lustre/llite/rw.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) 2002, 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  * lustre/llite/rw.c
  32  *
  33  * Lustre Lite I/O page cache routines shared by different kernel revs
  34  */
  35
  36 #include <linux/kernel.h>
  37 #include <linux/mm.h>
  38 #include <linux/string.h>
  39 #include <linux/stat.h>
  40 #include <linux/errno.h>
  41 #include <linux/unistd.h>
  42 #include <linux/writeback.h>
  43 #include <asm/uaccess.h>
  44
  45 #include <linux/fs.h>
  46 #include <linux/file.h>
  47 #include <linux/stat.h>
  48 #include <asm/uaccess.h>
  49 #include <linux/mm.h>
  50 #include <linux/pagemap.h>
  51 /* current_is_kswapd() */
  52 #include <linux/swap.h>
  53 #include <linux/task_io_accounting_ops.h>
  54
  55 #define DEBUG_SUBSYSTEM S_LLITE
  56
  57 #include <obd_cksum.h>
  58 #include "llite_internal.h"
  59 #include <lustre_compat.h>
  60
  61 static void ll_ra_stats_inc_sbi(struct ll_sb_info *sbi, enum ra_stat which);
  62
  63 /**
  64  * Get readahead pages from the filesystem readahead pool of the client for a
  65  * thread.
  66  *
  67  * /param sbi superblock for filesystem readahead state ll_ra_info
  68  * /param ria per-thread readahead state
  69  * /param pages number of pages requested for readahead for the thread.
  70  *
  71  * WARNING: This algorithm is used to reduce contention on sbi->ll_lock.
  72  * It should work well if the ra_max_pages is much greater than the single
  73  * file's read-ahead window, and not too many threads contending for
  74  * these readahead pages.
  75  *
  76  * TODO: There may be a 'global sync problem' if many threads are trying
  77  * to get an ra budget that is larger than the remaining readahead pages
  78  * and reach here at exactly the same time. They will compute /a ret to
  79  * consume the remaining pages, but will fail at atomic_add_return() and
  80  * get a zero ra window, although there is still ra space remaining. - Jay */
  81
  82 static unsigned long ll_ra_count_get(struct ll_sb_info *sbi,
  83                                      struct ra_io_arg *ria,
  84                                      unsigned long pages,
  85                                      unsigned long pages_min)
  86 {
  87         struct ll_ra_info *ra = &sbi->ll_ra_info;
  88         long ret;
  89
  90         ENTRY;
  91
  92         /**
  93          * Don't try readahead agreesively if we are limited
  94          * LRU pages, otherwise, it could cause deadlock.
  95          */
  96         pages = min(sbi->ll_cache->ccc_lru_max >> 2, pages);
  97
  98         /*
  99          * If read-ahead pages left are less than 1M, do not do read-ahead,
 100          * otherwise it will form small read RPC(< 1M), which hurt server
 101          * performance a lot.
 102          */
 103         ret = min(ra->ra_max_pages - atomic_read(&ra->ra_cur_pages),
 104                   pages);
 105         if (ret < 0 || ret < min_t(long, PTLRPC_MAX_BRW_PAGES, pages))
 106                 GOTO(out, ret = 0);
 107
 108         if (atomic_add_return(ret, &ra->ra_cur_pages) > ra->ra_max_pages) {
 109                 atomic_sub(ret, &ra->ra_cur_pages);
 110                 ret = 0;
 111         }
 112
 113 out:
 114         if (ret < pages_min) {
 115                 /* override ra limit for maximum performance */
 116                 atomic_add(pages_min - ret, &ra->ra_cur_pages);
 117                 ret = pages_min;
 118         }
 119         RETURN(ret);
 120 }
 121
 122 void ll_ra_count_put(struct ll_sb_info *sbi, unsigned long pages)
 123 {
 124         struct ll_ra_info *ra = &sbi->ll_ra_info;
 125         atomic_sub(pages, &ra->ra_cur_pages);
 126 }
 127
 128 static void ll_ra_stats_inc_sbi(struct ll_sb_info *sbi, enum ra_stat which)
 129 {
 130         LASSERTF(which < _NR_RA_STAT, "which: %u\n", which);
 131         lprocfs_counter_incr(sbi->ll_ra_stats, which);
 132 }
 133
 134 static inline bool ll_readahead_enabled(struct ll_sb_info *sbi)
 135 {
 136         return sbi->ll_ra_info.ra_max_pages_per_file > 0 &&
 137                 sbi->ll_ra_info.ra_max_pages > 0;
 138 }
 139
 140 void ll_ra_stats_inc(struct inode *inode, enum ra_stat which)
 141 {
 142         struct ll_sb_info *sbi = ll_i2sbi(inode);
 143         ll_ra_stats_inc_sbi(sbi, which);
 144 }
 145
 146 #define RAS_CDEBUG(ras) \
 147         CDEBUG(D_READA,                                                      \
 148                "lre %llu cr %lu cb %llu wsi %lu wp %lu nra %lu rpc %lu "     \
 149                "r %lu csr %lu so %llu sb %llu sl %llu lr %lu\n",             \
 150                ras->ras_last_read_end_bytes, ras->ras_consecutive_requests,  \
 151                ras->ras_consecutive_bytes, ras->ras_window_start_idx,        \
 152                ras->ras_window_pages, ras->ras_next_readahead_idx,           \
 153                ras->ras_rpc_pages, ras->ras_requests,                        \
 154                ras->ras_consecutive_stride_requests, ras->ras_stride_offset, \
 155                ras->ras_stride_bytes, ras->ras_stride_length,                \
 156                ras->ras_async_last_readpage_idx)
 157
 158 static bool pos_in_window(loff_t pos, loff_t point,
 159                           unsigned long before, unsigned long after)
 160 {
 161         loff_t start = point - before;
 162         loff_t end = point + after;
 163
 164         if (start > point)
 165                 start = 0;
 166         if (end < point)
 167                 end = ~0;
 168
 169         return start <= pos && pos <= end;
 170 }
 171
 172 enum ll_ra_page_hint {
 173         MAYNEED = 0, /* this page possibly accessed soon */
 174         WILLNEED /* this page is gurateed to be needed */
 175 };
 176
 177 /**
 178  * Initiates read-ahead of a page with given index.
 179  *
 180  * \retval +ve: page was already uptodate so it will be skipped
 181  *              from being added;
 182  * \retval -ve: page wasn't added to \a queue for error;
 183  * \retval   0: page was added into \a queue for read ahead.
 184  */
 185 static int ll_read_ahead_page(const struct lu_env *env, struct cl_io *io,
 186                               struct cl_page_list *queue, pgoff_t index,
 187                               enum ll_ra_page_hint hint)
 188 {
 189         struct cl_object *clob  = io->ci_obj;
 190         struct inode     *inode = vvp_object_inode(clob);
 191         struct page      *vmpage = NULL;
 192         struct cl_page   *page;
 193         struct vvp_page  *vpg;
 194         enum ra_stat      which = _NR_RA_STAT; /* keep gcc happy */
 195         int               rc    = 0;
 196         const char       *msg   = NULL;
 197
 198         ENTRY;
 199
 200         switch (hint) {
 201         case MAYNEED:
 202                 vmpage = grab_cache_page_nowait(inode->i_mapping, index);
 203                 if (vmpage == NULL) {
 204                         which = RA_STAT_FAILED_GRAB_PAGE;
 205                         msg   = "g_c_p_n failed";
 206                         GOTO(out, rc = -EBUSY);
 207                 }
 208                 break;
 209         case WILLNEED:
 210                 vmpage = find_or_create_page(inode->i_mapping, index,
 211                                              GFP_NOFS);
 212                 if (vmpage == NULL)
 213                         GOTO(out, rc = -ENOMEM);
 214                 break;
 215         default:
 216                 /* should not come here */
 217                 GOTO(out, rc = -EINVAL);
 218         }
 219
 220         /* Check if vmpage was truncated or reclaimed */
 221         if (vmpage->mapping != inode->i_mapping) {
 222                 which = RA_STAT_WRONG_GRAB_PAGE;
 223                 msg   = "g_c_p_n returned invalid page";
 224                 GOTO(out, rc = -EBUSY);
 225         }
 226
 227         page = cl_page_find(env, clob, vmpage->index, vmpage, CPT_CACHEABLE);
 228         if (IS_ERR(page)) {
 229                 which = RA_STAT_FAILED_GRAB_PAGE;
 230                 msg   = "cl_page_find failed";
 231                 GOTO(out, rc = PTR_ERR(page));
 232         }
 233
 234         lu_ref_add(&page->cp_reference, "ra", current);
 235         cl_page_assume(env, io, page);
 236         vpg = cl2vvp_page(cl_object_page_slice(clob, page));
 237         if (!vpg->vpg_defer_uptodate && !PageUptodate(vmpage)) {
 238                 vpg->vpg_defer_uptodate = 1;
 239                 vpg->vpg_ra_used = 0;
 240                 cl_page_list_add(queue, page);
 241         } else {
 242                 /* skip completed pages */
 243                 cl_page_unassume(env, io, page);
 244                 /* This page is already uptodate, returning a positive number
 245                  * to tell the callers about this */
 246                 rc = 1;
 247         }
 248
 249         lu_ref_del(&page->cp_reference, "ra", current);
 250         cl_page_put(env, page);
 251
 252 out:
 253         if (vmpage != NULL) {
 254                 if (rc != 0)
 255                         unlock_page(vmpage);
 256                 put_page(vmpage);
 257         }
 258         if (msg != NULL && hint == MAYNEED) {
 259                 ll_ra_stats_inc(inode, which);
 260                 CDEBUG(D_READA, "%s\n", msg);
 261
 262         }
 263
 264         RETURN(rc);
 265 }
 266
 267 #define RIA_DEBUG(ria)                                                  \
 268         CDEBUG(D_READA, "rs %lu re %lu ro %llu rl %llu rb %llu\n",      \
 269                ria->ria_start_idx, ria->ria_end_idx, ria->ria_stoff,    \
 270                ria->ria_length, ria->ria_bytes)
 271
 272 static inline int stride_io_mode(struct ll_readahead_state *ras)
 273 {
 274         return ras->ras_consecutive_stride_requests > 1;
 275 }
 276
 277 /* The function calculates how many bytes will be read in
 278  * [off, off + length], in such stride IO area,
 279  * stride_offset = st_off, stride_lengh = st_len,
 280  * stride_bytes = st_bytes
 281  *
 282  *   |------------------|*****|------------------|*****|------------|*****|....
 283  * st_off
 284  *   |--- st_bytes     ---|
 285  *   |-----     st_len   -----|
 286  *
 287  *              How many bytes it should read in such pattern
 288  *              |-------------------------------------------------------------|
 289  *              off
 290  *              |<------                  length                      ------->|
 291  *
 292  *          =   |<----->|  +  |-------------------------------------| +   |---|
 293  *             start_left                 st_bytes * i                 end_left
 294  */
 295 static loff_t stride_byte_count(loff_t st_off, loff_t st_len, loff_t st_bytes,
 296                                 loff_t off, loff_t length)
 297 {
 298         u64 start = off > st_off ? off - st_off : 0;
 299         u64 end = off + length > st_off ? off + length - st_off : 0;
 300         u64 start_left;
 301         u64 end_left;
 302         u64 bytes_count;
 303
 304         if (st_len == 0 || length == 0 || end == 0)
 305                 return length;
 306
 307         start = div64_u64_rem(start, st_len, &start_left);
 308         if (start_left < st_bytes)
 309                 start_left = st_bytes - start_left;
 310         else
 311                 start_left = 0;
 312
 313         end = div64_u64_rem(end, st_len, &end_left);
 314         if (end_left > st_bytes)
 315                 end_left = st_bytes;
 316
 317         CDEBUG(D_READA, "start %llu, end %llu start_left %llu end_left %llu\n",
 318                start, end, start_left, end_left);
 319
 320         if (start == end)
 321                 bytes_count = end_left - (st_bytes - start_left);
 322         else
 323                 bytes_count = start_left +
 324                         st_bytes * (end - start - 1) + end_left;
 325
 326         CDEBUG(D_READA,
 327                "st_off %llu, st_len %llu st_bytes %llu off %llu length %llu bytescount %llu\n",
 328                st_off, st_len, st_bytes, off, length, bytes_count);
 329
 330         return bytes_count;
 331 }
 332
 333 static unsigned long ria_page_count(struct ra_io_arg *ria)
 334 {
 335         loff_t length_bytes = ria->ria_end_idx >= ria->ria_start_idx ?
 336                 (loff_t)(ria->ria_end_idx -
 337                          ria->ria_start_idx + 1) << PAGE_SHIFT : 0;
 338         loff_t bytes_count;
 339
 340         if (ria->ria_length > ria->ria_bytes && ria->ria_bytes &&
 341             (ria->ria_length & ~PAGE_MASK || ria->ria_bytes & ~PAGE_MASK ||
 342              ria->ria_stoff & ~PAGE_MASK)) {
 343                 /* Over-estimate un-aligned page stride read */
 344                 unsigned long pg_count = ((ria->ria_bytes +
 345                                            PAGE_SIZE - 1) >> PAGE_SHIFT) + 1;
 346                 pg_count *= length_bytes / ria->ria_length + 1;
 347
 348                 return pg_count;
 349         }
 350         bytes_count = stride_byte_count(ria->ria_stoff, ria->ria_length,
 351                                         ria->ria_bytes,
 352                                         (loff_t)ria->ria_start_idx<<PAGE_SHIFT,
 353                                         length_bytes);
 354         return (bytes_count + PAGE_SIZE - 1) >> PAGE_SHIFT;
 355 }
 356
 357 static pgoff_t ras_align(struct ll_readahead_state *ras, pgoff_t index)
 358 {
 359         unsigned opt_size = min(ras->ras_window_pages, ras->ras_rpc_pages);
 360
 361         if (opt_size == 0)
 362                 opt_size = 1;
 363         return index - (index % opt_size);
 364 }
 365
 366 /* Check whether the index is in the defined ra-window */
 367 static bool ras_inside_ra_window(pgoff_t idx, struct ra_io_arg *ria)
 368 {
 369         loff_t pos = (loff_t)idx << PAGE_SHIFT;
 370
 371         /* If ria_length == ria_bytes, it means non-stride I/O mode,
 372          * idx should always inside read-ahead window in this case
 373          * For stride I/O mode, just check whether the idx is inside
 374          * the ria_bytes.
 375          */
 376         if (ria->ria_length == 0 || ria->ria_length == ria->ria_bytes)
 377                 return true;
 378
 379         if (pos >= ria->ria_stoff) {
 380                 u64 offset;
 381
 382                 div64_u64_rem(pos - ria->ria_stoff, ria->ria_length, &offset);
 383
 384                 if (offset < ria->ria_bytes ||
 385                     (ria->ria_length - offset) < PAGE_SIZE)
 386                         return true;
 387         } else if (pos + PAGE_SIZE > ria->ria_stoff) {
 388                 return true;
 389         }
 390
 391         return false;
 392 }
 393
 394 static unsigned long
 395 ll_read_ahead_pages(const struct lu_env *env, struct cl_io *io,
 396                     struct cl_page_list *queue, struct ll_readahead_state *ras,
 397                     struct ra_io_arg *ria, pgoff_t *ra_end, pgoff_t skip_index)
 398 {
 399         struct cl_read_ahead ra = { 0 };
 400         /* busy page count is per stride */
 401         int rc = 0, count = 0, busy_page_count = 0;
 402         pgoff_t page_idx;
 403
 404         LASSERT(ria != NULL);
 405         RIA_DEBUG(ria);
 406
 407         for (page_idx = ria->ria_start_idx;
 408              page_idx <= ria->ria_end_idx && ria->ria_reserved > 0;
 409              page_idx++) {
 410                 if (skip_index && page_idx == skip_index)
 411                         continue;
 412                 if (ras_inside_ra_window(page_idx, ria)) {
 413                         if (ra.cra_end_idx == 0 || ra.cra_end_idx < page_idx) {
 414                                 pgoff_t end_idx;
 415
 416                                 /*
 417                                  * Do not shrink ria_end_idx at any case until
 418                                  * the minimum end of current read is covered.
 419                                  *
 420                                  * Do not extend read lock accross stripe if
 421                                  * lock contention detected.
 422                                  */
 423                                 if (ra.cra_contention &&
 424                                     page_idx > ria->ria_end_idx_min) {
 425                                         ria->ria_end_idx = *ra_end;
 426                                         break;
 427                                 }
 428
 429                                 cl_read_ahead_release(env, &ra);
 430
 431                                 rc = cl_io_read_ahead(env, io, page_idx, &ra);
 432                                 if (rc < 0)
 433                                         break;
 434
 435                                  /*
 436                                   * Only shrink ria_end_idx if the matched
 437                                   * LDLM lock doesn't cover more.
 438                                   */
 439                                 if (page_idx > ra.cra_end_idx) {
 440                                         ria->ria_end_idx = ra.cra_end_idx;
 441                                         break;
 442                                 }
 443
 444                                 CDEBUG(D_READA, "idx: %lu, ra: %lu, rpc: %lu\n",
 445                                        page_idx, ra.cra_end_idx,
 446                                        ra.cra_rpc_pages);
 447                                 LASSERTF(ra.cra_end_idx >= page_idx,
 448                                          "object: %p, indcies %lu / %lu\n",
 449                                          io->ci_obj, ra.cra_end_idx, page_idx);
 450                                 /* update read ahead RPC size.
 451                                  * NB: it's racy but doesn't matter */
 452                                 if (ras->ras_rpc_pages != ra.cra_rpc_pages &&
 453                                     ra.cra_rpc_pages > 0)
 454                                         ras->ras_rpc_pages = ra.cra_rpc_pages;
 455                                 if (!skip_index) {
 456                                         /* trim it to align with optimal RPC size */
 457                                         end_idx = ras_align(ras, ria->ria_end_idx + 1);
 458                                         if (end_idx > 0 && !ria->ria_eof)
 459                                                 ria->ria_end_idx = end_idx - 1;
 460                                 }
 461                                 if (ria->ria_end_idx < ria->ria_end_idx_min)
 462                                         ria->ria_end_idx = ria->ria_end_idx_min;
 463                         }
 464                         if (page_idx > ria->ria_end_idx)
 465                                 break;
 466
 467                         /* If the page is inside the read-ahead window */
 468                         rc = ll_read_ahead_page(env, io, queue, page_idx,
 469                                                 MAYNEED);
 470                         if (rc < 0 && rc != -EBUSY)
 471                                 break;
 472                         if (rc == -EBUSY) {
 473                                 busy_page_count++;
 474                                 CDEBUG(D_READA,
 475                                        "skip busy page: %lu\n", page_idx);
 476                                 /* For page unaligned readahead the first
 477                                  * last pages of each region can be read by
 478                                  * another reader on the same node, and so
 479                                  * may be busy. So only stop for > 2 busy
 480                                  * pages. */
 481                                 if (busy_page_count > 2)
 482                                         break;
 483                         }
 484
 485                         *ra_end = page_idx;
 486                         /* Only subtract from reserve & count the page if we
 487                          * really did readahead on that page. */
 488                         if (rc == 0) {
 489                                 ria->ria_reserved--;
 490                                 count++;
 491                         }
 492                 } else if (stride_io_mode(ras)) {
 493                         /* If it is not in the read-ahead window, and it is
 494                          * read-ahead mode, then check whether it should skip
 495                          * the stride gap.
 496                          */
 497                         loff_t pos = (loff_t)page_idx << PAGE_SHIFT;
 498                         u64 offset;
 499
 500                         div64_u64_rem(pos - ria->ria_stoff, ria->ria_length,
 501                                       &offset);
 502                         if (offset >= ria->ria_bytes) {
 503                                 pos += (ria->ria_length - offset);
 504                                 if ((pos >> PAGE_SHIFT) >= page_idx + 1)
 505                                         page_idx = (pos >> PAGE_SHIFT) - 1;
 506                                 busy_page_count = 0;
 507                                 CDEBUG(D_READA,
 508                                        "Stride: jump %llu pages to %lu\n",
 509                                        ria->ria_length - offset, page_idx);
 510                                 continue;
 511                         }
 512                 }
 513         }
 514
 515         cl_read_ahead_release(env, &ra);
 516
 517         return count;
 518 }
 519
 520 static void ll_readahead_work_free(struct ll_readahead_work *work)
 521 {
 522         fput(work->lrw_file);
 523         OBD_FREE_PTR(work);
 524 }
 525
 526 static void ll_readahead_handle_work(struct work_struct *wq);
 527 static void ll_readahead_work_add(struct inode *inode,
 528                                   struct ll_readahead_work *work)
 529 {
 530         INIT_WORK(&work->lrw_readahead_work, ll_readahead_handle_work);
 531         queue_work(ll_i2sbi(inode)->ll_ra_info.ll_readahead_wq,
 532                    &work->lrw_readahead_work);
 533 }
 534
 535 static int ll_readahead_file_kms(const struct lu_env *env,
 536                                 struct cl_io *io, __u64 *kms)
 537 {
 538         struct cl_object *clob;
 539         struct inode *inode;
 540         struct cl_attr *attr = vvp_env_thread_attr(env);
 541         int ret;
 542
 543         clob = io->ci_obj;
 544         inode = vvp_object_inode(clob);
 545
 546         cl_object_attr_lock(clob);
 547         ret = cl_object_attr_get(env, clob, attr);
 548         cl_object_attr_unlock(clob);
 549
 550         if (ret != 0)
 551                 RETURN(ret);
 552
 553         *kms = attr->cat_kms;
 554         return 0;
 555 }
 556
 557 static void ll_readahead_handle_work(struct work_struct *wq)
 558 {
 559         struct ll_readahead_work *work;
 560         struct lu_env *env;
 561         __u16 refcheck;
 562         struct ra_io_arg *ria;
 563         struct inode *inode;
 564         struct ll_file_data *fd;
 565         struct ll_readahead_state *ras;
 566         struct cl_io *io;
 567         struct cl_2queue *queue;
 568         pgoff_t ra_end_idx = 0;
 569         unsigned long pages, pages_min = 0;
 570         struct file *file;
 571         __u64 kms;
 572         int rc;
 573         pgoff_t eof_index;
 574         struct ll_sb_info *sbi;
 575
 576         work = container_of(wq, struct ll_readahead_work,
 577                             lrw_readahead_work);
 578         fd = work->lrw_file->private_data;
 579         ras = &fd->fd_ras;
 580         file = work->lrw_file;
 581         inode = file_inode(file);
 582         sbi = ll_i2sbi(inode);
 583
 584         env = cl_env_alloc(&refcheck, LCT_NOREF);
 585         if (IS_ERR(env))
 586                 GOTO(out_free_work, rc = PTR_ERR(env));
 587
 588         io = vvp_env_thread_io(env);
 589         ll_io_init(io, file, CIT_READ, NULL);
 590
 591         rc = ll_readahead_file_kms(env, io, &kms);
 592         if (rc != 0)
 593                 GOTO(out_put_env, rc);
 594
 595         if (kms == 0) {
 596                 ll_ra_stats_inc(inode, RA_STAT_ZERO_LEN);
 597                 GOTO(out_put_env, rc = 0);
 598         }
 599
 600         ria = &ll_env_info(env)->lti_ria;
 601         memset(ria, 0, sizeof(*ria));
 602
 603         ria->ria_start_idx = work->lrw_start_idx;
 604         /* Truncate RA window to end of file */
 605         eof_index = (pgoff_t)(kms - 1) >> PAGE_SHIFT;
 606         if (eof_index <= work->lrw_end_idx) {
 607                 work->lrw_end_idx = eof_index;
 608                 ria->ria_eof = true;
 609         }
 610         if (work->lrw_end_idx <= work->lrw_start_idx)
 611                 GOTO(out_put_env, rc = 0);
 612
 613         ria->ria_end_idx = work->lrw_end_idx;
 614         pages = ria->ria_end_idx - ria->ria_start_idx + 1;
 615         ria->ria_reserved = ll_ra_count_get(sbi, ria,
 616                                             ria_page_count(ria), pages_min);
 617
 618         CDEBUG(D_READA,
 619                "async reserved pages: %lu/%lu/%lu, ra_cur %d, ra_max %lu\n",
 620                ria->ria_reserved, pages, pages_min,
 621                atomic_read(&ll_i2sbi(inode)->ll_ra_info.ra_cur_pages),
 622                ll_i2sbi(inode)->ll_ra_info.ra_max_pages);
 623
 624         if (ria->ria_reserved < pages) {
 625                 ll_ra_stats_inc(inode, RA_STAT_MAX_IN_FLIGHT);
 626                 if (PAGES_TO_MiB(ria->ria_reserved) < 1) {
 627                         ll_ra_count_put(ll_i2sbi(inode), ria->ria_reserved);
 628                         GOTO(out_put_env, rc = 0);
 629                 }
 630         }
 631
 632         rc = cl_io_rw_init(env, io, CIT_READ, ria->ria_start_idx, pages);
 633         if (rc)
 634                 GOTO(out_put_env, rc);
 635
 636         /* overwrite jobid inited in vvp_io_init() */
 637         if (strncmp(ll_i2info(inode)->lli_jobid, work->lrw_jobid,
 638                     sizeof(work->lrw_jobid)))
 639                 memcpy(ll_i2info(inode)->lli_jobid, work->lrw_jobid,
 640                        sizeof(work->lrw_jobid));
 641
 642         vvp_env_io(env)->vui_fd = fd;
 643         io->ci_state = CIS_LOCKED;
 644         io->ci_async_readahead = true;
 645         rc = cl_io_start(env, io);
 646         if (rc)
 647                 GOTO(out_io_fini, rc);
 648
 649         queue = &io->ci_queue;
 650         cl_2queue_init(queue);
 651
 652         rc = ll_read_ahead_pages(env, io, &queue->c2_qin, ras, ria,
 653                                  &ra_end_idx, 0);
 654         if (ria->ria_reserved != 0)
 655                 ll_ra_count_put(ll_i2sbi(inode), ria->ria_reserved);
 656         if (queue->c2_qin.pl_nr > 0) {
 657                 int count = queue->c2_qin.pl_nr;
 658
 659                 rc = cl_io_submit_rw(env, io, CRT_READ, queue);
 660                 if (rc == 0)
 661                         task_io_account_read(PAGE_SIZE * count);
 662         }
 663         if (ria->ria_end_idx == ra_end_idx && ra_end_idx == (kms >> PAGE_SHIFT))
 664                 ll_ra_stats_inc(inode, RA_STAT_EOF);
 665
 666         if (ra_end_idx != ria->ria_end_idx)
 667                 ll_ra_stats_inc(inode, RA_STAT_FAILED_REACH_END);
 668
 669         /* TODO: discard all pages until page reinit route is implemented */
 670         cl_page_list_discard(env, io, &queue->c2_qin);
 671
 672         /* Unlock unsent read pages in case of error. */
 673         cl_page_list_disown(env, io, &queue->c2_qin);
 674
 675         cl_2queue_fini(env, queue);
 676 out_io_fini:
 677         cl_io_end(env, io);
 678         cl_io_fini(env, io);
 679 out_put_env:
 680         cl_env_put(env, &refcheck);
 681 out_free_work:
 682         if (ra_end_idx > 0)
 683                 ll_ra_stats_inc_sbi(ll_i2sbi(inode), RA_STAT_ASYNC);
 684         atomic_dec(&sbi->ll_ra_info.ra_async_inflight);
 685         ll_readahead_work_free(work);
 686 }
 687
 688 static int ll_readahead(const struct lu_env *env, struct cl_io *io,
 689                         struct cl_page_list *queue,
 690                         struct ll_readahead_state *ras, bool hit,
 691                         struct file *file, pgoff_t skip_index)
 692 {
 693         struct vvp_io *vio = vvp_env_io(env);
 694         struct ll_thread_info *lti = ll_env_info(env);
 695         unsigned long pages, pages_min = 0;
 696         pgoff_t ra_end_idx = 0, start_idx = 0, end_idx = 0;
 697         struct inode *inode;
 698         struct ra_io_arg *ria = &lti->lti_ria;
 699         struct cl_object *clob;
 700         int ret = 0;
 701         __u64 kms;
 702         struct ll_sb_info *sbi;
 703         struct ll_ra_info *ra;
 704
 705         ENTRY;
 706
 707         ENTRY;
 708
 709         clob = io->ci_obj;
 710         inode = vvp_object_inode(clob);
 711         sbi = ll_i2sbi(inode);
 712         ra = &sbi->ll_ra_info;
 713
 714         /**
 715          * In case we have a limited max_cached_mb, readahead
 716          * should be stopped if it have run out of all LRU slots.
 717          */
 718         if (atomic_read(&ra->ra_cur_pages) >= sbi->ll_cache->ccc_lru_max) {
 719                 ll_ra_stats_inc(inode, RA_STAT_MAX_IN_FLIGHT);
 720                 RETURN(0);
 721         }
 722
 723         memset(ria, 0, sizeof(*ria));
 724         ret = ll_readahead_file_kms(env, io, &kms);
 725         if (ret != 0)
 726                 RETURN(ret);
 727
 728         if (kms == 0) {
 729                 ll_ra_stats_inc(inode, RA_STAT_ZERO_LEN);
 730                 RETURN(0);
 731         }
 732
 733         spin_lock(&ras->ras_lock);
 734
 735         /**
 736          * Note: other thread might rollback the ras_next_readahead_idx,
 737          * if it can not get the full size of prepared pages, see the
 738          * end of this function. For stride read ahead, it needs to
 739          * make sure the offset is no less than ras_stride_offset,
 740          * so that stride read ahead can work correctly.
 741          */
 742         if (stride_io_mode(ras))
 743                 start_idx = max_t(pgoff_t, ras->ras_next_readahead_idx,
 744                                   ras->ras_stride_offset >> PAGE_SHIFT);
 745         else
 746                 start_idx = ras->ras_next_readahead_idx;
 747
 748         if (ras->ras_window_pages > 0)
 749                 end_idx = ras->ras_window_start_idx + ras->ras_window_pages - 1;
 750
 751         if (skip_index)
 752                 end_idx = start_idx + ras->ras_window_pages - 1;
 753
 754         /* Enlarge the RA window to encompass the full read */
 755         if (vio->vui_ra_valid &&
 756             end_idx < vio->vui_ra_start_idx + vio->vui_ra_pages - 1)
 757                 end_idx = vio->vui_ra_start_idx + vio->vui_ra_pages - 1;
 758
 759         if (end_idx != 0) {
 760                 pgoff_t eof_index;
 761
 762                 /* Truncate RA window to end of file */
 763                 eof_index = (pgoff_t)((kms - 1) >> PAGE_SHIFT);
 764                 if (eof_index <= end_idx) {
 765                         end_idx = eof_index;
 766                         ria->ria_eof = true;
 767                 }
 768         }
 769         ria->ria_start_idx = start_idx;
 770         ria->ria_end_idx = end_idx;
 771         /* If stride I/O mode is detected, get stride window*/
 772         if (stride_io_mode(ras)) {
 773                 ria->ria_stoff = ras->ras_stride_offset;
 774                 ria->ria_length = ras->ras_stride_length;
 775                 ria->ria_bytes = ras->ras_stride_bytes;
 776         }
 777         spin_unlock(&ras->ras_lock);
 778
 779         if (end_idx == 0) {
 780                 ll_ra_stats_inc(inode, RA_STAT_ZERO_WINDOW);
 781                 RETURN(0);
 782         }
 783         pages = ria_page_count(ria);
 784         if (pages == 0) {
 785                 ll_ra_stats_inc(inode, RA_STAT_ZERO_WINDOW);
 786                 RETURN(0);
 787         }
 788
 789         RAS_CDEBUG(ras);
 790         CDEBUG(D_READA, DFID": ria: %lu/%lu, bead: %lu/%lu, hit: %d\n",
 791                PFID(lu_object_fid(&clob->co_lu)),
 792                ria->ria_start_idx, ria->ria_end_idx,
 793                vio->vui_ra_valid ? vio->vui_ra_start_idx : 0,
 794                vio->vui_ra_valid ? vio->vui_ra_pages : 0,
 795                hit);
 796
 797         /* at least to extend the readahead window to cover current read */
 798         if (!hit && vio->vui_ra_valid &&
 799             vio->vui_ra_start_idx + vio->vui_ra_pages > ria->ria_start_idx) {
 800                 ria->ria_end_idx_min =
 801                         vio->vui_ra_start_idx + vio->vui_ra_pages - 1;
 802                 pages_min = vio->vui_ra_start_idx + vio->vui_ra_pages -
 803                                 ria->ria_start_idx;
 804                  /**
 805                   * For performance reason, exceeding @ra_max_pages
 806                   * are allowed, but this should be limited with RPC
 807                   * size in case a large block size read issued. Trim
 808                   * to RPC boundary.
 809                   */
 810                 pages_min = min(pages_min, ras->ras_rpc_pages -
 811                                 (ria->ria_start_idx % ras->ras_rpc_pages));
 812         }
 813
 814         /* don't over reserved for mmap range read */
 815         if (skip_index)
 816                 pages_min = 0;
 817
 818         ria->ria_reserved = ll_ra_count_get(ll_i2sbi(inode), ria, pages,
 819                                             pages_min);
 820         if (ria->ria_reserved < pages)
 821                 ll_ra_stats_inc(inode, RA_STAT_MAX_IN_FLIGHT);
 822
 823         CDEBUG(D_READA, "reserved pages: %lu/%lu/%lu, ra_cur %d, ra_max %lu\n",
 824                ria->ria_reserved, pages, pages_min,
 825                atomic_read(&ll_i2sbi(inode)->ll_ra_info.ra_cur_pages),
 826                ll_i2sbi(inode)->ll_ra_info.ra_max_pages);
 827
 828         ret = ll_read_ahead_pages(env, io, queue, ras, ria, &ra_end_idx,
 829                                   skip_index);
 830         if (ria->ria_reserved != 0)
 831                 ll_ra_count_put(ll_i2sbi(inode), ria->ria_reserved);
 832
 833         if (ra_end_idx == end_idx && ra_end_idx == (kms >> PAGE_SHIFT))
 834                 ll_ra_stats_inc(inode, RA_STAT_EOF);
 835
 836         CDEBUG(D_READA,
 837                "ra_end_idx = %lu end_idx = %lu stride end = %lu pages = %d\n",
 838                ra_end_idx, end_idx, ria->ria_end_idx, ret);
 839
 840         if (ra_end_idx != end_idx)
 841                 ll_ra_stats_inc(inode, RA_STAT_FAILED_REACH_END);
 842         if (ra_end_idx > 0) {
 843                 /* update the ras so that the next read-ahead tries from
 844                  * where we left off. */
 845                 spin_lock(&ras->ras_lock);
 846                 ras->ras_next_readahead_idx = ra_end_idx + 1;
 847                 spin_unlock(&ras->ras_lock);
 848                 RAS_CDEBUG(ras);
 849         }
 850
 851         RETURN(ret);
 852 }
 853
 854 static int ll_readpages(const struct lu_env *env, struct cl_io *io,
 855                         struct cl_page_list *queue,
 856                         pgoff_t start, pgoff_t end)
 857 {
 858         int ret = 0;
 859         __u64 kms;
 860         pgoff_t page_idx;
 861         int count = 0;
 862
 863         ENTRY;
 864
 865         ret = ll_readahead_file_kms(env, io, &kms);
 866         if (ret != 0)
 867                 RETURN(ret);
 868
 869         if (kms == 0)
 870                 RETURN(0);
 871
 872         if (end != 0) {
 873                 unsigned long end_index;
 874
 875                 end_index = (unsigned long)((kms - 1) >> PAGE_SHIFT);
 876                 if (end_index <= end)
 877                         end = end_index;
 878         }
 879
 880         for (page_idx = start; page_idx <= end; page_idx++) {
 881                 ret= ll_read_ahead_page(env, io, queue, page_idx,
 882                                         WILLNEED);
 883                 if (ret < 0)
 884                         break;
 885                 else if (ret == 0) /* ret 1 is already uptodate */
 886                         count++;
 887         }
 888
 889         RETURN(count > 0 ? count : ret);
 890 }
 891
 892 static void ras_set_start(struct ll_readahead_state *ras, pgoff_t index)
 893 {
 894         ras->ras_window_start_idx = ras_align(ras, index);
 895 }
 896
 897 /* called with the ras_lock held or from places where it doesn't matter */
 898 static void ras_reset(struct ll_readahead_state *ras, pgoff_t index)
 899 {
 900         ras->ras_consecutive_requests = 0;
 901         ras->ras_consecutive_bytes = 0;
 902         ras->ras_window_pages = 0;
 903         ras_set_start(ras, index);
 904         ras->ras_next_readahead_idx = max(ras->ras_window_start_idx, index + 1);
 905
 906         RAS_CDEBUG(ras);
 907 }
 908
 909 /* called with the ras_lock held or from places where it doesn't matter */
 910 static void ras_stride_reset(struct ll_readahead_state *ras)
 911 {
 912         ras->ras_consecutive_stride_requests = 0;
 913         ras->ras_stride_length = 0;
 914         ras->ras_stride_bytes = 0;
 915         RAS_CDEBUG(ras);
 916 }
 917
 918 void ll_readahead_init(struct inode *inode, struct ll_readahead_state *ras)
 919 {
 920         spin_lock_init(&ras->ras_lock);
 921         ras->ras_rpc_pages = PTLRPC_MAX_BRW_PAGES;
 922         ras_reset(ras, 0);
 923         ras->ras_last_read_end_bytes = 0;
 924         ras->ras_requests = 0;
 925         ras->ras_range_min_start_idx = 0;
 926         ras->ras_range_max_end_idx = 0;
 927         ras->ras_range_requests = 0;
 928         ras->ras_last_range_pages = 0;
 929 }
 930
 931 /*
 932  * Check whether the read request is in the stride window.
 933  * If it is in the stride window, return true, otherwise return false.
 934  */
 935 static bool read_in_stride_window(struct ll_readahead_state *ras,
 936                                   loff_t pos, loff_t count)
 937 {
 938         loff_t stride_gap;
 939
 940         if (ras->ras_stride_length == 0 || ras->ras_stride_bytes == 0 ||
 941             ras->ras_stride_bytes == ras->ras_stride_length)
 942                 return false;
 943
 944         stride_gap = pos - ras->ras_last_read_end_bytes - 1;
 945
 946         /* If it is contiguous read */
 947         if (stride_gap == 0)
 948                 return ras->ras_consecutive_bytes + count <=
 949                         ras->ras_stride_bytes;
 950
 951         /* Otherwise check the stride by itself */
 952         return (ras->ras_stride_length - ras->ras_stride_bytes) == stride_gap &&
 953                 ras->ras_consecutive_bytes == ras->ras_stride_bytes &&
 954                 count <= ras->ras_stride_bytes;
 955 }
 956
 957 static void ras_init_stride_detector(struct ll_readahead_state *ras,
 958                                      loff_t pos, loff_t count)
 959 {
 960         loff_t stride_gap = pos - ras->ras_last_read_end_bytes - 1;
 961
 962         LASSERT(ras->ras_consecutive_stride_requests == 0);
 963
 964         if (pos <= ras->ras_last_read_end_bytes) {
 965                 /*Reset stride window for forward read*/
 966                 ras_stride_reset(ras);
 967                 return;
 968         }
 969
 970         ras->ras_stride_bytes = ras->ras_consecutive_bytes;
 971         ras->ras_stride_length = stride_gap + ras->ras_consecutive_bytes;
 972         ras->ras_consecutive_stride_requests++;
 973         ras->ras_stride_offset = pos;
 974
 975         RAS_CDEBUG(ras);
 976 }
 977
 978 static unsigned long
 979 stride_page_count(struct ll_readahead_state *ras, loff_t len)
 980 {
 981         loff_t bytes_count =
 982                 stride_byte_count(ras->ras_stride_offset,
 983                                   ras->ras_stride_length, ras->ras_stride_bytes,
 984                                   ras->ras_window_start_idx << PAGE_SHIFT, len);
 985
 986         return (bytes_count + PAGE_SIZE - 1) >> PAGE_SHIFT;
 987 }
 988
 989 /* Stride Read-ahead window will be increased inc_len according to
 990  * stride I/O pattern */
 991 static void ras_stride_increase_window(struct ll_readahead_state *ras,
 992                                        struct ll_ra_info *ra, loff_t inc_bytes)
 993 {
 994         loff_t window_bytes, stride_bytes;
 995         u64 left_bytes;
 996         u64 step;
 997         loff_t end;
 998
 999         /* temporarily store in page units to reduce LASSERT() cost below */
1000         end = ras->ras_window_start_idx + ras->ras_window_pages;
1001
1002         LASSERT(ras->ras_stride_length > 0);
1003         LASSERTF(end >= (ras->ras_stride_offset >> PAGE_SHIFT),
1004                  "window_start_idx %lu, window_pages %lu stride_offset %llu\n",
1005                  ras->ras_window_start_idx, ras->ras_window_pages,
1006                  ras->ras_stride_offset);
1007
1008         end <<= PAGE_SHIFT;
1009         if (end <= ras->ras_stride_offset)
1010                 stride_bytes = 0;
1011         else
1012                 stride_bytes = end - ras->ras_stride_offset;
1013
1014         div64_u64_rem(stride_bytes, ras->ras_stride_length, &left_bytes);
1015         window_bytes = (ras->ras_window_pages << PAGE_SHIFT);
1016         if (left_bytes < ras->ras_stride_bytes) {
1017                 if (ras->ras_stride_bytes - left_bytes >= inc_bytes) {
1018                         window_bytes += inc_bytes;
1019                         goto out;
1020                 } else {
1021                         window_bytes += (ras->ras_stride_bytes - left_bytes);
1022                         inc_bytes -= (ras->ras_stride_bytes - left_bytes);
1023                 }
1024         } else {
1025                 window_bytes += (ras->ras_stride_length - left_bytes);
1026         }
1027
1028         LASSERT(ras->ras_stride_bytes != 0);
1029
1030         step = div64_u64_rem(inc_bytes, ras->ras_stride_bytes, &left_bytes);
1031
1032         window_bytes += step * ras->ras_stride_length + left_bytes;
1033         LASSERT(window_bytes > 0);
1034
1035 out:
1036         if (stride_page_count(ras, window_bytes) <=
1037             ra->ra_max_pages_per_file || ras->ras_window_pages == 0)
1038                 ras->ras_window_pages = (window_bytes >> PAGE_SHIFT);
1039
1040         LASSERT(ras->ras_window_pages > 0);
1041
1042         RAS_CDEBUG(ras);
1043 }
1044
1045 static void ras_increase_window(struct inode *inode,
1046                                 struct ll_readahead_state *ras,
1047                                 struct ll_ra_info *ra)
1048 {
1049         /* The stretch of ra-window should be aligned with max rpc_size
1050          * but current clio architecture does not support retrieve such
1051          * information from lower layer. FIXME later
1052          */
1053         if (stride_io_mode(ras)) {
1054                 ras_stride_increase_window(ras, ra,
1055                                       (loff_t)ras->ras_rpc_pages << PAGE_SHIFT);
1056         } else {
1057                 pgoff_t window_pages;
1058
1059                 window_pages = min(ras->ras_window_pages + ras->ras_rpc_pages,
1060                                    ra->ra_max_pages_per_file);
1061                 if (window_pages < ras->ras_rpc_pages)
1062                         ras->ras_window_pages = window_pages;
1063                 else
1064                         ras->ras_window_pages = ras_align(ras, window_pages);
1065         }
1066 }
1067
1068 /**
1069  * Seek within 8 pages are considered as sequential read for now.
1070  */
1071 static inline bool is_loose_seq_read(struct ll_readahead_state *ras, loff_t pos)
1072 {
1073         return pos_in_window(pos, ras->ras_last_read_end_bytes,
1074                              8UL << PAGE_SHIFT, 8UL << PAGE_SHIFT);
1075 }
1076
1077 static inline bool is_loose_mmap_read(struct ll_sb_info *sbi,
1078                                       struct ll_readahead_state *ras,
1079                                       unsigned long pos)
1080 {
1081         unsigned long range_pages = sbi->ll_ra_info.ra_range_pages;
1082
1083         return pos_in_window(pos, ras->ras_last_read_end_bytes,
1084                              range_pages << PAGE_SHIFT,
1085                              range_pages << PAGE_SHIFT);
1086 }
1087
1088 /**
1089  * We have observed slow mmap read performances for some
1090  * applications. The problem is if access pattern is neither
1091  * sequential nor stride, but could be still adjacent in a
1092  * small range and then seek a random position.
1093  *
1094  * So the pattern could be something like this:
1095  *
1096  * [1M data] [hole] [0.5M data] [hole] [0.7M data] [1M data]
1097  *
1098  *
1099  * Every time an application reads mmap data, it may not only
1100  * read a single 4KB page, but aslo a cluster of nearby pages in
1101  * a range(e.g. 1MB) of the first page after a cache miss.
1102  *
1103  * The readahead engine is modified to track the range size of
1104  * a cluster of mmap reads, so that after a seek and/or cache miss,
1105  * the range size is used to efficiently prefetch multiple pages
1106  * in a single RPC rather than many small RPCs.
1107  */
1108 static void ras_detect_cluster_range(struct ll_readahead_state *ras,
1109                                      struct ll_sb_info *sbi,
1110                                      unsigned long pos, unsigned long count)
1111 {
1112         pgoff_t last_pages, pages;
1113         pgoff_t end_idx = (pos + count - 1) >> PAGE_SHIFT;
1114
1115         last_pages = ras->ras_range_max_end_idx -
1116                         ras->ras_range_min_start_idx + 1;
1117         /* First time come here */
1118         if (!ras->ras_range_max_end_idx)
1119                 goto out;
1120
1121         /* Random or Stride read */
1122         if (!is_loose_mmap_read(sbi, ras, pos))
1123                 goto out;
1124
1125         ras->ras_range_requests++;
1126         if (ras->ras_range_max_end_idx < end_idx)
1127                 ras->ras_range_max_end_idx = end_idx;
1128
1129         if (ras->ras_range_min_start_idx > (pos >> PAGE_SHIFT))
1130                 ras->ras_range_min_start_idx = pos >> PAGE_SHIFT;
1131
1132         /* Out of range, consider it as random or stride */
1133         pages = ras->ras_range_max_end_idx -
1134                         ras->ras_range_min_start_idx + 1;
1135         if (pages <= sbi->ll_ra_info.ra_range_pages)
1136                 return;
1137 out:
1138         ras->ras_last_range_pages = last_pages;
1139         ras->ras_range_requests = 0;
1140         ras->ras_range_min_start_idx = pos >> PAGE_SHIFT;
1141         ras->ras_range_max_end_idx = end_idx;
1142 }
1143
1144 static void ras_detect_read_pattern(struct ll_readahead_state *ras,
1145                                     struct ll_sb_info *sbi,
1146                                     loff_t pos, size_t count, bool mmap)
1147 {
1148         bool stride_detect = false;
1149         pgoff_t index = pos >> PAGE_SHIFT;
1150
1151         /*
1152          * Reset the read-ahead window in two cases. First when the app seeks
1153          * or reads to some other part of the file. Secondly if we get a
1154          * read-ahead miss that we think we've previously issued. This can
1155          * be a symptom of there being so many read-ahead pages that the VM
1156          * is reclaiming it before we get to it.
1157          */
1158         if (!is_loose_seq_read(ras, pos)) {
1159                 /* Check whether it is in stride I/O mode */
1160                 if (!read_in_stride_window(ras, pos, count)) {
1161                         if (ras->ras_consecutive_stride_requests == 0)
1162                                 ras_init_stride_detector(ras, pos, count);
1163                         else
1164                                 ras_stride_reset(ras);
1165                         ras->ras_consecutive_bytes = 0;
1166                         ras_reset(ras, index);
1167                 } else {
1168                         ras->ras_consecutive_bytes = 0;
1169                         ras->ras_consecutive_requests = 0;
1170                         if (++ras->ras_consecutive_stride_requests > 1)
1171                                 stride_detect = true;
1172                         RAS_CDEBUG(ras);
1173                 }
1174                 ll_ra_stats_inc_sbi(sbi, RA_STAT_DISTANT_READPAGE);
1175         } else if (stride_io_mode(ras)) {
1176                 /*
1177                  * If this is contiguous read but in stride I/O mode
1178                  * currently, check whether stride step still is valid,
1179                  * if invalid, it will reset the stride ra window to
1180                  * be zero.
1181                  */
1182                 if (!read_in_stride_window(ras, pos, count)) {
1183                         ras_stride_reset(ras);
1184                         ras->ras_window_pages = 0;
1185                         ras->ras_next_readahead_idx = index;
1186                 }
1187         }
1188
1189         ras->ras_consecutive_bytes += count;
1190         if (mmap) {
1191                 pgoff_t idx = ras->ras_consecutive_bytes >> PAGE_SHIFT;
1192                 unsigned long ra_range_pages =
1193                                 max_t(unsigned long, RA_MIN_MMAP_RANGE_PAGES,
1194                                       sbi->ll_ra_info.ra_range_pages);
1195
1196                 if ((idx >= ra_range_pages &&
1197                      idx % ra_range_pages == 0) || stride_detect)
1198                         ras->ras_need_increase_window = true;
1199         } else if ((ras->ras_consecutive_requests > 1 || stride_detect)) {
1200                 ras->ras_need_increase_window = true;
1201         }
1202
1203         ras->ras_last_read_end_bytes = pos + count - 1;
1204 }
1205
1206 void ll_ras_enter(struct file *f, loff_t pos, size_t count)
1207 {
1208         struct ll_file_data *fd = f->private_data;
1209         struct ll_readahead_state *ras = &fd->fd_ras;
1210         struct inode *inode = file_inode(f);
1211         unsigned long index = pos >> PAGE_SHIFT;
1212         struct ll_sb_info *sbi = ll_i2sbi(inode);
1213
1214         spin_lock(&ras->ras_lock);
1215         ras->ras_requests++;
1216         ras->ras_consecutive_requests++;
1217         ras->ras_need_increase_window = false;
1218         ras->ras_no_miss_check = false;
1219         /*
1220          * On the second access to a file smaller than the tunable
1221          * ra_max_read_ahead_whole_pages trigger RA on all pages in the
1222          * file up to ra_max_pages_per_file.  This is simply a best effort
1223          * and only occurs once per open file. Normal RA behavior is reverted
1224          * to for subsequent IO.
1225          */
1226         if (ras->ras_requests >= 2) {
1227                 __u64 kms_pages;
1228                 struct ll_ra_info *ra = &sbi->ll_ra_info;
1229
1230                 kms_pages = (i_size_read(inode) + PAGE_SIZE - 1) >>
1231                             PAGE_SHIFT;
1232
1233                 CDEBUG(D_READA, "kmsp %llu mwp %lu mp %lu\n", kms_pages,
1234                        ra->ra_max_read_ahead_whole_pages,
1235                        ra->ra_max_pages_per_file);
1236
1237                 if (kms_pages &&
1238                     kms_pages <= ra->ra_max_read_ahead_whole_pages) {
1239                         ras->ras_window_start_idx = 0;
1240                         ras->ras_next_readahead_idx = index + 1;
1241                         ras->ras_window_pages = min(ra->ra_max_pages_per_file,
1242                                             ra->ra_max_read_ahead_whole_pages);
1243                         ras->ras_no_miss_check = true;
1244                         GOTO(out_unlock, 0);
1245                 }
1246         }
1247         ras_detect_read_pattern(ras, sbi, pos, count, false);
1248 out_unlock:
1249         spin_unlock(&ras->ras_lock);
1250 }
1251
1252 static bool index_in_stride_window(struct ll_readahead_state *ras,
1253                                    pgoff_t index)
1254 {
1255         loff_t pos = (loff_t)index << PAGE_SHIFT;
1256
1257         if (ras->ras_stride_length == 0 || ras->ras_stride_bytes == 0 ||
1258             ras->ras_stride_bytes == ras->ras_stride_length)
1259                 return false;
1260
1261         if (pos >= ras->ras_stride_offset) {
1262                 u64 offset;
1263
1264                 div64_u64_rem(pos - ras->ras_stride_offset,
1265                               ras->ras_stride_length, &offset);
1266                 if (offset < ras->ras_stride_bytes ||
1267                     ras->ras_stride_length - offset < PAGE_SIZE)
1268                         return true;
1269         } else if (ras->ras_stride_offset - pos < PAGE_SIZE) {
1270                 return true;
1271         }
1272
1273         return false;
1274 }
1275
1276 /*
1277  * ll_ras_enter() is used to detect read pattern according to pos and count.
1278  *
1279  * ras_update() is used to detect cache miss and
1280  * reset window or increase window accordingly
1281  */
1282 static void ras_update(struct ll_sb_info *sbi, struct inode *inode,
1283                        struct ll_readahead_state *ras, pgoff_t index,
1284                        enum ras_update_flags flags, struct cl_io *io)
1285 {
1286         struct ll_ra_info *ra = &sbi->ll_ra_info;
1287         bool hit = flags & LL_RAS_HIT;
1288
1289         ENTRY;
1290         spin_lock(&ras->ras_lock);
1291
1292         if (!hit)
1293                 CDEBUG(D_READA, DFID " pages at %lu miss.\n",
1294                        PFID(ll_inode2fid(inode)), index);
1295         ll_ra_stats_inc_sbi(sbi, hit ? RA_STAT_HIT : RA_STAT_MISS);
1296
1297         /*
1298          * The readahead window has been expanded to cover whole
1299          * file size, we don't care whether ra miss happen or not.
1300          * Because we will read whole file to page cache even if
1301          * some pages missed.
1302          */
1303         if (ras->ras_no_miss_check)
1304                 GOTO(out_unlock, 0);
1305
1306         if (io && io->ci_rand_read)
1307                 GOTO(out_unlock, 0);
1308
1309         if (io && io->ci_seq_read) {
1310                 if (!hit) {
1311                         /* to avoid many small read RPC here */
1312                         ras->ras_window_pages = sbi->ll_ra_info.ra_range_pages;
1313                         ll_ra_stats_inc_sbi(sbi, RA_STAT_MMAP_RANGE_READ);
1314                 }
1315                 goto skip;
1316         }
1317
1318         if (flags & LL_RAS_MMAP) {
1319                 unsigned long ra_pages;
1320
1321                 ras_detect_cluster_range(ras, sbi, index << PAGE_SHIFT,
1322                                          PAGE_SIZE);
1323                 ras_detect_read_pattern(ras, sbi, (loff_t)index << PAGE_SHIFT,
1324                                         PAGE_SIZE, true);
1325
1326                 /* we did not detect anything but we could prefetch */
1327                 if (!ras->ras_need_increase_window &&
1328                     ras->ras_window_pages <= sbi->ll_ra_info.ra_range_pages &&
1329                     ras->ras_range_requests >= 2) {
1330                         if (!hit) {
1331                                 ra_pages = max_t(unsigned long,
1332                                         RA_MIN_MMAP_RANGE_PAGES,
1333                                         ras->ras_last_range_pages);
1334                                 if (index < ra_pages / 2)
1335                                         index = 0;
1336                                 else
1337                                         index -= ra_pages / 2;
1338                                 ras->ras_window_pages = ra_pages;
1339                                 ll_ra_stats_inc_sbi(sbi,
1340                                         RA_STAT_MMAP_RANGE_READ);
1341                         } else {
1342                                 ras->ras_window_pages = 0;
1343                         }
1344                         goto skip;
1345                 }
1346         }
1347
1348         if (!hit && ras->ras_window_pages &&
1349             index < ras->ras_next_readahead_idx &&
1350             pos_in_window(index, ras->ras_window_start_idx, 0,
1351                           ras->ras_window_pages)) {
1352                 ll_ra_stats_inc_sbi(sbi, RA_STAT_MISS_IN_WINDOW);
1353                 ras->ras_need_increase_window = false;
1354
1355                 if (index_in_stride_window(ras, index) &&
1356                     stride_io_mode(ras)) {
1357                         /*
1358                          * if (index != ras->ras_last_readpage + 1)
1359                          *      ras->ras_consecutive_pages = 0;
1360                          */
1361                         ras_reset(ras, index);
1362
1363                         /*
1364                          * If stride-RA hit cache miss, the stride
1365                          * detector will not be reset to avoid the
1366                          * overhead of redetecting read-ahead mode,
1367                          * but on the condition that the stride window
1368                          * is still intersect with normal sequential
1369                          * read-ahead window.
1370                          */
1371                         if (ras->ras_window_start_idx < ras->ras_stride_offset)
1372                                 ras_stride_reset(ras);
1373                         RAS_CDEBUG(ras);
1374                 } else {
1375                         /*
1376                          * Reset both stride window and normal RA
1377                          * window.
1378                          */
1379                         ras_reset(ras, index);
1380                         /* ras->ras_consecutive_pages++; */
1381                         ras->ras_consecutive_bytes = 0;
1382                         ras_stride_reset(ras);
1383                         GOTO(out_unlock, 0);
1384                 }
1385         }
1386
1387 skip:
1388         ras_set_start(ras, index);
1389
1390         if (stride_io_mode(ras)) {
1391                 /* Since stride readahead is sentivite to the offset
1392                  * of read-ahead, so we use original offset here,
1393                  * instead of ras_window_start_idx, which is RPC aligned.
1394                  */
1395                 ras->ras_next_readahead_idx = max(index + 1,
1396                                                   ras->ras_next_readahead_idx);
1397                 ras->ras_window_start_idx =
1398                                 max_t(pgoff_t, ras->ras_window_start_idx,
1399                                       ras->ras_stride_offset >> PAGE_SHIFT);
1400         } else {
1401                 if (ras->ras_next_readahead_idx < ras->ras_window_start_idx)
1402                         ras->ras_next_readahead_idx = ras->ras_window_start_idx;
1403                 if (!hit)
1404                         ras->ras_next_readahead_idx = index + 1;
1405         }
1406
1407         if (ras->ras_need_increase_window) {
1408                 ras_increase_window(inode, ras, ra);
1409                 ras->ras_need_increase_window = false;
1410         }
1411
1412         EXIT;
1413 out_unlock:
1414         spin_unlock(&ras->ras_lock);
1415 }
1416
1417 int ll_writepage(struct page *vmpage, struct writeback_control *wbc)
1418 {
1419         struct inode           *inode = vmpage->mapping->host;
1420         struct ll_inode_info   *lli   = ll_i2info(inode);
1421         struct lu_env          *env;
1422         struct cl_io           *io;
1423         struct cl_page         *page;
1424         struct cl_object       *clob;
1425         bool redirtied = false;
1426         bool unlocked = false;
1427         int result;
1428         __u16 refcheck;
1429         ENTRY;
1430
1431         LASSERT(PageLocked(vmpage));
1432         LASSERT(!PageWriteback(vmpage));
1433
1434         LASSERT(ll_i2dtexp(inode) != NULL);
1435
1436         env = cl_env_get(&refcheck);
1437         if (IS_ERR(env))
1438                 GOTO(out, result = PTR_ERR(env));
1439
1440         clob  = ll_i2info(inode)->lli_clob;
1441         LASSERT(clob != NULL);
1442
1443         io = vvp_env_thread_io(env);
1444         io->ci_obj = clob;
1445         io->ci_ignore_layout = 1;
1446         result = cl_io_init(env, io, CIT_MISC, clob);
1447         if (result == 0) {
1448                 page = cl_page_find(env, clob, vmpage->index,
1449                                     vmpage, CPT_CACHEABLE);
1450                 if (!IS_ERR(page)) {
1451                         lu_ref_add(&page->cp_reference, "writepage",
1452                                    current);
1453                         cl_page_assume(env, io, page);
1454                         result = cl_page_flush(env, io, page);
1455                         if (result != 0) {
1456                                 /*
1457                                  * Re-dirty page on error so it retries write,
1458                                  * but not in case when IO has actually
1459                                  * occurred and completed with an error.
1460                                  */
1461                                 if (!PageError(vmpage)) {
1462                                         redirty_page_for_writepage(wbc, vmpage);
1463                                         result = 0;
1464                                         redirtied = true;
1465                                 }
1466                         }
1467                         cl_page_disown(env, io, page);
1468                         unlocked = true;
1469                         lu_ref_del(&page->cp_reference,
1470                                    "writepage", current);
1471                         cl_page_put(env, page);
1472                 } else {
1473                         result = PTR_ERR(page);
1474                 }
1475         }
1476         cl_io_fini(env, io);
1477
1478         if (redirtied && wbc->sync_mode == WB_SYNC_ALL) {
1479                 loff_t offset = cl_offset(clob, vmpage->index);
1480
1481                 /* Flush page failed because the extent is being written out.
1482                  * Wait for the write of extent to be finished to avoid
1483                  * breaking kernel which assumes ->writepage should mark
1484                  * PageWriteback or clean the page. */
1485                 result = cl_sync_file_range(inode, offset,
1486                                             offset + PAGE_SIZE - 1,
1487                                             CL_FSYNC_LOCAL, 1);
1488                 if (result > 0) {
1489                         /* actually we may have written more than one page.
1490                          * decreasing this page because the caller will count
1491                          * it. */
1492                         wbc->nr_to_write -= result - 1;
1493                         result = 0;
1494                 }
1495         }
1496
1497         cl_env_put(env, &refcheck);
1498         GOTO(out, result);
1499
1500 out:
1501         if (result < 0) {
1502                 if (!lli->lli_async_rc)
1503                         lli->lli_async_rc = result;
1504                 SetPageError(vmpage);
1505                 if (!unlocked)
1506                         unlock_page(vmpage);
1507         }
1508         return result;
1509 }
1510
1511 int ll_writepages(struct address_space *mapping, struct writeback_control *wbc)
1512 {
1513         struct inode *inode = mapping->host;
1514         loff_t start;
1515         loff_t end;
1516         enum cl_fsync_mode mode;
1517         int range_whole = 0;
1518         int result;
1519         ENTRY;
1520
1521         if (wbc->range_cyclic) {
1522                 start = (loff_t)mapping->writeback_index << PAGE_SHIFT;
1523                 end = OBD_OBJECT_EOF;
1524         } else {
1525                 start = wbc->range_start;
1526                 end = wbc->range_end;
1527                 if (end == LLONG_MAX) {
1528                         end = OBD_OBJECT_EOF;
1529                         range_whole = start == 0;
1530                 }
1531         }
1532
1533         mode = CL_FSYNC_NONE;
1534         if (wbc->sync_mode == WB_SYNC_ALL)
1535                 mode = CL_FSYNC_LOCAL;
1536
1537         if (ll_i2info(inode)->lli_clob == NULL)
1538                 RETURN(0);
1539
1540         /* for directio, it would call writepages() to evict cached pages
1541          * inside the IO context of write, which will cause deadlock at
1542          * layout_conf since it waits for active IOs to complete. */
1543         result = cl_sync_file_range(inode, start, end, mode, 1);
1544         if (result > 0) {
1545                 wbc->nr_to_write -= result;
1546                 result = 0;
1547          }
1548
1549         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0)) {
1550                 if (end == OBD_OBJECT_EOF)
1551                         mapping->writeback_index = 0;
1552                 else
1553                         mapping->writeback_index = (end >> PAGE_SHIFT) + 1;
1554         }
1555         RETURN(result);
1556 }
1557
1558 struct ll_cl_context *ll_cl_find(struct file *file)
1559 {
1560         struct ll_file_data *fd = file->private_data;
1561         struct ll_cl_context *lcc;
1562         struct ll_cl_context *found = NULL;
1563
1564         read_lock(&fd->fd_lock);
1565         list_for_each_entry(lcc, &fd->fd_lccs, lcc_list) {
1566                 if (lcc->lcc_cookie == current) {
1567                         found = lcc;
1568                         break;
1569                 }
1570         }
1571         read_unlock(&fd->fd_lock);
1572
1573         return found;
1574 }
1575
1576 void ll_cl_add(struct file *file, const struct lu_env *env, struct cl_io *io,
1577                enum lcc_type type)
1578 {
1579         struct ll_file_data *fd = file->private_data;
1580         struct ll_cl_context *lcc = &ll_env_info(env)->lti_io_ctx;
1581
1582         memset(lcc, 0, sizeof(*lcc));
1583         INIT_LIST_HEAD(&lcc->lcc_list);
1584         lcc->lcc_cookie = current;
1585         lcc->lcc_env = env;
1586         lcc->lcc_io = io;
1587         lcc->lcc_type = type;
1588
1589         write_lock(&fd->fd_lock);
1590         list_add(&lcc->lcc_list, &fd->fd_lccs);
1591         write_unlock(&fd->fd_lock);
1592 }
1593
1594 void ll_cl_remove(struct file *file, const struct lu_env *env)
1595 {
1596         struct ll_file_data *fd = file->private_data;
1597         struct ll_cl_context *lcc = &ll_env_info(env)->lti_io_ctx;
1598
1599         write_lock(&fd->fd_lock);
1600         list_del_init(&lcc->lcc_list);
1601         write_unlock(&fd->fd_lock);
1602 }
1603
1604 int ll_io_read_page(const struct lu_env *env, struct cl_io *io,
1605                            struct cl_page *page, struct file *file)
1606 {
1607         struct inode              *inode  = vvp_object_inode(page->cp_obj);
1608         struct ll_sb_info         *sbi    = ll_i2sbi(inode);
1609         struct ll_file_data       *fd     = NULL;
1610         struct ll_readahead_state *ras    = NULL;
1611         struct cl_2queue          *queue  = &io->ci_queue;
1612         struct cl_sync_io         *anchor = NULL;
1613         struct vvp_page           *vpg;
1614         int                        rc = 0, rc2 = 0;
1615         bool                       uptodate;
1616         pgoff_t io_start_index;
1617         pgoff_t io_end_index;
1618         ENTRY;
1619
1620         if (file) {
1621                 fd = file->private_data;
1622                 ras = &fd->fd_ras;
1623         }
1624
1625         vpg = cl2vvp_page(cl_object_page_slice(page->cp_obj, page));
1626         uptodate = vpg->vpg_defer_uptodate;
1627
1628         if (ll_readahead_enabled(sbi) && !vpg->vpg_ra_updated && ras) {
1629                 struct vvp_io *vio = vvp_env_io(env);
1630                 enum ras_update_flags flags = 0;
1631
1632                 if (uptodate)
1633                         flags |= LL_RAS_HIT;
1634                 if (!vio->vui_ra_valid)
1635                         flags |= LL_RAS_MMAP;
1636                 ras_update(sbi, inode, ras, vvp_index(vpg), flags, io);
1637         }
1638
1639         cl_2queue_init(queue);
1640         if (uptodate) {
1641                 vpg->vpg_ra_used = 1;
1642                 cl_page_export(env, page, 1);
1643                 cl_page_disown(env, io, page);
1644         } else {
1645                 anchor = &vvp_env_info(env)->vti_anchor;
1646                 cl_sync_io_init(anchor, 1);
1647                 page->cp_sync_io = anchor;
1648
1649                 cl_2queue_add(queue, page);
1650         }
1651
1652         io_start_index = cl_index(io->ci_obj, io->u.ci_rw.crw_pos);
1653         io_end_index = cl_index(io->ci_obj, io->u.ci_rw.crw_pos +
1654                                 io->u.ci_rw.crw_count - 1);
1655         if (ll_readahead_enabled(sbi) && ras && !io->ci_rand_read) {
1656                 pgoff_t skip_index = 0;
1657
1658                 if (ras->ras_next_readahead_idx < vvp_index(vpg))
1659                         skip_index = vvp_index(vpg);
1660                 rc2 = ll_readahead(env, io, &queue->c2_qin, ras,
1661                                    uptodate, file, skip_index);
1662                 CDEBUG(D_READA, DFID " %d pages read ahead at %lu\n",
1663                        PFID(ll_inode2fid(inode)), rc2, vvp_index(vpg));
1664         } else if (vvp_index(vpg) == io_start_index &&
1665                    io_end_index - io_start_index > 0) {
1666                 rc2 = ll_readpages(env, io, &queue->c2_qin, io_start_index + 1,
1667                                    io_end_index);
1668                 CDEBUG(D_READA, DFID " %d pages read at %lu\n",
1669                        PFID(ll_inode2fid(inode)), rc2, vvp_index(vpg));
1670         }
1671
1672         if (queue->c2_qin.pl_nr > 0) {
1673                 int count = queue->c2_qin.pl_nr;
1674                 rc = cl_io_submit_rw(env, io, CRT_READ, queue);
1675                 if (rc == 0)
1676                         task_io_account_read(PAGE_SIZE * count);
1677         }
1678
1679
1680         if (anchor != NULL && !cl_page_is_owned(page, io)) { /* have sent */
1681                 rc = cl_sync_io_wait(env, anchor, 0);
1682
1683                 cl_page_assume(env, io, page);
1684                 cl_page_list_del(env, &queue->c2_qout, page);
1685
1686                 if (!PageUptodate(cl_page_vmpage(page))) {
1687                         /* Failed to read a mirror, discard this page so that
1688                          * new page can be created with new mirror.
1689                          *
1690                          * TODO: this is not needed after page reinit
1691                          * route is implemented */
1692                         cl_page_discard(env, io, page);
1693                 }
1694                 cl_page_disown(env, io, page);
1695         }
1696
1697         /* TODO: discard all pages until page reinit route is implemented */
1698         cl_page_list_discard(env, io, &queue->c2_qin);
1699
1700         /* Unlock unsent read pages in case of error. */
1701         cl_page_list_disown(env, io, &queue->c2_qin);
1702
1703         cl_2queue_fini(env, queue);
1704
1705         RETURN(rc);
1706 }
1707
1708 /*
1709  * Possible return value:
1710  * 0 no async readahead triggered and fast read could not be used.
1711  * 1 no async readahead, but fast read could be used.
1712  * 2 async readahead triggered and fast read could be used too.
1713  * < 0 on error.
1714  */
1715 static int kickoff_async_readahead(struct file *file, unsigned long pages)
1716 {
1717         struct ll_readahead_work *lrw;
1718         struct inode *inode = file_inode(file);
1719         struct ll_sb_info *sbi = ll_i2sbi(inode);
1720         struct ll_file_data *fd = file->private_data;
1721         struct ll_readahead_state *ras = &fd->fd_ras;
1722         struct ll_ra_info *ra = &sbi->ll_ra_info;
1723         unsigned long throttle;
1724         pgoff_t start_idx = ras_align(ras, ras->ras_next_readahead_idx);
1725         pgoff_t end_idx = start_idx + pages - 1;
1726
1727         /**
1728          * In case we have a limited max_cached_mb, readahead
1729          * should be stopped if it have run out of all LRU slots.
1730          */
1731         if (atomic_read(&ra->ra_cur_pages) >= sbi->ll_cache->ccc_lru_max) {
1732                 ll_ra_stats_inc(inode, RA_STAT_MAX_IN_FLIGHT);
1733                 return 0;
1734         }
1735
1736         throttle = min(ra->ra_async_pages_per_file_threshold,
1737                        ra->ra_max_pages_per_file);
1738         /*
1739          * If this is strided i/o or the window is smaller than the
1740          * throttle limit, we do not do async readahead. Otherwise,
1741          * we do async readahead, allowing the user thread to do fast i/o.
1742          */
1743         if (stride_io_mode(ras) || !throttle ||
1744             ras->ras_window_pages < throttle ||
1745             atomic_read(&ra->ra_async_inflight) > ra->ra_async_max_active)
1746                 return 0;
1747
1748         if ((atomic_read(&ra->ra_cur_pages) + pages) > ra->ra_max_pages)
1749                 return 0;
1750
1751         if (ras->ras_async_last_readpage_idx == start_idx)
1752                 return 1;
1753
1754         /* ll_readahead_work_free() free it */
1755         OBD_ALLOC_PTR(lrw);
1756         if (lrw) {
1757                 atomic_inc(&sbi->ll_ra_info.ra_async_inflight);
1758                 lrw->lrw_file = get_file(file);
1759                 lrw->lrw_start_idx = start_idx;
1760                 lrw->lrw_end_idx = end_idx;
1761                 spin_lock(&ras->ras_lock);
1762                 ras->ras_next_readahead_idx = end_idx + 1;
1763                 ras->ras_async_last_readpage_idx = start_idx;
1764                 spin_unlock(&ras->ras_lock);
1765                 memcpy(lrw->lrw_jobid, ll_i2info(inode)->lli_jobid,
1766                        sizeof(lrw->lrw_jobid));
1767                 ll_readahead_work_add(inode, lrw);
1768         } else {
1769                 return -ENOMEM;
1770         }
1771
1772         return 2;
1773 }
1774
1775 /*
1776  * Check if we can issue a readahead RPC, if that is
1777  * the case, we can't do fast IO because we will need
1778  * a cl_io to issue the RPC.
1779  */
1780 static bool ll_use_fast_io(struct file *file,
1781                            struct ll_readahead_state *ras, pgoff_t index)
1782 {
1783         unsigned long fast_read_pages =
1784                 max(RA_REMAIN_WINDOW_MIN, ras->ras_rpc_pages);
1785         loff_t skip_pages;
1786         loff_t stride_bytes = ras->ras_stride_bytes;
1787
1788         if (stride_io_mode(ras) && stride_bytes) {
1789                 skip_pages = (ras->ras_stride_length +
1790                         ras->ras_stride_bytes - 1) / stride_bytes;
1791                 skip_pages *= fast_read_pages;
1792         } else {
1793                 skip_pages = fast_read_pages;
1794         }
1795
1796         if (ras->ras_window_start_idx + ras->ras_window_pages <
1797             ras->ras_next_readahead_idx + skip_pages ||
1798             kickoff_async_readahead(file, fast_read_pages) > 0)
1799                 return true;
1800
1801         return false;
1802 }
1803
1804 int ll_readpage(struct file *file, struct page *vmpage)
1805 {
1806         struct inode *inode = file_inode(file);
1807         struct cl_object *clob = ll_i2info(inode)->lli_clob;
1808         struct ll_cl_context *lcc;
1809         const struct lu_env  *env = NULL;
1810         struct cl_io   *io = NULL;
1811         struct cl_page *page;
1812         struct ll_sb_info *sbi = ll_i2sbi(inode);
1813         int result;
1814         ENTRY;
1815
1816         lcc = ll_cl_find(file);
1817         if (lcc != NULL) {
1818                 env = lcc->lcc_env;
1819                 io  = lcc->lcc_io;
1820         }
1821
1822         if (io == NULL) { /* fast read */
1823                 struct inode *inode = file_inode(file);
1824                 struct ll_file_data *fd = file->private_data;
1825                 struct ll_readahead_state *ras = &fd->fd_ras;
1826                 struct lu_env  *local_env = NULL;
1827                 struct vvp_page *vpg;
1828
1829                 result = -ENODATA;
1830
1831                 /* TODO: need to verify the layout version to make sure
1832                  * the page is not invalid due to layout change. */
1833                 page = cl_vmpage_page(vmpage, clob);
1834                 if (page == NULL) {
1835                         unlock_page(vmpage);
1836                         ll_ra_stats_inc_sbi(sbi, RA_STAT_FAILED_FAST_READ);
1837                         RETURN(result);
1838                 }
1839
1840                 vpg = cl2vvp_page(cl_object_page_slice(page->cp_obj, page));
1841                 if (vpg->vpg_defer_uptodate) {
1842                         enum ras_update_flags flags = LL_RAS_HIT;
1843
1844                         if (lcc && lcc->lcc_type == LCC_MMAP)
1845                                 flags |= LL_RAS_MMAP;
1846
1847                         /* For fast read, it updates read ahead state only
1848                          * if the page is hit in cache because non cache page
1849                          * case will be handled by slow read later. */
1850                         ras_update(sbi, inode, ras, vvp_index(vpg), flags, io);
1851                         /* avoid duplicate ras_update() call */
1852                         vpg->vpg_ra_updated = 1;
1853
1854                         if (ll_use_fast_io(file, ras, vvp_index(vpg)))
1855                                 result = 0;
1856                 }
1857
1858                 if (!env) {
1859                         local_env = cl_env_percpu_get();
1860                         env = local_env;
1861                 }
1862
1863                 /* export the page and skip io stack */
1864                 if (result == 0) {
1865                         vpg->vpg_ra_used = 1;
1866                         cl_page_export(env, page, 1);
1867                 } else {
1868                         ll_ra_stats_inc_sbi(sbi, RA_STAT_FAILED_FAST_READ);
1869                 }
1870                 /* release page refcount before unlocking the page to ensure
1871                  * the object won't be destroyed in the calling path of
1872                  * cl_page_put(). Please see comment in ll_releasepage(). */
1873                 cl_page_put(env, page);
1874                 unlock_page(vmpage);
1875                 if (local_env)
1876                         cl_env_percpu_put(local_env);
1877
1878                 RETURN(result);
1879         }
1880
1881         /**
1882          * Direct read can fall back to buffered read, but DIO is done
1883          * with lockless i/o, and buffered requires LDLM locking, so in
1884          * this case we must restart without lockless.
1885          */
1886         if (file->f_flags & O_DIRECT &&
1887             lcc && lcc->lcc_type == LCC_RW &&
1888             !io->ci_dio_lock) {
1889                 unlock_page(vmpage);
1890                 io->ci_dio_lock = 1;
1891                 io->ci_need_restart = 1;
1892                 RETURN(-ENOLCK);
1893         }
1894
1895         LASSERT(io->ci_state == CIS_IO_GOING);
1896         page = cl_page_find(env, clob, vmpage->index, vmpage, CPT_CACHEABLE);
1897         if (!IS_ERR(page)) {
1898                 LASSERT(page->cp_type == CPT_CACHEABLE);
1899                 if (likely(!PageUptodate(vmpage))) {
1900                         cl_page_assume(env, io, page);
1901
1902                         result = ll_io_read_page(env, io, page, file);
1903                 } else {
1904                         /* Page from a non-object file. */
1905                         unlock_page(vmpage);
1906                         result = 0;
1907                 }
1908                 cl_page_put(env, page);
1909         } else {
1910                 unlock_page(vmpage);
1911                 result = PTR_ERR(page);
1912         }
1913         RETURN(result);
1914 }