Loading...
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
18static 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
35unsigned long memremap_compat_align(void)
36{
37 return SUBSECTION_SIZE;
38}
39EXPORT_SYMBOL_GPL(memremap_compat_align);
40#endif
41
42#ifdef CONFIG_FS_DAX
43DEFINE_STATIC_KEY_FALSE(devmap_managed_key);
44EXPORT_SYMBOL(devmap_managed_key);
45
46static 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
52static 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
58static void devmap_managed_enable_get(struct dev_pagemap *pgmap)
59{
60}
61static void devmap_managed_enable_put(struct dev_pagemap *pgmap)
62{
63}
64#endif /* CONFIG_FS_DAX */
65
66static 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
73static 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
83bool 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
98static 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
105static 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
111static 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
137void 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}
156EXPORT_SYMBOL_GPL(memunmap_pages);
157
158static void devm_memremap_pages_release(void *data)
159{
160 memunmap_pages(data);
161}
162
163static 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
170static 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
276err_add_memory:
277 if (!is_private)
278 kasan_remove_zero_shadow(__va(range->start), range_len(range));
279err_kasan:
280 untrack_pfn(NULL, PHYS_PFN(range->start), range_len(range), true);
281err_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 */
292void *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}
381EXPORT_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 */
403void *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}
418EXPORT_SYMBOL_GPL(devm_memremap_pages);
419
420void devm_memunmap_pages(struct device *dev, struct dev_pagemap *pgmap)
421{
422 devm_release_action(dev, devm_memremap_pages_release, pgmap);
423}
424EXPORT_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 */
434struct 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}
457EXPORT_SYMBOL_GPL(get_dev_pagemap);
458
459void free_zone_device_page(struct page *page)
460{
461 if (WARN_ON_ONCE(!page->pgmap->ops || !page->pgmap->ops->page_free))
462 return;
463
464 mem_cgroup_uncharge(page_folio(page));
465
466 /*
467 * Note: we don't expect anonymous compound pages yet. Once supported
468 * and we could PTE-map them similar to THP, we'd have to clear
469 * PG_anon_exclusive on all tail pages.
470 */
471 VM_BUG_ON_PAGE(PageAnon(page) && PageCompound(page), page);
472 if (PageAnon(page))
473 __ClearPageAnonExclusive(page);
474
475 /*
476 * When a device managed page is freed, the folio->mapping field
477 * may still contain a (stale) mapping value. For example, the
478 * lower bits of folio->mapping may still identify the folio as an
479 * anonymous folio. Ultimately, this entire field is just stale
480 * and wrong, and it will cause errors if not cleared.
481 *
482 * For other types of ZONE_DEVICE pages, migration is either
483 * handled differently or not done at all, so there is no need
484 * to clear page->mapping.
485 */
486 page->mapping = NULL;
487 page->pgmap->ops->page_free(page);
488
489 if (page->pgmap->type != MEMORY_DEVICE_PRIVATE &&
490 page->pgmap->type != MEMORY_DEVICE_COHERENT)
491 /*
492 * Reset the page count to 1 to prepare for handing out the page
493 * again.
494 */
495 set_page_count(page, 1);
496 else
497 put_dev_pagemap(page->pgmap);
498}
499
500void zone_device_page_init(struct page *page)
501{
502 /*
503 * Drivers shouldn't be allocating pages after calling
504 * memunmap_pages().
505 */
506 WARN_ON_ONCE(!percpu_ref_tryget_live(&page->pgmap->ref));
507 set_page_count(page, 1);
508 lock_page(page);
509}
510EXPORT_SYMBOL_GPL(zone_device_page_init);
511
512#ifdef CONFIG_FS_DAX
513bool __put_devmap_managed_page_refs(struct page *page, int refs)
514{
515 if (page->pgmap->type != MEMORY_DEVICE_FS_DAX)
516 return false;
517
518 /*
519 * fsdax page refcounts are 1-based, rather than 0-based: if
520 * refcount is 1, then the page is free and the refcount is
521 * stable because nobody holds a reference on the page.
522 */
523 if (page_ref_sub_return(page, refs) == 1)
524 wake_up_var(&page->_refcount);
525 return true;
526}
527EXPORT_SYMBOL(__put_devmap_managed_page_refs);
528#endif /* CONFIG_FS_DAX */
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/mm.h>
8#include <linux/pfn_t.h>
9#include <linux/swap.h>
10#include <linux/swapops.h>
11#include <linux/types.h>
12#include <linux/wait_bit.h>
13#include <linux/xarray.h>
14
15static DEFINE_XARRAY(pgmap_array);
16
17#ifdef CONFIG_DEV_PAGEMAP_OPS
18DEFINE_STATIC_KEY_FALSE(devmap_managed_key);
19EXPORT_SYMBOL(devmap_managed_key);
20static atomic_t devmap_managed_enable;
21
22static void devmap_managed_enable_put(void)
23{
24 if (atomic_dec_and_test(&devmap_managed_enable))
25 static_branch_disable(&devmap_managed_key);
26}
27
28static int devmap_managed_enable_get(struct dev_pagemap *pgmap)
29{
30 if (!pgmap->ops || !pgmap->ops->page_free) {
31 WARN(1, "Missing page_free method\n");
32 return -EINVAL;
33 }
34
35 if (atomic_inc_return(&devmap_managed_enable) == 1)
36 static_branch_enable(&devmap_managed_key);
37 return 0;
38}
39#else
40static int devmap_managed_enable_get(struct dev_pagemap *pgmap)
41{
42 return -EINVAL;
43}
44static void devmap_managed_enable_put(void)
45{
46}
47#endif /* CONFIG_DEV_PAGEMAP_OPS */
48
49static void pgmap_array_delete(struct resource *res)
50{
51 xa_store_range(&pgmap_array, PHYS_PFN(res->start), PHYS_PFN(res->end),
52 NULL, GFP_KERNEL);
53 synchronize_rcu();
54}
55
56static unsigned long pfn_first(struct dev_pagemap *pgmap)
57{
58 return PHYS_PFN(pgmap->res.start) +
59 vmem_altmap_offset(pgmap_altmap(pgmap));
60}
61
62static unsigned long pfn_end(struct dev_pagemap *pgmap)
63{
64 const struct resource *res = &pgmap->res;
65
66 return (res->start + resource_size(res)) >> PAGE_SHIFT;
67}
68
69static unsigned long pfn_next(unsigned long pfn)
70{
71 if (pfn % 1024 == 0)
72 cond_resched();
73 return pfn + 1;
74}
75
76#define for_each_device_pfn(pfn, map) \
77 for (pfn = pfn_first(map); pfn < pfn_end(map); pfn = pfn_next(pfn))
78
79static void dev_pagemap_kill(struct dev_pagemap *pgmap)
80{
81 if (pgmap->ops && pgmap->ops->kill)
82 pgmap->ops->kill(pgmap);
83 else
84 percpu_ref_kill(pgmap->ref);
85}
86
87static void dev_pagemap_cleanup(struct dev_pagemap *pgmap)
88{
89 if (pgmap->ops && pgmap->ops->cleanup) {
90 pgmap->ops->cleanup(pgmap);
91 } else {
92 wait_for_completion(&pgmap->done);
93 percpu_ref_exit(pgmap->ref);
94 }
95 /*
96 * Undo the pgmap ref assignment for the internal case as the
97 * caller may re-enable the same pgmap.
98 */
99 if (pgmap->ref == &pgmap->internal_ref)
100 pgmap->ref = NULL;
101}
102
103void memunmap_pages(struct dev_pagemap *pgmap)
104{
105 struct resource *res = &pgmap->res;
106 struct page *first_page;
107 unsigned long pfn;
108 int nid;
109
110 dev_pagemap_kill(pgmap);
111 for_each_device_pfn(pfn, pgmap)
112 put_page(pfn_to_page(pfn));
113 dev_pagemap_cleanup(pgmap);
114
115 /* make sure to access a memmap that was actually initialized */
116 first_page = pfn_to_page(pfn_first(pgmap));
117
118 /* pages are dead and unused, undo the arch mapping */
119 nid = page_to_nid(first_page);
120
121 mem_hotplug_begin();
122 if (pgmap->type == MEMORY_DEVICE_PRIVATE) {
123 __remove_pages(page_zone(first_page), PHYS_PFN(res->start),
124 PHYS_PFN(resource_size(res)), NULL);
125 } else {
126 arch_remove_memory(nid, res->start, resource_size(res),
127 pgmap_altmap(pgmap));
128 kasan_remove_zero_shadow(__va(res->start), resource_size(res));
129 }
130 mem_hotplug_done();
131
132 untrack_pfn(NULL, PHYS_PFN(res->start), resource_size(res));
133 pgmap_array_delete(res);
134 WARN_ONCE(pgmap->altmap.alloc, "failed to free all reserved pages\n");
135 devmap_managed_enable_put();
136}
137EXPORT_SYMBOL_GPL(memunmap_pages);
138
139static void devm_memremap_pages_release(void *data)
140{
141 memunmap_pages(data);
142}
143
144static void dev_pagemap_percpu_release(struct percpu_ref *ref)
145{
146 struct dev_pagemap *pgmap =
147 container_of(ref, struct dev_pagemap, internal_ref);
148
149 complete(&pgmap->done);
150}
151
152/*
153 * Not device managed version of dev_memremap_pages, undone by
154 * memunmap_pages(). Please use dev_memremap_pages if you have a struct
155 * device available.
156 */
157void *memremap_pages(struct dev_pagemap *pgmap, int nid)
158{
159 struct resource *res = &pgmap->res;
160 struct dev_pagemap *conflict_pgmap;
161 struct mhp_restrictions restrictions = {
162 /*
163 * We do not want any optional features only our own memmap
164 */
165 .altmap = pgmap_altmap(pgmap),
166 };
167 pgprot_t pgprot = PAGE_KERNEL;
168 int error, is_ram;
169 bool need_devmap_managed = true;
170
171 switch (pgmap->type) {
172 case MEMORY_DEVICE_PRIVATE:
173 if (!IS_ENABLED(CONFIG_DEVICE_PRIVATE)) {
174 WARN(1, "Device private memory not supported\n");
175 return ERR_PTR(-EINVAL);
176 }
177 if (!pgmap->ops || !pgmap->ops->migrate_to_ram) {
178 WARN(1, "Missing migrate_to_ram method\n");
179 return ERR_PTR(-EINVAL);
180 }
181 break;
182 case MEMORY_DEVICE_FS_DAX:
183 if (!IS_ENABLED(CONFIG_ZONE_DEVICE) ||
184 IS_ENABLED(CONFIG_FS_DAX_LIMITED)) {
185 WARN(1, "File system DAX not supported\n");
186 return ERR_PTR(-EINVAL);
187 }
188 break;
189 case MEMORY_DEVICE_DEVDAX:
190 case MEMORY_DEVICE_PCI_P2PDMA:
191 need_devmap_managed = false;
192 break;
193 default:
194 WARN(1, "Invalid pgmap type %d\n", pgmap->type);
195 break;
196 }
197
198 if (!pgmap->ref) {
199 if (pgmap->ops && (pgmap->ops->kill || pgmap->ops->cleanup))
200 return ERR_PTR(-EINVAL);
201
202 init_completion(&pgmap->done);
203 error = percpu_ref_init(&pgmap->internal_ref,
204 dev_pagemap_percpu_release, 0, GFP_KERNEL);
205 if (error)
206 return ERR_PTR(error);
207 pgmap->ref = &pgmap->internal_ref;
208 } else {
209 if (!pgmap->ops || !pgmap->ops->kill || !pgmap->ops->cleanup) {
210 WARN(1, "Missing reference count teardown definition\n");
211 return ERR_PTR(-EINVAL);
212 }
213 }
214
215 if (need_devmap_managed) {
216 error = devmap_managed_enable_get(pgmap);
217 if (error)
218 return ERR_PTR(error);
219 }
220
221 conflict_pgmap = get_dev_pagemap(PHYS_PFN(res->start), NULL);
222 if (conflict_pgmap) {
223 WARN(1, "Conflicting mapping in same section\n");
224 put_dev_pagemap(conflict_pgmap);
225 error = -ENOMEM;
226 goto err_array;
227 }
228
229 conflict_pgmap = get_dev_pagemap(PHYS_PFN(res->end), NULL);
230 if (conflict_pgmap) {
231 WARN(1, "Conflicting mapping in same section\n");
232 put_dev_pagemap(conflict_pgmap);
233 error = -ENOMEM;
234 goto err_array;
235 }
236
237 is_ram = region_intersects(res->start, resource_size(res),
238 IORESOURCE_SYSTEM_RAM, IORES_DESC_NONE);
239
240 if (is_ram != REGION_DISJOINT) {
241 WARN_ONCE(1, "%s attempted on %s region %pr\n", __func__,
242 is_ram == REGION_MIXED ? "mixed" : "ram", res);
243 error = -ENXIO;
244 goto err_array;
245 }
246
247 error = xa_err(xa_store_range(&pgmap_array, PHYS_PFN(res->start),
248 PHYS_PFN(res->end), pgmap, GFP_KERNEL));
249 if (error)
250 goto err_array;
251
252 if (nid < 0)
253 nid = numa_mem_id();
254
255 error = track_pfn_remap(NULL, &pgprot, PHYS_PFN(res->start), 0,
256 resource_size(res));
257 if (error)
258 goto err_pfn_remap;
259
260 mem_hotplug_begin();
261
262 /*
263 * For device private memory we call add_pages() as we only need to
264 * allocate and initialize struct page for the device memory. More-
265 * over the device memory is un-accessible thus we do not want to
266 * create a linear mapping for the memory like arch_add_memory()
267 * would do.
268 *
269 * For all other device memory types, which are accessible by
270 * the CPU, we do want the linear mapping and thus use
271 * arch_add_memory().
272 */
273 if (pgmap->type == MEMORY_DEVICE_PRIVATE) {
274 error = add_pages(nid, PHYS_PFN(res->start),
275 PHYS_PFN(resource_size(res)), &restrictions);
276 } else {
277 error = kasan_add_zero_shadow(__va(res->start), resource_size(res));
278 if (error) {
279 mem_hotplug_done();
280 goto err_kasan;
281 }
282
283 error = arch_add_memory(nid, res->start, resource_size(res),
284 &restrictions);
285 }
286
287 if (!error) {
288 struct zone *zone;
289
290 zone = &NODE_DATA(nid)->node_zones[ZONE_DEVICE];
291 move_pfn_range_to_zone(zone, PHYS_PFN(res->start),
292 PHYS_PFN(resource_size(res)), restrictions.altmap);
293 }
294
295 mem_hotplug_done();
296 if (error)
297 goto err_add_memory;
298
299 /*
300 * Initialization of the pages has been deferred until now in order
301 * to allow us to do the work while not holding the hotplug lock.
302 */
303 memmap_init_zone_device(&NODE_DATA(nid)->node_zones[ZONE_DEVICE],
304 PHYS_PFN(res->start),
305 PHYS_PFN(resource_size(res)), pgmap);
306 percpu_ref_get_many(pgmap->ref, pfn_end(pgmap) - pfn_first(pgmap));
307 return __va(res->start);
308
309 err_add_memory:
310 kasan_remove_zero_shadow(__va(res->start), resource_size(res));
311 err_kasan:
312 untrack_pfn(NULL, PHYS_PFN(res->start), resource_size(res));
313 err_pfn_remap:
314 pgmap_array_delete(res);
315 err_array:
316 dev_pagemap_kill(pgmap);
317 dev_pagemap_cleanup(pgmap);
318 devmap_managed_enable_put();
319 return ERR_PTR(error);
320}
321EXPORT_SYMBOL_GPL(memremap_pages);
322
323/**
324 * devm_memremap_pages - remap and provide memmap backing for the given resource
325 * @dev: hosting device for @res
326 * @pgmap: pointer to a struct dev_pagemap
327 *
328 * Notes:
329 * 1/ At a minimum the res and type members of @pgmap must be initialized
330 * by the caller before passing it to this function
331 *
332 * 2/ The altmap field may optionally be initialized, in which case
333 * PGMAP_ALTMAP_VALID must be set in pgmap->flags.
334 *
335 * 3/ The ref field may optionally be provided, in which pgmap->ref must be
336 * 'live' on entry and will be killed and reaped at
337 * devm_memremap_pages_release() time, or if this routine fails.
338 *
339 * 4/ res is expected to be a host memory range that could feasibly be
340 * treated as a "System RAM" range, i.e. not a device mmio range, but
341 * this is not enforced.
342 */
343void *devm_memremap_pages(struct device *dev, struct dev_pagemap *pgmap)
344{
345 int error;
346 void *ret;
347
348 ret = memremap_pages(pgmap, dev_to_node(dev));
349 if (IS_ERR(ret))
350 return ret;
351
352 error = devm_add_action_or_reset(dev, devm_memremap_pages_release,
353 pgmap);
354 if (error)
355 return ERR_PTR(error);
356 return ret;
357}
358EXPORT_SYMBOL_GPL(devm_memremap_pages);
359
360void devm_memunmap_pages(struct device *dev, struct dev_pagemap *pgmap)
361{
362 devm_release_action(dev, devm_memremap_pages_release, pgmap);
363}
364EXPORT_SYMBOL_GPL(devm_memunmap_pages);
365
366unsigned long vmem_altmap_offset(struct vmem_altmap *altmap)
367{
368 /* number of pfns from base where pfn_to_page() is valid */
369 if (altmap)
370 return altmap->reserve + altmap->free;
371 return 0;
372}
373
374void vmem_altmap_free(struct vmem_altmap *altmap, unsigned long nr_pfns)
375{
376 altmap->alloc -= nr_pfns;
377}
378
379/**
380 * get_dev_pagemap() - take a new live reference on the dev_pagemap for @pfn
381 * @pfn: page frame number to lookup page_map
382 * @pgmap: optional known pgmap that already has a reference
383 *
384 * If @pgmap is non-NULL and covers @pfn it will be returned as-is. If @pgmap
385 * is non-NULL but does not cover @pfn the reference to it will be released.
386 */
387struct dev_pagemap *get_dev_pagemap(unsigned long pfn,
388 struct dev_pagemap *pgmap)
389{
390 resource_size_t phys = PFN_PHYS(pfn);
391
392 /*
393 * In the cached case we're already holding a live reference.
394 */
395 if (pgmap) {
396 if (phys >= pgmap->res.start && phys <= pgmap->res.end)
397 return pgmap;
398 put_dev_pagemap(pgmap);
399 }
400
401 /* fall back to slow path lookup */
402 rcu_read_lock();
403 pgmap = xa_load(&pgmap_array, PHYS_PFN(phys));
404 if (pgmap && !percpu_ref_tryget_live(pgmap->ref))
405 pgmap = NULL;
406 rcu_read_unlock();
407
408 return pgmap;
409}
410EXPORT_SYMBOL_GPL(get_dev_pagemap);
411
412#ifdef CONFIG_DEV_PAGEMAP_OPS
413void __put_devmap_managed_page(struct page *page)
414{
415 int count = page_ref_dec_return(page);
416
417 /*
418 * If refcount is 1 then page is freed and refcount is stable as nobody
419 * holds a reference on the page.
420 */
421 if (count == 1) {
422 /* Clear Active bit in case of parallel mark_page_accessed */
423 __ClearPageActive(page);
424 __ClearPageWaiters(page);
425
426 mem_cgroup_uncharge(page);
427
428 /*
429 * When a device_private page is freed, the page->mapping field
430 * may still contain a (stale) mapping value. For example, the
431 * lower bits of page->mapping may still identify the page as
432 * an anonymous page. Ultimately, this entire field is just
433 * stale and wrong, and it will cause errors if not cleared.
434 * One example is:
435 *
436 * migrate_vma_pages()
437 * migrate_vma_insert_page()
438 * page_add_new_anon_rmap()
439 * __page_set_anon_rmap()
440 * ...checks page->mapping, via PageAnon(page) call,
441 * and incorrectly concludes that the page is an
442 * anonymous page. Therefore, it incorrectly,
443 * silently fails to set up the new anon rmap.
444 *
445 * For other types of ZONE_DEVICE pages, migration is either
446 * handled differently or not done at all, so there is no need
447 * to clear page->mapping.
448 */
449 if (is_device_private_page(page))
450 page->mapping = NULL;
451
452 page->pgmap->ops->page_free(page);
453 } else if (!count)
454 __put_page(page);
455}
456EXPORT_SYMBOL(__put_devmap_managed_page);
457#endif /* CONFIG_DEV_PAGEMAP_OPS */