Whamcloud - gitweb
LU-6698 kernel: kernel update RHEL 6.6 [2.6.32-504.23.4.el6]
[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.sun.com/software/products/lustre/docs/GPLv2.pdf
19  *
20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21  * CA 95054 USA or visit www.sun.com if you need additional information or
22  * have any questions.
23  *
24  * GPL HEADER END
25  */
26 /*
27  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
28  * Use is subject to license terms.
29  *
30  * Copyright (c) 2011, 2014, Intel Corporation.
31  */
32 /*
33  * This file is part of Lustre, http://www.lustre.org/
34  * Lustre is a trademark of Sun Microsystems, Inc.
35  *
36  * Implementation of cl_page for OSC layer.
37  *
38  *   Author: Nikita Danilov <nikita.danilov@sun.com>
39  *   Author: Jinshan Xiong <jinshan.xiong@intel.com>
40  */
41
42 #define DEBUG_SUBSYSTEM S_OSC
43
44 #include "osc_cl_internal.h"
45
46 static void osc_lru_del(struct client_obd *cli, struct osc_page *opg);
47 static void osc_lru_use(struct client_obd *cli, struct osc_page *opg);
48 static int osc_lru_reserve(const struct lu_env *env, struct osc_object *obj,
49                            struct osc_page *opg);
50
51 /** \addtogroup osc
52  *  @{
53  */
54
55 /*
56  * Comment out osc_page_protected because it may sleep inside the
57  * the client_obd_list_lock.
58  * client_obd_list_lock -> osc_ap_completion -> osc_completion ->
59  *   -> osc_page_protected -> osc_page_is_dlocked -> osc_match_base
60  *   -> ldlm_lock_match -> sptlrpc_import_check_ctx -> sleep.
61  */
62 #if 0
63 static int osc_page_is_dlocked(const struct lu_env *env,
64                                const struct osc_page *opg,
65                                enum cl_lock_mode mode, int pending, int unref)
66 {
67         struct cl_page         *page;
68         struct osc_object      *obj;
69         struct osc_thread_info *info;
70         struct ldlm_res_id     *resname;
71         struct lustre_handle   *lockh;
72         ldlm_policy_data_t     *policy;
73         ldlm_mode_t             dlmmode;
74         __u64                   flags;
75
76         might_sleep();
77
78         info = osc_env_info(env);
79         resname = &info->oti_resname;
80         policy = &info->oti_policy;
81         lockh = &info->oti_handle;
82         page = opg->ops_cl.cpl_page;
83         obj = cl2osc(opg->ops_cl.cpl_obj);
84
85         flags = LDLM_FL_TEST_LOCK | LDLM_FL_BLOCK_GRANTED;
86         if (pending)
87                 flags |= LDLM_FL_CBPENDING;
88
89         dlmmode = osc_cl_lock2ldlm(mode) | LCK_PW;
90         osc_lock_build_res(env, obj, resname);
91         osc_index2policy(policy, page->cp_obj, osc_index(opg), osc_index(opg));
92         return osc_match_base(osc_export(obj), resname, LDLM_EXTENT, policy,
93                               dlmmode, &flags, NULL, lockh, unref);
94 }
95
96 /**
97  * Checks an invariant that a page in the cache is covered by a lock, as
98  * needed.
99  */
100 static int osc_page_protected(const struct lu_env *env,
101                               const struct osc_page *opg,
102                               enum cl_lock_mode mode, int unref)
103 {
104         struct cl_object_header *hdr;
105         struct cl_lock          *scan;
106         struct cl_page          *page;
107         struct cl_lock_descr    *descr;
108         int result;
109
110         LINVRNT(!opg->ops_temp);
111
112         page = opg->ops_cl.cpl_page;
113         if (page->cp_owner != NULL &&
114             cl_io_top(page->cp_owner)->ci_lockreq == CILR_NEVER)
115                 /*
116                  * If IO is done without locks (liblustre, or lloop), lock is
117                  * not required.
118                  */
119                 result = 1;
120         else
121                 /* otherwise check for a DLM lock */
122         result = osc_page_is_dlocked(env, opg, mode, 1, unref);
123         if (result == 0) {
124                 /* maybe this page is a part of a lockless io? */
125                 hdr = cl_object_header(opg->ops_cl.cpl_obj);
126                 descr = &osc_env_info(env)->oti_descr;
127                 descr->cld_mode = mode;
128                 descr->cld_start = osc_index(opg);
129                 descr->cld_end   = osc_index(opg);
130                 spin_lock(&hdr->coh_lock_guard);
131                 list_for_each_entry(scan, &hdr->coh_locks, cll_linkage) {
132                         /*
133                          * Lock-less sub-lock has to be either in HELD state
134                          * (when io is actively going on), or in CACHED state,
135                          * when top-lock is being unlocked:
136                          * cl_io_unlock()->cl_unuse()->...->lov_lock_unuse().
137                          */
138                         if ((scan->cll_state == CLS_HELD ||
139                              scan->cll_state == CLS_CACHED) &&
140                             cl_lock_ext_match(&scan->cll_descr, descr)) {
141                                 struct osc_lock *olck;
142
143                                 olck = osc_lock_at(scan);
144                                 result = osc_lock_is_lockless(olck);
145                                 break;
146                         }
147                 }
148                 spin_unlock(&hdr->coh_lock_guard);
149         }
150         return result;
151 }
152 #else
153 static int osc_page_protected(const struct lu_env *env,
154                               const struct osc_page *opg,
155                               enum cl_lock_mode mode, int unref)
156 {
157         return 1;
158 }
159 #endif
160
161 /*****************************************************************************
162  *
163  * Page operations.
164  *
165  */
166 static void osc_page_transfer_get(struct osc_page *opg, const char *label)
167 {
168         struct cl_page *page = opg->ops_cl.cpl_page;
169
170         LASSERT(!opg->ops_transfer_pinned);
171         cl_page_get(page);
172         lu_ref_add_atomic(&page->cp_reference, label, page);
173         opg->ops_transfer_pinned = 1;
174 }
175
176 static void osc_page_transfer_put(const struct lu_env *env,
177                                   struct osc_page *opg)
178 {
179         struct cl_page *page = opg->ops_cl.cpl_page;
180
181         if (opg->ops_transfer_pinned) {
182                 opg->ops_transfer_pinned = 0;
183                 lu_ref_del(&page->cp_reference, "transfer", page);
184                 cl_page_put(env, page);
185         }
186 }
187
188 /**
189  * This is called once for every page when it is submitted for a transfer
190  * either opportunistic (osc_page_cache_add()), or immediate
191  * (osc_page_submit()).
192  */
193 static void osc_page_transfer_add(const struct lu_env *env,
194                                   struct osc_page *opg, enum cl_req_type crt)
195 {
196         struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
197
198         /* ops_lru and ops_inflight share the same field, so take it from LRU
199          * first and then use it as inflight. */
200         osc_lru_use(osc_cli(obj), opg);
201
202         spin_lock(&obj->oo_seatbelt);
203         list_add(&opg->ops_inflight, &obj->oo_inflight[crt]);
204         opg->ops_submitter = current;
205         spin_unlock(&obj->oo_seatbelt);
206 }
207
208 int osc_page_cache_add(const struct lu_env *env,
209                         const struct cl_page_slice *slice, struct cl_io *io)
210 {
211         struct osc_page *opg = cl2osc_page(slice);
212         int result;
213         ENTRY;
214
215         LINVRNT(osc_page_protected(env, opg, CLM_WRITE, 0));
216
217         osc_page_transfer_get(opg, "transfer\0cache");
218         result = osc_queue_async_io(env, io, opg);
219         if (result != 0)
220                 osc_page_transfer_put(env, opg);
221         else
222                 osc_page_transfer_add(env, opg, CRT_WRITE);
223
224         RETURN(result);
225 }
226
227 void osc_index2policy(ldlm_policy_data_t *policy, const struct cl_object *obj,
228                       pgoff_t start, pgoff_t end)
229 {
230         memset(policy, 0, sizeof *policy);
231         policy->l_extent.start = cl_offset(obj, start);
232         policy->l_extent.end   = cl_offset(obj, end + 1) - 1;
233 }
234
235 static const char *osc_list(struct list_head *head)
236 {
237         return list_empty(head) ? "-" : "+";
238 }
239
240 static inline cfs_time_t osc_submit_duration(struct osc_page *opg)
241 {
242         if (opg->ops_submit_time == 0)
243                 return 0;
244
245         return (cfs_time_current() - opg->ops_submit_time);
246 }
247
248 static int osc_page_print(const struct lu_env *env,
249                           const struct cl_page_slice *slice,
250                           void *cookie, lu_printer_t printer)
251 {
252         struct osc_page       *opg = cl2osc_page(slice);
253         struct osc_async_page *oap = &opg->ops_oap;
254         struct osc_object     *obj = cl2osc(slice->cpl_obj);
255         struct client_obd     *cli = &osc_export(obj)->exp_obd->u.cli;
256
257         return (*printer)(env, cookie, LUSTRE_OSC_NAME"-page@%p %lu: "
258                           "1< %#x %d %u %s %s > "
259                           "2< "LPD64" %u %u %#x %#x | %p %p %p > "
260                           "3< %s %p %d %lu %d > "
261                           "4< %d %d %d %lu %s | %s %s %s %s > "
262                           "5< %s %s %s %s | %d %s | %d %s %s>\n",
263                           opg, osc_index(opg),
264                           /* 1 */
265                           oap->oap_magic, oap->oap_cmd,
266                           oap->oap_interrupted,
267                           osc_list(&oap->oap_pending_item),
268                           osc_list(&oap->oap_rpc_item),
269                           /* 2 */
270                           oap->oap_obj_off, oap->oap_page_off, oap->oap_count,
271                           oap->oap_async_flags, oap->oap_brw_flags,
272                           oap->oap_request, oap->oap_cli, obj,
273                           /* 3 */
274                           osc_list(&opg->ops_inflight),
275                           opg->ops_submitter, opg->ops_transfer_pinned,
276                           osc_submit_duration(opg), opg->ops_srvlock,
277                           /* 4 */
278                           cli->cl_r_in_flight, cli->cl_w_in_flight,
279                           cli->cl_max_rpcs_in_flight,
280                           cli->cl_avail_grant,
281                           osc_list(&cli->cl_cache_waiters),
282                           osc_list(&cli->cl_loi_ready_list),
283                           osc_list(&cli->cl_loi_hp_ready_list),
284                           osc_list(&cli->cl_loi_write_list),
285                           osc_list(&cli->cl_loi_read_list),
286                           /* 5 */
287                           osc_list(&obj->oo_ready_item),
288                           osc_list(&obj->oo_hp_ready_item),
289                           osc_list(&obj->oo_write_item),
290                           osc_list(&obj->oo_read_item),
291                           atomic_read(&obj->oo_nr_reads),
292                           osc_list(&obj->oo_reading_exts),
293                           atomic_read(&obj->oo_nr_writes),
294                           osc_list(&obj->oo_hp_exts),
295                           osc_list(&obj->oo_urgent_exts));
296 }
297
298 static void osc_page_delete(const struct lu_env *env,
299                             const struct cl_page_slice *slice)
300 {
301         struct osc_page   *opg = cl2osc_page(slice);
302         struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
303         int rc;
304
305         LINVRNT(opg->ops_temp || osc_page_protected(env, opg, CLM_READ, 1));
306
307         ENTRY;
308         CDEBUG(D_TRACE, "%p\n", opg);
309         osc_page_transfer_put(env, opg);
310         rc = osc_teardown_async_page(env, obj, opg);
311         if (rc) {
312                 CL_PAGE_DEBUG(D_ERROR, env, slice->cpl_page,
313                               "Trying to teardown failed: %d\n", rc);
314                 LASSERT(0);
315         }
316
317         spin_lock(&obj->oo_seatbelt);
318         if (opg->ops_submitter != NULL) {
319                 LASSERT(!list_empty(&opg->ops_inflight));
320                 list_del_init(&opg->ops_inflight);
321                 opg->ops_submitter = NULL;
322         }
323         spin_unlock(&obj->oo_seatbelt);
324
325         osc_lru_del(osc_cli(obj), opg);
326
327         if (slice->cpl_page->cp_type == CPT_CACHEABLE) {
328                 void *value;
329
330                 spin_lock(&obj->oo_tree_lock);
331                 value = radix_tree_delete(&obj->oo_tree, osc_index(opg));
332                 if (value != NULL)
333                         --obj->oo_npages;
334                 spin_unlock(&obj->oo_tree_lock);
335
336                 LASSERT(ergo(value != NULL, value == opg));
337         }
338
339         EXIT;
340 }
341
342 static void osc_page_clip(const struct lu_env *env,
343                           const struct cl_page_slice *slice,
344                           int from, int to)
345 {
346         struct osc_page       *opg = cl2osc_page(slice);
347         struct osc_async_page *oap = &opg->ops_oap;
348
349         LINVRNT(osc_page_protected(env, opg, CLM_READ, 0));
350
351         opg->ops_from = from;
352         opg->ops_to   = to;
353         spin_lock(&oap->oap_lock);
354         oap->oap_async_flags |= ASYNC_COUNT_STABLE;
355         spin_unlock(&oap->oap_lock);
356 }
357
358 static int osc_page_cancel(const struct lu_env *env,
359                            const struct cl_page_slice *slice)
360 {
361         struct osc_page *opg = cl2osc_page(slice);
362         int rc = 0;
363
364         LINVRNT(osc_page_protected(env, opg, CLM_READ, 0));
365
366         /* Check if the transferring against this page
367          * is completed, or not even queued. */
368         if (opg->ops_transfer_pinned)
369                 /* FIXME: may not be interrupted.. */
370                 rc = osc_cancel_async_page(env, opg);
371         LASSERT(ergo(rc == 0, opg->ops_transfer_pinned == 0));
372         return rc;
373 }
374
375 static int osc_page_flush(const struct lu_env *env,
376                           const struct cl_page_slice *slice,
377                           struct cl_io *io)
378 {
379         struct osc_page *opg = cl2osc_page(slice);
380         int rc = 0;
381         ENTRY;
382         rc = osc_flush_async_page(env, io, opg);
383         RETURN(rc);
384 }
385
386 static const struct cl_page_operations osc_page_ops = {
387         .cpo_print         = osc_page_print,
388         .cpo_delete        = osc_page_delete,
389         .cpo_clip           = osc_page_clip,
390         .cpo_cancel         = osc_page_cancel,
391         .cpo_flush          = osc_page_flush
392 };
393
394 int osc_page_init(const struct lu_env *env, struct cl_object *obj,
395                   struct cl_page *page, pgoff_t index)
396 {
397         struct osc_object *osc = cl2osc(obj);
398         struct osc_page   *opg = cl_object_page_slice(obj, page);
399         int result;
400
401         opg->ops_from = 0;
402         opg->ops_to   = PAGE_CACHE_SIZE;
403
404         result = osc_prep_async_page(osc, opg, page->cp_vmpage,
405                                      cl_offset(obj, index));
406         if (result == 0) {
407                 struct osc_io *oio = osc_env_io(env);
408                 opg->ops_srvlock = osc_io_srvlock(oio);
409                 cl_page_slice_add(page, &opg->ops_cl, obj, index,
410                                   &osc_page_ops);
411         }
412         /*
413          * Cannot assert osc_page_protected() here as read-ahead
414          * creates temporary pages outside of a lock.
415          */
416 #ifdef CONFIG_LUSTRE_DEBUG_EXPENSIVE_CHECK
417         opg->ops_temp = !osc_page_protected(env, opg, CLM_READ, 1);
418 #endif
419         /* ops_inflight and ops_lru are the same field, but it doesn't
420          * hurt to initialize it twice :-) */
421         INIT_LIST_HEAD(&opg->ops_inflight);
422         INIT_LIST_HEAD(&opg->ops_lru);
423
424         /* reserve an LRU space for this page */
425         if (page->cp_type == CPT_CACHEABLE && result == 0) {
426                 result = osc_lru_reserve(env, osc, opg);
427                 if (result == 0) {
428                         spin_lock(&osc->oo_tree_lock);
429                         result = radix_tree_insert(&osc->oo_tree, index, opg);
430                         if (result == 0)
431                                 ++osc->oo_npages;
432                         spin_unlock(&osc->oo_tree_lock);
433                         LASSERT(result == 0);
434                 }
435         }
436
437         return result;
438 }
439
440 /**
441  * Helper function called by osc_io_submit() for every page in an immediate
442  * transfer (i.e., transferred synchronously).
443  */
444 void osc_page_submit(const struct lu_env *env, struct osc_page *opg,
445                      enum cl_req_type crt, int brw_flags)
446 {
447         struct osc_async_page *oap = &opg->ops_oap;
448         struct osc_object     *obj = oap->oap_obj;
449
450         LINVRNT(osc_page_protected(env, opg,
451                                    crt == CRT_WRITE ? CLM_WRITE : CLM_READ, 1));
452
453         LASSERTF(oap->oap_magic == OAP_MAGIC, "Bad oap magic: oap %p, "
454                  "magic 0x%x\n", oap, oap->oap_magic);
455         LASSERT(oap->oap_async_flags & ASYNC_READY);
456         LASSERT(oap->oap_async_flags & ASYNC_COUNT_STABLE);
457
458         oap->oap_cmd       = crt == CRT_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ;
459         oap->oap_page_off  = opg->ops_from;
460         oap->oap_count     = opg->ops_to - opg->ops_from;
461         oap->oap_brw_flags = OBD_BRW_SYNC | brw_flags;
462
463         if (!client_is_remote(osc_export(obj)) &&
464                         cfs_capable(CFS_CAP_SYS_RESOURCE)) {
465                 oap->oap_brw_flags |= OBD_BRW_NOQUOTA;
466                 oap->oap_cmd |= OBD_BRW_NOQUOTA;
467         }
468
469         opg->ops_submit_time = cfs_time_current();
470         osc_page_transfer_get(opg, "transfer\0imm");
471         osc_page_transfer_add(env, opg, crt);
472 }
473
474 /* --------------- LRU page management ------------------ */
475
476 /* OSC is a natural place to manage LRU pages as applications are specialized
477  * to write OSC by OSC. Ideally, if one OSC is used more frequently it should
478  * occupy more LRU slots. On the other hand, we should avoid using up all LRU
479  * slots (client_obd::cl_lru_left) otherwise process has to be put into sleep
480  * for free LRU slots - this will be very bad so the algorithm requires each
481  * OSC to free slots voluntarily to maintain a reasonable number of free slots
482  * at any time.
483  */
484
485 static DECLARE_WAIT_QUEUE_HEAD(osc_lru_waitq);
486 /* LRU pages are freed in batch mode. OSC should at least free this
487  * number of pages to avoid running out of LRU budget, and.. */
488 static const int lru_shrink_min = 2 << (20 - PAGE_CACHE_SHIFT); /* 2M */
489 /* free this number at most otherwise it will take too long time to finsih. */
490 static const int lru_shrink_max = 8 << (20 - PAGE_CACHE_SHIFT); /* 8M */
491
492 /* Check if we can free LRU slots from this OSC. If there exists LRU waiters,
493  * we should free slots aggressively. In this way, slots are freed in a steady
494  * step to maintain fairness among OSCs.
495  *
496  * Return how many LRU pages should be freed. */
497 static int osc_cache_too_much(struct client_obd *cli)
498 {
499         struct cl_client_cache *cache = cli->cl_cache;
500         long pages = atomic_long_read(&cli->cl_lru_in_list);
501         unsigned long budget;
502
503         LASSERT(cache != NULL);
504         budget = cache->ccc_lru_max / (atomic_read(&cache->ccc_users) - 2);
505
506         /* if it's going to run out LRU slots, we should free some, but not
507          * too much to maintain faireness among OSCs. */
508         if (atomic_long_read(cli->cl_lru_left) < cache->ccc_lru_max >> 4) {
509                 if (pages >= budget)
510                         return lru_shrink_max;
511                 else if (pages >= budget / 2)
512                         return lru_shrink_min;
513 #if 0
514         } else if (pages >= budget * 2)
515                 return lru_shrink_min;
516 #endif
517         }
518         return 0;
519 }
520
521 int lru_queue_work(const struct lu_env *env, void *data)
522 {
523         struct client_obd *cli = data;
524
525         CDEBUG(D_CACHE, "Run LRU work for client obd %p.\n", cli);
526
527         if (osc_cache_too_much(cli))
528                 osc_lru_shrink(env, cli, lru_shrink_max, true);
529
530         RETURN(0);
531 }
532
533 void osc_lru_add_batch(struct client_obd *cli, struct list_head *plist)
534 {
535         struct list_head lru = LIST_HEAD_INIT(lru);
536         struct osc_async_page *oap;
537         long npages = 0;
538
539         list_for_each_entry(oap, plist, oap_pending_item) {
540                 struct osc_page *opg = oap2osc_page(oap);
541
542                 if (!opg->ops_in_lru)
543                         continue;
544
545                 ++npages;
546                 LASSERT(list_empty(&opg->ops_lru));
547                 list_add(&opg->ops_lru, &lru);
548         }
549
550         if (npages > 0) {
551                 spin_lock(&cli->cl_lru_list_lock);
552                 list_splice_tail(&lru, &cli->cl_lru_list);
553                 atomic_long_sub(npages, &cli->cl_lru_busy);
554                 atomic_long_add(npages, &cli->cl_lru_in_list);
555                 spin_unlock(&cli->cl_lru_list_lock);
556
557                 /* XXX: May set force to be true for better performance */
558                 if (osc_cache_too_much(cli))
559                         (void)ptlrpcd_queue_work(cli->cl_lru_work);
560         }
561 }
562
563 static void __osc_lru_del(struct client_obd *cli, struct osc_page *opg)
564 {
565         LASSERT(atomic_long_read(&cli->cl_lru_in_list) > 0);
566         list_del_init(&opg->ops_lru);
567         atomic_long_dec(&cli->cl_lru_in_list);
568 }
569
570 /**
571  * Page is being destroyed. The page may be not in LRU list, if the transfer
572  * has never finished(error occurred).
573  */
574 static void osc_lru_del(struct client_obd *cli, struct osc_page *opg)
575 {
576         if (opg->ops_in_lru) {
577                 spin_lock(&cli->cl_lru_list_lock);
578                 if (!list_empty(&opg->ops_lru)) {
579                         __osc_lru_del(cli, opg);
580                 } else {
581                         LASSERT(atomic_long_read(&cli->cl_lru_busy) > 0);
582                         atomic_long_dec(&cli->cl_lru_busy);
583                 }
584                 spin_unlock(&cli->cl_lru_list_lock);
585
586                 atomic_long_inc(cli->cl_lru_left);
587                 /* this is a great place to release more LRU pages if
588                  * this osc occupies too many LRU pages and kernel is
589                  * stealing one of them. */
590                 if (!memory_pressure_get())
591                         (void)ptlrpcd_queue_work(cli->cl_lru_work);
592                 wake_up(&osc_lru_waitq);
593         } else {
594                 LASSERT(list_empty(&opg->ops_lru));
595         }
596 }
597
598 /**
599  * Delete page from LRUlist for redirty.
600  */
601 static void osc_lru_use(struct client_obd *cli, struct osc_page *opg)
602 {
603         /* If page is being transferred for the first time,
604          * ops_lru should be empty */
605         if (opg->ops_in_lru && !list_empty(&opg->ops_lru)) {
606                 spin_lock(&cli->cl_lru_list_lock);
607                 __osc_lru_del(cli, opg);
608                 spin_unlock(&cli->cl_lru_list_lock);
609                 atomic_long_inc(&cli->cl_lru_busy);
610         }
611 }
612
613 static void discard_pagevec(const struct lu_env *env, struct cl_io *io,
614                                 struct cl_page **pvec, int max_index)
615 {
616         int i;
617
618         for (i = 0; i < max_index; i++) {
619                 struct cl_page *page = pvec[i];
620
621                 LASSERT(cl_page_is_owned(page, io));
622                 cl_page_discard(env, io, page);
623                 cl_page_disown(env, io, page);
624                 cl_page_put(env, page);
625
626                 pvec[i] = NULL;
627         }
628 }
629
630 /**
631  * Check if a cl_page can be released, i.e, it's not being used.
632  *
633  * If unstable account is turned on, bulk transfer may hold one refcount
634  * for recovery so we need to check vmpage refcount as well; otherwise,
635  * even we can destroy cl_page but the corresponding vmpage can't be reused.
636  */
637 static inline bool lru_page_busy(struct client_obd *cli, struct cl_page *page)
638 {
639         if (cl_page_in_use_noref(page))
640                 return true;
641
642         if (cli->cl_cache->ccc_unstable_check) {
643                 struct page *vmpage = cl_page_vmpage(page);
644
645                 /* vmpage have two known users: cl_page and VM page cache */
646                 if (page_count(vmpage) - page_mapcount(vmpage) > 2)
647                         return true;
648         }
649         return false;
650 }
651
652 /**
653  * Drop @target of pages from LRU at most.
654  */
655 long osc_lru_shrink(const struct lu_env *env, struct client_obd *cli,
656                    long target, bool force)
657 {
658         struct cl_io *io;
659         struct cl_object *clobj = NULL;
660         struct cl_page **pvec;
661         struct osc_page *opg;
662         long count = 0;
663         int maxscan = 0;
664         int index = 0;
665         int rc = 0;
666         ENTRY;
667
668         LASSERT(atomic_long_read(&cli->cl_lru_in_list) >= 0);
669         if (atomic_long_read(&cli->cl_lru_in_list) == 0 || target <= 0)
670                 RETURN(0);
671
672         if (!force) {
673                 if (atomic_read(&cli->cl_lru_shrinkers) > 0)
674                         RETURN(-EBUSY);
675
676                 if (atomic_inc_return(&cli->cl_lru_shrinkers) > 1) {
677                         atomic_dec(&cli->cl_lru_shrinkers);
678                         RETURN(-EBUSY);
679                 }
680         } else {
681                 atomic_inc(&cli->cl_lru_shrinkers);
682         }
683
684         pvec = (struct cl_page **)osc_env_info(env)->oti_pvec;
685         io = &osc_env_info(env)->oti_io;
686
687         spin_lock(&cli->cl_lru_list_lock);
688         maxscan = min(target << 1, atomic_long_read(&cli->cl_lru_in_list));
689         while (!list_empty(&cli->cl_lru_list)) {
690                 struct cl_page *page;
691                 bool will_free = false;
692
693                 if (--maxscan < 0)
694                         break;
695
696                 opg = list_entry(cli->cl_lru_list.next, struct osc_page,
697                                  ops_lru);
698                 page = opg->ops_cl.cpl_page;
699                 if (lru_page_busy(cli, page)) {
700                         list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
701                         continue;
702                 }
703
704                 LASSERT(page->cp_obj != NULL);
705                 if (clobj != page->cp_obj) {
706                         struct cl_object *tmp = page->cp_obj;
707
708                         cl_object_get(tmp);
709                         spin_unlock(&cli->cl_lru_list_lock);
710
711                         if (clobj != NULL) {
712                                 discard_pagevec(env, io, pvec, index);
713                                 index = 0;
714
715                                 cl_io_fini(env, io);
716                                 cl_object_put(env, clobj);
717                                 clobj = NULL;
718                         }
719
720                         clobj = tmp;
721                         io->ci_obj = clobj;
722                         io->ci_ignore_layout = 1;
723                         rc = cl_io_init(env, io, CIT_MISC, clobj);
724
725                         spin_lock(&cli->cl_lru_list_lock);
726
727                         if (rc != 0)
728                                 break;
729
730                         ++maxscan;
731                         continue;
732                 }
733
734                 if (cl_page_own_try(env, io, page) == 0) {
735                         if (!lru_page_busy(cli, page)) {
736                                 /* remove it from lru list earlier to avoid
737                                  * lock contention */
738                                 __osc_lru_del(cli, opg);
739                                 opg->ops_in_lru = 0; /* will be discarded */
740
741                                 cl_page_get(page);
742                                 will_free = true;
743                         } else {
744                                 cl_page_disown(env, io, page);
745                         }
746                 }
747
748                 if (!will_free) {
749                         list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
750                         continue;
751                 }
752
753                 /* Don't discard and free the page with cl_lru_list held */
754                 pvec[index++] = page;
755                 if (unlikely(index == OTI_PVEC_SIZE)) {
756                         spin_unlock(&cli->cl_lru_list_lock);
757                         discard_pagevec(env, io, pvec, index);
758                         index = 0;
759
760                         spin_lock(&cli->cl_lru_list_lock);
761                 }
762
763                 if (++count >= target)
764                         break;
765         }
766         spin_unlock(&cli->cl_lru_list_lock);
767
768         if (clobj != NULL) {
769                 discard_pagevec(env, io, pvec, index);
770
771                 cl_io_fini(env, io);
772                 cl_object_put(env, clobj);
773         }
774
775         atomic_dec(&cli->cl_lru_shrinkers);
776         if (count > 0) {
777                 atomic_long_add(count, cli->cl_lru_left);
778                 wake_up_all(&osc_lru_waitq);
779         }
780         RETURN(count > 0 ? count : rc);
781 }
782
783 long osc_lru_reclaim(struct client_obd *cli)
784 {
785         struct cl_env_nest nest;
786         struct lu_env *env;
787         struct cl_client_cache *cache = cli->cl_cache;
788         long rc = 0;
789         int max_scans;
790         ENTRY;
791
792         LASSERT(cache != NULL);
793
794         env = cl_env_nested_get(&nest);
795         if (IS_ERR(env))
796                 RETURN(rc);
797
798         rc = osc_lru_shrink(env, cli, osc_cache_too_much(cli), false);
799         if (rc != 0) {
800                 if (rc == -EBUSY)
801                         rc = 0;
802
803                 CDEBUG(D_CACHE, "%s: Free %ld pages from own LRU: %p.\n",
804                         cli->cl_import->imp_obd->obd_name, rc, cli);
805                 GOTO(out, rc);
806         }
807
808         CDEBUG(D_CACHE, "%s: cli %p no free slots, pages: %ld, busy: %ld.\n",
809                 cli->cl_import->imp_obd->obd_name, cli,
810                 atomic_long_read(&cli->cl_lru_in_list),
811                 atomic_long_read(&cli->cl_lru_busy));
812
813         /* Reclaim LRU slots from other client_obd as it can't free enough
814          * from its own. This should rarely happen. */
815         spin_lock(&cache->ccc_lru_lock);
816         LASSERT(!list_empty(&cache->ccc_lru));
817
818         cache->ccc_lru_shrinkers++;
819         list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
820
821         max_scans = atomic_read(&cache->ccc_users) - 2;
822         while (--max_scans > 0 && !list_empty(&cache->ccc_lru)) {
823                 cli = list_entry(cache->ccc_lru.next, struct client_obd,
824                                  cl_lru_osc);
825
826                 CDEBUG(D_CACHE, "%s: cli %p LRU pages: %ld, busy: %ld.\n",
827                         cli->cl_import->imp_obd->obd_name, cli,
828                         atomic_long_read(&cli->cl_lru_in_list),
829                         atomic_long_read(&cli->cl_lru_busy));
830
831                 list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
832                 if (osc_cache_too_much(cli) > 0) {
833                         spin_unlock(&cache->ccc_lru_lock);
834
835                         rc = osc_lru_shrink(env, cli, osc_cache_too_much(cli),
836                                             true);
837                         spin_lock(&cache->ccc_lru_lock);
838                         if (rc != 0)
839                                 break;
840                 }
841         }
842         spin_unlock(&cache->ccc_lru_lock);
843
844 out:
845         cl_env_nested_put(&nest, env);
846         CDEBUG(D_CACHE, "%s: cli %p freed %ld pages.\n",
847                 cli->cl_import->imp_obd->obd_name, cli, rc);
848         return rc;
849 }
850
851 /**
852  * osc_lru_reserve() is called to reserve an LRU slot for a cl_page.
853  *
854  * Usually the LRU slots are reserved in osc_io_iter_rw_init().
855  * Only in the case that the LRU slots are in extreme shortage, it should
856  * have reserved enough slots for an IO.
857  */
858 static int osc_lru_reserve(const struct lu_env *env, struct osc_object *obj,
859                            struct osc_page *opg)
860 {
861         struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
862         struct osc_io *oio = osc_env_io(env);
863         struct client_obd *cli = osc_cli(obj);
864         int rc = 0;
865         ENTRY;
866
867         if (cli->cl_cache == NULL) /* shall not be in LRU */
868                 RETURN(0);
869
870         if (oio->oi_lru_reserved > 0) {
871                 --oio->oi_lru_reserved;
872                 goto out;
873         }
874
875         LASSERT(atomic_long_read(cli->cl_lru_left) >= 0);
876         while (!atomic_long_add_unless(cli->cl_lru_left, -1, 0)) {
877
878                 /* run out of LRU spaces, try to drop some by itself */
879                 rc = osc_lru_reclaim(cli);
880                 if (rc < 0)
881                         break;
882                 if (rc > 0)
883                         continue;
884
885                 cond_resched();
886                 rc = l_wait_event(osc_lru_waitq,
887                                 atomic_long_read(cli->cl_lru_left) > 0,
888                                 &lwi);
889                 if (rc < 0)
890                         break;
891         }
892
893 out:
894         if (rc >= 0) {
895                 atomic_long_inc(&cli->cl_lru_busy);
896                 opg->ops_in_lru = 1;
897                 rc = 0;
898         }
899
900         RETURN(rc);
901 }
902
903 /**
904  * Atomic operations are expensive. We accumulate the accounting for the
905  * same page zone to get better performance.
906  * In practice this can work pretty good because the pages in the same RPC
907  * are likely from the same page zone.
908  */
909 static inline void unstable_page_accounting(struct ptlrpc_bulk_desc *desc,
910                                             int factor)
911 {
912         int page_count = desc->bd_iov_count;
913         void *zone = NULL;
914         int count = 0;
915         int i;
916
917         LASSERT(ptlrpc_is_bulk_desc_kiov(desc->bd_type));
918
919         for (i = 0; i < page_count; i++) {
920                 void *pz = page_zone(BD_GET_KIOV(desc, i).kiov_page);
921
922                 if (likely(pz == zone)) {
923                         ++count;
924                         continue;
925                 }
926
927                 if (count > 0) {
928                         mod_zone_page_state(zone, NR_UNSTABLE_NFS,
929                                             factor * count);
930                         count = 0;
931                 }
932                 zone = pz;
933                 ++count;
934         }
935         if (count > 0)
936                 mod_zone_page_state(zone, NR_UNSTABLE_NFS, factor * count);
937 }
938
939 static inline void add_unstable_page_accounting(struct ptlrpc_bulk_desc *desc)
940 {
941         unstable_page_accounting(desc, 1);
942 }
943
944 static inline void dec_unstable_page_accounting(struct ptlrpc_bulk_desc *desc)
945 {
946         unstable_page_accounting(desc, -1);
947 }
948
949 /**
950  * Performs "unstable" page accounting. This function balances the
951  * increment operations performed in osc_inc_unstable_pages. It is
952  * registered as the RPC request callback, and is executed when the
953  * bulk RPC is committed on the server. Thus at this point, the pages
954  * involved in the bulk transfer are no longer considered unstable.
955  *
956  * If this function is called, the request should have been committed
957  * or req:rq_unstable must have been set; it implies that the unstable
958  * statistic have been added.
959  */
960 void osc_dec_unstable_pages(struct ptlrpc_request *req)
961 {
962         struct ptlrpc_bulk_desc *desc       = req->rq_bulk;
963         struct client_obd       *cli        = &req->rq_import->imp_obd->u.cli;
964         int                      page_count = desc->bd_iov_count;
965         long                     unstable_count;
966
967         LASSERT(page_count >= 0);
968         dec_unstable_page_accounting(desc);
969
970         unstable_count = atomic_long_sub_return(page_count,
971                                                 &cli->cl_unstable_count);
972         LASSERT(unstable_count >= 0);
973
974         unstable_count = atomic_long_sub_return(page_count,
975                                            &cli->cl_cache->ccc_unstable_nr);
976         LASSERT(unstable_count >= 0);
977         if (unstable_count == 0)
978                 wake_up_all(&cli->cl_cache->ccc_unstable_waitq);
979
980         if (osc_cache_too_much(cli))
981                 (void)ptlrpcd_queue_work(cli->cl_lru_work);
982 }
983
984 /**
985  * "unstable" page accounting. See: osc_dec_unstable_pages.
986  */
987 void osc_inc_unstable_pages(struct ptlrpc_request *req)
988 {
989         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
990         struct client_obd       *cli  = &req->rq_import->imp_obd->u.cli;
991         long                     page_count = desc->bd_iov_count;
992
993         /* No unstable page tracking */
994         if (cli->cl_cache == NULL || !cli->cl_cache->ccc_unstable_check)
995                 return;
996
997         add_unstable_page_accounting(desc);
998         atomic_long_add(page_count, &cli->cl_unstable_count);
999         atomic_long_add(page_count, &cli->cl_cache->ccc_unstable_nr);
1000
1001         /* If the request has already been committed (i.e. brw_commit
1002          * called via rq_commit_cb), we need to undo the unstable page
1003          * increments we just performed because rq_commit_cb wont be
1004          * called again. */
1005         spin_lock(&req->rq_lock);
1006         if (unlikely(req->rq_committed)) {
1007                 spin_unlock(&req->rq_lock);
1008
1009                 osc_dec_unstable_pages(req);
1010         } else {
1011                 req->rq_unstable = 1;
1012                 spin_unlock(&req->rq_lock);
1013         }
1014 }
1015
1016 /**
1017  * Check if it piggybacks SOFT_SYNC flag to OST from this OSC.
1018  * This function will be called by every BRW RPC so it's critical
1019  * to make this function fast.
1020  */
1021 bool osc_over_unstable_soft_limit(struct client_obd *cli)
1022 {
1023         long unstable_nr, osc_unstable_count;
1024
1025         /* Can't check cli->cl_unstable_count, therefore, no soft limit */
1026         if (cli->cl_cache == NULL || !cli->cl_cache->ccc_unstable_check)
1027                 return false;
1028
1029         osc_unstable_count = atomic_long_read(&cli->cl_unstable_count);
1030         unstable_nr = atomic_long_read(&cli->cl_cache->ccc_unstable_nr);
1031
1032         CDEBUG(D_CACHE,
1033                "%s: cli: %p unstable pages: %lu, osc unstable pages: %lu\n",
1034                cli->cl_import->imp_obd->obd_name, cli,
1035                unstable_nr, osc_unstable_count);
1036
1037         /* If the LRU slots are in shortage - 25% remaining AND this OSC
1038          * has one full RPC window of unstable pages, it's a good chance
1039          * to piggyback a SOFT_SYNC flag.
1040          * Please notice that the OST won't take immediate response for the
1041          * SOFT_SYNC request so active OSCs will have more chance to carry
1042          * the flag, this is reasonable. */
1043         return unstable_nr > cli->cl_cache->ccc_lru_max >> 2 &&
1044                osc_unstable_count > cli->cl_max_pages_per_rpc *
1045                                     cli->cl_max_rpcs_in_flight;
1046 }
1047
1048 /** @} osc */