mm/memremap.c

   1 /* SPDX-License-Identifier: GPL-2.0 */
   2 /* Copyright(c) 2015 Intel Corporation. All rights reserved. */
   3 #include <linux/device.h>
   4 #include <linux/io.h>
   5 #include <linux/kasan.h>
   6 #include <linux/memory_hotplug.h>
   7 #include <linux/mm.h>
   8 #include <linux/pfn_t.h>
   9 #include <linux/swap.h>
  10 #include <linux/swapops.h>
  11 #include <linux/types.h>
  12 #include <linux/wait_bit.h>
  13 #include <linux/xarray.h>
  14
  15 static DEFINE_XARRAY(pgmap_array);
  16 #define SECTION_MASK ~((1UL << PA_SECTION_SHIFT) - 1)
  17 #define SECTION_SIZE (1UL << PA_SECTION_SHIFT)
  18
  19 #ifdef CONFIG_DEV_PAGEMAP_OPS
  20 DEFINE_STATIC_KEY_FALSE(devmap_managed_key);
  21 EXPORT_SYMBOL(devmap_managed_key);
  22 static atomic_t devmap_managed_enable;
  23
  24 static void devmap_managed_enable_put(void *data)
  25 {
  26         if (atomic_dec_and_test(&devmap_managed_enable))
  27                 static_branch_disable(&devmap_managed_key);
  28 }
  29
  30 static int devmap_managed_enable_get(struct device *dev, struct dev_pagemap *pgmap)
  31 {
  32         if (!pgmap->ops || !pgmap->ops->page_free) {
  33                 WARN(1, "Missing page_free method\n");
  34                 return -EINVAL;
  35         }
  36
  37         if (atomic_inc_return(&devmap_managed_enable) == 1)
  38                 static_branch_enable(&devmap_managed_key);
  39         return devm_add_action_or_reset(dev, devmap_managed_enable_put, NULL);
  40 }
  41 #else
  42 static int devmap_managed_enable_get(struct device *dev, struct dev_pagemap *pgmap)
  43 {
  44         return -EINVAL;
  45 }
  46 #endif /* CONFIG_DEV_PAGEMAP_OPS */
  47
  48 static void pgmap_array_delete(struct resource *res)
  49 {
  50         xa_store_range(&pgmap_array, PHYS_PFN(res->start), PHYS_PFN(res->end),
  51                         NULL, GFP_KERNEL);
  52         synchronize_rcu();
  53 }
  54
  55 static unsigned long pfn_first(struct dev_pagemap *pgmap)
  56 {
  57         return PHYS_PFN(pgmap->res.start) +
  58                 vmem_altmap_offset(pgmap_altmap(pgmap));
  59 }
  60
  61 static unsigned long pfn_end(struct dev_pagemap *pgmap)
  62 {
  63         const struct resource *res = &pgmap->res;
  64
  65         return (res->start + resource_size(res)) >> PAGE_SHIFT;
  66 }
  67
  68 static unsigned long pfn_next(unsigned long pfn)
  69 {
  70         if (pfn % 1024 == 0)
  71                 cond_resched();
  72         return pfn + 1;
  73 }
  74
  75 #define for_each_device_pfn(pfn, map) \
  76         for (pfn = pfn_first(map); pfn < pfn_end(map); pfn = pfn_next(pfn))
  77
  78 static void dev_pagemap_kill(struct dev_pagemap *pgmap)
  79 {
  80         if (pgmap->ops && pgmap->ops->kill)
  81                 pgmap->ops->kill(pgmap);
  82         else
  83                 percpu_ref_kill(pgmap->ref);
  84 }
  85
  86 static void dev_pagemap_cleanup(struct dev_pagemap *pgmap)
  87 {
  88         if (pgmap->ops && pgmap->ops->cleanup) {
  89                 pgmap->ops->cleanup(pgmap);
  90         } else {
  91                 wait_for_completion(&pgmap->done);
  92                 percpu_ref_exit(pgmap->ref);
  93         }
  94         /*
  95          * Undo the pgmap ref assignment for the internal case as the
  96          * caller may re-enable the same pgmap.
  97          */
  98         if (pgmap->ref == &pgmap->internal_ref)
  99                 pgmap->ref = NULL;
 100 }
 101
 102 static void devm_memremap_pages_release(void *data)
 103 {
 104         struct dev_pagemap *pgmap = data;
 105         struct device *dev = pgmap->dev;
 106         struct resource *res = &pgmap->res;
 107         unsigned long pfn;
 108         int nid;
 109
 110         dev_pagemap_kill(pgmap);
 111         for_each_device_pfn(pfn, pgmap)
 112                 put_page(pfn_to_page(pfn));
 113         dev_pagemap_cleanup(pgmap);
 114
 115         /* pages are dead and unused, undo the arch mapping */
 116         nid = page_to_nid(pfn_to_page(PHYS_PFN(res->start)));
 117
 118         mem_hotplug_begin();
 119         if (pgmap->type == MEMORY_DEVICE_PRIVATE) {
 120                 pfn = PHYS_PFN(res->start);
 121                 __remove_pages(page_zone(pfn_to_page(pfn)), pfn,
 122                                  PHYS_PFN(resource_size(res)), NULL);
 123         } else {
 124                 arch_remove_memory(nid, res->start, resource_size(res),
 125                                 pgmap_altmap(pgmap));
 126                 kasan_remove_zero_shadow(__va(res->start), resource_size(res));
 127         }
 128         mem_hotplug_done();
 129
 130         untrack_pfn(NULL, PHYS_PFN(res->start), resource_size(res));
 131         pgmap_array_delete(res);
 132         dev_WARN_ONCE(dev, pgmap->altmap.alloc,
 133                       "%s: failed to free all reserved pages\n", __func__);
 134 }
 135
 136 static void dev_pagemap_percpu_release(struct percpu_ref *ref)
 137 {
 138         struct dev_pagemap *pgmap =
 139                 container_of(ref, struct dev_pagemap, internal_ref);
 140
 141         complete(&pgmap->done);
 142 }
 143
 144 /**
 145  * devm_memremap_pages - remap and provide memmap backing for the given resource
 146  * @dev: hosting device for @res
 147  * @pgmap: pointer to a struct dev_pagemap
 148  *
 149  * Notes:
 150  * 1/ At a minimum the res and type members of @pgmap must be initialized
 151  *    by the caller before passing it to this function
 152  *
 153  * 2/ The altmap field may optionally be initialized, in which case
 154  *    PGMAP_ALTMAP_VALID must be set in pgmap->flags.
 155  *
 156  * 3/ The ref field may optionally be provided, in which pgmap->ref must be
 157  *    'live' on entry and will be killed and reaped at
 158  *    devm_memremap_pages_release() time, or if this routine fails.
 159  *
 160  * 4/ res is expected to be a host memory range that could feasibly be
 161  *    treated as a "System RAM" range, i.e. not a device mmio range, but
 162  *    this is not enforced.
 163  */
 164 void *devm_memremap_pages(struct device *dev, struct dev_pagemap *pgmap)
 165 {
 166         struct resource *res = &pgmap->res;
 167         struct dev_pagemap *conflict_pgmap;
 168         struct mhp_restrictions restrictions = {
 169                 /*
 170                  * We do not want any optional features only our own memmap
 171                  */
 172                 .altmap = pgmap_altmap(pgmap),
 173         };
 174         pgprot_t pgprot = PAGE_KERNEL;
 175         int error, nid, is_ram;
 176         bool need_devmap_managed = true;
 177
 178         switch (pgmap->type) {
 179         case MEMORY_DEVICE_PRIVATE:
 180                 if (!IS_ENABLED(CONFIG_DEVICE_PRIVATE)) {
 181                         WARN(1, "Device private memory not supported\n");
 182                         return ERR_PTR(-EINVAL);
 183                 }
 184                 if (!pgmap->ops || !pgmap->ops->migrate_to_ram) {
 185                         WARN(1, "Missing migrate_to_ram method\n");
 186                         return ERR_PTR(-EINVAL);
 187                 }
 188                 break;
 189         case MEMORY_DEVICE_FS_DAX:
 190                 if (!IS_ENABLED(CONFIG_ZONE_DEVICE) ||
 191                     IS_ENABLED(CONFIG_FS_DAX_LIMITED)) {
 192                         WARN(1, "File system DAX not supported\n");
 193                         return ERR_PTR(-EINVAL);
 194                 }
 195                 break;
 196         case MEMORY_DEVICE_DEVDAX:
 197         case MEMORY_DEVICE_PCI_P2PDMA:
 198                 need_devmap_managed = false;
 199                 break;
 200         default:
 201                 WARN(1, "Invalid pgmap type %d\n", pgmap->type);
 202                 break;
 203         }
 204
 205         if (!pgmap->ref) {
 206                 if (pgmap->ops && (pgmap->ops->kill || pgmap->ops->cleanup))
 207                         return ERR_PTR(-EINVAL);
 208
 209                 init_completion(&pgmap->done);
 210                 error = percpu_ref_init(&pgmap->internal_ref,
 211                                 dev_pagemap_percpu_release, 0, GFP_KERNEL);
 212                 if (error)
 213                         return ERR_PTR(error);
 214                 pgmap->ref = &pgmap->internal_ref;
 215         } else {
 216                 if (!pgmap->ops || !pgmap->ops->kill || !pgmap->ops->cleanup) {
 217                         WARN(1, "Missing reference count teardown definition\n");
 218                         return ERR_PTR(-EINVAL);
 219                 }
 220         }
 221
 222         if (need_devmap_managed) {
 223                 error = devmap_managed_enable_get(dev, pgmap);
 224                 if (error)
 225                         return ERR_PTR(error);
 226         }
 227
 228         conflict_pgmap = get_dev_pagemap(PHYS_PFN(res->start), NULL);
 229         if (conflict_pgmap) {
 230                 dev_WARN(dev, "Conflicting mapping in same section\n");
 231                 put_dev_pagemap(conflict_pgmap);
 232                 error = -ENOMEM;
 233                 goto err_array;
 234         }
 235
 236         conflict_pgmap = get_dev_pagemap(PHYS_PFN(res->end), NULL);
 237         if (conflict_pgmap) {
 238                 dev_WARN(dev, "Conflicting mapping in same section\n");
 239                 put_dev_pagemap(conflict_pgmap);
 240                 error = -ENOMEM;
 241                 goto err_array;
 242         }
 243
 244         is_ram = region_intersects(res->start, resource_size(res),
 245                 IORESOURCE_SYSTEM_RAM, IORES_DESC_NONE);
 246
 247         if (is_ram != REGION_DISJOINT) {
 248                 WARN_ONCE(1, "%s attempted on %s region %pr\n", __func__,
 249                                 is_ram == REGION_MIXED ? "mixed" : "ram", res);
 250                 error = -ENXIO;
 251                 goto err_array;
 252         }
 253
 254         pgmap->dev = dev;
 255
 256         error = xa_err(xa_store_range(&pgmap_array, PHYS_PFN(res->start),
 257                                 PHYS_PFN(res->end), pgmap, GFP_KERNEL));
 258         if (error)
 259                 goto err_array;
 260
 261         nid = dev_to_node(dev);
 262         if (nid < 0)
 263                 nid = numa_mem_id();
 264
 265         error = track_pfn_remap(NULL, &pgprot, PHYS_PFN(res->start), 0,
 266                         resource_size(res));
 267         if (error)
 268                 goto err_pfn_remap;
 269
 270         mem_hotplug_begin();
 271
 272         /*
 273          * For device private memory we call add_pages() as we only need to
 274          * allocate and initialize struct page for the device memory. More-
 275          * over the device memory is un-accessible thus we do not want to
 276          * create a linear mapping for the memory like arch_add_memory()
 277          * would do.
 278          *
 279          * For all other device memory types, which are accessible by
 280          * the CPU, we do want the linear mapping and thus use
 281          * arch_add_memory().
 282          */
 283         if (pgmap->type == MEMORY_DEVICE_PRIVATE) {
 284                 error = add_pages(nid, PHYS_PFN(res->start),
 285                                 PHYS_PFN(resource_size(res)), &restrictions);
 286         } else {
 287                 error = kasan_add_zero_shadow(__va(res->start), resource_size(res));
 288                 if (error) {
 289                         mem_hotplug_done();
 290                         goto err_kasan;
 291                 }
 292
 293                 error = arch_add_memory(nid, res->start, resource_size(res),
 294                                         &restrictions);
 295         }
 296
 297         if (!error) {
 298                 struct zone *zone;
 299
 300                 zone = &NODE_DATA(nid)->node_zones[ZONE_DEVICE];
 301                 move_pfn_range_to_zone(zone, PHYS_PFN(res->start),
 302                                 PHYS_PFN(resource_size(res)), restrictions.altmap);
 303         }
 304
 305         mem_hotplug_done();
 306         if (error)
 307                 goto err_add_memory;
 308
 309         /*
 310          * Initialization of the pages has been deferred until now in order
 311          * to allow us to do the work while not holding the hotplug lock.
 312          */
 313         memmap_init_zone_device(&NODE_DATA(nid)->node_zones[ZONE_DEVICE],
 314                                 PHYS_PFN(res->start),
 315                                 PHYS_PFN(resource_size(res)), pgmap);
 316         percpu_ref_get_many(pgmap->ref, pfn_end(pgmap) - pfn_first(pgmap));
 317
 318         error = devm_add_action_or_reset(dev, devm_memremap_pages_release,
 319                         pgmap);
 320         if (error)
 321                 return ERR_PTR(error);
 322
 323         return __va(res->start);
 324
 325  err_add_memory:
 326         kasan_remove_zero_shadow(__va(res->start), resource_size(res));
 327  err_kasan:
 328         untrack_pfn(NULL, PHYS_PFN(res->start), resource_size(res));
 329  err_pfn_remap:
 330         pgmap_array_delete(res);
 331  err_array:
 332         dev_pagemap_kill(pgmap);
 333         dev_pagemap_cleanup(pgmap);
 334         return ERR_PTR(error);
 335 }
 336 EXPORT_SYMBOL_GPL(devm_memremap_pages);
 337
 338 void devm_memunmap_pages(struct device *dev, struct dev_pagemap *pgmap)
 339 {
 340         devm_release_action(dev, devm_memremap_pages_release, pgmap);
 341 }
 342 EXPORT_SYMBOL_GPL(devm_memunmap_pages);
 343
 344 unsigned long vmem_altmap_offset(struct vmem_altmap *altmap)
 345 {
 346         /* number of pfns from base where pfn_to_page() is valid */
 347         if (altmap)
 348                 return altmap->reserve + altmap->free;
 349         return 0;
 350 }
 351
 352 void vmem_altmap_free(struct vmem_altmap *altmap, unsigned long nr_pfns)
 353 {
 354         altmap->alloc -= nr_pfns;
 355 }
 356
 357 /**
 358  * get_dev_pagemap() - take a new live reference on the dev_pagemap for @pfn
 359  * @pfn: page frame number to lookup page_map
 360  * @pgmap: optional known pgmap that already has a reference
 361  *
 362  * If @pgmap is non-NULL and covers @pfn it will be returned as-is.  If @pgmap
 363  * is non-NULL but does not cover @pfn the reference to it will be released.
 364  */
 365 struct dev_pagemap *get_dev_pagemap(unsigned long pfn,
 366                 struct dev_pagemap *pgmap)
 367 {
 368         resource_size_t phys = PFN_PHYS(pfn);
 369
 370         /*
 371          * In the cached case we're already holding a live reference.
 372          */
 373         if (pgmap) {
 374                 if (phys >= pgmap->res.start && phys <= pgmap->res.end)
 375                         return pgmap;
 376                 put_dev_pagemap(pgmap);
 377         }
 378
 379         /* fall back to slow path lookup */
 380         rcu_read_lock();
 381         pgmap = xa_load(&pgmap_array, PHYS_PFN(phys));
 382         if (pgmap && !percpu_ref_tryget_live(pgmap->ref))
 383                 pgmap = NULL;
 384         rcu_read_unlock();
 385
 386         return pgmap;
 387 }
 388 EXPORT_SYMBOL_GPL(get_dev_pagemap);
 389
 390 #ifdef CONFIG_DEV_PAGEMAP_OPS
 391 void __put_devmap_managed_page(struct page *page)
 392 {
 393         int count = page_ref_dec_return(page);
 394
 395         /*
 396          * If refcount is 1 then page is freed and refcount is stable as nobody
 397          * holds a reference on the page.
 398          */
 399         if (count == 1) {
 400                 /* Clear Active bit in case of parallel mark_page_accessed */
 401                 __ClearPageActive(page);
 402                 __ClearPageWaiters(page);
 403
 404                 mem_cgroup_uncharge(page);
 405
 406                 /*
 407                  * When a device_private page is freed, the page->mapping field
 408                  * may still contain a (stale) mapping value. For example, the
 409                  * lower bits of page->mapping may still identify the page as
 410                  * an anonymous page. Ultimately, this entire field is just
 411                  * stale and wrong, and it will cause errors if not cleared.
 412                  * One example is:
 413                  *
 414                  *  migrate_vma_pages()
 415                  *    migrate_vma_insert_page()
 416                  *      page_add_new_anon_rmap()
 417                  *        __page_set_anon_rmap()
 418                  *          ...checks page->mapping, via PageAnon(page) call,
 419                  *            and incorrectly concludes that the page is an
 420                  *            anonymous page. Therefore, it incorrectly,
 421                  *            silently fails to set up the new anon rmap.
 422                  *
 423                  * For other types of ZONE_DEVICE pages, migration is either
 424                  * handled differently or not done at all, so there is no need
 425                  * to clear page->mapping.
 426                  */
 427                 if (is_device_private_page(page))
 428                         page->mapping = NULL;
 429
 430                 page->pgmap->ops->page_free(page);
 431         } else if (!count)
 432                 __put_page(page);
 433 }
 434 EXPORT_SYMBOL(__put_devmap_managed_page);
 435 #endif /* CONFIG_DEV_PAGEMAP_OPS */