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
 305 /**
 306  * Helper function called by osc_io_submit() for every page in an immediate
 307  * transfer (i.e., transferred synchronously).
 308  */
 309 void osc_page_submit(const struct lu_env *env, struct osc_page *opg,
 310                      enum cl_req_type crt, int brw_flags)
 311 {
 312         struct osc_async_page *oap = &opg->ops_oap;
 313
 314         LASSERTF(oap->oap_magic == OAP_MAGIC, "Bad oap magic: oap %p, "
 315                  "magic 0x%x\n", oap, oap->oap_magic);
 316         LASSERT(oap->oap_async_flags & ASYNC_READY);
 317         LASSERT(oap->oap_async_flags & ASYNC_COUNT_STABLE);
 318
 319         oap->oap_cmd       = crt == CRT_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ;
 320         oap->oap_page_off  = opg->ops_from;
 321         oap->oap_count     = opg->ops_to - opg->ops_from;
 322         oap->oap_brw_flags = OBD_BRW_SYNC | brw_flags;
 323
 324         if (cfs_capable(CFS_CAP_SYS_RESOURCE)) {
 325                 oap->oap_brw_flags |= OBD_BRW_NOQUOTA;
 326                 oap->oap_cmd |= OBD_BRW_NOQUOTA;
 327         }
 328
 329         opg->ops_submit_time = cfs_time_current();
 330         osc_page_transfer_get(opg, "transfer\0imm");
 331         osc_page_transfer_add(env, opg, crt);
 332 }
 333
 334 /* --------------- LRU page management ------------------ */
 335
 336 /* OSC is a natural place to manage LRU pages as applications are specialized
 337  * to write OSC by OSC. Ideally, if one OSC is used more frequently it should
 338  * occupy more LRU slots. On the other hand, we should avoid using up all LRU
 339  * slots (client_obd::cl_lru_left) otherwise process has to be put into sleep
 340  * for free LRU slots - this will be very bad so the algorithm requires each
 341  * OSC to free slots voluntarily to maintain a reasonable number of free slots
 342  * at any time.
 343  */
 344
 345 static DECLARE_WAIT_QUEUE_HEAD(osc_lru_waitq);
 346
 347 /**
 348  * LRU pages are freed in batch mode. OSC should at least free this
 349  * number of pages to avoid running out of LRU slots.
 350  */
 351 static inline int lru_shrink_min(struct client_obd *cli)
 352 {
 353         return cli->cl_max_pages_per_rpc * 2;
 354 }
 355
 356 /**
 357  * free this number at most otherwise it will take too long time to finsih.
 358  */
 359 static inline int lru_shrink_max(struct client_obd *cli)
 360 {
 361         return cli->cl_max_pages_per_rpc * cli->cl_max_rpcs_in_flight;
 362 }
 363
 364 /**
 365  * Check if we can free LRU slots from this OSC. If there exists LRU waiters,
 366  * we should free slots aggressively. In this way, slots are freed in a steady
 367  * step to maintain fairness among OSCs.
 368  *
 369  * Return how many LRU pages should be freed.
 370  */
 371 static int osc_cache_too_much(struct client_obd *cli)
 372 {
 373         struct cl_client_cache *cache = cli->cl_cache;
 374         long pages = atomic_long_read(&cli->cl_lru_in_list);
 375         unsigned long budget;
 376
 377         LASSERT(cache != NULL);
 378         budget = cache->ccc_lru_max / (atomic_read(&cache->ccc_users) - 2);
 379
 380         /* if it's going to run out LRU slots, we should free some, but not
 381          * too much to maintain faireness among OSCs. */
 382         if (atomic_long_read(cli->cl_lru_left) < cache->ccc_lru_max >> 2) {
 383                 if (pages >= budget)
 384                         return lru_shrink_max(cli);
 385                 else if (pages >= budget / 2)
 386                         return lru_shrink_min(cli);
 387         } else {
 388                 time64_t duration = ktime_get_real_seconds();
 389                 long timediff;
 390
 391                 /* knock out pages by duration of no IO activity */
 392                 duration -= cli->cl_lru_last_used;
 393                 /*
 394                  * The difference shouldn't be more than 70 years
 395                  * so we can safely case to a long. Round to
 396                  * approximately 1 minute.
 397                  */
 398                 timediff = (long)(duration >> 6);
 399                 if (timediff > 0 && pages >= budget / timediff)
 400                         return lru_shrink_min(cli);
 401         }
 402         return 0;
 403 }
 404
 405 int lru_queue_work(const struct lu_env *env, void *data)
 406 {
 407         struct client_obd *cli = data;
 408         int count;
 409
 410         CDEBUG(D_CACHE, "%s: run LRU work for client obd\n", cli_name(cli));
 411         count = osc_cache_too_much(cli);
 412         if (count > 0) {
 413                 int rc = osc_lru_shrink(env, cli, count, false);
 414
 415                 CDEBUG(D_CACHE, "%s: shrank %d/%d pages from client obd\n",
 416                        cli_name(cli), rc, count);
 417                 if (rc >= count) {
 418                         CDEBUG(D_CACHE, "%s: queue again\n", cli_name(cli));
 419                         ptlrpcd_queue_work(cli->cl_lru_work);
 420                 }
 421         }
 422
 423         RETURN(0);
 424 }
 425
 426 void osc_lru_add_batch(struct client_obd *cli, struct list_head *plist)
 427 {
 428         struct list_head lru = LIST_HEAD_INIT(lru);
 429         struct osc_async_page *oap;
 430         long npages = 0;
 431
 432         list_for_each_entry(oap, plist, oap_pending_item) {
 433                 struct osc_page *opg = oap2osc_page(oap);
 434
 435                 if (!opg->ops_in_lru)
 436                         continue;
 437
 438                 ++npages;
 439                 LASSERT(list_empty(&opg->ops_lru));
 440                 list_add(&opg->ops_lru, &lru);
 441         }
 442
 443         if (npages > 0) {
 444                 spin_lock(&cli->cl_lru_list_lock);
 445                 list_splice_tail(&lru, &cli->cl_lru_list);
 446                 atomic_long_sub(npages, &cli->cl_lru_busy);
 447                 atomic_long_add(npages, &cli->cl_lru_in_list);
 448                 cli->cl_lru_last_used = ktime_get_real_seconds();
 449                 spin_unlock(&cli->cl_lru_list_lock);
 450
 451                 if (waitqueue_active(&osc_lru_waitq))
 452                         (void)ptlrpcd_queue_work(cli->cl_lru_work);
 453         }
 454 }
 455
 456 static void __osc_lru_del(struct client_obd *cli, struct osc_page *opg)
 457 {
 458         LASSERT(atomic_long_read(&cli->cl_lru_in_list) > 0);
 459         list_del_init(&opg->ops_lru);
 460         atomic_long_dec(&cli->cl_lru_in_list);
 461 }
 462
 463 /**
 464  * Page is being destroyed. The page may be not in LRU list, if the transfer
 465  * has never finished(error occurred).
 466  */
 467 static void osc_lru_del(struct client_obd *cli, struct osc_page *opg)
 468 {
 469         if (opg->ops_in_lru) {
 470                 spin_lock(&cli->cl_lru_list_lock);
 471                 if (!list_empty(&opg->ops_lru)) {
 472                         __osc_lru_del(cli, opg);
 473                 } else {
 474                         LASSERT(atomic_long_read(&cli->cl_lru_busy) > 0);
 475                         atomic_long_dec(&cli->cl_lru_busy);
 476                 }
 477                 spin_unlock(&cli->cl_lru_list_lock);
 478
 479                 atomic_long_inc(cli->cl_lru_left);
 480                 /* this is a great place to release more LRU pages if
 481                  * this osc occupies too many LRU pages and kernel is
 482                  * stealing one of them. */
 483                 if (osc_cache_too_much(cli)) {
 484                         CDEBUG(D_CACHE, "%s: queue LRU work\n", cli_name(cli));
 485                         (void)ptlrpcd_queue_work(cli->cl_lru_work);
 486                 }
 487                 wake_up(&osc_lru_waitq);
 488         } else {
 489                 LASSERT(list_empty(&opg->ops_lru));
 490         }
 491 }
 492
 493 /**
 494  * Delete page from LRU list for redirty.
 495  */
 496 static void osc_lru_use(struct client_obd *cli, struct osc_page *opg)
 497 {
 498         /* If page is being transferred for the first time,
 499          * ops_lru should be empty */
 500         if (opg->ops_in_lru) {
 501                 spin_lock(&cli->cl_lru_list_lock);
 502                 if (!list_empty(&opg->ops_lru)) {
 503                         __osc_lru_del(cli, opg);
 504                         atomic_long_inc(&cli->cl_lru_busy);
 505                 }
 506                 spin_unlock(&cli->cl_lru_list_lock);
 507         }
 508 }
 509
 510 static void discard_pagevec(const struct lu_env *env, struct cl_io *io,
 511                                 struct cl_page **pvec, int max_index)
 512 {
 513         int i;
 514
 515         for (i = 0; i < max_index; i++) {
 516                 struct cl_page *page = pvec[i];
 517
 518                 LASSERT(cl_page_is_owned(page, io));
 519                 cl_page_delete(env, page);
 520                 cl_page_discard(env, io, page);
 521                 cl_page_disown(env, io, page);
 522                 cl_page_put(env, page);
 523
 524                 pvec[i] = NULL;
 525         }
 526 }
 527
 528 /**
 529  * Check if a cl_page can be released, i.e, it's not being used.
 530  *
 531  * If unstable account is turned on, bulk transfer may hold one refcount
 532  * for recovery so we need to check vmpage refcount as well; otherwise,
 533  * even we can destroy cl_page but the corresponding vmpage can't be reused.
 534  */
 535 static inline bool lru_page_busy(struct client_obd *cli, struct cl_page *page)
 536 {
 537         if (cl_page_in_use_noref(page))
 538                 return true;
 539
 540         if (cli->cl_cache->ccc_unstable_check) {
 541                 struct page *vmpage = cl_page_vmpage(page);
 542
 543                 /* vmpage have two known users: cl_page and VM page cache */
 544                 if (page_count(vmpage) - page_mapcount(vmpage) > 2)
 545                         return true;
 546         }
 547         return false;
 548 }
 549
 550 /**
 551  * Drop @target of pages from LRU at most.
 552  */
 553 long osc_lru_shrink(const struct lu_env *env, struct client_obd *cli,
 554                    long target, bool force)
 555 {
 556         struct cl_io *io;
 557         struct cl_object *clobj = NULL;
 558         struct cl_page **pvec;
 559         struct osc_page *opg;
 560         long count = 0;
 561         int maxscan = 0;
 562         int index = 0;
 563         int rc = 0;
 564         ENTRY;
 565
 566         LASSERT(atomic_long_read(&cli->cl_lru_in_list) >= 0);
 567         if (atomic_long_read(&cli->cl_lru_in_list) == 0 || target <= 0)
 568                 RETURN(0);
 569
 570         CDEBUG(D_CACHE, "%s: shrinkers: %d, force: %d\n",
 571                cli_name(cli), atomic_read(&cli->cl_lru_shrinkers), force);
 572         if (!force) {
 573                 if (atomic_read(&cli->cl_lru_shrinkers) > 0)
 574                         RETURN(-EBUSY);
 575
 576                 if (atomic_inc_return(&cli->cl_lru_shrinkers) > 1) {
 577                         atomic_dec(&cli->cl_lru_shrinkers);
 578                         RETURN(-EBUSY);
 579                 }
 580         } else {
 581                 atomic_inc(&cli->cl_lru_shrinkers);
 582         }
 583
 584         pvec = (struct cl_page **)osc_env_info(env)->oti_pvec;
 585         io = &osc_env_info(env)->oti_io;
 586
 587         spin_lock(&cli->cl_lru_list_lock);
 588         if (force)
 589                 cli->cl_lru_reclaim++;
 590         maxscan = min(target << 1, atomic_long_read(&cli->cl_lru_in_list));
 591         while (!list_empty(&cli->cl_lru_list)) {
 592                 struct cl_page *page;
 593                 bool will_free = false;
 594
 595                 if (!force && atomic_read(&cli->cl_lru_shrinkers) > 1)
 596                         break;
 597
 598                 if (--maxscan < 0)
 599                         break;
 600
 601                 opg = list_entry(cli->cl_lru_list.next, struct osc_page,
 602                                  ops_lru);
 603                 page = opg->ops_cl.cpl_page;
 604                 if (lru_page_busy(cli, page)) {
 605                         list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
 606                         continue;
 607                 }
 608
 609                 LASSERT(page->cp_obj != NULL);
 610                 if (clobj != page->cp_obj) {
 611                         struct cl_object *tmp = page->cp_obj;
 612
 613                         cl_object_get(tmp);
 614                         spin_unlock(&cli->cl_lru_list_lock);
 615
 616                         if (clobj != NULL) {
 617                                 discard_pagevec(env, io, pvec, index);
 618                                 index = 0;
 619
 620                                 cl_io_fini(env, io);
 621                                 cl_object_put(env, clobj);
 622                                 clobj = NULL;
 623                         }
 624
 625                         clobj = tmp;
 626                         io->ci_obj = clobj;
 627                         io->ci_ignore_layout = 1;
 628                         rc = cl_io_init(env, io, CIT_MISC, clobj);
 629
 630                         spin_lock(&cli->cl_lru_list_lock);
 631
 632                         if (rc != 0)
 633                                 break;
 634
 635                         ++maxscan;
 636                         continue;
 637                 }
 638
 639                 if (cl_page_own_try(env, io, page) == 0) {
 640                         if (!lru_page_busy(cli, page)) {
 641                                 /* remove it from lru list earlier to avoid
 642                                  * lock contention */
 643                                 __osc_lru_del(cli, opg);
 644                                 opg->ops_in_lru = 0; /* will be discarded */
 645
 646                                 cl_page_get(page);
 647                                 will_free = true;
 648                         } else {
 649                                 cl_page_disown(env, io, page);
 650                         }
 651                 }
 652
 653                 if (!will_free) {
 654                         list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
 655                         continue;
 656                 }
 657
 658                 /* Don't discard and free the page with cl_lru_list held */
 659                 pvec[index++] = page;
 660                 if (unlikely(index == OTI_PVEC_SIZE)) {
 661                         spin_unlock(&cli->cl_lru_list_lock);
 662                         discard_pagevec(env, io, pvec, index);
 663                         index = 0;
 664
 665                         spin_lock(&cli->cl_lru_list_lock);
 666                 }
 667
 668                 if (++count >= target)
 669                         break;
 670         }
 671         spin_unlock(&cli->cl_lru_list_lock);
 672
 673         if (clobj != NULL) {
 674                 discard_pagevec(env, io, pvec, index);
 675
 676                 cl_io_fini(env, io);
 677                 cl_object_put(env, clobj);
 678         }
 679
 680         atomic_dec(&cli->cl_lru_shrinkers);
 681         if (count > 0) {
 682                 atomic_long_add(count, cli->cl_lru_left);
 683                 wake_up_all(&osc_lru_waitq);
 684         }
 685         RETURN(count > 0 ? count : rc);
 686 }
 687
 688 /**
 689  * Reclaim LRU pages by an IO thread. The caller wants to reclaim at least
 690  * \@npages of LRU slots. For performance consideration, it's better to drop
 691  * LRU pages in batch. Therefore, the actual number is adjusted at least
 692  * max_pages_per_rpc.
 693  */
 694 static long osc_lru_reclaim(struct client_obd *cli, unsigned long npages)
 695 {
 696         struct lu_env *env;
 697         struct cl_client_cache *cache = cli->cl_cache;
 698         int max_scans;
 699         __u16 refcheck;
 700         long rc = 0;
 701         ENTRY;
 702
 703         LASSERT(cache != NULL);
 704
 705         env = cl_env_get(&refcheck);
 706         if (IS_ERR(env))
 707                 RETURN(rc);
 708
 709         npages = max_t(int, npages, cli->cl_max_pages_per_rpc);
 710         CDEBUG(D_CACHE, "%s: start to reclaim %ld pages from LRU\n",
 711                cli_name(cli), npages);
 712         rc = osc_lru_shrink(env, cli, npages, true);
 713         if (rc >= npages) {
 714                 CDEBUG(D_CACHE, "%s: reclaimed %ld/%ld pages from LRU\n",
 715                        cli_name(cli), rc, npages);
 716                 if (osc_cache_too_much(cli) > 0)
 717                         ptlrpcd_queue_work(cli->cl_lru_work);
 718                 GOTO(out, rc);
 719         } else if (rc > 0) {
 720                 npages -= rc;
 721         }
 722
 723         CDEBUG(D_CACHE, "%s: cli %p no free slots, pages: %ld/%ld, want: %ld\n",
 724                 cli_name(cli), cli, atomic_long_read(&cli->cl_lru_in_list),
 725                 atomic_long_read(&cli->cl_lru_busy), npages);
 726
 727         /* Reclaim LRU slots from other client_obd as it can't free enough
 728          * from its own. This should rarely happen. */
 729         spin_lock(&cache->ccc_lru_lock);
 730         LASSERT(!list_empty(&cache->ccc_lru));
 731
 732         cache->ccc_lru_shrinkers++;
 733         list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
 734
 735         max_scans = atomic_read(&cache->ccc_users) - 2;
 736         while (--max_scans > 0 && !list_empty(&cache->ccc_lru)) {
 737                 cli = list_entry(cache->ccc_lru.next, struct client_obd,
 738                                  cl_lru_osc);
 739
 740                 CDEBUG(D_CACHE, "%s: cli %p LRU pages: %ld, busy: %ld.\n",
 741                         cli_name(cli), cli,
 742                         atomic_long_read(&cli->cl_lru_in_list),
 743                         atomic_long_read(&cli->cl_lru_busy));
 744
 745                 list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
 746                 if (osc_cache_too_much(cli) > 0) {
 747                         spin_unlock(&cache->ccc_lru_lock);
 748
 749                         rc = osc_lru_shrink(env, cli, npages, true);
 750                         spin_lock(&cache->ccc_lru_lock);
 751                         if (rc >= npages)
 752                                 break;
 753                         if (rc > 0)
 754                                 npages -= rc;
 755                 }
 756         }
 757         spin_unlock(&cache->ccc_lru_lock);
 758
 759 out:
 760         cl_env_put(env, &refcheck);
 761         CDEBUG(D_CACHE, "%s: cli %p freed %ld pages.\n",
 762                 cli_name(cli), cli, rc);
 763         return rc;
 764 }
 765
 766 /**
 767  * osc_lru_alloc() is called to allocate an LRU slot for a cl_page.
 768  *
 769  * Usually the LRU slots are reserved in osc_io_iter_rw_init().
 770  * Only in the case that the LRU slots are in extreme shortage, it should
 771  * have reserved enough slots for an IO.
 772  */
 773 static int osc_lru_alloc(const struct lu_env *env, struct client_obd *cli,
 774                          struct osc_page *opg)
 775 {
 776         struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
 777         struct osc_io *oio = osc_env_io(env);
 778         int rc = 0;
 779         ENTRY;
 780
 781         if (cli->cl_cache == NULL) /* shall not be in LRU */
 782                 RETURN(0);
 783
 784         if (oio->oi_lru_reserved > 0) {
 785                 --oio->oi_lru_reserved;
 786                 goto out;
 787         }
 788
 789         LASSERT(atomic_long_read(cli->cl_lru_left) >= 0);
 790         while (!atomic_long_add_unless(cli->cl_lru_left, -1, 0)) {
 791                 /* run out of LRU spaces, try to drop some by itself */
 792                 rc = osc_lru_reclaim(cli, 1);
 793                 if (rc < 0)
 794                         break;
 795                 if (rc > 0)
 796                         continue;
 797
 798                 cond_resched();
 799                 rc = l_wait_event(osc_lru_waitq,
 800                                 atomic_long_read(cli->cl_lru_left) > 0,
 801                                 &lwi);
 802                 if (rc < 0)
 803                         break;
 804         }
 805
 806 out:
 807         if (rc >= 0) {
 808                 atomic_long_inc(&cli->cl_lru_busy);
 809                 opg->ops_in_lru = 1;
 810                 rc = 0;
 811         }
 812
 813         RETURN(rc);
 814 }
 815
 816 /**
 817  * osc_lru_reserve() is called to reserve enough LRU slots for I/O.
 818  *
 819  * The benefit of doing this is to reduce contention against atomic counter
 820  * cl_lru_left by changing it from per-page access to per-IO access.
 821  */
 822 unsigned long osc_lru_reserve(struct client_obd *cli, unsigned long npages)
 823 {
 824         unsigned long reserved = 0;
 825         unsigned long max_pages;
 826         unsigned long c;
 827
 828         /* reserve a full RPC window at most to avoid that a thread accidentally
 829          * consumes too many LRU slots */
 830         max_pages = cli->cl_max_pages_per_rpc * cli->cl_max_rpcs_in_flight;
 831         if (npages > max_pages)
 832                 npages = max_pages;
 833
 834         c = atomic_long_read(cli->cl_lru_left);
 835         if (c < npages && osc_lru_reclaim(cli, npages) > 0)
 836                 c = atomic_long_read(cli->cl_lru_left);
 837         while (c >= npages) {
 838                 if (c == atomic_long_cmpxchg(cli->cl_lru_left, c, c - npages)) {
 839                         reserved = npages;
 840                         break;
 841                 }
 842                 c = atomic_long_read(cli->cl_lru_left);
 843         }
 844         if (atomic_long_read(cli->cl_lru_left) < max_pages) {
 845                 /* If there aren't enough pages in the per-OSC LRU then
 846                  * wake up the LRU thread to try and clear out space, so
 847                  * we don't block if pages are being dirtied quickly. */
 848                 CDEBUG(D_CACHE, "%s: queue LRU, left: %lu/%ld.\n",
 849                        cli_name(cli), atomic_long_read(cli->cl_lru_left),
 850                        max_pages);
 851                 (void)ptlrpcd_queue_work(cli->cl_lru_work);
 852         }
 853
 854         return reserved;
 855 }
 856
 857 /**
 858  * osc_lru_unreserve() is called to unreserve LRU slots.
 859  *
 860  * LRU slots reserved by osc_lru_reserve() may have entries left due to several
 861  * reasons such as page already existing or I/O error. Those reserved slots
 862  * should be freed by calling this function.
 863  */
 864 void osc_lru_unreserve(struct client_obd *cli, unsigned long npages)
 865 {
 866         atomic_long_add(npages, cli->cl_lru_left);
 867         wake_up_all(&osc_lru_waitq);
 868 }
 869
 870 /**
 871  * Atomic operations are expensive. We accumulate the accounting for the
 872  * same page zone to get better performance.
 873  * In practice this can work pretty good because the pages in the same RPC
 874  * are likely from the same page zone.
 875  */
 876 static inline void unstable_page_accounting(struct ptlrpc_bulk_desc *desc,
 877                                             int factor)
 878 {
 879         int page_count = desc->bd_iov_count;
 880         void *zone = NULL;
 881         int count = 0;
 882         int i;
 883
 884         LASSERT(ptlrpc_is_bulk_desc_kiov(desc->bd_type));
 885
 886         for (i = 0; i < page_count; i++) {
 887                 void *pz = page_zone(BD_GET_KIOV(desc, i).kiov_page);
 888
 889                 if (likely(pz == zone)) {
 890                         ++count;
 891                         continue;
 892                 }
 893
 894                 if (count > 0) {
 895                         mod_zone_page_state(zone, NR_UNSTABLE_NFS,
 896                                             factor * count);
 897                         count = 0;
 898                 }
 899                 zone = pz;
 900                 ++count;
 901         }
 902         if (count > 0)
 903                 mod_zone_page_state(zone, NR_UNSTABLE_NFS, factor * count);
 904 }
 905
 906 static inline void add_unstable_page_accounting(struct ptlrpc_bulk_desc *desc)
 907 {
 908         unstable_page_accounting(desc, 1);
 909 }
 910
 911 static inline void dec_unstable_page_accounting(struct ptlrpc_bulk_desc *desc)
 912 {
 913         unstable_page_accounting(desc, -1);
 914 }
 915
 916 /**
 917  * Performs "unstable" page accounting. This function balances the
 918  * increment operations performed in osc_inc_unstable_pages. It is
 919  * registered as the RPC request callback, and is executed when the
 920  * bulk RPC is committed on the server. Thus at this point, the pages
 921  * involved in the bulk transfer are no longer considered unstable.
 922  *
 923  * If this function is called, the request should have been committed
 924  * or req:rq_unstable must have been set; it implies that the unstable
 925  * statistic have been added.
 926  */
 927 void osc_dec_unstable_pages(struct ptlrpc_request *req)
 928 {
 929         struct ptlrpc_bulk_desc *desc       = req->rq_bulk;
 930         struct client_obd       *cli        = &req->rq_import->imp_obd->u.cli;
 931         int                      page_count = desc->bd_iov_count;
 932         long                     unstable_count;
 933
 934         LASSERT(page_count >= 0);
 935         dec_unstable_page_accounting(desc);
 936
 937         unstable_count = atomic_long_sub_return(page_count,
 938                                                 &cli->cl_unstable_count);
 939         LASSERT(unstable_count >= 0);
 940
 941         unstable_count = atomic_long_sub_return(page_count,
 942                                            &cli->cl_cache->ccc_unstable_nr);
 943         LASSERT(unstable_count >= 0);
 944         if (unstable_count == 0)
 945                 wake_up_all(&cli->cl_cache->ccc_unstable_waitq);
 946
 947         if (waitqueue_active(&osc_lru_waitq))
 948                 (void)ptlrpcd_queue_work(cli->cl_lru_work);
 949 }
 950
 951 /**
 952  * "unstable" page accounting. See: osc_dec_unstable_pages.
 953  */
 954 void osc_inc_unstable_pages(struct ptlrpc_request *req)
 955 {
 956         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
 957         struct client_obd       *cli  = &req->rq_import->imp_obd->u.cli;
 958         long                     page_count = desc->bd_iov_count;
 959
 960         /* No unstable page tracking */
 961         if (cli->cl_cache == NULL || !cli->cl_cache->ccc_unstable_check)
 962                 return;
 963
 964         add_unstable_page_accounting(desc);
 965         atomic_long_add(page_count, &cli->cl_unstable_count);
 966         atomic_long_add(page_count, &cli->cl_cache->ccc_unstable_nr);
 967
 968         /* If the request has already been committed (i.e. brw_commit
 969          * called via rq_commit_cb), we need to undo the unstable page
 970          * increments we just performed because rq_commit_cb wont be
 971          * called again. */
 972         spin_lock(&req->rq_lock);
 973         if (unlikely(req->rq_committed)) {
 974                 spin_unlock(&req->rq_lock);
 975
 976                 osc_dec_unstable_pages(req);
 977         } else {
 978                 req->rq_unstable = 1;
 979                 spin_unlock(&req->rq_lock);
 980         }
 981 }
 982
 983 /**
 984  * Check if it piggybacks SOFT_SYNC flag to OST from this OSC.
 985  * This function will be called by every BRW RPC so it's critical
 986  * to make this function fast.
 987  */
 988 bool osc_over_unstable_soft_limit(struct client_obd *cli)
 989 {
 990         long unstable_nr, osc_unstable_count;
 991
 992         /* Can't check cli->cl_unstable_count, therefore, no soft limit */
 993         if (cli->cl_cache == NULL || !cli->cl_cache->ccc_unstable_check)
 994                 return false;
 995
 996         osc_unstable_count = atomic_long_read(&cli->cl_unstable_count);
 997         unstable_nr = atomic_long_read(&cli->cl_cache->ccc_unstable_nr);
 998
 999         CDEBUG(D_CACHE,
1000                "%s: cli: %p unstable pages: %lu, osc unstable pages: %lu\n",
1001                cli_name(cli), cli, unstable_nr, osc_unstable_count);
1002
1003         /* If the LRU slots are in shortage - 25% remaining AND this OSC
1004          * has one full RPC window of unstable pages, it's a good chance
1005          * to piggyback a SOFT_SYNC flag.
1006          * Please notice that the OST won't take immediate response for the
1007          * SOFT_SYNC request so active OSCs will have more chance to carry
1008          * the flag, this is reasonable. */
1009         return unstable_nr > cli->cl_cache->ccc_lru_max >> 2 &&
1010                osc_unstable_count > cli->cl_max_pages_per_rpc *
1011                                     cli->cl_max_rpcs_in_flight;
1012 }
1013
1014 /**
1015  * Return how many LRU pages in the cache of all OSC devices
1016  *
1017  * \retval      return # of cached LRU pages times reclaimation tendency
1018  * \retval      SHRINK_STOP if it cannot do any scanning in this time
1019  */
1020 unsigned long osc_cache_shrink_count(struct shrinker *sk,
1021                                      struct shrink_control *sc)
1022 {
1023         struct client_obd *cli;
1024         unsigned long cached = 0;
1025
1026         spin_lock(&osc_shrink_lock);
1027         list_for_each_entry(cli, &osc_shrink_list, cl_shrink_list)
1028                 cached += atomic_long_read(&cli->cl_lru_in_list);
1029         spin_unlock(&osc_shrink_lock);
1030
1031         return (cached  * sysctl_vfs_cache_pressure) / 100;
1032 }
1033
1034 /**
1035  * Scan and try to reclaim sc->nr_to_scan cached LRU pages
1036  *
1037  * \retval      number of cached LRU pages reclaimed
1038  * \retval      SHRINK_STOP if it cannot do any scanning in this time
1039  *
1040  * Linux kernel will loop calling this shrinker scan routine with
1041  * sc->nr_to_scan = SHRINK_BATCH(128 for now) until kernel got enough memory.
1042  *
1043  * If sc->nr_to_scan is 0, the VM is querying the cache size, we don't need
1044  * to scan and try to reclaim LRU pages, just return 0 and
1045  * osc_cache_shrink_count() will report the LRU page number.
1046  */
1047 unsigned long osc_cache_shrink_scan(struct shrinker *sk,
1048                                     struct shrink_control *sc)
1049 {
1050         struct client_obd *cli;
1051         struct client_obd *stop_anchor = NULL;
1052         struct lu_env *env;
1053         long shrank = 0;
1054         int rc;
1055         __u16 refcheck;
1056
1057         if (sc->nr_to_scan == 0)
1058                 return 0;
1059
1060         if (!(sc->gfp_mask & __GFP_FS))
1061                 return SHRINK_STOP;
1062
1063         env = cl_env_get(&refcheck);
1064         if (IS_ERR(env))
1065                 return SHRINK_STOP;
1066
1067         spin_lock(&osc_shrink_lock);
1068         while (!list_empty(&osc_shrink_list)) {
1069                 cli = list_entry(osc_shrink_list.next, struct client_obd,
1070                                  cl_shrink_list);
1071
1072                 if (stop_anchor == NULL)
1073                         stop_anchor = cli;
1074                 else if (cli == stop_anchor)
1075                         break;
1076
1077                 list_move_tail(&cli->cl_shrink_list, &osc_shrink_list);
1078                 spin_unlock(&osc_shrink_lock);
1079
1080                 /* shrink no more than max_pages_per_rpc for an OSC */
1081                 rc = osc_lru_shrink(env, cli, (sc->nr_to_scan - shrank) >
1082                                     cli->cl_max_pages_per_rpc ?
1083                                     cli->cl_max_pages_per_rpc :
1084                                     sc->nr_to_scan - shrank, true);
1085                 if (rc > 0)
1086                         shrank += rc;
1087
1088                 if (shrank >= sc->nr_to_scan)
1089                         goto out;
1090
1091                 spin_lock(&osc_shrink_lock);
1092         }
1093         spin_unlock(&osc_shrink_lock);
1094
1095 out:
1096         cl_env_put(env, &refcheck);
1097
1098         return shrank;
1099 }
1100
1101 /** @} osc */