Linux Audio

Check our new training course

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}
v6.13.7
   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
  71static int sysctl_enable_soft_offline __read_mostly = 1;
  72
  73atomic_long_t num_poisoned_pages __read_mostly = ATOMIC_LONG_INIT(0);
  74
  75static bool hw_memory_failure __read_mostly = false;
  76
  77static DEFINE_MUTEX(mf_mutex);
  78
  79void num_poisoned_pages_inc(unsigned long pfn)
  80{
  81	atomic_long_inc(&num_poisoned_pages);
  82	memblk_nr_poison_inc(pfn);
  83}
  84
  85void num_poisoned_pages_sub(unsigned long pfn, long i)
  86{
  87	atomic_long_sub(i, &num_poisoned_pages);
  88	if (pfn != -1UL)
  89		memblk_nr_poison_sub(pfn, i);
  90}
  91
  92/**
  93 * MF_ATTR_RO - Create sysfs entry for each memory failure statistics.
  94 * @_name: name of the file in the per NUMA sysfs directory.
  95 */
  96#define MF_ATTR_RO(_name)					\
  97static ssize_t _name##_show(struct device *dev,			\
  98			    struct device_attribute *attr,	\
  99			    char *buf)				\
 100{								\
 101	struct memory_failure_stats *mf_stats =			\
 102		&NODE_DATA(dev->id)->mf_stats;			\
 103	return sysfs_emit(buf, "%lu\n", mf_stats->_name);	\
 104}								\
 105static DEVICE_ATTR_RO(_name)
 106
 107MF_ATTR_RO(total);
 108MF_ATTR_RO(ignored);
 109MF_ATTR_RO(failed);
 110MF_ATTR_RO(delayed);
 111MF_ATTR_RO(recovered);
 112
 113static struct attribute *memory_failure_attr[] = {
 114	&dev_attr_total.attr,
 115	&dev_attr_ignored.attr,
 116	&dev_attr_failed.attr,
 117	&dev_attr_delayed.attr,
 118	&dev_attr_recovered.attr,
 119	NULL,
 120};
 121
 122const struct attribute_group memory_failure_attr_group = {
 123	.name = "memory_failure",
 124	.attrs = memory_failure_attr,
 125};
 126
 127static struct ctl_table memory_failure_table[] = {
 128	{
 129		.procname	= "memory_failure_early_kill",
 130		.data		= &sysctl_memory_failure_early_kill,
 131		.maxlen		= sizeof(sysctl_memory_failure_early_kill),
 132		.mode		= 0644,
 133		.proc_handler	= proc_dointvec_minmax,
 134		.extra1		= SYSCTL_ZERO,
 135		.extra2		= SYSCTL_ONE,
 136	},
 137	{
 138		.procname	= "memory_failure_recovery",
 139		.data		= &sysctl_memory_failure_recovery,
 140		.maxlen		= sizeof(sysctl_memory_failure_recovery),
 141		.mode		= 0644,
 142		.proc_handler	= proc_dointvec_minmax,
 143		.extra1		= SYSCTL_ZERO,
 144		.extra2		= SYSCTL_ONE,
 145	},
 146	{
 147		.procname	= "enable_soft_offline",
 148		.data		= &sysctl_enable_soft_offline,
 149		.maxlen		= sizeof(sysctl_enable_soft_offline),
 150		.mode		= 0644,
 151		.proc_handler	= proc_dointvec_minmax,
 152		.extra1		= SYSCTL_ZERO,
 153		.extra2		= SYSCTL_ONE,
 154	}
 155};
 156
 157/*
 158 * Return values:
 159 *   1:   the page is dissolved (if needed) and taken off from buddy,
 160 *   0:   the page is dissolved (if needed) and not taken off from buddy,
 161 *   < 0: failed to dissolve.
 162 */
 163static int __page_handle_poison(struct page *page)
 164{
 165	int ret;
 166
 167	/*
 168	 * zone_pcp_disable() can't be used here. It will
 169	 * hold pcp_batch_high_lock and dissolve_free_hugetlb_folio() might hold
 170	 * cpu_hotplug_lock via static_key_slow_dec() when hugetlb vmemmap
 171	 * optimization is enabled. This will break current lock dependency
 172	 * chain and leads to deadlock.
 173	 * Disabling pcp before dissolving the page was a deterministic
 174	 * approach because we made sure that those pages cannot end up in any
 175	 * PCP list. Draining PCP lists expels those pages to the buddy system,
 176	 * but nothing guarantees that those pages do not get back to a PCP
 177	 * queue if we need to refill those.
 178	 */
 179	ret = dissolve_free_hugetlb_folio(page_folio(page));
 180	if (!ret) {
 181		drain_all_pages(page_zone(page));
 182		ret = take_page_off_buddy(page);
 183	}
 184
 185	return ret;
 186}
 187
 188static bool page_handle_poison(struct page *page, bool hugepage_or_freepage, bool release)
 189{
 190	if (hugepage_or_freepage) {
 191		/*
 192		 * Doing this check for free pages is also fine since
 193		 * dissolve_free_hugetlb_folio() returns 0 for non-hugetlb folios as well.
 194		 */
 195		if (__page_handle_poison(page) <= 0)
 196			/*
 197			 * We could fail to take off the target page from buddy
 198			 * for example due to racy page allocation, but that's
 199			 * acceptable because soft-offlined page is not broken
 200			 * and if someone really want to use it, they should
 201			 * take it.
 202			 */
 203			return false;
 204	}
 205
 206	SetPageHWPoison(page);
 207	if (release)
 208		put_page(page);
 209	page_ref_inc(page);
 210	num_poisoned_pages_inc(page_to_pfn(page));
 211
 212	return true;
 213}
 214
 215#if IS_ENABLED(CONFIG_HWPOISON_INJECT)
 216
 217u32 hwpoison_filter_enable = 0;
 218u32 hwpoison_filter_dev_major = ~0U;
 219u32 hwpoison_filter_dev_minor = ~0U;
 220u64 hwpoison_filter_flags_mask;
 221u64 hwpoison_filter_flags_value;
 222EXPORT_SYMBOL_GPL(hwpoison_filter_enable);
 223EXPORT_SYMBOL_GPL(hwpoison_filter_dev_major);
 224EXPORT_SYMBOL_GPL(hwpoison_filter_dev_minor);
 225EXPORT_SYMBOL_GPL(hwpoison_filter_flags_mask);
 226EXPORT_SYMBOL_GPL(hwpoison_filter_flags_value);
 227
 228static int hwpoison_filter_dev(struct page *p)
 229{
 230	struct folio *folio = page_folio(p);
 231	struct address_space *mapping;
 232	dev_t dev;
 233
 234	if (hwpoison_filter_dev_major == ~0U &&
 235	    hwpoison_filter_dev_minor == ~0U)
 236		return 0;
 237
 238	mapping = folio_mapping(folio);
 
 
 
 
 
 
 239	if (mapping == NULL || mapping->host == NULL)
 240		return -EINVAL;
 241
 242	dev = mapping->host->i_sb->s_dev;
 243	if (hwpoison_filter_dev_major != ~0U &&
 244	    hwpoison_filter_dev_major != MAJOR(dev))
 245		return -EINVAL;
 246	if (hwpoison_filter_dev_minor != ~0U &&
 247	    hwpoison_filter_dev_minor != MINOR(dev))
 248		return -EINVAL;
 249
 250	return 0;
 251}
 252
 253static int hwpoison_filter_flags(struct page *p)
 254{
 255	if (!hwpoison_filter_flags_mask)
 256		return 0;
 257
 258	if ((stable_page_flags(p) & hwpoison_filter_flags_mask) ==
 259				    hwpoison_filter_flags_value)
 260		return 0;
 261	else
 262		return -EINVAL;
 263}
 264
 265/*
 266 * This allows stress tests to limit test scope to a collection of tasks
 267 * by putting them under some memcg. This prevents killing unrelated/important
 268 * processes such as /sbin/init. Note that the target task may share clean
 269 * pages with init (eg. libc text), which is harmless. If the target task
 270 * share _dirty_ pages with another task B, the test scheme must make sure B
 271 * is also included in the memcg. At last, due to race conditions this filter
 272 * can only guarantee that the page either belongs to the memcg tasks, or is
 273 * a freed page.
 274 */
 275#ifdef CONFIG_MEMCG
 276u64 hwpoison_filter_memcg;
 277EXPORT_SYMBOL_GPL(hwpoison_filter_memcg);
 278static int hwpoison_filter_task(struct page *p)
 279{
 
 
 
 
 280	if (!hwpoison_filter_memcg)
 281		return 0;
 282
 283	if (page_cgroup_ino(p) != hwpoison_filter_memcg)
 
 
 
 
 
 
 
 
 
 
 
 
 284		return -EINVAL;
 285
 286	return 0;
 287}
 288#else
 289static int hwpoison_filter_task(struct page *p) { return 0; }
 290#endif
 291
 292int hwpoison_filter(struct page *p)
 293{
 294	if (!hwpoison_filter_enable)
 295		return 0;
 296
 297	if (hwpoison_filter_dev(p))
 298		return -EINVAL;
 299
 300	if (hwpoison_filter_flags(p))
 301		return -EINVAL;
 302
 303	if (hwpoison_filter_task(p))
 304		return -EINVAL;
 305
 306	return 0;
 307}
 308EXPORT_SYMBOL_GPL(hwpoison_filter);
 309#else
 310int hwpoison_filter(struct page *p)
 311{
 312	return 0;
 313}
 314#endif
 315
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 316/*
 317 * Kill all processes that have a poisoned page mapped and then isolate
 318 * the page.
 319 *
 320 * General strategy:
 321 * Find all processes having the page mapped and kill them.
 322 * But we keep a page reference around so that the page is not
 323 * actually freed yet.
 324 * Then stash the page away
 325 *
 326 * There's no convenient way to get back to mapped processes
 327 * from the VMAs. So do a brute-force search over all
 328 * running processes.
 329 *
 330 * Remember that machine checks are not common (or rather
 331 * if they are common you have other problems), so this shouldn't
 332 * be a performance issue.
 333 *
 334 * Also there are some races possible while we get from the
 335 * error detection to actually handle it.
 336 */
 337
 338struct to_kill {
 339	struct list_head nd;
 340	struct task_struct *tsk;
 341	unsigned long addr;
 342	short size_shift;
 343};
 344
 345/*
 346 * Send all the processes who have the page mapped a signal.
 347 * ``action optional'' if they are not immediately affected by the error
 348 * ``action required'' if error happened in current execution context
 349 */
 350static int kill_proc(struct to_kill *tk, unsigned long pfn, int flags)
 351{
 352	struct task_struct *t = tk->tsk;
 353	short addr_lsb = tk->size_shift;
 354	int ret = 0;
 355
 356	pr_err("%#lx: Sending SIGBUS to %s:%d due to hardware memory corruption\n",
 357			pfn, t->comm, task_pid_nr(t));
 358
 359	if ((flags & MF_ACTION_REQUIRED) && (t == current))
 360		ret = force_sig_mceerr(BUS_MCEERR_AR,
 361				 (void __user *)tk->addr, addr_lsb);
 362	else
 363		/*
 364		 * Signal other processes sharing the page if they have
 365		 * PF_MCE_EARLY set.
 366		 * Don't use force here, it's convenient if the signal
 367		 * can be temporarily blocked.
 368		 */
 369		ret = send_sig_mceerr(BUS_MCEERR_AO, (void __user *)tk->addr,
 370				      addr_lsb, t);
 371	if (ret < 0)
 372		pr_info("Error sending signal to %s:%d: %d\n",
 373			t->comm, task_pid_nr(t), ret);
 374	return ret;
 375}
 376
 377/*
 378 * Unknown page type encountered. Try to check whether it can turn PageLRU by
 379 * lru_add_drain_all.
 380 */
 381void shake_folio(struct folio *folio)
 382{
 383	if (folio_test_hugetlb(folio))
 384		return;
 385	/*
 386	 * TODO: Could shrink slab caches here if a lightweight range-based
 387	 * shrinker will be available.
 388	 */
 389	if (folio_test_slab(folio))
 390		return;
 391
 392	lru_add_drain_all();
 393}
 394EXPORT_SYMBOL_GPL(shake_folio);
 395
 396static void shake_page(struct page *page)
 397{
 398	shake_folio(page_folio(page));
 399}
 400
 401static unsigned long dev_pagemap_mapping_shift(struct vm_area_struct *vma,
 402		unsigned long address)
 403{
 404	unsigned long ret = 0;
 405	pgd_t *pgd;
 406	p4d_t *p4d;
 407	pud_t *pud;
 408	pmd_t *pmd;
 409	pte_t *pte;
 410	pte_t ptent;
 411
 412	VM_BUG_ON_VMA(address == -EFAULT, vma);
 413	pgd = pgd_offset(vma->vm_mm, address);
 414	if (!pgd_present(*pgd))
 415		return 0;
 416	p4d = p4d_offset(pgd, address);
 417	if (!p4d_present(*p4d))
 418		return 0;
 419	pud = pud_offset(p4d, address);
 420	if (!pud_present(*pud))
 421		return 0;
 422	if (pud_devmap(*pud))
 423		return PUD_SHIFT;
 424	pmd = pmd_offset(pud, address);
 425	if (!pmd_present(*pmd))
 426		return 0;
 427	if (pmd_devmap(*pmd))
 428		return PMD_SHIFT;
 429	pte = pte_offset_map(pmd, address);
 430	if (!pte)
 431		return 0;
 432	ptent = ptep_get(pte);
 433	if (pte_present(ptent) && pte_devmap(ptent))
 434		ret = PAGE_SHIFT;
 435	pte_unmap(pte);
 436	return ret;
 437}
 438
 439/*
 440 * Failure handling: if we can't find or can't kill a process there's
 441 * not much we can do.	We just print a message and ignore otherwise.
 442 */
 443
 444/*
 445 * Schedule a process for later kill.
 446 * Uses GFP_ATOMIC allocations to avoid potential recursions in the VM.
 
 447 */
 448static void __add_to_kill(struct task_struct *tsk, const struct page *p,
 449			  struct vm_area_struct *vma, struct list_head *to_kill,
 450			  unsigned long addr)
 
 451{
 452	struct to_kill *tk;
 453
 454	tk = kmalloc(sizeof(struct to_kill), GFP_ATOMIC);
 455	if (!tk) {
 456		pr_err("Out of memory while machine check handling\n");
 457		return;
 
 
 
 
 
 
 458	}
 459
 460	tk->addr = addr;
 461	if (is_zone_device_page(p))
 462		tk->size_shift = dev_pagemap_mapping_shift(vma, tk->addr);
 463	else
 464		tk->size_shift = folio_shift(page_folio(p));
 465
 466	/*
 467	 * Send SIGKILL if "tk->addr == -EFAULT". Also, as
 468	 * "tk->size_shift" is always non-zero for !is_zone_device_page(),
 469	 * so "tk->size_shift == 0" effectively checks no mapping on
 470	 * ZONE_DEVICE. Indeed, when a devdax page is mmapped N times
 471	 * to a process' address space, it's possible not all N VMAs
 472	 * contain mappings for the page, but at least one VMA does.
 473	 * Only deliver SIGBUS with payload derived from the VMA that
 474	 * has a mapping for the page.
 475	 */
 476	if (tk->addr == -EFAULT) {
 477		pr_info("Unable to find user space address %lx in %s\n",
 478			page_to_pfn(p), tsk->comm);
 479	} else if (tk->size_shift == 0) {
 480		kfree(tk);
 481		return;
 482	}
 483
 484	get_task_struct(tsk);
 485	tk->tsk = tsk;
 486	list_add_tail(&tk->nd, to_kill);
 487}
 488
 489static void add_to_kill_anon_file(struct task_struct *tsk, const struct page *p,
 490		struct vm_area_struct *vma, struct list_head *to_kill,
 491		unsigned long addr)
 492{
 493	if (addr == -EFAULT)
 494		return;
 495	__add_to_kill(tsk, p, vma, to_kill, addr);
 496}
 497
 498#ifdef CONFIG_KSM
 499static bool task_in_to_kill_list(struct list_head *to_kill,
 500				 struct task_struct *tsk)
 501{
 502	struct to_kill *tk, *next;
 503
 504	list_for_each_entry_safe(tk, next, to_kill, nd) {
 505		if (tk->tsk == tsk)
 506			return true;
 507	}
 508
 509	return false;
 510}
 511
 512void add_to_kill_ksm(struct task_struct *tsk, const struct page *p,
 513		     struct vm_area_struct *vma, struct list_head *to_kill,
 514		     unsigned long addr)
 515{
 516	if (!task_in_to_kill_list(to_kill, tsk))
 517		__add_to_kill(tsk, p, vma, to_kill, addr);
 518}
 519#endif
 520/*
 521 * Kill the processes that have been collected earlier.
 522 *
 523 * Only do anything when FORCEKILL is set, otherwise just free the
 524 * list (this is used for clean pages which do not need killing)
 
 
 525 */
 526static void kill_procs(struct list_head *to_kill, int forcekill,
 527		unsigned long pfn, int flags)
 528{
 529	struct to_kill *tk, *next;
 530
 531	list_for_each_entry_safe(tk, next, to_kill, nd) {
 532		if (forcekill) {
 533			if (tk->addr == -EFAULT) {
 534				pr_err("%#lx: forcibly killing %s:%d because of failure to unmap corrupted page\n",
 535				       pfn, tk->tsk->comm, task_pid_nr(tk->tsk));
 536				do_send_sig_info(SIGKILL, SEND_SIG_PRIV,
 537						 tk->tsk, PIDTYPE_PID);
 
 
 
 
 
 538			}
 539
 540			/*
 541			 * In theory the process could have mapped
 542			 * something else on the address in-between. We could
 543			 * check for that, but we need to tell the
 544			 * process anyways.
 545			 */
 546			else if (kill_proc(tk, pfn, flags) < 0)
 547				pr_err("%#lx: Cannot send advisory machine check signal to %s:%d\n",
 548				       pfn, tk->tsk->comm, task_pid_nr(tk->tsk));
 
 
 549		}
 550		list_del(&tk->nd);
 551		put_task_struct(tk->tsk);
 552		kfree(tk);
 553	}
 554}
 555
 556/*
 557 * Find a dedicated thread which is supposed to handle SIGBUS(BUS_MCEERR_AO)
 558 * on behalf of the thread group. Return task_struct of the (first found)
 559 * dedicated thread if found, and return NULL otherwise.
 560 *
 561 * We already hold rcu lock in the caller, so we don't have to call
 562 * rcu_read_lock/unlock() in this function.
 563 */
 564static struct task_struct *find_early_kill_thread(struct task_struct *tsk)
 565{
 566	struct task_struct *t;
 567
 568	for_each_thread(tsk, t) {
 569		if (t->flags & PF_MCE_PROCESS) {
 570			if (t->flags & PF_MCE_EARLY)
 571				return t;
 572		} else {
 573			if (sysctl_memory_failure_early_kill)
 574				return t;
 575		}
 576	}
 577	return NULL;
 578}
 579
 580/*
 581 * Determine whether a given process is "early kill" process which expects
 582 * to be signaled when some page under the process is hwpoisoned.
 583 * Return task_struct of the dedicated thread (main thread unless explicitly
 584 * specified) if the process is "early kill" and otherwise returns NULL.
 585 *
 586 * Note that the above is true for Action Optional case. For Action Required
 587 * case, it's only meaningful to the current thread which need to be signaled
 588 * with SIGBUS, this error is Action Optional for other non current
 589 * processes sharing the same error page,if the process is "early kill", the
 590 * task_struct of the dedicated thread will also be returned.
 591 */
 592struct task_struct *task_early_kill(struct task_struct *tsk, int force_early)
 593{
 594	if (!tsk->mm)
 595		return NULL;
 596	/*
 597	 * Comparing ->mm here because current task might represent
 598	 * a subthread, while tsk always points to the main thread.
 599	 */
 600	if (force_early && tsk->mm == current->mm)
 601		return current;
 602
 603	return find_early_kill_thread(tsk);
 604}
 605
 606/*
 607 * Collect processes when the error hit an anonymous page.
 608 */
 609static void collect_procs_anon(const struct folio *folio,
 610		const struct page *page, struct list_head *to_kill,
 611		int force_early)
 612{
 
 613	struct task_struct *tsk;
 614	struct anon_vma *av;
 615	pgoff_t pgoff;
 616
 617	av = folio_lock_anon_vma_read(folio, NULL);
 618	if (av == NULL)	/* Not actually mapped anymore */
 619		return;
 620
 621	pgoff = page_pgoff(folio, page);
 622	rcu_read_lock();
 623	for_each_process(tsk) {
 624		struct vm_area_struct *vma;
 625		struct anon_vma_chain *vmac;
 626		struct task_struct *t = task_early_kill(tsk, force_early);
 627		unsigned long addr;
 628
 629		if (!t)
 630			continue;
 631		anon_vma_interval_tree_foreach(vmac, &av->rb_root,
 632					       pgoff, pgoff) {
 633			vma = vmac->vma;
 634			if (vma->vm_mm != t->mm)
 635				continue;
 636			addr = page_mapped_in_vma(page, vma);
 637			add_to_kill_anon_file(t, page, vma, to_kill, addr);
 638		}
 639	}
 640	rcu_read_unlock();
 641	anon_vma_unlock_read(av);
 642}
 643
 644/*
 645 * Collect processes when the error hit a file mapped page.
 646 */
 647static void collect_procs_file(const struct folio *folio,
 648		const struct page *page, struct list_head *to_kill,
 649		int force_early)
 650{
 651	struct vm_area_struct *vma;
 652	struct task_struct *tsk;
 653	struct address_space *mapping = folio->mapping;
 654	pgoff_t pgoff;
 655
 656	i_mmap_lock_read(mapping);
 657	rcu_read_lock();
 658	pgoff = page_pgoff(folio, page);
 659	for_each_process(tsk) {
 660		struct task_struct *t = task_early_kill(tsk, force_early);
 661		unsigned long addr;
 662
 663		if (!t)
 664			continue;
 665		vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff,
 
 666				      pgoff) {
 667			/*
 668			 * Send early kill signal to tasks where a vma covers
 669			 * the page but the corrupted page is not necessarily
 670			 * mapped in its pte.
 671			 * Assume applications who requested early kill want
 672			 * to be informed of all such data corruptions.
 673			 */
 674			if (vma->vm_mm != t->mm)
 675				continue;
 676			addr = page_address_in_vma(folio, page, vma);
 677			add_to_kill_anon_file(t, page, vma, to_kill, addr);
 678		}
 679	}
 680	rcu_read_unlock();
 681	i_mmap_unlock_read(mapping);
 682}
 683
 684#ifdef CONFIG_FS_DAX
 685static void add_to_kill_fsdax(struct task_struct *tsk, const struct page *p,
 686			      struct vm_area_struct *vma,
 687			      struct list_head *to_kill, pgoff_t pgoff)
 688{
 689	unsigned long addr = vma_address(vma, pgoff, 1);
 690	__add_to_kill(tsk, p, vma, to_kill, addr);
 691}
 692
 693/*
 694 * Collect processes when the error hit a fsdax page.
 
 
 
 695 */
 696static void collect_procs_fsdax(const struct page *page,
 697		struct address_space *mapping, pgoff_t pgoff,
 698		struct list_head *to_kill, bool pre_remove)
 699{
 700	struct vm_area_struct *vma;
 701	struct task_struct *tsk;
 702
 703	i_mmap_lock_read(mapping);
 704	rcu_read_lock();
 705	for_each_process(tsk) {
 706		struct task_struct *t = tsk;
 707
 708		/*
 709		 * Search for all tasks while MF_MEM_PRE_REMOVE is set, because
 710		 * the current may not be the one accessing the fsdax page.
 711		 * Otherwise, search for the current task.
 712		 */
 713		if (!pre_remove)
 714			t = task_early_kill(tsk, true);
 715		if (!t)
 716			continue;
 717		vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
 718			if (vma->vm_mm == t->mm)
 719				add_to_kill_fsdax(t, page, vma, to_kill, pgoff);
 720		}
 721	}
 722	rcu_read_unlock();
 723	i_mmap_unlock_read(mapping);
 724}
 725#endif /* CONFIG_FS_DAX */
 726
 727/*
 728 * Collect the processes who have the corrupted page mapped to kill.
 729 */
 730static void collect_procs(const struct folio *folio, const struct page *page,
 731		struct list_head *tokill, int force_early)
 732{
 733	if (!folio->mapping)
 734		return;
 735	if (unlikely(folio_test_ksm(folio)))
 736		collect_procs_ksm(folio, page, tokill, force_early);
 737	else if (folio_test_anon(folio))
 738		collect_procs_anon(folio, page, tokill, force_early);
 739	else
 740		collect_procs_file(folio, page, tokill, force_early);
 741}
 742
 743struct hwpoison_walk {
 744	struct to_kill tk;
 745	unsigned long pfn;
 746	int flags;
 747};
 748
 749static void set_to_kill(struct to_kill *tk, unsigned long addr, short shift)
 750{
 751	tk->addr = addr;
 752	tk->size_shift = shift;
 753}
 754
 755static int check_hwpoisoned_entry(pte_t pte, unsigned long addr, short shift,
 756				unsigned long poisoned_pfn, struct to_kill *tk)
 757{
 758	unsigned long pfn = 0;
 759
 760	if (pte_present(pte)) {
 761		pfn = pte_pfn(pte);
 762	} else {
 763		swp_entry_t swp = pte_to_swp_entry(pte);
 764
 765		if (is_hwpoison_entry(swp))
 766			pfn = swp_offset_pfn(swp);
 767	}
 768
 769	if (!pfn || pfn != poisoned_pfn)
 770		return 0;
 771
 772	set_to_kill(tk, addr, shift);
 773	return 1;
 774}
 775
 776#ifdef CONFIG_TRANSPARENT_HUGEPAGE
 777static int check_hwpoisoned_pmd_entry(pmd_t *pmdp, unsigned long addr,
 778				      struct hwpoison_walk *hwp)
 779{
 780	pmd_t pmd = *pmdp;
 781	unsigned long pfn;
 782	unsigned long hwpoison_vaddr;
 783
 784	if (!pmd_present(pmd))
 785		return 0;
 786	pfn = pmd_pfn(pmd);
 787	if (pfn <= hwp->pfn && hwp->pfn < pfn + HPAGE_PMD_NR) {
 788		hwpoison_vaddr = addr + ((hwp->pfn - pfn) << PAGE_SHIFT);
 789		set_to_kill(&hwp->tk, hwpoison_vaddr, PAGE_SHIFT);
 790		return 1;
 791	}
 792	return 0;
 793}
 794#else
 795static int check_hwpoisoned_pmd_entry(pmd_t *pmdp, unsigned long addr,
 796				      struct hwpoison_walk *hwp)
 797{
 798	return 0;
 799}
 800#endif
 801
 802static int hwpoison_pte_range(pmd_t *pmdp, unsigned long addr,
 803			      unsigned long end, struct mm_walk *walk)
 804{
 805	struct hwpoison_walk *hwp = walk->private;
 806	int ret = 0;
 807	pte_t *ptep, *mapped_pte;
 808	spinlock_t *ptl;
 809
 810	ptl = pmd_trans_huge_lock(pmdp, walk->vma);
 811	if (ptl) {
 812		ret = check_hwpoisoned_pmd_entry(pmdp, addr, hwp);
 813		spin_unlock(ptl);
 814		goto out;
 815	}
 816
 817	mapped_pte = ptep = pte_offset_map_lock(walk->vma->vm_mm, pmdp,
 818						addr, &ptl);
 819	if (!ptep)
 820		goto out;
 821
 822	for (; addr != end; ptep++, addr += PAGE_SIZE) {
 823		ret = check_hwpoisoned_entry(ptep_get(ptep), addr, PAGE_SHIFT,
 824					     hwp->pfn, &hwp->tk);
 825		if (ret == 1)
 826			break;
 827	}
 828	pte_unmap_unlock(mapped_pte, ptl);
 829out:
 830	cond_resched();
 831	return ret;
 832}
 833
 834#ifdef CONFIG_HUGETLB_PAGE
 835static int hwpoison_hugetlb_range(pte_t *ptep, unsigned long hmask,
 836			    unsigned long addr, unsigned long end,
 837			    struct mm_walk *walk)
 838{
 839	struct hwpoison_walk *hwp = walk->private;
 840	pte_t pte = huge_ptep_get(walk->mm, addr, ptep);
 841	struct hstate *h = hstate_vma(walk->vma);
 842
 843	return check_hwpoisoned_entry(pte, addr, huge_page_shift(h),
 844				      hwp->pfn, &hwp->tk);
 845}
 846#else
 847#define hwpoison_hugetlb_range	NULL
 848#endif
 849
 850static const struct mm_walk_ops hwpoison_walk_ops = {
 851	.pmd_entry = hwpoison_pte_range,
 852	.hugetlb_entry = hwpoison_hugetlb_range,
 853	.walk_lock = PGWALK_RDLOCK,
 854};
 855
 856/*
 857 * Sends SIGBUS to the current process with error info.
 858 *
 859 * This function is intended to handle "Action Required" MCEs on already
 860 * hardware poisoned pages. They could happen, for example, when
 861 * memory_failure() failed to unmap the error page at the first call, or
 862 * when multiple local machine checks happened on different CPUs.
 863 *
 864 * MCE handler currently has no easy access to the error virtual address,
 865 * so this function walks page table to find it. The returned virtual address
 866 * is proper in most cases, but it could be wrong when the application
 867 * process has multiple entries mapping the error page.
 868 */
 869static int kill_accessing_process(struct task_struct *p, unsigned long pfn,
 870				  int flags)
 871{
 872	int ret;
 873	struct hwpoison_walk priv = {
 874		.pfn = pfn,
 875	};
 876	priv.tk.tsk = p;
 877
 878	if (!p->mm)
 879		return -EFAULT;
 
 
 
 
 880
 881	mmap_read_lock(p->mm);
 882	ret = walk_page_range(p->mm, 0, TASK_SIZE, &hwpoison_walk_ops,
 883			      (void *)&priv);
 884	if (ret == 1 && priv.tk.addr)
 885		kill_proc(&priv.tk, pfn, flags);
 886	else
 887		ret = 0;
 888	mmap_read_unlock(p->mm);
 889	return ret > 0 ? -EHWPOISON : -EFAULT;
 890}
 891
 892/*
 893 * MF_IGNORED - The m-f() handler marks the page as PG_hwpoisoned'ed.
 894 * But it could not do more to isolate the page from being accessed again,
 895 * nor does it kill the process. This is extremely rare and one of the
 896 * potential causes is that the page state has been changed due to
 897 * underlying race condition. This is the most severe outcomes.
 898 *
 899 * MF_FAILED - The m-f() handler marks the page as PG_hwpoisoned'ed.
 900 * It should have killed the process, but it can't isolate the page,
 901 * due to conditions such as extra pin, unmap failure, etc. Accessing
 902 * the page again may trigger another MCE and the process will be killed
 903 * by the m-f() handler immediately.
 904 *
 905 * MF_DELAYED - The m-f() handler marks the page as PG_hwpoisoned'ed.
 906 * The page is unmapped, and is removed from the LRU or file mapping.
 907 * An attempt to access the page again will trigger page fault and the
 908 * PF handler will kill the process.
 909 *
 910 * MF_RECOVERED - The m-f() handler marks the page as PG_hwpoisoned'ed.
 911 * The page has been completely isolated, that is, unmapped, taken out of
 912 * the buddy system, or hole-punnched out of the file mapping.
 913 */
 914static const char *action_name[] = {
 915	[MF_IGNORED] = "Ignored",
 916	[MF_FAILED] = "Failed",
 917	[MF_DELAYED] = "Delayed",
 918	[MF_RECOVERED] = "Recovered",
 919};
 920
 921static const char * const action_page_types[] = {
 922	[MF_MSG_KERNEL]			= "reserved kernel page",
 923	[MF_MSG_KERNEL_HIGH_ORDER]	= "high-order kernel page",
 924	[MF_MSG_HUGE]			= "huge page",
 925	[MF_MSG_FREE_HUGE]		= "free huge page",
 926	[MF_MSG_GET_HWPOISON]		= "get hwpoison page",
 927	[MF_MSG_UNMAP_FAILED]		= "unmapping failed page",
 928	[MF_MSG_DIRTY_SWAPCACHE]	= "dirty swapcache page",
 929	[MF_MSG_CLEAN_SWAPCACHE]	= "clean swapcache page",
 930	[MF_MSG_DIRTY_MLOCKED_LRU]	= "dirty mlocked LRU page",
 931	[MF_MSG_CLEAN_MLOCKED_LRU]	= "clean mlocked LRU page",
 932	[MF_MSG_DIRTY_UNEVICTABLE_LRU]	= "dirty unevictable LRU page",
 933	[MF_MSG_CLEAN_UNEVICTABLE_LRU]	= "clean unevictable LRU page",
 934	[MF_MSG_DIRTY_LRU]		= "dirty LRU page",
 935	[MF_MSG_CLEAN_LRU]		= "clean LRU page",
 936	[MF_MSG_TRUNCATED_LRU]		= "already truncated LRU page",
 937	[MF_MSG_BUDDY]			= "free buddy page",
 938	[MF_MSG_DAX]			= "dax page",
 939	[MF_MSG_UNSPLIT_THP]		= "unsplit thp",
 940	[MF_MSG_ALREADY_POISONED]	= "already poisoned",
 941	[MF_MSG_UNKNOWN]		= "unknown page",
 942};
 943
 944/*
 945 * XXX: It is possible that a page is isolated from LRU cache,
 946 * and then kept in swap cache or failed to remove from page cache.
 947 * The page count will stop it from being freed by unpoison.
 948 * Stress tests should be aware of this memory leak problem.
 949 */
 950static int delete_from_lru_cache(struct folio *folio)
 951{
 952	if (folio_isolate_lru(folio)) {
 953		/*
 954		 * Clear sensible page flags, so that the buddy system won't
 955		 * complain when the folio is unpoison-and-freed.
 956		 */
 957		folio_clear_active(folio);
 958		folio_clear_unevictable(folio);
 959
 960		/*
 961		 * Poisoned page might never drop its ref count to 0 so we have
 962		 * to uncharge it manually from its memcg.
 963		 */
 964		mem_cgroup_uncharge(folio);
 965
 966		/*
 967		 * drop the refcount elevated by folio_isolate_lru()
 968		 */
 969		folio_put(folio);
 970		return 0;
 971	}
 972	return -EIO;
 973}
 974
 975static int truncate_error_folio(struct folio *folio, unsigned long pfn,
 976				struct address_space *mapping)
 977{
 978	int ret = MF_FAILED;
 979
 980	if (mapping->a_ops->error_remove_folio) {
 981		int err = mapping->a_ops->error_remove_folio(mapping, folio);
 982
 983		if (err != 0)
 984			pr_info("%#lx: Failed to punch page: %d\n", pfn, err);
 985		else if (!filemap_release_folio(folio, GFP_NOIO))
 986			pr_info("%#lx: failed to release buffers\n", pfn);
 987		else
 988			ret = MF_RECOVERED;
 989	} else {
 990		/*
 991		 * If the file system doesn't support it just invalidate
 992		 * This fails on dirty or anything with private pages
 993		 */
 994		if (mapping_evict_folio(mapping, folio))
 995			ret = MF_RECOVERED;
 996		else
 997			pr_info("%#lx: Failed to invalidate\n",	pfn);
 998	}
 999
1000	return ret;
1001}
1002
1003struct page_state {
1004	unsigned long mask;
1005	unsigned long res;
1006	enum mf_action_page_type type;
1007
1008	/* Callback ->action() has to unlock the relevant page inside it. */
1009	int (*action)(struct page_state *ps, struct page *p);
1010};
1011
1012/*
1013 * Return true if page is still referenced by others, otherwise return
1014 * false.
1015 *
1016 * The extra_pins is true when one extra refcount is expected.
1017 */
1018static bool has_extra_refcount(struct page_state *ps, struct page *p,
1019			       bool extra_pins)
1020{
1021	int count = page_count(p) - 1;
1022
1023	if (extra_pins)
1024		count -= folio_nr_pages(page_folio(p));
1025
1026	if (count > 0) {
1027		pr_err("%#lx: %s still referenced by %d users\n",
1028		       page_to_pfn(p), action_page_types[ps->type], count);
1029		return true;
1030	}
1031
1032	return false;
1033}
1034
1035/*
1036 * Error hit kernel page.
1037 * Do nothing, try to be lucky and not touch this instead. For a few cases we
1038 * could be more sophisticated.
1039 */
1040static int me_kernel(struct page_state *ps, struct page *p)
1041{
1042	unlock_page(p);
1043	return MF_IGNORED;
1044}
1045
1046/*
1047 * Page in unknown state. Do nothing.
1048 * This is a catch-all in case we fail to make sense of the page state.
1049 */
1050static int me_unknown(struct page_state *ps, struct page *p)
1051{
1052	pr_err("%#lx: Unknown page state\n", page_to_pfn(p));
1053	unlock_page(p);
1054	return MF_IGNORED;
1055}
1056
1057/*
1058 * Clean (or cleaned) page cache page.
1059 */
1060static int me_pagecache_clean(struct page_state *ps, struct page *p)
1061{
1062	struct folio *folio = page_folio(p);
1063	int ret;
1064	struct address_space *mapping;
1065	bool extra_pins;
1066
1067	delete_from_lru_cache(folio);
1068
1069	/*
1070	 * For anonymous folios the only reference left
1071	 * should be the one m_f() holds.
1072	 */
1073	if (folio_test_anon(folio)) {
1074		ret = MF_RECOVERED;
1075		goto out;
1076	}
1077
1078	/*
1079	 * Now truncate the page in the page cache. This is really
1080	 * more like a "temporary hole punch"
1081	 * Don't do this for block devices when someone else
1082	 * has a reference, because it could be file system metadata
1083	 * and that's not safe to truncate.
1084	 */
1085	mapping = folio_mapping(folio);
1086	if (!mapping) {
1087		/* Folio has been torn down in the meantime */
1088		ret = MF_FAILED;
1089		goto out;
 
1090	}
1091
1092	/*
1093	 * The shmem page is kept in page cache instead of truncating
1094	 * so is expected to have an extra refcount after error-handling.
1095	 */
1096	extra_pins = shmem_mapping(mapping);
1097
1098	/*
1099	 * Truncation is a bit tricky. Enable it per file system for now.
1100	 *
1101	 * Open: to take i_rwsem or not for this? Right now we don't.
1102	 */
1103	ret = truncate_error_folio(folio, page_to_pfn(p), mapping);
1104	if (has_extra_refcount(ps, p, extra_pins))
1105		ret = MF_FAILED;
1106
1107out:
1108	folio_unlock(folio);
1109
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1110	return ret;
1111}
1112
1113/*
1114 * Dirty pagecache page
1115 * Issues: when the error hit a hole page the error is not properly
1116 * propagated.
1117 */
1118static int me_pagecache_dirty(struct page_state *ps, struct page *p)
1119{
1120	struct folio *folio = page_folio(p);
1121	struct address_space *mapping = folio_mapping(folio);
1122
 
1123	/* TBD: print more information about the file. */
1124	if (mapping) {
1125		/*
1126		 * IO error will be reported by write(), fsync(), etc.
1127		 * who check the mapping.
1128		 * This way the application knows that something went
1129		 * wrong with its dirty file data.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1130		 */
1131		mapping_set_error(mapping, -EIO);
1132	}
1133
1134	return me_pagecache_clean(ps, p);
1135}
1136
1137/*
1138 * Clean and dirty swap cache.
1139 *
1140 * Dirty swap cache page is tricky to handle. The page could live both in page
1141 * table and swap cache(ie. page is freshly swapped in). So it could be
1142 * referenced concurrently by 2 types of PTEs:
1143 * normal PTEs and swap PTEs. We try to handle them consistently by calling
1144 * try_to_unmap(!TTU_HWPOISON) to convert the normal PTEs to swap PTEs,
1145 * and then
1146 *      - clear dirty bit to prevent IO
1147 *      - remove from LRU
1148 *      - but keep in the swap cache, so that when we return to it on
1149 *        a later page fault, we know the application is accessing
1150 *        corrupted data and shall be killed (we installed simple
1151 *        interception code in do_swap_page to catch it).
1152 *
1153 * Clean swap cache pages can be directly isolated. A later page fault will
1154 * bring in the known good data from disk.
1155 */
1156static int me_swapcache_dirty(struct page_state *ps, struct page *p)
1157{
1158	struct folio *folio = page_folio(p);
1159	int ret;
1160	bool extra_pins = false;
1161
1162	folio_clear_dirty(folio);
1163	/* Trigger EIO in shmem: */
1164	folio_clear_uptodate(folio);
1165
1166	ret = delete_from_lru_cache(folio) ? MF_FAILED : MF_DELAYED;
1167	folio_unlock(folio);
1168
1169	if (ret == MF_DELAYED)
1170		extra_pins = true;
1171
1172	if (has_extra_refcount(ps, p, extra_pins))
1173		ret = MF_FAILED;
1174
1175	return ret;
1176}
1177
1178static int me_swapcache_clean(struct page_state *ps, struct page *p)
1179{
1180	struct folio *folio = page_folio(p);
1181	int ret;
1182
1183	delete_from_swap_cache(folio);
1184
1185	ret = delete_from_lru_cache(folio) ? MF_FAILED : MF_RECOVERED;
1186	folio_unlock(folio);
1187
1188	if (has_extra_refcount(ps, p, false))
1189		ret = MF_FAILED;
1190
1191	return ret;
1192}
1193
1194/*
1195 * Huge pages. Needs work.
1196 * Issues:
1197 * - Error on hugepage is contained in hugepage unit (not in raw page unit.)
1198 *   To narrow down kill region to one page, we need to break up pmd.
1199 */
1200static int me_huge_page(struct page_state *ps, struct page *p)
1201{
1202	struct folio *folio = page_folio(p);
1203	int res;
1204	struct address_space *mapping;
1205	bool extra_pins = false;
1206
1207	mapping = folio_mapping(folio);
1208	if (mapping) {
1209		res = truncate_error_folio(folio, page_to_pfn(p), mapping);
1210		/* The page is kept in page cache. */
1211		extra_pins = true;
1212		folio_unlock(folio);
1213	} else {
1214		folio_unlock(folio);
1215		/*
1216		 * migration entry prevents later access on error hugepage,
1217		 * so we can free and dissolve it into buddy to save healthy
1218		 * subpages.
1219		 */
1220		folio_put(folio);
1221		if (__page_handle_poison(p) > 0) {
1222			page_ref_inc(p);
1223			res = MF_RECOVERED;
1224		} else {
1225			res = MF_FAILED;
1226		}
1227	}
1228
1229	if (has_extra_refcount(ps, p, extra_pins))
1230		res = MF_FAILED;
1231
1232	return res;
1233}
1234
1235/*
1236 * Various page states we can handle.
1237 *
1238 * A page state is defined by its current page->flags bits.
1239 * The table matches them in order and calls the right handler.
1240 *
1241 * This is quite tricky because we can access page at any time
1242 * in its live cycle, so all accesses have to be extremely careful.
1243 *
1244 * This is not complete. More states could be added.
1245 * For any missing state don't attempt recovery.
1246 */
1247
1248#define dirty		(1UL << PG_dirty)
1249#define sc		((1UL << PG_swapcache) | (1UL << PG_swapbacked))
1250#define unevict		(1UL << PG_unevictable)
1251#define mlock		(1UL << PG_mlocked)
 
1252#define lru		(1UL << PG_lru)
 
1253#define head		(1UL << PG_head)
 
 
 
1254#define reserved	(1UL << PG_reserved)
1255
1256static struct page_state error_states[] = {
1257	{ reserved,	reserved,	MF_MSG_KERNEL,	me_kernel },
 
 
 
 
 
1258	/*
1259	 * free pages are specially detected outside this table:
1260	 * PG_buddy pages only make a small fraction of all free pages.
1261	 */
1262
1263	{ head,		head,		MF_MSG_HUGE,		me_huge_page },
 
 
 
 
 
 
 
 
 
 
 
 
1264
1265	{ sc|dirty,	sc|dirty,	MF_MSG_DIRTY_SWAPCACHE,	me_swapcache_dirty },
1266	{ sc|dirty,	sc,		MF_MSG_CLEAN_SWAPCACHE,	me_swapcache_clean },
1267
1268	{ mlock|dirty,	mlock|dirty,	MF_MSG_DIRTY_MLOCKED_LRU,	me_pagecache_dirty },
1269	{ mlock|dirty,	mlock,		MF_MSG_CLEAN_MLOCKED_LRU,	me_pagecache_clean },
1270
1271	{ unevict|dirty, unevict|dirty,	MF_MSG_DIRTY_UNEVICTABLE_LRU,	me_pagecache_dirty },
1272	{ unevict|dirty, unevict,	MF_MSG_CLEAN_UNEVICTABLE_LRU,	me_pagecache_clean },
1273
1274	{ lru|dirty,	lru|dirty,	MF_MSG_DIRTY_LRU,	me_pagecache_dirty },
1275	{ lru|dirty,	lru,		MF_MSG_CLEAN_LRU,	me_pagecache_clean },
1276
1277	/*
1278	 * Catchall entry: must be at end.
1279	 */
1280	{ 0,		0,		MF_MSG_UNKNOWN,	me_unknown },
1281};
1282
1283#undef dirty
1284#undef sc
1285#undef unevict
1286#undef mlock
 
1287#undef lru
 
1288#undef head
 
 
 
1289#undef reserved
1290
1291static void update_per_node_mf_stats(unsigned long pfn,
1292				     enum mf_result result)
1293{
1294	int nid = MAX_NUMNODES;
1295	struct memory_failure_stats *mf_stats = NULL;
1296
1297	nid = pfn_to_nid(pfn);
1298	if (unlikely(nid < 0 || nid >= MAX_NUMNODES)) {
1299		WARN_ONCE(1, "Memory failure: pfn=%#lx, invalid nid=%d", pfn, nid);
1300		return;
1301	}
1302
1303	mf_stats = &NODE_DATA(nid)->mf_stats;
1304	switch (result) {
1305	case MF_IGNORED:
1306		++mf_stats->ignored;
1307		break;
1308	case MF_FAILED:
1309		++mf_stats->failed;
1310		break;
1311	case MF_DELAYED:
1312		++mf_stats->delayed;
1313		break;
1314	case MF_RECOVERED:
1315		++mf_stats->recovered;
1316		break;
1317	default:
1318		WARN_ONCE(1, "Memory failure: mf_result=%d is not properly handled", result);
1319		break;
1320	}
1321	++mf_stats->total;
1322}
1323
1324/*
1325 * "Dirty/Clean" indication is not 100% accurate due to the possibility of
1326 * setting PG_dirty outside page lock. See also comment above set_page_dirty().
1327 */
1328static int action_result(unsigned long pfn, enum mf_action_page_type type,
1329			 enum mf_result result)
1330{
1331	trace_memory_failure_event(pfn, type, result);
1332
1333	num_poisoned_pages_inc(pfn);
1334
1335	update_per_node_mf_stats(pfn, result);
1336
1337	pr_err("%#lx: recovery action for %s: %s\n",
1338		pfn, action_page_types[type], action_name[result]);
1339
1340	return (result == MF_RECOVERED || result == MF_DELAYED) ? 0 : -EBUSY;
1341}
1342
1343static int page_action(struct page_state *ps, struct page *p,
1344			unsigned long pfn)
1345{
1346	int result;
 
 
 
 
1347
1348	/* page p should be unlocked after returning from ps->action().  */
1349	result = ps->action(ps, p);
 
 
 
 
 
 
 
1350
1351	/* Could do more checks here if page looks ok */
1352	/*
1353	 * Could adjust zone counters here to correct for the missing page.
1354	 */
1355
1356	return action_result(pfn, ps->type, result);
1357}
1358
1359static inline bool PageHWPoisonTakenOff(struct page *page)
1360{
1361	return PageHWPoison(page) && page_private(page) == MAGIC_HWPOISON;
1362}
1363
1364void SetPageHWPoisonTakenOff(struct page *page)
1365{
1366	set_page_private(page, MAGIC_HWPOISON);
1367}
1368
1369void ClearPageHWPoisonTakenOff(struct page *page)
1370{
1371	if (PageHWPoison(page))
1372		set_page_private(page, 0);
1373}
1374
1375/*
1376 * Return true if a page type of a given page is supported by hwpoison
1377 * mechanism (while handling could fail), otherwise false.  This function
1378 * does not return true for hugetlb or device memory pages, so it's assumed
1379 * to be called only in the context where we never have such pages.
1380 */
1381static inline bool HWPoisonHandlable(struct page *page, unsigned long flags)
 
1382{
1383	if (PageSlab(page))
1384		return false;
 
 
 
 
 
1385
1386	/* Soft offline could migrate non-LRU movable pages */
1387	if ((flags & MF_SOFT_OFFLINE) && __PageMovable(page))
1388		return true;
1389
1390	return PageLRU(page) || is_free_buddy_page(page);
1391}
1392
1393static int __get_hwpoison_page(struct page *page, unsigned long flags)
1394{
1395	struct folio *folio = page_folio(page);
1396	int ret = 0;
1397	bool hugetlb = false;
1398
1399	ret = get_hwpoison_hugetlb_folio(folio, &hugetlb, false);
1400	if (hugetlb) {
1401		/* Make sure hugetlb demotion did not happen from under us. */
1402		if (folio == page_folio(page))
1403			return ret;
1404		if (ret > 0) {
1405			folio_put(folio);
1406			folio = page_folio(page);
1407		}
1408	}
1409
1410	/*
1411	 * This check prevents from calling folio_try_get() for any
1412	 * unsupported type of folio in order to reduce the risk of unexpected
1413	 * races caused by taking a folio refcount.
1414	 */
1415	if (!HWPoisonHandlable(&folio->page, flags))
1416		return -EBUSY;
1417
1418	if (folio_try_get(folio)) {
1419		if (folio == page_folio(page))
1420			return 1;
1421
1422		pr_info("%#lx cannot catch tail\n", page_to_pfn(page));
1423		folio_put(folio);
1424	}
1425
1426	return 0;
1427}
1428
1429#define GET_PAGE_MAX_RETRY_NUM 3
1430
1431static int get_any_page(struct page *p, unsigned long flags)
1432{
1433	int ret = 0, pass = 0;
1434	bool count_increased = false;
1435
1436	if (flags & MF_COUNT_INCREASED)
1437		count_increased = true;
1438
1439try_again:
1440	if (!count_increased) {
1441		ret = __get_hwpoison_page(p, flags);
1442		if (!ret) {
1443			if (page_count(p)) {
1444				/* We raced with an allocation, retry. */
1445				if (pass++ < GET_PAGE_MAX_RETRY_NUM)
1446					goto try_again;
1447				ret = -EBUSY;
1448			} else if (!PageHuge(p) && !is_free_buddy_page(p)) {
1449				/* We raced with put_page, retry. */
1450				if (pass++ < GET_PAGE_MAX_RETRY_NUM)
1451					goto try_again;
1452				ret = -EIO;
1453			}
1454			goto out;
1455		} else if (ret == -EBUSY) {
1456			/*
1457			 * We raced with (possibly temporary) unhandlable
1458			 * page, retry.
1459			 */
1460			if (pass++ < 3) {
1461				shake_page(p);
1462				goto try_again;
1463			}
1464			ret = -EIO;
1465			goto out;
1466		}
1467	}
1468
1469	if (PageHuge(p) || HWPoisonHandlable(p, flags)) {
1470		ret = 1;
1471	} else {
1472		/*
1473		 * A page we cannot handle. Check whether we can turn
1474		 * it into something we can handle.
1475		 */
1476		if (pass++ < GET_PAGE_MAX_RETRY_NUM) {
1477			put_page(p);
1478			shake_page(p);
1479			count_increased = false;
1480			goto try_again;
1481		}
1482		put_page(p);
1483		ret = -EIO;
1484	}
1485out:
1486	if (ret == -EIO)
1487		pr_err("%#lx: unhandlable page.\n", page_to_pfn(p));
1488
1489	return ret;
1490}
1491
1492static int __get_unpoison_page(struct page *page)
1493{
1494	struct folio *folio = page_folio(page);
1495	int ret = 0;
1496	bool hugetlb = false;
1497
1498	ret = get_hwpoison_hugetlb_folio(folio, &hugetlb, true);
1499	if (hugetlb) {
1500		/* Make sure hugetlb demotion did not happen from under us. */
1501		if (folio == page_folio(page))
1502			return ret;
1503		if (ret > 0)
1504			folio_put(folio);
1505	}
1506
1507	/*
1508	 * PageHWPoisonTakenOff pages are not only marked as PG_hwpoison,
1509	 * but also isolated from buddy freelist, so need to identify the
1510	 * state and have to cancel both operations to unpoison.
1511	 */
1512	if (PageHWPoisonTakenOff(page))
1513		return -EHWPOISON;
1514
1515	return get_page_unless_zero(page) ? 1 : 0;
1516}
1517
1518/**
1519 * get_hwpoison_page() - Get refcount for memory error handling
1520 * @p:		Raw error page (hit by memory error)
1521 * @flags:	Flags controlling behavior of error handling
1522 *
1523 * get_hwpoison_page() takes a page refcount of an error page to handle memory
1524 * error on it, after checking that the error page is in a well-defined state
1525 * (defined as a page-type we can successfully handle the memory error on it,
1526 * such as LRU page and hugetlb page).
1527 *
1528 * Memory error handling could be triggered at any time on any type of page,
1529 * so it's prone to race with typical memory management lifecycle (like
1530 * allocation and free).  So to avoid such races, get_hwpoison_page() takes
1531 * extra care for the error page's state (as done in __get_hwpoison_page()),
1532 * and has some retry logic in get_any_page().
1533 *
1534 * When called from unpoison_memory(), the caller should already ensure that
1535 * the given page has PG_hwpoison. So it's never reused for other page
1536 * allocations, and __get_unpoison_page() never races with them.
1537 *
1538 * Return: 0 on failure or free buddy (hugetlb) page,
1539 *         1 on success for in-use pages in a well-defined state,
1540 *         -EIO for pages on which we can not handle memory errors,
1541 *         -EBUSY when get_hwpoison_page() has raced with page lifecycle
1542 *         operations like allocation and free,
1543 *         -EHWPOISON when the page is hwpoisoned and taken off from buddy.
1544 */
1545static int get_hwpoison_page(struct page *p, unsigned long flags)
1546{
1547	int ret;
1548
1549	zone_pcp_disable(page_zone(p));
1550	if (flags & MF_UNPOISON)
1551		ret = __get_unpoison_page(p);
1552	else
1553		ret = get_any_page(p, flags);
1554	zone_pcp_enable(page_zone(p));
1555
1556	return ret;
1557}
1558
1559int unmap_poisoned_folio(struct folio *folio, unsigned long pfn, bool must_kill)
1560{
1561	enum ttu_flags ttu = TTU_IGNORE_MLOCK | TTU_SYNC | TTU_HWPOISON;
1562	struct address_space *mapping;
1563
1564	if (folio_test_swapcache(folio)) {
1565		pr_err("%#lx: keeping poisoned page in swap cache\n", pfn);
1566		ttu &= ~TTU_HWPOISON;
1567	}
1568
1569	/*
1570	 * Propagate the dirty bit from PTEs to struct page first, because we
1571	 * need this to decide if we should kill or just drop the page.
1572	 * XXX: the dirty test could be racy: set_page_dirty() may not always
1573	 * be called inside page lock (it's recommended but not enforced).
1574	 */
1575	mapping = folio_mapping(folio);
1576	if (!must_kill && !folio_test_dirty(folio) && mapping &&
1577	    mapping_can_writeback(mapping)) {
1578		if (folio_mkclean(folio)) {
1579			folio_set_dirty(folio);
1580		} else {
1581			ttu &= ~TTU_HWPOISON;
1582			pr_info("%#lx: corrupted page was clean: dropped without side effects\n",
 
 
1583				pfn);
1584		}
1585	}
1586
1587	if (folio_test_hugetlb(folio) && !folio_test_anon(folio)) {
 
 
 
 
 
 
 
 
1588		/*
1589		 * For hugetlb folios in shared mappings, try_to_unmap
1590		 * could potentially call huge_pmd_unshare.  Because of
1591		 * this, take semaphore in write mode here and set
1592		 * TTU_RMAP_LOCKED to indicate we have taken the lock
1593		 * at this higher level.
 
 
 
1594		 */
1595		mapping = hugetlb_folio_mapping_lock_write(folio);
1596		if (!mapping) {
1597			pr_info("%#lx: could not lock mapping for mapped hugetlb folio\n",
1598				folio_pfn(folio));
1599			return -EBUSY;
 
 
 
 
 
 
 
 
 
 
 
1600		}
1601
1602		try_to_unmap(folio, ttu|TTU_RMAP_LOCKED);
1603		i_mmap_unlock_write(mapping);
1604	} else {
1605		try_to_unmap(folio, ttu);
1606	}
1607
1608	return folio_mapped(folio) ? -EBUSY : 0;
1609}
1610
1611/*
1612 * Do all that is necessary to remove user space mappings. Unmap
1613 * the pages and send SIGBUS to the processes if the data was dirty.
1614 */
1615static bool hwpoison_user_mappings(struct folio *folio, struct page *p,
1616		unsigned long pfn, int flags)
1617{
1618	LIST_HEAD(tokill);
1619	bool unmap_success;
1620	int forcekill;
1621	bool mlocked = folio_test_mlocked(folio);
1622
1623	/*
1624	 * Here we are interested only in user-mapped pages, so skip any
1625	 * other types of pages.
1626	 */
1627	if (folio_test_reserved(folio) || folio_test_slab(folio) ||
1628	    folio_test_pgtable(folio) || folio_test_offline(folio))
1629		return true;
1630	if (!(folio_test_lru(folio) || folio_test_hugetlb(folio)))
1631		return true;
1632
1633	/*
1634	 * This check implies we don't kill processes if their pages
1635	 * are in the swap cache early. Those are always late kills.
1636	 */
1637	if (!folio_mapped(folio))
1638		return true;
1639
1640	/*
1641	 * First collect all the processes that have the page
1642	 * mapped in dirty form.  This has to be done before try_to_unmap,
1643	 * because ttu takes the rmap data structures down.
 
 
 
1644	 */
1645	collect_procs(folio, p, &tokill, flags & MF_ACTION_REQUIRED);
 
1646
1647	unmap_success = !unmap_poisoned_folio(folio, pfn, flags & MF_MUST_KILL);
1648	if (!unmap_success)
1649		pr_err("%#lx: failed to unmap page (folio mapcount=%d)\n",
1650		       pfn, folio_mapcount(folio));
1651
1652	/*
1653	 * try_to_unmap() might put mlocked page in lru cache, so call
1654	 * shake_page() again to ensure that it's flushed.
1655	 */
1656	if (mlocked)
1657		shake_folio(folio);
 
1658
1659	/*
1660	 * Now that the dirty bit has been propagated to the
1661	 * struct page and all unmaps done we can decide if
1662	 * killing is needed or not.  Only kill when the page
1663	 * was dirty or the process is not restartable,
1664	 * otherwise the tokill list is merely
1665	 * freed.  When there was a problem unmapping earlier
1666	 * use a more force-full uncatchable kill to prevent
1667	 * any accesses to the poisoned memory.
1668	 */
1669	forcekill = folio_test_dirty(folio) || (flags & MF_MUST_KILL) ||
1670		    !unmap_success;
1671	kill_procs(&tokill, forcekill, pfn, flags);
1672
1673	return unmap_success;
1674}
1675
1676static int identify_page_state(unsigned long pfn, struct page *p,
1677				unsigned long page_flags)
1678{
1679	struct page_state *ps;
1680
1681	/*
1682	 * The first check uses the current page flags which may not have any
1683	 * relevant information. The second check with the saved page flags is
1684	 * carried out only if the first check can't determine the page status.
1685	 */
1686	for (ps = error_states;; ps++)
1687		if ((p->flags & ps->mask) == ps->res)
1688			break;
1689
1690	page_flags |= (p->flags & (1UL << PG_dirty));
1691
1692	if (!ps->mask)
1693		for (ps = error_states;; ps++)
1694			if ((page_flags & ps->mask) == ps->res)
1695				break;
1696	return page_action(ps, p, pfn);
1697}
1698
1699/*
1700 * When 'release' is 'false', it means that if thp split has failed,
1701 * there is still more to do, hence the page refcount we took earlier
1702 * is still needed.
1703 */
1704static int try_to_split_thp_page(struct page *page, bool release)
1705{
1706	int ret;
1707
1708	lock_page(page);
1709	ret = split_huge_page(page);
1710	unlock_page(page);
1711
1712	if (ret && release)
1713		put_page(page);
1714
1715	return ret;
1716}
1717
1718static void unmap_and_kill(struct list_head *to_kill, unsigned long pfn,
1719		struct address_space *mapping, pgoff_t index, int flags)
1720{
1721	struct to_kill *tk;
1722	unsigned long size = 0;
 
 
 
1723
1724	list_for_each_entry(tk, to_kill, nd)
1725		if (tk->size_shift)
1726			size = max(size, 1UL << tk->size_shift);
1727
1728	if (size) {
1729		/*
1730		 * Unmap the largest mapping to avoid breaking up device-dax
1731		 * mappings which are constant size. The actual size of the
1732		 * mapping being torn down is communicated in siginfo, see
1733		 * kill_proc()
1734		 */
1735		loff_t start = ((loff_t)index << PAGE_SHIFT) & ~(size - 1);
1736
1737		unmap_mapping_range(mapping, start, size, 0);
 
 
 
 
1738	}
1739
1740	kill_procs(to_kill, flags & MF_MUST_KILL, pfn, flags);
1741}
1742
1743/*
1744 * Only dev_pagemap pages get here, such as fsdax when the filesystem
1745 * either do not claim or fails to claim a hwpoison event, or devdax.
1746 * The fsdax pages are initialized per base page, and the devdax pages
1747 * could be initialized either as base pages, or as compound pages with
1748 * vmemmap optimization enabled. Devdax is simplistic in its dealing with
1749 * hwpoison, such that, if a subpage of a compound page is poisoned,
1750 * simply mark the compound head page is by far sufficient.
1751 */
1752static int mf_generic_kill_procs(unsigned long long pfn, int flags,
1753		struct dev_pagemap *pgmap)
1754{
1755	struct folio *folio = pfn_folio(pfn);
1756	LIST_HEAD(to_kill);
1757	dax_entry_t cookie;
1758	int rc = 0;
1759
1760	/*
1761	 * Prevent the inode from being freed while we are interrogating
1762	 * the address_space, typically this would be handled by
1763	 * lock_page(), but dax pages do not use the page lock. This
1764	 * also prevents changes to the mapping of this pfn until
1765	 * poison signaling is complete.
 
 
 
 
 
 
 
1766	 */
1767	cookie = dax_lock_folio(folio);
1768	if (!cookie)
1769		return -EBUSY;
1770
1771	if (hwpoison_filter(&folio->page)) {
1772		rc = -EOPNOTSUPP;
1773		goto unlock;
1774	}
1775
1776	switch (pgmap->type) {
1777	case MEMORY_DEVICE_PRIVATE:
1778	case MEMORY_DEVICE_COHERENT:
1779		/*
1780		 * TODO: Handle device pages which may need coordination
1781		 * with device-side memory.
1782		 */
1783		rc = -ENXIO;
1784		goto unlock;
1785	default:
1786		break;
 
 
 
 
 
 
 
1787	}
1788
1789	/*
1790	 * Use this flag as an indication that the dax page has been
1791	 * remapped UC to prevent speculative consumption of poison.
 
 
 
 
1792	 */
1793	SetPageHWPoison(&folio->page);
1794
1795	/*
1796	 * Unlike System-RAM there is no possibility to swap in a
1797	 * different physical page at a given virtual address, so all
1798	 * userspace consumption of ZONE_DEVICE memory necessitates
1799	 * SIGBUS (i.e. MF_MUST_KILL)
1800	 */
1801	flags |= MF_ACTION_REQUIRED | MF_MUST_KILL;
1802	collect_procs(folio, &folio->page, &to_kill, true);
1803
1804	unmap_and_kill(&to_kill, pfn, folio->mapping, folio->index, flags);
1805unlock:
1806	dax_unlock_folio(folio, cookie);
1807	return rc;
1808}
1809
1810#ifdef CONFIG_FS_DAX
1811/**
1812 * mf_dax_kill_procs - Collect and kill processes who are using this file range
1813 * @mapping:	address_space of the file in use
1814 * @index:	start pgoff of the range within the file
1815 * @count:	length of the range, in unit of PAGE_SIZE
1816 * @mf_flags:	memory failure flags
1817 */
1818int mf_dax_kill_procs(struct address_space *mapping, pgoff_t index,
1819		unsigned long count, int mf_flags)
1820{
1821	LIST_HEAD(to_kill);
1822	dax_entry_t cookie;
1823	struct page *page;
1824	size_t end = index + count;
1825	bool pre_remove = mf_flags & MF_MEM_PRE_REMOVE;
1826
1827	mf_flags |= MF_ACTION_REQUIRED | MF_MUST_KILL;
1828
1829	for (; index < end; index++) {
1830		page = NULL;
1831		cookie = dax_lock_mapping_entry(mapping, index, &page);
1832		if (!cookie)
1833			return -EBUSY;
1834		if (!page)
1835			goto unlock;
1836
1837		if (!pre_remove)
1838			SetPageHWPoison(page);
1839
1840		/*
1841		 * The pre_remove case is revoking access, the memory is still
1842		 * good and could theoretically be put back into service.
1843		 */
1844		collect_procs_fsdax(page, mapping, index, &to_kill, pre_remove);
1845		unmap_and_kill(&to_kill, page_to_pfn(page), mapping,
1846				index, mf_flags);
1847unlock:
1848		dax_unlock_mapping_entry(mapping, index, cookie);
1849	}
1850	return 0;
1851}
1852EXPORT_SYMBOL_GPL(mf_dax_kill_procs);
1853#endif /* CONFIG_FS_DAX */
1854
1855#ifdef CONFIG_HUGETLB_PAGE
1856
1857/*
1858 * Struct raw_hwp_page represents information about "raw error page",
1859 * constructing singly linked list from ->_hugetlb_hwpoison field of folio.
1860 */
1861struct raw_hwp_page {
1862	struct llist_node node;
1863	struct page *page;
1864};
1865
1866static inline struct llist_head *raw_hwp_list_head(struct folio *folio)
1867{
1868	return (struct llist_head *)&folio->_hugetlb_hwpoison;
1869}
1870
1871bool is_raw_hwpoison_page_in_hugepage(struct page *page)
1872{
1873	struct llist_head *raw_hwp_head;
1874	struct raw_hwp_page *p;
1875	struct folio *folio = page_folio(page);
1876	bool ret = false;
1877
1878	if (!folio_test_hwpoison(folio))
1879		return false;
1880
1881	if (!folio_test_hugetlb(folio))
1882		return PageHWPoison(page);
1883
1884	/*
1885	 * When RawHwpUnreliable is set, kernel lost track of which subpages
1886	 * are HWPOISON. So return as if ALL subpages are HWPOISONed.
 
1887	 */
1888	if (folio_test_hugetlb_raw_hwp_unreliable(folio))
1889		return true;
1890
1891	mutex_lock(&mf_mutex);
1892
1893	raw_hwp_head = raw_hwp_list_head(folio);
1894	llist_for_each_entry(p, raw_hwp_head->first, node) {
1895		if (page == p->page) {
1896			ret = true;
1897			break;
1898		}
1899	}
1900
1901	mutex_unlock(&mf_mutex);
1902
1903	return ret;
1904}
1905
1906static unsigned long __folio_free_raw_hwp(struct folio *folio, bool move_flag)
1907{
1908	struct llist_node *head;
1909	struct raw_hwp_page *p, *next;
1910	unsigned long count = 0;
1911
1912	head = llist_del_all(raw_hwp_list_head(folio));
1913	llist_for_each_entry_safe(p, next, head, node) {
1914		if (move_flag)
1915			SetPageHWPoison(p->page);
1916		else
1917			num_poisoned_pages_sub(page_to_pfn(p->page), 1);
1918		kfree(p);
1919		count++;
1920	}
1921	return count;
1922}
1923
1924static int folio_set_hugetlb_hwpoison(struct folio *folio, struct page *page)
1925{
1926	struct llist_head *head;
1927	struct raw_hwp_page *raw_hwp;
1928	struct raw_hwp_page *p;
1929	int ret = folio_test_set_hwpoison(folio) ? -EHWPOISON : 0;
1930
1931	/*
1932	 * Once the hwpoison hugepage has lost reliable raw error info,
1933	 * there is little meaning to keep additional error info precisely,
1934	 * so skip to add additional raw error info.
1935	 */
1936	if (folio_test_hugetlb_raw_hwp_unreliable(folio))
1937		return -EHWPOISON;
1938	head = raw_hwp_list_head(folio);
1939	llist_for_each_entry(p, head->first, node) {
1940		if (p->page == page)
1941			return -EHWPOISON;
1942	}
1943
1944	raw_hwp = kmalloc(sizeof(struct raw_hwp_page), GFP_ATOMIC);
1945	if (raw_hwp) {
1946		raw_hwp->page = page;
1947		llist_add(&raw_hwp->node, head);
1948		/* the first error event will be counted in action_result(). */
1949		if (ret)
1950			num_poisoned_pages_inc(page_to_pfn(page));
1951	} else {
1952		/*
1953		 * Failed to save raw error info.  We no longer trace all
1954		 * hwpoisoned subpages, and we need refuse to free/dissolve
1955		 * this hwpoisoned hugepage.
1956		 */
1957		folio_set_hugetlb_raw_hwp_unreliable(folio);
1958		/*
1959		 * Once hugetlb_raw_hwp_unreliable is set, raw_hwp_page is not
1960		 * used any more, so free it.
1961		 */
1962		__folio_free_raw_hwp(folio, false);
1963	}
1964	return ret;
1965}
1966
1967static unsigned long folio_free_raw_hwp(struct folio *folio, bool move_flag)
1968{
1969	/*
1970	 * hugetlb_vmemmap_optimized hugepages can't be freed because struct
1971	 * pages for tail pages are required but they don't exist.
1972	 */
1973	if (move_flag && folio_test_hugetlb_vmemmap_optimized(folio))
 
 
 
 
1974		return 0;
1975
1976	/*
1977	 * hugetlb_raw_hwp_unreliable hugepages shouldn't be unpoisoned by
1978	 * definition.
 
 
1979	 */
1980	if (folio_test_hugetlb_raw_hwp_unreliable(folio))
1981		return 0;
1982
1983	return __folio_free_raw_hwp(folio, move_flag);
1984}
1985
1986void folio_clear_hugetlb_hwpoison(struct folio *folio)
1987{
1988	if (folio_test_hugetlb_raw_hwp_unreliable(folio))
1989		return;
1990	if (folio_test_hugetlb_vmemmap_optimized(folio))
1991		return;
1992	folio_clear_hwpoison(folio);
1993	folio_free_raw_hwp(folio, true);
1994}
1995
1996/*
1997 * Called from hugetlb code with hugetlb_lock held.
1998 *
1999 * Return values:
2000 *   0             - free hugepage
2001 *   1             - in-use hugepage
2002 *   2             - not a hugepage
2003 *   -EBUSY        - the hugepage is busy (try to retry)
2004 *   -EHWPOISON    - the hugepage is already hwpoisoned
2005 */
2006int __get_huge_page_for_hwpoison(unsigned long pfn, int flags,
2007				 bool *migratable_cleared)
2008{
2009	struct page *page = pfn_to_page(pfn);
2010	struct folio *folio = page_folio(page);
2011	int ret = 2;	/* fallback to normal page handling */
2012	bool count_increased = false;
2013
2014	if (!folio_test_hugetlb(folio))
2015		goto out;
2016
2017	if (flags & MF_COUNT_INCREASED) {
2018		ret = 1;
2019		count_increased = true;
2020	} else if (folio_test_hugetlb_freed(folio)) {
2021		ret = 0;
2022	} else if (folio_test_hugetlb_migratable(folio)) {
2023		ret = folio_try_get(folio);
2024		if (ret)
2025			count_increased = true;
2026	} else {
2027		ret = -EBUSY;
2028		if (!(flags & MF_NO_RETRY))
2029			goto out;
2030	}
2031
2032	if (folio_set_hugetlb_hwpoison(folio, page)) {
2033		ret = -EHWPOISON;
2034		goto out;
2035	}
2036
2037	/*
2038	 * Clearing hugetlb_migratable for hwpoisoned hugepages to prevent them
2039	 * from being migrated by memory hotremove.
2040	 */
2041	if (count_increased && folio_test_hugetlb_migratable(folio)) {
2042		folio_clear_hugetlb_migratable(folio);
2043		*migratable_cleared = true;
 
2044	}
2045
2046	return ret;
2047out:
2048	if (count_increased)
2049		folio_put(folio);
2050	return ret;
2051}
2052
2053/*
2054 * Taking refcount of hugetlb pages needs extra care about race conditions
2055 * with basic operations like hugepage allocation/free/demotion.
2056 * So some of prechecks for hwpoison (pinning, and testing/setting
2057 * PageHWPoison) should be done in single hugetlb_lock range.
2058 */
2059static int try_memory_failure_hugetlb(unsigned long pfn, int flags, int *hugetlb)
2060{
2061	int res;
2062	struct page *p = pfn_to_page(pfn);
2063	struct folio *folio;
2064	unsigned long page_flags;
2065	bool migratable_cleared = false;
2066
2067	*hugetlb = 1;
2068retry:
2069	res = get_huge_page_for_hwpoison(pfn, flags, &migratable_cleared);
2070	if (res == 2) { /* fallback to normal page handling */
2071		*hugetlb = 0;
2072		return 0;
2073	} else if (res == -EHWPOISON) {
2074		pr_err("%#lx: already hardware poisoned\n", pfn);
2075		if (flags & MF_ACTION_REQUIRED) {
2076			folio = page_folio(p);
2077			res = kill_accessing_process(current, folio_pfn(folio), flags);
2078			action_result(pfn, MF_MSG_ALREADY_POISONED, MF_FAILED);
2079		}
2080		return res;
2081	} else if (res == -EBUSY) {
2082		if (!(flags & MF_NO_RETRY)) {
2083			flags |= MF_NO_RETRY;
2084			goto retry;
2085		}
2086		return action_result(pfn, MF_MSG_GET_HWPOISON, MF_IGNORED);
2087	}
2088
2089	folio = page_folio(p);
2090	folio_lock(folio);
2091
2092	if (hwpoison_filter(p)) {
2093		folio_clear_hugetlb_hwpoison(folio);
2094		if (migratable_cleared)
2095			folio_set_hugetlb_migratable(folio);
2096		folio_unlock(folio);
2097		if (res == 1)
2098			folio_put(folio);
2099		return -EOPNOTSUPP;
2100	}
2101
2102	/*
2103	 * Handling free hugepage.  The possible race with hugepage allocation
2104	 * or demotion can be prevented by PageHWPoison flag.
2105	 */
2106	if (res == 0) {
2107		folio_unlock(folio);
2108		if (__page_handle_poison(p) > 0) {
2109			page_ref_inc(p);
2110			res = MF_RECOVERED;
2111		} else {
2112			res = MF_FAILED;
2113		}
2114		return action_result(pfn, MF_MSG_FREE_HUGE, res);
2115	}
2116
2117	page_flags = folio->flags;
2118
2119	if (!hwpoison_user_mappings(folio, p, pfn, flags)) {
2120		folio_unlock(folio);
2121		return action_result(pfn, MF_MSG_UNMAP_FAILED, MF_FAILED);
2122	}
2123
2124	return identify_page_state(pfn, p, page_flags);
2125}
2126
2127#else
2128static inline int try_memory_failure_hugetlb(unsigned long pfn, int flags, int *hugetlb)
2129{
2130	return 0;
2131}
2132
2133static inline unsigned long folio_free_raw_hwp(struct folio *folio, bool flag)
2134{
2135	return 0;
2136}
2137#endif	/* CONFIG_HUGETLB_PAGE */
2138
2139/* Drop the extra refcount in case we come from madvise() */
2140static void put_ref_page(unsigned long pfn, int flags)
2141{
2142	if (!(flags & MF_COUNT_INCREASED))
2143		return;
2144
2145	put_page(pfn_to_page(pfn));
2146}
2147
2148static int memory_failure_dev_pagemap(unsigned long pfn, int flags,
2149		struct dev_pagemap *pgmap)
2150{
2151	int rc = -ENXIO;
2152
2153	/* device metadata space is not recoverable */
2154	if (!pgmap_pfn_valid(pgmap, pfn))
2155		goto out;
2156
2157	/*
2158	 * Call driver's implementation to handle the memory failure, otherwise
2159	 * fall back to generic handler.
2160	 */
2161	if (pgmap_has_memory_failure(pgmap)) {
2162		rc = pgmap->ops->memory_failure(pgmap, pfn, 1, flags);
2163		/*
2164		 * Fall back to generic handler too if operation is not
2165		 * supported inside the driver/device/filesystem.
2166		 */
2167		if (rc != -EOPNOTSUPP)
2168			goto out;
2169	}
2170
2171	rc = mf_generic_kill_procs(pfn, flags, pgmap);
2172out:
2173	/* drop pgmap ref acquired in caller */
2174	put_dev_pagemap(pgmap);
2175	if (rc != -EOPNOTSUPP)
2176		action_result(pfn, MF_MSG_DAX, rc ? MF_FAILED : MF_RECOVERED);
2177	return rc;
2178}
2179
2180/*
2181 * The calling condition is as such: thp split failed, page might have
2182 * been RDMA pinned, not much can be done for recovery.
2183 * But a SIGBUS should be delivered with vaddr provided so that the user
2184 * application has a chance to recover. Also, application processes'
2185 * election for MCE early killed will be honored.
2186 */
2187static void kill_procs_now(struct page *p, unsigned long pfn, int flags,
2188				struct folio *folio)
2189{
2190	LIST_HEAD(tokill);
2191
2192	collect_procs(folio, p, &tokill, flags & MF_ACTION_REQUIRED);
2193	kill_procs(&tokill, true, pfn, flags);
2194}
 
2195
2196/**
2197 * memory_failure - Handle memory failure of a page.
2198 * @pfn: Page Number of the corrupted page
2199 * @flags: fine tune action taken
2200 *
2201 * This function is called by the low level machine check code
2202 * of an architecture when it detects hardware memory corruption
2203 * of a page. It tries its best to recover, which includes
2204 * dropping pages, killing processes etc.
2205 *
2206 * The function is primarily of use for corruptions that
2207 * happen outside the current execution context (e.g. when
2208 * detected by a background scrubber)
2209 *
2210 * Must run in process context (e.g. a work queue) with interrupts
2211 * enabled and no spinlocks held.
2212 *
2213 * Return: 0 for successfully handled the memory error,
2214 *         -EOPNOTSUPP for hwpoison_filter() filtered the error event,
2215 *         < 0(except -EOPNOTSUPP) on failure.
2216 */
2217int memory_failure(unsigned long pfn, int flags)
2218{
2219	struct page *p;
2220	struct folio *folio;
2221	struct dev_pagemap *pgmap;
2222	int res = 0;
2223	unsigned long page_flags;
2224	bool retry = true;
2225	int hugetlb = 0;
2226
2227	if (!sysctl_memory_failure_recovery)
2228		panic("Memory failure on page %lx", pfn);
2229
2230	mutex_lock(&mf_mutex);
2231
2232	if (!(flags & MF_SW_SIMULATED))
2233		hw_memory_failure = true;
2234
2235	p = pfn_to_online_page(pfn);
2236	if (!p) {
2237		res = arch_memory_failure(pfn, flags);
2238		if (res == 0)
2239			goto unlock_mutex;
2240
2241		if (pfn_valid(pfn)) {
2242			pgmap = get_dev_pagemap(pfn, NULL);
2243			put_ref_page(pfn, flags);
2244			if (pgmap) {
2245				res = memory_failure_dev_pagemap(pfn, flags,
2246								 pgmap);
2247				goto unlock_mutex;
2248			}
2249		}
2250		pr_err("%#lx: memory outside kernel control\n", pfn);
2251		res = -ENXIO;
2252		goto unlock_mutex;
2253	}
2254
2255try_again:
2256	res = try_memory_failure_hugetlb(pfn, flags, &hugetlb);
2257	if (hugetlb)
2258		goto unlock_mutex;
2259
2260	if (TestSetPageHWPoison(p)) {
2261		pr_err("%#lx: already hardware poisoned\n", pfn);
2262		res = -EHWPOISON;
2263		if (flags & MF_ACTION_REQUIRED)
2264			res = kill_accessing_process(current, pfn, flags);
2265		if (flags & MF_COUNT_INCREASED)
2266			put_page(p);
2267		action_result(pfn, MF_MSG_ALREADY_POISONED, MF_FAILED);
2268		goto unlock_mutex;
2269	}
2270
2271	/*
2272	 * We need/can do nothing about count=0 pages.
2273	 * 1) it's a free page, and therefore in safe hand:
2274	 *    check_new_page() will be the gate keeper.
2275	 * 2) it's part of a non-compound high order page.
2276	 *    Implies some kernel user: cannot stop them from
2277	 *    R/W the page; let's pray that the page has been
2278	 *    used and will be freed some time later.
2279	 * In fact it's dangerous to directly bump up page count from 0,
2280	 * that may make page_ref_freeze()/page_ref_unfreeze() mismatch.
2281	 */
2282	if (!(flags & MF_COUNT_INCREASED)) {
2283		res = get_hwpoison_page(p, flags);
2284		if (!res) {
2285			if (is_free_buddy_page(p)) {
2286				if (take_page_off_buddy(p)) {
2287					page_ref_inc(p);
2288					res = MF_RECOVERED;
2289				} else {
2290					/* We lost the race, try again */
2291					if (retry) {
2292						ClearPageHWPoison(p);
2293						retry = false;
2294						goto try_again;
2295					}
2296					res = MF_FAILED;
2297				}
2298				res = action_result(pfn, MF_MSG_BUDDY, res);
2299			} else {
2300				res = action_result(pfn, MF_MSG_KERNEL_HIGH_ORDER, MF_IGNORED);
2301			}
2302			goto unlock_mutex;
2303		} else if (res < 0) {
2304			res = action_result(pfn, MF_MSG_GET_HWPOISON, MF_IGNORED);
2305			goto unlock_mutex;
2306		}
2307	}
2308
2309	folio = page_folio(p);
2310
2311	/* filter pages that are protected from hwpoison test by users */
2312	folio_lock(folio);
2313	if (hwpoison_filter(p)) {
2314		ClearPageHWPoison(p);
2315		folio_unlock(folio);
2316		folio_put(folio);
2317		res = -EOPNOTSUPP;
2318		goto unlock_mutex;
2319	}
2320	folio_unlock(folio);
2321
2322	if (folio_test_large(folio)) {
2323		/*
2324		 * The flag must be set after the refcount is bumped
2325		 * otherwise it may race with THP split.
2326		 * And the flag can't be set in get_hwpoison_page() since
2327		 * it is called by soft offline too and it is just called
2328		 * for !MF_COUNT_INCREASED.  So here seems to be the best
2329		 * place.
2330		 *
2331		 * Don't need care about the above error handling paths for
2332		 * get_hwpoison_page() since they handle either free page
2333		 * or unhandlable page.  The refcount is bumped iff the
2334		 * page is a valid handlable page.
2335		 */
2336		folio_set_has_hwpoisoned(folio);
2337		if (try_to_split_thp_page(p, false) < 0) {
2338			res = -EHWPOISON;
2339			kill_procs_now(p, pfn, flags, folio);
2340			put_page(p);
2341			action_result(pfn, MF_MSG_UNSPLIT_THP, MF_FAILED);
2342			goto unlock_mutex;
2343		}
2344		VM_BUG_ON_PAGE(!page_count(p), p);
2345		folio = page_folio(p);
2346	}
2347
2348	/*
2349	 * We ignore non-LRU pages for good reasons.
2350	 * - PG_locked is only well defined for LRU pages and a few others
2351	 * - to avoid races with __SetPageLocked()
2352	 * - to avoid races with __SetPageSlab*() (and more non-atomic ops)
2353	 * The check (unnecessarily) ignores LRU pages being isolated and
2354	 * walked by the page reclaim code, however that's not a big loss.
2355	 */
2356	shake_folio(folio);
2357
2358	folio_lock(folio);
2359
2360	/*
2361	 * We're only intended to deal with the non-Compound page here.
2362	 * The page cannot become compound pages again as folio has been
2363	 * splited and extra refcnt is held.
2364	 */
2365	WARN_ON(folio_test_large(folio));
2366
2367	/*
2368	 * We use page flags to determine what action should be taken, but
2369	 * the flags can be modified by the error containment action.  One
2370	 * example is an mlocked page, where PG_mlocked is cleared by
2371	 * folio_remove_rmap_*() in try_to_unmap_one(). So to determine page
2372	 * status correctly, we save a copy of the page flags at this time.
2373	 */
2374	page_flags = folio->flags;
2375
2376	/*
2377	 * __munlock_folio() may clear a writeback folio's LRU flag without
2378	 * the folio lock. We need to wait for writeback completion for this
2379	 * folio or it may trigger a vfs BUG while evicting inode.
2380	 */
2381	if (!folio_test_lru(folio) && !folio_test_writeback(folio))
2382		goto identify_page_state;
2383
2384	/*
2385	 * It's very difficult to mess with pages currently under IO
2386	 * and in many cases impossible, so we just avoid it here.
2387	 */
2388	folio_wait_writeback(folio);
2389
2390	/*
2391	 * Now take care of user space mappings.
2392	 * Abort on fail: __filemap_remove_folio() assumes unmapped page.
2393	 */
2394	if (!hwpoison_user_mappings(folio, p, pfn, flags)) {
2395		res = action_result(pfn, MF_MSG_UNMAP_FAILED, MF_FAILED);
2396		goto unlock_page;
2397	}
2398
2399	/*
2400	 * Torn down by someone else?
2401	 */
2402	if (folio_test_lru(folio) && !folio_test_swapcache(folio) &&
2403	    folio->mapping == NULL) {
2404		res = action_result(pfn, MF_MSG_TRUNCATED_LRU, MF_IGNORED);
2405		goto unlock_page;
2406	}
2407
2408identify_page_state:
2409	res = identify_page_state(pfn, p, page_flags);
2410	mutex_unlock(&mf_mutex);
2411	return res;
2412unlock_page:
2413	folio_unlock(folio);
2414unlock_mutex:
2415	mutex_unlock(&mf_mutex);
2416	return res;
2417}
2418EXPORT_SYMBOL_GPL(memory_failure);
2419
2420#define MEMORY_FAILURE_FIFO_ORDER	4
2421#define MEMORY_FAILURE_FIFO_SIZE	(1 << MEMORY_FAILURE_FIFO_ORDER)
2422
2423struct memory_failure_entry {
2424	unsigned long pfn;
 
2425	int flags;
2426};
2427
2428struct memory_failure_cpu {
2429	DECLARE_KFIFO(fifo, struct memory_failure_entry,
2430		      MEMORY_FAILURE_FIFO_SIZE);
2431	raw_spinlock_t lock;
2432	struct work_struct work;
2433};
2434
2435static DEFINE_PER_CPU(struct memory_failure_cpu, memory_failure_cpu);
2436
2437/**
2438 * memory_failure_queue - Schedule handling memory failure of a page.
2439 * @pfn: Page Number of the corrupted page
 
2440 * @flags: Flags for memory failure handling
2441 *
2442 * This function is called by the low level hardware error handler
2443 * when it detects hardware memory corruption of a page. It schedules
2444 * the recovering of error page, including dropping pages, killing
2445 * processes etc.
2446 *
2447 * The function is primarily of use for corruptions that
2448 * happen outside the current execution context (e.g. when
2449 * detected by a background scrubber)
2450 *
2451 * Can run in IRQ context.
2452 */
2453void memory_failure_queue(unsigned long pfn, int flags)
2454{
2455	struct memory_failure_cpu *mf_cpu;
2456	unsigned long proc_flags;
2457	bool buffer_overflow;
2458	struct memory_failure_entry entry = {
2459		.pfn =		pfn,
 
2460		.flags =	flags,
2461	};
2462
2463	mf_cpu = &get_cpu_var(memory_failure_cpu);
2464	raw_spin_lock_irqsave(&mf_cpu->lock, proc_flags);
2465	buffer_overflow = !kfifo_put(&mf_cpu->fifo, entry);
2466	if (!buffer_overflow)
2467		schedule_work_on(smp_processor_id(), &mf_cpu->work);
2468	raw_spin_unlock_irqrestore(&mf_cpu->lock, proc_flags);
 
 
 
2469	put_cpu_var(memory_failure_cpu);
2470	if (buffer_overflow)
2471		pr_err("buffer overflow when queuing memory failure at %#lx\n",
2472		       pfn);
2473}
2474EXPORT_SYMBOL_GPL(memory_failure_queue);
2475
2476static void memory_failure_work_func(struct work_struct *work)
2477{
2478	struct memory_failure_cpu *mf_cpu;
2479	struct memory_failure_entry entry = { 0, };
2480	unsigned long proc_flags;
2481	int gotten;
2482
2483	mf_cpu = container_of(work, struct memory_failure_cpu, work);
2484	for (;;) {
2485		raw_spin_lock_irqsave(&mf_cpu->lock, proc_flags);
2486		gotten = kfifo_get(&mf_cpu->fifo, &entry);
2487		raw_spin_unlock_irqrestore(&mf_cpu->lock, proc_flags);
2488		if (!gotten)
2489			break;
2490		if (entry.flags & MF_SOFT_OFFLINE)
2491			soft_offline_page(entry.pfn, entry.flags);
2492		else
2493			memory_failure(entry.pfn, entry.flags);
2494	}
2495}
2496
2497/*
2498 * Process memory_failure work queued on the specified CPU.
2499 * Used to avoid return-to-userspace racing with the memory_failure workqueue.
2500 */
2501void memory_failure_queue_kick(int cpu)
2502{
2503	struct memory_failure_cpu *mf_cpu;
2504
2505	mf_cpu = &per_cpu(memory_failure_cpu, cpu);
2506	cancel_work_sync(&mf_cpu->work);
2507	memory_failure_work_func(&mf_cpu->work);
2508}
2509
2510static int __init memory_failure_init(void)
2511{
2512	struct memory_failure_cpu *mf_cpu;
2513	int cpu;
2514
2515	for_each_possible_cpu(cpu) {
2516		mf_cpu = &per_cpu(memory_failure_cpu, cpu);
2517		raw_spin_lock_init(&mf_cpu->lock);
2518		INIT_KFIFO(mf_cpu->fifo);
2519		INIT_WORK(&mf_cpu->work, memory_failure_work_func);
2520	}
2521
2522	register_sysctl_init("vm", memory_failure_table);
2523
2524	return 0;
2525}
2526core_initcall(memory_failure_init);
2527
2528#undef pr_fmt
2529#define pr_fmt(fmt)	"Unpoison: " fmt
2530#define unpoison_pr_info(fmt, pfn, rs)			\
2531({							\
2532	if (__ratelimit(rs))				\
2533		pr_info(fmt, pfn);			\
2534})
2535
2536/**
2537 * unpoison_memory - Unpoison a previously poisoned page
2538 * @pfn: Page number of the to be unpoisoned page
2539 *
2540 * Software-unpoison a page that has been poisoned by
2541 * memory_failure() earlier.
2542 *
2543 * This is only done on the software-level, so it only works
2544 * for linux injected failures, not real hardware failures
2545 *
2546 * Returns 0 for success, otherwise -errno.
2547 */
2548int unpoison_memory(unsigned long pfn)
2549{
2550	struct folio *folio;
2551	struct page *p;
2552	int ret = -EBUSY, ghp;
2553	unsigned long count;
2554	bool huge = false;
2555	static DEFINE_RATELIMIT_STATE(unpoison_rs, DEFAULT_RATELIMIT_INTERVAL,
2556					DEFAULT_RATELIMIT_BURST);
2557
2558	if (!pfn_valid(pfn))
2559		return -ENXIO;
2560
2561	p = pfn_to_page(pfn);
2562	folio = page_folio(p);
2563
2564	mutex_lock(&mf_mutex);
 
 
 
2565
2566	if (hw_memory_failure) {
2567		unpoison_pr_info("%#lx: disabled after HW memory failure\n",
2568				 pfn, &unpoison_rs);
2569		ret = -EOPNOTSUPP;
2570		goto unlock_mutex;
2571	}
2572
2573	if (is_huge_zero_folio(folio)) {
2574		unpoison_pr_info("%#lx: huge zero page is not supported\n",
2575				 pfn, &unpoison_rs);
2576		ret = -EOPNOTSUPP;
2577		goto unlock_mutex;
 
 
 
 
 
2578	}
2579
2580	if (!PageHWPoison(p)) {
2581		unpoison_pr_info("%#lx: page was already unpoisoned\n",
2582				 pfn, &unpoison_rs);
2583		goto unlock_mutex;
 
 
 
 
 
 
 
 
 
2584	}
 
 
 
 
 
2585
2586	if (folio_ref_count(folio) > 1) {
2587		unpoison_pr_info("%#lx: someone grabs the hwpoison page\n",
2588				 pfn, &unpoison_rs);
2589		goto unlock_mutex;
2590	}
 
 
 
 
 
 
 
 
2591
2592	if (folio_test_slab(folio) || folio_test_pgtable(folio) ||
2593	    folio_test_reserved(folio) || folio_test_offline(folio))
2594		goto unlock_mutex;
 
 
 
 
 
 
2595
2596	if (folio_mapped(folio)) {
2597		unpoison_pr_info("%#lx: someone maps the hwpoison page\n",
2598				 pfn, &unpoison_rs);
2599		goto unlock_mutex;
2600	}
2601
2602	if (folio_mapping(folio)) {
2603		unpoison_pr_info("%#lx: the hwpoison page has non-NULL mapping\n",
2604				 pfn, &unpoison_rs);
2605		goto unlock_mutex;
2606	}
2607
2608	ghp = get_hwpoison_page(p, MF_UNPOISON);
2609	if (!ghp) {
2610		if (folio_test_hugetlb(folio)) {
2611			huge = true;
2612			count = folio_free_raw_hwp(folio, false);
2613			if (count == 0)
2614				goto unlock_mutex;
2615		}
2616		ret = folio_test_clear_hwpoison(folio) ? 0 : -EBUSY;
2617	} else if (ghp < 0) {
2618		if (ghp == -EHWPOISON) {
2619			ret = put_page_back_buddy(p) ? 0 : -EBUSY;
 
 
 
 
 
 
2620		} else {
2621			ret = ghp;
2622			unpoison_pr_info("%#lx: failed to grab page\n",
2623					 pfn, &unpoison_rs);
2624		}
2625	} else {
2626		if (folio_test_hugetlb(folio)) {
2627			huge = true;
2628			count = folio_free_raw_hwp(folio, false);
2629			if (count == 0) {
2630				folio_put(folio);
2631				goto unlock_mutex;
2632			}
2633		}
2634
2635		folio_put(folio);
2636		if (TestClearPageHWPoison(p)) {
2637			folio_put(folio);
2638			ret = 0;
2639		}
2640	}
2641
2642unlock_mutex:
2643	mutex_unlock(&mf_mutex);
2644	if (!ret) {
2645		if (!huge)
2646			num_poisoned_pages_sub(pfn, 1);
2647		unpoison_pr_info("%#lx: software-unpoisoned page\n",
2648				 page_to_pfn(p), &unpoison_rs);
2649	}
 
 
2650	return ret;
2651}
2652EXPORT_SYMBOL(unpoison_memory);
2653
2654#undef pr_fmt
2655#define pr_fmt(fmt) "Soft offline: " fmt
2656
2657/*
2658 * soft_offline_in_use_page handles hugetlb-pages and non-hugetlb pages.
2659 * If the page is a non-dirty unmapped page-cache page, it simply invalidates.
2660 * If the page is mapped, it migrates the contents over.
2661 */
2662static int soft_offline_in_use_page(struct page *page)
2663{
2664	long ret = 0;
2665	unsigned long pfn = page_to_pfn(page);
2666	struct folio *folio = page_folio(page);
2667	char const *msg_page[] = {"page", "hugepage"};
2668	bool huge = folio_test_hugetlb(folio);
2669	bool isolated;
2670	LIST_HEAD(pagelist);
2671	struct migration_target_control mtc = {
2672		.nid = NUMA_NO_NODE,
2673		.gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL,
2674		.reason = MR_MEMORY_FAILURE,
2675	};
2676
2677	if (!huge && folio_test_large(folio)) {
2678		if (try_to_split_thp_page(page, true)) {
2679			pr_info("%#lx: thp split failed\n", pfn);
2680			return -EBUSY;
2681		}
2682		folio = page_folio(page);
2683	}
2684
2685	folio_lock(folio);
2686	if (!huge)
2687		folio_wait_writeback(folio);
2688	if (PageHWPoison(page)) {
2689		folio_unlock(folio);
2690		folio_put(folio);
2691		pr_info("%#lx: page already poisoned\n", pfn);
2692		return 0;
2693	}
2694
2695	if (!huge && folio_test_lru(folio) && !folio_test_swapcache(folio))
2696		/*
2697		 * Try to invalidate first. This should work for
2698		 * non dirty unmapped page cache pages.
2699		 */
2700		ret = mapping_evict_folio(folio_mapping(folio), folio);
2701	folio_unlock(folio);
2702
 
 
 
2703	if (ret) {
2704		pr_info("%#lx: invalidated\n", pfn);
2705		page_handle_poison(page, false, true);
2706		return 0;
2707	}
2708
2709	isolated = isolate_folio_to_list(folio, &pagelist);
2710
2711	/*
2712	 * If we succeed to isolate the folio, we grabbed another refcount on
2713	 * the folio, so we can safely drop the one we got from get_any_page().
2714	 * If we failed to isolate the folio, it means that we cannot go further
2715	 * and we will return an error, so drop the reference we got from
2716	 * get_any_page() as well.
2717	 */
2718	folio_put(folio);
2719
2720	if (isolated) {
2721		ret = migrate_pages(&pagelist, alloc_migration_target, NULL,
2722			(unsigned long)&mtc, MIGRATE_SYNC, MR_MEMORY_FAILURE, NULL);
2723		if (!ret) {
2724			bool release = !huge;
2725
2726			if (!page_handle_poison(page, huge, release))
2727				ret = -EBUSY;
2728		} else {
2729			if (!list_empty(&pagelist))
2730				putback_movable_pages(&pagelist);
2731
2732			pr_info("%#lx: %s migration failed %ld, type %pGp\n",
2733				pfn, msg_page[huge], ret, &page->flags);
2734			if (ret > 0)
2735				ret = -EBUSY;
2736		}
2737	} else {
2738		pr_info("%#lx: %s isolation failed, page count %d, type %pGp\n",
2739			pfn, msg_page[huge], page_count(page), &page->flags);
2740		ret = -EBUSY;
2741	}
 
 
 
 
 
 
2742	return ret;
2743}
2744
2745/**
2746 * soft_offline_page - Soft offline a page.
2747 * @pfn: pfn to soft-offline
2748 * @flags: flags. Same as memory_failure().
2749 *
2750 * Returns 0 on success,
2751 *         -EOPNOTSUPP for hwpoison_filter() filtered the error event, or
2752 *         disabled by /proc/sys/vm/enable_soft_offline,
2753 *         < 0 otherwise negated errno.
2754 *
2755 * Soft offline a page, by migration or invalidation,
2756 * without killing anything. This is for the case when
2757 * a page is not corrupted yet (so it's still valid to access),
2758 * but has had a number of corrected errors and is better taken
2759 * out.
2760 *
2761 * The actual policy on when to do that is maintained by
2762 * user space.
2763 *
2764 * This should never impact any application or cause data loss,
2765 * however it might take some time.
2766 *
2767 * This is not a 100% solution for all memory, but tries to be
2768 * ``good enough'' for the majority of memory.
2769 */
2770int soft_offline_page(unsigned long pfn, int flags)
2771{
2772	int ret;
2773	bool try_again = true;
2774	struct page *page;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2775
2776	if (!pfn_valid(pfn)) {
2777		WARN_ON_ONCE(flags & MF_COUNT_INCREASED);
2778		return -ENXIO;
 
 
 
 
 
2779	}
2780
2781	/* Only online pages can be soft-offlined (esp., not ZONE_DEVICE). */
2782	page = pfn_to_online_page(pfn);
2783	if (!page) {
2784		put_ref_page(pfn, flags);
2785		return -EIO;
2786	}
2787
2788	if (!sysctl_enable_soft_offline) {
2789		pr_info_once("disabled by /proc/sys/vm/enable_soft_offline\n");
2790		put_ref_page(pfn, flags);
2791		return -EOPNOTSUPP;
2792	}
2793
2794	mutex_lock(&mf_mutex);
2795
 
 
 
2796	if (PageHWPoison(page)) {
2797		pr_info("%#lx: page already poisoned\n", pfn);
2798		put_ref_page(pfn, flags);
2799		mutex_unlock(&mf_mutex);
2800		return 0;
2801	}
2802
2803retry:
2804	get_online_mems();
2805	ret = get_hwpoison_page(page, flags | MF_SOFT_OFFLINE);
2806	put_online_mems();
2807
2808	if (hwpoison_filter(page)) {
2809		if (ret > 0)
2810			put_page(page);
2811
2812		mutex_unlock(&mf_mutex);
2813		return -EOPNOTSUPP;
 
 
 
 
2814	}
2815
2816	if (ret > 0) {
2817		ret = soft_offline_in_use_page(page);
2818	} else if (ret == 0) {
2819		if (!page_handle_poison(page, true, false)) {
2820			if (try_again) {
2821				try_again = false;
2822				flags &= ~MF_COUNT_INCREASED;
2823				goto retry;
2824			}
2825			ret = -EBUSY;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2826		}
 
 
 
2827	}
 
 
2828
2829	mutex_unlock(&mf_mutex);
2830
 
 
2831	return ret;
2832}