kernel/memremap.c

   1 /*
   2  * Copyright(c) 2015 Intel Corporation. All rights reserved.
   3  *
   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.
   7  *
   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.
  12  */
  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>
  18 #include <linux/io.h>
  19 #include <linux/mm.h>
  20 #include <linux/memory_hotplug.h>
  21
  22 #ifndef ioremap_cache
  23 /* temporary while we convert existing ioremap_cache users to memremap */
  24 __weak void __iomem *ioremap_cache(resource_size_t offset, unsigned long size)
  25 {
  26         return ioremap(offset, size);
  27 }
  28 #endif
  29
  30 static void *try_ram_remap(resource_size_t offset, size_t size)
  31 {
  32         struct page *page = pfn_to_page(offset >> PAGE_SHIFT);
  33
  34         /* In the simple case just return the existing linear address */
  35         if (!PageHighMem(page))
  36                 return __va(offset);
  37         return NULL; /* fallback to ioremap_cache */
  38 }
  39
  40 /**
  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
  45  *
  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.
  49  *
  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.
  55  *
  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.
  60  */
  61 void *memremap(resource_size_t offset, size_t size, unsigned long flags)
  62 {
  63         int is_ram = region_intersects(offset, size, "System RAM");
  64         void *addr = NULL;
  65
  66         if (is_ram == REGION_MIXED) {
  67                 WARN_ONCE(1, "memremap attempted on mixed range %pa size: %#lx\n",
  68                                 &offset, (unsigned long) size);
  69                 return NULL;
  70         }
  71
  72         /* Try all mapping types requested until one returns non-NULL */
  73         if (flags & MEMREMAP_WB) {
  74                 flags &= ~MEMREMAP_WB;
  75                 /*
  76                  * MEMREMAP_WB is special in that it can be satisifed
  77                  * from the direct map.  Some archs depend on the
  78                  * capability of memremap() to autodetect cases where
  79                  * the requested range is potentially in "System RAM"
  80                  */
  81                 if (is_ram == REGION_INTERSECTS)
  82                         addr = try_ram_remap(offset, size);
  83                 if (!addr)
  84                         addr = ioremap_cache(offset, size);
  85         }
  86
  87         /*
  88          * If we don't have a mapping yet and more request flags are
  89          * pending then we will be attempting to establish a new virtual
  90          * address mapping.  Enforce that this mapping is not aliasing
  91          * "System RAM"
  92          */
  93         if (!addr && is_ram == REGION_INTERSECTS && flags) {
  94                 WARN_ONCE(1, "memremap attempted on ram %pa size: %#lx\n",
  95                                 &offset, (unsigned long) size);
  96                 return NULL;
  97         }
  98
  99         if (!addr && (flags & MEMREMAP_WT)) {
 100                 flags &= ~MEMREMAP_WT;
 101                 addr = ioremap_wt(offset, size);
 102         }
 103
 104         return addr;
 105 }
 106 EXPORT_SYMBOL(memremap);
 107
 108 void memunmap(void *addr)
 109 {
 110         if (is_vmalloc_addr(addr))
 111                 iounmap((void __iomem *) addr);
 112 }
 113 EXPORT_SYMBOL(memunmap);
 114
 115 static void devm_memremap_release(struct device *dev, void *res)
 116 {
 117         memunmap(res);
 118 }
 119
 120 static int devm_memremap_match(struct device *dev, void *res, void *match_data)
 121 {
 122         return *(void **)res == match_data;
 123 }
 124
 125 void *devm_memremap(struct device *dev, resource_size_t offset,
 126                 size_t size, unsigned long flags)
 127 {
 128         void **ptr, *addr;
 129
 130         ptr = devres_alloc_node(devm_memremap_release, sizeof(*ptr), GFP_KERNEL,
 131                         dev_to_node(dev));
 132         if (!ptr)
 133                 return ERR_PTR(-ENOMEM);
 134
 135         addr = memremap(offset, size, flags);
 136         if (addr) {
 137                 *ptr = addr;
 138                 devres_add(dev, ptr);
 139         } else
 140                 devres_free(ptr);
 141
 142         return addr;
 143 }
 144 EXPORT_SYMBOL(devm_memremap);
 145
 146 void devm_memunmap(struct device *dev, void *addr)
 147 {
 148         WARN_ON(devres_release(dev, devm_memremap_release,
 149                                 devm_memremap_match, addr));
 150 }
 151 EXPORT_SYMBOL(devm_memunmap);
 152
 153 pfn_t phys_to_pfn_t(dma_addr_t addr, unsigned long flags)
 154 {
 155         return __pfn_to_pfn_t(addr >> PAGE_SHIFT, flags);
 156 }
 157 EXPORT_SYMBOL(phys_to_pfn_t);
 158
 159 #ifdef CONFIG_ZONE_DEVICE
 160 static DEFINE_MUTEX(pgmap_lock);
 161 static RADIX_TREE(pgmap_radix, GFP_KERNEL);
 162 #define SECTION_MASK ~((1UL << PA_SECTION_SHIFT) - 1)
 163 #define SECTION_SIZE (1UL << PA_SECTION_SHIFT)
 164
 165 struct page_map {
 166         struct resource res;
 167         struct percpu_ref *ref;
 168         struct dev_pagemap pgmap;
 169         struct vmem_altmap altmap;
 170 };
 171
 172 void get_zone_device_page(struct page *page)
 173 {
 174         percpu_ref_get(page->pgmap->ref);
 175 }
 176 EXPORT_SYMBOL(get_zone_device_page);
 177
 178 void put_zone_device_page(struct page *page)
 179 {
 180         put_dev_pagemap(page->pgmap);
 181 }
 182 EXPORT_SYMBOL(put_zone_device_page);
 183
 184 static void pgmap_radix_release(struct resource *res)
 185 {
 186         resource_size_t key;
 187
 188         mutex_lock(&pgmap_lock);
 189         for (key = res->start; key <= res->end; key += SECTION_SIZE)
 190                 radix_tree_delete(&pgmap_radix, key >> PA_SECTION_SHIFT);
 191         mutex_unlock(&pgmap_lock);
 192 }
 193
 194 static unsigned long pfn_first(struct page_map *page_map)
 195 {
 196         struct dev_pagemap *pgmap = &page_map->pgmap;
 197         const struct resource *res = &page_map->res;
 198         struct vmem_altmap *altmap = pgmap->altmap;
 199         unsigned long pfn;
 200
 201         pfn = res->start >> PAGE_SHIFT;
 202         if (altmap)
 203                 pfn += vmem_altmap_offset(altmap);
 204         return pfn;
 205 }
 206
 207 static unsigned long pfn_end(struct page_map *page_map)
 208 {
 209         const struct resource *res = &page_map->res;
 210
 211         return (res->start + resource_size(res)) >> PAGE_SHIFT;
 212 }
 213
 214 #define for_each_device_pfn(pfn, map) \
 215         for (pfn = pfn_first(map); pfn < pfn_end(map); pfn++)
 216
 217 static void devm_memremap_pages_release(struct device *dev, void *data)
 218 {
 219         struct page_map *page_map = data;
 220         struct resource *res = &page_map->res;
 221         resource_size_t align_start, align_size;
 222         struct dev_pagemap *pgmap = &page_map->pgmap;
 223
 224         if (percpu_ref_tryget_live(pgmap->ref)) {
 225                 dev_WARN(dev, "%s: page mapping is still live!\n", __func__);
 226                 percpu_ref_put(pgmap->ref);
 227         }
 228
 229         pgmap_radix_release(res);
 230
 231         /* pages are dead and unused, undo the arch mapping */
 232         align_start = res->start & ~(SECTION_SIZE - 1);
 233         align_size = ALIGN(resource_size(res), SECTION_SIZE);
 234         arch_remove_memory(align_start, align_size);
 235         dev_WARN_ONCE(dev, pgmap->altmap && pgmap->altmap->alloc,
 236                         "%s: failed to free all reserved pages\n", __func__);
 237 }
 238
 239 /* assumes rcu_read_lock() held at entry */
 240 struct dev_pagemap *find_dev_pagemap(resource_size_t phys)
 241 {
 242         struct page_map *page_map;
 243
 244         WARN_ON_ONCE(!rcu_read_lock_held());
 245
 246         page_map = radix_tree_lookup(&pgmap_radix, phys >> PA_SECTION_SHIFT);
 247         return page_map ? &page_map->pgmap : NULL;
 248 }
 249
 250 /**
 251  * devm_memremap_pages - remap and provide memmap backing for the given resource
 252  * @dev: hosting device for @res
 253  * @res: "host memory" address range
 254  * @ref: a live per-cpu reference count
 255  * @altmap: optional descriptor for allocating the memmap from @res
 256  *
 257  * Notes:
 258  * 1/ @ref must be 'live' on entry and 'dead' before devm_memunmap_pages() time
 259  *    (or devm release event).
 260  *
 261  * 2/ @res is expected to be a host memory range that could feasibly be
 262  *    treated as a "System RAM" range, i.e. not a device mmio range, but
 263  *    this is not enforced.
 264  */
 265 void *devm_memremap_pages(struct device *dev, struct resource *res,
 266                 struct percpu_ref *ref, struct vmem_altmap *altmap)
 267 {
 268         int is_ram = region_intersects(res->start, resource_size(res),
 269                         "System RAM");
 270         resource_size_t key, align_start, align_size;
 271         struct dev_pagemap *pgmap;
 272         struct page_map *page_map;
 273         unsigned long pfn;
 274         int error, nid;
 275
 276         if (is_ram == REGION_MIXED) {
 277                 WARN_ONCE(1, "%s attempted on mixed region %pr\n",
 278                                 __func__, res);
 279                 return ERR_PTR(-ENXIO);
 280         }
 281
 282         if (is_ram == REGION_INTERSECTS)
 283                 return __va(res->start);
 284
 285         if (altmap && !IS_ENABLED(CONFIG_SPARSEMEM_VMEMMAP)) {
 286                 dev_err(dev, "%s: altmap requires CONFIG_SPARSEMEM_VMEMMAP=y\n",
 287                                 __func__);
 288                 return ERR_PTR(-ENXIO);
 289         }
 290
 291         if (!ref)
 292                 return ERR_PTR(-EINVAL);
 293
 294         page_map = devres_alloc_node(devm_memremap_pages_release,
 295                         sizeof(*page_map), GFP_KERNEL, dev_to_node(dev));
 296         if (!page_map)
 297                 return ERR_PTR(-ENOMEM);
 298         pgmap = &page_map->pgmap;
 299
 300         memcpy(&page_map->res, res, sizeof(*res));
 301
 302         pgmap->dev = dev;
 303         if (altmap) {
 304                 memcpy(&page_map->altmap, altmap, sizeof(*altmap));
 305                 pgmap->altmap = &page_map->altmap;
 306         }
 307         pgmap->ref = ref;
 308         pgmap->res = &page_map->res;
 309
 310         mutex_lock(&pgmap_lock);
 311         error = 0;
 312         for (key = res->start; key <= res->end; key += SECTION_SIZE) {
 313                 struct dev_pagemap *dup;
 314
 315                 rcu_read_lock();
 316                 dup = find_dev_pagemap(key);
 317                 rcu_read_unlock();
 318                 if (dup) {
 319                         dev_err(dev, "%s: %pr collides with mapping for %s\n",
 320                                         __func__, res, dev_name(dup->dev));
 321                         error = -EBUSY;
 322                         break;
 323                 }
 324                 error = radix_tree_insert(&pgmap_radix, key >> PA_SECTION_SHIFT,
 325                                 page_map);
 326                 if (error) {
 327                         dev_err(dev, "%s: failed: %d\n", __func__, error);
 328                         break;
 329                 }
 330         }
 331         mutex_unlock(&pgmap_lock);
 332         if (error)
 333                 goto err_radix;
 334
 335         nid = dev_to_node(dev);
 336         if (nid < 0)
 337                 nid = numa_mem_id();
 338
 339         align_start = res->start & ~(SECTION_SIZE - 1);
 340         align_size = ALIGN(resource_size(res), SECTION_SIZE);
 341         error = arch_add_memory(nid, align_start, align_size, true);
 342         if (error)
 343                 goto err_add_memory;
 344
 345         for_each_device_pfn(pfn, page_map) {
 346                 struct page *page = pfn_to_page(pfn);
 347
 348                 /* ZONE_DEVICE pages must never appear on a slab lru */
 349                 list_force_poison(&page->lru);
 350                 page->pgmap = pgmap;
 351         }
 352         devres_add(dev, page_map);
 353         return __va(res->start);
 354
 355  err_add_memory:
 356  err_radix:
 357         pgmap_radix_release(res);
 358         devres_free(page_map);
 359         return ERR_PTR(error);
 360 }
 361 EXPORT_SYMBOL(devm_memremap_pages);
 362
 363 unsigned long vmem_altmap_offset(struct vmem_altmap *altmap)
 364 {
 365         /* number of pfns from base where pfn_to_page() is valid */
 366         return altmap->reserve + altmap->free;
 367 }
 368
 369 void vmem_altmap_free(struct vmem_altmap *altmap, unsigned long nr_pfns)
 370 {
 371         altmap->alloc -= nr_pfns;
 372 }
 373
 374 #ifdef CONFIG_SPARSEMEM_VMEMMAP
 375 struct vmem_altmap *to_vmem_altmap(unsigned long memmap_start)
 376 {
 377         /*
 378          * 'memmap_start' is the virtual address for the first "struct
 379          * page" in this range of the vmemmap array.  In the case of
 380          * CONFIG_SPARSE_VMEMMAP a page_to_pfn conversion is simple
 381          * pointer arithmetic, so we can perform this to_vmem_altmap()
 382          * conversion without concern for the initialization state of
 383          * the struct page fields.
 384          */
 385         struct page *page = (struct page *) memmap_start;
 386         struct dev_pagemap *pgmap;
 387
 388         /*
 389          * Uncoditionally retrieve a dev_pagemap associated with the
 390          * given physical address, this is only for use in the
 391          * arch_{add|remove}_memory() for setting up and tearing down
 392          * the memmap.
 393          */
 394         rcu_read_lock();
 395         pgmap = find_dev_pagemap(__pfn_to_phys(page_to_pfn(page)));
 396         rcu_read_unlock();
 397
 398         return pgmap ? pgmap->altmap : NULL;
 399 }
 400 #endif /* CONFIG_SPARSEMEM_VMEMMAP */
 401 #endif /* CONFIG_ZONE_DEVICE */