Commit | Line | Data |
---|---|---|
d7e09d03 PT |
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.sun.com/software/products/lustre/docs/GPLv2.pdf | |
19 | * | |
20 | * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, | |
21 | * CA 95054 USA or visit www.sun.com if you need additional information or | |
22 | * have any questions. | |
23 | * | |
24 | * GPL HEADER END | |
25 | */ | |
26 | /* | |
27 | * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved. | |
28 | * Use is subject to license terms. | |
29 | * | |
1dc563a6 | 30 | * Copyright (c) 2011, 2015, Intel Corporation. |
d7e09d03 PT |
31 | */ |
32 | /* | |
33 | * This file is part of Lustre, http://www.lustre.org/ | |
34 | * Lustre is a trademark of Sun Microsystems, Inc. | |
35 | * | |
36 | * lustre/llite/rw.c | |
37 | * | |
38 | * Lustre Lite I/O page cache routines shared by different kernel revs | |
39 | */ | |
40 | ||
41 | #include <linux/kernel.h> | |
42 | #include <linux/mm.h> | |
43 | #include <linux/string.h> | |
44 | #include <linux/stat.h> | |
45 | #include <linux/errno.h> | |
46 | #include <linux/unistd.h> | |
47 | #include <linux/writeback.h> | |
e8fd99fd | 48 | #include <linux/uaccess.h> |
d7e09d03 PT |
49 | |
50 | #include <linux/fs.h> | |
d7e09d03 PT |
51 | #include <linux/pagemap.h> |
52 | /* current_is_kswapd() */ | |
53 | #include <linux/swap.h> | |
54 | ||
55 | #define DEBUG_SUBSYSTEM S_LLITE | |
56 | ||
67a235f5 GKH |
57 | #include "../include/lustre_lite.h" |
58 | #include "../include/obd_cksum.h" | |
d7e09d03 | 59 | #include "llite_internal.h" |
67a235f5 | 60 | #include "../include/linux/lustre_compat25.h" |
d7e09d03 PT |
61 | |
62 | /** | |
63 | * Finalizes cl-data before exiting typical address_space operation. Dual to | |
64 | * ll_cl_init(). | |
65 | */ | |
66 | static void ll_cl_fini(struct ll_cl_context *lcc) | |
67 | { | |
68 | struct lu_env *env = lcc->lcc_env; | |
69 | struct cl_io *io = lcc->lcc_io; | |
70 | struct cl_page *page = lcc->lcc_page; | |
71 | ||
72 | LASSERT(lcc->lcc_cookie == current); | |
6e16818b | 73 | LASSERT(env); |
d7e09d03 | 74 | |
6e16818b | 75 | if (page) { |
d7e09d03 PT |
76 | lu_ref_del(&page->cp_reference, "cl_io", io); |
77 | cl_page_put(env, page); | |
78 | } | |
79 | ||
d7e09d03 PT |
80 | cl_env_put(env, &lcc->lcc_refcheck); |
81 | } | |
82 | ||
83 | /** | |
84 | * Initializes common cl-data at the typical address_space operation entry | |
85 | * point. | |
86 | */ | |
87 | static struct ll_cl_context *ll_cl_init(struct file *file, | |
88 | struct page *vmpage, int create) | |
89 | { | |
90 | struct ll_cl_context *lcc; | |
91 | struct lu_env *env; | |
92 | struct cl_io *io; | |
93 | struct cl_object *clob; | |
94 | struct ccc_io *cio; | |
95 | ||
96 | int refcheck; | |
97 | int result = 0; | |
98 | ||
99 | clob = ll_i2info(vmpage->mapping->host)->lli_clob; | |
6e16818b | 100 | LASSERT(clob); |
d7e09d03 PT |
101 | |
102 | env = cl_env_get(&refcheck); | |
103 | if (IS_ERR(env)) | |
1dc8548c | 104 | return ERR_CAST(env); |
d7e09d03 PT |
105 | |
106 | lcc = &vvp_env_info(env)->vti_io_ctx; | |
107 | memset(lcc, 0, sizeof(*lcc)); | |
108 | lcc->lcc_env = env; | |
109 | lcc->lcc_refcheck = refcheck; | |
110 | lcc->lcc_cookie = current; | |
111 | ||
112 | cio = ccc_env_io(env); | |
113 | io = cio->cui_cl.cis_io; | |
6e16818b | 114 | if (!io && create) { |
d7e09d03 PT |
115 | struct inode *inode = vmpage->mapping->host; |
116 | loff_t pos; | |
117 | ||
5955102c AV |
118 | if (inode_trylock(inode)) { |
119 | inode_unlock((inode)); | |
d7e09d03 PT |
120 | |
121 | /* this is too bad. Someone is trying to write the | |
122 | * page w/o holding inode mutex. This means we can | |
c0894c6c OD |
123 | * add dirty pages into cache during truncate |
124 | */ | |
2d00bd17 JP |
125 | CERROR("Proc %s is dirtying page w/o inode lock, this will break truncate\n", |
126 | current->comm); | |
5d4450c4 | 127 | dump_stack(); |
d7e09d03 PT |
128 | LBUG(); |
129 | return ERR_PTR(-EIO); | |
130 | } | |
131 | ||
132 | /* | |
74c0da19 | 133 | * Loop-back driver calls ->prepare_write(). |
d7e09d03 PT |
134 | * methods directly, bypassing file system ->write() operation, |
135 | * so cl_io has to be created here. | |
136 | */ | |
137 | io = ccc_env_thread_io(env); | |
138 | ll_io_init(io, file, 1); | |
139 | ||
140 | /* No lock at all for this kind of IO - we can't do it because | |
141 | * we have held page lock, it would cause deadlock. | |
142 | * XXX: This causes poor performance to loop device - One page | |
143 | * per RPC. | |
144 | * In order to get better performance, users should use | |
145 | * lloop driver instead. | |
146 | */ | |
147 | io->ci_lockreq = CILR_NEVER; | |
148 | ||
09cbfeaf | 149 | pos = vmpage->index << PAGE_SHIFT; |
d7e09d03 PT |
150 | |
151 | /* Create a temp IO to serve write. */ | |
09cbfeaf | 152 | result = cl_io_rw_init(env, io, CIT_WRITE, pos, PAGE_SIZE); |
d7e09d03 PT |
153 | if (result == 0) { |
154 | cio->cui_fd = LUSTRE_FPRIVATE(file); | |
b42b15fd | 155 | cio->cui_iter = NULL; |
d7e09d03 PT |
156 | result = cl_io_iter_init(env, io); |
157 | if (result == 0) { | |
158 | result = cl_io_lock(env, io); | |
159 | if (result == 0) | |
160 | result = cl_io_start(env, io); | |
161 | } | |
162 | } else | |
163 | result = io->ci_result; | |
d7e09d03 PT |
164 | } |
165 | ||
166 | lcc->lcc_io = io; | |
6e16818b | 167 | if (!io) |
d7e09d03 PT |
168 | result = -EIO; |
169 | if (result == 0) { | |
170 | struct cl_page *page; | |
171 | ||
d7e09d03 PT |
172 | LASSERT(io->ci_state == CIS_IO_GOING); |
173 | LASSERT(cio->cui_fd == LUSTRE_FPRIVATE(file)); | |
174 | page = cl_page_find(env, clob, vmpage->index, vmpage, | |
175 | CPT_CACHEABLE); | |
176 | if (!IS_ERR(page)) { | |
177 | lcc->lcc_page = page; | |
178 | lu_ref_add(&page->cp_reference, "cl_io", io); | |
179 | result = 0; | |
180 | } else | |
181 | result = PTR_ERR(page); | |
182 | } | |
183 | if (result) { | |
184 | ll_cl_fini(lcc); | |
185 | lcc = ERR_PTR(result); | |
186 | } | |
187 | ||
188 | CDEBUG(D_VFSTRACE, "%lu@"DFID" -> %d %p %p\n", | |
189 | vmpage->index, PFID(lu_object_fid(&clob->co_lu)), result, | |
190 | env, io); | |
191 | return lcc; | |
192 | } | |
193 | ||
194 | static struct ll_cl_context *ll_cl_get(void) | |
195 | { | |
196 | struct ll_cl_context *lcc; | |
197 | struct lu_env *env; | |
198 | int refcheck; | |
199 | ||
200 | env = cl_env_get(&refcheck); | |
201 | LASSERT(!IS_ERR(env)); | |
202 | lcc = &vvp_env_info(env)->vti_io_ctx; | |
203 | LASSERT(env == lcc->lcc_env); | |
204 | LASSERT(current == lcc->lcc_cookie); | |
205 | cl_env_put(env, &refcheck); | |
206 | ||
207 | /* env has got in ll_cl_init, so it is still usable. */ | |
208 | return lcc; | |
209 | } | |
210 | ||
211 | /** | |
212 | * ->prepare_write() address space operation called by generic_file_write() | |
213 | * for every page during write. | |
214 | */ | |
215 | int ll_prepare_write(struct file *file, struct page *vmpage, unsigned from, | |
216 | unsigned to) | |
217 | { | |
218 | struct ll_cl_context *lcc; | |
219 | int result; | |
d7e09d03 PT |
220 | |
221 | lcc = ll_cl_init(file, vmpage, 1); | |
222 | if (!IS_ERR(lcc)) { | |
223 | struct lu_env *env = lcc->lcc_env; | |
224 | struct cl_io *io = lcc->lcc_io; | |
225 | struct cl_page *page = lcc->lcc_page; | |
226 | ||
227 | cl_page_assume(env, io, page); | |
228 | ||
229 | result = cl_io_prepare_write(env, io, page, from, to); | |
230 | if (result == 0) { | |
231 | /* | |
232 | * Add a reference, so that page is not evicted from | |
233 | * the cache until ->commit_write() is called. | |
234 | */ | |
235 | cl_page_get(page); | |
236 | lu_ref_add(&page->cp_reference, "prepare_write", | |
237 | current); | |
238 | } else { | |
239 | cl_page_unassume(env, io, page); | |
240 | ll_cl_fini(lcc); | |
241 | } | |
242 | /* returning 0 in prepare assumes commit must be called | |
c0894c6c OD |
243 | * afterwards |
244 | */ | |
d7e09d03 PT |
245 | } else { |
246 | result = PTR_ERR(lcc); | |
247 | } | |
0a3bdb00 | 248 | return result; |
d7e09d03 PT |
249 | } |
250 | ||
251 | int ll_commit_write(struct file *file, struct page *vmpage, unsigned from, | |
252 | unsigned to) | |
253 | { | |
254 | struct ll_cl_context *lcc; | |
255 | struct lu_env *env; | |
256 | struct cl_io *io; | |
257 | struct cl_page *page; | |
258 | int result = 0; | |
d7e09d03 PT |
259 | |
260 | lcc = ll_cl_get(); | |
261 | env = lcc->lcc_env; | |
262 | page = lcc->lcc_page; | |
263 | io = lcc->lcc_io; | |
264 | ||
265 | LASSERT(cl_page_is_owned(page, io)); | |
266 | LASSERT(from <= to); | |
267 | if (from != to) /* handle short write case. */ | |
268 | result = cl_io_commit_write(env, io, page, from, to); | |
269 | if (cl_page_is_owned(page, io)) | |
270 | cl_page_unassume(env, io, page); | |
271 | ||
272 | /* | |
273 | * Release reference acquired by ll_prepare_write(). | |
274 | */ | |
275 | lu_ref_del(&page->cp_reference, "prepare_write", current); | |
276 | cl_page_put(env, page); | |
277 | ll_cl_fini(lcc); | |
0a3bdb00 | 278 | return result; |
d7e09d03 PT |
279 | } |
280 | ||
d7e09d03 PT |
281 | static void ll_ra_stats_inc_sbi(struct ll_sb_info *sbi, enum ra_stat which); |
282 | ||
283 | /** | |
284 | * Get readahead pages from the filesystem readahead pool of the client for a | |
285 | * thread. | |
286 | * | |
287 | * /param sbi superblock for filesystem readahead state ll_ra_info | |
288 | * /param ria per-thread readahead state | |
289 | * /param pages number of pages requested for readahead for the thread. | |
290 | * | |
291 | * WARNING: This algorithm is used to reduce contention on sbi->ll_lock. | |
292 | * It should work well if the ra_max_pages is much greater than the single | |
293 | * file's read-ahead window, and not too many threads contending for | |
294 | * these readahead pages. | |
295 | * | |
296 | * TODO: There may be a 'global sync problem' if many threads are trying | |
297 | * to get an ra budget that is larger than the remaining readahead pages | |
298 | * and reach here at exactly the same time. They will compute /a ret to | |
299 | * consume the remaining pages, but will fail at atomic_add_return() and | |
c0894c6c OD |
300 | * get a zero ra window, although there is still ra space remaining. - Jay |
301 | */ | |
d7e09d03 PT |
302 | static unsigned long ll_ra_count_get(struct ll_sb_info *sbi, |
303 | struct ra_io_arg *ria, | |
304 | unsigned long pages) | |
305 | { | |
306 | struct ll_ra_info *ra = &sbi->ll_ra_info; | |
307 | long ret; | |
d7e09d03 PT |
308 | |
309 | /* If read-ahead pages left are less than 1M, do not do read-ahead, | |
310 | * otherwise it will form small read RPC(< 1M), which hurt server | |
c0894c6c OD |
311 | * performance a lot. |
312 | */ | |
d7e09d03 | 313 | ret = min(ra->ra_max_pages - atomic_read(&ra->ra_cur_pages), pages); |
34e1f2bb JL |
314 | if (ret < 0 || ret < min_t(long, PTLRPC_MAX_BRW_PAGES, pages)) { |
315 | ret = 0; | |
316 | goto out; | |
317 | } | |
d7e09d03 PT |
318 | |
319 | /* If the non-strided (ria_pages == 0) readahead window | |
320 | * (ria_start + ret) has grown across an RPC boundary, then trim | |
321 | * readahead size by the amount beyond the RPC so it ends on an | |
322 | * RPC boundary. If the readahead window is already ending on | |
323 | * an RPC boundary (beyond_rpc == 0), or smaller than a full | |
324 | * RPC (beyond_rpc < ret) the readahead size is unchanged. | |
325 | * The (beyond_rpc != 0) check is skipped since the conditional | |
326 | * branch is more expensive than subtracting zero from the result. | |
327 | * | |
328 | * Strided read is left unaligned to avoid small fragments beyond | |
c0894c6c OD |
329 | * the RPC boundary from needing an extra read RPC. |
330 | */ | |
d7e09d03 PT |
331 | if (ria->ria_pages == 0) { |
332 | long beyond_rpc = (ria->ria_start + ret) % PTLRPC_MAX_BRW_PAGES; | |
50ffcb7e | 333 | |
d7e09d03 PT |
334 | if (/* beyond_rpc != 0 && */ beyond_rpc < ret) |
335 | ret -= beyond_rpc; | |
336 | } | |
337 | ||
338 | if (atomic_add_return(ret, &ra->ra_cur_pages) > ra->ra_max_pages) { | |
339 | atomic_sub(ret, &ra->ra_cur_pages); | |
340 | ret = 0; | |
341 | } | |
342 | ||
343 | out: | |
0a3bdb00 | 344 | return ret; |
d7e09d03 PT |
345 | } |
346 | ||
347 | void ll_ra_count_put(struct ll_sb_info *sbi, unsigned long len) | |
348 | { | |
349 | struct ll_ra_info *ra = &sbi->ll_ra_info; | |
50ffcb7e | 350 | |
d7e09d03 PT |
351 | atomic_sub(len, &ra->ra_cur_pages); |
352 | } | |
353 | ||
354 | static void ll_ra_stats_inc_sbi(struct ll_sb_info *sbi, enum ra_stat which) | |
355 | { | |
356 | LASSERTF(which >= 0 && which < _NR_RA_STAT, "which: %u\n", which); | |
357 | lprocfs_counter_incr(sbi->ll_ra_stats, which); | |
358 | } | |
359 | ||
360 | void ll_ra_stats_inc(struct address_space *mapping, enum ra_stat which) | |
361 | { | |
362 | struct ll_sb_info *sbi = ll_i2sbi(mapping->host); | |
50ffcb7e | 363 | |
d7e09d03 PT |
364 | ll_ra_stats_inc_sbi(sbi, which); |
365 | } | |
366 | ||
367 | #define RAS_CDEBUG(ras) \ | |
368 | CDEBUG(D_READA, \ | |
369 | "lrp %lu cr %lu cp %lu ws %lu wl %lu nra %lu r %lu ri %lu" \ | |
72a87fca | 370 | "csr %lu sf %lu sp %lu sl %lu\n", \ |
d7e09d03 PT |
371 | ras->ras_last_readpage, ras->ras_consecutive_requests, \ |
372 | ras->ras_consecutive_pages, ras->ras_window_start, \ | |
373 | ras->ras_window_len, ras->ras_next_readahead, \ | |
374 | ras->ras_requests, ras->ras_request_index, \ | |
375 | ras->ras_consecutive_stride_requests, ras->ras_stride_offset, \ | |
376 | ras->ras_stride_pages, ras->ras_stride_length) | |
377 | ||
378 | static int index_in_window(unsigned long index, unsigned long point, | |
379 | unsigned long before, unsigned long after) | |
380 | { | |
381 | unsigned long start = point - before, end = point + after; | |
382 | ||
383 | if (start > point) | |
defa220f | 384 | start = 0; |
d7e09d03 | 385 | if (end < point) |
defa220f | 386 | end = ~0; |
d7e09d03 PT |
387 | |
388 | return start <= index && index <= end; | |
389 | } | |
390 | ||
391 | static struct ll_readahead_state *ll_ras_get(struct file *f) | |
392 | { | |
393 | struct ll_file_data *fd; | |
394 | ||
395 | fd = LUSTRE_FPRIVATE(f); | |
396 | return &fd->fd_ras; | |
397 | } | |
398 | ||
399 | void ll_ra_read_in(struct file *f, struct ll_ra_read *rar) | |
400 | { | |
401 | struct ll_readahead_state *ras; | |
402 | ||
403 | ras = ll_ras_get(f); | |
404 | ||
405 | spin_lock(&ras->ras_lock); | |
406 | ras->ras_requests++; | |
407 | ras->ras_request_index = 0; | |
408 | ras->ras_consecutive_requests++; | |
409 | rar->lrr_reader = current; | |
410 | ||
411 | list_add(&rar->lrr_linkage, &ras->ras_read_beads); | |
412 | spin_unlock(&ras->ras_lock); | |
413 | } | |
414 | ||
415 | void ll_ra_read_ex(struct file *f, struct ll_ra_read *rar) | |
416 | { | |
417 | struct ll_readahead_state *ras; | |
418 | ||
419 | ras = ll_ras_get(f); | |
420 | ||
421 | spin_lock(&ras->ras_lock); | |
422 | list_del_init(&rar->lrr_linkage); | |
423 | spin_unlock(&ras->ras_lock); | |
424 | } | |
425 | ||
d7e09d03 PT |
426 | static int cl_read_ahead_page(const struct lu_env *env, struct cl_io *io, |
427 | struct cl_page_list *queue, struct cl_page *page, | |
428 | struct page *vmpage) | |
429 | { | |
430 | struct ccc_page *cp; | |
431 | int rc; | |
432 | ||
d7e09d03 PT |
433 | rc = 0; |
434 | cl_page_assume(env, io, page); | |
435 | lu_ref_add(&page->cp_reference, "ra", current); | |
436 | cp = cl2ccc_page(cl_page_at(page, &vvp_device_type)); | |
437 | if (!cp->cpg_defer_uptodate && !PageUptodate(vmpage)) { | |
438 | rc = cl_page_is_under_lock(env, io, page); | |
439 | if (rc == -EBUSY) { | |
440 | cp->cpg_defer_uptodate = 1; | |
441 | cp->cpg_ra_used = 0; | |
442 | cl_page_list_add(queue, page); | |
443 | rc = 1; | |
444 | } else { | |
445 | cl_page_delete(env, page); | |
446 | rc = -ENOLCK; | |
447 | } | |
448 | } else { | |
449 | /* skip completed pages */ | |
450 | cl_page_unassume(env, io, page); | |
451 | } | |
452 | lu_ref_del(&page->cp_reference, "ra", current); | |
453 | cl_page_put(env, page); | |
0a3bdb00 | 454 | return rc; |
d7e09d03 PT |
455 | } |
456 | ||
457 | /** | |
458 | * Initiates read-ahead of a page with given index. | |
459 | * | |
460 | * \retval +ve: page was added to \a queue. | |
461 | * | |
462 | * \retval -ENOLCK: there is no extent lock for this part of a file, stop | |
463 | * read-ahead. | |
464 | * | |
465 | * \retval -ve, 0: page wasn't added to \a queue for other reason. | |
466 | */ | |
467 | static int ll_read_ahead_page(const struct lu_env *env, struct cl_io *io, | |
468 | struct cl_page_list *queue, | |
469 | pgoff_t index, struct address_space *mapping) | |
470 | { | |
471 | struct page *vmpage; | |
472 | struct cl_object *clob = ll_i2info(mapping->host)->lli_clob; | |
473 | struct cl_page *page; | |
474 | enum ra_stat which = _NR_RA_STAT; /* keep gcc happy */ | |
d7e09d03 PT |
475 | int rc = 0; |
476 | const char *msg = NULL; | |
477 | ||
d7e09d03 | 478 | vmpage = grab_cache_page_nowait(mapping, index); |
6e16818b | 479 | if (vmpage) { |
d7e09d03 PT |
480 | /* Check if vmpage was truncated or reclaimed */ |
481 | if (vmpage->mapping == mapping) { | |
482 | page = cl_page_find(env, clob, vmpage->index, | |
483 | vmpage, CPT_CACHEABLE); | |
484 | if (!IS_ERR(page)) { | |
485 | rc = cl_read_ahead_page(env, io, queue, | |
486 | page, vmpage); | |
487 | if (rc == -ENOLCK) { | |
488 | which = RA_STAT_FAILED_MATCH; | |
489 | msg = "lock match failed"; | |
490 | } | |
491 | } else { | |
492 | which = RA_STAT_FAILED_GRAB_PAGE; | |
493 | msg = "cl_page_find failed"; | |
494 | } | |
495 | } else { | |
496 | which = RA_STAT_WRONG_GRAB_PAGE; | |
497 | msg = "g_c_p_n returned invalid page"; | |
498 | } | |
499 | if (rc != 1) | |
500 | unlock_page(vmpage); | |
09cbfeaf | 501 | put_page(vmpage); |
d7e09d03 PT |
502 | } else { |
503 | which = RA_STAT_FAILED_GRAB_PAGE; | |
504 | msg = "g_c_p_n failed"; | |
505 | } | |
6e16818b | 506 | if (msg) { |
d7e09d03 PT |
507 | ll_ra_stats_inc(mapping, which); |
508 | CDEBUG(D_READA, "%s\n", msg); | |
509 | } | |
0a3bdb00 | 510 | return rc; |
d7e09d03 PT |
511 | } |
512 | ||
513 | #define RIA_DEBUG(ria) \ | |
514 | CDEBUG(D_READA, "rs %lu re %lu ro %lu rl %lu rp %lu\n", \ | |
515 | ria->ria_start, ria->ria_end, ria->ria_stoff, ria->ria_length,\ | |
516 | ria->ria_pages) | |
517 | ||
518 | /* Limit this to the blocksize instead of PTLRPC_BRW_MAX_SIZE, since we don't | |
519 | * know what the actual RPC size is. If this needs to change, it makes more | |
520 | * sense to tune the i_blkbits value for the file based on the OSTs it is | |
c0894c6c OD |
521 | * striped over, rather than having a constant value for all files here. |
522 | */ | |
d7e09d03 | 523 | |
ea1754a0 | 524 | /* RAS_INCREASE_STEP should be (1UL << (inode->i_blkbits - PAGE_SHIFT)). |
d0a0acc3 | 525 | * Temporarily set RAS_INCREASE_STEP to 1MB. After 4MB RPC is enabled |
d7e09d03 PT |
526 | * by default, this should be adjusted corresponding with max_read_ahead_mb |
527 | * and max_read_ahead_per_file_mb otherwise the readahead budget can be used | |
c0894c6c OD |
528 | * up quickly which will affect read performance significantly. See LU-2816 |
529 | */ | |
09cbfeaf | 530 | #define RAS_INCREASE_STEP(inode) (ONE_MB_BRW_SIZE >> PAGE_SHIFT) |
d7e09d03 PT |
531 | |
532 | static inline int stride_io_mode(struct ll_readahead_state *ras) | |
533 | { | |
534 | return ras->ras_consecutive_stride_requests > 1; | |
535 | } | |
c9f6bb96 | 536 | |
d7e09d03 PT |
537 | /* The function calculates how much pages will be read in |
538 | * [off, off + length], in such stride IO area, | |
d0a0acc3 | 539 | * stride_offset = st_off, stride_length = st_len, |
d7e09d03 PT |
540 | * stride_pages = st_pgs |
541 | * | |
542 | * |------------------|*****|------------------|*****|------------|*****|.... | |
543 | * st_off | |
544 | * |--- st_pgs ---| | |
545 | * |----- st_len -----| | |
546 | * | |
547 | * How many pages it should read in such pattern | |
548 | * |-------------------------------------------------------------| | |
549 | * off | |
550 | * |<------ length ------->| | |
551 | * | |
552 | * = |<----->| + |-------------------------------------| + |---| | |
553 | * start_left st_pgs * i end_left | |
554 | */ | |
555 | static unsigned long | |
556 | stride_pg_count(pgoff_t st_off, unsigned long st_len, unsigned long st_pgs, | |
557 | unsigned long off, unsigned long length) | |
558 | { | |
559 | __u64 start = off > st_off ? off - st_off : 0; | |
560 | __u64 end = off + length > st_off ? off + length - st_off : 0; | |
561 | unsigned long start_left = 0; | |
562 | unsigned long end_left = 0; | |
563 | unsigned long pg_count; | |
564 | ||
565 | if (st_len == 0 || length == 0 || end == 0) | |
566 | return length; | |
567 | ||
568 | start_left = do_div(start, st_len); | |
569 | if (start_left < st_pgs) | |
570 | start_left = st_pgs - start_left; | |
571 | else | |
572 | start_left = 0; | |
573 | ||
574 | end_left = do_div(end, st_len); | |
575 | if (end_left > st_pgs) | |
576 | end_left = st_pgs; | |
577 | ||
72a87fca | 578 | CDEBUG(D_READA, "start %llu, end %llu start_left %lu end_left %lu\n", |
d7e09d03 PT |
579 | start, end, start_left, end_left); |
580 | ||
581 | if (start == end) | |
582 | pg_count = end_left - (st_pgs - start_left); | |
583 | else | |
584 | pg_count = start_left + st_pgs * (end - start - 1) + end_left; | |
585 | ||
2d00bd17 JP |
586 | CDEBUG(D_READA, "st_off %lu, st_len %lu st_pgs %lu off %lu length %lu pgcount %lu\n", |
587 | st_off, st_len, st_pgs, off, length, pg_count); | |
d7e09d03 PT |
588 | |
589 | return pg_count; | |
590 | } | |
591 | ||
592 | static int ria_page_count(struct ra_io_arg *ria) | |
593 | { | |
594 | __u64 length = ria->ria_end >= ria->ria_start ? | |
595 | ria->ria_end - ria->ria_start + 1 : 0; | |
596 | ||
597 | return stride_pg_count(ria->ria_stoff, ria->ria_length, | |
598 | ria->ria_pages, ria->ria_start, | |
599 | length); | |
600 | } | |
601 | ||
602 | /*Check whether the index is in the defined ra-window */ | |
603 | static int ras_inside_ra_window(unsigned long idx, struct ra_io_arg *ria) | |
604 | { | |
605 | /* If ria_length == ria_pages, it means non-stride I/O mode, | |
606 | * idx should always inside read-ahead window in this case | |
607 | * For stride I/O mode, just check whether the idx is inside | |
c0894c6c OD |
608 | * the ria_pages. |
609 | */ | |
d7e09d03 PT |
610 | return ria->ria_length == 0 || ria->ria_length == ria->ria_pages || |
611 | (idx >= ria->ria_stoff && (idx - ria->ria_stoff) % | |
612 | ria->ria_length < ria->ria_pages); | |
613 | } | |
614 | ||
615 | static int ll_read_ahead_pages(const struct lu_env *env, | |
616 | struct cl_io *io, struct cl_page_list *queue, | |
617 | struct ra_io_arg *ria, | |
618 | unsigned long *reserved_pages, | |
619 | struct address_space *mapping, | |
620 | unsigned long *ra_end) | |
621 | { | |
622 | int rc, count = 0, stride_ria; | |
623 | unsigned long page_idx; | |
624 | ||
6e16818b | 625 | LASSERT(ria); |
d7e09d03 PT |
626 | RIA_DEBUG(ria); |
627 | ||
628 | stride_ria = ria->ria_length > ria->ria_pages && ria->ria_pages > 0; | |
e15ba45d OD |
629 | for (page_idx = ria->ria_start; |
630 | page_idx <= ria->ria_end && *reserved_pages > 0; page_idx++) { | |
d7e09d03 PT |
631 | if (ras_inside_ra_window(page_idx, ria)) { |
632 | /* If the page is inside the read-ahead window*/ | |
633 | rc = ll_read_ahead_page(env, io, queue, | |
634 | page_idx, mapping); | |
635 | if (rc == 1) { | |
636 | (*reserved_pages)--; | |
b2952d62 | 637 | count++; |
d7e09d03 PT |
638 | } else if (rc == -ENOLCK) |
639 | break; | |
640 | } else if (stride_ria) { | |
641 | /* If it is not in the read-ahead window, and it is | |
642 | * read-ahead mode, then check whether it should skip | |
c0894c6c OD |
643 | * the stride gap |
644 | */ | |
d7e09d03 PT |
645 | pgoff_t offset; |
646 | /* FIXME: This assertion only is valid when it is for | |
647 | * forward read-ahead, it will be fixed when backward | |
c0894c6c OD |
648 | * read-ahead is implemented |
649 | */ | |
2d00bd17 JP |
650 | LASSERTF(page_idx > ria->ria_stoff, "Invalid page_idx %lu rs %lu re %lu ro %lu rl %lu rp %lu\n", |
651 | page_idx, | |
652 | ria->ria_start, ria->ria_end, ria->ria_stoff, | |
653 | ria->ria_length, ria->ria_pages); | |
d7e09d03 PT |
654 | offset = page_idx - ria->ria_stoff; |
655 | offset = offset % (ria->ria_length); | |
656 | if (offset > ria->ria_pages) { | |
657 | page_idx += ria->ria_length - offset; | |
72a87fca | 658 | CDEBUG(D_READA, "i %lu skip %lu\n", page_idx, |
d7e09d03 PT |
659 | ria->ria_length - offset); |
660 | continue; | |
661 | } | |
662 | } | |
663 | } | |
664 | *ra_end = page_idx; | |
665 | return count; | |
666 | } | |
667 | ||
668 | int ll_readahead(const struct lu_env *env, struct cl_io *io, | |
669 | struct ll_readahead_state *ras, struct address_space *mapping, | |
670 | struct cl_page_list *queue, int flags) | |
671 | { | |
672 | struct vvp_io *vio = vvp_env_io(env); | |
673 | struct vvp_thread_info *vti = vvp_env_info(env); | |
674 | struct cl_attr *attr = ccc_env_thread_attr(env); | |
675 | unsigned long start = 0, end = 0, reserved; | |
676 | unsigned long ra_end, len; | |
677 | struct inode *inode; | |
678 | struct ll_ra_read *bead; | |
679 | struct ra_io_arg *ria = &vti->vti_ria; | |
680 | struct ll_inode_info *lli; | |
681 | struct cl_object *clob; | |
682 | int ret = 0; | |
683 | __u64 kms; | |
d7e09d03 PT |
684 | |
685 | inode = mapping->host; | |
686 | lli = ll_i2info(inode); | |
687 | clob = lli->lli_clob; | |
688 | ||
ec83e611 | 689 | memset(ria, 0, sizeof(*ria)); |
d7e09d03 PT |
690 | |
691 | cl_object_attr_lock(clob); | |
692 | ret = cl_object_attr_get(env, clob, attr); | |
693 | cl_object_attr_unlock(clob); | |
694 | ||
695 | if (ret != 0) | |
0a3bdb00 | 696 | return ret; |
d7e09d03 PT |
697 | kms = attr->cat_kms; |
698 | if (kms == 0) { | |
699 | ll_ra_stats_inc(mapping, RA_STAT_ZERO_LEN); | |
0a3bdb00 | 700 | return 0; |
d7e09d03 PT |
701 | } |
702 | ||
703 | spin_lock(&ras->ras_lock); | |
704 | if (vio->cui_ra_window_set) | |
705 | bead = &vio->cui_bead; | |
706 | else | |
707 | bead = NULL; | |
708 | ||
709 | /* Enlarge the RA window to encompass the full read */ | |
6e16818b | 710 | if (bead && ras->ras_window_start + ras->ras_window_len < |
d7e09d03 PT |
711 | bead->lrr_start + bead->lrr_count) { |
712 | ras->ras_window_len = bead->lrr_start + bead->lrr_count - | |
713 | ras->ras_window_start; | |
714 | } | |
715 | /* Reserve a part of the read-ahead window that we'll be issuing */ | |
716 | if (ras->ras_window_len) { | |
717 | start = ras->ras_next_readahead; | |
718 | end = ras->ras_window_start + ras->ras_window_len - 1; | |
719 | } | |
720 | if (end != 0) { | |
721 | unsigned long rpc_boundary; | |
722 | /* | |
723 | * Align RA window to an optimal boundary. | |
724 | * | |
725 | * XXX This would be better to align to cl_max_pages_per_rpc | |
726 | * instead of PTLRPC_MAX_BRW_PAGES, because the RPC size may | |
727 | * be aligned to the RAID stripe size in the future and that | |
728 | * is more important than the RPC size. | |
729 | */ | |
730 | /* Note: we only trim the RPC, instead of extending the RPC | |
731 | * to the boundary, so to avoid reading too much pages during | |
c0894c6c OD |
732 | * random reading. |
733 | */ | |
b6ee3824 | 734 | rpc_boundary = (end + 1) & (~(PTLRPC_MAX_BRW_PAGES - 1)); |
d7e09d03 PT |
735 | if (rpc_boundary > 0) |
736 | rpc_boundary--; | |
737 | ||
738 | if (rpc_boundary > start) | |
739 | end = rpc_boundary; | |
740 | ||
741 | /* Truncate RA window to end of file */ | |
09cbfeaf | 742 | end = min(end, (unsigned long)((kms - 1) >> PAGE_SHIFT)); |
d7e09d03 PT |
743 | |
744 | ras->ras_next_readahead = max(end, end + 1); | |
745 | RAS_CDEBUG(ras); | |
746 | } | |
747 | ria->ria_start = start; | |
748 | ria->ria_end = end; | |
749 | /* If stride I/O mode is detected, get stride window*/ | |
750 | if (stride_io_mode(ras)) { | |
751 | ria->ria_stoff = ras->ras_stride_offset; | |
752 | ria->ria_length = ras->ras_stride_length; | |
753 | ria->ria_pages = ras->ras_stride_pages; | |
754 | } | |
755 | spin_unlock(&ras->ras_lock); | |
756 | ||
757 | if (end == 0) { | |
758 | ll_ra_stats_inc(mapping, RA_STAT_ZERO_WINDOW); | |
0a3bdb00 | 759 | return 0; |
d7e09d03 PT |
760 | } |
761 | len = ria_page_count(ria); | |
762 | if (len == 0) | |
0a3bdb00 | 763 | return 0; |
d7e09d03 PT |
764 | |
765 | reserved = ll_ra_count_get(ll_i2sbi(inode), ria, len); | |
766 | if (reserved < len) | |
767 | ll_ra_stats_inc(mapping, RA_STAT_MAX_IN_FLIGHT); | |
768 | ||
769 | CDEBUG(D_READA, "reserved page %lu ra_cur %d ra_max %lu\n", reserved, | |
770 | atomic_read(&ll_i2sbi(inode)->ll_ra_info.ra_cur_pages), | |
771 | ll_i2sbi(inode)->ll_ra_info.ra_max_pages); | |
772 | ||
773 | ret = ll_read_ahead_pages(env, io, queue, | |
774 | ria, &reserved, mapping, &ra_end); | |
775 | ||
d7e09d03 PT |
776 | if (reserved != 0) |
777 | ll_ra_count_put(ll_i2sbi(inode), reserved); | |
778 | ||
09cbfeaf | 779 | if (ra_end == end + 1 && ra_end == (kms >> PAGE_SHIFT)) |
d7e09d03 PT |
780 | ll_ra_stats_inc(mapping, RA_STAT_EOF); |
781 | ||
782 | /* if we didn't get to the end of the region we reserved from | |
783 | * the ras we need to go back and update the ras so that the | |
784 | * next read-ahead tries from where we left off. we only do so | |
785 | * if the region we failed to issue read-ahead on is still ahead | |
c0894c6c OD |
786 | * of the app and behind the next index to start read-ahead from |
787 | */ | |
72a87fca | 788 | CDEBUG(D_READA, "ra_end %lu end %lu stride end %lu\n", |
d7e09d03 PT |
789 | ra_end, end, ria->ria_end); |
790 | ||
791 | if (ra_end != end + 1) { | |
792 | spin_lock(&ras->ras_lock); | |
793 | if (ra_end < ras->ras_next_readahead && | |
794 | index_in_window(ra_end, ras->ras_window_start, 0, | |
795 | ras->ras_window_len)) { | |
796 | ras->ras_next_readahead = ra_end; | |
797 | RAS_CDEBUG(ras); | |
798 | } | |
799 | spin_unlock(&ras->ras_lock); | |
800 | } | |
801 | ||
0a3bdb00 | 802 | return ret; |
d7e09d03 PT |
803 | } |
804 | ||
805 | static void ras_set_start(struct inode *inode, struct ll_readahead_state *ras, | |
806 | unsigned long index) | |
807 | { | |
808 | ras->ras_window_start = index & (~(RAS_INCREASE_STEP(inode) - 1)); | |
809 | } | |
810 | ||
811 | /* called with the ras_lock held or from places where it doesn't matter */ | |
812 | static void ras_reset(struct inode *inode, struct ll_readahead_state *ras, | |
813 | unsigned long index) | |
814 | { | |
815 | ras->ras_last_readpage = index; | |
816 | ras->ras_consecutive_requests = 0; | |
817 | ras->ras_consecutive_pages = 0; | |
818 | ras->ras_window_len = 0; | |
819 | ras_set_start(inode, ras, index); | |
820 | ras->ras_next_readahead = max(ras->ras_window_start, index); | |
821 | ||
822 | RAS_CDEBUG(ras); | |
823 | } | |
824 | ||
825 | /* called with the ras_lock held or from places where it doesn't matter */ | |
826 | static void ras_stride_reset(struct ll_readahead_state *ras) | |
827 | { | |
828 | ras->ras_consecutive_stride_requests = 0; | |
829 | ras->ras_stride_length = 0; | |
830 | ras->ras_stride_pages = 0; | |
831 | RAS_CDEBUG(ras); | |
832 | } | |
833 | ||
834 | void ll_readahead_init(struct inode *inode, struct ll_readahead_state *ras) | |
835 | { | |
836 | spin_lock_init(&ras->ras_lock); | |
837 | ras_reset(inode, ras, 0); | |
838 | ras->ras_requests = 0; | |
839 | INIT_LIST_HEAD(&ras->ras_read_beads); | |
840 | } | |
841 | ||
842 | /* | |
843 | * Check whether the read request is in the stride window. | |
844 | * If it is in the stride window, return 1, otherwise return 0. | |
845 | */ | |
846 | static int index_in_stride_window(struct ll_readahead_state *ras, | |
847 | unsigned long index) | |
848 | { | |
849 | unsigned long stride_gap; | |
850 | ||
851 | if (ras->ras_stride_length == 0 || ras->ras_stride_pages == 0 || | |
852 | ras->ras_stride_pages == ras->ras_stride_length) | |
853 | return 0; | |
854 | ||
855 | stride_gap = index - ras->ras_last_readpage - 1; | |
856 | ||
857 | /* If it is contiguous read */ | |
858 | if (stride_gap == 0) | |
859 | return ras->ras_consecutive_pages + 1 <= ras->ras_stride_pages; | |
860 | ||
861 | /* Otherwise check the stride by itself */ | |
862 | return (ras->ras_stride_length - ras->ras_stride_pages) == stride_gap && | |
863 | ras->ras_consecutive_pages == ras->ras_stride_pages; | |
864 | } | |
865 | ||
866 | static void ras_update_stride_detector(struct ll_readahead_state *ras, | |
867 | unsigned long index) | |
868 | { | |
869 | unsigned long stride_gap = index - ras->ras_last_readpage - 1; | |
870 | ||
871 | if (!stride_io_mode(ras) && (stride_gap != 0 || | |
e15ba45d | 872 | ras->ras_consecutive_stride_requests == 0)) { |
d7e09d03 | 873 | ras->ras_stride_pages = ras->ras_consecutive_pages; |
b2952d62 | 874 | ras->ras_stride_length = stride_gap+ras->ras_consecutive_pages; |
d7e09d03 PT |
875 | } |
876 | LASSERT(ras->ras_request_index == 0); | |
877 | LASSERT(ras->ras_consecutive_stride_requests == 0); | |
878 | ||
879 | if (index <= ras->ras_last_readpage) { | |
880 | /*Reset stride window for forward read*/ | |
881 | ras_stride_reset(ras); | |
882 | return; | |
883 | } | |
884 | ||
885 | ras->ras_stride_pages = ras->ras_consecutive_pages; | |
b2952d62 | 886 | ras->ras_stride_length = stride_gap+ras->ras_consecutive_pages; |
d7e09d03 PT |
887 | |
888 | RAS_CDEBUG(ras); | |
889 | return; | |
890 | } | |
891 | ||
d7e09d03 | 892 | /* Stride Read-ahead window will be increased inc_len according to |
c0894c6c OD |
893 | * stride I/O pattern |
894 | */ | |
d7e09d03 PT |
895 | static void ras_stride_increase_window(struct ll_readahead_state *ras, |
896 | struct ll_ra_info *ra, | |
897 | unsigned long inc_len) | |
898 | { | |
899 | unsigned long left, step, window_len; | |
900 | unsigned long stride_len; | |
901 | ||
902 | LASSERT(ras->ras_stride_length > 0); | |
903 | LASSERTF(ras->ras_window_start + ras->ras_window_len | |
2d00bd17 JP |
904 | >= ras->ras_stride_offset, "window_start %lu, window_len %lu stride_offset %lu\n", |
905 | ras->ras_window_start, | |
d7e09d03 PT |
906 | ras->ras_window_len, ras->ras_stride_offset); |
907 | ||
908 | stride_len = ras->ras_window_start + ras->ras_window_len - | |
909 | ras->ras_stride_offset; | |
910 | ||
911 | left = stride_len % ras->ras_stride_length; | |
912 | window_len = ras->ras_window_len - left; | |
913 | ||
914 | if (left < ras->ras_stride_pages) | |
915 | left += inc_len; | |
916 | else | |
917 | left = ras->ras_stride_pages + inc_len; | |
918 | ||
919 | LASSERT(ras->ras_stride_pages != 0); | |
920 | ||
921 | step = left / ras->ras_stride_pages; | |
922 | left %= ras->ras_stride_pages; | |
923 | ||
924 | window_len += step * ras->ras_stride_length + left; | |
925 | ||
464e5947 SB |
926 | if (stride_pg_count(ras->ras_stride_offset, ras->ras_stride_length, |
927 | ras->ras_stride_pages, ras->ras_stride_offset, | |
928 | window_len) <= ra->ra_max_pages_per_file) | |
d7e09d03 PT |
929 | ras->ras_window_len = window_len; |
930 | ||
931 | RAS_CDEBUG(ras); | |
932 | } | |
933 | ||
934 | static void ras_increase_window(struct inode *inode, | |
935 | struct ll_readahead_state *ras, | |
936 | struct ll_ra_info *ra) | |
937 | { | |
938 | /* The stretch of ra-window should be aligned with max rpc_size | |
939 | * but current clio architecture does not support retrieve such | |
940 | * information from lower layer. FIXME later | |
941 | */ | |
942 | if (stride_io_mode(ras)) | |
943 | ras_stride_increase_window(ras, ra, RAS_INCREASE_STEP(inode)); | |
944 | else | |
945 | ras->ras_window_len = min(ras->ras_window_len + | |
946 | RAS_INCREASE_STEP(inode), | |
947 | ra->ra_max_pages_per_file); | |
948 | } | |
949 | ||
950 | void ras_update(struct ll_sb_info *sbi, struct inode *inode, | |
951 | struct ll_readahead_state *ras, unsigned long index, | |
952 | unsigned hit) | |
953 | { | |
954 | struct ll_ra_info *ra = &sbi->ll_ra_info; | |
955 | int zero = 0, stride_detect = 0, ra_miss = 0; | |
d7e09d03 PT |
956 | |
957 | spin_lock(&ras->ras_lock); | |
958 | ||
959 | ll_ra_stats_inc_sbi(sbi, hit ? RA_STAT_HIT : RA_STAT_MISS); | |
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 is | |
c0894c6c OD |
965 | * reclaiming it before we get to it. |
966 | */ | |
d7e09d03 PT |
967 | if (!index_in_window(index, ras->ras_last_readpage, 8, 8)) { |
968 | zero = 1; | |
969 | ll_ra_stats_inc_sbi(sbi, RA_STAT_DISTANT_READPAGE); | |
970 | } else if (!hit && ras->ras_window_len && | |
971 | index < ras->ras_next_readahead && | |
972 | index_in_window(index, ras->ras_window_start, 0, | |
973 | ras->ras_window_len)) { | |
974 | ra_miss = 1; | |
975 | ll_ra_stats_inc_sbi(sbi, RA_STAT_MISS_IN_WINDOW); | |
976 | } | |
977 | ||
978 | /* On the second access to a file smaller than the tunable | |
979 | * ra_max_read_ahead_whole_pages trigger RA on all pages in the | |
980 | * file up to ra_max_pages_per_file. This is simply a best effort | |
981 | * and only occurs once per open file. Normal RA behavior is reverted | |
982 | * to for subsequent IO. The mmap case does not increment | |
c0894c6c OD |
983 | * ras_requests and thus can never trigger this behavior. |
984 | */ | |
d7e09d03 PT |
985 | if (ras->ras_requests == 2 && !ras->ras_request_index) { |
986 | __u64 kms_pages; | |
987 | ||
09cbfeaf KS |
988 | kms_pages = (i_size_read(inode) + PAGE_SIZE - 1) >> |
989 | PAGE_SHIFT; | |
d7e09d03 | 990 | |
b0f5aad5 | 991 | CDEBUG(D_READA, "kmsp %llu mwp %lu mp %lu\n", kms_pages, |
d7e09d03 PT |
992 | ra->ra_max_read_ahead_whole_pages, ra->ra_max_pages_per_file); |
993 | ||
994 | if (kms_pages && | |
995 | kms_pages <= ra->ra_max_read_ahead_whole_pages) { | |
996 | ras->ras_window_start = 0; | |
997 | ras->ras_last_readpage = 0; | |
998 | ras->ras_next_readahead = 0; | |
999 | ras->ras_window_len = min(ra->ra_max_pages_per_file, | |
1000 | ra->ra_max_read_ahead_whole_pages); | |
34e1f2bb | 1001 | goto out_unlock; |
d7e09d03 PT |
1002 | } |
1003 | } | |
1004 | if (zero) { | |
1005 | /* check whether it is in stride I/O mode*/ | |
1006 | if (!index_in_stride_window(ras, index)) { | |
1007 | if (ras->ras_consecutive_stride_requests == 0 && | |
1008 | ras->ras_request_index == 0) { | |
1009 | ras_update_stride_detector(ras, index); | |
1010 | ras->ras_consecutive_stride_requests++; | |
1011 | } else { | |
1012 | ras_stride_reset(ras); | |
1013 | } | |
1014 | ras_reset(inode, ras, index); | |
1015 | ras->ras_consecutive_pages++; | |
34e1f2bb | 1016 | goto out_unlock; |
d7e09d03 PT |
1017 | } else { |
1018 | ras->ras_consecutive_pages = 0; | |
1019 | ras->ras_consecutive_requests = 0; | |
1020 | if (++ras->ras_consecutive_stride_requests > 1) | |
1021 | stride_detect = 1; | |
1022 | RAS_CDEBUG(ras); | |
1023 | } | |
1024 | } else { | |
1025 | if (ra_miss) { | |
1026 | if (index_in_stride_window(ras, index) && | |
1027 | stride_io_mode(ras)) { | |
1028 | /*If stride-RA hit cache miss, the stride dector | |
1029 | *will not be reset to avoid the overhead of | |
c0894c6c OD |
1030 | *redetecting read-ahead mode |
1031 | */ | |
d7e09d03 PT |
1032 | if (index != ras->ras_last_readpage + 1) |
1033 | ras->ras_consecutive_pages = 0; | |
1034 | ras_reset(inode, ras, index); | |
1035 | RAS_CDEBUG(ras); | |
1036 | } else { | |
1037 | /* Reset both stride window and normal RA | |
c0894c6c OD |
1038 | * window |
1039 | */ | |
d7e09d03 PT |
1040 | ras_reset(inode, ras, index); |
1041 | ras->ras_consecutive_pages++; | |
1042 | ras_stride_reset(ras); | |
34e1f2bb | 1043 | goto out_unlock; |
d7e09d03 PT |
1044 | } |
1045 | } else if (stride_io_mode(ras)) { | |
1046 | /* If this is contiguous read but in stride I/O mode | |
1047 | * currently, check whether stride step still is valid, | |
c0894c6c OD |
1048 | * if invalid, it will reset the stride ra window |
1049 | */ | |
d7e09d03 PT |
1050 | if (!index_in_stride_window(ras, index)) { |
1051 | /* Shrink stride read-ahead window to be zero */ | |
1052 | ras_stride_reset(ras); | |
1053 | ras->ras_window_len = 0; | |
1054 | ras->ras_next_readahead = index; | |
1055 | } | |
1056 | } | |
1057 | } | |
1058 | ras->ras_consecutive_pages++; | |
1059 | ras->ras_last_readpage = index; | |
1060 | ras_set_start(inode, ras, index); | |
1061 | ||
1062 | if (stride_io_mode(ras)) | |
d0a0acc3 | 1063 | /* Since stride readahead is sensitive to the offset |
d7e09d03 | 1064 | * of read-ahead, so we use original offset here, |
c0894c6c OD |
1065 | * instead of ras_window_start, which is RPC aligned |
1066 | */ | |
d7e09d03 PT |
1067 | ras->ras_next_readahead = max(index, ras->ras_next_readahead); |
1068 | else | |
1069 | ras->ras_next_readahead = max(ras->ras_window_start, | |
1070 | ras->ras_next_readahead); | |
1071 | RAS_CDEBUG(ras); | |
1072 | ||
1073 | /* Trigger RA in the mmap case where ras_consecutive_requests | |
c0894c6c OD |
1074 | * is not incremented and thus can't be used to trigger RA |
1075 | */ | |
d7e09d03 PT |
1076 | if (!ras->ras_window_len && ras->ras_consecutive_pages == 4) { |
1077 | ras->ras_window_len = RAS_INCREASE_STEP(inode); | |
34e1f2bb | 1078 | goto out_unlock; |
d7e09d03 PT |
1079 | } |
1080 | ||
1081 | /* Initially reset the stride window offset to next_readahead*/ | |
1082 | if (ras->ras_consecutive_stride_requests == 2 && stride_detect) { | |
1083 | /** | |
1084 | * Once stride IO mode is detected, next_readahead should be | |
1085 | * reset to make sure next_readahead > stride offset | |
1086 | */ | |
1087 | ras->ras_next_readahead = max(index, ras->ras_next_readahead); | |
1088 | ras->ras_stride_offset = index; | |
1089 | ras->ras_window_len = RAS_INCREASE_STEP(inode); | |
1090 | } | |
1091 | ||
1092 | /* The initial ras_window_len is set to the request size. To avoid | |
1093 | * uselessly reading and discarding pages for random IO the window is | |
1094 | * only increased once per consecutive request received. */ | |
1095 | if ((ras->ras_consecutive_requests > 1 || stride_detect) && | |
1096 | !ras->ras_request_index) | |
1097 | ras_increase_window(inode, ras, ra); | |
d7e09d03 PT |
1098 | out_unlock: |
1099 | RAS_CDEBUG(ras); | |
1100 | ras->ras_request_index++; | |
1101 | spin_unlock(&ras->ras_lock); | |
1102 | return; | |
1103 | } | |
1104 | ||
1105 | int ll_writepage(struct page *vmpage, struct writeback_control *wbc) | |
1106 | { | |
1107 | struct inode *inode = vmpage->mapping->host; | |
1108 | struct ll_inode_info *lli = ll_i2info(inode); | |
1109 | struct lu_env *env; | |
1110 | struct cl_io *io; | |
1111 | struct cl_page *page; | |
1112 | struct cl_object *clob; | |
1113 | struct cl_env_nest nest; | |
1114 | bool redirtied = false; | |
1115 | bool unlocked = false; | |
1116 | int result; | |
d7e09d03 PT |
1117 | |
1118 | LASSERT(PageLocked(vmpage)); | |
1119 | LASSERT(!PageWriteback(vmpage)); | |
1120 | ||
6e16818b | 1121 | LASSERT(ll_i2dtexp(inode)); |
d7e09d03 PT |
1122 | |
1123 | env = cl_env_nested_get(&nest); | |
34e1f2bb JL |
1124 | if (IS_ERR(env)) { |
1125 | result = PTR_ERR(env); | |
1126 | goto out; | |
1127 | } | |
d7e09d03 PT |
1128 | |
1129 | clob = ll_i2info(inode)->lli_clob; | |
6e16818b | 1130 | LASSERT(clob); |
d7e09d03 PT |
1131 | |
1132 | io = ccc_env_thread_io(env); | |
1133 | io->ci_obj = clob; | |
1134 | io->ci_ignore_layout = 1; | |
1135 | result = cl_io_init(env, io, CIT_MISC, clob); | |
1136 | if (result == 0) { | |
1137 | page = cl_page_find(env, clob, vmpage->index, | |
1138 | vmpage, CPT_CACHEABLE); | |
1139 | if (!IS_ERR(page)) { | |
1140 | lu_ref_add(&page->cp_reference, "writepage", | |
1141 | current); | |
1142 | cl_page_assume(env, io, page); | |
1143 | result = cl_page_flush(env, io, page); | |
1144 | if (result != 0) { | |
1145 | /* | |
1146 | * Re-dirty page on error so it retries write, | |
1147 | * but not in case when IO has actually | |
1148 | * occurred and completed with an error. | |
1149 | */ | |
1150 | if (!PageError(vmpage)) { | |
1151 | redirty_page_for_writepage(wbc, vmpage); | |
1152 | result = 0; | |
1153 | redirtied = true; | |
1154 | } | |
1155 | } | |
1156 | cl_page_disown(env, io, page); | |
1157 | unlocked = true; | |
1158 | lu_ref_del(&page->cp_reference, | |
1159 | "writepage", current); | |
1160 | cl_page_put(env, page); | |
1161 | } else { | |
1162 | result = PTR_ERR(page); | |
1163 | } | |
1164 | } | |
1165 | cl_io_fini(env, io); | |
1166 | ||
1167 | if (redirtied && wbc->sync_mode == WB_SYNC_ALL) { | |
1168 | loff_t offset = cl_offset(clob, vmpage->index); | |
1169 | ||
1170 | /* Flush page failed because the extent is being written out. | |
1171 | * Wait for the write of extent to be finished to avoid | |
1172 | * breaking kernel which assumes ->writepage should mark | |
c0894c6c OD |
1173 | * PageWriteback or clean the page. |
1174 | */ | |
d7e09d03 | 1175 | result = cl_sync_file_range(inode, offset, |
09cbfeaf | 1176 | offset + PAGE_SIZE - 1, |
65fb55d1 | 1177 | CL_FSYNC_LOCAL, 1); |
d7e09d03 PT |
1178 | if (result > 0) { |
1179 | /* actually we may have written more than one page. | |
1180 | * decreasing this page because the caller will count | |
c0894c6c OD |
1181 | * it. |
1182 | */ | |
d7e09d03 PT |
1183 | wbc->nr_to_write -= result - 1; |
1184 | result = 0; | |
1185 | } | |
1186 | } | |
1187 | ||
1188 | cl_env_nested_put(&nest, env); | |
34e1f2bb | 1189 | goto out; |
d7e09d03 PT |
1190 | |
1191 | out: | |
1192 | if (result < 0) { | |
1193 | if (!lli->lli_async_rc) | |
1194 | lli->lli_async_rc = result; | |
1195 | SetPageError(vmpage); | |
1196 | if (!unlocked) | |
1197 | unlock_page(vmpage); | |
1198 | } | |
1199 | return result; | |
1200 | } | |
1201 | ||
1202 | int ll_writepages(struct address_space *mapping, struct writeback_control *wbc) | |
1203 | { | |
1204 | struct inode *inode = mapping->host; | |
65fb55d1 | 1205 | struct ll_sb_info *sbi = ll_i2sbi(inode); |
d7e09d03 PT |
1206 | loff_t start; |
1207 | loff_t end; | |
1208 | enum cl_fsync_mode mode; | |
1209 | int range_whole = 0; | |
1210 | int result; | |
65fb55d1 | 1211 | int ignore_layout = 0; |
d7e09d03 PT |
1212 | |
1213 | if (wbc->range_cyclic) { | |
09cbfeaf | 1214 | start = mapping->writeback_index << PAGE_SHIFT; |
d7e09d03 PT |
1215 | end = OBD_OBJECT_EOF; |
1216 | } else { | |
1217 | start = wbc->range_start; | |
1218 | end = wbc->range_end; | |
1219 | if (end == LLONG_MAX) { | |
1220 | end = OBD_OBJECT_EOF; | |
1221 | range_whole = start == 0; | |
1222 | } | |
1223 | } | |
1224 | ||
1225 | mode = CL_FSYNC_NONE; | |
1226 | if (wbc->sync_mode == WB_SYNC_ALL) | |
1227 | mode = CL_FSYNC_LOCAL; | |
1228 | ||
65fb55d1 NY |
1229 | if (sbi->ll_umounting) |
1230 | /* if the mountpoint is being umounted, all pages have to be | |
1231 | * evicted to avoid hitting LBUG when truncate_inode_pages() | |
c0894c6c OD |
1232 | * is called later on. |
1233 | */ | |
65fb55d1 NY |
1234 | ignore_layout = 1; |
1235 | result = cl_sync_file_range(inode, start, end, mode, ignore_layout); | |
d7e09d03 PT |
1236 | if (result > 0) { |
1237 | wbc->nr_to_write -= result; | |
1238 | result = 0; | |
1239 | } | |
1240 | ||
1241 | if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0)) { | |
1242 | if (end == OBD_OBJECT_EOF) | |
1243 | end = i_size_read(inode); | |
09cbfeaf | 1244 | mapping->writeback_index = (end >> PAGE_SHIFT) + 1; |
d7e09d03 | 1245 | } |
0a3bdb00 | 1246 | return result; |
d7e09d03 PT |
1247 | } |
1248 | ||
1249 | int ll_readpage(struct file *file, struct page *vmpage) | |
1250 | { | |
1251 | struct ll_cl_context *lcc; | |
1252 | int result; | |
d7e09d03 PT |
1253 | |
1254 | lcc = ll_cl_init(file, vmpage, 0); | |
1255 | if (!IS_ERR(lcc)) { | |
1256 | struct lu_env *env = lcc->lcc_env; | |
1257 | struct cl_io *io = lcc->lcc_io; | |
1258 | struct cl_page *page = lcc->lcc_page; | |
1259 | ||
1260 | LASSERT(page->cp_type == CPT_CACHEABLE); | |
1261 | if (likely(!PageUptodate(vmpage))) { | |
1262 | cl_page_assume(env, io, page); | |
1263 | result = cl_io_read_page(env, io, page); | |
1264 | } else { | |
1265 | /* Page from a non-object file. */ | |
1266 | unlock_page(vmpage); | |
1267 | result = 0; | |
1268 | } | |
1269 | ll_cl_fini(lcc); | |
1270 | } else { | |
1271 | unlock_page(vmpage); | |
1272 | result = PTR_ERR(lcc); | |
1273 | } | |
0a3bdb00 | 1274 | return result; |
d7e09d03 | 1275 | } |