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         WARN_ON_ONCE(pages_min > pages);
  93         /**
  94          * Don't try readahead aggresively if we are limited
  95          * LRU pages, otherwise, it could cause deadlock.
  96          */
  97         pages = min(sbi->ll_cache->ccc_lru_max >> 2, pages);
  98         /**
  99          * if this happen, we reserve more pages than needed,
 100          * this will make us leak @ra_cur_pages, because
 101          * ll_ra_count_put() acutally freed @pages.
 102          */
 103         if (unlikely(pages_min > pages))
 104                 pages_min = pages;
 105
 106         /*
 107          * If read-ahead pages left are less than 1M, do not do read-ahead,
 108          * otherwise it will form small read RPC(< 1M), which hurt server
 109          * performance a lot.
 110          */
 111         ret = min(ra->ra_max_pages - atomic_read(&ra->ra_cur_pages),
 112                   pages);
 113         if (ret < 0 || ret < min_t(long, PTLRPC_MAX_BRW_PAGES, pages))
 114                 GOTO(out, ret = 0);
 115
 116         if (atomic_add_return(ret, &ra->ra_cur_pages) > ra->ra_max_pages) {
 117                 atomic_sub(ret, &ra->ra_cur_pages);
 118                 ret = 0;
 119         }
 120
 121 out:
 122         if (ret < pages_min) {
 123                 /* override ra limit for maximum performance */
 124                 atomic_add(pages_min - ret, &ra->ra_cur_pages);
 125                 ret = pages_min;
 126         }
 127         RETURN(ret);
 128 }
 129
 130 void ll_ra_count_put(struct ll_sb_info *sbi, unsigned long pages)
 131 {
 132         struct ll_ra_info *ra = &sbi->ll_ra_info;
 133         atomic_sub(pages, &ra->ra_cur_pages);
 134 }
 135
 136 static void ll_ra_stats_inc_sbi(struct ll_sb_info *sbi, enum ra_stat which)
 137 {
 138         LASSERTF(which < _NR_RA_STAT, "which: %u\n", which);
 139         lprocfs_counter_incr(sbi->ll_ra_stats, which);
 140 }
 141
 142 static inline bool ll_readahead_enabled(struct ll_sb_info *sbi)
 143 {
 144         return sbi->ll_ra_info.ra_max_pages_per_file > 0 &&
 145                 sbi->ll_ra_info.ra_max_pages > 0;
 146 }
 147
 148 void ll_ra_stats_inc(struct inode *inode, enum ra_stat which)
 149 {
 150         struct ll_sb_info *sbi = ll_i2sbi(inode);
 151         ll_ra_stats_inc_sbi(sbi, which);
 152 }
 153
 154 #define RAS_CDEBUG(ras) \
 155         CDEBUG(D_READA,                                                      \
 156                "lre %llu cr %lu cb %llu wsi %lu wp %lu nra %lu rpc %lu "     \
 157                "r %lu csr %lu so %llu sb %llu sl %llu lr %lu\n",             \
 158                ras->ras_last_read_end_bytes, ras->ras_consecutive_requests,  \
 159                ras->ras_consecutive_bytes, ras->ras_window_start_idx,        \
 160                ras->ras_window_pages, ras->ras_next_readahead_idx,           \
 161                ras->ras_rpc_pages, ras->ras_requests,                        \
 162                ras->ras_consecutive_stride_requests, ras->ras_stride_offset, \
 163                ras->ras_stride_bytes, ras->ras_stride_length,                \
 164                ras->ras_async_last_readpage_idx)
 165
 166 static bool pos_in_window(loff_t pos, loff_t point,
 167                           unsigned long before, unsigned long after)
 168 {
 169         loff_t start = point - before;
 170         loff_t end = point + after;
 171
 172         if (start > point)
 173                 start = 0;
 174         if (end < point)
 175                 end = ~0;
 176
 177         return start <= pos && pos <= end;
 178 }
 179
 180 enum ll_ra_page_hint {
 181         MAYNEED = 0, /* this page possibly accessed soon */
 182         WILLNEED /* this page is gurateed to be needed */
 183 };
 184
 185 /**
 186  * Initiates read-ahead of a page with given index.
 187  *
 188  * \retval +ve: page was already uptodate so it will be skipped
 189  *              from being added;
 190  * \retval -ve: page wasn't added to \a queue for error;
 191  * \retval   0: page was added into \a queue for read ahead.
 192  */
 193 static int ll_read_ahead_page(const struct lu_env *env, struct cl_io *io,
 194                               struct cl_page_list *queue, pgoff_t index,
 195                               enum ll_ra_page_hint hint)
 196 {
 197         struct cl_object *clob  = io->ci_obj;
 198         struct inode     *inode = vvp_object_inode(clob);
 199         struct page      *vmpage = NULL;
 200         struct cl_page   *page;
 201         struct vvp_page  *vpg;
 202         enum ra_stat      which = _NR_RA_STAT; /* keep gcc happy */
 203         int               rc    = 0;
 204         const char       *msg   = NULL;
 205
 206         ENTRY;
 207
 208         switch (hint) {
 209         case MAYNEED:
 210                 vmpage = grab_cache_page_nowait(inode->i_mapping, index);
 211                 if (vmpage == NULL) {
 212                         which = RA_STAT_FAILED_GRAB_PAGE;
 213                         msg   = "g_c_p_n failed";
 214                         GOTO(out, rc = -EBUSY);
 215                 }
 216                 break;
 217         case WILLNEED:
 218                 vmpage = find_or_create_page(inode->i_mapping, index,
 219                                              GFP_NOFS);
 220                 if (vmpage == NULL)
 221                         GOTO(out, rc = -ENOMEM);
 222                 break;
 223         default:
 224                 /* should not come here */
 225                 GOTO(out, rc = -EINVAL);
 226         }
 227
 228         /* Check if vmpage was truncated or reclaimed */
 229         if (vmpage->mapping != inode->i_mapping) {
 230                 which = RA_STAT_WRONG_GRAB_PAGE;
 231                 msg   = "g_c_p_n returned invalid page";
 232                 GOTO(out, rc = -EBUSY);
 233         }
 234
 235         page = cl_page_find(env, clob, vmpage->index, vmpage, CPT_CACHEABLE);
 236         if (IS_ERR(page)) {
 237                 which = RA_STAT_FAILED_GRAB_PAGE;
 238                 msg   = "cl_page_find failed";
 239                 GOTO(out, rc = PTR_ERR(page));
 240         }
 241
 242         lu_ref_add(&page->cp_reference, "ra", current);
 243         cl_page_assume(env, io, page);
 244         vpg = cl2vvp_page(cl_object_page_slice(clob, page));
 245         if (!vpg->vpg_defer_uptodate && !PageUptodate(vmpage)) {
 246                 if (hint == MAYNEED) {
 247                         vpg->vpg_defer_uptodate = 1;
 248                         vpg->vpg_ra_used = 0;
 249                 }
 250                 cl_page_list_add(queue, page, true);
 251         } else {
 252                 /* skip completed pages */
 253                 cl_page_unassume(env, io, page);
 254                 /* This page is already uptodate, returning a positive number
 255                  * to tell the callers about this */
 256                 rc = 1;
 257         }
 258
 259         lu_ref_del(&page->cp_reference, "ra", current);
 260         cl_page_put(env, page);
 261
 262 out:
 263         if (vmpage != NULL) {
 264                 if (rc != 0)
 265                         unlock_page(vmpage);
 266                 put_page(vmpage);
 267         }
 268         if (msg != NULL && hint == MAYNEED) {
 269                 ll_ra_stats_inc(inode, which);
 270                 CDEBUG(D_READA, "%s\n", msg);
 271
 272         }
 273
 274         RETURN(rc);
 275 }
 276
 277 #define RIA_DEBUG(ria)                                                  \
 278         CDEBUG(D_READA, "rs %lu re %lu ro %llu rl %llu rb %llu\n",      \
 279                ria->ria_start_idx, ria->ria_end_idx, ria->ria_stoff,    \
 280                ria->ria_length, ria->ria_bytes)
 281
 282 static inline int stride_io_mode(struct ll_readahead_state *ras)
 283 {
 284         return ras->ras_consecutive_stride_requests > 1;
 285 }
 286
 287 /* The function calculates how many bytes will be read in
 288  * [off, off + length], in such stride IO area,
 289  * stride_offset = st_off, stride_lengh = st_len,
 290  * stride_bytes = st_bytes
 291  *
 292  *   |------------------|*****|------------------|*****|------------|*****|....
 293  * st_off
 294  *   |--- st_bytes     ---|
 295  *   |-----     st_len   -----|
 296  *
 297  *              How many bytes it should read in such pattern
 298  *              |-------------------------------------------------------------|
 299  *              off
 300  *              |<------                  length                      ------->|
 301  *
 302  *          =   |<----->|  +  |-------------------------------------| +   |---|
 303  *             start_left                 st_bytes * i                 end_left
 304  */
 305 static loff_t stride_byte_count(loff_t st_off, loff_t st_len, loff_t st_bytes,
 306                                 loff_t off, loff_t length)
 307 {
 308         u64 start = off > st_off ? off - st_off : 0;
 309         u64 end = off + length > st_off ? off + length - st_off : 0;
 310         u64 start_left;
 311         u64 end_left;
 312         u64 bytes_count;
 313
 314         if (st_len == 0 || length == 0 || end == 0)
 315                 return length;
 316
 317         start = div64_u64_rem(start, st_len, &start_left);
 318         if (start_left < st_bytes)
 319                 start_left = st_bytes - start_left;
 320         else
 321                 start_left = 0;
 322
 323         end = div64_u64_rem(end, st_len, &end_left);
 324         if (end_left > st_bytes)
 325                 end_left = st_bytes;
 326
 327         CDEBUG(D_READA, "start %llu, end %llu start_left %llu end_left %llu\n",
 328                start, end, start_left, end_left);
 329
 330         if (start == end)
 331                 bytes_count = end_left - (st_bytes - start_left);
 332         else
 333                 bytes_count = start_left +
 334                         st_bytes * (end - start - 1) + end_left;
 335
 336         CDEBUG(D_READA,
 337                "st_off %llu, st_len %llu st_bytes %llu off %llu length %llu bytescount %llu\n",
 338                st_off, st_len, st_bytes, off, length, bytes_count);
 339
 340         return bytes_count;
 341 }
 342
 343 static unsigned long ria_page_count(struct ra_io_arg *ria)
 344 {
 345         loff_t length_bytes = ria->ria_end_idx >= ria->ria_start_idx ?
 346                 (loff_t)(ria->ria_end_idx -
 347                          ria->ria_start_idx + 1) << PAGE_SHIFT : 0;
 348         loff_t bytes_count;
 349
 350         if (ria->ria_length > ria->ria_bytes && ria->ria_bytes &&
 351             (ria->ria_length & ~PAGE_MASK || ria->ria_bytes & ~PAGE_MASK ||
 352              ria->ria_stoff & ~PAGE_MASK)) {
 353                 /* Over-estimate un-aligned page stride read */
 354                 unsigned long pg_count = ((ria->ria_bytes +
 355                                            PAGE_SIZE - 1) >> PAGE_SHIFT) + 1;
 356                 pg_count *= length_bytes / ria->ria_length + 1;
 357
 358                 return pg_count;
 359         }
 360         bytes_count = stride_byte_count(ria->ria_stoff, ria->ria_length,
 361                                         ria->ria_bytes,
 362                                         (loff_t)ria->ria_start_idx<<PAGE_SHIFT,
 363                                         length_bytes);
 364         return (bytes_count + PAGE_SIZE - 1) >> PAGE_SHIFT;
 365 }
 366
 367 static pgoff_t ras_align(struct ll_readahead_state *ras, pgoff_t index)
 368 {
 369         unsigned opt_size = min(ras->ras_window_pages, ras->ras_rpc_pages);
 370
 371         if (opt_size == 0)
 372                 opt_size = 1;
 373         return index - (index % opt_size);
 374 }
 375
 376 /* Check whether the index is in the defined ra-window */
 377 static bool ras_inside_ra_window(pgoff_t idx, struct ra_io_arg *ria)
 378 {
 379         loff_t pos = (loff_t)idx << PAGE_SHIFT;
 380
 381         /* If ria_length == ria_bytes, it means non-stride I/O mode,
 382          * idx should always inside read-ahead window in this case
 383          * For stride I/O mode, just check whether the idx is inside
 384          * the ria_bytes.
 385          */
 386         if (ria->ria_length == 0 || ria->ria_length == ria->ria_bytes)
 387                 return true;
 388
 389         if (pos >= ria->ria_stoff) {
 390                 u64 offset;
 391
 392                 div64_u64_rem(pos - ria->ria_stoff, ria->ria_length, &offset);
 393
 394                 if (offset < ria->ria_bytes ||
 395                     (ria->ria_length - offset) < PAGE_SIZE)
 396                         return true;
 397         } else if (pos + PAGE_SIZE > ria->ria_stoff) {
 398                 return true;
 399         }
 400
 401         return false;
 402 }
 403
 404 static unsigned long
 405 ll_read_ahead_pages(const struct lu_env *env, struct cl_io *io,
 406                     struct cl_page_list *queue, struct ll_readahead_state *ras,
 407                     struct ra_io_arg *ria, pgoff_t *ra_end, pgoff_t skip_index)
 408 {
 409         struct cl_read_ahead ra = { 0 };
 410         /* busy page count is per stride */
 411         int rc = 0, count = 0, busy_page_count = 0;
 412         pgoff_t page_idx;
 413
 414         LASSERT(ria != NULL);
 415         RIA_DEBUG(ria);
 416
 417         for (page_idx = ria->ria_start_idx;
 418              page_idx <= ria->ria_end_idx && ria->ria_reserved > 0;
 419              page_idx++) {
 420                 if (skip_index && page_idx == skip_index)
 421                         continue;
 422                 if (ras_inside_ra_window(page_idx, ria)) {
 423                         if (ra.cra_end_idx == 0 || ra.cra_end_idx < page_idx) {
 424                                 pgoff_t end_idx;
 425
 426                                 /*
 427                                  * Do not shrink ria_end_idx at any case until
 428                                  * the minimum end of current read is covered.
 429                                  *
 430                                  * Do not extend read lock accross stripe if
 431                                  * lock contention detected.
 432                                  */
 433                                 if (ra.cra_contention &&
 434                                     page_idx > ria->ria_end_idx_min) {
 435                                         ria->ria_end_idx = *ra_end;
 436                                         break;
 437                                 }
 438
 439                                 cl_read_ahead_release(env, &ra);
 440
 441                                 rc = cl_io_read_ahead(env, io, page_idx, &ra);
 442                                 if (rc < 0)
 443                                         break;
 444
 445                                  /*
 446                                   * Only shrink ria_end_idx if the matched
 447                                   * LDLM lock doesn't cover more.
 448                                   */
 449                                 if (page_idx > ra.cra_end_idx) {
 450                                         ria->ria_end_idx = ra.cra_end_idx;
 451                                         break;
 452                                 }
 453
 454                                 CDEBUG(D_READA, "idx: %lu, ra: %lu, rpc: %lu\n",
 455                                        page_idx, ra.cra_end_idx,
 456                                        ra.cra_rpc_pages);
 457                                 LASSERTF(ra.cra_end_idx >= page_idx,
 458                                          "object: %p, indcies %lu / %lu\n",
 459                                          io->ci_obj, ra.cra_end_idx, page_idx);
 460                                 /* update read ahead RPC size.
 461                                  * NB: it's racy but doesn't matter */
 462                                 if (ras->ras_rpc_pages != ra.cra_rpc_pages &&
 463                                     ra.cra_rpc_pages > 0)
 464                                         ras->ras_rpc_pages = ra.cra_rpc_pages;
 465                                 if (!skip_index) {
 466                                         /* trim it to align with optimal RPC size */
 467                                         end_idx = ras_align(ras, ria->ria_end_idx + 1);
 468                                         if (end_idx > 0 && !ria->ria_eof)
 469                                                 ria->ria_end_idx = end_idx - 1;
 470                                 }
 471                                 if (ria->ria_end_idx < ria->ria_end_idx_min)
 472                                         ria->ria_end_idx = ria->ria_end_idx_min;
 473                         }
 474                         if (page_idx > ria->ria_end_idx)
 475                                 break;
 476
 477                         /* If the page is inside the read-ahead window */
 478                         rc = ll_read_ahead_page(env, io, queue, page_idx,
 479                                                 MAYNEED);
 480                         if (rc < 0 && rc != -EBUSY)
 481                                 break;
 482                         if (rc == -EBUSY) {
 483                                 busy_page_count++;
 484                                 CDEBUG(D_READA,
 485                                        "skip busy page: %lu\n", page_idx);
 486                                 /* For page unaligned readahead the first
 487                                  * last pages of each region can be read by
 488                                  * another reader on the same node, and so
 489                                  * may be busy. So only stop for > 2 busy
 490                                  * pages. */
 491                                 if (busy_page_count > 2)
 492                                         break;
 493                         }
 494
 495                         *ra_end = page_idx;
 496                         /* Only subtract from reserve & count the page if we
 497                          * really did readahead on that page. */
 498                         if (rc == 0) {
 499                                 ria->ria_reserved--;
 500                                 count++;
 501                         }
 502                 } else if (stride_io_mode(ras)) {
 503                         /* If it is not in the read-ahead window, and it is
 504                          * read-ahead mode, then check whether it should skip
 505                          * the stride gap.
 506                          */
 507                         loff_t pos = (loff_t)page_idx << PAGE_SHIFT;
 508                         u64 offset;
 509
 510                         div64_u64_rem(pos - ria->ria_stoff, ria->ria_length,
 511                                       &offset);
 512                         if (offset >= ria->ria_bytes) {
 513                                 pos += (ria->ria_length - offset);
 514                                 if ((pos >> PAGE_SHIFT) >= page_idx + 1)
 515                                         page_idx = (pos >> PAGE_SHIFT) - 1;
 516                                 busy_page_count = 0;
 517                                 CDEBUG(D_READA,
 518                                        "Stride: jump %llu pages to %lu\n",
 519                                        ria->ria_length - offset, page_idx);
 520                                 continue;
 521                         }
 522                 }
 523         }
 524
 525         cl_read_ahead_release(env, &ra);
 526
 527         return count;
 528 }
 529
 530 static void ll_readahead_work_free(struct ll_readahead_work *work)
 531 {
 532         fput(work->lrw_file);
 533         OBD_FREE_PTR(work);
 534 }
 535
 536 static void ll_readahead_handle_work(struct work_struct *wq);
 537 static void ll_readahead_work_add(struct inode *inode,
 538                                   struct ll_readahead_work *work)
 539 {
 540         INIT_WORK(&work->lrw_readahead_work, ll_readahead_handle_work);
 541         queue_work(ll_i2sbi(inode)->ll_ra_info.ll_readahead_wq,
 542                    &work->lrw_readahead_work);
 543 }
 544
 545 static int ll_readahead_file_kms(const struct lu_env *env,
 546                                 struct cl_io *io, __u64 *kms)
 547 {
 548         struct cl_object *clob;
 549         struct inode *inode;
 550         struct cl_attr *attr = vvp_env_thread_attr(env);
 551         int ret;
 552
 553         clob = io->ci_obj;
 554         inode = vvp_object_inode(clob);
 555
 556         cl_object_attr_lock(clob);
 557         ret = cl_object_attr_get(env, clob, attr);
 558         cl_object_attr_unlock(clob);
 559
 560         if (ret != 0)
 561                 RETURN(ret);
 562
 563         *kms = attr->cat_kms;
 564         return 0;
 565 }
 566
 567 static void ll_readahead_handle_work(struct work_struct *wq)
 568 {
 569         struct ll_readahead_work *work;
 570         struct lu_env *env;
 571         __u16 refcheck;
 572         struct ra_io_arg *ria;
 573         struct inode *inode;
 574         struct ll_file_data *fd;
 575         struct ll_readahead_state *ras;
 576         struct cl_io *io;
 577         struct cl_2queue *queue;
 578         pgoff_t ra_end_idx = 0;
 579         unsigned long pages, pages_min = 0;
 580         struct file *file;
 581         __u64 kms;
 582         int rc;
 583         pgoff_t eof_index;
 584         struct ll_sb_info *sbi;
 585
 586         work = container_of(wq, struct ll_readahead_work,
 587                             lrw_readahead_work);
 588         fd = work->lrw_file->private_data;
 589         ras = &fd->fd_ras;
 590         file = work->lrw_file;
 591         inode = file_inode(file);
 592         sbi = ll_i2sbi(inode);
 593
 594         CDEBUG(D_READA|D_IOTRACE,
 595                "%s: async ra from %lu to %lu triggered by user pid %d\n",
 596                file_dentry(file)->d_name.name, work->lrw_start_idx,
 597                work->lrw_end_idx, work->lrw_user_pid);
 598
 599         env = cl_env_alloc(&refcheck, LCT_NOREF);
 600         if (IS_ERR(env))
 601                 GOTO(out_free_work, rc = PTR_ERR(env));
 602
 603         io = vvp_env_thread_io(env);
 604         ll_io_init(io, file, CIT_READ, NULL);
 605
 606         rc = ll_readahead_file_kms(env, io, &kms);
 607         if (rc != 0)
 608                 GOTO(out_put_env, rc);
 609
 610         if (kms == 0) {
 611                 ll_ra_stats_inc(inode, RA_STAT_ZERO_LEN);
 612                 GOTO(out_put_env, rc = 0);
 613         }
 614
 615         ria = &ll_env_info(env)->lti_ria;
 616         memset(ria, 0, sizeof(*ria));
 617
 618         ria->ria_start_idx = work->lrw_start_idx;
 619         /* Truncate RA window to end of file */
 620         eof_index = (pgoff_t)(kms - 1) >> PAGE_SHIFT;
 621         if (eof_index <= work->lrw_end_idx) {
 622                 work->lrw_end_idx = eof_index;
 623                 ria->ria_eof = true;
 624         }
 625         if (work->lrw_end_idx <= work->lrw_start_idx)
 626                 GOTO(out_put_env, rc = 0);
 627
 628         ria->ria_end_idx = work->lrw_end_idx;
 629         pages = ria->ria_end_idx - ria->ria_start_idx + 1;
 630         ria->ria_reserved = ll_ra_count_get(sbi, ria,
 631                                             ria_page_count(ria), pages_min);
 632
 633         CDEBUG(D_READA,
 634                "async reserved pages: %lu/%lu/%lu, ra_cur %d, ra_max %lu\n",
 635                ria->ria_reserved, pages, pages_min,
 636                atomic_read(&ll_i2sbi(inode)->ll_ra_info.ra_cur_pages),
 637                ll_i2sbi(inode)->ll_ra_info.ra_max_pages);
 638
 639         if (ria->ria_reserved < pages) {
 640                 ll_ra_stats_inc(inode, RA_STAT_MAX_IN_FLIGHT);
 641                 if (PAGES_TO_MiB(ria->ria_reserved) < 1) {
 642                         ll_ra_count_put(ll_i2sbi(inode), ria->ria_reserved);
 643                         GOTO(out_put_env, rc = 0);
 644                 }
 645         }
 646
 647         rc = cl_io_rw_init(env, io, CIT_READ, ria->ria_start_idx, pages);
 648         if (rc)
 649                 GOTO(out_put_env, rc);
 650
 651         /* overwrite jobid inited in vvp_io_init() */
 652         if (strncmp(ll_i2info(inode)->lli_jobid, work->lrw_jobid,
 653                     sizeof(work->lrw_jobid)))
 654                 memcpy(ll_i2info(inode)->lli_jobid, work->lrw_jobid,
 655                        sizeof(work->lrw_jobid));
 656
 657         vvp_env_io(env)->vui_fd = fd;
 658         io->ci_state = CIS_LOCKED;
 659         io->ci_async_readahead = true;
 660         rc = cl_io_start(env, io);
 661         if (rc)
 662                 GOTO(out_io_fini, rc);
 663
 664         queue = &io->ci_queue;
 665         cl_2queue_init(queue);
 666
 667         rc = ll_read_ahead_pages(env, io, &queue->c2_qin, ras, ria,
 668                                  &ra_end_idx, 0);
 669         if (ria->ria_reserved != 0)
 670                 ll_ra_count_put(ll_i2sbi(inode), ria->ria_reserved);
 671         if (queue->c2_qin.pl_nr > 0) {
 672                 int count = queue->c2_qin.pl_nr;
 673
 674                 rc = cl_io_submit_rw(env, io, CRT_READ, queue);
 675                 if (rc == 0)
 676                         task_io_account_read(PAGE_SIZE * count);
 677         }
 678         if (ria->ria_end_idx == ra_end_idx && ra_end_idx == (kms >> PAGE_SHIFT))
 679                 ll_ra_stats_inc(inode, RA_STAT_EOF);
 680
 681         if (ra_end_idx != ria->ria_end_idx)
 682                 ll_ra_stats_inc(inode, RA_STAT_FAILED_REACH_END);
 683
 684         /* TODO: discard all pages until page reinit route is implemented */
 685         cl_page_list_discard(env, io, &queue->c2_qin);
 686
 687         /* Unlock unsent read pages in case of error. */
 688         cl_page_list_disown(env, io, &queue->c2_qin);
 689
 690         cl_2queue_fini(env, queue);
 691 out_io_fini:
 692         cl_io_end(env, io);
 693         cl_io_fini(env, io);
 694 out_put_env:
 695         cl_env_put(env, &refcheck);
 696 out_free_work:
 697         if (ra_end_idx > 0)
 698                 ll_ra_stats_inc_sbi(ll_i2sbi(inode), RA_STAT_ASYNC);
 699         atomic_dec(&sbi->ll_ra_info.ra_async_inflight);
 700         ll_readahead_work_free(work);
 701 }
 702
 703 static int ll_readahead(const struct lu_env *env, struct cl_io *io,
 704                         struct cl_page_list *queue,
 705                         struct ll_readahead_state *ras, bool hit,
 706                         struct file *file, pgoff_t skip_index)
 707 {
 708         struct vvp_io *vio = vvp_env_io(env);
 709         struct ll_thread_info *lti = ll_env_info(env);
 710         unsigned long pages, pages_min = 0;
 711         pgoff_t ra_end_idx = 0, start_idx = 0, end_idx = 0;
 712         struct inode *inode;
 713         struct ra_io_arg *ria = &lti->lti_ria;
 714         struct cl_object *clob;
 715         int ret = 0;
 716         __u64 kms;
 717         struct ll_sb_info *sbi;
 718         struct ll_ra_info *ra;
 719
 720         ENTRY;
 721
 722         ENTRY;
 723
 724         clob = io->ci_obj;
 725         inode = vvp_object_inode(clob);
 726         sbi = ll_i2sbi(inode);
 727         ra = &sbi->ll_ra_info;
 728
 729         /**
 730          * In case we have a limited max_cached_mb, readahead
 731          * should be stopped if it have run out of all LRU slots.
 732          */
 733         if (atomic_read(&ra->ra_cur_pages) >= sbi->ll_cache->ccc_lru_max) {
 734                 ll_ra_stats_inc(inode, RA_STAT_MAX_IN_FLIGHT);
 735                 RETURN(0);
 736         }
 737
 738         memset(ria, 0, sizeof(*ria));
 739         ret = ll_readahead_file_kms(env, io, &kms);
 740         if (ret != 0)
 741                 RETURN(ret);
 742
 743         if (kms == 0) {
 744                 ll_ra_stats_inc(inode, RA_STAT_ZERO_LEN);
 745                 RETURN(0);
 746         }
 747
 748         spin_lock(&ras->ras_lock);
 749
 750         /**
 751          * Note: other thread might rollback the ras_next_readahead_idx,
 752          * if it can not get the full size of prepared pages, see the
 753          * end of this function. For stride read ahead, it needs to
 754          * make sure the offset is no less than ras_stride_offset,
 755          * so that stride read ahead can work correctly.
 756          */
 757         if (stride_io_mode(ras))
 758                 start_idx = max_t(pgoff_t, ras->ras_next_readahead_idx,
 759                                   ras->ras_stride_offset >> PAGE_SHIFT);
 760         else
 761                 start_idx = ras->ras_next_readahead_idx;
 762
 763         if (ras->ras_window_pages > 0)
 764                 end_idx = ras->ras_window_start_idx + ras->ras_window_pages - 1;
 765
 766         if (skip_index)
 767                 end_idx = start_idx + ras->ras_window_pages - 1;
 768
 769         /* Enlarge the RA window to encompass the full read */
 770         if (vio->vui_ra_valid &&
 771             end_idx < vio->vui_ra_start_idx + vio->vui_ra_pages - 1)
 772                 end_idx = vio->vui_ra_start_idx + vio->vui_ra_pages - 1;
 773
 774         if (end_idx != 0) {
 775                 pgoff_t eof_index;
 776
 777                 /* Truncate RA window to end of file */
 778                 eof_index = (pgoff_t)((kms - 1) >> PAGE_SHIFT);
 779                 if (eof_index <= end_idx) {
 780                         end_idx = eof_index;
 781                         ria->ria_eof = true;
 782                 }
 783         }
 784         ria->ria_start_idx = start_idx;
 785         ria->ria_end_idx = end_idx;
 786         /* If stride I/O mode is detected, get stride window*/
 787         if (stride_io_mode(ras)) {
 788                 ria->ria_stoff = ras->ras_stride_offset;
 789                 ria->ria_length = ras->ras_stride_length;
 790                 ria->ria_bytes = ras->ras_stride_bytes;
 791         }
 792         spin_unlock(&ras->ras_lock);
 793
 794         if (end_idx == 0) {
 795                 ll_ra_stats_inc(inode, RA_STAT_ZERO_WINDOW);
 796                 RETURN(0);
 797         }
 798         pages = ria_page_count(ria);
 799         if (pages == 0) {
 800                 ll_ra_stats_inc(inode, RA_STAT_ZERO_WINDOW);
 801                 RETURN(0);
 802         }
 803
 804         RAS_CDEBUG(ras);
 805         CDEBUG(D_READA, DFID": ria: %lu/%lu, bead: %lu/%lu, hit: %d\n",
 806                PFID(lu_object_fid(&clob->co_lu)),
 807                ria->ria_start_idx, ria->ria_end_idx,
 808                vio->vui_ra_valid ? vio->vui_ra_start_idx : 0,
 809                vio->vui_ra_valid ? vio->vui_ra_pages : 0,
 810                hit);
 811
 812         /* at least to extend the readahead window to cover current read */
 813         if (!hit && vio->vui_ra_valid &&
 814             vio->vui_ra_start_idx + vio->vui_ra_pages > ria->ria_start_idx) {
 815                 ria->ria_end_idx_min =
 816                         vio->vui_ra_start_idx + vio->vui_ra_pages - 1;
 817                 pages_min = vio->vui_ra_start_idx + vio->vui_ra_pages -
 818                                 ria->ria_start_idx;
 819                  /**
 820                   * For performance reason, exceeding @ra_max_pages
 821                   * are allowed, but this should be limited with RPC
 822                   * size in case a large block size read issued. Trim
 823                   * to RPC boundary.
 824                   */
 825                 pages_min = min(pages_min, ras->ras_rpc_pages -
 826                                 (ria->ria_start_idx % ras->ras_rpc_pages));
 827         }
 828
 829         /* don't over reserved for mmap range read */
 830         if (skip_index)
 831                 pages_min = 0;
 832         if (pages_min > pages)
 833                 pages = pages_min;
 834         ria->ria_reserved = ll_ra_count_get(ll_i2sbi(inode), ria, pages,
 835                                             pages_min);
 836         if (ria->ria_reserved < pages)
 837                 ll_ra_stats_inc(inode, RA_STAT_MAX_IN_FLIGHT);
 838
 839         CDEBUG(D_READA, "reserved pages: %lu/%lu/%lu, ra_cur %d, ra_max %lu\n",
 840                ria->ria_reserved, pages, pages_min,
 841                atomic_read(&ll_i2sbi(inode)->ll_ra_info.ra_cur_pages),
 842                ll_i2sbi(inode)->ll_ra_info.ra_max_pages);
 843
 844         ret = ll_read_ahead_pages(env, io, queue, ras, ria, &ra_end_idx,
 845                                   skip_index);
 846         if (ria->ria_reserved != 0)
 847                 ll_ra_count_put(ll_i2sbi(inode), ria->ria_reserved);
 848
 849         if (ra_end_idx == end_idx && ra_end_idx == (kms >> PAGE_SHIFT))
 850                 ll_ra_stats_inc(inode, RA_STAT_EOF);
 851
 852         CDEBUG(D_READA,
 853                "ra_end_idx = %lu end_idx = %lu stride end = %lu pages = %d\n",
 854                ra_end_idx, end_idx, ria->ria_end_idx, ret);
 855
 856         if (ra_end_idx != end_idx)
 857                 ll_ra_stats_inc(inode, RA_STAT_FAILED_REACH_END);
 858         if (ra_end_idx > 0) {
 859                 /* update the ras so that the next read-ahead tries from
 860                  * where we left off. */
 861                 spin_lock(&ras->ras_lock);
 862                 ras->ras_next_readahead_idx = ra_end_idx + 1;
 863                 spin_unlock(&ras->ras_lock);
 864                 RAS_CDEBUG(ras);
 865         }
 866
 867         RETURN(ret);
 868 }
 869
 870 static int ll_readpages(const struct lu_env *env, struct cl_io *io,
 871                         struct cl_page_list *queue,
 872                         pgoff_t start, pgoff_t end)
 873 {
 874         int ret = 0;
 875         __u64 kms;
 876         pgoff_t page_idx;
 877         int count = 0;
 878
 879         ENTRY;
 880
 881         ret = ll_readahead_file_kms(env, io, &kms);
 882         if (ret != 0)
 883                 RETURN(ret);
 884
 885         if (kms == 0)
 886                 RETURN(0);
 887
 888         if (end != 0) {
 889                 unsigned long end_index;
 890
 891                 end_index = (unsigned long)((kms - 1) >> PAGE_SHIFT);
 892                 if (end_index <= end)
 893                         end = end_index;
 894         }
 895
 896         for (page_idx = start; page_idx <= end; page_idx++) {
 897                 ret= ll_read_ahead_page(env, io, queue, page_idx,
 898                                         WILLNEED);
 899                 if (ret < 0)
 900                         break;
 901                 else if (ret == 0) /* ret 1 is already uptodate */
 902                         count++;
 903         }
 904
 905         RETURN(count > 0 ? count : ret);
 906 }
 907
 908 static void ras_set_start(struct ll_readahead_state *ras, pgoff_t index)
 909 {
 910         ras->ras_window_start_idx = ras_align(ras, index);
 911 }
 912
 913 /* called with the ras_lock held or from places where it doesn't matter */
 914 static void ras_reset(struct ll_readahead_state *ras, pgoff_t index)
 915 {
 916         ras->ras_consecutive_requests = 0;
 917         ras->ras_consecutive_bytes = 0;
 918         ras->ras_window_pages = 0;
 919         ras_set_start(ras, index);
 920         ras->ras_next_readahead_idx = max(ras->ras_window_start_idx, index + 1);
 921
 922         RAS_CDEBUG(ras);
 923 }
 924
 925 /* called with the ras_lock held or from places where it doesn't matter */
 926 static void ras_stride_reset(struct ll_readahead_state *ras)
 927 {
 928         ras->ras_consecutive_stride_requests = 0;
 929         ras->ras_stride_length = 0;
 930         ras->ras_stride_bytes = 0;
 931         RAS_CDEBUG(ras);
 932 }
 933
 934 void ll_readahead_init(struct inode *inode, struct ll_readahead_state *ras)
 935 {
 936         spin_lock_init(&ras->ras_lock);
 937         ras->ras_rpc_pages = PTLRPC_MAX_BRW_PAGES;
 938         ras_reset(ras, 0);
 939         ras->ras_last_read_end_bytes = 0;
 940         ras->ras_requests = 0;
 941         ras->ras_range_min_start_idx = 0;
 942         ras->ras_range_max_end_idx = 0;
 943         ras->ras_range_requests = 0;
 944         ras->ras_last_range_pages = 0;
 945 }
 946
 947 /*
 948  * Check whether the read request is in the stride window.
 949  * If it is in the stride window, return true, otherwise return false.
 950  */
 951 static bool read_in_stride_window(struct ll_readahead_state *ras,
 952                                   loff_t pos, loff_t count)
 953 {
 954         loff_t stride_gap;
 955
 956         if (ras->ras_stride_length == 0 || ras->ras_stride_bytes == 0 ||
 957             ras->ras_stride_bytes == ras->ras_stride_length)
 958                 return false;
 959
 960         stride_gap = pos - ras->ras_last_read_end_bytes - 1;
 961
 962         /* If it is contiguous read */
 963         if (stride_gap == 0)
 964                 return ras->ras_consecutive_bytes + count <=
 965                         ras->ras_stride_bytes;
 966
 967         /* Otherwise check the stride by itself */
 968         return (ras->ras_stride_length - ras->ras_stride_bytes) == stride_gap &&
 969                 ras->ras_consecutive_bytes == ras->ras_stride_bytes &&
 970                 count <= ras->ras_stride_bytes;
 971 }
 972
 973 static void ras_init_stride_detector(struct ll_readahead_state *ras,
 974                                      loff_t pos, loff_t count)
 975 {
 976         loff_t stride_gap = pos - ras->ras_last_read_end_bytes - 1;
 977
 978         LASSERT(ras->ras_consecutive_stride_requests == 0);
 979
 980         if (pos <= ras->ras_last_read_end_bytes) {
 981                 /*Reset stride window for forward read*/
 982                 ras_stride_reset(ras);
 983                 return;
 984         }
 985
 986         ras->ras_stride_bytes = ras->ras_consecutive_bytes;
 987         ras->ras_stride_length = stride_gap + ras->ras_consecutive_bytes;
 988         ras->ras_consecutive_stride_requests++;
 989         ras->ras_stride_offset = pos;
 990
 991         RAS_CDEBUG(ras);
 992 }
 993
 994 static unsigned long
 995 stride_page_count(struct ll_readahead_state *ras, loff_t len)
 996 {
 997         loff_t bytes_count =
 998                 stride_byte_count(ras->ras_stride_offset,
 999                                   ras->ras_stride_length, ras->ras_stride_bytes,
1000                                   ras->ras_window_start_idx << PAGE_SHIFT, len);
1001
1002         return (bytes_count + PAGE_SIZE - 1) >> PAGE_SHIFT;
1003 }
1004
1005 /* Stride Read-ahead window will be increased inc_len according to
1006  * stride I/O pattern */
1007 static void ras_stride_increase_window(struct ll_readahead_state *ras,
1008                                        struct ll_ra_info *ra, loff_t inc_bytes)
1009 {
1010         loff_t window_bytes, stride_bytes;
1011         u64 left_bytes;
1012         u64 step;
1013         loff_t end;
1014
1015         /* temporarily store in page units to reduce LASSERT() cost below */
1016         end = ras->ras_window_start_idx + ras->ras_window_pages;
1017
1018         LASSERT(ras->ras_stride_length > 0);
1019         LASSERTF(end >= (ras->ras_stride_offset >> PAGE_SHIFT),
1020                  "window_start_idx %lu, window_pages %lu stride_offset %llu\n",
1021                  ras->ras_window_start_idx, ras->ras_window_pages,
1022                  ras->ras_stride_offset);
1023
1024         end <<= PAGE_SHIFT;
1025         if (end <= ras->ras_stride_offset)
1026                 stride_bytes = 0;
1027         else
1028                 stride_bytes = end - ras->ras_stride_offset;
1029
1030         div64_u64_rem(stride_bytes, ras->ras_stride_length, &left_bytes);
1031         window_bytes = (ras->ras_window_pages << PAGE_SHIFT);
1032         if (left_bytes < ras->ras_stride_bytes) {
1033                 if (ras->ras_stride_bytes - left_bytes >= inc_bytes) {
1034                         window_bytes += inc_bytes;
1035                         goto out;
1036                 } else {
1037                         window_bytes += (ras->ras_stride_bytes - left_bytes);
1038                         inc_bytes -= (ras->ras_stride_bytes - left_bytes);
1039                 }
1040         } else {
1041                 window_bytes += (ras->ras_stride_length - left_bytes);
1042         }
1043
1044         LASSERT(ras->ras_stride_bytes != 0);
1045
1046         step = div64_u64_rem(inc_bytes, ras->ras_stride_bytes, &left_bytes);
1047
1048         window_bytes += step * ras->ras_stride_length + left_bytes;
1049         LASSERT(window_bytes > 0);
1050
1051 out:
1052         if (stride_page_count(ras, window_bytes) <=
1053             ra->ra_max_pages_per_file || ras->ras_window_pages == 0)
1054                 ras->ras_window_pages = (window_bytes >> PAGE_SHIFT);
1055
1056         LASSERT(ras->ras_window_pages > 0);
1057
1058         RAS_CDEBUG(ras);
1059 }
1060
1061 static void ras_increase_window(struct inode *inode,
1062                                 struct ll_readahead_state *ras,
1063                                 struct ll_ra_info *ra)
1064 {
1065         /* The stretch of ra-window should be aligned with max rpc_size
1066          * but current clio architecture does not support retrieve such
1067          * information from lower layer. FIXME later
1068          */
1069         if (stride_io_mode(ras)) {
1070                 ras_stride_increase_window(ras, ra,
1071                                       (loff_t)ras->ras_rpc_pages << PAGE_SHIFT);
1072         } else {
1073                 pgoff_t window_pages;
1074
1075                 window_pages = min(ras->ras_window_pages + ras->ras_rpc_pages,
1076                                    ra->ra_max_pages_per_file);
1077                 if (window_pages < ras->ras_rpc_pages)
1078                         ras->ras_window_pages = window_pages;
1079                 else
1080                         ras->ras_window_pages = ras_align(ras, window_pages);
1081         }
1082 }
1083
1084 /**
1085  * Seek within 8 pages are considered as sequential read for now.
1086  */
1087 static inline bool is_loose_seq_read(struct ll_readahead_state *ras, loff_t pos)
1088 {
1089         return pos_in_window(pos, ras->ras_last_read_end_bytes,
1090                              8UL << PAGE_SHIFT, 8UL << PAGE_SHIFT);
1091 }
1092
1093 static inline bool is_loose_mmap_read(struct ll_sb_info *sbi,
1094                                       struct ll_readahead_state *ras,
1095                                       unsigned long pos)
1096 {
1097         unsigned long range_pages = sbi->ll_ra_info.ra_range_pages;
1098
1099         return pos_in_window(pos, ras->ras_last_read_end_bytes,
1100                              range_pages << PAGE_SHIFT,
1101                              range_pages << PAGE_SHIFT);
1102 }
1103
1104 /**
1105  * We have observed slow mmap read performances for some
1106  * applications. The problem is if access pattern is neither
1107  * sequential nor stride, but could be still adjacent in a
1108  * small range and then seek a random position.
1109  *
1110  * So the pattern could be something like this:
1111  *
1112  * [1M data] [hole] [0.5M data] [hole] [0.7M data] [1M data]
1113  *
1114  *
1115  * Every time an application reads mmap data, it may not only
1116  * read a single 4KB page, but aslo a cluster of nearby pages in
1117  * a range(e.g. 1MB) of the first page after a cache miss.
1118  *
1119  * The readahead engine is modified to track the range size of
1120  * a cluster of mmap reads, so that after a seek and/or cache miss,
1121  * the range size is used to efficiently prefetch multiple pages
1122  * in a single RPC rather than many small RPCs.
1123  */
1124 static void ras_detect_cluster_range(struct ll_readahead_state *ras,
1125                                      struct ll_sb_info *sbi,
1126                                      unsigned long pos, unsigned long count)
1127 {
1128         pgoff_t last_pages, pages;
1129         pgoff_t end_idx = (pos + count - 1) >> PAGE_SHIFT;
1130
1131         last_pages = ras->ras_range_max_end_idx -
1132                         ras->ras_range_min_start_idx + 1;
1133         /* First time come here */
1134         if (!ras->ras_range_max_end_idx)
1135                 goto out;
1136
1137         /* Random or Stride read */
1138         if (!is_loose_mmap_read(sbi, ras, pos))
1139                 goto out;
1140
1141         ras->ras_range_requests++;
1142         if (ras->ras_range_max_end_idx < end_idx)
1143                 ras->ras_range_max_end_idx = end_idx;
1144
1145         if (ras->ras_range_min_start_idx > (pos >> PAGE_SHIFT))
1146                 ras->ras_range_min_start_idx = pos >> PAGE_SHIFT;
1147
1148         /* Out of range, consider it as random or stride */
1149         pages = ras->ras_range_max_end_idx -
1150                         ras->ras_range_min_start_idx + 1;
1151         if (pages <= sbi->ll_ra_info.ra_range_pages)
1152                 return;
1153 out:
1154         ras->ras_last_range_pages = last_pages;
1155         ras->ras_range_requests = 0;
1156         ras->ras_range_min_start_idx = pos >> PAGE_SHIFT;
1157         ras->ras_range_max_end_idx = end_idx;
1158 }
1159
1160 static void ras_detect_read_pattern(struct ll_readahead_state *ras,
1161                                     struct ll_sb_info *sbi,
1162                                     loff_t pos, size_t count, bool mmap)
1163 {
1164         bool stride_detect = false;
1165         pgoff_t index = pos >> PAGE_SHIFT;
1166
1167         /*
1168          * Reset the read-ahead window in two cases. First when the app seeks
1169          * or reads to some other part of the file. Secondly if we get a
1170          * read-ahead miss that we think we've previously issued. This can
1171          * be a symptom of there being so many read-ahead pages that the VM
1172          * is reclaiming it before we get to it.
1173          */
1174         if (!is_loose_seq_read(ras, pos)) {
1175                 /* Check whether it is in stride I/O mode */
1176                 if (!read_in_stride_window(ras, pos, count)) {
1177                         if (ras->ras_consecutive_stride_requests == 0)
1178                                 ras_init_stride_detector(ras, pos, count);
1179                         else
1180                                 ras_stride_reset(ras);
1181                         ras->ras_consecutive_bytes = 0;
1182                         ras_reset(ras, index);
1183                 } else {
1184                         ras->ras_consecutive_bytes = 0;
1185                         ras->ras_consecutive_requests = 0;
1186                         if (++ras->ras_consecutive_stride_requests > 1)
1187                                 stride_detect = true;
1188                         RAS_CDEBUG(ras);
1189                 }
1190                 ll_ra_stats_inc_sbi(sbi, RA_STAT_DISTANT_READPAGE);
1191         } else if (stride_io_mode(ras)) {
1192                 /*
1193                  * If this is contiguous read but in stride I/O mode
1194                  * currently, check whether stride step still is valid,
1195                  * if invalid, it will reset the stride ra window to
1196                  * be zero.
1197                  */
1198                 if (!read_in_stride_window(ras, pos, count)) {
1199                         ras_stride_reset(ras);
1200                         ras->ras_window_pages = 0;
1201                         ras->ras_next_readahead_idx = index;
1202                 }
1203         }
1204
1205         ras->ras_consecutive_bytes += count;
1206         if (mmap) {
1207                 pgoff_t idx = ras->ras_consecutive_bytes >> PAGE_SHIFT;
1208                 unsigned long ra_range_pages =
1209                                 max_t(unsigned long, RA_MIN_MMAP_RANGE_PAGES,
1210                                       sbi->ll_ra_info.ra_range_pages);
1211
1212                 if ((idx >= ra_range_pages &&
1213                      idx % ra_range_pages == 0) || stride_detect)
1214                         ras->ras_need_increase_window = true;
1215         } else if ((ras->ras_consecutive_requests > 1 || stride_detect)) {
1216                 ras->ras_need_increase_window = true;
1217         }
1218
1219         ras->ras_last_read_end_bytes = pos + count - 1;
1220 }
1221
1222 void ll_ras_enter(struct file *f, loff_t pos, size_t count)
1223 {
1224         struct ll_file_data *fd = f->private_data;
1225         struct ll_readahead_state *ras = &fd->fd_ras;
1226         struct inode *inode = file_inode(f);
1227         unsigned long index = pos >> PAGE_SHIFT;
1228         struct ll_sb_info *sbi = ll_i2sbi(inode);
1229
1230         spin_lock(&ras->ras_lock);
1231         ras->ras_requests++;
1232         ras->ras_consecutive_requests++;
1233         ras->ras_need_increase_window = false;
1234         ras->ras_no_miss_check = false;
1235         /*
1236          * On the second access to a file smaller than the tunable
1237          * ra_max_read_ahead_whole_pages trigger RA on all pages in the
1238          * file up to ra_max_pages_per_file.  This is simply a best effort
1239          * and only occurs once per open file. Normal RA behavior is reverted
1240          * to for subsequent IO.
1241          */
1242         if (ras->ras_requests >= 2) {
1243                 __u64 kms_pages;
1244                 struct ll_ra_info *ra = &sbi->ll_ra_info;
1245
1246                 kms_pages = (i_size_read(inode) + PAGE_SIZE - 1) >>
1247                             PAGE_SHIFT;
1248
1249                 CDEBUG(D_READA, "kmsp %llu mwp %lu mp %lu\n", kms_pages,
1250                        ra->ra_max_read_ahead_whole_pages,
1251                        ra->ra_max_pages_per_file);
1252
1253                 if (kms_pages &&
1254                     kms_pages <= ra->ra_max_read_ahead_whole_pages) {
1255                         ras->ras_window_start_idx = 0;
1256                         ras->ras_next_readahead_idx = index + 1;
1257                         ras->ras_window_pages = min(ra->ra_max_pages_per_file,
1258                                             ra->ra_max_read_ahead_whole_pages);
1259                         ras->ras_no_miss_check = true;
1260                         GOTO(out_unlock, 0);
1261                 }
1262         }
1263         ras_detect_read_pattern(ras, sbi, pos, count, false);
1264 out_unlock:
1265         spin_unlock(&ras->ras_lock);
1266 }
1267
1268 static bool index_in_stride_window(struct ll_readahead_state *ras,
1269                                    pgoff_t index)
1270 {
1271         loff_t pos = (loff_t)index << PAGE_SHIFT;
1272
1273         if (ras->ras_stride_length == 0 || ras->ras_stride_bytes == 0 ||
1274             ras->ras_stride_bytes == ras->ras_stride_length)
1275                 return false;
1276
1277         if (pos >= ras->ras_stride_offset) {
1278                 u64 offset;
1279
1280                 div64_u64_rem(pos - ras->ras_stride_offset,
1281                               ras->ras_stride_length, &offset);
1282                 if (offset < ras->ras_stride_bytes ||
1283                     ras->ras_stride_length - offset < PAGE_SIZE)
1284                         return true;
1285         } else if (ras->ras_stride_offset - pos < PAGE_SIZE) {
1286                 return true;
1287         }
1288
1289         return false;
1290 }
1291
1292 /*
1293  * ll_ras_enter() is used to detect read pattern according to pos and count.
1294  *
1295  * ras_update() is used to detect cache miss and
1296  * reset window or increase window accordingly
1297  */
1298 static void ras_update(struct ll_sb_info *sbi, struct inode *inode,
1299                        struct ll_readahead_state *ras, pgoff_t index,
1300                        enum ras_update_flags flags, struct cl_io *io)
1301 {
1302         struct ll_ra_info *ra = &sbi->ll_ra_info;
1303         bool hit = flags & LL_RAS_HIT;
1304
1305         ENTRY;
1306         spin_lock(&ras->ras_lock);
1307
1308         if (!hit)
1309                 CDEBUG(D_READA|D_IOTRACE, DFID " pages at %lu miss.\n",
1310                        PFID(ll_inode2fid(inode)), index);
1311         ll_ra_stats_inc_sbi(sbi, hit ? RA_STAT_HIT : RA_STAT_MISS);
1312
1313         /*
1314          * The readahead window has been expanded to cover whole
1315          * file size, we don't care whether ra miss happen or not.
1316          * Because we will read whole file to page cache even if
1317          * some pages missed.
1318          */
1319         if (ras->ras_no_miss_check)
1320                 GOTO(out_unlock, 0);
1321
1322         if (io && io->ci_rand_read)
1323                 GOTO(out_unlock, 0);
1324
1325         if (io && io->ci_seq_read) {
1326                 if (!hit) {
1327                         /* to avoid many small read RPC here */
1328                         ras->ras_window_pages = sbi->ll_ra_info.ra_range_pages;
1329                         ll_ra_stats_inc_sbi(sbi, RA_STAT_MMAP_RANGE_READ);
1330                 }
1331                 goto skip;
1332         }
1333
1334         if (flags & LL_RAS_MMAP) {
1335                 unsigned long ra_pages;
1336
1337                 ras_detect_cluster_range(ras, sbi, index << PAGE_SHIFT,
1338                                          PAGE_SIZE);
1339                 ras_detect_read_pattern(ras, sbi, (loff_t)index << PAGE_SHIFT,
1340                                         PAGE_SIZE, true);
1341
1342                 /* we did not detect anything but we could prefetch */
1343                 if (!ras->ras_need_increase_window &&
1344                     ras->ras_window_pages <= sbi->ll_ra_info.ra_range_pages &&
1345                     ras->ras_range_requests >= 2) {
1346                         if (!hit) {
1347                                 ra_pages = max_t(unsigned long,
1348                                         RA_MIN_MMAP_RANGE_PAGES,
1349                                         ras->ras_last_range_pages);
1350                                 if (index < ra_pages / 2)
1351                                         index = 0;
1352                                 else
1353                                         index -= ra_pages / 2;
1354                                 ras->ras_window_pages = ra_pages;
1355                                 ll_ra_stats_inc_sbi(sbi,
1356                                         RA_STAT_MMAP_RANGE_READ);
1357                         } else {
1358                                 ras->ras_window_pages = 0;
1359                         }
1360                         goto skip;
1361                 }
1362         }
1363
1364         if (!hit && ras->ras_window_pages &&
1365             index < ras->ras_next_readahead_idx &&
1366             pos_in_window(index, ras->ras_window_start_idx, 0,
1367                           ras->ras_window_pages)) {
1368                 ll_ra_stats_inc_sbi(sbi, RA_STAT_MISS_IN_WINDOW);
1369                 ras->ras_need_increase_window = false;
1370
1371                 if (index_in_stride_window(ras, index) &&
1372                     stride_io_mode(ras)) {
1373                         /*
1374                          * if (index != ras->ras_last_readpage + 1)
1375                          *      ras->ras_consecutive_pages = 0;
1376                          */
1377                         ras_reset(ras, index);
1378
1379                         /*
1380                          * If stride-RA hit cache miss, the stride
1381                          * detector will not be reset to avoid the
1382                          * overhead of redetecting read-ahead mode,
1383                          * but on the condition that the stride window
1384                          * is still intersect with normal sequential
1385                          * read-ahead window.
1386                          */
1387                         if (ras->ras_window_start_idx < ras->ras_stride_offset)
1388                                 ras_stride_reset(ras);
1389                         RAS_CDEBUG(ras);
1390                 } else {
1391                         /*
1392                          * Reset both stride window and normal RA
1393                          * window.
1394                          */
1395                         ras_reset(ras, index);
1396                         /* ras->ras_consecutive_pages++; */
1397                         ras->ras_consecutive_bytes = 0;
1398                         ras_stride_reset(ras);
1399                         GOTO(out_unlock, 0);
1400                 }
1401         }
1402
1403 skip:
1404         ras_set_start(ras, index);
1405
1406         if (stride_io_mode(ras)) {
1407                 /* Since stride readahead is sentivite to the offset
1408                  * of read-ahead, so we use original offset here,
1409                  * instead of ras_window_start_idx, which is RPC aligned.
1410                  */
1411                 ras->ras_next_readahead_idx = max(index + 1,
1412                                                   ras->ras_next_readahead_idx);
1413                 ras->ras_window_start_idx =
1414                                 max_t(pgoff_t, ras->ras_window_start_idx,
1415                                       ras->ras_stride_offset >> PAGE_SHIFT);
1416         } else {
1417                 if (ras->ras_next_readahead_idx < ras->ras_window_start_idx)
1418                         ras->ras_next_readahead_idx = ras->ras_window_start_idx;
1419                 if (!hit)
1420                         ras->ras_next_readahead_idx = index + 1;
1421         }
1422
1423         if (ras->ras_need_increase_window) {
1424                 ras_increase_window(inode, ras, ra);
1425                 ras->ras_need_increase_window = false;
1426         }
1427
1428         EXIT;
1429 out_unlock:
1430         spin_unlock(&ras->ras_lock);
1431 }
1432
1433 int ll_writepage(struct page *vmpage, struct writeback_control *wbc)
1434 {
1435         struct inode           *inode = vmpage->mapping->host;
1436         struct ll_inode_info   *lli   = ll_i2info(inode);
1437         struct lu_env          *env;
1438         struct cl_io           *io;
1439         struct cl_page         *page;
1440         struct cl_object       *clob;
1441         bool redirtied = false;
1442         bool unlocked = false;
1443         int result;
1444         __u16 refcheck;
1445         ENTRY;
1446
1447         LASSERT(PageLocked(vmpage));
1448         LASSERT(!PageWriteback(vmpage));
1449
1450         LASSERT(ll_i2dtexp(inode) != NULL);
1451
1452         env = cl_env_get(&refcheck);
1453         if (IS_ERR(env))
1454                 GOTO(out, result = PTR_ERR(env));
1455
1456         clob  = ll_i2info(inode)->lli_clob;
1457         LASSERT(clob != NULL);
1458
1459         io = vvp_env_thread_io(env);
1460         io->ci_obj = clob;
1461         io->ci_ignore_layout = 1;
1462         result = cl_io_init(env, io, CIT_MISC, clob);
1463         if (result == 0) {
1464                 page = cl_page_find(env, clob, vmpage->index,
1465                                     vmpage, CPT_CACHEABLE);
1466                 if (!IS_ERR(page)) {
1467                         lu_ref_add(&page->cp_reference, "writepage",
1468                                    current);
1469                         cl_page_assume(env, io, page);
1470                         result = cl_page_flush(env, io, page);
1471                         if (result != 0) {
1472                                 /*
1473                                  * Re-dirty page on error so it retries write,
1474                                  * but not in case when IO has actually
1475                                  * occurred and completed with an error.
1476                                  */
1477                                 if (!PageError(vmpage)) {
1478                                         redirty_page_for_writepage(wbc, vmpage);
1479                                         result = 0;
1480                                         redirtied = true;
1481                                 }
1482                         }
1483                         cl_page_disown(env, io, page);
1484                         unlocked = true;
1485                         lu_ref_del(&page->cp_reference,
1486                                    "writepage", current);
1487                         cl_page_put(env, page);
1488                 } else {
1489                         result = PTR_ERR(page);
1490                 }
1491         }
1492         cl_io_fini(env, io);
1493
1494         if (redirtied && wbc->sync_mode == WB_SYNC_ALL) {
1495                 loff_t offset = cl_offset(clob, vmpage->index);
1496
1497                 /* Flush page failed because the extent is being written out.
1498                  * Wait for the write of extent to be finished to avoid
1499                  * breaking kernel which assumes ->writepage should mark
1500                  * PageWriteback or clean the page. */
1501                 result = cl_sync_file_range(inode, offset,
1502                                             offset + PAGE_SIZE - 1,
1503                                             CL_FSYNC_LOCAL, 1);
1504                 if (result > 0) {
1505                         /* actually we may have written more than one page.
1506                          * decreasing this page because the caller will count
1507                          * it. */
1508                         wbc->nr_to_write -= result - 1;
1509                         result = 0;
1510                 }
1511         }
1512
1513         cl_env_put(env, &refcheck);
1514         GOTO(out, result);
1515
1516 out:
1517         if (result < 0) {
1518                 if (!lli->lli_async_rc)
1519                         lli->lli_async_rc = result;
1520                 SetPageError(vmpage);
1521                 if (!unlocked)
1522                         unlock_page(vmpage);
1523         }
1524         return result;
1525 }
1526
1527 int ll_writepages(struct address_space *mapping, struct writeback_control *wbc)
1528 {
1529         struct inode *inode = mapping->host;
1530         loff_t start;
1531         loff_t end;
1532         enum cl_fsync_mode mode;
1533         int range_whole = 0;
1534         int result;
1535         ENTRY;
1536
1537         if (wbc->range_cyclic) {
1538                 start = (loff_t)mapping->writeback_index << PAGE_SHIFT;
1539                 end = OBD_OBJECT_EOF;
1540         } else {
1541                 start = wbc->range_start;
1542                 end = wbc->range_end;
1543                 if (end == LLONG_MAX) {
1544                         end = OBD_OBJECT_EOF;
1545                         range_whole = start == 0;
1546                 }
1547         }
1548
1549         mode = CL_FSYNC_NONE;
1550         if (wbc->sync_mode == WB_SYNC_ALL)
1551                 mode = CL_FSYNC_LOCAL;
1552
1553         if (ll_i2info(inode)->lli_clob == NULL)
1554                 RETURN(0);
1555
1556         /* for directio, it would call writepages() to evict cached pages
1557          * inside the IO context of write, which will cause deadlock at
1558          * layout_conf since it waits for active IOs to complete. */
1559         result = cl_sync_file_range(inode, start, end, mode, 1);
1560         if (result > 0) {
1561                 wbc->nr_to_write -= result;
1562                 result = 0;
1563          }
1564
1565         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0)) {
1566                 if (end == OBD_OBJECT_EOF)
1567                         mapping->writeback_index = 0;
1568                 else
1569                         mapping->writeback_index = (end >> PAGE_SHIFT) + 1;
1570         }
1571         RETURN(result);
1572 }
1573
1574 struct ll_cl_context *ll_cl_find(struct inode *inode)
1575 {
1576         struct ll_inode_info *lli = ll_i2info(inode);
1577         struct ll_cl_context *lcc;
1578         struct ll_cl_context *found = NULL;
1579
1580         read_lock(&lli->lli_lock);
1581         list_for_each_entry(lcc, &lli->lli_lccs, lcc_list) {
1582                 if (lcc->lcc_cookie == current) {
1583                         found = lcc;
1584                         break;
1585                 }
1586         }
1587         read_unlock(&lli->lli_lock);
1588
1589         return found;
1590 }
1591
1592 void ll_cl_add(struct inode *inode, const struct lu_env *env, struct cl_io *io,
1593                enum lcc_type type)
1594 {
1595         struct ll_inode_info *lli = ll_i2info(inode);
1596         struct ll_cl_context *lcc = &ll_env_info(env)->lti_io_ctx;
1597
1598         memset(lcc, 0, sizeof(*lcc));
1599         INIT_LIST_HEAD(&lcc->lcc_list);
1600         lcc->lcc_cookie = current;
1601         lcc->lcc_env = env;
1602         lcc->lcc_io = io;
1603         lcc->lcc_type = type;
1604
1605         write_lock(&lli->lli_lock);
1606         list_add(&lcc->lcc_list, &lli->lli_lccs);
1607         write_unlock(&lli->lli_lock);
1608 }
1609
1610 void ll_cl_remove(struct inode *inode, const struct lu_env *env)
1611 {
1612         struct ll_inode_info *lli = ll_i2info(inode);
1613         struct ll_cl_context *lcc = &ll_env_info(env)->lti_io_ctx;
1614
1615         write_lock(&lli->lli_lock);
1616         list_del_init(&lcc->lcc_list);
1617         write_unlock(&lli->lli_lock);
1618 }
1619
1620 int ll_io_read_page(const struct lu_env *env, struct cl_io *io,
1621                            struct cl_page *page, struct file *file)
1622 {
1623         struct inode              *inode  = vvp_object_inode(page->cp_obj);
1624         struct ll_sb_info         *sbi    = ll_i2sbi(inode);
1625         struct ll_file_data       *fd     = NULL;
1626         struct ll_readahead_state *ras    = NULL;
1627         struct cl_2queue          *queue  = &io->ci_queue;
1628         struct cl_sync_io         *anchor = NULL;
1629         struct vvp_page           *vpg;
1630         int                        rc = 0, rc2 = 0;
1631         bool                       uptodate;
1632         pgoff_t io_start_index;
1633         pgoff_t io_end_index;
1634         ENTRY;
1635
1636         if (file) {
1637                 fd = file->private_data;
1638                 ras = &fd->fd_ras;
1639         }
1640
1641         vpg = cl2vvp_page(cl_object_page_slice(page->cp_obj, page));
1642         uptodate = vpg->vpg_defer_uptodate;
1643
1644         if (ll_readahead_enabled(sbi) && !vpg->vpg_ra_updated && ras) {
1645                 struct vvp_io *vio = vvp_env_io(env);
1646                 enum ras_update_flags flags = 0;
1647
1648                 if (uptodate)
1649                         flags |= LL_RAS_HIT;
1650                 if (!vio->vui_ra_valid)
1651                         flags |= LL_RAS_MMAP;
1652                 ras_update(sbi, inode, ras, vvp_index(vpg), flags, io);
1653         }
1654
1655         cl_2queue_init(queue);
1656         if (uptodate) {
1657                 vpg->vpg_ra_used = 1;
1658                 cl_page_export(env, page, 1);
1659                 cl_page_disown(env, io, page);
1660         } else {
1661                 anchor = &vvp_env_info(env)->vti_anchor;
1662                 cl_sync_io_init(anchor, 1);
1663                 page->cp_sync_io = anchor;
1664
1665                 cl_2queue_add(queue, page, true);
1666         }
1667
1668         io_start_index = cl_index(io->ci_obj, io->u.ci_rw.crw_pos);
1669         io_end_index = cl_index(io->ci_obj, io->u.ci_rw.crw_pos +
1670                                 io->u.ci_rw.crw_count - 1);
1671         if (ll_readahead_enabled(sbi) && ras && !io->ci_rand_read) {
1672                 pgoff_t skip_index = 0;
1673
1674                 if (ras->ras_next_readahead_idx < vvp_index(vpg))
1675                         skip_index = vvp_index(vpg);
1676                 rc2 = ll_readahead(env, io, &queue->c2_qin, ras,
1677                                    uptodate, file, skip_index);
1678                 CDEBUG(D_READA|D_IOTRACE, DFID " %d pages read ahead at %lu\n",
1679                        PFID(ll_inode2fid(inode)), rc2, vvp_index(vpg));
1680         } else if (vvp_index(vpg) == io_start_index &&
1681                    io_end_index - io_start_index > 0) {
1682                 rc2 = ll_readpages(env, io, &queue->c2_qin, io_start_index + 1,
1683                                    io_end_index);
1684                 CDEBUG(D_READA, DFID " %d pages read at %lu\n",
1685                        PFID(ll_inode2fid(inode)), rc2, vvp_index(vpg));
1686         }
1687
1688         if (queue->c2_qin.pl_nr > 0) {
1689                 int count = queue->c2_qin.pl_nr;
1690                 rc = cl_io_submit_rw(env, io, CRT_READ, queue);
1691                 if (rc == 0)
1692                         task_io_account_read(PAGE_SIZE * count);
1693         }
1694
1695
1696         if (anchor != NULL && !cl_page_is_owned(page, io)) { /* have sent */
1697                 rc = cl_sync_io_wait(env, anchor, 0);
1698
1699                 cl_page_assume(env, io, page);
1700                 cl_page_list_del(env, &queue->c2_qout, page);
1701
1702                 if (!PageUptodate(cl_page_vmpage(page))) {
1703                         /* Failed to read a mirror, discard this page so that
1704                          * new page can be created with new mirror.
1705                          *
1706                          * TODO: this is not needed after page reinit
1707                          * route is implemented */
1708                         cl_page_discard(env, io, page);
1709                 }
1710                 cl_page_disown(env, io, page);
1711         }
1712
1713         /* TODO: discard all pages until page reinit route is implemented */
1714         cl_page_list_discard(env, io, &queue->c2_qin);
1715
1716         /* Unlock unsent read pages in case of error. */
1717         cl_page_list_disown(env, io, &queue->c2_qin);
1718
1719         cl_2queue_fini(env, queue);
1720
1721         RETURN(rc);
1722 }
1723
1724 /*
1725  * Possible return value:
1726  * 0 no async readahead triggered and fast read could not be used.
1727  * 1 no async readahead, but fast read could be used.
1728  * 2 async readahead triggered and fast read could be used too.
1729  * < 0 on error.
1730  */
1731 static int kickoff_async_readahead(struct file *file, unsigned long pages)
1732 {
1733         struct ll_readahead_work *lrw;
1734         struct inode *inode = file_inode(file);
1735         struct ll_sb_info *sbi = ll_i2sbi(inode);
1736         struct ll_file_data *fd = file->private_data;
1737         struct ll_readahead_state *ras = &fd->fd_ras;
1738         struct ll_ra_info *ra = &sbi->ll_ra_info;
1739         unsigned long throttle;
1740         pgoff_t start_idx = ras_align(ras, ras->ras_next_readahead_idx);
1741         pgoff_t end_idx = start_idx + pages - 1;
1742
1743         /**
1744          * In case we have a limited max_cached_mb, readahead
1745          * should be stopped if it have run out of all LRU slots.
1746          */
1747         if (atomic_read(&ra->ra_cur_pages) >= sbi->ll_cache->ccc_lru_max) {
1748                 ll_ra_stats_inc(inode, RA_STAT_MAX_IN_FLIGHT);
1749                 return 0;
1750         }
1751
1752         throttle = min(ra->ra_async_pages_per_file_threshold,
1753                        ra->ra_max_pages_per_file);
1754         /*
1755          * If this is strided i/o or the window is smaller than the
1756          * throttle limit, we do not do async readahead. Otherwise,
1757          * we do async readahead, allowing the user thread to do fast i/o.
1758          */
1759         if (stride_io_mode(ras) || !throttle ||
1760             ras->ras_window_pages < throttle ||
1761             atomic_read(&ra->ra_async_inflight) > ra->ra_async_max_active)
1762                 return 0;
1763
1764         if ((atomic_read(&ra->ra_cur_pages) + pages) > ra->ra_max_pages)
1765                 return 0;
1766
1767         if (ras->ras_async_last_readpage_idx == start_idx)
1768                 return 1;
1769
1770         /* ll_readahead_work_free() free it */
1771         OBD_ALLOC_PTR(lrw);
1772         if (lrw) {
1773                 atomic_inc(&sbi->ll_ra_info.ra_async_inflight);
1774                 lrw->lrw_file = get_file(file);
1775                 lrw->lrw_start_idx = start_idx;
1776                 lrw->lrw_end_idx = end_idx;
1777                 lrw->lrw_user_pid = current->pid;
1778                 spin_lock(&ras->ras_lock);
1779                 ras->ras_next_readahead_idx = end_idx + 1;
1780                 ras->ras_async_last_readpage_idx = start_idx;
1781                 spin_unlock(&ras->ras_lock);
1782                 memcpy(lrw->lrw_jobid, ll_i2info(inode)->lli_jobid,
1783                        sizeof(lrw->lrw_jobid));
1784                 ll_readahead_work_add(inode, lrw);
1785         } else {
1786                 return -ENOMEM;
1787         }
1788
1789         return 2;
1790 }
1791
1792 /*
1793  * Check if we can issue a readahead RPC, if that is
1794  * the case, we can't do fast IO because we will need
1795  * a cl_io to issue the RPC.
1796  */
1797 static bool ll_use_fast_io(struct file *file,
1798                            struct ll_readahead_state *ras, pgoff_t index)
1799 {
1800         unsigned long fast_read_pages =
1801                 max(RA_REMAIN_WINDOW_MIN, ras->ras_rpc_pages);
1802         loff_t skip_pages;
1803         loff_t stride_bytes = ras->ras_stride_bytes;
1804
1805         if (stride_io_mode(ras) && stride_bytes) {
1806                 skip_pages = (ras->ras_stride_length +
1807                         ras->ras_stride_bytes - 1) / stride_bytes;
1808                 skip_pages *= fast_read_pages;
1809         } else {
1810                 skip_pages = fast_read_pages;
1811         }
1812
1813         if (ras->ras_window_start_idx + ras->ras_window_pages <
1814             ras->ras_next_readahead_idx + skip_pages ||
1815             kickoff_async_readahead(file, fast_read_pages) > 0)
1816                 return true;
1817
1818         return false;
1819 }
1820
1821 int ll_readpage(struct file *file, struct page *vmpage)
1822 {
1823         struct inode *inode = file_inode(file);
1824         struct cl_object *clob = ll_i2info(inode)->lli_clob;
1825         struct ll_cl_context *lcc;
1826         const struct lu_env  *env = NULL;
1827         struct cl_io   *io = NULL;
1828         struct cl_page *page;
1829         struct ll_sb_info *sbi = ll_i2sbi(inode);
1830         int result;
1831         ENTRY;
1832
1833         lcc = ll_cl_find(inode);
1834         if (lcc != NULL) {
1835                 env = lcc->lcc_env;
1836                 io  = lcc->lcc_io;
1837         }
1838
1839         if (io == NULL) { /* fast read */
1840                 struct inode *inode = file_inode(file);
1841                 struct ll_file_data *fd = file->private_data;
1842                 struct ll_readahead_state *ras = &fd->fd_ras;
1843                 struct lu_env  *local_env = NULL;
1844                 struct vvp_page *vpg;
1845
1846                 result = -ENODATA;
1847
1848                 /* TODO: need to verify the layout version to make sure
1849                  * the page is not invalid due to layout change. */
1850                 page = cl_vmpage_page(vmpage, clob);
1851                 if (page == NULL) {
1852                         unlock_page(vmpage);
1853                         ll_ra_stats_inc_sbi(sbi, RA_STAT_FAILED_FAST_READ);
1854                         RETURN(result);
1855                 }
1856
1857                 vpg = cl2vvp_page(cl_object_page_slice(page->cp_obj, page));
1858                 if (vpg->vpg_defer_uptodate) {
1859                         enum ras_update_flags flags = LL_RAS_HIT;
1860
1861                         if (lcc && lcc->lcc_type == LCC_MMAP)
1862                                 flags |= LL_RAS_MMAP;
1863
1864                         /* For fast read, it updates read ahead state only
1865                          * if the page is hit in cache because non cache page
1866                          * case will be handled by slow read later. */
1867                         ras_update(sbi, inode, ras, vvp_index(vpg), flags, io);
1868                         /* avoid duplicate ras_update() call */
1869                         vpg->vpg_ra_updated = 1;
1870
1871                         if (ll_use_fast_io(file, ras, vvp_index(vpg)))
1872                                 result = 0;
1873                 }
1874
1875                 if (!env) {
1876                         local_env = cl_env_percpu_get();
1877                         env = local_env;
1878                 }
1879
1880                 /* export the page and skip io stack */
1881                 if (result == 0) {
1882                         vpg->vpg_ra_used = 1;
1883                         cl_page_export(env, page, 1);
1884                 } else {
1885                         ll_ra_stats_inc_sbi(sbi, RA_STAT_FAILED_FAST_READ);
1886                 }
1887                 /* release page refcount before unlocking the page to ensure
1888                  * the object won't be destroyed in the calling path of
1889                  * cl_page_put(). Please see comment in ll_releasepage(). */
1890                 cl_page_put(env, page);
1891                 unlock_page(vmpage);
1892                 if (local_env)
1893                         cl_env_percpu_put(local_env);
1894
1895                 RETURN(result);
1896         }
1897
1898         /**
1899          * Direct read can fall back to buffered read, but DIO is done
1900          * with lockless i/o, and buffered requires LDLM locking, so in
1901          * this case we must restart without lockless.
1902          */
1903         if (file->f_flags & O_DIRECT &&
1904             lcc && lcc->lcc_type == LCC_RW &&
1905             !io->ci_dio_lock) {
1906                 unlock_page(vmpage);
1907                 io->ci_dio_lock = 1;
1908                 io->ci_need_restart = 1;
1909                 RETURN(-ENOLCK);
1910         }
1911
1912         LASSERT(io->ci_state == CIS_IO_GOING);
1913         page = cl_page_find(env, clob, vmpage->index, vmpage, CPT_CACHEABLE);
1914         if (!IS_ERR(page)) {
1915                 LASSERT(page->cp_type == CPT_CACHEABLE);
1916                 if (likely(!PageUptodate(vmpage))) {
1917                         cl_page_assume(env, io, page);
1918
1919                         result = ll_io_read_page(env, io, page, file);
1920                 } else {
1921                         /* Page from a non-object file. */
1922                         unlock_page(vmpage);
1923                         result = 0;
1924                 }
1925                 cl_page_put(env, page);
1926         } else {
1927                 unlock_page(vmpage);
1928                 result = PTR_ERR(page);
1929         }
1930         RETURN(result);
1931 }