lustre/osc/osc_page.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2014, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * Implementation of cl_page for OSC layer.
  37  *
  38  *   Author: Nikita Danilov <nikita.danilov@sun.com>
  39  *   Author: Jinshan Xiong <jinshan.xiong@intel.com>
  40  */
  41
  42 #define DEBUG_SUBSYSTEM S_OSC
  43
  44 #include "osc_cl_internal.h"
  45
  46 static void osc_lru_del(struct client_obd *cli, struct osc_page *opg);
  47 static void osc_lru_use(struct client_obd *cli, struct osc_page *opg);
  48 static int osc_lru_reserve(const struct lu_env *env, struct osc_object *obj,
  49                            struct osc_page *opg);
  50
  51 /** \addtogroup osc
  52  *  @{
  53  */
  54
  55 /*
  56  * Comment out osc_page_protected because it may sleep inside the
  57  * the client_obd_list_lock.
  58  * client_obd_list_lock -> osc_ap_completion -> osc_completion ->
  59  *   -> osc_page_protected -> osc_page_is_dlocked -> osc_match_base
  60  *   -> ldlm_lock_match -> sptlrpc_import_check_ctx -> sleep.
  61  */
  62 #if 0
  63 static int osc_page_is_dlocked(const struct lu_env *env,
  64                                const struct osc_page *opg,
  65                                enum cl_lock_mode mode, int pending, int unref)
  66 {
  67         struct cl_page         *page;
  68         struct osc_object      *obj;
  69         struct osc_thread_info *info;
  70         struct ldlm_res_id     *resname;
  71         struct lustre_handle   *lockh;
  72         ldlm_policy_data_t     *policy;
  73         ldlm_mode_t             dlmmode;
  74         __u64                   flags;
  75
  76         might_sleep();
  77
  78         info = osc_env_info(env);
  79         resname = &info->oti_resname;
  80         policy = &info->oti_policy;
  81         lockh = &info->oti_handle;
  82         page = opg->ops_cl.cpl_page;
  83         obj = cl2osc(opg->ops_cl.cpl_obj);
  84
  85         flags = LDLM_FL_TEST_LOCK | LDLM_FL_BLOCK_GRANTED;
  86         if (pending)
  87                 flags |= LDLM_FL_CBPENDING;
  88
  89         dlmmode = osc_cl_lock2ldlm(mode) | LCK_PW;
  90         osc_lock_build_res(env, obj, resname);
  91         osc_index2policy(policy, page->cp_obj, osc_index(opg), osc_index(opg));
  92         return osc_match_base(osc_export(obj), resname, LDLM_EXTENT, policy,
  93                               dlmmode, &flags, NULL, lockh, unref);
  94 }
  95
  96 /**
  97  * Checks an invariant that a page in the cache is covered by a lock, as
  98  * needed.
  99  */
 100 static int osc_page_protected(const struct lu_env *env,
 101                               const struct osc_page *opg,
 102                               enum cl_lock_mode mode, int unref)
 103 {
 104         struct cl_object_header *hdr;
 105         struct cl_lock          *scan;
 106         struct cl_page          *page;
 107         struct cl_lock_descr    *descr;
 108         int result;
 109
 110         LINVRNT(!opg->ops_temp);
 111
 112         page = opg->ops_cl.cpl_page;
 113         if (page->cp_owner != NULL &&
 114             cl_io_top(page->cp_owner)->ci_lockreq == CILR_NEVER)
 115                 /*
 116                  * If IO is done without locks (liblustre, or lloop), lock is
 117                  * not required.
 118                  */
 119                 result = 1;
 120         else
 121                 /* otherwise check for a DLM lock */
 122         result = osc_page_is_dlocked(env, opg, mode, 1, unref);
 123         if (result == 0) {
 124                 /* maybe this page is a part of a lockless io? */
 125                 hdr = cl_object_header(opg->ops_cl.cpl_obj);
 126                 descr = &osc_env_info(env)->oti_descr;
 127                 descr->cld_mode = mode;
 128                 descr->cld_start = osc_index(opg);
 129                 descr->cld_end   = osc_index(opg);
 130                 spin_lock(&hdr->coh_lock_guard);
 131                 list_for_each_entry(scan, &hdr->coh_locks, cll_linkage) {
 132                         /*
 133                          * Lock-less sub-lock has to be either in HELD state
 134                          * (when io is actively going on), or in CACHED state,
 135                          * when top-lock is being unlocked:
 136                          * cl_io_unlock()->cl_unuse()->...->lov_lock_unuse().
 137                          */
 138                         if ((scan->cll_state == CLS_HELD ||
 139                              scan->cll_state == CLS_CACHED) &&
 140                             cl_lock_ext_match(&scan->cll_descr, descr)) {
 141                                 struct osc_lock *olck;
 142
 143                                 olck = osc_lock_at(scan);
 144                                 result = osc_lock_is_lockless(olck);
 145                                 break;
 146                         }
 147                 }
 148                 spin_unlock(&hdr->coh_lock_guard);
 149         }
 150         return result;
 151 }
 152 #else
 153 static int osc_page_protected(const struct lu_env *env,
 154                               const struct osc_page *opg,
 155                               enum cl_lock_mode mode, int unref)
 156 {
 157         return 1;
 158 }
 159 #endif
 160
 161 /*****************************************************************************
 162  *
 163  * Page operations.
 164  *
 165  */
 166 static void osc_page_transfer_get(struct osc_page *opg, const char *label)
 167 {
 168         struct cl_page *page = opg->ops_cl.cpl_page;
 169
 170         LASSERT(!opg->ops_transfer_pinned);
 171         cl_page_get(page);
 172         lu_ref_add_atomic(&page->cp_reference, label, page);
 173         opg->ops_transfer_pinned = 1;
 174 }
 175
 176 static void osc_page_transfer_put(const struct lu_env *env,
 177                                   struct osc_page *opg)
 178 {
 179         struct cl_page *page = opg->ops_cl.cpl_page;
 180
 181         if (opg->ops_transfer_pinned) {
 182                 opg->ops_transfer_pinned = 0;
 183                 lu_ref_del(&page->cp_reference, "transfer", page);
 184                 cl_page_put(env, page);
 185         }
 186 }
 187
 188 /**
 189  * This is called once for every page when it is submitted for a transfer
 190  * either opportunistic (osc_page_cache_add()), or immediate
 191  * (osc_page_submit()).
 192  */
 193 static void osc_page_transfer_add(const struct lu_env *env,
 194                                   struct osc_page *opg, enum cl_req_type crt)
 195 {
 196         struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
 197
 198         /* ops_lru and ops_inflight share the same field, so take it from LRU
 199          * first and then use it as inflight. */
 200         osc_lru_use(osc_cli(obj), opg);
 201
 202         spin_lock(&obj->oo_seatbelt);
 203         list_add(&opg->ops_inflight, &obj->oo_inflight[crt]);
 204         opg->ops_submitter = current;
 205         spin_unlock(&obj->oo_seatbelt);
 206 }
 207
 208 int osc_page_cache_add(const struct lu_env *env,
 209                         const struct cl_page_slice *slice, struct cl_io *io)
 210 {
 211         struct osc_page *opg = cl2osc_page(slice);
 212         int result;
 213         ENTRY;
 214
 215         LINVRNT(osc_page_protected(env, opg, CLM_WRITE, 0));
 216
 217         osc_page_transfer_get(opg, "transfer\0cache");
 218         result = osc_queue_async_io(env, io, opg);
 219         if (result != 0)
 220                 osc_page_transfer_put(env, opg);
 221         else
 222                 osc_page_transfer_add(env, opg, CRT_WRITE);
 223
 224         RETURN(result);
 225 }
 226
 227 void osc_index2policy(ldlm_policy_data_t *policy, const struct cl_object *obj,
 228                       pgoff_t start, pgoff_t end)
 229 {
 230         memset(policy, 0, sizeof *policy);
 231         policy->l_extent.start = cl_offset(obj, start);
 232         policy->l_extent.end   = cl_offset(obj, end + 1) - 1;
 233 }
 234
 235 static int osc_page_is_under_lock(const struct lu_env *env,
 236                                   const struct cl_page_slice *slice,
 237                                   struct cl_io *unused, pgoff_t *max_index)
 238 {
 239         struct osc_page         *opg = cl2osc_page(slice);
 240         struct ldlm_lock        *dlmlock;
 241         int                     result = -ENODATA;
 242
 243         ENTRY;
 244         dlmlock = osc_dlmlock_at_pgoff(env, cl2osc(slice->cpl_obj),
 245                                        osc_index(opg), 1, 0);
 246         if (dlmlock != NULL) {
 247                 *max_index = cl_index(slice->cpl_obj,
 248                                       dlmlock->l_policy_data.l_extent.end);
 249                 LDLM_LOCK_PUT(dlmlock);
 250                 result = 0;
 251         }
 252         RETURN(result);
 253 }
 254
 255 static const char *osc_list(struct list_head *head)
 256 {
 257         return list_empty(head) ? "-" : "+";
 258 }
 259
 260 static inline cfs_time_t osc_submit_duration(struct osc_page *opg)
 261 {
 262         if (opg->ops_submit_time == 0)
 263                 return 0;
 264
 265         return (cfs_time_current() - opg->ops_submit_time);
 266 }
 267
 268 static int osc_page_print(const struct lu_env *env,
 269                           const struct cl_page_slice *slice,
 270                           void *cookie, lu_printer_t printer)
 271 {
 272         struct osc_page       *opg = cl2osc_page(slice);
 273         struct osc_async_page *oap = &opg->ops_oap;
 274         struct osc_object     *obj = cl2osc(slice->cpl_obj);
 275         struct client_obd     *cli = &osc_export(obj)->exp_obd->u.cli;
 276
 277         return (*printer)(env, cookie, LUSTRE_OSC_NAME"-page@%p %lu: "
 278                           "1< %#x %d %u %s %s > "
 279                           "2< "LPU64" %u %u %#x %#x | %p %p %p > "
 280                           "3< %s %p %d %lu %d > "
 281                           "4< %d %d %d %lu %s | %s %s %s %s > "
 282                           "5< %s %s %s %s | %d %s | %d %s %s>\n",
 283                           opg, osc_index(opg),
 284                           /* 1 */
 285                           oap->oap_magic, oap->oap_cmd,
 286                           oap->oap_interrupted,
 287                           osc_list(&oap->oap_pending_item),
 288                           osc_list(&oap->oap_rpc_item),
 289                           /* 2 */
 290                           oap->oap_obj_off, oap->oap_page_off, oap->oap_count,
 291                           oap->oap_async_flags, oap->oap_brw_flags,
 292                           oap->oap_request, oap->oap_cli, obj,
 293                           /* 3 */
 294                           osc_list(&opg->ops_inflight),
 295                           opg->ops_submitter, opg->ops_transfer_pinned,
 296                           osc_submit_duration(opg), opg->ops_srvlock,
 297                           /* 4 */
 298                           cli->cl_r_in_flight, cli->cl_w_in_flight,
 299                           cli->cl_max_rpcs_in_flight,
 300                           cli->cl_avail_grant,
 301                           osc_list(&cli->cl_cache_waiters),
 302                           osc_list(&cli->cl_loi_ready_list),
 303                           osc_list(&cli->cl_loi_hp_ready_list),
 304                           osc_list(&cli->cl_loi_write_list),
 305                           osc_list(&cli->cl_loi_read_list),
 306                           /* 5 */
 307                           osc_list(&obj->oo_ready_item),
 308                           osc_list(&obj->oo_hp_ready_item),
 309                           osc_list(&obj->oo_write_item),
 310                           osc_list(&obj->oo_read_item),
 311                           atomic_read(&obj->oo_nr_reads),
 312                           osc_list(&obj->oo_reading_exts),
 313                           atomic_read(&obj->oo_nr_writes),
 314                           osc_list(&obj->oo_hp_exts),
 315                           osc_list(&obj->oo_urgent_exts));
 316 }
 317
 318 static void osc_page_delete(const struct lu_env *env,
 319                             const struct cl_page_slice *slice)
 320 {
 321         struct osc_page   *opg = cl2osc_page(slice);
 322         struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
 323         int rc;
 324
 325         LINVRNT(opg->ops_temp || osc_page_protected(env, opg, CLM_READ, 1));
 326
 327         ENTRY;
 328         CDEBUG(D_TRACE, "%p\n", opg);
 329         osc_page_transfer_put(env, opg);
 330         rc = osc_teardown_async_page(env, obj, opg);
 331         if (rc) {
 332                 CL_PAGE_DEBUG(D_ERROR, env, slice->cpl_page,
 333                               "Trying to teardown failed: %d\n", rc);
 334                 LASSERT(0);
 335         }
 336
 337         spin_lock(&obj->oo_seatbelt);
 338         if (opg->ops_submitter != NULL) {
 339                 LASSERT(!list_empty(&opg->ops_inflight));
 340                 list_del_init(&opg->ops_inflight);
 341                 opg->ops_submitter = NULL;
 342         }
 343         spin_unlock(&obj->oo_seatbelt);
 344
 345         osc_lru_del(osc_cli(obj), opg);
 346
 347         if (slice->cpl_page->cp_type == CPT_CACHEABLE) {
 348                 void *value;
 349
 350                 spin_lock(&obj->oo_tree_lock);
 351                 value = radix_tree_delete(&obj->oo_tree, osc_index(opg));
 352                 if (value != NULL)
 353                         --obj->oo_npages;
 354                 spin_unlock(&obj->oo_tree_lock);
 355
 356                 LASSERT(ergo(value != NULL, value == opg));
 357         }
 358
 359         EXIT;
 360 }
 361
 362 static void osc_page_clip(const struct lu_env *env,
 363                           const struct cl_page_slice *slice,
 364                           int from, int to)
 365 {
 366         struct osc_page       *opg = cl2osc_page(slice);
 367         struct osc_async_page *oap = &opg->ops_oap;
 368
 369         LINVRNT(osc_page_protected(env, opg, CLM_READ, 0));
 370
 371         opg->ops_from = from;
 372         opg->ops_to   = to;
 373         spin_lock(&oap->oap_lock);
 374         oap->oap_async_flags |= ASYNC_COUNT_STABLE;
 375         spin_unlock(&oap->oap_lock);
 376 }
 377
 378 static int osc_page_cancel(const struct lu_env *env,
 379                            const struct cl_page_slice *slice)
 380 {
 381         struct osc_page *opg = cl2osc_page(slice);
 382         int rc = 0;
 383
 384         LINVRNT(osc_page_protected(env, opg, CLM_READ, 0));
 385
 386         /* Check if the transferring against this page
 387          * is completed, or not even queued. */
 388         if (opg->ops_transfer_pinned)
 389                 /* FIXME: may not be interrupted.. */
 390                 rc = osc_cancel_async_page(env, opg);
 391         LASSERT(ergo(rc == 0, opg->ops_transfer_pinned == 0));
 392         return rc;
 393 }
 394
 395 static int osc_page_flush(const struct lu_env *env,
 396                           const struct cl_page_slice *slice,
 397                           struct cl_io *io)
 398 {
 399         struct osc_page *opg = cl2osc_page(slice);
 400         int rc = 0;
 401         ENTRY;
 402         rc = osc_flush_async_page(env, io, opg);
 403         RETURN(rc);
 404 }
 405
 406 static const struct cl_page_operations osc_page_ops = {
 407         .cpo_print         = osc_page_print,
 408         .cpo_delete        = osc_page_delete,
 409         .cpo_is_under_lock = osc_page_is_under_lock,
 410         .cpo_clip           = osc_page_clip,
 411         .cpo_cancel         = osc_page_cancel,
 412         .cpo_flush          = osc_page_flush
 413 };
 414
 415 int osc_page_init(const struct lu_env *env, struct cl_object *obj,
 416                   struct cl_page *page, pgoff_t index)
 417 {
 418         struct osc_object *osc = cl2osc(obj);
 419         struct osc_page   *opg = cl_object_page_slice(obj, page);
 420         int result;
 421
 422         opg->ops_from = 0;
 423         opg->ops_to   = PAGE_CACHE_SIZE;
 424
 425         result = osc_prep_async_page(osc, opg, page->cp_vmpage,
 426                                      cl_offset(obj, index));
 427         if (result == 0) {
 428                 struct osc_io *oio = osc_env_io(env);
 429                 opg->ops_srvlock = osc_io_srvlock(oio);
 430                 cl_page_slice_add(page, &opg->ops_cl, obj, index,
 431                                   &osc_page_ops);
 432         }
 433         /*
 434          * Cannot assert osc_page_protected() here as read-ahead
 435          * creates temporary pages outside of a lock.
 436          */
 437 #ifdef CONFIG_LUSTRE_DEBUG_EXPENSIVE_CHECK
 438         opg->ops_temp = !osc_page_protected(env, opg, CLM_READ, 1);
 439 #endif
 440         /* ops_inflight and ops_lru are the same field, but it doesn't
 441          * hurt to initialize it twice :-) */
 442         INIT_LIST_HEAD(&opg->ops_inflight);
 443         INIT_LIST_HEAD(&opg->ops_lru);
 444
 445         /* reserve an LRU space for this page */
 446         if (page->cp_type == CPT_CACHEABLE && result == 0) {
 447                 result = osc_lru_reserve(env, osc, opg);
 448                 if (result == 0) {
 449                         spin_lock(&osc->oo_tree_lock);
 450                         result = radix_tree_insert(&osc->oo_tree, index, opg);
 451                         if (result == 0)
 452                                 ++osc->oo_npages;
 453                         spin_unlock(&osc->oo_tree_lock);
 454                         LASSERT(result == 0);
 455                 }
 456         }
 457
 458         return result;
 459 }
 460
 461 /**
 462  * Helper function called by osc_io_submit() for every page in an immediate
 463  * transfer (i.e., transferred synchronously).
 464  */
 465 void osc_page_submit(const struct lu_env *env, struct osc_page *opg,
 466                      enum cl_req_type crt, int brw_flags)
 467 {
 468         struct osc_async_page *oap = &opg->ops_oap;
 469         struct osc_object     *obj = oap->oap_obj;
 470
 471         LINVRNT(osc_page_protected(env, opg,
 472                                    crt == CRT_WRITE ? CLM_WRITE : CLM_READ, 1));
 473
 474         LASSERTF(oap->oap_magic == OAP_MAGIC, "Bad oap magic: oap %p, "
 475                  "magic 0x%x\n", oap, oap->oap_magic);
 476         LASSERT(oap->oap_async_flags & ASYNC_READY);
 477         LASSERT(oap->oap_async_flags & ASYNC_COUNT_STABLE);
 478
 479         oap->oap_cmd       = crt == CRT_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ;
 480         oap->oap_page_off  = opg->ops_from;
 481         oap->oap_count     = opg->ops_to - opg->ops_from;
 482         oap->oap_brw_flags = OBD_BRW_SYNC | brw_flags;
 483
 484         if (!client_is_remote(osc_export(obj)) &&
 485                         cfs_capable(CFS_CAP_SYS_RESOURCE)) {
 486                 oap->oap_brw_flags |= OBD_BRW_NOQUOTA;
 487                 oap->oap_cmd |= OBD_BRW_NOQUOTA;
 488         }
 489
 490         opg->ops_submit_time = cfs_time_current();
 491         osc_page_transfer_get(opg, "transfer\0imm");
 492         osc_page_transfer_add(env, opg, crt);
 493 }
 494
 495 /* --------------- LRU page management ------------------ */
 496
 497 /* OSC is a natural place to manage LRU pages as applications are specialized
 498  * to write OSC by OSC. Ideally, if one OSC is used more frequently it should
 499  * occupy more LRU slots. On the other hand, we should avoid using up all LRU
 500  * slots (client_obd::cl_lru_left) otherwise process has to be put into sleep
 501  * for free LRU slots - this will be very bad so the algorithm requires each
 502  * OSC to free slots voluntarily to maintain a reasonable number of free slots
 503  * at any time.
 504  */
 505
 506 static CFS_DECL_WAITQ(osc_lru_waitq);
 507 /* LRU pages are freed in batch mode. OSC should at least free this
 508  * number of pages to avoid running out of LRU budget, and.. */
 509 static const int lru_shrink_min = 2 << (20 - PAGE_CACHE_SHIFT); /* 2M */
 510 /* free this number at most otherwise it will take too long time to finsih. */
 511 static const int lru_shrink_max = 8 << (20 - PAGE_CACHE_SHIFT); /* 8M */
 512
 513 /* Check if we can free LRU slots from this OSC. If there exists LRU waiters,
 514  * we should free slots aggressively. In this way, slots are freed in a steady
 515  * step to maintain fairness among OSCs.
 516  *
 517  * Return how many LRU pages should be freed. */
 518 static int osc_cache_too_much(struct client_obd *cli)
 519 {
 520         struct cl_client_cache *cache = cli->cl_cache;
 521         long pages = atomic_long_read(&cli->cl_lru_in_list);
 522         unsigned long budget;
 523
 524         budget = cache->ccc_lru_max / atomic_read(&cache->ccc_users);
 525
 526         /* if it's going to run out LRU slots, we should free some, but not
 527          * too much to maintain faireness among OSCs. */
 528         if (atomic_long_read(cli->cl_lru_left) < cache->ccc_lru_max >> 4) {
 529                 if (pages >= budget)
 530                         return lru_shrink_max;
 531                 else if (pages >= budget / 2)
 532                         return lru_shrink_min;
 533         } else if (pages >= budget * 2)
 534                 return lru_shrink_min;
 535         return 0;
 536 }
 537
 538 int lru_queue_work(const struct lu_env *env, void *data)
 539 {
 540         struct client_obd *cli = data;
 541
 542         CDEBUG(D_CACHE, "Run LRU work for client obd %p.\n", cli);
 543
 544         if (osc_cache_too_much(cli))
 545                 osc_lru_shrink(env, cli, lru_shrink_max, true);
 546
 547         RETURN(0);
 548 }
 549
 550 void osc_lru_add_batch(struct client_obd *cli, struct list_head *plist)
 551 {
 552         struct list_head lru = LIST_HEAD_INIT(lru);
 553         struct osc_async_page *oap;
 554         long npages = 0;
 555
 556         list_for_each_entry(oap, plist, oap_pending_item) {
 557                 struct osc_page *opg = oap2osc_page(oap);
 558
 559                 if (!opg->ops_in_lru)
 560                         continue;
 561
 562                 ++npages;
 563                 LASSERT(list_empty(&opg->ops_lru));
 564                 list_add(&opg->ops_lru, &lru);
 565         }
 566
 567         if (npages > 0) {
 568                 spin_lock(&cli->cl_lru_list_lock);
 569                 list_splice_tail(&lru, &cli->cl_lru_list);
 570                 atomic_long_sub(npages, &cli->cl_lru_busy);
 571                 atomic_long_add(npages, &cli->cl_lru_in_list);
 572                 spin_unlock(&cli->cl_lru_list_lock);
 573
 574                 /* XXX: May set force to be true for better performance */
 575                 if (osc_cache_too_much(cli))
 576                         (void)ptlrpcd_queue_work(cli->cl_lru_work);
 577         }
 578 }
 579
 580 static void __osc_lru_del(struct client_obd *cli, struct osc_page *opg)
 581 {
 582         LASSERT(atomic_long_read(&cli->cl_lru_in_list) > 0);
 583         list_del_init(&opg->ops_lru);
 584         atomic_long_dec(&cli->cl_lru_in_list);
 585 }
 586
 587 /**
 588  * Page is being destroyed. The page may be not in LRU list, if the transfer
 589  * has never finished(error occurred).
 590  */
 591 static void osc_lru_del(struct client_obd *cli, struct osc_page *opg)
 592 {
 593         if (opg->ops_in_lru) {
 594                 spin_lock(&cli->cl_lru_list_lock);
 595                 if (!list_empty(&opg->ops_lru)) {
 596                         __osc_lru_del(cli, opg);
 597                 } else {
 598                         LASSERT(atomic_long_read(&cli->cl_lru_busy) > 0);
 599                         atomic_long_dec(&cli->cl_lru_busy);
 600                 }
 601                 spin_unlock(&cli->cl_lru_list_lock);
 602
 603                 atomic_long_inc(cli->cl_lru_left);
 604                 /* this is a great place to release more LRU pages if
 605                  * this osc occupies too many LRU pages and kernel is
 606                  * stealing one of them. */
 607                 if (!memory_pressure_get())
 608                         (void)ptlrpcd_queue_work(cli->cl_lru_work);
 609                 wake_up(&osc_lru_waitq);
 610         } else {
 611                 LASSERT(list_empty(&opg->ops_lru));
 612         }
 613 }
 614
 615 /**
 616  * Delete page from LRUlist for redirty.
 617  */
 618 static void osc_lru_use(struct client_obd *cli, struct osc_page *opg)
 619 {
 620         /* If page is being transfered for the first time,
 621          * ops_lru should be empty */
 622         if (opg->ops_in_lru && !list_empty(&opg->ops_lru)) {
 623                 spin_lock(&cli->cl_lru_list_lock);
 624                 __osc_lru_del(cli, opg);
 625                 spin_unlock(&cli->cl_lru_list_lock);
 626                 atomic_long_inc(&cli->cl_lru_busy);
 627         }
 628 }
 629
 630 static void discard_pagevec(const struct lu_env *env, struct cl_io *io,
 631                                 struct cl_page **pvec, int max_index)
 632 {
 633         int i;
 634
 635         for (i = 0; i < max_index; i++) {
 636                 struct cl_page *page = pvec[i];
 637
 638                 LASSERT(cl_page_is_owned(page, io));
 639                 cl_page_discard(env, io, page);
 640                 cl_page_disown(env, io, page);
 641                 cl_page_put(env, page);
 642
 643                 pvec[i] = NULL;
 644         }
 645 }
 646
 647 /**
 648  * Check if a cl_page can be released, i.e, it's not being used.
 649  *
 650  * If unstable account is turned on, bulk transfer may hold one refcount
 651  * for recovery so we need to check vmpage refcount as well; otherwise,
 652  * even we can destroy cl_page but the corresponding vmpage can't be reused.
 653  */
 654 static inline bool lru_page_busy(struct client_obd *cli, struct cl_page *page)
 655 {
 656         if (cl_page_in_use_noref(page))
 657                 return true;
 658
 659         if (cli->cl_cache->ccc_unstable_check) {
 660                 struct page *vmpage = cl_page_vmpage(page);
 661
 662                 /* vmpage have two known users: cl_page and VM page cache */
 663                 if (page_count(vmpage) - page_mapcount(vmpage) > 2)
 664                         return true;
 665         }
 666         return false;
 667 }
 668
 669 /**
 670  * Drop @target of pages from LRU at most.
 671  */
 672 long osc_lru_shrink(const struct lu_env *env, struct client_obd *cli,
 673                    long target, bool force)
 674 {
 675         struct cl_io *io;
 676         struct cl_object *clobj = NULL;
 677         struct cl_page **pvec;
 678         struct osc_page *opg;
 679         long count = 0;
 680         int maxscan = 0;
 681         int index = 0;
 682         int rc = 0;
 683         ENTRY;
 684
 685         LASSERT(atomic_long_read(&cli->cl_lru_in_list) >= 0);
 686         if (atomic_long_read(&cli->cl_lru_in_list) == 0 || target <= 0)
 687                 RETURN(0);
 688
 689         if (!force) {
 690                 if (atomic_read(&cli->cl_lru_shrinkers) > 0)
 691                         RETURN(-EBUSY);
 692
 693                 if (atomic_inc_return(&cli->cl_lru_shrinkers) > 1) {
 694                         atomic_dec(&cli->cl_lru_shrinkers);
 695                         RETURN(-EBUSY);
 696                 }
 697         } else {
 698                 atomic_inc(&cli->cl_lru_shrinkers);
 699         }
 700
 701         pvec = (struct cl_page **)osc_env_info(env)->oti_pvec;
 702         io = &osc_env_info(env)->oti_io;
 703
 704         spin_lock(&cli->cl_lru_list_lock);
 705         maxscan = min(target << 1, atomic_long_read(&cli->cl_lru_in_list));
 706         while (!list_empty(&cli->cl_lru_list)) {
 707                 struct cl_page *page;
 708                 bool will_free = false;
 709
 710                 if (--maxscan < 0)
 711                         break;
 712
 713                 opg = list_entry(cli->cl_lru_list.next, struct osc_page,
 714                                  ops_lru);
 715                 page = opg->ops_cl.cpl_page;
 716                 if (lru_page_busy(cli, page)) {
 717                         list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
 718                         continue;
 719                 }
 720
 721                 LASSERT(page->cp_obj != NULL);
 722                 if (clobj != page->cp_obj) {
 723                         struct cl_object *tmp = page->cp_obj;
 724
 725                         cl_object_get(tmp);
 726                         spin_unlock(&cli->cl_lru_list_lock);
 727
 728                         if (clobj != NULL) {
 729                                 discard_pagevec(env, io, pvec, index);
 730                                 index = 0;
 731
 732                                 cl_io_fini(env, io);
 733                                 cl_object_put(env, clobj);
 734                                 clobj = NULL;
 735                         }
 736
 737                         clobj = tmp;
 738                         io->ci_obj = clobj;
 739                         io->ci_ignore_layout = 1;
 740                         rc = cl_io_init(env, io, CIT_MISC, clobj);
 741
 742                         spin_lock(&cli->cl_lru_list_lock);
 743
 744                         if (rc != 0)
 745                                 break;
 746
 747                         ++maxscan;
 748                         continue;
 749                 }
 750
 751                 if (cl_page_own_try(env, io, page) == 0) {
 752                         if (!lru_page_busy(cli, page)) {
 753                                 /* remove it from lru list earlier to avoid
 754                                  * lock contention */
 755                                 __osc_lru_del(cli, opg);
 756                                 opg->ops_in_lru = 0; /* will be discarded */
 757
 758                                 cl_page_get(page);
 759                                 will_free = true;
 760                         } else {
 761                                 cl_page_disown(env, io, page);
 762                         }
 763                 }
 764
 765                 if (!will_free) {
 766                         list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
 767                         continue;
 768                 }
 769
 770                 /* Don't discard and free the page with cl_lru_list held */
 771                 pvec[index++] = page;
 772                 if (unlikely(index == OTI_PVEC_SIZE)) {
 773                         spin_unlock(&cli->cl_lru_list_lock);
 774                         discard_pagevec(env, io, pvec, index);
 775                         index = 0;
 776
 777                         spin_lock(&cli->cl_lru_list_lock);
 778                 }
 779
 780                 if (++count >= target)
 781                         break;
 782         }
 783         spin_unlock(&cli->cl_lru_list_lock);
 784
 785         if (clobj != NULL) {
 786                 discard_pagevec(env, io, pvec, index);
 787
 788                 cl_io_fini(env, io);
 789                 cl_object_put(env, clobj);
 790         }
 791
 792         atomic_dec(&cli->cl_lru_shrinkers);
 793         if (count > 0) {
 794                 atomic_long_add(count, cli->cl_lru_left);
 795                 wake_up_all(&osc_lru_waitq);
 796         }
 797         RETURN(count > 0 ? count : rc);
 798 }
 799
 800 long osc_lru_reclaim(struct client_obd *cli)
 801 {
 802         struct cl_env_nest nest;
 803         struct lu_env *env;
 804         struct cl_client_cache *cache = cli->cl_cache;
 805         long rc = 0;
 806         int max_scans;
 807         ENTRY;
 808
 809         LASSERT(cache != NULL);
 810         LASSERT(!list_empty(&cache->ccc_lru));
 811
 812         env = cl_env_nested_get(&nest);
 813         if (IS_ERR(env))
 814                 RETURN(rc);
 815
 816         rc = osc_lru_shrink(env, cli, osc_cache_too_much(cli), false);
 817         if (rc != 0) {
 818                 if (rc == -EBUSY)
 819                         rc = 0;
 820
 821                 CDEBUG(D_CACHE, "%s: Free %ld pages from own LRU: %p.\n",
 822                         cli->cl_import->imp_obd->obd_name, rc, cli);
 823                 GOTO(out, rc);
 824         }
 825
 826         CDEBUG(D_CACHE, "%s: cli %p no free slots, pages: %ld, busy: %ld.\n",
 827                 cli->cl_import->imp_obd->obd_name, cli,
 828                 atomic_long_read(&cli->cl_lru_in_list),
 829                 atomic_long_read(&cli->cl_lru_busy));
 830
 831         /* Reclaim LRU slots from other client_obd as it can't free enough
 832          * from its own. This should rarely happen. */
 833         spin_lock(&cache->ccc_lru_lock);
 834         cache->ccc_lru_shrinkers++;
 835         list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
 836
 837         max_scans = atomic_read(&cache->ccc_users);
 838         while (--max_scans > 0 && !list_empty(&cache->ccc_lru)) {
 839                 cli = list_entry(cache->ccc_lru.next, struct client_obd,
 840                                  cl_lru_osc);
 841
 842                 CDEBUG(D_CACHE, "%s: cli %p LRU pages: %ld, busy: %ld.\n",
 843                         cli->cl_import->imp_obd->obd_name, cli,
 844                         atomic_long_read(&cli->cl_lru_in_list),
 845                         atomic_long_read(&cli->cl_lru_busy));
 846
 847                 list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
 848                 if (osc_cache_too_much(cli) > 0) {
 849                         spin_unlock(&cache->ccc_lru_lock);
 850
 851                         rc = osc_lru_shrink(env, cli, osc_cache_too_much(cli),
 852                                             true);
 853                         spin_lock(&cache->ccc_lru_lock);
 854                         if (rc != 0)
 855                                 break;
 856                 }
 857         }
 858         spin_unlock(&cache->ccc_lru_lock);
 859
 860 out:
 861         cl_env_nested_put(&nest, env);
 862         CDEBUG(D_CACHE, "%s: cli %p freed %ld pages.\n",
 863                 cli->cl_import->imp_obd->obd_name, cli, rc);
 864         return rc;
 865 }
 866
 867 /**
 868  * osc_lru_reserve() is called to reserve an LRU slot for a cl_page.
 869  *
 870  * Usually the LRU slots are reserved in osc_io_iter_rw_init().
 871  * Only in the case that the LRU slots are in extreme shortage, it should
 872  * have reserved enough slots for an IO.
 873  */
 874 static int osc_lru_reserve(const struct lu_env *env, struct osc_object *obj,
 875                            struct osc_page *opg)
 876 {
 877         struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
 878         struct osc_io *oio = osc_env_io(env);
 879         struct client_obd *cli = osc_cli(obj);
 880         int rc = 0;
 881         ENTRY;
 882
 883         if (cli->cl_cache == NULL) /* shall not be in LRU */
 884                 RETURN(0);
 885
 886         if (oio->oi_lru_reserved > 0) {
 887                 --oio->oi_lru_reserved;
 888                 goto out;
 889         }
 890
 891         LASSERT(atomic_long_read(cli->cl_lru_left) >= 0);
 892         while (!atomic_long_add_unless(cli->cl_lru_left, -1, 0)) {
 893
 894                 /* run out of LRU spaces, try to drop some by itself */
 895                 rc = osc_lru_reclaim(cli);
 896                 if (rc < 0)
 897                         break;
 898                 if (rc > 0)
 899                         continue;
 900
 901                 cond_resched();
 902                 rc = l_wait_event(osc_lru_waitq,
 903                                 atomic_long_read(cli->cl_lru_left) > 0,
 904                                 &lwi);
 905                 if (rc < 0)
 906                         break;
 907         }
 908
 909 out:
 910         if (rc >= 0) {
 911                 atomic_long_inc(&cli->cl_lru_busy);
 912                 opg->ops_in_lru = 1;
 913                 rc = 0;
 914         }
 915
 916         RETURN(rc);
 917 }
 918
 919 /**
 920  * Atomic operations are expensive. We accumulate the accounting for the
 921  * same page zone to get better performance.
 922  * In practice this can work pretty good because the pages in the same RPC
 923  * are likely from the same page zone.
 924  */
 925 static inline void unstable_page_accounting(struct ptlrpc_bulk_desc *desc,
 926                                             int factor)
 927 {
 928         obd_count page_count = desc->bd_iov_count;
 929         void *zone = NULL;
 930         int count = 0;
 931         int i;
 932
 933         for (i = 0; i < page_count; i++) {
 934                 void *pz = page_zone(desc->bd_iov[i].kiov_page);
 935
 936                 if (likely(pz == zone)) {
 937                         ++count;
 938                         continue;
 939                 }
 940
 941                 if (count > 0) {
 942                         mod_zone_page_state(zone, NR_UNSTABLE_NFS,
 943                                             factor * count);
 944                         count = 0;
 945                 }
 946                 zone = pz;
 947                 ++count;
 948         }
 949         if (count > 0)
 950                 mod_zone_page_state(zone, NR_UNSTABLE_NFS, factor * count);
 951 }
 952
 953 static inline void add_unstable_page_accounting(struct ptlrpc_bulk_desc *desc)
 954 {
 955         unstable_page_accounting(desc, 1);
 956 }
 957
 958 static inline void dec_unstable_page_accounting(struct ptlrpc_bulk_desc *desc)
 959 {
 960         unstable_page_accounting(desc, -1);
 961 }
 962
 963 /**
 964  * Performs "unstable" page accounting. This function balances the
 965  * increment operations performed in osc_inc_unstable_pages. It is
 966  * registered as the RPC request callback, and is executed when the
 967  * bulk RPC is committed on the server. Thus at this point, the pages
 968  * involved in the bulk transfer are no longer considered unstable.
 969  *
 970  * If this function is called, the request should have been committed
 971  * or req:rq_unstable must have been set; it implies that the unstable
 972  * statistic have been added.
 973  */
 974 void osc_dec_unstable_pages(struct ptlrpc_request *req)
 975 {
 976         struct ptlrpc_bulk_desc *desc       = req->rq_bulk;
 977         struct client_obd       *cli        = &req->rq_import->imp_obd->u.cli;
 978         int                      page_count = desc->bd_iov_count;
 979         long                     unstable_count;
 980
 981         LASSERT(page_count >= 0);
 982         dec_unstable_page_accounting(desc);
 983
 984         unstable_count = atomic_long_sub_return(page_count,
 985                                                 &cli->cl_unstable_count);
 986         LASSERT(unstable_count >= 0);
 987
 988         unstable_count = atomic_long_sub_return(page_count,
 989                                            &cli->cl_cache->ccc_unstable_nr);
 990         LASSERT(unstable_count >= 0);
 991         if (unstable_count == 0)
 992                 wake_up_all(&cli->cl_cache->ccc_unstable_waitq);
 993
 994         if (osc_cache_too_much(cli))
 995                 (void)ptlrpcd_queue_work(cli->cl_lru_work);
 996 }
 997
 998 /**
 999  * "unstable" page accounting. See: osc_dec_unstable_pages.
1000  */
1001 void osc_inc_unstable_pages(struct ptlrpc_request *req)
1002 {
1003         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
1004         struct client_obd       *cli  = &req->rq_import->imp_obd->u.cli;
1005         long                     page_count = desc->bd_iov_count;
1006
1007         /* No unstable page tracking */
1008         if (cli->cl_cache == NULL || !cli->cl_cache->ccc_unstable_check)
1009                 return;
1010
1011         add_unstable_page_accounting(desc);
1012         atomic_long_add(page_count, &cli->cl_unstable_count);
1013         atomic_long_add(page_count, &cli->cl_cache->ccc_unstable_nr);
1014
1015         /* If the request has already been committed (i.e. brw_commit
1016          * called via rq_commit_cb), we need to undo the unstable page
1017          * increments we just performed because rq_commit_cb wont be
1018          * called again. */
1019         spin_lock(&req->rq_lock);
1020         if (unlikely(req->rq_committed)) {
1021                 spin_unlock(&req->rq_lock);
1022
1023                 osc_dec_unstable_pages(req);
1024         } else {
1025                 req->rq_unstable = 1;
1026                 spin_unlock(&req->rq_lock);
1027         }
1028 }
1029
1030 /**
1031  * Check if it piggybacks SOFT_SYNC flag to OST from this OSC.
1032  * This function will be called by every BRW RPC so it's critical
1033  * to make this function fast.
1034  */
1035 bool osc_over_unstable_soft_limit(struct client_obd *cli)
1036 {
1037         long unstable_nr, osc_unstable_count;
1038
1039         /* Can't check cli->cl_unstable_count, therefore, no soft limit */
1040         if (cli->cl_cache == NULL || !cli->cl_cache->ccc_unstable_check)
1041                 return false;
1042
1043         osc_unstable_count = atomic_long_read(&cli->cl_unstable_count);
1044         unstable_nr = atomic_long_read(&cli->cl_cache->ccc_unstable_nr);
1045
1046         CDEBUG(D_CACHE,
1047                "%s: cli: %p unstable pages: %lu, osc unstable pages: %lu\n",
1048                cli->cl_import->imp_obd->obd_name, cli,
1049                unstable_nr, osc_unstable_count);
1050
1051         /* If the LRU slots are in shortage - 25% remaining AND this OSC
1052          * has one full RPC window of unstable pages, it's a good chance
1053          * to piggyback a SOFT_SYNC flag.
1054          * Please notice that the OST won't take immediate response for the
1055          * SOFT_SYNC request so active OSCs will have more chance to carry
1056          * the flag, this is reasonable. */
1057         return unstable_nr > cli->cl_cache->ccc_lru_max >> 2 &&
1058                osc_unstable_count > cli->cl_max_pages_per_rpc *
1059                                     cli->cl_max_rpcs_in_flight;
1060 }
1061
1062 /** @} osc */