drivers/staging/lustre/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, 2015, 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  */
  40
  41 #define DEBUG_SUBSYSTEM S_OSC
  42
  43 #include "osc_cl_internal.h"
  44
  45 static void osc_lru_del(struct client_obd *cli, struct osc_page *opg, bool del);
  46 static void osc_lru_add(struct client_obd *cli, struct osc_page *opg);
  47 static int osc_lru_reserve(const struct lu_env *env, struct osc_object *obj,
  48                            struct osc_page *opg);
  49
  50 /** \addtogroup osc
  51  *  @{
  52  */
  53
  54 static int osc_page_protected(const struct lu_env *env,
  55                               const struct osc_page *opg,
  56                               enum cl_lock_mode mode, int unref)
  57 {
  58         return 1;
  59 }
  60
  61 /*****************************************************************************
  62  *
  63  * Page operations.
  64  *
  65  */
  66 static void osc_page_fini(const struct lu_env *env,
  67                           struct cl_page_slice *slice)
  68 {
  69         struct osc_page *opg = cl2osc_page(slice);
  70
  71         CDEBUG(D_TRACE, "%p\n", opg);
  72         LASSERT(!opg->ops_lock);
  73 }
  74
  75 static void osc_page_transfer_get(struct osc_page *opg, const char *label)
  76 {
  77         struct cl_page *page = cl_page_top(opg->ops_cl.cpl_page);
  78
  79         LASSERT(!opg->ops_transfer_pinned);
  80         cl_page_get(page);
  81         lu_ref_add_atomic(&page->cp_reference, label, page);
  82         opg->ops_transfer_pinned = 1;
  83 }
  84
  85 static void osc_page_transfer_put(const struct lu_env *env,
  86                                   struct osc_page *opg)
  87 {
  88         struct cl_page *page = cl_page_top(opg->ops_cl.cpl_page);
  89
  90         if (opg->ops_transfer_pinned) {
  91                 lu_ref_del(&page->cp_reference, "transfer", page);
  92                 opg->ops_transfer_pinned = 0;
  93                 cl_page_put(env, page);
  94         }
  95 }
  96
  97 /**
  98  * This is called once for every page when it is submitted for a transfer
  99  * either opportunistic (osc_page_cache_add()), or immediate
 100  * (osc_page_submit()).
 101  */
 102 static void osc_page_transfer_add(const struct lu_env *env,
 103                                   struct osc_page *opg, enum cl_req_type crt)
 104 {
 105         struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
 106
 107         /* ops_lru and ops_inflight share the same field, so take it from LRU
 108          * first and then use it as inflight.
 109          */
 110         osc_lru_del(osc_cli(obj), opg, false);
 111
 112         spin_lock(&obj->oo_seatbelt);
 113         list_add(&opg->ops_inflight, &obj->oo_inflight[crt]);
 114         opg->ops_submitter = current;
 115         spin_unlock(&obj->oo_seatbelt);
 116 }
 117
 118 static int osc_page_cache_add(const struct lu_env *env,
 119                               const struct cl_page_slice *slice,
 120                               struct cl_io *io)
 121 {
 122         struct osc_io *oio = osc_env_io(env);
 123         struct osc_page *opg = cl2osc_page(slice);
 124         int result;
 125
 126         LINVRNT(osc_page_protected(env, opg, CLM_WRITE, 0));
 127
 128         osc_page_transfer_get(opg, "transfer\0cache");
 129         result = osc_queue_async_io(env, io, opg);
 130         if (result != 0)
 131                 osc_page_transfer_put(env, opg);
 132         else
 133                 osc_page_transfer_add(env, opg, CRT_WRITE);
 134
 135         /* for sync write, kernel will wait for this page to be flushed before
 136          * osc_io_end() is called, so release it earlier.
 137          * for mkwrite(), it's known there is no further pages.
 138          */
 139         if (cl_io_is_sync_write(io) || cl_io_is_mkwrite(io)) {
 140                 if (oio->oi_active) {
 141                         osc_extent_release(env, oio->oi_active);
 142                         oio->oi_active = NULL;
 143                 }
 144         }
 145
 146         return result;
 147 }
 148
 149 void osc_index2policy(ldlm_policy_data_t *policy, const struct cl_object *obj,
 150                       pgoff_t start, pgoff_t end)
 151 {
 152         memset(policy, 0, sizeof(*policy));
 153         policy->l_extent.start = cl_offset(obj, start);
 154         policy->l_extent.end = cl_offset(obj, end + 1) - 1;
 155 }
 156
 157 static int osc_page_addref_lock(const struct lu_env *env,
 158                                 struct osc_page *opg,
 159                                 struct cl_lock *lock)
 160 {
 161         struct osc_lock *olock;
 162         int rc;
 163
 164         LASSERT(!opg->ops_lock);
 165
 166         olock = osc_lock_at(lock);
 167         if (atomic_inc_return(&olock->ols_pageref) <= 0) {
 168                 atomic_dec(&olock->ols_pageref);
 169                 rc = -ENODATA;
 170         } else {
 171                 cl_lock_get(lock);
 172                 opg->ops_lock = lock;
 173                 rc = 0;
 174         }
 175         return rc;
 176 }
 177
 178 static void osc_page_putref_lock(const struct lu_env *env,
 179                                  struct osc_page *opg)
 180 {
 181         struct cl_lock *lock = opg->ops_lock;
 182         struct osc_lock *olock;
 183
 184         LASSERT(lock);
 185         olock = osc_lock_at(lock);
 186
 187         atomic_dec(&olock->ols_pageref);
 188         opg->ops_lock = NULL;
 189
 190         cl_lock_put(env, lock);
 191 }
 192
 193 static int osc_page_is_under_lock(const struct lu_env *env,
 194                                   const struct cl_page_slice *slice,
 195                                   struct cl_io *unused)
 196 {
 197         struct cl_lock *lock;
 198         int result = -ENODATA;
 199
 200         lock = cl_lock_at_page(env, slice->cpl_obj, slice->cpl_page,
 201                                NULL, 1, 0);
 202         if (lock) {
 203                 if (osc_page_addref_lock(env, cl2osc_page(slice), lock) == 0)
 204                         result = -EBUSY;
 205                 cl_lock_put(env, lock);
 206         }
 207         return result;
 208 }
 209
 210 static void osc_page_disown(const struct lu_env *env,
 211                             const struct cl_page_slice *slice,
 212                             struct cl_io *io)
 213 {
 214         struct osc_page *opg = cl2osc_page(slice);
 215
 216         if (unlikely(opg->ops_lock))
 217                 osc_page_putref_lock(env, opg);
 218 }
 219
 220 static void osc_page_completion_read(const struct lu_env *env,
 221                                      const struct cl_page_slice *slice,
 222                                      int ioret)
 223 {
 224         struct osc_page *opg = cl2osc_page(slice);
 225         struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
 226
 227         if (likely(opg->ops_lock))
 228                 osc_page_putref_lock(env, opg);
 229         osc_lru_add(osc_cli(obj), opg);
 230 }
 231
 232 static void osc_page_completion_write(const struct lu_env *env,
 233                                       const struct cl_page_slice *slice,
 234                                       int ioret)
 235 {
 236         struct osc_page *opg = cl2osc_page(slice);
 237         struct osc_object *obj = cl2osc(slice->cpl_obj);
 238
 239         osc_lru_add(osc_cli(obj), opg);
 240 }
 241
 242 static int osc_page_fail(const struct lu_env *env,
 243                          const struct cl_page_slice *slice,
 244                          struct cl_io *unused)
 245 {
 246         /*
 247          * Cached read?
 248          */
 249         LBUG();
 250         return 0;
 251 }
 252
 253 static const char *osc_list(struct list_head *head)
 254 {
 255         return list_empty(head) ? "-" : "+";
 256 }
 257
 258 static inline unsigned long osc_submit_duration(struct osc_page *opg)
 259 {
 260         if (opg->ops_submit_time == 0)
 261                 return 0;
 262
 263         return (cfs_time_current() - opg->ops_submit_time);
 264 }
 265
 266 static int osc_page_print(const struct lu_env *env,
 267                           const struct cl_page_slice *slice,
 268                           void *cookie, lu_printer_t printer)
 269 {
 270         struct osc_page *opg = cl2osc_page(slice);
 271         struct osc_async_page *oap = &opg->ops_oap;
 272         struct osc_object *obj = cl2osc(slice->cpl_obj);
 273         struct client_obd *cli = &osc_export(obj)->exp_obd->u.cli;
 274
 275         return (*printer)(env, cookie, LUSTRE_OSC_NAME "-page@%p: 1< %#x %d %u %s %s > 2< %llu %u %u %#x %#x | %p %p %p > 3< %s %p %d %lu %d > 4< %d %d %d %lu %s | %s %s %s %s > 5< %s %s %s %s | %d %s | %d %s %s>\n",
 276                           opg,
 277                           /* 1 */
 278                           oap->oap_magic, oap->oap_cmd,
 279                           oap->oap_interrupted,
 280                           osc_list(&oap->oap_pending_item),
 281                           osc_list(&oap->oap_rpc_item),
 282                           /* 2 */
 283                           oap->oap_obj_off, oap->oap_page_off, oap->oap_count,
 284                           oap->oap_async_flags, oap->oap_brw_flags,
 285                           oap->oap_request, oap->oap_cli, obj,
 286                           /* 3 */
 287                           osc_list(&opg->ops_inflight),
 288                           opg->ops_submitter, opg->ops_transfer_pinned,
 289                           osc_submit_duration(opg), opg->ops_srvlock,
 290                           /* 4 */
 291                           cli->cl_r_in_flight, cli->cl_w_in_flight,
 292                           cli->cl_max_rpcs_in_flight,
 293                           cli->cl_avail_grant,
 294                           osc_list(&cli->cl_cache_waiters),
 295                           osc_list(&cli->cl_loi_ready_list),
 296                           osc_list(&cli->cl_loi_hp_ready_list),
 297                           osc_list(&cli->cl_loi_write_list),
 298                           osc_list(&cli->cl_loi_read_list),
 299                           /* 5 */
 300                           osc_list(&obj->oo_ready_item),
 301                           osc_list(&obj->oo_hp_ready_item),
 302                           osc_list(&obj->oo_write_item),
 303                           osc_list(&obj->oo_read_item),
 304                           atomic_read(&obj->oo_nr_reads),
 305                           osc_list(&obj->oo_reading_exts),
 306                           atomic_read(&obj->oo_nr_writes),
 307                           osc_list(&obj->oo_hp_exts),
 308                           osc_list(&obj->oo_urgent_exts));
 309 }
 310
 311 static void osc_page_delete(const struct lu_env *env,
 312                             const struct cl_page_slice *slice)
 313 {
 314         struct osc_page *opg = cl2osc_page(slice);
 315         struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
 316         int rc;
 317
 318         LINVRNT(opg->ops_temp || osc_page_protected(env, opg, CLM_READ, 1));
 319
 320         CDEBUG(D_TRACE, "%p\n", opg);
 321         osc_page_transfer_put(env, opg);
 322         rc = osc_teardown_async_page(env, obj, opg);
 323         if (rc) {
 324                 CL_PAGE_DEBUG(D_ERROR, env, cl_page_top(slice->cpl_page),
 325                               "Trying to teardown failed: %d\n", rc);
 326                 LASSERT(0);
 327         }
 328
 329         spin_lock(&obj->oo_seatbelt);
 330         if (opg->ops_submitter) {
 331                 LASSERT(!list_empty(&opg->ops_inflight));
 332                 list_del_init(&opg->ops_inflight);
 333                 opg->ops_submitter = NULL;
 334         }
 335         spin_unlock(&obj->oo_seatbelt);
 336
 337         osc_lru_del(osc_cli(obj), opg, true);
 338 }
 339
 340 static void osc_page_clip(const struct lu_env *env,
 341                           const struct cl_page_slice *slice, int from, int to)
 342 {
 343         struct osc_page *opg = cl2osc_page(slice);
 344         struct osc_async_page *oap = &opg->ops_oap;
 345
 346         LINVRNT(osc_page_protected(env, opg, CLM_READ, 0));
 347
 348         opg->ops_from = from;
 349         opg->ops_to = to;
 350         spin_lock(&oap->oap_lock);
 351         oap->oap_async_flags |= ASYNC_COUNT_STABLE;
 352         spin_unlock(&oap->oap_lock);
 353 }
 354
 355 static int osc_page_cancel(const struct lu_env *env,
 356                            const struct cl_page_slice *slice)
 357 {
 358         struct osc_page *opg = cl2osc_page(slice);
 359         int rc = 0;
 360
 361         LINVRNT(osc_page_protected(env, opg, CLM_READ, 0));
 362
 363         /* Check if the transferring against this page
 364          * is completed, or not even queued.
 365          */
 366         if (opg->ops_transfer_pinned)
 367                 /* FIXME: may not be interrupted.. */
 368                 rc = osc_cancel_async_page(env, opg);
 369         LASSERT(ergo(rc == 0, opg->ops_transfer_pinned == 0));
 370         return rc;
 371 }
 372
 373 static int osc_page_flush(const struct lu_env *env,
 374                           const struct cl_page_slice *slice,
 375                           struct cl_io *io)
 376 {
 377         struct osc_page *opg = cl2osc_page(slice);
 378         int rc;
 379
 380         rc = osc_flush_async_page(env, io, opg);
 381         return rc;
 382 }
 383
 384 static const struct cl_page_operations osc_page_ops = {
 385         .cpo_fini         = osc_page_fini,
 386         .cpo_print       = osc_page_print,
 387         .cpo_delete     = osc_page_delete,
 388         .cpo_is_under_lock = osc_page_is_under_lock,
 389         .cpo_disown     = osc_page_disown,
 390         .io = {
 391                 [CRT_READ] = {
 392                         .cpo_cache_add  = osc_page_fail,
 393                         .cpo_completion = osc_page_completion_read
 394                 },
 395                 [CRT_WRITE] = {
 396                         .cpo_cache_add  = osc_page_cache_add,
 397                         .cpo_completion = osc_page_completion_write
 398                 }
 399         },
 400         .cpo_clip          = osc_page_clip,
 401         .cpo_cancel      = osc_page_cancel,
 402         .cpo_flush        = osc_page_flush
 403 };
 404
 405 int osc_page_init(const struct lu_env *env, struct cl_object *obj,
 406                   struct cl_page *page, struct page *vmpage)
 407 {
 408         struct osc_object *osc = cl2osc(obj);
 409         struct osc_page *opg = cl_object_page_slice(obj, page);
 410         int result;
 411
 412         opg->ops_from = 0;
 413         opg->ops_to = PAGE_SIZE;
 414
 415         result = osc_prep_async_page(osc, opg, vmpage,
 416                                      cl_offset(obj, page->cp_index));
 417         if (result == 0) {
 418                 struct osc_io *oio = osc_env_io(env);
 419
 420                 opg->ops_srvlock = osc_io_srvlock(oio);
 421                 cl_page_slice_add(page, &opg->ops_cl, obj, &osc_page_ops);
 422         }
 423         /*
 424          * Cannot assert osc_page_protected() here as read-ahead
 425          * creates temporary pages outside of a lock.
 426          */
 427         /* ops_inflight and ops_lru are the same field, but it doesn't
 428          * hurt to initialize it twice :-)
 429          */
 430         INIT_LIST_HEAD(&opg->ops_inflight);
 431         INIT_LIST_HEAD(&opg->ops_lru);
 432
 433         /* reserve an LRU space for this page */
 434         if (page->cp_type == CPT_CACHEABLE && result == 0)
 435                 result = osc_lru_reserve(env, osc, opg);
 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, magic 0x%x\n",
 454                  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 = brw_flags | OBD_BRW_SYNC;
 462
 463         if (!client_is_remote(osc_export(obj)) &&
 464             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 static atomic_t osc_lru_waiters = ATOMIC_INIT(0);
 487 /* LRU pages are freed in batch mode. OSC should at least free this
 488  * number of pages to avoid running out of LRU budget, and..
 489  */
 490 static const int lru_shrink_min = 2 << (20 - PAGE_SHIFT);  /* 2M */
 491 /* free this number at most otherwise it will take too long time to finish. */
 492 static const int lru_shrink_max = 32 << (20 - PAGE_SHIFT); /* 32M */
 493
 494 /* Check if we can free LRU slots from this OSC. If there exists LRU waiters,
 495  * we should free slots aggressively. In this way, slots are freed in a steady
 496  * step to maintain fairness among OSCs.
 497  *
 498  * Return how many LRU pages should be freed.
 499  */
 500 static int osc_cache_too_much(struct client_obd *cli)
 501 {
 502         struct cl_client_cache *cache = cli->cl_cache;
 503         int pages = atomic_read(&cli->cl_lru_in_list) >> 1;
 504
 505         if (atomic_read(&osc_lru_waiters) > 0 &&
 506             atomic_read(cli->cl_lru_left) < lru_shrink_max)
 507                 /* drop lru pages aggressively */
 508                 return min(pages, lru_shrink_max);
 509
 510         /* if it's going to run out LRU slots, we should free some, but not
 511          * too much to maintain fairness among OSCs.
 512          */
 513         if (atomic_read(cli->cl_lru_left) < cache->ccc_lru_max >> 4) {
 514                 unsigned long tmp;
 515
 516                 tmp = cache->ccc_lru_max / atomic_read(&cache->ccc_users);
 517                 if (pages > tmp)
 518                         return min(pages, lru_shrink_max);
 519
 520                 return pages > lru_shrink_min ? lru_shrink_min : 0;
 521         }
 522
 523         return 0;
 524 }
 525
 526 /* Return how many pages are not discarded in @pvec. */
 527 static int discard_pagevec(const struct lu_env *env, struct cl_io *io,
 528                            struct cl_page **pvec, int max_index)
 529 {
 530         int count;
 531         int i;
 532
 533         for (count = 0, i = 0; i < max_index; i++) {
 534                 struct cl_page *page = pvec[i];
 535
 536                 if (cl_page_own_try(env, io, page) == 0) {
 537                         /* free LRU page only if nobody is using it.
 538                          * This check is necessary to avoid freeing the pages
 539                          * having already been removed from LRU and pinned
 540                          * for IO.
 541                          */
 542                         if (!cl_page_in_use(page)) {
 543                                 cl_page_unmap(env, io, page);
 544                                 cl_page_discard(env, io, page);
 545                                 ++count;
 546                         }
 547                         cl_page_disown(env, io, page);
 548                 }
 549                 cl_page_put(env, page);
 550                 pvec[i] = NULL;
 551         }
 552         return max_index - count;
 553 }
 554
 555 /**
 556  * Drop @target of pages from LRU at most.
 557  */
 558 int osc_lru_shrink(struct client_obd *cli, int target)
 559 {
 560         struct cl_env_nest nest;
 561         struct lu_env *env;
 562         struct cl_io *io;
 563         struct cl_object *clobj = NULL;
 564         struct cl_page **pvec;
 565         struct osc_page *opg;
 566         struct osc_page *temp;
 567         int maxscan = 0;
 568         int count = 0;
 569         int index = 0;
 570         int rc = 0;
 571
 572         LASSERT(atomic_read(&cli->cl_lru_in_list) >= 0);
 573         if (atomic_read(&cli->cl_lru_in_list) == 0 || target <= 0)
 574                 return 0;
 575
 576         env = cl_env_nested_get(&nest);
 577         if (IS_ERR(env))
 578                 return PTR_ERR(env);
 579
 580         pvec = osc_env_info(env)->oti_pvec;
 581         io = &osc_env_info(env)->oti_io;
 582
 583         client_obd_list_lock(&cli->cl_lru_list_lock);
 584         atomic_inc(&cli->cl_lru_shrinkers);
 585         maxscan = min(target << 1, atomic_read(&cli->cl_lru_in_list));
 586         list_for_each_entry_safe(opg, temp, &cli->cl_lru_list, ops_lru) {
 587                 struct cl_page *page;
 588
 589                 if (--maxscan < 0)
 590                         break;
 591
 592                 page = cl_page_top(opg->ops_cl.cpl_page);
 593                 if (cl_page_in_use_noref(page)) {
 594                         list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
 595                         continue;
 596                 }
 597
 598                 LASSERT(page->cp_obj);
 599                 if (clobj != page->cp_obj) {
 600                         struct cl_object *tmp = page->cp_obj;
 601
 602                         cl_object_get(tmp);
 603                         client_obd_list_unlock(&cli->cl_lru_list_lock);
 604
 605                         if (clobj) {
 606                                 count -= discard_pagevec(env, io, pvec, index);
 607                                 index = 0;
 608
 609                                 cl_io_fini(env, io);
 610                                 cl_object_put(env, clobj);
 611                                 clobj = NULL;
 612                         }
 613
 614                         clobj = tmp;
 615                         io->ci_obj = clobj;
 616                         io->ci_ignore_layout = 1;
 617                         rc = cl_io_init(env, io, CIT_MISC, clobj);
 618
 619                         client_obd_list_lock(&cli->cl_lru_list_lock);
 620
 621                         if (rc != 0)
 622                                 break;
 623
 624                         ++maxscan;
 625                         continue;
 626                 }
 627
 628                 /* move this page to the end of list as it will be discarded
 629                  * soon. The page will be finally removed from LRU list in
 630                  * osc_page_delete().
 631                  */
 632                 list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
 633
 634                 /* it's okay to grab a refcount here w/o holding lock because
 635                  * it has to grab cl_lru_list_lock to delete the page.
 636                  */
 637                 cl_page_get(page);
 638                 pvec[index++] = page;
 639                 if (++count >= target)
 640                         break;
 641
 642                 if (unlikely(index == OTI_PVEC_SIZE)) {
 643                         client_obd_list_unlock(&cli->cl_lru_list_lock);
 644                         count -= discard_pagevec(env, io, pvec, index);
 645                         index = 0;
 646
 647                         client_obd_list_lock(&cli->cl_lru_list_lock);
 648                 }
 649         }
 650         client_obd_list_unlock(&cli->cl_lru_list_lock);
 651
 652         if (clobj) {
 653                 count -= discard_pagevec(env, io, pvec, index);
 654
 655                 cl_io_fini(env, io);
 656                 cl_object_put(env, clobj);
 657         }
 658         cl_env_nested_put(&nest, env);
 659
 660         atomic_dec(&cli->cl_lru_shrinkers);
 661         return count > 0 ? count : rc;
 662 }
 663
 664 static void osc_lru_add(struct client_obd *cli, struct osc_page *opg)
 665 {
 666         bool wakeup = false;
 667
 668         if (!opg->ops_in_lru)
 669                 return;
 670
 671         atomic_dec(&cli->cl_lru_busy);
 672         client_obd_list_lock(&cli->cl_lru_list_lock);
 673         if (list_empty(&opg->ops_lru)) {
 674                 list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
 675                 atomic_inc_return(&cli->cl_lru_in_list);
 676                 wakeup = atomic_read(&osc_lru_waiters) > 0;
 677         }
 678         client_obd_list_unlock(&cli->cl_lru_list_lock);
 679
 680         if (wakeup) {
 681                 osc_lru_shrink(cli, osc_cache_too_much(cli));
 682                 wake_up_all(&osc_lru_waitq);
 683         }
 684 }
 685
 686 /* delete page from LRUlist. The page can be deleted from LRUlist for two
 687  * reasons: redirtied or deleted from page cache.
 688  */
 689 static void osc_lru_del(struct client_obd *cli, struct osc_page *opg, bool del)
 690 {
 691         if (opg->ops_in_lru) {
 692                 client_obd_list_lock(&cli->cl_lru_list_lock);
 693                 if (!list_empty(&opg->ops_lru)) {
 694                         LASSERT(atomic_read(&cli->cl_lru_in_list) > 0);
 695                         list_del_init(&opg->ops_lru);
 696                         atomic_dec(&cli->cl_lru_in_list);
 697                         if (!del)
 698                                 atomic_inc(&cli->cl_lru_busy);
 699                 } else if (del) {
 700                         LASSERT(atomic_read(&cli->cl_lru_busy) > 0);
 701                         atomic_dec(&cli->cl_lru_busy);
 702                 }
 703                 client_obd_list_unlock(&cli->cl_lru_list_lock);
 704                 if (del) {
 705                         atomic_inc(cli->cl_lru_left);
 706                         /* this is a great place to release more LRU pages if
 707                          * this osc occupies too many LRU pages and kernel is
 708                          * stealing one of them.
 709                          * cl_lru_shrinkers is to avoid recursive call in case
 710                          * we're already in the context of osc_lru_shrink().
 711                          */
 712                         if (atomic_read(&cli->cl_lru_shrinkers) == 0 &&
 713                             !memory_pressure_get())
 714                                 osc_lru_shrink(cli, osc_cache_too_much(cli));
 715                         wake_up(&osc_lru_waitq);
 716                 }
 717         } else {
 718                 LASSERT(list_empty(&opg->ops_lru));
 719         }
 720 }
 721
 722 static inline int max_to_shrink(struct client_obd *cli)
 723 {
 724         return min(atomic_read(&cli->cl_lru_in_list) >> 1, lru_shrink_max);
 725 }
 726
 727 static int osc_lru_reclaim(struct client_obd *cli)
 728 {
 729         struct cl_client_cache *cache = cli->cl_cache;
 730         int max_scans;
 731         int rc;
 732
 733         LASSERT(cache);
 734
 735         rc = osc_lru_shrink(cli, lru_shrink_min);
 736         if (rc != 0) {
 737                 CDEBUG(D_CACHE, "%s: Free %d pages from own LRU: %p.\n",
 738                        cli->cl_import->imp_obd->obd_name, rc, cli);
 739                 return rc;
 740         }
 741
 742         CDEBUG(D_CACHE, "%s: cli %p no free slots, pages: %d, busy: %d.\n",
 743                cli->cl_import->imp_obd->obd_name, cli,
 744                atomic_read(&cli->cl_lru_in_list),
 745                atomic_read(&cli->cl_lru_busy));
 746
 747         /* Reclaim LRU slots from other client_obd as it can't free enough
 748          * from its own. This should rarely happen.
 749          */
 750         spin_lock(&cache->ccc_lru_lock);
 751         LASSERT(!list_empty(&cache->ccc_lru));
 752
 753         cache->ccc_lru_shrinkers++;
 754         list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
 755
 756         max_scans = atomic_read(&cache->ccc_users);
 757         while (--max_scans > 0 && !list_empty(&cache->ccc_lru)) {
 758                 cli = list_entry(cache->ccc_lru.next, struct client_obd,
 759                                  cl_lru_osc);
 760
 761                 CDEBUG(D_CACHE, "%s: cli %p LRU pages: %d, busy: %d.\n",
 762                        cli->cl_import->imp_obd->obd_name, cli,
 763                        atomic_read(&cli->cl_lru_in_list),
 764                        atomic_read(&cli->cl_lru_busy));
 765
 766                 list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
 767                 if (atomic_read(&cli->cl_lru_in_list) > 0) {
 768                         spin_unlock(&cache->ccc_lru_lock);
 769
 770                         rc = osc_lru_shrink(cli, max_to_shrink(cli));
 771                         spin_lock(&cache->ccc_lru_lock);
 772                         if (rc != 0)
 773                                 break;
 774                 }
 775         }
 776         spin_unlock(&cache->ccc_lru_lock);
 777
 778         CDEBUG(D_CACHE, "%s: cli %p freed %d pages.\n",
 779                cli->cl_import->imp_obd->obd_name, cli, rc);
 780         return rc;
 781 }
 782
 783 static int osc_lru_reserve(const struct lu_env *env, struct osc_object *obj,
 784                            struct osc_page *opg)
 785 {
 786         struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
 787         struct client_obd *cli = osc_cli(obj);
 788         int rc = 0;
 789
 790         if (!cli->cl_cache) /* shall not be in LRU */
 791                 return 0;
 792
 793         LASSERT(atomic_read(cli->cl_lru_left) >= 0);
 794         while (!atomic_add_unless(cli->cl_lru_left, -1, 0)) {
 795                 int gen;
 796
 797                 /* run out of LRU spaces, try to drop some by itself */
 798                 rc = osc_lru_reclaim(cli);
 799                 if (rc < 0)
 800                         break;
 801                 if (rc > 0)
 802                         continue;
 803
 804                 cond_resched();
 805
 806                 /* slowest case, all of caching pages are busy, notifying
 807                  * other OSCs that we're lack of LRU slots.
 808                  */
 809                 atomic_inc(&osc_lru_waiters);
 810
 811                 gen = atomic_read(&cli->cl_lru_in_list);
 812                 rc = l_wait_event(osc_lru_waitq,
 813                                   atomic_read(cli->cl_lru_left) > 0 ||
 814                                   (atomic_read(&cli->cl_lru_in_list) > 0 &&
 815                                    gen != atomic_read(&cli->cl_lru_in_list)),
 816                                   &lwi);
 817
 818                 atomic_dec(&osc_lru_waiters);
 819                 if (rc < 0)
 820                         break;
 821         }
 822
 823         if (rc >= 0) {
 824                 atomic_inc(&cli->cl_lru_busy);
 825                 opg->ops_in_lru = 1;
 826                 rc = 0;
 827         }
 828
 829         return rc;
 830 }
 831
 832 /** @} osc */