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/signal.h>
  44#include <linux/sched/task.h>
  45#include <linux/ksm.h>
  46#include <linux/rmap.h>
  47#include <linux/export.h>
  48#include <linux/pagemap.h>
  49#include <linux/swap.h>
  50#include <linux/backing-dev.h>
  51#include <linux/migrate.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,
 182			unsigned long pfn, struct page *page, int flags)
 183{
 184	short addr_lsb;
 185	int ret;
 186
 187	pr_err("Memory failure: %#lx: Killing %s:%d due to hardware memory corruption\n",
 
 188		pfn, t->comm, t->pid);
 189	addr_lsb = compound_order(compound_head(page)) + PAGE_SHIFT;
 
 
 
 
 
 
 190
 191	if ((flags & MF_ACTION_REQUIRED) && t->mm == current->mm) {
 192		ret = force_sig_mceerr(BUS_MCEERR_AR, (void __user *)addr,
 193				       addr_lsb, current);
 194	} else {
 195		/*
 196		 * Don't use force here, it's convenient if the signal
 197		 * can be temporarily blocked.
 198		 * This could cause a loop when the user sets SIGBUS
 199		 * to SIG_IGN, but hopefully no one will do that?
 200		 */
 201		ret = send_sig_mceerr(BUS_MCEERR_AO, (void __user *)addr,
 202				      addr_lsb, t);  /* synchronous? */
 203	}
 204	if (ret < 0)
 205		pr_info("Memory failure: Error sending signal to %s:%d: %d\n",
 206			t->comm, t->pid, ret);
 207	return ret;
 208}
 209
 210/*
 211 * When a unknown page type is encountered drain as many buffers as possible
 212 * in the hope to turn the page into a LRU or free page, which we can handle.
 213 */
 214void shake_page(struct page *p, int access)
 215{
 216	if (PageHuge(p))
 217		return;
 218
 219	if (!PageSlab(p)) {
 220		lru_add_drain_all();
 221		if (PageLRU(p))
 222			return;
 223		drain_all_pages(page_zone(p));
 224		if (PageLRU(p) || is_free_buddy_page(p))
 225			return;
 226	}
 227
 228	/*
 229	 * Only call shrink_node_slabs here (which would also shrink
 230	 * other caches) if access is not potentially fatal.
 231	 */
 232	if (access)
 233		drop_slab_node(page_to_nid(p));
 
 
 
 
 
 
 
 
 
 
 
 
 234}
 235EXPORT_SYMBOL_GPL(shake_page);
 236
 237/*
 238 * Kill all processes that have a poisoned page mapped and then isolate
 239 * the page.
 240 *
 241 * General strategy:
 242 * Find all processes having the page mapped and kill them.
 243 * But we keep a page reference around so that the page is not
 244 * actually freed yet.
 245 * Then stash the page away
 246 *
 247 * There's no convenient way to get back to mapped processes
 248 * from the VMAs. So do a brute-force search over all
 249 * running processes.
 250 *
 251 * Remember that machine checks are not common (or rather
 252 * if they are common you have other problems), so this shouldn't
 253 * be a performance issue.
 254 *
 255 * Also there are some races possible while we get from the
 256 * error detection to actually handle it.
 257 */
 258
 259struct to_kill {
 260	struct list_head nd;
 261	struct task_struct *tsk;
 262	unsigned long addr;
 263	char addr_valid;
 264};
 265
 266/*
 267 * Failure handling: if we can't find or can't kill a process there's
 268 * not much we can do.	We just print a message and ignore otherwise.
 269 */
 270
 271/*
 272 * Schedule a process for later kill.
 273 * Uses GFP_ATOMIC allocations to avoid potential recursions in the VM.
 274 * TBD would GFP_NOIO be enough?
 275 */
 276static void add_to_kill(struct task_struct *tsk, struct page *p,
 277		       struct vm_area_struct *vma,
 278		       struct list_head *to_kill,
 279		       struct to_kill **tkc)
 280{
 281	struct to_kill *tk;
 282
 283	if (*tkc) {
 284		tk = *tkc;
 285		*tkc = NULL;
 286	} else {
 287		tk = kmalloc(sizeof(struct to_kill), GFP_ATOMIC);
 288		if (!tk) {
 289			pr_err("Memory failure: Out of memory while machine check handling\n");
 
 290			return;
 291		}
 292	}
 293	tk->addr = page_address_in_vma(p, vma);
 294	tk->addr_valid = 1;
 295
 296	/*
 297	 * In theory we don't have to kill when the page was
 298	 * munmaped. But it could be also a mremap. Since that's
 299	 * likely very rare kill anyways just out of paranoia, but use
 300	 * a SIGKILL because the error is not contained anymore.
 301	 */
 302	if (tk->addr == -EFAULT) {
 303		pr_info("Memory failure: Unable to find user space address %lx in %s\n",
 304			page_to_pfn(p), tsk->comm);
 305		tk->addr_valid = 0;
 306	}
 307	get_task_struct(tsk);
 308	tk->tsk = tsk;
 309	list_add_tail(&tk->nd, to_kill);
 310}
 311
 312/*
 313 * Kill the processes that have been collected earlier.
 314 *
 315 * Only do anything when DOIT is set, otherwise just free the list
 316 * (this is used for clean pages which do not need killing)
 317 * Also when FAIL is set do a force kill because something went
 318 * wrong earlier.
 319 */
 320static void kill_procs(struct list_head *to_kill, int forcekill,
 321			  bool fail, struct page *page, unsigned long pfn,
 322			  int flags)
 323{
 324	struct to_kill *tk, *next;
 325
 326	list_for_each_entry_safe (tk, next, to_kill, nd) {
 327		if (forcekill) {
 328			/*
 329			 * In case something went wrong with munmapping
 330			 * make sure the process doesn't catch the
 331			 * signal and then access the memory. Just kill it.
 332			 */
 333			if (fail || tk->addr_valid == 0) {
 334				pr_err("Memory failure: %#lx: forcibly killing %s:%d because of failure to unmap corrupted page\n",
 335				       pfn, tk->tsk->comm, tk->tsk->pid);
 
 336				force_sig(SIGKILL, tk->tsk);
 337			}
 338
 339			/*
 340			 * In theory the process could have mapped
 341			 * something else on the address in-between. We could
 342			 * check for that, but we need to tell the
 343			 * process anyways.
 344			 */
 345			else if (kill_proc(tk->tsk, tk->addr,
 346					      pfn, page, flags) < 0)
 347				pr_err("Memory failure: %#lx: Cannot send advisory machine check signal to %s:%d\n",
 348				       pfn, tk->tsk->comm, tk->tsk->pid);
 
 349		}
 350		put_task_struct(tk->tsk);
 351		kfree(tk);
 352	}
 353}
 354
 355/*
 356 * Find a dedicated thread which is supposed to handle SIGBUS(BUS_MCEERR_AO)
 357 * on behalf of the thread group. Return task_struct of the (first found)
 358 * dedicated thread if found, and return NULL otherwise.
 359 *
 360 * We already hold read_lock(&tasklist_lock) in the caller, so we don't
 361 * have to call rcu_read_lock/unlock() in this function.
 362 */
 363static struct task_struct *find_early_kill_thread(struct task_struct *tsk)
 364{
 365	struct task_struct *t;
 366
 367	for_each_thread(tsk, t)
 368		if ((t->flags & PF_MCE_PROCESS) && (t->flags & PF_MCE_EARLY))
 369			return t;
 370	return NULL;
 371}
 372
 373/*
 374 * Determine whether a given process is "early kill" process which expects
 375 * to be signaled when some page under the process is hwpoisoned.
 376 * Return task_struct of the dedicated thread (main thread unless explicitly
 377 * specified) if the process is "early kill," and otherwise returns NULL.
 378 */
 379static struct task_struct *task_early_kill(struct task_struct *tsk,
 380					   int force_early)
 381{
 382	struct task_struct *t;
 383	if (!tsk->mm)
 384		return NULL;
 385	if (force_early)
 386		return tsk;
 387	t = find_early_kill_thread(tsk);
 388	if (t)
 389		return t;
 390	if (sysctl_memory_failure_early_kill)
 391		return tsk;
 392	return NULL;
 393}
 394
 395/*
 396 * Collect processes when the error hit an anonymous page.
 397 */
 398static void collect_procs_anon(struct page *page, struct list_head *to_kill,
 399			      struct to_kill **tkc, int force_early)
 400{
 401	struct vm_area_struct *vma;
 402	struct task_struct *tsk;
 403	struct anon_vma *av;
 404	pgoff_t pgoff;
 405
 406	av = page_lock_anon_vma_read(page);
 407	if (av == NULL)	/* Not actually mapped anymore */
 408		return;
 409
 410	pgoff = page_to_pgoff(page);
 411	read_lock(&tasklist_lock);
 412	for_each_process (tsk) {
 413		struct anon_vma_chain *vmac;
 414		struct task_struct *t = task_early_kill(tsk, force_early);
 415
 416		if (!t)
 417			continue;
 418		anon_vma_interval_tree_foreach(vmac, &av->rb_root,
 419					       pgoff, pgoff) {
 420			vma = vmac->vma;
 421			if (!page_mapped_in_vma(page, vma))
 422				continue;
 423			if (vma->vm_mm == t->mm)
 424				add_to_kill(t, page, vma, to_kill, tkc);
 425		}
 426	}
 427	read_unlock(&tasklist_lock);
 428	page_unlock_anon_vma_read(av);
 429}
 430
 431/*
 432 * Collect processes when the error hit a file mapped page.
 433 */
 434static void collect_procs_file(struct page *page, struct list_head *to_kill,
 435			      struct to_kill **tkc, int force_early)
 436{
 437	struct vm_area_struct *vma;
 438	struct task_struct *tsk;
 439	struct address_space *mapping = page->mapping;
 440
 441	i_mmap_lock_read(mapping);
 442	read_lock(&tasklist_lock);
 443	for_each_process(tsk) {
 444		pgoff_t pgoff = page_to_pgoff(page);
 445		struct task_struct *t = task_early_kill(tsk, force_early);
 446
 447		if (!t)
 448			continue;
 
 449		vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff,
 450				      pgoff) {
 451			/*
 452			 * Send early kill signal to tasks where a vma covers
 453			 * the page but the corrupted page is not necessarily
 454			 * mapped it in its pte.
 455			 * Assume applications who requested early kill want
 456			 * to be informed of all such data corruptions.
 457			 */
 458			if (vma->vm_mm == t->mm)
 459				add_to_kill(t, page, vma, to_kill, tkc);
 460		}
 461	}
 462	read_unlock(&tasklist_lock);
 463	i_mmap_unlock_read(mapping);
 464}
 465
 466/*
 467 * Collect the processes who have the corrupted page mapped to kill.
 468 * This is done in two steps for locking reasons.
 469 * First preallocate one tokill structure outside the spin locks,
 470 * so that we can kill at least one process reasonably reliable.
 471 */
 472static void collect_procs(struct page *page, struct list_head *tokill,
 473				int force_early)
 474{
 475	struct to_kill *tk;
 476
 477	if (!page->mapping)
 478		return;
 479
 480	tk = kmalloc(sizeof(struct to_kill), GFP_NOIO);
 481	if (!tk)
 482		return;
 483	if (PageAnon(page))
 484		collect_procs_anon(page, tokill, &tk, force_early);
 485	else
 486		collect_procs_file(page, tokill, &tk, force_early);
 487	kfree(tk);
 488}
 489
 490static const char *action_name[] = {
 491	[MF_IGNORED] = "Ignored",
 492	[MF_FAILED] = "Failed",
 493	[MF_DELAYED] = "Delayed",
 494	[MF_RECOVERED] = "Recovered",
 
 
 
 
 495};
 496
 497static const char * const action_page_types[] = {
 498	[MF_MSG_KERNEL]			= "reserved kernel page",
 499	[MF_MSG_KERNEL_HIGH_ORDER]	= "high-order kernel page",
 500	[MF_MSG_SLAB]			= "kernel slab page",
 501	[MF_MSG_DIFFERENT_COMPOUND]	= "different compound page after locking",
 502	[MF_MSG_POISONED_HUGE]		= "huge page already hardware poisoned",
 503	[MF_MSG_HUGE]			= "huge page",
 504	[MF_MSG_FREE_HUGE]		= "free huge page",
 505	[MF_MSG_NON_PMD_HUGE]		= "non-pmd-sized huge page",
 506	[MF_MSG_UNMAP_FAILED]		= "unmapping failed page",
 507	[MF_MSG_DIRTY_SWAPCACHE]	= "dirty swapcache page",
 508	[MF_MSG_CLEAN_SWAPCACHE]	= "clean swapcache page",
 509	[MF_MSG_DIRTY_MLOCKED_LRU]	= "dirty mlocked LRU page",
 510	[MF_MSG_CLEAN_MLOCKED_LRU]	= "clean mlocked LRU page",
 511	[MF_MSG_DIRTY_UNEVICTABLE_LRU]	= "dirty unevictable LRU page",
 512	[MF_MSG_CLEAN_UNEVICTABLE_LRU]	= "clean unevictable LRU page",
 513	[MF_MSG_DIRTY_LRU]		= "dirty LRU page",
 514	[MF_MSG_CLEAN_LRU]		= "clean LRU page",
 515	[MF_MSG_TRUNCATED_LRU]		= "already truncated LRU page",
 516	[MF_MSG_BUDDY]			= "free buddy page",
 517	[MF_MSG_BUDDY_2ND]		= "free buddy page (2nd try)",
 518	[MF_MSG_UNKNOWN]		= "unknown page",
 519};
 520
 521/*
 522 * XXX: It is possible that a page is isolated from LRU cache,
 523 * and then kept in swap cache or failed to remove from page cache.
 524 * The page count will stop it from being freed by unpoison.
 525 * Stress tests should be aware of this memory leak problem.
 526 */
 527static int delete_from_lru_cache(struct page *p)
 528{
 529	if (!isolate_lru_page(p)) {
 530		/*
 531		 * Clear sensible page flags, so that the buddy system won't
 532		 * complain when the page is unpoison-and-freed.
 533		 */
 534		ClearPageActive(p);
 535		ClearPageUnevictable(p);
 536
 537		/*
 538		 * Poisoned page might never drop its ref count to 0 so we have
 539		 * to uncharge it manually from its memcg.
 540		 */
 541		mem_cgroup_uncharge(p);
 542
 543		/*
 544		 * drop the page count elevated by isolate_lru_page()
 545		 */
 546		put_page(p);
 547		return 0;
 548	}
 549	return -EIO;
 550}
 551
 552static int truncate_error_page(struct page *p, unsigned long pfn,
 553				struct address_space *mapping)
 554{
 555	int ret = MF_FAILED;
 556
 557	if (mapping->a_ops->error_remove_page) {
 558		int err = mapping->a_ops->error_remove_page(mapping, p);
 559
 560		if (err != 0) {
 561			pr_info("Memory failure: %#lx: Failed to punch page: %d\n",
 562				pfn, err);
 563		} else if (page_has_private(p) &&
 564			   !try_to_release_page(p, GFP_NOIO)) {
 565			pr_info("Memory failure: %#lx: failed to release buffers\n",
 566				pfn);
 567		} else {
 568			ret = MF_RECOVERED;
 569		}
 570	} else {
 571		/*
 572		 * If the file system doesn't support it just invalidate
 573		 * This fails on dirty or anything with private pages
 574		 */
 575		if (invalidate_inode_page(p))
 576			ret = MF_RECOVERED;
 577		else
 578			pr_info("Memory failure: %#lx: Failed to invalidate\n",
 579				pfn);
 580	}
 581
 582	return ret;
 583}
 584
 585/*
 586 * Error hit kernel page.
 587 * Do nothing, try to be lucky and not touch this instead. For a few cases we
 588 * could be more sophisticated.
 589 */
 590static int me_kernel(struct page *p, unsigned long pfn)
 591{
 592	return MF_IGNORED;
 593}
 594
 595/*
 596 * Page in unknown state. Do nothing.
 597 */
 598static int me_unknown(struct page *p, unsigned long pfn)
 599{
 600	pr_err("Memory failure: %#lx: Unknown page state\n", pfn);
 601	return MF_FAILED;
 602}
 603
 604/*
 605 * Clean (or cleaned) page cache page.
 606 */
 607static int me_pagecache_clean(struct page *p, unsigned long pfn)
 608{
 
 
 609	struct address_space *mapping;
 610
 611	delete_from_lru_cache(p);
 612
 613	/*
 614	 * For anonymous pages we're done the only reference left
 615	 * should be the one m_f() holds.
 616	 */
 617	if (PageAnon(p))
 618		return MF_RECOVERED;
 619
 620	/*
 621	 * Now truncate the page in the page cache. This is really
 622	 * more like a "temporary hole punch"
 623	 * Don't do this for block devices when someone else
 624	 * has a reference, because it could be file system metadata
 625	 * and that's not safe to truncate.
 626	 */
 627	mapping = page_mapping(p);
 628	if (!mapping) {
 629		/*
 630		 * Page has been teared down in the meanwhile
 631		 */
 632		return MF_FAILED;
 633	}
 634
 635	/*
 636	 * Truncation is a bit tricky. Enable it per file system for now.
 637	 *
 638	 * Open: to take i_mutex or not for this? Right now we don't.
 639	 */
 640	return truncate_error_page(p, pfn, mapping);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 641}
 642
 643/*
 644 * Dirty pagecache page
 645 * Issues: when the error hit a hole page the error is not properly
 646 * propagated.
 647 */
 648static int me_pagecache_dirty(struct page *p, unsigned long pfn)
 649{
 650	struct address_space *mapping = page_mapping(p);
 651
 652	SetPageError(p);
 653	/* TBD: print more information about the file. */
 654	if (mapping) {
 655		/*
 656		 * IO error will be reported by write(), fsync(), etc.
 657		 * who check the mapping.
 658		 * This way the application knows that something went
 659		 * wrong with its dirty file data.
 660		 *
 661		 * There's one open issue:
 662		 *
 663		 * The EIO will be only reported on the next IO
 664		 * operation and then cleared through the IO map.
 665		 * Normally Linux has two mechanisms to pass IO error
 666		 * first through the AS_EIO flag in the address space
 667		 * and then through the PageError flag in the page.
 668		 * Since we drop pages on memory failure handling the
 669		 * only mechanism open to use is through AS_AIO.
 670		 *
 671		 * This has the disadvantage that it gets cleared on
 672		 * the first operation that returns an error, while
 673		 * the PageError bit is more sticky and only cleared
 674		 * when the page is reread or dropped.  If an
 675		 * application assumes it will always get error on
 676		 * fsync, but does other operations on the fd before
 677		 * and the page is dropped between then the error
 678		 * will not be properly reported.
 679		 *
 680		 * This can already happen even without hwpoisoned
 681		 * pages: first on metadata IO errors (which only
 682		 * report through AS_EIO) or when the page is dropped
 683		 * at the wrong time.
 684		 *
 685		 * So right now we assume that the application DTRT on
 686		 * the first EIO, but we're not worse than other parts
 687		 * of the kernel.
 688		 */
 689		mapping_set_error(mapping, -EIO);
 690	}
 691
 692	return me_pagecache_clean(p, pfn);
 693}
 694
 695/*
 696 * Clean and dirty swap cache.
 697 *
 698 * Dirty swap cache page is tricky to handle. The page could live both in page
 699 * cache and swap cache(ie. page is freshly swapped in). So it could be
 700 * referenced concurrently by 2 types of PTEs:
 701 * normal PTEs and swap PTEs. We try to handle them consistently by calling
 702 * try_to_unmap(TTU_IGNORE_HWPOISON) to convert the normal PTEs to swap PTEs,
 703 * and then
 704 *      - clear dirty bit to prevent IO
 705 *      - remove from LRU
 706 *      - but keep in the swap cache, so that when we return to it on
 707 *        a later page fault, we know the application is accessing
 708 *        corrupted data and shall be killed (we installed simple
 709 *        interception code in do_swap_page to catch it).
 710 *
 711 * Clean swap cache pages can be directly isolated. A later page fault will
 712 * bring in the known good data from disk.
 713 */
 714static int me_swapcache_dirty(struct page *p, unsigned long pfn)
 715{
 716	ClearPageDirty(p);
 717	/* Trigger EIO in shmem: */
 718	ClearPageUptodate(p);
 719
 720	if (!delete_from_lru_cache(p))
 721		return MF_DELAYED;
 722	else
 723		return MF_FAILED;
 724}
 725
 726static int me_swapcache_clean(struct page *p, unsigned long pfn)
 727{
 728	delete_from_swap_cache(p);
 729
 730	if (!delete_from_lru_cache(p))
 731		return MF_RECOVERED;
 732	else
 733		return MF_FAILED;
 734}
 735
 736/*
 737 * Huge pages. Needs work.
 738 * Issues:
 739 * - Error on hugepage is contained in hugepage unit (not in raw page unit.)
 740 *   To narrow down kill region to one page, we need to break up pmd.
 741 */
 742static int me_huge_page(struct page *p, unsigned long pfn)
 743{
 744	int res = 0;
 745	struct page *hpage = compound_head(p);
 746	struct address_space *mapping;
 747
 748	if (!PageHuge(hpage))
 749		return MF_DELAYED;
 750
 751	mapping = page_mapping(hpage);
 752	if (mapping) {
 753		res = truncate_error_page(hpage, pfn, mapping);
 754	} else {
 755		unlock_page(hpage);
 756		/*
 757		 * migration entry prevents later access on error anonymous
 758		 * hugepage, so we can free and dissolve it into buddy to
 759		 * save healthy subpages.
 760		 */
 761		if (PageAnon(hpage))
 762			put_page(hpage);
 763		dissolve_free_huge_page(p);
 764		res = MF_RECOVERED;
 765		lock_page(hpage);
 766	}
 767
 768	return res;
 769}
 770
 771/*
 772 * Various page states we can handle.
 773 *
 774 * A page state is defined by its current page->flags bits.
 775 * The table matches them in order and calls the right handler.
 776 *
 777 * This is quite tricky because we can access page at any time
 778 * in its live cycle, so all accesses have to be extremely careful.
 779 *
 780 * This is not complete. More states could be added.
 781 * For any missing state don't attempt recovery.
 782 */
 783
 784#define dirty		(1UL << PG_dirty)
 785#define sc		((1UL << PG_swapcache) | (1UL << PG_swapbacked))
 786#define unevict		(1UL << PG_unevictable)
 787#define mlock		(1UL << PG_mlocked)
 788#define writeback	(1UL << PG_writeback)
 789#define lru		(1UL << PG_lru)
 
 790#define head		(1UL << PG_head)
 
 
 791#define slab		(1UL << PG_slab)
 792#define reserved	(1UL << PG_reserved)
 793
 794static struct page_state {
 795	unsigned long mask;
 796	unsigned long res;
 797	enum mf_action_page_type type;
 798	int (*action)(struct page *p, unsigned long pfn);
 799} error_states[] = {
 800	{ reserved,	reserved,	MF_MSG_KERNEL,	me_kernel },
 801	/*
 802	 * free pages are specially detected outside this table:
 803	 * PG_buddy pages only make a small fraction of all free pages.
 804	 */
 805
 806	/*
 807	 * Could in theory check if slab page is free or if we can drop
 808	 * currently unused objects without touching them. But just
 809	 * treat it as standard kernel for now.
 810	 */
 811	{ slab,		slab,		MF_MSG_SLAB,	me_kernel },
 812
 813	{ head,		head,		MF_MSG_HUGE,		me_huge_page },
 
 
 
 
 
 814
 815	{ sc|dirty,	sc|dirty,	MF_MSG_DIRTY_SWAPCACHE,	me_swapcache_dirty },
 816	{ sc|dirty,	sc,		MF_MSG_CLEAN_SWAPCACHE,	me_swapcache_clean },
 817
 818	{ mlock|dirty,	mlock|dirty,	MF_MSG_DIRTY_MLOCKED_LRU,	me_pagecache_dirty },
 819	{ mlock|dirty,	mlock,		MF_MSG_CLEAN_MLOCKED_LRU,	me_pagecache_clean },
 820
 821	{ unevict|dirty, unevict|dirty,	MF_MSG_DIRTY_UNEVICTABLE_LRU,	me_pagecache_dirty },
 822	{ unevict|dirty, unevict,	MF_MSG_CLEAN_UNEVICTABLE_LRU,	me_pagecache_clean },
 823
 824	{ lru|dirty,	lru|dirty,	MF_MSG_DIRTY_LRU,	me_pagecache_dirty },
 825	{ lru|dirty,	lru,		MF_MSG_CLEAN_LRU,	me_pagecache_clean },
 826
 827	/*
 828	 * Catchall entry: must be at end.
 829	 */
 830	{ 0,		0,		MF_MSG_UNKNOWN,	me_unknown },
 831};
 832
 833#undef dirty
 834#undef sc
 835#undef unevict
 836#undef mlock
 837#undef writeback
 838#undef lru
 
 839#undef head
 
 
 840#undef slab
 841#undef reserved
 842
 843/*
 844 * "Dirty/Clean" indication is not 100% accurate due to the possibility of
 845 * setting PG_dirty outside page lock. See also comment above set_page_dirty().
 846 */
 847static void action_result(unsigned long pfn, enum mf_action_page_type type,
 848			  enum mf_result result)
 849{
 850	trace_memory_failure_event(pfn, type, result);
 851
 852	pr_err("Memory failure: %#lx: recovery action for %s: %s\n",
 853		pfn, action_page_types[type], action_name[result]);
 854}
 855
 856static int page_action(struct page_state *ps, struct page *p,
 857			unsigned long pfn)
 858{
 859	int result;
 860	int count;
 861
 862	result = ps->action(p, pfn);
 
 863
 864	count = page_count(p) - 1;
 865	if (ps->action == me_swapcache_dirty && result == MF_DELAYED)
 866		count--;
 867	if (count > 0) {
 868		pr_err("Memory failure: %#lx: %s still referenced by %d users\n",
 869		       pfn, action_page_types[ps->type], count);
 870		result = MF_FAILED;
 
 871	}
 872	action_result(pfn, ps->type, result);
 873
 874	/* Could do more checks here if page looks ok */
 875	/*
 876	 * Could adjust zone counters here to correct for the missing page.
 877	 */
 878
 879	return (result == MF_RECOVERED || result == MF_DELAYED) ? 0 : -EBUSY;
 880}
 881
 882/**
 883 * get_hwpoison_page() - Get refcount for memory error handling:
 884 * @page:	raw error page (hit by memory error)
 885 *
 886 * Return: return 0 if failed to grab the refcount, otherwise true (some
 887 * non-zero value.)
 888 */
 889int get_hwpoison_page(struct page *page)
 890{
 891	struct page *head = compound_head(page);
 892
 893	if (!PageHuge(head) && PageTransHuge(head)) {
 894		/*
 895		 * Non anonymous thp exists only in allocation/free time. We
 896		 * can't handle such a case correctly, so let's give it up.
 897		 * This should be better than triggering BUG_ON when kernel
 898		 * tries to touch the "partially handled" page.
 899		 */
 900		if (!PageAnon(head)) {
 901			pr_err("Memory failure: %#lx: non anonymous thp\n",
 902				page_to_pfn(page));
 903			return 0;
 904		}
 905	}
 906
 907	if (get_page_unless_zero(head)) {
 908		if (head == compound_head(page))
 909			return 1;
 910
 911		pr_info("Memory failure: %#lx cannot catch tail\n",
 912			page_to_pfn(page));
 913		put_page(head);
 914	}
 915
 916	return 0;
 917}
 918EXPORT_SYMBOL_GPL(get_hwpoison_page);
 919
 920/*
 921 * Do all that is necessary to remove user space mappings. Unmap
 922 * the pages and send SIGBUS to the processes if the data was dirty.
 923 */
 924static bool hwpoison_user_mappings(struct page *p, unsigned long pfn,
 925				  int flags, struct page **hpagep)
 926{
 927	enum ttu_flags ttu = TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS;
 928	struct address_space *mapping;
 929	LIST_HEAD(tokill);
 930	bool unmap_success;
 931	int kill = 1, forcekill;
 932	struct page *hpage = *hpagep;
 933	bool mlocked = PageMlocked(hpage);
 934
 935	/*
 936	 * Here we are interested only in user-mapped pages, so skip any
 937	 * other types of pages.
 938	 */
 939	if (PageReserved(p) || PageSlab(p))
 940		return true;
 941	if (!(PageLRU(hpage) || PageHuge(p)))
 942		return true;
 943
 944	/*
 945	 * This check implies we don't kill processes if their pages
 946	 * are in the swap cache early. Those are always late kills.
 947	 */
 948	if (!page_mapped(hpage))
 949		return true;
 950
 951	if (PageKsm(p)) {
 952		pr_err("Memory failure: %#lx: can't handle KSM pages.\n", pfn);
 953		return false;
 954	}
 955
 956	if (PageSwapCache(p)) {
 957		pr_err("Memory failure: %#lx: keeping poisoned page in swap cache\n",
 958			pfn);
 959		ttu |= TTU_IGNORE_HWPOISON;
 960	}
 961
 962	/*
 963	 * Propagate the dirty bit from PTEs to struct page first, because we
 964	 * need this to decide if we should kill or just drop the page.
 965	 * XXX: the dirty test could be racy: set_page_dirty() may not always
 966	 * be called inside page lock (it's recommended but not enforced).
 967	 */
 968	mapping = page_mapping(hpage);
 969	if (!(flags & MF_MUST_KILL) && !PageDirty(hpage) && mapping &&
 970	    mapping_cap_writeback_dirty(mapping)) {
 971		if (page_mkclean(hpage)) {
 972			SetPageDirty(hpage);
 973		} else {
 974			kill = 0;
 975			ttu |= TTU_IGNORE_HWPOISON;
 976			pr_info("Memory failure: %#lx: corrupted page was clean: dropped without side effects\n",
 
 977				pfn);
 978		}
 979	}
 980
 981	/*
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 982	 * First collect all the processes that have the page
 983	 * mapped in dirty form.  This has to be done before try_to_unmap,
 984	 * because ttu takes the rmap data structures down.
 985	 *
 986	 * Error handling: We ignore errors here because
 987	 * there's nothing that can be done.
 988	 */
 989	if (kill)
 990		collect_procs(hpage, &tokill, flags & MF_ACTION_REQUIRED);
 991
 992	unmap_success = try_to_unmap(hpage, ttu);
 993	if (!unmap_success)
 994		pr_err("Memory failure: %#lx: failed to unmap page (mapcount=%d)\n",
 995		       pfn, page_mapcount(hpage));
 996
 997	/*
 998	 * try_to_unmap() might put mlocked page in lru cache, so call
 999	 * shake_page() again to ensure that it's flushed.
1000	 */
1001	if (mlocked)
1002		shake_page(hpage, 0);
1003
1004	/*
1005	 * Now that the dirty bit has been propagated to the
1006	 * struct page and all unmaps done we can decide if
1007	 * killing is needed or not.  Only kill when the page
1008	 * was dirty or the process is not restartable,
1009	 * otherwise the tokill list is merely
1010	 * freed.  When there was a problem unmapping earlier
1011	 * use a more force-full uncatchable kill to prevent
1012	 * any accesses to the poisoned memory.
1013	 */
1014	forcekill = PageDirty(hpage) || (flags & MF_MUST_KILL);
1015	kill_procs(&tokill, forcekill, !unmap_success, p, pfn, flags);
 
1016
1017	return unmap_success;
1018}
1019
1020static int identify_page_state(unsigned long pfn, struct page *p,
1021				unsigned long page_flags)
1022{
1023	struct page_state *ps;
1024
1025	/*
1026	 * The first check uses the current page flags which may not have any
1027	 * relevant information. The second check with the saved page flags is
1028	 * carried out only if the first check can't determine the page status.
1029	 */
1030	for (ps = error_states;; ps++)
1031		if ((p->flags & ps->mask) == ps->res)
1032			break;
1033
1034	page_flags |= (p->flags & (1UL << PG_dirty));
1035
1036	if (!ps->mask)
1037		for (ps = error_states;; ps++)
1038			if ((page_flags & ps->mask) == ps->res)
1039				break;
1040	return page_action(ps, p, pfn);
1041}
1042
1043static int memory_failure_hugetlb(unsigned long pfn, int flags)
1044{
1045	struct page *p = pfn_to_page(pfn);
1046	struct page *head = compound_head(p);
1047	int res;
1048	unsigned long page_flags;
1049
1050	if (TestSetPageHWPoison(head)) {
1051		pr_err("Memory failure: %#lx: already hardware poisoned\n",
1052		       pfn);
1053		return 0;
1054	}
1055
1056	num_poisoned_pages_inc();
1057
1058	if (!(flags & MF_COUNT_INCREASED) && !get_hwpoison_page(p)) {
1059		/*
1060		 * Check "filter hit" and "race with other subpage."
1061		 */
1062		lock_page(head);
1063		if (PageHWPoison(head)) {
1064			if ((hwpoison_filter(p) && TestClearPageHWPoison(p))
1065			    || (p != head && TestSetPageHWPoison(head))) {
1066				num_poisoned_pages_dec();
1067				unlock_page(head);
1068				return 0;
1069			}
1070		}
1071		unlock_page(head);
1072		dissolve_free_huge_page(p);
1073		action_result(pfn, MF_MSG_FREE_HUGE, MF_DELAYED);
1074		return 0;
1075	}
1076
1077	lock_page(head);
1078	page_flags = head->flags;
1079
1080	if (!PageHWPoison(head)) {
1081		pr_err("Memory failure: %#lx: just unpoisoned\n", pfn);
1082		num_poisoned_pages_dec();
1083		unlock_page(head);
1084		put_hwpoison_page(head);
1085		return 0;
1086	}
1087
1088	/*
1089	 * TODO: hwpoison for pud-sized hugetlb doesn't work right now, so
1090	 * simply disable it. In order to make it work properly, we need
1091	 * make sure that:
1092	 *  - conversion of a pud that maps an error hugetlb into hwpoison
1093	 *    entry properly works, and
1094	 *  - other mm code walking over page table is aware of pud-aligned
1095	 *    hwpoison entries.
1096	 */
1097	if (huge_page_size(page_hstate(head)) > PMD_SIZE) {
1098		action_result(pfn, MF_MSG_NON_PMD_HUGE, MF_IGNORED);
1099		res = -EBUSY;
1100		goto out;
1101	}
1102
1103	if (!hwpoison_user_mappings(p, pfn, flags, &head)) {
1104		action_result(pfn, MF_MSG_UNMAP_FAILED, MF_IGNORED);
1105		res = -EBUSY;
1106		goto out;
1107	}
1108
1109	res = identify_page_state(pfn, p, page_flags);
1110out:
1111	unlock_page(head);
1112	return res;
1113}
1114
1115/**
1116 * memory_failure - Handle memory failure of a page.
1117 * @pfn: Page Number of the corrupted page
 
1118 * @flags: fine tune action taken
1119 *
1120 * This function is called by the low level machine check code
1121 * of an architecture when it detects hardware memory corruption
1122 * of a page. It tries its best to recover, which includes
1123 * dropping pages, killing processes etc.
1124 *
1125 * The function is primarily of use for corruptions that
1126 * happen outside the current execution context (e.g. when
1127 * detected by a background scrubber)
1128 *
1129 * Must run in process context (e.g. a work queue) with interrupts
1130 * enabled and no spinlocks hold.
1131 */
1132int memory_failure(unsigned long pfn, int flags)
1133{
 
1134	struct page *p;
1135	struct page *hpage;
1136	struct page *orig_head;
1137	int res;
 
1138	unsigned long page_flags;
1139
1140	if (!sysctl_memory_failure_recovery)
1141		panic("Memory failure on page %lx", pfn);
1142
1143	if (!pfn_valid(pfn)) {
1144		pr_err("Memory failure: %#lx: memory outside kernel control\n",
1145			pfn);
 
1146		return -ENXIO;
1147	}
1148
1149	p = pfn_to_page(pfn);
1150	if (PageHuge(p))
1151		return memory_failure_hugetlb(pfn, flags);
1152	if (TestSetPageHWPoison(p)) {
1153		pr_err("Memory failure: %#lx: already hardware poisoned\n",
1154			pfn);
1155		return 0;
1156	}
1157
1158	orig_head = hpage = compound_head(p);
1159	num_poisoned_pages_inc();
 
 
 
 
 
 
 
 
 
 
1160
1161	/*
1162	 * We need/can do nothing about count=0 pages.
1163	 * 1) it's a free page, and therefore in safe hand:
1164	 *    prep_new_page() will be the gate keeper.
1165	 * 2) it's part of a non-compound high order page.
 
 
 
1166	 *    Implies some kernel user: cannot stop them from
1167	 *    R/W the page; let's pray that the page has been
1168	 *    used and will be freed some time later.
1169	 * In fact it's dangerous to directly bump up page count from 0,
1170	 * that may make page_freeze_refs()/page_unfreeze_refs() mismatch.
1171	 */
1172	if (!(flags & MF_COUNT_INCREASED) && !get_hwpoison_page(p)) {
 
1173		if (is_free_buddy_page(p)) {
1174			action_result(pfn, MF_MSG_BUDDY, MF_DELAYED);
1175			return 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1176		} else {
1177			action_result(pfn, MF_MSG_KERNEL_HIGH_ORDER, MF_IGNORED);
1178			return -EBUSY;
1179		}
1180	}
1181
1182	if (PageTransHuge(hpage)) {
1183		lock_page(p);
1184		if (!PageAnon(p) || unlikely(split_huge_page(p))) {
1185			unlock_page(p);
1186			if (!PageAnon(p))
1187				pr_err("Memory failure: %#lx: non anonymous thp\n",
1188					pfn);
1189			else
1190				pr_err("Memory failure: %#lx: thp split failed\n",
1191					pfn);
1192			if (TestClearPageHWPoison(p))
1193				num_poisoned_pages_dec();
1194			put_hwpoison_page(p);
1195			return -EBUSY;
1196		}
1197		unlock_page(p);
1198		VM_BUG_ON_PAGE(!page_count(p), p);
1199		hpage = compound_head(p);
1200	}
1201
1202	/*
1203	 * We ignore non-LRU pages for good reasons.
1204	 * - PG_locked is only well defined for LRU pages and a few others
1205	 * - to avoid races with __SetPageLocked()
1206	 * - to avoid races with __SetPageSlab*() (and more non-atomic ops)
1207	 * The check (unnecessarily) ignores LRU pages being isolated and
1208	 * walked by the page reclaim code, however that's not a big loss.
1209	 */
1210	shake_page(p, 0);
1211	/* shake_page could have turned it free. */
1212	if (!PageLRU(p) && is_free_buddy_page(p)) {
1213		if (flags & MF_COUNT_INCREASED)
1214			action_result(pfn, MF_MSG_BUDDY, MF_DELAYED);
1215		else
1216			action_result(pfn, MF_MSG_BUDDY_2ND, MF_DELAYED);
1217		return 0;
 
 
 
 
 
 
 
 
 
 
1218	}
1219
1220	lock_page(p);
1221
1222	/*
1223	 * The page could have changed compound pages during the locking.
1224	 * If this happens just bail out.
 
1225	 */
1226	if (PageCompound(p) && compound_head(p) != orig_head) {
1227		action_result(pfn, MF_MSG_DIFFERENT_COMPOUND, MF_IGNORED);
1228		res = -EBUSY;
1229		goto out;
1230	}
1231
1232	/*
1233	 * We use page flags to determine what action should be taken, but
1234	 * the flags can be modified by the error containment action.  One
1235	 * example is an mlocked page, where PG_mlocked is cleared by
1236	 * page_remove_rmap() in try_to_unmap_one(). So to determine page status
1237	 * correctly, we save a copy of the page flags at this time.
1238	 */
1239	if (PageHuge(p))
1240		page_flags = hpage->flags;
1241	else
1242		page_flags = p->flags;
1243
1244	/*
1245	 * unpoison always clear PG_hwpoison inside page lock
1246	 */
1247	if (!PageHWPoison(p)) {
1248		pr_err("Memory failure: %#lx: just unpoisoned\n", pfn);
1249		num_poisoned_pages_dec();
1250		unlock_page(p);
1251		put_hwpoison_page(p);
1252		return 0;
1253	}
1254	if (hwpoison_filter(p)) {
1255		if (TestClearPageHWPoison(p))
1256			num_poisoned_pages_dec();
1257		unlock_page(p);
1258		put_hwpoison_page(p);
1259		return 0;
1260	}
1261
1262	if (!PageTransTail(p) && !PageLRU(p))
1263		goto identify_page_state;
1264
1265	/*
1266	 * It's very difficult to mess with pages currently under IO
1267	 * and in many cases impossible, so we just avoid it here.
 
 
 
 
 
 
 
 
 
 
 
 
 
1268	 */
 
 
 
1269	wait_on_page_writeback(p);
1270
1271	/*
1272	 * Now take care of user space mappings.
1273	 * Abort on fail: __delete_from_page_cache() assumes unmapped page.
1274	 *
1275	 * When the raw error page is thp tail page, hpage points to the raw
1276	 * page after thp split.
1277	 */
1278	if (!hwpoison_user_mappings(p, pfn, flags, &hpage)) {
1279		action_result(pfn, MF_MSG_UNMAP_FAILED, MF_IGNORED);
 
1280		res = -EBUSY;
1281		goto out;
1282	}
1283
1284	/*
1285	 * Torn down by someone else?
1286	 */
1287	if (PageLRU(p) && !PageSwapCache(p) && p->mapping == NULL) {
1288		action_result(pfn, MF_MSG_TRUNCATED_LRU, MF_IGNORED);
1289		res = -EBUSY;
1290		goto out;
1291	}
1292
1293identify_page_state:
1294	res = identify_page_state(pfn, p, page_flags);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1295out:
1296	unlock_page(p);
1297	return res;
1298}
1299EXPORT_SYMBOL_GPL(memory_failure);
1300
1301#define MEMORY_FAILURE_FIFO_ORDER	4
1302#define MEMORY_FAILURE_FIFO_SIZE	(1 << MEMORY_FAILURE_FIFO_ORDER)
1303
1304struct memory_failure_entry {
1305	unsigned long pfn;
 
1306	int flags;
1307};
1308
1309struct memory_failure_cpu {
1310	DECLARE_KFIFO(fifo, struct memory_failure_entry,
1311		      MEMORY_FAILURE_FIFO_SIZE);
1312	spinlock_t lock;
1313	struct work_struct work;
1314};
1315
1316static DEFINE_PER_CPU(struct memory_failure_cpu, memory_failure_cpu);
1317
1318/**
1319 * memory_failure_queue - Schedule handling memory failure of a page.
1320 * @pfn: Page Number of the corrupted page
 
1321 * @flags: Flags for memory failure handling
1322 *
1323 * This function is called by the low level hardware error handler
1324 * when it detects hardware memory corruption of a page. It schedules
1325 * the recovering of error page, including dropping pages, killing
1326 * processes etc.
1327 *
1328 * The function is primarily of use for corruptions that
1329 * happen outside the current execution context (e.g. when
1330 * detected by a background scrubber)
1331 *
1332 * Can run in IRQ context.
1333 */
1334void memory_failure_queue(unsigned long pfn, int flags)
1335{
1336	struct memory_failure_cpu *mf_cpu;
1337	unsigned long proc_flags;
1338	struct memory_failure_entry entry = {
1339		.pfn =		pfn,
 
1340		.flags =	flags,
1341	};
1342
1343	mf_cpu = &get_cpu_var(memory_failure_cpu);
1344	spin_lock_irqsave(&mf_cpu->lock, proc_flags);
1345	if (kfifo_put(&mf_cpu->fifo, entry))
1346		schedule_work_on(smp_processor_id(), &mf_cpu->work);
1347	else
1348		pr_err("Memory failure: buffer overflow when queuing memory failure at %#lx\n",
1349		       pfn);
1350	spin_unlock_irqrestore(&mf_cpu->lock, proc_flags);
1351	put_cpu_var(memory_failure_cpu);
1352}
1353EXPORT_SYMBOL_GPL(memory_failure_queue);
1354
1355static void memory_failure_work_func(struct work_struct *work)
1356{
1357	struct memory_failure_cpu *mf_cpu;
1358	struct memory_failure_entry entry = { 0, };
1359	unsigned long proc_flags;
1360	int gotten;
1361
1362	mf_cpu = this_cpu_ptr(&memory_failure_cpu);
1363	for (;;) {
1364		spin_lock_irqsave(&mf_cpu->lock, proc_flags);
1365		gotten = kfifo_get(&mf_cpu->fifo, &entry);
1366		spin_unlock_irqrestore(&mf_cpu->lock, proc_flags);
1367		if (!gotten)
1368			break;
1369		if (entry.flags & MF_SOFT_OFFLINE)
1370			soft_offline_page(pfn_to_page(entry.pfn), entry.flags);
1371		else
1372			memory_failure(entry.pfn, entry.flags);
1373	}
1374}
1375
1376static int __init memory_failure_init(void)
1377{
1378	struct memory_failure_cpu *mf_cpu;
1379	int cpu;
1380
1381	for_each_possible_cpu(cpu) {
1382		mf_cpu = &per_cpu(memory_failure_cpu, cpu);
1383		spin_lock_init(&mf_cpu->lock);
1384		INIT_KFIFO(mf_cpu->fifo);
1385		INIT_WORK(&mf_cpu->work, memory_failure_work_func);
1386	}
1387
1388	return 0;
1389}
1390core_initcall(memory_failure_init);
1391
1392#define unpoison_pr_info(fmt, pfn, rs)			\
1393({							\
1394	if (__ratelimit(rs))				\
1395		pr_info(fmt, pfn);			\
1396})
1397
1398/**
1399 * unpoison_memory - Unpoison a previously poisoned page
1400 * @pfn: Page number of the to be unpoisoned page
1401 *
1402 * Software-unpoison a page that has been poisoned by
1403 * memory_failure() earlier.
1404 *
1405 * This is only done on the software-level, so it only works
1406 * for linux injected failures, not real hardware failures
1407 *
1408 * Returns 0 for success, otherwise -errno.
1409 */
1410int unpoison_memory(unsigned long pfn)
1411{
1412	struct page *page;
1413	struct page *p;
1414	int freeit = 0;
1415	static DEFINE_RATELIMIT_STATE(unpoison_rs, DEFAULT_RATELIMIT_INTERVAL,
1416					DEFAULT_RATELIMIT_BURST);
1417
1418	if (!pfn_valid(pfn))
1419		return -ENXIO;
1420
1421	p = pfn_to_page(pfn);
1422	page = compound_head(p);
1423
1424	if (!PageHWPoison(p)) {
1425		unpoison_pr_info("Unpoison: Page was already unpoisoned %#lx\n",
1426				 pfn, &unpoison_rs);
1427		return 0;
1428	}
1429
1430	if (page_count(page) > 1) {
1431		unpoison_pr_info("Unpoison: Someone grabs the hwpoison page %#lx\n",
1432				 pfn, &unpoison_rs);
1433		return 0;
1434	}
1435
1436	if (page_mapped(page)) {
1437		unpoison_pr_info("Unpoison: Someone maps the hwpoison page %#lx\n",
1438				 pfn, &unpoison_rs);
1439		return 0;
1440	}
1441
1442	if (page_mapping(page)) {
1443		unpoison_pr_info("Unpoison: the hwpoison page has non-NULL mapping %#lx\n",
1444				 pfn, &unpoison_rs);
1445		return 0;
1446	}
1447
1448	/*
1449	 * unpoison_memory() can encounter thp only when the thp is being
1450	 * worked by memory_failure() and the page lock is not held yet.
1451	 * In such case, we yield to memory_failure() and make unpoison fail.
1452	 */
1453	if (!PageHuge(page) && PageTransHuge(page)) {
1454		unpoison_pr_info("Unpoison: Memory failure is now running on %#lx\n",
1455				 pfn, &unpoison_rs);
1456		return 0;
1457	}
1458
1459	if (!get_hwpoison_page(p)) {
 
 
 
 
 
 
 
 
 
 
 
 
1460		if (TestClearPageHWPoison(p))
1461			num_poisoned_pages_dec();
1462		unpoison_pr_info("Unpoison: Software-unpoisoned free page %#lx\n",
1463				 pfn, &unpoison_rs);
1464		return 0;
1465	}
1466
1467	lock_page(page);
1468	/*
1469	 * This test is racy because PG_hwpoison is set outside of page lock.
1470	 * That's acceptable because that won't trigger kernel panic. Instead,
1471	 * the PG_hwpoison page will be caught and isolated on the entrance to
1472	 * the free buddy page pool.
1473	 */
1474	if (TestClearPageHWPoison(page)) {
1475		unpoison_pr_info("Unpoison: Software-unpoisoned page %#lx\n",
1476				 pfn, &unpoison_rs);
1477		num_poisoned_pages_dec();
1478		freeit = 1;
 
 
1479	}
1480	unlock_page(page);
1481
1482	put_hwpoison_page(page);
1483	if (freeit && !(pfn == my_zero_pfn(0) && page_count(p) == 1))
1484		put_hwpoison_page(page);
1485
1486	return 0;
1487}
1488EXPORT_SYMBOL(unpoison_memory);
1489
1490static struct page *new_page(struct page *p, unsigned long private)
1491{
1492	int nid = page_to_nid(p);
1493
1494	return new_page_nodemask(p, nid, &node_states[N_MEMORY]);
 
 
 
1495}
1496
1497/*
1498 * Safely get reference count of an arbitrary page.
1499 * Returns 0 for a free page, -EIO for a zero refcount page
1500 * that is not free, and 1 for any other page type.
1501 * For 1 the page is returned with increased page count, otherwise not.
1502 */
1503static int __get_any_page(struct page *p, unsigned long pfn, int flags)
1504{
1505	int ret;
1506
1507	if (flags & MF_COUNT_INCREASED)
1508		return 1;
1509
1510	/*
1511	 * When the target page is a free hugepage, just remove it
1512	 * from free hugepage list.
1513	 */
1514	if (!get_hwpoison_page(p)) {
1515		if (PageHuge(p)) {
1516			pr_info("%s: %#lx free huge page\n", __func__, pfn);
1517			ret = 0;
1518		} else if (is_free_buddy_page(p)) {
1519			pr_info("%s: %#lx free buddy page\n", __func__, pfn);
1520			ret = 0;
1521		} else {
1522			pr_info("%s: %#lx: unknown zero refcount page type %lx\n",
1523				__func__, pfn, p->flags);
1524			ret = -EIO;
1525		}
1526	} else {
1527		/* Not a free page */
1528		ret = 1;
1529	}
1530	return ret;
1531}
1532
1533static int get_any_page(struct page *page, unsigned long pfn, int flags)
1534{
1535	int ret = __get_any_page(page, pfn, flags);
1536
1537	if (ret == 1 && !PageHuge(page) &&
1538	    !PageLRU(page) && !__PageMovable(page)) {
1539		/*
1540		 * Try to free it.
1541		 */
1542		put_hwpoison_page(page);
1543		shake_page(page, 1);
1544
1545		/*
1546		 * Did it turn free?
1547		 */
1548		ret = __get_any_page(page, pfn, 0);
1549		if (ret == 1 && !PageLRU(page)) {
1550			/* Drop page reference which is from __get_any_page() */
1551			put_hwpoison_page(page);
1552			pr_info("soft_offline: %#lx: unknown non LRU page type %lx (%pGp)\n",
1553				pfn, page->flags, &page->flags);
1554			return -EIO;
1555		}
1556	}
1557	return ret;
1558}
1559
1560static int soft_offline_huge_page(struct page *page, int flags)
1561{
1562	int ret;
1563	unsigned long pfn = page_to_pfn(page);
1564	struct page *hpage = compound_head(page);
1565	LIST_HEAD(pagelist);
1566
1567	/*
1568	 * This double-check of PageHWPoison is to avoid the race with
1569	 * memory_failure(). See also comment in __soft_offline_page().
1570	 */
1571	lock_page(hpage);
1572	if (PageHWPoison(hpage)) {
1573		unlock_page(hpage);
1574		put_hwpoison_page(hpage);
1575		pr_info("soft offline: %#lx hugepage already poisoned\n", pfn);
1576		return -EBUSY;
1577	}
1578	unlock_page(hpage);
1579
1580	ret = isolate_huge_page(hpage, &pagelist);
1581	/*
1582	 * get_any_page() and isolate_huge_page() takes a refcount each,
1583	 * so need to drop one here.
1584	 */
1585	put_hwpoison_page(hpage);
1586	if (!ret) {
1587		pr_info("soft offline: %#lx hugepage failed to isolate\n", pfn);
1588		return -EBUSY;
1589	}
1590
1591	ret = migrate_pages(&pagelist, new_page, NULL, MPOL_MF_MOVE_ALL,
1592				MIGRATE_SYNC, MR_MEMORY_FAILURE);
1593	if (ret) {
1594		pr_info("soft offline: %#lx: hugepage migration failed %d, type %lx (%pGp)\n",
1595			pfn, ret, page->flags, &page->flags);
1596		if (!list_empty(&pagelist))
1597			putback_movable_pages(&pagelist);
 
 
 
 
1598		if (ret > 0)
1599			ret = -EIO;
1600	} else {
1601		if (PageHuge(page))
1602			dissolve_free_huge_page(page);
 
 
 
 
 
 
 
 
1603	}
1604	return ret;
1605}
1606
1607static int __soft_offline_page(struct page *page, int flags)
1608{
1609	int ret;
1610	unsigned long pfn = page_to_pfn(page);
1611
1612	/*
1613	 * Check PageHWPoison again inside page lock because PageHWPoison
1614	 * is set by memory_failure() outside page lock. Note that
1615	 * memory_failure() also double-checks PageHWPoison inside page lock,
1616	 * so there's no race between soft_offline_page() and memory_failure().
1617	 */
1618	lock_page(page);
1619	wait_on_page_writeback(page);
1620	if (PageHWPoison(page)) {
1621		unlock_page(page);
1622		put_hwpoison_page(page);
1623		pr_info("soft offline: %#lx page already poisoned\n", pfn);
1624		return -EBUSY;
1625	}
1626	/*
1627	 * Try to invalidate first. This should work for
1628	 * non dirty unmapped page cache pages.
1629	 */
1630	ret = invalidate_inode_page(page);
1631	unlock_page(page);
1632	/*
1633	 * RED-PEN would be better to keep it isolated here, but we
1634	 * would need to fix isolation locking first.
1635	 */
1636	if (ret == 1) {
1637		put_hwpoison_page(page);
1638		pr_info("soft_offline: %#lx: invalidated\n", pfn);
1639		SetPageHWPoison(page);
1640		num_poisoned_pages_inc();
1641		return 0;
1642	}
1643
1644	/*
1645	 * Simple invalidation didn't work.
1646	 * Try to migrate to a new page instead. migrate.c
1647	 * handles a large number of cases for us.
1648	 */
1649	if (PageLRU(page))
1650		ret = isolate_lru_page(page);
1651	else
1652		ret = isolate_movable_page(page, ISOLATE_UNEVICTABLE);
1653	/*
1654	 * Drop page reference which is came from get_any_page()
1655	 * successful isolate_lru_page() already took another one.
1656	 */
1657	put_hwpoison_page(page);
1658	if (!ret) {
1659		LIST_HEAD(pagelist);
1660		/*
1661		 * After isolated lru page, the PageLRU will be cleared,
1662		 * so use !__PageMovable instead for LRU page's mapping
1663		 * cannot have PAGE_MAPPING_MOVABLE.
1664		 */
1665		if (!__PageMovable(page))
1666			inc_node_page_state(page, NR_ISOLATED_ANON +
1667						page_is_file_cache(page));
1668		list_add(&page->lru, &pagelist);
1669		ret = migrate_pages(&pagelist, new_page, NULL, MPOL_MF_MOVE_ALL,
1670					MIGRATE_SYNC, MR_MEMORY_FAILURE);
1671		if (ret) {
1672			if (!list_empty(&pagelist))
1673				putback_movable_pages(&pagelist);
 
 
 
 
1674
1675			pr_info("soft offline: %#lx: migration failed %d, type %lx (%pGp)\n",
1676				pfn, ret, page->flags, &page->flags);
1677			if (ret > 0)
1678				ret = -EIO;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1679		}
1680	} else {
1681		pr_info("soft offline: %#lx: isolation failed: %d, page count %d, type %lx (%pGp)\n",
1682			pfn, ret, page_count(page), page->flags, &page->flags);
1683	}
1684	return ret;
1685}
1686
1687static int soft_offline_in_use_page(struct page *page, int flags)
1688{
1689	int ret;
1690	struct page *hpage = compound_head(page);
1691
1692	if (!PageHuge(page) && PageTransHuge(hpage)) {
1693		lock_page(hpage);
1694		if (!PageAnon(hpage) || unlikely(split_huge_page(hpage))) {
1695			unlock_page(hpage);
1696			if (!PageAnon(hpage))
1697				pr_info("soft offline: %#lx: non anonymous thp\n", page_to_pfn(page));
1698			else
1699				pr_info("soft offline: %#lx: thp split failed\n", page_to_pfn(page));
1700			put_hwpoison_page(hpage);
1701			return -EBUSY;
1702		}
1703		unlock_page(hpage);
1704		get_hwpoison_page(page);
1705		put_hwpoison_page(hpage);
1706	}
1707
1708	if (PageHuge(page))
1709		ret = soft_offline_huge_page(page, flags);
1710	else
1711		ret = __soft_offline_page(page, flags);
1712
1713	return ret;
1714}
1715
1716static void soft_offline_free_page(struct page *page)
1717{
1718	struct page *head = compound_head(page);
1719
1720	if (!TestSetPageHWPoison(head)) {
1721		num_poisoned_pages_inc();
1722		if (PageHuge(head))
1723			dissolve_free_huge_page(page);
1724	}
1725}
1726
1727/**
1728 * soft_offline_page - Soft offline a page.
1729 * @page: page to offline
1730 * @flags: flags. Same as memory_failure().
1731 *
1732 * Returns 0 on success, otherwise negated errno.
1733 *
1734 * Soft offline a page, by migration or invalidation,
1735 * without killing anything. This is for the case when
1736 * a page is not corrupted yet (so it's still valid to access),
1737 * but has had a number of corrected errors and is better taken
1738 * out.
1739 *
1740 * The actual policy on when to do that is maintained by
1741 * user space.
1742 *
1743 * This should never impact any application or cause data loss,
1744 * however it might take some time.
1745 *
1746 * This is not a 100% solution for all memory, but tries to be
1747 * ``good enough'' for the majority of memory.
1748 */
1749int soft_offline_page(struct page *page, int flags)
1750{
1751	int ret;
1752	unsigned long pfn = page_to_pfn(page);
 
1753
1754	if (PageHWPoison(page)) {
1755		pr_info("soft offline: %#lx page already poisoned\n", pfn);
1756		if (flags & MF_COUNT_INCREASED)
1757			put_hwpoison_page(page);
1758		return -EBUSY;
1759	}
 
 
 
 
 
 
 
1760
1761	get_online_mems();
1762	ret = get_any_page(page, pfn, flags);
1763	put_online_mems();
 
 
1764
1765	if (ret > 0)
1766		ret = soft_offline_in_use_page(page, flags);
1767	else if (ret == 0)
1768		soft_offline_free_page(page);
 
 
 
1769
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1770	return ret;
1771}