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/mmzone.h>
  11 #include <linux/swapops.h>
  12 #include <linux/types.h>
  13 #include <linux/wait_bit.h>
  14 #include <linux/xarray.h>
  15
  16 static DEFINE_XARRAY(pgmap_array);
  17
  18 /*
  19  * The memremap() and memremap_pages() interfaces are alternately used
  20  * to map persistent memory namespaces. These interfaces place different
  21  * constraints on the alignment and size of the mapping (namespace).
  22  * memremap() can map individual PAGE_SIZE pages. memremap_pages() can
  23  * only map subsections (2MB), and at least one architecture (PowerPC)
  24  * the minimum mapping granularity of memremap_pages() is 16MB.
  25  *
  26  * The role of memremap_compat_align() is to communicate the minimum
  27  * arch supported alignment of a namespace such that it can freely
  28  * switch modes without violating the arch constraint. Namely, do not
  29  * allow a namespace to be PAGE_SIZE aligned since that namespace may be
  30  * reconfigured into a mode that requires SUBSECTION_SIZE alignment.
  31  */
  32 #ifndef CONFIG_ARCH_HAS_MEMREMAP_COMPAT_ALIGN
  33 unsigned long memremap_compat_align(void)
  34 {
  35         return SUBSECTION_SIZE;
  36 }
  37 EXPORT_SYMBOL_GPL(memremap_compat_align);
  38 #endif
  39
  40 #ifdef CONFIG_DEV_PAGEMAP_OPS
  41 DEFINE_STATIC_KEY_FALSE(devmap_managed_key);
  42 EXPORT_SYMBOL(devmap_managed_key);
  43
  44 static void devmap_managed_enable_put(struct dev_pagemap *pgmap)
  45 {
  46         if (pgmap->type == MEMORY_DEVICE_PRIVATE ||
  47             pgmap->type == MEMORY_DEVICE_FS_DAX)
  48                 static_branch_dec(&devmap_managed_key);
  49 }
  50
  51 static void devmap_managed_enable_get(struct dev_pagemap *pgmap)
  52 {
  53         if (pgmap->type == MEMORY_DEVICE_PRIVATE ||
  54             pgmap->type == MEMORY_DEVICE_FS_DAX)
  55                 static_branch_inc(&devmap_managed_key);
  56 }
  57 #else
  58 static void devmap_managed_enable_get(struct dev_pagemap *pgmap)
  59 {
  60 }
  61 static void devmap_managed_enable_put(struct dev_pagemap *pgmap)
  62 {
  63 }
  64 #endif /* CONFIG_DEV_PAGEMAP_OPS */
  65
  66 static void pgmap_array_delete(struct range *range)
  67 {
  68         xa_store_range(&pgmap_array, PHYS_PFN(range->start), PHYS_PFN(range->end),
  69                         NULL, GFP_KERNEL);
  70         synchronize_rcu();
  71 }
  72
  73 static unsigned long pfn_first(struct dev_pagemap *pgmap, int range_id)
  74 {
  75         struct range *range = &pgmap->ranges[range_id];
  76         unsigned long pfn = PHYS_PFN(range->start);
  77
  78         if (range_id)
  79                 return pfn;
  80         return pfn + vmem_altmap_offset(pgmap_altmap(pgmap));
  81 }
  82
  83 static unsigned long pfn_end(struct dev_pagemap *pgmap, int range_id)
  84 {
  85         const struct range *range = &pgmap->ranges[range_id];
  86
  87         return (range->start + range_len(range)) >> PAGE_SHIFT;
  88 }
  89
  90 static unsigned long pfn_next(unsigned long pfn)
  91 {
  92         if (pfn % 1024 == 0)
  93                 cond_resched();
  94         return pfn + 1;
  95 }
  96
  97 #define for_each_device_pfn(pfn, map, i) \
  98         for (pfn = pfn_first(map, i); pfn < pfn_end(map, i); pfn = pfn_next(pfn))
  99
 100 static void dev_pagemap_kill(struct dev_pagemap *pgmap)
 101 {
 102         if (pgmap->ops && pgmap->ops->kill)
 103                 pgmap->ops->kill(pgmap);
 104         else
 105                 percpu_ref_kill(pgmap->ref);
 106 }
 107
 108 static void dev_pagemap_cleanup(struct dev_pagemap *pgmap)
 109 {
 110         if (pgmap->ops && pgmap->ops->cleanup) {
 111                 pgmap->ops->cleanup(pgmap);
 112         } else {
 113                 wait_for_completion(&pgmap->done);
 114                 percpu_ref_exit(pgmap->ref);
 115         }
 116         /*
 117          * Undo the pgmap ref assignment for the internal case as the
 118          * caller may re-enable the same pgmap.
 119          */
 120         if (pgmap->ref == &pgmap->internal_ref)
 121                 pgmap->ref = NULL;
 122 }
 123
 124 static void pageunmap_range(struct dev_pagemap *pgmap, int range_id)
 125 {
 126         struct range *range = &pgmap->ranges[range_id];
 127         struct page *first_page;
 128         int nid;
 129
 130         /* make sure to access a memmap that was actually initialized */
 131         first_page = pfn_to_page(pfn_first(pgmap, range_id));
 132
 133         /* pages are dead and unused, undo the arch mapping */
 134         nid = page_to_nid(first_page);
 135
 136         mem_hotplug_begin();
 137         remove_pfn_range_from_zone(page_zone(first_page), PHYS_PFN(range->start),
 138                                    PHYS_PFN(range_len(range)));
 139         if (pgmap->type == MEMORY_DEVICE_PRIVATE) {
 140                 __remove_pages(PHYS_PFN(range->start),
 141                                PHYS_PFN(range_len(range)), NULL);
 142         } else {
 143                 arch_remove_memory(nid, range->start, range_len(range),
 144                                 pgmap_altmap(pgmap));
 145                 kasan_remove_zero_shadow(__va(range->start), range_len(range));
 146         }
 147         mem_hotplug_done();
 148
 149         untrack_pfn(NULL, PHYS_PFN(range->start), range_len(range));
 150         pgmap_array_delete(range);
 151 }
 152
 153 void memunmap_pages(struct dev_pagemap *pgmap)
 154 {
 155         unsigned long pfn;
 156         int i;
 157
 158         dev_pagemap_kill(pgmap);
 159         for (i = 0; i < pgmap->nr_range; i++)
 160                 for_each_device_pfn(pfn, pgmap, i)
 161                         put_page(pfn_to_page(pfn));
 162         dev_pagemap_cleanup(pgmap);
 163
 164         for (i = 0; i < pgmap->nr_range; i++)
 165                 pageunmap_range(pgmap, i);
 166
 167         WARN_ONCE(pgmap->altmap.alloc, "failed to free all reserved pages\n");
 168         devmap_managed_enable_put(pgmap);
 169 }
 170 EXPORT_SYMBOL_GPL(memunmap_pages);
 171
 172 static void devm_memremap_pages_release(void *data)
 173 {
 174         memunmap_pages(data);
 175 }
 176
 177 static void dev_pagemap_percpu_release(struct percpu_ref *ref)
 178 {
 179         struct dev_pagemap *pgmap =
 180                 container_of(ref, struct dev_pagemap, internal_ref);
 181
 182         complete(&pgmap->done);
 183 }
 184
 185 static int pagemap_range(struct dev_pagemap *pgmap, struct mhp_params *params,
 186                 int range_id, int nid)
 187 {
 188         struct range *range = &pgmap->ranges[range_id];
 189         struct dev_pagemap *conflict_pgmap;
 190         int error, is_ram;
 191
 192         if (WARN_ONCE(pgmap_altmap(pgmap) && range_id > 0,
 193                                 "altmap not supported for multiple ranges\n"))
 194                 return -EINVAL;
 195
 196         conflict_pgmap = get_dev_pagemap(PHYS_PFN(range->start), NULL);
 197         if (conflict_pgmap) {
 198                 WARN(1, "Conflicting mapping in same section\n");
 199                 put_dev_pagemap(conflict_pgmap);
 200                 return -ENOMEM;
 201         }
 202
 203         conflict_pgmap = get_dev_pagemap(PHYS_PFN(range->end), NULL);
 204         if (conflict_pgmap) {
 205                 WARN(1, "Conflicting mapping in same section\n");
 206                 put_dev_pagemap(conflict_pgmap);
 207                 return -ENOMEM;
 208         }
 209
 210         is_ram = region_intersects(range->start, range_len(range),
 211                 IORESOURCE_SYSTEM_RAM, IORES_DESC_NONE);
 212
 213         if (is_ram != REGION_DISJOINT) {
 214                 WARN_ONCE(1, "attempted on %s region %#llx-%#llx\n",
 215                                 is_ram == REGION_MIXED ? "mixed" : "ram",
 216                                 range->start, range->end);
 217                 return -ENXIO;
 218         }
 219
 220         error = xa_err(xa_store_range(&pgmap_array, PHYS_PFN(range->start),
 221                                 PHYS_PFN(range->end), pgmap, GFP_KERNEL));
 222         if (error)
 223                 return error;
 224
 225         if (nid < 0)
 226                 nid = numa_mem_id();
 227
 228         error = track_pfn_remap(NULL, &params->pgprot, PHYS_PFN(range->start), 0,
 229                         range_len(range));
 230         if (error)
 231                 goto err_pfn_remap;
 232
 233         mem_hotplug_begin();
 234
 235         /*
 236          * For device private memory we call add_pages() as we only need to
 237          * allocate and initialize struct page for the device memory. More-
 238          * over the device memory is un-accessible thus we do not want to
 239          * create a linear mapping for the memory like arch_add_memory()
 240          * would do.
 241          *
 242          * For all other device memory types, which are accessible by
 243          * the CPU, we do want the linear mapping and thus use
 244          * arch_add_memory().
 245          */
 246         if (pgmap->type == MEMORY_DEVICE_PRIVATE) {
 247                 error = add_pages(nid, PHYS_PFN(range->start),
 248                                 PHYS_PFN(range_len(range)), params);
 249         } else {
 250                 error = kasan_add_zero_shadow(__va(range->start), range_len(range));
 251                 if (error) {
 252                         mem_hotplug_done();
 253                         goto err_kasan;
 254                 }
 255
 256                 error = arch_add_memory(nid, range->start, range_len(range),
 257                                         params);
 258         }
 259
 260         if (!error) {
 261                 struct zone *zone;
 262
 263                 zone = &NODE_DATA(nid)->node_zones[ZONE_DEVICE];
 264                 move_pfn_range_to_zone(zone, PHYS_PFN(range->start),
 265                                 PHYS_PFN(range_len(range)), params->altmap,
 266                                 MIGRATE_MOVABLE);
 267         }
 268
 269         mem_hotplug_done();
 270         if (error)
 271                 goto err_add_memory;
 272
 273         /*
 274          * Initialization of the pages has been deferred until now in order
 275          * to allow us to do the work while not holding the hotplug lock.
 276          */
 277         memmap_init_zone_device(&NODE_DATA(nid)->node_zones[ZONE_DEVICE],
 278                                 PHYS_PFN(range->start),
 279                                 PHYS_PFN(range_len(range)), pgmap);
 280         percpu_ref_get_many(pgmap->ref, pfn_end(pgmap, range_id)
 281                         - pfn_first(pgmap, range_id));
 282         return 0;
 283
 284 err_add_memory:
 285         kasan_remove_zero_shadow(__va(range->start), range_len(range));
 286 err_kasan:
 287         untrack_pfn(NULL, PHYS_PFN(range->start), range_len(range));
 288 err_pfn_remap:
 289         pgmap_array_delete(range);
 290         return error;
 291 }
 292
 293
 294 /*
 295  * Not device managed version of dev_memremap_pages, undone by
 296  * memunmap_pages().  Please use dev_memremap_pages if you have a struct
 297  * device available.
 298  */
 299 void *memremap_pages(struct dev_pagemap *pgmap, int nid)
 300 {
 301         struct mhp_params params = {
 302                 .altmap = pgmap_altmap(pgmap),
 303                 .pgprot = PAGE_KERNEL,
 304         };
 305         const int nr_range = pgmap->nr_range;
 306         int error, i;
 307
 308         if (WARN_ONCE(!nr_range, "nr_range must be specified\n"))
 309                 return ERR_PTR(-EINVAL);
 310
 311         switch (pgmap->type) {
 312         case MEMORY_DEVICE_PRIVATE:
 313                 if (!IS_ENABLED(CONFIG_DEVICE_PRIVATE)) {
 314                         WARN(1, "Device private memory not supported\n");
 315                         return ERR_PTR(-EINVAL);
 316                 }
 317                 if (!pgmap->ops || !pgmap->ops->migrate_to_ram) {
 318                         WARN(1, "Missing migrate_to_ram method\n");
 319                         return ERR_PTR(-EINVAL);
 320                 }
 321                 if (!pgmap->ops->page_free) {
 322                         WARN(1, "Missing page_free method\n");
 323                         return ERR_PTR(-EINVAL);
 324                 }
 325                 if (!pgmap->owner) {
 326                         WARN(1, "Missing owner\n");
 327                         return ERR_PTR(-EINVAL);
 328                 }
 329                 break;
 330         case MEMORY_DEVICE_FS_DAX:
 331                 if (!IS_ENABLED(CONFIG_ZONE_DEVICE) ||
 332                     IS_ENABLED(CONFIG_FS_DAX_LIMITED)) {
 333                         WARN(1, "File system DAX not supported\n");
 334                         return ERR_PTR(-EINVAL);
 335                 }
 336                 break;
 337         case MEMORY_DEVICE_GENERIC:
 338                 break;
 339         case MEMORY_DEVICE_PCI_P2PDMA:
 340                 params.pgprot = pgprot_noncached(params.pgprot);
 341                 break;
 342         default:
 343                 WARN(1, "Invalid pgmap type %d\n", pgmap->type);
 344                 break;
 345         }
 346
 347         if (!pgmap->ref) {
 348                 if (pgmap->ops && (pgmap->ops->kill || pgmap->ops->cleanup))
 349                         return ERR_PTR(-EINVAL);
 350
 351                 init_completion(&pgmap->done);
 352                 error = percpu_ref_init(&pgmap->internal_ref,
 353                                 dev_pagemap_percpu_release, 0, GFP_KERNEL);
 354                 if (error)
 355                         return ERR_PTR(error);
 356                 pgmap->ref = &pgmap->internal_ref;
 357         } else {
 358                 if (!pgmap->ops || !pgmap->ops->kill || !pgmap->ops->cleanup) {
 359                         WARN(1, "Missing reference count teardown definition\n");
 360                         return ERR_PTR(-EINVAL);
 361                 }
 362         }
 363
 364         devmap_managed_enable_get(pgmap);
 365
 366         /*
 367          * Clear the pgmap nr_range as it will be incremented for each
 368          * successfully processed range. This communicates how many
 369          * regions to unwind in the abort case.
 370          */
 371         pgmap->nr_range = 0;
 372         error = 0;
 373         for (i = 0; i < nr_range; i++) {
 374                 error = pagemap_range(pgmap, &params, i, nid);
 375                 if (error)
 376                         break;
 377                 pgmap->nr_range++;
 378         }
 379
 380         if (i < nr_range) {
 381                 memunmap_pages(pgmap);
 382                 pgmap->nr_range = nr_range;
 383                 return ERR_PTR(error);
 384         }
 385
 386         return __va(pgmap->ranges[0].start);
 387 }
 388 EXPORT_SYMBOL_GPL(memremap_pages);
 389
 390 /**
 391  * devm_memremap_pages - remap and provide memmap backing for the given resource
 392  * @dev: hosting device for @res
 393  * @pgmap: pointer to a struct dev_pagemap
 394  *
 395  * Notes:
 396  * 1/ At a minimum the res and type members of @pgmap must be initialized
 397  *    by the caller before passing it to this function
 398  *
 399  * 2/ The altmap field may optionally be initialized, in which case
 400  *    PGMAP_ALTMAP_VALID must be set in pgmap->flags.
 401  *
 402  * 3/ The ref field may optionally be provided, in which pgmap->ref must be
 403  *    'live' on entry and will be killed and reaped at
 404  *    devm_memremap_pages_release() time, or if this routine fails.
 405  *
 406  * 4/ range is expected to be a host memory range that could feasibly be
 407  *    treated as a "System RAM" range, i.e. not a device mmio range, but
 408  *    this is not enforced.
 409  */
 410 void *devm_memremap_pages(struct device *dev, struct dev_pagemap *pgmap)
 411 {
 412         int error;
 413         void *ret;
 414
 415         ret = memremap_pages(pgmap, dev_to_node(dev));
 416         if (IS_ERR(ret))
 417                 return ret;
 418
 419         error = devm_add_action_or_reset(dev, devm_memremap_pages_release,
 420                         pgmap);
 421         if (error)
 422                 return ERR_PTR(error);
 423         return ret;
 424 }
 425 EXPORT_SYMBOL_GPL(devm_memremap_pages);
 426
 427 void devm_memunmap_pages(struct device *dev, struct dev_pagemap *pgmap)
 428 {
 429         devm_release_action(dev, devm_memremap_pages_release, pgmap);
 430 }
 431 EXPORT_SYMBOL_GPL(devm_memunmap_pages);
 432
 433 unsigned long vmem_altmap_offset(struct vmem_altmap *altmap)
 434 {
 435         /* number of pfns from base where pfn_to_page() is valid */
 436         if (altmap)
 437                 return altmap->reserve + altmap->free;
 438         return 0;
 439 }
 440
 441 void vmem_altmap_free(struct vmem_altmap *altmap, unsigned long nr_pfns)
 442 {
 443         altmap->alloc -= nr_pfns;
 444 }
 445
 446 /**
 447  * get_dev_pagemap() - take a new live reference on the dev_pagemap for @pfn
 448  * @pfn: page frame number to lookup page_map
 449  * @pgmap: optional known pgmap that already has a reference
 450  *
 451  * If @pgmap is non-NULL and covers @pfn it will be returned as-is.  If @pgmap
 452  * is non-NULL but does not cover @pfn the reference to it will be released.
 453  */
 454 struct dev_pagemap *get_dev_pagemap(unsigned long pfn,
 455                 struct dev_pagemap *pgmap)
 456 {
 457         resource_size_t phys = PFN_PHYS(pfn);
 458
 459         /*
 460          * In the cached case we're already holding a live reference.
 461          */
 462         if (pgmap) {
 463                 if (phys >= pgmap->range.start && phys <= pgmap->range.end)
 464                         return pgmap;
 465                 put_dev_pagemap(pgmap);
 466         }
 467
 468         /* fall back to slow path lookup */
 469         rcu_read_lock();
 470         pgmap = xa_load(&pgmap_array, PHYS_PFN(phys));
 471         if (pgmap && !percpu_ref_tryget_live(pgmap->ref))
 472                 pgmap = NULL;
 473         rcu_read_unlock();
 474
 475         return pgmap;
 476 }
 477 EXPORT_SYMBOL_GPL(get_dev_pagemap);
 478
 479 #ifdef CONFIG_DEV_PAGEMAP_OPS
 480 void free_devmap_managed_page(struct page *page)
 481 {
 482         /* notify page idle for dax */
 483         if (!is_device_private_page(page)) {
 484                 wake_up_var(&page->_refcount);
 485                 return;
 486         }
 487
 488         __ClearPageWaiters(page);
 489
 490         mem_cgroup_uncharge(page);
 491
 492         /*
 493          * When a device_private page is freed, the page->mapping field
 494          * may still contain a (stale) mapping value. For example, the
 495          * lower bits of page->mapping may still identify the page as an
 496          * anonymous page. Ultimately, this entire field is just stale
 497          * and wrong, and it will cause errors if not cleared.  One
 498          * example is:
 499          *
 500          *  migrate_vma_pages()
 501          *    migrate_vma_insert_page()
 502          *      page_add_new_anon_rmap()
 503          *        __page_set_anon_rmap()
 504          *          ...checks page->mapping, via PageAnon(page) call,
 505          *            and incorrectly concludes that the page is an
 506          *            anonymous page. Therefore, it incorrectly,
 507          *            silently fails to set up the new anon rmap.
 508          *
 509          * For other types of ZONE_DEVICE pages, migration is either
 510          * handled differently or not done at all, so there is no need
 511          * to clear page->mapping.
 512          */
 513         page->mapping = NULL;
 514         page->pgmap->ops->page_free(page);
 515 }
 516 #endif /* CONFIG_DEV_PAGEMAP_OPS */