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);
17 #ifdef CONFIG_DEV_PAGEMAP_OPS
18 DEFINE_STATIC_KEY_FALSE(devmap_managed_key);
19 EXPORT_SYMBOL(devmap_managed_key);
20 static atomic_t devmap_managed_enable;
22 static void devmap_managed_enable_put(void)
24 if (atomic_dec_and_test(&devmap_managed_enable))
25 static_branch_disable(&devmap_managed_key);
28 static int devmap_managed_enable_get(struct dev_pagemap *pgmap)
30 if (!pgmap->ops || !pgmap->ops->page_free) {
31 WARN(1, "Missing page_free method\n");
35 if (atomic_inc_return(&devmap_managed_enable) == 1)
36 static_branch_enable(&devmap_managed_key);
40 static int devmap_managed_enable_get(struct dev_pagemap *pgmap)
44 static void devmap_managed_enable_put(void)
47 #endif /* CONFIG_DEV_PAGEMAP_OPS */
49 static void pgmap_array_delete(struct resource *res)
51 xa_store_range(&pgmap_array, PHYS_PFN(res->start), PHYS_PFN(res->end),
56 static unsigned long pfn_first(struct dev_pagemap *pgmap)
58 return PHYS_PFN(pgmap->res.start) +
59 vmem_altmap_offset(pgmap_altmap(pgmap));
62 static unsigned long pfn_end(struct dev_pagemap *pgmap)
64 const struct resource *res = &pgmap->res;
66 return (res->start + resource_size(res)) >> PAGE_SHIFT;
69 static unsigned long pfn_next(unsigned long pfn)
76 #define for_each_device_pfn(pfn, map) \
77 for (pfn = pfn_first(map); pfn < pfn_end(map); pfn = pfn_next(pfn))
79 static void dev_pagemap_kill(struct dev_pagemap *pgmap)
81 if (pgmap->ops && pgmap->ops->kill)
82 pgmap->ops->kill(pgmap);
84 percpu_ref_kill(pgmap->ref);
87 static void dev_pagemap_cleanup(struct dev_pagemap *pgmap)
89 if (pgmap->ops && pgmap->ops->cleanup) {
90 pgmap->ops->cleanup(pgmap);
92 wait_for_completion(&pgmap->done);
93 percpu_ref_exit(pgmap->ref);
96 * Undo the pgmap ref assignment for the internal case as the
97 * caller may re-enable the same pgmap.
99 if (pgmap->ref == &pgmap->internal_ref)
103 void memunmap_pages(struct dev_pagemap *pgmap)
105 struct resource *res = &pgmap->res;
109 dev_pagemap_kill(pgmap);
110 for_each_device_pfn(pfn, pgmap)
111 put_page(pfn_to_page(pfn));
112 dev_pagemap_cleanup(pgmap);
114 /* pages are dead and unused, undo the arch mapping */
115 nid = page_to_nid(pfn_to_page(PHYS_PFN(res->start)));
118 if (pgmap->type == MEMORY_DEVICE_PRIVATE) {
119 pfn = PHYS_PFN(res->start);
120 __remove_pages(page_zone(pfn_to_page(pfn)), pfn,
121 PHYS_PFN(resource_size(res)), NULL);
123 arch_remove_memory(nid, res->start, resource_size(res),
124 pgmap_altmap(pgmap));
125 kasan_remove_zero_shadow(__va(res->start), resource_size(res));
129 untrack_pfn(NULL, PHYS_PFN(res->start), resource_size(res));
130 pgmap_array_delete(res);
131 WARN_ONCE(pgmap->altmap.alloc, "failed to free all reserved pages\n");
132 devmap_managed_enable_put();
134 EXPORT_SYMBOL_GPL(memunmap_pages);
136 static void devm_memremap_pages_release(void *data)
138 memunmap_pages(data);
141 static void dev_pagemap_percpu_release(struct percpu_ref *ref)
143 struct dev_pagemap *pgmap =
144 container_of(ref, struct dev_pagemap, internal_ref);
146 complete(&pgmap->done);
150 * Not device managed version of dev_memremap_pages, undone by
151 * memunmap_pages(). Please use dev_memremap_pages if you have a struct
154 void *memremap_pages(struct dev_pagemap *pgmap, int nid)
156 struct resource *res = &pgmap->res;
157 struct dev_pagemap *conflict_pgmap;
158 struct mhp_restrictions restrictions = {
160 * We do not want any optional features only our own memmap
162 .altmap = pgmap_altmap(pgmap),
164 pgprot_t pgprot = PAGE_KERNEL;
166 bool need_devmap_managed = true;
168 switch (pgmap->type) {
169 case MEMORY_DEVICE_PRIVATE:
170 if (!IS_ENABLED(CONFIG_DEVICE_PRIVATE)) {
171 WARN(1, "Device private memory not supported\n");
172 return ERR_PTR(-EINVAL);
174 if (!pgmap->ops || !pgmap->ops->migrate_to_ram) {
175 WARN(1, "Missing migrate_to_ram method\n");
176 return ERR_PTR(-EINVAL);
179 case MEMORY_DEVICE_FS_DAX:
180 if (!IS_ENABLED(CONFIG_ZONE_DEVICE) ||
181 IS_ENABLED(CONFIG_FS_DAX_LIMITED)) {
182 WARN(1, "File system DAX not supported\n");
183 return ERR_PTR(-EINVAL);
186 case MEMORY_DEVICE_DEVDAX:
187 case MEMORY_DEVICE_PCI_P2PDMA:
188 need_devmap_managed = false;
191 WARN(1, "Invalid pgmap type %d\n", pgmap->type);
196 if (pgmap->ops && (pgmap->ops->kill || pgmap->ops->cleanup))
197 return ERR_PTR(-EINVAL);
199 init_completion(&pgmap->done);
200 error = percpu_ref_init(&pgmap->internal_ref,
201 dev_pagemap_percpu_release, 0, GFP_KERNEL);
203 return ERR_PTR(error);
204 pgmap->ref = &pgmap->internal_ref;
206 if (!pgmap->ops || !pgmap->ops->kill || !pgmap->ops->cleanup) {
207 WARN(1, "Missing reference count teardown definition\n");
208 return ERR_PTR(-EINVAL);
212 if (need_devmap_managed) {
213 error = devmap_managed_enable_get(pgmap);
215 return ERR_PTR(error);
218 conflict_pgmap = get_dev_pagemap(PHYS_PFN(res->start), NULL);
219 if (conflict_pgmap) {
220 WARN(1, "Conflicting mapping in same section\n");
221 put_dev_pagemap(conflict_pgmap);
226 conflict_pgmap = get_dev_pagemap(PHYS_PFN(res->end), NULL);
227 if (conflict_pgmap) {
228 WARN(1, "Conflicting mapping in same section\n");
229 put_dev_pagemap(conflict_pgmap);
234 is_ram = region_intersects(res->start, resource_size(res),
235 IORESOURCE_SYSTEM_RAM, IORES_DESC_NONE);
237 if (is_ram != REGION_DISJOINT) {
238 WARN_ONCE(1, "%s attempted on %s region %pr\n", __func__,
239 is_ram == REGION_MIXED ? "mixed" : "ram", res);
244 error = xa_err(xa_store_range(&pgmap_array, PHYS_PFN(res->start),
245 PHYS_PFN(res->end), pgmap, GFP_KERNEL));
252 error = track_pfn_remap(NULL, &pgprot, PHYS_PFN(res->start), 0,
260 * For device private memory we call add_pages() as we only need to
261 * allocate and initialize struct page for the device memory. More-
262 * over the device memory is un-accessible thus we do not want to
263 * create a linear mapping for the memory like arch_add_memory()
266 * For all other device memory types, which are accessible by
267 * the CPU, we do want the linear mapping and thus use
270 if (pgmap->type == MEMORY_DEVICE_PRIVATE) {
271 error = add_pages(nid, PHYS_PFN(res->start),
272 PHYS_PFN(resource_size(res)), &restrictions);
274 error = kasan_add_zero_shadow(__va(res->start), resource_size(res));
280 error = arch_add_memory(nid, res->start, resource_size(res),
287 zone = &NODE_DATA(nid)->node_zones[ZONE_DEVICE];
288 move_pfn_range_to_zone(zone, PHYS_PFN(res->start),
289 PHYS_PFN(resource_size(res)), restrictions.altmap);
297 * Initialization of the pages has been deferred until now in order
298 * to allow us to do the work while not holding the hotplug lock.
300 memmap_init_zone_device(&NODE_DATA(nid)->node_zones[ZONE_DEVICE],
301 PHYS_PFN(res->start),
302 PHYS_PFN(resource_size(res)), pgmap);
303 percpu_ref_get_many(pgmap->ref, pfn_end(pgmap) - pfn_first(pgmap));
304 return __va(res->start);
307 kasan_remove_zero_shadow(__va(res->start), resource_size(res));
309 untrack_pfn(NULL, PHYS_PFN(res->start), resource_size(res));
311 pgmap_array_delete(res);
313 dev_pagemap_kill(pgmap);
314 dev_pagemap_cleanup(pgmap);
315 devmap_managed_enable_put();
316 return ERR_PTR(error);
318 EXPORT_SYMBOL_GPL(memremap_pages);
321 * devm_memremap_pages - remap and provide memmap backing for the given resource
322 * @dev: hosting device for @res
323 * @pgmap: pointer to a struct dev_pagemap
326 * 1/ At a minimum the res and type members of @pgmap must be initialized
327 * by the caller before passing it to this function
329 * 2/ The altmap field may optionally be initialized, in which case
330 * PGMAP_ALTMAP_VALID must be set in pgmap->flags.
332 * 3/ The ref field may optionally be provided, in which pgmap->ref must be
333 * 'live' on entry and will be killed and reaped at
334 * devm_memremap_pages_release() time, or if this routine fails.
336 * 4/ res is expected to be a host memory range that could feasibly be
337 * treated as a "System RAM" range, i.e. not a device mmio range, but
338 * this is not enforced.
340 void *devm_memremap_pages(struct device *dev, struct dev_pagemap *pgmap)
345 ret = memremap_pages(pgmap, dev_to_node(dev));
349 error = devm_add_action_or_reset(dev, devm_memremap_pages_release,
352 return ERR_PTR(error);
355 EXPORT_SYMBOL_GPL(devm_memremap_pages);
357 void devm_memunmap_pages(struct device *dev, struct dev_pagemap *pgmap)
359 devm_release_action(dev, devm_memremap_pages_release, pgmap);
361 EXPORT_SYMBOL_GPL(devm_memunmap_pages);
363 unsigned long vmem_altmap_offset(struct vmem_altmap *altmap)
365 /* number of pfns from base where pfn_to_page() is valid */
367 return altmap->reserve + altmap->free;
371 void vmem_altmap_free(struct vmem_altmap *altmap, unsigned long nr_pfns)
373 altmap->alloc -= nr_pfns;
377 * get_dev_pagemap() - take a new live reference on the dev_pagemap for @pfn
378 * @pfn: page frame number to lookup page_map
379 * @pgmap: optional known pgmap that already has a reference
381 * If @pgmap is non-NULL and covers @pfn it will be returned as-is. If @pgmap
382 * is non-NULL but does not cover @pfn the reference to it will be released.
384 struct dev_pagemap *get_dev_pagemap(unsigned long pfn,
385 struct dev_pagemap *pgmap)
387 resource_size_t phys = PFN_PHYS(pfn);
390 * In the cached case we're already holding a live reference.
393 if (phys >= pgmap->res.start && phys <= pgmap->res.end)
395 put_dev_pagemap(pgmap);
398 /* fall back to slow path lookup */
400 pgmap = xa_load(&pgmap_array, PHYS_PFN(phys));
401 if (pgmap && !percpu_ref_tryget_live(pgmap->ref))
407 EXPORT_SYMBOL_GPL(get_dev_pagemap);
409 #ifdef CONFIG_DEV_PAGEMAP_OPS
410 void __put_devmap_managed_page(struct page *page)
412 int count = page_ref_dec_return(page);
415 * If refcount is 1 then page is freed and refcount is stable as nobody
416 * holds a reference on the page.
419 /* Clear Active bit in case of parallel mark_page_accessed */
420 __ClearPageActive(page);
421 __ClearPageWaiters(page);
423 mem_cgroup_uncharge(page);
426 * When a device_private page is freed, the page->mapping field
427 * may still contain a (stale) mapping value. For example, the
428 * lower bits of page->mapping may still identify the page as
429 * an anonymous page. Ultimately, this entire field is just
430 * stale and wrong, and it will cause errors if not cleared.
433 * migrate_vma_pages()
434 * migrate_vma_insert_page()
435 * page_add_new_anon_rmap()
436 * __page_set_anon_rmap()
437 * ...checks page->mapping, via PageAnon(page) call,
438 * and incorrectly concludes that the page is an
439 * anonymous page. Therefore, it incorrectly,
440 * silently fails to set up the new anon rmap.
442 * For other types of ZONE_DEVICE pages, migration is either
443 * handled differently or not done at all, so there is no need
444 * to clear page->mapping.
446 if (is_device_private_page(page))
447 page->mapping = NULL;
449 page->pgmap->ops->page_free(page);
453 EXPORT_SYMBOL(__put_devmap_managed_page);
454 #endif /* CONFIG_DEV_PAGEMAP_OPS */