Whamcloud - gitweb
d731d4aecc4f9f99db731adcea1623a73d4da771
[fs/lustre-release.git] / lustre / osd-ldiskfs / osd_io.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.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) 2010, Oracle and/or its affiliates. All rights reserved.
28  * Use is subject to license terms.
29  *
30  * Copyright (c) 2012, 2014, Intel Corporation.
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/osd/osd_io.c
37  *
38  * body operations
39  *
40  * Author: Nikita Danilov <nikita@clusterfs.com>
41  * Author: Alex Zhuravlev <bzzz@whamcloud.com>
42  *
43  */
44
45 /* LUSTRE_VERSION_CODE */
46 #include <lustre_ver.h>
47 /* prerequisite for linux/xattr.h */
48 #include <linux/types.h>
49 /* prerequisite for linux/xattr.h */
50 #include <linux/fs.h>
51
52 /*
53  * struct OBD_{ALLOC,FREE}*()
54  * OBD_FAIL_CHECK
55  */
56 #include <obd_support.h>
57
58 #include "osd_internal.h"
59
60 /* ext_depth() */
61 #include <ldiskfs/ldiskfs_extents.h>
62
63 static int __osd_init_iobuf(struct osd_device *d, struct osd_iobuf *iobuf,
64                             int rw, int line, int pages)
65 {
66         int blocks, i;
67
68         LASSERTF(iobuf->dr_elapsed_valid == 0,
69                  "iobuf %p, reqs %d, rw %d, line %d\n", iobuf,
70                  atomic_read(&iobuf->dr_numreqs), iobuf->dr_rw,
71                  iobuf->dr_init_at);
72         LASSERT(pages <= PTLRPC_MAX_BRW_PAGES);
73
74         init_waitqueue_head(&iobuf->dr_wait);
75         atomic_set(&iobuf->dr_numreqs, 0);
76         iobuf->dr_npages = 0;
77         iobuf->dr_error = 0;
78         iobuf->dr_dev = d;
79         iobuf->dr_frags = 0;
80         iobuf->dr_elapsed = 0;
81         /* must be counted before, so assert */
82         iobuf->dr_rw = rw;
83         iobuf->dr_init_at = line;
84
85         blocks = pages * (PAGE_CACHE_SIZE >> osd_sb(d)->s_blocksize_bits);
86         if (iobuf->dr_bl_buf.lb_len >= blocks * sizeof(iobuf->dr_blocks[0])) {
87                 LASSERT(iobuf->dr_pg_buf.lb_len >=
88                         pages * sizeof(iobuf->dr_pages[0]));
89                 return 0;
90         }
91
92         /* start with 1MB for 4K blocks */
93         i = 256;
94         while (i <= PTLRPC_MAX_BRW_PAGES && i < pages)
95                 i <<= 1;
96
97         CDEBUG(D_OTHER, "realloc %u for %u (%u) pages\n",
98                (unsigned)(pages * sizeof(iobuf->dr_pages[0])), i, pages);
99         pages = i;
100         blocks = pages * (PAGE_CACHE_SIZE >> osd_sb(d)->s_blocksize_bits);
101         iobuf->dr_max_pages = 0;
102         CDEBUG(D_OTHER, "realloc %u for %u blocks\n",
103                (unsigned)(blocks * sizeof(iobuf->dr_blocks[0])), blocks);
104
105         lu_buf_realloc(&iobuf->dr_bl_buf, blocks * sizeof(iobuf->dr_blocks[0]));
106         iobuf->dr_blocks = iobuf->dr_bl_buf.lb_buf;
107         if (unlikely(iobuf->dr_blocks == NULL))
108                 return -ENOMEM;
109
110         lu_buf_realloc(&iobuf->dr_pg_buf, pages * sizeof(iobuf->dr_pages[0]));
111         iobuf->dr_pages = iobuf->dr_pg_buf.lb_buf;
112         if (unlikely(iobuf->dr_pages == NULL))
113                 return -ENOMEM;
114
115         iobuf->dr_max_pages = pages;
116
117         return 0;
118 }
119 #define osd_init_iobuf(dev, iobuf, rw, pages) \
120         __osd_init_iobuf(dev, iobuf, rw, __LINE__, pages)
121
122 static void osd_iobuf_add_page(struct osd_iobuf *iobuf, struct page *page)
123 {
124         LASSERT(iobuf->dr_npages < iobuf->dr_max_pages);
125         iobuf->dr_pages[iobuf->dr_npages++] = page;
126 }
127
128 void osd_fini_iobuf(struct osd_device *d, struct osd_iobuf *iobuf)
129 {
130         int rw = iobuf->dr_rw;
131
132         if (iobuf->dr_elapsed_valid) {
133                 iobuf->dr_elapsed_valid = 0;
134                 LASSERT(iobuf->dr_dev == d);
135                 LASSERT(iobuf->dr_frags > 0);
136                 lprocfs_oh_tally(&d->od_brw_stats.
137                                  hist[BRW_R_DIO_FRAGS+rw],
138                                  iobuf->dr_frags);
139                 lprocfs_oh_tally_log2(&d->od_brw_stats.hist[BRW_R_IO_TIME+rw],
140                                       iobuf->dr_elapsed);
141         }
142 }
143
144 #ifndef REQ_WRITE /* pre-2.6.35 */
145 #define __REQ_WRITE BIO_RW
146 #endif
147
148 static void dio_complete_routine(struct bio *bio, int error)
149 {
150         struct osd_iobuf *iobuf = bio->bi_private;
151 #ifdef HAVE_BVEC_ITER
152         struct bvec_iter iter;
153         struct bio_vec bvl;
154 #else
155         int iter;
156         struct bio_vec *bvl;
157 #endif
158
159         /* CAVEAT EMPTOR: possibly in IRQ context
160          * DO NOT record procfs stats here!!! */
161
162         if (unlikely(iobuf == NULL)) {
163                 CERROR("***** bio->bi_private is NULL!  This should never "
164                        "happen.  Normally, I would crash here, but instead I "
165                        "will dump the bio contents to the console.  Please "
166                        "report this to <https://jira.hpdd.intel.com/> , along "
167                        "with any interesting messages leading up to this point "
168                        "(like SCSI errors, perhaps).  Because bi_private is "
169                        "NULL, I can't wake up the thread that initiated this "
170                        "IO - you will probably have to reboot this node.\n");
171                 CERROR("bi_next: %p, bi_flags: %lx, bi_rw: %lu, bi_vcnt: %d, "
172                        "bi_idx: %d, bi->size: %d, bi_end_io: %p, bi_cnt: %d, "
173                        "bi_private: %p\n", bio->bi_next, bio->bi_flags,
174                         bio->bi_rw, bio->bi_vcnt, bio_idx(bio),
175                         bio_sectors(bio) << 9, bio->bi_end_io,
176                         atomic_read(&bio->bi_cnt), bio->bi_private);
177                 return;
178         }
179
180         /* the check is outside of the cycle for performance reason -bzzz */
181         if (!test_bit(__REQ_WRITE, &bio->bi_rw)) {
182                 bio_for_each_segment(bvl, bio, iter) {
183                         if (likely(error == 0))
184                                 SetPageUptodate(bvec_iter_page(&bvl, iter));
185                         LASSERT(PageLocked(bvec_iter_page(&bvl, iter)));
186                 }
187                 atomic_dec(&iobuf->dr_dev->od_r_in_flight);
188         } else {
189                 atomic_dec(&iobuf->dr_dev->od_w_in_flight);
190         }
191
192         /* any real error is good enough -bzzz */
193         if (error != 0 && iobuf->dr_error == 0)
194                 iobuf->dr_error = error;
195
196         /*
197          * set dr_elapsed before dr_numreqs turns to 0, otherwise
198          * it's possible that service thread will see dr_numreqs
199          * is zero, but dr_elapsed is not set yet, leading to lost
200          * data in this processing and an assertion in a subsequent
201          * call to OSD.
202          */
203         if (atomic_read(&iobuf->dr_numreqs) == 1) {
204                 iobuf->dr_elapsed = jiffies - iobuf->dr_start_time;
205                 iobuf->dr_elapsed_valid = 1;
206         }
207         if (atomic_dec_and_test(&iobuf->dr_numreqs))
208                 wake_up(&iobuf->dr_wait);
209
210         /* Completed bios used to be chained off iobuf->dr_bios and freed in
211          * filter_clear_dreq().  It was then possible to exhaust the biovec-256
212          * mempool when serious on-disk fragmentation was encountered,
213          * deadlocking the OST.  The bios are now released as soon as complete
214          * so the pool cannot be exhausted while IOs are competing. bug 10076 */
215         bio_put(bio);
216 }
217
218 static void record_start_io(struct osd_iobuf *iobuf, int size)
219 {
220         struct osd_device    *osd = iobuf->dr_dev;
221         struct obd_histogram *h = osd->od_brw_stats.hist;
222
223         iobuf->dr_frags++;
224         atomic_inc(&iobuf->dr_numreqs);
225
226         if (iobuf->dr_rw == 0) {
227                 atomic_inc(&osd->od_r_in_flight);
228                 lprocfs_oh_tally(&h[BRW_R_RPC_HIST],
229                                  atomic_read(&osd->od_r_in_flight));
230                 lprocfs_oh_tally_log2(&h[BRW_R_DISK_IOSIZE], size);
231         } else if (iobuf->dr_rw == 1) {
232                 atomic_inc(&osd->od_w_in_flight);
233                 lprocfs_oh_tally(&h[BRW_W_RPC_HIST],
234                                  atomic_read(&osd->od_w_in_flight));
235                 lprocfs_oh_tally_log2(&h[BRW_W_DISK_IOSIZE], size);
236         } else {
237                 LBUG();
238         }
239 }
240
241 static void osd_submit_bio(int rw, struct bio *bio)
242 {
243         LASSERTF(rw == 0 || rw == 1, "%x\n", rw);
244         if (rw == 0)
245                 submit_bio(READ, bio);
246         else
247                 submit_bio(WRITE, bio);
248 }
249
250 static int can_be_merged(struct bio *bio, sector_t sector)
251 {
252         if (bio == NULL)
253                 return 0;
254
255         return bio_end_sector(bio) == sector ? 1 : 0;
256 }
257
258 static int osd_do_bio(struct osd_device *osd, struct inode *inode,
259                       struct osd_iobuf *iobuf)
260 {
261         int            blocks_per_page = PAGE_CACHE_SIZE >> inode->i_blkbits;
262         struct page  **pages = iobuf->dr_pages;
263         int            npages = iobuf->dr_npages;
264         unsigned long *blocks = iobuf->dr_blocks;
265         int            total_blocks = npages * blocks_per_page;
266         int            sector_bits = inode->i_sb->s_blocksize_bits - 9;
267         unsigned int   blocksize = inode->i_sb->s_blocksize;
268         struct bio    *bio = NULL;
269         struct page   *page;
270         unsigned int   page_offset;
271         sector_t       sector;
272         int            nblocks;
273         int            block_idx;
274         int            page_idx;
275         int            i;
276         int            rc = 0;
277         ENTRY;
278
279         LASSERT(iobuf->dr_npages == npages);
280
281         osd_brw_stats_update(osd, iobuf);
282         iobuf->dr_start_time = cfs_time_current();
283
284         for (page_idx = 0, block_idx = 0;
285              page_idx < npages;
286              page_idx++, block_idx += blocks_per_page) {
287
288                 page = pages[page_idx];
289                 LASSERT(block_idx + blocks_per_page <= total_blocks);
290
291                 for (i = 0, page_offset = 0;
292                      i < blocks_per_page;
293                      i += nblocks, page_offset += blocksize * nblocks) {
294
295                         nblocks = 1;
296
297                         if (blocks[block_idx + i] == 0) {  /* hole */
298                                 LASSERTF(iobuf->dr_rw == 0,
299                                          "page_idx %u, block_idx %u, i %u\n",
300                                          page_idx, block_idx, i);
301                                 memset(kmap(page) + page_offset, 0, blocksize);
302                                 kunmap(page);
303                                 continue;
304                         }
305
306                         sector = (sector_t)blocks[block_idx + i] << sector_bits;
307
308                         /* Additional contiguous file blocks? */
309                         while (i + nblocks < blocks_per_page &&
310                                (sector + (nblocks << sector_bits)) ==
311                                ((sector_t)blocks[block_idx + i + nblocks] <<
312                                 sector_bits))
313                                 nblocks++;
314
315                         if (bio != NULL &&
316                             can_be_merged(bio, sector) &&
317                             bio_add_page(bio, page,
318                                          blocksize * nblocks, page_offset) != 0)
319                                 continue;       /* added this frag OK */
320
321                         if (bio != NULL) {
322                                 struct request_queue *q =
323                                         bdev_get_queue(bio->bi_bdev);
324                                 unsigned int bi_size = bio_sectors(bio) << 9;
325
326                                 /* Dang! I have to fragment this I/O */
327                                 CDEBUG(D_INODE, "bio++ sz %d vcnt %d(%d) "
328                                        "sectors %d(%d) psg %d(%d) hsg %d(%d)\n",
329                                        bi_size, bio->bi_vcnt, bio->bi_max_vecs,
330                                        bio_sectors(bio),
331                                        queue_max_sectors(q),
332                                        bio_phys_segments(q, bio),
333                                        queue_max_phys_segments(q),
334                                        0, queue_max_hw_segments(q));
335                                 record_start_io(iobuf, bi_size);
336                                 osd_submit_bio(iobuf->dr_rw, bio);
337                         }
338
339                         /* allocate new bio */
340                         bio = bio_alloc(GFP_NOIO, min(BIO_MAX_PAGES,
341                                                       (npages - page_idx) *
342                                                       blocks_per_page));
343                         if (bio == NULL) {
344                                 CERROR("Can't allocate bio %u*%u = %u pages\n",
345                                        (npages - page_idx), blocks_per_page,
346                                        (npages - page_idx) * blocks_per_page);
347                                 rc = -ENOMEM;
348                                 goto out;
349                         }
350
351                         bio->bi_bdev = inode->i_sb->s_bdev;
352                         bio_set_sector(bio, sector);
353                         bio->bi_rw = (iobuf->dr_rw == 0) ? READ : WRITE;
354                         bio->bi_end_io = dio_complete_routine;
355                         bio->bi_private = iobuf;
356
357                         rc = bio_add_page(bio, page,
358                                           blocksize * nblocks, page_offset);
359                         LASSERT(rc != 0);
360                 }
361         }
362
363         if (bio != NULL) {
364                 record_start_io(iobuf, bio_sectors(bio) << 9);
365                 osd_submit_bio(iobuf->dr_rw, bio);
366                 rc = 0;
367         }
368
369 out:
370         /* in order to achieve better IO throughput, we don't wait for writes
371          * completion here. instead we proceed with transaction commit in
372          * parallel and wait for IO completion once transaction is stopped
373          * see osd_trans_stop() for more details -bzzz */
374         if (iobuf->dr_rw == 0) {
375                 wait_event(iobuf->dr_wait,
376                            atomic_read(&iobuf->dr_numreqs) == 0);
377                 osd_fini_iobuf(osd, iobuf);
378         }
379
380         if (rc == 0)
381                 rc = iobuf->dr_error;
382         RETURN(rc);
383 }
384
385 static int osd_map_remote_to_local(loff_t offset, ssize_t len, int *nrpages,
386                                    struct niobuf_local *lnb)
387 {
388         ENTRY;
389
390         *nrpages = 0;
391
392         while (len > 0) {
393                 int poff = offset & (PAGE_CACHE_SIZE - 1);
394                 int plen = PAGE_CACHE_SIZE - poff;
395
396                 if (plen > len)
397                         plen = len;
398                 lnb->lnb_file_offset = offset;
399                 lnb->lnb_page_offset = poff;
400                 lnb->lnb_len = plen;
401                 /* lnb->lnb_flags = rnb->rnb_flags; */
402                 lnb->lnb_flags = 0;
403                 lnb->lnb_page = NULL;
404                 lnb->lnb_rc = 0;
405
406                 LASSERTF(plen <= len, "plen %u, len %lld\n", plen,
407                          (long long) len);
408                 offset += plen;
409                 len -= plen;
410                 lnb++;
411                 (*nrpages)++;
412         }
413
414         RETURN(0);
415 }
416
417 static struct page *osd_get_page(struct dt_object *dt, loff_t offset, int rw)
418 {
419         struct inode      *inode = osd_dt_obj(dt)->oo_inode;
420         struct osd_device *d = osd_obj2dev(osd_dt_obj(dt));
421         struct page       *page;
422
423         LASSERT(inode);
424
425         page = find_or_create_page(inode->i_mapping, offset >> PAGE_CACHE_SHIFT,
426                                    GFP_NOFS | __GFP_HIGHMEM);
427         if (unlikely(page == NULL))
428                 lprocfs_counter_add(d->od_stats, LPROC_OSD_NO_PAGE, 1);
429
430         return page;
431 }
432
433 /*
434  * there are following "locks":
435  * journal_start
436  * i_mutex
437  * page lock
438  *
439  * osd write path:
440  *  - lock page(s)
441  *  - journal_start
442  *  - truncate_sem
443  *
444  * ext4 vmtruncate:
445  *  - lock pages, unlock
446  *  - journal_start
447  *  - lock partial page
448  *  - i_data_sem
449  *
450  */
451
452 /**
453  * Unlock and release pages loaded by osd_bufs_get()
454  *
455  * Unlock \a npages pages from \a lnb and drop the refcount on them.
456  *
457  * \param env           thread execution environment
458  * \param dt            dt object undergoing IO (OSD object + methods)
459  * \param lnb           array of pages undergoing IO
460  * \param npages        number of pages in \a lnb
461  *
462  * \retval 0            always
463  */
464 static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt,
465                         struct niobuf_local *lnb, int npages)
466 {
467         int i;
468
469         for (i = 0; i < npages; i++) {
470                 if (lnb[i].lnb_page == NULL)
471                         continue;
472                 LASSERT(PageLocked(lnb[i].lnb_page));
473                 unlock_page(lnb[i].lnb_page);
474                 page_cache_release(lnb[i].lnb_page);
475                 lu_object_put(env, &dt->do_lu);
476                 lnb[i].lnb_page = NULL;
477         }
478
479         RETURN(0);
480 }
481
482 /**
483  * Load and lock pages undergoing IO
484  *
485  * Pages as described in the \a lnb array are fetched (from disk or cache)
486  * and locked for IO by the caller.
487  *
488  * DLM locking protects us from write and truncate competing for same region,
489  * but partial-page truncate can leave dirty pages in the cache for ldiskfs.
490  * It's possible the writeout on a such a page is in progress when we access
491  * it. It's also possible that during this writeout we put new (partial) data
492  * into the page, but won't be able to proceed in filter_commitrw_write().
493  * Therefore, just wait for writeout completion as it should be rare enough.
494  *
495  * \param env           thread execution environment
496  * \param dt            dt object undergoing IO (OSD object + methods)
497  * \param pos           byte offset of IO start
498  * \param len           number of bytes of IO
499  * \param lnb           array of extents undergoing IO
500  * \param rw            read or write operation?
501  * \param capa          capabilities
502  *
503  * \retval pages        (zero or more) loaded successfully
504  * \retval -ENOMEM      on memory/page allocation error
505  */
506 static int osd_bufs_get(const struct lu_env *env, struct dt_object *dt,
507                         loff_t pos, ssize_t len, struct niobuf_local *lnb,
508                         int rw, struct lustre_capa *capa)
509 {
510         struct osd_object   *obj    = osd_dt_obj(dt);
511         int npages, i, rc = 0;
512
513         LASSERT(obj->oo_inode);
514
515         osd_map_remote_to_local(pos, len, &npages, lnb);
516
517         for (i = 0; i < npages; i++, lnb++) {
518                 lnb->lnb_page = osd_get_page(dt, lnb->lnb_file_offset, rw);
519                 if (lnb->lnb_page == NULL)
520                         GOTO(cleanup, rc = -ENOMEM);
521
522                 wait_on_page_writeback(lnb->lnb_page);
523                 BUG_ON(PageWriteback(lnb->lnb_page));
524
525                 lu_object_get(&dt->do_lu);
526         }
527
528         RETURN(i);
529
530 cleanup:
531         if (i > 0)
532                 osd_bufs_put(env, dt, lnb - i, i);
533         return rc;
534 }
535
536 #ifndef HAVE_LDISKFS_MAP_BLOCKS
537
538 #ifdef HAVE_EXT_PBLOCK /* Name changed to ext4_ext_pblock for kernel 2.6.35 */
539 #define ldiskfs_ext_pblock(ex) ext_pblock((ex))
540 #endif
541
542 struct bpointers {
543         unsigned long *blocks;
544         unsigned long start;
545         int num;
546         int init_num;
547         int create;
548 };
549
550 static long ldiskfs_ext_find_goal(struct inode *inode,
551                                   struct ldiskfs_ext_path *path,
552                                   unsigned long block, int *aflags)
553 {
554         struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
555         unsigned long bg_start;
556         unsigned long colour;
557         int depth;
558
559         if (path) {
560                 struct ldiskfs_extent *ex;
561                 depth = path->p_depth;
562
563                 /* try to predict block placement */
564                 if ((ex = path[depth].p_ext))
565                         return ldiskfs_ext_pblock(ex) +
566                                 (block - le32_to_cpu(ex->ee_block));
567
568                 /* it looks index is empty
569                  * try to find starting from index itself */
570                 if (path[depth].p_bh)
571                         return path[depth].p_bh->b_blocknr;
572         }
573
574         /* OK. use inode's group */
575         bg_start = (ei->i_block_group * LDISKFS_BLOCKS_PER_GROUP(inode->i_sb)) +
576                 le32_to_cpu(LDISKFS_SB(inode->i_sb)->s_es->s_first_data_block);
577         colour = (current->pid % 16) *
578                 (LDISKFS_BLOCKS_PER_GROUP(inode->i_sb) / 16);
579         return bg_start + colour + block;
580 }
581
582 static unsigned long new_blocks(handle_t *handle, struct inode *inode,
583                                 struct ldiskfs_ext_path *path,
584                                 unsigned long block, unsigned long *count,
585                                 int *err)
586 {
587         struct ldiskfs_allocation_request ar;
588         unsigned long pblock;
589         int aflags;
590
591         /* find neighbour allocated blocks */
592         ar.lleft = block;
593         *err = ldiskfs_ext_search_left(inode, path, &ar.lleft, &ar.pleft);
594         if (*err)
595                 return 0;
596         ar.lright = block;
597         *err = ldiskfs_ext_search_right(inode, path, &ar.lright, &ar.pright);
598         if (*err)
599                 return 0;
600
601         /* allocate new block */
602         ar.goal = ldiskfs_ext_find_goal(inode, path, block, &aflags);
603         ar.inode = inode;
604         ar.logical = block;
605         ar.len = *count;
606         ar.flags = LDISKFS_MB_HINT_DATA;
607         pblock = ldiskfs_mb_new_blocks(handle, &ar, err);
608         *count = ar.len;
609         return pblock;
610 }
611
612 static int ldiskfs_ext_new_extent_cb(struct inode *inode,
613                                      struct ldiskfs_ext_path *path,
614                                      struct ldiskfs_ext_cache *cex,
615 #ifdef HAVE_EXT_PREPARE_CB_EXTENT
616                                      struct ldiskfs_extent *ex,
617 #endif
618                                      void *cbdata)
619 {
620         struct bpointers *bp = cbdata;
621         struct ldiskfs_extent nex;
622         unsigned long pblock;
623         unsigned long tgen;
624         int err, i;
625         unsigned long count;
626         handle_t *handle;
627
628 #ifdef LDISKFS_EXT_CACHE_EXTENT /* until kernel 2.6.37 */
629         if (cex->ec_type == LDISKFS_EXT_CACHE_EXTENT) {
630 #else
631         if ((cex->ec_len != 0) && (cex->ec_start != 0)) {
632 #endif
633                 err = EXT_CONTINUE;
634                 goto map;
635         }
636
637         if (bp->create == 0) {
638                 i = 0;
639                 if (cex->ec_block < bp->start)
640                         i = bp->start - cex->ec_block;
641                 if (i >= cex->ec_len)
642                         CERROR("nothing to do?! i = %d, e_num = %u\n",
643                                         i, cex->ec_len);
644                 for (; i < cex->ec_len && bp->num; i++) {
645                         *(bp->blocks) = 0;
646                         bp->blocks++;
647                         bp->num--;
648                         bp->start++;
649                 }
650
651                 return EXT_CONTINUE;
652         }
653
654         tgen = LDISKFS_I(inode)->i_ext_generation;
655         count = ldiskfs_ext_calc_credits_for_insert(inode, path);
656
657         handle = osd_journal_start(inode, LDISKFS_HT_MISC,
658                                    count + LDISKFS_ALLOC_NEEDED + 1);
659         if (IS_ERR(handle)) {
660                 return PTR_ERR(handle);
661         }
662
663         if (tgen != LDISKFS_I(inode)->i_ext_generation) {
664                 /* the tree has changed. so path can be invalid at moment */
665                 ldiskfs_journal_stop(handle);
666                 return EXT_REPEAT;
667         }
668
669         /* In 2.6.32 kernel, ldiskfs_ext_walk_space()'s callback func is not
670          * protected by i_data_sem as whole. so we patch it to store
671          * generation to path and now verify the tree hasn't changed */
672         down_write((&LDISKFS_I(inode)->i_data_sem));
673
674         /* validate extent, make sure the extent tree does not changed */
675         if (LDISKFS_I(inode)->i_ext_generation != path[0].p_generation) {
676                 /* cex is invalid, try again */
677                 up_write(&LDISKFS_I(inode)->i_data_sem);
678                 ldiskfs_journal_stop(handle);
679                 return EXT_REPEAT;
680         }
681
682         count = cex->ec_len;
683         pblock = new_blocks(handle, inode, path, cex->ec_block, &count, &err);
684         if (!pblock)
685                 goto out;
686         BUG_ON(count > cex->ec_len);
687
688         /* insert new extent */
689         nex.ee_block = cpu_to_le32(cex->ec_block);
690         ldiskfs_ext_store_pblock(&nex, pblock);
691         nex.ee_len = cpu_to_le16(count);
692         err = ldiskfs_ext_insert_extent(handle, inode, path, &nex, 0);
693         if (err) {
694                 /* free data blocks we just allocated */
695                 /* not a good idea to call discard here directly,
696                  * but otherwise we'd need to call it every free() */
697                 ldiskfs_discard_preallocations(inode);
698 #ifdef HAVE_EXT_FREE_BLOCK_WITH_BUFFER_HEAD /* Introduced in 2.6.32-rc7 */
699                 ldiskfs_free_blocks(handle, inode, NULL,
700                                     ldiskfs_ext_pblock(&nex),
701                                     le16_to_cpu(nex.ee_len), 0);
702 #else
703                 ldiskfs_free_blocks(handle, inode, ldiskfs_ext_pblock(&nex),
704                                     le16_to_cpu(nex.ee_len), 0);
705 #endif
706                 goto out;
707         }
708
709         /*
710          * Putting len of the actual extent we just inserted,
711          * we are asking ldiskfs_ext_walk_space() to continue
712          * scaning after that block
713          */
714         cex->ec_len = le16_to_cpu(nex.ee_len);
715         cex->ec_start = ldiskfs_ext_pblock(&nex);
716         BUG_ON(le16_to_cpu(nex.ee_len) == 0);
717         BUG_ON(le32_to_cpu(nex.ee_block) != cex->ec_block);
718
719 out:
720         up_write((&LDISKFS_I(inode)->i_data_sem));
721         ldiskfs_journal_stop(handle);
722 map:
723         if (err >= 0) {
724                 /* map blocks */
725                 if (bp->num == 0) {
726                         CERROR("hmm. why do we find this extent?\n");
727                         CERROR("initial space: %lu:%u\n",
728                                 bp->start, bp->init_num);
729 #ifdef LDISKFS_EXT_CACHE_EXTENT /* until kernel 2.6.37 */
730                         CERROR("current extent: %u/%u/%llu %d\n",
731                                 cex->ec_block, cex->ec_len,
732                                 (unsigned long long)cex->ec_start,
733                                 cex->ec_type);
734 #else
735                         CERROR("current extent: %u/%u/%llu\n",
736                                 cex->ec_block, cex->ec_len,
737                                 (unsigned long long)cex->ec_start);
738 #endif
739                 }
740                 i = 0;
741                 if (cex->ec_block < bp->start)
742                         i = bp->start - cex->ec_block;
743                 if (i >= cex->ec_len)
744                         CERROR("nothing to do?! i = %d, e_num = %u\n",
745                                         i, cex->ec_len);
746                 for (; i < cex->ec_len && bp->num; i++) {
747                         *(bp->blocks) = cex->ec_start + i;
748 #ifdef LDISKFS_EXT_CACHE_EXTENT /* until kernel 2.6.37 */
749                         if (cex->ec_type != LDISKFS_EXT_CACHE_EXTENT) {
750 #else
751                         if ((cex->ec_len == 0) || (cex->ec_start == 0)) {
752 #endif
753                                 /* unmap any possible underlying metadata from
754                                  * the block device mapping.  bug 6998. */
755                                 unmap_underlying_metadata(inode->i_sb->s_bdev,
756                                                           *(bp->blocks));
757                         }
758                         bp->blocks++;
759                         bp->num--;
760                         bp->start++;
761                 }
762         }
763         return err;
764 }
765
766 static int osd_ldiskfs_map_nblocks(struct inode *inode, unsigned long block,
767                                    unsigned long num, unsigned long *blocks,
768                                    int create)
769 {
770         struct bpointers bp;
771         int err;
772
773         CDEBUG(D_OTHER, "blocks %lu-%lu requested for inode %u\n",
774                block, block + num - 1, (unsigned) inode->i_ino);
775
776         bp.blocks = blocks;
777         bp.start = block;
778         bp.init_num = bp.num = num;
779         bp.create = create;
780
781         err = ldiskfs_ext_walk_space(inode, block, num,
782                                          ldiskfs_ext_new_extent_cb, &bp);
783         ldiskfs_ext_invalidate_cache(inode);
784
785         return err;
786 }
787
788 static int osd_ldiskfs_map_bm_inode_pages(struct inode *inode,
789                                           struct page **page, int pages,
790                                           unsigned long *blocks, int create)
791 {
792         int blocks_per_page = PAGE_CACHE_SIZE >> inode->i_blkbits;
793         unsigned long *b;
794         int rc = 0, i;
795
796         for (i = 0, b = blocks; i < pages; i++, page++) {
797                 rc = ldiskfs_map_inode_page(inode, *page, b, create);
798                 if (rc) {
799                         CERROR("ino %lu, blk %lu create %d: rc %d\n",
800                                inode->i_ino, *b, create, rc);
801                         break;
802                 }
803                 b += blocks_per_page;
804         }
805         return rc;
806 }
807
808 static int osd_ldiskfs_map_ext_inode_pages(struct inode *inode,
809                                            struct page **page,
810                                            int pages, unsigned long *blocks,
811                                            int create)
812 {
813         int blocks_per_page = PAGE_CACHE_SIZE >> inode->i_blkbits;
814         int rc = 0, i = 0;
815         struct page *fp = NULL;
816         int clen = 0;
817
818         CDEBUG(D_OTHER, "inode %lu: map %d pages from %lu\n",
819                 inode->i_ino, pages, (*page)->index);
820
821         /* pages are sorted already. so, we just have to find
822          * contig. space and process them properly */
823         while (i < pages) {
824                 if (fp == NULL) {
825                         /* start new extent */
826                         fp = *page++;
827                         clen = 1;
828                         i++;
829                         continue;
830                 } else if (fp->index + clen == (*page)->index) {
831                         /* continue the extent */
832                         page++;
833                         clen++;
834                         i++;
835                         continue;
836                 }
837
838                 /* process found extent */
839                 rc = osd_ldiskfs_map_nblocks(inode, fp->index * blocks_per_page,
840                                              clen * blocks_per_page, blocks,
841                                              create);
842                 if (rc)
843                         GOTO(cleanup, rc);
844
845                 /* look for next extent */
846                 fp = NULL;
847                 blocks += blocks_per_page * clen;
848         }
849
850         if (fp)
851                 rc = osd_ldiskfs_map_nblocks(inode, fp->index * blocks_per_page,
852                                              clen * blocks_per_page, blocks,
853                                              create);
854 cleanup:
855         return rc;
856 }
857
858 static int osd_ldiskfs_map_inode_pages(struct inode *inode, struct page **page,
859                                        int pages, unsigned long *blocks,
860                                        int create)
861 {
862         int rc;
863
864         if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL) {
865                 rc = osd_ldiskfs_map_ext_inode_pages(inode, page, pages,
866                                                      blocks, create);
867                 return rc;
868         }
869         rc = osd_ldiskfs_map_bm_inode_pages(inode, page, pages, blocks, create);
870
871         return rc;
872 }
873 #else
874 static int osd_ldiskfs_map_inode_pages(struct inode *inode, struct page **page,
875                                        int pages, unsigned long *blocks,
876                                        int create)
877 {
878         int blocks_per_page = PAGE_CACHE_SIZE >> inode->i_blkbits;
879         int rc = 0, i = 0;
880         struct page *fp = NULL;
881         int clen = 0;
882
883         CDEBUG(D_OTHER, "inode %lu: map %d pages from %lu\n",
884                 inode->i_ino, pages, (*page)->index);
885
886         /* pages are sorted already. so, we just have to find
887          * contig. space and process them properly */
888         while (i < pages) {
889                 long blen, total = 0;
890                 handle_t *handle = NULL;
891                 struct ldiskfs_map_blocks map = { 0 };
892
893                 if (fp == NULL) { /* start new extent */
894                         fp = *page++;
895                         clen = 1;
896                         if (++i != pages)
897                                 continue;
898                 } else if (fp->index + clen == (*page)->index) {
899                         /* continue the extent */
900                         page++;
901                         clen++;
902                         if (++i != pages)
903                                 continue;
904                 }
905                 /* process found extent */
906                 map.m_lblk = fp->index * blocks_per_page;
907                 map.m_len = blen = clen * blocks_per_page;
908                 if (create) {
909                         create = LDISKFS_GET_BLOCKS_CREATE;
910                         handle = ldiskfs_journal_current_handle();
911                         LASSERT(handle != NULL);
912                 }
913 cont_map:
914                 rc = ldiskfs_map_blocks(handle, inode, &map, create);
915                 if (rc >= 0) {
916                         int c = 0;
917                         for (; total < blen && c < map.m_len; c++, total++) {
918                                 if (rc == 0) {
919                                         *(blocks + total) = 0;
920                                         total++;
921                                         break;
922                                 } else {
923                                         *(blocks + total) = map.m_pblk + c;
924                                         /* unmap any possible underlying
925                                          * metadata from the block device
926                                          * mapping.  bug 6998. */
927                                         if ((map.m_flags & LDISKFS_MAP_NEW) &&
928                                             create)
929                                                 unmap_underlying_metadata(
930                                                         inode->i_sb->s_bdev,
931                                                         map.m_pblk + c);
932                                 }
933                         }
934                         rc = 0;
935                 }
936                 if (rc == 0 && total < blen) {
937                         map.m_lblk = fp->index * blocks_per_page + total;
938                         map.m_len = blen - total;
939                         goto cont_map;
940                 }
941                 if (rc != 0)
942                         GOTO(cleanup, rc);
943
944                 /* look for next extent */
945                 fp = NULL;
946                 blocks += blocks_per_page * clen;
947         }
948 cleanup:
949         return rc;
950 }
951 #endif /* HAVE_LDISKFS_MAP_BLOCKS */
952
953 static int osd_write_prep(const struct lu_env *env, struct dt_object *dt,
954                           struct niobuf_local *lnb, int npages)
955 {
956         struct osd_thread_info *oti   = osd_oti_get(env);
957         struct osd_iobuf       *iobuf = &oti->oti_iobuf;
958         struct inode           *inode = osd_dt_obj(dt)->oo_inode;
959         struct osd_device      *osd   = osd_obj2dev(osd_dt_obj(dt));
960         struct timeval          start;
961         struct timeval          end;
962         unsigned long           timediff;
963         ssize_t                 isize;
964         __s64                   maxidx;
965         int                     rc = 0;
966         int                     i;
967         int                     cache = 0;
968
969         LASSERT(inode);
970
971         rc = osd_init_iobuf(osd, iobuf, 0, npages);
972         if (unlikely(rc != 0))
973                 RETURN(rc);
974
975         isize = i_size_read(inode);
976         maxidx = ((isize + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT) - 1;
977
978         if (osd->od_writethrough_cache)
979                 cache = 1;
980         if (isize > osd->od_readcache_max_filesize)
981                 cache = 0;
982
983         do_gettimeofday(&start);
984         for (i = 0; i < npages; i++) {
985
986                 if (cache == 0)
987                         generic_error_remove_page(inode->i_mapping,
988                                                   lnb[i].lnb_page);
989
990                 /*
991                  * till commit the content of the page is undefined
992                  * we'll set it uptodate once bulk is done. otherwise
993                  * subsequent reads can access non-stable data
994                  */
995                 ClearPageUptodate(lnb[i].lnb_page);
996
997                 if (lnb[i].lnb_len == PAGE_CACHE_SIZE)
998                         continue;
999
1000                 if (maxidx >= lnb[i].lnb_page->index) {
1001                         osd_iobuf_add_page(iobuf, lnb[i].lnb_page);
1002                 } else {
1003                         long off;
1004                         char *p = kmap(lnb[i].lnb_page);
1005
1006                         off = lnb[i].lnb_page_offset;
1007                         if (off)
1008                                 memset(p, 0, off);
1009                         off = (lnb[i].lnb_page_offset + lnb[i].lnb_len) &
1010                               ~CFS_PAGE_MASK;
1011                         if (off)
1012                                 memset(p + off, 0, PAGE_CACHE_SIZE - off);
1013                         kunmap(lnb[i].lnb_page);
1014                 }
1015         }
1016         do_gettimeofday(&end);
1017         timediff = cfs_timeval_sub(&end, &start, NULL);
1018         lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
1019
1020         if (iobuf->dr_npages) {
1021                 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1022                                                  iobuf->dr_npages,
1023                                                  iobuf->dr_blocks, 0);
1024                 if (likely(rc == 0)) {
1025                         rc = osd_do_bio(osd, inode, iobuf);
1026                         /* do IO stats for preparation reads */
1027                         osd_fini_iobuf(osd, iobuf);
1028                 }
1029         }
1030         RETURN(rc);
1031 }
1032
1033 /* Check if a block is allocated or not */
1034 static int osd_is_mapped(struct inode *inode, u64 offset)
1035 {
1036         sector_t (*fs_bmap)(struct address_space *, sector_t);
1037
1038         fs_bmap = inode->i_mapping->a_ops->bmap;
1039
1040         /* We can't know if we are overwriting or not */
1041         if (unlikely(fs_bmap == NULL))
1042                 return 0;
1043
1044         if (i_size_read(inode) == 0)
1045                 return 0;
1046
1047         /* Beyond EOF, must not be mapped */
1048         if (((i_size_read(inode) - 1) >> inode->i_blkbits) <
1049             (offset >> inode->i_blkbits))
1050                 return 0;
1051
1052         if (fs_bmap(inode->i_mapping, offset >> inode->i_blkbits) == 0)
1053                 return 0;
1054
1055         return 1;
1056 }
1057
1058 static int osd_declare_write_commit(const struct lu_env *env,
1059                                     struct dt_object *dt,
1060                                     struct niobuf_local *lnb, int npages,
1061                                     struct thandle *handle)
1062 {
1063         const struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1064         struct inode            *inode = osd_dt_obj(dt)->oo_inode;
1065         struct osd_thandle      *oh;
1066         int                      extents = 1;
1067         int                      depth;
1068         int                      i;
1069         int                      newblocks;
1070         int                      rc = 0;
1071         int                      flags = 0;
1072         bool                     ignore_quota = false;
1073         long long                quota_space = 0;
1074         ENTRY;
1075
1076         LASSERT(handle != NULL);
1077         oh = container_of0(handle, struct osd_thandle, ot_super);
1078         LASSERT(oh->ot_handle == NULL);
1079
1080         newblocks = npages;
1081
1082         /* calculate number of extents (probably better to pass nb) */
1083         for (i = 0; i < npages; i++) {
1084                 if (i && lnb[i].lnb_file_offset !=
1085                     lnb[i - 1].lnb_file_offset + lnb[i - 1].lnb_len)
1086                         extents++;
1087
1088                 if (!osd_is_mapped(inode, lnb[i].lnb_file_offset))
1089                         quota_space += PAGE_CACHE_SIZE;
1090
1091                 /* ignore quota for the whole request if any page is from
1092                  * client cache or written by root.
1093                  *
1094                  * XXX once we drop the 1.8 client support, the checking
1095                  * for whether page is from cache can be simplified as:
1096                  * !(lnb[i].flags & OBD_BRW_SYNC)
1097                  *
1098                  * XXX we could handle this on per-lnb basis as done by
1099                  * grant. */
1100                 if ((lnb[i].lnb_flags & OBD_BRW_NOQUOTA) ||
1101                     (lnb[i].lnb_flags & (OBD_BRW_FROM_GRANT | OBD_BRW_SYNC)) ==
1102                     OBD_BRW_FROM_GRANT)
1103                         ignore_quota = true;
1104         }
1105
1106         /*
1107          * each extent can go into new leaf causing a split
1108          * 5 is max tree depth: inode + 4 index blocks
1109          * with blockmaps, depth is 3 at most
1110          */
1111         if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL) {
1112                 /*
1113                  * many concurrent threads may grow tree by the time
1114                  * our transaction starts. so, consider 2 is a min depth
1115                  */
1116                 depth = ext_depth(inode);
1117                 depth = max(depth, 1) + 1;
1118                 newblocks += depth;
1119                 oh->ot_credits++; /* inode */
1120                 oh->ot_credits += depth * 2 * extents;
1121         } else {
1122                 depth = 3;
1123                 newblocks += depth;
1124                 oh->ot_credits++; /* inode */
1125                 oh->ot_credits += depth * extents;
1126         }
1127
1128         /* quota space for metadata blocks */
1129         quota_space += depth * extents * LDISKFS_BLOCK_SIZE(osd_sb(osd));
1130
1131         /* quota space should be reported in 1K blocks */
1132         quota_space = toqb(quota_space);
1133
1134         /* each new block can go in different group (bitmap + gd) */
1135
1136         /* we can't dirty more bitmap blocks than exist */
1137         if (newblocks > LDISKFS_SB(osd_sb(osd))->s_groups_count)
1138                 oh->ot_credits += LDISKFS_SB(osd_sb(osd))->s_groups_count;
1139         else
1140                 oh->ot_credits += newblocks;
1141
1142         /* we can't dirty more gd blocks than exist */
1143         if (newblocks > LDISKFS_SB(osd_sb(osd))->s_gdb_count)
1144                 oh->ot_credits += LDISKFS_SB(osd_sb(osd))->s_gdb_count;
1145         else
1146                 oh->ot_credits += newblocks;
1147
1148         /* make sure the over quota flags were not set */
1149         lnb[0].lnb_flags &= ~(OBD_BRW_OVER_USRQUOTA | OBD_BRW_OVER_GRPQUOTA);
1150
1151         rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
1152                                    quota_space, oh, osd_dt_obj(dt), true,
1153                                    &flags, ignore_quota);
1154
1155         /* we need only to store the overquota flags in the first lnb for
1156          * now, once we support multiple objects BRW, this code needs be
1157          * revised. */
1158         if (flags & QUOTA_FL_OVER_USRQUOTA)
1159                 lnb[0].lnb_flags |= OBD_BRW_OVER_USRQUOTA;
1160         if (flags & QUOTA_FL_OVER_GRPQUOTA)
1161                 lnb[0].lnb_flags |= OBD_BRW_OVER_GRPQUOTA;
1162
1163         RETURN(rc);
1164 }
1165
1166 /* Check if a block is allocated or not */
1167 static int osd_write_commit(const struct lu_env *env, struct dt_object *dt,
1168                             struct niobuf_local *lnb, int npages,
1169                             struct thandle *thandle)
1170 {
1171         struct osd_thread_info *oti = osd_oti_get(env);
1172         struct osd_iobuf *iobuf = &oti->oti_iobuf;
1173         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1174         struct osd_device  *osd = osd_obj2dev(osd_dt_obj(dt));
1175         loff_t isize;
1176         int rc = 0, i;
1177
1178         LASSERT(inode);
1179
1180         rc = osd_init_iobuf(osd, iobuf, 1, npages);
1181         if (unlikely(rc != 0))
1182                 RETURN(rc);
1183
1184         isize = i_size_read(inode);
1185         ll_vfs_dq_init(inode);
1186
1187         for (i = 0; i < npages; i++) {
1188                 if (lnb[i].lnb_rc == -ENOSPC &&
1189                     osd_is_mapped(inode, lnb[i].lnb_file_offset)) {
1190                         /* Allow the write to proceed if overwriting an
1191                          * existing block */
1192                         lnb[i].lnb_rc = 0;
1193                 }
1194
1195                 if (lnb[i].lnb_rc) { /* ENOSPC, network RPC error, etc. */
1196                         CDEBUG(D_INODE, "Skipping [%d] == %d\n", i,
1197                                lnb[i].lnb_rc);
1198                         LASSERT(lnb[i].lnb_page);
1199                         generic_error_remove_page(inode->i_mapping,
1200                                                   lnb[i].lnb_page);
1201                         continue;
1202                 }
1203
1204                 LASSERT(PageLocked(lnb[i].lnb_page));
1205                 LASSERT(!PageWriteback(lnb[i].lnb_page));
1206
1207                 if (lnb[i].lnb_file_offset + lnb[i].lnb_len > isize)
1208                         isize = lnb[i].lnb_file_offset + lnb[i].lnb_len;
1209
1210                 /*
1211                  * Since write and truncate are serialized by oo_sem, even
1212                  * partial-page truncate should not leave dirty pages in the
1213                  * page cache.
1214                  */
1215                 LASSERT(!PageDirty(lnb[i].lnb_page));
1216
1217                 SetPageUptodate(lnb[i].lnb_page);
1218
1219                 osd_iobuf_add_page(iobuf, lnb[i].lnb_page);
1220         }
1221
1222         if (OBD_FAIL_CHECK(OBD_FAIL_OST_MAPBLK_ENOSPC)) {
1223                 rc = -ENOSPC;
1224         } else if (iobuf->dr_npages > 0) {
1225                 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1226                                                  iobuf->dr_npages,
1227                                                  iobuf->dr_blocks, 1);
1228         } else {
1229                 /* no pages to write, no transno is needed */
1230                 thandle->th_local = 1;
1231         }
1232
1233         if (likely(rc == 0)) {
1234                 if (isize > i_size_read(inode)) {
1235                         i_size_write(inode, isize);
1236                         LDISKFS_I(inode)->i_disksize = isize;
1237                         ll_dirty_inode(inode, I_DIRTY_DATASYNC);
1238                 }
1239
1240                 rc = osd_do_bio(osd, inode, iobuf);
1241                 /* we don't do stats here as in read path because
1242                  * write is async: we'll do this in osd_put_bufs() */
1243         } else {
1244                 osd_fini_iobuf(osd, iobuf);
1245         }
1246
1247         if (unlikely(rc != 0)) {
1248                 /* if write fails, we should drop pages from the cache */
1249                 for (i = 0; i < npages; i++) {
1250                         if (lnb[i].lnb_page == NULL)
1251                                 continue;
1252                         LASSERT(PageLocked(lnb[i].lnb_page));
1253                         generic_error_remove_page(inode->i_mapping,
1254                                                   lnb[i].lnb_page);
1255                 }
1256         }
1257
1258         RETURN(rc);
1259 }
1260
1261 static int osd_read_prep(const struct lu_env *env, struct dt_object *dt,
1262                          struct niobuf_local *lnb, int npages)
1263 {
1264         struct osd_thread_info *oti = osd_oti_get(env);
1265         struct osd_iobuf *iobuf = &oti->oti_iobuf;
1266         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1267         struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1268         struct timeval start, end;
1269         unsigned long timediff;
1270         int rc = 0, i, m = 0, cache = 0, cache_hits = 0, cache_misses = 0;
1271
1272         LASSERT(inode);
1273
1274         rc = osd_init_iobuf(osd, iobuf, 0, npages);
1275         if (unlikely(rc != 0))
1276                 RETURN(rc);
1277
1278         if (osd->od_read_cache)
1279                 cache = 1;
1280         if (i_size_read(inode) > osd->od_readcache_max_filesize)
1281                 cache = 0;
1282
1283         do_gettimeofday(&start);
1284         for (i = 0; i < npages; i++) {
1285
1286                 if (i_size_read(inode) <= lnb[i].lnb_file_offset)
1287                         /* If there's no more data, abort early.
1288                          * lnb->lnb_rc == 0, so it's easy to detect later. */
1289                         break;
1290
1291                 if (i_size_read(inode) <
1292                     lnb[i].lnb_file_offset + lnb[i].lnb_len - 1)
1293                         lnb[i].lnb_rc = i_size_read(inode) -
1294                                 lnb[i].lnb_file_offset;
1295                 else
1296                         lnb[i].lnb_rc = lnb[i].lnb_len;
1297                 m += lnb[i].lnb_len;
1298
1299                 if (PageUptodate(lnb[i].lnb_page)) {
1300                         cache_hits++;
1301                 } else {
1302                         cache_misses++;
1303                         osd_iobuf_add_page(iobuf, lnb[i].lnb_page);
1304                 }
1305
1306                 if (cache == 0)
1307                         generic_error_remove_page(inode->i_mapping,
1308                                                   lnb[i].lnb_page);
1309         }
1310         do_gettimeofday(&end);
1311         timediff = cfs_timeval_sub(&end, &start, NULL);
1312         lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
1313
1314         if (cache_hits != 0)
1315                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_HIT,
1316                                     cache_hits);
1317         if (cache_misses != 0)
1318                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_MISS,
1319                                     cache_misses);
1320         if (cache_hits + cache_misses != 0)
1321                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_ACCESS,
1322                                     cache_hits + cache_misses);
1323
1324         if (iobuf->dr_npages) {
1325                 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1326                                                  iobuf->dr_npages,
1327                                                  iobuf->dr_blocks, 0);
1328                 rc = osd_do_bio(osd, inode, iobuf);
1329
1330                 /* IO stats will be done in osd_bufs_put() */
1331         }
1332
1333         RETURN(rc);
1334 }
1335
1336 /*
1337  * XXX: Another layering violation for now.
1338  *
1339  * We don't want to use ->f_op->read methods, because generic file write
1340  *
1341  *         - serializes on ->i_sem, and
1342  *
1343  *         - does a lot of extra work like balance_dirty_pages(),
1344  *
1345  * which doesn't work for globally shared files like /last_rcvd.
1346  */
1347 static int osd_ldiskfs_readlink(struct inode *inode, char *buffer, int buflen)
1348 {
1349         struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
1350
1351         memcpy(buffer, (char *)ei->i_data, buflen);
1352
1353         return  buflen;
1354 }
1355
1356 int osd_ldiskfs_read(struct inode *inode, void *buf, int size, loff_t *offs)
1357 {
1358         struct buffer_head *bh;
1359         unsigned long block;
1360         int osize;
1361         int blocksize;
1362         int csize;
1363         int boffs;
1364         int err;
1365
1366         /* prevent reading after eof */
1367         spin_lock(&inode->i_lock);
1368         if (i_size_read(inode) < *offs + size) {
1369                 loff_t diff = i_size_read(inode) - *offs;
1370                 spin_unlock(&inode->i_lock);
1371                 if (diff < 0) {
1372                         CDEBUG(D_EXT2, "size %llu is too short to read @%llu\n",
1373                                i_size_read(inode), *offs);
1374                         return -EBADR;
1375                 } else if (diff == 0) {
1376                         return 0;
1377                 } else {
1378                         size = diff;
1379                 }
1380         } else {
1381                 spin_unlock(&inode->i_lock);
1382         }
1383
1384         blocksize = 1 << inode->i_blkbits;
1385         osize = size;
1386         while (size > 0) {
1387                 block = *offs >> inode->i_blkbits;
1388                 boffs = *offs & (blocksize - 1);
1389                 csize = min(blocksize - boffs, size);
1390                 bh = ldiskfs_bread(NULL, inode, block, 0, &err);
1391                 if (err != 0) {
1392                         CERROR("%s: can't read %u@%llu on ino %lu: rc = %d\n",
1393                                LDISKFS_SB(inode->i_sb)->s_es->s_volume_name,
1394                                csize, *offs, inode->i_ino, err);
1395                         if (bh != NULL)
1396                                 brelse(bh);
1397                         return err;
1398                 }
1399
1400                 if (bh != NULL) {
1401                         memcpy(buf, bh->b_data + boffs, csize);
1402                         brelse(bh);
1403                 } else {
1404                         memset(buf, 0, csize);
1405                 }
1406
1407                 *offs += csize;
1408                 buf += csize;
1409                 size -= csize;
1410         }
1411         return osize;
1412 }
1413
1414 static ssize_t osd_read(const struct lu_env *env, struct dt_object *dt,
1415                         struct lu_buf *buf, loff_t *pos,
1416                         struct lustre_capa *capa)
1417 {
1418         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1419         int           rc;
1420
1421         if (osd_object_auth(env, dt, capa, CAPA_OPC_BODY_READ))
1422                 RETURN(-EACCES);
1423
1424         /* Read small symlink from inode body as we need to maintain correct
1425          * on-disk symlinks for ldiskfs.
1426          */
1427         if (S_ISLNK(dt->do_lu.lo_header->loh_attr) &&
1428             (buf->lb_len < sizeof(LDISKFS_I(inode)->i_data)))
1429                 rc = osd_ldiskfs_readlink(inode, buf->lb_buf, buf->lb_len);
1430         else
1431                 rc = osd_ldiskfs_read(inode, buf->lb_buf, buf->lb_len, pos);
1432
1433         return rc;
1434 }
1435
1436 static inline int osd_extents_enabled(struct super_block *sb,
1437                                       struct inode *inode)
1438 {
1439         if (inode != NULL) {
1440                 if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL)
1441                         return 1;
1442         } else if (test_opt(sb, EXTENTS)) {
1443                 return 1;
1444         }
1445         return 0;
1446 }
1447
1448 static inline int osd_calc_bkmap_credits(struct super_block *sb,
1449                                          struct inode *inode,
1450                                          const loff_t size,
1451                                          const loff_t pos,
1452                                          const int blocks)
1453 {
1454         int credits, bits, bs, i;
1455
1456         bits = sb->s_blocksize_bits;
1457         bs = 1 << bits;
1458
1459         /* legacy blockmap: 3 levels * 3 (bitmap,gd,itself)
1460          * we do not expect blockmaps on the large files,
1461          * so let's shrink it to 2 levels (4GB files) */
1462
1463         /* this is default reservation: 2 levels */
1464         credits = (blocks + 2) * 3;
1465
1466         /* actual offset is unknown, hard to optimize */
1467         if (pos == -1)
1468                 return credits;
1469
1470         /* now check for few specific cases to optimize */
1471         if (pos + size <= LDISKFS_NDIR_BLOCKS * bs) {
1472                 /* no indirects */
1473                 credits = blocks;
1474                 /* allocate if not allocated */
1475                 if (inode == NULL) {
1476                         credits += blocks * 2;
1477                         return credits;
1478                 }
1479                 for (i = (pos >> bits); i < (pos >> bits) + blocks; i++) {
1480                         LASSERT(i < LDISKFS_NDIR_BLOCKS);
1481                         if (LDISKFS_I(inode)->i_data[i] == 0)
1482                                 credits += 2;
1483                 }
1484         } else if (pos + size <= (LDISKFS_NDIR_BLOCKS + 1024) * bs) {
1485                 /* single indirect */
1486                 credits = blocks * 3;
1487                 /* probably indirect block has been allocated already */
1488                 if (!inode || LDISKFS_I(inode)->i_data[LDISKFS_IND_BLOCK])
1489                         credits += 3;
1490         }
1491
1492         return credits;
1493 }
1494
1495 static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt,
1496                                  const struct lu_buf *buf, loff_t _pos,
1497                                  struct thandle *handle)
1498 {
1499         struct osd_object  *obj  = osd_dt_obj(dt);
1500         struct inode       *inode = obj->oo_inode;
1501         struct super_block *sb = osd_sb(osd_obj2dev(obj));
1502         struct osd_thandle *oh;
1503         int                 rc = 0, est = 0, credits, blocks, allocated = 0;
1504         int                 bits, bs;
1505         int                 depth, size;
1506         loff_t              pos;
1507         ENTRY;
1508
1509         LASSERT(buf != NULL);
1510         LASSERT(handle != NULL);
1511
1512         oh = container_of0(handle, struct osd_thandle, ot_super);
1513         LASSERT(oh->ot_handle == NULL);
1514
1515         size = buf->lb_len;
1516         bits = sb->s_blocksize_bits;
1517         bs = 1 << bits;
1518
1519         if (_pos == -1) {
1520                 /* if this is an append, then we
1521                  * should expect cross-block record */
1522                 pos = 0;
1523         } else {
1524                 pos = _pos;
1525         }
1526
1527         /* blocks to modify */
1528         blocks = ((pos + size + bs - 1) >> bits) - (pos >> bits);
1529         LASSERT(blocks > 0);
1530
1531         if (inode != NULL && _pos != -1) {
1532                 /* object size in blocks */
1533                 est = (i_size_read(inode) + bs - 1) >> bits;
1534                 allocated = inode->i_blocks >> (bits - 9);
1535                 if (pos + size <= i_size_read(inode) && est <= allocated) {
1536                         /* looks like an overwrite, no need to modify tree */
1537                         credits = blocks;
1538                         /* no need to modify i_size */
1539                         goto out;
1540                 }
1541         }
1542
1543         if (osd_extents_enabled(sb, inode)) {
1544                 /*
1545                  * many concurrent threads may grow tree by the time
1546                  * our transaction starts. so, consider 2 is a min depth
1547                  * for every level we may need to allocate a new block
1548                  * and take some entries from the old one. so, 3 blocks
1549                  * to allocate (bitmap, gd, itself) + old block - 4 per
1550                  * level.
1551                  */
1552                 depth = inode != NULL ? ext_depth(inode) : 0;
1553                 depth = max(depth, 1) + 1;
1554                 credits = depth;
1555                 /* if not append, then split may need to modify
1556                  * existing blocks moving entries into the new ones */
1557                 if (_pos == -1)
1558                         credits += depth;
1559                 /* blocks to store data: bitmap,gd,itself */
1560                 credits += blocks * 3;
1561         } else {
1562                 credits = osd_calc_bkmap_credits(sb, inode, size, _pos, blocks);
1563         }
1564         /* if inode is created as part of the transaction,
1565          * then it's counted already by the creation method */
1566         if (inode != NULL)
1567                 credits++;
1568
1569 out:
1570
1571         osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
1572
1573         /* dt_declare_write() is usually called for system objects, such
1574          * as llog or last_rcvd files. We needn't enforce quota on those
1575          * objects, so always set the lqi_space as 0. */
1576         if (inode != NULL)
1577                 rc = osd_declare_inode_qid(env, i_uid_read(inode),
1578                                            i_gid_read(inode), 0, oh, obj, true,
1579                                            NULL, false);
1580         RETURN(rc);
1581 }
1582
1583 static int osd_ldiskfs_writelink(struct inode *inode, char *buffer, int buflen)
1584 {
1585         /* LU-2634: clear the extent format for fast symlink */
1586         ldiskfs_clear_inode_flag(inode, LDISKFS_INODE_EXTENTS);
1587
1588         memcpy((char *)&LDISKFS_I(inode)->i_data, (char *)buffer, buflen);
1589         LDISKFS_I(inode)->i_disksize = buflen;
1590         i_size_write(inode, buflen);
1591         ll_dirty_inode(inode, I_DIRTY_DATASYNC);
1592
1593         return 0;
1594 }
1595
1596 int osd_ldiskfs_write_record(struct inode *inode, void *buf, int bufsize,
1597                              int write_NUL, loff_t *offs, handle_t *handle)
1598 {
1599         struct buffer_head *bh        = NULL;
1600         loff_t              offset    = *offs;
1601         loff_t              new_size  = i_size_read(inode);
1602         unsigned long       block;
1603         int                 blocksize = 1 << inode->i_blkbits;
1604         int                 err = 0;
1605         int                 size;
1606         int                 boffs;
1607         int                 dirty_inode = 0;
1608
1609         if (write_NUL) {
1610                 /*
1611                  * long symlink write does not count the NUL terminator in
1612                  * bufsize, we write it, and the inode's file size does not
1613                  * count the NUL terminator as well.
1614                  */
1615                 ((char *)buf)[bufsize] = '\0';
1616                 ++bufsize;
1617         }
1618         while (bufsize > 0) {
1619                 if (bh != NULL)
1620                         brelse(bh);
1621
1622                 block = offset >> inode->i_blkbits;
1623                 boffs = offset & (blocksize - 1);
1624                 size = min(blocksize - boffs, bufsize);
1625                 bh = ldiskfs_bread(handle, inode, block, 1, &err);
1626                 if (!bh) {
1627                         CERROR("%s: error reading offset %llu (block %lu): "
1628                                "rc = %d\n",
1629                                inode->i_sb->s_id, offset, block, err);
1630                         break;
1631                 }
1632
1633                 err = ldiskfs_journal_get_write_access(handle, bh);
1634                 if (err) {
1635                         CERROR("journal_get_write_access() returned error %d\n",
1636                                err);
1637                         break;
1638                 }
1639                 LASSERTF(boffs + size <= bh->b_size,
1640                          "boffs %d size %d bh->b_size %lu\n",
1641                          boffs, size, (unsigned long)bh->b_size);
1642                 memcpy(bh->b_data + boffs, buf, size);
1643                 err = ldiskfs_handle_dirty_metadata(handle, NULL, bh);
1644                 if (err)
1645                         break;
1646
1647                 if (offset + size > new_size)
1648                         new_size = offset + size;
1649                 offset += size;
1650                 bufsize -= size;
1651                 buf += size;
1652         }
1653         if (bh)
1654                 brelse(bh);
1655
1656         if (write_NUL)
1657                 --new_size;
1658         /* correct in-core and on-disk sizes */
1659         if (new_size > i_size_read(inode)) {
1660                 spin_lock(&inode->i_lock);
1661                 if (new_size > i_size_read(inode))
1662                         i_size_write(inode, new_size);
1663                 if (i_size_read(inode) > LDISKFS_I(inode)->i_disksize) {
1664                         LDISKFS_I(inode)->i_disksize = i_size_read(inode);
1665                         dirty_inode = 1;
1666                 }
1667                 spin_unlock(&inode->i_lock);
1668                 if (dirty_inode)
1669                         ll_dirty_inode(inode, I_DIRTY_DATASYNC);
1670         }
1671
1672         if (err == 0)
1673                 *offs = offset;
1674         return err;
1675 }
1676
1677 static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt,
1678                          const struct lu_buf *buf, loff_t *pos,
1679                          struct thandle *handle, struct lustre_capa *capa,
1680                          int ignore_quota)
1681 {
1682         struct inode            *inode = osd_dt_obj(dt)->oo_inode;
1683         struct osd_thandle      *oh;
1684         ssize_t                 result;
1685         int                     is_link;
1686
1687         LASSERT(dt_object_exists(dt));
1688
1689         if (osd_object_auth(env, dt, capa, CAPA_OPC_BODY_WRITE))
1690                 return -EACCES;
1691
1692         LASSERT(handle != NULL);
1693         LASSERT(inode != NULL);
1694         ll_vfs_dq_init(inode);
1695
1696         /* XXX: don't check: one declared chunk can be used many times */
1697         /* osd_trans_exec_op(env, handle, OSD_OT_WRITE); */
1698
1699         oh = container_of(handle, struct osd_thandle, ot_super);
1700         LASSERT(oh->ot_handle->h_transaction != NULL);
1701         /* Write small symlink to inode body as we need to maintain correct
1702          * on-disk symlinks for ldiskfs.
1703          * Note: the buf->lb_buf contains a NUL terminator while buf->lb_len
1704          * does not count it in.
1705          */
1706         is_link = S_ISLNK(dt->do_lu.lo_header->loh_attr);
1707         if (is_link && (buf->lb_len < sizeof(LDISKFS_I(inode)->i_data)))
1708                 result = osd_ldiskfs_writelink(inode, buf->lb_buf, buf->lb_len);
1709         else
1710                 result = osd_ldiskfs_write_record(inode, buf->lb_buf,
1711                                                   buf->lb_len, is_link, pos,
1712                                                   oh->ot_handle);
1713         if (result == 0)
1714                 result = buf->lb_len;
1715         return result;
1716 }
1717
1718 static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt,
1719                              __u64 start, __u64 end, struct thandle *th)
1720 {
1721         struct osd_thandle *oh;
1722         struct inode       *inode;
1723         int                 rc;
1724         ENTRY;
1725
1726         LASSERT(th);
1727         oh = container_of(th, struct osd_thandle, ot_super);
1728
1729         /*
1730          * we don't need to reserve credits for whole truncate
1731          * it's not possible as truncate may need to free too many
1732          * blocks and that won't fit a single transaction. instead
1733          * we reserve credits to change i_size and put inode onto
1734          * orphan list. if needed truncate will extend or restart
1735          * transaction
1736          */
1737         osd_trans_declare_op(env, oh, OSD_OT_PUNCH,
1738                              osd_dto_credits_noquota[DTO_ATTR_SET_BASE] + 3);
1739
1740         inode = osd_dt_obj(dt)->oo_inode;
1741         LASSERT(inode);
1742
1743         rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
1744                                    0, oh, osd_dt_obj(dt), true, NULL, false);
1745         RETURN(rc);
1746 }
1747
1748 static int osd_punch(const struct lu_env *env, struct dt_object *dt,
1749                      __u64 start, __u64 end, struct thandle *th,
1750                      struct lustre_capa *capa)
1751 {
1752         struct osd_thandle *oh;
1753         struct osd_object  *obj = osd_dt_obj(dt);
1754         struct inode       *inode = obj->oo_inode;
1755         handle_t           *h;
1756         tid_t               tid;
1757         int                rc = 0, rc2 = 0;
1758         ENTRY;
1759
1760         LASSERT(end == OBD_OBJECT_EOF);
1761         LASSERT(dt_object_exists(dt));
1762         LASSERT(osd_invariant(obj));
1763         LASSERT(inode != NULL);
1764         ll_vfs_dq_init(inode);
1765
1766         LASSERT(th);
1767         oh = container_of(th, struct osd_thandle, ot_super);
1768         LASSERT(oh->ot_handle->h_transaction != NULL);
1769
1770         osd_trans_exec_op(env, th, OSD_OT_PUNCH);
1771
1772         tid = oh->ot_handle->h_transaction->t_tid;
1773
1774         i_size_write(inode, start);
1775         ll_truncate_pagecache(inode, start);
1776 #ifdef HAVE_INODEOPS_TRUNCATE
1777         if (inode->i_op->truncate) {
1778                 inode->i_op->truncate(inode);
1779         } else
1780 #endif
1781                 ldiskfs_truncate(inode);
1782
1783         /*
1784          * For a partial-page truncate, flush the page to disk immediately to
1785          * avoid data corruption during direct disk write.  b=17397
1786          */
1787         if ((start & ~CFS_PAGE_MASK) != 0)
1788                 rc = filemap_fdatawrite_range(inode->i_mapping, start, start+1);
1789
1790         h = journal_current_handle();
1791         LASSERT(h != NULL);
1792         LASSERT(h == oh->ot_handle);
1793
1794         if (tid != h->h_transaction->t_tid) {
1795                 int credits = oh->ot_credits;
1796                 /*
1797                  * transaction has changed during truncate
1798                  * we need to restart the handle with our credits
1799                  */
1800                 if (h->h_buffer_credits < credits) {
1801                         if (ldiskfs_journal_extend(h, credits))
1802                                 rc2 = ldiskfs_journal_restart(h, credits);
1803                 }
1804         }
1805
1806         RETURN(rc == 0 ? rc2 : rc);
1807 }
1808
1809 static int osd_fiemap_get(const struct lu_env *env, struct dt_object *dt,
1810                           struct ll_user_fiemap *fm)
1811 {
1812         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1813         struct osd_thread_info *info   = osd_oti_get(env);
1814         struct dentry          *dentry = &info->oti_obj_dentry;
1815         struct file            *file   = &info->oti_file;
1816         mm_segment_t            saved_fs;
1817         int rc;
1818
1819         LASSERT(inode);
1820         dentry->d_inode = inode;
1821         dentry->d_sb = inode->i_sb;
1822         file->f_dentry = dentry;
1823         file->f_mapping = inode->i_mapping;
1824         file->f_op = inode->i_fop;
1825         set_file_inode(file, inode);
1826
1827         saved_fs = get_fs();
1828         set_fs(get_ds());
1829         /* ldiskfs_ioctl does not have a inode argument */
1830         if (inode->i_fop->unlocked_ioctl)
1831                 rc = inode->i_fop->unlocked_ioctl(file, FSFILT_IOC_FIEMAP,
1832                                                   (long)fm);
1833         else
1834                 rc = -ENOTTY;
1835         set_fs(saved_fs);
1836         return rc;
1837 }
1838
1839 /*
1840  * in some cases we may need declare methods for objects being created
1841  * e.g., when we create symlink
1842  */
1843 const struct dt_body_operations osd_body_ops_new = {
1844         .dbo_declare_write = osd_declare_write,
1845 };
1846
1847 const struct dt_body_operations osd_body_ops = {
1848         .dbo_read                 = osd_read,
1849         .dbo_declare_write        = osd_declare_write,
1850         .dbo_write                = osd_write,
1851         .dbo_bufs_get             = osd_bufs_get,
1852         .dbo_bufs_put             = osd_bufs_put,
1853         .dbo_write_prep           = osd_write_prep,
1854         .dbo_declare_write_commit = osd_declare_write_commit,
1855         .dbo_write_commit         = osd_write_commit,
1856         .dbo_read_prep            = osd_read_prep,
1857         .dbo_declare_punch         = osd_declare_punch,
1858         .dbo_punch                 = osd_punch,
1859         .dbo_fiemap_get           = osd_fiemap_get,
1860 };
1861