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                 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 void osc_page_clip(const struct lu_env *env, const struct cl_page_slice *slice,
 363                    int from, int to)
 364 {
 365         struct osc_page       *opg = cl2osc_page(slice);
 366         struct osc_async_page *oap = &opg->ops_oap;
 367
 368         LINVRNT(osc_page_protected(env, opg, CLM_READ, 0));
 369
 370         opg->ops_from = from;
 371         opg->ops_to   = to;
 372         spin_lock(&oap->oap_lock);
 373         oap->oap_async_flags |= ASYNC_COUNT_STABLE;
 374         spin_unlock(&oap->oap_lock);
 375 }
 376
 377 static int osc_page_cancel(const struct lu_env *env,
 378                            const struct cl_page_slice *slice)
 379 {
 380         struct osc_page *opg = cl2osc_page(slice);
 381         int rc = 0;
 382
 383         LINVRNT(osc_page_protected(env, opg, CLM_READ, 0));
 384
 385         /* Check if the transferring against this page
 386          * is completed, or not even queued. */
 387         if (opg->ops_transfer_pinned)
 388                 /* FIXME: may not be interrupted.. */
 389                 rc = osc_cancel_async_page(env, opg);
 390         LASSERT(ergo(rc == 0, opg->ops_transfer_pinned == 0));
 391         return rc;
 392 }
 393
 394 static int osc_page_flush(const struct lu_env *env,
 395                           const struct cl_page_slice *slice,
 396                           struct cl_io *io)
 397 {
 398         struct osc_page *opg = cl2osc_page(slice);
 399         int rc = 0;
 400         ENTRY;
 401         rc = osc_flush_async_page(env, io, opg);
 402         RETURN(rc);
 403 }
 404
 405 static const struct cl_page_operations osc_page_ops = {
 406         .cpo_print         = osc_page_print,
 407         .cpo_delete        = osc_page_delete,
 408         .cpo_is_under_lock = osc_page_is_under_lock,
 409         .cpo_clip           = osc_page_clip,
 410         .cpo_cancel         = osc_page_cancel,
 411         .cpo_flush          = osc_page_flush
 412 };
 413
 414 int osc_page_init(const struct lu_env *env, struct cl_object *obj,
 415                   struct cl_page *page, pgoff_t index)
 416 {
 417         struct osc_object *osc = cl2osc(obj);
 418         struct osc_page   *opg = cl_object_page_slice(obj, page);
 419         int result;
 420
 421         opg->ops_from = 0;
 422         opg->ops_to   = PAGE_CACHE_SIZE;
 423
 424         result = osc_prep_async_page(osc, opg, page->cp_vmpage,
 425                                      cl_offset(obj, index));
 426         if (result == 0) {
 427                 struct osc_io *oio = osc_env_io(env);
 428                 opg->ops_srvlock = osc_io_srvlock(oio);
 429                 cl_page_slice_add(page, &opg->ops_cl, obj, index,
 430                                   &osc_page_ops);
 431         }
 432         /*
 433          * Cannot assert osc_page_protected() here as read-ahead
 434          * creates temporary pages outside of a lock.
 435          */
 436 #ifdef CONFIG_LUSTRE_DEBUG_EXPENSIVE_CHECK
 437         opg->ops_temp = !osc_page_protected(env, opg, CLM_READ, 1);
 438 #endif
 439         /* ops_inflight and ops_lru are the same field, but it doesn't
 440          * hurt to initialize it twice :-) */
 441         INIT_LIST_HEAD(&opg->ops_inflight);
 442         INIT_LIST_HEAD(&opg->ops_lru);
 443
 444         /* reserve an LRU space for this page */
 445         if (page->cp_type == CPT_CACHEABLE && result == 0) {
 446                 result = osc_lru_reserve(env, osc, opg);
 447                 if (result == 0) {
 448                         spin_lock(&osc->oo_tree_lock);
 449                         result = radix_tree_insert(&osc->oo_tree, index, opg);
 450                         if (result == 0)
 451                                 ++osc->oo_npages;
 452                         spin_unlock(&osc->oo_tree_lock);
 453                         LASSERT(result == 0);
 454                 }
 455         }
 456
 457         return result;
 458 }
 459
 460 /**
 461  * Helper function called by osc_io_submit() for every page in an immediate
 462  * transfer (i.e., transferred synchronously).
 463  */
 464 void osc_page_submit(const struct lu_env *env, struct osc_page *opg,
 465                      enum cl_req_type crt, int brw_flags)
 466 {
 467         struct osc_async_page *oap = &opg->ops_oap;
 468         struct osc_object     *obj = oap->oap_obj;
 469
 470         LINVRNT(osc_page_protected(env, opg,
 471                                    crt == CRT_WRITE ? CLM_WRITE : CLM_READ, 1));
 472
 473         LASSERTF(oap->oap_magic == OAP_MAGIC, "Bad oap magic: oap %p, "
 474                  "magic 0x%x\n", oap, oap->oap_magic);
 475         LASSERT(oap->oap_async_flags & ASYNC_READY);
 476         LASSERT(oap->oap_async_flags & ASYNC_COUNT_STABLE);
 477
 478         oap->oap_cmd       = crt == CRT_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ;
 479         oap->oap_page_off  = opg->ops_from;
 480         oap->oap_count     = opg->ops_to - opg->ops_from;
 481         oap->oap_brw_flags = OBD_BRW_SYNC | brw_flags;
 482
 483         if (!client_is_remote(osc_export(obj)) &&
 484                         cfs_capable(CFS_CAP_SYS_RESOURCE)) {
 485                 oap->oap_brw_flags |= OBD_BRW_NOQUOTA;
 486                 oap->oap_cmd |= OBD_BRW_NOQUOTA;
 487         }
 488
 489         opg->ops_submit_time = cfs_time_current();
 490         osc_page_transfer_get(opg, "transfer\0imm");
 491         osc_page_transfer_add(env, opg, crt);
 492 }
 493
 494 /* --------------- LRU page management ------------------ */
 495
 496 /* OSC is a natural place to manage LRU pages as applications are specialized
 497  * to write OSC by OSC. Ideally, if one OSC is used more frequently it should
 498  * occupy more LRU slots. On the other hand, we should avoid using up all LRU
 499  * slots (client_obd::cl_lru_left) otherwise process has to be put into sleep
 500  * for free LRU slots - this will be very bad so the algorithm requires each
 501  * OSC to free slots voluntarily to maintain a reasonable number of free slots
 502  * at any time.
 503  */
 504
 505 static CFS_DECL_WAITQ(osc_lru_waitq);
 506 /* LRU pages are freed in batch mode. OSC should at least free this
 507  * number of pages to avoid running out of LRU budget, and.. */
 508 static const int lru_shrink_min = 2 << (20 - PAGE_CACHE_SHIFT); /* 2M */
 509 /* free this number at most otherwise it will take too long time to finsih. */
 510 static const int lru_shrink_max = 8 << (20 - PAGE_CACHE_SHIFT); /* 8M */
 511
 512 /* Check if we can free LRU slots from this OSC. If there exists LRU waiters,
 513  * we should free slots aggressively. In this way, slots are freed in a steady
 514  * step to maintain fairness among OSCs.
 515  *
 516  * Return how many LRU pages should be freed. */
 517 static int osc_cache_too_much(struct client_obd *cli)
 518 {
 519         struct cl_client_cache *cache = cli->cl_cache;
 520         long pages = atomic_long_read(&cli->cl_lru_in_list);
 521         unsigned long budget;
 522
 523         budget = cache->ccc_lru_max / atomic_read(&cache->ccc_users);
 524
 525         /* if it's going to run out LRU slots, we should free some, but not
 526          * too much to maintain faireness among OSCs. */
 527         if (atomic_long_read(cli->cl_lru_left) < cache->ccc_lru_max >> 4) {
 528                 if (pages >= budget)
 529                         return lru_shrink_max;
 530                 else if (pages >= budget / 2)
 531                         return lru_shrink_min;
 532         } else if (pages >= budget * 2)
 533                 return lru_shrink_min;
 534         return 0;
 535 }
 536
 537 int lru_queue_work(const struct lu_env *env, void *data)
 538 {
 539         struct client_obd *cli = data;
 540
 541         CDEBUG(D_CACHE, "Run LRU work for client obd %p.\n", cli);
 542
 543         if (osc_cache_too_much(cli))
 544                 osc_lru_shrink(env, cli, lru_shrink_max, true);
 545
 546         RETURN(0);
 547 }
 548
 549 void osc_lru_add_batch(struct client_obd *cli, struct list_head *plist)
 550 {
 551         struct list_head lru = LIST_HEAD_INIT(lru);
 552         struct osc_async_page *oap;
 553         long npages = 0;
 554
 555         list_for_each_entry(oap, plist, oap_pending_item) {
 556                 struct osc_page *opg = oap2osc_page(oap);
 557
 558                 if (!opg->ops_in_lru)
 559                         continue;
 560
 561                 ++npages;
 562                 LASSERT(list_empty(&opg->ops_lru));
 563                 list_add(&opg->ops_lru, &lru);
 564         }
 565
 566         if (npages > 0) {
 567                 spin_lock(&cli->cl_lru_list_lock);
 568                 list_splice_tail(&lru, &cli->cl_lru_list);
 569                 atomic_long_sub(npages, &cli->cl_lru_busy);
 570                 atomic_long_add(npages, &cli->cl_lru_in_list);
 571                 spin_unlock(&cli->cl_lru_list_lock);
 572
 573                 /* XXX: May set force to be true for better performance */
 574                 if (osc_cache_too_much(cli))
 575                         (void)ptlrpcd_queue_work(cli->cl_lru_work);
 576         }
 577 }
 578
 579 static void __osc_lru_del(struct client_obd *cli, struct osc_page *opg)
 580 {
 581         LASSERT(atomic_long_read(&cli->cl_lru_in_list) > 0);
 582         list_del_init(&opg->ops_lru);
 583         atomic_long_dec(&cli->cl_lru_in_list);
 584 }
 585
 586 /**
 587  * Page is being destroyed. The page may be not in LRU list, if the transfer
 588  * has never finished(error occurred).
 589  */
 590 static void osc_lru_del(struct client_obd *cli, struct osc_page *opg)
 591 {
 592         if (opg->ops_in_lru) {
 593                 spin_lock(&cli->cl_lru_list_lock);
 594                 if (!list_empty(&opg->ops_lru)) {
 595                         __osc_lru_del(cli, opg);
 596                 } else {
 597                         LASSERT(atomic_long_read(&cli->cl_lru_busy) > 0);
 598                         atomic_long_dec(&cli->cl_lru_busy);
 599                 }
 600                 spin_unlock(&cli->cl_lru_list_lock);
 601
 602                 atomic_long_inc(cli->cl_lru_left);
 603                 /* this is a great place to release more LRU pages if
 604                  * this osc occupies too many LRU pages and kernel is
 605                  * stealing one of them. */
 606                 if (!memory_pressure_get())
 607                         (void)ptlrpcd_queue_work(cli->cl_lru_work);
 608                 wake_up(&osc_lru_waitq);
 609         } else {
 610                 LASSERT(list_empty(&opg->ops_lru));
 611         }
 612 }
 613
 614 /**
 615  * Delete page from LRUlist for redirty.
 616  */
 617 static void osc_lru_use(struct client_obd *cli, struct osc_page *opg)
 618 {
 619         /* If page is being transfered for the first time,
 620          * ops_lru should be empty */
 621         if (opg->ops_in_lru && !list_empty(&opg->ops_lru)) {
 622                 spin_lock(&cli->cl_lru_list_lock);
 623                 __osc_lru_del(cli, opg);
 624                 spin_unlock(&cli->cl_lru_list_lock);
 625                 atomic_long_inc(&cli->cl_lru_busy);
 626         }
 627 }
 628
 629 static void discard_pagevec(const struct lu_env *env, struct cl_io *io,
 630                                 struct cl_page **pvec, int max_index)
 631 {
 632         int i;
 633
 634         for (i = 0; i < max_index; i++) {
 635                 struct cl_page *page = pvec[i];
 636
 637                 LASSERT(cl_page_is_owned(page, io));
 638                 cl_page_discard(env, io, page);
 639                 cl_page_disown(env, io, page);
 640                 cl_page_put(env, page);
 641
 642                 pvec[i] = NULL;
 643         }
 644 }
 645
 646 /**
 647  * Check if a cl_page can be released, i.e, it's not being used.
 648  *
 649  * If unstable account is turned on, bulk transfer may hold one refcount
 650  * for recovery so we need to check vmpage refcount as well; otherwise,
 651  * even we can destroy cl_page but the corresponding vmpage can't be reused.
 652  */
 653 static inline bool lru_page_busy(struct client_obd *cli, struct cl_page *page)
 654 {
 655         if (cl_page_in_use_noref(page))
 656                 return true;
 657
 658         if (cli->cl_cache->ccc_unstable_check) {
 659                 struct page *vmpage = cl_page_vmpage(page);
 660
 661                 /* vmpage have two known users: cl_page and VM page cache */
 662                 if (page_count(vmpage) - page_mapcount(vmpage) > 2)
 663                         return true;
 664         }
 665         return false;
 666 }
 667
 668 /**
 669  * Drop @target of pages from LRU at most.
 670  */
 671 long osc_lru_shrink(const struct lu_env *env, struct client_obd *cli,
 672                    long target, bool force)
 673 {
 674         struct cl_io *io;
 675         struct cl_object *clobj = NULL;
 676         struct cl_page **pvec;
 677         struct osc_page *opg;
 678         long count = 0;
 679         int maxscan = 0;
 680         int index = 0;
 681         int rc = 0;
 682         ENTRY;
 683
 684         LASSERT(atomic_long_read(&cli->cl_lru_in_list) >= 0);
 685         if (atomic_long_read(&cli->cl_lru_in_list) == 0 || target <= 0)
 686                 RETURN(0);
 687
 688         if (!force) {
 689                 if (atomic_read(&cli->cl_lru_shrinkers) > 0)
 690                         RETURN(-EBUSY);
 691
 692                 if (atomic_inc_return(&cli->cl_lru_shrinkers) > 1) {
 693                         atomic_dec(&cli->cl_lru_shrinkers);
 694                         RETURN(-EBUSY);
 695                 }
 696         } else {
 697                 atomic_inc(&cli->cl_lru_shrinkers);
 698         }
 699
 700         pvec = (struct cl_page **)osc_env_info(env)->oti_pvec;
 701         io = &osc_env_info(env)->oti_io;
 702
 703         spin_lock(&cli->cl_lru_list_lock);
 704         maxscan = min(target << 1, atomic_long_read(&cli->cl_lru_in_list));
 705         while (!list_empty(&cli->cl_lru_list)) {
 706                 struct cl_page *page;
 707                 bool will_free = false;
 708
 709                 if (--maxscan < 0)
 710                         break;
 711
 712                 opg = list_entry(cli->cl_lru_list.next, struct osc_page,
 713                                  ops_lru);
 714                 page = opg->ops_cl.cpl_page;
 715                 if (lru_page_busy(cli, page)) {
 716                         list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
 717                         continue;
 718                 }
 719
 720                 LASSERT(page->cp_obj != NULL);
 721                 if (clobj != page->cp_obj) {
 722                         struct cl_object *tmp = page->cp_obj;
 723
 724                         cl_object_get(tmp);
 725                         spin_unlock(&cli->cl_lru_list_lock);
 726
 727                         if (clobj != NULL) {
 728                                 discard_pagevec(env, io, pvec, index);
 729                                 index = 0;
 730
 731                                 cl_io_fini(env, io);
 732                                 cl_object_put(env, clobj);
 733                                 clobj = NULL;
 734                         }
 735
 736                         clobj = tmp;
 737                         io->ci_obj = clobj;
 738                         io->ci_ignore_layout = 1;
 739                         rc = cl_io_init(env, io, CIT_MISC, clobj);
 740
 741                         spin_lock(&cli->cl_lru_list_lock);
 742
 743                         if (rc != 0)
 744                                 break;
 745
 746                         ++maxscan;
 747                         continue;
 748                 }
 749
 750                 if (cl_page_own_try(env, io, page) == 0) {
 751                         if (!lru_page_busy(cli, page)) {
 752                                 /* remove it from lru list earlier to avoid
 753                                  * lock contention */
 754                                 __osc_lru_del(cli, opg);
 755                                 opg->ops_in_lru = 0; /* will be discarded */
 756
 757                                 cl_page_get(page);
 758                                 will_free = true;
 759                         } else {
 760                                 cl_page_disown(env, io, page);
 761                         }
 762                 }
 763
 764                 if (!will_free) {
 765                         list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
 766                         continue;
 767                 }
 768
 769                 /* Don't discard and free the page with cl_lru_list held */
 770                 pvec[index++] = page;
 771                 if (unlikely(index == OTI_PVEC_SIZE)) {
 772                         spin_unlock(&cli->cl_lru_list_lock);
 773                         discard_pagevec(env, io, pvec, index);
 774                         index = 0;
 775
 776                         spin_lock(&cli->cl_lru_list_lock);
 777                 }
 778
 779                 if (++count >= target)
 780                         break;
 781         }
 782         spin_unlock(&cli->cl_lru_list_lock);
 783
 784         if (clobj != NULL) {
 785                 discard_pagevec(env, io, pvec, index);
 786
 787                 cl_io_fini(env, io);
 788                 cl_object_put(env, clobj);
 789         }
 790
 791         atomic_dec(&cli->cl_lru_shrinkers);
 792         if (count > 0) {
 793                 atomic_long_add(count, cli->cl_lru_left);
 794                 wake_up_all(&osc_lru_waitq);
 795         }
 796         RETURN(count > 0 ? count : rc);
 797 }
 798
 799 long osc_lru_reclaim(struct client_obd *cli)
 800 {
 801         struct cl_env_nest nest;
 802         struct lu_env *env;
 803         struct cl_client_cache *cache = cli->cl_cache;
 804         long rc = 0;
 805         int max_scans;
 806         ENTRY;
 807
 808         LASSERT(cache != NULL);
 809         LASSERT(!list_empty(&cache->ccc_lru));
 810
 811         env = cl_env_nested_get(&nest);
 812         if (IS_ERR(env))
 813                 RETURN(rc);
 814
 815         rc = osc_lru_shrink(env, cli, osc_cache_too_much(cli), false);
 816         if (rc != 0) {
 817                 if (rc == -EBUSY)
 818                         rc = 0;
 819
 820                 CDEBUG(D_CACHE, "%s: Free %ld pages from own LRU: %p.\n",
 821                         cli->cl_import->imp_obd->obd_name, rc, cli);
 822                 GOTO(out, rc);
 823         }
 824
 825         CDEBUG(D_CACHE, "%s: cli %p no free slots, pages: %ld, busy: %ld.\n",
 826                 cli->cl_import->imp_obd->obd_name, cli,
 827                 atomic_long_read(&cli->cl_lru_in_list),
 828                 atomic_long_read(&cli->cl_lru_busy));
 829
 830         /* Reclaim LRU slots from other client_obd as it can't free enough
 831          * from its own. This should rarely happen. */
 832         spin_lock(&cache->ccc_lru_lock);
 833         cache->ccc_lru_shrinkers++;
 834         list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
 835
 836         max_scans = atomic_read(&cache->ccc_users);
 837         while (--max_scans > 0 && !list_empty(&cache->ccc_lru)) {
 838                 cli = list_entry(cache->ccc_lru.next, struct client_obd,
 839                                  cl_lru_osc);
 840
 841                 CDEBUG(D_CACHE, "%s: cli %p LRU pages: %ld, busy: %ld.\n",
 842                         cli->cl_import->imp_obd->obd_name, cli,
 843                         atomic_long_read(&cli->cl_lru_in_list),
 844                         atomic_long_read(&cli->cl_lru_busy));
 845
 846                 list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
 847                 if (osc_cache_too_much(cli) > 0) {
 848                         spin_unlock(&cache->ccc_lru_lock);
 849
 850                         rc = osc_lru_shrink(env, cli, osc_cache_too_much(cli),
 851                                             true);
 852                         spin_lock(&cache->ccc_lru_lock);
 853                         if (rc != 0)
 854                                 break;
 855                 }
 856         }
 857         spin_unlock(&cache->ccc_lru_lock);
 858
 859 out:
 860         cl_env_nested_put(&nest, env);
 861         CDEBUG(D_CACHE, "%s: cli %p freed %ld pages.\n",
 862                 cli->cl_import->imp_obd->obd_name, cli, rc);
 863         return rc;
 864 }
 865
 866 /**
 867  * osc_lru_reserve() is called to reserve an LRU slot for a cl_page.
 868  *
 869  * Usually the LRU slots are reserved in osc_io_iter_rw_init().
 870  * Only in the case that the LRU slots are in extreme shortage, it should
 871  * have reserved enough slots for an IO.
 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 osc_io *oio = osc_env_io(env);
 878         struct client_obd *cli = osc_cli(obj);
 879         int rc = 0;
 880         ENTRY;
 881
 882         if (cli->cl_cache == NULL) /* shall not be in LRU */
 883                 RETURN(0);
 884
 885         if (oio->oi_lru_reserved > 0) {
 886                 --oio->oi_lru_reserved;
 887                 goto out;
 888         }
 889
 890         LASSERT(atomic_long_read(cli->cl_lru_left) >= 0);
 891         while (!atomic_long_add_unless(cli->cl_lru_left, -1, 0)) {
 892
 893                 /* run out of LRU spaces, try to drop some by itself */
 894                 rc = osc_lru_reclaim(cli);
 895                 if (rc < 0)
 896                         break;
 897                 if (rc > 0)
 898                         continue;
 899
 900                 cond_resched();
 901                 rc = l_wait_event(osc_lru_waitq,
 902                                 atomic_long_read(cli->cl_lru_left) > 0,
 903                                 &lwi);
 904                 if (rc < 0)
 905                         break;
 906         }
 907
 908 out:
 909         if (rc >= 0) {
 910                 atomic_long_inc(&cli->cl_lru_busy);
 911                 opg->ops_in_lru = 1;
 912                 rc = 0;
 913         }
 914
 915         RETURN(rc);
 916 }
 917
 918 /**
 919  * Atomic operations are expensive. We accumulate the accounting for the
 920  * same page zone to get better performance.
 921  * In practice this can work pretty good because the pages in the same RPC
 922  * are likely from the same page zone.
 923  */
 924 static inline void unstable_page_accounting(struct ptlrpc_bulk_desc *desc,
 925                                             int factor)
 926 {
 927         obd_count page_count = desc->bd_iov_count;
 928         void *zone = NULL;
 929         int count = 0;
 930         int i;
 931
 932         for (i = 0; i < page_count; i++) {
 933                 void *pz = page_zone(desc->bd_iov[i].kiov_page);
 934
 935                 if (likely(pz == zone)) {
 936                         ++count;
 937                         continue;
 938                 }
 939
 940                 if (count > 0) {
 941                         mod_zone_page_state(zone, NR_UNSTABLE_NFS,
 942                                             factor * count);
 943                         count = 0;
 944                 }
 945                 zone = pz;
 946                 ++count;
 947         }
 948         if (count > 0)
 949                 mod_zone_page_state(zone, NR_UNSTABLE_NFS, factor * count);
 950 }
 951
 952 static inline void add_unstable_page_accounting(struct ptlrpc_bulk_desc *desc)
 953 {
 954         unstable_page_accounting(desc, 1);
 955 }
 956
 957 static inline void dec_unstable_page_accounting(struct ptlrpc_bulk_desc *desc)
 958 {
 959         unstable_page_accounting(desc, -1);
 960 }
 961
 962 /**
 963  * Performs "unstable" page accounting. This function balances the
 964  * increment operations performed in osc_inc_unstable_pages. It is
 965  * registered as the RPC request callback, and is executed when the
 966  * bulk RPC is committed on the server. Thus at this point, the pages
 967  * involved in the bulk transfer are no longer considered unstable.
 968  *
 969  * If this function is called, the request should have been committed
 970  * or req:rq_unstable must have been set; it implies that the unstable
 971  * statistic have been added.
 972  */
 973 void osc_dec_unstable_pages(struct ptlrpc_request *req)
 974 {
 975         struct ptlrpc_bulk_desc *desc       = req->rq_bulk;
 976         struct client_obd       *cli        = &req->rq_import->imp_obd->u.cli;
 977         int                      page_count = desc->bd_iov_count;
 978         long                     unstable_count;
 979
 980         LASSERT(page_count >= 0);
 981         dec_unstable_page_accounting(desc);
 982
 983         unstable_count = atomic_long_sub_return(page_count,
 984                                                 &cli->cl_unstable_count);
 985         LASSERT(unstable_count >= 0);
 986
 987         unstable_count = atomic_long_sub_return(page_count,
 988                                            &cli->cl_cache->ccc_unstable_nr);
 989         LASSERT(unstable_count >= 0);
 990         if (unstable_count == 0)
 991                 wake_up_all(&cli->cl_cache->ccc_unstable_waitq);
 992
 993         if (osc_cache_too_much(cli))
 994                 (void)ptlrpcd_queue_work(cli->cl_lru_work);
 995 }
 996
 997 /**
 998  * "unstable" page accounting. See: osc_dec_unstable_pages.
 999  */
1000 void osc_inc_unstable_pages(struct ptlrpc_request *req)
1001 {
1002         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
1003         struct client_obd       *cli  = &req->rq_import->imp_obd->u.cli;
1004         long                     page_count = desc->bd_iov_count;
1005
1006         /* No unstable page tracking */
1007         if (cli->cl_cache == NULL || !cli->cl_cache->ccc_unstable_check)
1008                 return;
1009
1010         add_unstable_page_accounting(desc);
1011         atomic_long_add(page_count, &cli->cl_unstable_count);
1012         atomic_long_add(page_count, &cli->cl_cache->ccc_unstable_nr);
1013
1014         /* If the request has already been committed (i.e. brw_commit
1015          * called via rq_commit_cb), we need to undo the unstable page
1016          * increments we just performed because rq_commit_cb wont be
1017          * called again. */
1018         spin_lock(&req->rq_lock);
1019         if (unlikely(req->rq_committed)) {
1020                 spin_unlock(&req->rq_lock);
1021
1022                 osc_dec_unstable_pages(req);
1023         } else {
1024                 req->rq_unstable = 1;
1025                 spin_unlock(&req->rq_lock);
1026         }
1027 }
1028
1029 /**
1030  * Check if it piggybacks SOFT_SYNC flag to OST from this OSC.
1031  * This function will be called by every BRW RPC so it's critical
1032  * to make this function fast.
1033  */
1034 bool osc_over_unstable_soft_limit(struct client_obd *cli)
1035 {
1036         long unstable_nr, osc_unstable_count;
1037
1038         /* Can't check cli->cl_unstable_count, therefore, no soft limit */
1039         if (cli->cl_cache == NULL || !cli->cl_cache->ccc_unstable_check)
1040                 return false;
1041
1042         osc_unstable_count = atomic_long_read(&cli->cl_unstable_count);
1043         unstable_nr = atomic_long_read(&cli->cl_cache->ccc_unstable_nr);
1044
1045         CDEBUG(D_CACHE,
1046                "%s: cli: %p unstable pages: %lu, osc unstable pages: %lu\n",
1047                cli->cl_import->imp_obd->obd_name, cli,
1048                unstable_nr, osc_unstable_count);
1049
1050         /* If the LRU slots are in shortage - 25% remaining AND this OSC
1051          * has one full RPC window of unstable pages, it's a good chance
1052          * to piggyback a SOFT_SYNC flag.
1053          * Please notice that the OST won't take immediate response for the
1054          * SOFT_SYNC request so active OSCs will have more chance to carry
1055          * the flag, this is reasonable. */
1056         return unstable_nr > cli->cl_cache->ccc_lru_max >> 2 &&
1057                osc_unstable_count > cli->cl_max_pages_per_rpc *
1058                                     cli->cl_max_rpcs_in_flight;
1059 }
1060
1061 /** @} osc */