1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3 * Copyright 2013 Red Hat Inc.
4 *
5 * Authors: Jérôme Glisse <jglisse@redhat.com>
6 */
7 /*
8 * Refer to include/linux/hmm.h for information about heterogeneous memory
9 * management or HMM for short.
10 */
11 #include <linux/pagewalk.h>
12 #include <linux/hmm.h>
13 #include <linux/init.h>
14 #include <linux/rmap.h>
15 #include <linux/swap.h>
16 #include <linux/slab.h>
17 #include <linux/sched.h>
18 #include <linux/mmzone.h>
19 #include <linux/pagemap.h>
20 #include <linux/swapops.h>
21 #include <linux/hugetlb.h>
22 #include <linux/memremap.h>
23 #include <linux/sched/mm.h>
24 #include <linux/jump_label.h>
25 #include <linux/dma-mapping.h>
26 #include <linux/mmu_notifier.h>
27 #include <linux/memory_hotplug.h>
28
29 #include "internal.h"
30
31 struct hmm_vma_walk {
32 struct hmm_range *range;
33 unsigned long last;
34 };
35
36 enum {
37 HMM_NEED_FAULT = 1 << 0,
38 HMM_NEED_WRITE_FAULT = 1 << 1,
39 HMM_NEED_ALL_BITS = HMM_NEED_FAULT | HMM_NEED_WRITE_FAULT,
40 };
41
42 static int hmm_pfns_fill(unsigned long addr, unsigned long end,
43 struct hmm_range *range, unsigned long cpu_flags)
44 {
45 unsigned long i = (addr - range->start) >> PAGE_SHIFT;
46
47 for (; addr < end; addr += PAGE_SIZE, i++)
48 range->hmm_pfns[i] = cpu_flags;
49 return 0;
50 }
51
52 /*
53 * hmm_vma_fault() - fault in a range lacking valid pmd or pte(s)
54 * @addr: range virtual start address (inclusive)
55 * @end: range virtual end address (exclusive)
56 * @required_fault: HMM_NEED_* flags
57 * @walk: mm_walk structure
58 * Return: -EBUSY after page fault, or page fault error
59 *
60 * This function will be called whenever pmd_none() or pte_none() returns true,
61 * or whenever there is no page directory covering the virtual address range.
62 */
63 static int hmm_vma_fault(unsigned long addr, unsigned long end,
64 unsigned int required_fault, struct mm_walk *walk)
65 {
66 struct hmm_vma_walk *hmm_vma_walk = walk->private;
67 struct vm_area_struct *vma = walk->vma;
68 unsigned int fault_flags = FAULT_FLAG_REMOTE;
69
70 WARN_ON_ONCE(!required_fault);
71 hmm_vma_walk->last = addr;
72
73 if (required_fault & HMM_NEED_WRITE_FAULT) {
74 if (!(vma->vm_flags & VM_WRITE))
75 return -EPERM;
76 fault_flags |= FAULT_FLAG_WRITE;
77 }
78
79 for (; addr < end; addr += PAGE_SIZE)
80 if (handle_mm_fault(vma, addr, fault_flags, NULL) &
81 VM_FAULT_ERROR)
82 return -EFAULT;
83 return -EBUSY;
84 }
85
86 static unsigned int hmm_pte_need_fault(const struct hmm_vma_walk *hmm_vma_walk,
87 unsigned long pfn_req_flags,
88 unsigned long cpu_flags)
89 {
90 struct hmm_range *range = hmm_vma_walk->range;
91
92 /*
93 * So we not only consider the individual per page request we also
94 * consider the default flags requested for the range. The API can
95 * be used 2 ways. The first one where the HMM user coalesces
96 * multiple page faults into one request and sets flags per pfn for
97 * those faults. The second one where the HMM user wants to pre-
98 * fault a range with specific flags. For the latter one it is a
99 * waste to have the user pre-fill the pfn arrays with a default
100 * flags value.
101 */
102 pfn_req_flags &= range->pfn_flags_mask;
103 pfn_req_flags |= range->default_flags;
104
105 /* We aren't ask to do anything ... */
106 if (!(pfn_req_flags & HMM_PFN_REQ_FAULT))
107 return 0;
108
109 /* Need to write fault ? */
110 if ((pfn_req_flags & HMM_PFN_REQ_WRITE) &&
111 !(cpu_flags & HMM_PFN_WRITE))
112 return HMM_NEED_FAULT | HMM_NEED_WRITE_FAULT;
113
114 /* If CPU page table is not valid then we need to fault */
115 if (!(cpu_flags & HMM_PFN_VALID))
116 return HMM_NEED_FAULT;
117 return 0;
118 }
119
120 static unsigned int
121 hmm_range_need_fault(const struct hmm_vma_walk *hmm_vma_walk,
122 const unsigned long hmm_pfns[], unsigned long npages,
123 unsigned long cpu_flags)
124 {
125 struct hmm_range *range = hmm_vma_walk->range;
126 unsigned int required_fault = 0;
127 unsigned long i;
128
129 /*
130 * If the default flags do not request to fault pages, and the mask does
131 * not allow for individual pages to be faulted, then
132 * hmm_pte_need_fault() will always return 0.
133 */
134 if (!((range->default_flags | range->pfn_flags_mask) &
135 HMM_PFN_REQ_FAULT))
136 return 0;
137
138 for (i = 0; i < npages; ++i) {
139 required_fault |= hmm_pte_need_fault(hmm_vma_walk, hmm_pfns[i],
140 cpu_flags);
141 if (required_fault == HMM_NEED_ALL_BITS)
142 return required_fault;
143 }
144 return required_fault;
145 }
146
147 static int hmm_vma_walk_hole(unsigned long addr, unsigned long end,
148 __always_unused int depth, struct mm_walk *walk)
149 {
150 struct hmm_vma_walk *hmm_vma_walk = walk->private;
151 struct hmm_range *range = hmm_vma_walk->range;
152 unsigned int required_fault;
153 unsigned long i, npages;
154 unsigned long *hmm_pfns;
155
156 i = (addr - range->start) >> PAGE_SHIFT;
157 npages = (end - addr) >> PAGE_SHIFT;
158 hmm_pfns = &range->hmm_pfns[i];
159 required_fault =
160 hmm_range_need_fault(hmm_vma_walk, hmm_pfns, npages, 0);
161 if (!walk->vma) {
162 if (required_fault)
163 return -EFAULT;
164 return hmm_pfns_fill(addr, end, range, HMM_PFN_ERROR);
165 }
166 if (required_fault)
167 return hmm_vma_fault(addr, end, required_fault, walk);
168 return hmm_pfns_fill(addr, end, range, 0);
169 }
170
171 static inline unsigned long hmm_pfn_flags_order(unsigned long order)
172 {
173 return order << HMM_PFN_ORDER_SHIFT;
174 }
175
176 static inline unsigned long pmd_to_hmm_pfn_flags(struct hmm_range *range,
177 pmd_t pmd)
178 {
179 if (pmd_protnone(pmd))
180 return 0;
181 return (pmd_write(pmd) ? (HMM_PFN_VALID | HMM_PFN_WRITE) :
182 HMM_PFN_VALID) |
183 hmm_pfn_flags_order(PMD_SHIFT - PAGE_SHIFT);
184 }
185
186 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
187 static int hmm_vma_handle_pmd(struct mm_walk *walk, unsigned long addr,
188 unsigned long end, unsigned long hmm_pfns[],
189 pmd_t pmd)
190 {
191 struct hmm_vma_walk *hmm_vma_walk = walk->private;
192 struct hmm_range *range = hmm_vma_walk->range;
193 unsigned long pfn, npages, i;
194 unsigned int required_fault;
195 unsigned long cpu_flags;
196
197 npages = (end - addr) >> PAGE_SHIFT;
198 cpu_flags = pmd_to_hmm_pfn_flags(range, pmd);
199 required_fault =
200 hmm_range_need_fault(hmm_vma_walk, hmm_pfns, npages, cpu_flags);
201 if (required_fault)
202 return hmm_vma_fault(addr, end, required_fault, walk);
203
204 pfn = pmd_pfn(pmd) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
205 for (i = 0; addr < end; addr += PAGE_SIZE, i++, pfn++)
206 hmm_pfns[i] = pfn | cpu_flags;
207 return 0;
208 }
209 #else /* CONFIG_TRANSPARENT_HUGEPAGE */
210 /* stub to allow the code below to compile */
211 int hmm_vma_handle_pmd(struct mm_walk *walk, unsigned long addr,
212 unsigned long end, unsigned long hmm_pfns[], pmd_t pmd);
213 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
214
215 static inline unsigned long pte_to_hmm_pfn_flags(struct hmm_range *range,
216 pte_t pte)
217 {
218 if (pte_none(pte) || !pte_present(pte) || pte_protnone(pte))
219 return 0;
220 return pte_write(pte) ? (HMM_PFN_VALID | HMM_PFN_WRITE) : HMM_PFN_VALID;
221 }
222
223 static int hmm_vma_handle_pte(struct mm_walk *walk, unsigned long addr,
224 unsigned long end, pmd_t *pmdp, pte_t *ptep,
225 unsigned long *hmm_pfn)
226 {
227 struct hmm_vma_walk *hmm_vma_walk = walk->private;
228 struct hmm_range *range = hmm_vma_walk->range;
229 unsigned int required_fault;
230 unsigned long cpu_flags;
231 pte_t pte = ptep_get(ptep);
232 uint64_t pfn_req_flags = *hmm_pfn;
233
234 if (pte_none_mostly(pte)) {
235 required_fault =
236 hmm_pte_need_fault(hmm_vma_walk, pfn_req_flags, 0);
237 if (required_fault)
238 goto fault;
239 *hmm_pfn = 0;
240 return 0;
241 }
242
243 if (!pte_present(pte)) {
244 swp_entry_t entry = pte_to_swp_entry(pte);
245
246 /*
247 * Don't fault in device private pages owned by the caller,
248 * just report the PFN.
249 */
250 if (is_device_private_entry(entry) &&
251 pfn_swap_entry_to_page(entry)->pgmap->owner ==
252 range->dev_private_owner) {
253 cpu_flags = HMM_PFN_VALID;
254 if (is_writable_device_private_entry(entry))
255 cpu_flags |= HMM_PFN_WRITE;
256 *hmm_pfn = swp_offset_pfn(entry) | cpu_flags;
257 return 0;
258 }
259
260 required_fault =
261 hmm_pte_need_fault(hmm_vma_walk, pfn_req_flags, 0);
262 if (!required_fault) {
263 *hmm_pfn = 0;
264 return 0;
265 }
266
267 if (!non_swap_entry(entry))
268 goto fault;
269
270 if (is_device_private_entry(entry))
271 goto fault;
272
273 if (is_device_exclusive_entry(entry))
274 goto fault;
275
276 if (is_migration_entry(entry)) {
277 pte_unmap(ptep);
278 hmm_vma_walk->last = addr;
279 migration_entry_wait(walk->mm, pmdp, addr);
280 return -EBUSY;
281 }
282
283 /* Report error for everything else */
284 pte_unmap(ptep);
285 return -EFAULT;
286 }
287
288 cpu_flags = pte_to_hmm_pfn_flags(range, pte);
289 required_fault =
290 hmm_pte_need_fault(hmm_vma_walk, pfn_req_flags, cpu_flags);
291 if (required_fault)
292 goto fault;
293
294 /*
295 * Bypass devmap pte such as DAX page when all pfn requested
296 * flags(pfn_req_flags) are fulfilled.
297 * Since each architecture defines a struct page for the zero page, just
298 * fall through and treat it like a normal page.
299 */
300 if (!vm_normal_page(walk->vma, addr, pte) &&
301 !pte_devmap(pte) &&
302 !is_zero_pfn(pte_pfn(pte))) {
303 if (hmm_pte_need_fault(hmm_vma_walk, pfn_req_flags, 0)) {
304 pte_unmap(ptep);
305 return -EFAULT;
306 }
307 *hmm_pfn = HMM_PFN_ERROR;
308 return 0;
309 }
310
311 *hmm_pfn = pte_pfn(pte) | cpu_flags;
312 return 0;
313
314 fault:
315 pte_unmap(ptep);
316 /* Fault any virtual address we were asked to fault */
317 return hmm_vma_fault(addr, end, required_fault, walk);
318 }
319
320 static int hmm_vma_walk_pmd(pmd_t *pmdp,
321 unsigned long start,
322 unsigned long end,
323 struct mm_walk *walk)
324 {
325 struct hmm_vma_walk *hmm_vma_walk = walk->private;
326 struct hmm_range *range = hmm_vma_walk->range;
327 unsigned long *hmm_pfns =
328 &range->hmm_pfns[(start - range->start) >> PAGE_SHIFT];
329 unsigned long npages = (end - start) >> PAGE_SHIFT;
330 unsigned long addr = start;
331 pte_t *ptep;
332 pmd_t pmd;
333
334 again:
335 pmd = pmdp_get_lockless(pmdp);
336 if (pmd_none(pmd))
337 return hmm_vma_walk_hole(start, end, -1, walk);
338
339 if (thp_migration_supported() && is_pmd_migration_entry(pmd)) {
340 if (hmm_range_need_fault(hmm_vma_walk, hmm_pfns, npages, 0)) {
341 hmm_vma_walk->last = addr;
342 pmd_migration_entry_wait(walk->mm, pmdp);
343 return -EBUSY;
344 }
345 return hmm_pfns_fill(start, end, range, 0);
346 }
347
348 if (!pmd_present(pmd)) {
349 if (hmm_range_need_fault(hmm_vma_walk, hmm_pfns, npages, 0))
350 return -EFAULT;
351 return hmm_pfns_fill(start, end, range, HMM_PFN_ERROR);
352 }
353
354 if (pmd_devmap(pmd) || pmd_trans_huge(pmd)) {
355 /*
356 * No need to take pmd_lock here, even if some other thread
357 * is splitting the huge pmd we will get that event through
358 * mmu_notifier callback.
359 *
360 * So just read pmd value and check again it's a transparent
361 * huge or device mapping one and compute corresponding pfn
362 * values.
363 */
364 pmd = pmdp_get_lockless(pmdp);
365 if (!pmd_devmap(pmd) && !pmd_trans_huge(pmd))
366 goto again;
367
368 return hmm_vma_handle_pmd(walk, addr, end, hmm_pfns, pmd);
369 }
370
371 /*
372 * We have handled all the valid cases above ie either none, migration,
373 * huge or transparent huge. At this point either it is a valid pmd
374 * entry pointing to pte directory or it is a bad pmd that will not
375 * recover.
376 */
377 if (pmd_bad(pmd)) {
378 if (hmm_range_need_fault(hmm_vma_walk, hmm_pfns, npages, 0))
379 return -EFAULT;
380 return hmm_pfns_fill(start, end, range, HMM_PFN_ERROR);
381 }
382
383 ptep = pte_offset_map(pmdp, addr);
384 if (!ptep)
385 goto again;
386 for (; addr < end; addr += PAGE_SIZE, ptep++, hmm_pfns++) {
387 int r;
388
389 r = hmm_vma_handle_pte(walk, addr, end, pmdp, ptep, hmm_pfns);
390 if (r) {
391 /* hmm_vma_handle_pte() did pte_unmap() */
392 return r;
393 }
394 }
395 pte_unmap(ptep - 1);
396 return 0;
397 }
398
399 #if defined(CONFIG_ARCH_HAS_PTE_DEVMAP) && \
400 defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD)
401 static inline unsigned long pud_to_hmm_pfn_flags(struct hmm_range *range,
402 pud_t pud)
403 {
404 if (!pud_present(pud))
405 return 0;
406 return (pud_write(pud) ? (HMM_PFN_VALID | HMM_PFN_WRITE) :
407 HMM_PFN_VALID) |
408 hmm_pfn_flags_order(PUD_SHIFT - PAGE_SHIFT);
409 }
410
411 static int hmm_vma_walk_pud(pud_t *pudp, unsigned long start, unsigned long end,
412 struct mm_walk *walk)
413 {
414 struct hmm_vma_walk *hmm_vma_walk = walk->private;
415 struct hmm_range *range = hmm_vma_walk->range;
416 unsigned long addr = start;
417 pud_t pud;
418 spinlock_t *ptl = pud_trans_huge_lock(pudp, walk->vma);
419
420 if (!ptl)
421 return 0;
422
423 /* Normally we don't want to split the huge page */
424 walk->action = ACTION_CONTINUE;
425
426 pud = READ_ONCE(*pudp);
427 if (!pud_present(pud)) {
428 spin_unlock(ptl);
429 return hmm_vma_walk_hole(start, end, -1, walk);
430 }
431
432 if (pud_leaf(pud) && pud_devmap(pud)) {
433 unsigned long i, npages, pfn;
434 unsigned int required_fault;
435 unsigned long *hmm_pfns;
436 unsigned long cpu_flags;
437
438 i = (addr - range->start) >> PAGE_SHIFT;
439 npages = (end - addr) >> PAGE_SHIFT;
440 hmm_pfns = &range->hmm_pfns[i];
441
442 cpu_flags = pud_to_hmm_pfn_flags(range, pud);
443 required_fault = hmm_range_need_fault(hmm_vma_walk, hmm_pfns,
444 npages, cpu_flags);
445 if (required_fault) {
446 spin_unlock(ptl);
447 return hmm_vma_fault(addr, end, required_fault, walk);
448 }
449
450 pfn = pud_pfn(pud) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
451 for (i = 0; i < npages; ++i, ++pfn)
452 hmm_pfns[i] = pfn | cpu_flags;
453 goto out_unlock;
454 }
455
456 /* Ask for the PUD to be split */
457 walk->action = ACTION_SUBTREE;
458
459 out_unlock:
460 spin_unlock(ptl);
461 return 0;
462 }
463 #else
464 #define hmm_vma_walk_pud NULL
465 #endif
466
467 #ifdef CONFIG_HUGETLB_PAGE
468 static int hmm_vma_walk_hugetlb_entry(pte_t *pte, unsigned long hmask,
469 unsigned long start, unsigned long end,
470 struct mm_walk *walk)
471 {
472 unsigned long addr = start, i, pfn;
473 struct hmm_vma_walk *hmm_vma_walk = walk->private;
474 struct hmm_range *range = hmm_vma_walk->range;
475 struct vm_area_struct *vma = walk->vma;
476 unsigned int required_fault;
477 unsigned long pfn_req_flags;
478 unsigned long cpu_flags;
479 spinlock_t *ptl;
480 pte_t entry;
481
482 ptl = huge_pte_lock(hstate_vma(vma), walk->mm, pte);
483 entry = huge_ptep_get(walk->mm, addr, pte);
484
485 i = (start - range->start) >> PAGE_SHIFT;
486 pfn_req_flags = range->hmm_pfns[i];
487 cpu_flags = pte_to_hmm_pfn_flags(range, entry) |
488 hmm_pfn_flags_order(huge_page_order(hstate_vma(vma)));
489 required_fault =
490 hmm_pte_need_fault(hmm_vma_walk, pfn_req_flags, cpu_flags);
491 if (required_fault) {
492 int ret;
493
494 spin_unlock(ptl);
495 hugetlb_vma_unlock_read(vma);
496 /*
497 * Avoid deadlock: drop the vma lock before calling
498 * hmm_vma_fault(), which will itself potentially take and
499 * drop the vma lock. This is also correct from a
500 * protection point of view, because there is no further
501 * use here of either pte or ptl after dropping the vma
502 * lock.
503 */
504 ret = hmm_vma_fault(addr, end, required_fault, walk);
505 hugetlb_vma_lock_read(vma);
506 return ret;
507 }
508
509 pfn = pte_pfn(entry) + ((start & ~hmask) >> PAGE_SHIFT);
510 for (; addr < end; addr += PAGE_SIZE, i++, pfn++)
511 range->hmm_pfns[i] = pfn | cpu_flags;
512
513 spin_unlock(ptl);
514 return 0;
515 }
516 #else
517 #define hmm_vma_walk_hugetlb_entry NULL
518 #endif /* CONFIG_HUGETLB_PAGE */
519
520 static int hmm_vma_walk_test(unsigned long start, unsigned long end,
521 struct mm_walk *walk)
522 {
523 struct hmm_vma_walk *hmm_vma_walk = walk->private;
524 struct hmm_range *range = hmm_vma_walk->range;
525 struct vm_area_struct *vma = walk->vma;
526
527 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)) &&
528 vma->vm_flags & VM_READ)
529 return 0;
530
531 /*
532 * vma ranges that don't have struct page backing them or map I/O
533 * devices directly cannot be handled by hmm_range_fault().
534 *
535 * If the vma does not allow read access, then assume that it does not
536 * allow write access either. HMM does not support architectures that
537 * allow write without read.
538 *
539 * If a fault is requested for an unsupported range then it is a hard
540 * failure.
541 */
542 if (hmm_range_need_fault(hmm_vma_walk,
543 range->hmm_pfns +
544 ((start - range->start) >> PAGE_SHIFT),
545 (end - start) >> PAGE_SHIFT, 0))
546 return -EFAULT;
547
548 hmm_pfns_fill(start, end, range, HMM_PFN_ERROR);
549
550 /* Skip this vma and continue processing the next vma. */
551 return 1;
552 }
553
554 static const struct mm_walk_ops hmm_walk_ops = {
555 .pud_entry = hmm_vma_walk_pud,
556 .pmd_entry = hmm_vma_walk_pmd,
557 .pte_hole = hmm_vma_walk_hole,
558 .hugetlb_entry = hmm_vma_walk_hugetlb_entry,
559 .test_walk = hmm_vma_walk_test,
560 .walk_lock = PGWALK_RDLOCK,
561 };
562
563 /**
564 * hmm_range_fault - try to fault some address in a virtual address range
565 * @range: argument structure
566 *
567 * Returns 0 on success or one of the following error codes:
568 *
569 * -EINVAL: Invalid arguments or mm or virtual address is in an invalid vma
570 * (e.g., device file vma).
571 * -ENOMEM: Out of memory.
572 * -EPERM: Invalid permission (e.g., asking for write and range is read
573 * only).
574 * -EBUSY: The range has been invalidated and the caller needs to wait for
575 * the invalidation to finish.
576 * -EFAULT: A page was requested to be valid and could not be made valid
577 * ie it has no backing VMA or it is illegal to access
578 *
579 * This is similar to get_user_pages(), except that it can read the page tables
580 * without mutating them (ie causing faults).
581 */
582 int hmm_range_fault(struct hmm_range *range)
583 {
584 struct hmm_vma_walk hmm_vma_walk = {
585 .range = range,
586 .last = range->start,
587 };
588 struct mm_struct *mm = range->notifier->mm;
589 int ret;
590
591 mmap_assert_locked(mm);
592
593 do {
594 /* If range is no longer valid force retry. */
595 if (mmu_interval_check_retry(range->notifier,
596 range->notifier_seq))
597 return -EBUSY;
598 ret = walk_page_range(mm, hmm_vma_walk.last, range->end,
599 &hmm_walk_ops, &hmm_vma_walk);
600 /*
601 * When -EBUSY is returned the loop restarts with
602 * hmm_vma_walk.last set to an address that has not been stored
603 * in pfns. All entries < last in the pfn array are set to their
604 * output, and all >= are still at their input values.
605 */
606 } while (ret == -EBUSY);
607 return ret;
608 }
609 EXPORT_SYMBOL(hmm_range_fault);
610
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.