1 // SPDX-License-Identifier: GPL-2.0
2 /* Copyright(c) 2015 Intel Corporation. All rights reserved. */
3 #include <linux/device.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/mmzone.h>
11 #include <linux/swapops.h>
12 #include <linux/types.h>
13 #include <linux/wait_bit.h>
14 #include <linux/xarray.h>
17 static DEFINE_XARRAY(pgmap_array);
20 * The memremap() and memremap_pages() interfaces are alternately used
21 * to map persistent memory namespaces. These interfaces place different
22 * constraints on the alignment and size of the mapping (namespace).
23 * memremap() can map individual PAGE_SIZE pages. memremap_pages() can
24 * only map subsections (2MB), and at least one architecture (PowerPC)
25 * the minimum mapping granularity of memremap_pages() is 16MB.
27 * The role of memremap_compat_align() is to communicate the minimum
28 * arch supported alignment of a namespace such that it can freely
29 * switch modes without violating the arch constraint. Namely, do not
30 * allow a namespace to be PAGE_SIZE aligned since that namespace may be
31 * reconfigured into a mode that requires SUBSECTION_SIZE alignment.
33 #ifndef CONFIG_ARCH_HAS_MEMREMAP_COMPAT_ALIGN
34 unsigned long memremap_compat_align(void)
36 return SUBSECTION_SIZE;
38 EXPORT_SYMBOL_GPL(memremap_compat_align);
42 DEFINE_STATIC_KEY_FALSE(devmap_managed_key);
43 EXPORT_SYMBOL(devmap_managed_key);
45 static void devmap_managed_enable_put(struct dev_pagemap *pgmap)
47 if (pgmap->type == MEMORY_DEVICE_FS_DAX)
48 static_branch_dec(&devmap_managed_key);
51 static void devmap_managed_enable_get(struct dev_pagemap *pgmap)
53 if (pgmap->type == MEMORY_DEVICE_FS_DAX)
54 static_branch_inc(&devmap_managed_key);
57 static void devmap_managed_enable_get(struct dev_pagemap *pgmap)
60 static void devmap_managed_enable_put(struct dev_pagemap *pgmap)
63 #endif /* CONFIG_FS_DAX */
65 static void pgmap_array_delete(struct range *range)
67 xa_store_range(&pgmap_array, PHYS_PFN(range->start), PHYS_PFN(range->end),
72 static unsigned long pfn_first(struct dev_pagemap *pgmap, int range_id)
74 struct range *range = &pgmap->ranges[range_id];
75 unsigned long pfn = PHYS_PFN(range->start);
79 return pfn + vmem_altmap_offset(pgmap_altmap(pgmap));
82 bool pgmap_pfn_valid(struct dev_pagemap *pgmap, unsigned long pfn)
86 for (i = 0; i < pgmap->nr_range; i++) {
87 struct range *range = &pgmap->ranges[i];
89 if (pfn >= PHYS_PFN(range->start) &&
90 pfn <= PHYS_PFN(range->end))
91 return pfn >= pfn_first(pgmap, i);
97 static unsigned long pfn_end(struct dev_pagemap *pgmap, int range_id)
99 const struct range *range = &pgmap->ranges[range_id];
101 return (range->start + range_len(range)) >> PAGE_SHIFT;
104 static unsigned long pfn_len(struct dev_pagemap *pgmap, unsigned long range_id)
106 return (pfn_end(pgmap, range_id) -
107 pfn_first(pgmap, range_id)) >> pgmap->vmemmap_shift;
110 static void pageunmap_range(struct dev_pagemap *pgmap, int range_id)
112 struct range *range = &pgmap->ranges[range_id];
113 struct page *first_page;
115 /* make sure to access a memmap that was actually initialized */
116 first_page = pfn_to_page(pfn_first(pgmap, range_id));
118 /* pages are dead and unused, undo the arch mapping */
120 remove_pfn_range_from_zone(page_zone(first_page), PHYS_PFN(range->start),
121 PHYS_PFN(range_len(range)));
122 if (pgmap->type == MEMORY_DEVICE_PRIVATE) {
123 __remove_pages(PHYS_PFN(range->start),
124 PHYS_PFN(range_len(range)), NULL);
126 arch_remove_memory(range->start, range_len(range),
127 pgmap_altmap(pgmap));
128 kasan_remove_zero_shadow(__va(range->start), range_len(range));
132 untrack_pfn(NULL, PHYS_PFN(range->start), range_len(range));
133 pgmap_array_delete(range);
136 void memunmap_pages(struct dev_pagemap *pgmap)
140 percpu_ref_kill(&pgmap->ref);
141 for (i = 0; i < pgmap->nr_range; i++)
142 percpu_ref_put_many(&pgmap->ref, pfn_len(pgmap, i));
143 wait_for_completion(&pgmap->done);
144 percpu_ref_exit(&pgmap->ref);
146 for (i = 0; i < pgmap->nr_range; i++)
147 pageunmap_range(pgmap, i);
149 WARN_ONCE(pgmap->altmap.alloc, "failed to free all reserved pages\n");
150 devmap_managed_enable_put(pgmap);
152 EXPORT_SYMBOL_GPL(memunmap_pages);
154 static void devm_memremap_pages_release(void *data)
156 memunmap_pages(data);
159 static void dev_pagemap_percpu_release(struct percpu_ref *ref)
161 struct dev_pagemap *pgmap = container_of(ref, struct dev_pagemap, ref);
163 complete(&pgmap->done);
166 static int pagemap_range(struct dev_pagemap *pgmap, struct mhp_params *params,
167 int range_id, int nid)
169 const bool is_private = pgmap->type == MEMORY_DEVICE_PRIVATE;
170 struct range *range = &pgmap->ranges[range_id];
171 struct dev_pagemap *conflict_pgmap;
174 if (WARN_ONCE(pgmap_altmap(pgmap) && range_id > 0,
175 "altmap not supported for multiple ranges\n"))
178 conflict_pgmap = get_dev_pagemap(PHYS_PFN(range->start), NULL);
179 if (conflict_pgmap) {
180 WARN(1, "Conflicting mapping in same section\n");
181 put_dev_pagemap(conflict_pgmap);
185 conflict_pgmap = get_dev_pagemap(PHYS_PFN(range->end), NULL);
186 if (conflict_pgmap) {
187 WARN(1, "Conflicting mapping in same section\n");
188 put_dev_pagemap(conflict_pgmap);
192 is_ram = region_intersects(range->start, range_len(range),
193 IORESOURCE_SYSTEM_RAM, IORES_DESC_NONE);
195 if (is_ram != REGION_DISJOINT) {
196 WARN_ONCE(1, "attempted on %s region %#llx-%#llx\n",
197 is_ram == REGION_MIXED ? "mixed" : "ram",
198 range->start, range->end);
202 error = xa_err(xa_store_range(&pgmap_array, PHYS_PFN(range->start),
203 PHYS_PFN(range->end), pgmap, GFP_KERNEL));
210 error = track_pfn_remap(NULL, ¶ms->pgprot, PHYS_PFN(range->start), 0,
215 if (!mhp_range_allowed(range->start, range_len(range), !is_private)) {
223 * For device private memory we call add_pages() as we only need to
224 * allocate and initialize struct page for the device memory. More-
225 * over the device memory is un-accessible thus we do not want to
226 * create a linear mapping for the memory like arch_add_memory()
229 * For all other device memory types, which are accessible by
230 * the CPU, we do want the linear mapping and thus use
234 error = add_pages(nid, PHYS_PFN(range->start),
235 PHYS_PFN(range_len(range)), params);
237 error = kasan_add_zero_shadow(__va(range->start), range_len(range));
243 error = arch_add_memory(nid, range->start, range_len(range),
250 zone = &NODE_DATA(nid)->node_zones[ZONE_DEVICE];
251 move_pfn_range_to_zone(zone, PHYS_PFN(range->start),
252 PHYS_PFN(range_len(range)), params->altmap,
261 * Initialization of the pages has been deferred until now in order
262 * to allow us to do the work while not holding the hotplug lock.
264 memmap_init_zone_device(&NODE_DATA(nid)->node_zones[ZONE_DEVICE],
265 PHYS_PFN(range->start),
266 PHYS_PFN(range_len(range)), pgmap);
267 percpu_ref_get_many(&pgmap->ref, pfn_len(pgmap, range_id));
272 kasan_remove_zero_shadow(__va(range->start), range_len(range));
274 untrack_pfn(NULL, PHYS_PFN(range->start), range_len(range));
276 pgmap_array_delete(range);
282 * Not device managed version of dev_memremap_pages, undone by
283 * memunmap_pages(). Please use dev_memremap_pages if you have a struct
286 void *memremap_pages(struct dev_pagemap *pgmap, int nid)
288 struct mhp_params params = {
289 .altmap = pgmap_altmap(pgmap),
291 .pgprot = PAGE_KERNEL,
293 const int nr_range = pgmap->nr_range;
296 if (WARN_ONCE(!nr_range, "nr_range must be specified\n"))
297 return ERR_PTR(-EINVAL);
299 switch (pgmap->type) {
300 case MEMORY_DEVICE_PRIVATE:
301 if (!IS_ENABLED(CONFIG_DEVICE_PRIVATE)) {
302 WARN(1, "Device private memory not supported\n");
303 return ERR_PTR(-EINVAL);
305 if (!pgmap->ops || !pgmap->ops->migrate_to_ram) {
306 WARN(1, "Missing migrate_to_ram method\n");
307 return ERR_PTR(-EINVAL);
309 if (!pgmap->ops->page_free) {
310 WARN(1, "Missing page_free method\n");
311 return ERR_PTR(-EINVAL);
314 WARN(1, "Missing owner\n");
315 return ERR_PTR(-EINVAL);
318 case MEMORY_DEVICE_FS_DAX:
319 if (IS_ENABLED(CONFIG_FS_DAX_LIMITED)) {
320 WARN(1, "File system DAX not supported\n");
321 return ERR_PTR(-EINVAL);
324 case MEMORY_DEVICE_GENERIC:
326 case MEMORY_DEVICE_PCI_P2PDMA:
327 params.pgprot = pgprot_noncached(params.pgprot);
330 WARN(1, "Invalid pgmap type %d\n", pgmap->type);
334 init_completion(&pgmap->done);
335 error = percpu_ref_init(&pgmap->ref, dev_pagemap_percpu_release, 0,
338 return ERR_PTR(error);
340 devmap_managed_enable_get(pgmap);
343 * Clear the pgmap nr_range as it will be incremented for each
344 * successfully processed range. This communicates how many
345 * regions to unwind in the abort case.
349 for (i = 0; i < nr_range; i++) {
350 error = pagemap_range(pgmap, ¶ms, i, nid);
357 memunmap_pages(pgmap);
358 pgmap->nr_range = nr_range;
359 return ERR_PTR(error);
362 return __va(pgmap->ranges[0].start);
364 EXPORT_SYMBOL_GPL(memremap_pages);
367 * devm_memremap_pages - remap and provide memmap backing for the given resource
368 * @dev: hosting device for @res
369 * @pgmap: pointer to a struct dev_pagemap
372 * 1/ At a minimum the res and type members of @pgmap must be initialized
373 * by the caller before passing it to this function
375 * 2/ The altmap field may optionally be initialized, in which case
376 * PGMAP_ALTMAP_VALID must be set in pgmap->flags.
378 * 3/ The ref field may optionally be provided, in which pgmap->ref must be
379 * 'live' on entry and will be killed and reaped at
380 * devm_memremap_pages_release() time, or if this routine fails.
382 * 4/ range is expected to be a host memory range that could feasibly be
383 * treated as a "System RAM" range, i.e. not a device mmio range, but
384 * this is not enforced.
386 void *devm_memremap_pages(struct device *dev, struct dev_pagemap *pgmap)
391 ret = memremap_pages(pgmap, dev_to_node(dev));
395 error = devm_add_action_or_reset(dev, devm_memremap_pages_release,
398 return ERR_PTR(error);
401 EXPORT_SYMBOL_GPL(devm_memremap_pages);
403 void devm_memunmap_pages(struct device *dev, struct dev_pagemap *pgmap)
405 devm_release_action(dev, devm_memremap_pages_release, pgmap);
407 EXPORT_SYMBOL_GPL(devm_memunmap_pages);
409 unsigned long vmem_altmap_offset(struct vmem_altmap *altmap)
411 /* number of pfns from base where pfn_to_page() is valid */
413 return altmap->reserve + altmap->free;
417 void vmem_altmap_free(struct vmem_altmap *altmap, unsigned long nr_pfns)
419 altmap->alloc -= nr_pfns;
423 * get_dev_pagemap() - take a new live reference on the dev_pagemap for @pfn
424 * @pfn: page frame number to lookup page_map
425 * @pgmap: optional known pgmap that already has a reference
427 * If @pgmap is non-NULL and covers @pfn it will be returned as-is. If @pgmap
428 * is non-NULL but does not cover @pfn the reference to it will be released.
430 struct dev_pagemap *get_dev_pagemap(unsigned long pfn,
431 struct dev_pagemap *pgmap)
433 resource_size_t phys = PFN_PHYS(pfn);
436 * In the cached case we're already holding a live reference.
439 if (phys >= pgmap->range.start && phys <= pgmap->range.end)
441 put_dev_pagemap(pgmap);
444 /* fall back to slow path lookup */
446 pgmap = xa_load(&pgmap_array, PHYS_PFN(phys));
447 if (pgmap && !percpu_ref_tryget_live(&pgmap->ref))
453 EXPORT_SYMBOL_GPL(get_dev_pagemap);
455 void free_zone_device_page(struct page *page)
457 if (WARN_ON_ONCE(!page->pgmap->ops || !page->pgmap->ops->page_free))
460 mem_cgroup_uncharge(page_folio(page));
463 * Note: we don't expect anonymous compound pages yet. Once supported
464 * and we could PTE-map them similar to THP, we'd have to clear
465 * PG_anon_exclusive on all tail pages.
467 VM_BUG_ON_PAGE(PageAnon(page) && PageCompound(page), page);
469 __ClearPageAnonExclusive(page);
472 * When a device managed page is freed, the page->mapping field
473 * may still contain a (stale) mapping value. For example, the
474 * lower bits of page->mapping may still identify the page as an
475 * anonymous page. Ultimately, this entire field is just stale
476 * and wrong, and it will cause errors if not cleared. One
479 * migrate_vma_pages()
480 * migrate_vma_insert_page()
481 * page_add_new_anon_rmap()
482 * __page_set_anon_rmap()
483 * ...checks page->mapping, via PageAnon(page) call,
484 * and incorrectly concludes that the page is an
485 * anonymous page. Therefore, it incorrectly,
486 * silently fails to set up the new anon rmap.
488 * For other types of ZONE_DEVICE pages, migration is either
489 * handled differently or not done at all, so there is no need
490 * to clear page->mapping.
492 page->mapping = NULL;
493 page->pgmap->ops->page_free(page);
496 * Reset the page count to 1 to prepare for handing out the page again.
498 set_page_count(page, 1);
502 bool __put_devmap_managed_page(struct page *page)
504 if (page->pgmap->type != MEMORY_DEVICE_FS_DAX)
508 * fsdax page refcounts are 1-based, rather than 0-based: if
509 * refcount is 1, then the page is free and the refcount is
510 * stable because nobody holds a reference on the page.
512 if (page_ref_dec_return(page) == 1)
513 wake_up_var(&page->_refcount);
516 EXPORT_SYMBOL(__put_devmap_managed_page);
517 #endif /* CONFIG_FS_DAX */