lustre/osc/osc_page.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2015, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * Implementation of cl_page for OSC layer.
  37  *
  38  *   Author: Nikita Danilov <nikita.danilov@sun.com>
  39  *   Author: Jinshan Xiong <jinshan.xiong@intel.com>
  40  */
  41
  42 #define DEBUG_SUBSYSTEM S_OSC
  43
  44 #include "osc_cl_internal.h"
  45
  46 static void osc_lru_del(struct client_obd *cli, struct osc_page *opg);
  47 static void osc_lru_use(struct client_obd *cli, struct osc_page *opg);
  48 static int osc_lru_alloc(const struct lu_env *env, struct client_obd *cli,
  49                          struct osc_page *opg);
  50
  51 /** \addtogroup osc
  52  *  @{
  53  */
  54
  55 /*
  56  * Page operations.
  57  */
  58 static void osc_page_transfer_get(struct osc_page *opg, const char *label)
  59 {
  60         struct cl_page *page = opg->ops_cl.cpl_page;
  61
  62         LASSERT(!opg->ops_transfer_pinned);
  63         cl_page_get(page);
  64         lu_ref_add_atomic(&page->cp_reference, label, page);
  65         opg->ops_transfer_pinned = 1;
  66 }
  67
  68 static void osc_page_transfer_put(const struct lu_env *env,
  69                                   struct osc_page *opg)
  70 {
  71         struct cl_page *page = opg->ops_cl.cpl_page;
  72
  73         if (opg->ops_transfer_pinned) {
  74                 opg->ops_transfer_pinned = 0;
  75                 lu_ref_del(&page->cp_reference, "transfer", page);
  76                 cl_page_put(env, page);
  77         }
  78 }
  79
  80 /**
  81  * This is called once for every page when it is submitted for a transfer
  82  * either opportunistic (osc_page_cache_add()), or immediate
  83  * (osc_page_submit()).
  84  */
  85 static void osc_page_transfer_add(const struct lu_env *env,
  86                                   struct osc_page *opg, enum cl_req_type crt)
  87 {
  88         struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
  89
  90         /* ops_lru and ops_inflight share the same field, so take it from LRU
  91          * first and then use it as inflight. */
  92         osc_lru_use(osc_cli(obj), opg);
  93 }
  94
  95 int osc_page_cache_add(const struct lu_env *env,
  96                         const struct cl_page_slice *slice, struct cl_io *io)
  97 {
  98         struct osc_page *opg = cl2osc_page(slice);
  99         int result;
 100         ENTRY;
 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(union ldlm_policy_data *policy,
 113                       const struct cl_object *obj, 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 const char *osc_list(struct list_head *head)
 121 {
 122         return list_empty(head) ? "-" : "+";
 123 }
 124
 125 static inline cfs_time_t osc_submit_duration(struct osc_page *opg)
 126 {
 127         if (opg->ops_submit_time == 0)
 128                 return 0;
 129
 130         return (cfs_time_current() - opg->ops_submit_time);
 131 }
 132
 133 static int osc_page_print(const struct lu_env *env,
 134                           const struct cl_page_slice *slice,
 135                           void *cookie, lu_printer_t printer)
 136 {
 137         struct osc_page       *opg = cl2osc_page(slice);
 138         struct osc_async_page *oap = &opg->ops_oap;
 139         struct osc_object     *obj = cl2osc(slice->cpl_obj);
 140         struct client_obd     *cli = &osc_export(obj)->exp_obd->u.cli;
 141
 142         return (*printer)(env, cookie, LUSTRE_OSC_NAME"-page@%p %lu: "
 143                           "1< %#x %d %u %s %s > "
 144                           "2< "LPD64" %u %u %#x %#x | %p %p %p > "
 145                           "3< %d %lu %d > "
 146                           "4< %d %d %d %lu %s | %s %s %s %s > "
 147                           "5< %s %s %s %s | %d %s | %d %s %s>\n",
 148                           opg, osc_index(opg),
 149                           /* 1 */
 150                           oap->oap_magic, oap->oap_cmd,
 151                           oap->oap_interrupted,
 152                           osc_list(&oap->oap_pending_item),
 153                           osc_list(&oap->oap_rpc_item),
 154                           /* 2 */
 155                           oap->oap_obj_off, oap->oap_page_off, oap->oap_count,
 156                           oap->oap_async_flags, oap->oap_brw_flags,
 157                           oap->oap_request, oap->oap_cli, obj,
 158                           /* 3 */
 159                           opg->ops_transfer_pinned,
 160                           osc_submit_duration(opg), opg->ops_srvlock,
 161                           /* 4 */
 162                           cli->cl_r_in_flight, cli->cl_w_in_flight,
 163                           cli->cl_max_rpcs_in_flight,
 164                           cli->cl_avail_grant,
 165                           osc_list(&cli->cl_cache_waiters),
 166                           osc_list(&cli->cl_loi_ready_list),
 167                           osc_list(&cli->cl_loi_hp_ready_list),
 168                           osc_list(&cli->cl_loi_write_list),
 169                           osc_list(&cli->cl_loi_read_list),
 170                           /* 5 */
 171                           osc_list(&obj->oo_ready_item),
 172                           osc_list(&obj->oo_hp_ready_item),
 173                           osc_list(&obj->oo_write_item),
 174                           osc_list(&obj->oo_read_item),
 175                           atomic_read(&obj->oo_nr_reads),
 176                           osc_list(&obj->oo_reading_exts),
 177                           atomic_read(&obj->oo_nr_writes),
 178                           osc_list(&obj->oo_hp_exts),
 179                           osc_list(&obj->oo_urgent_exts));
 180 }
 181
 182 static void osc_page_delete(const struct lu_env *env,
 183                             const struct cl_page_slice *slice)
 184 {
 185         struct osc_page   *opg = cl2osc_page(slice);
 186         struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
 187         int rc;
 188
 189         ENTRY;
 190         CDEBUG(D_TRACE, "%p\n", opg);
 191         osc_page_transfer_put(env, opg);
 192         rc = osc_teardown_async_page(env, obj, opg);
 193         if (rc) {
 194                 CL_PAGE_DEBUG(D_ERROR, env, slice->cpl_page,
 195                               "Trying to teardown failed: %d\n", rc);
 196                 LASSERT(0);
 197         }
 198
 199         osc_lru_del(osc_cli(obj), opg);
 200
 201         if (slice->cpl_page->cp_type == CPT_CACHEABLE) {
 202                 void *value;
 203
 204                 spin_lock(&obj->oo_tree_lock);
 205                 value = radix_tree_delete(&obj->oo_tree, osc_index(opg));
 206                 if (value != NULL)
 207                         --obj->oo_npages;
 208                 spin_unlock(&obj->oo_tree_lock);
 209
 210                 LASSERT(ergo(value != NULL, value == opg));
 211         }
 212
 213         EXIT;
 214 }
 215
 216 static void osc_page_clip(const struct lu_env *env,
 217                           const struct cl_page_slice *slice,
 218                           int from, int to)
 219 {
 220         struct osc_page       *opg = cl2osc_page(slice);
 221         struct osc_async_page *oap = &opg->ops_oap;
 222
 223         opg->ops_from = from;
 224         opg->ops_to   = to;
 225         spin_lock(&oap->oap_lock);
 226         oap->oap_async_flags |= ASYNC_COUNT_STABLE;
 227         spin_unlock(&oap->oap_lock);
 228 }
 229
 230 static int osc_page_cancel(const struct lu_env *env,
 231                            const struct cl_page_slice *slice)
 232 {
 233         struct osc_page *opg = cl2osc_page(slice);
 234         int rc = 0;
 235
 236         /* Check if the transferring against this page
 237          * is completed, or not even queued. */
 238         if (opg->ops_transfer_pinned)
 239                 /* FIXME: may not be interrupted.. */
 240                 rc = osc_cancel_async_page(env, opg);
 241         LASSERT(ergo(rc == 0, opg->ops_transfer_pinned == 0));
 242         return rc;
 243 }
 244
 245 static int osc_page_flush(const struct lu_env *env,
 246                           const struct cl_page_slice *slice,
 247                           struct cl_io *io)
 248 {
 249         struct osc_page *opg = cl2osc_page(slice);
 250         int rc = 0;
 251         ENTRY;
 252         rc = osc_flush_async_page(env, io, opg);
 253         RETURN(rc);
 254 }
 255
 256 static const struct cl_page_operations osc_page_ops = {
 257         .cpo_print         = osc_page_print,
 258         .cpo_delete        = osc_page_delete,
 259         .cpo_clip           = osc_page_clip,
 260         .cpo_cancel         = osc_page_cancel,
 261         .cpo_flush          = osc_page_flush
 262 };
 263
 264 int osc_page_init(const struct lu_env *env, struct cl_object *obj,
 265                   struct cl_page *page, pgoff_t index)
 266 {
 267         struct osc_object *osc = cl2osc(obj);
 268         struct osc_page   *opg = cl_object_page_slice(obj, page);
 269         int result;
 270
 271         opg->ops_from = 0;
 272         opg->ops_to   = PAGE_CACHE_SIZE;
 273
 274         result = osc_prep_async_page(osc, opg, page->cp_vmpage,
 275                                      cl_offset(obj, index));
 276         if (result == 0) {
 277                 struct osc_io *oio = osc_env_io(env);
 278                 opg->ops_srvlock = osc_io_srvlock(oio);
 279                 cl_page_slice_add(page, &opg->ops_cl, obj, index,
 280                                   &osc_page_ops);
 281         }
 282         INIT_LIST_HEAD(&opg->ops_lru);
 283
 284         /* reserve an LRU space for this page */
 285         if (page->cp_type == CPT_CACHEABLE && result == 0) {
 286                 result = osc_lru_alloc(env, osc_cli(osc), opg);
 287                 if (result == 0) {
 288                         spin_lock(&osc->oo_tree_lock);
 289                         result = radix_tree_insert(&osc->oo_tree, index, opg);
 290                         if (result == 0)
 291                                 ++osc->oo_npages;
 292                         spin_unlock(&osc->oo_tree_lock);
 293                         LASSERT(result == 0);
 294                 }
 295         }
 296
 297         return result;
 298 }
 299
 300 /**
 301  * Helper function called by osc_io_submit() for every page in an immediate
 302  * transfer (i.e., transferred synchronously).
 303  */
 304 void osc_page_submit(const struct lu_env *env, struct osc_page *opg,
 305                      enum cl_req_type crt, int brw_flags)
 306 {
 307         struct osc_async_page *oap = &opg->ops_oap;
 308
 309         LASSERTF(oap->oap_magic == OAP_MAGIC, "Bad oap magic: oap %p, "
 310                  "magic 0x%x\n", oap, oap->oap_magic);
 311         LASSERT(oap->oap_async_flags & ASYNC_READY);
 312         LASSERT(oap->oap_async_flags & ASYNC_COUNT_STABLE);
 313
 314         oap->oap_cmd       = crt == CRT_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ;
 315         oap->oap_page_off  = opg->ops_from;
 316         oap->oap_count     = opg->ops_to - opg->ops_from;
 317         oap->oap_brw_flags = OBD_BRW_SYNC | brw_flags;
 318
 319         if (cfs_capable(CFS_CAP_SYS_RESOURCE)) {
 320                 oap->oap_brw_flags |= OBD_BRW_NOQUOTA;
 321                 oap->oap_cmd |= OBD_BRW_NOQUOTA;
 322         }
 323
 324         opg->ops_submit_time = cfs_time_current();
 325         osc_page_transfer_get(opg, "transfer\0imm");
 326         osc_page_transfer_add(env, opg, crt);
 327 }
 328
 329 /* --------------- LRU page management ------------------ */
 330
 331 /* OSC is a natural place to manage LRU pages as applications are specialized
 332  * to write OSC by OSC. Ideally, if one OSC is used more frequently it should
 333  * occupy more LRU slots. On the other hand, we should avoid using up all LRU
 334  * slots (client_obd::cl_lru_left) otherwise process has to be put into sleep
 335  * for free LRU slots - this will be very bad so the algorithm requires each
 336  * OSC to free slots voluntarily to maintain a reasonable number of free slots
 337  * at any time.
 338  */
 339
 340 static DECLARE_WAIT_QUEUE_HEAD(osc_lru_waitq);
 341
 342 /**
 343  * LRU pages are freed in batch mode. OSC should at least free this
 344  * number of pages to avoid running out of LRU slots.
 345  */
 346 static inline int lru_shrink_min(struct client_obd *cli)
 347 {
 348         return cli->cl_max_pages_per_rpc * 2;
 349 }
 350
 351 /**
 352  * free this number at most otherwise it will take too long time to finsih.
 353  */
 354 static inline int lru_shrink_max(struct client_obd *cli)
 355 {
 356         return cli->cl_max_pages_per_rpc * cli->cl_max_rpcs_in_flight;
 357 }
 358
 359 /**
 360  * Check if we can free LRU slots from this OSC. If there exists LRU waiters,
 361  * we should free slots aggressively. In this way, slots are freed in a steady
 362  * step to maintain fairness among OSCs.
 363  *
 364  * Return how many LRU pages should be freed.
 365  */
 366 static int osc_cache_too_much(struct client_obd *cli)
 367 {
 368         struct cl_client_cache *cache = cli->cl_cache;
 369         long pages = atomic_long_read(&cli->cl_lru_in_list);
 370         unsigned long budget;
 371
 372         LASSERT(cache != NULL);
 373         budget = cache->ccc_lru_max / (atomic_read(&cache->ccc_users) - 2);
 374
 375         /* if it's going to run out LRU slots, we should free some, but not
 376          * too much to maintain faireness among OSCs. */
 377         if (atomic_long_read(cli->cl_lru_left) < cache->ccc_lru_max >> 2) {
 378                 if (pages >= budget)
 379                         return lru_shrink_max(cli);
 380                 else if (pages >= budget / 2)
 381                         return lru_shrink_min(cli);
 382         } else {
 383                 int duration = cfs_time_current_sec() - cli->cl_lru_last_used;
 384
 385                 /* knock out pages by duration of no IO activity */
 386                 duration >>= 6; /* approximately 1 minute */
 387                 if (duration > 0 && pages >= budget / duration)
 388                         return lru_shrink_min(cli);
 389         }
 390         return 0;
 391 }
 392
 393 int lru_queue_work(const struct lu_env *env, void *data)
 394 {
 395         struct client_obd *cli = data;
 396         int count;
 397
 398         CDEBUG(D_CACHE, "%s: run LRU work for client obd\n", cli_name(cli));
 399         count = osc_cache_too_much(cli);
 400         if (count > 0) {
 401                 int rc = osc_lru_shrink(env, cli, count, false);
 402
 403                 CDEBUG(D_CACHE, "%s: shrank %d/%d pages from client obd\n",
 404                        cli_name(cli), rc, count);
 405                 if (rc >= count) {
 406                         CDEBUG(D_CACHE, "%s: queue again\n", cli_name(cli));
 407                         ptlrpcd_queue_work(cli->cl_lru_work);
 408                 }
 409         }
 410
 411         RETURN(0);
 412 }
 413
 414 void osc_lru_add_batch(struct client_obd *cli, struct list_head *plist)
 415 {
 416         struct list_head lru = LIST_HEAD_INIT(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                 cli->cl_lru_last_used = cfs_time_current_sec();
 437                 spin_unlock(&cli->cl_lru_list_lock);
 438
 439                 if (waitqueue_active(&osc_lru_waitq))
 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                 if (osc_cache_too_much(cli)) {
 472                         CDEBUG(D_CACHE, "%s: queue LRU work\n", cli_name(cli));
 473                         (void)ptlrpcd_queue_work(cli->cl_lru_work);
 474                 }
 475                 wake_up(&osc_lru_waitq);
 476         } else {
 477                 LASSERT(list_empty(&opg->ops_lru));
 478         }
 479 }
 480
 481 /**
 482  * Delete page from LRUlist for redirty.
 483  */
 484 static void osc_lru_use(struct client_obd *cli, struct osc_page *opg)
 485 {
 486         /* If page is being transferred for the first time,
 487          * ops_lru should be empty */
 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_delete(env, page);
 506                 cl_page_discard(env, io, page);
 507                 cl_page_disown(env, io, page);
 508                 cl_page_put(env, page);
 509
 510                 pvec[i] = NULL;
 511         }
 512 }
 513
 514 /**
 515  * Check if a cl_page can be released, i.e, it's not being used.
 516  *
 517  * If unstable account is turned on, bulk transfer may hold one refcount
 518  * for recovery so we need to check vmpage refcount as well; otherwise,
 519  * even we can destroy cl_page but the corresponding vmpage can't be reused.
 520  */
 521 static inline bool lru_page_busy(struct client_obd *cli, struct cl_page *page)
 522 {
 523         if (cl_page_in_use_noref(page))
 524                 return true;
 525
 526         if (cli->cl_cache->ccc_unstable_check) {
 527                 struct page *vmpage = cl_page_vmpage(page);
 528
 529                 /* vmpage have two known users: cl_page and VM page cache */
 530                 if (page_count(vmpage) - page_mapcount(vmpage) > 2)
 531                         return true;
 532         }
 533         return false;
 534 }
 535
 536 /**
 537  * Drop @target of pages from LRU at most.
 538  */
 539 long osc_lru_shrink(const struct lu_env *env, struct client_obd *cli,
 540                    long target, bool force)
 541 {
 542         struct cl_io *io;
 543         struct cl_object *clobj = NULL;
 544         struct cl_page **pvec;
 545         struct osc_page *opg;
 546         long count = 0;
 547         int maxscan = 0;
 548         int index = 0;
 549         int rc = 0;
 550         ENTRY;
 551
 552         LASSERT(atomic_long_read(&cli->cl_lru_in_list) >= 0);
 553         if (atomic_long_read(&cli->cl_lru_in_list) == 0 || target <= 0)
 554                 RETURN(0);
 555
 556         CDEBUG(D_CACHE, "%s: shrinkers: %d, force: %d\n",
 557                cli_name(cli), atomic_read(&cli->cl_lru_shrinkers), force);
 558         if (!force) {
 559                 if (atomic_read(&cli->cl_lru_shrinkers) > 0)
 560                         RETURN(-EBUSY);
 561
 562                 if (atomic_inc_return(&cli->cl_lru_shrinkers) > 1) {
 563                         atomic_dec(&cli->cl_lru_shrinkers);
 564                         RETURN(-EBUSY);
 565                 }
 566         } else {
 567                 atomic_inc(&cli->cl_lru_shrinkers);
 568         }
 569
 570         pvec = (struct cl_page **)osc_env_info(env)->oti_pvec;
 571         io = &osc_env_info(env)->oti_io;
 572
 573         spin_lock(&cli->cl_lru_list_lock);
 574         if (force)
 575                 cli->cl_lru_reclaim++;
 576         maxscan = min(target << 1, atomic_long_read(&cli->cl_lru_in_list));
 577         while (!list_empty(&cli->cl_lru_list)) {
 578                 struct cl_page *page;
 579                 bool will_free = false;
 580
 581                 if (!force && atomic_read(&cli->cl_lru_shrinkers) > 1)
 582                         break;
 583
 584                 if (--maxscan < 0)
 585                         break;
 586
 587                 opg = list_entry(cli->cl_lru_list.next, struct osc_page,
 588                                  ops_lru);
 589                 page = opg->ops_cl.cpl_page;
 590                 if (lru_page_busy(cli, page)) {
 591                         list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
 592                         continue;
 593                 }
 594
 595                 LASSERT(page->cp_obj != NULL);
 596                 if (clobj != page->cp_obj) {
 597                         struct cl_object *tmp = page->cp_obj;
 598
 599                         cl_object_get(tmp);
 600                         spin_unlock(&cli->cl_lru_list_lock);
 601
 602                         if (clobj != NULL) {
 603                                 discard_pagevec(env, io, pvec, index);
 604                                 index = 0;
 605
 606                                 cl_io_fini(env, io);
 607                                 cl_object_put(env, clobj);
 608                                 clobj = NULL;
 609                         }
 610
 611                         clobj = tmp;
 612                         io->ci_obj = clobj;
 613                         io->ci_ignore_layout = 1;
 614                         rc = cl_io_init(env, io, CIT_MISC, clobj);
 615
 616                         spin_lock(&cli->cl_lru_list_lock);
 617
 618                         if (rc != 0)
 619                                 break;
 620
 621                         ++maxscan;
 622                         continue;
 623                 }
 624
 625                 if (cl_page_own_try(env, io, page) == 0) {
 626                         if (!lru_page_busy(cli, page)) {
 627                                 /* remove it from lru list earlier to avoid
 628                                  * lock contention */
 629                                 __osc_lru_del(cli, opg);
 630                                 opg->ops_in_lru = 0; /* will be discarded */
 631
 632                                 cl_page_get(page);
 633                                 will_free = true;
 634                         } else {
 635                                 cl_page_disown(env, io, page);
 636                         }
 637                 }
 638
 639                 if (!will_free) {
 640                         list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
 641                         continue;
 642                 }
 643
 644                 /* Don't discard and free the page with cl_lru_list held */
 645                 pvec[index++] = page;
 646                 if (unlikely(index == OTI_PVEC_SIZE)) {
 647                         spin_unlock(&cli->cl_lru_list_lock);
 648                         discard_pagevec(env, io, pvec, index);
 649                         index = 0;
 650
 651                         spin_lock(&cli->cl_lru_list_lock);
 652                 }
 653
 654                 if (++count >= target)
 655                         break;
 656         }
 657         spin_unlock(&cli->cl_lru_list_lock);
 658
 659         if (clobj != NULL) {
 660                 discard_pagevec(env, io, pvec, index);
 661
 662                 cl_io_fini(env, io);
 663                 cl_object_put(env, clobj);
 664         }
 665
 666         atomic_dec(&cli->cl_lru_shrinkers);
 667         if (count > 0) {
 668                 atomic_long_add(count, cli->cl_lru_left);
 669                 wake_up_all(&osc_lru_waitq);
 670         }
 671         RETURN(count > 0 ? count : rc);
 672 }
 673
 674 /**
 675  * Reclaim LRU pages by an IO thread. The caller wants to reclaim at least
 676  * \@npages of LRU slots. For performance consideration, it's better to drop
 677  * LRU pages in batch. Therefore, the actual number is adjusted at least
 678  * max_pages_per_rpc.
 679  */
 680 static long osc_lru_reclaim(struct client_obd *cli, unsigned long npages)
 681 {
 682         struct lu_env *env;
 683         struct cl_client_cache *cache = cli->cl_cache;
 684         int max_scans;
 685         __u16 refcheck;
 686         long rc = 0;
 687         ENTRY;
 688
 689         LASSERT(cache != NULL);
 690
 691         env = cl_env_get(&refcheck);
 692         if (IS_ERR(env))
 693                 RETURN(rc);
 694
 695         npages = max_t(int, npages, cli->cl_max_pages_per_rpc);
 696         CDEBUG(D_CACHE, "%s: start to reclaim %ld pages from LRU\n",
 697                cli_name(cli), npages);
 698         rc = osc_lru_shrink(env, cli, npages, true);
 699         if (rc >= npages) {
 700                 CDEBUG(D_CACHE, "%s: reclaimed %ld/%ld pages from LRU\n",
 701                        cli_name(cli), rc, npages);
 702                 if (osc_cache_too_much(cli) > 0)
 703                         ptlrpcd_queue_work(cli->cl_lru_work);
 704                 GOTO(out, rc);
 705         } else if (rc > 0) {
 706                 npages -= rc;
 707         }
 708
 709         CDEBUG(D_CACHE, "%s: cli %p no free slots, pages: %ld/%ld, want: %ld\n",
 710                 cli_name(cli), cli, atomic_long_read(&cli->cl_lru_in_list),
 711                 atomic_long_read(&cli->cl_lru_busy), npages);
 712
 713         /* Reclaim LRU slots from other client_obd as it can't free enough
 714          * from its own. This should rarely happen. */
 715         spin_lock(&cache->ccc_lru_lock);
 716         LASSERT(!list_empty(&cache->ccc_lru));
 717
 718         cache->ccc_lru_shrinkers++;
 719         list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
 720
 721         max_scans = atomic_read(&cache->ccc_users) - 2;
 722         while (--max_scans > 0 && !list_empty(&cache->ccc_lru)) {
 723                 cli = list_entry(cache->ccc_lru.next, struct client_obd,
 724                                  cl_lru_osc);
 725
 726                 CDEBUG(D_CACHE, "%s: cli %p LRU pages: %ld, busy: %ld.\n",
 727                         cli_name(cli), cli,
 728                         atomic_long_read(&cli->cl_lru_in_list),
 729                         atomic_long_read(&cli->cl_lru_busy));
 730
 731                 list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
 732                 if (osc_cache_too_much(cli) > 0) {
 733                         spin_unlock(&cache->ccc_lru_lock);
 734
 735                         rc = osc_lru_shrink(env, cli, npages, true);
 736                         spin_lock(&cache->ccc_lru_lock);
 737                         if (rc >= npages)
 738                                 break;
 739                         if (rc > 0)
 740                                 npages -= rc;
 741                 }
 742         }
 743         spin_unlock(&cache->ccc_lru_lock);
 744
 745 out:
 746         cl_env_put(env, &refcheck);
 747         CDEBUG(D_CACHE, "%s: cli %p freed %ld pages.\n",
 748                 cli_name(cli), cli, rc);
 749         return rc;
 750 }
 751
 752 /**
 753  * osc_lru_alloc() is called to allocate an LRU slot for a cl_page.
 754  *
 755  * Usually the LRU slots are reserved in osc_io_iter_rw_init().
 756  * Only in the case that the LRU slots are in extreme shortage, it should
 757  * have reserved enough slots for an IO.
 758  */
 759 static int osc_lru_alloc(const struct lu_env *env, struct client_obd *cli,
 760                          struct osc_page *opg)
 761 {
 762         struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
 763         struct osc_io *oio = osc_env_io(env);
 764         int rc = 0;
 765         ENTRY;
 766
 767         if (cli->cl_cache == NULL) /* shall not be in LRU */
 768                 RETURN(0);
 769
 770         if (oio->oi_lru_reserved > 0) {
 771                 --oio->oi_lru_reserved;
 772                 goto out;
 773         }
 774
 775         LASSERT(atomic_long_read(cli->cl_lru_left) >= 0);
 776         while (!atomic_long_add_unless(cli->cl_lru_left, -1, 0)) {
 777                 /* run out of LRU spaces, try to drop some by itself */
 778                 rc = osc_lru_reclaim(cli, 1);
 779                 if (rc < 0)
 780                         break;
 781                 if (rc > 0)
 782                         continue;
 783
 784                 cond_resched();
 785                 rc = l_wait_event(osc_lru_waitq,
 786                                 atomic_long_read(cli->cl_lru_left) > 0,
 787                                 &lwi);
 788                 if (rc < 0)
 789                         break;
 790         }
 791
 792 out:
 793         if (rc >= 0) {
 794                 atomic_long_inc(&cli->cl_lru_busy);
 795                 opg->ops_in_lru = 1;
 796                 rc = 0;
 797         }
 798
 799         RETURN(rc);
 800 }
 801
 802 /**
 803  * osc_lru_reserve() is called to reserve enough LRU slots for I/O.
 804  *
 805  * The benefit of doing this is to reduce contention against atomic counter
 806  * cl_lru_left by changing it from per-page access to per-IO access.
 807  */
 808 unsigned long osc_lru_reserve(struct client_obd *cli, unsigned long npages)
 809 {
 810         unsigned long reserved = 0;
 811         unsigned long max_pages;
 812         unsigned long c;
 813
 814         /* reserve a full RPC window at most to avoid that a thread accidentally
 815          * consumes too many LRU slots */
 816         max_pages = cli->cl_max_pages_per_rpc * cli->cl_max_rpcs_in_flight;
 817         if (npages > max_pages)
 818                 npages = max_pages;
 819
 820         c = atomic_long_read(cli->cl_lru_left);
 821         if (c < npages && osc_lru_reclaim(cli, npages) > 0)
 822                 c = atomic_long_read(cli->cl_lru_left);
 823         while (c >= npages) {
 824                 if (c == atomic_long_cmpxchg(cli->cl_lru_left, c, c - npages)) {
 825                         reserved = npages;
 826                         break;
 827                 }
 828                 c = atomic_long_read(cli->cl_lru_left);
 829         }
 830         if (atomic_long_read(cli->cl_lru_left) < max_pages) {
 831                 /* If there aren't enough pages in the per-OSC LRU then
 832                  * wake up the LRU thread to try and clear out space, so
 833                  * we don't block if pages are being dirtied quickly. */
 834                 CDEBUG(D_CACHE, "%s: queue LRU, left: %lu/%ld.\n",
 835                        cli_name(cli), atomic_long_read(cli->cl_lru_left),
 836                        max_pages);
 837                 (void)ptlrpcd_queue_work(cli->cl_lru_work);
 838         }
 839
 840         return reserved;
 841 }
 842
 843 /**
 844  * osc_lru_unreserve() is called to unreserve LRU slots.
 845  *
 846  * LRU slots reserved by osc_lru_reserve() may have entries left due to several
 847  * reasons such as page already existing or I/O error. Those reserved slots
 848  * should be freed by calling this function.
 849  */
 850 void osc_lru_unreserve(struct client_obd *cli, unsigned long npages)
 851 {
 852         atomic_long_add(npages, cli->cl_lru_left);
 853         wake_up_all(&osc_lru_waitq);
 854 }
 855
 856 /**
 857  * Atomic operations are expensive. We accumulate the accounting for the
 858  * same page zone to get better performance.
 859  * In practice this can work pretty good because the pages in the same RPC
 860  * are likely from the same page zone.
 861  */
 862 static inline void unstable_page_accounting(struct ptlrpc_bulk_desc *desc,
 863                                             int factor)
 864 {
 865         int page_count = desc->bd_iov_count;
 866         void *zone = NULL;
 867         int count = 0;
 868         int i;
 869
 870         LASSERT(ptlrpc_is_bulk_desc_kiov(desc->bd_type));
 871
 872         for (i = 0; i < page_count; i++) {
 873                 void *pz = page_zone(BD_GET_KIOV(desc, i).kiov_page);
 874
 875                 if (likely(pz == zone)) {
 876                         ++count;
 877                         continue;
 878                 }
 879
 880                 if (count > 0) {
 881                         mod_zone_page_state(zone, NR_UNSTABLE_NFS,
 882                                             factor * count);
 883                         count = 0;
 884                 }
 885                 zone = pz;
 886                 ++count;
 887         }
 888         if (count > 0)
 889                 mod_zone_page_state(zone, NR_UNSTABLE_NFS, factor * count);
 890 }
 891
 892 static inline void add_unstable_page_accounting(struct ptlrpc_bulk_desc *desc)
 893 {
 894         unstable_page_accounting(desc, 1);
 895 }
 896
 897 static inline void dec_unstable_page_accounting(struct ptlrpc_bulk_desc *desc)
 898 {
 899         unstable_page_accounting(desc, -1);
 900 }
 901
 902 /**
 903  * Performs "unstable" page accounting. This function balances the
 904  * increment operations performed in osc_inc_unstable_pages. It is
 905  * registered as the RPC request callback, and is executed when the
 906  * bulk RPC is committed on the server. Thus at this point, the pages
 907  * involved in the bulk transfer are no longer considered unstable.
 908  *
 909  * If this function is called, the request should have been committed
 910  * or req:rq_unstable must have been set; it implies that the unstable
 911  * statistic have been added.
 912  */
 913 void osc_dec_unstable_pages(struct ptlrpc_request *req)
 914 {
 915         struct ptlrpc_bulk_desc *desc       = req->rq_bulk;
 916         struct client_obd       *cli        = &req->rq_import->imp_obd->u.cli;
 917         int                      page_count = desc->bd_iov_count;
 918         long                     unstable_count;
 919
 920         LASSERT(page_count >= 0);
 921         dec_unstable_page_accounting(desc);
 922
 923         unstable_count = atomic_long_sub_return(page_count,
 924                                                 &cli->cl_unstable_count);
 925         LASSERT(unstable_count >= 0);
 926
 927         unstable_count = atomic_long_sub_return(page_count,
 928                                            &cli->cl_cache->ccc_unstable_nr);
 929         LASSERT(unstable_count >= 0);
 930         if (unstable_count == 0)
 931                 wake_up_all(&cli->cl_cache->ccc_unstable_waitq);
 932
 933         if (waitqueue_active(&osc_lru_waitq))
 934                 (void)ptlrpcd_queue_work(cli->cl_lru_work);
 935 }
 936
 937 /**
 938  * "unstable" page accounting. See: osc_dec_unstable_pages.
 939  */
 940 void osc_inc_unstable_pages(struct ptlrpc_request *req)
 941 {
 942         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
 943         struct client_obd       *cli  = &req->rq_import->imp_obd->u.cli;
 944         long                     page_count = desc->bd_iov_count;
 945
 946         /* No unstable page tracking */
 947         if (cli->cl_cache == NULL || !cli->cl_cache->ccc_unstable_check)
 948                 return;
 949
 950         add_unstable_page_accounting(desc);
 951         atomic_long_add(page_count, &cli->cl_unstable_count);
 952         atomic_long_add(page_count, &cli->cl_cache->ccc_unstable_nr);
 953
 954         /* If the request has already been committed (i.e. brw_commit
 955          * called via rq_commit_cb), we need to undo the unstable page
 956          * increments we just performed because rq_commit_cb wont be
 957          * called again. */
 958         spin_lock(&req->rq_lock);
 959         if (unlikely(req->rq_committed)) {
 960                 spin_unlock(&req->rq_lock);
 961
 962                 osc_dec_unstable_pages(req);
 963         } else {
 964                 req->rq_unstable = 1;
 965                 spin_unlock(&req->rq_lock);
 966         }
 967 }
 968
 969 /**
 970  * Check if it piggybacks SOFT_SYNC flag to OST from this OSC.
 971  * This function will be called by every BRW RPC so it's critical
 972  * to make this function fast.
 973  */
 974 bool osc_over_unstable_soft_limit(struct client_obd *cli)
 975 {
 976         long unstable_nr, osc_unstable_count;
 977
 978         /* Can't check cli->cl_unstable_count, therefore, no soft limit */
 979         if (cli->cl_cache == NULL || !cli->cl_cache->ccc_unstable_check)
 980                 return false;
 981
 982         osc_unstable_count = atomic_long_read(&cli->cl_unstable_count);
 983         unstable_nr = atomic_long_read(&cli->cl_cache->ccc_unstable_nr);
 984
 985         CDEBUG(D_CACHE,
 986                "%s: cli: %p unstable pages: %lu, osc unstable pages: %lu\n",
 987                cli_name(cli), cli, unstable_nr, osc_unstable_count);
 988
 989         /* If the LRU slots are in shortage - 25% remaining AND this OSC
 990          * has one full RPC window of unstable pages, it's a good chance
 991          * to piggyback a SOFT_SYNC flag.
 992          * Please notice that the OST won't take immediate response for the
 993          * SOFT_SYNC request so active OSCs will have more chance to carry
 994          * the flag, this is reasonable. */
 995         return unstable_nr > cli->cl_cache->ccc_lru_max >> 2 &&
 996                osc_unstable_count > cli->cl_max_pages_per_rpc *
 997                                     cli->cl_max_rpcs_in_flight;
 998 }
 999
1000 /**
1001  * Return how many LRU pages in the cache of all OSC devices
1002  *
1003  * \retval      return # of cached LRU pages times reclaimation tendency
1004  * \retval      SHRINK_STOP if it cannot do any scanning in this time
1005  */
1006 unsigned long osc_cache_shrink_count(struct shrinker *sk,
1007                                      struct shrink_control *sc)
1008 {
1009         struct client_obd *cli;
1010         unsigned long cached = 0;
1011
1012         spin_lock(&osc_shrink_lock);
1013         list_for_each_entry(cli, &osc_shrink_list, cl_shrink_list)
1014                 cached += atomic_long_read(&cli->cl_lru_in_list);
1015         spin_unlock(&osc_shrink_lock);
1016
1017         return (cached  * sysctl_vfs_cache_pressure) / 100;
1018 }
1019
1020 /**
1021  * Scan and try to reclaim sc->nr_to_scan cached LRU pages
1022  *
1023  * \retval      number of cached LRU pages reclaimed
1024  * \retval      SHRINK_STOP if it cannot do any scanning in this time
1025  *
1026  * Linux kernel will loop calling this shrinker scan routine with
1027  * sc->nr_to_scan = SHRINK_BATCH(128 for now) until kernel got enough memory.
1028  *
1029  * If sc->nr_to_scan is 0, the VM is querying the cache size, we don't need
1030  * to scan and try to reclaim LRU pages, just return 0 and
1031  * osc_cache_shrink_count() will report the LRU page number.
1032  */
1033 unsigned long osc_cache_shrink_scan(struct shrinker *sk,
1034                                     struct shrink_control *sc)
1035 {
1036         struct client_obd *cli;
1037         struct client_obd *stop_anchor = NULL;
1038         struct lu_env *env;
1039         long shrank = 0;
1040         int rc;
1041         __u16 refcheck;
1042
1043         if (sc->nr_to_scan == 0)
1044                 return 0;
1045
1046         if (!(sc->gfp_mask & __GFP_FS))
1047                 return SHRINK_STOP;
1048
1049         env = cl_env_get(&refcheck);
1050         if (IS_ERR(env))
1051                 return SHRINK_STOP;
1052
1053         spin_lock(&osc_shrink_lock);
1054         while (!list_empty(&osc_shrink_list)) {
1055                 cli = list_entry(osc_shrink_list.next, struct client_obd,
1056                                  cl_shrink_list);
1057
1058                 if (stop_anchor == NULL)
1059                         stop_anchor = cli;
1060                 else if (cli == stop_anchor)
1061                         break;
1062
1063                 list_move_tail(&cli->cl_shrink_list, &osc_shrink_list);
1064                 spin_unlock(&osc_shrink_lock);
1065
1066                 /* shrink no more than max_pages_per_rpc for an OSC */
1067                 rc = osc_lru_shrink(env, cli, (sc->nr_to_scan - shrank) >
1068                                     cli->cl_max_pages_per_rpc ?
1069                                     cli->cl_max_pages_per_rpc :
1070                                     sc->nr_to_scan - shrank, true);
1071                 if (rc > 0)
1072                         shrank += rc;
1073
1074                 if (shrank >= sc->nr_to_scan)
1075                         goto out;
1076
1077                 spin_lock(&osc_shrink_lock);
1078         }
1079         spin_unlock(&osc_shrink_lock);
1080
1081 out:
1082         cl_env_put(env, &refcheck);
1083
1084         return shrank;
1085 }
1086
1087 /** @} osc */