]>
Commit | Line | Data |
---|---|---|
d2912cb1 | 1 | // SPDX-License-Identifier: GPL-2.0-only |
73fa0d10 AW |
2 | /* |
3 | * VFIO: IOMMU DMA mapping support for Type1 IOMMU | |
4 | * | |
5 | * Copyright (C) 2012 Red Hat, Inc. All rights reserved. | |
6 | * Author: Alex Williamson <alex.williamson@redhat.com> | |
7 | * | |
73fa0d10 AW |
8 | * Derived from original vfio: |
9 | * Copyright 2010 Cisco Systems, Inc. All rights reserved. | |
10 | * Author: Tom Lyon, pugs@cisco.com | |
11 | * | |
12 | * We arbitrarily define a Type1 IOMMU as one matching the below code. | |
13 | * It could be called the x86 IOMMU as it's designed for AMD-Vi & Intel | |
14 | * VT-d, but that makes it harder to re-use as theoretically anyone | |
15 | * implementing a similar IOMMU could make use of this. We expect the | |
16 | * IOMMU to support the IOMMU API and have few to no restrictions around | |
17 | * the IOVA range that can be mapped. The Type1 IOMMU is currently | |
18 | * optimized for relatively static mappings of a userspace process with | |
19 | * userpsace pages pinned into memory. We also assume devices and IOMMU | |
20 | * domains are PCI based as the IOMMU API is still centered around a | |
21 | * device/bus interface rather than a group interface. | |
22 | */ | |
23 | ||
24 | #include <linux/compat.h> | |
25 | #include <linux/device.h> | |
26 | #include <linux/fs.h> | |
27 | #include <linux/iommu.h> | |
28 | #include <linux/module.h> | |
29 | #include <linux/mm.h> | |
4dbe59a6 | 30 | #include <linux/kthread.h> |
cd9b2268 | 31 | #include <linux/rbtree.h> |
3f07c014 | 32 | #include <linux/sched/signal.h> |
6e84f315 | 33 | #include <linux/sched/mm.h> |
73fa0d10 AW |
34 | #include <linux/slab.h> |
35 | #include <linux/uaccess.h> | |
36 | #include <linux/vfio.h> | |
37 | #include <linux/workqueue.h> | |
a54eb550 | 38 | #include <linux/mdev.h> |
c086de81 | 39 | #include <linux/notifier.h> |
5d704992 | 40 | #include <linux/dma-iommu.h> |
9d72f87b | 41 | #include <linux/irqdomain.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 | ||
49285593 AW |
59 | static unsigned int dma_entry_limit __read_mostly = U16_MAX; |
60 | module_param_named(dma_entry_limit, dma_entry_limit, uint, 0644); | |
61 | MODULE_PARM_DESC(dma_entry_limit, | |
62 | "Maximum number of user DMA mappings per container (65535)."); | |
63 | ||
73fa0d10 | 64 | struct vfio_iommu { |
1ef3e2bc | 65 | struct list_head domain_list; |
1108696a | 66 | struct list_head iova_list; |
a54eb550 | 67 | struct vfio_domain *external_domain; /* domain for external user */ |
73fa0d10 | 68 | struct mutex lock; |
cd9b2268 | 69 | struct rb_root dma_list; |
c086de81 | 70 | struct blocking_notifier_head notifier; |
49285593 | 71 | unsigned int dma_avail; |
cade075f | 72 | uint64_t pgsize_bitmap; |
f5c9eceb WD |
73 | bool v2; |
74 | bool nesting; | |
d6a4c185 | 75 | bool dirty_page_tracking; |
95fc87b4 | 76 | bool pinned_page_dirty_scope; |
1ef3e2bc AW |
77 | }; |
78 | ||
79 | struct vfio_domain { | |
80 | struct iommu_domain *domain; | |
81 | struct list_head next; | |
73fa0d10 | 82 | struct list_head group_list; |
1ef3e2bc | 83 | int prot; /* IOMMU_CACHE */ |
6fe1010d | 84 | bool fgsp; /* Fine-grained super pages */ |
73fa0d10 AW |
85 | }; |
86 | ||
87 | struct vfio_dma { | |
cd9b2268 | 88 | struct rb_node node; |
73fa0d10 AW |
89 | dma_addr_t iova; /* Device address */ |
90 | unsigned long vaddr; /* Process virtual addr */ | |
166fd7d9 | 91 | size_t size; /* Map size (bytes) */ |
73fa0d10 | 92 | int prot; /* IOMMU_READ/WRITE */ |
a54eb550 | 93 | bool iommu_mapped; |
48d8476b | 94 | bool lock_cap; /* capable(CAP_IPC_LOCK) */ |
8f0d5bb9 | 95 | struct task_struct *task; |
a54eb550 | 96 | struct rb_root pfn_list; /* Ex-user pinned pfn list */ |
d6a4c185 | 97 | unsigned long *bitmap; |
73fa0d10 AW |
98 | }; |
99 | ||
100 | struct vfio_group { | |
101 | struct iommu_group *iommu_group; | |
102 | struct list_head next; | |
7bd50f0c | 103 | bool mdev_group; /* An mdev group */ |
95fc87b4 | 104 | bool pinned_page_dirty_scope; |
73fa0d10 AW |
105 | }; |
106 | ||
1108696a SK |
107 | struct vfio_iova { |
108 | struct list_head list; | |
109 | dma_addr_t start; | |
110 | dma_addr_t end; | |
111 | }; | |
112 | ||
a54eb550 KW |
113 | /* |
114 | * Guest RAM pinning working set or DMA target | |
115 | */ | |
116 | struct vfio_pfn { | |
117 | struct rb_node node; | |
118 | dma_addr_t iova; /* Device address */ | |
119 | unsigned long pfn; /* Host pfn */ | |
65817085 | 120 | unsigned int ref_count; |
a54eb550 KW |
121 | }; |
122 | ||
6bd06f5a SS |
123 | struct vfio_regions { |
124 | struct list_head list; | |
125 | dma_addr_t iova; | |
126 | phys_addr_t phys; | |
127 | size_t len; | |
128 | }; | |
129 | ||
a54eb550 KW |
130 | #define IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu) \ |
131 | (!list_empty(&iommu->domain_list)) | |
132 | ||
d6a4c185 KW |
133 | #define DIRTY_BITMAP_BYTES(n) (ALIGN(n, BITS_PER_TYPE(u64)) / BITS_PER_BYTE) |
134 | ||
135 | /* | |
136 | * Input argument of number of bits to bitmap_set() is unsigned integer, which | |
137 | * further casts to signed integer for unaligned multi-bit operation, | |
138 | * __bitmap_set(). | |
139 | * Then maximum bitmap size supported is 2^31 bits divided by 2^3 bits/byte, | |
140 | * that is 2^28 (256 MB) which maps to 2^31 * 2^12 = 2^43 (8TB) on 4K page | |
141 | * system. | |
142 | */ | |
143 | #define DIRTY_BITMAP_PAGES_MAX ((u64)INT_MAX) | |
144 | #define DIRTY_BITMAP_SIZE_MAX DIRTY_BITMAP_BYTES(DIRTY_BITMAP_PAGES_MAX) | |
145 | ||
a54eb550 KW |
146 | static int put_pfn(unsigned long pfn, int prot); |
147 | ||
95fc87b4 KW |
148 | static struct vfio_group *vfio_iommu_find_iommu_group(struct vfio_iommu *iommu, |
149 | struct iommu_group *iommu_group); | |
150 | ||
151 | static void update_pinned_page_dirty_scope(struct vfio_iommu *iommu); | |
73fa0d10 AW |
152 | /* |
153 | * This code handles mapping and unmapping of user data buffers | |
154 | * into DMA'ble space using the IOMMU | |
155 | */ | |
156 | ||
cd9b2268 AW |
157 | static struct vfio_dma *vfio_find_dma(struct vfio_iommu *iommu, |
158 | dma_addr_t start, size_t size) | |
159 | { | |
160 | struct rb_node *node = iommu->dma_list.rb_node; | |
161 | ||
162 | while (node) { | |
163 | struct vfio_dma *dma = rb_entry(node, struct vfio_dma, node); | |
164 | ||
165 | if (start + size <= dma->iova) | |
166 | node = node->rb_left; | |
166fd7d9 | 167 | else if (start >= dma->iova + dma->size) |
cd9b2268 AW |
168 | node = node->rb_right; |
169 | else | |
170 | return dma; | |
171 | } | |
172 | ||
173 | return NULL; | |
174 | } | |
175 | ||
1ef3e2bc | 176 | static void vfio_link_dma(struct vfio_iommu *iommu, struct vfio_dma *new) |
cd9b2268 AW |
177 | { |
178 | struct rb_node **link = &iommu->dma_list.rb_node, *parent = NULL; | |
179 | struct vfio_dma *dma; | |
180 | ||
181 | while (*link) { | |
182 | parent = *link; | |
183 | dma = rb_entry(parent, struct vfio_dma, node); | |
184 | ||
166fd7d9 | 185 | if (new->iova + new->size <= dma->iova) |
cd9b2268 AW |
186 | link = &(*link)->rb_left; |
187 | else | |
188 | link = &(*link)->rb_right; | |
189 | } | |
190 | ||
191 | rb_link_node(&new->node, parent, link); | |
192 | rb_insert_color(&new->node, &iommu->dma_list); | |
193 | } | |
194 | ||
1ef3e2bc | 195 | static void vfio_unlink_dma(struct vfio_iommu *iommu, struct vfio_dma *old) |
cd9b2268 AW |
196 | { |
197 | rb_erase(&old->node, &iommu->dma_list); | |
198 | } | |
199 | ||
d6a4c185 KW |
200 | |
201 | static int vfio_dma_bitmap_alloc(struct vfio_dma *dma, size_t pgsize) | |
202 | { | |
203 | uint64_t npages = dma->size / pgsize; | |
204 | ||
205 | if (npages > DIRTY_BITMAP_PAGES_MAX) | |
206 | return -EINVAL; | |
207 | ||
208 | /* | |
209 | * Allocate extra 64 bits that are used to calculate shift required for | |
210 | * bitmap_shift_left() to manipulate and club unaligned number of pages | |
211 | * in adjacent vfio_dma ranges. | |
212 | */ | |
213 | dma->bitmap = kvzalloc(DIRTY_BITMAP_BYTES(npages) + sizeof(u64), | |
214 | GFP_KERNEL); | |
215 | if (!dma->bitmap) | |
216 | return -ENOMEM; | |
217 | ||
218 | return 0; | |
219 | } | |
220 | ||
221 | static void vfio_dma_bitmap_free(struct vfio_dma *dma) | |
222 | { | |
223 | kfree(dma->bitmap); | |
224 | dma->bitmap = NULL; | |
225 | } | |
226 | ||
227 | static void vfio_dma_populate_bitmap(struct vfio_dma *dma, size_t pgsize) | |
228 | { | |
229 | struct rb_node *p; | |
cd0bb41e | 230 | unsigned long pgshift = __ffs(pgsize); |
d6a4c185 KW |
231 | |
232 | for (p = rb_first(&dma->pfn_list); p; p = rb_next(p)) { | |
233 | struct vfio_pfn *vpfn = rb_entry(p, struct vfio_pfn, node); | |
234 | ||
cd0bb41e | 235 | bitmap_set(dma->bitmap, (vpfn->iova - dma->iova) >> pgshift, 1); |
d6a4c185 KW |
236 | } |
237 | } | |
238 | ||
239 | static int vfio_dma_bitmap_alloc_all(struct vfio_iommu *iommu, size_t pgsize) | |
240 | { | |
241 | struct rb_node *n; | |
242 | ||
243 | for (n = rb_first(&iommu->dma_list); n; n = rb_next(n)) { | |
244 | struct vfio_dma *dma = rb_entry(n, struct vfio_dma, node); | |
245 | int ret; | |
246 | ||
247 | ret = vfio_dma_bitmap_alloc(dma, pgsize); | |
248 | if (ret) { | |
249 | struct rb_node *p; | |
250 | ||
251 | for (p = rb_prev(n); p; p = rb_prev(p)) { | |
252 | struct vfio_dma *dma = rb_entry(n, | |
253 | struct vfio_dma, node); | |
254 | ||
255 | vfio_dma_bitmap_free(dma); | |
256 | } | |
257 | return ret; | |
258 | } | |
259 | vfio_dma_populate_bitmap(dma, pgsize); | |
260 | } | |
261 | return 0; | |
262 | } | |
263 | ||
264 | static void vfio_dma_bitmap_free_all(struct vfio_iommu *iommu) | |
265 | { | |
266 | struct rb_node *n; | |
267 | ||
268 | for (n = rb_first(&iommu->dma_list); n; n = rb_next(n)) { | |
269 | struct vfio_dma *dma = rb_entry(n, struct vfio_dma, node); | |
270 | ||
271 | vfio_dma_bitmap_free(dma); | |
272 | } | |
273 | } | |
274 | ||
a54eb550 KW |
275 | /* |
276 | * Helper Functions for host iova-pfn list | |
277 | */ | |
278 | static struct vfio_pfn *vfio_find_vpfn(struct vfio_dma *dma, dma_addr_t iova) | |
279 | { | |
280 | struct vfio_pfn *vpfn; | |
281 | struct rb_node *node = dma->pfn_list.rb_node; | |
282 | ||
283 | while (node) { | |
284 | vpfn = rb_entry(node, struct vfio_pfn, node); | |
285 | ||
286 | if (iova < vpfn->iova) | |
287 | node = node->rb_left; | |
288 | else if (iova > vpfn->iova) | |
289 | node = node->rb_right; | |
290 | else | |
291 | return vpfn; | |
292 | } | |
293 | return NULL; | |
294 | } | |
295 | ||
296 | static void vfio_link_pfn(struct vfio_dma *dma, | |
297 | struct vfio_pfn *new) | |
298 | { | |
299 | struct rb_node **link, *parent = NULL; | |
300 | struct vfio_pfn *vpfn; | |
301 | ||
302 | link = &dma->pfn_list.rb_node; | |
303 | while (*link) { | |
304 | parent = *link; | |
305 | vpfn = rb_entry(parent, struct vfio_pfn, node); | |
306 | ||
307 | if (new->iova < vpfn->iova) | |
308 | link = &(*link)->rb_left; | |
309 | else | |
310 | link = &(*link)->rb_right; | |
311 | } | |
312 | ||
313 | rb_link_node(&new->node, parent, link); | |
314 | rb_insert_color(&new->node, &dma->pfn_list); | |
315 | } | |
316 | ||
317 | static void vfio_unlink_pfn(struct vfio_dma *dma, struct vfio_pfn *old) | |
318 | { | |
319 | rb_erase(&old->node, &dma->pfn_list); | |
320 | } | |
321 | ||
322 | static int vfio_add_to_pfn_list(struct vfio_dma *dma, dma_addr_t iova, | |
323 | unsigned long pfn) | |
324 | { | |
325 | struct vfio_pfn *vpfn; | |
326 | ||
327 | vpfn = kzalloc(sizeof(*vpfn), GFP_KERNEL); | |
328 | if (!vpfn) | |
329 | return -ENOMEM; | |
330 | ||
331 | vpfn->iova = iova; | |
332 | vpfn->pfn = pfn; | |
65817085 | 333 | vpfn->ref_count = 1; |
a54eb550 KW |
334 | vfio_link_pfn(dma, vpfn); |
335 | return 0; | |
336 | } | |
337 | ||
338 | static void vfio_remove_from_pfn_list(struct vfio_dma *dma, | |
339 | struct vfio_pfn *vpfn) | |
340 | { | |
341 | vfio_unlink_pfn(dma, vpfn); | |
342 | kfree(vpfn); | |
343 | } | |
344 | ||
345 | static struct vfio_pfn *vfio_iova_get_vfio_pfn(struct vfio_dma *dma, | |
346 | unsigned long iova) | |
347 | { | |
348 | struct vfio_pfn *vpfn = vfio_find_vpfn(dma, iova); | |
349 | ||
350 | if (vpfn) | |
65817085 | 351 | vpfn->ref_count++; |
a54eb550 KW |
352 | return vpfn; |
353 | } | |
354 | ||
355 | static int vfio_iova_put_vfio_pfn(struct vfio_dma *dma, struct vfio_pfn *vpfn) | |
356 | { | |
357 | int ret = 0; | |
358 | ||
65817085 KW |
359 | vpfn->ref_count--; |
360 | if (!vpfn->ref_count) { | |
a54eb550 KW |
361 | ret = put_pfn(vpfn->pfn, dma->prot); |
362 | vfio_remove_from_pfn_list(dma, vpfn); | |
363 | } | |
364 | return ret; | |
365 | } | |
366 | ||
48d8476b | 367 | static int vfio_lock_acct(struct vfio_dma *dma, long npage, bool async) |
73fa0d10 | 368 | { |
73fa0d10 | 369 | struct mm_struct *mm; |
0cfef2b7 | 370 | int ret; |
73fa0d10 | 371 | |
3624a248 | 372 | if (!npage) |
0cfef2b7 | 373 | return 0; |
3624a248 | 374 | |
48d8476b | 375 | mm = async ? get_task_mm(dma->task) : dma->task->mm; |
3624a248 | 376 | if (!mm) |
0cfef2b7 | 377 | return -ESRCH; /* process exited */ |
73fa0d10 | 378 | |
d8ed45c5 | 379 | ret = mmap_write_lock_killable(mm); |
0cfef2b7 | 380 | if (!ret) { |
79eb597c DJ |
381 | ret = __account_locked_vm(mm, abs(npage), npage > 0, dma->task, |
382 | dma->lock_cap); | |
d8ed45c5 | 383 | mmap_write_unlock(mm); |
6c38c055 AW |
384 | } |
385 | ||
48d8476b | 386 | if (async) |
3624a248 | 387 | mmput(mm); |
0cfef2b7 AW |
388 | |
389 | return ret; | |
73fa0d10 AW |
390 | } |
391 | ||
392 | /* | |
393 | * Some mappings aren't backed by a struct page, for example an mmap'd | |
394 | * MMIO range for our own or another device. These use a different | |
395 | * pfn conversion and shouldn't be tracked as locked pages. | |
026948f0 BL |
396 | * For compound pages, any driver that sets the reserved bit in head |
397 | * page needs to set the reserved bit in all subpages to be safe. | |
73fa0d10 AW |
398 | */ |
399 | static bool is_invalid_reserved_pfn(unsigned long pfn) | |
400 | { | |
026948f0 BL |
401 | if (pfn_valid(pfn)) |
402 | return PageReserved(pfn_to_page(pfn)); | |
73fa0d10 AW |
403 | |
404 | return true; | |
405 | } | |
406 | ||
407 | static int put_pfn(unsigned long pfn, int prot) | |
408 | { | |
409 | if (!is_invalid_reserved_pfn(pfn)) { | |
410 | struct page *page = pfn_to_page(pfn); | |
19fed0da | 411 | |
f1f6a7dd | 412 | unpin_user_pages_dirty_lock(&page, 1, prot & IOMMU_WRITE); |
73fa0d10 AW |
413 | return 1; |
414 | } | |
415 | return 0; | |
416 | } | |
417 | ||
41311242 AW |
418 | static int follow_fault_pfn(struct vm_area_struct *vma, struct mm_struct *mm, |
419 | unsigned long vaddr, unsigned long *pfn, | |
420 | bool write_fault) | |
421 | { | |
422 | int ret; | |
423 | ||
424 | ret = follow_pfn(vma, vaddr, pfn); | |
425 | if (ret) { | |
426 | bool unlocked = false; | |
427 | ||
64019a2e | 428 | ret = fixup_user_fault(mm, vaddr, |
41311242 AW |
429 | FAULT_FLAG_REMOTE | |
430 | (write_fault ? FAULT_FLAG_WRITE : 0), | |
431 | &unlocked); | |
432 | if (unlocked) | |
433 | return -EAGAIN; | |
434 | ||
435 | if (ret) | |
436 | return ret; | |
437 | ||
438 | ret = follow_pfn(vma, vaddr, pfn); | |
439 | } | |
440 | ||
441 | return ret; | |
442 | } | |
443 | ||
ea85cf35 KW |
444 | static int vaddr_get_pfn(struct mm_struct *mm, unsigned long vaddr, |
445 | int prot, unsigned long *pfn) | |
73fa0d10 AW |
446 | { |
447 | struct page *page[1]; | |
448 | struct vm_area_struct *vma; | |
bb94b55a | 449 | unsigned int flags = 0; |
ea85cf35 | 450 | int ret; |
73fa0d10 | 451 | |
bb94b55a JG |
452 | if (prot & IOMMU_WRITE) |
453 | flags |= FOLL_WRITE; | |
454 | ||
d8ed45c5 | 455 | mmap_read_lock(mm); |
64019a2e | 456 | ret = pin_user_pages_remote(mm, vaddr, 1, flags | FOLL_LONGTERM, |
3567813e | 457 | page, NULL, NULL); |
ea85cf35 | 458 | if (ret == 1) { |
73fa0d10 | 459 | *pfn = page_to_pfn(page[0]); |
3567813e JH |
460 | ret = 0; |
461 | goto done; | |
73fa0d10 AW |
462 | } |
463 | ||
6cf5354c AK |
464 | vaddr = untagged_addr(vaddr); |
465 | ||
41311242 | 466 | retry: |
ea85cf35 | 467 | vma = find_vma_intersection(mm, vaddr, vaddr + 1); |
73fa0d10 AW |
468 | |
469 | if (vma && vma->vm_flags & VM_PFNMAP) { | |
41311242 AW |
470 | ret = follow_fault_pfn(vma, mm, vaddr, pfn, prot & IOMMU_WRITE); |
471 | if (ret == -EAGAIN) | |
472 | goto retry; | |
473 | ||
474 | if (!ret && !is_invalid_reserved_pfn(*pfn)) | |
475 | ret = -EFAULT; | |
73fa0d10 | 476 | } |
3567813e | 477 | done: |
d8ed45c5 | 478 | mmap_read_unlock(mm); |
73fa0d10 AW |
479 | return ret; |
480 | } | |
481 | ||
166fd7d9 AW |
482 | /* |
483 | * Attempt to pin pages. We really don't want to track all the pfns and | |
484 | * the iommu can only map chunks of consecutive pfns anyway, so get the | |
485 | * first page and all consecutive pages with the same locking. | |
486 | */ | |
8f0d5bb9 | 487 | static long vfio_pin_pages_remote(struct vfio_dma *dma, unsigned long vaddr, |
7cb671e7 | 488 | long npage, unsigned long *pfn_base, |
48d8476b | 489 | unsigned long limit) |
73fa0d10 | 490 | { |
7cb671e7 | 491 | unsigned long pfn = 0; |
6c38c055 | 492 | long ret, pinned = 0, lock_acct = 0; |
89c29def | 493 | bool rsvd; |
a54eb550 | 494 | dma_addr_t iova = vaddr - dma->vaddr + dma->iova; |
73fa0d10 | 495 | |
6c38c055 AW |
496 | /* This code path is only user initiated */ |
497 | if (!current->mm) | |
166fd7d9 | 498 | return -ENODEV; |
73fa0d10 | 499 | |
6c38c055 | 500 | ret = vaddr_get_pfn(current->mm, vaddr, dma->prot, pfn_base); |
166fd7d9 | 501 | if (ret) |
6c38c055 | 502 | return ret; |
73fa0d10 | 503 | |
6c38c055 | 504 | pinned++; |
89c29def | 505 | rsvd = is_invalid_reserved_pfn(*pfn_base); |
73fa0d10 | 506 | |
a54eb550 KW |
507 | /* |
508 | * Reserved pages aren't counted against the user, externally pinned | |
509 | * pages are already counted against the user. | |
510 | */ | |
89c29def | 511 | if (!rsvd && !vfio_find_vpfn(dma, iova)) { |
48d8476b | 512 | if (!dma->lock_cap && current->mm->locked_vm + 1 > limit) { |
a54eb550 KW |
513 | put_pfn(*pfn_base, dma->prot); |
514 | pr_warn("%s: RLIMIT_MEMLOCK (%ld) exceeded\n", __func__, | |
515 | limit << PAGE_SHIFT); | |
6c38c055 | 516 | return -ENOMEM; |
a54eb550 KW |
517 | } |
518 | lock_acct++; | |
5c6c2b21 AW |
519 | } |
520 | ||
6c38c055 AW |
521 | if (unlikely(disable_hugepages)) |
522 | goto out; | |
73fa0d10 | 523 | |
6c38c055 AW |
524 | /* Lock all the consecutive pages from pfn_base */ |
525 | for (vaddr += PAGE_SIZE, iova += PAGE_SIZE; pinned < npage; | |
526 | pinned++, vaddr += PAGE_SIZE, iova += PAGE_SIZE) { | |
6c38c055 AW |
527 | ret = vaddr_get_pfn(current->mm, vaddr, dma->prot, &pfn); |
528 | if (ret) | |
529 | break; | |
530 | ||
89c29def AW |
531 | if (pfn != *pfn_base + pinned || |
532 | rsvd != is_invalid_reserved_pfn(pfn)) { | |
6c38c055 AW |
533 | put_pfn(pfn, dma->prot); |
534 | break; | |
535 | } | |
166fd7d9 | 536 | |
89c29def | 537 | if (!rsvd && !vfio_find_vpfn(dma, iova)) { |
48d8476b | 538 | if (!dma->lock_cap && |
6c38c055 | 539 | current->mm->locked_vm + lock_acct + 1 > limit) { |
a54eb550 | 540 | put_pfn(pfn, dma->prot); |
6c38c055 AW |
541 | pr_warn("%s: RLIMIT_MEMLOCK (%ld) exceeded\n", |
542 | __func__, limit << PAGE_SHIFT); | |
0cfef2b7 AW |
543 | ret = -ENOMEM; |
544 | goto unpin_out; | |
a54eb550 | 545 | } |
6c38c055 | 546 | lock_acct++; |
166fd7d9 AW |
547 | } |
548 | } | |
549 | ||
6c38c055 | 550 | out: |
48d8476b | 551 | ret = vfio_lock_acct(dma, lock_acct, false); |
0cfef2b7 AW |
552 | |
553 | unpin_out: | |
554 | if (ret) { | |
89c29def AW |
555 | if (!rsvd) { |
556 | for (pfn = *pfn_base ; pinned ; pfn++, pinned--) | |
557 | put_pfn(pfn, dma->prot); | |
558 | } | |
0cfef2b7 AW |
559 | |
560 | return ret; | |
561 | } | |
166fd7d9 | 562 | |
6c38c055 | 563 | return pinned; |
166fd7d9 AW |
564 | } |
565 | ||
a54eb550 KW |
566 | static long vfio_unpin_pages_remote(struct vfio_dma *dma, dma_addr_t iova, |
567 | unsigned long pfn, long npage, | |
568 | bool do_accounting) | |
166fd7d9 | 569 | { |
a54eb550 | 570 | long unlocked = 0, locked = 0; |
166fd7d9 AW |
571 | long i; |
572 | ||
6c38c055 | 573 | for (i = 0; i < npage; i++, iova += PAGE_SIZE) { |
a54eb550 KW |
574 | if (put_pfn(pfn++, dma->prot)) { |
575 | unlocked++; | |
6c38c055 | 576 | if (vfio_find_vpfn(dma, iova)) |
a54eb550 KW |
577 | locked++; |
578 | } | |
579 | } | |
580 | ||
581 | if (do_accounting) | |
48d8476b | 582 | vfio_lock_acct(dma, locked - unlocked, true); |
a54eb550 KW |
583 | |
584 | return unlocked; | |
585 | } | |
586 | ||
587 | static int vfio_pin_page_external(struct vfio_dma *dma, unsigned long vaddr, | |
588 | unsigned long *pfn_base, bool do_accounting) | |
589 | { | |
a54eb550 KW |
590 | struct mm_struct *mm; |
591 | int ret; | |
a54eb550 KW |
592 | |
593 | mm = get_task_mm(dma->task); | |
594 | if (!mm) | |
595 | return -ENODEV; | |
596 | ||
597 | ret = vaddr_get_pfn(mm, vaddr, dma->prot, pfn_base); | |
80dbe1fb | 598 | if (!ret && do_accounting && !is_invalid_reserved_pfn(*pfn_base)) { |
48d8476b | 599 | ret = vfio_lock_acct(dma, 1, true); |
0cfef2b7 AW |
600 | if (ret) { |
601 | put_pfn(*pfn_base, dma->prot); | |
80dbe1fb AW |
602 | if (ret == -ENOMEM) |
603 | pr_warn("%s: Task %s (%d) RLIMIT_MEMLOCK " | |
604 | "(%ld) exceeded\n", __func__, | |
605 | dma->task->comm, task_pid_nr(dma->task), | |
606 | task_rlimit(dma->task, RLIMIT_MEMLOCK)); | |
0cfef2b7 AW |
607 | } |
608 | } | |
609 | ||
a54eb550 KW |
610 | mmput(mm); |
611 | return ret; | |
612 | } | |
613 | ||
614 | static int vfio_unpin_page_external(struct vfio_dma *dma, dma_addr_t iova, | |
615 | bool do_accounting) | |
616 | { | |
617 | int unlocked; | |
618 | struct vfio_pfn *vpfn = vfio_find_vpfn(dma, iova); | |
619 | ||
620 | if (!vpfn) | |
621 | return 0; | |
622 | ||
623 | unlocked = vfio_iova_put_vfio_pfn(dma, vpfn); | |
166fd7d9 AW |
624 | |
625 | if (do_accounting) | |
48d8476b | 626 | vfio_lock_acct(dma, -unlocked, true); |
166fd7d9 AW |
627 | |
628 | return unlocked; | |
629 | } | |
630 | ||
a54eb550 | 631 | static int vfio_iommu_type1_pin_pages(void *iommu_data, |
95fc87b4 | 632 | struct iommu_group *iommu_group, |
a54eb550 KW |
633 | unsigned long *user_pfn, |
634 | int npage, int prot, | |
635 | unsigned long *phys_pfn) | |
636 | { | |
637 | struct vfio_iommu *iommu = iommu_data; | |
95fc87b4 | 638 | struct vfio_group *group; |
a54eb550 KW |
639 | int i, j, ret; |
640 | unsigned long remote_vaddr; | |
641 | struct vfio_dma *dma; | |
642 | bool do_accounting; | |
643 | ||
644 | if (!iommu || !user_pfn || !phys_pfn) | |
645 | return -EINVAL; | |
646 | ||
647 | /* Supported for v2 version only */ | |
648 | if (!iommu->v2) | |
649 | return -EACCES; | |
650 | ||
651 | mutex_lock(&iommu->lock); | |
652 | ||
c086de81 | 653 | /* Fail if notifier list is empty */ |
be068fa2 | 654 | if (!iommu->notifier.head) { |
a54eb550 KW |
655 | ret = -EINVAL; |
656 | goto pin_done; | |
657 | } | |
658 | ||
659 | /* | |
660 | * If iommu capable domain exist in the container then all pages are | |
661 | * already pinned and accounted. Accouting should be done if there is no | |
662 | * iommu capable domain in the container. | |
663 | */ | |
664 | do_accounting = !IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu); | |
665 | ||
666 | for (i = 0; i < npage; i++) { | |
667 | dma_addr_t iova; | |
668 | struct vfio_pfn *vpfn; | |
669 | ||
670 | iova = user_pfn[i] << PAGE_SHIFT; | |
2b8bb1d7 | 671 | dma = vfio_find_dma(iommu, iova, PAGE_SIZE); |
a54eb550 KW |
672 | if (!dma) { |
673 | ret = -EINVAL; | |
674 | goto pin_unwind; | |
675 | } | |
676 | ||
677 | if ((dma->prot & prot) != prot) { | |
678 | ret = -EPERM; | |
679 | goto pin_unwind; | |
680 | } | |
681 | ||
682 | vpfn = vfio_iova_get_vfio_pfn(dma, iova); | |
683 | if (vpfn) { | |
684 | phys_pfn[i] = vpfn->pfn; | |
685 | continue; | |
686 | } | |
687 | ||
0ea971f8 | 688 | remote_vaddr = dma->vaddr + (iova - dma->iova); |
a54eb550 KW |
689 | ret = vfio_pin_page_external(dma, remote_vaddr, &phys_pfn[i], |
690 | do_accounting); | |
80dbe1fb | 691 | if (ret) |
a54eb550 | 692 | goto pin_unwind; |
a54eb550 KW |
693 | |
694 | ret = vfio_add_to_pfn_list(dma, iova, phys_pfn[i]); | |
695 | if (ret) { | |
2e6cfd49 XX |
696 | if (put_pfn(phys_pfn[i], dma->prot) && do_accounting) |
697 | vfio_lock_acct(dma, -1, true); | |
a54eb550 KW |
698 | goto pin_unwind; |
699 | } | |
d6a4c185 KW |
700 | |
701 | if (iommu->dirty_page_tracking) { | |
702 | unsigned long pgshift = __ffs(iommu->pgsize_bitmap); | |
703 | ||
704 | /* | |
705 | * Bitmap populated with the smallest supported page | |
706 | * size | |
707 | */ | |
708 | bitmap_set(dma->bitmap, | |
709 | (iova - dma->iova) >> pgshift, 1); | |
710 | } | |
a54eb550 | 711 | } |
a54eb550 | 712 | ret = i; |
95fc87b4 KW |
713 | |
714 | group = vfio_iommu_find_iommu_group(iommu, iommu_group); | |
715 | if (!group->pinned_page_dirty_scope) { | |
716 | group->pinned_page_dirty_scope = true; | |
717 | update_pinned_page_dirty_scope(iommu); | |
718 | } | |
719 | ||
a54eb550 KW |
720 | goto pin_done; |
721 | ||
722 | pin_unwind: | |
723 | phys_pfn[i] = 0; | |
724 | for (j = 0; j < i; j++) { | |
725 | dma_addr_t iova; | |
726 | ||
727 | iova = user_pfn[j] << PAGE_SHIFT; | |
2b8bb1d7 | 728 | dma = vfio_find_dma(iommu, iova, PAGE_SIZE); |
a54eb550 KW |
729 | vfio_unpin_page_external(dma, iova, do_accounting); |
730 | phys_pfn[j] = 0; | |
731 | } | |
732 | pin_done: | |
733 | mutex_unlock(&iommu->lock); | |
734 | return ret; | |
735 | } | |
736 | ||
737 | static int vfio_iommu_type1_unpin_pages(void *iommu_data, | |
738 | unsigned long *user_pfn, | |
739 | int npage) | |
740 | { | |
741 | struct vfio_iommu *iommu = iommu_data; | |
742 | bool do_accounting; | |
743 | int i; | |
744 | ||
745 | if (!iommu || !user_pfn) | |
746 | return -EINVAL; | |
747 | ||
748 | /* Supported for v2 version only */ | |
749 | if (!iommu->v2) | |
750 | return -EACCES; | |
751 | ||
752 | mutex_lock(&iommu->lock); | |
753 | ||
a54eb550 KW |
754 | do_accounting = !IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu); |
755 | for (i = 0; i < npage; i++) { | |
756 | struct vfio_dma *dma; | |
757 | dma_addr_t iova; | |
758 | ||
759 | iova = user_pfn[i] << PAGE_SHIFT; | |
2b8bb1d7 | 760 | dma = vfio_find_dma(iommu, iova, PAGE_SIZE); |
a54eb550 KW |
761 | if (!dma) |
762 | goto unpin_exit; | |
763 | vfio_unpin_page_external(dma, iova, do_accounting); | |
764 | } | |
765 | ||
766 | unpin_exit: | |
767 | mutex_unlock(&iommu->lock); | |
768 | return i > npage ? npage : (i > 0 ? i : -EINVAL); | |
769 | } | |
770 | ||
6bd06f5a | 771 | static long vfio_sync_unpin(struct vfio_dma *dma, struct vfio_domain *domain, |
a7d20dc1 WD |
772 | struct list_head *regions, |
773 | struct iommu_iotlb_gather *iotlb_gather) | |
6bd06f5a SS |
774 | { |
775 | long unlocked = 0; | |
776 | struct vfio_regions *entry, *next; | |
777 | ||
aae4c8e2 | 778 | iommu_iotlb_sync(domain->domain, iotlb_gather); |
6bd06f5a SS |
779 | |
780 | list_for_each_entry_safe(entry, next, regions, list) { | |
781 | unlocked += vfio_unpin_pages_remote(dma, | |
782 | entry->iova, | |
783 | entry->phys >> PAGE_SHIFT, | |
784 | entry->len >> PAGE_SHIFT, | |
785 | false); | |
786 | list_del(&entry->list); | |
787 | kfree(entry); | |
788 | } | |
789 | ||
790 | cond_resched(); | |
791 | ||
792 | return unlocked; | |
793 | } | |
794 | ||
795 | /* | |
796 | * Generally, VFIO needs to unpin remote pages after each IOTLB flush. | |
797 | * Therefore, when using IOTLB flush sync interface, VFIO need to keep track | |
798 | * of these regions (currently using a list). | |
799 | * | |
800 | * This value specifies maximum number of regions for each IOTLB flush sync. | |
801 | */ | |
802 | #define VFIO_IOMMU_TLB_SYNC_MAX 512 | |
803 | ||
804 | static size_t unmap_unpin_fast(struct vfio_domain *domain, | |
805 | struct vfio_dma *dma, dma_addr_t *iova, | |
806 | size_t len, phys_addr_t phys, long *unlocked, | |
807 | struct list_head *unmapped_list, | |
a7d20dc1 WD |
808 | int *unmapped_cnt, |
809 | struct iommu_iotlb_gather *iotlb_gather) | |
6bd06f5a SS |
810 | { |
811 | size_t unmapped = 0; | |
812 | struct vfio_regions *entry = kzalloc(sizeof(*entry), GFP_KERNEL); | |
813 | ||
814 | if (entry) { | |
a7d20dc1 WD |
815 | unmapped = iommu_unmap_fast(domain->domain, *iova, len, |
816 | iotlb_gather); | |
6bd06f5a SS |
817 | |
818 | if (!unmapped) { | |
819 | kfree(entry); | |
820 | } else { | |
6bd06f5a SS |
821 | entry->iova = *iova; |
822 | entry->phys = phys; | |
823 | entry->len = unmapped; | |
824 | list_add_tail(&entry->list, unmapped_list); | |
825 | ||
826 | *iova += unmapped; | |
827 | (*unmapped_cnt)++; | |
828 | } | |
829 | } | |
830 | ||
831 | /* | |
832 | * Sync if the number of fast-unmap regions hits the limit | |
833 | * or in case of errors. | |
834 | */ | |
835 | if (*unmapped_cnt >= VFIO_IOMMU_TLB_SYNC_MAX || !unmapped) { | |
a7d20dc1 WD |
836 | *unlocked += vfio_sync_unpin(dma, domain, unmapped_list, |
837 | iotlb_gather); | |
6bd06f5a SS |
838 | *unmapped_cnt = 0; |
839 | } | |
840 | ||
841 | return unmapped; | |
842 | } | |
843 | ||
844 | static size_t unmap_unpin_slow(struct vfio_domain *domain, | |
845 | struct vfio_dma *dma, dma_addr_t *iova, | |
846 | size_t len, phys_addr_t phys, | |
847 | long *unlocked) | |
848 | { | |
849 | size_t unmapped = iommu_unmap(domain->domain, *iova, len); | |
850 | ||
851 | if (unmapped) { | |
852 | *unlocked += vfio_unpin_pages_remote(dma, *iova, | |
853 | phys >> PAGE_SHIFT, | |
854 | unmapped >> PAGE_SHIFT, | |
855 | false); | |
856 | *iova += unmapped; | |
857 | cond_resched(); | |
858 | } | |
859 | return unmapped; | |
860 | } | |
861 | ||
a54eb550 KW |
862 | static long vfio_unmap_unpin(struct vfio_iommu *iommu, struct vfio_dma *dma, |
863 | bool do_accounting) | |
166fd7d9 | 864 | { |
1ef3e2bc AW |
865 | dma_addr_t iova = dma->iova, end = dma->iova + dma->size; |
866 | struct vfio_domain *domain, *d; | |
6bd06f5a | 867 | LIST_HEAD(unmapped_region_list); |
a7d20dc1 | 868 | struct iommu_iotlb_gather iotlb_gather; |
6bd06f5a | 869 | int unmapped_region_cnt = 0; |
166fd7d9 AW |
870 | long unlocked = 0; |
871 | ||
1ef3e2bc | 872 | if (!dma->size) |
a54eb550 KW |
873 | return 0; |
874 | ||
875 | if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu)) | |
876 | return 0; | |
877 | ||
1ef3e2bc AW |
878 | /* |
879 | * We use the IOMMU to track the physical addresses, otherwise we'd | |
880 | * need a much more complicated tracking system. Unfortunately that | |
881 | * means we need to use one of the iommu domains to figure out the | |
882 | * pfns to unpin. The rest need to be unmapped in advance so we have | |
883 | * no iommu translations remaining when the pages are unpinned. | |
884 | */ | |
885 | domain = d = list_first_entry(&iommu->domain_list, | |
886 | struct vfio_domain, next); | |
887 | ||
c5e66887 | 888 | list_for_each_entry_continue(d, &iommu->domain_list, next) { |
1ef3e2bc | 889 | iommu_unmap(d->domain, dma->iova, dma->size); |
c5e66887 AW |
890 | cond_resched(); |
891 | } | |
1ef3e2bc | 892 | |
a7d20dc1 | 893 | iommu_iotlb_gather_init(&iotlb_gather); |
166fd7d9 | 894 | while (iova < end) { |
6fe1010d AW |
895 | size_t unmapped, len; |
896 | phys_addr_t phys, next; | |
166fd7d9 | 897 | |
1ef3e2bc | 898 | phys = iommu_iova_to_phys(domain->domain, iova); |
166fd7d9 AW |
899 | if (WARN_ON(!phys)) { |
900 | iova += PAGE_SIZE; | |
901 | continue; | |
73fa0d10 | 902 | } |
166fd7d9 | 903 | |
6fe1010d AW |
904 | /* |
905 | * To optimize for fewer iommu_unmap() calls, each of which | |
906 | * may require hardware cache flushing, try to find the | |
907 | * largest contiguous physical memory chunk to unmap. | |
908 | */ | |
909 | for (len = PAGE_SIZE; | |
910 | !domain->fgsp && iova + len < end; len += PAGE_SIZE) { | |
911 | next = iommu_iova_to_phys(domain->domain, iova + len); | |
912 | if (next != phys + len) | |
913 | break; | |
914 | } | |
915 | ||
6bd06f5a SS |
916 | /* |
917 | * First, try to use fast unmap/unpin. In case of failure, | |
918 | * switch to slow unmap/unpin path. | |
919 | */ | |
920 | unmapped = unmap_unpin_fast(domain, dma, &iova, len, phys, | |
921 | &unlocked, &unmapped_region_list, | |
a7d20dc1 WD |
922 | &unmapped_region_cnt, |
923 | &iotlb_gather); | |
6bd06f5a SS |
924 | if (!unmapped) { |
925 | unmapped = unmap_unpin_slow(domain, dma, &iova, len, | |
926 | phys, &unlocked); | |
927 | if (WARN_ON(!unmapped)) | |
928 | break; | |
929 | } | |
73fa0d10 | 930 | } |
166fd7d9 | 931 | |
a54eb550 | 932 | dma->iommu_mapped = false; |
6bd06f5a | 933 | |
a7d20dc1 WD |
934 | if (unmapped_region_cnt) { |
935 | unlocked += vfio_sync_unpin(dma, domain, &unmapped_region_list, | |
936 | &iotlb_gather); | |
937 | } | |
6bd06f5a | 938 | |
a54eb550 | 939 | if (do_accounting) { |
48d8476b | 940 | vfio_lock_acct(dma, -unlocked, true); |
a54eb550 KW |
941 | return 0; |
942 | } | |
943 | return unlocked; | |
73fa0d10 AW |
944 | } |
945 | ||
1ef3e2bc | 946 | static void vfio_remove_dma(struct vfio_iommu *iommu, struct vfio_dma *dma) |
73fa0d10 | 947 | { |
a54eb550 | 948 | vfio_unmap_unpin(iommu, dma, true); |
1ef3e2bc | 949 | vfio_unlink_dma(iommu, dma); |
8f0d5bb9 | 950 | put_task_struct(dma->task); |
d6a4c185 | 951 | vfio_dma_bitmap_free(dma); |
1ef3e2bc | 952 | kfree(dma); |
49285593 | 953 | iommu->dma_avail++; |
1ef3e2bc | 954 | } |
73fa0d10 | 955 | |
cade075f | 956 | static void vfio_update_pgsize_bitmap(struct vfio_iommu *iommu) |
1ef3e2bc AW |
957 | { |
958 | struct vfio_domain *domain; | |
166fd7d9 | 959 | |
cade075f KW |
960 | iommu->pgsize_bitmap = ULONG_MAX; |
961 | ||
1ef3e2bc | 962 | list_for_each_entry(domain, &iommu->domain_list, next) |
cade075f | 963 | iommu->pgsize_bitmap &= domain->domain->pgsize_bitmap; |
73fa0d10 | 964 | |
4644321f EA |
965 | /* |
966 | * In case the IOMMU supports page sizes smaller than PAGE_SIZE | |
967 | * we pretend PAGE_SIZE is supported and hide sub-PAGE_SIZE sizes. | |
968 | * That way the user will be able to map/unmap buffers whose size/ | |
969 | * start address is aligned with PAGE_SIZE. Pinning code uses that | |
970 | * granularity while iommu driver can use the sub-PAGE_SIZE size | |
971 | * to map the buffer. | |
972 | */ | |
cade075f KW |
973 | if (iommu->pgsize_bitmap & ~PAGE_MASK) { |
974 | iommu->pgsize_bitmap &= PAGE_MASK; | |
975 | iommu->pgsize_bitmap |= PAGE_SIZE; | |
4644321f | 976 | } |
73fa0d10 AW |
977 | } |
978 | ||
95fc87b4 KW |
979 | static int update_user_bitmap(u64 __user *bitmap, struct vfio_iommu *iommu, |
980 | struct vfio_dma *dma, dma_addr_t base_iova, | |
981 | size_t pgsize) | |
d6a4c185 KW |
982 | { |
983 | unsigned long pgshift = __ffs(pgsize); | |
984 | unsigned long nbits = dma->size >> pgshift; | |
985 | unsigned long bit_offset = (dma->iova - base_iova) >> pgshift; | |
986 | unsigned long copy_offset = bit_offset / BITS_PER_LONG; | |
987 | unsigned long shift = bit_offset % BITS_PER_LONG; | |
988 | unsigned long leftover; | |
989 | ||
95fc87b4 KW |
990 | /* |
991 | * mark all pages dirty if any IOMMU capable device is not able | |
992 | * to report dirty pages and all pages are pinned and mapped. | |
993 | */ | |
994 | if (!iommu->pinned_page_dirty_scope && dma->iommu_mapped) | |
d6a4c185 KW |
995 | bitmap_set(dma->bitmap, 0, nbits); |
996 | ||
997 | if (shift) { | |
998 | bitmap_shift_left(dma->bitmap, dma->bitmap, shift, | |
999 | nbits + shift); | |
1000 | ||
1001 | if (copy_from_user(&leftover, | |
c8e9df47 | 1002 | (void __user *)(bitmap + copy_offset), |
d6a4c185 KW |
1003 | sizeof(leftover))) |
1004 | return -EFAULT; | |
1005 | ||
1006 | bitmap_or(dma->bitmap, dma->bitmap, &leftover, shift); | |
1007 | } | |
1008 | ||
c8e9df47 | 1009 | if (copy_to_user((void __user *)(bitmap + copy_offset), dma->bitmap, |
d6a4c185 KW |
1010 | DIRTY_BITMAP_BYTES(nbits + shift))) |
1011 | return -EFAULT; | |
1012 | ||
1013 | return 0; | |
1014 | } | |
1015 | ||
1016 | static int vfio_iova_dirty_bitmap(u64 __user *bitmap, struct vfio_iommu *iommu, | |
1017 | dma_addr_t iova, size_t size, size_t pgsize) | |
1018 | { | |
1019 | struct vfio_dma *dma; | |
1020 | struct rb_node *n; | |
1021 | unsigned long pgshift = __ffs(pgsize); | |
1022 | int ret; | |
1023 | ||
1024 | /* | |
1025 | * GET_BITMAP request must fully cover vfio_dma mappings. Multiple | |
1026 | * vfio_dma mappings may be clubbed by specifying large ranges, but | |
1027 | * there must not be any previous mappings bisected by the range. | |
1028 | * An error will be returned if these conditions are not met. | |
1029 | */ | |
1030 | dma = vfio_find_dma(iommu, iova, 1); | |
1031 | if (dma && dma->iova != iova) | |
1032 | return -EINVAL; | |
1033 | ||
1034 | dma = vfio_find_dma(iommu, iova + size - 1, 0); | |
1035 | if (dma && dma->iova + dma->size != iova + size) | |
1036 | return -EINVAL; | |
1037 | ||
1038 | for (n = rb_first(&iommu->dma_list); n; n = rb_next(n)) { | |
1039 | struct vfio_dma *dma = rb_entry(n, struct vfio_dma, node); | |
1040 | ||
1041 | if (dma->iova < iova) | |
1042 | continue; | |
1043 | ||
1044 | if (dma->iova > iova + size - 1) | |
1045 | break; | |
1046 | ||
95fc87b4 | 1047 | ret = update_user_bitmap(bitmap, iommu, dma, iova, pgsize); |
d6a4c185 KW |
1048 | if (ret) |
1049 | return ret; | |
1050 | ||
1051 | /* | |
1052 | * Re-populate bitmap to include all pinned pages which are | |
1053 | * considered as dirty but exclude pages which are unpinned and | |
1054 | * pages which are marked dirty by vfio_dma_rw() | |
1055 | */ | |
1056 | bitmap_clear(dma->bitmap, 0, dma->size >> pgshift); | |
1057 | vfio_dma_populate_bitmap(dma, pgsize); | |
1058 | } | |
1059 | return 0; | |
1060 | } | |
1061 | ||
1062 | static int verify_bitmap_size(uint64_t npages, uint64_t bitmap_size) | |
1063 | { | |
1064 | if (!npages || !bitmap_size || (bitmap_size > DIRTY_BITMAP_SIZE_MAX) || | |
1065 | (bitmap_size < DIRTY_BITMAP_BYTES(npages))) | |
1066 | return -EINVAL; | |
1067 | ||
1068 | return 0; | |
1069 | } | |
1070 | ||
73fa0d10 | 1071 | static int vfio_dma_do_unmap(struct vfio_iommu *iommu, |
331e33d2 KW |
1072 | struct vfio_iommu_type1_dma_unmap *unmap, |
1073 | struct vfio_bitmap *bitmap) | |
73fa0d10 | 1074 | { |
c086de81 | 1075 | struct vfio_dma *dma, *dma_last = NULL; |
331e33d2 | 1076 | size_t unmapped = 0, pgsize; |
c086de81 | 1077 | int ret = 0, retries = 0; |
331e33d2 | 1078 | unsigned long pgshift; |
73fa0d10 | 1079 | |
cade075f KW |
1080 | mutex_lock(&iommu->lock); |
1081 | ||
331e33d2 KW |
1082 | pgshift = __ffs(iommu->pgsize_bitmap); |
1083 | pgsize = (size_t)1 << pgshift; | |
cade075f | 1084 | |
331e33d2 | 1085 | if (unmap->iova & (pgsize - 1)) { |
cade075f KW |
1086 | ret = -EINVAL; |
1087 | goto unlock; | |
1088 | } | |
1089 | ||
331e33d2 | 1090 | if (!unmap->size || unmap->size & (pgsize - 1)) { |
cade075f KW |
1091 | ret = -EINVAL; |
1092 | goto unlock; | |
1093 | } | |
73fa0d10 | 1094 | |
58fec830 | 1095 | if (unmap->iova + unmap->size - 1 < unmap->iova || |
cade075f KW |
1096 | unmap->size > SIZE_MAX) { |
1097 | ret = -EINVAL; | |
1098 | goto unlock; | |
1099 | } | |
73fa0d10 | 1100 | |
331e33d2 KW |
1101 | /* When dirty tracking is enabled, allow only min supported pgsize */ |
1102 | if ((unmap->flags & VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP) && | |
1103 | (!iommu->dirty_page_tracking || (bitmap->pgsize != pgsize))) { | |
1104 | ret = -EINVAL; | |
1105 | goto unlock; | |
1106 | } | |
73fa0d10 | 1107 | |
331e33d2 KW |
1108 | WARN_ON((pgsize - 1) & PAGE_MASK); |
1109 | again: | |
1ef3e2bc AW |
1110 | /* |
1111 | * vfio-iommu-type1 (v1) - User mappings were coalesced together to | |
1112 | * avoid tracking individual mappings. This means that the granularity | |
1113 | * of the original mapping was lost and the user was allowed to attempt | |
1114 | * to unmap any range. Depending on the contiguousness of physical | |
1115 | * memory and page sizes supported by the IOMMU, arbitrary unmaps may | |
1116 | * or may not have worked. We only guaranteed unmap granularity | |
1117 | * matching the original mapping; even though it was untracked here, | |
1118 | * the original mappings are reflected in IOMMU mappings. This | |
1119 | * resulted in a couple unusual behaviors. First, if a range is not | |
1120 | * able to be unmapped, ex. a set of 4k pages that was mapped as a | |
1121 | * 2M hugepage into the IOMMU, the unmap ioctl returns success but with | |
1122 | * a zero sized unmap. Also, if an unmap request overlaps the first | |
1123 | * address of a hugepage, the IOMMU will unmap the entire hugepage. | |
1124 | * This also returns success and the returned unmap size reflects the | |
1125 | * actual size unmapped. | |
1126 | * | |
1127 | * We attempt to maintain compatibility with this "v1" interface, but | |
1128 | * we take control out of the hands of the IOMMU. Therefore, an unmap | |
1129 | * request offset from the beginning of the original mapping will | |
1130 | * return success with zero sized unmap. And an unmap request covering | |
1131 | * the first iova of mapping will unmap the entire range. | |
1132 | * | |
1133 | * The v2 version of this interface intends to be more deterministic. | |
1134 | * Unmap requests must fully cover previous mappings. Multiple | |
1135 | * mappings may still be unmaped by specifying large ranges, but there | |
1136 | * must not be any previous mappings bisected by the range. An error | |
1137 | * will be returned if these conditions are not met. The v2 interface | |
1138 | * will only return success and a size of zero if there were no | |
1139 | * mappings within the range. | |
1140 | */ | |
1141 | if (iommu->v2) { | |
7c03f428 | 1142 | dma = vfio_find_dma(iommu, unmap->iova, 1); |
1ef3e2bc AW |
1143 | if (dma && dma->iova != unmap->iova) { |
1144 | ret = -EINVAL; | |
1145 | goto unlock; | |
1146 | } | |
1147 | dma = vfio_find_dma(iommu, unmap->iova + unmap->size - 1, 0); | |
1148 | if (dma && dma->iova + dma->size != unmap->iova + unmap->size) { | |
1149 | ret = -EINVAL; | |
1150 | goto unlock; | |
1151 | } | |
1152 | } | |
1153 | ||
166fd7d9 | 1154 | while ((dma = vfio_find_dma(iommu, unmap->iova, unmap->size))) { |
1ef3e2bc | 1155 | if (!iommu->v2 && unmap->iova > dma->iova) |
166fd7d9 | 1156 | break; |
8f0d5bb9 KW |
1157 | /* |
1158 | * Task with same address space who mapped this iova range is | |
1159 | * allowed to unmap the iova range. | |
1160 | */ | |
1161 | if (dma->task->mm != current->mm) | |
1162 | break; | |
c086de81 KW |
1163 | |
1164 | if (!RB_EMPTY_ROOT(&dma->pfn_list)) { | |
1165 | struct vfio_iommu_type1_dma_unmap nb_unmap; | |
1166 | ||
1167 | if (dma_last == dma) { | |
1168 | BUG_ON(++retries > 10); | |
1169 | } else { | |
1170 | dma_last = dma; | |
1171 | retries = 0; | |
1172 | } | |
1173 | ||
1174 | nb_unmap.iova = dma->iova; | |
1175 | nb_unmap.size = dma->size; | |
1176 | ||
1177 | /* | |
1178 | * Notify anyone (mdev vendor drivers) to invalidate and | |
1179 | * unmap iovas within the range we're about to unmap. | |
1180 | * Vendor drivers MUST unpin pages in response to an | |
1181 | * invalidation. | |
1182 | */ | |
1183 | mutex_unlock(&iommu->lock); | |
1184 | blocking_notifier_call_chain(&iommu->notifier, | |
1185 | VFIO_IOMMU_NOTIFY_DMA_UNMAP, | |
1186 | &nb_unmap); | |
cade075f | 1187 | mutex_lock(&iommu->lock); |
c086de81 KW |
1188 | goto again; |
1189 | } | |
331e33d2 KW |
1190 | |
1191 | if (unmap->flags & VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP) { | |
95fc87b4 | 1192 | ret = update_user_bitmap(bitmap->data, iommu, dma, |
331e33d2 KW |
1193 | unmap->iova, pgsize); |
1194 | if (ret) | |
1195 | break; | |
1196 | } | |
1197 | ||
1ef3e2bc AW |
1198 | unmapped += dma->size; |
1199 | vfio_remove_dma(iommu, dma); | |
166fd7d9 | 1200 | } |
cd9b2268 | 1201 | |
1ef3e2bc | 1202 | unlock: |
73fa0d10 | 1203 | mutex_unlock(&iommu->lock); |
166fd7d9 | 1204 | |
1ef3e2bc | 1205 | /* Report how much was unmapped */ |
166fd7d9 AW |
1206 | unmap->size = unmapped; |
1207 | ||
1208 | return ret; | |
1209 | } | |
1210 | ||
1ef3e2bc AW |
1211 | static int vfio_iommu_map(struct vfio_iommu *iommu, dma_addr_t iova, |
1212 | unsigned long pfn, long npage, int prot) | |
1213 | { | |
1214 | struct vfio_domain *d; | |
1215 | int ret; | |
1216 | ||
1217 | list_for_each_entry(d, &iommu->domain_list, next) { | |
1218 | ret = iommu_map(d->domain, iova, (phys_addr_t)pfn << PAGE_SHIFT, | |
1219 | npage << PAGE_SHIFT, prot | d->prot); | |
7a30423a JR |
1220 | if (ret) |
1221 | goto unwind; | |
c5e66887 AW |
1222 | |
1223 | cond_resched(); | |
1ef3e2bc AW |
1224 | } |
1225 | ||
1226 | return 0; | |
1227 | ||
1228 | unwind: | |
e1907d67 | 1229 | list_for_each_entry_continue_reverse(d, &iommu->domain_list, next) { |
1ef3e2bc | 1230 | iommu_unmap(d->domain, iova, npage << PAGE_SHIFT); |
e1907d67 XZ |
1231 | cond_resched(); |
1232 | } | |
166fd7d9 | 1233 | |
cd9b2268 | 1234 | return ret; |
73fa0d10 AW |
1235 | } |
1236 | ||
8f0d5bb9 KW |
1237 | static int vfio_pin_map_dma(struct vfio_iommu *iommu, struct vfio_dma *dma, |
1238 | size_t map_size) | |
1239 | { | |
1240 | dma_addr_t iova = dma->iova; | |
1241 | unsigned long vaddr = dma->vaddr; | |
1242 | size_t size = map_size; | |
1243 | long npage; | |
7cb671e7 | 1244 | unsigned long pfn, limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT; |
8f0d5bb9 KW |
1245 | int ret = 0; |
1246 | ||
1247 | while (size) { | |
1248 | /* Pin a contiguous chunk of memory */ | |
1249 | npage = vfio_pin_pages_remote(dma, vaddr + dma->size, | |
48d8476b | 1250 | size >> PAGE_SHIFT, &pfn, limit); |
8f0d5bb9 KW |
1251 | if (npage <= 0) { |
1252 | WARN_ON(!npage); | |
1253 | ret = (int)npage; | |
1254 | break; | |
1255 | } | |
1256 | ||
1257 | /* Map it! */ | |
1258 | ret = vfio_iommu_map(iommu, iova + dma->size, pfn, npage, | |
1259 | dma->prot); | |
1260 | if (ret) { | |
a54eb550 KW |
1261 | vfio_unpin_pages_remote(dma, iova + dma->size, pfn, |
1262 | npage, true); | |
8f0d5bb9 KW |
1263 | break; |
1264 | } | |
1265 | ||
1266 | size -= npage << PAGE_SHIFT; | |
1267 | dma->size += npage << PAGE_SHIFT; | |
1268 | } | |
1269 | ||
a54eb550 KW |
1270 | dma->iommu_mapped = true; |
1271 | ||
8f0d5bb9 KW |
1272 | if (ret) |
1273 | vfio_remove_dma(iommu, dma); | |
1274 | ||
1275 | return ret; | |
1276 | } | |
1277 | ||
9b77e5c7 SK |
1278 | /* |
1279 | * Check dma map request is within a valid iova range | |
1280 | */ | |
1281 | static bool vfio_iommu_iova_dma_valid(struct vfio_iommu *iommu, | |
1282 | dma_addr_t start, dma_addr_t end) | |
1283 | { | |
1284 | struct list_head *iova = &iommu->iova_list; | |
1285 | struct vfio_iova *node; | |
1286 | ||
1287 | list_for_each_entry(node, iova, list) { | |
1288 | if (start >= node->start && end <= node->end) | |
1289 | return true; | |
1290 | } | |
1291 | ||
1292 | /* | |
1293 | * Check for list_empty() as well since a container with | |
1294 | * a single mdev device will have an empty list. | |
1295 | */ | |
1296 | return list_empty(iova); | |
1297 | } | |
1298 | ||
73fa0d10 AW |
1299 | static int vfio_dma_do_map(struct vfio_iommu *iommu, |
1300 | struct vfio_iommu_type1_dma_map *map) | |
1301 | { | |
c8dbca16 | 1302 | dma_addr_t iova = map->iova; |
166fd7d9 | 1303 | unsigned long vaddr = map->vaddr; |
73fa0d10 AW |
1304 | size_t size = map->size; |
1305 | int ret = 0, prot = 0; | |
d6a4c185 | 1306 | size_t pgsize; |
1ef3e2bc | 1307 | struct vfio_dma *dma; |
166fd7d9 | 1308 | |
c8dbca16 AW |
1309 | /* Verify that none of our __u64 fields overflow */ |
1310 | if (map->size != size || map->vaddr != vaddr || map->iova != iova) | |
1311 | return -EINVAL; | |
73fa0d10 | 1312 | |
73fa0d10 AW |
1313 | /* READ/WRITE from device perspective */ |
1314 | if (map->flags & VFIO_DMA_MAP_FLAG_WRITE) | |
1315 | prot |= IOMMU_WRITE; | |
1316 | if (map->flags & VFIO_DMA_MAP_FLAG_READ) | |
1317 | prot |= IOMMU_READ; | |
1318 | ||
cade075f | 1319 | mutex_lock(&iommu->lock); |
73fa0d10 | 1320 | |
d6a4c185 | 1321 | pgsize = (size_t)1 << __ffs(iommu->pgsize_bitmap); |
73fa0d10 | 1322 | |
d6a4c185 | 1323 | WARN_ON((pgsize - 1) & PAGE_MASK); |
cade075f | 1324 | |
d6a4c185 | 1325 | if (!prot || !size || (size | iova | vaddr) & (pgsize - 1)) { |
cade075f KW |
1326 | ret = -EINVAL; |
1327 | goto out_unlock; | |
1328 | } | |
1329 | ||
1330 | /* Don't allow IOVA or virtual address wrap */ | |
1331 | if (iova + size - 1 < iova || vaddr + size - 1 < vaddr) { | |
1332 | ret = -EINVAL; | |
1333 | goto out_unlock; | |
1334 | } | |
73fa0d10 | 1335 | |
c8dbca16 | 1336 | if (vfio_find_dma(iommu, iova, size)) { |
8f0d5bb9 KW |
1337 | ret = -EEXIST; |
1338 | goto out_unlock; | |
73fa0d10 AW |
1339 | } |
1340 | ||
49285593 AW |
1341 | if (!iommu->dma_avail) { |
1342 | ret = -ENOSPC; | |
1343 | goto out_unlock; | |
1344 | } | |
1345 | ||
9b77e5c7 SK |
1346 | if (!vfio_iommu_iova_dma_valid(iommu, iova, iova + size - 1)) { |
1347 | ret = -EINVAL; | |
1348 | goto out_unlock; | |
1349 | } | |
1350 | ||
1ef3e2bc AW |
1351 | dma = kzalloc(sizeof(*dma), GFP_KERNEL); |
1352 | if (!dma) { | |
8f0d5bb9 KW |
1353 | ret = -ENOMEM; |
1354 | goto out_unlock; | |
1ef3e2bc AW |
1355 | } |
1356 | ||
49285593 | 1357 | iommu->dma_avail--; |
c8dbca16 AW |
1358 | dma->iova = iova; |
1359 | dma->vaddr = vaddr; | |
1ef3e2bc | 1360 | dma->prot = prot; |
48d8476b AW |
1361 | |
1362 | /* | |
1363 | * We need to be able to both add to a task's locked memory and test | |
1364 | * against the locked memory limit and we need to be able to do both | |
1365 | * outside of this call path as pinning can be asynchronous via the | |
1366 | * external interfaces for mdev devices. RLIMIT_MEMLOCK requires a | |
1367 | * task_struct and VM locked pages requires an mm_struct, however | |
1368 | * holding an indefinite mm reference is not recommended, therefore we | |
1369 | * only hold a reference to a task. We could hold a reference to | |
1370 | * current, however QEMU uses this call path through vCPU threads, | |
1371 | * which can be killed resulting in a NULL mm and failure in the unmap | |
1372 | * path when called via a different thread. Avoid this problem by | |
1373 | * using the group_leader as threads within the same group require | |
1374 | * both CLONE_THREAD and CLONE_VM and will therefore use the same | |
1375 | * mm_struct. | |
1376 | * | |
1377 | * Previously we also used the task for testing CAP_IPC_LOCK at the | |
1378 | * time of pinning and accounting, however has_capability() makes use | |
1379 | * of real_cred, a copy-on-write field, so we can't guarantee that it | |
1380 | * matches group_leader, or in fact that it might not change by the | |
1381 | * time it's evaluated. If a process were to call MAP_DMA with | |
1382 | * CAP_IPC_LOCK but later drop it, it doesn't make sense that they | |
1383 | * possibly see different results for an iommu_mapped vfio_dma vs | |
1384 | * externally mapped. Therefore track CAP_IPC_LOCK in vfio_dma at the | |
1385 | * time of calling MAP_DMA. | |
1386 | */ | |
1387 | get_task_struct(current->group_leader); | |
1388 | dma->task = current->group_leader; | |
1389 | dma->lock_cap = capable(CAP_IPC_LOCK); | |
1390 | ||
a54eb550 | 1391 | dma->pfn_list = RB_ROOT; |
166fd7d9 | 1392 | |
1ef3e2bc AW |
1393 | /* Insert zero-sized and grow as we map chunks of it */ |
1394 | vfio_link_dma(iommu, dma); | |
166fd7d9 | 1395 | |
a54eb550 KW |
1396 | /* Don't pin and map if container doesn't contain IOMMU capable domain*/ |
1397 | if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu)) | |
1398 | dma->size = size; | |
1399 | else | |
1400 | ret = vfio_pin_map_dma(iommu, dma, size); | |
1401 | ||
d6a4c185 KW |
1402 | if (!ret && iommu->dirty_page_tracking) { |
1403 | ret = vfio_dma_bitmap_alloc(dma, pgsize); | |
1404 | if (ret) | |
1405 | vfio_remove_dma(iommu, dma); | |
1406 | } | |
1407 | ||
8f0d5bb9 | 1408 | out_unlock: |
1ef3e2bc AW |
1409 | mutex_unlock(&iommu->lock); |
1410 | return ret; | |
1411 | } | |
1412 | ||
1413 | static int vfio_bus_type(struct device *dev, void *data) | |
1414 | { | |
1415 | struct bus_type **bus = data; | |
1416 | ||
1417 | if (*bus && *bus != dev->bus) | |
1418 | return -EINVAL; | |
1419 | ||
1420 | *bus = dev->bus; | |
1421 | ||
1422 | return 0; | |
1423 | } | |
1424 | ||
1425 | static int vfio_iommu_replay(struct vfio_iommu *iommu, | |
1426 | struct vfio_domain *domain) | |
1427 | { | |
aae7a75a | 1428 | struct vfio_domain *d = NULL; |
1ef3e2bc | 1429 | struct rb_node *n; |
7cb671e7 | 1430 | unsigned long limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT; |
1ef3e2bc AW |
1431 | int ret; |
1432 | ||
1433 | /* Arbitrarily pick the first domain in the list for lookups */ | |
aae7a75a AW |
1434 | if (!list_empty(&iommu->domain_list)) |
1435 | d = list_first_entry(&iommu->domain_list, | |
1436 | struct vfio_domain, next); | |
1437 | ||
1ef3e2bc AW |
1438 | n = rb_first(&iommu->dma_list); |
1439 | ||
1ef3e2bc AW |
1440 | for (; n; n = rb_next(n)) { |
1441 | struct vfio_dma *dma; | |
1442 | dma_addr_t iova; | |
1443 | ||
1444 | dma = rb_entry(n, struct vfio_dma, node); | |
1445 | iova = dma->iova; | |
1446 | ||
1447 | while (iova < dma->iova + dma->size) { | |
a54eb550 | 1448 | phys_addr_t phys; |
1ef3e2bc | 1449 | size_t size; |
73fa0d10 | 1450 | |
a54eb550 KW |
1451 | if (dma->iommu_mapped) { |
1452 | phys_addr_t p; | |
1453 | dma_addr_t i; | |
1454 | ||
aae7a75a AW |
1455 | if (WARN_ON(!d)) { /* mapped w/o a domain?! */ |
1456 | ret = -EINVAL; | |
1457 | goto unwind; | |
1458 | } | |
1459 | ||
a54eb550 KW |
1460 | phys = iommu_iova_to_phys(d->domain, iova); |
1461 | ||
1462 | if (WARN_ON(!phys)) { | |
1463 | iova += PAGE_SIZE; | |
1464 | continue; | |
1465 | } | |
1466 | ||
1467 | size = PAGE_SIZE; | |
1468 | p = phys + size; | |
1469 | i = iova + size; | |
1470 | while (i < dma->iova + dma->size && | |
1471 | p == iommu_iova_to_phys(d->domain, i)) { | |
1472 | size += PAGE_SIZE; | |
1473 | p += PAGE_SIZE; | |
1474 | i += PAGE_SIZE; | |
1475 | } | |
1476 | } else { | |
1477 | unsigned long pfn; | |
1478 | unsigned long vaddr = dma->vaddr + | |
1479 | (iova - dma->iova); | |
1480 | size_t n = dma->iova + dma->size - iova; | |
1481 | long npage; | |
1482 | ||
1483 | npage = vfio_pin_pages_remote(dma, vaddr, | |
1484 | n >> PAGE_SHIFT, | |
48d8476b | 1485 | &pfn, limit); |
a54eb550 KW |
1486 | if (npage <= 0) { |
1487 | WARN_ON(!npage); | |
1488 | ret = (int)npage; | |
aae7a75a | 1489 | goto unwind; |
a54eb550 KW |
1490 | } |
1491 | ||
1492 | phys = pfn << PAGE_SHIFT; | |
1493 | size = npage << PAGE_SHIFT; | |
166fd7d9 AW |
1494 | } |
1495 | ||
1ef3e2bc AW |
1496 | ret = iommu_map(domain->domain, iova, phys, |
1497 | size, dma->prot | domain->prot); | |
aae7a75a AW |
1498 | if (ret) { |
1499 | if (!dma->iommu_mapped) | |
1500 | vfio_unpin_pages_remote(dma, iova, | |
1501 | phys >> PAGE_SHIFT, | |
1502 | size >> PAGE_SHIFT, | |
1503 | true); | |
1504 | goto unwind; | |
1505 | } | |
d93b3ac0 | 1506 | |
1ef3e2bc AW |
1507 | iova += size; |
1508 | } | |
aae7a75a AW |
1509 | } |
1510 | ||
1511 | /* All dmas are now mapped, defer to second tree walk for unwind */ | |
1512 | for (n = rb_first(&iommu->dma_list); n; n = rb_next(n)) { | |
1513 | struct vfio_dma *dma = rb_entry(n, struct vfio_dma, node); | |
1514 | ||
a54eb550 | 1515 | dma->iommu_mapped = true; |
166fd7d9 | 1516 | } |
aae7a75a | 1517 | |
1ef3e2bc | 1518 | return 0; |
aae7a75a AW |
1519 | |
1520 | unwind: | |
1521 | for (; n; n = rb_prev(n)) { | |
1522 | struct vfio_dma *dma = rb_entry(n, struct vfio_dma, node); | |
1523 | dma_addr_t iova; | |
1524 | ||
1525 | if (dma->iommu_mapped) { | |
1526 | iommu_unmap(domain->domain, dma->iova, dma->size); | |
1527 | continue; | |
1528 | } | |
1529 | ||
1530 | iova = dma->iova; | |
1531 | while (iova < dma->iova + dma->size) { | |
1532 | phys_addr_t phys, p; | |
1533 | size_t size; | |
1534 | dma_addr_t i; | |
1535 | ||
1536 | phys = iommu_iova_to_phys(domain->domain, iova); | |
1537 | if (!phys) { | |
1538 | iova += PAGE_SIZE; | |
1539 | continue; | |
1540 | } | |
1541 | ||
1542 | size = PAGE_SIZE; | |
1543 | p = phys + size; | |
1544 | i = iova + size; | |
1545 | while (i < dma->iova + dma->size && | |
1546 | p == iommu_iova_to_phys(domain->domain, i)) { | |
1547 | size += PAGE_SIZE; | |
1548 | p += PAGE_SIZE; | |
1549 | i += PAGE_SIZE; | |
1550 | } | |
1551 | ||
1552 | iommu_unmap(domain->domain, iova, size); | |
1553 | vfio_unpin_pages_remote(dma, iova, phys >> PAGE_SHIFT, | |
1554 | size >> PAGE_SHIFT, true); | |
1555 | } | |
1556 | } | |
1557 | ||
1558 | return ret; | |
73fa0d10 AW |
1559 | } |
1560 | ||
6fe1010d AW |
1561 | /* |
1562 | * We change our unmap behavior slightly depending on whether the IOMMU | |
1563 | * supports fine-grained superpages. IOMMUs like AMD-Vi will use a superpage | |
1564 | * for practically any contiguous power-of-two mapping we give it. This means | |
1565 | * we don't need to look for contiguous chunks ourselves to make unmapping | |
1566 | * more efficient. On IOMMUs with coarse-grained super pages, like Intel VT-d | |
1567 | * with discrete 2M/1G/512G/1T superpages, identifying contiguous chunks | |
1568 | * significantly boosts non-hugetlbfs mappings and doesn't seem to hurt when | |
1569 | * hugetlbfs is in use. | |
1570 | */ | |
1571 | static void vfio_test_domain_fgsp(struct vfio_domain *domain) | |
1572 | { | |
1573 | struct page *pages; | |
1574 | int ret, order = get_order(PAGE_SIZE * 2); | |
1575 | ||
1576 | pages = alloc_pages(GFP_KERNEL | __GFP_ZERO, order); | |
1577 | if (!pages) | |
1578 | return; | |
1579 | ||
1580 | ret = iommu_map(domain->domain, 0, page_to_phys(pages), PAGE_SIZE * 2, | |
1581 | IOMMU_READ | IOMMU_WRITE | domain->prot); | |
1582 | if (!ret) { | |
1583 | size_t unmapped = iommu_unmap(domain->domain, 0, PAGE_SIZE); | |
1584 | ||
1585 | if (unmapped == PAGE_SIZE) | |
1586 | iommu_unmap(domain->domain, PAGE_SIZE, PAGE_SIZE); | |
1587 | else | |
1588 | domain->fgsp = true; | |
1589 | } | |
1590 | ||
1591 | __free_pages(pages, order); | |
1592 | } | |
1593 | ||
7896c998 KW |
1594 | static struct vfio_group *find_iommu_group(struct vfio_domain *domain, |
1595 | struct iommu_group *iommu_group) | |
1596 | { | |
1597 | struct vfio_group *g; | |
1598 | ||
1599 | list_for_each_entry(g, &domain->group_list, next) { | |
1600 | if (g->iommu_group == iommu_group) | |
1601 | return g; | |
1602 | } | |
1603 | ||
1604 | return NULL; | |
1605 | } | |
1606 | ||
95fc87b4 KW |
1607 | static struct vfio_group *vfio_iommu_find_iommu_group(struct vfio_iommu *iommu, |
1608 | struct iommu_group *iommu_group) | |
1609 | { | |
1610 | struct vfio_domain *domain; | |
1611 | struct vfio_group *group = NULL; | |
1612 | ||
1613 | list_for_each_entry(domain, &iommu->domain_list, next) { | |
1614 | group = find_iommu_group(domain, iommu_group); | |
1615 | if (group) | |
1616 | return group; | |
1617 | } | |
1618 | ||
1619 | if (iommu->external_domain) | |
1620 | group = find_iommu_group(iommu->external_domain, iommu_group); | |
1621 | ||
1622 | return group; | |
1623 | } | |
1624 | ||
1625 | static void update_pinned_page_dirty_scope(struct vfio_iommu *iommu) | |
1626 | { | |
1627 | struct vfio_domain *domain; | |
1628 | struct vfio_group *group; | |
1629 | ||
1630 | list_for_each_entry(domain, &iommu->domain_list, next) { | |
1631 | list_for_each_entry(group, &domain->group_list, next) { | |
1632 | if (!group->pinned_page_dirty_scope) { | |
1633 | iommu->pinned_page_dirty_scope = false; | |
1634 | return; | |
1635 | } | |
1636 | } | |
1637 | } | |
1638 | ||
1639 | if (iommu->external_domain) { | |
1640 | domain = iommu->external_domain; | |
1641 | list_for_each_entry(group, &domain->group_list, next) { | |
1642 | if (!group->pinned_page_dirty_scope) { | |
1643 | iommu->pinned_page_dirty_scope = false; | |
1644 | return; | |
1645 | } | |
1646 | } | |
1647 | } | |
1648 | ||
1649 | iommu->pinned_page_dirty_scope = true; | |
1650 | } | |
1651 | ||
b09d6e47 SK |
1652 | static bool vfio_iommu_has_sw_msi(struct list_head *group_resv_regions, |
1653 | phys_addr_t *base) | |
5d704992 | 1654 | { |
b09d6e47 | 1655 | struct iommu_resv_region *region; |
5d704992 EA |
1656 | bool ret = false; |
1657 | ||
b09d6e47 | 1658 | list_for_each_entry(region, group_resv_regions, list) { |
f203f7f1 RM |
1659 | /* |
1660 | * The presence of any 'real' MSI regions should take | |
1661 | * precedence over the software-managed one if the | |
1662 | * IOMMU driver happens to advertise both types. | |
1663 | */ | |
1664 | if (region->type == IOMMU_RESV_MSI) { | |
1665 | ret = false; | |
1666 | break; | |
1667 | } | |
1668 | ||
9d3a4de4 | 1669 | if (region->type == IOMMU_RESV_SW_MSI) { |
5d704992 EA |
1670 | *base = region->start; |
1671 | ret = true; | |
5d704992 EA |
1672 | } |
1673 | } | |
b09d6e47 | 1674 | |
5d704992 EA |
1675 | return ret; |
1676 | } | |
1677 | ||
7bd50f0c LB |
1678 | static struct device *vfio_mdev_get_iommu_device(struct device *dev) |
1679 | { | |
1680 | struct device *(*fn)(struct device *dev); | |
1681 | struct device *iommu_device; | |
1682 | ||
1683 | fn = symbol_get(mdev_get_iommu_device); | |
1684 | if (fn) { | |
1685 | iommu_device = fn(dev); | |
1686 | symbol_put(mdev_get_iommu_device); | |
1687 | ||
1688 | return iommu_device; | |
1689 | } | |
1690 | ||
1691 | return NULL; | |
1692 | } | |
1693 | ||
1694 | static int vfio_mdev_attach_domain(struct device *dev, void *data) | |
1695 | { | |
1696 | struct iommu_domain *domain = data; | |
1697 | struct device *iommu_device; | |
1698 | ||
1699 | iommu_device = vfio_mdev_get_iommu_device(dev); | |
1700 | if (iommu_device) { | |
1701 | if (iommu_dev_feature_enabled(iommu_device, IOMMU_DEV_FEAT_AUX)) | |
1702 | return iommu_aux_attach_device(domain, iommu_device); | |
1703 | else | |
1704 | return iommu_attach_device(domain, iommu_device); | |
1705 | } | |
1706 | ||
1707 | return -EINVAL; | |
1708 | } | |
1709 | ||
1710 | static int vfio_mdev_detach_domain(struct device *dev, void *data) | |
1711 | { | |
1712 | struct iommu_domain *domain = data; | |
1713 | struct device *iommu_device; | |
1714 | ||
1715 | iommu_device = vfio_mdev_get_iommu_device(dev); | |
1716 | if (iommu_device) { | |
1717 | if (iommu_dev_feature_enabled(iommu_device, IOMMU_DEV_FEAT_AUX)) | |
1718 | iommu_aux_detach_device(domain, iommu_device); | |
1719 | else | |
1720 | iommu_detach_device(domain, iommu_device); | |
1721 | } | |
1722 | ||
1723 | return 0; | |
1724 | } | |
1725 | ||
1726 | static int vfio_iommu_attach_group(struct vfio_domain *domain, | |
1727 | struct vfio_group *group) | |
1728 | { | |
1729 | if (group->mdev_group) | |
1730 | return iommu_group_for_each_dev(group->iommu_group, | |
1731 | domain->domain, | |
1732 | vfio_mdev_attach_domain); | |
1733 | else | |
1734 | return iommu_attach_group(domain->domain, group->iommu_group); | |
1735 | } | |
1736 | ||
1737 | static void vfio_iommu_detach_group(struct vfio_domain *domain, | |
1738 | struct vfio_group *group) | |
1739 | { | |
1740 | if (group->mdev_group) | |
1741 | iommu_group_for_each_dev(group->iommu_group, domain->domain, | |
1742 | vfio_mdev_detach_domain); | |
1743 | else | |
1744 | iommu_detach_group(domain->domain, group->iommu_group); | |
1745 | } | |
1746 | ||
be068fa2 LB |
1747 | static bool vfio_bus_is_mdev(struct bus_type *bus) |
1748 | { | |
1749 | struct bus_type *mdev_bus; | |
1750 | bool ret = false; | |
1751 | ||
1752 | mdev_bus = symbol_get(mdev_bus_type); | |
1753 | if (mdev_bus) { | |
1754 | ret = (bus == mdev_bus); | |
1755 | symbol_put(mdev_bus_type); | |
1756 | } | |
1757 | ||
1758 | return ret; | |
1759 | } | |
1760 | ||
1761 | static int vfio_mdev_iommu_device(struct device *dev, void *data) | |
1762 | { | |
1763 | struct device **old = data, *new; | |
1764 | ||
1765 | new = vfio_mdev_get_iommu_device(dev); | |
1766 | if (!new || (*old && *old != new)) | |
1767 | return -EINVAL; | |
1768 | ||
1769 | *old = new; | |
1770 | ||
1771 | return 0; | |
1772 | } | |
1773 | ||
1108696a SK |
1774 | /* |
1775 | * This is a helper function to insert an address range to iova list. | |
1776 | * The list is initially created with a single entry corresponding to | |
1777 | * the IOMMU domain geometry to which the device group is attached. | |
1778 | * The list aperture gets modified when a new domain is added to the | |
1779 | * container if the new aperture doesn't conflict with the current one | |
1780 | * or with any existing dma mappings. The list is also modified to | |
1781 | * exclude any reserved regions associated with the device group. | |
1782 | */ | |
1783 | static int vfio_iommu_iova_insert(struct list_head *head, | |
1784 | dma_addr_t start, dma_addr_t end) | |
1785 | { | |
1786 | struct vfio_iova *region; | |
1787 | ||
1788 | region = kmalloc(sizeof(*region), GFP_KERNEL); | |
1789 | if (!region) | |
1790 | return -ENOMEM; | |
1791 | ||
1792 | INIT_LIST_HEAD(®ion->list); | |
1793 | region->start = start; | |
1794 | region->end = end; | |
1795 | ||
1796 | list_add_tail(®ion->list, head); | |
1797 | return 0; | |
1798 | } | |
1799 | ||
1800 | /* | |
1801 | * Check the new iommu aperture conflicts with existing aper or with any | |
1802 | * existing dma mappings. | |
1803 | */ | |
1804 | static bool vfio_iommu_aper_conflict(struct vfio_iommu *iommu, | |
1805 | dma_addr_t start, dma_addr_t end) | |
1806 | { | |
1807 | struct vfio_iova *first, *last; | |
1808 | struct list_head *iova = &iommu->iova_list; | |
1809 | ||
1810 | if (list_empty(iova)) | |
1811 | return false; | |
1812 | ||
1813 | /* Disjoint sets, return conflict */ | |
1814 | first = list_first_entry(iova, struct vfio_iova, list); | |
1815 | last = list_last_entry(iova, struct vfio_iova, list); | |
1816 | if (start > last->end || end < first->start) | |
1817 | return true; | |
1818 | ||
1819 | /* Check for any existing dma mappings below the new start */ | |
1820 | if (start > first->start) { | |
1821 | if (vfio_find_dma(iommu, first->start, start - first->start)) | |
1822 | return true; | |
1823 | } | |
1824 | ||
1825 | /* Check for any existing dma mappings beyond the new end */ | |
1826 | if (end < last->end) { | |
1827 | if (vfio_find_dma(iommu, end + 1, last->end - end)) | |
1828 | return true; | |
1829 | } | |
1830 | ||
1831 | return false; | |
1832 | } | |
1833 | ||
1834 | /* | |
1835 | * Resize iommu iova aperture window. This is called only if the new | |
1836 | * aperture has no conflict with existing aperture and dma mappings. | |
1837 | */ | |
1838 | static int vfio_iommu_aper_resize(struct list_head *iova, | |
1839 | dma_addr_t start, dma_addr_t end) | |
1840 | { | |
1841 | struct vfio_iova *node, *next; | |
1842 | ||
1843 | if (list_empty(iova)) | |
1844 | return vfio_iommu_iova_insert(iova, start, end); | |
1845 | ||
1846 | /* Adjust iova list start */ | |
1847 | list_for_each_entry_safe(node, next, iova, list) { | |
1848 | if (start < node->start) | |
1849 | break; | |
1850 | if (start >= node->start && start < node->end) { | |
1851 | node->start = start; | |
1852 | break; | |
1853 | } | |
1854 | /* Delete nodes before new start */ | |
1855 | list_del(&node->list); | |
1856 | kfree(node); | |
1857 | } | |
1858 | ||
1859 | /* Adjust iova list end */ | |
1860 | list_for_each_entry_safe(node, next, iova, list) { | |
1861 | if (end > node->end) | |
1862 | continue; | |
1863 | if (end > node->start && end <= node->end) { | |
1864 | node->end = end; | |
1865 | continue; | |
1866 | } | |
1867 | /* Delete nodes after new end */ | |
1868 | list_del(&node->list); | |
1869 | kfree(node); | |
1870 | } | |
1871 | ||
1872 | return 0; | |
1873 | } | |
1874 | ||
af029169 SK |
1875 | /* |
1876 | * Check reserved region conflicts with existing dma mappings | |
1877 | */ | |
1878 | static bool vfio_iommu_resv_conflict(struct vfio_iommu *iommu, | |
1879 | struct list_head *resv_regions) | |
1880 | { | |
1881 | struct iommu_resv_region *region; | |
1882 | ||
1883 | /* Check for conflict with existing dma mappings */ | |
1884 | list_for_each_entry(region, resv_regions, list) { | |
1885 | if (region->type == IOMMU_RESV_DIRECT_RELAXABLE) | |
1886 | continue; | |
1887 | ||
1888 | if (vfio_find_dma(iommu, region->start, region->length)) | |
1889 | return true; | |
1890 | } | |
1891 | ||
1892 | return false; | |
1893 | } | |
1894 | ||
1895 | /* | |
1896 | * Check iova region overlap with reserved regions and | |
1897 | * exclude them from the iommu iova range | |
1898 | */ | |
1899 | static int vfio_iommu_resv_exclude(struct list_head *iova, | |
1900 | struct list_head *resv_regions) | |
1901 | { | |
1902 | struct iommu_resv_region *resv; | |
1903 | struct vfio_iova *n, *next; | |
1904 | ||
1905 | list_for_each_entry(resv, resv_regions, list) { | |
1906 | phys_addr_t start, end; | |
1907 | ||
1908 | if (resv->type == IOMMU_RESV_DIRECT_RELAXABLE) | |
1909 | continue; | |
1910 | ||
1911 | start = resv->start; | |
1912 | end = resv->start + resv->length - 1; | |
1913 | ||
1914 | list_for_each_entry_safe(n, next, iova, list) { | |
1915 | int ret = 0; | |
1916 | ||
1917 | /* No overlap */ | |
1918 | if (start > n->end || end < n->start) | |
1919 | continue; | |
1920 | /* | |
1921 | * Insert a new node if current node overlaps with the | |
1922 | * reserve region to exlude that from valid iova range. | |
1923 | * Note that, new node is inserted before the current | |
1924 | * node and finally the current node is deleted keeping | |
1925 | * the list updated and sorted. | |
1926 | */ | |
1927 | if (start > n->start) | |
1928 | ret = vfio_iommu_iova_insert(&n->list, n->start, | |
1929 | start - 1); | |
1930 | if (!ret && end < n->end) | |
1931 | ret = vfio_iommu_iova_insert(&n->list, end + 1, | |
1932 | n->end); | |
1933 | if (ret) | |
1934 | return ret; | |
1935 | ||
1936 | list_del(&n->list); | |
1937 | kfree(n); | |
1938 | } | |
1939 | } | |
1940 | ||
1941 | if (list_empty(iova)) | |
1942 | return -EINVAL; | |
1943 | ||
1944 | return 0; | |
1945 | } | |
1946 | ||
1947 | static void vfio_iommu_resv_free(struct list_head *resv_regions) | |
1948 | { | |
1949 | struct iommu_resv_region *n, *next; | |
1950 | ||
1951 | list_for_each_entry_safe(n, next, resv_regions, list) { | |
1952 | list_del(&n->list); | |
1953 | kfree(n); | |
1954 | } | |
1955 | } | |
1956 | ||
1108696a SK |
1957 | static void vfio_iommu_iova_free(struct list_head *iova) |
1958 | { | |
1959 | struct vfio_iova *n, *next; | |
1960 | ||
1961 | list_for_each_entry_safe(n, next, iova, list) { | |
1962 | list_del(&n->list); | |
1963 | kfree(n); | |
1964 | } | |
1965 | } | |
1966 | ||
1967 | static int vfio_iommu_iova_get_copy(struct vfio_iommu *iommu, | |
1968 | struct list_head *iova_copy) | |
1969 | { | |
1970 | struct list_head *iova = &iommu->iova_list; | |
1971 | struct vfio_iova *n; | |
1972 | int ret; | |
1973 | ||
1974 | list_for_each_entry(n, iova, list) { | |
1975 | ret = vfio_iommu_iova_insert(iova_copy, n->start, n->end); | |
1976 | if (ret) | |
1977 | goto out_free; | |
1978 | } | |
1979 | ||
1980 | return 0; | |
1981 | ||
1982 | out_free: | |
1983 | vfio_iommu_iova_free(iova_copy); | |
1984 | return ret; | |
1985 | } | |
1986 | ||
1987 | static void vfio_iommu_iova_insert_copy(struct vfio_iommu *iommu, | |
1988 | struct list_head *iova_copy) | |
1989 | { | |
1990 | struct list_head *iova = &iommu->iova_list; | |
1991 | ||
1992 | vfio_iommu_iova_free(iova); | |
1993 | ||
1994 | list_splice_tail(iova_copy, iova); | |
1995 | } | |
572f64c7 | 1996 | |
73fa0d10 AW |
1997 | static int vfio_iommu_type1_attach_group(void *iommu_data, |
1998 | struct iommu_group *iommu_group) | |
1999 | { | |
2000 | struct vfio_iommu *iommu = iommu_data; | |
7896c998 | 2001 | struct vfio_group *group; |
1ef3e2bc | 2002 | struct vfio_domain *domain, *d; |
be068fa2 | 2003 | struct bus_type *bus = NULL; |
73fa0d10 | 2004 | int ret; |
9d72f87b | 2005 | bool resv_msi, msi_remap; |
95f89e09 | 2006 | phys_addr_t resv_msi_base = 0; |
1108696a SK |
2007 | struct iommu_domain_geometry geo; |
2008 | LIST_HEAD(iova_copy); | |
af029169 | 2009 | LIST_HEAD(group_resv_regions); |
73fa0d10 | 2010 | |
73fa0d10 AW |
2011 | mutex_lock(&iommu->lock); |
2012 | ||
572f64c7 ZY |
2013 | /* Check for duplicates */ |
2014 | if (vfio_iommu_find_iommu_group(iommu, iommu_group)) { | |
2015 | mutex_unlock(&iommu->lock); | |
2016 | return -EINVAL; | |
a54eb550 KW |
2017 | } |
2018 | ||
1ef3e2bc AW |
2019 | group = kzalloc(sizeof(*group), GFP_KERNEL); |
2020 | domain = kzalloc(sizeof(*domain), GFP_KERNEL); | |
2021 | if (!group || !domain) { | |
2022 | ret = -ENOMEM; | |
2023 | goto out_free; | |
2024 | } | |
2025 | ||
2026 | group->iommu_group = iommu_group; | |
2027 | ||
2028 | /* Determine bus_type in order to allocate a domain */ | |
2029 | ret = iommu_group_for_each_dev(iommu_group, &bus, vfio_bus_type); | |
2030 | if (ret) | |
2031 | goto out_free; | |
2032 | ||
be068fa2 LB |
2033 | if (vfio_bus_is_mdev(bus)) { |
2034 | struct device *iommu_device = NULL; | |
a54eb550 | 2035 | |
be068fa2 LB |
2036 | group->mdev_group = true; |
2037 | ||
2038 | /* Determine the isolation type */ | |
2039 | ret = iommu_group_for_each_dev(iommu_group, &iommu_device, | |
2040 | vfio_mdev_iommu_device); | |
2041 | if (ret || !iommu_device) { | |
a54eb550 KW |
2042 | if (!iommu->external_domain) { |
2043 | INIT_LIST_HEAD(&domain->group_list); | |
2044 | iommu->external_domain = domain; | |
cade075f | 2045 | vfio_update_pgsize_bitmap(iommu); |
be068fa2 | 2046 | } else { |
a54eb550 | 2047 | kfree(domain); |
be068fa2 | 2048 | } |
a54eb550 KW |
2049 | |
2050 | list_add(&group->next, | |
2051 | &iommu->external_domain->group_list); | |
95fc87b4 KW |
2052 | /* |
2053 | * Non-iommu backed group cannot dirty memory directly, | |
2054 | * it can only use interfaces that provide dirty | |
2055 | * tracking. | |
2056 | * The iommu scope can only be promoted with the | |
2057 | * addition of a dirty tracking group. | |
2058 | */ | |
2059 | group->pinned_page_dirty_scope = true; | |
2060 | if (!iommu->pinned_page_dirty_scope) | |
2061 | update_pinned_page_dirty_scope(iommu); | |
a54eb550 | 2062 | mutex_unlock(&iommu->lock); |
be068fa2 | 2063 | |
a54eb550 KW |
2064 | return 0; |
2065 | } | |
be068fa2 LB |
2066 | |
2067 | bus = iommu_device->bus; | |
a54eb550 KW |
2068 | } |
2069 | ||
1ef3e2bc AW |
2070 | domain->domain = iommu_domain_alloc(bus); |
2071 | if (!domain->domain) { | |
2072 | ret = -EIO; | |
2073 | goto out_free; | |
2074 | } | |
2075 | ||
f5c9eceb WD |
2076 | if (iommu->nesting) { |
2077 | int attr = 1; | |
2078 | ||
2079 | ret = iommu_domain_set_attr(domain->domain, DOMAIN_ATTR_NESTING, | |
2080 | &attr); | |
2081 | if (ret) | |
2082 | goto out_domain; | |
2083 | } | |
2084 | ||
7bd50f0c | 2085 | ret = vfio_iommu_attach_group(domain, group); |
1ef3e2bc AW |
2086 | if (ret) |
2087 | goto out_domain; | |
2088 | ||
1108696a SK |
2089 | /* Get aperture info */ |
2090 | iommu_domain_get_attr(domain->domain, DOMAIN_ATTR_GEOMETRY, &geo); | |
2091 | ||
2092 | if (vfio_iommu_aper_conflict(iommu, geo.aperture_start, | |
2093 | geo.aperture_end)) { | |
2094 | ret = -EINVAL; | |
2095 | goto out_detach; | |
2096 | } | |
2097 | ||
af029169 SK |
2098 | ret = iommu_get_group_resv_regions(iommu_group, &group_resv_regions); |
2099 | if (ret) | |
2100 | goto out_detach; | |
2101 | ||
2102 | if (vfio_iommu_resv_conflict(iommu, &group_resv_regions)) { | |
2103 | ret = -EINVAL; | |
2104 | goto out_detach; | |
2105 | } | |
2106 | ||
1108696a SK |
2107 | /* |
2108 | * We don't want to work on the original iova list as the list | |
2109 | * gets modified and in case of failure we have to retain the | |
2110 | * original list. Get a copy here. | |
2111 | */ | |
2112 | ret = vfio_iommu_iova_get_copy(iommu, &iova_copy); | |
2113 | if (ret) | |
2114 | goto out_detach; | |
2115 | ||
2116 | ret = vfio_iommu_aper_resize(&iova_copy, geo.aperture_start, | |
2117 | geo.aperture_end); | |
2118 | if (ret) | |
2119 | goto out_detach; | |
2120 | ||
af029169 SK |
2121 | ret = vfio_iommu_resv_exclude(&iova_copy, &group_resv_regions); |
2122 | if (ret) | |
2123 | goto out_detach; | |
2124 | ||
b09d6e47 | 2125 | resv_msi = vfio_iommu_has_sw_msi(&group_resv_regions, &resv_msi_base); |
5d704992 | 2126 | |
1ef3e2bc AW |
2127 | INIT_LIST_HEAD(&domain->group_list); |
2128 | list_add(&group->next, &domain->group_list); | |
2129 | ||
db406cc0 RM |
2130 | msi_remap = irq_domain_check_msi_remap() || |
2131 | iommu_capable(bus, IOMMU_CAP_INTR_REMAP); | |
9d72f87b EA |
2132 | |
2133 | if (!allow_unsafe_interrupts && !msi_remap) { | |
1ef3e2bc AW |
2134 | pr_warn("%s: No interrupt remapping support. Use the module param \"allow_unsafe_interrupts\" to enable VFIO IOMMU support on this platform\n", |
2135 | __func__); | |
2136 | ret = -EPERM; | |
2137 | goto out_detach; | |
2138 | } | |
2139 | ||
eb165f05 | 2140 | if (iommu_capable(bus, IOMMU_CAP_CACHE_COHERENCY)) |
1ef3e2bc AW |
2141 | domain->prot |= IOMMU_CACHE; |
2142 | ||
73fa0d10 | 2143 | /* |
1ef3e2bc AW |
2144 | * Try to match an existing compatible domain. We don't want to |
2145 | * preclude an IOMMU driver supporting multiple bus_types and being | |
2146 | * able to include different bus_types in the same IOMMU domain, so | |
2147 | * we test whether the domains use the same iommu_ops rather than | |
2148 | * testing if they're on the same bus_type. | |
73fa0d10 | 2149 | */ |
1ef3e2bc AW |
2150 | list_for_each_entry(d, &iommu->domain_list, next) { |
2151 | if (d->domain->ops == domain->domain->ops && | |
2152 | d->prot == domain->prot) { | |
7bd50f0c LB |
2153 | vfio_iommu_detach_group(domain, group); |
2154 | if (!vfio_iommu_attach_group(d, group)) { | |
1ef3e2bc AW |
2155 | list_add(&group->next, &d->group_list); |
2156 | iommu_domain_free(domain->domain); | |
2157 | kfree(domain); | |
1108696a | 2158 | goto done; |
1ef3e2bc AW |
2159 | } |
2160 | ||
7bd50f0c | 2161 | ret = vfio_iommu_attach_group(domain, group); |
1ef3e2bc AW |
2162 | if (ret) |
2163 | goto out_domain; | |
2164 | } | |
73fa0d10 AW |
2165 | } |
2166 | ||
6fe1010d AW |
2167 | vfio_test_domain_fgsp(domain); |
2168 | ||
1ef3e2bc AW |
2169 | /* replay mappings on new domains */ |
2170 | ret = vfio_iommu_replay(iommu, domain); | |
2171 | if (ret) | |
2172 | goto out_detach; | |
2173 | ||
2c9f1af5 WY |
2174 | if (resv_msi) { |
2175 | ret = iommu_get_msi_cookie(domain->domain, resv_msi_base); | |
f44efca0 | 2176 | if (ret && ret != -ENODEV) |
2c9f1af5 WY |
2177 | goto out_detach; |
2178 | } | |
5d704992 | 2179 | |
1ef3e2bc | 2180 | list_add(&domain->next, &iommu->domain_list); |
cade075f | 2181 | vfio_update_pgsize_bitmap(iommu); |
1108696a SK |
2182 | done: |
2183 | /* Delete the old one and insert new iova list */ | |
2184 | vfio_iommu_iova_insert_copy(iommu, &iova_copy); | |
95fc87b4 KW |
2185 | |
2186 | /* | |
2187 | * An iommu backed group can dirty memory directly and therefore | |
2188 | * demotes the iommu scope until it declares itself dirty tracking | |
2189 | * capable via the page pinning interface. | |
2190 | */ | |
2191 | iommu->pinned_page_dirty_scope = false; | |
73fa0d10 | 2192 | mutex_unlock(&iommu->lock); |
af029169 | 2193 | vfio_iommu_resv_free(&group_resv_regions); |
73fa0d10 AW |
2194 | |
2195 | return 0; | |
1ef3e2bc AW |
2196 | |
2197 | out_detach: | |
7bd50f0c | 2198 | vfio_iommu_detach_group(domain, group); |
1ef3e2bc AW |
2199 | out_domain: |
2200 | iommu_domain_free(domain->domain); | |
1108696a | 2201 | vfio_iommu_iova_free(&iova_copy); |
af029169 | 2202 | vfio_iommu_resv_free(&group_resv_regions); |
1ef3e2bc AW |
2203 | out_free: |
2204 | kfree(domain); | |
2205 | kfree(group); | |
2206 | mutex_unlock(&iommu->lock); | |
2207 | return ret; | |
2208 | } | |
2209 | ||
2210 | static void vfio_iommu_unmap_unpin_all(struct vfio_iommu *iommu) | |
2211 | { | |
2212 | struct rb_node *node; | |
2213 | ||
2214 | while ((node = rb_first(&iommu->dma_list))) | |
2215 | vfio_remove_dma(iommu, rb_entry(node, struct vfio_dma, node)); | |
73fa0d10 AW |
2216 | } |
2217 | ||
a54eb550 KW |
2218 | static void vfio_iommu_unmap_unpin_reaccount(struct vfio_iommu *iommu) |
2219 | { | |
2220 | struct rb_node *n, *p; | |
2221 | ||
2222 | n = rb_first(&iommu->dma_list); | |
2223 | for (; n; n = rb_next(n)) { | |
2224 | struct vfio_dma *dma; | |
2225 | long locked = 0, unlocked = 0; | |
2226 | ||
2227 | dma = rb_entry(n, struct vfio_dma, node); | |
2228 | unlocked += vfio_unmap_unpin(iommu, dma, false); | |
2229 | p = rb_first(&dma->pfn_list); | |
2230 | for (; p; p = rb_next(p)) { | |
2231 | struct vfio_pfn *vpfn = rb_entry(p, struct vfio_pfn, | |
2232 | node); | |
2233 | ||
2234 | if (!is_invalid_reserved_pfn(vpfn->pfn)) | |
2235 | locked++; | |
2236 | } | |
48d8476b | 2237 | vfio_lock_acct(dma, locked - unlocked, true); |
a54eb550 KW |
2238 | } |
2239 | } | |
2240 | ||
2241 | static void vfio_sanity_check_pfn_list(struct vfio_iommu *iommu) | |
2242 | { | |
2243 | struct rb_node *n; | |
2244 | ||
2245 | n = rb_first(&iommu->dma_list); | |
2246 | for (; n; n = rb_next(n)) { | |
2247 | struct vfio_dma *dma; | |
2248 | ||
2249 | dma = rb_entry(n, struct vfio_dma, node); | |
2250 | ||
2251 | if (WARN_ON(!RB_EMPTY_ROOT(&dma->pfn_list))) | |
2252 | break; | |
2253 | } | |
3cedd7d7 KW |
2254 | /* mdev vendor driver must unregister notifier */ |
2255 | WARN_ON(iommu->notifier.head); | |
a54eb550 KW |
2256 | } |
2257 | ||
f45daadf SK |
2258 | /* |
2259 | * Called when a domain is removed in detach. It is possible that | |
2260 | * the removed domain decided the iova aperture window. Modify the | |
2261 | * iova aperture with the smallest window among existing domains. | |
2262 | */ | |
2263 | static void vfio_iommu_aper_expand(struct vfio_iommu *iommu, | |
2264 | struct list_head *iova_copy) | |
2265 | { | |
2266 | struct vfio_domain *domain; | |
2267 | struct iommu_domain_geometry geo; | |
2268 | struct vfio_iova *node; | |
2269 | dma_addr_t start = 0; | |
2270 | dma_addr_t end = (dma_addr_t)~0; | |
2271 | ||
2272 | if (list_empty(iova_copy)) | |
2273 | return; | |
2274 | ||
2275 | list_for_each_entry(domain, &iommu->domain_list, next) { | |
2276 | iommu_domain_get_attr(domain->domain, DOMAIN_ATTR_GEOMETRY, | |
2277 | &geo); | |
2278 | if (geo.aperture_start > start) | |
2279 | start = geo.aperture_start; | |
2280 | if (geo.aperture_end < end) | |
2281 | end = geo.aperture_end; | |
2282 | } | |
2283 | ||
2284 | /* Modify aperture limits. The new aper is either same or bigger */ | |
2285 | node = list_first_entry(iova_copy, struct vfio_iova, list); | |
2286 | node->start = start; | |
2287 | node = list_last_entry(iova_copy, struct vfio_iova, list); | |
2288 | node->end = end; | |
2289 | } | |
2290 | ||
2291 | /* | |
2292 | * Called when a group is detached. The reserved regions for that | |
2293 | * group can be part of valid iova now. But since reserved regions | |
2294 | * may be duplicated among groups, populate the iova valid regions | |
2295 | * list again. | |
2296 | */ | |
2297 | static int vfio_iommu_resv_refresh(struct vfio_iommu *iommu, | |
2298 | struct list_head *iova_copy) | |
2299 | { | |
2300 | struct vfio_domain *d; | |
2301 | struct vfio_group *g; | |
2302 | struct vfio_iova *node; | |
2303 | dma_addr_t start, end; | |
2304 | LIST_HEAD(resv_regions); | |
2305 | int ret; | |
2306 | ||
2307 | if (list_empty(iova_copy)) | |
2308 | return -EINVAL; | |
2309 | ||
2310 | list_for_each_entry(d, &iommu->domain_list, next) { | |
2311 | list_for_each_entry(g, &d->group_list, next) { | |
2312 | ret = iommu_get_group_resv_regions(g->iommu_group, | |
2313 | &resv_regions); | |
2314 | if (ret) | |
2315 | goto done; | |
2316 | } | |
2317 | } | |
2318 | ||
2319 | node = list_first_entry(iova_copy, struct vfio_iova, list); | |
2320 | start = node->start; | |
2321 | node = list_last_entry(iova_copy, struct vfio_iova, list); | |
2322 | end = node->end; | |
2323 | ||
2324 | /* purge the iova list and create new one */ | |
2325 | vfio_iommu_iova_free(iova_copy); | |
2326 | ||
2327 | ret = vfio_iommu_aper_resize(iova_copy, start, end); | |
2328 | if (ret) | |
2329 | goto done; | |
2330 | ||
2331 | /* Exclude current reserved regions from iova ranges */ | |
2332 | ret = vfio_iommu_resv_exclude(iova_copy, &resv_regions); | |
2333 | done: | |
2334 | vfio_iommu_resv_free(&resv_regions); | |
2335 | return ret; | |
2336 | } | |
2337 | ||
73fa0d10 AW |
2338 | static void vfio_iommu_type1_detach_group(void *iommu_data, |
2339 | struct iommu_group *iommu_group) | |
2340 | { | |
2341 | struct vfio_iommu *iommu = iommu_data; | |
1ef3e2bc | 2342 | struct vfio_domain *domain; |
73fa0d10 | 2343 | struct vfio_group *group; |
95fc87b4 | 2344 | bool update_dirty_scope = false; |
f45daadf | 2345 | LIST_HEAD(iova_copy); |
73fa0d10 AW |
2346 | |
2347 | mutex_lock(&iommu->lock); | |
2348 | ||
a54eb550 KW |
2349 | if (iommu->external_domain) { |
2350 | group = find_iommu_group(iommu->external_domain, iommu_group); | |
2351 | if (group) { | |
95fc87b4 | 2352 | update_dirty_scope = !group->pinned_page_dirty_scope; |
a54eb550 KW |
2353 | list_del(&group->next); |
2354 | kfree(group); | |
2355 | ||
2356 | if (list_empty(&iommu->external_domain->group_list)) { | |
2357 | vfio_sanity_check_pfn_list(iommu); | |
2358 | ||
2359 | if (!IS_IOMMU_CAP_DOMAIN_IN_CONTAINER(iommu)) | |
2360 | vfio_iommu_unmap_unpin_all(iommu); | |
2361 | ||
2362 | kfree(iommu->external_domain); | |
2363 | iommu->external_domain = NULL; | |
2364 | } | |
2365 | goto detach_group_done; | |
2366 | } | |
2367 | } | |
2368 | ||
f45daadf SK |
2369 | /* |
2370 | * Get a copy of iova list. This will be used to update | |
2371 | * and to replace the current one later. Please note that | |
2372 | * we will leave the original list as it is if update fails. | |
2373 | */ | |
2374 | vfio_iommu_iova_get_copy(iommu, &iova_copy); | |
2375 | ||
1ef3e2bc | 2376 | list_for_each_entry(domain, &iommu->domain_list, next) { |
7896c998 KW |
2377 | group = find_iommu_group(domain, iommu_group); |
2378 | if (!group) | |
2379 | continue; | |
1ef3e2bc | 2380 | |
7bd50f0c | 2381 | vfio_iommu_detach_group(domain, group); |
95fc87b4 | 2382 | update_dirty_scope = !group->pinned_page_dirty_scope; |
7896c998 KW |
2383 | list_del(&group->next); |
2384 | kfree(group); | |
2385 | /* | |
a54eb550 KW |
2386 | * Group ownership provides privilege, if the group list is |
2387 | * empty, the domain goes away. If it's the last domain with | |
2388 | * iommu and external domain doesn't exist, then all the | |
2389 | * mappings go away too. If it's the last domain with iommu and | |
2390 | * external domain exist, update accounting | |
7896c998 KW |
2391 | */ |
2392 | if (list_empty(&domain->group_list)) { | |
a54eb550 KW |
2393 | if (list_is_singular(&iommu->domain_list)) { |
2394 | if (!iommu->external_domain) | |
2395 | vfio_iommu_unmap_unpin_all(iommu); | |
2396 | else | |
2397 | vfio_iommu_unmap_unpin_reaccount(iommu); | |
2398 | } | |
7896c998 KW |
2399 | iommu_domain_free(domain->domain); |
2400 | list_del(&domain->next); | |
2401 | kfree(domain); | |
f45daadf | 2402 | vfio_iommu_aper_expand(iommu, &iova_copy); |
cade075f | 2403 | vfio_update_pgsize_bitmap(iommu); |
73fa0d10 | 2404 | } |
a54eb550 | 2405 | break; |
73fa0d10 AW |
2406 | } |
2407 | ||
f45daadf SK |
2408 | if (!vfio_iommu_resv_refresh(iommu, &iova_copy)) |
2409 | vfio_iommu_iova_insert_copy(iommu, &iova_copy); | |
2410 | else | |
2411 | vfio_iommu_iova_free(&iova_copy); | |
2412 | ||
a54eb550 | 2413 | detach_group_done: |
95fc87b4 KW |
2414 | /* |
2415 | * Removal of a group without dirty tracking may allow the iommu scope | |
2416 | * to be promoted. | |
2417 | */ | |
2418 | if (update_dirty_scope) | |
2419 | update_pinned_page_dirty_scope(iommu); | |
73fa0d10 AW |
2420 | mutex_unlock(&iommu->lock); |
2421 | } | |
2422 | ||
2423 | static void *vfio_iommu_type1_open(unsigned long arg) | |
2424 | { | |
2425 | struct vfio_iommu *iommu; | |
2426 | ||
73fa0d10 AW |
2427 | iommu = kzalloc(sizeof(*iommu), GFP_KERNEL); |
2428 | if (!iommu) | |
2429 | return ERR_PTR(-ENOMEM); | |
2430 | ||
f5c9eceb WD |
2431 | switch (arg) { |
2432 | case VFIO_TYPE1_IOMMU: | |
2433 | break; | |
2434 | case VFIO_TYPE1_NESTING_IOMMU: | |
2435 | iommu->nesting = true; | |
df561f66 | 2436 | fallthrough; |
f5c9eceb WD |
2437 | case VFIO_TYPE1v2_IOMMU: |
2438 | iommu->v2 = true; | |
2439 | break; | |
2440 | default: | |
2441 | kfree(iommu); | |
2442 | return ERR_PTR(-EINVAL); | |
2443 | } | |
2444 | ||
1ef3e2bc | 2445 | INIT_LIST_HEAD(&iommu->domain_list); |
1108696a | 2446 | INIT_LIST_HEAD(&iommu->iova_list); |
cd9b2268 | 2447 | iommu->dma_list = RB_ROOT; |
49285593 | 2448 | iommu->dma_avail = dma_entry_limit; |
73fa0d10 | 2449 | mutex_init(&iommu->lock); |
c086de81 | 2450 | BLOCKING_INIT_NOTIFIER_HEAD(&iommu->notifier); |
73fa0d10 AW |
2451 | |
2452 | return iommu; | |
2453 | } | |
2454 | ||
a54eb550 KW |
2455 | static void vfio_release_domain(struct vfio_domain *domain, bool external) |
2456 | { | |
2457 | struct vfio_group *group, *group_tmp; | |
2458 | ||
2459 | list_for_each_entry_safe(group, group_tmp, | |
2460 | &domain->group_list, next) { | |
2461 | if (!external) | |
7bd50f0c | 2462 | vfio_iommu_detach_group(domain, group); |
a54eb550 KW |
2463 | list_del(&group->next); |
2464 | kfree(group); | |
2465 | } | |
2466 | ||
2467 | if (!external) | |
2468 | iommu_domain_free(domain->domain); | |
2469 | } | |
2470 | ||
73fa0d10 AW |
2471 | static void vfio_iommu_type1_release(void *iommu_data) |
2472 | { | |
2473 | struct vfio_iommu *iommu = iommu_data; | |
1ef3e2bc | 2474 | struct vfio_domain *domain, *domain_tmp; |
a54eb550 KW |
2475 | |
2476 | if (iommu->external_domain) { | |
2477 | vfio_release_domain(iommu->external_domain, true); | |
2478 | vfio_sanity_check_pfn_list(iommu); | |
2479 | kfree(iommu->external_domain); | |
2480 | } | |
73fa0d10 | 2481 | |
1ef3e2bc | 2482 | vfio_iommu_unmap_unpin_all(iommu); |
73fa0d10 | 2483 | |
1ef3e2bc AW |
2484 | list_for_each_entry_safe(domain, domain_tmp, |
2485 | &iommu->domain_list, next) { | |
a54eb550 | 2486 | vfio_release_domain(domain, false); |
1ef3e2bc AW |
2487 | list_del(&domain->next); |
2488 | kfree(domain); | |
73fa0d10 | 2489 | } |
1108696a SK |
2490 | |
2491 | vfio_iommu_iova_free(&iommu->iova_list); | |
2492 | ||
73fa0d10 AW |
2493 | kfree(iommu); |
2494 | } | |
2495 | ||
aa429318 AW |
2496 | static int vfio_domains_have_iommu_cache(struct vfio_iommu *iommu) |
2497 | { | |
2498 | struct vfio_domain *domain; | |
2499 | int ret = 1; | |
2500 | ||
2501 | mutex_lock(&iommu->lock); | |
2502 | list_for_each_entry(domain, &iommu->domain_list, next) { | |
2503 | if (!(domain->prot & IOMMU_CACHE)) { | |
2504 | ret = 0; | |
f5bfdbf2 | 2505 | break; |
aa429318 | 2506 | } |
73fa0d10 | 2507 | } |
aa429318 | 2508 | mutex_unlock(&iommu->lock); |
73fa0d10 | 2509 | |
aa429318 | 2510 | return ret; |
73fa0d10 AW |
2511 | } |
2512 | ||
ccd59dce LY |
2513 | static int vfio_iommu_type1_check_extension(struct vfio_iommu *iommu, |
2514 | unsigned long arg) | |
2515 | { | |
2516 | switch (arg) { | |
2517 | case VFIO_TYPE1_IOMMU: | |
2518 | case VFIO_TYPE1v2_IOMMU: | |
2519 | case VFIO_TYPE1_NESTING_IOMMU: | |
2520 | return 1; | |
2521 | case VFIO_DMA_CC_IOMMU: | |
2522 | if (!iommu) | |
2523 | return 0; | |
2524 | return vfio_domains_have_iommu_cache(iommu); | |
2525 | default: | |
2526 | return 0; | |
2527 | } | |
2528 | } | |
2529 | ||
a7170720 SK |
2530 | static int vfio_iommu_iova_add_cap(struct vfio_info_cap *caps, |
2531 | struct vfio_iommu_type1_info_cap_iova_range *cap_iovas, | |
2532 | size_t size) | |
2533 | { | |
2534 | struct vfio_info_cap_header *header; | |
2535 | struct vfio_iommu_type1_info_cap_iova_range *iova_cap; | |
2536 | ||
2537 | header = vfio_info_cap_add(caps, size, | |
2538 | VFIO_IOMMU_TYPE1_INFO_CAP_IOVA_RANGE, 1); | |
2539 | if (IS_ERR(header)) | |
2540 | return PTR_ERR(header); | |
2541 | ||
2542 | iova_cap = container_of(header, | |
2543 | struct vfio_iommu_type1_info_cap_iova_range, | |
2544 | header); | |
2545 | iova_cap->nr_iovas = cap_iovas->nr_iovas; | |
2546 | memcpy(iova_cap->iova_ranges, cap_iovas->iova_ranges, | |
2547 | cap_iovas->nr_iovas * sizeof(*cap_iovas->iova_ranges)); | |
2548 | return 0; | |
2549 | } | |
2550 | ||
2551 | static int vfio_iommu_iova_build_caps(struct vfio_iommu *iommu, | |
2552 | struct vfio_info_cap *caps) | |
2553 | { | |
2554 | struct vfio_iommu_type1_info_cap_iova_range *cap_iovas; | |
2555 | struct vfio_iova *iova; | |
2556 | size_t size; | |
2557 | int iovas = 0, i = 0, ret; | |
2558 | ||
a7170720 SK |
2559 | list_for_each_entry(iova, &iommu->iova_list, list) |
2560 | iovas++; | |
2561 | ||
2562 | if (!iovas) { | |
2563 | /* | |
2564 | * Return 0 as a container with a single mdev device | |
2565 | * will have an empty list | |
2566 | */ | |
cade075f | 2567 | return 0; |
a7170720 SK |
2568 | } |
2569 | ||
2570 | size = sizeof(*cap_iovas) + (iovas * sizeof(*cap_iovas->iova_ranges)); | |
2571 | ||
2572 | cap_iovas = kzalloc(size, GFP_KERNEL); | |
cade075f KW |
2573 | if (!cap_iovas) |
2574 | return -ENOMEM; | |
a7170720 SK |
2575 | |
2576 | cap_iovas->nr_iovas = iovas; | |
2577 | ||
2578 | list_for_each_entry(iova, &iommu->iova_list, list) { | |
2579 | cap_iovas->iova_ranges[i].start = iova->start; | |
2580 | cap_iovas->iova_ranges[i].end = iova->end; | |
2581 | i++; | |
2582 | } | |
2583 | ||
2584 | ret = vfio_iommu_iova_add_cap(caps, cap_iovas, size); | |
2585 | ||
2586 | kfree(cap_iovas); | |
a7170720 SK |
2587 | return ret; |
2588 | } | |
2589 | ||
ad721705 KW |
2590 | static int vfio_iommu_migration_build_caps(struct vfio_iommu *iommu, |
2591 | struct vfio_info_cap *caps) | |
2592 | { | |
2593 | struct vfio_iommu_type1_info_cap_migration cap_mig; | |
2594 | ||
2595 | cap_mig.header.id = VFIO_IOMMU_TYPE1_INFO_CAP_MIGRATION; | |
2596 | cap_mig.header.version = 1; | |
2597 | ||
2598 | cap_mig.flags = 0; | |
2599 | /* support minimum pgsize */ | |
2600 | cap_mig.pgsize_bitmap = (size_t)1 << __ffs(iommu->pgsize_bitmap); | |
2601 | cap_mig.max_dirty_bitmap_size = DIRTY_BITMAP_SIZE_MAX; | |
2602 | ||
2603 | return vfio_info_add_capability(caps, &cap_mig.header, sizeof(cap_mig)); | |
2604 | } | |
2605 | ||
7d6e1329 MR |
2606 | static int vfio_iommu_dma_avail_build_caps(struct vfio_iommu *iommu, |
2607 | struct vfio_info_cap *caps) | |
2608 | { | |
2609 | struct vfio_iommu_type1_info_dma_avail cap_dma_avail; | |
2610 | ||
2611 | cap_dma_avail.header.id = VFIO_IOMMU_TYPE1_INFO_DMA_AVAIL; | |
2612 | cap_dma_avail.header.version = 1; | |
2613 | ||
2614 | cap_dma_avail.avail = iommu->dma_avail; | |
2615 | ||
2616 | return vfio_info_add_capability(caps, &cap_dma_avail.header, | |
2617 | sizeof(cap_dma_avail)); | |
2618 | } | |
2619 | ||
ccd59dce LY |
2620 | static int vfio_iommu_type1_get_info(struct vfio_iommu *iommu, |
2621 | unsigned long arg) | |
73fa0d10 | 2622 | { |
ccd59dce | 2623 | struct vfio_iommu_type1_info info; |
73fa0d10 | 2624 | unsigned long minsz; |
ccd59dce LY |
2625 | struct vfio_info_cap caps = { .buf = NULL, .size = 0 }; |
2626 | unsigned long capsz; | |
2627 | int ret; | |
73fa0d10 | 2628 | |
ccd59dce | 2629 | minsz = offsetofend(struct vfio_iommu_type1_info, iova_pgsizes); |
73fa0d10 | 2630 | |
ccd59dce LY |
2631 | /* For backward compatibility, cannot require this */ |
2632 | capsz = offsetofend(struct vfio_iommu_type1_info, cap_offset); | |
a7170720 | 2633 | |
ccd59dce LY |
2634 | if (copy_from_user(&info, (void __user *)arg, minsz)) |
2635 | return -EFAULT; | |
73fa0d10 | 2636 | |
ccd59dce LY |
2637 | if (info.argsz < minsz) |
2638 | return -EINVAL; | |
73fa0d10 | 2639 | |
ccd59dce LY |
2640 | if (info.argsz >= capsz) { |
2641 | minsz = capsz; | |
2642 | info.cap_offset = 0; /* output, no-recopy necessary */ | |
2643 | } | |
a7170720 | 2644 | |
ccd59dce LY |
2645 | mutex_lock(&iommu->lock); |
2646 | info.flags = VFIO_IOMMU_INFO_PGSIZES; | |
73fa0d10 | 2647 | |
ccd59dce | 2648 | info.iova_pgsizes = iommu->pgsize_bitmap; |
73fa0d10 | 2649 | |
ccd59dce | 2650 | ret = vfio_iommu_migration_build_caps(iommu, &caps); |
ad721705 | 2651 | |
7d6e1329 MR |
2652 | if (!ret) |
2653 | ret = vfio_iommu_dma_avail_build_caps(iommu, &caps); | |
2654 | ||
ccd59dce LY |
2655 | if (!ret) |
2656 | ret = vfio_iommu_iova_build_caps(iommu, &caps); | |
ad721705 | 2657 | |
ccd59dce | 2658 | mutex_unlock(&iommu->lock); |
ad721705 | 2659 | |
ccd59dce LY |
2660 | if (ret) |
2661 | return ret; | |
a7170720 | 2662 | |
ccd59dce LY |
2663 | if (caps.size) { |
2664 | info.flags |= VFIO_IOMMU_INFO_CAPS; | |
a7170720 | 2665 | |
ccd59dce LY |
2666 | if (info.argsz < sizeof(info) + caps.size) { |
2667 | info.argsz = sizeof(info) + caps.size; | |
2668 | } else { | |
2669 | vfio_info_cap_shift(&caps, sizeof(info)); | |
2670 | if (copy_to_user((void __user *)arg + | |
2671 | sizeof(info), caps.buf, | |
2672 | caps.size)) { | |
2673 | kfree(caps.buf); | |
2674 | return -EFAULT; | |
a7170720 | 2675 | } |
ccd59dce | 2676 | info.cap_offset = sizeof(info); |
a7170720 SK |
2677 | } |
2678 | ||
ccd59dce LY |
2679 | kfree(caps.buf); |
2680 | } | |
73fa0d10 | 2681 | |
ccd59dce LY |
2682 | return copy_to_user((void __user *)arg, &info, minsz) ? |
2683 | -EFAULT : 0; | |
2684 | } | |
73fa0d10 | 2685 | |
ccd59dce LY |
2686 | static int vfio_iommu_type1_map_dma(struct vfio_iommu *iommu, |
2687 | unsigned long arg) | |
2688 | { | |
2689 | struct vfio_iommu_type1_dma_map map; | |
2690 | unsigned long minsz; | |
2691 | uint32_t mask = VFIO_DMA_MAP_FLAG_READ | VFIO_DMA_MAP_FLAG_WRITE; | |
73fa0d10 | 2692 | |
ccd59dce | 2693 | minsz = offsetofend(struct vfio_iommu_type1_dma_map, size); |
73fa0d10 | 2694 | |
ccd59dce LY |
2695 | if (copy_from_user(&map, (void __user *)arg, minsz)) |
2696 | return -EFAULT; | |
73fa0d10 | 2697 | |
ccd59dce LY |
2698 | if (map.argsz < minsz || map.flags & ~mask) |
2699 | return -EINVAL; | |
73fa0d10 | 2700 | |
ccd59dce LY |
2701 | return vfio_dma_do_map(iommu, &map); |
2702 | } | |
73fa0d10 | 2703 | |
ccd59dce LY |
2704 | static int vfio_iommu_type1_unmap_dma(struct vfio_iommu *iommu, |
2705 | unsigned long arg) | |
2706 | { | |
2707 | struct vfio_iommu_type1_dma_unmap unmap; | |
2708 | struct vfio_bitmap bitmap = { 0 }; | |
2709 | unsigned long minsz; | |
2710 | int ret; | |
73fa0d10 | 2711 | |
ccd59dce | 2712 | minsz = offsetofend(struct vfio_iommu_type1_dma_unmap, size); |
73fa0d10 | 2713 | |
ccd59dce LY |
2714 | if (copy_from_user(&unmap, (void __user *)arg, minsz)) |
2715 | return -EFAULT; | |
73fa0d10 | 2716 | |
ccd59dce LY |
2717 | if (unmap.argsz < minsz || |
2718 | unmap.flags & ~VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP) | |
2719 | return -EINVAL; | |
331e33d2 | 2720 | |
ccd59dce LY |
2721 | if (unmap.flags & VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP) { |
2722 | unsigned long pgshift; | |
331e33d2 | 2723 | |
ccd59dce LY |
2724 | if (unmap.argsz < (minsz + sizeof(bitmap))) |
2725 | return -EINVAL; | |
331e33d2 | 2726 | |
ccd59dce LY |
2727 | if (copy_from_user(&bitmap, |
2728 | (void __user *)(arg + minsz), | |
2729 | sizeof(bitmap))) | |
2730 | return -EFAULT; | |
331e33d2 | 2731 | |
ccd59dce LY |
2732 | if (!access_ok((void __user *)bitmap.data, bitmap.size)) |
2733 | return -EINVAL; | |
331e33d2 | 2734 | |
ccd59dce LY |
2735 | pgshift = __ffs(bitmap.pgsize); |
2736 | ret = verify_bitmap_size(unmap.size >> pgshift, | |
2737 | bitmap.size); | |
166fd7d9 AW |
2738 | if (ret) |
2739 | return ret; | |
ccd59dce LY |
2740 | } |
2741 | ||
2742 | ret = vfio_dma_do_unmap(iommu, &unmap, &bitmap); | |
2743 | if (ret) | |
2744 | return ret; | |
166fd7d9 | 2745 | |
ccd59dce | 2746 | return copy_to_user((void __user *)arg, &unmap, minsz) ? |
8160c4e4 | 2747 | -EFAULT : 0; |
ccd59dce | 2748 | } |
d6a4c185 | 2749 | |
ccd59dce LY |
2750 | static int vfio_iommu_type1_dirty_pages(struct vfio_iommu *iommu, |
2751 | unsigned long arg) | |
2752 | { | |
2753 | struct vfio_iommu_type1_dirty_bitmap dirty; | |
2754 | uint32_t mask = VFIO_IOMMU_DIRTY_PAGES_FLAG_START | | |
2755 | VFIO_IOMMU_DIRTY_PAGES_FLAG_STOP | | |
2756 | VFIO_IOMMU_DIRTY_PAGES_FLAG_GET_BITMAP; | |
2757 | unsigned long minsz; | |
2758 | int ret = 0; | |
d6a4c185 | 2759 | |
ccd59dce LY |
2760 | if (!iommu->v2) |
2761 | return -EACCES; | |
d6a4c185 | 2762 | |
ccd59dce | 2763 | minsz = offsetofend(struct vfio_iommu_type1_dirty_bitmap, flags); |
d6a4c185 | 2764 | |
ccd59dce LY |
2765 | if (copy_from_user(&dirty, (void __user *)arg, minsz)) |
2766 | return -EFAULT; | |
d6a4c185 | 2767 | |
ccd59dce LY |
2768 | if (dirty.argsz < minsz || dirty.flags & ~mask) |
2769 | return -EINVAL; | |
d6a4c185 | 2770 | |
ccd59dce LY |
2771 | /* only one flag should be set at a time */ |
2772 | if (__ffs(dirty.flags) != __fls(dirty.flags)) | |
2773 | return -EINVAL; | |
d6a4c185 | 2774 | |
ccd59dce LY |
2775 | if (dirty.flags & VFIO_IOMMU_DIRTY_PAGES_FLAG_START) { |
2776 | size_t pgsize; | |
d6a4c185 | 2777 | |
ccd59dce LY |
2778 | mutex_lock(&iommu->lock); |
2779 | pgsize = 1 << __ffs(iommu->pgsize_bitmap); | |
2780 | if (!iommu->dirty_page_tracking) { | |
2781 | ret = vfio_dma_bitmap_alloc_all(iommu, pgsize); | |
2782 | if (!ret) | |
2783 | iommu->dirty_page_tracking = true; | |
2784 | } | |
2785 | mutex_unlock(&iommu->lock); | |
2786 | return ret; | |
2787 | } else if (dirty.flags & VFIO_IOMMU_DIRTY_PAGES_FLAG_STOP) { | |
2788 | mutex_lock(&iommu->lock); | |
2789 | if (iommu->dirty_page_tracking) { | |
2790 | iommu->dirty_page_tracking = false; | |
2791 | vfio_dma_bitmap_free_all(iommu); | |
2792 | } | |
2793 | mutex_unlock(&iommu->lock); | |
2794 | return 0; | |
2795 | } else if (dirty.flags & VFIO_IOMMU_DIRTY_PAGES_FLAG_GET_BITMAP) { | |
2796 | struct vfio_iommu_type1_dirty_bitmap_get range; | |
2797 | unsigned long pgshift; | |
2798 | size_t data_size = dirty.argsz - minsz; | |
2799 | size_t iommu_pgsize; | |
d6a4c185 | 2800 | |
ccd59dce LY |
2801 | if (!data_size || data_size < sizeof(range)) |
2802 | return -EINVAL; | |
d6a4c185 | 2803 | |
ccd59dce LY |
2804 | if (copy_from_user(&range, (void __user *)(arg + minsz), |
2805 | sizeof(range))) | |
2806 | return -EFAULT; | |
d6a4c185 | 2807 | |
ccd59dce LY |
2808 | if (range.iova + range.size < range.iova) |
2809 | return -EINVAL; | |
2810 | if (!access_ok((void __user *)range.bitmap.data, | |
2811 | range.bitmap.size)) | |
2812 | return -EINVAL; | |
d6a4c185 | 2813 | |
ccd59dce LY |
2814 | pgshift = __ffs(range.bitmap.pgsize); |
2815 | ret = verify_bitmap_size(range.size >> pgshift, | |
2816 | range.bitmap.size); | |
2817 | if (ret) | |
2818 | return ret; | |
d6a4c185 | 2819 | |
ccd59dce | 2820 | mutex_lock(&iommu->lock); |
d6a4c185 | 2821 | |
ccd59dce LY |
2822 | iommu_pgsize = (size_t)1 << __ffs(iommu->pgsize_bitmap); |
2823 | ||
2824 | /* allow only smallest supported pgsize */ | |
2825 | if (range.bitmap.pgsize != iommu_pgsize) { | |
2826 | ret = -EINVAL; | |
2827 | goto out_unlock; | |
2828 | } | |
2829 | if (range.iova & (iommu_pgsize - 1)) { | |
2830 | ret = -EINVAL; | |
2831 | goto out_unlock; | |
2832 | } | |
2833 | if (!range.size || range.size & (iommu_pgsize - 1)) { | |
2834 | ret = -EINVAL; | |
2835 | goto out_unlock; | |
d6a4c185 | 2836 | } |
ccd59dce LY |
2837 | |
2838 | if (iommu->dirty_page_tracking) | |
2839 | ret = vfio_iova_dirty_bitmap(range.bitmap.data, | |
2840 | iommu, range.iova, | |
2841 | range.size, | |
2842 | range.bitmap.pgsize); | |
2843 | else | |
2844 | ret = -EINVAL; | |
2845 | out_unlock: | |
2846 | mutex_unlock(&iommu->lock); | |
2847 | ||
2848 | return ret; | |
73fa0d10 AW |
2849 | } |
2850 | ||
ccd59dce LY |
2851 | return -EINVAL; |
2852 | } | |
2853 | ||
2854 | static long vfio_iommu_type1_ioctl(void *iommu_data, | |
2855 | unsigned int cmd, unsigned long arg) | |
2856 | { | |
2857 | struct vfio_iommu *iommu = iommu_data; | |
2858 | ||
2859 | switch (cmd) { | |
2860 | case VFIO_CHECK_EXTENSION: | |
2861 | return vfio_iommu_type1_check_extension(iommu, arg); | |
2862 | case VFIO_IOMMU_GET_INFO: | |
2863 | return vfio_iommu_type1_get_info(iommu, arg); | |
2864 | case VFIO_IOMMU_MAP_DMA: | |
2865 | return vfio_iommu_type1_map_dma(iommu, arg); | |
2866 | case VFIO_IOMMU_UNMAP_DMA: | |
2867 | return vfio_iommu_type1_unmap_dma(iommu, arg); | |
2868 | case VFIO_IOMMU_DIRTY_PAGES: | |
2869 | return vfio_iommu_type1_dirty_pages(iommu, arg); | |
2870 | default: | |
2871 | return -ENOTTY; | |
2872 | } | |
73fa0d10 AW |
2873 | } |
2874 | ||
c086de81 | 2875 | static int vfio_iommu_type1_register_notifier(void *iommu_data, |
22195cbd | 2876 | unsigned long *events, |
c086de81 KW |
2877 | struct notifier_block *nb) |
2878 | { | |
2879 | struct vfio_iommu *iommu = iommu_data; | |
2880 | ||
22195cbd JS |
2881 | /* clear known events */ |
2882 | *events &= ~VFIO_IOMMU_NOTIFY_DMA_UNMAP; | |
2883 | ||
2884 | /* refuse to register if still events remaining */ | |
2885 | if (*events) | |
2886 | return -EINVAL; | |
2887 | ||
c086de81 KW |
2888 | return blocking_notifier_chain_register(&iommu->notifier, nb); |
2889 | } | |
2890 | ||
2891 | static int vfio_iommu_type1_unregister_notifier(void *iommu_data, | |
2892 | struct notifier_block *nb) | |
2893 | { | |
2894 | struct vfio_iommu *iommu = iommu_data; | |
2895 | ||
2896 | return blocking_notifier_chain_unregister(&iommu->notifier, nb); | |
2897 | } | |
2898 | ||
8d46c0cc YZ |
2899 | static int vfio_iommu_type1_dma_rw_chunk(struct vfio_iommu *iommu, |
2900 | dma_addr_t user_iova, void *data, | |
2901 | size_t count, bool write, | |
2902 | size_t *copied) | |
2903 | { | |
2904 | struct mm_struct *mm; | |
2905 | unsigned long vaddr; | |
2906 | struct vfio_dma *dma; | |
2907 | bool kthread = current->mm == NULL; | |
2908 | size_t offset; | |
2909 | ||
2910 | *copied = 0; | |
2911 | ||
2912 | dma = vfio_find_dma(iommu, user_iova, 1); | |
2913 | if (!dma) | |
2914 | return -EINVAL; | |
2915 | ||
2916 | if ((write && !(dma->prot & IOMMU_WRITE)) || | |
2917 | !(dma->prot & IOMMU_READ)) | |
2918 | return -EPERM; | |
2919 | ||
2920 | mm = get_task_mm(dma->task); | |
2921 | ||
2922 | if (!mm) | |
2923 | return -EPERM; | |
2924 | ||
2925 | if (kthread) | |
f5678e7f | 2926 | kthread_use_mm(mm); |
8d46c0cc YZ |
2927 | else if (current->mm != mm) |
2928 | goto out; | |
2929 | ||
2930 | offset = user_iova - dma->iova; | |
2931 | ||
2932 | if (count > dma->size - offset) | |
2933 | count = dma->size - offset; | |
2934 | ||
2935 | vaddr = dma->vaddr + offset; | |
2936 | ||
d6a4c185 | 2937 | if (write) { |
205323b8 | 2938 | *copied = copy_to_user((void __user *)vaddr, data, |
8d46c0cc | 2939 | count) ? 0 : count; |
d6a4c185 KW |
2940 | if (*copied && iommu->dirty_page_tracking) { |
2941 | unsigned long pgshift = __ffs(iommu->pgsize_bitmap); | |
2942 | /* | |
2943 | * Bitmap populated with the smallest supported page | |
2944 | * size | |
2945 | */ | |
2946 | bitmap_set(dma->bitmap, offset >> pgshift, | |
2c5af985 YZ |
2947 | ((offset + *copied - 1) >> pgshift) - |
2948 | (offset >> pgshift) + 1); | |
d6a4c185 KW |
2949 | } |
2950 | } else | |
205323b8 | 2951 | *copied = copy_from_user(data, (void __user *)vaddr, |
8d46c0cc YZ |
2952 | count) ? 0 : count; |
2953 | if (kthread) | |
f5678e7f | 2954 | kthread_unuse_mm(mm); |
8d46c0cc YZ |
2955 | out: |
2956 | mmput(mm); | |
2957 | return *copied ? 0 : -EFAULT; | |
2958 | } | |
2959 | ||
2960 | static int vfio_iommu_type1_dma_rw(void *iommu_data, dma_addr_t user_iova, | |
2961 | void *data, size_t count, bool write) | |
2962 | { | |
2963 | struct vfio_iommu *iommu = iommu_data; | |
2964 | int ret = 0; | |
2965 | size_t done; | |
2966 | ||
2967 | mutex_lock(&iommu->lock); | |
2968 | while (count > 0) { | |
2969 | ret = vfio_iommu_type1_dma_rw_chunk(iommu, user_iova, data, | |
2970 | count, write, &done); | |
2971 | if (ret) | |
2972 | break; | |
2973 | ||
2974 | count -= done; | |
2975 | data += done; | |
2976 | user_iova += done; | |
2977 | } | |
2978 | ||
2979 | mutex_unlock(&iommu->lock); | |
2980 | return ret; | |
2981 | } | |
2982 | ||
73fa0d10 | 2983 | static const struct vfio_iommu_driver_ops vfio_iommu_driver_ops_type1 = { |
c086de81 KW |
2984 | .name = "vfio-iommu-type1", |
2985 | .owner = THIS_MODULE, | |
2986 | .open = vfio_iommu_type1_open, | |
2987 | .release = vfio_iommu_type1_release, | |
2988 | .ioctl = vfio_iommu_type1_ioctl, | |
2989 | .attach_group = vfio_iommu_type1_attach_group, | |
2990 | .detach_group = vfio_iommu_type1_detach_group, | |
2991 | .pin_pages = vfio_iommu_type1_pin_pages, | |
2992 | .unpin_pages = vfio_iommu_type1_unpin_pages, | |
2993 | .register_notifier = vfio_iommu_type1_register_notifier, | |
2994 | .unregister_notifier = vfio_iommu_type1_unregister_notifier, | |
8d46c0cc | 2995 | .dma_rw = vfio_iommu_type1_dma_rw, |
73fa0d10 AW |
2996 | }; |
2997 | ||
2998 | static int __init vfio_iommu_type1_init(void) | |
2999 | { | |
73fa0d10 AW |
3000 | return vfio_register_iommu_driver(&vfio_iommu_driver_ops_type1); |
3001 | } | |
3002 | ||
3003 | static void __exit vfio_iommu_type1_cleanup(void) | |
3004 | { | |
3005 | vfio_unregister_iommu_driver(&vfio_iommu_driver_ops_type1); | |
3006 | } | |
3007 | ||
3008 | module_init(vfio_iommu_type1_init); | |
3009 | module_exit(vfio_iommu_type1_cleanup); | |
3010 | ||
3011 | MODULE_VERSION(DRIVER_VERSION); | |
3012 | MODULE_LICENSE("GPL v2"); | |
3013 | MODULE_AUTHOR(DRIVER_AUTHOR); | |
3014 | MODULE_DESCRIPTION(DRIVER_DESC); |