Loading...
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
31struct hmm_vma_walk {
32 struct hmm_range *range;
33 unsigned long last;
34};
35
36enum {
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
42static 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 */
63static 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
86static 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
120static unsigned int
121hmm_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
147static 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
171static inline unsigned long hmm_pfn_flags_order(unsigned long order)
172{
173 return order << HMM_PFN_ORDER_SHIFT;
174}
175
176static 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
187static 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 */
211int 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
215static 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
223static 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;
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
314fault:
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
320static 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
334again:
335 pmd = READ_ONCE(*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 for (; addr < end; addr += PAGE_SIZE, ptep++, hmm_pfns++) {
385 int r;
386
387 r = hmm_vma_handle_pte(walk, addr, end, pmdp, ptep, hmm_pfns);
388 if (r) {
389 /* hmm_vma_handle_pte() did pte_unmap() */
390 return r;
391 }
392 }
393 pte_unmap(ptep - 1);
394 return 0;
395}
396
397#if defined(CONFIG_ARCH_HAS_PTE_DEVMAP) && \
398 defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD)
399static inline unsigned long pud_to_hmm_pfn_flags(struct hmm_range *range,
400 pud_t pud)
401{
402 if (!pud_present(pud))
403 return 0;
404 return (pud_write(pud) ? (HMM_PFN_VALID | HMM_PFN_WRITE) :
405 HMM_PFN_VALID) |
406 hmm_pfn_flags_order(PUD_SHIFT - PAGE_SHIFT);
407}
408
409static int hmm_vma_walk_pud(pud_t *pudp, unsigned long start, unsigned long end,
410 struct mm_walk *walk)
411{
412 struct hmm_vma_walk *hmm_vma_walk = walk->private;
413 struct hmm_range *range = hmm_vma_walk->range;
414 unsigned long addr = start;
415 pud_t pud;
416 spinlock_t *ptl = pud_trans_huge_lock(pudp, walk->vma);
417
418 if (!ptl)
419 return 0;
420
421 /* Normally we don't want to split the huge page */
422 walk->action = ACTION_CONTINUE;
423
424 pud = READ_ONCE(*pudp);
425 if (pud_none(pud)) {
426 spin_unlock(ptl);
427 return hmm_vma_walk_hole(start, end, -1, walk);
428 }
429
430 if (pud_huge(pud) && pud_devmap(pud)) {
431 unsigned long i, npages, pfn;
432 unsigned int required_fault;
433 unsigned long *hmm_pfns;
434 unsigned long cpu_flags;
435
436 if (!pud_present(pud)) {
437 spin_unlock(ptl);
438 return hmm_vma_walk_hole(start, end, -1, walk);
439 }
440
441 i = (addr - range->start) >> PAGE_SHIFT;
442 npages = (end - addr) >> PAGE_SHIFT;
443 hmm_pfns = &range->hmm_pfns[i];
444
445 cpu_flags = pud_to_hmm_pfn_flags(range, pud);
446 required_fault = hmm_range_need_fault(hmm_vma_walk, hmm_pfns,
447 npages, cpu_flags);
448 if (required_fault) {
449 spin_unlock(ptl);
450 return hmm_vma_fault(addr, end, required_fault, walk);
451 }
452
453 pfn = pud_pfn(pud) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
454 for (i = 0; i < npages; ++i, ++pfn)
455 hmm_pfns[i] = pfn | cpu_flags;
456 goto out_unlock;
457 }
458
459 /* Ask for the PUD to be split */
460 walk->action = ACTION_SUBTREE;
461
462out_unlock:
463 spin_unlock(ptl);
464 return 0;
465}
466#else
467#define hmm_vma_walk_pud NULL
468#endif
469
470#ifdef CONFIG_HUGETLB_PAGE
471static int hmm_vma_walk_hugetlb_entry(pte_t *pte, unsigned long hmask,
472 unsigned long start, unsigned long end,
473 struct mm_walk *walk)
474{
475 unsigned long addr = start, i, pfn;
476 struct hmm_vma_walk *hmm_vma_walk = walk->private;
477 struct hmm_range *range = hmm_vma_walk->range;
478 struct vm_area_struct *vma = walk->vma;
479 unsigned int required_fault;
480 unsigned long pfn_req_flags;
481 unsigned long cpu_flags;
482 spinlock_t *ptl;
483 pte_t entry;
484
485 ptl = huge_pte_lock(hstate_vma(vma), walk->mm, pte);
486 entry = huge_ptep_get(pte);
487
488 i = (start - range->start) >> PAGE_SHIFT;
489 pfn_req_flags = range->hmm_pfns[i];
490 cpu_flags = pte_to_hmm_pfn_flags(range, entry) |
491 hmm_pfn_flags_order(huge_page_order(hstate_vma(vma)));
492 required_fault =
493 hmm_pte_need_fault(hmm_vma_walk, pfn_req_flags, cpu_flags);
494 if (required_fault) {
495 spin_unlock(ptl);
496 return hmm_vma_fault(addr, end, required_fault, walk);
497 }
498
499 pfn = pte_pfn(entry) + ((start & ~hmask) >> PAGE_SHIFT);
500 for (; addr < end; addr += PAGE_SIZE, i++, pfn++)
501 range->hmm_pfns[i] = pfn | cpu_flags;
502
503 spin_unlock(ptl);
504 return 0;
505}
506#else
507#define hmm_vma_walk_hugetlb_entry NULL
508#endif /* CONFIG_HUGETLB_PAGE */
509
510static int hmm_vma_walk_test(unsigned long start, unsigned long end,
511 struct mm_walk *walk)
512{
513 struct hmm_vma_walk *hmm_vma_walk = walk->private;
514 struct hmm_range *range = hmm_vma_walk->range;
515 struct vm_area_struct *vma = walk->vma;
516
517 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)) &&
518 vma->vm_flags & VM_READ)
519 return 0;
520
521 /*
522 * vma ranges that don't have struct page backing them or map I/O
523 * devices directly cannot be handled by hmm_range_fault().
524 *
525 * If the vma does not allow read access, then assume that it does not
526 * allow write access either. HMM does not support architectures that
527 * allow write without read.
528 *
529 * If a fault is requested for an unsupported range then it is a hard
530 * failure.
531 */
532 if (hmm_range_need_fault(hmm_vma_walk,
533 range->hmm_pfns +
534 ((start - range->start) >> PAGE_SHIFT),
535 (end - start) >> PAGE_SHIFT, 0))
536 return -EFAULT;
537
538 hmm_pfns_fill(start, end, range, HMM_PFN_ERROR);
539
540 /* Skip this vma and continue processing the next vma. */
541 return 1;
542}
543
544static const struct mm_walk_ops hmm_walk_ops = {
545 .pud_entry = hmm_vma_walk_pud,
546 .pmd_entry = hmm_vma_walk_pmd,
547 .pte_hole = hmm_vma_walk_hole,
548 .hugetlb_entry = hmm_vma_walk_hugetlb_entry,
549 .test_walk = hmm_vma_walk_test,
550};
551
552/**
553 * hmm_range_fault - try to fault some address in a virtual address range
554 * @range: argument structure
555 *
556 * Returns 0 on success or one of the following error codes:
557 *
558 * -EINVAL: Invalid arguments or mm or virtual address is in an invalid vma
559 * (e.g., device file vma).
560 * -ENOMEM: Out of memory.
561 * -EPERM: Invalid permission (e.g., asking for write and range is read
562 * only).
563 * -EBUSY: The range has been invalidated and the caller needs to wait for
564 * the invalidation to finish.
565 * -EFAULT: A page was requested to be valid and could not be made valid
566 * ie it has no backing VMA or it is illegal to access
567 *
568 * This is similar to get_user_pages(), except that it can read the page tables
569 * without mutating them (ie causing faults).
570 */
571int hmm_range_fault(struct hmm_range *range)
572{
573 struct hmm_vma_walk hmm_vma_walk = {
574 .range = range,
575 .last = range->start,
576 };
577 struct mm_struct *mm = range->notifier->mm;
578 int ret;
579
580 mmap_assert_locked(mm);
581
582 do {
583 /* If range is no longer valid force retry. */
584 if (mmu_interval_check_retry(range->notifier,
585 range->notifier_seq))
586 return -EBUSY;
587 ret = walk_page_range(mm, hmm_vma_walk.last, range->end,
588 &hmm_walk_ops, &hmm_vma_walk);
589 /*
590 * When -EBUSY is returned the loop restarts with
591 * hmm_vma_walk.last set to an address that has not been stored
592 * in pfns. All entries < last in the pfn array are set to their
593 * output, and all >= are still at their input values.
594 */
595 } while (ret == -EBUSY);
596 return ret;
597}
598EXPORT_SYMBOL(hmm_range_fault);
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
31struct hmm_vma_walk {
32 struct hmm_range *range;
33 unsigned long last;
34};
35
36enum {
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
42static 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 */
63static 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
86static 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
120static unsigned int
121hmm_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
147static 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
171static inline unsigned long hmm_pfn_flags_order(unsigned long order)
172{
173 return order << HMM_PFN_ORDER_SHIFT;
174}
175
176static 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
187static 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 */
211int 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
215static inline bool hmm_is_device_private_entry(struct hmm_range *range,
216 swp_entry_t entry)
217{
218 return is_device_private_entry(entry) &&
219 pfn_swap_entry_to_page(entry)->pgmap->owner ==
220 range->dev_private_owner;
221}
222
223static inline unsigned long pte_to_hmm_pfn_flags(struct hmm_range *range,
224 pte_t pte)
225{
226 if (pte_none(pte) || !pte_present(pte) || pte_protnone(pte))
227 return 0;
228 return pte_write(pte) ? (HMM_PFN_VALID | HMM_PFN_WRITE) : HMM_PFN_VALID;
229}
230
231static int hmm_vma_handle_pte(struct mm_walk *walk, unsigned long addr,
232 unsigned long end, pmd_t *pmdp, pte_t *ptep,
233 unsigned long *hmm_pfn)
234{
235 struct hmm_vma_walk *hmm_vma_walk = walk->private;
236 struct hmm_range *range = hmm_vma_walk->range;
237 unsigned int required_fault;
238 unsigned long cpu_flags;
239 pte_t pte = *ptep;
240 uint64_t pfn_req_flags = *hmm_pfn;
241
242 if (pte_none(pte)) {
243 required_fault =
244 hmm_pte_need_fault(hmm_vma_walk, pfn_req_flags, 0);
245 if (required_fault)
246 goto fault;
247 *hmm_pfn = 0;
248 return 0;
249 }
250
251 if (!pte_present(pte)) {
252 swp_entry_t entry = pte_to_swp_entry(pte);
253
254 /*
255 * Never fault in device private pages, but just report
256 * the PFN even if not present.
257 */
258 if (hmm_is_device_private_entry(range, entry)) {
259 cpu_flags = HMM_PFN_VALID;
260 if (is_writable_device_private_entry(entry))
261 cpu_flags |= HMM_PFN_WRITE;
262 *hmm_pfn = swp_offset(entry) | cpu_flags;
263 return 0;
264 }
265
266 required_fault =
267 hmm_pte_need_fault(hmm_vma_walk, pfn_req_flags, 0);
268 if (!required_fault) {
269 *hmm_pfn = 0;
270 return 0;
271 }
272
273 if (!non_swap_entry(entry))
274 goto fault;
275
276 if (is_device_exclusive_entry(entry))
277 goto fault;
278
279 if (is_migration_entry(entry)) {
280 pte_unmap(ptep);
281 hmm_vma_walk->last = addr;
282 migration_entry_wait(walk->mm, pmdp, addr);
283 return -EBUSY;
284 }
285
286 /* Report error for everything else */
287 pte_unmap(ptep);
288 return -EFAULT;
289 }
290
291 cpu_flags = pte_to_hmm_pfn_flags(range, pte);
292 required_fault =
293 hmm_pte_need_fault(hmm_vma_walk, pfn_req_flags, cpu_flags);
294 if (required_fault)
295 goto fault;
296
297 /*
298 * Bypass devmap pte such as DAX page when all pfn requested
299 * flags(pfn_req_flags) are fulfilled.
300 * Since each architecture defines a struct page for the zero page, just
301 * fall through and treat it like a normal page.
302 */
303 if (pte_special(pte) && !pte_devmap(pte) &&
304 !is_zero_pfn(pte_pfn(pte))) {
305 if (hmm_pte_need_fault(hmm_vma_walk, pfn_req_flags, 0)) {
306 pte_unmap(ptep);
307 return -EFAULT;
308 }
309 *hmm_pfn = HMM_PFN_ERROR;
310 return 0;
311 }
312
313 *hmm_pfn = pte_pfn(pte) | cpu_flags;
314 return 0;
315
316fault:
317 pte_unmap(ptep);
318 /* Fault any virtual address we were asked to fault */
319 return hmm_vma_fault(addr, end, required_fault, walk);
320}
321
322static int hmm_vma_walk_pmd(pmd_t *pmdp,
323 unsigned long start,
324 unsigned long end,
325 struct mm_walk *walk)
326{
327 struct hmm_vma_walk *hmm_vma_walk = walk->private;
328 struct hmm_range *range = hmm_vma_walk->range;
329 unsigned long *hmm_pfns =
330 &range->hmm_pfns[(start - range->start) >> PAGE_SHIFT];
331 unsigned long npages = (end - start) >> PAGE_SHIFT;
332 unsigned long addr = start;
333 pte_t *ptep;
334 pmd_t pmd;
335
336again:
337 pmd = READ_ONCE(*pmdp);
338 if (pmd_none(pmd))
339 return hmm_vma_walk_hole(start, end, -1, walk);
340
341 if (thp_migration_supported() && is_pmd_migration_entry(pmd)) {
342 if (hmm_range_need_fault(hmm_vma_walk, hmm_pfns, npages, 0)) {
343 hmm_vma_walk->last = addr;
344 pmd_migration_entry_wait(walk->mm, pmdp);
345 return -EBUSY;
346 }
347 return hmm_pfns_fill(start, end, range, 0);
348 }
349
350 if (!pmd_present(pmd)) {
351 if (hmm_range_need_fault(hmm_vma_walk, hmm_pfns, npages, 0))
352 return -EFAULT;
353 return hmm_pfns_fill(start, end, range, HMM_PFN_ERROR);
354 }
355
356 if (pmd_devmap(pmd) || pmd_trans_huge(pmd)) {
357 /*
358 * No need to take pmd_lock here, even if some other thread
359 * is splitting the huge pmd we will get that event through
360 * mmu_notifier callback.
361 *
362 * So just read pmd value and check again it's a transparent
363 * huge or device mapping one and compute corresponding pfn
364 * values.
365 */
366 pmd = pmd_read_atomic(pmdp);
367 barrier();
368 if (!pmd_devmap(pmd) && !pmd_trans_huge(pmd))
369 goto again;
370
371 return hmm_vma_handle_pmd(walk, addr, end, hmm_pfns, pmd);
372 }
373
374 /*
375 * We have handled all the valid cases above ie either none, migration,
376 * huge or transparent huge. At this point either it is a valid pmd
377 * entry pointing to pte directory or it is a bad pmd that will not
378 * recover.
379 */
380 if (pmd_bad(pmd)) {
381 if (hmm_range_need_fault(hmm_vma_walk, hmm_pfns, npages, 0))
382 return -EFAULT;
383 return hmm_pfns_fill(start, end, range, HMM_PFN_ERROR);
384 }
385
386 ptep = pte_offset_map(pmdp, addr);
387 for (; addr < end; addr += PAGE_SIZE, ptep++, hmm_pfns++) {
388 int r;
389
390 r = hmm_vma_handle_pte(walk, addr, end, pmdp, ptep, hmm_pfns);
391 if (r) {
392 /* hmm_vma_handle_pte() did pte_unmap() */
393 return r;
394 }
395 }
396 pte_unmap(ptep - 1);
397 return 0;
398}
399
400#if defined(CONFIG_ARCH_HAS_PTE_DEVMAP) && \
401 defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD)
402static inline unsigned long pud_to_hmm_pfn_flags(struct hmm_range *range,
403 pud_t pud)
404{
405 if (!pud_present(pud))
406 return 0;
407 return (pud_write(pud) ? (HMM_PFN_VALID | HMM_PFN_WRITE) :
408 HMM_PFN_VALID) |
409 hmm_pfn_flags_order(PUD_SHIFT - PAGE_SHIFT);
410}
411
412static int hmm_vma_walk_pud(pud_t *pudp, unsigned long start, unsigned long end,
413 struct mm_walk *walk)
414{
415 struct hmm_vma_walk *hmm_vma_walk = walk->private;
416 struct hmm_range *range = hmm_vma_walk->range;
417 unsigned long addr = start;
418 pud_t pud;
419 int ret = 0;
420 spinlock_t *ptl = pud_trans_huge_lock(pudp, walk->vma);
421
422 if (!ptl)
423 return 0;
424
425 /* Normally we don't want to split the huge page */
426 walk->action = ACTION_CONTINUE;
427
428 pud = READ_ONCE(*pudp);
429 if (pud_none(pud)) {
430 spin_unlock(ptl);
431 return hmm_vma_walk_hole(start, end, -1, walk);
432 }
433
434 if (pud_huge(pud) && pud_devmap(pud)) {
435 unsigned long i, npages, pfn;
436 unsigned int required_fault;
437 unsigned long *hmm_pfns;
438 unsigned long cpu_flags;
439
440 if (!pud_present(pud)) {
441 spin_unlock(ptl);
442 return hmm_vma_walk_hole(start, end, -1, walk);
443 }
444
445 i = (addr - range->start) >> PAGE_SHIFT;
446 npages = (end - addr) >> PAGE_SHIFT;
447 hmm_pfns = &range->hmm_pfns[i];
448
449 cpu_flags = pud_to_hmm_pfn_flags(range, pud);
450 required_fault = hmm_range_need_fault(hmm_vma_walk, hmm_pfns,
451 npages, cpu_flags);
452 if (required_fault) {
453 spin_unlock(ptl);
454 return hmm_vma_fault(addr, end, required_fault, walk);
455 }
456
457 pfn = pud_pfn(pud) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
458 for (i = 0; i < npages; ++i, ++pfn)
459 hmm_pfns[i] = pfn | cpu_flags;
460 goto out_unlock;
461 }
462
463 /* Ask for the PUD to be split */
464 walk->action = ACTION_SUBTREE;
465
466out_unlock:
467 spin_unlock(ptl);
468 return ret;
469}
470#else
471#define hmm_vma_walk_pud NULL
472#endif
473
474#ifdef CONFIG_HUGETLB_PAGE
475static int hmm_vma_walk_hugetlb_entry(pte_t *pte, unsigned long hmask,
476 unsigned long start, unsigned long end,
477 struct mm_walk *walk)
478{
479 unsigned long addr = start, i, pfn;
480 struct hmm_vma_walk *hmm_vma_walk = walk->private;
481 struct hmm_range *range = hmm_vma_walk->range;
482 struct vm_area_struct *vma = walk->vma;
483 unsigned int required_fault;
484 unsigned long pfn_req_flags;
485 unsigned long cpu_flags;
486 spinlock_t *ptl;
487 pte_t entry;
488
489 ptl = huge_pte_lock(hstate_vma(vma), walk->mm, pte);
490 entry = huge_ptep_get(pte);
491
492 i = (start - range->start) >> PAGE_SHIFT;
493 pfn_req_flags = range->hmm_pfns[i];
494 cpu_flags = pte_to_hmm_pfn_flags(range, entry) |
495 hmm_pfn_flags_order(huge_page_order(hstate_vma(vma)));
496 required_fault =
497 hmm_pte_need_fault(hmm_vma_walk, pfn_req_flags, cpu_flags);
498 if (required_fault) {
499 spin_unlock(ptl);
500 return hmm_vma_fault(addr, end, required_fault, walk);
501 }
502
503 pfn = pte_pfn(entry) + ((start & ~hmask) >> PAGE_SHIFT);
504 for (; addr < end; addr += PAGE_SIZE, i++, pfn++)
505 range->hmm_pfns[i] = pfn | cpu_flags;
506
507 spin_unlock(ptl);
508 return 0;
509}
510#else
511#define hmm_vma_walk_hugetlb_entry NULL
512#endif /* CONFIG_HUGETLB_PAGE */
513
514static int hmm_vma_walk_test(unsigned long start, unsigned long end,
515 struct mm_walk *walk)
516{
517 struct hmm_vma_walk *hmm_vma_walk = walk->private;
518 struct hmm_range *range = hmm_vma_walk->range;
519 struct vm_area_struct *vma = walk->vma;
520
521 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP | VM_MIXEDMAP)) &&
522 vma->vm_flags & VM_READ)
523 return 0;
524
525 /*
526 * vma ranges that don't have struct page backing them or map I/O
527 * devices directly cannot be handled by hmm_range_fault().
528 *
529 * If the vma does not allow read access, then assume that it does not
530 * allow write access either. HMM does not support architectures that
531 * allow write without read.
532 *
533 * If a fault is requested for an unsupported range then it is a hard
534 * failure.
535 */
536 if (hmm_range_need_fault(hmm_vma_walk,
537 range->hmm_pfns +
538 ((start - range->start) >> PAGE_SHIFT),
539 (end - start) >> PAGE_SHIFT, 0))
540 return -EFAULT;
541
542 hmm_pfns_fill(start, end, range, HMM_PFN_ERROR);
543
544 /* Skip this vma and continue processing the next vma. */
545 return 1;
546}
547
548static const struct mm_walk_ops hmm_walk_ops = {
549 .pud_entry = hmm_vma_walk_pud,
550 .pmd_entry = hmm_vma_walk_pmd,
551 .pte_hole = hmm_vma_walk_hole,
552 .hugetlb_entry = hmm_vma_walk_hugetlb_entry,
553 .test_walk = hmm_vma_walk_test,
554};
555
556/**
557 * hmm_range_fault - try to fault some address in a virtual address range
558 * @range: argument structure
559 *
560 * Returns 0 on success or one of the following error codes:
561 *
562 * -EINVAL: Invalid arguments or mm or virtual address is in an invalid vma
563 * (e.g., device file vma).
564 * -ENOMEM: Out of memory.
565 * -EPERM: Invalid permission (e.g., asking for write and range is read
566 * only).
567 * -EBUSY: The range has been invalidated and the caller needs to wait for
568 * the invalidation to finish.
569 * -EFAULT: A page was requested to be valid and could not be made valid
570 * ie it has no backing VMA or it is illegal to access
571 *
572 * This is similar to get_user_pages(), except that it can read the page tables
573 * without mutating them (ie causing faults).
574 */
575int hmm_range_fault(struct hmm_range *range)
576{
577 struct hmm_vma_walk hmm_vma_walk = {
578 .range = range,
579 .last = range->start,
580 };
581 struct mm_struct *mm = range->notifier->mm;
582 int ret;
583
584 mmap_assert_locked(mm);
585
586 do {
587 /* If range is no longer valid force retry. */
588 if (mmu_interval_check_retry(range->notifier,
589 range->notifier_seq))
590 return -EBUSY;
591 ret = walk_page_range(mm, hmm_vma_walk.last, range->end,
592 &hmm_walk_ops, &hmm_vma_walk);
593 /*
594 * When -EBUSY is returned the loop restarts with
595 * hmm_vma_walk.last set to an address that has not been stored
596 * in pfns. All entries < last in the pfn array are set to their
597 * output, and all >= are still at their input values.
598 */
599 } while (ret == -EBUSY);
600 return ret;
601}
602EXPORT_SYMBOL(hmm_range_fault);