blob: 320545b98ff55997029476f32e09dbf5d4f5f009 [file] [log] [blame]
Jérôme Glisse133ff0e2017-09-08 16:11:23 -07001/*
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>
Jérôme Glisse858b54d2017-09-08 16:12:02 -070022#include <linux/init.h>
Jérôme Glisseda4c3c72017-09-08 16:11:31 -070023#include <linux/rmap.h>
24#include <linux/swap.h>
Jérôme Glisse133ff0e2017-09-08 16:11:23 -070025#include <linux/slab.h>
26#include <linux/sched.h>
Jérôme Glisse4ef589d2017-09-08 16:11:58 -070027#include <linux/mmzone.h>
28#include <linux/pagemap.h>
Jérôme Glisseda4c3c72017-09-08 16:11:31 -070029#include <linux/swapops.h>
30#include <linux/hugetlb.h>
Jérôme Glisse4ef589d2017-09-08 16:11:58 -070031#include <linux/memremap.h>
Jérôme Glisse7b2d55d22017-09-08 16:11:46 -070032#include <linux/jump_label.h>
Jérôme Glissec0b12402017-09-08 16:11:27 -070033#include <linux/mmu_notifier.h>
Jérôme Glisse4ef589d2017-09-08 16:11:58 -070034#include <linux/memory_hotplug.h>
35
36#define PA_SECTION_SIZE (1UL << PA_SECTION_SHIFT)
Jérôme Glisse133ff0e2017-09-08 16:11:23 -070037
Jérôme Glisse6b368cd2017-09-08 16:12:32 -070038#if defined(CONFIG_DEVICE_PRIVATE) || defined(CONFIG_DEVICE_PUBLIC)
Jérôme Glisse7b2d55d22017-09-08 16:11:46 -070039/*
40 * Device private memory see HMM (Documentation/vm/hmm.txt) or hmm.h
41 */
42DEFINE_STATIC_KEY_FALSE(device_private_key);
43EXPORT_SYMBOL(device_private_key);
Jérôme Glisse6b368cd2017-09-08 16:12:32 -070044#endif /* CONFIG_DEVICE_PRIVATE || CONFIG_DEVICE_PUBLIC */
Jérôme Glisse7b2d55d22017-09-08 16:11:46 -070045
46
Jérôme Glisse6b368cd2017-09-08 16:12:32 -070047#if IS_ENABLED(CONFIG_HMM_MIRROR)
Jérôme Glissec0b12402017-09-08 16:11:27 -070048static const struct mmu_notifier_ops hmm_mmu_notifier_ops;
49
Jérôme Glisse133ff0e2017-09-08 16:11:23 -070050/*
51 * struct hmm - HMM per mm struct
52 *
53 * @mm: mm struct this HMM struct is bound to
Jérôme Glisseda4c3c72017-09-08 16:11:31 -070054 * @lock: lock protecting ranges list
Jérôme Glissec0b12402017-09-08 16:11:27 -070055 * @sequence: we track updates to the CPU page table with a sequence number
Jérôme Glisseda4c3c72017-09-08 16:11:31 -070056 * @ranges: list of range being snapshotted
Jérôme Glissec0b12402017-09-08 16:11:27 -070057 * @mirrors: list of mirrors for this mm
58 * @mmu_notifier: mmu notifier to track updates to CPU page table
59 * @mirrors_sem: read/write semaphore protecting the mirrors list
Jérôme Glisse133ff0e2017-09-08 16:11:23 -070060 */
61struct hmm {
62 struct mm_struct *mm;
Jérôme Glisseda4c3c72017-09-08 16:11:31 -070063 spinlock_t lock;
Jérôme Glissec0b12402017-09-08 16:11:27 -070064 atomic_t sequence;
Jérôme Glisseda4c3c72017-09-08 16:11:31 -070065 struct list_head ranges;
Jérôme Glissec0b12402017-09-08 16:11:27 -070066 struct list_head mirrors;
67 struct mmu_notifier mmu_notifier;
68 struct rw_semaphore mirrors_sem;
Jérôme Glisse133ff0e2017-09-08 16:11:23 -070069};
70
71/*
72 * hmm_register - register HMM against an mm (HMM internal)
73 *
74 * @mm: mm struct to attach to
75 *
76 * This is not intended to be used directly by device drivers. It allocates an
77 * HMM struct if mm does not have one, and initializes it.
78 */
79static struct hmm *hmm_register(struct mm_struct *mm)
80{
Jérôme Glissec0b12402017-09-08 16:11:27 -070081 struct hmm *hmm = READ_ONCE(mm->hmm);
82 bool cleanup = false;
Jérôme Glisse133ff0e2017-09-08 16:11:23 -070083
84 /*
85 * The hmm struct can only be freed once the mm_struct goes away,
86 * hence we should always have pre-allocated an new hmm struct
87 * above.
88 */
Jérôme Glissec0b12402017-09-08 16:11:27 -070089 if (hmm)
90 return hmm;
91
92 hmm = kmalloc(sizeof(*hmm), GFP_KERNEL);
93 if (!hmm)
94 return NULL;
95 INIT_LIST_HEAD(&hmm->mirrors);
96 init_rwsem(&hmm->mirrors_sem);
97 atomic_set(&hmm->sequence, 0);
98 hmm->mmu_notifier.ops = NULL;
Jérôme Glisseda4c3c72017-09-08 16:11:31 -070099 INIT_LIST_HEAD(&hmm->ranges);
100 spin_lock_init(&hmm->lock);
Jérôme Glissec0b12402017-09-08 16:11:27 -0700101 hmm->mm = mm;
102
103 /*
104 * We should only get here if hold the mmap_sem in write mode ie on
105 * registration of first mirror through hmm_mirror_register()
106 */
107 hmm->mmu_notifier.ops = &hmm_mmu_notifier_ops;
108 if (__mmu_notifier_register(&hmm->mmu_notifier, mm)) {
109 kfree(hmm);
110 return NULL;
111 }
112
113 spin_lock(&mm->page_table_lock);
114 if (!mm->hmm)
115 mm->hmm = hmm;
116 else
117 cleanup = true;
118 spin_unlock(&mm->page_table_lock);
119
120 if (cleanup) {
121 mmu_notifier_unregister(&hmm->mmu_notifier, mm);
122 kfree(hmm);
123 }
124
Jérôme Glisse133ff0e2017-09-08 16:11:23 -0700125 return mm->hmm;
126}
127
128void hmm_mm_destroy(struct mm_struct *mm)
129{
130 kfree(mm->hmm);
131}
Jérôme Glissec0b12402017-09-08 16:11:27 -0700132
Jérôme Glissec0b12402017-09-08 16:11:27 -0700133static 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;
Jérôme Glisseda4c3c72017-09-08 16:11:31 -0700139 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);
Jérôme Glissec0b12402017-09-08 16:11:27 -0700155
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
163static 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
175static 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
187static 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 */
203int 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}
219EXPORT_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 */
228void 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}
236EXPORT_SYMBOL(hmm_mirror_unregister);
Jérôme Glisseda4c3c72017-09-08 16:11:31 -0700237
Jérôme Glisse74eee182017-09-08 16:11:35 -0700238struct hmm_vma_walk {
239 struct hmm_range *range;
240 unsigned long last;
241 bool fault;
242 bool block;
243 bool write;
244};
245
246static 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
Jérôme Glisseda4c3c72017-09-08 16:11:31 -0700268static 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
276static 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
Jérôme Glisse74eee182017-09-08 16:11:35 -0700291static 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
Jérôme Glisseda4c3c72017-09-08 16:11:31 -0700299static int hmm_vma_walk_hole(unsigned long addr,
300 unsigned long end,
301 struct mm_walk *walk)
302{
Jérôme Glisse74eee182017-09-08 16:11:35 -0700303 struct hmm_vma_walk *hmm_vma_walk = walk->private;
304 struct hmm_range *range = hmm_vma_walk->range;
Jérôme Glisseda4c3c72017-09-08 16:11:31 -0700305 hmm_pfn_t *pfns = range->pfns;
306 unsigned long i;
307
Jérôme Glisse74eee182017-09-08 16:11:35 -0700308 hmm_vma_walk->last = addr;
Jérôme Glisseda4c3c72017-09-08 16:11:31 -0700309 i = (addr - range->start) >> PAGE_SHIFT;
Jérôme Glisse74eee182017-09-08 16:11:35 -0700310 for (; addr < end; addr += PAGE_SIZE, i++) {
Jérôme Glisseda4c3c72017-09-08 16:11:31 -0700311 pfns[i] = HMM_PFN_EMPTY;
Jérôme Glisse74eee182017-09-08 16:11:35 -0700312 if (hmm_vma_walk->fault) {
313 int ret;
Jérôme Glisseda4c3c72017-09-08 16:11:31 -0700314
Jérôme Glisse74eee182017-09-08 16:11:35 -0700315 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;
Jérôme Glisseda4c3c72017-09-08 16:11:31 -0700322}
323
324static int hmm_vma_walk_clear(unsigned long addr,
325 unsigned long end,
326 struct mm_walk *walk)
327{
Jérôme Glisse74eee182017-09-08 16:11:35 -0700328 struct hmm_vma_walk *hmm_vma_walk = walk->private;
329 struct hmm_range *range = hmm_vma_walk->range;
Jérôme Glisseda4c3c72017-09-08 16:11:31 -0700330 hmm_pfn_t *pfns = range->pfns;
331 unsigned long i;
332
Jérôme Glisse74eee182017-09-08 16:11:35 -0700333 hmm_vma_walk->last = addr;
Jérôme Glisseda4c3c72017-09-08 16:11:31 -0700334 i = (addr - range->start) >> PAGE_SHIFT;
Jérôme Glisse74eee182017-09-08 16:11:35 -0700335 for (; addr < end; addr += PAGE_SIZE, i++) {
Jérôme Glisseda4c3c72017-09-08 16:11:31 -0700336 pfns[i] = 0;
Jérôme Glisse74eee182017-09-08 16:11:35 -0700337 if (hmm_vma_walk->fault) {
338 int ret;
Jérôme Glisseda4c3c72017-09-08 16:11:31 -0700339
Jérôme Glisse74eee182017-09-08 16:11:35 -0700340 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;
Jérôme Glisseda4c3c72017-09-08 16:11:31 -0700347}
348
349static int hmm_vma_walk_pmd(pmd_t *pmdp,
350 unsigned long start,
351 unsigned long end,
352 struct mm_walk *walk)
353{
Jérôme Glisse74eee182017-09-08 16:11:35 -0700354 struct hmm_vma_walk *hmm_vma_walk = walk->private;
355 struct hmm_range *range = hmm_vma_walk->range;
Jérôme Glisseda4c3c72017-09-08 16:11:31 -0700356 struct vm_area_struct *vma = walk->vma;
357 hmm_pfn_t *pfns = range->pfns;
358 unsigned long addr = start, i;
Jérôme Glisse74eee182017-09-08 16:11:35 -0700359 bool write_fault;
Jérôme Glisseda4c3c72017-09-08 16:11:31 -0700360 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;
Jérôme Glisse74eee182017-09-08 16:11:35 -0700365 write_fault = hmm_vma_walk->fault & hmm_vma_walk->write;
Jérôme Glisseda4c3c72017-09-08 16:11:31 -0700366
367again:
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
Linus Torvaldsf6f37322017-12-15 18:53:22 -0800394 if (write_fault && !pmd_write(pmd))
Jérôme Glisse74eee182017-09-08 16:11:35 -0700395 return hmm_vma_walk_clear(start, end, walk);
396
Jérôme Glisseda4c3c72017-09-08 16:11:31 -0700397 pfn = pmd_pfn(pmd) + pte_index(addr);
Linus Torvaldsf6f37322017-12-15 18:53:22 -0800398 flag |= pmd_write(pmd) ? HMM_PFN_WRITE : 0;
Jérôme Glisseda4c3c72017-09-08 16:11:31 -0700399 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
Jérôme Glisse74eee182017-09-08 16:11:35 -0700413 if (pte_none(pte)) {
Jérôme Glisseda4c3c72017-09-08 16:11:31 -0700414 pfns[i] = HMM_PFN_EMPTY;
Jérôme Glisse74eee182017-09-08 16:11:35 -0700415 if (hmm_vma_walk->fault)
416 goto fault;
Jérôme Glisseda4c3c72017-09-08 16:11:31 -0700417 continue;
418 }
419
Jérôme Glisse74eee182017-09-08 16:11:35 -0700420 if (!pte_present(pte)) {
Ralph Campbell8d63e4c2018-01-31 16:20:30 -0800421 swp_entry_t entry = pte_to_swp_entry(pte);
Jérôme Glisse74eee182017-09-08 16:11:35 -0700422
423 if (!non_swap_entry(entry)) {
424 if (hmm_vma_walk->fault)
425 goto fault;
426 continue;
427 }
428
Jérôme Glisse74eee182017-09-08 16:11:35 -0700429 /*
430 * This is a special swap entry, ignore migration, use
431 * device and report anything else as error.
432 */
Jérôme Glisse4ef589d2017-09-08 16:11:58 -0700433 if (is_device_private_entry(entry)) {
434 pfns[i] = hmm_pfn_t_from_pfn(swp_offset(entry));
435 if (is_write_device_private_entry(entry)) {
436 pfns[i] |= HMM_PFN_WRITE;
437 } else if (write_fault)
438 goto fault;
439 pfns[i] |= HMM_PFN_DEVICE_UNADDRESSABLE;
440 pfns[i] |= flag;
441 } else if (is_migration_entry(entry)) {
Jérôme Glisse74eee182017-09-08 16:11:35 -0700442 if (hmm_vma_walk->fault) {
443 pte_unmap(ptep);
444 hmm_vma_walk->last = addr;
445 migration_entry_wait(vma->vm_mm,
446 pmdp, addr);
447 return -EAGAIN;
448 }
449 continue;
450 } else {
451 /* Report error for everything else */
452 pfns[i] = HMM_PFN_ERROR;
453 }
454 continue;
455 }
456
Linus Torvaldsf6f37322017-12-15 18:53:22 -0800457 if (write_fault && !pte_write(pte))
Jérôme Glisse74eee182017-09-08 16:11:35 -0700458 goto fault;
459
Jérôme Glisseda4c3c72017-09-08 16:11:31 -0700460 pfns[i] = hmm_pfn_t_from_pfn(pte_pfn(pte)) | flag;
Linus Torvaldsf6f37322017-12-15 18:53:22 -0800461 pfns[i] |= pte_write(pte) ? HMM_PFN_WRITE : 0;
Jérôme Glisse74eee182017-09-08 16:11:35 -0700462 continue;
463
464fault:
465 pte_unmap(ptep);
466 /* Fault all pages in range */
467 return hmm_vma_walk_clear(start, end, walk);
Jérôme Glisseda4c3c72017-09-08 16:11:31 -0700468 }
469 pte_unmap(ptep - 1);
470
471 return 0;
472}
473
474/*
475 * hmm_vma_get_pfns() - snapshot CPU page table for a range of virtual addresses
476 * @vma: virtual memory area containing the virtual address range
477 * @range: used to track snapshot validity
478 * @start: range virtual start address (inclusive)
479 * @end: range virtual end address (exclusive)
480 * @entries: array of hmm_pfn_t: provided by the caller, filled in by function
481 * Returns: -EINVAL if invalid argument, -ENOMEM out of memory, 0 success
482 *
483 * This snapshots the CPU page table for a range of virtual addresses. Snapshot
484 * validity is tracked by range struct. See hmm_vma_range_done() for further
485 * information.
486 *
487 * The range struct is initialized here. It tracks the CPU page table, but only
488 * if the function returns success (0), in which case the caller must then call
489 * hmm_vma_range_done() to stop CPU page table update tracking on this range.
490 *
491 * NOT CALLING hmm_vma_range_done() IF FUNCTION RETURNS 0 WILL LEAD TO SERIOUS
492 * MEMORY CORRUPTION ! YOU HAVE BEEN WARNED !
493 */
494int hmm_vma_get_pfns(struct vm_area_struct *vma,
495 struct hmm_range *range,
496 unsigned long start,
497 unsigned long end,
498 hmm_pfn_t *pfns)
499{
Jérôme Glisse74eee182017-09-08 16:11:35 -0700500 struct hmm_vma_walk hmm_vma_walk;
Jérôme Glisseda4c3c72017-09-08 16:11:31 -0700501 struct mm_walk mm_walk;
502 struct hmm *hmm;
503
504 /* FIXME support hugetlb fs */
505 if (is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_SPECIAL)) {
506 hmm_pfns_special(pfns, start, end);
507 return -EINVAL;
508 }
509
510 /* Sanity check, this really should not happen ! */
511 if (start < vma->vm_start || start >= vma->vm_end)
512 return -EINVAL;
513 if (end < vma->vm_start || end > vma->vm_end)
514 return -EINVAL;
515
516 hmm = hmm_register(vma->vm_mm);
517 if (!hmm)
518 return -ENOMEM;
519 /* Caller must have registered a mirror, via hmm_mirror_register() ! */
520 if (!hmm->mmu_notifier.ops)
521 return -EINVAL;
522
523 /* Initialize range to track CPU page table update */
524 range->start = start;
525 range->pfns = pfns;
526 range->end = end;
527 spin_lock(&hmm->lock);
528 range->valid = true;
529 list_add_rcu(&range->list, &hmm->ranges);
530 spin_unlock(&hmm->lock);
531
Jérôme Glisse74eee182017-09-08 16:11:35 -0700532 hmm_vma_walk.fault = false;
533 hmm_vma_walk.range = range;
534 mm_walk.private = &hmm_vma_walk;
535
Jérôme Glisseda4c3c72017-09-08 16:11:31 -0700536 mm_walk.vma = vma;
537 mm_walk.mm = vma->vm_mm;
Jérôme Glisseda4c3c72017-09-08 16:11:31 -0700538 mm_walk.pte_entry = NULL;
539 mm_walk.test_walk = NULL;
540 mm_walk.hugetlb_entry = NULL;
541 mm_walk.pmd_entry = hmm_vma_walk_pmd;
542 mm_walk.pte_hole = hmm_vma_walk_hole;
543
544 walk_page_range(start, end, &mm_walk);
Jérôme Glisseda4c3c72017-09-08 16:11:31 -0700545 return 0;
546}
547EXPORT_SYMBOL(hmm_vma_get_pfns);
548
549/*
550 * hmm_vma_range_done() - stop tracking change to CPU page table over a range
551 * @vma: virtual memory area containing the virtual address range
552 * @range: range being tracked
553 * Returns: false if range data has been invalidated, true otherwise
554 *
555 * Range struct is used to track updates to the CPU page table after a call to
556 * either hmm_vma_get_pfns() or hmm_vma_fault(). Once the device driver is done
557 * using the data, or wants to lock updates to the data it got from those
558 * functions, it must call the hmm_vma_range_done() function, which will then
559 * stop tracking CPU page table updates.
560 *
561 * Note that device driver must still implement general CPU page table update
562 * tracking either by using hmm_mirror (see hmm_mirror_register()) or by using
563 * the mmu_notifier API directly.
564 *
565 * CPU page table update tracking done through hmm_range is only temporary and
566 * to be used while trying to duplicate CPU page table contents for a range of
567 * virtual addresses.
568 *
569 * There are two ways to use this :
570 * again:
Jérôme Glisse74eee182017-09-08 16:11:35 -0700571 * hmm_vma_get_pfns(vma, range, start, end, pfns); or hmm_vma_fault(...);
Jérôme Glisseda4c3c72017-09-08 16:11:31 -0700572 * trans = device_build_page_table_update_transaction(pfns);
573 * device_page_table_lock();
574 * if (!hmm_vma_range_done(vma, range)) {
575 * device_page_table_unlock();
576 * goto again;
577 * }
578 * device_commit_transaction(trans);
579 * device_page_table_unlock();
580 *
581 * Or:
Jérôme Glisse74eee182017-09-08 16:11:35 -0700582 * hmm_vma_get_pfns(vma, range, start, end, pfns); or hmm_vma_fault(...);
Jérôme Glisseda4c3c72017-09-08 16:11:31 -0700583 * device_page_table_lock();
584 * hmm_vma_range_done(vma, range);
585 * device_update_page_table(pfns);
586 * device_page_table_unlock();
587 */
588bool hmm_vma_range_done(struct vm_area_struct *vma, struct hmm_range *range)
589{
590 unsigned long npages = (range->end - range->start) >> PAGE_SHIFT;
591 struct hmm *hmm;
592
593 if (range->end <= range->start) {
594 BUG();
595 return false;
596 }
597
598 hmm = hmm_register(vma->vm_mm);
599 if (!hmm) {
600 memset(range->pfns, 0, sizeof(*range->pfns) * npages);
601 return false;
602 }
603
604 spin_lock(&hmm->lock);
605 list_del_rcu(&range->list);
606 spin_unlock(&hmm->lock);
607
608 return range->valid;
609}
610EXPORT_SYMBOL(hmm_vma_range_done);
Jérôme Glisse74eee182017-09-08 16:11:35 -0700611
612/*
613 * hmm_vma_fault() - try to fault some address in a virtual address range
614 * @vma: virtual memory area containing the virtual address range
615 * @range: use to track pfns array content validity
616 * @start: fault range virtual start address (inclusive)
617 * @end: fault range virtual end address (exclusive)
618 * @pfns: array of hmm_pfn_t, only entry with fault flag set will be faulted
619 * @write: is it a write fault
620 * @block: allow blocking on fault (if true it sleeps and do not drop mmap_sem)
621 * Returns: 0 success, error otherwise (-EAGAIN means mmap_sem have been drop)
622 *
623 * This is similar to a regular CPU page fault except that it will not trigger
624 * any memory migration if the memory being faulted is not accessible by CPUs.
625 *
626 * On error, for one virtual address in the range, the function will set the
627 * hmm_pfn_t error flag for the corresponding pfn entry.
628 *
629 * Expected use pattern:
630 * retry:
631 * down_read(&mm->mmap_sem);
632 * // Find vma and address device wants to fault, initialize hmm_pfn_t
633 * // array accordingly
634 * ret = hmm_vma_fault(vma, start, end, pfns, allow_retry);
635 * switch (ret) {
636 * case -EAGAIN:
637 * hmm_vma_range_done(vma, range);
638 * // You might want to rate limit or yield to play nicely, you may
639 * // also commit any valid pfn in the array assuming that you are
640 * // getting true from hmm_vma_range_monitor_end()
641 * goto retry;
642 * case 0:
643 * break;
644 * default:
645 * // Handle error !
646 * up_read(&mm->mmap_sem)
647 * return;
648 * }
649 * // Take device driver lock that serialize device page table update
650 * driver_lock_device_page_table_update();
651 * hmm_vma_range_done(vma, range);
652 * // Commit pfns we got from hmm_vma_fault()
653 * driver_unlock_device_page_table_update();
654 * up_read(&mm->mmap_sem)
655 *
656 * YOU MUST CALL hmm_vma_range_done() AFTER THIS FUNCTION RETURN SUCCESS (0)
657 * BEFORE FREEING THE range struct OR YOU WILL HAVE SERIOUS MEMORY CORRUPTION !
658 *
659 * YOU HAVE BEEN WARNED !
660 */
661int hmm_vma_fault(struct vm_area_struct *vma,
662 struct hmm_range *range,
663 unsigned long start,
664 unsigned long end,
665 hmm_pfn_t *pfns,
666 bool write,
667 bool block)
668{
669 struct hmm_vma_walk hmm_vma_walk;
670 struct mm_walk mm_walk;
671 struct hmm *hmm;
672 int ret;
673
674 /* Sanity check, this really should not happen ! */
675 if (start < vma->vm_start || start >= vma->vm_end)
676 return -EINVAL;
677 if (end < vma->vm_start || end > vma->vm_end)
678 return -EINVAL;
679
680 hmm = hmm_register(vma->vm_mm);
681 if (!hmm) {
682 hmm_pfns_clear(pfns, start, end);
683 return -ENOMEM;
684 }
685 /* Caller must have registered a mirror using hmm_mirror_register() */
686 if (!hmm->mmu_notifier.ops)
687 return -EINVAL;
688
689 /* Initialize range to track CPU page table update */
690 range->start = start;
691 range->pfns = pfns;
692 range->end = end;
693 spin_lock(&hmm->lock);
694 range->valid = true;
695 list_add_rcu(&range->list, &hmm->ranges);
696 spin_unlock(&hmm->lock);
697
698 /* FIXME support hugetlb fs */
699 if (is_vm_hugetlb_page(vma) || (vma->vm_flags & VM_SPECIAL)) {
700 hmm_pfns_special(pfns, start, end);
701 return 0;
702 }
703
704 hmm_vma_walk.fault = true;
705 hmm_vma_walk.write = write;
706 hmm_vma_walk.block = block;
707 hmm_vma_walk.range = range;
708 mm_walk.private = &hmm_vma_walk;
709 hmm_vma_walk.last = range->start;
710
711 mm_walk.vma = vma;
712 mm_walk.mm = vma->vm_mm;
713 mm_walk.pte_entry = NULL;
714 mm_walk.test_walk = NULL;
715 mm_walk.hugetlb_entry = NULL;
716 mm_walk.pmd_entry = hmm_vma_walk_pmd;
717 mm_walk.pte_hole = hmm_vma_walk_hole;
718
719 do {
720 ret = walk_page_range(start, end, &mm_walk);
721 start = hmm_vma_walk.last;
722 } while (ret == -EAGAIN);
723
724 if (ret) {
725 unsigned long i;
726
727 i = (hmm_vma_walk.last - range->start) >> PAGE_SHIFT;
728 hmm_pfns_clear(&pfns[i], hmm_vma_walk.last, end);
729 hmm_vma_range_done(vma, range);
730 }
731 return ret;
732}
733EXPORT_SYMBOL(hmm_vma_fault);
Jérôme Glissec0b12402017-09-08 16:11:27 -0700734#endif /* IS_ENABLED(CONFIG_HMM_MIRROR) */
Jérôme Glisse4ef589d2017-09-08 16:11:58 -0700735
736
Jérôme Glissedf6ad692017-09-08 16:12:24 -0700737#if IS_ENABLED(CONFIG_DEVICE_PRIVATE) || IS_ENABLED(CONFIG_DEVICE_PUBLIC)
Jérôme Glisse4ef589d2017-09-08 16:11:58 -0700738struct page *hmm_vma_alloc_locked_page(struct vm_area_struct *vma,
739 unsigned long addr)
740{
741 struct page *page;
742
743 page = alloc_page_vma(GFP_HIGHUSER, vma, addr);
744 if (!page)
745 return NULL;
746 lock_page(page);
747 return page;
748}
749EXPORT_SYMBOL(hmm_vma_alloc_locked_page);
750
751
752static void hmm_devmem_ref_release(struct percpu_ref *ref)
753{
754 struct hmm_devmem *devmem;
755
756 devmem = container_of(ref, struct hmm_devmem, ref);
757 complete(&devmem->completion);
758}
759
760static void hmm_devmem_ref_exit(void *data)
761{
762 struct percpu_ref *ref = data;
763 struct hmm_devmem *devmem;
764
765 devmem = container_of(ref, struct hmm_devmem, ref);
766 percpu_ref_exit(ref);
767 devm_remove_action(devmem->device, &hmm_devmem_ref_exit, data);
768}
769
770static void hmm_devmem_ref_kill(void *data)
771{
772 struct percpu_ref *ref = data;
773 struct hmm_devmem *devmem;
774
775 devmem = container_of(ref, struct hmm_devmem, ref);
776 percpu_ref_kill(ref);
777 wait_for_completion(&devmem->completion);
778 devm_remove_action(devmem->device, &hmm_devmem_ref_kill, data);
779}
780
781static int hmm_devmem_fault(struct vm_area_struct *vma,
782 unsigned long addr,
783 const struct page *page,
784 unsigned int flags,
785 pmd_t *pmdp)
786{
787 struct hmm_devmem *devmem = page->pgmap->data;
788
789 return devmem->ops->fault(devmem, vma, addr, page, flags, pmdp);
790}
791
792static void hmm_devmem_free(struct page *page, void *data)
793{
794 struct hmm_devmem *devmem = data;
795
796 devmem->ops->free(devmem, page);
797}
798
799static DEFINE_MUTEX(hmm_devmem_lock);
800static RADIX_TREE(hmm_devmem_radix, GFP_KERNEL);
801
802static void hmm_devmem_radix_release(struct resource *resource)
803{
Colin Ian Kingfec11bc2017-11-15 17:38:52 -0800804 resource_size_t key, align_start, align_size;
Jérôme Glisse4ef589d2017-09-08 16:11:58 -0700805
806 align_start = resource->start & ~(PA_SECTION_SIZE - 1);
807 align_size = ALIGN(resource_size(resource), PA_SECTION_SIZE);
Jérôme Glisse4ef589d2017-09-08 16:11:58 -0700808
809 mutex_lock(&hmm_devmem_lock);
810 for (key = resource->start;
811 key <= resource->end;
812 key += PA_SECTION_SIZE)
813 radix_tree_delete(&hmm_devmem_radix, key >> PA_SECTION_SHIFT);
814 mutex_unlock(&hmm_devmem_lock);
815}
816
817static void hmm_devmem_release(struct device *dev, void *data)
818{
819 struct hmm_devmem *devmem = data;
820 struct resource *resource = devmem->resource;
821 unsigned long start_pfn, npages;
822 struct zone *zone;
823 struct page *page;
824
825 if (percpu_ref_tryget_live(&devmem->ref)) {
826 dev_WARN(dev, "%s: page mapping is still live!\n", __func__);
827 percpu_ref_put(&devmem->ref);
828 }
829
830 /* pages are dead and unused, undo the arch mapping */
831 start_pfn = (resource->start & ~(PA_SECTION_SIZE - 1)) >> PAGE_SHIFT;
832 npages = ALIGN(resource_size(resource), PA_SECTION_SIZE) >> PAGE_SHIFT;
833
834 page = pfn_to_page(start_pfn);
835 zone = page_zone(page);
836
837 mem_hotplug_begin();
Jérôme Glissed3df0a42017-09-08 16:12:28 -0700838 if (resource->desc == IORES_DESC_DEVICE_PRIVATE_MEMORY)
Christoph Hellwigda024512017-12-29 08:53:55 +0100839 __remove_pages(zone, start_pfn, npages, NULL);
Jérôme Glissed3df0a42017-09-08 16:12:28 -0700840 else
841 arch_remove_memory(start_pfn << PAGE_SHIFT,
Christoph Hellwigda024512017-12-29 08:53:55 +0100842 npages << PAGE_SHIFT, NULL);
Jérôme Glisse4ef589d2017-09-08 16:11:58 -0700843 mem_hotplug_done();
844
845 hmm_devmem_radix_release(resource);
846}
847
848static struct hmm_devmem *hmm_devmem_find(resource_size_t phys)
849{
850 WARN_ON_ONCE(!rcu_read_lock_held());
851
852 return radix_tree_lookup(&hmm_devmem_radix, phys >> PA_SECTION_SHIFT);
853}
854
855static int hmm_devmem_pages_create(struct hmm_devmem *devmem)
856{
857 resource_size_t key, align_start, align_size, align_end;
858 struct device *device = devmem->device;
859 int ret, nid, is_ram;
860 unsigned long pfn;
861
862 align_start = devmem->resource->start & ~(PA_SECTION_SIZE - 1);
863 align_size = ALIGN(devmem->resource->start +
864 resource_size(devmem->resource),
865 PA_SECTION_SIZE) - align_start;
866
867 is_ram = region_intersects(align_start, align_size,
868 IORESOURCE_SYSTEM_RAM,
869 IORES_DESC_NONE);
870 if (is_ram == REGION_MIXED) {
871 WARN_ONCE(1, "%s attempted on mixed region %pr\n",
872 __func__, devmem->resource);
873 return -ENXIO;
874 }
875 if (is_ram == REGION_INTERSECTS)
876 return -ENXIO;
877
Jérôme Glissed3df0a42017-09-08 16:12:28 -0700878 if (devmem->resource->desc == IORES_DESC_DEVICE_PUBLIC_MEMORY)
879 devmem->pagemap.type = MEMORY_DEVICE_PUBLIC;
880 else
881 devmem->pagemap.type = MEMORY_DEVICE_PRIVATE;
882
Logan Gunthorpee7744aa2017-12-29 08:54:04 +0100883 devmem->pagemap.res = *devmem->resource;
Jérôme Glisse4ef589d2017-09-08 16:11:58 -0700884 devmem->pagemap.page_fault = hmm_devmem_fault;
885 devmem->pagemap.page_free = hmm_devmem_free;
886 devmem->pagemap.dev = devmem->device;
887 devmem->pagemap.ref = &devmem->ref;
888 devmem->pagemap.data = devmem;
889
890 mutex_lock(&hmm_devmem_lock);
891 align_end = align_start + align_size - 1;
892 for (key = align_start; key <= align_end; key += PA_SECTION_SIZE) {
893 struct hmm_devmem *dup;
894
895 rcu_read_lock();
896 dup = hmm_devmem_find(key);
897 rcu_read_unlock();
898 if (dup) {
899 dev_err(device, "%s: collides with mapping for %s\n",
900 __func__, dev_name(dup->device));
901 mutex_unlock(&hmm_devmem_lock);
902 ret = -EBUSY;
903 goto error;
904 }
905 ret = radix_tree_insert(&hmm_devmem_radix,
906 key >> PA_SECTION_SHIFT,
907 devmem);
908 if (ret) {
909 dev_err(device, "%s: failed: %d\n", __func__, ret);
910 mutex_unlock(&hmm_devmem_lock);
911 goto error_radix;
912 }
913 }
914 mutex_unlock(&hmm_devmem_lock);
915
916 nid = dev_to_node(device);
917 if (nid < 0)
918 nid = numa_mem_id();
919
920 mem_hotplug_begin();
921 /*
922 * For device private memory we call add_pages() as we only need to
923 * allocate and initialize struct page for the device memory. More-
924 * over the device memory is un-accessible thus we do not want to
925 * create a linear mapping for the memory like arch_add_memory()
926 * would do.
Jérôme Glissed3df0a42017-09-08 16:12:28 -0700927 *
928 * For device public memory, which is accesible by the CPU, we do
929 * want the linear mapping and thus use arch_add_memory().
Jérôme Glisse4ef589d2017-09-08 16:11:58 -0700930 */
Jérôme Glissed3df0a42017-09-08 16:12:28 -0700931 if (devmem->pagemap.type == MEMORY_DEVICE_PUBLIC)
Christoph Hellwig24e6d5a2017-12-29 08:53:53 +0100932 ret = arch_add_memory(nid, align_start, align_size, NULL,
933 false);
Jérôme Glissed3df0a42017-09-08 16:12:28 -0700934 else
935 ret = add_pages(nid, align_start >> PAGE_SHIFT,
Christoph Hellwig24e6d5a2017-12-29 08:53:53 +0100936 align_size >> PAGE_SHIFT, NULL, false);
Jérôme Glisse4ef589d2017-09-08 16:11:58 -0700937 if (ret) {
938 mem_hotplug_done();
939 goto error_add_memory;
940 }
941 move_pfn_range_to_zone(&NODE_DATA(nid)->node_zones[ZONE_DEVICE],
942 align_start >> PAGE_SHIFT,
Christoph Hellwiga99583e2017-12-29 08:53:57 +0100943 align_size >> PAGE_SHIFT, NULL);
Jérôme Glisse4ef589d2017-09-08 16:11:58 -0700944 mem_hotplug_done();
945
946 for (pfn = devmem->pfn_first; pfn < devmem->pfn_last; pfn++) {
947 struct page *page = pfn_to_page(pfn);
948
949 page->pgmap = &devmem->pagemap;
950 }
951 return 0;
952
953error_add_memory:
954 untrack_pfn(NULL, PHYS_PFN(align_start), align_size);
955error_radix:
956 hmm_devmem_radix_release(devmem->resource);
957error:
958 return ret;
959}
960
961static int hmm_devmem_match(struct device *dev, void *data, void *match_data)
962{
963 struct hmm_devmem *devmem = data;
964
965 return devmem->resource == match_data;
966}
967
968static void hmm_devmem_pages_remove(struct hmm_devmem *devmem)
969{
970 devres_release(devmem->device, &hmm_devmem_release,
971 &hmm_devmem_match, devmem->resource);
972}
973
974/*
975 * hmm_devmem_add() - hotplug ZONE_DEVICE memory for device memory
976 *
977 * @ops: memory event device driver callback (see struct hmm_devmem_ops)
978 * @device: device struct to bind the resource too
979 * @size: size in bytes of the device memory to add
980 * Returns: pointer to new hmm_devmem struct ERR_PTR otherwise
981 *
982 * This function first finds an empty range of physical address big enough to
983 * contain the new resource, and then hotplugs it as ZONE_DEVICE memory, which
984 * in turn allocates struct pages. It does not do anything beyond that; all
985 * events affecting the memory will go through the various callbacks provided
986 * by hmm_devmem_ops struct.
987 *
988 * Device driver should call this function during device initialization and
989 * is then responsible of memory management. HMM only provides helpers.
990 */
991struct hmm_devmem *hmm_devmem_add(const struct hmm_devmem_ops *ops,
992 struct device *device,
993 unsigned long size)
994{
995 struct hmm_devmem *devmem;
996 resource_size_t addr;
997 int ret;
998
999 static_branch_enable(&device_private_key);
1000
1001 devmem = devres_alloc_node(&hmm_devmem_release, sizeof(*devmem),
1002 GFP_KERNEL, dev_to_node(device));
1003 if (!devmem)
1004 return ERR_PTR(-ENOMEM);
1005
1006 init_completion(&devmem->completion);
1007 devmem->pfn_first = -1UL;
1008 devmem->pfn_last = -1UL;
1009 devmem->resource = NULL;
1010 devmem->device = device;
1011 devmem->ops = ops;
1012
1013 ret = percpu_ref_init(&devmem->ref, &hmm_devmem_ref_release,
1014 0, GFP_KERNEL);
1015 if (ret)
1016 goto error_percpu_ref;
1017
1018 ret = devm_add_action(device, hmm_devmem_ref_exit, &devmem->ref);
1019 if (ret)
1020 goto error_devm_add_action;
1021
1022 size = ALIGN(size, PA_SECTION_SIZE);
1023 addr = min((unsigned long)iomem_resource.end,
1024 (1UL << MAX_PHYSMEM_BITS) - 1);
1025 addr = addr - size + 1UL;
1026
1027 /*
1028 * FIXME add a new helper to quickly walk resource tree and find free
1029 * range
1030 *
1031 * FIXME what about ioport_resource resource ?
1032 */
1033 for (; addr > size && addr >= iomem_resource.start; addr -= size) {
1034 ret = region_intersects(addr, size, 0, IORES_DESC_NONE);
1035 if (ret != REGION_DISJOINT)
1036 continue;
1037
1038 devmem->resource = devm_request_mem_region(device, addr, size,
1039 dev_name(device));
1040 if (!devmem->resource) {
1041 ret = -ENOMEM;
1042 goto error_no_resource;
1043 }
1044 break;
1045 }
1046 if (!devmem->resource) {
1047 ret = -ERANGE;
1048 goto error_no_resource;
1049 }
1050
1051 devmem->resource->desc = IORES_DESC_DEVICE_PRIVATE_MEMORY;
1052 devmem->pfn_first = devmem->resource->start >> PAGE_SHIFT;
1053 devmem->pfn_last = devmem->pfn_first +
1054 (resource_size(devmem->resource) >> PAGE_SHIFT);
1055
1056 ret = hmm_devmem_pages_create(devmem);
1057 if (ret)
1058 goto error_pages;
1059
1060 devres_add(device, devmem);
1061
1062 ret = devm_add_action(device, hmm_devmem_ref_kill, &devmem->ref);
1063 if (ret) {
1064 hmm_devmem_remove(devmem);
1065 return ERR_PTR(ret);
1066 }
1067
1068 return devmem;
1069
1070error_pages:
1071 devm_release_mem_region(device, devmem->resource->start,
1072 resource_size(devmem->resource));
1073error_no_resource:
1074error_devm_add_action:
1075 hmm_devmem_ref_kill(&devmem->ref);
1076 hmm_devmem_ref_exit(&devmem->ref);
1077error_percpu_ref:
1078 devres_free(devmem);
1079 return ERR_PTR(ret);
1080}
1081EXPORT_SYMBOL(hmm_devmem_add);
1082
Jérôme Glissed3df0a42017-09-08 16:12:28 -07001083struct hmm_devmem *hmm_devmem_add_resource(const struct hmm_devmem_ops *ops,
1084 struct device *device,
1085 struct resource *res)
1086{
1087 struct hmm_devmem *devmem;
1088 int ret;
1089
1090 if (res->desc != IORES_DESC_DEVICE_PUBLIC_MEMORY)
1091 return ERR_PTR(-EINVAL);
1092
1093 static_branch_enable(&device_private_key);
1094
1095 devmem = devres_alloc_node(&hmm_devmem_release, sizeof(*devmem),
1096 GFP_KERNEL, dev_to_node(device));
1097 if (!devmem)
1098 return ERR_PTR(-ENOMEM);
1099
1100 init_completion(&devmem->completion);
1101 devmem->pfn_first = -1UL;
1102 devmem->pfn_last = -1UL;
1103 devmem->resource = res;
1104 devmem->device = device;
1105 devmem->ops = ops;
1106
1107 ret = percpu_ref_init(&devmem->ref, &hmm_devmem_ref_release,
1108 0, GFP_KERNEL);
1109 if (ret)
1110 goto error_percpu_ref;
1111
1112 ret = devm_add_action(device, hmm_devmem_ref_exit, &devmem->ref);
1113 if (ret)
1114 goto error_devm_add_action;
1115
1116
1117 devmem->pfn_first = devmem->resource->start >> PAGE_SHIFT;
1118 devmem->pfn_last = devmem->pfn_first +
1119 (resource_size(devmem->resource) >> PAGE_SHIFT);
1120
1121 ret = hmm_devmem_pages_create(devmem);
1122 if (ret)
1123 goto error_devm_add_action;
1124
1125 devres_add(device, devmem);
1126
1127 ret = devm_add_action(device, hmm_devmem_ref_kill, &devmem->ref);
1128 if (ret) {
1129 hmm_devmem_remove(devmem);
1130 return ERR_PTR(ret);
1131 }
1132
1133 return devmem;
1134
1135error_devm_add_action:
1136 hmm_devmem_ref_kill(&devmem->ref);
1137 hmm_devmem_ref_exit(&devmem->ref);
1138error_percpu_ref:
1139 devres_free(devmem);
1140 return ERR_PTR(ret);
1141}
1142EXPORT_SYMBOL(hmm_devmem_add_resource);
1143
Jérôme Glisse4ef589d2017-09-08 16:11:58 -07001144/*
1145 * hmm_devmem_remove() - remove device memory (kill and free ZONE_DEVICE)
1146 *
1147 * @devmem: hmm_devmem struct use to track and manage the ZONE_DEVICE memory
1148 *
1149 * This will hot-unplug memory that was hotplugged by hmm_devmem_add on behalf
1150 * of the device driver. It will free struct page and remove the resource that
1151 * reserved the physical address range for this device memory.
1152 */
1153void hmm_devmem_remove(struct hmm_devmem *devmem)
1154{
1155 resource_size_t start, size;
1156 struct device *device;
Jérôme Glissed3df0a42017-09-08 16:12:28 -07001157 bool cdm = false;
Jérôme Glisse4ef589d2017-09-08 16:11:58 -07001158
1159 if (!devmem)
1160 return;
1161
1162 device = devmem->device;
1163 start = devmem->resource->start;
1164 size = resource_size(devmem->resource);
1165
Jérôme Glissed3df0a42017-09-08 16:12:28 -07001166 cdm = devmem->resource->desc == IORES_DESC_DEVICE_PUBLIC_MEMORY;
Jérôme Glisse4ef589d2017-09-08 16:11:58 -07001167 hmm_devmem_ref_kill(&devmem->ref);
1168 hmm_devmem_ref_exit(&devmem->ref);
1169 hmm_devmem_pages_remove(devmem);
1170
Jérôme Glissed3df0a42017-09-08 16:12:28 -07001171 if (!cdm)
1172 devm_release_mem_region(device, start, size);
Jérôme Glisse4ef589d2017-09-08 16:11:58 -07001173}
1174EXPORT_SYMBOL(hmm_devmem_remove);
Jérôme Glisse858b54d2017-09-08 16:12:02 -07001175
1176/*
1177 * A device driver that wants to handle multiple devices memory through a
1178 * single fake device can use hmm_device to do so. This is purely a helper
1179 * and it is not needed to make use of any HMM functionality.
1180 */
1181#define HMM_DEVICE_MAX 256
1182
1183static DECLARE_BITMAP(hmm_device_mask, HMM_DEVICE_MAX);
1184static DEFINE_SPINLOCK(hmm_device_lock);
1185static struct class *hmm_device_class;
1186static dev_t hmm_device_devt;
1187
1188static void hmm_device_release(struct device *device)
1189{
1190 struct hmm_device *hmm_device;
1191
1192 hmm_device = container_of(device, struct hmm_device, device);
1193 spin_lock(&hmm_device_lock);
1194 clear_bit(hmm_device->minor, hmm_device_mask);
1195 spin_unlock(&hmm_device_lock);
1196
1197 kfree(hmm_device);
1198}
1199
1200struct hmm_device *hmm_device_new(void *drvdata)
1201{
1202 struct hmm_device *hmm_device;
1203
1204 hmm_device = kzalloc(sizeof(*hmm_device), GFP_KERNEL);
1205 if (!hmm_device)
1206 return ERR_PTR(-ENOMEM);
1207
1208 spin_lock(&hmm_device_lock);
1209 hmm_device->minor = find_first_zero_bit(hmm_device_mask, HMM_DEVICE_MAX);
1210 if (hmm_device->minor >= HMM_DEVICE_MAX) {
1211 spin_unlock(&hmm_device_lock);
1212 kfree(hmm_device);
1213 return ERR_PTR(-EBUSY);
1214 }
1215 set_bit(hmm_device->minor, hmm_device_mask);
1216 spin_unlock(&hmm_device_lock);
1217
1218 dev_set_name(&hmm_device->device, "hmm_device%d", hmm_device->minor);
1219 hmm_device->device.devt = MKDEV(MAJOR(hmm_device_devt),
1220 hmm_device->minor);
1221 hmm_device->device.release = hmm_device_release;
1222 dev_set_drvdata(&hmm_device->device, drvdata);
1223 hmm_device->device.class = hmm_device_class;
1224 device_initialize(&hmm_device->device);
1225
1226 return hmm_device;
1227}
1228EXPORT_SYMBOL(hmm_device_new);
1229
1230void hmm_device_put(struct hmm_device *hmm_device)
1231{
1232 put_device(&hmm_device->device);
1233}
1234EXPORT_SYMBOL(hmm_device_put);
1235
1236static int __init hmm_init(void)
1237{
1238 int ret;
1239
1240 ret = alloc_chrdev_region(&hmm_device_devt, 0,
1241 HMM_DEVICE_MAX,
1242 "hmm_device");
1243 if (ret)
1244 return ret;
1245
1246 hmm_device_class = class_create(THIS_MODULE, "hmm_device");
1247 if (IS_ERR(hmm_device_class)) {
1248 unregister_chrdev_region(hmm_device_devt, HMM_DEVICE_MAX);
1249 return PTR_ERR(hmm_device_class);
1250 }
1251 return 0;
1252}
1253
1254device_initcall(hmm_init);
Jérôme Glissedf6ad692017-09-08 16:12:24 -07001255#endif /* CONFIG_DEVICE_PRIVATE || CONFIG_DEVICE_PUBLIC */