]>
Commit | Line | Data |
---|---|---|
133ff0ea JG |
1 | /* |
2 | * Copyright 2013 Red Hat Inc. | |
3 | * | |
4 | * This program is free software; you can redistribute it and/or modify | |
5 | * it under the terms of the GNU General Public License as published by | |
6 | * the Free Software Foundation; either version 2 of the License, or | |
7 | * (at your option) any later version. | |
8 | * | |
9 | * This program is distributed in the hope that it will be useful, | |
10 | * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
11 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
12 | * GNU General Public License for more details. | |
13 | * | |
14 | * Authors: Jérôme Glisse <jglisse@redhat.com> | |
15 | */ | |
16 | /* | |
17 | * Refer to include/linux/hmm.h for information about heterogeneous memory | |
18 | * management or HMM for short. | |
19 | */ | |
20 | #include <linux/mm.h> | |
21 | #include <linux/hmm.h> | |
da4c3c73 JG |
22 | #include <linux/rmap.h> |
23 | #include <linux/swap.h> | |
133ff0ea JG |
24 | #include <linux/slab.h> |
25 | #include <linux/sched.h> | |
da4c3c73 JG |
26 | #include <linux/swapops.h> |
27 | #include <linux/hugetlb.h> | |
7b2d55d2 | 28 | #include <linux/jump_label.h> |
c0b12405 | 29 | #include <linux/mmu_notifier.h> |
133ff0ea JG |
30 | |
31 | ||
7b2d55d2 JG |
32 | /* |
33 | * Device private memory see HMM (Documentation/vm/hmm.txt) or hmm.h | |
34 | */ | |
35 | DEFINE_STATIC_KEY_FALSE(device_private_key); | |
36 | EXPORT_SYMBOL(device_private_key); | |
37 | ||
38 | ||
133ff0ea | 39 | #ifdef CONFIG_HMM |
c0b12405 JG |
40 | static const struct mmu_notifier_ops hmm_mmu_notifier_ops; |
41 | ||
133ff0ea JG |
42 | /* |
43 | * struct hmm - HMM per mm struct | |
44 | * | |
45 | * @mm: mm struct this HMM struct is bound to | |
da4c3c73 | 46 | * @lock: lock protecting ranges list |
c0b12405 | 47 | * @sequence: we track updates to the CPU page table with a sequence number |
da4c3c73 | 48 | * @ranges: list of range being snapshotted |
c0b12405 JG |
49 | * @mirrors: list of mirrors for this mm |
50 | * @mmu_notifier: mmu notifier to track updates to CPU page table | |
51 | * @mirrors_sem: read/write semaphore protecting the mirrors list | |
133ff0ea JG |
52 | */ |
53 | struct hmm { | |
54 | struct mm_struct *mm; | |
da4c3c73 | 55 | spinlock_t lock; |
c0b12405 | 56 | atomic_t sequence; |
da4c3c73 | 57 | struct list_head ranges; |
c0b12405 JG |
58 | struct list_head mirrors; |
59 | struct mmu_notifier mmu_notifier; | |
60 | struct rw_semaphore mirrors_sem; | |
133ff0ea JG |
61 | }; |
62 | ||
63 | /* | |
64 | * hmm_register - register HMM against an mm (HMM internal) | |
65 | * | |
66 | * @mm: mm struct to attach to | |
67 | * | |
68 | * This is not intended to be used directly by device drivers. It allocates an | |
69 | * HMM struct if mm does not have one, and initializes it. | |
70 | */ | |
71 | static struct hmm *hmm_register(struct mm_struct *mm) | |
72 | { | |
c0b12405 JG |
73 | struct hmm *hmm = READ_ONCE(mm->hmm); |
74 | bool cleanup = false; | |
133ff0ea JG |
75 | |
76 | /* | |
77 | * The hmm struct can only be freed once the mm_struct goes away, | |
78 | * hence we should always have pre-allocated an new hmm struct | |
79 | * above. | |
80 | */ | |
c0b12405 JG |
81 | if (hmm) |
82 | return hmm; | |
83 | ||
84 | hmm = kmalloc(sizeof(*hmm), GFP_KERNEL); | |
85 | if (!hmm) | |
86 | return NULL; | |
87 | INIT_LIST_HEAD(&hmm->mirrors); | |
88 | init_rwsem(&hmm->mirrors_sem); | |
89 | atomic_set(&hmm->sequence, 0); | |
90 | hmm->mmu_notifier.ops = NULL; | |
da4c3c73 JG |
91 | INIT_LIST_HEAD(&hmm->ranges); |
92 | spin_lock_init(&hmm->lock); | |
c0b12405 JG |
93 | hmm->mm = mm; |
94 | ||
95 | /* | |
96 | * We should only get here if hold the mmap_sem in write mode ie on | |
97 | * registration of first mirror through hmm_mirror_register() | |
98 | */ | |
99 | hmm->mmu_notifier.ops = &hmm_mmu_notifier_ops; | |
100 | if (__mmu_notifier_register(&hmm->mmu_notifier, mm)) { | |
101 | kfree(hmm); | |
102 | return NULL; | |
103 | } | |
104 | ||
105 | spin_lock(&mm->page_table_lock); | |
106 | if (!mm->hmm) | |
107 | mm->hmm = hmm; | |
108 | else | |
109 | cleanup = true; | |
110 | spin_unlock(&mm->page_table_lock); | |
111 | ||
112 | if (cleanup) { | |
113 | mmu_notifier_unregister(&hmm->mmu_notifier, mm); | |
114 | kfree(hmm); | |
115 | } | |
116 | ||
133ff0ea JG |
117 | return mm->hmm; |
118 | } | |
119 | ||
120 | void hmm_mm_destroy(struct mm_struct *mm) | |
121 | { | |
122 | kfree(mm->hmm); | |
123 | } | |
124 | #endif /* CONFIG_HMM */ | |
c0b12405 JG |
125 | |
126 | #if IS_ENABLED(CONFIG_HMM_MIRROR) | |
127 | static void hmm_invalidate_range(struct hmm *hmm, | |
128 | enum hmm_update_type action, | |
129 | unsigned long start, | |
130 | unsigned long end) | |
131 | { | |
132 | struct hmm_mirror *mirror; | |
da4c3c73 JG |
133 | struct hmm_range *range; |
134 | ||
135 | spin_lock(&hmm->lock); | |
136 | list_for_each_entry(range, &hmm->ranges, list) { | |
137 | unsigned long addr, idx, npages; | |
138 | ||
139 | if (end < range->start || start >= range->end) | |
140 | continue; | |
141 | ||
142 | range->valid = false; | |
143 | addr = max(start, range->start); | |
144 | idx = (addr - range->start) >> PAGE_SHIFT; | |
145 | npages = (min(range->end, end) - addr) >> PAGE_SHIFT; | |
146 | memset(&range->pfns[idx], 0, sizeof(*range->pfns) * npages); | |
147 | } | |
148 | spin_unlock(&hmm->lock); | |
c0b12405 JG |
149 | |
150 | down_read(&hmm->mirrors_sem); | |
151 | list_for_each_entry(mirror, &hmm->mirrors, list) | |
152 | mirror->ops->sync_cpu_device_pagetables(mirror, action, | |
153 | start, end); | |
154 | up_read(&hmm->mirrors_sem); | |
155 | } | |
156 | ||
157 | static void hmm_invalidate_range_start(struct mmu_notifier *mn, | |
158 | struct mm_struct *mm, | |
159 | unsigned long start, | |
160 | unsigned long end) | |
161 | { | |
162 | struct hmm *hmm = mm->hmm; | |
163 | ||
164 | VM_BUG_ON(!hmm); | |
165 | ||
166 | atomic_inc(&hmm->sequence); | |
167 | } | |
168 | ||
169 | static void hmm_invalidate_range_end(struct mmu_notifier *mn, | |
170 | struct mm_struct *mm, | |
171 | unsigned long start, | |
172 | unsigned long end) | |
173 | { | |
174 | struct hmm *hmm = mm->hmm; | |
175 | ||
176 | VM_BUG_ON(!hmm); | |
177 | ||
178 | hmm_invalidate_range(mm->hmm, HMM_UPDATE_INVALIDATE, start, end); | |
179 | } | |
180 | ||
181 | static const struct mmu_notifier_ops hmm_mmu_notifier_ops = { | |
182 | .invalidate_range_start = hmm_invalidate_range_start, | |
183 | .invalidate_range_end = hmm_invalidate_range_end, | |
184 | }; | |
185 | ||
186 | /* | |
187 | * hmm_mirror_register() - register a mirror against an mm | |
188 | * | |
189 | * @mirror: new mirror struct to register | |
190 | * @mm: mm to register against | |
191 | * | |
192 | * To start mirroring a process address space, the device driver must register | |
193 | * an HMM mirror struct. | |
194 | * | |
195 | * THE mm->mmap_sem MUST BE HELD IN WRITE MODE ! | |
196 | */ | |
197 | int hmm_mirror_register(struct hmm_mirror *mirror, struct mm_struct *mm) | |
198 | { | |
199 | /* Sanity check */ | |
200 | if (!mm || !mirror || !mirror->ops) | |
201 | return -EINVAL; | |
202 | ||
203 | mirror->hmm = hmm_register(mm); | |
204 | if (!mirror->hmm) | |
205 | return -ENOMEM; | |
206 | ||
207 | down_write(&mirror->hmm->mirrors_sem); | |
208 | list_add(&mirror->list, &mirror->hmm->mirrors); | |
209 | up_write(&mirror->hmm->mirrors_sem); | |
210 | ||
211 | return 0; | |
212 | } | |
213 | EXPORT_SYMBOL(hmm_mirror_register); | |
214 | ||
215 | /* | |
216 | * hmm_mirror_unregister() - unregister a mirror | |
217 | * | |
218 | * @mirror: new mirror struct to register | |
219 | * | |
220 | * Stop mirroring a process address space, and cleanup. | |
221 | */ | |
222 | void hmm_mirror_unregister(struct hmm_mirror *mirror) | |
223 | { | |
224 | struct hmm *hmm = mirror->hmm; | |
225 | ||
226 | down_write(&hmm->mirrors_sem); | |
227 | list_del(&mirror->list); | |
228 | up_write(&hmm->mirrors_sem); | |
229 | } | |
230 | EXPORT_SYMBOL(hmm_mirror_unregister); | |
da4c3c73 | 231 | |
74eee180 JG |
232 | struct hmm_vma_walk { |
233 | struct hmm_range *range; | |
234 | unsigned long last; | |
235 | bool fault; | |
236 | bool block; | |
237 | bool write; | |
238 | }; | |
239 | ||
240 | static int hmm_vma_do_fault(struct mm_walk *walk, | |
241 | unsigned long addr, | |
242 | hmm_pfn_t *pfn) | |
243 | { | |
244 | unsigned int flags = FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_REMOTE; | |
245 | struct hmm_vma_walk *hmm_vma_walk = walk->private; | |
246 | struct vm_area_struct *vma = walk->vma; | |
247 | int r; | |
248 | ||
249 | flags |= hmm_vma_walk->block ? 0 : FAULT_FLAG_ALLOW_RETRY; | |
250 | flags |= hmm_vma_walk->write ? FAULT_FLAG_WRITE : 0; | |
251 | r = handle_mm_fault(vma, addr, flags); | |
252 | if (r & VM_FAULT_RETRY) | |
253 | return -EBUSY; | |
254 | if (r & VM_FAULT_ERROR) { | |
255 | *pfn = HMM_PFN_ERROR; | |
256 | return -EFAULT; | |
257 | } | |
258 | ||
259 | return -EAGAIN; | |
260 | } | |
261 | ||
da4c3c73 JG |
262 | static void hmm_pfns_special(hmm_pfn_t *pfns, |
263 | unsigned long addr, | |
264 | unsigned long end) | |
265 | { | |
266 | for (; addr < end; addr += PAGE_SIZE, pfns++) | |
267 | *pfns = HMM_PFN_SPECIAL; | |
268 | } | |
269 | ||
270 | static int hmm_pfns_bad(unsigned long addr, | |
271 | unsigned long end, | |
272 | struct mm_walk *walk) | |
273 | { | |
274 | struct hmm_range *range = walk->private; | |
275 | hmm_pfn_t *pfns = range->pfns; | |
276 | unsigned long i; | |
277 | ||
278 | i = (addr - range->start) >> PAGE_SHIFT; | |
279 | for (; addr < end; addr += PAGE_SIZE, i++) | |
280 | pfns[i] = HMM_PFN_ERROR; | |
281 | ||
282 | return 0; | |
283 | } | |
284 | ||
74eee180 JG |
285 | static void hmm_pfns_clear(hmm_pfn_t *pfns, |
286 | unsigned long addr, | |
287 | unsigned long end) | |
288 | { | |
289 | for (; addr < end; addr += PAGE_SIZE, pfns++) | |
290 | *pfns = 0; | |
291 | } | |
292 | ||
da4c3c73 JG |
293 | static int hmm_vma_walk_hole(unsigned long addr, |
294 | unsigned long end, | |
295 | struct mm_walk *walk) | |
296 | { | |
74eee180 JG |
297 | struct hmm_vma_walk *hmm_vma_walk = walk->private; |
298 | struct hmm_range *range = hmm_vma_walk->range; | |
da4c3c73 JG |
299 | hmm_pfn_t *pfns = range->pfns; |
300 | unsigned long i; | |
301 | ||
74eee180 | 302 | hmm_vma_walk->last = addr; |
da4c3c73 | 303 | i = (addr - range->start) >> PAGE_SHIFT; |
74eee180 | 304 | for (; addr < end; addr += PAGE_SIZE, i++) { |
da4c3c73 | 305 | pfns[i] = HMM_PFN_EMPTY; |
74eee180 JG |
306 | if (hmm_vma_walk->fault) { |
307 | int ret; | |
da4c3c73 | 308 | |
74eee180 JG |
309 | ret = hmm_vma_do_fault(walk, addr, &pfns[i]); |
310 | if (ret != -EAGAIN) | |
311 | return ret; | |
312 | } | |
313 | } | |
314 | ||
315 | return hmm_vma_walk->fault ? -EAGAIN : 0; | |
da4c3c73 JG |
316 | } |
317 | ||
318 | static int hmm_vma_walk_clear(unsigned long addr, | |
319 | unsigned long end, | |
320 | struct mm_walk *walk) | |
321 | { | |
74eee180 JG |
322 | struct hmm_vma_walk *hmm_vma_walk = walk->private; |
323 | struct hmm_range *range = hmm_vma_walk->range; | |
da4c3c73 JG |
324 | hmm_pfn_t *pfns = range->pfns; |
325 | unsigned long i; | |
326 | ||
74eee180 | 327 | hmm_vma_walk->last = addr; |
da4c3c73 | 328 | i = (addr - range->start) >> PAGE_SHIFT; |
74eee180 | 329 | for (; addr < end; addr += PAGE_SIZE, i++) { |
da4c3c73 | 330 | pfns[i] = 0; |
74eee180 JG |
331 | if (hmm_vma_walk->fault) { |
332 | int ret; | |
da4c3c73 | 333 | |
74eee180 JG |
334 | ret = hmm_vma_do_fault(walk, addr, &pfns[i]); |
335 | if (ret != -EAGAIN) | |
336 | return ret; | |
337 | } | |
338 | } | |
339 | ||
340 | return hmm_vma_walk->fault ? -EAGAIN : 0; | |
da4c3c73 JG |
341 | } |
342 | ||
343 | static int hmm_vma_walk_pmd(pmd_t *pmdp, | |
344 | unsigned long start, | |
345 | unsigned long end, | |
346 | struct mm_walk *walk) | |
347 | { | |
74eee180 JG |
348 | struct hmm_vma_walk *hmm_vma_walk = walk->private; |
349 | struct hmm_range *range = hmm_vma_walk->range; | |
da4c3c73 JG |
350 | struct vm_area_struct *vma = walk->vma; |
351 | hmm_pfn_t *pfns = range->pfns; | |
352 | unsigned long addr = start, i; | |
74eee180 | 353 | bool write_fault; |
da4c3c73 JG |
354 | hmm_pfn_t flag; |
355 | pte_t *ptep; | |
356 | ||
357 | i = (addr - range->start) >> PAGE_SHIFT; | |
358 | flag = vma->vm_flags & VM_READ ? HMM_PFN_READ : 0; | |
74eee180 | 359 | write_fault = hmm_vma_walk->fault & hmm_vma_walk->write; |
da4c3c73 JG |
360 | |
361 | again: | |
362 | if (pmd_none(*pmdp)) | |
363 | return hmm_vma_walk_hole(start, end, walk); | |
364 | ||
365 | if (pmd_huge(*pmdp) && vma->vm_flags & VM_HUGETLB) | |
366 | return hmm_pfns_bad(start, end, walk); | |
367 | ||
368 | if (pmd_devmap(*pmdp) || pmd_trans_huge(*pmdp)) { | |
369 | unsigned long pfn; | |
370 | pmd_t pmd; | |
371 | ||
372 | /* | |
373 | * No need to take pmd_lock here, even if some other threads | |
374 | * is splitting the huge pmd we will get that event through | |
375 | * mmu_notifier callback. | |
376 | * | |
377 | * So just read pmd value and check again its a transparent | |
378 | * huge or device mapping one and compute corresponding pfn | |
379 | * values. | |
380 | */ | |
381 | pmd = pmd_read_atomic(pmdp); | |
382 | barrier(); | |
383 | if (!pmd_devmap(pmd) && !pmd_trans_huge(pmd)) | |
384 | goto again; | |
385 | if (pmd_protnone(pmd)) | |
386 | return hmm_vma_walk_clear(start, end, walk); | |
387 | ||
74eee180 JG |
388 | if (write_fault && !pmd_write(pmd)) |
389 | return hmm_vma_walk_clear(start, end, walk); | |
390 | ||
da4c3c73 JG |
391 | pfn = pmd_pfn(pmd) + pte_index(addr); |
392 | flag |= pmd_write(pmd) ? HMM_PFN_WRITE : 0; | |
393 | for (; addr < end; addr += PAGE_SIZE, i++, pfn++) | |
394 | pfns[i] = hmm_pfn_t_from_pfn(pfn) | flag; | |
395 | return 0; | |
396 | } | |
397 | ||
398 | if (pmd_bad(*pmdp)) | |
399 | return hmm_pfns_bad(start, end, walk); | |
400 | ||
401 | ptep = pte_offset_map(pmdp, addr); | |
402 | for (; addr < end; addr += PAGE_SIZE, ptep++, i++) { | |
403 | pte_t pte = *ptep; | |
404 | ||
405 | pfns[i] = 0; | |
406 | ||
74eee180 | 407 | if (pte_none(pte)) { |
da4c3c73 | 408 | pfns[i] = HMM_PFN_EMPTY; |
74eee180 JG |
409 | if (hmm_vma_walk->fault) |
410 | goto fault; | |
da4c3c73 JG |
411 | continue; |
412 | } | |
413 | ||
74eee180 JG |
414 | if (!pte_present(pte)) { |
415 | swp_entry_t entry; | |
416 | ||
417 | if (!non_swap_entry(entry)) { | |
418 | if (hmm_vma_walk->fault) | |
419 | goto fault; | |
420 | continue; | |
421 | } | |
422 | ||
423 | entry = pte_to_swp_entry(pte); | |
424 | ||
425 | /* | |
426 | * This is a special swap entry, ignore migration, use | |
427 | * device and report anything else as error. | |
428 | */ | |
429 | if (is_migration_entry(entry)) { | |
430 | if (hmm_vma_walk->fault) { | |
431 | pte_unmap(ptep); | |
432 | hmm_vma_walk->last = addr; | |
433 | migration_entry_wait(vma->vm_mm, | |
434 | pmdp, addr); | |
435 | return -EAGAIN; | |
436 | } | |
437 | continue; | |
438 | } else { | |
439 | /* Report error for everything else */ | |
440 | pfns[i] = HMM_PFN_ERROR; | |
441 | } | |
442 | continue; | |
443 | } | |
444 | ||
445 | if (write_fault && !pte_write(pte)) | |
446 | goto fault; | |
447 | ||
da4c3c73 JG |
448 | pfns[i] = hmm_pfn_t_from_pfn(pte_pfn(pte)) | flag; |
449 | pfns[i] |= pte_write(pte) ? HMM_PFN_WRITE : 0; | |
74eee180 JG |
450 | continue; |
451 | ||
452 | fault: | |
453 | pte_unmap(ptep); | |
454 | /* Fault all pages in range */ | |
455 | return hmm_vma_walk_clear(start, end, walk); | |
da4c3c73 JG |
456 | } |
457 | pte_unmap(ptep - 1); | |
458 | ||
459 | return 0; | |
460 | } | |
461 | ||
462 | /* | |
463 | * hmm_vma_get_pfns() - snapshot CPU page table for a range of virtual addresses | |
464 | * @vma: virtual memory area containing the virtual address range | |
465 | * @range: used to track snapshot validity | |
466 | * @start: range virtual start address (inclusive) | |
467 | * @end: range virtual end address (exclusive) | |
468 | * @entries: array of hmm_pfn_t: provided by the caller, filled in by function | |
469 | * Returns: -EINVAL if invalid argument, -ENOMEM out of memory, 0 success | |
470 | * | |
471 | * This snapshots the CPU page table for a range of virtual addresses. Snapshot | |
472 | * validity is tracked by range struct. See hmm_vma_range_done() for further | |
473 | * information. | |
474 | * | |
475 | * The range struct is initialized here. It tracks the CPU page table, but only | |
476 | * if the function returns success (0), in which case the caller must then call | |
477 | * hmm_vma_range_done() to stop CPU page table update tracking on this range. | |
478 | * | |
479 | * NOT CALLING hmm_vma_range_done() IF FUNCTION RETURNS 0 WILL LEAD TO SERIOUS | |
480 | * MEMORY CORRUPTION ! YOU HAVE BEEN WARNED ! | |
481 | */ | |
482 | int hmm_vma_get_pfns(struct vm_area_struct *vma, | |
483 | struct hmm_range *range, | |
484 | unsigned long start, | |
485 | unsigned long end, | |
486 | hmm_pfn_t *pfns) | |
487 | { | |
74eee180 | 488 | struct hmm_vma_walk hmm_vma_walk; |
da4c3c73 JG |
489 | struct mm_walk mm_walk; |
490 | struct hmm *hmm; | |
491 | ||
492 | /* FIXME support hugetlb fs */ | |
493 | if (is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_SPECIAL)) { | |
494 | hmm_pfns_special(pfns, start, end); | |
495 | return -EINVAL; | |
496 | } | |
497 | ||
498 | /* Sanity check, this really should not happen ! */ | |
499 | if (start < vma->vm_start || start >= vma->vm_end) | |
500 | return -EINVAL; | |
501 | if (end < vma->vm_start || end > vma->vm_end) | |
502 | return -EINVAL; | |
503 | ||
504 | hmm = hmm_register(vma->vm_mm); | |
505 | if (!hmm) | |
506 | return -ENOMEM; | |
507 | /* Caller must have registered a mirror, via hmm_mirror_register() ! */ | |
508 | if (!hmm->mmu_notifier.ops) | |
509 | return -EINVAL; | |
510 | ||
511 | /* Initialize range to track CPU page table update */ | |
512 | range->start = start; | |
513 | range->pfns = pfns; | |
514 | range->end = end; | |
515 | spin_lock(&hmm->lock); | |
516 | range->valid = true; | |
517 | list_add_rcu(&range->list, &hmm->ranges); | |
518 | spin_unlock(&hmm->lock); | |
519 | ||
74eee180 JG |
520 | hmm_vma_walk.fault = false; |
521 | hmm_vma_walk.range = range; | |
522 | mm_walk.private = &hmm_vma_walk; | |
523 | ||
da4c3c73 JG |
524 | mm_walk.vma = vma; |
525 | mm_walk.mm = vma->vm_mm; | |
da4c3c73 JG |
526 | mm_walk.pte_entry = NULL; |
527 | mm_walk.test_walk = NULL; | |
528 | mm_walk.hugetlb_entry = NULL; | |
529 | mm_walk.pmd_entry = hmm_vma_walk_pmd; | |
530 | mm_walk.pte_hole = hmm_vma_walk_hole; | |
531 | ||
532 | walk_page_range(start, end, &mm_walk); | |
da4c3c73 JG |
533 | return 0; |
534 | } | |
535 | EXPORT_SYMBOL(hmm_vma_get_pfns); | |
536 | ||
537 | /* | |
538 | * hmm_vma_range_done() - stop tracking change to CPU page table over a range | |
539 | * @vma: virtual memory area containing the virtual address range | |
540 | * @range: range being tracked | |
541 | * Returns: false if range data has been invalidated, true otherwise | |
542 | * | |
543 | * Range struct is used to track updates to the CPU page table after a call to | |
544 | * either hmm_vma_get_pfns() or hmm_vma_fault(). Once the device driver is done | |
545 | * using the data, or wants to lock updates to the data it got from those | |
546 | * functions, it must call the hmm_vma_range_done() function, which will then | |
547 | * stop tracking CPU page table updates. | |
548 | * | |
549 | * Note that device driver must still implement general CPU page table update | |
550 | * tracking either by using hmm_mirror (see hmm_mirror_register()) or by using | |
551 | * the mmu_notifier API directly. | |
552 | * | |
553 | * CPU page table update tracking done through hmm_range is only temporary and | |
554 | * to be used while trying to duplicate CPU page table contents for a range of | |
555 | * virtual addresses. | |
556 | * | |
557 | * There are two ways to use this : | |
558 | * again: | |
74eee180 | 559 | * hmm_vma_get_pfns(vma, range, start, end, pfns); or hmm_vma_fault(...); |
da4c3c73 JG |
560 | * trans = device_build_page_table_update_transaction(pfns); |
561 | * device_page_table_lock(); | |
562 | * if (!hmm_vma_range_done(vma, range)) { | |
563 | * device_page_table_unlock(); | |
564 | * goto again; | |
565 | * } | |
566 | * device_commit_transaction(trans); | |
567 | * device_page_table_unlock(); | |
568 | * | |
569 | * Or: | |
74eee180 | 570 | * hmm_vma_get_pfns(vma, range, start, end, pfns); or hmm_vma_fault(...); |
da4c3c73 JG |
571 | * device_page_table_lock(); |
572 | * hmm_vma_range_done(vma, range); | |
573 | * device_update_page_table(pfns); | |
574 | * device_page_table_unlock(); | |
575 | */ | |
576 | bool hmm_vma_range_done(struct vm_area_struct *vma, struct hmm_range *range) | |
577 | { | |
578 | unsigned long npages = (range->end - range->start) >> PAGE_SHIFT; | |
579 | struct hmm *hmm; | |
580 | ||
581 | if (range->end <= range->start) { | |
582 | BUG(); | |
583 | return false; | |
584 | } | |
585 | ||
586 | hmm = hmm_register(vma->vm_mm); | |
587 | if (!hmm) { | |
588 | memset(range->pfns, 0, sizeof(*range->pfns) * npages); | |
589 | return false; | |
590 | } | |
591 | ||
592 | spin_lock(&hmm->lock); | |
593 | list_del_rcu(&range->list); | |
594 | spin_unlock(&hmm->lock); | |
595 | ||
596 | return range->valid; | |
597 | } | |
598 | EXPORT_SYMBOL(hmm_vma_range_done); | |
74eee180 JG |
599 | |
600 | /* | |
601 | * hmm_vma_fault() - try to fault some address in a virtual address range | |
602 | * @vma: virtual memory area containing the virtual address range | |
603 | * @range: use to track pfns array content validity | |
604 | * @start: fault range virtual start address (inclusive) | |
605 | * @end: fault range virtual end address (exclusive) | |
606 | * @pfns: array of hmm_pfn_t, only entry with fault flag set will be faulted | |
607 | * @write: is it a write fault | |
608 | * @block: allow blocking on fault (if true it sleeps and do not drop mmap_sem) | |
609 | * Returns: 0 success, error otherwise (-EAGAIN means mmap_sem have been drop) | |
610 | * | |
611 | * This is similar to a regular CPU page fault except that it will not trigger | |
612 | * any memory migration if the memory being faulted is not accessible by CPUs. | |
613 | * | |
614 | * On error, for one virtual address in the range, the function will set the | |
615 | * hmm_pfn_t error flag for the corresponding pfn entry. | |
616 | * | |
617 | * Expected use pattern: | |
618 | * retry: | |
619 | * down_read(&mm->mmap_sem); | |
620 | * // Find vma and address device wants to fault, initialize hmm_pfn_t | |
621 | * // array accordingly | |
622 | * ret = hmm_vma_fault(vma, start, end, pfns, allow_retry); | |
623 | * switch (ret) { | |
624 | * case -EAGAIN: | |
625 | * hmm_vma_range_done(vma, range); | |
626 | * // You might want to rate limit or yield to play nicely, you may | |
627 | * // also commit any valid pfn in the array assuming that you are | |
628 | * // getting true from hmm_vma_range_monitor_end() | |
629 | * goto retry; | |
630 | * case 0: | |
631 | * break; | |
632 | * default: | |
633 | * // Handle error ! | |
634 | * up_read(&mm->mmap_sem) | |
635 | * return; | |
636 | * } | |
637 | * // Take device driver lock that serialize device page table update | |
638 | * driver_lock_device_page_table_update(); | |
639 | * hmm_vma_range_done(vma, range); | |
640 | * // Commit pfns we got from hmm_vma_fault() | |
641 | * driver_unlock_device_page_table_update(); | |
642 | * up_read(&mm->mmap_sem) | |
643 | * | |
644 | * YOU MUST CALL hmm_vma_range_done() AFTER THIS FUNCTION RETURN SUCCESS (0) | |
645 | * BEFORE FREEING THE range struct OR YOU WILL HAVE SERIOUS MEMORY CORRUPTION ! | |
646 | * | |
647 | * YOU HAVE BEEN WARNED ! | |
648 | */ | |
649 | int hmm_vma_fault(struct vm_area_struct *vma, | |
650 | struct hmm_range *range, | |
651 | unsigned long start, | |
652 | unsigned long end, | |
653 | hmm_pfn_t *pfns, | |
654 | bool write, | |
655 | bool block) | |
656 | { | |
657 | struct hmm_vma_walk hmm_vma_walk; | |
658 | struct mm_walk mm_walk; | |
659 | struct hmm *hmm; | |
660 | int ret; | |
661 | ||
662 | /* Sanity check, this really should not happen ! */ | |
663 | if (start < vma->vm_start || start >= vma->vm_end) | |
664 | return -EINVAL; | |
665 | if (end < vma->vm_start || end > vma->vm_end) | |
666 | return -EINVAL; | |
667 | ||
668 | hmm = hmm_register(vma->vm_mm); | |
669 | if (!hmm) { | |
670 | hmm_pfns_clear(pfns, start, end); | |
671 | return -ENOMEM; | |
672 | } | |
673 | /* Caller must have registered a mirror using hmm_mirror_register() */ | |
674 | if (!hmm->mmu_notifier.ops) | |
675 | return -EINVAL; | |
676 | ||
677 | /* Initialize range to track CPU page table update */ | |
678 | range->start = start; | |
679 | range->pfns = pfns; | |
680 | range->end = end; | |
681 | spin_lock(&hmm->lock); | |
682 | range->valid = true; | |
683 | list_add_rcu(&range->list, &hmm->ranges); | |
684 | spin_unlock(&hmm->lock); | |
685 | ||
686 | /* FIXME support hugetlb fs */ | |
687 | if (is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_SPECIAL)) { | |
688 | hmm_pfns_special(pfns, start, end); | |
689 | return 0; | |
690 | } | |
691 | ||
692 | hmm_vma_walk.fault = true; | |
693 | hmm_vma_walk.write = write; | |
694 | hmm_vma_walk.block = block; | |
695 | hmm_vma_walk.range = range; | |
696 | mm_walk.private = &hmm_vma_walk; | |
697 | hmm_vma_walk.last = range->start; | |
698 | ||
699 | mm_walk.vma = vma; | |
700 | mm_walk.mm = vma->vm_mm; | |
701 | mm_walk.pte_entry = NULL; | |
702 | mm_walk.test_walk = NULL; | |
703 | mm_walk.hugetlb_entry = NULL; | |
704 | mm_walk.pmd_entry = hmm_vma_walk_pmd; | |
705 | mm_walk.pte_hole = hmm_vma_walk_hole; | |
706 | ||
707 | do { | |
708 | ret = walk_page_range(start, end, &mm_walk); | |
709 | start = hmm_vma_walk.last; | |
710 | } while (ret == -EAGAIN); | |
711 | ||
712 | if (ret) { | |
713 | unsigned long i; | |
714 | ||
715 | i = (hmm_vma_walk.last - range->start) >> PAGE_SHIFT; | |
716 | hmm_pfns_clear(&pfns[i], hmm_vma_walk.last, end); | |
717 | hmm_vma_range_done(vma, range); | |
718 | } | |
719 | return ret; | |
720 | } | |
721 | EXPORT_SYMBOL(hmm_vma_fault); | |
c0b12405 | 722 | #endif /* IS_ENABLED(CONFIG_HMM_MIRROR) */ |