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 #ifndef arch_memremap_wb
  31 static void *arch_memremap_wb(resource_size_t offset, unsigned long size)
  32 {
  33         return (__force void *)ioremap_cache(offset, size);
  34 }
  35 #endif
  36
  37 #ifndef arch_memremap_can_ram_remap
  38 static bool arch_memremap_can_ram_remap(resource_size_t offset, size_t size,
  39                                         unsigned long flags)
  40 {
  41         return true;
  42 }
  43 #endif
  44
  45 static void *try_ram_remap(resource_size_t offset, size_t size,
  46                            unsigned long flags)
  47 {
  48         unsigned long pfn = PHYS_PFN(offset);
  49
  50         /* In the simple case just return the existing linear address */
  51         if (pfn_valid(pfn) && !PageHighMem(pfn_to_page(pfn)) &&
  52             arch_memremap_can_ram_remap(offset, size, flags))
  53                 return __va(offset);
  54
  55         return NULL; /* fallback to arch_memremap_wb */
  56 }
  57
  58 /**
  59  * memremap() - remap an iomem_resource as cacheable memory
  60  * @offset: iomem resource start address
  61  * @size: size of remap
  62  * @flags: any of MEMREMAP_WB, MEMREMAP_WT, MEMREMAP_WC,
  63  *                MEMREMAP_ENC, MEMREMAP_DEC
  64  *
  65  * memremap() is "ioremap" for cases where it is known that the resource
  66  * being mapped does not have i/o side effects and the __iomem
  67  * annotation is not applicable. In the case of multiple flags, the different
  68  * mapping types will be attempted in the order listed below until one of
  69  * them succeeds.
  70  *
  71  * MEMREMAP_WB - matches the default mapping for System RAM on
  72  * the architecture.  This is usually a read-allocate write-back cache.
  73  * Morever, if MEMREMAP_WB is specified and the requested remap region is RAM
  74  * memremap() will bypass establishing a new mapping and instead return
  75  * a pointer into the direct map.
  76  *
  77  * MEMREMAP_WT - establish a mapping whereby writes either bypass the
  78  * cache or are written through to memory and never exist in a
  79  * cache-dirty state with respect to program visibility.  Attempts to
  80  * map System RAM with this mapping type will fail.
  81  *
  82  * MEMREMAP_WC - establish a writecombine mapping, whereby writes may
  83  * be coalesced together (e.g. in the CPU's write buffers), but is otherwise
  84  * uncached. Attempts to map System RAM with this mapping type will fail.
  85  */
  86 void *memremap(resource_size_t offset, size_t size, unsigned long flags)
  87 {
  88         int is_ram = region_intersects(offset, size,
  89                                        IORESOURCE_SYSTEM_RAM, IORES_DESC_NONE);
  90         void *addr = NULL;
  91
  92         if (!flags)
  93                 return NULL;
  94
  95         if (is_ram == REGION_MIXED) {
  96                 WARN_ONCE(1, "memremap attempted on mixed range %pa size: %#lx\n",
  97                                 &offset, (unsigned long) size);
  98                 return NULL;
  99         }
 100
 101         /* Try all mapping types requested until one returns non-NULL */
 102         if (flags & MEMREMAP_WB) {
 103                 /*
 104                  * MEMREMAP_WB is special in that it can be satisifed
 105                  * from the direct map.  Some archs depend on the
 106                  * capability of memremap() to autodetect cases where
 107                  * the requested range is potentially in System RAM.
 108                  */
 109                 if (is_ram == REGION_INTERSECTS)
 110                         addr = try_ram_remap(offset, size, flags);
 111                 if (!addr)
 112                         addr = arch_memremap_wb(offset, size);
 113         }
 114
 115         /*
 116          * If we don't have a mapping yet and other request flags are
 117          * present then we will be attempting to establish a new virtual
 118          * address mapping.  Enforce that this mapping is not aliasing
 119          * System RAM.
 120          */
 121         if (!addr && is_ram == REGION_INTERSECTS && flags != MEMREMAP_WB) {
 122                 WARN_ONCE(1, "memremap attempted on ram %pa size: %#lx\n",
 123                                 &offset, (unsigned long) size);
 124                 return NULL;
 125         }
 126
 127         if (!addr && (flags & MEMREMAP_WT))
 128                 addr = ioremap_wt(offset, size);
 129
 130         if (!addr && (flags & MEMREMAP_WC))
 131                 addr = ioremap_wc(offset, size);
 132
 133         return addr;
 134 }
 135 EXPORT_SYMBOL(memremap);
 136
 137 void memunmap(void *addr)
 138 {
 139         if (is_vmalloc_addr(addr))
 140                 iounmap((void __iomem *) addr);
 141 }
 142 EXPORT_SYMBOL(memunmap);
 143
 144 static void devm_memremap_release(struct device *dev, void *res)
 145 {
 146         memunmap(*(void **)res);
 147 }
 148
 149 static int devm_memremap_match(struct device *dev, void *res, void *match_data)
 150 {
 151         return *(void **)res == match_data;
 152 }
 153
 154 void *devm_memremap(struct device *dev, resource_size_t offset,
 155                 size_t size, unsigned long flags)
 156 {
 157         void **ptr, *addr;
 158
 159         ptr = devres_alloc_node(devm_memremap_release, sizeof(*ptr), GFP_KERNEL,
 160                         dev_to_node(dev));
 161         if (!ptr)
 162                 return ERR_PTR(-ENOMEM);
 163
 164         addr = memremap(offset, size, flags);
 165         if (addr) {
 166                 *ptr = addr;
 167                 devres_add(dev, ptr);
 168         } else {
 169                 devres_free(ptr);
 170                 return ERR_PTR(-ENXIO);
 171         }
 172
 173         return addr;
 174 }
 175 EXPORT_SYMBOL(devm_memremap);
 176
 177 void devm_memunmap(struct device *dev, void *addr)
 178 {
 179         WARN_ON(devres_release(dev, devm_memremap_release,
 180                                 devm_memremap_match, addr));
 181 }
 182 EXPORT_SYMBOL(devm_memunmap);
 183
 184 #ifdef CONFIG_ZONE_DEVICE
 185 static DEFINE_MUTEX(pgmap_lock);
 186 static RADIX_TREE(pgmap_radix, GFP_KERNEL);
 187 #define SECTION_MASK ~((1UL << PA_SECTION_SHIFT) - 1)
 188 #define SECTION_SIZE (1UL << PA_SECTION_SHIFT)
 189
 190 struct page_map {
 191         struct resource res;
 192         struct percpu_ref *ref;
 193         struct dev_pagemap pgmap;
 194         struct vmem_altmap altmap;
 195 };
 196
 197 static void pgmap_radix_release(struct resource *res)
 198 {
 199         resource_size_t key, align_start, align_size, align_end;
 200
 201         align_start = res->start & ~(SECTION_SIZE - 1);
 202         align_size = ALIGN(resource_size(res), SECTION_SIZE);
 203         align_end = align_start + align_size - 1;
 204
 205         mutex_lock(&pgmap_lock);
 206         for (key = res->start; key <= res->end; key += SECTION_SIZE)
 207                 radix_tree_delete(&pgmap_radix, key >> PA_SECTION_SHIFT);
 208         mutex_unlock(&pgmap_lock);
 209 }
 210
 211 static unsigned long pfn_first(struct page_map *page_map)
 212 {
 213         struct dev_pagemap *pgmap = &page_map->pgmap;
 214         const struct resource *res = &page_map->res;
 215         struct vmem_altmap *altmap = pgmap->altmap;
 216         unsigned long pfn;
 217
 218         pfn = res->start >> PAGE_SHIFT;
 219         if (altmap)
 220                 pfn += vmem_altmap_offset(altmap);
 221         return pfn;
 222 }
 223
 224 static unsigned long pfn_end(struct page_map *page_map)
 225 {
 226         const struct resource *res = &page_map->res;
 227
 228         return (res->start + resource_size(res)) >> PAGE_SHIFT;
 229 }
 230
 231 #define for_each_device_pfn(pfn, map) \
 232         for (pfn = pfn_first(map); pfn < pfn_end(map); pfn++)
 233
 234 static void devm_memremap_pages_release(struct device *dev, void *data)
 235 {
 236         struct page_map *page_map = data;
 237         struct resource *res = &page_map->res;
 238         resource_size_t align_start, align_size;
 239         struct dev_pagemap *pgmap = &page_map->pgmap;
 240         unsigned long pfn;
 241
 242         for_each_device_pfn(pfn, page_map)
 243                 put_page(pfn_to_page(pfn));
 244
 245         if (percpu_ref_tryget_live(pgmap->ref)) {
 246                 dev_WARN(dev, "%s: page mapping is still live!\n", __func__);
 247                 percpu_ref_put(pgmap->ref);
 248         }
 249
 250         /* pages are dead and unused, undo the arch mapping */
 251         align_start = res->start & ~(SECTION_SIZE - 1);
 252         align_size = ALIGN(resource_size(res), SECTION_SIZE);
 253
 254         mem_hotplug_begin();
 255         arch_remove_memory(align_start, align_size);
 256         mem_hotplug_done();
 257
 258         untrack_pfn(NULL, PHYS_PFN(align_start), align_size);
 259         pgmap_radix_release(res);
 260         dev_WARN_ONCE(dev, pgmap->altmap && pgmap->altmap->alloc,
 261                         "%s: failed to free all reserved pages\n", __func__);
 262 }
 263
 264 /* assumes rcu_read_lock() held at entry */
 265 struct dev_pagemap *find_dev_pagemap(resource_size_t phys)
 266 {
 267         struct page_map *page_map;
 268
 269         WARN_ON_ONCE(!rcu_read_lock_held());
 270
 271         page_map = radix_tree_lookup(&pgmap_radix, phys >> PA_SECTION_SHIFT);
 272         return page_map ? &page_map->pgmap : NULL;
 273 }
 274
 275 /**
 276  * devm_memremap_pages - remap and provide memmap backing for the given resource
 277  * @dev: hosting device for @res
 278  * @res: "host memory" address range
 279  * @ref: a live per-cpu reference count
 280  * @altmap: optional descriptor for allocating the memmap from @res
 281  *
 282  * Notes:
 283  * 1/ @ref must be 'live' on entry and 'dead' before devm_memunmap_pages() time
 284  *    (or devm release event). The expected order of events is that @ref has
 285  *    been through percpu_ref_kill() before devm_memremap_pages_release(). The
 286  *    wait for the completion of all references being dropped and
 287  *    percpu_ref_exit() must occur after devm_memremap_pages_release().
 288  *
 289  * 2/ @res is expected to be a host memory range that could feasibly be
 290  *    treated as a "System RAM" range, i.e. not a device mmio range, but
 291  *    this is not enforced.
 292  */
 293 void *devm_memremap_pages(struct device *dev, struct resource *res,
 294                 struct percpu_ref *ref, struct vmem_altmap *altmap)
 295 {
 296         resource_size_t key, align_start, align_size, align_end;
 297         pgprot_t pgprot = PAGE_KERNEL;
 298         struct dev_pagemap *pgmap;
 299         struct page_map *page_map;
 300         int error, nid, is_ram;
 301         unsigned long pfn;
 302
 303         align_start = res->start & ~(SECTION_SIZE - 1);
 304         align_size = ALIGN(res->start + resource_size(res), SECTION_SIZE)
 305                 - align_start;
 306         is_ram = region_intersects(align_start, align_size,
 307                 IORESOURCE_SYSTEM_RAM, IORES_DESC_NONE);
 308
 309         if (is_ram == REGION_MIXED) {
 310                 WARN_ONCE(1, "%s attempted on mixed region %pr\n",
 311                                 __func__, res);
 312                 return ERR_PTR(-ENXIO);
 313         }
 314
 315         if (is_ram == REGION_INTERSECTS)
 316                 return __va(res->start);
 317
 318         if (!ref)
 319                 return ERR_PTR(-EINVAL);
 320
 321         page_map = devres_alloc_node(devm_memremap_pages_release,
 322                         sizeof(*page_map), GFP_KERNEL, dev_to_node(dev));
 323         if (!page_map)
 324                 return ERR_PTR(-ENOMEM);
 325         pgmap = &page_map->pgmap;
 326
 327         memcpy(&page_map->res, res, sizeof(*res));
 328
 329         pgmap->dev = dev;
 330         if (altmap) {
 331                 memcpy(&page_map->altmap, altmap, sizeof(*altmap));
 332                 pgmap->altmap = &page_map->altmap;
 333         }
 334         pgmap->ref = ref;
 335         pgmap->res = &page_map->res;
 336
 337         mutex_lock(&pgmap_lock);
 338         error = 0;
 339         align_end = align_start + align_size - 1;
 340         for (key = align_start; key <= align_end; key += SECTION_SIZE) {
 341                 struct dev_pagemap *dup;
 342
 343                 rcu_read_lock();
 344                 dup = find_dev_pagemap(key);
 345                 rcu_read_unlock();
 346                 if (dup) {
 347                         dev_err(dev, "%s: %pr collides with mapping for %s\n",
 348                                         __func__, res, dev_name(dup->dev));
 349                         error = -EBUSY;
 350                         break;
 351                 }
 352                 error = radix_tree_insert(&pgmap_radix, key >> PA_SECTION_SHIFT,
 353                                 page_map);
 354                 if (error) {
 355                         dev_err(dev, "%s: failed: %d\n", __func__, error);
 356                         break;
 357                 }
 358         }
 359         mutex_unlock(&pgmap_lock);
 360         if (error)
 361                 goto err_radix;
 362
 363         nid = dev_to_node(dev);
 364         if (nid < 0)
 365                 nid = numa_mem_id();
 366
 367         error = track_pfn_remap(NULL, &pgprot, PHYS_PFN(align_start), 0,
 368                         align_size);
 369         if (error)
 370                 goto err_pfn_remap;
 371
 372         mem_hotplug_begin();
 373         error = arch_add_memory(nid, align_start, align_size, false);
 374         if (!error)
 375                 move_pfn_range_to_zone(&NODE_DATA(nid)->node_zones[ZONE_DEVICE],
 376                                         align_start >> PAGE_SHIFT,
 377                                         align_size >> PAGE_SHIFT);
 378         mem_hotplug_done();
 379         if (error)
 380                 goto err_add_memory;
 381
 382         for_each_device_pfn(pfn, page_map) {
 383                 struct page *page = pfn_to_page(pfn);
 384
 385                 /*
 386                  * ZONE_DEVICE pages union ->lru with a ->pgmap back
 387                  * pointer.  It is a bug if a ZONE_DEVICE page is ever
 388                  * freed or placed on a driver-private list.  Seed the
 389                  * storage with LIST_POISON* values.
 390                  */
 391                 list_del(&page->lru);
 392                 page->pgmap = pgmap;
 393                 percpu_ref_get(ref);
 394         }
 395         devres_add(dev, page_map);
 396         return __va(res->start);
 397
 398  err_add_memory:
 399         untrack_pfn(NULL, PHYS_PFN(align_start), align_size);
 400  err_pfn_remap:
 401  err_radix:
 402         pgmap_radix_release(res);
 403         devres_free(page_map);
 404         return ERR_PTR(error);
 405 }
 406 EXPORT_SYMBOL(devm_memremap_pages);
 407
 408 unsigned long vmem_altmap_offset(struct vmem_altmap *altmap)
 409 {
 410         /* number of pfns from base where pfn_to_page() is valid */
 411         return altmap->reserve + altmap->free;
 412 }
 413
 414 void vmem_altmap_free(struct vmem_altmap *altmap, unsigned long nr_pfns)
 415 {
 416         altmap->alloc -= nr_pfns;
 417 }
 418
 419 struct vmem_altmap *to_vmem_altmap(unsigned long memmap_start)
 420 {
 421         /*
 422          * 'memmap_start' is the virtual address for the first "struct
 423          * page" in this range of the vmemmap array.  In the case of
 424          * CONFIG_SPARSEMEM_VMEMMAP a page_to_pfn conversion is simple
 425          * pointer arithmetic, so we can perform this to_vmem_altmap()
 426          * conversion without concern for the initialization state of
 427          * the struct page fields.
 428          */
 429         struct page *page = (struct page *) memmap_start;
 430         struct dev_pagemap *pgmap;
 431
 432         /*
 433          * Unconditionally retrieve a dev_pagemap associated with the
 434          * given physical address, this is only for use in the
 435          * arch_{add|remove}_memory() for setting up and tearing down
 436          * the memmap.
 437          */
 438         rcu_read_lock();
 439         pgmap = find_dev_pagemap(__pfn_to_phys(page_to_pfn(page)));
 440         rcu_read_unlock();
 441
 442         return pgmap ? pgmap->altmap : NULL;
 443 }
 444 #endif /* CONFIG_ZONE_DEVICE */