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->ops || !pgmap->ops->page_free) {
  31                 WARN(1, "Missing page_free method\n");
  32                 return -EINVAL;
  33         }
  34
  35         if (atomic_inc_return(&devmap_managed_enable) == 1)
  36                 static_branch_enable(&devmap_managed_key);
  37         return 0;
  38 }
  39 #else
  40 static int devmap_managed_enable_get(struct dev_pagemap *pgmap)
  41 {
  42         return -EINVAL;
  43 }
  44 static void devmap_managed_enable_put(void)
  45 {
  46 }
  47 #endif /* CONFIG_DEV_PAGEMAP_OPS */
  48
  49 static void pgmap_array_delete(struct resource *res)
  50 {
  51         xa_store_range(&pgmap_array, PHYS_PFN(res->start), PHYS_PFN(res->end),
  52                         NULL, GFP_KERNEL);
  53         synchronize_rcu();
  54 }
  55
  56 static unsigned long pfn_first(struct dev_pagemap *pgmap)
  57 {
  58         return PHYS_PFN(pgmap->res.start) +
  59                 vmem_altmap_offset(pgmap_altmap(pgmap));
  60 }
  61
  62 static unsigned long pfn_end(struct dev_pagemap *pgmap)
  63 {
  64         const struct resource *res = &pgmap->res;
  65
  66         return (res->start + resource_size(res)) >> PAGE_SHIFT;
  67 }
  68
  69 static unsigned long pfn_next(unsigned long pfn)
  70 {
  71         if (pfn % 1024 == 0)
  72                 cond_resched();
  73         return pfn + 1;
  74 }
  75
  76 #define for_each_device_pfn(pfn, map) \
  77         for (pfn = pfn_first(map); pfn < pfn_end(map); pfn = pfn_next(pfn))
  78
  79 static void dev_pagemap_kill(struct dev_pagemap *pgmap)
  80 {
  81         if (pgmap->ops && pgmap->ops->kill)
  82                 pgmap->ops->kill(pgmap);
  83         else
  84                 percpu_ref_kill(pgmap->ref);
  85 }
  86
  87 static void dev_pagemap_cleanup(struct dev_pagemap *pgmap)
  88 {
  89         if (pgmap->ops && pgmap->ops->cleanup) {
  90                 pgmap->ops->cleanup(pgmap);
  91         } else {
  92                 wait_for_completion(&pgmap->done);
  93                 percpu_ref_exit(pgmap->ref);
  94         }
  95         /*
  96          * Undo the pgmap ref assignment for the internal case as the
  97          * caller may re-enable the same pgmap.
  98          */
  99         if (pgmap->ref == &pgmap->internal_ref)
 100                 pgmap->ref = NULL;
 101 }
 102
 103 void memunmap_pages(struct dev_pagemap *pgmap)
 104 {
 105         struct resource *res = &pgmap->res;
 106         struct page *first_page;
 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         /* make sure to access a memmap that was actually initialized */
 116         first_page = pfn_to_page(pfn_first(pgmap));
 117
 118         /* pages are dead and unused, undo the arch mapping */
 119         nid = page_to_nid(first_page);
 120
 121         mem_hotplug_begin();
 122         if (pgmap->type == MEMORY_DEVICE_PRIVATE) {
 123                 __remove_pages(PHYS_PFN(res->start),
 124                                PHYS_PFN(resource_size(res)), NULL);
 125         } else {
 126                 arch_remove_memory(nid, res->start, resource_size(res),
 127                                 pgmap_altmap(pgmap));
 128                 kasan_remove_zero_shadow(__va(res->start), resource_size(res));
 129         }
 130         mem_hotplug_done();
 131
 132         untrack_pfn(NULL, PHYS_PFN(res->start), resource_size(res));
 133         pgmap_array_delete(res);
 134         WARN_ONCE(pgmap->altmap.alloc, "failed to free all reserved pages\n");
 135         devmap_managed_enable_put();
 136 }
 137 EXPORT_SYMBOL_GPL(memunmap_pages);
 138
 139 static void devm_memremap_pages_release(void *data)
 140 {
 141         memunmap_pages(data);
 142 }
 143
 144 static void dev_pagemap_percpu_release(struct percpu_ref *ref)
 145 {
 146         struct dev_pagemap *pgmap =
 147                 container_of(ref, struct dev_pagemap, internal_ref);
 148
 149         complete(&pgmap->done);
 150 }
 151
 152 /*
 153  * Not device managed version of dev_memremap_pages, undone by
 154  * memunmap_pages().  Please use dev_memremap_pages if you have a struct
 155  * device available.
 156  */
 157 void *memremap_pages(struct dev_pagemap *pgmap, int nid)
 158 {
 159         struct resource *res = &pgmap->res;
 160         struct dev_pagemap *conflict_pgmap;
 161         struct mhp_restrictions restrictions = {
 162                 /*
 163                  * We do not want any optional features only our own memmap
 164                  */
 165                 .altmap = pgmap_altmap(pgmap),
 166         };
 167         pgprot_t pgprot = PAGE_KERNEL;
 168         int error, is_ram;
 169         bool need_devmap_managed = true;
 170
 171         switch (pgmap->type) {
 172         case MEMORY_DEVICE_PRIVATE:
 173                 if (!IS_ENABLED(CONFIG_DEVICE_PRIVATE)) {
 174                         WARN(1, "Device private memory not supported\n");
 175                         return ERR_PTR(-EINVAL);
 176                 }
 177                 if (!pgmap->ops || !pgmap->ops->migrate_to_ram) {
 178                         WARN(1, "Missing migrate_to_ram method\n");
 179                         return ERR_PTR(-EINVAL);
 180                 }
 181                 break;
 182         case MEMORY_DEVICE_FS_DAX:
 183                 if (!IS_ENABLED(CONFIG_ZONE_DEVICE) ||
 184                     IS_ENABLED(CONFIG_FS_DAX_LIMITED)) {
 185                         WARN(1, "File system DAX not supported\n");
 186                         return ERR_PTR(-EINVAL);
 187                 }
 188                 break;
 189         case MEMORY_DEVICE_DEVDAX:
 190         case MEMORY_DEVICE_PCI_P2PDMA:
 191                 need_devmap_managed = false;
 192                 break;
 193         default:
 194                 WARN(1, "Invalid pgmap type %d\n", pgmap->type);
 195                 break;
 196         }
 197
 198         if (!pgmap->ref) {
 199                 if (pgmap->ops && (pgmap->ops->kill || pgmap->ops->cleanup))
 200                         return ERR_PTR(-EINVAL);
 201
 202                 init_completion(&pgmap->done);
 203                 error = percpu_ref_init(&pgmap->internal_ref,
 204                                 dev_pagemap_percpu_release, 0, GFP_KERNEL);
 205                 if (error)
 206                         return ERR_PTR(error);
 207                 pgmap->ref = &pgmap->internal_ref;
 208         } else {
 209                 if (!pgmap->ops || !pgmap->ops->kill || !pgmap->ops->cleanup) {
 210                         WARN(1, "Missing reference count teardown definition\n");
 211                         return ERR_PTR(-EINVAL);
 212                 }
 213         }
 214
 215         if (need_devmap_managed) {
 216                 error = devmap_managed_enable_get(pgmap);
 217                 if (error)
 218                         return ERR_PTR(error);
 219         }
 220
 221         conflict_pgmap = get_dev_pagemap(PHYS_PFN(res->start), NULL);
 222         if (conflict_pgmap) {
 223                 WARN(1, "Conflicting mapping in same section\n");
 224                 put_dev_pagemap(conflict_pgmap);
 225                 error = -ENOMEM;
 226                 goto err_array;
 227         }
 228
 229         conflict_pgmap = get_dev_pagemap(PHYS_PFN(res->end), NULL);
 230         if (conflict_pgmap) {
 231                 WARN(1, "Conflicting mapping in same section\n");
 232                 put_dev_pagemap(conflict_pgmap);
 233                 error = -ENOMEM;
 234                 goto err_array;
 235         }
 236
 237         is_ram = region_intersects(res->start, resource_size(res),
 238                 IORESOURCE_SYSTEM_RAM, IORES_DESC_NONE);
 239
 240         if (is_ram != REGION_DISJOINT) {
 241                 WARN_ONCE(1, "%s attempted on %s region %pr\n", __func__,
 242                                 is_ram == REGION_MIXED ? "mixed" : "ram", res);
 243                 error = -ENXIO;
 244                 goto err_array;
 245         }
 246
 247         error = xa_err(xa_store_range(&pgmap_array, PHYS_PFN(res->start),
 248                                 PHYS_PFN(res->end), pgmap, GFP_KERNEL));
 249         if (error)
 250                 goto err_array;
 251
 252         if (nid < 0)
 253                 nid = numa_mem_id();
 254
 255         error = track_pfn_remap(NULL, &pgprot, PHYS_PFN(res->start), 0,
 256                         resource_size(res));
 257         if (error)
 258                 goto err_pfn_remap;
 259
 260         mem_hotplug_begin();
 261
 262         /*
 263          * For device private memory we call add_pages() as we only need to
 264          * allocate and initialize struct page for the device memory. More-
 265          * over the device memory is un-accessible thus we do not want to
 266          * create a linear mapping for the memory like arch_add_memory()
 267          * would do.
 268          *
 269          * For all other device memory types, which are accessible by
 270          * the CPU, we do want the linear mapping and thus use
 271          * arch_add_memory().
 272          */
 273         if (pgmap->type == MEMORY_DEVICE_PRIVATE) {
 274                 error = add_pages(nid, PHYS_PFN(res->start),
 275                                 PHYS_PFN(resource_size(res)), &restrictions);
 276         } else {
 277                 error = kasan_add_zero_shadow(__va(res->start), resource_size(res));
 278                 if (error) {
 279                         mem_hotplug_done();
 280                         goto err_kasan;
 281                 }
 282
 283                 error = arch_add_memory(nid, res->start, resource_size(res),
 284                                         &restrictions);
 285         }
 286
 287         if (!error) {
 288                 struct zone *zone;
 289
 290                 zone = &NODE_DATA(nid)->node_zones[ZONE_DEVICE];
 291                 move_pfn_range_to_zone(zone, PHYS_PFN(res->start),
 292                                 PHYS_PFN(resource_size(res)), restrictions.altmap);
 293         }
 294
 295         mem_hotplug_done();
 296         if (error)
 297                 goto err_add_memory;
 298
 299         /*
 300          * Initialization of the pages has been deferred until now in order
 301          * to allow us to do the work while not holding the hotplug lock.
 302          */
 303         memmap_init_zone_device(&NODE_DATA(nid)->node_zones[ZONE_DEVICE],
 304                                 PHYS_PFN(res->start),
 305                                 PHYS_PFN(resource_size(res)), pgmap);
 306         percpu_ref_get_many(pgmap->ref, pfn_end(pgmap) - pfn_first(pgmap));
 307         return __va(res->start);
 308
 309  err_add_memory:
 310         kasan_remove_zero_shadow(__va(res->start), resource_size(res));
 311  err_kasan:
 312         untrack_pfn(NULL, PHYS_PFN(res->start), resource_size(res));
 313  err_pfn_remap:
 314         pgmap_array_delete(res);
 315  err_array:
 316         dev_pagemap_kill(pgmap);
 317         dev_pagemap_cleanup(pgmap);
 318         devmap_managed_enable_put();
 319         return ERR_PTR(error);
 320 }
 321 EXPORT_SYMBOL_GPL(memremap_pages);
 322
 323 /**
 324  * devm_memremap_pages - remap and provide memmap backing for the given resource
 325  * @dev: hosting device for @res
 326  * @pgmap: pointer to a struct dev_pagemap
 327  *
 328  * Notes:
 329  * 1/ At a minimum the res and type members of @pgmap must be initialized
 330  *    by the caller before passing it to this function
 331  *
 332  * 2/ The altmap field may optionally be initialized, in which case
 333  *    PGMAP_ALTMAP_VALID must be set in pgmap->flags.
 334  *
 335  * 3/ The ref field may optionally be provided, in which pgmap->ref must be
 336  *    'live' on entry and will be killed and reaped at
 337  *    devm_memremap_pages_release() time, or if this routine fails.
 338  *
 339  * 4/ res is expected to be a host memory range that could feasibly be
 340  *    treated as a "System RAM" range, i.e. not a device mmio range, but
 341  *    this is not enforced.
 342  */
 343 void *devm_memremap_pages(struct device *dev, struct dev_pagemap *pgmap)
 344 {
 345         int error;
 346         void *ret;
 347
 348         ret = memremap_pages(pgmap, dev_to_node(dev));
 349         if (IS_ERR(ret))
 350                 return ret;
 351
 352         error = devm_add_action_or_reset(dev, devm_memremap_pages_release,
 353                         pgmap);
 354         if (error)
 355                 return ERR_PTR(error);
 356         return ret;
 357 }
 358 EXPORT_SYMBOL_GPL(devm_memremap_pages);
 359
 360 void devm_memunmap_pages(struct device *dev, struct dev_pagemap *pgmap)
 361 {
 362         devm_release_action(dev, devm_memremap_pages_release, pgmap);
 363 }
 364 EXPORT_SYMBOL_GPL(devm_memunmap_pages);
 365
 366 unsigned long vmem_altmap_offset(struct vmem_altmap *altmap)
 367 {
 368         /* number of pfns from base where pfn_to_page() is valid */
 369         if (altmap)
 370                 return altmap->reserve + altmap->free;
 371         return 0;
 372 }
 373
 374 void vmem_altmap_free(struct vmem_altmap *altmap, unsigned long nr_pfns)
 375 {
 376         altmap->alloc -= nr_pfns;
 377 }
 378
 379 /**
 380  * get_dev_pagemap() - take a new live reference on the dev_pagemap for @pfn
 381  * @pfn: page frame number to lookup page_map
 382  * @pgmap: optional known pgmap that already has a reference
 383  *
 384  * If @pgmap is non-NULL and covers @pfn it will be returned as-is.  If @pgmap
 385  * is non-NULL but does not cover @pfn the reference to it will be released.
 386  */
 387 struct dev_pagemap *get_dev_pagemap(unsigned long pfn,
 388                 struct dev_pagemap *pgmap)
 389 {
 390         resource_size_t phys = PFN_PHYS(pfn);
 391
 392         /*
 393          * In the cached case we're already holding a live reference.
 394          */
 395         if (pgmap) {
 396                 if (phys >= pgmap->res.start && phys <= pgmap->res.end)
 397                         return pgmap;
 398                 put_dev_pagemap(pgmap);
 399         }
 400
 401         /* fall back to slow path lookup */
 402         rcu_read_lock();
 403         pgmap = xa_load(&pgmap_array, PHYS_PFN(phys));
 404         if (pgmap && !percpu_ref_tryget_live(pgmap->ref))
 405                 pgmap = NULL;
 406         rcu_read_unlock();
 407
 408         return pgmap;
 409 }
 410 EXPORT_SYMBOL_GPL(get_dev_pagemap);
 411
 412 #ifdef CONFIG_DEV_PAGEMAP_OPS
 413 void __put_devmap_managed_page(struct page *page)
 414 {
 415         int count = page_ref_dec_return(page);
 416
 417         /*
 418          * If refcount is 1 then page is freed and refcount is stable as nobody
 419          * holds a reference on the page.
 420          */
 421         if (count == 1) {
 422                 /* Clear Active bit in case of parallel mark_page_accessed */
 423                 __ClearPageActive(page);
 424                 __ClearPageWaiters(page);
 425
 426                 mem_cgroup_uncharge(page);
 427
 428                 /*
 429                  * When a device_private page is freed, the page->mapping field
 430                  * may still contain a (stale) mapping value. For example, the
 431                  * lower bits of page->mapping may still identify the page as
 432                  * an anonymous page. Ultimately, this entire field is just
 433                  * stale and wrong, and it will cause errors if not cleared.
 434                  * One example is:
 435                  *
 436                  *  migrate_vma_pages()
 437                  *    migrate_vma_insert_page()
 438                  *      page_add_new_anon_rmap()
 439                  *        __page_set_anon_rmap()
 440                  *          ...checks page->mapping, via PageAnon(page) call,
 441                  *            and incorrectly concludes that the page is an
 442                  *            anonymous page. Therefore, it incorrectly,
 443                  *            silently fails to set up the new anon rmap.
 444                  *
 445                  * For other types of ZONE_DEVICE pages, migration is either
 446                  * handled differently or not done at all, so there is no need
 447                  * to clear page->mapping.
 448                  */
 449                 if (is_device_private_page(page))
 450                         page->mapping = NULL;
 451
 452                 page->pgmap->ops->page_free(page);
 453         } else if (!count)
 454                 __put_page(page);
 455 }
 456 EXPORT_SYMBOL(__put_devmap_managed_page);
 457 #endif /* CONFIG_DEV_PAGEMAP_OPS */