lustre/osc/osc_page.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  *
  31  * Implementation of cl_page for OSC layer.
  32  *
  33  *   Author: Nikita Danilov <nikita.danilov@sun.com>
  34  *   Author: Jinshan Xiong <jinshan.xiong@intel.com>
  35  */
  36
  37 #define DEBUG_SUBSYSTEM S_OSC
  38 #include <lustre_osc.h>
  39
  40 #include "osc_internal.h"
  41
  42 static void osc_lru_del(struct client_obd *cli, struct osc_page *opg);
  43 static void osc_lru_use(struct client_obd *cli, struct osc_page *opg);
  44 static int osc_lru_alloc(const struct lu_env *env, struct client_obd *cli,
  45                          struct osc_page *opg);
  46
  47 /** \addtogroup osc
  48  *  @{
  49  */
  50
  51 /*
  52  * Page operations.
  53  */
  54 static void osc_page_transfer_get(struct osc_page *opg, const char *label)
  55 {
  56         struct cl_page *page = opg->ops_cl.cpl_page;
  57
  58         LASSERT(!opg->ops_transfer_pinned);
  59         cl_page_get(page);
  60         lu_ref_add_atomic(&page->cp_reference, label, page);
  61         opg->ops_transfer_pinned = 1;
  62 }
  63
  64 static void osc_page_transfer_put(const struct lu_env *env,
  65                                   struct osc_page *opg)
  66 {
  67         struct cl_page *page = opg->ops_cl.cpl_page;
  68
  69         if (opg->ops_transfer_pinned) {
  70                 opg->ops_transfer_pinned = 0;
  71                 lu_ref_del(&page->cp_reference, "transfer", page);
  72                 cl_page_put(env, page);
  73         }
  74 }
  75
  76 /**
  77  * This is called once for every page when it is submitted for a transfer
  78  * either opportunistic (osc_page_cache_add()), or immediate
  79  * (osc_page_submit()).
  80  */
  81 static void osc_page_transfer_add(const struct lu_env *env,
  82                                   struct osc_page *opg, enum cl_req_type crt)
  83 {
  84         struct osc_object *obj = osc_page_object(opg);
  85
  86         osc_lru_use(osc_cli(obj), opg);
  87 }
  88
  89 int osc_page_cache_add(const struct lu_env *env, struct osc_page *opg,
  90                        struct cl_io *io, cl_commit_cbt cb)
  91 {
  92         int result;
  93         ENTRY;
  94
  95         osc_page_transfer_get(opg, "transfer\0cache");
  96         result = osc_queue_async_io(env, io, opg, cb);
  97         if (result != 0)
  98                 osc_page_transfer_put(env, opg);
  99         else
 100                 osc_page_transfer_add(env, opg, CRT_WRITE);
 101
 102         RETURN(result);
 103 }
 104
 105 void osc_index2policy(union ldlm_policy_data *policy,
 106                       const struct cl_object *obj, pgoff_t start, pgoff_t end)
 107 {
 108         memset(policy, 0, sizeof *policy);
 109         policy->l_extent.start = cl_offset(obj, start);
 110         policy->l_extent.end   = cl_offset(obj, end + 1) - 1;
 111 }
 112
 113 static inline s64 osc_submit_duration(struct osc_page *opg)
 114 {
 115         if (ktime_to_ns(opg->ops_submit_time) == 0)
 116                 return 0;
 117
 118         return ktime_ms_delta(ktime_get(), opg->ops_submit_time);
 119 }
 120
 121 static int osc_page_print(const struct lu_env *env,
 122                           const struct cl_page_slice *slice,
 123                           void *cookie, lu_printer_t printer)
 124 {
 125         struct osc_page *opg = cl2osc_page(slice);
 126         struct osc_async_page *oap = &opg->ops_oap;
 127         struct osc_object *obj = osc_page_object(opg);
 128         struct client_obd *cli = &osc_export(obj)->exp_obd->u.cli;
 129
 130         return (*printer)(env, cookie, LUSTRE_OSC_NAME"-page@%p %lu: "
 131                           "1< %#x %d %c %c > "
 132                           "2< %lld %u %u %#x %#x | %p %p %p > "
 133                           "3< %d %lld %d > "
 134                           "4< %d %d %d %lu %c | %c %c %c %c > "
 135                           "5< %c %c %c %c | %d %c | %d %c %c>\n",
 136                           opg, osc_index(opg),
 137                           /* 1 */
 138                           oap->oap_magic, oap->oap_cmd,
 139                           list_empty_marker(&oap->oap_pending_item),
 140                           list_empty_marker(&oap->oap_rpc_item),
 141                           /* 2 */
 142                           oap->oap_obj_off, oap->oap_page_off, oap->oap_count,
 143                           oap->oap_async_flags, oap->oap_brw_flags,
 144                           oap->oap_request, cli, obj,
 145                           /* 3 */
 146                           opg->ops_transfer_pinned,
 147                           osc_submit_duration(opg), opg->ops_srvlock,
 148                           /* 4 */
 149                           cli->cl_r_in_flight, cli->cl_w_in_flight,
 150                           cli->cl_max_rpcs_in_flight,
 151                           cli->cl_avail_grant,
 152                           waitqueue_active(&cli->cl_cache_waiters) ? '+' : '-',
 153                           list_empty_marker(&cli->cl_loi_ready_list),
 154                           list_empty_marker(&cli->cl_loi_hp_ready_list),
 155                           list_empty_marker(&cli->cl_loi_write_list),
 156                           list_empty_marker(&cli->cl_loi_read_list),
 157                           /* 5 */
 158                           list_empty_marker(&obj->oo_ready_item),
 159                           list_empty_marker(&obj->oo_hp_ready_item),
 160                           list_empty_marker(&obj->oo_write_item),
 161                           list_empty_marker(&obj->oo_read_item),
 162                           atomic_read(&obj->oo_nr_reads),
 163                           list_empty_marker(&obj->oo_reading_exts),
 164                           atomic_read(&obj->oo_nr_writes),
 165                           list_empty_marker(&obj->oo_hp_exts),
 166                           list_empty_marker(&obj->oo_urgent_exts));
 167 }
 168
 169 static void osc_page_delete(const struct lu_env *env,
 170                             const struct cl_page_slice *slice)
 171 {
 172         struct osc_page   *opg = cl2osc_page(slice);
 173         struct osc_object *obj = osc_page_object(opg);
 174         int rc;
 175
 176         ENTRY;
 177         CDEBUG(D_TRACE, "%p\n", opg);
 178         osc_page_transfer_put(env, opg);
 179         rc = osc_teardown_async_page(env, obj, opg);
 180         if (rc) {
 181                 CL_PAGE_DEBUG(D_ERROR, env, slice->cpl_page,
 182                               "Trying to teardown failed: %d\n", rc);
 183                 LASSERT(0);
 184         }
 185
 186         osc_lru_del(osc_cli(obj), opg);
 187
 188         if (slice->cpl_page->cp_type == CPT_CACHEABLE) {
 189                 void *value = NULL;
 190
 191                 spin_lock(&obj->oo_tree_lock);
 192                 if (opg->ops_intree) {
 193                         value = radix_tree_delete(&obj->oo_tree,
 194                                                   osc_index(opg));
 195                         if (value != NULL) {
 196                                 --obj->oo_npages;
 197                                 opg->ops_intree = 0;
 198                         }
 199                 }
 200                 spin_unlock(&obj->oo_tree_lock);
 201
 202                 LASSERT(ergo(value != NULL, value == opg));
 203         }
 204
 205         EXIT;
 206 }
 207
 208 static void osc_page_clip(const struct lu_env *env,
 209                           const struct cl_page_slice *slice,
 210                           int from, int to)
 211 {
 212         struct osc_page       *opg = cl2osc_page(slice);
 213         struct osc_async_page *oap = &opg->ops_oap;
 214
 215         opg->ops_from = from;
 216         /* argument @to is exclusive, but @ops_to is inclusive */
 217         opg->ops_to   = to - 1;
 218         /* This isn't really necessary for transient pages, but we also don't
 219          * call clip on transient pages often, so it's OK.
 220          */
 221         spin_lock(&oap->oap_lock);
 222         oap->oap_async_flags |= ASYNC_COUNT_STABLE;
 223         spin_unlock(&oap->oap_lock);
 224 }
 225
 226 static int osc_page_flush(const struct lu_env *env,
 227                           const struct cl_page_slice *slice,
 228                           struct cl_io *io)
 229 {
 230         struct osc_page *opg = cl2osc_page(slice);
 231         int rc = 0;
 232         ENTRY;
 233         rc = osc_flush_async_page(env, io, opg);
 234         RETURN(rc);
 235 }
 236
 237 static void osc_page_touch(const struct lu_env *env,
 238                           const struct cl_page_slice *slice, size_t to)
 239 {
 240         struct osc_page *opg = cl2osc_page(slice);
 241         struct cl_object *obj = osc2cl(osc_page_object(opg));
 242
 243         osc_page_touch_at(env, obj, osc_index(opg), to);
 244 }
 245
 246 static const struct cl_page_operations osc_page_ops = {
 247         .cpo_print         = osc_page_print,
 248         .cpo_delete        = osc_page_delete,
 249         .cpo_clip           = osc_page_clip,
 250         .cpo_flush          = osc_page_flush,
 251         .cpo_page_touch    = osc_page_touch,
 252 };
 253
 254 int osc_page_init(const struct lu_env *env, struct cl_object *obj,
 255                   struct cl_page *cl_page, pgoff_t index)
 256 {
 257         struct osc_object *osc = cl2osc(obj);
 258         struct osc_page *opg = cl_object_page_slice(obj, cl_page);
 259         struct osc_io *oio = osc_env_io(env);
 260         int result;
 261
 262         opg->ops_from = 0;
 263         opg->ops_to = PAGE_SIZE - 1;
 264
 265         INIT_LIST_HEAD(&opg->ops_lru);
 266
 267         result = osc_prep_async_page(osc, opg, cl_page, cl_offset(obj, index));
 268         if (result != 0)
 269                 return result;
 270
 271         opg->ops_srvlock = osc_io_srvlock(oio);
 272         cl_page_slice_add(cl_page, &opg->ops_cl, obj, &osc_page_ops);
 273
 274         /* reserve an LRU space for this page */
 275         if (cl_page->cp_type == CPT_CACHEABLE) {
 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, ktime_t submit_time)
 304 {
 305         struct osc_io *oio = osc_env_io(env);
 306         struct osc_async_page *oap = &opg->ops_oap;
 307
 308         LASSERTF(oap->oap_magic == OAP_MAGIC, "Bad oap magic: oap %p, "
 309                  "magic 0x%x\n", oap, oap->oap_magic);
 310         LASSERT(oap->oap_async_flags & ASYNC_READY);
 311         LASSERT(oap->oap_async_flags & ASYNC_COUNT_STABLE);
 312
 313         oap->oap_cmd = crt == CRT_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ;
 314         oap->oap_page_off = opg->ops_from;
 315         oap->oap_count = opg->ops_to - opg->ops_from + 1;
 316         oap->oap_brw_flags = OBD_BRW_SYNC | brw_flags;
 317
 318         if (oio->oi_cap_sys_resource) {
 319                 oap->oap_brw_flags |= OBD_BRW_SYS_RESOURCE;
 320                 oap->oap_cmd |= OBD_BRW_SYS_RESOURCE;
 321         }
 322
 323         opg->ops_submit_time = submit_time;
 324         osc_page_transfer_get(opg, "transfer\0imm");
 325         osc_page_transfer_add(env, opg, crt);
 326 }
 327
 328 /* --------------- LRU page management ------------------ */
 329
 330 /* OSC is a natural place to manage LRU pages as applications are specialized
 331  * to write OSC by OSC. Ideally, if one OSC is used more frequently it should
 332  * occupy more LRU slots. On the other hand, we should avoid using up all LRU
 333  * slots (client_obd::cl_lru_left) otherwise process has to be put into sleep
 334  * for free LRU slots - this will be very bad so the algorithm requires each
 335  * OSC to free slots voluntarily to maintain a reasonable number of free slots
 336  * at any time.
 337  */
 338
 339 static DECLARE_WAIT_QUEUE_HEAD(osc_lru_waitq);
 340
 341 /**
 342  * LRU pages are freed in batch mode. OSC should at least free this
 343  * number of pages to avoid running out of LRU slots.
 344  */
 345 static inline int lru_shrink_min(struct client_obd *cli)
 346 {
 347         return cli->cl_max_pages_per_rpc * 2;
 348 }
 349
 350 /**
 351  * free this number at most otherwise it will take too long time to finsih.
 352  */
 353 static inline int lru_shrink_max(struct client_obd *cli)
 354 {
 355         return cli->cl_max_pages_per_rpc * cli->cl_max_rpcs_in_flight;
 356 }
 357
 358 /**
 359  * Check if we can free LRU slots from this OSC. If there exists LRU waiters,
 360  * we should free slots aggressively. In this way, slots are freed in a steady
 361  * step to maintain fairness among OSCs.
 362  *
 363  * Return how many LRU pages should be freed.
 364  */
 365 static int osc_cache_too_much(struct client_obd *cli)
 366 {
 367         struct cl_client_cache *cache = cli->cl_cache;
 368         long pages = atomic_long_read(&cli->cl_lru_in_list);
 369         unsigned long budget;
 370
 371         LASSERT(cache != NULL);
 372         budget = cache->ccc_lru_max / (atomic_read(&cache->ccc_users) - 2);
 373
 374         /* if it's going to run out LRU slots, we should free some, but not
 375          * too much to maintain faireness among OSCs. */
 376         if (atomic_long_read(cli->cl_lru_left) < cache->ccc_lru_max >> 2) {
 377                 if (pages >= budget)
 378                         return lru_shrink_max(cli);
 379                 else if (pages >= budget / 2)
 380                         return lru_shrink_min(cli);
 381         } else {
 382                 time64_t duration = ktime_get_real_seconds();
 383                 long timediff;
 384
 385                 /* knock out pages by duration of no IO activity */
 386                 duration -= cli->cl_lru_last_used;
 387                 /*
 388                  * The difference shouldn't be more than 70 years
 389                  * so we can safely case to a long. Round to
 390                  * approximately 1 minute.
 391                  */
 392                 timediff = (long)(duration >> 6);
 393                 if (timediff > 0 && pages >= budget / timediff)
 394                         return lru_shrink_min(cli);
 395         }
 396         return 0;
 397 }
 398
 399 int lru_queue_work(const struct lu_env *env, void *data)
 400 {
 401         struct client_obd *cli = data;
 402         int count;
 403
 404         CDEBUG(D_CACHE, "%s: run LRU work for client obd\n", cli_name(cli));
 405         count = osc_cache_too_much(cli);
 406         if (count > 0) {
 407                 int rc = osc_lru_shrink(env, cli, count, false);
 408
 409                 CDEBUG(D_CACHE, "%s: shrank %d/%d pages from client obd\n",
 410                        cli_name(cli), rc, count);
 411                 if (rc >= count) {
 412                         CDEBUG(D_CACHE, "%s: queue again\n", cli_name(cli));
 413                         ptlrpcd_queue_work(cli->cl_lru_work);
 414                 }
 415         }
 416
 417         RETURN(0);
 418 }
 419
 420 void osc_lru_add_batch(struct client_obd *cli, struct list_head *plist)
 421 {
 422         LIST_HEAD(lru);
 423         struct osc_async_page *oap;
 424         long npages = 0;
 425
 426         list_for_each_entry(oap, plist, oap_pending_item) {
 427                 struct osc_page *opg = oap2osc_page(oap);
 428
 429                 if (!opg->ops_in_lru)
 430                         continue;
 431
 432                 ++npages;
 433                 LASSERT(list_empty(&opg->ops_lru));
 434                 list_add(&opg->ops_lru, &lru);
 435         }
 436
 437         if (npages > 0) {
 438                 spin_lock(&cli->cl_lru_list_lock);
 439                 list_splice_tail(&lru, &cli->cl_lru_list);
 440                 atomic_long_sub(npages, &cli->cl_lru_busy);
 441                 atomic_long_add(npages, &cli->cl_lru_in_list);
 442                 cli->cl_lru_last_used = ktime_get_real_seconds();
 443                 spin_unlock(&cli->cl_lru_list_lock);
 444
 445                 if (waitqueue_active(&osc_lru_waitq))
 446                         (void)ptlrpcd_queue_work(cli->cl_lru_work);
 447         }
 448 }
 449
 450 static void __osc_lru_del(struct client_obd *cli, struct osc_page *opg)
 451 {
 452         LASSERT(atomic_long_read(&cli->cl_lru_in_list) > 0);
 453         list_del_init(&opg->ops_lru);
 454         atomic_long_dec(&cli->cl_lru_in_list);
 455 }
 456
 457 /**
 458  * Page is being destroyed. The page may be not in LRU list, if the transfer
 459  * has never finished(error occurred).
 460  */
 461 static void osc_lru_del(struct client_obd *cli, struct osc_page *opg)
 462 {
 463         if (opg->ops_in_lru) {
 464                 spin_lock(&cli->cl_lru_list_lock);
 465                 if (!list_empty(&opg->ops_lru)) {
 466                         __osc_lru_del(cli, opg);
 467                 } else {
 468                         LASSERT(atomic_long_read(&cli->cl_lru_busy) > 0);
 469                         atomic_long_dec(&cli->cl_lru_busy);
 470                 }
 471                 spin_unlock(&cli->cl_lru_list_lock);
 472
 473                 atomic_long_inc(cli->cl_lru_left);
 474                 /* this is a great place to release more LRU pages if
 475                  * this osc occupies too many LRU pages and kernel is
 476                  * stealing one of them. */
 477                 if (osc_cache_too_much(cli)) {
 478                         CDEBUG(D_CACHE, "%s: queue LRU work\n", cli_name(cli));
 479                         (void)ptlrpcd_queue_work(cli->cl_lru_work);
 480                 }
 481                 wake_up(&osc_lru_waitq);
 482         } else {
 483                 LASSERT(list_empty(&opg->ops_lru));
 484         }
 485 }
 486
 487 /**
 488  * Delete page from LRU list for redirty.
 489  */
 490 static void osc_lru_use(struct client_obd *cli, struct osc_page *opg)
 491 {
 492         /* If page is being transferred for the first time,
 493          * ops_lru should be empty */
 494         if (opg->ops_in_lru) {
 495                 if (list_empty(&opg->ops_lru))
 496                         return;
 497                 spin_lock(&cli->cl_lru_list_lock);
 498                 if (!list_empty(&opg->ops_lru)) {
 499                         __osc_lru_del(cli, opg);
 500                         atomic_long_inc(&cli->cl_lru_busy);
 501                 }
 502                 spin_unlock(&cli->cl_lru_list_lock);
 503         }
 504 }
 505
 506 static void discard_pagevec(const struct lu_env *env, struct cl_io *io,
 507                                 struct cl_page **pvec, int max_index)
 508 {
 509         struct pagevec *pagevec = &osc_env_info(env)->oti_pagevec;
 510         int i;
 511
 512         ll_pagevec_init(pagevec, 0);
 513         for (i = 0; i < max_index; i++) {
 514                 struct cl_page *page = pvec[i];
 515
 516                 LASSERT(cl_page_is_owned(page, io));
 517                 cl_page_delete(env, page);
 518                 cl_page_discard(env, io, page);
 519                 cl_page_disown(env, io, page);
 520                 cl_pagevec_put(env, page, pagevec);
 521
 522                 pvec[i] = NULL;
 523         }
 524         pagevec_release(pagevec);
 525 }
 526
 527 /**
 528  * Check if a cl_page can be released, i.e, it's not being used.
 529  *
 530  * If unstable account is turned on, bulk transfer may hold one refcount
 531  * for recovery so we need to check vmpage refcount as well; otherwise,
 532  * even we can destroy cl_page but the corresponding vmpage can't be reused.
 533  */
 534 static inline bool lru_page_busy(struct client_obd *cli, struct cl_page *page)
 535 {
 536         if (cl_page_in_use_noref(page))
 537                 return true;
 538
 539         if (cli->cl_cache->ccc_unstable_check) {
 540                 struct page *vmpage = cl_page_vmpage(page);
 541
 542                 /* this check is racy because the vmpage is not locked, but
 543                  * that's OK - the code which does the actual page release
 544                  * checks this again before releasing
 545                  *
 546                  * vmpage have two known users: cl_page and VM page cache
 547                  */
 548                 if (vmpage_in_use(vmpage, 0))
 549                         return true;
 550         }
 551         return false;
 552 }
 553
 554 /**
 555  * Drop @target of pages from LRU at most.
 556  */
 557 long osc_lru_shrink(const struct lu_env *env, struct client_obd *cli,
 558                    long target, bool force)
 559 {
 560         struct cl_io *io;
 561         struct cl_object *clobj = NULL;
 562         struct cl_page **pvec;
 563         struct osc_page *opg;
 564         long count = 0;
 565         int maxscan = 0;
 566         int index = 0;
 567         int rc = 0;
 568         ENTRY;
 569
 570         LASSERT(atomic_long_read(&cli->cl_lru_in_list) >= 0);
 571         if (atomic_long_read(&cli->cl_lru_in_list) == 0 || target <= 0)
 572                 RETURN(0);
 573
 574         CDEBUG(D_CACHE, "%s: shrinkers: %d, force: %d\n",
 575                cli_name(cli), atomic_read(&cli->cl_lru_shrinkers), force);
 576         if (!force) {
 577                 if (atomic_read(&cli->cl_lru_shrinkers) > 0)
 578                         RETURN(-EBUSY);
 579
 580                 if (atomic_inc_return(&cli->cl_lru_shrinkers) > 1) {
 581                         atomic_dec(&cli->cl_lru_shrinkers);
 582                         RETURN(-EBUSY);
 583                 }
 584         } else {
 585                 atomic_inc(&cli->cl_lru_shrinkers);
 586         }
 587
 588         pvec = (struct cl_page **)osc_env_info(env)->oti_pvec;
 589         io = osc_env_thread_io(env);
 590
 591         spin_lock(&cli->cl_lru_list_lock);
 592         if (force)
 593                 cli->cl_lru_reclaim++;
 594         maxscan = min(target << 1, atomic_long_read(&cli->cl_lru_in_list));
 595         while (!list_empty(&cli->cl_lru_list)) {
 596                 struct cl_page *page;
 597                 bool will_free = false;
 598
 599                 if (!force && atomic_read(&cli->cl_lru_shrinkers) > 1)
 600                         break;
 601
 602                 if (--maxscan < 0)
 603                         break;
 604
 605                 opg = list_first_entry(&cli->cl_lru_list, struct osc_page,
 606                                        ops_lru);
 607                 page = opg->ops_cl.cpl_page;
 608                 if (lru_page_busy(cli, page)) {
 609                         list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
 610                         continue;
 611                 }
 612
 613                 LASSERT(page->cp_obj != NULL);
 614                 if (clobj != page->cp_obj) {
 615                         struct cl_object *tmp = page->cp_obj;
 616
 617                         cl_object_get(tmp);
 618                         spin_unlock(&cli->cl_lru_list_lock);
 619
 620                         if (clobj != NULL) {
 621                                 discard_pagevec(env, io, pvec, index);
 622                                 index = 0;
 623
 624                                 cl_io_fini(env, io);
 625                                 cl_object_put(env, clobj);
 626                                 clobj = NULL;
 627                         }
 628
 629                         clobj = tmp;
 630                         io->ci_obj = clobj;
 631                         io->ci_ignore_layout = 1;
 632                         rc = cl_io_init(env, io, CIT_MISC, clobj);
 633
 634                         spin_lock(&cli->cl_lru_list_lock);
 635
 636                         if (rc != 0)
 637                                 break;
 638
 639                         ++maxscan;
 640                         continue;
 641                 }
 642
 643                 if (cl_page_own_try(env, io, page) == 0) {
 644                         if (!lru_page_busy(cli, page)) {
 645                                 /* remove it from lru list earlier to avoid
 646                                  * lock contention */
 647                                 __osc_lru_del(cli, opg);
 648                                 opg->ops_in_lru = 0; /* will be discarded */
 649
 650                                 cl_page_get(page);
 651                                 will_free = true;
 652                         } else {
 653                                 cl_page_disown(env, io, page);
 654                         }
 655                 }
 656
 657                 if (!will_free) {
 658                         list_move_tail(&opg->ops_lru, &cli->cl_lru_list);
 659                         continue;
 660                 }
 661
 662                 /* Don't discard and free the page with cl_lru_list held */
 663                 pvec[index++] = page;
 664                 if (unlikely(index == OTI_PVEC_SIZE)) {
 665                         spin_unlock(&cli->cl_lru_list_lock);
 666                         discard_pagevec(env, io, pvec, index);
 667                         index = 0;
 668
 669                         spin_lock(&cli->cl_lru_list_lock);
 670                 }
 671
 672                 if (++count >= target)
 673                         break;
 674         }
 675         spin_unlock(&cli->cl_lru_list_lock);
 676
 677         if (clobj != NULL) {
 678                 discard_pagevec(env, io, pvec, index);
 679
 680                 cl_io_fini(env, io);
 681                 cl_object_put(env, clobj);
 682         }
 683
 684         atomic_dec(&cli->cl_lru_shrinkers);
 685         if (count > 0) {
 686                 atomic_long_add(count, cli->cl_lru_left);
 687                 wake_up(&osc_lru_waitq);
 688         }
 689         RETURN(count > 0 ? count : rc);
 690 }
 691 EXPORT_SYMBOL(osc_lru_shrink);
 692
 693 /**
 694  * Reclaim LRU pages by an IO thread. The caller wants to reclaim at least
 695  * \@npages of LRU slots. For performance consideration, it's better to drop
 696  * LRU pages in batch. Therefore, the actual number is adjusted at least
 697  * max_pages_per_rpc.
 698  */
 699 static long osc_lru_reclaim(struct client_obd *cli, unsigned long npages)
 700 {
 701         struct lu_env *env;
 702         struct cl_client_cache *cache = cli->cl_cache;
 703         struct client_obd *scan;
 704         int max_scans;
 705         __u16 refcheck;
 706         long rc = 0;
 707         ENTRY;
 708
 709         LASSERT(cache != NULL);
 710
 711         env = cl_env_get(&refcheck);
 712         if (IS_ERR(env))
 713                 RETURN(rc);
 714
 715         npages = max_t(int, npages, cli->cl_max_pages_per_rpc);
 716         CDEBUG(D_CACHE, "%s: start to reclaim %ld pages from LRU\n",
 717                cli_name(cli), npages);
 718         rc = osc_lru_shrink(env, cli, npages, true);
 719         if (rc >= npages) {
 720                 CDEBUG(D_CACHE, "%s: reclaimed %ld/%ld pages from LRU\n",
 721                        cli_name(cli), rc, npages);
 722                 if (osc_cache_too_much(cli) > 0)
 723                         ptlrpcd_queue_work(cli->cl_lru_work);
 724                 GOTO(out, rc);
 725         } else if (rc > 0) {
 726                 npages -= rc;
 727         }
 728
 729         CDEBUG(D_CACHE, "%s: cli %p no free slots, pages: %ld/%ld, want: %ld\n",
 730                 cli_name(cli), cli, atomic_long_read(&cli->cl_lru_in_list),
 731                 atomic_long_read(&cli->cl_lru_busy), npages);
 732
 733         /* Reclaim LRU slots from other client_obd as it can't free enough
 734          * from its own. This should rarely happen. */
 735         spin_lock(&cache->ccc_lru_lock);
 736         LASSERT(!list_empty(&cache->ccc_lru));
 737
 738         cache->ccc_lru_shrinkers++;
 739         list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru);
 740
 741         max_scans = atomic_read(&cache->ccc_users) - 2;
 742         while (--max_scans > 0 &&
 743                (scan = list_first_entry_or_null(&cache->ccc_lru,
 744                                                   struct client_obd,
 745                                                   cl_lru_osc)) != NULL) {
 746                 CDEBUG(D_CACHE, "%s: cli %p LRU pages: %ld, busy: %ld.\n",
 747                        cli_name(scan), scan,
 748                        atomic_long_read(&scan->cl_lru_in_list),
 749                        atomic_long_read(&scan->cl_lru_busy));
 750
 751                 list_move_tail(&scan->cl_lru_osc, &cache->ccc_lru);
 752                 if (osc_cache_too_much(scan) > 0) {
 753                         spin_unlock(&cache->ccc_lru_lock);
 754
 755                         rc = osc_lru_shrink(env, scan, npages, true);
 756                         spin_lock(&cache->ccc_lru_lock);
 757                         if (rc >= npages)
 758                                 break;
 759                         if (rc > 0)
 760                                 npages -= rc;
 761                 }
 762         }
 763         spin_unlock(&cache->ccc_lru_lock);
 764
 765 out:
 766         cl_env_put(env, &refcheck);
 767         CDEBUG(D_CACHE, "%s: cli %p freed %ld pages.\n",
 768                cli_name(cli), cli, rc);
 769         return rc;
 770 }
 771
 772 /**
 773  * osc_lru_alloc() is called to allocate an LRU slot for a cl_page.
 774  *
 775  * Usually the LRU slots are reserved in osc_io_iter_rw_init().
 776  * Only in the case that the LRU slots are in extreme shortage, it should
 777  * have reserved enough slots for an IO.
 778  */
 779 static int osc_lru_alloc(const struct lu_env *env, struct client_obd *cli,
 780                          struct osc_page *opg)
 781 {
 782         struct osc_io *oio = osc_env_io(env);
 783         int rc = 0;
 784
 785         ENTRY;
 786
 787         if (cli->cl_cache == NULL) /* shall not be in LRU */
 788                 RETURN(0);
 789
 790         if (oio->oi_lru_reserved > 0) {
 791                 --oio->oi_lru_reserved;
 792                 goto out;
 793         }
 794
 795         LASSERT(atomic_long_read(cli->cl_lru_left) >= 0);
 796         while (!atomic_long_add_unless(cli->cl_lru_left, -1, 0)) {
 797                 /* run out of LRU spaces, try to drop some by itself */
 798                 rc = osc_lru_reclaim(cli, 1);
 799                 if (rc < 0)
 800                         break;
 801                 if (rc > 0)
 802                         continue;
 803                 /* IO issued by readahead, don't try hard */
 804                 if (oio->oi_is_readahead) {
 805                         if (atomic_long_read(cli->cl_lru_left) > 0)
 806                                 continue;
 807                         rc = -EBUSY;
 808                         break;
 809                 }
 810
 811                 cond_resched();
 812                 rc = l_wait_event_abortable(
 813                         osc_lru_waitq,
 814                         atomic_long_read(cli->cl_lru_left) > 0);
 815                 if (rc < 0) {
 816                         rc = -EINTR;
 817                         break;
 818                 }
 819         }
 820
 821 out:
 822         if (rc >= 0) {
 823                 atomic_long_inc(&cli->cl_lru_busy);
 824                 opg->ops_in_lru = 1;
 825                 rc = 0;
 826         }
 827
 828         RETURN(rc);
 829 }
 830
 831 /**
 832  * osc_lru_reserve() is called to reserve enough LRU slots for I/O.
 833  *
 834  * The benefit of doing this is to reduce contention against atomic counter
 835  * cl_lru_left by changing it from per-page access to per-IO access.
 836  */
 837 unsigned long osc_lru_reserve(struct client_obd *cli, unsigned long npages)
 838 {
 839         unsigned long reserved = 0;
 840         unsigned long max_pages;
 841         unsigned long c;
 842         int rc;
 843
 844 again:
 845         c = atomic_long_read(cli->cl_lru_left);
 846         if (c < npages && osc_lru_reclaim(cli, npages) > 0)
 847                 c = atomic_long_read(cli->cl_lru_left);
 848
 849         if (c < npages) {
 850                 /*
 851                  * Trigger writeback in the hope some LRU slot could
 852                  * be freed.
 853                  */
 854                 rc = ptlrpcd_queue_work(cli->cl_writeback_work);
 855                 if (rc)
 856                         return 0;
 857         }
 858
 859         while (c >= npages) {
 860                 if (c == atomic_long_cmpxchg(cli->cl_lru_left, c, c - npages)) {
 861                         reserved = npages;
 862                         break;
 863                 }
 864                 c = atomic_long_read(cli->cl_lru_left);
 865         }
 866
 867         if (reserved != npages) {
 868                 cond_resched();
 869                 rc = l_wait_event_abortable(
 870                         osc_lru_waitq,
 871                         atomic_long_read(cli->cl_lru_left) > 0);
 872                 goto again;
 873         }
 874
 875         max_pages = cli->cl_max_pages_per_rpc * cli->cl_max_rpcs_in_flight;
 876         if (atomic_long_read(cli->cl_lru_left) < max_pages) {
 877                 /* If there aren't enough pages in the per-OSC LRU then
 878                  * wake up the LRU thread to try and clear out space, so
 879                  * we don't block if pages are being dirtied quickly. */
 880                 CDEBUG(D_CACHE, "%s: queue LRU, left: %lu/%ld.\n",
 881                        cli_name(cli), atomic_long_read(cli->cl_lru_left),
 882                        max_pages);
 883                 (void)ptlrpcd_queue_work(cli->cl_lru_work);
 884         }
 885
 886         return reserved;
 887 }
 888
 889 /**
 890  * osc_lru_unreserve() is called to unreserve LRU slots.
 891  *
 892  * LRU slots reserved by osc_lru_reserve() may have entries left due to several
 893  * reasons such as page already existing or I/O error. Those reserved slots
 894  * should be freed by calling this function.
 895  */
 896 void osc_lru_unreserve(struct client_obd *cli, unsigned long npages)
 897 {
 898         atomic_long_add(npages, cli->cl_lru_left);
 899         wake_up(&osc_lru_waitq);
 900 }
 901
 902 /**
 903  * Atomic operations are expensive. We accumulate the accounting for the
 904  * same page zone to get better performance.
 905  * In practice this can work pretty good because the pages in the same RPC
 906  * are likely from the same page zone.
 907  */
 908 #ifdef HAVE_NR_UNSTABLE_NFS
 909 /* Old kernels use a separate counter for unstable pages,
 910  * newer kernels treat them like any other writeback.
 911  */
 912 #define NR_WRITEBACK NR_UNSTABLE_NFS
 913 #endif
 914
 915 static inline void unstable_page_accounting(struct ptlrpc_bulk_desc *desc,
 916                                             struct osc_brw_async_args *aa,
 917                                             int factor)
 918 {
 919         int page_count;
 920         void *zone = NULL;
 921         int count = 0;
 922         int i;
 923
 924         if (desc != NULL) {
 925                 page_count = desc->bd_iov_count;
 926         } else {
 927                 page_count = aa->aa_page_count;
 928         }
 929
 930         for (i = 0; i < page_count; i++) {
 931                 void *pz;
 932                 if (desc)
 933                         pz = page_zone(desc->bd_vec[i].bv_page);
 934                 else
 935                         pz = page_zone(aa->aa_ppga[i]->pg);
 936
 937                 if (likely(pz == zone)) {
 938                         ++count;
 939                         continue;
 940                 }
 941
 942                 if (count > 0) {
 943                         mod_zone_page_state(zone, NR_WRITEBACK,
 944                                             factor * count);
 945                         count = 0;
 946                 }
 947                 zone = pz;
 948                 ++count;
 949         }
 950         if (count > 0)
 951                 mod_zone_page_state(zone, NR_WRITEBACK, factor * count);
 952 }
 953
 954 static inline void add_unstable_page_accounting(struct ptlrpc_bulk_desc *desc,
 955                                                 struct osc_brw_async_args *aa)
 956 {
 957         unstable_page_accounting(desc, aa, 1);
 958 }
 959
 960 static inline void dec_unstable_page_accounting(struct ptlrpc_bulk_desc *desc,
 961                                                 struct osc_brw_async_args *aa)
 962 {
 963         unstable_page_accounting(desc, aa, -1);
 964 }
 965
 966 /**
 967  * Performs "unstable" page accounting. This function balances the
 968  * increment operations performed in osc_inc_unstable_pages. It is
 969  * registered as the RPC request callback, and is executed when the
 970  * bulk RPC is committed on the server. Thus at this point, the pages
 971  * involved in the bulk transfer are no longer considered unstable.
 972  *
 973  * If this function is called, the request should have been committed
 974  * or req:rq_unstable must have been set; it implies that the unstable
 975  * statistic have been added.
 976  */
 977 void osc_dec_unstable_pages(struct ptlrpc_request *req)
 978 {
 979         struct ptlrpc_bulk_desc *desc       = req->rq_bulk;
 980         struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
 981         struct client_obd       *cli        = &req->rq_import->imp_obd->u.cli;
 982         int                      page_count;
 983         long                     unstable_count;
 984
 985         if (desc)
 986                 page_count = desc->bd_iov_count;
 987         else
 988                 page_count = aa->aa_page_count;
 989
 990         LASSERT(page_count >= 0);
 991
 992         dec_unstable_page_accounting(desc, aa);
 993
 994         unstable_count = atomic_long_sub_return(page_count,
 995                                                 &cli->cl_unstable_count);
 996         LASSERT(unstable_count >= 0);
 997
 998         unstable_count = atomic_long_sub_return(page_count,
 999                                            &cli->cl_cache->ccc_unstable_nr);
1000         LASSERT(unstable_count >= 0);
1001         if (unstable_count == 0)
1002                 wake_up(&cli->cl_cache->ccc_unstable_waitq);
1003
1004         if (waitqueue_active(&osc_lru_waitq))
1005                 (void)ptlrpcd_queue_work(cli->cl_lru_work);
1006 }
1007
1008 /**
1009  * "unstable" page accounting. See: osc_dec_unstable_pages.
1010  */
1011 void osc_inc_unstable_pages(struct ptlrpc_request *req)
1012 {
1013         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
1014         struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
1015         struct client_obd       *cli  = &req->rq_import->imp_obd->u.cli;
1016         long                     page_count;
1017
1018         /* No unstable page tracking */
1019         if (cli->cl_cache == NULL || !cli->cl_cache->ccc_unstable_check)
1020                 return;
1021
1022         if (desc)
1023                 page_count = desc->bd_iov_count;
1024         else
1025                 page_count = aa->aa_page_count;
1026
1027         add_unstable_page_accounting(desc, aa);
1028         atomic_long_add(page_count, &cli->cl_unstable_count);
1029         atomic_long_add(page_count, &cli->cl_cache->ccc_unstable_nr);
1030
1031         /* If the request has already been committed (i.e. brw_commit
1032          * called via rq_commit_cb), we need to undo the unstable page
1033          * increments we just performed because rq_commit_cb wont be
1034          * called again. */
1035         spin_lock(&req->rq_lock);
1036         if (unlikely(req->rq_committed)) {
1037                 spin_unlock(&req->rq_lock);
1038
1039                 osc_dec_unstable_pages(req);
1040         } else {
1041                 req->rq_unstable = 1;
1042                 spin_unlock(&req->rq_lock);
1043         }
1044 }
1045
1046 /**
1047  * Check if it piggybacks SOFT_SYNC flag to OST from this OSC.
1048  * This function will be called by every BRW RPC so it's critical
1049  * to make this function fast.
1050  */
1051 bool osc_over_unstable_soft_limit(struct client_obd *cli)
1052 {
1053         long unstable_nr, osc_unstable_count;
1054
1055         /* Can't check cli->cl_unstable_count, therefore, no soft limit */
1056         if (cli->cl_cache == NULL || !cli->cl_cache->ccc_unstable_check)
1057                 return false;
1058
1059         osc_unstable_count = atomic_long_read(&cli->cl_unstable_count);
1060         unstable_nr = atomic_long_read(&cli->cl_cache->ccc_unstable_nr);
1061
1062         CDEBUG(D_CACHE,
1063                "%s: cli: %p unstable pages: %lu, osc unstable pages: %lu\n",
1064                cli_name(cli), cli, unstable_nr, osc_unstable_count);
1065
1066         /* If the LRU slots are in shortage - 25% remaining AND this OSC
1067          * has one full RPC window of unstable pages, it's a good chance
1068          * to piggyback a SOFT_SYNC flag.
1069          * Please notice that the OST won't take immediate response for the
1070          * SOFT_SYNC request so active OSCs will have more chance to carry
1071          * the flag, this is reasonable. */
1072         return unstable_nr > cli->cl_cache->ccc_lru_max >> 2 &&
1073                osc_unstable_count > cli->cl_max_pages_per_rpc *
1074                                     cli->cl_max_rpcs_in_flight;
1075 }
1076
1077 /**
1078  * Return how many LRU pages in the cache of all OSC devices
1079  *
1080  * \retval      return # of cached LRU pages times reclaimation tendency
1081  * \retval      SHRINK_STOP if it cannot do any scanning in this time
1082  */
1083 unsigned long osc_cache_shrink_count(struct shrinker *sk,
1084                                      struct shrink_control *sc)
1085 {
1086         struct client_obd *cli;
1087         unsigned long cached = 0;
1088
1089         spin_lock(&osc_shrink_lock);
1090         list_for_each_entry(cli, &osc_shrink_list, cl_shrink_list)
1091                 cached += atomic_long_read(&cli->cl_lru_in_list);
1092         spin_unlock(&osc_shrink_lock);
1093
1094         return (cached  * sysctl_vfs_cache_pressure) / 100;
1095 }
1096
1097 /**
1098  * Scan and try to reclaim sc->nr_to_scan cached LRU pages
1099  *
1100  * \retval      number of cached LRU pages reclaimed
1101  * \retval      SHRINK_STOP if it cannot do any scanning in this time
1102  *
1103  * Linux kernel will loop calling this shrinker scan routine with
1104  * sc->nr_to_scan = SHRINK_BATCH(128 for now) until kernel got enough memory.
1105  *
1106  * If sc->nr_to_scan is 0, the VM is querying the cache size, we don't need
1107  * to scan and try to reclaim LRU pages, just return 0 and
1108  * osc_cache_shrink_count() will report the LRU page number.
1109  */
1110 unsigned long osc_cache_shrink_scan(struct shrinker *sk,
1111                                     struct shrink_control *sc)
1112 {
1113         struct client_obd *cli;
1114         struct client_obd *stop_anchor = NULL;
1115         struct lu_env *env;
1116         long shrank = 0;
1117         int rc;
1118         __u16 refcheck;
1119
1120         if (sc->nr_to_scan == 0)
1121                 return 0;
1122
1123         if (!(sc->gfp_mask & __GFP_FS))
1124                 return SHRINK_STOP;
1125
1126         env = cl_env_get(&refcheck);
1127         if (IS_ERR(env))
1128                 return SHRINK_STOP;
1129
1130         spin_lock(&osc_shrink_lock);
1131         while ((cli = list_first_entry_or_null(&osc_shrink_list,
1132                                                struct client_obd,
1133                                                cl_shrink_list)) != NULL) {
1134                 if (stop_anchor == NULL)
1135                         stop_anchor = cli;
1136                 else if (cli == stop_anchor)
1137                         break;
1138
1139                 list_move_tail(&cli->cl_shrink_list, &osc_shrink_list);
1140                 spin_unlock(&osc_shrink_lock);
1141
1142                 /* shrink no more than max_pages_per_rpc for an OSC */
1143                 rc = osc_lru_shrink(env, cli, (sc->nr_to_scan - shrank) >
1144                                     cli->cl_max_pages_per_rpc ?
1145                                     cli->cl_max_pages_per_rpc :
1146                                     sc->nr_to_scan - shrank, true);
1147                 if (rc > 0)
1148                         shrank += rc;
1149
1150                 if (shrank >= sc->nr_to_scan)
1151                         goto out;
1152
1153                 spin_lock(&osc_shrink_lock);
1154         }
1155         spin_unlock(&osc_shrink_lock);
1156
1157 out:
1158         cl_env_put(env, &refcheck);
1159
1160         return shrank;
1161 }
1162
1163 /** @} osc */