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
 173         if (slice->cpl_page->cp_type == CPT_CACHEABLE) {
 174                 void *value = NULL;
 175
 176                 rc = osc_teardown_async_page(env, obj, opg);
 177                 if (rc) {
 178                         CL_PAGE_DEBUG(D_ERROR, env, slice->cpl_page,
 179                                       "Trying to teardown failed: %d\n", rc);
 180                         LASSERT(0);
 181                 }
 182
 183                 osc_lru_del(osc_cli(obj), opg);
 184
 185                 spin_lock(&obj->oo_tree_lock);
 186                 if (opg->ops_intree) {
 187                         value = radix_tree_delete(&obj->oo_tree,
 188                                                   osc_index(opg));
 189                         if (value != NULL) {
 190                                 --obj->oo_npages;
 191                                 opg->ops_intree = 0;
 192                         }
 193                 }
 194                 spin_unlock(&obj->oo_tree_lock);
 195
 196                 LASSERT(ergo(value != NULL, value == opg));
 197         }
 198
 199         EXIT;
 200 }
 201
 202 static void osc_page_clip(const struct lu_env *env,
 203                           const struct cl_page_slice *slice,
 204                           int from, int to)
 205 {
 206         struct osc_page       *opg = cl2osc_page(slice);
 207         struct osc_async_page *oap = &opg->ops_oap;
 208
 209         CDEBUG(D_CACHE, "from %d, to %d\n", from, to);
 210
 211         opg->ops_from = from;
 212         /* argument @to is exclusive, but @ops_to is inclusive */
 213         opg->ops_to   = to - 1;
 214         oap->oap_async_flags |= ASYNC_COUNT_STABLE;
 215 }
 216
 217 static int osc_page_flush(const struct lu_env *env,
 218                           const struct cl_page_slice *slice,
 219                           struct cl_io *io)
 220 {
 221         struct osc_page *opg = cl2osc_page(slice);
 222         int rc = 0;
 223         ENTRY;
 224         rc = osc_flush_async_page(env, io, opg);
 225         RETURN(rc);
 226 }
 227
 228 static void osc_page_touch(const struct lu_env *env,
 229                           const struct cl_page_slice *slice, size_t to)
 230 {
 231         struct osc_page *opg = cl2osc_page(slice);
 232         struct cl_object *obj = osc2cl(osc_page_object(opg));
 233
 234         osc_page_touch_at(env, obj, osc_index(opg), to);
 235 }
 236
 237 static const struct cl_page_operations osc_transient_page_ops = {
 238         .cpo_print         = osc_page_print,
 239         .cpo_delete        = osc_page_delete,
 240         .cpo_clip           = osc_page_clip,
 241 };
 242
 243 static const struct cl_page_operations osc_page_ops = {
 244         .cpo_print         = osc_page_print,
 245         .cpo_delete        = osc_page_delete,
 246         .cpo_clip           = osc_page_clip,
 247         .cpo_flush          = osc_page_flush,
 248         .cpo_page_touch    = osc_page_touch,
 249 };
 250
 251 int osc_page_init(const struct lu_env *env, struct cl_object *obj,
 252                   struct cl_page *cl_page, pgoff_t index)
 253 {
 254         struct osc_object *osc = cl2osc(obj);
 255         struct osc_page *opg = cl_object_page_slice(obj, cl_page);
 256         struct osc_io *oio = osc_env_io(env);
 257         int result;
 258
 259         opg->ops_from = 0;
 260         opg->ops_to = PAGE_SIZE - 1;
 261
 262         INIT_LIST_HEAD(&opg->ops_lru);
 263
 264         result = osc_prep_async_page(osc, opg, cl_page, index << PAGE_SHIFT);
 265         if (result != 0)
 266                 return result;
 267
 268         opg->ops_srvlock = osc_io_srvlock(oio);
 269
 270         if (cl_page->cp_type == CPT_TRANSIENT) {
 271                 cl_page_slice_add(cl_page, &opg->ops_cl, obj,
 272                                   &osc_transient_page_ops);
 273         } else if (cl_page->cp_type == CPT_CACHEABLE) {
 274                 cl_page_slice_add(cl_page, &opg->ops_cl, obj, &osc_page_ops);
 275                 /* reserve an LRU space for this page */
 276                 result = osc_lru_alloc(env, osc_cli(osc), opg);
 277                 if (result == 0) {
 278                         result = radix_tree_preload(GFP_NOFS);
 279                         if (result == 0) {
 280                                 spin_lock(&osc->oo_tree_lock);
 281                                 result = radix_tree_insert(&osc->oo_tree,
 282                                                            index, opg);
 283                                 if (result == 0) {
 284                                         ++osc->oo_npages;
 285                                         opg->ops_intree = 1;
 286                                 }
 287                                 spin_unlock(&osc->oo_tree_lock);
 288
 289                                 radix_tree_preload_end();
 290                         }
 291                 }
 292         }
 293
 294         return result;
 295 }
 296 EXPORT_SYMBOL(osc_page_init);
 297
 298 /**
 299  * Helper function called by osc_io_submit() for every page in an immediate
 300  * transfer (i.e., transferred synchronously).
 301  */
 302 void osc_page_submit(const struct lu_env *env, struct osc_page *opg,
 303                      enum cl_req_type crt, int brw_flags)
 304 {
 305         struct osc_io *oio = osc_env_io(env);
 306         struct osc_async_page *oap = &opg->ops_oap;
 307
 308         LASSERT(oap->oap_async_flags & ASYNC_READY);
 309         LASSERT(oap->oap_async_flags & ASYNC_COUNT_STABLE);
 310
 311         oap->oap_cmd = crt == CRT_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ;
 312         oap->oap_page_off = opg->ops_from;
 313         oap->oap_count = opg->ops_to - opg->ops_from + 1;
 314         oap->oap_brw_flags = OBD_BRW_SYNC | brw_flags;
 315
 316         if (oio->oi_cap_sys_resource)
 317                 oap->oap_brw_flags |= OBD_BRW_SYS_RESOURCE;
 318
 319         osc_page_transfer_get(opg, "transfer\0imm");
 320         osc_page_transfer_add(env, opg, crt);
 321 }
 322
 323 /* --------------- LRU page management ------------------ */
 324
 325 /* OSC is a natural place to manage LRU pages as applications are specialized
 326  * to write OSC by OSC. Ideally, if one OSC is used more frequently it should
 327  * occupy more LRU slots. On the other hand, we should avoid using up all LRU
 328  * slots (client_obd::cl_lru_left) otherwise process has to be put into sleep
 329  * for free LRU slots - this will be very bad so the algorithm requires each
 330  * OSC to free slots voluntarily to maintain a reasonable number of free slots
 331  * at any time.
 332  */
 333
 334 static DECLARE_WAIT_QUEUE_HEAD(osc_lru_waitq);
 335
 336 /**
 337  * LRU pages are freed in batch mode. OSC should at least free this
 338  * number of pages to avoid running out of LRU slots.
 339  */
 340 static inline int lru_shrink_min(struct client_obd *cli)
 341 {
 342         return cli->cl_max_pages_per_rpc * 2;
 343 }
 344
 345 /**
 346  * free this number at most otherwise it will take too long time to finsih.
 347  */
 348 static inline int lru_shrink_max(struct client_obd *cli)
 349 {
 350         return cli->cl_max_pages_per_rpc * cli->cl_max_rpcs_in_flight;
 351 }
 352
 353 /**
 354  * Check if we can free LRU slots from this OSC. If there exists LRU waiters,
 355  * we should free slots aggressively. In this way, slots are freed in a steady
 356  * step to maintain fairness among OSCs.
 357  *
 358  * Return how many LRU pages should be freed.
 359  */
 360 static int osc_cache_too_much(struct client_obd *cli)
 361 {
 362         struct cl_client_cache *cache = cli->cl_cache;
 363         long pages = atomic_long_read(&cli->cl_lru_in_list);
 364         unsigned long budget;
 365
 366         LASSERT(cache != NULL);
 367         budget = cache->ccc_lru_max / (refcount_read(&cache->ccc_users) - 2);
 368
 369         /* if it's going to run out LRU slots, we should free some, but not
 370          * too much to maintain faireness among OSCs. */
 371         if (atomic_long_read(cli->cl_lru_left) < cache->ccc_lru_max >> 2) {
 372                 if (pages >= budget)
 373                         return lru_shrink_max(cli);
 374                 else if (pages >= budget / 2)
 375                         return lru_shrink_min(cli);
 376         } else {
 377                 time64_t duration = ktime_get_real_seconds();
 378                 long timediff;
 379
 380                 /* knock out pages by duration of no IO activity */
 381                 duration -= cli->cl_lru_last_used;
 382                 /*
 383                  * The difference shouldn't be more than 70 years
 384                  * so we can safely case to a long. Round to
 385                  * approximately 1 minute.
 386                  */
 387                 timediff = (long)(duration >> 6);
 388                 if (timediff > 0 && pages >= budget / timediff)
 389                         return lru_shrink_min(cli);
 390         }
 391         return 0;
 392 }
 393
 394 int lru_queue_work(const struct lu_env *env, void *data)
 395 {
 396         struct client_obd *cli = data;
 397         int count;
 398
 399         CDEBUG(D_CACHE, "%s: run LRU work for client obd\n", cli_name(cli));
 400         count = osc_cache_too_much(cli);
 401         if (count > 0) {
 402                 int rc = osc_lru_shrink(env, cli, count, false);
 403
 404                 CDEBUG(D_CACHE, "%s: shrank %d/%d pages from client obd\n",
 405                        cli_name(cli), rc, count);
 406                 if (rc >= count) {
 407                         CDEBUG(D_CACHE, "%s: queue again\n", cli_name(cli));
 408                         ptlrpcd_queue_work(cli->cl_lru_work);
 409                 }
 410         }
 411
 412         RETURN(0);
 413 }
 414
 415 void osc_lru_add_batch(struct client_obd *cli, struct list_head *plist)
 416 {
 417         LIST_HEAD(lru);
 418         struct osc_async_page *oap;
 419         long npages = 0;
 420
 421         list_for_each_entry(oap, plist, oap_pending_item) {
 422                 struct osc_page *opg = oap2osc_page(oap);
 423
 424                 if (!opg->ops_in_lru)
 425                         continue;
 426
 427                 ++npages;
 428                 LASSERT(list_empty(&opg->ops_lru));
 429                 list_add(&opg->ops_lru, &lru);
 430         }
 431
 432         if (npages > 0) {
 433                 spin_lock(&cli->cl_lru_list_lock);
 434                 list_splice_tail(&lru, &cli->cl_lru_list);
 435                 atomic_long_sub(npages, &cli->cl_lru_busy);
 436                 atomic_long_add(npages, &cli->cl_lru_in_list);
 437                 cli->cl_lru_last_used = ktime_get_real_seconds();
 438                 spin_unlock(&cli->cl_lru_list_lock);
 439
 440                 if (waitqueue_active(&osc_lru_waitq))
 441                         (void)ptlrpcd_queue_work(cli->cl_lru_work);
 442         }
 443 }
 444
 445 static void __osc_lru_del(struct client_obd *cli, struct osc_page *opg)
 446 {
 447         LASSERT(atomic_long_read(&cli->cl_lru_in_list) > 0);
 448         list_del_init(&opg->ops_lru);
 449         atomic_long_dec(&cli->cl_lru_in_list);
 450 }
 451
 452 /**
 453  * Page is being destroyed. The page may be not in LRU list, if the transfer
 454  * has never finished(error occurred).
 455  */
 456 static void osc_lru_del(struct client_obd *cli, struct osc_page *opg)
 457 {
 458         if (opg->ops_in_lru) {
 459                 spin_lock(&cli->cl_lru_list_lock);
 460                 if (!list_empty(&opg->ops_lru)) {
 461                         __osc_lru_del(cli, opg);
 462                 } else {
 463                         LASSERT(atomic_long_read(&cli->cl_lru_busy) > 0);
 464                         atomic_long_dec(&cli->cl_lru_busy);
 465                 }
 466                 spin_unlock(&cli->cl_lru_list_lock);
 467
 468                 atomic_long_inc(cli->cl_lru_left);
 469                 /* this is a great place to release more LRU pages if
 470                  * this osc occupies too many LRU pages and kernel is
 471                  * stealing one of them. */
 472                 if (osc_cache_too_much(cli)) {
 473                         CDEBUG(D_CACHE, "%s: queue LRU work\n", cli_name(cli));
 474                         (void)ptlrpcd_queue_work(cli->cl_lru_work);
 475                 }
 476                 wake_up(&osc_lru_waitq);
 477         } else {
 478                 LASSERT(list_empty(&opg->ops_lru));
 479         }
 480 }
 481
 482 /**
 483  * Delete page from LRU list for redirty.
 484  */
 485 static void osc_lru_use(struct client_obd *cli, struct osc_page *opg)
 486 {
 487         /* If page is being transferred for the first time,
 488          * ops_lru should be empty */
 489         if (opg->ops_in_lru) {
 490                 if (list_empty(&opg->ops_lru))
 491                         return;
 492                 spin_lock(&cli->cl_lru_list_lock);
 493                 if (!list_empty(&opg->ops_lru)) {
 494                         __osc_lru_del(cli, opg);
 495                         atomic_long_inc(&cli->cl_lru_busy);
 496                 }
 497                 spin_unlock(&cli->cl_lru_list_lock);
 498         }
 499 }
 500
 501 static void discard_cl_pages(const struct lu_env *env, struct cl_io *io,
 502                              struct cl_page **pvec, int max_index)
 503 {
 504         struct folio_batch *fbatch = &osc_env_info(env)->oti_fbatch;
 505         int i;
 506
 507         ll_folio_batch_init(fbatch, 0);
 508         for (i = 0; i < max_index; i++) {
 509                 struct cl_page *page = pvec[i];
 510
 511                 LASSERT(page->cp_type != CPT_TRANSIENT);
 512                 LASSERT(cl_page_is_owned(page, io));
 513                 cl_page_discard(env, io, page);
 514                 cl_page_disown(env, io, page);
 515                 cl_batch_put(env, page, fbatch);
 516
 517                 pvec[i] = NULL;
 518         }
 519         folio_batch_release(fbatch);
 520 }
 521
 522 /**
 523  * Check if a cl_page can be released, i.e, it's not being used.
 524  *
 525  * If unstable account is turned on, bulk transfer may hold one refcount
 526  * for recovery so we need to check vmpage refcount as well; otherwise,
 527  * even we can destroy cl_page but the corresponding vmpage can't be reused.
 528  */
 529 static inline bool lru_page_busy(struct client_obd *cli, struct cl_page *page)
 530 {
 531         if (cl_page_in_use_noref(page))
 532                 return true;
 533
 534         if (cli->cl_cache->ccc_unstable_check) {
 535                 struct page *vmpage = cl_page_vmpage(page);
 536
 537                 /* vmpage have two known users: cl_page and VM page cache */
 538                 if (page_count(vmpage) - page_mapcount(vmpage) > 2)
 539                         return true;
 540         }
 541         return false;
 542 }
 543
 544 /**
 545  * Drop @target of pages from LRU at most.
 546  */
 547 long osc_lru_shrink(const struct lu_env *env, struct client_obd *cli,
 548                    long target, bool force)
 549 {
 550         struct cl_io *io;
 551         struct cl_object *clobj = NULL;
 552         struct cl_page **pvec;
 553         struct osc_page *opg;
 554         long count = 0;
 555         int maxscan = 0;
 556         int index = 0;
 557         int rc = 0;
 558         ENTRY;
 559
 560         LASSERT(atomic_long_read(&cli->cl_lru_in_list) >= 0);
 561         if (atomic_long_read(&cli->cl_lru_in_list) == 0 || target <= 0)
 562                 RETURN(0);
 563
 564         CDEBUG(D_CACHE, "%s: shrinkers: %d, force: %d\n",
 565                cli_name(cli), atomic_read(&cli->cl_lru_shrinkers), force);
 566         if (!force) {
 567                 if (atomic_read(&cli->cl_lru_shrinkers) > 0)
 568                         RETURN(-EBUSY);
 569
 570                 if (atomic_inc_return(&cli->cl_lru_shrinkers) > 1) {
 571                         atomic_dec(&cli->cl_lru_shrinkers);
 572                         RETURN(-EBUSY);
 573                 }
 574         } else {
 575                 atomic_inc(&cli->cl_lru_shrinkers);
 576         }
 577
 578         pvec = (struct cl_page **)osc_env_info(env)->oti_pvec;
 579         io = osc_env_thread_io(env);
 580
 581         spin_lock(&cli->cl_lru_list_lock);
 582         if (force)
 583                 cli->cl_lru_reclaim++;
 584         maxscan = min(target << 1, atomic_long_read(&cli->cl_lru_in_list));
 585         while (!list_empty(&cli->cl_lru_list)) {
 586                 struct cl_page *page;
 587                 bool will_free = false;
 588
 589                 if (!force && atomic_read(&cli->cl_lru_shrinkers) > 1)
 590                         break;
 591
 592                 if (--maxscan < 0)
 593                         break;
 594
 595                 opg = list_first_entry(&cli->cl_lru_list, struct osc_page,
 596                                        ops_lru);
 597                 page = opg->ops_cl.cpl_page;
 598                 if (lru_page_busy(cli, page)) {
 599                         list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
 600                         continue;
 601                 }
 602
 603                 LASSERT(page->cp_obj != NULL);
 604                 if (clobj != page->cp_obj) {
 605                         struct cl_object *tmp = page->cp_obj;
 606
 607                         cl_object_get(tmp);
 608                         spin_unlock(&cli->cl_lru_list_lock);
 609
 610                         if (clobj != NULL) {
 611                                 discard_cl_pages(env, io, pvec, index);
 612                                 index = 0;
 613
 614                                 cl_io_fini(env, io);
 615                                 cl_object_put(env, clobj);
 616                                 clobj = NULL;
 617                                 cond_resched();
 618                         }
 619
 620                         clobj = tmp;
 621                         io->ci_obj = clobj;
 622                         io->ci_ignore_layout = 1;
 623                         rc = cl_io_init(env, io, CIT_MISC, clobj);
 624
 625                         spin_lock(&cli->cl_lru_list_lock);
 626
 627                         if (rc != 0)
 628                                 break;
 629
 630                         ++maxscan;
 631                         continue;
 632                 }
 633
 634                 if (cl_page_own_try(env, io, page) == 0) {
 635                         if (!lru_page_busy(cli, page)) {
 636                                 /* remove it from lru list earlier to avoid
 637                                  * lock contention */
 638                                 __osc_lru_del(cli, opg);
 639                                 opg->ops_in_lru = 0; /* will be discarded */
 640
 641                                 cl_page_get(page);
 642                                 will_free = true;
 643                         } else {
 644                                 cl_page_disown(env, io, page);
 645                         }
 646                 }
 647
 648                 if (!will_free) {
 649                         list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
 650                         continue;
 651                 }
 652
 653                 /* Don't discard and free the page with cl_lru_list held */
 654                 pvec[index++] = page;
 655                 if (unlikely(index == OTI_PVEC_SIZE)) {
 656                         spin_unlock(&cli->cl_lru_list_lock);
 657                         discard_cl_pages(env, io, pvec, index);
 658                         index = 0;
 659
 660                         spin_lock(&cli->cl_lru_list_lock);
 661                 }
 662
 663                 if (++count >= target)
 664                         break;
 665         }
 666         spin_unlock(&cli->cl_lru_list_lock);
 667
 668         if (clobj != NULL) {
 669                 discard_cl_pages(env, io, pvec, index);
 670
 671                 cl_io_fini(env, io);
 672                 cl_object_put(env, clobj);
 673                 cond_resched();
 674         }
 675
 676         atomic_dec(&cli->cl_lru_shrinkers);
 677         if (count > 0) {
 678                 atomic_long_add(count, cli->cl_lru_left);
 679                 wake_up(&osc_lru_waitq);
 680         }
 681         RETURN(count > 0 ? count : rc);
 682 }
 683 EXPORT_SYMBOL(osc_lru_shrink);
 684
 685 /**
 686  * Reclaim LRU pages by an IO thread. The caller wants to reclaim at least
 687  * \@npages of LRU slots. For performance consideration, it's better to drop
 688  * LRU pages in batch. Therefore, the actual number is adjusted at least
 689  * max_pages_per_rpc.
 690  */
 691 static long osc_lru_reclaim(struct client_obd *cli, unsigned long npages)
 692 {
 693         struct lu_env *env;
 694         struct cl_client_cache *cache = cli->cl_cache;
 695         struct client_obd *scan;
 696         int max_scans;
 697         __u16 refcheck;
 698         long rc = 0;
 699         ENTRY;
 700
 701         LASSERT(cache != NULL);
 702
 703         env = cl_env_get(&refcheck);
 704         if (IS_ERR(env))
 705                 RETURN(rc);
 706
 707         npages = max_t(int, npages, cli->cl_max_pages_per_rpc);
 708         CDEBUG(D_CACHE, "%s: start to reclaim %ld pages from LRU\n",
 709                cli_name(cli), npages);
 710         rc = osc_lru_shrink(env, cli, npages, true);
 711         if (rc >= npages) {
 712                 CDEBUG(D_CACHE, "%s: reclaimed %ld/%ld pages from LRU\n",
 713                        cli_name(cli), rc, npages);
 714                 if (osc_cache_too_much(cli) > 0)
 715                         ptlrpcd_queue_work(cli->cl_lru_work);
 716                 GOTO(out, rc);
 717         } else if (rc > 0) {
 718                 npages -= rc;
 719         }
 720
 721         CDEBUG(D_CACHE, "%s: cli %p no free slots, pages: %ld/%ld, want: %ld\n",
 722                 cli_name(cli), cli, atomic_long_read(&cli->cl_lru_in_list),
 723                 atomic_long_read(&cli->cl_lru_busy), npages);
 724
 725         /* Reclaim LRU slots from other client_obd as it can't free enough
 726          * from its own. This should rarely happen. */
 727         spin_lock(&cache->ccc_lru_lock);
 728         LASSERT(!list_empty(&cache->ccc_lru));
 729
 730         cache->ccc_lru_shrinkers++;
 731         list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
 732
 733         max_scans = refcount_read(&cache->ccc_users) - 2;
 734         while (--max_scans > 0 &&
 735                (scan = list_first_entry_or_null(&cache->ccc_lru,
 736                                                   struct client_obd,
 737                                                   cl_lru_osc)) != NULL) {
 738                 CDEBUG(D_CACHE, "%s: cli %p LRU pages: %ld, busy: %ld.\n",
 739                        cli_name(scan), scan,
 740                        atomic_long_read(&scan->cl_lru_in_list),
 741                        atomic_long_read(&scan->cl_lru_busy));
 742
 743                 list_move_tail(&scan->cl_lru_osc, &cache->ccc_lru);
 744                 if (osc_cache_too_much(scan) > 0) {
 745                         spin_unlock(&cache->ccc_lru_lock);
 746
 747                         rc = osc_lru_shrink(env, scan, npages, true);
 748                         spin_lock(&cache->ccc_lru_lock);
 749                         if (rc >= npages)
 750                                 break;
 751                         if (rc > 0)
 752                                 npages -= rc;
 753                 }
 754         }
 755         spin_unlock(&cache->ccc_lru_lock);
 756
 757 out:
 758         cl_env_put(env, &refcheck);
 759         CDEBUG(D_CACHE, "%s: cli %p freed %ld pages.\n",
 760                cli_name(cli), cli, rc);
 761         return rc;
 762 }
 763
 764 /**
 765  * osc_lru_alloc() is called to allocate an LRU slot for a cl_page.
 766  *
 767  * Usually the LRU slots are reserved in osc_io_iter_rw_init().
 768  * Only in the case that the LRU slots are in extreme shortage, it should
 769  * have reserved enough slots for an IO.
 770  */
 771 static int osc_lru_alloc(const struct lu_env *env, struct client_obd *cli,
 772                          struct osc_page *opg)
 773 {
 774         struct osc_io *oio = osc_env_io(env);
 775         int rc = 0;
 776
 777         ENTRY;
 778
 779         if (cli->cl_cache == NULL) /* shall not be in LRU */
 780                 RETURN(0);
 781
 782         if (oio->oi_lru_reserved > 0) {
 783                 --oio->oi_lru_reserved;
 784                 goto out;
 785         }
 786
 787         LASSERT(atomic_long_read(cli->cl_lru_left) >= 0);
 788         while (!atomic_long_add_unless(cli->cl_lru_left, -1, 0)) {
 789                 /* run out of LRU spaces, try to drop some by itself */
 790                 rc = osc_lru_reclaim(cli, 1);
 791                 if (rc < 0)
 792                         break;
 793                 if (rc > 0)
 794                         continue;
 795                 /* IO issued by readahead, don't try hard */
 796                 if (oio->oi_is_readahead) {
 797                         if (atomic_long_read(cli->cl_lru_left) > 0)
 798                                 continue;
 799                         rc = -EBUSY;
 800                         break;
 801                 }
 802
 803                 cond_resched();
 804                 rc = l_wait_event_abortable(
 805                         osc_lru_waitq,
 806                         atomic_long_read(cli->cl_lru_left) > 0);
 807                 if (rc < 0) {
 808                         rc = -EINTR;
 809                         break;
 810                 }
 811         }
 812
 813 out:
 814         if (rc >= 0) {
 815                 atomic_long_inc(&cli->cl_lru_busy);
 816                 opg->ops_in_lru = 1;
 817                 rc = 0;
 818         }
 819
 820         RETURN(rc);
 821 }
 822
 823 /**
 824  * osc_lru_reserve() is called to reserve enough LRU slots for I/O.
 825  *
 826  * The benefit of doing this is to reduce contention against atomic counter
 827  * cl_lru_left by changing it from per-page access to per-IO access.
 828  */
 829 unsigned long osc_lru_reserve(struct client_obd *cli, unsigned long npages)
 830 {
 831         unsigned long reserved = 0;
 832         unsigned long max_pages;
 833         unsigned long c;
 834         int rc;
 835
 836 again:
 837         c = atomic_long_read(cli->cl_lru_left);
 838         if (c < npages && osc_lru_reclaim(cli, npages) > 0)
 839                 c = atomic_long_read(cli->cl_lru_left);
 840
 841         if (c < npages) {
 842                 /*
 843                  * Trigger writeback in the hope some LRU slot could
 844                  * be freed.
 845                  */
 846                 rc = ptlrpcd_queue_work(cli->cl_writeback_work);
 847                 if (rc)
 848                         return 0;
 849         }
 850
 851         while (c >= npages) {
 852                 if (c == atomic_long_cmpxchg(cli->cl_lru_left, c, c - npages)) {
 853                         reserved = npages;
 854                         break;
 855                 }
 856                 c = atomic_long_read(cli->cl_lru_left);
 857         }
 858
 859         if (reserved != npages) {
 860                 cond_resched();
 861                 rc = l_wait_event_abortable(
 862                         osc_lru_waitq,
 863                         atomic_long_read(cli->cl_lru_left) > 0);
 864                 goto again;
 865         }
 866
 867         max_pages = cli->cl_max_pages_per_rpc * cli->cl_max_rpcs_in_flight;
 868         if (atomic_long_read(cli->cl_lru_left) < max_pages) {
 869                 /* If there aren't enough pages in the per-OSC LRU then
 870                  * wake up the LRU thread to try and clear out space, so
 871                  * we don't block if pages are being dirtied quickly. */
 872                 CDEBUG(D_CACHE, "%s: queue LRU, left: %lu/%ld.\n",
 873                        cli_name(cli), atomic_long_read(cli->cl_lru_left),
 874                        max_pages);
 875                 (void)ptlrpcd_queue_work(cli->cl_lru_work);
 876         }
 877
 878         return reserved;
 879 }
 880
 881 /**
 882  * osc_lru_unreserve() is called to unreserve LRU slots.
 883  *
 884  * LRU slots reserved by osc_lru_reserve() may have entries left due to several
 885  * reasons such as page already existing or I/O error. Those reserved slots
 886  * should be freed by calling this function.
 887  */
 888 void osc_lru_unreserve(struct client_obd *cli, unsigned long npages)
 889 {
 890         atomic_long_add(npages, cli->cl_lru_left);
 891         wake_up(&osc_lru_waitq);
 892 }
 893
 894 /**
 895  * Atomic operations are expensive. We accumulate the accounting for the
 896  * same page zone to get better performance.
 897  * In practice this can work pretty good because the pages in the same RPC
 898  * are likely from the same page zone.
 899  */
 900 #ifdef HAVE_NR_UNSTABLE_NFS
 901 /* Old kernels use a separate counter for unstable pages,
 902  * newer kernels treat them like any other writeback.
 903  * (see Linux commit: v5.7-467-g8d92890bd6b8)
 904  */
 905 #define NR_ZONE_WRITE_PENDING           ((enum zone_stat_item)NR_UNSTABLE_NFS)
 906 #elif !defined(HAVE_NR_ZONE_WRITE_PENDING)
 907 #define NR_ZONE_WRITE_PENDING           ((enum zone_stat_item)NR_WRITEBACK)
 908 #endif
 909
 910 static inline void unstable_page_accounting(struct ptlrpc_bulk_desc *desc,
 911                                             int factor)
 912 {
 913         int page_count;
 914         void *zone = NULL;
 915         int count = 0;
 916         int i;
 917
 918         ENTRY;
 919
 920         page_count = desc->bd_iov_count;
 921
 922         CDEBUG(D_PAGE, "%s %d unstable pages\n",
 923                factor == 1 ? "adding" : "removing", page_count);
 924
 925         for (i = 0; i < page_count; i++) {
 926                 void *pz = page_zone(desc->bd_vec[i].bv_page);
 927
 928                 if (likely(pz == zone)) {
 929                         ++count;
 930                         continue;
 931                 }
 932
 933                 if (count > 0) {
 934                         mod_zone_page_state(zone, NR_ZONE_WRITE_PENDING,
 935                                             factor * count);
 936                         count = 0;
 937                 }
 938                 zone = pz;
 939                 ++count;
 940         }
 941         if (count > 0)
 942                 mod_zone_page_state(zone, NR_ZONE_WRITE_PENDING,
 943                                     factor * count);
 944
 945         EXIT;
 946 }
 947
 948 static inline void add_unstable_pages(struct ptlrpc_bulk_desc *desc)
 949 {
 950         unstable_page_accounting(desc, 1);
 951 }
 952
 953 static inline void dec_unstable_pages(struct ptlrpc_bulk_desc *desc)
 954 {
 955         unstable_page_accounting(desc, -1);
 956 }
 957
 958 /**
 959  * Performs "unstable" page accounting. This function balances the
 960  * increment operations performed in osc_inc_unstable_pages. It is
 961  * registered as the RPC request callback, and is executed when the
 962  * bulk RPC is committed on the server. Thus at this point, the pages
 963  * involved in the bulk transfer are no longer considered unstable.
 964  *
 965  * If this function is called, the request should have been committed
 966  * or req:rq_unstable must have been set; it implies that the unstable
 967  * statistic have been added.
 968  */
 969 void osc_dec_unstable_pages(struct ptlrpc_request *req)
 970 {
 971         struct ptlrpc_bulk_desc *desc       = req->rq_bulk;
 972         struct client_obd       *cli        = &req->rq_import->imp_obd->u.cli;
 973         int                      page_count;
 974         long                     unstable_count;
 975
 976         /* no desc means short io, which doesn't have separate unstable pages,
 977          * it's just using space inside the RPC itself
 978          */
 979         if (!desc)
 980                 return;
 981
 982         page_count = desc->bd_iov_count;
 983
 984         LASSERT(page_count >= 0);
 985
 986         dec_unstable_pages(desc);
 987
 988         unstable_count = atomic_long_sub_return(page_count,
 989                                                 &cli->cl_unstable_count);
 990         LASSERT(unstable_count >= 0);
 991
 992         unstable_count = atomic_long_sub_return(page_count,
 993                                            &cli->cl_cache->ccc_unstable_nr);
 994         LASSERT(unstable_count >= 0);
 995
 996         if (waitqueue_active(&osc_lru_waitq))
 997                 (void)ptlrpcd_queue_work(cli->cl_lru_work);
 998 }
 999
1000 /**
1001  * "unstable" page accounting. See: osc_dec_unstable_pages.
1002  */
1003 void osc_inc_unstable_pages(struct ptlrpc_request *req)
1004 {
1005         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
1006         struct client_obd       *cli  = &req->rq_import->imp_obd->u.cli;
1007         long                     page_count;
1008
1009         /* No unstable page tracking */
1010         if (cli->cl_cache == NULL || !cli->cl_cache->ccc_unstable_check)
1011                 return;
1012
1013         /* no desc means short io, which doesn't have separate unstable pages,
1014          * it's just using space inside the RPC itself
1015          */
1016         if (!desc)
1017                 return;
1018
1019         page_count = desc->bd_iov_count;
1020
1021         add_unstable_pages(desc);
1022         atomic_long_add(page_count, &cli->cl_unstable_count);
1023         atomic_long_add(page_count, &cli->cl_cache->ccc_unstable_nr);
1024
1025         /* If the request has already been committed (i.e. brw_commit
1026          * called via rq_commit_cb), we need to undo the unstable page
1027          * increments we just performed because rq_commit_cb wont be
1028          * called again. */
1029         spin_lock(&req->rq_lock);
1030         if (unlikely(req->rq_committed)) {
1031                 spin_unlock(&req->rq_lock);
1032
1033                 osc_dec_unstable_pages(req);
1034         } else {
1035                 req->rq_unstable = 1;
1036                 spin_unlock(&req->rq_lock);
1037         }
1038 }
1039
1040 /**
1041  * Check if it piggybacks SOFT_SYNC flag to OST from this OSC.
1042  * This function will be called by every BRW RPC so it's critical
1043  * to make this function fast.
1044  */
1045 bool osc_over_unstable_soft_limit(struct client_obd *cli)
1046 {
1047         long unstable_nr, osc_unstable_count;
1048
1049         /* Can't check cli->cl_unstable_count, therefore, no soft limit */
1050         if (cli->cl_cache == NULL || !cli->cl_cache->ccc_unstable_check)
1051                 return false;
1052
1053         osc_unstable_count = atomic_long_read(&cli->cl_unstable_count);
1054         unstable_nr = atomic_long_read(&cli->cl_cache->ccc_unstable_nr);
1055
1056         CDEBUG(D_CACHE,
1057                "%s: cli: %p unstable pages: %lu, osc unstable pages: %lu\n",
1058                cli_name(cli), cli, unstable_nr, osc_unstable_count);
1059
1060         /* If the LRU slots are in shortage - 25% remaining AND this OSC
1061          * has one full RPC window of unstable pages, it's a good chance
1062          * to piggyback a SOFT_SYNC flag.
1063          * Please notice that the OST won't take immediate response for the
1064          * SOFT_SYNC request so active OSCs will have more chance to carry
1065          * the flag, this is reasonable. */
1066         return unstable_nr > cli->cl_cache->ccc_lru_max >> 2 &&
1067                osc_unstable_count > cli->cl_max_pages_per_rpc *
1068                                     cli->cl_max_rpcs_in_flight;
1069 }
1070
1071 /**
1072  * Return how many LRU pages in the cache of all OSC devices
1073  *
1074  * \retval      return # of cached LRU pages times reclaimation tendency
1075  * \retval      SHRINK_STOP if it cannot do any scanning in this time
1076  */
1077 unsigned long osc_cache_shrink_count(struct shrinker *sk,
1078                                      struct shrink_control *sc)
1079 {
1080         struct client_obd *cli;
1081         unsigned long cached = 0;
1082
1083         if (!osc_page_cache_shrink_enabled)
1084                 return 0;
1085
1086         spin_lock(&osc_shrink_lock);
1087         list_for_each_entry(cli, &osc_shrink_list, cl_shrink_list)
1088                 cached += atomic_long_read(&cli->cl_lru_in_list);
1089         spin_unlock(&osc_shrink_lock);
1090
1091         return (cached  * sysctl_vfs_cache_pressure) / 100;
1092 }
1093
1094 /**
1095  * Scan and try to reclaim sc->nr_to_scan cached LRU pages
1096  *
1097  * \retval      number of cached LRU pages reclaimed
1098  * \retval      SHRINK_STOP if it cannot do any scanning in this time
1099  *
1100  * Linux kernel will loop calling this shrinker scan routine with
1101  * sc->nr_to_scan = SHRINK_BATCH(128 for now) until kernel got enough memory.
1102  *
1103  * If sc->nr_to_scan is 0, the VM is querying the cache size, we don't need
1104  * to scan and try to reclaim LRU pages, just return 0 and
1105  * osc_cache_shrink_count() will report the LRU page number.
1106  */
1107 unsigned long osc_cache_shrink_scan(struct shrinker *sk,
1108                                     struct shrink_control *sc)
1109 {
1110         struct client_obd *cli;
1111         struct client_obd *stop_anchor = NULL;
1112         struct lu_env *env;
1113         long shrank = 0;
1114         int rc;
1115         __u16 refcheck;
1116
1117         if (sc->nr_to_scan == 0)
1118                 return 0;
1119
1120         if (!(sc->gfp_mask & __GFP_FS))
1121                 return SHRINK_STOP;
1122
1123         env = cl_env_get(&refcheck);
1124         if (IS_ERR(env))
1125                 return SHRINK_STOP;
1126
1127         spin_lock(&osc_shrink_lock);
1128         while ((cli = list_first_entry_or_null(&osc_shrink_list,
1129                                                struct client_obd,
1130                                                cl_shrink_list)) != NULL) {
1131                 if (stop_anchor == NULL)
1132                         stop_anchor = cli;
1133                 else if (cli == stop_anchor)
1134                         break;
1135
1136                 list_move_tail(&cli->cl_shrink_list, &osc_shrink_list);
1137                 spin_unlock(&osc_shrink_lock);
1138
1139                 /* shrink no more than max_pages_per_rpc for an OSC */
1140                 rc = osc_lru_shrink(env, cli, (sc->nr_to_scan - shrank) >
1141                                     cli->cl_max_pages_per_rpc ?
1142                                     cli->cl_max_pages_per_rpc :
1143                                     sc->nr_to_scan - shrank, true);
1144                 if (rc > 0)
1145                         shrank += rc;
1146
1147                 if (shrank >= sc->nr_to_scan)
1148                         goto out;
1149
1150                 spin_lock(&osc_shrink_lock);
1151         }
1152         spin_unlock(&osc_shrink_lock);
1153
1154 out:
1155         cl_env_put(env, &refcheck);
1156
1157         return shrank;
1158 }
1159
1160 /** @} osc */