kernel/memremap.c

   1 /* SPDX-License-Identifier: GPL-2.0 */
   2 /* Copyright(c) 2015 Intel Corporation. All rights reserved. */
   3 #include <linux/radix-tree.h>
   4 #include <linux/device.h>
   5 #include <linux/types.h>
   6 #include <linux/pfn_t.h>
   7 #include <linux/io.h>
   8 #include <linux/mm.h>
   9 #include <linux/memory_hotplug.h>
  10 #include <linux/swap.h>
  11 #include <linux/swapops.h>
  12 #include <linux/wait_bit.h>
  13
  14 static DEFINE_MUTEX(pgmap_lock);
  15 static RADIX_TREE(pgmap_radix, GFP_KERNEL);
  16 #define SECTION_MASK ~((1UL << PA_SECTION_SHIFT) - 1)
  17 #define SECTION_SIZE (1UL << PA_SECTION_SHIFT)
  18
  19 static unsigned long order_at(struct resource *res, unsigned long pgoff)
  20 {
  21         unsigned long phys_pgoff = PHYS_PFN(res->start) + pgoff;
  22         unsigned long nr_pages, mask;
  23
  24         nr_pages = PHYS_PFN(resource_size(res));
  25         if (nr_pages == pgoff)
  26                 return ULONG_MAX;
  27
  28         /*
  29          * What is the largest aligned power-of-2 range available from
  30          * this resource pgoff to the end of the resource range,
  31          * considering the alignment of the current pgoff?
  32          */
  33         mask = phys_pgoff | rounddown_pow_of_two(nr_pages - pgoff);
  34         if (!mask)
  35                 return ULONG_MAX;
  36
  37         return find_first_bit(&mask, BITS_PER_LONG);
  38 }
  39
  40 #define foreach_order_pgoff(res, order, pgoff) \
  41         for (pgoff = 0, order = order_at((res), pgoff); order < ULONG_MAX; \
  42                         pgoff += 1UL << order, order = order_at((res), pgoff))
  43
  44 #if IS_ENABLED(CONFIG_DEVICE_PRIVATE)
  45 int device_private_entry_fault(struct vm_area_struct *vma,
  46                        unsigned long addr,
  47                        swp_entry_t entry,
  48                        unsigned int flags,
  49                        pmd_t *pmdp)
  50 {
  51         struct page *page = device_private_entry_to_page(entry);
  52
  53         /*
  54          * The page_fault() callback must migrate page back to system memory
  55          * so that CPU can access it. This might fail for various reasons
  56          * (device issue, device was unsafely unplugged, ...). When such
  57          * error conditions happen, the callback must return VM_FAULT_SIGBUS.
  58          *
  59          * Note that because memory cgroup charges are accounted to the device
  60          * memory, this should never fail because of memory restrictions (but
  61          * allocation of regular system page might still fail because we are
  62          * out of memory).
  63          *
  64          * There is a more in-depth description of what that callback can and
  65          * cannot do, in include/linux/memremap.h
  66          */
  67         return page->pgmap->page_fault(vma, addr, page, flags, pmdp);
  68 }
  69 EXPORT_SYMBOL(device_private_entry_fault);
  70 #endif /* CONFIG_DEVICE_PRIVATE */
  71
  72 static void pgmap_radix_release(struct resource *res, unsigned long end_pgoff)
  73 {
  74         unsigned long pgoff, order;
  75
  76         mutex_lock(&pgmap_lock);
  77         foreach_order_pgoff(res, order, pgoff) {
  78                 if (pgoff >= end_pgoff)
  79                         break;
  80                 radix_tree_delete(&pgmap_radix, PHYS_PFN(res->start) + pgoff);
  81         }
  82         mutex_unlock(&pgmap_lock);
  83
  84         synchronize_rcu();
  85 }
  86
  87 static unsigned long pfn_first(struct dev_pagemap *pgmap)
  88 {
  89         const struct resource *res = &pgmap->res;
  90         struct vmem_altmap *altmap = &pgmap->altmap;
  91         unsigned long pfn;
  92
  93         pfn = res->start >> PAGE_SHIFT;
  94         if (pgmap->altmap_valid)
  95                 pfn += vmem_altmap_offset(altmap);
  96         return pfn;
  97 }
  98
  99 static unsigned long pfn_end(struct dev_pagemap *pgmap)
 100 {
 101         const struct resource *res = &pgmap->res;
 102
 103         return (res->start + resource_size(res)) >> PAGE_SHIFT;
 104 }
 105
 106 static unsigned long pfn_next(unsigned long pfn)
 107 {
 108         if (pfn % 1024 == 0)
 109                 cond_resched();
 110         return pfn + 1;
 111 }
 112
 113 #define for_each_device_pfn(pfn, map) \
 114         for (pfn = pfn_first(map); pfn < pfn_end(map); pfn = pfn_next(pfn))
 115
 116 static void devm_memremap_pages_release(void *data)
 117 {
 118         struct dev_pagemap *pgmap = data;
 119         struct device *dev = pgmap->dev;
 120         struct resource *res = &pgmap->res;
 121         resource_size_t align_start, align_size;
 122         unsigned long pfn;
 123
 124         for_each_device_pfn(pfn, pgmap)
 125                 put_page(pfn_to_page(pfn));
 126
 127         if (percpu_ref_tryget_live(pgmap->ref)) {
 128                 dev_WARN(dev, "%s: page mapping is still live!\n", __func__);
 129                 percpu_ref_put(pgmap->ref);
 130         }
 131
 132         /* pages are dead and unused, undo the arch mapping */
 133         align_start = res->start & ~(SECTION_SIZE - 1);
 134         align_size = ALIGN(res->start + resource_size(res), SECTION_SIZE)
 135                 - align_start;
 136
 137         mem_hotplug_begin();
 138         arch_remove_memory(align_start, align_size, pgmap->altmap_valid ?
 139                         &pgmap->altmap : NULL);
 140         mem_hotplug_done();
 141
 142         untrack_pfn(NULL, PHYS_PFN(align_start), align_size);
 143         pgmap_radix_release(res, -1);
 144         dev_WARN_ONCE(dev, pgmap->altmap.alloc,
 145                       "%s: failed to free all reserved pages\n", __func__);
 146 }
 147
 148 /**
 149  * devm_memremap_pages - remap and provide memmap backing for the given resource
 150  * @dev: hosting device for @res
 151  * @pgmap: pointer to a struct dev_pgmap
 152  *
 153  * Notes:
 154  * 1/ At a minimum the res, ref and type members of @pgmap must be initialized
 155  *    by the caller before passing it to this function
 156  *
 157  * 2/ The altmap field may optionally be initialized, in which case altmap_valid
 158  *    must be set to true
 159  *
 160  * 3/ pgmap.ref must be 'live' on entry and 'dead' before devm_memunmap_pages()
 161  *    time (or devm release event). The expected order of events is that ref has
 162  *    been through percpu_ref_kill() before devm_memremap_pages_release(). The
 163  *    wait for the completion of all references being dropped and
 164  *    percpu_ref_exit() must occur after devm_memremap_pages_release().
 165  *
 166  * 4/ res is expected to be a host memory range that could feasibly be
 167  *    treated as a "System RAM" range, i.e. not a device mmio range, but
 168  *    this is not enforced.
 169  */
 170 void *devm_memremap_pages(struct device *dev, struct dev_pagemap *pgmap)
 171 {
 172         resource_size_t align_start, align_size, align_end;
 173         struct vmem_altmap *altmap = pgmap->altmap_valid ?
 174                         &pgmap->altmap : NULL;
 175         struct resource *res = &pgmap->res;
 176         unsigned long pfn, pgoff, order;
 177         pgprot_t pgprot = PAGE_KERNEL;
 178         int error, nid, is_ram;
 179
 180         align_start = res->start & ~(SECTION_SIZE - 1);
 181         align_size = ALIGN(res->start + resource_size(res), SECTION_SIZE)
 182                 - align_start;
 183         is_ram = region_intersects(align_start, align_size,
 184                 IORESOURCE_SYSTEM_RAM, IORES_DESC_NONE);
 185
 186         if (is_ram == REGION_MIXED) {
 187                 WARN_ONCE(1, "%s attempted on mixed region %pr\n",
 188                                 __func__, res);
 189                 return ERR_PTR(-ENXIO);
 190         }
 191
 192         if (is_ram == REGION_INTERSECTS)
 193                 return __va(res->start);
 194
 195         if (!pgmap->ref)
 196                 return ERR_PTR(-EINVAL);
 197
 198         pgmap->dev = dev;
 199
 200         mutex_lock(&pgmap_lock);
 201         error = 0;
 202         align_end = align_start + align_size - 1;
 203
 204         foreach_order_pgoff(res, order, pgoff) {
 205                 error = __radix_tree_insert(&pgmap_radix,
 206                                 PHYS_PFN(res->start) + pgoff, order, pgmap);
 207                 if (error) {
 208                         dev_err(dev, "%s: failed: %d\n", __func__, error);
 209                         break;
 210                 }
 211         }
 212         mutex_unlock(&pgmap_lock);
 213         if (error)
 214                 goto err_radix;
 215
 216         nid = dev_to_node(dev);
 217         if (nid < 0)
 218                 nid = numa_mem_id();
 219
 220         error = track_pfn_remap(NULL, &pgprot, PHYS_PFN(align_start), 0,
 221                         align_size);
 222         if (error)
 223                 goto err_pfn_remap;
 224
 225         mem_hotplug_begin();
 226         error = arch_add_memory(nid, align_start, align_size, altmap, false);
 227         if (!error)
 228                 move_pfn_range_to_zone(&NODE_DATA(nid)->node_zones[ZONE_DEVICE],
 229                                         align_start >> PAGE_SHIFT,
 230                                         align_size >> PAGE_SHIFT, altmap);
 231         mem_hotplug_done();
 232         if (error)
 233                 goto err_add_memory;
 234
 235         for_each_device_pfn(pfn, pgmap) {
 236                 struct page *page = pfn_to_page(pfn);
 237
 238                 /*
 239                  * ZONE_DEVICE pages union ->lru with a ->pgmap back
 240                  * pointer.  It is a bug if a ZONE_DEVICE page is ever
 241                  * freed or placed on a driver-private list.  Seed the
 242                  * storage with LIST_POISON* values.
 243                  */
 244                 list_del(&page->lru);
 245                 page->pgmap = pgmap;
 246                 percpu_ref_get(pgmap->ref);
 247         }
 248
 249         devm_add_action(dev, devm_memremap_pages_release, pgmap);
 250
 251         return __va(res->start);
 252
 253  err_add_memory:
 254         untrack_pfn(NULL, PHYS_PFN(align_start), align_size);
 255  err_pfn_remap:
 256  err_radix:
 257         pgmap_radix_release(res, pgoff);
 258         return ERR_PTR(error);
 259 }
 260 EXPORT_SYMBOL(devm_memremap_pages);
 261
 262 unsigned long vmem_altmap_offset(struct vmem_altmap *altmap)
 263 {
 264         /* number of pfns from base where pfn_to_page() is valid */
 265         return altmap->reserve + altmap->free;
 266 }
 267
 268 void vmem_altmap_free(struct vmem_altmap *altmap, unsigned long nr_pfns)
 269 {
 270         altmap->alloc -= nr_pfns;
 271 }
 272
 273 /**
 274  * get_dev_pagemap() - take a new live reference on the dev_pagemap for @pfn
 275  * @pfn: page frame number to lookup page_map
 276  * @pgmap: optional known pgmap that already has a reference
 277  *
 278  * If @pgmap is non-NULL and covers @pfn it will be returned as-is.  If @pgmap
 279  * is non-NULL but does not cover @pfn the reference to it will be released.
 280  */
 281 struct dev_pagemap *get_dev_pagemap(unsigned long pfn,
 282                 struct dev_pagemap *pgmap)
 283 {
 284         resource_size_t phys = PFN_PHYS(pfn);
 285
 286         /*
 287          * In the cached case we're already holding a live reference.
 288          */
 289         if (pgmap) {
 290                 if (phys >= pgmap->res.start && phys <= pgmap->res.end)
 291                         return pgmap;
 292                 put_dev_pagemap(pgmap);
 293         }
 294
 295         /* fall back to slow path lookup */
 296         rcu_read_lock();
 297         pgmap = radix_tree_lookup(&pgmap_radix, PHYS_PFN(phys));
 298         if (pgmap && !percpu_ref_tryget_live(pgmap->ref))
 299                 pgmap = NULL;
 300         rcu_read_unlock();
 301
 302         return pgmap;
 303 }
 304 EXPORT_SYMBOL_GPL(get_dev_pagemap);
 305
 306 #ifdef CONFIG_DEV_PAGEMAP_OPS
 307 DEFINE_STATIC_KEY_FALSE(devmap_managed_key);
 308 EXPORT_SYMBOL_GPL(devmap_managed_key);
 309 static atomic_t devmap_enable;
 310
 311 /*
 312  * Toggle the static key for ->page_free() callbacks when dev_pagemap
 313  * pages go idle.
 314  */
 315 void dev_pagemap_get_ops(void)
 316 {
 317         if (atomic_inc_return(&devmap_enable) == 1)
 318                 static_branch_enable(&devmap_managed_key);
 319 }
 320 EXPORT_SYMBOL_GPL(dev_pagemap_get_ops);
 321
 322 void dev_pagemap_put_ops(void)
 323 {
 324         if (atomic_dec_and_test(&devmap_enable))
 325                 static_branch_disable(&devmap_managed_key);
 326 }
 327 EXPORT_SYMBOL_GPL(dev_pagemap_put_ops);
 328
 329 void __put_devmap_managed_page(struct page *page)
 330 {
 331         int count = page_ref_dec_return(page);
 332
 333         /*
 334          * If refcount is 1 then page is freed and refcount is stable as nobody
 335          * holds a reference on the page.
 336          */
 337         if (count == 1) {
 338                 /* Clear Active bit in case of parallel mark_page_accessed */
 339                 __ClearPageActive(page);
 340                 __ClearPageWaiters(page);
 341
 342                 page->mapping = NULL;
 343                 mem_cgroup_uncharge(page);
 344
 345                 page->pgmap->page_free(page, page->pgmap->data);
 346         } else if (!count)
 347                 __put_page(page);
 348 }
 349 EXPORT_SYMBOL_GPL(__put_devmap_managed_page);
 350 #endif /* CONFIG_DEV_PAGEMAP_OPS */