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  *   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_alloc(const struct lu_env *env, struct client_obd *cli,
  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         union ldlm_policy_data *policy;
  73         enum ldlm_mode          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
 203 int osc_page_cache_add(const struct lu_env *env,
 204                         const struct cl_page_slice *slice, struct cl_io *io)
 205 {
 206         struct osc_page *opg = cl2osc_page(slice);
 207         int result;
 208         ENTRY;
 209
 210         LINVRNT(osc_page_protected(env, opg, CLM_WRITE, 0));
 211
 212         osc_page_transfer_get(opg, "transfer\0cache");
 213         result = osc_queue_async_io(env, io, opg);
 214         if (result != 0)
 215                 osc_page_transfer_put(env, opg);
 216         else
 217                 osc_page_transfer_add(env, opg, CRT_WRITE);
 218
 219         RETURN(result);
 220 }
 221
 222 void osc_index2policy(union ldlm_policy_data *policy,
 223                       const struct cl_object *obj, pgoff_t start, pgoff_t end)
 224 {
 225         memset(policy, 0, sizeof *policy);
 226         policy->l_extent.start = cl_offset(obj, start);
 227         policy->l_extent.end   = cl_offset(obj, end + 1) - 1;
 228 }
 229
 230 static const char *osc_list(struct list_head *head)
 231 {
 232         return list_empty(head) ? "-" : "+";
 233 }
 234
 235 static inline cfs_time_t osc_submit_duration(struct osc_page *opg)
 236 {
 237         if (opg->ops_submit_time == 0)
 238                 return 0;
 239
 240         return (cfs_time_current() - opg->ops_submit_time);
 241 }
 242
 243 static int osc_page_print(const struct lu_env *env,
 244                           const struct cl_page_slice *slice,
 245                           void *cookie, lu_printer_t printer)
 246 {
 247         struct osc_page       *opg = cl2osc_page(slice);
 248         struct osc_async_page *oap = &opg->ops_oap;
 249         struct osc_object     *obj = cl2osc(slice->cpl_obj);
 250         struct client_obd     *cli = &osc_export(obj)->exp_obd->u.cli;
 251
 252         return (*printer)(env, cookie, LUSTRE_OSC_NAME"-page@%p %lu: "
 253                           "1< %#x %d %u %s %s > "
 254                           "2< "LPD64" %u %u %#x %#x | %p %p %p > "
 255                           "3< %d %lu %d > "
 256                           "4< %d %d %d %lu %s | %s %s %s %s > "
 257                           "5< %s %s %s %s | %d %s | %d %s %s>\n",
 258                           opg, osc_index(opg),
 259                           /* 1 */
 260                           oap->oap_magic, oap->oap_cmd,
 261                           oap->oap_interrupted,
 262                           osc_list(&oap->oap_pending_item),
 263                           osc_list(&oap->oap_rpc_item),
 264                           /* 2 */
 265                           oap->oap_obj_off, oap->oap_page_off, oap->oap_count,
 266                           oap->oap_async_flags, oap->oap_brw_flags,
 267                           oap->oap_request, oap->oap_cli, obj,
 268                           /* 3 */
 269                           opg->ops_transfer_pinned,
 270                           osc_submit_duration(opg), opg->ops_srvlock,
 271                           /* 4 */
 272                           cli->cl_r_in_flight, cli->cl_w_in_flight,
 273                           cli->cl_max_rpcs_in_flight,
 274                           cli->cl_avail_grant,
 275                           osc_list(&cli->cl_cache_waiters),
 276                           osc_list(&cli->cl_loi_ready_list),
 277                           osc_list(&cli->cl_loi_hp_ready_list),
 278                           osc_list(&cli->cl_loi_write_list),
 279                           osc_list(&cli->cl_loi_read_list),
 280                           /* 5 */
 281                           osc_list(&obj->oo_ready_item),
 282                           osc_list(&obj->oo_hp_ready_item),
 283                           osc_list(&obj->oo_write_item),
 284                           osc_list(&obj->oo_read_item),
 285                           atomic_read(&obj->oo_nr_reads),
 286                           osc_list(&obj->oo_reading_exts),
 287                           atomic_read(&obj->oo_nr_writes),
 288                           osc_list(&obj->oo_hp_exts),
 289                           osc_list(&obj->oo_urgent_exts));
 290 }
 291
 292 static void osc_page_delete(const struct lu_env *env,
 293                             const struct cl_page_slice *slice)
 294 {
 295         struct osc_page   *opg = cl2osc_page(slice);
 296         struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
 297         int rc;
 298
 299         LINVRNT(opg->ops_temp || osc_page_protected(env, opg, CLM_READ, 1));
 300
 301         ENTRY;
 302         CDEBUG(D_TRACE, "%p\n", opg);
 303         osc_page_transfer_put(env, opg);
 304         rc = osc_teardown_async_page(env, obj, opg);
 305         if (rc) {
 306                 CL_PAGE_DEBUG(D_ERROR, env, slice->cpl_page,
 307                               "Trying to teardown failed: %d\n", rc);
 308                 LASSERT(0);
 309         }
 310
 311         osc_lru_del(osc_cli(obj), opg);
 312
 313         if (slice->cpl_page->cp_type == CPT_CACHEABLE) {
 314                 void *value;
 315
 316                 spin_lock(&obj->oo_tree_lock);
 317                 value = radix_tree_delete(&obj->oo_tree, osc_index(opg));
 318                 if (value != NULL)
 319                         --obj->oo_npages;
 320                 spin_unlock(&obj->oo_tree_lock);
 321
 322                 LASSERT(ergo(value != NULL, value == opg));
 323         }
 324
 325         EXIT;
 326 }
 327
 328 static void osc_page_clip(const struct lu_env *env,
 329                           const struct cl_page_slice *slice,
 330                           int from, int to)
 331 {
 332         struct osc_page       *opg = cl2osc_page(slice);
 333         struct osc_async_page *oap = &opg->ops_oap;
 334
 335         LINVRNT(osc_page_protected(env, opg, CLM_READ, 0));
 336
 337         opg->ops_from = from;
 338         opg->ops_to   = to;
 339         spin_lock(&oap->oap_lock);
 340         oap->oap_async_flags |= ASYNC_COUNT_STABLE;
 341         spin_unlock(&oap->oap_lock);
 342 }
 343
 344 static int osc_page_cancel(const struct lu_env *env,
 345                            const struct cl_page_slice *slice)
 346 {
 347         struct osc_page *opg = cl2osc_page(slice);
 348         int rc = 0;
 349
 350         LINVRNT(osc_page_protected(env, opg, CLM_READ, 0));
 351
 352         /* Check if the transferring against this page
 353          * is completed, or not even queued. */
 354         if (opg->ops_transfer_pinned)
 355                 /* FIXME: may not be interrupted.. */
 356                 rc = osc_cancel_async_page(env, opg);
 357         LASSERT(ergo(rc == 0, opg->ops_transfer_pinned == 0));
 358         return rc;
 359 }
 360
 361 static int osc_page_flush(const struct lu_env *env,
 362                           const struct cl_page_slice *slice,
 363                           struct cl_io *io)
 364 {
 365         struct osc_page *opg = cl2osc_page(slice);
 366         int rc = 0;
 367         ENTRY;
 368         rc = osc_flush_async_page(env, io, opg);
 369         RETURN(rc);
 370 }
 371
 372 static const struct cl_page_operations osc_page_ops = {
 373         .cpo_print         = osc_page_print,
 374         .cpo_delete        = osc_page_delete,
 375         .cpo_clip           = osc_page_clip,
 376         .cpo_cancel         = osc_page_cancel,
 377         .cpo_flush          = osc_page_flush
 378 };
 379
 380 int osc_page_init(const struct lu_env *env, struct cl_object *obj,
 381                   struct cl_page *page, pgoff_t index)
 382 {
 383         struct osc_object *osc = cl2osc(obj);
 384         struct osc_page   *opg = cl_object_page_slice(obj, page);
 385         int result;
 386
 387         opg->ops_from = 0;
 388         opg->ops_to   = PAGE_CACHE_SIZE;
 389
 390         result = osc_prep_async_page(osc, opg, page->cp_vmpage,
 391                                      cl_offset(obj, index));
 392         if (result == 0) {
 393                 struct osc_io *oio = osc_env_io(env);
 394                 opg->ops_srvlock = osc_io_srvlock(oio);
 395                 cl_page_slice_add(page, &opg->ops_cl, obj, index,
 396                                   &osc_page_ops);
 397         }
 398         /*
 399          * Cannot assert osc_page_protected() here as read-ahead
 400          * creates temporary pages outside of a lock.
 401          */
 402 #ifdef CONFIG_LUSTRE_DEBUG_EXPENSIVE_CHECK
 403         opg->ops_temp = !osc_page_protected(env, opg, CLM_READ, 1);
 404 #endif
 405         INIT_LIST_HEAD(&opg->ops_lru);
 406
 407         /* reserve an LRU space for this page */
 408         if (page->cp_type == CPT_CACHEABLE && result == 0) {
 409                 result = osc_lru_alloc(env, osc_cli(osc), opg);
 410                 if (result == 0) {
 411                         spin_lock(&osc->oo_tree_lock);
 412                         result = radix_tree_insert(&osc->oo_tree, index, opg);
 413                         if (result == 0)
 414                                 ++osc->oo_npages;
 415                         spin_unlock(&osc->oo_tree_lock);
 416                         LASSERT(result == 0);
 417                 }
 418         }
 419
 420         return result;
 421 }
 422
 423 /**
 424  * Helper function called by osc_io_submit() for every page in an immediate
 425  * transfer (i.e., transferred synchronously).
 426  */
 427 void osc_page_submit(const struct lu_env *env, struct osc_page *opg,
 428                      enum cl_req_type crt, int brw_flags)
 429 {
 430         struct osc_async_page *oap = &opg->ops_oap;
 431         struct osc_object     *obj = oap->oap_obj;
 432
 433         LINVRNT(osc_page_protected(env, opg,
 434                                    crt == CRT_WRITE ? CLM_WRITE : CLM_READ, 1));
 435
 436         LASSERTF(oap->oap_magic == OAP_MAGIC, "Bad oap magic: oap %p, "
 437                  "magic 0x%x\n", oap, oap->oap_magic);
 438         LASSERT(oap->oap_async_flags & ASYNC_READY);
 439         LASSERT(oap->oap_async_flags & ASYNC_COUNT_STABLE);
 440
 441         oap->oap_cmd       = crt == CRT_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ;
 442         oap->oap_page_off  = opg->ops_from;
 443         oap->oap_count     = opg->ops_to - opg->ops_from;
 444         oap->oap_brw_flags = OBD_BRW_SYNC | brw_flags;
 445
 446         if (!client_is_remote(osc_export(obj)) &&
 447                         cfs_capable(CFS_CAP_SYS_RESOURCE)) {
 448                 oap->oap_brw_flags |= OBD_BRW_NOQUOTA;
 449                 oap->oap_cmd |= OBD_BRW_NOQUOTA;
 450         }
 451
 452         opg->ops_submit_time = cfs_time_current();
 453         osc_page_transfer_get(opg, "transfer\0imm");
 454         osc_page_transfer_add(env, opg, crt);
 455 }
 456
 457 /* --------------- LRU page management ------------------ */
 458
 459 /* OSC is a natural place to manage LRU pages as applications are specialized
 460  * to write OSC by OSC. Ideally, if one OSC is used more frequently it should
 461  * occupy more LRU slots. On the other hand, we should avoid using up all LRU
 462  * slots (client_obd::cl_lru_left) otherwise process has to be put into sleep
 463  * for free LRU slots - this will be very bad so the algorithm requires each
 464  * OSC to free slots voluntarily to maintain a reasonable number of free slots
 465  * at any time.
 466  */
 467
 468 static DECLARE_WAIT_QUEUE_HEAD(osc_lru_waitq);
 469
 470 /**
 471  * LRU pages are freed in batch mode. OSC should at least free this
 472  * number of pages to avoid running out of LRU slots.
 473  */
 474 static inline int lru_shrink_min(struct client_obd *cli)
 475 {
 476         return cli->cl_max_pages_per_rpc * 2;
 477 }
 478
 479 /**
 480  * free this number at most otherwise it will take too long time to finsih.
 481  */
 482 static inline int lru_shrink_max(struct client_obd *cli)
 483 {
 484         return cli->cl_max_pages_per_rpc * cli->cl_max_rpcs_in_flight;
 485 }
 486
 487 /**
 488  * Check if we can free LRU slots from this OSC. If there exists LRU waiters,
 489  * we should free slots aggressively. In this way, slots are freed in a steady
 490  * step to maintain fairness among OSCs.
 491  *
 492  * Return how many LRU pages should be freed.
 493  */
 494 static int osc_cache_too_much(struct client_obd *cli)
 495 {
 496         struct cl_client_cache *cache = cli->cl_cache;
 497         long pages = atomic_long_read(&cli->cl_lru_in_list);
 498         unsigned long budget;
 499
 500         LASSERT(cache != NULL);
 501         budget = cache->ccc_lru_max / (atomic_read(&cache->ccc_users) - 2);
 502
 503         /* if it's going to run out LRU slots, we should free some, but not
 504          * too much to maintain faireness among OSCs. */
 505         if (atomic_long_read(cli->cl_lru_left) < cache->ccc_lru_max >> 2) {
 506                 if (pages >= budget)
 507                         return lru_shrink_max(cli);
 508                 else if (pages >= budget / 2)
 509                         return lru_shrink_min(cli);
 510         } else {
 511                 int duration = cfs_time_current_sec() - cli->cl_lru_last_used;
 512
 513                 /* knock out pages by duration of no IO activity */
 514                 duration >>= 6; /* approximately 1 minute */
 515                 if (duration > 0 && pages >= budget / duration)
 516                         return lru_shrink_min(cli);
 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         int count;
 525
 526         CDEBUG(D_CACHE, "%s: run LRU work for client obd\n", cli_name(cli));
 527         count = osc_cache_too_much(cli);
 528         if (count > 0) {
 529                 int rc = osc_lru_shrink(env, cli, count, false);
 530
 531                 CDEBUG(D_CACHE, "%s: shrank %d/%d pages from client obd\n",
 532                        cli_name(cli), rc, count);
 533                 if (rc >= count) {
 534                         CDEBUG(D_CACHE, "%s: queue again\n", cli_name(cli));
 535                         ptlrpcd_queue_work(cli->cl_lru_work);
 536                 }
 537         }
 538
 539         RETURN(0);
 540 }
 541
 542 void osc_lru_add_batch(struct client_obd *cli, struct list_head *plist)
 543 {
 544         struct list_head lru = LIST_HEAD_INIT(lru);
 545         struct osc_async_page *oap;
 546         long npages = 0;
 547
 548         list_for_each_entry(oap, plist, oap_pending_item) {
 549                 struct osc_page *opg = oap2osc_page(oap);
 550
 551                 if (!opg->ops_in_lru)
 552                         continue;
 553
 554                 ++npages;
 555                 LASSERT(list_empty(&opg->ops_lru));
 556                 list_add(&opg->ops_lru, &lru);
 557         }
 558
 559         if (npages > 0) {
 560                 spin_lock(&cli->cl_lru_list_lock);
 561                 list_splice_tail(&lru, &cli->cl_lru_list);
 562                 atomic_long_sub(npages, &cli->cl_lru_busy);
 563                 atomic_long_add(npages, &cli->cl_lru_in_list);
 564                 cli->cl_lru_last_used = cfs_time_current_sec();
 565                 spin_unlock(&cli->cl_lru_list_lock);
 566
 567                 if (waitqueue_active(&osc_lru_waitq))
 568                         (void)ptlrpcd_queue_work(cli->cl_lru_work);
 569         }
 570 }
 571
 572 static void __osc_lru_del(struct client_obd *cli, struct osc_page *opg)
 573 {
 574         LASSERT(atomic_long_read(&cli->cl_lru_in_list) > 0);
 575         list_del_init(&opg->ops_lru);
 576         atomic_long_dec(&cli->cl_lru_in_list);
 577 }
 578
 579 /**
 580  * Page is being destroyed. The page may be not in LRU list, if the transfer
 581  * has never finished(error occurred).
 582  */
 583 static void osc_lru_del(struct client_obd *cli, struct osc_page *opg)
 584 {
 585         if (opg->ops_in_lru) {
 586                 spin_lock(&cli->cl_lru_list_lock);
 587                 if (!list_empty(&opg->ops_lru)) {
 588                         __osc_lru_del(cli, opg);
 589                 } else {
 590                         LASSERT(atomic_long_read(&cli->cl_lru_busy) > 0);
 591                         atomic_long_dec(&cli->cl_lru_busy);
 592                 }
 593                 spin_unlock(&cli->cl_lru_list_lock);
 594
 595                 atomic_long_inc(cli->cl_lru_left);
 596                 /* this is a great place to release more LRU pages if
 597                  * this osc occupies too many LRU pages and kernel is
 598                  * stealing one of them. */
 599                 if (osc_cache_too_much(cli)) {
 600                         CDEBUG(D_CACHE, "%s: queue LRU work\n", cli_name(cli));
 601                         (void)ptlrpcd_queue_work(cli->cl_lru_work);
 602                 }
 603                 wake_up(&osc_lru_waitq);
 604         } else {
 605                 LASSERT(list_empty(&opg->ops_lru));
 606         }
 607 }
 608
 609 /**
 610  * Delete page from LRUlist for redirty.
 611  */
 612 static void osc_lru_use(struct client_obd *cli, struct osc_page *opg)
 613 {
 614         /* If page is being transferred for the first time,
 615          * ops_lru should be empty */
 616         if (opg->ops_in_lru && !list_empty(&opg->ops_lru)) {
 617                 spin_lock(&cli->cl_lru_list_lock);
 618                 __osc_lru_del(cli, opg);
 619                 spin_unlock(&cli->cl_lru_list_lock);
 620                 atomic_long_inc(&cli->cl_lru_busy);
 621         }
 622 }
 623
 624 static void discard_pagevec(const struct lu_env *env, struct cl_io *io,
 625                                 struct cl_page **pvec, int max_index)
 626 {
 627         int i;
 628
 629         for (i = 0; i < max_index; i++) {
 630                 struct cl_page *page = pvec[i];
 631
 632                 LASSERT(cl_page_is_owned(page, io));
 633                 cl_page_delete(env, page);
 634                 cl_page_discard(env, io, page);
 635                 cl_page_disown(env, io, page);
 636                 cl_page_put(env, page);
 637
 638                 pvec[i] = NULL;
 639         }
 640 }
 641
 642 /**
 643  * Check if a cl_page can be released, i.e, it's not being used.
 644  *
 645  * If unstable account is turned on, bulk transfer may hold one refcount
 646  * for recovery so we need to check vmpage refcount as well; otherwise,
 647  * even we can destroy cl_page but the corresponding vmpage can't be reused.
 648  */
 649 static inline bool lru_page_busy(struct client_obd *cli, struct cl_page *page)
 650 {
 651         if (cl_page_in_use_noref(page))
 652                 return true;
 653
 654         if (cli->cl_cache->ccc_unstable_check) {
 655                 struct page *vmpage = cl_page_vmpage(page);
 656
 657                 /* vmpage have two known users: cl_page and VM page cache */
 658                 if (page_count(vmpage) - page_mapcount(vmpage) > 2)
 659                         return true;
 660         }
 661         return false;
 662 }
 663
 664 /**
 665  * Drop @target of pages from LRU at most.
 666  */
 667 long osc_lru_shrink(const struct lu_env *env, struct client_obd *cli,
 668                    long target, bool force)
 669 {
 670         struct cl_io *io;
 671         struct cl_object *clobj = NULL;
 672         struct cl_page **pvec;
 673         struct osc_page *opg;
 674         long count = 0;
 675         int maxscan = 0;
 676         int index = 0;
 677         int rc = 0;
 678         ENTRY;
 679
 680         LASSERT(atomic_long_read(&cli->cl_lru_in_list) >= 0);
 681         if (atomic_long_read(&cli->cl_lru_in_list) == 0 || target <= 0)
 682                 RETURN(0);
 683
 684         CDEBUG(D_CACHE, "%s: shrinkers: %d, force: %d\n",
 685                cli_name(cli), atomic_read(&cli->cl_lru_shrinkers), force);
 686         if (!force) {
 687                 if (atomic_read(&cli->cl_lru_shrinkers) > 0)
 688                         RETURN(-EBUSY);
 689
 690                 if (atomic_inc_return(&cli->cl_lru_shrinkers) > 1) {
 691                         atomic_dec(&cli->cl_lru_shrinkers);
 692                         RETURN(-EBUSY);
 693                 }
 694         } else {
 695                 atomic_inc(&cli->cl_lru_shrinkers);
 696         }
 697
 698         pvec = (struct cl_page **)osc_env_info(env)->oti_pvec;
 699         io = &osc_env_info(env)->oti_io;
 700
 701         spin_lock(&cli->cl_lru_list_lock);
 702         if (force)
 703                 cli->cl_lru_reclaim++;
 704         maxscan = min(target << 1, atomic_long_read(&cli->cl_lru_in_list));
 705         while (!list_empty(&cli->cl_lru_list)) {
 706                 struct cl_page *page;
 707                 bool will_free = false;
 708
 709                 if (!force && atomic_read(&cli->cl_lru_shrinkers) > 1)
 710                         break;
 711
 712                 if (--maxscan < 0)
 713                         break;
 714
 715                 opg = list_entry(cli->cl_lru_list.next, struct osc_page,
 716                                  ops_lru);
 717                 page = opg->ops_cl.cpl_page;
 718                 if (lru_page_busy(cli, page)) {
 719                         list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
 720                         continue;
 721                 }
 722
 723                 LASSERT(page->cp_obj != NULL);
 724                 if (clobj != page->cp_obj) {
 725                         struct cl_object *tmp = page->cp_obj;
 726
 727                         cl_object_get(tmp);
 728                         spin_unlock(&cli->cl_lru_list_lock);
 729
 730                         if (clobj != NULL) {
 731                                 discard_pagevec(env, io, pvec, index);
 732                                 index = 0;
 733
 734                                 cl_io_fini(env, io);
 735                                 cl_object_put(env, clobj);
 736                                 clobj = NULL;
 737                         }
 738
 739                         clobj = tmp;
 740                         io->ci_obj = clobj;
 741                         io->ci_ignore_layout = 1;
 742                         rc = cl_io_init(env, io, CIT_MISC, clobj);
 743
 744                         spin_lock(&cli->cl_lru_list_lock);
 745
 746                         if (rc != 0)
 747                                 break;
 748
 749                         ++maxscan;
 750                         continue;
 751                 }
 752
 753                 if (cl_page_own_try(env, io, page) == 0) {
 754                         if (!lru_page_busy(cli, page)) {
 755                                 /* remove it from lru list earlier to avoid
 756                                  * lock contention */
 757                                 __osc_lru_del(cli, opg);
 758                                 opg->ops_in_lru = 0; /* will be discarded */
 759
 760                                 cl_page_get(page);
 761                                 will_free = true;
 762                         } else {
 763                                 cl_page_disown(env, io, page);
 764                         }
 765                 }
 766
 767                 if (!will_free) {
 768                         list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
 769                         continue;
 770                 }
 771
 772                 /* Don't discard and free the page with cl_lru_list held */
 773                 pvec[index++] = page;
 774                 if (unlikely(index == OTI_PVEC_SIZE)) {
 775                         spin_unlock(&cli->cl_lru_list_lock);
 776                         discard_pagevec(env, io, pvec, index);
 777                         index = 0;
 778
 779                         spin_lock(&cli->cl_lru_list_lock);
 780                 }
 781
 782                 if (++count >= target)
 783                         break;
 784         }
 785         spin_unlock(&cli->cl_lru_list_lock);
 786
 787         if (clobj != NULL) {
 788                 discard_pagevec(env, io, pvec, index);
 789
 790                 cl_io_fini(env, io);
 791                 cl_object_put(env, clobj);
 792         }
 793
 794         atomic_dec(&cli->cl_lru_shrinkers);
 795         if (count > 0) {
 796                 atomic_long_add(count, cli->cl_lru_left);
 797                 wake_up_all(&osc_lru_waitq);
 798         }
 799         RETURN(count > 0 ? count : rc);
 800 }
 801
 802 /**
 803  * Reclaim LRU pages by an IO thread. The caller wants to reclaim at least
 804  * \@npages of LRU slots. For performance consideration, it's better to drop
 805  * LRU pages in batch. Therefore, the actual number is adjusted at least
 806  * max_pages_per_rpc.
 807  */
 808 static long osc_lru_reclaim(struct client_obd *cli, unsigned long npages)
 809 {
 810         struct lu_env *env;
 811         struct cl_client_cache *cache = cli->cl_cache;
 812         int max_scans;
 813         __u16 refcheck;
 814         long rc = 0;
 815         ENTRY;
 816
 817         LASSERT(cache != NULL);
 818
 819         env = cl_env_get(&refcheck);
 820         if (IS_ERR(env))
 821                 RETURN(rc);
 822
 823         npages = max_t(int, npages, cli->cl_max_pages_per_rpc);
 824         CDEBUG(D_CACHE, "%s: start to reclaim %ld pages from LRU\n",
 825                cli_name(cli), npages);
 826         rc = osc_lru_shrink(env, cli, npages, true);
 827         if (rc >= npages) {
 828                 CDEBUG(D_CACHE, "%s: reclaimed %ld/%ld pages from LRU\n",
 829                        cli_name(cli), rc, npages);
 830                 if (osc_cache_too_much(cli) > 0)
 831                         ptlrpcd_queue_work(cli->cl_lru_work);
 832                 GOTO(out, rc);
 833         } else if (rc > 0) {
 834                 npages -= rc;
 835         }
 836
 837         CDEBUG(D_CACHE, "%s: cli %p no free slots, pages: %ld/%ld, want: %ld\n",
 838                 cli_name(cli), cli, atomic_long_read(&cli->cl_lru_in_list),
 839                 atomic_long_read(&cli->cl_lru_busy), npages);
 840
 841         /* Reclaim LRU slots from other client_obd as it can't free enough
 842          * from its own. This should rarely happen. */
 843         spin_lock(&cache->ccc_lru_lock);
 844         LASSERT(!list_empty(&cache->ccc_lru));
 845
 846         cache->ccc_lru_shrinkers++;
 847         list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
 848
 849         max_scans = atomic_read(&cache->ccc_users) - 2;
 850         while (--max_scans > 0 && !list_empty(&cache->ccc_lru)) {
 851                 cli = list_entry(cache->ccc_lru.next, struct client_obd,
 852                                  cl_lru_osc);
 853
 854                 CDEBUG(D_CACHE, "%s: cli %p LRU pages: %ld, busy: %ld.\n",
 855                         cli_name(cli), cli,
 856                         atomic_long_read(&cli->cl_lru_in_list),
 857                         atomic_long_read(&cli->cl_lru_busy));
 858
 859                 list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
 860                 if (osc_cache_too_much(cli) > 0) {
 861                         spin_unlock(&cache->ccc_lru_lock);
 862
 863                         rc = osc_lru_shrink(env, cli, npages, true);
 864                         spin_lock(&cache->ccc_lru_lock);
 865                         if (rc >= npages)
 866                                 break;
 867                         if (rc > 0)
 868                                 npages -= rc;
 869                 }
 870         }
 871         spin_unlock(&cache->ccc_lru_lock);
 872
 873 out:
 874         cl_env_put(env, &refcheck);
 875         CDEBUG(D_CACHE, "%s: cli %p freed %ld pages.\n",
 876                 cli_name(cli), cli, rc);
 877         return rc;
 878 }
 879
 880 /**
 881  * osc_lru_alloc() is called to allocate an LRU slot for a cl_page.
 882  *
 883  * Usually the LRU slots are reserved in osc_io_iter_rw_init().
 884  * Only in the case that the LRU slots are in extreme shortage, it should
 885  * have reserved enough slots for an IO.
 886  */
 887 static int osc_lru_alloc(const struct lu_env *env, struct client_obd *cli,
 888                          struct osc_page *opg)
 889 {
 890         struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
 891         struct osc_io *oio = osc_env_io(env);
 892         int rc = 0;
 893         ENTRY;
 894
 895         if (cli->cl_cache == NULL) /* shall not be in LRU */
 896                 RETURN(0);
 897
 898         if (oio->oi_lru_reserved > 0) {
 899                 --oio->oi_lru_reserved;
 900                 goto out;
 901         }
 902
 903         LASSERT(atomic_long_read(cli->cl_lru_left) >= 0);
 904         while (!atomic_long_add_unless(cli->cl_lru_left, -1, 0)) {
 905                 /* run out of LRU spaces, try to drop some by itself */
 906                 rc = osc_lru_reclaim(cli, 1);
 907                 if (rc < 0)
 908                         break;
 909                 if (rc > 0)
 910                         continue;
 911
 912                 cond_resched();
 913                 rc = l_wait_event(osc_lru_waitq,
 914                                 atomic_long_read(cli->cl_lru_left) > 0,
 915                                 &lwi);
 916                 if (rc < 0)
 917                         break;
 918         }
 919
 920 out:
 921         if (rc >= 0) {
 922                 atomic_long_inc(&cli->cl_lru_busy);
 923                 opg->ops_in_lru = 1;
 924                 rc = 0;
 925         }
 926
 927         RETURN(rc);
 928 }
 929
 930 /**
 931  * osc_lru_reserve() is called to reserve enough LRU slots for I/O.
 932  *
 933  * The benefit of doing this is to reduce contention against atomic counter
 934  * cl_lru_left by changing it from per-page access to per-IO access.
 935  */
 936 unsigned long osc_lru_reserve(struct client_obd *cli, unsigned long npages)
 937 {
 938         unsigned long reserved = 0;
 939         unsigned long max_pages;
 940         unsigned long c;
 941
 942         /* reserve a full RPC window at most to avoid that a thread accidentally
 943          * consumes too many LRU slots */
 944         max_pages = cli->cl_max_pages_per_rpc * cli->cl_max_rpcs_in_flight;
 945         if (npages > max_pages)
 946                 npages = max_pages;
 947
 948         c = atomic_long_read(cli->cl_lru_left);
 949         if (c < npages && osc_lru_reclaim(cli, npages) > 0)
 950                 c = atomic_long_read(cli->cl_lru_left);
 951         while (c >= npages) {
 952                 if (c == atomic_long_cmpxchg(cli->cl_lru_left, c, c - npages)) {
 953                         reserved = npages;
 954                         break;
 955                 }
 956                 c = atomic_long_read(cli->cl_lru_left);
 957         }
 958         if (atomic_long_read(cli->cl_lru_left) < max_pages) {
 959                 /* If there aren't enough pages in the per-OSC LRU then
 960                  * wake up the LRU thread to try and clear out space, so
 961                  * we don't block if pages are being dirtied quickly. */
 962                 CDEBUG(D_CACHE, "%s: queue LRU, left: %lu/%ld.\n",
 963                        cli_name(cli), atomic_long_read(cli->cl_lru_left),
 964                        max_pages);
 965                 (void)ptlrpcd_queue_work(cli->cl_lru_work);
 966         }
 967
 968         return reserved;
 969 }
 970
 971 /**
 972  * osc_lru_unreserve() is called to unreserve LRU slots.
 973  *
 974  * LRU slots reserved by osc_lru_reserve() may have entries left due to several
 975  * reasons such as page already existing or I/O error. Those reserved slots
 976  * should be freed by calling this function.
 977  */
 978 void osc_lru_unreserve(struct client_obd *cli, unsigned long npages)
 979 {
 980         atomic_long_add(npages, cli->cl_lru_left);
 981         wake_up_all(&osc_lru_waitq);
 982 }
 983
 984 /**
 985  * Atomic operations are expensive. We accumulate the accounting for the
 986  * same page zone to get better performance.
 987  * In practice this can work pretty good because the pages in the same RPC
 988  * are likely from the same page zone.
 989  */
 990 static inline void unstable_page_accounting(struct ptlrpc_bulk_desc *desc,
 991                                             int factor)
 992 {
 993         int page_count = desc->bd_iov_count;
 994         void *zone = NULL;
 995         int count = 0;
 996         int i;
 997
 998         LASSERT(ptlrpc_is_bulk_desc_kiov(desc->bd_type));
 999
1000         for (i = 0; i < page_count; i++) {
1001                 void *pz = page_zone(BD_GET_KIOV(desc, i).kiov_page);
1002
1003                 if (likely(pz == zone)) {
1004                         ++count;
1005                         continue;
1006                 }
1007
1008                 if (count > 0) {
1009                         mod_zone_page_state(zone, NR_UNSTABLE_NFS,
1010                                             factor * count);
1011                         count = 0;
1012                 }
1013                 zone = pz;
1014                 ++count;
1015         }
1016         if (count > 0)
1017                 mod_zone_page_state(zone, NR_UNSTABLE_NFS, factor * count);
1018 }
1019
1020 static inline void add_unstable_page_accounting(struct ptlrpc_bulk_desc *desc)
1021 {
1022         unstable_page_accounting(desc, 1);
1023 }
1024
1025 static inline void dec_unstable_page_accounting(struct ptlrpc_bulk_desc *desc)
1026 {
1027         unstable_page_accounting(desc, -1);
1028 }
1029
1030 /**
1031  * Performs "unstable" page accounting. This function balances the
1032  * increment operations performed in osc_inc_unstable_pages. It is
1033  * registered as the RPC request callback, and is executed when the
1034  * bulk RPC is committed on the server. Thus at this point, the pages
1035  * involved in the bulk transfer are no longer considered unstable.
1036  *
1037  * If this function is called, the request should have been committed
1038  * or req:rq_unstable must have been set; it implies that the unstable
1039  * statistic have been added.
1040  */
1041 void osc_dec_unstable_pages(struct ptlrpc_request *req)
1042 {
1043         struct ptlrpc_bulk_desc *desc       = req->rq_bulk;
1044         struct client_obd       *cli        = &req->rq_import->imp_obd->u.cli;
1045         int                      page_count = desc->bd_iov_count;
1046         long                     unstable_count;
1047
1048         LASSERT(page_count >= 0);
1049         dec_unstable_page_accounting(desc);
1050
1051         unstable_count = atomic_long_sub_return(page_count,
1052                                                 &cli->cl_unstable_count);
1053         LASSERT(unstable_count >= 0);
1054
1055         unstable_count = atomic_long_sub_return(page_count,
1056                                            &cli->cl_cache->ccc_unstable_nr);
1057         LASSERT(unstable_count >= 0);
1058         if (unstable_count == 0)
1059                 wake_up_all(&cli->cl_cache->ccc_unstable_waitq);
1060
1061         if (waitqueue_active(&osc_lru_waitq))
1062                 (void)ptlrpcd_queue_work(cli->cl_lru_work);
1063 }
1064
1065 /**
1066  * "unstable" page accounting. See: osc_dec_unstable_pages.
1067  */
1068 void osc_inc_unstable_pages(struct ptlrpc_request *req)
1069 {
1070         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
1071         struct client_obd       *cli  = &req->rq_import->imp_obd->u.cli;
1072         long                     page_count = desc->bd_iov_count;
1073
1074         /* No unstable page tracking */
1075         if (cli->cl_cache == NULL || !cli->cl_cache->ccc_unstable_check)
1076                 return;
1077
1078         add_unstable_page_accounting(desc);
1079         atomic_long_add(page_count, &cli->cl_unstable_count);
1080         atomic_long_add(page_count, &cli->cl_cache->ccc_unstable_nr);
1081
1082         /* If the request has already been committed (i.e. brw_commit
1083          * called via rq_commit_cb), we need to undo the unstable page
1084          * increments we just performed because rq_commit_cb wont be
1085          * called again. */
1086         spin_lock(&req->rq_lock);
1087         if (unlikely(req->rq_committed)) {
1088                 spin_unlock(&req->rq_lock);
1089
1090                 osc_dec_unstable_pages(req);
1091         } else {
1092                 req->rq_unstable = 1;
1093                 spin_unlock(&req->rq_lock);
1094         }
1095 }
1096
1097 /**
1098  * Check if it piggybacks SOFT_SYNC flag to OST from this OSC.
1099  * This function will be called by every BRW RPC so it's critical
1100  * to make this function fast.
1101  */
1102 bool osc_over_unstable_soft_limit(struct client_obd *cli)
1103 {
1104         long unstable_nr, osc_unstable_count;
1105
1106         /* Can't check cli->cl_unstable_count, therefore, no soft limit */
1107         if (cli->cl_cache == NULL || !cli->cl_cache->ccc_unstable_check)
1108                 return false;
1109
1110         osc_unstable_count = atomic_long_read(&cli->cl_unstable_count);
1111         unstable_nr = atomic_long_read(&cli->cl_cache->ccc_unstable_nr);
1112
1113         CDEBUG(D_CACHE,
1114                "%s: cli: %p unstable pages: %lu, osc unstable pages: %lu\n",
1115                cli_name(cli), cli, unstable_nr, osc_unstable_count);
1116
1117         /* If the LRU slots are in shortage - 25% remaining AND this OSC
1118          * has one full RPC window of unstable pages, it's a good chance
1119          * to piggyback a SOFT_SYNC flag.
1120          * Please notice that the OST won't take immediate response for the
1121          * SOFT_SYNC request so active OSCs will have more chance to carry
1122          * the flag, this is reasonable. */
1123         return unstable_nr > cli->cl_cache->ccc_lru_max >> 2 &&
1124                osc_unstable_count > cli->cl_max_pages_per_rpc *
1125                                     cli->cl_max_rpcs_in_flight;
1126 }
1127
1128 /**
1129  * Return how many LRU pages in the cache of all OSC devices
1130  *
1131  * \retval      return # of cached LRU pages times reclaimation tendency
1132  * \retval      SHRINK_STOP if it cannot do any scanning in this time
1133  */
1134 unsigned long osc_cache_shrink_count(struct shrinker *sk,
1135                                      struct shrink_control *sc)
1136 {
1137         struct client_obd *cli;
1138         unsigned long cached = 0;
1139
1140         spin_lock(&osc_shrink_lock);
1141         list_for_each_entry(cli, &osc_shrink_list, cl_shrink_list)
1142                 cached += atomic_long_read(&cli->cl_lru_in_list);
1143         spin_unlock(&osc_shrink_lock);
1144
1145         return (cached  * sysctl_vfs_cache_pressure) / 100;
1146 }
1147
1148 /**
1149  * Scan and try to reclaim sc->nr_to_scan cached LRU pages
1150  *
1151  * \retval      number of cached LRU pages reclaimed
1152  * \retval      SHRINK_STOP if it cannot do any scanning in this time
1153  *
1154  * Linux kernel will loop calling this shrinker scan routine with
1155  * sc->nr_to_scan = SHRINK_BATCH(128 for now) until kernel got enough memory.
1156  *
1157  * If sc->nr_to_scan is 0, the VM is querying the cache size, we don't need
1158  * to scan and try to reclaim LRU pages, just return 0 and
1159  * osc_cache_shrink_count() will report the LRU page number.
1160  */
1161 unsigned long osc_cache_shrink_scan(struct shrinker *sk,
1162                                     struct shrink_control *sc)
1163 {
1164         struct client_obd *cli;
1165         struct client_obd *stop_anchor = NULL;
1166         struct lu_env *env;
1167         long shrank = 0;
1168         int rc;
1169         __u16 refcheck;
1170
1171         if (sc->nr_to_scan == 0)
1172                 return 0;
1173
1174         if (!(sc->gfp_mask & __GFP_FS))
1175                 return SHRINK_STOP;
1176
1177         env = cl_env_get(&refcheck);
1178         if (IS_ERR(env))
1179                 return SHRINK_STOP;
1180
1181         spin_lock(&osc_shrink_lock);
1182         while (!list_empty(&osc_shrink_list)) {
1183                 cli = list_entry(osc_shrink_list.next, struct client_obd,
1184                                  cl_shrink_list);
1185
1186                 if (stop_anchor == NULL)
1187                         stop_anchor = cli;
1188                 else if (cli == stop_anchor)
1189                         break;
1190
1191                 list_move_tail(&cli->cl_shrink_list, &osc_shrink_list);
1192                 spin_unlock(&osc_shrink_lock);
1193
1194                 /* shrink no more than max_pages_per_rpc for an OSC */
1195                 rc = osc_lru_shrink(env, cli, (sc->nr_to_scan - shrank) >
1196                                     cli->cl_max_pages_per_rpc ?
1197                                     cli->cl_max_pages_per_rpc :
1198                                     sc->nr_to_scan - shrank, true);
1199                 if (rc > 0)
1200                         shrank += rc;
1201
1202                 if (shrank >= sc->nr_to_scan)
1203                         goto out;
1204
1205                 spin_lock(&osc_shrink_lock);
1206         }
1207         spin_unlock(&osc_shrink_lock);
1208
1209 out:
1210         cl_env_put(env, &refcheck);
1211
1212         return shrank;
1213 }
1214
1215 /** @} osc */