linux/mm/memory-failure.c
<<
>>
Prefs
   1/*
   2 * Copyright (C) 2008, 2009 Intel Corporation
   3 * Authors: Andi Kleen, Fengguang Wu
   4 *
   5 * This software may be redistributed and/or modified under the terms of
   6 * the GNU General Public License ("GPL") version 2 only as published by the
   7 * Free Software Foundation.
   8 *
   9 * High level machine check handler. Handles pages reported by the
  10 * hardware as being corrupted usually due to a multi-bit ECC memory or cache
  11 * failure.
  12 * 
  13 * In addition there is a "soft offline" entry point that allows stop using
  14 * not-yet-corrupted-by-suspicious pages without killing anything.
  15 *
  16 * Handles page cache pages in various states.  The tricky part
  17 * here is that we can access any page asynchronously in respect to 
  18 * other VM users, because memory failures could happen anytime and 
  19 * anywhere. This could violate some of their assumptions. This is why 
  20 * this code has to be extremely careful. Generally it tries to use 
  21 * normal locking rules, as in get the standard locks, even if that means 
  22 * the error handling takes potentially a long time.
  23 * 
  24 * There are several operations here with exponential complexity because
  25 * of unsuitable VM data structures. For example the operation to map back 
  26 * from RMAP chains to processes has to walk the complete process list and 
  27 * has non linear complexity with the number. But since memory corruptions
  28 * are rare we hope to get away with this. This avoids impacting the core 
  29 * VM.
  30 */
  31
  32/*
  33 * Notebook:
  34 * - hugetlb needs more code
  35 * - kcore/oldmem/vmcore/mem/kmem check for hwpoison pages
  36 * - pass bad pages to kdump next kernel
  37 */
  38#include <linux/kernel.h>
  39#include <linux/mm.h>
  40#include <linux/page-flags.h>
  41#include <linux/kernel-page-flags.h>
  42#include <linux/sched.h>
  43#include <linux/ksm.h>
  44#include <linux/rmap.h>
  45#include <linux/export.h>
  46#include <linux/pagemap.h>
  47#include <linux/swap.h>
  48#include <linux/backing-dev.h>
  49#include <linux/migrate.h>
  50#include <linux/page-isolation.h>
  51#include <linux/suspend.h>
  52#include <linux/slab.h>
  53#include <linux/swapops.h>
  54#include <linux/hugetlb.h>
  55#include <linux/memory_hotplug.h>
  56#include <linux/mm_inline.h>
  57#include <linux/kfifo.h>
  58#include "internal.h"
  59
  60int sysctl_memory_failure_early_kill __read_mostly = 0;
  61
  62int sysctl_memory_failure_recovery __read_mostly = 1;
  63
  64atomic_long_t mce_bad_pages __read_mostly = ATOMIC_LONG_INIT(0);
  65
  66#if defined(CONFIG_HWPOISON_INJECT) || defined(CONFIG_HWPOISON_INJECT_MODULE)
  67
  68u32 hwpoison_filter_enable = 0;
  69u32 hwpoison_filter_dev_major = ~0U;
  70u32 hwpoison_filter_dev_minor = ~0U;
  71u64 hwpoison_filter_flags_mask;
  72u64 hwpoison_filter_flags_value;
  73EXPORT_SYMBOL_GPL(hwpoison_filter_enable);
  74EXPORT_SYMBOL_GPL(hwpoison_filter_dev_major);
  75EXPORT_SYMBOL_GPL(hwpoison_filter_dev_minor);
  76EXPORT_SYMBOL_GPL(hwpoison_filter_flags_mask);
  77EXPORT_SYMBOL_GPL(hwpoison_filter_flags_value);
  78
  79static int hwpoison_filter_dev(struct page *p)
  80{
  81        struct address_space *mapping;
  82        dev_t dev;
  83
  84        if (hwpoison_filter_dev_major == ~0U &&
  85            hwpoison_filter_dev_minor == ~0U)
  86                return 0;
  87
  88        /*
  89         * page_mapping() does not accept slab pages.
  90         */
  91        if (PageSlab(p))
  92                return -EINVAL;
  93
  94        mapping = page_mapping(p);
  95        if (mapping == NULL || mapping->host == NULL)
  96                return -EINVAL;
  97
  98        dev = mapping->host->i_sb->s_dev;
  99        if (hwpoison_filter_dev_major != ~0U &&
 100            hwpoison_filter_dev_major != MAJOR(dev))
 101                return -EINVAL;
 102        if (hwpoison_filter_dev_minor != ~0U &&
 103            hwpoison_filter_dev_minor != MINOR(dev))
 104                return -EINVAL;
 105
 106        return 0;
 107}
 108
 109static int hwpoison_filter_flags(struct page *p)
 110{
 111        if (!hwpoison_filter_flags_mask)
 112                return 0;
 113
 114        if ((stable_page_flags(p) & hwpoison_filter_flags_mask) ==
 115                                    hwpoison_filter_flags_value)
 116                return 0;
 117        else
 118                return -EINVAL;
 119}
 120
 121/*
 122 * This allows stress tests to limit test scope to a collection of tasks
 123 * by putting them under some memcg. This prevents killing unrelated/important
 124 * processes such as /sbin/init. Note that the target task may share clean
 125 * pages with init (eg. libc text), which is harmless. If the target task
 126 * share _dirty_ pages with another task B, the test scheme must make sure B
 127 * is also included in the memcg. At last, due to race conditions this filter
 128 * can only guarantee that the page either belongs to the memcg tasks, or is
 129 * a freed page.
 130 */
 131#ifdef  CONFIG_MEMCG_SWAP
 132u64 hwpoison_filter_memcg;
 133EXPORT_SYMBOL_GPL(hwpoison_filter_memcg);
 134static int hwpoison_filter_task(struct page *p)
 135{
 136        struct mem_cgroup *mem;
 137        struct cgroup_subsys_state *css;
 138        unsigned long ino;
 139
 140        if (!hwpoison_filter_memcg)
 141                return 0;
 142
 143        mem = try_get_mem_cgroup_from_page(p);
 144        if (!mem)
 145                return -EINVAL;
 146
 147        css = mem_cgroup_css(mem);
 148        /* root_mem_cgroup has NULL dentries */
 149        if (!css->cgroup->dentry)
 150                return -EINVAL;
 151
 152        ino = css->cgroup->dentry->d_inode->i_ino;
 153        css_put(css);
 154
 155        if (ino != hwpoison_filter_memcg)
 156                return -EINVAL;
 157
 158        return 0;
 159}
 160#else
 161static int hwpoison_filter_task(struct page *p) { return 0; }
 162#endif
 163
 164int hwpoison_filter(struct page *p)
 165{
 166        if (!hwpoison_filter_enable)
 167                return 0;
 168
 169        if (hwpoison_filter_dev(p))
 170                return -EINVAL;
 171
 172        if (hwpoison_filter_flags(p))
 173                return -EINVAL;
 174
 175        if (hwpoison_filter_task(p))
 176                return -EINVAL;
 177
 178        return 0;
 179}
 180#else
 181int hwpoison_filter(struct page *p)
 182{
 183        return 0;
 184}
 185#endif
 186
 187EXPORT_SYMBOL_GPL(hwpoison_filter);
 188
 189/*
 190 * Send all the processes who have the page mapped a signal.
 191 * ``action optional'' if they are not immediately affected by the error
 192 * ``action required'' if error happened in current execution context
 193 */
 194static int kill_proc(struct task_struct *t, unsigned long addr, int trapno,
 195                        unsigned long pfn, struct page *page, int flags)
 196{
 197        struct siginfo si;
 198        int ret;
 199
 200        printk(KERN_ERR
 201                "MCE %#lx: Killing %s:%d due to hardware memory corruption\n",
 202                pfn, t->comm, t->pid);
 203        si.si_signo = SIGBUS;
 204        si.si_errno = 0;
 205        si.si_addr = (void *)addr;
 206#ifdef __ARCH_SI_TRAPNO
 207        si.si_trapno = trapno;
 208#endif
 209        si.si_addr_lsb = compound_trans_order(compound_head(page)) + PAGE_SHIFT;
 210
 211        if ((flags & MF_ACTION_REQUIRED) && t == current) {
 212                si.si_code = BUS_MCEERR_AR;
 213                ret = force_sig_info(SIGBUS, &si, t);
 214        } else {
 215                /*
 216                 * Don't use force here, it's convenient if the signal
 217                 * can be temporarily blocked.
 218                 * This could cause a loop when the user sets SIGBUS
 219                 * to SIG_IGN, but hopefully no one will do that?
 220                 */
 221                si.si_code = BUS_MCEERR_AO;
 222                ret = send_sig_info(SIGBUS, &si, t);  /* synchronous? */
 223        }
 224        if (ret < 0)
 225                printk(KERN_INFO "MCE: Error sending signal to %s:%d: %d\n",
 226                       t->comm, t->pid, ret);
 227        return ret;
 228}
 229
 230/*
 231 * When a unknown page type is encountered drain as many buffers as possible
 232 * in the hope to turn the page into a LRU or free page, which we can handle.
 233 */
 234void shake_page(struct page *p, int access)
 235{
 236        if (!PageSlab(p)) {
 237                lru_add_drain_all();
 238                if (PageLRU(p))
 239                        return;
 240                drain_all_pages();
 241                if (PageLRU(p) || is_free_buddy_page(p))
 242                        return;
 243        }
 244
 245        /*
 246         * Only call shrink_slab here (which would also shrink other caches) if
 247         * access is not potentially fatal.
 248         */
 249        if (access) {
 250                int nr;
 251                do {
 252                        struct shrink_control shrink = {
 253                                .gfp_mask = GFP_KERNEL,
 254                        };
 255
 256                        nr = shrink_slab(&shrink, 1000, 1000);
 257                        if (page_count(p) == 1)
 258                                break;
 259                } while (nr > 10);
 260        }
 261}
 262EXPORT_SYMBOL_GPL(shake_page);
 263
 264/*
 265 * Kill all processes that have a poisoned page mapped and then isolate
 266 * the page.
 267 *
 268 * General strategy:
 269 * Find all processes having the page mapped and kill them.
 270 * But we keep a page reference around so that the page is not
 271 * actually freed yet.
 272 * Then stash the page away
 273 *
 274 * There's no convenient way to get back to mapped processes
 275 * from the VMAs. So do a brute-force search over all
 276 * running processes.
 277 *
 278 * Remember that machine checks are not common (or rather
 279 * if they are common you have other problems), so this shouldn't
 280 * be a performance issue.
 281 *
 282 * Also there are some races possible while we get from the
 283 * error detection to actually handle it.
 284 */
 285
 286struct to_kill {
 287        struct list_head nd;
 288        struct task_struct *tsk;
 289        unsigned long addr;
 290        char addr_valid;
 291};
 292
 293/*
 294 * Failure handling: if we can't find or can't kill a process there's
 295 * not much we can do.  We just print a message and ignore otherwise.
 296 */
 297
 298/*
 299 * Schedule a process for later kill.
 300 * Uses GFP_ATOMIC allocations to avoid potential recursions in the VM.
 301 * TBD would GFP_NOIO be enough?
 302 */
 303static void add_to_kill(struct task_struct *tsk, struct page *p,
 304                       struct vm_area_struct *vma,
 305                       struct list_head *to_kill,
 306                       struct to_kill **tkc)
 307{
 308        struct to_kill *tk;
 309
 310        if (*tkc) {
 311                tk = *tkc;
 312                *tkc = NULL;
 313        } else {
 314                tk = kmalloc(sizeof(struct to_kill), GFP_ATOMIC);
 315                if (!tk) {
 316                        printk(KERN_ERR
 317                "MCE: Out of memory while machine check handling\n");
 318                        return;
 319                }
 320        }
 321        tk->addr = page_address_in_vma(p, vma);
 322        tk->addr_valid = 1;
 323
 324        /*
 325         * In theory we don't have to kill when the page was
 326         * munmaped. But it could be also a mremap. Since that's
 327         * likely very rare kill anyways just out of paranoia, but use
 328         * a SIGKILL because the error is not contained anymore.
 329         */
 330        if (tk->addr == -EFAULT) {
 331                pr_info("MCE: Unable to find user space address %lx in %s\n",
 332                        page_to_pfn(p), tsk->comm);
 333                tk->addr_valid = 0;
 334        }
 335        get_task_struct(tsk);
 336        tk->tsk = tsk;
 337        list_add_tail(&tk->nd, to_kill);
 338}
 339
 340/*
 341 * Kill the processes that have been collected earlier.
 342 *
 343 * Only do anything when DOIT is set, otherwise just free the list
 344 * (this is used for clean pages which do not need killing)
 345 * Also when FAIL is set do a force kill because something went
 346 * wrong earlier.
 347 */
 348static void kill_procs(struct list_head *to_kill, int forcekill, int trapno,
 349                          int fail, struct page *page, unsigned long pfn,
 350                          int flags)
 351{
 352        struct to_kill *tk, *next;
 353
 354        list_for_each_entry_safe (tk, next, to_kill, nd) {
 355                if (forcekill) {
 356                        /*
 357                         * In case something went wrong with munmapping
 358                         * make sure the process doesn't catch the
 359                         * signal and then access the memory. Just kill it.
 360                         */
 361                        if (fail || tk->addr_valid == 0) {
 362                                printk(KERN_ERR
 363                "MCE %#lx: forcibly killing %s:%d because of failure to unmap corrupted page\n",
 364                                        pfn, tk->tsk->comm, tk->tsk->pid);
 365                                force_sig(SIGKILL, tk->tsk);
 366                        }
 367
 368                        /*
 369                         * In theory the process could have mapped
 370                         * something else on the address in-between. We could
 371                         * check for that, but we need to tell the
 372                         * process anyways.
 373                         */
 374                        else if (kill_proc(tk->tsk, tk->addr, trapno,
 375                                              pfn, page, flags) < 0)
 376                                printk(KERN_ERR
 377                "MCE %#lx: Cannot send advisory machine check signal to %s:%d\n",
 378                                        pfn, tk->tsk->comm, tk->tsk->pid);
 379                }
 380                put_task_struct(tk->tsk);
 381                kfree(tk);
 382        }
 383}
 384
 385static int task_early_kill(struct task_struct *tsk)
 386{
 387        if (!tsk->mm)
 388                return 0;
 389        if (tsk->flags & PF_MCE_PROCESS)
 390                return !!(tsk->flags & PF_MCE_EARLY);
 391        return sysctl_memory_failure_early_kill;
 392}
 393
 394/*
 395 * Collect processes when the error hit an anonymous page.
 396 */
 397static void collect_procs_anon(struct page *page, struct list_head *to_kill,
 398                              struct to_kill **tkc)
 399{
 400        struct vm_area_struct *vma;
 401        struct task_struct *tsk;
 402        struct anon_vma *av;
 403
 404        av = page_lock_anon_vma(page);
 405        if (av == NULL) /* Not actually mapped anymore */
 406                return;
 407
 408        read_lock(&tasklist_lock);
 409        for_each_process (tsk) {
 410                struct anon_vma_chain *vmac;
 411
 412                if (!task_early_kill(tsk))
 413                        continue;
 414                list_for_each_entry(vmac, &av->head, same_anon_vma) {
 415                        vma = vmac->vma;
 416                        if (!page_mapped_in_vma(page, vma))
 417                                continue;
 418                        if (vma->vm_mm == tsk->mm)
 419                                add_to_kill(tsk, page, vma, to_kill, tkc);
 420                }
 421        }
 422        read_unlock(&tasklist_lock);
 423        page_unlock_anon_vma(av);
 424}
 425
 426/*
 427 * Collect processes when the error hit a file mapped page.
 428 */
 429static void collect_procs_file(struct page *page, struct list_head *to_kill,
 430                              struct to_kill **tkc)
 431{
 432        struct vm_area_struct *vma;
 433        struct task_struct *tsk;
 434        struct prio_tree_iter iter;
 435        struct address_space *mapping = page->mapping;
 436
 437        mutex_lock(&mapping->i_mmap_mutex);
 438        read_lock(&tasklist_lock);
 439        for_each_process(tsk) {
 440                pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
 441
 442                if (!task_early_kill(tsk))
 443                        continue;
 444
 445                vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff,
 446                                      pgoff) {
 447                        /*
 448                         * Send early kill signal to tasks where a vma covers
 449                         * the page but the corrupted page is not necessarily
 450                         * mapped it in its pte.
 451                         * Assume applications who requested early kill want
 452                         * to be informed of all such data corruptions.
 453                         */
 454                        if (vma->vm_mm == tsk->mm)
 455                                add_to_kill(tsk, page, vma, to_kill, tkc);
 456                }
 457        }
 458        read_unlock(&tasklist_lock);
 459        mutex_unlock(&mapping->i_mmap_mutex);
 460}
 461
 462/*
 463 * Collect the processes who have the corrupted page mapped to kill.
 464 * This is done in two steps for locking reasons.
 465 * First preallocate one tokill structure outside the spin locks,
 466 * so that we can kill at least one process reasonably reliable.
 467 */
 468static void collect_procs(struct page *page, struct list_head *tokill)
 469{
 470        struct to_kill *tk;
 471
 472        if (!page->mapping)
 473                return;
 474
 475        tk = kmalloc(sizeof(struct to_kill), GFP_NOIO);
 476        if (!tk)
 477                return;
 478        if (PageAnon(page))
 479                collect_procs_anon(page, tokill, &tk);
 480        else
 481                collect_procs_file(page, tokill, &tk);
 482        kfree(tk);
 483}
 484
 485/*
 486 * Error handlers for various types of pages.
 487 */
 488
 489enum outcome {
 490        IGNORED,        /* Error: cannot be handled */
 491        FAILED,         /* Error: handling failed */
 492        DELAYED,        /* Will be handled later */
 493        RECOVERED,      /* Successfully recovered */
 494};
 495
 496static const char *action_name[] = {
 497        [IGNORED] = "Ignored",
 498        [FAILED] = "Failed",
 499        [DELAYED] = "Delayed",
 500        [RECOVERED] = "Recovered",
 501};
 502
 503/*
 504 * XXX: It is possible that a page is isolated from LRU cache,
 505 * and then kept in swap cache or failed to remove from page cache.
 506 * The page count will stop it from being freed by unpoison.
 507 * Stress tests should be aware of this memory leak problem.
 508 */
 509static int delete_from_lru_cache(struct page *p)
 510{
 511        if (!isolate_lru_page(p)) {
 512                /*
 513                 * Clear sensible page flags, so that the buddy system won't
 514                 * complain when the page is unpoison-and-freed.
 515                 */
 516                ClearPageActive(p);
 517                ClearPageUnevictable(p);
 518                /*
 519                 * drop the page count elevated by isolate_lru_page()
 520                 */
 521                page_cache_release(p);
 522                return 0;
 523        }
 524        return -EIO;
 525}
 526
 527/*
 528 * Error hit kernel page.
 529 * Do nothing, try to be lucky and not touch this instead. For a few cases we
 530 * could be more sophisticated.
 531 */
 532static int me_kernel(struct page *p, unsigned long pfn)
 533{
 534        return IGNORED;
 535}
 536
 537/*
 538 * Page in unknown state. Do nothing.
 539 */
 540static int me_unknown(struct page *p, unsigned long pfn)
 541{
 542        printk(KERN_ERR "MCE %#lx: Unknown page state\n", pfn);
 543        return FAILED;
 544}
 545
 546/*
 547 * Clean (or cleaned) page cache page.
 548 */
 549static int me_pagecache_clean(struct page *p, unsigned long pfn)
 550{
 551        int err;
 552        int ret = FAILED;
 553        struct address_space *mapping;
 554
 555        delete_from_lru_cache(p);
 556
 557        /*
 558         * For anonymous pages we're done the only reference left
 559         * should be the one m_f() holds.
 560         */
 561        if (PageAnon(p))
 562                return RECOVERED;
 563
 564        /*
 565         * Now truncate the page in the page cache. This is really
 566         * more like a "temporary hole punch"
 567         * Don't do this for block devices when someone else
 568         * has a reference, because it could be file system metadata
 569         * and that's not safe to truncate.
 570         */
 571        mapping = page_mapping(p);
 572        if (!mapping) {
 573                /*
 574                 * Page has been teared down in the meanwhile
 575                 */
 576                return FAILED;
 577        }
 578
 579        /*
 580         * Truncation is a bit tricky. Enable it per file system for now.
 581         *
 582         * Open: to take i_mutex or not for this? Right now we don't.
 583         */
 584        if (mapping->a_ops->error_remove_page) {
 585                err = mapping->a_ops->error_remove_page(mapping, p);
 586                if (err != 0) {
 587                        printk(KERN_INFO "MCE %#lx: Failed to punch page: %d\n",
 588                                        pfn, err);
 589                } else if (page_has_private(p) &&
 590                                !try_to_release_page(p, GFP_NOIO)) {
 591                        pr_info("MCE %#lx: failed to release buffers\n", pfn);
 592                } else {
 593                        ret = RECOVERED;
 594                }
 595        } else {
 596                /*
 597                 * If the file system doesn't support it just invalidate
 598                 * This fails on dirty or anything with private pages
 599                 */
 600                if (invalidate_inode_page(p))
 601                        ret = RECOVERED;
 602                else
 603                        printk(KERN_INFO "MCE %#lx: Failed to invalidate\n",
 604                                pfn);
 605        }
 606        return ret;
 607}
 608
 609/*
 610 * Dirty cache page page
 611 * Issues: when the error hit a hole page the error is not properly
 612 * propagated.
 613 */
 614static int me_pagecache_dirty(struct page *p, unsigned long pfn)
 615{
 616        struct address_space *mapping = page_mapping(p);
 617
 618        SetPageError(p);
 619        /* TBD: print more information about the file. */
 620        if (mapping) {
 621                /*
 622                 * IO error will be reported by write(), fsync(), etc.
 623                 * who check the mapping.
 624                 * This way the application knows that something went
 625                 * wrong with its dirty file data.
 626                 *
 627                 * There's one open issue:
 628                 *
 629                 * The EIO will be only reported on the next IO
 630                 * operation and then cleared through the IO map.
 631                 * Normally Linux has two mechanisms to pass IO error
 632                 * first through the AS_EIO flag in the address space
 633                 * and then through the PageError flag in the page.
 634                 * Since we drop pages on memory failure handling the
 635                 * only mechanism open to use is through AS_AIO.
 636                 *
 637                 * This has the disadvantage that it gets cleared on
 638                 * the first operation that returns an error, while
 639                 * the PageError bit is more sticky and only cleared
 640                 * when the page is reread or dropped.  If an
 641                 * application assumes it will always get error on
 642                 * fsync, but does other operations on the fd before
 643                 * and the page is dropped between then the error
 644                 * will not be properly reported.
 645                 *
 646                 * This can already happen even without hwpoisoned
 647                 * pages: first on metadata IO errors (which only
 648                 * report through AS_EIO) or when the page is dropped
 649                 * at the wrong time.
 650                 *
 651                 * So right now we assume that the application DTRT on
 652                 * the first EIO, but we're not worse than other parts
 653                 * of the kernel.
 654                 */
 655                mapping_set_error(mapping, EIO);
 656        }
 657
 658        return me_pagecache_clean(p, pfn);
 659}
 660
 661/*
 662 * Clean and dirty swap cache.
 663 *
 664 * Dirty swap cache page is tricky to handle. The page could live both in page
 665 * cache and swap cache(ie. page is freshly swapped in). So it could be
 666 * referenced concurrently by 2 types of PTEs:
 667 * normal PTEs and swap PTEs. We try to handle them consistently by calling
 668 * try_to_unmap(TTU_IGNORE_HWPOISON) to convert the normal PTEs to swap PTEs,
 669 * and then
 670 *      - clear dirty bit to prevent IO
 671 *      - remove from LRU
 672 *      - but keep in the swap cache, so that when we return to it on
 673 *        a later page fault, we know the application is accessing
 674 *        corrupted data and shall be killed (we installed simple
 675 *        interception code in do_swap_page to catch it).
 676 *
 677 * Clean swap cache pages can be directly isolated. A later page fault will
 678 * bring in the known good data from disk.
 679 */
 680static int me_swapcache_dirty(struct page *p, unsigned long pfn)
 681{
 682        ClearPageDirty(p);
 683        /* Trigger EIO in shmem: */
 684        ClearPageUptodate(p);
 685
 686        if (!delete_from_lru_cache(p))
 687                return DELAYED;
 688        else
 689                return FAILED;
 690}
 691
 692static int me_swapcache_clean(struct page *p, unsigned long pfn)
 693{
 694        delete_from_swap_cache(p);
 695
 696        if (!delete_from_lru_cache(p))
 697                return RECOVERED;
 698        else
 699                return FAILED;
 700}
 701
 702/*
 703 * Huge pages. Needs work.
 704 * Issues:
 705 * - Error on hugepage is contained in hugepage unit (not in raw page unit.)
 706 *   To narrow down kill region to one page, we need to break up pmd.
 707 */
 708static int me_huge_page(struct page *p, unsigned long pfn)
 709{
 710        int res = 0;
 711        struct page *hpage = compound_head(p);
 712        /*
 713         * We can safely recover from error on free or reserved (i.e.
 714         * not in-use) hugepage by dequeuing it from freelist.
 715         * To check whether a hugepage is in-use or not, we can't use
 716         * page->lru because it can be used in other hugepage operations,
 717         * such as __unmap_hugepage_range() and gather_surplus_pages().
 718         * So instead we use page_mapping() and PageAnon().
 719         * We assume that this function is called with page lock held,
 720         * so there is no race between isolation and mapping/unmapping.
 721         */
 722        if (!(page_mapping(hpage) || PageAnon(hpage))) {
 723                res = dequeue_hwpoisoned_huge_page(hpage);
 724                if (!res)
 725                        return RECOVERED;
 726        }
 727        return DELAYED;
 728}
 729
 730/*
 731 * Various page states we can handle.
 732 *
 733 * A page state is defined by its current page->flags bits.
 734 * The table matches them in order and calls the right handler.
 735 *
 736 * This is quite tricky because we can access page at any time
 737 * in its live cycle, so all accesses have to be extremely careful.
 738 *
 739 * This is not complete. More states could be added.
 740 * For any missing state don't attempt recovery.
 741 */
 742
 743#define dirty           (1UL << PG_dirty)
 744#define sc              (1UL << PG_swapcache)
 745#define unevict         (1UL << PG_unevictable)
 746#define mlock           (1UL << PG_mlocked)
 747#define writeback       (1UL << PG_writeback)
 748#define lru             (1UL << PG_lru)
 749#define swapbacked      (1UL << PG_swapbacked)
 750#define head            (1UL << PG_head)
 751#define tail            (1UL << PG_tail)
 752#define compound        (1UL << PG_compound)
 753#define slab            (1UL << PG_slab)
 754#define reserved        (1UL << PG_reserved)
 755
 756static struct page_state {
 757        unsigned long mask;
 758        unsigned long res;
 759        char *msg;
 760        int (*action)(struct page *p, unsigned long pfn);
 761} error_states[] = {
 762        { reserved,     reserved,       "reserved kernel",      me_kernel },
 763        /*
 764         * free pages are specially detected outside this table:
 765         * PG_buddy pages only make a small fraction of all free pages.
 766         */
 767
 768        /*
 769         * Could in theory check if slab page is free or if we can drop
 770         * currently unused objects without touching them. But just
 771         * treat it as standard kernel for now.
 772         */
 773        { slab,         slab,           "kernel slab",  me_kernel },
 774
 775#ifdef CONFIG_PAGEFLAGS_EXTENDED
 776        { head,         head,           "huge",         me_huge_page },
 777        { tail,         tail,           "huge",         me_huge_page },
 778#else
 779        { compound,     compound,       "huge",         me_huge_page },
 780#endif
 781
 782        { sc|dirty,     sc|dirty,       "swapcache",    me_swapcache_dirty },
 783        { sc|dirty,     sc,             "swapcache",    me_swapcache_clean },
 784
 785        { unevict|dirty, unevict|dirty, "unevictable LRU", me_pagecache_dirty},
 786        { unevict,      unevict,        "unevictable LRU", me_pagecache_clean},
 787
 788        { mlock|dirty,  mlock|dirty,    "mlocked LRU",  me_pagecache_dirty },
 789        { mlock,        mlock,          "mlocked LRU",  me_pagecache_clean },
 790
 791        { lru|dirty,    lru|dirty,      "LRU",          me_pagecache_dirty },
 792        { lru|dirty,    lru,            "clean LRU",    me_pagecache_clean },
 793
 794        /*
 795         * Catchall entry: must be at end.
 796         */
 797        { 0,            0,              "unknown page state",   me_unknown },
 798};
 799
 800#undef dirty
 801#undef sc
 802#undef unevict
 803#undef mlock
 804#undef writeback
 805#undef lru
 806#undef swapbacked
 807#undef head
 808#undef tail
 809#undef compound
 810#undef slab
 811#undef reserved
 812
 813static void action_result(unsigned long pfn, char *msg, int result)
 814{
 815        struct page *page = pfn_to_page(pfn);
 816
 817        printk(KERN_ERR "MCE %#lx: %s%s page recovery: %s\n",
 818                pfn,
 819                PageDirty(page) ? "dirty " : "",
 820                msg, action_name[result]);
 821}
 822
 823static int page_action(struct page_state *ps, struct page *p,
 824                        unsigned long pfn)
 825{
 826        int result;
 827        int count;
 828
 829        result = ps->action(p, pfn);
 830        action_result(pfn, ps->msg, result);
 831
 832        count = page_count(p) - 1;
 833        if (ps->action == me_swapcache_dirty && result == DELAYED)
 834                count--;
 835        if (count != 0) {
 836                printk(KERN_ERR
 837                       "MCE %#lx: %s page still referenced by %d users\n",
 838                       pfn, ps->msg, count);
 839                result = FAILED;
 840        }
 841
 842        /* Could do more checks here if page looks ok */
 843        /*
 844         * Could adjust zone counters here to correct for the missing page.
 845         */
 846
 847        return (result == RECOVERED || result == DELAYED) ? 0 : -EBUSY;
 848}
 849
 850/*
 851 * Do all that is necessary to remove user space mappings. Unmap
 852 * the pages and send SIGBUS to the processes if the data was dirty.
 853 */
 854static int hwpoison_user_mappings(struct page *p, unsigned long pfn,
 855                                  int trapno, int flags)
 856{
 857        enum ttu_flags ttu = TTU_UNMAP | TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS;
 858        struct address_space *mapping;
 859        LIST_HEAD(tokill);
 860        int ret;
 861        int kill = 1, forcekill;
 862        struct page *hpage = compound_head(p);
 863        struct page *ppage;
 864
 865        if (PageReserved(p) || PageSlab(p))
 866                return SWAP_SUCCESS;
 867
 868        /*
 869         * This check implies we don't kill processes if their pages
 870         * are in the swap cache early. Those are always late kills.
 871         */
 872        if (!page_mapped(hpage))
 873                return SWAP_SUCCESS;
 874
 875        if (PageKsm(p))
 876                return SWAP_FAIL;
 877
 878        if (PageSwapCache(p)) {
 879                printk(KERN_ERR
 880                       "MCE %#lx: keeping poisoned page in swap cache\n", pfn);
 881                ttu |= TTU_IGNORE_HWPOISON;
 882        }
 883
 884        /*
 885         * Propagate the dirty bit from PTEs to struct page first, because we
 886         * need this to decide if we should kill or just drop the page.
 887         * XXX: the dirty test could be racy: set_page_dirty() may not always
 888         * be called inside page lock (it's recommended but not enforced).
 889         */
 890        mapping = page_mapping(hpage);
 891        if (!(flags & MF_MUST_KILL) && !PageDirty(hpage) && mapping &&
 892            mapping_cap_writeback_dirty(mapping)) {
 893                if (page_mkclean(hpage)) {
 894                        SetPageDirty(hpage);
 895                } else {
 896                        kill = 0;
 897                        ttu |= TTU_IGNORE_HWPOISON;
 898                        printk(KERN_INFO
 899        "MCE %#lx: corrupted page was clean: dropped without side effects\n",
 900                                pfn);
 901                }
 902        }
 903
 904        /*
 905         * ppage: poisoned page
 906         *   if p is regular page(4k page)
 907         *        ppage == real poisoned page;
 908         *   else p is hugetlb or THP, ppage == head page.
 909         */
 910        ppage = hpage;
 911
 912        if (PageTransHuge(hpage)) {
 913                /*
 914                 * Verify that this isn't a hugetlbfs head page, the check for
 915                 * PageAnon is just for avoid tripping a split_huge_page
 916                 * internal debug check, as split_huge_page refuses to deal with
 917                 * anything that isn't an anon page. PageAnon can't go away fro
 918                 * under us because we hold a refcount on the hpage, without a
 919                 * refcount on the hpage. split_huge_page can't be safely called
 920                 * in the first place, having a refcount on the tail isn't
 921                 * enough * to be safe.
 922                 */
 923                if (!PageHuge(hpage) && PageAnon(hpage)) {
 924                        if (unlikely(split_huge_page(hpage))) {
 925                                /*
 926                                 * FIXME: if splitting THP is failed, it is
 927                                 * better to stop the following operation rather
 928                                 * than causing panic by unmapping. System might
 929                                 * survive if the page is freed later.
 930                                 */
 931                                printk(KERN_INFO
 932                                        "MCE %#lx: failed to split THP\n", pfn);
 933
 934                                BUG_ON(!PageHWPoison(p));
 935                                return SWAP_FAIL;
 936                        }
 937                        /* THP is split, so ppage should be the real poisoned page. */
 938                        ppage = p;
 939                }
 940        }
 941
 942        /*
 943         * First collect all the processes that have the page
 944         * mapped in dirty form.  This has to be done before try_to_unmap,
 945         * because ttu takes the rmap data structures down.
 946         *
 947         * Error handling: We ignore errors here because
 948         * there's nothing that can be done.
 949         */
 950        if (kill)
 951                collect_procs(ppage, &tokill);
 952
 953        if (hpage != ppage)
 954                lock_page(ppage);
 955
 956        ret = try_to_unmap(ppage, ttu);
 957        if (ret != SWAP_SUCCESS)
 958                printk(KERN_ERR "MCE %#lx: failed to unmap page (mapcount=%d)\n",
 959                                pfn, page_mapcount(ppage));
 960
 961        if (hpage != ppage)
 962                unlock_page(ppage);
 963
 964        /*
 965         * Now that the dirty bit has been propagated to the
 966         * struct page and all unmaps done we can decide if
 967         * killing is needed or not.  Only kill when the page
 968         * was dirty or the process is not restartable,
 969         * otherwise the tokill list is merely
 970         * freed.  When there was a problem unmapping earlier
 971         * use a more force-full uncatchable kill to prevent
 972         * any accesses to the poisoned memory.
 973         */
 974        forcekill = PageDirty(ppage) || (flags & MF_MUST_KILL);
 975        kill_procs(&tokill, forcekill, trapno,
 976                      ret != SWAP_SUCCESS, p, pfn, flags);
 977
 978        return ret;
 979}
 980
 981static void set_page_hwpoison_huge_page(struct page *hpage)
 982{
 983        int i;
 984        int nr_pages = 1 << compound_trans_order(hpage);
 985        for (i = 0; i < nr_pages; i++)
 986                SetPageHWPoison(hpage + i);
 987}
 988
 989static void clear_page_hwpoison_huge_page(struct page *hpage)
 990{
 991        int i;
 992        int nr_pages = 1 << compound_trans_order(hpage);
 993        for (i = 0; i < nr_pages; i++)
 994                ClearPageHWPoison(hpage + i);
 995}
 996
 997/**
 998 * memory_failure - Handle memory failure of a page.
 999 * @pfn: Page Number of the corrupted page
1000 * @trapno: Trap number reported in the signal to user space.
1001 * @flags: fine tune action taken
1002 *
1003 * This function is called by the low level machine check code
1004 * of an architecture when it detects hardware memory corruption
1005 * of a page. It tries its best to recover, which includes
1006 * dropping pages, killing processes etc.
1007 *
1008 * The function is primarily of use for corruptions that
1009 * happen outside the current execution context (e.g. when
1010 * detected by a background scrubber)
1011 *
1012 * Must run in process context (e.g. a work queue) with interrupts
1013 * enabled and no spinlocks hold.
1014 */
1015int memory_failure(unsigned long pfn, int trapno, int flags)
1016{
1017        struct page_state *ps;
1018        struct page *p;
1019        struct page *hpage;
1020        int res;
1021        unsigned int nr_pages;
1022
1023        if (!sysctl_memory_failure_recovery)
1024                panic("Memory failure from trap %d on page %lx", trapno, pfn);
1025
1026        if (!pfn_valid(pfn)) {
1027                printk(KERN_ERR
1028                       "MCE %#lx: memory outside kernel control\n",
1029                       pfn);
1030                return -ENXIO;
1031        }
1032
1033        p = pfn_to_page(pfn);
1034        hpage = compound_head(p);
1035        if (TestSetPageHWPoison(p)) {
1036                printk(KERN_ERR "MCE %#lx: already hardware poisoned\n", pfn);
1037                return 0;
1038        }
1039
1040        nr_pages = 1 << compound_trans_order(hpage);
1041        atomic_long_add(nr_pages, &mce_bad_pages);
1042
1043        /*
1044         * We need/can do nothing about count=0 pages.
1045         * 1) it's a free page, and therefore in safe hand:
1046         *    prep_new_page() will be the gate keeper.
1047         * 2) it's a free hugepage, which is also safe:
1048         *    an affected hugepage will be dequeued from hugepage freelist,
1049         *    so there's no concern about reusing it ever after.
1050         * 3) it's part of a non-compound high order page.
1051         *    Implies some kernel user: cannot stop them from
1052         *    R/W the page; let's pray that the page has been
1053         *    used and will be freed some time later.
1054         * In fact it's dangerous to directly bump up page count from 0,
1055         * that may make page_freeze_refs()/page_unfreeze_refs() mismatch.
1056         */
1057        if (!(flags & MF_COUNT_INCREASED) &&
1058                !get_page_unless_zero(hpage)) {
1059                if (is_free_buddy_page(p)) {
1060                        action_result(pfn, "free buddy", DELAYED);
1061                        return 0;
1062                } else if (PageHuge(hpage)) {
1063                        /*
1064                         * Check "just unpoisoned", "filter hit", and
1065                         * "race with other subpage."
1066                         */
1067                        lock_page(hpage);
1068                        if (!PageHWPoison(hpage)
1069                            || (hwpoison_filter(p) && TestClearPageHWPoison(p))
1070                            || (p != hpage && TestSetPageHWPoison(hpage))) {
1071                                atomic_long_sub(nr_pages, &mce_bad_pages);
1072                                return 0;
1073                        }
1074                        set_page_hwpoison_huge_page(hpage);
1075                        res = dequeue_hwpoisoned_huge_page(hpage);
1076                        action_result(pfn, "free huge",
1077                                      res ? IGNORED : DELAYED);
1078                        unlock_page(hpage);
1079                        return res;
1080                } else {
1081                        action_result(pfn, "high order kernel", IGNORED);
1082                        return -EBUSY;
1083                }
1084        }
1085
1086        /*
1087         * We ignore non-LRU pages for good reasons.
1088         * - PG_locked is only well defined for LRU pages and a few others
1089         * - to avoid races with __set_page_locked()
1090         * - to avoid races with __SetPageSlab*() (and more non-atomic ops)
1091         * The check (unnecessarily) ignores LRU pages being isolated and
1092         * walked by the page reclaim code, however that's not a big loss.
1093         */
1094        if (!PageHuge(p) && !PageTransTail(p)) {
1095                if (!PageLRU(p))
1096                        shake_page(p, 0);
1097                if (!PageLRU(p)) {
1098                        /*
1099                         * shake_page could have turned it free.
1100                         */
1101                        if (is_free_buddy_page(p)) {
1102                                action_result(pfn, "free buddy, 2nd try",
1103                                                DELAYED);
1104                                return 0;
1105                        }
1106                        action_result(pfn, "non LRU", IGNORED);
1107                        put_page(p);
1108                        return -EBUSY;
1109                }
1110        }
1111
1112        /*
1113         * Lock the page and wait for writeback to finish.
1114         * It's very difficult to mess with pages currently under IO
1115         * and in many cases impossible, so we just avoid it here.
1116         */
1117        lock_page(hpage);
1118
1119        /*
1120         * unpoison always clear PG_hwpoison inside page lock
1121         */
1122        if (!PageHWPoison(p)) {
1123                printk(KERN_ERR "MCE %#lx: just unpoisoned\n", pfn);
1124                res = 0;
1125                goto out;
1126        }
1127        if (hwpoison_filter(p)) {
1128                if (TestClearPageHWPoison(p))
1129                        atomic_long_sub(nr_pages, &mce_bad_pages);
1130                unlock_page(hpage);
1131                put_page(hpage);
1132                return 0;
1133        }
1134
1135        /*
1136         * For error on the tail page, we should set PG_hwpoison
1137         * on the head page to show that the hugepage is hwpoisoned
1138         */
1139        if (PageHuge(p) && PageTail(p) && TestSetPageHWPoison(hpage)) {
1140                action_result(pfn, "hugepage already hardware poisoned",
1141                                IGNORED);
1142                unlock_page(hpage);
1143                put_page(hpage);
1144                return 0;
1145        }
1146        /*
1147         * Set PG_hwpoison on all pages in an error hugepage,
1148         * because containment is done in hugepage unit for now.
1149         * Since we have done TestSetPageHWPoison() for the head page with
1150         * page lock held, we can safely set PG_hwpoison bits on tail pages.
1151         */
1152        if (PageHuge(p))
1153                set_page_hwpoison_huge_page(hpage);
1154
1155        wait_on_page_writeback(p);
1156
1157        /*
1158         * Now take care of user space mappings.
1159         * Abort on fail: __delete_from_page_cache() assumes unmapped page.
1160         */
1161        if (hwpoison_user_mappings(p, pfn, trapno, flags) != SWAP_SUCCESS) {
1162                printk(KERN_ERR "MCE %#lx: cannot unmap page, give up\n", pfn);
1163                res = -EBUSY;
1164                goto out;
1165        }
1166
1167        /*
1168         * Torn down by someone else?
1169         */
1170        if (PageLRU(p) && !PageSwapCache(p) && p->mapping == NULL) {
1171                action_result(pfn, "already truncated LRU", IGNORED);
1172                res = -EBUSY;
1173                goto out;
1174        }
1175
1176        res = -EBUSY;
1177        for (ps = error_states;; ps++) {
1178                if ((p->flags & ps->mask) == ps->res) {
1179                        res = page_action(ps, p, pfn);
1180                        break;
1181                }
1182        }
1183out:
1184        unlock_page(hpage);
1185        return res;
1186}
1187EXPORT_SYMBOL_GPL(memory_failure);
1188
1189#define MEMORY_FAILURE_FIFO_ORDER       4
1190#define MEMORY_FAILURE_FIFO_SIZE        (1 << MEMORY_FAILURE_FIFO_ORDER)
1191
1192struct memory_failure_entry {
1193        unsigned long pfn;
1194        int trapno;
1195        int flags;
1196};
1197
1198struct memory_failure_cpu {
1199        DECLARE_KFIFO(fifo, struct memory_failure_entry,
1200                      MEMORY_FAILURE_FIFO_SIZE);
1201        spinlock_t lock;
1202        struct work_struct work;
1203};
1204
1205static DEFINE_PER_CPU(struct memory_failure_cpu, memory_failure_cpu);
1206
1207/**
1208 * memory_failure_queue - Schedule handling memory failure of a page.
1209 * @pfn: Page Number of the corrupted page
1210 * @trapno: Trap number reported in the signal to user space.
1211 * @flags: Flags for memory failure handling
1212 *
1213 * This function is called by the low level hardware error handler
1214 * when it detects hardware memory corruption of a page. It schedules
1215 * the recovering of error page, including dropping pages, killing
1216 * processes etc.
1217 *
1218 * The function is primarily of use for corruptions that
1219 * happen outside the current execution context (e.g. when
1220 * detected by a background scrubber)
1221 *
1222 * Can run in IRQ context.
1223 */
1224void memory_failure_queue(unsigned long pfn, int trapno, int flags)
1225{
1226        struct memory_failure_cpu *mf_cpu;
1227        unsigned long proc_flags;
1228        struct memory_failure_entry entry = {
1229                .pfn =          pfn,
1230                .trapno =       trapno,
1231                .flags =        flags,
1232        };
1233
1234        mf_cpu = &get_cpu_var(memory_failure_cpu);
1235        spin_lock_irqsave(&mf_cpu->lock, proc_flags);
1236        if (kfifo_put(&mf_cpu->fifo, &entry))
1237                schedule_work_on(smp_processor_id(), &mf_cpu->work);
1238        else
1239                pr_err("Memory failure: buffer overflow when queuing memory failure at 0x%#lx\n",
1240                       pfn);
1241        spin_unlock_irqrestore(&mf_cpu->lock, proc_flags);
1242        put_cpu_var(memory_failure_cpu);
1243}
1244EXPORT_SYMBOL_GPL(memory_failure_queue);
1245
1246static void memory_failure_work_func(struct work_struct *work)
1247{
1248        struct memory_failure_cpu *mf_cpu;
1249        struct memory_failure_entry entry = { 0, };
1250        unsigned long proc_flags;
1251        int gotten;
1252
1253        mf_cpu = &__get_cpu_var(memory_failure_cpu);
1254        for (;;) {
1255                spin_lock_irqsave(&mf_cpu->lock, proc_flags);
1256                gotten = kfifo_get(&mf_cpu->fifo, &entry);
1257                spin_unlock_irqrestore(&mf_cpu->lock, proc_flags);
1258                if (!gotten)
1259                        break;
1260                memory_failure(entry.pfn, entry.trapno, entry.flags);
1261        }
1262}
1263
1264static int __init memory_failure_init(void)
1265{
1266        struct memory_failure_cpu *mf_cpu;
1267        int cpu;
1268
1269        for_each_possible_cpu(cpu) {
1270                mf_cpu = &per_cpu(memory_failure_cpu, cpu);
1271                spin_lock_init(&mf_cpu->lock);
1272                INIT_KFIFO(mf_cpu->fifo);
1273                INIT_WORK(&mf_cpu->work, memory_failure_work_func);
1274        }
1275
1276        return 0;
1277}
1278core_initcall(memory_failure_init);
1279
1280/**
1281 * unpoison_memory - Unpoison a previously poisoned page
1282 * @pfn: Page number of the to be unpoisoned page
1283 *
1284 * Software-unpoison a page that has been poisoned by
1285 * memory_failure() earlier.
1286 *
1287 * This is only done on the software-level, so it only works
1288 * for linux injected failures, not real hardware failures
1289 *
1290 * Returns 0 for success, otherwise -errno.
1291 */
1292int unpoison_memory(unsigned long pfn)
1293{
1294        struct page *page;
1295        struct page *p;
1296        int freeit = 0;
1297        unsigned int nr_pages;
1298
1299        if (!pfn_valid(pfn))
1300                return -ENXIO;
1301
1302        p = pfn_to_page(pfn);
1303        page = compound_head(p);
1304
1305        if (!PageHWPoison(p)) {
1306                pr_info("MCE: Page was already unpoisoned %#lx\n", pfn);
1307                return 0;
1308        }
1309
1310        nr_pages = 1 << compound_trans_order(page);
1311
1312        if (!get_page_unless_zero(page)) {
1313                /*
1314                 * Since HWPoisoned hugepage should have non-zero refcount,
1315                 * race between memory failure and unpoison seems to happen.
1316                 * In such case unpoison fails and memory failure runs
1317                 * to the end.
1318                 */
1319                if (PageHuge(page)) {
1320                        pr_info("MCE: Memory failure is now running on free hugepage %#lx\n", pfn);
1321                        return 0;
1322                }
1323                if (TestClearPageHWPoison(p))
1324                        atomic_long_sub(nr_pages, &mce_bad_pages);
1325                pr_info("MCE: Software-unpoisoned free page %#lx\n", pfn);
1326                return 0;
1327        }
1328
1329        lock_page(page);
1330        /*
1331         * This test is racy because PG_hwpoison is set outside of page lock.
1332         * That's acceptable because that won't trigger kernel panic. Instead,
1333         * the PG_hwpoison page will be caught and isolated on the entrance to
1334         * the free buddy page pool.
1335         */
1336        if (TestClearPageHWPoison(page)) {
1337                pr_info("MCE: Software-unpoisoned page %#lx\n", pfn);
1338                atomic_long_sub(nr_pages, &mce_bad_pages);
1339                freeit = 1;
1340                if (PageHuge(page))
1341                        clear_page_hwpoison_huge_page(page);
1342        }
1343        unlock_page(page);
1344
1345        put_page(page);
1346        if (freeit)
1347                put_page(page);
1348
1349        return 0;
1350}
1351EXPORT_SYMBOL(unpoison_memory);
1352
1353static struct page *new_page(struct page *p, unsigned long private, int **x)
1354{
1355        int nid = page_to_nid(p);
1356        if (PageHuge(p))
1357                return alloc_huge_page_node(page_hstate(compound_head(p)),
1358                                                   nid);
1359        else
1360                return alloc_pages_exact_node(nid, GFP_HIGHUSER_MOVABLE, 0);
1361}
1362
1363/*
1364 * Safely get reference count of an arbitrary page.
1365 * Returns 0 for a free page, -EIO for a zero refcount page
1366 * that is not free, and 1 for any other page type.
1367 * For 1 the page is returned with increased page count, otherwise not.
1368 */
1369static int get_any_page(struct page *p, unsigned long pfn, int flags)
1370{
1371        int ret;
1372
1373        if (flags & MF_COUNT_INCREASED)
1374                return 1;
1375
1376        /*
1377         * The lock_memory_hotplug prevents a race with memory hotplug.
1378         * This is a big hammer, a better would be nicer.
1379         */
1380        lock_memory_hotplug();
1381
1382        /*
1383         * Isolate the page, so that it doesn't get reallocated if it
1384         * was free.
1385         */
1386        set_migratetype_isolate(p);
1387        /*
1388         * When the target page is a free hugepage, just remove it
1389         * from free hugepage list.
1390         */
1391        if (!get_page_unless_zero(compound_head(p))) {
1392                if (PageHuge(p)) {
1393                        pr_info("%s: %#lx free huge page\n", __func__, pfn);
1394                        ret = dequeue_hwpoisoned_huge_page(compound_head(p));
1395                } else if (is_free_buddy_page(p)) {
1396                        pr_info("%s: %#lx free buddy page\n", __func__, pfn);
1397                        /* Set hwpoison bit while page is still isolated */
1398                        SetPageHWPoison(p);
1399                        ret = 0;
1400                } else {
1401                        pr_info("%s: %#lx: unknown zero refcount page type %lx\n",
1402                                __func__, pfn, p->flags);
1403                        ret = -EIO;
1404                }
1405        } else {
1406                /* Not a free page */
1407                ret = 1;
1408        }
1409        unset_migratetype_isolate(p, MIGRATE_MOVABLE);
1410        unlock_memory_hotplug();
1411        return ret;
1412}
1413
1414static int soft_offline_huge_page(struct page *page, int flags)
1415{
1416        int ret;
1417        unsigned long pfn = page_to_pfn(page);
1418        struct page *hpage = compound_head(page);
1419
1420        ret = get_any_page(page, pfn, flags);
1421        if (ret < 0)
1422                return ret;
1423        if (ret == 0)
1424                goto done;
1425
1426        if (PageHWPoison(hpage)) {
1427                put_page(hpage);
1428                pr_info("soft offline: %#lx hugepage already poisoned\n", pfn);
1429                return -EBUSY;
1430        }
1431
1432        /* Keep page count to indicate a given hugepage is isolated. */
1433        ret = migrate_huge_page(hpage, new_page, MPOL_MF_MOVE_ALL, false,
1434                                MIGRATE_SYNC);
1435        put_page(hpage);
1436        if (ret) {
1437                pr_info("soft offline: %#lx: migration failed %d, type %lx\n",
1438                        pfn, ret, page->flags);
1439                return ret;
1440        }
1441done:
1442        if (!PageHWPoison(hpage))
1443                atomic_long_add(1 << compound_trans_order(hpage),
1444                                &mce_bad_pages);
1445        set_page_hwpoison_huge_page(hpage);
1446        dequeue_hwpoisoned_huge_page(hpage);
1447        /* keep elevated page count for bad page */
1448        return ret;
1449}
1450
1451/**
1452 * soft_offline_page - Soft offline a page.
1453 * @page: page to offline
1454 * @flags: flags. Same as memory_failure().
1455 *
1456 * Returns 0 on success, otherwise negated errno.
1457 *
1458 * Soft offline a page, by migration or invalidation,
1459 * without killing anything. This is for the case when
1460 * a page is not corrupted yet (so it's still valid to access),
1461 * but has had a number of corrected errors and is better taken
1462 * out.
1463 *
1464 * The actual policy on when to do that is maintained by
1465 * user space.
1466 *
1467 * This should never impact any application or cause data loss,
1468 * however it might take some time.
1469 *
1470 * This is not a 100% solution for all memory, but tries to be
1471 * ``good enough'' for the majority of memory.
1472 */
1473int soft_offline_page(struct page *page, int flags)
1474{
1475        int ret;
1476        unsigned long pfn = page_to_pfn(page);
1477        struct page *hpage = compound_trans_head(page);
1478
1479        if (PageHuge(page))
1480                return soft_offline_huge_page(page, flags);
1481        if (PageTransHuge(hpage)) {
1482                if (PageAnon(hpage) && unlikely(split_huge_page(hpage))) {
1483                        pr_info("soft offline: %#lx: failed to split THP\n",
1484                                pfn);
1485                        return -EBUSY;
1486                }
1487        }
1488
1489        ret = get_any_page(page, pfn, flags);
1490        if (ret < 0)
1491                return ret;
1492        if (ret == 0)
1493                goto done;
1494
1495        /*
1496         * Page cache page we can handle?
1497         */
1498        if (!PageLRU(page)) {
1499                /*
1500                 * Try to free it.
1501                 */
1502                put_page(page);
1503                shake_page(page, 1);
1504
1505                /*
1506                 * Did it turn free?
1507                 */
1508                ret = get_any_page(page, pfn, 0);
1509                if (ret < 0)
1510                        return ret;
1511                if (ret == 0)
1512                        goto done;
1513        }
1514        if (!PageLRU(page)) {
1515                pr_info("soft_offline: %#lx: unknown non LRU page type %lx\n",
1516                        pfn, page->flags);
1517                return -EIO;
1518        }
1519
1520        lock_page(page);
1521        wait_on_page_writeback(page);
1522
1523        /*
1524         * Synchronized using the page lock with memory_failure()
1525         */
1526        if (PageHWPoison(page)) {
1527                unlock_page(page);
1528                put_page(page);
1529                pr_info("soft offline: %#lx page already poisoned\n", pfn);
1530                return -EBUSY;
1531        }
1532
1533        /*
1534         * Try to invalidate first. This should work for
1535         * non dirty unmapped page cache pages.
1536         */
1537        ret = invalidate_inode_page(page);
1538        unlock_page(page);
1539        /*
1540         * RED-PEN would be better to keep it isolated here, but we
1541         * would need to fix isolation locking first.
1542         */
1543        if (ret == 1) {
1544                put_page(page);
1545                ret = 0;
1546                pr_info("soft_offline: %#lx: invalidated\n", pfn);
1547                goto done;
1548        }
1549
1550        /*
1551         * Simple invalidation didn't work.
1552         * Try to migrate to a new page instead. migrate.c
1553         * handles a large number of cases for us.
1554         */
1555        ret = isolate_lru_page(page);
1556        /*
1557         * Drop page reference which is came from get_any_page()
1558         * successful isolate_lru_page() already took another one.
1559         */
1560        put_page(page);
1561        if (!ret) {
1562                LIST_HEAD(pagelist);
1563                inc_zone_page_state(page, NR_ISOLATED_ANON +
1564                                            page_is_file_cache(page));
1565                list_add(&page->lru, &pagelist);
1566                ret = migrate_pages(&pagelist, new_page, MPOL_MF_MOVE_ALL,
1567                                                        false, MIGRATE_SYNC);
1568                if (ret) {
1569                        putback_lru_pages(&pagelist);
1570                        pr_info("soft offline: %#lx: migration failed %d, type %lx\n",
1571                                pfn, ret, page->flags);
1572                        if (ret > 0)
1573                                ret = -EIO;
1574                }
1575        } else {
1576                pr_info("soft offline: %#lx: isolation failed: %d, page count %d, type %lx\n",
1577                        pfn, ret, page_count(page), page->flags);
1578        }
1579        if (ret)
1580                return ret;
1581
1582done:
1583        atomic_long_add(1, &mce_bad_pages);
1584        SetPageHWPoison(page);
1585        /* keep elevated page count for bad page */
1586        return ret;
1587}
1588
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.