2 * Copyright(c) 2015 Intel Corporation. All rights reserved.
4 * This program is free software; you can redistribute it and/or modify
5 * it under the terms of version 2 of the GNU General Public License as
6 * published by the Free Software Foundation.
8 * This program is distributed in the hope that it will be useful, but
9 * WITHOUT ANY WARRANTY; without even the implied warranty of
10 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
11 * General Public License for more details.
13 #include <linux/radix-tree.h>
14 #include <linux/memremap.h>
15 #include <linux/device.h>
16 #include <linux/types.h>
17 #include <linux/pfn_t.h>
20 #include <linux/memory_hotplug.h>
23 /* temporary while we convert existing ioremap_cache users to memremap */
24 __weak void __iomem *ioremap_cache(resource_size_t offset, unsigned long size)
26 return ioremap(offset, size);
30 static void *try_ram_remap(resource_size_t offset, size_t size)
32 struct page *page = pfn_to_page(offset >> PAGE_SHIFT);
34 /* In the simple case just return the existing linear address */
35 if (!PageHighMem(page))
37 return NULL; /* fallback to ioremap_cache */
41 * memremap() - remap an iomem_resource as cacheable memory
42 * @offset: iomem resource start address
43 * @size: size of remap
44 * @flags: either MEMREMAP_WB or MEMREMAP_WT
46 * memremap() is "ioremap" for cases where it is known that the resource
47 * being mapped does not have i/o side effects and the __iomem
48 * annotation is not applicable.
50 * MEMREMAP_WB - matches the default mapping for System RAM on
51 * the architecture. This is usually a read-allocate write-back cache.
52 * Morever, if MEMREMAP_WB is specified and the requested remap region is RAM
53 * memremap() will bypass establishing a new mapping and instead return
54 * a pointer into the direct map.
56 * MEMREMAP_WT - establish a mapping whereby writes either bypass the
57 * cache or are written through to memory and never exist in a
58 * cache-dirty state with respect to program visibility. Attempts to
59 * map System RAM with this mapping type will fail.
61 void *memremap(resource_size_t offset, size_t size, unsigned long flags)
63 int is_ram = region_intersects(offset, size,
64 IORESOURCE_SYSTEM_RAM, IORES_DESC_NONE);
67 if (is_ram == REGION_MIXED) {
68 WARN_ONCE(1, "memremap attempted on mixed range %pa size: %#lx\n",
69 &offset, (unsigned long) size);
73 /* Try all mapping types requested until one returns non-NULL */
74 if (flags & MEMREMAP_WB) {
75 flags &= ~MEMREMAP_WB;
77 * MEMREMAP_WB is special in that it can be satisifed
78 * from the direct map. Some archs depend on the
79 * capability of memremap() to autodetect cases where
80 * the requested range is potentially in System RAM.
82 if (is_ram == REGION_INTERSECTS)
83 addr = try_ram_remap(offset, size);
85 addr = ioremap_cache(offset, size);
89 * If we don't have a mapping yet and more request flags are
90 * pending then we will be attempting to establish a new virtual
91 * address mapping. Enforce that this mapping is not aliasing
94 if (!addr && is_ram == REGION_INTERSECTS && flags) {
95 WARN_ONCE(1, "memremap attempted on ram %pa size: %#lx\n",
96 &offset, (unsigned long) size);
100 if (!addr && (flags & MEMREMAP_WT)) {
101 flags &= ~MEMREMAP_WT;
102 addr = ioremap_wt(offset, size);
107 EXPORT_SYMBOL(memremap);
109 void memunmap(void *addr)
111 if (is_vmalloc_addr(addr))
112 iounmap((void __iomem *) addr);
114 EXPORT_SYMBOL(memunmap);
116 static void devm_memremap_release(struct device *dev, void *res)
121 static int devm_memremap_match(struct device *dev, void *res, void *match_data)
123 return *(void **)res == match_data;
126 void *devm_memremap(struct device *dev, resource_size_t offset,
127 size_t size, unsigned long flags)
131 ptr = devres_alloc_node(devm_memremap_release, sizeof(*ptr), GFP_KERNEL,
134 return ERR_PTR(-ENOMEM);
136 addr = memremap(offset, size, flags);
139 devres_add(dev, ptr);
145 EXPORT_SYMBOL(devm_memremap);
147 void devm_memunmap(struct device *dev, void *addr)
149 WARN_ON(devres_release(dev, devm_memremap_release,
150 devm_memremap_match, addr));
152 EXPORT_SYMBOL(devm_memunmap);
154 pfn_t phys_to_pfn_t(phys_addr_t addr, unsigned long flags)
156 return __pfn_to_pfn_t(addr >> PAGE_SHIFT, flags);
158 EXPORT_SYMBOL(phys_to_pfn_t);
160 #ifdef CONFIG_ZONE_DEVICE
161 static DEFINE_MUTEX(pgmap_lock);
162 static RADIX_TREE(pgmap_radix, GFP_KERNEL);
163 #define SECTION_MASK ~((1UL << PA_SECTION_SHIFT) - 1)
164 #define SECTION_SIZE (1UL << PA_SECTION_SHIFT)
168 struct percpu_ref *ref;
169 struct dev_pagemap pgmap;
170 struct vmem_altmap altmap;
173 void get_zone_device_page(struct page *page)
175 percpu_ref_get(page->pgmap->ref);
177 EXPORT_SYMBOL(get_zone_device_page);
179 void put_zone_device_page(struct page *page)
181 put_dev_pagemap(page->pgmap);
183 EXPORT_SYMBOL(put_zone_device_page);
185 static void pgmap_radix_release(struct resource *res)
187 resource_size_t key, align_start, align_size, align_end;
189 align_start = res->start & ~(SECTION_SIZE - 1);
190 align_size = ALIGN(resource_size(res), SECTION_SIZE);
191 align_end = align_start + align_size - 1;
193 mutex_lock(&pgmap_lock);
194 for (key = res->start; key <= res->end; key += SECTION_SIZE)
195 radix_tree_delete(&pgmap_radix, key >> PA_SECTION_SHIFT);
196 mutex_unlock(&pgmap_lock);
199 static unsigned long pfn_first(struct page_map *page_map)
201 struct dev_pagemap *pgmap = &page_map->pgmap;
202 const struct resource *res = &page_map->res;
203 struct vmem_altmap *altmap = pgmap->altmap;
206 pfn = res->start >> PAGE_SHIFT;
208 pfn += vmem_altmap_offset(altmap);
212 static unsigned long pfn_end(struct page_map *page_map)
214 const struct resource *res = &page_map->res;
216 return (res->start + resource_size(res)) >> PAGE_SHIFT;
219 #define for_each_device_pfn(pfn, map) \
220 for (pfn = pfn_first(map); pfn < pfn_end(map); pfn++)
222 static void devm_memremap_pages_release(struct device *dev, void *data)
224 struct page_map *page_map = data;
225 struct resource *res = &page_map->res;
226 resource_size_t align_start, align_size;
227 struct dev_pagemap *pgmap = &page_map->pgmap;
229 if (percpu_ref_tryget_live(pgmap->ref)) {
230 dev_WARN(dev, "%s: page mapping is still live!\n", __func__);
231 percpu_ref_put(pgmap->ref);
234 /* pages are dead and unused, undo the arch mapping */
235 align_start = res->start & ~(SECTION_SIZE - 1);
236 align_size = ALIGN(resource_size(res), SECTION_SIZE);
237 arch_remove_memory(align_start, align_size);
238 pgmap_radix_release(res);
239 dev_WARN_ONCE(dev, pgmap->altmap && pgmap->altmap->alloc,
240 "%s: failed to free all reserved pages\n", __func__);
243 /* assumes rcu_read_lock() held at entry */
244 struct dev_pagemap *find_dev_pagemap(resource_size_t phys)
246 struct page_map *page_map;
248 WARN_ON_ONCE(!rcu_read_lock_held());
250 page_map = radix_tree_lookup(&pgmap_radix, phys >> PA_SECTION_SHIFT);
251 return page_map ? &page_map->pgmap : NULL;
255 * devm_memremap_pages - remap and provide memmap backing for the given resource
256 * @dev: hosting device for @res
257 * @res: "host memory" address range
258 * @ref: a live per-cpu reference count
259 * @altmap: optional descriptor for allocating the memmap from @res
262 * 1/ @ref must be 'live' on entry and 'dead' before devm_memunmap_pages() time
263 * (or devm release event).
265 * 2/ @res is expected to be a host memory range that could feasibly be
266 * treated as a "System RAM" range, i.e. not a device mmio range, but
267 * this is not enforced.
269 void *devm_memremap_pages(struct device *dev, struct resource *res,
270 struct percpu_ref *ref, struct vmem_altmap *altmap)
272 int is_ram = region_intersects(res->start, resource_size(res),
273 IORESOURCE_SYSTEM_RAM, IORES_DESC_NONE);
274 resource_size_t key, align_start, align_size, align_end;
275 struct dev_pagemap *pgmap;
276 struct page_map *page_map;
280 if (is_ram == REGION_MIXED) {
281 WARN_ONCE(1, "%s attempted on mixed region %pr\n",
283 return ERR_PTR(-ENXIO);
286 if (is_ram == REGION_INTERSECTS)
287 return __va(res->start);
289 if (altmap && !IS_ENABLED(CONFIG_SPARSEMEM_VMEMMAP)) {
290 dev_err(dev, "%s: altmap requires CONFIG_SPARSEMEM_VMEMMAP=y\n",
292 return ERR_PTR(-ENXIO);
296 return ERR_PTR(-EINVAL);
298 page_map = devres_alloc_node(devm_memremap_pages_release,
299 sizeof(*page_map), GFP_KERNEL, dev_to_node(dev));
301 return ERR_PTR(-ENOMEM);
302 pgmap = &page_map->pgmap;
304 memcpy(&page_map->res, res, sizeof(*res));
308 memcpy(&page_map->altmap, altmap, sizeof(*altmap));
309 pgmap->altmap = &page_map->altmap;
312 pgmap->res = &page_map->res;
314 mutex_lock(&pgmap_lock);
316 align_start = res->start & ~(SECTION_SIZE - 1);
317 align_size = ALIGN(resource_size(res), SECTION_SIZE);
318 align_end = align_start + align_size - 1;
319 for (key = align_start; key <= align_end; key += SECTION_SIZE) {
320 struct dev_pagemap *dup;
323 dup = find_dev_pagemap(key);
326 dev_err(dev, "%s: %pr collides with mapping for %s\n",
327 __func__, res, dev_name(dup->dev));
331 error = radix_tree_insert(&pgmap_radix, key >> PA_SECTION_SHIFT,
334 dev_err(dev, "%s: failed: %d\n", __func__, error);
338 mutex_unlock(&pgmap_lock);
342 nid = dev_to_node(dev);
346 error = arch_add_memory(nid, align_start, align_size, true);
350 for_each_device_pfn(pfn, page_map) {
351 struct page *page = pfn_to_page(pfn);
353 /* ZONE_DEVICE pages must never appear on a slab lru */
354 list_force_poison(&page->lru);
357 devres_add(dev, page_map);
358 return __va(res->start);
362 pgmap_radix_release(res);
363 devres_free(page_map);
364 return ERR_PTR(error);
366 EXPORT_SYMBOL(devm_memremap_pages);
368 unsigned long vmem_altmap_offset(struct vmem_altmap *altmap)
370 /* number of pfns from base where pfn_to_page() is valid */
371 return altmap->reserve + altmap->free;
374 void vmem_altmap_free(struct vmem_altmap *altmap, unsigned long nr_pfns)
376 altmap->alloc -= nr_pfns;
379 #ifdef CONFIG_SPARSEMEM_VMEMMAP
380 struct vmem_altmap *to_vmem_altmap(unsigned long memmap_start)
383 * 'memmap_start' is the virtual address for the first "struct
384 * page" in this range of the vmemmap array. In the case of
385 * CONFIG_SPARSE_VMEMMAP a page_to_pfn conversion is simple
386 * pointer arithmetic, so we can perform this to_vmem_altmap()
387 * conversion without concern for the initialization state of
388 * the struct page fields.
390 struct page *page = (struct page *) memmap_start;
391 struct dev_pagemap *pgmap;
394 * Uncoditionally retrieve a dev_pagemap associated with the
395 * given physical address, this is only for use in the
396 * arch_{add|remove}_memory() for setting up and tearing down
400 pgmap = find_dev_pagemap(__pfn_to_phys(page_to_pfn(page)));
403 return pgmap ? pgmap->altmap : NULL;
405 #endif /* CONFIG_SPARSEMEM_VMEMMAP */
406 #endif /* CONFIG_ZONE_DEVICE */