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, 2012, 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 /*
  55  * Comment out osc_page_protected because it may sleep inside the
  56  * the client_obd_list_lock.
  57  * client_obd_list_lock -> osc_ap_completion -> osc_completion ->
  58  *   -> osc_page_protected -> osc_page_is_dlocked -> osc_match_base
  59  *   -> ldlm_lock_match -> sptlrpc_import_check_ctx -> sleep.
  60  */
  61 #if 0
  62 static int osc_page_is_dlocked(const struct lu_env *env,
  63                                const struct osc_page *opg,
  64                                enum cl_lock_mode mode, int pending, int unref)
  65 {
  66         struct cl_page         *page;
  67         struct osc_object      *obj;
  68         struct osc_thread_info *info;
  69         struct ldlm_res_id     *resname;
  70         struct lustre_handle   *lockh;
  71         ldlm_policy_data_t     *policy;
  72         ldlm_mode_t             dlmmode;
  73         int                     flags;
  74
  75         cfs_might_sleep();
  76
  77         info = osc_env_info(env);
  78         resname = &info->oti_resname;
  79         policy = &info->oti_policy;
  80         lockh = &info->oti_handle;
  81         page = opg->ops_cl.cpl_page;
  82         obj = cl2osc(opg->ops_cl.cpl_obj);
  83
  84         flags = LDLM_FL_TEST_LOCK | LDLM_FL_BLOCK_GRANTED;
  85         if (pending)
  86                 flags |= LDLM_FL_CBPENDING;
  87
  88         dlmmode = osc_cl_lock2ldlm(mode) | LCK_PW;
  89         osc_lock_build_res(env, obj, resname);
  90         osc_index2policy(policy, page->cp_obj, page->cp_index, page->cp_index);
  91         return osc_match_base(osc_export(obj), resname, LDLM_EXTENT, policy,
  92                               dlmmode, &flags, NULL, lockh, unref);
  93 }
  94
  95 /**
  96  * Checks an invariant that a page in the cache is covered by a lock, as
  97  * needed.
  98  */
  99 static int osc_page_protected(const struct lu_env *env,
 100                               const struct osc_page *opg,
 101                               enum cl_lock_mode mode, int unref)
 102 {
 103         struct cl_object_header *hdr;
 104         struct cl_lock          *scan;
 105         struct cl_page          *page;
 106         struct cl_lock_descr    *descr;
 107         int result;
 108
 109         LINVRNT(!opg->ops_temp);
 110
 111         page = opg->ops_cl.cpl_page;
 112         if (page->cp_owner != NULL &&
 113             cl_io_top(page->cp_owner)->ci_lockreq == CILR_NEVER)
 114                 /*
 115                  * If IO is done without locks (liblustre, or lloop), lock is
 116                  * not required.
 117                  */
 118                 result = 1;
 119         else
 120                 /* otherwise check for a DLM lock */
 121         result = osc_page_is_dlocked(env, opg, mode, 1, unref);
 122         if (result == 0) {
 123                 /* maybe this page is a part of a lockless io? */
 124                 hdr = cl_object_header(opg->ops_cl.cpl_obj);
 125                 descr = &osc_env_info(env)->oti_descr;
 126                 descr->cld_mode = mode;
 127                 descr->cld_start = page->cp_index;
 128                 descr->cld_end   = page->cp_index;
 129                 spin_lock(&hdr->coh_lock_guard);
 130                 cfs_list_for_each_entry(scan, &hdr->coh_locks, cll_linkage) {
 131                         /*
 132                          * Lock-less sub-lock has to be either in HELD state
 133                          * (when io is actively going on), or in CACHED state,
 134                          * when top-lock is being unlocked:
 135                          * cl_io_unlock()->cl_unuse()->...->lov_lock_unuse().
 136                          */
 137                         if ((scan->cll_state == CLS_HELD ||
 138                              scan->cll_state == CLS_CACHED) &&
 139                             cl_lock_ext_match(&scan->cll_descr, descr)) {
 140                                 struct osc_lock *olck;
 141
 142                                 olck = osc_lock_at(scan);
 143                                 result = osc_lock_is_lockless(olck);
 144                                 break;
 145                         }
 146                 }
 147                 spin_unlock(&hdr->coh_lock_guard);
 148         }
 149         return result;
 150 }
 151 #else
 152 static int osc_page_protected(const struct lu_env *env,
 153                               const struct osc_page *opg,
 154                               enum cl_lock_mode mode, int unref)
 155 {
 156         return 1;
 157 }
 158 #endif
 159
 160 /*****************************************************************************
 161  *
 162  * Page operations.
 163  *
 164  */
 165 static void osc_page_fini(const struct lu_env *env,
 166                           struct cl_page_slice *slice)
 167 {
 168         struct osc_page *opg = cl2osc_page(slice);
 169         CDEBUG(D_TRACE, "%p\n", opg);
 170         LASSERT(opg->ops_lock == NULL);
 171 }
 172
 173 static void osc_page_transfer_get(struct osc_page *opg, const char *label)
 174 {
 175         struct cl_page *page = cl_page_top(opg->ops_cl.cpl_page);
 176
 177         LASSERT(!opg->ops_transfer_pinned);
 178         cl_page_get(page);
 179         lu_ref_add_atomic(&page->cp_reference, label, page);
 180         opg->ops_transfer_pinned = 1;
 181 }
 182
 183 static void osc_page_transfer_put(const struct lu_env *env,
 184                                   struct osc_page *opg)
 185 {
 186         struct cl_page *page = cl_page_top(opg->ops_cl.cpl_page);
 187
 188         if (opg->ops_transfer_pinned) {
 189                 lu_ref_del(&page->cp_reference, "transfer", page);
 190                 opg->ops_transfer_pinned = 0;
 191                 cl_page_put(env, page);
 192         }
 193 }
 194
 195 /**
 196  * This is called once for every page when it is submitted for a transfer
 197  * either opportunistic (osc_page_cache_add()), or immediate
 198  * (osc_page_submit()).
 199  */
 200 static void osc_page_transfer_add(const struct lu_env *env,
 201                                   struct osc_page *opg, enum cl_req_type crt)
 202 {
 203         struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
 204
 205         /* ops_lru and ops_inflight share the same field, so take it from LRU
 206          * first and then use it as inflight. */
 207         osc_lru_del(osc_cli(obj), opg, false);
 208
 209         spin_lock(&obj->oo_seatbelt);
 210         cfs_list_add(&opg->ops_inflight, &obj->oo_inflight[crt]);
 211         opg->ops_submitter = cfs_current();
 212         spin_unlock(&obj->oo_seatbelt);
 213 }
 214
 215 static int osc_page_cache_add(const struct lu_env *env,
 216                               const struct cl_page_slice *slice,
 217                               struct cl_io *io)
 218 {
 219         struct osc_io   *oio = osc_env_io(env);
 220         struct osc_page *opg = cl2osc_page(slice);
 221         int result;
 222         ENTRY;
 223
 224         LINVRNT(osc_page_protected(env, opg, CLM_WRITE, 0));
 225
 226         osc_page_transfer_get(opg, "transfer\0cache");
 227         result = osc_queue_async_io(env, io, opg);
 228         if (result != 0)
 229                 osc_page_transfer_put(env, opg);
 230         else
 231                 osc_page_transfer_add(env, opg, CRT_WRITE);
 232
 233         /* for sync write, kernel will wait for this page to be flushed before
 234          * osc_io_end() is called, so release it earlier.
 235          * for mkwrite(), it's known there is no further pages. */
 236         if (cl_io_is_sync_write(io) || cl_io_is_mkwrite(io)) {
 237                 if (oio->oi_active != NULL) {
 238                         osc_extent_release(env, oio->oi_active);
 239                         oio->oi_active = NULL;
 240                 }
 241         }
 242
 243         RETURN(result);
 244 }
 245
 246 void osc_index2policy(ldlm_policy_data_t *policy, const struct cl_object *obj,
 247                       pgoff_t start, pgoff_t end)
 248 {
 249         memset(policy, 0, sizeof *policy);
 250         policy->l_extent.start = cl_offset(obj, start);
 251         policy->l_extent.end   = cl_offset(obj, end + 1) - 1;
 252 }
 253
 254 static int osc_page_addref_lock(const struct lu_env *env,
 255                                 struct osc_page *opg,
 256                                 struct cl_lock *lock)
 257 {
 258         struct osc_lock *olock;
 259         int              rc;
 260
 261         LASSERT(opg->ops_lock == NULL);
 262
 263         olock = osc_lock_at(lock);
 264         if (cfs_atomic_inc_return(&olock->ols_pageref) <= 0) {
 265                 cfs_atomic_dec(&olock->ols_pageref);
 266                 rc = -ENODATA;
 267         } else {
 268                 cl_lock_get(lock);
 269                 opg->ops_lock = lock;
 270                 rc = 0;
 271         }
 272         return rc;
 273 }
 274
 275 static void osc_page_putref_lock(const struct lu_env *env,
 276                                  struct osc_page *opg)
 277 {
 278         struct cl_lock  *lock = opg->ops_lock;
 279         struct osc_lock *olock;
 280
 281         LASSERT(lock != NULL);
 282         olock = osc_lock_at(lock);
 283
 284         cfs_atomic_dec(&olock->ols_pageref);
 285         opg->ops_lock = NULL;
 286
 287         cl_lock_put(env, lock);
 288 }
 289
 290 static int osc_page_is_under_lock(const struct lu_env *env,
 291                                   const struct cl_page_slice *slice,
 292                                   struct cl_io *unused)
 293 {
 294         struct cl_lock *lock;
 295         int             result = -ENODATA;
 296
 297         ENTRY;
 298         lock = cl_lock_at_page(env, slice->cpl_obj, slice->cpl_page,
 299                                NULL, 1, 0);
 300         if (lock != NULL) {
 301                 if (osc_page_addref_lock(env, cl2osc_page(slice), lock) == 0)
 302                         result = -EBUSY;
 303                 cl_lock_put(env, lock);
 304         }
 305         RETURN(result);
 306 }
 307
 308 static void osc_page_disown(const struct lu_env *env,
 309                             const struct cl_page_slice *slice,
 310                             struct cl_io *io)
 311 {
 312         struct osc_page *opg = cl2osc_page(slice);
 313
 314         if (unlikely(opg->ops_lock))
 315                 osc_page_putref_lock(env, opg);
 316 }
 317
 318 static void osc_page_completion_read(const struct lu_env *env,
 319                                      const struct cl_page_slice *slice,
 320                                      int ioret)
 321 {
 322         struct osc_page   *opg = cl2osc_page(slice);
 323         struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
 324
 325         if (likely(opg->ops_lock))
 326                 osc_page_putref_lock(env, opg);
 327         osc_lru_add(osc_cli(obj), opg);
 328 }
 329
 330 static void osc_page_completion_write(const struct lu_env *env,
 331                                       const struct cl_page_slice *slice,
 332                                       int ioret)
 333 {
 334         struct osc_page   *opg = cl2osc_page(slice);
 335         struct osc_object *obj = cl2osc(slice->cpl_obj);
 336
 337         osc_lru_add(osc_cli(obj), opg);
 338 }
 339
 340 static int osc_page_fail(const struct lu_env *env,
 341                          const struct cl_page_slice *slice,
 342                          struct cl_io *unused)
 343 {
 344         /*
 345          * Cached read?
 346          */
 347         LBUG();
 348         return 0;
 349 }
 350
 351
 352 static const char *osc_list(cfs_list_t *head)
 353 {
 354         return cfs_list_empty(head) ? "-" : "+";
 355 }
 356
 357 static inline cfs_time_t osc_submit_duration(struct osc_page *opg)
 358 {
 359         if (opg->ops_submit_time == 0)
 360                 return 0;
 361
 362         return (cfs_time_current() - opg->ops_submit_time);
 363 }
 364
 365 static int osc_page_print(const struct lu_env *env,
 366                           const struct cl_page_slice *slice,
 367                           void *cookie, lu_printer_t printer)
 368 {
 369         struct osc_page       *opg = cl2osc_page(slice);
 370         struct osc_async_page *oap = &opg->ops_oap;
 371         struct osc_object     *obj = cl2osc(slice->cpl_obj);
 372         struct client_obd     *cli = &osc_export(obj)->exp_obd->u.cli;
 373
 374         return (*printer)(env, cookie, LUSTRE_OSC_NAME"-page@%p: "
 375                           "1< %#x %d %u %s %s > "
 376                           "2< "LPU64" %u %u %#x %#x | %p %p %p > "
 377                           "3< %s %p %d %lu %d > "
 378                           "4< %d %d %d %lu %s | %s %s %s %s > "
 379                           "5< %s %s %s %s | %d %s | %d %s %s>\n",
 380                           opg,
 381                           /* 1 */
 382                           oap->oap_magic, oap->oap_cmd,
 383                           oap->oap_interrupted,
 384                           osc_list(&oap->oap_pending_item),
 385                           osc_list(&oap->oap_rpc_item),
 386                           /* 2 */
 387                           oap->oap_obj_off, oap->oap_page_off, oap->oap_count,
 388                           oap->oap_async_flags, oap->oap_brw_flags,
 389                           oap->oap_request, oap->oap_cli, obj,
 390                           /* 3 */
 391                           osc_list(&opg->ops_inflight),
 392                           opg->ops_submitter, opg->ops_transfer_pinned,
 393                           osc_submit_duration(opg), opg->ops_srvlock,
 394                           /* 4 */
 395                           cli->cl_r_in_flight, cli->cl_w_in_flight,
 396                           cli->cl_max_rpcs_in_flight,
 397                           cli->cl_avail_grant,
 398                           osc_list(&cli->cl_cache_waiters),
 399                           osc_list(&cli->cl_loi_ready_list),
 400                           osc_list(&cli->cl_loi_hp_ready_list),
 401                           osc_list(&cli->cl_loi_write_list),
 402                           osc_list(&cli->cl_loi_read_list),
 403                           /* 5 */
 404                           osc_list(&obj->oo_ready_item),
 405                           osc_list(&obj->oo_hp_ready_item),
 406                           osc_list(&obj->oo_write_item),
 407                           osc_list(&obj->oo_read_item),
 408                           cfs_atomic_read(&obj->oo_nr_reads),
 409                           osc_list(&obj->oo_reading_exts),
 410                           cfs_atomic_read(&obj->oo_nr_writes),
 411                           osc_list(&obj->oo_hp_exts),
 412                           osc_list(&obj->oo_urgent_exts));
 413 }
 414
 415 static void osc_page_delete(const struct lu_env *env,
 416                             const struct cl_page_slice *slice)
 417 {
 418         struct osc_page   *opg = cl2osc_page(slice);
 419         struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
 420         int rc;
 421
 422         LINVRNT(opg->ops_temp || osc_page_protected(env, opg, CLM_READ, 1));
 423
 424         ENTRY;
 425         CDEBUG(D_TRACE, "%p\n", opg);
 426         osc_page_transfer_put(env, opg);
 427         rc = osc_teardown_async_page(env, obj, opg);
 428         if (rc) {
 429                 CL_PAGE_DEBUG(D_ERROR, env, cl_page_top(slice->cpl_page),
 430                               "Trying to teardown failed: %d\n", rc);
 431                 LASSERT(0);
 432         }
 433
 434         spin_lock(&obj->oo_seatbelt);
 435         if (opg->ops_submitter != NULL) {
 436                 LASSERT(!cfs_list_empty(&opg->ops_inflight));
 437                 cfs_list_del_init(&opg->ops_inflight);
 438                 opg->ops_submitter = NULL;
 439         }
 440         spin_unlock(&obj->oo_seatbelt);
 441
 442         osc_lru_del(osc_cli(obj), opg, true);
 443         EXIT;
 444 }
 445
 446 void osc_page_clip(const struct lu_env *env, const struct cl_page_slice *slice,
 447                    int from, int to)
 448 {
 449         struct osc_page       *opg = cl2osc_page(slice);
 450         struct osc_async_page *oap = &opg->ops_oap;
 451
 452         LINVRNT(osc_page_protected(env, opg, CLM_READ, 0));
 453
 454         opg->ops_from = from;
 455         opg->ops_to   = to;
 456         spin_lock(&oap->oap_lock);
 457         oap->oap_async_flags |= ASYNC_COUNT_STABLE;
 458         spin_unlock(&oap->oap_lock);
 459 }
 460
 461 static int osc_page_cancel(const struct lu_env *env,
 462                            const struct cl_page_slice *slice)
 463 {
 464         struct osc_page *opg = cl2osc_page(slice);
 465         int rc = 0;
 466
 467         LINVRNT(osc_page_protected(env, opg, CLM_READ, 0));
 468
 469         /* Check if the transferring against this page
 470          * is completed, or not even queued. */
 471         if (opg->ops_transfer_pinned)
 472                 /* FIXME: may not be interrupted.. */
 473                 rc = osc_cancel_async_page(env, opg);
 474         LASSERT(ergo(rc == 0, opg->ops_transfer_pinned == 0));
 475         return rc;
 476 }
 477
 478 static int osc_page_flush(const struct lu_env *env,
 479                           const struct cl_page_slice *slice,
 480                           struct cl_io *io)
 481 {
 482         struct osc_page *opg = cl2osc_page(slice);
 483         int rc = 0;
 484         ENTRY;
 485         rc = osc_flush_async_page(env, io, opg);
 486         RETURN(rc);
 487 }
 488
 489 static const struct cl_page_operations osc_page_ops = {
 490         .cpo_fini          = osc_page_fini,
 491         .cpo_print         = osc_page_print,
 492         .cpo_delete        = osc_page_delete,
 493         .cpo_is_under_lock = osc_page_is_under_lock,
 494         .cpo_disown        = osc_page_disown,
 495         .io = {
 496                 [CRT_READ] = {
 497                         .cpo_cache_add  = osc_page_fail,
 498                         .cpo_completion = osc_page_completion_read
 499                 },
 500                 [CRT_WRITE] = {
 501                         .cpo_cache_add  = osc_page_cache_add,
 502                         .cpo_completion = osc_page_completion_write
 503                 }
 504         },
 505         .cpo_clip           = osc_page_clip,
 506         .cpo_cancel         = osc_page_cancel,
 507         .cpo_flush          = osc_page_flush
 508 };
 509
 510 int osc_page_init(const struct lu_env *env, struct cl_object *obj,
 511                 struct cl_page *page, cfs_page_t *vmpage)
 512 {
 513         struct osc_object *osc = cl2osc(obj);
 514         struct osc_page   *opg = cl_object_page_slice(obj, page);
 515         int result;
 516
 517         opg->ops_from = 0;
 518         opg->ops_to   = CFS_PAGE_SIZE;
 519
 520         result = osc_prep_async_page(osc, opg, vmpage,
 521                                         cl_offset(obj, page->cp_index));
 522         if (result == 0) {
 523                 struct osc_io *oio = osc_env_io(env);
 524                 opg->ops_srvlock = osc_io_srvlock(oio);
 525                 cl_page_slice_add(page, &opg->ops_cl, obj,
 526                                 &osc_page_ops);
 527         }
 528         /*
 529          * Cannot assert osc_page_protected() here as read-ahead
 530          * creates temporary pages outside of a lock.
 531          */
 532 #ifdef INVARIANT_CHECK
 533         opg->ops_temp = !osc_page_protected(env, opg, CLM_READ, 1);
 534 #endif
 535         /* ops_inflight and ops_lru are the same field, but it doesn't
 536          * hurt to initialize it twice :-) */
 537         CFS_INIT_LIST_HEAD(&opg->ops_inflight);
 538         CFS_INIT_LIST_HEAD(&opg->ops_lru);
 539
 540         /* reserve an LRU space for this page */
 541         if (page->cp_type == CPT_CACHEABLE && result == 0)
 542                 result = osc_lru_reserve(env, osc, opg);
 543
 544         return result;
 545 }
 546
 547 /**
 548  * Helper function called by osc_io_submit() for every page in an immediate
 549  * transfer (i.e., transferred synchronously).
 550  */
 551 void osc_page_submit(const struct lu_env *env, struct osc_page *opg,
 552                      enum cl_req_type crt, int brw_flags)
 553 {
 554         struct osc_async_page *oap = &opg->ops_oap;
 555         struct osc_object     *obj = oap->oap_obj;
 556
 557         LINVRNT(osc_page_protected(env, opg,
 558                                    crt == CRT_WRITE ? CLM_WRITE : CLM_READ, 1));
 559
 560         LASSERTF(oap->oap_magic == OAP_MAGIC, "Bad oap magic: oap %p, "
 561                  "magic 0x%x\n", oap, oap->oap_magic);
 562         LASSERT(oap->oap_async_flags & ASYNC_READY);
 563         LASSERT(oap->oap_async_flags & ASYNC_COUNT_STABLE);
 564
 565         oap->oap_cmd       = crt == CRT_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ;
 566         oap->oap_page_off  = opg->ops_from;
 567         oap->oap_count     = opg->ops_to - opg->ops_from;
 568         oap->oap_brw_flags = OBD_BRW_SYNC | brw_flags;
 569
 570         if (!client_is_remote(osc_export(obj)) &&
 571                         cfs_capable(CFS_CAP_SYS_RESOURCE)) {
 572                 oap->oap_brw_flags |= OBD_BRW_NOQUOTA;
 573                 oap->oap_cmd |= OBD_BRW_NOQUOTA;
 574         }
 575
 576         opg->ops_submit_time = cfs_time_current();
 577         osc_page_transfer_get(opg, "transfer\0imm");
 578         osc_page_transfer_add(env, opg, crt);
 579 }
 580
 581 /* --------------- LRU page management ------------------ */
 582
 583 /* OSC is a natural place to manage LRU pages as applications are specialized
 584  * to write OSC by OSC. Ideally, if one OSC is used more frequently it should
 585  * occupy more LRU slots. On the other hand, we should avoid using up all LRU
 586  * slots (client_obd::cl_lru_left) otherwise process has to be put into sleep
 587  * for free LRU slots - this will be very bad so the algorithm requires each
 588  * OSC to free slots voluntarily to maintain a reasonable number of free slots
 589  * at any time.
 590  */
 591
 592 static CFS_DECL_WAITQ(osc_lru_waitq);
 593 static cfs_atomic_t osc_lru_waiters = CFS_ATOMIC_INIT(0);
 594 /* LRU pages are freed in batch mode. OSC should at least free this
 595  * number of pages to avoid running out of LRU budget, and.. */
 596 static const int lru_shrink_min = 2 << (20 - CFS_PAGE_SHIFT);  /* 2M */
 597 /* free this number at most otherwise it will take too long time to finsih. */
 598 static const int lru_shrink_max = 32 << (20 - CFS_PAGE_SHIFT); /* 32M */
 599
 600 /* Check if we can free LRU slots from this OSC. If there exists LRU waiters,
 601  * we should free slots aggressively. In this way, slots are freed in a steady
 602  * step to maintain fairness among OSCs.
 603  *
 604  * Return how many LRU pages should be freed. */
 605 static int osc_cache_too_much(struct client_obd *cli)
 606 {
 607         struct cl_client_cache *cache = cli->cl_cache;
 608         int pages = cfs_atomic_read(&cli->cl_lru_in_list) >> 1;
 609
 610         if (cfs_atomic_read(&osc_lru_waiters) > 0 &&
 611             cfs_atomic_read(cli->cl_lru_left) < lru_shrink_max)
 612                 /* drop lru pages aggressively */
 613                 return min(pages, lru_shrink_max);
 614
 615         /* if it's going to run out LRU slots, we should free some, but not
 616          * too much to maintain faireness among OSCs. */
 617         if (cfs_atomic_read(cli->cl_lru_left) < cache->ccc_lru_max >> 4) {
 618                 unsigned long tmp;
 619
 620                 tmp = cache->ccc_lru_max / cfs_atomic_read(&cache->ccc_users);
 621                 if (pages > tmp)
 622                         return min(pages, lru_shrink_max);
 623
 624                 return pages > lru_shrink_min ? lru_shrink_min : 0;
 625         }
 626
 627         return 0;
 628 }
 629
 630 /* Return how many pages are not discarded in @pvec. */
 631 static int discard_pagevec(const struct lu_env *env, struct cl_io *io,
 632                            struct cl_page **pvec, int max_index)
 633 {
 634         int count;
 635         int i;
 636
 637         for (count = 0, i = 0; i < max_index; i++) {
 638                 struct cl_page *page = pvec[i];
 639                 if (cl_page_own_try(env, io, page) == 0) {
 640                         /* free LRU page only if nobody is using it.
 641                          * This check is necessary to avoid freeing the pages
 642                          * having already been removed from LRU and pinned
 643                          * for IO. */
 644                         if (!cl_page_in_use(page)) {
 645                                 cl_page_unmap(env, io, page);
 646                                 cl_page_discard(env, io, page);
 647                                 ++count;
 648                         }
 649                         cl_page_disown(env, io, page);
 650                 }
 651                 cl_page_put(env, page);
 652                 pvec[i] = NULL;
 653         }
 654         return max_index - count;
 655 }
 656
 657 /**
 658  * Drop @target of pages from LRU at most.
 659  */
 660 int osc_lru_shrink(struct client_obd *cli, int target)
 661 {
 662         struct cl_env_nest nest;
 663         struct lu_env *env;
 664         struct cl_io *io;
 665         struct cl_object *clobj = NULL;
 666         struct cl_page **pvec;
 667         struct osc_page *opg;
 668         int maxscan = 0;
 669         int count = 0;
 670         int index = 0;
 671         int rc = 0;
 672         ENTRY;
 673
 674         LASSERT(cfs_atomic_read(&cli->cl_lru_in_list) >= 0);
 675         if (cfs_atomic_read(&cli->cl_lru_in_list) == 0 || target <= 0)
 676                 RETURN(0);
 677
 678         env = cl_env_nested_get(&nest);
 679         if (IS_ERR(env))
 680                 RETURN(PTR_ERR(env));
 681
 682         pvec = osc_env_info(env)->oti_pvec;
 683         io = &osc_env_info(env)->oti_io;
 684
 685         client_obd_list_lock(&cli->cl_lru_list_lock);
 686         cfs_atomic_inc(&cli->cl_lru_shrinkers);
 687         maxscan = min(target << 1, cfs_atomic_read(&cli->cl_lru_in_list));
 688         while (!cfs_list_empty(&cli->cl_lru_list)) {
 689                 struct cl_page *page;
 690
 691                 if (--maxscan < 0)
 692                         break;
 693
 694                 opg = cfs_list_entry(cli->cl_lru_list.next, struct osc_page,
 695                                      ops_lru);
 696                 page = cl_page_top(opg->ops_cl.cpl_page);
 697                 if (cl_page_in_use_noref(page)) {
 698                         cfs_list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
 699                         continue;
 700                 }
 701
 702                 LASSERT(page->cp_obj != NULL);
 703                 if (clobj != page->cp_obj) {
 704                         struct cl_object *tmp = page->cp_obj;
 705
 706                         cl_object_get(tmp);
 707                         client_obd_list_unlock(&cli->cl_lru_list_lock);
 708
 709                         if (clobj != NULL) {
 710                                 count -= discard_pagevec(env, io, pvec, index);
 711                                 index = 0;
 712
 713                                 cl_io_fini(env, io);
 714                                 cl_object_put(env, clobj);
 715                                 clobj = NULL;
 716                         }
 717
 718                         clobj = tmp;
 719                         io->ci_obj = clobj;
 720                         rc = cl_io_init(env, io, CIT_MISC, clobj);
 721                         if (rc != 0)
 722                                 break;
 723
 724                         ++maxscan;
 725                         client_obd_list_lock(&cli->cl_lru_list_lock);
 726                         continue;
 727                 }
 728
 729                 /* move this page to the end of list as it will be discarded
 730                  * soon. The page will be finally removed from LRU list in
 731                  * osc_page_delete().  */
 732                 cfs_list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
 733
 734                 /* it's okay to grab a refcount here w/o holding lock because
 735                  * it has to grab cl_lru_list_lock to delete the page. */
 736                 cl_page_get(page);
 737                 pvec[index++] = page;
 738                 if (++count >= target)
 739                         break;
 740
 741                 if (unlikely(index == OTI_PVEC_SIZE)) {
 742                         client_obd_list_unlock(&cli->cl_lru_list_lock);
 743                         count -= discard_pagevec(env, io, pvec, index);
 744                         index = 0;
 745
 746                         client_obd_list_lock(&cli->cl_lru_list_lock);
 747                 }
 748         }
 749         client_obd_list_unlock(&cli->cl_lru_list_lock);
 750
 751         if (clobj != NULL) {
 752                 count -= discard_pagevec(env, io, pvec, index);
 753
 754                 cl_io_fini(env, io);
 755                 cl_object_put(env, clobj);
 756         }
 757         cl_env_nested_put(&nest, env);
 758
 759         cfs_atomic_dec(&cli->cl_lru_shrinkers);
 760         RETURN(count > 0 ? count : rc);
 761 }
 762
 763 static void osc_lru_add(struct client_obd *cli, struct osc_page *opg)
 764 {
 765         bool wakeup = false;
 766
 767         if (!opg->ops_in_lru)
 768                 return;
 769
 770         cfs_atomic_dec(&cli->cl_lru_busy);
 771         client_obd_list_lock(&cli->cl_lru_list_lock);
 772         if (cfs_list_empty(&opg->ops_lru)) {
 773                 cfs_list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
 774                 cfs_atomic_inc_return(&cli->cl_lru_in_list);
 775                 wakeup = cfs_atomic_read(&osc_lru_waiters) > 0;
 776         }
 777         client_obd_list_unlock(&cli->cl_lru_list_lock);
 778
 779         if (wakeup)
 780                 cfs_waitq_broadcast(&osc_lru_waitq);
 781 }
 782
 783 /* delete page from LRUlist. The page can be deleted from LRUlist for two
 784  * reasons: redirtied or deleted from page cache. */
 785 static void osc_lru_del(struct client_obd *cli, struct osc_page *opg, bool del)
 786 {
 787         if (opg->ops_in_lru) {
 788                 client_obd_list_lock(&cli->cl_lru_list_lock);
 789                 if (!cfs_list_empty(&opg->ops_lru)) {
 790                         LASSERT(cfs_atomic_read(&cli->cl_lru_in_list) > 0);
 791                         cfs_list_del_init(&opg->ops_lru);
 792                         cfs_atomic_dec(&cli->cl_lru_in_list);
 793                         if (!del)
 794                                 cfs_atomic_inc(&cli->cl_lru_busy);
 795                 } else if (del) {
 796                         LASSERT(cfs_atomic_read(&cli->cl_lru_busy) > 0);
 797                         cfs_atomic_dec(&cli->cl_lru_busy);
 798                 }
 799                 client_obd_list_unlock(&cli->cl_lru_list_lock);
 800                 if (del) {
 801                         cfs_atomic_inc(cli->cl_lru_left);
 802                         /* this is a great place to release more LRU pages if
 803                          * this osc occupies too many LRU pages and kernel is
 804                          * stealing one of them.
 805                          * cl_lru_shrinkers is to avoid recursive call in case
 806                          * we're already in the context of osc_lru_shrink(). */
 807                         if (cfs_atomic_read(&cli->cl_lru_shrinkers) == 0)
 808                                 osc_lru_shrink(cli, osc_cache_too_much(cli));
 809                         cfs_waitq_signal(&osc_lru_waitq);
 810                 }
 811         } else {
 812                 LASSERT(cfs_list_empty(&opg->ops_lru));
 813         }
 814 }
 815
 816 static int osc_lru_reclaim(struct client_obd *cli)
 817 {
 818         struct cl_client_cache *cache = cli->cl_cache;
 819         struct client_obd *victim;
 820         struct client_obd *tmp;
 821         int rc;
 822
 823         LASSERT(cache != NULL);
 824         LASSERT(!cfs_list_empty(&cache->ccc_lru));
 825
 826         rc = osc_lru_shrink(cli, lru_shrink_min);
 827         if (rc > 0) {
 828                 CDEBUG(D_CACHE, "%s: Free %d pages from own LRU: %p.\n",
 829                         cli->cl_import->imp_obd->obd_name, rc, cli);
 830                 return rc;
 831         }
 832
 833         CDEBUG(D_CACHE, "%s: cli %p no free slots, pages: %d, busy: %d.\n",
 834                 cli->cl_import->imp_obd->obd_name, cli,
 835                 cfs_atomic_read(&cli->cl_lru_in_list),
 836                 cfs_atomic_read(&cli->cl_lru_busy));
 837
 838         /* Reclaim LRU slots from other client_obd as it can't free enough
 839          * from its own. This should rarely happen. */
 840         spin_lock(&cache->ccc_lru_lock);
 841         cache->ccc_lru_shrinkers++;
 842         cfs_list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
 843         cfs_list_for_each_entry_safe(victim, tmp, &cache->ccc_lru, cl_lru_osc) {
 844                 if (victim == cli)
 845                         break;
 846
 847                 CDEBUG(D_CACHE, "%s: cli %p LRU pages: %d, busy: %d.\n",
 848                         victim->cl_import->imp_obd->obd_name, victim,
 849                         cfs_atomic_read(&victim->cl_lru_in_list),
 850                         cfs_atomic_read(&victim->cl_lru_busy));
 851
 852                 cfs_list_move_tail(&victim->cl_lru_osc, &cache->ccc_lru);
 853                 if (cfs_atomic_read(&victim->cl_lru_in_list) > 0)
 854                         break;
 855         }
 856         spin_unlock(&cache->ccc_lru_lock);
 857         if (victim == cli) {
 858                 CDEBUG(D_CACHE, "%s: can't get any free LRU slots.\n",
 859                         cli->cl_import->imp_obd->obd_name);
 860                 return 0;
 861         }
 862
 863         rc = osc_lru_shrink(victim,
 864                             min(cfs_atomic_read(&victim->cl_lru_in_list) >> 1,
 865                                 lru_shrink_max));
 866
 867         CDEBUG(D_CACHE, "%s: Free %d pages from other cli: %p.\n",
 868                 cli->cl_import->imp_obd->obd_name, rc, victim);
 869
 870         return rc;
 871 }
 872
 873 static int osc_lru_reserve(const struct lu_env *env, struct osc_object *obj,
 874                            struct osc_page *opg)
 875 {
 876         struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
 877         struct client_obd *cli = osc_cli(obj);
 878         int rc = 0;
 879         ENTRY;
 880
 881         if (cli->cl_cache == NULL) /* shall not be in LRU */
 882                 RETURN(0);
 883
 884         LASSERT(cfs_atomic_read(cli->cl_lru_left) >= 0);
 885         while (!cfs_atomic_add_unless(cli->cl_lru_left, -1, 0)) {
 886                 int gen;
 887
 888                 /* run out of LRU spaces, try to drop some by itself */
 889                 rc = osc_lru_reclaim(cli);
 890                 if (rc < 0)
 891                         break;
 892                 if (rc > 0)
 893                         continue;
 894
 895                 cfs_cond_resched();
 896
 897                 /* slowest case, all of caching pages are busy, notifying
 898                  * other OSCs that we're lack of LRU slots. */
 899                 cfs_atomic_inc(&osc_lru_waiters);
 900
 901                 gen = cfs_atomic_read(&cli->cl_lru_in_list);
 902                 rc = l_wait_event(osc_lru_waitq,
 903                                 cfs_atomic_read(cli->cl_lru_left) > 0 ||
 904                                 (cfs_atomic_read(&cli->cl_lru_in_list) > 0 &&
 905                                  gen != cfs_atomic_read(&cli->cl_lru_in_list)),
 906                                 &lwi);
 907
 908                 cfs_atomic_dec(&osc_lru_waiters);
 909                 if (rc < 0)
 910                         break;
 911         }
 912
 913         if (rc >= 0) {
 914                 cfs_atomic_inc(&cli->cl_lru_busy);
 915                 opg->ops_in_lru = 1;
 916                 rc = 0;
 917         }
 918
 919         RETURN(rc);
 920 }
 921
 922 /** @} osc */