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>
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>
15 static DEFINE_XARRAY(pgmap_array);
16 #define SECTION_MASK ~((1UL << PA_SECTION_SHIFT) - 1)
17 #define SECTION_SIZE (1UL << PA_SECTION_SHIFT)
19 #ifdef CONFIG_DEV_PAGEMAP_OPS
20 DEFINE_STATIC_KEY_FALSE(devmap_managed_key);
21 EXPORT_SYMBOL(devmap_managed_key);
22 static atomic_t devmap_managed_enable;
24 static void devmap_managed_enable_put(void)
26 if (atomic_dec_and_test(&devmap_managed_enable))
27 static_branch_disable(&devmap_managed_key);
30 static int devmap_managed_enable_get(struct dev_pagemap *pgmap)
32 if (!pgmap->ops || !pgmap->ops->page_free) {
33 WARN(1, "Missing page_free method\n");
37 if (atomic_inc_return(&devmap_managed_enable) == 1)
38 static_branch_enable(&devmap_managed_key);
42 static int devmap_managed_enable_get(struct dev_pagemap *pgmap)
46 static void devmap_managed_enable_put(void)
49 #endif /* CONFIG_DEV_PAGEMAP_OPS */
51 static void pgmap_array_delete(struct resource *res)
53 xa_store_range(&pgmap_array, PHYS_PFN(res->start), PHYS_PFN(res->end),
58 static unsigned long pfn_first(struct dev_pagemap *pgmap)
60 return PHYS_PFN(pgmap->res.start) +
61 vmem_altmap_offset(pgmap_altmap(pgmap));
64 static unsigned long pfn_end(struct dev_pagemap *pgmap)
66 const struct resource *res = &pgmap->res;
68 return (res->start + resource_size(res)) >> PAGE_SHIFT;
71 static unsigned long pfn_next(unsigned long pfn)
78 #define for_each_device_pfn(pfn, map) \
79 for (pfn = pfn_first(map); pfn < pfn_end(map); pfn = pfn_next(pfn))
81 static void dev_pagemap_kill(struct dev_pagemap *pgmap)
83 if (pgmap->ops && pgmap->ops->kill)
84 pgmap->ops->kill(pgmap);
86 percpu_ref_kill(pgmap->ref);
89 static void dev_pagemap_cleanup(struct dev_pagemap *pgmap)
91 if (pgmap->ops && pgmap->ops->cleanup) {
92 pgmap->ops->cleanup(pgmap);
94 wait_for_completion(&pgmap->done);
95 percpu_ref_exit(pgmap->ref);
98 * Undo the pgmap ref assignment for the internal case as the
99 * caller may re-enable the same pgmap.
101 if (pgmap->ref == &pgmap->internal_ref)
105 void memunmap_pages(struct dev_pagemap *pgmap)
107 struct resource *res = &pgmap->res;
111 dev_pagemap_kill(pgmap);
112 for_each_device_pfn(pfn, pgmap)
113 put_page(pfn_to_page(pfn));
114 dev_pagemap_cleanup(pgmap);
116 /* pages are dead and unused, undo the arch mapping */
117 nid = page_to_nid(pfn_to_page(PHYS_PFN(res->start)));
120 if (pgmap->type == MEMORY_DEVICE_PRIVATE) {
121 pfn = PHYS_PFN(res->start);
122 __remove_pages(page_zone(pfn_to_page(pfn)), pfn,
123 PHYS_PFN(resource_size(res)), NULL);
125 arch_remove_memory(nid, res->start, resource_size(res),
126 pgmap_altmap(pgmap));
127 kasan_remove_zero_shadow(__va(res->start), resource_size(res));
131 untrack_pfn(NULL, PHYS_PFN(res->start), resource_size(res));
132 pgmap_array_delete(res);
133 WARN_ONCE(pgmap->altmap.alloc, "failed to free all reserved pages\n");
134 devmap_managed_enable_put();
136 EXPORT_SYMBOL_GPL(memunmap_pages);
138 static void devm_memremap_pages_release(void *data)
140 memunmap_pages(data);
143 static void dev_pagemap_percpu_release(struct percpu_ref *ref)
145 struct dev_pagemap *pgmap =
146 container_of(ref, struct dev_pagemap, internal_ref);
148 complete(&pgmap->done);
152 * Not device managed version of dev_memremap_pages, undone by
153 * memunmap_pages(). Please use dev_memremap_pages if you have a struct
156 void *memremap_pages(struct dev_pagemap *pgmap, int nid)
158 struct resource *res = &pgmap->res;
159 struct dev_pagemap *conflict_pgmap;
160 struct mhp_restrictions restrictions = {
162 * We do not want any optional features only our own memmap
164 .altmap = pgmap_altmap(pgmap),
166 pgprot_t pgprot = PAGE_KERNEL;
168 bool need_devmap_managed = true;
170 switch (pgmap->type) {
171 case MEMORY_DEVICE_PRIVATE:
172 if (!IS_ENABLED(CONFIG_DEVICE_PRIVATE)) {
173 WARN(1, "Device private memory not supported\n");
174 return ERR_PTR(-EINVAL);
176 if (!pgmap->ops || !pgmap->ops->migrate_to_ram) {
177 WARN(1, "Missing migrate_to_ram method\n");
178 return ERR_PTR(-EINVAL);
181 case MEMORY_DEVICE_FS_DAX:
182 if (!IS_ENABLED(CONFIG_ZONE_DEVICE) ||
183 IS_ENABLED(CONFIG_FS_DAX_LIMITED)) {
184 WARN(1, "File system DAX not supported\n");
185 return ERR_PTR(-EINVAL);
188 case MEMORY_DEVICE_DEVDAX:
189 case MEMORY_DEVICE_PCI_P2PDMA:
190 need_devmap_managed = false;
193 WARN(1, "Invalid pgmap type %d\n", pgmap->type);
198 if (pgmap->ops && (pgmap->ops->kill || pgmap->ops->cleanup))
199 return ERR_PTR(-EINVAL);
201 init_completion(&pgmap->done);
202 error = percpu_ref_init(&pgmap->internal_ref,
203 dev_pagemap_percpu_release, 0, GFP_KERNEL);
205 return ERR_PTR(error);
206 pgmap->ref = &pgmap->internal_ref;
208 if (!pgmap->ops || !pgmap->ops->kill || !pgmap->ops->cleanup) {
209 WARN(1, "Missing reference count teardown definition\n");
210 return ERR_PTR(-EINVAL);
214 if (need_devmap_managed) {
215 error = devmap_managed_enable_get(pgmap);
217 return ERR_PTR(error);
220 conflict_pgmap = get_dev_pagemap(PHYS_PFN(res->start), NULL);
221 if (conflict_pgmap) {
222 WARN(1, "Conflicting mapping in same section\n");
223 put_dev_pagemap(conflict_pgmap);
228 conflict_pgmap = get_dev_pagemap(PHYS_PFN(res->end), NULL);
229 if (conflict_pgmap) {
230 WARN(1, "Conflicting mapping in same section\n");
231 put_dev_pagemap(conflict_pgmap);
236 is_ram = region_intersects(res->start, resource_size(res),
237 IORESOURCE_SYSTEM_RAM, IORES_DESC_NONE);
239 if (is_ram != REGION_DISJOINT) {
240 WARN_ONCE(1, "%s attempted on %s region %pr\n", __func__,
241 is_ram == REGION_MIXED ? "mixed" : "ram", res);
246 error = xa_err(xa_store_range(&pgmap_array, PHYS_PFN(res->start),
247 PHYS_PFN(res->end), pgmap, GFP_KERNEL));
254 error = track_pfn_remap(NULL, &pgprot, PHYS_PFN(res->start), 0,
262 * For device private memory we call add_pages() as we only need to
263 * allocate and initialize struct page for the device memory. More-
264 * over the device memory is un-accessible thus we do not want to
265 * create a linear mapping for the memory like arch_add_memory()
268 * For all other device memory types, which are accessible by
269 * the CPU, we do want the linear mapping and thus use
272 if (pgmap->type == MEMORY_DEVICE_PRIVATE) {
273 error = add_pages(nid, PHYS_PFN(res->start),
274 PHYS_PFN(resource_size(res)), &restrictions);
276 error = kasan_add_zero_shadow(__va(res->start), resource_size(res));
282 error = arch_add_memory(nid, res->start, resource_size(res),
289 zone = &NODE_DATA(nid)->node_zones[ZONE_DEVICE];
290 move_pfn_range_to_zone(zone, PHYS_PFN(res->start),
291 PHYS_PFN(resource_size(res)), restrictions.altmap);
299 * Initialization of the pages has been deferred until now in order
300 * to allow us to do the work while not holding the hotplug lock.
302 memmap_init_zone_device(&NODE_DATA(nid)->node_zones[ZONE_DEVICE],
303 PHYS_PFN(res->start),
304 PHYS_PFN(resource_size(res)), pgmap);
305 percpu_ref_get_many(pgmap->ref, pfn_end(pgmap) - pfn_first(pgmap));
306 return __va(res->start);
309 kasan_remove_zero_shadow(__va(res->start), resource_size(res));
311 untrack_pfn(NULL, PHYS_PFN(res->start), resource_size(res));
313 pgmap_array_delete(res);
315 dev_pagemap_kill(pgmap);
316 dev_pagemap_cleanup(pgmap);
317 devmap_managed_enable_put();
318 return ERR_PTR(error);
320 EXPORT_SYMBOL_GPL(memremap_pages);
323 * devm_memremap_pages - remap and provide memmap backing for the given resource
324 * @dev: hosting device for @res
325 * @pgmap: pointer to a struct dev_pagemap
328 * 1/ At a minimum the res and type members of @pgmap must be initialized
329 * by the caller before passing it to this function
331 * 2/ The altmap field may optionally be initialized, in which case
332 * PGMAP_ALTMAP_VALID must be set in pgmap->flags.
334 * 3/ The ref field may optionally be provided, in which pgmap->ref must be
335 * 'live' on entry and will be killed and reaped at
336 * devm_memremap_pages_release() time, or if this routine fails.
338 * 4/ res is expected to be a host memory range that could feasibly be
339 * treated as a "System RAM" range, i.e. not a device mmio range, but
340 * this is not enforced.
342 void *devm_memremap_pages(struct device *dev, struct dev_pagemap *pgmap)
347 ret = memremap_pages(pgmap, dev_to_node(dev));
351 error = devm_add_action_or_reset(dev, devm_memremap_pages_release,
354 return ERR_PTR(error);
357 EXPORT_SYMBOL_GPL(devm_memremap_pages);
359 void devm_memunmap_pages(struct device *dev, struct dev_pagemap *pgmap)
361 devm_release_action(dev, devm_memremap_pages_release, pgmap);
363 EXPORT_SYMBOL_GPL(devm_memunmap_pages);
365 unsigned long vmem_altmap_offset(struct vmem_altmap *altmap)
367 /* number of pfns from base where pfn_to_page() is valid */
369 return altmap->reserve + altmap->free;
373 void vmem_altmap_free(struct vmem_altmap *altmap, unsigned long nr_pfns)
375 altmap->alloc -= nr_pfns;
379 * get_dev_pagemap() - take a new live reference on the dev_pagemap for @pfn
380 * @pfn: page frame number to lookup page_map
381 * @pgmap: optional known pgmap that already has a reference
383 * If @pgmap is non-NULL and covers @pfn it will be returned as-is. If @pgmap
384 * is non-NULL but does not cover @pfn the reference to it will be released.
386 struct dev_pagemap *get_dev_pagemap(unsigned long pfn,
387 struct dev_pagemap *pgmap)
389 resource_size_t phys = PFN_PHYS(pfn);
392 * In the cached case we're already holding a live reference.
395 if (phys >= pgmap->res.start && phys <= pgmap->res.end)
397 put_dev_pagemap(pgmap);
400 /* fall back to slow path lookup */
402 pgmap = xa_load(&pgmap_array, PHYS_PFN(phys));
403 if (pgmap && !percpu_ref_tryget_live(pgmap->ref))
409 EXPORT_SYMBOL_GPL(get_dev_pagemap);
411 #ifdef CONFIG_DEV_PAGEMAP_OPS
412 void __put_devmap_managed_page(struct page *page)
414 int count = page_ref_dec_return(page);
417 * If refcount is 1 then page is freed and refcount is stable as nobody
418 * holds a reference on the page.
421 /* Clear Active bit in case of parallel mark_page_accessed */
422 __ClearPageActive(page);
423 __ClearPageWaiters(page);
425 mem_cgroup_uncharge(page);
428 * When a device_private page is freed, the page->mapping field
429 * may still contain a (stale) mapping value. For example, the
430 * lower bits of page->mapping may still identify the page as
431 * an anonymous page. Ultimately, this entire field is just
432 * stale and wrong, and it will cause errors if not cleared.
435 * migrate_vma_pages()
436 * migrate_vma_insert_page()
437 * page_add_new_anon_rmap()
438 * __page_set_anon_rmap()
439 * ...checks page->mapping, via PageAnon(page) call,
440 * and incorrectly concludes that the page is an
441 * anonymous page. Therefore, it incorrectly,
442 * silently fails to set up the new anon rmap.
444 * For other types of ZONE_DEVICE pages, migration is either
445 * handled differently or not done at all, so there is no need
446 * to clear page->mapping.
448 if (is_device_private_page(page))
449 page->mapping = NULL;
451 page->pgmap->ops->page_free(page);
455 EXPORT_SYMBOL(__put_devmap_managed_page);
456 #endif /* CONFIG_DEV_PAGEMAP_OPS */