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 #if IS_ENABLED(CONFIG_DEVICE_PRIVATE)
  20 vm_fault_t device_private_entry_fault(struct vm_area_struct *vma,
  21                        unsigned long addr,
  22                        swp_entry_t entry,
  23                        unsigned int flags,
  24                        pmd_t *pmdp)
  25 {
  26         struct page *page = device_private_entry_to_page(entry);
  27
  28         /*
  29          * The page_fault() callback must migrate page back to system memory
  30          * so that CPU can access it. This might fail for various reasons
  31          * (device issue, device was unsafely unplugged, ...). When such
  32          * error conditions happen, the callback must return VM_FAULT_SIGBUS.
  33          *
  34          * Note that because memory cgroup charges are accounted to the device
  35          * memory, this should never fail because of memory restrictions (but
  36          * allocation of regular system page might still fail because we are
  37          * out of memory).
  38          *
  39          * There is a more in-depth description of what that callback can and
  40          * cannot do, in include/linux/memremap.h
  41          */
  42         return page->pgmap->page_fault(vma, addr, page, flags, pmdp);
  43 }
  44 EXPORT_SYMBOL(device_private_entry_fault);
  45 #endif /* CONFIG_DEVICE_PRIVATE */
  46
  47 static void pgmap_array_delete(struct resource *res)
  48 {
  49         xa_store_range(&pgmap_array, PHYS_PFN(res->start), PHYS_PFN(res->end),
  50                         NULL, GFP_KERNEL);
  51         synchronize_rcu();
  52 }
  53
  54 static unsigned long pfn_first(struct dev_pagemap *pgmap)
  55 {
  56         const struct resource *res = &pgmap->res;
  57         struct vmem_altmap *altmap = &pgmap->altmap;
  58         unsigned long pfn;
  59
  60         pfn = res->start >> PAGE_SHIFT;
  61         if (pgmap->altmap_valid)
  62                 pfn += vmem_altmap_offset(altmap);
  63         return pfn;
  64 }
  65
  66 static unsigned long pfn_end(struct dev_pagemap *pgmap)
  67 {
  68         const struct resource *res = &pgmap->res;
  69
  70         return (res->start + resource_size(res)) >> PAGE_SHIFT;
  71 }
  72
  73 static unsigned long pfn_next(unsigned long pfn)
  74 {
  75         if (pfn % 1024 == 0)
  76                 cond_resched();
  77         return pfn + 1;
  78 }
  79
  80 #define for_each_device_pfn(pfn, map) \
  81         for (pfn = pfn_first(map); pfn < pfn_end(map); pfn = pfn_next(pfn))
  82
  83 static void devm_memremap_pages_release(void *data)
  84 {
  85         struct dev_pagemap *pgmap = data;
  86         struct device *dev = pgmap->dev;
  87         struct resource *res = &pgmap->res;
  88         resource_size_t align_start, align_size;
  89         unsigned long pfn;
  90         int nid;
  91
  92         pgmap->kill(pgmap->ref);
  93         for_each_device_pfn(pfn, pgmap)
  94                 put_page(pfn_to_page(pfn));
  95
  96         /* pages are dead and unused, undo the arch mapping */
  97         align_start = res->start & ~(SECTION_SIZE - 1);
  98         align_size = ALIGN(res->start + resource_size(res), SECTION_SIZE)
  99                 - align_start;
 100
 101         nid = page_to_nid(pfn_to_page(align_start >> PAGE_SHIFT));
 102
 103         mem_hotplug_begin();
 104         if (pgmap->type == MEMORY_DEVICE_PRIVATE) {
 105                 pfn = align_start >> PAGE_SHIFT;
 106                 __remove_pages(page_zone(pfn_to_page(pfn)), pfn,
 107                                 align_size >> PAGE_SHIFT, NULL);
 108         } else {
 109                 arch_remove_memory(nid, align_start, align_size,
 110                                 pgmap->altmap_valid ? &pgmap->altmap : NULL);
 111                 kasan_remove_zero_shadow(__va(align_start), align_size);
 112         }
 113         mem_hotplug_done();
 114
 115         untrack_pfn(NULL, PHYS_PFN(align_start), align_size);
 116         pgmap_array_delete(res);
 117         dev_WARN_ONCE(dev, pgmap->altmap.alloc,
 118                       "%s: failed to free all reserved pages\n", __func__);
 119 }
 120
 121 /**
 122  * devm_memremap_pages - remap and provide memmap backing for the given resource
 123  * @dev: hosting device for @res
 124  * @pgmap: pointer to a struct dev_pagemap
 125  *
 126  * Notes:
 127  * 1/ At a minimum the res, ref and type members of @pgmap must be initialized
 128  *    by the caller before passing it to this function
 129  *
 130  * 2/ The altmap field may optionally be initialized, in which case altmap_valid
 131  *    must be set to true
 132  *
 133  * 3/ pgmap->ref must be 'live' on entry and will be killed at
 134  *    devm_memremap_pages_release() time, or if this routine fails.
 135  *
 136  * 4/ res is expected to be a host memory range that could feasibly be
 137  *    treated as a "System RAM" range, i.e. not a device mmio range, but
 138  *    this is not enforced.
 139  */
 140 void *devm_memremap_pages(struct device *dev, struct dev_pagemap *pgmap)
 141 {
 142         resource_size_t align_start, align_size, align_end;
 143         struct vmem_altmap *altmap = pgmap->altmap_valid ?
 144                         &pgmap->altmap : NULL;
 145         struct resource *res = &pgmap->res;
 146         struct dev_pagemap *conflict_pgmap;
 147         pgprot_t pgprot = PAGE_KERNEL;
 148         int error, nid, is_ram;
 149
 150         if (!pgmap->ref || !pgmap->kill)
 151                 return ERR_PTR(-EINVAL);
 152
 153         align_start = res->start & ~(SECTION_SIZE - 1);
 154         align_size = ALIGN(res->start + resource_size(res), SECTION_SIZE)
 155                 - align_start;
 156         align_end = align_start + align_size - 1;
 157
 158         conflict_pgmap = get_dev_pagemap(PHYS_PFN(align_start), NULL);
 159         if (conflict_pgmap) {
 160                 dev_WARN(dev, "Conflicting mapping in same section\n");
 161                 put_dev_pagemap(conflict_pgmap);
 162                 return ERR_PTR(-ENOMEM);
 163         }
 164
 165         conflict_pgmap = get_dev_pagemap(PHYS_PFN(align_end), NULL);
 166         if (conflict_pgmap) {
 167                 dev_WARN(dev, "Conflicting mapping in same section\n");
 168                 put_dev_pagemap(conflict_pgmap);
 169                 return ERR_PTR(-ENOMEM);
 170         }
 171
 172         is_ram = region_intersects(align_start, align_size,
 173                 IORESOURCE_SYSTEM_RAM, IORES_DESC_NONE);
 174
 175         if (is_ram != REGION_DISJOINT) {
 176                 WARN_ONCE(1, "%s attempted on %s region %pr\n", __func__,
 177                                 is_ram == REGION_MIXED ? "mixed" : "ram", res);
 178                 error = -ENXIO;
 179                 goto err_array;
 180         }
 181
 182         pgmap->dev = dev;
 183
 184         error = xa_err(xa_store_range(&pgmap_array, PHYS_PFN(res->start),
 185                                 PHYS_PFN(res->end), pgmap, GFP_KERNEL));
 186         if (error)
 187                 goto err_array;
 188
 189         nid = dev_to_node(dev);
 190         if (nid < 0)
 191                 nid = numa_mem_id();
 192
 193         error = track_pfn_remap(NULL, &pgprot, PHYS_PFN(align_start), 0,
 194                         align_size);
 195         if (error)
 196                 goto err_pfn_remap;
 197
 198         mem_hotplug_begin();
 199
 200         /*
 201          * For device private memory we call add_pages() as we only need to
 202          * allocate and initialize struct page for the device memory. More-
 203          * over the device memory is un-accessible thus we do not want to
 204          * create a linear mapping for the memory like arch_add_memory()
 205          * would do.
 206          *
 207          * For all other device memory types, which are accessible by
 208          * the CPU, we do want the linear mapping and thus use
 209          * arch_add_memory().
 210          */
 211         if (pgmap->type == MEMORY_DEVICE_PRIVATE) {
 212                 error = add_pages(nid, align_start >> PAGE_SHIFT,
 213                                 align_size >> PAGE_SHIFT, NULL, false);
 214         } else {
 215                 error = kasan_add_zero_shadow(__va(align_start), align_size);
 216                 if (error) {
 217                         mem_hotplug_done();
 218                         goto err_kasan;
 219                 }
 220
 221                 error = arch_add_memory(nid, align_start, align_size, altmap,
 222                                 false);
 223         }
 224
 225         if (!error) {
 226                 struct zone *zone;
 227
 228                 zone = &NODE_DATA(nid)->node_zones[ZONE_DEVICE];
 229                 move_pfn_range_to_zone(zone, align_start >> PAGE_SHIFT,
 230                                 align_size >> PAGE_SHIFT, altmap);
 231         }
 232
 233         mem_hotplug_done();
 234         if (error)
 235                 goto err_add_memory;
 236
 237         /*
 238          * Initialization of the pages has been deferred until now in order
 239          * to allow us to do the work while not holding the hotplug lock.
 240          */
 241         memmap_init_zone_device(&NODE_DATA(nid)->node_zones[ZONE_DEVICE],
 242                                 align_start >> PAGE_SHIFT,
 243                                 align_size >> PAGE_SHIFT, pgmap);
 244         percpu_ref_get_many(pgmap->ref, pfn_end(pgmap) - pfn_first(pgmap));
 245
 246         error = devm_add_action_or_reset(dev, devm_memremap_pages_release,
 247                         pgmap);
 248         if (error)
 249                 return ERR_PTR(error);
 250
 251         return __va(res->start);
 252
 253  err_add_memory:
 254         kasan_remove_zero_shadow(__va(align_start), align_size);
 255  err_kasan:
 256         untrack_pfn(NULL, PHYS_PFN(align_start), align_size);
 257  err_pfn_remap:
 258         pgmap_array_delete(res);
 259  err_array:
 260         pgmap->kill(pgmap->ref);
 261         return ERR_PTR(error);
 262 }
 263 EXPORT_SYMBOL_GPL(devm_memremap_pages);
 264
 265 unsigned long vmem_altmap_offset(struct vmem_altmap *altmap)
 266 {
 267         /* number of pfns from base where pfn_to_page() is valid */
 268         return altmap->reserve + altmap->free;
 269 }
 270
 271 void vmem_altmap_free(struct vmem_altmap *altmap, unsigned long nr_pfns)
 272 {
 273         altmap->alloc -= nr_pfns;
 274 }
 275
 276 /**
 277  * get_dev_pagemap() - take a new live reference on the dev_pagemap for @pfn
 278  * @pfn: page frame number to lookup page_map
 279  * @pgmap: optional known pgmap that already has a reference
 280  *
 281  * If @pgmap is non-NULL and covers @pfn it will be returned as-is.  If @pgmap
 282  * is non-NULL but does not cover @pfn the reference to it will be released.
 283  */
 284 struct dev_pagemap *get_dev_pagemap(unsigned long pfn,
 285                 struct dev_pagemap *pgmap)
 286 {
 287         resource_size_t phys = PFN_PHYS(pfn);
 288
 289         /*
 290          * In the cached case we're already holding a live reference.
 291          */
 292         if (pgmap) {
 293                 if (phys >= pgmap->res.start && phys <= pgmap->res.end)
 294                         return pgmap;
 295                 put_dev_pagemap(pgmap);
 296         }
 297
 298         /* fall back to slow path lookup */
 299         rcu_read_lock();
 300         pgmap = xa_load(&pgmap_array, PHYS_PFN(phys));
 301         if (pgmap && !percpu_ref_tryget_live(pgmap->ref))
 302                 pgmap = NULL;
 303         rcu_read_unlock();
 304
 305         return pgmap;
 306 }
 307 EXPORT_SYMBOL_GPL(get_dev_pagemap);
 308
 309 #ifdef CONFIG_DEV_PAGEMAP_OPS
 310 DEFINE_STATIC_KEY_FALSE(devmap_managed_key);
 311 EXPORT_SYMBOL(devmap_managed_key);
 312 static atomic_t devmap_enable;
 313
 314 /*
 315  * Toggle the static key for ->page_free() callbacks when dev_pagemap
 316  * pages go idle.
 317  */
 318 void dev_pagemap_get_ops(void)
 319 {
 320         if (atomic_inc_return(&devmap_enable) == 1)
 321                 static_branch_enable(&devmap_managed_key);
 322 }
 323 EXPORT_SYMBOL_GPL(dev_pagemap_get_ops);
 324
 325 void dev_pagemap_put_ops(void)
 326 {
 327         if (atomic_dec_and_test(&devmap_enable))
 328                 static_branch_disable(&devmap_managed_key);
 329 }
 330 EXPORT_SYMBOL_GPL(dev_pagemap_put_ops);
 331
 332 void __put_devmap_managed_page(struct page *page)
 333 {
 334         int count = page_ref_dec_return(page);
 335
 336         /*
 337          * If refcount is 1 then page is freed and refcount is stable as nobody
 338          * holds a reference on the page.
 339          */
 340         if (count == 1) {
 341                 /* Clear Active bit in case of parallel mark_page_accessed */
 342                 __ClearPageActive(page);
 343                 __ClearPageWaiters(page);
 344
 345                 mem_cgroup_uncharge(page);
 346
 347                 page->pgmap->page_free(page, page->pgmap->data);
 348         } else if (!count)
 349                 __put_page(page);
 350 }
 351 EXPORT_SYMBOL(__put_devmap_managed_page);
 352 #endif /* CONFIG_DEV_PAGEMAP_OPS */