lustre/osc/osc_page.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2015, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * Implementation of cl_page for OSC layer.
  37  *
  38  *   Author: Nikita Danilov <nikita.danilov@sun.com>
  39  *   Author: Jinshan Xiong <jinshan.xiong@intel.com>
  40  */
  41
  42 #define DEBUG_SUBSYSTEM S_OSC
  43
  44 #include "osc_cl_internal.h"
  45
  46 static void osc_lru_del(struct client_obd *cli, struct osc_page *opg);
  47 static void osc_lru_use(struct client_obd *cli, struct osc_page *opg);
  48 static int osc_lru_alloc(const struct lu_env *env, struct client_obd *cli,
  49                          struct osc_page *opg);
  50
  51 /** \addtogroup osc
  52  *  @{
  53  */
  54
  55 /*
  56  * Comment out osc_page_protected because it may sleep inside the
  57  * the client_obd_list_lock.
  58  * client_obd_list_lock -> osc_ap_completion -> osc_completion ->
  59  *   -> osc_page_protected -> osc_page_is_dlocked -> osc_match_base
  60  *   -> ldlm_lock_match -> sptlrpc_import_check_ctx -> sleep.
  61  */
  62 #if 0
  63 static int osc_page_is_dlocked(const struct lu_env *env,
  64                                const struct osc_page *opg,
  65                                enum cl_lock_mode mode, int pending, int unref)
  66 {
  67         struct cl_page         *page;
  68         struct osc_object      *obj;
  69         struct osc_thread_info *info;
  70         struct ldlm_res_id     *resname;
  71         struct lustre_handle   *lockh;
  72         union ldlm_policy_data *policy;
  73         enum ldlm_mode          dlmmode;
  74         __u64                   flags;
  75
  76         might_sleep();
  77
  78         info = osc_env_info(env);
  79         resname = &info->oti_resname;
  80         policy = &info->oti_policy;
  81         lockh = &info->oti_handle;
  82         page = opg->ops_cl.cpl_page;
  83         obj = cl2osc(opg->ops_cl.cpl_obj);
  84
  85         flags = LDLM_FL_TEST_LOCK | LDLM_FL_BLOCK_GRANTED;
  86         if (pending)
  87                 flags |= LDLM_FL_CBPENDING;
  88
  89         dlmmode = osc_cl_lock2ldlm(mode) | LCK_PW;
  90         osc_lock_build_res(env, obj, resname);
  91         osc_index2policy(policy, page->cp_obj, osc_index(opg), osc_index(opg));
  92         return osc_match_base(osc_export(obj), resname, LDLM_EXTENT, policy,
  93                               dlmmode, &flags, NULL, lockh, unref);
  94 }
  95
  96 /**
  97  * Checks an invariant that a page in the cache is covered by a lock, as
  98  * needed.
  99  */
 100 static int osc_page_protected(const struct lu_env *env,
 101                               const struct osc_page *opg,
 102                               enum cl_lock_mode mode, int unref)
 103 {
 104         struct cl_object_header *hdr;
 105         struct cl_lock          *scan;
 106         struct cl_page          *page;
 107         struct cl_lock_descr    *descr;
 108         int result;
 109
 110         LINVRNT(!opg->ops_temp);
 111
 112         page = opg->ops_cl.cpl_page;
 113         if (page->cp_owner != NULL &&
 114             cl_io_top(page->cp_owner)->ci_lockreq == CILR_NEVER)
 115                 /*
 116                  * If IO is done without locks (liblustre, or lloop), lock is
 117                  * not required.
 118                  */
 119                 result = 1;
 120         else
 121                 /* otherwise check for a DLM lock */
 122         result = osc_page_is_dlocked(env, opg, mode, 1, unref);
 123         if (result == 0) {
 124                 /* maybe this page is a part of a lockless io? */
 125                 hdr = cl_object_header(opg->ops_cl.cpl_obj);
 126                 descr = &osc_env_info(env)->oti_descr;
 127                 descr->cld_mode = mode;
 128                 descr->cld_start = osc_index(opg);
 129                 descr->cld_end   = osc_index(opg);
 130                 spin_lock(&hdr->coh_lock_guard);
 131                 list_for_each_entry(scan, &hdr->coh_locks, cll_linkage) {
 132                         /*
 133                          * Lock-less sub-lock has to be either in HELD state
 134                          * (when io is actively going on), or in CACHED state,
 135                          * when top-lock is being unlocked:
 136                          * cl_io_unlock()->cl_unuse()->...->lov_lock_unuse().
 137                          */
 138                         if ((scan->cll_state == CLS_HELD ||
 139                              scan->cll_state == CLS_CACHED) &&
 140                             cl_lock_ext_match(&scan->cll_descr, descr)) {
 141                                 struct osc_lock *olck;
 142
 143                                 olck = osc_lock_at(scan);
 144                                 result = osc_lock_is_lockless(olck);
 145                                 break;
 146                         }
 147                 }
 148                 spin_unlock(&hdr->coh_lock_guard);
 149         }
 150         return result;
 151 }
 152 #else
 153 static int osc_page_protected(const struct lu_env *env,
 154                               const struct osc_page *opg,
 155                               enum cl_lock_mode mode, int unref)
 156 {
 157         return 1;
 158 }
 159 #endif
 160
 161 /*****************************************************************************
 162  *
 163  * Page operations.
 164  *
 165  */
 166 static void osc_page_transfer_get(struct osc_page *opg, const char *label)
 167 {
 168         struct cl_page *page = opg->ops_cl.cpl_page;
 169
 170         LASSERT(!opg->ops_transfer_pinned);
 171         cl_page_get(page);
 172         lu_ref_add_atomic(&page->cp_reference, label, page);
 173         opg->ops_transfer_pinned = 1;
 174 }
 175
 176 static void osc_page_transfer_put(const struct lu_env *env,
 177                                   struct osc_page *opg)
 178 {
 179         struct cl_page *page = opg->ops_cl.cpl_page;
 180
 181         if (opg->ops_transfer_pinned) {
 182                 opg->ops_transfer_pinned = 0;
 183                 lu_ref_del(&page->cp_reference, "transfer", page);
 184                 cl_page_put(env, page);
 185         }
 186 }
 187
 188 /**
 189  * This is called once for every page when it is submitted for a transfer
 190  * either opportunistic (osc_page_cache_add()), or immediate
 191  * (osc_page_submit()).
 192  */
 193 static void osc_page_transfer_add(const struct lu_env *env,
 194                                   struct osc_page *opg, enum cl_req_type crt)
 195 {
 196         struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
 197
 198         /* ops_lru and ops_inflight share the same field, so take it from LRU
 199          * first and then use it as inflight. */
 200         osc_lru_use(osc_cli(obj), opg);
 201 }
 202
 203 int osc_page_cache_add(const struct lu_env *env,
 204                         const struct cl_page_slice *slice, struct cl_io *io)
 205 {
 206         struct osc_page *opg = cl2osc_page(slice);
 207         int result;
 208         ENTRY;
 209
 210         LINVRNT(osc_page_protected(env, opg, CLM_WRITE, 0));
 211
 212         osc_page_transfer_get(opg, "transfer\0cache");
 213         result = osc_queue_async_io(env, io, opg);
 214         if (result != 0)
 215                 osc_page_transfer_put(env, opg);
 216         else
 217                 osc_page_transfer_add(env, opg, CRT_WRITE);
 218
 219         RETURN(result);
 220 }
 221
 222 void osc_index2policy(union ldlm_policy_data *policy,
 223                       const struct cl_object *obj, pgoff_t start, pgoff_t end)
 224 {
 225         memset(policy, 0, sizeof *policy);
 226         policy->l_extent.start = cl_offset(obj, start);
 227         policy->l_extent.end   = cl_offset(obj, end + 1) - 1;
 228 }
 229
 230 static const char *osc_list(struct list_head *head)
 231 {
 232         return list_empty(head) ? "-" : "+";
 233 }
 234
 235 static inline cfs_time_t osc_submit_duration(struct osc_page *opg)
 236 {
 237         if (opg->ops_submit_time == 0)
 238                 return 0;
 239
 240         return (cfs_time_current() - opg->ops_submit_time);
 241 }
 242
 243 static int osc_page_print(const struct lu_env *env,
 244                           const struct cl_page_slice *slice,
 245                           void *cookie, lu_printer_t printer)
 246 {
 247         struct osc_page       *opg = cl2osc_page(slice);
 248         struct osc_async_page *oap = &opg->ops_oap;
 249         struct osc_object     *obj = cl2osc(slice->cpl_obj);
 250         struct client_obd     *cli = &osc_export(obj)->exp_obd->u.cli;
 251
 252         return (*printer)(env, cookie, LUSTRE_OSC_NAME"-page@%p %lu: "
 253                           "1< %#x %d %u %s %s > "
 254                           "2< "LPD64" %u %u %#x %#x | %p %p %p > "
 255                           "3< %d %lu %d > "
 256                           "4< %d %d %d %lu %s | %s %s %s %s > "
 257                           "5< %s %s %s %s | %d %s | %d %s %s>\n",
 258                           opg, osc_index(opg),
 259                           /* 1 */
 260                           oap->oap_magic, oap->oap_cmd,
 261                           oap->oap_interrupted,
 262                           osc_list(&oap->oap_pending_item),
 263                           osc_list(&oap->oap_rpc_item),
 264                           /* 2 */
 265                           oap->oap_obj_off, oap->oap_page_off, oap->oap_count,
 266                           oap->oap_async_flags, oap->oap_brw_flags,
 267                           oap->oap_request, oap->oap_cli, obj,
 268                           /* 3 */
 269                           opg->ops_transfer_pinned,
 270                           osc_submit_duration(opg), opg->ops_srvlock,
 271                           /* 4 */
 272                           cli->cl_r_in_flight, cli->cl_w_in_flight,
 273                           cli->cl_max_rpcs_in_flight,
 274                           cli->cl_avail_grant,
 275                           osc_list(&cli->cl_cache_waiters),
 276                           osc_list(&cli->cl_loi_ready_list),
 277                           osc_list(&cli->cl_loi_hp_ready_list),
 278                           osc_list(&cli->cl_loi_write_list),
 279                           osc_list(&cli->cl_loi_read_list),
 280                           /* 5 */
 281                           osc_list(&obj->oo_ready_item),
 282                           osc_list(&obj->oo_hp_ready_item),
 283                           osc_list(&obj->oo_write_item),
 284                           osc_list(&obj->oo_read_item),
 285                           atomic_read(&obj->oo_nr_reads),
 286                           osc_list(&obj->oo_reading_exts),
 287                           atomic_read(&obj->oo_nr_writes),
 288                           osc_list(&obj->oo_hp_exts),
 289                           osc_list(&obj->oo_urgent_exts));
 290 }
 291
 292 static void osc_page_delete(const struct lu_env *env,
 293                             const struct cl_page_slice *slice)
 294 {
 295         struct osc_page   *opg = cl2osc_page(slice);
 296         struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
 297         int rc;
 298
 299         LINVRNT(opg->ops_temp || osc_page_protected(env, opg, CLM_READ, 1));
 300
 301         ENTRY;
 302         CDEBUG(D_TRACE, "%p\n", opg);
 303         osc_page_transfer_put(env, opg);
 304         rc = osc_teardown_async_page(env, obj, opg);
 305         if (rc) {
 306                 CL_PAGE_DEBUG(D_ERROR, env, slice->cpl_page,
 307                               "Trying to teardown failed: %d\n", rc);
 308                 LASSERT(0);
 309         }
 310
 311         osc_lru_del(osc_cli(obj), opg);
 312
 313         if (slice->cpl_page->cp_type == CPT_CACHEABLE) {
 314                 void *value;
 315
 316                 spin_lock(&obj->oo_tree_lock);
 317                 value = radix_tree_delete(&obj->oo_tree, osc_index(opg));
 318                 if (value != NULL)
 319                         --obj->oo_npages;
 320                 spin_unlock(&obj->oo_tree_lock);
 321
 322                 LASSERT(ergo(value != NULL, value == opg));
 323         }
 324
 325         EXIT;
 326 }
 327
 328 static void osc_page_clip(const struct lu_env *env,
 329                           const struct cl_page_slice *slice,
 330                           int from, int to)
 331 {
 332         struct osc_page       *opg = cl2osc_page(slice);
 333         struct osc_async_page *oap = &opg->ops_oap;
 334
 335         LINVRNT(osc_page_protected(env, opg, CLM_READ, 0));
 336
 337         opg->ops_from = from;
 338         opg->ops_to   = to;
 339         spin_lock(&oap->oap_lock);
 340         oap->oap_async_flags |= ASYNC_COUNT_STABLE;
 341         spin_unlock(&oap->oap_lock);
 342 }
 343
 344 static int osc_page_cancel(const struct lu_env *env,
 345                            const struct cl_page_slice *slice)
 346 {
 347         struct osc_page *opg = cl2osc_page(slice);
 348         int rc = 0;
 349
 350         LINVRNT(osc_page_protected(env, opg, CLM_READ, 0));
 351
 352         /* Check if the transferring against this page
 353          * is completed, or not even queued. */
 354         if (opg->ops_transfer_pinned)
 355                 /* FIXME: may not be interrupted.. */
 356                 rc = osc_cancel_async_page(env, opg);
 357         LASSERT(ergo(rc == 0, opg->ops_transfer_pinned == 0));
 358         return rc;
 359 }
 360
 361 static int osc_page_flush(const struct lu_env *env,
 362                           const struct cl_page_slice *slice,
 363                           struct cl_io *io)
 364 {
 365         struct osc_page *opg = cl2osc_page(slice);
 366         int rc = 0;
 367         ENTRY;
 368         rc = osc_flush_async_page(env, io, opg);
 369         RETURN(rc);
 370 }
 371
 372 static const struct cl_page_operations osc_page_ops = {
 373         .cpo_print         = osc_page_print,
 374         .cpo_delete        = osc_page_delete,
 375         .cpo_clip           = osc_page_clip,
 376         .cpo_cancel         = osc_page_cancel,
 377         .cpo_flush          = osc_page_flush
 378 };
 379
 380 int osc_page_init(const struct lu_env *env, struct cl_object *obj,
 381                   struct cl_page *page, pgoff_t index)
 382 {
 383         struct osc_object *osc = cl2osc(obj);
 384         struct osc_page   *opg = cl_object_page_slice(obj, page);
 385         int result;
 386
 387         opg->ops_from = 0;
 388         opg->ops_to   = PAGE_CACHE_SIZE;
 389
 390         result = osc_prep_async_page(osc, opg, page->cp_vmpage,
 391                                      cl_offset(obj, index));
 392         if (result == 0) {
 393                 struct osc_io *oio = osc_env_io(env);
 394                 opg->ops_srvlock = osc_io_srvlock(oio);
 395                 cl_page_slice_add(page, &opg->ops_cl, obj, index,
 396                                   &osc_page_ops);
 397         }
 398         /*
 399          * Cannot assert osc_page_protected() here as read-ahead
 400          * creates temporary pages outside of a lock.
 401          */
 402 #ifdef CONFIG_LUSTRE_DEBUG_EXPENSIVE_CHECK
 403         opg->ops_temp = !osc_page_protected(env, opg, CLM_READ, 1);
 404 #endif
 405         INIT_LIST_HEAD(&opg->ops_lru);
 406
 407         /* reserve an LRU space for this page */
 408         if (page->cp_type == CPT_CACHEABLE && result == 0) {
 409                 result = osc_lru_alloc(env, osc_cli(osc), opg);
 410                 if (result == 0) {
 411                         spin_lock(&osc->oo_tree_lock);
 412                         result = radix_tree_insert(&osc->oo_tree, index, opg);
 413                         if (result == 0)
 414                                 ++osc->oo_npages;
 415                         spin_unlock(&osc->oo_tree_lock);
 416                         LASSERT(result == 0);
 417                 }
 418         }
 419
 420         return result;
 421 }
 422
 423 /**
 424  * Helper function called by osc_io_submit() for every page in an immediate
 425  * transfer (i.e., transferred synchronously).
 426  */
 427 void osc_page_submit(const struct lu_env *env, struct osc_page *opg,
 428                      enum cl_req_type crt, int brw_flags)
 429 {
 430         struct osc_async_page *oap = &opg->ops_oap;
 431
 432         LINVRNT(osc_page_protected(env, opg,
 433                                    crt == CRT_WRITE ? CLM_WRITE : CLM_READ, 1));
 434
 435         LASSERTF(oap->oap_magic == OAP_MAGIC, "Bad oap magic: oap %p, "
 436                  "magic 0x%x\n", oap, oap->oap_magic);
 437         LASSERT(oap->oap_async_flags & ASYNC_READY);
 438         LASSERT(oap->oap_async_flags & ASYNC_COUNT_STABLE);
 439
 440         oap->oap_cmd       = crt == CRT_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ;
 441         oap->oap_page_off  = opg->ops_from;
 442         oap->oap_count     = opg->ops_to - opg->ops_from;
 443         oap->oap_brw_flags = OBD_BRW_SYNC | brw_flags;
 444
 445         if (cfs_capable(CFS_CAP_SYS_RESOURCE)) {
 446                 oap->oap_brw_flags |= OBD_BRW_NOQUOTA;
 447                 oap->oap_cmd |= OBD_BRW_NOQUOTA;
 448         }
 449
 450         opg->ops_submit_time = cfs_time_current();
 451         osc_page_transfer_get(opg, "transfer\0imm");
 452         osc_page_transfer_add(env, opg, crt);
 453 }
 454
 455 /* --------------- LRU page management ------------------ */
 456
 457 /* OSC is a natural place to manage LRU pages as applications are specialized
 458  * to write OSC by OSC. Ideally, if one OSC is used more frequently it should
 459  * occupy more LRU slots. On the other hand, we should avoid using up all LRU
 460  * slots (client_obd::cl_lru_left) otherwise process has to be put into sleep
 461  * for free LRU slots - this will be very bad so the algorithm requires each
 462  * OSC to free slots voluntarily to maintain a reasonable number of free slots
 463  * at any time.
 464  */
 465
 466 static DECLARE_WAIT_QUEUE_HEAD(osc_lru_waitq);
 467
 468 /**
 469  * LRU pages are freed in batch mode. OSC should at least free this
 470  * number of pages to avoid running out of LRU slots.
 471  */
 472 static inline int lru_shrink_min(struct client_obd *cli)
 473 {
 474         return cli->cl_max_pages_per_rpc * 2;
 475 }
 476
 477 /**
 478  * free this number at most otherwise it will take too long time to finsih.
 479  */
 480 static inline int lru_shrink_max(struct client_obd *cli)
 481 {
 482         return cli->cl_max_pages_per_rpc * cli->cl_max_rpcs_in_flight;
 483 }
 484
 485 /**
 486  * Check if we can free LRU slots from this OSC. If there exists LRU waiters,
 487  * we should free slots aggressively. In this way, slots are freed in a steady
 488  * step to maintain fairness among OSCs.
 489  *
 490  * Return how many LRU pages should be freed.
 491  */
 492 static int osc_cache_too_much(struct client_obd *cli)
 493 {
 494         struct cl_client_cache *cache = cli->cl_cache;
 495         long pages = atomic_long_read(&cli->cl_lru_in_list);
 496         unsigned long budget;
 497
 498         LASSERT(cache != NULL);
 499         budget = cache->ccc_lru_max / (atomic_read(&cache->ccc_users) - 2);
 500
 501         /* if it's going to run out LRU slots, we should free some, but not
 502          * too much to maintain faireness among OSCs. */
 503         if (atomic_long_read(cli->cl_lru_left) < cache->ccc_lru_max >> 2) {
 504                 if (pages >= budget)
 505                         return lru_shrink_max(cli);
 506                 else if (pages >= budget / 2)
 507                         return lru_shrink_min(cli);
 508         } else {
 509                 int duration = cfs_time_current_sec() - cli->cl_lru_last_used;
 510
 511                 /* knock out pages by duration of no IO activity */
 512                 duration >>= 6; /* approximately 1 minute */
 513                 if (duration > 0 && pages >= budget / duration)
 514                         return lru_shrink_min(cli);
 515         }
 516         return 0;
 517 }
 518
 519 int lru_queue_work(const struct lu_env *env, void *data)
 520 {
 521         struct client_obd *cli = data;
 522         int count;
 523
 524         CDEBUG(D_CACHE, "%s: run LRU work for client obd\n", cli_name(cli));
 525         count = osc_cache_too_much(cli);
 526         if (count > 0) {
 527                 int rc = osc_lru_shrink(env, cli, count, false);
 528
 529                 CDEBUG(D_CACHE, "%s: shrank %d/%d pages from client obd\n",
 530                        cli_name(cli), rc, count);
 531                 if (rc >= count) {
 532                         CDEBUG(D_CACHE, "%s: queue again\n", cli_name(cli));
 533                         ptlrpcd_queue_work(cli->cl_lru_work);
 534                 }
 535         }
 536
 537         RETURN(0);
 538 }
 539
 540 void osc_lru_add_batch(struct client_obd *cli, struct list_head *plist)
 541 {
 542         struct list_head lru = LIST_HEAD_INIT(lru);
 543         struct osc_async_page *oap;
 544         long npages = 0;
 545
 546         list_for_each_entry(oap, plist, oap_pending_item) {
 547                 struct osc_page *opg = oap2osc_page(oap);
 548
 549                 if (!opg->ops_in_lru)
 550                         continue;
 551
 552                 ++npages;
 553                 LASSERT(list_empty(&opg->ops_lru));
 554                 list_add(&opg->ops_lru, &lru);
 555         }
 556
 557         if (npages > 0) {
 558                 spin_lock(&cli->cl_lru_list_lock);
 559                 list_splice_tail(&lru, &cli->cl_lru_list);
 560                 atomic_long_sub(npages, &cli->cl_lru_busy);
 561                 atomic_long_add(npages, &cli->cl_lru_in_list);
 562                 cli->cl_lru_last_used = cfs_time_current_sec();
 563                 spin_unlock(&cli->cl_lru_list_lock);
 564
 565                 if (waitqueue_active(&osc_lru_waitq))
 566                         (void)ptlrpcd_queue_work(cli->cl_lru_work);
 567         }
 568 }
 569
 570 static void __osc_lru_del(struct client_obd *cli, struct osc_page *opg)
 571 {
 572         LASSERT(atomic_long_read(&cli->cl_lru_in_list) > 0);
 573         list_del_init(&opg->ops_lru);
 574         atomic_long_dec(&cli->cl_lru_in_list);
 575 }
 576
 577 /**
 578  * Page is being destroyed. The page may be not in LRU list, if the transfer
 579  * has never finished(error occurred).
 580  */
 581 static void osc_lru_del(struct client_obd *cli, struct osc_page *opg)
 582 {
 583         if (opg->ops_in_lru) {
 584                 spin_lock(&cli->cl_lru_list_lock);
 585                 if (!list_empty(&opg->ops_lru)) {
 586                         __osc_lru_del(cli, opg);
 587                 } else {
 588                         LASSERT(atomic_long_read(&cli->cl_lru_busy) > 0);
 589                         atomic_long_dec(&cli->cl_lru_busy);
 590                 }
 591                 spin_unlock(&cli->cl_lru_list_lock);
 592
 593                 atomic_long_inc(cli->cl_lru_left);
 594                 /* this is a great place to release more LRU pages if
 595                  * this osc occupies too many LRU pages and kernel is
 596                  * stealing one of them. */
 597                 if (osc_cache_too_much(cli)) {
 598                         CDEBUG(D_CACHE, "%s: queue LRU work\n", cli_name(cli));
 599                         (void)ptlrpcd_queue_work(cli->cl_lru_work);
 600                 }
 601                 wake_up(&osc_lru_waitq);
 602         } else {
 603                 LASSERT(list_empty(&opg->ops_lru));
 604         }
 605 }
 606
 607 /**
 608  * Delete page from LRUlist for redirty.
 609  */
 610 static void osc_lru_use(struct client_obd *cli, struct osc_page *opg)
 611 {
 612         /* If page is being transferred for the first time,
 613          * ops_lru should be empty */
 614         if (opg->ops_in_lru && !list_empty(&opg->ops_lru)) {
 615                 spin_lock(&cli->cl_lru_list_lock);
 616                 __osc_lru_del(cli, opg);
 617                 spin_unlock(&cli->cl_lru_list_lock);
 618                 atomic_long_inc(&cli->cl_lru_busy);
 619         }
 620 }
 621
 622 static void discard_pagevec(const struct lu_env *env, struct cl_io *io,
 623                                 struct cl_page **pvec, int max_index)
 624 {
 625         int i;
 626
 627         for (i = 0; i < max_index; i++) {
 628                 struct cl_page *page = pvec[i];
 629
 630                 LASSERT(cl_page_is_owned(page, io));
 631                 cl_page_delete(env, page);
 632                 cl_page_discard(env, io, page);
 633                 cl_page_disown(env, io, page);
 634                 cl_page_put(env, page);
 635
 636                 pvec[i] = NULL;
 637         }
 638 }
 639
 640 /**
 641  * Check if a cl_page can be released, i.e, it's not being used.
 642  *
 643  * If unstable account is turned on, bulk transfer may hold one refcount
 644  * for recovery so we need to check vmpage refcount as well; otherwise,
 645  * even we can destroy cl_page but the corresponding vmpage can't be reused.
 646  */
 647 static inline bool lru_page_busy(struct client_obd *cli, struct cl_page *page)
 648 {
 649         if (cl_page_in_use_noref(page))
 650                 return true;
 651
 652         if (cli->cl_cache->ccc_unstable_check) {
 653                 struct page *vmpage = cl_page_vmpage(page);
 654
 655                 /* vmpage have two known users: cl_page and VM page cache */
 656                 if (page_count(vmpage) - page_mapcount(vmpage) > 2)
 657                         return true;
 658         }
 659         return false;
 660 }
 661
 662 /**
 663  * Drop @target of pages from LRU at most.
 664  */
 665 long osc_lru_shrink(const struct lu_env *env, struct client_obd *cli,
 666                    long target, bool force)
 667 {
 668         struct cl_io *io;
 669         struct cl_object *clobj = NULL;
 670         struct cl_page **pvec;
 671         struct osc_page *opg;
 672         long count = 0;
 673         int maxscan = 0;
 674         int index = 0;
 675         int rc = 0;
 676         ENTRY;
 677
 678         LASSERT(atomic_long_read(&cli->cl_lru_in_list) >= 0);
 679         if (atomic_long_read(&cli->cl_lru_in_list) == 0 || target <= 0)
 680                 RETURN(0);
 681
 682         CDEBUG(D_CACHE, "%s: shrinkers: %d, force: %d\n",
 683                cli_name(cli), atomic_read(&cli->cl_lru_shrinkers), force);
 684         if (!force) {
 685                 if (atomic_read(&cli->cl_lru_shrinkers) > 0)
 686                         RETURN(-EBUSY);
 687
 688                 if (atomic_inc_return(&cli->cl_lru_shrinkers) > 1) {
 689                         atomic_dec(&cli->cl_lru_shrinkers);
 690                         RETURN(-EBUSY);
 691                 }
 692         } else {
 693                 atomic_inc(&cli->cl_lru_shrinkers);
 694         }
 695
 696         pvec = (struct cl_page **)osc_env_info(env)->oti_pvec;
 697         io = &osc_env_info(env)->oti_io;
 698
 699         spin_lock(&cli->cl_lru_list_lock);
 700         if (force)
 701                 cli->cl_lru_reclaim++;
 702         maxscan = min(target << 1, atomic_long_read(&cli->cl_lru_in_list));
 703         while (!list_empty(&cli->cl_lru_list)) {
 704                 struct cl_page *page;
 705                 bool will_free = false;
 706
 707                 if (!force && atomic_read(&cli->cl_lru_shrinkers) > 1)
 708                         break;
 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 /**
 801  * Reclaim LRU pages by an IO thread. The caller wants to reclaim at least
 802  * \@npages of LRU slots. For performance consideration, it's better to drop
 803  * LRU pages in batch. Therefore, the actual number is adjusted at least
 804  * max_pages_per_rpc.
 805  */
 806 static long osc_lru_reclaim(struct client_obd *cli, unsigned long npages)
 807 {
 808         struct lu_env *env;
 809         struct cl_client_cache *cache = cli->cl_cache;
 810         int max_scans;
 811         __u16 refcheck;
 812         long rc = 0;
 813         ENTRY;
 814
 815         LASSERT(cache != NULL);
 816
 817         env = cl_env_get(&refcheck);
 818         if (IS_ERR(env))
 819                 RETURN(rc);
 820
 821         npages = max_t(int, npages, cli->cl_max_pages_per_rpc);
 822         CDEBUG(D_CACHE, "%s: start to reclaim %ld pages from LRU\n",
 823                cli_name(cli), npages);
 824         rc = osc_lru_shrink(env, cli, npages, true);
 825         if (rc >= npages) {
 826                 CDEBUG(D_CACHE, "%s: reclaimed %ld/%ld pages from LRU\n",
 827                        cli_name(cli), rc, npages);
 828                 if (osc_cache_too_much(cli) > 0)
 829                         ptlrpcd_queue_work(cli->cl_lru_work);
 830                 GOTO(out, rc);
 831         } else if (rc > 0) {
 832                 npages -= rc;
 833         }
 834
 835         CDEBUG(D_CACHE, "%s: cli %p no free slots, pages: %ld/%ld, want: %ld\n",
 836                 cli_name(cli), cli, atomic_long_read(&cli->cl_lru_in_list),
 837                 atomic_long_read(&cli->cl_lru_busy), npages);
 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         LASSERT(!list_empty(&cache->ccc_lru));
 843
 844         cache->ccc_lru_shrinkers++;
 845         list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
 846
 847         max_scans = atomic_read(&cache->ccc_users) - 2;
 848         while (--max_scans > 0 && !list_empty(&cache->ccc_lru)) {
 849                 cli = list_entry(cache->ccc_lru.next, struct client_obd,
 850                                  cl_lru_osc);
 851
 852                 CDEBUG(D_CACHE, "%s: cli %p LRU pages: %ld, busy: %ld.\n",
 853                         cli_name(cli), cli,
 854                         atomic_long_read(&cli->cl_lru_in_list),
 855                         atomic_long_read(&cli->cl_lru_busy));
 856
 857                 list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
 858                 if (osc_cache_too_much(cli) > 0) {
 859                         spin_unlock(&cache->ccc_lru_lock);
 860
 861                         rc = osc_lru_shrink(env, cli, npages, true);
 862                         spin_lock(&cache->ccc_lru_lock);
 863                         if (rc >= npages)
 864                                 break;
 865                         if (rc > 0)
 866                                 npages -= rc;
 867                 }
 868         }
 869         spin_unlock(&cache->ccc_lru_lock);
 870
 871 out:
 872         cl_env_put(env, &refcheck);
 873         CDEBUG(D_CACHE, "%s: cli %p freed %ld pages.\n",
 874                 cli_name(cli), cli, rc);
 875         return rc;
 876 }
 877
 878 /**
 879  * osc_lru_alloc() is called to allocate an LRU slot for a cl_page.
 880  *
 881  * Usually the LRU slots are reserved in osc_io_iter_rw_init().
 882  * Only in the case that the LRU slots are in extreme shortage, it should
 883  * have reserved enough slots for an IO.
 884  */
 885 static int osc_lru_alloc(const struct lu_env *env, struct client_obd *cli,
 886                          struct osc_page *opg)
 887 {
 888         struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
 889         struct osc_io *oio = osc_env_io(env);
 890         int rc = 0;
 891         ENTRY;
 892
 893         if (cli->cl_cache == NULL) /* shall not be in LRU */
 894                 RETURN(0);
 895
 896         if (oio->oi_lru_reserved > 0) {
 897                 --oio->oi_lru_reserved;
 898                 goto out;
 899         }
 900
 901         LASSERT(atomic_long_read(cli->cl_lru_left) >= 0);
 902         while (!atomic_long_add_unless(cli->cl_lru_left, -1, 0)) {
 903                 /* run out of LRU spaces, try to drop some by itself */
 904                 rc = osc_lru_reclaim(cli, 1);
 905                 if (rc < 0)
 906                         break;
 907                 if (rc > 0)
 908                         continue;
 909
 910                 cond_resched();
 911                 rc = l_wait_event(osc_lru_waitq,
 912                                 atomic_long_read(cli->cl_lru_left) > 0,
 913                                 &lwi);
 914                 if (rc < 0)
 915                         break;
 916         }
 917
 918 out:
 919         if (rc >= 0) {
 920                 atomic_long_inc(&cli->cl_lru_busy);
 921                 opg->ops_in_lru = 1;
 922                 rc = 0;
 923         }
 924
 925         RETURN(rc);
 926 }
 927
 928 /**
 929  * osc_lru_reserve() is called to reserve enough LRU slots for I/O.
 930  *
 931  * The benefit of doing this is to reduce contention against atomic counter
 932  * cl_lru_left by changing it from per-page access to per-IO access.
 933  */
 934 unsigned long osc_lru_reserve(struct client_obd *cli, unsigned long npages)
 935 {
 936         unsigned long reserved = 0;
 937         unsigned long max_pages;
 938         unsigned long c;
 939
 940         /* reserve a full RPC window at most to avoid that a thread accidentally
 941          * consumes too many LRU slots */
 942         max_pages = cli->cl_max_pages_per_rpc * cli->cl_max_rpcs_in_flight;
 943         if (npages > max_pages)
 944                 npages = max_pages;
 945
 946         c = atomic_long_read(cli->cl_lru_left);
 947         if (c < npages && osc_lru_reclaim(cli, npages) > 0)
 948                 c = atomic_long_read(cli->cl_lru_left);
 949         while (c >= npages) {
 950                 if (c == atomic_long_cmpxchg(cli->cl_lru_left, c, c - npages)) {
 951                         reserved = npages;
 952                         break;
 953                 }
 954                 c = atomic_long_read(cli->cl_lru_left);
 955         }
 956         if (atomic_long_read(cli->cl_lru_left) < max_pages) {
 957                 /* If there aren't enough pages in the per-OSC LRU then
 958                  * wake up the LRU thread to try and clear out space, so
 959                  * we don't block if pages are being dirtied quickly. */
 960                 CDEBUG(D_CACHE, "%s: queue LRU, left: %lu/%ld.\n",
 961                        cli_name(cli), atomic_long_read(cli->cl_lru_left),
 962                        max_pages);
 963                 (void)ptlrpcd_queue_work(cli->cl_lru_work);
 964         }
 965
 966         return reserved;
 967 }
 968
 969 /**
 970  * osc_lru_unreserve() is called to unreserve LRU slots.
 971  *
 972  * LRU slots reserved by osc_lru_reserve() may have entries left due to several
 973  * reasons such as page already existing or I/O error. Those reserved slots
 974  * should be freed by calling this function.
 975  */
 976 void osc_lru_unreserve(struct client_obd *cli, unsigned long npages)
 977 {
 978         atomic_long_add(npages, cli->cl_lru_left);
 979         wake_up_all(&osc_lru_waitq);
 980 }
 981
 982 /**
 983  * Atomic operations are expensive. We accumulate the accounting for the
 984  * same page zone to get better performance.
 985  * In practice this can work pretty good because the pages in the same RPC
 986  * are likely from the same page zone.
 987  */
 988 static inline void unstable_page_accounting(struct ptlrpc_bulk_desc *desc,
 989                                             int factor)
 990 {
 991         int page_count = desc->bd_iov_count;
 992         void *zone = NULL;
 993         int count = 0;
 994         int i;
 995
 996         LASSERT(ptlrpc_is_bulk_desc_kiov(desc->bd_type));
 997
 998         for (i = 0; i < page_count; i++) {
 999                 void *pz = page_zone(BD_GET_KIOV(desc, i).kiov_page);
1000
1001                 if (likely(pz == zone)) {
1002                         ++count;
1003                         continue;
1004                 }
1005
1006                 if (count > 0) {
1007                         mod_zone_page_state(zone, NR_UNSTABLE_NFS,
1008                                             factor * count);
1009                         count = 0;
1010                 }
1011                 zone = pz;
1012                 ++count;
1013         }
1014         if (count > 0)
1015                 mod_zone_page_state(zone, NR_UNSTABLE_NFS, factor * count);
1016 }
1017
1018 static inline void add_unstable_page_accounting(struct ptlrpc_bulk_desc *desc)
1019 {
1020         unstable_page_accounting(desc, 1);
1021 }
1022
1023 static inline void dec_unstable_page_accounting(struct ptlrpc_bulk_desc *desc)
1024 {
1025         unstable_page_accounting(desc, -1);
1026 }
1027
1028 /**
1029  * Performs "unstable" page accounting. This function balances the
1030  * increment operations performed in osc_inc_unstable_pages. It is
1031  * registered as the RPC request callback, and is executed when the
1032  * bulk RPC is committed on the server. Thus at this point, the pages
1033  * involved in the bulk transfer are no longer considered unstable.
1034  *
1035  * If this function is called, the request should have been committed
1036  * or req:rq_unstable must have been set; it implies that the unstable
1037  * statistic have been added.
1038  */
1039 void osc_dec_unstable_pages(struct ptlrpc_request *req)
1040 {
1041         struct ptlrpc_bulk_desc *desc       = req->rq_bulk;
1042         struct client_obd       *cli        = &req->rq_import->imp_obd->u.cli;
1043         int                      page_count = desc->bd_iov_count;
1044         long                     unstable_count;
1045
1046         LASSERT(page_count >= 0);
1047         dec_unstable_page_accounting(desc);
1048
1049         unstable_count = atomic_long_sub_return(page_count,
1050                                                 &cli->cl_unstable_count);
1051         LASSERT(unstable_count >= 0);
1052
1053         unstable_count = atomic_long_sub_return(page_count,
1054                                            &cli->cl_cache->ccc_unstable_nr);
1055         LASSERT(unstable_count >= 0);
1056         if (unstable_count == 0)
1057                 wake_up_all(&cli->cl_cache->ccc_unstable_waitq);
1058
1059         if (waitqueue_active(&osc_lru_waitq))
1060                 (void)ptlrpcd_queue_work(cli->cl_lru_work);
1061 }
1062
1063 /**
1064  * "unstable" page accounting. See: osc_dec_unstable_pages.
1065  */
1066 void osc_inc_unstable_pages(struct ptlrpc_request *req)
1067 {
1068         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
1069         struct client_obd       *cli  = &req->rq_import->imp_obd->u.cli;
1070         long                     page_count = desc->bd_iov_count;
1071
1072         /* No unstable page tracking */
1073         if (cli->cl_cache == NULL || !cli->cl_cache->ccc_unstable_check)
1074                 return;
1075
1076         add_unstable_page_accounting(desc);
1077         atomic_long_add(page_count, &cli->cl_unstable_count);
1078         atomic_long_add(page_count, &cli->cl_cache->ccc_unstable_nr);
1079
1080         /* If the request has already been committed (i.e. brw_commit
1081          * called via rq_commit_cb), we need to undo the unstable page
1082          * increments we just performed because rq_commit_cb wont be
1083          * called again. */
1084         spin_lock(&req->rq_lock);
1085         if (unlikely(req->rq_committed)) {
1086                 spin_unlock(&req->rq_lock);
1087
1088                 osc_dec_unstable_pages(req);
1089         } else {
1090                 req->rq_unstable = 1;
1091                 spin_unlock(&req->rq_lock);
1092         }
1093 }
1094
1095 /**
1096  * Check if it piggybacks SOFT_SYNC flag to OST from this OSC.
1097  * This function will be called by every BRW RPC so it's critical
1098  * to make this function fast.
1099  */
1100 bool osc_over_unstable_soft_limit(struct client_obd *cli)
1101 {
1102         long unstable_nr, osc_unstable_count;
1103
1104         /* Can't check cli->cl_unstable_count, therefore, no soft limit */
1105         if (cli->cl_cache == NULL || !cli->cl_cache->ccc_unstable_check)
1106                 return false;
1107
1108         osc_unstable_count = atomic_long_read(&cli->cl_unstable_count);
1109         unstable_nr = atomic_long_read(&cli->cl_cache->ccc_unstable_nr);
1110
1111         CDEBUG(D_CACHE,
1112                "%s: cli: %p unstable pages: %lu, osc unstable pages: %lu\n",
1113                cli_name(cli), cli, unstable_nr, osc_unstable_count);
1114
1115         /* If the LRU slots are in shortage - 25% remaining AND this OSC
1116          * has one full RPC window of unstable pages, it's a good chance
1117          * to piggyback a SOFT_SYNC flag.
1118          * Please notice that the OST won't take immediate response for the
1119          * SOFT_SYNC request so active OSCs will have more chance to carry
1120          * the flag, this is reasonable. */
1121         return unstable_nr > cli->cl_cache->ccc_lru_max >> 2 &&
1122                osc_unstable_count > cli->cl_max_pages_per_rpc *
1123                                     cli->cl_max_rpcs_in_flight;
1124 }
1125
1126 /**
1127  * Return how many LRU pages in the cache of all OSC devices
1128  *
1129  * \retval      return # of cached LRU pages times reclaimation tendency
1130  * \retval      SHRINK_STOP if it cannot do any scanning in this time
1131  */
1132 unsigned long osc_cache_shrink_count(struct shrinker *sk,
1133                                      struct shrink_control *sc)
1134 {
1135         struct client_obd *cli;
1136         unsigned long cached = 0;
1137
1138         spin_lock(&osc_shrink_lock);
1139         list_for_each_entry(cli, &osc_shrink_list, cl_shrink_list)
1140                 cached += atomic_long_read(&cli->cl_lru_in_list);
1141         spin_unlock(&osc_shrink_lock);
1142
1143         return (cached  * sysctl_vfs_cache_pressure) / 100;
1144 }
1145
1146 /**
1147  * Scan and try to reclaim sc->nr_to_scan cached LRU pages
1148  *
1149  * \retval      number of cached LRU pages reclaimed
1150  * \retval      SHRINK_STOP if it cannot do any scanning in this time
1151  *
1152  * Linux kernel will loop calling this shrinker scan routine with
1153  * sc->nr_to_scan = SHRINK_BATCH(128 for now) until kernel got enough memory.
1154  *
1155  * If sc->nr_to_scan is 0, the VM is querying the cache size, we don't need
1156  * to scan and try to reclaim LRU pages, just return 0 and
1157  * osc_cache_shrink_count() will report the LRU page number.
1158  */
1159 unsigned long osc_cache_shrink_scan(struct shrinker *sk,
1160                                     struct shrink_control *sc)
1161 {
1162         struct client_obd *cli;
1163         struct client_obd *stop_anchor = NULL;
1164         struct lu_env *env;
1165         long shrank = 0;
1166         int rc;
1167         __u16 refcheck;
1168
1169         if (sc->nr_to_scan == 0)
1170                 return 0;
1171
1172         if (!(sc->gfp_mask & __GFP_FS))
1173                 return SHRINK_STOP;
1174
1175         env = cl_env_get(&refcheck);
1176         if (IS_ERR(env))
1177                 return SHRINK_STOP;
1178
1179         spin_lock(&osc_shrink_lock);
1180         while (!list_empty(&osc_shrink_list)) {
1181                 cli = list_entry(osc_shrink_list.next, struct client_obd,
1182                                  cl_shrink_list);
1183
1184                 if (stop_anchor == NULL)
1185                         stop_anchor = cli;
1186                 else if (cli == stop_anchor)
1187                         break;
1188
1189                 list_move_tail(&cli->cl_shrink_list, &osc_shrink_list);
1190                 spin_unlock(&osc_shrink_lock);
1191
1192                 /* shrink no more than max_pages_per_rpc for an OSC */
1193                 rc = osc_lru_shrink(env, cli, (sc->nr_to_scan - shrank) >
1194                                     cli->cl_max_pages_per_rpc ?
1195                                     cli->cl_max_pages_per_rpc :
1196                                     sc->nr_to_scan - shrank, true);
1197                 if (rc > 0)
1198                         shrank += rc;
1199
1200                 if (shrank >= sc->nr_to_scan)
1201                         goto out;
1202
1203                 spin_lock(&osc_shrink_lock);
1204         }
1205         spin_unlock(&osc_shrink_lock);
1206
1207 out:
1208         cl_env_put(env, &refcheck);
1209
1210         return shrank;
1211 }
1212
1213 /** @} osc */