Linux Audio

Check our new training course

In-person Linux kernel drivers training

Jun 16-20, 2025
Register
Loading...
v3.1
 
   1/*
   2 * Copyright (C) 2008, 2009 Intel Corporation
   3 * Authors: Andi Kleen, Fengguang Wu
   4 *
   5 * This software may be redistributed and/or modified under the terms of
   6 * the GNU General Public License ("GPL") version 2 only as published by the
   7 * Free Software Foundation.
   8 *
   9 * High level machine check handler. Handles pages reported by the
  10 * hardware as being corrupted usually due to a multi-bit ECC memory or cache
  11 * failure.
  12 * 
  13 * In addition there is a "soft offline" entry point that allows stop using
  14 * not-yet-corrupted-by-suspicious pages without killing anything.
  15 *
  16 * Handles page cache pages in various states.	The tricky part
  17 * here is that we can access any page asynchronously in respect to 
  18 * other VM users, because memory failures could happen anytime and 
  19 * anywhere. This could violate some of their assumptions. This is why 
  20 * this code has to be extremely careful. Generally it tries to use 
  21 * normal locking rules, as in get the standard locks, even if that means 
  22 * the error handling takes potentially a long time.
  23 * 
 
 
 
 
 
 
 
 
  24 * There are several operations here with exponential complexity because
  25 * of unsuitable VM data structures. For example the operation to map back 
  26 * from RMAP chains to processes has to walk the complete process list and 
  27 * has non linear complexity with the number. But since memory corruptions
  28 * are rare we hope to get away with this. This avoids impacting the core 
  29 * VM.
  30 */
  31
  32/*
  33 * Notebook:
  34 * - hugetlb needs more code
  35 * - kcore/oldmem/vmcore/mem/kmem check for hwpoison pages
  36 * - pass bad pages to kdump next kernel
  37 */
  38#include <linux/kernel.h>
  39#include <linux/mm.h>
  40#include <linux/page-flags.h>
  41#include <linux/kernel-page-flags.h>
  42#include <linux/sched.h>
 
  43#include <linux/ksm.h>
  44#include <linux/rmap.h>
 
  45#include <linux/pagemap.h>
  46#include <linux/swap.h>
  47#include <linux/backing-dev.h>
  48#include <linux/migrate.h>
  49#include <linux/page-isolation.h>
  50#include <linux/suspend.h>
  51#include <linux/slab.h>
  52#include <linux/swapops.h>
  53#include <linux/hugetlb.h>
  54#include <linux/memory_hotplug.h>
  55#include <linux/mm_inline.h>
 
  56#include <linux/kfifo.h>
 
 
 
 
 
  57#include "internal.h"
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  58
  59int sysctl_memory_failure_early_kill __read_mostly = 0;
 
 
 
  60
  61int sysctl_memory_failure_recovery __read_mostly = 1;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  62
  63atomic_long_t mce_bad_pages __read_mostly = ATOMIC_LONG_INIT(0);
 
 
 
 
 
 
 
 
  64
  65#if defined(CONFIG_HWPOISON_INJECT) || defined(CONFIG_HWPOISON_INJECT_MODULE)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  66
  67u32 hwpoison_filter_enable = 0;
  68u32 hwpoison_filter_dev_major = ~0U;
  69u32 hwpoison_filter_dev_minor = ~0U;
  70u64 hwpoison_filter_flags_mask;
  71u64 hwpoison_filter_flags_value;
  72EXPORT_SYMBOL_GPL(hwpoison_filter_enable);
  73EXPORT_SYMBOL_GPL(hwpoison_filter_dev_major);
  74EXPORT_SYMBOL_GPL(hwpoison_filter_dev_minor);
  75EXPORT_SYMBOL_GPL(hwpoison_filter_flags_mask);
  76EXPORT_SYMBOL_GPL(hwpoison_filter_flags_value);
  77
  78static int hwpoison_filter_dev(struct page *p)
  79{
  80	struct address_space *mapping;
  81	dev_t dev;
  82
  83	if (hwpoison_filter_dev_major == ~0U &&
  84	    hwpoison_filter_dev_minor == ~0U)
  85		return 0;
  86
  87	/*
  88	 * page_mapping() does not accept slab pages.
  89	 */
  90	if (PageSlab(p))
  91		return -EINVAL;
  92
  93	mapping = page_mapping(p);
  94	if (mapping == NULL || mapping->host == NULL)
  95		return -EINVAL;
  96
  97	dev = mapping->host->i_sb->s_dev;
  98	if (hwpoison_filter_dev_major != ~0U &&
  99	    hwpoison_filter_dev_major != MAJOR(dev))
 100		return -EINVAL;
 101	if (hwpoison_filter_dev_minor != ~0U &&
 102	    hwpoison_filter_dev_minor != MINOR(dev))
 103		return -EINVAL;
 104
 105	return 0;
 106}
 107
 108static int hwpoison_filter_flags(struct page *p)
 109{
 110	if (!hwpoison_filter_flags_mask)
 111		return 0;
 112
 113	if ((stable_page_flags(p) & hwpoison_filter_flags_mask) ==
 114				    hwpoison_filter_flags_value)
 115		return 0;
 116	else
 117		return -EINVAL;
 118}
 119
 120/*
 121 * This allows stress tests to limit test scope to a collection of tasks
 122 * by putting them under some memcg. This prevents killing unrelated/important
 123 * processes such as /sbin/init. Note that the target task may share clean
 124 * pages with init (eg. libc text), which is harmless. If the target task
 125 * share _dirty_ pages with another task B, the test scheme must make sure B
 126 * is also included in the memcg. At last, due to race conditions this filter
 127 * can only guarantee that the page either belongs to the memcg tasks, or is
 128 * a freed page.
 129 */
 130#ifdef	CONFIG_CGROUP_MEM_RES_CTLR_SWAP
 131u64 hwpoison_filter_memcg;
 132EXPORT_SYMBOL_GPL(hwpoison_filter_memcg);
 133static int hwpoison_filter_task(struct page *p)
 134{
 135	struct mem_cgroup *mem;
 136	struct cgroup_subsys_state *css;
 137	unsigned long ino;
 138
 139	if (!hwpoison_filter_memcg)
 140		return 0;
 141
 142	mem = try_get_mem_cgroup_from_page(p);
 143	if (!mem)
 144		return -EINVAL;
 145
 146	css = mem_cgroup_css(mem);
 147	/* root_mem_cgroup has NULL dentries */
 148	if (!css->cgroup->dentry)
 149		return -EINVAL;
 150
 151	ino = css->cgroup->dentry->d_inode->i_ino;
 152	css_put(css);
 153
 154	if (ino != hwpoison_filter_memcg)
 155		return -EINVAL;
 156
 157	return 0;
 158}
 159#else
 160static int hwpoison_filter_task(struct page *p) { return 0; }
 161#endif
 162
 163int hwpoison_filter(struct page *p)
 164{
 165	if (!hwpoison_filter_enable)
 166		return 0;
 167
 168	if (hwpoison_filter_dev(p))
 169		return -EINVAL;
 170
 171	if (hwpoison_filter_flags(p))
 172		return -EINVAL;
 173
 174	if (hwpoison_filter_task(p))
 175		return -EINVAL;
 176
 177	return 0;
 178}
 179#else
 180int hwpoison_filter(struct page *p)
 181{
 182	return 0;
 183}
 184#endif
 185
 186EXPORT_SYMBOL_GPL(hwpoison_filter);
 187
 188/*
 189 * Send all the processes who have the page mapped an ``action optional''
 190 * signal.
 191 */
 192static int kill_proc_ao(struct task_struct *t, unsigned long addr, int trapno,
 193			unsigned long pfn, struct page *page)
 194{
 195	struct siginfo si;
 196	int ret;
 197
 198	printk(KERN_ERR
 199		"MCE %#lx: Killing %s:%d early due to hardware memory corruption\n",
 200		pfn, t->comm, t->pid);
 201	si.si_signo = SIGBUS;
 202	si.si_errno = 0;
 203	si.si_code = BUS_MCEERR_AO;
 204	si.si_addr = (void *)addr;
 205#ifdef __ARCH_SI_TRAPNO
 206	si.si_trapno = trapno;
 207#endif
 208	si.si_addr_lsb = compound_trans_order(compound_head(page)) + PAGE_SHIFT;
 209	/*
 210	 * Don't use force here, it's convenient if the signal
 211	 * can be temporarily blocked.
 212	 * This could cause a loop when the user sets SIGBUS
 213	 * to SIG_IGN, but hopefully no one will do that?
 214	 */
 215	ret = send_sig_info(SIGBUS, &si, t);  /* synchronous? */
 216	if (ret < 0)
 217		printk(KERN_INFO "MCE: Error sending signal to %s:%d: %d\n",
 218		       t->comm, t->pid, ret);
 219	return ret;
 220}
 221
 222/*
 223 * When a unknown page type is encountered drain as many buffers as possible
 224 * in the hope to turn the page into a LRU or free page, which we can handle.
 225 */
 226void shake_page(struct page *p, int access)
 227{
 228	if (!PageSlab(p)) {
 229		lru_add_drain_all();
 230		if (PageLRU(p))
 231			return;
 232		drain_all_pages();
 233		if (PageLRU(p) || is_free_buddy_page(p))
 234			return;
 235	}
 236
 237	/*
 238	 * Only call shrink_slab here (which would also shrink other caches) if
 239	 * access is not potentially fatal.
 240	 */
 241	if (access) {
 242		int nr;
 243		do {
 244			struct shrink_control shrink = {
 245				.gfp_mask = GFP_KERNEL,
 246			};
 247
 248			nr = shrink_slab(&shrink, 1000, 1000);
 249			if (page_count(p) == 1)
 250				break;
 251		} while (nr > 10);
 252	}
 253}
 254EXPORT_SYMBOL_GPL(shake_page);
 255
 256/*
 257 * Kill all processes that have a poisoned page mapped and then isolate
 258 * the page.
 259 *
 260 * General strategy:
 261 * Find all processes having the page mapped and kill them.
 262 * But we keep a page reference around so that the page is not
 263 * actually freed yet.
 264 * Then stash the page away
 265 *
 266 * There's no convenient way to get back to mapped processes
 267 * from the VMAs. So do a brute-force search over all
 268 * running processes.
 269 *
 270 * Remember that machine checks are not common (or rather
 271 * if they are common you have other problems), so this shouldn't
 272 * be a performance issue.
 273 *
 274 * Also there are some races possible while we get from the
 275 * error detection to actually handle it.
 276 */
 277
 278struct to_kill {
 279	struct list_head nd;
 280	struct task_struct *tsk;
 281	unsigned long addr;
 282	char addr_valid;
 283};
 284
 285/*
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 286 * Failure handling: if we can't find or can't kill a process there's
 287 * not much we can do.	We just print a message and ignore otherwise.
 288 */
 289
 
 
 290/*
 291 * Schedule a process for later kill.
 292 * Uses GFP_ATOMIC allocations to avoid potential recursions in the VM.
 293 * TBD would GFP_NOIO be enough?
 294 */
 295static void add_to_kill(struct task_struct *tsk, struct page *p,
 296		       struct vm_area_struct *vma,
 297		       struct list_head *to_kill,
 298		       struct to_kill **tkc)
 
 
 
 
 299{
 300	struct to_kill *tk;
 301
 302	if (*tkc) {
 303		tk = *tkc;
 304		*tkc = NULL;
 305	} else {
 306		tk = kmalloc(sizeof(struct to_kill), GFP_ATOMIC);
 307		if (!tk) {
 308			printk(KERN_ERR
 309		"MCE: Out of memory while machine check handling\n");
 310			return;
 311		}
 312	}
 313	tk->addr = page_address_in_vma(p, vma);
 314	tk->addr_valid = 1;
 315
 316	/*
 317	 * In theory we don't have to kill when the page was
 318	 * munmaped. But it could be also a mremap. Since that's
 319	 * likely very rare kill anyways just out of paranoia, but use
 320	 * a SIGKILL because the error is not contained anymore.
 
 
 
 
 
 
 
 
 
 
 
 
 321	 */
 322	if (tk->addr == -EFAULT) {
 323		pr_info("MCE: Unable to find user space address %lx in %s\n",
 324			page_to_pfn(p), tsk->comm);
 325		tk->addr_valid = 0;
 
 
 326	}
 
 327	get_task_struct(tsk);
 328	tk->tsk = tsk;
 329	list_add_tail(&tk->nd, to_kill);
 330}
 331
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 332/*
 333 * Kill the processes that have been collected earlier.
 334 *
 335 * Only do anything when DOIT is set, otherwise just free the list
 336 * (this is used for clean pages which do not need killing)
 337 * Also when FAIL is set do a force kill because something went
 338 * wrong earlier.
 339 */
 340static void kill_procs_ao(struct list_head *to_kill, int doit, int trapno,
 341			  int fail, struct page *page, unsigned long pfn)
 342{
 343	struct to_kill *tk, *next;
 344
 345	list_for_each_entry_safe (tk, next, to_kill, nd) {
 346		if (doit) {
 347			/*
 348			 * In case something went wrong with munmapping
 349			 * make sure the process doesn't catch the
 350			 * signal and then access the memory. Just kill it.
 351			 */
 352			if (fail || tk->addr_valid == 0) {
 353				printk(KERN_ERR
 354		"MCE %#lx: forcibly killing %s:%d because of failure to unmap corrupted page\n",
 355					pfn, tk->tsk->comm, tk->tsk->pid);
 356				force_sig(SIGKILL, tk->tsk);
 357			}
 358
 359			/*
 360			 * In theory the process could have mapped
 361			 * something else on the address in-between. We could
 362			 * check for that, but we need to tell the
 363			 * process anyways.
 364			 */
 365			else if (kill_proc_ao(tk->tsk, tk->addr, trapno,
 366					      pfn, page) < 0)
 367				printk(KERN_ERR
 368		"MCE %#lx: Cannot send advisory machine check signal to %s:%d\n",
 369					pfn, tk->tsk->comm, tk->tsk->pid);
 370		}
 
 371		put_task_struct(tk->tsk);
 372		kfree(tk);
 373	}
 374}
 375
 376static int task_early_kill(struct task_struct *tsk)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 377{
 378	if (!tsk->mm)
 379		return 0;
 380	if (tsk->flags & PF_MCE_PROCESS)
 381		return !!(tsk->flags & PF_MCE_EARLY);
 382	return sysctl_memory_failure_early_kill;
 
 
 
 
 
 383}
 384
 385/*
 386 * Collect processes when the error hit an anonymous page.
 387 */
 388static void collect_procs_anon(struct page *page, struct list_head *to_kill,
 389			      struct to_kill **tkc)
 390{
 391	struct vm_area_struct *vma;
 392	struct task_struct *tsk;
 393	struct anon_vma *av;
 
 394
 395	av = page_lock_anon_vma(page);
 396	if (av == NULL)	/* Not actually mapped anymore */
 397		return;
 398
 399	read_lock(&tasklist_lock);
 400	for_each_process (tsk) {
 
 401		struct anon_vma_chain *vmac;
 
 402
 403		if (!task_early_kill(tsk))
 404			continue;
 405		list_for_each_entry(vmac, &av->head, same_anon_vma) {
 
 406			vma = vmac->vma;
 
 
 407			if (!page_mapped_in_vma(page, vma))
 408				continue;
 409			if (vma->vm_mm == tsk->mm)
 410				add_to_kill(tsk, page, vma, to_kill, tkc);
 411		}
 412	}
 413	read_unlock(&tasklist_lock);
 414	page_unlock_anon_vma(av);
 415}
 416
 417/*
 418 * Collect processes when the error hit a file mapped page.
 419 */
 420static void collect_procs_file(struct page *page, struct list_head *to_kill,
 421			      struct to_kill **tkc)
 422{
 423	struct vm_area_struct *vma;
 424	struct task_struct *tsk;
 425	struct prio_tree_iter iter;
 426	struct address_space *mapping = page->mapping;
 427
 428	mutex_lock(&mapping->i_mmap_mutex);
 429	read_lock(&tasklist_lock);
 
 430	for_each_process(tsk) {
 431		pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
 432
 433		if (!task_early_kill(tsk))
 434			continue;
 435
 436		vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, pgoff,
 437				      pgoff) {
 438			/*
 439			 * Send early kill signal to tasks where a vma covers
 440			 * the page but the corrupted page is not necessarily
 441			 * mapped it in its pte.
 442			 * Assume applications who requested early kill want
 443			 * to be informed of all such data corruptions.
 444			 */
 445			if (vma->vm_mm == tsk->mm)
 446				add_to_kill(tsk, page, vma, to_kill, tkc);
 447		}
 448	}
 449	read_unlock(&tasklist_lock);
 450	mutex_unlock(&mapping->i_mmap_mutex);
 
 
 
 
 
 
 
 
 451}
 452
 453/*
 454 * Collect the processes who have the corrupted page mapped to kill.
 455 * This is done in two steps for locking reasons.
 456 * First preallocate one tokill structure outside the spin locks,
 457 * so that we can kill at least one process reasonably reliable.
 458 */
 459static void collect_procs(struct page *page, struct list_head *tokill)
 
 
 460{
 461	struct to_kill *tk;
 
 462
 463	if (!page->mapping)
 464		return;
 
 
 465
 466	tk = kmalloc(sizeof(struct to_kill), GFP_NOIO);
 467	if (!tk)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 468		return;
 469	if (PageAnon(page))
 470		collect_procs_anon(page, tokill, &tk);
 
 
 471	else
 472		collect_procs_file(page, tokill, &tk);
 473	kfree(tk);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 474}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 475
 476/*
 477 * Error handlers for various types of pages.
 
 
 
 
 
 
 
 
 
 
 478 */
 
 
 
 
 
 
 
 
 479
 480enum outcome {
 481	IGNORED,	/* Error: cannot be handled */
 482	FAILED,		/* Error: handling failed */
 483	DELAYED,	/* Will be handled later */
 484	RECOVERED,	/* Successfully recovered */
 485};
 
 
 
 
 
 
 
 486
 487static const char *action_name[] = {
 488	[IGNORED] = "Ignored",
 489	[FAILED] = "Failed",
 490	[DELAYED] = "Delayed",
 491	[RECOVERED] = "Recovered",
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 492};
 493
 494/*
 495 * XXX: It is possible that a page is isolated from LRU cache,
 496 * and then kept in swap cache or failed to remove from page cache.
 497 * The page count will stop it from being freed by unpoison.
 498 * Stress tests should be aware of this memory leak problem.
 499 */
 500static int delete_from_lru_cache(struct page *p)
 501{
 502	if (!isolate_lru_page(p)) {
 503		/*
 504		 * Clear sensible page flags, so that the buddy system won't
 505		 * complain when the page is unpoison-and-freed.
 506		 */
 507		ClearPageActive(p);
 508		ClearPageUnevictable(p);
 
 509		/*
 510		 * drop the page count elevated by isolate_lru_page()
 
 511		 */
 512		page_cache_release(p);
 
 
 
 
 
 513		return 0;
 514	}
 515	return -EIO;
 516}
 517
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 518/*
 519 * Error hit kernel page.
 520 * Do nothing, try to be lucky and not touch this instead. For a few cases we
 521 * could be more sophisticated.
 522 */
 523static int me_kernel(struct page *p, unsigned long pfn)
 524{
 525	return IGNORED;
 
 526}
 527
 528/*
 529 * Page in unknown state. Do nothing.
 530 */
 531static int me_unknown(struct page *p, unsigned long pfn)
 532{
 533	printk(KERN_ERR "MCE %#lx: Unknown page state\n", pfn);
 534	return FAILED;
 
 535}
 536
 537/*
 538 * Clean (or cleaned) page cache page.
 539 */
 540static int me_pagecache_clean(struct page *p, unsigned long pfn)
 541{
 542	int err;
 543	int ret = FAILED;
 544	struct address_space *mapping;
 
 545
 546	delete_from_lru_cache(p);
 547
 548	/*
 549	 * For anonymous pages we're done the only reference left
 550	 * should be the one m_f() holds.
 551	 */
 552	if (PageAnon(p))
 553		return RECOVERED;
 
 
 554
 555	/*
 556	 * Now truncate the page in the page cache. This is really
 557	 * more like a "temporary hole punch"
 558	 * Don't do this for block devices when someone else
 559	 * has a reference, because it could be file system metadata
 560	 * and that's not safe to truncate.
 561	 */
 562	mapping = page_mapping(p);
 563	if (!mapping) {
 564		/*
 565		 * Page has been teared down in the meanwhile
 566		 */
 567		return FAILED;
 568	}
 569
 570	/*
 
 
 
 
 
 
 571	 * Truncation is a bit tricky. Enable it per file system for now.
 572	 *
 573	 * Open: to take i_mutex or not for this? Right now we don't.
 574	 */
 575	if (mapping->a_ops->error_remove_page) {
 576		err = mapping->a_ops->error_remove_page(mapping, p);
 577		if (err != 0) {
 578			printk(KERN_INFO "MCE %#lx: Failed to punch page: %d\n",
 579					pfn, err);
 580		} else if (page_has_private(p) &&
 581				!try_to_release_page(p, GFP_NOIO)) {
 582			pr_info("MCE %#lx: failed to release buffers\n", pfn);
 583		} else {
 584			ret = RECOVERED;
 585		}
 586	} else {
 587		/*
 588		 * If the file system doesn't support it just invalidate
 589		 * This fails on dirty or anything with private pages
 590		 */
 591		if (invalidate_inode_page(p))
 592			ret = RECOVERED;
 593		else
 594			printk(KERN_INFO "MCE %#lx: Failed to invalidate\n",
 595				pfn);
 596	}
 597	return ret;
 598}
 599
 600/*
 601 * Dirty cache page page
 602 * Issues: when the error hit a hole page the error is not properly
 603 * propagated.
 604 */
 605static int me_pagecache_dirty(struct page *p, unsigned long pfn)
 606{
 607	struct address_space *mapping = page_mapping(p);
 608
 609	SetPageError(p);
 610	/* TBD: print more information about the file. */
 611	if (mapping) {
 612		/*
 613		 * IO error will be reported by write(), fsync(), etc.
 614		 * who check the mapping.
 615		 * This way the application knows that something went
 616		 * wrong with its dirty file data.
 617		 *
 618		 * There's one open issue:
 619		 *
 620		 * The EIO will be only reported on the next IO
 621		 * operation and then cleared through the IO map.
 622		 * Normally Linux has two mechanisms to pass IO error
 623		 * first through the AS_EIO flag in the address space
 624		 * and then through the PageError flag in the page.
 625		 * Since we drop pages on memory failure handling the
 626		 * only mechanism open to use is through AS_AIO.
 627		 *
 628		 * This has the disadvantage that it gets cleared on
 629		 * the first operation that returns an error, while
 630		 * the PageError bit is more sticky and only cleared
 631		 * when the page is reread or dropped.  If an
 632		 * application assumes it will always get error on
 633		 * fsync, but does other operations on the fd before
 634		 * and the page is dropped between then the error
 635		 * will not be properly reported.
 636		 *
 637		 * This can already happen even without hwpoisoned
 638		 * pages: first on metadata IO errors (which only
 639		 * report through AS_EIO) or when the page is dropped
 640		 * at the wrong time.
 641		 *
 642		 * So right now we assume that the application DTRT on
 643		 * the first EIO, but we're not worse than other parts
 644		 * of the kernel.
 645		 */
 646		mapping_set_error(mapping, EIO);
 647	}
 648
 649	return me_pagecache_clean(p, pfn);
 650}
 651
 652/*
 653 * Clean and dirty swap cache.
 654 *
 655 * Dirty swap cache page is tricky to handle. The page could live both in page
 656 * cache and swap cache(ie. page is freshly swapped in). So it could be
 657 * referenced concurrently by 2 types of PTEs:
 658 * normal PTEs and swap PTEs. We try to handle them consistently by calling
 659 * try_to_unmap(TTU_IGNORE_HWPOISON) to convert the normal PTEs to swap PTEs,
 660 * and then
 661 *      - clear dirty bit to prevent IO
 662 *      - remove from LRU
 663 *      - but keep in the swap cache, so that when we return to it on
 664 *        a later page fault, we know the application is accessing
 665 *        corrupted data and shall be killed (we installed simple
 666 *        interception code in do_swap_page to catch it).
 667 *
 668 * Clean swap cache pages can be directly isolated. A later page fault will
 669 * bring in the known good data from disk.
 670 */
 671static int me_swapcache_dirty(struct page *p, unsigned long pfn)
 672{
 673	ClearPageDirty(p);
 
 
 
 
 674	/* Trigger EIO in shmem: */
 675	ClearPageUptodate(p);
 676
 677	if (!delete_from_lru_cache(p))
 678		return DELAYED;
 679	else
 680		return FAILED;
 
 
 
 
 
 
 681}
 682
 683static int me_swapcache_clean(struct page *p, unsigned long pfn)
 684{
 685	delete_from_swap_cache(p);
 
 686
 687	if (!delete_from_lru_cache(p))
 688		return RECOVERED;
 689	else
 690		return FAILED;
 
 
 
 
 
 691}
 692
 693/*
 694 * Huge pages. Needs work.
 695 * Issues:
 696 * - Error on hugepage is contained in hugepage unit (not in raw page unit.)
 697 *   To narrow down kill region to one page, we need to break up pmd.
 698 */
 699static int me_huge_page(struct page *p, unsigned long pfn)
 700{
 701	int res = 0;
 702	struct page *hpage = compound_head(p);
 703	/*
 704	 * We can safely recover from error on free or reserved (i.e.
 705	 * not in-use) hugepage by dequeuing it from freelist.
 706	 * To check whether a hugepage is in-use or not, we can't use
 707	 * page->lru because it can be used in other hugepage operations,
 708	 * such as __unmap_hugepage_range() and gather_surplus_pages().
 709	 * So instead we use page_mapping() and PageAnon().
 710	 * We assume that this function is called with page lock held,
 711	 * so there is no race between isolation and mapping/unmapping.
 712	 */
 713	if (!(page_mapping(hpage) || PageAnon(hpage))) {
 714		res = dequeue_hwpoisoned_huge_page(hpage);
 715		if (!res)
 716			return RECOVERED;
 
 
 
 
 
 
 
 
 
 717	}
 718	return DELAYED;
 
 
 
 
 719}
 720
 721/*
 722 * Various page states we can handle.
 723 *
 724 * A page state is defined by its current page->flags bits.
 725 * The table matches them in order and calls the right handler.
 726 *
 727 * This is quite tricky because we can access page at any time
 728 * in its live cycle, so all accesses have to be extremely careful.
 729 *
 730 * This is not complete. More states could be added.
 731 * For any missing state don't attempt recovery.
 732 */
 733
 734#define dirty		(1UL << PG_dirty)
 735#define sc		(1UL << PG_swapcache)
 736#define unevict		(1UL << PG_unevictable)
 737#define mlock		(1UL << PG_mlocked)
 738#define writeback	(1UL << PG_writeback)
 739#define lru		(1UL << PG_lru)
 740#define swapbacked	(1UL << PG_swapbacked)
 741#define head		(1UL << PG_head)
 742#define tail		(1UL << PG_tail)
 743#define compound	(1UL << PG_compound)
 744#define slab		(1UL << PG_slab)
 745#define reserved	(1UL << PG_reserved)
 746
 747static struct page_state {
 748	unsigned long mask;
 749	unsigned long res;
 750	char *msg;
 751	int (*action)(struct page *p, unsigned long pfn);
 752} error_states[] = {
 753	{ reserved,	reserved,	"reserved kernel",	me_kernel },
 754	/*
 755	 * free pages are specially detected outside this table:
 756	 * PG_buddy pages only make a small fraction of all free pages.
 757	 */
 758
 759	/*
 760	 * Could in theory check if slab page is free or if we can drop
 761	 * currently unused objects without touching them. But just
 762	 * treat it as standard kernel for now.
 763	 */
 764	{ slab,		slab,		"kernel slab",	me_kernel },
 765
 766#ifdef CONFIG_PAGEFLAGS_EXTENDED
 767	{ head,		head,		"huge",		me_huge_page },
 768	{ tail,		tail,		"huge",		me_huge_page },
 769#else
 770	{ compound,	compound,	"huge",		me_huge_page },
 771#endif
 772
 773	{ sc|dirty,	sc|dirty,	"swapcache",	me_swapcache_dirty },
 774	{ sc|dirty,	sc,		"swapcache",	me_swapcache_clean },
 775
 776	{ unevict|dirty, unevict|dirty,	"unevictable LRU", me_pagecache_dirty},
 777	{ unevict,	unevict,	"unevictable LRU", me_pagecache_clean},
 778
 779	{ mlock|dirty,	mlock|dirty,	"mlocked LRU",	me_pagecache_dirty },
 780	{ mlock,	mlock,		"mlocked LRU",	me_pagecache_clean },
 781
 782	{ lru|dirty,	lru|dirty,	"LRU",		me_pagecache_dirty },
 783	{ lru|dirty,	lru,		"clean LRU",	me_pagecache_clean },
 784
 785	/*
 786	 * Catchall entry: must be at end.
 787	 */
 788	{ 0,		0,		"unknown page state",	me_unknown },
 789};
 790
 791#undef dirty
 792#undef sc
 793#undef unevict
 794#undef mlock
 795#undef writeback
 796#undef lru
 797#undef swapbacked
 798#undef head
 799#undef tail
 800#undef compound
 801#undef slab
 802#undef reserved
 803
 804static void action_result(unsigned long pfn, char *msg, int result)
 
 805{
 806	struct page *page = pfn_to_page(pfn);
 
 
 
 
 
 
 
 807
 808	printk(KERN_ERR "MCE %#lx: %s%s page recovery: %s\n",
 809		pfn,
 810		PageDirty(page) ? "dirty " : "",
 811		msg, action_name[result]);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 812}
 813
 814static int page_action(struct page_state *ps, struct page *p,
 815			unsigned long pfn)
 816{
 817	int result;
 818	int count;
 819
 820	result = ps->action(p, pfn);
 821	action_result(pfn, ps->msg, result);
 822
 823	count = page_count(p) - 1;
 824	if (ps->action == me_swapcache_dirty && result == DELAYED)
 825		count--;
 826	if (count != 0) {
 827		printk(KERN_ERR
 828		       "MCE %#lx: %s page still referenced by %d users\n",
 829		       pfn, ps->msg, count);
 830		result = FAILED;
 831	}
 832
 833	/* Could do more checks here if page looks ok */
 834	/*
 835	 * Could adjust zone counters here to correct for the missing page.
 836	 */
 837
 838	return (result == RECOVERED || result == DELAYED) ? 0 : -EBUSY;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 839}
 840
 841/*
 842 * Do all that is necessary to remove user space mappings. Unmap
 843 * the pages and send SIGBUS to the processes if the data was dirty.
 844 */
 845static int hwpoison_user_mappings(struct page *p, unsigned long pfn,
 846				  int trapno)
 847{
 848	enum ttu_flags ttu = TTU_UNMAP | TTU_IGNORE_MLOCK | TTU_IGNORE_ACCESS;
 
 849	struct address_space *mapping;
 850	LIST_HEAD(tokill);
 851	int ret;
 852	int kill = 1;
 853	struct page *hpage = compound_head(p);
 854	struct page *ppage;
 855
 856	if (PageReserved(p) || PageSlab(p))
 857		return SWAP_SUCCESS;
 
 
 
 
 
 
 858
 859	/*
 860	 * This check implies we don't kill processes if their pages
 861	 * are in the swap cache early. Those are always late kills.
 862	 */
 863	if (!page_mapped(hpage))
 864		return SWAP_SUCCESS;
 865
 866	if (PageKsm(p))
 867		return SWAP_FAIL;
 868
 869	if (PageSwapCache(p)) {
 870		printk(KERN_ERR
 871		       "MCE %#lx: keeping poisoned page in swap cache\n", pfn);
 872		ttu |= TTU_IGNORE_HWPOISON;
 873	}
 874
 875	/*
 876	 * Propagate the dirty bit from PTEs to struct page first, because we
 877	 * need this to decide if we should kill or just drop the page.
 878	 * XXX: the dirty test could be racy: set_page_dirty() may not always
 879	 * be called inside page lock (it's recommended but not enforced).
 880	 */
 881	mapping = page_mapping(hpage);
 882	if (!PageDirty(hpage) && mapping &&
 883	    mapping_cap_writeback_dirty(mapping)) {
 884		if (page_mkclean(hpage)) {
 885			SetPageDirty(hpage);
 886		} else {
 887			kill = 0;
 888			ttu |= TTU_IGNORE_HWPOISON;
 889			printk(KERN_INFO
 890	"MCE %#lx: corrupted page was clean: dropped without side effects\n",
 891				pfn);
 892		}
 893	}
 894
 895	/*
 896	 * ppage: poisoned page
 897	 *   if p is regular page(4k page)
 898	 *        ppage == real poisoned page;
 899	 *   else p is hugetlb or THP, ppage == head page.
 900	 */
 901	ppage = hpage;
 902
 903	if (PageTransHuge(hpage)) {
 904		/*
 905		 * Verify that this isn't a hugetlbfs head page, the check for
 906		 * PageAnon is just for avoid tripping a split_huge_page
 907		 * internal debug check, as split_huge_page refuses to deal with
 908		 * anything that isn't an anon page. PageAnon can't go away fro
 909		 * under us because we hold a refcount on the hpage, without a
 910		 * refcount on the hpage. split_huge_page can't be safely called
 911		 * in the first place, having a refcount on the tail isn't
 912		 * enough * to be safe.
 913		 */
 914		if (!PageHuge(hpage) && PageAnon(hpage)) {
 915			if (unlikely(split_huge_page(hpage))) {
 916				/*
 917				 * FIXME: if splitting THP is failed, it is
 918				 * better to stop the following operation rather
 919				 * than causing panic by unmapping. System might
 920				 * survive if the page is freed later.
 921				 */
 922				printk(KERN_INFO
 923					"MCE %#lx: failed to split THP\n", pfn);
 924
 925				BUG_ON(!PageHWPoison(p));
 926				return SWAP_FAIL;
 927			}
 928			/* THP is split, so ppage should be the real poisoned page. */
 929			ppage = p;
 930		}
 931	}
 932
 
 
 
 
 
 933	/*
 934	 * First collect all the processes that have the page
 935	 * mapped in dirty form.  This has to be done before try_to_unmap,
 936	 * because ttu takes the rmap data structures down.
 937	 *
 938	 * Error handling: We ignore errors here because
 939	 * there's nothing that can be done.
 940	 */
 941	if (kill)
 942		collect_procs(ppage, &tokill);
 943
 944	if (hpage != ppage)
 945		lock_page(ppage);
 946
 947	ret = try_to_unmap(ppage, ttu);
 948	if (ret != SWAP_SUCCESS)
 949		printk(KERN_ERR "MCE %#lx: failed to unmap page (mapcount=%d)\n",
 950				pfn, page_mapcount(ppage));
 951
 952	if (hpage != ppage)
 953		unlock_page(ppage);
 954
 955	/*
 956	 * Now that the dirty bit has been propagated to the
 957	 * struct page and all unmaps done we can decide if
 958	 * killing is needed or not.  Only kill when the page
 959	 * was dirty, otherwise the tokill list is merely
 
 960	 * freed.  When there was a problem unmapping earlier
 961	 * use a more force-full uncatchable kill to prevent
 962	 * any accesses to the poisoned memory.
 963	 */
 964	kill_procs_ao(&tokill, !!PageDirty(ppage), trapno,
 965		      ret != SWAP_SUCCESS, p, pfn);
 
 966
 967	return ret;
 968}
 969
 970static void set_page_hwpoison_huge_page(struct page *hpage)
 
 971{
 972	int i;
 973	int nr_pages = 1 << compound_trans_order(hpage);
 974	for (i = 0; i < nr_pages; i++)
 975		SetPageHWPoison(hpage + i);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 976}
 977
 978static void clear_page_hwpoison_huge_page(struct page *hpage)
 979{
 980	int i;
 981	int nr_pages = 1 << compound_trans_order(hpage);
 982	for (i = 0; i < nr_pages; i++)
 983		ClearPageHWPoison(hpage + i);
 
 
 
 
 
 
 984}
 985
 986int __memory_failure(unsigned long pfn, int trapno, int flags)
 
 987{
 988	struct page_state *ps;
 989	struct page *p;
 990	struct page *hpage;
 991	int res;
 992	unsigned int nr_pages;
 993
 994	if (!sysctl_memory_failure_recovery)
 995		panic("Memory failure from trap %d on page %lx", trapno, pfn);
 
 996
 997	if (!pfn_valid(pfn)) {
 998		printk(KERN_ERR
 999		       "MCE %#lx: memory outside kernel control\n",
1000		       pfn);
1001		return -ENXIO;
1002	}
 
 
1003
1004	p = pfn_to_page(pfn);
1005	hpage = compound_head(p);
1006	if (TestSetPageHWPoison(p)) {
1007		printk(KERN_ERR "MCE %#lx: already hardware poisoned\n", pfn);
1008		return 0;
1009	}
1010
1011	nr_pages = 1 << compound_trans_order(hpage);
1012	atomic_long_add(nr_pages, &mce_bad_pages);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1013
1014	/*
1015	 * We need/can do nothing about count=0 pages.
1016	 * 1) it's a free page, and therefore in safe hand:
1017	 *    prep_new_page() will be the gate keeper.
1018	 * 2) it's a free hugepage, which is also safe:
1019	 *    an affected hugepage will be dequeued from hugepage freelist,
1020	 *    so there's no concern about reusing it ever after.
1021	 * 3) it's part of a non-compound high order page.
1022	 *    Implies some kernel user: cannot stop them from
1023	 *    R/W the page; let's pray that the page has been
1024	 *    used and will be freed some time later.
1025	 * In fact it's dangerous to directly bump up page count from 0,
1026	 * that may make page_freeze_refs()/page_unfreeze_refs() mismatch.
1027	 */
1028	if (!(flags & MF_COUNT_INCREASED) &&
1029		!get_page_unless_zero(hpage)) {
1030		if (is_free_buddy_page(p)) {
1031			action_result(pfn, "free buddy", DELAYED);
1032			return 0;
1033		} else if (PageHuge(hpage)) {
1034			/*
1035			 * Check "just unpoisoned", "filter hit", and
1036			 * "race with other subpage."
1037			 */
1038			lock_page(hpage);
1039			if (!PageHWPoison(hpage)
1040			    || (hwpoison_filter(p) && TestClearPageHWPoison(p))
1041			    || (p != hpage && TestSetPageHWPoison(hpage))) {
1042				atomic_long_sub(nr_pages, &mce_bad_pages);
1043				return 0;
1044			}
1045			set_page_hwpoison_huge_page(hpage);
1046			res = dequeue_hwpoisoned_huge_page(hpage);
1047			action_result(pfn, "free huge",
1048				      res ? IGNORED : DELAYED);
1049			unlock_page(hpage);
1050			return res;
1051		} else {
1052			action_result(pfn, "high order kernel", IGNORED);
1053			return -EBUSY;
1054		}
1055	}
1056
1057	/*
1058	 * We ignore non-LRU pages for good reasons.
1059	 * - PG_locked is only well defined for LRU pages and a few others
1060	 * - to avoid races with __set_page_locked()
1061	 * - to avoid races with __SetPageSlab*() (and more non-atomic ops)
1062	 * The check (unnecessarily) ignores LRU pages being isolated and
1063	 * walked by the page reclaim code, however that's not a big loss.
1064	 */
1065	if (!PageHuge(p) && !PageTransCompound(p)) {
1066		if (!PageLRU(p))
1067			shake_page(p, 0);
1068		if (!PageLRU(p)) {
1069			/*
1070			 * shake_page could have turned it free.
1071			 */
1072			if (is_free_buddy_page(p)) {
1073				action_result(pfn, "free buddy, 2nd try",
1074						DELAYED);
1075				return 0;
1076			}
1077			action_result(pfn, "non LRU", IGNORED);
1078			put_page(p);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1079			return -EBUSY;
1080		}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1081	}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1082
1083	/*
1084	 * Lock the page and wait for writeback to finish.
1085	 * It's very difficult to mess with pages currently under IO
1086	 * and in many cases impossible, so we just avoid it here.
1087	 */
1088	lock_page(hpage);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1089
1090	/*
1091	 * unpoison always clear PG_hwpoison inside page lock
 
 
1092	 */
1093	if (!PageHWPoison(p)) {
1094		printk(KERN_ERR "MCE %#lx: just unpoisoned\n", pfn);
1095		res = 0;
1096		goto out;
 
 
1097	}
1098	if (hwpoison_filter(p)) {
1099		if (TestClearPageHWPoison(p))
1100			atomic_long_sub(nr_pages, &mce_bad_pages);
1101		unlock_page(hpage);
1102		put_page(hpage);
1103		return 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1104	}
 
 
1105
 
 
1106	/*
1107	 * For error on the tail page, we should set PG_hwpoison
1108	 * on the head page to show that the hugepage is hwpoisoned
1109	 */
1110	if (PageHuge(p) && PageTail(p) && TestSetPageHWPoison(hpage)) {
1111		action_result(pfn, "hugepage already hardware poisoned",
1112				IGNORED);
1113		unlock_page(hpage);
1114		put_page(hpage);
1115		return 0;
1116	}
1117	/*
1118	 * Set PG_hwpoison on all pages in an error hugepage,
1119	 * because containment is done in hugepage unit for now.
1120	 * Since we have done TestSetPageHWPoison() for the head page with
1121	 * page lock held, we can safely set PG_hwpoison bits on tail pages.
1122	 */
1123	if (PageHuge(p))
1124		set_page_hwpoison_huge_page(hpage);
1125
1126	wait_on_page_writeback(p);
 
1127
1128	/*
1129	 * Now take care of user space mappings.
1130	 * Abort on fail: __delete_from_page_cache() assumes unmapped page.
1131	 */
1132	if (hwpoison_user_mappings(p, pfn, trapno) != SWAP_SUCCESS) {
1133		printk(KERN_ERR "MCE %#lx: cannot unmap page, give up\n", pfn);
1134		res = -EBUSY;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1135		goto out;
1136	}
1137
1138	/*
1139	 * Torn down by someone else?
 
1140	 */
1141	if (PageLRU(p) && !PageSwapCache(p) && p->mapping == NULL) {
1142		action_result(pfn, "already truncated LRU", IGNORED);
1143		res = -EBUSY;
1144		goto out;
1145	}
1146
1147	res = -EBUSY;
1148	for (ps = error_states;; ps++) {
1149		if ((p->flags & ps->mask) == ps->res) {
1150			res = page_action(ps, p, pfn);
1151			break;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1152		}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1153	}
 
 
1154out:
1155	unlock_page(hpage);
1156	return res;
 
 
 
1157}
1158EXPORT_SYMBOL_GPL(__memory_failure);
1159
1160/**
1161 * memory_failure - Handle memory failure of a page.
1162 * @pfn: Page Number of the corrupted page
1163 * @trapno: Trap number reported in the signal to user space.
1164 *
1165 * This function is called by the low level machine check code
1166 * of an architecture when it detects hardware memory corruption
1167 * of a page. It tries its best to recover, which includes
1168 * dropping pages, killing processes etc.
1169 *
1170 * The function is primarily of use for corruptions that
1171 * happen outside the current execution context (e.g. when
1172 * detected by a background scrubber)
1173 *
1174 * Must run in process context (e.g. a work queue) with interrupts
1175 * enabled and no spinlocks hold.
 
 
 
 
1176 */
1177void memory_failure(unsigned long pfn, int trapno)
1178{
1179	__memory_failure(pfn, trapno, 0);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1180}
 
1181
1182#define MEMORY_FAILURE_FIFO_ORDER	4
1183#define MEMORY_FAILURE_FIFO_SIZE	(1 << MEMORY_FAILURE_FIFO_ORDER)
1184
1185struct memory_failure_entry {
1186	unsigned long pfn;
1187	int trapno;
1188	int flags;
1189};
1190
1191struct memory_failure_cpu {
1192	DECLARE_KFIFO(fifo, struct memory_failure_entry,
1193		      MEMORY_FAILURE_FIFO_SIZE);
1194	spinlock_t lock;
1195	struct work_struct work;
1196};
1197
1198static DEFINE_PER_CPU(struct memory_failure_cpu, memory_failure_cpu);
1199
1200/**
1201 * memory_failure_queue - Schedule handling memory failure of a page.
1202 * @pfn: Page Number of the corrupted page
1203 * @trapno: Trap number reported in the signal to user space.
1204 * @flags: Flags for memory failure handling
1205 *
1206 * This function is called by the low level hardware error handler
1207 * when it detects hardware memory corruption of a page. It schedules
1208 * the recovering of error page, including dropping pages, killing
1209 * processes etc.
1210 *
1211 * The function is primarily of use for corruptions that
1212 * happen outside the current execution context (e.g. when
1213 * detected by a background scrubber)
1214 *
1215 * Can run in IRQ context.
1216 */
1217void memory_failure_queue(unsigned long pfn, int trapno, int flags)
1218{
1219	struct memory_failure_cpu *mf_cpu;
1220	unsigned long proc_flags;
1221	struct memory_failure_entry entry = {
1222		.pfn =		pfn,
1223		.trapno =	trapno,
1224		.flags =	flags,
1225	};
1226
1227	mf_cpu = &get_cpu_var(memory_failure_cpu);
1228	spin_lock_irqsave(&mf_cpu->lock, proc_flags);
1229	if (kfifo_put(&mf_cpu->fifo, &entry))
1230		schedule_work_on(smp_processor_id(), &mf_cpu->work);
1231	else
1232		pr_err("Memory failure: buffer overflow when queuing memory failure at 0x%#lx\n",
1233		       pfn);
1234	spin_unlock_irqrestore(&mf_cpu->lock, proc_flags);
1235	put_cpu_var(memory_failure_cpu);
1236}
1237EXPORT_SYMBOL_GPL(memory_failure_queue);
1238
1239static void memory_failure_work_func(struct work_struct *work)
1240{
1241	struct memory_failure_cpu *mf_cpu;
1242	struct memory_failure_entry entry = { 0, };
1243	unsigned long proc_flags;
1244	int gotten;
1245
1246	mf_cpu = &__get_cpu_var(memory_failure_cpu);
1247	for (;;) {
1248		spin_lock_irqsave(&mf_cpu->lock, proc_flags);
1249		gotten = kfifo_get(&mf_cpu->fifo, &entry);
1250		spin_unlock_irqrestore(&mf_cpu->lock, proc_flags);
1251		if (!gotten)
1252			break;
1253		__memory_failure(entry.pfn, entry.trapno, entry.flags);
 
 
 
1254	}
1255}
1256
 
 
 
 
 
 
 
 
 
 
 
 
 
1257static int __init memory_failure_init(void)
1258{
1259	struct memory_failure_cpu *mf_cpu;
1260	int cpu;
1261
1262	for_each_possible_cpu(cpu) {
1263		mf_cpu = &per_cpu(memory_failure_cpu, cpu);
1264		spin_lock_init(&mf_cpu->lock);
1265		INIT_KFIFO(mf_cpu->fifo);
1266		INIT_WORK(&mf_cpu->work, memory_failure_work_func);
1267	}
1268
 
 
1269	return 0;
1270}
1271core_initcall(memory_failure_init);
1272
 
 
 
 
 
 
 
 
1273/**
1274 * unpoison_memory - Unpoison a previously poisoned page
1275 * @pfn: Page number of the to be unpoisoned page
1276 *
1277 * Software-unpoison a page that has been poisoned by
1278 * memory_failure() earlier.
1279 *
1280 * This is only done on the software-level, so it only works
1281 * for linux injected failures, not real hardware failures
1282 *
1283 * Returns 0 for success, otherwise -errno.
1284 */
1285int unpoison_memory(unsigned long pfn)
1286{
1287	struct page *page;
1288	struct page *p;
1289	int freeit = 0;
1290	unsigned int nr_pages;
 
 
 
1291
1292	if (!pfn_valid(pfn))
1293		return -ENXIO;
1294
1295	p = pfn_to_page(pfn);
1296	page = compound_head(p);
1297
1298	if (!PageHWPoison(p)) {
1299		pr_info("MCE: Page was already unpoisoned %#lx\n", pfn);
1300		return 0;
 
 
 
 
1301	}
1302
1303	nr_pages = 1 << compound_trans_order(page);
 
 
 
 
1304
1305	if (!get_page_unless_zero(page)) {
1306		/*
1307		 * Since HWPoisoned hugepage should have non-zero refcount,
1308		 * race between memory failure and unpoison seems to happen.
1309		 * In such case unpoison fails and memory failure runs
1310		 * to the end.
1311		 */
1312		if (PageHuge(page)) {
1313			pr_debug("MCE: Memory failure is now running on free hugepage %#lx\n", pfn);
1314			return 0;
1315		}
1316		if (TestClearPageHWPoison(p))
1317			atomic_long_sub(nr_pages, &mce_bad_pages);
1318		pr_info("MCE: Software-unpoisoned free page %#lx\n", pfn);
1319		return 0;
1320	}
1321
1322	lock_page(page);
 
 
 
1323	/*
1324	 * This test is racy because PG_hwpoison is set outside of page lock.
1325	 * That's acceptable because that won't trigger kernel panic. Instead,
1326	 * the PG_hwpoison page will be caught and isolated on the entrance to
1327	 * the free buddy page pool.
1328	 */
1329	if (TestClearPageHWPoison(page)) {
1330		pr_info("MCE: Software-unpoisoned page %#lx\n", pfn);
1331		atomic_long_sub(nr_pages, &mce_bad_pages);
1332		freeit = 1;
1333		if (PageHuge(page))
1334			clear_page_hwpoison_huge_page(page);
1335	}
1336	unlock_page(page);
1337
1338	put_page(page);
1339	if (freeit)
1340		put_page(page);
 
 
1341
1342	return 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1343}
1344EXPORT_SYMBOL(unpoison_memory);
1345
1346static struct page *new_page(struct page *p, unsigned long private, int **x)
1347{
1348	int nid = page_to_nid(p);
1349	if (PageHuge(p))
1350		return alloc_huge_page_node(page_hstate(compound_head(p)),
1351						   nid);
1352	else
1353		return alloc_pages_exact_node(nid, GFP_HIGHUSER_MOVABLE, 0);
1354}
1355
1356/*
1357 * Safely get reference count of an arbitrary page.
1358 * Returns 0 for a free page, -EIO for a zero refcount page
1359 * that is not free, and 1 for any other page type.
1360 * For 1 the page is returned with increased page count, otherwise not.
1361 */
1362static int get_any_page(struct page *p, unsigned long pfn, int flags)
1363{
1364	int ret;
1365
1366	if (flags & MF_COUNT_INCREASED)
1367		return 1;
 
 
 
1368
1369	/*
1370	 * The lock_memory_hotplug prevents a race with memory hotplug.
1371	 * This is a big hammer, a better would be nicer.
1372	 */
1373	lock_memory_hotplug();
 
 
1374
1375	/*
1376	 * Isolate the page, so that it doesn't get reallocated if it
1377	 * was free.
1378	 */
1379	set_migratetype_isolate(p);
1380	/*
1381	 * When the target page is a free hugepage, just remove it
1382	 * from free hugepage list.
1383	 */
1384	if (!get_page_unless_zero(compound_head(p))) {
1385		if (PageHuge(p)) {
1386			pr_info("get_any_page: %#lx free huge page\n", pfn);
1387			ret = dequeue_hwpoisoned_huge_page(compound_head(p));
1388		} else if (is_free_buddy_page(p)) {
1389			pr_info("get_any_page: %#lx free buddy page\n", pfn);
1390			/* Set hwpoison bit while page is still isolated */
1391			SetPageHWPoison(p);
1392			ret = 0;
1393		} else {
1394			pr_info("get_any_page: %#lx: unknown zero refcount page type %lx\n",
1395				pfn, p->flags);
1396			ret = -EIO;
1397		}
1398	} else {
1399		/* Not a free page */
1400		ret = 1;
1401	}
1402	unset_migratetype_isolate(p);
1403	unlock_memory_hotplug();
1404	return ret;
1405}
1406
1407static int soft_offline_huge_page(struct page *page, int flags)
 
 
 
 
 
1408{
1409	int ret;
1410	unsigned long pfn = page_to_pfn(page);
1411	struct page *hpage = compound_head(page);
 
 
1412	LIST_HEAD(pagelist);
 
 
 
 
1413
1414	ret = get_any_page(page, pfn, flags);
1415	if (ret < 0)
1416		return ret;
1417	if (ret == 0)
1418		goto done;
1419
1420	if (PageHWPoison(hpage)) {
1421		put_page(hpage);
1422		pr_debug("soft offline: %#lx hugepage already poisoned\n", pfn);
1423		return -EBUSY;
1424	}
1425
1426	/* Keep page count to indicate a given hugepage is isolated. */
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1427
1428	list_add(&hpage->lru, &pagelist);
1429	ret = migrate_huge_pages(&pagelist, new_page, MPOL_MF_MOVE_ALL, 0,
1430				true);
1431	if (ret) {
1432		struct page *page1, *page2;
1433		list_for_each_entry_safe(page1, page2, &pagelist, lru)
1434			put_page(page1);
 
1435
1436		pr_debug("soft offline: %#lx: migration failed %d, type %lx\n",
1437			 pfn, ret, page->flags);
1438		if (ret > 0)
1439			ret = -EIO;
1440		return ret;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1441	}
1442done:
1443	if (!PageHWPoison(hpage))
1444		atomic_long_add(1 << compound_trans_order(hpage), &mce_bad_pages);
1445	set_page_hwpoison_huge_page(hpage);
1446	dequeue_hwpoisoned_huge_page(hpage);
1447	/* keep elevated page count for bad page */
1448	return ret;
1449}
1450
1451/**
1452 * soft_offline_page - Soft offline a page.
1453 * @page: page to offline
1454 * @flags: flags. Same as memory_failure().
1455 *
1456 * Returns 0 on success, otherwise negated errno.
 
 
1457 *
1458 * Soft offline a page, by migration or invalidation,
1459 * without killing anything. This is for the case when
1460 * a page is not corrupted yet (so it's still valid to access),
1461 * but has had a number of corrected errors and is better taken
1462 * out.
1463 *
1464 * The actual policy on when to do that is maintained by
1465 * user space.
1466 *
1467 * This should never impact any application or cause data loss,
1468 * however it might take some time.
1469 *
1470 * This is not a 100% solution for all memory, but tries to be
1471 * ``good enough'' for the majority of memory.
1472 */
1473int soft_offline_page(struct page *page, int flags)
1474{
1475	int ret;
1476	unsigned long pfn = page_to_pfn(page);
1477
1478	if (PageHuge(page))
1479		return soft_offline_huge_page(page, flags);
1480
1481	ret = get_any_page(page, pfn, flags);
1482	if (ret < 0)
1483		return ret;
1484	if (ret == 0)
1485		goto done;
1486
1487	/*
1488	 * Page cache page we can handle?
1489	 */
1490	if (!PageLRU(page)) {
1491		/*
1492		 * Try to free it.
1493		 */
1494		put_page(page);
1495		shake_page(page, 1);
1496
1497		/*
1498		 * Did it turn free?
1499		 */
1500		ret = get_any_page(page, pfn, 0);
1501		if (ret < 0)
1502			return ret;
1503		if (ret == 0)
1504			goto done;
1505	}
1506	if (!PageLRU(page)) {
1507		pr_info("soft_offline: %#lx: unknown non LRU page type %lx\n",
1508				pfn, page->flags);
 
 
1509		return -EIO;
1510	}
1511
1512	lock_page(page);
1513	wait_on_page_writeback(page);
1514
1515	/*
1516	 * Synchronized using the page lock with memory_failure()
1517	 */
1518	if (PageHWPoison(page)) {
1519		unlock_page(page);
1520		put_page(page);
1521		pr_info("soft offline: %#lx page already poisoned\n", pfn);
1522		return -EBUSY;
1523	}
1524
1525	/*
1526	 * Try to invalidate first. This should work for
1527	 * non dirty unmapped page cache pages.
1528	 */
1529	ret = invalidate_inode_page(page);
1530	unlock_page(page);
1531	/*
1532	 * RED-PEN would be better to keep it isolated here, but we
1533	 * would need to fix isolation locking first.
1534	 */
1535	if (ret == 1) {
1536		put_page(page);
1537		ret = 0;
1538		pr_info("soft_offline: %#lx: invalidated\n", pfn);
1539		goto done;
1540	}
1541
1542	/*
1543	 * Simple invalidation didn't work.
1544	 * Try to migrate to a new page instead. migrate.c
1545	 * handles a large number of cases for us.
1546	 */
1547	ret = isolate_lru_page(page);
1548	/*
1549	 * Drop page reference which is came from get_any_page()
1550	 * successful isolate_lru_page() already took another one.
1551	 */
1552	put_page(page);
1553	if (!ret) {
1554		LIST_HEAD(pagelist);
1555		inc_zone_page_state(page, NR_ISOLATED_ANON +
1556					    page_is_file_cache(page));
1557		list_add(&page->lru, &pagelist);
1558		ret = migrate_pages(&pagelist, new_page, MPOL_MF_MOVE_ALL,
1559								0, true);
1560		if (ret) {
1561			putback_lru_pages(&pagelist);
1562			pr_info("soft offline: %#lx: migration failed %d, type %lx\n",
1563				pfn, ret, page->flags);
1564			if (ret > 0)
1565				ret = -EIO;
1566		}
1567	} else {
1568		pr_info("soft offline: %#lx: isolation failed: %d, page count %d, type %lx\n",
1569				pfn, ret, page_count(page), page->flags);
1570	}
1571	if (ret)
1572		return ret;
1573
1574done:
1575	atomic_long_add(1, &mce_bad_pages);
1576	SetPageHWPoison(page);
1577	/* keep elevated page count for bad page */
1578	return ret;
1579}
v6.8
   1// SPDX-License-Identifier: GPL-2.0-only
   2/*
   3 * Copyright (C) 2008, 2009 Intel Corporation
   4 * Authors: Andi Kleen, Fengguang Wu
   5 *
 
 
 
 
   6 * High level machine check handler. Handles pages reported by the
   7 * hardware as being corrupted usually due to a multi-bit ECC memory or cache
   8 * failure.
   9 *
  10 * In addition there is a "soft offline" entry point that allows stop using
  11 * not-yet-corrupted-by-suspicious pages without killing anything.
  12 *
  13 * Handles page cache pages in various states.	The tricky part
  14 * here is that we can access any page asynchronously in respect to
  15 * other VM users, because memory failures could happen anytime and
  16 * anywhere. This could violate some of their assumptions. This is why
  17 * this code has to be extremely careful. Generally it tries to use
  18 * normal locking rules, as in get the standard locks, even if that means
  19 * the error handling takes potentially a long time.
  20 *
  21 * It can be very tempting to add handling for obscure cases here.
  22 * In general any code for handling new cases should only be added iff:
  23 * - You know how to test it.
  24 * - You have a test that can be added to mce-test
  25 *   https://git.kernel.org/cgit/utils/cpu/mce/mce-test.git/
  26 * - The case actually shows up as a frequent (top 10) page state in
  27 *   tools/mm/page-types when running a real workload.
  28 *
  29 * There are several operations here with exponential complexity because
  30 * of unsuitable VM data structures. For example the operation to map back
  31 * from RMAP chains to processes has to walk the complete process list and
  32 * has non linear complexity with the number. But since memory corruptions
  33 * are rare we hope to get away with this. This avoids impacting the core
  34 * VM.
  35 */
  36
  37#define pr_fmt(fmt) "Memory failure: " fmt
  38
 
 
 
 
  39#include <linux/kernel.h>
  40#include <linux/mm.h>
  41#include <linux/page-flags.h>
  42#include <linux/sched/signal.h>
  43#include <linux/sched/task.h>
  44#include <linux/dax.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/slab.h>
  53#include <linux/swapops.h>
  54#include <linux/hugetlb.h>
  55#include <linux/memory_hotplug.h>
  56#include <linux/mm_inline.h>
  57#include <linux/memremap.h>
  58#include <linux/kfifo.h>
  59#include <linux/ratelimit.h>
  60#include <linux/pagewalk.h>
  61#include <linux/shmem_fs.h>
  62#include <linux/sysctl.h>
  63#include "swap.h"
  64#include "internal.h"
  65#include "ras/ras_event.h"
  66
  67static int sysctl_memory_failure_early_kill __read_mostly;
  68
  69static int sysctl_memory_failure_recovery __read_mostly = 1;
  70
  71atomic_long_t num_poisoned_pages __read_mostly = ATOMIC_LONG_INIT(0);
  72
  73static bool hw_memory_failure __read_mostly = false;
  74
  75static DEFINE_MUTEX(mf_mutex);
  76
  77void num_poisoned_pages_inc(unsigned long pfn)
  78{
  79	atomic_long_inc(&num_poisoned_pages);
  80	memblk_nr_poison_inc(pfn);
  81}
  82
  83void num_poisoned_pages_sub(unsigned long pfn, long i)
  84{
  85	atomic_long_sub(i, &num_poisoned_pages);
  86	if (pfn != -1UL)
  87		memblk_nr_poison_sub(pfn, i);
  88}
  89
  90/**
  91 * MF_ATTR_RO - Create sysfs entry for each memory failure statistics.
  92 * @_name: name of the file in the per NUMA sysfs directory.
  93 */
  94#define MF_ATTR_RO(_name)					\
  95static ssize_t _name##_show(struct device *dev,			\
  96			    struct device_attribute *attr,	\
  97			    char *buf)				\
  98{								\
  99	struct memory_failure_stats *mf_stats =			\
 100		&NODE_DATA(dev->id)->mf_stats;			\
 101	return sprintf(buf, "%lu\n", mf_stats->_name);		\
 102}								\
 103static DEVICE_ATTR_RO(_name)
 104
 105MF_ATTR_RO(total);
 106MF_ATTR_RO(ignored);
 107MF_ATTR_RO(failed);
 108MF_ATTR_RO(delayed);
 109MF_ATTR_RO(recovered);
 110
 111static struct attribute *memory_failure_attr[] = {
 112	&dev_attr_total.attr,
 113	&dev_attr_ignored.attr,
 114	&dev_attr_failed.attr,
 115	&dev_attr_delayed.attr,
 116	&dev_attr_recovered.attr,
 117	NULL,
 118};
 119
 120const struct attribute_group memory_failure_attr_group = {
 121	.name = "memory_failure",
 122	.attrs = memory_failure_attr,
 123};
 124
 125static struct ctl_table memory_failure_table[] = {
 126	{
 127		.procname	= "memory_failure_early_kill",
 128		.data		= &sysctl_memory_failure_early_kill,
 129		.maxlen		= sizeof(sysctl_memory_failure_early_kill),
 130		.mode		= 0644,
 131		.proc_handler	= proc_dointvec_minmax,
 132		.extra1		= SYSCTL_ZERO,
 133		.extra2		= SYSCTL_ONE,
 134	},
 135	{
 136		.procname	= "memory_failure_recovery",
 137		.data		= &sysctl_memory_failure_recovery,
 138		.maxlen		= sizeof(sysctl_memory_failure_recovery),
 139		.mode		= 0644,
 140		.proc_handler	= proc_dointvec_minmax,
 141		.extra1		= SYSCTL_ZERO,
 142		.extra2		= SYSCTL_ONE,
 143	},
 144	{ }
 145};
 146
 147/*
 148 * Return values:
 149 *   1:   the page is dissolved (if needed) and taken off from buddy,
 150 *   0:   the page is dissolved (if needed) and not taken off from buddy,
 151 *   < 0: failed to dissolve.
 152 */
 153static int __page_handle_poison(struct page *page)
 154{
 155	int ret;
 156
 157	zone_pcp_disable(page_zone(page));
 158	ret = dissolve_free_huge_page(page);
 159	if (!ret)
 160		ret = take_page_off_buddy(page);
 161	zone_pcp_enable(page_zone(page));
 162
 163	return ret;
 164}
 165
 166static bool page_handle_poison(struct page *page, bool hugepage_or_freepage, bool release)
 167{
 168	if (hugepage_or_freepage) {
 169		/*
 170		 * Doing this check for free pages is also fine since dissolve_free_huge_page
 171		 * returns 0 for non-hugetlb pages as well.
 172		 */
 173		if (__page_handle_poison(page) <= 0)
 174			/*
 175			 * We could fail to take off the target page from buddy
 176			 * for example due to racy page allocation, but that's
 177			 * acceptable because soft-offlined page is not broken
 178			 * and if someone really want to use it, they should
 179			 * take it.
 180			 */
 181			return false;
 182	}
 183
 184	SetPageHWPoison(page);
 185	if (release)
 186		put_page(page);
 187	page_ref_inc(page);
 188	num_poisoned_pages_inc(page_to_pfn(page));
 189
 190	return true;
 191}
 192
 193#if IS_ENABLED(CONFIG_HWPOISON_INJECT)
 194
 195u32 hwpoison_filter_enable = 0;
 196u32 hwpoison_filter_dev_major = ~0U;
 197u32 hwpoison_filter_dev_minor = ~0U;
 198u64 hwpoison_filter_flags_mask;
 199u64 hwpoison_filter_flags_value;
 200EXPORT_SYMBOL_GPL(hwpoison_filter_enable);
 201EXPORT_SYMBOL_GPL(hwpoison_filter_dev_major);
 202EXPORT_SYMBOL_GPL(hwpoison_filter_dev_minor);
 203EXPORT_SYMBOL_GPL(hwpoison_filter_flags_mask);
 204EXPORT_SYMBOL_GPL(hwpoison_filter_flags_value);
 205
 206static int hwpoison_filter_dev(struct page *p)
 207{
 208	struct address_space *mapping;
 209	dev_t dev;
 210
 211	if (hwpoison_filter_dev_major == ~0U &&
 212	    hwpoison_filter_dev_minor == ~0U)
 213		return 0;
 214
 
 
 
 
 
 
 215	mapping = page_mapping(p);
 216	if (mapping == NULL || mapping->host == NULL)
 217		return -EINVAL;
 218
 219	dev = mapping->host->i_sb->s_dev;
 220	if (hwpoison_filter_dev_major != ~0U &&
 221	    hwpoison_filter_dev_major != MAJOR(dev))
 222		return -EINVAL;
 223	if (hwpoison_filter_dev_minor != ~0U &&
 224	    hwpoison_filter_dev_minor != MINOR(dev))
 225		return -EINVAL;
 226
 227	return 0;
 228}
 229
 230static int hwpoison_filter_flags(struct page *p)
 231{
 232	if (!hwpoison_filter_flags_mask)
 233		return 0;
 234
 235	if ((stable_page_flags(p) & hwpoison_filter_flags_mask) ==
 236				    hwpoison_filter_flags_value)
 237		return 0;
 238	else
 239		return -EINVAL;
 240}
 241
 242/*
 243 * This allows stress tests to limit test scope to a collection of tasks
 244 * by putting them under some memcg. This prevents killing unrelated/important
 245 * processes such as /sbin/init. Note that the target task may share clean
 246 * pages with init (eg. libc text), which is harmless. If the target task
 247 * share _dirty_ pages with another task B, the test scheme must make sure B
 248 * is also included in the memcg. At last, due to race conditions this filter
 249 * can only guarantee that the page either belongs to the memcg tasks, or is
 250 * a freed page.
 251 */
 252#ifdef CONFIG_MEMCG
 253u64 hwpoison_filter_memcg;
 254EXPORT_SYMBOL_GPL(hwpoison_filter_memcg);
 255static int hwpoison_filter_task(struct page *p)
 256{
 
 
 
 
 257	if (!hwpoison_filter_memcg)
 258		return 0;
 259
 260	if (page_cgroup_ino(p) != hwpoison_filter_memcg)
 
 
 
 
 
 
 
 
 
 
 
 
 261		return -EINVAL;
 262
 263	return 0;
 264}
 265#else
 266static int hwpoison_filter_task(struct page *p) { return 0; }
 267#endif
 268
 269int hwpoison_filter(struct page *p)
 270{
 271	if (!hwpoison_filter_enable)
 272		return 0;
 273
 274	if (hwpoison_filter_dev(p))
 275		return -EINVAL;
 276
 277	if (hwpoison_filter_flags(p))
 278		return -EINVAL;
 279
 280	if (hwpoison_filter_task(p))
 281		return -EINVAL;
 282
 283	return 0;
 284}
 285#else
 286int hwpoison_filter(struct page *p)
 287{
 288	return 0;
 289}
 290#endif
 291
 292EXPORT_SYMBOL_GPL(hwpoison_filter);
 293
 294/*
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 295 * Kill all processes that have a poisoned page mapped and then isolate
 296 * the page.
 297 *
 298 * General strategy:
 299 * Find all processes having the page mapped and kill them.
 300 * But we keep a page reference around so that the page is not
 301 * actually freed yet.
 302 * Then stash the page away
 303 *
 304 * There's no convenient way to get back to mapped processes
 305 * from the VMAs. So do a brute-force search over all
 306 * running processes.
 307 *
 308 * Remember that machine checks are not common (or rather
 309 * if they are common you have other problems), so this shouldn't
 310 * be a performance issue.
 311 *
 312 * Also there are some races possible while we get from the
 313 * error detection to actually handle it.
 314 */
 315
 316struct to_kill {
 317	struct list_head nd;
 318	struct task_struct *tsk;
 319	unsigned long addr;
 320	short size_shift;
 321};
 322
 323/*
 324 * Send all the processes who have the page mapped a signal.
 325 * ``action optional'' if they are not immediately affected by the error
 326 * ``action required'' if error happened in current execution context
 327 */
 328static int kill_proc(struct to_kill *tk, unsigned long pfn, int flags)
 329{
 330	struct task_struct *t = tk->tsk;
 331	short addr_lsb = tk->size_shift;
 332	int ret = 0;
 333
 334	pr_err("%#lx: Sending SIGBUS to %s:%d due to hardware memory corruption\n",
 335			pfn, t->comm, t->pid);
 336
 337	if ((flags & MF_ACTION_REQUIRED) && (t == current))
 338		ret = force_sig_mceerr(BUS_MCEERR_AR,
 339				 (void __user *)tk->addr, addr_lsb);
 340	else
 341		/*
 342		 * Signal other processes sharing the page if they have
 343		 * PF_MCE_EARLY set.
 344		 * Don't use force here, it's convenient if the signal
 345		 * can be temporarily blocked.
 346		 * This could cause a loop when the user sets SIGBUS
 347		 * to SIG_IGN, but hopefully no one will do that?
 348		 */
 349		ret = send_sig_mceerr(BUS_MCEERR_AO, (void __user *)tk->addr,
 350				      addr_lsb, t);
 351	if (ret < 0)
 352		pr_info("Error sending signal to %s:%d: %d\n",
 353			t->comm, t->pid, ret);
 354	return ret;
 355}
 356
 357/*
 358 * Unknown page type encountered. Try to check whether it can turn PageLRU by
 359 * lru_add_drain_all.
 360 */
 361void shake_page(struct page *p)
 362{
 363	if (PageHuge(p))
 364		return;
 365	/*
 366	 * TODO: Could shrink slab caches here if a lightweight range-based
 367	 * shrinker will be available.
 368	 */
 369	if (PageSlab(p))
 370		return;
 371
 372	lru_add_drain_all();
 373}
 374EXPORT_SYMBOL_GPL(shake_page);
 375
 376static unsigned long dev_pagemap_mapping_shift(struct vm_area_struct *vma,
 377		unsigned long address)
 378{
 379	unsigned long ret = 0;
 380	pgd_t *pgd;
 381	p4d_t *p4d;
 382	pud_t *pud;
 383	pmd_t *pmd;
 384	pte_t *pte;
 385	pte_t ptent;
 386
 387	VM_BUG_ON_VMA(address == -EFAULT, vma);
 388	pgd = pgd_offset(vma->vm_mm, address);
 389	if (!pgd_present(*pgd))
 390		return 0;
 391	p4d = p4d_offset(pgd, address);
 392	if (!p4d_present(*p4d))
 393		return 0;
 394	pud = pud_offset(p4d, address);
 395	if (!pud_present(*pud))
 396		return 0;
 397	if (pud_devmap(*pud))
 398		return PUD_SHIFT;
 399	pmd = pmd_offset(pud, address);
 400	if (!pmd_present(*pmd))
 401		return 0;
 402	if (pmd_devmap(*pmd))
 403		return PMD_SHIFT;
 404	pte = pte_offset_map(pmd, address);
 405	if (!pte)
 406		return 0;
 407	ptent = ptep_get(pte);
 408	if (pte_present(ptent) && pte_devmap(ptent))
 409		ret = PAGE_SHIFT;
 410	pte_unmap(pte);
 411	return ret;
 412}
 413
 414/*
 415 * Failure handling: if we can't find or can't kill a process there's
 416 * not much we can do.	We just print a message and ignore otherwise.
 417 */
 418
 419#define FSDAX_INVALID_PGOFF ULONG_MAX
 420
 421/*
 422 * Schedule a process for later kill.
 423 * Uses GFP_ATOMIC allocations to avoid potential recursions in the VM.
 424 *
 425 * Note: @fsdax_pgoff is used only when @p is a fsdax page and a
 426 * filesystem with a memory failure handler has claimed the
 427 * memory_failure event. In all other cases, page->index and
 428 * page->mapping are sufficient for mapping the page back to its
 429 * corresponding user virtual address.
 430 */
 431static void __add_to_kill(struct task_struct *tsk, struct page *p,
 432			  struct vm_area_struct *vma, struct list_head *to_kill,
 433			  unsigned long ksm_addr, pgoff_t fsdax_pgoff)
 434{
 435	struct to_kill *tk;
 436
 437	tk = kmalloc(sizeof(struct to_kill), GFP_ATOMIC);
 438	if (!tk) {
 439		pr_err("Out of memory while machine check handling\n");
 440		return;
 
 
 
 
 
 
 441	}
 
 
 442
 443	tk->addr = ksm_addr ? ksm_addr : page_address_in_vma(p, vma);
 444	if (is_zone_device_page(p)) {
 445		if (fsdax_pgoff != FSDAX_INVALID_PGOFF)
 446			tk->addr = vma_pgoff_address(fsdax_pgoff, 1, vma);
 447		tk->size_shift = dev_pagemap_mapping_shift(vma, tk->addr);
 448	} else
 449		tk->size_shift = page_shift(compound_head(p));
 450
 451	/*
 452	 * Send SIGKILL if "tk->addr == -EFAULT". Also, as
 453	 * "tk->size_shift" is always non-zero for !is_zone_device_page(),
 454	 * so "tk->size_shift == 0" effectively checks no mapping on
 455	 * ZONE_DEVICE. Indeed, when a devdax page is mmapped N times
 456	 * to a process' address space, it's possible not all N VMAs
 457	 * contain mappings for the page, but at least one VMA does.
 458	 * Only deliver SIGBUS with payload derived from the VMA that
 459	 * has a mapping for the page.
 460	 */
 461	if (tk->addr == -EFAULT) {
 462		pr_info("Unable to find user space address %lx in %s\n",
 463			page_to_pfn(p), tsk->comm);
 464	} else if (tk->size_shift == 0) {
 465		kfree(tk);
 466		return;
 467	}
 468
 469	get_task_struct(tsk);
 470	tk->tsk = tsk;
 471	list_add_tail(&tk->nd, to_kill);
 472}
 473
 474static void add_to_kill_anon_file(struct task_struct *tsk, struct page *p,
 475				  struct vm_area_struct *vma,
 476				  struct list_head *to_kill)
 477{
 478	__add_to_kill(tsk, p, vma, to_kill, 0, FSDAX_INVALID_PGOFF);
 479}
 480
 481#ifdef CONFIG_KSM
 482static bool task_in_to_kill_list(struct list_head *to_kill,
 483				 struct task_struct *tsk)
 484{
 485	struct to_kill *tk, *next;
 486
 487	list_for_each_entry_safe(tk, next, to_kill, nd) {
 488		if (tk->tsk == tsk)
 489			return true;
 490	}
 491
 492	return false;
 493}
 494void add_to_kill_ksm(struct task_struct *tsk, struct page *p,
 495		     struct vm_area_struct *vma, struct list_head *to_kill,
 496		     unsigned long ksm_addr)
 497{
 498	if (!task_in_to_kill_list(to_kill, tsk))
 499		__add_to_kill(tsk, p, vma, to_kill, ksm_addr, FSDAX_INVALID_PGOFF);
 500}
 501#endif
 502/*
 503 * Kill the processes that have been collected earlier.
 504 *
 505 * Only do anything when FORCEKILL is set, otherwise just free the
 506 * list (this is used for clean pages which do not need killing)
 507 * Also when FAIL is set do a force kill because something went
 508 * wrong earlier.
 509 */
 510static void kill_procs(struct list_head *to_kill, int forcekill, bool fail,
 511		unsigned long pfn, int flags)
 512{
 513	struct to_kill *tk, *next;
 514
 515	list_for_each_entry_safe(tk, next, to_kill, nd) {
 516		if (forcekill) {
 517			/*
 518			 * In case something went wrong with munmapping
 519			 * make sure the process doesn't catch the
 520			 * signal and then access the memory. Just kill it.
 521			 */
 522			if (fail || tk->addr == -EFAULT) {
 523				pr_err("%#lx: forcibly killing %s:%d because of failure to unmap corrupted page\n",
 524				       pfn, tk->tsk->comm, tk->tsk->pid);
 525				do_send_sig_info(SIGKILL, SEND_SIG_PRIV,
 526						 tk->tsk, PIDTYPE_PID);
 527			}
 528
 529			/*
 530			 * In theory the process could have mapped
 531			 * something else on the address in-between. We could
 532			 * check for that, but we need to tell the
 533			 * process anyways.
 534			 */
 535			else if (kill_proc(tk, pfn, flags) < 0)
 536				pr_err("%#lx: Cannot send advisory machine check signal to %s:%d\n",
 537				       pfn, tk->tsk->comm, tk->tsk->pid);
 
 
 538		}
 539		list_del(&tk->nd);
 540		put_task_struct(tk->tsk);
 541		kfree(tk);
 542	}
 543}
 544
 545/*
 546 * Find a dedicated thread which is supposed to handle SIGBUS(BUS_MCEERR_AO)
 547 * on behalf of the thread group. Return task_struct of the (first found)
 548 * dedicated thread if found, and return NULL otherwise.
 549 *
 550 * We already hold rcu lock in the caller, so we don't have to call
 551 * rcu_read_lock/unlock() in this function.
 552 */
 553static struct task_struct *find_early_kill_thread(struct task_struct *tsk)
 554{
 555	struct task_struct *t;
 556
 557	for_each_thread(tsk, t) {
 558		if (t->flags & PF_MCE_PROCESS) {
 559			if (t->flags & PF_MCE_EARLY)
 560				return t;
 561		} else {
 562			if (sysctl_memory_failure_early_kill)
 563				return t;
 564		}
 565	}
 566	return NULL;
 567}
 568
 569/*
 570 * Determine whether a given process is "early kill" process which expects
 571 * to be signaled when some page under the process is hwpoisoned.
 572 * Return task_struct of the dedicated thread (main thread unless explicitly
 573 * specified) if the process is "early kill" and otherwise returns NULL.
 574 *
 575 * Note that the above is true for Action Optional case. For Action Required
 576 * case, it's only meaningful to the current thread which need to be signaled
 577 * with SIGBUS, this error is Action Optional for other non current
 578 * processes sharing the same error page,if the process is "early kill", the
 579 * task_struct of the dedicated thread will also be returned.
 580 */
 581struct task_struct *task_early_kill(struct task_struct *tsk, int force_early)
 582{
 583	if (!tsk->mm)
 584		return NULL;
 585	/*
 586	 * Comparing ->mm here because current task might represent
 587	 * a subthread, while tsk always points to the main thread.
 588	 */
 589	if (force_early && tsk->mm == current->mm)
 590		return current;
 591
 592	return find_early_kill_thread(tsk);
 593}
 594
 595/*
 596 * Collect processes when the error hit an anonymous page.
 597 */
 598static void collect_procs_anon(struct folio *folio, struct page *page,
 599		struct list_head *to_kill, int force_early)
 600{
 601	struct vm_area_struct *vma;
 602	struct task_struct *tsk;
 603	struct anon_vma *av;
 604	pgoff_t pgoff;
 605
 606	av = folio_lock_anon_vma_read(folio, NULL);
 607	if (av == NULL)	/* Not actually mapped anymore */
 608		return;
 609
 610	pgoff = page_to_pgoff(page);
 611	rcu_read_lock();
 612	for_each_process(tsk) {
 613		struct anon_vma_chain *vmac;
 614		struct task_struct *t = task_early_kill(tsk, force_early);
 615
 616		if (!t)
 617			continue;
 618		anon_vma_interval_tree_foreach(vmac, &av->rb_root,
 619					       pgoff, pgoff) {
 620			vma = vmac->vma;
 621			if (vma->vm_mm != t->mm)
 622				continue;
 623			if (!page_mapped_in_vma(page, vma))
 624				continue;
 625			add_to_kill_anon_file(t, page, vma, to_kill);
 
 626		}
 627	}
 628	rcu_read_unlock();
 629	anon_vma_unlock_read(av);
 630}
 631
 632/*
 633 * Collect processes when the error hit a file mapped page.
 634 */
 635static void collect_procs_file(struct folio *folio, struct page *page,
 636		struct list_head *to_kill, int force_early)
 637{
 638	struct vm_area_struct *vma;
 639	struct task_struct *tsk;
 640	struct address_space *mapping = folio->mapping;
 641	pgoff_t pgoff;
 642
 643	i_mmap_lock_read(mapping);
 644	rcu_read_lock();
 645	pgoff = page_to_pgoff(page);
 646	for_each_process(tsk) {
 647		struct task_struct *t = task_early_kill(tsk, force_early);
 648
 649		if (!t)
 650			continue;
 651		vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff,
 
 652				      pgoff) {
 653			/*
 654			 * Send early kill signal to tasks where a vma covers
 655			 * the page but the corrupted page is not necessarily
 656			 * mapped in its pte.
 657			 * Assume applications who requested early kill want
 658			 * to be informed of all such data corruptions.
 659			 */
 660			if (vma->vm_mm == t->mm)
 661				add_to_kill_anon_file(t, page, vma, to_kill);
 662		}
 663	}
 664	rcu_read_unlock();
 665	i_mmap_unlock_read(mapping);
 666}
 667
 668#ifdef CONFIG_FS_DAX
 669static void add_to_kill_fsdax(struct task_struct *tsk, struct page *p,
 670			      struct vm_area_struct *vma,
 671			      struct list_head *to_kill, pgoff_t pgoff)
 672{
 673	__add_to_kill(tsk, p, vma, to_kill, 0, pgoff);
 674}
 675
 676/*
 677 * Collect processes when the error hit a fsdax page.
 
 
 
 678 */
 679static void collect_procs_fsdax(struct page *page,
 680		struct address_space *mapping, pgoff_t pgoff,
 681		struct list_head *to_kill, bool pre_remove)
 682{
 683	struct vm_area_struct *vma;
 684	struct task_struct *tsk;
 685
 686	i_mmap_lock_read(mapping);
 687	rcu_read_lock();
 688	for_each_process(tsk) {
 689		struct task_struct *t = tsk;
 690
 691		/*
 692		 * Search for all tasks while MF_MEM_PRE_REMOVE is set, because
 693		 * the current may not be the one accessing the fsdax page.
 694		 * Otherwise, search for the current task.
 695		 */
 696		if (!pre_remove)
 697			t = task_early_kill(tsk, true);
 698		if (!t)
 699			continue;
 700		vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
 701			if (vma->vm_mm == t->mm)
 702				add_to_kill_fsdax(t, page, vma, to_kill, pgoff);
 703		}
 704	}
 705	rcu_read_unlock();
 706	i_mmap_unlock_read(mapping);
 707}
 708#endif /* CONFIG_FS_DAX */
 709
 710/*
 711 * Collect the processes who have the corrupted page mapped to kill.
 712 */
 713static void collect_procs(struct folio *folio, struct page *page,
 714		struct list_head *tokill, int force_early)
 715{
 716	if (!folio->mapping)
 717		return;
 718	if (unlikely(PageKsm(page)))
 719		collect_procs_ksm(page, tokill, force_early);
 720	else if (PageAnon(page))
 721		collect_procs_anon(folio, page, tokill, force_early);
 722	else
 723		collect_procs_file(folio, page, tokill, force_early);
 724}
 725
 726struct hwpoison_walk {
 727	struct to_kill tk;
 728	unsigned long pfn;
 729	int flags;
 730};
 731
 732static void set_to_kill(struct to_kill *tk, unsigned long addr, short shift)
 733{
 734	tk->addr = addr;
 735	tk->size_shift = shift;
 736}
 737
 738static int check_hwpoisoned_entry(pte_t pte, unsigned long addr, short shift,
 739				unsigned long poisoned_pfn, struct to_kill *tk)
 740{
 741	unsigned long pfn = 0;
 742
 743	if (pte_present(pte)) {
 744		pfn = pte_pfn(pte);
 745	} else {
 746		swp_entry_t swp = pte_to_swp_entry(pte);
 747
 748		if (is_hwpoison_entry(swp))
 749			pfn = swp_offset_pfn(swp);
 750	}
 751
 752	if (!pfn || pfn != poisoned_pfn)
 753		return 0;
 754
 755	set_to_kill(tk, addr, shift);
 756	return 1;
 757}
 758
 759#ifdef CONFIG_TRANSPARENT_HUGEPAGE
 760static int check_hwpoisoned_pmd_entry(pmd_t *pmdp, unsigned long addr,
 761				      struct hwpoison_walk *hwp)
 762{
 763	pmd_t pmd = *pmdp;
 764	unsigned long pfn;
 765	unsigned long hwpoison_vaddr;
 766
 767	if (!pmd_present(pmd))
 768		return 0;
 769	pfn = pmd_pfn(pmd);
 770	if (pfn <= hwp->pfn && hwp->pfn < pfn + HPAGE_PMD_NR) {
 771		hwpoison_vaddr = addr + ((hwp->pfn - pfn) << PAGE_SHIFT);
 772		set_to_kill(&hwp->tk, hwpoison_vaddr, PAGE_SHIFT);
 773		return 1;
 774	}
 775	return 0;
 776}
 777#else
 778static int check_hwpoisoned_pmd_entry(pmd_t *pmdp, unsigned long addr,
 779				      struct hwpoison_walk *hwp)
 780{
 781	return 0;
 782}
 783#endif
 784
 785static int hwpoison_pte_range(pmd_t *pmdp, unsigned long addr,
 786			      unsigned long end, struct mm_walk *walk)
 787{
 788	struct hwpoison_walk *hwp = walk->private;
 789	int ret = 0;
 790	pte_t *ptep, *mapped_pte;
 791	spinlock_t *ptl;
 792
 793	ptl = pmd_trans_huge_lock(pmdp, walk->vma);
 794	if (ptl) {
 795		ret = check_hwpoisoned_pmd_entry(pmdp, addr, hwp);
 796		spin_unlock(ptl);
 797		goto out;
 798	}
 799
 800	mapped_pte = ptep = pte_offset_map_lock(walk->vma->vm_mm, pmdp,
 801						addr, &ptl);
 802	if (!ptep)
 803		goto out;
 804
 805	for (; addr != end; ptep++, addr += PAGE_SIZE) {
 806		ret = check_hwpoisoned_entry(ptep_get(ptep), addr, PAGE_SHIFT,
 807					     hwp->pfn, &hwp->tk);
 808		if (ret == 1)
 809			break;
 810	}
 811	pte_unmap_unlock(mapped_pte, ptl);
 812out:
 813	cond_resched();
 814	return ret;
 815}
 816
 817#ifdef CONFIG_HUGETLB_PAGE
 818static int hwpoison_hugetlb_range(pte_t *ptep, unsigned long hmask,
 819			    unsigned long addr, unsigned long end,
 820			    struct mm_walk *walk)
 821{
 822	struct hwpoison_walk *hwp = walk->private;
 823	pte_t pte = huge_ptep_get(ptep);
 824	struct hstate *h = hstate_vma(walk->vma);
 825
 826	return check_hwpoisoned_entry(pte, addr, huge_page_shift(h),
 827				      hwp->pfn, &hwp->tk);
 828}
 829#else
 830#define hwpoison_hugetlb_range	NULL
 831#endif
 832
 833static const struct mm_walk_ops hwpoison_walk_ops = {
 834	.pmd_entry = hwpoison_pte_range,
 835	.hugetlb_entry = hwpoison_hugetlb_range,
 836	.walk_lock = PGWALK_RDLOCK,
 837};
 838
 839/*
 840 * Sends SIGBUS to the current process with error info.
 841 *
 842 * This function is intended to handle "Action Required" MCEs on already
 843 * hardware poisoned pages. They could happen, for example, when
 844 * memory_failure() failed to unmap the error page at the first call, or
 845 * when multiple local machine checks happened on different CPUs.
 846 *
 847 * MCE handler currently has no easy access to the error virtual address,
 848 * so this function walks page table to find it. The returned virtual address
 849 * is proper in most cases, but it could be wrong when the application
 850 * process has multiple entries mapping the error page.
 851 */
 852static int kill_accessing_process(struct task_struct *p, unsigned long pfn,
 853				  int flags)
 854{
 855	int ret;
 856	struct hwpoison_walk priv = {
 857		.pfn = pfn,
 858	};
 859	priv.tk.tsk = p;
 860
 861	if (!p->mm)
 862		return -EFAULT;
 863
 864	mmap_read_lock(p->mm);
 865	ret = walk_page_range(p->mm, 0, TASK_SIZE, &hwpoison_walk_ops,
 866			      (void *)&priv);
 867	if (ret == 1 && priv.tk.addr)
 868		kill_proc(&priv.tk, pfn, flags);
 869	else
 870		ret = 0;
 871	mmap_read_unlock(p->mm);
 872	return ret > 0 ? -EHWPOISON : -EFAULT;
 873}
 874
 875static const char *action_name[] = {
 876	[MF_IGNORED] = "Ignored",
 877	[MF_FAILED] = "Failed",
 878	[MF_DELAYED] = "Delayed",
 879	[MF_RECOVERED] = "Recovered",
 880};
 881
 882static const char * const action_page_types[] = {
 883	[MF_MSG_KERNEL]			= "reserved kernel page",
 884	[MF_MSG_KERNEL_HIGH_ORDER]	= "high-order kernel page",
 885	[MF_MSG_SLAB]			= "kernel slab page",
 886	[MF_MSG_DIFFERENT_COMPOUND]	= "different compound page after locking",
 887	[MF_MSG_HUGE]			= "huge page",
 888	[MF_MSG_FREE_HUGE]		= "free huge page",
 889	[MF_MSG_UNMAP_FAILED]		= "unmapping failed page",
 890	[MF_MSG_DIRTY_SWAPCACHE]	= "dirty swapcache page",
 891	[MF_MSG_CLEAN_SWAPCACHE]	= "clean swapcache page",
 892	[MF_MSG_DIRTY_MLOCKED_LRU]	= "dirty mlocked LRU page",
 893	[MF_MSG_CLEAN_MLOCKED_LRU]	= "clean mlocked LRU page",
 894	[MF_MSG_DIRTY_UNEVICTABLE_LRU]	= "dirty unevictable LRU page",
 895	[MF_MSG_CLEAN_UNEVICTABLE_LRU]	= "clean unevictable LRU page",
 896	[MF_MSG_DIRTY_LRU]		= "dirty LRU page",
 897	[MF_MSG_CLEAN_LRU]		= "clean LRU page",
 898	[MF_MSG_TRUNCATED_LRU]		= "already truncated LRU page",
 899	[MF_MSG_BUDDY]			= "free buddy page",
 900	[MF_MSG_DAX]			= "dax page",
 901	[MF_MSG_UNSPLIT_THP]		= "unsplit thp",
 902	[MF_MSG_UNKNOWN]		= "unknown page",
 903};
 904
 905/*
 906 * XXX: It is possible that a page is isolated from LRU cache,
 907 * and then kept in swap cache or failed to remove from page cache.
 908 * The page count will stop it from being freed by unpoison.
 909 * Stress tests should be aware of this memory leak problem.
 910 */
 911static int delete_from_lru_cache(struct folio *folio)
 912{
 913	if (folio_isolate_lru(folio)) {
 914		/*
 915		 * Clear sensible page flags, so that the buddy system won't
 916		 * complain when the folio is unpoison-and-freed.
 917		 */
 918		folio_clear_active(folio);
 919		folio_clear_unevictable(folio);
 920
 921		/*
 922		 * Poisoned page might never drop its ref count to 0 so we have
 923		 * to uncharge it manually from its memcg.
 924		 */
 925		mem_cgroup_uncharge(folio);
 926
 927		/*
 928		 * drop the refcount elevated by folio_isolate_lru()
 929		 */
 930		folio_put(folio);
 931		return 0;
 932	}
 933	return -EIO;
 934}
 935
 936static int truncate_error_folio(struct folio *folio, unsigned long pfn,
 937				struct address_space *mapping)
 938{
 939	int ret = MF_FAILED;
 940
 941	if (mapping->a_ops->error_remove_folio) {
 942		int err = mapping->a_ops->error_remove_folio(mapping, folio);
 943
 944		if (err != 0)
 945			pr_info("%#lx: Failed to punch page: %d\n", pfn, err);
 946		else if (!filemap_release_folio(folio, GFP_NOIO))
 947			pr_info("%#lx: failed to release buffers\n", pfn);
 948		else
 949			ret = MF_RECOVERED;
 950	} else {
 951		/*
 952		 * If the file system doesn't support it just invalidate
 953		 * This fails on dirty or anything with private pages
 954		 */
 955		if (mapping_evict_folio(mapping, folio))
 956			ret = MF_RECOVERED;
 957		else
 958			pr_info("%#lx: Failed to invalidate\n",	pfn);
 959	}
 960
 961	return ret;
 962}
 963
 964struct page_state {
 965	unsigned long mask;
 966	unsigned long res;
 967	enum mf_action_page_type type;
 968
 969	/* Callback ->action() has to unlock the relevant page inside it. */
 970	int (*action)(struct page_state *ps, struct page *p);
 971};
 972
 973/*
 974 * Return true if page is still referenced by others, otherwise return
 975 * false.
 976 *
 977 * The extra_pins is true when one extra refcount is expected.
 978 */
 979static bool has_extra_refcount(struct page_state *ps, struct page *p,
 980			       bool extra_pins)
 981{
 982	int count = page_count(p) - 1;
 983
 984	if (extra_pins)
 985		count -= folio_nr_pages(page_folio(p));
 986
 987	if (count > 0) {
 988		pr_err("%#lx: %s still referenced by %d users\n",
 989		       page_to_pfn(p), action_page_types[ps->type], count);
 990		return true;
 991	}
 992
 993	return false;
 994}
 995
 996/*
 997 * Error hit kernel page.
 998 * Do nothing, try to be lucky and not touch this instead. For a few cases we
 999 * could be more sophisticated.
1000 */
1001static int me_kernel(struct page_state *ps, struct page *p)
1002{
1003	unlock_page(p);
1004	return MF_IGNORED;
1005}
1006
1007/*
1008 * Page in unknown state. Do nothing.
1009 */
1010static int me_unknown(struct page_state *ps, struct page *p)
1011{
1012	pr_err("%#lx: Unknown page state\n", page_to_pfn(p));
1013	unlock_page(p);
1014	return MF_FAILED;
1015}
1016
1017/*
1018 * Clean (or cleaned) page cache page.
1019 */
1020static int me_pagecache_clean(struct page_state *ps, struct page *p)
1021{
1022	struct folio *folio = page_folio(p);
1023	int ret;
1024	struct address_space *mapping;
1025	bool extra_pins;
1026
1027	delete_from_lru_cache(folio);
1028
1029	/*
1030	 * For anonymous folios the only reference left
1031	 * should be the one m_f() holds.
1032	 */
1033	if (folio_test_anon(folio)) {
1034		ret = MF_RECOVERED;
1035		goto out;
1036	}
1037
1038	/*
1039	 * Now truncate the page in the page cache. This is really
1040	 * more like a "temporary hole punch"
1041	 * Don't do this for block devices when someone else
1042	 * has a reference, because it could be file system metadata
1043	 * and that's not safe to truncate.
1044	 */
1045	mapping = folio_mapping(folio);
1046	if (!mapping) {
1047		/* Folio has been torn down in the meantime */
1048		ret = MF_FAILED;
1049		goto out;
 
1050	}
1051
1052	/*
1053	 * The shmem page is kept in page cache instead of truncating
1054	 * so is expected to have an extra refcount after error-handling.
1055	 */
1056	extra_pins = shmem_mapping(mapping);
1057
1058	/*
1059	 * Truncation is a bit tricky. Enable it per file system for now.
1060	 *
1061	 * Open: to take i_rwsem or not for this? Right now we don't.
1062	 */
1063	ret = truncate_error_folio(folio, page_to_pfn(p), mapping);
1064	if (has_extra_refcount(ps, p, extra_pins))
1065		ret = MF_FAILED;
1066
1067out:
1068	folio_unlock(folio);
1069
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1070	return ret;
1071}
1072
1073/*
1074 * Dirty pagecache page
1075 * Issues: when the error hit a hole page the error is not properly
1076 * propagated.
1077 */
1078static int me_pagecache_dirty(struct page_state *ps, struct page *p)
1079{
1080	struct address_space *mapping = page_mapping(p);
1081
1082	SetPageError(p);
1083	/* TBD: print more information about the file. */
1084	if (mapping) {
1085		/*
1086		 * IO error will be reported by write(), fsync(), etc.
1087		 * who check the mapping.
1088		 * This way the application knows that something went
1089		 * wrong with its dirty file data.
1090		 *
1091		 * There's one open issue:
1092		 *
1093		 * The EIO will be only reported on the next IO
1094		 * operation and then cleared through the IO map.
1095		 * Normally Linux has two mechanisms to pass IO error
1096		 * first through the AS_EIO flag in the address space
1097		 * and then through the PageError flag in the page.
1098		 * Since we drop pages on memory failure handling the
1099		 * only mechanism open to use is through AS_AIO.
1100		 *
1101		 * This has the disadvantage that it gets cleared on
1102		 * the first operation that returns an error, while
1103		 * the PageError bit is more sticky and only cleared
1104		 * when the page is reread or dropped.  If an
1105		 * application assumes it will always get error on
1106		 * fsync, but does other operations on the fd before
1107		 * and the page is dropped between then the error
1108		 * will not be properly reported.
1109		 *
1110		 * This can already happen even without hwpoisoned
1111		 * pages: first on metadata IO errors (which only
1112		 * report through AS_EIO) or when the page is dropped
1113		 * at the wrong time.
1114		 *
1115		 * So right now we assume that the application DTRT on
1116		 * the first EIO, but we're not worse than other parts
1117		 * of the kernel.
1118		 */
1119		mapping_set_error(mapping, -EIO);
1120	}
1121
1122	return me_pagecache_clean(ps, p);
1123}
1124
1125/*
1126 * Clean and dirty swap cache.
1127 *
1128 * Dirty swap cache page is tricky to handle. The page could live both in page
1129 * cache and swap cache(ie. page is freshly swapped in). So it could be
1130 * referenced concurrently by 2 types of PTEs:
1131 * normal PTEs and swap PTEs. We try to handle them consistently by calling
1132 * try_to_unmap(!TTU_HWPOISON) to convert the normal PTEs to swap PTEs,
1133 * and then
1134 *      - clear dirty bit to prevent IO
1135 *      - remove from LRU
1136 *      - but keep in the swap cache, so that when we return to it on
1137 *        a later page fault, we know the application is accessing
1138 *        corrupted data and shall be killed (we installed simple
1139 *        interception code in do_swap_page to catch it).
1140 *
1141 * Clean swap cache pages can be directly isolated. A later page fault will
1142 * bring in the known good data from disk.
1143 */
1144static int me_swapcache_dirty(struct page_state *ps, struct page *p)
1145{
1146	struct folio *folio = page_folio(p);
1147	int ret;
1148	bool extra_pins = false;
1149
1150	folio_clear_dirty(folio);
1151	/* Trigger EIO in shmem: */
1152	folio_clear_uptodate(folio);
1153
1154	ret = delete_from_lru_cache(folio) ? MF_FAILED : MF_DELAYED;
1155	folio_unlock(folio);
1156
1157	if (ret == MF_DELAYED)
1158		extra_pins = true;
1159
1160	if (has_extra_refcount(ps, p, extra_pins))
1161		ret = MF_FAILED;
1162
1163	return ret;
1164}
1165
1166static int me_swapcache_clean(struct page_state *ps, struct page *p)
1167{
1168	struct folio *folio = page_folio(p);
1169	int ret;
1170
1171	delete_from_swap_cache(folio);
1172
1173	ret = delete_from_lru_cache(folio) ? MF_FAILED : MF_RECOVERED;
1174	folio_unlock(folio);
1175
1176	if (has_extra_refcount(ps, p, false))
1177		ret = MF_FAILED;
1178
1179	return ret;
1180}
1181
1182/*
1183 * Huge pages. Needs work.
1184 * Issues:
1185 * - Error on hugepage is contained in hugepage unit (not in raw page unit.)
1186 *   To narrow down kill region to one page, we need to break up pmd.
1187 */
1188static int me_huge_page(struct page_state *ps, struct page *p)
1189{
1190	struct folio *folio = page_folio(p);
1191	int res;
1192	struct address_space *mapping;
1193	bool extra_pins = false;
1194
1195	mapping = folio_mapping(folio);
1196	if (mapping) {
1197		res = truncate_error_folio(folio, page_to_pfn(p), mapping);
1198		/* The page is kept in page cache. */
1199		extra_pins = true;
1200		folio_unlock(folio);
1201	} else {
1202		folio_unlock(folio);
1203		/*
1204		 * migration entry prevents later access on error hugepage,
1205		 * so we can free and dissolve it into buddy to save healthy
1206		 * subpages.
1207		 */
1208		folio_put(folio);
1209		if (__page_handle_poison(p) >= 0) {
1210			page_ref_inc(p);
1211			res = MF_RECOVERED;
1212		} else {
1213			res = MF_FAILED;
1214		}
1215	}
1216
1217	if (has_extra_refcount(ps, p, extra_pins))
1218		res = MF_FAILED;
1219
1220	return res;
1221}
1222
1223/*
1224 * Various page states we can handle.
1225 *
1226 * A page state is defined by its current page->flags bits.
1227 * The table matches them in order and calls the right handler.
1228 *
1229 * This is quite tricky because we can access page at any time
1230 * in its live cycle, so all accesses have to be extremely careful.
1231 *
1232 * This is not complete. More states could be added.
1233 * For any missing state don't attempt recovery.
1234 */
1235
1236#define dirty		(1UL << PG_dirty)
1237#define sc		((1UL << PG_swapcache) | (1UL << PG_swapbacked))
1238#define unevict		(1UL << PG_unevictable)
1239#define mlock		(1UL << PG_mlocked)
 
1240#define lru		(1UL << PG_lru)
 
1241#define head		(1UL << PG_head)
 
 
1242#define slab		(1UL << PG_slab)
1243#define reserved	(1UL << PG_reserved)
1244
1245static struct page_state error_states[] = {
1246	{ reserved,	reserved,	MF_MSG_KERNEL,	me_kernel },
 
 
 
 
 
1247	/*
1248	 * free pages are specially detected outside this table:
1249	 * PG_buddy pages only make a small fraction of all free pages.
1250	 */
1251
1252	/*
1253	 * Could in theory check if slab page is free or if we can drop
1254	 * currently unused objects without touching them. But just
1255	 * treat it as standard kernel for now.
1256	 */
1257	{ slab,		slab,		MF_MSG_SLAB,	me_kernel },
1258
1259	{ head,		head,		MF_MSG_HUGE,		me_huge_page },
 
 
 
 
 
1260
1261	{ sc|dirty,	sc|dirty,	MF_MSG_DIRTY_SWAPCACHE,	me_swapcache_dirty },
1262	{ sc|dirty,	sc,		MF_MSG_CLEAN_SWAPCACHE,	me_swapcache_clean },
1263
1264	{ mlock|dirty,	mlock|dirty,	MF_MSG_DIRTY_MLOCKED_LRU,	me_pagecache_dirty },
1265	{ mlock|dirty,	mlock,		MF_MSG_CLEAN_MLOCKED_LRU,	me_pagecache_clean },
1266
1267	{ unevict|dirty, unevict|dirty,	MF_MSG_DIRTY_UNEVICTABLE_LRU,	me_pagecache_dirty },
1268	{ unevict|dirty, unevict,	MF_MSG_CLEAN_UNEVICTABLE_LRU,	me_pagecache_clean },
1269
1270	{ lru|dirty,	lru|dirty,	MF_MSG_DIRTY_LRU,	me_pagecache_dirty },
1271	{ lru|dirty,	lru,		MF_MSG_CLEAN_LRU,	me_pagecache_clean },
1272
1273	/*
1274	 * Catchall entry: must be at end.
1275	 */
1276	{ 0,		0,		MF_MSG_UNKNOWN,	me_unknown },
1277};
1278
1279#undef dirty
1280#undef sc
1281#undef unevict
1282#undef mlock
 
1283#undef lru
 
1284#undef head
 
 
1285#undef slab
1286#undef reserved
1287
1288static void update_per_node_mf_stats(unsigned long pfn,
1289				     enum mf_result result)
1290{
1291	int nid = MAX_NUMNODES;
1292	struct memory_failure_stats *mf_stats = NULL;
1293
1294	nid = pfn_to_nid(pfn);
1295	if (unlikely(nid < 0 || nid >= MAX_NUMNODES)) {
1296		WARN_ONCE(1, "Memory failure: pfn=%#lx, invalid nid=%d", pfn, nid);
1297		return;
1298	}
1299
1300	mf_stats = &NODE_DATA(nid)->mf_stats;
1301	switch (result) {
1302	case MF_IGNORED:
1303		++mf_stats->ignored;
1304		break;
1305	case MF_FAILED:
1306		++mf_stats->failed;
1307		break;
1308	case MF_DELAYED:
1309		++mf_stats->delayed;
1310		break;
1311	case MF_RECOVERED:
1312		++mf_stats->recovered;
1313		break;
1314	default:
1315		WARN_ONCE(1, "Memory failure: mf_result=%d is not properly handled", result);
1316		break;
1317	}
1318	++mf_stats->total;
1319}
1320
1321/*
1322 * "Dirty/Clean" indication is not 100% accurate due to the possibility of
1323 * setting PG_dirty outside page lock. See also comment above set_page_dirty().
1324 */
1325static int action_result(unsigned long pfn, enum mf_action_page_type type,
1326			 enum mf_result result)
1327{
1328	trace_memory_failure_event(pfn, type, result);
1329
1330	num_poisoned_pages_inc(pfn);
1331
1332	update_per_node_mf_stats(pfn, result);
1333
1334	pr_err("%#lx: recovery action for %s: %s\n",
1335		pfn, action_page_types[type], action_name[result]);
1336
1337	return (result == MF_RECOVERED || result == MF_DELAYED) ? 0 : -EBUSY;
1338}
1339
1340static int page_action(struct page_state *ps, struct page *p,
1341			unsigned long pfn)
1342{
1343	int result;
 
 
 
 
1344
1345	/* page p should be unlocked after returning from ps->action().  */
1346	result = ps->action(ps, p);
 
 
 
 
 
 
 
1347
1348	/* Could do more checks here if page looks ok */
1349	/*
1350	 * Could adjust zone counters here to correct for the missing page.
1351	 */
1352
1353	return action_result(pfn, ps->type, result);
1354}
1355
1356static inline bool PageHWPoisonTakenOff(struct page *page)
1357{
1358	return PageHWPoison(page) && page_private(page) == MAGIC_HWPOISON;
1359}
1360
1361void SetPageHWPoisonTakenOff(struct page *page)
1362{
1363	set_page_private(page, MAGIC_HWPOISON);
1364}
1365
1366void ClearPageHWPoisonTakenOff(struct page *page)
1367{
1368	if (PageHWPoison(page))
1369		set_page_private(page, 0);
1370}
1371
1372/*
1373 * Return true if a page type of a given page is supported by hwpoison
1374 * mechanism (while handling could fail), otherwise false.  This function
1375 * does not return true for hugetlb or device memory pages, so it's assumed
1376 * to be called only in the context where we never have such pages.
1377 */
1378static inline bool HWPoisonHandlable(struct page *page, unsigned long flags)
1379{
1380	if (PageSlab(page))
1381		return false;
1382
1383	/* Soft offline could migrate non-LRU movable pages */
1384	if ((flags & MF_SOFT_OFFLINE) && __PageMovable(page))
1385		return true;
1386
1387	return PageLRU(page) || is_free_buddy_page(page);
1388}
1389
1390static int __get_hwpoison_page(struct page *page, unsigned long flags)
1391{
1392	struct folio *folio = page_folio(page);
1393	int ret = 0;
1394	bool hugetlb = false;
1395
1396	ret = get_hwpoison_hugetlb_folio(folio, &hugetlb, false);
1397	if (hugetlb) {
1398		/* Make sure hugetlb demotion did not happen from under us. */
1399		if (folio == page_folio(page))
1400			return ret;
1401		if (ret > 0) {
1402			folio_put(folio);
1403			folio = page_folio(page);
1404		}
1405	}
1406
1407	/*
1408	 * This check prevents from calling folio_try_get() for any
1409	 * unsupported type of folio in order to reduce the risk of unexpected
1410	 * races caused by taking a folio refcount.
1411	 */
1412	if (!HWPoisonHandlable(&folio->page, flags))
1413		return -EBUSY;
1414
1415	if (folio_try_get(folio)) {
1416		if (folio == page_folio(page))
1417			return 1;
1418
1419		pr_info("%#lx cannot catch tail\n", page_to_pfn(page));
1420		folio_put(folio);
1421	}
1422
1423	return 0;
1424}
1425
1426static int get_any_page(struct page *p, unsigned long flags)
1427{
1428	int ret = 0, pass = 0;
1429	bool count_increased = false;
1430
1431	if (flags & MF_COUNT_INCREASED)
1432		count_increased = true;
1433
1434try_again:
1435	if (!count_increased) {
1436		ret = __get_hwpoison_page(p, flags);
1437		if (!ret) {
1438			if (page_count(p)) {
1439				/* We raced with an allocation, retry. */
1440				if (pass++ < 3)
1441					goto try_again;
1442				ret = -EBUSY;
1443			} else if (!PageHuge(p) && !is_free_buddy_page(p)) {
1444				/* We raced with put_page, retry. */
1445				if (pass++ < 3)
1446					goto try_again;
1447				ret = -EIO;
1448			}
1449			goto out;
1450		} else if (ret == -EBUSY) {
1451			/*
1452			 * We raced with (possibly temporary) unhandlable
1453			 * page, retry.
1454			 */
1455			if (pass++ < 3) {
1456				shake_page(p);
1457				goto try_again;
1458			}
1459			ret = -EIO;
1460			goto out;
1461		}
1462	}
1463
1464	if (PageHuge(p) || HWPoisonHandlable(p, flags)) {
1465		ret = 1;
1466	} else {
1467		/*
1468		 * A page we cannot handle. Check whether we can turn
1469		 * it into something we can handle.
1470		 */
1471		if (pass++ < 3) {
1472			put_page(p);
1473			shake_page(p);
1474			count_increased = false;
1475			goto try_again;
1476		}
1477		put_page(p);
1478		ret = -EIO;
1479	}
1480out:
1481	if (ret == -EIO)
1482		pr_err("%#lx: unhandlable page.\n", page_to_pfn(p));
1483
1484	return ret;
1485}
1486
1487static int __get_unpoison_page(struct page *page)
1488{
1489	struct folio *folio = page_folio(page);
1490	int ret = 0;
1491	bool hugetlb = false;
1492
1493	ret = get_hwpoison_hugetlb_folio(folio, &hugetlb, true);
1494	if (hugetlb) {
1495		/* Make sure hugetlb demotion did not happen from under us. */
1496		if (folio == page_folio(page))
1497			return ret;
1498		if (ret > 0)
1499			folio_put(folio);
1500	}
1501
1502	/*
1503	 * PageHWPoisonTakenOff pages are not only marked as PG_hwpoison,
1504	 * but also isolated from buddy freelist, so need to identify the
1505	 * state and have to cancel both operations to unpoison.
1506	 */
1507	if (PageHWPoisonTakenOff(page))
1508		return -EHWPOISON;
1509
1510	return get_page_unless_zero(page) ? 1 : 0;
1511}
1512
1513/**
1514 * get_hwpoison_page() - Get refcount for memory error handling
1515 * @p:		Raw error page (hit by memory error)
1516 * @flags:	Flags controlling behavior of error handling
1517 *
1518 * get_hwpoison_page() takes a page refcount of an error page to handle memory
1519 * error on it, after checking that the error page is in a well-defined state
1520 * (defined as a page-type we can successfully handle the memory error on it,
1521 * such as LRU page and hugetlb page).
1522 *
1523 * Memory error handling could be triggered at any time on any type of page,
1524 * so it's prone to race with typical memory management lifecycle (like
1525 * allocation and free).  So to avoid such races, get_hwpoison_page() takes
1526 * extra care for the error page's state (as done in __get_hwpoison_page()),
1527 * and has some retry logic in get_any_page().
1528 *
1529 * When called from unpoison_memory(), the caller should already ensure that
1530 * the given page has PG_hwpoison. So it's never reused for other page
1531 * allocations, and __get_unpoison_page() never races with them.
1532 *
1533 * Return: 0 on failure,
1534 *         1 on success for in-use pages in a well-defined state,
1535 *         -EIO for pages on which we can not handle memory errors,
1536 *         -EBUSY when get_hwpoison_page() has raced with page lifecycle
1537 *         operations like allocation and free,
1538 *         -EHWPOISON when the page is hwpoisoned and taken off from buddy.
1539 */
1540static int get_hwpoison_page(struct page *p, unsigned long flags)
1541{
1542	int ret;
1543
1544	zone_pcp_disable(page_zone(p));
1545	if (flags & MF_UNPOISON)
1546		ret = __get_unpoison_page(p);
1547	else
1548		ret = get_any_page(p, flags);
1549	zone_pcp_enable(page_zone(p));
1550
1551	return ret;
1552}
1553
1554/*
1555 * Do all that is necessary to remove user space mappings. Unmap
1556 * the pages and send SIGBUS to the processes if the data was dirty.
1557 */
1558static bool hwpoison_user_mappings(struct page *p, unsigned long pfn,
1559				  int flags, struct page *hpage)
1560{
1561	struct folio *folio = page_folio(hpage);
1562	enum ttu_flags ttu = TTU_IGNORE_MLOCK | TTU_SYNC | TTU_HWPOISON;
1563	struct address_space *mapping;
1564	LIST_HEAD(tokill);
1565	bool unmap_success;
1566	int forcekill;
1567	bool mlocked = PageMlocked(hpage);
 
1568
1569	/*
1570	 * Here we are interested only in user-mapped pages, so skip any
1571	 * other types of pages.
1572	 */
1573	if (PageReserved(p) || PageSlab(p) || PageTable(p) || PageOffline(p))
1574		return true;
1575	if (!(PageLRU(hpage) || PageHuge(p)))
1576		return true;
1577
1578	/*
1579	 * This check implies we don't kill processes if their pages
1580	 * are in the swap cache early. Those are always late kills.
1581	 */
1582	if (!page_mapped(p))
1583		return true;
 
 
 
1584
1585	if (PageSwapCache(p)) {
1586		pr_err("%#lx: keeping poisoned page in swap cache\n", pfn);
1587		ttu &= ~TTU_HWPOISON;
 
1588	}
1589
1590	/*
1591	 * Propagate the dirty bit from PTEs to struct page first, because we
1592	 * need this to decide if we should kill or just drop the page.
1593	 * XXX: the dirty test could be racy: set_page_dirty() may not always
1594	 * be called inside page lock (it's recommended but not enforced).
1595	 */
1596	mapping = page_mapping(hpage);
1597	if (!(flags & MF_MUST_KILL) && !PageDirty(hpage) && mapping &&
1598	    mapping_can_writeback(mapping)) {
1599		if (page_mkclean(hpage)) {
1600			SetPageDirty(hpage);
1601		} else {
1602			ttu &= ~TTU_HWPOISON;
1603			pr_info("%#lx: corrupted page was clean: dropped without side effects\n",
 
 
1604				pfn);
1605		}
1606	}
1607
1608	/*
1609	 * First collect all the processes that have the page
1610	 * mapped in dirty form.  This has to be done before try_to_unmap,
1611	 * because ttu takes the rmap data structures down.
 
1612	 */
1613	collect_procs(folio, p, &tokill, flags & MF_ACTION_REQUIRED);
1614
1615	if (PageHuge(hpage) && !PageAnon(hpage)) {
1616		/*
1617		 * For hugetlb pages in shared mappings, try_to_unmap
1618		 * could potentially call huge_pmd_unshare.  Because of
1619		 * this, take semaphore in write mode here and set
1620		 * TTU_RMAP_LOCKED to indicate we have taken the lock
1621		 * at this higher level.
 
 
 
1622		 */
1623		mapping = hugetlb_page_mapping_lock_write(hpage);
1624		if (mapping) {
1625			try_to_unmap(folio, ttu|TTU_RMAP_LOCKED);
1626			i_mmap_unlock_write(mapping);
1627		} else
1628			pr_info("%#lx: could not lock mapping for mapped huge page\n", pfn);
1629	} else {
1630		try_to_unmap(folio, ttu);
 
 
 
 
 
 
 
 
 
1631	}
1632
1633	unmap_success = !page_mapped(p);
1634	if (!unmap_success)
1635		pr_err("%#lx: failed to unmap page (mapcount=%d)\n",
1636		       pfn, page_mapcount(p));
1637
1638	/*
1639	 * try_to_unmap() might put mlocked page in lru cache, so call
1640	 * shake_page() again to ensure that it's flushed.
 
 
 
 
1641	 */
1642	if (mlocked)
1643		shake_page(hpage);
 
 
 
 
 
 
 
 
 
 
 
1644
1645	/*
1646	 * Now that the dirty bit has been propagated to the
1647	 * struct page and all unmaps done we can decide if
1648	 * killing is needed or not.  Only kill when the page
1649	 * was dirty or the process is not restartable,
1650	 * otherwise the tokill list is merely
1651	 * freed.  When there was a problem unmapping earlier
1652	 * use a more force-full uncatchable kill to prevent
1653	 * any accesses to the poisoned memory.
1654	 */
1655	forcekill = PageDirty(hpage) || (flags & MF_MUST_KILL) ||
1656		    !unmap_success;
1657	kill_procs(&tokill, forcekill, !unmap_success, pfn, flags);
1658
1659	return unmap_success;
1660}
1661
1662static int identify_page_state(unsigned long pfn, struct page *p,
1663				unsigned long page_flags)
1664{
1665	struct page_state *ps;
1666
1667	/*
1668	 * The first check uses the current page flags which may not have any
1669	 * relevant information. The second check with the saved page flags is
1670	 * carried out only if the first check can't determine the page status.
1671	 */
1672	for (ps = error_states;; ps++)
1673		if ((p->flags & ps->mask) == ps->res)
1674			break;
1675
1676	page_flags |= (p->flags & (1UL << PG_dirty));
1677
1678	if (!ps->mask)
1679		for (ps = error_states;; ps++)
1680			if ((page_flags & ps->mask) == ps->res)
1681				break;
1682	return page_action(ps, p, pfn);
1683}
1684
1685static int try_to_split_thp_page(struct page *page)
1686{
1687	int ret;
1688
1689	lock_page(page);
1690	ret = split_huge_page(page);
1691	unlock_page(page);
1692
1693	if (unlikely(ret))
1694		put_page(page);
1695
1696	return ret;
1697}
1698
1699static void unmap_and_kill(struct list_head *to_kill, unsigned long pfn,
1700		struct address_space *mapping, pgoff_t index, int flags)
1701{
1702	struct to_kill *tk;
1703	unsigned long size = 0;
 
 
 
1704
1705	list_for_each_entry(tk, to_kill, nd)
1706		if (tk->size_shift)
1707			size = max(size, 1UL << tk->size_shift);
1708
1709	if (size) {
1710		/*
1711		 * Unmap the largest mapping to avoid breaking up device-dax
1712		 * mappings which are constant size. The actual size of the
1713		 * mapping being torn down is communicated in siginfo, see
1714		 * kill_proc()
1715		 */
1716		loff_t start = ((loff_t)index << PAGE_SHIFT) & ~(size - 1);
1717
1718		unmap_mapping_range(mapping, start, size, 0);
 
 
 
 
1719	}
1720
1721	kill_procs(to_kill, flags & MF_MUST_KILL, false, pfn, flags);
1722}
1723
1724/*
1725 * Only dev_pagemap pages get here, such as fsdax when the filesystem
1726 * either do not claim or fails to claim a hwpoison event, or devdax.
1727 * The fsdax pages are initialized per base page, and the devdax pages
1728 * could be initialized either as base pages, or as compound pages with
1729 * vmemmap optimization enabled. Devdax is simplistic in its dealing with
1730 * hwpoison, such that, if a subpage of a compound page is poisoned,
1731 * simply mark the compound head page is by far sufficient.
1732 */
1733static int mf_generic_kill_procs(unsigned long long pfn, int flags,
1734		struct dev_pagemap *pgmap)
1735{
1736	struct folio *folio = pfn_folio(pfn);
1737	LIST_HEAD(to_kill);
1738	dax_entry_t cookie;
1739	int rc = 0;
1740
1741	/*
1742	 * Prevent the inode from being freed while we are interrogating
1743	 * the address_space, typically this would be handled by
1744	 * lock_page(), but dax pages do not use the page lock. This
1745	 * also prevents changes to the mapping of this pfn until
1746	 * poison signaling is complete.
 
 
 
 
 
 
 
1747	 */
1748	cookie = dax_lock_folio(folio);
1749	if (!cookie)
1750		return -EBUSY;
1751
1752	if (hwpoison_filter(&folio->page)) {
1753		rc = -EOPNOTSUPP;
1754		goto unlock;
1755	}
1756
1757	switch (pgmap->type) {
1758	case MEMORY_DEVICE_PRIVATE:
1759	case MEMORY_DEVICE_COHERENT:
1760		/*
1761		 * TODO: Handle device pages which may need coordination
1762		 * with device-side memory.
1763		 */
1764		rc = -ENXIO;
1765		goto unlock;
1766	default:
1767		break;
 
 
 
 
 
 
 
1768	}
1769
1770	/*
1771	 * Use this flag as an indication that the dax page has been
1772	 * remapped UC to prevent speculative consumption of poison.
 
 
 
 
1773	 */
1774	SetPageHWPoison(&folio->page);
1775
1776	/*
1777	 * Unlike System-RAM there is no possibility to swap in a
1778	 * different physical page at a given virtual address, so all
1779	 * userspace consumption of ZONE_DEVICE memory necessitates
1780	 * SIGBUS (i.e. MF_MUST_KILL)
1781	 */
1782	flags |= MF_ACTION_REQUIRED | MF_MUST_KILL;
1783	collect_procs(folio, &folio->page, &to_kill, true);
1784
1785	unmap_and_kill(&to_kill, pfn, folio->mapping, folio->index, flags);
1786unlock:
1787	dax_unlock_folio(folio, cookie);
1788	return rc;
1789}
1790
1791#ifdef CONFIG_FS_DAX
1792/**
1793 * mf_dax_kill_procs - Collect and kill processes who are using this file range
1794 * @mapping:	address_space of the file in use
1795 * @index:	start pgoff of the range within the file
1796 * @count:	length of the range, in unit of PAGE_SIZE
1797 * @mf_flags:	memory failure flags
1798 */
1799int mf_dax_kill_procs(struct address_space *mapping, pgoff_t index,
1800		unsigned long count, int mf_flags)
1801{
1802	LIST_HEAD(to_kill);
1803	dax_entry_t cookie;
1804	struct page *page;
1805	size_t end = index + count;
1806	bool pre_remove = mf_flags & MF_MEM_PRE_REMOVE;
1807
1808	mf_flags |= MF_ACTION_REQUIRED | MF_MUST_KILL;
1809
1810	for (; index < end; index++) {
1811		page = NULL;
1812		cookie = dax_lock_mapping_entry(mapping, index, &page);
1813		if (!cookie)
1814			return -EBUSY;
1815		if (!page)
1816			goto unlock;
1817
1818		if (!pre_remove)
1819			SetPageHWPoison(page);
1820
1821		/*
1822		 * The pre_remove case is revoking access, the memory is still
1823		 * good and could theoretically be put back into service.
1824		 */
1825		collect_procs_fsdax(page, mapping, index, &to_kill, pre_remove);
1826		unmap_and_kill(&to_kill, page_to_pfn(page), mapping,
1827				index, mf_flags);
1828unlock:
1829		dax_unlock_mapping_entry(mapping, index, cookie);
1830	}
1831	return 0;
1832}
1833EXPORT_SYMBOL_GPL(mf_dax_kill_procs);
1834#endif /* CONFIG_FS_DAX */
1835
1836#ifdef CONFIG_HUGETLB_PAGE
1837
1838/*
1839 * Struct raw_hwp_page represents information about "raw error page",
1840 * constructing singly linked list from ->_hugetlb_hwpoison field of folio.
1841 */
1842struct raw_hwp_page {
1843	struct llist_node node;
1844	struct page *page;
1845};
1846
1847static inline struct llist_head *raw_hwp_list_head(struct folio *folio)
1848{
1849	return (struct llist_head *)&folio->_hugetlb_hwpoison;
1850}
1851
1852bool is_raw_hwpoison_page_in_hugepage(struct page *page)
1853{
1854	struct llist_head *raw_hwp_head;
1855	struct raw_hwp_page *p;
1856	struct folio *folio = page_folio(page);
1857	bool ret = false;
1858
1859	if (!folio_test_hwpoison(folio))
1860		return false;
1861
1862	if (!folio_test_hugetlb(folio))
1863		return PageHWPoison(page);
1864
1865	/*
1866	 * When RawHwpUnreliable is set, kernel lost track of which subpages
1867	 * are HWPOISON. So return as if ALL subpages are HWPOISONed.
 
1868	 */
1869	if (folio_test_hugetlb_raw_hwp_unreliable(folio))
1870		return true;
1871
1872	mutex_lock(&mf_mutex);
1873
1874	raw_hwp_head = raw_hwp_list_head(folio);
1875	llist_for_each_entry(p, raw_hwp_head->first, node) {
1876		if (page == p->page) {
1877			ret = true;
1878			break;
1879		}
1880	}
1881
1882	mutex_unlock(&mf_mutex);
1883
1884	return ret;
1885}
1886
1887static unsigned long __folio_free_raw_hwp(struct folio *folio, bool move_flag)
1888{
1889	struct llist_node *head;
1890	struct raw_hwp_page *p, *next;
1891	unsigned long count = 0;
1892
1893	head = llist_del_all(raw_hwp_list_head(folio));
1894	llist_for_each_entry_safe(p, next, head, node) {
1895		if (move_flag)
1896			SetPageHWPoison(p->page);
1897		else
1898			num_poisoned_pages_sub(page_to_pfn(p->page), 1);
1899		kfree(p);
1900		count++;
1901	}
1902	return count;
1903}
1904
1905static int folio_set_hugetlb_hwpoison(struct folio *folio, struct page *page)
1906{
1907	struct llist_head *head;
1908	struct raw_hwp_page *raw_hwp;
1909	struct raw_hwp_page *p, *next;
1910	int ret = folio_test_set_hwpoison(folio) ? -EHWPOISON : 0;
1911
1912	/*
1913	 * Once the hwpoison hugepage has lost reliable raw error info,
1914	 * there is little meaning to keep additional error info precisely,
1915	 * so skip to add additional raw error info.
1916	 */
1917	if (folio_test_hugetlb_raw_hwp_unreliable(folio))
1918		return -EHWPOISON;
1919	head = raw_hwp_list_head(folio);
1920	llist_for_each_entry_safe(p, next, head->first, node) {
1921		if (p->page == page)
1922			return -EHWPOISON;
1923	}
1924
1925	raw_hwp = kmalloc(sizeof(struct raw_hwp_page), GFP_ATOMIC);
1926	if (raw_hwp) {
1927		raw_hwp->page = page;
1928		llist_add(&raw_hwp->node, head);
1929		/* the first error event will be counted in action_result(). */
1930		if (ret)
1931			num_poisoned_pages_inc(page_to_pfn(page));
1932	} else {
1933		/*
1934		 * Failed to save raw error info.  We no longer trace all
1935		 * hwpoisoned subpages, and we need refuse to free/dissolve
1936		 * this hwpoisoned hugepage.
1937		 */
1938		folio_set_hugetlb_raw_hwp_unreliable(folio);
1939		/*
1940		 * Once hugetlb_raw_hwp_unreliable is set, raw_hwp_page is not
1941		 * used any more, so free it.
1942		 */
1943		__folio_free_raw_hwp(folio, false);
1944	}
1945	return ret;
1946}
1947
1948static unsigned long folio_free_raw_hwp(struct folio *folio, bool move_flag)
1949{
1950	/*
1951	 * hugetlb_vmemmap_optimized hugepages can't be freed because struct
1952	 * pages for tail pages are required but they don't exist.
1953	 */
1954	if (move_flag && folio_test_hugetlb_vmemmap_optimized(folio))
 
 
 
 
1955		return 0;
1956
1957	/*
1958	 * hugetlb_raw_hwp_unreliable hugepages shouldn't be unpoisoned by
1959	 * definition.
 
 
1960	 */
1961	if (folio_test_hugetlb_raw_hwp_unreliable(folio))
1962		return 0;
1963
1964	return __folio_free_raw_hwp(folio, move_flag);
1965}
1966
1967void folio_clear_hugetlb_hwpoison(struct folio *folio)
1968{
1969	if (folio_test_hugetlb_raw_hwp_unreliable(folio))
1970		return;
1971	if (folio_test_hugetlb_vmemmap_optimized(folio))
1972		return;
1973	folio_clear_hwpoison(folio);
1974	folio_free_raw_hwp(folio, true);
1975}
1976
1977/*
1978 * Called from hugetlb code with hugetlb_lock held.
1979 *
1980 * Return values:
1981 *   0             - free hugepage
1982 *   1             - in-use hugepage
1983 *   2             - not a hugepage
1984 *   -EBUSY        - the hugepage is busy (try to retry)
1985 *   -EHWPOISON    - the hugepage is already hwpoisoned
1986 */
1987int __get_huge_page_for_hwpoison(unsigned long pfn, int flags,
1988				 bool *migratable_cleared)
1989{
1990	struct page *page = pfn_to_page(pfn);
1991	struct folio *folio = page_folio(page);
1992	int ret = 2;	/* fallback to normal page handling */
1993	bool count_increased = false;
1994
1995	if (!folio_test_hugetlb(folio))
1996		goto out;
1997
1998	if (flags & MF_COUNT_INCREASED) {
1999		ret = 1;
2000		count_increased = true;
2001	} else if (folio_test_hugetlb_freed(folio)) {
2002		ret = 0;
2003	} else if (folio_test_hugetlb_migratable(folio)) {
2004		ret = folio_try_get(folio);
2005		if (ret)
2006			count_increased = true;
2007	} else {
2008		ret = -EBUSY;
2009		if (!(flags & MF_NO_RETRY))
2010			goto out;
2011	}
2012
2013	if (folio_set_hugetlb_hwpoison(folio, page)) {
2014		ret = -EHWPOISON;
2015		goto out;
2016	}
2017
2018	/*
2019	 * Clearing hugetlb_migratable for hwpoisoned hugepages to prevent them
2020	 * from being migrated by memory hotremove.
2021	 */
2022	if (count_increased && folio_test_hugetlb_migratable(folio)) {
2023		folio_clear_hugetlb_migratable(folio);
2024		*migratable_cleared = true;
 
2025	}
2026
2027	return ret;
2028out:
2029	if (count_increased)
2030		folio_put(folio);
2031	return ret;
2032}
2033
2034/*
2035 * Taking refcount of hugetlb pages needs extra care about race conditions
2036 * with basic operations like hugepage allocation/free/demotion.
2037 * So some of prechecks for hwpoison (pinning, and testing/setting
2038 * PageHWPoison) should be done in single hugetlb_lock range.
2039 */
2040static int try_memory_failure_hugetlb(unsigned long pfn, int flags, int *hugetlb)
2041{
2042	int res;
2043	struct page *p = pfn_to_page(pfn);
2044	struct folio *folio;
2045	unsigned long page_flags;
2046	bool migratable_cleared = false;
2047
2048	*hugetlb = 1;
2049retry:
2050	res = get_huge_page_for_hwpoison(pfn, flags, &migratable_cleared);
2051	if (res == 2) { /* fallback to normal page handling */
2052		*hugetlb = 0;
2053		return 0;
2054	} else if (res == -EHWPOISON) {
2055		pr_err("%#lx: already hardware poisoned\n", pfn);
2056		if (flags & MF_ACTION_REQUIRED) {
2057			folio = page_folio(p);
2058			res = kill_accessing_process(current, folio_pfn(folio), flags);
2059		}
2060		return res;
2061	} else if (res == -EBUSY) {
2062		if (!(flags & MF_NO_RETRY)) {
2063			flags |= MF_NO_RETRY;
2064			goto retry;
2065		}
2066		return action_result(pfn, MF_MSG_UNKNOWN, MF_IGNORED);
2067	}
2068
2069	folio = page_folio(p);
2070	folio_lock(folio);
2071
2072	if (hwpoison_filter(p)) {
2073		folio_clear_hugetlb_hwpoison(folio);
2074		if (migratable_cleared)
2075			folio_set_hugetlb_migratable(folio);
2076		folio_unlock(folio);
2077		if (res == 1)
2078			folio_put(folio);
2079		return -EOPNOTSUPP;
2080	}
2081
2082	/*
2083	 * Handling free hugepage.  The possible race with hugepage allocation
2084	 * or demotion can be prevented by PageHWPoison flag.
2085	 */
2086	if (res == 0) {
2087		folio_unlock(folio);
2088		if (__page_handle_poison(p) >= 0) {
2089			page_ref_inc(p);
2090			res = MF_RECOVERED;
2091		} else {
2092			res = MF_FAILED;
2093		}
2094		return action_result(pfn, MF_MSG_FREE_HUGE, res);
2095	}
2096
2097	page_flags = folio->flags;
2098
2099	if (!hwpoison_user_mappings(p, pfn, flags, &folio->page)) {
2100		folio_unlock(folio);
2101		return action_result(pfn, MF_MSG_UNMAP_FAILED, MF_IGNORED);
2102	}
2103
2104	return identify_page_state(pfn, p, page_flags);
2105}
2106
2107#else
2108static inline int try_memory_failure_hugetlb(unsigned long pfn, int flags, int *hugetlb)
2109{
2110	return 0;
2111}
2112
2113static inline unsigned long folio_free_raw_hwp(struct folio *folio, bool flag)
2114{
2115	return 0;
2116}
2117#endif	/* CONFIG_HUGETLB_PAGE */
2118
2119/* Drop the extra refcount in case we come from madvise() */
2120static void put_ref_page(unsigned long pfn, int flags)
2121{
2122	struct page *page;
2123
2124	if (!(flags & MF_COUNT_INCREASED))
2125		return;
2126
2127	page = pfn_to_page(pfn);
2128	if (page)
2129		put_page(page);
2130}
2131
2132static int memory_failure_dev_pagemap(unsigned long pfn, int flags,
2133		struct dev_pagemap *pgmap)
2134{
2135	int rc = -ENXIO;
2136
2137	/* device metadata space is not recoverable */
2138	if (!pgmap_pfn_valid(pgmap, pfn))
2139		goto out;
2140
2141	/*
2142	 * Call driver's implementation to handle the memory failure, otherwise
2143	 * fall back to generic handler.
2144	 */
2145	if (pgmap_has_memory_failure(pgmap)) {
2146		rc = pgmap->ops->memory_failure(pgmap, pfn, 1, flags);
2147		/*
2148		 * Fall back to generic handler too if operation is not
2149		 * supported inside the driver/device/filesystem.
2150		 */
2151		if (rc != -EOPNOTSUPP)
2152			goto out;
2153	}
2154
2155	rc = mf_generic_kill_procs(pfn, flags, pgmap);
2156out:
2157	/* drop pgmap ref acquired in caller */
2158	put_dev_pagemap(pgmap);
2159	if (rc != -EOPNOTSUPP)
2160		action_result(pfn, MF_MSG_DAX, rc ? MF_FAILED : MF_RECOVERED);
2161	return rc;
2162}
 
2163
2164/**
2165 * memory_failure - Handle memory failure of a page.
2166 * @pfn: Page Number of the corrupted page
2167 * @flags: fine tune action taken
2168 *
2169 * This function is called by the low level machine check code
2170 * of an architecture when it detects hardware memory corruption
2171 * of a page. It tries its best to recover, which includes
2172 * dropping pages, killing processes etc.
2173 *
2174 * The function is primarily of use for corruptions that
2175 * happen outside the current execution context (e.g. when
2176 * detected by a background scrubber)
2177 *
2178 * Must run in process context (e.g. a work queue) with interrupts
2179 * enabled and no spinlocks held.
2180 *
2181 * Return: 0 for successfully handled the memory error,
2182 *         -EOPNOTSUPP for hwpoison_filter() filtered the error event,
2183 *         < 0(except -EOPNOTSUPP) on failure.
2184 */
2185int memory_failure(unsigned long pfn, int flags)
2186{
2187	struct page *p;
2188	struct page *hpage;
2189	struct dev_pagemap *pgmap;
2190	int res = 0;
2191	unsigned long page_flags;
2192	bool retry = true;
2193	int hugetlb = 0;
2194
2195	if (!sysctl_memory_failure_recovery)
2196		panic("Memory failure on page %lx", pfn);
2197
2198	mutex_lock(&mf_mutex);
2199
2200	if (!(flags & MF_SW_SIMULATED))
2201		hw_memory_failure = true;
2202
2203	p = pfn_to_online_page(pfn);
2204	if (!p) {
2205		res = arch_memory_failure(pfn, flags);
2206		if (res == 0)
2207			goto unlock_mutex;
2208
2209		if (pfn_valid(pfn)) {
2210			pgmap = get_dev_pagemap(pfn, NULL);
2211			put_ref_page(pfn, flags);
2212			if (pgmap) {
2213				res = memory_failure_dev_pagemap(pfn, flags,
2214								 pgmap);
2215				goto unlock_mutex;
2216			}
2217		}
2218		pr_err("%#lx: memory outside kernel control\n", pfn);
2219		res = -ENXIO;
2220		goto unlock_mutex;
2221	}
2222
2223try_again:
2224	res = try_memory_failure_hugetlb(pfn, flags, &hugetlb);
2225	if (hugetlb)
2226		goto unlock_mutex;
2227
2228	if (TestSetPageHWPoison(p)) {
2229		pr_err("%#lx: already hardware poisoned\n", pfn);
2230		res = -EHWPOISON;
2231		if (flags & MF_ACTION_REQUIRED)
2232			res = kill_accessing_process(current, pfn, flags);
2233		if (flags & MF_COUNT_INCREASED)
2234			put_page(p);
2235		goto unlock_mutex;
2236	}
2237
2238	/*
2239	 * We need/can do nothing about count=0 pages.
2240	 * 1) it's a free page, and therefore in safe hand:
2241	 *    check_new_page() will be the gate keeper.
2242	 * 2) it's part of a non-compound high order page.
2243	 *    Implies some kernel user: cannot stop them from
2244	 *    R/W the page; let's pray that the page has been
2245	 *    used and will be freed some time later.
2246	 * In fact it's dangerous to directly bump up page count from 0,
2247	 * that may make page_ref_freeze()/page_ref_unfreeze() mismatch.
2248	 */
2249	if (!(flags & MF_COUNT_INCREASED)) {
2250		res = get_hwpoison_page(p, flags);
2251		if (!res) {
2252			if (is_free_buddy_page(p)) {
2253				if (take_page_off_buddy(p)) {
2254					page_ref_inc(p);
2255					res = MF_RECOVERED;
2256				} else {
2257					/* We lost the race, try again */
2258					if (retry) {
2259						ClearPageHWPoison(p);
2260						retry = false;
2261						goto try_again;
2262					}
2263					res = MF_FAILED;
2264				}
2265				res = action_result(pfn, MF_MSG_BUDDY, res);
2266			} else {
2267				res = action_result(pfn, MF_MSG_KERNEL_HIGH_ORDER, MF_IGNORED);
2268			}
2269			goto unlock_mutex;
2270		} else if (res < 0) {
2271			res = action_result(pfn, MF_MSG_UNKNOWN, MF_IGNORED);
2272			goto unlock_mutex;
2273		}
2274	}
2275
2276	hpage = compound_head(p);
2277	if (PageTransHuge(hpage)) {
2278		/*
2279		 * The flag must be set after the refcount is bumped
2280		 * otherwise it may race with THP split.
2281		 * And the flag can't be set in get_hwpoison_page() since
2282		 * it is called by soft offline too and it is just called
2283		 * for !MF_COUNT_INCREASED.  So here seems to be the best
2284		 * place.
2285		 *
2286		 * Don't need care about the above error handling paths for
2287		 * get_hwpoison_page() since they handle either free page
2288		 * or unhandlable page.  The refcount is bumped iff the
2289		 * page is a valid handlable page.
2290		 */
2291		SetPageHasHWPoisoned(hpage);
2292		if (try_to_split_thp_page(p) < 0) {
2293			res = action_result(pfn, MF_MSG_UNSPLIT_THP, MF_IGNORED);
2294			goto unlock_mutex;
2295		}
2296		VM_BUG_ON_PAGE(!page_count(p), p);
2297	}
2298
2299	/*
2300	 * We ignore non-LRU pages for good reasons.
2301	 * - PG_locked is only well defined for LRU pages and a few others
2302	 * - to avoid races with __SetPageLocked()
2303	 * - to avoid races with __SetPageSlab*() (and more non-atomic ops)
2304	 * The check (unnecessarily) ignores LRU pages being isolated and
2305	 * walked by the page reclaim code, however that's not a big loss.
2306	 */
2307	shake_page(p);
2308
2309	lock_page(p);
2310
2311	/*
2312	 * We're only intended to deal with the non-Compound page here.
2313	 * However, the page could have changed compound pages due to
2314	 * race window. If this happens, we could try again to hopefully
2315	 * handle the page next round.
2316	 */
2317	if (PageCompound(p)) {
2318		if (retry) {
2319			ClearPageHWPoison(p);
2320			unlock_page(p);
2321			put_page(p);
2322			flags &= ~MF_COUNT_INCREASED;
2323			retry = false;
2324			goto try_again;
2325		}
2326		res = action_result(pfn, MF_MSG_DIFFERENT_COMPOUND, MF_IGNORED);
2327		goto unlock_page;
2328	}
2329
2330	/*
2331	 * We use page flags to determine what action should be taken, but
2332	 * the flags can be modified by the error containment action.  One
2333	 * example is an mlocked page, where PG_mlocked is cleared by
2334	 * folio_remove_rmap_*() in try_to_unmap_one(). So to determine page
2335	 * status correctly, we save a copy of the page flags at this time.
2336	 */
2337	page_flags = p->flags;
2338
2339	if (hwpoison_filter(p)) {
2340		ClearPageHWPoison(p);
2341		unlock_page(p);
2342		put_page(p);
2343		res = -EOPNOTSUPP;
2344		goto unlock_mutex;
2345	}
2346
2347	/*
2348	 * __munlock_folio() may clear a writeback page's LRU flag without
2349	 * page_lock. We need wait writeback completion for this page or it
2350	 * may trigger vfs BUG while evict inode.
2351	 */
2352	if (!PageLRU(p) && !PageWriteback(p))
2353		goto identify_page_state;
2354
2355	/*
2356	 * It's very difficult to mess with pages currently under IO
2357	 * and in many cases impossible, so we just avoid it here.
2358	 */
2359	wait_on_page_writeback(p);
2360
2361	/*
2362	 * Now take care of user space mappings.
2363	 * Abort on fail: __filemap_remove_folio() assumes unmapped page.
2364	 */
2365	if (!hwpoison_user_mappings(p, pfn, flags, p)) {
2366		res = action_result(pfn, MF_MSG_UNMAP_FAILED, MF_IGNORED);
2367		goto unlock_page;
2368	}
2369
2370	/*
2371	 * Torn down by someone else?
2372	 */
2373	if (PageLRU(p) && !PageSwapCache(p) && p->mapping == NULL) {
2374		res = action_result(pfn, MF_MSG_TRUNCATED_LRU, MF_IGNORED);
2375		goto unlock_page;
2376	}
2377
2378identify_page_state:
2379	res = identify_page_state(pfn, p, page_flags);
2380	mutex_unlock(&mf_mutex);
2381	return res;
2382unlock_page:
2383	unlock_page(p);
2384unlock_mutex:
2385	mutex_unlock(&mf_mutex);
2386	return res;
2387}
2388EXPORT_SYMBOL_GPL(memory_failure);
2389
2390#define MEMORY_FAILURE_FIFO_ORDER	4
2391#define MEMORY_FAILURE_FIFO_SIZE	(1 << MEMORY_FAILURE_FIFO_ORDER)
2392
2393struct memory_failure_entry {
2394	unsigned long pfn;
 
2395	int flags;
2396};
2397
2398struct memory_failure_cpu {
2399	DECLARE_KFIFO(fifo, struct memory_failure_entry,
2400		      MEMORY_FAILURE_FIFO_SIZE);
2401	spinlock_t lock;
2402	struct work_struct work;
2403};
2404
2405static DEFINE_PER_CPU(struct memory_failure_cpu, memory_failure_cpu);
2406
2407/**
2408 * memory_failure_queue - Schedule handling memory failure of a page.
2409 * @pfn: Page Number of the corrupted page
 
2410 * @flags: Flags for memory failure handling
2411 *
2412 * This function is called by the low level hardware error handler
2413 * when it detects hardware memory corruption of a page. It schedules
2414 * the recovering of error page, including dropping pages, killing
2415 * processes etc.
2416 *
2417 * The function is primarily of use for corruptions that
2418 * happen outside the current execution context (e.g. when
2419 * detected by a background scrubber)
2420 *
2421 * Can run in IRQ context.
2422 */
2423void memory_failure_queue(unsigned long pfn, int flags)
2424{
2425	struct memory_failure_cpu *mf_cpu;
2426	unsigned long proc_flags;
2427	struct memory_failure_entry entry = {
2428		.pfn =		pfn,
 
2429		.flags =	flags,
2430	};
2431
2432	mf_cpu = &get_cpu_var(memory_failure_cpu);
2433	spin_lock_irqsave(&mf_cpu->lock, proc_flags);
2434	if (kfifo_put(&mf_cpu->fifo, entry))
2435		schedule_work_on(smp_processor_id(), &mf_cpu->work);
2436	else
2437		pr_err("buffer overflow when queuing memory failure at %#lx\n",
2438		       pfn);
2439	spin_unlock_irqrestore(&mf_cpu->lock, proc_flags);
2440	put_cpu_var(memory_failure_cpu);
2441}
2442EXPORT_SYMBOL_GPL(memory_failure_queue);
2443
2444static void memory_failure_work_func(struct work_struct *work)
2445{
2446	struct memory_failure_cpu *mf_cpu;
2447	struct memory_failure_entry entry = { 0, };
2448	unsigned long proc_flags;
2449	int gotten;
2450
2451	mf_cpu = container_of(work, struct memory_failure_cpu, work);
2452	for (;;) {
2453		spin_lock_irqsave(&mf_cpu->lock, proc_flags);
2454		gotten = kfifo_get(&mf_cpu->fifo, &entry);
2455		spin_unlock_irqrestore(&mf_cpu->lock, proc_flags);
2456		if (!gotten)
2457			break;
2458		if (entry.flags & MF_SOFT_OFFLINE)
2459			soft_offline_page(entry.pfn, entry.flags);
2460		else
2461			memory_failure(entry.pfn, entry.flags);
2462	}
2463}
2464
2465/*
2466 * Process memory_failure work queued on the specified CPU.
2467 * Used to avoid return-to-userspace racing with the memory_failure workqueue.
2468 */
2469void memory_failure_queue_kick(int cpu)
2470{
2471	struct memory_failure_cpu *mf_cpu;
2472
2473	mf_cpu = &per_cpu(memory_failure_cpu, cpu);
2474	cancel_work_sync(&mf_cpu->work);
2475	memory_failure_work_func(&mf_cpu->work);
2476}
2477
2478static int __init memory_failure_init(void)
2479{
2480	struct memory_failure_cpu *mf_cpu;
2481	int cpu;
2482
2483	for_each_possible_cpu(cpu) {
2484		mf_cpu = &per_cpu(memory_failure_cpu, cpu);
2485		spin_lock_init(&mf_cpu->lock);
2486		INIT_KFIFO(mf_cpu->fifo);
2487		INIT_WORK(&mf_cpu->work, memory_failure_work_func);
2488	}
2489
2490	register_sysctl_init("vm", memory_failure_table);
2491
2492	return 0;
2493}
2494core_initcall(memory_failure_init);
2495
2496#undef pr_fmt
2497#define pr_fmt(fmt)	"" fmt
2498#define unpoison_pr_info(fmt, pfn, rs)			\
2499({							\
2500	if (__ratelimit(rs))				\
2501		pr_info(fmt, pfn);			\
2502})
2503
2504/**
2505 * unpoison_memory - Unpoison a previously poisoned page
2506 * @pfn: Page number of the to be unpoisoned page
2507 *
2508 * Software-unpoison a page that has been poisoned by
2509 * memory_failure() earlier.
2510 *
2511 * This is only done on the software-level, so it only works
2512 * for linux injected failures, not real hardware failures
2513 *
2514 * Returns 0 for success, otherwise -errno.
2515 */
2516int unpoison_memory(unsigned long pfn)
2517{
2518	struct folio *folio;
2519	struct page *p;
2520	int ret = -EBUSY, ghp;
2521	unsigned long count = 1;
2522	bool huge = false;
2523	static DEFINE_RATELIMIT_STATE(unpoison_rs, DEFAULT_RATELIMIT_INTERVAL,
2524					DEFAULT_RATELIMIT_BURST);
2525
2526	if (!pfn_valid(pfn))
2527		return -ENXIO;
2528
2529	p = pfn_to_page(pfn);
2530	folio = page_folio(p);
2531
2532	mutex_lock(&mf_mutex);
2533
2534	if (hw_memory_failure) {
2535		unpoison_pr_info("Unpoison: Disabled after HW memory failure %#lx\n",
2536				 pfn, &unpoison_rs);
2537		ret = -EOPNOTSUPP;
2538		goto unlock_mutex;
2539	}
2540
2541	if (!PageHWPoison(p)) {
2542		unpoison_pr_info("Unpoison: Page was already unpoisoned %#lx\n",
2543				 pfn, &unpoison_rs);
2544		goto unlock_mutex;
2545	}
2546
2547	if (folio_ref_count(folio) > 1) {
2548		unpoison_pr_info("Unpoison: Someone grabs the hwpoison page %#lx\n",
2549				 pfn, &unpoison_rs);
2550		goto unlock_mutex;
 
 
 
 
 
 
 
 
 
 
 
2551	}
2552
2553	if (folio_test_slab(folio) || PageTable(&folio->page) ||
2554	    folio_test_reserved(folio) || PageOffline(&folio->page))
2555		goto unlock_mutex;
2556
2557	/*
2558	 * Note that folio->_mapcount is overloaded in SLAB, so the simple test
2559	 * in folio_mapped() has to be done after folio_test_slab() is checked.
2560	 */
2561	if (folio_mapped(folio)) {
2562		unpoison_pr_info("Unpoison: Someone maps the hwpoison page %#lx\n",
2563				 pfn, &unpoison_rs);
2564		goto unlock_mutex;
 
 
 
 
2565	}
 
2566
2567	if (folio_mapping(folio)) {
2568		unpoison_pr_info("Unpoison: the hwpoison page has non-NULL mapping %#lx\n",
2569				 pfn, &unpoison_rs);
2570		goto unlock_mutex;
2571	}
2572
2573	ghp = get_hwpoison_page(p, MF_UNPOISON);
2574	if (!ghp) {
2575		if (PageHuge(p)) {
2576			huge = true;
2577			count = folio_free_raw_hwp(folio, false);
2578			if (count == 0)
2579				goto unlock_mutex;
2580		}
2581		ret = folio_test_clear_hwpoison(folio) ? 0 : -EBUSY;
2582	} else if (ghp < 0) {
2583		if (ghp == -EHWPOISON) {
2584			ret = put_page_back_buddy(p) ? 0 : -EBUSY;
2585		} else {
2586			ret = ghp;
2587			unpoison_pr_info("Unpoison: failed to grab page %#lx\n",
2588					 pfn, &unpoison_rs);
2589		}
2590	} else {
2591		if (PageHuge(p)) {
2592			huge = true;
2593			count = folio_free_raw_hwp(folio, false);
2594			if (count == 0) {
2595				folio_put(folio);
2596				goto unlock_mutex;
2597			}
2598		}
2599
2600		folio_put(folio);
2601		if (TestClearPageHWPoison(p)) {
2602			folio_put(folio);
2603			ret = 0;
2604		}
2605	}
2606
2607unlock_mutex:
2608	mutex_unlock(&mf_mutex);
2609	if (!ret) {
2610		if (!huge)
2611			num_poisoned_pages_sub(pfn, 1);
2612		unpoison_pr_info("Unpoison: Software-unpoisoned page %#lx\n",
2613				 page_to_pfn(p), &unpoison_rs);
2614	}
2615	return ret;
2616}
2617EXPORT_SYMBOL(unpoison_memory);
2618
2619static bool mf_isolate_folio(struct folio *folio, struct list_head *pagelist)
2620{
2621	bool isolated = false;
 
 
 
 
 
 
2622
2623	if (folio_test_hugetlb(folio)) {
2624		isolated = isolate_hugetlb(folio, pagelist);
2625	} else {
2626		bool lru = !__folio_test_movable(folio);
 
 
 
 
 
2627
2628		if (lru)
2629			isolated = folio_isolate_lru(folio);
2630		else
2631			isolated = isolate_movable_page(&folio->page,
2632							ISOLATE_UNEVICTABLE);
2633
2634		if (isolated) {
2635			list_add(&folio->lru, pagelist);
2636			if (lru)
2637				node_stat_add_folio(folio, NR_ISOLATED_ANON +
2638						    folio_is_file_lru(folio));
2639		}
2640	}
2641
2642	/*
2643	 * If we succeed to isolate the folio, we grabbed another refcount on
2644	 * the folio, so we can safely drop the one we got from get_any_page().
2645	 * If we failed to isolate the folio, it means that we cannot go further
2646	 * and we will return an error, so drop the reference we got from
2647	 * get_any_page() as well.
 
 
2648	 */
2649	folio_put(folio);
2650	return isolated;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2651}
2652
2653/*
2654 * soft_offline_in_use_page handles hugetlb-pages and non-hugetlb pages.
2655 * If the page is a non-dirty unmapped page-cache page, it simply invalidates.
2656 * If the page is mapped, it migrates the contents over.
2657 */
2658static int soft_offline_in_use_page(struct page *page)
2659{
2660	long ret = 0;
2661	unsigned long pfn = page_to_pfn(page);
2662	struct folio *folio = page_folio(page);
2663	char const *msg_page[] = {"page", "hugepage"};
2664	bool huge = folio_test_hugetlb(folio);
2665	LIST_HEAD(pagelist);
2666	struct migration_target_control mtc = {
2667		.nid = NUMA_NO_NODE,
2668		.gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL,
2669	};
2670
2671	if (!huge && folio_test_large(folio)) {
2672		if (try_to_split_thp_page(page)) {
2673			pr_info("soft offline: %#lx: thp split failed\n", pfn);
2674			return -EBUSY;
2675		}
2676		folio = page_folio(page);
 
 
 
 
2677	}
2678
2679	folio_lock(folio);
2680	if (!huge)
2681		folio_wait_writeback(folio);
2682	if (PageHWPoison(page)) {
2683		folio_unlock(folio);
2684		folio_put(folio);
2685		pr_info("soft offline: %#lx page already poisoned\n", pfn);
2686		return 0;
2687	}
2688
2689	if (!huge && folio_test_lru(folio) && !folio_test_swapcache(folio))
2690		/*
2691		 * Try to invalidate first. This should work for
2692		 * non dirty unmapped page cache pages.
2693		 */
2694		ret = mapping_evict_folio(folio_mapping(folio), folio);
2695	folio_unlock(folio);
2696
 
 
 
2697	if (ret) {
2698		pr_info("soft_offline: %#lx: invalidated\n", pfn);
2699		page_handle_poison(page, false, true);
2700		return 0;
2701	}
2702
2703	if (mf_isolate_folio(folio, &pagelist)) {
2704		ret = migrate_pages(&pagelist, alloc_migration_target, NULL,
2705			(unsigned long)&mtc, MIGRATE_SYNC, MR_MEMORY_FAILURE, NULL);
2706		if (!ret) {
2707			bool release = !huge;
2708
2709			if (!page_handle_poison(page, huge, release))
2710				ret = -EBUSY;
2711		} else {
2712			if (!list_empty(&pagelist))
2713				putback_movable_pages(&pagelist);
2714
2715			pr_info("soft offline: %#lx: %s migration failed %ld, type %pGp\n",
2716				pfn, msg_page[huge], ret, &page->flags);
2717			if (ret > 0)
2718				ret = -EBUSY;
2719		}
2720	} else {
2721		pr_info("soft offline: %#lx: %s isolation failed, page count %d, type %pGp\n",
2722			pfn, msg_page[huge], page_count(page), &page->flags);
2723		ret = -EBUSY;
2724	}
 
 
 
 
 
 
2725	return ret;
2726}
2727
2728/**
2729 * soft_offline_page - Soft offline a page.
2730 * @pfn: pfn to soft-offline
2731 * @flags: flags. Same as memory_failure().
2732 *
2733 * Returns 0 on success
2734 *         -EOPNOTSUPP for hwpoison_filter() filtered the error event
2735 *         < 0 otherwise negated errno.
2736 *
2737 * Soft offline a page, by migration or invalidation,
2738 * without killing anything. This is for the case when
2739 * a page is not corrupted yet (so it's still valid to access),
2740 * but has had a number of corrected errors and is better taken
2741 * out.
2742 *
2743 * The actual policy on when to do that is maintained by
2744 * user space.
2745 *
2746 * This should never impact any application or cause data loss,
2747 * however it might take some time.
2748 *
2749 * This is not a 100% solution for all memory, but tries to be
2750 * ``good enough'' for the majority of memory.
2751 */
2752int soft_offline_page(unsigned long pfn, int flags)
2753{
2754	int ret;
2755	bool try_again = true;
2756	struct page *page;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2757
2758	if (!pfn_valid(pfn)) {
2759		WARN_ON_ONCE(flags & MF_COUNT_INCREASED);
2760		return -ENXIO;
 
 
 
 
 
2761	}
2762
2763	/* Only online pages can be soft-offlined (esp., not ZONE_DEVICE). */
2764	page = pfn_to_online_page(pfn);
2765	if (!page) {
2766		put_ref_page(pfn, flags);
2767		return -EIO;
2768	}
2769
2770	mutex_lock(&mf_mutex);
 
2771
 
 
 
2772	if (PageHWPoison(page)) {
2773		pr_info("%s: %#lx page already poisoned\n", __func__, pfn);
2774		put_ref_page(pfn, flags);
2775		mutex_unlock(&mf_mutex);
2776		return 0;
2777	}
2778
2779retry:
2780	get_online_mems();
2781	ret = get_hwpoison_page(page, flags | MF_SOFT_OFFLINE);
2782	put_online_mems();
2783
2784	if (hwpoison_filter(page)) {
2785		if (ret > 0)
2786			put_page(page);
2787
2788		mutex_unlock(&mf_mutex);
2789		return -EOPNOTSUPP;
 
 
 
 
2790	}
2791
2792	if (ret > 0) {
2793		ret = soft_offline_in_use_page(page);
2794	} else if (ret == 0) {
2795		if (!page_handle_poison(page, true, false)) {
2796			if (try_again) {
2797				try_again = false;
2798				flags &= ~MF_COUNT_INCREASED;
2799				goto retry;
2800			}
2801			ret = -EBUSY;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2802		}
 
 
 
2803	}
 
 
2804
2805	mutex_unlock(&mf_mutex);
2806
 
 
2807	return ret;
2808}