Whamcloud - gitweb
7a7dc1d08f70a24a87b6e686e4fe60390159dc95
[fs/lustre-release.git] / lustre / osc / osc_page.c
1 /*
2  * GPL HEADER START
3  *
4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
5  *
6  * This program is free software; you can redistribute it and/or modify
7  * it under the terms of the GNU General Public License version 2 only,
8  * as published by the Free Software Foundation.
9  *
10  * This program is distributed in the hope that it will be useful, but
11  * WITHOUT ANY WARRANTY; without even the implied warranty of
12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
13  * General Public License version 2 for more details (a copy is included
14  * in the LICENSE file that accompanied this code).
15  *
16  * You should have received a copy of the GNU General Public License
17  * version 2 along with this program; If not, see
18  * http://www.gnu.org/licenses/gpl-2.0.html
19  *
20  * GPL HEADER END
21  */
22 /*
23  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
24  * Use is subject to license terms.
25  *
26  * Copyright (c) 2011, 2017, Intel Corporation.
27  */
28 /*
29  * This file is part of Lustre, http://www.lustre.org/
30  * Lustre is a trademark of Sun Microsystems, Inc.
31  *
32  * Implementation of cl_page for OSC layer.
33  *
34  *   Author: Nikita Danilov <nikita.danilov@sun.com>
35  *   Author: Jinshan Xiong <jinshan.xiong@intel.com>
36  */
37
38 #define DEBUG_SUBSYSTEM S_OSC
39 #include <lustre_osc.h>
40
41 #include "osc_internal.h"
42
43 static void osc_lru_del(struct client_obd *cli, struct osc_page *opg);
44 static void osc_lru_use(struct client_obd *cli, struct osc_page *opg);
45 static int osc_lru_alloc(const struct lu_env *env, struct client_obd *cli,
46                          struct osc_page *opg);
47
48 /** \addtogroup osc
49  *  @{
50  */
51
52 /*
53  * Page operations.
54  */
55 static void osc_page_transfer_get(struct osc_page *opg, const char *label)
56 {
57         struct cl_page *page = opg->ops_cl.cpl_page;
58
59         LASSERT(!opg->ops_transfer_pinned);
60         cl_page_get(page);
61         lu_ref_add_atomic(&page->cp_reference, label, page);
62         opg->ops_transfer_pinned = 1;
63 }
64
65 static void osc_page_transfer_put(const struct lu_env *env,
66                                   struct osc_page *opg)
67 {
68         struct cl_page *page = opg->ops_cl.cpl_page;
69
70         if (opg->ops_transfer_pinned) {
71                 opg->ops_transfer_pinned = 0;
72                 lu_ref_del(&page->cp_reference, "transfer", page);
73                 cl_page_put(env, page);
74         }
75 }
76
77 /**
78  * This is called once for every page when it is submitted for a transfer
79  * either opportunistic (osc_page_cache_add()), or immediate
80  * (osc_page_submit()).
81  */
82 static void osc_page_transfer_add(const struct lu_env *env,
83                                   struct osc_page *opg, enum cl_req_type crt)
84 {
85         struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
86
87         /* ops_lru and ops_inflight share the same field, so take it from LRU
88          * first and then use it as inflight. */
89         osc_lru_use(osc_cli(obj), opg);
90 }
91
92 int osc_page_cache_add(const struct lu_env *env,
93                         const struct cl_page_slice *slice, struct cl_io *io)
94 {
95         struct osc_page *opg = cl2osc_page(slice);
96         int result;
97         ENTRY;
98
99         osc_page_transfer_get(opg, "transfer\0cache");
100         result = osc_queue_async_io(env, io, opg);
101         if (result != 0)
102                 osc_page_transfer_put(env, opg);
103         else
104                 osc_page_transfer_add(env, opg, CRT_WRITE);
105
106         RETURN(result);
107 }
108
109 void osc_index2policy(union ldlm_policy_data *policy,
110                       const struct cl_object *obj, pgoff_t start, pgoff_t end)
111 {
112         memset(policy, 0, sizeof *policy);
113         policy->l_extent.start = cl_offset(obj, start);
114         policy->l_extent.end   = cl_offset(obj, end + 1) - 1;
115 }
116
117 static const char *osc_list(struct list_head *head)
118 {
119         return list_empty(head) ? "-" : "+";
120 }
121
122 static inline s64 osc_submit_duration(struct osc_page *opg)
123 {
124         if (ktime_to_ns(opg->ops_submit_time) == 0)
125                 return 0;
126
127         return ktime_ms_delta(ktime_get(), opg->ops_submit_time);
128 }
129
130 static int osc_page_print(const struct lu_env *env,
131                           const struct cl_page_slice *slice,
132                           void *cookie, lu_printer_t printer)
133 {
134         struct osc_page       *opg = cl2osc_page(slice);
135         struct osc_async_page *oap = &opg->ops_oap;
136         struct osc_object     *obj = cl2osc(slice->cpl_obj);
137         struct client_obd     *cli = &osc_export(obj)->exp_obd->u.cli;
138
139         return (*printer)(env, cookie, LUSTRE_OSC_NAME"-page@%p %lu: "
140                           "1< %#x %d %u %s %s > "
141                           "2< %lld %u %u %#x %#x | %p %p %p > "
142                           "3< %d %lld %d > "
143                           "4< %d %d %d %lu %s | %s %s %s %s > "
144                           "5< %s %s %s %s | %d %s | %d %s %s>\n",
145                           opg, osc_index(opg),
146                           /* 1 */
147                           oap->oap_magic, oap->oap_cmd,
148                           oap->oap_interrupted,
149                           osc_list(&oap->oap_pending_item),
150                           osc_list(&oap->oap_rpc_item),
151                           /* 2 */
152                           oap->oap_obj_off, oap->oap_page_off, oap->oap_count,
153                           oap->oap_async_flags, oap->oap_brw_flags,
154                           oap->oap_request, oap->oap_cli, obj,
155                           /* 3 */
156                           opg->ops_transfer_pinned,
157                           osc_submit_duration(opg), opg->ops_srvlock,
158                           /* 4 */
159                           cli->cl_r_in_flight, cli->cl_w_in_flight,
160                           cli->cl_max_rpcs_in_flight,
161                           cli->cl_avail_grant,
162                           osc_list(&cli->cl_cache_waiters),
163                           osc_list(&cli->cl_loi_ready_list),
164                           osc_list(&cli->cl_loi_hp_ready_list),
165                           osc_list(&cli->cl_loi_write_list),
166                           osc_list(&cli->cl_loi_read_list),
167                           /* 5 */
168                           osc_list(&obj->oo_ready_item),
169                           osc_list(&obj->oo_hp_ready_item),
170                           osc_list(&obj->oo_write_item),
171                           osc_list(&obj->oo_read_item),
172                           atomic_read(&obj->oo_nr_reads),
173                           osc_list(&obj->oo_reading_exts),
174                           atomic_read(&obj->oo_nr_writes),
175                           osc_list(&obj->oo_hp_exts),
176                           osc_list(&obj->oo_urgent_exts));
177 }
178
179 static void osc_page_delete(const struct lu_env *env,
180                             const struct cl_page_slice *slice)
181 {
182         struct osc_page   *opg = cl2osc_page(slice);
183         struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
184         int rc;
185
186         ENTRY;
187         CDEBUG(D_TRACE, "%p\n", opg);
188         osc_page_transfer_put(env, opg);
189         rc = osc_teardown_async_page(env, obj, opg);
190         if (rc) {
191                 CL_PAGE_DEBUG(D_ERROR, env, slice->cpl_page,
192                               "Trying to teardown failed: %d\n", rc);
193                 LASSERT(0);
194         }
195
196         osc_lru_del(osc_cli(obj), opg);
197
198         if (slice->cpl_page->cp_type == CPT_CACHEABLE) {
199                 void *value = NULL;
200
201                 spin_lock(&obj->oo_tree_lock);
202                 if (opg->ops_intree) {
203                         value = radix_tree_delete(&obj->oo_tree,
204                                                   osc_index(opg));
205                         if (value != NULL) {
206                                 --obj->oo_npages;
207                                 opg->ops_intree = 0;
208                         }
209                 }
210                 spin_unlock(&obj->oo_tree_lock);
211
212                 LASSERT(ergo(value != NULL, value == opg));
213         }
214
215         EXIT;
216 }
217
218 static void osc_page_clip(const struct lu_env *env,
219                           const struct cl_page_slice *slice,
220                           int from, int to)
221 {
222         struct osc_page       *opg = cl2osc_page(slice);
223         struct osc_async_page *oap = &opg->ops_oap;
224
225         opg->ops_from = from;
226         opg->ops_to   = to;
227         spin_lock(&oap->oap_lock);
228         oap->oap_async_flags |= ASYNC_COUNT_STABLE;
229         spin_unlock(&oap->oap_lock);
230 }
231
232 static int osc_page_cancel(const struct lu_env *env,
233                            const struct cl_page_slice *slice)
234 {
235         struct osc_page *opg = cl2osc_page(slice);
236         int rc = 0;
237
238         /* Check if the transferring against this page
239          * is completed, or not even queued. */
240         if (opg->ops_transfer_pinned)
241                 /* FIXME: may not be interrupted.. */
242                 rc = osc_cancel_async_page(env, opg);
243         LASSERT(ergo(rc == 0, opg->ops_transfer_pinned == 0));
244         return rc;
245 }
246
247 static int osc_page_flush(const struct lu_env *env,
248                           const struct cl_page_slice *slice,
249                           struct cl_io *io)
250 {
251         struct osc_page *opg = cl2osc_page(slice);
252         int rc = 0;
253         ENTRY;
254         rc = osc_flush_async_page(env, io, opg);
255         RETURN(rc);
256 }
257
258 static void osc_page_touch(const struct lu_env *env,
259                           const struct cl_page_slice *slice, size_t to)
260 {
261         struct osc_page *opg = cl2osc_page(slice);
262         struct cl_object *obj = opg->ops_cl.cpl_obj;
263
264         osc_page_touch_at(env, obj, osc_index(opg), to);
265 }
266
267 static const struct cl_page_operations osc_page_ops = {
268         .cpo_print         = osc_page_print,
269         .cpo_delete        = osc_page_delete,
270         .cpo_clip           = osc_page_clip,
271         .cpo_cancel         = osc_page_cancel,
272         .cpo_flush          = osc_page_flush,
273         .cpo_page_touch    = osc_page_touch,
274 };
275
276 int osc_page_init(const struct lu_env *env, struct cl_object *obj,
277                   struct cl_page *page, pgoff_t index)
278 {
279         struct osc_object *osc = cl2osc(obj);
280         struct osc_page   *opg = cl_object_page_slice(obj, page);
281         struct osc_io *oio = osc_env_io(env);
282         int result;
283
284         opg->ops_from = 0;
285         opg->ops_to   = PAGE_SIZE;
286
287         INIT_LIST_HEAD(&opg->ops_lru);
288
289         result = osc_prep_async_page(osc, opg, page->cp_vmpage,
290                                      cl_offset(obj, index));
291         if (result != 0)
292                 return result;
293
294         opg->ops_srvlock = osc_io_srvlock(oio);
295         cl_page_slice_add(page, &opg->ops_cl, obj, index,
296                           &osc_page_ops);
297
298
299         /* reserve an LRU space for this page */
300         if (page->cp_type == CPT_CACHEABLE) {
301                 result = osc_lru_alloc(env, osc_cli(osc), opg);
302                 if (result == 0) {
303                         result = radix_tree_preload(GFP_NOFS);
304                         if (result == 0) {
305                                 spin_lock(&osc->oo_tree_lock);
306                                 result = radix_tree_insert(&osc->oo_tree,
307                                                            index, opg);
308                                 if (result == 0) {
309                                         ++osc->oo_npages;
310                                         opg->ops_intree = 1;
311                                 }
312                                 spin_unlock(&osc->oo_tree_lock);
313
314                                 radix_tree_preload_end();
315                         }
316                 }
317         }
318
319         return result;
320 }
321 EXPORT_SYMBOL(osc_page_init);
322
323 /**
324  * Helper function called by osc_io_submit() for every page in an immediate
325  * transfer (i.e., transferred synchronously).
326  */
327 void osc_page_submit(const struct lu_env *env, struct osc_page *opg,
328                      enum cl_req_type crt, int brw_flags)
329 {
330         struct osc_io *oio = osc_env_io(env);
331         struct osc_async_page *oap = &opg->ops_oap;
332
333         LASSERTF(oap->oap_magic == OAP_MAGIC, "Bad oap magic: oap %p, "
334                  "magic 0x%x\n", oap, oap->oap_magic);
335         LASSERT(oap->oap_async_flags & ASYNC_READY);
336         LASSERT(oap->oap_async_flags & ASYNC_COUNT_STABLE);
337
338         oap->oap_cmd       = crt == CRT_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ;
339         oap->oap_page_off  = opg->ops_from;
340         oap->oap_count     = opg->ops_to - opg->ops_from;
341         oap->oap_brw_flags = OBD_BRW_SYNC | brw_flags;
342
343         if (oio->oi_cap_sys_resource) {
344                 oap->oap_brw_flags |= OBD_BRW_NOQUOTA;
345                 oap->oap_cmd |= OBD_BRW_NOQUOTA;
346         }
347
348         opg->ops_submit_time = ktime_get();
349         osc_page_transfer_get(opg, "transfer\0imm");
350         osc_page_transfer_add(env, opg, crt);
351 }
352
353 /* --------------- LRU page management ------------------ */
354
355 /* OSC is a natural place to manage LRU pages as applications are specialized
356  * to write OSC by OSC. Ideally, if one OSC is used more frequently it should
357  * occupy more LRU slots. On the other hand, we should avoid using up all LRU
358  * slots (client_obd::cl_lru_left) otherwise process has to be put into sleep
359  * for free LRU slots - this will be very bad so the algorithm requires each
360  * OSC to free slots voluntarily to maintain a reasonable number of free slots
361  * at any time.
362  */
363
364 static DECLARE_WAIT_QUEUE_HEAD(osc_lru_waitq);
365
366 /**
367  * LRU pages are freed in batch mode. OSC should at least free this
368  * number of pages to avoid running out of LRU slots.
369  */
370 static inline int lru_shrink_min(struct client_obd *cli)
371 {
372         return cli->cl_max_pages_per_rpc * 2;
373 }
374
375 /**
376  * free this number at most otherwise it will take too long time to finsih.
377  */
378 static inline int lru_shrink_max(struct client_obd *cli)
379 {
380         return cli->cl_max_pages_per_rpc * cli->cl_max_rpcs_in_flight;
381 }
382
383 /**
384  * Check if we can free LRU slots from this OSC. If there exists LRU waiters,
385  * we should free slots aggressively. In this way, slots are freed in a steady
386  * step to maintain fairness among OSCs.
387  *
388  * Return how many LRU pages should be freed.
389  */
390 static int osc_cache_too_much(struct client_obd *cli)
391 {
392         struct cl_client_cache *cache = cli->cl_cache;
393         long pages = atomic_long_read(&cli->cl_lru_in_list);
394         unsigned long budget;
395
396         LASSERT(cache != NULL);
397         budget = cache->ccc_lru_max / (atomic_read(&cache->ccc_users) - 2);
398
399         /* if it's going to run out LRU slots, we should free some, but not
400          * too much to maintain faireness among OSCs. */
401         if (atomic_long_read(cli->cl_lru_left) < cache->ccc_lru_max >> 2) {
402                 if (pages >= budget)
403                         return lru_shrink_max(cli);
404                 else if (pages >= budget / 2)
405                         return lru_shrink_min(cli);
406         } else {
407                 time64_t duration = ktime_get_real_seconds();
408                 long timediff;
409
410                 /* knock out pages by duration of no IO activity */
411                 duration -= cli->cl_lru_last_used;
412                 /*
413                  * The difference shouldn't be more than 70 years
414                  * so we can safely case to a long. Round to
415                  * approximately 1 minute.
416                  */
417                 timediff = (long)(duration >> 6);
418                 if (timediff > 0 && pages >= budget / timediff)
419                         return lru_shrink_min(cli);
420         }
421         return 0;
422 }
423
424 int lru_queue_work(const struct lu_env *env, void *data)
425 {
426         struct client_obd *cli = data;
427         int count;
428
429         CDEBUG(D_CACHE, "%s: run LRU work for client obd\n", cli_name(cli));
430         count = osc_cache_too_much(cli);
431         if (count > 0) {
432                 int rc = osc_lru_shrink(env, cli, count, false);
433
434                 CDEBUG(D_CACHE, "%s: shrank %d/%d pages from client obd\n",
435                        cli_name(cli), rc, count);
436                 if (rc >= count) {
437                         CDEBUG(D_CACHE, "%s: queue again\n", cli_name(cli));
438                         ptlrpcd_queue_work(cli->cl_lru_work);
439                 }
440         }
441
442         RETURN(0);
443 }
444
445 void osc_lru_add_batch(struct client_obd *cli, struct list_head *plist)
446 {
447         struct list_head lru = LIST_HEAD_INIT(lru);
448         struct osc_async_page *oap;
449         long npages = 0;
450
451         list_for_each_entry(oap, plist, oap_pending_item) {
452                 struct osc_page *opg = oap2osc_page(oap);
453
454                 if (!opg->ops_in_lru)
455                         continue;
456
457                 ++npages;
458                 LASSERT(list_empty(&opg->ops_lru));
459                 list_add(&opg->ops_lru, &lru);
460         }
461
462         if (npages > 0) {
463                 spin_lock(&cli->cl_lru_list_lock);
464                 list_splice_tail(&lru, &cli->cl_lru_list);
465                 atomic_long_sub(npages, &cli->cl_lru_busy);
466                 atomic_long_add(npages, &cli->cl_lru_in_list);
467                 cli->cl_lru_last_used = ktime_get_real_seconds();
468                 spin_unlock(&cli->cl_lru_list_lock);
469
470                 if (waitqueue_active(&osc_lru_waitq))
471                         (void)ptlrpcd_queue_work(cli->cl_lru_work);
472         }
473 }
474
475 static void __osc_lru_del(struct client_obd *cli, struct osc_page *opg)
476 {
477         LASSERT(atomic_long_read(&cli->cl_lru_in_list) > 0);
478         list_del_init(&opg->ops_lru);
479         atomic_long_dec(&cli->cl_lru_in_list);
480 }
481
482 /**
483  * Page is being destroyed. The page may be not in LRU list, if the transfer
484  * has never finished(error occurred).
485  */
486 static void osc_lru_del(struct client_obd *cli, struct osc_page *opg)
487 {
488         if (opg->ops_in_lru) {
489                 spin_lock(&cli->cl_lru_list_lock);
490                 if (!list_empty(&opg->ops_lru)) {
491                         __osc_lru_del(cli, opg);
492                 } else {
493                         LASSERT(atomic_long_read(&cli->cl_lru_busy) > 0);
494                         atomic_long_dec(&cli->cl_lru_busy);
495                 }
496                 spin_unlock(&cli->cl_lru_list_lock);
497
498                 atomic_long_inc(cli->cl_lru_left);
499                 /* this is a great place to release more LRU pages if
500                  * this osc occupies too many LRU pages and kernel is
501                  * stealing one of them. */
502                 if (osc_cache_too_much(cli)) {
503                         CDEBUG(D_CACHE, "%s: queue LRU work\n", cli_name(cli));
504                         (void)ptlrpcd_queue_work(cli->cl_lru_work);
505                 }
506                 wake_up(&osc_lru_waitq);
507         } else {
508                 LASSERT(list_empty(&opg->ops_lru));
509         }
510 }
511
512 /**
513  * Delete page from LRU list for redirty.
514  */
515 static void osc_lru_use(struct client_obd *cli, struct osc_page *opg)
516 {
517         /* If page is being transferred for the first time,
518          * ops_lru should be empty */
519         if (opg->ops_in_lru) {
520                 if (list_empty(&opg->ops_lru))
521                         return;
522                 spin_lock(&cli->cl_lru_list_lock);
523                 if (!list_empty(&opg->ops_lru)) {
524                         __osc_lru_del(cli, opg);
525                         atomic_long_inc(&cli->cl_lru_busy);
526                 }
527                 spin_unlock(&cli->cl_lru_list_lock);
528         }
529 }
530
531 static void discard_pagevec(const struct lu_env *env, struct cl_io *io,
532                                 struct cl_page **pvec, int max_index)
533 {
534         struct pagevec *pagevec = &osc_env_info(env)->oti_pagevec;
535         int i;
536
537         ll_pagevec_init(pagevec, 0);
538         for (i = 0; i < max_index; i++) {
539                 struct cl_page *page = pvec[i];
540
541                 LASSERT(cl_page_is_owned(page, io));
542                 cl_page_delete(env, page);
543                 cl_page_discard(env, io, page);
544                 cl_page_disown(env, io, page);
545                 cl_pagevec_put(env, page, pagevec);
546
547                 pvec[i] = NULL;
548         }
549         pagevec_release(pagevec);
550 }
551
552 /**
553  * Check if a cl_page can be released, i.e, it's not being used.
554  *
555  * If unstable account is turned on, bulk transfer may hold one refcount
556  * for recovery so we need to check vmpage refcount as well; otherwise,
557  * even we can destroy cl_page but the corresponding vmpage can't be reused.
558  */
559 static inline bool lru_page_busy(struct client_obd *cli, struct cl_page *page)
560 {
561         if (cl_page_in_use_noref(page))
562                 return true;
563
564         if (cli->cl_cache->ccc_unstable_check) {
565                 struct page *vmpage = cl_page_vmpage(page);
566
567                 /* vmpage have two known users: cl_page and VM page cache */
568                 if (page_count(vmpage) - page_mapcount(vmpage) > 2)
569                         return true;
570         }
571         return false;
572 }
573
574 /**
575  * Drop @target of pages from LRU at most.
576  */
577 long osc_lru_shrink(const struct lu_env *env, struct client_obd *cli,
578                    long target, bool force)
579 {
580         struct cl_io *io;
581         struct cl_object *clobj = NULL;
582         struct cl_page **pvec;
583         struct osc_page *opg;
584         long count = 0;
585         int maxscan = 0;
586         int index = 0;
587         int rc = 0;
588         ENTRY;
589
590         LASSERT(atomic_long_read(&cli->cl_lru_in_list) >= 0);
591         if (atomic_long_read(&cli->cl_lru_in_list) == 0 || target <= 0)
592                 RETURN(0);
593
594         CDEBUG(D_CACHE, "%s: shrinkers: %d, force: %d\n",
595                cli_name(cli), atomic_read(&cli->cl_lru_shrinkers), force);
596         if (!force) {
597                 if (atomic_read(&cli->cl_lru_shrinkers) > 0)
598                         RETURN(-EBUSY);
599
600                 if (atomic_inc_return(&cli->cl_lru_shrinkers) > 1) {
601                         atomic_dec(&cli->cl_lru_shrinkers);
602                         RETURN(-EBUSY);
603                 }
604         } else {
605                 atomic_inc(&cli->cl_lru_shrinkers);
606         }
607
608         pvec = (struct cl_page **)osc_env_info(env)->oti_pvec;
609         io = osc_env_thread_io(env);
610
611         spin_lock(&cli->cl_lru_list_lock);
612         if (force)
613                 cli->cl_lru_reclaim++;
614         maxscan = min(target << 1, atomic_long_read(&cli->cl_lru_in_list));
615         while (!list_empty(&cli->cl_lru_list)) {
616                 struct cl_page *page;
617                 bool will_free = false;
618
619                 if (!force && atomic_read(&cli->cl_lru_shrinkers) > 1)
620                         break;
621
622                 if (--maxscan < 0)
623                         break;
624
625                 opg = list_entry(cli->cl_lru_list.next, struct osc_page,
626                                  ops_lru);
627                 page = opg->ops_cl.cpl_page;
628                 if (lru_page_busy(cli, page)) {
629                         list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
630                         continue;
631                 }
632
633                 LASSERT(page->cp_obj != NULL);
634                 if (clobj != page->cp_obj) {
635                         struct cl_object *tmp = page->cp_obj;
636
637                         cl_object_get(tmp);
638                         spin_unlock(&cli->cl_lru_list_lock);
639
640                         if (clobj != NULL) {
641                                 discard_pagevec(env, io, pvec, index);
642                                 index = 0;
643
644                                 cl_io_fini(env, io);
645                                 cl_object_put(env, clobj);
646                                 clobj = NULL;
647                         }
648
649                         clobj = tmp;
650                         io->ci_obj = clobj;
651                         io->ci_ignore_layout = 1;
652                         rc = cl_io_init(env, io, CIT_MISC, clobj);
653
654                         spin_lock(&cli->cl_lru_list_lock);
655
656                         if (rc != 0)
657                                 break;
658
659                         ++maxscan;
660                         continue;
661                 }
662
663                 if (cl_page_own_try(env, io, page) == 0) {
664                         if (!lru_page_busy(cli, page)) {
665                                 /* remove it from lru list earlier to avoid
666                                  * lock contention */
667                                 __osc_lru_del(cli, opg);
668                                 opg->ops_in_lru = 0; /* will be discarded */
669
670                                 cl_page_get(page);
671                                 will_free = true;
672                         } else {
673                                 cl_page_disown(env, io, page);
674                         }
675                 }
676
677                 if (!will_free) {
678                         list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
679                         continue;
680                 }
681
682                 /* Don't discard and free the page with cl_lru_list held */
683                 pvec[index++] = page;
684                 if (unlikely(index == OTI_PVEC_SIZE)) {
685                         spin_unlock(&cli->cl_lru_list_lock);
686                         discard_pagevec(env, io, pvec, index);
687                         index = 0;
688
689                         spin_lock(&cli->cl_lru_list_lock);
690                 }
691
692                 if (++count >= target)
693                         break;
694         }
695         spin_unlock(&cli->cl_lru_list_lock);
696
697         if (clobj != NULL) {
698                 discard_pagevec(env, io, pvec, index);
699
700                 cl_io_fini(env, io);
701                 cl_object_put(env, clobj);
702         }
703
704         atomic_dec(&cli->cl_lru_shrinkers);
705         if (count > 0) {
706                 atomic_long_add(count, cli->cl_lru_left);
707                 wake_up_all(&osc_lru_waitq);
708         }
709         RETURN(count > 0 ? count : rc);
710 }
711 EXPORT_SYMBOL(osc_lru_shrink);
712
713 /**
714  * Reclaim LRU pages by an IO thread. The caller wants to reclaim at least
715  * \@npages of LRU slots. For performance consideration, it's better to drop
716  * LRU pages in batch. Therefore, the actual number is adjusted at least
717  * max_pages_per_rpc.
718  */
719 static long osc_lru_reclaim(struct client_obd *cli, unsigned long npages)
720 {
721         struct lu_env *env;
722         struct cl_client_cache *cache = cli->cl_cache;
723         int max_scans;
724         __u16 refcheck;
725         long rc = 0;
726         ENTRY;
727
728         LASSERT(cache != NULL);
729
730         env = cl_env_get(&refcheck);
731         if (IS_ERR(env))
732                 RETURN(rc);
733
734         npages = max_t(int, npages, cli->cl_max_pages_per_rpc);
735         CDEBUG(D_CACHE, "%s: start to reclaim %ld pages from LRU\n",
736                cli_name(cli), npages);
737         rc = osc_lru_shrink(env, cli, npages, true);
738         if (rc >= npages) {
739                 CDEBUG(D_CACHE, "%s: reclaimed %ld/%ld pages from LRU\n",
740                        cli_name(cli), rc, npages);
741                 if (osc_cache_too_much(cli) > 0)
742                         ptlrpcd_queue_work(cli->cl_lru_work);
743                 GOTO(out, rc);
744         } else if (rc > 0) {
745                 npages -= rc;
746         }
747
748         CDEBUG(D_CACHE, "%s: cli %p no free slots, pages: %ld/%ld, want: %ld\n",
749                 cli_name(cli), cli, atomic_long_read(&cli->cl_lru_in_list),
750                 atomic_long_read(&cli->cl_lru_busy), npages);
751
752         /* Reclaim LRU slots from other client_obd as it can't free enough
753          * from its own. This should rarely happen. */
754         spin_lock(&cache->ccc_lru_lock);
755         LASSERT(!list_empty(&cache->ccc_lru));
756
757         cache->ccc_lru_shrinkers++;
758         list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
759
760         max_scans = atomic_read(&cache->ccc_users) - 2;
761         while (--max_scans > 0 && !list_empty(&cache->ccc_lru)) {
762                 cli = list_entry(cache->ccc_lru.next, struct client_obd,
763                                  cl_lru_osc);
764
765                 CDEBUG(D_CACHE, "%s: cli %p LRU pages: %ld, busy: %ld.\n",
766                         cli_name(cli), cli,
767                         atomic_long_read(&cli->cl_lru_in_list),
768                         atomic_long_read(&cli->cl_lru_busy));
769
770                 list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
771                 if (osc_cache_too_much(cli) > 0) {
772                         spin_unlock(&cache->ccc_lru_lock);
773
774                         rc = osc_lru_shrink(env, cli, npages, true);
775                         spin_lock(&cache->ccc_lru_lock);
776                         if (rc >= npages)
777                                 break;
778                         if (rc > 0)
779                                 npages -= rc;
780                 }
781         }
782         spin_unlock(&cache->ccc_lru_lock);
783
784 out:
785         cl_env_put(env, &refcheck);
786         CDEBUG(D_CACHE, "%s: cli %p freed %ld pages.\n",
787                 cli_name(cli), cli, rc);
788         return rc;
789 }
790
791 /**
792  * osc_lru_alloc() is called to allocate an LRU slot for a cl_page.
793  *
794  * Usually the LRU slots are reserved in osc_io_iter_rw_init().
795  * Only in the case that the LRU slots are in extreme shortage, it should
796  * have reserved enough slots for an IO.
797  */
798 static int osc_lru_alloc(const struct lu_env *env, struct client_obd *cli,
799                          struct osc_page *opg)
800 {
801         struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
802         struct osc_io *oio = osc_env_io(env);
803         int rc = 0;
804
805         ENTRY;
806
807         if (cli->cl_cache == NULL) /* shall not be in LRU */
808                 RETURN(0);
809
810         if (oio->oi_lru_reserved > 0) {
811                 --oio->oi_lru_reserved;
812                 goto out;
813         }
814
815         LASSERT(atomic_long_read(cli->cl_lru_left) >= 0);
816         while (!atomic_long_add_unless(cli->cl_lru_left, -1, 0)) {
817                 /* run out of LRU spaces, try to drop some by itself */
818                 rc = osc_lru_reclaim(cli, 1);
819                 if (rc < 0)
820                         break;
821                 if (rc > 0)
822                         continue;
823
824                 cond_resched();
825                 rc = l_wait_event(osc_lru_waitq,
826                                 atomic_long_read(cli->cl_lru_left) > 0,
827                                 &lwi);
828                 if (rc < 0)
829                         break;
830         }
831
832 out:
833         if (rc >= 0) {
834                 atomic_long_inc(&cli->cl_lru_busy);
835                 opg->ops_in_lru = 1;
836                 rc = 0;
837         }
838
839         RETURN(rc);
840 }
841
842 /**
843  * osc_lru_reserve() is called to reserve enough LRU slots for I/O.
844  *
845  * The benefit of doing this is to reduce contention against atomic counter
846  * cl_lru_left by changing it from per-page access to per-IO access.
847  */
848 unsigned long osc_lru_reserve(struct client_obd *cli, unsigned long npages)
849 {
850         unsigned long reserved = 0;
851         unsigned long max_pages;
852         unsigned long c;
853
854         /* reserve a full RPC window at most to avoid that a thread accidentally
855          * consumes too many LRU slots */
856         max_pages = cli->cl_max_pages_per_rpc * cli->cl_max_rpcs_in_flight;
857         if (npages > max_pages)
858                 npages = max_pages;
859
860         c = atomic_long_read(cli->cl_lru_left);
861         if (c < npages && osc_lru_reclaim(cli, npages) > 0)
862                 c = atomic_long_read(cli->cl_lru_left);
863         while (c >= npages) {
864                 if (c == atomic_long_cmpxchg(cli->cl_lru_left, c, c - npages)) {
865                         reserved = npages;
866                         break;
867                 }
868                 c = atomic_long_read(cli->cl_lru_left);
869         }
870         if (atomic_long_read(cli->cl_lru_left) < max_pages) {
871                 /* If there aren't enough pages in the per-OSC LRU then
872                  * wake up the LRU thread to try and clear out space, so
873                  * we don't block if pages are being dirtied quickly. */
874                 CDEBUG(D_CACHE, "%s: queue LRU, left: %lu/%ld.\n",
875                        cli_name(cli), atomic_long_read(cli->cl_lru_left),
876                        max_pages);
877                 (void)ptlrpcd_queue_work(cli->cl_lru_work);
878         }
879
880         return reserved;
881 }
882
883 /**
884  * osc_lru_unreserve() is called to unreserve LRU slots.
885  *
886  * LRU slots reserved by osc_lru_reserve() may have entries left due to several
887  * reasons such as page already existing or I/O error. Those reserved slots
888  * should be freed by calling this function.
889  */
890 void osc_lru_unreserve(struct client_obd *cli, unsigned long npages)
891 {
892         atomic_long_add(npages, cli->cl_lru_left);
893         wake_up_all(&osc_lru_waitq);
894 }
895
896 /**
897  * Atomic operations are expensive. We accumulate the accounting for the
898  * same page zone to get better performance.
899  * In practice this can work pretty good because the pages in the same RPC
900  * are likely from the same page zone.
901  */
902 static inline void unstable_page_accounting(struct ptlrpc_bulk_desc *desc,
903                                             struct osc_brw_async_args *aa,
904                                             int factor)
905 {
906         int page_count;
907         void *zone = NULL;
908         int count = 0;
909         int i;
910
911         if (desc != NULL) {
912                 LASSERT(ptlrpc_is_bulk_desc_kiov(desc->bd_type));
913                 page_count = desc->bd_iov_count;
914         } else {
915                 page_count = aa->aa_page_count;
916         }
917
918         for (i = 0; i < page_count; i++) {
919                 void *pz;
920                 if (desc)
921                         pz = page_zone(BD_GET_KIOV(desc, i).kiov_page);
922                 else
923                         pz = page_zone(aa->aa_ppga[i]->pg);
924
925                 if (likely(pz == zone)) {
926                         ++count;
927                         continue;
928                 }
929
930                 if (count > 0) {
931                         mod_zone_page_state(zone, NR_UNSTABLE_NFS,
932                                             factor * count);
933                         count = 0;
934                 }
935                 zone = pz;
936                 ++count;
937         }
938         if (count > 0)
939                 mod_zone_page_state(zone, NR_UNSTABLE_NFS, factor * count);
940 }
941
942 static inline void add_unstable_page_accounting(struct ptlrpc_bulk_desc *desc,
943                                                 struct osc_brw_async_args *aa)
944 {
945         unstable_page_accounting(desc, aa, 1);
946 }
947
948 static inline void dec_unstable_page_accounting(struct ptlrpc_bulk_desc *desc,
949                                                 struct osc_brw_async_args *aa)
950 {
951         unstable_page_accounting(desc, aa, -1);
952 }
953
954 /**
955  * Performs "unstable" page accounting. This function balances the
956  * increment operations performed in osc_inc_unstable_pages. It is
957  * registered as the RPC request callback, and is executed when the
958  * bulk RPC is committed on the server. Thus at this point, the pages
959  * involved in the bulk transfer are no longer considered unstable.
960  *
961  * If this function is called, the request should have been committed
962  * or req:rq_unstable must have been set; it implies that the unstable
963  * statistic have been added.
964  */
965 void osc_dec_unstable_pages(struct ptlrpc_request *req)
966 {
967         struct ptlrpc_bulk_desc *desc       = req->rq_bulk;
968         struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
969         struct client_obd       *cli        = &req->rq_import->imp_obd->u.cli;
970         int                      page_count;
971         long                     unstable_count;
972
973         if (desc)
974                 page_count = desc->bd_iov_count;
975         else
976                 page_count = aa->aa_page_count;
977
978         LASSERT(page_count >= 0);
979
980         dec_unstable_page_accounting(desc, aa);
981
982         unstable_count = atomic_long_sub_return(page_count,
983                                                 &cli->cl_unstable_count);
984         LASSERT(unstable_count >= 0);
985
986         unstable_count = atomic_long_sub_return(page_count,
987                                            &cli->cl_cache->ccc_unstable_nr);
988         LASSERT(unstable_count >= 0);
989         if (unstable_count == 0)
990                 wake_up_all(&cli->cl_cache->ccc_unstable_waitq);
991
992         if (waitqueue_active(&osc_lru_waitq))
993                 (void)ptlrpcd_queue_work(cli->cl_lru_work);
994 }
995
996 /**
997  * "unstable" page accounting. See: osc_dec_unstable_pages.
998  */
999 void osc_inc_unstable_pages(struct ptlrpc_request *req)
1000 {
1001         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
1002         struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
1003         struct client_obd       *cli  = &req->rq_import->imp_obd->u.cli;
1004         long                     page_count;
1005
1006         /* No unstable page tracking */
1007         if (cli->cl_cache == NULL || !cli->cl_cache->ccc_unstable_check)
1008                 return;
1009
1010         if (desc)
1011                 page_count = desc->bd_iov_count;
1012         else
1013                 page_count = aa->aa_page_count;
1014
1015         add_unstable_page_accounting(desc, aa);
1016         atomic_long_add(page_count, &cli->cl_unstable_count);
1017         atomic_long_add(page_count, &cli->cl_cache->ccc_unstable_nr);
1018
1019         /* If the request has already been committed (i.e. brw_commit
1020          * called via rq_commit_cb), we need to undo the unstable page
1021          * increments we just performed because rq_commit_cb wont be
1022          * called again. */
1023         spin_lock(&req->rq_lock);
1024         if (unlikely(req->rq_committed)) {
1025                 spin_unlock(&req->rq_lock);
1026
1027                 osc_dec_unstable_pages(req);
1028         } else {
1029                 req->rq_unstable = 1;
1030                 spin_unlock(&req->rq_lock);
1031         }
1032 }
1033
1034 /**
1035  * Check if it piggybacks SOFT_SYNC flag to OST from this OSC.
1036  * This function will be called by every BRW RPC so it's critical
1037  * to make this function fast.
1038  */
1039 bool osc_over_unstable_soft_limit(struct client_obd *cli)
1040 {
1041         long unstable_nr, osc_unstable_count;
1042
1043         /* Can't check cli->cl_unstable_count, therefore, no soft limit */
1044         if (cli->cl_cache == NULL || !cli->cl_cache->ccc_unstable_check)
1045                 return false;
1046
1047         osc_unstable_count = atomic_long_read(&cli->cl_unstable_count);
1048         unstable_nr = atomic_long_read(&cli->cl_cache->ccc_unstable_nr);
1049
1050         CDEBUG(D_CACHE,
1051                "%s: cli: %p unstable pages: %lu, osc unstable pages: %lu\n",
1052                cli_name(cli), cli, unstable_nr, osc_unstable_count);
1053
1054         /* If the LRU slots are in shortage - 25% remaining AND this OSC
1055          * has one full RPC window of unstable pages, it's a good chance
1056          * to piggyback a SOFT_SYNC flag.
1057          * Please notice that the OST won't take immediate response for the
1058          * SOFT_SYNC request so active OSCs will have more chance to carry
1059          * the flag, this is reasonable. */
1060         return unstable_nr > cli->cl_cache->ccc_lru_max >> 2 &&
1061                osc_unstable_count > cli->cl_max_pages_per_rpc *
1062                                     cli->cl_max_rpcs_in_flight;
1063 }
1064
1065 /**
1066  * Return how many LRU pages in the cache of all OSC devices
1067  *
1068  * \retval      return # of cached LRU pages times reclaimation tendency
1069  * \retval      SHRINK_STOP if it cannot do any scanning in this time
1070  */
1071 unsigned long osc_cache_shrink_count(struct shrinker *sk,
1072                                      struct shrink_control *sc)
1073 {
1074         struct client_obd *cli;
1075         unsigned long cached = 0;
1076
1077         spin_lock(&osc_shrink_lock);
1078         list_for_each_entry(cli, &osc_shrink_list, cl_shrink_list)
1079                 cached += atomic_long_read(&cli->cl_lru_in_list);
1080         spin_unlock(&osc_shrink_lock);
1081
1082         return (cached  * sysctl_vfs_cache_pressure) / 100;
1083 }
1084
1085 /**
1086  * Scan and try to reclaim sc->nr_to_scan cached LRU pages
1087  *
1088  * \retval      number of cached LRU pages reclaimed
1089  * \retval      SHRINK_STOP if it cannot do any scanning in this time
1090  *
1091  * Linux kernel will loop calling this shrinker scan routine with
1092  * sc->nr_to_scan = SHRINK_BATCH(128 for now) until kernel got enough memory.
1093  *
1094  * If sc->nr_to_scan is 0, the VM is querying the cache size, we don't need
1095  * to scan and try to reclaim LRU pages, just return 0 and
1096  * osc_cache_shrink_count() will report the LRU page number.
1097  */
1098 unsigned long osc_cache_shrink_scan(struct shrinker *sk,
1099                                     struct shrink_control *sc)
1100 {
1101         struct client_obd *cli;
1102         struct client_obd *stop_anchor = NULL;
1103         struct lu_env *env;
1104         long shrank = 0;
1105         int rc;
1106         __u16 refcheck;
1107
1108         if (sc->nr_to_scan == 0)
1109                 return 0;
1110
1111         if (!(sc->gfp_mask & __GFP_FS))
1112                 return SHRINK_STOP;
1113
1114         env = cl_env_get(&refcheck);
1115         if (IS_ERR(env))
1116                 return SHRINK_STOP;
1117
1118         spin_lock(&osc_shrink_lock);
1119         while (!list_empty(&osc_shrink_list)) {
1120                 cli = list_entry(osc_shrink_list.next, struct client_obd,
1121                                  cl_shrink_list);
1122
1123                 if (stop_anchor == NULL)
1124                         stop_anchor = cli;
1125                 else if (cli == stop_anchor)
1126                         break;
1127
1128                 list_move_tail(&cli->cl_shrink_list, &osc_shrink_list);
1129                 spin_unlock(&osc_shrink_lock);
1130
1131                 /* shrink no more than max_pages_per_rpc for an OSC */
1132                 rc = osc_lru_shrink(env, cli, (sc->nr_to_scan - shrank) >
1133                                     cli->cl_max_pages_per_rpc ?
1134                                     cli->cl_max_pages_per_rpc :
1135                                     sc->nr_to_scan - shrank, true);
1136                 if (rc > 0)
1137                         shrank += rc;
1138
1139                 if (shrank >= sc->nr_to_scan)
1140                         goto out;
1141
1142                 spin_lock(&osc_shrink_lock);
1143         }
1144         spin_unlock(&osc_shrink_lock);
1145
1146 out:
1147         cl_env_put(env, &refcheck);
1148
1149         return shrank;
1150 }
1151
1152 /** @} osc */