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
  17 #ifdef CONFIG_DEV_PAGEMAP_OPS
  18 DEFINE_STATIC_KEY_FALSE(devmap_managed_key);
  19 EXPORT_SYMBOL(devmap_managed_key);
  20 static atomic_t devmap_managed_enable;
  21
  22 static void devmap_managed_enable_put(void)
  23 {
  24         if (atomic_dec_and_test(&devmap_managed_enable))
  25                 static_branch_disable(&devmap_managed_key);
  26 }
  27
  28 static int devmap_managed_enable_get(struct dev_pagemap *pgmap)
  29 {
  30         if (pgmap->type == MEMORY_DEVICE_PRIVATE &&
  31             (!pgmap->ops || !pgmap->ops->page_free)) {
  32                 WARN(1, "Missing page_free method\n");
  33                 return -EINVAL;
  34         }
  35
  36         if (atomic_inc_return(&devmap_managed_enable) == 1)
  37                 static_branch_enable(&devmap_managed_key);
  38         return 0;
  39 }
  40 #else
  41 static int devmap_managed_enable_get(struct dev_pagemap *pgmap)
  42 {
  43         return -EINVAL;
  44 }
  45 static void devmap_managed_enable_put(void)
  46 {
  47 }
  48 #endif /* CONFIG_DEV_PAGEMAP_OPS */
  49
  50 static void pgmap_array_delete(struct resource *res)
  51 {
  52         xa_store_range(&pgmap_array, PHYS_PFN(res->start), PHYS_PFN(res->end),
  53                         NULL, GFP_KERNEL);
  54         synchronize_rcu();
  55 }
  56
  57 static unsigned long pfn_first(struct dev_pagemap *pgmap)
  58 {
  59         return PHYS_PFN(pgmap->res.start) +
  60                 vmem_altmap_offset(pgmap_altmap(pgmap));
  61 }
  62
  63 static unsigned long pfn_end(struct dev_pagemap *pgmap)
  64 {
  65         const struct resource *res = &pgmap->res;
  66
  67         return (res->start + resource_size(res)) >> PAGE_SHIFT;
  68 }
  69
  70 static unsigned long pfn_next(unsigned long pfn)
  71 {
  72         if (pfn % 1024 == 0)
  73                 cond_resched();
  74         return pfn + 1;
  75 }
  76
  77 #define for_each_device_pfn(pfn, map) \
  78         for (pfn = pfn_first(map); pfn < pfn_end(map); pfn = pfn_next(pfn))
  79
  80 static void dev_pagemap_kill(struct dev_pagemap *pgmap)
  81 {
  82         if (pgmap->ops && pgmap->ops->kill)
  83                 pgmap->ops->kill(pgmap);
  84         else
  85                 percpu_ref_kill(pgmap->ref);
  86 }
  87
  88 static void dev_pagemap_cleanup(struct dev_pagemap *pgmap)
  89 {
  90         if (pgmap->ops && pgmap->ops->cleanup) {
  91                 pgmap->ops->cleanup(pgmap);
  92         } else {
  93                 wait_for_completion(&pgmap->done);
  94                 percpu_ref_exit(pgmap->ref);
  95         }
  96         /*
  97          * Undo the pgmap ref assignment for the internal case as the
  98          * caller may re-enable the same pgmap.
  99          */
 100         if (pgmap->ref == &pgmap->internal_ref)
 101                 pgmap->ref = NULL;
 102 }
 103
 104 void memunmap_pages(struct dev_pagemap *pgmap)
 105 {
 106         struct resource *res = &pgmap->res;
 107         struct page *first_page;
 108         unsigned long pfn;
 109         int nid;
 110
 111         dev_pagemap_kill(pgmap);
 112         for_each_device_pfn(pfn, pgmap)
 113                 put_page(pfn_to_page(pfn));
 114         dev_pagemap_cleanup(pgmap);
 115
 116         /* make sure to access a memmap that was actually initialized */
 117         first_page = pfn_to_page(pfn_first(pgmap));
 118
 119         /* pages are dead and unused, undo the arch mapping */
 120         nid = page_to_nid(first_page);
 121
 122         mem_hotplug_begin();
 123         remove_pfn_range_from_zone(page_zone(first_page), PHYS_PFN(res->start),
 124                                    PHYS_PFN(resource_size(res)));
 125         if (pgmap->type == MEMORY_DEVICE_PRIVATE) {
 126                 __remove_pages(PHYS_PFN(res->start),
 127                                PHYS_PFN(resource_size(res)), NULL);
 128         } else {
 129                 arch_remove_memory(nid, res->start, resource_size(res),
 130                                 pgmap_altmap(pgmap));
 131                 kasan_remove_zero_shadow(__va(res->start), resource_size(res));
 132         }
 133         mem_hotplug_done();
 134
 135         untrack_pfn(NULL, PHYS_PFN(res->start), resource_size(res));
 136         pgmap_array_delete(res);
 137         WARN_ONCE(pgmap->altmap.alloc, "failed to free all reserved pages\n");
 138         devmap_managed_enable_put();
 139 }
 140 EXPORT_SYMBOL_GPL(memunmap_pages);
 141
 142 static void devm_memremap_pages_release(void *data)
 143 {
 144         memunmap_pages(data);
 145 }
 146
 147 static void dev_pagemap_percpu_release(struct percpu_ref *ref)
 148 {
 149         struct dev_pagemap *pgmap =
 150                 container_of(ref, struct dev_pagemap, internal_ref);
 151
 152         complete(&pgmap->done);
 153 }
 154
 155 /*
 156  * Not device managed version of dev_memremap_pages, undone by
 157  * memunmap_pages().  Please use dev_memremap_pages if you have a struct
 158  * device available.
 159  */
 160 void *memremap_pages(struct dev_pagemap *pgmap, int nid)
 161 {
 162         struct resource *res = &pgmap->res;
 163         struct dev_pagemap *conflict_pgmap;
 164         struct mhp_restrictions restrictions = {
 165                 /*
 166                  * We do not want any optional features only our own memmap
 167                  */
 168                 .altmap = pgmap_altmap(pgmap),
 169         };
 170         pgprot_t pgprot = PAGE_KERNEL;
 171         int error, is_ram;
 172         bool need_devmap_managed = true;
 173
 174         switch (pgmap->type) {
 175         case MEMORY_DEVICE_PRIVATE:
 176                 if (!IS_ENABLED(CONFIG_DEVICE_PRIVATE)) {
 177                         WARN(1, "Device private memory not supported\n");
 178                         return ERR_PTR(-EINVAL);
 179                 }
 180                 if (!pgmap->ops || !pgmap->ops->migrate_to_ram) {
 181                         WARN(1, "Missing migrate_to_ram method\n");
 182                         return ERR_PTR(-EINVAL);
 183                 }
 184                 break;
 185         case MEMORY_DEVICE_FS_DAX:
 186                 if (!IS_ENABLED(CONFIG_ZONE_DEVICE) ||
 187                     IS_ENABLED(CONFIG_FS_DAX_LIMITED)) {
 188                         WARN(1, "File system DAX not supported\n");
 189                         return ERR_PTR(-EINVAL);
 190                 }
 191                 break;
 192         case MEMORY_DEVICE_DEVDAX:
 193         case MEMORY_DEVICE_PCI_P2PDMA:
 194                 need_devmap_managed = false;
 195                 break;
 196         default:
 197                 WARN(1, "Invalid pgmap type %d\n", pgmap->type);
 198                 break;
 199         }
 200
 201         if (!pgmap->ref) {
 202                 if (pgmap->ops && (pgmap->ops->kill || pgmap->ops->cleanup))
 203                         return ERR_PTR(-EINVAL);
 204
 205                 init_completion(&pgmap->done);
 206                 error = percpu_ref_init(&pgmap->internal_ref,
 207                                 dev_pagemap_percpu_release, 0, GFP_KERNEL);
 208                 if (error)
 209                         return ERR_PTR(error);
 210                 pgmap->ref = &pgmap->internal_ref;
 211         } else {
 212                 if (!pgmap->ops || !pgmap->ops->kill || !pgmap->ops->cleanup) {
 213                         WARN(1, "Missing reference count teardown definition\n");
 214                         return ERR_PTR(-EINVAL);
 215                 }
 216         }
 217
 218         if (need_devmap_managed) {
 219                 error = devmap_managed_enable_get(pgmap);
 220                 if (error)
 221                         return ERR_PTR(error);
 222         }
 223
 224         conflict_pgmap = get_dev_pagemap(PHYS_PFN(res->start), NULL);
 225         if (conflict_pgmap) {
 226                 WARN(1, "Conflicting mapping in same section\n");
 227                 put_dev_pagemap(conflict_pgmap);
 228                 error = -ENOMEM;
 229                 goto err_array;
 230         }
 231
 232         conflict_pgmap = get_dev_pagemap(PHYS_PFN(res->end), NULL);
 233         if (conflict_pgmap) {
 234                 WARN(1, "Conflicting mapping in same section\n");
 235                 put_dev_pagemap(conflict_pgmap);
 236                 error = -ENOMEM;
 237                 goto err_array;
 238         }
 239
 240         is_ram = region_intersects(res->start, resource_size(res),
 241                 IORESOURCE_SYSTEM_RAM, IORES_DESC_NONE);
 242
 243         if (is_ram != REGION_DISJOINT) {
 244                 WARN_ONCE(1, "%s attempted on %s region %pr\n", __func__,
 245                                 is_ram == REGION_MIXED ? "mixed" : "ram", res);
 246                 error = -ENXIO;
 247                 goto err_array;
 248         }
 249
 250         error = xa_err(xa_store_range(&pgmap_array, PHYS_PFN(res->start),
 251                                 PHYS_PFN(res->end), pgmap, GFP_KERNEL));
 252         if (error)
 253                 goto err_array;
 254
 255         if (nid < 0)
 256                 nid = numa_mem_id();
 257
 258         error = track_pfn_remap(NULL, &pgprot, PHYS_PFN(res->start), 0,
 259                         resource_size(res));
 260         if (error)
 261                 goto err_pfn_remap;
 262
 263         mem_hotplug_begin();
 264
 265         /*
 266          * For device private memory we call add_pages() as we only need to
 267          * allocate and initialize struct page for the device memory. More-
 268          * over the device memory is un-accessible thus we do not want to
 269          * create a linear mapping for the memory like arch_add_memory()
 270          * would do.
 271          *
 272          * For all other device memory types, which are accessible by
 273          * the CPU, we do want the linear mapping and thus use
 274          * arch_add_memory().
 275          */
 276         if (pgmap->type == MEMORY_DEVICE_PRIVATE) {
 277                 error = add_pages(nid, PHYS_PFN(res->start),
 278                                 PHYS_PFN(resource_size(res)), &restrictions);
 279         } else {
 280                 error = kasan_add_zero_shadow(__va(res->start), resource_size(res));
 281                 if (error) {
 282                         mem_hotplug_done();
 283                         goto err_kasan;
 284                 }
 285
 286                 error = arch_add_memory(nid, res->start, resource_size(res),
 287                                         &restrictions);
 288         }
 289
 290         if (!error) {
 291                 struct zone *zone;
 292
 293                 zone = &NODE_DATA(nid)->node_zones[ZONE_DEVICE];
 294                 move_pfn_range_to_zone(zone, PHYS_PFN(res->start),
 295                                 PHYS_PFN(resource_size(res)), restrictions.altmap);
 296         }
 297
 298         mem_hotplug_done();
 299         if (error)
 300                 goto err_add_memory;
 301
 302         /*
 303          * Initialization of the pages has been deferred until now in order
 304          * to allow us to do the work while not holding the hotplug lock.
 305          */
 306         memmap_init_zone_device(&NODE_DATA(nid)->node_zones[ZONE_DEVICE],
 307                                 PHYS_PFN(res->start),
 308                                 PHYS_PFN(resource_size(res)), pgmap);
 309         percpu_ref_get_many(pgmap->ref, pfn_end(pgmap) - pfn_first(pgmap));
 310         return __va(res->start);
 311
 312  err_add_memory:
 313         kasan_remove_zero_shadow(__va(res->start), resource_size(res));
 314  err_kasan:
 315         untrack_pfn(NULL, PHYS_PFN(res->start), resource_size(res));
 316  err_pfn_remap:
 317         pgmap_array_delete(res);
 318  err_array:
 319         dev_pagemap_kill(pgmap);
 320         dev_pagemap_cleanup(pgmap);
 321         devmap_managed_enable_put();
 322         return ERR_PTR(error);
 323 }
 324 EXPORT_SYMBOL_GPL(memremap_pages);
 325
 326 /**
 327  * devm_memremap_pages - remap and provide memmap backing for the given resource
 328  * @dev: hosting device for @res
 329  * @pgmap: pointer to a struct dev_pagemap
 330  *
 331  * Notes:
 332  * 1/ At a minimum the res and type members of @pgmap must be initialized
 333  *    by the caller before passing it to this function
 334  *
 335  * 2/ The altmap field may optionally be initialized, in which case
 336  *    PGMAP_ALTMAP_VALID must be set in pgmap->flags.
 337  *
 338  * 3/ The ref field may optionally be provided, in which pgmap->ref must be
 339  *    'live' on entry and will be killed and reaped at
 340  *    devm_memremap_pages_release() time, or if this routine fails.
 341  *
 342  * 4/ res is expected to be a host memory range that could feasibly be
 343  *    treated as a "System RAM" range, i.e. not a device mmio range, but
 344  *    this is not enforced.
 345  */
 346 void *devm_memremap_pages(struct device *dev, struct dev_pagemap *pgmap)
 347 {
 348         int error;
 349         void *ret;
 350
 351         ret = memremap_pages(pgmap, dev_to_node(dev));
 352         if (IS_ERR(ret))
 353                 return ret;
 354
 355         error = devm_add_action_or_reset(dev, devm_memremap_pages_release,
 356                         pgmap);
 357         if (error)
 358                 return ERR_PTR(error);
 359         return ret;
 360 }
 361 EXPORT_SYMBOL_GPL(devm_memremap_pages);
 362
 363 void devm_memunmap_pages(struct device *dev, struct dev_pagemap *pgmap)
 364 {
 365         devm_release_action(dev, devm_memremap_pages_release, pgmap);
 366 }
 367 EXPORT_SYMBOL_GPL(devm_memunmap_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         if (altmap)
 373                 return altmap->reserve + altmap->free;
 374         return 0;
 375 }
 376
 377 void vmem_altmap_free(struct vmem_altmap *altmap, unsigned long nr_pfns)
 378 {
 379         altmap->alloc -= nr_pfns;
 380 }
 381
 382 /**
 383  * get_dev_pagemap() - take a new live reference on the dev_pagemap for @pfn
 384  * @pfn: page frame number to lookup page_map
 385  * @pgmap: optional known pgmap that already has a reference
 386  *
 387  * If @pgmap is non-NULL and covers @pfn it will be returned as-is.  If @pgmap
 388  * is non-NULL but does not cover @pfn the reference to it will be released.
 389  */
 390 struct dev_pagemap *get_dev_pagemap(unsigned long pfn,
 391                 struct dev_pagemap *pgmap)
 392 {
 393         resource_size_t phys = PFN_PHYS(pfn);
 394
 395         /*
 396          * In the cached case we're already holding a live reference.
 397          */
 398         if (pgmap) {
 399                 if (phys >= pgmap->res.start && phys <= pgmap->res.end)
 400                         return pgmap;
 401                 put_dev_pagemap(pgmap);
 402         }
 403
 404         /* fall back to slow path lookup */
 405         rcu_read_lock();
 406         pgmap = xa_load(&pgmap_array, PHYS_PFN(phys));
 407         if (pgmap && !percpu_ref_tryget_live(pgmap->ref))
 408                 pgmap = NULL;
 409         rcu_read_unlock();
 410
 411         return pgmap;
 412 }
 413 EXPORT_SYMBOL_GPL(get_dev_pagemap);
 414
 415 #ifdef CONFIG_DEV_PAGEMAP_OPS
 416 void free_devmap_managed_page(struct page *page)
 417 {
 418         /* notify page idle for dax */
 419         if (!is_device_private_page(page)) {
 420                 wake_up_var(&page->_refcount);
 421                 return;
 422         }
 423
 424         /* Clear Active bit in case of parallel mark_page_accessed */
 425         __ClearPageActive(page);
 426         __ClearPageWaiters(page);
 427
 428         mem_cgroup_uncharge(page);
 429
 430         /*
 431          * When a device_private page is freed, the page->mapping field
 432          * may still contain a (stale) mapping value. For example, the
 433          * lower bits of page->mapping may still identify the page as an
 434          * anonymous page. Ultimately, this entire field is just stale
 435          * and wrong, and it will cause errors if not cleared.  One
 436          * example is:
 437          *
 438          *  migrate_vma_pages()
 439          *    migrate_vma_insert_page()
 440          *      page_add_new_anon_rmap()
 441          *        __page_set_anon_rmap()
 442          *          ...checks page->mapping, via PageAnon(page) call,
 443          *            and incorrectly concludes that the page is an
 444          *            anonymous page. Therefore, it incorrectly,
 445          *            silently fails to set up the new anon rmap.
 446          *
 447          * For other types of ZONE_DEVICE pages, migration is either
 448          * handled differently or not done at all, so there is no need
 449          * to clear page->mapping.
 450          */
 451         page->mapping = NULL;
 452         page->pgmap->ops->page_free(page);
 453 }
 454 #endif /* CONFIG_DEV_PAGEMAP_OPS */