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.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 2016, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * Implementation of cl_page for OSC layer.
  33  *
  34  *   Author: Nikita Danilov <nikita.danilov@sun.com>
  35  *   Author: Jinshan Xiong <jinshan.xiong@intel.com>
  36  */
  37
  38 #define DEBUG_SUBSYSTEM S_OSC
  39 #include <lustre_osc.h>
  40
  41 #include "osc_internal.h"
  42
  43 static void osc_lru_del(struct client_obd *cli, struct osc_page *opg);
  44 static void osc_lru_use(struct client_obd *cli, struct osc_page *opg);
  45 static int osc_lru_alloc(const struct lu_env *env, struct client_obd *cli,
  46                          struct osc_page *opg);
  47
  48 /** \addtogroup osc
  49  *  @{
  50  */
  51
  52 /*
  53  * Page operations.
  54  */
  55 static void osc_page_transfer_get(struct osc_page *opg, const char *label)
  56 {
  57         struct cl_page *page = opg->ops_cl.cpl_page;
  58
  59         LASSERT(!opg->ops_transfer_pinned);
  60         cl_page_get(page);
  61         lu_ref_add_atomic(&page->cp_reference, label, page);
  62         opg->ops_transfer_pinned = 1;
  63 }
  64
  65 static void osc_page_transfer_put(const struct lu_env *env,
  66                                   struct osc_page *opg)
  67 {
  68         struct cl_page *page = opg->ops_cl.cpl_page;
  69
  70         if (opg->ops_transfer_pinned) {
  71                 opg->ops_transfer_pinned = 0;
  72                 lu_ref_del(&page->cp_reference, "transfer", page);
  73                 cl_page_put(env, page);
  74         }
  75 }
  76
  77 /**
  78  * This is called once for every page when it is submitted for a transfer
  79  * either opportunistic (osc_page_cache_add()), or immediate
  80  * (osc_page_submit()).
  81  */
  82 static void osc_page_transfer_add(const struct lu_env *env,
  83                                   struct osc_page *opg, enum cl_req_type crt)
  84 {
  85         struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
  86
  87         /* ops_lru and ops_inflight share the same field, so take it from LRU
  88          * first and then use it as inflight. */
  89         osc_lru_use(osc_cli(obj), opg);
  90 }
  91
  92 int osc_page_cache_add(const struct lu_env *env,
  93                         const struct cl_page_slice *slice, struct cl_io *io)
  94 {
  95         struct osc_page *opg = cl2osc_page(slice);
  96         int result;
  97         ENTRY;
  98
  99         osc_page_transfer_get(opg, "transfer\0cache");
 100         result = osc_queue_async_io(env, io, opg);
 101         if (result != 0)
 102                 osc_page_transfer_put(env, opg);
 103         else
 104                 osc_page_transfer_add(env, opg, CRT_WRITE);
 105
 106         RETURN(result);
 107 }
 108
 109 void osc_index2policy(union ldlm_policy_data *policy,
 110                       const struct cl_object *obj, pgoff_t start, pgoff_t end)
 111 {
 112         memset(policy, 0, sizeof *policy);
 113         policy->l_extent.start = cl_offset(obj, start);
 114         policy->l_extent.end   = cl_offset(obj, end + 1) - 1;
 115 }
 116
 117 static const char *osc_list(struct list_head *head)
 118 {
 119         return list_empty(head) ? "-" : "+";
 120 }
 121
 122 static inline cfs_time_t osc_submit_duration(struct osc_page *opg)
 123 {
 124         if (opg->ops_submit_time == 0)
 125                 return 0;
 126
 127         return (cfs_time_current() - opg->ops_submit_time);
 128 }
 129
 130 static int osc_page_print(const struct lu_env *env,
 131                           const struct cl_page_slice *slice,
 132                           void *cookie, lu_printer_t printer)
 133 {
 134         struct osc_page       *opg = cl2osc_page(slice);
 135         struct osc_async_page *oap = &opg->ops_oap;
 136         struct osc_object     *obj = cl2osc(slice->cpl_obj);
 137         struct client_obd     *cli = &osc_export(obj)->exp_obd->u.cli;
 138
 139         return (*printer)(env, cookie, LUSTRE_OSC_NAME"-page@%p %lu: "
 140                           "1< %#x %d %u %s %s > "
 141                           "2< %lld %u %u %#x %#x | %p %p %p > "
 142                           "3< %d %lu %d > "
 143                           "4< %d %d %d %lu %s | %s %s %s %s > "
 144                           "5< %s %s %s %s | %d %s | %d %s %s>\n",
 145                           opg, osc_index(opg),
 146                           /* 1 */
 147                           oap->oap_magic, oap->oap_cmd,
 148                           oap->oap_interrupted,
 149                           osc_list(&oap->oap_pending_item),
 150                           osc_list(&oap->oap_rpc_item),
 151                           /* 2 */
 152                           oap->oap_obj_off, oap->oap_page_off, oap->oap_count,
 153                           oap->oap_async_flags, oap->oap_brw_flags,
 154                           oap->oap_request, oap->oap_cli, obj,
 155                           /* 3 */
 156                           opg->ops_transfer_pinned,
 157                           osc_submit_duration(opg), opg->ops_srvlock,
 158                           /* 4 */
 159                           cli->cl_r_in_flight, cli->cl_w_in_flight,
 160                           cli->cl_max_rpcs_in_flight,
 161                           cli->cl_avail_grant,
 162                           osc_list(&cli->cl_cache_waiters),
 163                           osc_list(&cli->cl_loi_ready_list),
 164                           osc_list(&cli->cl_loi_hp_ready_list),
 165                           osc_list(&cli->cl_loi_write_list),
 166                           osc_list(&cli->cl_loi_read_list),
 167                           /* 5 */
 168                           osc_list(&obj->oo_ready_item),
 169                           osc_list(&obj->oo_hp_ready_item),
 170                           osc_list(&obj->oo_write_item),
 171                           osc_list(&obj->oo_read_item),
 172                           atomic_read(&obj->oo_nr_reads),
 173                           osc_list(&obj->oo_reading_exts),
 174                           atomic_read(&obj->oo_nr_writes),
 175                           osc_list(&obj->oo_hp_exts),
 176                           osc_list(&obj->oo_urgent_exts));
 177 }
 178
 179 static void osc_page_delete(const struct lu_env *env,
 180                             const struct cl_page_slice *slice)
 181 {
 182         struct osc_page   *opg = cl2osc_page(slice);
 183         struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
 184         int rc;
 185
 186         ENTRY;
 187         CDEBUG(D_TRACE, "%p\n", opg);
 188         osc_page_transfer_put(env, opg);
 189         rc = osc_teardown_async_page(env, obj, opg);
 190         if (rc) {
 191                 CL_PAGE_DEBUG(D_ERROR, env, slice->cpl_page,
 192                               "Trying to teardown failed: %d\n", rc);
 193                 LASSERT(0);
 194         }
 195
 196         osc_lru_del(osc_cli(obj), opg);
 197
 198         if (slice->cpl_page->cp_type == CPT_CACHEABLE) {
 199                 void *value;
 200
 201                 spin_lock(&obj->oo_tree_lock);
 202                 value = radix_tree_delete(&obj->oo_tree, osc_index(opg));
 203                 if (value != NULL)
 204                         --obj->oo_npages;
 205                 spin_unlock(&obj->oo_tree_lock);
 206
 207                 LASSERT(ergo(value != NULL, value == opg));
 208         }
 209
 210         EXIT;
 211 }
 212
 213 static void osc_page_clip(const struct lu_env *env,
 214                           const struct cl_page_slice *slice,
 215                           int from, int to)
 216 {
 217         struct osc_page       *opg = cl2osc_page(slice);
 218         struct osc_async_page *oap = &opg->ops_oap;
 219
 220         opg->ops_from = from;
 221         opg->ops_to   = to;
 222         spin_lock(&oap->oap_lock);
 223         oap->oap_async_flags |= ASYNC_COUNT_STABLE;
 224         spin_unlock(&oap->oap_lock);
 225 }
 226
 227 static int osc_page_cancel(const struct lu_env *env,
 228                            const struct cl_page_slice *slice)
 229 {
 230         struct osc_page *opg = cl2osc_page(slice);
 231         int rc = 0;
 232
 233         /* Check if the transferring against this page
 234          * is completed, or not even queued. */
 235         if (opg->ops_transfer_pinned)
 236                 /* FIXME: may not be interrupted.. */
 237                 rc = osc_cancel_async_page(env, opg);
 238         LASSERT(ergo(rc == 0, opg->ops_transfer_pinned == 0));
 239         return rc;
 240 }
 241
 242 static int osc_page_flush(const struct lu_env *env,
 243                           const struct cl_page_slice *slice,
 244                           struct cl_io *io)
 245 {
 246         struct osc_page *opg = cl2osc_page(slice);
 247         int rc = 0;
 248         ENTRY;
 249         rc = osc_flush_async_page(env, io, opg);
 250         RETURN(rc);
 251 }
 252
 253 static const struct cl_page_operations osc_page_ops = {
 254         .cpo_print         = osc_page_print,
 255         .cpo_delete        = osc_page_delete,
 256         .cpo_clip           = osc_page_clip,
 257         .cpo_cancel         = osc_page_cancel,
 258         .cpo_flush          = osc_page_flush
 259 };
 260
 261 int osc_page_init(const struct lu_env *env, struct cl_object *obj,
 262                   struct cl_page *page, pgoff_t index)
 263 {
 264         struct osc_object *osc = cl2osc(obj);
 265         struct osc_page   *opg = cl_object_page_slice(obj, page);
 266         struct osc_io *oio = osc_env_io(env);
 267         int result;
 268
 269         opg->ops_from = 0;
 270         opg->ops_to   = PAGE_SIZE;
 271
 272         INIT_LIST_HEAD(&opg->ops_lru);
 273
 274         result = osc_prep_async_page(osc, opg, page->cp_vmpage,
 275                                      cl_offset(obj, index));
 276         if (result != 0)
 277                 return result;
 278
 279         opg->ops_srvlock = osc_io_srvlock(oio);
 280         cl_page_slice_add(page, &opg->ops_cl, obj, index,
 281                           &osc_page_ops);
 282
 283
 284         /* reserve an LRU space for this page */
 285         if (page->cp_type == CPT_CACHEABLE) {
 286                 result = osc_lru_alloc(env, osc_cli(osc), opg);
 287                 if (result == 0) {
 288                         result = radix_tree_preload(GFP_NOFS);
 289                         if (result == 0) {
 290                                 spin_lock(&osc->oo_tree_lock);
 291                                 result = radix_tree_insert(&osc->oo_tree,
 292                                                            index, opg);
 293                                 if (result == 0)
 294                                         ++osc->oo_npages;
 295                                 spin_unlock(&osc->oo_tree_lock);
 296
 297                                 radix_tree_preload_end();
 298                         }
 299                 }
 300         }
 301
 302         return result;
 303 }
 304 EXPORT_SYMBOL(osc_page_init);
 305
 306 /**
 307  * Helper function called by osc_io_submit() for every page in an immediate
 308  * transfer (i.e., transferred synchronously).
 309  */
 310 void osc_page_submit(const struct lu_env *env, struct osc_page *opg,
 311                      enum cl_req_type crt, int brw_flags)
 312 {
 313         struct osc_async_page *oap = &opg->ops_oap;
 314
 315         LASSERTF(oap->oap_magic == OAP_MAGIC, "Bad oap magic: oap %p, "
 316                  "magic 0x%x\n", oap, oap->oap_magic);
 317         LASSERT(oap->oap_async_flags & ASYNC_READY);
 318         LASSERT(oap->oap_async_flags & ASYNC_COUNT_STABLE);
 319
 320         oap->oap_cmd       = crt == CRT_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ;
 321         oap->oap_page_off  = opg->ops_from;
 322         oap->oap_count     = opg->ops_to - opg->ops_from;
 323         oap->oap_brw_flags = OBD_BRW_SYNC | brw_flags;
 324
 325         if (cfs_capable(CFS_CAP_SYS_RESOURCE)) {
 326                 oap->oap_brw_flags |= OBD_BRW_NOQUOTA;
 327                 oap->oap_cmd |= OBD_BRW_NOQUOTA;
 328         }
 329
 330         opg->ops_submit_time = cfs_time_current();
 331         osc_page_transfer_get(opg, "transfer\0imm");
 332         osc_page_transfer_add(env, opg, crt);
 333 }
 334
 335 /* --------------- LRU page management ------------------ */
 336
 337 /* OSC is a natural place to manage LRU pages as applications are specialized
 338  * to write OSC by OSC. Ideally, if one OSC is used more frequently it should
 339  * occupy more LRU slots. On the other hand, we should avoid using up all LRU
 340  * slots (client_obd::cl_lru_left) otherwise process has to be put into sleep
 341  * for free LRU slots - this will be very bad so the algorithm requires each
 342  * OSC to free slots voluntarily to maintain a reasonable number of free slots
 343  * at any time.
 344  */
 345
 346 static DECLARE_WAIT_QUEUE_HEAD(osc_lru_waitq);
 347
 348 /**
 349  * LRU pages are freed in batch mode. OSC should at least free this
 350  * number of pages to avoid running out of LRU slots.
 351  */
 352 static inline int lru_shrink_min(struct client_obd *cli)
 353 {
 354         return cli->cl_max_pages_per_rpc * 2;
 355 }
 356
 357 /**
 358  * free this number at most otherwise it will take too long time to finsih.
 359  */
 360 static inline int lru_shrink_max(struct client_obd *cli)
 361 {
 362         return cli->cl_max_pages_per_rpc * cli->cl_max_rpcs_in_flight;
 363 }
 364
 365 /**
 366  * Check if we can free LRU slots from this OSC. If there exists LRU waiters,
 367  * we should free slots aggressively. In this way, slots are freed in a steady
 368  * step to maintain fairness among OSCs.
 369  *
 370  * Return how many LRU pages should be freed.
 371  */
 372 static int osc_cache_too_much(struct client_obd *cli)
 373 {
 374         struct cl_client_cache *cache = cli->cl_cache;
 375         long pages = atomic_long_read(&cli->cl_lru_in_list);
 376         unsigned long budget;
 377
 378         LASSERT(cache != NULL);
 379         budget = cache->ccc_lru_max / (atomic_read(&cache->ccc_users) - 2);
 380
 381         /* if it's going to run out LRU slots, we should free some, but not
 382          * too much to maintain faireness among OSCs. */
 383         if (atomic_long_read(cli->cl_lru_left) < cache->ccc_lru_max >> 2) {
 384                 if (pages >= budget)
 385                         return lru_shrink_max(cli);
 386                 else if (pages >= budget / 2)
 387                         return lru_shrink_min(cli);
 388         } else {
 389                 time64_t duration = ktime_get_real_seconds();
 390                 long timediff;
 391
 392                 /* knock out pages by duration of no IO activity */
 393                 duration -= cli->cl_lru_last_used;
 394                 /*
 395                  * The difference shouldn't be more than 70 years
 396                  * so we can safely case to a long. Round to
 397                  * approximately 1 minute.
 398                  */
 399                 timediff = (long)(duration >> 6);
 400                 if (timediff > 0 && pages >= budget / timediff)
 401                         return lru_shrink_min(cli);
 402         }
 403         return 0;
 404 }
 405
 406 int lru_queue_work(const struct lu_env *env, void *data)
 407 {
 408         struct client_obd *cli = data;
 409         int count;
 410
 411         CDEBUG(D_CACHE, "%s: run LRU work for client obd\n", cli_name(cli));
 412         count = osc_cache_too_much(cli);
 413         if (count > 0) {
 414                 int rc = osc_lru_shrink(env, cli, count, false);
 415
 416                 CDEBUG(D_CACHE, "%s: shrank %d/%d pages from client obd\n",
 417                        cli_name(cli), rc, count);
 418                 if (rc >= count) {
 419                         CDEBUG(D_CACHE, "%s: queue again\n", cli_name(cli));
 420                         ptlrpcd_queue_work(cli->cl_lru_work);
 421                 }
 422         }
 423
 424         RETURN(0);
 425 }
 426
 427 void osc_lru_add_batch(struct client_obd *cli, struct list_head *plist)
 428 {
 429         struct list_head lru = LIST_HEAD_INIT(lru);
 430         struct osc_async_page *oap;
 431         long npages = 0;
 432
 433         list_for_each_entry(oap, plist, oap_pending_item) {
 434                 struct osc_page *opg = oap2osc_page(oap);
 435
 436                 if (!opg->ops_in_lru)
 437                         continue;
 438
 439                 ++npages;
 440                 LASSERT(list_empty(&opg->ops_lru));
 441                 list_add(&opg->ops_lru, &lru);
 442         }
 443
 444         if (npages > 0) {
 445                 spin_lock(&cli->cl_lru_list_lock);
 446                 list_splice_tail(&lru, &cli->cl_lru_list);
 447                 atomic_long_sub(npages, &cli->cl_lru_busy);
 448                 atomic_long_add(npages, &cli->cl_lru_in_list);
 449                 cli->cl_lru_last_used = ktime_get_real_seconds();
 450                 spin_unlock(&cli->cl_lru_list_lock);
 451
 452                 if (waitqueue_active(&osc_lru_waitq))
 453                         (void)ptlrpcd_queue_work(cli->cl_lru_work);
 454         }
 455 }
 456
 457 static void __osc_lru_del(struct client_obd *cli, struct osc_page *opg)
 458 {
 459         LASSERT(atomic_long_read(&cli->cl_lru_in_list) > 0);
 460         list_del_init(&opg->ops_lru);
 461         atomic_long_dec(&cli->cl_lru_in_list);
 462 }
 463
 464 /**
 465  * Page is being destroyed. The page may be not in LRU list, if the transfer
 466  * has never finished(error occurred).
 467  */
 468 static void osc_lru_del(struct client_obd *cli, struct osc_page *opg)
 469 {
 470         if (opg->ops_in_lru) {
 471                 spin_lock(&cli->cl_lru_list_lock);
 472                 if (!list_empty(&opg->ops_lru)) {
 473                         __osc_lru_del(cli, opg);
 474                 } else {
 475                         LASSERT(atomic_long_read(&cli->cl_lru_busy) > 0);
 476                         atomic_long_dec(&cli->cl_lru_busy);
 477                 }
 478                 spin_unlock(&cli->cl_lru_list_lock);
 479
 480                 atomic_long_inc(cli->cl_lru_left);
 481                 /* this is a great place to release more LRU pages if
 482                  * this osc occupies too many LRU pages and kernel is
 483                  * stealing one of them. */
 484                 if (osc_cache_too_much(cli)) {
 485                         CDEBUG(D_CACHE, "%s: queue LRU work\n", cli_name(cli));
 486                         (void)ptlrpcd_queue_work(cli->cl_lru_work);
 487                 }
 488                 wake_up(&osc_lru_waitq);
 489         } else {
 490                 LASSERT(list_empty(&opg->ops_lru));
 491         }
 492 }
 493
 494 /**
 495  * Delete page from LRU list for redirty.
 496  */
 497 static void osc_lru_use(struct client_obd *cli, struct osc_page *opg)
 498 {
 499         /* If page is being transferred for the first time,
 500          * ops_lru should be empty */
 501         if (opg->ops_in_lru) {
 502                 spin_lock(&cli->cl_lru_list_lock);
 503                 if (!list_empty(&opg->ops_lru)) {
 504                         __osc_lru_del(cli, opg);
 505                         atomic_long_inc(&cli->cl_lru_busy);
 506                 }
 507                 spin_unlock(&cli->cl_lru_list_lock);
 508         }
 509 }
 510
 511 static void discard_pagevec(const struct lu_env *env, struct cl_io *io,
 512                                 struct cl_page **pvec, int max_index)
 513 {
 514         int i;
 515
 516         for (i = 0; i < max_index; i++) {
 517                 struct cl_page *page = pvec[i];
 518
 519                 LASSERT(cl_page_is_owned(page, io));
 520                 cl_page_delete(env, page);
 521                 cl_page_discard(env, io, page);
 522                 cl_page_disown(env, io, page);
 523                 cl_page_put(env, page);
 524
 525                 pvec[i] = NULL;
 526         }
 527 }
 528
 529 /**
 530  * Check if a cl_page can be released, i.e, it's not being used.
 531  *
 532  * If unstable account is turned on, bulk transfer may hold one refcount
 533  * for recovery so we need to check vmpage refcount as well; otherwise,
 534  * even we can destroy cl_page but the corresponding vmpage can't be reused.
 535  */
 536 static inline bool lru_page_busy(struct client_obd *cli, struct cl_page *page)
 537 {
 538         if (cl_page_in_use_noref(page))
 539                 return true;
 540
 541         if (cli->cl_cache->ccc_unstable_check) {
 542                 struct page *vmpage = cl_page_vmpage(page);
 543
 544                 /* vmpage have two known users: cl_page and VM page cache */
 545                 if (page_count(vmpage) - page_mapcount(vmpage) > 2)
 546                         return true;
 547         }
 548         return false;
 549 }
 550
 551 /**
 552  * Drop @target of pages from LRU at most.
 553  */
 554 long osc_lru_shrink(const struct lu_env *env, struct client_obd *cli,
 555                    long target, bool force)
 556 {
 557         struct cl_io *io;
 558         struct cl_object *clobj = NULL;
 559         struct cl_page **pvec;
 560         struct osc_page *opg;
 561         long count = 0;
 562         int maxscan = 0;
 563         int index = 0;
 564         int rc = 0;
 565         ENTRY;
 566
 567         LASSERT(atomic_long_read(&cli->cl_lru_in_list) >= 0);
 568         if (atomic_long_read(&cli->cl_lru_in_list) == 0 || target <= 0)
 569                 RETURN(0);
 570
 571         CDEBUG(D_CACHE, "%s: shrinkers: %d, force: %d\n",
 572                cli_name(cli), atomic_read(&cli->cl_lru_shrinkers), force);
 573         if (!force) {
 574                 if (atomic_read(&cli->cl_lru_shrinkers) > 0)
 575                         RETURN(-EBUSY);
 576
 577                 if (atomic_inc_return(&cli->cl_lru_shrinkers) > 1) {
 578                         atomic_dec(&cli->cl_lru_shrinkers);
 579                         RETURN(-EBUSY);
 580                 }
 581         } else {
 582                 atomic_inc(&cli->cl_lru_shrinkers);
 583         }
 584
 585         pvec = (struct cl_page **)osc_env_info(env)->oti_pvec;
 586         io = osc_env_thread_io(env);
 587
 588         spin_lock(&cli->cl_lru_list_lock);
 589         if (force)
 590                 cli->cl_lru_reclaim++;
 591         maxscan = min(target << 1, atomic_long_read(&cli->cl_lru_in_list));
 592         while (!list_empty(&cli->cl_lru_list)) {
 593                 struct cl_page *page;
 594                 bool will_free = false;
 595
 596                 if (!force && atomic_read(&cli->cl_lru_shrinkers) > 1)
 597                         break;
 598
 599                 if (--maxscan < 0)
 600                         break;
 601
 602                 opg = list_entry(cli->cl_lru_list.next, struct osc_page,
 603                                  ops_lru);
 604                 page = opg->ops_cl.cpl_page;
 605                 if (lru_page_busy(cli, page)) {
 606                         list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
 607                         continue;
 608                 }
 609
 610                 LASSERT(page->cp_obj != NULL);
 611                 if (clobj != page->cp_obj) {
 612                         struct cl_object *tmp = page->cp_obj;
 613
 614                         cl_object_get(tmp);
 615                         spin_unlock(&cli->cl_lru_list_lock);
 616
 617                         if (clobj != NULL) {
 618                                 discard_pagevec(env, io, pvec, index);
 619                                 index = 0;
 620
 621                                 cl_io_fini(env, io);
 622                                 cl_object_put(env, clobj);
 623                                 clobj = NULL;
 624                         }
 625
 626                         clobj = tmp;
 627                         io->ci_obj = clobj;
 628                         io->ci_ignore_layout = 1;
 629                         rc = cl_io_init(env, io, CIT_MISC, clobj);
 630
 631                         spin_lock(&cli->cl_lru_list_lock);
 632
 633                         if (rc != 0)
 634                                 break;
 635
 636                         ++maxscan;
 637                         continue;
 638                 }
 639
 640                 if (cl_page_own_try(env, io, page) == 0) {
 641                         if (!lru_page_busy(cli, page)) {
 642                                 /* remove it from lru list earlier to avoid
 643                                  * lock contention */
 644                                 __osc_lru_del(cli, opg);
 645                                 opg->ops_in_lru = 0; /* will be discarded */
 646
 647                                 cl_page_get(page);
 648                                 will_free = true;
 649                         } else {
 650                                 cl_page_disown(env, io, page);
 651                         }
 652                 }
 653
 654                 if (!will_free) {
 655                         list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
 656                         continue;
 657                 }
 658
 659                 /* Don't discard and free the page with cl_lru_list held */
 660                 pvec[index++] = page;
 661                 if (unlikely(index == OTI_PVEC_SIZE)) {
 662                         spin_unlock(&cli->cl_lru_list_lock);
 663                         discard_pagevec(env, io, pvec, index);
 664                         index = 0;
 665
 666                         spin_lock(&cli->cl_lru_list_lock);
 667                 }
 668
 669                 if (++count >= target)
 670                         break;
 671         }
 672         spin_unlock(&cli->cl_lru_list_lock);
 673
 674         if (clobj != NULL) {
 675                 discard_pagevec(env, io, pvec, index);
 676
 677                 cl_io_fini(env, io);
 678                 cl_object_put(env, clobj);
 679         }
 680
 681         atomic_dec(&cli->cl_lru_shrinkers);
 682         if (count > 0) {
 683                 atomic_long_add(count, cli->cl_lru_left);
 684                 wake_up_all(&osc_lru_waitq);
 685         }
 686         RETURN(count > 0 ? count : rc);
 687 }
 688 EXPORT_SYMBOL(osc_lru_shrink);
 689
 690 /**
 691  * Reclaim LRU pages by an IO thread. The caller wants to reclaim at least
 692  * \@npages of LRU slots. For performance consideration, it's better to drop
 693  * LRU pages in batch. Therefore, the actual number is adjusted at least
 694  * max_pages_per_rpc.
 695  */
 696 static long osc_lru_reclaim(struct client_obd *cli, unsigned long npages)
 697 {
 698         struct lu_env *env;
 699         struct cl_client_cache *cache = cli->cl_cache;
 700         int max_scans;
 701         __u16 refcheck;
 702         long rc = 0;
 703         ENTRY;
 704
 705         LASSERT(cache != NULL);
 706
 707         env = cl_env_get(&refcheck);
 708         if (IS_ERR(env))
 709                 RETURN(rc);
 710
 711         npages = max_t(int, npages, cli->cl_max_pages_per_rpc);
 712         CDEBUG(D_CACHE, "%s: start to reclaim %ld pages from LRU\n",
 713                cli_name(cli), npages);
 714         rc = osc_lru_shrink(env, cli, npages, true);
 715         if (rc >= npages) {
 716                 CDEBUG(D_CACHE, "%s: reclaimed %ld/%ld pages from LRU\n",
 717                        cli_name(cli), rc, npages);
 718                 if (osc_cache_too_much(cli) > 0)
 719                         ptlrpcd_queue_work(cli->cl_lru_work);
 720                 GOTO(out, rc);
 721         } else if (rc > 0) {
 722                 npages -= rc;
 723         }
 724
 725         CDEBUG(D_CACHE, "%s: cli %p no free slots, pages: %ld/%ld, want: %ld\n",
 726                 cli_name(cli), cli, atomic_long_read(&cli->cl_lru_in_list),
 727                 atomic_long_read(&cli->cl_lru_busy), npages);
 728
 729         /* Reclaim LRU slots from other client_obd as it can't free enough
 730          * from its own. This should rarely happen. */
 731         spin_lock(&cache->ccc_lru_lock);
 732         LASSERT(!list_empty(&cache->ccc_lru));
 733
 734         cache->ccc_lru_shrinkers++;
 735         list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
 736
 737         max_scans = atomic_read(&cache->ccc_users) - 2;
 738         while (--max_scans > 0 && !list_empty(&cache->ccc_lru)) {
 739                 cli = list_entry(cache->ccc_lru.next, struct client_obd,
 740                                  cl_lru_osc);
 741
 742                 CDEBUG(D_CACHE, "%s: cli %p LRU pages: %ld, busy: %ld.\n",
 743                         cli_name(cli), cli,
 744                         atomic_long_read(&cli->cl_lru_in_list),
 745                         atomic_long_read(&cli->cl_lru_busy));
 746
 747                 list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
 748                 if (osc_cache_too_much(cli) > 0) {
 749                         spin_unlock(&cache->ccc_lru_lock);
 750
 751                         rc = osc_lru_shrink(env, cli, npages, true);
 752                         spin_lock(&cache->ccc_lru_lock);
 753                         if (rc >= npages)
 754                                 break;
 755                         if (rc > 0)
 756                                 npages -= rc;
 757                 }
 758         }
 759         spin_unlock(&cache->ccc_lru_lock);
 760
 761 out:
 762         cl_env_put(env, &refcheck);
 763         CDEBUG(D_CACHE, "%s: cli %p freed %ld pages.\n",
 764                 cli_name(cli), cli, rc);
 765         return rc;
 766 }
 767
 768 /**
 769  * osc_lru_alloc() is called to allocate an LRU slot for a cl_page.
 770  *
 771  * Usually the LRU slots are reserved in osc_io_iter_rw_init().
 772  * Only in the case that the LRU slots are in extreme shortage, it should
 773  * have reserved enough slots for an IO.
 774  */
 775 static int osc_lru_alloc(const struct lu_env *env, struct client_obd *cli,
 776                          struct osc_page *opg)
 777 {
 778         struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
 779         struct osc_io *oio = osc_env_io(env);
 780         int rc = 0;
 781
 782         ENTRY;
 783
 784         if (cli->cl_cache == NULL) /* shall not be in LRU */
 785                 RETURN(0);
 786
 787         if (oio->oi_lru_reserved > 0) {
 788                 --oio->oi_lru_reserved;
 789                 goto out;
 790         }
 791
 792         LASSERT(atomic_long_read(cli->cl_lru_left) >= 0);
 793         while (!atomic_long_add_unless(cli->cl_lru_left, -1, 0)) {
 794                 /* run out of LRU spaces, try to drop some by itself */
 795                 rc = osc_lru_reclaim(cli, 1);
 796                 if (rc < 0)
 797                         break;
 798                 if (rc > 0)
 799                         continue;
 800
 801                 cond_resched();
 802                 rc = l_wait_event(osc_lru_waitq,
 803                                 atomic_long_read(cli->cl_lru_left) > 0,
 804                                 &lwi);
 805                 if (rc < 0)
 806                         break;
 807         }
 808
 809 out:
 810         if (rc >= 0) {
 811                 atomic_long_inc(&cli->cl_lru_busy);
 812                 opg->ops_in_lru = 1;
 813                 rc = 0;
 814         }
 815
 816         RETURN(rc);
 817 }
 818
 819 /**
 820  * osc_lru_reserve() is called to reserve enough LRU slots for I/O.
 821  *
 822  * The benefit of doing this is to reduce contention against atomic counter
 823  * cl_lru_left by changing it from per-page access to per-IO access.
 824  */
 825 unsigned long osc_lru_reserve(struct client_obd *cli, unsigned long npages)
 826 {
 827         unsigned long reserved = 0;
 828         unsigned long max_pages;
 829         unsigned long c;
 830
 831         /* reserve a full RPC window at most to avoid that a thread accidentally
 832          * consumes too many LRU slots */
 833         max_pages = cli->cl_max_pages_per_rpc * cli->cl_max_rpcs_in_flight;
 834         if (npages > max_pages)
 835                 npages = max_pages;
 836
 837         c = atomic_long_read(cli->cl_lru_left);
 838         if (c < npages && osc_lru_reclaim(cli, npages) > 0)
 839                 c = atomic_long_read(cli->cl_lru_left);
 840         while (c >= npages) {
 841                 if (c == atomic_long_cmpxchg(cli->cl_lru_left, c, c - npages)) {
 842                         reserved = npages;
 843                         break;
 844                 }
 845                 c = atomic_long_read(cli->cl_lru_left);
 846         }
 847         if (atomic_long_read(cli->cl_lru_left) < max_pages) {
 848                 /* If there aren't enough pages in the per-OSC LRU then
 849                  * wake up the LRU thread to try and clear out space, so
 850                  * we don't block if pages are being dirtied quickly. */
 851                 CDEBUG(D_CACHE, "%s: queue LRU, left: %lu/%ld.\n",
 852                        cli_name(cli), atomic_long_read(cli->cl_lru_left),
 853                        max_pages);
 854                 (void)ptlrpcd_queue_work(cli->cl_lru_work);
 855         }
 856
 857         return reserved;
 858 }
 859
 860 /**
 861  * osc_lru_unreserve() is called to unreserve LRU slots.
 862  *
 863  * LRU slots reserved by osc_lru_reserve() may have entries left due to several
 864  * reasons such as page already existing or I/O error. Those reserved slots
 865  * should be freed by calling this function.
 866  */
 867 void osc_lru_unreserve(struct client_obd *cli, unsigned long npages)
 868 {
 869         atomic_long_add(npages, cli->cl_lru_left);
 870         wake_up_all(&osc_lru_waitq);
 871 }
 872
 873 /**
 874  * Atomic operations are expensive. We accumulate the accounting for the
 875  * same page zone to get better performance.
 876  * In practice this can work pretty good because the pages in the same RPC
 877  * are likely from the same page zone.
 878  */
 879 static inline void unstable_page_accounting(struct ptlrpc_bulk_desc *desc,
 880                                             struct osc_brw_async_args *aa,
 881                                             int factor)
 882 {
 883         int page_count;
 884         void *zone = NULL;
 885         int count = 0;
 886         int i;
 887
 888         if (desc != NULL) {
 889                 LASSERT(ptlrpc_is_bulk_desc_kiov(desc->bd_type));
 890                 page_count = desc->bd_iov_count;
 891         } else {
 892                 page_count = aa->aa_page_count;
 893         }
 894
 895         for (i = 0; i < page_count; i++) {
 896                 void *pz;
 897                 if (desc)
 898                         pz = page_zone(BD_GET_KIOV(desc, i).kiov_page);
 899                 else
 900                         pz = page_zone(aa->aa_ppga[i]->pg);
 901
 902                 if (likely(pz == zone)) {
 903                         ++count;
 904                         continue;
 905                 }
 906
 907                 if (count > 0) {
 908                         mod_zone_page_state(zone, NR_UNSTABLE_NFS,
 909                                             factor * count);
 910                         count = 0;
 911                 }
 912                 zone = pz;
 913                 ++count;
 914         }
 915         if (count > 0)
 916                 mod_zone_page_state(zone, NR_UNSTABLE_NFS, factor * count);
 917 }
 918
 919 static inline void add_unstable_page_accounting(struct ptlrpc_bulk_desc *desc,
 920                                                 struct osc_brw_async_args *aa)
 921 {
 922         unstable_page_accounting(desc, aa, 1);
 923 }
 924
 925 static inline void dec_unstable_page_accounting(struct ptlrpc_bulk_desc *desc,
 926                                                 struct osc_brw_async_args *aa)
 927 {
 928         unstable_page_accounting(desc, aa, -1);
 929 }
 930
 931 /**
 932  * Performs "unstable" page accounting. This function balances the
 933  * increment operations performed in osc_inc_unstable_pages. It is
 934  * registered as the RPC request callback, and is executed when the
 935  * bulk RPC is committed on the server. Thus at this point, the pages
 936  * involved in the bulk transfer are no longer considered unstable.
 937  *
 938  * If this function is called, the request should have been committed
 939  * or req:rq_unstable must have been set; it implies that the unstable
 940  * statistic have been added.
 941  */
 942 void osc_dec_unstable_pages(struct ptlrpc_request *req)
 943 {
 944         struct ptlrpc_bulk_desc *desc       = req->rq_bulk;
 945         struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
 946         struct client_obd       *cli        = &req->rq_import->imp_obd->u.cli;
 947         int                      page_count;
 948         long                     unstable_count;
 949
 950         if (desc)
 951                 page_count = desc->bd_iov_count;
 952         else
 953                 page_count = aa->aa_page_count;
 954
 955         LASSERT(page_count >= 0);
 956
 957         dec_unstable_page_accounting(desc, aa);
 958
 959         unstable_count = atomic_long_sub_return(page_count,
 960                                                 &cli->cl_unstable_count);
 961         LASSERT(unstable_count >= 0);
 962
 963         unstable_count = atomic_long_sub_return(page_count,
 964                                            &cli->cl_cache->ccc_unstable_nr);
 965         LASSERT(unstable_count >= 0);
 966         if (unstable_count == 0)
 967                 wake_up_all(&cli->cl_cache->ccc_unstable_waitq);
 968
 969         if (waitqueue_active(&osc_lru_waitq))
 970                 (void)ptlrpcd_queue_work(cli->cl_lru_work);
 971 }
 972
 973 /**
 974  * "unstable" page accounting. See: osc_dec_unstable_pages.
 975  */
 976 void osc_inc_unstable_pages(struct ptlrpc_request *req)
 977 {
 978         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
 979         struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
 980         struct client_obd       *cli  = &req->rq_import->imp_obd->u.cli;
 981         long                     page_count;
 982
 983         /* No unstable page tracking */
 984         if (cli->cl_cache == NULL || !cli->cl_cache->ccc_unstable_check)
 985                 return;
 986
 987         if (desc)
 988                 page_count = desc->bd_iov_count;
 989         else
 990                 page_count = aa->aa_page_count;
 991
 992         add_unstable_page_accounting(desc, aa);
 993         atomic_long_add(page_count, &cli->cl_unstable_count);
 994         atomic_long_add(page_count, &cli->cl_cache->ccc_unstable_nr);
 995
 996         /* If the request has already been committed (i.e. brw_commit
 997          * called via rq_commit_cb), we need to undo the unstable page
 998          * increments we just performed because rq_commit_cb wont be
 999          * called again. */
1000         spin_lock(&req->rq_lock);
1001         if (unlikely(req->rq_committed)) {
1002                 spin_unlock(&req->rq_lock);
1003
1004                 osc_dec_unstable_pages(req);
1005         } else {
1006                 req->rq_unstable = 1;
1007                 spin_unlock(&req->rq_lock);
1008         }
1009 }
1010
1011 /**
1012  * Check if it piggybacks SOFT_SYNC flag to OST from this OSC.
1013  * This function will be called by every BRW RPC so it's critical
1014  * to make this function fast.
1015  */
1016 bool osc_over_unstable_soft_limit(struct client_obd *cli)
1017 {
1018         long unstable_nr, osc_unstable_count;
1019
1020         /* Can't check cli->cl_unstable_count, therefore, no soft limit */
1021         if (cli->cl_cache == NULL || !cli->cl_cache->ccc_unstable_check)
1022                 return false;
1023
1024         osc_unstable_count = atomic_long_read(&cli->cl_unstable_count);
1025         unstable_nr = atomic_long_read(&cli->cl_cache->ccc_unstable_nr);
1026
1027         CDEBUG(D_CACHE,
1028                "%s: cli: %p unstable pages: %lu, osc unstable pages: %lu\n",
1029                cli_name(cli), cli, unstable_nr, osc_unstable_count);
1030
1031         /* If the LRU slots are in shortage - 25% remaining AND this OSC
1032          * has one full RPC window of unstable pages, it's a good chance
1033          * to piggyback a SOFT_SYNC flag.
1034          * Please notice that the OST won't take immediate response for the
1035          * SOFT_SYNC request so active OSCs will have more chance to carry
1036          * the flag, this is reasonable. */
1037         return unstable_nr > cli->cl_cache->ccc_lru_max >> 2 &&
1038                osc_unstable_count > cli->cl_max_pages_per_rpc *
1039                                     cli->cl_max_rpcs_in_flight;
1040 }
1041
1042 /**
1043  * Return how many LRU pages in the cache of all OSC devices
1044  *
1045  * \retval      return # of cached LRU pages times reclaimation tendency
1046  * \retval      SHRINK_STOP if it cannot do any scanning in this time
1047  */
1048 unsigned long osc_cache_shrink_count(struct shrinker *sk,
1049                                      struct shrink_control *sc)
1050 {
1051         struct client_obd *cli;
1052         unsigned long cached = 0;
1053
1054         spin_lock(&osc_shrink_lock);
1055         list_for_each_entry(cli, &osc_shrink_list, cl_shrink_list)
1056                 cached += atomic_long_read(&cli->cl_lru_in_list);
1057         spin_unlock(&osc_shrink_lock);
1058
1059         return (cached  * sysctl_vfs_cache_pressure) / 100;
1060 }
1061
1062 /**
1063  * Scan and try to reclaim sc->nr_to_scan cached LRU pages
1064  *
1065  * \retval      number of cached LRU pages reclaimed
1066  * \retval      SHRINK_STOP if it cannot do any scanning in this time
1067  *
1068  * Linux kernel will loop calling this shrinker scan routine with
1069  * sc->nr_to_scan = SHRINK_BATCH(128 for now) until kernel got enough memory.
1070  *
1071  * If sc->nr_to_scan is 0, the VM is querying the cache size, we don't need
1072  * to scan and try to reclaim LRU pages, just return 0 and
1073  * osc_cache_shrink_count() will report the LRU page number.
1074  */
1075 unsigned long osc_cache_shrink_scan(struct shrinker *sk,
1076                                     struct shrink_control *sc)
1077 {
1078         struct client_obd *cli;
1079         struct client_obd *stop_anchor = NULL;
1080         struct lu_env *env;
1081         long shrank = 0;
1082         int rc;
1083         __u16 refcheck;
1084
1085         if (sc->nr_to_scan == 0)
1086                 return 0;
1087
1088         if (!(sc->gfp_mask & __GFP_FS))
1089                 return SHRINK_STOP;
1090
1091         env = cl_env_get(&refcheck);
1092         if (IS_ERR(env))
1093                 return SHRINK_STOP;
1094
1095         spin_lock(&osc_shrink_lock);
1096         while (!list_empty(&osc_shrink_list)) {
1097                 cli = list_entry(osc_shrink_list.next, struct client_obd,
1098                                  cl_shrink_list);
1099
1100                 if (stop_anchor == NULL)
1101                         stop_anchor = cli;
1102                 else if (cli == stop_anchor)
1103                         break;
1104
1105                 list_move_tail(&cli->cl_shrink_list, &osc_shrink_list);
1106                 spin_unlock(&osc_shrink_lock);
1107
1108                 /* shrink no more than max_pages_per_rpc for an OSC */
1109                 rc = osc_lru_shrink(env, cli, (sc->nr_to_scan - shrank) >
1110                                     cli->cl_max_pages_per_rpc ?
1111                                     cli->cl_max_pages_per_rpc :
1112                                     sc->nr_to_scan - shrank, true);
1113                 if (rc > 0)
1114                         shrank += rc;
1115
1116                 if (shrank >= sc->nr_to_scan)
1117                         goto out;
1118
1119                 spin_lock(&osc_shrink_lock);
1120         }
1121         spin_unlock(&osc_shrink_lock);
1122
1123 out:
1124         cl_env_put(env, &refcheck);
1125
1126         return shrank;
1127 }
1128
1129 /** @} osc */