Linux Audio

Check our new training course

Loading...
v6.2
  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);
v5.9
  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
 
 
 29struct hmm_vma_walk {
 30	struct hmm_range	*range;
 31	unsigned long		last;
 32};
 33
 34enum {
 35	HMM_NEED_FAULT = 1 << 0,
 36	HMM_NEED_WRITE_FAULT = 1 << 1,
 37	HMM_NEED_ALL_BITS = HMM_NEED_FAULT | HMM_NEED_WRITE_FAULT,
 38};
 39
 40static int hmm_pfns_fill(unsigned long addr, unsigned long end,
 41			 struct hmm_range *range, unsigned long cpu_flags)
 42{
 43	unsigned long i = (addr - range->start) >> PAGE_SHIFT;
 44
 45	for (; addr < end; addr += PAGE_SIZE, i++)
 46		range->hmm_pfns[i] = cpu_flags;
 47	return 0;
 48}
 49
 50/*
 51 * hmm_vma_fault() - fault in a range lacking valid pmd or pte(s)
 52 * @addr: range virtual start address (inclusive)
 53 * @end: range virtual end address (exclusive)
 54 * @required_fault: HMM_NEED_* flags
 55 * @walk: mm_walk structure
 56 * Return: -EBUSY after page fault, or page fault error
 57 *
 58 * This function will be called whenever pmd_none() or pte_none() returns true,
 59 * or whenever there is no page directory covering the virtual address range.
 60 */
 61static int hmm_vma_fault(unsigned long addr, unsigned long end,
 62			 unsigned int required_fault, struct mm_walk *walk)
 63{
 64	struct hmm_vma_walk *hmm_vma_walk = walk->private;
 65	struct vm_area_struct *vma = walk->vma;
 66	unsigned int fault_flags = FAULT_FLAG_REMOTE;
 67
 68	WARN_ON_ONCE(!required_fault);
 69	hmm_vma_walk->last = addr;
 70
 71	if (required_fault & HMM_NEED_WRITE_FAULT) {
 72		if (!(vma->vm_flags & VM_WRITE))
 73			return -EPERM;
 74		fault_flags |= FAULT_FLAG_WRITE;
 75	}
 76
 77	for (; addr < end; addr += PAGE_SIZE)
 78		if (handle_mm_fault(vma, addr, fault_flags, NULL) &
 79		    VM_FAULT_ERROR)
 80			return -EFAULT;
 81	return -EBUSY;
 82}
 83
 84static unsigned int hmm_pte_need_fault(const struct hmm_vma_walk *hmm_vma_walk,
 85				       unsigned long pfn_req_flags,
 86				       unsigned long cpu_flags)
 87{
 88	struct hmm_range *range = hmm_vma_walk->range;
 89
 90	/*
 91	 * So we not only consider the individual per page request we also
 92	 * consider the default flags requested for the range. The API can
 93	 * be used 2 ways. The first one where the HMM user coalesces
 94	 * multiple page faults into one request and sets flags per pfn for
 95	 * those faults. The second one where the HMM user wants to pre-
 96	 * fault a range with specific flags. For the latter one it is a
 97	 * waste to have the user pre-fill the pfn arrays with a default
 98	 * flags value.
 99	 */
100	pfn_req_flags &= range->pfn_flags_mask;
101	pfn_req_flags |= range->default_flags;
102
103	/* We aren't ask to do anything ... */
104	if (!(pfn_req_flags & HMM_PFN_REQ_FAULT))
105		return 0;
106
107	/* Need to write fault ? */
108	if ((pfn_req_flags & HMM_PFN_REQ_WRITE) &&
109	    !(cpu_flags & HMM_PFN_WRITE))
110		return HMM_NEED_FAULT | HMM_NEED_WRITE_FAULT;
111
112	/* If CPU page table is not valid then we need to fault */
113	if (!(cpu_flags & HMM_PFN_VALID))
114		return HMM_NEED_FAULT;
115	return 0;
116}
117
118static unsigned int
119hmm_range_need_fault(const struct hmm_vma_walk *hmm_vma_walk,
120		     const unsigned long hmm_pfns[], unsigned long npages,
121		     unsigned long cpu_flags)
122{
123	struct hmm_range *range = hmm_vma_walk->range;
124	unsigned int required_fault = 0;
125	unsigned long i;
126
127	/*
128	 * If the default flags do not request to fault pages, and the mask does
129	 * not allow for individual pages to be faulted, then
130	 * hmm_pte_need_fault() will always return 0.
131	 */
132	if (!((range->default_flags | range->pfn_flags_mask) &
133	      HMM_PFN_REQ_FAULT))
134		return 0;
135
136	for (i = 0; i < npages; ++i) {
137		required_fault |= hmm_pte_need_fault(hmm_vma_walk, hmm_pfns[i],
138						     cpu_flags);
139		if (required_fault == HMM_NEED_ALL_BITS)
140			return required_fault;
141	}
142	return required_fault;
143}
144
145static int hmm_vma_walk_hole(unsigned long addr, unsigned long end,
146			     __always_unused int depth, struct mm_walk *walk)
147{
148	struct hmm_vma_walk *hmm_vma_walk = walk->private;
149	struct hmm_range *range = hmm_vma_walk->range;
150	unsigned int required_fault;
151	unsigned long i, npages;
152	unsigned long *hmm_pfns;
153
154	i = (addr - range->start) >> PAGE_SHIFT;
155	npages = (end - addr) >> PAGE_SHIFT;
156	hmm_pfns = &range->hmm_pfns[i];
157	required_fault =
158		hmm_range_need_fault(hmm_vma_walk, hmm_pfns, npages, 0);
159	if (!walk->vma) {
160		if (required_fault)
161			return -EFAULT;
162		return hmm_pfns_fill(addr, end, range, HMM_PFN_ERROR);
163	}
164	if (required_fault)
165		return hmm_vma_fault(addr, end, required_fault, walk);
166	return hmm_pfns_fill(addr, end, range, 0);
167}
168
169static inline unsigned long hmm_pfn_flags_order(unsigned long order)
170{
171	return order << HMM_PFN_ORDER_SHIFT;
172}
173
174static inline unsigned long pmd_to_hmm_pfn_flags(struct hmm_range *range,
175						 pmd_t pmd)
176{
177	if (pmd_protnone(pmd))
178		return 0;
179	return (pmd_write(pmd) ? (HMM_PFN_VALID | HMM_PFN_WRITE) :
180				 HMM_PFN_VALID) |
181	       hmm_pfn_flags_order(PMD_SHIFT - PAGE_SHIFT);
182}
183
184#ifdef CONFIG_TRANSPARENT_HUGEPAGE
185static int hmm_vma_handle_pmd(struct mm_walk *walk, unsigned long addr,
186			      unsigned long end, unsigned long hmm_pfns[],
187			      pmd_t pmd)
188{
189	struct hmm_vma_walk *hmm_vma_walk = walk->private;
190	struct hmm_range *range = hmm_vma_walk->range;
191	unsigned long pfn, npages, i;
192	unsigned int required_fault;
193	unsigned long cpu_flags;
194
195	npages = (end - addr) >> PAGE_SHIFT;
196	cpu_flags = pmd_to_hmm_pfn_flags(range, pmd);
197	required_fault =
198		hmm_range_need_fault(hmm_vma_walk, hmm_pfns, npages, cpu_flags);
199	if (required_fault)
200		return hmm_vma_fault(addr, end, required_fault, walk);
201
202	pfn = pmd_pfn(pmd) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
203	for (i = 0; addr < end; addr += PAGE_SIZE, i++, pfn++)
204		hmm_pfns[i] = pfn | cpu_flags;
205	return 0;
206}
207#else /* CONFIG_TRANSPARENT_HUGEPAGE */
208/* stub to allow the code below to compile */
209int hmm_vma_handle_pmd(struct mm_walk *walk, unsigned long addr,
210		unsigned long end, unsigned long hmm_pfns[], pmd_t pmd);
211#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
212
213static inline bool hmm_is_device_private_entry(struct hmm_range *range,
214		swp_entry_t entry)
215{
216	return is_device_private_entry(entry) &&
217		device_private_entry_to_page(entry)->pgmap->owner ==
218		range->dev_private_owner;
219}
220
221static inline unsigned long pte_to_hmm_pfn_flags(struct hmm_range *range,
222						 pte_t pte)
223{
224	if (pte_none(pte) || !pte_present(pte) || pte_protnone(pte))
225		return 0;
226	return pte_write(pte) ? (HMM_PFN_VALID | HMM_PFN_WRITE) : HMM_PFN_VALID;
227}
228
229static int hmm_vma_handle_pte(struct mm_walk *walk, unsigned long addr,
230			      unsigned long end, pmd_t *pmdp, pte_t *ptep,
231			      unsigned long *hmm_pfn)
232{
233	struct hmm_vma_walk *hmm_vma_walk = walk->private;
234	struct hmm_range *range = hmm_vma_walk->range;
235	unsigned int required_fault;
236	unsigned long cpu_flags;
237	pte_t pte = *ptep;
238	uint64_t pfn_req_flags = *hmm_pfn;
239
240	if (pte_none(pte)) {
241		required_fault =
242			hmm_pte_need_fault(hmm_vma_walk, pfn_req_flags, 0);
243		if (required_fault)
244			goto fault;
245		*hmm_pfn = 0;
246		return 0;
247	}
248
249	if (!pte_present(pte)) {
250		swp_entry_t entry = pte_to_swp_entry(pte);
251
252		/*
253		 * Never fault in device private pages, but just report
254		 * the PFN even if not present.
255		 */
256		if (hmm_is_device_private_entry(range, entry)) {
 
 
257			cpu_flags = HMM_PFN_VALID;
258			if (is_write_device_private_entry(entry))
259				cpu_flags |= HMM_PFN_WRITE;
260			*hmm_pfn = device_private_entry_to_pfn(entry) |
261					cpu_flags;
262			return 0;
263		}
264
265		required_fault =
266			hmm_pte_need_fault(hmm_vma_walk, pfn_req_flags, 0);
267		if (!required_fault) {
268			*hmm_pfn = 0;
269			return 0;
270		}
271
272		if (!non_swap_entry(entry))
273			goto fault;
274
 
 
 
 
 
 
275		if (is_migration_entry(entry)) {
276			pte_unmap(ptep);
277			hmm_vma_walk->last = addr;
278			migration_entry_wait(walk->mm, pmdp, addr);
279			return -EBUSY;
280		}
281
282		/* Report error for everything else */
283		pte_unmap(ptep);
284		return -EFAULT;
285	}
286
287	cpu_flags = pte_to_hmm_pfn_flags(range, pte);
288	required_fault =
289		hmm_pte_need_fault(hmm_vma_walk, pfn_req_flags, cpu_flags);
290	if (required_fault)
291		goto fault;
292
293	/*
 
 
294	 * Since each architecture defines a struct page for the zero page, just
295	 * fall through and treat it like a normal page.
296	 */
297	if (pte_special(pte) && !is_zero_pfn(pte_pfn(pte))) {
 
 
298		if (hmm_pte_need_fault(hmm_vma_walk, pfn_req_flags, 0)) {
299			pte_unmap(ptep);
300			return -EFAULT;
301		}
302		*hmm_pfn = HMM_PFN_ERROR;
303		return 0;
304	}
305
306	*hmm_pfn = pte_pfn(pte) | cpu_flags;
307	return 0;
308
309fault:
310	pte_unmap(ptep);
311	/* Fault any virtual address we were asked to fault */
312	return hmm_vma_fault(addr, end, required_fault, walk);
313}
314
315static int hmm_vma_walk_pmd(pmd_t *pmdp,
316			    unsigned long start,
317			    unsigned long end,
318			    struct mm_walk *walk)
319{
320	struct hmm_vma_walk *hmm_vma_walk = walk->private;
321	struct hmm_range *range = hmm_vma_walk->range;
322	unsigned long *hmm_pfns =
323		&range->hmm_pfns[(start - range->start) >> PAGE_SHIFT];
324	unsigned long npages = (end - start) >> PAGE_SHIFT;
325	unsigned long addr = start;
326	pte_t *ptep;
327	pmd_t pmd;
328
329again:
330	pmd = READ_ONCE(*pmdp);
331	if (pmd_none(pmd))
332		return hmm_vma_walk_hole(start, end, -1, walk);
333
334	if (thp_migration_supported() && is_pmd_migration_entry(pmd)) {
335		if (hmm_range_need_fault(hmm_vma_walk, hmm_pfns, npages, 0)) {
336			hmm_vma_walk->last = addr;
337			pmd_migration_entry_wait(walk->mm, pmdp);
338			return -EBUSY;
339		}
340		return hmm_pfns_fill(start, end, range, 0);
341	}
342
343	if (!pmd_present(pmd)) {
344		if (hmm_range_need_fault(hmm_vma_walk, hmm_pfns, npages, 0))
345			return -EFAULT;
346		return hmm_pfns_fill(start, end, range, HMM_PFN_ERROR);
347	}
348
349	if (pmd_devmap(pmd) || pmd_trans_huge(pmd)) {
350		/*
351		 * No need to take pmd_lock here, even if some other thread
352		 * is splitting the huge pmd we will get that event through
353		 * mmu_notifier callback.
354		 *
355		 * So just read pmd value and check again it's a transparent
356		 * huge or device mapping one and compute corresponding pfn
357		 * values.
358		 */
359		pmd = pmd_read_atomic(pmdp);
360		barrier();
361		if (!pmd_devmap(pmd) && !pmd_trans_huge(pmd))
362			goto again;
363
364		return hmm_vma_handle_pmd(walk, addr, end, hmm_pfns, pmd);
365	}
366
367	/*
368	 * We have handled all the valid cases above ie either none, migration,
369	 * huge or transparent huge. At this point either it is a valid pmd
370	 * entry pointing to pte directory or it is a bad pmd that will not
371	 * recover.
372	 */
373	if (pmd_bad(pmd)) {
374		if (hmm_range_need_fault(hmm_vma_walk, hmm_pfns, npages, 0))
375			return -EFAULT;
376		return hmm_pfns_fill(start, end, range, HMM_PFN_ERROR);
377	}
378
379	ptep = pte_offset_map(pmdp, addr);
380	for (; addr < end; addr += PAGE_SIZE, ptep++, hmm_pfns++) {
381		int r;
382
383		r = hmm_vma_handle_pte(walk, addr, end, pmdp, ptep, hmm_pfns);
384		if (r) {
385			/* hmm_vma_handle_pte() did pte_unmap() */
386			return r;
387		}
388	}
389	pte_unmap(ptep - 1);
390	return 0;
391}
392
393#if defined(CONFIG_ARCH_HAS_PTE_DEVMAP) && \
394    defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD)
395static inline unsigned long pud_to_hmm_pfn_flags(struct hmm_range *range,
396						 pud_t pud)
397{
398	if (!pud_present(pud))
399		return 0;
400	return (pud_write(pud) ? (HMM_PFN_VALID | HMM_PFN_WRITE) :
401				 HMM_PFN_VALID) |
402	       hmm_pfn_flags_order(PUD_SHIFT - PAGE_SHIFT);
403}
404
405static int hmm_vma_walk_pud(pud_t *pudp, unsigned long start, unsigned long end,
406		struct mm_walk *walk)
407{
408	struct hmm_vma_walk *hmm_vma_walk = walk->private;
409	struct hmm_range *range = hmm_vma_walk->range;
410	unsigned long addr = start;
411	pud_t pud;
412	int ret = 0;
413	spinlock_t *ptl = pud_trans_huge_lock(pudp, walk->vma);
414
415	if (!ptl)
416		return 0;
417
418	/* Normally we don't want to split the huge page */
419	walk->action = ACTION_CONTINUE;
420
421	pud = READ_ONCE(*pudp);
422	if (pud_none(pud)) {
423		spin_unlock(ptl);
424		return hmm_vma_walk_hole(start, end, -1, walk);
425	}
426
427	if (pud_huge(pud) && pud_devmap(pud)) {
428		unsigned long i, npages, pfn;
429		unsigned int required_fault;
430		unsigned long *hmm_pfns;
431		unsigned long cpu_flags;
432
433		if (!pud_present(pud)) {
434			spin_unlock(ptl);
435			return hmm_vma_walk_hole(start, end, -1, walk);
436		}
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
459out_unlock:
460	spin_unlock(ptl);
461	return ret;
462}
463#else
464#define hmm_vma_walk_pud	NULL
465#endif
466
467#ifdef CONFIG_HUGETLB_PAGE
468static 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(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		spin_unlock(ptl);
493		return hmm_vma_fault(addr, end, required_fault, walk);
494	}
495
496	pfn = pte_pfn(entry) + ((start & ~hmask) >> PAGE_SHIFT);
497	for (; addr < end; addr += PAGE_SIZE, i++, pfn++)
498		range->hmm_pfns[i] = pfn | cpu_flags;
499
500	spin_unlock(ptl);
501	return 0;
502}
503#else
504#define hmm_vma_walk_hugetlb_entry NULL
505#endif /* CONFIG_HUGETLB_PAGE */
506
507static int hmm_vma_walk_test(unsigned long start, unsigned long end,
508			     struct mm_walk *walk)
509{
510	struct hmm_vma_walk *hmm_vma_walk = walk->private;
511	struct hmm_range *range = hmm_vma_walk->range;
512	struct vm_area_struct *vma = walk->vma;
513
514	if (!(vma->vm_flags & (VM_IO | VM_PFNMAP | VM_MIXEDMAP)) &&
515	    vma->vm_flags & VM_READ)
516		return 0;
517
518	/*
519	 * vma ranges that don't have struct page backing them or map I/O
520	 * devices directly cannot be handled by hmm_range_fault().
521	 *
522	 * If the vma does not allow read access, then assume that it does not
523	 * allow write access either. HMM does not support architectures that
524	 * allow write without read.
525	 *
526	 * If a fault is requested for an unsupported range then it is a hard
527	 * failure.
528	 */
529	if (hmm_range_need_fault(hmm_vma_walk,
530				 range->hmm_pfns +
531					 ((start - range->start) >> PAGE_SHIFT),
532				 (end - start) >> PAGE_SHIFT, 0))
533		return -EFAULT;
534
535	hmm_pfns_fill(start, end, range, HMM_PFN_ERROR);
536
537	/* Skip this vma and continue processing the next vma. */
538	return 1;
539}
540
541static const struct mm_walk_ops hmm_walk_ops = {
542	.pud_entry	= hmm_vma_walk_pud,
543	.pmd_entry	= hmm_vma_walk_pmd,
544	.pte_hole	= hmm_vma_walk_hole,
545	.hugetlb_entry	= hmm_vma_walk_hugetlb_entry,
546	.test_walk	= hmm_vma_walk_test,
547};
548
549/**
550 * hmm_range_fault - try to fault some address in a virtual address range
551 * @range:	argument structure
552 *
553 * Returns 0 on success or one of the following error codes:
554 *
555 * -EINVAL:	Invalid arguments or mm or virtual address is in an invalid vma
556 *		(e.g., device file vma).
557 * -ENOMEM:	Out of memory.
558 * -EPERM:	Invalid permission (e.g., asking for write and range is read
559 *		only).
560 * -EBUSY:	The range has been invalidated and the caller needs to wait for
561 *		the invalidation to finish.
562 * -EFAULT:     A page was requested to be valid and could not be made valid
563 *              ie it has no backing VMA or it is illegal to access
564 *
565 * This is similar to get_user_pages(), except that it can read the page tables
566 * without mutating them (ie causing faults).
567 */
568int hmm_range_fault(struct hmm_range *range)
569{
570	struct hmm_vma_walk hmm_vma_walk = {
571		.range = range,
572		.last = range->start,
573	};
574	struct mm_struct *mm = range->notifier->mm;
575	int ret;
576
577	mmap_assert_locked(mm);
578
579	do {
580		/* If range is no longer valid force retry. */
581		if (mmu_interval_check_retry(range->notifier,
582					     range->notifier_seq))
583			return -EBUSY;
584		ret = walk_page_range(mm, hmm_vma_walk.last, range->end,
585				      &hmm_walk_ops, &hmm_vma_walk);
586		/*
587		 * When -EBUSY is returned the loop restarts with
588		 * hmm_vma_walk.last set to an address that has not been stored
589		 * in pfns. All entries < last in the pfn array are set to their
590		 * output, and all >= are still at their input values.
591		 */
592	} while (ret == -EBUSY);
593	return ret;
594}
595EXPORT_SYMBOL(hmm_range_fault);