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