mm/hmm/devmem: device memory hotplug using ZONE_DEVICE
[sfrench/cifs-2.6.git] / mm / hmm.c
1 /*
2  * Copyright 2013 Red Hat Inc.
3  *
4  * This program is free software; you can redistribute it and/or modify
5  * it under the terms of the GNU General Public License as published by
6  * the Free Software Foundation; either version 2 of the License, or
7  * (at your option) any later version.
8  *
9  * This program is distributed in the hope that it will be useful,
10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
12  * GNU General Public License for more details.
13  *
14  * Authors: Jérôme Glisse <jglisse@redhat.com>
15  */
16 /*
17  * Refer to include/linux/hmm.h for information about heterogeneous memory
18  * management or HMM for short.
19  */
20 #include <linux/mm.h>
21 #include <linux/hmm.h>
22 #include <linux/rmap.h>
23 #include <linux/swap.h>
24 #include <linux/slab.h>
25 #include <linux/sched.h>
26 #include <linux/mmzone.h>
27 #include <linux/pagemap.h>
28 #include <linux/swapops.h>
29 #include <linux/hugetlb.h>
30 #include <linux/memremap.h>
31 #include <linux/jump_label.h>
32 #include <linux/mmu_notifier.h>
33 #include <linux/memory_hotplug.h>
34
35 #define PA_SECTION_SIZE (1UL << PA_SECTION_SHIFT)
36
37
38 /*
39  * Device private memory see HMM (Documentation/vm/hmm.txt) or hmm.h
40  */
41 DEFINE_STATIC_KEY_FALSE(device_private_key);
42 EXPORT_SYMBOL(device_private_key);
43
44
45 #ifdef CONFIG_HMM
46 static const struct mmu_notifier_ops hmm_mmu_notifier_ops;
47
48 /*
49  * struct hmm - HMM per mm struct
50  *
51  * @mm: mm struct this HMM struct is bound to
52  * @lock: lock protecting ranges list
53  * @sequence: we track updates to the CPU page table with a sequence number
54  * @ranges: list of range being snapshotted
55  * @mirrors: list of mirrors for this mm
56  * @mmu_notifier: mmu notifier to track updates to CPU page table
57  * @mirrors_sem: read/write semaphore protecting the mirrors list
58  */
59 struct hmm {
60         struct mm_struct        *mm;
61         spinlock_t              lock;
62         atomic_t                sequence;
63         struct list_head        ranges;
64         struct list_head        mirrors;
65         struct mmu_notifier     mmu_notifier;
66         struct rw_semaphore     mirrors_sem;
67 };
68
69 /*
70  * hmm_register - register HMM against an mm (HMM internal)
71  *
72  * @mm: mm struct to attach to
73  *
74  * This is not intended to be used directly by device drivers. It allocates an
75  * HMM struct if mm does not have one, and initializes it.
76  */
77 static struct hmm *hmm_register(struct mm_struct *mm)
78 {
79         struct hmm *hmm = READ_ONCE(mm->hmm);
80         bool cleanup = false;
81
82         /*
83          * The hmm struct can only be freed once the mm_struct goes away,
84          * hence we should always have pre-allocated an new hmm struct
85          * above.
86          */
87         if (hmm)
88                 return hmm;
89
90         hmm = kmalloc(sizeof(*hmm), GFP_KERNEL);
91         if (!hmm)
92                 return NULL;
93         INIT_LIST_HEAD(&hmm->mirrors);
94         init_rwsem(&hmm->mirrors_sem);
95         atomic_set(&hmm->sequence, 0);
96         hmm->mmu_notifier.ops = NULL;
97         INIT_LIST_HEAD(&hmm->ranges);
98         spin_lock_init(&hmm->lock);
99         hmm->mm = mm;
100
101         /*
102          * We should only get here if hold the mmap_sem in write mode ie on
103          * registration of first mirror through hmm_mirror_register()
104          */
105         hmm->mmu_notifier.ops = &hmm_mmu_notifier_ops;
106         if (__mmu_notifier_register(&hmm->mmu_notifier, mm)) {
107                 kfree(hmm);
108                 return NULL;
109         }
110
111         spin_lock(&mm->page_table_lock);
112         if (!mm->hmm)
113                 mm->hmm = hmm;
114         else
115                 cleanup = true;
116         spin_unlock(&mm->page_table_lock);
117
118         if (cleanup) {
119                 mmu_notifier_unregister(&hmm->mmu_notifier, mm);
120                 kfree(hmm);
121         }
122
123         return mm->hmm;
124 }
125
126 void hmm_mm_destroy(struct mm_struct *mm)
127 {
128         kfree(mm->hmm);
129 }
130 #endif /* CONFIG_HMM */
131
132 #if IS_ENABLED(CONFIG_HMM_MIRROR)
133 static void hmm_invalidate_range(struct hmm *hmm,
134                                  enum hmm_update_type action,
135                                  unsigned long start,
136                                  unsigned long end)
137 {
138         struct hmm_mirror *mirror;
139         struct hmm_range *range;
140
141         spin_lock(&hmm->lock);
142         list_for_each_entry(range, &hmm->ranges, list) {
143                 unsigned long addr, idx, npages;
144
145                 if (end < range->start || start >= range->end)
146                         continue;
147
148                 range->valid = false;
149                 addr = max(start, range->start);
150                 idx = (addr - range->start) >> PAGE_SHIFT;
151                 npages = (min(range->end, end) - addr) >> PAGE_SHIFT;
152                 memset(&range->pfns[idx], 0, sizeof(*range->pfns) * npages);
153         }
154         spin_unlock(&hmm->lock);
155
156         down_read(&hmm->mirrors_sem);
157         list_for_each_entry(mirror, &hmm->mirrors, list)
158                 mirror->ops->sync_cpu_device_pagetables(mirror, action,
159                                                         start, end);
160         up_read(&hmm->mirrors_sem);
161 }
162
163 static void hmm_invalidate_range_start(struct mmu_notifier *mn,
164                                        struct mm_struct *mm,
165                                        unsigned long start,
166                                        unsigned long end)
167 {
168         struct hmm *hmm = mm->hmm;
169
170         VM_BUG_ON(!hmm);
171
172         atomic_inc(&hmm->sequence);
173 }
174
175 static void hmm_invalidate_range_end(struct mmu_notifier *mn,
176                                      struct mm_struct *mm,
177                                      unsigned long start,
178                                      unsigned long end)
179 {
180         struct hmm *hmm = mm->hmm;
181
182         VM_BUG_ON(!hmm);
183
184         hmm_invalidate_range(mm->hmm, HMM_UPDATE_INVALIDATE, start, end);
185 }
186
187 static const struct mmu_notifier_ops hmm_mmu_notifier_ops = {
188         .invalidate_range_start = hmm_invalidate_range_start,
189         .invalidate_range_end   = hmm_invalidate_range_end,
190 };
191
192 /*
193  * hmm_mirror_register() - register a mirror against an mm
194  *
195  * @mirror: new mirror struct to register
196  * @mm: mm to register against
197  *
198  * To start mirroring a process address space, the device driver must register
199  * an HMM mirror struct.
200  *
201  * THE mm->mmap_sem MUST BE HELD IN WRITE MODE !
202  */
203 int hmm_mirror_register(struct hmm_mirror *mirror, struct mm_struct *mm)
204 {
205         /* Sanity check */
206         if (!mm || !mirror || !mirror->ops)
207                 return -EINVAL;
208
209         mirror->hmm = hmm_register(mm);
210         if (!mirror->hmm)
211                 return -ENOMEM;
212
213         down_write(&mirror->hmm->mirrors_sem);
214         list_add(&mirror->list, &mirror->hmm->mirrors);
215         up_write(&mirror->hmm->mirrors_sem);
216
217         return 0;
218 }
219 EXPORT_SYMBOL(hmm_mirror_register);
220
221 /*
222  * hmm_mirror_unregister() - unregister a mirror
223  *
224  * @mirror: new mirror struct to register
225  *
226  * Stop mirroring a process address space, and cleanup.
227  */
228 void hmm_mirror_unregister(struct hmm_mirror *mirror)
229 {
230         struct hmm *hmm = mirror->hmm;
231
232         down_write(&hmm->mirrors_sem);
233         list_del(&mirror->list);
234         up_write(&hmm->mirrors_sem);
235 }
236 EXPORT_SYMBOL(hmm_mirror_unregister);
237
238 struct hmm_vma_walk {
239         struct hmm_range        *range;
240         unsigned long           last;
241         bool                    fault;
242         bool                    block;
243         bool                    write;
244 };
245
246 static int hmm_vma_do_fault(struct mm_walk *walk,
247                             unsigned long addr,
248                             hmm_pfn_t *pfn)
249 {
250         unsigned int flags = FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_REMOTE;
251         struct hmm_vma_walk *hmm_vma_walk = walk->private;
252         struct vm_area_struct *vma = walk->vma;
253         int r;
254
255         flags |= hmm_vma_walk->block ? 0 : FAULT_FLAG_ALLOW_RETRY;
256         flags |= hmm_vma_walk->write ? FAULT_FLAG_WRITE : 0;
257         r = handle_mm_fault(vma, addr, flags);
258         if (r & VM_FAULT_RETRY)
259                 return -EBUSY;
260         if (r & VM_FAULT_ERROR) {
261                 *pfn = HMM_PFN_ERROR;
262                 return -EFAULT;
263         }
264
265         return -EAGAIN;
266 }
267
268 static void hmm_pfns_special(hmm_pfn_t *pfns,
269                              unsigned long addr,
270                              unsigned long end)
271 {
272         for (; addr < end; addr += PAGE_SIZE, pfns++)
273                 *pfns = HMM_PFN_SPECIAL;
274 }
275
276 static int hmm_pfns_bad(unsigned long addr,
277                         unsigned long end,
278                         struct mm_walk *walk)
279 {
280         struct hmm_range *range = walk->private;
281         hmm_pfn_t *pfns = range->pfns;
282         unsigned long i;
283
284         i = (addr - range->start) >> PAGE_SHIFT;
285         for (; addr < end; addr += PAGE_SIZE, i++)
286                 pfns[i] = HMM_PFN_ERROR;
287
288         return 0;
289 }
290
291 static void hmm_pfns_clear(hmm_pfn_t *pfns,
292                            unsigned long addr,
293                            unsigned long end)
294 {
295         for (; addr < end; addr += PAGE_SIZE, pfns++)
296                 *pfns = 0;
297 }
298
299 static int hmm_vma_walk_hole(unsigned long addr,
300                              unsigned long end,
301                              struct mm_walk *walk)
302 {
303         struct hmm_vma_walk *hmm_vma_walk = walk->private;
304         struct hmm_range *range = hmm_vma_walk->range;
305         hmm_pfn_t *pfns = range->pfns;
306         unsigned long i;
307
308         hmm_vma_walk->last = addr;
309         i = (addr - range->start) >> PAGE_SHIFT;
310         for (; addr < end; addr += PAGE_SIZE, i++) {
311                 pfns[i] = HMM_PFN_EMPTY;
312                 if (hmm_vma_walk->fault) {
313                         int ret;
314
315                         ret = hmm_vma_do_fault(walk, addr, &pfns[i]);
316                         if (ret != -EAGAIN)
317                                 return ret;
318                 }
319         }
320
321         return hmm_vma_walk->fault ? -EAGAIN : 0;
322 }
323
324 static int hmm_vma_walk_clear(unsigned long addr,
325                               unsigned long end,
326                               struct mm_walk *walk)
327 {
328         struct hmm_vma_walk *hmm_vma_walk = walk->private;
329         struct hmm_range *range = hmm_vma_walk->range;
330         hmm_pfn_t *pfns = range->pfns;
331         unsigned long i;
332
333         hmm_vma_walk->last = addr;
334         i = (addr - range->start) >> PAGE_SHIFT;
335         for (; addr < end; addr += PAGE_SIZE, i++) {
336                 pfns[i] = 0;
337                 if (hmm_vma_walk->fault) {
338                         int ret;
339
340                         ret = hmm_vma_do_fault(walk, addr, &pfns[i]);
341                         if (ret != -EAGAIN)
342                                 return ret;
343                 }
344         }
345
346         return hmm_vma_walk->fault ? -EAGAIN : 0;
347 }
348
349 static int hmm_vma_walk_pmd(pmd_t *pmdp,
350                             unsigned long start,
351                             unsigned long end,
352                             struct mm_walk *walk)
353 {
354         struct hmm_vma_walk *hmm_vma_walk = walk->private;
355         struct hmm_range *range = hmm_vma_walk->range;
356         struct vm_area_struct *vma = walk->vma;
357         hmm_pfn_t *pfns = range->pfns;
358         unsigned long addr = start, i;
359         bool write_fault;
360         hmm_pfn_t flag;
361         pte_t *ptep;
362
363         i = (addr - range->start) >> PAGE_SHIFT;
364         flag = vma->vm_flags & VM_READ ? HMM_PFN_READ : 0;
365         write_fault = hmm_vma_walk->fault & hmm_vma_walk->write;
366
367 again:
368         if (pmd_none(*pmdp))
369                 return hmm_vma_walk_hole(start, end, walk);
370
371         if (pmd_huge(*pmdp) && vma->vm_flags & VM_HUGETLB)
372                 return hmm_pfns_bad(start, end, walk);
373
374         if (pmd_devmap(*pmdp) || pmd_trans_huge(*pmdp)) {
375                 unsigned long pfn;
376                 pmd_t pmd;
377
378                 /*
379                  * No need to take pmd_lock here, even if some other threads
380                  * is splitting the huge pmd we will get that event through
381                  * mmu_notifier callback.
382                  *
383                  * So just read pmd value and check again its a transparent
384                  * huge or device mapping one and compute corresponding pfn
385                  * values.
386                  */
387                 pmd = pmd_read_atomic(pmdp);
388                 barrier();
389                 if (!pmd_devmap(pmd) && !pmd_trans_huge(pmd))
390                         goto again;
391                 if (pmd_protnone(pmd))
392                         return hmm_vma_walk_clear(start, end, walk);
393
394                 if (write_fault && !pmd_write(pmd))
395                         return hmm_vma_walk_clear(start, end, walk);
396
397                 pfn = pmd_pfn(pmd) + pte_index(addr);
398                 flag |= pmd_write(pmd) ? HMM_PFN_WRITE : 0;
399                 for (; addr < end; addr += PAGE_SIZE, i++, pfn++)
400                         pfns[i] = hmm_pfn_t_from_pfn(pfn) | flag;
401                 return 0;
402         }
403
404         if (pmd_bad(*pmdp))
405                 return hmm_pfns_bad(start, end, walk);
406
407         ptep = pte_offset_map(pmdp, addr);
408         for (; addr < end; addr += PAGE_SIZE, ptep++, i++) {
409                 pte_t pte = *ptep;
410
411                 pfns[i] = 0;
412
413                 if (pte_none(pte)) {
414                         pfns[i] = HMM_PFN_EMPTY;
415                         if (hmm_vma_walk->fault)
416                                 goto fault;
417                         continue;
418                 }
419
420                 if (!pte_present(pte)) {
421                         swp_entry_t entry;
422
423                         if (!non_swap_entry(entry)) {
424                                 if (hmm_vma_walk->fault)
425                                         goto fault;
426                                 continue;
427                         }
428
429                         entry = pte_to_swp_entry(pte);
430
431                         /*
432                          * This is a special swap entry, ignore migration, use
433                          * device and report anything else as error.
434                          */
435                         if (is_device_private_entry(entry)) {
436                                 pfns[i] = hmm_pfn_t_from_pfn(swp_offset(entry));
437                                 if (is_write_device_private_entry(entry)) {
438                                         pfns[i] |= HMM_PFN_WRITE;
439                                 } else if (write_fault)
440                                         goto fault;
441                                 pfns[i] |= HMM_PFN_DEVICE_UNADDRESSABLE;
442                                 pfns[i] |= flag;
443                         } else if (is_migration_entry(entry)) {
444                                 if (hmm_vma_walk->fault) {
445                                         pte_unmap(ptep);
446                                         hmm_vma_walk->last = addr;
447                                         migration_entry_wait(vma->vm_mm,
448                                                              pmdp, addr);
449                                         return -EAGAIN;
450                                 }
451                                 continue;
452                         } else {
453                                 /* Report error for everything else */
454                                 pfns[i] = HMM_PFN_ERROR;
455                         }
456                         continue;
457                 }
458
459                 if (write_fault && !pte_write(pte))
460                         goto fault;
461
462                 pfns[i] = hmm_pfn_t_from_pfn(pte_pfn(pte)) | flag;
463                 pfns[i] |= pte_write(pte) ? HMM_PFN_WRITE : 0;
464                 continue;
465
466 fault:
467                 pte_unmap(ptep);
468                 /* Fault all pages in range */
469                 return hmm_vma_walk_clear(start, end, walk);
470         }
471         pte_unmap(ptep - 1);
472
473         return 0;
474 }
475
476 /*
477  * hmm_vma_get_pfns() - snapshot CPU page table for a range of virtual addresses
478  * @vma: virtual memory area containing the virtual address range
479  * @range: used to track snapshot validity
480  * @start: range virtual start address (inclusive)
481  * @end: range virtual end address (exclusive)
482  * @entries: array of hmm_pfn_t: provided by the caller, filled in by function
483  * Returns: -EINVAL if invalid argument, -ENOMEM out of memory, 0 success
484  *
485  * This snapshots the CPU page table for a range of virtual addresses. Snapshot
486  * validity is tracked by range struct. See hmm_vma_range_done() for further
487  * information.
488  *
489  * The range struct is initialized here. It tracks the CPU page table, but only
490  * if the function returns success (0), in which case the caller must then call
491  * hmm_vma_range_done() to stop CPU page table update tracking on this range.
492  *
493  * NOT CALLING hmm_vma_range_done() IF FUNCTION RETURNS 0 WILL LEAD TO SERIOUS
494  * MEMORY CORRUPTION ! YOU HAVE BEEN WARNED !
495  */
496 int hmm_vma_get_pfns(struct vm_area_struct *vma,
497                      struct hmm_range *range,
498                      unsigned long start,
499                      unsigned long end,
500                      hmm_pfn_t *pfns)
501 {
502         struct hmm_vma_walk hmm_vma_walk;
503         struct mm_walk mm_walk;
504         struct hmm *hmm;
505
506         /* FIXME support hugetlb fs */
507         if (is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_SPECIAL)) {
508                 hmm_pfns_special(pfns, start, end);
509                 return -EINVAL;
510         }
511
512         /* Sanity check, this really should not happen ! */
513         if (start < vma->vm_start || start >= vma->vm_end)
514                 return -EINVAL;
515         if (end < vma->vm_start || end > vma->vm_end)
516                 return -EINVAL;
517
518         hmm = hmm_register(vma->vm_mm);
519         if (!hmm)
520                 return -ENOMEM;
521         /* Caller must have registered a mirror, via hmm_mirror_register() ! */
522         if (!hmm->mmu_notifier.ops)
523                 return -EINVAL;
524
525         /* Initialize range to track CPU page table update */
526         range->start = start;
527         range->pfns = pfns;
528         range->end = end;
529         spin_lock(&hmm->lock);
530         range->valid = true;
531         list_add_rcu(&range->list, &hmm->ranges);
532         spin_unlock(&hmm->lock);
533
534         hmm_vma_walk.fault = false;
535         hmm_vma_walk.range = range;
536         mm_walk.private = &hmm_vma_walk;
537
538         mm_walk.vma = vma;
539         mm_walk.mm = vma->vm_mm;
540         mm_walk.pte_entry = NULL;
541         mm_walk.test_walk = NULL;
542         mm_walk.hugetlb_entry = NULL;
543         mm_walk.pmd_entry = hmm_vma_walk_pmd;
544         mm_walk.pte_hole = hmm_vma_walk_hole;
545
546         walk_page_range(start, end, &mm_walk);
547         return 0;
548 }
549 EXPORT_SYMBOL(hmm_vma_get_pfns);
550
551 /*
552  * hmm_vma_range_done() - stop tracking change to CPU page table over a range
553  * @vma: virtual memory area containing the virtual address range
554  * @range: range being tracked
555  * Returns: false if range data has been invalidated, true otherwise
556  *
557  * Range struct is used to track updates to the CPU page table after a call to
558  * either hmm_vma_get_pfns() or hmm_vma_fault(). Once the device driver is done
559  * using the data,  or wants to lock updates to the data it got from those
560  * functions, it must call the hmm_vma_range_done() function, which will then
561  * stop tracking CPU page table updates.
562  *
563  * Note that device driver must still implement general CPU page table update
564  * tracking either by using hmm_mirror (see hmm_mirror_register()) or by using
565  * the mmu_notifier API directly.
566  *
567  * CPU page table update tracking done through hmm_range is only temporary and
568  * to be used while trying to duplicate CPU page table contents for a range of
569  * virtual addresses.
570  *
571  * There are two ways to use this :
572  * again:
573  *   hmm_vma_get_pfns(vma, range, start, end, pfns); or hmm_vma_fault(...);
574  *   trans = device_build_page_table_update_transaction(pfns);
575  *   device_page_table_lock();
576  *   if (!hmm_vma_range_done(vma, range)) {
577  *     device_page_table_unlock();
578  *     goto again;
579  *   }
580  *   device_commit_transaction(trans);
581  *   device_page_table_unlock();
582  *
583  * Or:
584  *   hmm_vma_get_pfns(vma, range, start, end, pfns); or hmm_vma_fault(...);
585  *   device_page_table_lock();
586  *   hmm_vma_range_done(vma, range);
587  *   device_update_page_table(pfns);
588  *   device_page_table_unlock();
589  */
590 bool hmm_vma_range_done(struct vm_area_struct *vma, struct hmm_range *range)
591 {
592         unsigned long npages = (range->end - range->start) >> PAGE_SHIFT;
593         struct hmm *hmm;
594
595         if (range->end <= range->start) {
596                 BUG();
597                 return false;
598         }
599
600         hmm = hmm_register(vma->vm_mm);
601         if (!hmm) {
602                 memset(range->pfns, 0, sizeof(*range->pfns) * npages);
603                 return false;
604         }
605
606         spin_lock(&hmm->lock);
607         list_del_rcu(&range->list);
608         spin_unlock(&hmm->lock);
609
610         return range->valid;
611 }
612 EXPORT_SYMBOL(hmm_vma_range_done);
613
614 /*
615  * hmm_vma_fault() - try to fault some address in a virtual address range
616  * @vma: virtual memory area containing the virtual address range
617  * @range: use to track pfns array content validity
618  * @start: fault range virtual start address (inclusive)
619  * @end: fault range virtual end address (exclusive)
620  * @pfns: array of hmm_pfn_t, only entry with fault flag set will be faulted
621  * @write: is it a write fault
622  * @block: allow blocking on fault (if true it sleeps and do not drop mmap_sem)
623  * Returns: 0 success, error otherwise (-EAGAIN means mmap_sem have been drop)
624  *
625  * This is similar to a regular CPU page fault except that it will not trigger
626  * any memory migration if the memory being faulted is not accessible by CPUs.
627  *
628  * On error, for one virtual address in the range, the function will set the
629  * hmm_pfn_t error flag for the corresponding pfn entry.
630  *
631  * Expected use pattern:
632  * retry:
633  *   down_read(&mm->mmap_sem);
634  *   // Find vma and address device wants to fault, initialize hmm_pfn_t
635  *   // array accordingly
636  *   ret = hmm_vma_fault(vma, start, end, pfns, allow_retry);
637  *   switch (ret) {
638  *   case -EAGAIN:
639  *     hmm_vma_range_done(vma, range);
640  *     // You might want to rate limit or yield to play nicely, you may
641  *     // also commit any valid pfn in the array assuming that you are
642  *     // getting true from hmm_vma_range_monitor_end()
643  *     goto retry;
644  *   case 0:
645  *     break;
646  *   default:
647  *     // Handle error !
648  *     up_read(&mm->mmap_sem)
649  *     return;
650  *   }
651  *   // Take device driver lock that serialize device page table update
652  *   driver_lock_device_page_table_update();
653  *   hmm_vma_range_done(vma, range);
654  *   // Commit pfns we got from hmm_vma_fault()
655  *   driver_unlock_device_page_table_update();
656  *   up_read(&mm->mmap_sem)
657  *
658  * YOU MUST CALL hmm_vma_range_done() AFTER THIS FUNCTION RETURN SUCCESS (0)
659  * BEFORE FREEING THE range struct OR YOU WILL HAVE SERIOUS MEMORY CORRUPTION !
660  *
661  * YOU HAVE BEEN WARNED !
662  */
663 int hmm_vma_fault(struct vm_area_struct *vma,
664                   struct hmm_range *range,
665                   unsigned long start,
666                   unsigned long end,
667                   hmm_pfn_t *pfns,
668                   bool write,
669                   bool block)
670 {
671         struct hmm_vma_walk hmm_vma_walk;
672         struct mm_walk mm_walk;
673         struct hmm *hmm;
674         int ret;
675
676         /* Sanity check, this really should not happen ! */
677         if (start < vma->vm_start || start >= vma->vm_end)
678                 return -EINVAL;
679         if (end < vma->vm_start || end > vma->vm_end)
680                 return -EINVAL;
681
682         hmm = hmm_register(vma->vm_mm);
683         if (!hmm) {
684                 hmm_pfns_clear(pfns, start, end);
685                 return -ENOMEM;
686         }
687         /* Caller must have registered a mirror using hmm_mirror_register() */
688         if (!hmm->mmu_notifier.ops)
689                 return -EINVAL;
690
691         /* Initialize range to track CPU page table update */
692         range->start = start;
693         range->pfns = pfns;
694         range->end = end;
695         spin_lock(&hmm->lock);
696         range->valid = true;
697         list_add_rcu(&range->list, &hmm->ranges);
698         spin_unlock(&hmm->lock);
699
700         /* FIXME support hugetlb fs */
701         if (is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_SPECIAL)) {
702                 hmm_pfns_special(pfns, start, end);
703                 return 0;
704         }
705
706         hmm_vma_walk.fault = true;
707         hmm_vma_walk.write = write;
708         hmm_vma_walk.block = block;
709         hmm_vma_walk.range = range;
710         mm_walk.private = &hmm_vma_walk;
711         hmm_vma_walk.last = range->start;
712
713         mm_walk.vma = vma;
714         mm_walk.mm = vma->vm_mm;
715         mm_walk.pte_entry = NULL;
716         mm_walk.test_walk = NULL;
717         mm_walk.hugetlb_entry = NULL;
718         mm_walk.pmd_entry = hmm_vma_walk_pmd;
719         mm_walk.pte_hole = hmm_vma_walk_hole;
720
721         do {
722                 ret = walk_page_range(start, end, &mm_walk);
723                 start = hmm_vma_walk.last;
724         } while (ret == -EAGAIN);
725
726         if (ret) {
727                 unsigned long i;
728
729                 i = (hmm_vma_walk.last - range->start) >> PAGE_SHIFT;
730                 hmm_pfns_clear(&pfns[i], hmm_vma_walk.last, end);
731                 hmm_vma_range_done(vma, range);
732         }
733         return ret;
734 }
735 EXPORT_SYMBOL(hmm_vma_fault);
736 #endif /* IS_ENABLED(CONFIG_HMM_MIRROR) */
737
738
739 #if IS_ENABLED(CONFIG_DEVICE_PRIVATE)
740 struct page *hmm_vma_alloc_locked_page(struct vm_area_struct *vma,
741                                        unsigned long addr)
742 {
743         struct page *page;
744
745         page = alloc_page_vma(GFP_HIGHUSER, vma, addr);
746         if (!page)
747                 return NULL;
748         lock_page(page);
749         return page;
750 }
751 EXPORT_SYMBOL(hmm_vma_alloc_locked_page);
752
753
754 static void hmm_devmem_ref_release(struct percpu_ref *ref)
755 {
756         struct hmm_devmem *devmem;
757
758         devmem = container_of(ref, struct hmm_devmem, ref);
759         complete(&devmem->completion);
760 }
761
762 static void hmm_devmem_ref_exit(void *data)
763 {
764         struct percpu_ref *ref = data;
765         struct hmm_devmem *devmem;
766
767         devmem = container_of(ref, struct hmm_devmem, ref);
768         percpu_ref_exit(ref);
769         devm_remove_action(devmem->device, &hmm_devmem_ref_exit, data);
770 }
771
772 static void hmm_devmem_ref_kill(void *data)
773 {
774         struct percpu_ref *ref = data;
775         struct hmm_devmem *devmem;
776
777         devmem = container_of(ref, struct hmm_devmem, ref);
778         percpu_ref_kill(ref);
779         wait_for_completion(&devmem->completion);
780         devm_remove_action(devmem->device, &hmm_devmem_ref_kill, data);
781 }
782
783 static int hmm_devmem_fault(struct vm_area_struct *vma,
784                             unsigned long addr,
785                             const struct page *page,
786                             unsigned int flags,
787                             pmd_t *pmdp)
788 {
789         struct hmm_devmem *devmem = page->pgmap->data;
790
791         return devmem->ops->fault(devmem, vma, addr, page, flags, pmdp);
792 }
793
794 static void hmm_devmem_free(struct page *page, void *data)
795 {
796         struct hmm_devmem *devmem = data;
797
798         devmem->ops->free(devmem, page);
799 }
800
801 static DEFINE_MUTEX(hmm_devmem_lock);
802 static RADIX_TREE(hmm_devmem_radix, GFP_KERNEL);
803
804 static void hmm_devmem_radix_release(struct resource *resource)
805 {
806         resource_size_t key, align_start, align_size, align_end;
807
808         align_start = resource->start & ~(PA_SECTION_SIZE - 1);
809         align_size = ALIGN(resource_size(resource), PA_SECTION_SIZE);
810         align_end = align_start + align_size - 1;
811
812         mutex_lock(&hmm_devmem_lock);
813         for (key = resource->start;
814              key <= resource->end;
815              key += PA_SECTION_SIZE)
816                 radix_tree_delete(&hmm_devmem_radix, key >> PA_SECTION_SHIFT);
817         mutex_unlock(&hmm_devmem_lock);
818 }
819
820 static void hmm_devmem_release(struct device *dev, void *data)
821 {
822         struct hmm_devmem *devmem = data;
823         struct resource *resource = devmem->resource;
824         unsigned long start_pfn, npages;
825         struct zone *zone;
826         struct page *page;
827
828         if (percpu_ref_tryget_live(&devmem->ref)) {
829                 dev_WARN(dev, "%s: page mapping is still live!\n", __func__);
830                 percpu_ref_put(&devmem->ref);
831         }
832
833         /* pages are dead and unused, undo the arch mapping */
834         start_pfn = (resource->start & ~(PA_SECTION_SIZE - 1)) >> PAGE_SHIFT;
835         npages = ALIGN(resource_size(resource), PA_SECTION_SIZE) >> PAGE_SHIFT;
836
837         page = pfn_to_page(start_pfn);
838         zone = page_zone(page);
839
840         mem_hotplug_begin();
841         __remove_pages(zone, start_pfn, npages);
842         mem_hotplug_done();
843
844         hmm_devmem_radix_release(resource);
845 }
846
847 static struct hmm_devmem *hmm_devmem_find(resource_size_t phys)
848 {
849         WARN_ON_ONCE(!rcu_read_lock_held());
850
851         return radix_tree_lookup(&hmm_devmem_radix, phys >> PA_SECTION_SHIFT);
852 }
853
854 static int hmm_devmem_pages_create(struct hmm_devmem *devmem)
855 {
856         resource_size_t key, align_start, align_size, align_end;
857         struct device *device = devmem->device;
858         int ret, nid, is_ram;
859         unsigned long pfn;
860
861         align_start = devmem->resource->start & ~(PA_SECTION_SIZE - 1);
862         align_size = ALIGN(devmem->resource->start +
863                            resource_size(devmem->resource),
864                            PA_SECTION_SIZE) - align_start;
865
866         is_ram = region_intersects(align_start, align_size,
867                                    IORESOURCE_SYSTEM_RAM,
868                                    IORES_DESC_NONE);
869         if (is_ram == REGION_MIXED) {
870                 WARN_ONCE(1, "%s attempted on mixed region %pr\n",
871                                 __func__, devmem->resource);
872                 return -ENXIO;
873         }
874         if (is_ram == REGION_INTERSECTS)
875                 return -ENXIO;
876
877         devmem->pagemap.type = MEMORY_DEVICE_PRIVATE;
878         devmem->pagemap.res = devmem->resource;
879         devmem->pagemap.page_fault = hmm_devmem_fault;
880         devmem->pagemap.page_free = hmm_devmem_free;
881         devmem->pagemap.dev = devmem->device;
882         devmem->pagemap.ref = &devmem->ref;
883         devmem->pagemap.data = devmem;
884
885         mutex_lock(&hmm_devmem_lock);
886         align_end = align_start + align_size - 1;
887         for (key = align_start; key <= align_end; key += PA_SECTION_SIZE) {
888                 struct hmm_devmem *dup;
889
890                 rcu_read_lock();
891                 dup = hmm_devmem_find(key);
892                 rcu_read_unlock();
893                 if (dup) {
894                         dev_err(device, "%s: collides with mapping for %s\n",
895                                 __func__, dev_name(dup->device));
896                         mutex_unlock(&hmm_devmem_lock);
897                         ret = -EBUSY;
898                         goto error;
899                 }
900                 ret = radix_tree_insert(&hmm_devmem_radix,
901                                         key >> PA_SECTION_SHIFT,
902                                         devmem);
903                 if (ret) {
904                         dev_err(device, "%s: failed: %d\n", __func__, ret);
905                         mutex_unlock(&hmm_devmem_lock);
906                         goto error_radix;
907                 }
908         }
909         mutex_unlock(&hmm_devmem_lock);
910
911         nid = dev_to_node(device);
912         if (nid < 0)
913                 nid = numa_mem_id();
914
915         mem_hotplug_begin();
916         /*
917          * For device private memory we call add_pages() as we only need to
918          * allocate and initialize struct page for the device memory. More-
919          * over the device memory is un-accessible thus we do not want to
920          * create a linear mapping for the memory like arch_add_memory()
921          * would do.
922          */
923         ret = add_pages(nid, align_start >> PAGE_SHIFT,
924                         align_size >> PAGE_SHIFT, false);
925         if (ret) {
926                 mem_hotplug_done();
927                 goto error_add_memory;
928         }
929         move_pfn_range_to_zone(&NODE_DATA(nid)->node_zones[ZONE_DEVICE],
930                                 align_start >> PAGE_SHIFT,
931                                 align_size >> PAGE_SHIFT);
932         mem_hotplug_done();
933
934         for (pfn = devmem->pfn_first; pfn < devmem->pfn_last; pfn++) {
935                 struct page *page = pfn_to_page(pfn);
936
937                 page->pgmap = &devmem->pagemap;
938         }
939         return 0;
940
941 error_add_memory:
942         untrack_pfn(NULL, PHYS_PFN(align_start), align_size);
943 error_radix:
944         hmm_devmem_radix_release(devmem->resource);
945 error:
946         return ret;
947 }
948
949 static int hmm_devmem_match(struct device *dev, void *data, void *match_data)
950 {
951         struct hmm_devmem *devmem = data;
952
953         return devmem->resource == match_data;
954 }
955
956 static void hmm_devmem_pages_remove(struct hmm_devmem *devmem)
957 {
958         devres_release(devmem->device, &hmm_devmem_release,
959                        &hmm_devmem_match, devmem->resource);
960 }
961
962 /*
963  * hmm_devmem_add() - hotplug ZONE_DEVICE memory for device memory
964  *
965  * @ops: memory event device driver callback (see struct hmm_devmem_ops)
966  * @device: device struct to bind the resource too
967  * @size: size in bytes of the device memory to add
968  * Returns: pointer to new hmm_devmem struct ERR_PTR otherwise
969  *
970  * This function first finds an empty range of physical address big enough to
971  * contain the new resource, and then hotplugs it as ZONE_DEVICE memory, which
972  * in turn allocates struct pages. It does not do anything beyond that; all
973  * events affecting the memory will go through the various callbacks provided
974  * by hmm_devmem_ops struct.
975  *
976  * Device driver should call this function during device initialization and
977  * is then responsible of memory management. HMM only provides helpers.
978  */
979 struct hmm_devmem *hmm_devmem_add(const struct hmm_devmem_ops *ops,
980                                   struct device *device,
981                                   unsigned long size)
982 {
983         struct hmm_devmem *devmem;
984         resource_size_t addr;
985         int ret;
986
987         static_branch_enable(&device_private_key);
988
989         devmem = devres_alloc_node(&hmm_devmem_release, sizeof(*devmem),
990                                    GFP_KERNEL, dev_to_node(device));
991         if (!devmem)
992                 return ERR_PTR(-ENOMEM);
993
994         init_completion(&devmem->completion);
995         devmem->pfn_first = -1UL;
996         devmem->pfn_last = -1UL;
997         devmem->resource = NULL;
998         devmem->device = device;
999         devmem->ops = ops;
1000
1001         ret = percpu_ref_init(&devmem->ref, &hmm_devmem_ref_release,
1002                               0, GFP_KERNEL);
1003         if (ret)
1004                 goto error_percpu_ref;
1005
1006         ret = devm_add_action(device, hmm_devmem_ref_exit, &devmem->ref);
1007         if (ret)
1008                 goto error_devm_add_action;
1009
1010         size = ALIGN(size, PA_SECTION_SIZE);
1011         addr = min((unsigned long)iomem_resource.end,
1012                    (1UL << MAX_PHYSMEM_BITS) - 1);
1013         addr = addr - size + 1UL;
1014
1015         /*
1016          * FIXME add a new helper to quickly walk resource tree and find free
1017          * range
1018          *
1019          * FIXME what about ioport_resource resource ?
1020          */
1021         for (; addr > size && addr >= iomem_resource.start; addr -= size) {
1022                 ret = region_intersects(addr, size, 0, IORES_DESC_NONE);
1023                 if (ret != REGION_DISJOINT)
1024                         continue;
1025
1026                 devmem->resource = devm_request_mem_region(device, addr, size,
1027                                                            dev_name(device));
1028                 if (!devmem->resource) {
1029                         ret = -ENOMEM;
1030                         goto error_no_resource;
1031                 }
1032                 break;
1033         }
1034         if (!devmem->resource) {
1035                 ret = -ERANGE;
1036                 goto error_no_resource;
1037         }
1038
1039         devmem->resource->desc = IORES_DESC_DEVICE_PRIVATE_MEMORY;
1040         devmem->pfn_first = devmem->resource->start >> PAGE_SHIFT;
1041         devmem->pfn_last = devmem->pfn_first +
1042                            (resource_size(devmem->resource) >> PAGE_SHIFT);
1043
1044         ret = hmm_devmem_pages_create(devmem);
1045         if (ret)
1046                 goto error_pages;
1047
1048         devres_add(device, devmem);
1049
1050         ret = devm_add_action(device, hmm_devmem_ref_kill, &devmem->ref);
1051         if (ret) {
1052                 hmm_devmem_remove(devmem);
1053                 return ERR_PTR(ret);
1054         }
1055
1056         return devmem;
1057
1058 error_pages:
1059         devm_release_mem_region(device, devmem->resource->start,
1060                                 resource_size(devmem->resource));
1061 error_no_resource:
1062 error_devm_add_action:
1063         hmm_devmem_ref_kill(&devmem->ref);
1064         hmm_devmem_ref_exit(&devmem->ref);
1065 error_percpu_ref:
1066         devres_free(devmem);
1067         return ERR_PTR(ret);
1068 }
1069 EXPORT_SYMBOL(hmm_devmem_add);
1070
1071 /*
1072  * hmm_devmem_remove() - remove device memory (kill and free ZONE_DEVICE)
1073  *
1074  * @devmem: hmm_devmem struct use to track and manage the ZONE_DEVICE memory
1075  *
1076  * This will hot-unplug memory that was hotplugged by hmm_devmem_add on behalf
1077  * of the device driver. It will free struct page and remove the resource that
1078  * reserved the physical address range for this device memory.
1079  */
1080 void hmm_devmem_remove(struct hmm_devmem *devmem)
1081 {
1082         resource_size_t start, size;
1083         struct device *device;
1084
1085         if (!devmem)
1086                 return;
1087
1088         device = devmem->device;
1089         start = devmem->resource->start;
1090         size = resource_size(devmem->resource);
1091
1092         hmm_devmem_ref_kill(&devmem->ref);
1093         hmm_devmem_ref_exit(&devmem->ref);
1094         hmm_devmem_pages_remove(devmem);
1095
1096         devm_release_mem_region(device, start, size);
1097 }
1098 EXPORT_SYMBOL(hmm_devmem_remove);
1099 #endif /* IS_ENABLED(CONFIG_DEVICE_PRIVATE) */