1 /*
2 * CDDL HEADER START
3 *
4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
7 *
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
12 *
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 *
19 * CDDL HEADER END
20 */
21 /*
22 * Copyright 2010 Sun Microsystems, Inc. All rights reserved.
23 * Use is subject to license terms.
24 */
25
26 /*
27 * Page Retire - Big Theory Statement.
28 *
29 * This file handles removing sections of faulty memory from use when the
30 * user land FMA Diagnosis Engine requests that a page be removed or when
31 * a CE or UE is detected by the hardware.
32 *
33 * In the bad old days, the kernel side of Page Retire did a lot of the work
34 * on its own. Now, with the DE keeping track of errors, the kernel side is
35 * rather simple minded on most platforms.
36 *
37 * Errors are all reflected to the DE, and after digesting the error and
38 * looking at all previously reported errors, the DE decides what should
39 * be done about the current error. If the DE wants a particular page to
40 * be retired, then the kernel page retire code is invoked via an ioctl.
41 * On non-FMA platforms, the ue_drain and ce_drain paths ends up calling
42 * page retire to handle the error. Since page retire is just a simple
43 * mechanism it doesn't need to differentiate between the different callers.
44 *
45 * The p_toxic field in the page_t is used to indicate which errors have
46 * occurred and what action has been taken on a given page. Because errors are
47 * reported without regard to the locked state of a page, no locks are used
48 * to SET the error bits in p_toxic. However, in order to clear the error
49 * bits, the page_t must be held exclusively locked.
50 *
51 * When page_retire() is called, it must be able to acquire locks, sleep, etc.
52 * It must not be called from high-level interrupt context.
53 *
54 * Depending on how the requested page is being used at the time of the retire
55 * request (and on the availability of sufficient system resources), the page
56 * may be retired immediately, or just marked for retirement later. For
57 * example, locked pages are marked, while free pages are retired. Multiple
58 * requests may be made to retire the same page, although there is no need
59 * to: once the p_toxic flags are set, the page will be retired as soon as it
60 * can be exclusively locked.
61 *
62 * The retire mechanism is driven centrally out of page_unlock(). To expedite
63 * the retirement of pages, further requests for SE_SHARED locks are denied
64 * as long as a page retirement is pending. In addition, as long as pages are
65 * pending retirement a background thread runs periodically trying to retire
66 * those pages. Pages which could not be retired while the system is running
67 * are scrubbed prior to rebooting to avoid latent errors on the next boot.
68 *
69 * UE pages without persistent errors are scrubbed and returned to service.
70 * Recidivist pages, as well as FMA-directed requests for retirement, result
71 * in the page being taken out of service. Once the decision is made to take
72 * a page out of service, the page is cleared, hashed onto the retired_pages
73 * vnode, marked as retired, and it is unlocked. No other requesters (except
74 * for unretire) are allowed to lock retired pages.
75 *
76 * The public routines return (sadly) 0 if they worked and a non-zero error
77 * value if something went wrong. This is done for the ioctl side of the
78 * world to allow errors to be reflected all the way out to user land. The
79 * non-zero values are explained in comments atop each function.
80 */
81
82 /*
83 * Things to fix:
84 *
85 * 1. Trying to retire non-relocatable kvp pages may result in a
86 * quagmire. This is because seg_kmem() no longer keeps its pages locked,
87 * and calls page_lookup() in the free path; since kvp pages are modified
88 * and don't have a usable backing store, page_retire() can't do anything
89 * with them, and we'll keep denying the lock to seg_kmem_free() in a
90 * vicious cycle. To prevent that, we don't deny locks to kvp pages, and
91 * hence only try to retire a page from page_unlock() in the free path.
92 * Since most kernel pages are indefinitely held anyway, and don't
93 * participate in I/O, this is of little consequence.
94 *
95 * 2. Low memory situations will be interesting. If we don't have
96 * enough memory for page_relocate() to succeed, we won't be able to
97 * retire dirty pages; nobody will be able to push them out to disk
98 * either, since we aggressively deny the page lock. We could change
99 * fsflush so it can recognize this situation, grab the lock, and push
100 * the page out, where we'll catch it in the free path and retire it.
101 *
102 * 3. Beware of places that have code like this in them:
103 *
104 * if (! page_tryupgrade(pp)) {
105 * page_unlock(pp);
106 * while (! page_lock(pp, SE_EXCL, NULL, P_RECLAIM)) {
107 * / *NOTHING* /
108 * }
109 * }
110 * page_free(pp);
111 *
112 * The problem is that pp can change identity right after the
113 * page_unlock() call. In particular, page_retire() can step in
114 * there, change pp's identity, and hash pp onto the retired_vnode.
115 *
116 * Of course, other functions besides page_retire() can have the
117 * same effect. A kmem reader can waltz by, set up a mapping to the
118 * page, and then unlock the page. Page_free() will then go castors
119 * up. So if anybody is doing this, it's already a bug.
120 *
121 * 4. mdboot()'s call into page_retire_mdboot() should probably be
122 * moved lower. Where the call is made now, we can get into trouble
123 * by scrubbing a kernel page that is then accessed later.
124 */
125
126 #include <sys/types.h>
127 #include <sys/param.h>
128 #include <sys/systm.h>
129 #include <sys/mman.h>
130 #include <sys/vnode.h>
131 #include <sys/vfs_opreg.h>
132 #include <sys/cmn_err.h>
133 #include <sys/ksynch.h>
134 #include <sys/thread.h>
135 #include <sys/disp.h>
136 #include <sys/ontrap.h>
137 #include <sys/vmsystm.h>
138 #include <sys/mem_config.h>
139 #include <sys/atomic.h>
140 #include <sys/callb.h>
141 #include <sys/kobj.h>
142 #include <vm/page.h>
143 #include <vm/vm_dep.h>
144 #include <vm/as.h>
145 #include <vm/hat.h>
146 #include <vm/seg_kmem.h>
147
148 /*
149 * vnode for all pages which are retired from the VM system;
150 */
151 vnode_t *retired_pages;
152
153 static int page_retire_pp_finish(page_t *, void *, uint_t);
154
155 /*
156 * Make a list of all of the pages that have been marked for retirement
157 * but are not yet retired. At system shutdown, we will scrub all of the
158 * pages in the list in case there are outstanding UEs. Then, we
159 * cross-check this list against the number of pages that are yet to be
160 * retired, and if we find inconsistencies, we scan every page_t in the
161 * whole system looking for any pages that need to be scrubbed for UEs.
162 * The background thread also uses this queue to determine which pages
163 * it should keep trying to retire.
164 */
165 #ifdef DEBUG
166 #define PR_PENDING_QMAX 32
167 #else /* DEBUG */
168 #define PR_PENDING_QMAX 256
169 #endif /* DEBUG */
170 page_t *pr_pending_q[PR_PENDING_QMAX];
171 kmutex_t pr_q_mutex;
172
173 /*
174 * Page retire global kstats
175 */
176 struct page_retire_kstat {
177 kstat_named_t pr_retired;
178 kstat_named_t pr_requested;
179 kstat_named_t pr_requested_free;
180 kstat_named_t pr_enqueue_fail;
181 kstat_named_t pr_dequeue_fail;
182 kstat_named_t pr_pending;
183 kstat_named_t pr_pending_kas;
184 kstat_named_t pr_failed;
185 kstat_named_t pr_failed_kernel;
186 kstat_named_t pr_limit;
187 kstat_named_t pr_limit_exceeded;
188 kstat_named_t pr_fma;
189 kstat_named_t pr_mce;
190 kstat_named_t pr_ue;
191 kstat_named_t pr_ue_cleared_retire;
192 kstat_named_t pr_ue_cleared_free;
193 kstat_named_t pr_ue_persistent;
194 kstat_named_t pr_unretired;
195 };
196
197 static struct page_retire_kstat page_retire_kstat = {
198 { "pages_retired", KSTAT_DATA_UINT64},
199 { "pages_retire_request", KSTAT_DATA_UINT64},
200 { "pages_retire_request_free", KSTAT_DATA_UINT64},
201 { "pages_notenqueued", KSTAT_DATA_UINT64},
202 { "pages_notdequeued", KSTAT_DATA_UINT64},
203 { "pages_pending", KSTAT_DATA_UINT64},
204 { "pages_pending_kas", KSTAT_DATA_UINT64},
205 { "pages_deferred", KSTAT_DATA_UINT64},
206 { "pages_deferred_kernel", KSTAT_DATA_UINT64},
207 { "pages_limit", KSTAT_DATA_UINT64},
208 { "pages_limit_exceeded", KSTAT_DATA_UINT64},
209 { "pages_fma", KSTAT_DATA_UINT64},
210 { "pages_multiple_ce", KSTAT_DATA_UINT64},
211 { "pages_ue", KSTAT_DATA_UINT64},
212 { "pages_ue_cleared_retired", KSTAT_DATA_UINT64},
213 { "pages_ue_cleared_freed", KSTAT_DATA_UINT64},
214 { "pages_ue_persistent", KSTAT_DATA_UINT64},
215 { "pages_unretired", KSTAT_DATA_UINT64},
216 };
217
218 static kstat_t *page_retire_ksp = NULL;
219
220 #define PR_INCR_KSTAT(stat) \
221 atomic_inc_64(&(page_retire_kstat.stat.value.ui64))
222 #define PR_DECR_KSTAT(stat) \
223 atomic_dec_64(&(page_retire_kstat.stat.value.ui64))
224
225 #define PR_KSTAT_RETIRED_CE (page_retire_kstat.pr_mce.value.ui64)
226 #define PR_KSTAT_RETIRED_FMA (page_retire_kstat.pr_fma.value.ui64)
227 #define PR_KSTAT_RETIRED_NOTUE (PR_KSTAT_RETIRED_CE + PR_KSTAT_RETIRED_FMA)
228 #define PR_KSTAT_PENDING (page_retire_kstat.pr_pending.value.ui64)
229 #define PR_KSTAT_PENDING_KAS (page_retire_kstat.pr_pending_kas.value.ui64)
230 #define PR_KSTAT_EQFAIL (page_retire_kstat.pr_enqueue_fail.value.ui64)
231 #define PR_KSTAT_DQFAIL (page_retire_kstat.pr_dequeue_fail.value.ui64)
232
233 /*
234 * page retire kstats to list all retired pages
235 */
236 static int pr_list_kstat_update(kstat_t *ksp, int rw);
237 static int pr_list_kstat_snapshot(kstat_t *ksp, void *buf, int rw);
238 kmutex_t pr_list_kstat_mutex;
239
240 /*
241 * Limit the number of multiple CE page retires.
242 * The default is 0.1% of physmem, or 1 in 1000 pages. This is set in
243 * basis points, where 100 basis points equals one percent.
244 */
245 #define MCE_BPT 10
246 uint64_t max_pages_retired_bps = MCE_BPT;
247 #define PAGE_RETIRE_LIMIT ((physmem * max_pages_retired_bps) / 10000)
248
249 /*
250 * Control over the verbosity of page retirement.
251 *
252 * When set to zero (the default), no messages will be printed.
253 * When set to one, summary messages will be printed.
254 * When set > one, all messages will be printed.
255 *
256 * A value of one will trigger detailed messages for retirement operations,
257 * and is intended as a platform tunable for processors where FMA's DE does
258 * not run (e.g., spitfire). Values > one are intended for debugging only.
259 */
260 int page_retire_messages = 0;
261
262 /*
263 * Control whether or not we return scrubbed UE pages to service.
264 * By default we do not since FMA wants to run its diagnostics first
265 * and then ask us to unretire the page if it passes. Non-FMA platforms
266 * may set this to zero so we will only retire recidivist pages. It should
267 * not be changed by the user.
268 */
269 int page_retire_first_ue = 1;
270
271 /*
272 * Master enable for page retire. This prevents a CE or UE early in boot
273 * from trying to retire a page before page_retire_init() has finished
274 * setting things up. This is internal only and is not a tunable!
275 */
276 static int pr_enable = 0;
277
278 static void (*memscrub_notify_func)(uint64_t);
279
280 #ifdef DEBUG
281 struct page_retire_debug {
282 int prd_dup1;
283 int prd_dup2;
284 int prd_qdup;
285 int prd_noaction;
286 int prd_queued;
287 int prd_notqueued;
288 int prd_dequeue;
289 int prd_top;
290 int prd_locked;
291 int prd_reloc;
292 int prd_relocfail;
293 int prd_mod;
294 int prd_mod_late;
295 int prd_kern;
296 int prd_free;
297 int prd_noreclaim;
298 int prd_hashout;
299 int prd_fma;
300 int prd_uescrubbed;
301 int prd_uenotscrubbed;
302 int prd_mce;
303 int prd_prlocked;
304 int prd_prnotlocked;
305 int prd_prretired;
306 int prd_ulocked;
307 int prd_unotretired;
308 int prd_udestroy;
309 int prd_uhashout;
310 int prd_uunretired;
311 int prd_unotlocked;
312 int prd_checkhit;
313 int prd_checkmiss_pend;
314 int prd_checkmiss_noerr;
315 int prd_tctop;
316 int prd_tclocked;
317 int prd_hunt;
318 int prd_dohunt;
319 int prd_earlyhunt;
320 int prd_latehunt;
321 int prd_nofreedemote;
322 int prd_nodemote;
323 int prd_demoted;
324 } pr_debug;
325
326 #define PR_DEBUG(foo) ((pr_debug.foo)++)
327
328 /*
329 * A type histogram. We record the incidence of the various toxic
330 * flag combinations along with the interesting page attributes. The
331 * goal is to get as many combinations as we can while driving all
332 * pr_debug values nonzero (indicating we've exercised all possible
333 * code paths across all possible page types). Not all combinations
334 * will make sense -- e.g. PRT_MOD|PRT_KERNEL.
335 *
336 * pr_type offset bit encoding (when examining with a debugger):
337 *
338 * PRT_NAMED - 0x4
339 * PRT_KERNEL - 0x8
340 * PRT_FREE - 0x10
341 * PRT_MOD - 0x20
342 * PRT_FMA - 0x0
343 * PRT_MCE - 0x40
344 * PRT_UE - 0x80
345 */
346
347 #define PRT_NAMED 0x01
348 #define PRT_KERNEL 0x02
349 #define PRT_FREE 0x04
350 #define PRT_MOD 0x08
351 #define PRT_FMA 0x00 /* yes, this is not a mistake */
352 #define PRT_MCE 0x10
353 #define PRT_UE 0x20
354 #define PRT_ALL 0x3F
355
356 int pr_types[PRT_ALL+1];
357
358 #define PR_TYPES(pp) { \
359 int whichtype = 0; \
360 if (pp->p_vnode) \
361 whichtype |= PRT_NAMED; \
362 if (PP_ISKAS(pp)) \
363 whichtype |= PRT_KERNEL; \
364 if (PP_ISFREE(pp)) \
365 whichtype |= PRT_FREE; \
366 if (hat_ismod(pp)) \
367 whichtype |= PRT_MOD; \
368 if (pp->p_toxic & PR_UE) \
369 whichtype |= PRT_UE; \
370 if (pp->p_toxic & PR_MCE) \
371 whichtype |= PRT_MCE; \
372 pr_types[whichtype]++; \
373 }
374
375 int recl_calls;
376 int recl_mtbf = 3;
377 int reloc_calls;
378 int reloc_mtbf = 7;
379 int pr_calls;
380 int pr_mtbf = 15;
381
382 #define MTBF(v, f) (((++(v)) & (f)) != (f))
383
384 #else /* DEBUG */
385
386 #define PR_DEBUG(foo) /* nothing */
387 #define PR_TYPES(foo) /* nothing */
388 #define MTBF(v, f) (1)
389
390 #endif /* DEBUG */
391
392 /*
393 * page_retire_done() - completion processing
394 *
395 * Used by the page_retire code for common completion processing.
396 * It keeps track of how many times a given result has happened,
397 * and writes out an occasional message.
398 *
399 * May be called with a NULL pp (PRD_INVALID_PA case).
400 */
401 #define PRD_INVALID_KEY -1
402 #define PRD_SUCCESS 0
403 #define PRD_PENDING 1
404 #define PRD_FAILED 2
405 #define PRD_DUPLICATE 3
406 #define PRD_INVALID_PA 4
407 #define PRD_LIMIT 5
408 #define PRD_UE_SCRUBBED 6
409 #define PRD_UNR_SUCCESS 7
410 #define PRD_UNR_CANTLOCK 8
411 #define PRD_UNR_NOT 9
412
413 typedef struct page_retire_op {
414 int pr_key; /* one of the PRD_* defines from above */
415 int pr_count; /* How many times this has happened */
416 int pr_retval; /* return value */
417 int pr_msglvl; /* message level - when to print */
418 char *pr_message; /* Cryptic message for field service */
419 } page_retire_op_t;
420
421 static page_retire_op_t page_retire_ops[] = {
422 /* key count retval msglvl message */
423 {PRD_SUCCESS, 0, 0, 1,
424 "Page 0x%08x.%08x removed from service"},
425 {PRD_PENDING, 0, EAGAIN, 2,
426 "Page 0x%08x.%08x will be retired on free"},
427 {PRD_FAILED, 0, EAGAIN, 0, NULL},
428 {PRD_DUPLICATE, 0, EIO, 2,
429 "Page 0x%08x.%08x already retired or pending"},
430 {PRD_INVALID_PA, 0, EINVAL, 2,
431 "PA 0x%08x.%08x is not a relocatable page"},
432 {PRD_LIMIT, 0, 0, 1,
433 "Page 0x%08x.%08x not retired due to limit exceeded"},
434 {PRD_UE_SCRUBBED, 0, 0, 1,
435 "Previously reported error on page 0x%08x.%08x cleared"},
436 {PRD_UNR_SUCCESS, 0, 0, 1,
437 "Page 0x%08x.%08x returned to service"},
438 {PRD_UNR_CANTLOCK, 0, EAGAIN, 2,
439 "Page 0x%08x.%08x could not be unretired"},
440 {PRD_UNR_NOT, 0, EIO, 2,
441 "Page 0x%08x.%08x is not retired"},
442 {PRD_INVALID_KEY, 0, 0, 0, NULL} /* MUST BE LAST! */
443 };
444
445 /*
446 * print a message if page_retire_messages is true.
447 */
448 #define PR_MESSAGE(debuglvl, msglvl, msg, pa) \
449 { \
450 uint64_t p = (uint64_t)pa; \
451 if (page_retire_messages >= msglvl && msg != NULL) { \
452 cmn_err(debuglvl, msg, \
453 (uint32_t)(p >> 32), (uint32_t)p); \
454 } \
455 }
456
457 /*
458 * Note that multiple bits may be set in a single settoxic operation.
459 * May be called without the page locked.
460 */
461 void
page_settoxic(page_t * pp,uchar_t bits)462 page_settoxic(page_t *pp, uchar_t bits)
463 {
464 atomic_or_8(&pp->p_toxic, bits);
465 }
466
467 /*
468 * Note that multiple bits may cleared in a single clrtoxic operation.
469 * Must be called with the page exclusively locked to prevent races which
470 * may attempt to retire a page without any toxic bits set.
471 * Note that the PR_CAPTURE bit can be cleared without the exclusive lock
472 * being held as there is a separate mutex which protects that bit.
473 */
474 void
page_clrtoxic(page_t * pp,uchar_t bits)475 page_clrtoxic(page_t *pp, uchar_t bits)
476 {
477 ASSERT((bits & PR_CAPTURE) || PAGE_EXCL(pp));
478 atomic_and_8(&pp->p_toxic, ~bits);
479 }
480
481 /*
482 * Prints any page retire messages to the user, and decides what
483 * error code is appropriate for the condition reported.
484 */
485 static int
page_retire_done(page_t * pp,int code)486 page_retire_done(page_t *pp, int code)
487 {
488 page_retire_op_t *prop;
489 uint64_t pa = 0;
490 int i;
491
492 if (pp != NULL) {
493 pa = mmu_ptob((uint64_t)pp->p_pagenum);
494 }
495
496 prop = NULL;
497 for (i = 0; page_retire_ops[i].pr_key != PRD_INVALID_KEY; i++) {
498 if (page_retire_ops[i].pr_key == code) {
499 prop = &page_retire_ops[i];
500 break;
501 }
502 }
503
504 #ifdef DEBUG
505 if (page_retire_ops[i].pr_key == PRD_INVALID_KEY) {
506 cmn_err(CE_PANIC, "page_retire_done: Invalid opcode %d", code);
507 }
508 #endif
509
510 ASSERT(prop->pr_key == code);
511
512 prop->pr_count++;
513
514 PR_MESSAGE(CE_NOTE, prop->pr_msglvl, prop->pr_message, pa);
515 if (pp != NULL) {
516 page_settoxic(pp, PR_MSG);
517 }
518
519 return (prop->pr_retval);
520 }
521
522 /*
523 * Act like page_destroy(), but instead of freeing the page, hash it onto
524 * the retired_pages vnode, and mark it retired.
525 *
526 * For fun, we try to scrub the page until it's squeaky clean.
527 * availrmem is adjusted here.
528 */
529 static void
page_retire_destroy(page_t * pp)530 page_retire_destroy(page_t *pp)
531 {
532 u_offset_t off = (u_offset_t)((uintptr_t)pp);
533
534 ASSERT(PAGE_EXCL(pp));
535 ASSERT(!PP_ISFREE(pp));
536 ASSERT(pp->p_szc == 0);
537 ASSERT(!hat_page_is_mapped(pp));
538 ASSERT(!pp->p_vnode);
539
540 page_clr_all_props(pp);
541 pagescrub(pp, 0, MMU_PAGESIZE);
542
543 pp->p_next = NULL;
544 pp->p_prev = NULL;
545 if (page_hashin(pp, retired_pages, off, NULL) == 0) {
546 cmn_err(CE_PANIC, "retired page %p hashin failed", (void *)pp);
547 }
548
549 page_settoxic(pp, PR_RETIRED);
550 PR_INCR_KSTAT(pr_retired);
551
552 if (pp->p_toxic & PR_FMA) {
553 PR_INCR_KSTAT(pr_fma);
554 } else if (pp->p_toxic & PR_UE) {
555 PR_INCR_KSTAT(pr_ue);
556 } else {
557 PR_INCR_KSTAT(pr_mce);
558 }
559
560 mutex_enter(&freemem_lock);
561 availrmem--;
562 mutex_exit(&freemem_lock);
563
564 page_unlock(pp);
565 }
566
567 /*
568 * Check whether the number of pages which have been retired already exceeds
569 * the maximum allowable percentage of memory which may be retired.
570 *
571 * Returns 1 if the limit has been exceeded.
572 */
573 static int
page_retire_limit(void)574 page_retire_limit(void)
575 {
576 if (PR_KSTAT_RETIRED_NOTUE >= (uint64_t)PAGE_RETIRE_LIMIT) {
577 PR_INCR_KSTAT(pr_limit_exceeded);
578 return (1);
579 }
580
581 return (0);
582 }
583
584 #define MSG_DM "Data Mismatch occurred at PA 0x%08x.%08x" \
585 "[ 0x%x != 0x%x ] while attempting to clear previously " \
586 "reported error; page removed from service"
587
588 #define MSG_UE "Uncorrectable Error occurred at PA 0x%08x.%08x while " \
589 "attempting to clear previously reported error; page removed " \
590 "from service"
591
592 /*
593 * Attempt to clear a UE from a page.
594 * Returns 1 if the error has been successfully cleared.
595 */
596 static int
page_clear_transient_ue(page_t * pp)597 page_clear_transient_ue(page_t *pp)
598 {
599 caddr_t kaddr;
600 uint8_t rb, wb;
601 uint64_t pa;
602 uint32_t pa_hi, pa_lo;
603 on_trap_data_t otd;
604 int errors = 0;
605 int i;
606
607 ASSERT(PAGE_EXCL(pp));
608 ASSERT(PP_PR_REQ(pp));
609 ASSERT(pp->p_szc == 0);
610 ASSERT(!hat_page_is_mapped(pp));
611
612 /*
613 * Clear the page and attempt to clear the UE. If we trap
614 * on the next access to the page, we know the UE has recurred.
615 */
616 pagescrub(pp, 0, PAGESIZE);
617
618 /*
619 * Map the page and write a bunch of bit patterns to compare
620 * what we wrote with what we read back. This isn't a perfect
621 * test but it should be good enough to catch most of the
622 * recurring UEs. If this fails to catch a recurrent UE, we'll
623 * retire the page the next time we see a UE on the page.
624 */
625 kaddr = ppmapin(pp, PROT_READ|PROT_WRITE, (caddr_t)-1);
626
627 pa = ptob((uint64_t)page_pptonum(pp));
628 pa_hi = (uint32_t)(pa >> 32);
629 pa_lo = (uint32_t)pa;
630
631 /*
632 * Disable preemption to prevent the off chance that
633 * we migrate while in the middle of running through
634 * the bit pattern and run on a different processor
635 * than what we started on.
636 */
637 kpreempt_disable();
638
639 /*
640 * Fill the page with each (0x00 - 0xFF] bit pattern, flushing
641 * the cache in between reading and writing. We do this under
642 * on_trap() protection to avoid recursion.
643 */
644 if (on_trap(&otd, OT_DATA_EC)) {
645 PR_MESSAGE(CE_WARN, 1, MSG_UE, pa);
646 errors = 1;
647 } else {
648 for (wb = 0xff; wb > 0; wb--) {
649 for (i = 0; i < PAGESIZE; i++) {
650 kaddr[i] = wb;
651 }
652
653 sync_data_memory(kaddr, PAGESIZE);
654
655 for (i = 0; i < PAGESIZE; i++) {
656 rb = kaddr[i];
657 if (rb != wb) {
658 /*
659 * We had a mismatch without a trap.
660 * Uh-oh. Something is really wrong
661 * with this system.
662 */
663 if (page_retire_messages) {
664 cmn_err(CE_WARN, MSG_DM,
665 pa_hi, pa_lo, rb, wb);
666 }
667 errors = 1;
668 goto out; /* double break */
669 }
670 }
671 }
672 }
673 out:
674 no_trap();
675 kpreempt_enable();
676 ppmapout(kaddr);
677
678 return (errors ? 0 : 1);
679 }
680
681 /*
682 * Try to clear a page_t with a single UE. If the UE was transient, it is
683 * returned to service, and we return 1. Otherwise we return 0 meaning
684 * that further processing is required to retire the page.
685 */
686 static int
page_retire_transient_ue(page_t * pp)687 page_retire_transient_ue(page_t *pp)
688 {
689 ASSERT(PAGE_EXCL(pp));
690 ASSERT(!hat_page_is_mapped(pp));
691
692 /*
693 * If this page is a repeat offender, retire him under the
694 * "two strikes and you're out" rule. The caller is responsible
695 * for scrubbing the page to try to clear the error.
696 */
697 if (pp->p_toxic & PR_UE_SCRUBBED) {
698 PR_INCR_KSTAT(pr_ue_persistent);
699 return (0);
700 }
701
702 if (page_clear_transient_ue(pp)) {
703 /*
704 * We set the PR_SCRUBBED_UE bit; if we ever see this
705 * page again, we will retire it, no questions asked.
706 */
707 page_settoxic(pp, PR_UE_SCRUBBED);
708
709 if (page_retire_first_ue) {
710 PR_INCR_KSTAT(pr_ue_cleared_retire);
711 return (0);
712 } else {
713 PR_INCR_KSTAT(pr_ue_cleared_free);
714
715 page_clrtoxic(pp, PR_UE | PR_MCE | PR_MSG);
716
717 /* LINTED: CONSTCOND */
718 VN_DISPOSE(pp, B_FREE, 1, kcred);
719 return (1);
720 }
721 }
722
723 PR_INCR_KSTAT(pr_ue_persistent);
724 return (0);
725 }
726
727 /*
728 * Update the statistics dynamically when our kstat is read.
729 */
730 static int
page_retire_kstat_update(kstat_t * ksp,int rw)731 page_retire_kstat_update(kstat_t *ksp, int rw)
732 {
733 struct page_retire_kstat *pr;
734
735 if (ksp == NULL)
736 return (EINVAL);
737
738 switch (rw) {
739
740 case KSTAT_READ:
741 pr = (struct page_retire_kstat *)ksp->ks_data;
742 ASSERT(pr == &page_retire_kstat);
743 pr->pr_limit.value.ui64 = PAGE_RETIRE_LIMIT;
744 return (0);
745
746 case KSTAT_WRITE:
747 return (EACCES);
748
749 default:
750 return (EINVAL);
751 }
752 /*NOTREACHED*/
753 }
754
755 static int
pr_list_kstat_update(kstat_t * ksp,int rw)756 pr_list_kstat_update(kstat_t *ksp, int rw)
757 {
758 uint_t count;
759 page_t *pp;
760 kmutex_t *vphm;
761
762 if (rw == KSTAT_WRITE)
763 return (EACCES);
764
765 vphm = page_vnode_mutex(retired_pages);
766 mutex_enter(vphm);
767 /* Needs to be under a lock so that for loop will work right */
768 if (retired_pages->v_pages == NULL) {
769 mutex_exit(vphm);
770 ksp->ks_ndata = 0;
771 ksp->ks_data_size = 0;
772 return (0);
773 }
774
775 count = 1;
776 for (pp = retired_pages->v_pages->p_vpnext;
777 pp != retired_pages->v_pages; pp = pp->p_vpnext) {
778 count++;
779 }
780 mutex_exit(vphm);
781
782 ksp->ks_ndata = count;
783 ksp->ks_data_size = count * 2 * sizeof (uint64_t);
784
785 return (0);
786 }
787
788 /*
789 * all spans will be pagesize and no coalescing will be done with the
790 * list produced.
791 */
792 static int
pr_list_kstat_snapshot(kstat_t * ksp,void * buf,int rw)793 pr_list_kstat_snapshot(kstat_t *ksp, void *buf, int rw)
794 {
795 kmutex_t *vphm;
796 page_t *pp;
797 struct memunit {
798 uint64_t address;
799 uint64_t size;
800 } *kspmem;
801
802 if (rw == KSTAT_WRITE)
803 return (EACCES);
804
805 ksp->ks_snaptime = gethrtime();
806
807 kspmem = (struct memunit *)buf;
808
809 vphm = page_vnode_mutex(retired_pages);
810 mutex_enter(vphm);
811 pp = retired_pages->v_pages;
812 if (((caddr_t)kspmem >= (caddr_t)buf + ksp->ks_data_size) ||
813 (pp == NULL)) {
814 mutex_exit(vphm);
815 return (0);
816 }
817 kspmem->address = ptob(pp->p_pagenum);
818 kspmem->size = PAGESIZE;
819 kspmem++;
820 for (pp = pp->p_vpnext; pp != retired_pages->v_pages;
821 pp = pp->p_vpnext, kspmem++) {
822 if ((caddr_t)kspmem >= (caddr_t)buf + ksp->ks_data_size)
823 break;
824 kspmem->address = ptob(pp->p_pagenum);
825 kspmem->size = PAGESIZE;
826 }
827 mutex_exit(vphm);
828
829 return (0);
830 }
831
832 /*
833 * page_retire_pend_count -- helper function for page_capture_thread,
834 * returns the number of pages pending retirement.
835 */
836 uint64_t
page_retire_pend_count(void)837 page_retire_pend_count(void)
838 {
839 return (PR_KSTAT_PENDING);
840 }
841
842 uint64_t
page_retire_pend_kas_count(void)843 page_retire_pend_kas_count(void)
844 {
845 return (PR_KSTAT_PENDING_KAS);
846 }
847
848 void
page_retire_incr_pend_count(void * datap)849 page_retire_incr_pend_count(void *datap)
850 {
851 PR_INCR_KSTAT(pr_pending);
852
853 if ((datap == &kvp) || (datap == &zvp)) {
854 PR_INCR_KSTAT(pr_pending_kas);
855 }
856 }
857
858 void
page_retire_decr_pend_count(void * datap)859 page_retire_decr_pend_count(void *datap)
860 {
861 PR_DECR_KSTAT(pr_pending);
862
863 if ((datap == &kvp) || (datap == &zvp)) {
864 PR_DECR_KSTAT(pr_pending_kas);
865 }
866 }
867
868 /*
869 * Initialize the page retire mechanism:
870 *
871 * - Establish the correctable error retire limit.
872 * - Initialize locks.
873 * - Build the retired_pages vnode.
874 * - Set up the kstats.
875 * - Fire off the background thread.
876 * - Tell page_retire() it's OK to start retiring pages.
877 */
878 void
page_retire_init(void)879 page_retire_init(void)
880 {
881 const fs_operation_def_t retired_vnodeops_template[] = {
882 { NULL, NULL }
883 };
884 struct vnodeops *vops;
885 kstat_t *ksp;
886
887 const uint_t page_retire_ndata =
888 sizeof (page_retire_kstat) / sizeof (kstat_named_t);
889
890 ASSERT(page_retire_ksp == NULL);
891
892 if (max_pages_retired_bps <= 0) {
893 max_pages_retired_bps = MCE_BPT;
894 }
895
896 mutex_init(&pr_q_mutex, NULL, MUTEX_DEFAULT, NULL);
897
898 retired_pages = vn_alloc(KM_SLEEP);
899 if (vn_make_ops("retired_pages", retired_vnodeops_template, &vops)) {
900 cmn_err(CE_PANIC,
901 "page_retired_init: can't make retired vnodeops");
902 }
903 vn_setops(retired_pages, vops);
904
905 if ((page_retire_ksp = kstat_create("unix", 0, "page_retire",
906 "misc", KSTAT_TYPE_NAMED, page_retire_ndata,
907 KSTAT_FLAG_VIRTUAL)) == NULL) {
908 cmn_err(CE_WARN, "kstat_create for page_retire failed");
909 } else {
910 page_retire_ksp->ks_data = (void *)&page_retire_kstat;
911 page_retire_ksp->ks_update = page_retire_kstat_update;
912 kstat_install(page_retire_ksp);
913 }
914
915 mutex_init(&pr_list_kstat_mutex, NULL, MUTEX_DEFAULT, NULL);
916 ksp = kstat_create("unix", 0, "page_retire_list", "misc",
917 KSTAT_TYPE_RAW, 0, KSTAT_FLAG_VAR_SIZE | KSTAT_FLAG_VIRTUAL);
918 if (ksp != NULL) {
919 ksp->ks_update = pr_list_kstat_update;
920 ksp->ks_snapshot = pr_list_kstat_snapshot;
921 ksp->ks_lock = &pr_list_kstat_mutex;
922 kstat_install(ksp);
923 }
924
925 memscrub_notify_func =
926 (void(*)(uint64_t))kobj_getsymvalue("memscrub_notify", 0);
927
928 page_capture_register_callback(PC_RETIRE, -1, page_retire_pp_finish);
929 pr_enable = 1;
930 }
931
932 /*
933 * page_retire_hunt() callback for the retire thread.
934 */
935 static void
page_retire_thread_cb(page_t * pp)936 page_retire_thread_cb(page_t *pp)
937 {
938 PR_DEBUG(prd_tctop);
939 if (!PP_ISKAS(pp) && page_trylock(pp, SE_EXCL)) {
940 PR_DEBUG(prd_tclocked);
941 page_unlock(pp);
942 }
943 }
944
945 /*
946 * Callback used by page_trycapture() to finish off retiring a page.
947 * The page has already been cleaned and we've been given sole access to
948 * it.
949 * Always returns 0 to indicate that callback succeded as the callback never
950 * fails to finish retiring the given page.
951 */
952 /*ARGSUSED*/
953 static int
page_retire_pp_finish(page_t * pp,void * notused,uint_t flags)954 page_retire_pp_finish(page_t *pp, void *notused, uint_t flags)
955 {
956 int toxic;
957
958 ASSERT(PAGE_EXCL(pp));
959 ASSERT(pp->p_iolock_state == 0);
960 ASSERT(pp->p_szc == 0);
961
962 toxic = pp->p_toxic;
963
964 /*
965 * The problem page is locked, demoted, unmapped, not free,
966 * hashed out, and not COW or mlocked (whew!).
967 *
968 * Now we select our ammunition, take it around back, and shoot it.
969 */
970 if (toxic & PR_UE) {
971 ue_error:
972 if (page_retire_transient_ue(pp)) {
973 PR_DEBUG(prd_uescrubbed);
974 (void) page_retire_done(pp, PRD_UE_SCRUBBED);
975 } else {
976 PR_DEBUG(prd_uenotscrubbed);
977 page_retire_destroy(pp);
978 (void) page_retire_done(pp, PRD_SUCCESS);
979 }
980 return (0);
981 } else if (toxic & PR_FMA) {
982 PR_DEBUG(prd_fma);
983 page_retire_destroy(pp);
984 (void) page_retire_done(pp, PRD_SUCCESS);
985 return (0);
986 } else if (toxic & PR_MCE) {
987 PR_DEBUG(prd_mce);
988 page_retire_destroy(pp);
989 (void) page_retire_done(pp, PRD_SUCCESS);
990 return (0);
991 }
992
993 /*
994 * When page_retire_first_ue is set to zero and a UE occurs which is
995 * transient, it's possible that we clear some flags set by a second
996 * UE error on the page which occurs while the first is currently being
997 * handled and thus we need to handle the case where none of the above
998 * are set. In this instance, PR_UE_SCRUBBED should be set and thus
999 * we should execute the UE code above.
1000 */
1001 if (toxic & PR_UE_SCRUBBED) {
1002 goto ue_error;
1003 }
1004
1005 /*
1006 * It's impossible to get here.
1007 */
1008 panic("bad toxic flags 0x%x in page_retire_pp_finish\n", toxic);
1009 return (0);
1010 }
1011
1012 /*
1013 * page_retire() - the front door in to retire a page.
1014 *
1015 * Ideally, page_retire() would instantly retire the requested page.
1016 * Unfortunately, some pages are locked or otherwise tied up and cannot be
1017 * retired right away. We use the page capture logic to deal with this
1018 * situation as it will continuously try to retire the page in the background
1019 * if the first attempt fails. Success is determined by looking to see whether
1020 * the page has been retired after the page_trycapture() attempt.
1021 *
1022 * Returns:
1023 *
1024 * - 0 on success,
1025 * - EINVAL when the PA is whacko,
1026 * - EIO if the page is already retired or already pending retirement, or
1027 * - EAGAIN if the page could not be _immediately_ retired but is pending.
1028 */
1029 int
page_retire(uint64_t pa,uchar_t reason)1030 page_retire(uint64_t pa, uchar_t reason)
1031 {
1032 page_t *pp;
1033
1034 ASSERT(reason & PR_REASONS); /* there must be a reason */
1035 ASSERT(!(reason & ~PR_REASONS)); /* but no other bits */
1036
1037 pp = page_numtopp_nolock(mmu_btop(pa));
1038 if (pp == NULL) {
1039 PR_MESSAGE(CE_WARN, 1, "Cannot schedule clearing of error on"
1040 " page 0x%08x.%08x; page is not relocatable memory", pa);
1041 return (page_retire_done(pp, PRD_INVALID_PA));
1042 }
1043 if (PP_RETIRED(pp)) {
1044 PR_DEBUG(prd_dup1);
1045 return (page_retire_done(pp, PRD_DUPLICATE));
1046 }
1047
1048 if (memscrub_notify_func != NULL) {
1049 (void) memscrub_notify_func(pa);
1050 }
1051
1052 if ((reason & PR_UE) && !PP_TOXIC(pp)) {
1053 PR_MESSAGE(CE_NOTE, 1, "Scheduling clearing of error on"
1054 " page 0x%08x.%08x", pa);
1055 } else if (PP_PR_REQ(pp)) {
1056 PR_DEBUG(prd_dup2);
1057 return (page_retire_done(pp, PRD_DUPLICATE));
1058 } else {
1059 PR_MESSAGE(CE_NOTE, 1, "Scheduling removal of"
1060 " page 0x%08x.%08x", pa);
1061 }
1062
1063 /* Avoid setting toxic bits in the first place */
1064 if ((reason & (PR_FMA | PR_MCE)) && !(reason & PR_UE) &&
1065 page_retire_limit()) {
1066 return (page_retire_done(pp, PRD_LIMIT));
1067 }
1068
1069 if (MTBF(pr_calls, pr_mtbf)) {
1070 page_settoxic(pp, reason);
1071 if (page_trycapture(pp, 0, CAPTURE_RETIRE, pp->p_vnode) == 0) {
1072 PR_DEBUG(prd_prlocked);
1073 } else {
1074 PR_DEBUG(prd_prnotlocked);
1075 }
1076 } else {
1077 PR_DEBUG(prd_prnotlocked);
1078 }
1079
1080 if (PP_RETIRED(pp)) {
1081 PR_DEBUG(prd_prretired);
1082 return (0);
1083 } else {
1084 cv_signal(&pc_cv);
1085 PR_INCR_KSTAT(pr_failed);
1086
1087 if (pp->p_toxic & PR_MSG) {
1088 return (page_retire_done(pp, PRD_FAILED));
1089 } else {
1090 return (page_retire_done(pp, PRD_PENDING));
1091 }
1092 }
1093 }
1094
1095 /*
1096 * Take a retired page off the retired-pages vnode and clear the toxic flags.
1097 * If "free" is nonzero, lock it and put it back on the freelist. If "free"
1098 * is zero, the caller already holds SE_EXCL lock so we simply unretire it
1099 * and don't do anything else with it.
1100 *
1101 * Any unretire messages are printed from this routine.
1102 *
1103 * Returns 0 if page pp was unretired; else an error code.
1104 *
1105 * If flags is:
1106 * PR_UNR_FREE - lock the page, clear the toxic flags and free it
1107 * to the freelist.
1108 * PR_UNR_TEMP - lock the page, unretire it, leave the toxic
1109 * bits set as is and return it to the caller.
1110 * PR_UNR_CLEAN - page is SE_EXCL locked, unretire it, clear the
1111 * toxic flags and return it to caller as is.
1112 */
1113 int
page_unretire_pp(page_t * pp,int flags)1114 page_unretire_pp(page_t *pp, int flags)
1115 {
1116 /*
1117 * To be retired, a page has to be hashed onto the retired_pages vnode
1118 * and have PR_RETIRED set in p_toxic.
1119 */
1120 if (flags == PR_UNR_CLEAN ||
1121 page_try_reclaim_lock(pp, SE_EXCL, SE_RETIRED)) {
1122 ASSERT(PAGE_EXCL(pp));
1123 PR_DEBUG(prd_ulocked);
1124 if (!PP_RETIRED(pp)) {
1125 PR_DEBUG(prd_unotretired);
1126 page_unlock(pp);
1127 return (page_retire_done(pp, PRD_UNR_NOT));
1128 }
1129
1130 PR_MESSAGE(CE_NOTE, 1, "unretiring retired"
1131 " page 0x%08x.%08x", mmu_ptob((uint64_t)pp->p_pagenum));
1132 if (pp->p_toxic & PR_FMA) {
1133 PR_DECR_KSTAT(pr_fma);
1134 } else if (pp->p_toxic & PR_UE) {
1135 PR_DECR_KSTAT(pr_ue);
1136 } else {
1137 PR_DECR_KSTAT(pr_mce);
1138 }
1139
1140 if (flags == PR_UNR_TEMP)
1141 page_clrtoxic(pp, PR_RETIRED);
1142 else
1143 page_clrtoxic(pp, PR_TOXICFLAGS);
1144
1145 if (flags == PR_UNR_FREE) {
1146 PR_DEBUG(prd_udestroy);
1147 page_destroy(pp, 0);
1148 } else {
1149 PR_DEBUG(prd_uhashout);
1150 page_hashout(pp, NULL);
1151 }
1152
1153 mutex_enter(&freemem_lock);
1154 availrmem++;
1155 mutex_exit(&freemem_lock);
1156
1157 PR_DEBUG(prd_uunretired);
1158 PR_DECR_KSTAT(pr_retired);
1159 PR_INCR_KSTAT(pr_unretired);
1160 return (page_retire_done(pp, PRD_UNR_SUCCESS));
1161 }
1162 PR_DEBUG(prd_unotlocked);
1163 return (page_retire_done(pp, PRD_UNR_CANTLOCK));
1164 }
1165
1166 /*
1167 * Return a page to service by moving it from the retired_pages vnode
1168 * onto the freelist.
1169 *
1170 * Called from mmioctl_page_retire() on behalf of the FMA DE.
1171 *
1172 * Returns:
1173 *
1174 * - 0 if the page is unretired,
1175 * - EAGAIN if the pp can not be locked,
1176 * - EINVAL if the PA is whacko, and
1177 * - EIO if the pp is not retired.
1178 */
1179 int
page_unretire(uint64_t pa)1180 page_unretire(uint64_t pa)
1181 {
1182 page_t *pp;
1183
1184 pp = page_numtopp_nolock(mmu_btop(pa));
1185 if (pp == NULL) {
1186 return (page_retire_done(pp, PRD_INVALID_PA));
1187 }
1188
1189 return (page_unretire_pp(pp, PR_UNR_FREE));
1190 }
1191
1192 /*
1193 * Test a page to see if it is retired. If errors is non-NULL, the toxic
1194 * bits of the page are returned. Returns 0 on success, error code on failure.
1195 */
1196 int
page_retire_check_pp(page_t * pp,uint64_t * errors)1197 page_retire_check_pp(page_t *pp, uint64_t *errors)
1198 {
1199 int rc;
1200
1201 if (PP_RETIRED(pp)) {
1202 PR_DEBUG(prd_checkhit);
1203 rc = 0;
1204 } else if (PP_PR_REQ(pp)) {
1205 PR_DEBUG(prd_checkmiss_pend);
1206 rc = EAGAIN;
1207 } else {
1208 PR_DEBUG(prd_checkmiss_noerr);
1209 rc = EIO;
1210 }
1211
1212 /*
1213 * We have magically arranged the bit values returned to fmd(1M)
1214 * to line up with the FMA, MCE, and UE bits of the page_t.
1215 */
1216 if (errors) {
1217 uint64_t toxic = (uint64_t)(pp->p_toxic & PR_ERRMASK);
1218 if (toxic & PR_UE_SCRUBBED) {
1219 toxic &= ~PR_UE_SCRUBBED;
1220 toxic |= PR_UE;
1221 }
1222 *errors = toxic;
1223 }
1224
1225 return (rc);
1226 }
1227
1228 /*
1229 * Test to see if the page_t for a given PA is retired, and return the
1230 * hardware errors we have seen on the page if requested.
1231 *
1232 * Called from mmioctl_page_retire on behalf of the FMA DE.
1233 *
1234 * Returns:
1235 *
1236 * - 0 if the page is retired,
1237 * - EIO if the page is not retired and has no errors,
1238 * - EAGAIN if the page is not retired but is pending; and
1239 * - EINVAL if the PA is whacko.
1240 */
1241 int
page_retire_check(uint64_t pa,uint64_t * errors)1242 page_retire_check(uint64_t pa, uint64_t *errors)
1243 {
1244 page_t *pp;
1245
1246 if (errors) {
1247 *errors = 0;
1248 }
1249
1250 pp = page_numtopp_nolock(mmu_btop(pa));
1251 if (pp == NULL) {
1252 return (page_retire_done(pp, PRD_INVALID_PA));
1253 }
1254
1255 return (page_retire_check_pp(pp, errors));
1256 }
1257
1258 /*
1259 * Page retire self-test. For now, it always returns 0.
1260 */
1261 int
page_retire_test(void)1262 page_retire_test(void)
1263 {
1264 page_t *first, *pp, *cpp, *cpp2, *lpp;
1265
1266 /*
1267 * Tests the corner case where a large page can't be retired
1268 * because one of the constituent pages is locked. We mark
1269 * one page to be retired and try to retire it, and mark the
1270 * other page to be retired but don't try to retire it, so
1271 * that page_unlock() in the failure path will recurse and try
1272 * to retire THAT page. This is the worst possible situation
1273 * we can get ourselves into.
1274 */
1275 memsegs_lock(0);
1276 pp = first = page_first();
1277 do {
1278 if (pp->p_szc && PP_PAGEROOT(pp) == pp) {
1279 cpp = pp + 1;
1280 lpp = PP_ISFREE(pp)? pp : pp + 2;
1281 cpp2 = pp + 3;
1282 if (!page_trylock(lpp, pp == lpp? SE_EXCL : SE_SHARED))
1283 continue;
1284 if (!page_trylock(cpp, SE_EXCL)) {
1285 page_unlock(lpp);
1286 continue;
1287 }
1288
1289 /* fails */
1290 (void) page_retire(ptob(cpp->p_pagenum), PR_FMA);
1291
1292 page_unlock(lpp);
1293 page_unlock(cpp);
1294 (void) page_retire(ptob(cpp->p_pagenum), PR_FMA);
1295 (void) page_retire(ptob(cpp2->p_pagenum), PR_FMA);
1296 }
1297 } while ((pp = page_next(pp)) != first);
1298 memsegs_unlock(0);
1299
1300 return (0);
1301 }
1302