1 // SPDX-License-Identifier: GPL-2.0
2 /* Copyright(c) 2015 Intel Corporation. All rights reserved. */
3 #include <linux/device.h>
4 #include <linux/io.h>
5 #include <linux/kasan.h>
6 #include <linux/memory_hotplug.h>
7 #include <linux/memremap.h>
8 #include <linux/pfn_t.h>
9 #include <linux/swap.h>
10 #include <linux/mm.h>
11 #include <linux/mmzone.h>
12 #include <linux/swapops.h>
13 #include <linux/types.h>
14 #include <linux/wait_bit.h>
15 #include <linux/xarray.h>
16 #include "internal.h"
17
18 static DEFINE_XARRAY(pgmap_array);
19
20 /*
21 * The memremap() and memremap_pages() interfaces are alternately used
22 * to map persistent memory namespaces. These interfaces place different
23 * constraints on the alignment and size of the mapping (namespace).
24 * memremap() can map individual PAGE_SIZE pages. memremap_pages() can
25 * only map subsections (2MB), and at least one architecture (PowerPC)
26 * the minimum mapping granularity of memremap_pages() is 16MB.
27 *
28 * The role of memremap_compat_align() is to communicate the minimum
29 * arch supported alignment of a namespace such that it can freely
30 * switch modes without violating the arch constraint. Namely, do not
31 * allow a namespace to be PAGE_SIZE aligned since that namespace may be
32 * reconfigured into a mode that requires SUBSECTION_SIZE alignment.
33 */
34 #ifndef CONFIG_ARCH_HAS_MEMREMAP_COMPAT_ALIGN
35 unsigned long memremap_compat_align(void)
36 {
37 return SUBSECTION_SIZE;
38 }
39 EXPORT_SYMBOL_GPL(memremap_compat_align);
40 #endif
41
42 #ifdef CONFIG_FS_DAX
43 DEFINE_STATIC_KEY_FALSE(devmap_managed_key);
44 EXPORT_SYMBOL(devmap_managed_key);
45
46 static void devmap_managed_enable_put(struct dev_pagemap *pgmap)
47 {
48 if (pgmap->type == MEMORY_DEVICE_FS_DAX)
49 static_branch_dec(&devmap_managed_key);
50 }
51
52 static void devmap_managed_enable_get(struct dev_pagemap *pgmap)
53 {
54 if (pgmap->type == MEMORY_DEVICE_FS_DAX)
55 static_branch_inc(&devmap_managed_key);
56 }
57 #else
58 static void devmap_managed_enable_get(struct dev_pagemap *pgmap)
59 {
60 }
61 static void devmap_managed_enable_put(struct dev_pagemap *pgmap)
62 {
63 }
64 #endif /* CONFIG_FS_DAX */
65
66 static void pgmap_array_delete(struct range *range)
67 {
68 xa_store_range(&pgmap_array, PHYS_PFN(range->start), PHYS_PFN(range->end),
69 NULL, GFP_KERNEL);
70 synchronize_rcu();
71 }
72
73 static unsigned long pfn_first(struct dev_pagemap *pgmap, int range_id)
74 {
75 struct range *range = &pgmap->ranges[range_id];
76 unsigned long pfn = PHYS_PFN(range->start);
77
78 if (range_id)
79 return pfn;
80 return pfn + vmem_altmap_offset(pgmap_altmap(pgmap));
81 }
82
83 bool pgmap_pfn_valid(struct dev_pagemap *pgmap, unsigned long pfn)
84 {
85 int i;
86
87 for (i = 0; i < pgmap->nr_range; i++) {
88 struct range *range = &pgmap->ranges[i];
89
90 if (pfn >= PHYS_PFN(range->start) &&
91 pfn <= PHYS_PFN(range->end))
92 return pfn >= pfn_first(pgmap, i);
93 }
94
95 return false;
96 }
97
98 static unsigned long pfn_end(struct dev_pagemap *pgmap, int range_id)
99 {
100 const struct range *range = &pgmap->ranges[range_id];
101
102 return (range->start + range_len(range)) >> PAGE_SHIFT;
103 }
104
105 static unsigned long pfn_len(struct dev_pagemap *pgmap, unsigned long range_id)
106 {
107 return (pfn_end(pgmap, range_id) -
108 pfn_first(pgmap, range_id)) >> pgmap->vmemmap_shift;
109 }
110
111 static void pageunmap_range(struct dev_pagemap *pgmap, int range_id)
112 {
113 struct range *range = &pgmap->ranges[range_id];
114 struct page *first_page;
115
116 /* make sure to access a memmap that was actually initialized */
117 first_page = pfn_to_page(pfn_first(pgmap, range_id));
118
119 /* pages are dead and unused, undo the arch mapping */
120 mem_hotplug_begin();
121 remove_pfn_range_from_zone(page_zone(first_page), PHYS_PFN(range->start),
122 PHYS_PFN(range_len(range)));
123 if (pgmap->type == MEMORY_DEVICE_PRIVATE) {
124 __remove_pages(PHYS_PFN(range->start),
125 PHYS_PFN(range_len(range)), NULL);
126 } else {
127 arch_remove_memory(range->start, range_len(range),
128 pgmap_altmap(pgmap));
129 kasan_remove_zero_shadow(__va(range->start), range_len(range));
130 }
131 mem_hotplug_done();
132
133 untrack_pfn(NULL, PHYS_PFN(range->start), range_len(range), true);
134 pgmap_array_delete(range);
135 }
136
137 void memunmap_pages(struct dev_pagemap *pgmap)
138 {
139 int i;
140
141 percpu_ref_kill(&pgmap->ref);
142 if (pgmap->type != MEMORY_DEVICE_PRIVATE &&
143 pgmap->type != MEMORY_DEVICE_COHERENT)
144 for (i = 0; i < pgmap->nr_range; i++)
145 percpu_ref_put_many(&pgmap->ref, pfn_len(pgmap, i));
146
147 wait_for_completion(&pgmap->done);
148
149 for (i = 0; i < pgmap->nr_range; i++)
150 pageunmap_range(pgmap, i);
151 percpu_ref_exit(&pgmap->ref);
152
153 WARN_ONCE(pgmap->altmap.alloc, "failed to free all reserved pages\n");
154 devmap_managed_enable_put(pgmap);
155 }
156 EXPORT_SYMBOL_GPL(memunmap_pages);
157
158 static void devm_memremap_pages_release(void *data)
159 {
160 memunmap_pages(data);
161 }
162
163 static void dev_pagemap_percpu_release(struct percpu_ref *ref)
164 {
165 struct dev_pagemap *pgmap = container_of(ref, struct dev_pagemap, ref);
166
167 complete(&pgmap->done);
168 }
169
170 static int pagemap_range(struct dev_pagemap *pgmap, struct mhp_params *params,
171 int range_id, int nid)
172 {
173 const bool is_private = pgmap->type == MEMORY_DEVICE_PRIVATE;
174 struct range *range = &pgmap->ranges[range_id];
175 struct dev_pagemap *conflict_pgmap;
176 int error, is_ram;
177
178 if (WARN_ONCE(pgmap_altmap(pgmap) && range_id > 0,
179 "altmap not supported for multiple ranges\n"))
180 return -EINVAL;
181
182 conflict_pgmap = get_dev_pagemap(PHYS_PFN(range->start), NULL);
183 if (conflict_pgmap) {
184 WARN(1, "Conflicting mapping in same section\n");
185 put_dev_pagemap(conflict_pgmap);
186 return -ENOMEM;
187 }
188
189 conflict_pgmap = get_dev_pagemap(PHYS_PFN(range->end), NULL);
190 if (conflict_pgmap) {
191 WARN(1, "Conflicting mapping in same section\n");
192 put_dev_pagemap(conflict_pgmap);
193 return -ENOMEM;
194 }
195
196 is_ram = region_intersects(range->start, range_len(range),
197 IORESOURCE_SYSTEM_RAM, IORES_DESC_NONE);
198
199 if (is_ram != REGION_DISJOINT) {
200 WARN_ONCE(1, "attempted on %s region %#llx-%#llx\n",
201 is_ram == REGION_MIXED ? "mixed" : "ram",
202 range->start, range->end);
203 return -ENXIO;
204 }
205
206 error = xa_err(xa_store_range(&pgmap_array, PHYS_PFN(range->start),
207 PHYS_PFN(range->end), pgmap, GFP_KERNEL));
208 if (error)
209 return error;
210
211 if (nid < 0)
212 nid = numa_mem_id();
213
214 error = track_pfn_remap(NULL, ¶ms->pgprot, PHYS_PFN(range->start), 0,
215 range_len(range));
216 if (error)
217 goto err_pfn_remap;
218
219 if (!mhp_range_allowed(range->start, range_len(range), !is_private)) {
220 error = -EINVAL;
221 goto err_kasan;
222 }
223
224 mem_hotplug_begin();
225
226 /*
227 * For device private memory we call add_pages() as we only need to
228 * allocate and initialize struct page for the device memory. More-
229 * over the device memory is un-accessible thus we do not want to
230 * create a linear mapping for the memory like arch_add_memory()
231 * would do.
232 *
233 * For all other device memory types, which are accessible by
234 * the CPU, we do want the linear mapping and thus use
235 * arch_add_memory().
236 */
237 if (is_private) {
238 error = add_pages(nid, PHYS_PFN(range->start),
239 PHYS_PFN(range_len(range)), params);
240 } else {
241 error = kasan_add_zero_shadow(__va(range->start), range_len(range));
242 if (error) {
243 mem_hotplug_done();
244 goto err_kasan;
245 }
246
247 error = arch_add_memory(nid, range->start, range_len(range),
248 params);
249 }
250
251 if (!error) {
252 struct zone *zone;
253
254 zone = &NODE_DATA(nid)->node_zones[ZONE_DEVICE];
255 move_pfn_range_to_zone(zone, PHYS_PFN(range->start),
256 PHYS_PFN(range_len(range)), params->altmap,
257 MIGRATE_MOVABLE);
258 }
259
260 mem_hotplug_done();
261 if (error)
262 goto err_add_memory;
263
264 /*
265 * Initialization of the pages has been deferred until now in order
266 * to allow us to do the work while not holding the hotplug lock.
267 */
268 memmap_init_zone_device(&NODE_DATA(nid)->node_zones[ZONE_DEVICE],
269 PHYS_PFN(range->start),
270 PHYS_PFN(range_len(range)), pgmap);
271 if (pgmap->type != MEMORY_DEVICE_PRIVATE &&
272 pgmap->type != MEMORY_DEVICE_COHERENT)
273 percpu_ref_get_many(&pgmap->ref, pfn_len(pgmap, range_id));
274 return 0;
275
276 err_add_memory:
277 if (!is_private)
278 kasan_remove_zero_shadow(__va(range->start), range_len(range));
279 err_kasan:
280 untrack_pfn(NULL, PHYS_PFN(range->start), range_len(range), true);
281 err_pfn_remap:
282 pgmap_array_delete(range);
283 return error;
284 }
285
286
287 /*
288 * Not device managed version of devm_memremap_pages, undone by
289 * memunmap_pages(). Please use devm_memremap_pages if you have a struct
290 * device available.
291 */
292 void *memremap_pages(struct dev_pagemap *pgmap, int nid)
293 {
294 struct mhp_params params = {
295 .altmap = pgmap_altmap(pgmap),
296 .pgmap = pgmap,
297 .pgprot = PAGE_KERNEL,
298 };
299 const int nr_range = pgmap->nr_range;
300 int error, i;
301
302 if (WARN_ONCE(!nr_range, "nr_range must be specified\n"))
303 return ERR_PTR(-EINVAL);
304
305 switch (pgmap->type) {
306 case MEMORY_DEVICE_PRIVATE:
307 if (!IS_ENABLED(CONFIG_DEVICE_PRIVATE)) {
308 WARN(1, "Device private memory not supported\n");
309 return ERR_PTR(-EINVAL);
310 }
311 if (!pgmap->ops || !pgmap->ops->migrate_to_ram) {
312 WARN(1, "Missing migrate_to_ram method\n");
313 return ERR_PTR(-EINVAL);
314 }
315 if (!pgmap->ops->page_free) {
316 WARN(1, "Missing page_free method\n");
317 return ERR_PTR(-EINVAL);
318 }
319 if (!pgmap->owner) {
320 WARN(1, "Missing owner\n");
321 return ERR_PTR(-EINVAL);
322 }
323 break;
324 case MEMORY_DEVICE_COHERENT:
325 if (!pgmap->ops->page_free) {
326 WARN(1, "Missing page_free method\n");
327 return ERR_PTR(-EINVAL);
328 }
329 if (!pgmap->owner) {
330 WARN(1, "Missing owner\n");
331 return ERR_PTR(-EINVAL);
332 }
333 break;
334 case MEMORY_DEVICE_FS_DAX:
335 if (IS_ENABLED(CONFIG_FS_DAX_LIMITED)) {
336 WARN(1, "File system DAX not supported\n");
337 return ERR_PTR(-EINVAL);
338 }
339 params.pgprot = pgprot_decrypted(params.pgprot);
340 break;
341 case MEMORY_DEVICE_GENERIC:
342 break;
343 case MEMORY_DEVICE_PCI_P2PDMA:
344 params.pgprot = pgprot_noncached(params.pgprot);
345 break;
346 default:
347 WARN(1, "Invalid pgmap type %d\n", pgmap->type);
348 break;
349 }
350
351 init_completion(&pgmap->done);
352 error = percpu_ref_init(&pgmap->ref, dev_pagemap_percpu_release, 0,
353 GFP_KERNEL);
354 if (error)
355 return ERR_PTR(error);
356
357 devmap_managed_enable_get(pgmap);
358
359 /*
360 * Clear the pgmap nr_range as it will be incremented for each
361 * successfully processed range. This communicates how many
362 * regions to unwind in the abort case.
363 */
364 pgmap->nr_range = 0;
365 error = 0;
366 for (i = 0; i < nr_range; i++) {
367 error = pagemap_range(pgmap, ¶ms, i, nid);
368 if (error)
369 break;
370 pgmap->nr_range++;
371 }
372
373 if (i < nr_range) {
374 memunmap_pages(pgmap);
375 pgmap->nr_range = nr_range;
376 return ERR_PTR(error);
377 }
378
379 return __va(pgmap->ranges[0].start);
380 }
381 EXPORT_SYMBOL_GPL(memremap_pages);
382
383 /**
384 * devm_memremap_pages - remap and provide memmap backing for the given resource
385 * @dev: hosting device for @res
386 * @pgmap: pointer to a struct dev_pagemap
387 *
388 * Notes:
389 * 1/ At a minimum the range and type members of @pgmap must be initialized
390 * by the caller before passing it to this function
391 *
392 * 2/ The altmap field may optionally be initialized, in which case
393 * PGMAP_ALTMAP_VALID must be set in pgmap->flags.
394 *
395 * 3/ The ref field may optionally be provided, in which pgmap->ref must be
396 * 'live' on entry and will be killed and reaped at
397 * devm_memremap_pages_release() time, or if this routine fails.
398 *
399 * 4/ range is expected to be a host memory range that could feasibly be
400 * treated as a "System RAM" range, i.e. not a device mmio range, but
401 * this is not enforced.
402 */
403 void *devm_memremap_pages(struct device *dev, struct dev_pagemap *pgmap)
404 {
405 int error;
406 void *ret;
407
408 ret = memremap_pages(pgmap, dev_to_node(dev));
409 if (IS_ERR(ret))
410 return ret;
411
412 error = devm_add_action_or_reset(dev, devm_memremap_pages_release,
413 pgmap);
414 if (error)
415 return ERR_PTR(error);
416 return ret;
417 }
418 EXPORT_SYMBOL_GPL(devm_memremap_pages);
419
420 void devm_memunmap_pages(struct device *dev, struct dev_pagemap *pgmap)
421 {
422 devm_release_action(dev, devm_memremap_pages_release, pgmap);
423 }
424 EXPORT_SYMBOL_GPL(devm_memunmap_pages);
425
426 /**
427 * get_dev_pagemap() - take a new live reference on the dev_pagemap for @pfn
428 * @pfn: page frame number to lookup page_map
429 * @pgmap: optional known pgmap that already has a reference
430 *
431 * If @pgmap is non-NULL and covers @pfn it will be returned as-is. If @pgmap
432 * is non-NULL but does not cover @pfn the reference to it will be released.
433 */
434 struct dev_pagemap *get_dev_pagemap(unsigned long pfn,
435 struct dev_pagemap *pgmap)
436 {
437 resource_size_t phys = PFN_PHYS(pfn);
438
439 /*
440 * In the cached case we're already holding a live reference.
441 */
442 if (pgmap) {
443 if (phys >= pgmap->range.start && phys <= pgmap->range.end)
444 return pgmap;
445 put_dev_pagemap(pgmap);
446 }
447
448 /* fall back to slow path lookup */
449 rcu_read_lock();
450 pgmap = xa_load(&pgmap_array, PHYS_PFN(phys));
451 if (pgmap && !percpu_ref_tryget_live_rcu(&pgmap->ref))
452 pgmap = NULL;
453 rcu_read_unlock();
454
455 return pgmap;
456 }
457 EXPORT_SYMBOL_GPL(get_dev_pagemap);
458
459 void free_zone_device_folio(struct folio *folio)
460 {
461 if (WARN_ON_ONCE(!folio->page.pgmap->ops ||
462 !folio->page.pgmap->ops->page_free))
463 return;
464
465 mem_cgroup_uncharge(folio);
466
467 /*
468 * Note: we don't expect anonymous compound pages yet. Once supported
469 * and we could PTE-map them similar to THP, we'd have to clear
470 * PG_anon_exclusive on all tail pages.
471 */
472 if (folio_test_anon(folio)) {
473 VM_BUG_ON_FOLIO(folio_test_large(folio), folio);
474 __ClearPageAnonExclusive(folio_page(folio, 0));
475 }
476
477 /*
478 * When a device managed page is freed, the folio->mapping field
479 * may still contain a (stale) mapping value. For example, the
480 * lower bits of folio->mapping may still identify the folio as an
481 * anonymous folio. Ultimately, this entire field is just stale
482 * and wrong, and it will cause errors if not cleared.
483 *
484 * For other types of ZONE_DEVICE pages, migration is either
485 * handled differently or not done at all, so there is no need
486 * to clear folio->mapping.
487 */
488 folio->mapping = NULL;
489 folio->page.pgmap->ops->page_free(folio_page(folio, 0));
490
491 if (folio->page.pgmap->type != MEMORY_DEVICE_PRIVATE &&
492 folio->page.pgmap->type != MEMORY_DEVICE_COHERENT)
493 /*
494 * Reset the refcount to 1 to prepare for handing out the page
495 * again.
496 */
497 folio_set_count(folio, 1);
498 else
499 put_dev_pagemap(folio->page.pgmap);
500 }
501
502 void zone_device_page_init(struct page *page)
503 {
504 /*
505 * Drivers shouldn't be allocating pages after calling
506 * memunmap_pages().
507 */
508 WARN_ON_ONCE(!percpu_ref_tryget_live(&page->pgmap->ref));
509 set_page_count(page, 1);
510 lock_page(page);
511 }
512 EXPORT_SYMBOL_GPL(zone_device_page_init);
513
514 #ifdef CONFIG_FS_DAX
515 bool __put_devmap_managed_folio_refs(struct folio *folio, int refs)
516 {
517 if (folio->page.pgmap->type != MEMORY_DEVICE_FS_DAX)
518 return false;
519
520 /*
521 * fsdax page refcounts are 1-based, rather than 0-based: if
522 * refcount is 1, then the page is free and the refcount is
523 * stable because nobody holds a reference on the page.
524 */
525 if (folio_ref_sub_return(folio, refs) == 1)
526 wake_up_var(&folio->_refcount);
527 return true;
528 }
529 EXPORT_SYMBOL(__put_devmap_managed_folio_refs);
530 #endif /* CONFIG_FS_DAX */
531
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.