1 // SPDX-License-Identifier: GPL-2.0-or-later 1
2 /*
3 * Copyright 2013 Red Hat Inc.
4 *
5 * Authors: Jérôme Glisse <jglisse@redhat.co
6 */
7 /*
8 * Refer to include/linux/hmm.h for informatio
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 | H
40 };
41
42 static int hmm_pfns_fill(unsigned long addr, u
43 struct hmm_range *ran
44 {
45 unsigned long i = (addr - range->start
46
47 for (; addr < end; addr += PAGE_SIZE,
48 range->hmm_pfns[i] = cpu_flags
49 return 0;
50 }
51
52 /*
53 * hmm_vma_fault() - fault in a range lacking
54 * @addr: range virtual start address (inclusi
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 fa
59 *
60 * This function will be called whenever pmd_n
61 * or whenever there is no page directory cove
62 */
63 static int hmm_vma_fault(unsigned long addr, u
64 unsigned int required
65 {
66 struct hmm_vma_walk *hmm_vma_walk = wa
67 struct vm_area_struct *vma = walk->vma
68 unsigned int fault_flags = FAULT_FLAG_
69
70 WARN_ON_ONCE(!required_fault);
71 hmm_vma_walk->last = addr;
72
73 if (required_fault & HMM_NEED_WRITE_FA
74 if (!(vma->vm_flags & VM_WRITE
75 return -EPERM;
76 fault_flags |= FAULT_FLAG_WRIT
77 }
78
79 for (; addr < end; addr += PAGE_SIZE)
80 if (handle_mm_fault(vma, addr,
81 VM_FAULT_ERROR)
82 return -EFAULT;
83 return -EBUSY;
84 }
85
86 static unsigned int hmm_pte_need_fault(const s
87 unsigne
88 unsigne
89 {
90 struct hmm_range *range = hmm_vma_walk
91
92 /*
93 * So we not only consider the individ
94 * consider the default flags requeste
95 * be used 2 ways. The first one where
96 * multiple page faults into one reque
97 * those faults. The second one where
98 * fault a range with specific flags.
99 * waste to have the user pre-fill the
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_FAUL
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_NE
113
114 /* If CPU page table is not valid then
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
122 const unsigned long hmm_p
123 unsigned long cpu_flags)
124 {
125 struct hmm_range *range = hmm_vma_walk
126 unsigned int required_fault = 0;
127 unsigned long i;
128
129 /*
130 * If the default flags do not request
131 * not allow for individual pages to b
132 * hmm_pte_need_fault() will always re
133 */
134 if (!((range->default_flags | range->p
135 HMM_PFN_REQ_FAULT))
136 return 0;
137
138 for (i = 0; i < npages; ++i) {
139 required_fault |= hmm_pte_need
140
141 if (required_fault == HMM_NEED
142 return required_fault;
143 }
144 return required_fault;
145 }
146
147 static int hmm_vma_walk_hole(unsigned long add
148 __always_unused i
149 {
150 struct hmm_vma_walk *hmm_vma_walk = wa
151 struct hmm_range *range = hmm_vma_walk
152 unsigned int required_fault;
153 unsigned long i, npages;
154 unsigned long *hmm_pfns;
155
156 i = (addr - range->start) >> PAGE_SHIF
157 npages = (end - addr) >> PAGE_SHIFT;
158 hmm_pfns = &range->hmm_pfns[i];
159 required_fault =
160 hmm_range_need_fault(hmm_vma_w
161 if (!walk->vma) {
162 if (required_fault)
163 return -EFAULT;
164 return hmm_pfns_fill(addr, end
165 }
166 if (required_fault)
167 return hmm_vma_fault(addr, end
168 return hmm_pfns_fill(addr, end, range,
169 }
170
171 static inline unsigned long hmm_pfn_flags_orde
172 {
173 return order << HMM_PFN_ORDER_SHIFT;
174 }
175
176 static inline unsigned long pmd_to_hmm_pfn_fla
177
178 {
179 if (pmd_protnone(pmd))
180 return 0;
181 return (pmd_write(pmd) ? (HMM_PFN_VALI
182 HMM_PFN_VALID
183 hmm_pfn_flags_order(PMD_SHIFT -
184 }
185
186 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
187 static int hmm_vma_handle_pmd(struct mm_walk *
188 unsigned long en
189 pmd_t pmd)
190 {
191 struct hmm_vma_walk *hmm_vma_walk = wa
192 struct hmm_range *range = hmm_vma_walk
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
199 required_fault =
200 hmm_range_need_fault(hmm_vma_w
201 if (required_fault)
202 return hmm_vma_fault(addr, end
203
204 pfn = pmd_pfn(pmd) + ((addr & ~PMD_MAS
205 for (i = 0; addr < end; addr += PAGE_S
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, u
212 unsigned long end, unsigned lo
213 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
214
215 static inline unsigned long pte_to_hmm_pfn_fla
216
217 {
218 if (pte_none(pte) || !pte_present(pte)
219 return 0;
220 return pte_write(pte) ? (HMM_PFN_VALID
221 }
222
223 static int hmm_vma_handle_pte(struct mm_walk *
224 unsigned long en
225 unsigned long *h
226 {
227 struct hmm_vma_walk *hmm_vma_walk = wa
228 struct hmm_range *range = hmm_vma_walk
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
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
245
246 /*
247 * Don't fault in device priva
248 * just report the PFN.
249 */
250 if (is_device_private_entry(en
251 pfn_swap_entry_to_page(ent
252 range->dev_private_owner)
253 cpu_flags = HMM_PFN_VA
254 if (is_writable_device
255 cpu_flags |= H
256 *hmm_pfn = swp_offset_
257 return 0;
258 }
259
260 required_fault =
261 hmm_pte_need_fault(hmm
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(en
271 goto fault;
272
273 if (is_device_exclusive_entry(
274 goto fault;
275
276 if (is_migration_entry(entry))
277 pte_unmap(ptep);
278 hmm_vma_walk->last = a
279 migration_entry_wait(w
280 return -EBUSY;
281 }
282
283 /* Report error for everything
284 pte_unmap(ptep);
285 return -EFAULT;
286 }
287
288 cpu_flags = pte_to_hmm_pfn_flags(range
289 required_fault =
290 hmm_pte_need_fault(hmm_vma_wal
291 if (required_fault)
292 goto fault;
293
294 /*
295 * Bypass devmap pte such as DAX page
296 * flags(pfn_req_flags) are fulfilled.
297 * Since each architecture defines a s
298 * fall through and treat it like a no
299 */
300 if (!vm_normal_page(walk->vma, addr, p
301 !pte_devmap(pte) &&
302 !is_zero_pfn(pte_pfn(pte))) {
303 if (hmm_pte_need_fault(hmm_vma
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 a
317 return hmm_vma_fault(addr, end, requir
318 }
319
320 static int hmm_vma_walk_pmd(pmd_t *pmdp,
321 unsigned long star
322 unsigned long end,
323 struct mm_walk *wa
324 {
325 struct hmm_vma_walk *hmm_vma_walk = wa
326 struct hmm_range *range = hmm_vma_walk
327 unsigned long *hmm_pfns =
328 &range->hmm_pfns[(start - rang
329 unsigned long npages = (end - start) >
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
338
339 if (thp_migration_supported() && is_pm
340 if (hmm_range_need_fault(hmm_v
341 hmm_vma_walk->last = a
342 pmd_migration_entry_wa
343 return -EBUSY;
344 }
345 return hmm_pfns_fill(start, en
346 }
347
348 if (!pmd_present(pmd)) {
349 if (hmm_range_need_fault(hmm_v
350 return -EFAULT;
351 return hmm_pfns_fill(start, en
352 }
353
354 if (pmd_devmap(pmd) || pmd_trans_huge(
355 /*
356 * No need to take pmd_lock he
357 * is splitting the huge pmd w
358 * mmu_notifier callback.
359 *
360 * So just read pmd value and
361 * huge or device mapping one
362 * values.
363 */
364 pmd = pmdp_get_lockless(pmdp);
365 if (!pmd_devmap(pmd) && !pmd_t
366 goto again;
367
368 return hmm_vma_handle_pmd(walk
369 }
370
371 /*
372 * We have handled all the valid cases
373 * huge or transparent huge. At this p
374 * entry pointing to pte directory or
375 * recover.
376 */
377 if (pmd_bad(pmd)) {
378 if (hmm_range_need_fault(hmm_v
379 return -EFAULT;
380 return hmm_pfns_fill(start, en
381 }
382
383 ptep = pte_offset_map(pmdp, addr);
384 if (!ptep)
385 goto again;
386 for (; addr < end; addr += PAGE_SIZE,
387 int r;
388
389 r = hmm_vma_handle_pte(walk, a
390 if (r) {
391 /* hmm_vma_handle_pte(
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_HUGEP
401 static inline unsigned long pud_to_hmm_pfn_fla
402
403 {
404 if (!pud_present(pud))
405 return 0;
406 return (pud_write(pud) ? (HMM_PFN_VALI
407 HMM_PFN_VALID
408 hmm_pfn_flags_order(PUD_SHIFT -
409 }
410
411 static int hmm_vma_walk_pud(pud_t *pudp, unsig
412 struct mm_walk *walk)
413 {
414 struct hmm_vma_walk *hmm_vma_walk = wa
415 struct hmm_range *range = hmm_vma_walk
416 unsigned long addr = start;
417 pud_t pud;
418 spinlock_t *ptl = pud_trans_huge_lock(
419
420 if (!ptl)
421 return 0;
422
423 /* Normally we don't want to split the
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
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) >> P
439 npages = (end - addr) >> PAGE_
440 hmm_pfns = &range->hmm_pfns[i]
441
442 cpu_flags = pud_to_hmm_pfn_fla
443 required_fault = hmm_range_nee
444
445 if (required_fault) {
446 spin_unlock(ptl);
447 return hmm_vma_fault(a
448 }
449
450 pfn = pud_pfn(pud) + ((addr &
451 for (i = 0; i < npages; ++i, +
452 hmm_pfns[i] = pfn | cp
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 *p
469 unsigned
470 struct m
471 {
472 unsigned long addr = start, i, pfn;
473 struct hmm_vma_walk *hmm_vma_walk = wa
474 struct hmm_range *range = hmm_vma_walk
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), w
483 entry = huge_ptep_get(walk->mm, addr,
484
485 i = (start - range->start) >> PAGE_SHI
486 pfn_req_flags = range->hmm_pfns[i];
487 cpu_flags = pte_to_hmm_pfn_flags(range
488 hmm_pfn_flags_order(huge_p
489 required_fault =
490 hmm_pte_need_fault(hmm_vma_wal
491 if (required_fault) {
492 int ret;
493
494 spin_unlock(ptl);
495 hugetlb_vma_unlock_read(vma);
496 /*
497 * Avoid deadlock: drop the vm
498 * hmm_vma_fault(), which will
499 * drop the vma lock. This is
500 * protection point of view, b
501 * use here of either pte or p
502 * lock.
503 */
504 ret = hmm_vma_fault(addr, end,
505 hugetlb_vma_lock_read(vma);
506 return ret;
507 }
508
509 pfn = pte_pfn(entry) + ((start & ~hmas
510 for (; addr < end; addr += PAGE_SIZE,
511 range->hmm_pfns[i] = pfn | cpu
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 sta
521 struct mm_walk *w
522 {
523 struct hmm_vma_walk *hmm_vma_walk = wa
524 struct hmm_range *range = hmm_vma_walk
525 struct vm_area_struct *vma = walk->vma
526
527 if (!(vma->vm_flags & (VM_IO | VM_PFNM
528 vma->vm_flags & VM_READ)
529 return 0;
530
531 /*
532 * vma ranges that don't have struct p
533 * devices directly cannot be handled
534 *
535 * If the vma does not allow read acce
536 * allow write access either. HMM does
537 * allow write without read.
538 *
539 * If a fault is requested for an unsu
540 * failure.
541 */
542 if (hmm_range_need_fault(hmm_vma_walk,
543 range->hmm_pf
544 ((sta
545 (end - start)
546 return -EFAULT;
547
548 hmm_pfns_fill(start, end, range, HMM_P
549
550 /* Skip this vma and continue processi
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
559 .test_walk = hmm_vma_walk_test,
560 .walk_lock = PGWALK_RDLOCK,
561 };
562
563 /**
564 * hmm_range_fault - try to fault some address
565 * @range: argument structure
566 *
567 * Returns 0 on success or one of the followin
568 *
569 * -EINVAL: Invalid arguments or mm or vir
570 * (e.g., device file vma).
571 * -ENOMEM: Out of memory.
572 * -EPERM: Invalid permission (e.g., aski
573 * only).
574 * -EBUSY: The range has been invalidated
575 * the invalidation to finish.
576 * -EFAULT: A page was requested to be val
577 * ie it has no backing VMA or it
578 *
579 * This is similar to get_user_pages(), except
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
589 int ret;
590
591 mmap_assert_locked(mm);
592
593 do {
594 /* If range is no longer valid
595 if (mmu_interval_check_retry(r
596 r
597 return -EBUSY;
598 ret = walk_page_range(mm, hmm_
599 &hmm_wal
600 /*
601 * When -EBUSY is returned the
602 * hmm_vma_walk.last set to an
603 * in pfns. All entries < last
604 * output, and all >= are stil
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.