drivers/staging/lustre/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, 2015, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * Implementation of cl_page for OSC layer.
  33  *
  34  *   Author: Nikita Danilov <nikita.danilov@sun.com>
  35  *   Author: Jinshan Xiong <jinshan.xiong@intel.com>
  36  */
  37
  38 #define DEBUG_SUBSYSTEM S_OSC
  39
  40 #include <linux/math64.h>
  41 #include "osc_cl_internal.h"
  42
  43 static void osc_lru_del(struct client_obd *cli, struct osc_page *opg);
  44 static void osc_lru_use(struct client_obd *cli, struct osc_page *opg);
  45 static int osc_lru_alloc(const struct lu_env *env, struct client_obd *cli,
  46                          struct osc_page *opg);
  47
  48 /** \addtogroup osc
  49  *  @{
  50  */
  51
  52 /*****************************************************************************
  53  *
  54  * Page operations.
  55  *
  56  */
  57 static void osc_page_transfer_get(struct osc_page *opg, const char *label)
  58 {
  59         struct cl_page *page = opg->ops_cl.cpl_page;
  60
  61         LASSERT(!opg->ops_transfer_pinned);
  62         cl_page_get(page);
  63         lu_ref_add_atomic(&page->cp_reference, label, page);
  64         opg->ops_transfer_pinned = 1;
  65 }
  66
  67 static void osc_page_transfer_put(const struct lu_env *env,
  68                                   struct osc_page *opg)
  69 {
  70         struct cl_page *page = opg->ops_cl.cpl_page;
  71
  72         if (opg->ops_transfer_pinned) {
  73                 opg->ops_transfer_pinned = 0;
  74                 lu_ref_del(&page->cp_reference, "transfer", page);
  75                 cl_page_put(env, page);
  76         }
  77 }
  78
  79 /**
  80  * This is called once for every page when it is submitted for a transfer
  81  * either opportunistic (osc_page_cache_add()), or immediate
  82  * (osc_page_submit()).
  83  */
  84 static void osc_page_transfer_add(const struct lu_env *env,
  85                                   struct osc_page *opg, enum cl_req_type crt)
  86 {
  87         struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
  88
  89         osc_lru_use(osc_cli(obj), opg);
  90 }
  91
  92 int osc_page_cache_add(const struct lu_env *env,
  93                        const struct cl_page_slice *slice, struct cl_io *io)
  94 {
  95         struct osc_page *opg = cl2osc_page(slice);
  96         int result;
  97
  98         osc_page_transfer_get(opg, "transfer\0cache");
  99         result = osc_queue_async_io(env, io, opg);
 100         if (result != 0)
 101                 osc_page_transfer_put(env, opg);
 102         else
 103                 osc_page_transfer_add(env, opg, CRT_WRITE);
 104
 105         return result;
 106 }
 107
 108 void osc_index2policy(union ldlm_policy_data *policy,
 109                       const struct cl_object *obj,
 110                       pgoff_t start, pgoff_t end)
 111 {
 112         memset(policy, 0, sizeof(*policy));
 113         policy->l_extent.start = cl_offset(obj, start);
 114         policy->l_extent.end = cl_offset(obj, end + 1) - 1;
 115 }
 116
 117 static const char *osc_list(struct list_head *head)
 118 {
 119         return list_empty(head) ? "-" : "+";
 120 }
 121
 122 static inline unsigned long osc_submit_duration(struct osc_page *opg)
 123 {
 124         if (opg->ops_submit_time == 0)
 125                 return 0;
 126
 127         return (cfs_time_current() - opg->ops_submit_time);
 128 }
 129
 130 static int osc_page_print(const struct lu_env *env,
 131                           const struct cl_page_slice *slice,
 132                           void *cookie, lu_printer_t printer)
 133 {
 134         struct osc_page *opg = cl2osc_page(slice);
 135         struct osc_async_page *oap = &opg->ops_oap;
 136         struct osc_object *obj = cl2osc(slice->cpl_obj);
 137         struct client_obd *cli = &osc_export(obj)->exp_obd->u.cli;
 138
 139         return (*printer)(env, cookie, LUSTRE_OSC_NAME "-page@%p %lu: 1< %#x %d %u %s %s > 2< %llu %u %u %#x %#x | %p %p %p > 3< %d %lu %d > 4< %d %d %d %lu %s | %s %s %s %s > 5< %s %s %s %s | %d %s | %d %s %s>\n",
 140                           opg, osc_index(opg),
 141                           /* 1 */
 142                           oap->oap_magic, oap->oap_cmd,
 143                           oap->oap_interrupted,
 144                           osc_list(&oap->oap_pending_item),
 145                           osc_list(&oap->oap_rpc_item),
 146                           /* 2 */
 147                           oap->oap_obj_off, oap->oap_page_off, oap->oap_count,
 148                           oap->oap_async_flags, oap->oap_brw_flags,
 149                           oap->oap_request, oap->oap_cli, obj,
 150                           /* 3 */
 151                           opg->ops_transfer_pinned,
 152                           osc_submit_duration(opg), opg->ops_srvlock,
 153                           /* 4 */
 154                           cli->cl_r_in_flight, cli->cl_w_in_flight,
 155                           cli->cl_max_rpcs_in_flight,
 156                           cli->cl_avail_grant,
 157                           osc_list(&cli->cl_cache_waiters),
 158                           osc_list(&cli->cl_loi_ready_list),
 159                           osc_list(&cli->cl_loi_hp_ready_list),
 160                           osc_list(&cli->cl_loi_write_list),
 161                           osc_list(&cli->cl_loi_read_list),
 162                           /* 5 */
 163                           osc_list(&obj->oo_ready_item),
 164                           osc_list(&obj->oo_hp_ready_item),
 165                           osc_list(&obj->oo_write_item),
 166                           osc_list(&obj->oo_read_item),
 167                           atomic_read(&obj->oo_nr_reads),
 168                           osc_list(&obj->oo_reading_exts),
 169                           atomic_read(&obj->oo_nr_writes),
 170                           osc_list(&obj->oo_hp_exts),
 171                           osc_list(&obj->oo_urgent_exts));
 172 }
 173
 174 static void osc_page_delete(const struct lu_env *env,
 175                             const struct cl_page_slice *slice)
 176 {
 177         struct osc_page *opg = cl2osc_page(slice);
 178         struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
 179         int rc;
 180
 181         CDEBUG(D_TRACE, "%p\n", opg);
 182         osc_page_transfer_put(env, opg);
 183         rc = osc_teardown_async_page(env, obj, opg);
 184         if (rc) {
 185                 CL_PAGE_DEBUG(D_ERROR, env, slice->cpl_page,
 186                               "Trying to teardown failed: %d\n", rc);
 187                 LASSERT(0);
 188         }
 189
 190         osc_lru_del(osc_cli(obj), opg);
 191
 192         if (slice->cpl_page->cp_type == CPT_CACHEABLE) {
 193                 void *value;
 194
 195                 spin_lock(&obj->oo_tree_lock);
 196                 value = radix_tree_delete(&obj->oo_tree, osc_index(opg));
 197                 if (value)
 198                         --obj->oo_npages;
 199                 spin_unlock(&obj->oo_tree_lock);
 200
 201                 LASSERT(ergo(value, value == opg));
 202         }
 203 }
 204
 205 static void osc_page_clip(const struct lu_env *env,
 206                           const struct cl_page_slice *slice, int from, int to)
 207 {
 208         struct osc_page *opg = cl2osc_page(slice);
 209         struct osc_async_page *oap = &opg->ops_oap;
 210
 211         opg->ops_from = from;
 212         opg->ops_to = to;
 213         spin_lock(&oap->oap_lock);
 214         oap->oap_async_flags |= ASYNC_COUNT_STABLE;
 215         spin_unlock(&oap->oap_lock);
 216 }
 217
 218 static int osc_page_cancel(const struct lu_env *env,
 219                            const struct cl_page_slice *slice)
 220 {
 221         struct osc_page *opg = cl2osc_page(slice);
 222         int rc = 0;
 223
 224         /* Check if the transferring against this page
 225          * is completed, or not even queued.
 226          */
 227         if (opg->ops_transfer_pinned)
 228                 /* FIXME: may not be interrupted.. */
 229                 rc = osc_cancel_async_page(env, opg);
 230         LASSERT(ergo(rc == 0, opg->ops_transfer_pinned == 0));
 231         return rc;
 232 }
 233
 234 static int osc_page_flush(const struct lu_env *env,
 235                           const struct cl_page_slice *slice,
 236                           struct cl_io *io)
 237 {
 238         struct osc_page *opg = cl2osc_page(slice);
 239         int rc;
 240
 241         rc = osc_flush_async_page(env, io, opg);
 242         return rc;
 243 }
 244
 245 static const struct cl_page_operations osc_page_ops = {
 246         .cpo_print       = osc_page_print,
 247         .cpo_delete     = osc_page_delete,
 248         .cpo_clip          = osc_page_clip,
 249         .cpo_cancel      = osc_page_cancel,
 250         .cpo_flush        = osc_page_flush
 251 };
 252
 253 int osc_page_init(const struct lu_env *env, struct cl_object *obj,
 254                   struct cl_page *page, pgoff_t index)
 255 {
 256         struct osc_object *osc = cl2osc(obj);
 257         struct osc_page *opg = cl_object_page_slice(obj, page);
 258         int result;
 259
 260         opg->ops_from = 0;
 261         opg->ops_to = PAGE_SIZE;
 262
 263         result = osc_prep_async_page(osc, opg, page->cp_vmpage,
 264                                      cl_offset(obj, index));
 265         if (result == 0) {
 266                 struct osc_io *oio = osc_env_io(env);
 267
 268                 opg->ops_srvlock = osc_io_srvlock(oio);
 269                 cl_page_slice_add(page, &opg->ops_cl, obj, index,
 270                                   &osc_page_ops);
 271         }
 272         INIT_LIST_HEAD(&opg->ops_lru);
 273
 274         /* reserve an LRU space for this page */
 275         if (page->cp_type == CPT_CACHEABLE && result == 0) {
 276                 result = osc_lru_alloc(env, osc_cli(osc), opg);
 277                 if (result == 0) {
 278                         spin_lock(&osc->oo_tree_lock);
 279                         result = radix_tree_insert(&osc->oo_tree, index, opg);
 280                         if (result == 0)
 281                                 ++osc->oo_npages;
 282                         spin_unlock(&osc->oo_tree_lock);
 283                         LASSERT(result == 0);
 284                 }
 285         }
 286
 287         return result;
 288 }
 289
 290 /**
 291  * Helper function called by osc_io_submit() for every page in an immediate
 292  * transfer (i.e., transferred synchronously).
 293  */
 294 void osc_page_submit(const struct lu_env *env, struct osc_page *opg,
 295                      enum cl_req_type crt, int brw_flags)
 296 {
 297         struct osc_async_page *oap = &opg->ops_oap;
 298
 299         LASSERTF(oap->oap_magic == OAP_MAGIC, "Bad oap magic: oap %p, magic 0x%x\n",
 300                  oap, oap->oap_magic);
 301         LASSERT(oap->oap_async_flags & ASYNC_READY);
 302         LASSERT(oap->oap_async_flags & ASYNC_COUNT_STABLE);
 303
 304         oap->oap_cmd = crt == CRT_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ;
 305         oap->oap_page_off = opg->ops_from;
 306         oap->oap_count = opg->ops_to - opg->ops_from;
 307         oap->oap_brw_flags = brw_flags | OBD_BRW_SYNC;
 308
 309         if (capable(CFS_CAP_SYS_RESOURCE)) {
 310                 oap->oap_brw_flags |= OBD_BRW_NOQUOTA;
 311                 oap->oap_cmd |= OBD_BRW_NOQUOTA;
 312         }
 313
 314         opg->ops_submit_time = cfs_time_current();
 315         osc_page_transfer_get(opg, "transfer\0imm");
 316         osc_page_transfer_add(env, opg, crt);
 317 }
 318
 319 /* --------------- LRU page management ------------------ */
 320
 321 /* OSC is a natural place to manage LRU pages as applications are specialized
 322  * to write OSC by OSC. Ideally, if one OSC is used more frequently it should
 323  * occupy more LRU slots. On the other hand, we should avoid using up all LRU
 324  * slots (client_obd::cl_lru_left) otherwise process has to be put into sleep
 325  * for free LRU slots - this will be very bad so the algorithm requires each
 326  * OSC to free slots voluntarily to maintain a reasonable number of free slots
 327  * at any time.
 328  */
 329 static DECLARE_WAIT_QUEUE_HEAD(osc_lru_waitq);
 330
 331 /**
 332  * LRU pages are freed in batch mode. OSC should at least free this
 333  * number of pages to avoid running out of LRU slots.
 334  */
 335 static inline int lru_shrink_min(struct client_obd *cli)
 336 {
 337         return cli->cl_max_pages_per_rpc * 2;
 338 }
 339
 340 /**
 341  * free this number at most otherwise it will take too long time to finish.
 342  */
 343 static inline int lru_shrink_max(struct client_obd *cli)
 344 {
 345         return cli->cl_max_pages_per_rpc * cli->cl_max_rpcs_in_flight;
 346 }
 347
 348 /**
 349  * Check if we can free LRU slots from this OSC. If there exists LRU waiters,
 350  * we should free slots aggressively. In this way, slots are freed in a steady
 351  * step to maintain fairness among OSCs.
 352  *
 353  * Return how many LRU pages should be freed.
 354  */
 355 static int osc_cache_too_much(struct client_obd *cli)
 356 {
 357         struct cl_client_cache *cache = cli->cl_cache;
 358         long pages = atomic_long_read(&cli->cl_lru_in_list);
 359         unsigned long budget;
 360
 361         budget = cache->ccc_lru_max / (atomic_read(&cache->ccc_users) - 2);
 362
 363         /* if it's going to run out LRU slots, we should free some, but not
 364          * too much to maintain fairness among OSCs.
 365          */
 366         if (atomic_long_read(cli->cl_lru_left) < cache->ccc_lru_max >> 2) {
 367                 if (pages >= budget)
 368                         return lru_shrink_max(cli);
 369                 else if (pages >= budget / 2)
 370                         return lru_shrink_min(cli);
 371         } else {
 372                 time64_t duration = ktime_get_real_seconds();
 373                 long timediff;
 374
 375                 /* knock out pages by duration of no IO activity */
 376                 duration -= cli->cl_lru_last_used;
 377                 /*
 378                  * The difference shouldn't be more than 70 years
 379                  * so we can safely case to a long. Round to
 380                  * approximately 1 minute.
 381                  */
 382                 timediff = (long)(duration >> 6);
 383                 if (timediff > 0 && pages >= budget / timediff)
 384                         return lru_shrink_min(cli);
 385         }
 386         return 0;
 387 }
 388
 389 int lru_queue_work(const struct lu_env *env, void *data)
 390 {
 391         struct client_obd *cli = data;
 392         int count;
 393
 394         CDEBUG(D_CACHE, "%s: run LRU work for client obd\n", cli_name(cli));
 395
 396         count = osc_cache_too_much(cli);
 397         if (count > 0) {
 398                 int rc = osc_lru_shrink(env, cli, count, false);
 399
 400                 CDEBUG(D_CACHE, "%s: shrank %d/%d pages from client obd\n",
 401                        cli_name(cli), rc, count);
 402                 if (rc >= count) {
 403                         CDEBUG(D_CACHE, "%s: queue again\n", cli_name(cli));
 404                         ptlrpcd_queue_work(cli->cl_lru_work);
 405                 }
 406         }
 407
 408         return 0;
 409 }
 410
 411 void osc_lru_add_batch(struct client_obd *cli, struct list_head *plist)
 412 {
 413         LIST_HEAD(lru);
 414         struct osc_async_page *oap;
 415         long npages = 0;
 416
 417         list_for_each_entry(oap, plist, oap_pending_item) {
 418                 struct osc_page *opg = oap2osc_page(oap);
 419
 420                 if (!opg->ops_in_lru)
 421                         continue;
 422
 423                 ++npages;
 424                 LASSERT(list_empty(&opg->ops_lru));
 425                 list_add(&opg->ops_lru, &lru);
 426         }
 427
 428         if (npages > 0) {
 429                 spin_lock(&cli->cl_lru_list_lock);
 430                 list_splice_tail(&lru, &cli->cl_lru_list);
 431                 atomic_long_sub(npages, &cli->cl_lru_busy);
 432                 atomic_long_add(npages, &cli->cl_lru_in_list);
 433                 cli->cl_lru_last_used = ktime_get_real_seconds();
 434                 spin_unlock(&cli->cl_lru_list_lock);
 435
 436                 if (waitqueue_active(&osc_lru_waitq))
 437                         (void)ptlrpcd_queue_work(cli->cl_lru_work);
 438         }
 439 }
 440
 441 static void __osc_lru_del(struct client_obd *cli, struct osc_page *opg)
 442 {
 443         LASSERT(atomic_long_read(&cli->cl_lru_in_list) > 0);
 444         list_del_init(&opg->ops_lru);
 445         atomic_long_dec(&cli->cl_lru_in_list);
 446 }
 447
 448 /**
 449  * Page is being destroyed. The page may be not in LRU list, if the transfer
 450  * has never finished(error occurred).
 451  */
 452 static void osc_lru_del(struct client_obd *cli, struct osc_page *opg)
 453 {
 454         if (opg->ops_in_lru) {
 455                 spin_lock(&cli->cl_lru_list_lock);
 456                 if (!list_empty(&opg->ops_lru)) {
 457                         __osc_lru_del(cli, opg);
 458                 } else {
 459                         LASSERT(atomic_long_read(&cli->cl_lru_busy) > 0);
 460                         atomic_long_dec(&cli->cl_lru_busy);
 461                 }
 462                 spin_unlock(&cli->cl_lru_list_lock);
 463
 464                 atomic_long_inc(cli->cl_lru_left);
 465                 /* this is a great place to release more LRU pages if
 466                  * this osc occupies too many LRU pages and kernel is
 467                  * stealing one of them.
 468                  */
 469                 if (osc_cache_too_much(cli)) {
 470                         CDEBUG(D_CACHE, "%s: queue LRU work\n", cli_name(cli));
 471                         (void)ptlrpcd_queue_work(cli->cl_lru_work);
 472                 }
 473                 wake_up(&osc_lru_waitq);
 474         } else {
 475                 LASSERT(list_empty(&opg->ops_lru));
 476         }
 477 }
 478
 479 /**
 480  * Delete page from LRUlist for redirty.
 481  */
 482 static void osc_lru_use(struct client_obd *cli, struct osc_page *opg)
 483 {
 484         /* If page is being transferred for the first time,
 485          * ops_lru should be empty
 486          */
 487         if (opg->ops_in_lru && !list_empty(&opg->ops_lru)) {
 488                 spin_lock(&cli->cl_lru_list_lock);
 489                 __osc_lru_del(cli, opg);
 490                 spin_unlock(&cli->cl_lru_list_lock);
 491                 atomic_long_inc(&cli->cl_lru_busy);
 492         }
 493 }
 494
 495 static void discard_pagevec(const struct lu_env *env, struct cl_io *io,
 496                             struct cl_page **pvec, int max_index)
 497 {
 498         int i;
 499
 500         for (i = 0; i < max_index; i++) {
 501                 struct cl_page *page = pvec[i];
 502
 503                 LASSERT(cl_page_is_owned(page, io));
 504                 cl_page_delete(env, page);
 505                 cl_page_discard(env, io, page);
 506                 cl_page_disown(env, io, page);
 507                 cl_page_put(env, page);
 508
 509                 pvec[i] = NULL;
 510         }
 511 }
 512
 513 /**
 514  * Check if a cl_page can be released, i.e, it's not being used.
 515  *
 516  * If unstable account is turned on, bulk transfer may hold one refcount
 517  * for recovery so we need to check vmpage refcount as well; otherwise,
 518  * even we can destroy cl_page but the corresponding vmpage can't be reused.
 519  */
 520 static inline bool lru_page_busy(struct client_obd *cli, struct cl_page *page)
 521 {
 522         if (cl_page_in_use_noref(page))
 523                 return true;
 524
 525         if (cli->cl_cache->ccc_unstable_check) {
 526                 struct page *vmpage = cl_page_vmpage(page);
 527
 528                 /* vmpage have two known users: cl_page and VM page cache */
 529                 if (page_count(vmpage) - page_mapcount(vmpage) > 2)
 530                         return true;
 531         }
 532         return false;
 533 }
 534
 535 /**
 536  * Drop @target of pages from LRU at most.
 537  */
 538 long osc_lru_shrink(const struct lu_env *env, struct client_obd *cli,
 539                     long target, bool force)
 540 {
 541         struct cl_io *io;
 542         struct cl_object *clobj = NULL;
 543         struct cl_page **pvec;
 544         struct osc_page *opg;
 545         int maxscan = 0;
 546         long count = 0;
 547         int index = 0;
 548         int rc = 0;
 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_info(env)->oti_io;
 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_entry(cli->cl_lru_list.next, 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);
 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) {
 601                                 discard_pagevec(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                                  */
 628                                 __osc_lru_del(cli, opg);
 629                                 opg->ops_in_lru = 0; /* will be discarded */
 630
 631                                 cl_page_get(page);
 632                                 will_free = true;
 633                         } else {
 634                                 cl_page_disown(env, io, page);
 635                         }
 636                 }
 637
 638                 if (!will_free) {
 639                         list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
 640                         continue;
 641                 }
 642
 643                 /* Don't discard and free the page with cl_lru_list held */
 644                 pvec[index++] = page;
 645                 if (unlikely(index == OTI_PVEC_SIZE)) {
 646                         spin_unlock(&cli->cl_lru_list_lock);
 647                         discard_pagevec(env, io, pvec, index);
 648                         index = 0;
 649
 650                         spin_lock(&cli->cl_lru_list_lock);
 651                 }
 652
 653                 if (++count >= target)
 654                         break;
 655         }
 656         spin_unlock(&cli->cl_lru_list_lock);
 657
 658         if (clobj) {
 659                 discard_pagevec(env, io, pvec, index);
 660
 661                 cl_io_fini(env, io);
 662                 cl_object_put(env, clobj);
 663         }
 664
 665         atomic_dec(&cli->cl_lru_shrinkers);
 666         if (count > 0) {
 667                 atomic_long_add(count, cli->cl_lru_left);
 668                 wake_up_all(&osc_lru_waitq);
 669         }
 670         return count > 0 ? count : rc;
 671 }
 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         int max_scans;
 684         u16 refcheck;
 685         long rc = 0;
 686
 687         LASSERT(cache);
 688
 689         env = cl_env_get(&refcheck);
 690         if (IS_ERR(env))
 691                 return 0;
 692
 693         npages = max_t(int, npages, cli->cl_max_pages_per_rpc);
 694         CDEBUG(D_CACHE, "%s: start to reclaim %ld pages from LRU\n",
 695                cli_name(cli), npages);
 696         rc = osc_lru_shrink(env, cli, npages, true);
 697         if (rc >= npages) {
 698                 CDEBUG(D_CACHE, "%s: reclaimed %ld/%ld pages from LRU\n",
 699                        cli_name(cli), rc, npages);
 700                 if (osc_cache_too_much(cli) > 0)
 701                         ptlrpcd_queue_work(cli->cl_lru_work);
 702                 goto out;
 703         } else if (rc > 0) {
 704                 npages -= rc;
 705         }
 706
 707         CDEBUG(D_CACHE, "%s: cli %p no free slots, pages: %ld/%ld, want: %ld\n",
 708                cli_name(cli), cli, atomic_long_read(&cli->cl_lru_in_list),
 709                atomic_long_read(&cli->cl_lru_busy), npages);
 710
 711         /* Reclaim LRU slots from other client_obd as it can't free enough
 712          * from its own. This should rarely happen.
 713          */
 714         spin_lock(&cache->ccc_lru_lock);
 715         LASSERT(!list_empty(&cache->ccc_lru));
 716
 717         cache->ccc_lru_shrinkers++;
 718         list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
 719
 720         max_scans = atomic_read(&cache->ccc_users) - 2;
 721         while (--max_scans > 0 && !list_empty(&cache->ccc_lru)) {
 722                 cli = list_entry(cache->ccc_lru.next, struct client_obd,
 723                                  cl_lru_osc);
 724
 725                 CDEBUG(D_CACHE, "%s: cli %p LRU pages: %ld, busy: %ld.\n",
 726                        cli_name(cli), cli,
 727                        atomic_long_read(&cli->cl_lru_in_list),
 728                        atomic_long_read(&cli->cl_lru_busy));
 729
 730                 list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
 731                 if (osc_cache_too_much(cli) > 0) {
 732                         spin_unlock(&cache->ccc_lru_lock);
 733
 734                         rc = osc_lru_shrink(env, cli, npages, true);
 735                         spin_lock(&cache->ccc_lru_lock);
 736                         if (rc >= npages)
 737                                 break;
 738                         if (rc > 0)
 739                                 npages -= rc;
 740                 }
 741         }
 742         spin_unlock(&cache->ccc_lru_lock);
 743
 744 out:
 745         cl_env_put(env, &refcheck);
 746         CDEBUG(D_CACHE, "%s: cli %p freed %ld pages.\n",
 747                cli_name(cli), cli, rc);
 748         return rc;
 749 }
 750
 751 /**
 752  * osc_lru_alloc() is called to reserve an LRU slot for a cl_page.
 753  *
 754  * Usually the LRU slots are reserved in osc_io_iter_rw_init().
 755  * Only in the case that the LRU slots are in extreme shortage, it should
 756  * have reserved enough slots for an IO.
 757  */
 758 static int osc_lru_alloc(const struct lu_env *env, struct client_obd *cli,
 759                          struct osc_page *opg)
 760 {
 761         struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
 762         struct osc_io *oio = osc_env_io(env);
 763         int rc = 0;
 764
 765         if (!cli->cl_cache) /* shall not be in LRU */
 766                 return 0;
 767
 768         if (oio->oi_lru_reserved > 0) {
 769                 --oio->oi_lru_reserved;
 770                 goto out;
 771         }
 772
 773         LASSERT(atomic_long_read(cli->cl_lru_left) >= 0);
 774         while (!atomic_long_add_unless(cli->cl_lru_left, -1, 0)) {
 775                 /* run out of LRU spaces, try to drop some by itself */
 776                 rc = osc_lru_reclaim(cli, 1);
 777                 if (rc < 0)
 778                         break;
 779                 if (rc > 0)
 780                         continue;
 781
 782                 cond_resched();
 783
 784                 rc = l_wait_event(osc_lru_waitq,
 785                                   atomic_long_read(cli->cl_lru_left) > 0,
 786                                   &lwi);
 787
 788                 if (rc < 0)
 789                         break;
 790         }
 791
 792 out:
 793         if (rc >= 0) {
 794                 atomic_long_inc(&cli->cl_lru_busy);
 795                 opg->ops_in_lru = 1;
 796                 rc = 0;
 797         }
 798
 799         return rc;
 800 }
 801
 802 /**
 803  * osc_lru_reserve() is called to reserve enough LRU slots for I/O.
 804  *
 805  * The benefit of doing this is to reduce contention against atomic counter
 806  * cl_lru_left by changing it from per-page access to per-IO access.
 807  */
 808 unsigned long osc_lru_reserve(struct client_obd *cli, unsigned long npages)
 809 {
 810         unsigned long reserved = 0;
 811         unsigned long max_pages;
 812         unsigned long c;
 813
 814         /*
 815          * reserve a full RPC window at most to avoid that a thread accidentally
 816          * consumes too many LRU slots
 817          */
 818         max_pages = cli->cl_max_pages_per_rpc * cli->cl_max_rpcs_in_flight;
 819         if (npages > max_pages)
 820                 npages = max_pages;
 821
 822         c = atomic_long_read(cli->cl_lru_left);
 823         if (c < npages && osc_lru_reclaim(cli, npages) > 0)
 824                 c = atomic_long_read(cli->cl_lru_left);
 825         while (c >= npages) {
 826                 if (c == atomic_long_cmpxchg(cli->cl_lru_left, c, c - npages)) {
 827                         reserved = npages;
 828                         break;
 829                 }
 830                 c = atomic_long_read(cli->cl_lru_left);
 831         }
 832         if (atomic_long_read(cli->cl_lru_left) < max_pages) {
 833                 /*
 834                  * If there aren't enough pages in the per-OSC LRU then
 835                  * wake up the LRU thread to try and clear out space, so
 836                  * we don't block if pages are being dirtied quickly.
 837                  */
 838                 CDEBUG(D_CACHE, "%s: queue LRU, left: %lu/%ld.\n",
 839                        cli_name(cli), atomic_long_read(cli->cl_lru_left),
 840                        max_pages);
 841                 (void)ptlrpcd_queue_work(cli->cl_lru_work);
 842         }
 843
 844         return reserved;
 845 }
 846
 847 /**
 848  * osc_lru_unreserve() is called to unreserve LRU slots.
 849  *
 850  * LRU slots reserved by osc_lru_reserve() may have entries left due to several
 851  * reasons such as page already existing or I/O error. Those reserved slots
 852  * should be freed by calling this function.
 853  */
 854 void osc_lru_unreserve(struct client_obd *cli, unsigned long npages)
 855 {
 856         atomic_long_add(npages, cli->cl_lru_left);
 857         wake_up_all(&osc_lru_waitq);
 858 }
 859
 860 /**
 861  * Atomic operations are expensive. We accumulate the accounting for the
 862  * same page pgdat to get better performance.
 863  * In practice this can work pretty good because the pages in the same RPC
 864  * are likely from the same page zone.
 865  */
 866 static inline void unstable_page_accounting(struct ptlrpc_bulk_desc *desc,
 867                                             int factor)
 868 {
 869         int page_count = desc->bd_iov_count;
 870         pg_data_t *last = NULL;
 871         int count = 0;
 872         int i;
 873
 874         LASSERT(ptlrpc_is_bulk_desc_kiov(desc->bd_type));
 875
 876         for (i = 0; i < page_count; i++) {
 877                 pg_data_t *pgdat = page_pgdat(BD_GET_KIOV(desc, i).bv_page);
 878
 879                 if (likely(pgdat == last)) {
 880                         ++count;
 881                         continue;
 882                 }
 883
 884                 if (count > 0) {
 885                         mod_node_page_state(pgdat, NR_UNSTABLE_NFS,
 886                                             factor * count);
 887                         count = 0;
 888                 }
 889                 last = pgdat;
 890                 ++count;
 891         }
 892         if (count > 0)
 893                 mod_node_page_state(last, NR_UNSTABLE_NFS, factor * count);
 894 }
 895
 896 static inline void add_unstable_page_accounting(struct ptlrpc_bulk_desc *desc)
 897 {
 898         unstable_page_accounting(desc, 1);
 899 }
 900
 901 static inline void dec_unstable_page_accounting(struct ptlrpc_bulk_desc *desc)
 902 {
 903         unstable_page_accounting(desc, -1);
 904 }
 905
 906 /**
 907  * Performs "unstable" page accounting. This function balances the
 908  * increment operations performed in osc_inc_unstable_pages. It is
 909  * registered as the RPC request callback, and is executed when the
 910  * bulk RPC is committed on the server. Thus at this point, the pages
 911  * involved in the bulk transfer are no longer considered unstable.
 912  *
 913  * If this function is called, the request should have been committed
 914  * or req:rq_unstable must have been set; it implies that the unstable
 915  * statistic have been added.
 916  */
 917 void osc_dec_unstable_pages(struct ptlrpc_request *req)
 918 {
 919         struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
 920         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
 921         int page_count = desc->bd_iov_count;
 922         long unstable_count;
 923
 924         LASSERT(page_count >= 0);
 925         dec_unstable_page_accounting(desc);
 926
 927         unstable_count = atomic_long_sub_return(page_count,
 928                                                 &cli->cl_unstable_count);
 929         LASSERT(unstable_count >= 0);
 930
 931         unstable_count = atomic_long_sub_return(page_count,
 932                                                 &cli->cl_cache->ccc_unstable_nr);
 933         LASSERT(unstable_count >= 0);
 934         if (!unstable_count)
 935                 wake_up_all(&cli->cl_cache->ccc_unstable_waitq);
 936
 937         if (waitqueue_active(&osc_lru_waitq))
 938                 (void)ptlrpcd_queue_work(cli->cl_lru_work);
 939 }
 940
 941 /**
 942  * "unstable" page accounting. See: osc_dec_unstable_pages.
 943  */
 944 void osc_inc_unstable_pages(struct ptlrpc_request *req)
 945 {
 946         struct client_obd *cli  = &req->rq_import->imp_obd->u.cli;
 947         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
 948         long page_count = desc->bd_iov_count;
 949
 950         /* No unstable page tracking */
 951         if (!cli->cl_cache || !cli->cl_cache->ccc_unstable_check)
 952                 return;
 953
 954         add_unstable_page_accounting(desc);
 955         atomic_long_add(page_count, &cli->cl_unstable_count);
 956         atomic_long_add(page_count, &cli->cl_cache->ccc_unstable_nr);
 957
 958         /*
 959          * If the request has already been committed (i.e. brw_commit
 960          * called via rq_commit_cb), we need to undo the unstable page
 961          * increments we just performed because rq_commit_cb wont be
 962          * called again.
 963          */
 964         spin_lock(&req->rq_lock);
 965         if (unlikely(req->rq_committed)) {
 966                 spin_unlock(&req->rq_lock);
 967
 968                 osc_dec_unstable_pages(req);
 969         } else {
 970                 req->rq_unstable = 1;
 971                 spin_unlock(&req->rq_lock);
 972         }
 973 }
 974
 975 /**
 976  * Check if it piggybacks SOFT_SYNC flag to OST from this OSC.
 977  * This function will be called by every BRW RPC so it's critical
 978  * to make this function fast.
 979  */
 980 bool osc_over_unstable_soft_limit(struct client_obd *cli)
 981 {
 982         long unstable_nr, osc_unstable_count;
 983
 984         /* Can't check cli->cl_unstable_count, therefore, no soft limit */
 985         if (!cli->cl_cache || !cli->cl_cache->ccc_unstable_check)
 986                 return false;
 987
 988         osc_unstable_count = atomic_long_read(&cli->cl_unstable_count);
 989         unstable_nr = atomic_long_read(&cli->cl_cache->ccc_unstable_nr);
 990
 991         CDEBUG(D_CACHE,
 992                "%s: cli: %p unstable pages: %lu, osc unstable pages: %lu\n",
 993                cli_name(cli), cli, unstable_nr, osc_unstable_count);
 994
 995         /*
 996          * If the LRU slots are in shortage - 25% remaining AND this OSC
 997          * has one full RPC window of unstable pages, it's a good chance
 998          * to piggyback a SOFT_SYNC flag.
 999          * Please notice that the OST won't take immediate response for the
1000          * SOFT_SYNC request so active OSCs will have more chance to carry
1001          * the flag, this is reasonable.
1002          */
1003         return unstable_nr > cli->cl_cache->ccc_lru_max >> 2 &&
1004                osc_unstable_count > cli->cl_max_pages_per_rpc *
1005                                     cli->cl_max_rpcs_in_flight;
1006 }
1007
1008 /**
1009  * Return how many LRU pages in the cache of all OSC devices
1010  *
1011  * Return:      return # of cached LRU pages times reclaimation tendency
1012  *              SHRINK_STOP if it cannot do any scanning in this time
1013  */
1014 unsigned long osc_cache_shrink_count(struct shrinker *sk,
1015                                      struct shrink_control *sc)
1016 {
1017         struct client_obd *cli;
1018         unsigned long cached = 0;
1019
1020         spin_lock(&osc_shrink_lock);
1021         list_for_each_entry(cli, &osc_shrink_list, cl_shrink_list)
1022                 cached += atomic_long_read(&cli->cl_lru_in_list);
1023         spin_unlock(&osc_shrink_lock);
1024
1025         return (cached  * sysctl_vfs_cache_pressure) / 100;
1026 }
1027
1028 /**
1029  * Scan and try to reclaim sc->nr_to_scan cached LRU pages
1030  *
1031  * Return:      number of cached LRU pages reclaimed
1032  *              SHRINK_STOP if it cannot do any scanning in this time
1033  *
1034  * Linux kernel will loop calling this shrinker scan routine with
1035  * sc->nr_to_scan = SHRINK_BATCH(128 for now) until kernel got enough memory.
1036  *
1037  * If sc->nr_to_scan is 0, the VM is querying the cache size, we don't need
1038  * to scan and try to reclaim LRU pages, just return 0 and
1039  * osc_cache_shrink_count() will report the LRU page number.
1040  */
1041 unsigned long osc_cache_shrink_scan(struct shrinker *sk,
1042                                     struct shrink_control *sc)
1043 {
1044         struct client_obd *stop_anchor = NULL;
1045         struct client_obd *cli;
1046         struct lu_env *env;
1047         long shrank = 0;
1048         u16 refcheck;
1049         int rc;
1050
1051         if (!sc->nr_to_scan)
1052                 return 0;
1053
1054         if (!(sc->gfp_mask & __GFP_FS))
1055                 return SHRINK_STOP;
1056
1057         env = cl_env_get(&refcheck);
1058         if (IS_ERR(env))
1059                 return SHRINK_STOP;
1060
1061         spin_lock(&osc_shrink_lock);
1062         while (!list_empty(&osc_shrink_list)) {
1063                 cli = list_entry(osc_shrink_list.next, struct client_obd,
1064                                  cl_shrink_list);
1065
1066                 if (!stop_anchor)
1067                         stop_anchor = cli;
1068                 else if (cli == stop_anchor)
1069                         break;
1070
1071                 list_move_tail(&cli->cl_shrink_list, &osc_shrink_list);
1072                 spin_unlock(&osc_shrink_lock);
1073
1074                 /* shrink no more than max_pages_per_rpc for an OSC */
1075                 rc = osc_lru_shrink(env, cli, (sc->nr_to_scan - shrank) >
1076                                     cli->cl_max_pages_per_rpc ?
1077                                     cli->cl_max_pages_per_rpc :
1078                                     sc->nr_to_scan - shrank, true);
1079                 if (rc > 0)
1080                         shrank += rc;
1081
1082                 if (shrank >= sc->nr_to_scan)
1083                         goto out;
1084
1085                 spin_lock(&osc_shrink_lock);
1086         }
1087         spin_unlock(&osc_shrink_lock);
1088
1089 out:
1090         cl_env_put(env, &refcheck);
1091
1092         return shrank;
1093 }
1094
1095 /** @} osc */