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   *cp;
 201         enum ra_stat      which = _NR_RA_STAT; /* keep gcc happy */
 202         int               rc    = 0;
 203         const char       *msg   = NULL;
 204
 205         ENTRY;
 206
 207         switch (hint) {
 208         case MAYNEED:
 209                 vmpage = grab_cache_page_nowait(inode->i_mapping, index);
 210                 if (vmpage == NULL) {
 211                         which = RA_STAT_FAILED_GRAB_PAGE;
 212                         msg   = "g_c_p_n failed";
 213                         GOTO(out, rc = -EBUSY);
 214                 }
 215                 break;
 216         case WILLNEED:
 217                 vmpage = find_or_create_page(inode->i_mapping, index,
 218                                              GFP_NOFS);
 219                 if (vmpage == NULL)
 220                         GOTO(out, rc = -ENOMEM);
 221                 break;
 222         default:
 223                 /* should not come here */
 224                 GOTO(out, rc = -EINVAL);
 225         }
 226
 227         /* Check if vmpage was truncated or reclaimed */
 228         if (vmpage->mapping != inode->i_mapping) {
 229                 which = RA_STAT_WRONG_GRAB_PAGE;
 230                 msg   = "g_c_p_n returned invalid page";
 231                 GOTO(out, rc = -EBUSY);
 232         }
 233
 234         cp = cl_page_find(env, clob, vmpage->index, vmpage, CPT_CACHEABLE);
 235         if (IS_ERR(cp)) {
 236                 which = RA_STAT_FAILED_GRAB_PAGE;
 237                 msg   = "cl_page_find failed";
 238                 GOTO(out, rc = PTR_ERR(cp));
 239         }
 240
 241         lu_ref_add(&cp->cp_reference, "ra", current);
 242         cl_page_assume(env, io, cp);
 243
 244         if (!cp->cp_defer_uptodate && !PageUptodate(vmpage)) {
 245                 if (hint == MAYNEED) {
 246                         cp->cp_defer_uptodate = 1;
 247                         cp->cp_ra_used = 0;
 248                 }
 249
 250                 cl_page_list_add(queue, cp, true);
 251         } else {
 252                 /* skip completed pages */
 253                 cl_page_unassume(env, io, cp);
 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(&cp->cp_reference, "ra", current);
 260         cl_page_put(env, cp);
 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:"DFID": async ra from %lu to %lu triggered by user pid %d\n",
 596                file_dentry(file)->d_name.name, PFID(ll_inode2fid(inode)),
 597                work->lrw_start_idx, 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, &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                         pgoff_t *start_idx)
 708 {
 709         struct vvp_io *vio = vvp_env_io(env);
 710         struct ll_thread_info *lti = ll_env_info(env);
 711         unsigned long pages, pages_min = 0;
 712         pgoff_t ra_end_idx = 0, end_idx = 0;
 713         struct inode *inode;
 714         struct ra_io_arg *ria = &lti->lti_ria;
 715         struct cl_object *clob;
 716         int ret = 0;
 717         __u64 kms;
 718         struct ll_sb_info *sbi;
 719         struct ll_ra_info *ra;
 720
 721         ENTRY;
 722
 723         clob = io->ci_obj;
 724         inode = vvp_object_inode(clob);
 725         sbi = ll_i2sbi(inode);
 726         ra = &sbi->ll_ra_info;
 727
 728         /**
 729          * In case we have a limited max_cached_mb, readahead
 730          * should be stopped if it have run out of all LRU slots.
 731          */
 732         if (atomic_read(&ra->ra_cur_pages) >= sbi->ll_cache->ccc_lru_max) {
 733                 ll_ra_stats_inc(inode, RA_STAT_MAX_IN_FLIGHT);
 734                 RETURN(0);
 735         }
 736
 737         memset(ria, 0, sizeof(*ria));
 738         ret = ll_readahead_file_kms(env, io, &kms);
 739         if (ret != 0)
 740                 RETURN(ret);
 741
 742         if (kms == 0) {
 743                 ll_ra_stats_inc(inode, RA_STAT_ZERO_LEN);
 744                 RETURN(0);
 745         }
 746
 747         spin_lock(&ras->ras_lock);
 748
 749         /**
 750          * Note: other thread might rollback the ras_next_readahead_idx,
 751          * if it can not get the full size of prepared pages, see the
 752          * end of this function. For stride read ahead, it needs to
 753          * make sure the offset is no less than ras_stride_offset,
 754          * so that stride read ahead can work correctly.
 755          */
 756         if (stride_io_mode(ras))
 757                 *start_idx = max_t(pgoff_t, ras->ras_next_readahead_idx,
 758                                   ras->ras_stride_offset >> PAGE_SHIFT);
 759         else
 760                 *start_idx = ras->ras_next_readahead_idx;
 761
 762         if (ras->ras_window_pages > 0)
 763                 end_idx = ras->ras_window_start_idx + ras->ras_window_pages - 1;
 764
 765         if (skip_index)
 766                 end_idx = *start_idx + ras->ras_window_pages - 1;
 767
 768         /* Enlarge the RA window to encompass the full read */
 769         if (vio->vui_ra_valid &&
 770             end_idx < vio->vui_ra_start_idx + vio->vui_ra_pages - 1)
 771                 end_idx = vio->vui_ra_start_idx + vio->vui_ra_pages - 1;
 772
 773         if (end_idx != 0) {
 774                 pgoff_t eof_index;
 775
 776                 /* Truncate RA window to end of file */
 777                 eof_index = (pgoff_t)((kms - 1) >> PAGE_SHIFT);
 778                 if (eof_index <= end_idx) {
 779                         end_idx = eof_index;
 780                         ria->ria_eof = true;
 781                 }
 782         }
 783         ria->ria_start_idx = *start_idx;
 784         ria->ria_end_idx = end_idx;
 785         /* If stride I/O mode is detected, get stride window*/
 786         if (stride_io_mode(ras)) {
 787                 ria->ria_stoff = ras->ras_stride_offset;
 788                 ria->ria_length = ras->ras_stride_length;
 789                 ria->ria_bytes = ras->ras_stride_bytes;
 790         }
 791         spin_unlock(&ras->ras_lock);
 792
 793         if (end_idx == 0) {
 794                 ll_ra_stats_inc(inode, RA_STAT_ZERO_WINDOW);
 795                 RETURN(0);
 796         }
 797         pages = ria_page_count(ria);
 798         if (pages == 0) {
 799                 ll_ra_stats_inc(inode, RA_STAT_ZERO_WINDOW);
 800                 RETURN(0);
 801         }
 802
 803         RAS_CDEBUG(ras);
 804         CDEBUG(D_READA, DFID": ria: %lu/%lu, bead: %lu/%lu, hit: %d\n",
 805                PFID(lu_object_fid(&clob->co_lu)),
 806                ria->ria_start_idx, ria->ria_end_idx,
 807                vio->vui_ra_valid ? vio->vui_ra_start_idx : 0,
 808                vio->vui_ra_valid ? vio->vui_ra_pages : 0,
 809                hit);
 810
 811         /* at least to extend the readahead window to cover current read */
 812         if (!hit && vio->vui_ra_valid &&
 813             vio->vui_ra_start_idx + vio->vui_ra_pages > ria->ria_start_idx) {
 814                 ria->ria_end_idx_min =
 815                         vio->vui_ra_start_idx + vio->vui_ra_pages - 1;
 816                 pages_min = vio->vui_ra_start_idx + vio->vui_ra_pages -
 817                                 ria->ria_start_idx;
 818                  /**
 819                   * For performance reason, exceeding @ra_max_pages
 820                   * are allowed, but this should be limited with RPC
 821                   * size in case a large block size read issued. Trim
 822                   * to RPC boundary.
 823                   */
 824                 pages_min = min(pages_min, ras->ras_rpc_pages -
 825                                 (ria->ria_start_idx % ras->ras_rpc_pages));
 826         }
 827
 828         /* don't over reserved for mmap range read */
 829         if (skip_index)
 830                 pages_min = 0;
 831         if (pages_min > pages)
 832                 pages = pages_min;
 833         ria->ria_reserved = ll_ra_count_get(ll_i2sbi(inode), ria, pages,
 834                                             pages_min);
 835         if (ria->ria_reserved < pages)
 836                 ll_ra_stats_inc(inode, RA_STAT_MAX_IN_FLIGHT);
 837
 838         CDEBUG(D_READA, "reserved pages: %lu/%lu/%lu, ra_cur %d, ra_max %lu\n",
 839                ria->ria_reserved, pages, pages_min,
 840                atomic_read(&ll_i2sbi(inode)->ll_ra_info.ra_cur_pages),
 841                ll_i2sbi(inode)->ll_ra_info.ra_max_pages);
 842
 843         ret = ll_read_ahead_pages(env, io, queue, ras, ria, &ra_end_idx,
 844                                   skip_index);
 845         if (ria->ria_reserved != 0)
 846                 ll_ra_count_put(ll_i2sbi(inode), ria->ria_reserved);
 847
 848         if (ra_end_idx == end_idx && ra_end_idx == (kms >> PAGE_SHIFT))
 849                 ll_ra_stats_inc(inode, RA_STAT_EOF);
 850
 851         CDEBUG(D_READA,
 852                "ra_end_idx = %lu end_idx = %lu stride end = %lu pages = %d\n",
 853                ra_end_idx, end_idx, ria->ria_end_idx, ret);
 854
 855         if (ra_end_idx != end_idx)
 856                 ll_ra_stats_inc(inode, RA_STAT_FAILED_REACH_END);
 857         if (ra_end_idx > 0) {
 858                 /* update the ras so that the next read-ahead tries from
 859                  * where we left off. */
 860                 spin_lock(&ras->ras_lock);
 861                 ras->ras_next_readahead_idx = ra_end_idx + 1;
 862                 spin_unlock(&ras->ras_lock);
 863                 RAS_CDEBUG(ras);
 864         }
 865
 866         RETURN(ret);
 867 }
 868
 869 static int ll_readpages(const struct lu_env *env, struct cl_io *io,
 870                         struct cl_page_list *queue,
 871                         pgoff_t start, pgoff_t end)
 872 {
 873         int ret = 0;
 874         __u64 kms;
 875         pgoff_t page_idx;
 876         int count = 0;
 877
 878         ENTRY;
 879
 880         ret = ll_readahead_file_kms(env, io, &kms);
 881         if (ret != 0)
 882                 RETURN(ret);
 883
 884         if (kms == 0)
 885                 RETURN(0);
 886
 887         if (end != 0) {
 888                 unsigned long end_index;
 889
 890                 end_index = (unsigned long)((kms - 1) >> PAGE_SHIFT);
 891                 if (end_index <= end)
 892                         end = end_index;
 893         }
 894
 895         for (page_idx = start; page_idx <= end; page_idx++) {
 896                 ret= ll_read_ahead_page(env, io, queue, page_idx,
 897                                         WILLNEED);
 898                 if (ret < 0)
 899                         break;
 900                 else if (ret == 0) /* ret 1 is already uptodate */
 901                         count++;
 902         }
 903
 904         RETURN(count > 0 ? count : ret);
 905 }
 906
 907 static void ras_set_start(struct ll_readahead_state *ras, pgoff_t index)
 908 {
 909         ras->ras_window_start_idx = ras_align(ras, index);
 910 }
 911
 912 /* called with the ras_lock held or from places where it doesn't matter */
 913 static void ras_reset(struct ll_readahead_state *ras, pgoff_t index)
 914 {
 915         ras->ras_consecutive_requests = 0;
 916         ras->ras_consecutive_bytes = 0;
 917         ras->ras_window_pages = 0;
 918         ras_set_start(ras, index);
 919         ras->ras_next_readahead_idx = max(ras->ras_window_start_idx, index + 1);
 920
 921         RAS_CDEBUG(ras);
 922 }
 923
 924 /* called with the ras_lock held or from places where it doesn't matter */
 925 static void ras_stride_reset(struct ll_readahead_state *ras)
 926 {
 927         ras->ras_consecutive_stride_requests = 0;
 928         ras->ras_stride_length = 0;
 929         ras->ras_stride_bytes = 0;
 930         RAS_CDEBUG(ras);
 931 }
 932
 933 void ll_readahead_init(struct inode *inode, struct ll_readahead_state *ras)
 934 {
 935         spin_lock_init(&ras->ras_lock);
 936         ras->ras_rpc_pages = PTLRPC_MAX_BRW_PAGES;
 937         ras_reset(ras, 0);
 938         ras->ras_last_read_end_bytes = 0;
 939         ras->ras_requests = 0;
 940         ras->ras_range_min_start_idx = 0;
 941         ras->ras_range_max_end_idx = 0;
 942         ras->ras_range_requests = 0;
 943         ras->ras_last_range_pages = 0;
 944 }
 945
 946 /*
 947  * Check whether the read request is in the stride window.
 948  * If it is in the stride window, return true, otherwise return false.
 949  */
 950 static bool read_in_stride_window(struct ll_readahead_state *ras,
 951                                   loff_t pos, loff_t count)
 952 {
 953         loff_t stride_gap;
 954
 955         if (ras->ras_stride_length == 0 || ras->ras_stride_bytes == 0 ||
 956             ras->ras_stride_bytes == ras->ras_stride_length)
 957                 return false;
 958
 959         stride_gap = pos - ras->ras_last_read_end_bytes - 1;
 960
 961         /* If it is contiguous read */
 962         if (stride_gap == 0)
 963                 return ras->ras_consecutive_bytes + count <=
 964                         ras->ras_stride_bytes;
 965
 966         /* Otherwise check the stride by itself */
 967         return (ras->ras_stride_length - ras->ras_stride_bytes) == stride_gap &&
 968                 ras->ras_consecutive_bytes == ras->ras_stride_bytes &&
 969                 count <= ras->ras_stride_bytes;
 970 }
 971
 972 static void ras_init_stride_detector(struct ll_readahead_state *ras,
 973                                      loff_t pos, loff_t count)
 974 {
 975         loff_t stride_gap = pos - ras->ras_last_read_end_bytes - 1;
 976
 977         LASSERT(ras->ras_consecutive_stride_requests == 0);
 978
 979         if (pos <= ras->ras_last_read_end_bytes) {
 980                 /*Reset stride window for forward read*/
 981                 ras_stride_reset(ras);
 982                 return;
 983         }
 984
 985         ras->ras_stride_bytes = ras->ras_consecutive_bytes;
 986         ras->ras_stride_length = stride_gap + ras->ras_consecutive_bytes;
 987         ras->ras_consecutive_stride_requests++;
 988         ras->ras_stride_offset = pos;
 989
 990         RAS_CDEBUG(ras);
 991 }
 992
 993 static unsigned long
 994 stride_page_count(struct ll_readahead_state *ras, loff_t len)
 995 {
 996         loff_t bytes_count =
 997                 stride_byte_count(ras->ras_stride_offset,
 998                                   ras->ras_stride_length, ras->ras_stride_bytes,
 999                                   ras->ras_window_start_idx << PAGE_SHIFT, len);
1000
1001         return (bytes_count + PAGE_SIZE - 1) >> PAGE_SHIFT;
1002 }
1003
1004 /* Stride Read-ahead window will be increased inc_len according to
1005  * stride I/O pattern */
1006 static void ras_stride_increase_window(struct ll_readahead_state *ras,
1007                                        struct ll_ra_info *ra, loff_t inc_bytes)
1008 {
1009         loff_t window_bytes, stride_bytes;
1010         u64 left_bytes;
1011         u64 step;
1012         loff_t end;
1013
1014         /* temporarily store in page units to reduce LASSERT() cost below */
1015         end = ras->ras_window_start_idx + ras->ras_window_pages;
1016
1017         LASSERT(ras->ras_stride_length > 0);
1018         LASSERTF(end >= (ras->ras_stride_offset >> PAGE_SHIFT),
1019                  "window_start_idx %lu, window_pages %lu stride_offset %llu\n",
1020                  ras->ras_window_start_idx, ras->ras_window_pages,
1021                  ras->ras_stride_offset);
1022
1023         end <<= PAGE_SHIFT;
1024         if (end <= ras->ras_stride_offset)
1025                 stride_bytes = 0;
1026         else
1027                 stride_bytes = end - ras->ras_stride_offset;
1028
1029         div64_u64_rem(stride_bytes, ras->ras_stride_length, &left_bytes);
1030         window_bytes = (ras->ras_window_pages << PAGE_SHIFT);
1031         if (left_bytes < ras->ras_stride_bytes) {
1032                 if (ras->ras_stride_bytes - left_bytes >= inc_bytes) {
1033                         window_bytes += inc_bytes;
1034                         goto out;
1035                 } else {
1036                         window_bytes += (ras->ras_stride_bytes - left_bytes);
1037                         inc_bytes -= (ras->ras_stride_bytes - left_bytes);
1038                 }
1039         } else {
1040                 window_bytes += (ras->ras_stride_length - left_bytes);
1041         }
1042
1043         LASSERT(ras->ras_stride_bytes != 0);
1044
1045         step = div64_u64_rem(inc_bytes, ras->ras_stride_bytes, &left_bytes);
1046
1047         window_bytes += step * ras->ras_stride_length + left_bytes;
1048         LASSERT(window_bytes > 0);
1049
1050 out:
1051         if (stride_page_count(ras, window_bytes) <=
1052             ra->ra_max_pages_per_file || ras->ras_window_pages == 0)
1053                 ras->ras_window_pages = (window_bytes >> PAGE_SHIFT);
1054
1055         LASSERT(ras->ras_window_pages > 0);
1056
1057         RAS_CDEBUG(ras);
1058 }
1059
1060 static void ras_increase_window(struct inode *inode,
1061                                 struct ll_readahead_state *ras,
1062                                 struct ll_ra_info *ra)
1063 {
1064         /* The stretch of ra-window should be aligned with max rpc_size
1065          * but current clio architecture does not support retrieve such
1066          * information from lower layer. FIXME later
1067          */
1068         if (stride_io_mode(ras)) {
1069                 ras_stride_increase_window(ras, ra,
1070                                       (loff_t)ras->ras_rpc_pages << PAGE_SHIFT);
1071         } else {
1072                 pgoff_t window_pages;
1073
1074                 window_pages = min(ras->ras_window_pages + ras->ras_rpc_pages,
1075                                    ra->ra_max_pages_per_file);
1076                 if (window_pages < ras->ras_rpc_pages)
1077                         ras->ras_window_pages = window_pages;
1078                 else
1079                         ras->ras_window_pages = ras_align(ras, window_pages);
1080         }
1081 }
1082
1083 /**
1084  * Seek within 8 pages are considered as sequential read for now.
1085  */
1086 static inline bool is_loose_seq_read(struct ll_readahead_state *ras, loff_t pos)
1087 {
1088         return pos_in_window(pos, ras->ras_last_read_end_bytes,
1089                              8UL << PAGE_SHIFT, 8UL << PAGE_SHIFT);
1090 }
1091
1092 static inline bool is_loose_mmap_read(struct ll_sb_info *sbi,
1093                                       struct ll_readahead_state *ras,
1094                                       unsigned long pos)
1095 {
1096         unsigned long range_pages = sbi->ll_ra_info.ra_range_pages;
1097
1098         return pos_in_window(pos, ras->ras_last_read_end_bytes,
1099                              range_pages << PAGE_SHIFT,
1100                              range_pages << PAGE_SHIFT);
1101 }
1102
1103 /**
1104  * We have observed slow mmap read performances for some
1105  * applications. The problem is if access pattern is neither
1106  * sequential nor stride, but could be still adjacent in a
1107  * small range and then seek a random position.
1108  *
1109  * So the pattern could be something like this:
1110  *
1111  * [1M data] [hole] [0.5M data] [hole] [0.7M data] [1M data]
1112  *
1113  *
1114  * Every time an application reads mmap data, it may not only
1115  * read a single 4KB page, but aslo a cluster of nearby pages in
1116  * a range(e.g. 1MB) of the first page after a cache miss.
1117  *
1118  * The readahead engine is modified to track the range size of
1119  * a cluster of mmap reads, so that after a seek and/or cache miss,
1120  * the range size is used to efficiently prefetch multiple pages
1121  * in a single RPC rather than many small RPCs.
1122  */
1123 static void ras_detect_cluster_range(struct ll_readahead_state *ras,
1124                                      struct ll_sb_info *sbi,
1125                                      unsigned long pos, unsigned long count)
1126 {
1127         pgoff_t last_pages, pages;
1128         pgoff_t end_idx = (pos + count - 1) >> PAGE_SHIFT;
1129
1130         last_pages = ras->ras_range_max_end_idx -
1131                         ras->ras_range_min_start_idx + 1;
1132         /* First time come here */
1133         if (!ras->ras_range_max_end_idx)
1134                 goto out;
1135
1136         /* Random or Stride read */
1137         if (!is_loose_mmap_read(sbi, ras, pos))
1138                 goto out;
1139
1140         ras->ras_range_requests++;
1141         if (ras->ras_range_max_end_idx < end_idx)
1142                 ras->ras_range_max_end_idx = end_idx;
1143
1144         if (ras->ras_range_min_start_idx > (pos >> PAGE_SHIFT))
1145                 ras->ras_range_min_start_idx = pos >> PAGE_SHIFT;
1146
1147         /* Out of range, consider it as random or stride */
1148         pages = ras->ras_range_max_end_idx -
1149                         ras->ras_range_min_start_idx + 1;
1150         if (pages <= sbi->ll_ra_info.ra_range_pages)
1151                 return;
1152 out:
1153         ras->ras_last_range_pages = last_pages;
1154         ras->ras_range_requests = 0;
1155         ras->ras_range_min_start_idx = pos >> PAGE_SHIFT;
1156         ras->ras_range_max_end_idx = end_idx;
1157 }
1158
1159 static void ras_detect_read_pattern(struct ll_readahead_state *ras,
1160                                     struct ll_sb_info *sbi,
1161                                     loff_t pos, size_t count, bool mmap)
1162 {
1163         bool stride_detect = false;
1164         pgoff_t index = pos >> PAGE_SHIFT;
1165
1166         /*
1167          * Reset the read-ahead window in two cases. First when the app seeks
1168          * or reads to some other part of the file. Secondly if we get a
1169          * read-ahead miss that we think we've previously issued. This can
1170          * be a symptom of there being so many read-ahead pages that the VM
1171          * is reclaiming it before we get to it.
1172          */
1173         if (!is_loose_seq_read(ras, pos)) {
1174                 /* Check whether it is in stride I/O mode */
1175                 if (!read_in_stride_window(ras, pos, count)) {
1176                         if (ras->ras_consecutive_stride_requests == 0)
1177                                 ras_init_stride_detector(ras, pos, count);
1178                         else
1179                                 ras_stride_reset(ras);
1180                         ras->ras_consecutive_bytes = 0;
1181                         ras_reset(ras, index);
1182                 } else {
1183                         ras->ras_consecutive_bytes = 0;
1184                         ras->ras_consecutive_requests = 0;
1185                         if (++ras->ras_consecutive_stride_requests > 1)
1186                                 stride_detect = true;
1187                         RAS_CDEBUG(ras);
1188                 }
1189                 ll_ra_stats_inc_sbi(sbi, RA_STAT_DISTANT_READPAGE);
1190         } else if (stride_io_mode(ras)) {
1191                 /*
1192                  * If this is contiguous read but in stride I/O mode
1193                  * currently, check whether stride step still is valid,
1194                  * if invalid, it will reset the stride ra window to
1195                  * be zero.
1196                  */
1197                 if (!read_in_stride_window(ras, pos, count)) {
1198                         ras_stride_reset(ras);
1199                         ras->ras_window_pages = 0;
1200                         ras->ras_next_readahead_idx = index;
1201                 }
1202         }
1203
1204         ras->ras_consecutive_bytes += count;
1205         if (mmap) {
1206                 pgoff_t idx = ras->ras_consecutive_bytes >> PAGE_SHIFT;
1207                 unsigned long ra_range_pages =
1208                                 max_t(unsigned long, RA_MIN_MMAP_RANGE_PAGES,
1209                                       sbi->ll_ra_info.ra_range_pages);
1210
1211                 if ((idx >= ra_range_pages &&
1212                      idx % ra_range_pages == 0) || stride_detect)
1213                         ras->ras_need_increase_window = true;
1214         } else if ((ras->ras_consecutive_requests > 1 || stride_detect)) {
1215                 ras->ras_need_increase_window = true;
1216         }
1217
1218         ras->ras_last_read_end_bytes = pos + count - 1;
1219 }
1220
1221 void ll_ras_enter(struct file *f, loff_t pos, size_t count)
1222 {
1223         struct ll_file_data *fd = f->private_data;
1224         struct ll_readahead_state *ras = &fd->fd_ras;
1225         struct inode *inode = file_inode(f);
1226         unsigned long index = pos >> PAGE_SHIFT;
1227         struct ll_sb_info *sbi = ll_i2sbi(inode);
1228
1229         spin_lock(&ras->ras_lock);
1230         ras->ras_requests++;
1231         ras->ras_consecutive_requests++;
1232         ras->ras_need_increase_window = false;
1233         ras->ras_no_miss_check = false;
1234         /*
1235          * On the second access to a file smaller than the tunable
1236          * ra_max_read_ahead_whole_pages trigger RA on all pages in the
1237          * file up to ra_max_pages_per_file.  This is simply a best effort
1238          * and only occurs once per open file. Normal RA behavior is reverted
1239          * to for subsequent IO.
1240          */
1241         if (ras->ras_requests >= 2) {
1242                 __u64 kms_pages;
1243                 struct ll_ra_info *ra = &sbi->ll_ra_info;
1244
1245                 kms_pages = (i_size_read(inode) + PAGE_SIZE - 1) >>
1246                             PAGE_SHIFT;
1247
1248                 CDEBUG(D_READA, "kmsp %llu mwp %lu mp %lu\n", kms_pages,
1249                        ra->ra_max_read_ahead_whole_pages,
1250                        ra->ra_max_pages_per_file);
1251
1252                 if (kms_pages &&
1253                     kms_pages <= ra->ra_max_read_ahead_whole_pages) {
1254                         ras->ras_window_start_idx = 0;
1255                         ras->ras_next_readahead_idx = index + 1;
1256                         ras->ras_window_pages = min(ra->ra_max_pages_per_file,
1257                                             ra->ra_max_read_ahead_whole_pages);
1258                         ras->ras_no_miss_check = true;
1259                         GOTO(out_unlock, 0);
1260                 }
1261         }
1262         ras_detect_read_pattern(ras, sbi, pos, count, false);
1263 out_unlock:
1264         spin_unlock(&ras->ras_lock);
1265 }
1266
1267 static bool index_in_stride_window(struct ll_readahead_state *ras,
1268                                    pgoff_t index)
1269 {
1270         loff_t pos = (loff_t)index << PAGE_SHIFT;
1271
1272         if (ras->ras_stride_length == 0 || ras->ras_stride_bytes == 0 ||
1273             ras->ras_stride_bytes == ras->ras_stride_length)
1274                 return false;
1275
1276         if (pos >= ras->ras_stride_offset) {
1277                 u64 offset;
1278
1279                 div64_u64_rem(pos - ras->ras_stride_offset,
1280                               ras->ras_stride_length, &offset);
1281                 if (offset < ras->ras_stride_bytes ||
1282                     ras->ras_stride_length - offset < PAGE_SIZE)
1283                         return true;
1284         } else if (ras->ras_stride_offset - pos < PAGE_SIZE) {
1285                 return true;
1286         }
1287
1288         return false;
1289 }
1290
1291 /*
1292  * ll_ras_enter() is used to detect read pattern according to pos and count.
1293  *
1294  * ras_update() is used to detect cache miss and
1295  * reset window or increase window accordingly
1296  */
1297 static void ras_update(struct ll_sb_info *sbi, struct inode *inode,
1298                        struct ll_readahead_state *ras, pgoff_t index,
1299                        enum ras_update_flags flags, struct cl_io *io)
1300 {
1301         struct ll_ra_info *ra = &sbi->ll_ra_info;
1302         bool hit = flags & LL_RAS_HIT;
1303
1304         ENTRY;
1305         spin_lock(&ras->ras_lock);
1306
1307         if (!hit)
1308                 CDEBUG(D_READA|D_IOTRACE, DFID " pages at %lu miss.\n",
1309                        PFID(ll_inode2fid(inode)), index);
1310         ll_ra_stats_inc_sbi(sbi, hit ? RA_STAT_HIT : RA_STAT_MISS);
1311
1312         /*
1313          * The readahead window has been expanded to cover whole
1314          * file size, we don't care whether ra miss happen or not.
1315          * Because we will read whole file to page cache even if
1316          * some pages missed.
1317          */
1318         if (ras->ras_no_miss_check)
1319                 GOTO(out_unlock, 0);
1320
1321         if (io && io->ci_rand_read)
1322                 GOTO(out_unlock, 0);
1323
1324         if (io && io->ci_seq_read) {
1325                 if (!hit) {
1326                         /* to avoid many small read RPC here */
1327                         ras->ras_window_pages = sbi->ll_ra_info.ra_range_pages;
1328                         ll_ra_stats_inc_sbi(sbi, RA_STAT_MMAP_RANGE_READ);
1329                 }
1330                 goto skip;
1331         }
1332
1333         if (flags & LL_RAS_MMAP) {
1334                 unsigned long ra_pages;
1335
1336                 ras_detect_cluster_range(ras, sbi, index << PAGE_SHIFT,
1337                                          PAGE_SIZE);
1338                 ras_detect_read_pattern(ras, sbi, (loff_t)index << PAGE_SHIFT,
1339                                         PAGE_SIZE, true);
1340
1341                 /* we did not detect anything but we could prefetch */
1342                 if (!ras->ras_need_increase_window &&
1343                     ras->ras_window_pages <= sbi->ll_ra_info.ra_range_pages &&
1344                     ras->ras_range_requests >= 2) {
1345                         if (!hit) {
1346                                 ra_pages = max_t(unsigned long,
1347                                         RA_MIN_MMAP_RANGE_PAGES,
1348                                         ras->ras_last_range_pages);
1349                                 if (index < ra_pages / 2)
1350                                         index = 0;
1351                                 else
1352                                         index -= ra_pages / 2;
1353                                 ras->ras_window_pages = ra_pages;
1354                                 ll_ra_stats_inc_sbi(sbi,
1355                                         RA_STAT_MMAP_RANGE_READ);
1356                         } else {
1357                                 ras->ras_window_pages = 0;
1358                         }
1359                         goto skip;
1360                 }
1361         }
1362
1363         if (!hit && ras->ras_window_pages &&
1364             index < ras->ras_next_readahead_idx &&
1365             pos_in_window(index, ras->ras_window_start_idx, 0,
1366                           ras->ras_window_pages)) {
1367                 ll_ra_stats_inc_sbi(sbi, RA_STAT_MISS_IN_WINDOW);
1368                 ras->ras_need_increase_window = false;
1369
1370                 if (index_in_stride_window(ras, index) &&
1371                     stride_io_mode(ras)) {
1372                         /*
1373                          * if (index != ras->ras_last_readpage + 1)
1374                          *      ras->ras_consecutive_pages = 0;
1375                          */
1376                         ras_reset(ras, index);
1377
1378                         /*
1379                          * If stride-RA hit cache miss, the stride
1380                          * detector will not be reset to avoid the
1381                          * overhead of redetecting read-ahead mode,
1382                          * but on the condition that the stride window
1383                          * is still intersect with normal sequential
1384                          * read-ahead window.
1385                          */
1386                         if (ras->ras_window_start_idx < ras->ras_stride_offset)
1387                                 ras_stride_reset(ras);
1388                         RAS_CDEBUG(ras);
1389                 } else {
1390                         /*
1391                          * Reset both stride window and normal RA
1392                          * window.
1393                          */
1394                         ras_reset(ras, index);
1395                         /* ras->ras_consecutive_pages++; */
1396                         ras->ras_consecutive_bytes = 0;
1397                         ras_stride_reset(ras);
1398                         GOTO(out_unlock, 0);
1399                 }
1400         }
1401
1402 skip:
1403         ras_set_start(ras, index);
1404
1405         if (stride_io_mode(ras)) {
1406                 /* Since stride readahead is sentivite to the offset
1407                  * of read-ahead, so we use original offset here,
1408                  * instead of ras_window_start_idx, which is RPC aligned.
1409                  */
1410                 ras->ras_next_readahead_idx = max(index + 1,
1411                                                   ras->ras_next_readahead_idx);
1412                 ras->ras_window_start_idx =
1413                                 max_t(pgoff_t, ras->ras_window_start_idx,
1414                                       ras->ras_stride_offset >> PAGE_SHIFT);
1415         } else {
1416                 if (ras->ras_next_readahead_idx < ras->ras_window_start_idx)
1417                         ras->ras_next_readahead_idx = ras->ras_window_start_idx;
1418                 if (!hit)
1419                         ras->ras_next_readahead_idx = index + 1;
1420         }
1421
1422         if (ras->ras_need_increase_window) {
1423                 ras_increase_window(inode, ras, ra);
1424                 ras->ras_need_increase_window = false;
1425         }
1426
1427         EXIT;
1428 out_unlock:
1429         spin_unlock(&ras->ras_lock);
1430 }
1431
1432 int ll_writepage(struct page *vmpage, struct writeback_control *wbc)
1433 {
1434         struct inode           *inode = vmpage->mapping->host;
1435         struct ll_inode_info   *lli   = ll_i2info(inode);
1436         struct lu_env          *env;
1437         struct cl_io           *io;
1438         struct cl_page         *page;
1439         struct cl_object       *clob;
1440         bool redirtied = false;
1441         bool unlocked = false;
1442         int result;
1443         __u16 refcheck;
1444         ENTRY;
1445
1446         LASSERT(PageLocked(vmpage));
1447         LASSERT(!PageWriteback(vmpage));
1448
1449         LASSERT(ll_i2dtexp(inode) != NULL);
1450
1451         env = cl_env_get(&refcheck);
1452         if (IS_ERR(env))
1453                 GOTO(out, result = PTR_ERR(env));
1454
1455         clob  = ll_i2info(inode)->lli_clob;
1456         LASSERT(clob != NULL);
1457
1458         io = vvp_env_thread_io(env);
1459         io->ci_obj = clob;
1460         io->ci_ignore_layout = 1;
1461         result = cl_io_init(env, io, CIT_MISC, clob);
1462         if (result == 0) {
1463                 page = cl_page_find(env, clob, vmpage->index,
1464                                     vmpage, CPT_CACHEABLE);
1465                 if (!IS_ERR(page)) {
1466                         lu_ref_add(&page->cp_reference, "writepage",
1467                                    current);
1468                         cl_page_assume(env, io, page);
1469                         result = cl_page_flush(env, io, page);
1470                         if (result != 0) {
1471                                 /*
1472                                  * Re-dirty page on error so it retries write,
1473                                  * but not in case when IO has actually
1474                                  * occurred and completed with an error.
1475                                  */
1476                                 if (!PageError(vmpage)) {
1477                                         redirty_page_for_writepage(wbc, vmpage);
1478                                         result = 0;
1479                                         redirtied = true;
1480                                 }
1481                         }
1482                         cl_page_disown(env, io, page);
1483                         unlocked = true;
1484                         lu_ref_del(&page->cp_reference,
1485                                    "writepage", current);
1486                         cl_page_put(env, page);
1487                 } else {
1488                         result = PTR_ERR(page);
1489                 }
1490         }
1491         cl_io_fini(env, io);
1492
1493         if (redirtied && wbc->sync_mode == WB_SYNC_ALL) {
1494                 loff_t offset = cl_offset(clob, vmpage->index);
1495
1496                 /* Flush page failed because the extent is being written out.
1497                  * Wait for the write of extent to be finished to avoid
1498                  * breaking kernel which assumes ->writepage should mark
1499                  * PageWriteback or clean the page. */
1500                 result = cl_sync_file_range(inode, offset,
1501                                             offset + PAGE_SIZE - 1,
1502                                             CL_FSYNC_LOCAL, 1);
1503                 if (result > 0) {
1504                         /* actually we may have written more than one page.
1505                          * decreasing this page because the caller will count
1506                          * it. */
1507                         wbc->nr_to_write -= result - 1;
1508                         result = 0;
1509                 }
1510         }
1511
1512         cl_env_put(env, &refcheck);
1513         GOTO(out, result);
1514
1515 out:
1516         if (result < 0) {
1517                 if (!lli->lli_async_rc)
1518                         lli->lli_async_rc = result;
1519                 SetPageError(vmpage);
1520                 if (!unlocked)
1521                         unlock_page(vmpage);
1522         }
1523         return result;
1524 }
1525
1526 int ll_writepages(struct address_space *mapping, struct writeback_control *wbc)
1527 {
1528         struct inode *inode = mapping->host;
1529         loff_t start;
1530         loff_t end;
1531         enum cl_fsync_mode mode;
1532         int range_whole = 0;
1533         int result;
1534         ENTRY;
1535
1536         if (wbc->range_cyclic) {
1537                 start = (loff_t)mapping->writeback_index << PAGE_SHIFT;
1538                 end = OBD_OBJECT_EOF;
1539         } else {
1540                 start = wbc->range_start;
1541                 end = wbc->range_end;
1542                 if (end == LLONG_MAX) {
1543                         end = OBD_OBJECT_EOF;
1544                         range_whole = start == 0;
1545                 }
1546         }
1547
1548         mode = CL_FSYNC_NONE;
1549         if (wbc->sync_mode == WB_SYNC_ALL)
1550                 mode = CL_FSYNC_LOCAL;
1551
1552         if (ll_i2info(inode)->lli_clob == NULL)
1553                 RETURN(0);
1554
1555         /* for directio, it would call writepages() to evict cached pages
1556          * inside the IO context of write, which will cause deadlock at
1557          * layout_conf since it waits for active IOs to complete. */
1558         result = cl_sync_file_range(inode, start, end, mode, 1);
1559         if (result > 0) {
1560                 wbc->nr_to_write -= result;
1561                 result = 0;
1562          }
1563
1564         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0)) {
1565                 if (end == OBD_OBJECT_EOF)
1566                         mapping->writeback_index = 0;
1567                 else
1568                         mapping->writeback_index = (end >> PAGE_SHIFT) + 1;
1569         }
1570         RETURN(result);
1571 }
1572
1573 struct ll_cl_context *ll_cl_find(struct inode *inode)
1574 {
1575         struct ll_inode_info *lli = ll_i2info(inode);
1576         struct ll_cl_context *lcc;
1577         struct ll_cl_context *found = NULL;
1578
1579         read_lock(&lli->lli_lock);
1580         list_for_each_entry(lcc, &lli->lli_lccs, lcc_list) {
1581                 if (lcc->lcc_cookie == current) {
1582                         found = lcc;
1583                         break;
1584                 }
1585         }
1586         read_unlock(&lli->lli_lock);
1587
1588         return found;
1589 }
1590
1591 void ll_cl_add(struct inode *inode, const struct lu_env *env, struct cl_io *io,
1592                enum lcc_type type)
1593 {
1594         struct ll_inode_info *lli = ll_i2info(inode);
1595         struct ll_cl_context *lcc = &ll_env_info(env)->lti_io_ctx;
1596
1597         memset(lcc, 0, sizeof(*lcc));
1598         INIT_LIST_HEAD(&lcc->lcc_list);
1599         lcc->lcc_cookie = current;
1600         lcc->lcc_env = env;
1601         lcc->lcc_io = io;
1602         lcc->lcc_type = type;
1603
1604         write_lock(&lli->lli_lock);
1605         list_add(&lcc->lcc_list, &lli->lli_lccs);
1606         write_unlock(&lli->lli_lock);
1607 }
1608
1609 void ll_cl_remove(struct inode *inode, const struct lu_env *env)
1610 {
1611         struct ll_inode_info *lli = ll_i2info(inode);
1612         struct ll_cl_context *lcc = &ll_env_info(env)->lti_io_ctx;
1613
1614         write_lock(&lli->lli_lock);
1615         list_del_init(&lcc->lcc_list);
1616         write_unlock(&lli->lli_lock);
1617 }
1618
1619 int ll_io_read_page(const struct lu_env *env, struct cl_io *io,
1620                            struct cl_page *page, struct file *file)
1621 {
1622         struct inode              *inode  = vvp_object_inode(page->cp_obj);
1623         struct ll_sb_info         *sbi    = ll_i2sbi(inode);
1624         struct ll_file_data       *fd     = NULL;
1625         struct ll_readahead_state *ras    = NULL;
1626         struct cl_2queue          *queue  = &io->ci_queue;
1627         struct cl_sync_io         *anchor = NULL;
1628         struct vvp_page           *vpg;
1629         int                        rc = 0, rc2 = 0;
1630         bool                       uptodate;
1631         struct vvp_io *vio = vvp_env_io(env);
1632         bool mmap = !vio->vui_ra_valid;
1633         pgoff_t ra_start_index = 0;
1634         pgoff_t io_start_index;
1635         pgoff_t io_end_index;
1636         bool unlockpage = true;
1637         ENTRY;
1638
1639         if (file) {
1640                 fd = file->private_data;
1641                 ras = &fd->fd_ras;
1642         }
1643
1644         /* PagePrivate2 is set in ll_io_zero_page() to tell us the vmpage
1645          * must not be unlocked after processing.
1646          */
1647         if (page->cp_vmpage && PagePrivate2(page->cp_vmpage))
1648                 unlockpage = false;
1649
1650         vpg = cl2vvp_page(cl_object_page_slice(page->cp_obj, page));
1651         uptodate = page->cp_defer_uptodate;
1652
1653         if (ll_readahead_enabled(sbi) && !page->cp_ra_updated && ras) {
1654                 enum ras_update_flags flags = 0;
1655
1656                 if (uptodate)
1657                         flags |= LL_RAS_HIT;
1658                 if (mmap)
1659                         flags |= LL_RAS_MMAP;
1660                 ras_update(sbi, inode, ras, vvp_index(vpg), flags, io);
1661         }
1662
1663         cl_2queue_init(queue);
1664         if (uptodate) {
1665                 page->cp_ra_used = 1;
1666                 SetPageUptodate(page->cp_vmpage);
1667                 cl_page_disown(env, io, page);
1668         } else {
1669                 anchor = &vvp_env_info(env)->vti_anchor;
1670                 cl_sync_io_init(anchor, 1);
1671                 page->cp_sync_io = anchor;
1672
1673                 cl_2queue_add(queue, page, true);
1674         }
1675
1676         /* mmap does not set the ci_rw fields */
1677         if (!mmap) {
1678                 io_start_index = cl_index(io->ci_obj, io->u.ci_rw.crw_pos);
1679                 io_end_index = cl_index(io->ci_obj, io->u.ci_rw.crw_pos +
1680                                         io->u.ci_rw.crw_count - 1);
1681         } else {
1682                 io_start_index = vvp_index(vpg);
1683                 io_end_index = vvp_index(vpg);
1684         }
1685
1686         if (ll_readahead_enabled(sbi) && ras && !io->ci_rand_read) {
1687                 pgoff_t skip_index = 0;
1688
1689                 if (ras->ras_next_readahead_idx < vvp_index(vpg))
1690                         skip_index = vvp_index(vpg);
1691                 rc2 = ll_readahead(env, io, &queue->c2_qin, ras,
1692                                    uptodate, file, skip_index,
1693                                    &ra_start_index);
1694                 /* to keep iotrace clean, we only print here if we actually
1695                  * read pages
1696                  */
1697                 CDEBUG(D_READA | (rc2 ? D_IOTRACE : 0),
1698                        DFID " %d pages read ahead at %lu, triggered by user read at %lu, stride offset %lld, stride length %lld, stride bytes %lld\n",
1699                        PFID(ll_inode2fid(inode)), rc2, ra_start_index,
1700                        vvp_index(vpg), ras->ras_stride_offset,
1701                        ras->ras_stride_length, ras->ras_stride_bytes);
1702
1703         } else if (vvp_index(vpg) == io_start_index &&
1704                    io_end_index - io_start_index > 0) {
1705                 rc2 = ll_readpages(env, io, &queue->c2_qin, io_start_index + 1,
1706                                    io_end_index);
1707                 CDEBUG(D_READA, DFID " %d pages read at %lu\n",
1708                        PFID(ll_inode2fid(inode)), rc2, vvp_index(vpg));
1709         }
1710
1711         if (queue->c2_qin.pl_nr > 0) {
1712                 int count = queue->c2_qin.pl_nr;
1713                 rc = cl_io_submit_rw(env, io, CRT_READ, queue);
1714                 if (rc == 0)
1715                         task_io_account_read(PAGE_SIZE * count);
1716         }
1717
1718
1719         if (anchor != NULL && !cl_page_is_owned(page, io)) { /* have sent */
1720                 rc = cl_sync_io_wait(env, anchor, 0);
1721
1722                 cl_page_assume(env, io, page);
1723                 cl_page_list_del(env, &queue->c2_qout, page);
1724
1725                 if (!PageUptodate(cl_page_vmpage(page))) {
1726                         /* Failed to read a mirror, discard this page so that
1727                          * new page can be created with new mirror.
1728                          *
1729                          * TODO: this is not needed after page reinit
1730                          * route is implemented */
1731                         cl_page_discard(env, io, page);
1732                 }
1733                 if (unlockpage)
1734                         cl_page_disown(env, io, page);
1735         }
1736
1737         /* TODO: discard all pages until page reinit route is implemented */
1738         cl_page_list_discard(env, io, &queue->c2_qin);
1739
1740         /* Unlock unsent read pages in case of error. */
1741         cl_page_list_disown(env, &queue->c2_qin);
1742
1743         cl_2queue_fini(env, queue);
1744
1745         RETURN(rc);
1746 }
1747
1748 /*
1749  * Possible return value:
1750  * 0 no async readahead triggered and fast read could not be used.
1751  * 1 no async readahead, but fast read could be used.
1752  * 2 async readahead triggered and fast read could be used too.
1753  * < 0 on error.
1754  */
1755 static int kickoff_async_readahead(struct file *file, unsigned long pages)
1756 {
1757         struct ll_readahead_work *lrw;
1758         struct inode *inode = file_inode(file);
1759         struct ll_sb_info *sbi = ll_i2sbi(inode);
1760         struct ll_file_data *fd = file->private_data;
1761         struct ll_readahead_state *ras = &fd->fd_ras;
1762         struct ll_ra_info *ra = &sbi->ll_ra_info;
1763         unsigned long throttle;
1764         pgoff_t start_idx = ras_align(ras, ras->ras_next_readahead_idx);
1765         pgoff_t end_idx = start_idx + pages - 1;
1766
1767         /**
1768          * In case we have a limited max_cached_mb, readahead
1769          * should be stopped if it have run out of all LRU slots.
1770          */
1771         if (atomic_read(&ra->ra_cur_pages) >= sbi->ll_cache->ccc_lru_max) {
1772                 ll_ra_stats_inc(inode, RA_STAT_MAX_IN_FLIGHT);
1773                 return 0;
1774         }
1775
1776         throttle = min(ra->ra_async_pages_per_file_threshold,
1777                        ra->ra_max_pages_per_file);
1778         /*
1779          * If this is strided i/o or the window is smaller than the
1780          * throttle limit, we do not do async readahead. Otherwise,
1781          * we do async readahead, allowing the user thread to do fast i/o.
1782          */
1783         if (stride_io_mode(ras) || !throttle ||
1784             ras->ras_window_pages < throttle ||
1785             atomic_read(&ra->ra_async_inflight) > ra->ra_async_max_active)
1786                 return 0;
1787
1788         if ((atomic_read(&ra->ra_cur_pages) + pages) > ra->ra_max_pages)
1789                 return 0;
1790
1791         if (ras->ras_async_last_readpage_idx == start_idx)
1792                 return 1;
1793
1794         /* ll_readahead_work_free() free it */
1795         OBD_ALLOC_PTR(lrw);
1796         if (lrw) {
1797                 atomic_inc(&sbi->ll_ra_info.ra_async_inflight);
1798                 lrw->lrw_file = get_file(file);
1799                 lrw->lrw_start_idx = start_idx;
1800                 lrw->lrw_end_idx = end_idx;
1801                 lrw->lrw_user_pid = current->pid;
1802                 spin_lock(&ras->ras_lock);
1803                 ras->ras_next_readahead_idx = end_idx + 1;
1804                 ras->ras_async_last_readpage_idx = start_idx;
1805                 spin_unlock(&ras->ras_lock);
1806                 memcpy(lrw->lrw_jobid, ll_i2info(inode)->lli_jobid,
1807                        sizeof(lrw->lrw_jobid));
1808                 ll_readahead_work_add(inode, lrw);
1809         } else {
1810                 return -ENOMEM;
1811         }
1812
1813         return 2;
1814 }
1815
1816 /*
1817  * Check if we can issue a readahead RPC, if that is
1818  * the case, we can't do fast IO because we will need
1819  * a cl_io to issue the RPC.
1820  */
1821 static bool ll_use_fast_io(struct file *file,
1822                            struct ll_readahead_state *ras, pgoff_t index)
1823 {
1824         unsigned long fast_read_pages =
1825                 max(RA_REMAIN_WINDOW_MIN, ras->ras_rpc_pages);
1826         loff_t skip_pages;
1827         loff_t stride_bytes = ras->ras_stride_bytes;
1828
1829         if (stride_io_mode(ras) && stride_bytes) {
1830                 skip_pages = (ras->ras_stride_length +
1831                         ras->ras_stride_bytes - 1) / stride_bytes;
1832                 skip_pages *= fast_read_pages;
1833         } else {
1834                 skip_pages = fast_read_pages;
1835         }
1836
1837         if (ras->ras_window_start_idx + ras->ras_window_pages <
1838             ras->ras_next_readahead_idx + skip_pages ||
1839             kickoff_async_readahead(file, fast_read_pages) > 0)
1840                 return true;
1841
1842         return false;
1843 }
1844
1845 int ll_readpage(struct file *file, struct page *vmpage)
1846 {
1847         struct inode *inode = file_inode(file);
1848         struct cl_object *clob = ll_i2info(inode)->lli_clob;
1849         struct ll_cl_context *lcc;
1850         const struct lu_env  *env = NULL;
1851         struct cl_io   *io = NULL;
1852         struct cl_page *page;
1853         struct ll_sb_info *sbi = ll_i2sbi(inode);
1854         int result;
1855         ENTRY;
1856
1857         lcc = ll_cl_find(inode);
1858         if (lcc != NULL) {
1859                 env = lcc->lcc_env;
1860                 io  = lcc->lcc_io;
1861         }
1862
1863         if (io == NULL) { /* fast read */
1864                 struct inode *inode = file_inode(file);
1865                 struct ll_file_data *fd = file->private_data;
1866                 struct ll_readahead_state *ras = &fd->fd_ras;
1867                 struct lu_env  *local_env = NULL;
1868                 struct vvp_page *vpg;
1869
1870                 result = -ENODATA;
1871
1872                 /* TODO: need to verify the layout version to make sure
1873                  * the page is not invalid due to layout change. */
1874                 page = cl_vmpage_page(vmpage, clob);
1875                 if (page == NULL) {
1876                         unlock_page(vmpage);
1877                         ll_ra_stats_inc_sbi(sbi, RA_STAT_FAILED_FAST_READ);
1878                         RETURN(result);
1879                 }
1880
1881                 vpg = cl2vvp_page(cl_object_page_slice(page->cp_obj, page));
1882                 if (page->cp_defer_uptodate) {
1883                         enum ras_update_flags flags = LL_RAS_HIT;
1884
1885                         if (lcc && lcc->lcc_type == LCC_MMAP)
1886                                 flags |= LL_RAS_MMAP;
1887
1888                         /* For fast read, it updates read ahead state only
1889                          * if the page is hit in cache because non cache page
1890                          * case will be handled by slow read later. */
1891                         ras_update(sbi, inode, ras, vvp_index(vpg), flags, io);
1892                         /* avoid duplicate ras_update() call */
1893                         page->cp_ra_updated = 1;
1894
1895                         if (ll_use_fast_io(file, ras, vvp_index(vpg)))
1896                                 result = 0;
1897                 }
1898
1899                 if (!env) {
1900                         local_env = cl_env_percpu_get();
1901                         env = local_env;
1902                 }
1903
1904                 /* export the page and skip io stack */
1905                 if (result == 0) {
1906                         page->cp_ra_used = 1;
1907                         SetPageUptodate(vmpage);
1908                 } else {
1909                         ll_ra_stats_inc_sbi(sbi, RA_STAT_FAILED_FAST_READ);
1910                 }
1911
1912                 /* release page refcount before unlocking the page to ensure
1913                  * the object won't be destroyed in the calling path of
1914                  * cl_page_put(). Please see comment in ll_releasepage(). */
1915                 cl_page_put(env, page);
1916                 unlock_page(vmpage);
1917                 if (local_env)
1918                         cl_env_percpu_put(local_env);
1919
1920                 RETURN(result);
1921         }
1922
1923         /**
1924          * Direct read can fall back to buffered read, but DIO is done
1925          * with lockless i/o, and buffered requires LDLM locking, so in
1926          * this case we must restart without lockless.
1927          */
1928         if (file->f_flags & O_DIRECT &&
1929             lcc && lcc->lcc_type == LCC_RW &&
1930             !io->ci_dio_lock) {
1931                 unlock_page(vmpage);
1932                 io->ci_dio_lock = 1;
1933                 io->ci_need_restart = 1;
1934                 RETURN(-ENOLCK);
1935         }
1936
1937         LASSERT(io->ci_state == CIS_IO_GOING);
1938         page = cl_page_find(env, clob, vmpage->index, vmpage, CPT_CACHEABLE);
1939         if (!IS_ERR(page)) {
1940                 LASSERT(page->cp_type == CPT_CACHEABLE);
1941                 if (likely(!PageUptodate(vmpage))) {
1942                         cl_page_assume(env, io, page);
1943
1944                         result = ll_io_read_page(env, io, page, file);
1945                 } else {
1946                         /* Page from a non-object file. */
1947                         unlock_page(vmpage);
1948                         result = 0;
1949                 }
1950                 cl_page_put(env, page);
1951         } else {
1952                 unlock_page(vmpage);
1953                 result = PTR_ERR(page);
1954         }
1955         RETURN(result);
1956 }