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_object *osc,
  90                        struct osc_page *opg, struct cl_io *io,
  91                        cl_commit_cbt cb)
  92 {
  93         int result;
  94         ENTRY;
  95
  96         osc_page_transfer_get(opg, "transfer\0cache");
  97         result = osc_queue_async_io(env, io, osc, opg, cb);
  98         if (result != 0)
  99                 osc_page_transfer_put(env, opg);
 100         else
 101                 osc_page_transfer_add(env, opg, CRT_WRITE);
 102
 103         RETURN(result);
 104 }
 105
 106 void osc_index2policy(union ldlm_policy_data *policy,
 107                       const struct cl_object *obj, pgoff_t start, pgoff_t end)
 108 {
 109         memset(policy, 0, sizeof *policy);
 110         policy->l_extent.start = start << PAGE_SHIFT;
 111         policy->l_extent.end = ((end + 1) << PAGE_SHIFT) - 1;
 112 }
 113
 114 static int osc_page_print(const struct lu_env *env,
 115                           const struct cl_page_slice *slice,
 116                           void *cookie, lu_printer_t printer)
 117 {
 118         struct osc_page *opg = cl2osc_page(slice);
 119         struct osc_async_page *oap = &opg->ops_oap;
 120         struct osc_object *obj = osc_page_object(opg);
 121         struct client_obd *cli = &osc_export(obj)->exp_obd->u.cli;
 122
 123         return (*printer)(env, cookie, LUSTRE_OSC_NAME"-page@%p %lu: "
 124                           "1< %d %c %c > "
 125                           "2< %lld %u %u %#x %#x | %p %p > "
 126                           "3< %d %d > "
 127                           "4< %d %d %d %lu %c | %c %c %c %c > "
 128                           "5< %c %c %c %c | %d %c | %d %c %c>\n",
 129                           opg, osc_index(opg),
 130                           /* 1 */
 131                           oap->oap_cmd,
 132                           list_empty_marker(&oap->oap_pending_item),
 133                           list_empty_marker(&oap->oap_rpc_item),
 134                           /* 2 */
 135                           oap->oap_obj_off, oap->oap_page_off, oap->oap_count,
 136                           oap->oap_async_flags, oap->oap_brw_flags,
 137                           cli, obj,
 138                           /* 3 */
 139                           opg->ops_transfer_pinned,
 140                           opg->ops_srvlock,
 141                           /* 4 */
 142                           cli->cl_r_in_flight, cli->cl_w_in_flight,
 143                           cli->cl_max_rpcs_in_flight,
 144                           cli->cl_avail_grant,
 145                           waitqueue_active(&cli->cl_cache_waiters) ? '+' : '-',
 146                           list_empty_marker(&cli->cl_loi_ready_list),
 147                           list_empty_marker(&cli->cl_loi_hp_ready_list),
 148                           list_empty_marker(&cli->cl_loi_write_list),
 149                           list_empty_marker(&cli->cl_loi_read_list),
 150                           /* 5 */
 151                           list_empty_marker(&obj->oo_ready_item),
 152                           list_empty_marker(&obj->oo_hp_ready_item),
 153                           list_empty_marker(&obj->oo_write_item),
 154                           list_empty_marker(&obj->oo_read_item),
 155                           atomic_read(&obj->oo_nr_reads),
 156                           list_empty_marker(&obj->oo_reading_exts),
 157                           atomic_read(&obj->oo_nr_writes),
 158                           list_empty_marker(&obj->oo_hp_exts),
 159                           list_empty_marker(&obj->oo_urgent_exts));
 160 }
 161
 162 static void osc_page_delete(const struct lu_env *env,
 163                             const struct cl_page_slice *slice)
 164 {
 165         struct osc_page   *opg = cl2osc_page(slice);
 166         struct osc_object *obj = osc_page_object(opg);
 167         void *value = NULL;
 168         int rc;
 169
 170         ENTRY;
 171         CDEBUG(D_TRACE, "%p\n", opg);
 172
 173         osc_page_transfer_put(env, opg);
 174
 175         rc = osc_teardown_async_page(env, obj, opg);
 176         if (rc) {
 177                 CL_PAGE_DEBUG(D_ERROR, env, slice->cpl_page,
 178                               "Trying to teardown failed: %d\n", rc);
 179                 LASSERT(0);
 180         }
 181
 182         osc_lru_del(osc_cli(obj), opg);
 183
 184         spin_lock(&obj->oo_tree_lock);
 185         if (opg->ops_intree) {
 186                 value = radix_tree_delete(&obj->oo_tree,
 187                                           osc_index(opg));
 188                 if (value != NULL) {
 189                         --obj->oo_npages;
 190                         opg->ops_intree = 0;
 191                 }
 192         }
 193         spin_unlock(&obj->oo_tree_lock);
 194
 195         LASSERT(ergo(value != NULL, value == opg));
 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         CDEBUG(D_CACHE, "from %d, to %d\n", from, to);
 208
 209         opg->ops_from = from;
 210         /* argument @to is exclusive, but @ops_to is inclusive */
 211         opg->ops_to   = to - 1;
 212         oap->oap_async_flags |= ASYNC_COUNT_STABLE;
 213 }
 214
 215 static int osc_page_flush(const struct lu_env *env,
 216                           const struct cl_page_slice *slice,
 217                           struct cl_io *io)
 218 {
 219         struct osc_page *opg = cl2osc_page(slice);
 220         int rc = 0;
 221         ENTRY;
 222         rc = osc_flush_async_page(env, io, opg);
 223         RETURN(rc);
 224 }
 225
 226 static void osc_page_touch(const struct lu_env *env,
 227                           const struct cl_page_slice *slice, size_t to)
 228 {
 229         struct osc_page *opg = cl2osc_page(slice);
 230         struct cl_object *obj = osc2cl(osc_page_object(opg));
 231
 232         osc_page_touch_at(env, obj, osc_index(opg), to);
 233 }
 234
 235 static const struct cl_page_operations osc_transient_page_ops = {
 236         .cpo_print         = osc_page_print,
 237         .cpo_clip           = osc_page_clip,
 238 };
 239
 240 static const struct cl_page_operations osc_page_ops = {
 241         .cpo_print         = osc_page_print,
 242         .cpo_delete        = osc_page_delete,
 243         .cpo_clip           = osc_page_clip,
 244         .cpo_flush          = osc_page_flush,
 245         .cpo_page_touch    = osc_page_touch,
 246 };
 247
 248 int osc_page_init(const struct lu_env *env, struct cl_object *obj,
 249                   struct cl_page *cl_page, pgoff_t index)
 250 {
 251         struct osc_object *osc = cl2osc(obj);
 252         struct osc_page *opg = cl_object_page_slice(obj, cl_page);
 253         struct osc_io *oio = osc_env_io(env);
 254         int result;
 255
 256         opg->ops_from = 0;
 257         opg->ops_to = PAGE_SIZE - 1;
 258
 259         INIT_LIST_HEAD(&opg->ops_lru);
 260
 261         result = osc_prep_async_page(osc, opg, cl_page, index << PAGE_SHIFT);
 262         if (result != 0)
 263                 return result;
 264
 265         opg->ops_srvlock = osc_io_srvlock(oio);
 266
 267         if (cl_page->cp_type == CPT_TRANSIENT) {
 268                 cl_page_slice_add(cl_page, &opg->ops_cl, obj,
 269                                   &osc_transient_page_ops);
 270         } else if (cl_page->cp_type == CPT_CACHEABLE) {
 271                 cl_page_slice_add(cl_page, &opg->ops_cl, obj, &osc_page_ops);
 272                 /* reserve an LRU space for this page */
 273                 result = osc_lru_alloc(env, osc_cli(osc), opg);
 274                 if (result == 0) {
 275                         result = radix_tree_preload(GFP_NOFS);
 276                         if (result == 0) {
 277                                 spin_lock(&osc->oo_tree_lock);
 278                                 result = radix_tree_insert(&osc->oo_tree,
 279                                                            index, opg);
 280                                 if (result == 0) {
 281                                         ++osc->oo_npages;
 282                                         opg->ops_intree = 1;
 283                                 }
 284                                 spin_unlock(&osc->oo_tree_lock);
 285
 286                                 radix_tree_preload_end();
 287                         }
 288                 }
 289         }
 290
 291         return result;
 292 }
 293 EXPORT_SYMBOL(osc_page_init);
 294
 295 /**
 296  * Helper function called by osc_io_submit() for every page in an immediate
 297  * transfer (i.e., transferred synchronously).
 298  */
 299 void osc_page_submit(const struct lu_env *env, struct osc_page *opg,
 300                      enum cl_req_type crt, int brw_flags)
 301 {
 302         struct osc_io *oio = osc_env_io(env);
 303         struct osc_async_page *oap = &opg->ops_oap;
 304         struct cl_page *page = opg->ops_cl.cpl_page;
 305
 306         LASSERT(oap->oap_async_flags & ASYNC_READY);
 307         LASSERT(oap->oap_async_flags & ASYNC_COUNT_STABLE);
 308
 309         oap->oap_cmd = crt == CRT_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ;
 310         oap->oap_page_off = opg->ops_from;
 311         oap->oap_count = opg->ops_to - opg->ops_from + 1;
 312         oap->oap_brw_flags = OBD_BRW_SYNC | brw_flags;
 313
 314         if (oio->oi_cap_sys_resource)
 315                 oap->oap_brw_flags |= OBD_BRW_SYS_RESOURCE;
 316
 317         if (page->cp_type != CPT_TRANSIENT)
 318                 osc_page_transfer_get(opg, "transfer\0imm");
 319         osc_page_transfer_add(env, opg, crt);
 320 }
 321
 322 /* --------------- LRU page management ------------------ */
 323
 324 /* OSC is a natural place to manage LRU pages as applications are specialized
 325  * to write OSC by OSC. Ideally, if one OSC is used more frequently it should
 326  * occupy more LRU slots. On the other hand, we should avoid using up all LRU
 327  * slots (client_obd::cl_lru_left) otherwise process has to be put into sleep
 328  * for free LRU slots - this will be very bad so the algorithm requires each
 329  * OSC to free slots voluntarily to maintain a reasonable number of free slots
 330  * at any time.
 331  */
 332
 333 static DECLARE_WAIT_QUEUE_HEAD(osc_lru_waitq);
 334
 335 /**
 336  * LRU pages are freed in batch mode. OSC should at least free this
 337  * number of pages to avoid running out of LRU slots.
 338  */
 339 static inline int lru_shrink_min(struct client_obd *cli)
 340 {
 341         return cli->cl_max_pages_per_rpc * 2;
 342 }
 343
 344 /**
 345  * free this number at most otherwise it will take too long time to finsih.
 346  */
 347 static inline int lru_shrink_max(struct client_obd *cli)
 348 {
 349         return cli->cl_max_pages_per_rpc * cli->cl_max_rpcs_in_flight;
 350 }
 351
 352 /**
 353  * Check if we can free LRU slots from this OSC. If there exists LRU waiters,
 354  * we should free slots aggressively. In this way, slots are freed in a steady
 355  * step to maintain fairness among OSCs.
 356  *
 357  * Return how many LRU pages should be freed.
 358  */
 359 static int osc_cache_too_much(struct client_obd *cli)
 360 {
 361         struct cl_client_cache *cache = cli->cl_cache;
 362         long pages = atomic_long_read(&cli->cl_lru_in_list);
 363         unsigned long budget;
 364
 365         LASSERT(cache != NULL);
 366         budget = cache->ccc_lru_max / (refcount_read(&cache->ccc_users) - 2);
 367
 368         /* if it's going to run out LRU slots, we should free some, but not
 369          * too much to maintain faireness among OSCs. */
 370         if (atomic_long_read(cli->cl_lru_left) < cache->ccc_lru_max >> 2) {
 371                 if (pages >= budget)
 372                         return lru_shrink_max(cli);
 373                 else if (pages >= budget / 2)
 374                         return lru_shrink_min(cli);
 375         } else {
 376                 time64_t duration = ktime_get_real_seconds();
 377                 long timediff;
 378
 379                 /* knock out pages by duration of no IO activity */
 380                 duration -= cli->cl_lru_last_used;
 381                 /*
 382                  * The difference shouldn't be more than 70 years
 383                  * so we can safely case to a long. Round to
 384                  * approximately 1 minute.
 385                  */
 386                 timediff = (long)(duration >> 6);
 387                 if (timediff > 0 && pages >= budget / timediff)
 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         LIST_HEAD(lru);
 417         struct osc_async_page *oap;
 418         long npages = 0;
 419
 420         list_for_each_entry(oap, plist, oap_pending_item) {
 421                 struct osc_page *opg = oap2osc_page(oap);
 422
 423                 if (!opg->ops_in_lru)
 424                         continue;
 425
 426                 ++npages;
 427                 LASSERT(list_empty(&opg->ops_lru));
 428                 list_add(&opg->ops_lru, &lru);
 429         }
 430
 431         if (npages > 0) {
 432                 spin_lock(&cli->cl_lru_list_lock);
 433                 list_splice_tail(&lru, &cli->cl_lru_list);
 434                 atomic_long_sub(npages, &cli->cl_lru_busy);
 435                 atomic_long_add(npages, &cli->cl_lru_in_list);
 436                 cli->cl_lru_last_used = ktime_get_real_seconds();
 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 LRU list 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) {
 489                 if (list_empty(&opg->ops_lru))
 490                         return;
 491                 spin_lock(&cli->cl_lru_list_lock);
 492                 if (!list_empty(&opg->ops_lru)) {
 493                         __osc_lru_del(cli, opg);
 494                         atomic_long_inc(&cli->cl_lru_busy);
 495                 }
 496                 spin_unlock(&cli->cl_lru_list_lock);
 497         }
 498 }
 499
 500 static void discard_cl_pages(const struct lu_env *env, struct cl_io *io,
 501                              struct cl_page **pvec, int max_index)
 502 {
 503         struct folio_batch *fbatch = &osc_env_info(env)->oti_fbatch;
 504         int i;
 505
 506         ll_folio_batch_init(fbatch, 0);
 507         for (i = 0; i < max_index; i++) {
 508                 struct cl_page *page = pvec[i];
 509
 510                 LASSERT(page->cp_type != CPT_TRANSIENT);
 511                 LASSERT(cl_page_is_owned(page, io));
 512                 cl_page_discard(env, io, page);
 513                 cl_page_disown(env, io, page);
 514                 cl_batch_put(env, page, fbatch);
 515
 516                 pvec[i] = NULL;
 517         }
 518         folio_batch_release(fbatch);
 519 }
 520
 521 /**
 522  * Check if a cl_page can be released, i.e, it's not being used.
 523  *
 524  * If unstable account is turned on, bulk transfer may hold one refcount
 525  * for recovery so we need to check vmpage refcount as well; otherwise,
 526  * even we can destroy cl_page but the corresponding vmpage can't be reused.
 527  */
 528 static inline bool lru_page_busy(struct client_obd *cli, struct cl_page *page)
 529 {
 530         if (cl_page_in_use_noref(page))
 531                 return true;
 532
 533         if (cli->cl_cache->ccc_unstable_check) {
 534                 struct page *vmpage = cl_page_vmpage(page);
 535
 536                 /* vmpage have two known users: cl_page and VM page cache */
 537                 if (page_count(vmpage) - page_mapcount(vmpage) > 2)
 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_cl_pages(env, io, pvec, index);
 611                                 index = 0;
 612
 613                                 cl_io_fini(env, io);
 614                                 cl_object_put(env, clobj);
 615                                 clobj = NULL;
 616                                 cond_resched();
 617                         }
 618
 619                         clobj = tmp;
 620                         io->ci_obj = clobj;
 621                         io->ci_ignore_layout = 1;
 622                         rc = cl_io_init(env, io, CIT_MISC, clobj);
 623
 624                         spin_lock(&cli->cl_lru_list_lock);
 625
 626                         if (rc != 0)
 627                                 break;
 628
 629                         ++maxscan;
 630                         continue;
 631                 }
 632
 633                 if (cl_page_own_try(env, io, page) == 0) {
 634                         if (!lru_page_busy(cli, page)) {
 635                                 /* remove it from lru list earlier to avoid
 636                                  * lock contention */
 637                                 __osc_lru_del(cli, opg);
 638                                 opg->ops_in_lru = 0; /* will be discarded */
 639
 640                                 cl_page_get(page);
 641                                 will_free = true;
 642                         } else {
 643                                 cl_page_disown(env, io, page);
 644                         }
 645                 }
 646
 647                 if (!will_free) {
 648                         list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
 649                         continue;
 650                 }
 651
 652                 /* Don't discard and free the page with cl_lru_list held */
 653                 pvec[index++] = page;
 654                 if (unlikely(index == OTI_PVEC_SIZE)) {
 655                         spin_unlock(&cli->cl_lru_list_lock);
 656                         discard_cl_pages(env, io, pvec, index);
 657                         index = 0;
 658
 659                         spin_lock(&cli->cl_lru_list_lock);
 660                 }
 661
 662                 if (++count >= target)
 663                         break;
 664         }
 665         spin_unlock(&cli->cl_lru_list_lock);
 666
 667         if (clobj != NULL) {
 668                 discard_cl_pages(env, io, pvec, index);
 669
 670                 cl_io_fini(env, io);
 671                 cl_object_put(env, clobj);
 672                 cond_resched();
 673         }
 674
 675         atomic_dec(&cli->cl_lru_shrinkers);
 676         if (count > 0) {
 677                 atomic_long_add(count, cli->cl_lru_left);
 678                 wake_up(&osc_lru_waitq);
 679         }
 680         RETURN(count > 0 ? count : rc);
 681 }
 682 EXPORT_SYMBOL(osc_lru_shrink);
 683
 684 /**
 685  * Reclaim LRU pages by an IO thread. The caller wants to reclaim at least
 686  * \@npages of LRU slots. For performance consideration, it's better to drop
 687  * LRU pages in batch. Therefore, the actual number is adjusted at least
 688  * max_pages_per_rpc.
 689  */
 690 static long osc_lru_reclaim(struct client_obd *cli, unsigned long npages)
 691 {
 692         struct lu_env *env;
 693         struct cl_client_cache *cache = cli->cl_cache;
 694         struct client_obd *scan;
 695         int max_scans;
 696         __u16 refcheck;
 697         long rc = 0;
 698         ENTRY;
 699
 700         LASSERT(cache != NULL);
 701
 702         env = cl_env_get(&refcheck);
 703         if (IS_ERR(env))
 704                 RETURN(rc);
 705
 706         npages = max_t(int, npages, cli->cl_max_pages_per_rpc);
 707         CDEBUG(D_CACHE, "%s: start to reclaim %ld pages from LRU\n",
 708                cli_name(cli), npages);
 709         rc = osc_lru_shrink(env, cli, npages, true);
 710         if (rc >= npages) {
 711                 CDEBUG(D_CACHE, "%s: reclaimed %ld/%ld pages from LRU\n",
 712                        cli_name(cli), rc, npages);
 713                 if (osc_cache_too_much(cli) > 0)
 714                         ptlrpcd_queue_work(cli->cl_lru_work);
 715                 GOTO(out, rc);
 716         } else if (rc > 0) {
 717                 npages -= rc;
 718         }
 719
 720         CDEBUG(D_CACHE, "%s: cli %p no free slots, pages: %ld/%ld, want: %ld\n",
 721                 cli_name(cli), cli, atomic_long_read(&cli->cl_lru_in_list),
 722                 atomic_long_read(&cli->cl_lru_busy), npages);
 723
 724         /* Reclaim LRU slots from other client_obd as it can't free enough
 725          * from its own. This should rarely happen. */
 726         spin_lock(&cache->ccc_lru_lock);
 727         LASSERT(!list_empty(&cache->ccc_lru));
 728
 729         cache->ccc_lru_shrinkers++;
 730         list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
 731
 732         max_scans = refcount_read(&cache->ccc_users) - 2;
 733         while (--max_scans > 0 &&
 734                (scan = list_first_entry_or_null(&cache->ccc_lru,
 735                                                   struct client_obd,
 736                                                   cl_lru_osc)) != NULL) {
 737                 CDEBUG(D_CACHE, "%s: cli %p LRU pages: %ld, busy: %ld.\n",
 738                        cli_name(scan), scan,
 739                        atomic_long_read(&scan->cl_lru_in_list),
 740                        atomic_long_read(&scan->cl_lru_busy));
 741
 742                 list_move_tail(&scan->cl_lru_osc, &cache->ccc_lru);
 743                 if (osc_cache_too_much(scan) > 0) {
 744                         spin_unlock(&cache->ccc_lru_lock);
 745
 746                         rc = osc_lru_shrink(env, scan, npages, true);
 747                         spin_lock(&cache->ccc_lru_lock);
 748                         if (rc >= npages)
 749                                 break;
 750                         if (rc > 0)
 751                                 npages -= rc;
 752                 }
 753         }
 754         spin_unlock(&cache->ccc_lru_lock);
 755
 756 out:
 757         cl_env_put(env, &refcheck);
 758         CDEBUG(D_CACHE, "%s: cli %p freed %ld pages.\n",
 759                cli_name(cli), cli, rc);
 760         return rc;
 761 }
 762
 763 /**
 764  * osc_lru_alloc() is called to allocate an LRU slot for a cl_page.
 765  *
 766  * Usually the LRU slots are reserved in osc_io_iter_rw_init().
 767  * Only in the case that the LRU slots are in extreme shortage, it should
 768  * have reserved enough slots for an IO.
 769  */
 770 static int osc_lru_alloc(const struct lu_env *env, struct client_obd *cli,
 771                          struct osc_page *opg)
 772 {
 773         struct osc_io *oio = osc_env_io(env);
 774         int rc = 0;
 775
 776         ENTRY;
 777
 778         if (cli->cl_cache == NULL) /* shall not be in LRU */
 779                 RETURN(0);
 780
 781         if (oio->oi_lru_reserved > 0) {
 782                 --oio->oi_lru_reserved;
 783                 goto out;
 784         }
 785
 786         LASSERT(atomic_long_read(cli->cl_lru_left) >= 0);
 787         while (!atomic_long_add_unless(cli->cl_lru_left, -1, 0)) {
 788                 /* run out of LRU spaces, try to drop some by itself */
 789                 rc = osc_lru_reclaim(cli, 1);
 790                 if (rc < 0)
 791                         break;
 792                 if (rc > 0)
 793                         continue;
 794                 /* IO issued by readahead, don't try hard */
 795                 if (oio->oi_is_readahead) {
 796                         if (atomic_long_read(cli->cl_lru_left) > 0)
 797                                 continue;
 798                         rc = -EBUSY;
 799                         break;
 800                 }
 801
 802                 cond_resched();
 803                 rc = l_wait_event_abortable(
 804                         osc_lru_waitq,
 805                         atomic_long_read(cli->cl_lru_left) > 0);
 806                 if (rc < 0) {
 807                         rc = -EINTR;
 808                         break;
 809                 }
 810         }
 811
 812 out:
 813         if (rc >= 0) {
 814                 atomic_long_inc(&cli->cl_lru_busy);
 815                 opg->ops_in_lru = 1;
 816                 rc = 0;
 817         }
 818
 819         RETURN(rc);
 820 }
 821
 822 /**
 823  * osc_lru_reserve() is called to reserve enough LRU slots for I/O.
 824  *
 825  * The benefit of doing this is to reduce contention against atomic counter
 826  * cl_lru_left by changing it from per-page access to per-IO access.
 827  */
 828 unsigned long osc_lru_reserve(struct client_obd *cli, unsigned long npages)
 829 {
 830         unsigned long reserved = 0;
 831         unsigned long max_pages;
 832         unsigned long c;
 833         int rc;
 834
 835 again:
 836         c = atomic_long_read(cli->cl_lru_left);
 837         if (c < npages && osc_lru_reclaim(cli, npages) > 0)
 838                 c = atomic_long_read(cli->cl_lru_left);
 839
 840         if (c < npages) {
 841                 /*
 842                  * Trigger writeback in the hope some LRU slot could
 843                  * be freed.
 844                  */
 845                 rc = ptlrpcd_queue_work(cli->cl_writeback_work);
 846                 if (rc)
 847                         return 0;
 848         }
 849
 850         while (c >= npages) {
 851                 if (c == atomic_long_cmpxchg(cli->cl_lru_left, c, c - npages)) {
 852                         reserved = npages;
 853                         break;
 854                 }
 855                 c = atomic_long_read(cli->cl_lru_left);
 856         }
 857
 858         if (reserved != npages) {
 859                 cond_resched();
 860                 rc = l_wait_event_abortable(
 861                         osc_lru_waitq,
 862                         atomic_long_read(cli->cl_lru_left) > 0);
 863                 goto again;
 864         }
 865
 866         max_pages = cli->cl_max_pages_per_rpc * cli->cl_max_rpcs_in_flight;
 867         if (atomic_long_read(cli->cl_lru_left) < max_pages) {
 868                 /* If there aren't enough pages in the per-OSC LRU then
 869                  * wake up the LRU thread to try and clear out space, so
 870                  * we don't block if pages are being dirtied quickly. */
 871                 CDEBUG(D_CACHE, "%s: queue LRU, left: %lu/%ld.\n",
 872                        cli_name(cli), atomic_long_read(cli->cl_lru_left),
 873                        max_pages);
 874                 (void)ptlrpcd_queue_work(cli->cl_lru_work);
 875         }
 876
 877         return reserved;
 878 }
 879
 880 /**
 881  * osc_lru_unreserve() is called to unreserve LRU slots.
 882  *
 883  * LRU slots reserved by osc_lru_reserve() may have entries left due to several
 884  * reasons such as page already existing or I/O error. Those reserved slots
 885  * should be freed by calling this function.
 886  */
 887 void osc_lru_unreserve(struct client_obd *cli, unsigned long npages)
 888 {
 889         atomic_long_add(npages, cli->cl_lru_left);
 890         wake_up(&osc_lru_waitq);
 891 }
 892
 893 /**
 894  * Atomic operations are expensive. We accumulate the accounting for the
 895  * same page zone to get better performance.
 896  * In practice this can work pretty good because the pages in the same RPC
 897  * are likely from the same page zone.
 898  */
 899 #ifdef HAVE_NR_UNSTABLE_NFS
 900 /* Old kernels use a separate counter for unstable pages,
 901  * newer kernels treat them like any other writeback.
 902  * (see Linux commit: v5.7-467-g8d92890bd6b8)
 903  */
 904 #define NR_ZONE_WRITE_PENDING           ((enum zone_stat_item)NR_UNSTABLE_NFS)
 905 #elif !defined(HAVE_NR_ZONE_WRITE_PENDING)
 906 #define NR_ZONE_WRITE_PENDING           ((enum zone_stat_item)NR_WRITEBACK)
 907 #endif
 908
 909 static inline void unstable_page_accounting(struct ptlrpc_bulk_desc *desc,
 910                                             int factor)
 911 {
 912         int page_count;
 913         void *zone = NULL;
 914         int count = 0;
 915         int i;
 916
 917         ENTRY;
 918
 919         page_count = desc->bd_iov_count;
 920
 921         CDEBUG(D_PAGE, "%s %d unstable pages\n",
 922                factor == 1 ? "adding" : "removing", page_count);
 923
 924         for (i = 0; i < page_count; i++) {
 925                 void *pz = page_zone(desc->bd_vec[i].bv_page);
 926
 927                 if (likely(pz == zone)) {
 928                         ++count;
 929                         continue;
 930                 }
 931
 932                 if (count > 0) {
 933                         mod_zone_page_state(zone, NR_ZONE_WRITE_PENDING,
 934                                             factor * count);
 935                         count = 0;
 936                 }
 937                 zone = pz;
 938                 ++count;
 939         }
 940         if (count > 0)
 941                 mod_zone_page_state(zone, NR_ZONE_WRITE_PENDING,
 942                                     factor * count);
 943
 944         EXIT;
 945 }
 946
 947 static inline void add_unstable_pages(struct ptlrpc_bulk_desc *desc)
 948 {
 949         unstable_page_accounting(desc, 1);
 950 }
 951
 952 static inline void dec_unstable_pages(struct ptlrpc_bulk_desc *desc)
 953 {
 954         unstable_page_accounting(desc, -1);
 955 }
 956
 957 /**
 958  * Performs "unstable" page accounting. This function balances the
 959  * increment operations performed in osc_inc_unstable_pages. It is
 960  * registered as the RPC request callback, and is executed when the
 961  * bulk RPC is committed on the server. Thus at this point, the pages
 962  * involved in the bulk transfer are no longer considered unstable.
 963  *
 964  * If this function is called, the request should have been committed
 965  * or req:rq_unstable must have been set; it implies that the unstable
 966  * statistic have been added.
 967  */
 968 void osc_dec_unstable_pages(struct ptlrpc_request *req)
 969 {
 970         struct ptlrpc_bulk_desc *desc       = req->rq_bulk;
 971         struct client_obd       *cli        = &req->rq_import->imp_obd->u.cli;
 972         int                      page_count;
 973         long                     unstable_count;
 974
 975         /* no desc means short io, which doesn't have separate unstable pages,
 976          * it's just using space inside the RPC itself
 977          */
 978         if (!desc)
 979                 return;
 980
 981         page_count = desc->bd_iov_count;
 982
 983         LASSERT(page_count >= 0);
 984
 985         dec_unstable_pages(desc);
 986
 987         unstable_count = atomic_long_sub_return(page_count,
 988                                                 &cli->cl_unstable_count);
 989         LASSERT(unstable_count >= 0);
 990
 991         unstable_count = atomic_long_sub_return(page_count,
 992                                            &cli->cl_cache->ccc_unstable_nr);
 993         LASSERT(unstable_count >= 0);
 994
 995         if (waitqueue_active(&osc_lru_waitq))
 996                 (void)ptlrpcd_queue_work(cli->cl_lru_work);
 997 }
 998
 999 /**
1000  * "unstable" page accounting. See: osc_dec_unstable_pages.
1001  */
1002 void osc_inc_unstable_pages(struct ptlrpc_request *req)
1003 {
1004         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
1005         struct client_obd       *cli  = &req->rq_import->imp_obd->u.cli;
1006         long                     page_count;
1007
1008         /* No unstable page tracking */
1009         if (cli->cl_cache == NULL || !cli->cl_cache->ccc_unstable_check)
1010                 return;
1011
1012         /* no desc means short io, which doesn't have separate unstable pages,
1013          * it's just using space inside the RPC itself
1014          */
1015         if (!desc)
1016                 return;
1017
1018         page_count = desc->bd_iov_count;
1019
1020         add_unstable_pages(desc);
1021         atomic_long_add(page_count, &cli->cl_unstable_count);
1022         atomic_long_add(page_count, &cli->cl_cache->ccc_unstable_nr);
1023
1024         /* If the request has already been committed (i.e. brw_commit
1025          * called via rq_commit_cb), we need to undo the unstable page
1026          * increments we just performed because rq_commit_cb wont be
1027          * called again. */
1028         spin_lock(&req->rq_lock);
1029         if (unlikely(req->rq_committed)) {
1030                 spin_unlock(&req->rq_lock);
1031
1032                 osc_dec_unstable_pages(req);
1033         } else {
1034                 req->rq_unstable = 1;
1035                 spin_unlock(&req->rq_lock);
1036         }
1037 }
1038
1039 /**
1040  * Check if it piggybacks SOFT_SYNC flag to OST from this OSC.
1041  * This function will be called by every BRW RPC so it's critical
1042  * to make this function fast.
1043  */
1044 bool osc_over_unstable_soft_limit(struct client_obd *cli)
1045 {
1046         long unstable_nr, osc_unstable_count;
1047
1048         /* Can't check cli->cl_unstable_count, therefore, no soft limit */
1049         if (cli->cl_cache == NULL || !cli->cl_cache->ccc_unstable_check)
1050                 return false;
1051
1052         osc_unstable_count = atomic_long_read(&cli->cl_unstable_count);
1053         unstable_nr = atomic_long_read(&cli->cl_cache->ccc_unstable_nr);
1054
1055         CDEBUG(D_CACHE,
1056                "%s: cli: %p unstable pages: %lu, osc unstable pages: %lu\n",
1057                cli_name(cli), cli, unstable_nr, osc_unstable_count);
1058
1059         /* If the LRU slots are in shortage - 25% remaining AND this OSC
1060          * has one full RPC window of unstable pages, it's a good chance
1061          * to piggyback a SOFT_SYNC flag.
1062          * Please notice that the OST won't take immediate response for the
1063          * SOFT_SYNC request so active OSCs will have more chance to carry
1064          * the flag, this is reasonable. */
1065         return unstable_nr > cli->cl_cache->ccc_lru_max >> 2 &&
1066                osc_unstable_count > cli->cl_max_pages_per_rpc *
1067                                     cli->cl_max_rpcs_in_flight;
1068 }
1069
1070 /**
1071  * Return how many LRU pages in the cache of all OSC devices
1072  *
1073  * \retval      return # of cached LRU pages times reclaimation tendency
1074  * \retval      SHRINK_STOP if it cannot do any scanning in this time
1075  */
1076 unsigned long osc_cache_shrink_count(struct shrinker *sk,
1077                                      struct shrink_control *sc)
1078 {
1079         struct client_obd *cli;
1080         unsigned long cached = 0;
1081
1082         if (!osc_page_cache_shrink_enabled)
1083                 return 0;
1084
1085         spin_lock(&osc_shrink_lock);
1086         list_for_each_entry(cli, &osc_shrink_list, cl_shrink_list)
1087                 cached += atomic_long_read(&cli->cl_lru_in_list);
1088         spin_unlock(&osc_shrink_lock);
1089
1090         return (cached  * sysctl_vfs_cache_pressure) / 100;
1091 }
1092
1093 /**
1094  * Scan and try to reclaim sc->nr_to_scan cached LRU pages
1095  *
1096  * \retval      number of cached LRU pages reclaimed
1097  * \retval      SHRINK_STOP if it cannot do any scanning in this time
1098  *
1099  * Linux kernel will loop calling this shrinker scan routine with
1100  * sc->nr_to_scan = SHRINK_BATCH(128 for now) until kernel got enough memory.
1101  *
1102  * If sc->nr_to_scan is 0, the VM is querying the cache size, we don't need
1103  * to scan and try to reclaim LRU pages, just return 0 and
1104  * osc_cache_shrink_count() will report the LRU page number.
1105  */
1106 unsigned long osc_cache_shrink_scan(struct shrinker *sk,
1107                                     struct shrink_control *sc)
1108 {
1109         struct client_obd *cli;
1110         struct client_obd *stop_anchor = NULL;
1111         struct lu_env *env;
1112         long shrank = 0;
1113         int rc;
1114         __u16 refcheck;
1115
1116         if (sc->nr_to_scan == 0)
1117                 return 0;
1118
1119         if (!(sc->gfp_mask & __GFP_FS))
1120                 return SHRINK_STOP;
1121
1122         env = cl_env_get(&refcheck);
1123         if (IS_ERR(env))
1124                 return SHRINK_STOP;
1125
1126         spin_lock(&osc_shrink_lock);
1127         while ((cli = list_first_entry_or_null(&osc_shrink_list,
1128                                                struct client_obd,
1129                                                cl_shrink_list)) != NULL) {
1130                 if (stop_anchor == NULL)
1131                         stop_anchor = cli;
1132                 else if (cli == stop_anchor)
1133                         break;
1134
1135                 list_move_tail(&cli->cl_shrink_list, &osc_shrink_list);
1136                 spin_unlock(&osc_shrink_lock);
1137
1138                 /* shrink no more than max_pages_per_rpc for an OSC */
1139                 rc = osc_lru_shrink(env, cli, (sc->nr_to_scan - shrank) >
1140                                     cli->cl_max_pages_per_rpc ?
1141                                     cli->cl_max_pages_per_rpc :
1142                                     sc->nr_to_scan - shrank, true);
1143                 if (rc > 0)
1144                         shrank += rc;
1145
1146                 if (shrank >= sc->nr_to_scan)
1147                         goto out;
1148
1149                 spin_lock(&osc_shrink_lock);
1150         }
1151         spin_unlock(&osc_shrink_lock);
1152
1153 out:
1154         cl_env_put(env, &refcheck);
1155
1156         return shrank;
1157 }
1158
1159 /** @} osc */