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  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * lustre/llite/rw.c
  33  *
  34  * Lustre Lite I/O page cache routines shared by different kernel revs
  35  */
  36
  37 #include <linux/kernel.h>
  38 #include <linux/mm.h>
  39 #include <linux/string.h>
  40 #include <linux/stat.h>
  41 #include <linux/errno.h>
  42 #include <linux/unistd.h>
  43 #include <linux/writeback.h>
  44 #include <asm/uaccess.h>
  45
  46 #include <linux/fs.h>
  47 #include <linux/file.h>
  48 #include <linux/stat.h>
  49 #include <asm/uaccess.h>
  50 #include <linux/mm.h>
  51 #include <linux/pagemap.h>
  52 /* current_is_kswapd() */
  53 #include <linux/swap.h>
  54 #include <linux/task_io_accounting_ops.h>
  55
  56 #define DEBUG_SUBSYSTEM S_LLITE
  57
  58 #include <obd_cksum.h>
  59 #include "llite_internal.h"
  60 #include <lustre_compat.h>
  61
  62 static void ll_ra_stats_inc_sbi(struct ll_sb_info *sbi, enum ra_stat which);
  63
  64 /**
  65  * Get readahead pages from the filesystem readahead pool of the client for a
  66  * thread.
  67  *
  68  * /param sbi superblock for filesystem readahead state ll_ra_info
  69  * /param ria per-thread readahead state
  70  * /param pages number of pages requested for readahead for the thread.
  71  *
  72  * WARNING: This algorithm is used to reduce contention on sbi->ll_lock.
  73  * It should work well if the ra_max_pages is much greater than the single
  74  * file's read-ahead window, and not too many threads contending for
  75  * these readahead pages.
  76  *
  77  * TODO: There may be a 'global sync problem' if many threads are trying
  78  * to get an ra budget that is larger than the remaining readahead pages
  79  * and reach here at exactly the same time. They will compute /a ret to
  80  * consume the remaining pages, but will fail at atomic_add_return() and
  81  * get a zero ra window, although there is still ra space remaining. - Jay */
  82
  83 static unsigned long ll_ra_count_get(struct ll_sb_info *sbi,
  84                                      struct ra_io_arg *ria,
  85                                      unsigned long pages,
  86                                      unsigned long pages_min)
  87 {
  88         struct ll_ra_info *ra = &sbi->ll_ra_info;
  89         long ret;
  90         ENTRY;
  91
  92         /* If read-ahead pages left are less than 1M, do not do read-ahead,
  93          * otherwise it will form small read RPC(< 1M), which hurt server
  94          * performance a lot. */
  95         ret = min(ra->ra_max_pages - atomic_read(&ra->ra_cur_pages),
  96                   pages);
  97         if (ret < 0 || ret < min_t(long, PTLRPC_MAX_BRW_PAGES, pages))
  98                 GOTO(out, ret = 0);
  99
 100         if (atomic_add_return(ret, &ra->ra_cur_pages) > ra->ra_max_pages) {
 101                 atomic_sub(ret, &ra->ra_cur_pages);
 102                 ret = 0;
 103         }
 104
 105 out:
 106         if (ret < pages_min) {
 107                 /* override ra limit for maximum performance */
 108                 atomic_add(pages_min - ret, &ra->ra_cur_pages);
 109                 ret = pages_min;
 110         }
 111         RETURN(ret);
 112 }
 113
 114 void ll_ra_count_put(struct ll_sb_info *sbi, unsigned long pages)
 115 {
 116         struct ll_ra_info *ra = &sbi->ll_ra_info;
 117         atomic_sub(pages, &ra->ra_cur_pages);
 118 }
 119
 120 static void ll_ra_stats_inc_sbi(struct ll_sb_info *sbi, enum ra_stat which)
 121 {
 122         LASSERTF(which < _NR_RA_STAT, "which: %u\n", which);
 123         lprocfs_counter_incr(sbi->ll_ra_stats, which);
 124 }
 125
 126 static inline bool ll_readahead_enabled(struct ll_sb_info *sbi)
 127 {
 128         return sbi->ll_ra_info.ra_max_pages_per_file > 0 &&
 129                 sbi->ll_ra_info.ra_max_pages > 0;
 130 }
 131
 132 void ll_ra_stats_inc(struct inode *inode, enum ra_stat which)
 133 {
 134         struct ll_sb_info *sbi = ll_i2sbi(inode);
 135         ll_ra_stats_inc_sbi(sbi, which);
 136 }
 137
 138 #define RAS_CDEBUG(ras) \
 139         CDEBUG(D_READA,                                                      \
 140                "lre %llu cr %lu cb %llu wsi %lu wp %lu nra %lu rpc %lu "     \
 141                "r %lu csr %lu so %llu sb %llu sl %llu lr %lu\n",             \
 142                ras->ras_last_read_end_bytes, ras->ras_consecutive_requests,  \
 143                ras->ras_consecutive_bytes, ras->ras_window_start_idx,        \
 144                ras->ras_window_pages, ras->ras_next_readahead_idx,           \
 145                ras->ras_rpc_pages, ras->ras_requests,                        \
 146                ras->ras_consecutive_stride_requests, ras->ras_stride_offset, \
 147                ras->ras_stride_bytes, ras->ras_stride_length,                \
 148                ras->ras_async_last_readpage_idx)
 149
 150 static bool pos_in_window(loff_t pos, loff_t point,
 151                           unsigned long before, unsigned long after)
 152 {
 153         loff_t start = point - before;
 154         loff_t end = point + after;
 155
 156         if (start > point)
 157                 start = 0;
 158         if (end < point)
 159                 end = ~0;
 160
 161         return start <= pos && pos <= end;
 162 }
 163
 164 enum ll_ra_page_hint {
 165         MAYNEED = 0, /* this page possibly accessed soon */
 166         WILLNEED /* this page is gurateed to be needed */
 167 };
 168
 169 /**
 170  * Initiates read-ahead of a page with given index.
 171  *
 172  * \retval +ve: page was already uptodate so it will be skipped
 173  *              from being added;
 174  * \retval -ve: page wasn't added to \a queue for error;
 175  * \retval   0: page was added into \a queue for read ahead.
 176  */
 177 static int ll_read_ahead_page(const struct lu_env *env, struct cl_io *io,
 178                               struct cl_page_list *queue, pgoff_t index,
 179                               enum ll_ra_page_hint hint)
 180 {
 181         struct cl_object *clob  = io->ci_obj;
 182         struct inode     *inode = vvp_object_inode(clob);
 183         struct page      *vmpage = NULL;
 184         struct cl_page   *page;
 185         struct vvp_page  *vpg;
 186         enum ra_stat      which = _NR_RA_STAT; /* keep gcc happy */
 187         int               rc    = 0;
 188         const char       *msg   = NULL;
 189
 190         ENTRY;
 191
 192         switch (hint) {
 193         case MAYNEED:
 194                 vmpage = grab_cache_page_nowait(inode->i_mapping, index);
 195                 if (vmpage == NULL) {
 196                         which = RA_STAT_FAILED_GRAB_PAGE;
 197                         msg   = "g_c_p_n failed";
 198                         GOTO(out, rc = -EBUSY);
 199                 }
 200                 break;
 201         case WILLNEED:
 202                 vmpage = find_or_create_page(inode->i_mapping, index,
 203                                              GFP_NOFS);
 204                 if (vmpage == NULL)
 205                         GOTO(out, rc = -ENOMEM);
 206                 break;
 207         default:
 208                 /* should not come here */
 209                 GOTO(out, rc = -EINVAL);
 210         }
 211
 212         /* Check if vmpage was truncated or reclaimed */
 213         if (vmpage->mapping != inode->i_mapping) {
 214                 which = RA_STAT_WRONG_GRAB_PAGE;
 215                 msg   = "g_c_p_n returned invalid page";
 216                 GOTO(out, rc = -EBUSY);
 217         }
 218
 219         page = cl_page_find(env, clob, vmpage->index, vmpage, CPT_CACHEABLE);
 220         if (IS_ERR(page)) {
 221                 which = RA_STAT_FAILED_GRAB_PAGE;
 222                 msg   = "cl_page_find failed";
 223                 GOTO(out, rc = PTR_ERR(page));
 224         }
 225
 226         lu_ref_add(&page->cp_reference, "ra", current);
 227         cl_page_assume(env, io, page);
 228         vpg = cl2vvp_page(cl_object_page_slice(clob, page));
 229         if (!vpg->vpg_defer_uptodate && !PageUptodate(vmpage)) {
 230                 vpg->vpg_defer_uptodate = 1;
 231                 vpg->vpg_ra_used = 0;
 232                 cl_page_list_add(queue, page);
 233         } else {
 234                 /* skip completed pages */
 235                 cl_page_unassume(env, io, page);
 236                 /* This page is already uptodate, returning a positive number
 237                  * to tell the callers about this */
 238                 rc = 1;
 239         }
 240
 241         lu_ref_del(&page->cp_reference, "ra", current);
 242         cl_page_put(env, page);
 243
 244 out:
 245         if (vmpage != NULL) {
 246                 if (rc != 0)
 247                         unlock_page(vmpage);
 248                 put_page(vmpage);
 249         }
 250         if (msg != NULL && hint == MAYNEED) {
 251                 ll_ra_stats_inc(inode, which);
 252                 CDEBUG(D_READA, "%s\n", msg);
 253
 254         }
 255
 256         RETURN(rc);
 257 }
 258
 259 #define RIA_DEBUG(ria)                                                  \
 260         CDEBUG(D_READA, "rs %lu re %lu ro %llu rl %llu rb %llu\n",      \
 261                ria->ria_start_idx, ria->ria_end_idx, ria->ria_stoff,    \
 262                ria->ria_length, ria->ria_bytes)
 263
 264 static inline int stride_io_mode(struct ll_readahead_state *ras)
 265 {
 266         return ras->ras_consecutive_stride_requests > 1;
 267 }
 268
 269 /* The function calculates how many bytes will be read in
 270  * [off, off + length], in such stride IO area,
 271  * stride_offset = st_off, stride_lengh = st_len,
 272  * stride_bytes = st_bytes
 273  *
 274  *   |------------------|*****|------------------|*****|------------|*****|....
 275  * st_off
 276  *   |--- st_bytes     ---|
 277  *   |-----     st_len   -----|
 278  *
 279  *              How many bytes it should read in such pattern
 280  *              |-------------------------------------------------------------|
 281  *              off
 282  *              |<------                  length                      ------->|
 283  *
 284  *          =   |<----->|  +  |-------------------------------------| +   |---|
 285  *             start_left                 st_bytes * i                 end_left
 286  */
 287 static loff_t stride_byte_count(loff_t st_off, loff_t st_len, loff_t st_bytes,
 288                                 loff_t off, loff_t length)
 289 {
 290         u64 start = off > st_off ? off - st_off : 0;
 291         u64 end = off + length > st_off ? off + length - st_off : 0;
 292         u64 start_left;
 293         u64 end_left;
 294         u64 bytes_count;
 295
 296         if (st_len == 0 || length == 0 || end == 0)
 297                 return length;
 298
 299         start = div64_u64_rem(start, st_len, &start_left);
 300         if (start_left < st_bytes)
 301                 start_left = st_bytes - start_left;
 302         else
 303                 start_left = 0;
 304
 305         end = div64_u64_rem(end, st_len, &end_left);
 306         if (end_left > st_bytes)
 307                 end_left = st_bytes;
 308
 309         CDEBUG(D_READA, "start %llu, end %llu start_left %llu end_left %llu\n",
 310                start, end, start_left, end_left);
 311
 312         if (start == end)
 313                 bytes_count = end_left - (st_bytes - start_left);
 314         else
 315                 bytes_count = start_left +
 316                         st_bytes * (end - start - 1) + end_left;
 317
 318         CDEBUG(D_READA,
 319                "st_off %llu, st_len %llu st_bytes %llu off %llu length %llu bytescount %llu\n",
 320                st_off, st_len, st_bytes, off, length, bytes_count);
 321
 322         return bytes_count;
 323 }
 324
 325 static unsigned long ria_page_count(struct ra_io_arg *ria)
 326 {
 327         loff_t length_bytes = ria->ria_end_idx >= ria->ria_start_idx ?
 328                 (loff_t)(ria->ria_end_idx -
 329                          ria->ria_start_idx + 1) << PAGE_SHIFT : 0;
 330         loff_t bytes_count;
 331
 332         if (ria->ria_length > ria->ria_bytes && ria->ria_bytes &&
 333             (ria->ria_length & ~PAGE_MASK || ria->ria_bytes & ~PAGE_MASK ||
 334              ria->ria_stoff & ~PAGE_MASK)) {
 335                 /* Over-estimate un-aligned page stride read */
 336                 unsigned long pg_count = ((ria->ria_bytes +
 337                                            PAGE_SIZE - 1) >> PAGE_SHIFT) + 1;
 338                 pg_count *= length_bytes / ria->ria_length + 1;
 339
 340                 return pg_count;
 341         }
 342         bytes_count = stride_byte_count(ria->ria_stoff, ria->ria_length,
 343                                         ria->ria_bytes,
 344                                         (loff_t)ria->ria_start_idx<<PAGE_SHIFT,
 345                                         length_bytes);
 346         return (bytes_count + PAGE_SIZE - 1) >> PAGE_SHIFT;
 347 }
 348
 349 static pgoff_t ras_align(struct ll_readahead_state *ras, pgoff_t index)
 350 {
 351         return index - (index % ras->ras_rpc_pages);
 352 }
 353
 354 /* Check whether the index is in the defined ra-window */
 355 static bool ras_inside_ra_window(pgoff_t idx, struct ra_io_arg *ria)
 356 {
 357         loff_t pos = (loff_t)idx << PAGE_SHIFT;
 358
 359         /* If ria_length == ria_bytes, it means non-stride I/O mode,
 360          * idx should always inside read-ahead window in this case
 361          * For stride I/O mode, just check whether the idx is inside
 362          * the ria_bytes.
 363          */
 364         if (ria->ria_length == 0 || ria->ria_length == ria->ria_bytes)
 365                 return true;
 366
 367         if (pos >= ria->ria_stoff) {
 368                 u64 offset;
 369
 370                 div64_u64_rem(pos - ria->ria_stoff, ria->ria_length, &offset);
 371
 372                 if (offset < ria->ria_bytes ||
 373                     (ria->ria_length - offset) < PAGE_SIZE)
 374                         return true;
 375         } else if (pos + PAGE_SIZE > ria->ria_stoff) {
 376                 return true;
 377         }
 378
 379         return false;
 380 }
 381
 382 static unsigned long
 383 ll_read_ahead_pages(const struct lu_env *env, struct cl_io *io,
 384                     struct cl_page_list *queue, struct ll_readahead_state *ras,
 385                     struct ra_io_arg *ria, pgoff_t *ra_end)
 386 {
 387         struct cl_read_ahead ra = { 0 };
 388         /* busy page count is per stride */
 389         int rc = 0, count = 0, busy_page_count = 0;
 390         pgoff_t page_idx;
 391
 392         LASSERT(ria != NULL);
 393         RIA_DEBUG(ria);
 394
 395         for (page_idx = ria->ria_start_idx;
 396              page_idx <= ria->ria_end_idx && ria->ria_reserved > 0;
 397              page_idx++) {
 398                 if (ras_inside_ra_window(page_idx, ria)) {
 399                         if (ra.cra_end_idx == 0 || ra.cra_end_idx < page_idx) {
 400                                 pgoff_t end_idx;
 401
 402                                 /*
 403                                  * Do not shrink ria_end_idx at any case until
 404                                  * the minimum end of current read is covered.
 405                                  *
 406                                  * Do not extend read lock accross stripe if
 407                                  * lock contention detected.
 408                                  */
 409                                 if (ra.cra_contention &&
 410                                     page_idx > ria->ria_end_idx_min) {
 411                                         ria->ria_end_idx = *ra_end;
 412                                         break;
 413                                 }
 414
 415                                 cl_read_ahead_release(env, &ra);
 416
 417                                 rc = cl_io_read_ahead(env, io, page_idx, &ra);
 418                                 if (rc < 0)
 419                                         break;
 420
 421                                  /*
 422                                   * Only shrink ria_end_idx if the matched
 423                                   * LDLM lock doesn't cover more.
 424                                   */
 425                                 if (page_idx > ra.cra_end_idx) {
 426                                         ria->ria_end_idx = ra.cra_end_idx;
 427                                         break;
 428                                 }
 429
 430                                 CDEBUG(D_READA, "idx: %lu, ra: %lu, rpc: %lu\n",
 431                                        page_idx, ra.cra_end_idx,
 432                                        ra.cra_rpc_pages);
 433                                 LASSERTF(ra.cra_end_idx >= page_idx,
 434                                          "object: %p, indcies %lu / %lu\n",
 435                                          io->ci_obj, ra.cra_end_idx, page_idx);
 436                                 /* update read ahead RPC size.
 437                                  * NB: it's racy but doesn't matter */
 438                                 if (ras->ras_rpc_pages != ra.cra_rpc_pages &&
 439                                     ra.cra_rpc_pages > 0)
 440                                         ras->ras_rpc_pages = ra.cra_rpc_pages;
 441                                 /* trim it to align with optimal RPC size */
 442                                 end_idx = ras_align(ras, ria->ria_end_idx + 1);
 443                                 if (end_idx > 0 && !ria->ria_eof)
 444                                         ria->ria_end_idx = end_idx - 1;
 445                                 if (ria->ria_end_idx < ria->ria_end_idx_min)
 446                                         ria->ria_end_idx = ria->ria_end_idx_min;
 447                         }
 448                         if (page_idx > ria->ria_end_idx)
 449                                 break;
 450
 451                         /* If the page is inside the read-ahead window */
 452                         rc = ll_read_ahead_page(env, io, queue, page_idx,
 453                                                 MAYNEED);
 454                         if (rc < 0 && rc != -EBUSY)
 455                                 break;
 456                         if (rc == -EBUSY) {
 457                                 busy_page_count++;
 458                                 CDEBUG(D_READA,
 459                                        "skip busy page: %lu\n", page_idx);
 460                                 /* For page unaligned readahead the first
 461                                  * last pages of each region can be read by
 462                                  * another reader on the same node, and so
 463                                  * may be busy. So only stop for > 2 busy
 464                                  * pages. */
 465                                 if (busy_page_count > 2)
 466                                         break;
 467                         }
 468
 469                         *ra_end = page_idx;
 470                         /* Only subtract from reserve & count the page if we
 471                          * really did readahead on that page. */
 472                         if (rc == 0) {
 473                                 ria->ria_reserved--;
 474                                 count++;
 475                         }
 476                 } else if (stride_io_mode(ras)) {
 477                         /* If it is not in the read-ahead window, and it is
 478                          * read-ahead mode, then check whether it should skip
 479                          * the stride gap.
 480                          */
 481                         loff_t pos = (loff_t)page_idx << PAGE_SHIFT;
 482                         u64 offset;
 483
 484                         div64_u64_rem(pos - ria->ria_stoff, ria->ria_length,
 485                                       &offset);
 486                         if (offset >= ria->ria_bytes) {
 487                                 pos += (ria->ria_length - offset);
 488                                 if ((pos >> PAGE_SHIFT) >= page_idx + 1)
 489                                         page_idx = (pos >> PAGE_SHIFT) - 1;
 490                                 busy_page_count = 0;
 491                                 CDEBUG(D_READA,
 492                                        "Stride: jump %llu pages to %lu\n",
 493                                        ria->ria_length - offset, page_idx);
 494                                 continue;
 495                         }
 496                 }
 497         }
 498
 499         cl_read_ahead_release(env, &ra);
 500
 501         return count;
 502 }
 503
 504 static void ll_readahead_work_free(struct ll_readahead_work *work)
 505 {
 506         fput(work->lrw_file);
 507         OBD_FREE_PTR(work);
 508 }
 509
 510 static void ll_readahead_handle_work(struct work_struct *wq);
 511 static void ll_readahead_work_add(struct inode *inode,
 512                                   struct ll_readahead_work *work)
 513 {
 514         INIT_WORK(&work->lrw_readahead_work, ll_readahead_handle_work);
 515         queue_work(ll_i2sbi(inode)->ll_ra_info.ll_readahead_wq,
 516                    &work->lrw_readahead_work);
 517 }
 518
 519 static int ll_readahead_file_kms(const struct lu_env *env,
 520                                 struct cl_io *io, __u64 *kms)
 521 {
 522         struct cl_object *clob;
 523         struct inode *inode;
 524         struct cl_attr *attr = vvp_env_thread_attr(env);
 525         int ret;
 526
 527         clob = io->ci_obj;
 528         inode = vvp_object_inode(clob);
 529
 530         cl_object_attr_lock(clob);
 531         ret = cl_object_attr_get(env, clob, attr);
 532         cl_object_attr_unlock(clob);
 533
 534         if (ret != 0)
 535                 RETURN(ret);
 536
 537         *kms = attr->cat_kms;
 538         return 0;
 539 }
 540
 541 static void ll_readahead_handle_work(struct work_struct *wq)
 542 {
 543         struct ll_readahead_work *work;
 544         struct lu_env *env;
 545         __u16 refcheck;
 546         struct ra_io_arg *ria;
 547         struct inode *inode;
 548         struct ll_file_data *fd;
 549         struct ll_readahead_state *ras;
 550         struct cl_io *io;
 551         struct cl_2queue *queue;
 552         pgoff_t ra_end_idx = 0;
 553         unsigned long pages, pages_min = 0;
 554         struct file *file;
 555         __u64 kms;
 556         int rc;
 557         pgoff_t eof_index;
 558         struct ll_sb_info *sbi;
 559
 560         work = container_of(wq, struct ll_readahead_work,
 561                             lrw_readahead_work);
 562         fd = work->lrw_file->private_data;
 563         ras = &fd->fd_ras;
 564         file = work->lrw_file;
 565         inode = file_inode(file);
 566         sbi = ll_i2sbi(inode);
 567
 568         env = cl_env_alloc(&refcheck, LCT_NOREF);
 569         if (IS_ERR(env))
 570                 GOTO(out_free_work, rc = PTR_ERR(env));
 571
 572         io = vvp_env_thread_io(env);
 573         ll_io_init(io, file, CIT_READ, NULL);
 574
 575         rc = ll_readahead_file_kms(env, io, &kms);
 576         if (rc != 0)
 577                 GOTO(out_put_env, rc);
 578
 579         if (kms == 0) {
 580                 ll_ra_stats_inc(inode, RA_STAT_ZERO_LEN);
 581                 GOTO(out_put_env, rc = 0);
 582         }
 583
 584         ria = &ll_env_info(env)->lti_ria;
 585         memset(ria, 0, sizeof(*ria));
 586
 587         ria->ria_start_idx = work->lrw_start_idx;
 588         /* Truncate RA window to end of file */
 589         eof_index = (pgoff_t)(kms - 1) >> PAGE_SHIFT;
 590         if (eof_index <= work->lrw_end_idx) {
 591                 work->lrw_end_idx = eof_index;
 592                 ria->ria_eof = true;
 593         }
 594         if (work->lrw_end_idx <= work->lrw_start_idx)
 595                 GOTO(out_put_env, rc = 0);
 596
 597         ria->ria_end_idx = work->lrw_end_idx;
 598         pages = ria->ria_end_idx - ria->ria_start_idx + 1;
 599         ria->ria_reserved = ll_ra_count_get(sbi, ria,
 600                                             ria_page_count(ria), pages_min);
 601
 602         CDEBUG(D_READA,
 603                "async reserved pages: %lu/%lu/%lu, ra_cur %d, ra_max %lu\n",
 604                ria->ria_reserved, pages, pages_min,
 605                atomic_read(&ll_i2sbi(inode)->ll_ra_info.ra_cur_pages),
 606                ll_i2sbi(inode)->ll_ra_info.ra_max_pages);
 607
 608         if (ria->ria_reserved < pages) {
 609                 ll_ra_stats_inc(inode, RA_STAT_MAX_IN_FLIGHT);
 610                 if (PAGES_TO_MiB(ria->ria_reserved) < 1) {
 611                         ll_ra_count_put(ll_i2sbi(inode), ria->ria_reserved);
 612                         GOTO(out_put_env, rc = 0);
 613                 }
 614         }
 615
 616         rc = cl_io_rw_init(env, io, CIT_READ, ria->ria_start_idx, pages);
 617         if (rc)
 618                 GOTO(out_put_env, rc);
 619
 620         /* overwrite jobid inited in vvp_io_init() */
 621         if (strncmp(ll_i2info(inode)->lli_jobid, work->lrw_jobid,
 622                     sizeof(work->lrw_jobid)))
 623                 memcpy(ll_i2info(inode)->lli_jobid, work->lrw_jobid,
 624                        sizeof(work->lrw_jobid));
 625
 626         vvp_env_io(env)->vui_io_subtype = IO_NORMAL;
 627         vvp_env_io(env)->vui_fd = fd;
 628         io->ci_state = CIS_LOCKED;
 629         io->ci_async_readahead = true;
 630         rc = cl_io_start(env, io);
 631         if (rc)
 632                 GOTO(out_io_fini, rc);
 633
 634         queue = &io->ci_queue;
 635         cl_2queue_init(queue);
 636
 637         rc = ll_read_ahead_pages(env, io, &queue->c2_qin, ras, ria,
 638                                  &ra_end_idx);
 639         if (ria->ria_reserved != 0)
 640                 ll_ra_count_put(ll_i2sbi(inode), ria->ria_reserved);
 641         if (queue->c2_qin.pl_nr > 0) {
 642                 int count = queue->c2_qin.pl_nr;
 643
 644                 rc = cl_io_submit_rw(env, io, CRT_READ, queue);
 645                 if (rc == 0)
 646                         task_io_account_read(PAGE_SIZE * count);
 647         }
 648         if (ria->ria_end_idx == ra_end_idx && ra_end_idx == (kms >> PAGE_SHIFT))
 649                 ll_ra_stats_inc(inode, RA_STAT_EOF);
 650
 651         if (ra_end_idx != ria->ria_end_idx)
 652                 ll_ra_stats_inc(inode, RA_STAT_FAILED_REACH_END);
 653
 654         /* TODO: discard all pages until page reinit route is implemented */
 655         cl_page_list_discard(env, io, &queue->c2_qin);
 656
 657         /* Unlock unsent read pages in case of error. */
 658         cl_page_list_disown(env, io, &queue->c2_qin);
 659
 660         cl_2queue_fini(env, queue);
 661 out_io_fini:
 662         cl_io_end(env, io);
 663         cl_io_fini(env, io);
 664 out_put_env:
 665         cl_env_put(env, &refcheck);
 666 out_free_work:
 667         if (ra_end_idx > 0)
 668                 ll_ra_stats_inc_sbi(ll_i2sbi(inode), RA_STAT_ASYNC);
 669         atomic_dec(&sbi->ll_ra_info.ra_async_inflight);
 670         ll_readahead_work_free(work);
 671 }
 672
 673 static int ll_readahead(const struct lu_env *env, struct cl_io *io,
 674                         struct cl_page_list *queue,
 675                         struct ll_readahead_state *ras, bool hit,
 676                         struct file *file)
 677 {
 678         struct vvp_io *vio = vvp_env_io(env);
 679         struct ll_thread_info *lti = ll_env_info(env);
 680         unsigned long pages, pages_min = 0;
 681         pgoff_t ra_end_idx = 0, start_idx = 0, end_idx = 0;
 682         struct inode *inode;
 683         struct ra_io_arg *ria = &lti->lti_ria;
 684         struct cl_object *clob;
 685         int ret = 0;
 686         __u64 kms;
 687         ENTRY;
 688
 689         clob = io->ci_obj;
 690         inode = vvp_object_inode(clob);
 691
 692         memset(ria, 0, sizeof(*ria));
 693         ret = ll_readahead_file_kms(env, io, &kms);
 694         if (ret != 0)
 695                 RETURN(ret);
 696
 697         if (kms == 0) {
 698                 ll_ra_stats_inc(inode, RA_STAT_ZERO_LEN);
 699                 RETURN(0);
 700         }
 701
 702         spin_lock(&ras->ras_lock);
 703
 704         /**
 705          * Note: other thread might rollback the ras_next_readahead_idx,
 706          * if it can not get the full size of prepared pages, see the
 707          * end of this function. For stride read ahead, it needs to
 708          * make sure the offset is no less than ras_stride_offset,
 709          * so that stride read ahead can work correctly.
 710          */
 711         if (stride_io_mode(ras))
 712                 start_idx = max_t(pgoff_t, ras->ras_next_readahead_idx,
 713                                   ras->ras_stride_offset >> PAGE_SHIFT);
 714         else
 715                 start_idx = ras->ras_next_readahead_idx;
 716
 717         if (ras->ras_window_pages > 0)
 718                 end_idx = ras->ras_window_start_idx + ras->ras_window_pages - 1;
 719
 720         /* Enlarge the RA window to encompass the full read */
 721         if (vio->vui_ra_valid &&
 722             end_idx < vio->vui_ra_start_idx + vio->vui_ra_pages - 1)
 723                 end_idx = vio->vui_ra_start_idx + vio->vui_ra_pages - 1;
 724
 725         if (end_idx != 0) {
 726                 pgoff_t eof_index;
 727
 728                 /* Truncate RA window to end of file */
 729                 eof_index = (pgoff_t)((kms - 1) >> PAGE_SHIFT);
 730                 if (eof_index <= end_idx) {
 731                         end_idx = eof_index;
 732                         ria->ria_eof = true;
 733                 }
 734         }
 735         ria->ria_start_idx = start_idx;
 736         ria->ria_end_idx = end_idx;
 737         /* If stride I/O mode is detected, get stride window*/
 738         if (stride_io_mode(ras)) {
 739                 ria->ria_stoff = ras->ras_stride_offset;
 740                 ria->ria_length = ras->ras_stride_length;
 741                 ria->ria_bytes = ras->ras_stride_bytes;
 742         }
 743         spin_unlock(&ras->ras_lock);
 744
 745         if (end_idx == 0) {
 746                 ll_ra_stats_inc(inode, RA_STAT_ZERO_WINDOW);
 747                 RETURN(0);
 748         }
 749         pages = ria_page_count(ria);
 750         if (pages == 0) {
 751                 ll_ra_stats_inc(inode, RA_STAT_ZERO_WINDOW);
 752                 RETURN(0);
 753         }
 754
 755         RAS_CDEBUG(ras);
 756         CDEBUG(D_READA, DFID": ria: %lu/%lu, bead: %lu/%lu, hit: %d\n",
 757                PFID(lu_object_fid(&clob->co_lu)),
 758                ria->ria_start_idx, ria->ria_end_idx,
 759                vio->vui_ra_valid ? vio->vui_ra_start_idx : 0,
 760                vio->vui_ra_valid ? vio->vui_ra_pages : 0,
 761                hit);
 762
 763         /* at least to extend the readahead window to cover current read */
 764         if (!hit && vio->vui_ra_valid &&
 765             vio->vui_ra_start_idx + vio->vui_ra_pages > ria->ria_start_idx) {
 766                 ria->ria_end_idx_min =
 767                         vio->vui_ra_start_idx + vio->vui_ra_pages - 1;
 768                 pages_min = vio->vui_ra_start_idx + vio->vui_ra_pages -
 769                                 ria->ria_start_idx;
 770         }
 771
 772         ria->ria_reserved = ll_ra_count_get(ll_i2sbi(inode), ria, pages,
 773                                             pages_min);
 774         if (ria->ria_reserved < pages)
 775                 ll_ra_stats_inc(inode, RA_STAT_MAX_IN_FLIGHT);
 776
 777         CDEBUG(D_READA, "reserved pages: %lu/%lu/%lu, ra_cur %d, ra_max %lu\n",
 778                ria->ria_reserved, pages, pages_min,
 779                atomic_read(&ll_i2sbi(inode)->ll_ra_info.ra_cur_pages),
 780                ll_i2sbi(inode)->ll_ra_info.ra_max_pages);
 781
 782         ret = ll_read_ahead_pages(env, io, queue, ras, ria, &ra_end_idx);
 783
 784         if (ria->ria_reserved != 0)
 785                 ll_ra_count_put(ll_i2sbi(inode), ria->ria_reserved);
 786
 787         if (ra_end_idx == end_idx && ra_end_idx == (kms >> PAGE_SHIFT))
 788                 ll_ra_stats_inc(inode, RA_STAT_EOF);
 789
 790         CDEBUG(D_READA,
 791                "ra_end_idx = %lu end_idx = %lu stride end = %lu pages = %d\n",
 792                ra_end_idx, end_idx, ria->ria_end_idx, ret);
 793
 794         if (ra_end_idx != end_idx)
 795                 ll_ra_stats_inc(inode, RA_STAT_FAILED_REACH_END);
 796         if (ra_end_idx > 0) {
 797                 /* update the ras so that the next read-ahead tries from
 798                  * where we left off. */
 799                 spin_lock(&ras->ras_lock);
 800                 ras->ras_next_readahead_idx = ra_end_idx + 1;
 801                 spin_unlock(&ras->ras_lock);
 802                 RAS_CDEBUG(ras);
 803         }
 804
 805         RETURN(ret);
 806 }
 807
 808 static int ll_readpages(const struct lu_env *env, struct cl_io *io,
 809                         struct cl_page_list *queue,
 810                         pgoff_t start, pgoff_t end)
 811 {
 812         int ret = 0;
 813         __u64 kms;
 814         pgoff_t page_idx;
 815         int count = 0;
 816
 817         ENTRY;
 818
 819         ret = ll_readahead_file_kms(env, io, &kms);
 820         if (ret != 0)
 821                 RETURN(ret);
 822
 823         if (kms == 0)
 824                 RETURN(0);
 825
 826         if (end != 0) {
 827                 unsigned long end_index;
 828
 829                 end_index = (unsigned long)((kms - 1) >> PAGE_SHIFT);
 830                 if (end_index <= end)
 831                         end = end_index;
 832         }
 833
 834         for (page_idx = start; page_idx <= end; page_idx++) {
 835                 ret= ll_read_ahead_page(env, io, queue, page_idx,
 836                                         WILLNEED);
 837                 if (ret < 0)
 838                         break;
 839                 else if (ret == 0) /* ret 1 is already uptodate */
 840                         count++;
 841         }
 842
 843         RETURN(count > 0 ? count : ret);
 844 }
 845
 846 static void ras_set_start(struct ll_readahead_state *ras, pgoff_t index)
 847 {
 848         ras->ras_window_start_idx = ras_align(ras, index);
 849 }
 850
 851 /* called with the ras_lock held or from places where it doesn't matter */
 852 static void ras_reset(struct ll_readahead_state *ras, pgoff_t index)
 853 {
 854         ras->ras_consecutive_requests = 0;
 855         ras->ras_consecutive_bytes = 0;
 856         ras->ras_window_pages = 0;
 857         ras_set_start(ras, index);
 858         ras->ras_next_readahead_idx = max(ras->ras_window_start_idx, index + 1);
 859
 860         RAS_CDEBUG(ras);
 861 }
 862
 863 /* called with the ras_lock held or from places where it doesn't matter */
 864 static void ras_stride_reset(struct ll_readahead_state *ras)
 865 {
 866         ras->ras_consecutive_stride_requests = 0;
 867         ras->ras_stride_length = 0;
 868         ras->ras_stride_bytes = 0;
 869         RAS_CDEBUG(ras);
 870 }
 871
 872 void ll_readahead_init(struct inode *inode, struct ll_readahead_state *ras)
 873 {
 874         spin_lock_init(&ras->ras_lock);
 875         ras->ras_rpc_pages = PTLRPC_MAX_BRW_PAGES;
 876         ras_reset(ras, 0);
 877         ras->ras_last_read_end_bytes = 0;
 878         ras->ras_requests = 0;
 879 }
 880
 881 /*
 882  * Check whether the read request is in the stride window.
 883  * If it is in the stride window, return true, otherwise return false.
 884  */
 885 static bool read_in_stride_window(struct ll_readahead_state *ras,
 886                                   loff_t pos, loff_t count)
 887 {
 888         loff_t stride_gap;
 889
 890         if (ras->ras_stride_length == 0 || ras->ras_stride_bytes == 0 ||
 891             ras->ras_stride_bytes == ras->ras_stride_length)
 892                 return false;
 893
 894         stride_gap = pos - ras->ras_last_read_end_bytes - 1;
 895
 896         /* If it is contiguous read */
 897         if (stride_gap == 0)
 898                 return ras->ras_consecutive_bytes + count <=
 899                         ras->ras_stride_bytes;
 900
 901         /* Otherwise check the stride by itself */
 902         return (ras->ras_stride_length - ras->ras_stride_bytes) == stride_gap &&
 903                 ras->ras_consecutive_bytes == ras->ras_stride_bytes &&
 904                 count <= ras->ras_stride_bytes;
 905 }
 906
 907 static void ras_init_stride_detector(struct ll_readahead_state *ras,
 908                                      loff_t pos, loff_t count)
 909 {
 910         loff_t stride_gap = pos - ras->ras_last_read_end_bytes - 1;
 911
 912         LASSERT(ras->ras_consecutive_stride_requests == 0);
 913
 914         if (pos <= ras->ras_last_read_end_bytes) {
 915                 /*Reset stride window for forward read*/
 916                 ras_stride_reset(ras);
 917                 return;
 918         }
 919
 920         ras->ras_stride_bytes = ras->ras_consecutive_bytes;
 921         ras->ras_stride_length = stride_gap + ras->ras_consecutive_bytes;
 922         ras->ras_consecutive_stride_requests++;
 923         ras->ras_stride_offset = pos;
 924
 925         RAS_CDEBUG(ras);
 926 }
 927
 928 static unsigned long
 929 stride_page_count(struct ll_readahead_state *ras, loff_t len)
 930 {
 931         loff_t bytes_count =
 932                 stride_byte_count(ras->ras_stride_offset,
 933                                   ras->ras_stride_length, ras->ras_stride_bytes,
 934                                   ras->ras_window_start_idx << PAGE_SHIFT, len);
 935
 936         return (bytes_count + PAGE_SIZE - 1) >> PAGE_SHIFT;
 937 }
 938
 939 /* Stride Read-ahead window will be increased inc_len according to
 940  * stride I/O pattern */
 941 static void ras_stride_increase_window(struct ll_readahead_state *ras,
 942                                        struct ll_ra_info *ra, loff_t inc_bytes)
 943 {
 944         loff_t window_bytes, stride_bytes;
 945         u64 left_bytes;
 946         u64 step;
 947         loff_t end;
 948
 949         /* temporarily store in page units to reduce LASSERT() cost below */
 950         end = ras->ras_window_start_idx + ras->ras_window_pages;
 951
 952         LASSERT(ras->ras_stride_length > 0);
 953         LASSERTF(end >= (ras->ras_stride_offset >> PAGE_SHIFT),
 954                  "window_start_idx %lu, window_pages %lu stride_offset %llu\n",
 955                  ras->ras_window_start_idx, ras->ras_window_pages,
 956                  ras->ras_stride_offset);
 957
 958         end <<= PAGE_SHIFT;
 959         if (end <= ras->ras_stride_offset)
 960                 stride_bytes = 0;
 961         else
 962                 stride_bytes = end - ras->ras_stride_offset;
 963
 964         div64_u64_rem(stride_bytes, ras->ras_stride_length, &left_bytes);
 965         window_bytes = (ras->ras_window_pages << PAGE_SHIFT);
 966         if (left_bytes < ras->ras_stride_bytes) {
 967                 if (ras->ras_stride_bytes - left_bytes >= inc_bytes) {
 968                         window_bytes += inc_bytes;
 969                         goto out;
 970                 } else {
 971                         window_bytes += (ras->ras_stride_bytes - left_bytes);
 972                         inc_bytes -= (ras->ras_stride_bytes - left_bytes);
 973                 }
 974         } else {
 975                 window_bytes += (ras->ras_stride_length - left_bytes);
 976         }
 977
 978         LASSERT(ras->ras_stride_bytes != 0);
 979
 980         step = div64_u64_rem(inc_bytes, ras->ras_stride_bytes, &left_bytes);
 981
 982         window_bytes += step * ras->ras_stride_length + left_bytes;
 983         LASSERT(window_bytes > 0);
 984
 985 out:
 986         if (stride_page_count(ras, window_bytes) <=
 987             ra->ra_max_pages_per_file || ras->ras_window_pages == 0)
 988                 ras->ras_window_pages = (window_bytes >> PAGE_SHIFT);
 989
 990         LASSERT(ras->ras_window_pages > 0);
 991
 992         RAS_CDEBUG(ras);
 993 }
 994
 995 static void ras_increase_window(struct inode *inode,
 996                                 struct ll_readahead_state *ras,
 997                                 struct ll_ra_info *ra)
 998 {
 999         /* The stretch of ra-window should be aligned with max rpc_size
1000          * but current clio architecture does not support retrieve such
1001          * information from lower layer. FIXME later
1002          */
1003         if (stride_io_mode(ras)) {
1004                 ras_stride_increase_window(ras, ra,
1005                                       (loff_t)ras->ras_rpc_pages << PAGE_SHIFT);
1006         } else {
1007                 pgoff_t window_pages;
1008
1009                 window_pages = min(ras->ras_window_pages + ras->ras_rpc_pages,
1010                                    ra->ra_max_pages_per_file);
1011                 if (window_pages < ras->ras_rpc_pages)
1012                         ras->ras_window_pages = window_pages;
1013                 else
1014                         ras->ras_window_pages = ras_align(ras, window_pages);
1015         }
1016 }
1017
1018 /**
1019  * Seek within 8 pages are considered as sequential read for now.
1020  */
1021 static inline bool is_loose_seq_read(struct ll_readahead_state *ras, loff_t pos)
1022 {
1023         return pos_in_window(pos, ras->ras_last_read_end_bytes,
1024                              8UL << PAGE_SHIFT, 8UL << PAGE_SHIFT);
1025 }
1026
1027 static void ras_detect_read_pattern(struct ll_readahead_state *ras,
1028                                     struct ll_sb_info *sbi,
1029                                     loff_t pos, size_t count, bool mmap)
1030 {
1031         bool stride_detect = false;
1032         pgoff_t index = pos >> PAGE_SHIFT;
1033
1034         /*
1035          * Reset the read-ahead window in two cases. First when the app seeks
1036          * or reads to some other part of the file. Secondly if we get a
1037          * read-ahead miss that we think we've previously issued. This can
1038          * be a symptom of there being so many read-ahead pages that the VM
1039          * is reclaiming it before we get to it.
1040          */
1041         if (!is_loose_seq_read(ras, pos)) {
1042                 /* Check whether it is in stride I/O mode */
1043                 if (!read_in_stride_window(ras, pos, count)) {
1044                         if (ras->ras_consecutive_stride_requests == 0)
1045                                 ras_init_stride_detector(ras, pos, count);
1046                         else
1047                                 ras_stride_reset(ras);
1048                         ras->ras_consecutive_bytes = 0;
1049                         ras_reset(ras, index);
1050                 } else {
1051                         ras->ras_consecutive_bytes = 0;
1052                         ras->ras_consecutive_requests = 0;
1053                         if (++ras->ras_consecutive_stride_requests > 1)
1054                                 stride_detect = true;
1055                         RAS_CDEBUG(ras);
1056                 }
1057                 ll_ra_stats_inc_sbi(sbi, RA_STAT_DISTANT_READPAGE);
1058         } else if (stride_io_mode(ras)) {
1059                 /*
1060                  * If this is contiguous read but in stride I/O mode
1061                  * currently, check whether stride step still is valid,
1062                  * if invalid, it will reset the stride ra window to
1063                  * be zero.
1064                  */
1065                 if (!read_in_stride_window(ras, pos, count)) {
1066                         ras_stride_reset(ras);
1067                         ras->ras_window_pages = 0;
1068                         ras->ras_next_readahead_idx = index;
1069                 }
1070         }
1071
1072         ras->ras_consecutive_bytes += count;
1073         if (mmap) {
1074                 pgoff_t idx = ras->ras_consecutive_bytes >> PAGE_SHIFT;
1075
1076                 if ((idx >= 4 && (idx & 3UL) == 0) || stride_detect)
1077                         ras->ras_need_increase_window = true;
1078         } else if ((ras->ras_consecutive_requests > 1 || stride_detect)) {
1079                 ras->ras_need_increase_window = true;
1080         }
1081
1082         ras->ras_last_read_end_bytes = pos + count - 1;
1083 }
1084
1085 void ll_ras_enter(struct file *f, loff_t pos, size_t count)
1086 {
1087         struct ll_file_data *fd = f->private_data;
1088         struct ll_readahead_state *ras = &fd->fd_ras;
1089         struct inode *inode = file_inode(f);
1090         unsigned long index = pos >> PAGE_SHIFT;
1091         struct ll_sb_info *sbi = ll_i2sbi(inode);
1092
1093         spin_lock(&ras->ras_lock);
1094         ras->ras_requests++;
1095         ras->ras_consecutive_requests++;
1096         ras->ras_need_increase_window = false;
1097         ras->ras_no_miss_check = false;
1098         /*
1099          * On the second access to a file smaller than the tunable
1100          * ra_max_read_ahead_whole_pages trigger RA on all pages in the
1101          * file up to ra_max_pages_per_file.  This is simply a best effort
1102          * and only occurs once per open file. Normal RA behavior is reverted
1103          * to for subsequent IO.
1104          */
1105         if (ras->ras_requests >= 2) {
1106                 __u64 kms_pages;
1107                 struct ll_ra_info *ra = &sbi->ll_ra_info;
1108
1109                 kms_pages = (i_size_read(inode) + PAGE_SIZE - 1) >>
1110                             PAGE_SHIFT;
1111
1112                 CDEBUG(D_READA, "kmsp %llu mwp %lu mp %lu\n", kms_pages,
1113                        ra->ra_max_read_ahead_whole_pages,
1114                        ra->ra_max_pages_per_file);
1115
1116                 if (kms_pages &&
1117                     kms_pages <= ra->ra_max_read_ahead_whole_pages) {
1118                         ras->ras_window_start_idx = 0;
1119                         ras->ras_next_readahead_idx = index + 1;
1120                         ras->ras_window_pages = min(ra->ra_max_pages_per_file,
1121                                             ra->ra_max_read_ahead_whole_pages);
1122                         ras->ras_no_miss_check = true;
1123                         GOTO(out_unlock, 0);
1124                 }
1125         }
1126         ras_detect_read_pattern(ras, sbi, pos, count, false);
1127 out_unlock:
1128         spin_unlock(&ras->ras_lock);
1129 }
1130
1131 static bool index_in_stride_window(struct ll_readahead_state *ras,
1132                                    pgoff_t index)
1133 {
1134         loff_t pos = (loff_t)index << PAGE_SHIFT;
1135
1136         if (ras->ras_stride_length == 0 || ras->ras_stride_bytes == 0 ||
1137             ras->ras_stride_bytes == ras->ras_stride_length)
1138                 return false;
1139
1140         if (pos >= ras->ras_stride_offset) {
1141                 u64 offset;
1142
1143                 div64_u64_rem(pos - ras->ras_stride_offset,
1144                               ras->ras_stride_length, &offset);
1145                 if (offset < ras->ras_stride_bytes ||
1146                     ras->ras_stride_length - offset < PAGE_SIZE)
1147                         return true;
1148         } else if (ras->ras_stride_offset - pos < PAGE_SIZE) {
1149                 return true;
1150         }
1151
1152         return false;
1153 }
1154
1155 /*
1156  * ll_ras_enter() is used to detect read pattern according to pos and count.
1157  *
1158  * ras_update() is used to detect cache miss and
1159  * reset window or increase window accordingly
1160  */
1161 static void ras_update(struct ll_sb_info *sbi, struct inode *inode,
1162                        struct ll_readahead_state *ras, pgoff_t index,
1163                        enum ras_update_flags flags)
1164 {
1165         struct ll_ra_info *ra = &sbi->ll_ra_info;
1166         bool hit = flags & LL_RAS_HIT;
1167
1168         ENTRY;
1169         spin_lock(&ras->ras_lock);
1170
1171         if (!hit)
1172                 CDEBUG(D_READA, DFID " pages at %lu miss.\n",
1173                        PFID(ll_inode2fid(inode)), index);
1174         ll_ra_stats_inc_sbi(sbi, hit ? RA_STAT_HIT : RA_STAT_MISS);
1175
1176         /*
1177          * The readahead window has been expanded to cover whole
1178          * file size, we don't care whether ra miss happen or not.
1179          * Because we will read whole file to page cache even if
1180          * some pages missed.
1181          */
1182         if (ras->ras_no_miss_check)
1183                 GOTO(out_unlock, 0);
1184
1185         if (flags & LL_RAS_MMAP)
1186                 ras_detect_read_pattern(ras, sbi, (loff_t)index << PAGE_SHIFT,
1187                                         PAGE_SIZE, true);
1188
1189         if (!hit && ras->ras_window_pages &&
1190             index < ras->ras_next_readahead_idx &&
1191             pos_in_window(index, ras->ras_window_start_idx, 0,
1192                           ras->ras_window_pages)) {
1193                 ll_ra_stats_inc_sbi(sbi, RA_STAT_MISS_IN_WINDOW);
1194                 ras->ras_need_increase_window = false;
1195
1196                 if (index_in_stride_window(ras, index) &&
1197                     stride_io_mode(ras)) {
1198                         /*
1199                          * if (index != ras->ras_last_readpage + 1)
1200                          *      ras->ras_consecutive_pages = 0;
1201                          */
1202                         ras_reset(ras, index);
1203
1204                         /*
1205                          * If stride-RA hit cache miss, the stride
1206                          * detector will not be reset to avoid the
1207                          * overhead of redetecting read-ahead mode,
1208                          * but on the condition that the stride window
1209                          * is still intersect with normal sequential
1210                          * read-ahead window.
1211                          */
1212                         if (ras->ras_window_start_idx < ras->ras_stride_offset)
1213                                 ras_stride_reset(ras);
1214                         RAS_CDEBUG(ras);
1215                 } else {
1216                         /*
1217                          * Reset both stride window and normal RA
1218                          * window.
1219                          */
1220                         ras_reset(ras, index);
1221                         /* ras->ras_consecutive_pages++; */
1222                         ras->ras_consecutive_bytes = 0;
1223                         ras_stride_reset(ras);
1224                         GOTO(out_unlock, 0);
1225                 }
1226         }
1227         ras_set_start(ras, index);
1228
1229         if (stride_io_mode(ras)) {
1230                 /* Since stride readahead is sentivite to the offset
1231                  * of read-ahead, so we use original offset here,
1232                  * instead of ras_window_start_idx, which is RPC aligned.
1233                  */
1234                 ras->ras_next_readahead_idx = max(index + 1,
1235                                                   ras->ras_next_readahead_idx);
1236                 ras->ras_window_start_idx =
1237                                 max_t(pgoff_t, ras->ras_window_start_idx,
1238                                       ras->ras_stride_offset >> PAGE_SHIFT);
1239         } else {
1240                 if (ras->ras_next_readahead_idx < ras->ras_window_start_idx)
1241                         ras->ras_next_readahead_idx = ras->ras_window_start_idx;
1242                 if (!hit)
1243                         ras->ras_next_readahead_idx = index + 1;
1244         }
1245
1246         if (ras->ras_need_increase_window) {
1247                 ras_increase_window(inode, ras, ra);
1248                 ras->ras_need_increase_window = false;
1249         }
1250
1251         EXIT;
1252 out_unlock:
1253         spin_unlock(&ras->ras_lock);
1254 }
1255
1256 int ll_writepage(struct page *vmpage, struct writeback_control *wbc)
1257 {
1258         struct inode           *inode = vmpage->mapping->host;
1259         struct ll_inode_info   *lli   = ll_i2info(inode);
1260         struct lu_env          *env;
1261         struct cl_io           *io;
1262         struct cl_page         *page;
1263         struct cl_object       *clob;
1264         bool redirtied = false;
1265         bool unlocked = false;
1266         int result;
1267         __u16 refcheck;
1268         ENTRY;
1269
1270         LASSERT(PageLocked(vmpage));
1271         LASSERT(!PageWriteback(vmpage));
1272
1273         LASSERT(ll_i2dtexp(inode) != NULL);
1274
1275         env = cl_env_get(&refcheck);
1276         if (IS_ERR(env))
1277                 GOTO(out, result = PTR_ERR(env));
1278
1279         clob  = ll_i2info(inode)->lli_clob;
1280         LASSERT(clob != NULL);
1281
1282         io = vvp_env_thread_io(env);
1283         io->ci_obj = clob;
1284         io->ci_ignore_layout = 1;
1285         result = cl_io_init(env, io, CIT_MISC, clob);
1286         if (result == 0) {
1287                 page = cl_page_find(env, clob, vmpage->index,
1288                                     vmpage, CPT_CACHEABLE);
1289                 if (!IS_ERR(page)) {
1290                         lu_ref_add(&page->cp_reference, "writepage",
1291                                    current);
1292                         cl_page_assume(env, io, page);
1293                         result = cl_page_flush(env, io, page);
1294                         if (result != 0) {
1295                                 /*
1296                                  * Re-dirty page on error so it retries write,
1297                                  * but not in case when IO has actually
1298                                  * occurred and completed with an error.
1299                                  */
1300                                 if (!PageError(vmpage)) {
1301                                         redirty_page_for_writepage(wbc, vmpage);
1302                                         result = 0;
1303                                         redirtied = true;
1304                                 }
1305                         }
1306                         cl_page_disown(env, io, page);
1307                         unlocked = true;
1308                         lu_ref_del(&page->cp_reference,
1309                                    "writepage", current);
1310                         cl_page_put(env, page);
1311                 } else {
1312                         result = PTR_ERR(page);
1313                 }
1314         }
1315         cl_io_fini(env, io);
1316
1317         if (redirtied && wbc->sync_mode == WB_SYNC_ALL) {
1318                 loff_t offset = cl_offset(clob, vmpage->index);
1319
1320                 /* Flush page failed because the extent is being written out.
1321                  * Wait for the write of extent to be finished to avoid
1322                  * breaking kernel which assumes ->writepage should mark
1323                  * PageWriteback or clean the page. */
1324                 result = cl_sync_file_range(inode, offset,
1325                                             offset + PAGE_SIZE - 1,
1326                                             CL_FSYNC_LOCAL, 1);
1327                 if (result > 0) {
1328                         /* actually we may have written more than one page.
1329                          * decreasing this page because the caller will count
1330                          * it. */
1331                         wbc->nr_to_write -= result - 1;
1332                         result = 0;
1333                 }
1334         }
1335
1336         cl_env_put(env, &refcheck);
1337         GOTO(out, result);
1338
1339 out:
1340         if (result < 0) {
1341                 if (!lli->lli_async_rc)
1342                         lli->lli_async_rc = result;
1343                 SetPageError(vmpage);
1344                 if (!unlocked)
1345                         unlock_page(vmpage);
1346         }
1347         return result;
1348 }
1349
1350 int ll_writepages(struct address_space *mapping, struct writeback_control *wbc)
1351 {
1352         struct inode *inode = mapping->host;
1353         loff_t start;
1354         loff_t end;
1355         enum cl_fsync_mode mode;
1356         int range_whole = 0;
1357         int result;
1358         ENTRY;
1359
1360         if (wbc->range_cyclic) {
1361                 start = (loff_t)mapping->writeback_index << PAGE_SHIFT;
1362                 end = OBD_OBJECT_EOF;
1363         } else {
1364                 start = wbc->range_start;
1365                 end = wbc->range_end;
1366                 if (end == LLONG_MAX) {
1367                         end = OBD_OBJECT_EOF;
1368                         range_whole = start == 0;
1369                 }
1370         }
1371
1372         mode = CL_FSYNC_NONE;
1373         if (wbc->sync_mode == WB_SYNC_ALL)
1374                 mode = CL_FSYNC_LOCAL;
1375
1376         if (ll_i2info(inode)->lli_clob == NULL)
1377                 RETURN(0);
1378
1379         /* for directio, it would call writepages() to evict cached pages
1380          * inside the IO context of write, which will cause deadlock at
1381          * layout_conf since it waits for active IOs to complete. */
1382         result = cl_sync_file_range(inode, start, end, mode, 1);
1383         if (result > 0) {
1384                 wbc->nr_to_write -= result;
1385                 result = 0;
1386          }
1387
1388         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0)) {
1389                 if (end == OBD_OBJECT_EOF)
1390                         mapping->writeback_index = 0;
1391                 else
1392                         mapping->writeback_index = (end >> PAGE_SHIFT) + 1;
1393         }
1394         RETURN(result);
1395 }
1396
1397 struct ll_cl_context *ll_cl_find(struct file *file)
1398 {
1399         struct ll_file_data *fd = file->private_data;
1400         struct ll_cl_context *lcc;
1401         struct ll_cl_context *found = NULL;
1402
1403         read_lock(&fd->fd_lock);
1404         list_for_each_entry(lcc, &fd->fd_lccs, lcc_list) {
1405                 if (lcc->lcc_cookie == current) {
1406                         found = lcc;
1407                         break;
1408                 }
1409         }
1410         read_unlock(&fd->fd_lock);
1411
1412         return found;
1413 }
1414
1415 void ll_cl_add(struct file *file, const struct lu_env *env, struct cl_io *io,
1416                enum lcc_type type)
1417 {
1418         struct ll_file_data *fd = file->private_data;
1419         struct ll_cl_context *lcc = &ll_env_info(env)->lti_io_ctx;
1420
1421         memset(lcc, 0, sizeof(*lcc));
1422         INIT_LIST_HEAD(&lcc->lcc_list);
1423         lcc->lcc_cookie = current;
1424         lcc->lcc_env = env;
1425         lcc->lcc_io = io;
1426         lcc->lcc_type = type;
1427
1428         write_lock(&fd->fd_lock);
1429         list_add(&lcc->lcc_list, &fd->fd_lccs);
1430         write_unlock(&fd->fd_lock);
1431 }
1432
1433 void ll_cl_remove(struct file *file, const struct lu_env *env)
1434 {
1435         struct ll_file_data *fd = file->private_data;
1436         struct ll_cl_context *lcc = &ll_env_info(env)->lti_io_ctx;
1437
1438         write_lock(&fd->fd_lock);
1439         list_del_init(&lcc->lcc_list);
1440         write_unlock(&fd->fd_lock);
1441 }
1442
1443 int ll_io_read_page(const struct lu_env *env, struct cl_io *io,
1444                            struct cl_page *page, struct file *file)
1445 {
1446         struct inode              *inode  = vvp_object_inode(page->cp_obj);
1447         struct ll_sb_info         *sbi    = ll_i2sbi(inode);
1448         struct ll_file_data       *fd     = file->private_data;
1449         struct ll_readahead_state *ras    = &fd->fd_ras;
1450         struct cl_2queue          *queue  = &io->ci_queue;
1451         struct cl_sync_io         *anchor = NULL;
1452         struct vvp_page           *vpg;
1453         int                        rc = 0, rc2 = 0;
1454         bool                       uptodate;
1455         pgoff_t io_start_index;
1456         pgoff_t io_end_index;
1457         ENTRY;
1458
1459         vpg = cl2vvp_page(cl_object_page_slice(page->cp_obj, page));
1460         uptodate = vpg->vpg_defer_uptodate;
1461
1462         if (ll_readahead_enabled(sbi) && !vpg->vpg_ra_updated) {
1463                 struct vvp_io *vio = vvp_env_io(env);
1464                 enum ras_update_flags flags = 0;
1465
1466                 if (uptodate)
1467                         flags |= LL_RAS_HIT;
1468                 if (!vio->vui_ra_valid)
1469                         flags |= LL_RAS_MMAP;
1470                 ras_update(sbi, inode, ras, vvp_index(vpg), flags);
1471         }
1472
1473         cl_2queue_init(queue);
1474         if (uptodate) {
1475                 vpg->vpg_ra_used = 1;
1476                 cl_page_export(env, page, 1);
1477                 cl_page_disown(env, io, page);
1478         } else {
1479                 anchor = &vvp_env_info(env)->vti_anchor;
1480                 cl_sync_io_init(anchor, 1);
1481                 page->cp_sync_io = anchor;
1482
1483                 cl_2queue_add(queue, page);
1484         }
1485
1486         io_start_index = cl_index(io->ci_obj, io->u.ci_rw.crw_pos);
1487         io_end_index = cl_index(io->ci_obj, io->u.ci_rw.crw_pos +
1488                                 io->u.ci_rw.crw_count - 1);
1489         if (ll_readahead_enabled(sbi)) {
1490                 rc2 = ll_readahead(env, io, &queue->c2_qin, ras,
1491                                    uptodate, file);
1492                 CDEBUG(D_READA, DFID " %d pages read ahead at %lu\n",
1493                        PFID(ll_inode2fid(inode)), rc2, vvp_index(vpg));
1494         } else if (vvp_index(vpg) == io_start_index &&
1495                    io_end_index - io_start_index > 0) {
1496                 rc2 = ll_readpages(env, io, &queue->c2_qin, io_start_index + 1,
1497                                    io_end_index);
1498                 CDEBUG(D_READA, DFID " %d pages read at %lu\n",
1499                        PFID(ll_inode2fid(inode)), rc2, vvp_index(vpg));
1500         }
1501
1502         if (queue->c2_qin.pl_nr > 0) {
1503                 int count = queue->c2_qin.pl_nr;
1504                 rc = cl_io_submit_rw(env, io, CRT_READ, queue);
1505                 if (rc == 0)
1506                         task_io_account_read(PAGE_SIZE * count);
1507         }
1508
1509
1510         if (anchor != NULL && !cl_page_is_owned(page, io)) { /* have sent */
1511                 rc = cl_sync_io_wait(env, anchor, 0);
1512
1513                 cl_page_assume(env, io, page);
1514                 cl_page_list_del(env, &queue->c2_qout, page);
1515
1516                 if (!PageUptodate(cl_page_vmpage(page))) {
1517                         /* Failed to read a mirror, discard this page so that
1518                          * new page can be created with new mirror.
1519                          *
1520                          * TODO: this is not needed after page reinit
1521                          * route is implemented */
1522                         cl_page_discard(env, io, page);
1523                 }
1524                 cl_page_disown(env, io, page);
1525         }
1526
1527         /* TODO: discard all pages until page reinit route is implemented */
1528         cl_page_list_discard(env, io, &queue->c2_qin);
1529
1530         /* Unlock unsent read pages in case of error. */
1531         cl_page_list_disown(env, io, &queue->c2_qin);
1532
1533         cl_2queue_fini(env, queue);
1534
1535         RETURN(rc);
1536 }
1537
1538 /*
1539  * Possible return value:
1540  * 0 no async readahead triggered and fast read could not be used.
1541  * 1 no async readahead, but fast read could be used.
1542  * 2 async readahead triggered and fast read could be used too.
1543  * < 0 on error.
1544  */
1545 static int kickoff_async_readahead(struct file *file, unsigned long pages)
1546 {
1547         struct ll_readahead_work *lrw;
1548         struct inode *inode = file_inode(file);
1549         struct ll_sb_info *sbi = ll_i2sbi(inode);
1550         struct ll_file_data *fd = file->private_data;
1551         struct ll_readahead_state *ras = &fd->fd_ras;
1552         struct ll_ra_info *ra = &sbi->ll_ra_info;
1553         unsigned long throttle;
1554         pgoff_t start_idx = ras_align(ras, ras->ras_next_readahead_idx);
1555         pgoff_t end_idx = start_idx + pages - 1;
1556
1557         throttle = min(ra->ra_async_pages_per_file_threshold,
1558                        ra->ra_max_pages_per_file);
1559         /*
1560          * If this is strided i/o or the window is smaller than the
1561          * throttle limit, we do not do async readahead. Otherwise,
1562          * we do async readahead, allowing the user thread to do fast i/o.
1563          */
1564         if (stride_io_mode(ras) || !throttle ||
1565             ras->ras_window_pages < throttle ||
1566             atomic_read(&ra->ra_async_inflight) > ra->ra_async_max_active)
1567                 return 0;
1568
1569         if ((atomic_read(&ra->ra_cur_pages) + pages) > ra->ra_max_pages)
1570                 return 0;
1571
1572         if (ras->ras_async_last_readpage_idx == start_idx)
1573                 return 1;
1574
1575         /* ll_readahead_work_free() free it */
1576         OBD_ALLOC_PTR(lrw);
1577         if (lrw) {
1578                 atomic_inc(&sbi->ll_ra_info.ra_async_inflight);
1579                 lrw->lrw_file = get_file(file);
1580                 lrw->lrw_start_idx = start_idx;
1581                 lrw->lrw_end_idx = end_idx;
1582                 spin_lock(&ras->ras_lock);
1583                 ras->ras_next_readahead_idx = end_idx + 1;
1584                 ras->ras_async_last_readpage_idx = start_idx;
1585                 spin_unlock(&ras->ras_lock);
1586                 memcpy(lrw->lrw_jobid, ll_i2info(inode)->lli_jobid,
1587                        sizeof(lrw->lrw_jobid));
1588                 ll_readahead_work_add(inode, lrw);
1589         } else {
1590                 return -ENOMEM;
1591         }
1592
1593         return 2;
1594 }
1595
1596 /*
1597  * Check if we can issue a readahead RPC, if that is
1598  * the case, we can't do fast IO because we will need
1599  * a cl_io to issue the RPC.
1600  */
1601 static bool ll_use_fast_io(struct file *file,
1602                            struct ll_readahead_state *ras, pgoff_t index)
1603 {
1604         unsigned long fast_read_pages =
1605                 max(RA_REMAIN_WINDOW_MIN, ras->ras_rpc_pages);
1606         loff_t skip_pages;
1607         loff_t stride_bytes = ras->ras_stride_bytes;
1608
1609         if (stride_io_mode(ras) && stride_bytes) {
1610                 skip_pages = (ras->ras_stride_length +
1611                         ras->ras_stride_bytes - 1) / stride_bytes;
1612                 skip_pages *= fast_read_pages;
1613         } else {
1614                 skip_pages = fast_read_pages;
1615         }
1616
1617         if (ras->ras_window_start_idx + ras->ras_window_pages <
1618             ras->ras_next_readahead_idx + skip_pages ||
1619             kickoff_async_readahead(file, fast_read_pages) > 0)
1620                 return true;
1621
1622         return false;
1623 }
1624
1625 int ll_readpage(struct file *file, struct page *vmpage)
1626 {
1627         struct inode *inode = file_inode(file);
1628         struct cl_object *clob = ll_i2info(inode)->lli_clob;
1629         struct ll_cl_context *lcc;
1630         const struct lu_env  *env = NULL;
1631         struct cl_io   *io = NULL;
1632         struct cl_page *page;
1633         struct ll_sb_info *sbi = ll_i2sbi(inode);
1634         int result;
1635         ENTRY;
1636
1637         lcc = ll_cl_find(file);
1638         if (lcc != NULL) {
1639                 env = lcc->lcc_env;
1640                 io  = lcc->lcc_io;
1641         }
1642
1643         if (io == NULL) { /* fast read */
1644                 struct inode *inode = file_inode(file);
1645                 struct ll_file_data *fd = file->private_data;
1646                 struct ll_readahead_state *ras = &fd->fd_ras;
1647                 struct lu_env  *local_env = NULL;
1648                 struct vvp_page *vpg;
1649
1650                 result = -ENODATA;
1651
1652                 /* TODO: need to verify the layout version to make sure
1653                  * the page is not invalid due to layout change. */
1654                 page = cl_vmpage_page(vmpage, clob);
1655                 if (page == NULL) {
1656                         unlock_page(vmpage);
1657                         ll_ra_stats_inc_sbi(sbi, RA_STAT_FAILED_FAST_READ);
1658                         RETURN(result);
1659                 }
1660
1661                 vpg = cl2vvp_page(cl_object_page_slice(page->cp_obj, page));
1662                 if (vpg->vpg_defer_uptodate) {
1663                         enum ras_update_flags flags = LL_RAS_HIT;
1664
1665                         if (lcc && lcc->lcc_type == LCC_MMAP)
1666                                 flags |= LL_RAS_MMAP;
1667
1668                         /* For fast read, it updates read ahead state only
1669                          * if the page is hit in cache because non cache page
1670                          * case will be handled by slow read later. */
1671                         ras_update(sbi, inode, ras, vvp_index(vpg), flags);
1672                         /* avoid duplicate ras_update() call */
1673                         vpg->vpg_ra_updated = 1;
1674
1675                         if (ll_use_fast_io(file, ras, vvp_index(vpg)))
1676                                 result = 0;
1677                 }
1678
1679                 if (!env) {
1680                         local_env = cl_env_percpu_get();
1681                         env = local_env;
1682                 }
1683
1684                 /* export the page and skip io stack */
1685                 if (result == 0) {
1686                         vpg->vpg_ra_used = 1;
1687                         cl_page_export(env, page, 1);
1688                 } else {
1689                         ll_ra_stats_inc_sbi(sbi, RA_STAT_FAILED_FAST_READ);
1690                 }
1691                 /* release page refcount before unlocking the page to ensure
1692                  * the object won't be destroyed in the calling path of
1693                  * cl_page_put(). Please see comment in ll_releasepage(). */
1694                 cl_page_put(env, page);
1695                 unlock_page(vmpage);
1696                 if (local_env)
1697                         cl_env_percpu_put(local_env);
1698
1699                 RETURN(result);
1700         }
1701
1702         /**
1703          * Direct read can fall back to buffered read, but DIO is done
1704          * with lockless i/o, and buffered requires LDLM locking, so in
1705          * this case we must restart without lockless.
1706          */
1707         if (file->f_flags & O_DIRECT &&
1708             lcc && lcc->lcc_type == LCC_RW &&
1709             !io->ci_ignore_lockless) {
1710                 unlock_page(vmpage);
1711                 io->ci_ignore_lockless = 1;
1712                 io->ci_need_restart = 1;
1713                 RETURN(-ENOLCK);
1714         }
1715
1716         LASSERT(io->ci_state == CIS_IO_GOING);
1717         page = cl_page_find(env, clob, vmpage->index, vmpage, CPT_CACHEABLE);
1718         if (!IS_ERR(page)) {
1719                 LASSERT(page->cp_type == CPT_CACHEABLE);
1720                 if (likely(!PageUptodate(vmpage))) {
1721                         cl_page_assume(env, io, page);
1722
1723                         result = ll_io_read_page(env, io, page, file);
1724                 } else {
1725                         /* Page from a non-object file. */
1726                         unlock_page(vmpage);
1727                         result = 0;
1728                 }
1729                 cl_page_put(env, page);
1730         } else {
1731                 unlock_page(vmpage);
1732                 result = PTR_ERR(page);
1733         }
1734         RETURN(result);
1735 }