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, 2013, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * Implementation of cl_page for OSC layer.
  37  *
  38  *   Author: Nikita Danilov <nikita.danilov@sun.com>
  39  *   Author: Jinshan Xiong <jinshan.xiong@intel.com>
  40  */
  41
  42 #define DEBUG_SUBSYSTEM S_OSC
  43
  44 #include "osc_cl_internal.h"
  45
  46 static void osc_lru_del(struct client_obd *cli, struct osc_page *opg);
  47 static void osc_lru_use(struct client_obd *cli, struct osc_page *opg);
  48 static int osc_lru_reserve(const struct lu_env *env, struct osc_object *obj,
  49                            struct osc_page *opg);
  50
  51 /** \addtogroup osc
  52  *  @{
  53  */
  54
  55 /*
  56  * Comment out osc_page_protected because it may sleep inside the
  57  * the client_obd_list_lock.
  58  * client_obd_list_lock -> osc_ap_completion -> osc_completion ->
  59  *   -> osc_page_protected -> osc_page_is_dlocked -> osc_match_base
  60  *   -> ldlm_lock_match -> sptlrpc_import_check_ctx -> sleep.
  61  */
  62 #if 0
  63 static int osc_page_is_dlocked(const struct lu_env *env,
  64                                const struct osc_page *opg,
  65                                enum cl_lock_mode mode, int pending, int unref)
  66 {
  67         struct cl_page         *page;
  68         struct osc_object      *obj;
  69         struct osc_thread_info *info;
  70         struct ldlm_res_id     *resname;
  71         struct lustre_handle   *lockh;
  72         ldlm_policy_data_t     *policy;
  73         ldlm_mode_t             dlmmode;
  74         __u64                   flags;
  75
  76         might_sleep();
  77
  78         info = osc_env_info(env);
  79         resname = &info->oti_resname;
  80         policy = &info->oti_policy;
  81         lockh = &info->oti_handle;
  82         page = opg->ops_cl.cpl_page;
  83         obj = cl2osc(opg->ops_cl.cpl_obj);
  84
  85         flags = LDLM_FL_TEST_LOCK | LDLM_FL_BLOCK_GRANTED;
  86         if (pending)
  87                 flags |= LDLM_FL_CBPENDING;
  88
  89         dlmmode = osc_cl_lock2ldlm(mode) | LCK_PW;
  90         osc_lock_build_res(env, obj, resname);
  91         osc_index2policy(policy, page->cp_obj, osc_index(opg), osc_index(opg));
  92         return osc_match_base(osc_export(obj), resname, LDLM_EXTENT, policy,
  93                               dlmmode, &flags, NULL, lockh, unref);
  94 }
  95
  96 /**
  97  * Checks an invariant that a page in the cache is covered by a lock, as
  98  * needed.
  99  */
 100 static int osc_page_protected(const struct lu_env *env,
 101                               const struct osc_page *opg,
 102                               enum cl_lock_mode mode, int unref)
 103 {
 104         struct cl_object_header *hdr;
 105         struct cl_lock          *scan;
 106         struct cl_page          *page;
 107         struct cl_lock_descr    *descr;
 108         int result;
 109
 110         LINVRNT(!opg->ops_temp);
 111
 112         page = opg->ops_cl.cpl_page;
 113         if (page->cp_owner != NULL &&
 114             cl_io_top(page->cp_owner)->ci_lockreq == CILR_NEVER)
 115                 /*
 116                  * If IO is done without locks (liblustre, or lloop), lock is
 117                  * not required.
 118                  */
 119                 result = 1;
 120         else
 121                 /* otherwise check for a DLM lock */
 122         result = osc_page_is_dlocked(env, opg, mode, 1, unref);
 123         if (result == 0) {
 124                 /* maybe this page is a part of a lockless io? */
 125                 hdr = cl_object_header(opg->ops_cl.cpl_obj);
 126                 descr = &osc_env_info(env)->oti_descr;
 127                 descr->cld_mode = mode;
 128                 descr->cld_start = osc_index(opg);
 129                 descr->cld_end   = osc_index(opg);
 130                 spin_lock(&hdr->coh_lock_guard);
 131                 cfs_list_for_each_entry(scan, &hdr->coh_locks, cll_linkage) {
 132                         /*
 133                          * Lock-less sub-lock has to be either in HELD state
 134                          * (when io is actively going on), or in CACHED state,
 135                          * when top-lock is being unlocked:
 136                          * cl_io_unlock()->cl_unuse()->...->lov_lock_unuse().
 137                          */
 138                         if ((scan->cll_state == CLS_HELD ||
 139                              scan->cll_state == CLS_CACHED) &&
 140                             cl_lock_ext_match(&scan->cll_descr, descr)) {
 141                                 struct osc_lock *olck;
 142
 143                                 olck = osc_lock_at(scan);
 144                                 result = osc_lock_is_lockless(olck);
 145                                 break;
 146                         }
 147                 }
 148                 spin_unlock(&hdr->coh_lock_guard);
 149         }
 150         return result;
 151 }
 152 #else
 153 static int osc_page_protected(const struct lu_env *env,
 154                               const struct osc_page *opg,
 155                               enum cl_lock_mode mode, int unref)
 156 {
 157         return 1;
 158 }
 159 #endif
 160
 161 /*****************************************************************************
 162  *
 163  * Page operations.
 164  *
 165  */
 166 static void osc_page_transfer_get(struct osc_page *opg, const char *label)
 167 {
 168         struct cl_page *page = opg->ops_cl.cpl_page;
 169
 170         LASSERT(!opg->ops_transfer_pinned);
 171         cl_page_get(page);
 172         lu_ref_add_atomic(&page->cp_reference, label, page);
 173         opg->ops_transfer_pinned = 1;
 174 }
 175
 176 static void osc_page_transfer_put(const struct lu_env *env,
 177                                   struct osc_page *opg)
 178 {
 179         struct cl_page *page = opg->ops_cl.cpl_page;
 180
 181         if (opg->ops_transfer_pinned) {
 182                 opg->ops_transfer_pinned = 0;
 183                 lu_ref_del(&page->cp_reference, "transfer", page);
 184                 cl_page_put(env, page);
 185         }
 186 }
 187
 188 /**
 189  * This is called once for every page when it is submitted for a transfer
 190  * either opportunistic (osc_page_cache_add()), or immediate
 191  * (osc_page_submit()).
 192  */
 193 static void osc_page_transfer_add(const struct lu_env *env,
 194                                   struct osc_page *opg, enum cl_req_type crt)
 195 {
 196         struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
 197
 198         /* ops_lru and ops_inflight share the same field, so take it from LRU
 199          * first and then use it as inflight. */
 200         osc_lru_use(osc_cli(obj), opg);
 201
 202         spin_lock(&obj->oo_seatbelt);
 203         cfs_list_add(&opg->ops_inflight, &obj->oo_inflight[crt]);
 204         opg->ops_submitter = current;
 205         spin_unlock(&obj->oo_seatbelt);
 206 }
 207
 208 int osc_page_cache_add(const struct lu_env *env,
 209                         const struct cl_page_slice *slice, struct cl_io *io)
 210 {
 211         struct osc_page *opg = cl2osc_page(slice);
 212         int result;
 213         ENTRY;
 214
 215         LINVRNT(osc_page_protected(env, opg, CLM_WRITE, 0));
 216
 217         osc_page_transfer_get(opg, "transfer\0cache");
 218         result = osc_queue_async_io(env, io, opg);
 219         if (result != 0)
 220                 osc_page_transfer_put(env, opg);
 221         else
 222                 osc_page_transfer_add(env, opg, CRT_WRITE);
 223
 224         RETURN(result);
 225 }
 226
 227 void osc_index2policy(ldlm_policy_data_t *policy, const struct cl_object *obj,
 228                       pgoff_t start, pgoff_t end)
 229 {
 230         memset(policy, 0, sizeof *policy);
 231         policy->l_extent.start = cl_offset(obj, start);
 232         policy->l_extent.end   = cl_offset(obj, end + 1) - 1;
 233 }
 234
 235 static int osc_page_is_under_lock(const struct lu_env *env,
 236                                   const struct cl_page_slice *slice,
 237                                   struct cl_io *unused)
 238 {
 239         struct osc_page *opg = cl2osc_page(slice);
 240         struct cl_lock *lock;
 241         int             result = -ENODATA;
 242
 243         ENTRY;
 244         lock = cl_lock_at_pgoff(env, slice->cpl_obj, osc_index(opg),
 245                         NULL, 1, 0);
 246         if (lock != NULL) {
 247                 cl_lock_put(env, lock);
 248                 result = -EBUSY;
 249         }
 250         RETURN(result);
 251 }
 252
 253 static const char *osc_list(cfs_list_t *head)
 254 {
 255         return cfs_list_empty(head) ? "-" : "+";
 256 }
 257
 258 static inline cfs_time_t 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 %lu: "
 276                           "1< %#x %d %u %s %s > "
 277                           "2< "LPU64" %u %u %#x %#x | %p %p %p > "
 278                           "3< %s %p %d %lu %d > "
 279                           "4< %d %d %d %lu %s | %s %s %s %s > "
 280                           "5< %s %s %s %s | %d %s | %d %s %s>\n",
 281                           opg, osc_index(opg),
 282                           /* 1 */
 283                           oap->oap_magic, oap->oap_cmd,
 284                           oap->oap_interrupted,
 285                           osc_list(&oap->oap_pending_item),
 286                           osc_list(&oap->oap_rpc_item),
 287                           /* 2 */
 288                           oap->oap_obj_off, oap->oap_page_off, oap->oap_count,
 289                           oap->oap_async_flags, oap->oap_brw_flags,
 290                           oap->oap_request, oap->oap_cli, obj,
 291                           /* 3 */
 292                           osc_list(&opg->ops_inflight),
 293                           opg->ops_submitter, opg->ops_transfer_pinned,
 294                           osc_submit_duration(opg), opg->ops_srvlock,
 295                           /* 4 */
 296                           cli->cl_r_in_flight, cli->cl_w_in_flight,
 297                           cli->cl_max_rpcs_in_flight,
 298                           cli->cl_avail_grant,
 299                           osc_list(&cli->cl_cache_waiters),
 300                           osc_list(&cli->cl_loi_ready_list),
 301                           osc_list(&cli->cl_loi_hp_ready_list),
 302                           osc_list(&cli->cl_loi_write_list),
 303                           osc_list(&cli->cl_loi_read_list),
 304                           /* 5 */
 305                           osc_list(&obj->oo_ready_item),
 306                           osc_list(&obj->oo_hp_ready_item),
 307                           osc_list(&obj->oo_write_item),
 308                           osc_list(&obj->oo_read_item),
 309                           cfs_atomic_read(&obj->oo_nr_reads),
 310                           osc_list(&obj->oo_reading_exts),
 311                           cfs_atomic_read(&obj->oo_nr_writes),
 312                           osc_list(&obj->oo_hp_exts),
 313                           osc_list(&obj->oo_urgent_exts));
 314 }
 315
 316 static void osc_page_delete(const struct lu_env *env,
 317                             const struct cl_page_slice *slice)
 318 {
 319         struct osc_page   *opg = cl2osc_page(slice);
 320         struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
 321         int rc;
 322
 323         LINVRNT(opg->ops_temp || osc_page_protected(env, opg, CLM_READ, 1));
 324
 325         ENTRY;
 326         CDEBUG(D_TRACE, "%p\n", opg);
 327         osc_page_transfer_put(env, opg);
 328         rc = osc_teardown_async_page(env, obj, opg);
 329         if (rc) {
 330                 CL_PAGE_DEBUG(D_ERROR, env, slice->cpl_page,
 331                               "Trying to teardown failed: %d\n", rc);
 332                 LASSERT(0);
 333         }
 334
 335         spin_lock(&obj->oo_seatbelt);
 336         if (opg->ops_submitter != NULL) {
 337                 LASSERT(!cfs_list_empty(&opg->ops_inflight));
 338                 cfs_list_del_init(&opg->ops_inflight);
 339                 opg->ops_submitter = NULL;
 340         }
 341         spin_unlock(&obj->oo_seatbelt);
 342
 343         osc_lru_del(osc_cli(obj), opg);
 344
 345         if (slice->cpl_page->cp_type == CPT_CACHEABLE) {
 346                 void *value;
 347
 348                 spin_lock(&obj->oo_tree_lock);
 349                 value = radix_tree_delete(&obj->oo_tree, osc_index(opg));
 350                 if (value != NULL)
 351                         --obj->oo_npages;
 352                 spin_unlock(&obj->oo_tree_lock);
 353
 354                 LASSERT(ergo(value != NULL, value == opg));
 355         }
 356
 357         EXIT;
 358 }
 359
 360 void osc_page_clip(const struct lu_env *env, const struct cl_page_slice *slice,
 361                    int from, int to)
 362 {
 363         struct osc_page       *opg = cl2osc_page(slice);
 364         struct osc_async_page *oap = &opg->ops_oap;
 365
 366         LINVRNT(osc_page_protected(env, opg, CLM_READ, 0));
 367
 368         opg->ops_from = from;
 369         opg->ops_to   = to;
 370         spin_lock(&oap->oap_lock);
 371         oap->oap_async_flags |= ASYNC_COUNT_STABLE;
 372         spin_unlock(&oap->oap_lock);
 373 }
 374
 375 static int osc_page_cancel(const struct lu_env *env,
 376                            const struct cl_page_slice *slice)
 377 {
 378         struct osc_page *opg = cl2osc_page(slice);
 379         int rc = 0;
 380
 381         LINVRNT(osc_page_protected(env, opg, CLM_READ, 0));
 382
 383         /* Check if the transferring against this page
 384          * is completed, or not even queued. */
 385         if (opg->ops_transfer_pinned)
 386                 /* FIXME: may not be interrupted.. */
 387                 rc = osc_cancel_async_page(env, opg);
 388         LASSERT(ergo(rc == 0, opg->ops_transfer_pinned == 0));
 389         return rc;
 390 }
 391
 392 static int osc_page_flush(const struct lu_env *env,
 393                           const struct cl_page_slice *slice,
 394                           struct cl_io *io)
 395 {
 396         struct osc_page *opg = cl2osc_page(slice);
 397         int rc = 0;
 398         ENTRY;
 399         rc = osc_flush_async_page(env, io, opg);
 400         RETURN(rc);
 401 }
 402
 403 static const struct cl_page_operations osc_page_ops = {
 404         .cpo_print         = osc_page_print,
 405         .cpo_delete        = osc_page_delete,
 406         .cpo_is_under_lock = osc_page_is_under_lock,
 407         .cpo_clip           = osc_page_clip,
 408         .cpo_cancel         = osc_page_cancel,
 409         .cpo_flush          = osc_page_flush
 410 };
 411
 412 int osc_page_init(const struct lu_env *env, struct cl_object *obj,
 413                   struct cl_page *page, pgoff_t index)
 414 {
 415         struct osc_object *osc = cl2osc(obj);
 416         struct osc_page   *opg = cl_object_page_slice(obj, page);
 417         int result;
 418
 419         opg->ops_from = 0;
 420         opg->ops_to   = PAGE_CACHE_SIZE;
 421         opg->ops_cl.cpl_index = index;
 422
 423         result = osc_prep_async_page(osc, opg, page->cp_vmpage,
 424                                      cl_offset(obj, index));
 425         if (result == 0) {
 426                 struct osc_io *oio = osc_env_io(env);
 427                 opg->ops_srvlock = osc_io_srvlock(oio);
 428                 cl_page_slice_add(page, &opg->ops_cl, obj,
 429                                 &osc_page_ops);
 430         }
 431         /*
 432          * Cannot assert osc_page_protected() here as read-ahead
 433          * creates temporary pages outside of a lock.
 434          */
 435 #ifdef CONFIG_LUSTRE_DEBUG_EXPENSIVE_CHECK
 436         opg->ops_temp = !osc_page_protected(env, opg, CLM_READ, 1);
 437 #endif
 438         /* ops_inflight and ops_lru are the same field, but it doesn't
 439          * hurt to initialize it twice :-) */
 440         CFS_INIT_LIST_HEAD(&opg->ops_inflight);
 441         CFS_INIT_LIST_HEAD(&opg->ops_lru);
 442
 443         /* reserve an LRU space for this page */
 444         if (page->cp_type == CPT_CACHEABLE && result == 0) {
 445                 result = osc_lru_reserve(env, osc, opg);
 446                 if (result == 0) {
 447                         spin_lock(&osc->oo_tree_lock);
 448                         result = radix_tree_insert(&osc->oo_tree, index, opg);
 449                         if (result == 0)
 450                                 ++osc->oo_npages;
 451                         spin_unlock(&osc->oo_tree_lock);
 452                         LASSERT(result == 0);
 453                 }
 454         }
 455
 456         return result;
 457 }
 458
 459 int osc_over_unstable_soft_limit(struct client_obd *cli)
 460 {
 461         long obd_upages, obd_dpages, osc_upages;
 462
 463         /* Can't check cli->cl_unstable_count, therefore, no soft limit */
 464         if (cli == NULL)
 465                 return 0;
 466
 467         obd_upages = cfs_atomic_read(&obd_unstable_pages);
 468         obd_dpages = cfs_atomic_read(&obd_dirty_pages);
 469
 470         osc_upages = cfs_atomic_read(&cli->cl_unstable_count);
 471
 472         /* obd_max_dirty_pages is the max number of (dirty + unstable)
 473          * pages allowed at any given time. To simulate an unstable page
 474          * only limit, we subtract the current number of dirty pages
 475          * from this max. This difference is roughly the amount of pages
 476          * currently available for unstable pages. Thus, the soft limit
 477          * is half of that difference. Check osc_upages to ensure we don't
 478          * set SOFT_SYNC for OSCs without any outstanding unstable pages. */
 479         return osc_upages != 0 &&
 480                obd_upages >= (obd_max_dirty_pages - obd_dpages) / 2;
 481 }
 482
 483 /**
 484  * Helper function called by osc_io_submit() for every page in an immediate
 485  * transfer (i.e., transferred synchronously).
 486  */
 487 void osc_page_submit(const struct lu_env *env, struct osc_page *opg,
 488                      enum cl_req_type crt, int brw_flags)
 489 {
 490         struct osc_async_page *oap = &opg->ops_oap;
 491         struct osc_object     *obj = oap->oap_obj;
 492
 493         LINVRNT(osc_page_protected(env, opg,
 494                                    crt == CRT_WRITE ? CLM_WRITE : CLM_READ, 1));
 495
 496         LASSERTF(oap->oap_magic == OAP_MAGIC, "Bad oap magic: oap %p, "
 497                  "magic 0x%x\n", oap, oap->oap_magic);
 498         LASSERT(oap->oap_async_flags & ASYNC_READY);
 499         LASSERT(oap->oap_async_flags & ASYNC_COUNT_STABLE);
 500
 501         oap->oap_cmd       = crt == CRT_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ;
 502         oap->oap_page_off  = opg->ops_from;
 503         oap->oap_count     = opg->ops_to - opg->ops_from;
 504         oap->oap_brw_flags = OBD_BRW_SYNC | brw_flags;
 505
 506         if (osc_over_unstable_soft_limit(oap->oap_cli))
 507                 oap->oap_brw_flags |= OBD_BRW_SOFT_SYNC;
 508
 509         if (!client_is_remote(osc_export(obj)) &&
 510                         cfs_capable(CFS_CAP_SYS_RESOURCE)) {
 511                 oap->oap_brw_flags |= OBD_BRW_NOQUOTA;
 512                 oap->oap_cmd |= OBD_BRW_NOQUOTA;
 513         }
 514
 515         opg->ops_submit_time = cfs_time_current();
 516         osc_page_transfer_get(opg, "transfer\0imm");
 517         osc_page_transfer_add(env, opg, crt);
 518 }
 519
 520 /* --------------- LRU page management ------------------ */
 521
 522 /* OSC is a natural place to manage LRU pages as applications are specialized
 523  * to write OSC by OSC. Ideally, if one OSC is used more frequently it should
 524  * occupy more LRU slots. On the other hand, we should avoid using up all LRU
 525  * slots (client_obd::cl_lru_left) otherwise process has to be put into sleep
 526  * for free LRU slots - this will be very bad so the algorithm requires each
 527  * OSC to free slots voluntarily to maintain a reasonable number of free slots
 528  * at any time.
 529  */
 530
 531 static CFS_DECL_WAITQ(osc_lru_waitq);
 532 /* LRU pages are freed in batch mode. OSC should at least free this
 533  * number of pages to avoid running out of LRU budget, and.. */
 534 static const int lru_shrink_min = 2 << (20 - PAGE_CACHE_SHIFT); /* 2M */
 535 /* free this number at most otherwise it will take too long time to finsih. */
 536 static const int lru_shrink_max = 8 << (20 - PAGE_CACHE_SHIFT); /* 8M */
 537
 538 /* Check if we can free LRU slots from this OSC. If there exists LRU waiters,
 539  * we should free slots aggressively. In this way, slots are freed in a steady
 540  * step to maintain fairness among OSCs.
 541  *
 542  * Return how many LRU pages should be freed. */
 543 static int osc_cache_too_much(struct client_obd *cli)
 544 {
 545         struct cl_client_cache *cache = cli->cl_cache;
 546         int pages = cfs_atomic_read(&cli->cl_lru_in_list);
 547         unsigned long budget;
 548
 549         budget = cache->ccc_lru_max / cfs_atomic_read(&cache->ccc_users);
 550
 551         /* if it's going to run out LRU slots, we should free some, but not
 552          * too much to maintain faireness among OSCs. */
 553         if (cfs_atomic_read(cli->cl_lru_left) < cache->ccc_lru_max >> 4) {
 554                 if (pages >= budget)
 555                         return lru_shrink_max;
 556                 else if (pages >= budget / 2)
 557                         return lru_shrink_min;
 558         } else if (pages >= budget * 2)
 559                 return lru_shrink_min;
 560         return 0;
 561 }
 562
 563 int lru_queue_work(const struct lu_env *env, void *data)
 564 {
 565         struct client_obd *cli = data;
 566
 567         CDEBUG(D_CACHE, "Run LRU work for client obd %p.\n", cli);
 568
 569         if (osc_cache_too_much(cli))
 570                 osc_lru_shrink(env, cli, lru_shrink_max, true);
 571
 572         RETURN(0);
 573 }
 574
 575 void osc_lru_add_batch(struct client_obd *cli, cfs_list_t *plist)
 576 {
 577         CFS_LIST_HEAD(lru);
 578         struct osc_async_page *oap;
 579         int npages = 0;
 580
 581         cfs_list_for_each_entry(oap, plist, oap_pending_item) {
 582                 struct osc_page *opg = oap2osc_page(oap);
 583
 584                 if (!opg->ops_in_lru)
 585                         continue;
 586
 587                 ++npages;
 588                 LASSERT(cfs_list_empty(&opg->ops_lru));
 589                 cfs_list_add(&opg->ops_lru, &lru);
 590         }
 591
 592         if (npages > 0) {
 593                 client_obd_list_lock(&cli->cl_lru_list_lock);
 594                 cfs_list_splice_tail(&lru, &cli->cl_lru_list);
 595                 cfs_atomic_sub(npages, &cli->cl_lru_busy);
 596                 cfs_atomic_add(npages, &cli->cl_lru_in_list);
 597                 client_obd_list_unlock(&cli->cl_lru_list_lock);
 598
 599                 /* XXX: May set force to be true for better performance */
 600                 if (osc_cache_too_much(cli))
 601                         (void)ptlrpcd_queue_work(cli->cl_lru_work);
 602         }
 603 }
 604
 605 static void __osc_lru_del(struct client_obd *cli, struct osc_page *opg)
 606 {
 607         LASSERT(cfs_atomic_read(&cli->cl_lru_in_list) > 0);
 608         cfs_list_del_init(&opg->ops_lru);
 609         cfs_atomic_dec(&cli->cl_lru_in_list);
 610 }
 611
 612 /**
 613  * Page is being destroyed. The page may be not in LRU list, if the transfer
 614  * has never finished(error occurred).
 615  */
 616 static void osc_lru_del(struct client_obd *cli, struct osc_page *opg)
 617 {
 618         if (opg->ops_in_lru) {
 619                 client_obd_list_lock(&cli->cl_lru_list_lock);
 620                 if (!cfs_list_empty(&opg->ops_lru)) {
 621                         __osc_lru_del(cli, opg);
 622                 } else {
 623                         LASSERT(cfs_atomic_read(&cli->cl_lru_busy) > 0);
 624                         cfs_atomic_dec(&cli->cl_lru_busy);
 625                 }
 626                 client_obd_list_unlock(&cli->cl_lru_list_lock);
 627
 628                 cfs_atomic_inc(cli->cl_lru_left);
 629                 /* this is a great place to release more LRU pages if
 630                  * this osc occupies too many LRU pages and kernel is
 631                  * stealing one of them. */
 632                 if (!memory_pressure_get())
 633                         (void)ptlrpcd_queue_work(cli->cl_lru_work);
 634                 wake_up(&osc_lru_waitq);
 635         } else {
 636                 LASSERT(cfs_list_empty(&opg->ops_lru));
 637         }
 638 }
 639
 640 /**
 641  * Delete page from LRUlist for redirty.
 642  */
 643 static void osc_lru_use(struct client_obd *cli, struct osc_page *opg)
 644 {
 645         /* If page is being transfered for the first time,
 646          * ops_lru should be empty */
 647         if (opg->ops_in_lru && !cfs_list_empty(&opg->ops_lru)) {
 648                 client_obd_list_lock(&cli->cl_lru_list_lock);
 649                 __osc_lru_del(cli, opg);
 650                 client_obd_list_unlock(&cli->cl_lru_list_lock);
 651                 cfs_atomic_inc(&cli->cl_lru_busy);
 652         }
 653 }
 654
 655 static void discard_pagevec(const struct lu_env *env, struct cl_io *io,
 656                                 struct cl_page **pvec, int max_index)
 657 {
 658         int i;
 659
 660         for (i = 0; i < max_index; i++) {
 661                 struct cl_page *page = pvec[i];
 662
 663                 LASSERT(cl_page_is_owned(page, io));
 664                 cl_page_discard(env, io, page);
 665                 cl_page_disown(env, io, page);
 666                 cl_page_put(env, page);
 667
 668                 pvec[i] = NULL;
 669         }
 670 }
 671
 672 /**
 673  * Drop @target of pages from LRU at most.
 674  */
 675 int osc_lru_shrink(const struct lu_env *env, struct client_obd *cli,
 676                    int target, bool force)
 677 {
 678         struct cl_io *io;
 679         struct cl_object *clobj = NULL;
 680         struct cl_page **pvec;
 681         struct osc_page *opg;
 682         int maxscan = 0;
 683         int count = 0;
 684         int index = 0;
 685         int rc = 0;
 686         ENTRY;
 687
 688         LASSERT(cfs_atomic_read(&cli->cl_lru_in_list) >= 0);
 689         if (cfs_atomic_read(&cli->cl_lru_in_list) == 0 || target <= 0)
 690                 RETURN(0);
 691
 692         if (!force) {
 693                 if (cfs_atomic_read(&cli->cl_lru_shrinkers) > 0)
 694                         RETURN(-EBUSY);
 695
 696                 if (cfs_atomic_inc_return(&cli->cl_lru_shrinkers) > 1) {
 697                         cfs_atomic_dec(&cli->cl_lru_shrinkers);
 698                         RETURN(-EBUSY);
 699                 }
 700         } else {
 701                 cfs_atomic_inc(&cli->cl_lru_shrinkers);
 702         }
 703
 704         pvec = (struct cl_page **)osc_env_info(env)->oti_pvec;
 705         io = &osc_env_info(env)->oti_io;
 706
 707         client_obd_list_lock(&cli->cl_lru_list_lock);
 708         maxscan = min(target << 1, cfs_atomic_read(&cli->cl_lru_in_list));
 709         while (!cfs_list_empty(&cli->cl_lru_list)) {
 710                 struct cl_page *page;
 711                 bool will_free = false;
 712
 713                 if (--maxscan < 0)
 714                         break;
 715
 716                 opg = cfs_list_entry(cli->cl_lru_list.next, struct osc_page,
 717                                      ops_lru);
 718                 page = opg->ops_cl.cpl_page;
 719                 if (cl_page_in_use_noref(page)) {
 720                         cfs_list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
 721                         continue;
 722                 }
 723
 724                 LASSERT(page->cp_obj != NULL);
 725                 if (clobj != page->cp_obj) {
 726                         struct cl_object *tmp = page->cp_obj;
 727
 728                         cl_object_get(tmp);
 729                         client_obd_list_unlock(&cli->cl_lru_list_lock);
 730
 731                         if (clobj != NULL) {
 732                                 discard_pagevec(env, io, pvec, index);
 733                                 index = 0;
 734
 735                                 cl_io_fini(env, io);
 736                                 cl_object_put(env, clobj);
 737                                 clobj = NULL;
 738                         }
 739
 740                         clobj = tmp;
 741                         io->ci_obj = clobj;
 742                         io->ci_ignore_layout = 1;
 743                         rc = cl_io_init(env, io, CIT_MISC, clobj);
 744
 745                         client_obd_list_lock(&cli->cl_lru_list_lock);
 746
 747                         if (rc != 0)
 748                                 break;
 749
 750                         ++maxscan;
 751                         continue;
 752                 }
 753
 754                 if (cl_page_own_try(env, io, page) == 0) {
 755                         if (!cl_page_in_use_noref(page)) {
 756                                 /* remove it from lru list earlier to avoid
 757                                  * lock contention */
 758                                 __osc_lru_del(cli, opg);
 759                                 opg->ops_in_lru = 0; /* will be discarded */
 760
 761                                 cl_page_get(page);
 762                                 will_free = true;
 763                         } else {
 764                                 cl_page_disown(env, io, page);
 765                         }
 766                 }
 767
 768                 if (!will_free) {
 769                         cfs_list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
 770                         continue;
 771                 }
 772
 773                 /* Don't discard and free the page with cl_lru_list held */
 774                 pvec[index++] = page;
 775                 if (unlikely(index == OTI_PVEC_SIZE)) {
 776                         client_obd_list_unlock(&cli->cl_lru_list_lock);
 777                         discard_pagevec(env, io, pvec, index);
 778                         index = 0;
 779
 780                         client_obd_list_lock(&cli->cl_lru_list_lock);
 781                 }
 782
 783                 if (++count >= target)
 784                         break;
 785         }
 786         client_obd_list_unlock(&cli->cl_lru_list_lock);
 787
 788         if (clobj != NULL) {
 789                 discard_pagevec(env, io, pvec, index);
 790
 791                 cl_io_fini(env, io);
 792                 cl_object_put(env, clobj);
 793         }
 794
 795         cfs_atomic_dec(&cli->cl_lru_shrinkers);
 796         if (count > 0) {
 797                 cfs_atomic_add(count, cli->cl_lru_left);
 798                 wake_up_all(&osc_lru_waitq);
 799         }
 800         RETURN(count > 0 ? count : rc);
 801 }
 802
 803 static inline int max_to_shrink(struct client_obd *cli)
 804 {
 805         return min(cfs_atomic_read(&cli->cl_lru_in_list) >> 1, lru_shrink_max);
 806 }
 807
 808 int osc_lru_reclaim(struct client_obd *cli)
 809 {
 810         struct cl_env_nest nest;
 811         struct lu_env *env;
 812         struct cl_client_cache *cache = cli->cl_cache;
 813         int max_scans;
 814         int rc = 0;
 815         ENTRY;
 816
 817         LASSERT(cache != NULL);
 818         LASSERT(!cfs_list_empty(&cache->ccc_lru));
 819
 820         env = cl_env_nested_get(&nest);
 821         if (IS_ERR(env))
 822                 RETURN(rc);
 823
 824         rc = osc_lru_shrink(env, cli, osc_cache_too_much(cli), false);
 825         if (rc != 0) {
 826                 if (rc == -EBUSY)
 827                         rc = 0;
 828
 829                 CDEBUG(D_CACHE, "%s: Free %d pages from own LRU: %p.\n",
 830                         cli->cl_import->imp_obd->obd_name, rc, cli);
 831                 GOTO(out, rc);
 832         }
 833
 834         CDEBUG(D_CACHE, "%s: cli %p no free slots, pages: %d, busy: %d.\n",
 835                 cli->cl_import->imp_obd->obd_name, cli,
 836                 cfs_atomic_read(&cli->cl_lru_in_list),
 837                 cfs_atomic_read(&cli->cl_lru_busy));
 838
 839         /* Reclaim LRU slots from other client_obd as it can't free enough
 840          * from its own. This should rarely happen. */
 841         spin_lock(&cache->ccc_lru_lock);
 842         cache->ccc_lru_shrinkers++;
 843         cfs_list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
 844
 845         max_scans = cfs_atomic_read(&cache->ccc_users);
 846         while (--max_scans > 0 && !cfs_list_empty(&cache->ccc_lru)) {
 847                 cli = cfs_list_entry(cache->ccc_lru.next, struct client_obd,
 848                                         cl_lru_osc);
 849
 850                 CDEBUG(D_CACHE, "%s: cli %p LRU pages: %d, busy: %d.\n",
 851                         cli->cl_import->imp_obd->obd_name, cli,
 852                         cfs_atomic_read(&cli->cl_lru_in_list),
 853                         cfs_atomic_read(&cli->cl_lru_busy));
 854
 855                 cfs_list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
 856                 if (osc_cache_too_much(cli) > 0) {
 857                         spin_unlock(&cache->ccc_lru_lock);
 858
 859                         rc = osc_lru_shrink(env, cli, osc_cache_too_much(cli),
 860                                             true);
 861                         spin_lock(&cache->ccc_lru_lock);
 862                         if (rc != 0)
 863                                 break;
 864                 }
 865         }
 866         spin_unlock(&cache->ccc_lru_lock);
 867
 868 out:
 869         cl_env_nested_put(&nest, env);
 870         CDEBUG(D_CACHE, "%s: cli %p freed %d pages.\n",
 871                 cli->cl_import->imp_obd->obd_name, cli, rc);
 872         return rc;
 873 }
 874
 875 static int osc_lru_reserve(const struct lu_env *env, struct osc_object *obj,
 876                            struct osc_page *opg)
 877 {
 878         struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
 879         struct osc_io *oio = osc_env_io(env);
 880         struct client_obd *cli = osc_cli(obj);
 881         int rc = 0;
 882         ENTRY;
 883
 884         if (cli->cl_cache == NULL) /* shall not be in LRU */
 885                 RETURN(0);
 886
 887         if (oio->oi_lru_reserved > 0) {
 888                 --oio->oi_lru_reserved;
 889                 goto out;
 890         }
 891
 892         LASSERT(cfs_atomic_read(cli->cl_lru_left) >= 0);
 893         while (!cfs_atomic_add_unless(cli->cl_lru_left, -1, 0)) {
 894
 895                 /* run out of LRU spaces, try to drop some by itself */
 896                 rc = osc_lru_reclaim(cli);
 897                 if (rc < 0)
 898                         break;
 899                 if (rc > 0)
 900                         continue;
 901
 902                 cond_resched();
 903                 rc = l_wait_event(osc_lru_waitq,
 904                                 cfs_atomic_read(cli->cl_lru_left) > 0,
 905                                 &lwi);
 906                 if (rc < 0)
 907                         break;
 908         }
 909
 910 out:
 911         if (rc >= 0) {
 912                 cfs_atomic_inc(&cli->cl_lru_busy);
 913                 opg->ops_in_lru = 1;
 914                 rc = 0;
 915         }
 916
 917         RETURN(rc);
 918 }
 919
 920 /** @} osc */