Whamcloud - gitweb
9095b51ba786a57d8205f6daa2b880237a500782
[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.gnu.org/licenses/gpl-2.0.html
19  *
20  * GPL HEADER END
21  */
22 /*
23  * Copyright (c) 2010, Oracle and/or its affiliates. All rights reserved.
24  * Use is subject to license terms.
25  *
26  * Copyright (c) 2012, 2017, Intel Corporation.
27  */
28 /*
29  * This file is part of Lustre, http://www.lustre.org/
30  * Lustre is a trademark of Sun Microsystems, Inc.
31  *
32  * lustre/osd/osd_io.c
33  *
34  * body operations
35  *
36  * Author: Nikita Danilov <nikita@clusterfs.com>
37  * Author: Alex Zhuravlev <bzzz@whamcloud.com>
38  *
39  */
40
41 /* prerequisite for linux/xattr.h */
42 #include <linux/types.h>
43 /* prerequisite for linux/xattr.h */
44 #include <linux/fs.h>
45 #include <linux/mm.h>
46 #include <linux/pagevec.h>
47
48 /*
49  * struct OBD_{ALLOC,FREE}*()
50  * OBD_FAIL_CHECK
51  */
52 #include <obd_support.h>
53
54 #include "osd_internal.h"
55
56 /* ext_depth() */
57 #include <ldiskfs/ldiskfs_extents.h>
58
59 static inline bool osd_use_page_cache(struct osd_device *d)
60 {
61         /* do not use pagecache if write and read caching are disabled */
62         if (d->od_writethrough_cache + d->od_read_cache == 0)
63                 return false;
64         /* use pagecache by default */
65         return true;
66 }
67
68 static int __osd_init_iobuf(struct osd_device *d, struct osd_iobuf *iobuf,
69                             int rw, int line, int pages)
70 {
71         int blocks, i;
72
73         LASSERTF(iobuf->dr_elapsed_valid == 0,
74                  "iobuf %p, reqs %d, rw %d, line %d\n", iobuf,
75                  atomic_read(&iobuf->dr_numreqs), iobuf->dr_rw,
76                  iobuf->dr_init_at);
77         LASSERT(pages <= PTLRPC_MAX_BRW_PAGES);
78
79         init_waitqueue_head(&iobuf->dr_wait);
80         atomic_set(&iobuf->dr_numreqs, 0);
81         iobuf->dr_npages = 0;
82         iobuf->dr_error = 0;
83         iobuf->dr_dev = d;
84         iobuf->dr_frags = 0;
85         iobuf->dr_elapsed = ktime_set(0, 0);
86         /* must be counted before, so assert */
87         iobuf->dr_rw = rw;
88         iobuf->dr_init_at = line;
89
90         blocks = pages * (PAGE_SIZE >> osd_sb(d)->s_blocksize_bits);
91         if (iobuf->dr_bl_buf.lb_len >= blocks * sizeof(iobuf->dr_blocks[0])) {
92                 LASSERT(iobuf->dr_pg_buf.lb_len >=
93                         pages * sizeof(iobuf->dr_pages[0]));
94                 return 0;
95         }
96
97         /* start with 1MB for 4K blocks */
98         i = 256;
99         while (i <= PTLRPC_MAX_BRW_PAGES && i < pages)
100                 i <<= 1;
101
102         CDEBUG(D_OTHER, "realloc %u for %u (%u) pages\n",
103                (unsigned)(pages * sizeof(iobuf->dr_pages[0])), i, pages);
104         pages = i;
105         blocks = pages * (PAGE_SIZE >> osd_sb(d)->s_blocksize_bits);
106         iobuf->dr_max_pages = 0;
107         CDEBUG(D_OTHER, "realloc %u for %u blocks\n",
108                (unsigned)(blocks * sizeof(iobuf->dr_blocks[0])), blocks);
109
110         lu_buf_realloc(&iobuf->dr_bl_buf, blocks * sizeof(iobuf->dr_blocks[0]));
111         iobuf->dr_blocks = iobuf->dr_bl_buf.lb_buf;
112         if (unlikely(iobuf->dr_blocks == NULL))
113                 return -ENOMEM;
114
115         lu_buf_realloc(&iobuf->dr_pg_buf, pages * sizeof(iobuf->dr_pages[0]));
116         iobuf->dr_pages = iobuf->dr_pg_buf.lb_buf;
117         if (unlikely(iobuf->dr_pages == NULL))
118                 return -ENOMEM;
119
120         lu_buf_realloc(&iobuf->dr_lnb_buf,
121                        pages * sizeof(iobuf->dr_lnbs[0]));
122         iobuf->dr_lnbs = iobuf->dr_lnb_buf.lb_buf;
123         if (unlikely(iobuf->dr_lnbs == NULL))
124                 return -ENOMEM;
125
126         iobuf->dr_max_pages = pages;
127
128         return 0;
129 }
130 #define osd_init_iobuf(dev, iobuf, rw, pages) \
131         __osd_init_iobuf(dev, iobuf, rw, __LINE__, pages)
132
133 static void osd_iobuf_add_page(struct osd_iobuf *iobuf,
134                                struct niobuf_local *lnb)
135 {
136         LASSERT(iobuf->dr_npages < iobuf->dr_max_pages);
137         iobuf->dr_pages[iobuf->dr_npages] = lnb->lnb_page;
138         iobuf->dr_lnbs[iobuf->dr_npages] = lnb;
139         iobuf->dr_npages++;
140 }
141
142 void osd_fini_iobuf(struct osd_device *d, struct osd_iobuf *iobuf)
143 {
144         int rw = iobuf->dr_rw;
145
146         if (iobuf->dr_elapsed_valid) {
147                 iobuf->dr_elapsed_valid = 0;
148                 LASSERT(iobuf->dr_dev == d);
149                 LASSERT(iobuf->dr_frags > 0);
150                 lprocfs_oh_tally(&d->od_brw_stats.
151                                  hist[BRW_R_DIO_FRAGS+rw],
152                                  iobuf->dr_frags);
153                 lprocfs_oh_tally_log2(&d->od_brw_stats.hist[BRW_R_IO_TIME+rw],
154                                       ktime_to_ms(iobuf->dr_elapsed));
155         }
156 }
157
158 #ifdef HAVE_BIO_ENDIO_USES_ONE_ARG
159 static void dio_complete_routine(struct bio *bio)
160 {
161         int error = bio->bi_status;
162 #else
163 static void dio_complete_routine(struct bio *bio, int error)
164 {
165 #endif
166         struct osd_iobuf *iobuf = bio->bi_private;
167         struct bio_vec *bvl;
168
169         /* CAVEAT EMPTOR: possibly in IRQ context
170          * DO NOT record procfs stats here!!! */
171
172         if (unlikely(iobuf == NULL)) {
173                 CERROR("***** bio->bi_private is NULL!  This should never "
174                        "happen.  Normally, I would crash here, but instead I "
175                        "will dump the bio contents to the console.  Please "
176                        "report this to <https://jira.whamcloud.com/> , along "
177                        "with any interesting messages leading up to this point "
178                        "(like SCSI errors, perhaps).  Because bi_private is "
179                        "NULL, I can't wake up the thread that initiated this "
180                        "IO - you will probably have to reboot this node.\n");
181                 CERROR("bi_next: %p, bi_flags: %lx, " __stringify(bi_opf)
182                        ": %x, bi_vcnt: %d, bi_idx: %d, bi->size: %d, bi_end_io: %p, bi_cnt: %d, bi_private: %p\n",
183                        bio->bi_next, (unsigned long)bio->bi_flags,
184                        (unsigned int)bio->bi_opf, bio->bi_vcnt, bio_idx(bio),
185                        bio_sectors(bio) << 9, bio->bi_end_io,
186                        atomic_read(&bio->__bi_cnt),
187                        bio->bi_private);
188                 return;
189         }
190
191         /* the check is outside of the cycle for performance reason -bzzz */
192         if (!bio_data_dir(bio)) {
193                 DECLARE_BVEC_ITER_ALL(iter_all);
194
195                 bio_for_each_segment_all(bvl, bio, iter_all) {
196                         if (likely(error == 0))
197                                 SetPageUptodate(bvl_to_page(bvl));
198                         LASSERT(PageLocked(bvl_to_page(bvl)));
199                 }
200                 atomic_dec(&iobuf->dr_dev->od_r_in_flight);
201         } else {
202                 atomic_dec(&iobuf->dr_dev->od_w_in_flight);
203         }
204
205         /* any real error is good enough -bzzz */
206         if (error != 0 && iobuf->dr_error == 0)
207                 iobuf->dr_error = error;
208
209         /*
210          * set dr_elapsed before dr_numreqs turns to 0, otherwise
211          * it's possible that service thread will see dr_numreqs
212          * is zero, but dr_elapsed is not set yet, leading to lost
213          * data in this processing and an assertion in a subsequent
214          * call to OSD.
215          */
216         if (atomic_read(&iobuf->dr_numreqs) == 1) {
217                 ktime_t now = ktime_get();
218
219                 iobuf->dr_elapsed = ktime_sub(now, iobuf->dr_start_time);
220                 iobuf->dr_elapsed_valid = 1;
221         }
222         if (atomic_dec_and_test(&iobuf->dr_numreqs))
223                 wake_up(&iobuf->dr_wait);
224
225         /* Completed bios used to be chained off iobuf->dr_bios and freed in
226          * filter_clear_dreq().  It was then possible to exhaust the biovec-256
227          * mempool when serious on-disk fragmentation was encountered,
228          * deadlocking the OST.  The bios are now released as soon as complete
229          * so the pool cannot be exhausted while IOs are competing. bug 10076 */
230         bio_put(bio);
231 }
232
233 static void record_start_io(struct osd_iobuf *iobuf, int size)
234 {
235         struct osd_device    *osd = iobuf->dr_dev;
236         struct obd_histogram *h = osd->od_brw_stats.hist;
237
238         iobuf->dr_frags++;
239         atomic_inc(&iobuf->dr_numreqs);
240
241         if (iobuf->dr_rw == 0) {
242                 atomic_inc(&osd->od_r_in_flight);
243                 lprocfs_oh_tally(&h[BRW_R_RPC_HIST],
244                                  atomic_read(&osd->od_r_in_flight));
245                 lprocfs_oh_tally_log2(&h[BRW_R_DISK_IOSIZE], size);
246         } else if (iobuf->dr_rw == 1) {
247                 atomic_inc(&osd->od_w_in_flight);
248                 lprocfs_oh_tally(&h[BRW_W_RPC_HIST],
249                                  atomic_read(&osd->od_w_in_flight));
250                 lprocfs_oh_tally_log2(&h[BRW_W_DISK_IOSIZE], size);
251         } else {
252                 LBUG();
253         }
254 }
255
256 static void osd_submit_bio(int rw, struct bio *bio)
257 {
258         LASSERTF(rw == 0 || rw == 1, "%x\n", rw);
259 #ifdef HAVE_SUBMIT_BIO_2ARGS
260         submit_bio(rw ? WRITE : READ, bio);
261 #else
262         bio->bi_opf |= rw;
263         submit_bio(bio);
264 #endif
265 }
266
267 static int can_be_merged(struct bio *bio, sector_t sector)
268 {
269         if (bio == NULL)
270                 return 0;
271
272         return bio_end_sector(bio) == sector ? 1 : 0;
273 }
274
275 #if IS_ENABLED(CONFIG_BLK_DEV_INTEGRITY)
276 /*
277  * This function will change the data written, thus it should only be
278  * used when checking data integrity feature
279  */
280 static void bio_integrity_fault_inject(struct bio *bio)
281 {
282         struct bio_vec *bvec;
283         DECLARE_BVEC_ITER_ALL(iter_all);
284         void *kaddr;
285         char *addr;
286
287         bio_for_each_segment_all(bvec, bio, iter_all) {
288                 struct page *page = bvec->bv_page;
289
290                 kaddr = kmap(page);
291                 addr = kaddr;
292                 *addr = ~(*addr);
293                 kunmap(page);
294                 break;
295         }
296 }
297
298 static int bio_dif_compare(__u16 *expected_guard_buf, void *bio_prot_buf,
299                            unsigned int sectors, int tuple_size)
300 {
301         __u16 *expected_guard;
302         __u16 *bio_guard;
303         int i;
304
305         expected_guard = expected_guard_buf;
306         for (i = 0; i < sectors; i++) {
307                 bio_guard = (__u16 *)bio_prot_buf;
308                 if (*bio_guard != *expected_guard) {
309                         CERROR("unexpected guard tags on sector %d "
310                                "expected guard %u, bio guard "
311                                "%u, sectors %u, tuple size %d\n",
312                                i, *expected_guard, *bio_guard, sectors,
313                                tuple_size);
314                         return -EIO;
315                 }
316                 expected_guard++;
317                 bio_prot_buf += tuple_size;
318         }
319         return 0;
320 }
321
322 static int osd_bio_integrity_compare(struct bio *bio, struct block_device *bdev,
323                                      struct osd_iobuf *iobuf, int index)
324 {
325         struct blk_integrity *bi = bdev_get_integrity(bdev);
326         struct bio_integrity_payload *bip = bio->bi_integrity;
327         struct niobuf_local *lnb;
328         unsigned short sector_size = blk_integrity_interval(bi);
329         void *bio_prot_buf = page_address(bip->bip_vec->bv_page) +
330                 bip->bip_vec->bv_offset;
331         struct bio_vec *bv;
332         sector_t sector = bio_start_sector(bio);
333         unsigned int sectors, total;
334         DECLARE_BVEC_ITER_ALL(iter_all);
335         __u16 *expected_guard;
336         int rc;
337
338         total = 0;
339         bio_for_each_segment_all(bv, bio, iter_all) {
340                 lnb = iobuf->dr_lnbs[index];
341                 expected_guard = lnb->lnb_guards;
342                 sectors = bv->bv_len / sector_size;
343                 if (lnb->lnb_guard_rpc) {
344                         rc = bio_dif_compare(expected_guard, bio_prot_buf,
345                                              sectors, bi->tuple_size);
346                         if (rc)
347                                 return rc;
348                 }
349
350                 sector += sectors;
351                 bio_prot_buf += sectors * bi->tuple_size;
352                 total += sectors * bi->tuple_size;
353                 LASSERT(total <= bip_size(bio->bi_integrity));
354                 index++;
355         }
356         return 0;
357 }
358
359 static int osd_bio_integrity_handle(struct osd_device *osd, struct bio *bio,
360                                     struct osd_iobuf *iobuf,
361                                     int start_page_idx, bool fault_inject,
362                                     bool integrity_enabled)
363 {
364         struct super_block *sb = osd_sb(osd);
365         integrity_gen_fn *generate_fn = NULL;
366         integrity_vrfy_fn *verify_fn = NULL;
367         int rc;
368
369         ENTRY;
370
371         if (!integrity_enabled)
372                 RETURN(0);
373
374         rc = osd_get_integrity_profile(osd, &generate_fn, &verify_fn);
375         if (rc)
376                 RETURN(rc);
377
378         rc = bio_integrity_prep_fn(bio, generate_fn, verify_fn);
379         if (rc)
380                 RETURN(rc);
381
382         /* Verify and inject fault only when writing */
383         if (iobuf->dr_rw == 1) {
384                 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_OST_INTEGRITY_CMP))) {
385                         rc = osd_bio_integrity_compare(bio, sb->s_bdev, iobuf,
386                                                        start_page_idx);
387                         if (rc)
388                                 RETURN(rc);
389                 }
390
391                 if (unlikely(fault_inject))
392                         bio_integrity_fault_inject(bio);
393         }
394
395         RETURN(0);
396 }
397
398 #ifdef HAVE_BIO_INTEGRITY_PREP_FN
399 #  ifdef HAVE_BIO_ENDIO_USES_ONE_ARG
400 static void dio_integrity_complete_routine(struct bio *bio)
401 #  else
402 static void dio_integrity_complete_routine(struct bio *bio, int error)
403 #  endif
404 {
405         struct osd_bio_private *bio_private = bio->bi_private;
406
407         bio->bi_private = bio_private->obp_iobuf;
408         osd_dio_complete_routine(bio, error);
409
410         OBD_FREE_PTR(bio_private);
411 }
412 #endif /* HAVE_BIO_INTEGRITY_PREP_FN */
413 #else  /* !CONFIG_BLK_DEV_INTEGRITY */
414 #define osd_bio_integrity_handle(osd, bio, iobuf, start_page_idx, \
415                                  fault_inject, integrity_enabled) 0
416 #endif /* CONFIG_BLK_DEV_INTEGRITY */
417
418 static int osd_bio_init(struct bio *bio, struct osd_iobuf *iobuf,
419                         bool integrity_enabled, int start_page_idx,
420                         struct osd_bio_private **pprivate)
421 {
422         ENTRY;
423
424         *pprivate = NULL;
425
426 #ifdef HAVE_BIO_INTEGRITY_PREP_FN
427         if (integrity_enabled) {
428                 struct osd_bio_private *bio_private = NULL;
429
430                 OBD_ALLOC_GFP(bio_private, sizeof(*bio_private), GFP_NOIO);
431                 if (bio_private == NULL)
432                         RETURN(-ENOMEM);
433                 bio->bi_end_io = dio_integrity_complete_routine;
434                 bio->bi_private = bio_private;
435                 bio_private->obp_start_page_idx = start_page_idx;
436                 bio_private->obp_iobuf = iobuf;
437                 *pprivate = bio_private;
438         } else
439 #endif
440         {
441                 bio->bi_end_io = dio_complete_routine;
442                 bio->bi_private = iobuf;
443         }
444
445         RETURN(0);
446 }
447
448 static int osd_do_bio(struct osd_device *osd, struct inode *inode,
449                       struct osd_iobuf *iobuf)
450 {
451         int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
452         struct page **pages = iobuf->dr_pages;
453         int npages = iobuf->dr_npages;
454         sector_t *blocks = iobuf->dr_blocks;
455         int total_blocks = npages * blocks_per_page;
456         struct super_block *sb = inode->i_sb;
457         int sector_bits = sb->s_blocksize_bits - 9;
458         unsigned int blocksize = sb->s_blocksize;
459         struct block_device *bdev = sb->s_bdev;
460         struct osd_bio_private *bio_private = NULL;
461         struct bio *bio = NULL;
462         int bio_start_page_idx;
463         struct page *page;
464         unsigned int page_offset;
465         sector_t sector;
466         int nblocks;
467         int block_idx;
468         int page_idx;
469         int i;
470         int rc = 0;
471         bool fault_inject;
472         bool integrity_enabled;
473         struct blk_plug plug;
474         ENTRY;
475
476         fault_inject = OBD_FAIL_CHECK(OBD_FAIL_OST_INTEGRITY_FAULT);
477         LASSERT(iobuf->dr_npages == npages);
478
479         integrity_enabled = bdev_integrity_enabled(bdev, iobuf->dr_rw);
480
481         osd_brw_stats_update(osd, iobuf);
482         iobuf->dr_start_time = ktime_get();
483
484         blk_start_plug(&plug);
485         for (page_idx = 0, block_idx = 0;
486              page_idx < npages;
487              page_idx++, block_idx += blocks_per_page) {
488
489                 page = pages[page_idx];
490                 LASSERT(block_idx + blocks_per_page <= total_blocks);
491
492                 for (i = 0, page_offset = 0;
493                      i < blocks_per_page;
494                      i += nblocks, page_offset += blocksize * nblocks) {
495
496                         nblocks = 1;
497
498                         if (blocks[block_idx + i] == 0) {  /* hole */
499                                 LASSERTF(iobuf->dr_rw == 0,
500                                          "page_idx %u, block_idx %u, i %u\n",
501                                          page_idx, block_idx, i);
502                                 memset(kmap(page) + page_offset, 0, blocksize);
503                                 kunmap(page);
504                                 continue;
505                         }
506
507                         sector = (sector_t)blocks[block_idx + i] << sector_bits;
508
509                         /* Additional contiguous file blocks? */
510                         while (i + nblocks < blocks_per_page &&
511                                (sector + (nblocks << sector_bits)) ==
512                                ((sector_t)blocks[block_idx + i + nblocks] <<
513                                 sector_bits))
514                                 nblocks++;
515
516                         if (bio != NULL &&
517                             can_be_merged(bio, sector) &&
518                             bio_add_page(bio, page,
519                                          blocksize * nblocks, page_offset) != 0)
520                                 continue;       /* added this frag OK */
521
522                         if (bio != NULL) {
523                                 struct request_queue *q = bio_get_queue(bio);
524                                 unsigned int bi_size = bio_sectors(bio) << 9;
525
526                                 /* Dang! I have to fragment this I/O */
527                                 CDEBUG(D_INODE,
528                                        "bio++ sz %d vcnt %d(%d) sectors %d(%d) psg %d(%d)\n",
529                                        bi_size, bio->bi_vcnt, bio->bi_max_vecs,
530                                        bio_sectors(bio),
531                                        queue_max_sectors(q),
532                                        osd_bio_nr_segs(bio),
533                                        queue_max_segments(q));
534                                 rc = osd_bio_integrity_handle(osd, bio,
535                                         iobuf, bio_start_page_idx,
536                                         fault_inject, integrity_enabled);
537                                 if (rc) {
538                                         bio_put(bio);
539                                         goto out;
540                                 }
541
542                                 record_start_io(iobuf, bi_size);
543                                 osd_submit_bio(iobuf->dr_rw, bio);
544                         }
545
546                         bio_start_page_idx = page_idx;
547                         /* allocate new bio */
548                         bio = bio_alloc(GFP_NOIO, min(BIO_MAX_PAGES,
549                                                       (npages - page_idx) *
550                                                       blocks_per_page));
551                         if (bio == NULL) {
552                                 CERROR("Can't allocate bio %u*%u = %u pages\n",
553                                        (npages - page_idx), blocks_per_page,
554                                        (npages - page_idx) * blocks_per_page);
555                                 rc = -ENOMEM;
556                                 goto out;
557                         }
558
559                         bio_set_dev(bio, bdev);
560                         bio_set_sector(bio, sector);
561                         bio->bi_opf = iobuf->dr_rw ? WRITE : READ;
562                         rc = osd_bio_init(bio, iobuf, integrity_enabled,
563                                           bio_start_page_idx, &bio_private);
564                         if (rc) {
565                                 bio_put(bio);
566                                 goto out;
567                         }
568
569                         rc = bio_add_page(bio, page,
570                                           blocksize * nblocks, page_offset);
571                         LASSERT(rc != 0);
572                 }
573         }
574
575         if (bio != NULL) {
576                 rc = osd_bio_integrity_handle(osd, bio, iobuf,
577                                               bio_start_page_idx,
578                                               fault_inject,
579                                               integrity_enabled);
580                 if (rc) {
581                         bio_put(bio);
582                         goto out;
583                 }
584
585                 record_start_io(iobuf, bio_sectors(bio) << 9);
586                 osd_submit_bio(iobuf->dr_rw, bio);
587                 rc = 0;
588         }
589
590 out:
591         blk_finish_plug(&plug);
592
593         /* in order to achieve better IO throughput, we don't wait for writes
594          * completion here. instead we proceed with transaction commit in
595          * parallel and wait for IO completion once transaction is stopped
596          * see osd_trans_stop() for more details -bzzz */
597         if (iobuf->dr_rw == 0 || fault_inject) {
598                 wait_event(iobuf->dr_wait,
599                            atomic_read(&iobuf->dr_numreqs) == 0);
600                 osd_fini_iobuf(osd, iobuf);
601         }
602
603         if (rc == 0) {
604                 rc = iobuf->dr_error;
605         } else {
606                 if (bio_private)
607                         OBD_FREE_PTR(bio_private);
608         }
609
610         RETURN(rc);
611 }
612
613 static int osd_map_remote_to_local(loff_t offset, ssize_t len, int *nrpages,
614                                    struct niobuf_local *lnb, int maxlnb)
615 {
616         int rc = 0;
617         ENTRY;
618
619         *nrpages = 0;
620
621         while (len > 0) {
622                 int poff = offset & (PAGE_SIZE - 1);
623                 int plen = PAGE_SIZE - poff;
624
625                 if (*nrpages >= maxlnb) {
626                         rc = -EOVERFLOW;
627                         break;
628                 }
629
630                 if (plen > len)
631                         plen = len;
632                 lnb->lnb_file_offset = offset;
633                 lnb->lnb_page_offset = poff;
634                 lnb->lnb_len = plen;
635                 /* lnb->lnb_flags = rnb->rnb_flags; */
636                 lnb->lnb_flags = 0;
637                 lnb->lnb_page = NULL;
638                 lnb->lnb_rc = 0;
639                 lnb->lnb_guard_rpc = 0;
640                 lnb->lnb_guard_disk = 0;
641                 lnb->lnb_locked = 0;
642
643                 LASSERTF(plen <= len, "plen %u, len %lld\n", plen,
644                          (long long) len);
645                 offset += plen;
646                 len -= plen;
647                 lnb++;
648                 (*nrpages)++;
649         }
650
651         RETURN(rc);
652 }
653
654 static struct page *osd_get_page(const struct lu_env *env, struct dt_object *dt,
655                                  loff_t offset, gfp_t gfp_mask, bool cache)
656 {
657         struct osd_thread_info *oti = osd_oti_get(env);
658         struct inode *inode = osd_dt_obj(dt)->oo_inode;
659         struct osd_device *d = osd_obj2dev(osd_dt_obj(dt));
660         struct page *page;
661         int cur;
662
663         LASSERT(inode);
664
665         if (cache) {
666                 page = find_or_create_page(inode->i_mapping,
667                                            offset >> PAGE_SHIFT, gfp_mask);
668
669                 if (likely(page))
670                         LASSERT(!test_bit(PG_private_2, &page->flags));
671                 else
672                         lprocfs_counter_add(d->od_stats, LPROC_OSD_NO_PAGE, 1);
673
674                 return page;
675         }
676
677         if (inode->i_mapping->nrpages) {
678                 /* consult with pagecache, but do not create new pages */
679                 /* this is normally used once */
680                 page = find_lock_page(inode->i_mapping, offset >> PAGE_SHIFT);
681                 if (page)
682                         return page;
683         }
684
685         LASSERT(oti->oti_dio_pages);
686         cur = oti->oti_dio_pages_used;
687
688         if (unlikely(!oti->oti_dio_pages[cur])) {
689                 LASSERT(cur < PTLRPC_MAX_BRW_PAGES);
690                 page = alloc_page(gfp_mask);
691                 if (!page)
692                         return NULL;
693                 oti->oti_dio_pages[cur] = page;
694         }
695
696         page = oti->oti_dio_pages[cur];
697         LASSERT(!test_bit(PG_private_2, &page->flags));
698         set_bit(PG_private_2, &page->flags);
699         oti->oti_dio_pages_used++;
700
701         LASSERT(!PageLocked(page));
702         lock_page(page);
703
704         LASSERT(!page->mapping);
705         LASSERT(!PageWriteback(page));
706         ClearPageUptodate(page);
707
708         page->index = offset >> PAGE_SHIFT;
709
710         return page;
711 }
712
713 /*
714  * there are following "locks":
715  * journal_start
716  * i_mutex
717  * page lock
718  *
719  * osd write path:
720  *  - lock page(s)
721  *  - journal_start
722  *  - truncate_sem
723  *
724  * ext4 vmtruncate:
725  *  - lock pages, unlock
726  *  - journal_start
727  *  - lock partial page
728  *  - i_data_sem
729  *
730  */
731
732 /**
733  * Unlock and release pages loaded by osd_bufs_get()
734  *
735  * Unlock \a npages pages from \a lnb and drop the refcount on them.
736  *
737  * \param env           thread execution environment
738  * \param dt            dt object undergoing IO (OSD object + methods)
739  * \param lnb           array of pages undergoing IO
740  * \param npages        number of pages in \a lnb
741  *
742  * \retval 0            always
743  */
744 static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt,
745                         struct niobuf_local *lnb, int npages)
746 {
747         struct osd_thread_info *oti = osd_oti_get(env);
748         struct pagevec pvec;
749         int i;
750
751         ll_pagevec_init(&pvec, 0);
752
753         for (i = 0; i < npages; i++) {
754                 struct page *page = lnb[i].lnb_page;
755
756                 if (page == NULL)
757                         continue;
758
759                 /* if the page isn't cached, then reset uptodate
760                  * to prevent reuse */
761                 if (test_bit(PG_private_2, &page->flags)) {
762                         clear_bit(PG_private_2, &page->flags);
763                         ClearPageUptodate(page);
764                         if (lnb[i].lnb_locked)
765                                 unlock_page(page);
766                         oti->oti_dio_pages_used--;
767                 } else {
768                         if (lnb[i].lnb_locked)
769                                 unlock_page(page);
770                         if (pagevec_add(&pvec, page) == 0)
771                                 pagevec_release(&pvec);
772                 }
773                 dt_object_put(env, dt);
774
775                 lnb[i].lnb_page = NULL;
776         }
777
778         LASSERTF(oti->oti_dio_pages_used == 0, "%d\n", oti->oti_dio_pages_used);
779
780         /* Release any partial pagevec */
781         pagevec_release(&pvec);
782
783         RETURN(0);
784 }
785
786 /**
787  * Load and lock pages undergoing IO
788  *
789  * Pages as described in the \a lnb array are fetched (from disk or cache)
790  * and locked for IO by the caller.
791  *
792  * DLM locking protects us from write and truncate competing for same region,
793  * but partial-page truncate can leave dirty pages in the cache for ldiskfs.
794  * It's possible the writeout on a such a page is in progress when we access
795  * it. It's also possible that during this writeout we put new (partial) data
796  * into the page, but won't be able to proceed in filter_commitrw_write().
797  * Therefore, just wait for writeout completion as it should be rare enough.
798  *
799  * \param env           thread execution environment
800  * \param dt            dt object undergoing IO (OSD object + methods)
801  * \param pos           byte offset of IO start
802  * \param len           number of bytes of IO
803  * \param lnb           array of extents undergoing IO
804  * \param rw            read or write operation, and other flags
805  * \param capa          capabilities
806  *
807  * \retval pages        (zero or more) loaded successfully
808  * \retval -ENOMEM      on memory/page allocation error
809  */
810 static int osd_bufs_get(const struct lu_env *env, struct dt_object *dt,
811                         loff_t pos, ssize_t len, struct niobuf_local *lnb,
812                         int maxlnb, enum dt_bufs_type rw)
813 {
814         struct osd_thread_info *oti = osd_oti_get(env);
815         struct osd_object *obj = osd_dt_obj(dt);
816         struct osd_device *osd   = osd_obj2dev(obj);
817         int npages, i, iosize, rc = 0;
818         bool cache, write;
819         loff_t fsize;
820         gfp_t gfp_mask;
821
822         LASSERT(obj->oo_inode);
823
824         rc = osd_map_remote_to_local(pos, len, &npages, lnb, maxlnb);
825         if (rc)
826                 RETURN(rc);
827
828         write = rw & DT_BUFS_TYPE_WRITE;
829
830         fsize = lnb[npages - 1].lnb_file_offset + lnb[npages - 1].lnb_len;
831         iosize = fsize - lnb[0].lnb_file_offset;
832         fsize = max(fsize, i_size_read(obj->oo_inode));
833
834         cache = rw & DT_BUFS_TYPE_READAHEAD;
835         if (cache)
836                 goto bypass_checks;
837
838         cache = osd_use_page_cache(osd);
839         while (cache) {
840                 if (write) {
841                         if (!osd->od_writethrough_cache) {
842                                 cache = false;
843                                 break;
844                         }
845                         if (iosize > osd->od_writethrough_max_iosize) {
846                                 cache = false;
847                                 break;
848                         }
849                 } else {
850                         if (!osd->od_read_cache) {
851                                 cache = false;
852                                 break;
853                         }
854                         if (iosize > osd->od_readcache_max_iosize) {
855                                 cache = false;
856                                 break;
857                         }
858                 }
859                 /* don't use cache on large files */
860                 if (osd->od_readcache_max_filesize &&
861                     fsize > osd->od_readcache_max_filesize)
862                         cache = false;
863                 break;
864         }
865
866 bypass_checks:
867         if (!cache && unlikely(!oti->oti_dio_pages)) {
868                 OBD_ALLOC(oti->oti_dio_pages,
869                           sizeof(struct page *) * PTLRPC_MAX_BRW_PAGES);
870                 if (!oti->oti_dio_pages)
871                         return -ENOMEM;
872         }
873
874         /* this could also try less hard for DT_BUFS_TYPE_READAHEAD pages */
875         gfp_mask = rw & DT_BUFS_TYPE_LOCAL ? (GFP_NOFS | __GFP_HIGHMEM) :
876                                              GFP_HIGHUSER;
877         for (i = 0; i < npages; i++, lnb++) {
878                 lnb->lnb_page = osd_get_page(env, dt, lnb->lnb_file_offset,
879                                              gfp_mask, cache);
880                 if (lnb->lnb_page == NULL)
881                         GOTO(cleanup, rc = -ENOMEM);
882
883                 lnb->lnb_locked = 1;
884                 wait_on_page_writeback(lnb->lnb_page);
885                 BUG_ON(PageWriteback(lnb->lnb_page));
886
887                 lu_object_get(&dt->do_lu);
888         }
889
890 #if 0
891         /* XXX: this version doesn't invalidate cached pages, but use them */
892         if (!cache && write && obj->oo_inode->i_mapping->nrpages) {
893                 /* do not allow data aliasing, invalidate pagecache */
894                 /* XXX: can be quite expensive in mixed case */
895                 invalidate_mapping_pages(obj->oo_inode->i_mapping,
896                                 lnb[0].lnb_file_offset >> PAGE_SHIFT,
897                                 lnb[npages - 1].lnb_file_offset >> PAGE_SHIFT);
898         }
899 #endif
900
901         RETURN(i);
902
903 cleanup:
904         if (i > 0)
905                 osd_bufs_put(env, dt, lnb - i, i);
906         return rc;
907 }
908
909 static int osd_ldiskfs_map_inode_pages(struct inode *inode, struct page **page,
910                                        int pages, sector_t *blocks,
911                                        int create)
912 {
913         int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
914         int rc = 0, i = 0;
915         struct page *fp = NULL;
916         int clen = 0;
917         pgoff_t max_page_index;
918         handle_t *handle = NULL;
919
920         max_page_index = inode->i_sb->s_maxbytes >> PAGE_SHIFT;
921
922         CDEBUG(D_OTHER, "inode %lu: map %d pages from %lu\n",
923                 inode->i_ino, pages, (*page)->index);
924
925         if (create) {
926                 create = LDISKFS_GET_BLOCKS_CREATE;
927                 handle = ldiskfs_journal_current_handle();
928                 LASSERT(handle != NULL);
929                 rc = osd_attach_jinode(inode);
930                 if (rc)
931                         return rc;
932         }
933         /* pages are sorted already. so, we just have to find
934          * contig. space and process them properly */
935         while (i < pages) {
936                 long blen, total = 0;
937                 struct ldiskfs_map_blocks map = { 0 };
938
939                 if (fp == NULL) { /* start new extent */
940                         fp = *page++;
941                         clen = 1;
942                         if (++i != pages)
943                                 continue;
944                 } else if (fp->index + clen == (*page)->index) {
945                         /* continue the extent */
946                         page++;
947                         clen++;
948                         if (++i != pages)
949                                 continue;
950                 }
951                 if (fp->index + clen >= max_page_index)
952                         GOTO(cleanup, rc = -EFBIG);
953                 /* process found extent */
954                 map.m_lblk = fp->index * blocks_per_page;
955                 map.m_len = blen = clen * blocks_per_page;
956 cont_map:
957                 rc = ldiskfs_map_blocks(handle, inode, &map, create);
958                 if (rc >= 0) {
959                         int c = 0;
960                         for (; total < blen && c < map.m_len; c++, total++) {
961                                 if (rc == 0) {
962                                         *(blocks + total) = 0;
963                                         total++;
964                                         break;
965                                 } else {
966                                         *(blocks + total) = map.m_pblk + c;
967                                         /* unmap any possible underlying
968                                          * metadata from the block device
969                                          * mapping.  bug 6998. */
970                                         if ((map.m_flags & LDISKFS_MAP_NEW) &&
971                                             create)
972                                                 clean_bdev_aliases(
973                                                         inode->i_sb->s_bdev,
974                                                         map.m_pblk + c, 1);
975                                 }
976                         }
977                         rc = 0;
978                 }
979                 if (rc == 0 && total < blen) {
980                         map.m_lblk = fp->index * blocks_per_page + total;
981                         map.m_len = blen - total;
982                         goto cont_map;
983                 }
984                 if (rc != 0)
985                         GOTO(cleanup, rc);
986
987                 /* look for next extent */
988                 fp = NULL;
989                 blocks += blocks_per_page * clen;
990         }
991 cleanup:
992         return rc;
993 }
994
995 static int osd_write_prep(const struct lu_env *env, struct dt_object *dt,
996                           struct niobuf_local *lnb, int npages)
997 {
998         struct osd_thread_info *oti   = osd_oti_get(env);
999         struct osd_iobuf       *iobuf = &oti->oti_iobuf;
1000         struct inode           *inode = osd_dt_obj(dt)->oo_inode;
1001         struct osd_device      *osd   = osd_obj2dev(osd_dt_obj(dt));
1002         ktime_t start, end;
1003         s64 timediff;
1004         ssize_t isize;
1005         __s64  maxidx;
1006         int i, rc = 0;
1007
1008         LASSERT(inode);
1009
1010         rc = osd_init_iobuf(osd, iobuf, 0, npages);
1011         if (unlikely(rc != 0))
1012                 RETURN(rc);
1013
1014         isize = i_size_read(inode);
1015         maxidx = ((isize + PAGE_SIZE - 1) >> PAGE_SHIFT) - 1;
1016
1017         start = ktime_get();
1018         for (i = 0; i < npages; i++) {
1019
1020                 /*
1021                  * till commit the content of the page is undefined
1022                  * we'll set it uptodate once bulk is done. otherwise
1023                  * subsequent reads can access non-stable data
1024                  */
1025                 ClearPageUptodate(lnb[i].lnb_page);
1026
1027                 if (lnb[i].lnb_len == PAGE_SIZE)
1028                         continue;
1029
1030                 if (maxidx >= lnb[i].lnb_page->index) {
1031                         osd_iobuf_add_page(iobuf, &lnb[i]);
1032                 } else {
1033                         long off;
1034                         char *p = kmap(lnb[i].lnb_page);
1035
1036                         off = lnb[i].lnb_page_offset;
1037                         if (off)
1038                                 memset(p, 0, off);
1039                         off = (lnb[i].lnb_page_offset + lnb[i].lnb_len) &
1040                               ~PAGE_MASK;
1041                         if (off)
1042                                 memset(p + off, 0, PAGE_SIZE - off);
1043                         kunmap(lnb[i].lnb_page);
1044                 }
1045         }
1046         end = ktime_get();
1047         timediff = ktime_us_delta(end, start);
1048         lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
1049
1050         if (iobuf->dr_npages) {
1051                 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1052                                                  iobuf->dr_npages,
1053                                                  iobuf->dr_blocks, 0);
1054                 if (likely(rc == 0)) {
1055                         rc = osd_do_bio(osd, inode, iobuf);
1056                         /* do IO stats for preparation reads */
1057                         osd_fini_iobuf(osd, iobuf);
1058                 }
1059         }
1060         RETURN(rc);
1061 }
1062
1063 struct osd_fextent {
1064         sector_t        start;
1065         sector_t        end;
1066         unsigned int    mapped:1;
1067 };
1068
1069 static int osd_is_mapped(struct dt_object *dt, __u64 offset,
1070                          struct osd_fextent *cached_extent)
1071 {
1072         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1073         sector_t block = offset >> inode->i_blkbits;
1074         sector_t start;
1075         struct fiemap_extent_info fei = { 0 };
1076         struct fiemap_extent fe = { 0 };
1077         mm_segment_t saved_fs;
1078         int rc;
1079
1080         if (block >= cached_extent->start && block < cached_extent->end)
1081                 return cached_extent->mapped;
1082
1083         if (i_size_read(inode) == 0)
1084                 return 0;
1085
1086         /* Beyond EOF, must not be mapped */
1087         if (((i_size_read(inode) - 1) >> inode->i_blkbits) < block)
1088                 return 0;
1089
1090         fei.fi_extents_max = 1;
1091         fei.fi_extents_start = &fe;
1092
1093         saved_fs = get_fs();
1094         set_fs(KERNEL_DS);
1095         rc = inode->i_op->fiemap(inode, &fei, offset, FIEMAP_MAX_OFFSET-offset);
1096         set_fs(saved_fs);
1097         if (rc != 0)
1098                 return 0;
1099
1100         start = fe.fe_logical >> inode->i_blkbits;
1101
1102         if (start > block) {
1103                 cached_extent->start = block;
1104                 cached_extent->end = start;
1105                 cached_extent->mapped = 0;
1106         } else {
1107                 cached_extent->start = start;
1108                 cached_extent->end = (fe.fe_logical + fe.fe_length) >>
1109                                       inode->i_blkbits;
1110                 cached_extent->mapped = 1;
1111         }
1112
1113         return cached_extent->mapped;
1114 }
1115
1116 static int osd_declare_write_commit(const struct lu_env *env,
1117                                     struct dt_object *dt,
1118                                     struct niobuf_local *lnb, int npages,
1119                                     struct thandle *handle)
1120 {
1121         const struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1122         struct inode            *inode = osd_dt_obj(dt)->oo_inode;
1123         struct osd_thandle      *oh;
1124         int                     extents = 1;
1125         int                     depth;
1126         int                     i;
1127         int                     newblocks;
1128         int                     rc = 0;
1129         int                     credits = 0;
1130         long long               quota_space = 0;
1131         struct osd_fextent      extent = { 0 };
1132         enum osd_quota_local_flags local_flags = 0;
1133         enum osd_qid_declare_flags declare_flags = OSD_QID_BLK;
1134         ENTRY;
1135
1136         LASSERT(handle != NULL);
1137         oh = container_of0(handle, struct osd_thandle, ot_super);
1138         LASSERT(oh->ot_handle == NULL);
1139
1140         newblocks = npages;
1141
1142         /* calculate number of extents (probably better to pass nb) */
1143         for (i = 0; i < npages; i++) {
1144                 if (i && lnb[i].lnb_file_offset !=
1145                     lnb[i - 1].lnb_file_offset + lnb[i - 1].lnb_len)
1146                         extents++;
1147
1148                 if (osd_is_mapped(dt, lnb[i].lnb_file_offset, &extent))
1149                         lnb[i].lnb_flags |= OBD_BRW_MAPPED;
1150                 else
1151                         quota_space += PAGE_SIZE;
1152
1153                 /* ignore quota for the whole request if any page is from
1154                  * client cache or written by root.
1155                  *
1156                  * XXX once we drop the 1.8 client support, the checking
1157                  * for whether page is from cache can be simplified as:
1158                  * !(lnb[i].flags & OBD_BRW_SYNC)
1159                  *
1160                  * XXX we could handle this on per-lnb basis as done by
1161                  * grant. */
1162                 if ((lnb[i].lnb_flags & OBD_BRW_NOQUOTA) ||
1163                     (lnb[i].lnb_flags & (OBD_BRW_FROM_GRANT | OBD_BRW_SYNC)) ==
1164                     OBD_BRW_FROM_GRANT)
1165                         declare_flags |= OSD_QID_FORCE;
1166         }
1167
1168         /*
1169          * each extent can go into new leaf causing a split
1170          * 5 is max tree depth: inode + 4 index blocks
1171          * with blockmaps, depth is 3 at most
1172          */
1173         if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL) {
1174                 /*
1175                  * many concurrent threads may grow tree by the time
1176                  * our transaction starts. so, consider 2 is a min depth
1177                  */
1178                 depth = ext_depth(inode);
1179                 depth = max(depth, 1) + 1;
1180                 newblocks += depth;
1181                 credits++; /* inode */
1182                 credits += depth * 2 * extents;
1183         } else {
1184                 depth = 3;
1185                 newblocks += depth;
1186                 credits++; /* inode */
1187                 credits += depth * extents;
1188         }
1189
1190         /* quota space for metadata blocks */
1191         quota_space += depth * extents * LDISKFS_BLOCK_SIZE(osd_sb(osd));
1192
1193         /* quota space should be reported in 1K blocks */
1194         quota_space = toqb(quota_space);
1195
1196         /* each new block can go in different group (bitmap + gd) */
1197
1198         /* we can't dirty more bitmap blocks than exist */
1199         if (newblocks > LDISKFS_SB(osd_sb(osd))->s_groups_count)
1200                 credits += LDISKFS_SB(osd_sb(osd))->s_groups_count;
1201         else
1202                 credits += newblocks;
1203
1204         /* we can't dirty more gd blocks than exist */
1205         if (newblocks > LDISKFS_SB(osd_sb(osd))->s_gdb_count)
1206                 credits += LDISKFS_SB(osd_sb(osd))->s_gdb_count;
1207         else
1208                 credits += newblocks;
1209
1210         osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
1211
1212         /* make sure the over quota flags were not set */
1213         lnb[0].lnb_flags &= ~OBD_BRW_OVER_ALLQUOTA;
1214
1215         rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
1216                                    i_projid_read(inode), quota_space, oh,
1217                                    osd_dt_obj(dt), &local_flags, declare_flags);
1218
1219         /* we need only to store the overquota flags in the first lnb for
1220          * now, once we support multiple objects BRW, this code needs be
1221          * revised. */
1222         if (local_flags & QUOTA_FL_OVER_USRQUOTA)
1223                 lnb[0].lnb_flags |= OBD_BRW_OVER_USRQUOTA;
1224         if (local_flags & QUOTA_FL_OVER_GRPQUOTA)
1225                 lnb[0].lnb_flags |= OBD_BRW_OVER_GRPQUOTA;
1226         if (local_flags & QUOTA_FL_OVER_PRJQUOTA)
1227                 lnb[0].lnb_flags |= OBD_BRW_OVER_PRJQUOTA;
1228
1229         if (rc == 0)
1230                 rc = osd_trunc_lock(osd_dt_obj(dt), oh, true);
1231
1232         RETURN(rc);
1233 }
1234
1235 /* Check if a block is allocated or not */
1236 static int osd_write_commit(const struct lu_env *env, struct dt_object *dt,
1237                             struct niobuf_local *lnb, int npages,
1238                             struct thandle *thandle)
1239 {
1240         struct osd_thread_info *oti = osd_oti_get(env);
1241         struct osd_iobuf *iobuf = &oti->oti_iobuf;
1242         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1243         struct osd_device  *osd = osd_obj2dev(osd_dt_obj(dt));
1244         loff_t isize;
1245         int rc = 0, i;
1246
1247         LASSERT(inode);
1248
1249         rc = osd_init_iobuf(osd, iobuf, 1, npages);
1250         if (unlikely(rc != 0))
1251                 RETURN(rc);
1252
1253         isize = i_size_read(inode);
1254         dquot_initialize(inode);
1255
1256         for (i = 0; i < npages; i++) {
1257                 if (lnb[i].lnb_rc == -ENOSPC &&
1258                     (lnb[i].lnb_flags & OBD_BRW_MAPPED)) {
1259                         /* Allow the write to proceed if overwriting an
1260                          * existing block */
1261                         lnb[i].lnb_rc = 0;
1262                 }
1263
1264                 if (lnb[i].lnb_rc) { /* ENOSPC, network RPC error, etc. */
1265                         CDEBUG(D_INODE, "Skipping [%d] == %d\n", i,
1266                                lnb[i].lnb_rc);
1267                         LASSERT(lnb[i].lnb_page);
1268                         generic_error_remove_page(inode->i_mapping,
1269                                                   lnb[i].lnb_page);
1270                         continue;
1271                 }
1272
1273                 LASSERT(PageLocked(lnb[i].lnb_page));
1274                 LASSERT(!PageWriteback(lnb[i].lnb_page));
1275
1276                 if (lnb[i].lnb_file_offset + lnb[i].lnb_len > isize)
1277                         isize = lnb[i].lnb_file_offset + lnb[i].lnb_len;
1278
1279                 /*
1280                  * Since write and truncate are serialized by oo_sem, even
1281                  * partial-page truncate should not leave dirty pages in the
1282                  * page cache.
1283                  */
1284                 LASSERT(!PageDirty(lnb[i].lnb_page));
1285
1286                 SetPageUptodate(lnb[i].lnb_page);
1287
1288                 osd_iobuf_add_page(iobuf, &lnb[i]);
1289         }
1290
1291         osd_trans_exec_op(env, thandle, OSD_OT_WRITE);
1292
1293         if (OBD_FAIL_CHECK(OBD_FAIL_OST_MAPBLK_ENOSPC)) {
1294                 rc = -ENOSPC;
1295         } else if (iobuf->dr_npages > 0) {
1296                 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1297                                                  iobuf->dr_npages,
1298                                                  iobuf->dr_blocks, 1);
1299         } else {
1300                 /* no pages to write, no transno is needed */
1301                 thandle->th_local = 1;
1302         }
1303
1304         if (likely(rc == 0)) {
1305                 spin_lock(&inode->i_lock);
1306                 if (isize > i_size_read(inode)) {
1307                         i_size_write(inode, isize);
1308                         LDISKFS_I(inode)->i_disksize = isize;
1309                         spin_unlock(&inode->i_lock);
1310                         ll_dirty_inode(inode, I_DIRTY_DATASYNC);
1311                 } else {
1312                         spin_unlock(&inode->i_lock);
1313                 }
1314
1315                 rc = osd_do_bio(osd, inode, iobuf);
1316                 /* we don't do stats here as in read path because
1317                  * write is async: we'll do this in osd_put_bufs() */
1318         } else {
1319                 osd_fini_iobuf(osd, iobuf);
1320         }
1321
1322         osd_trans_exec_check(env, thandle, OSD_OT_WRITE);
1323
1324         if (unlikely(rc != 0)) {
1325                 /* if write fails, we should drop pages from the cache */
1326                 for (i = 0; i < npages; i++) {
1327                         if (lnb[i].lnb_page == NULL)
1328                                 continue;
1329                         LASSERT(PageLocked(lnb[i].lnb_page));
1330                         generic_error_remove_page(inode->i_mapping,
1331                                                   lnb[i].lnb_page);
1332                 }
1333         }
1334
1335         RETURN(rc);
1336 }
1337
1338 static int osd_read_prep(const struct lu_env *env, struct dt_object *dt,
1339                          struct niobuf_local *lnb, int npages)
1340 {
1341         struct osd_thread_info *oti = osd_oti_get(env);
1342         struct osd_iobuf *iobuf = &oti->oti_iobuf;
1343         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1344         struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1345         int rc = 0, i, cache_hits = 0, cache_misses = 0;
1346         ktime_t start, end;
1347         s64 timediff;
1348         loff_t isize;
1349
1350         LASSERT(inode);
1351
1352         rc = osd_init_iobuf(osd, iobuf, 0, npages);
1353         if (unlikely(rc != 0))
1354                 RETURN(rc);
1355
1356         isize = i_size_read(inode);
1357
1358         start = ktime_get();
1359         for (i = 0; i < npages; i++) {
1360
1361                 if (isize <= lnb[i].lnb_file_offset)
1362                         /* If there's no more data, abort early.
1363                          * lnb->lnb_rc == 0, so it's easy to detect later. */
1364                         break;
1365
1366                 /* instead of looking if we go beyong isize, send complete
1367                  * pages all the time
1368                  */
1369                 lnb[i].lnb_rc = lnb[i].lnb_len;
1370
1371                 /* Bypass disk read if fail_loc is set properly */
1372                 if (OBD_FAIL_CHECK(OBD_FAIL_OST_FAKE_RW))
1373                         SetPageUptodate(lnb[i].lnb_page);
1374
1375                 if (PageUptodate(lnb[i].lnb_page)) {
1376                         cache_hits++;
1377                         unlock_page(lnb[i].lnb_page);
1378                 } else {
1379                         cache_misses++;
1380                         osd_iobuf_add_page(iobuf, &lnb[i]);
1381                 }
1382                 /* no need to unlock in osd_bufs_put(), the sooner page is
1383                  * unlocked, the earlier another client can access it.
1384                  * notice real unlock_page() can be called few lines
1385                  * below after osd_do_bio(). lnb is a per-thread, so it's
1386                  * fine to have PG_locked and lnb_locked inconsistent here */
1387                 lnb[i].lnb_locked = 0;
1388         }
1389         end = ktime_get();
1390         timediff = ktime_us_delta(end, start);
1391         lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
1392
1393         if (cache_hits != 0)
1394                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_HIT,
1395                                     cache_hits);
1396         if (cache_misses != 0)
1397                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_MISS,
1398                                     cache_misses);
1399         if (cache_hits + cache_misses != 0)
1400                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_ACCESS,
1401                                     cache_hits + cache_misses);
1402
1403         if (iobuf->dr_npages) {
1404                 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1405                                                  iobuf->dr_npages,
1406                                                  iobuf->dr_blocks, 0);
1407                 rc = osd_do_bio(osd, inode, iobuf);
1408
1409                 /* IO stats will be done in osd_bufs_put() */
1410
1411                 /* early release to let others read data during the bulk */
1412                 for (i = 0; i < iobuf->dr_npages; i++) {
1413                         LASSERT(PageLocked(iobuf->dr_pages[i]));
1414                         unlock_page(iobuf->dr_pages[i]);
1415                 }
1416         }
1417
1418         RETURN(rc);
1419 }
1420
1421 /*
1422  * XXX: Another layering violation for now.
1423  *
1424  * We don't want to use ->f_op->read methods, because generic file write
1425  *
1426  *         - serializes on ->i_sem, and
1427  *
1428  *         - does a lot of extra work like balance_dirty_pages(),
1429  *
1430  * which doesn't work for globally shared files like /last_rcvd.
1431  */
1432 static int osd_ldiskfs_readlink(struct inode *inode, char *buffer, int buflen)
1433 {
1434         struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
1435
1436         memcpy(buffer, (char *)ei->i_data, buflen);
1437
1438         return  buflen;
1439 }
1440
1441 int osd_ldiskfs_read(struct inode *inode, void *buf, int size, loff_t *offs)
1442 {
1443         struct buffer_head *bh;
1444         unsigned long block;
1445         int osize;
1446         int blocksize;
1447         int csize;
1448         int boffs;
1449
1450         /* prevent reading after eof */
1451         spin_lock(&inode->i_lock);
1452         if (i_size_read(inode) < *offs + size) {
1453                 loff_t diff = i_size_read(inode) - *offs;
1454                 spin_unlock(&inode->i_lock);
1455                 if (diff < 0) {
1456                         CDEBUG(D_OTHER,
1457                                "size %llu is too short to read @%llu\n",
1458                                i_size_read(inode), *offs);
1459                         return -EBADR;
1460                 } else if (diff == 0) {
1461                         return 0;
1462                 } else {
1463                         size = diff;
1464                 }
1465         } else {
1466                 spin_unlock(&inode->i_lock);
1467         }
1468
1469         blocksize = 1 << inode->i_blkbits;
1470         osize = size;
1471         while (size > 0) {
1472                 block = *offs >> inode->i_blkbits;
1473                 boffs = *offs & (blocksize - 1);
1474                 csize = min(blocksize - boffs, size);
1475                 bh = __ldiskfs_bread(NULL, inode, block, 0);
1476                 if (IS_ERR(bh)) {
1477                         CERROR("%s: can't read %u@%llu on ino %lu: "
1478                                "rc = %ld\n", osd_ino2name(inode),
1479                                csize, *offs, inode->i_ino,
1480                                PTR_ERR(bh));
1481                         return PTR_ERR(bh);
1482                 }
1483
1484                 if (bh != NULL) {
1485                         memcpy(buf, bh->b_data + boffs, csize);
1486                         brelse(bh);
1487                 } else {
1488                         memset(buf, 0, csize);
1489                 }
1490
1491                 *offs += csize;
1492                 buf += csize;
1493                 size -= csize;
1494         }
1495         return osize;
1496 }
1497
1498 static ssize_t osd_read(const struct lu_env *env, struct dt_object *dt,
1499                         struct lu_buf *buf, loff_t *pos)
1500 {
1501         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1502         int           rc;
1503
1504         /* Read small symlink from inode body as we need to maintain correct
1505          * on-disk symlinks for ldiskfs.
1506          */
1507         if (S_ISLNK(dt->do_lu.lo_header->loh_attr) &&
1508             (buf->lb_len < sizeof(LDISKFS_I(inode)->i_data)))
1509                 rc = osd_ldiskfs_readlink(inode, buf->lb_buf, buf->lb_len);
1510         else
1511                 rc = osd_ldiskfs_read(inode, buf->lb_buf, buf->lb_len, pos);
1512
1513         return rc;
1514 }
1515
1516 static inline int osd_extents_enabled(struct super_block *sb,
1517                                       struct inode *inode)
1518 {
1519         if (inode != NULL) {
1520                 if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL)
1521                         return 1;
1522         } else if (ldiskfs_has_feature_extents(sb)) {
1523                 return 1;
1524         }
1525         return 0;
1526 }
1527
1528 int osd_calc_bkmap_credits(struct super_block *sb, struct inode *inode,
1529                            const loff_t size, const loff_t pos,
1530                            const int blocks)
1531 {
1532         int credits, bits, bs, i;
1533
1534         bits = sb->s_blocksize_bits;
1535         bs = 1 << bits;
1536
1537         /* legacy blockmap: 3 levels * 3 (bitmap,gd,itself)
1538          * we do not expect blockmaps on the large files,
1539          * so let's shrink it to 2 levels (4GB files) */
1540
1541         /* this is default reservation: 2 levels */
1542         credits = (blocks + 2) * 3;
1543
1544         /* actual offset is unknown, hard to optimize */
1545         if (pos == -1)
1546                 return credits;
1547
1548         /* now check for few specific cases to optimize */
1549         if (pos + size <= LDISKFS_NDIR_BLOCKS * bs) {
1550                 /* no indirects */
1551                 credits = blocks;
1552                 /* allocate if not allocated */
1553                 if (inode == NULL) {
1554                         credits += blocks * 2;
1555                         return credits;
1556                 }
1557                 for (i = (pos >> bits); i < (pos >> bits) + blocks; i++) {
1558                         LASSERT(i < LDISKFS_NDIR_BLOCKS);
1559                         if (LDISKFS_I(inode)->i_data[i] == 0)
1560                                 credits += 2;
1561                 }
1562         } else if (pos + size <= (LDISKFS_NDIR_BLOCKS + 1024) * bs) {
1563                 /* single indirect */
1564                 credits = blocks * 3;
1565                 if (inode == NULL ||
1566                     LDISKFS_I(inode)->i_data[LDISKFS_IND_BLOCK] == 0)
1567                         credits += 3;
1568                 else
1569                         /* The indirect block may be modified. */
1570                         credits += 1;
1571         }
1572
1573         return credits;
1574 }
1575
1576 static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt,
1577                                  const struct lu_buf *buf, loff_t _pos,
1578                                  struct thandle *handle)
1579 {
1580         struct osd_object  *obj  = osd_dt_obj(dt);
1581         struct inode       *inode = obj->oo_inode;
1582         struct super_block *sb = osd_sb(osd_obj2dev(obj));
1583         struct osd_thandle *oh;
1584         int                 rc = 0, est = 0, credits, blocks, allocated = 0;
1585         int                 bits, bs;
1586         int                 depth, size;
1587         loff_t              pos;
1588         ENTRY;
1589
1590         LASSERT(buf != NULL);
1591         LASSERT(handle != NULL);
1592
1593         oh = container_of0(handle, struct osd_thandle, ot_super);
1594         LASSERT(oh->ot_handle == NULL);
1595
1596         size = buf->lb_len;
1597         bits = sb->s_blocksize_bits;
1598         bs = 1 << bits;
1599
1600         if (_pos == -1) {
1601                 /* if this is an append, then we
1602                  * should expect cross-block record */
1603                 pos = 0;
1604         } else {
1605                 pos = _pos;
1606         }
1607
1608         /* blocks to modify */
1609         blocks = ((pos + size + bs - 1) >> bits) - (pos >> bits);
1610         LASSERT(blocks > 0);
1611
1612         if (inode != NULL && _pos != -1) {
1613                 /* object size in blocks */
1614                 est = (i_size_read(inode) + bs - 1) >> bits;
1615                 allocated = inode->i_blocks >> (bits - 9);
1616                 if (pos + size <= i_size_read(inode) && est <= allocated) {
1617                         /* looks like an overwrite, no need to modify tree */
1618                         credits = blocks;
1619                         /* no need to modify i_size */
1620                         goto out;
1621                 }
1622         }
1623
1624         if (osd_extents_enabled(sb, inode)) {
1625                 /*
1626                  * many concurrent threads may grow tree by the time
1627                  * our transaction starts. so, consider 2 is a min depth
1628                  * for every level we may need to allocate a new block
1629                  * and take some entries from the old one. so, 3 blocks
1630                  * to allocate (bitmap, gd, itself) + old block - 4 per
1631                  * level.
1632                  */
1633                 depth = inode != NULL ? ext_depth(inode) : 0;
1634                 depth = max(depth, 1) + 1;
1635                 credits = depth;
1636                 /* if not append, then split may need to modify
1637                  * existing blocks moving entries into the new ones */
1638                 if (_pos != -1)
1639                         credits += depth;
1640                 /* blocks to store data: bitmap,gd,itself */
1641                 credits += blocks * 3;
1642         } else {
1643                 credits = osd_calc_bkmap_credits(sb, inode, size, _pos, blocks);
1644         }
1645         /* if inode is created as part of the transaction,
1646          * then it's counted already by the creation method */
1647         if (inode != NULL)
1648                 credits++;
1649
1650 out:
1651
1652         osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
1653
1654         /* dt_declare_write() is usually called for system objects, such
1655          * as llog or last_rcvd files. We needn't enforce quota on those
1656          * objects, so always set the lqi_space as 0. */
1657         if (inode != NULL)
1658                 rc = osd_declare_inode_qid(env, i_uid_read(inode),
1659                                            i_gid_read(inode),
1660                                            i_projid_read(inode), 0,
1661                                            oh, obj, NULL, OSD_QID_BLK);
1662
1663         if (rc == 0)
1664                 rc = osd_trunc_lock(obj, oh, true);
1665
1666         RETURN(rc);
1667 }
1668
1669 static int osd_ldiskfs_writelink(struct inode *inode, char *buffer, int buflen)
1670 {
1671         /* LU-2634: clear the extent format for fast symlink */
1672         ldiskfs_clear_inode_flag(inode, LDISKFS_INODE_EXTENTS);
1673
1674         memcpy((char *)&LDISKFS_I(inode)->i_data, (char *)buffer, buflen);
1675         spin_lock(&inode->i_lock);
1676         LDISKFS_I(inode)->i_disksize = buflen;
1677         i_size_write(inode, buflen);
1678         spin_unlock(&inode->i_lock);
1679         ll_dirty_inode(inode, I_DIRTY_DATASYNC);
1680
1681         return 0;
1682 }
1683
1684 static int osd_ldiskfs_write_record(struct dt_object *dt, void *buf,
1685                                     int bufsize, int write_NUL, loff_t *offs,
1686                                     handle_t *handle)
1687 {
1688         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1689         struct buffer_head *bh        = NULL;
1690         loff_t              offset    = *offs;
1691         loff_t              new_size  = i_size_read(inode);
1692         unsigned long       block;
1693         int                 blocksize = 1 << inode->i_blkbits;
1694         int                 err = 0;
1695         int                 size;
1696         int                 boffs;
1697         int                 dirty_inode = 0;
1698         struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
1699         bool create, sparse;
1700
1701         if (write_NUL) {
1702                 /*
1703                  * long symlink write does not count the NUL terminator in
1704                  * bufsize, we write it, and the inode's file size does not
1705                  * count the NUL terminator as well.
1706                  */
1707                 ((char *)buf)[bufsize] = '\0';
1708                 ++bufsize;
1709         }
1710
1711         /* sparse checking is racy, but sparse is very rare case, leave as is */
1712         sparse = (new_size > 0 && (inode->i_blocks >> (inode->i_blkbits - 9)) <
1713                   ((new_size - 1) >> inode->i_blkbits) + 1);
1714
1715         while (bufsize > 0) {
1716                 int credits = handle->h_buffer_credits;
1717                 bool sync;
1718                 unsigned long last_block = (new_size == 0) ? 0 :
1719                                            (new_size - 1) >> inode->i_blkbits;
1720
1721                 if (bh)
1722                         brelse(bh);
1723
1724                 block = offset >> inode->i_blkbits;
1725                 boffs = offset & (blocksize - 1);
1726                 size = min(blocksize - boffs, bufsize);
1727                 sync = (block > last_block || new_size == 0 || sparse);
1728
1729                 if (sync)
1730                         down(&ei->i_append_sem);
1731
1732                 bh = __ldiskfs_bread(handle, inode, block, 0);
1733
1734                 if (unlikely(IS_ERR_OR_NULL(bh) && !sync))
1735                         CWARN("%s: adding bh without locking off %llu (block %lu, "
1736                               "size %d, offs %llu)\n", inode->i_sb->s_id,
1737                               offset, block, bufsize, *offs);
1738
1739                 if (IS_ERR_OR_NULL(bh)) {
1740                         struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1741                         int flags = LDISKFS_GET_BLOCKS_CREATE;
1742
1743                         /* while the file system is being mounted, avoid
1744                          * preallocation otherwise mount can take a long
1745                          * time as mballoc cache is cold.
1746                          * XXX: this is a workaround until we have a proper
1747                          *      fix in mballoc
1748                          * XXX: works with extent-based files only */
1749                         if (!osd->od_cl_seq)
1750                                 flags |= LDISKFS_GET_BLOCKS_NO_NORMALIZE;
1751                         bh = __ldiskfs_bread(handle, inode, block, flags);
1752                         create = true;
1753                 } else {
1754                         if (sync)
1755                                 up(&ei->i_append_sem);
1756                         create = false;
1757                 }
1758                 if (IS_ERR_OR_NULL(bh)) {
1759                         if (bh == NULL) {
1760                                 err = -EIO;
1761                         } else {
1762                                 err = PTR_ERR(bh);
1763                                 bh = NULL;
1764                         }
1765
1766                         CERROR("%s: error reading offset %llu (block %lu, "
1767                                "size %d, offs %llu), credits %d/%d: rc = %d\n",
1768                                inode->i_sb->s_id, offset, block, bufsize, *offs,
1769                                credits, handle->h_buffer_credits, err);
1770                         break;
1771                 }
1772
1773                 err = ldiskfs_journal_get_write_access(handle, bh);
1774                 if (err) {
1775                         CERROR("journal_get_write_access() returned error %d\n",
1776                                err);
1777                         break;
1778                 }
1779                 LASSERTF(boffs + size <= bh->b_size,
1780                          "boffs %d size %d bh->b_size %lu\n",
1781                          boffs, size, (unsigned long)bh->b_size);
1782                 if (create) {
1783                         memset(bh->b_data, 0, bh->b_size);
1784                         if (sync)
1785                                 up(&ei->i_append_sem);
1786                 }
1787                 memcpy(bh->b_data + boffs, buf, size);
1788                 err = ldiskfs_handle_dirty_metadata(handle, NULL, bh);
1789                 if (err)
1790                         break;
1791
1792                 if (offset + size > new_size)
1793                         new_size = offset + size;
1794                 offset += size;
1795                 bufsize -= size;
1796                 buf += size;
1797         }
1798         if (bh)
1799                 brelse(bh);
1800
1801         if (write_NUL)
1802                 --new_size;
1803         /* correct in-core and on-disk sizes */
1804         if (new_size > i_size_read(inode)) {
1805                 spin_lock(&inode->i_lock);
1806                 if (new_size > i_size_read(inode))
1807                         i_size_write(inode, new_size);
1808                 if (i_size_read(inode) > LDISKFS_I(inode)->i_disksize) {
1809                         LDISKFS_I(inode)->i_disksize = i_size_read(inode);
1810                         dirty_inode = 1;
1811                 }
1812                 spin_unlock(&inode->i_lock);
1813                 if (dirty_inode)
1814                         ll_dirty_inode(inode, I_DIRTY_DATASYNC);
1815         }
1816
1817         if (err == 0)
1818                 *offs = offset;
1819         return err;
1820 }
1821
1822 static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt,
1823                          const struct lu_buf *buf, loff_t *pos,
1824                          struct thandle *handle)
1825 {
1826         struct inode            *inode = osd_dt_obj(dt)->oo_inode;
1827         struct osd_thandle      *oh;
1828         ssize_t                 result;
1829         int                     is_link;
1830
1831         LASSERT(dt_object_exists(dt));
1832
1833         LASSERT(handle != NULL);
1834         LASSERT(inode != NULL);
1835         dquot_initialize(inode);
1836
1837         /* XXX: don't check: one declared chunk can be used many times */
1838         /* osd_trans_exec_op(env, handle, OSD_OT_WRITE); */
1839
1840         oh = container_of(handle, struct osd_thandle, ot_super);
1841         LASSERT(oh->ot_handle->h_transaction != NULL);
1842         osd_trans_exec_op(env, handle, OSD_OT_WRITE);
1843
1844         /* Write small symlink to inode body as we need to maintain correct
1845          * on-disk symlinks for ldiskfs.
1846          * Note: the buf->lb_buf contains a NUL terminator while buf->lb_len
1847          * does not count it in.
1848          */
1849         is_link = S_ISLNK(dt->do_lu.lo_header->loh_attr);
1850         if (is_link && (buf->lb_len < sizeof(LDISKFS_I(inode)->i_data)))
1851                 result = osd_ldiskfs_writelink(inode, buf->lb_buf, buf->lb_len);
1852         else
1853                 result = osd_ldiskfs_write_record(dt, buf->lb_buf, buf->lb_len,
1854                                                   is_link, pos, oh->ot_handle);
1855         if (result == 0)
1856                 result = buf->lb_len;
1857
1858         osd_trans_exec_check(env, handle, OSD_OT_WRITE);
1859
1860         return result;
1861 }
1862
1863 static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt,
1864                              __u64 start, __u64 end, struct thandle *th)
1865 {
1866         struct osd_thandle *oh;
1867         struct inode       *inode;
1868         int                 rc;
1869         ENTRY;
1870
1871         LASSERT(th);
1872         oh = container_of(th, struct osd_thandle, ot_super);
1873
1874         /*
1875          * we don't need to reserve credits for whole truncate
1876          * it's not possible as truncate may need to free too many
1877          * blocks and that won't fit a single transaction. instead
1878          * we reserve credits to change i_size and put inode onto
1879          * orphan list. if needed truncate will extend or restart
1880          * transaction
1881          */
1882         osd_trans_declare_op(env, oh, OSD_OT_PUNCH,
1883                              osd_dto_credits_noquota[DTO_ATTR_SET_BASE] + 3);
1884
1885         inode = osd_dt_obj(dt)->oo_inode;
1886         LASSERT(inode);
1887
1888         rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
1889                                    i_projid_read(inode), 0, oh, osd_dt_obj(dt),
1890                                    NULL, OSD_QID_BLK);
1891
1892         if (rc == 0)
1893                 rc = osd_trunc_lock(osd_dt_obj(dt), oh, false);
1894
1895         RETURN(rc);
1896 }
1897
1898 static int osd_punch(const struct lu_env *env, struct dt_object *dt,
1899                      __u64 start, __u64 end, struct thandle *th)
1900 {
1901         struct osd_object *obj = osd_dt_obj(dt);
1902         struct osd_device *osd = osd_obj2dev(obj);
1903         struct inode *inode = obj->oo_inode;
1904         struct osd_access_lock *al;
1905         struct osd_thandle *oh;
1906         int rc = 0, found = 0;
1907         bool grow = false;
1908         ENTRY;
1909
1910         LASSERT(end == OBD_OBJECT_EOF);
1911         LASSERT(dt_object_exists(dt));
1912         LASSERT(osd_invariant(obj));
1913         LASSERT(inode != NULL);
1914         dquot_initialize(inode);
1915
1916         LASSERT(th);
1917         oh = container_of(th, struct osd_thandle, ot_super);
1918         LASSERT(oh->ot_handle->h_transaction != NULL);
1919
1920         /* we used to skip truncate to current size to
1921          * optimize truncates on OST. with DoM we can
1922          * get attr_set to set specific size (MDS_REINT)
1923          * and then get truncate RPC which essentially
1924          * would be skipped. this is bad.. so, disable
1925          * this optimization on MDS till the client stop
1926          * to sent MDS_REINT (LU-11033) -bzzz */
1927         if (osd->od_is_ost && i_size_read(inode) == start)
1928                 RETURN(0);
1929
1930         osd_trans_exec_op(env, th, OSD_OT_PUNCH);
1931
1932         spin_lock(&inode->i_lock);
1933         if (i_size_read(inode) < start)
1934                 grow = true;
1935         i_size_write(inode, start);
1936         spin_unlock(&inode->i_lock);
1937         ll_truncate_pagecache(inode, start);
1938
1939         /* optimize grow case */
1940         if (grow) {
1941                 osd_execute_truncate(obj);
1942                 GOTO(out, rc);
1943         }
1944
1945         /* add to orphan list to ensure truncate completion
1946          * if this transaction succeed. ldiskfs_truncate()
1947          * will take the inode out of the list */
1948         rc = ldiskfs_orphan_add(oh->ot_handle, inode);
1949         if (rc != 0)
1950                 GOTO(out, rc);
1951
1952         list_for_each_entry(al, &oh->ot_trunc_locks, tl_list) {
1953                 if (obj != al->tl_obj)
1954                         continue;
1955                 LASSERT(al->tl_shared == 0);
1956                 found = 1;
1957                 /* do actual truncate in osd_trans_stop() */
1958                 al->tl_truncate = 1;
1959                 break;
1960         }
1961         LASSERT(found);
1962
1963 out:
1964         RETURN(rc);
1965 }
1966
1967 static int fiemap_check_ranges(struct inode *inode,
1968                                u64 start, u64 len, u64 *new_len)
1969 {
1970         loff_t maxbytes;
1971
1972         *new_len = len;
1973
1974         if (len == 0)
1975                 return -EINVAL;
1976
1977         if (ldiskfs_test_inode_flag(inode, LDISKFS_INODE_EXTENTS))
1978                 maxbytes = inode->i_sb->s_maxbytes;
1979         else
1980                 maxbytes = LDISKFS_SB(inode->i_sb)->s_bitmap_maxbytes;
1981
1982         if (start > maxbytes)
1983                 return -EFBIG;
1984
1985         /*
1986          * Shrink request scope to what the fs can actually handle.
1987          */
1988         if (len > maxbytes || (maxbytes - len) < start)
1989                 *new_len = maxbytes - start;
1990
1991         return 0;
1992 }
1993
1994 /* So that the fiemap access checks can't overflow on 32 bit machines. */
1995 #define FIEMAP_MAX_EXTENTS     (UINT_MAX / sizeof(struct fiemap_extent))
1996
1997 static int osd_fiemap_get(const struct lu_env *env, struct dt_object *dt,
1998                           struct fiemap *fm)
1999 {
2000         struct fiemap_extent_info fieinfo = {0, };
2001         struct inode *inode = osd_dt_obj(dt)->oo_inode;
2002         u64 len;
2003         int rc;
2004         mm_segment_t cur_fs;
2005
2006         LASSERT(inode);
2007         if (inode->i_op->fiemap == NULL)
2008                 return -EOPNOTSUPP;
2009
2010         if (fm->fm_extent_count > FIEMAP_MAX_EXTENTS)
2011                 return -EINVAL;
2012
2013         rc = fiemap_check_ranges(inode, fm->fm_start, fm->fm_length, &len);
2014         if (rc)
2015                 return rc;
2016
2017         fieinfo.fi_flags = fm->fm_flags;
2018         fieinfo.fi_extents_max = fm->fm_extent_count;
2019         fieinfo.fi_extents_start = fm->fm_extents;
2020
2021         if (fieinfo.fi_flags & FIEMAP_FLAG_SYNC)
2022                 filemap_write_and_wait(inode->i_mapping);
2023
2024         /* Save previous value address limit */
2025         cur_fs = get_fs();
2026         /* Set the address limit of the kernel */
2027         set_fs(KERNEL_DS);
2028
2029         rc = inode->i_op->fiemap(inode, &fieinfo, fm->fm_start, len);
2030         fm->fm_flags = fieinfo.fi_flags;
2031         fm->fm_mapped_extents = fieinfo.fi_extents_mapped;
2032
2033         /* Restore the previous address limt */
2034         set_fs(cur_fs);
2035
2036         return rc;
2037 }
2038
2039 static int osd_ladvise(const struct lu_env *env, struct dt_object *dt,
2040                        __u64 start, __u64 end, enum lu_ladvise_type advice)
2041 {
2042         struct osd_object *obj = osd_dt_obj(dt);
2043         int rc = 0;
2044         ENTRY;
2045
2046         switch (advice) {
2047         case LU_LADVISE_DONTNEED:
2048                 if (end)
2049                         invalidate_mapping_pages(obj->oo_inode->i_mapping,
2050                                                  start >> PAGE_SHIFT,
2051                                                  (end - 1) >> PAGE_SHIFT);
2052                 break;
2053         default:
2054                 rc = -ENOTSUPP;
2055                 break;
2056         }
2057
2058         RETURN(rc);
2059 }
2060
2061 /*
2062  * in some cases we may need declare methods for objects being created
2063  * e.g., when we create symlink
2064  */
2065 const struct dt_body_operations osd_body_ops_new = {
2066         .dbo_declare_write = osd_declare_write,
2067 };
2068
2069 const struct dt_body_operations osd_body_ops = {
2070         .dbo_read                       = osd_read,
2071         .dbo_declare_write              = osd_declare_write,
2072         .dbo_write                      = osd_write,
2073         .dbo_bufs_get                   = osd_bufs_get,
2074         .dbo_bufs_put                   = osd_bufs_put,
2075         .dbo_write_prep                 = osd_write_prep,
2076         .dbo_declare_write_commit       = osd_declare_write_commit,
2077         .dbo_write_commit               = osd_write_commit,
2078         .dbo_read_prep                  = osd_read_prep,
2079         .dbo_declare_punch              = osd_declare_punch,
2080         .dbo_punch                      = osd_punch,
2081         .dbo_fiemap_get                 = osd_fiemap_get,
2082         .dbo_ladvise                    = osd_ladvise,
2083 };
2084
2085 /**
2086  * Get a truncate lock
2087  *
2088  * In order to take multi-transaction truncate out of main transaction we let
2089  * the caller grab a lock on the object passed. the lock can be shared (for
2090  * writes) and exclusive (for truncate). It's not allowed to mix truncate
2091  * and write in the same transaction handle (do not confuse with big ldiskfs
2092  * transaction containing lots of handles).
2093  * The lock must be taken at declaration.
2094  *
2095  * \param obj           object to lock
2096  * \oh                  transaction
2097  * \shared              shared or exclusive
2098  *
2099  * \retval 0            lock is granted
2100  * \retval -NOMEM       no memory to allocate lock
2101  */
2102 int osd_trunc_lock(struct osd_object *obj, struct osd_thandle *oh, bool shared)
2103 {
2104         struct osd_access_lock *al, *tmp;
2105
2106         LASSERT(obj);
2107         LASSERT(oh);
2108
2109         list_for_each_entry(tmp, &oh->ot_trunc_locks, tl_list) {
2110                 if (tmp->tl_obj != obj)
2111                         continue;
2112                 LASSERT(tmp->tl_shared == shared);
2113                 /* found same lock */
2114                 return 0;
2115         }
2116
2117         OBD_ALLOC_PTR(al);
2118         if (unlikely(al == NULL))
2119                 return -ENOMEM;
2120         al->tl_obj = obj;
2121         al->tl_truncate = false;
2122         if (shared)
2123                 down_read(&obj->oo_ext_idx_sem);
2124         else
2125                 down_write(&obj->oo_ext_idx_sem);
2126         al->tl_shared = shared;
2127
2128         list_add(&al->tl_list, &oh->ot_trunc_locks);
2129
2130         return 0;
2131 }
2132
2133 void osd_trunc_unlock_all(struct list_head *list)
2134 {
2135         struct osd_access_lock *al, *tmp;
2136         list_for_each_entry_safe(al, tmp, list, tl_list) {
2137                 if (al->tl_shared)
2138                         up_read(&al->tl_obj->oo_ext_idx_sem);
2139                 else
2140                         up_write(&al->tl_obj->oo_ext_idx_sem);
2141                 list_del(&al->tl_list);
2142                 OBD_FREE_PTR(al);
2143         }
2144 }
2145
2146 void osd_execute_truncate(struct osd_object *obj)
2147 {
2148         struct osd_device *d = osd_obj2dev(obj);
2149         struct inode *inode = obj->oo_inode;
2150         __u64 size;
2151
2152         /* simulate crash before (in the middle) of delayed truncate */
2153         if (OBD_FAIL_CHECK(OBD_FAIL_OSD_FAIL_AT_TRUNCATE)) {
2154                 struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
2155                 struct ldiskfs_sb_info *sbi = LDISKFS_SB(inode->i_sb);
2156
2157                 mutex_lock(&sbi->s_orphan_lock);
2158                 list_del_init(&ei->i_orphan);
2159                 mutex_unlock(&sbi->s_orphan_lock);
2160                 return;
2161         }
2162
2163         inode_lock(inode);
2164         ldiskfs_truncate(inode);
2165         inode_unlock(inode);
2166
2167         /*
2168          * For a partial-page truncate, flush the page to disk immediately to
2169          * avoid data corruption during direct disk write.  b=17397
2170          */
2171         size = i_size_read(inode);
2172         if ((size & ~PAGE_MASK) == 0)
2173                 return;
2174         if (osd_use_page_cache(d)) {
2175                 filemap_fdatawrite_range(inode->i_mapping, size, size + 1);
2176         } else {
2177                 /* Notice we use "wait" version to ensure I/O is complete */
2178                 filemap_write_and_wait_range(inode->i_mapping, size, size + 1);
2179                 invalidate_mapping_pages(inode->i_mapping, size >> PAGE_SHIFT,
2180                                          size >> PAGE_SHIFT);
2181         }
2182 }
2183
2184 void osd_process_truncates(struct list_head *list)
2185 {
2186         struct osd_access_lock *al;
2187
2188         LASSERT(journal_current_handle() == NULL);
2189
2190         list_for_each_entry(al, list, tl_list) {
2191                 if (al->tl_shared)
2192                         continue;
2193                 if (!al->tl_truncate)
2194                         continue;
2195                 osd_execute_truncate(al->tl_obj);
2196         }
2197 }