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 #include <linux/hmm.h>
  15
  16 static DEFINE_XARRAY(pgmap_array);
  17 #define SECTION_MASK ~((1UL << PA_SECTION_SHIFT) - 1)
  18 #define SECTION_SIZE (1UL << PA_SECTION_SHIFT)
  19
  20 #if IS_ENABLED(CONFIG_DEVICE_PRIVATE)
  21 vm_fault_t device_private_entry_fault(struct vm_area_struct *vma,
  22                        unsigned long addr,
  23                        swp_entry_t entry,
  24                        unsigned int flags,
  25                        pmd_t *pmdp)
  26 {
  27         struct page *page = device_private_entry_to_page(entry);
  28         struct hmm_devmem *devmem;
  29
  30         devmem = container_of(page->pgmap, typeof(*devmem), pagemap);
  31
  32         /*
  33          * The page_fault() callback must migrate page back to system memory
  34          * so that CPU can access it. This might fail for various reasons
  35          * (device issue, device was unsafely unplugged, ...). When such
  36          * error conditions happen, the callback must return VM_FAULT_SIGBUS.
  37          *
  38          * Note that because memory cgroup charges are accounted to the device
  39          * memory, this should never fail because of memory restrictions (but
  40          * allocation of regular system page might still fail because we are
  41          * out of memory).
  42          *
  43          * There is a more in-depth description of what that callback can and
  44          * cannot do, in include/linux/memremap.h
  45          */
  46         return devmem->page_fault(vma, addr, page, flags, pmdp);
  47 }
  48 #endif /* CONFIG_DEVICE_PRIVATE */
  49
  50 static void pgmap_array_delete(struct resource *res)
  51 {
  52         xa_store_range(&pgmap_array, PHYS_PFN(res->start), PHYS_PFN(res->end),
  53                         NULL, GFP_KERNEL);
  54         synchronize_rcu();
  55 }
  56
  57 static unsigned long pfn_first(struct dev_pagemap *pgmap)
  58 {
  59         const struct resource *res = &pgmap->res;
  60         struct vmem_altmap *altmap = &pgmap->altmap;
  61         unsigned long pfn;
  62
  63         pfn = res->start >> PAGE_SHIFT;
  64         if (pgmap->altmap_valid)
  65                 pfn += vmem_altmap_offset(altmap);
  66         return pfn;
  67 }
  68
  69 static unsigned long pfn_end(struct dev_pagemap *pgmap)
  70 {
  71         const struct resource *res = &pgmap->res;
  72
  73         return (res->start + resource_size(res)) >> PAGE_SHIFT;
  74 }
  75
  76 static unsigned long pfn_next(unsigned long pfn)
  77 {
  78         if (pfn % 1024 == 0)
  79                 cond_resched();
  80         return pfn + 1;
  81 }
  82
  83 #define for_each_device_pfn(pfn, map) \
  84         for (pfn = pfn_first(map); pfn < pfn_end(map); pfn = pfn_next(pfn))
  85
  86 static void devm_memremap_pages_release(void *data)
  87 {
  88         struct dev_pagemap *pgmap = data;
  89         struct device *dev = pgmap->dev;
  90         struct resource *res = &pgmap->res;
  91         resource_size_t align_start, align_size;
  92         unsigned long pfn;
  93         int nid;
  94
  95         pgmap->kill(pgmap->ref);
  96         for_each_device_pfn(pfn, pgmap)
  97                 put_page(pfn_to_page(pfn));
  98         pgmap->cleanup(pgmap->ref);
  99
 100         /* pages are dead and unused, undo the arch mapping */
 101         align_start = res->start & ~(SECTION_SIZE - 1);
 102         align_size = ALIGN(res->start + resource_size(res), SECTION_SIZE)
 103                 - align_start;
 104
 105         nid = page_to_nid(pfn_to_page(align_start >> PAGE_SHIFT));
 106
 107         mem_hotplug_begin();
 108         if (pgmap->type == MEMORY_DEVICE_PRIVATE) {
 109                 pfn = align_start >> PAGE_SHIFT;
 110                 __remove_pages(page_zone(pfn_to_page(pfn)), pfn,
 111                                 align_size >> PAGE_SHIFT, NULL);
 112         } else {
 113                 arch_remove_memory(nid, align_start, align_size,
 114                                 pgmap->altmap_valid ? &pgmap->altmap : NULL);
 115                 kasan_remove_zero_shadow(__va(align_start), align_size);
 116         }
 117         mem_hotplug_done();
 118
 119         untrack_pfn(NULL, PHYS_PFN(align_start), align_size);
 120         pgmap_array_delete(res);
 121         dev_WARN_ONCE(dev, pgmap->altmap.alloc,
 122                       "%s: failed to free all reserved pages\n", __func__);
 123 }
 124
 125 /**
 126  * devm_memremap_pages - remap and provide memmap backing for the given resource
 127  * @dev: hosting device for @res
 128  * @pgmap: pointer to a struct dev_pagemap
 129  *
 130  * Notes:
 131  * 1/ At a minimum the res, ref and type members of @pgmap must be initialized
 132  *    by the caller before passing it to this function
 133  *
 134  * 2/ The altmap field may optionally be initialized, in which case altmap_valid
 135  *    must be set to true
 136  *
 137  * 3/ pgmap->ref must be 'live' on entry and will be killed and reaped
 138  *    at devm_memremap_pages_release() time, or if this routine fails.
 139  *
 140  * 4/ res is expected to be a host memory range that could feasibly be
 141  *    treated as a "System RAM" range, i.e. not a device mmio range, but
 142  *    this is not enforced.
 143  */
 144 void *devm_memremap_pages(struct device *dev, struct dev_pagemap *pgmap)
 145 {
 146         resource_size_t align_start, align_size, align_end;
 147         struct vmem_altmap *altmap = pgmap->altmap_valid ?
 148                         &pgmap->altmap : NULL;
 149         struct resource *res = &pgmap->res;
 150         struct dev_pagemap *conflict_pgmap;
 151         struct mhp_restrictions restrictions = {
 152                 /*
 153                  * We do not want any optional features only our own memmap
 154                 */
 155                 .altmap = altmap,
 156         };
 157         pgprot_t pgprot = PAGE_KERNEL;
 158         int error, nid, is_ram;
 159
 160         if (!pgmap->ref || !pgmap->kill || !pgmap->cleanup) {
 161                 WARN(1, "Missing reference count teardown definition\n");
 162                 return ERR_PTR(-EINVAL);
 163         }
 164
 165         align_start = res->start & ~(SECTION_SIZE - 1);
 166         align_size = ALIGN(res->start + resource_size(res), SECTION_SIZE)
 167                 - align_start;
 168         align_end = align_start + align_size - 1;
 169
 170         conflict_pgmap = get_dev_pagemap(PHYS_PFN(align_start), NULL);
 171         if (conflict_pgmap) {
 172                 dev_WARN(dev, "Conflicting mapping in same section\n");
 173                 put_dev_pagemap(conflict_pgmap);
 174                 error = -ENOMEM;
 175                 goto err_array;
 176         }
 177
 178         conflict_pgmap = get_dev_pagemap(PHYS_PFN(align_end), NULL);
 179         if (conflict_pgmap) {
 180                 dev_WARN(dev, "Conflicting mapping in same section\n");
 181                 put_dev_pagemap(conflict_pgmap);
 182                 error = -ENOMEM;
 183                 goto err_array;
 184         }
 185
 186         is_ram = region_intersects(align_start, align_size,
 187                 IORESOURCE_SYSTEM_RAM, IORES_DESC_NONE);
 188
 189         if (is_ram != REGION_DISJOINT) {
 190                 WARN_ONCE(1, "%s attempted on %s region %pr\n", __func__,
 191                                 is_ram == REGION_MIXED ? "mixed" : "ram", res);
 192                 error = -ENXIO;
 193                 goto err_array;
 194         }
 195
 196         pgmap->dev = dev;
 197
 198         error = xa_err(xa_store_range(&pgmap_array, PHYS_PFN(res->start),
 199                                 PHYS_PFN(res->end), pgmap, GFP_KERNEL));
 200         if (error)
 201                 goto err_array;
 202
 203         nid = dev_to_node(dev);
 204         if (nid < 0)
 205                 nid = numa_mem_id();
 206
 207         error = track_pfn_remap(NULL, &pgprot, PHYS_PFN(align_start), 0,
 208                         align_size);
 209         if (error)
 210                 goto err_pfn_remap;
 211
 212         mem_hotplug_begin();
 213
 214         /*
 215          * For device private memory we call add_pages() as we only need to
 216          * allocate and initialize struct page for the device memory. More-
 217          * over the device memory is un-accessible thus we do not want to
 218          * create a linear mapping for the memory like arch_add_memory()
 219          * would do.
 220          *
 221          * For all other device memory types, which are accessible by
 222          * the CPU, we do want the linear mapping and thus use
 223          * arch_add_memory().
 224          */
 225         if (pgmap->type == MEMORY_DEVICE_PRIVATE) {
 226                 error = add_pages(nid, align_start >> PAGE_SHIFT,
 227                                 align_size >> PAGE_SHIFT, &restrictions);
 228         } else {
 229                 error = kasan_add_zero_shadow(__va(align_start), align_size);
 230                 if (error) {
 231                         mem_hotplug_done();
 232                         goto err_kasan;
 233                 }
 234
 235                 error = arch_add_memory(nid, align_start, align_size,
 236                                         &restrictions);
 237         }
 238
 239         if (!error) {
 240                 struct zone *zone;
 241
 242                 zone = &NODE_DATA(nid)->node_zones[ZONE_DEVICE];
 243                 move_pfn_range_to_zone(zone, align_start >> PAGE_SHIFT,
 244                                 align_size >> PAGE_SHIFT, altmap);
 245         }
 246
 247         mem_hotplug_done();
 248         if (error)
 249                 goto err_add_memory;
 250
 251         /*
 252          * Initialization of the pages has been deferred until now in order
 253          * to allow us to do the work while not holding the hotplug lock.
 254          */
 255         memmap_init_zone_device(&NODE_DATA(nid)->node_zones[ZONE_DEVICE],
 256                                 align_start >> PAGE_SHIFT,
 257                                 align_size >> PAGE_SHIFT, pgmap);
 258         percpu_ref_get_many(pgmap->ref, pfn_end(pgmap) - pfn_first(pgmap));
 259
 260         error = devm_add_action_or_reset(dev, devm_memremap_pages_release,
 261                         pgmap);
 262         if (error)
 263                 return ERR_PTR(error);
 264
 265         return __va(res->start);
 266
 267  err_add_memory:
 268         kasan_remove_zero_shadow(__va(align_start), align_size);
 269  err_kasan:
 270         untrack_pfn(NULL, PHYS_PFN(align_start), align_size);
 271  err_pfn_remap:
 272         pgmap_array_delete(res);
 273  err_array:
 274         pgmap->kill(pgmap->ref);
 275         pgmap->cleanup(pgmap->ref);
 276
 277         return ERR_PTR(error);
 278 }
 279 EXPORT_SYMBOL_GPL(devm_memremap_pages);
 280
 281 void devm_memunmap_pages(struct device *dev, struct dev_pagemap *pgmap)
 282 {
 283         devm_release_action(dev, devm_memremap_pages_release, pgmap);
 284 }
 285 EXPORT_SYMBOL_GPL(devm_memunmap_pages);
 286
 287 unsigned long vmem_altmap_offset(struct vmem_altmap *altmap)
 288 {
 289         /* number of pfns from base where pfn_to_page() is valid */
 290         return altmap->reserve + altmap->free;
 291 }
 292
 293 void vmem_altmap_free(struct vmem_altmap *altmap, unsigned long nr_pfns)
 294 {
 295         altmap->alloc -= nr_pfns;
 296 }
 297
 298 /**
 299  * get_dev_pagemap() - take a new live reference on the dev_pagemap for @pfn
 300  * @pfn: page frame number to lookup page_map
 301  * @pgmap: optional known pgmap that already has a reference
 302  *
 303  * If @pgmap is non-NULL and covers @pfn it will be returned as-is.  If @pgmap
 304  * is non-NULL but does not cover @pfn the reference to it will be released.
 305  */
 306 struct dev_pagemap *get_dev_pagemap(unsigned long pfn,
 307                 struct dev_pagemap *pgmap)
 308 {
 309         resource_size_t phys = PFN_PHYS(pfn);
 310
 311         /*
 312          * In the cached case we're already holding a live reference.
 313          */
 314         if (pgmap) {
 315                 if (phys >= pgmap->res.start && phys <= pgmap->res.end)
 316                         return pgmap;
 317                 put_dev_pagemap(pgmap);
 318         }
 319
 320         /* fall back to slow path lookup */
 321         rcu_read_lock();
 322         pgmap = xa_load(&pgmap_array, PHYS_PFN(phys));
 323         if (pgmap && !percpu_ref_tryget_live(pgmap->ref))
 324                 pgmap = NULL;
 325         rcu_read_unlock();
 326
 327         return pgmap;
 328 }
 329 EXPORT_SYMBOL_GPL(get_dev_pagemap);
 330
 331 #ifdef CONFIG_DEV_PAGEMAP_OPS
 332 DEFINE_STATIC_KEY_FALSE(devmap_managed_key);
 333 EXPORT_SYMBOL(devmap_managed_key);
 334 static atomic_t devmap_enable;
 335
 336 /*
 337  * Toggle the static key for ->page_free() callbacks when dev_pagemap
 338  * pages go idle.
 339  */
 340 void dev_pagemap_get_ops(void)
 341 {
 342         if (atomic_inc_return(&devmap_enable) == 1)
 343                 static_branch_enable(&devmap_managed_key);
 344 }
 345 EXPORT_SYMBOL_GPL(dev_pagemap_get_ops);
 346
 347 void dev_pagemap_put_ops(void)
 348 {
 349         if (atomic_dec_and_test(&devmap_enable))
 350                 static_branch_disable(&devmap_managed_key);
 351 }
 352 EXPORT_SYMBOL_GPL(dev_pagemap_put_ops);
 353
 354 void __put_devmap_managed_page(struct page *page)
 355 {
 356         int count = page_ref_dec_return(page);
 357
 358         /*
 359          * If refcount is 1 then page is freed and refcount is stable as nobody
 360          * holds a reference on the page.
 361          */
 362         if (count == 1) {
 363                 /* Clear Active bit in case of parallel mark_page_accessed */
 364                 __ClearPageActive(page);
 365                 __ClearPageWaiters(page);
 366
 367                 mem_cgroup_uncharge(page);
 368
 369                 page->pgmap->page_free(page, page->pgmap->data);
 370         } else if (!count)
 371                 __put_page(page);
 372 }
 373 EXPORT_SYMBOL(__put_devmap_managed_page);
 374 #endif /* CONFIG_DEV_PAGEMAP_OPS */