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, 2015, 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
  40 #include "osc_cl_internal.h"
  41
  42 static void osc_lru_del(struct client_obd *cli, struct osc_page *opg);
  43 static void osc_lru_use(struct client_obd *cli, struct osc_page *opg);
  44 static int osc_lru_alloc(const struct lu_env *env, struct client_obd *cli,
  45                          struct osc_page *opg);
  46
  47 /** \addtogroup osc
  48  *  @{
  49  */
  50
  51 /*
  52  * Page operations.
  53  */
  54 static void osc_page_transfer_get(struct osc_page *opg, const char *label)
  55 {
  56         struct cl_page *page = opg->ops_cl.cpl_page;
  57
  58         LASSERT(!opg->ops_transfer_pinned);
  59         cl_page_get(page);
  60         lu_ref_add_atomic(&page->cp_reference, label, page);
  61         opg->ops_transfer_pinned = 1;
  62 }
  63
  64 static void osc_page_transfer_put(const struct lu_env *env,
  65                                   struct osc_page *opg)
  66 {
  67         struct cl_page *page = opg->ops_cl.cpl_page;
  68
  69         if (opg->ops_transfer_pinned) {
  70                 opg->ops_transfer_pinned = 0;
  71                 lu_ref_del(&page->cp_reference, "transfer", page);
  72                 cl_page_put(env, page);
  73         }
  74 }
  75
  76 /**
  77  * This is called once for every page when it is submitted for a transfer
  78  * either opportunistic (osc_page_cache_add()), or immediate
  79  * (osc_page_submit()).
  80  */
  81 static void osc_page_transfer_add(const struct lu_env *env,
  82                                   struct osc_page *opg, enum cl_req_type crt)
  83 {
  84         struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
  85
  86         /* ops_lru and ops_inflight share the same field, so take it from LRU
  87          * first and then use it as inflight. */
  88         osc_lru_use(osc_cli(obj), opg);
  89 }
  90
  91 int osc_page_cache_add(const struct lu_env *env,
  92                         const struct cl_page_slice *slice, struct cl_io *io)
  93 {
  94         struct osc_page *opg = cl2osc_page(slice);
  95         int result;
  96         ENTRY;
  97
  98         osc_page_transfer_get(opg, "transfer\0cache");
  99         result = osc_queue_async_io(env, io, opg);
 100         if (result != 0)
 101                 osc_page_transfer_put(env, opg);
 102         else
 103                 osc_page_transfer_add(env, opg, CRT_WRITE);
 104
 105         RETURN(result);
 106 }
 107
 108 void osc_index2policy(union ldlm_policy_data *policy,
 109                       const struct cl_object *obj, pgoff_t start, pgoff_t end)
 110 {
 111         memset(policy, 0, sizeof *policy);
 112         policy->l_extent.start = cl_offset(obj, start);
 113         policy->l_extent.end   = cl_offset(obj, end + 1) - 1;
 114 }
 115
 116 static const char *osc_list(struct list_head *head)
 117 {
 118         return list_empty(head) ? "-" : "+";
 119 }
 120
 121 static inline cfs_time_t osc_submit_duration(struct osc_page *opg)
 122 {
 123         if (opg->ops_submit_time == 0)
 124                 return 0;
 125
 126         return (cfs_time_current() - opg->ops_submit_time);
 127 }
 128
 129 static int osc_page_print(const struct lu_env *env,
 130                           const struct cl_page_slice *slice,
 131                           void *cookie, lu_printer_t printer)
 132 {
 133         struct osc_page       *opg = cl2osc_page(slice);
 134         struct osc_async_page *oap = &opg->ops_oap;
 135         struct osc_object     *obj = cl2osc(slice->cpl_obj);
 136         struct client_obd     *cli = &osc_export(obj)->exp_obd->u.cli;
 137
 138         return (*printer)(env, cookie, LUSTRE_OSC_NAME"-page@%p %lu: "
 139                           "1< %#x %d %u %s %s > "
 140                           "2< %lld %u %u %#x %#x | %p %p %p > "
 141                           "3< %d %lu %d > "
 142                           "4< %d %d %d %lu %s | %s %s %s %s > "
 143                           "5< %s %s %s %s | %d %s | %d %s %s>\n",
 144                           opg, osc_index(opg),
 145                           /* 1 */
 146                           oap->oap_magic, oap->oap_cmd,
 147                           oap->oap_interrupted,
 148                           osc_list(&oap->oap_pending_item),
 149                           osc_list(&oap->oap_rpc_item),
 150                           /* 2 */
 151                           oap->oap_obj_off, oap->oap_page_off, oap->oap_count,
 152                           oap->oap_async_flags, oap->oap_brw_flags,
 153                           oap->oap_request, oap->oap_cli, obj,
 154                           /* 3 */
 155                           opg->ops_transfer_pinned,
 156                           osc_submit_duration(opg), opg->ops_srvlock,
 157                           /* 4 */
 158                           cli->cl_r_in_flight, cli->cl_w_in_flight,
 159                           cli->cl_max_rpcs_in_flight,
 160                           cli->cl_avail_grant,
 161                           osc_list(&cli->cl_cache_waiters),
 162                           osc_list(&cli->cl_loi_ready_list),
 163                           osc_list(&cli->cl_loi_hp_ready_list),
 164                           osc_list(&cli->cl_loi_write_list),
 165                           osc_list(&cli->cl_loi_read_list),
 166                           /* 5 */
 167                           osc_list(&obj->oo_ready_item),
 168                           osc_list(&obj->oo_hp_ready_item),
 169                           osc_list(&obj->oo_write_item),
 170                           osc_list(&obj->oo_read_item),
 171                           atomic_read(&obj->oo_nr_reads),
 172                           osc_list(&obj->oo_reading_exts),
 173                           atomic_read(&obj->oo_nr_writes),
 174                           osc_list(&obj->oo_hp_exts),
 175                           osc_list(&obj->oo_urgent_exts));
 176 }
 177
 178 static void osc_page_delete(const struct lu_env *env,
 179                             const struct cl_page_slice *slice)
 180 {
 181         struct osc_page   *opg = cl2osc_page(slice);
 182         struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
 183         int rc;
 184
 185         ENTRY;
 186         CDEBUG(D_TRACE, "%p\n", opg);
 187         osc_page_transfer_put(env, opg);
 188         rc = osc_teardown_async_page(env, obj, opg);
 189         if (rc) {
 190                 CL_PAGE_DEBUG(D_ERROR, env, slice->cpl_page,
 191                               "Trying to teardown failed: %d\n", rc);
 192                 LASSERT(0);
 193         }
 194
 195         osc_lru_del(osc_cli(obj), opg);
 196
 197         if (slice->cpl_page->cp_type == CPT_CACHEABLE) {
 198                 void *value;
 199
 200                 spin_lock(&obj->oo_tree_lock);
 201                 value = radix_tree_delete(&obj->oo_tree, osc_index(opg));
 202                 if (value != NULL)
 203                         --obj->oo_npages;
 204                 spin_unlock(&obj->oo_tree_lock);
 205
 206                 LASSERT(ergo(value != NULL, value == opg));
 207         }
 208
 209         EXIT;
 210 }
 211
 212 static void osc_page_clip(const struct lu_env *env,
 213                           const struct cl_page_slice *slice,
 214                           int from, int to)
 215 {
 216         struct osc_page       *opg = cl2osc_page(slice);
 217         struct osc_async_page *oap = &opg->ops_oap;
 218
 219         opg->ops_from = from;
 220         opg->ops_to   = to;
 221         spin_lock(&oap->oap_lock);
 222         oap->oap_async_flags |= ASYNC_COUNT_STABLE;
 223         spin_unlock(&oap->oap_lock);
 224 }
 225
 226 static int osc_page_cancel(const struct lu_env *env,
 227                            const struct cl_page_slice *slice)
 228 {
 229         struct osc_page *opg = cl2osc_page(slice);
 230         int rc = 0;
 231
 232         /* Check if the transferring against this page
 233          * is completed, or not even queued. */
 234         if (opg->ops_transfer_pinned)
 235                 /* FIXME: may not be interrupted.. */
 236                 rc = osc_cancel_async_page(env, opg);
 237         LASSERT(ergo(rc == 0, opg->ops_transfer_pinned == 0));
 238         return rc;
 239 }
 240
 241 static int osc_page_flush(const struct lu_env *env,
 242                           const struct cl_page_slice *slice,
 243                           struct cl_io *io)
 244 {
 245         struct osc_page *opg = cl2osc_page(slice);
 246         int rc = 0;
 247         ENTRY;
 248         rc = osc_flush_async_page(env, io, opg);
 249         RETURN(rc);
 250 }
 251
 252 static const struct cl_page_operations osc_page_ops = {
 253         .cpo_print         = osc_page_print,
 254         .cpo_delete        = osc_page_delete,
 255         .cpo_clip           = osc_page_clip,
 256         .cpo_cancel         = osc_page_cancel,
 257         .cpo_flush          = osc_page_flush
 258 };
 259
 260 int osc_page_init(const struct lu_env *env, struct cl_object *obj,
 261                   struct cl_page *page, pgoff_t index)
 262 {
 263         struct osc_object *osc = cl2osc(obj);
 264         struct osc_page   *opg = cl_object_page_slice(obj, page);
 265         int result;
 266
 267         opg->ops_from = 0;
 268         opg->ops_to   = PAGE_SIZE;
 269
 270         result = osc_prep_async_page(osc, opg, page->cp_vmpage,
 271                                      cl_offset(obj, index));
 272         if (result == 0) {
 273                 struct osc_io *oio = osc_env_io(env);
 274                 opg->ops_srvlock = osc_io_srvlock(oio);
 275                 cl_page_slice_add(page, &opg->ops_cl, obj, index,
 276                                   &osc_page_ops);
 277         }
 278         INIT_LIST_HEAD(&opg->ops_lru);
 279
 280         /* reserve an LRU space for this page */
 281         if (page->cp_type == CPT_CACHEABLE && result == 0) {
 282                 result = osc_lru_alloc(env, osc_cli(osc), opg);
 283                 if (result == 0) {
 284                         spin_lock(&osc->oo_tree_lock);
 285                         result = radix_tree_insert(&osc->oo_tree, index, opg);
 286                         if (result == 0)
 287                                 ++osc->oo_npages;
 288                         spin_unlock(&osc->oo_tree_lock);
 289                         LASSERT(result == 0);
 290                 }
 291         }
 292
 293         return result;
 294 }
 295
 296 /**
 297  * Helper function called by osc_io_submit() for every page in an immediate
 298  * transfer (i.e., transferred synchronously).
 299  */
 300 void osc_page_submit(const struct lu_env *env, struct osc_page *opg,
 301                      enum cl_req_type crt, int brw_flags)
 302 {
 303         struct osc_async_page *oap = &opg->ops_oap;
 304
 305         LASSERTF(oap->oap_magic == OAP_MAGIC, "Bad oap magic: oap %p, "
 306                  "magic 0x%x\n", oap, oap->oap_magic);
 307         LASSERT(oap->oap_async_flags & ASYNC_READY);
 308         LASSERT(oap->oap_async_flags & ASYNC_COUNT_STABLE);
 309
 310         oap->oap_cmd       = crt == CRT_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ;
 311         oap->oap_page_off  = opg->ops_from;
 312         oap->oap_count     = opg->ops_to - opg->ops_from;
 313         oap->oap_brw_flags = OBD_BRW_SYNC | brw_flags;
 314
 315         if (cfs_capable(CFS_CAP_SYS_RESOURCE)) {
 316                 oap->oap_brw_flags |= OBD_BRW_NOQUOTA;
 317                 oap->oap_cmd |= OBD_BRW_NOQUOTA;
 318         }
 319
 320         opg->ops_submit_time = cfs_time_current();
 321         osc_page_transfer_get(opg, "transfer\0imm");
 322         osc_page_transfer_add(env, opg, crt);
 323 }
 324
 325 /* --------------- LRU page management ------------------ */
 326
 327 /* OSC is a natural place to manage LRU pages as applications are specialized
 328  * to write OSC by OSC. Ideally, if one OSC is used more frequently it should
 329  * occupy more LRU slots. On the other hand, we should avoid using up all LRU
 330  * slots (client_obd::cl_lru_left) otherwise process has to be put into sleep
 331  * for free LRU slots - this will be very bad so the algorithm requires each
 332  * OSC to free slots voluntarily to maintain a reasonable number of free slots
 333  * at any time.
 334  */
 335
 336 static DECLARE_WAIT_QUEUE_HEAD(osc_lru_waitq);
 337
 338 /**
 339  * LRU pages are freed in batch mode. OSC should at least free this
 340  * number of pages to avoid running out of LRU slots.
 341  */
 342 static inline int lru_shrink_min(struct client_obd *cli)
 343 {
 344         return cli->cl_max_pages_per_rpc * 2;
 345 }
 346
 347 /**
 348  * free this number at most otherwise it will take too long time to finsih.
 349  */
 350 static inline int lru_shrink_max(struct client_obd *cli)
 351 {
 352         return cli->cl_max_pages_per_rpc * cli->cl_max_rpcs_in_flight;
 353 }
 354
 355 /**
 356  * Check if we can free LRU slots from this OSC. If there exists LRU waiters,
 357  * we should free slots aggressively. In this way, slots are freed in a steady
 358  * step to maintain fairness among OSCs.
 359  *
 360  * Return how many LRU pages should be freed.
 361  */
 362 static int osc_cache_too_much(struct client_obd *cli)
 363 {
 364         struct cl_client_cache *cache = cli->cl_cache;
 365         long pages = atomic_long_read(&cli->cl_lru_in_list);
 366         unsigned long budget;
 367
 368         LASSERT(cache != NULL);
 369         budget = cache->ccc_lru_max / (atomic_read(&cache->ccc_users) - 2);
 370
 371         /* if it's going to run out LRU slots, we should free some, but not
 372          * too much to maintain faireness among OSCs. */
 373         if (atomic_long_read(cli->cl_lru_left) < cache->ccc_lru_max >> 2) {
 374                 if (pages >= budget)
 375                         return lru_shrink_max(cli);
 376                 else if (pages >= budget / 2)
 377                         return lru_shrink_min(cli);
 378         } else {
 379                 int duration = cfs_time_current_sec() - cli->cl_lru_last_used;
 380
 381                 /* knock out pages by duration of no IO activity */
 382                 duration >>= 6; /* approximately 1 minute */
 383                 if (duration > 0 && pages >= budget / duration)
 384                         return lru_shrink_min(cli);
 385         }
 386         return 0;
 387 }
 388
 389 int lru_queue_work(const struct lu_env *env, void *data)
 390 {
 391         struct client_obd *cli = data;
 392         int count;
 393
 394         CDEBUG(D_CACHE, "%s: run LRU work for client obd\n", cli_name(cli));
 395         count = osc_cache_too_much(cli);
 396         if (count > 0) {
 397                 int rc = osc_lru_shrink(env, cli, count, false);
 398
 399                 CDEBUG(D_CACHE, "%s: shrank %d/%d pages from client obd\n",
 400                        cli_name(cli), rc, count);
 401                 if (rc >= count) {
 402                         CDEBUG(D_CACHE, "%s: queue again\n", cli_name(cli));
 403                         ptlrpcd_queue_work(cli->cl_lru_work);
 404                 }
 405         }
 406
 407         RETURN(0);
 408 }
 409
 410 void osc_lru_add_batch(struct client_obd *cli, struct list_head *plist)
 411 {
 412         struct list_head lru = LIST_HEAD_INIT(lru);
 413         struct osc_async_page *oap;
 414         long npages = 0;
 415
 416         list_for_each_entry(oap, plist, oap_pending_item) {
 417                 struct osc_page *opg = oap2osc_page(oap);
 418
 419                 if (!opg->ops_in_lru)
 420                         continue;
 421
 422                 ++npages;
 423                 LASSERT(list_empty(&opg->ops_lru));
 424                 list_add(&opg->ops_lru, &lru);
 425         }
 426
 427         if (npages > 0) {
 428                 spin_lock(&cli->cl_lru_list_lock);
 429                 list_splice_tail(&lru, &cli->cl_lru_list);
 430                 atomic_long_sub(npages, &cli->cl_lru_busy);
 431                 atomic_long_add(npages, &cli->cl_lru_in_list);
 432                 cli->cl_lru_last_used = cfs_time_current_sec();
 433                 spin_unlock(&cli->cl_lru_list_lock);
 434
 435                 if (waitqueue_active(&osc_lru_waitq))
 436                         (void)ptlrpcd_queue_work(cli->cl_lru_work);
 437         }
 438 }
 439
 440 static void __osc_lru_del(struct client_obd *cli, struct osc_page *opg)
 441 {
 442         LASSERT(atomic_long_read(&cli->cl_lru_in_list) > 0);
 443         list_del_init(&opg->ops_lru);
 444         atomic_long_dec(&cli->cl_lru_in_list);
 445 }
 446
 447 /**
 448  * Page is being destroyed. The page may be not in LRU list, if the transfer
 449  * has never finished(error occurred).
 450  */
 451 static void osc_lru_del(struct client_obd *cli, struct osc_page *opg)
 452 {
 453         if (opg->ops_in_lru) {
 454                 spin_lock(&cli->cl_lru_list_lock);
 455                 if (!list_empty(&opg->ops_lru)) {
 456                         __osc_lru_del(cli, opg);
 457                 } else {
 458                         LASSERT(atomic_long_read(&cli->cl_lru_busy) > 0);
 459                         atomic_long_dec(&cli->cl_lru_busy);
 460                 }
 461                 spin_unlock(&cli->cl_lru_list_lock);
 462
 463                 atomic_long_inc(cli->cl_lru_left);
 464                 /* this is a great place to release more LRU pages if
 465                  * this osc occupies too many LRU pages and kernel is
 466                  * stealing one of them. */
 467                 if (osc_cache_too_much(cli)) {
 468                         CDEBUG(D_CACHE, "%s: queue LRU work\n", cli_name(cli));
 469                         (void)ptlrpcd_queue_work(cli->cl_lru_work);
 470                 }
 471                 wake_up(&osc_lru_waitq);
 472         } else {
 473                 LASSERT(list_empty(&opg->ops_lru));
 474         }
 475 }
 476
 477 /**
 478  * Delete page from LRUlist for redirty.
 479  */
 480 static void osc_lru_use(struct client_obd *cli, struct osc_page *opg)
 481 {
 482         /* If page is being transferred for the first time,
 483          * ops_lru should be empty */
 484         if (opg->ops_in_lru && !list_empty(&opg->ops_lru)) {
 485                 spin_lock(&cli->cl_lru_list_lock);
 486                 __osc_lru_del(cli, opg);
 487                 spin_unlock(&cli->cl_lru_list_lock);
 488                 atomic_long_inc(&cli->cl_lru_busy);
 489         }
 490 }
 491
 492 static void discard_pagevec(const struct lu_env *env, struct cl_io *io,
 493                                 struct cl_page **pvec, int max_index)
 494 {
 495         int i;
 496
 497         for (i = 0; i < max_index; i++) {
 498                 struct cl_page *page = pvec[i];
 499
 500                 LASSERT(cl_page_is_owned(page, io));
 501                 cl_page_delete(env, page);
 502                 cl_page_discard(env, io, page);
 503                 cl_page_disown(env, io, page);
 504                 cl_page_put(env, page);
 505
 506                 pvec[i] = NULL;
 507         }
 508 }
 509
 510 /**
 511  * Check if a cl_page can be released, i.e, it's not being used.
 512  *
 513  * If unstable account is turned on, bulk transfer may hold one refcount
 514  * for recovery so we need to check vmpage refcount as well; otherwise,
 515  * even we can destroy cl_page but the corresponding vmpage can't be reused.
 516  */
 517 static inline bool lru_page_busy(struct client_obd *cli, struct cl_page *page)
 518 {
 519         if (cl_page_in_use_noref(page))
 520                 return true;
 521
 522         if (cli->cl_cache->ccc_unstable_check) {
 523                 struct page *vmpage = cl_page_vmpage(page);
 524
 525                 /* vmpage have two known users: cl_page and VM page cache */
 526                 if (page_count(vmpage) - page_mapcount(vmpage) > 2)
 527                         return true;
 528         }
 529         return false;
 530 }
 531
 532 /**
 533  * Drop @target of pages from LRU at most.
 534  */
 535 long osc_lru_shrink(const struct lu_env *env, struct client_obd *cli,
 536                    long target, bool force)
 537 {
 538         struct cl_io *io;
 539         struct cl_object *clobj = NULL;
 540         struct cl_page **pvec;
 541         struct osc_page *opg;
 542         long count = 0;
 543         int maxscan = 0;
 544         int index = 0;
 545         int rc = 0;
 546         ENTRY;
 547
 548         LASSERT(atomic_long_read(&cli->cl_lru_in_list) >= 0);
 549         if (atomic_long_read(&cli->cl_lru_in_list) == 0 || target <= 0)
 550                 RETURN(0);
 551
 552         CDEBUG(D_CACHE, "%s: shrinkers: %d, force: %d\n",
 553                cli_name(cli), atomic_read(&cli->cl_lru_shrinkers), force);
 554         if (!force) {
 555                 if (atomic_read(&cli->cl_lru_shrinkers) > 0)
 556                         RETURN(-EBUSY);
 557
 558                 if (atomic_inc_return(&cli->cl_lru_shrinkers) > 1) {
 559                         atomic_dec(&cli->cl_lru_shrinkers);
 560                         RETURN(-EBUSY);
 561                 }
 562         } else {
 563                 atomic_inc(&cli->cl_lru_shrinkers);
 564         }
 565
 566         pvec = (struct cl_page **)osc_env_info(env)->oti_pvec;
 567         io = &osc_env_info(env)->oti_io;
 568
 569         spin_lock(&cli->cl_lru_list_lock);
 570         if (force)
 571                 cli->cl_lru_reclaim++;
 572         maxscan = min(target << 1, atomic_long_read(&cli->cl_lru_in_list));
 573         while (!list_empty(&cli->cl_lru_list)) {
 574                 struct cl_page *page;
 575                 bool will_free = false;
 576
 577                 if (!force && atomic_read(&cli->cl_lru_shrinkers) > 1)
 578                         break;
 579
 580                 if (--maxscan < 0)
 581                         break;
 582
 583                 opg = list_entry(cli->cl_lru_list.next, struct osc_page,
 584                                  ops_lru);
 585                 page = opg->ops_cl.cpl_page;
 586                 if (lru_page_busy(cli, page)) {
 587                         list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
 588                         continue;
 589                 }
 590
 591                 LASSERT(page->cp_obj != NULL);
 592                 if (clobj != page->cp_obj) {
 593                         struct cl_object *tmp = page->cp_obj;
 594
 595                         cl_object_get(tmp);
 596                         spin_unlock(&cli->cl_lru_list_lock);
 597
 598                         if (clobj != NULL) {
 599                                 discard_pagevec(env, io, pvec, index);
 600                                 index = 0;
 601
 602                                 cl_io_fini(env, io);
 603                                 cl_object_put(env, clobj);
 604                                 clobj = NULL;
 605                         }
 606
 607                         clobj = tmp;
 608                         io->ci_obj = clobj;
 609                         io->ci_ignore_layout = 1;
 610                         rc = cl_io_init(env, io, CIT_MISC, clobj);
 611
 612                         spin_lock(&cli->cl_lru_list_lock);
 613
 614                         if (rc != 0)
 615                                 break;
 616
 617                         ++maxscan;
 618                         continue;
 619                 }
 620
 621                 if (cl_page_own_try(env, io, page) == 0) {
 622                         if (!lru_page_busy(cli, page)) {
 623                                 /* remove it from lru list earlier to avoid
 624                                  * lock contention */
 625                                 __osc_lru_del(cli, opg);
 626                                 opg->ops_in_lru = 0; /* will be discarded */
 627
 628                                 cl_page_get(page);
 629                                 will_free = true;
 630                         } else {
 631                                 cl_page_disown(env, io, page);
 632                         }
 633                 }
 634
 635                 if (!will_free) {
 636                         list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
 637                         continue;
 638                 }
 639
 640                 /* Don't discard and free the page with cl_lru_list held */
 641                 pvec[index++] = page;
 642                 if (unlikely(index == OTI_PVEC_SIZE)) {
 643                         spin_unlock(&cli->cl_lru_list_lock);
 644                         discard_pagevec(env, io, pvec, index);
 645                         index = 0;
 646
 647                         spin_lock(&cli->cl_lru_list_lock);
 648                 }
 649
 650                 if (++count >= target)
 651                         break;
 652         }
 653         spin_unlock(&cli->cl_lru_list_lock);
 654
 655         if (clobj != NULL) {
 656                 discard_pagevec(env, io, pvec, index);
 657
 658                 cl_io_fini(env, io);
 659                 cl_object_put(env, clobj);
 660         }
 661
 662         atomic_dec(&cli->cl_lru_shrinkers);
 663         if (count > 0) {
 664                 atomic_long_add(count, cli->cl_lru_left);
 665                 wake_up_all(&osc_lru_waitq);
 666         }
 667         RETURN(count > 0 ? count : rc);
 668 }
 669
 670 /**
 671  * Reclaim LRU pages by an IO thread. The caller wants to reclaim at least
 672  * \@npages of LRU slots. For performance consideration, it's better to drop
 673  * LRU pages in batch. Therefore, the actual number is adjusted at least
 674  * max_pages_per_rpc.
 675  */
 676 static long osc_lru_reclaim(struct client_obd *cli, unsigned long npages)
 677 {
 678         struct lu_env *env;
 679         struct cl_client_cache *cache = cli->cl_cache;
 680         int max_scans;
 681         __u16 refcheck;
 682         long rc = 0;
 683         ENTRY;
 684
 685         LASSERT(cache != NULL);
 686
 687         env = cl_env_get(&refcheck);
 688         if (IS_ERR(env))
 689                 RETURN(rc);
 690
 691         npages = max_t(int, npages, cli->cl_max_pages_per_rpc);
 692         CDEBUG(D_CACHE, "%s: start to reclaim %ld pages from LRU\n",
 693                cli_name(cli), npages);
 694         rc = osc_lru_shrink(env, cli, npages, true);
 695         if (rc >= npages) {
 696                 CDEBUG(D_CACHE, "%s: reclaimed %ld/%ld pages from LRU\n",
 697                        cli_name(cli), rc, npages);
 698                 if (osc_cache_too_much(cli) > 0)
 699                         ptlrpcd_queue_work(cli->cl_lru_work);
 700                 GOTO(out, rc);
 701         } else if (rc > 0) {
 702                 npages -= rc;
 703         }
 704
 705         CDEBUG(D_CACHE, "%s: cli %p no free slots, pages: %ld/%ld, want: %ld\n",
 706                 cli_name(cli), cli, atomic_long_read(&cli->cl_lru_in_list),
 707                 atomic_long_read(&cli->cl_lru_busy), npages);
 708
 709         /* Reclaim LRU slots from other client_obd as it can't free enough
 710          * from its own. This should rarely happen. */
 711         spin_lock(&cache->ccc_lru_lock);
 712         LASSERT(!list_empty(&cache->ccc_lru));
 713
 714         cache->ccc_lru_shrinkers++;
 715         list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
 716
 717         max_scans = atomic_read(&cache->ccc_users) - 2;
 718         while (--max_scans > 0 && !list_empty(&cache->ccc_lru)) {
 719                 cli = list_entry(cache->ccc_lru.next, struct client_obd,
 720                                  cl_lru_osc);
 721
 722                 CDEBUG(D_CACHE, "%s: cli %p LRU pages: %ld, busy: %ld.\n",
 723                         cli_name(cli), cli,
 724                         atomic_long_read(&cli->cl_lru_in_list),
 725                         atomic_long_read(&cli->cl_lru_busy));
 726
 727                 list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
 728                 if (osc_cache_too_much(cli) > 0) {
 729                         spin_unlock(&cache->ccc_lru_lock);
 730
 731                         rc = osc_lru_shrink(env, cli, npages, true);
 732                         spin_lock(&cache->ccc_lru_lock);
 733                         if (rc >= npages)
 734                                 break;
 735                         if (rc > 0)
 736                                 npages -= rc;
 737                 }
 738         }
 739         spin_unlock(&cache->ccc_lru_lock);
 740
 741 out:
 742         cl_env_put(env, &refcheck);
 743         CDEBUG(D_CACHE, "%s: cli %p freed %ld pages.\n",
 744                 cli_name(cli), cli, rc);
 745         return rc;
 746 }
 747
 748 /**
 749  * osc_lru_alloc() is called to allocate an LRU slot for a cl_page.
 750  *
 751  * Usually the LRU slots are reserved in osc_io_iter_rw_init().
 752  * Only in the case that the LRU slots are in extreme shortage, it should
 753  * have reserved enough slots for an IO.
 754  */
 755 static int osc_lru_alloc(const struct lu_env *env, struct client_obd *cli,
 756                          struct osc_page *opg)
 757 {
 758         struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
 759         struct osc_io *oio = osc_env_io(env);
 760         int rc = 0;
 761         ENTRY;
 762
 763         if (cli->cl_cache == NULL) /* shall not be in LRU */
 764                 RETURN(0);
 765
 766         if (oio->oi_lru_reserved > 0) {
 767                 --oio->oi_lru_reserved;
 768                 goto out;
 769         }
 770
 771         LASSERT(atomic_long_read(cli->cl_lru_left) >= 0);
 772         while (!atomic_long_add_unless(cli->cl_lru_left, -1, 0)) {
 773                 /* run out of LRU spaces, try to drop some by itself */
 774                 rc = osc_lru_reclaim(cli, 1);
 775                 if (rc < 0)
 776                         break;
 777                 if (rc > 0)
 778                         continue;
 779
 780                 cond_resched();
 781                 rc = l_wait_event(osc_lru_waitq,
 782                                 atomic_long_read(cli->cl_lru_left) > 0,
 783                                 &lwi);
 784                 if (rc < 0)
 785                         break;
 786         }
 787
 788 out:
 789         if (rc >= 0) {
 790                 atomic_long_inc(&cli->cl_lru_busy);
 791                 opg->ops_in_lru = 1;
 792                 rc = 0;
 793         }
 794
 795         RETURN(rc);
 796 }
 797
 798 /**
 799  * osc_lru_reserve() is called to reserve enough LRU slots for I/O.
 800  *
 801  * The benefit of doing this is to reduce contention against atomic counter
 802  * cl_lru_left by changing it from per-page access to per-IO access.
 803  */
 804 unsigned long osc_lru_reserve(struct client_obd *cli, unsigned long npages)
 805 {
 806         unsigned long reserved = 0;
 807         unsigned long max_pages;
 808         unsigned long c;
 809
 810         /* reserve a full RPC window at most to avoid that a thread accidentally
 811          * consumes too many LRU slots */
 812         max_pages = cli->cl_max_pages_per_rpc * cli->cl_max_rpcs_in_flight;
 813         if (npages > max_pages)
 814                 npages = max_pages;
 815
 816         c = atomic_long_read(cli->cl_lru_left);
 817         if (c < npages && osc_lru_reclaim(cli, npages) > 0)
 818                 c = atomic_long_read(cli->cl_lru_left);
 819         while (c >= npages) {
 820                 if (c == atomic_long_cmpxchg(cli->cl_lru_left, c, c - npages)) {
 821                         reserved = npages;
 822                         break;
 823                 }
 824                 c = atomic_long_read(cli->cl_lru_left);
 825         }
 826         if (atomic_long_read(cli->cl_lru_left) < max_pages) {
 827                 /* If there aren't enough pages in the per-OSC LRU then
 828                  * wake up the LRU thread to try and clear out space, so
 829                  * we don't block if pages are being dirtied quickly. */
 830                 CDEBUG(D_CACHE, "%s: queue LRU, left: %lu/%ld.\n",
 831                        cli_name(cli), atomic_long_read(cli->cl_lru_left),
 832                        max_pages);
 833                 (void)ptlrpcd_queue_work(cli->cl_lru_work);
 834         }
 835
 836         return reserved;
 837 }
 838
 839 /**
 840  * osc_lru_unreserve() is called to unreserve LRU slots.
 841  *
 842  * LRU slots reserved by osc_lru_reserve() may have entries left due to several
 843  * reasons such as page already existing or I/O error. Those reserved slots
 844  * should be freed by calling this function.
 845  */
 846 void osc_lru_unreserve(struct client_obd *cli, unsigned long npages)
 847 {
 848         atomic_long_add(npages, cli->cl_lru_left);
 849         wake_up_all(&osc_lru_waitq);
 850 }
 851
 852 /**
 853  * Atomic operations are expensive. We accumulate the accounting for the
 854  * same page zone to get better performance.
 855  * In practice this can work pretty good because the pages in the same RPC
 856  * are likely from the same page zone.
 857  */
 858 static inline void unstable_page_accounting(struct ptlrpc_bulk_desc *desc,
 859                                             int factor)
 860 {
 861         int page_count = desc->bd_iov_count;
 862         void *zone = NULL;
 863         int count = 0;
 864         int i;
 865
 866         LASSERT(ptlrpc_is_bulk_desc_kiov(desc->bd_type));
 867
 868         for (i = 0; i < page_count; i++) {
 869                 void *pz = page_zone(BD_GET_KIOV(desc, i).kiov_page);
 870
 871                 if (likely(pz == zone)) {
 872                         ++count;
 873                         continue;
 874                 }
 875
 876                 if (count > 0) {
 877                         mod_zone_page_state(zone, NR_UNSTABLE_NFS,
 878                                             factor * count);
 879                         count = 0;
 880                 }
 881                 zone = pz;
 882                 ++count;
 883         }
 884         if (count > 0)
 885                 mod_zone_page_state(zone, NR_UNSTABLE_NFS, factor * count);
 886 }
 887
 888 static inline void add_unstable_page_accounting(struct ptlrpc_bulk_desc *desc)
 889 {
 890         unstable_page_accounting(desc, 1);
 891 }
 892
 893 static inline void dec_unstable_page_accounting(struct ptlrpc_bulk_desc *desc)
 894 {
 895         unstable_page_accounting(desc, -1);
 896 }
 897
 898 /**
 899  * Performs "unstable" page accounting. This function balances the
 900  * increment operations performed in osc_inc_unstable_pages. It is
 901  * registered as the RPC request callback, and is executed when the
 902  * bulk RPC is committed on the server. Thus at this point, the pages
 903  * involved in the bulk transfer are no longer considered unstable.
 904  *
 905  * If this function is called, the request should have been committed
 906  * or req:rq_unstable must have been set; it implies that the unstable
 907  * statistic have been added.
 908  */
 909 void osc_dec_unstable_pages(struct ptlrpc_request *req)
 910 {
 911         struct ptlrpc_bulk_desc *desc       = req->rq_bulk;
 912         struct client_obd       *cli        = &req->rq_import->imp_obd->u.cli;
 913         int                      page_count = desc->bd_iov_count;
 914         long                     unstable_count;
 915
 916         LASSERT(page_count >= 0);
 917         dec_unstable_page_accounting(desc);
 918
 919         unstable_count = atomic_long_sub_return(page_count,
 920                                                 &cli->cl_unstable_count);
 921         LASSERT(unstable_count >= 0);
 922
 923         unstable_count = atomic_long_sub_return(page_count,
 924                                            &cli->cl_cache->ccc_unstable_nr);
 925         LASSERT(unstable_count >= 0);
 926         if (unstable_count == 0)
 927                 wake_up_all(&cli->cl_cache->ccc_unstable_waitq);
 928
 929         if (waitqueue_active(&osc_lru_waitq))
 930                 (void)ptlrpcd_queue_work(cli->cl_lru_work);
 931 }
 932
 933 /**
 934  * "unstable" page accounting. See: osc_dec_unstable_pages.
 935  */
 936 void osc_inc_unstable_pages(struct ptlrpc_request *req)
 937 {
 938         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
 939         struct client_obd       *cli  = &req->rq_import->imp_obd->u.cli;
 940         long                     page_count = desc->bd_iov_count;
 941
 942         /* No unstable page tracking */
 943         if (cli->cl_cache == NULL || !cli->cl_cache->ccc_unstable_check)
 944                 return;
 945
 946         add_unstable_page_accounting(desc);
 947         atomic_long_add(page_count, &cli->cl_unstable_count);
 948         atomic_long_add(page_count, &cli->cl_cache->ccc_unstable_nr);
 949
 950         /* If the request has already been committed (i.e. brw_commit
 951          * called via rq_commit_cb), we need to undo the unstable page
 952          * increments we just performed because rq_commit_cb wont be
 953          * called again. */
 954         spin_lock(&req->rq_lock);
 955         if (unlikely(req->rq_committed)) {
 956                 spin_unlock(&req->rq_lock);
 957
 958                 osc_dec_unstable_pages(req);
 959         } else {
 960                 req->rq_unstable = 1;
 961                 spin_unlock(&req->rq_lock);
 962         }
 963 }
 964
 965 /**
 966  * Check if it piggybacks SOFT_SYNC flag to OST from this OSC.
 967  * This function will be called by every BRW RPC so it's critical
 968  * to make this function fast.
 969  */
 970 bool osc_over_unstable_soft_limit(struct client_obd *cli)
 971 {
 972         long unstable_nr, osc_unstable_count;
 973
 974         /* Can't check cli->cl_unstable_count, therefore, no soft limit */
 975         if (cli->cl_cache == NULL || !cli->cl_cache->ccc_unstable_check)
 976                 return false;
 977
 978         osc_unstable_count = atomic_long_read(&cli->cl_unstable_count);
 979         unstable_nr = atomic_long_read(&cli->cl_cache->ccc_unstable_nr);
 980
 981         CDEBUG(D_CACHE,
 982                "%s: cli: %p unstable pages: %lu, osc unstable pages: %lu\n",
 983                cli_name(cli), cli, unstable_nr, osc_unstable_count);
 984
 985         /* If the LRU slots are in shortage - 25% remaining AND this OSC
 986          * has one full RPC window of unstable pages, it's a good chance
 987          * to piggyback a SOFT_SYNC flag.
 988          * Please notice that the OST won't take immediate response for the
 989          * SOFT_SYNC request so active OSCs will have more chance to carry
 990          * the flag, this is reasonable. */
 991         return unstable_nr > cli->cl_cache->ccc_lru_max >> 2 &&
 992                osc_unstable_count > cli->cl_max_pages_per_rpc *
 993                                     cli->cl_max_rpcs_in_flight;
 994 }
 995
 996 /**
 997  * Return how many LRU pages in the cache of all OSC devices
 998  *
 999  * \retval      return # of cached LRU pages times reclaimation tendency
1000  * \retval      SHRINK_STOP if it cannot do any scanning in this time
1001  */
1002 unsigned long osc_cache_shrink_count(struct shrinker *sk,
1003                                      struct shrink_control *sc)
1004 {
1005         struct client_obd *cli;
1006         unsigned long cached = 0;
1007
1008         spin_lock(&osc_shrink_lock);
1009         list_for_each_entry(cli, &osc_shrink_list, cl_shrink_list)
1010                 cached += atomic_long_read(&cli->cl_lru_in_list);
1011         spin_unlock(&osc_shrink_lock);
1012
1013         return (cached  * sysctl_vfs_cache_pressure) / 100;
1014 }
1015
1016 /**
1017  * Scan and try to reclaim sc->nr_to_scan cached LRU pages
1018  *
1019  * \retval      number of cached LRU pages reclaimed
1020  * \retval      SHRINK_STOP if it cannot do any scanning in this time
1021  *
1022  * Linux kernel will loop calling this shrinker scan routine with
1023  * sc->nr_to_scan = SHRINK_BATCH(128 for now) until kernel got enough memory.
1024  *
1025  * If sc->nr_to_scan is 0, the VM is querying the cache size, we don't need
1026  * to scan and try to reclaim LRU pages, just return 0 and
1027  * osc_cache_shrink_count() will report the LRU page number.
1028  */
1029 unsigned long osc_cache_shrink_scan(struct shrinker *sk,
1030                                     struct shrink_control *sc)
1031 {
1032         struct client_obd *cli;
1033         struct client_obd *stop_anchor = NULL;
1034         struct lu_env *env;
1035         long shrank = 0;
1036         int rc;
1037         __u16 refcheck;
1038
1039         if (sc->nr_to_scan == 0)
1040                 return 0;
1041
1042         if (!(sc->gfp_mask & __GFP_FS))
1043                 return SHRINK_STOP;
1044
1045         env = cl_env_get(&refcheck);
1046         if (IS_ERR(env))
1047                 return SHRINK_STOP;
1048
1049         spin_lock(&osc_shrink_lock);
1050         while (!list_empty(&osc_shrink_list)) {
1051                 cli = list_entry(osc_shrink_list.next, struct client_obd,
1052                                  cl_shrink_list);
1053
1054                 if (stop_anchor == NULL)
1055                         stop_anchor = cli;
1056                 else if (cli == stop_anchor)
1057                         break;
1058
1059                 list_move_tail(&cli->cl_shrink_list, &osc_shrink_list);
1060                 spin_unlock(&osc_shrink_lock);
1061
1062                 /* shrink no more than max_pages_per_rpc for an OSC */
1063                 rc = osc_lru_shrink(env, cli, (sc->nr_to_scan - shrank) >
1064                                     cli->cl_max_pages_per_rpc ?
1065                                     cli->cl_max_pages_per_rpc :
1066                                     sc->nr_to_scan - shrank, true);
1067                 if (rc > 0)
1068                         shrank += rc;
1069
1070                 if (shrank >= sc->nr_to_scan)
1071                         goto out;
1072
1073                 spin_lock(&osc_shrink_lock);
1074         }
1075         spin_unlock(&osc_shrink_lock);
1076
1077 out:
1078         cl_env_put(env, &refcheck);
1079
1080         return shrank;
1081 }
1082
1083 /** @} osc */