]>
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 | * | |
f813f219 | 14 | * Authors: Jérôme Glisse <jglisse@redhat.com> |
133ff0ea JG |
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> | |
858b54da | 22 | #include <linux/init.h> |
da4c3c73 JG |
23 | #include <linux/rmap.h> |
24 | #include <linux/swap.h> | |
133ff0ea JG |
25 | #include <linux/slab.h> |
26 | #include <linux/sched.h> | |
4ef589dc JG |
27 | #include <linux/mmzone.h> |
28 | #include <linux/pagemap.h> | |
da4c3c73 JG |
29 | #include <linux/swapops.h> |
30 | #include <linux/hugetlb.h> | |
4ef589dc | 31 | #include <linux/memremap.h> |
7b2d55d2 | 32 | #include <linux/jump_label.h> |
55c0ece8 | 33 | #include <linux/dma-mapping.h> |
c0b12405 | 34 | #include <linux/mmu_notifier.h> |
4ef589dc JG |
35 | #include <linux/memory_hotplug.h> |
36 | ||
37 | #define PA_SECTION_SIZE (1UL << PA_SECTION_SHIFT) | |
133ff0ea | 38 | |
6b368cd4 | 39 | #if IS_ENABLED(CONFIG_HMM_MIRROR) |
c0b12405 JG |
40 | static const struct mmu_notifier_ops hmm_mmu_notifier_ops; |
41 | ||
704f3f2c JG |
42 | static inline struct hmm *mm_get_hmm(struct mm_struct *mm) |
43 | { | |
44 | struct hmm *hmm = READ_ONCE(mm->hmm); | |
45 | ||
46 | if (hmm && kref_get_unless_zero(&hmm->kref)) | |
47 | return hmm; | |
48 | ||
49 | return NULL; | |
50 | } | |
51 | ||
52 | /** | |
53 | * hmm_get_or_create - register HMM against an mm (HMM internal) | |
133ff0ea JG |
54 | * |
55 | * @mm: mm struct to attach to | |
704f3f2c JG |
56 | * Returns: returns an HMM object, either by referencing the existing |
57 | * (per-process) object, or by creating a new one. | |
133ff0ea | 58 | * |
704f3f2c JG |
59 | * This is not intended to be used directly by device drivers. If mm already |
60 | * has an HMM struct then it get a reference on it and returns it. Otherwise | |
61 | * it allocates an HMM struct, initializes it, associate it with the mm and | |
62 | * returns it. | |
133ff0ea | 63 | */ |
704f3f2c | 64 | static struct hmm *hmm_get_or_create(struct mm_struct *mm) |
133ff0ea | 65 | { |
704f3f2c | 66 | struct hmm *hmm = mm_get_hmm(mm); |
c0b12405 | 67 | bool cleanup = false; |
133ff0ea | 68 | |
c0b12405 JG |
69 | if (hmm) |
70 | return hmm; | |
71 | ||
72 | hmm = kmalloc(sizeof(*hmm), GFP_KERNEL); | |
73 | if (!hmm) | |
74 | return NULL; | |
a3e0d41c | 75 | init_waitqueue_head(&hmm->wq); |
c0b12405 JG |
76 | INIT_LIST_HEAD(&hmm->mirrors); |
77 | init_rwsem(&hmm->mirrors_sem); | |
c0b12405 | 78 | hmm->mmu_notifier.ops = NULL; |
da4c3c73 | 79 | INIT_LIST_HEAD(&hmm->ranges); |
a3e0d41c | 80 | mutex_init(&hmm->lock); |
704f3f2c | 81 | kref_init(&hmm->kref); |
a3e0d41c JG |
82 | hmm->notifiers = 0; |
83 | hmm->dead = false; | |
c0b12405 JG |
84 | hmm->mm = mm; |
85 | ||
c0b12405 JG |
86 | spin_lock(&mm->page_table_lock); |
87 | if (!mm->hmm) | |
88 | mm->hmm = hmm; | |
89 | else | |
90 | cleanup = true; | |
91 | spin_unlock(&mm->page_table_lock); | |
92 | ||
86a2d598 RC |
93 | if (cleanup) |
94 | goto error; | |
95 | ||
96 | /* | |
97 | * We should only get here if hold the mmap_sem in write mode ie on | |
98 | * registration of first mirror through hmm_mirror_register() | |
99 | */ | |
100 | hmm->mmu_notifier.ops = &hmm_mmu_notifier_ops; | |
101 | if (__mmu_notifier_register(&hmm->mmu_notifier, mm)) | |
102 | goto error_mm; | |
c0b12405 | 103 | |
704f3f2c | 104 | return hmm; |
86a2d598 RC |
105 | |
106 | error_mm: | |
107 | spin_lock(&mm->page_table_lock); | |
108 | if (mm->hmm == hmm) | |
109 | mm->hmm = NULL; | |
110 | spin_unlock(&mm->page_table_lock); | |
111 | error: | |
112 | kfree(hmm); | |
113 | return NULL; | |
133ff0ea JG |
114 | } |
115 | ||
704f3f2c JG |
116 | static void hmm_free(struct kref *kref) |
117 | { | |
118 | struct hmm *hmm = container_of(kref, struct hmm, kref); | |
119 | struct mm_struct *mm = hmm->mm; | |
120 | ||
121 | mmu_notifier_unregister_no_release(&hmm->mmu_notifier, mm); | |
122 | ||
123 | spin_lock(&mm->page_table_lock); | |
124 | if (mm->hmm == hmm) | |
125 | mm->hmm = NULL; | |
126 | spin_unlock(&mm->page_table_lock); | |
127 | ||
128 | kfree(hmm); | |
129 | } | |
130 | ||
131 | static inline void hmm_put(struct hmm *hmm) | |
132 | { | |
133 | kref_put(&hmm->kref, hmm_free); | |
134 | } | |
135 | ||
133ff0ea JG |
136 | void hmm_mm_destroy(struct mm_struct *mm) |
137 | { | |
704f3f2c JG |
138 | struct hmm *hmm; |
139 | ||
140 | spin_lock(&mm->page_table_lock); | |
141 | hmm = mm_get_hmm(mm); | |
142 | mm->hmm = NULL; | |
143 | if (hmm) { | |
144 | hmm->mm = NULL; | |
a3e0d41c | 145 | hmm->dead = true; |
704f3f2c JG |
146 | spin_unlock(&mm->page_table_lock); |
147 | hmm_put(hmm); | |
148 | return; | |
149 | } | |
150 | ||
151 | spin_unlock(&mm->page_table_lock); | |
133ff0ea | 152 | } |
c0b12405 | 153 | |
a3e0d41c | 154 | static void hmm_release(struct mmu_notifier *mn, struct mm_struct *mm) |
c0b12405 | 155 | { |
a3e0d41c | 156 | struct hmm *hmm = mm_get_hmm(mm); |
c0b12405 | 157 | struct hmm_mirror *mirror; |
da4c3c73 JG |
158 | struct hmm_range *range; |
159 | ||
a3e0d41c JG |
160 | /* Report this HMM as dying. */ |
161 | hmm->dead = true; | |
da4c3c73 | 162 | |
a3e0d41c JG |
163 | /* Wake-up everyone waiting on any range. */ |
164 | mutex_lock(&hmm->lock); | |
165 | list_for_each_entry(range, &hmm->ranges, list) { | |
da4c3c73 | 166 | range->valid = false; |
da4c3c73 | 167 | } |
a3e0d41c JG |
168 | wake_up_all(&hmm->wq); |
169 | mutex_unlock(&hmm->lock); | |
e1401513 RC |
170 | |
171 | down_write(&hmm->mirrors_sem); | |
172 | mirror = list_first_entry_or_null(&hmm->mirrors, struct hmm_mirror, | |
173 | list); | |
174 | while (mirror) { | |
175 | list_del_init(&mirror->list); | |
176 | if (mirror->ops->release) { | |
177 | /* | |
178 | * Drop mirrors_sem so callback can wait on any pending | |
179 | * work that might itself trigger mmu_notifier callback | |
180 | * and thus would deadlock with us. | |
181 | */ | |
182 | up_write(&hmm->mirrors_sem); | |
183 | mirror->ops->release(mirror); | |
184 | down_write(&hmm->mirrors_sem); | |
185 | } | |
186 | mirror = list_first_entry_or_null(&hmm->mirrors, | |
187 | struct hmm_mirror, list); | |
188 | } | |
189 | up_write(&hmm->mirrors_sem); | |
704f3f2c JG |
190 | |
191 | hmm_put(hmm); | |
e1401513 RC |
192 | } |
193 | ||
93065ac7 | 194 | static int hmm_invalidate_range_start(struct mmu_notifier *mn, |
a3e0d41c | 195 | const struct mmu_notifier_range *nrange) |
c0b12405 | 196 | { |
a3e0d41c JG |
197 | struct hmm *hmm = mm_get_hmm(nrange->mm); |
198 | struct hmm_mirror *mirror; | |
ec131b2d | 199 | struct hmm_update update; |
a3e0d41c JG |
200 | struct hmm_range *range; |
201 | int ret = 0; | |
c0b12405 JG |
202 | |
203 | VM_BUG_ON(!hmm); | |
204 | ||
a3e0d41c JG |
205 | update.start = nrange->start; |
206 | update.end = nrange->end; | |
ec131b2d | 207 | update.event = HMM_UPDATE_INVALIDATE; |
a3e0d41c JG |
208 | update.blockable = nrange->blockable; |
209 | ||
210 | if (nrange->blockable) | |
211 | mutex_lock(&hmm->lock); | |
212 | else if (!mutex_trylock(&hmm->lock)) { | |
213 | ret = -EAGAIN; | |
214 | goto out; | |
215 | } | |
216 | hmm->notifiers++; | |
217 | list_for_each_entry(range, &hmm->ranges, list) { | |
218 | if (update.end < range->start || update.start >= range->end) | |
219 | continue; | |
220 | ||
221 | range->valid = false; | |
222 | } | |
223 | mutex_unlock(&hmm->lock); | |
224 | ||
225 | if (nrange->blockable) | |
226 | down_read(&hmm->mirrors_sem); | |
227 | else if (!down_read_trylock(&hmm->mirrors_sem)) { | |
228 | ret = -EAGAIN; | |
229 | goto out; | |
230 | } | |
231 | list_for_each_entry(mirror, &hmm->mirrors, list) { | |
232 | int ret; | |
233 | ||
234 | ret = mirror->ops->sync_cpu_device_pagetables(mirror, &update); | |
235 | if (!update.blockable && ret == -EAGAIN) { | |
236 | up_read(&hmm->mirrors_sem); | |
237 | ret = -EAGAIN; | |
238 | goto out; | |
239 | } | |
240 | } | |
241 | up_read(&hmm->mirrors_sem); | |
242 | ||
243 | out: | |
704f3f2c JG |
244 | hmm_put(hmm); |
245 | return ret; | |
c0b12405 JG |
246 | } |
247 | ||
248 | static void hmm_invalidate_range_end(struct mmu_notifier *mn, | |
a3e0d41c | 249 | const struct mmu_notifier_range *nrange) |
c0b12405 | 250 | { |
a3e0d41c | 251 | struct hmm *hmm = mm_get_hmm(nrange->mm); |
c0b12405 JG |
252 | |
253 | VM_BUG_ON(!hmm); | |
254 | ||
a3e0d41c JG |
255 | mutex_lock(&hmm->lock); |
256 | hmm->notifiers--; | |
257 | if (!hmm->notifiers) { | |
258 | struct hmm_range *range; | |
259 | ||
260 | list_for_each_entry(range, &hmm->ranges, list) { | |
261 | if (range->valid) | |
262 | continue; | |
263 | range->valid = true; | |
264 | } | |
265 | wake_up_all(&hmm->wq); | |
266 | } | |
267 | mutex_unlock(&hmm->lock); | |
268 | ||
704f3f2c | 269 | hmm_put(hmm); |
c0b12405 JG |
270 | } |
271 | ||
272 | static const struct mmu_notifier_ops hmm_mmu_notifier_ops = { | |
e1401513 | 273 | .release = hmm_release, |
c0b12405 JG |
274 | .invalidate_range_start = hmm_invalidate_range_start, |
275 | .invalidate_range_end = hmm_invalidate_range_end, | |
276 | }; | |
277 | ||
278 | /* | |
279 | * hmm_mirror_register() - register a mirror against an mm | |
280 | * | |
281 | * @mirror: new mirror struct to register | |
282 | * @mm: mm to register against | |
283 | * | |
284 | * To start mirroring a process address space, the device driver must register | |
285 | * an HMM mirror struct. | |
286 | * | |
287 | * THE mm->mmap_sem MUST BE HELD IN WRITE MODE ! | |
288 | */ | |
289 | int hmm_mirror_register(struct hmm_mirror *mirror, struct mm_struct *mm) | |
290 | { | |
291 | /* Sanity check */ | |
292 | if (!mm || !mirror || !mirror->ops) | |
293 | return -EINVAL; | |
294 | ||
704f3f2c | 295 | mirror->hmm = hmm_get_or_create(mm); |
c0b12405 JG |
296 | if (!mirror->hmm) |
297 | return -ENOMEM; | |
298 | ||
299 | down_write(&mirror->hmm->mirrors_sem); | |
704f3f2c JG |
300 | list_add(&mirror->list, &mirror->hmm->mirrors); |
301 | up_write(&mirror->hmm->mirrors_sem); | |
c0b12405 JG |
302 | |
303 | return 0; | |
304 | } | |
305 | EXPORT_SYMBOL(hmm_mirror_register); | |
306 | ||
307 | /* | |
308 | * hmm_mirror_unregister() - unregister a mirror | |
309 | * | |
310 | * @mirror: new mirror struct to register | |
311 | * | |
312 | * Stop mirroring a process address space, and cleanup. | |
313 | */ | |
314 | void hmm_mirror_unregister(struct hmm_mirror *mirror) | |
315 | { | |
704f3f2c | 316 | struct hmm *hmm = READ_ONCE(mirror->hmm); |
c01cbba2 | 317 | |
704f3f2c | 318 | if (hmm == NULL) |
c01cbba2 | 319 | return; |
c0b12405 JG |
320 | |
321 | down_write(&hmm->mirrors_sem); | |
e1401513 | 322 | list_del_init(&mirror->list); |
704f3f2c | 323 | /* To protect us against double unregister ... */ |
c01cbba2 | 324 | mirror->hmm = NULL; |
c0b12405 | 325 | up_write(&hmm->mirrors_sem); |
c01cbba2 | 326 | |
704f3f2c | 327 | hmm_put(hmm); |
c0b12405 JG |
328 | } |
329 | EXPORT_SYMBOL(hmm_mirror_unregister); | |
da4c3c73 | 330 | |
74eee180 JG |
331 | struct hmm_vma_walk { |
332 | struct hmm_range *range; | |
992de9a8 | 333 | struct dev_pagemap *pgmap; |
74eee180 JG |
334 | unsigned long last; |
335 | bool fault; | |
336 | bool block; | |
74eee180 JG |
337 | }; |
338 | ||
2aee09d8 JG |
339 | static int hmm_vma_do_fault(struct mm_walk *walk, unsigned long addr, |
340 | bool write_fault, uint64_t *pfn) | |
74eee180 JG |
341 | { |
342 | unsigned int flags = FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_REMOTE; | |
343 | struct hmm_vma_walk *hmm_vma_walk = walk->private; | |
f88a1e90 | 344 | struct hmm_range *range = hmm_vma_walk->range; |
74eee180 | 345 | struct vm_area_struct *vma = walk->vma; |
50a7ca3c | 346 | vm_fault_t ret; |
74eee180 JG |
347 | |
348 | flags |= hmm_vma_walk->block ? 0 : FAULT_FLAG_ALLOW_RETRY; | |
2aee09d8 | 349 | flags |= write_fault ? FAULT_FLAG_WRITE : 0; |
50a7ca3c SJ |
350 | ret = handle_mm_fault(vma, addr, flags); |
351 | if (ret & VM_FAULT_RETRY) | |
73231612 | 352 | return -EAGAIN; |
50a7ca3c | 353 | if (ret & VM_FAULT_ERROR) { |
f88a1e90 | 354 | *pfn = range->values[HMM_PFN_ERROR]; |
74eee180 JG |
355 | return -EFAULT; |
356 | } | |
357 | ||
73231612 | 358 | return -EBUSY; |
74eee180 JG |
359 | } |
360 | ||
da4c3c73 JG |
361 | static int hmm_pfns_bad(unsigned long addr, |
362 | unsigned long end, | |
363 | struct mm_walk *walk) | |
364 | { | |
c719547f JG |
365 | struct hmm_vma_walk *hmm_vma_walk = walk->private; |
366 | struct hmm_range *range = hmm_vma_walk->range; | |
ff05c0c6 | 367 | uint64_t *pfns = range->pfns; |
da4c3c73 JG |
368 | unsigned long i; |
369 | ||
370 | i = (addr - range->start) >> PAGE_SHIFT; | |
371 | for (; addr < end; addr += PAGE_SIZE, i++) | |
f88a1e90 | 372 | pfns[i] = range->values[HMM_PFN_ERROR]; |
da4c3c73 JG |
373 | |
374 | return 0; | |
375 | } | |
376 | ||
5504ed29 JG |
377 | /* |
378 | * hmm_vma_walk_hole() - handle a range lacking valid pmd or pte(s) | |
379 | * @start: range virtual start address (inclusive) | |
380 | * @end: range virtual end address (exclusive) | |
2aee09d8 JG |
381 | * @fault: should we fault or not ? |
382 | * @write_fault: write fault ? | |
5504ed29 | 383 | * @walk: mm_walk structure |
73231612 | 384 | * Returns: 0 on success, -EBUSY after page fault, or page fault error |
5504ed29 JG |
385 | * |
386 | * This function will be called whenever pmd_none() or pte_none() returns true, | |
387 | * or whenever there is no page directory covering the virtual address range. | |
388 | */ | |
2aee09d8 JG |
389 | static int hmm_vma_walk_hole_(unsigned long addr, unsigned long end, |
390 | bool fault, bool write_fault, | |
391 | struct mm_walk *walk) | |
da4c3c73 | 392 | { |
74eee180 JG |
393 | struct hmm_vma_walk *hmm_vma_walk = walk->private; |
394 | struct hmm_range *range = hmm_vma_walk->range; | |
ff05c0c6 | 395 | uint64_t *pfns = range->pfns; |
63d5066f | 396 | unsigned long i, page_size; |
da4c3c73 | 397 | |
74eee180 | 398 | hmm_vma_walk->last = addr; |
63d5066f JG |
399 | page_size = hmm_range_page_size(range); |
400 | i = (addr - range->start) >> range->page_shift; | |
401 | ||
402 | for (; addr < end; addr += page_size, i++) { | |
f88a1e90 | 403 | pfns[i] = range->values[HMM_PFN_NONE]; |
2aee09d8 | 404 | if (fault || write_fault) { |
74eee180 | 405 | int ret; |
da4c3c73 | 406 | |
2aee09d8 JG |
407 | ret = hmm_vma_do_fault(walk, addr, write_fault, |
408 | &pfns[i]); | |
73231612 | 409 | if (ret != -EBUSY) |
74eee180 JG |
410 | return ret; |
411 | } | |
412 | } | |
413 | ||
73231612 | 414 | return (fault || write_fault) ? -EBUSY : 0; |
2aee09d8 JG |
415 | } |
416 | ||
417 | static inline void hmm_pte_need_fault(const struct hmm_vma_walk *hmm_vma_walk, | |
418 | uint64_t pfns, uint64_t cpu_flags, | |
419 | bool *fault, bool *write_fault) | |
420 | { | |
f88a1e90 JG |
421 | struct hmm_range *range = hmm_vma_walk->range; |
422 | ||
2aee09d8 JG |
423 | if (!hmm_vma_walk->fault) |
424 | return; | |
425 | ||
023a019a JG |
426 | /* |
427 | * So we not only consider the individual per page request we also | |
428 | * consider the default flags requested for the range. The API can | |
429 | * be use in 2 fashions. The first one where the HMM user coalesce | |
430 | * multiple page fault into one request and set flags per pfns for | |
431 | * of those faults. The second one where the HMM user want to pre- | |
432 | * fault a range with specific flags. For the latter one it is a | |
433 | * waste to have the user pre-fill the pfn arrays with a default | |
434 | * flags value. | |
435 | */ | |
436 | pfns = (pfns & range->pfn_flags_mask) | range->default_flags; | |
437 | ||
2aee09d8 | 438 | /* We aren't ask to do anything ... */ |
f88a1e90 | 439 | if (!(pfns & range->flags[HMM_PFN_VALID])) |
2aee09d8 | 440 | return; |
f88a1e90 JG |
441 | /* If this is device memory than only fault if explicitly requested */ |
442 | if ((cpu_flags & range->flags[HMM_PFN_DEVICE_PRIVATE])) { | |
443 | /* Do we fault on device memory ? */ | |
444 | if (pfns & range->flags[HMM_PFN_DEVICE_PRIVATE]) { | |
445 | *write_fault = pfns & range->flags[HMM_PFN_WRITE]; | |
446 | *fault = true; | |
447 | } | |
2aee09d8 JG |
448 | return; |
449 | } | |
f88a1e90 JG |
450 | |
451 | /* If CPU page table is not valid then we need to fault */ | |
452 | *fault = !(cpu_flags & range->flags[HMM_PFN_VALID]); | |
453 | /* Need to write fault ? */ | |
454 | if ((pfns & range->flags[HMM_PFN_WRITE]) && | |
455 | !(cpu_flags & range->flags[HMM_PFN_WRITE])) { | |
456 | *write_fault = true; | |
2aee09d8 JG |
457 | *fault = true; |
458 | } | |
459 | } | |
460 | ||
461 | static void hmm_range_need_fault(const struct hmm_vma_walk *hmm_vma_walk, | |
462 | const uint64_t *pfns, unsigned long npages, | |
463 | uint64_t cpu_flags, bool *fault, | |
464 | bool *write_fault) | |
465 | { | |
466 | unsigned long i; | |
467 | ||
468 | if (!hmm_vma_walk->fault) { | |
469 | *fault = *write_fault = false; | |
470 | return; | |
471 | } | |
472 | ||
a3e0d41c | 473 | *fault = *write_fault = false; |
2aee09d8 JG |
474 | for (i = 0; i < npages; ++i) { |
475 | hmm_pte_need_fault(hmm_vma_walk, pfns[i], cpu_flags, | |
476 | fault, write_fault); | |
a3e0d41c | 477 | if ((*write_fault)) |
2aee09d8 JG |
478 | return; |
479 | } | |
480 | } | |
481 | ||
482 | static int hmm_vma_walk_hole(unsigned long addr, unsigned long end, | |
483 | struct mm_walk *walk) | |
484 | { | |
485 | struct hmm_vma_walk *hmm_vma_walk = walk->private; | |
486 | struct hmm_range *range = hmm_vma_walk->range; | |
487 | bool fault, write_fault; | |
488 | unsigned long i, npages; | |
489 | uint64_t *pfns; | |
490 | ||
491 | i = (addr - range->start) >> PAGE_SHIFT; | |
492 | npages = (end - addr) >> PAGE_SHIFT; | |
493 | pfns = &range->pfns[i]; | |
494 | hmm_range_need_fault(hmm_vma_walk, pfns, npages, | |
495 | 0, &fault, &write_fault); | |
496 | return hmm_vma_walk_hole_(addr, end, fault, write_fault, walk); | |
497 | } | |
498 | ||
f88a1e90 | 499 | static inline uint64_t pmd_to_hmm_pfn_flags(struct hmm_range *range, pmd_t pmd) |
2aee09d8 JG |
500 | { |
501 | if (pmd_protnone(pmd)) | |
502 | return 0; | |
f88a1e90 JG |
503 | return pmd_write(pmd) ? range->flags[HMM_PFN_VALID] | |
504 | range->flags[HMM_PFN_WRITE] : | |
505 | range->flags[HMM_PFN_VALID]; | |
da4c3c73 JG |
506 | } |
507 | ||
992de9a8 JG |
508 | static inline uint64_t pud_to_hmm_pfn_flags(struct hmm_range *range, pud_t pud) |
509 | { | |
510 | if (!pud_present(pud)) | |
511 | return 0; | |
512 | return pud_write(pud) ? range->flags[HMM_PFN_VALID] | | |
513 | range->flags[HMM_PFN_WRITE] : | |
514 | range->flags[HMM_PFN_VALID]; | |
515 | } | |
516 | ||
53f5c3f4 JG |
517 | static int hmm_vma_handle_pmd(struct mm_walk *walk, |
518 | unsigned long addr, | |
519 | unsigned long end, | |
520 | uint64_t *pfns, | |
521 | pmd_t pmd) | |
522 | { | |
992de9a8 | 523 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE |
53f5c3f4 | 524 | struct hmm_vma_walk *hmm_vma_walk = walk->private; |
f88a1e90 | 525 | struct hmm_range *range = hmm_vma_walk->range; |
2aee09d8 | 526 | unsigned long pfn, npages, i; |
2aee09d8 | 527 | bool fault, write_fault; |
f88a1e90 | 528 | uint64_t cpu_flags; |
53f5c3f4 | 529 | |
2aee09d8 | 530 | npages = (end - addr) >> PAGE_SHIFT; |
f88a1e90 | 531 | cpu_flags = pmd_to_hmm_pfn_flags(range, pmd); |
2aee09d8 JG |
532 | hmm_range_need_fault(hmm_vma_walk, pfns, npages, cpu_flags, |
533 | &fault, &write_fault); | |
53f5c3f4 | 534 | |
2aee09d8 JG |
535 | if (pmd_protnone(pmd) || fault || write_fault) |
536 | return hmm_vma_walk_hole_(addr, end, fault, write_fault, walk); | |
53f5c3f4 JG |
537 | |
538 | pfn = pmd_pfn(pmd) + pte_index(addr); | |
992de9a8 JG |
539 | for (i = 0; addr < end; addr += PAGE_SIZE, i++, pfn++) { |
540 | if (pmd_devmap(pmd)) { | |
541 | hmm_vma_walk->pgmap = get_dev_pagemap(pfn, | |
542 | hmm_vma_walk->pgmap); | |
543 | if (unlikely(!hmm_vma_walk->pgmap)) | |
544 | return -EBUSY; | |
545 | } | |
f88a1e90 | 546 | pfns[i] = hmm_pfn_from_pfn(range, pfn) | cpu_flags; |
992de9a8 JG |
547 | } |
548 | if (hmm_vma_walk->pgmap) { | |
549 | put_dev_pagemap(hmm_vma_walk->pgmap); | |
550 | hmm_vma_walk->pgmap = NULL; | |
551 | } | |
53f5c3f4 JG |
552 | hmm_vma_walk->last = end; |
553 | return 0; | |
992de9a8 JG |
554 | #else |
555 | /* If THP is not enabled then we should never reach that code ! */ | |
556 | return -EINVAL; | |
557 | #endif | |
53f5c3f4 JG |
558 | } |
559 | ||
f88a1e90 | 560 | static inline uint64_t pte_to_hmm_pfn_flags(struct hmm_range *range, pte_t pte) |
2aee09d8 JG |
561 | { |
562 | if (pte_none(pte) || !pte_present(pte)) | |
563 | return 0; | |
f88a1e90 JG |
564 | return pte_write(pte) ? range->flags[HMM_PFN_VALID] | |
565 | range->flags[HMM_PFN_WRITE] : | |
566 | range->flags[HMM_PFN_VALID]; | |
2aee09d8 JG |
567 | } |
568 | ||
53f5c3f4 JG |
569 | static int hmm_vma_handle_pte(struct mm_walk *walk, unsigned long addr, |
570 | unsigned long end, pmd_t *pmdp, pte_t *ptep, | |
571 | uint64_t *pfn) | |
572 | { | |
573 | struct hmm_vma_walk *hmm_vma_walk = walk->private; | |
f88a1e90 | 574 | struct hmm_range *range = hmm_vma_walk->range; |
53f5c3f4 | 575 | struct vm_area_struct *vma = walk->vma; |
2aee09d8 JG |
576 | bool fault, write_fault; |
577 | uint64_t cpu_flags; | |
53f5c3f4 | 578 | pte_t pte = *ptep; |
f88a1e90 | 579 | uint64_t orig_pfn = *pfn; |
53f5c3f4 | 580 | |
f88a1e90 | 581 | *pfn = range->values[HMM_PFN_NONE]; |
73231612 | 582 | fault = write_fault = false; |
53f5c3f4 JG |
583 | |
584 | if (pte_none(pte)) { | |
73231612 JG |
585 | hmm_pte_need_fault(hmm_vma_walk, orig_pfn, 0, |
586 | &fault, &write_fault); | |
2aee09d8 | 587 | if (fault || write_fault) |
53f5c3f4 JG |
588 | goto fault; |
589 | return 0; | |
590 | } | |
591 | ||
592 | if (!pte_present(pte)) { | |
593 | swp_entry_t entry = pte_to_swp_entry(pte); | |
594 | ||
595 | if (!non_swap_entry(entry)) { | |
2aee09d8 | 596 | if (fault || write_fault) |
53f5c3f4 JG |
597 | goto fault; |
598 | return 0; | |
599 | } | |
600 | ||
601 | /* | |
602 | * This is a special swap entry, ignore migration, use | |
603 | * device and report anything else as error. | |
604 | */ | |
605 | if (is_device_private_entry(entry)) { | |
f88a1e90 JG |
606 | cpu_flags = range->flags[HMM_PFN_VALID] | |
607 | range->flags[HMM_PFN_DEVICE_PRIVATE]; | |
2aee09d8 | 608 | cpu_flags |= is_write_device_private_entry(entry) ? |
f88a1e90 JG |
609 | range->flags[HMM_PFN_WRITE] : 0; |
610 | hmm_pte_need_fault(hmm_vma_walk, orig_pfn, cpu_flags, | |
611 | &fault, &write_fault); | |
612 | if (fault || write_fault) | |
613 | goto fault; | |
614 | *pfn = hmm_pfn_from_pfn(range, swp_offset(entry)); | |
615 | *pfn |= cpu_flags; | |
53f5c3f4 JG |
616 | return 0; |
617 | } | |
618 | ||
619 | if (is_migration_entry(entry)) { | |
2aee09d8 | 620 | if (fault || write_fault) { |
53f5c3f4 JG |
621 | pte_unmap(ptep); |
622 | hmm_vma_walk->last = addr; | |
623 | migration_entry_wait(vma->vm_mm, | |
2aee09d8 | 624 | pmdp, addr); |
73231612 | 625 | return -EBUSY; |
53f5c3f4 JG |
626 | } |
627 | return 0; | |
628 | } | |
629 | ||
630 | /* Report error for everything else */ | |
f88a1e90 | 631 | *pfn = range->values[HMM_PFN_ERROR]; |
53f5c3f4 | 632 | return -EFAULT; |
73231612 JG |
633 | } else { |
634 | cpu_flags = pte_to_hmm_pfn_flags(range, pte); | |
635 | hmm_pte_need_fault(hmm_vma_walk, orig_pfn, cpu_flags, | |
636 | &fault, &write_fault); | |
53f5c3f4 JG |
637 | } |
638 | ||
2aee09d8 | 639 | if (fault || write_fault) |
53f5c3f4 JG |
640 | goto fault; |
641 | ||
992de9a8 JG |
642 | if (pte_devmap(pte)) { |
643 | hmm_vma_walk->pgmap = get_dev_pagemap(pte_pfn(pte), | |
644 | hmm_vma_walk->pgmap); | |
645 | if (unlikely(!hmm_vma_walk->pgmap)) | |
646 | return -EBUSY; | |
647 | } else if (IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL) && pte_special(pte)) { | |
648 | *pfn = range->values[HMM_PFN_SPECIAL]; | |
649 | return -EFAULT; | |
650 | } | |
651 | ||
f88a1e90 | 652 | *pfn = hmm_pfn_from_pfn(range, pte_pfn(pte)) | cpu_flags; |
53f5c3f4 JG |
653 | return 0; |
654 | ||
655 | fault: | |
992de9a8 JG |
656 | if (hmm_vma_walk->pgmap) { |
657 | put_dev_pagemap(hmm_vma_walk->pgmap); | |
658 | hmm_vma_walk->pgmap = NULL; | |
659 | } | |
53f5c3f4 JG |
660 | pte_unmap(ptep); |
661 | /* Fault any virtual address we were asked to fault */ | |
2aee09d8 | 662 | return hmm_vma_walk_hole_(addr, end, fault, write_fault, walk); |
53f5c3f4 JG |
663 | } |
664 | ||
da4c3c73 JG |
665 | static int hmm_vma_walk_pmd(pmd_t *pmdp, |
666 | unsigned long start, | |
667 | unsigned long end, | |
668 | struct mm_walk *walk) | |
669 | { | |
74eee180 JG |
670 | struct hmm_vma_walk *hmm_vma_walk = walk->private; |
671 | struct hmm_range *range = hmm_vma_walk->range; | |
d08faca0 | 672 | struct vm_area_struct *vma = walk->vma; |
ff05c0c6 | 673 | uint64_t *pfns = range->pfns; |
da4c3c73 | 674 | unsigned long addr = start, i; |
da4c3c73 | 675 | pte_t *ptep; |
d08faca0 | 676 | pmd_t pmd; |
da4c3c73 | 677 | |
da4c3c73 JG |
678 | |
679 | again: | |
d08faca0 JG |
680 | pmd = READ_ONCE(*pmdp); |
681 | if (pmd_none(pmd)) | |
da4c3c73 JG |
682 | return hmm_vma_walk_hole(start, end, walk); |
683 | ||
d08faca0 | 684 | if (pmd_huge(pmd) && (range->vma->vm_flags & VM_HUGETLB)) |
da4c3c73 JG |
685 | return hmm_pfns_bad(start, end, walk); |
686 | ||
d08faca0 JG |
687 | if (thp_migration_supported() && is_pmd_migration_entry(pmd)) { |
688 | bool fault, write_fault; | |
689 | unsigned long npages; | |
690 | uint64_t *pfns; | |
691 | ||
692 | i = (addr - range->start) >> PAGE_SHIFT; | |
693 | npages = (end - addr) >> PAGE_SHIFT; | |
694 | pfns = &range->pfns[i]; | |
695 | ||
696 | hmm_range_need_fault(hmm_vma_walk, pfns, npages, | |
697 | 0, &fault, &write_fault); | |
698 | if (fault || write_fault) { | |
699 | hmm_vma_walk->last = addr; | |
700 | pmd_migration_entry_wait(vma->vm_mm, pmdp); | |
73231612 | 701 | return -EBUSY; |
d08faca0 JG |
702 | } |
703 | return 0; | |
704 | } else if (!pmd_present(pmd)) | |
705 | return hmm_pfns_bad(start, end, walk); | |
da4c3c73 | 706 | |
d08faca0 | 707 | if (pmd_devmap(pmd) || pmd_trans_huge(pmd)) { |
da4c3c73 JG |
708 | /* |
709 | * No need to take pmd_lock here, even if some other threads | |
710 | * is splitting the huge pmd we will get that event through | |
711 | * mmu_notifier callback. | |
712 | * | |
713 | * So just read pmd value and check again its a transparent | |
714 | * huge or device mapping one and compute corresponding pfn | |
715 | * values. | |
716 | */ | |
717 | pmd = pmd_read_atomic(pmdp); | |
718 | barrier(); | |
719 | if (!pmd_devmap(pmd) && !pmd_trans_huge(pmd)) | |
720 | goto again; | |
74eee180 | 721 | |
d08faca0 | 722 | i = (addr - range->start) >> PAGE_SHIFT; |
53f5c3f4 | 723 | return hmm_vma_handle_pmd(walk, addr, end, &pfns[i], pmd); |
da4c3c73 JG |
724 | } |
725 | ||
d08faca0 JG |
726 | /* |
727 | * We have handled all the valid case above ie either none, migration, | |
728 | * huge or transparent huge. At this point either it is a valid pmd | |
729 | * entry pointing to pte directory or it is a bad pmd that will not | |
730 | * recover. | |
731 | */ | |
732 | if (pmd_bad(pmd)) | |
da4c3c73 JG |
733 | return hmm_pfns_bad(start, end, walk); |
734 | ||
735 | ptep = pte_offset_map(pmdp, addr); | |
d08faca0 | 736 | i = (addr - range->start) >> PAGE_SHIFT; |
da4c3c73 | 737 | for (; addr < end; addr += PAGE_SIZE, ptep++, i++) { |
53f5c3f4 | 738 | int r; |
74eee180 | 739 | |
53f5c3f4 JG |
740 | r = hmm_vma_handle_pte(walk, addr, end, pmdp, ptep, &pfns[i]); |
741 | if (r) { | |
742 | /* hmm_vma_handle_pte() did unmap pte directory */ | |
743 | hmm_vma_walk->last = addr; | |
744 | return r; | |
74eee180 | 745 | } |
da4c3c73 | 746 | } |
992de9a8 JG |
747 | if (hmm_vma_walk->pgmap) { |
748 | /* | |
749 | * We do put_dev_pagemap() here and not in hmm_vma_handle_pte() | |
750 | * so that we can leverage get_dev_pagemap() optimization which | |
751 | * will not re-take a reference on a pgmap if we already have | |
752 | * one. | |
753 | */ | |
754 | put_dev_pagemap(hmm_vma_walk->pgmap); | |
755 | hmm_vma_walk->pgmap = NULL; | |
756 | } | |
da4c3c73 JG |
757 | pte_unmap(ptep - 1); |
758 | ||
53f5c3f4 | 759 | hmm_vma_walk->last = addr; |
da4c3c73 JG |
760 | return 0; |
761 | } | |
762 | ||
992de9a8 JG |
763 | static int hmm_vma_walk_pud(pud_t *pudp, |
764 | unsigned long start, | |
765 | unsigned long end, | |
766 | struct mm_walk *walk) | |
767 | { | |
768 | struct hmm_vma_walk *hmm_vma_walk = walk->private; | |
769 | struct hmm_range *range = hmm_vma_walk->range; | |
770 | unsigned long addr = start, next; | |
771 | pmd_t *pmdp; | |
772 | pud_t pud; | |
773 | int ret; | |
774 | ||
775 | again: | |
776 | pud = READ_ONCE(*pudp); | |
777 | if (pud_none(pud)) | |
778 | return hmm_vma_walk_hole(start, end, walk); | |
779 | ||
780 | if (pud_huge(pud) && pud_devmap(pud)) { | |
781 | unsigned long i, npages, pfn; | |
782 | uint64_t *pfns, cpu_flags; | |
783 | bool fault, write_fault; | |
784 | ||
785 | if (!pud_present(pud)) | |
786 | return hmm_vma_walk_hole(start, end, walk); | |
787 | ||
788 | i = (addr - range->start) >> PAGE_SHIFT; | |
789 | npages = (end - addr) >> PAGE_SHIFT; | |
790 | pfns = &range->pfns[i]; | |
791 | ||
792 | cpu_flags = pud_to_hmm_pfn_flags(range, pud); | |
793 | hmm_range_need_fault(hmm_vma_walk, pfns, npages, | |
794 | cpu_flags, &fault, &write_fault); | |
795 | if (fault || write_fault) | |
796 | return hmm_vma_walk_hole_(addr, end, fault, | |
797 | write_fault, walk); | |
798 | ||
799 | #ifdef CONFIG_HUGETLB_PAGE | |
800 | pfn = pud_pfn(pud) + ((addr & ~PUD_MASK) >> PAGE_SHIFT); | |
801 | for (i = 0; i < npages; ++i, ++pfn) { | |
802 | hmm_vma_walk->pgmap = get_dev_pagemap(pfn, | |
803 | hmm_vma_walk->pgmap); | |
804 | if (unlikely(!hmm_vma_walk->pgmap)) | |
805 | return -EBUSY; | |
806 | pfns[i] = hmm_pfn_from_pfn(range, pfn) | cpu_flags; | |
807 | } | |
808 | if (hmm_vma_walk->pgmap) { | |
809 | put_dev_pagemap(hmm_vma_walk->pgmap); | |
810 | hmm_vma_walk->pgmap = NULL; | |
811 | } | |
812 | hmm_vma_walk->last = end; | |
813 | return 0; | |
814 | #else | |
815 | return -EINVAL; | |
816 | #endif | |
817 | } | |
818 | ||
819 | split_huge_pud(walk->vma, pudp, addr); | |
820 | if (pud_none(*pudp)) | |
821 | goto again; | |
822 | ||
823 | pmdp = pmd_offset(pudp, addr); | |
824 | do { | |
825 | next = pmd_addr_end(addr, end); | |
826 | ret = hmm_vma_walk_pmd(pmdp, addr, next, walk); | |
827 | if (ret) | |
828 | return ret; | |
829 | } while (pmdp++, addr = next, addr != end); | |
830 | ||
831 | return 0; | |
832 | } | |
833 | ||
63d5066f JG |
834 | static int hmm_vma_walk_hugetlb_entry(pte_t *pte, unsigned long hmask, |
835 | unsigned long start, unsigned long end, | |
836 | struct mm_walk *walk) | |
837 | { | |
838 | #ifdef CONFIG_HUGETLB_PAGE | |
839 | unsigned long addr = start, i, pfn, mask, size, pfn_inc; | |
840 | struct hmm_vma_walk *hmm_vma_walk = walk->private; | |
841 | struct hmm_range *range = hmm_vma_walk->range; | |
842 | struct vm_area_struct *vma = walk->vma; | |
843 | struct hstate *h = hstate_vma(vma); | |
844 | uint64_t orig_pfn, cpu_flags; | |
845 | bool fault, write_fault; | |
846 | spinlock_t *ptl; | |
847 | pte_t entry; | |
848 | int ret = 0; | |
849 | ||
850 | size = 1UL << huge_page_shift(h); | |
851 | mask = size - 1; | |
852 | if (range->page_shift != PAGE_SHIFT) { | |
853 | /* Make sure we are looking at full page. */ | |
854 | if (start & mask) | |
855 | return -EINVAL; | |
856 | if (end < (start + size)) | |
857 | return -EINVAL; | |
858 | pfn_inc = size >> PAGE_SHIFT; | |
859 | } else { | |
860 | pfn_inc = 1; | |
861 | size = PAGE_SIZE; | |
862 | } | |
863 | ||
864 | ||
865 | ptl = huge_pte_lock(hstate_vma(walk->vma), walk->mm, pte); | |
866 | entry = huge_ptep_get(pte); | |
867 | ||
868 | i = (start - range->start) >> range->page_shift; | |
869 | orig_pfn = range->pfns[i]; | |
870 | range->pfns[i] = range->values[HMM_PFN_NONE]; | |
871 | cpu_flags = pte_to_hmm_pfn_flags(range, entry); | |
872 | fault = write_fault = false; | |
873 | hmm_pte_need_fault(hmm_vma_walk, orig_pfn, cpu_flags, | |
874 | &fault, &write_fault); | |
875 | if (fault || write_fault) { | |
876 | ret = -ENOENT; | |
877 | goto unlock; | |
878 | } | |
879 | ||
880 | pfn = pte_pfn(entry) + ((start & mask) >> range->page_shift); | |
881 | for (; addr < end; addr += size, i++, pfn += pfn_inc) | |
882 | range->pfns[i] = hmm_pfn_from_pfn(range, pfn) | cpu_flags; | |
883 | hmm_vma_walk->last = end; | |
884 | ||
885 | unlock: | |
886 | spin_unlock(ptl); | |
887 | ||
888 | if (ret == -ENOENT) | |
889 | return hmm_vma_walk_hole_(addr, end, fault, write_fault, walk); | |
890 | ||
891 | return ret; | |
892 | #else /* CONFIG_HUGETLB_PAGE */ | |
893 | return -EINVAL; | |
894 | #endif | |
895 | } | |
896 | ||
f88a1e90 JG |
897 | static void hmm_pfns_clear(struct hmm_range *range, |
898 | uint64_t *pfns, | |
33cd47dc JG |
899 | unsigned long addr, |
900 | unsigned long end) | |
901 | { | |
902 | for (; addr < end; addr += PAGE_SIZE, pfns++) | |
f88a1e90 | 903 | *pfns = range->values[HMM_PFN_NONE]; |
33cd47dc JG |
904 | } |
905 | ||
da4c3c73 | 906 | /* |
a3e0d41c | 907 | * hmm_range_register() - start tracking change to CPU page table over a range |
25f23a0c | 908 | * @range: range |
a3e0d41c JG |
909 | * @mm: the mm struct for the range of virtual address |
910 | * @start: start virtual address (inclusive) | |
911 | * @end: end virtual address (exclusive) | |
63d5066f | 912 | * @page_shift: expect page shift for the range |
a3e0d41c | 913 | * Returns 0 on success, -EFAULT if the address space is no longer valid |
25f23a0c | 914 | * |
a3e0d41c | 915 | * Track updates to the CPU page table see include/linux/hmm.h |
da4c3c73 | 916 | */ |
a3e0d41c JG |
917 | int hmm_range_register(struct hmm_range *range, |
918 | struct mm_struct *mm, | |
919 | unsigned long start, | |
63d5066f JG |
920 | unsigned long end, |
921 | unsigned page_shift) | |
da4c3c73 | 922 | { |
63d5066f JG |
923 | unsigned long mask = ((1UL << page_shift) - 1UL); |
924 | ||
a3e0d41c | 925 | range->valid = false; |
704f3f2c JG |
926 | range->hmm = NULL; |
927 | ||
63d5066f JG |
928 | if ((start & mask) || (end & mask)) |
929 | return -EINVAL; | |
930 | if (start >= end) | |
da4c3c73 JG |
931 | return -EINVAL; |
932 | ||
63d5066f | 933 | range->page_shift = page_shift; |
a3e0d41c JG |
934 | range->start = start; |
935 | range->end = end; | |
936 | ||
937 | range->hmm = hmm_get_or_create(mm); | |
938 | if (!range->hmm) | |
939 | return -EFAULT; | |
704f3f2c JG |
940 | |
941 | /* Check if hmm_mm_destroy() was call. */ | |
a3e0d41c JG |
942 | if (range->hmm->mm == NULL || range->hmm->dead) { |
943 | hmm_put(range->hmm); | |
944 | return -EFAULT; | |
704f3f2c | 945 | } |
da4c3c73 | 946 | |
a3e0d41c JG |
947 | /* Initialize range to track CPU page table update */ |
948 | mutex_lock(&range->hmm->lock); | |
855ce7d2 | 949 | |
a3e0d41c | 950 | list_add_rcu(&range->list, &range->hmm->ranges); |
86586a41 | 951 | |
704f3f2c | 952 | /* |
a3e0d41c JG |
953 | * If there are any concurrent notifiers we have to wait for them for |
954 | * the range to be valid (see hmm_range_wait_until_valid()). | |
704f3f2c | 955 | */ |
a3e0d41c JG |
956 | if (!range->hmm->notifiers) |
957 | range->valid = true; | |
958 | mutex_unlock(&range->hmm->lock); | |
959 | ||
960 | return 0; | |
da4c3c73 | 961 | } |
a3e0d41c | 962 | EXPORT_SYMBOL(hmm_range_register); |
da4c3c73 JG |
963 | |
964 | /* | |
a3e0d41c JG |
965 | * hmm_range_unregister() - stop tracking change to CPU page table over a range |
966 | * @range: range | |
da4c3c73 JG |
967 | * |
968 | * Range struct is used to track updates to the CPU page table after a call to | |
a3e0d41c | 969 | * hmm_range_register(). See include/linux/hmm.h for how to use it. |
da4c3c73 | 970 | */ |
a3e0d41c | 971 | void hmm_range_unregister(struct hmm_range *range) |
da4c3c73 | 972 | { |
704f3f2c | 973 | /* Sanity check this really should not happen. */ |
a3e0d41c JG |
974 | if (range->hmm == NULL || range->end <= range->start) |
975 | return; | |
da4c3c73 | 976 | |
a3e0d41c | 977 | mutex_lock(&range->hmm->lock); |
da4c3c73 | 978 | list_del_rcu(&range->list); |
a3e0d41c | 979 | mutex_unlock(&range->hmm->lock); |
da4c3c73 | 980 | |
a3e0d41c JG |
981 | /* Drop reference taken by hmm_range_register() */ |
982 | range->valid = false; | |
704f3f2c JG |
983 | hmm_put(range->hmm); |
984 | range->hmm = NULL; | |
da4c3c73 | 985 | } |
a3e0d41c JG |
986 | EXPORT_SYMBOL(hmm_range_unregister); |
987 | ||
988 | /* | |
989 | * hmm_range_snapshot() - snapshot CPU page table for a range | |
990 | * @range: range | |
991 | * Returns: -EINVAL if invalid argument, -ENOMEM out of memory, -EPERM invalid | |
992 | * permission (for instance asking for write and range is read only), | |
993 | * -EAGAIN if you need to retry, -EFAULT invalid (ie either no valid | |
994 | * vma or it is illegal to access that range), number of valid pages | |
995 | * in range->pfns[] (from range start address). | |
996 | * | |
997 | * This snapshots the CPU page table for a range of virtual addresses. Snapshot | |
998 | * validity is tracked by range struct. See in include/linux/hmm.h for example | |
999 | * on how to use. | |
1000 | */ | |
1001 | long hmm_range_snapshot(struct hmm_range *range) | |
1002 | { | |
63d5066f | 1003 | const unsigned long device_vma = VM_IO | VM_PFNMAP | VM_MIXEDMAP; |
a3e0d41c JG |
1004 | unsigned long start = range->start, end; |
1005 | struct hmm_vma_walk hmm_vma_walk; | |
1006 | struct hmm *hmm = range->hmm; | |
1007 | struct vm_area_struct *vma; | |
1008 | struct mm_walk mm_walk; | |
1009 | ||
1010 | /* Check if hmm_mm_destroy() was call. */ | |
1011 | if (hmm->mm == NULL || hmm->dead) | |
1012 | return -EFAULT; | |
1013 | ||
1014 | do { | |
1015 | /* If range is no longer valid force retry. */ | |
1016 | if (!range->valid) | |
1017 | return -EAGAIN; | |
1018 | ||
1019 | vma = find_vma(hmm->mm, start); | |
63d5066f | 1020 | if (vma == NULL || (vma->vm_flags & device_vma)) |
a3e0d41c JG |
1021 | return -EFAULT; |
1022 | ||
63d5066f JG |
1023 | if (is_vm_hugetlb_page(vma)) { |
1024 | struct hstate *h = hstate_vma(vma); | |
1025 | ||
1026 | if (huge_page_shift(h) != range->page_shift && | |
1027 | range->page_shift != PAGE_SHIFT) | |
1028 | return -EINVAL; | |
1029 | } else { | |
1030 | if (range->page_shift != PAGE_SHIFT) | |
1031 | return -EINVAL; | |
1032 | } | |
1033 | ||
a3e0d41c JG |
1034 | if (!(vma->vm_flags & VM_READ)) { |
1035 | /* | |
1036 | * If vma do not allow read access, then assume that it | |
1037 | * does not allow write access, either. HMM does not | |
1038 | * support architecture that allow write without read. | |
1039 | */ | |
1040 | hmm_pfns_clear(range, range->pfns, | |
1041 | range->start, range->end); | |
1042 | return -EPERM; | |
1043 | } | |
1044 | ||
1045 | range->vma = vma; | |
992de9a8 | 1046 | hmm_vma_walk.pgmap = NULL; |
a3e0d41c JG |
1047 | hmm_vma_walk.last = start; |
1048 | hmm_vma_walk.fault = false; | |
1049 | hmm_vma_walk.range = range; | |
1050 | mm_walk.private = &hmm_vma_walk; | |
1051 | end = min(range->end, vma->vm_end); | |
1052 | ||
1053 | mm_walk.vma = vma; | |
1054 | mm_walk.mm = vma->vm_mm; | |
1055 | mm_walk.pte_entry = NULL; | |
1056 | mm_walk.test_walk = NULL; | |
1057 | mm_walk.hugetlb_entry = NULL; | |
992de9a8 | 1058 | mm_walk.pud_entry = hmm_vma_walk_pud; |
a3e0d41c JG |
1059 | mm_walk.pmd_entry = hmm_vma_walk_pmd; |
1060 | mm_walk.pte_hole = hmm_vma_walk_hole; | |
63d5066f | 1061 | mm_walk.hugetlb_entry = hmm_vma_walk_hugetlb_entry; |
a3e0d41c JG |
1062 | |
1063 | walk_page_range(start, end, &mm_walk); | |
1064 | start = end; | |
1065 | } while (start < range->end); | |
1066 | ||
1067 | return (hmm_vma_walk.last - range->start) >> PAGE_SHIFT; | |
1068 | } | |
1069 | EXPORT_SYMBOL(hmm_range_snapshot); | |
74eee180 JG |
1070 | |
1071 | /* | |
73231612 | 1072 | * hmm_range_fault() - try to fault some address in a virtual address range |
08232a45 | 1073 | * @range: range being faulted |
74eee180 | 1074 | * @block: allow blocking on fault (if true it sleeps and do not drop mmap_sem) |
73231612 JG |
1075 | * Returns: number of valid pages in range->pfns[] (from range start |
1076 | * address). This may be zero. If the return value is negative, | |
1077 | * then one of the following values may be returned: | |
1078 | * | |
1079 | * -EINVAL invalid arguments or mm or virtual address are in an | |
63d5066f | 1080 | * invalid vma (for instance device file vma). |
73231612 JG |
1081 | * -ENOMEM: Out of memory. |
1082 | * -EPERM: Invalid permission (for instance asking for write and | |
1083 | * range is read only). | |
1084 | * -EAGAIN: If you need to retry and mmap_sem was drop. This can only | |
1085 | * happens if block argument is false. | |
1086 | * -EBUSY: If the the range is being invalidated and you should wait | |
1087 | * for invalidation to finish. | |
1088 | * -EFAULT: Invalid (ie either no valid vma or it is illegal to access | |
1089 | * that range), number of valid pages in range->pfns[] (from | |
1090 | * range start address). | |
74eee180 JG |
1091 | * |
1092 | * This is similar to a regular CPU page fault except that it will not trigger | |
73231612 JG |
1093 | * any memory migration if the memory being faulted is not accessible by CPUs |
1094 | * and caller does not ask for migration. | |
74eee180 | 1095 | * |
ff05c0c6 JG |
1096 | * On error, for one virtual address in the range, the function will mark the |
1097 | * corresponding HMM pfn entry with an error flag. | |
74eee180 | 1098 | */ |
73231612 | 1099 | long hmm_range_fault(struct hmm_range *range, bool block) |
74eee180 | 1100 | { |
63d5066f | 1101 | const unsigned long device_vma = VM_IO | VM_PFNMAP | VM_MIXEDMAP; |
a3e0d41c | 1102 | unsigned long start = range->start, end; |
74eee180 | 1103 | struct hmm_vma_walk hmm_vma_walk; |
a3e0d41c JG |
1104 | struct hmm *hmm = range->hmm; |
1105 | struct vm_area_struct *vma; | |
74eee180 | 1106 | struct mm_walk mm_walk; |
74eee180 JG |
1107 | int ret; |
1108 | ||
a3e0d41c JG |
1109 | /* Check if hmm_mm_destroy() was call. */ |
1110 | if (hmm->mm == NULL || hmm->dead) | |
1111 | return -EFAULT; | |
704f3f2c | 1112 | |
a3e0d41c JG |
1113 | do { |
1114 | /* If range is no longer valid force retry. */ | |
1115 | if (!range->valid) { | |
1116 | up_read(&hmm->mm->mmap_sem); | |
1117 | return -EAGAIN; | |
1118 | } | |
74eee180 | 1119 | |
a3e0d41c | 1120 | vma = find_vma(hmm->mm, start); |
63d5066f | 1121 | if (vma == NULL || (vma->vm_flags & device_vma)) |
a3e0d41c | 1122 | return -EFAULT; |
704f3f2c | 1123 | |
63d5066f JG |
1124 | if (is_vm_hugetlb_page(vma)) { |
1125 | if (huge_page_shift(hstate_vma(vma)) != | |
1126 | range->page_shift && | |
1127 | range->page_shift != PAGE_SHIFT) | |
1128 | return -EINVAL; | |
1129 | } else { | |
1130 | if (range->page_shift != PAGE_SHIFT) | |
1131 | return -EINVAL; | |
1132 | } | |
1133 | ||
a3e0d41c JG |
1134 | if (!(vma->vm_flags & VM_READ)) { |
1135 | /* | |
1136 | * If vma do not allow read access, then assume that it | |
1137 | * does not allow write access, either. HMM does not | |
1138 | * support architecture that allow write without read. | |
1139 | */ | |
1140 | hmm_pfns_clear(range, range->pfns, | |
1141 | range->start, range->end); | |
1142 | return -EPERM; | |
1143 | } | |
74eee180 | 1144 | |
a3e0d41c | 1145 | range->vma = vma; |
992de9a8 | 1146 | hmm_vma_walk.pgmap = NULL; |
a3e0d41c JG |
1147 | hmm_vma_walk.last = start; |
1148 | hmm_vma_walk.fault = true; | |
1149 | hmm_vma_walk.block = block; | |
1150 | hmm_vma_walk.range = range; | |
1151 | mm_walk.private = &hmm_vma_walk; | |
1152 | end = min(range->end, vma->vm_end); | |
1153 | ||
1154 | mm_walk.vma = vma; | |
1155 | mm_walk.mm = vma->vm_mm; | |
1156 | mm_walk.pte_entry = NULL; | |
1157 | mm_walk.test_walk = NULL; | |
1158 | mm_walk.hugetlb_entry = NULL; | |
992de9a8 | 1159 | mm_walk.pud_entry = hmm_vma_walk_pud; |
a3e0d41c JG |
1160 | mm_walk.pmd_entry = hmm_vma_walk_pmd; |
1161 | mm_walk.pte_hole = hmm_vma_walk_hole; | |
63d5066f | 1162 | mm_walk.hugetlb_entry = hmm_vma_walk_hugetlb_entry; |
a3e0d41c JG |
1163 | |
1164 | do { | |
1165 | ret = walk_page_range(start, end, &mm_walk); | |
1166 | start = hmm_vma_walk.last; | |
1167 | ||
1168 | /* Keep trying while the range is valid. */ | |
1169 | } while (ret == -EBUSY && range->valid); | |
1170 | ||
1171 | if (ret) { | |
1172 | unsigned long i; | |
1173 | ||
1174 | i = (hmm_vma_walk.last - range->start) >> PAGE_SHIFT; | |
1175 | hmm_pfns_clear(range, &range->pfns[i], | |
1176 | hmm_vma_walk.last, range->end); | |
1177 | return ret; | |
1178 | } | |
1179 | start = end; | |
74eee180 | 1180 | |
a3e0d41c | 1181 | } while (start < range->end); |
704f3f2c | 1182 | |
73231612 | 1183 | return (hmm_vma_walk.last - range->start) >> PAGE_SHIFT; |
74eee180 | 1184 | } |
73231612 | 1185 | EXPORT_SYMBOL(hmm_range_fault); |
55c0ece8 JG |
1186 | |
1187 | /** | |
1188 | * hmm_range_dma_map() - hmm_range_fault() and dma map page all in one. | |
1189 | * @range: range being faulted | |
1190 | * @device: device against to dma map page to | |
1191 | * @daddrs: dma address of mapped pages | |
1192 | * @block: allow blocking on fault (if true it sleeps and do not drop mmap_sem) | |
1193 | * Returns: number of pages mapped on success, -EAGAIN if mmap_sem have been | |
1194 | * drop and you need to try again, some other error value otherwise | |
1195 | * | |
1196 | * Note same usage pattern as hmm_range_fault(). | |
1197 | */ | |
1198 | long hmm_range_dma_map(struct hmm_range *range, | |
1199 | struct device *device, | |
1200 | dma_addr_t *daddrs, | |
1201 | bool block) | |
1202 | { | |
1203 | unsigned long i, npages, mapped; | |
1204 | long ret; | |
1205 | ||
1206 | ret = hmm_range_fault(range, block); | |
1207 | if (ret <= 0) | |
1208 | return ret ? ret : -EBUSY; | |
1209 | ||
1210 | npages = (range->end - range->start) >> PAGE_SHIFT; | |
1211 | for (i = 0, mapped = 0; i < npages; ++i) { | |
1212 | enum dma_data_direction dir = DMA_TO_DEVICE; | |
1213 | struct page *page; | |
1214 | ||
1215 | /* | |
1216 | * FIXME need to update DMA API to provide invalid DMA address | |
1217 | * value instead of a function to test dma address value. This | |
1218 | * would remove lot of dumb code duplicated accross many arch. | |
1219 | * | |
1220 | * For now setting it to 0 here is good enough as the pfns[] | |
1221 | * value is what is use to check what is valid and what isn't. | |
1222 | */ | |
1223 | daddrs[i] = 0; | |
1224 | ||
1225 | page = hmm_pfn_to_page(range, range->pfns[i]); | |
1226 | if (page == NULL) | |
1227 | continue; | |
1228 | ||
1229 | /* Check if range is being invalidated */ | |
1230 | if (!range->valid) { | |
1231 | ret = -EBUSY; | |
1232 | goto unmap; | |
1233 | } | |
1234 | ||
1235 | /* If it is read and write than map bi-directional. */ | |
1236 | if (range->pfns[i] & range->flags[HMM_PFN_WRITE]) | |
1237 | dir = DMA_BIDIRECTIONAL; | |
1238 | ||
1239 | daddrs[i] = dma_map_page(device, page, 0, PAGE_SIZE, dir); | |
1240 | if (dma_mapping_error(device, daddrs[i])) { | |
1241 | ret = -EFAULT; | |
1242 | goto unmap; | |
1243 | } | |
1244 | ||
1245 | mapped++; | |
1246 | } | |
1247 | ||
1248 | return mapped; | |
1249 | ||
1250 | unmap: | |
1251 | for (npages = i, i = 0; (i < npages) && mapped; ++i) { | |
1252 | enum dma_data_direction dir = DMA_TO_DEVICE; | |
1253 | struct page *page; | |
1254 | ||
1255 | page = hmm_pfn_to_page(range, range->pfns[i]); | |
1256 | if (page == NULL) | |
1257 | continue; | |
1258 | ||
1259 | if (dma_mapping_error(device, daddrs[i])) | |
1260 | continue; | |
1261 | ||
1262 | /* If it is read and write than map bi-directional. */ | |
1263 | if (range->pfns[i] & range->flags[HMM_PFN_WRITE]) | |
1264 | dir = DMA_BIDIRECTIONAL; | |
1265 | ||
1266 | dma_unmap_page(device, daddrs[i], PAGE_SIZE, dir); | |
1267 | mapped--; | |
1268 | } | |
1269 | ||
1270 | return ret; | |
1271 | } | |
1272 | EXPORT_SYMBOL(hmm_range_dma_map); | |
1273 | ||
1274 | /** | |
1275 | * hmm_range_dma_unmap() - unmap range of that was map with hmm_range_dma_map() | |
1276 | * @range: range being unmapped | |
1277 | * @vma: the vma against which the range (optional) | |
1278 | * @device: device against which dma map was done | |
1279 | * @daddrs: dma address of mapped pages | |
1280 | * @dirty: dirty page if it had the write flag set | |
1281 | * Returns: number of page unmapped on success, -EINVAL otherwise | |
1282 | * | |
1283 | * Note that caller MUST abide by mmu notifier or use HMM mirror and abide | |
1284 | * to the sync_cpu_device_pagetables() callback so that it is safe here to | |
1285 | * call set_page_dirty(). Caller must also take appropriate locks to avoid | |
1286 | * concurrent mmu notifier or sync_cpu_device_pagetables() to make progress. | |
1287 | */ | |
1288 | long hmm_range_dma_unmap(struct hmm_range *range, | |
1289 | struct vm_area_struct *vma, | |
1290 | struct device *device, | |
1291 | dma_addr_t *daddrs, | |
1292 | bool dirty) | |
1293 | { | |
1294 | unsigned long i, npages; | |
1295 | long cpages = 0; | |
1296 | ||
1297 | /* Sanity check. */ | |
1298 | if (range->end <= range->start) | |
1299 | return -EINVAL; | |
1300 | if (!daddrs) | |
1301 | return -EINVAL; | |
1302 | if (!range->pfns) | |
1303 | return -EINVAL; | |
1304 | ||
1305 | npages = (range->end - range->start) >> PAGE_SHIFT; | |
1306 | for (i = 0; i < npages; ++i) { | |
1307 | enum dma_data_direction dir = DMA_TO_DEVICE; | |
1308 | struct page *page; | |
1309 | ||
1310 | page = hmm_pfn_to_page(range, range->pfns[i]); | |
1311 | if (page == NULL) | |
1312 | continue; | |
1313 | ||
1314 | /* If it is read and write than map bi-directional. */ | |
1315 | if (range->pfns[i] & range->flags[HMM_PFN_WRITE]) { | |
1316 | dir = DMA_BIDIRECTIONAL; | |
1317 | ||
1318 | /* | |
1319 | * See comments in function description on why it is | |
1320 | * safe here to call set_page_dirty() | |
1321 | */ | |
1322 | if (dirty) | |
1323 | set_page_dirty(page); | |
1324 | } | |
1325 | ||
1326 | /* Unmap and clear pfns/dma address */ | |
1327 | dma_unmap_page(device, daddrs[i], PAGE_SIZE, dir); | |
1328 | range->pfns[i] = range->values[HMM_PFN_NONE]; | |
1329 | /* FIXME see comments in hmm_vma_dma_map() */ | |
1330 | daddrs[i] = 0; | |
1331 | cpages++; | |
1332 | } | |
1333 | ||
1334 | return cpages; | |
1335 | } | |
1336 | EXPORT_SYMBOL(hmm_range_dma_unmap); | |
c0b12405 | 1337 | #endif /* IS_ENABLED(CONFIG_HMM_MIRROR) */ |
4ef589dc JG |
1338 | |
1339 | ||
df6ad698 | 1340 | #if IS_ENABLED(CONFIG_DEVICE_PRIVATE) || IS_ENABLED(CONFIG_DEVICE_PUBLIC) |
4ef589dc JG |
1341 | struct page *hmm_vma_alloc_locked_page(struct vm_area_struct *vma, |
1342 | unsigned long addr) | |
1343 | { | |
1344 | struct page *page; | |
1345 | ||
1346 | page = alloc_page_vma(GFP_HIGHUSER, vma, addr); | |
1347 | if (!page) | |
1348 | return NULL; | |
1349 | lock_page(page); | |
1350 | return page; | |
1351 | } | |
1352 | EXPORT_SYMBOL(hmm_vma_alloc_locked_page); | |
1353 | ||
1354 | ||
1355 | static void hmm_devmem_ref_release(struct percpu_ref *ref) | |
1356 | { | |
1357 | struct hmm_devmem *devmem; | |
1358 | ||
1359 | devmem = container_of(ref, struct hmm_devmem, ref); | |
1360 | complete(&devmem->completion); | |
1361 | } | |
1362 | ||
1363 | static void hmm_devmem_ref_exit(void *data) | |
1364 | { | |
1365 | struct percpu_ref *ref = data; | |
1366 | struct hmm_devmem *devmem; | |
1367 | ||
1368 | devmem = container_of(ref, struct hmm_devmem, ref); | |
bbecd94e | 1369 | wait_for_completion(&devmem->completion); |
4ef589dc | 1370 | percpu_ref_exit(ref); |
4ef589dc JG |
1371 | } |
1372 | ||
bbecd94e | 1373 | static void hmm_devmem_ref_kill(struct percpu_ref *ref) |
4ef589dc | 1374 | { |
4ef589dc | 1375 | percpu_ref_kill(ref); |
4ef589dc JG |
1376 | } |
1377 | ||
b57e622e | 1378 | static vm_fault_t hmm_devmem_fault(struct vm_area_struct *vma, |
4ef589dc JG |
1379 | unsigned long addr, |
1380 | const struct page *page, | |
1381 | unsigned int flags, | |
1382 | pmd_t *pmdp) | |
1383 | { | |
1384 | struct hmm_devmem *devmem = page->pgmap->data; | |
1385 | ||
1386 | return devmem->ops->fault(devmem, vma, addr, page, flags, pmdp); | |
1387 | } | |
1388 | ||
1389 | static void hmm_devmem_free(struct page *page, void *data) | |
1390 | { | |
1391 | struct hmm_devmem *devmem = data; | |
1392 | ||
2fa147bd DW |
1393 | page->mapping = NULL; |
1394 | ||
4ef589dc JG |
1395 | devmem->ops->free(devmem, page); |
1396 | } | |
1397 | ||
4ef589dc JG |
1398 | /* |
1399 | * hmm_devmem_add() - hotplug ZONE_DEVICE memory for device memory | |
1400 | * | |
1401 | * @ops: memory event device driver callback (see struct hmm_devmem_ops) | |
1402 | * @device: device struct to bind the resource too | |
1403 | * @size: size in bytes of the device memory to add | |
1404 | * Returns: pointer to new hmm_devmem struct ERR_PTR otherwise | |
1405 | * | |
1406 | * This function first finds an empty range of physical address big enough to | |
1407 | * contain the new resource, and then hotplugs it as ZONE_DEVICE memory, which | |
1408 | * in turn allocates struct pages. It does not do anything beyond that; all | |
1409 | * events affecting the memory will go through the various callbacks provided | |
1410 | * by hmm_devmem_ops struct. | |
1411 | * | |
1412 | * Device driver should call this function during device initialization and | |
1413 | * is then responsible of memory management. HMM only provides helpers. | |
1414 | */ | |
1415 | struct hmm_devmem *hmm_devmem_add(const struct hmm_devmem_ops *ops, | |
1416 | struct device *device, | |
1417 | unsigned long size) | |
1418 | { | |
1419 | struct hmm_devmem *devmem; | |
1420 | resource_size_t addr; | |
bbecd94e | 1421 | void *result; |
4ef589dc JG |
1422 | int ret; |
1423 | ||
e7638488 | 1424 | dev_pagemap_get_ops(); |
4ef589dc | 1425 | |
58ef15b7 | 1426 | devmem = devm_kzalloc(device, sizeof(*devmem), GFP_KERNEL); |
4ef589dc JG |
1427 | if (!devmem) |
1428 | return ERR_PTR(-ENOMEM); | |
1429 | ||
1430 | init_completion(&devmem->completion); | |
1431 | devmem->pfn_first = -1UL; | |
1432 | devmem->pfn_last = -1UL; | |
1433 | devmem->resource = NULL; | |
1434 | devmem->device = device; | |
1435 | devmem->ops = ops; | |
1436 | ||
1437 | ret = percpu_ref_init(&devmem->ref, &hmm_devmem_ref_release, | |
1438 | 0, GFP_KERNEL); | |
1439 | if (ret) | |
58ef15b7 | 1440 | return ERR_PTR(ret); |
4ef589dc | 1441 | |
58ef15b7 | 1442 | ret = devm_add_action_or_reset(device, hmm_devmem_ref_exit, &devmem->ref); |
4ef589dc | 1443 | if (ret) |
58ef15b7 | 1444 | return ERR_PTR(ret); |
4ef589dc JG |
1445 | |
1446 | size = ALIGN(size, PA_SECTION_SIZE); | |
1447 | addr = min((unsigned long)iomem_resource.end, | |
1448 | (1UL << MAX_PHYSMEM_BITS) - 1); | |
1449 | addr = addr - size + 1UL; | |
1450 | ||
1451 | /* | |
1452 | * FIXME add a new helper to quickly walk resource tree and find free | |
1453 | * range | |
1454 | * | |
1455 | * FIXME what about ioport_resource resource ? | |
1456 | */ | |
1457 | for (; addr > size && addr >= iomem_resource.start; addr -= size) { | |
1458 | ret = region_intersects(addr, size, 0, IORES_DESC_NONE); | |
1459 | if (ret != REGION_DISJOINT) | |
1460 | continue; | |
1461 | ||
1462 | devmem->resource = devm_request_mem_region(device, addr, size, | |
1463 | dev_name(device)); | |
58ef15b7 DW |
1464 | if (!devmem->resource) |
1465 | return ERR_PTR(-ENOMEM); | |
4ef589dc JG |
1466 | break; |
1467 | } | |
58ef15b7 DW |
1468 | if (!devmem->resource) |
1469 | return ERR_PTR(-ERANGE); | |
4ef589dc JG |
1470 | |
1471 | devmem->resource->desc = IORES_DESC_DEVICE_PRIVATE_MEMORY; | |
1472 | devmem->pfn_first = devmem->resource->start >> PAGE_SHIFT; | |
1473 | devmem->pfn_last = devmem->pfn_first + | |
1474 | (resource_size(devmem->resource) >> PAGE_SHIFT); | |
063a7d1d | 1475 | devmem->page_fault = hmm_devmem_fault; |
4ef589dc | 1476 | |
bbecd94e DW |
1477 | devmem->pagemap.type = MEMORY_DEVICE_PRIVATE; |
1478 | devmem->pagemap.res = *devmem->resource; | |
bbecd94e DW |
1479 | devmem->pagemap.page_free = hmm_devmem_free; |
1480 | devmem->pagemap.altmap_valid = false; | |
1481 | devmem->pagemap.ref = &devmem->ref; | |
1482 | devmem->pagemap.data = devmem; | |
1483 | devmem->pagemap.kill = hmm_devmem_ref_kill; | |
4ef589dc | 1484 | |
bbecd94e DW |
1485 | result = devm_memremap_pages(devmem->device, &devmem->pagemap); |
1486 | if (IS_ERR(result)) | |
1487 | return result; | |
4ef589dc | 1488 | return devmem; |
4ef589dc | 1489 | } |
02917e9f | 1490 | EXPORT_SYMBOL_GPL(hmm_devmem_add); |
4ef589dc | 1491 | |
d3df0a42 JG |
1492 | struct hmm_devmem *hmm_devmem_add_resource(const struct hmm_devmem_ops *ops, |
1493 | struct device *device, | |
1494 | struct resource *res) | |
1495 | { | |
1496 | struct hmm_devmem *devmem; | |
bbecd94e | 1497 | void *result; |
d3df0a42 JG |
1498 | int ret; |
1499 | ||
1500 | if (res->desc != IORES_DESC_DEVICE_PUBLIC_MEMORY) | |
1501 | return ERR_PTR(-EINVAL); | |
1502 | ||
e7638488 | 1503 | dev_pagemap_get_ops(); |
d3df0a42 | 1504 | |
58ef15b7 | 1505 | devmem = devm_kzalloc(device, sizeof(*devmem), GFP_KERNEL); |
d3df0a42 JG |
1506 | if (!devmem) |
1507 | return ERR_PTR(-ENOMEM); | |
1508 | ||
1509 | init_completion(&devmem->completion); | |
1510 | devmem->pfn_first = -1UL; | |
1511 | devmem->pfn_last = -1UL; | |
1512 | devmem->resource = res; | |
1513 | devmem->device = device; | |
1514 | devmem->ops = ops; | |
1515 | ||
1516 | ret = percpu_ref_init(&devmem->ref, &hmm_devmem_ref_release, | |
1517 | 0, GFP_KERNEL); | |
1518 | if (ret) | |
58ef15b7 | 1519 | return ERR_PTR(ret); |
d3df0a42 | 1520 | |
58ef15b7 DW |
1521 | ret = devm_add_action_or_reset(device, hmm_devmem_ref_exit, |
1522 | &devmem->ref); | |
d3df0a42 | 1523 | if (ret) |
58ef15b7 | 1524 | return ERR_PTR(ret); |
d3df0a42 JG |
1525 | |
1526 | devmem->pfn_first = devmem->resource->start >> PAGE_SHIFT; | |
1527 | devmem->pfn_last = devmem->pfn_first + | |
1528 | (resource_size(devmem->resource) >> PAGE_SHIFT); | |
063a7d1d | 1529 | devmem->page_fault = hmm_devmem_fault; |
d3df0a42 | 1530 | |
bbecd94e DW |
1531 | devmem->pagemap.type = MEMORY_DEVICE_PUBLIC; |
1532 | devmem->pagemap.res = *devmem->resource; | |
bbecd94e DW |
1533 | devmem->pagemap.page_free = hmm_devmem_free; |
1534 | devmem->pagemap.altmap_valid = false; | |
1535 | devmem->pagemap.ref = &devmem->ref; | |
1536 | devmem->pagemap.data = devmem; | |
1537 | devmem->pagemap.kill = hmm_devmem_ref_kill; | |
d3df0a42 | 1538 | |
bbecd94e DW |
1539 | result = devm_memremap_pages(devmem->device, &devmem->pagemap); |
1540 | if (IS_ERR(result)) | |
1541 | return result; | |
d3df0a42 | 1542 | return devmem; |
d3df0a42 | 1543 | } |
02917e9f | 1544 | EXPORT_SYMBOL_GPL(hmm_devmem_add_resource); |
d3df0a42 | 1545 | |
858b54da JG |
1546 | /* |
1547 | * A device driver that wants to handle multiple devices memory through a | |
1548 | * single fake device can use hmm_device to do so. This is purely a helper | |
1549 | * and it is not needed to make use of any HMM functionality. | |
1550 | */ | |
1551 | #define HMM_DEVICE_MAX 256 | |
1552 | ||
1553 | static DECLARE_BITMAP(hmm_device_mask, HMM_DEVICE_MAX); | |
1554 | static DEFINE_SPINLOCK(hmm_device_lock); | |
1555 | static struct class *hmm_device_class; | |
1556 | static dev_t hmm_device_devt; | |
1557 | ||
1558 | static void hmm_device_release(struct device *device) | |
1559 | { | |
1560 | struct hmm_device *hmm_device; | |
1561 | ||
1562 | hmm_device = container_of(device, struct hmm_device, device); | |
1563 | spin_lock(&hmm_device_lock); | |
1564 | clear_bit(hmm_device->minor, hmm_device_mask); | |
1565 | spin_unlock(&hmm_device_lock); | |
1566 | ||
1567 | kfree(hmm_device); | |
1568 | } | |
1569 | ||
1570 | struct hmm_device *hmm_device_new(void *drvdata) | |
1571 | { | |
1572 | struct hmm_device *hmm_device; | |
1573 | ||
1574 | hmm_device = kzalloc(sizeof(*hmm_device), GFP_KERNEL); | |
1575 | if (!hmm_device) | |
1576 | return ERR_PTR(-ENOMEM); | |
1577 | ||
1578 | spin_lock(&hmm_device_lock); | |
1579 | hmm_device->minor = find_first_zero_bit(hmm_device_mask, HMM_DEVICE_MAX); | |
1580 | if (hmm_device->minor >= HMM_DEVICE_MAX) { | |
1581 | spin_unlock(&hmm_device_lock); | |
1582 | kfree(hmm_device); | |
1583 | return ERR_PTR(-EBUSY); | |
1584 | } | |
1585 | set_bit(hmm_device->minor, hmm_device_mask); | |
1586 | spin_unlock(&hmm_device_lock); | |
1587 | ||
1588 | dev_set_name(&hmm_device->device, "hmm_device%d", hmm_device->minor); | |
1589 | hmm_device->device.devt = MKDEV(MAJOR(hmm_device_devt), | |
1590 | hmm_device->minor); | |
1591 | hmm_device->device.release = hmm_device_release; | |
1592 | dev_set_drvdata(&hmm_device->device, drvdata); | |
1593 | hmm_device->device.class = hmm_device_class; | |
1594 | device_initialize(&hmm_device->device); | |
1595 | ||
1596 | return hmm_device; | |
1597 | } | |
1598 | EXPORT_SYMBOL(hmm_device_new); | |
1599 | ||
1600 | void hmm_device_put(struct hmm_device *hmm_device) | |
1601 | { | |
1602 | put_device(&hmm_device->device); | |
1603 | } | |
1604 | EXPORT_SYMBOL(hmm_device_put); | |
1605 | ||
1606 | static int __init hmm_init(void) | |
1607 | { | |
1608 | int ret; | |
1609 | ||
1610 | ret = alloc_chrdev_region(&hmm_device_devt, 0, | |
1611 | HMM_DEVICE_MAX, | |
1612 | "hmm_device"); | |
1613 | if (ret) | |
1614 | return ret; | |
1615 | ||
1616 | hmm_device_class = class_create(THIS_MODULE, "hmm_device"); | |
1617 | if (IS_ERR(hmm_device_class)) { | |
1618 | unregister_chrdev_region(hmm_device_devt, HMM_DEVICE_MAX); | |
1619 | return PTR_ERR(hmm_device_class); | |
1620 | } | |
1621 | return 0; | |
1622 | } | |
1623 | ||
1624 | device_initcall(hmm_init); | |
df6ad698 | 1625 | #endif /* CONFIG_DEVICE_PRIVATE || CONFIG_DEVICE_PUBLIC */ |