kernel/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
  96 static void devm_memremap_pages_release(void *data)
  97 {
  98         struct dev_pagemap *pgmap = data;
  99         struct device *dev = pgmap->dev;
 100         struct resource *res = &pgmap->res;
 101         unsigned long pfn;
 102         int nid;
 103
 104         dev_pagemap_kill(pgmap);
 105         for_each_device_pfn(pfn, pgmap)
 106                 put_page(pfn_to_page(pfn));
 107         dev_pagemap_cleanup(pgmap);
 108
 109         /* pages are dead and unused, undo the arch mapping */
 110         nid = page_to_nid(pfn_to_page(PHYS_PFN(res->start)));
 111
 112         mem_hotplug_begin();
 113         if (pgmap->type == MEMORY_DEVICE_PRIVATE) {
 114                 pfn = PHYS_PFN(res->start);
 115                 __remove_pages(page_zone(pfn_to_page(pfn)), pfn,
 116                                  PHYS_PFN(resource_size(res)), NULL);
 117         } else {
 118                 arch_remove_memory(nid, res->start, resource_size(res),
 119                                 pgmap_altmap(pgmap));
 120                 kasan_remove_zero_shadow(__va(res->start), resource_size(res));
 121         }
 122         mem_hotplug_done();
 123
 124         untrack_pfn(NULL, PHYS_PFN(res->start), resource_size(res));
 125         pgmap_array_delete(res);
 126         dev_WARN_ONCE(dev, pgmap->altmap.alloc,
 127                       "%s: failed to free all reserved pages\n", __func__);
 128 }
 129
 130 static void dev_pagemap_percpu_release(struct percpu_ref *ref)
 131 {
 132         struct dev_pagemap *pgmap =
 133                 container_of(ref, struct dev_pagemap, internal_ref);
 134
 135         complete(&pgmap->done);
 136 }
 137
 138 /**
 139  * devm_memremap_pages - remap and provide memmap backing for the given resource
 140  * @dev: hosting device for @res
 141  * @pgmap: pointer to a struct dev_pagemap
 142  *
 143  * Notes:
 144  * 1/ At a minimum the res and type members of @pgmap must be initialized
 145  *    by the caller before passing it to this function
 146  *
 147  * 2/ The altmap field may optionally be initialized, in which case
 148  *    PGMAP_ALTMAP_VALID must be set in pgmap->flags.
 149  *
 150  * 3/ The ref field may optionally be provided, in which pgmap->ref must be
 151  *    'live' on entry and will be killed and reaped at
 152  *    devm_memremap_pages_release() time, or if this routine fails.
 153  *
 154  * 4/ res is expected to be a host memory range that could feasibly be
 155  *    treated as a "System RAM" range, i.e. not a device mmio range, but
 156  *    this is not enforced.
 157  */
 158 void *devm_memremap_pages(struct device *dev, struct dev_pagemap *pgmap)
 159 {
 160         struct resource *res = &pgmap->res;
 161         struct dev_pagemap *conflict_pgmap;
 162         struct mhp_restrictions restrictions = {
 163                 /*
 164                  * We do not want any optional features only our own memmap
 165                  */
 166                 .altmap = pgmap_altmap(pgmap),
 167         };
 168         pgprot_t pgprot = PAGE_KERNEL;
 169         int error, nid, is_ram;
 170         bool need_devmap_managed = true;
 171
 172         switch (pgmap->type) {
 173         case MEMORY_DEVICE_PRIVATE:
 174                 if (!IS_ENABLED(CONFIG_DEVICE_PRIVATE)) {
 175                         WARN(1, "Device private memory not supported\n");
 176                         return ERR_PTR(-EINVAL);
 177                 }
 178                 if (!pgmap->ops || !pgmap->ops->migrate_to_ram) {
 179                         WARN(1, "Missing migrate_to_ram method\n");
 180                         return ERR_PTR(-EINVAL);
 181                 }
 182                 break;
 183         case MEMORY_DEVICE_FS_DAX:
 184                 if (!IS_ENABLED(CONFIG_ZONE_DEVICE) ||
 185                     IS_ENABLED(CONFIG_FS_DAX_LIMITED)) {
 186                         WARN(1, "File system DAX not supported\n");
 187                         return ERR_PTR(-EINVAL);
 188                 }
 189                 break;
 190         case MEMORY_DEVICE_DEVDAX:
 191         case MEMORY_DEVICE_PCI_P2PDMA:
 192                 need_devmap_managed = false;
 193                 break;
 194         default:
 195                 WARN(1, "Invalid pgmap type %d\n", pgmap->type);
 196                 break;
 197         }
 198
 199         if (!pgmap->ref) {
 200                 if (pgmap->ops && (pgmap->ops->kill || pgmap->ops->cleanup))
 201                         return ERR_PTR(-EINVAL);
 202
 203                 init_completion(&pgmap->done);
 204                 error = percpu_ref_init(&pgmap->internal_ref,
 205                                 dev_pagemap_percpu_release, 0, GFP_KERNEL);
 206                 if (error)
 207                         return ERR_PTR(error);
 208                 pgmap->ref = &pgmap->internal_ref;
 209         } else {
 210                 if (!pgmap->ops || !pgmap->ops->kill || !pgmap->ops->cleanup) {
 211                         WARN(1, "Missing reference count teardown definition\n");
 212                         return ERR_PTR(-EINVAL);
 213                 }
 214         }
 215
 216         if (need_devmap_managed) {
 217                 error = devmap_managed_enable_get(dev, pgmap);
 218                 if (error)
 219                         return ERR_PTR(error);
 220         }
 221
 222         conflict_pgmap = get_dev_pagemap(PHYS_PFN(res->start), NULL);
 223         if (conflict_pgmap) {
 224                 dev_WARN(dev, "Conflicting mapping in same section\n");
 225                 put_dev_pagemap(conflict_pgmap);
 226                 error = -ENOMEM;
 227                 goto err_array;
 228         }
 229
 230         conflict_pgmap = get_dev_pagemap(PHYS_PFN(res->end), NULL);
 231         if (conflict_pgmap) {
 232                 dev_WARN(dev, "Conflicting mapping in same section\n");
 233                 put_dev_pagemap(conflict_pgmap);
 234                 error = -ENOMEM;
 235                 goto err_array;
 236         }
 237
 238         is_ram = region_intersects(res->start, resource_size(res),
 239                 IORESOURCE_SYSTEM_RAM, IORES_DESC_NONE);
 240
 241         if (is_ram != REGION_DISJOINT) {
 242                 WARN_ONCE(1, "%s attempted on %s region %pr\n", __func__,
 243                                 is_ram == REGION_MIXED ? "mixed" : "ram", res);
 244                 error = -ENXIO;
 245                 goto err_array;
 246         }
 247
 248         pgmap->dev = dev;
 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         nid = dev_to_node(dev);
 256         if (nid < 0)
 257                 nid = numa_mem_id();
 258
 259         error = track_pfn_remap(NULL, &pgprot, PHYS_PFN(res->start), 0,
 260                         resource_size(res));
 261         if (error)
 262                 goto err_pfn_remap;
 263
 264         mem_hotplug_begin();
 265
 266         /*
 267          * For device private memory we call add_pages() as we only need to
 268          * allocate and initialize struct page for the device memory. More-
 269          * over the device memory is un-accessible thus we do not want to
 270          * create a linear mapping for the memory like arch_add_memory()
 271          * would do.
 272          *
 273          * For all other device memory types, which are accessible by
 274          * the CPU, we do want the linear mapping and thus use
 275          * arch_add_memory().
 276          */
 277         if (pgmap->type == MEMORY_DEVICE_PRIVATE) {
 278                 error = add_pages(nid, PHYS_PFN(res->start),
 279                                 PHYS_PFN(resource_size(res)), &restrictions);
 280         } else {
 281                 error = kasan_add_zero_shadow(__va(res->start), resource_size(res));
 282                 if (error) {
 283                         mem_hotplug_done();
 284                         goto err_kasan;
 285                 }
 286
 287                 error = arch_add_memory(nid, res->start, resource_size(res),
 288                                         &restrictions);
 289         }
 290
 291         if (!error) {
 292                 struct zone *zone;
 293
 294                 zone = &NODE_DATA(nid)->node_zones[ZONE_DEVICE];
 295                 move_pfn_range_to_zone(zone, PHYS_PFN(res->start),
 296                                 PHYS_PFN(resource_size(res)), restrictions.altmap);
 297         }
 298
 299         mem_hotplug_done();
 300         if (error)
 301                 goto err_add_memory;
 302
 303         /*
 304          * Initialization of the pages has been deferred until now in order
 305          * to allow us to do the work while not holding the hotplug lock.
 306          */
 307         memmap_init_zone_device(&NODE_DATA(nid)->node_zones[ZONE_DEVICE],
 308                                 PHYS_PFN(res->start),
 309                                 PHYS_PFN(resource_size(res)), pgmap);
 310         percpu_ref_get_many(pgmap->ref, pfn_end(pgmap) - pfn_first(pgmap));
 311
 312         error = devm_add_action_or_reset(dev, devm_memremap_pages_release,
 313                         pgmap);
 314         if (error)
 315                 return ERR_PTR(error);
 316
 317         return __va(res->start);
 318
 319  err_add_memory:
 320         kasan_remove_zero_shadow(__va(res->start), resource_size(res));
 321  err_kasan:
 322         untrack_pfn(NULL, PHYS_PFN(res->start), resource_size(res));
 323  err_pfn_remap:
 324         pgmap_array_delete(res);
 325  err_array:
 326         dev_pagemap_kill(pgmap);
 327         dev_pagemap_cleanup(pgmap);
 328         return ERR_PTR(error);
 329 }
 330 EXPORT_SYMBOL_GPL(devm_memremap_pages);
 331
 332 void devm_memunmap_pages(struct device *dev, struct dev_pagemap *pgmap)
 333 {
 334         devm_release_action(dev, devm_memremap_pages_release, pgmap);
 335 }
 336 EXPORT_SYMBOL_GPL(devm_memunmap_pages);
 337
 338 unsigned long vmem_altmap_offset(struct vmem_altmap *altmap)
 339 {
 340         /* number of pfns from base where pfn_to_page() is valid */
 341         if (altmap)
 342                 return altmap->reserve + altmap->free;
 343         return 0;
 344 }
 345
 346 void vmem_altmap_free(struct vmem_altmap *altmap, unsigned long nr_pfns)
 347 {
 348         altmap->alloc -= nr_pfns;
 349 }
 350
 351 /**
 352  * get_dev_pagemap() - take a new live reference on the dev_pagemap for @pfn
 353  * @pfn: page frame number to lookup page_map
 354  * @pgmap: optional known pgmap that already has a reference
 355  *
 356  * If @pgmap is non-NULL and covers @pfn it will be returned as-is.  If @pgmap
 357  * is non-NULL but does not cover @pfn the reference to it will be released.
 358  */
 359 struct dev_pagemap *get_dev_pagemap(unsigned long pfn,
 360                 struct dev_pagemap *pgmap)
 361 {
 362         resource_size_t phys = PFN_PHYS(pfn);
 363
 364         /*
 365          * In the cached case we're already holding a live reference.
 366          */
 367         if (pgmap) {
 368                 if (phys >= pgmap->res.start && phys <= pgmap->res.end)
 369                         return pgmap;
 370                 put_dev_pagemap(pgmap);
 371         }
 372
 373         /* fall back to slow path lookup */
 374         rcu_read_lock();
 375         pgmap = xa_load(&pgmap_array, PHYS_PFN(phys));
 376         if (pgmap && !percpu_ref_tryget_live(pgmap->ref))
 377                 pgmap = NULL;
 378         rcu_read_unlock();
 379
 380         return pgmap;
 381 }
 382 EXPORT_SYMBOL_GPL(get_dev_pagemap);
 383
 384 #ifdef CONFIG_DEV_PAGEMAP_OPS
 385 void __put_devmap_managed_page(struct page *page)
 386 {
 387         int count = page_ref_dec_return(page);
 388
 389         /*
 390          * If refcount is 1 then page is freed and refcount is stable as nobody
 391          * holds a reference on the page.
 392          */
 393         if (count == 1) {
 394                 /* Clear Active bit in case of parallel mark_page_accessed */
 395                 __ClearPageActive(page);
 396                 __ClearPageWaiters(page);
 397
 398                 mem_cgroup_uncharge(page);
 399
 400                 page->pgmap->ops->page_free(page);
 401         } else if (!count)
 402                 __put_page(page);
 403 }
 404 EXPORT_SYMBOL(__put_devmap_managed_page);
 405 #endif /* CONFIG_DEV_PAGEMAP_OPS */