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