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/vm/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#include <linux/kernel.h>
  37#include <linux/mm.h>
  38#include <linux/page-flags.h>
  39#include <linux/kernel-page-flags.h>
  40#include <linux/sched/signal.h>
  41#include <linux/sched/task.h>
  42#include <linux/ksm.h>
  43#include <linux/rmap.h>
  44#include <linux/export.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/suspend.h>
  50#include <linux/slab.h>
  51#include <linux/swapops.h>
  52#include <linux/hugetlb.h>
  53#include <linux/memory_hotplug.h>
  54#include <linux/mm_inline.h>
  55#include <linux/memremap.h>
  56#include <linux/kfifo.h>
  57#include <linux/ratelimit.h>
  58#include <linux/page-isolation.h>
  59#include <linux/pagewalk.h>
  60#include "internal.h"
  61#include "ras/ras_event.h"
  62
  63int sysctl_memory_failure_early_kill __read_mostly = 0;
  64
  65int sysctl_memory_failure_recovery __read_mostly = 1;
  66
  67atomic_long_t num_poisoned_pages __read_mostly = ATOMIC_LONG_INIT(0);
  68
  69static bool __page_handle_poison(struct page *page)
  70{
  71	int ret;
  72
  73	zone_pcp_disable(page_zone(page));
  74	ret = dissolve_free_huge_page(page);
  75	if (!ret)
  76		ret = take_page_off_buddy(page);
  77	zone_pcp_enable(page_zone(page));
  78
  79	return ret > 0;
  80}
  81
  82static bool page_handle_poison(struct page *page, bool hugepage_or_freepage, bool release)
  83{
  84	if (hugepage_or_freepage) {
  85		/*
  86		 * Doing this check for free pages is also fine since dissolve_free_huge_page
  87		 * returns 0 for non-hugetlb pages as well.
  88		 */
  89		if (!__page_handle_poison(page))
  90			/*
  91			 * We could fail to take off the target page from buddy
  92			 * for example due to racy page allocation, but that's
  93			 * acceptable because soft-offlined page is not broken
  94			 * and if someone really want to use it, they should
  95			 * take it.
  96			 */
  97			return false;
  98	}
  99
 100	SetPageHWPoison(page);
 101	if (release)
 102		put_page(page);
 103	page_ref_inc(page);
 104	num_poisoned_pages_inc();
 105
 106	return true;
 107}
 108
 109#if defined(CONFIG_HWPOISON_INJECT) || defined(CONFIG_HWPOISON_INJECT_MODULE)
 110
 111u32 hwpoison_filter_enable = 0;
 112u32 hwpoison_filter_dev_major = ~0U;
 113u32 hwpoison_filter_dev_minor = ~0U;
 114u64 hwpoison_filter_flags_mask;
 115u64 hwpoison_filter_flags_value;
 116EXPORT_SYMBOL_GPL(hwpoison_filter_enable);
 117EXPORT_SYMBOL_GPL(hwpoison_filter_dev_major);
 118EXPORT_SYMBOL_GPL(hwpoison_filter_dev_minor);
 119EXPORT_SYMBOL_GPL(hwpoison_filter_flags_mask);
 120EXPORT_SYMBOL_GPL(hwpoison_filter_flags_value);
 121
 122static int hwpoison_filter_dev(struct page *p)
 123{
 124	struct address_space *mapping;
 125	dev_t dev;
 126
 127	if (hwpoison_filter_dev_major == ~0U &&
 128	    hwpoison_filter_dev_minor == ~0U)
 129		return 0;
 130
 131	/*
 132	 * page_mapping() does not accept slab pages.
 133	 */
 134	if (PageSlab(p))
 135		return -EINVAL;
 136
 137	mapping = page_mapping(p);
 138	if (mapping == NULL || mapping->host == NULL)
 139		return -EINVAL;
 140
 141	dev = mapping->host->i_sb->s_dev;
 142	if (hwpoison_filter_dev_major != ~0U &&
 143	    hwpoison_filter_dev_major != MAJOR(dev))
 144		return -EINVAL;
 145	if (hwpoison_filter_dev_minor != ~0U &&
 146	    hwpoison_filter_dev_minor != MINOR(dev))
 147		return -EINVAL;
 148
 149	return 0;
 150}
 151
 152static int hwpoison_filter_flags(struct page *p)
 153{
 154	if (!hwpoison_filter_flags_mask)
 155		return 0;
 156
 157	if ((stable_page_flags(p) & hwpoison_filter_flags_mask) ==
 158				    hwpoison_filter_flags_value)
 159		return 0;
 160	else
 161		return -EINVAL;
 162}
 163
 164/*
 165 * This allows stress tests to limit test scope to a collection of tasks
 166 * by putting them under some memcg. This prevents killing unrelated/important
 167 * processes such as /sbin/init. Note that the target task may share clean
 168 * pages with init (eg. libc text), which is harmless. If the target task
 169 * share _dirty_ pages with another task B, the test scheme must make sure B
 170 * is also included in the memcg. At last, due to race conditions this filter
 171 * can only guarantee that the page either belongs to the memcg tasks, or is
 172 * a freed page.
 173 */
 174#ifdef CONFIG_MEMCG
 175u64 hwpoison_filter_memcg;
 176EXPORT_SYMBOL_GPL(hwpoison_filter_memcg);
 177static int hwpoison_filter_task(struct page *p)
 178{
 179	if (!hwpoison_filter_memcg)
 180		return 0;
 181
 182	if (page_cgroup_ino(p) != hwpoison_filter_memcg)
 183		return -EINVAL;
 184
 185	return 0;
 186}
 187#else
 188static int hwpoison_filter_task(struct page *p) { return 0; }
 189#endif
 190
 191int hwpoison_filter(struct page *p)
 192{
 193	if (!hwpoison_filter_enable)
 194		return 0;
 195
 196	if (hwpoison_filter_dev(p))
 197		return -EINVAL;
 198
 199	if (hwpoison_filter_flags(p))
 200		return -EINVAL;
 201
 202	if (hwpoison_filter_task(p))
 203		return -EINVAL;
 204
 205	return 0;
 206}
 207#else
 208int hwpoison_filter(struct page *p)
 209{
 210	return 0;
 211}
 212#endif
 213
 214EXPORT_SYMBOL_GPL(hwpoison_filter);
 215
 216/*
 217 * Kill all processes that have a poisoned page mapped and then isolate
 218 * the page.
 219 *
 220 * General strategy:
 221 * Find all processes having the page mapped and kill them.
 222 * But we keep a page reference around so that the page is not
 223 * actually freed yet.
 224 * Then stash the page away
 225 *
 226 * There's no convenient way to get back to mapped processes
 227 * from the VMAs. So do a brute-force search over all
 228 * running processes.
 229 *
 230 * Remember that machine checks are not common (or rather
 231 * if they are common you have other problems), so this shouldn't
 232 * be a performance issue.
 233 *
 234 * Also there are some races possible while we get from the
 235 * error detection to actually handle it.
 236 */
 237
 238struct to_kill {
 239	struct list_head nd;
 240	struct task_struct *tsk;
 241	unsigned long addr;
 242	short size_shift;
 243};
 244
 245/*
 246 * Send all the processes who have the page mapped a signal.
 247 * ``action optional'' if they are not immediately affected by the error
 248 * ``action required'' if error happened in current execution context
 249 */
 250static int kill_proc(struct to_kill *tk, unsigned long pfn, int flags)
 251{
 252	struct task_struct *t = tk->tsk;
 253	short addr_lsb = tk->size_shift;
 254	int ret = 0;
 255
 256	pr_err("Memory failure: %#lx: Sending SIGBUS to %s:%d due to hardware memory corruption\n",
 257			pfn, t->comm, t->pid);
 258
 259	if (flags & MF_ACTION_REQUIRED) {
 260		if (t == current)
 261			ret = force_sig_mceerr(BUS_MCEERR_AR,
 262					 (void __user *)tk->addr, addr_lsb);
 263		else
 264			/* Signal other processes sharing the page if they have PF_MCE_EARLY set. */
 265			ret = send_sig_mceerr(BUS_MCEERR_AO, (void __user *)tk->addr,
 266				addr_lsb, t);
 267	} else {
 268		/*
 269		 * Don't use force here, it's convenient if the signal
 270		 * can be temporarily blocked.
 271		 * This could cause a loop when the user sets SIGBUS
 272		 * to SIG_IGN, but hopefully no one will do that?
 273		 */
 274		ret = send_sig_mceerr(BUS_MCEERR_AO, (void __user *)tk->addr,
 275				      addr_lsb, t);  /* synchronous? */
 276	}
 277	if (ret < 0)
 278		pr_info("Memory failure: Error sending signal to %s:%d: %d\n",
 279			t->comm, t->pid, ret);
 280	return ret;
 281}
 282
 283/*
 284 * Unknown page type encountered. Try to check whether it can turn PageLRU by
 285 * lru_add_drain_all, or a free page by reclaiming slabs when possible.
 286 */
 287void shake_page(struct page *p, int access)
 288{
 289	if (PageHuge(p))
 290		return;
 291
 292	if (!PageSlab(p)) {
 293		lru_add_drain_all();
 
 
 
 294		if (PageLRU(p) || is_free_buddy_page(p))
 295			return;
 296	}
 297
 298	/*
 299	 * Only call shrink_node_slabs here (which would also shrink
 300	 * other caches) if access is not potentially fatal.
 301	 */
 302	if (access)
 303		drop_slab_node(page_to_nid(p));
 304}
 305EXPORT_SYMBOL_GPL(shake_page);
 306
 307static unsigned long dev_pagemap_mapping_shift(struct page *page,
 308		struct vm_area_struct *vma)
 309{
 310	unsigned long address = vma_address(page, vma);
 311	pgd_t *pgd;
 312	p4d_t *p4d;
 313	pud_t *pud;
 314	pmd_t *pmd;
 315	pte_t *pte;
 316
 317	pgd = pgd_offset(vma->vm_mm, address);
 318	if (!pgd_present(*pgd))
 319		return 0;
 320	p4d = p4d_offset(pgd, address);
 321	if (!p4d_present(*p4d))
 322		return 0;
 323	pud = pud_offset(p4d, address);
 324	if (!pud_present(*pud))
 325		return 0;
 326	if (pud_devmap(*pud))
 327		return PUD_SHIFT;
 328	pmd = pmd_offset(pud, address);
 329	if (!pmd_present(*pmd))
 330		return 0;
 331	if (pmd_devmap(*pmd))
 332		return PMD_SHIFT;
 333	pte = pte_offset_map(pmd, address);
 334	if (!pte_present(*pte))
 335		return 0;
 336	if (pte_devmap(*pte))
 337		return PAGE_SHIFT;
 338	return 0;
 339}
 340
 341/*
 342 * Failure handling: if we can't find or can't kill a process there's
 343 * not much we can do.	We just print a message and ignore otherwise.
 344 */
 345
 346/*
 347 * Schedule a process for later kill.
 348 * Uses GFP_ATOMIC allocations to avoid potential recursions in the VM.
 
 349 */
 350static void add_to_kill(struct task_struct *tsk, struct page *p,
 351		       struct vm_area_struct *vma,
 352		       struct list_head *to_kill)
 
 353{
 354	struct to_kill *tk;
 355
 356	tk = kmalloc(sizeof(struct to_kill), GFP_ATOMIC);
 357	if (!tk) {
 358		pr_err("Memory failure: Out of memory while machine check handling\n");
 359		return;
 
 
 
 
 
 360	}
 361
 362	tk->addr = page_address_in_vma(p, vma);
 363	if (is_zone_device_page(p))
 364		tk->size_shift = dev_pagemap_mapping_shift(p, vma);
 365	else
 366		tk->size_shift = page_shift(compound_head(p));
 367
 368	/*
 369	 * Send SIGKILL if "tk->addr == -EFAULT". Also, as
 370	 * "tk->size_shift" is always non-zero for !is_zone_device_page(),
 371	 * so "tk->size_shift == 0" effectively checks no mapping on
 372	 * ZONE_DEVICE. Indeed, when a devdax page is mmapped N times
 373	 * to a process' address space, it's possible not all N VMAs
 374	 * contain mappings for the page, but at least one VMA does.
 375	 * Only deliver SIGBUS with payload derived from the VMA that
 376	 * has a mapping for the page.
 377	 */
 378	if (tk->addr == -EFAULT) {
 379		pr_info("Memory failure: Unable to find user space address %lx in %s\n",
 380			page_to_pfn(p), tsk->comm);
 381	} else if (tk->size_shift == 0) {
 382		kfree(tk);
 383		return;
 384	}
 385
 386	get_task_struct(tsk);
 387	tk->tsk = tsk;
 388	list_add_tail(&tk->nd, to_kill);
 389}
 390
 391/*
 392 * Kill the processes that have been collected earlier.
 393 *
 394 * Only do anything when DOIT is set, otherwise just free the list
 395 * (this is used for clean pages which do not need killing)
 396 * Also when FAIL is set do a force kill because something went
 397 * wrong earlier.
 398 */
 399static void kill_procs(struct list_head *to_kill, int forcekill, bool fail,
 400		unsigned long pfn, int flags)
 401{
 402	struct to_kill *tk, *next;
 403
 404	list_for_each_entry_safe (tk, next, to_kill, nd) {
 405		if (forcekill) {
 406			/*
 407			 * In case something went wrong with munmapping
 408			 * make sure the process doesn't catch the
 409			 * signal and then access the memory. Just kill it.
 410			 */
 411			if (fail || tk->addr == -EFAULT) {
 412				pr_err("Memory failure: %#lx: forcibly killing %s:%d because of failure to unmap corrupted page\n",
 413				       pfn, tk->tsk->comm, tk->tsk->pid);
 414				do_send_sig_info(SIGKILL, SEND_SIG_PRIV,
 415						 tk->tsk, PIDTYPE_PID);
 416			}
 417
 418			/*
 419			 * In theory the process could have mapped
 420			 * something else on the address in-between. We could
 421			 * check for that, but we need to tell the
 422			 * process anyways.
 423			 */
 424			else if (kill_proc(tk, pfn, flags) < 0)
 425				pr_err("Memory failure: %#lx: Cannot send advisory machine check signal to %s:%d\n",
 426				       pfn, tk->tsk->comm, tk->tsk->pid);
 427		}
 428		put_task_struct(tk->tsk);
 429		kfree(tk);
 430	}
 431}
 432
 433/*
 434 * Find a dedicated thread which is supposed to handle SIGBUS(BUS_MCEERR_AO)
 435 * on behalf of the thread group. Return task_struct of the (first found)
 436 * dedicated thread if found, and return NULL otherwise.
 437 *
 438 * We already hold read_lock(&tasklist_lock) in the caller, so we don't
 439 * have to call rcu_read_lock/unlock() in this function.
 440 */
 441static struct task_struct *find_early_kill_thread(struct task_struct *tsk)
 442{
 443	struct task_struct *t;
 444
 445	for_each_thread(tsk, t) {
 446		if (t->flags & PF_MCE_PROCESS) {
 447			if (t->flags & PF_MCE_EARLY)
 448				return t;
 449		} else {
 450			if (sysctl_memory_failure_early_kill)
 451				return t;
 452		}
 453	}
 454	return NULL;
 455}
 456
 457/*
 458 * Determine whether a given process is "early kill" process which expects
 459 * to be signaled when some page under the process is hwpoisoned.
 460 * Return task_struct of the dedicated thread (main thread unless explicitly
 461 * specified) if the process is "early kill" and otherwise returns NULL.
 462 *
 463 * Note that the above is true for Action Optional case. For Action Required
 464 * case, it's only meaningful to the current thread which need to be signaled
 465 * with SIGBUS, this error is Action Optional for other non current
 466 * processes sharing the same error page,if the process is "early kill", the
 467 * task_struct of the dedicated thread will also be returned.
 468 */
 469static struct task_struct *task_early_kill(struct task_struct *tsk,
 470					   int force_early)
 471{
 
 472	if (!tsk->mm)
 473		return NULL;
 474	/*
 475	 * Comparing ->mm here because current task might represent
 476	 * a subthread, while tsk always points to the main thread.
 477	 */
 478	if (force_early && tsk->mm == current->mm)
 479		return current;
 480
 481	return find_early_kill_thread(tsk);
 482}
 483
 484/*
 485 * Collect processes when the error hit an anonymous page.
 486 */
 487static void collect_procs_anon(struct page *page, struct list_head *to_kill,
 488				int force_early)
 489{
 490	struct vm_area_struct *vma;
 491	struct task_struct *tsk;
 492	struct anon_vma *av;
 493	pgoff_t pgoff;
 494
 495	av = page_lock_anon_vma_read(page);
 496	if (av == NULL)	/* Not actually mapped anymore */
 497		return;
 498
 499	pgoff = page_to_pgoff(page);
 500	read_lock(&tasklist_lock);
 501	for_each_process (tsk) {
 502		struct anon_vma_chain *vmac;
 503		struct task_struct *t = task_early_kill(tsk, force_early);
 504
 505		if (!t)
 506			continue;
 507		anon_vma_interval_tree_foreach(vmac, &av->rb_root,
 508					       pgoff, pgoff) {
 509			vma = vmac->vma;
 510			if (!page_mapped_in_vma(page, vma))
 511				continue;
 512			if (vma->vm_mm == t->mm)
 513				add_to_kill(t, page, vma, to_kill);
 514		}
 515	}
 516	read_unlock(&tasklist_lock);
 517	page_unlock_anon_vma_read(av);
 518}
 519
 520/*
 521 * Collect processes when the error hit a file mapped page.
 522 */
 523static void collect_procs_file(struct page *page, struct list_head *to_kill,
 524				int force_early)
 525{
 526	struct vm_area_struct *vma;
 527	struct task_struct *tsk;
 528	struct address_space *mapping = page->mapping;
 529	pgoff_t pgoff;
 530
 531	i_mmap_lock_read(mapping);
 532	read_lock(&tasklist_lock);
 533	pgoff = page_to_pgoff(page);
 534	for_each_process(tsk) {
 
 535		struct task_struct *t = task_early_kill(tsk, force_early);
 536
 537		if (!t)
 538			continue;
 539		vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff,
 540				      pgoff) {
 541			/*
 542			 * Send early kill signal to tasks where a vma covers
 543			 * the page but the corrupted page is not necessarily
 544			 * mapped it in its pte.
 545			 * Assume applications who requested early kill want
 546			 * to be informed of all such data corruptions.
 547			 */
 548			if (vma->vm_mm == t->mm)
 549				add_to_kill(t, page, vma, to_kill);
 550		}
 551	}
 552	read_unlock(&tasklist_lock);
 553	i_mmap_unlock_read(mapping);
 554}
 555
 556/*
 557 * Collect the processes who have the corrupted page mapped to kill.
 
 
 
 558 */
 559static void collect_procs(struct page *page, struct list_head *tokill,
 560				int force_early)
 561{
 
 
 562	if (!page->mapping)
 563		return;
 564
 
 
 
 565	if (PageAnon(page))
 566		collect_procs_anon(page, tokill, force_early);
 567	else
 568		collect_procs_file(page, tokill, force_early);
 569}
 570
 571struct hwp_walk {
 572	struct to_kill tk;
 573	unsigned long pfn;
 574	int flags;
 575};
 576
 577static void set_to_kill(struct to_kill *tk, unsigned long addr, short shift)
 578{
 579	tk->addr = addr;
 580	tk->size_shift = shift;
 581}
 582
 583static int check_hwpoisoned_entry(pte_t pte, unsigned long addr, short shift,
 584				unsigned long poisoned_pfn, struct to_kill *tk)
 585{
 586	unsigned long pfn = 0;
 587
 588	if (pte_present(pte)) {
 589		pfn = pte_pfn(pte);
 590	} else {
 591		swp_entry_t swp = pte_to_swp_entry(pte);
 592
 593		if (is_hwpoison_entry(swp))
 594			pfn = hwpoison_entry_to_pfn(swp);
 595	}
 596
 597	if (!pfn || pfn != poisoned_pfn)
 598		return 0;
 599
 600	set_to_kill(tk, addr, shift);
 601	return 1;
 602}
 603
 604#ifdef CONFIG_TRANSPARENT_HUGEPAGE
 605static int check_hwpoisoned_pmd_entry(pmd_t *pmdp, unsigned long addr,
 606				      struct hwp_walk *hwp)
 607{
 608	pmd_t pmd = *pmdp;
 609	unsigned long pfn;
 610	unsigned long hwpoison_vaddr;
 611
 612	if (!pmd_present(pmd))
 613		return 0;
 614	pfn = pmd_pfn(pmd);
 615	if (pfn <= hwp->pfn && hwp->pfn < pfn + HPAGE_PMD_NR) {
 616		hwpoison_vaddr = addr + ((hwp->pfn - pfn) << PAGE_SHIFT);
 617		set_to_kill(&hwp->tk, hwpoison_vaddr, PAGE_SHIFT);
 618		return 1;
 619	}
 620	return 0;
 621}
 622#else
 623static int check_hwpoisoned_pmd_entry(pmd_t *pmdp, unsigned long addr,
 624				      struct hwp_walk *hwp)
 625{
 626	return 0;
 627}
 628#endif
 629
 630static int hwpoison_pte_range(pmd_t *pmdp, unsigned long addr,
 631			      unsigned long end, struct mm_walk *walk)
 632{
 633	struct hwp_walk *hwp = (struct hwp_walk *)walk->private;
 634	int ret = 0;
 635	pte_t *ptep;
 636	spinlock_t *ptl;
 637
 638	ptl = pmd_trans_huge_lock(pmdp, walk->vma);
 639	if (ptl) {
 640		ret = check_hwpoisoned_pmd_entry(pmdp, addr, hwp);
 641		spin_unlock(ptl);
 642		goto out;
 643	}
 644
 645	if (pmd_trans_unstable(pmdp))
 646		goto out;
 647
 648	ptep = pte_offset_map_lock(walk->vma->vm_mm, pmdp, addr, &ptl);
 649	for (; addr != end; ptep++, addr += PAGE_SIZE) {
 650		ret = check_hwpoisoned_entry(*ptep, addr, PAGE_SHIFT,
 651					     hwp->pfn, &hwp->tk);
 652		if (ret == 1)
 653			break;
 654	}
 655	pte_unmap_unlock(ptep - 1, ptl);
 656out:
 657	cond_resched();
 658	return ret;
 659}
 660
 661#ifdef CONFIG_HUGETLB_PAGE
 662static int hwpoison_hugetlb_range(pte_t *ptep, unsigned long hmask,
 663			    unsigned long addr, unsigned long end,
 664			    struct mm_walk *walk)
 665{
 666	struct hwp_walk *hwp = (struct hwp_walk *)walk->private;
 667	pte_t pte = huge_ptep_get(ptep);
 668	struct hstate *h = hstate_vma(walk->vma);
 669
 670	return check_hwpoisoned_entry(pte, addr, huge_page_shift(h),
 671				      hwp->pfn, &hwp->tk);
 672}
 673#else
 674#define hwpoison_hugetlb_range	NULL
 675#endif
 676
 677static struct mm_walk_ops hwp_walk_ops = {
 678	.pmd_entry = hwpoison_pte_range,
 679	.hugetlb_entry = hwpoison_hugetlb_range,
 680};
 681
 682/*
 683 * Sends SIGBUS to the current process with error info.
 684 *
 685 * This function is intended to handle "Action Required" MCEs on already
 686 * hardware poisoned pages. They could happen, for example, when
 687 * memory_failure() failed to unmap the error page at the first call, or
 688 * when multiple local machine checks happened on different CPUs.
 689 *
 690 * MCE handler currently has no easy access to the error virtual address,
 691 * so this function walks page table to find it. The returned virtual address
 692 * is proper in most cases, but it could be wrong when the application
 693 * process has multiple entries mapping the error page.
 694 */
 695static int kill_accessing_process(struct task_struct *p, unsigned long pfn,
 696				  int flags)
 697{
 698	int ret;
 699	struct hwp_walk priv = {
 700		.pfn = pfn,
 701	};
 702	priv.tk.tsk = p;
 703
 704	mmap_read_lock(p->mm);
 705	ret = walk_page_range(p->mm, 0, TASK_SIZE, &hwp_walk_ops,
 706			      (void *)&priv);
 707	if (ret == 1 && priv.tk.addr)
 708		kill_proc(&priv.tk, pfn, flags);
 709	mmap_read_unlock(p->mm);
 710	return ret ? -EFAULT : -EHWPOISON;
 711}
 712
 713static const char *action_name[] = {
 714	[MF_IGNORED] = "Ignored",
 715	[MF_FAILED] = "Failed",
 716	[MF_DELAYED] = "Delayed",
 717	[MF_RECOVERED] = "Recovered",
 718};
 719
 720static const char * const action_page_types[] = {
 721	[MF_MSG_KERNEL]			= "reserved kernel page",
 722	[MF_MSG_KERNEL_HIGH_ORDER]	= "high-order kernel page",
 723	[MF_MSG_SLAB]			= "kernel slab page",
 724	[MF_MSG_DIFFERENT_COMPOUND]	= "different compound page after locking",
 725	[MF_MSG_POISONED_HUGE]		= "huge page already hardware poisoned",
 726	[MF_MSG_HUGE]			= "huge page",
 727	[MF_MSG_FREE_HUGE]		= "free huge page",
 728	[MF_MSG_NON_PMD_HUGE]		= "non-pmd-sized huge page",
 729	[MF_MSG_UNMAP_FAILED]		= "unmapping failed page",
 730	[MF_MSG_DIRTY_SWAPCACHE]	= "dirty swapcache page",
 731	[MF_MSG_CLEAN_SWAPCACHE]	= "clean swapcache page",
 732	[MF_MSG_DIRTY_MLOCKED_LRU]	= "dirty mlocked LRU page",
 733	[MF_MSG_CLEAN_MLOCKED_LRU]	= "clean mlocked LRU page",
 734	[MF_MSG_DIRTY_UNEVICTABLE_LRU]	= "dirty unevictable LRU page",
 735	[MF_MSG_CLEAN_UNEVICTABLE_LRU]	= "clean unevictable LRU page",
 736	[MF_MSG_DIRTY_LRU]		= "dirty LRU page",
 737	[MF_MSG_CLEAN_LRU]		= "clean LRU page",
 738	[MF_MSG_TRUNCATED_LRU]		= "already truncated LRU page",
 739	[MF_MSG_BUDDY]			= "free buddy page",
 740	[MF_MSG_BUDDY_2ND]		= "free buddy page (2nd try)",
 741	[MF_MSG_DAX]			= "dax page",
 742	[MF_MSG_UNSPLIT_THP]		= "unsplit thp",
 743	[MF_MSG_UNKNOWN]		= "unknown page",
 744};
 745
 746/*
 747 * XXX: It is possible that a page is isolated from LRU cache,
 748 * and then kept in swap cache or failed to remove from page cache.
 749 * The page count will stop it from being freed by unpoison.
 750 * Stress tests should be aware of this memory leak problem.
 751 */
 752static int delete_from_lru_cache(struct page *p)
 753{
 754	if (!isolate_lru_page(p)) {
 755		/*
 756		 * Clear sensible page flags, so that the buddy system won't
 757		 * complain when the page is unpoison-and-freed.
 758		 */
 759		ClearPageActive(p);
 760		ClearPageUnevictable(p);
 761
 762		/*
 763		 * Poisoned page might never drop its ref count to 0 so we have
 764		 * to uncharge it manually from its memcg.
 765		 */
 766		mem_cgroup_uncharge(p);
 767
 768		/*
 769		 * drop the page count elevated by isolate_lru_page()
 770		 */
 771		put_page(p);
 772		return 0;
 773	}
 774	return -EIO;
 775}
 776
 777static int truncate_error_page(struct page *p, unsigned long pfn,
 778				struct address_space *mapping)
 779{
 780	int ret = MF_FAILED;
 781
 782	if (mapping->a_ops->error_remove_page) {
 783		int err = mapping->a_ops->error_remove_page(mapping, p);
 784
 785		if (err != 0) {
 786			pr_info("Memory failure: %#lx: Failed to punch page: %d\n",
 787				pfn, err);
 788		} else if (page_has_private(p) &&
 789			   !try_to_release_page(p, GFP_NOIO)) {
 790			pr_info("Memory failure: %#lx: failed to release buffers\n",
 791				pfn);
 792		} else {
 793			ret = MF_RECOVERED;
 794		}
 795	} else {
 796		/*
 797		 * If the file system doesn't support it just invalidate
 798		 * This fails on dirty or anything with private pages
 799		 */
 800		if (invalidate_inode_page(p))
 801			ret = MF_RECOVERED;
 802		else
 803			pr_info("Memory failure: %#lx: Failed to invalidate\n",
 804				pfn);
 805	}
 806
 807	return ret;
 808}
 809
 810/*
 811 * Error hit kernel page.
 812 * Do nothing, try to be lucky and not touch this instead. For a few cases we
 813 * could be more sophisticated.
 814 */
 815static int me_kernel(struct page *p, unsigned long pfn)
 816{
 817	unlock_page(p);
 818	return MF_IGNORED;
 819}
 820
 821/*
 822 * Page in unknown state. Do nothing.
 823 */
 824static int me_unknown(struct page *p, unsigned long pfn)
 825{
 826	pr_err("Memory failure: %#lx: Unknown page state\n", pfn);
 827	unlock_page(p);
 828	return MF_FAILED;
 829}
 830
 831/*
 832 * Clean (or cleaned) page cache page.
 833 */
 834static int me_pagecache_clean(struct page *p, unsigned long pfn)
 835{
 836	int ret;
 837	struct address_space *mapping;
 838
 839	delete_from_lru_cache(p);
 840
 841	/*
 842	 * For anonymous pages we're done the only reference left
 843	 * should be the one m_f() holds.
 844	 */
 845	if (PageAnon(p)) {
 846		ret = MF_RECOVERED;
 847		goto out;
 848	}
 849
 850	/*
 851	 * Now truncate the page in the page cache. This is really
 852	 * more like a "temporary hole punch"
 853	 * Don't do this for block devices when someone else
 854	 * has a reference, because it could be file system metadata
 855	 * and that's not safe to truncate.
 856	 */
 857	mapping = page_mapping(p);
 858	if (!mapping) {
 859		/*
 860		 * Page has been teared down in the meanwhile
 861		 */
 862		ret = MF_FAILED;
 863		goto out;
 864	}
 865
 866	/*
 867	 * Truncation is a bit tricky. Enable it per file system for now.
 868	 *
 869	 * Open: to take i_mutex or not for this? Right now we don't.
 870	 */
 871	ret = truncate_error_page(p, pfn, mapping);
 872out:
 873	unlock_page(p);
 874	return ret;
 875}
 876
 877/*
 878 * Dirty pagecache page
 879 * Issues: when the error hit a hole page the error is not properly
 880 * propagated.
 881 */
 882static int me_pagecache_dirty(struct page *p, unsigned long pfn)
 883{
 884	struct address_space *mapping = page_mapping(p);
 885
 886	SetPageError(p);
 887	/* TBD: print more information about the file. */
 888	if (mapping) {
 889		/*
 890		 * IO error will be reported by write(), fsync(), etc.
 891		 * who check the mapping.
 892		 * This way the application knows that something went
 893		 * wrong with its dirty file data.
 894		 *
 895		 * There's one open issue:
 896		 *
 897		 * The EIO will be only reported on the next IO
 898		 * operation and then cleared through the IO map.
 899		 * Normally Linux has two mechanisms to pass IO error
 900		 * first through the AS_EIO flag in the address space
 901		 * and then through the PageError flag in the page.
 902		 * Since we drop pages on memory failure handling the
 903		 * only mechanism open to use is through AS_AIO.
 904		 *
 905		 * This has the disadvantage that it gets cleared on
 906		 * the first operation that returns an error, while
 907		 * the PageError bit is more sticky and only cleared
 908		 * when the page is reread or dropped.  If an
 909		 * application assumes it will always get error on
 910		 * fsync, but does other operations on the fd before
 911		 * and the page is dropped between then the error
 912		 * will not be properly reported.
 913		 *
 914		 * This can already happen even without hwpoisoned
 915		 * pages: first on metadata IO errors (which only
 916		 * report through AS_EIO) or when the page is dropped
 917		 * at the wrong time.
 918		 *
 919		 * So right now we assume that the application DTRT on
 920		 * the first EIO, but we're not worse than other parts
 921		 * of the kernel.
 922		 */
 923		mapping_set_error(mapping, -EIO);
 924	}
 925
 926	return me_pagecache_clean(p, pfn);
 927}
 928
 929/*
 930 * Clean and dirty swap cache.
 931 *
 932 * Dirty swap cache page is tricky to handle. The page could live both in page
 933 * cache and swap cache(ie. page is freshly swapped in). So it could be
 934 * referenced concurrently by 2 types of PTEs:
 935 * normal PTEs and swap PTEs. We try to handle them consistently by calling
 936 * try_to_unmap(TTU_IGNORE_HWPOISON) to convert the normal PTEs to swap PTEs,
 937 * and then
 938 *      - clear dirty bit to prevent IO
 939 *      - remove from LRU
 940 *      - but keep in the swap cache, so that when we return to it on
 941 *        a later page fault, we know the application is accessing
 942 *        corrupted data and shall be killed (we installed simple
 943 *        interception code in do_swap_page to catch it).
 944 *
 945 * Clean swap cache pages can be directly isolated. A later page fault will
 946 * bring in the known good data from disk.
 947 */
 948static int me_swapcache_dirty(struct page *p, unsigned long pfn)
 949{
 950	int ret;
 951
 952	ClearPageDirty(p);
 953	/* Trigger EIO in shmem: */
 954	ClearPageUptodate(p);
 955
 956	ret = delete_from_lru_cache(p) ? MF_FAILED : MF_DELAYED;
 957	unlock_page(p);
 958	return ret;
 
 959}
 960
 961static int me_swapcache_clean(struct page *p, unsigned long pfn)
 962{
 963	int ret;
 964
 965	delete_from_swap_cache(p);
 966
 967	ret = delete_from_lru_cache(p) ? MF_FAILED : MF_RECOVERED;
 968	unlock_page(p);
 969	return ret;
 
 970}
 971
 972/*
 973 * Huge pages. Needs work.
 974 * Issues:
 975 * - Error on hugepage is contained in hugepage unit (not in raw page unit.)
 976 *   To narrow down kill region to one page, we need to break up pmd.
 977 */
 978static int me_huge_page(struct page *p, unsigned long pfn)
 979{
 980	int res;
 981	struct page *hpage = compound_head(p);
 982	struct address_space *mapping;
 983
 984	if (!PageHuge(hpage))
 985		return MF_DELAYED;
 986
 987	mapping = page_mapping(hpage);
 988	if (mapping) {
 989		res = truncate_error_page(hpage, pfn, mapping);
 990		unlock_page(hpage);
 991	} else {
 992		res = MF_FAILED;
 993		unlock_page(hpage);
 994		/*
 995		 * migration entry prevents later access on error anonymous
 996		 * hugepage, so we can free and dissolve it into buddy to
 997		 * save healthy subpages.
 998		 */
 999		if (PageAnon(hpage))
1000			put_page(hpage);
1001		if (__page_handle_poison(p)) {
1002			page_ref_inc(p);
1003			res = MF_RECOVERED;
1004		}
1005	}
1006
1007	return res;
1008}
1009
1010/*
1011 * Various page states we can handle.
1012 *
1013 * A page state is defined by its current page->flags bits.
1014 * The table matches them in order and calls the right handler.
1015 *
1016 * This is quite tricky because we can access page at any time
1017 * in its live cycle, so all accesses have to be extremely careful.
1018 *
1019 * This is not complete. More states could be added.
1020 * For any missing state don't attempt recovery.
1021 */
1022
1023#define dirty		(1UL << PG_dirty)
1024#define sc		((1UL << PG_swapcache) | (1UL << PG_swapbacked))
1025#define unevict		(1UL << PG_unevictable)
1026#define mlock		(1UL << PG_mlocked)
 
1027#define lru		(1UL << PG_lru)
1028#define head		(1UL << PG_head)
1029#define slab		(1UL << PG_slab)
1030#define reserved	(1UL << PG_reserved)
1031
1032static struct page_state {
1033	unsigned long mask;
1034	unsigned long res;
1035	enum mf_action_page_type type;
1036
1037	/* Callback ->action() has to unlock the relevant page inside it. */
1038	int (*action)(struct page *p, unsigned long pfn);
1039} error_states[] = {
1040	{ reserved,	reserved,	MF_MSG_KERNEL,	me_kernel },
1041	/*
1042	 * free pages are specially detected outside this table:
1043	 * PG_buddy pages only make a small fraction of all free pages.
1044	 */
1045
1046	/*
1047	 * Could in theory check if slab page is free or if we can drop
1048	 * currently unused objects without touching them. But just
1049	 * treat it as standard kernel for now.
1050	 */
1051	{ slab,		slab,		MF_MSG_SLAB,	me_kernel },
1052
1053	{ head,		head,		MF_MSG_HUGE,		me_huge_page },
1054
1055	{ sc|dirty,	sc|dirty,	MF_MSG_DIRTY_SWAPCACHE,	me_swapcache_dirty },
1056	{ sc|dirty,	sc,		MF_MSG_CLEAN_SWAPCACHE,	me_swapcache_clean },
1057
1058	{ mlock|dirty,	mlock|dirty,	MF_MSG_DIRTY_MLOCKED_LRU,	me_pagecache_dirty },
1059	{ mlock|dirty,	mlock,		MF_MSG_CLEAN_MLOCKED_LRU,	me_pagecache_clean },
1060
1061	{ unevict|dirty, unevict|dirty,	MF_MSG_DIRTY_UNEVICTABLE_LRU,	me_pagecache_dirty },
1062	{ unevict|dirty, unevict,	MF_MSG_CLEAN_UNEVICTABLE_LRU,	me_pagecache_clean },
1063
1064	{ lru|dirty,	lru|dirty,	MF_MSG_DIRTY_LRU,	me_pagecache_dirty },
1065	{ lru|dirty,	lru,		MF_MSG_CLEAN_LRU,	me_pagecache_clean },
1066
1067	/*
1068	 * Catchall entry: must be at end.
1069	 */
1070	{ 0,		0,		MF_MSG_UNKNOWN,	me_unknown },
1071};
1072
1073#undef dirty
1074#undef sc
1075#undef unevict
1076#undef mlock
 
1077#undef lru
1078#undef head
1079#undef slab
1080#undef reserved
1081
1082/*
1083 * "Dirty/Clean" indication is not 100% accurate due to the possibility of
1084 * setting PG_dirty outside page lock. See also comment above set_page_dirty().
1085 */
1086static void action_result(unsigned long pfn, enum mf_action_page_type type,
1087			  enum mf_result result)
1088{
1089	trace_memory_failure_event(pfn, type, result);
1090
1091	pr_err("Memory failure: %#lx: recovery action for %s: %s\n",
1092		pfn, action_page_types[type], action_name[result]);
1093}
1094
1095static int page_action(struct page_state *ps, struct page *p,
1096			unsigned long pfn)
1097{
1098	int result;
1099	int count;
1100
1101	/* page p should be unlocked after returning from ps->action().  */
1102	result = ps->action(p, pfn);
1103
1104	count = page_count(p) - 1;
1105	if (ps->action == me_swapcache_dirty && result == MF_DELAYED)
1106		count--;
1107	if (count > 0) {
1108		pr_err("Memory failure: %#lx: %s still referenced by %d users\n",
1109		       pfn, action_page_types[ps->type], count);
1110		result = MF_FAILED;
1111	}
1112	action_result(pfn, ps->type, result);
1113
1114	/* Could do more checks here if page looks ok */
1115	/*
1116	 * Could adjust zone counters here to correct for the missing page.
1117	 */
1118
1119	return (result == MF_RECOVERED || result == MF_DELAYED) ? 0 : -EBUSY;
1120}
1121
1122/*
1123 * Return true if a page type of a given page is supported by hwpoison
1124 * mechanism (while handling could fail), otherwise false.  This function
1125 * does not return true for hugetlb or device memory pages, so it's assumed
1126 * to be called only in the context where we never have such pages.
 
1127 */
1128static inline bool HWPoisonHandlable(struct page *page)
1129{
1130	return PageLRU(page) || __PageMovable(page) || is_free_buddy_page(page);
1131}
1132
1133static int __get_hwpoison_page(struct page *page)
1134{
1135	struct page *head = compound_head(page);
1136	int ret = 0;
1137	bool hugetlb = false;
1138
1139	ret = get_hwpoison_huge_page(head, &hugetlb);
1140	if (hugetlb)
1141		return ret;
1142
1143	/*
1144	 * This check prevents from calling get_hwpoison_unless_zero()
1145	 * for any unsupported type of page in order to reduce the risk of
1146	 * unexpected races caused by taking a page refcount.
1147	 */
1148	if (!HWPoisonHandlable(head))
1149		return -EBUSY;
1150
1151	if (PageTransHuge(head)) {
1152		/*
1153		 * Non anonymous thp exists only in allocation/free time. We
1154		 * can't handle such a case correctly, so let's give it up.
1155		 * This should be better than triggering BUG_ON when kernel
1156		 * tries to touch the "partially handled" page.
1157		 */
1158		if (!PageAnon(head)) {
1159			pr_err("Memory failure: %#lx: non anonymous thp\n",
1160				page_to_pfn(page));
1161			return 0;
1162		}
1163	}
1164
1165	if (get_page_unless_zero(head)) {
1166		if (head == compound_head(page))
1167			return 1;
1168
1169		pr_info("Memory failure: %#lx cannot catch tail\n",
1170			page_to_pfn(page));
1171		put_page(head);
1172	}
1173
1174	return 0;
1175}
1176
1177static int get_any_page(struct page *p, unsigned long flags)
1178{
1179	int ret = 0, pass = 0;
1180	bool count_increased = false;
1181
1182	if (flags & MF_COUNT_INCREASED)
1183		count_increased = true;
1184
1185try_again:
1186	if (!count_increased) {
1187		ret = __get_hwpoison_page(p);
1188		if (!ret) {
1189			if (page_count(p)) {
1190				/* We raced with an allocation, retry. */
1191				if (pass++ < 3)
1192					goto try_again;
1193				ret = -EBUSY;
1194			} else if (!PageHuge(p) && !is_free_buddy_page(p)) {
1195				/* We raced with put_page, retry. */
1196				if (pass++ < 3)
1197					goto try_again;
1198				ret = -EIO;
1199			}
1200			goto out;
1201		} else if (ret == -EBUSY) {
1202			/*
1203			 * We raced with (possibly temporary) unhandlable
1204			 * page, retry.
1205			 */
1206			if (pass++ < 3) {
1207				shake_page(p, 1);
1208				goto try_again;
1209			}
1210			ret = -EIO;
1211			goto out;
1212		}
1213	}
1214
1215	if (PageHuge(p) || HWPoisonHandlable(p)) {
1216		ret = 1;
1217	} else {
1218		/*
1219		 * A page we cannot handle. Check whether we can turn
1220		 * it into something we can handle.
1221		 */
1222		if (pass++ < 3) {
1223			put_page(p);
1224			shake_page(p, 1);
1225			count_increased = false;
1226			goto try_again;
1227		}
1228		put_page(p);
1229		ret = -EIO;
1230	}
1231out:
1232	return ret;
1233}
1234
1235/**
1236 * get_hwpoison_page() - Get refcount for memory error handling
1237 * @p:		Raw error page (hit by memory error)
1238 * @flags:	Flags controlling behavior of error handling
1239 *
1240 * get_hwpoison_page() takes a page refcount of an error page to handle memory
1241 * error on it, after checking that the error page is in a well-defined state
1242 * (defined as a page-type we can successfully handle the memor error on it,
1243 * such as LRU page and hugetlb page).
1244 *
1245 * Memory error handling could be triggered at any time on any type of page,
1246 * so it's prone to race with typical memory management lifecycle (like
1247 * allocation and free).  So to avoid such races, get_hwpoison_page() takes
1248 * extra care for the error page's state (as done in __get_hwpoison_page()),
1249 * and has some retry logic in get_any_page().
1250 *
1251 * Return: 0 on failure,
1252 *         1 on success for in-use pages in a well-defined state,
1253 *         -EIO for pages on which we can not handle memory errors,
1254 *         -EBUSY when get_hwpoison_page() has raced with page lifecycle
1255 *         operations like allocation and free.
1256 */
1257static int get_hwpoison_page(struct page *p, unsigned long flags)
1258{
1259	int ret;
1260
1261	zone_pcp_disable(page_zone(p));
1262	ret = get_any_page(p, flags);
1263	zone_pcp_enable(page_zone(p));
1264
1265	return ret;
1266}
1267
1268/*
1269 * Do all that is necessary to remove user space mappings. Unmap
1270 * the pages and send SIGBUS to the processes if the data was dirty.
1271 */
1272static bool hwpoison_user_mappings(struct page *p, unsigned long pfn,
1273				  int flags, struct page **hpagep)
1274{
1275	enum ttu_flags ttu = TTU_IGNORE_MLOCK | TTU_SYNC;
1276	struct address_space *mapping;
1277	LIST_HEAD(tokill);
1278	bool unmap_success;
1279	int kill = 1, forcekill;
1280	struct page *hpage = *hpagep;
1281	bool mlocked = PageMlocked(hpage);
1282
1283	/*
1284	 * Here we are interested only in user-mapped pages, so skip any
1285	 * other types of pages.
1286	 */
1287	if (PageReserved(p) || PageSlab(p))
1288		return true;
1289	if (!(PageLRU(hpage) || PageHuge(p)))
1290		return true;
1291
1292	/*
1293	 * This check implies we don't kill processes if their pages
1294	 * are in the swap cache early. Those are always late kills.
1295	 */
1296	if (!page_mapped(hpage))
1297		return true;
1298
1299	if (PageKsm(p)) {
1300		pr_err("Memory failure: %#lx: can't handle KSM pages.\n", pfn);
1301		return false;
1302	}
1303
1304	if (PageSwapCache(p)) {
1305		pr_err("Memory failure: %#lx: keeping poisoned page in swap cache\n",
1306			pfn);
1307		ttu |= TTU_IGNORE_HWPOISON;
1308	}
1309
1310	/*
1311	 * Propagate the dirty bit from PTEs to struct page first, because we
1312	 * need this to decide if we should kill or just drop the page.
1313	 * XXX: the dirty test could be racy: set_page_dirty() may not always
1314	 * be called inside page lock (it's recommended but not enforced).
1315	 */
1316	mapping = page_mapping(hpage);
1317	if (!(flags & MF_MUST_KILL) && !PageDirty(hpage) && mapping &&
1318	    mapping_can_writeback(mapping)) {
1319		if (page_mkclean(hpage)) {
1320			SetPageDirty(hpage);
1321		} else {
1322			kill = 0;
1323			ttu |= TTU_IGNORE_HWPOISON;
1324			pr_info("Memory failure: %#lx: corrupted page was clean: dropped without side effects\n",
1325				pfn);
1326		}
1327	}
1328
1329	/*
1330	 * First collect all the processes that have the page
1331	 * mapped in dirty form.  This has to be done before try_to_unmap,
1332	 * because ttu takes the rmap data structures down.
1333	 *
1334	 * Error handling: We ignore errors here because
1335	 * there's nothing that can be done.
1336	 */
1337	if (kill)
1338		collect_procs(hpage, &tokill, flags & MF_ACTION_REQUIRED);
1339
1340	if (!PageHuge(hpage)) {
1341		try_to_unmap(hpage, ttu);
1342	} else {
1343		if (!PageAnon(hpage)) {
1344			/*
1345			 * For hugetlb pages in shared mappings, try_to_unmap
1346			 * could potentially call huge_pmd_unshare.  Because of
1347			 * this, take semaphore in write mode here and set
1348			 * TTU_RMAP_LOCKED to indicate we have taken the lock
1349			 * at this higher level.
1350			 */
1351			mapping = hugetlb_page_mapping_lock_write(hpage);
1352			if (mapping) {
1353				try_to_unmap(hpage, ttu|TTU_RMAP_LOCKED);
1354				i_mmap_unlock_write(mapping);
1355			} else
1356				pr_info("Memory failure: %#lx: could not lock mapping for mapped huge page\n", pfn);
1357		} else {
1358			try_to_unmap(hpage, ttu);
1359		}
1360	}
1361
1362	unmap_success = !page_mapped(hpage);
1363	if (!unmap_success)
1364		pr_err("Memory failure: %#lx: failed to unmap page (mapcount=%d)\n",
1365		       pfn, page_mapcount(hpage));
1366
1367	/*
1368	 * try_to_unmap() might put mlocked page in lru cache, so call
1369	 * shake_page() again to ensure that it's flushed.
1370	 */
1371	if (mlocked)
1372		shake_page(hpage, 0);
1373
1374	/*
1375	 * Now that the dirty bit has been propagated to the
1376	 * struct page and all unmaps done we can decide if
1377	 * killing is needed or not.  Only kill when the page
1378	 * was dirty or the process is not restartable,
1379	 * otherwise the tokill list is merely
1380	 * freed.  When there was a problem unmapping earlier
1381	 * use a more force-full uncatchable kill to prevent
1382	 * any accesses to the poisoned memory.
1383	 */
1384	forcekill = PageDirty(hpage) || (flags & MF_MUST_KILL);
1385	kill_procs(&tokill, forcekill, !unmap_success, pfn, flags);
1386
1387	return unmap_success;
1388}
1389
1390static int identify_page_state(unsigned long pfn, struct page *p,
1391				unsigned long page_flags)
1392{
1393	struct page_state *ps;
1394
1395	/*
1396	 * The first check uses the current page flags which may not have any
1397	 * relevant information. The second check with the saved page flags is
1398	 * carried out only if the first check can't determine the page status.
1399	 */
1400	for (ps = error_states;; ps++)
1401		if ((p->flags & ps->mask) == ps->res)
1402			break;
1403
1404	page_flags |= (p->flags & (1UL << PG_dirty));
1405
1406	if (!ps->mask)
1407		for (ps = error_states;; ps++)
1408			if ((page_flags & ps->mask) == ps->res)
1409				break;
1410	return page_action(ps, p, pfn);
1411}
1412
1413static int try_to_split_thp_page(struct page *page, const char *msg)
1414{
1415	lock_page(page);
1416	if (!PageAnon(page) || unlikely(split_huge_page(page))) {
1417		unsigned long pfn = page_to_pfn(page);
1418
1419		unlock_page(page);
1420		if (!PageAnon(page))
1421			pr_info("%s: %#lx: non anonymous thp\n", msg, pfn);
1422		else
1423			pr_info("%s: %#lx: thp split failed\n", msg, pfn);
1424		put_page(page);
1425		return -EBUSY;
1426	}
1427	unlock_page(page);
1428
1429	return 0;
1430}
1431
1432static int memory_failure_hugetlb(unsigned long pfn, int flags)
1433{
1434	struct page *p = pfn_to_page(pfn);
1435	struct page *head = compound_head(p);
1436	int res;
1437	unsigned long page_flags;
1438
1439	if (TestSetPageHWPoison(head)) {
1440		pr_err("Memory failure: %#lx: already hardware poisoned\n",
1441		       pfn);
1442		res = -EHWPOISON;
1443		if (flags & MF_ACTION_REQUIRED)
1444			res = kill_accessing_process(current, page_to_pfn(head), flags);
1445		return res;
1446	}
1447
1448	num_poisoned_pages_inc();
1449
1450	if (!(flags & MF_COUNT_INCREASED)) {
1451		res = get_hwpoison_page(p, flags);
1452		if (!res) {
1453			/*
1454			 * Check "filter hit" and "race with other subpage."
1455			 */
1456			lock_page(head);
1457			if (PageHWPoison(head)) {
1458				if ((hwpoison_filter(p) && TestClearPageHWPoison(p))
1459				    || (p != head && TestSetPageHWPoison(head))) {
1460					num_poisoned_pages_dec();
1461					unlock_page(head);
1462					return 0;
1463				}
1464			}
1465			unlock_page(head);
1466			res = MF_FAILED;
1467			if (__page_handle_poison(p)) {
1468				page_ref_inc(p);
1469				res = MF_RECOVERED;
1470			}
1471			action_result(pfn, MF_MSG_FREE_HUGE, res);
1472			return res == MF_RECOVERED ? 0 : -EBUSY;
1473		} else if (res < 0) {
1474			action_result(pfn, MF_MSG_UNKNOWN, MF_IGNORED);
1475			return -EBUSY;
1476		}
 
 
 
 
1477	}
1478
1479	lock_page(head);
1480	page_flags = head->flags;
1481
1482	if (!PageHWPoison(head)) {
1483		pr_err("Memory failure: %#lx: just unpoisoned\n", pfn);
1484		num_poisoned_pages_dec();
1485		unlock_page(head);
1486		put_page(head);
1487		return 0;
1488	}
1489
1490	/*
1491	 * TODO: hwpoison for pud-sized hugetlb doesn't work right now, so
1492	 * simply disable it. In order to make it work properly, we need
1493	 * make sure that:
1494	 *  - conversion of a pud that maps an error hugetlb into hwpoison
1495	 *    entry properly works, and
1496	 *  - other mm code walking over page table is aware of pud-aligned
1497	 *    hwpoison entries.
1498	 */
1499	if (huge_page_size(page_hstate(head)) > PMD_SIZE) {
1500		action_result(pfn, MF_MSG_NON_PMD_HUGE, MF_IGNORED);
1501		res = -EBUSY;
1502		goto out;
1503	}
1504
1505	if (!hwpoison_user_mappings(p, pfn, flags, &head)) {
1506		action_result(pfn, MF_MSG_UNMAP_FAILED, MF_IGNORED);
1507		res = -EBUSY;
1508		goto out;
1509	}
1510
1511	return identify_page_state(pfn, p, page_flags);
1512out:
1513	unlock_page(head);
1514	return res;
1515}
1516
1517static int memory_failure_dev_pagemap(unsigned long pfn, int flags,
1518		struct dev_pagemap *pgmap)
1519{
1520	struct page *page = pfn_to_page(pfn);
1521	const bool unmap_success = true;
1522	unsigned long size = 0;
1523	struct to_kill *tk;
1524	LIST_HEAD(tokill);
1525	int rc = -EBUSY;
1526	loff_t start;
1527	dax_entry_t cookie;
1528
1529	if (flags & MF_COUNT_INCREASED)
1530		/*
1531		 * Drop the extra refcount in case we come from madvise().
1532		 */
1533		put_page(page);
1534
1535	/* device metadata space is not recoverable */
1536	if (!pgmap_pfn_valid(pgmap, pfn)) {
1537		rc = -ENXIO;
1538		goto out;
1539	}
1540
1541	/*
1542	 * Prevent the inode from being freed while we are interrogating
1543	 * the address_space, typically this would be handled by
1544	 * lock_page(), but dax pages do not use the page lock. This
1545	 * also prevents changes to the mapping of this pfn until
1546	 * poison signaling is complete.
1547	 */
1548	cookie = dax_lock_page(page);
1549	if (!cookie)
1550		goto out;
1551
1552	if (hwpoison_filter(page)) {
1553		rc = 0;
1554		goto unlock;
1555	}
1556
1557	if (pgmap->type == MEMORY_DEVICE_PRIVATE) {
1558		/*
1559		 * TODO: Handle HMM pages which may need coordination
1560		 * with device-side memory.
1561		 */
1562		goto unlock;
1563	}
1564
1565	/*
1566	 * Use this flag as an indication that the dax page has been
1567	 * remapped UC to prevent speculative consumption of poison.
1568	 */
1569	SetPageHWPoison(page);
1570
1571	/*
1572	 * Unlike System-RAM there is no possibility to swap in a
1573	 * different physical page at a given virtual address, so all
1574	 * userspace consumption of ZONE_DEVICE memory necessitates
1575	 * SIGBUS (i.e. MF_MUST_KILL)
1576	 */
1577	flags |= MF_ACTION_REQUIRED | MF_MUST_KILL;
1578	collect_procs(page, &tokill, flags & MF_ACTION_REQUIRED);
1579
1580	list_for_each_entry(tk, &tokill, nd)
1581		if (tk->size_shift)
1582			size = max(size, 1UL << tk->size_shift);
1583	if (size) {
1584		/*
1585		 * Unmap the largest mapping to avoid breaking up
1586		 * device-dax mappings which are constant size. The
1587		 * actual size of the mapping being torn down is
1588		 * communicated in siginfo, see kill_proc()
1589		 */
1590		start = (page->index << PAGE_SHIFT) & ~(size - 1);
1591		unmap_mapping_range(page->mapping, start, size, 0);
1592	}
1593	kill_procs(&tokill, flags & MF_MUST_KILL, !unmap_success, pfn, flags);
1594	rc = 0;
1595unlock:
1596	dax_unlock_page(page, cookie);
1597out:
1598	/* drop pgmap ref acquired in caller */
1599	put_dev_pagemap(pgmap);
1600	action_result(pfn, MF_MSG_DAX, rc ? MF_FAILED : MF_RECOVERED);
1601	return rc;
1602}
1603
1604/**
1605 * memory_failure - Handle memory failure of a page.
1606 * @pfn: Page Number of the corrupted page
1607 * @flags: fine tune action taken
1608 *
1609 * This function is called by the low level machine check code
1610 * of an architecture when it detects hardware memory corruption
1611 * of a page. It tries its best to recover, which includes
1612 * dropping pages, killing processes etc.
1613 *
1614 * The function is primarily of use for corruptions that
1615 * happen outside the current execution context (e.g. when
1616 * detected by a background scrubber)
1617 *
1618 * Must run in process context (e.g. a work queue) with interrupts
1619 * enabled and no spinlocks hold.
1620 */
1621int memory_failure(unsigned long pfn, int flags)
1622{
1623	struct page *p;
1624	struct page *hpage;
1625	struct page *orig_head;
1626	struct dev_pagemap *pgmap;
1627	int res = 0;
1628	unsigned long page_flags;
1629	bool retry = true;
1630	static DEFINE_MUTEX(mf_mutex);
1631
1632	if (!sysctl_memory_failure_recovery)
1633		panic("Memory failure on page %lx", pfn);
1634
1635	p = pfn_to_online_page(pfn);
1636	if (!p) {
1637		if (pfn_valid(pfn)) {
1638			pgmap = get_dev_pagemap(pfn, NULL);
1639			if (pgmap)
1640				return memory_failure_dev_pagemap(pfn, flags,
1641								  pgmap);
1642		}
1643		pr_err("Memory failure: %#lx: memory outside kernel control\n",
1644			pfn);
1645		return -ENXIO;
1646	}
1647
1648	mutex_lock(&mf_mutex);
1649
1650try_again:
1651	if (PageHuge(p)) {
1652		res = memory_failure_hugetlb(pfn, flags);
1653		goto unlock_mutex;
1654	}
1655
1656	if (TestSetPageHWPoison(p)) {
1657		pr_err("Memory failure: %#lx: already hardware poisoned\n",
1658			pfn);
1659		res = -EHWPOISON;
1660		if (flags & MF_ACTION_REQUIRED)
1661			res = kill_accessing_process(current, pfn, flags);
1662		goto unlock_mutex;
1663	}
1664
1665	orig_head = hpage = compound_head(p);
1666	num_poisoned_pages_inc();
1667
1668	/*
1669	 * We need/can do nothing about count=0 pages.
1670	 * 1) it's a free page, and therefore in safe hand:
1671	 *    prep_new_page() will be the gate keeper.
1672	 * 2) it's part of a non-compound high order page.
1673	 *    Implies some kernel user: cannot stop them from
1674	 *    R/W the page; let's pray that the page has been
1675	 *    used and will be freed some time later.
1676	 * In fact it's dangerous to directly bump up page count from 0,
1677	 * that may make page_ref_freeze()/page_ref_unfreeze() mismatch.
1678	 */
1679	if (!(flags & MF_COUNT_INCREASED)) {
1680		res = get_hwpoison_page(p, flags);
1681		if (!res) {
1682			if (is_free_buddy_page(p)) {
1683				if (take_page_off_buddy(p)) {
1684					page_ref_inc(p);
1685					res = MF_RECOVERED;
1686				} else {
1687					/* We lost the race, try again */
1688					if (retry) {
1689						ClearPageHWPoison(p);
1690						num_poisoned_pages_dec();
1691						retry = false;
1692						goto try_again;
1693					}
1694					res = MF_FAILED;
1695				}
1696				action_result(pfn, MF_MSG_BUDDY, res);
1697				res = res == MF_RECOVERED ? 0 : -EBUSY;
1698			} else {
1699				action_result(pfn, MF_MSG_KERNEL_HIGH_ORDER, MF_IGNORED);
1700				res = -EBUSY;
1701			}
1702			goto unlock_mutex;
1703		} else if (res < 0) {
1704			action_result(pfn, MF_MSG_UNKNOWN, MF_IGNORED);
1705			res = -EBUSY;
1706			goto unlock_mutex;
1707		}
1708	}
1709
1710	if (PageTransHuge(hpage)) {
1711		if (try_to_split_thp_page(p, "Memory Failure") < 0) {
1712			action_result(pfn, MF_MSG_UNSPLIT_THP, MF_IGNORED);
1713			res = -EBUSY;
1714			goto unlock_mutex;
 
 
 
 
 
 
 
 
 
1715		}
 
1716		VM_BUG_ON_PAGE(!page_count(p), p);
 
1717	}
1718
1719	/*
1720	 * We ignore non-LRU pages for good reasons.
1721	 * - PG_locked is only well defined for LRU pages and a few others
1722	 * - to avoid races with __SetPageLocked()
1723	 * - to avoid races with __SetPageSlab*() (and more non-atomic ops)
1724	 * The check (unnecessarily) ignores LRU pages being isolated and
1725	 * walked by the page reclaim code, however that's not a big loss.
1726	 */
1727	shake_page(p, 0);
 
 
 
 
 
 
 
 
1728
1729	lock_page(p);
1730
1731	/*
1732	 * The page could have changed compound pages during the locking.
1733	 * If this happens just bail out.
1734	 */
1735	if (PageCompound(p) && compound_head(p) != orig_head) {
1736		action_result(pfn, MF_MSG_DIFFERENT_COMPOUND, MF_IGNORED);
1737		res = -EBUSY;
1738		goto unlock_page;
1739	}
1740
1741	/*
1742	 * We use page flags to determine what action should be taken, but
1743	 * the flags can be modified by the error containment action.  One
1744	 * example is an mlocked page, where PG_mlocked is cleared by
1745	 * page_remove_rmap() in try_to_unmap_one(). So to determine page status
1746	 * correctly, we save a copy of the page flags at this time.
1747	 */
1748	page_flags = p->flags;
 
 
 
1749
1750	/*
1751	 * unpoison always clear PG_hwpoison inside page lock
1752	 */
1753	if (!PageHWPoison(p)) {
1754		pr_err("Memory failure: %#lx: just unpoisoned\n", pfn);
1755		num_poisoned_pages_dec();
1756		unlock_page(p);
1757		put_page(p);
1758		goto unlock_mutex;
1759	}
1760	if (hwpoison_filter(p)) {
1761		if (TestClearPageHWPoison(p))
1762			num_poisoned_pages_dec();
1763		unlock_page(p);
1764		put_page(p);
1765		goto unlock_mutex;
1766	}
1767
1768	/*
1769	 * __munlock_pagevec may clear a writeback page's LRU flag without
1770	 * page_lock. We need wait writeback completion for this page or it
1771	 * may trigger vfs BUG while evict inode.
1772	 */
1773	if (!PageTransTail(p) && !PageLRU(p) && !PageWriteback(p))
1774		goto identify_page_state;
1775
1776	/*
1777	 * It's very difficult to mess with pages currently under IO
1778	 * and in many cases impossible, so we just avoid it here.
1779	 */
1780	wait_on_page_writeback(p);
1781
1782	/*
1783	 * Now take care of user space mappings.
1784	 * Abort on fail: __delete_from_page_cache() assumes unmapped page.
 
 
 
1785	 */
1786	if (!hwpoison_user_mappings(p, pfn, flags, &p)) {
1787		action_result(pfn, MF_MSG_UNMAP_FAILED, MF_IGNORED);
1788		res = -EBUSY;
1789		goto unlock_page;
1790	}
1791
1792	/*
1793	 * Torn down by someone else?
1794	 */
1795	if (PageLRU(p) && !PageSwapCache(p) && p->mapping == NULL) {
1796		action_result(pfn, MF_MSG_TRUNCATED_LRU, MF_IGNORED);
1797		res = -EBUSY;
1798		goto unlock_page;
1799	}
1800
1801identify_page_state:
1802	res = identify_page_state(pfn, p, page_flags);
1803	mutex_unlock(&mf_mutex);
1804	return res;
1805unlock_page:
1806	unlock_page(p);
1807unlock_mutex:
1808	mutex_unlock(&mf_mutex);
1809	return res;
1810}
1811EXPORT_SYMBOL_GPL(memory_failure);
1812
1813#define MEMORY_FAILURE_FIFO_ORDER	4
1814#define MEMORY_FAILURE_FIFO_SIZE	(1 << MEMORY_FAILURE_FIFO_ORDER)
1815
1816struct memory_failure_entry {
1817	unsigned long pfn;
1818	int flags;
1819};
1820
1821struct memory_failure_cpu {
1822	DECLARE_KFIFO(fifo, struct memory_failure_entry,
1823		      MEMORY_FAILURE_FIFO_SIZE);
1824	spinlock_t lock;
1825	struct work_struct work;
1826};
1827
1828static DEFINE_PER_CPU(struct memory_failure_cpu, memory_failure_cpu);
1829
1830/**
1831 * memory_failure_queue - Schedule handling memory failure of a page.
1832 * @pfn: Page Number of the corrupted page
1833 * @flags: Flags for memory failure handling
1834 *
1835 * This function is called by the low level hardware error handler
1836 * when it detects hardware memory corruption of a page. It schedules
1837 * the recovering of error page, including dropping pages, killing
1838 * processes etc.
1839 *
1840 * The function is primarily of use for corruptions that
1841 * happen outside the current execution context (e.g. when
1842 * detected by a background scrubber)
1843 *
1844 * Can run in IRQ context.
1845 */
1846void memory_failure_queue(unsigned long pfn, int flags)
1847{
1848	struct memory_failure_cpu *mf_cpu;
1849	unsigned long proc_flags;
1850	struct memory_failure_entry entry = {
1851		.pfn =		pfn,
1852		.flags =	flags,
1853	};
1854
1855	mf_cpu = &get_cpu_var(memory_failure_cpu);
1856	spin_lock_irqsave(&mf_cpu->lock, proc_flags);
1857	if (kfifo_put(&mf_cpu->fifo, entry))
1858		schedule_work_on(smp_processor_id(), &mf_cpu->work);
1859	else
1860		pr_err("Memory failure: buffer overflow when queuing memory failure at %#lx\n",
1861		       pfn);
1862	spin_unlock_irqrestore(&mf_cpu->lock, proc_flags);
1863	put_cpu_var(memory_failure_cpu);
1864}
1865EXPORT_SYMBOL_GPL(memory_failure_queue);
1866
1867static void memory_failure_work_func(struct work_struct *work)
1868{
1869	struct memory_failure_cpu *mf_cpu;
1870	struct memory_failure_entry entry = { 0, };
1871	unsigned long proc_flags;
1872	int gotten;
1873
1874	mf_cpu = container_of(work, struct memory_failure_cpu, work);
1875	for (;;) {
1876		spin_lock_irqsave(&mf_cpu->lock, proc_flags);
1877		gotten = kfifo_get(&mf_cpu->fifo, &entry);
1878		spin_unlock_irqrestore(&mf_cpu->lock, proc_flags);
1879		if (!gotten)
1880			break;
1881		if (entry.flags & MF_SOFT_OFFLINE)
1882			soft_offline_page(entry.pfn, entry.flags);
1883		else
1884			memory_failure(entry.pfn, entry.flags);
1885	}
1886}
1887
1888/*
1889 * Process memory_failure work queued on the specified CPU.
1890 * Used to avoid return-to-userspace racing with the memory_failure workqueue.
1891 */
1892void memory_failure_queue_kick(int cpu)
1893{
1894	struct memory_failure_cpu *mf_cpu;
1895
1896	mf_cpu = &per_cpu(memory_failure_cpu, cpu);
1897	cancel_work_sync(&mf_cpu->work);
1898	memory_failure_work_func(&mf_cpu->work);
1899}
1900
1901static int __init memory_failure_init(void)
1902{
1903	struct memory_failure_cpu *mf_cpu;
1904	int cpu;
1905
1906	for_each_possible_cpu(cpu) {
1907		mf_cpu = &per_cpu(memory_failure_cpu, cpu);
1908		spin_lock_init(&mf_cpu->lock);
1909		INIT_KFIFO(mf_cpu->fifo);
1910		INIT_WORK(&mf_cpu->work, memory_failure_work_func);
1911	}
1912
1913	return 0;
1914}
1915core_initcall(memory_failure_init);
1916
1917#define unpoison_pr_info(fmt, pfn, rs)			\
1918({							\
1919	if (__ratelimit(rs))				\
1920		pr_info(fmt, pfn);			\
1921})
1922
1923/**
1924 * unpoison_memory - Unpoison a previously poisoned page
1925 * @pfn: Page number of the to be unpoisoned page
1926 *
1927 * Software-unpoison a page that has been poisoned by
1928 * memory_failure() earlier.
1929 *
1930 * This is only done on the software-level, so it only works
1931 * for linux injected failures, not real hardware failures
1932 *
1933 * Returns 0 for success, otherwise -errno.
1934 */
1935int unpoison_memory(unsigned long pfn)
1936{
1937	struct page *page;
1938	struct page *p;
1939	int freeit = 0;
1940	unsigned long flags = 0;
1941	static DEFINE_RATELIMIT_STATE(unpoison_rs, DEFAULT_RATELIMIT_INTERVAL,
1942					DEFAULT_RATELIMIT_BURST);
1943
1944	if (!pfn_valid(pfn))
1945		return -ENXIO;
1946
1947	p = pfn_to_page(pfn);
1948	page = compound_head(p);
1949
1950	if (!PageHWPoison(p)) {
1951		unpoison_pr_info("Unpoison: Page was already unpoisoned %#lx\n",
1952				 pfn, &unpoison_rs);
1953		return 0;
1954	}
1955
1956	if (page_count(page) > 1) {
1957		unpoison_pr_info("Unpoison: Someone grabs the hwpoison page %#lx\n",
1958				 pfn, &unpoison_rs);
1959		return 0;
1960	}
1961
1962	if (page_mapped(page)) {
1963		unpoison_pr_info("Unpoison: Someone maps the hwpoison page %#lx\n",
1964				 pfn, &unpoison_rs);
1965		return 0;
1966	}
1967
1968	if (page_mapping(page)) {
1969		unpoison_pr_info("Unpoison: the hwpoison page has non-NULL mapping %#lx\n",
1970				 pfn, &unpoison_rs);
1971		return 0;
1972	}
1973
1974	/*
1975	 * unpoison_memory() can encounter thp only when the thp is being
1976	 * worked by memory_failure() and the page lock is not held yet.
1977	 * In such case, we yield to memory_failure() and make unpoison fail.
1978	 */
1979	if (!PageHuge(page) && PageTransHuge(page)) {
1980		unpoison_pr_info("Unpoison: Memory failure is now running on %#lx\n",
1981				 pfn, &unpoison_rs);
1982		return 0;
1983	}
1984
1985	if (!get_hwpoison_page(p, flags)) {
1986		if (TestClearPageHWPoison(p))
1987			num_poisoned_pages_dec();
1988		unpoison_pr_info("Unpoison: Software-unpoisoned free page %#lx\n",
1989				 pfn, &unpoison_rs);
1990		return 0;
1991	}
1992
1993	lock_page(page);
1994	/*
1995	 * This test is racy because PG_hwpoison is set outside of page lock.
1996	 * That's acceptable because that won't trigger kernel panic. Instead,
1997	 * the PG_hwpoison page will be caught and isolated on the entrance to
1998	 * the free buddy page pool.
1999	 */
2000	if (TestClearPageHWPoison(page)) {
2001		unpoison_pr_info("Unpoison: Software-unpoisoned page %#lx\n",
2002				 pfn, &unpoison_rs);
2003		num_poisoned_pages_dec();
2004		freeit = 1;
2005	}
2006	unlock_page(page);
2007
2008	put_page(page);
2009	if (freeit && !(pfn == my_zero_pfn(0) && page_count(p) == 1))
2010		put_page(page);
2011
2012	return 0;
2013}
2014EXPORT_SYMBOL(unpoison_memory);
2015
2016static bool isolate_page(struct page *page, struct list_head *pagelist)
2017{
2018	bool isolated = false;
2019	bool lru = PageLRU(page);
2020
2021	if (PageHuge(page)) {
2022		isolated = isolate_huge_page(page, pagelist);
2023	} else {
2024		if (lru)
2025			isolated = !isolate_lru_page(page);
2026		else
2027			isolated = !isolate_movable_page(page, ISOLATE_UNEVICTABLE);
2028
2029		if (isolated)
2030			list_add(&page->lru, pagelist);
2031	}
 
 
 
 
 
 
2032
2033	if (isolated && lru)
2034		inc_node_page_state(page, NR_ISOLATED_ANON +
2035				    page_is_file_lru(page));
2036
2037	/*
2038	 * If we succeed to isolate the page, we grabbed another refcount on
2039	 * the page, so we can safely drop the one we got from get_any_pages().
2040	 * If we failed to isolate the page, it means that we cannot go further
2041	 * and we will return an error, so drop the reference we got from
2042	 * get_any_pages() as well.
2043	 */
2044	put_page(page);
2045	return isolated;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2046}
2047
2048/*
2049 * __soft_offline_page handles hugetlb-pages and non-hugetlb pages.
2050 * If the page is a non-dirty unmapped page-cache page, it simply invalidates.
2051 * If the page is mapped, it migrates the contents over.
2052 */
2053static int __soft_offline_page(struct page *page)
2054{
2055	int ret = 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2056	unsigned long pfn = page_to_pfn(page);
2057	struct page *hpage = compound_head(page);
2058	char const *msg_page[] = {"page", "hugepage"};
2059	bool huge = PageHuge(page);
2060	LIST_HEAD(pagelist);
2061	struct migration_target_control mtc = {
2062		.nid = NUMA_NO_NODE,
2063		.gfp_mask = GFP_USER | __GFP_MOVABLE | __GFP_RETRY_MAYFAIL,
2064	};
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2065
2066	/*
2067	 * Check PageHWPoison again inside page lock because PageHWPoison
2068	 * is set by memory_failure() outside page lock. Note that
2069	 * memory_failure() also double-checks PageHWPoison inside page lock,
2070	 * so there's no race between soft_offline_page() and memory_failure().
2071	 */
2072	lock_page(page);
2073	if (!PageHuge(page))
2074		wait_on_page_writeback(page);
2075	if (PageHWPoison(page)) {
2076		unlock_page(page);
2077		put_page(page);
2078		pr_info("soft offline: %#lx page already poisoned\n", pfn);
2079		return 0;
2080	}
2081
2082	if (!PageHuge(page))
2083		/*
2084		 * Try to invalidate first. This should work for
2085		 * non dirty unmapped page cache pages.
2086		 */
2087		ret = invalidate_inode_page(page);
2088	unlock_page(page);
2089
2090	/*
2091	 * RED-PEN would be better to keep it isolated here, but we
2092	 * would need to fix isolation locking first.
2093	 */
2094	if (ret) {
 
2095		pr_info("soft_offline: %#lx: invalidated\n", pfn);
2096		page_handle_poison(page, false, true);
 
2097		return 0;
2098	}
2099
2100	if (isolate_page(hpage, &pagelist)) {
2101		ret = migrate_pages(&pagelist, alloc_migration_target, NULL,
2102			(unsigned long)&mtc, MIGRATE_SYNC, MR_MEMORY_FAILURE);
2103		if (!ret) {
2104			bool release = !huge;
2105
2106			if (!page_handle_poison(page, huge, release))
2107				ret = -EBUSY;
2108		} else {
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2109			if (!list_empty(&pagelist))
2110				putback_movable_pages(&pagelist);
2111
2112			pr_info("soft offline: %#lx: %s migration failed %d, type %lx (%pGp)\n",
2113				pfn, msg_page[huge], ret, page->flags, &page->flags);
2114			if (ret > 0)
2115				ret = -EBUSY;
2116		}
2117	} else {
2118		pr_info("soft offline: %#lx: %s isolation failed, page count %d, type %lx (%pGp)\n",
2119			pfn, msg_page[huge], page_count(page), page->flags, &page->flags);
2120		ret = -EBUSY;
2121	}
2122	return ret;
2123}
2124
2125static int soft_offline_in_use_page(struct page *page)
2126{
 
 
2127	struct page *hpage = compound_head(page);
2128
2129	if (!PageHuge(page) && PageTransHuge(hpage))
2130		if (try_to_split_thp_page(page, "soft offline") < 0)
 
 
 
 
 
 
 
2131			return -EBUSY;
2132	return __soft_offline_page(page);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2133}
2134
2135static int soft_offline_free_page(struct page *page)
2136{
2137	int rc = 0;
2138
2139	if (!page_handle_poison(page, true, false))
2140		rc = -EBUSY;
2141
 
 
 
 
 
 
2142	return rc;
2143}
2144
2145static void put_ref_page(struct page *page)
2146{
2147	if (page)
2148		put_page(page);
2149}
2150
2151/**
2152 * soft_offline_page - Soft offline a page.
2153 * @pfn: pfn to soft-offline
2154 * @flags: flags. Same as memory_failure().
2155 *
2156 * Returns 0 on success, otherwise negated errno.
2157 *
2158 * Soft offline a page, by migration or invalidation,
2159 * without killing anything. This is for the case when
2160 * a page is not corrupted yet (so it's still valid to access),
2161 * but has had a number of corrected errors and is better taken
2162 * out.
2163 *
2164 * The actual policy on when to do that is maintained by
2165 * user space.
2166 *
2167 * This should never impact any application or cause data loss,
2168 * however it might take some time.
2169 *
2170 * This is not a 100% solution for all memory, but tries to be
2171 * ``good enough'' for the majority of memory.
2172 */
2173int soft_offline_page(unsigned long pfn, int flags)
2174{
2175	int ret;
2176	bool try_again = true;
2177	struct page *page, *ref_page = NULL;
2178
2179	WARN_ON_ONCE(!pfn_valid(pfn) && (flags & MF_COUNT_INCREASED));
2180
2181	if (!pfn_valid(pfn))
2182		return -ENXIO;
2183	if (flags & MF_COUNT_INCREASED)
2184		ref_page = pfn_to_page(pfn);
2185
2186	/* Only online pages can be soft-offlined (esp., not ZONE_DEVICE). */
2187	page = pfn_to_online_page(pfn);
2188	if (!page) {
2189		put_ref_page(ref_page);
 
2190		return -EIO;
2191	}
2192
2193	if (PageHWPoison(page)) {
2194		pr_info("%s: %#lx page already poisoned\n", __func__, pfn);
2195		put_ref_page(ref_page);
2196		return 0;
 
2197	}
2198
2199retry:
2200	get_online_mems();
2201	ret = get_hwpoison_page(page, flags);
2202	put_online_mems();
2203
2204	if (ret > 0) {
2205		ret = soft_offline_in_use_page(page);
2206	} else if (ret == 0) {
2207		if (soft_offline_free_page(page) && try_again) {
2208			try_again = false;
2209			goto retry;
2210		}
2211	} else if (ret == -EIO) {
2212		pr_info("%s: %#lx: unknown page type: %lx (%pGp)\n",
2213			 __func__, pfn, page->flags, &page->flags);
2214	}
2215
2216	return ret;
2217}