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
  91         for_each_device_pfn(pfn, pgmap)
  92                 put_page(pfn_to_page(pfn));
  93
  94         if (percpu_ref_tryget_live(pgmap->ref)) {
  95                 dev_WARN(dev, "%s: page mapping is still live!\n", __func__);
  96                 percpu_ref_put(pgmap->ref);
  97         }
  98
  99         /* pages are dead and unused, undo the arch mapping */
 100         align_start = res->start & ~(SECTION_SIZE - 1);
 101         align_size = ALIGN(res->start + resource_size(res), SECTION_SIZE)
 102                 - align_start;
 103
 104         mem_hotplug_begin();
 105         arch_remove_memory(align_start, align_size, pgmap->altmap_valid ?
 106                         &pgmap->altmap : NULL);
 107         kasan_remove_zero_shadow(__va(align_start), align_size);
 108         mem_hotplug_done();
 109
 110         untrack_pfn(NULL, PHYS_PFN(align_start), align_size);
 111         pgmap_array_delete(res);
 112         dev_WARN_ONCE(dev, pgmap->altmap.alloc,
 113                       "%s: failed to free all reserved pages\n", __func__);
 114 }
 115
 116 /**
 117  * devm_memremap_pages - remap and provide memmap backing for the given resource
 118  * @dev: hosting device for @res
 119  * @pgmap: pointer to a struct dev_pgmap
 120  *
 121  * Notes:
 122  * 1/ At a minimum the res, ref and type members of @pgmap must be initialized
 123  *    by the caller before passing it to this function
 124  *
 125  * 2/ The altmap field may optionally be initialized, in which case altmap_valid
 126  *    must be set to true
 127  *
 128  * 3/ pgmap.ref must be 'live' on entry and 'dead' before devm_memunmap_pages()
 129  *    time (or devm release event). The expected order of events is that ref has
 130  *    been through percpu_ref_kill() before devm_memremap_pages_release(). The
 131  *    wait for the completion of all references being dropped and
 132  *    percpu_ref_exit() must occur after devm_memremap_pages_release().
 133  *
 134  * 4/ res is expected to be a host memory range that could feasibly be
 135  *    treated as a "System RAM" range, i.e. not a device mmio range, but
 136  *    this is not enforced.
 137  */
 138 void *devm_memremap_pages(struct device *dev, struct dev_pagemap *pgmap)
 139 {
 140         resource_size_t align_start, align_size, align_end;
 141         struct vmem_altmap *altmap = pgmap->altmap_valid ?
 142                         &pgmap->altmap : NULL;
 143         struct resource *res = &pgmap->res;
 144         struct dev_pagemap *conflict_pgmap;
 145         pgprot_t pgprot = PAGE_KERNEL;
 146         int error, nid, is_ram;
 147
 148         align_start = res->start & ~(SECTION_SIZE - 1);
 149         align_size = ALIGN(res->start + resource_size(res), SECTION_SIZE)
 150                 - align_start;
 151         align_end = align_start + align_size - 1;
 152
 153         conflict_pgmap = get_dev_pagemap(PHYS_PFN(align_start), NULL);
 154         if (conflict_pgmap) {
 155                 dev_WARN(dev, "Conflicting mapping in same section\n");
 156                 put_dev_pagemap(conflict_pgmap);
 157                 return ERR_PTR(-ENOMEM);
 158         }
 159
 160         conflict_pgmap = get_dev_pagemap(PHYS_PFN(align_end), NULL);
 161         if (conflict_pgmap) {
 162                 dev_WARN(dev, "Conflicting mapping in same section\n");
 163                 put_dev_pagemap(conflict_pgmap);
 164                 return ERR_PTR(-ENOMEM);
 165         }
 166
 167         is_ram = region_intersects(align_start, align_size,
 168                 IORESOURCE_SYSTEM_RAM, IORES_DESC_NONE);
 169
 170         if (is_ram == REGION_MIXED) {
 171                 WARN_ONCE(1, "%s attempted on mixed region %pr\n",
 172                                 __func__, res);
 173                 return ERR_PTR(-ENXIO);
 174         }
 175
 176         if (is_ram == REGION_INTERSECTS)
 177                 return __va(res->start);
 178
 179         if (!pgmap->ref)
 180                 return ERR_PTR(-EINVAL);
 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         error = kasan_add_zero_shadow(__va(align_start), align_size);
 200         if (error) {
 201                 mem_hotplug_done();
 202                 goto err_kasan;
 203         }
 204
 205         error = arch_add_memory(nid, align_start, align_size, altmap, false);
 206         if (!error)
 207                 move_pfn_range_to_zone(&NODE_DATA(nid)->node_zones[ZONE_DEVICE],
 208                                         align_start >> PAGE_SHIFT,
 209                                         align_size >> PAGE_SHIFT, altmap);
 210         mem_hotplug_done();
 211         if (error)
 212                 goto err_add_memory;
 213
 214         /*
 215          * Initialization of the pages has been deferred until now in order
 216          * to allow us to do the work while not holding the hotplug lock.
 217          */
 218         memmap_init_zone_device(&NODE_DATA(nid)->node_zones[ZONE_DEVICE],
 219                                 align_start >> PAGE_SHIFT,
 220                                 align_size >> PAGE_SHIFT, pgmap);
 221         percpu_ref_get_many(pgmap->ref, pfn_end(pgmap) - pfn_first(pgmap));
 222
 223         devm_add_action(dev, devm_memremap_pages_release, pgmap);
 224
 225         return __va(res->start);
 226
 227  err_add_memory:
 228         kasan_remove_zero_shadow(__va(align_start), align_size);
 229  err_kasan:
 230         untrack_pfn(NULL, PHYS_PFN(align_start), align_size);
 231  err_pfn_remap:
 232         pgmap_array_delete(res);
 233  err_array:
 234         return ERR_PTR(error);
 235 }
 236 EXPORT_SYMBOL(devm_memremap_pages);
 237
 238 unsigned long vmem_altmap_offset(struct vmem_altmap *altmap)
 239 {
 240         /* number of pfns from base where pfn_to_page() is valid */
 241         return altmap->reserve + altmap->free;
 242 }
 243
 244 void vmem_altmap_free(struct vmem_altmap *altmap, unsigned long nr_pfns)
 245 {
 246         altmap->alloc -= nr_pfns;
 247 }
 248
 249 /**
 250  * get_dev_pagemap() - take a new live reference on the dev_pagemap for @pfn
 251  * @pfn: page frame number to lookup page_map
 252  * @pgmap: optional known pgmap that already has a reference
 253  *
 254  * If @pgmap is non-NULL and covers @pfn it will be returned as-is.  If @pgmap
 255  * is non-NULL but does not cover @pfn the reference to it will be released.
 256  */
 257 struct dev_pagemap *get_dev_pagemap(unsigned long pfn,
 258                 struct dev_pagemap *pgmap)
 259 {
 260         resource_size_t phys = PFN_PHYS(pfn);
 261
 262         /*
 263          * In the cached case we're already holding a live reference.
 264          */
 265         if (pgmap) {
 266                 if (phys >= pgmap->res.start && phys <= pgmap->res.end)
 267                         return pgmap;
 268                 put_dev_pagemap(pgmap);
 269         }
 270
 271         /* fall back to slow path lookup */
 272         rcu_read_lock();
 273         pgmap = xa_load(&pgmap_array, PHYS_PFN(phys));
 274         if (pgmap && !percpu_ref_tryget_live(pgmap->ref))
 275                 pgmap = NULL;
 276         rcu_read_unlock();
 277
 278         return pgmap;
 279 }
 280 EXPORT_SYMBOL_GPL(get_dev_pagemap);
 281
 282 #ifdef CONFIG_DEV_PAGEMAP_OPS
 283 DEFINE_STATIC_KEY_FALSE(devmap_managed_key);
 284 EXPORT_SYMBOL(devmap_managed_key);
 285 static atomic_t devmap_enable;
 286
 287 /*
 288  * Toggle the static key for ->page_free() callbacks when dev_pagemap
 289  * pages go idle.
 290  */
 291 void dev_pagemap_get_ops(void)
 292 {
 293         if (atomic_inc_return(&devmap_enable) == 1)
 294                 static_branch_enable(&devmap_managed_key);
 295 }
 296 EXPORT_SYMBOL_GPL(dev_pagemap_get_ops);
 297
 298 void dev_pagemap_put_ops(void)
 299 {
 300         if (atomic_dec_and_test(&devmap_enable))
 301                 static_branch_disable(&devmap_managed_key);
 302 }
 303 EXPORT_SYMBOL_GPL(dev_pagemap_put_ops);
 304
 305 void __put_devmap_managed_page(struct page *page)
 306 {
 307         int count = page_ref_dec_return(page);
 308
 309         /*
 310          * If refcount is 1 then page is freed and refcount is stable as nobody
 311          * holds a reference on the page.
 312          */
 313         if (count == 1) {
 314                 /* Clear Active bit in case of parallel mark_page_accessed */
 315                 __ClearPageActive(page);
 316                 __ClearPageWaiters(page);
 317
 318                 mem_cgroup_uncharge(page);
 319
 320                 page->pgmap->page_free(page, page->pgmap->data);
 321         } else if (!count)
 322                 __put_page(page);
 323 }
 324 EXPORT_SYMBOL(__put_devmap_managed_page);
 325 #endif /* CONFIG_DEV_PAGEMAP_OPS */