Whamcloud - gitweb
LU-12748 readahead: limit async ra requests
[fs/lustre-release.git] / 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 void ll_ra_stats_inc(struct inode *inode, enum ra_stat which)
127 {
128         struct ll_sb_info *sbi = ll_i2sbi(inode);
129         ll_ra_stats_inc_sbi(sbi, which);
130 }
131
132 #define RAS_CDEBUG(ras) \
133         CDEBUG(D_READA,                                                      \
134                "lre %llu cr %lu cb %llu wsi %lu wp %lu nra %lu rpc %lu "     \
135                "r %lu csr %lu so %llu sb %llu sl %llu lr %lu\n",             \
136                ras->ras_last_read_end_bytes, ras->ras_consecutive_requests,  \
137                ras->ras_consecutive_bytes, ras->ras_window_start_idx,        \
138                ras->ras_window_pages, ras->ras_next_readahead_idx,           \
139                ras->ras_rpc_pages, ras->ras_requests,                        \
140                ras->ras_consecutive_stride_requests, ras->ras_stride_offset, \
141                ras->ras_stride_bytes, ras->ras_stride_length,                \
142                ras->ras_async_last_readpage_idx)
143
144 static bool pos_in_window(loff_t pos, loff_t point,
145                           unsigned long before, unsigned long after)
146 {
147         loff_t start = point - before;
148         loff_t end = point + after;
149
150         if (start > point)
151                 start = 0;
152         if (end < point)
153                 end = ~0;
154
155         return start <= pos && pos <= end;
156 }
157
158 /**
159  * Initiates read-ahead of a page with given index.
160  *
161  * \retval +ve: page was already uptodate so it will be skipped
162  *              from being added;
163  * \retval -ve: page wasn't added to \a queue for error;
164  * \retval   0: page was added into \a queue for read ahead.
165  */
166 static int ll_read_ahead_page(const struct lu_env *env, struct cl_io *io,
167                               struct cl_page_list *queue, pgoff_t index)
168 {
169         struct cl_object *clob  = io->ci_obj;
170         struct inode     *inode = vvp_object_inode(clob);
171         struct page      *vmpage;
172         struct cl_page   *page;
173         struct vvp_page  *vpg;
174         enum ra_stat      which = _NR_RA_STAT; /* keep gcc happy */
175         int               rc    = 0;
176         const char       *msg   = NULL;
177         ENTRY;
178
179         vmpage = grab_cache_page_nowait(inode->i_mapping, index);
180         if (vmpage == NULL) {
181                 which = RA_STAT_FAILED_GRAB_PAGE;
182                 msg   = "g_c_p_n failed";
183                 GOTO(out, rc = -EBUSY);
184         }
185
186         /* Check if vmpage was truncated or reclaimed */
187         if (vmpage->mapping != inode->i_mapping) {
188                 which = RA_STAT_WRONG_GRAB_PAGE;
189                 msg   = "g_c_p_n returned invalid page";
190                 GOTO(out, rc = -EBUSY);
191         }
192
193         page = cl_page_find(env, clob, vmpage->index, vmpage, CPT_CACHEABLE);
194         if (IS_ERR(page)) {
195                 which = RA_STAT_FAILED_GRAB_PAGE;
196                 msg   = "cl_page_find failed";
197                 GOTO(out, rc = PTR_ERR(page));
198         }
199
200         lu_ref_add(&page->cp_reference, "ra", current);
201         cl_page_assume(env, io, page);
202         vpg = cl2vvp_page(cl_object_page_slice(clob, page));
203         if (!vpg->vpg_defer_uptodate && !PageUptodate(vmpage)) {
204                 vpg->vpg_defer_uptodate = 1;
205                 vpg->vpg_ra_used = 0;
206                 cl_page_list_add(queue, page);
207         } else {
208                 /* skip completed pages */
209                 cl_page_unassume(env, io, page);
210                 /* This page is already uptodate, returning a positive number
211                  * to tell the callers about this */
212                 rc = 1;
213         }
214
215         lu_ref_del(&page->cp_reference, "ra", current);
216         cl_page_put(env, page);
217
218 out:
219         if (vmpage != NULL) {
220                 if (rc != 0)
221                         unlock_page(vmpage);
222                 put_page(vmpage);
223         }
224         if (msg != NULL) {
225                 ll_ra_stats_inc(inode, which);
226                 CDEBUG(D_READA, "%s\n", msg);
227
228         }
229
230         RETURN(rc);
231 }
232
233 #define RIA_DEBUG(ria)                                                  \
234         CDEBUG(D_READA, "rs %lu re %lu ro %llu rl %llu rb %llu\n",      \
235                ria->ria_start_idx, ria->ria_end_idx, ria->ria_stoff,    \
236                ria->ria_length, ria->ria_bytes)
237
238 static inline int stride_io_mode(struct ll_readahead_state *ras)
239 {
240         return ras->ras_consecutive_stride_requests > 1;
241 }
242
243 /* The function calculates how many bytes will be read in
244  * [off, off + length], in such stride IO area,
245  * stride_offset = st_off, stride_lengh = st_len,
246  * stride_bytes = st_bytes
247  *
248  *   |------------------|*****|------------------|*****|------------|*****|....
249  * st_off
250  *   |--- st_bytes     ---|
251  *   |-----     st_len   -----|
252  *
253  *              How many bytes it should read in such pattern
254  *              |-------------------------------------------------------------|
255  *              off
256  *              |<------                  length                      ------->|
257  *
258  *          =   |<----->|  +  |-------------------------------------| +   |---|
259  *             start_left                 st_bytes * i                 end_left
260  */
261 static loff_t stride_byte_count(loff_t st_off, loff_t st_len, loff_t st_bytes,
262                                 loff_t off, loff_t length)
263 {
264         u64 start = off > st_off ? off - st_off : 0;
265         u64 end = off + length > st_off ? off + length - st_off : 0;
266         u64 start_left;
267         u64 end_left;
268         u64 bytes_count;
269
270         if (st_len == 0 || length == 0 || end == 0)
271                 return length;
272
273         start = div64_u64_rem(start, st_len, &start_left);
274         if (start_left < st_bytes)
275                 start_left = st_bytes - start_left;
276         else
277                 start_left = 0;
278
279         end = div64_u64_rem(end, st_len, &end_left);
280         if (end_left > st_bytes)
281                 end_left = st_bytes;
282
283         CDEBUG(D_READA, "start %llu, end %llu start_left %llu end_left %llu\n",
284                start, end, start_left, end_left);
285
286         if (start == end)
287                 bytes_count = end_left - (st_bytes - start_left);
288         else
289                 bytes_count = start_left +
290                         st_bytes * (end - start - 1) + end_left;
291
292         CDEBUG(D_READA,
293                "st_off %llu, st_len %llu st_bytes %llu off %llu length %llu bytescount %llu\n",
294                st_off, st_len, st_bytes, off, length, bytes_count);
295
296         return bytes_count;
297 }
298
299 static unsigned long ria_page_count(struct ra_io_arg *ria)
300 {
301         loff_t length_bytes = ria->ria_end_idx >= ria->ria_start_idx ?
302                 (loff_t)(ria->ria_end_idx -
303                          ria->ria_start_idx + 1) << PAGE_SHIFT : 0;
304         loff_t bytes_count;
305
306         if (ria->ria_length > ria->ria_bytes && ria->ria_bytes &&
307             (ria->ria_length & ~PAGE_MASK || ria->ria_bytes & ~PAGE_MASK ||
308              ria->ria_stoff & ~PAGE_MASK)) {
309                 /* Over-estimate un-aligned page stride read */
310                 unsigned long pg_count = ((ria->ria_bytes +
311                                            PAGE_SIZE - 1) >> PAGE_SHIFT) + 1;
312                 pg_count *= length_bytes / ria->ria_length + 1;
313
314                 return pg_count;
315         }
316         bytes_count = stride_byte_count(ria->ria_stoff, ria->ria_length,
317                                         ria->ria_bytes,
318                                         (loff_t)ria->ria_start_idx<<PAGE_SHIFT,
319                                         length_bytes);
320         return (bytes_count + PAGE_SIZE - 1) >> PAGE_SHIFT;
321 }
322
323 static pgoff_t ras_align(struct ll_readahead_state *ras, pgoff_t index)
324 {
325         return index - (index % ras->ras_rpc_pages);
326 }
327
328 /* Check whether the index is in the defined ra-window */
329 static bool ras_inside_ra_window(pgoff_t idx, struct ra_io_arg *ria)
330 {
331         loff_t pos = (loff_t)idx << PAGE_SHIFT;
332
333         /* If ria_length == ria_bytes, it means non-stride I/O mode,
334          * idx should always inside read-ahead window in this case
335          * For stride I/O mode, just check whether the idx is inside
336          * the ria_bytes.
337          */
338         if (ria->ria_length == 0 || ria->ria_length == ria->ria_bytes)
339                 return true;
340
341         if (pos >= ria->ria_stoff) {
342                 u64 offset;
343
344                 div64_u64_rem(pos - ria->ria_stoff, ria->ria_length, &offset);
345
346                 if (offset < ria->ria_bytes ||
347                     (ria->ria_length - offset) < PAGE_SIZE)
348                         return true;
349         } else if (pos + PAGE_SIZE > ria->ria_stoff) {
350                 return true;
351         }
352
353         return false;
354 }
355
356 static unsigned long
357 ll_read_ahead_pages(const struct lu_env *env, struct cl_io *io,
358                     struct cl_page_list *queue, struct ll_readahead_state *ras,
359                     struct ra_io_arg *ria, pgoff_t *ra_end)
360 {
361         struct cl_read_ahead ra = { 0 };
362         /* busy page count is per stride */
363         int rc = 0, count = 0, busy_page_count = 0;
364         pgoff_t page_idx;
365
366         LASSERT(ria != NULL);
367         RIA_DEBUG(ria);
368
369         for (page_idx = ria->ria_start_idx;
370              page_idx <= ria->ria_end_idx && ria->ria_reserved > 0;
371              page_idx++) {
372                 if (ras_inside_ra_window(page_idx, ria)) {
373                         if (ra.cra_end_idx == 0 || ra.cra_end_idx < page_idx) {
374                                 pgoff_t end_idx;
375
376                                 /*
377                                  * Do not shrink ria_end_idx at any case until
378                                  * the minimum end of current read is covered.
379                                  *
380                                  * Do not extend read lock accross stripe if
381                                  * lock contention detected.
382                                  */
383                                 if (ra.cra_contention &&
384                                     page_idx > ria->ria_end_idx_min) {
385                                         ria->ria_end_idx = *ra_end;
386                                         break;
387                                 }
388
389                                 cl_read_ahead_release(env, &ra);
390
391                                 rc = cl_io_read_ahead(env, io, page_idx, &ra);
392                                 if (rc < 0)
393                                         break;
394
395                                  /*
396                                   * Only shrink ria_end_idx if the matched
397                                   * LDLM lock doesn't cover more.
398                                   */
399                                 if (page_idx > ra.cra_end_idx) {
400                                         ria->ria_end_idx = ra.cra_end_idx;
401                                         break;
402                                 }
403
404                                 CDEBUG(D_READA, "idx: %lu, ra: %lu, rpc: %lu\n",
405                                        page_idx, ra.cra_end_idx,
406                                        ra.cra_rpc_pages);
407                                 LASSERTF(ra.cra_end_idx >= page_idx,
408                                          "object: %p, indcies %lu / %lu\n",
409                                          io->ci_obj, ra.cra_end_idx, page_idx);
410                                 /* update read ahead RPC size.
411                                  * NB: it's racy but doesn't matter */
412                                 if (ras->ras_rpc_pages != ra.cra_rpc_pages &&
413                                     ra.cra_rpc_pages > 0)
414                                         ras->ras_rpc_pages = ra.cra_rpc_pages;
415                                 /* trim it to align with optimal RPC size */
416                                 end_idx = ras_align(ras, ria->ria_end_idx + 1);
417                                 if (end_idx > 0 && !ria->ria_eof)
418                                         ria->ria_end_idx = end_idx - 1;
419                                 if (ria->ria_end_idx < ria->ria_end_idx_min)
420                                         ria->ria_end_idx = ria->ria_end_idx_min;
421                         }
422                         if (page_idx > ria->ria_end_idx)
423                                 break;
424
425                         /* If the page is inside the read-ahead window */
426                         rc = ll_read_ahead_page(env, io, queue, page_idx);
427                         if (rc < 0 && rc != -EBUSY)
428                                 break;
429                         if (rc == -EBUSY) {
430                                 busy_page_count++;
431                                 CDEBUG(D_READA,
432                                        "skip busy page: %lu\n", page_idx);
433                                 /* For page unaligned readahead the first
434                                  * last pages of each region can be read by
435                                  * another reader on the same node, and so
436                                  * may be busy. So only stop for > 2 busy
437                                  * pages. */
438                                 if (busy_page_count > 2)
439                                         break;
440                         }
441
442                         *ra_end = page_idx;
443                         /* Only subtract from reserve & count the page if we
444                          * really did readahead on that page. */
445                         if (rc == 0) {
446                                 ria->ria_reserved--;
447                                 count++;
448                         }
449                 } else if (stride_io_mode(ras)) {
450                         /* If it is not in the read-ahead window, and it is
451                          * read-ahead mode, then check whether it should skip
452                          * the stride gap.
453                          */
454                         loff_t pos = (loff_t)page_idx << PAGE_SHIFT;
455                         u64 offset;
456
457                         div64_u64_rem(pos - ria->ria_stoff, ria->ria_length,
458                                       &offset);
459                         if (offset >= ria->ria_bytes) {
460                                 pos += (ria->ria_length - offset);
461                                 if ((pos >> PAGE_SHIFT) >= page_idx + 1)
462                                         page_idx = (pos >> PAGE_SHIFT) - 1;
463                                 busy_page_count = 0;
464                                 CDEBUG(D_READA,
465                                        "Stride: jump %llu pages to %lu\n",
466                                        ria->ria_length - offset, page_idx);
467                                 continue;
468                         }
469                 }
470         }
471
472         cl_read_ahead_release(env, &ra);
473
474         return count;
475 }
476
477 static void ll_readahead_work_free(struct ll_readahead_work *work)
478 {
479         fput(work->lrw_file);
480         OBD_FREE_PTR(work);
481 }
482
483 static void ll_readahead_handle_work(struct work_struct *wq);
484 static void ll_readahead_work_add(struct inode *inode,
485                                   struct ll_readahead_work *work)
486 {
487         INIT_WORK(&work->lrw_readahead_work, ll_readahead_handle_work);
488         queue_work(ll_i2sbi(inode)->ll_ra_info.ll_readahead_wq,
489                    &work->lrw_readahead_work);
490 }
491
492 static int ll_readahead_file_kms(const struct lu_env *env,
493                                 struct cl_io *io, __u64 *kms)
494 {
495         struct cl_object *clob;
496         struct inode *inode;
497         struct cl_attr *attr = vvp_env_thread_attr(env);
498         int ret;
499
500         clob = io->ci_obj;
501         inode = vvp_object_inode(clob);
502
503         cl_object_attr_lock(clob);
504         ret = cl_object_attr_get(env, clob, attr);
505         cl_object_attr_unlock(clob);
506
507         if (ret != 0)
508                 RETURN(ret);
509
510         *kms = attr->cat_kms;
511         return 0;
512 }
513
514 static void ll_readahead_handle_work(struct work_struct *wq)
515 {
516         struct ll_readahead_work *work;
517         struct lu_env *env;
518         __u16 refcheck;
519         struct ra_io_arg *ria;
520         struct inode *inode;
521         struct ll_file_data *fd;
522         struct ll_readahead_state *ras;
523         struct cl_io *io;
524         struct cl_2queue *queue;
525         pgoff_t ra_end_idx = 0;
526         unsigned long pages, pages_min = 0;
527         struct file *file;
528         __u64 kms;
529         int rc;
530         pgoff_t eof_index;
531         struct ll_sb_info *sbi;
532
533         work = container_of(wq, struct ll_readahead_work,
534                             lrw_readahead_work);
535         fd = work->lrw_file->private_data;
536         ras = &fd->fd_ras;
537         file = work->lrw_file;
538         inode = file_inode(file);
539         sbi = ll_i2sbi(inode);
540
541         env = cl_env_alloc(&refcheck, LCT_NOREF);
542         if (IS_ERR(env))
543                 GOTO(out_free_work, rc = PTR_ERR(env));
544
545         io = vvp_env_thread_io(env);
546         ll_io_init(io, file, CIT_READ, NULL);
547
548         rc = ll_readahead_file_kms(env, io, &kms);
549         if (rc != 0)
550                 GOTO(out_put_env, rc);
551
552         if (kms == 0) {
553                 ll_ra_stats_inc(inode, RA_STAT_ZERO_LEN);
554                 GOTO(out_put_env, rc = 0);
555         }
556
557         ria = &ll_env_info(env)->lti_ria;
558         memset(ria, 0, sizeof(*ria));
559
560         ria->ria_start_idx = work->lrw_start_idx;
561         /* Truncate RA window to end of file */
562         eof_index = (pgoff_t)(kms - 1) >> PAGE_SHIFT;
563         if (eof_index <= work->lrw_end_idx) {
564                 work->lrw_end_idx = eof_index;
565                 ria->ria_eof = true;
566         }
567         if (work->lrw_end_idx <= work->lrw_start_idx)
568                 GOTO(out_put_env, rc = 0);
569
570         ria->ria_end_idx = work->lrw_end_idx;
571         pages = ria->ria_end_idx - ria->ria_start_idx + 1;
572         ria->ria_reserved = ll_ra_count_get(sbi, ria,
573                                             ria_page_count(ria), pages_min);
574
575         CDEBUG(D_READA,
576                "async reserved pages: %lu/%lu/%lu, ra_cur %d, ra_max %lu\n",
577                ria->ria_reserved, pages, pages_min,
578                atomic_read(&ll_i2sbi(inode)->ll_ra_info.ra_cur_pages),
579                ll_i2sbi(inode)->ll_ra_info.ra_max_pages);
580
581         if (ria->ria_reserved < pages) {
582                 ll_ra_stats_inc(inode, RA_STAT_MAX_IN_FLIGHT);
583                 if (PAGES_TO_MiB(ria->ria_reserved) < 1) {
584                         ll_ra_count_put(ll_i2sbi(inode), ria->ria_reserved);
585                         GOTO(out_put_env, rc = 0);
586                 }
587         }
588
589         rc = cl_io_rw_init(env, io, CIT_READ, ria->ria_start_idx, pages);
590         if (rc)
591                 GOTO(out_put_env, rc);
592
593         vvp_env_io(env)->vui_io_subtype = IO_NORMAL;
594         vvp_env_io(env)->vui_fd = fd;
595         io->ci_state = CIS_LOCKED;
596         io->ci_async_readahead = true;
597         rc = cl_io_start(env, io);
598         if (rc)
599                 GOTO(out_io_fini, rc);
600
601         queue = &io->ci_queue;
602         cl_2queue_init(queue);
603
604         rc = ll_read_ahead_pages(env, io, &queue->c2_qin, ras, ria,
605                                  &ra_end_idx);
606         if (ria->ria_reserved != 0)
607                 ll_ra_count_put(ll_i2sbi(inode), ria->ria_reserved);
608         if (queue->c2_qin.pl_nr > 0) {
609                 int count = queue->c2_qin.pl_nr;
610
611                 rc = cl_io_submit_rw(env, io, CRT_READ, queue);
612                 if (rc == 0)
613                         task_io_account_read(PAGE_SIZE * count);
614         }
615         if (ria->ria_end_idx == ra_end_idx && ra_end_idx == (kms >> PAGE_SHIFT))
616                 ll_ra_stats_inc(inode, RA_STAT_EOF);
617
618         if (ra_end_idx != ria->ria_end_idx)
619                 ll_ra_stats_inc(inode, RA_STAT_FAILED_REACH_END);
620
621         /* TODO: discard all pages until page reinit route is implemented */
622         cl_page_list_discard(env, io, &queue->c2_qin);
623
624         /* Unlock unsent read pages in case of error. */
625         cl_page_list_disown(env, io, &queue->c2_qin);
626
627         cl_2queue_fini(env, queue);
628 out_io_fini:
629         cl_io_end(env, io);
630         cl_io_fini(env, io);
631 out_put_env:
632         cl_env_put(env, &refcheck);
633 out_free_work:
634         if (ra_end_idx > 0)
635                 ll_ra_stats_inc_sbi(ll_i2sbi(inode), RA_STAT_ASYNC);
636         atomic_dec(&sbi->ll_ra_info.ra_async_inflight);
637         ll_readahead_work_free(work);
638 }
639
640 static int ll_readahead(const struct lu_env *env, struct cl_io *io,
641                         struct cl_page_list *queue,
642                         struct ll_readahead_state *ras, bool hit,
643                         struct file *file)
644 {
645         struct vvp_io *vio = vvp_env_io(env);
646         struct ll_thread_info *lti = ll_env_info(env);
647         unsigned long pages, pages_min = 0;
648         pgoff_t ra_end_idx = 0, start_idx = 0, end_idx = 0;
649         struct inode *inode;
650         struct ra_io_arg *ria = &lti->lti_ria;
651         struct cl_object *clob;
652         int ret = 0;
653         __u64 kms;
654         ENTRY;
655
656         clob = io->ci_obj;
657         inode = vvp_object_inode(clob);
658
659         memset(ria, 0, sizeof(*ria));
660         ret = ll_readahead_file_kms(env, io, &kms);
661         if (ret != 0)
662                 RETURN(ret);
663
664         if (kms == 0) {
665                 ll_ra_stats_inc(inode, RA_STAT_ZERO_LEN);
666                 RETURN(0);
667         }
668
669         spin_lock(&ras->ras_lock);
670
671         /**
672          * Note: other thread might rollback the ras_next_readahead_idx,
673          * if it can not get the full size of prepared pages, see the
674          * end of this function. For stride read ahead, it needs to
675          * make sure the offset is no less than ras_stride_offset,
676          * so that stride read ahead can work correctly.
677          */
678         if (stride_io_mode(ras))
679                 start_idx = max_t(pgoff_t, ras->ras_next_readahead_idx,
680                                   ras->ras_stride_offset >> PAGE_SHIFT);
681         else
682                 start_idx = ras->ras_next_readahead_idx;
683
684         if (ras->ras_window_pages > 0)
685                 end_idx = ras->ras_window_start_idx + ras->ras_window_pages - 1;
686
687         /* Enlarge the RA window to encompass the full read */
688         if (vio->vui_ra_valid &&
689             end_idx < vio->vui_ra_start_idx + vio->vui_ra_pages - 1)
690                 end_idx = vio->vui_ra_start_idx + vio->vui_ra_pages - 1;
691
692         if (end_idx != 0) {
693                 pgoff_t eof_index;
694
695                 /* Truncate RA window to end of file */
696                 eof_index = (pgoff_t)((kms - 1) >> PAGE_SHIFT);
697                 if (eof_index <= end_idx) {
698                         end_idx = eof_index;
699                         ria->ria_eof = true;
700                 }
701         }
702         ria->ria_start_idx = start_idx;
703         ria->ria_end_idx = end_idx;
704         /* If stride I/O mode is detected, get stride window*/
705         if (stride_io_mode(ras)) {
706                 ria->ria_stoff = ras->ras_stride_offset;
707                 ria->ria_length = ras->ras_stride_length;
708                 ria->ria_bytes = ras->ras_stride_bytes;
709         }
710         spin_unlock(&ras->ras_lock);
711
712         if (end_idx == 0) {
713                 ll_ra_stats_inc(inode, RA_STAT_ZERO_WINDOW);
714                 RETURN(0);
715         }
716         pages = ria_page_count(ria);
717         if (pages == 0) {
718                 ll_ra_stats_inc(inode, RA_STAT_ZERO_WINDOW);
719                 RETURN(0);
720         }
721
722         RAS_CDEBUG(ras);
723         CDEBUG(D_READA, DFID": ria: %lu/%lu, bead: %lu/%lu, hit: %d\n",
724                PFID(lu_object_fid(&clob->co_lu)),
725                ria->ria_start_idx, ria->ria_end_idx,
726                vio->vui_ra_valid ? vio->vui_ra_start_idx : 0,
727                vio->vui_ra_valid ? vio->vui_ra_pages : 0,
728                hit);
729
730         /* at least to extend the readahead window to cover current read */
731         if (!hit && vio->vui_ra_valid &&
732             vio->vui_ra_start_idx + vio->vui_ra_pages > ria->ria_start_idx)
733                 ria->ria_end_idx_min =
734                         vio->vui_ra_start_idx + vio->vui_ra_pages - 1;
735
736         ria->ria_reserved = ll_ra_count_get(ll_i2sbi(inode), ria, pages,
737                                             pages_min);
738         if (ria->ria_reserved < pages)
739                 ll_ra_stats_inc(inode, RA_STAT_MAX_IN_FLIGHT);
740
741         CDEBUG(D_READA, "reserved pages: %lu/%lu/%lu, ra_cur %d, ra_max %lu\n",
742                ria->ria_reserved, pages, pages_min,
743                atomic_read(&ll_i2sbi(inode)->ll_ra_info.ra_cur_pages),
744                ll_i2sbi(inode)->ll_ra_info.ra_max_pages);
745
746         ret = ll_read_ahead_pages(env, io, queue, ras, ria, &ra_end_idx);
747
748         if (ria->ria_reserved != 0)
749                 ll_ra_count_put(ll_i2sbi(inode), ria->ria_reserved);
750
751         if (ra_end_idx == end_idx && ra_end_idx == (kms >> PAGE_SHIFT))
752                 ll_ra_stats_inc(inode, RA_STAT_EOF);
753
754         CDEBUG(D_READA,
755                "ra_end_idx = %lu end_idx = %lu stride end = %lu pages = %d\n",
756                ra_end_idx, end_idx, ria->ria_end_idx, ret);
757
758         if (ra_end_idx != end_idx)
759                 ll_ra_stats_inc(inode, RA_STAT_FAILED_REACH_END);
760         if (ra_end_idx > 0) {
761                 /* update the ras so that the next read-ahead tries from
762                  * where we left off. */
763                 spin_lock(&ras->ras_lock);
764                 ras->ras_next_readahead_idx = ra_end_idx + 1;
765                 spin_unlock(&ras->ras_lock);
766                 RAS_CDEBUG(ras);
767         }
768
769         RETURN(ret);
770 }
771
772 static void ras_set_start(struct ll_readahead_state *ras, pgoff_t index)
773 {
774         ras->ras_window_start_idx = ras_align(ras, index);
775 }
776
777 /* called with the ras_lock held or from places where it doesn't matter */
778 static void ras_reset(struct ll_readahead_state *ras, pgoff_t index)
779 {
780         ras->ras_consecutive_requests = 0;
781         ras->ras_consecutive_bytes = 0;
782         ras->ras_window_pages = 0;
783         ras_set_start(ras, index);
784         ras->ras_next_readahead_idx = max(ras->ras_window_start_idx, index + 1);
785
786         RAS_CDEBUG(ras);
787 }
788
789 /* called with the ras_lock held or from places where it doesn't matter */
790 static void ras_stride_reset(struct ll_readahead_state *ras)
791 {
792         ras->ras_consecutive_stride_requests = 0;
793         ras->ras_stride_length = 0;
794         ras->ras_stride_bytes = 0;
795         RAS_CDEBUG(ras);
796 }
797
798 void ll_readahead_init(struct inode *inode, struct ll_readahead_state *ras)
799 {
800         spin_lock_init(&ras->ras_lock);
801         ras->ras_rpc_pages = PTLRPC_MAX_BRW_PAGES;
802         ras_reset(ras, 0);
803         ras->ras_last_read_end_bytes = 0;
804         ras->ras_requests = 0;
805 }
806
807 /*
808  * Check whether the read request is in the stride window.
809  * If it is in the stride window, return true, otherwise return false.
810  */
811 static bool read_in_stride_window(struct ll_readahead_state *ras,
812                                   loff_t pos, loff_t count)
813 {
814         loff_t stride_gap;
815
816         if (ras->ras_stride_length == 0 || ras->ras_stride_bytes == 0 ||
817             ras->ras_stride_bytes == ras->ras_stride_length)
818                 return false;
819
820         stride_gap = pos - ras->ras_last_read_end_bytes - 1;
821
822         /* If it is contiguous read */
823         if (stride_gap == 0)
824                 return ras->ras_consecutive_bytes + count <=
825                         ras->ras_stride_bytes;
826
827         /* Otherwise check the stride by itself */
828         return (ras->ras_stride_length - ras->ras_stride_bytes) == stride_gap &&
829                 ras->ras_consecutive_bytes == ras->ras_stride_bytes &&
830                 count <= ras->ras_stride_bytes;
831 }
832
833 static void ras_init_stride_detector(struct ll_readahead_state *ras,
834                                      loff_t pos, loff_t count)
835 {
836         loff_t stride_gap = pos - ras->ras_last_read_end_bytes - 1;
837
838         LASSERT(ras->ras_consecutive_stride_requests == 0);
839
840         if (pos <= ras->ras_last_read_end_bytes) {
841                 /*Reset stride window for forward read*/
842                 ras_stride_reset(ras);
843                 return;
844         }
845
846         ras->ras_stride_bytes = ras->ras_consecutive_bytes;
847         ras->ras_stride_length = stride_gap + ras->ras_consecutive_bytes;
848         ras->ras_consecutive_stride_requests++;
849         ras->ras_stride_offset = pos;
850
851         RAS_CDEBUG(ras);
852 }
853
854 static unsigned long
855 stride_page_count(struct ll_readahead_state *ras, loff_t len)
856 {
857         loff_t bytes_count =
858                 stride_byte_count(ras->ras_stride_offset,
859                                   ras->ras_stride_length, ras->ras_stride_bytes,
860                                   ras->ras_window_start_idx << PAGE_SHIFT, len);
861
862         return (bytes_count + PAGE_SIZE - 1) >> PAGE_SHIFT;
863 }
864
865 /* Stride Read-ahead window will be increased inc_len according to
866  * stride I/O pattern */
867 static void ras_stride_increase_window(struct ll_readahead_state *ras,
868                                        struct ll_ra_info *ra, loff_t inc_bytes)
869 {
870         loff_t window_bytes, stride_bytes;
871         u64 left_bytes;
872         u64 step;
873         loff_t end;
874
875         /* temporarily store in page units to reduce LASSERT() cost below */
876         end = ras->ras_window_start_idx + ras->ras_window_pages;
877
878         LASSERT(ras->ras_stride_length > 0);
879         LASSERTF(end >= (ras->ras_stride_offset >> PAGE_SHIFT),
880                  "window_start_idx %lu, window_pages %lu stride_offset %llu\n",
881                  ras->ras_window_start_idx, ras->ras_window_pages,
882                  ras->ras_stride_offset);
883
884         end <<= PAGE_SHIFT;
885         if (end <= ras->ras_stride_offset)
886                 stride_bytes = 0;
887         else
888                 stride_bytes = end - ras->ras_stride_offset;
889
890         div64_u64_rem(stride_bytes, ras->ras_stride_length, &left_bytes);
891         window_bytes = (ras->ras_window_pages << PAGE_SHIFT);
892         if (left_bytes < ras->ras_stride_bytes) {
893                 if (ras->ras_stride_bytes - left_bytes >= inc_bytes) {
894                         window_bytes += inc_bytes;
895                         goto out;
896                 } else {
897                         window_bytes += (ras->ras_stride_bytes - left_bytes);
898                         inc_bytes -= (ras->ras_stride_bytes - left_bytes);
899                 }
900         } else {
901                 window_bytes += (ras->ras_stride_length - left_bytes);
902         }
903
904         LASSERT(ras->ras_stride_bytes != 0);
905
906         step = div64_u64_rem(inc_bytes, ras->ras_stride_bytes, &left_bytes);
907
908         window_bytes += step * ras->ras_stride_length + left_bytes;
909         LASSERT(window_bytes > 0);
910
911 out:
912         if (stride_page_count(ras, window_bytes) <=
913             ra->ra_max_pages_per_file || ras->ras_window_pages == 0)
914                 ras->ras_window_pages = (window_bytes >> PAGE_SHIFT);
915
916         LASSERT(ras->ras_window_pages > 0);
917
918         RAS_CDEBUG(ras);
919 }
920
921 static void ras_increase_window(struct inode *inode,
922                                 struct ll_readahead_state *ras,
923                                 struct ll_ra_info *ra)
924 {
925         /* The stretch of ra-window should be aligned with max rpc_size
926          * but current clio architecture does not support retrieve such
927          * information from lower layer. FIXME later
928          */
929         if (stride_io_mode(ras)) {
930                 ras_stride_increase_window(ras, ra,
931                                       (loff_t)ras->ras_rpc_pages << PAGE_SHIFT);
932         } else {
933                 pgoff_t window_pages;
934
935                 window_pages = min(ras->ras_window_pages + ras->ras_rpc_pages,
936                                    ra->ra_max_pages_per_file);
937                 if (window_pages < ras->ras_rpc_pages)
938                         ras->ras_window_pages = window_pages;
939                 else
940                         ras->ras_window_pages = ras_align(ras, window_pages);
941         }
942 }
943
944 /**
945  * Seek within 8 pages are considered as sequential read for now.
946  */
947 static inline bool is_loose_seq_read(struct ll_readahead_state *ras, loff_t pos)
948 {
949         return pos_in_window(pos, ras->ras_last_read_end_bytes,
950                              8UL << PAGE_SHIFT, 8UL << PAGE_SHIFT);
951 }
952
953 static void ras_detect_read_pattern(struct ll_readahead_state *ras,
954                                     struct ll_sb_info *sbi,
955                                     loff_t pos, size_t count, bool mmap)
956 {
957         bool stride_detect = false;
958         pgoff_t index = pos >> PAGE_SHIFT;
959
960         /*
961          * Reset the read-ahead window in two cases. First when the app seeks
962          * or reads to some other part of the file. Secondly if we get a
963          * read-ahead miss that we think we've previously issued. This can
964          * be a symptom of there being so many read-ahead pages that the VM
965          * is reclaiming it before we get to it.
966          */
967         if (!is_loose_seq_read(ras, pos)) {
968                 /* Check whether it is in stride I/O mode */
969                 if (!read_in_stride_window(ras, pos, count)) {
970                         if (ras->ras_consecutive_stride_requests == 0)
971                                 ras_init_stride_detector(ras, pos, count);
972                         else
973                                 ras_stride_reset(ras);
974                         ras->ras_consecutive_bytes = 0;
975                         ras_reset(ras, index);
976                 } else {
977                         ras->ras_consecutive_bytes = 0;
978                         ras->ras_consecutive_requests = 0;
979                         if (++ras->ras_consecutive_stride_requests > 1)
980                                 stride_detect = true;
981                         RAS_CDEBUG(ras);
982                 }
983                 ll_ra_stats_inc_sbi(sbi, RA_STAT_DISTANT_READPAGE);
984         } else if (stride_io_mode(ras)) {
985                 /*
986                  * If this is contiguous read but in stride I/O mode
987                  * currently, check whether stride step still is valid,
988                  * if invalid, it will reset the stride ra window to
989                  * be zero.
990                  */
991                 if (!read_in_stride_window(ras, pos, count)) {
992                         ras_stride_reset(ras);
993                         ras->ras_window_pages = 0;
994                         ras->ras_next_readahead_idx = index;
995                 }
996         }
997
998         ras->ras_consecutive_bytes += count;
999         if (mmap) {
1000                 pgoff_t idx = ras->ras_consecutive_bytes >> PAGE_SHIFT;
1001
1002                 if ((idx >= 4 && (idx & 3UL) == 0) || stride_detect)
1003                         ras->ras_need_increase_window = true;
1004         } else if ((ras->ras_consecutive_requests > 1 || stride_detect)) {
1005                 ras->ras_need_increase_window = true;
1006         }
1007
1008         ras->ras_last_read_end_bytes = pos + count - 1;
1009 }
1010
1011 void ll_ras_enter(struct file *f, loff_t pos, size_t count)
1012 {
1013         struct ll_file_data *fd = f->private_data;
1014         struct ll_readahead_state *ras = &fd->fd_ras;
1015         struct inode *inode = file_inode(f);
1016         unsigned long index = pos >> PAGE_SHIFT;
1017         struct ll_sb_info *sbi = ll_i2sbi(inode);
1018
1019         spin_lock(&ras->ras_lock);
1020         ras->ras_requests++;
1021         ras->ras_consecutive_requests++;
1022         ras->ras_need_increase_window = false;
1023         ras->ras_no_miss_check = false;
1024         /*
1025          * On the second access to a file smaller than the tunable
1026          * ra_max_read_ahead_whole_pages trigger RA on all pages in the
1027          * file up to ra_max_pages_per_file.  This is simply a best effort
1028          * and only occurs once per open file. Normal RA behavior is reverted
1029          * to for subsequent IO.
1030          */
1031         if (ras->ras_requests >= 2) {
1032                 __u64 kms_pages;
1033                 struct ll_ra_info *ra = &sbi->ll_ra_info;
1034
1035                 kms_pages = (i_size_read(inode) + PAGE_SIZE - 1) >>
1036                             PAGE_SHIFT;
1037
1038                 CDEBUG(D_READA, "kmsp %llu mwp %lu mp %lu\n", kms_pages,
1039                        ra->ra_max_read_ahead_whole_pages,
1040                        ra->ra_max_pages_per_file);
1041
1042                 if (kms_pages &&
1043                     kms_pages <= ra->ra_max_read_ahead_whole_pages) {
1044                         ras->ras_window_start_idx = 0;
1045                         ras->ras_next_readahead_idx = index + 1;
1046                         ras->ras_window_pages = min(ra->ra_max_pages_per_file,
1047                                             ra->ra_max_read_ahead_whole_pages);
1048                         ras->ras_no_miss_check = true;
1049                         GOTO(out_unlock, 0);
1050                 }
1051         }
1052         ras_detect_read_pattern(ras, sbi, pos, count, false);
1053 out_unlock:
1054         spin_unlock(&ras->ras_lock);
1055 }
1056
1057 static bool index_in_stride_window(struct ll_readahead_state *ras,
1058                                    pgoff_t index)
1059 {
1060         loff_t pos = (loff_t)index << PAGE_SHIFT;
1061
1062         if (ras->ras_stride_length == 0 || ras->ras_stride_bytes == 0 ||
1063             ras->ras_stride_bytes == ras->ras_stride_length)
1064                 return false;
1065
1066         if (pos >= ras->ras_stride_offset) {
1067                 u64 offset;
1068
1069                 div64_u64_rem(pos - ras->ras_stride_offset,
1070                               ras->ras_stride_length, &offset);
1071                 if (offset < ras->ras_stride_bytes ||
1072                     ras->ras_stride_length - offset < PAGE_SIZE)
1073                         return true;
1074         } else if (ras->ras_stride_offset - pos < PAGE_SIZE) {
1075                 return true;
1076         }
1077
1078         return false;
1079 }
1080
1081 /*
1082  * ll_ras_enter() is used to detect read pattern according to pos and count.
1083  *
1084  * ras_update() is used to detect cache miss and
1085  * reset window or increase window accordingly
1086  */
1087 static void ras_update(struct ll_sb_info *sbi, struct inode *inode,
1088                        struct ll_readahead_state *ras, pgoff_t index,
1089                        enum ras_update_flags flags)
1090 {
1091         struct ll_ra_info *ra = &sbi->ll_ra_info;
1092         bool hit = flags & LL_RAS_HIT;
1093
1094         ENTRY;
1095         spin_lock(&ras->ras_lock);
1096
1097         if (!hit)
1098                 CDEBUG(D_READA, DFID " pages at %lu miss.\n",
1099                        PFID(ll_inode2fid(inode)), index);
1100         ll_ra_stats_inc_sbi(sbi, hit ? RA_STAT_HIT : RA_STAT_MISS);
1101
1102         /*
1103          * The readahead window has been expanded to cover whole
1104          * file size, we don't care whether ra miss happen or not.
1105          * Because we will read whole file to page cache even if
1106          * some pages missed.
1107          */
1108         if (ras->ras_no_miss_check)
1109                 GOTO(out_unlock, 0);
1110
1111         if (flags & LL_RAS_MMAP)
1112                 ras_detect_read_pattern(ras, sbi, (loff_t)index << PAGE_SHIFT,
1113                                         PAGE_SIZE, true);
1114
1115         if (!hit && ras->ras_window_pages &&
1116             index < ras->ras_next_readahead_idx &&
1117             pos_in_window(index, ras->ras_window_start_idx, 0,
1118                           ras->ras_window_pages)) {
1119                 ll_ra_stats_inc_sbi(sbi, RA_STAT_MISS_IN_WINDOW);
1120                 ras->ras_need_increase_window = false;
1121
1122                 if (index_in_stride_window(ras, index) &&
1123                     stride_io_mode(ras)) {
1124                         /*
1125                          * if (index != ras->ras_last_readpage + 1)
1126                          *      ras->ras_consecutive_pages = 0;
1127                          */
1128                         ras_reset(ras, index);
1129
1130                         /*
1131                          * If stride-RA hit cache miss, the stride
1132                          * detector will not be reset to avoid the
1133                          * overhead of redetecting read-ahead mode,
1134                          * but on the condition that the stride window
1135                          * is still intersect with normal sequential
1136                          * read-ahead window.
1137                          */
1138                         if (ras->ras_window_start_idx < ras->ras_stride_offset)
1139                                 ras_stride_reset(ras);
1140                         RAS_CDEBUG(ras);
1141                 } else {
1142                         /*
1143                          * Reset both stride window and normal RA
1144                          * window.
1145                          */
1146                         ras_reset(ras, index);
1147                         /* ras->ras_consecutive_pages++; */
1148                         ras->ras_consecutive_bytes = 0;
1149                         ras_stride_reset(ras);
1150                         GOTO(out_unlock, 0);
1151                 }
1152         }
1153         ras_set_start(ras, index);
1154
1155         if (stride_io_mode(ras)) {
1156                 /* Since stride readahead is sentivite to the offset
1157                  * of read-ahead, so we use original offset here,
1158                  * instead of ras_window_start_idx, which is RPC aligned.
1159                  */
1160                 ras->ras_next_readahead_idx = max(index + 1,
1161                                                   ras->ras_next_readahead_idx);
1162                 ras->ras_window_start_idx =
1163                                 max_t(pgoff_t, ras->ras_window_start_idx,
1164                                       ras->ras_stride_offset >> PAGE_SHIFT);
1165         } else {
1166                 if (ras->ras_next_readahead_idx < ras->ras_window_start_idx)
1167                         ras->ras_next_readahead_idx = ras->ras_window_start_idx;
1168                 if (!hit)
1169                         ras->ras_next_readahead_idx = index + 1;
1170         }
1171
1172         if (ras->ras_need_increase_window) {
1173                 ras_increase_window(inode, ras, ra);
1174                 ras->ras_need_increase_window = false;
1175         }
1176
1177         EXIT;
1178 out_unlock:
1179         spin_unlock(&ras->ras_lock);
1180 }
1181
1182 int ll_writepage(struct page *vmpage, struct writeback_control *wbc)
1183 {
1184         struct inode           *inode = vmpage->mapping->host;
1185         struct ll_inode_info   *lli   = ll_i2info(inode);
1186         struct lu_env          *env;
1187         struct cl_io           *io;
1188         struct cl_page         *page;
1189         struct cl_object       *clob;
1190         bool redirtied = false;
1191         bool unlocked = false;
1192         int result;
1193         __u16 refcheck;
1194         ENTRY;
1195
1196         LASSERT(PageLocked(vmpage));
1197         LASSERT(!PageWriteback(vmpage));
1198
1199         LASSERT(ll_i2dtexp(inode) != NULL);
1200
1201         env = cl_env_get(&refcheck);
1202         if (IS_ERR(env))
1203                 GOTO(out, result = PTR_ERR(env));
1204
1205         clob  = ll_i2info(inode)->lli_clob;
1206         LASSERT(clob != NULL);
1207
1208         io = vvp_env_thread_io(env);
1209         io->ci_obj = clob;
1210         io->ci_ignore_layout = 1;
1211         result = cl_io_init(env, io, CIT_MISC, clob);
1212         if (result == 0) {
1213                 page = cl_page_find(env, clob, vmpage->index,
1214                                     vmpage, CPT_CACHEABLE);
1215                 if (!IS_ERR(page)) {
1216                         lu_ref_add(&page->cp_reference, "writepage",
1217                                    current);
1218                         cl_page_assume(env, io, page);
1219                         result = cl_page_flush(env, io, page);
1220                         if (result != 0) {
1221                                 /*
1222                                  * Re-dirty page on error so it retries write,
1223                                  * but not in case when IO has actually
1224                                  * occurred and completed with an error.
1225                                  */
1226                                 if (!PageError(vmpage)) {
1227                                         redirty_page_for_writepage(wbc, vmpage);
1228                                         result = 0;
1229                                         redirtied = true;
1230                                 }
1231                         }
1232                         cl_page_disown(env, io, page);
1233                         unlocked = true;
1234                         lu_ref_del(&page->cp_reference,
1235                                    "writepage", current);
1236                         cl_page_put(env, page);
1237                 } else {
1238                         result = PTR_ERR(page);
1239                 }
1240         }
1241         cl_io_fini(env, io);
1242
1243         if (redirtied && wbc->sync_mode == WB_SYNC_ALL) {
1244                 loff_t offset = cl_offset(clob, vmpage->index);
1245
1246                 /* Flush page failed because the extent is being written out.
1247                  * Wait for the write of extent to be finished to avoid
1248                  * breaking kernel which assumes ->writepage should mark
1249                  * PageWriteback or clean the page. */
1250                 result = cl_sync_file_range(inode, offset,
1251                                             offset + PAGE_SIZE - 1,
1252                                             CL_FSYNC_LOCAL, 1);
1253                 if (result > 0) {
1254                         /* actually we may have written more than one page.
1255                          * decreasing this page because the caller will count
1256                          * it. */
1257                         wbc->nr_to_write -= result - 1;
1258                         result = 0;
1259                 }
1260         }
1261
1262         cl_env_put(env, &refcheck);
1263         GOTO(out, result);
1264
1265 out:
1266         if (result < 0) {
1267                 if (!lli->lli_async_rc)
1268                         lli->lli_async_rc = result;
1269                 SetPageError(vmpage);
1270                 if (!unlocked)
1271                         unlock_page(vmpage);
1272         }
1273         return result;
1274 }
1275
1276 int ll_writepages(struct address_space *mapping, struct writeback_control *wbc)
1277 {
1278         struct inode *inode = mapping->host;
1279         loff_t start;
1280         loff_t end;
1281         enum cl_fsync_mode mode;
1282         int range_whole = 0;
1283         int result;
1284         ENTRY;
1285
1286         if (wbc->range_cyclic) {
1287                 start = (loff_t)mapping->writeback_index << PAGE_SHIFT;
1288                 end = OBD_OBJECT_EOF;
1289         } else {
1290                 start = wbc->range_start;
1291                 end = wbc->range_end;
1292                 if (end == LLONG_MAX) {
1293                         end = OBD_OBJECT_EOF;
1294                         range_whole = start == 0;
1295                 }
1296         }
1297
1298         mode = CL_FSYNC_NONE;
1299         if (wbc->sync_mode == WB_SYNC_ALL)
1300                 mode = CL_FSYNC_LOCAL;
1301
1302         if (ll_i2info(inode)->lli_clob == NULL)
1303                 RETURN(0);
1304
1305         /* for directio, it would call writepages() to evict cached pages
1306          * inside the IO context of write, which will cause deadlock at
1307          * layout_conf since it waits for active IOs to complete. */
1308         result = cl_sync_file_range(inode, start, end, mode, 1);
1309         if (result > 0) {
1310                 wbc->nr_to_write -= result;
1311                 result = 0;
1312          }
1313
1314         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0)) {
1315                 if (end == OBD_OBJECT_EOF)
1316                         mapping->writeback_index = 0;
1317                 else
1318                         mapping->writeback_index = (end >> PAGE_SHIFT) + 1;
1319         }
1320         RETURN(result);
1321 }
1322
1323 struct ll_cl_context *ll_cl_find(struct file *file)
1324 {
1325         struct ll_file_data *fd = file->private_data;
1326         struct ll_cl_context *lcc;
1327         struct ll_cl_context *found = NULL;
1328
1329         read_lock(&fd->fd_lock);
1330         list_for_each_entry(lcc, &fd->fd_lccs, lcc_list) {
1331                 if (lcc->lcc_cookie == current) {
1332                         found = lcc;
1333                         break;
1334                 }
1335         }
1336         read_unlock(&fd->fd_lock);
1337
1338         return found;
1339 }
1340
1341 void ll_cl_add(struct file *file, const struct lu_env *env, struct cl_io *io,
1342                enum lcc_type type)
1343 {
1344         struct ll_file_data *fd = file->private_data;
1345         struct ll_cl_context *lcc = &ll_env_info(env)->lti_io_ctx;
1346
1347         memset(lcc, 0, sizeof(*lcc));
1348         INIT_LIST_HEAD(&lcc->lcc_list);
1349         lcc->lcc_cookie = current;
1350         lcc->lcc_env = env;
1351         lcc->lcc_io = io;
1352         lcc->lcc_type = type;
1353
1354         write_lock(&fd->fd_lock);
1355         list_add(&lcc->lcc_list, &fd->fd_lccs);
1356         write_unlock(&fd->fd_lock);
1357 }
1358
1359 void ll_cl_remove(struct file *file, const struct lu_env *env)
1360 {
1361         struct ll_file_data *fd = file->private_data;
1362         struct ll_cl_context *lcc = &ll_env_info(env)->lti_io_ctx;
1363
1364         write_lock(&fd->fd_lock);
1365         list_del_init(&lcc->lcc_list);
1366         write_unlock(&fd->fd_lock);
1367 }
1368
1369 int ll_io_read_page(const struct lu_env *env, struct cl_io *io,
1370                            struct cl_page *page, struct file *file)
1371 {
1372         struct inode              *inode  = vvp_object_inode(page->cp_obj);
1373         struct ll_sb_info         *sbi    = ll_i2sbi(inode);
1374         struct ll_file_data       *fd     = file->private_data;
1375         struct ll_readahead_state *ras    = &fd->fd_ras;
1376         struct cl_2queue          *queue  = &io->ci_queue;
1377         struct cl_sync_io         *anchor = NULL;
1378         struct vvp_page           *vpg;
1379         int                        rc = 0;
1380         bool                       uptodate;
1381         ENTRY;
1382
1383         vpg = cl2vvp_page(cl_object_page_slice(page->cp_obj, page));
1384         uptodate = vpg->vpg_defer_uptodate;
1385
1386         if (sbi->ll_ra_info.ra_max_pages_per_file > 0 &&
1387             sbi->ll_ra_info.ra_max_pages > 0 &&
1388             !vpg->vpg_ra_updated) {
1389                 struct vvp_io *vio = vvp_env_io(env);
1390                 enum ras_update_flags flags = 0;
1391
1392                 if (uptodate)
1393                         flags |= LL_RAS_HIT;
1394                 if (!vio->vui_ra_valid)
1395                         flags |= LL_RAS_MMAP;
1396                 ras_update(sbi, inode, ras, vvp_index(vpg), flags);
1397         }
1398
1399         cl_2queue_init(queue);
1400         if (uptodate) {
1401                 vpg->vpg_ra_used = 1;
1402                 cl_page_export(env, page, 1);
1403                 cl_page_disown(env, io, page);
1404         } else {
1405                 anchor = &vvp_env_info(env)->vti_anchor;
1406                 cl_sync_io_init(anchor, 1);
1407                 page->cp_sync_io = anchor;
1408
1409                 cl_2queue_add(queue, page);
1410         }
1411
1412         if (sbi->ll_ra_info.ra_max_pages_per_file > 0 &&
1413             sbi->ll_ra_info.ra_max_pages > 0) {
1414                 int rc2;
1415
1416                 rc2 = ll_readahead(env, io, &queue->c2_qin, ras,
1417                                    uptodate, file);
1418                 CDEBUG(D_READA, DFID "%d pages read ahead at %lu\n",
1419                        PFID(ll_inode2fid(inode)), rc2, vvp_index(vpg));
1420         }
1421
1422         if (queue->c2_qin.pl_nr > 0) {
1423                 int count = queue->c2_qin.pl_nr;
1424                 rc = cl_io_submit_rw(env, io, CRT_READ, queue);
1425                 if (rc == 0)
1426                         task_io_account_read(PAGE_SIZE * count);
1427         }
1428
1429
1430         if (anchor != NULL && !cl_page_is_owned(page, io)) { /* have sent */
1431                 rc = cl_sync_io_wait(env, anchor, 0);
1432
1433                 cl_page_assume(env, io, page);
1434                 cl_page_list_del(env, &queue->c2_qout, page);
1435
1436                 if (!PageUptodate(cl_page_vmpage(page))) {
1437                         /* Failed to read a mirror, discard this page so that
1438                          * new page can be created with new mirror.
1439                          *
1440                          * TODO: this is not needed after page reinit
1441                          * route is implemented */
1442                         cl_page_discard(env, io, page);
1443                 }
1444                 cl_page_disown(env, io, page);
1445         }
1446
1447         /* TODO: discard all pages until page reinit route is implemented */
1448         cl_page_list_discard(env, io, &queue->c2_qin);
1449
1450         /* Unlock unsent read pages in case of error. */
1451         cl_page_list_disown(env, io, &queue->c2_qin);
1452
1453         cl_2queue_fini(env, queue);
1454
1455         RETURN(rc);
1456 }
1457
1458 /*
1459  * Possible return value:
1460  * 0 no async readahead triggered and fast read could not be used.
1461  * 1 no async readahead, but fast read could be used.
1462  * 2 async readahead triggered and fast read could be used too.
1463  * < 0 on error.
1464  */
1465 static int kickoff_async_readahead(struct file *file, unsigned long pages)
1466 {
1467         struct ll_readahead_work *lrw;
1468         struct inode *inode = file_inode(file);
1469         struct ll_sb_info *sbi = ll_i2sbi(inode);
1470         struct ll_file_data *fd = file->private_data;
1471         struct ll_readahead_state *ras = &fd->fd_ras;
1472         struct ll_ra_info *ra = &sbi->ll_ra_info;
1473         unsigned long throttle;
1474         pgoff_t start_idx = ras_align(ras, ras->ras_next_readahead_idx);
1475         pgoff_t end_idx = start_idx + pages - 1;
1476
1477         throttle = min(ra->ra_async_pages_per_file_threshold,
1478                        ra->ra_max_pages_per_file);
1479         /*
1480          * If this is strided i/o or the window is smaller than the
1481          * throttle limit, we do not do async readahead. Otherwise,
1482          * we do async readahead, allowing the user thread to do fast i/o.
1483          */
1484         if (stride_io_mode(ras) || !throttle ||
1485             ras->ras_window_pages < throttle ||
1486             atomic_read(&ra->ra_async_inflight) > ra->ra_async_max_active)
1487                 return 0;
1488
1489         if ((atomic_read(&ra->ra_cur_pages) + pages) > ra->ra_max_pages)
1490                 return 0;
1491
1492         if (ras->ras_async_last_readpage_idx == start_idx)
1493                 return 1;
1494
1495         /* ll_readahead_work_free() free it */
1496         OBD_ALLOC_PTR(lrw);
1497         if (lrw) {
1498                 atomic_inc(&sbi->ll_ra_info.ra_async_inflight);
1499                 lrw->lrw_file = get_file(file);
1500                 lrw->lrw_start_idx = start_idx;
1501                 lrw->lrw_end_idx = end_idx;
1502                 spin_lock(&ras->ras_lock);
1503                 ras->ras_next_readahead_idx = end_idx + 1;
1504                 ras->ras_async_last_readpage_idx = start_idx;
1505                 spin_unlock(&ras->ras_lock);
1506                 ll_readahead_work_add(inode, lrw);
1507         } else {
1508                 return -ENOMEM;
1509         }
1510
1511         return 2;
1512 }
1513
1514 /*
1515  * Check if we can issue a readahead RPC, if that is
1516  * the case, we can't do fast IO because we will need
1517  * a cl_io to issue the RPC.
1518  */
1519 static bool ll_use_fast_io(struct file *file,
1520                            struct ll_readahead_state *ras, pgoff_t index)
1521 {
1522         unsigned long fast_read_pages =
1523                 max(RA_REMAIN_WINDOW_MIN, ras->ras_rpc_pages);
1524         loff_t skip_pages;
1525
1526         if (stride_io_mode(ras)) {
1527                 skip_pages = (ras->ras_stride_length +
1528                         ras->ras_stride_bytes - 1) / ras->ras_stride_bytes;
1529                 skip_pages *= fast_read_pages;
1530         } else {
1531                 skip_pages = fast_read_pages;
1532         }
1533
1534         if (ras->ras_window_start_idx + ras->ras_window_pages <
1535             ras->ras_next_readahead_idx + skip_pages ||
1536             kickoff_async_readahead(file, fast_read_pages) > 0)
1537                 return true;
1538
1539         return false;
1540 }
1541
1542 int ll_readpage(struct file *file, struct page *vmpage)
1543 {
1544         struct inode *inode = file_inode(file);
1545         struct cl_object *clob = ll_i2info(inode)->lli_clob;
1546         struct ll_cl_context *lcc;
1547         const struct lu_env  *env = NULL;
1548         struct cl_io   *io = NULL;
1549         struct cl_page *page;
1550         struct ll_sb_info *sbi = ll_i2sbi(inode);
1551         int result;
1552         ENTRY;
1553
1554         lcc = ll_cl_find(file);
1555         if (lcc != NULL) {
1556                 env = lcc->lcc_env;
1557                 io  = lcc->lcc_io;
1558         }
1559
1560         if (io == NULL) { /* fast read */
1561                 struct inode *inode = file_inode(file);
1562                 struct ll_file_data *fd = file->private_data;
1563                 struct ll_readahead_state *ras = &fd->fd_ras;
1564                 struct lu_env  *local_env = NULL;
1565                 struct vvp_page *vpg;
1566
1567                 result = -ENODATA;
1568
1569                 /* TODO: need to verify the layout version to make sure
1570                  * the page is not invalid due to layout change. */
1571                 page = cl_vmpage_page(vmpage, clob);
1572                 if (page == NULL) {
1573                         unlock_page(vmpage);
1574                         ll_ra_stats_inc_sbi(sbi, RA_STAT_FAILED_FAST_READ);
1575                         RETURN(result);
1576                 }
1577
1578                 vpg = cl2vvp_page(cl_object_page_slice(page->cp_obj, page));
1579                 if (vpg->vpg_defer_uptodate) {
1580                         enum ras_update_flags flags = LL_RAS_HIT;
1581
1582                         if (lcc && lcc->lcc_type == LCC_MMAP)
1583                                 flags |= LL_RAS_MMAP;
1584
1585                         /* For fast read, it updates read ahead state only
1586                          * if the page is hit in cache because non cache page
1587                          * case will be handled by slow read later. */
1588                         ras_update(sbi, inode, ras, vvp_index(vpg), flags);
1589                         /* avoid duplicate ras_update() call */
1590                         vpg->vpg_ra_updated = 1;
1591
1592                         if (ll_use_fast_io(file, ras, vvp_index(vpg)))
1593                                 result = 0;
1594                 }
1595
1596                 if (!env) {
1597                         local_env = cl_env_percpu_get();
1598                         env = local_env;
1599                 }
1600
1601                 /* export the page and skip io stack */
1602                 if (result == 0) {
1603                         vpg->vpg_ra_used = 1;
1604                         cl_page_export(env, page, 1);
1605                 } else {
1606                         ll_ra_stats_inc_sbi(sbi, RA_STAT_FAILED_FAST_READ);
1607                 }
1608                 /* release page refcount before unlocking the page to ensure
1609                  * the object won't be destroyed in the calling path of
1610                  * cl_page_put(). Please see comment in ll_releasepage(). */
1611                 cl_page_put(env, page);
1612                 unlock_page(vmpage);
1613                 if (local_env)
1614                         cl_env_percpu_put(local_env);
1615
1616                 RETURN(result);
1617         }
1618
1619         /**
1620          * Direct read can fall back to buffered read, but DIO is done
1621          * with lockless i/o, and buffered requires LDLM locking, so in
1622          * this case we must restart without lockless.
1623          */
1624         if (file->f_flags & O_DIRECT &&
1625             lcc && lcc->lcc_type == LCC_RW &&
1626             !io->ci_ignore_lockless) {
1627                 unlock_page(vmpage);
1628                 io->ci_ignore_lockless = 1;
1629                 io->ci_need_restart = 1;
1630                 RETURN(-ENOLCK);
1631         }
1632
1633         LASSERT(io->ci_state == CIS_IO_GOING);
1634         page = cl_page_find(env, clob, vmpage->index, vmpage, CPT_CACHEABLE);
1635         if (!IS_ERR(page)) {
1636                 LASSERT(page->cp_type == CPT_CACHEABLE);
1637                 if (likely(!PageUptodate(vmpage))) {
1638                         cl_page_assume(env, io, page);
1639
1640                         result = ll_io_read_page(env, io, page, file);
1641                 } else {
1642                         /* Page from a non-object file. */
1643                         unlock_page(vmpage);
1644                         result = 0;
1645                 }
1646                 cl_page_put(env, page);
1647         } else {
1648                 unlock_page(vmpage);
1649                 result = PTR_ERR(page);
1650         }
1651         RETURN(result);
1652 }