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 num_poisoned_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        pgoff_t pgoff;
 404
 405        av = page_lock_anon_vma_read(page);
 406        if (av == NULL) /* Not actually mapped anymore */
 407                return;
 408
 409        pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
 410        read_lock(&tasklist_lock);
 411        for_each_process (tsk) {
 412                struct anon_vma_chain *vmac;
 413
 414                if (!task_early_kill(tsk))
 415                        continue;
 416                anon_vma_interval_tree_foreach(vmac, &av->rb_root,
 417                                               pgoff, pgoff) {
 418                        vma = vmac->vma;
 419                        if (!page_mapped_in_vma(page, vma))
 420                                continue;
 421                        if (vma->vm_mm == tsk->mm)
 422                                add_to_kill(tsk, page, vma, to_kill, tkc);
 423                }
 424        }
 425        read_unlock(&tasklist_lock);
 426        page_unlock_anon_vma_read(av);
 427}
 428
 429/*
 430 * Collect processes when the error hit a file mapped page.
 431 */
 432static void collect_procs_file(struct page *page, struct list_head *to_kill,
 433                              struct to_kill **tkc)
 434{
 435        struct vm_area_struct *vma;
 436        struct task_struct *tsk;
 437        struct address_space *mapping = page->mapping;
 438
 439        mutex_lock(&mapping->i_mmap_mutex);
 440        read_lock(&tasklist_lock);
 441        for_each_process(tsk) {
 442                pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
 443
 444                if (!task_early_kill(tsk))
 445                        continue;
 446
 447                vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff,
 448                                      pgoff) {
 449                        /*
 450                         * Send early kill signal to tasks where a vma covers
 451                         * the page but the corrupted page is not necessarily
 452                         * mapped it in its pte.
 453                         * Assume applications who requested early kill want
 454                         * to be informed of all such data corruptions.
 455                         */
 456                        if (vma->vm_mm == tsk->mm)
 457                                add_to_kill(tsk, page, vma, to_kill, tkc);
 458                }
 459        }
 460        read_unlock(&tasklist_lock);
 461        mutex_unlock(&mapping->i_mmap_mutex);
 462}
 463
 464/*
 465 * Collect the processes who have the corrupted page mapped to kill.
 466 * This is done in two steps for locking reasons.
 467 * First preallocate one tokill structure outside the spin locks,
 468 * so that we can kill at least one process reasonably reliable.
 469 */
 470static void collect_procs(struct page *page, struct list_head *tokill)
 471{
 472        struct to_kill *tk;
 473
 474        if (!page->mapping)
 475                return;
 476
 477        tk = kmalloc(sizeof(struct to_kill), GFP_NOIO);
 478        if (!tk)
 479                return;
 480        if (PageAnon(page))
 481                collect_procs_anon(page, tokill, &tk);
 482        else
 483                collect_procs_file(page, tokill, &tk);
 484        kfree(tk);
 485}
 486
 487/*
 488 * Error handlers for various types of pages.
 489 */
 490
 491enum outcome {
 492        IGNORED,        /* Error: cannot be handled */
 493        FAILED,         /* Error: handling failed */
 494        DELAYED,        /* Will be handled later */
 495        RECOVERED,      /* Successfully recovered */
 496};
 497
 498static const char *action_name[] = {
 499        [IGNORED] = "Ignored",
 500        [FAILED] = "Failed",
 501        [DELAYED] = "Delayed",
 502        [RECOVERED] = "Recovered",
 503};
 504
 505/*
 506 * XXX: It is possible that a page is isolated from LRU cache,
 507 * and then kept in swap cache or failed to remove from page cache.
 508 * The page count will stop it from being freed by unpoison.
 509 * Stress tests should be aware of this memory leak problem.
 510 */
 511static int delete_from_lru_cache(struct page *p)
 512{
 513        if (!isolate_lru_page(p)) {
 514                /*
 515                 * Clear sensible page flags, so that the buddy system won't
 516                 * complain when the page is unpoison-and-freed.
 517                 */
 518                ClearPageActive(p);
 519                ClearPageUnevictable(p);
 520                /*
 521                 * drop the page count elevated by isolate_lru_page()
 522                 */
 523                page_cache_release(p);
 524                return 0;
 525        }
 526        return -EIO;
 527}
 528
 529/*
 530 * Error hit kernel page.
 531 * Do nothing, try to be lucky and not touch this instead. For a few cases we
 532 * could be more sophisticated.
 533 */
 534static int me_kernel(struct page *p, unsigned long pfn)
 535{
 536        return IGNORED;
 537}
 538
 539/*
 540 * Page in unknown state. Do nothing.
 541 */
 542static int me_unknown(struct page *p, unsigned long pfn)
 543{
 544        printk(KERN_ERR "MCE %#lx: Unknown page state\n", pfn);
 545        return FAILED;
 546}
 547
 548/*
 549 * Clean (or cleaned) page cache page.
 550 */
 551static int me_pagecache_clean(struct page *p, unsigned long pfn)
 552{
 553        int err;
 554        int ret = FAILED;
 555        struct address_space *mapping;
 556
 557        delete_from_lru_cache(p);
 558
 559        /*
 560         * For anonymous pages we're done the only reference left
 561         * should be the one m_f() holds.
 562         */
 563        if (PageAnon(p))
 564                return RECOVERED;
 565
 566        /*
 567         * Now truncate the page in the page cache. This is really
 568         * more like a "temporary hole punch"
 569         * Don't do this for block devices when someone else
 570         * has a reference, because it could be file system metadata
 571         * and that's not safe to truncate.
 572         */
 573        mapping = page_mapping(p);
 574        if (!mapping) {
 575                /*
 576                 * Page has been teared down in the meanwhile
 577                 */
 578                return FAILED;
 579        }
 580
 581        /*
 582         * Truncation is a bit tricky. Enable it per file system for now.
 583         *
 584         * Open: to take i_mutex or not for this? Right now we don't.
 585         */
 586        if (mapping->a_ops->error_remove_page) {
 587                err = mapping->a_ops->error_remove_page(mapping, p);
 588                if (err != 0) {
 589                        printk(KERN_INFO "MCE %#lx: Failed to punch page: %d\n",
 590                                        pfn, err);
 591                } else if (page_has_private(p) &&
 592                                !try_to_release_page(p, GFP_NOIO)) {
 593                        pr_info("MCE %#lx: failed to release buffers\n", pfn);
 594                } else {
 595                        ret = RECOVERED;
 596                }
 597        } else {
 598                /*
 599                 * If the file system doesn't support it just invalidate
 600                 * This fails on dirty or anything with private pages
 601                 */
 602                if (invalidate_inode_page(p))
 603                        ret = RECOVERED;
 604                else
 605                        printk(KERN_INFO "MCE %#lx: Failed to invalidate\n",
 606                                pfn);
 607        }
 608        return ret;
 609}
 610
 611/*
 612 * Dirty cache page page
 613 * Issues: when the error hit a hole page the error is not properly
 614 * propagated.
 615 */
 616static int me_pagecache_dirty(struct page *p, unsigned long pfn)
 617{
 618        struct address_space *mapping = page_mapping(p);
 619
 620        SetPageError(p);
 621        /* TBD: print more information about the file. */
 622        if (mapping) {
 623                /*
 624                 * IO error will be reported by write(), fsync(), etc.
 625                 * who check the mapping.
 626                 * This way the application knows that something went
 627                 * wrong with its dirty file data.
 628                 *
 629                 * There's one open issue:
 630                 *
 631                 * The EIO will be only reported on the next IO
 632                 * operation and then cleared through the IO map.
 633                 * Normally Linux has two mechanisms to pass IO error
 634                 * first through the AS_EIO flag in the address space
 635                 * and then through the PageError flag in the page.
 636                 * Since we drop pages on memory failure handling the
 637                 * only mechanism open to use is through AS_AIO.
 638                 *
 639                 * This has the disadvantage that it gets cleared on
 640                 * the first operation that returns an error, while
 641                 * the PageError bit is more sticky and only cleared
 642                 * when the page is reread or dropped.  If an
 643                 * application assumes it will always get error on
 644                 * fsync, but does other operations on the fd before
 645                 * and the page is dropped between then the error
 646                 * will not be properly reported.
 647                 *
 648                 * This can already happen even without hwpoisoned
 649                 * pages: first on metadata IO errors (which only
 650                 * report through AS_EIO) or when the page is dropped
 651                 * at the wrong time.
 652                 *
 653                 * So right now we assume that the application DTRT on
 654                 * the first EIO, but we're not worse than other parts
 655                 * of the kernel.
 656                 */
 657                mapping_set_error(mapping, EIO);
 658        }
 659
 660        return me_pagecache_clean(p, pfn);
 661}
 662
 663/*
 664 * Clean and dirty swap cache.
 665 *
 666 * Dirty swap cache page is tricky to handle. The page could live both in page
 667 * cache and swap cache(ie. page is freshly swapped in). So it could be
 668 * referenced concurrently by 2 types of PTEs:
 669 * normal PTEs and swap PTEs. We try to handle them consistently by calling
 670 * try_to_unmap(TTU_IGNORE_HWPOISON) to convert the normal PTEs to swap PTEs,
 671 * and then
 672 *      - clear dirty bit to prevent IO
 673 *      - remove from LRU
 674 *      - but keep in the swap cache, so that when we return to it on
 675 *        a later page fault, we know the application is accessing
 676 *        corrupted data and shall be killed (we installed simple
 677 *        interception code in do_swap_page to catch it).
 678 *
 679 * Clean swap cache pages can be directly isolated. A later page fault will
 680 * bring in the known good data from disk.
 681 */
 682static int me_swapcache_dirty(struct page *p, unsigned long pfn)
 683{
 684        ClearPageDirty(p);
 685        /* Trigger EIO in shmem: */
 686        ClearPageUptodate(p);
 687
 688        if (!delete_from_lru_cache(p))
 689                return DELAYED;
 690        else
 691                return FAILED;
 692}
 693
 694static int me_swapcache_clean(struct page *p, unsigned long pfn)
 695{
 696        delete_from_swap_cache(p);
 697
 698        if (!delete_from_lru_cache(p))
 699                return RECOVERED;
 700        else
 701                return FAILED;
 702}
 703
 704/*
 705 * Huge pages. Needs work.
 706 * Issues:
 707 * - Error on hugepage is contained in hugepage unit (not in raw page unit.)
 708 *   To narrow down kill region to one page, we need to break up pmd.
 709 */
 710static int me_huge_page(struct page *p, unsigned long pfn)
 711{
 712        int res = 0;
 713        struct page *hpage = compound_head(p);
 714        /*
 715         * We can safely recover from error on free or reserved (i.e.
 716         * not in-use) hugepage by dequeuing it from freelist.
 717         * To check whether a hugepage is in-use or not, we can't use
 718         * page->lru because it can be used in other hugepage operations,
 719         * such as __unmap_hugepage_range() and gather_surplus_pages().
 720         * So instead we use page_mapping() and PageAnon().
 721         * We assume that this function is called with page lock held,
 722         * so there is no race between isolation and mapping/unmapping.
 723         */
 724        if (!(page_mapping(hpage) || PageAnon(hpage))) {
 725                res = dequeue_hwpoisoned_huge_page(hpage);
 726                if (!res)
 727                        return RECOVERED;
 728        }
 729        return DELAYED;
 730}
 731
 732/*
 733 * Various page states we can handle.
 734 *
 735 * A page state is defined by its current page->flags bits.
 736 * The table matches them in order and calls the right handler.
 737 *
 738 * This is quite tricky because we can access page at any time
 739 * in its live cycle, so all accesses have to be extremely careful.
 740 *
 741 * This is not complete. More states could be added.
 742 * For any missing state don't attempt recovery.
 743 */
 744
 745#define dirty           (1UL << PG_dirty)
 746#define sc              (1UL << PG_swapcache)
 747#define unevict         (1UL << PG_unevictable)
 748#define mlock           (1UL << PG_mlocked)
 749#define writeback       (1UL << PG_writeback)
 750#define lru             (1UL << PG_lru)
 751#define swapbacked      (1UL << PG_swapbacked)
 752#define head            (1UL << PG_head)
 753#define tail            (1UL << PG_tail)
 754#define compound        (1UL << PG_compound)
 755#define slab            (1UL << PG_slab)
 756#define reserved        (1UL << PG_reserved)
 757
 758static struct page_state {
 759        unsigned long mask;
 760        unsigned long res;
 761        char *msg;
 762        int (*action)(struct page *p, unsigned long pfn);
 763} error_states[] = {
 764        { reserved,     reserved,       "reserved kernel",      me_kernel },
 765        /*
 766         * free pages are specially detected outside this table:
 767         * PG_buddy pages only make a small fraction of all free pages.
 768         */
 769
 770        /*
 771         * Could in theory check if slab page is free or if we can drop
 772         * currently unused objects without touching them. But just
 773         * treat it as standard kernel for now.
 774         */
 775        { slab,         slab,           "kernel slab",  me_kernel },
 776
 777#ifdef CONFIG_PAGEFLAGS_EXTENDED
 778        { head,         head,           "huge",         me_huge_page },
 779        { tail,         tail,           "huge",         me_huge_page },
 780#else
 781        { compound,     compound,       "huge",         me_huge_page },
 782#endif
 783
 784        { sc|dirty,     sc|dirty,       "dirty swapcache",      me_swapcache_dirty },
 785        { sc|dirty,     sc,             "clean swapcache",      me_swapcache_clean },
 786
 787        { mlock|dirty,  mlock|dirty,    "dirty mlocked LRU",    me_pagecache_dirty },
 788        { mlock|dirty,  mlock,          "clean mlocked LRU",    me_pagecache_clean },
 789
 790        { unevict|dirty, unevict|dirty, "dirty unevictable LRU", me_pagecache_dirty },
 791        { unevict|dirty, unevict,       "clean unevictable LRU", me_pagecache_clean },
 792
 793        { lru|dirty,    lru|dirty,      "dirty LRU",    me_pagecache_dirty },
 794        { lru|dirty,    lru,            "clean LRU",    me_pagecache_clean },
 795
 796        /*
 797         * Catchall entry: must be at end.
 798         */
 799        { 0,            0,              "unknown page state",   me_unknown },
 800};
 801
 802#undef dirty
 803#undef sc
 804#undef unevict
 805#undef mlock
 806#undef writeback
 807#undef lru
 808#undef swapbacked
 809#undef head
 810#undef tail
 811#undef compound
 812#undef slab
 813#undef reserved
 814
 815/*
 816 * "Dirty/Clean" indication is not 100% accurate due to the possibility of
 817 * setting PG_dirty outside page lock. See also comment above set_page_dirty().
 818 */
 819static void action_result(unsigned long pfn, char *msg, int result)
 820{
 821        pr_err("MCE %#lx: %s page recovery: %s\n",
 822                pfn, msg, action_name[result]);
 823}
 824
 825static int page_action(struct page_state *ps, struct page *p,
 826                        unsigned long pfn)
 827{
 828        int result;
 829        int count;
 830
 831        result = ps->action(p, pfn);
 832        action_result(pfn, ps->msg, result);
 833
 834        count = page_count(p) - 1;
 835        if (ps->action == me_swapcache_dirty && result == DELAYED)
 836                count--;
 837        if (count != 0) {
 838                printk(KERN_ERR
 839                       "MCE %#lx: %s page still referenced by %d users\n",
 840                       pfn, ps->msg, count);
 841                result = FAILED;
 842        }
 843
 844        /* Could do more checks here if page looks ok */
 845        /*
 846         * Could adjust zone counters here to correct for the missing page.
 847         */
 848
 849        return (result == RECOVERED || result == DELAYED) ? 0 : -EBUSY;
 850}
 851
 852/*
 853 * Do all that is necessary to remove user space mappings. Unmap
 854 * the pages and send SIGBUS to the processes if the data was dirty.
 855 */
 856static int hwpoison_user_mappings(struct page *p, unsigned long pfn,
 857                                  int trapno, int flags)
 858{
 859        enum ttu_flags ttu = TTU_UNMAP | TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS;
 860        struct address_space *mapping;
 861        LIST_HEAD(tokill);
 862        int ret;
 863        int kill = 1, forcekill;
 864        struct page *hpage = compound_head(p);
 865        struct page *ppage;
 866
 867        if (PageReserved(p) || PageSlab(p))
 868                return SWAP_SUCCESS;
 869
 870        /*
 871         * This check implies we don't kill processes if their pages
 872         * are in the swap cache early. Those are always late kills.
 873         */
 874        if (!page_mapped(hpage))
 875                return SWAP_SUCCESS;
 876
 877        if (PageKsm(p))
 878                return SWAP_FAIL;
 879
 880        if (PageSwapCache(p)) {
 881                printk(KERN_ERR
 882                       "MCE %#lx: keeping poisoned page in swap cache\n", pfn);
 883                ttu |= TTU_IGNORE_HWPOISON;
 884        }
 885
 886        /*
 887         * Propagate the dirty bit from PTEs to struct page first, because we
 888         * need this to decide if we should kill or just drop the page.
 889         * XXX: the dirty test could be racy: set_page_dirty() may not always
 890         * be called inside page lock (it's recommended but not enforced).
 891         */
 892        mapping = page_mapping(hpage);
 893        if (!(flags & MF_MUST_KILL) && !PageDirty(hpage) && mapping &&
 894            mapping_cap_writeback_dirty(mapping)) {
 895                if (page_mkclean(hpage)) {
 896                        SetPageDirty(hpage);
 897                } else {
 898                        kill = 0;
 899                        ttu |= TTU_IGNORE_HWPOISON;
 900                        printk(KERN_INFO
 901        "MCE %#lx: corrupted page was clean: dropped without side effects\n",
 902                                pfn);
 903                }
 904        }
 905
 906        /*
 907         * ppage: poisoned page
 908         *   if p is regular page(4k page)
 909         *        ppage == real poisoned page;
 910         *   else p is hugetlb or THP, ppage == head page.
 911         */
 912        ppage = hpage;
 913
 914        if (PageTransHuge(hpage)) {
 915                /*
 916                 * Verify that this isn't a hugetlbfs head page, the check for
 917                 * PageAnon is just for avoid tripping a split_huge_page
 918                 * internal debug check, as split_huge_page refuses to deal with
 919                 * anything that isn't an anon page. PageAnon can't go away fro
 920                 * under us because we hold a refcount on the hpage, without a
 921                 * refcount on the hpage. split_huge_page can't be safely called
 922                 * in the first place, having a refcount on the tail isn't
 923                 * enough * to be safe.
 924                 */
 925                if (!PageHuge(hpage) && PageAnon(hpage)) {
 926                        if (unlikely(split_huge_page(hpage))) {
 927                                /*
 928                                 * FIXME: if splitting THP is failed, it is
 929                                 * better to stop the following operation rather
 930                                 * than causing panic by unmapping. System might
 931                                 * survive if the page is freed later.
 932                                 */
 933                                printk(KERN_INFO
 934                                        "MCE %#lx: failed to split THP\n", pfn);
 935
 936                                BUG_ON(!PageHWPoison(p));
 937                                return SWAP_FAIL;
 938                        }
 939                        /* THP is split, so ppage should be the real poisoned page. */
 940                        ppage = p;
 941                }
 942        }
 943
 944        /*
 945         * First collect all the processes that have the page
 946         * mapped in dirty form.  This has to be done before try_to_unmap,
 947         * because ttu takes the rmap data structures down.
 948         *
 949         * Error handling: We ignore errors here because
 950         * there's nothing that can be done.
 951         */
 952        if (kill)
 953                collect_procs(ppage, &tokill);
 954
 955        if (hpage != ppage)
 956                lock_page(ppage);
 957
 958        ret = try_to_unmap(ppage, ttu);
 959        if (ret != SWAP_SUCCESS)
 960                printk(KERN_ERR "MCE %#lx: failed to unmap page (mapcount=%d)\n",
 961                                pfn, page_mapcount(ppage));
 962
 963        if (hpage != ppage)
 964                unlock_page(ppage);
 965
 966        /*
 967         * Now that the dirty bit has been propagated to the
 968         * struct page and all unmaps done we can decide if
 969         * killing is needed or not.  Only kill when the page
 970         * was dirty or the process is not restartable,
 971         * otherwise the tokill list is merely
 972         * freed.  When there was a problem unmapping earlier
 973         * use a more force-full uncatchable kill to prevent
 974         * any accesses to the poisoned memory.
 975         */
 976        forcekill = PageDirty(ppage) || (flags & MF_MUST_KILL);
 977        kill_procs(&tokill, forcekill, trapno,
 978                      ret != SWAP_SUCCESS, p, pfn, flags);
 979
 980        return ret;
 981}
 982
 983static void set_page_hwpoison_huge_page(struct page *hpage)
 984{
 985        int i;
 986        int nr_pages = 1 << compound_trans_order(hpage);
 987        for (i = 0; i < nr_pages; i++)
 988                SetPageHWPoison(hpage + i);
 989}
 990
 991static void clear_page_hwpoison_huge_page(struct page *hpage)
 992{
 993        int i;
 994        int nr_pages = 1 << compound_trans_order(hpage);
 995        for (i = 0; i < nr_pages; i++)
 996                ClearPageHWPoison(hpage + i);
 997}
 998
 999/**
1000 * memory_failure - Handle memory failure of a page.
1001 * @pfn: Page Number of the corrupted page
1002 * @trapno: Trap number reported in the signal to user space.
1003 * @flags: fine tune action taken
1004 *
1005 * This function is called by the low level machine check code
1006 * of an architecture when it detects hardware memory corruption
1007 * of a page. It tries its best to recover, which includes
1008 * dropping pages, killing processes etc.
1009 *
1010 * The function is primarily of use for corruptions that
1011 * happen outside the current execution context (e.g. when
1012 * detected by a background scrubber)
1013 *
1014 * Must run in process context (e.g. a work queue) with interrupts
1015 * enabled and no spinlocks hold.
1016 */
1017int memory_failure(unsigned long pfn, int trapno, int flags)
1018{
1019        struct page_state *ps;
1020        struct page *p;
1021        struct page *hpage;
1022        int res;
1023        unsigned int nr_pages;
1024        unsigned long page_flags;
1025
1026        if (!sysctl_memory_failure_recovery)
1027                panic("Memory failure from trap %d on page %lx", trapno, pfn);
1028
1029        if (!pfn_valid(pfn)) {
1030                printk(KERN_ERR
1031                       "MCE %#lx: memory outside kernel control\n",
1032                       pfn);
1033                return -ENXIO;
1034        }
1035
1036        p = pfn_to_page(pfn);
1037        hpage = compound_head(p);
1038        if (TestSetPageHWPoison(p)) {
1039                printk(KERN_ERR "MCE %#lx: already hardware poisoned\n", pfn);
1040                return 0;
1041        }
1042
1043        /*
1044         * Currently errors on hugetlbfs pages are measured in hugepage units,
1045         * so nr_pages should be 1 << compound_order.  OTOH when errors are on
1046         * transparent hugepages, they are supposed to be split and error
1047         * measurement is done in normal page units.  So nr_pages should be one
1048         * in this case.
1049         */
1050        if (PageHuge(p))
1051                nr_pages = 1 << compound_order(hpage);
1052        else /* normal page or thp */
1053                nr_pages = 1;
1054        atomic_long_add(nr_pages, &num_poisoned_pages);
1055
1056        /*
1057         * We need/can do nothing about count=0 pages.
1058         * 1) it's a free page, and therefore in safe hand:
1059         *    prep_new_page() will be the gate keeper.
1060         * 2) it's a free hugepage, which is also safe:
1061         *    an affected hugepage will be dequeued from hugepage freelist,
1062         *    so there's no concern about reusing it ever after.
1063         * 3) it's part of a non-compound high order page.
1064         *    Implies some kernel user: cannot stop them from
1065         *    R/W the page; let's pray that the page has been
1066         *    used and will be freed some time later.
1067         * In fact it's dangerous to directly bump up page count from 0,
1068         * that may make page_freeze_refs()/page_unfreeze_refs() mismatch.
1069         */
1070        if (!(flags & MF_COUNT_INCREASED) &&
1071                !get_page_unless_zero(hpage)) {
1072                if (is_free_buddy_page(p)) {
1073                        action_result(pfn, "free buddy", DELAYED);
1074                        return 0;
1075                } else if (PageHuge(hpage)) {
1076                        /*
1077                         * Check "just unpoisoned", "filter hit", and
1078                         * "race with other subpage."
1079                         */
1080                        lock_page(hpage);
1081                        if (!PageHWPoison(hpage)
1082                            || (hwpoison_filter(p) && TestClearPageHWPoison(p))
1083                            || (p != hpage && TestSetPageHWPoison(hpage))) {
1084                                atomic_long_sub(nr_pages, &num_poisoned_pages);
1085                                return 0;
1086                        }
1087                        set_page_hwpoison_huge_page(hpage);
1088                        res = dequeue_hwpoisoned_huge_page(hpage);
1089                        action_result(pfn, "free huge",
1090                                      res ? IGNORED : DELAYED);
1091                        unlock_page(hpage);
1092                        return res;
1093                } else {
1094                        action_result(pfn, "high order kernel", IGNORED);
1095                        return -EBUSY;
1096                }
1097        }
1098
1099        /*
1100         * We ignore non-LRU pages for good reasons.
1101         * - PG_locked is only well defined for LRU pages and a few others
1102         * - to avoid races with __set_page_locked()
1103         * - to avoid races with __SetPageSlab*() (and more non-atomic ops)
1104         * The check (unnecessarily) ignores LRU pages being isolated and
1105         * walked by the page reclaim code, however that's not a big loss.
1106         */
1107        if (!PageHuge(p) && !PageTransTail(p)) {
1108                if (!PageLRU(p))
1109                        shake_page(p, 0);
1110                if (!PageLRU(p)) {
1111                        /*
1112                         * shake_page could have turned it free.
1113                         */
1114                        if (is_free_buddy_page(p)) {
1115                                action_result(pfn, "free buddy, 2nd try",
1116                                                DELAYED);
1117                                return 0;
1118                        }
1119                        action_result(pfn, "non LRU", IGNORED);
1120                        put_page(p);
1121                        return -EBUSY;
1122                }
1123        }
1124
1125        /*
1126         * Lock the page and wait for writeback to finish.
1127         * It's very difficult to mess with pages currently under IO
1128         * and in many cases impossible, so we just avoid it here.
1129         */
1130        lock_page(hpage);
1131
1132        /*
1133         * We use page flags to determine what action should be taken, but
1134         * the flags can be modified by the error containment action.  One
1135         * example is an mlocked page, where PG_mlocked is cleared by
1136         * page_remove_rmap() in try_to_unmap_one(). So to determine page status
1137         * correctly, we save a copy of the page flags at this time.
1138         */
1139        page_flags = p->flags;
1140
1141        /*
1142         * unpoison always clear PG_hwpoison inside page lock
1143         */
1144        if (!PageHWPoison(p)) {
1145                printk(KERN_ERR "MCE %#lx: just unpoisoned\n", pfn);
1146                res = 0;
1147                goto out;
1148        }
1149        if (hwpoison_filter(p)) {
1150                if (TestClearPageHWPoison(p))
1151                        atomic_long_sub(nr_pages, &num_poisoned_pages);
1152                unlock_page(hpage);
1153                put_page(hpage);
1154                return 0;
1155        }
1156
1157        /*
1158         * For error on the tail page, we should set PG_hwpoison
1159         * on the head page to show that the hugepage is hwpoisoned
1160         */
1161        if (PageHuge(p) && PageTail(p) && TestSetPageHWPoison(hpage)) {
1162                action_result(pfn, "hugepage already hardware poisoned",
1163                                IGNORED);
1164                unlock_page(hpage);
1165                put_page(hpage);
1166                return 0;
1167        }
1168        /*
1169         * Set PG_hwpoison on all pages in an error hugepage,
1170         * because containment is done in hugepage unit for now.
1171         * Since we have done TestSetPageHWPoison() for the head page with
1172         * page lock held, we can safely set PG_hwpoison bits on tail pages.
1173         */
1174        if (PageHuge(p))
1175                set_page_hwpoison_huge_page(hpage);
1176
1177        wait_on_page_writeback(p);
1178
1179        /*
1180         * Now take care of user space mappings.
1181         * Abort on fail: __delete_from_page_cache() assumes unmapped page.
1182         */
1183        if (hwpoison_user_mappings(p, pfn, trapno, flags) != SWAP_SUCCESS) {
1184                printk(KERN_ERR "MCE %#lx: cannot unmap page, give up\n", pfn);
1185                res = -EBUSY;
1186                goto out;
1187        }
1188
1189        /*
1190         * Torn down by someone else?
1191         */
1192        if (PageLRU(p) && !PageSwapCache(p) && p->mapping == NULL) {
1193                action_result(pfn, "already truncated LRU", IGNORED);
1194                res = -EBUSY;
1195                goto out;
1196        }
1197
1198        res = -EBUSY;
1199        /*
1200         * The first check uses the current page flags which may not have any
1201         * relevant information. The second check with the saved page flagss is
1202         * carried out only if the first check can't determine the page status.
1203         */
1204        for (ps = error_states;; ps++)
1205                if ((p->flags & ps->mask) == ps->res)
1206                        break;
1207        if (!ps->mask)
1208                for (ps = error_states;; ps++)
1209                        if ((page_flags & ps->mask) == ps->res)
1210                                break;
1211        res = page_action(ps, p, pfn);
1212out:
1213        unlock_page(hpage);
1214        return res;
1215}
1216EXPORT_SYMBOL_GPL(memory_failure);
1217
1218#define MEMORY_FAILURE_FIFO_ORDER       4
1219#define MEMORY_FAILURE_FIFO_SIZE        (1 << MEMORY_FAILURE_FIFO_ORDER)
1220
1221struct memory_failure_entry {
1222        unsigned long pfn;
1223        int trapno;
1224        int flags;
1225};
1226
1227struct memory_failure_cpu {
1228        DECLARE_KFIFO(fifo, struct memory_failure_entry,
1229                      MEMORY_FAILURE_FIFO_SIZE);
1230        spinlock_t lock;
1231        struct work_struct work;
1232};
1233
1234static DEFINE_PER_CPU(struct memory_failure_cpu, memory_failure_cpu);
1235
1236/**
1237 * memory_failure_queue - Schedule handling memory failure of a page.
1238 * @pfn: Page Number of the corrupted page
1239 * @trapno: Trap number reported in the signal to user space.
1240 * @flags: Flags for memory failure handling
1241 *
1242 * This function is called by the low level hardware error handler
1243 * when it detects hardware memory corruption of a page. It schedules
1244 * the recovering of error page, including dropping pages, killing
1245 * processes etc.
1246 *
1247 * The function is primarily of use for corruptions that
1248 * happen outside the current execution context (e.g. when
1249 * detected by a background scrubber)
1250 *
1251 * Can run in IRQ context.
1252 */
1253void memory_failure_queue(unsigned long pfn, int trapno, int flags)
1254{
1255        struct memory_failure_cpu *mf_cpu;
1256        unsigned long proc_flags;
1257        struct memory_failure_entry entry = {
1258                .pfn =          pfn,
1259                .trapno =       trapno,
1260                .flags =        flags,
1261        };
1262
1263        mf_cpu = &get_cpu_var(memory_failure_cpu);
1264        spin_lock_irqsave(&mf_cpu->lock, proc_flags);
1265        if (kfifo_put(&mf_cpu->fifo, &entry))
1266                schedule_work_on(smp_processor_id(), &mf_cpu->work);
1267        else
1268                pr_err("Memory failure: buffer overflow when queuing memory failure at 0x%#lx\n",
1269                       pfn);
1270        spin_unlock_irqrestore(&mf_cpu->lock, proc_flags);
1271        put_cpu_var(memory_failure_cpu);
1272}
1273EXPORT_SYMBOL_GPL(memory_failure_queue);
1274
1275static void memory_failure_work_func(struct work_struct *work)
1276{
1277        struct memory_failure_cpu *mf_cpu;
1278        struct memory_failure_entry entry = { 0, };
1279        unsigned long proc_flags;
1280        int gotten;
1281
1282        mf_cpu = &__get_cpu_var(memory_failure_cpu);
1283        for (;;) {
1284                spin_lock_irqsave(&mf_cpu->lock, proc_flags);
1285                gotten = kfifo_get(&mf_cpu->fifo, &entry);
1286                spin_unlock_irqrestore(&mf_cpu->lock, proc_flags);
1287                if (!gotten)
1288                        break;
1289                memory_failure(entry.pfn, entry.trapno, entry.flags);
1290        }
1291}
1292
1293static int __init memory_failure_init(void)
1294{
1295        struct memory_failure_cpu *mf_cpu;
1296        int cpu;
1297
1298        for_each_possible_cpu(cpu) {
1299                mf_cpu = &per_cpu(memory_failure_cpu, cpu);
1300                spin_lock_init(&mf_cpu->lock);
1301                INIT_KFIFO(mf_cpu->fifo);
1302                INIT_WORK(&mf_cpu->work, memory_failure_work_func);
1303        }
1304
1305        return 0;
1306}
1307core_initcall(memory_failure_init);
1308
1309/**
1310 * unpoison_memory - Unpoison a previously poisoned page
1311 * @pfn: Page number of the to be unpoisoned page
1312 *
1313 * Software-unpoison a page that has been poisoned by
1314 * memory_failure() earlier.
1315 *
1316 * This is only done on the software-level, so it only works
1317 * for linux injected failures, not real hardware failures
1318 *
1319 * Returns 0 for success, otherwise -errno.
1320 */
1321int unpoison_memory(unsigned long pfn)
1322{
1323        struct page *page;
1324        struct page *p;
1325        int freeit = 0;
1326        unsigned int nr_pages;
1327
1328        if (!pfn_valid(pfn))
1329                return -ENXIO;
1330
1331        p = pfn_to_page(pfn);
1332        page = compound_head(p);
1333
1334        if (!PageHWPoison(p)) {
1335                pr_info("MCE: Page was already unpoisoned %#lx\n", pfn);
1336                return 0;
1337        }
1338
1339        nr_pages = 1 << compound_trans_order(page);
1340
1341        if (!get_page_unless_zero(page)) {
1342                /*
1343                 * Since HWPoisoned hugepage should have non-zero refcount,
1344                 * race between memory failure and unpoison seems to happen.
1345                 * In such case unpoison fails and memory failure runs
1346                 * to the end.
1347                 */
1348                if (PageHuge(page)) {
1349                        pr_info("MCE: Memory failure is now running on free hugepage %#lx\n", pfn);
1350                        return 0;
1351                }
1352                if (TestClearPageHWPoison(p))
1353                        atomic_long_sub(nr_pages, &num_poisoned_pages);
1354                pr_info("MCE: Software-unpoisoned free page %#lx\n", pfn);
1355                return 0;
1356        }
1357
1358        lock_page(page);
1359        /*
1360         * This test is racy because PG_hwpoison is set outside of page lock.
1361         * That's acceptable because that won't trigger kernel panic. Instead,
1362         * the PG_hwpoison page will be caught and isolated on the entrance to
1363         * the free buddy page pool.
1364         */
1365        if (TestClearPageHWPoison(page)) {
1366                pr_info("MCE: Software-unpoisoned page %#lx\n", pfn);
1367                atomic_long_sub(nr_pages, &num_poisoned_pages);
1368                freeit = 1;
1369                if (PageHuge(page))
1370                        clear_page_hwpoison_huge_page(page);
1371        }
1372        unlock_page(page);
1373
1374        put_page(page);
1375        if (freeit)
1376                put_page(page);
1377
1378        return 0;
1379}
1380EXPORT_SYMBOL(unpoison_memory);
1381
1382static struct page *new_page(struct page *p, unsigned long private, int **x)
1383{
1384        int nid = page_to_nid(p);
1385        if (PageHuge(p))
1386                return alloc_huge_page_node(page_hstate(compound_head(p)),
1387                                                   nid);
1388        else
1389                return alloc_pages_exact_node(nid, GFP_HIGHUSER_MOVABLE, 0);
1390}
1391
1392/*
1393 * Safely get reference count of an arbitrary page.
1394 * Returns 0 for a free page, -EIO for a zero refcount page
1395 * that is not free, and 1 for any other page type.
1396 * For 1 the page is returned with increased page count, otherwise not.
1397 */
1398static int __get_any_page(struct page *p, unsigned long pfn, int flags)
1399{
1400        int ret;
1401
1402        if (flags & MF_COUNT_INCREASED)
1403                return 1;
1404
1405        /*
1406         * The lock_memory_hotplug prevents a race with memory hotplug.
1407         * This is a big hammer, a better would be nicer.
1408         */
1409        lock_memory_hotplug();
1410
1411        /*
1412         * Isolate the page, so that it doesn't get reallocated if it
1413         * was free. This flag should be kept set until the source page
1414         * is freed and PG_hwpoison on it is set.
1415         */
1416        set_migratetype_isolate(p, true);
1417        /*
1418         * When the target page is a free hugepage, just remove it
1419         * from free hugepage list.
1420         */
1421        if (!get_page_unless_zero(compound_head(p))) {
1422                if (PageHuge(p)) {
1423                        pr_info("%s: %#lx free huge page\n", __func__, pfn);
1424                        ret = 0;
1425                } else if (is_free_buddy_page(p)) {
1426                        pr_info("%s: %#lx free buddy page\n", __func__, pfn);
1427                        ret = 0;
1428                } else {
1429                        pr_info("%s: %#lx: unknown zero refcount page type %lx\n",
1430                                __func__, pfn, p->flags);
1431                        ret = -EIO;
1432                }
1433        } else {
1434                /* Not a free page */
1435                ret = 1;
1436        }
1437        unlock_memory_hotplug();
1438        return ret;
1439}
1440
1441static int get_any_page(struct page *page, unsigned long pfn, int flags)
1442{
1443        int ret = __get_any_page(page, pfn, flags);
1444
1445        if (ret == 1 && !PageHuge(page) && !PageLRU(page)) {
1446                /*
1447                 * Try to free it.
1448                 */
1449                put_page(page);
1450                shake_page(page, 1);
1451
1452                /*
1453                 * Did it turn free?
1454                 */
1455                ret = __get_any_page(page, pfn, 0);
1456                if (!PageLRU(page)) {
1457                        pr_info("soft_offline: %#lx: unknown non LRU page type %lx\n",
1458                                pfn, page->flags);
1459                        return -EIO;
1460                }
1461        }
1462        return ret;
1463}
1464
1465static int soft_offline_huge_page(struct page *page, int flags)
1466{
1467        int ret;
1468        unsigned long pfn = page_to_pfn(page);
1469        struct page *hpage = compound_head(page);
1470
1471        /*
1472         * This double-check of PageHWPoison is to avoid the race with
1473         * memory_failure(). See also comment in __soft_offline_page().
1474         */
1475        lock_page(hpage);
1476        if (PageHWPoison(hpage)) {
1477                unlock_page(hpage);
1478                put_page(hpage);
1479                pr_info("soft offline: %#lx hugepage already poisoned\n", pfn);
1480                return -EBUSY;
1481        }
1482        unlock_page(hpage);
1483
1484        /* Keep page count to indicate a given hugepage is isolated. */
1485        ret = migrate_huge_page(hpage, new_page, MPOL_MF_MOVE_ALL,
1486                                MIGRATE_SYNC);
1487        put_page(hpage);
1488        if (ret) {
1489                pr_info("soft offline: %#lx: migration failed %d, type %lx\n",
1490                        pfn, ret, page->flags);
1491        } else {
1492                set_page_hwpoison_huge_page(hpage);
1493                dequeue_hwpoisoned_huge_page(hpage);
1494                atomic_long_add(1 << compound_trans_order(hpage),
1495                                &num_poisoned_pages);
1496        }
1497        return ret;
1498}
1499
1500static int __soft_offline_page(struct page *page, int flags);
1501
1502/**
1503 * soft_offline_page - Soft offline a page.
1504 * @page: page to offline
1505 * @flags: flags. Same as memory_failure().
1506 *
1507 * Returns 0 on success, otherwise negated errno.
1508 *
1509 * Soft offline a page, by migration or invalidation,
1510 * without killing anything. This is for the case when
1511 * a page is not corrupted yet (so it's still valid to access),
1512 * but has had a number of corrected errors and is better taken
1513 * out.
1514 *
1515 * The actual policy on when to do that is maintained by
1516 * user space.
1517 *
1518 * This should never impact any application or cause data loss,
1519 * however it might take some time.
1520 *
1521 * This is not a 100% solution for all memory, but tries to be
1522 * ``good enough'' for the majority of memory.
1523 */
1524int soft_offline_page(struct page *page, int flags)
1525{
1526        int ret;
1527        unsigned long pfn = page_to_pfn(page);
1528        struct page *hpage = compound_trans_head(page);
1529
1530        if (PageHWPoison(page)) {
1531                pr_info("soft offline: %#lx page already poisoned\n", pfn);
1532                return -EBUSY;
1533        }
1534        if (!PageHuge(page) && PageTransHuge(hpage)) {
1535                if (PageAnon(hpage) && unlikely(split_huge_page(hpage))) {
1536                        pr_info("soft offline: %#lx: failed to split THP\n",
1537                                pfn);
1538                        return -EBUSY;
1539                }
1540        }
1541
1542        ret = get_any_page(page, pfn, flags);
1543        if (ret < 0)
1544                return ret;
1545        if (ret) { /* for in-use pages */
1546                if (PageHuge(page))
1547                        ret = soft_offline_huge_page(page, flags);
1548                else
1549                        ret = __soft_offline_page(page, flags);
1550        } else { /* for free pages */
1551                if (PageHuge(page)) {
1552                        set_page_hwpoison_huge_page(hpage);
1553                        dequeue_hwpoisoned_huge_page(hpage);
1554                        atomic_long_add(1 << compound_trans_order(hpage),
1555                                        &num_poisoned_pages);
1556                } else {
1557                        SetPageHWPoison(page);
1558                        atomic_long_inc(&num_poisoned_pages);
1559                }
1560        }
1561        unset_migratetype_isolate(page, MIGRATE_MOVABLE);
1562        return ret;
1563}
1564
1565static int __soft_offline_page(struct page *page, int flags)
1566{
1567        int ret;
1568        unsigned long pfn = page_to_pfn(page);
1569
1570        /*
1571         * Check PageHWPoison again inside page lock because PageHWPoison
1572         * is set by memory_failure() outside page lock. Note that
1573         * memory_failure() also double-checks PageHWPoison inside page lock,
1574         * so there's no race between soft_offline_page() and memory_failure().
1575         */
1576        lock_page(page);
1577        wait_on_page_writeback(page);
1578        if (PageHWPoison(page)) {
1579                unlock_page(page);
1580                put_page(page);
1581                pr_info("soft offline: %#lx page already poisoned\n", pfn);
1582                return -EBUSY;
1583        }
1584        /*
1585         * Try to invalidate first. This should work for
1586         * non dirty unmapped page cache pages.
1587         */
1588        ret = invalidate_inode_page(page);
1589        unlock_page(page);
1590        /*
1591         * RED-PEN would be better to keep it isolated here, but we
1592         * would need to fix isolation locking first.
1593         */
1594        if (ret == 1) {
1595                put_page(page);
1596                pr_info("soft_offline: %#lx: invalidated\n", pfn);
1597                SetPageHWPoison(page);
1598                atomic_long_inc(&num_poisoned_pages);
1599                return 0;
1600        }
1601
1602        /*
1603         * Simple invalidation didn't work.
1604         * Try to migrate to a new page instead. migrate.c
1605         * handles a large number of cases for us.
1606         */
1607        ret = isolate_lru_page(page);
1608        /*
1609         * Drop page reference which is came from get_any_page()
1610         * successful isolate_lru_page() already took another one.
1611         */
1612        put_page(page);
1613        if (!ret) {
1614                LIST_HEAD(pagelist);
1615                inc_zone_page_state(page, NR_ISOLATED_ANON +
1616                                        page_is_file_cache(page));
1617                list_add(&page->lru, &pagelist);
1618                ret = migrate_pages(&pagelist, new_page, MPOL_MF_MOVE_ALL,
1619                                        MIGRATE_SYNC, MR_MEMORY_FAILURE);
1620                if (ret) {
1621                        putback_lru_pages(&pagelist);
1622                        pr_info("soft offline: %#lx: migration failed %d, type %lx\n",
1623                                pfn, ret, page->flags);
1624                        if (ret > 0)
1625                                ret = -EIO;
1626                } else {
1627                        /*
1628                         * After page migration succeeds, the source page can
1629                         * be trapped in pagevec and actual freeing is delayed.
1630                         * Freeing code works differently based on PG_hwpoison,
1631                         * so there's a race. We need to make sure that the
1632                         * source page should be freed back to buddy before
1633                         * setting PG_hwpoison.
1634                         */
1635                        if (!is_free_buddy_page(page))
1636                                lru_add_drain_all();
1637                        if (!is_free_buddy_page(page))
1638                                drain_all_pages();
1639                        SetPageHWPoison(page);
1640                        if (!is_free_buddy_page(page))
1641                                pr_info("soft offline: %#lx: page leaked\n",
1642                                        pfn);
1643                        atomic_long_inc(&num_poisoned_pages);
1644                }
1645        } else {
1646                pr_info("soft offline: %#lx: isolation failed: %d, page count %d, type %lx\n",
1647                        pfn, ret, page_count(page), page->flags);
1648        }
1649        return ret;
1650}
1651
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.