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 * It can be very tempting to add handling for obscure cases here.
  25 * In general any code for handling new cases should only be added iff:
  26 * - You know how to test it.
  27 * - You have a test that can be added to mce-test
  28 *   https://git.kernel.org/cgit/utils/cpu/mce/mce-test.git/
  29 * - The case actually shows up as a frequent (top 10) page state in
  30 *   tools/vm/page-types when running a real workload.
  31 * 
  32 * There are several operations here with exponential complexity because
  33 * of unsuitable VM data structures. For example the operation to map back 
  34 * from RMAP chains to processes has to walk the complete process list and 
  35 * has non linear complexity with the number. But since memory corruptions
  36 * are rare we hope to get away with this. This avoids impacting the core 
  37 * VM.
  38 */
  39#include <linux/kernel.h>
  40#include <linux/mm.h>
  41#include <linux/page-flags.h>
  42#include <linux/kernel-page-flags.h>
  43#include <linux/sched.h>
 
  44#include <linux/ksm.h>
  45#include <linux/rmap.h>
  46#include <linux/export.h>
  47#include <linux/pagemap.h>
  48#include <linux/swap.h>
  49#include <linux/backing-dev.h>
  50#include <linux/migrate.h>
  51#include <linux/page-isolation.h>
  52#include <linux/suspend.h>
  53#include <linux/slab.h>
  54#include <linux/swapops.h>
  55#include <linux/hugetlb.h>
  56#include <linux/memory_hotplug.h>
  57#include <linux/mm_inline.h>
  58#include <linux/kfifo.h>
  59#include <linux/ratelimit.h>
  60#include "internal.h"
  61#include "ras/ras_event.h"
  62
  63int sysctl_memory_failure_early_kill __read_mostly = 0;
  64
  65int sysctl_memory_failure_recovery __read_mostly = 1;
  66
  67atomic_long_t num_poisoned_pages __read_mostly = ATOMIC_LONG_INIT(0);
  68
  69#if defined(CONFIG_HWPOISON_INJECT) || defined(CONFIG_HWPOISON_INJECT_MODULE)
  70
  71u32 hwpoison_filter_enable = 0;
  72u32 hwpoison_filter_dev_major = ~0U;
  73u32 hwpoison_filter_dev_minor = ~0U;
  74u64 hwpoison_filter_flags_mask;
  75u64 hwpoison_filter_flags_value;
  76EXPORT_SYMBOL_GPL(hwpoison_filter_enable);
  77EXPORT_SYMBOL_GPL(hwpoison_filter_dev_major);
  78EXPORT_SYMBOL_GPL(hwpoison_filter_dev_minor);
  79EXPORT_SYMBOL_GPL(hwpoison_filter_flags_mask);
  80EXPORT_SYMBOL_GPL(hwpoison_filter_flags_value);
  81
  82static int hwpoison_filter_dev(struct page *p)
  83{
  84	struct address_space *mapping;
  85	dev_t dev;
  86
  87	if (hwpoison_filter_dev_major == ~0U &&
  88	    hwpoison_filter_dev_minor == ~0U)
  89		return 0;
  90
  91	/*
  92	 * page_mapping() does not accept slab pages.
  93	 */
  94	if (PageSlab(p))
  95		return -EINVAL;
  96
  97	mapping = page_mapping(p);
  98	if (mapping == NULL || mapping->host == NULL)
  99		return -EINVAL;
 100
 101	dev = mapping->host->i_sb->s_dev;
 102	if (hwpoison_filter_dev_major != ~0U &&
 103	    hwpoison_filter_dev_major != MAJOR(dev))
 104		return -EINVAL;
 105	if (hwpoison_filter_dev_minor != ~0U &&
 106	    hwpoison_filter_dev_minor != MINOR(dev))
 107		return -EINVAL;
 108
 109	return 0;
 110}
 111
 112static int hwpoison_filter_flags(struct page *p)
 113{
 114	if (!hwpoison_filter_flags_mask)
 115		return 0;
 116
 117	if ((stable_page_flags(p) & hwpoison_filter_flags_mask) ==
 118				    hwpoison_filter_flags_value)
 119		return 0;
 120	else
 121		return -EINVAL;
 122}
 123
 124/*
 125 * This allows stress tests to limit test scope to a collection of tasks
 126 * by putting them under some memcg. This prevents killing unrelated/important
 127 * processes such as /sbin/init. Note that the target task may share clean
 128 * pages with init (eg. libc text), which is harmless. If the target task
 129 * share _dirty_ pages with another task B, the test scheme must make sure B
 130 * is also included in the memcg. At last, due to race conditions this filter
 131 * can only guarantee that the page either belongs to the memcg tasks, or is
 132 * a freed page.
 133 */
 134#ifdef CONFIG_MEMCG
 135u64 hwpoison_filter_memcg;
 136EXPORT_SYMBOL_GPL(hwpoison_filter_memcg);
 137static int hwpoison_filter_task(struct page *p)
 138{
 139	if (!hwpoison_filter_memcg)
 140		return 0;
 141
 142	if (page_cgroup_ino(p) != hwpoison_filter_memcg)
 143		return -EINVAL;
 144
 145	return 0;
 146}
 147#else
 148static int hwpoison_filter_task(struct page *p) { return 0; }
 149#endif
 150
 151int hwpoison_filter(struct page *p)
 152{
 153	if (!hwpoison_filter_enable)
 154		return 0;
 155
 156	if (hwpoison_filter_dev(p))
 157		return -EINVAL;
 158
 159	if (hwpoison_filter_flags(p))
 160		return -EINVAL;
 161
 162	if (hwpoison_filter_task(p))
 163		return -EINVAL;
 164
 165	return 0;
 166}
 167#else
 168int hwpoison_filter(struct page *p)
 169{
 170	return 0;
 171}
 172#endif
 173
 174EXPORT_SYMBOL_GPL(hwpoison_filter);
 175
 176/*
 177 * Send all the processes who have the page mapped a signal.
 178 * ``action optional'' if they are not immediately affected by the error
 179 * ``action required'' if error happened in current execution context
 180 */
 181static int kill_proc(struct task_struct *t, unsigned long addr, int trapno,
 182			unsigned long pfn, struct page *page, int flags)
 183{
 184	struct siginfo si;
 185	int ret;
 186
 187	pr_err("MCE %#lx: Killing %s:%d due to hardware memory corruption\n",
 188	       pfn, t->comm, t->pid);
 189	si.si_signo = SIGBUS;
 190	si.si_errno = 0;
 191	si.si_addr = (void *)addr;
 192#ifdef __ARCH_SI_TRAPNO
 193	si.si_trapno = trapno;
 194#endif
 195	si.si_addr_lsb = compound_order(compound_head(page)) + PAGE_SHIFT;
 196
 197	if ((flags & MF_ACTION_REQUIRED) && t->mm == current->mm) {
 198		si.si_code = BUS_MCEERR_AR;
 199		ret = force_sig_info(SIGBUS, &si, current);
 200	} else {
 201		/*
 202		 * Don't use force here, it's convenient if the signal
 203		 * can be temporarily blocked.
 204		 * This could cause a loop when the user sets SIGBUS
 205		 * to SIG_IGN, but hopefully no one will do that?
 206		 */
 207		si.si_code = BUS_MCEERR_AO;
 208		ret = send_sig_info(SIGBUS, &si, t);  /* synchronous? */
 209	}
 210	if (ret < 0)
 211		pr_info("MCE: Error sending signal to %s:%d: %d\n",
 212			t->comm, t->pid, ret);
 213	return ret;
 214}
 215
 216/*
 217 * When a unknown page type is encountered drain as many buffers as possible
 218 * in the hope to turn the page into a LRU or free page, which we can handle.
 219 */
 220void shake_page(struct page *p, int access)
 221{
 
 
 
 222	if (!PageSlab(p)) {
 223		lru_add_drain_all();
 224		if (PageLRU(p))
 225			return;
 226		drain_all_pages(page_zone(p));
 227		if (PageLRU(p) || is_free_buddy_page(p))
 228			return;
 229	}
 230
 231	/*
 232	 * Only call shrink_node_slabs here (which would also shrink
 233	 * other caches) if access is not potentially fatal.
 234	 */
 235	if (access)
 236		drop_slab_node(page_to_nid(p));
 237}
 238EXPORT_SYMBOL_GPL(shake_page);
 239
 240/*
 241 * Kill all processes that have a poisoned page mapped and then isolate
 242 * the page.
 243 *
 244 * General strategy:
 245 * Find all processes having the page mapped and kill them.
 246 * But we keep a page reference around so that the page is not
 247 * actually freed yet.
 248 * Then stash the page away
 249 *
 250 * There's no convenient way to get back to mapped processes
 251 * from the VMAs. So do a brute-force search over all
 252 * running processes.
 253 *
 254 * Remember that machine checks are not common (or rather
 255 * if they are common you have other problems), so this shouldn't
 256 * be a performance issue.
 257 *
 258 * Also there are some races possible while we get from the
 259 * error detection to actually handle it.
 260 */
 261
 262struct to_kill {
 263	struct list_head nd;
 264	struct task_struct *tsk;
 265	unsigned long addr;
 266	char addr_valid;
 267};
 268
 269/*
 270 * Failure handling: if we can't find or can't kill a process there's
 271 * not much we can do.	We just print a message and ignore otherwise.
 272 */
 273
 274/*
 275 * Schedule a process for later kill.
 276 * Uses GFP_ATOMIC allocations to avoid potential recursions in the VM.
 277 * TBD would GFP_NOIO be enough?
 278 */
 279static void add_to_kill(struct task_struct *tsk, struct page *p,
 280		       struct vm_area_struct *vma,
 281		       struct list_head *to_kill,
 282		       struct to_kill **tkc)
 283{
 284	struct to_kill *tk;
 285
 286	if (*tkc) {
 287		tk = *tkc;
 288		*tkc = NULL;
 289	} else {
 290		tk = kmalloc(sizeof(struct to_kill), GFP_ATOMIC);
 291		if (!tk) {
 292			pr_err("MCE: Out of memory while machine check handling\n");
 293			return;
 294		}
 295	}
 296	tk->addr = page_address_in_vma(p, vma);
 297	tk->addr_valid = 1;
 298
 299	/*
 300	 * In theory we don't have to kill when the page was
 301	 * munmaped. But it could be also a mremap. Since that's
 302	 * likely very rare kill anyways just out of paranoia, but use
 303	 * a SIGKILL because the error is not contained anymore.
 304	 */
 305	if (tk->addr == -EFAULT) {
 306		pr_info("MCE: Unable to find user space address %lx in %s\n",
 307			page_to_pfn(p), tsk->comm);
 308		tk->addr_valid = 0;
 309	}
 310	get_task_struct(tsk);
 311	tk->tsk = tsk;
 312	list_add_tail(&tk->nd, to_kill);
 313}
 314
 315/*
 316 * Kill the processes that have been collected earlier.
 317 *
 318 * Only do anything when DOIT is set, otherwise just free the list
 319 * (this is used for clean pages which do not need killing)
 320 * Also when FAIL is set do a force kill because something went
 321 * wrong earlier.
 322 */
 323static void kill_procs(struct list_head *to_kill, int forcekill, int trapno,
 324			  int fail, struct page *page, unsigned long pfn,
 325			  int flags)
 326{
 327	struct to_kill *tk, *next;
 328
 329	list_for_each_entry_safe (tk, next, to_kill, nd) {
 330		if (forcekill) {
 331			/*
 332			 * In case something went wrong with munmapping
 333			 * make sure the process doesn't catch the
 334			 * signal and then access the memory. Just kill it.
 335			 */
 336			if (fail || tk->addr_valid == 0) {
 337				pr_err("MCE %#lx: forcibly killing %s:%d because of failure to unmap corrupted page\n",
 338				       pfn, tk->tsk->comm, tk->tsk->pid);
 339				force_sig(SIGKILL, tk->tsk);
 340			}
 341
 342			/*
 343			 * In theory the process could have mapped
 344			 * something else on the address in-between. We could
 345			 * check for that, but we need to tell the
 346			 * process anyways.
 347			 */
 348			else if (kill_proc(tk->tsk, tk->addr, trapno,
 349					      pfn, page, flags) < 0)
 350				pr_err("MCE %#lx: Cannot send advisory machine check signal to %s:%d\n",
 351				       pfn, tk->tsk->comm, tk->tsk->pid);
 352		}
 353		put_task_struct(tk->tsk);
 354		kfree(tk);
 355	}
 356}
 357
 358/*
 359 * Find a dedicated thread which is supposed to handle SIGBUS(BUS_MCEERR_AO)
 360 * on behalf of the thread group. Return task_struct of the (first found)
 361 * dedicated thread if found, and return NULL otherwise.
 362 *
 363 * We already hold read_lock(&tasklist_lock) in the caller, so we don't
 364 * have to call rcu_read_lock/unlock() in this function.
 365 */
 366static struct task_struct *find_early_kill_thread(struct task_struct *tsk)
 367{
 368	struct task_struct *t;
 369
 370	for_each_thread(tsk, t)
 371		if ((t->flags & PF_MCE_PROCESS) && (t->flags & PF_MCE_EARLY))
 372			return t;
 373	return NULL;
 374}
 375
 376/*
 377 * Determine whether a given process is "early kill" process which expects
 378 * to be signaled when some page under the process is hwpoisoned.
 379 * Return task_struct of the dedicated thread (main thread unless explicitly
 380 * specified) if the process is "early kill," and otherwise returns NULL.
 381 */
 382static struct task_struct *task_early_kill(struct task_struct *tsk,
 383					   int force_early)
 384{
 385	struct task_struct *t;
 386	if (!tsk->mm)
 387		return NULL;
 388	if (force_early)
 389		return tsk;
 390	t = find_early_kill_thread(tsk);
 391	if (t)
 392		return t;
 393	if (sysctl_memory_failure_early_kill)
 394		return tsk;
 395	return NULL;
 396}
 397
 398/*
 399 * Collect processes when the error hit an anonymous page.
 400 */
 401static void collect_procs_anon(struct page *page, struct list_head *to_kill,
 402			      struct to_kill **tkc, int force_early)
 403{
 404	struct vm_area_struct *vma;
 405	struct task_struct *tsk;
 406	struct anon_vma *av;
 407	pgoff_t pgoff;
 408
 409	av = page_lock_anon_vma_read(page);
 410	if (av == NULL)	/* Not actually mapped anymore */
 411		return;
 412
 413	pgoff = page_to_pgoff(page);
 414	read_lock(&tasklist_lock);
 415	for_each_process (tsk) {
 416		struct anon_vma_chain *vmac;
 417		struct task_struct *t = task_early_kill(tsk, force_early);
 418
 419		if (!t)
 420			continue;
 421		anon_vma_interval_tree_foreach(vmac, &av->rb_root,
 422					       pgoff, pgoff) {
 423			vma = vmac->vma;
 424			if (!page_mapped_in_vma(page, vma))
 425				continue;
 426			if (vma->vm_mm == t->mm)
 427				add_to_kill(t, page, vma, to_kill, tkc);
 428		}
 429	}
 430	read_unlock(&tasklist_lock);
 431	page_unlock_anon_vma_read(av);
 432}
 433
 434/*
 435 * Collect processes when the error hit a file mapped page.
 436 */
 437static void collect_procs_file(struct page *page, struct list_head *to_kill,
 438			      struct to_kill **tkc, int force_early)
 439{
 440	struct vm_area_struct *vma;
 441	struct task_struct *tsk;
 442	struct address_space *mapping = page->mapping;
 443
 444	i_mmap_lock_read(mapping);
 445	read_lock(&tasklist_lock);
 446	for_each_process(tsk) {
 447		pgoff_t pgoff = page_to_pgoff(page);
 448		struct task_struct *t = task_early_kill(tsk, force_early);
 449
 450		if (!t)
 451			continue;
 452		vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff,
 453				      pgoff) {
 454			/*
 455			 * Send early kill signal to tasks where a vma covers
 456			 * the page but the corrupted page is not necessarily
 457			 * mapped it in its pte.
 458			 * Assume applications who requested early kill want
 459			 * to be informed of all such data corruptions.
 460			 */
 461			if (vma->vm_mm == t->mm)
 462				add_to_kill(t, page, vma, to_kill, tkc);
 463		}
 464	}
 465	read_unlock(&tasklist_lock);
 466	i_mmap_unlock_read(mapping);
 467}
 468
 469/*
 470 * Collect the processes who have the corrupted page mapped to kill.
 471 * This is done in two steps for locking reasons.
 472 * First preallocate one tokill structure outside the spin locks,
 473 * so that we can kill at least one process reasonably reliable.
 474 */
 475static void collect_procs(struct page *page, struct list_head *tokill,
 476				int force_early)
 477{
 478	struct to_kill *tk;
 479
 480	if (!page->mapping)
 481		return;
 482
 483	tk = kmalloc(sizeof(struct to_kill), GFP_NOIO);
 484	if (!tk)
 485		return;
 486	if (PageAnon(page))
 487		collect_procs_anon(page, tokill, &tk, force_early);
 488	else
 489		collect_procs_file(page, tokill, &tk, force_early);
 490	kfree(tk);
 491}
 492
 493static const char *action_name[] = {
 494	[MF_IGNORED] = "Ignored",
 495	[MF_FAILED] = "Failed",
 496	[MF_DELAYED] = "Delayed",
 497	[MF_RECOVERED] = "Recovered",
 498};
 499
 500static const char * const action_page_types[] = {
 501	[MF_MSG_KERNEL]			= "reserved kernel page",
 502	[MF_MSG_KERNEL_HIGH_ORDER]	= "high-order kernel page",
 503	[MF_MSG_SLAB]			= "kernel slab page",
 504	[MF_MSG_DIFFERENT_COMPOUND]	= "different compound page after locking",
 505	[MF_MSG_POISONED_HUGE]		= "huge page already hardware poisoned",
 506	[MF_MSG_HUGE]			= "huge page",
 507	[MF_MSG_FREE_HUGE]		= "free huge page",
 
 508	[MF_MSG_UNMAP_FAILED]		= "unmapping failed page",
 509	[MF_MSG_DIRTY_SWAPCACHE]	= "dirty swapcache page",
 510	[MF_MSG_CLEAN_SWAPCACHE]	= "clean swapcache page",
 511	[MF_MSG_DIRTY_MLOCKED_LRU]	= "dirty mlocked LRU page",
 512	[MF_MSG_CLEAN_MLOCKED_LRU]	= "clean mlocked LRU page",
 513	[MF_MSG_DIRTY_UNEVICTABLE_LRU]	= "dirty unevictable LRU page",
 514	[MF_MSG_CLEAN_UNEVICTABLE_LRU]	= "clean unevictable LRU page",
 515	[MF_MSG_DIRTY_LRU]		= "dirty LRU page",
 516	[MF_MSG_CLEAN_LRU]		= "clean LRU page",
 517	[MF_MSG_TRUNCATED_LRU]		= "already truncated LRU page",
 518	[MF_MSG_BUDDY]			= "free buddy page",
 519	[MF_MSG_BUDDY_2ND]		= "free buddy page (2nd try)",
 520	[MF_MSG_UNKNOWN]		= "unknown page",
 521};
 522
 523/*
 524 * XXX: It is possible that a page is isolated from LRU cache,
 525 * and then kept in swap cache or failed to remove from page cache.
 526 * The page count will stop it from being freed by unpoison.
 527 * Stress tests should be aware of this memory leak problem.
 528 */
 529static int delete_from_lru_cache(struct page *p)
 530{
 531	if (!isolate_lru_page(p)) {
 532		/*
 533		 * Clear sensible page flags, so that the buddy system won't
 534		 * complain when the page is unpoison-and-freed.
 535		 */
 536		ClearPageActive(p);
 537		ClearPageUnevictable(p);
 
 
 
 
 
 
 
 538		/*
 539		 * drop the page count elevated by isolate_lru_page()
 540		 */
 541		put_page(p);
 542		return 0;
 543	}
 544	return -EIO;
 545}
 546
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 547/*
 548 * Error hit kernel page.
 549 * Do nothing, try to be lucky and not touch this instead. For a few cases we
 550 * could be more sophisticated.
 551 */
 552static int me_kernel(struct page *p, unsigned long pfn)
 553{
 554	return MF_IGNORED;
 555}
 556
 557/*
 558 * Page in unknown state. Do nothing.
 559 */
 560static int me_unknown(struct page *p, unsigned long pfn)
 561{
 562	pr_err("MCE %#lx: Unknown page state\n", pfn);
 563	return MF_FAILED;
 564}
 565
 566/*
 567 * Clean (or cleaned) page cache page.
 568 */
 569static int me_pagecache_clean(struct page *p, unsigned long pfn)
 570{
 571	int err;
 572	int ret = MF_FAILED;
 573	struct address_space *mapping;
 574
 575	delete_from_lru_cache(p);
 576
 577	/*
 578	 * For anonymous pages we're done the only reference left
 579	 * should be the one m_f() holds.
 580	 */
 581	if (PageAnon(p))
 582		return MF_RECOVERED;
 583
 584	/*
 585	 * Now truncate the page in the page cache. This is really
 586	 * more like a "temporary hole punch"
 587	 * Don't do this for block devices when someone else
 588	 * has a reference, because it could be file system metadata
 589	 * and that's not safe to truncate.
 590	 */
 591	mapping = page_mapping(p);
 592	if (!mapping) {
 593		/*
 594		 * Page has been teared down in the meanwhile
 595		 */
 596		return MF_FAILED;
 597	}
 598
 599	/*
 600	 * Truncation is a bit tricky. Enable it per file system for now.
 601	 *
 602	 * Open: to take i_mutex or not for this? Right now we don't.
 603	 */
 604	if (mapping->a_ops->error_remove_page) {
 605		err = mapping->a_ops->error_remove_page(mapping, p);
 606		if (err != 0) {
 607			pr_info("MCE %#lx: Failed to punch page: %d\n",
 608				pfn, err);
 609		} else if (page_has_private(p) &&
 610				!try_to_release_page(p, GFP_NOIO)) {
 611			pr_info("MCE %#lx: failed to release buffers\n", pfn);
 612		} else {
 613			ret = MF_RECOVERED;
 614		}
 615	} else {
 616		/*
 617		 * If the file system doesn't support it just invalidate
 618		 * This fails on dirty or anything with private pages
 619		 */
 620		if (invalidate_inode_page(p))
 621			ret = MF_RECOVERED;
 622		else
 623			pr_info("MCE %#lx: Failed to invalidate\n", pfn);
 624	}
 625	return ret;
 626}
 627
 628/*
 629 * Dirty pagecache page
 630 * Issues: when the error hit a hole page the error is not properly
 631 * propagated.
 632 */
 633static int me_pagecache_dirty(struct page *p, unsigned long pfn)
 634{
 635	struct address_space *mapping = page_mapping(p);
 636
 637	SetPageError(p);
 638	/* TBD: print more information about the file. */
 639	if (mapping) {
 640		/*
 641		 * IO error will be reported by write(), fsync(), etc.
 642		 * who check the mapping.
 643		 * This way the application knows that something went
 644		 * wrong with its dirty file data.
 645		 *
 646		 * There's one open issue:
 647		 *
 648		 * The EIO will be only reported on the next IO
 649		 * operation and then cleared through the IO map.
 650		 * Normally Linux has two mechanisms to pass IO error
 651		 * first through the AS_EIO flag in the address space
 652		 * and then through the PageError flag in the page.
 653		 * Since we drop pages on memory failure handling the
 654		 * only mechanism open to use is through AS_AIO.
 655		 *
 656		 * This has the disadvantage that it gets cleared on
 657		 * the first operation that returns an error, while
 658		 * the PageError bit is more sticky and only cleared
 659		 * when the page is reread or dropped.  If an
 660		 * application assumes it will always get error on
 661		 * fsync, but does other operations on the fd before
 662		 * and the page is dropped between then the error
 663		 * will not be properly reported.
 664		 *
 665		 * This can already happen even without hwpoisoned
 666		 * pages: first on metadata IO errors (which only
 667		 * report through AS_EIO) or when the page is dropped
 668		 * at the wrong time.
 669		 *
 670		 * So right now we assume that the application DTRT on
 671		 * the first EIO, but we're not worse than other parts
 672		 * of the kernel.
 673		 */
 674		mapping_set_error(mapping, EIO);
 675	}
 676
 677	return me_pagecache_clean(p, pfn);
 678}
 679
 680/*
 681 * Clean and dirty swap cache.
 682 *
 683 * Dirty swap cache page is tricky to handle. The page could live both in page
 684 * cache and swap cache(ie. page is freshly swapped in). So it could be
 685 * referenced concurrently by 2 types of PTEs:
 686 * normal PTEs and swap PTEs. We try to handle them consistently by calling
 687 * try_to_unmap(TTU_IGNORE_HWPOISON) to convert the normal PTEs to swap PTEs,
 688 * and then
 689 *      - clear dirty bit to prevent IO
 690 *      - remove from LRU
 691 *      - but keep in the swap cache, so that when we return to it on
 692 *        a later page fault, we know the application is accessing
 693 *        corrupted data and shall be killed (we installed simple
 694 *        interception code in do_swap_page to catch it).
 695 *
 696 * Clean swap cache pages can be directly isolated. A later page fault will
 697 * bring in the known good data from disk.
 698 */
 699static int me_swapcache_dirty(struct page *p, unsigned long pfn)
 700{
 701	ClearPageDirty(p);
 702	/* Trigger EIO in shmem: */
 703	ClearPageUptodate(p);
 704
 705	if (!delete_from_lru_cache(p))
 706		return MF_DELAYED;
 707	else
 708		return MF_FAILED;
 709}
 710
 711static int me_swapcache_clean(struct page *p, unsigned long pfn)
 712{
 713	delete_from_swap_cache(p);
 714
 715	if (!delete_from_lru_cache(p))
 716		return MF_RECOVERED;
 717	else
 718		return MF_FAILED;
 719}
 720
 721/*
 722 * Huge pages. Needs work.
 723 * Issues:
 724 * - Error on hugepage is contained in hugepage unit (not in raw page unit.)
 725 *   To narrow down kill region to one page, we need to break up pmd.
 726 */
 727static int me_huge_page(struct page *p, unsigned long pfn)
 728{
 729	int res = 0;
 730	struct page *hpage = compound_head(p);
 
 731
 732	if (!PageHuge(hpage))
 733		return MF_DELAYED;
 734
 735	/*
 736	 * We can safely recover from error on free or reserved (i.e.
 737	 * not in-use) hugepage by dequeuing it from freelist.
 738	 * To check whether a hugepage is in-use or not, we can't use
 739	 * page->lru because it can be used in other hugepage operations,
 740	 * such as __unmap_hugepage_range() and gather_surplus_pages().
 741	 * So instead we use page_mapping() and PageAnon().
 742	 * We assume that this function is called with page lock held,
 743	 * so there is no race between isolation and mapping/unmapping.
 744	 */
 745	if (!(page_mapping(hpage) || PageAnon(hpage))) {
 746		res = dequeue_hwpoisoned_huge_page(hpage);
 747		if (!res)
 748			return MF_RECOVERED;
 
 749	}
 750	return MF_DELAYED;
 
 751}
 752
 753/*
 754 * Various page states we can handle.
 755 *
 756 * A page state is defined by its current page->flags bits.
 757 * The table matches them in order and calls the right handler.
 758 *
 759 * This is quite tricky because we can access page at any time
 760 * in its live cycle, so all accesses have to be extremely careful.
 761 *
 762 * This is not complete. More states could be added.
 763 * For any missing state don't attempt recovery.
 764 */
 765
 766#define dirty		(1UL << PG_dirty)
 767#define sc		(1UL << PG_swapcache)
 768#define unevict		(1UL << PG_unevictable)
 769#define mlock		(1UL << PG_mlocked)
 770#define writeback	(1UL << PG_writeback)
 771#define lru		(1UL << PG_lru)
 772#define swapbacked	(1UL << PG_swapbacked)
 773#define head		(1UL << PG_head)
 774#define slab		(1UL << PG_slab)
 775#define reserved	(1UL << PG_reserved)
 776
 777static struct page_state {
 778	unsigned long mask;
 779	unsigned long res;
 780	enum mf_action_page_type type;
 781	int (*action)(struct page *p, unsigned long pfn);
 782} error_states[] = {
 783	{ reserved,	reserved,	MF_MSG_KERNEL,	me_kernel },
 784	/*
 785	 * free pages are specially detected outside this table:
 786	 * PG_buddy pages only make a small fraction of all free pages.
 787	 */
 788
 789	/*
 790	 * Could in theory check if slab page is free or if we can drop
 791	 * currently unused objects without touching them. But just
 792	 * treat it as standard kernel for now.
 793	 */
 794	{ slab,		slab,		MF_MSG_SLAB,	me_kernel },
 795
 796	{ head,		head,		MF_MSG_HUGE,		me_huge_page },
 797
 798	{ sc|dirty,	sc|dirty,	MF_MSG_DIRTY_SWAPCACHE,	me_swapcache_dirty },
 799	{ sc|dirty,	sc,		MF_MSG_CLEAN_SWAPCACHE,	me_swapcache_clean },
 800
 801	{ mlock|dirty,	mlock|dirty,	MF_MSG_DIRTY_MLOCKED_LRU,	me_pagecache_dirty },
 802	{ mlock|dirty,	mlock,		MF_MSG_CLEAN_MLOCKED_LRU,	me_pagecache_clean },
 803
 804	{ unevict|dirty, unevict|dirty,	MF_MSG_DIRTY_UNEVICTABLE_LRU,	me_pagecache_dirty },
 805	{ unevict|dirty, unevict,	MF_MSG_CLEAN_UNEVICTABLE_LRU,	me_pagecache_clean },
 806
 807	{ lru|dirty,	lru|dirty,	MF_MSG_DIRTY_LRU,	me_pagecache_dirty },
 808	{ lru|dirty,	lru,		MF_MSG_CLEAN_LRU,	me_pagecache_clean },
 809
 810	/*
 811	 * Catchall entry: must be at end.
 812	 */
 813	{ 0,		0,		MF_MSG_UNKNOWN,	me_unknown },
 814};
 815
 816#undef dirty
 817#undef sc
 818#undef unevict
 819#undef mlock
 820#undef writeback
 821#undef lru
 822#undef swapbacked
 823#undef head
 824#undef slab
 825#undef reserved
 826
 827/*
 828 * "Dirty/Clean" indication is not 100% accurate due to the possibility of
 829 * setting PG_dirty outside page lock. See also comment above set_page_dirty().
 830 */
 831static void action_result(unsigned long pfn, enum mf_action_page_type type,
 832			  enum mf_result result)
 833{
 834	trace_memory_failure_event(pfn, type, result);
 835
 836	pr_err("MCE %#lx: recovery action for %s: %s\n",
 837		pfn, action_page_types[type], action_name[result]);
 838}
 839
 840static int page_action(struct page_state *ps, struct page *p,
 841			unsigned long pfn)
 842{
 843	int result;
 844	int count;
 845
 846	result = ps->action(p, pfn);
 847
 848	count = page_count(p) - 1;
 849	if (ps->action == me_swapcache_dirty && result == MF_DELAYED)
 850		count--;
 851	if (count != 0) {
 852		pr_err("MCE %#lx: %s still referenced by %d users\n",
 853		       pfn, action_page_types[ps->type], count);
 854		result = MF_FAILED;
 855	}
 856	action_result(pfn, ps->type, result);
 857
 858	/* Could do more checks here if page looks ok */
 859	/*
 860	 * Could adjust zone counters here to correct for the missing page.
 861	 */
 862
 863	return (result == MF_RECOVERED || result == MF_DELAYED) ? 0 : -EBUSY;
 864}
 865
 866/**
 867 * get_hwpoison_page() - Get refcount for memory error handling:
 868 * @page:	raw error page (hit by memory error)
 869 *
 870 * Return: return 0 if failed to grab the refcount, otherwise true (some
 871 * non-zero value.)
 872 */
 873int get_hwpoison_page(struct page *page)
 874{
 875	struct page *head = compound_head(page);
 876
 877	if (!PageHuge(head) && PageTransHuge(head)) {
 878		/*
 879		 * Non anonymous thp exists only in allocation/free time. We
 880		 * can't handle such a case correctly, so let's give it up.
 881		 * This should be better than triggering BUG_ON when kernel
 882		 * tries to touch the "partially handled" page.
 883		 */
 884		if (!PageAnon(head)) {
 885			pr_err("MCE: %#lx: non anonymous thp\n",
 886				page_to_pfn(page));
 887			return 0;
 888		}
 889	}
 890
 891	if (get_page_unless_zero(head)) {
 892		if (head == compound_head(page))
 893			return 1;
 894
 895		pr_info("MCE: %#lx cannot catch tail\n", page_to_pfn(page));
 
 896		put_page(head);
 897	}
 898
 899	return 0;
 900}
 901EXPORT_SYMBOL_GPL(get_hwpoison_page);
 902
 903/*
 904 * Do all that is necessary to remove user space mappings. Unmap
 905 * the pages and send SIGBUS to the processes if the data was dirty.
 906 */
 907static int hwpoison_user_mappings(struct page *p, unsigned long pfn,
 908				  int trapno, int flags, struct page **hpagep)
 909{
 910	enum ttu_flags ttu = TTU_UNMAP | TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS;
 911	struct address_space *mapping;
 912	LIST_HEAD(tokill);
 913	int ret;
 914	int kill = 1, forcekill;
 915	struct page *hpage = *hpagep;
 
 916
 917	/*
 918	 * Here we are interested only in user-mapped pages, so skip any
 919	 * other types of pages.
 920	 */
 921	if (PageReserved(p) || PageSlab(p))
 922		return SWAP_SUCCESS;
 923	if (!(PageLRU(hpage) || PageHuge(p)))
 924		return SWAP_SUCCESS;
 925
 926	/*
 927	 * This check implies we don't kill processes if their pages
 928	 * are in the swap cache early. Those are always late kills.
 929	 */
 930	if (!page_mapped(hpage))
 931		return SWAP_SUCCESS;
 932
 933	if (PageKsm(p)) {
 934		pr_err("MCE %#lx: can't handle KSM pages.\n", pfn);
 935		return SWAP_FAIL;
 936	}
 937
 938	if (PageSwapCache(p)) {
 939		pr_err("MCE %#lx: keeping poisoned page in swap cache\n", pfn);
 
 940		ttu |= TTU_IGNORE_HWPOISON;
 941	}
 942
 943	/*
 944	 * Propagate the dirty bit from PTEs to struct page first, because we
 945	 * need this to decide if we should kill or just drop the page.
 946	 * XXX: the dirty test could be racy: set_page_dirty() may not always
 947	 * be called inside page lock (it's recommended but not enforced).
 948	 */
 949	mapping = page_mapping(hpage);
 950	if (!(flags & MF_MUST_KILL) && !PageDirty(hpage) && mapping &&
 951	    mapping_cap_writeback_dirty(mapping)) {
 952		if (page_mkclean(hpage)) {
 953			SetPageDirty(hpage);
 954		} else {
 955			kill = 0;
 956			ttu |= TTU_IGNORE_HWPOISON;
 957			pr_info("MCE %#lx: corrupted page was clean: dropped without side effects\n",
 958				pfn);
 959		}
 960	}
 961
 962	/*
 963	 * First collect all the processes that have the page
 964	 * mapped in dirty form.  This has to be done before try_to_unmap,
 965	 * because ttu takes the rmap data structures down.
 966	 *
 967	 * Error handling: We ignore errors here because
 968	 * there's nothing that can be done.
 969	 */
 970	if (kill)
 971		collect_procs(hpage, &tokill, flags & MF_ACTION_REQUIRED);
 972
 973	ret = try_to_unmap(hpage, ttu);
 974	if (ret != SWAP_SUCCESS)
 975		pr_err("MCE %#lx: failed to unmap page (mapcount=%d)\n",
 976		       pfn, page_mapcount(hpage));
 977
 978	/*
 
 
 
 
 
 
 
 979	 * Now that the dirty bit has been propagated to the
 980	 * struct page and all unmaps done we can decide if
 981	 * killing is needed or not.  Only kill when the page
 982	 * was dirty or the process is not restartable,
 983	 * otherwise the tokill list is merely
 984	 * freed.  When there was a problem unmapping earlier
 985	 * use a more force-full uncatchable kill to prevent
 986	 * any accesses to the poisoned memory.
 987	 */
 988	forcekill = PageDirty(hpage) || (flags & MF_MUST_KILL);
 989	kill_procs(&tokill, forcekill, trapno,
 990		      ret != SWAP_SUCCESS, p, pfn, flags);
 991
 992	return ret;
 993}
 994
 995static void set_page_hwpoison_huge_page(struct page *hpage)
 
 996{
 997	int i;
 998	int nr_pages = 1 << compound_order(hpage);
 999	for (i = 0; i < nr_pages; i++)
1000		SetPageHWPoison(hpage + i);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1001}
1002
1003static void clear_page_hwpoison_huge_page(struct page *hpage)
1004{
1005	int i;
1006	int nr_pages = 1 << compound_order(hpage);
1007	for (i = 0; i < nr_pages; i++)
1008		ClearPageHWPoison(hpage + i);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1009}
1010
1011/**
1012 * memory_failure - Handle memory failure of a page.
1013 * @pfn: Page Number of the corrupted page
1014 * @trapno: Trap number reported in the signal to user space.
1015 * @flags: fine tune action taken
1016 *
1017 * This function is called by the low level machine check code
1018 * of an architecture when it detects hardware memory corruption
1019 * of a page. It tries its best to recover, which includes
1020 * dropping pages, killing processes etc.
1021 *
1022 * The function is primarily of use for corruptions that
1023 * happen outside the current execution context (e.g. when
1024 * detected by a background scrubber)
1025 *
1026 * Must run in process context (e.g. a work queue) with interrupts
1027 * enabled and no spinlocks hold.
1028 */
1029int memory_failure(unsigned long pfn, int trapno, int flags)
1030{
1031	struct page_state *ps;
1032	struct page *p;
1033	struct page *hpage;
1034	struct page *orig_head;
1035	int res;
1036	unsigned int nr_pages;
1037	unsigned long page_flags;
1038
1039	if (!sysctl_memory_failure_recovery)
1040		panic("Memory failure from trap %d on page %lx", trapno, pfn);
1041
1042	if (!pfn_valid(pfn)) {
1043		pr_err("MCE %#lx: memory outside kernel control\n", pfn);
 
1044		return -ENXIO;
1045	}
1046
1047	p = pfn_to_page(pfn);
1048	orig_head = hpage = compound_head(p);
 
1049	if (TestSetPageHWPoison(p)) {
1050		pr_err("MCE %#lx: already hardware poisoned\n", pfn);
 
1051		return 0;
1052	}
1053
1054	/*
1055	 * Currently errors on hugetlbfs pages are measured in hugepage units,
1056	 * so nr_pages should be 1 << compound_order.  OTOH when errors are on
1057	 * transparent hugepages, they are supposed to be split and error
1058	 * measurement is done in normal page units.  So nr_pages should be one
1059	 * in this case.
1060	 */
1061	if (PageHuge(p))
1062		nr_pages = 1 << compound_order(hpage);
1063	else /* normal page or thp */
1064		nr_pages = 1;
1065	num_poisoned_pages_add(nr_pages);
1066
1067	/*
1068	 * We need/can do nothing about count=0 pages.
1069	 * 1) it's a free page, and therefore in safe hand:
1070	 *    prep_new_page() will be the gate keeper.
1071	 * 2) it's a free hugepage, which is also safe:
1072	 *    an affected hugepage will be dequeued from hugepage freelist,
1073	 *    so there's no concern about reusing it ever after.
1074	 * 3) it's part of a non-compound high order page.
1075	 *    Implies some kernel user: cannot stop them from
1076	 *    R/W the page; let's pray that the page has been
1077	 *    used and will be freed some time later.
1078	 * In fact it's dangerous to directly bump up page count from 0,
1079	 * that may make page_freeze_refs()/page_unfreeze_refs() mismatch.
1080	 */
1081	if (!(flags & MF_COUNT_INCREASED) && !get_hwpoison_page(p)) {
1082		if (is_free_buddy_page(p)) {
1083			action_result(pfn, MF_MSG_BUDDY, MF_DELAYED);
1084			return 0;
1085		} else if (PageHuge(hpage)) {
1086			/*
1087			 * Check "filter hit" and "race with other subpage."
1088			 */
1089			lock_page(hpage);
1090			if (PageHWPoison(hpage)) {
1091				if ((hwpoison_filter(p) && TestClearPageHWPoison(p))
1092				    || (p != hpage && TestSetPageHWPoison(hpage))) {
1093					num_poisoned_pages_sub(nr_pages);
1094					unlock_page(hpage);
1095					return 0;
1096				}
1097			}
1098			set_page_hwpoison_huge_page(hpage);
1099			res = dequeue_hwpoisoned_huge_page(hpage);
1100			action_result(pfn, MF_MSG_FREE_HUGE,
1101				      res ? MF_IGNORED : MF_DELAYED);
1102			unlock_page(hpage);
1103			return res;
1104		} else {
1105			action_result(pfn, MF_MSG_KERNEL_HIGH_ORDER, MF_IGNORED);
1106			return -EBUSY;
1107		}
1108	}
1109
1110	if (!PageHuge(p) && PageTransHuge(hpage)) {
1111		lock_page(hpage);
1112		if (!PageAnon(hpage) || unlikely(split_huge_page(hpage))) {
1113			unlock_page(hpage);
1114			if (!PageAnon(hpage))
1115				pr_err("MCE: %#lx: non anonymous thp\n", pfn);
 
1116			else
1117				pr_err("MCE: %#lx: thp split failed\n", pfn);
 
1118			if (TestClearPageHWPoison(p))
1119				num_poisoned_pages_sub(nr_pages);
1120			put_hwpoison_page(p);
1121			return -EBUSY;
1122		}
1123		unlock_page(hpage);
1124		get_hwpoison_page(p);
1125		put_hwpoison_page(hpage);
1126		VM_BUG_ON_PAGE(!page_count(p), p);
1127		hpage = compound_head(p);
1128	}
1129
1130	/*
1131	 * We ignore non-LRU pages for good reasons.
1132	 * - PG_locked is only well defined for LRU pages and a few others
1133	 * - to avoid races with __SetPageLocked()
1134	 * - to avoid races with __SetPageSlab*() (and more non-atomic ops)
1135	 * The check (unnecessarily) ignores LRU pages being isolated and
1136	 * walked by the page reclaim code, however that's not a big loss.
1137	 */
1138	if (!PageHuge(p)) {
1139		if (!PageLRU(p))
1140			shake_page(p, 0);
1141		if (!PageLRU(p)) {
1142			/*
1143			 * shake_page could have turned it free.
1144			 */
1145			if (is_free_buddy_page(p)) {
1146				if (flags & MF_COUNT_INCREASED)
1147					action_result(pfn, MF_MSG_BUDDY, MF_DELAYED);
1148				else
1149					action_result(pfn, MF_MSG_BUDDY_2ND,
1150						      MF_DELAYED);
1151				return 0;
1152			}
1153		}
1154	}
1155
1156	lock_page(hpage);
1157
1158	/*
1159	 * The page could have changed compound pages during the locking.
1160	 * If this happens just bail out.
1161	 */
1162	if (PageCompound(p) && compound_head(p) != orig_head) {
1163		action_result(pfn, MF_MSG_DIFFERENT_COMPOUND, MF_IGNORED);
1164		res = -EBUSY;
1165		goto out;
1166	}
1167
1168	/*
1169	 * We use page flags to determine what action should be taken, but
1170	 * the flags can be modified by the error containment action.  One
1171	 * example is an mlocked page, where PG_mlocked is cleared by
1172	 * page_remove_rmap() in try_to_unmap_one(). So to determine page status
1173	 * correctly, we save a copy of the page flags at this time.
1174	 */
1175	page_flags = p->flags;
 
 
 
1176
1177	/*
1178	 * unpoison always clear PG_hwpoison inside page lock
1179	 */
1180	if (!PageHWPoison(p)) {
1181		pr_err("MCE %#lx: just unpoisoned\n", pfn);
1182		num_poisoned_pages_sub(nr_pages);
1183		unlock_page(hpage);
1184		put_hwpoison_page(hpage);
1185		return 0;
1186	}
1187	if (hwpoison_filter(p)) {
1188		if (TestClearPageHWPoison(p))
1189			num_poisoned_pages_sub(nr_pages);
1190		unlock_page(hpage);
1191		put_hwpoison_page(hpage);
1192		return 0;
1193	}
1194
1195	if (!PageHuge(p) && !PageTransTail(p) && !PageLRU(p))
1196		goto identify_page_state;
1197
1198	/*
1199	 * For error on the tail page, we should set PG_hwpoison
1200	 * on the head page to show that the hugepage is hwpoisoned
1201	 */
1202	if (PageHuge(p) && PageTail(p) && TestSetPageHWPoison(hpage)) {
1203		action_result(pfn, MF_MSG_POISONED_HUGE, MF_IGNORED);
1204		unlock_page(hpage);
1205		put_hwpoison_page(hpage);
1206		return 0;
1207	}
1208	/*
1209	 * Set PG_hwpoison on all pages in an error hugepage,
1210	 * because containment is done in hugepage unit for now.
1211	 * Since we have done TestSetPageHWPoison() for the head page with
1212	 * page lock held, we can safely set PG_hwpoison bits on tail pages.
1213	 */
1214	if (PageHuge(p))
1215		set_page_hwpoison_huge_page(hpage);
1216
1217	/*
1218	 * It's very difficult to mess with pages currently under IO
1219	 * and in many cases impossible, so we just avoid it here.
1220	 */
1221	wait_on_page_writeback(p);
1222
1223	/*
1224	 * Now take care of user space mappings.
1225	 * Abort on fail: __delete_from_page_cache() assumes unmapped page.
1226	 *
1227	 * When the raw error page is thp tail page, hpage points to the raw
1228	 * page after thp split.
1229	 */
1230	if (hwpoison_user_mappings(p, pfn, trapno, flags, &hpage)
1231	    != SWAP_SUCCESS) {
1232		action_result(pfn, MF_MSG_UNMAP_FAILED, MF_IGNORED);
1233		res = -EBUSY;
1234		goto out;
1235	}
1236
1237	/*
1238	 * Torn down by someone else?
1239	 */
1240	if (PageLRU(p) && !PageSwapCache(p) && p->mapping == NULL) {
1241		action_result(pfn, MF_MSG_TRUNCATED_LRU, MF_IGNORED);
1242		res = -EBUSY;
1243		goto out;
1244	}
1245
1246identify_page_state:
1247	res = -EBUSY;
1248	/*
1249	 * The first check uses the current page flags which may not have any
1250	 * relevant information. The second check with the saved page flagss is
1251	 * carried out only if the first check can't determine the page status.
1252	 */
1253	for (ps = error_states;; ps++)
1254		if ((p->flags & ps->mask) == ps->res)
1255			break;
1256
1257	page_flags |= (p->flags & (1UL << PG_dirty));
1258
1259	if (!ps->mask)
1260		for (ps = error_states;; ps++)
1261			if ((page_flags & ps->mask) == ps->res)
1262				break;
1263	res = page_action(ps, p, pfn);
1264out:
1265	unlock_page(hpage);
1266	return res;
1267}
1268EXPORT_SYMBOL_GPL(memory_failure);
1269
1270#define MEMORY_FAILURE_FIFO_ORDER	4
1271#define MEMORY_FAILURE_FIFO_SIZE	(1 << MEMORY_FAILURE_FIFO_ORDER)
1272
1273struct memory_failure_entry {
1274	unsigned long pfn;
1275	int trapno;
1276	int flags;
1277};
1278
1279struct memory_failure_cpu {
1280	DECLARE_KFIFO(fifo, struct memory_failure_entry,
1281		      MEMORY_FAILURE_FIFO_SIZE);
1282	spinlock_t lock;
1283	struct work_struct work;
1284};
1285
1286static DEFINE_PER_CPU(struct memory_failure_cpu, memory_failure_cpu);
1287
1288/**
1289 * memory_failure_queue - Schedule handling memory failure of a page.
1290 * @pfn: Page Number of the corrupted page
1291 * @trapno: Trap number reported in the signal to user space.
1292 * @flags: Flags for memory failure handling
1293 *
1294 * This function is called by the low level hardware error handler
1295 * when it detects hardware memory corruption of a page. It schedules
1296 * the recovering of error page, including dropping pages, killing
1297 * processes etc.
1298 *
1299 * The function is primarily of use for corruptions that
1300 * happen outside the current execution context (e.g. when
1301 * detected by a background scrubber)
1302 *
1303 * Can run in IRQ context.
1304 */
1305void memory_failure_queue(unsigned long pfn, int trapno, int flags)
1306{
1307	struct memory_failure_cpu *mf_cpu;
1308	unsigned long proc_flags;
1309	struct memory_failure_entry entry = {
1310		.pfn =		pfn,
1311		.trapno =	trapno,
1312		.flags =	flags,
1313	};
1314
1315	mf_cpu = &get_cpu_var(memory_failure_cpu);
1316	spin_lock_irqsave(&mf_cpu->lock, proc_flags);
1317	if (kfifo_put(&mf_cpu->fifo, entry))
1318		schedule_work_on(smp_processor_id(), &mf_cpu->work);
1319	else
1320		pr_err("Memory failure: buffer overflow when queuing memory failure at %#lx\n",
1321		       pfn);
1322	spin_unlock_irqrestore(&mf_cpu->lock, proc_flags);
1323	put_cpu_var(memory_failure_cpu);
1324}
1325EXPORT_SYMBOL_GPL(memory_failure_queue);
1326
1327static void memory_failure_work_func(struct work_struct *work)
1328{
1329	struct memory_failure_cpu *mf_cpu;
1330	struct memory_failure_entry entry = { 0, };
1331	unsigned long proc_flags;
1332	int gotten;
1333
1334	mf_cpu = this_cpu_ptr(&memory_failure_cpu);
1335	for (;;) {
1336		spin_lock_irqsave(&mf_cpu->lock, proc_flags);
1337		gotten = kfifo_get(&mf_cpu->fifo, &entry);
1338		spin_unlock_irqrestore(&mf_cpu->lock, proc_flags);
1339		if (!gotten)
1340			break;
1341		if (entry.flags & MF_SOFT_OFFLINE)
1342			soft_offline_page(pfn_to_page(entry.pfn), entry.flags);
1343		else
1344			memory_failure(entry.pfn, entry.trapno, entry.flags);
1345	}
1346}
1347
1348static int __init memory_failure_init(void)
1349{
1350	struct memory_failure_cpu *mf_cpu;
1351	int cpu;
1352
1353	for_each_possible_cpu(cpu) {
1354		mf_cpu = &per_cpu(memory_failure_cpu, cpu);
1355		spin_lock_init(&mf_cpu->lock);
1356		INIT_KFIFO(mf_cpu->fifo);
1357		INIT_WORK(&mf_cpu->work, memory_failure_work_func);
1358	}
1359
1360	return 0;
1361}
1362core_initcall(memory_failure_init);
1363
1364#define unpoison_pr_info(fmt, pfn, rs)			\
1365({							\
1366	if (__ratelimit(rs))				\
1367		pr_info(fmt, pfn);			\
1368})
1369
1370/**
1371 * unpoison_memory - Unpoison a previously poisoned page
1372 * @pfn: Page number of the to be unpoisoned page
1373 *
1374 * Software-unpoison a page that has been poisoned by
1375 * memory_failure() earlier.
1376 *
1377 * This is only done on the software-level, so it only works
1378 * for linux injected failures, not real hardware failures
1379 *
1380 * Returns 0 for success, otherwise -errno.
1381 */
1382int unpoison_memory(unsigned long pfn)
1383{
1384	struct page *page;
1385	struct page *p;
1386	int freeit = 0;
1387	unsigned int nr_pages;
1388	static DEFINE_RATELIMIT_STATE(unpoison_rs, DEFAULT_RATELIMIT_INTERVAL,
1389					DEFAULT_RATELIMIT_BURST);
1390
1391	if (!pfn_valid(pfn))
1392		return -ENXIO;
1393
1394	p = pfn_to_page(pfn);
1395	page = compound_head(p);
1396
1397	if (!PageHWPoison(p)) {
1398		unpoison_pr_info("MCE: Page was already unpoisoned %#lx\n",
1399				 pfn, &unpoison_rs);
1400		return 0;
1401	}
1402
1403	if (page_count(page) > 1) {
1404		unpoison_pr_info("MCE: Someone grabs the hwpoison page %#lx\n",
1405				 pfn, &unpoison_rs);
1406		return 0;
1407	}
1408
1409	if (page_mapped(page)) {
1410		unpoison_pr_info("MCE: Someone maps the hwpoison page %#lx\n",
1411				 pfn, &unpoison_rs);
1412		return 0;
1413	}
1414
1415	if (page_mapping(page)) {
1416		unpoison_pr_info("MCE: the hwpoison page has non-NULL mapping %#lx\n",
1417				 pfn, &unpoison_rs);
1418		return 0;
1419	}
1420
1421	/*
1422	 * unpoison_memory() can encounter thp only when the thp is being
1423	 * worked by memory_failure() and the page lock is not held yet.
1424	 * In such case, we yield to memory_failure() and make unpoison fail.
1425	 */
1426	if (!PageHuge(page) && PageTransHuge(page)) {
1427		unpoison_pr_info("MCE: Memory failure is now running on %#lx\n",
1428				 pfn, &unpoison_rs);
1429		return 0;
1430	}
1431
1432	nr_pages = 1 << compound_order(page);
1433
1434	if (!get_hwpoison_page(p)) {
1435		/*
1436		 * Since HWPoisoned hugepage should have non-zero refcount,
1437		 * race between memory failure and unpoison seems to happen.
1438		 * In such case unpoison fails and memory failure runs
1439		 * to the end.
1440		 */
1441		if (PageHuge(page)) {
1442			unpoison_pr_info("MCE: Memory failure is now running on free hugepage %#lx\n",
1443					 pfn, &unpoison_rs);
1444			return 0;
1445		}
1446		if (TestClearPageHWPoison(p))
1447			num_poisoned_pages_dec();
1448		unpoison_pr_info("MCE: Software-unpoisoned free page %#lx\n",
1449				 pfn, &unpoison_rs);
1450		return 0;
1451	}
1452
1453	lock_page(page);
1454	/*
1455	 * This test is racy because PG_hwpoison is set outside of page lock.
1456	 * That's acceptable because that won't trigger kernel panic. Instead,
1457	 * the PG_hwpoison page will be caught and isolated on the entrance to
1458	 * the free buddy page pool.
1459	 */
1460	if (TestClearPageHWPoison(page)) {
1461		unpoison_pr_info("MCE: Software-unpoisoned page %#lx\n",
1462				 pfn, &unpoison_rs);
1463		num_poisoned_pages_sub(nr_pages);
1464		freeit = 1;
1465		if (PageHuge(page))
1466			clear_page_hwpoison_huge_page(page);
1467	}
1468	unlock_page(page);
1469
1470	put_hwpoison_page(page);
1471	if (freeit && !(pfn == my_zero_pfn(0) && page_count(p) == 1))
1472		put_hwpoison_page(page);
1473
1474	return 0;
1475}
1476EXPORT_SYMBOL(unpoison_memory);
1477
1478static struct page *new_page(struct page *p, unsigned long private, int **x)
1479{
1480	int nid = page_to_nid(p);
1481	if (PageHuge(p))
1482		return alloc_huge_page_node(page_hstate(compound_head(p)),
1483						   nid);
1484	else
1485		return __alloc_pages_node(nid, GFP_HIGHUSER_MOVABLE, 0);
1486}
1487
1488/*
1489 * Safely get reference count of an arbitrary page.
1490 * Returns 0 for a free page, -EIO for a zero refcount page
1491 * that is not free, and 1 for any other page type.
1492 * For 1 the page is returned with increased page count, otherwise not.
1493 */
1494static int __get_any_page(struct page *p, unsigned long pfn, int flags)
1495{
1496	int ret;
1497
1498	if (flags & MF_COUNT_INCREASED)
1499		return 1;
1500
1501	/*
1502	 * When the target page is a free hugepage, just remove it
1503	 * from free hugepage list.
1504	 */
1505	if (!get_hwpoison_page(p)) {
1506		if (PageHuge(p)) {
1507			pr_info("%s: %#lx free huge page\n", __func__, pfn);
1508			ret = 0;
1509		} else if (is_free_buddy_page(p)) {
1510			pr_info("%s: %#lx free buddy page\n", __func__, pfn);
1511			ret = 0;
1512		} else {
1513			pr_info("%s: %#lx: unknown zero refcount page type %lx\n",
1514				__func__, pfn, p->flags);
1515			ret = -EIO;
1516		}
1517	} else {
1518		/* Not a free page */
1519		ret = 1;
1520	}
1521	return ret;
1522}
1523
1524static int get_any_page(struct page *page, unsigned long pfn, int flags)
1525{
1526	int ret = __get_any_page(page, pfn, flags);
1527
1528	if (ret == 1 && !PageHuge(page) && !PageLRU(page)) {
 
1529		/*
1530		 * Try to free it.
1531		 */
1532		put_hwpoison_page(page);
1533		shake_page(page, 1);
1534
1535		/*
1536		 * Did it turn free?
1537		 */
1538		ret = __get_any_page(page, pfn, 0);
1539		if (ret == 1 && !PageLRU(page)) {
1540			/* Drop page reference which is from __get_any_page() */
1541			put_hwpoison_page(page);
1542			pr_info("soft_offline: %#lx: unknown non LRU page type %lx\n",
1543				pfn, page->flags);
1544			return -EIO;
1545		}
1546	}
1547	return ret;
1548}
1549
1550static int soft_offline_huge_page(struct page *page, int flags)
1551{
1552	int ret;
1553	unsigned long pfn = page_to_pfn(page);
1554	struct page *hpage = compound_head(page);
1555	LIST_HEAD(pagelist);
1556
1557	/*
1558	 * This double-check of PageHWPoison is to avoid the race with
1559	 * memory_failure(). See also comment in __soft_offline_page().
1560	 */
1561	lock_page(hpage);
1562	if (PageHWPoison(hpage)) {
1563		unlock_page(hpage);
1564		put_hwpoison_page(hpage);
1565		pr_info("soft offline: %#lx hugepage already poisoned\n", pfn);
1566		return -EBUSY;
1567	}
1568	unlock_page(hpage);
1569
1570	ret = isolate_huge_page(hpage, &pagelist);
1571	/*
1572	 * get_any_page() and isolate_huge_page() takes a refcount each,
1573	 * so need to drop one here.
1574	 */
1575	put_hwpoison_page(hpage);
1576	if (!ret) {
1577		pr_info("soft offline: %#lx hugepage failed to isolate\n", pfn);
1578		return -EBUSY;
1579	}
1580
1581	ret = migrate_pages(&pagelist, new_page, NULL, MPOL_MF_MOVE_ALL,
1582				MIGRATE_SYNC, MR_MEMORY_FAILURE);
1583	if (ret) {
1584		pr_info("soft offline: %#lx: migration failed %d, type %lx\n",
1585			pfn, ret, page->flags);
1586		/*
1587		 * We know that soft_offline_huge_page() tries to migrate
1588		 * only one hugepage pointed to by hpage, so we need not
1589		 * run through the pagelist here.
1590		 */
1591		putback_active_hugepage(hpage);
1592		if (ret > 0)
1593			ret = -EIO;
1594	} else {
1595		/* overcommit hugetlb page will be freed to buddy */
1596		if (PageHuge(page)) {
1597			set_page_hwpoison_huge_page(hpage);
1598			dequeue_hwpoisoned_huge_page(hpage);
1599			num_poisoned_pages_add(1 << compound_order(hpage));
1600		} else {
1601			SetPageHWPoison(page);
1602			num_poisoned_pages_inc();
1603		}
1604	}
1605	return ret;
1606}
1607
1608static int __soft_offline_page(struct page *page, int flags)
1609{
1610	int ret;
1611	unsigned long pfn = page_to_pfn(page);
1612
1613	/*
1614	 * Check PageHWPoison again inside page lock because PageHWPoison
1615	 * is set by memory_failure() outside page lock. Note that
1616	 * memory_failure() also double-checks PageHWPoison inside page lock,
1617	 * so there's no race between soft_offline_page() and memory_failure().
1618	 */
1619	lock_page(page);
1620	wait_on_page_writeback(page);
1621	if (PageHWPoison(page)) {
1622		unlock_page(page);
1623		put_hwpoison_page(page);
1624		pr_info("soft offline: %#lx page already poisoned\n", pfn);
1625		return -EBUSY;
1626	}
1627	/*
1628	 * Try to invalidate first. This should work for
1629	 * non dirty unmapped page cache pages.
1630	 */
1631	ret = invalidate_inode_page(page);
1632	unlock_page(page);
1633	/*
1634	 * RED-PEN would be better to keep it isolated here, but we
1635	 * would need to fix isolation locking first.
1636	 */
1637	if (ret == 1) {
1638		put_hwpoison_page(page);
1639		pr_info("soft_offline: %#lx: invalidated\n", pfn);
1640		SetPageHWPoison(page);
1641		num_poisoned_pages_inc();
1642		return 0;
1643	}
1644
1645	/*
1646	 * Simple invalidation didn't work.
1647	 * Try to migrate to a new page instead. migrate.c
1648	 * handles a large number of cases for us.
1649	 */
1650	ret = isolate_lru_page(page);
 
 
 
1651	/*
1652	 * Drop page reference which is came from get_any_page()
1653	 * successful isolate_lru_page() already took another one.
1654	 */
1655	put_hwpoison_page(page);
1656	if (!ret) {
1657		LIST_HEAD(pagelist);
1658		inc_zone_page_state(page, NR_ISOLATED_ANON +
1659					page_is_file_cache(page));
 
 
 
 
 
 
1660		list_add(&page->lru, &pagelist);
1661		ret = migrate_pages(&pagelist, new_page, NULL, MPOL_MF_MOVE_ALL,
1662					MIGRATE_SYNC, MR_MEMORY_FAILURE);
1663		if (ret) {
1664			if (!list_empty(&pagelist)) {
1665				list_del(&page->lru);
1666				dec_zone_page_state(page, NR_ISOLATED_ANON +
1667						page_is_file_cache(page));
1668				putback_lru_page(page);
1669			}
1670
1671			pr_info("soft offline: %#lx: migration failed %d, type %lx\n",
1672				pfn, ret, page->flags);
1673			if (ret > 0)
1674				ret = -EIO;
1675		}
1676	} else {
1677		pr_info("soft offline: %#lx: isolation failed: %d, page count %d, type %lx\n",
1678			pfn, ret, page_count(page), page->flags);
1679	}
1680	return ret;
1681}
1682
1683static int soft_offline_in_use_page(struct page *page, int flags)
1684{
1685	int ret;
1686	struct page *hpage = compound_head(page);
1687
1688	if (!PageHuge(page) && PageTransHuge(hpage)) {
1689		lock_page(hpage);
1690		if (!PageAnon(hpage) || unlikely(split_huge_page(hpage))) {
1691			unlock_page(hpage);
1692			if (!PageAnon(hpage))
1693				pr_info("soft offline: %#lx: non anonymous thp\n", page_to_pfn(page));
1694			else
1695				pr_info("soft offline: %#lx: thp split failed\n", page_to_pfn(page));
1696			put_hwpoison_page(hpage);
1697			return -EBUSY;
1698		}
1699		unlock_page(hpage);
1700		get_hwpoison_page(page);
1701		put_hwpoison_page(hpage);
1702	}
1703
1704	if (PageHuge(page))
1705		ret = soft_offline_huge_page(page, flags);
1706	else
1707		ret = __soft_offline_page(page, flags);
1708
1709	return ret;
1710}
1711
1712static void soft_offline_free_page(struct page *page)
1713{
1714	if (PageHuge(page)) {
1715		struct page *hpage = compound_head(page);
1716
1717		set_page_hwpoison_huge_page(hpage);
1718		if (!dequeue_hwpoisoned_huge_page(hpage))
1719			num_poisoned_pages_add(1 << compound_order(hpage));
1720	} else {
1721		if (!TestSetPageHWPoison(page))
1722			num_poisoned_pages_inc();
1723	}
1724}
1725
1726/**
1727 * soft_offline_page - Soft offline a page.
1728 * @page: page to offline
1729 * @flags: flags. Same as memory_failure().
1730 *
1731 * Returns 0 on success, otherwise negated errno.
1732 *
1733 * Soft offline a page, by migration or invalidation,
1734 * without killing anything. This is for the case when
1735 * a page is not corrupted yet (so it's still valid to access),
1736 * but has had a number of corrected errors and is better taken
1737 * out.
1738 *
1739 * The actual policy on when to do that is maintained by
1740 * user space.
1741 *
1742 * This should never impact any application or cause data loss,
1743 * however it might take some time.
1744 *
1745 * This is not a 100% solution for all memory, but tries to be
1746 * ``good enough'' for the majority of memory.
1747 */
1748int soft_offline_page(struct page *page, int flags)
1749{
1750	int ret;
1751	unsigned long pfn = page_to_pfn(page);
1752
1753	if (PageHWPoison(page)) {
1754		pr_info("soft offline: %#lx page already poisoned\n", pfn);
1755		if (flags & MF_COUNT_INCREASED)
1756			put_hwpoison_page(page);
1757		return -EBUSY;
1758	}
1759
1760	get_online_mems();
1761	ret = get_any_page(page, pfn, flags);
1762	put_online_mems();
1763
1764	if (ret > 0)
1765		ret = soft_offline_in_use_page(page, flags);
1766	else if (ret == 0)
1767		soft_offline_free_page(page);
1768
1769	return ret;
1770}