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         int rc;
 168
 169         ENTRY;
 170         CDEBUG(D_TRACE, "%p\n", opg);
 171         osc_page_transfer_put(env, opg);
 172         rc = osc_teardown_async_page(env, obj, opg);
 173         if (rc) {
 174                 CL_PAGE_DEBUG(D_ERROR, env, slice->cpl_page,
 175                               "Trying to teardown failed: %d\n", rc);
 176                 LASSERT(0);
 177         }
 178
 179         osc_lru_del(osc_cli(obj), opg);
 180
 181         if (slice->cpl_page->cp_type == CPT_CACHEABLE) {
 182                 void *value = NULL;
 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
 198         EXIT;
 199 }
 200
 201 static void osc_page_clip(const struct lu_env *env,
 202                           const struct cl_page_slice *slice,
 203                           int from, int to)
 204 {
 205         struct osc_page       *opg = cl2osc_page(slice);
 206         struct osc_async_page *oap = &opg->ops_oap;
 207
 208         CDEBUG(D_CACHE, "from %d, to %d\n", from, to);
 209
 210         opg->ops_from = from;
 211         /* argument @to is exclusive, but @ops_to is inclusive */
 212         opg->ops_to   = to - 1;
 213         oap->oap_async_flags |= ASYNC_COUNT_STABLE;
 214 }
 215
 216 static int osc_page_flush(const struct lu_env *env,
 217                           const struct cl_page_slice *slice,
 218                           struct cl_io *io)
 219 {
 220         struct osc_page *opg = cl2osc_page(slice);
 221         int rc = 0;
 222         ENTRY;
 223         rc = osc_flush_async_page(env, io, opg);
 224         RETURN(rc);
 225 }
 226
 227 static void osc_page_touch(const struct lu_env *env,
 228                           const struct cl_page_slice *slice, size_t to)
 229 {
 230         struct osc_page *opg = cl2osc_page(slice);
 231         struct cl_object *obj = osc2cl(osc_page_object(opg));
 232
 233         osc_page_touch_at(env, obj, osc_index(opg), to);
 234 }
 235
 236 static const struct cl_page_operations osc_page_ops = {
 237         .cpo_print         = osc_page_print,
 238         .cpo_delete        = osc_page_delete,
 239         .cpo_clip           = osc_page_clip,
 240         .cpo_flush          = osc_page_flush,
 241         .cpo_page_touch    = osc_page_touch,
 242 };
 243
 244 int osc_page_init(const struct lu_env *env, struct cl_object *obj,
 245                   struct cl_page *cl_page, pgoff_t index)
 246 {
 247         struct osc_object *osc = cl2osc(obj);
 248         struct osc_page *opg = cl_object_page_slice(obj, cl_page);
 249         struct osc_io *oio = osc_env_io(env);
 250         int result;
 251
 252         opg->ops_from = 0;
 253         opg->ops_to = PAGE_SIZE - 1;
 254
 255         INIT_LIST_HEAD(&opg->ops_lru);
 256
 257         result = osc_prep_async_page(osc, opg, cl_page, index << PAGE_SHIFT);
 258         if (result != 0)
 259                 return result;
 260
 261         opg->ops_srvlock = osc_io_srvlock(oio);
 262         cl_page_slice_add(cl_page, &opg->ops_cl, obj, &osc_page_ops);
 263
 264         /* reserve an LRU space for this page */
 265         if (cl_page->cp_type == CPT_CACHEABLE) {
 266                 result = osc_lru_alloc(env, osc_cli(osc), opg);
 267                 if (result == 0) {
 268                         result = radix_tree_preload(GFP_NOFS);
 269                         if (result == 0) {
 270                                 spin_lock(&osc->oo_tree_lock);
 271                                 result = radix_tree_insert(&osc->oo_tree,
 272                                                            index, opg);
 273                                 if (result == 0) {
 274                                         ++osc->oo_npages;
 275                                         opg->ops_intree = 1;
 276                                 }
 277                                 spin_unlock(&osc->oo_tree_lock);
 278
 279                                 radix_tree_preload_end();
 280                         }
 281                 }
 282         }
 283
 284         return result;
 285 }
 286 EXPORT_SYMBOL(osc_page_init);
 287
 288 /**
 289  * Helper function called by osc_io_submit() for every page in an immediate
 290  * transfer (i.e., transferred synchronously).
 291  */
 292 void osc_page_submit(const struct lu_env *env, struct osc_page *opg,
 293                      enum cl_req_type crt, int brw_flags)
 294 {
 295         struct osc_io *oio = osc_env_io(env);
 296         struct osc_async_page *oap = &opg->ops_oap;
 297
 298         LASSERT(oap->oap_async_flags & ASYNC_READY);
 299         LASSERT(oap->oap_async_flags & ASYNC_COUNT_STABLE);
 300
 301         oap->oap_cmd = crt == CRT_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ;
 302         oap->oap_page_off = opg->ops_from;
 303         oap->oap_count = opg->ops_to - opg->ops_from + 1;
 304         oap->oap_brw_flags = OBD_BRW_SYNC | brw_flags;
 305
 306         if (oio->oi_cap_sys_resource)
 307                 oap->oap_brw_flags |= OBD_BRW_SYS_RESOURCE;
 308
 309         osc_page_transfer_get(opg, "transfer\0imm");
 310         osc_page_transfer_add(env, opg, crt);
 311 }
 312
 313 /* --------------- LRU page management ------------------ */
 314
 315 /* OSC is a natural place to manage LRU pages as applications are specialized
 316  * to write OSC by OSC. Ideally, if one OSC is used more frequently it should
 317  * occupy more LRU slots. On the other hand, we should avoid using up all LRU
 318  * slots (client_obd::cl_lru_left) otherwise process has to be put into sleep
 319  * for free LRU slots - this will be very bad so the algorithm requires each
 320  * OSC to free slots voluntarily to maintain a reasonable number of free slots
 321  * at any time.
 322  */
 323
 324 static DECLARE_WAIT_QUEUE_HEAD(osc_lru_waitq);
 325
 326 /**
 327  * LRU pages are freed in batch mode. OSC should at least free this
 328  * number of pages to avoid running out of LRU slots.
 329  */
 330 static inline int lru_shrink_min(struct client_obd *cli)
 331 {
 332         return cli->cl_max_pages_per_rpc * 2;
 333 }
 334
 335 /**
 336  * free this number at most otherwise it will take too long time to finsih.
 337  */
 338 static inline int lru_shrink_max(struct client_obd *cli)
 339 {
 340         return cli->cl_max_pages_per_rpc * cli->cl_max_rpcs_in_flight;
 341 }
 342
 343 /**
 344  * Check if we can free LRU slots from this OSC. If there exists LRU waiters,
 345  * we should free slots aggressively. In this way, slots are freed in a steady
 346  * step to maintain fairness among OSCs.
 347  *
 348  * Return how many LRU pages should be freed.
 349  */
 350 static int osc_cache_too_much(struct client_obd *cli)
 351 {
 352         struct cl_client_cache *cache = cli->cl_cache;
 353         long pages = atomic_long_read(&cli->cl_lru_in_list);
 354         unsigned long budget;
 355
 356         LASSERT(cache != NULL);
 357         budget = cache->ccc_lru_max / (refcount_read(&cache->ccc_users) - 2);
 358
 359         /* if it's going to run out LRU slots, we should free some, but not
 360          * too much to maintain faireness among OSCs. */
 361         if (atomic_long_read(cli->cl_lru_left) < cache->ccc_lru_max >> 2) {
 362                 if (pages >= budget)
 363                         return lru_shrink_max(cli);
 364                 else if (pages >= budget / 2)
 365                         return lru_shrink_min(cli);
 366         } else {
 367                 time64_t duration = ktime_get_real_seconds();
 368                 long timediff;
 369
 370                 /* knock out pages by duration of no IO activity */
 371                 duration -= cli->cl_lru_last_used;
 372                 /*
 373                  * The difference shouldn't be more than 70 years
 374                  * so we can safely case to a long. Round to
 375                  * approximately 1 minute.
 376                  */
 377                 timediff = (long)(duration >> 6);
 378                 if (timediff > 0 && pages >= budget / timediff)
 379                         return lru_shrink_min(cli);
 380         }
 381         return 0;
 382 }
 383
 384 int lru_queue_work(const struct lu_env *env, void *data)
 385 {
 386         struct client_obd *cli = data;
 387         int count;
 388
 389         CDEBUG(D_CACHE, "%s: run LRU work for client obd\n", cli_name(cli));
 390         count = osc_cache_too_much(cli);
 391         if (count > 0) {
 392                 int rc = osc_lru_shrink(env, cli, count, false);
 393
 394                 CDEBUG(D_CACHE, "%s: shrank %d/%d pages from client obd\n",
 395                        cli_name(cli), rc, count);
 396                 if (rc >= count) {
 397                         CDEBUG(D_CACHE, "%s: queue again\n", cli_name(cli));
 398                         ptlrpcd_queue_work(cli->cl_lru_work);
 399                 }
 400         }
 401
 402         RETURN(0);
 403 }
 404
 405 void osc_lru_add_batch(struct client_obd *cli, struct list_head *plist)
 406 {
 407         LIST_HEAD(lru);
 408         struct osc_async_page *oap;
 409         long npages = 0;
 410
 411         list_for_each_entry(oap, plist, oap_pending_item) {
 412                 struct osc_page *opg = oap2osc_page(oap);
 413
 414                 if (!opg->ops_in_lru)
 415                         continue;
 416
 417                 ++npages;
 418                 LASSERT(list_empty(&opg->ops_lru));
 419                 list_add(&opg->ops_lru, &lru);
 420         }
 421
 422         if (npages > 0) {
 423                 spin_lock(&cli->cl_lru_list_lock);
 424                 list_splice_tail(&lru, &cli->cl_lru_list);
 425                 atomic_long_sub(npages, &cli->cl_lru_busy);
 426                 atomic_long_add(npages, &cli->cl_lru_in_list);
 427                 cli->cl_lru_last_used = ktime_get_real_seconds();
 428                 spin_unlock(&cli->cl_lru_list_lock);
 429
 430                 if (waitqueue_active(&osc_lru_waitq))
 431                         (void)ptlrpcd_queue_work(cli->cl_lru_work);
 432         }
 433 }
 434
 435 static void __osc_lru_del(struct client_obd *cli, struct osc_page *opg)
 436 {
 437         LASSERT(atomic_long_read(&cli->cl_lru_in_list) > 0);
 438         list_del_init(&opg->ops_lru);
 439         atomic_long_dec(&cli->cl_lru_in_list);
 440 }
 441
 442 /**
 443  * Page is being destroyed. The page may be not in LRU list, if the transfer
 444  * has never finished(error occurred).
 445  */
 446 static void osc_lru_del(struct client_obd *cli, struct osc_page *opg)
 447 {
 448         if (opg->ops_in_lru) {
 449                 spin_lock(&cli->cl_lru_list_lock);
 450                 if (!list_empty(&opg->ops_lru)) {
 451                         __osc_lru_del(cli, opg);
 452                 } else {
 453                         LASSERT(atomic_long_read(&cli->cl_lru_busy) > 0);
 454                         atomic_long_dec(&cli->cl_lru_busy);
 455                 }
 456                 spin_unlock(&cli->cl_lru_list_lock);
 457
 458                 atomic_long_inc(cli->cl_lru_left);
 459                 /* this is a great place to release more LRU pages if
 460                  * this osc occupies too many LRU pages and kernel is
 461                  * stealing one of them. */
 462                 if (osc_cache_too_much(cli)) {
 463                         CDEBUG(D_CACHE, "%s: queue LRU work\n", cli_name(cli));
 464                         (void)ptlrpcd_queue_work(cli->cl_lru_work);
 465                 }
 466                 wake_up(&osc_lru_waitq);
 467         } else {
 468                 LASSERT(list_empty(&opg->ops_lru));
 469         }
 470 }
 471
 472 /**
 473  * Delete page from LRU list for redirty.
 474  */
 475 static void osc_lru_use(struct client_obd *cli, struct osc_page *opg)
 476 {
 477         /* If page is being transferred for the first time,
 478          * ops_lru should be empty */
 479         if (opg->ops_in_lru) {
 480                 if (list_empty(&opg->ops_lru))
 481                         return;
 482                 spin_lock(&cli->cl_lru_list_lock);
 483                 if (!list_empty(&opg->ops_lru)) {
 484                         __osc_lru_del(cli, opg);
 485                         atomic_long_inc(&cli->cl_lru_busy);
 486                 }
 487                 spin_unlock(&cli->cl_lru_list_lock);
 488         }
 489 }
 490
 491 static void discard_cl_pages(const struct lu_env *env, struct cl_io *io,
 492                              struct cl_page **pvec, int max_index)
 493 {
 494         struct folio_batch *fbatch = &osc_env_info(env)->oti_fbatch;
 495         int i;
 496
 497         ll_folio_batch_init(fbatch, 0);
 498         for (i = 0; i < max_index; i++) {
 499                 struct cl_page *page = pvec[i];
 500
 501                 LASSERT(page->cp_type != CPT_TRANSIENT);
 502                 LASSERT(cl_page_is_owned(page, io));
 503                 cl_page_discard(env, io, page);
 504                 cl_page_disown(env, io, page);
 505                 cl_batch_put(env, page, fbatch);
 506
 507                 pvec[i] = NULL;
 508         }
 509         folio_batch_release(fbatch);
 510 }
 511
 512 /**
 513  * Check if a cl_page can be released, i.e, it's not being used.
 514  *
 515  * If unstable account is turned on, bulk transfer may hold one refcount
 516  * for recovery so we need to check vmpage refcount as well; otherwise,
 517  * even we can destroy cl_page but the corresponding vmpage can't be reused.
 518  */
 519 static inline bool lru_page_busy(struct client_obd *cli, struct cl_page *page)
 520 {
 521         if (cl_page_in_use_noref(page))
 522                 return true;
 523
 524         if (cli->cl_cache->ccc_unstable_check) {
 525                 struct page *vmpage = cl_page_vmpage(page);
 526
 527                 /* vmpage have two known users: cl_page and VM page cache */
 528                 if (page_count(vmpage) - page_mapcount(vmpage) > 2)
 529                         return true;
 530         }
 531         return false;
 532 }
 533
 534 /**
 535  * Drop @target of pages from LRU at most.
 536  */
 537 long osc_lru_shrink(const struct lu_env *env, struct client_obd *cli,
 538                    long target, bool force)
 539 {
 540         struct cl_io *io;
 541         struct cl_object *clobj = NULL;
 542         struct cl_page **pvec;
 543         struct osc_page *opg;
 544         long count = 0;
 545         int maxscan = 0;
 546         int index = 0;
 547         int rc = 0;
 548         ENTRY;
 549
 550         LASSERT(atomic_long_read(&cli->cl_lru_in_list) >= 0);
 551         if (atomic_long_read(&cli->cl_lru_in_list) == 0 || target <= 0)
 552                 RETURN(0);
 553
 554         CDEBUG(D_CACHE, "%s: shrinkers: %d, force: %d\n",
 555                cli_name(cli), atomic_read(&cli->cl_lru_shrinkers), force);
 556         if (!force) {
 557                 if (atomic_read(&cli->cl_lru_shrinkers) > 0)
 558                         RETURN(-EBUSY);
 559
 560                 if (atomic_inc_return(&cli->cl_lru_shrinkers) > 1) {
 561                         atomic_dec(&cli->cl_lru_shrinkers);
 562                         RETURN(-EBUSY);
 563                 }
 564         } else {
 565                 atomic_inc(&cli->cl_lru_shrinkers);
 566         }
 567
 568         pvec = (struct cl_page **)osc_env_info(env)->oti_pvec;
 569         io = osc_env_thread_io(env);
 570
 571         spin_lock(&cli->cl_lru_list_lock);
 572         if (force)
 573                 cli->cl_lru_reclaim++;
 574         maxscan = min(target << 1, atomic_long_read(&cli->cl_lru_in_list));
 575         while (!list_empty(&cli->cl_lru_list)) {
 576                 struct cl_page *page;
 577                 bool will_free = false;
 578
 579                 if (!force && atomic_read(&cli->cl_lru_shrinkers) > 1)
 580                         break;
 581
 582                 if (--maxscan < 0)
 583                         break;
 584
 585                 opg = list_first_entry(&cli->cl_lru_list, struct osc_page,
 586                                        ops_lru);
 587                 page = opg->ops_cl.cpl_page;
 588                 if (lru_page_busy(cli, page)) {
 589                         list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
 590                         continue;
 591                 }
 592
 593                 LASSERT(page->cp_obj != NULL);
 594                 if (clobj != page->cp_obj) {
 595                         struct cl_object *tmp = page->cp_obj;
 596
 597                         cl_object_get(tmp);
 598                         spin_unlock(&cli->cl_lru_list_lock);
 599
 600                         if (clobj != NULL) {
 601                                 discard_cl_pages(env, io, pvec, index);
 602                                 index = 0;
 603
 604                                 cl_io_fini(env, io);
 605                                 cl_object_put(env, clobj);
 606                                 clobj = NULL;
 607                         }
 608
 609                         clobj = tmp;
 610                         io->ci_obj = clobj;
 611                         io->ci_ignore_layout = 1;
 612                         rc = cl_io_init(env, io, CIT_MISC, clobj);
 613
 614                         spin_lock(&cli->cl_lru_list_lock);
 615
 616                         if (rc != 0)
 617                                 break;
 618
 619                         ++maxscan;
 620                         continue;
 621                 }
 622
 623                 if (cl_page_own_try(env, io, page) == 0) {
 624                         if (!lru_page_busy(cli, page)) {
 625                                 /* remove it from lru list earlier to avoid
 626                                  * lock contention */
 627                                 __osc_lru_del(cli, opg);
 628                                 opg->ops_in_lru = 0; /* will be discarded */
 629
 630                                 cl_page_get(page);
 631                                 will_free = true;
 632                         } else {
 633                                 cl_page_disown(env, io, page);
 634                         }
 635                 }
 636
 637                 if (!will_free) {
 638                         list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
 639                         continue;
 640                 }
 641
 642                 /* Don't discard and free the page with cl_lru_list held */
 643                 pvec[index++] = page;
 644                 if (unlikely(index == OTI_PVEC_SIZE)) {
 645                         spin_unlock(&cli->cl_lru_list_lock);
 646                         discard_cl_pages(env, io, pvec, index);
 647                         index = 0;
 648
 649                         spin_lock(&cli->cl_lru_list_lock);
 650                 }
 651
 652                 if (++count >= target)
 653                         break;
 654         }
 655         spin_unlock(&cli->cl_lru_list_lock);
 656
 657         if (clobj != NULL) {
 658                 discard_cl_pages(env, io, pvec, index);
 659
 660                 cl_io_fini(env, io);
 661                 cl_object_put(env, clobj);
 662         }
 663
 664         atomic_dec(&cli->cl_lru_shrinkers);
 665         if (count > 0) {
 666                 atomic_long_add(count, cli->cl_lru_left);
 667                 wake_up(&osc_lru_waitq);
 668         }
 669         RETURN(count > 0 ? count : rc);
 670 }
 671 EXPORT_SYMBOL(osc_lru_shrink);
 672
 673 /**
 674  * Reclaim LRU pages by an IO thread. The caller wants to reclaim at least
 675  * \@npages of LRU slots. For performance consideration, it's better to drop
 676  * LRU pages in batch. Therefore, the actual number is adjusted at least
 677  * max_pages_per_rpc.
 678  */
 679 static long osc_lru_reclaim(struct client_obd *cli, unsigned long npages)
 680 {
 681         struct lu_env *env;
 682         struct cl_client_cache *cache = cli->cl_cache;
 683         struct client_obd *scan;
 684         int max_scans;
 685         __u16 refcheck;
 686         long rc = 0;
 687         ENTRY;
 688
 689         LASSERT(cache != NULL);
 690
 691         env = cl_env_get(&refcheck);
 692         if (IS_ERR(env))
 693                 RETURN(rc);
 694
 695         npages = max_t(int, npages, cli->cl_max_pages_per_rpc);
 696         CDEBUG(D_CACHE, "%s: start to reclaim %ld pages from LRU\n",
 697                cli_name(cli), npages);
 698         rc = osc_lru_shrink(env, cli, npages, true);
 699         if (rc >= npages) {
 700                 CDEBUG(D_CACHE, "%s: reclaimed %ld/%ld pages from LRU\n",
 701                        cli_name(cli), rc, npages);
 702                 if (osc_cache_too_much(cli) > 0)
 703                         ptlrpcd_queue_work(cli->cl_lru_work);
 704                 GOTO(out, rc);
 705         } else if (rc > 0) {
 706                 npages -= rc;
 707         }
 708
 709         CDEBUG(D_CACHE, "%s: cli %p no free slots, pages: %ld/%ld, want: %ld\n",
 710                 cli_name(cli), cli, atomic_long_read(&cli->cl_lru_in_list),
 711                 atomic_long_read(&cli->cl_lru_busy), npages);
 712
 713         /* Reclaim LRU slots from other client_obd as it can't free enough
 714          * from its own. This should rarely happen. */
 715         spin_lock(&cache->ccc_lru_lock);
 716         LASSERT(!list_empty(&cache->ccc_lru));
 717
 718         cache->ccc_lru_shrinkers++;
 719         list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
 720
 721         max_scans = refcount_read(&cache->ccc_users) - 2;
 722         while (--max_scans > 0 &&
 723                (scan = list_first_entry_or_null(&cache->ccc_lru,
 724                                                   struct client_obd,
 725                                                   cl_lru_osc)) != NULL) {
 726                 CDEBUG(D_CACHE, "%s: cli %p LRU pages: %ld, busy: %ld.\n",
 727                        cli_name(scan), scan,
 728                        atomic_long_read(&scan->cl_lru_in_list),
 729                        atomic_long_read(&scan->cl_lru_busy));
 730
 731                 list_move_tail(&scan->cl_lru_osc, &cache->ccc_lru);
 732                 if (osc_cache_too_much(scan) > 0) {
 733                         spin_unlock(&cache->ccc_lru_lock);
 734
 735                         rc = osc_lru_shrink(env, scan, npages, true);
 736                         spin_lock(&cache->ccc_lru_lock);
 737                         if (rc >= npages)
 738                                 break;
 739                         if (rc > 0)
 740                                 npages -= rc;
 741                 }
 742         }
 743         spin_unlock(&cache->ccc_lru_lock);
 744
 745 out:
 746         cl_env_put(env, &refcheck);
 747         CDEBUG(D_CACHE, "%s: cli %p freed %ld pages.\n",
 748                cli_name(cli), cli, rc);
 749         return rc;
 750 }
 751
 752 /**
 753  * osc_lru_alloc() is called to allocate an LRU slot for a cl_page.
 754  *
 755  * Usually the LRU slots are reserved in osc_io_iter_rw_init().
 756  * Only in the case that the LRU slots are in extreme shortage, it should
 757  * have reserved enough slots for an IO.
 758  */
 759 static int osc_lru_alloc(const struct lu_env *env, struct client_obd *cli,
 760                          struct osc_page *opg)
 761 {
 762         struct osc_io *oio = osc_env_io(env);
 763         int rc = 0;
 764
 765         ENTRY;
 766
 767         if (cli->cl_cache == NULL) /* shall not be in LRU */
 768                 RETURN(0);
 769
 770         if (oio->oi_lru_reserved > 0) {
 771                 --oio->oi_lru_reserved;
 772                 goto out;
 773         }
 774
 775         LASSERT(atomic_long_read(cli->cl_lru_left) >= 0);
 776         while (!atomic_long_add_unless(cli->cl_lru_left, -1, 0)) {
 777                 /* run out of LRU spaces, try to drop some by itself */
 778                 rc = osc_lru_reclaim(cli, 1);
 779                 if (rc < 0)
 780                         break;
 781                 if (rc > 0)
 782                         continue;
 783                 /* IO issued by readahead, don't try hard */
 784                 if (oio->oi_is_readahead) {
 785                         if (atomic_long_read(cli->cl_lru_left) > 0)
 786                                 continue;
 787                         rc = -EBUSY;
 788                         break;
 789                 }
 790
 791                 cond_resched();
 792                 rc = l_wait_event_abortable(
 793                         osc_lru_waitq,
 794                         atomic_long_read(cli->cl_lru_left) > 0);
 795                 if (rc < 0) {
 796                         rc = -EINTR;
 797                         break;
 798                 }
 799         }
 800
 801 out:
 802         if (rc >= 0) {
 803                 atomic_long_inc(&cli->cl_lru_busy);
 804                 opg->ops_in_lru = 1;
 805                 rc = 0;
 806         }
 807
 808         RETURN(rc);
 809 }
 810
 811 /**
 812  * osc_lru_reserve() is called to reserve enough LRU slots for I/O.
 813  *
 814  * The benefit of doing this is to reduce contention against atomic counter
 815  * cl_lru_left by changing it from per-page access to per-IO access.
 816  */
 817 unsigned long osc_lru_reserve(struct client_obd *cli, unsigned long npages)
 818 {
 819         unsigned long reserved = 0;
 820         unsigned long max_pages;
 821         unsigned long c;
 822         int rc;
 823
 824 again:
 825         c = atomic_long_read(cli->cl_lru_left);
 826         if (c < npages && osc_lru_reclaim(cli, npages) > 0)
 827                 c = atomic_long_read(cli->cl_lru_left);
 828
 829         if (c < npages) {
 830                 /*
 831                  * Trigger writeback in the hope some LRU slot could
 832                  * be freed.
 833                  */
 834                 rc = ptlrpcd_queue_work(cli->cl_writeback_work);
 835                 if (rc)
 836                         return 0;
 837         }
 838
 839         while (c >= npages) {
 840                 if (c == atomic_long_cmpxchg(cli->cl_lru_left, c, c - npages)) {
 841                         reserved = npages;
 842                         break;
 843                 }
 844                 c = atomic_long_read(cli->cl_lru_left);
 845         }
 846
 847         if (reserved != npages) {
 848                 cond_resched();
 849                 rc = l_wait_event_abortable(
 850                         osc_lru_waitq,
 851                         atomic_long_read(cli->cl_lru_left) > 0);
 852                 goto again;
 853         }
 854
 855         max_pages = cli->cl_max_pages_per_rpc * cli->cl_max_rpcs_in_flight;
 856         if (atomic_long_read(cli->cl_lru_left) < max_pages) {
 857                 /* If there aren't enough pages in the per-OSC LRU then
 858                  * wake up the LRU thread to try and clear out space, so
 859                  * we don't block if pages are being dirtied quickly. */
 860                 CDEBUG(D_CACHE, "%s: queue LRU, left: %lu/%ld.\n",
 861                        cli_name(cli), atomic_long_read(cli->cl_lru_left),
 862                        max_pages);
 863                 (void)ptlrpcd_queue_work(cli->cl_lru_work);
 864         }
 865
 866         return reserved;
 867 }
 868
 869 /**
 870  * osc_lru_unreserve() is called to unreserve LRU slots.
 871  *
 872  * LRU slots reserved by osc_lru_reserve() may have entries left due to several
 873  * reasons such as page already existing or I/O error. Those reserved slots
 874  * should be freed by calling this function.
 875  */
 876 void osc_lru_unreserve(struct client_obd *cli, unsigned long npages)
 877 {
 878         atomic_long_add(npages, cli->cl_lru_left);
 879         wake_up(&osc_lru_waitq);
 880 }
 881
 882 /**
 883  * Atomic operations are expensive. We accumulate the accounting for the
 884  * same page zone to get better performance.
 885  * In practice this can work pretty good because the pages in the same RPC
 886  * are likely from the same page zone.
 887  */
 888 #ifdef HAVE_NR_UNSTABLE_NFS
 889 /* Old kernels use a separate counter for unstable pages,
 890  * newer kernels treat them like any other writeback.
 891  * (see Linux commit: v5.7-467-g8d92890bd6b8)
 892  */
 893 #define NR_ZONE_WRITE_PENDING           ((enum zone_stat_item)NR_UNSTABLE_NFS)
 894 #elif !defined(HAVE_NR_ZONE_WRITE_PENDING)
 895 #define NR_ZONE_WRITE_PENDING           ((enum zone_stat_item)NR_WRITEBACK)
 896 #endif
 897
 898 static inline void unstable_page_accounting(struct ptlrpc_bulk_desc *desc,
 899                                             int factor)
 900 {
 901         int page_count;
 902         void *zone = NULL;
 903         int count = 0;
 904         int i;
 905
 906         ENTRY;
 907
 908         page_count = desc->bd_iov_count;
 909
 910         CDEBUG(D_PAGE, "%s %d unstable pages\n",
 911                factor == 1 ? "adding" : "removing", page_count);
 912
 913         for (i = 0; i < page_count; i++) {
 914                 void *pz = page_zone(desc->bd_vec[i].bv_page);
 915
 916                 if (likely(pz == zone)) {
 917                         ++count;
 918                         continue;
 919                 }
 920
 921                 if (count > 0) {
 922                         mod_zone_page_state(zone, NR_ZONE_WRITE_PENDING,
 923                                             factor * count);
 924                         count = 0;
 925                 }
 926                 zone = pz;
 927                 ++count;
 928         }
 929         if (count > 0)
 930                 mod_zone_page_state(zone, NR_ZONE_WRITE_PENDING,
 931                                     factor * count);
 932
 933         EXIT;
 934 }
 935
 936 static inline void add_unstable_pages(struct ptlrpc_bulk_desc *desc)
 937 {
 938         unstable_page_accounting(desc, 1);
 939 }
 940
 941 static inline void dec_unstable_pages(struct ptlrpc_bulk_desc *desc)
 942 {
 943         unstable_page_accounting(desc, -1);
 944 }
 945
 946 /**
 947  * Performs "unstable" page accounting. This function balances the
 948  * increment operations performed in osc_inc_unstable_pages. It is
 949  * registered as the RPC request callback, and is executed when the
 950  * bulk RPC is committed on the server. Thus at this point, the pages
 951  * involved in the bulk transfer are no longer considered unstable.
 952  *
 953  * If this function is called, the request should have been committed
 954  * or req:rq_unstable must have been set; it implies that the unstable
 955  * statistic have been added.
 956  */
 957 void osc_dec_unstable_pages(struct ptlrpc_request *req)
 958 {
 959         struct ptlrpc_bulk_desc *desc       = req->rq_bulk;
 960         struct client_obd       *cli        = &req->rq_import->imp_obd->u.cli;
 961         int                      page_count;
 962         long                     unstable_count;
 963
 964         /* no desc means short io, which doesn't have separate unstable pages,
 965          * it's just using space inside the RPC itself
 966          */
 967         if (!desc)
 968                 return;
 969
 970         page_count = desc->bd_iov_count;
 971
 972         LASSERT(page_count >= 0);
 973
 974         dec_unstable_pages(desc);
 975
 976         unstable_count = atomic_long_sub_return(page_count,
 977                                                 &cli->cl_unstable_count);
 978         LASSERT(unstable_count >= 0);
 979
 980         unstable_count = atomic_long_sub_return(page_count,
 981                                            &cli->cl_cache->ccc_unstable_nr);
 982         LASSERT(unstable_count >= 0);
 983
 984         if (waitqueue_active(&osc_lru_waitq))
 985                 (void)ptlrpcd_queue_work(cli->cl_lru_work);
 986 }
 987
 988 /**
 989  * "unstable" page accounting. See: osc_dec_unstable_pages.
 990  */
 991 void osc_inc_unstable_pages(struct ptlrpc_request *req)
 992 {
 993         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
 994         struct client_obd       *cli  = &req->rq_import->imp_obd->u.cli;
 995         long                     page_count;
 996
 997         /* No unstable page tracking */
 998         if (cli->cl_cache == NULL || !cli->cl_cache->ccc_unstable_check)
 999                 return;
1000
1001         /* no desc means short io, which doesn't have separate unstable pages,
1002          * it's just using space inside the RPC itself
1003          */
1004         if (!desc)
1005                 return;
1006
1007         page_count = desc->bd_iov_count;
1008
1009         add_unstable_pages(desc);
1010         atomic_long_add(page_count, &cli->cl_unstable_count);
1011         atomic_long_add(page_count, &cli->cl_cache->ccc_unstable_nr);
1012
1013         /* If the request has already been committed (i.e. brw_commit
1014          * called via rq_commit_cb), we need to undo the unstable page
1015          * increments we just performed because rq_commit_cb wont be
1016          * called again. */
1017         spin_lock(&req->rq_lock);
1018         if (unlikely(req->rq_committed)) {
1019                 spin_unlock(&req->rq_lock);
1020
1021                 osc_dec_unstable_pages(req);
1022         } else {
1023                 req->rq_unstable = 1;
1024                 spin_unlock(&req->rq_lock);
1025         }
1026 }
1027
1028 /**
1029  * Check if it piggybacks SOFT_SYNC flag to OST from this OSC.
1030  * This function will be called by every BRW RPC so it's critical
1031  * to make this function fast.
1032  */
1033 bool osc_over_unstable_soft_limit(struct client_obd *cli)
1034 {
1035         long unstable_nr, osc_unstable_count;
1036
1037         /* Can't check cli->cl_unstable_count, therefore, no soft limit */
1038         if (cli->cl_cache == NULL || !cli->cl_cache->ccc_unstable_check)
1039                 return false;
1040
1041         osc_unstable_count = atomic_long_read(&cli->cl_unstable_count);
1042         unstable_nr = atomic_long_read(&cli->cl_cache->ccc_unstable_nr);
1043
1044         CDEBUG(D_CACHE,
1045                "%s: cli: %p unstable pages: %lu, osc unstable pages: %lu\n",
1046                cli_name(cli), cli, unstable_nr, osc_unstable_count);
1047
1048         /* If the LRU slots are in shortage - 25% remaining AND this OSC
1049          * has one full RPC window of unstable pages, it's a good chance
1050          * to piggyback a SOFT_SYNC flag.
1051          * Please notice that the OST won't take immediate response for the
1052          * SOFT_SYNC request so active OSCs will have more chance to carry
1053          * the flag, this is reasonable. */
1054         return unstable_nr > cli->cl_cache->ccc_lru_max >> 2 &&
1055                osc_unstable_count > cli->cl_max_pages_per_rpc *
1056                                     cli->cl_max_rpcs_in_flight;
1057 }
1058
1059 /**
1060  * Return how many LRU pages in the cache of all OSC devices
1061  *
1062  * \retval      return # of cached LRU pages times reclaimation tendency
1063  * \retval      SHRINK_STOP if it cannot do any scanning in this time
1064  */
1065 unsigned long osc_cache_shrink_count(struct shrinker *sk,
1066                                      struct shrink_control *sc)
1067 {
1068         struct client_obd *cli;
1069         unsigned long cached = 0;
1070
1071         if (!osc_page_cache_shrink_enabled)
1072                 return 0;
1073
1074         spin_lock(&osc_shrink_lock);
1075         list_for_each_entry(cli, &osc_shrink_list, cl_shrink_list)
1076                 cached += atomic_long_read(&cli->cl_lru_in_list);
1077         spin_unlock(&osc_shrink_lock);
1078
1079         return (cached  * sysctl_vfs_cache_pressure) / 100;
1080 }
1081
1082 /**
1083  * Scan and try to reclaim sc->nr_to_scan cached LRU pages
1084  *
1085  * \retval      number of cached LRU pages reclaimed
1086  * \retval      SHRINK_STOP if it cannot do any scanning in this time
1087  *
1088  * Linux kernel will loop calling this shrinker scan routine with
1089  * sc->nr_to_scan = SHRINK_BATCH(128 for now) until kernel got enough memory.
1090  *
1091  * If sc->nr_to_scan is 0, the VM is querying the cache size, we don't need
1092  * to scan and try to reclaim LRU pages, just return 0 and
1093  * osc_cache_shrink_count() will report the LRU page number.
1094  */
1095 unsigned long osc_cache_shrink_scan(struct shrinker *sk,
1096                                     struct shrink_control *sc)
1097 {
1098         struct client_obd *cli;
1099         struct client_obd *stop_anchor = NULL;
1100         struct lu_env *env;
1101         long shrank = 0;
1102         int rc;
1103         __u16 refcheck;
1104
1105         if (sc->nr_to_scan == 0)
1106                 return 0;
1107
1108         if (!(sc->gfp_mask & __GFP_FS))
1109                 return SHRINK_STOP;
1110
1111         env = cl_env_get(&refcheck);
1112         if (IS_ERR(env))
1113                 return SHRINK_STOP;
1114
1115         spin_lock(&osc_shrink_lock);
1116         while ((cli = list_first_entry_or_null(&osc_shrink_list,
1117                                                struct client_obd,
1118                                                cl_shrink_list)) != NULL) {
1119                 if (stop_anchor == NULL)
1120                         stop_anchor = cli;
1121                 else if (cli == stop_anchor)
1122                         break;
1123
1124                 list_move_tail(&cli->cl_shrink_list, &osc_shrink_list);
1125                 spin_unlock(&osc_shrink_lock);
1126
1127                 /* shrink no more than max_pages_per_rpc for an OSC */
1128                 rc = osc_lru_shrink(env, cli, (sc->nr_to_scan - shrank) >
1129                                     cli->cl_max_pages_per_rpc ?
1130                                     cli->cl_max_pages_per_rpc :
1131                                     sc->nr_to_scan - shrank, true);
1132                 if (rc > 0)
1133                         shrank += rc;
1134
1135                 if (shrank >= sc->nr_to_scan)
1136                         goto out;
1137
1138                 spin_lock(&osc_shrink_lock);
1139         }
1140         spin_unlock(&osc_shrink_lock);
1141
1142 out:
1143         cl_env_put(env, &refcheck);
1144
1145         return shrank;
1146 }
1147
1148 /** @} osc */