drivers/staging/lustre/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_reserve(const struct lu_env *env, struct osc_object *obj,
  45                            struct osc_page *opg);
  46
  47 /** \addtogroup osc
  48  *  @{
  49  */
  50
  51 /*****************************************************************************
  52  *
  53  * Page operations.
  54  *
  55  */
  56 static void osc_page_transfer_get(struct osc_page *opg, const char *label)
  57 {
  58         struct cl_page *page = opg->ops_cl.cpl_page;
  59
  60         LASSERT(!opg->ops_transfer_pinned);
  61         cl_page_get(page);
  62         lu_ref_add_atomic(&page->cp_reference, label, page);
  63         opg->ops_transfer_pinned = 1;
  64 }
  65
  66 static void osc_page_transfer_put(const struct lu_env *env,
  67                                   struct osc_page *opg)
  68 {
  69         struct cl_page *page = opg->ops_cl.cpl_page;
  70
  71         if (opg->ops_transfer_pinned) {
  72                 opg->ops_transfer_pinned = 0;
  73                 lu_ref_del(&page->cp_reference, "transfer", page);
  74                 cl_page_put(env, page);
  75         }
  76 }
  77
  78 /**
  79  * This is called once for every page when it is submitted for a transfer
  80  * either opportunistic (osc_page_cache_add()), or immediate
  81  * (osc_page_submit()).
  82  */
  83 static void osc_page_transfer_add(const struct lu_env *env,
  84                                   struct osc_page *opg, enum cl_req_type crt)
  85 {
  86         struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
  87
  88         osc_lru_use(osc_cli(obj), opg);
  89
  90         spin_lock(&obj->oo_seatbelt);
  91         list_add(&opg->ops_inflight, &obj->oo_inflight[crt]);
  92         opg->ops_submitter = current;
  93         spin_unlock(&obj->oo_seatbelt);
  94 }
  95
  96 int osc_page_cache_add(const struct lu_env *env,
  97                        const struct cl_page_slice *slice, struct cl_io *io)
  98 {
  99         struct osc_page *opg = cl2osc_page(slice);
 100         int result;
 101
 102         osc_page_transfer_get(opg, "transfer\0cache");
 103         result = osc_queue_async_io(env, io, opg);
 104         if (result != 0)
 105                 osc_page_transfer_put(env, opg);
 106         else
 107                 osc_page_transfer_add(env, opg, CRT_WRITE);
 108
 109         return result;
 110 }
 111
 112 void osc_index2policy(ldlm_policy_data_t *policy, const struct cl_object *obj,
 113                       pgoff_t start, pgoff_t end)
 114 {
 115         memset(policy, 0, sizeof(*policy));
 116         policy->l_extent.start = cl_offset(obj, start);
 117         policy->l_extent.end = cl_offset(obj, end + 1) - 1;
 118 }
 119
 120 static int osc_page_is_under_lock(const struct lu_env *env,
 121                                   const struct cl_page_slice *slice,
 122                                   struct cl_io *unused, pgoff_t *max_index)
 123 {
 124         struct osc_page *opg = cl2osc_page(slice);
 125         struct ldlm_lock *dlmlock;
 126         int result = -ENODATA;
 127
 128         dlmlock = osc_dlmlock_at_pgoff(env, cl2osc(slice->cpl_obj),
 129                                        osc_index(opg), 1, 0);
 130         if (dlmlock) {
 131                 *max_index = cl_index(slice->cpl_obj,
 132                                       dlmlock->l_policy_data.l_extent.end);
 133                 LDLM_LOCK_PUT(dlmlock);
 134                 result = 0;
 135         }
 136         return result;
 137 }
 138
 139 static const char *osc_list(struct list_head *head)
 140 {
 141         return list_empty(head) ? "-" : "+";
 142 }
 143
 144 static inline unsigned long osc_submit_duration(struct osc_page *opg)
 145 {
 146         if (opg->ops_submit_time == 0)
 147                 return 0;
 148
 149         return (cfs_time_current() - opg->ops_submit_time);
 150 }
 151
 152 static int osc_page_print(const struct lu_env *env,
 153                           const struct cl_page_slice *slice,
 154                           void *cookie, lu_printer_t printer)
 155 {
 156         struct osc_page *opg = cl2osc_page(slice);
 157         struct osc_async_page *oap = &opg->ops_oap;
 158         struct osc_object *obj = cl2osc(slice->cpl_obj);
 159         struct client_obd *cli = &osc_export(obj)->exp_obd->u.cli;
 160
 161         return (*printer)(env, cookie, LUSTRE_OSC_NAME "-page@%p %lu: 1< %#x %d %u %s %s > 2< %llu %u %u %#x %#x | %p %p %p > 3< %s %p %d %lu %d > 4< %d %d %d %lu %s | %s %s %s %s > 5< %s %s %s %s | %d %s | %d %s %s>\n",
 162                           opg, osc_index(opg),
 163                           /* 1 */
 164                           oap->oap_magic, oap->oap_cmd,
 165                           oap->oap_interrupted,
 166                           osc_list(&oap->oap_pending_item),
 167                           osc_list(&oap->oap_rpc_item),
 168                           /* 2 */
 169                           oap->oap_obj_off, oap->oap_page_off, oap->oap_count,
 170                           oap->oap_async_flags, oap->oap_brw_flags,
 171                           oap->oap_request, oap->oap_cli, obj,
 172                           /* 3 */
 173                           osc_list(&opg->ops_inflight),
 174                           opg->ops_submitter, opg->ops_transfer_pinned,
 175                           osc_submit_duration(opg), opg->ops_srvlock,
 176                           /* 4 */
 177                           cli->cl_r_in_flight, cli->cl_w_in_flight,
 178                           cli->cl_max_rpcs_in_flight,
 179                           cli->cl_avail_grant,
 180                           osc_list(&cli->cl_cache_waiters),
 181                           osc_list(&cli->cl_loi_ready_list),
 182                           osc_list(&cli->cl_loi_hp_ready_list),
 183                           osc_list(&cli->cl_loi_write_list),
 184                           osc_list(&cli->cl_loi_read_list),
 185                           /* 5 */
 186                           osc_list(&obj->oo_ready_item),
 187                           osc_list(&obj->oo_hp_ready_item),
 188                           osc_list(&obj->oo_write_item),
 189                           osc_list(&obj->oo_read_item),
 190                           atomic_read(&obj->oo_nr_reads),
 191                           osc_list(&obj->oo_reading_exts),
 192                           atomic_read(&obj->oo_nr_writes),
 193                           osc_list(&obj->oo_hp_exts),
 194                           osc_list(&obj->oo_urgent_exts));
 195 }
 196
 197 static void osc_page_delete(const struct lu_env *env,
 198                             const struct cl_page_slice *slice)
 199 {
 200         struct osc_page *opg = cl2osc_page(slice);
 201         struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
 202         int rc;
 203
 204         CDEBUG(D_TRACE, "%p\n", opg);
 205         osc_page_transfer_put(env, opg);
 206         rc = osc_teardown_async_page(env, obj, opg);
 207         if (rc) {
 208                 CL_PAGE_DEBUG(D_ERROR, env, slice->cpl_page,
 209                               "Trying to teardown failed: %d\n", rc);
 210                 LASSERT(0);
 211         }
 212
 213         spin_lock(&obj->oo_seatbelt);
 214         if (opg->ops_submitter) {
 215                 LASSERT(!list_empty(&opg->ops_inflight));
 216                 list_del_init(&opg->ops_inflight);
 217                 opg->ops_submitter = NULL;
 218         }
 219         spin_unlock(&obj->oo_seatbelt);
 220
 221         osc_lru_del(osc_cli(obj), opg);
 222
 223         if (slice->cpl_page->cp_type == CPT_CACHEABLE) {
 224                 void *value;
 225
 226                 spin_lock(&obj->oo_tree_lock);
 227                 value = radix_tree_delete(&obj->oo_tree, osc_index(opg));
 228                 if (value)
 229                         --obj->oo_npages;
 230                 spin_unlock(&obj->oo_tree_lock);
 231
 232                 LASSERT(ergo(value, value == opg));
 233         }
 234 }
 235
 236 static void osc_page_clip(const struct lu_env *env,
 237                           const struct cl_page_slice *slice, int from, int to)
 238 {
 239         struct osc_page *opg = cl2osc_page(slice);
 240         struct osc_async_page *oap = &opg->ops_oap;
 241
 242         opg->ops_from = from;
 243         opg->ops_to = to;
 244         spin_lock(&oap->oap_lock);
 245         oap->oap_async_flags |= ASYNC_COUNT_STABLE;
 246         spin_unlock(&oap->oap_lock);
 247 }
 248
 249 static int osc_page_cancel(const struct lu_env *env,
 250                            const struct cl_page_slice *slice)
 251 {
 252         struct osc_page *opg = cl2osc_page(slice);
 253         int rc = 0;
 254
 255         /* Check if the transferring against this page
 256          * is completed, or not even queued.
 257          */
 258         if (opg->ops_transfer_pinned)
 259                 /* FIXME: may not be interrupted.. */
 260                 rc = osc_cancel_async_page(env, opg);
 261         LASSERT(ergo(rc == 0, opg->ops_transfer_pinned == 0));
 262         return rc;
 263 }
 264
 265 static int osc_page_flush(const struct lu_env *env,
 266                           const struct cl_page_slice *slice,
 267                           struct cl_io *io)
 268 {
 269         struct osc_page *opg = cl2osc_page(slice);
 270         int rc;
 271
 272         rc = osc_flush_async_page(env, io, opg);
 273         return rc;
 274 }
 275
 276 static const struct cl_page_operations osc_page_ops = {
 277         .cpo_print       = osc_page_print,
 278         .cpo_delete     = osc_page_delete,
 279         .cpo_is_under_lock = osc_page_is_under_lock,
 280         .cpo_clip          = osc_page_clip,
 281         .cpo_cancel      = osc_page_cancel,
 282         .cpo_flush        = osc_page_flush
 283 };
 284
 285 int osc_page_init(const struct lu_env *env, struct cl_object *obj,
 286                   struct cl_page *page, pgoff_t index)
 287 {
 288         struct osc_object *osc = cl2osc(obj);
 289         struct osc_page *opg = cl_object_page_slice(obj, page);
 290         int result;
 291
 292         opg->ops_from = 0;
 293         opg->ops_to = PAGE_SIZE;
 294
 295         result = osc_prep_async_page(osc, opg, page->cp_vmpage,
 296                                      cl_offset(obj, index));
 297         if (result == 0) {
 298                 struct osc_io *oio = osc_env_io(env);
 299
 300                 opg->ops_srvlock = osc_io_srvlock(oio);
 301                 cl_page_slice_add(page, &opg->ops_cl, obj, index,
 302                                   &osc_page_ops);
 303         }
 304         /* ops_inflight and ops_lru are the same field, but it doesn't
 305          * hurt to initialize it twice :-)
 306          */
 307         INIT_LIST_HEAD(&opg->ops_inflight);
 308         INIT_LIST_HEAD(&opg->ops_lru);
 309
 310         /* reserve an LRU space for this page */
 311         if (page->cp_type == CPT_CACHEABLE && result == 0) {
 312                 result = osc_lru_reserve(env, osc, opg);
 313                 if (result == 0) {
 314                         spin_lock(&osc->oo_tree_lock);
 315                         result = radix_tree_insert(&osc->oo_tree, index, opg);
 316                         if (result == 0)
 317                                 ++osc->oo_npages;
 318                         spin_unlock(&osc->oo_tree_lock);
 319                         LASSERT(result == 0);
 320                 }
 321         }
 322
 323         return result;
 324 }
 325
 326 /**
 327  * Helper function called by osc_io_submit() for every page in an immediate
 328  * transfer (i.e., transferred synchronously).
 329  */
 330 void osc_page_submit(const struct lu_env *env, struct osc_page *opg,
 331                      enum cl_req_type crt, int brw_flags)
 332 {
 333         struct osc_async_page *oap = &opg->ops_oap;
 334
 335         LASSERTF(oap->oap_magic == OAP_MAGIC, "Bad oap magic: oap %p, magic 0x%x\n",
 336                  oap, oap->oap_magic);
 337         LASSERT(oap->oap_async_flags & ASYNC_READY);
 338         LASSERT(oap->oap_async_flags & ASYNC_COUNT_STABLE);
 339
 340         oap->oap_cmd = crt == CRT_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ;
 341         oap->oap_page_off = opg->ops_from;
 342         oap->oap_count = opg->ops_to - opg->ops_from;
 343         oap->oap_brw_flags = brw_flags | OBD_BRW_SYNC;
 344
 345         if (capable(CFS_CAP_SYS_RESOURCE)) {
 346                 oap->oap_brw_flags |= OBD_BRW_NOQUOTA;
 347                 oap->oap_cmd |= OBD_BRW_NOQUOTA;
 348         }
 349
 350         opg->ops_submit_time = cfs_time_current();
 351         osc_page_transfer_get(opg, "transfer\0imm");
 352         osc_page_transfer_add(env, opg, crt);
 353 }
 354
 355 /* --------------- LRU page management ------------------ */
 356
 357 /* OSC is a natural place to manage LRU pages as applications are specialized
 358  * to write OSC by OSC. Ideally, if one OSC is used more frequently it should
 359  * occupy more LRU slots. On the other hand, we should avoid using up all LRU
 360  * slots (client_obd::cl_lru_left) otherwise process has to be put into sleep
 361  * for free LRU slots - this will be very bad so the algorithm requires each
 362  * OSC to free slots voluntarily to maintain a reasonable number of free slots
 363  * at any time.
 364  */
 365
 366 static DECLARE_WAIT_QUEUE_HEAD(osc_lru_waitq);
 367 /* LRU pages are freed in batch mode. OSC should at least free this
 368  * number of pages to avoid running out of LRU budget, and..
 369  */
 370 static const int lru_shrink_min = 2 << (20 - PAGE_SHIFT);  /* 2M */
 371 /* free this number at most otherwise it will take too long time to finish. */
 372 static const int lru_shrink_max = 8 << (20 - PAGE_SHIFT); /* 8M */
 373
 374 /* Check if we can free LRU slots from this OSC. If there exists LRU waiters,
 375  * we should free slots aggressively. In this way, slots are freed in a steady
 376  * step to maintain fairness among OSCs.
 377  *
 378  * Return how many LRU pages should be freed.
 379  */
 380 static int osc_cache_too_much(struct client_obd *cli)
 381 {
 382         struct cl_client_cache *cache = cli->cl_cache;
 383         long pages = atomic_long_read(&cli->cl_lru_in_list);
 384         unsigned long budget;
 385
 386         budget = cache->ccc_lru_max / (atomic_read(&cache->ccc_users) - 2);
 387
 388         /* if it's going to run out LRU slots, we should free some, but not
 389          * too much to maintain fairness among OSCs.
 390          */
 391         if (atomic_long_read(cli->cl_lru_left) < cache->ccc_lru_max >> 4) {
 392                 if (pages >= budget)
 393                         return lru_shrink_max;
 394                 else if (pages >= budget / 2)
 395                         return lru_shrink_min;
 396         } else if (pages >= budget * 2) {
 397                 return lru_shrink_min;
 398         }
 399         return 0;
 400 }
 401
 402 int lru_queue_work(const struct lu_env *env, void *data)
 403 {
 404         struct client_obd *cli = data;
 405
 406         CDEBUG(D_CACHE, "Run LRU work for client obd %p.\n", cli);
 407
 408         if (osc_cache_too_much(cli))
 409                 osc_lru_shrink(env, cli, lru_shrink_max, true);
 410
 411         return 0;
 412 }
 413
 414 void osc_lru_add_batch(struct client_obd *cli, struct list_head *plist)
 415 {
 416         LIST_HEAD(lru);
 417         struct osc_async_page *oap;
 418         long npages = 0;
 419
 420         list_for_each_entry(oap, plist, oap_pending_item) {
 421                 struct osc_page *opg = oap2osc_page(oap);
 422
 423                 if (!opg->ops_in_lru)
 424                         continue;
 425
 426                 ++npages;
 427                 LASSERT(list_empty(&opg->ops_lru));
 428                 list_add(&opg->ops_lru, &lru);
 429         }
 430
 431         if (npages > 0) {
 432                 spin_lock(&cli->cl_lru_list_lock);
 433                 list_splice_tail(&lru, &cli->cl_lru_list);
 434                 atomic_long_sub(npages, &cli->cl_lru_busy);
 435                 atomic_long_add(npages, &cli->cl_lru_in_list);
 436                 spin_unlock(&cli->cl_lru_list_lock);
 437
 438                 /* XXX: May set force to be true for better performance */
 439                 if (osc_cache_too_much(cli))
 440                         (void)ptlrpcd_queue_work(cli->cl_lru_work);
 441         }
 442 }
 443
 444 static void __osc_lru_del(struct client_obd *cli, struct osc_page *opg)
 445 {
 446         LASSERT(atomic_long_read(&cli->cl_lru_in_list) > 0);
 447         list_del_init(&opg->ops_lru);
 448         atomic_long_dec(&cli->cl_lru_in_list);
 449 }
 450
 451 /**
 452  * Page is being destroyed. The page may be not in LRU list, if the transfer
 453  * has never finished(error occurred).
 454  */
 455 static void osc_lru_del(struct client_obd *cli, struct osc_page *opg)
 456 {
 457         if (opg->ops_in_lru) {
 458                 spin_lock(&cli->cl_lru_list_lock);
 459                 if (!list_empty(&opg->ops_lru)) {
 460                         __osc_lru_del(cli, opg);
 461                 } else {
 462                         LASSERT(atomic_long_read(&cli->cl_lru_busy) > 0);
 463                         atomic_long_dec(&cli->cl_lru_busy);
 464                 }
 465                 spin_unlock(&cli->cl_lru_list_lock);
 466
 467                 atomic_long_inc(cli->cl_lru_left);
 468                 /* this is a great place to release more LRU pages if
 469                  * this osc occupies too many LRU pages and kernel is
 470                  * stealing one of them.
 471                  */
 472                 if (!memory_pressure_get())
 473                         (void)ptlrpcd_queue_work(cli->cl_lru_work);
 474                 wake_up(&osc_lru_waitq);
 475         } else {
 476                 LASSERT(list_empty(&opg->ops_lru));
 477         }
 478 }
 479
 480 /**
 481  * Delete page from LRUlist for redirty.
 482  */
 483 static void osc_lru_use(struct client_obd *cli, struct osc_page *opg)
 484 {
 485         /* If page is being transferred for the first time,
 486          * ops_lru should be empty
 487          */
 488         if (opg->ops_in_lru && !list_empty(&opg->ops_lru)) {
 489                 spin_lock(&cli->cl_lru_list_lock);
 490                 __osc_lru_del(cli, opg);
 491                 spin_unlock(&cli->cl_lru_list_lock);
 492                 atomic_long_inc(&cli->cl_lru_busy);
 493         }
 494 }
 495
 496 static void discard_pagevec(const struct lu_env *env, struct cl_io *io,
 497                             struct cl_page **pvec, int max_index)
 498 {
 499         int i;
 500
 501         for (i = 0; i < max_index; i++) {
 502                 struct cl_page *page = pvec[i];
 503
 504                 LASSERT(cl_page_is_owned(page, io));
 505                 cl_page_discard(env, io, page);
 506                 cl_page_disown(env, io, page);
 507                 cl_page_put(env, page);
 508
 509                 pvec[i] = NULL;
 510         }
 511 }
 512
 513 /**
 514  * Check if a cl_page can be released, i.e, it's not being used.
 515  *
 516  * If unstable account is turned on, bulk transfer may hold one refcount
 517  * for recovery so we need to check vmpage refcount as well; otherwise,
 518  * even we can destroy cl_page but the corresponding vmpage can't be reused.
 519  */
 520 static inline bool lru_page_busy(struct client_obd *cli, struct cl_page *page)
 521 {
 522         if (cl_page_in_use_noref(page))
 523                 return true;
 524
 525         if (cli->cl_cache->ccc_unstable_check) {
 526                 struct page *vmpage = cl_page_vmpage(page);
 527
 528                 /* vmpage have two known users: cl_page and VM page cache */
 529                 if (page_count(vmpage) - page_mapcount(vmpage) > 2)
 530                         return true;
 531         }
 532         return false;
 533 }
 534
 535 /**
 536  * Drop @target of pages from LRU at most.
 537  */
 538 long osc_lru_shrink(const struct lu_env *env, struct client_obd *cli,
 539                     long target, bool force)
 540 {
 541         struct cl_io *io;
 542         struct cl_object *clobj = NULL;
 543         struct cl_page **pvec;
 544         struct osc_page *opg;
 545         struct osc_page *temp;
 546         int maxscan = 0;
 547         long count = 0;
 548         int index = 0;
 549         int rc = 0;
 550
 551         LASSERT(atomic_long_read(&cli->cl_lru_in_list) >= 0);
 552         if (atomic_long_read(&cli->cl_lru_in_list) == 0 || target <= 0)
 553                 return 0;
 554
 555         if (!force) {
 556                 if (atomic_read(&cli->cl_lru_shrinkers) > 0)
 557                         return -EBUSY;
 558
 559                 if (atomic_inc_return(&cli->cl_lru_shrinkers) > 1) {
 560                         atomic_dec(&cli->cl_lru_shrinkers);
 561                         return -EBUSY;
 562                 }
 563         } else {
 564                 atomic_inc(&cli->cl_lru_shrinkers);
 565         }
 566
 567         pvec = (struct cl_page **)osc_env_info(env)->oti_pvec;
 568         io = &osc_env_info(env)->oti_io;
 569
 570         spin_lock(&cli->cl_lru_list_lock);
 571         maxscan = min(target << 1, atomic_long_read(&cli->cl_lru_in_list));
 572         list_for_each_entry_safe(opg, temp, &cli->cl_lru_list, ops_lru) {
 573                 struct cl_page *page;
 574                 bool will_free = false;
 575
 576                 if (--maxscan < 0)
 577                         break;
 578
 579                 page = opg->ops_cl.cpl_page;
 580                 if (lru_page_busy(cli, page)) {
 581                         list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
 582                         continue;
 583                 }
 584
 585                 LASSERT(page->cp_obj);
 586                 if (clobj != page->cp_obj) {
 587                         struct cl_object *tmp = page->cp_obj;
 588
 589                         cl_object_get(tmp);
 590                         spin_unlock(&cli->cl_lru_list_lock);
 591
 592                         if (clobj) {
 593                                 discard_pagevec(env, io, pvec, index);
 594                                 index = 0;
 595
 596                                 cl_io_fini(env, io);
 597                                 cl_object_put(env, clobj);
 598                                 clobj = NULL;
 599                         }
 600
 601                         clobj = tmp;
 602                         io->ci_obj = clobj;
 603                         io->ci_ignore_layout = 1;
 604                         rc = cl_io_init(env, io, CIT_MISC, clobj);
 605
 606                         spin_lock(&cli->cl_lru_list_lock);
 607
 608                         if (rc != 0)
 609                                 break;
 610
 611                         ++maxscan;
 612                         continue;
 613                 }
 614
 615                 if (cl_page_own_try(env, io, page) == 0) {
 616                         if (!lru_page_busy(cli, page)) {
 617                                 /* remove it from lru list earlier to avoid
 618                                  * lock contention
 619                                  */
 620                                 __osc_lru_del(cli, opg);
 621                                 opg->ops_in_lru = 0; /* will be discarded */
 622
 623                                 cl_page_get(page);
 624                                 will_free = true;
 625                         } else {
 626                                 cl_page_disown(env, io, page);
 627                         }
 628                 }
 629
 630                 if (!will_free) {
 631                         list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
 632                         continue;
 633                 }
 634
 635                 /* Don't discard and free the page with cl_lru_list held */
 636                 pvec[index++] = page;
 637                 if (unlikely(index == OTI_PVEC_SIZE)) {
 638                         spin_unlock(&cli->cl_lru_list_lock);
 639                         discard_pagevec(env, io, pvec, index);
 640                         index = 0;
 641
 642                         spin_lock(&cli->cl_lru_list_lock);
 643                 }
 644
 645                 if (++count >= target)
 646                         break;
 647         }
 648         spin_unlock(&cli->cl_lru_list_lock);
 649
 650         if (clobj) {
 651                 discard_pagevec(env, io, pvec, index);
 652
 653                 cl_io_fini(env, io);
 654                 cl_object_put(env, clobj);
 655         }
 656
 657         atomic_dec(&cli->cl_lru_shrinkers);
 658         if (count > 0) {
 659                 atomic_long_add(count, cli->cl_lru_left);
 660                 wake_up_all(&osc_lru_waitq);
 661         }
 662         return count > 0 ? count : rc;
 663 }
 664
 665 long osc_lru_reclaim(struct client_obd *cli)
 666 {
 667         struct cl_env_nest nest;
 668         struct lu_env *env;
 669         struct cl_client_cache *cache = cli->cl_cache;
 670         int max_scans;
 671         long rc = 0;
 672
 673         LASSERT(cache);
 674
 675         env = cl_env_nested_get(&nest);
 676         if (IS_ERR(env))
 677                 return 0;
 678
 679         rc = osc_lru_shrink(env, cli, osc_cache_too_much(cli), false);
 680         if (rc != 0) {
 681                 if (rc == -EBUSY)
 682                         rc = 0;
 683
 684                 CDEBUG(D_CACHE, "%s: Free %ld pages from own LRU: %p.\n",
 685                        cli->cl_import->imp_obd->obd_name, rc, cli);
 686                 goto out;
 687         }
 688
 689         CDEBUG(D_CACHE, "%s: cli %p no free slots, pages: %ld, busy: %ld.\n",
 690                cli->cl_import->imp_obd->obd_name, cli,
 691                atomic_long_read(&cli->cl_lru_in_list),
 692                atomic_long_read(&cli->cl_lru_busy));
 693
 694         /* Reclaim LRU slots from other client_obd as it can't free enough
 695          * from its own. This should rarely happen.
 696          */
 697         spin_lock(&cache->ccc_lru_lock);
 698         LASSERT(!list_empty(&cache->ccc_lru));
 699
 700         cache->ccc_lru_shrinkers++;
 701         list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
 702
 703         max_scans = atomic_read(&cache->ccc_users) - 2;
 704         while (--max_scans > 0 && !list_empty(&cache->ccc_lru)) {
 705                 cli = list_entry(cache->ccc_lru.next, struct client_obd,
 706                                  cl_lru_osc);
 707
 708                 CDEBUG(D_CACHE, "%s: cli %p LRU pages: %ld, busy: %ld.\n",
 709                        cli->cl_import->imp_obd->obd_name, cli,
 710                        atomic_long_read(&cli->cl_lru_in_list),
 711                        atomic_long_read(&cli->cl_lru_busy));
 712
 713                 list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
 714                 if (osc_cache_too_much(cli) > 0) {
 715                         spin_unlock(&cache->ccc_lru_lock);
 716
 717                         rc = osc_lru_shrink(env, cli, osc_cache_too_much(cli),
 718                                             true);
 719                         spin_lock(&cache->ccc_lru_lock);
 720                         if (rc != 0)
 721                                 break;
 722                 }
 723         }
 724         spin_unlock(&cache->ccc_lru_lock);
 725
 726 out:
 727         cl_env_nested_put(&nest, env);
 728         CDEBUG(D_CACHE, "%s: cli %p freed %ld pages.\n",
 729                cli->cl_import->imp_obd->obd_name, cli, rc);
 730         return rc;
 731 }
 732
 733 /**
 734  * osc_lru_reserve() is called to reserve an LRU slot for a cl_page.
 735  *
 736  * Usually the LRU slots are reserved in osc_io_iter_rw_init().
 737  * Only in the case that the LRU slots are in extreme shortage, it should
 738  * have reserved enough slots for an IO.
 739  */
 740 static int osc_lru_reserve(const struct lu_env *env, struct osc_object *obj,
 741                            struct osc_page *opg)
 742 {
 743         struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
 744         struct osc_io *oio = osc_env_io(env);
 745         struct client_obd *cli = osc_cli(obj);
 746         int rc = 0;
 747
 748         if (!cli->cl_cache) /* shall not be in LRU */
 749                 return 0;
 750
 751         if (oio->oi_lru_reserved > 0) {
 752                 --oio->oi_lru_reserved;
 753                 goto out;
 754         }
 755
 756         LASSERT(atomic_long_read(cli->cl_lru_left) >= 0);
 757         while (!atomic_long_add_unless(cli->cl_lru_left, -1, 0)) {
 758                 /* run out of LRU spaces, try to drop some by itself */
 759                 rc = osc_lru_reclaim(cli);
 760                 if (rc < 0)
 761                         break;
 762                 if (rc > 0)
 763                         continue;
 764
 765                 cond_resched();
 766
 767                 rc = l_wait_event(osc_lru_waitq,
 768                                   atomic_long_read(cli->cl_lru_left) > 0,
 769                                   &lwi);
 770
 771                 if (rc < 0)
 772                         break;
 773         }
 774
 775 out:
 776         if (rc >= 0) {
 777                 atomic_long_inc(&cli->cl_lru_busy);
 778                 opg->ops_in_lru = 1;
 779                 rc = 0;
 780         }
 781
 782         return rc;
 783 }
 784
 785 /**
 786  * Atomic operations are expensive. We accumulate the accounting for the
 787  * same page pgdat to get better performance.
 788  * In practice this can work pretty good because the pages in the same RPC
 789  * are likely from the same page zone.
 790  */
 791 static inline void unstable_page_accounting(struct ptlrpc_bulk_desc *desc,
 792                                             int factor)
 793 {
 794         int page_count = desc->bd_iov_count;
 795         pg_data_t *last = NULL;
 796         int count = 0;
 797         int i;
 798
 799         for (i = 0; i < page_count; i++) {
 800                 pg_data_t *pgdat = page_pgdat(desc->bd_iov[i].bv_page);
 801
 802                 if (likely(pgdat == last)) {
 803                         ++count;
 804                         continue;
 805                 }
 806
 807                 if (count > 0) {
 808                         mod_node_page_state(pgdat, NR_UNSTABLE_NFS,
 809                                             factor * count);
 810                         count = 0;
 811                 }
 812                 last = pgdat;
 813                 ++count;
 814         }
 815         if (count > 0)
 816                 mod_node_page_state(last, NR_UNSTABLE_NFS, factor * count);
 817 }
 818
 819 static inline void add_unstable_page_accounting(struct ptlrpc_bulk_desc *desc)
 820 {
 821         unstable_page_accounting(desc, 1);
 822 }
 823
 824 static inline void dec_unstable_page_accounting(struct ptlrpc_bulk_desc *desc)
 825 {
 826         unstable_page_accounting(desc, -1);
 827 }
 828
 829 /**
 830  * Performs "unstable" page accounting. This function balances the
 831  * increment operations performed in osc_inc_unstable_pages. It is
 832  * registered as the RPC request callback, and is executed when the
 833  * bulk RPC is committed on the server. Thus at this point, the pages
 834  * involved in the bulk transfer are no longer considered unstable.
 835  *
 836  * If this function is called, the request should have been committed
 837  * or req:rq_unstable must have been set; it implies that the unstable
 838  * statistic have been added.
 839  */
 840 void osc_dec_unstable_pages(struct ptlrpc_request *req)
 841 {
 842         struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
 843         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
 844         int page_count = desc->bd_iov_count;
 845         long unstable_count;
 846
 847         LASSERT(page_count >= 0);
 848         dec_unstable_page_accounting(desc);
 849
 850         unstable_count = atomic_long_sub_return(page_count,
 851                                                 &cli->cl_unstable_count);
 852         LASSERT(unstable_count >= 0);
 853
 854         unstable_count = atomic_long_sub_return(page_count,
 855                                                 &cli->cl_cache->ccc_unstable_nr);
 856         LASSERT(unstable_count >= 0);
 857         if (!unstable_count)
 858                 wake_up_all(&cli->cl_cache->ccc_unstable_waitq);
 859
 860         if (osc_cache_too_much(cli))
 861                 (void)ptlrpcd_queue_work(cli->cl_lru_work);
 862 }
 863
 864 /**
 865  * "unstable" page accounting. See: osc_dec_unstable_pages.
 866  */
 867 void osc_inc_unstable_pages(struct ptlrpc_request *req)
 868 {
 869         struct client_obd *cli  = &req->rq_import->imp_obd->u.cli;
 870         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
 871         long page_count = desc->bd_iov_count;
 872
 873         /* No unstable page tracking */
 874         if (!cli->cl_cache || !cli->cl_cache->ccc_unstable_check)
 875                 return;
 876
 877         add_unstable_page_accounting(desc);
 878         atomic_long_add(page_count, &cli->cl_unstable_count);
 879         atomic_long_add(page_count, &cli->cl_cache->ccc_unstable_nr);
 880
 881         /*
 882          * If the request has already been committed (i.e. brw_commit
 883          * called via rq_commit_cb), we need to undo the unstable page
 884          * increments we just performed because rq_commit_cb wont be
 885          * called again.
 886          */
 887         spin_lock(&req->rq_lock);
 888         if (unlikely(req->rq_committed)) {
 889                 spin_unlock(&req->rq_lock);
 890
 891                 osc_dec_unstable_pages(req);
 892         } else {
 893                 req->rq_unstable = 1;
 894                 spin_unlock(&req->rq_lock);
 895         }
 896 }
 897
 898 /**
 899  * Check if it piggybacks SOFT_SYNC flag to OST from this OSC.
 900  * This function will be called by every BRW RPC so it's critical
 901  * to make this function fast.
 902  */
 903 bool osc_over_unstable_soft_limit(struct client_obd *cli)
 904 {
 905         long unstable_nr, osc_unstable_count;
 906
 907         /* Can't check cli->cl_unstable_count, therefore, no soft limit */
 908         if (!cli->cl_cache || !cli->cl_cache->ccc_unstable_check)
 909                 return false;
 910
 911         osc_unstable_count = atomic_long_read(&cli->cl_unstable_count);
 912         unstable_nr = atomic_long_read(&cli->cl_cache->ccc_unstable_nr);
 913
 914         CDEBUG(D_CACHE,
 915                "%s: cli: %p unstable pages: %lu, osc unstable pages: %lu\n",
 916                cli->cl_import->imp_obd->obd_name, cli,
 917                unstable_nr, osc_unstable_count);
 918
 919         /*
 920          * If the LRU slots are in shortage - 25% remaining AND this OSC
 921          * has one full RPC window of unstable pages, it's a good chance
 922          * to piggyback a SOFT_SYNC flag.
 923          * Please notice that the OST won't take immediate response for the
 924          * SOFT_SYNC request so active OSCs will have more chance to carry
 925          * the flag, this is reasonable.
 926          */
 927         return unstable_nr > cli->cl_cache->ccc_lru_max >> 2 &&
 928                osc_unstable_count > cli->cl_max_pages_per_rpc *
 929                                     cli->cl_max_rpcs_in_flight;
 930 }
 931
 932 /** @} osc */