lustre/osc/osc_page.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 2016, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * Implementation of cl_page for OSC layer.
  33  *
  34  *   Author: Nikita Danilov <nikita.danilov@sun.com>
  35  *   Author: Jinshan Xiong <jinshan.xiong@intel.com>
  36  */
  37
  38 #define DEBUG_SUBSYSTEM S_OSC
  39
  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         struct osc_io *oio = osc_env_io(env);
 266         int result;
 267
 268         opg->ops_from = 0;
 269         opg->ops_to   = PAGE_SIZE;
 270
 271         INIT_LIST_HEAD(&opg->ops_lru);
 272
 273         result = osc_prep_async_page(osc, opg, page->cp_vmpage,
 274                                      cl_offset(obj, index));
 275         if (result != 0)
 276                 return result;
 277
 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
 283         /* reserve an LRU space for this page */
 284         if (page->cp_type == CPT_CACHEABLE) {
 285                 result = osc_lru_alloc(env, osc_cli(osc), opg);
 286                 if (result == 0) {
 287                         result = radix_tree_preload(GFP_NOFS);
 288                         if (result == 0) {
 289                                 spin_lock(&osc->oo_tree_lock);
 290                                 result = radix_tree_insert(&osc->oo_tree,
 291                                                            index, opg);
 292                                 if (result == 0)
 293                                         ++osc->oo_npages;
 294                                 spin_unlock(&osc->oo_tree_lock);
 295
 296                                 radix_tree_preload_end();
 297                         }
 298                 }
 299         }
 300
 301         return result;
 302 }
 303
 304 /**
 305  * Helper function called by osc_io_submit() for every page in an immediate
 306  * transfer (i.e., transferred synchronously).
 307  */
 308 void osc_page_submit(const struct lu_env *env, struct osc_page *opg,
 309                      enum cl_req_type crt, int brw_flags)
 310 {
 311         struct osc_async_page *oap = &opg->ops_oap;
 312
 313         LASSERTF(oap->oap_magic == OAP_MAGIC, "Bad oap magic: oap %p, "
 314                  "magic 0x%x\n", oap, oap->oap_magic);
 315         LASSERT(oap->oap_async_flags & ASYNC_READY);
 316         LASSERT(oap->oap_async_flags & ASYNC_COUNT_STABLE);
 317
 318         oap->oap_cmd       = crt == CRT_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ;
 319         oap->oap_page_off  = opg->ops_from;
 320         oap->oap_count     = opg->ops_to - opg->ops_from;
 321         oap->oap_brw_flags = OBD_BRW_SYNC | brw_flags;
 322
 323         if (cfs_capable(CFS_CAP_SYS_RESOURCE)) {
 324                 oap->oap_brw_flags |= OBD_BRW_NOQUOTA;
 325                 oap->oap_cmd |= OBD_BRW_NOQUOTA;
 326         }
 327
 328         opg->ops_submit_time = cfs_time_current();
 329         osc_page_transfer_get(opg, "transfer\0imm");
 330         osc_page_transfer_add(env, opg, crt);
 331 }
 332
 333 /* --------------- LRU page management ------------------ */
 334
 335 /* OSC is a natural place to manage LRU pages as applications are specialized
 336  * to write OSC by OSC. Ideally, if one OSC is used more frequently it should
 337  * occupy more LRU slots. On the other hand, we should avoid using up all LRU
 338  * slots (client_obd::cl_lru_left) otherwise process has to be put into sleep
 339  * for free LRU slots - this will be very bad so the algorithm requires each
 340  * OSC to free slots voluntarily to maintain a reasonable number of free slots
 341  * at any time.
 342  */
 343
 344 static DECLARE_WAIT_QUEUE_HEAD(osc_lru_waitq);
 345
 346 /**
 347  * LRU pages are freed in batch mode. OSC should at least free this
 348  * number of pages to avoid running out of LRU slots.
 349  */
 350 static inline int lru_shrink_min(struct client_obd *cli)
 351 {
 352         return cli->cl_max_pages_per_rpc * 2;
 353 }
 354
 355 /**
 356  * free this number at most otherwise it will take too long time to finsih.
 357  */
 358 static inline int lru_shrink_max(struct client_obd *cli)
 359 {
 360         return cli->cl_max_pages_per_rpc * cli->cl_max_rpcs_in_flight;
 361 }
 362
 363 /**
 364  * Check if we can free LRU slots from this OSC. If there exists LRU waiters,
 365  * we should free slots aggressively. In this way, slots are freed in a steady
 366  * step to maintain fairness among OSCs.
 367  *
 368  * Return how many LRU pages should be freed.
 369  */
 370 static int osc_cache_too_much(struct client_obd *cli)
 371 {
 372         struct cl_client_cache *cache = cli->cl_cache;
 373         long pages = atomic_long_read(&cli->cl_lru_in_list);
 374         unsigned long budget;
 375
 376         LASSERT(cache != NULL);
 377         budget = cache->ccc_lru_max / (atomic_read(&cache->ccc_users) - 2);
 378
 379         /* if it's going to run out LRU slots, we should free some, but not
 380          * too much to maintain faireness among OSCs. */
 381         if (atomic_long_read(cli->cl_lru_left) < cache->ccc_lru_max >> 2) {
 382                 if (pages >= budget)
 383                         return lru_shrink_max(cli);
 384                 else if (pages >= budget / 2)
 385                         return lru_shrink_min(cli);
 386         } else {
 387                 time64_t duration = ktime_get_real_seconds();
 388                 long timediff;
 389
 390                 /* knock out pages by duration of no IO activity */
 391                 duration -= cli->cl_lru_last_used;
 392                 /*
 393                  * The difference shouldn't be more than 70 years
 394                  * so we can safely case to a long. Round to
 395                  * approximately 1 minute.
 396                  */
 397                 timediff = (long)(duration >> 6);
 398                 if (timediff > 0 && pages >= budget / timediff)
 399                         return lru_shrink_min(cli);
 400         }
 401         return 0;
 402 }
 403
 404 int lru_queue_work(const struct lu_env *env, void *data)
 405 {
 406         struct client_obd *cli = data;
 407         int count;
 408
 409         CDEBUG(D_CACHE, "%s: run LRU work for client obd\n", cli_name(cli));
 410         count = osc_cache_too_much(cli);
 411         if (count > 0) {
 412                 int rc = osc_lru_shrink(env, cli, count, false);
 413
 414                 CDEBUG(D_CACHE, "%s: shrank %d/%d pages from client obd\n",
 415                        cli_name(cli), rc, count);
 416                 if (rc >= count) {
 417                         CDEBUG(D_CACHE, "%s: queue again\n", cli_name(cli));
 418                         ptlrpcd_queue_work(cli->cl_lru_work);
 419                 }
 420         }
 421
 422         RETURN(0);
 423 }
 424
 425 void osc_lru_add_batch(struct client_obd *cli, struct list_head *plist)
 426 {
 427         struct list_head lru = LIST_HEAD_INIT(lru);
 428         struct osc_async_page *oap;
 429         long npages = 0;
 430
 431         list_for_each_entry(oap, plist, oap_pending_item) {
 432                 struct osc_page *opg = oap2osc_page(oap);
 433
 434                 if (!opg->ops_in_lru)
 435                         continue;
 436
 437                 ++npages;
 438                 LASSERT(list_empty(&opg->ops_lru));
 439                 list_add(&opg->ops_lru, &lru);
 440         }
 441
 442         if (npages > 0) {
 443                 spin_lock(&cli->cl_lru_list_lock);
 444                 list_splice_tail(&lru, &cli->cl_lru_list);
 445                 atomic_long_sub(npages, &cli->cl_lru_busy);
 446                 atomic_long_add(npages, &cli->cl_lru_in_list);
 447                 cli->cl_lru_last_used = ktime_get_real_seconds();
 448                 spin_unlock(&cli->cl_lru_list_lock);
 449
 450                 if (waitqueue_active(&osc_lru_waitq))
 451                         (void)ptlrpcd_queue_work(cli->cl_lru_work);
 452         }
 453 }
 454
 455 static void __osc_lru_del(struct client_obd *cli, struct osc_page *opg)
 456 {
 457         LASSERT(atomic_long_read(&cli->cl_lru_in_list) > 0);
 458         list_del_init(&opg->ops_lru);
 459         atomic_long_dec(&cli->cl_lru_in_list);
 460 }
 461
 462 /**
 463  * Page is being destroyed. The page may be not in LRU list, if the transfer
 464  * has never finished(error occurred).
 465  */
 466 static void osc_lru_del(struct client_obd *cli, struct osc_page *opg)
 467 {
 468         if (opg->ops_in_lru) {
 469                 spin_lock(&cli->cl_lru_list_lock);
 470                 if (!list_empty(&opg->ops_lru)) {
 471                         __osc_lru_del(cli, opg);
 472                 } else {
 473                         LASSERT(atomic_long_read(&cli->cl_lru_busy) > 0);
 474                         atomic_long_dec(&cli->cl_lru_busy);
 475                 }
 476                 spin_unlock(&cli->cl_lru_list_lock);
 477
 478                 atomic_long_inc(cli->cl_lru_left);
 479                 /* this is a great place to release more LRU pages if
 480                  * this osc occupies too many LRU pages and kernel is
 481                  * stealing one of them. */
 482                 if (osc_cache_too_much(cli)) {
 483                         CDEBUG(D_CACHE, "%s: queue LRU work\n", cli_name(cli));
 484                         (void)ptlrpcd_queue_work(cli->cl_lru_work);
 485                 }
 486                 wake_up(&osc_lru_waitq);
 487         } else {
 488                 LASSERT(list_empty(&opg->ops_lru));
 489         }
 490 }
 491
 492 /**
 493  * Delete page from LRU list for redirty.
 494  */
 495 static void osc_lru_use(struct client_obd *cli, struct osc_page *opg)
 496 {
 497         /* If page is being transferred for the first time,
 498          * ops_lru should be empty */
 499         if (opg->ops_in_lru) {
 500                 spin_lock(&cli->cl_lru_list_lock);
 501                 if (!list_empty(&opg->ops_lru)) {
 502                         __osc_lru_del(cli, opg);
 503                         atomic_long_inc(&cli->cl_lru_busy);
 504                 }
 505                 spin_unlock(&cli->cl_lru_list_lock);
 506         }
 507 }
 508
 509 static void discard_pagevec(const struct lu_env *env, struct cl_io *io,
 510                                 struct cl_page **pvec, int max_index)
 511 {
 512         int i;
 513
 514         for (i = 0; i < max_index; i++) {
 515                 struct cl_page *page = pvec[i];
 516
 517                 LASSERT(cl_page_is_owned(page, io));
 518                 cl_page_delete(env, page);
 519                 cl_page_discard(env, io, page);
 520                 cl_page_disown(env, io, page);
 521                 cl_page_put(env, page);
 522
 523                 pvec[i] = NULL;
 524         }
 525 }
 526
 527 /**
 528  * Check if a cl_page can be released, i.e, it's not being used.
 529  *
 530  * If unstable account is turned on, bulk transfer may hold one refcount
 531  * for recovery so we need to check vmpage refcount as well; otherwise,
 532  * even we can destroy cl_page but the corresponding vmpage can't be reused.
 533  */
 534 static inline bool lru_page_busy(struct client_obd *cli, struct cl_page *page)
 535 {
 536         if (cl_page_in_use_noref(page))
 537                 return true;
 538
 539         if (cli->cl_cache->ccc_unstable_check) {
 540                 struct page *vmpage = cl_page_vmpage(page);
 541
 542                 /* vmpage have two known users: cl_page and VM page cache */
 543                 if (page_count(vmpage) - page_mapcount(vmpage) > 2)
 544                         return true;
 545         }
 546         return false;
 547 }
 548
 549 /**
 550  * Drop @target of pages from LRU at most.
 551  */
 552 long osc_lru_shrink(const struct lu_env *env, struct client_obd *cli,
 553                    long target, bool force)
 554 {
 555         struct cl_io *io;
 556         struct cl_object *clobj = NULL;
 557         struct cl_page **pvec;
 558         struct osc_page *opg;
 559         long count = 0;
 560         int maxscan = 0;
 561         int index = 0;
 562         int rc = 0;
 563         ENTRY;
 564
 565         LASSERT(atomic_long_read(&cli->cl_lru_in_list) >= 0);
 566         if (atomic_long_read(&cli->cl_lru_in_list) == 0 || target <= 0)
 567                 RETURN(0);
 568
 569         CDEBUG(D_CACHE, "%s: shrinkers: %d, force: %d\n",
 570                cli_name(cli), atomic_read(&cli->cl_lru_shrinkers), force);
 571         if (!force) {
 572                 if (atomic_read(&cli->cl_lru_shrinkers) > 0)
 573                         RETURN(-EBUSY);
 574
 575                 if (atomic_inc_return(&cli->cl_lru_shrinkers) > 1) {
 576                         atomic_dec(&cli->cl_lru_shrinkers);
 577                         RETURN(-EBUSY);
 578                 }
 579         } else {
 580                 atomic_inc(&cli->cl_lru_shrinkers);
 581         }
 582
 583         pvec = (struct cl_page **)osc_env_info(env)->oti_pvec;
 584         io = &osc_env_info(env)->oti_io;
 585
 586         spin_lock(&cli->cl_lru_list_lock);
 587         if (force)
 588                 cli->cl_lru_reclaim++;
 589         maxscan = min(target << 1, atomic_long_read(&cli->cl_lru_in_list));
 590         while (!list_empty(&cli->cl_lru_list)) {
 591                 struct cl_page *page;
 592                 bool will_free = false;
 593
 594                 if (!force && atomic_read(&cli->cl_lru_shrinkers) > 1)
 595                         break;
 596
 597                 if (--maxscan < 0)
 598                         break;
 599
 600                 opg = list_entry(cli->cl_lru_list.next, struct osc_page,
 601                                  ops_lru);
 602                 page = opg->ops_cl.cpl_page;
 603                 if (lru_page_busy(cli, page)) {
 604                         list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
 605                         continue;
 606                 }
 607
 608                 LASSERT(page->cp_obj != NULL);
 609                 if (clobj != page->cp_obj) {
 610                         struct cl_object *tmp = page->cp_obj;
 611
 612                         cl_object_get(tmp);
 613                         spin_unlock(&cli->cl_lru_list_lock);
 614
 615                         if (clobj != NULL) {
 616                                 discard_pagevec(env, io, pvec, index);
 617                                 index = 0;
 618
 619                                 cl_io_fini(env, io);
 620                                 cl_object_put(env, clobj);
 621                                 clobj = NULL;
 622                         }
 623
 624                         clobj = tmp;
 625                         io->ci_obj = clobj;
 626                         io->ci_ignore_layout = 1;
 627                         rc = cl_io_init(env, io, CIT_MISC, clobj);
 628
 629                         spin_lock(&cli->cl_lru_list_lock);
 630
 631                         if (rc != 0)
 632                                 break;
 633
 634                         ++maxscan;
 635                         continue;
 636                 }
 637
 638                 if (cl_page_own_try(env, io, page) == 0) {
 639                         if (!lru_page_busy(cli, page)) {
 640                                 /* remove it from lru list earlier to avoid
 641                                  * lock contention */
 642                                 __osc_lru_del(cli, opg);
 643                                 opg->ops_in_lru = 0; /* will be discarded */
 644
 645                                 cl_page_get(page);
 646                                 will_free = true;
 647                         } else {
 648                                 cl_page_disown(env, io, page);
 649                         }
 650                 }
 651
 652                 if (!will_free) {
 653                         list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
 654                         continue;
 655                 }
 656
 657                 /* Don't discard and free the page with cl_lru_list held */
 658                 pvec[index++] = page;
 659                 if (unlikely(index == OTI_PVEC_SIZE)) {
 660                         spin_unlock(&cli->cl_lru_list_lock);
 661                         discard_pagevec(env, io, pvec, index);
 662                         index = 0;
 663
 664                         spin_lock(&cli->cl_lru_list_lock);
 665                 }
 666
 667                 if (++count >= target)
 668                         break;
 669         }
 670         spin_unlock(&cli->cl_lru_list_lock);
 671
 672         if (clobj != NULL) {
 673                 discard_pagevec(env, io, pvec, index);
 674
 675                 cl_io_fini(env, io);
 676                 cl_object_put(env, clobj);
 677         }
 678
 679         atomic_dec(&cli->cl_lru_shrinkers);
 680         if (count > 0) {
 681                 atomic_long_add(count, cli->cl_lru_left);
 682                 wake_up_all(&osc_lru_waitq);
 683         }
 684         RETURN(count > 0 ? count : rc);
 685 }
 686
 687 /**
 688  * Reclaim LRU pages by an IO thread. The caller wants to reclaim at least
 689  * \@npages of LRU slots. For performance consideration, it's better to drop
 690  * LRU pages in batch. Therefore, the actual number is adjusted at least
 691  * max_pages_per_rpc.
 692  */
 693 static long osc_lru_reclaim(struct client_obd *cli, unsigned long npages)
 694 {
 695         struct lu_env *env;
 696         struct cl_client_cache *cache = cli->cl_cache;
 697         int max_scans;
 698         __u16 refcheck;
 699         long rc = 0;
 700         ENTRY;
 701
 702         LASSERT(cache != NULL);
 703
 704         env = cl_env_get(&refcheck);
 705         if (IS_ERR(env))
 706                 RETURN(rc);
 707
 708         npages = max_t(int, npages, cli->cl_max_pages_per_rpc);
 709         CDEBUG(D_CACHE, "%s: start to reclaim %ld pages from LRU\n",
 710                cli_name(cli), npages);
 711         rc = osc_lru_shrink(env, cli, npages, true);
 712         if (rc >= npages) {
 713                 CDEBUG(D_CACHE, "%s: reclaimed %ld/%ld pages from LRU\n",
 714                        cli_name(cli), rc, npages);
 715                 if (osc_cache_too_much(cli) > 0)
 716                         ptlrpcd_queue_work(cli->cl_lru_work);
 717                 GOTO(out, rc);
 718         } else if (rc > 0) {
 719                 npages -= rc;
 720         }
 721
 722         CDEBUG(D_CACHE, "%s: cli %p no free slots, pages: %ld/%ld, want: %ld\n",
 723                 cli_name(cli), cli, atomic_long_read(&cli->cl_lru_in_list),
 724                 atomic_long_read(&cli->cl_lru_busy), npages);
 725
 726         /* Reclaim LRU slots from other client_obd as it can't free enough
 727          * from its own. This should rarely happen. */
 728         spin_lock(&cache->ccc_lru_lock);
 729         LASSERT(!list_empty(&cache->ccc_lru));
 730
 731         cache->ccc_lru_shrinkers++;
 732         list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
 733
 734         max_scans = atomic_read(&cache->ccc_users) - 2;
 735         while (--max_scans > 0 && !list_empty(&cache->ccc_lru)) {
 736                 cli = list_entry(cache->ccc_lru.next, struct client_obd,
 737                                  cl_lru_osc);
 738
 739                 CDEBUG(D_CACHE, "%s: cli %p LRU pages: %ld, busy: %ld.\n",
 740                         cli_name(cli), cli,
 741                         atomic_long_read(&cli->cl_lru_in_list),
 742                         atomic_long_read(&cli->cl_lru_busy));
 743
 744                 list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
 745                 if (osc_cache_too_much(cli) > 0) {
 746                         spin_unlock(&cache->ccc_lru_lock);
 747
 748                         rc = osc_lru_shrink(env, cli, npages, true);
 749                         spin_lock(&cache->ccc_lru_lock);
 750                         if (rc >= npages)
 751                                 break;
 752                         if (rc > 0)
 753                                 npages -= rc;
 754                 }
 755         }
 756         spin_unlock(&cache->ccc_lru_lock);
 757
 758 out:
 759         cl_env_put(env, &refcheck);
 760         CDEBUG(D_CACHE, "%s: cli %p freed %ld pages.\n",
 761                 cli_name(cli), cli, rc);
 762         return rc;
 763 }
 764
 765 /**
 766  * osc_lru_alloc() is called to allocate an LRU slot for a cl_page.
 767  *
 768  * Usually the LRU slots are reserved in osc_io_iter_rw_init().
 769  * Only in the case that the LRU slots are in extreme shortage, it should
 770  * have reserved enough slots for an IO.
 771  */
 772 static int osc_lru_alloc(const struct lu_env *env, struct client_obd *cli,
 773                          struct osc_page *opg)
 774 {
 775         struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
 776         struct osc_io *oio = osc_env_io(env);
 777         int rc = 0;
 778         ENTRY;
 779
 780         if (cli->cl_cache == NULL) /* shall not be in LRU */
 781                 RETURN(0);
 782
 783         if (oio->oi_lru_reserved > 0) {
 784                 --oio->oi_lru_reserved;
 785                 goto out;
 786         }
 787
 788         LASSERT(atomic_long_read(cli->cl_lru_left) >= 0);
 789         while (!atomic_long_add_unless(cli->cl_lru_left, -1, 0)) {
 790                 /* run out of LRU spaces, try to drop some by itself */
 791                 rc = osc_lru_reclaim(cli, 1);
 792                 if (rc < 0)
 793                         break;
 794                 if (rc > 0)
 795                         continue;
 796
 797                 cond_resched();
 798                 rc = l_wait_event(osc_lru_waitq,
 799                                 atomic_long_read(cli->cl_lru_left) > 0,
 800                                 &lwi);
 801                 if (rc < 0)
 802                         break;
 803         }
 804
 805 out:
 806         if (rc >= 0) {
 807                 atomic_long_inc(&cli->cl_lru_busy);
 808                 opg->ops_in_lru = 1;
 809                 rc = 0;
 810         }
 811
 812         RETURN(rc);
 813 }
 814
 815 /**
 816  * osc_lru_reserve() is called to reserve enough LRU slots for I/O.
 817  *
 818  * The benefit of doing this is to reduce contention against atomic counter
 819  * cl_lru_left by changing it from per-page access to per-IO access.
 820  */
 821 unsigned long osc_lru_reserve(struct client_obd *cli, unsigned long npages)
 822 {
 823         unsigned long reserved = 0;
 824         unsigned long max_pages;
 825         unsigned long c;
 826
 827         /* reserve a full RPC window at most to avoid that a thread accidentally
 828          * consumes too many LRU slots */
 829         max_pages = cli->cl_max_pages_per_rpc * cli->cl_max_rpcs_in_flight;
 830         if (npages > max_pages)
 831                 npages = max_pages;
 832
 833         c = atomic_long_read(cli->cl_lru_left);
 834         if (c < npages && osc_lru_reclaim(cli, npages) > 0)
 835                 c = atomic_long_read(cli->cl_lru_left);
 836         while (c >= npages) {
 837                 if (c == atomic_long_cmpxchg(cli->cl_lru_left, c, c - npages)) {
 838                         reserved = npages;
 839                         break;
 840                 }
 841                 c = atomic_long_read(cli->cl_lru_left);
 842         }
 843         if (atomic_long_read(cli->cl_lru_left) < max_pages) {
 844                 /* If there aren't enough pages in the per-OSC LRU then
 845                  * wake up the LRU thread to try and clear out space, so
 846                  * we don't block if pages are being dirtied quickly. */
 847                 CDEBUG(D_CACHE, "%s: queue LRU, left: %lu/%ld.\n",
 848                        cli_name(cli), atomic_long_read(cli->cl_lru_left),
 849                        max_pages);
 850                 (void)ptlrpcd_queue_work(cli->cl_lru_work);
 851         }
 852
 853         return reserved;
 854 }
 855
 856 /**
 857  * osc_lru_unreserve() is called to unreserve LRU slots.
 858  *
 859  * LRU slots reserved by osc_lru_reserve() may have entries left due to several
 860  * reasons such as page already existing or I/O error. Those reserved slots
 861  * should be freed by calling this function.
 862  */
 863 void osc_lru_unreserve(struct client_obd *cli, unsigned long npages)
 864 {
 865         atomic_long_add(npages, cli->cl_lru_left);
 866         wake_up_all(&osc_lru_waitq);
 867 }
 868
 869 /**
 870  * Atomic operations are expensive. We accumulate the accounting for the
 871  * same page zone to get better performance.
 872  * In practice this can work pretty good because the pages in the same RPC
 873  * are likely from the same page zone.
 874  */
 875 static inline void unstable_page_accounting(struct ptlrpc_bulk_desc *desc,
 876                                             int factor)
 877 {
 878         int page_count = desc->bd_iov_count;
 879         void *zone = NULL;
 880         int count = 0;
 881         int i;
 882
 883         LASSERT(ptlrpc_is_bulk_desc_kiov(desc->bd_type));
 884
 885         for (i = 0; i < page_count; i++) {
 886                 void *pz = page_zone(BD_GET_KIOV(desc, i).kiov_page);
 887
 888                 if (likely(pz == zone)) {
 889                         ++count;
 890                         continue;
 891                 }
 892
 893                 if (count > 0) {
 894                         mod_zone_page_state(zone, NR_UNSTABLE_NFS,
 895                                             factor * count);
 896                         count = 0;
 897                 }
 898                 zone = pz;
 899                 ++count;
 900         }
 901         if (count > 0)
 902                 mod_zone_page_state(zone, NR_UNSTABLE_NFS, factor * count);
 903 }
 904
 905 static inline void add_unstable_page_accounting(struct ptlrpc_bulk_desc *desc)
 906 {
 907         unstable_page_accounting(desc, 1);
 908 }
 909
 910 static inline void dec_unstable_page_accounting(struct ptlrpc_bulk_desc *desc)
 911 {
 912         unstable_page_accounting(desc, -1);
 913 }
 914
 915 /**
 916  * Performs "unstable" page accounting. This function balances the
 917  * increment operations performed in osc_inc_unstable_pages. It is
 918  * registered as the RPC request callback, and is executed when the
 919  * bulk RPC is committed on the server. Thus at this point, the pages
 920  * involved in the bulk transfer are no longer considered unstable.
 921  *
 922  * If this function is called, the request should have been committed
 923  * or req:rq_unstable must have been set; it implies that the unstable
 924  * statistic have been added.
 925  */
 926 void osc_dec_unstable_pages(struct ptlrpc_request *req)
 927 {
 928         struct ptlrpc_bulk_desc *desc       = req->rq_bulk;
 929         struct client_obd       *cli        = &req->rq_import->imp_obd->u.cli;
 930         int                      page_count = desc->bd_iov_count;
 931         long                     unstable_count;
 932
 933         LASSERT(page_count >= 0);
 934         dec_unstable_page_accounting(desc);
 935
 936         unstable_count = atomic_long_sub_return(page_count,
 937                                                 &cli->cl_unstable_count);
 938         LASSERT(unstable_count >= 0);
 939
 940         unstable_count = atomic_long_sub_return(page_count,
 941                                            &cli->cl_cache->ccc_unstable_nr);
 942         LASSERT(unstable_count >= 0);
 943         if (unstable_count == 0)
 944                 wake_up_all(&cli->cl_cache->ccc_unstable_waitq);
 945
 946         if (waitqueue_active(&osc_lru_waitq))
 947                 (void)ptlrpcd_queue_work(cli->cl_lru_work);
 948 }
 949
 950 /**
 951  * "unstable" page accounting. See: osc_dec_unstable_pages.
 952  */
 953 void osc_inc_unstable_pages(struct ptlrpc_request *req)
 954 {
 955         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
 956         struct client_obd       *cli  = &req->rq_import->imp_obd->u.cli;
 957         long                     page_count = desc->bd_iov_count;
 958
 959         /* No unstable page tracking */
 960         if (cli->cl_cache == NULL || !cli->cl_cache->ccc_unstable_check)
 961                 return;
 962
 963         add_unstable_page_accounting(desc);
 964         atomic_long_add(page_count, &cli->cl_unstable_count);
 965         atomic_long_add(page_count, &cli->cl_cache->ccc_unstable_nr);
 966
 967         /* If the request has already been committed (i.e. brw_commit
 968          * called via rq_commit_cb), we need to undo the unstable page
 969          * increments we just performed because rq_commit_cb wont be
 970          * called again. */
 971         spin_lock(&req->rq_lock);
 972         if (unlikely(req->rq_committed)) {
 973                 spin_unlock(&req->rq_lock);
 974
 975                 osc_dec_unstable_pages(req);
 976         } else {
 977                 req->rq_unstable = 1;
 978                 spin_unlock(&req->rq_lock);
 979         }
 980 }
 981
 982 /**
 983  * Check if it piggybacks SOFT_SYNC flag to OST from this OSC.
 984  * This function will be called by every BRW RPC so it's critical
 985  * to make this function fast.
 986  */
 987 bool osc_over_unstable_soft_limit(struct client_obd *cli)
 988 {
 989         long unstable_nr, osc_unstable_count;
 990
 991         /* Can't check cli->cl_unstable_count, therefore, no soft limit */
 992         if (cli->cl_cache == NULL || !cli->cl_cache->ccc_unstable_check)
 993                 return false;
 994
 995         osc_unstable_count = atomic_long_read(&cli->cl_unstable_count);
 996         unstable_nr = atomic_long_read(&cli->cl_cache->ccc_unstable_nr);
 997
 998         CDEBUG(D_CACHE,
 999                "%s: cli: %p unstable pages: %lu, osc unstable pages: %lu\n",
1000                cli_name(cli), cli, unstable_nr, osc_unstable_count);
1001
1002         /* If the LRU slots are in shortage - 25% remaining AND this OSC
1003          * has one full RPC window of unstable pages, it's a good chance
1004          * to piggyback a SOFT_SYNC flag.
1005          * Please notice that the OST won't take immediate response for the
1006          * SOFT_SYNC request so active OSCs will have more chance to carry
1007          * the flag, this is reasonable. */
1008         return unstable_nr > cli->cl_cache->ccc_lru_max >> 2 &&
1009                osc_unstable_count > cli->cl_max_pages_per_rpc *
1010                                     cli->cl_max_rpcs_in_flight;
1011 }
1012
1013 /**
1014  * Return how many LRU pages in the cache of all OSC devices
1015  *
1016  * \retval      return # of cached LRU pages times reclaimation tendency
1017  * \retval      SHRINK_STOP if it cannot do any scanning in this time
1018  */
1019 unsigned long osc_cache_shrink_count(struct shrinker *sk,
1020                                      struct shrink_control *sc)
1021 {
1022         struct client_obd *cli;
1023         unsigned long cached = 0;
1024
1025         spin_lock(&osc_shrink_lock);
1026         list_for_each_entry(cli, &osc_shrink_list, cl_shrink_list)
1027                 cached += atomic_long_read(&cli->cl_lru_in_list);
1028         spin_unlock(&osc_shrink_lock);
1029
1030         return (cached  * sysctl_vfs_cache_pressure) / 100;
1031 }
1032
1033 /**
1034  * Scan and try to reclaim sc->nr_to_scan cached LRU pages
1035  *
1036  * \retval      number of cached LRU pages reclaimed
1037  * \retval      SHRINK_STOP if it cannot do any scanning in this time
1038  *
1039  * Linux kernel will loop calling this shrinker scan routine with
1040  * sc->nr_to_scan = SHRINK_BATCH(128 for now) until kernel got enough memory.
1041  *
1042  * If sc->nr_to_scan is 0, the VM is querying the cache size, we don't need
1043  * to scan and try to reclaim LRU pages, just return 0 and
1044  * osc_cache_shrink_count() will report the LRU page number.
1045  */
1046 unsigned long osc_cache_shrink_scan(struct shrinker *sk,
1047                                     struct shrink_control *sc)
1048 {
1049         struct client_obd *cli;
1050         struct client_obd *stop_anchor = NULL;
1051         struct lu_env *env;
1052         long shrank = 0;
1053         int rc;
1054         __u16 refcheck;
1055
1056         if (sc->nr_to_scan == 0)
1057                 return 0;
1058
1059         if (!(sc->gfp_mask & __GFP_FS))
1060                 return SHRINK_STOP;
1061
1062         env = cl_env_get(&refcheck);
1063         if (IS_ERR(env))
1064                 return SHRINK_STOP;
1065
1066         spin_lock(&osc_shrink_lock);
1067         while (!list_empty(&osc_shrink_list)) {
1068                 cli = list_entry(osc_shrink_list.next, struct client_obd,
1069                                  cl_shrink_list);
1070
1071                 if (stop_anchor == NULL)
1072                         stop_anchor = cli;
1073                 else if (cli == stop_anchor)
1074                         break;
1075
1076                 list_move_tail(&cli->cl_shrink_list, &osc_shrink_list);
1077                 spin_unlock(&osc_shrink_lock);
1078
1079                 /* shrink no more than max_pages_per_rpc for an OSC */
1080                 rc = osc_lru_shrink(env, cli, (sc->nr_to_scan - shrank) >
1081                                     cli->cl_max_pages_per_rpc ?
1082                                     cli->cl_max_pages_per_rpc :
1083                                     sc->nr_to_scan - shrank, true);
1084                 if (rc > 0)
1085                         shrank += rc;
1086
1087                 if (shrank >= sc->nr_to_scan)
1088                         goto out;
1089
1090                 spin_lock(&osc_shrink_lock);
1091         }
1092         spin_unlock(&osc_shrink_lock);
1093
1094 out:
1095         cl_env_put(env, &refcheck);
1096
1097         return shrank;
1098 }
1099
1100 /** @} osc */