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, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  *
  31  * Implementation of cl_page for OSC layer.
  32  *
  33  *   Author: Nikita Danilov <nikita.danilov@sun.com>
  34  *   Author: Jinshan Xiong <jinshan.xiong@intel.com>
  35  */
  36
  37 #define DEBUG_SUBSYSTEM S_OSC
  38 #include <lustre_osc.h>
  39
  40 #include "osc_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 = osc_page_object(opg);
  85
  86         osc_lru_use(osc_cli(obj), opg);
  87 }
  88
  89 int osc_page_cache_add(const struct lu_env *env, struct osc_page *opg,
  90                        struct cl_io *io, cl_commit_cbt cb)
  91 {
  92         int result;
  93         ENTRY;
  94
  95         osc_page_transfer_get(opg, "transfer\0cache");
  96         result = osc_queue_async_io(env, io, opg, cb);
  97         if (result != 0)
  98                 osc_page_transfer_put(env, opg);
  99         else
 100                 osc_page_transfer_add(env, opg, CRT_WRITE);
 101
 102         RETURN(result);
 103 }
 104
 105 void osc_index2policy(union ldlm_policy_data *policy,
 106                       const struct cl_object *obj, pgoff_t start, pgoff_t end)
 107 {
 108         memset(policy, 0, sizeof *policy);
 109         policy->l_extent.start = cl_offset(obj, start);
 110         policy->l_extent.end   = cl_offset(obj, end + 1) - 1;
 111 }
 112
 113 static int osc_page_print(const struct lu_env *env,
 114                           const struct cl_page_slice *slice,
 115                           void *cookie, lu_printer_t printer)
 116 {
 117         struct osc_page *opg = cl2osc_page(slice);
 118         struct osc_async_page *oap = &opg->ops_oap;
 119         struct osc_object *obj = osc_page_object(opg);
 120         struct client_obd *cli = &osc_export(obj)->exp_obd->u.cli;
 121
 122         return (*printer)(env, cookie, LUSTRE_OSC_NAME"-page@%p %lu: "
 123                           "1< %d %c %c > "
 124                           "2< %lld %u %u %#x %#x | %p %p %p > "
 125                           "3< %d %d > "
 126                           "4< %d %d %d %lu %c | %c %c %c %c > "
 127                           "5< %c %c %c %c | %d %c | %d %c %c>\n",
 128                           opg, osc_index(opg),
 129                           /* 1 */
 130                           oap->oap_cmd,
 131                           list_empty_marker(&oap->oap_pending_item),
 132                           list_empty_marker(&oap->oap_rpc_item),
 133                           /* 2 */
 134                           oap->oap_obj_off, oap->oap_page_off, oap->oap_count,
 135                           oap->oap_async_flags, oap->oap_brw_flags,
 136                           oap->oap_request, cli, obj,
 137                           /* 3 */
 138                           opg->ops_transfer_pinned,
 139                           opg->ops_srvlock,
 140                           /* 4 */
 141                           cli->cl_r_in_flight, cli->cl_w_in_flight,
 142                           cli->cl_max_rpcs_in_flight,
 143                           cli->cl_avail_grant,
 144                           waitqueue_active(&cli->cl_cache_waiters) ? '+' : '-',
 145                           list_empty_marker(&cli->cl_loi_ready_list),
 146                           list_empty_marker(&cli->cl_loi_hp_ready_list),
 147                           list_empty_marker(&cli->cl_loi_write_list),
 148                           list_empty_marker(&cli->cl_loi_read_list),
 149                           /* 5 */
 150                           list_empty_marker(&obj->oo_ready_item),
 151                           list_empty_marker(&obj->oo_hp_ready_item),
 152                           list_empty_marker(&obj->oo_write_item),
 153                           list_empty_marker(&obj->oo_read_item),
 154                           atomic_read(&obj->oo_nr_reads),
 155                           list_empty_marker(&obj->oo_reading_exts),
 156                           atomic_read(&obj->oo_nr_writes),
 157                           list_empty_marker(&obj->oo_hp_exts),
 158                           list_empty_marker(&obj->oo_urgent_exts));
 159 }
 160
 161 static void osc_page_delete(const struct lu_env *env,
 162                             const struct cl_page_slice *slice)
 163 {
 164         struct osc_page   *opg = cl2osc_page(slice);
 165         struct osc_object *obj = osc_page_object(opg);
 166         int rc;
 167
 168         ENTRY;
 169         CDEBUG(D_TRACE, "%p\n", opg);
 170         osc_page_transfer_put(env, opg);
 171         rc = osc_teardown_async_page(env, obj, opg);
 172         if (rc) {
 173                 CL_PAGE_DEBUG(D_ERROR, env, slice->cpl_page,
 174                               "Trying to teardown failed: %d\n", rc);
 175                 LASSERT(0);
 176         }
 177
 178         osc_lru_del(osc_cli(obj), opg);
 179
 180         if (slice->cpl_page->cp_type == CPT_CACHEABLE) {
 181                 void *value = NULL;
 182
 183                 spin_lock(&obj->oo_tree_lock);
 184                 if (opg->ops_intree) {
 185                         value = radix_tree_delete(&obj->oo_tree,
 186                                                   osc_index(opg));
 187                         if (value != NULL) {
 188                                 --obj->oo_npages;
 189                                 opg->ops_intree = 0;
 190                         }
 191                 }
 192                 spin_unlock(&obj->oo_tree_lock);
 193
 194                 LASSERT(ergo(value != NULL, value == opg));
 195         }
 196
 197         EXIT;
 198 }
 199
 200 static void osc_page_clip(const struct lu_env *env,
 201                           const struct cl_page_slice *slice,
 202                           int from, int to)
 203 {
 204         struct osc_page       *opg = cl2osc_page(slice);
 205         struct osc_async_page *oap = &opg->ops_oap;
 206
 207         opg->ops_from = from;
 208         /* argument @to is exclusive, but @ops_to is inclusive */
 209         opg->ops_to   = to - 1;
 210         /* This isn't really necessary for transient pages, but we also don't
 211          * call clip on transient pages often, so it's OK.
 212          */
 213         spin_lock(&oap->oap_lock);
 214         oap->oap_async_flags |= ASYNC_COUNT_STABLE;
 215         spin_unlock(&oap->oap_lock);
 216 }
 217
 218 static int osc_page_flush(const struct lu_env *env,
 219                           const struct cl_page_slice *slice,
 220                           struct cl_io *io)
 221 {
 222         struct osc_page *opg = cl2osc_page(slice);
 223         int rc = 0;
 224         ENTRY;
 225         rc = osc_flush_async_page(env, io, opg);
 226         RETURN(rc);
 227 }
 228
 229 static void osc_page_touch(const struct lu_env *env,
 230                           const struct cl_page_slice *slice, size_t to)
 231 {
 232         struct osc_page *opg = cl2osc_page(slice);
 233         struct cl_object *obj = osc2cl(osc_page_object(opg));
 234
 235         osc_page_touch_at(env, obj, osc_index(opg), to);
 236 }
 237
 238 static const struct cl_page_operations osc_page_ops = {
 239         .cpo_print         = osc_page_print,
 240         .cpo_delete        = osc_page_delete,
 241         .cpo_clip           = osc_page_clip,
 242         .cpo_flush          = osc_page_flush,
 243         .cpo_page_touch    = osc_page_touch,
 244 };
 245
 246 int osc_page_init(const struct lu_env *env, struct cl_object *obj,
 247                   struct cl_page *cl_page, pgoff_t index)
 248 {
 249         struct osc_object *osc = cl2osc(obj);
 250         struct osc_page *opg = cl_object_page_slice(obj, cl_page);
 251         struct osc_io *oio = osc_env_io(env);
 252         int result;
 253
 254         opg->ops_from = 0;
 255         opg->ops_to = PAGE_SIZE - 1;
 256
 257         INIT_LIST_HEAD(&opg->ops_lru);
 258
 259         result = osc_prep_async_page(osc, opg, cl_page, cl_offset(obj, index));
 260         if (result != 0)
 261                 return result;
 262
 263         opg->ops_srvlock = osc_io_srvlock(oio);
 264         cl_page_slice_add(cl_page, &opg->ops_cl, obj, &osc_page_ops);
 265
 266         /* reserve an LRU space for this page */
 267         if (cl_page->cp_type == CPT_CACHEABLE) {
 268                 result = osc_lru_alloc(env, osc_cli(osc), opg);
 269                 if (result == 0) {
 270                         result = radix_tree_preload(GFP_NOFS);
 271                         if (result == 0) {
 272                                 spin_lock(&osc->oo_tree_lock);
 273                                 result = radix_tree_insert(&osc->oo_tree,
 274                                                            index, opg);
 275                                 if (result == 0) {
 276                                         ++osc->oo_npages;
 277                                         opg->ops_intree = 1;
 278                                 }
 279                                 spin_unlock(&osc->oo_tree_lock);
 280
 281                                 radix_tree_preload_end();
 282                         }
 283                 }
 284         }
 285
 286         return result;
 287 }
 288 EXPORT_SYMBOL(osc_page_init);
 289
 290 /**
 291  * Helper function called by osc_io_submit() for every page in an immediate
 292  * transfer (i.e., transferred synchronously).
 293  */
 294 void osc_page_submit(const struct lu_env *env, struct osc_page *opg,
 295                      enum cl_req_type crt, int brw_flags)
 296 {
 297         struct osc_io *oio = osc_env_io(env);
 298         struct osc_async_page *oap = &opg->ops_oap;
 299
 300         LASSERT(oap->oap_async_flags & ASYNC_READY);
 301         LASSERT(oap->oap_async_flags & ASYNC_COUNT_STABLE);
 302
 303         oap->oap_cmd = crt == CRT_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ;
 304         oap->oap_page_off = opg->ops_from;
 305         oap->oap_count = opg->ops_to - opg->ops_from + 1;
 306         oap->oap_brw_flags = OBD_BRW_SYNC | brw_flags;
 307
 308         if (oio->oi_cap_sys_resource) {
 309                 oap->oap_brw_flags |= OBD_BRW_SYS_RESOURCE;
 310                 oap->oap_cmd |= OBD_BRW_SYS_RESOURCE;
 311         }
 312
 313         osc_page_transfer_get(opg, "transfer\0imm");
 314         osc_page_transfer_add(env, opg, crt);
 315 }
 316
 317 /* --------------- LRU page management ------------------ */
 318
 319 /* OSC is a natural place to manage LRU pages as applications are specialized
 320  * to write OSC by OSC. Ideally, if one OSC is used more frequently it should
 321  * occupy more LRU slots. On the other hand, we should avoid using up all LRU
 322  * slots (client_obd::cl_lru_left) otherwise process has to be put into sleep
 323  * for free LRU slots - this will be very bad so the algorithm requires each
 324  * OSC to free slots voluntarily to maintain a reasonable number of free slots
 325  * at any time.
 326  */
 327
 328 static DECLARE_WAIT_QUEUE_HEAD(osc_lru_waitq);
 329
 330 /**
 331  * LRU pages are freed in batch mode. OSC should at least free this
 332  * number of pages to avoid running out of LRU slots.
 333  */
 334 static inline int lru_shrink_min(struct client_obd *cli)
 335 {
 336         return cli->cl_max_pages_per_rpc * 2;
 337 }
 338
 339 /**
 340  * free this number at most otherwise it will take too long time to finsih.
 341  */
 342 static inline int lru_shrink_max(struct client_obd *cli)
 343 {
 344         return cli->cl_max_pages_per_rpc * cli->cl_max_rpcs_in_flight;
 345 }
 346
 347 /**
 348  * Check if we can free LRU slots from this OSC. If there exists LRU waiters,
 349  * we should free slots aggressively. In this way, slots are freed in a steady
 350  * step to maintain fairness among OSCs.
 351  *
 352  * Return how many LRU pages should be freed.
 353  */
 354 static int osc_cache_too_much(struct client_obd *cli)
 355 {
 356         struct cl_client_cache *cache = cli->cl_cache;
 357         long pages = atomic_long_read(&cli->cl_lru_in_list);
 358         unsigned long budget;
 359
 360         LASSERT(cache != NULL);
 361         budget = cache->ccc_lru_max / (atomic_read(&cache->ccc_users) - 2);
 362
 363         /* if it's going to run out LRU slots, we should free some, but not
 364          * too much to maintain faireness among OSCs. */
 365         if (atomic_long_read(cli->cl_lru_left) < cache->ccc_lru_max >> 2) {
 366                 if (pages >= budget)
 367                         return lru_shrink_max(cli);
 368                 else if (pages >= budget / 2)
 369                         return lru_shrink_min(cli);
 370         } else {
 371                 time64_t duration = ktime_get_real_seconds();
 372                 long timediff;
 373
 374                 /* knock out pages by duration of no IO activity */
 375                 duration -= cli->cl_lru_last_used;
 376                 /*
 377                  * The difference shouldn't be more than 70 years
 378                  * so we can safely case to a long. Round to
 379                  * approximately 1 minute.
 380                  */
 381                 timediff = (long)(duration >> 6);
 382                 if (timediff > 0 && pages >= budget / timediff)
 383                         return lru_shrink_min(cli);
 384         }
 385         return 0;
 386 }
 387
 388 int lru_queue_work(const struct lu_env *env, void *data)
 389 {
 390         struct client_obd *cli = data;
 391         int count;
 392
 393         CDEBUG(D_CACHE, "%s: run LRU work for client obd\n", cli_name(cli));
 394         count = osc_cache_too_much(cli);
 395         if (count > 0) {
 396                 int rc = osc_lru_shrink(env, cli, count, false);
 397
 398                 CDEBUG(D_CACHE, "%s: shrank %d/%d pages from client obd\n",
 399                        cli_name(cli), rc, count);
 400                 if (rc >= count) {
 401                         CDEBUG(D_CACHE, "%s: queue again\n", cli_name(cli));
 402                         ptlrpcd_queue_work(cli->cl_lru_work);
 403                 }
 404         }
 405
 406         RETURN(0);
 407 }
 408
 409 void osc_lru_add_batch(struct client_obd *cli, struct list_head *plist)
 410 {
 411         LIST_HEAD(lru);
 412         struct osc_async_page *oap;
 413         long npages = 0;
 414
 415         list_for_each_entry(oap, plist, oap_pending_item) {
 416                 struct osc_page *opg = oap2osc_page(oap);
 417
 418                 if (!opg->ops_in_lru)
 419                         continue;
 420
 421                 ++npages;
 422                 LASSERT(list_empty(&opg->ops_lru));
 423                 list_add(&opg->ops_lru, &lru);
 424         }
 425
 426         if (npages > 0) {
 427                 spin_lock(&cli->cl_lru_list_lock);
 428                 list_splice_tail(&lru, &cli->cl_lru_list);
 429                 atomic_long_sub(npages, &cli->cl_lru_busy);
 430                 atomic_long_add(npages, &cli->cl_lru_in_list);
 431                 cli->cl_lru_last_used = ktime_get_real_seconds();
 432                 spin_unlock(&cli->cl_lru_list_lock);
 433
 434                 if (waitqueue_active(&osc_lru_waitq))
 435                         (void)ptlrpcd_queue_work(cli->cl_lru_work);
 436         }
 437 }
 438
 439 static void __osc_lru_del(struct client_obd *cli, struct osc_page *opg)
 440 {
 441         LASSERT(atomic_long_read(&cli->cl_lru_in_list) > 0);
 442         list_del_init(&opg->ops_lru);
 443         atomic_long_dec(&cli->cl_lru_in_list);
 444 }
 445
 446 /**
 447  * Page is being destroyed. The page may be not in LRU list, if the transfer
 448  * has never finished(error occurred).
 449  */
 450 static void osc_lru_del(struct client_obd *cli, struct osc_page *opg)
 451 {
 452         if (opg->ops_in_lru) {
 453                 spin_lock(&cli->cl_lru_list_lock);
 454                 if (!list_empty(&opg->ops_lru)) {
 455                         __osc_lru_del(cli, opg);
 456                 } else {
 457                         LASSERT(atomic_long_read(&cli->cl_lru_busy) > 0);
 458                         atomic_long_dec(&cli->cl_lru_busy);
 459                 }
 460                 spin_unlock(&cli->cl_lru_list_lock);
 461
 462                 atomic_long_inc(cli->cl_lru_left);
 463                 /* this is a great place to release more LRU pages if
 464                  * this osc occupies too many LRU pages and kernel is
 465                  * stealing one of them. */
 466                 if (osc_cache_too_much(cli)) {
 467                         CDEBUG(D_CACHE, "%s: queue LRU work\n", cli_name(cli));
 468                         (void)ptlrpcd_queue_work(cli->cl_lru_work);
 469                 }
 470                 wake_up(&osc_lru_waitq);
 471         } else {
 472                 LASSERT(list_empty(&opg->ops_lru));
 473         }
 474 }
 475
 476 /**
 477  * Delete page from LRU list for redirty.
 478  */
 479 static void osc_lru_use(struct client_obd *cli, struct osc_page *opg)
 480 {
 481         /* If page is being transferred for the first time,
 482          * ops_lru should be empty */
 483         if (opg->ops_in_lru) {
 484                 if (list_empty(&opg->ops_lru))
 485                         return;
 486                 spin_lock(&cli->cl_lru_list_lock);
 487                 if (!list_empty(&opg->ops_lru)) {
 488                         __osc_lru_del(cli, opg);
 489                         atomic_long_inc(&cli->cl_lru_busy);
 490                 }
 491                 spin_unlock(&cli->cl_lru_list_lock);
 492         }
 493 }
 494
 495 static void discard_pagevec(const struct lu_env *env, struct cl_io *io,
 496                                 struct cl_page **pvec, int max_index)
 497 {
 498         struct pagevec *pagevec = &osc_env_info(env)->oti_pagevec;
 499         int i;
 500
 501         ll_pagevec_init(pagevec, 0);
 502         for (i = 0; i < max_index; i++) {
 503                 struct cl_page *page = pvec[i];
 504
 505                 LASSERT(cl_page_is_owned(page, io));
 506                 cl_page_delete(env, page);
 507                 cl_page_discard(env, io, page);
 508                 cl_page_disown(env, io, page);
 509                 cl_pagevec_put(env, page, pagevec);
 510
 511                 pvec[i] = NULL;
 512         }
 513         pagevec_release(pagevec);
 514 }
 515
 516 /**
 517  * Check if a cl_page can be released, i.e, it's not being used.
 518  *
 519  * If unstable account is turned on, bulk transfer may hold one refcount
 520  * for recovery so we need to check vmpage refcount as well; otherwise,
 521  * even we can destroy cl_page but the corresponding vmpage can't be reused.
 522  */
 523 static inline bool lru_page_busy(struct client_obd *cli, struct cl_page *page)
 524 {
 525         if (cl_page_in_use_noref(page))
 526                 return true;
 527
 528         if (cli->cl_cache->ccc_unstable_check) {
 529                 struct page *vmpage = cl_page_vmpage(page);
 530
 531                 /* this check is racy because the vmpage is not locked, but
 532                  * that's OK - the code which does the actual page release
 533                  * checks this again before releasing
 534                  *
 535                  * vmpage have two known users: cl_page and VM page cache
 536                  */
 537                 if (vmpage_in_use(vmpage, 0))
 538                         return true;
 539         }
 540         return false;
 541 }
 542
 543 /**
 544  * Drop @target of pages from LRU at most.
 545  */
 546 long osc_lru_shrink(const struct lu_env *env, struct client_obd *cli,
 547                    long target, bool force)
 548 {
 549         struct cl_io *io;
 550         struct cl_object *clobj = NULL;
 551         struct cl_page **pvec;
 552         struct osc_page *opg;
 553         long count = 0;
 554         int maxscan = 0;
 555         int index = 0;
 556         int rc = 0;
 557         ENTRY;
 558
 559         LASSERT(atomic_long_read(&cli->cl_lru_in_list) >= 0);
 560         if (atomic_long_read(&cli->cl_lru_in_list) == 0 || target <= 0)
 561                 RETURN(0);
 562
 563         CDEBUG(D_CACHE, "%s: shrinkers: %d, force: %d\n",
 564                cli_name(cli), atomic_read(&cli->cl_lru_shrinkers), force);
 565         if (!force) {
 566                 if (atomic_read(&cli->cl_lru_shrinkers) > 0)
 567                         RETURN(-EBUSY);
 568
 569                 if (atomic_inc_return(&cli->cl_lru_shrinkers) > 1) {
 570                         atomic_dec(&cli->cl_lru_shrinkers);
 571                         RETURN(-EBUSY);
 572                 }
 573         } else {
 574                 atomic_inc(&cli->cl_lru_shrinkers);
 575         }
 576
 577         pvec = (struct cl_page **)osc_env_info(env)->oti_pvec;
 578         io = osc_env_thread_io(env);
 579
 580         spin_lock(&cli->cl_lru_list_lock);
 581         if (force)
 582                 cli->cl_lru_reclaim++;
 583         maxscan = min(target << 1, atomic_long_read(&cli->cl_lru_in_list));
 584         while (!list_empty(&cli->cl_lru_list)) {
 585                 struct cl_page *page;
 586                 bool will_free = false;
 587
 588                 if (!force && atomic_read(&cli->cl_lru_shrinkers) > 1)
 589                         break;
 590
 591                 if (--maxscan < 0)
 592                         break;
 593
 594                 opg = list_first_entry(&cli->cl_lru_list, struct osc_page,
 595                                        ops_lru);
 596                 page = opg->ops_cl.cpl_page;
 597                 if (lru_page_busy(cli, page)) {
 598                         list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
 599                         continue;
 600                 }
 601
 602                 LASSERT(page->cp_obj != NULL);
 603                 if (clobj != page->cp_obj) {
 604                         struct cl_object *tmp = page->cp_obj;
 605
 606                         cl_object_get(tmp);
 607                         spin_unlock(&cli->cl_lru_list_lock);
 608
 609                         if (clobj != NULL) {
 610                                 discard_pagevec(env, io, pvec, index);
 611                                 index = 0;
 612
 613                                 cl_io_fini(env, io);
 614                                 cl_object_put(env, clobj);
 615                                 clobj = NULL;
 616                         }
 617
 618                         clobj = tmp;
 619                         io->ci_obj = clobj;
 620                         io->ci_ignore_layout = 1;
 621                         rc = cl_io_init(env, io, CIT_MISC, clobj);
 622
 623                         spin_lock(&cli->cl_lru_list_lock);
 624
 625                         if (rc != 0)
 626                                 break;
 627
 628                         ++maxscan;
 629                         continue;
 630                 }
 631
 632                 if (cl_page_own_try(env, io, page) == 0) {
 633                         if (!lru_page_busy(cli, page)) {
 634                                 /* remove it from lru list earlier to avoid
 635                                  * lock contention */
 636                                 __osc_lru_del(cli, opg);
 637                                 opg->ops_in_lru = 0; /* will be discarded */
 638
 639                                 cl_page_get(page);
 640                                 will_free = true;
 641                         } else {
 642                                 cl_page_disown(env, io, page);
 643                         }
 644                 }
 645
 646                 if (!will_free) {
 647                         list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
 648                         continue;
 649                 }
 650
 651                 /* Don't discard and free the page with cl_lru_list held */
 652                 pvec[index++] = page;
 653                 if (unlikely(index == OTI_PVEC_SIZE)) {
 654                         spin_unlock(&cli->cl_lru_list_lock);
 655                         discard_pagevec(env, io, pvec, index);
 656                         index = 0;
 657
 658                         spin_lock(&cli->cl_lru_list_lock);
 659                 }
 660
 661                 if (++count >= target)
 662                         break;
 663         }
 664         spin_unlock(&cli->cl_lru_list_lock);
 665
 666         if (clobj != NULL) {
 667                 discard_pagevec(env, io, pvec, index);
 668
 669                 cl_io_fini(env, io);
 670                 cl_object_put(env, clobj);
 671         }
 672
 673         atomic_dec(&cli->cl_lru_shrinkers);
 674         if (count > 0) {
 675                 atomic_long_add(count, cli->cl_lru_left);
 676                 wake_up(&osc_lru_waitq);
 677         }
 678         RETURN(count > 0 ? count : rc);
 679 }
 680 EXPORT_SYMBOL(osc_lru_shrink);
 681
 682 /**
 683  * Reclaim LRU pages by an IO thread. The caller wants to reclaim at least
 684  * \@npages of LRU slots. For performance consideration, it's better to drop
 685  * LRU pages in batch. Therefore, the actual number is adjusted at least
 686  * max_pages_per_rpc.
 687  */
 688 static long osc_lru_reclaim(struct client_obd *cli, unsigned long npages)
 689 {
 690         struct lu_env *env;
 691         struct cl_client_cache *cache = cli->cl_cache;
 692         struct client_obd *scan;
 693         int max_scans;
 694         __u16 refcheck;
 695         long rc = 0;
 696         ENTRY;
 697
 698         LASSERT(cache != NULL);
 699
 700         env = cl_env_get(&refcheck);
 701         if (IS_ERR(env))
 702                 RETURN(rc);
 703
 704         npages = max_t(int, npages, cli->cl_max_pages_per_rpc);
 705         CDEBUG(D_CACHE, "%s: start to reclaim %ld pages from LRU\n",
 706                cli_name(cli), npages);
 707         rc = osc_lru_shrink(env, cli, npages, true);
 708         if (rc >= npages) {
 709                 CDEBUG(D_CACHE, "%s: reclaimed %ld/%ld pages from LRU\n",
 710                        cli_name(cli), rc, npages);
 711                 if (osc_cache_too_much(cli) > 0)
 712                         ptlrpcd_queue_work(cli->cl_lru_work);
 713                 GOTO(out, rc);
 714         } else if (rc > 0) {
 715                 npages -= rc;
 716         }
 717
 718         CDEBUG(D_CACHE, "%s: cli %p no free slots, pages: %ld/%ld, want: %ld\n",
 719                 cli_name(cli), cli, atomic_long_read(&cli->cl_lru_in_list),
 720                 atomic_long_read(&cli->cl_lru_busy), npages);
 721
 722         /* Reclaim LRU slots from other client_obd as it can't free enough
 723          * from its own. This should rarely happen. */
 724         spin_lock(&cache->ccc_lru_lock);
 725         LASSERT(!list_empty(&cache->ccc_lru));
 726
 727         cache->ccc_lru_shrinkers++;
 728         list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
 729
 730         max_scans = atomic_read(&cache->ccc_users) - 2;
 731         while (--max_scans > 0 &&
 732                (scan = list_first_entry_or_null(&cache->ccc_lru,
 733                                                   struct client_obd,
 734                                                   cl_lru_osc)) != NULL) {
 735                 CDEBUG(D_CACHE, "%s: cli %p LRU pages: %ld, busy: %ld.\n",
 736                        cli_name(scan), scan,
 737                        atomic_long_read(&scan->cl_lru_in_list),
 738                        atomic_long_read(&scan->cl_lru_busy));
 739
 740                 list_move_tail(&scan->cl_lru_osc, &cache->ccc_lru);
 741                 if (osc_cache_too_much(scan) > 0) {
 742                         spin_unlock(&cache->ccc_lru_lock);
 743
 744                         rc = osc_lru_shrink(env, scan, npages, true);
 745                         spin_lock(&cache->ccc_lru_lock);
 746                         if (rc >= npages)
 747                                 break;
 748                         if (rc > 0)
 749                                 npages -= rc;
 750                 }
 751         }
 752         spin_unlock(&cache->ccc_lru_lock);
 753
 754 out:
 755         cl_env_put(env, &refcheck);
 756         CDEBUG(D_CACHE, "%s: cli %p freed %ld pages.\n",
 757                cli_name(cli), cli, rc);
 758         return rc;
 759 }
 760
 761 /**
 762  * osc_lru_alloc() is called to allocate an LRU slot for a cl_page.
 763  *
 764  * Usually the LRU slots are reserved in osc_io_iter_rw_init().
 765  * Only in the case that the LRU slots are in extreme shortage, it should
 766  * have reserved enough slots for an IO.
 767  */
 768 static int osc_lru_alloc(const struct lu_env *env, struct client_obd *cli,
 769                          struct osc_page *opg)
 770 {
 771         struct osc_io *oio = osc_env_io(env);
 772         int rc = 0;
 773
 774         ENTRY;
 775
 776         if (cli->cl_cache == NULL) /* shall not be in LRU */
 777                 RETURN(0);
 778
 779         if (oio->oi_lru_reserved > 0) {
 780                 --oio->oi_lru_reserved;
 781                 goto out;
 782         }
 783
 784         LASSERT(atomic_long_read(cli->cl_lru_left) >= 0);
 785         while (!atomic_long_add_unless(cli->cl_lru_left, -1, 0)) {
 786                 /* run out of LRU spaces, try to drop some by itself */
 787                 rc = osc_lru_reclaim(cli, 1);
 788                 if (rc < 0)
 789                         break;
 790                 if (rc > 0)
 791                         continue;
 792                 /* IO issued by readahead, don't try hard */
 793                 if (oio->oi_is_readahead) {
 794                         if (atomic_long_read(cli->cl_lru_left) > 0)
 795                                 continue;
 796                         rc = -EBUSY;
 797                         break;
 798                 }
 799
 800                 cond_resched();
 801                 rc = l_wait_event_abortable(
 802                         osc_lru_waitq,
 803                         atomic_long_read(cli->cl_lru_left) > 0);
 804                 if (rc < 0) {
 805                         rc = -EINTR;
 806                         break;
 807                 }
 808         }
 809
 810 out:
 811         if (rc >= 0) {
 812                 atomic_long_inc(&cli->cl_lru_busy);
 813                 opg->ops_in_lru = 1;
 814                 rc = 0;
 815         }
 816
 817         RETURN(rc);
 818 }
 819
 820 /**
 821  * osc_lru_reserve() is called to reserve enough LRU slots for I/O.
 822  *
 823  * The benefit of doing this is to reduce contention against atomic counter
 824  * cl_lru_left by changing it from per-page access to per-IO access.
 825  */
 826 unsigned long osc_lru_reserve(struct client_obd *cli, unsigned long npages)
 827 {
 828         unsigned long reserved = 0;
 829         unsigned long max_pages;
 830         unsigned long c;
 831         int rc;
 832
 833 again:
 834         c = atomic_long_read(cli->cl_lru_left);
 835         if (c < npages && osc_lru_reclaim(cli, npages) > 0)
 836                 c = atomic_long_read(cli->cl_lru_left);
 837
 838         if (c < npages) {
 839                 /*
 840                  * Trigger writeback in the hope some LRU slot could
 841                  * be freed.
 842                  */
 843                 rc = ptlrpcd_queue_work(cli->cl_writeback_work);
 844                 if (rc)
 845                         return 0;
 846         }
 847
 848         while (c >= npages) {
 849                 if (c == atomic_long_cmpxchg(cli->cl_lru_left, c, c - npages)) {
 850                         reserved = npages;
 851                         break;
 852                 }
 853                 c = atomic_long_read(cli->cl_lru_left);
 854         }
 855
 856         if (reserved != npages) {
 857                 cond_resched();
 858                 rc = l_wait_event_abortable(
 859                         osc_lru_waitq,
 860                         atomic_long_read(cli->cl_lru_left) > 0);
 861                 goto again;
 862         }
 863
 864         max_pages = cli->cl_max_pages_per_rpc * cli->cl_max_rpcs_in_flight;
 865         if (atomic_long_read(cli->cl_lru_left) < max_pages) {
 866                 /* If there aren't enough pages in the per-OSC LRU then
 867                  * wake up the LRU thread to try and clear out space, so
 868                  * we don't block if pages are being dirtied quickly. */
 869                 CDEBUG(D_CACHE, "%s: queue LRU, left: %lu/%ld.\n",
 870                        cli_name(cli), atomic_long_read(cli->cl_lru_left),
 871                        max_pages);
 872                 (void)ptlrpcd_queue_work(cli->cl_lru_work);
 873         }
 874
 875         return reserved;
 876 }
 877
 878 /**
 879  * osc_lru_unreserve() is called to unreserve LRU slots.
 880  *
 881  * LRU slots reserved by osc_lru_reserve() may have entries left due to several
 882  * reasons such as page already existing or I/O error. Those reserved slots
 883  * should be freed by calling this function.
 884  */
 885 void osc_lru_unreserve(struct client_obd *cli, unsigned long npages)
 886 {
 887         atomic_long_add(npages, cli->cl_lru_left);
 888         wake_up(&osc_lru_waitq);
 889 }
 890
 891 /**
 892  * Atomic operations are expensive. We accumulate the accounting for the
 893  * same page zone to get better performance.
 894  * In practice this can work pretty good because the pages in the same RPC
 895  * are likely from the same page zone.
 896  */
 897 #ifdef HAVE_NR_UNSTABLE_NFS
 898 /* Old kernels use a separate counter for unstable pages,
 899  * newer kernels treat them like any other writeback.
 900  */
 901 #define NR_WRITEBACK NR_UNSTABLE_NFS
 902 #endif
 903
 904 static inline void unstable_page_accounting(struct ptlrpc_bulk_desc *desc,
 905                                             struct osc_brw_async_args *aa,
 906                                             int factor)
 907 {
 908         int page_count;
 909         void *zone = NULL;
 910         int count = 0;
 911         int i;
 912
 913         if (desc != NULL) {
 914                 page_count = desc->bd_iov_count;
 915         } else {
 916                 page_count = aa->aa_page_count;
 917         }
 918
 919         for (i = 0; i < page_count; i++) {
 920                 void *pz;
 921                 if (desc)
 922                         pz = page_zone(desc->bd_vec[i].bv_page);
 923                 else
 924                         pz = page_zone(aa->aa_ppga[i]->pg);
 925
 926                 if (likely(pz == zone)) {
 927                         ++count;
 928                         continue;
 929                 }
 930
 931                 if (count > 0) {
 932                         mod_zone_page_state(zone, NR_WRITEBACK,
 933                                             factor * count);
 934                         count = 0;
 935                 }
 936                 zone = pz;
 937                 ++count;
 938         }
 939         if (count > 0)
 940                 mod_zone_page_state(zone, NR_WRITEBACK, factor * count);
 941 }
 942
 943 static inline void add_unstable_page_accounting(struct ptlrpc_bulk_desc *desc,
 944                                                 struct osc_brw_async_args *aa)
 945 {
 946         unstable_page_accounting(desc, aa, 1);
 947 }
 948
 949 static inline void dec_unstable_page_accounting(struct ptlrpc_bulk_desc *desc,
 950                                                 struct osc_brw_async_args *aa)
 951 {
 952         unstable_page_accounting(desc, aa, -1);
 953 }
 954
 955 /**
 956  * Performs "unstable" page accounting. This function balances the
 957  * increment operations performed in osc_inc_unstable_pages. It is
 958  * registered as the RPC request callback, and is executed when the
 959  * bulk RPC is committed on the server. Thus at this point, the pages
 960  * involved in the bulk transfer are no longer considered unstable.
 961  *
 962  * If this function is called, the request should have been committed
 963  * or req:rq_unstable must have been set; it implies that the unstable
 964  * statistic have been added.
 965  */
 966 void osc_dec_unstable_pages(struct ptlrpc_request *req)
 967 {
 968         struct ptlrpc_bulk_desc *desc       = req->rq_bulk;
 969         struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
 970         struct client_obd       *cli        = &req->rq_import->imp_obd->u.cli;
 971         int                      page_count;
 972         long                     unstable_count;
 973
 974         if (desc)
 975                 page_count = desc->bd_iov_count;
 976         else
 977                 page_count = aa->aa_page_count;
 978
 979         LASSERT(page_count >= 0);
 980
 981         dec_unstable_page_accounting(desc, aa);
 982
 983         unstable_count = atomic_long_sub_return(page_count,
 984                                                 &cli->cl_unstable_count);
 985         LASSERT(unstable_count >= 0);
 986
 987         unstable_count = atomic_long_sub_return(page_count,
 988                                            &cli->cl_cache->ccc_unstable_nr);
 989         LASSERT(unstable_count >= 0);
 990         if (unstable_count == 0)
 991                 wake_up(&cli->cl_cache->ccc_unstable_waitq);
 992
 993         if (waitqueue_active(&osc_lru_waitq))
 994                 (void)ptlrpcd_queue_work(cli->cl_lru_work);
 995 }
 996
 997 /**
 998  * "unstable" page accounting. See: osc_dec_unstable_pages.
 999  */
1000 void osc_inc_unstable_pages(struct ptlrpc_request *req)
1001 {
1002         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
1003         struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
1004         struct client_obd       *cli  = &req->rq_import->imp_obd->u.cli;
1005         long                     page_count;
1006
1007         /* No unstable page tracking */
1008         if (cli->cl_cache == NULL || !cli->cl_cache->ccc_unstable_check)
1009                 return;
1010
1011         if (desc)
1012                 page_count = desc->bd_iov_count;
1013         else
1014                 page_count = aa->aa_page_count;
1015
1016         add_unstable_page_accounting(desc, aa);
1017         atomic_long_add(page_count, &cli->cl_unstable_count);
1018         atomic_long_add(page_count, &cli->cl_cache->ccc_unstable_nr);
1019
1020         /* If the request has already been committed (i.e. brw_commit
1021          * called via rq_commit_cb), we need to undo the unstable page
1022          * increments we just performed because rq_commit_cb wont be
1023          * called again. */
1024         spin_lock(&req->rq_lock);
1025         if (unlikely(req->rq_committed)) {
1026                 spin_unlock(&req->rq_lock);
1027
1028                 osc_dec_unstable_pages(req);
1029         } else {
1030                 req->rq_unstable = 1;
1031                 spin_unlock(&req->rq_lock);
1032         }
1033 }
1034
1035 /**
1036  * Check if it piggybacks SOFT_SYNC flag to OST from this OSC.
1037  * This function will be called by every BRW RPC so it's critical
1038  * to make this function fast.
1039  */
1040 bool osc_over_unstable_soft_limit(struct client_obd *cli)
1041 {
1042         long unstable_nr, osc_unstable_count;
1043
1044         /* Can't check cli->cl_unstable_count, therefore, no soft limit */
1045         if (cli->cl_cache == NULL || !cli->cl_cache->ccc_unstable_check)
1046                 return false;
1047
1048         osc_unstable_count = atomic_long_read(&cli->cl_unstable_count);
1049         unstable_nr = atomic_long_read(&cli->cl_cache->ccc_unstable_nr);
1050
1051         CDEBUG(D_CACHE,
1052                "%s: cli: %p unstable pages: %lu, osc unstable pages: %lu\n",
1053                cli_name(cli), cli, unstable_nr, osc_unstable_count);
1054
1055         /* If the LRU slots are in shortage - 25% remaining AND this OSC
1056          * has one full RPC window of unstable pages, it's a good chance
1057          * to piggyback a SOFT_SYNC flag.
1058          * Please notice that the OST won't take immediate response for the
1059          * SOFT_SYNC request so active OSCs will have more chance to carry
1060          * the flag, this is reasonable. */
1061         return unstable_nr > cli->cl_cache->ccc_lru_max >> 2 &&
1062                osc_unstable_count > cli->cl_max_pages_per_rpc *
1063                                     cli->cl_max_rpcs_in_flight;
1064 }
1065
1066 /**
1067  * Return how many LRU pages in the cache of all OSC devices
1068  *
1069  * \retval      return # of cached LRU pages times reclaimation tendency
1070  * \retval      SHRINK_STOP if it cannot do any scanning in this time
1071  */
1072 unsigned long osc_cache_shrink_count(struct shrinker *sk,
1073                                      struct shrink_control *sc)
1074 {
1075         struct client_obd *cli;
1076         unsigned long cached = 0;
1077
1078         spin_lock(&osc_shrink_lock);
1079         list_for_each_entry(cli, &osc_shrink_list, cl_shrink_list)
1080                 cached += atomic_long_read(&cli->cl_lru_in_list);
1081         spin_unlock(&osc_shrink_lock);
1082
1083         return (cached  * sysctl_vfs_cache_pressure) / 100;
1084 }
1085
1086 /**
1087  * Scan and try to reclaim sc->nr_to_scan cached LRU pages
1088  *
1089  * \retval      number of cached LRU pages reclaimed
1090  * \retval      SHRINK_STOP if it cannot do any scanning in this time
1091  *
1092  * Linux kernel will loop calling this shrinker scan routine with
1093  * sc->nr_to_scan = SHRINK_BATCH(128 for now) until kernel got enough memory.
1094  *
1095  * If sc->nr_to_scan is 0, the VM is querying the cache size, we don't need
1096  * to scan and try to reclaim LRU pages, just return 0 and
1097  * osc_cache_shrink_count() will report the LRU page number.
1098  */
1099 unsigned long osc_cache_shrink_scan(struct shrinker *sk,
1100                                     struct shrink_control *sc)
1101 {
1102         struct client_obd *cli;
1103         struct client_obd *stop_anchor = NULL;
1104         struct lu_env *env;
1105         long shrank = 0;
1106         int rc;
1107         __u16 refcheck;
1108
1109         if (sc->nr_to_scan == 0)
1110                 return 0;
1111
1112         if (!(sc->gfp_mask & __GFP_FS))
1113                 return SHRINK_STOP;
1114
1115         env = cl_env_get(&refcheck);
1116         if (IS_ERR(env))
1117                 return SHRINK_STOP;
1118
1119         spin_lock(&osc_shrink_lock);
1120         while ((cli = list_first_entry_or_null(&osc_shrink_list,
1121                                                struct client_obd,
1122                                                cl_shrink_list)) != NULL) {
1123                 if (stop_anchor == NULL)
1124                         stop_anchor = cli;
1125                 else if (cli == stop_anchor)
1126                         break;
1127
1128                 list_move_tail(&cli->cl_shrink_list, &osc_shrink_list);
1129                 spin_unlock(&osc_shrink_lock);
1130
1131                 /* shrink no more than max_pages_per_rpc for an OSC */
1132                 rc = osc_lru_shrink(env, cli, (sc->nr_to_scan - shrank) >
1133                                     cli->cl_max_pages_per_rpc ?
1134                                     cli->cl_max_pages_per_rpc :
1135                                     sc->nr_to_scan - shrank, true);
1136                 if (rc > 0)
1137                         shrank += rc;
1138
1139                 if (shrank >= sc->nr_to_scan)
1140                         goto out;
1141
1142                 spin_lock(&osc_shrink_lock);
1143         }
1144         spin_unlock(&osc_shrink_lock);
1145
1146 out:
1147         cl_env_put(env, &refcheck);
1148
1149         return shrank;
1150 }
1151
1152 /** @} osc */