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(*(void **)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                 return ERR_PTR(-ENXIO);
 142         }
 143
 144         return addr;
 145 }
 146 EXPORT_SYMBOL(devm_memremap);
 147
 148 void devm_memunmap(struct device *dev, void *addr)
 149 {
 150         WARN_ON(devres_release(dev, devm_memremap_release,
 151                                 devm_memremap_match, addr));
 152 }
 153 EXPORT_SYMBOL(devm_memunmap);
 154
 155 pfn_t phys_to_pfn_t(phys_addr_t addr, u64 flags)
 156 {
 157         return __pfn_to_pfn_t(addr >> PAGE_SHIFT, flags);
 158 }
 159 EXPORT_SYMBOL(phys_to_pfn_t);
 160
 161 #ifdef CONFIG_ZONE_DEVICE
 162 static DEFINE_MUTEX(pgmap_lock);
 163 static RADIX_TREE(pgmap_radix, GFP_KERNEL);
 164 #define SECTION_MASK ~((1UL << PA_SECTION_SHIFT) - 1)
 165 #define SECTION_SIZE (1UL << PA_SECTION_SHIFT)
 166
 167 struct page_map {
 168         struct resource res;
 169         struct percpu_ref *ref;
 170         struct dev_pagemap pgmap;
 171         struct vmem_altmap altmap;
 172 };
 173
 174 void get_zone_device_page(struct page *page)
 175 {
 176         percpu_ref_get(page->pgmap->ref);
 177 }
 178 EXPORT_SYMBOL(get_zone_device_page);
 179
 180 void put_zone_device_page(struct page *page)
 181 {
 182         put_dev_pagemap(page->pgmap);
 183 }
 184 EXPORT_SYMBOL(put_zone_device_page);
 185
 186 static void pgmap_radix_release(struct resource *res)
 187 {
 188         resource_size_t key, align_start, align_size, align_end;
 189
 190         align_start = res->start & ~(SECTION_SIZE - 1);
 191         align_size = ALIGN(resource_size(res), SECTION_SIZE);
 192         align_end = align_start + align_size - 1;
 193
 194         mutex_lock(&pgmap_lock);
 195         for (key = res->start; key <= res->end; key += SECTION_SIZE)
 196                 radix_tree_delete(&pgmap_radix, key >> PA_SECTION_SHIFT);
 197         mutex_unlock(&pgmap_lock);
 198 }
 199
 200 static unsigned long pfn_first(struct page_map *page_map)
 201 {
 202         struct dev_pagemap *pgmap = &page_map->pgmap;
 203         const struct resource *res = &page_map->res;
 204         struct vmem_altmap *altmap = pgmap->altmap;
 205         unsigned long pfn;
 206
 207         pfn = res->start >> PAGE_SHIFT;
 208         if (altmap)
 209                 pfn += vmem_altmap_offset(altmap);
 210         return pfn;
 211 }
 212
 213 static unsigned long pfn_end(struct page_map *page_map)
 214 {
 215         const struct resource *res = &page_map->res;
 216
 217         return (res->start + resource_size(res)) >> PAGE_SHIFT;
 218 }
 219
 220 #define for_each_device_pfn(pfn, map) \
 221         for (pfn = pfn_first(map); pfn < pfn_end(map); pfn++)
 222
 223 static void devm_memremap_pages_release(struct device *dev, void *data)
 224 {
 225         struct page_map *page_map = data;
 226         struct resource *res = &page_map->res;
 227         resource_size_t align_start, align_size;
 228         struct dev_pagemap *pgmap = &page_map->pgmap;
 229
 230         if (percpu_ref_tryget_live(pgmap->ref)) {
 231                 dev_WARN(dev, "%s: page mapping is still live!\n", __func__);
 232                 percpu_ref_put(pgmap->ref);
 233         }
 234
 235         /* pages are dead and unused, undo the arch mapping */
 236         align_start = res->start & ~(SECTION_SIZE - 1);
 237         align_size = ALIGN(resource_size(res), SECTION_SIZE);
 238         arch_remove_memory(align_start, align_size);
 239         pgmap_radix_release(res);
 240         dev_WARN_ONCE(dev, pgmap->altmap && pgmap->altmap->alloc,
 241                         "%s: failed to free all reserved pages\n", __func__);
 242 }
 243
 244 /* assumes rcu_read_lock() held at entry */
 245 struct dev_pagemap *find_dev_pagemap(resource_size_t phys)
 246 {
 247         struct page_map *page_map;
 248
 249         WARN_ON_ONCE(!rcu_read_lock_held());
 250
 251         page_map = radix_tree_lookup(&pgmap_radix, phys >> PA_SECTION_SHIFT);
 252         return page_map ? &page_map->pgmap : NULL;
 253 }
 254
 255 /**
 256  * devm_memremap_pages - remap and provide memmap backing for the given resource
 257  * @dev: hosting device for @res
 258  * @res: "host memory" address range
 259  * @ref: a live per-cpu reference count
 260  * @altmap: optional descriptor for allocating the memmap from @res
 261  *
 262  * Notes:
 263  * 1/ @ref must be 'live' on entry and 'dead' before devm_memunmap_pages() time
 264  *    (or devm release event).
 265  *
 266  * 2/ @res is expected to be a host memory range that could feasibly be
 267  *    treated as a "System RAM" range, i.e. not a device mmio range, but
 268  *    this is not enforced.
 269  */
 270 void *devm_memremap_pages(struct device *dev, struct resource *res,
 271                 struct percpu_ref *ref, struct vmem_altmap *altmap)
 272 {
 273         int is_ram = region_intersects(res->start, resource_size(res),
 274                         "System RAM");
 275         resource_size_t key, align_start, align_size, align_end;
 276         struct dev_pagemap *pgmap;
 277         struct page_map *page_map;
 278         unsigned long pfn;
 279         int error, nid;
 280
 281         if (is_ram == REGION_MIXED) {
 282                 WARN_ONCE(1, "%s attempted on mixed region %pr\n",
 283                                 __func__, res);
 284                 return ERR_PTR(-ENXIO);
 285         }
 286
 287         if (is_ram == REGION_INTERSECTS)
 288                 return __va(res->start);
 289
 290         if (altmap && !IS_ENABLED(CONFIG_SPARSEMEM_VMEMMAP)) {
 291                 dev_err(dev, "%s: altmap requires CONFIG_SPARSEMEM_VMEMMAP=y\n",
 292                                 __func__);
 293                 return ERR_PTR(-ENXIO);
 294         }
 295
 296         if (!ref)
 297                 return ERR_PTR(-EINVAL);
 298
 299         page_map = devres_alloc_node(devm_memremap_pages_release,
 300                         sizeof(*page_map), GFP_KERNEL, dev_to_node(dev));
 301         if (!page_map)
 302                 return ERR_PTR(-ENOMEM);
 303         pgmap = &page_map->pgmap;
 304
 305         memcpy(&page_map->res, res, sizeof(*res));
 306
 307         pgmap->dev = dev;
 308         if (altmap) {
 309                 memcpy(&page_map->altmap, altmap, sizeof(*altmap));
 310                 pgmap->altmap = &page_map->altmap;
 311         }
 312         pgmap->ref = ref;
 313         pgmap->res = &page_map->res;
 314
 315         mutex_lock(&pgmap_lock);
 316         error = 0;
 317         align_start = res->start & ~(SECTION_SIZE - 1);
 318         align_size = ALIGN(resource_size(res), SECTION_SIZE);
 319         align_end = align_start + align_size - 1;
 320         for (key = align_start; key <= align_end; key += SECTION_SIZE) {
 321                 struct dev_pagemap *dup;
 322
 323                 rcu_read_lock();
 324                 dup = find_dev_pagemap(key);
 325                 rcu_read_unlock();
 326                 if (dup) {
 327                         dev_err(dev, "%s: %pr collides with mapping for %s\n",
 328                                         __func__, res, dev_name(dup->dev));
 329                         error = -EBUSY;
 330                         break;
 331                 }
 332                 error = radix_tree_insert(&pgmap_radix, key >> PA_SECTION_SHIFT,
 333                                 page_map);
 334                 if (error) {
 335                         dev_err(dev, "%s: failed: %d\n", __func__, error);
 336                         break;
 337                 }
 338         }
 339         mutex_unlock(&pgmap_lock);
 340         if (error)
 341                 goto err_radix;
 342
 343         nid = dev_to_node(dev);
 344         if (nid < 0)
 345                 nid = numa_mem_id();
 346
 347         error = arch_add_memory(nid, align_start, align_size, true);
 348         if (error)
 349                 goto err_add_memory;
 350
 351         for_each_device_pfn(pfn, page_map) {
 352                 struct page *page = pfn_to_page(pfn);
 353
 354                 /* ZONE_DEVICE pages must never appear on a slab lru */
 355                 list_force_poison(&page->lru);
 356                 page->pgmap = pgmap;
 357         }
 358         devres_add(dev, page_map);
 359         return __va(res->start);
 360
 361  err_add_memory:
 362  err_radix:
 363         pgmap_radix_release(res);
 364         devres_free(page_map);
 365         return ERR_PTR(error);
 366 }
 367 EXPORT_SYMBOL(devm_memremap_pages);
 368
 369 unsigned long vmem_altmap_offset(struct vmem_altmap *altmap)
 370 {
 371         /* number of pfns from base where pfn_to_page() is valid */
 372         return altmap->reserve + altmap->free;
 373 }
 374
 375 void vmem_altmap_free(struct vmem_altmap *altmap, unsigned long nr_pfns)
 376 {
 377         altmap->alloc -= nr_pfns;
 378 }
 379
 380 #ifdef CONFIG_SPARSEMEM_VMEMMAP
 381 struct vmem_altmap *to_vmem_altmap(unsigned long memmap_start)
 382 {
 383         /*
 384          * 'memmap_start' is the virtual address for the first "struct
 385          * page" in this range of the vmemmap array.  In the case of
 386          * CONFIG_SPARSE_VMEMMAP a page_to_pfn conversion is simple
 387          * pointer arithmetic, so we can perform this to_vmem_altmap()
 388          * conversion without concern for the initialization state of
 389          * the struct page fields.
 390          */
 391         struct page *page = (struct page *) memmap_start;
 392         struct dev_pagemap *pgmap;
 393
 394         /*
 395          * Uncoditionally retrieve a dev_pagemap associated with the
 396          * given physical address, this is only for use in the
 397          * arch_{add|remove}_memory() for setting up and tearing down
 398          * the memmap.
 399          */
 400         rcu_read_lock();
 401         pgmap = find_dev_pagemap(__pfn_to_phys(page_to_pfn(page)));
 402         rcu_read_unlock();
 403
 404         return pgmap ? pgmap->altmap : NULL;
 405 }
 406 #endif /* CONFIG_SPARSEMEM_VMEMMAP */
 407 #endif /* CONFIG_ZONE_DEVICE */