linux/mm/memremap.c
<<
>>
Prefs
   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/mmzone.h>
  11#include <linux/swapops.h>
  12#include <linux/types.h>
  13#include <linux/wait_bit.h>
  14#include <linux/xarray.h>
  15
  16static DEFINE_XARRAY(pgmap_array);
  17
  18/*
  19 * The memremap() and memremap_pages() interfaces are alternately used
  20 * to map persistent memory namespaces. These interfaces place different
  21 * constraints on the alignment and size of the mapping (namespace).
  22 * memremap() can map individual PAGE_SIZE pages. memremap_pages() can
  23 * only map subsections (2MB), and at least one architecture (PowerPC)
  24 * the minimum mapping granularity of memremap_pages() is 16MB.
  25 *
  26 * The role of memremap_compat_align() is to communicate the minimum
  27 * arch supported alignment of a namespace such that it can freely
  28 * switch modes without violating the arch constraint. Namely, do not
  29 * allow a namespace to be PAGE_SIZE aligned since that namespace may be
  30 * reconfigured into a mode that requires SUBSECTION_SIZE alignment.
  31 */
  32#ifndef CONFIG_ARCH_HAS_MEMREMAP_COMPAT_ALIGN
  33unsigned long memremap_compat_align(void)
  34{
  35        return SUBSECTION_SIZE;
  36}
  37EXPORT_SYMBOL_GPL(memremap_compat_align);
  38#endif
  39
  40#ifdef CONFIG_DEV_PAGEMAP_OPS
  41DEFINE_STATIC_KEY_FALSE(devmap_managed_key);
  42EXPORT_SYMBOL(devmap_managed_key);
  43
  44static void devmap_managed_enable_put(struct dev_pagemap *pgmap)
  45{
  46        if (pgmap->type == MEMORY_DEVICE_PRIVATE ||
  47            pgmap->type == MEMORY_DEVICE_FS_DAX)
  48                static_branch_dec(&devmap_managed_key);
  49}
  50
  51static void devmap_managed_enable_get(struct dev_pagemap *pgmap)
  52{
  53        if (pgmap->type == MEMORY_DEVICE_PRIVATE ||
  54            pgmap->type == MEMORY_DEVICE_FS_DAX)
  55                static_branch_inc(&devmap_managed_key);
  56}
  57#else
  58static void devmap_managed_enable_get(struct dev_pagemap *pgmap)
  59{
  60}
  61static void devmap_managed_enable_put(struct dev_pagemap *pgmap)
  62{
  63}
  64#endif /* CONFIG_DEV_PAGEMAP_OPS */
  65
  66static void pgmap_array_delete(struct range *range)
  67{
  68        xa_store_range(&pgmap_array, PHYS_PFN(range->start), PHYS_PFN(range->end),
  69                        NULL, GFP_KERNEL);
  70        synchronize_rcu();
  71}
  72
  73static unsigned long pfn_first(struct dev_pagemap *pgmap, int range_id)
  74{
  75        struct range *range = &pgmap->ranges[range_id];
  76        unsigned long pfn = PHYS_PFN(range->start);
  77
  78        if (range_id)
  79                return pfn;
  80        return pfn + vmem_altmap_offset(pgmap_altmap(pgmap));
  81}
  82
  83bool pgmap_pfn_valid(struct dev_pagemap *pgmap, unsigned long pfn)
  84{
  85        int i;
  86
  87        for (i = 0; i < pgmap->nr_range; i++) {
  88                struct range *range = &pgmap->ranges[i];
  89
  90                if (pfn >= PHYS_PFN(range->start) &&
  91                    pfn <= PHYS_PFN(range->end))
  92                        return pfn >= pfn_first(pgmap, i);
  93        }
  94
  95        return false;
  96}
  97
  98static unsigned long pfn_end(struct dev_pagemap *pgmap, int range_id)
  99{
 100        const struct range *range = &pgmap->ranges[range_id];
 101
 102        return (range->start + range_len(range)) >> PAGE_SHIFT;
 103}
 104
 105static unsigned long pfn_next(unsigned long pfn)
 106{
 107        if (pfn % 1024 == 0)
 108                cond_resched();
 109        return pfn + 1;
 110}
 111
 112#define for_each_device_pfn(pfn, map, i) \
 113        for (pfn = pfn_first(map, i); pfn < pfn_end(map, i); pfn = pfn_next(pfn))
 114
 115static void dev_pagemap_kill(struct dev_pagemap *pgmap)
 116{
 117        if (pgmap->ops && pgmap->ops->kill)
 118                pgmap->ops->kill(pgmap);
 119        else
 120                percpu_ref_kill(pgmap->ref);
 121}
 122
 123static void dev_pagemap_cleanup(struct dev_pagemap *pgmap)
 124{
 125        if (pgmap->ops && pgmap->ops->cleanup) {
 126                pgmap->ops->cleanup(pgmap);
 127        } else {
 128                wait_for_completion(&pgmap->done);
 129                percpu_ref_exit(pgmap->ref);
 130        }
 131        /*
 132         * Undo the pgmap ref assignment for the internal case as the
 133         * caller may re-enable the same pgmap.
 134         */
 135        if (pgmap->ref == &pgmap->internal_ref)
 136                pgmap->ref = NULL;
 137}
 138
 139static void pageunmap_range(struct dev_pagemap *pgmap, int range_id)
 140{
 141        struct range *range = &pgmap->ranges[range_id];
 142        struct page *first_page;
 143        int nid;
 144
 145        /* make sure to access a memmap that was actually initialized */
 146        first_page = pfn_to_page(pfn_first(pgmap, range_id));
 147
 148        /* pages are dead and unused, undo the arch mapping */
 149        nid = page_to_nid(first_page);
 150
 151        mem_hotplug_begin();
 152        remove_pfn_range_from_zone(page_zone(first_page), PHYS_PFN(range->start),
 153                                   PHYS_PFN(range_len(range)));
 154        if (pgmap->type == MEMORY_DEVICE_PRIVATE) {
 155                __remove_pages(PHYS_PFN(range->start),
 156                               PHYS_PFN(range_len(range)), NULL);
 157        } else {
 158                arch_remove_memory(nid, range->start, range_len(range),
 159                                pgmap_altmap(pgmap));
 160                kasan_remove_zero_shadow(__va(range->start), range_len(range));
 161        }
 162        mem_hotplug_done();
 163
 164        untrack_pfn(NULL, PHYS_PFN(range->start), range_len(range));
 165        pgmap_array_delete(range);
 166}
 167
 168void memunmap_pages(struct dev_pagemap *pgmap)
 169{
 170        unsigned long pfn;
 171        int i;
 172
 173        dev_pagemap_kill(pgmap);
 174        for (i = 0; i < pgmap->nr_range; i++)
 175                for_each_device_pfn(pfn, pgmap, i)
 176                        put_page(pfn_to_page(pfn));
 177        dev_pagemap_cleanup(pgmap);
 178
 179        for (i = 0; i < pgmap->nr_range; i++)
 180                pageunmap_range(pgmap, i);
 181
 182        WARN_ONCE(pgmap->altmap.alloc, "failed to free all reserved pages\n");
 183        devmap_managed_enable_put(pgmap);
 184}
 185EXPORT_SYMBOL_GPL(memunmap_pages);
 186
 187static void devm_memremap_pages_release(void *data)
 188{
 189        memunmap_pages(data);
 190}
 191
 192static void dev_pagemap_percpu_release(struct percpu_ref *ref)
 193{
 194        struct dev_pagemap *pgmap =
 195                container_of(ref, struct dev_pagemap, internal_ref);
 196
 197        complete(&pgmap->done);
 198}
 199
 200static int pagemap_range(struct dev_pagemap *pgmap, struct mhp_params *params,
 201                int range_id, int nid)
 202{
 203        const bool is_private = pgmap->type == MEMORY_DEVICE_PRIVATE;
 204        struct range *range = &pgmap->ranges[range_id];
 205        struct dev_pagemap *conflict_pgmap;
 206        int error, is_ram;
 207
 208        if (WARN_ONCE(pgmap_altmap(pgmap) && range_id > 0,
 209                                "altmap not supported for multiple ranges\n"))
 210                return -EINVAL;
 211
 212        conflict_pgmap = get_dev_pagemap(PHYS_PFN(range->start), NULL);
 213        if (conflict_pgmap) {
 214                WARN(1, "Conflicting mapping in same section\n");
 215                put_dev_pagemap(conflict_pgmap);
 216                return -ENOMEM;
 217        }
 218
 219        conflict_pgmap = get_dev_pagemap(PHYS_PFN(range->end), NULL);
 220        if (conflict_pgmap) {
 221                WARN(1, "Conflicting mapping in same section\n");
 222                put_dev_pagemap(conflict_pgmap);
 223                return -ENOMEM;
 224        }
 225
 226        is_ram = region_intersects(range->start, range_len(range),
 227                IORESOURCE_SYSTEM_RAM, IORES_DESC_NONE);
 228
 229        if (is_ram != REGION_DISJOINT) {
 230                WARN_ONCE(1, "attempted on %s region %#llx-%#llx\n",
 231                                is_ram == REGION_MIXED ? "mixed" : "ram",
 232                                range->start, range->end);
 233                return -ENXIO;
 234        }
 235
 236        error = xa_err(xa_store_range(&pgmap_array, PHYS_PFN(range->start),
 237                                PHYS_PFN(range->end), pgmap, GFP_KERNEL));
 238        if (error)
 239                return error;
 240
 241        if (nid < 0)
 242                nid = numa_mem_id();
 243
 244        error = track_pfn_remap(NULL, &params->pgprot, PHYS_PFN(range->start), 0,
 245                        range_len(range));
 246        if (error)
 247                goto err_pfn_remap;
 248
 249        if (!mhp_range_allowed(range->start, range_len(range), !is_private)) {
 250                error = -EINVAL;
 251                goto err_pfn_remap;
 252        }
 253
 254        mem_hotplug_begin();
 255
 256        /*
 257         * For device private memory we call add_pages() as we only need to
 258         * allocate and initialize struct page for the device memory. More-
 259         * over the device memory is un-accessible thus we do not want to
 260         * create a linear mapping for the memory like arch_add_memory()
 261         * would do.
 262         *
 263         * For all other device memory types, which are accessible by
 264         * the CPU, we do want the linear mapping and thus use
 265         * arch_add_memory().
 266         */
 267        if (is_private) {
 268                error = add_pages(nid, PHYS_PFN(range->start),
 269                                PHYS_PFN(range_len(range)), params);
 270        } else {
 271                error = kasan_add_zero_shadow(__va(range->start), range_len(range));
 272                if (error) {
 273                        mem_hotplug_done();
 274                        goto err_kasan;
 275                }
 276
 277                error = arch_add_memory(nid, range->start, range_len(range),
 278                                        params);
 279        }
 280
 281        if (!error) {
 282                struct zone *zone;
 283
 284                zone = &NODE_DATA(nid)->node_zones[ZONE_DEVICE];
 285                move_pfn_range_to_zone(zone, PHYS_PFN(range->start),
 286                                PHYS_PFN(range_len(range)), params->altmap,
 287                                MIGRATE_MOVABLE);
 288        }
 289
 290        mem_hotplug_done();
 291        if (error)
 292                goto err_add_memory;
 293
 294        /*
 295         * Initialization of the pages has been deferred until now in order
 296         * to allow us to do the work while not holding the hotplug lock.
 297         */
 298        memmap_init_zone_device(&NODE_DATA(nid)->node_zones[ZONE_DEVICE],
 299                                PHYS_PFN(range->start),
 300                                PHYS_PFN(range_len(range)), pgmap);
 301        percpu_ref_get_many(pgmap->ref, pfn_end(pgmap, range_id)
 302                        - pfn_first(pgmap, range_id));
 303        return 0;
 304
 305err_add_memory:
 306        kasan_remove_zero_shadow(__va(range->start), range_len(range));
 307err_kasan:
 308        untrack_pfn(NULL, PHYS_PFN(range->start), range_len(range));
 309err_pfn_remap:
 310        pgmap_array_delete(range);
 311        return error;
 312}
 313
 314
 315/*
 316 * Not device managed version of dev_memremap_pages, undone by
 317 * memunmap_pages().  Please use dev_memremap_pages if you have a struct
 318 * device available.
 319 */
 320void *memremap_pages(struct dev_pagemap *pgmap, int nid)
 321{
 322        struct mhp_params params = {
 323                .altmap = pgmap_altmap(pgmap),
 324                .pgprot = PAGE_KERNEL,
 325        };
 326        const int nr_range = pgmap->nr_range;
 327        int error, i;
 328
 329        if (WARN_ONCE(!nr_range, "nr_range must be specified\n"))
 330                return ERR_PTR(-EINVAL);
 331
 332        switch (pgmap->type) {
 333        case MEMORY_DEVICE_PRIVATE:
 334                if (!IS_ENABLED(CONFIG_DEVICE_PRIVATE)) {
 335                        WARN(1, "Device private memory not supported\n");
 336                        return ERR_PTR(-EINVAL);
 337                }
 338                if (!pgmap->ops || !pgmap->ops->migrate_to_ram) {
 339                        WARN(1, "Missing migrate_to_ram method\n");
 340                        return ERR_PTR(-EINVAL);
 341                }
 342                if (!pgmap->ops->page_free) {
 343                        WARN(1, "Missing page_free method\n");
 344                        return ERR_PTR(-EINVAL);
 345                }
 346                if (!pgmap->owner) {
 347                        WARN(1, "Missing owner\n");
 348                        return ERR_PTR(-EINVAL);
 349                }
 350                break;
 351        case MEMORY_DEVICE_FS_DAX:
 352                if (!IS_ENABLED(CONFIG_ZONE_DEVICE) ||
 353                    IS_ENABLED(CONFIG_FS_DAX_LIMITED)) {
 354                        WARN(1, "File system DAX not supported\n");
 355                        return ERR_PTR(-EINVAL);
 356                }
 357                break;
 358        case MEMORY_DEVICE_GENERIC:
 359                break;
 360        case MEMORY_DEVICE_PCI_P2PDMA:
 361                params.pgprot = pgprot_noncached(params.pgprot);
 362                break;
 363        default:
 364                WARN(1, "Invalid pgmap type %d\n", pgmap->type);
 365                break;
 366        }
 367
 368        if (!pgmap->ref) {
 369                if (pgmap->ops && (pgmap->ops->kill || pgmap->ops->cleanup))
 370                        return ERR_PTR(-EINVAL);
 371
 372                init_completion(&pgmap->done);
 373                error = percpu_ref_init(&pgmap->internal_ref,
 374                                dev_pagemap_percpu_release, 0, GFP_KERNEL);
 375                if (error)
 376                        return ERR_PTR(error);
 377                pgmap->ref = &pgmap->internal_ref;
 378        } else {
 379                if (!pgmap->ops || !pgmap->ops->kill || !pgmap->ops->cleanup) {
 380                        WARN(1, "Missing reference count teardown definition\n");
 381                        return ERR_PTR(-EINVAL);
 382                }
 383        }
 384
 385        devmap_managed_enable_get(pgmap);
 386
 387        /*
 388         * Clear the pgmap nr_range as it will be incremented for each
 389         * successfully processed range. This communicates how many
 390         * regions to unwind in the abort case.
 391         */
 392        pgmap->nr_range = 0;
 393        error = 0;
 394        for (i = 0; i < nr_range; i++) {
 395                error = pagemap_range(pgmap, &params, i, nid);
 396                if (error)
 397                        break;
 398                pgmap->nr_range++;
 399        }
 400
 401        if (i < nr_range) {
 402                memunmap_pages(pgmap);
 403                pgmap->nr_range = nr_range;
 404                return ERR_PTR(error);
 405        }
 406
 407        return __va(pgmap->ranges[0].start);
 408}
 409EXPORT_SYMBOL_GPL(memremap_pages);
 410
 411/**
 412 * devm_memremap_pages - remap and provide memmap backing for the given resource
 413 * @dev: hosting device for @res
 414 * @pgmap: pointer to a struct dev_pagemap
 415 *
 416 * Notes:
 417 * 1/ At a minimum the res and type members of @pgmap must be initialized
 418 *    by the caller before passing it to this function
 419 *
 420 * 2/ The altmap field may optionally be initialized, in which case
 421 *    PGMAP_ALTMAP_VALID must be set in pgmap->flags.
 422 *
 423 * 3/ The ref field may optionally be provided, in which pgmap->ref must be
 424 *    'live' on entry and will be killed and reaped at
 425 *    devm_memremap_pages_release() time, or if this routine fails.
 426 *
 427 * 4/ range is expected to be a host memory range that could feasibly be
 428 *    treated as a "System RAM" range, i.e. not a device mmio range, but
 429 *    this is not enforced.
 430 */
 431void *devm_memremap_pages(struct device *dev, struct dev_pagemap *pgmap)
 432{
 433        int error;
 434        void *ret;
 435
 436        ret = memremap_pages(pgmap, dev_to_node(dev));
 437        if (IS_ERR(ret))
 438                return ret;
 439
 440        error = devm_add_action_or_reset(dev, devm_memremap_pages_release,
 441                        pgmap);
 442        if (error)
 443                return ERR_PTR(error);
 444        return ret;
 445}
 446EXPORT_SYMBOL_GPL(devm_memremap_pages);
 447
 448void devm_memunmap_pages(struct device *dev, struct dev_pagemap *pgmap)
 449{
 450        devm_release_action(dev, devm_memremap_pages_release, pgmap);
 451}
 452EXPORT_SYMBOL_GPL(devm_memunmap_pages);
 453
 454unsigned long vmem_altmap_offset(struct vmem_altmap *altmap)
 455{
 456        /* number of pfns from base where pfn_to_page() is valid */
 457        if (altmap)
 458                return altmap->reserve + altmap->free;
 459        return 0;
 460}
 461
 462void vmem_altmap_free(struct vmem_altmap *altmap, unsigned long nr_pfns)
 463{
 464        altmap->alloc -= nr_pfns;
 465}
 466
 467/**
 468 * get_dev_pagemap() - take a new live reference on the dev_pagemap for @pfn
 469 * @pfn: page frame number to lookup page_map
 470 * @pgmap: optional known pgmap that already has a reference
 471 *
 472 * If @pgmap is non-NULL and covers @pfn it will be returned as-is.  If @pgmap
 473 * is non-NULL but does not cover @pfn the reference to it will be released.
 474 */
 475struct dev_pagemap *get_dev_pagemap(unsigned long pfn,
 476                struct dev_pagemap *pgmap)
 477{
 478        resource_size_t phys = PFN_PHYS(pfn);
 479
 480        /*
 481         * In the cached case we're already holding a live reference.
 482         */
 483        if (pgmap) {
 484                if (phys >= pgmap->range.start && phys <= pgmap->range.end)
 485                        return pgmap;
 486                put_dev_pagemap(pgmap);
 487        }
 488
 489        /* fall back to slow path lookup */
 490        rcu_read_lock();
 491        pgmap = xa_load(&pgmap_array, PHYS_PFN(phys));
 492        if (pgmap && !percpu_ref_tryget_live(pgmap->ref))
 493                pgmap = NULL;
 494        rcu_read_unlock();
 495
 496        return pgmap;
 497}
 498EXPORT_SYMBOL_GPL(get_dev_pagemap);
 499
 500#ifdef CONFIG_DEV_PAGEMAP_OPS
 501void free_devmap_managed_page(struct page *page)
 502{
 503        /* notify page idle for dax */
 504        if (!is_device_private_page(page)) {
 505                wake_up_var(&page->_refcount);
 506                return;
 507        }
 508
 509        __ClearPageWaiters(page);
 510
 511        mem_cgroup_uncharge(page);
 512
 513        /*
 514         * When a device_private page is freed, the page->mapping field
 515         * may still contain a (stale) mapping value. For example, the
 516         * lower bits of page->mapping may still identify the page as an
 517         * anonymous page. Ultimately, this entire field is just stale
 518         * and wrong, and it will cause errors if not cleared.  One
 519         * example is:
 520         *
 521         *  migrate_vma_pages()
 522         *    migrate_vma_insert_page()
 523         *      page_add_new_anon_rmap()
 524         *        __page_set_anon_rmap()
 525         *          ...checks page->mapping, via PageAnon(page) call,
 526         *            and incorrectly concludes that the page is an
 527         *            anonymous page. Therefore, it incorrectly,
 528         *            silently fails to set up the new anon rmap.
 529         *
 530         * For other types of ZONE_DEVICE pages, migration is either
 531         * handled differently or not done at all, so there is no need
 532         * to clear page->mapping.
 533         */
 534        page->mapping = NULL;
 535        page->pgmap->ops->page_free(page);
 536}
 537#endif /* CONFIG_DEV_PAGEMAP_OPS */
 538