | static int put_pfn(unsigned long pfn, int prot) | |
335 | { | |
336 | if (!is_invalid_reserved_pfn(pfn)) { | |
337 | struct page *page = pfn_to_page(pfn); | |
338 | if (prot & IOMMU_WRITE) | |
339 | SetPageDirty(page); | |
340 | put_page(page); | |
341 | return 1; | |
342 | } | |
343 | return 0; | |
344 | } | |
345 | ||
ea85cf35 KW |
346 | static int vaddr_get_pfn(struct mm_struct *mm, unsigned long vaddr, |
347 | int prot, unsigned long *pfn) | |
73fa0d10 AW |
348 | { |
349 | struct page *page[1]; | |
350 | struct vm_area_struct *vma; | |
94db151d | 351 | struct vm_area_struct *vmas[1]; |
bb94b55a | 352 | unsigned int flags = 0; |
ea85cf35 | 353 | int ret; |
73fa0d10 | 354 | |
bb94b55a JG |
355 | if (prot & IOMMU_WRITE) |
356 | flags |= FOLL_WRITE; | |
357 | ||
358 | down_read(&mm->mmap_sem); | |
ea85cf35 | 359 | if (mm == current->mm) { |
bb94b55a | 360 | ret = get_user_pages_longterm(vaddr, 1, flags, page, vmas); |
ea85cf35 | 361 | } else { |
ea85cf35 | 362 | ret = get_user_pages_remote(NULL, mm, vaddr, 1, flags, page, |
94db151d DW |
363 | vmas, NULL); |
364 | /* | |
365 | * The lifetime of a vaddr_get_pfn() page pin is | |
366 | * userspace-controlled. In the fs-dax case this could | |
367 | * lead to indefinite stalls in filesystem operations. | |
368 | * Disallow attempts to pin fs-dax pages via this | |
369 | * interface. | |
370 | */ | |
371 | if (ret > 0 && vma_is_fsdax(vmas[0])) { | |
372 | ret = -EOPNOTSUPP; | |
373 | put_page(page[0]); | |
374 | } | |
ea85cf35 | 375 | } |
bb94b55a | 376 | up_read(&mm->mmap_sem); |
ea85cf35 KW |
377 | |
378 | if (ret == 1) { | |
73fa0d10 AW |
379 | *pfn = page_to_pfn(page[0]); |
380 | return 0; | |
381 | } | |
382 | ||
ea85cf35 | 383 | down_read(&mm->mmap_sem); |
73fa0d10 | 384 | |
ea85cf35 | 385 | vma = find_vma_intersection(mm, vaddr, vaddr + 1); |
73fa0d10 AW |
386 | |
387 | if (vma && vma->vm_flags & VM_PFNMAP) { | |
388 | *pfn = ((vaddr - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff; | |
389 | if (is_invalid_reserved_pfn(*pfn)) | |
390 | ret = 0; | |
391 | } | |
392 | ||
ea85cf35 | 393 | up_read(&mm->mmap_sem); |
73fa0d10 AW |
394 | return ret; |
395 | } | |
396 | ||
166fd7d9 AW |
397 | /* |
398 | * Attempt to pin pages. We really don't want to track all the pfns and | |
399 | * the iommu can only map chunks of consecutive pfns anyway, so get the | |
400 | * first page and all consecutive pages with the same locking. | |
401 | */ | |
8f0d5bb9 | 402 | static long vfio_pin_pages_remote(struct vfio_dma *dma, unsigned long vaddr, |
7cb671e7 | 403 | long npage, unsigned long *pfn_base, |
48d8476b | 404 | unsigned long limit) |
73fa0d10 | 405 | { |
7cb671e7 | 406 | unsigned long pfn = 0; |
6c38c055 | 407 | long ret, pinned = 0, lock_acct = 0; |
89c29def | 408 | bool rsvd; |
a54eb550 | 409 | dma_addr_t iova = vaddr - dma->vaddr + dma->iova; |
73fa0d10 | 410 | |
6c38c055 AW |
411 | /* This code path is only user initiated */ |
412 | if (!current->mm) | |
166fd7d9 | 413 | return -ENODEV; |
73fa0d10 | 414 | |
6c38c055 | 415 | ret = vaddr_get_pfn(current->mm, vaddr, dma->prot, pfn_base); |
166fd7d9 | 416 | if (ret) |
6c38c055 | 417 | return ret; |
73fa0d10 | 418 | |
6c38c055 | 419 | pinned++; |
89c29def | 420 | rsvd = is_invalid_reserved_pfn(*pfn_base); |
73fa0d10 | 421 | |
a54eb550 KW |
422 | /* |
423 | * Reserved pages aren't counted against the user, externally pinned | |
424 | * pages are already counted against the user. | |
425 | */ | |
89c29def | 426 | if (!rsvd && !vfio_find_vpfn(dma, iova)) { |
48d8476b | 427 | if (!dma->lock_cap && current->mm->locked_vm + 1 > limit) { |
a54eb550 KW |
428 | put_pfn(*pfn_base, dma->prot); |
429 | pr_warn("%s: RLIMIT_MEMLOCK (%ld) exceeded\n", __func__, | |
430 | limit << PAGE_SHIFT); | |
6c38c055 | 431 | return -ENOMEM; |
a54eb550 KW |
432 | } |
433 | lock_acct++; | |
5c6c2b21 AW |
434 | } |
435 | ||
6c38c055 AW |
436 | if (unlikely(disable_hugepages)) |
437 | goto out; | |
73fa0d10 | 438 | |
6c38c055 AW |
439 | /* Lock all the consecutive pages from pfn_base */ |
440 | for (vaddr += PAGE_SIZE, iova += PAGE_SIZE; pinned < npage; | |
441 | pinned++, vaddr += PAGE_SIZE, iova += PAGE_SIZE) { | |
6c38c055 AW |
442 | ret = vaddr_get_pfn(current->mm, vaddr, dma->prot, &pfn); |
443 | if (ret) | |
444 | break; | |
445 | ||
89c29def AW |
446 | if (pfn != *pfn_base + pinned || |
447 | rsvd != is_invalid_reserved_pfn(pfn)) { | |
6c38c055 AW |
448 | put_pfn(pfn, dma->prot); |
449 | break; | |
450 | } | |
166fd7d9 | 451 | |
89c29def | 452 | if (!rsvd && !vfio_find_vpfn(dma, iova)) { |
48d8476b | 453 | if (!dma->lock_cap && |
6c38c055 | 454 | current->mm->locked_vm + lock_acct + 1 > limit) { |
a54eb550 | 455 | put_pfn(pfn, dma->prot); |
6c38c055 AW |
456 | pr_warn("%s: RLIMIT_MEMLOCK (%ld) exceeded\n", |
457 | __func__, limit << PAGE_SHIFT); | |
0cfef2b7 AW |
458 | ret = -ENOMEM; |
459 | goto unpin_out; | |
a54eb550 | 460 | } |
6c38c055 | 461 | lock_acct++; |
166fd7d9 AW |
462 | } |
463 | } | |
464 | ||
6c38c055 | 465 | out: |
48d8476b | 466 | ret = vfio_lock_acct(dma, lock_acct, false); |
0cfef2b7 AW |
467 | |
468 | unpin_out: | |
469 | if (ret) { | |
89c29def AW |
470 | if (!rsvd) { |
471 | for (pfn = *pfn_base ; pinned ; pfn++, pinned--) | |
472 | put_pfn(pfn, dma->prot); | |
473 | } | |
0cfef2b7 AW |
474 | |
475 | return ret; | |
476 | } | |
166fd7d9 | 477 | |
6c38c055 | 478 | return pinned; |
166fd7d9 AW |
479 | } |
480 | ||
a54eb550 KW |
481 | static long vfio_unpin_pages_remote(struct vfio_dma *dma, dma_addr_t iova, |
482 | unsigned long pfn, long npage, | |
483 | bool do_accounting) | |
166fd7d9 | 484 | { |
a54eb550 | 485 | long unlocked = 0, locked = 0; |
166fd7d9 AW |
486 | long i; |
487 | ||
6c38c055 | 488 | for (i = 0; i < npage; i++, iova += PAGE_SIZE) { |
a54eb550 KW |
489 | if (put_pfn(pfn++, dma->prot)) { |
490 | unlocked++; | |
6c38c055 | 491 | if (vfio_find_vpfn(dma, iova)) |
a54eb550 KW |
492 | locked++; |
493 | } | |
494 | } | |
495 | ||
496 | if (do_accounting) | |
48d8476b | 497 | vfio_lock_acct(dma, locked - unlocked, true); |
a54eb550 KW |
498 | |
499 | return unlocked; | |
500 | } | |
501 | ||
502 | static int vfio_pin_page_external(struct vfio_dma *dma, unsigned long vaddr, | |
503 | unsigned long *pfn_base, bool do_accounting) | |
504 | { | |
a54eb550 KW |
505 | struct mm_struct *mm; |
506 | int ret; | |
a54eb550 KW |
507 | |
508 | mm = get_task_mm(dma->task); | |
509 | if (!mm) | |
510 | return -ENODEV; | |
511 | ||
512 | ret = vaddr_get_pfn(mm, vaddr, dma->prot, pfn_base); | |
80dbe1fb | 513 | if (!ret && do_accounting && !is_invalid_reserved_pfn(*pfn_base)) { |
48d8476b | 514 | ret = vfio_lock_acct(dma, 1, true); |
0cfef2b7 AW |
515 | if (ret) { |
516 | put_pfn(*pfn_base, dma->prot); | |
80dbe1fb AW |
517 | if (ret == -ENOMEM) |
518 | pr_warn("%s: Task %s (%d) RLIMIT_MEMLOCK " | |
519 | "(%ld) exceeded\n", __func__, | |
520 | dma->task->comm, task_pid_nr(dma->task), | |
521 | task_rlimit(dma->task, RLIMIT_MEMLOCK)); | |
0cfef2b7 AW |
522 | } |
523 | } | |
524 | ||
a54eb550 KW |
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) | |
48d8476b | 541 | vfio_lock_acct(dma, -unlocked, true); |
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); | |
80dbe1fb | 604 | if (ret) |
a54eb550 | 605 | goto pin_unwind; |
a54eb550 KW |
606 | |
607 | ret = vfio_add_to_pfn_list(dma, iova, phys_pfn[i]); | |
608 | if (ret) { | |
609 | vfio_unpin_page_external(dma, iova, do_accounting); | |
610 | goto pin_unwind; | |
611 | } | |
612 | } | |
613 | ||
614 | ret = i; | |
615 | goto pin_done; | |
616 | ||
617 | pin_unwind: | |
618 | phys_pfn[i] = 0; | |
619 | for (j = 0; j < i; j++) { | |
620 | dma_addr_t iova; | |
621 | ||
622 | iova = user_pfn[j] << PAGE_SHIFT; | |
2b8bb1d7 | 623 | dma = vfio_find_dma(iommu, iova, PAGE_SIZE); |
a54eb550 KW |
624 | vfio_unpin_page_external(dma, iova, do_accounting); |
625 | phys_pfn[j] = 0; | |
626 | } | |
627 | pin_done: | |
628 | mutex_unlock(&iommu->lock); | |
629 | return ret; | |
630 | } | |
631 | ||
632 | static int vfio_iommu_type1_unpin_pages(void *iommu_data, | |
633 | unsigned long *user_pfn, | |
634 | int npage) | |
635 | { | |
636 | struct vfio_iommu *iommu = iommu_data; | |
637 | bool do_accounting; | |
638 | int i; | |
639 | ||
640 | if (!iommu || !user_pfn) | |
641 | return -EINVAL; | |
642 | ||
643 | /* Supported for v2 version only */ | |
644 | if (!iommu->v2) | |
645 | return -EACCES; | |
646 | ||
647 | mutex_lock(&iommu->lock); | |
648 | ||
649 | if (!iommu->external_domain) { | |
650 | mutex_unlock(&iommu->lock); | |
651 | return -EINVAL; | |
652 | } | |
653 | ||
654 | do_accounting = !IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu); | |
655 | for (i = 0; i < npage; i++) { | |
656 | struct vfio_dma *dma; | |
657 | dma_addr_t iova; | |
658 | ||
659 | iova = user_pfn[i] << PAGE_SHIFT; | |
2b8bb1d7 | 660 | dma = vfio_find_dma(iommu, iova, PAGE_SIZE); |
a54eb550 KW |
661 | if (!dma) |
662 | goto unpin_exit; | |
663 | vfio_unpin_page_external(dma, iova, do_accounting); | |
664 | } | |
665 | ||
666 | unpin_exit: | |
667 | mutex_unlock(&iommu->lock); | |
668 | return i > npage ? npage : (i > 0 ? i : -EINVAL); | |
669 | } | |
670 | ||
6bd06f5a SS |
671 | static long vfio_sync_unpin(struct vfio_dma *dma, struct vfio_domain *domain, |
672 | struct list_head *regions) | |
673 | { | |
674 | long unlocked = 0; | |
675 | struct vfio_regions *entry, *next; | |
676 | ||
677 | iommu_tlb_sync(domain->domain); | |
678 | ||
679 | list_for_each_entry_safe(entry, next, regions, list) { | |
680 | unlocked += vfio_unpin_pages_remote(dma, | |
681 | entry->iova, | |
682 | entry->phys >> PAGE_SHIFT, | |
683 | entry->len >> PAGE_SHIFT, | |
684 | false); | |
685 | list_del(&entry->list); | |
686 | kfree(entry); | |
687 | } | |
688 | ||
689 | cond_resched(); | |
690 | ||
691 | return unlocked; | |
692 | } | |
693 | ||
694 | /* | |
695 | * Generally, VFIO needs to unpin remote pages after each IOTLB flush. | |
696 | * Therefore, when using IOTLB flush sync interface, VFIO need to keep track | |
697 | * of these regions (currently using a list). | |
698 | * | |
699 | * This value specifies maximum number of regions for each IOTLB flush sync. | |
700 | */ | |
701 | #define VFIO_IOMMU_TLB_SYNC_MAX 512 | |
702 | ||
703 | static size_t unmap_unpin_fast(struct vfio_domain *domain, | |
704 | struct vfio_dma *dma, dma_addr_t *iova, | |
705 | size_t len, phys_addr_t phys, long *unlocked, | |
706 | struct list_head *unmapped_list, | |
707 | int *unmapped_cnt) | |
708 | { | |
709 | size_t unmapped = 0; | |
710 | struct vfio_regions *entry = kzalloc(sizeof(*entry), GFP_KERNEL); | |
711 | ||
712 | if (entry) { | |
713 | unmapped = iommu_unmap_fast(domain->domain, *iova, len); | |
714 | ||
715 | if (!unmapped) { | |
716 | kfree(entry); | |
717 | } else { | |
718 | iommu_tlb_range_add(domain->domain, *iova, unmapped); | |
719 | entry->iova = *iova; | |
720 | entry->phys = phys; | |
721 | entry->len = unmapped; | |
722 | list_add_tail(&entry->list, unmapped_list); | |
723 | ||
724 | *iova += unmapped; | |
725 | (*unmapped_cnt)++; | |
726 | } | |
727 | } | |
728 | ||
729 | /* | |
730 | * Sync if the number of fast-unmap regions hits the limit | |
731 | * or in case of errors. | |
732 | */ | |
733 | if (*unmapped_cnt >= VFIO_IOMMU_TLB_SYNC_MAX || !unmapped) { | |
734 | *unlocked += vfio_sync_unpin(dma, domain, | |
735 | unmapped_list); | |
736 | *unmapped_cnt = 0; | |
737 | } | |
738 | ||
739 | return unmapped; | |
740 | } | |
741 | ||
742 | static size_t unmap_unpin_slow(struct vfio_domain *domain, | |
743 | struct vfio_dma *dma, dma_addr_t *iova, | |
744 | size_t len, phys_addr_t phys, | |
745 | long *unlocked) | |
746 | { | |
747 | size_t unmapped = iommu_unmap(domain->domain, *iova, len); | |
748 | ||
749 | if (unmapped) { | |
750 | *unlocked += vfio_unpin_pages_remote(dma, *iova, | |
751 | phys >> PAGE_SHIFT, | |
752 | unmapped >> PAGE_SHIFT, | |
753 | false); | |
754 | *iova += unmapped; | |
755 | cond_resched(); | |
756 | } | |
757 | return unmapped; | |
758 | } | |
759 | ||
a54eb550 KW |
760 | static long vfio_unmap_unpin(struct vfio_iommu *iommu, struct vfio_dma *dma, |
761 | bool do_accounting) | |
166fd7d9 | 762 | { |
1ef3e2bc AW |
763 | dma_addr_t iova = dma->iova, end = dma->iova + dma->size; |
764 | struct vfio_domain *domain, *d; | |
6bd06f5a SS |
765 | LIST_HEAD(unmapped_region_list); |
766 | int unmapped_region_cnt = 0; | |
166fd7d9 AW |
767 | long unlocked = 0; |
768 | ||
1ef3e2bc | 769 | if (!dma->size) |
a54eb550 KW |
770 | return 0; |
771 | ||
772 | if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu)) | |
773 | return 0; | |
774 | ||
1ef3e2bc AW |
775 | /* |
776 | * We use the IOMMU to track the physical addresses, otherwise we'd | |
777 | * need a much more complicated tracking system. Unfortunately that | |
778 | * means we need to use one of the iommu domains to figure out the | |
779 | * pfns to unpin. The rest need to be unmapped in advance so we have | |
780 | * no iommu translations remaining when the pages are unpinned. | |
781 | */ | |
782 | domain = d = list_first_entry(&iommu->domain_list, | |
783 | struct vfio_domain, next); | |
784 | ||
c5e66887 | 785 | list_for_each_entry_continue(d, &iommu->domain_list, next) { |
1ef3e2bc | 786 | iommu_unmap(d->domain, dma->iova, dma->size); |
c5e66887 AW |
787 | cond_resched(); |
788 | } | |
1ef3e2bc | 789 | |
166fd7d9 | 790 | while (iova < end) { |
6fe1010d AW |
791 | size_t unmapped, len; |
792 | phys_addr_t phys, next; | |
166fd7d9 | 793 | |
1ef3e2bc | 794 | phys = iommu_iova_to_phys(domain->domain, iova); |
166fd7d9 AW |
795 | if (WARN_ON(!phys)) { |
796 | iova += PAGE_SIZE; | |
797 | continue; | |
73fa0d10 | 798 | } |
166fd7d9 | 799 | |
6fe1010d AW |
800 | /* |
801 | * To optimize for fewer iommu_unmap() calls, each of which | |
802 | * may require hardware cache flushing, try to find the | |
803 | * largest contiguous physical memory chunk to unmap. | |
804 | */ | |
805 | for (len = PAGE_SIZE; | |
806 | !domain->fgsp && iova + len < end; len += PAGE_SIZE) { | |
807 | next = iommu_iova_to_phys(domain->domain, iova + len); | |
808 | if (next != phys + len) | |
809 | break; | |
810 | } | |
811 | ||
6bd06f5a SS |
812 | /* |
813 | * First, try to use fast unmap/unpin. In case of failure, | |
814 | * switch to slow unmap/unpin path. | |
815 | */ | |
816 | unmapped = unmap_unpin_fast(domain, dma, &iova, len, phys, | |
817 | &unlocked, &unmapped_region_list, | |
818 | &unmapped_region_cnt); | |
819 | if (!unmapped) { | |
820 | unmapped = unmap_unpin_slow(domain, dma, &iova, len, | |
821 | phys, &unlocked); | |
822 | if (WARN_ON(!unmapped)) | |
823 | break; | |
824 | } | |
73fa0d10 | 825 | } |
166fd7d9 | 826 | |
a54eb550 | 827 | dma->iommu_mapped = false; |
6bd06f5a SS |
828 | |
829 | if (unmapped_region_cnt) | |
830 | unlocked += vfio_sync_unpin(dma, domain, &unmapped_region_list); | |
831 | ||
a54eb550 | 832 | if (do_accounting) { |
48d8476b | 833 | vfio_lock_acct(dma, -unlocked, true); |
a54eb550 KW |
834 | return 0; |
835 | } | |
836 | return unlocked; | |
73fa0d10 AW |
837 | } |
838 | ||
1ef3e2bc | 839 | static void vfio_remove_dma(struct vfio_iommu *iommu, struct vfio_dma *dma) |
73fa0d10 | 840 | { |
a54eb550 | 841 | vfio_unmap_unpin(iommu, dma, true); |
1ef3e2bc | 842 | vfio_unlink_dma(iommu, dma); |
8f0d5bb9 | 843 | put_task_struct(dma->task); |
1ef3e2bc | 844 | kfree(dma); |
49285593 | 845 | iommu->dma_avail++; |
1ef3e2bc | 846 | } |
73fa0d10 | 847 | |
1ef3e2bc AW |
848 | static unsigned long vfio_pgsize_bitmap(struct vfio_iommu *iommu) |
849 | { | |
850 | struct vfio_domain *domain; | |
4644321f | 851 | unsigned long bitmap = ULONG_MAX; |
166fd7d9 | 852 | |
1ef3e2bc AW |
853 | mutex_lock(&iommu->lock); |
854 | list_for_each_entry(domain, &iommu->domain_list, next) | |
d16e0faa | 855 | bitmap &= domain->domain->pgsize_bitmap; |
1ef3e2bc | 856 | mutex_unlock(&iommu->lock); |
73fa0d10 | 857 | |
4644321f EA |
858 | /* |
859 | * In case the IOMMU supports page sizes smaller than PAGE_SIZE | |
860 | * we pretend PAGE_SIZE is supported and hide sub-PAGE_SIZE sizes. | |
861 | * That way the user will be able to map/unmap buffers whose size/ | |
862 | * start address is aligned with PAGE_SIZE. Pinning code uses that | |
863 | * granularity while iommu driver can use the sub-PAGE_SIZE size | |
864 | * to map the buffer. | |
865 | */ | |
866 | if (bitmap & ~PAGE_MASK) { | |
867 | bitmap &= PAGE_MASK; | |
868 | bitmap |= PAGE_SIZE; | |
869 | } | |
870 | ||
1ef3e2bc | 871 | return bitmap; |
73fa0d10 AW |
872 | } |
873 | ||
874 | static int vfio_dma_do_unmap(struct vfio_iommu *iommu, | |
875 | struct vfio_iommu_type1_dma_unmap *unmap) | |
876 | { | |
73fa0d10 | 877 | uint64_t mask; |
c086de81 | 878 | struct vfio_dma *dma, *dma_last = NULL; |
1ef3e2bc | 879 | size_t unmapped = 0; |
c086de81 | 880 | int ret = 0, retries = 0; |
73fa0d10 | 881 | |
1ef3e2bc | 882 | mask = ((uint64_t)1 << __ffs(vfio_pgsize_bitmap(iommu))) - 1; |
73fa0d10 AW |
883 | |
884 | if (unmap->iova & mask) | |
885 | return -EINVAL; | |
f5bfdbf2 | 886 | if (!unmap->size || unmap->size & mask) |
73fa0d10 | 887 | return -EINVAL; |
58fec830 | 888 | if (unmap->iova + unmap->size - 1 < unmap->iova || |
71a7d3d7 DC |
889 | unmap->size > SIZE_MAX) |
890 | return -EINVAL; | |
73fa0d10 | 891 | |
73fa0d10 | 892 | WARN_ON(mask & PAGE_MASK); |
c086de81 | 893 | again: |
73fa0d10 AW |
894 | mutex_lock(&iommu->lock); |
895 | ||
1ef3e2bc AW |
896 | /* |
897 | * vfio-iommu-type1 (v1) - User mappings were coalesced together to | |
898 | * avoid tracking individual mappings. This means that the granularity | |
899 | * of the original mapping was lost and the user was allowed to attempt | |
900 | * to unmap any range. Depending on the contiguousness of physical | |
901 | * memory and page sizes supported by the IOMMU, arbitrary unmaps may | |
902 | * or may not have worked. We only guaranteed unmap granularity | |
903 | * matching the original mapping; even though it was untracked here, | |
904 | * the original mappings are reflected in IOMMU mappings. This | |
905 | * resulted in a couple unusual behaviors. First, if a range is not | |
906 | * able to be unmapped, ex. a set of 4k pages that was mapped as a | |
907 | * 2M hugepage into the IOMMU, the unmap ioctl returns success but with | |
908 | * a zero sized unmap. Also, if an unmap request overlaps the first | |
909 | * address of a hugepage, the IOMMU will unmap the entire hugepage. | |
910 | * This also returns success and the returned unmap size reflects the | |
911 | * actual size unmapped. | |
912 | * | |
913 | * We attempt to maintain compatibility with this "v1" interface, but | |
914 | * we take control out of the hands of the IOMMU. Therefore, an unmap | |
915 | * request offset from the beginning of the original mapping will | |
916 | * return success with zero sized unmap. And an unmap request covering | |
917 | * the first iova of mapping will unmap the entire range. | |
918 | * | |
919 | * The v2 version of this interface intends to be more deterministic. | |
920 | * Unmap requests must fully cover previous mappings. Multiple | |
921 | * mappings may still be unmaped by specifying large ranges, but there | |
922 | * must not be any previous mappings bisected by the range. An error | |
923 | * will be returned if these conditions are not met. The v2 interface | |
924 | * will only return success and a size of zero if there were no | |
925 | * mappings within the range. | |
926 | */ | |
927 | if (iommu->v2) { | |
7c03f428 | 928 | dma = vfio_find_dma(iommu, unmap->iova, 1); |
1ef3e2bc AW |
929 | if (dma && dma->iova != unmap->iova) { |
930 | ret = -EINVAL; | |
931 | goto unlock; | |
932 | } | |
933 | dma = vfio_find_dma(iommu, unmap->iova + unmap->size - 1, 0); | |
934 | if (dma && dma->iova + dma->size != unmap->iova + unmap->size) { | |
935 | ret = -EINVAL; | |
936 | goto unlock; | |
937 | } | |
938 | } | |
939 | ||
166fd7d9 | 940 | while ((dma = vfio_find_dma(iommu, unmap->iova, unmap->size))) { |
1ef3e2bc | 941 | if (!iommu->v2 && unmap->iova > dma->iova) |
166fd7d9 | 942 | break; |
8f0d5bb9 KW |
943 | /* |
944 | * Task with same address space who mapped this iova range is | |
945 | * allowed to unmap the iova range. | |
946 | */ | |
947 | if (dma->task->mm != current->mm) | |
948 | break; | |
c086de81 KW |
949 | |
950 | if (!RB_EMPTY_ROOT(&dma->pfn_list)) { | |
951 | struct vfio_iommu_type1_dma_unmap nb_unmap; | |
952 | ||
953 | if (dma_last == dma) { | |
954 | BUG_ON(++retries > 10); | |
955 | } else { | |
956 | dma_last = dma; | |
957 | retries = 0; | |
958 | } | |
959 | ||
960 | nb_unmap.iova = dma->iova; | |
961 | nb_unmap.size = dma->size; | |
962 | ||
963 | /* | |
964 | * Notify anyone (mdev vendor drivers) to invalidate and | |
965 | * unmap iovas within the range we're about to unmap. | |
966 | * Vendor drivers MUST unpin pages in response to an | |
967 | * invalidation. | |
968 | */ | |
969 | mutex_unlock(&iommu->lock); | |
970 | blocking_notifier_call_chain(&iommu->notifier, | |
971 | VFIO_IOMMU_NOTIFY_DMA_UNMAP, | |
972 | &nb_unmap); | |
973 | goto again; | |
974 | } | |
1ef3e2bc AW |
975 | unmapped += dma->size; |
976 | vfio_remove_dma(iommu, dma); | |
166fd7d9 | 977 | } |
cd9b2268 | 978 | |
1ef3e2bc | 979 | unlock: |
73fa0d10 | 980 | mutex_unlock(&iommu->lock); |
166fd7d9 | 981 | |
1ef3e2bc | 982 | /* Report how much was unmapped */ |
166fd7d9 AW |
983 | unmap->size = unmapped; |
984 | ||
985 | return ret; | |
986 | } | |
987 | ||
1ef3e2bc AW |
988 | static int vfio_iommu_map(struct vfio_iommu *iommu, dma_addr_t iova, |
989 | unsigned long pfn, long npage, int prot) | |
990 | { | |
991 | struct vfio_domain *d; | |
992 | int ret; | |
993 | ||
994 | list_for_each_entry(d, &iommu->domain_list, next) { | |
995 | ret = iommu_map(d->domain, iova, (phys_addr_t)pfn << PAGE_SHIFT, | |
996 | npage << PAGE_SHIFT, prot | d->prot); | |
7a30423a JR |
997 | if (ret) |
998 | goto unwind; | |
c5e66887 AW |
999 | |
1000 | cond_resched(); | |
1ef3e2bc AW |
1001 | } |
1002 | ||
1003 | return 0; | |
1004 | ||
1005 | unwind: | |
1006 | list_for_each_entry_continue_reverse(d, &iommu->domain_list, next) | |
1007 | iommu_unmap(d->domain, iova, npage << PAGE_SHIFT); | |
166fd7d9 | 1008 | |
cd9b2268 | 1009 | return ret; |
73fa0d10 AW |
1010 | } |
1011 | ||
8f0d5bb9 KW |
1012 | static int vfio_pin_map_dma(struct vfio_iommu *iommu, struct vfio_dma *dma, |
1013 | size_t map_size) | |
1014 | { | |
1015 | dma_addr_t iova = dma->iova; | |
1016 | unsigned long vaddr = dma->vaddr; | |
1017 | size_t size = map_size; | |
1018 | long npage; | |
7cb671e7 | 1019 | unsigned long pfn, limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT; |
8f0d5bb9 KW |
1020 | int ret = 0; |
1021 | ||
1022 | while (size) { | |
1023 | /* Pin a contiguous chunk of memory */ | |
1024 | npage = vfio_pin_pages_remote(dma, vaddr + dma->size, | |
48d8476b | 1025 | size >> PAGE_SHIFT, &pfn, limit); |
8f0d5bb9 KW |
1026 | if (npage <= 0) { |
1027 | WARN_ON(!npage); | |
1028 | ret = (int)npage; | |
1029 | break; | |
1030 | } | |
1031 | ||
1032 | /* Map it! */ | |
1033 | ret = vfio_iommu_map(iommu, iova + dma->size, pfn, npage, | |
1034 | dma->prot); | |
1035 | if (ret) { | |
a54eb550 KW |
1036 | vfio_unpin_pages_remote(dma, iova + dma->size, pfn, |
1037 | npage, true); | |
8f0d5bb9 KW |
1038 | break; |
1039 | } | |
1040 | ||
1041 | size -= npage << PAGE_SHIFT; | |
1042 | dma->size += npage << PAGE_SHIFT; | |
1043 | } | |
1044 | ||
a54eb550 KW |
1045 | dma->iommu_mapped = true; |
1046 | ||
8f0d5bb9 KW |
1047 | if (ret) |
1048 | vfio_remove_dma(iommu, dma); | |
1049 | ||
1050 | return ret; | |
1051 | } | |
1052 | ||
73fa0d10 AW |
1053 | static int vfio_dma_do_map(struct vfio_iommu *iommu, |
1054 | struct vfio_iommu_type1_dma_map *map) | |
1055 | { | |
c8dbca16 | 1056 | dma_addr_t iova = map->iova; |
166fd7d9 | 1057 | unsigned long vaddr = map->vaddr; |
73fa0d10 AW |
1058 | size_t size = map->size; |
1059 | int ret = 0, prot = 0; | |
1060 | uint64_t mask; | |
1ef3e2bc | 1061 | struct vfio_dma *dma; |
166fd7d9 | 1062 | |
c8dbca16 AW |
1063 | /* Verify that none of our __u64 fields overflow */ |
1064 | if (map->size != size || map->vaddr != vaddr || map->iova != iova) | |
1065 | return -EINVAL; | |
73fa0d10 | 1066 | |
1ef3e2bc | 1067 | mask = ((uint64_t)1 << __ffs(vfio_pgsize_bitmap(iommu))) - 1; |
73fa0d10 | 1068 | |
c8dbca16 AW |
1069 | WARN_ON(mask & PAGE_MASK); |
1070 | ||
73fa0d10 AW |
1071 | /* READ/WRITE from device perspective */ |
1072 | if (map->flags & VFIO_DMA_MAP_FLAG_WRITE) | |
1073 | prot |= IOMMU_WRITE; | |
1074 | if (map->flags & VFIO_DMA_MAP_FLAG_READ) | |
1075 | prot |= IOMMU_READ; | |
1076 | ||
c8dbca16 | 1077 | if (!prot || !size || (size | iova | vaddr) & mask) |
73fa0d10 AW |
1078 | return -EINVAL; |
1079 | ||
c8dbca16 AW |
1080 | /* Don't allow IOVA or virtual address wrap */ |
1081 | if (iova + size - 1 < iova || vaddr + size - 1 < vaddr) | |
73fa0d10 AW |
1082 | return -EINVAL; |
1083 | ||
1084 | mutex_lock(&iommu->lock); | |
1085 | ||
c8dbca16 | 1086 | if (vfio_find_dma(iommu, iova, size)) { |
8f0d5bb9 KW |
1087 | ret = -EEXIST; |
1088 | goto out_unlock; | |
73fa0d10 AW |
1089 | } |
1090 | ||
49285593 AW |
1091 | if (!iommu->dma_avail) { |
1092 | ret = -ENOSPC; | |
1093 | goto out_unlock; | |
1094 | } | |
1095 | ||
1ef3e2bc AW |
1096 | dma = kzalloc(sizeof(*dma), GFP_KERNEL); |
1097 | if (!dma) { | |
8f0d5bb9 KW |
1098 | ret = -ENOMEM; |
1099 | goto out_unlock; | |
1ef3e2bc AW |
1100 | } |
1101 | ||
49285593 | 1102 | iommu->dma_avail--; |
c8dbca16 AW |
1103 | dma->iova = iova; |
1104 | dma->vaddr = vaddr; | |
1ef3e2bc | 1105 | dma->prot = prot; |
48d8476b AW |
1106 | |
1107 | /* | |
1108 | * We need to be able to both add to a task's locked memory and test | |
1109 | * against the locked memory limit and we need to be able to do both | |
1110 | * outside of this call path as pinning can be asynchronous via the | |
1111 | * external interfaces for mdev devices. RLIMIT_MEMLOCK requires a | |
1112 | * task_struct and VM locked pages requires an mm_struct, however | |
1113 | * holding an indefinite mm reference is not recommended, therefore we | |
1114 | * only hold a reference to a task. We could hold a reference to | |
1115 | * current, however QEMU uses this call path through vCPU threads, | |
1116 | * which can be killed resulting in a NULL mm and failure in the unmap | |
1117 | * path when called via a different thread. Avoid this problem by | |
1118 | * using the group_leader as threads within the same group require | |
1119 | * both CLONE_THREAD and CLONE_VM and will therefore use the same | |
1120 | * mm_struct. | |
1121 | * | |
1122 | * Previously we also used the task for testing CAP_IPC_LOCK at the | |
1123 | * time of pinning and accounting, however has_capability() makes use | |
1124 | * of real_cred, a copy-on-write field, so we can't guarantee that it | |
1125 | * matches group_leader, or in fact that it might not change by the | |
1126 | * time it's evaluated. If a process were to call MAP_DMA with | |
1127 | * CAP_IPC_LOCK but later drop it, it doesn't make sense that they | |
1128 | * possibly see different results for an iommu_mapped vfio_dma vs | |
1129 | * externally mapped. Therefore track CAP_IPC_LOCK in vfio_dma at the | |
1130 | * time of calling MAP_DMA. | |
1131 | */ | |
1132 | get_task_struct(current->group_leader); | |
1133 | dma->task = current->group_leader; | |
1134 | dma->lock_cap = capable(CAP_IPC_LOCK); | |
1135 | ||
a54eb550 | 1136 | dma->pfn_list = RB_ROOT; |
166fd7d9 | 1137 | |
1ef3e2bc AW |
1138 | /* Insert zero-sized and grow as we map chunks of it */ |
1139 | vfio_link_dma(iommu, dma); | |
166fd7d9 | 1140 | |
a54eb550 KW |
1141 | /* Don't pin and map if container doesn't contain IOMMU capable domain*/ |
1142 | if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu)) | |
1143 | dma->size = size; | |
1144 | else | |
1145 | ret = vfio_pin_map_dma(iommu, dma, size); | |
1146 | ||
8f0d5bb9 | 1147 | out_unlock: |
1ef3e2bc AW |
1148 | mutex_unlock(&iommu->lock); |
1149 | return ret; | |
1150 | } | |
1151 | ||
1152 | static int vfio_bus_type(struct device *dev, void *data) | |
1153 | { | |
1154 | struct bus_type **bus = data; | |
1155 | ||
1156 | if (*bus && *bus != dev->bus) | |
1157 | return -EINVAL; | |
1158 | ||
1159 | *bus = dev->bus; | |
1160 | ||
1161 | return 0; | |
1162 | } | |
1163 | ||
1164 | static int vfio_iommu_replay(struct vfio_iommu *iommu, | |
1165 | struct vfio_domain *domain) | |
1166 | { | |
1167 | struct vfio_domain *d; | |
1168 | struct rb_node *n; | |
7cb671e7 | 1169 | unsigned long limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT; |
1ef3e2bc AW |
1170 | int ret; |
1171 | ||
1172 | /* Arbitrarily pick the first domain in the list for lookups */ | |
1173 | d = list_first_entry(&iommu->domain_list, struct vfio_domain, next); | |
1174 | n = rb_first(&iommu->dma_list); | |
1175 | ||
1ef3e2bc AW |
1176 | for (; n; n = rb_next(n)) { |
1177 | struct vfio_dma *dma; | |
1178 | dma_addr_t iova; | |
1179 | ||
1180 | dma = rb_entry(n, struct vfio_dma, node); | |
1181 | iova = dma->iova; | |
1182 | ||
1183 | while (iova < dma->iova + dma->size) { | |
a54eb550 | 1184 | phys_addr_t phys; |
1ef3e2bc | 1185 | size_t size; |
73fa0d10 | 1186 | |
a54eb550 KW |
1187 | if (dma->iommu_mapped) { |
1188 | phys_addr_t p; | |
1189 | dma_addr_t i; | |
1190 | ||
1191 | phys = iommu_iova_to_phys(d->domain, iova); | |
1192 | ||
1193 | if (WARN_ON(!phys)) { | |
1194 | iova += PAGE_SIZE; | |
1195 | continue; | |
1196 | } | |
1197 | ||
1198 | size = PAGE_SIZE; | |
1199 | p = phys + size; | |
1200 | i = iova + size; | |
1201 | while (i < dma->iova + dma->size && | |
1202 | p == iommu_iova_to_phys(d->domain, i)) { | |
1203 | size += PAGE_SIZE; | |
1204 | p += PAGE_SIZE; | |
1205 | i += PAGE_SIZE; | |
1206 | } | |
1207 | } else { | |
1208 | unsigned long pfn; | |
1209 | unsigned long vaddr = dma->vaddr + | |
1210 | (iova - dma->iova); | |
1211 | size_t n = dma->iova + dma->size - iova; | |
1212 | long npage; | |
1213 | ||
1214 | npage = vfio_pin_pages_remote(dma, vaddr, | |
1215 | n >> PAGE_SHIFT, | |
48d8476b | 1216 | &pfn, limit); |
a54eb550 KW |
1217 | if (npage <= 0) { |
1218 | WARN_ON(!npage); | |
1219 | ret = (int)npage; | |
1220 | return ret; | |
1221 | } | |
1222 | ||
1223 | phys = pfn << PAGE_SHIFT; | |
1224 | size = npage << PAGE_SHIFT; | |
166fd7d9 AW |
1225 | } |
1226 | ||
1ef3e2bc AW |
1227 | ret = iommu_map(domain->domain, iova, phys, |
1228 | size, dma->prot | domain->prot); | |
1229 | if (ret) | |
1230 | return ret; | |
d93b3ac0 | 1231 | |
1ef3e2bc AW |
1232 | iova += size; |
1233 | } | |
a54eb550 | 1234 | dma->iommu_mapped = true; |
166fd7d9 | 1235 | } |
1ef3e2bc | 1236 | return 0; |
73fa0d10 AW |
1237 | } |
1238 | ||
6fe1010d AW |
1239 | /* |
1240 | * We change our unmap behavior slightly depending on whether the IOMMU | |
1241 | * supports fine-grained superpages. IOMMUs like AMD-Vi will use a superpage | |
1242 | * for practically any contiguous power-of-two mapping we give it. This means | |
1243 | * we don't need to look for contiguous chunks ourselves to make unmapping | |
1244 | * more efficient. On IOMMUs with coarse-grained super pages, like Intel VT-d | |
1245 | * with discrete 2M/1G/512G/1T superpages, identifying contiguous chunks | |
1246 | * significantly boosts non-hugetlbfs mappings and doesn't seem to hurt when | |
1247 | * hugetlbfs is in use. | |
1248 | */ | |
1249 | static void vfio_test_domain_fgsp(struct vfio_domain *domain) | |
1250 | { | |
1251 | struct page *pages; | |
1252 | int ret, order = get_order(PAGE_SIZE * 2); | |
1253 | ||
1254 | pages = alloc_pages(GFP_KERNEL | __GFP_ZERO, order); | |
1255 | if (!pages) | |
1256 | return; | |
1257 | ||
1258 | ret = iommu_map(domain->domain, 0, page_to_phys(pages), PAGE_SIZE * 2, | |
1259 | IOMMU_READ | IOMMU_WRITE | domain->prot); | |
1260 | if (!ret) { | |
1261 | size_t unmapped = iommu_unmap(domain->domain, 0, PAGE_SIZE); | |
1262 | ||
1263 | if (unmapped == PAGE_SIZE) | |
1264 | iommu_unmap(domain->domain, PAGE_SIZE, PAGE_SIZE); | |
1265 | else | |
1266 | domain->fgsp = true; | |
1267 | } | |
1268 | ||
1269 | __free_pages(pages, order); | |
1270 | } | |
1271 | ||
7896c998 KW |
1272 | static struct vfio_group *find_iommu_group(struct vfio_domain *domain, |
1273 | struct iommu_group *iommu_group) | |
1274 | { | |
1275 | struct vfio_group *g; | |
1276 | ||
1277 | list_for_each_entry(g, &domain->group_list, next) { | |
1278 | if (g->iommu_group == iommu_group) | |
1279 | return g; | |
1280 | } | |
1281 | ||
1282 | return NULL; | |
1283 | } | |
1284 | ||
9d3a4de4 | 1285 | static bool vfio_iommu_has_sw_msi(struct iommu_group *group, phys_addr_t *base) |
5d704992 EA |
1286 | { |
1287 | struct list_head group_resv_regions; | |
1288 | struct iommu_resv_region *region, *next; | |
1289 | bool ret = false; | |
1290 | ||
1291 | INIT_LIST_HEAD(&group_resv_regions); | |
1292 | iommu_get_group_resv_regions(group, &group_resv_regions); | |
1293 | list_for_each_entry(region, &group_resv_regions, list) { | |
f203f7f1 RM |
1294 | /* |
1295 | * The presence of any 'real' MSI regions should take | |
1296 | * precedence over the software-managed one if the | |
1297 | * IOMMU driver happens to advertise both types. | |
1298 | */ | |
1299 | if (region->type == IOMMU_RESV_MSI) { | |
1300 | ret = false; | |
1301 | break; | |
1302 | } | |
1303 | ||
9d3a4de4 | 1304 | if (region->type == IOMMU_RESV_SW_MSI) { |
5d704992 EA |
1305 | *base = region->start; |
1306 | ret = true; | |
5d704992 EA |
1307 | } |
1308 | } | |
5d704992 EA |
1309 | list_for_each_entry_safe(region, next, &group_resv_regions, list) |
1310 | kfree(region); | |
1311 | return ret; | |
1312 | } | |
1313 | ||
73fa0d10 AW |
1314 | static int vfio_iommu_type1_attach_group(void *iommu_data, |
1315 | struct iommu_group *iommu_group) | |
1316 | { | |
1317 | struct vfio_iommu *iommu = iommu_data; | |
7896c998 | 1318 | struct vfio_group *group; |
1ef3e2bc | 1319 | struct vfio_domain *domain, *d; |
a54eb550 | 1320 | struct bus_type *bus = NULL, *mdev_bus; |
73fa0d10 | 1321 | int ret; |
9d72f87b | 1322 | bool resv_msi, msi_remap; |
5d704992 | 1323 | phys_addr_t resv_msi_base; |
73fa0d10 | 1324 | |
73fa0d10 AW |
1325 | mutex_lock(&iommu->lock); |
1326 | ||
1ef3e2bc | 1327 | list_for_each_entry(d, &iommu->domain_list, next) { |
7896c998 | 1328 | if (find_iommu_group(d, iommu_group)) { |
73fa0d10 | 1329 | mutex_unlock(&iommu->lock); |
73fa0d10 AW |
1330 | return -EINVAL; |
1331 | } | |
1332 | } | |
1333 | ||
a54eb550 KW |
1334 | if (iommu->external_domain) { |
1335 | if (find_iommu_group(iommu->external_domain, iommu_group)) { | |
1336 | mutex_unlock(&iommu->lock); | |
1337 | return -EINVAL; | |
1338 | } | |
1339 | } | |
1340 | ||
1ef3e2bc AW |
1341 | group = kzalloc(sizeof(*group), GFP_KERNEL); |
1342 | domain = kzalloc(sizeof(*domain), GFP_KERNEL); | |
1343 | if (!group || !domain) { | |
1344 | ret = -ENOMEM; | |
1345 | goto out_free; | |
1346 | } | |
1347 | ||
1348 | group->iommu_group = iommu_group; | |
1349 | ||
1350 | /* Determine bus_type in order to allocate a domain */ | |
1351 | ret = iommu_group_for_each_dev(iommu_group, &bus, vfio_bus_type); | |
1352 | if (ret) | |
1353 | goto out_free; | |
1354 | ||
a54eb550 KW |
1355 | mdev_bus = symbol_get(mdev_bus_type); |
1356 | ||
1357 | if (mdev_bus) { | |
1358 | if ((bus == mdev_bus) && !iommu_present(bus)) { | |
1359 | symbol_put(mdev_bus_type); | |
1360 | if (!iommu->external_domain) { | |
1361 | INIT_LIST_HEAD(&domain->group_list); | |
1362 | iommu->external_domain = domain; | |
1363 | } else | |
1364 | kfree(domain); | |
1365 | ||
1366 | list_add(&group->next, | |
1367 | &iommu->external_domain->group_list); | |
1368 | mutex_unlock(&iommu->lock); | |
1369 | return 0; | |
1370 | } | |
1371 | symbol_put(mdev_bus_type); | |
1372 | } | |
1373 | ||
1ef3e2bc AW |
1374 | domain->domain = iommu_domain_alloc(bus); |
1375 | if (!domain->domain) { | |
1376 | ret = -EIO; | |
1377 | goto out_free; | |
1378 | } | |
1379 | ||
f5c9eceb WD |
1380 | if (iommu->nesting) { |
1381 | int attr = 1; | |
1382 | ||
1383 | ret = iommu_domain_set_attr(domain->domain, DOMAIN_ATTR_NESTING, | |
1384 | &attr); | |
1385 | if (ret) | |
1386 | goto out_domain; | |
1387 | } | |
1388 | ||
1ef3e2bc AW |
1389 | ret = iommu_attach_group(domain->domain, iommu_group); |
1390 | if (ret) | |
1391 | goto out_domain; | |
1392 | ||
9d3a4de4 | 1393 | resv_msi = vfio_iommu_has_sw_msi(iommu_group, &resv_msi_base); |
5d704992 | 1394 | |
1ef3e2bc AW |
1395 | INIT_LIST_HEAD(&domain->group_list); |
1396 | list_add(&group->next, &domain->group_list); | |
1397 | ||
db406cc0 RM |
1398 | msi_remap = irq_domain_check_msi_remap() || |
1399 | iommu_capable(bus, IOMMU_CAP_INTR_REMAP); | |
9d72f87b EA |
1400 | |
1401 | if (!allow_unsafe_interrupts && !msi_remap) { | |
1ef3e2bc AW |
1402 | pr_warn("%s: No interrupt remapping support. Use the module param \"allow_unsafe_interrupts\" to enable VFIO IOMMU support on this platform\n", |
1403 | __func__); | |
1404 | ret = -EPERM; | |
1405 | goto out_detach; | |
1406 | } | |
1407 | ||
eb165f05 | 1408 | if (iommu_capable(bus, IOMMU_CAP_CACHE_COHERENCY)) |
1ef3e2bc AW |
1409 | domain->prot |= IOMMU_CACHE; |
1410 | ||
73fa0d10 | 1411 | /* |
1ef3e2bc AW |
1412 | * Try to match an existing compatible domain. We don't want to |
1413 | * preclude an IOMMU driver supporting multiple bus_types and being | |
1414 | * able to include different bus_types in the same IOMMU domain, so | |
1415 | * we test whether the domains use the same iommu_ops rather than | |
1416 | * testing if they're on the same bus_type. | |
73fa0d10 | 1417 | */ |
1ef3e2bc AW |
1418 | list_for_each_entry(d, &iommu->domain_list, next) { |
1419 | if (d->domain->ops == domain->domain->ops && | |
1420 | d->prot == domain->prot) { | |
1421 | iommu_detach_group(domain->domain, iommu_group); | |
1422 | if (!iommu_attach_group(d->domain, iommu_group)) { | |
1423 | list_add(&group->next, &d->group_list); | |
1424 | iommu_domain_free(domain->domain); | |
1425 | kfree(domain); | |
1426 | mutex_unlock(&iommu->lock); | |
1427 | return 0; | |
1428 | } | |
1429 | ||
1430 | ret = iommu_attach_group(domain->domain, iommu_group); | |
1431 | if (ret) | |
1432 | goto out_domain; | |
1433 | } | |
73fa0d10 AW |
1434 | } |
1435 | ||
6fe1010d AW |
1436 | vfio_test_domain_fgsp(domain); |
1437 | ||
1ef3e2bc AW |
1438 | /* replay mappings on new domains */ |
1439 | ret = vfio_iommu_replay(iommu, domain); | |
1440 | if (ret) | |
1441 | goto out_detach; | |
1442 | ||
2c9f1af5 WY |
1443 | if (resv_msi) { |
1444 | ret = iommu_get_msi_cookie(domain->domain, resv_msi_base); | |
1445 | if (ret) | |
1446 | goto out_detach; | |
1447 | } | |
5d704992 | 1448 | |
1ef3e2bc | 1449 | list_add(&domain->next, &iommu->domain_list); |
73fa0d10 AW |
1450 | |
1451 | mutex_unlock(&iommu->lock); | |
1452 | ||
1453 | return 0; | |
1ef3e2bc AW |
1454 | |
1455 | out_detach: | |
1456 | iommu_detach_group(domain->domain, iommu_group); | |
1457 | out_domain: | |
1458 | iommu_domain_free(domain->domain); | |
1459 | out_free: | |
1460 | kfree(domain); | |
1461 | kfree(group); | |
1462 | mutex_unlock(&iommu->lock); | |
1463 | return ret; | |
1464 | } | |
1465 | ||
1466 | static void vfio_iommu_unmap_unpin_all(struct vfio_iommu *iommu) | |
1467 | { | |
1468 | struct rb_node *node; | |
1469 | ||
1470 | while ((node = rb_first(&iommu->dma_list))) | |
1471 | vfio_remove_dma(iommu, rb_entry(node, struct vfio_dma, node)); | |
73fa0d10 AW |
1472 | } |
1473 | ||
a54eb550 KW |
1474 | static void vfio_iommu_unmap_unpin_reaccount(struct vfio_iommu *iommu) |
1475 | { | |
1476 | struct rb_node *n, *p; | |
1477 | ||
1478 | n = rb_first(&iommu->dma_list); | |
1479 | for (; n; n = rb_next(n)) { | |
1480 | struct vfio_dma *dma; | |
1481 | long locked = 0, unlocked = 0; | |
1482 | ||
1483 | dma = rb_entry(n, struct vfio_dma, node); | |
1484 | unlocked += vfio_unmap_unpin(iommu, dma, false); | |
1485 | p = rb_first(&dma->pfn_list); | |
1486 | for (; p; p = rb_next(p)) { | |
1487 | struct vfio_pfn *vpfn = rb_entry(p, struct vfio_pfn, | |
1488 | node); | |
1489 | ||
1490 | if (!is_invalid_reserved_pfn(vpfn->pfn)) | |
1491 | locked++; | |
1492 | } | |
48d8476b | 1493 | vfio_lock_acct(dma, locked - unlocked, true); |
a54eb550 KW |
1494 | } |
1495 | } | |
1496 | ||
1497 | static void vfio_sanity_check_pfn_list(struct vfio_iommu *iommu) | |
1498 | { | |
1499 | struct rb_node *n; | |
1500 | ||
1501 | n = rb_first(&iommu->dma_list); | |
1502 | for (; n; n = rb_next(n)) { | |
1503 | struct vfio_dma *dma; | |
1504 | ||
1505 | dma = rb_entry(n, struct vfio_dma, node); | |
1506 | ||
1507 | if (WARN_ON(!RB_EMPTY_ROOT(&dma->pfn_list))) | |
1508 | break; | |
1509 | } | |
3cedd7d7 KW |
1510 | /* mdev vendor driver must unregister notifier */ |
1511 | WARN_ON(iommu->notifier.head); | |
a54eb550 KW |
1512 | } |
1513 | ||
73fa0d10 AW |
1514 | static void vfio_iommu_type1_detach_group(void *iommu_data, |
1515 | struct iommu_group *iommu_group) | |
1516 | { | |
1517 | struct vfio_iommu *iommu = iommu_data; | |
1ef3e2bc | 1518 | struct vfio_domain *domain; |
73fa0d10 AW |
1519 | struct vfio_group *group; |
1520 | ||
1521 | mutex_lock(&iommu->lock); | |
1522 | ||
a54eb550 KW |
1523 | if (iommu->external_domain) { |
1524 | group = find_iommu_group(iommu->external_domain, iommu_group); | |
1525 | if (group) { | |
1526 | list_del(&group->next); | |
1527 | kfree(group); | |
1528 | ||
1529 | if (list_empty(&iommu->external_domain->group_list)) { | |
1530 | vfio_sanity_check_pfn_list(iommu); | |
1531 | ||
1532 | if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu)) | |
1533 | vfio_iommu_unmap_unpin_all(iommu); | |
1534 | ||
1535 | kfree(iommu->external_domain); | |
1536 | iommu->external_domain = NULL; | |
1537 | } | |
1538 | goto detach_group_done; | |
1539 | } | |
1540 | } | |
1541 | ||
1ef3e2bc | 1542 | list_for_each_entry(domain, &iommu->domain_list, next) { |
7896c998 KW |
1543 | group = find_iommu_group(domain, iommu_group); |
1544 | if (!group) | |
1545 | continue; | |
1ef3e2bc | 1546 | |
7896c998 KW |
1547 | iommu_detach_group(domain->domain, iommu_group); |
1548 | list_del(&group->next); | |
1549 | kfree(group); | |
1550 | /* | |
a54eb550 KW |
1551 | * Group ownership provides privilege, if the group list is |
1552 | * empty, the domain goes away. If it's the last domain with | |
1553 | * iommu and external domain doesn't exist, then all the | |
1554 | * mappings go away too. If it's the last domain with iommu and | |
1555 | * external domain exist, update accounting | |
7896c998 KW |
1556 | */ |
1557 | if (list_empty(&domain->group_list)) { | |
a54eb550 KW |
1558 | if (list_is_singular(&iommu->domain_list)) { |
1559 | if (!iommu->external_domain) | |
1560 | vfio_iommu_unmap_unpin_all(iommu); | |
1561 | else | |
1562 | vfio_iommu_unmap_unpin_reaccount(iommu); | |
1563 | } | |
7896c998 KW |
1564 | iommu_domain_free(domain->domain); |
1565 | list_del(&domain->next); | |
1566 | kfree(domain); | |
73fa0d10 | 1567 | } |
a54eb550 | 1568 | break; |
73fa0d10 AW |
1569 | } |
1570 | ||
a54eb550 | 1571 | detach_group_done: |
73fa0d10 AW |
1572 | mutex_unlock(&iommu->lock); |
1573 | } | |
1574 | ||
1575 | static void *vfio_iommu_type1_open(unsigned long arg) | |
1576 | { | |
1577 | struct vfio_iommu *iommu; | |
1578 | ||
73fa0d10 AW |
1579 | iommu = kzalloc(sizeof(*iommu), GFP_KERNEL); |
1580 | if (!iommu) | |
1581 | return ERR_PTR(-ENOMEM); | |
1582 | ||
f5c9eceb WD |
1583 | switch (arg) { |
1584 | case VFIO_TYPE1_IOMMU: | |
1585 | break; | |
1586 | case VFIO_TYPE1_NESTING_IOMMU: | |
1587 | iommu->nesting = true; | |
544c05a6 | 1588 | /* fall through */ |
f5c9eceb WD |
1589 | case VFIO_TYPE1v2_IOMMU: |
1590 | iommu->v2 = true; | |
1591 | break; | |
1592 | default: | |
1593 | kfree(iommu); | |
1594 | return ERR_PTR(-EINVAL); | |
1595 | } | |
1596 | ||
1ef3e2bc | 1597 | INIT_LIST_HEAD(&iommu->domain_list); |
cd9b2268 | 1598 | iommu->dma_list = RB_ROOT; |
49285593 | 1599 | iommu->dma_avail = dma_entry_limit; |
73fa0d10 | 1600 | mutex_init(&iommu->lock); |
c086de81 | 1601 | BLOCKING_INIT_NOTIFIER_HEAD(&iommu->notifier); |
73fa0d10 AW |
1602 | |
1603 | return iommu; | |
1604 | } | |
1605 | ||
a54eb550 KW |
1606 | static void vfio_release_domain(struct vfio_domain *domain, bool external) |
1607 | { | |
1608 | struct vfio_group *group, *group_tmp; | |
1609 | ||
1610 | list_for_each_entry_safe(group, group_tmp, | |
1611 | &domain->group_list, next) { | |
1612 | if (!external) | |
1613 | iommu_detach_group(domain->domain, group->iommu_group); | |
1614 | list_del(&group->next); | |
1615 | kfree(group); | |
1616 | } | |
1617 | ||
1618 | if (!external) | |
1619 | iommu_domain_free(domain->domain); | |
1620 | } | |
1621 | ||
73fa0d10 AW |
1622 | static void vfio_iommu_type1_release(void *iommu_data) |
1623 | { | |
1624 | struct vfio_iommu *iommu = iommu_data; | |
1ef3e2bc | 1625 | struct vfio_domain *domain, *domain_tmp; |
a54eb550 KW |
1626 | |
1627 | if (iommu->external_domain) { | |
1628 | vfio_release_domain(iommu->external_domain, true); | |
1629 | vfio_sanity_check_pfn_list(iommu); | |
1630 | kfree(iommu->external_domain); | |
1631 | } | |
73fa0d10 | 1632 | |
1ef3e2bc | 1633 | vfio_iommu_unmap_unpin_all(iommu); |
73fa0d10 | 1634 | |
1ef3e2bc AW |
1635 | list_for_each_entry_safe(domain, domain_tmp, |
1636 | &iommu->domain_list, next) { | |
a54eb550 | 1637 | vfio_release_domain(domain, false); |
1ef3e2bc AW |
1638 | list_del(&domain->next); |
1639 | kfree(domain); | |
73fa0d10 | 1640 | } |
73fa0d10 AW |
1641 | kfree(iommu); |
1642 | } | |
1643 | ||
aa429318 AW |
1644 | static int vfio_domains_have_iommu_cache(struct vfio_iommu *iommu) |
1645 | { | |
1646 | struct vfio_domain *domain; | |
1647 | int ret = 1; | |
1648 | ||
1649 | mutex_lock(&iommu->lock); | |
1650 | list_for_each_entry(domain, &iommu->domain_list, next) { | |
1651 | if (!(domain->prot & IOMMU_CACHE)) { | |
1652 | ret = 0; | |
f5bfdbf2 | 1653 | break; |
aa429318 | 1654 | } |
73fa0d10 | 1655 | } |
aa429318 | 1656 | mutex_unlock(&iommu->lock); |
73fa0d10 | 1657 | |
aa429318 | 1658 | return ret; |
73fa0d10 AW |
1659 | } |
1660 | ||
1661 | static long vfio_iommu_type1_ioctl(void *iommu_data, | |
1662 | unsigned int cmd, unsigned long arg) | |
1663 | { | |
1664 | struct vfio_iommu *iommu = iommu_data; | |
1665 | unsigned long minsz; | |
1666 | ||
1667 | if (cmd == VFIO_CHECK_EXTENSION) { | |
1668 | switch (arg) { | |
1669 | case VFIO_TYPE1_IOMMU: | |
1ef3e2bc | 1670 | case VFIO_TYPE1v2_IOMMU: |
f5c9eceb | 1671 | case VFIO_TYPE1_NESTING_IOMMU: |
73fa0d10 | 1672 | return 1; |
aa429318 AW |
1673 | case VFIO_DMA_CC_IOMMU: |
1674 | if (!iommu) | |
1675 | return 0; | |
1676 | return vfio_domains_have_iommu_cache(iommu); | |
73fa0d10 AW |
1677 | default: |
1678 | return 0; | |
1679 | } | |
1680 | } else if (cmd == VFIO_IOMMU_GET_INFO) { | |
1681 | struct vfio_iommu_type1_info info; | |
1682 | ||
1683 | minsz = offsetofend(struct vfio_iommu_type1_info, iova_pgsizes); | |
1684 | ||
1685 | if (copy_from_user(&info, (void __user *)arg, minsz)) | |
1686 | return -EFAULT; | |
1687 | ||
1688 | if (info.argsz < minsz) | |
1689 | return -EINVAL; | |
1690 | ||
d4f50ee2 | 1691 | info.flags = VFIO_IOMMU_INFO_PGSIZES; |
73fa0d10 | 1692 | |
1ef3e2bc | 1693 | info.iova_pgsizes = vfio_pgsize_bitmap(iommu); |
73fa0d10 | 1694 | |
8160c4e4 MT |
1695 | return copy_to_user((void __user *)arg, &info, minsz) ? |
1696 | -EFAULT : 0; | |
73fa0d10 AW |
1697 | |
1698 | } else if (cmd == VFIO_IOMMU_MAP_DMA) { | |
1699 | struct vfio_iommu_type1_dma_map map; | |
1700 | uint32_t mask = VFIO_DMA_MAP_FLAG_READ | | |
1701 | VFIO_DMA_MAP_FLAG_WRITE; | |
1702 | ||
1703 | minsz = offsetofend(struct vfio_iommu_type1_dma_map, size); | |
1704 | ||
1705 | if (copy_from_user(&map, (void __user *)arg, minsz)) | |
1706 | return -EFAULT; | |
1707 | ||
1708 | if (map.argsz < minsz || map.flags & ~mask) | |
1709 | return -EINVAL; | |
1710 | ||
1711 | return vfio_dma_do_map(iommu, &map); | |
1712 | ||
1713 | } else if (cmd == VFIO_IOMMU_UNMAP_DMA) { | |
1714 | struct vfio_iommu_type1_dma_unmap unmap; | |
166fd7d9 | 1715 | long ret; |
73fa0d10 AW |
1716 | |
1717 | minsz = offsetofend(struct vfio_iommu_type1_dma_unmap, size); | |
1718 | ||
1719 | if (copy_from_user(&unmap, (void __user *)arg, minsz)) | |
1720 | return -EFAULT; | |
1721 | ||
1722 | if (unmap.argsz < minsz || unmap.flags) | |
1723 | return -EINVAL; | |
1724 | ||
166fd7d9 AW |
1725 | ret = vfio_dma_do_unmap(iommu, &unmap); |
1726 | if (ret) | |
1727 | return ret; | |
1728 | ||
8160c4e4 MT |
1729 | return copy_to_user((void __user *)arg, &unmap, minsz) ? |
1730 | -EFAULT : 0; | |
73fa0d10 AW |
1731 | } |
1732 | ||
1733 | return -ENOTTY; | |
1734 | } | |
1735 | ||
c086de81 | 1736 | static int vfio_iommu_type1_register_notifier(void *iommu_data, |
22195cbd | 1737 | unsigned long *events, |
c086de81 KW |
1738 | struct notifier_block *nb) |
1739 | { | |
1740 | struct vfio_iommu *iommu = iommu_data; | |
1741 | ||
22195cbd JS |
1742 | /* clear known events */ |
1743 | *events &= ~VFIO_IOMMU_NOTIFY_DMA_UNMAP; | |
1744 | ||
1745 | /* refuse to register if still events remaining */ | |
1746 | if (*events) | |
1747 | return -EINVAL; | |
1748 | ||
c086de81 KW |
1749 | return blocking_notifier_chain_register(&iommu->notifier, nb); |
1750 | } | |
1751 | ||
1752 | static int vfio_iommu_type1_unregister_notifier(void *iommu_data, | |
1753 | struct notifier_block *nb) | |
1754 | { | |
1755 | struct vfio_iommu *iommu = iommu_data; | |
1756 | ||
1757 | return blocking_notifier_chain_unregister(&iommu->notifier, nb); | |
1758 | } | |
1759 | ||
73fa0d10 | 1760 | static const struct vfio_iommu_driver_ops vfio_iommu_driver_ops_type1 = { |
c086de81 KW |
1761 | .name = "vfio-iommu-type1", |
1762 | .owner = THIS_MODULE, | |
1763 | .open = vfio_iommu_type1_open, | |
1764 | .release = vfio_iommu_type1_release, | |
1765 | .ioctl = vfio_iommu_type1_ioctl, | |
1766 | .attach_group = vfio_iommu_type1_attach_group, | |
1767 | .detach_group = vfio_iommu_type1_detach_group, | |
1768 | .pin_pages = vfio_iommu_type1_pin_pages, | |
1769 | .unpin_pages = vfio_iommu_type1_unpin_pages, | |
1770 | .register_notifier = vfio_iommu_type1_register_notifier, | |
1771 | .unregister_notifier = vfio_iommu_type1_unregister_notifier, | |
73fa0d10 AW |
1772 | }; |
1773 | ||
1774 | static int __init vfio_iommu_type1_init(void) | |
1775 | { | |
73fa0d10 AW |
1776 | return vfio_register_iommu_driver(&vfio_iommu_driver_ops_type1); |
1777 | } | |
1778 | ||
1779 | static void __exit vfio_iommu_type1_cleanup(void) | |
1780 | { | |
1781 | vfio_unregister_iommu_driver(&vfio_iommu_driver_ops_type1); | |
1782 | } | |
1783 | ||
1784 | module_init(vfio_iommu_type1_init); | |
1785 | module_exit(vfio_iommu_type1_cleanup); | |
1786 | ||
1787 | MODULE_VERSION(DRIVER_VERSION); | |
1788 | MODULE_LICENSE("GPL v2"); | |
1789 | MODULE_AUTHOR(DRIVER_AUTHOR); | |
1790 | MODULE_DESCRIPTION(DRIVER_DESC); |