Whamcloud - gitweb
LU-12275 sec: deal with encrypted object size
[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(!PagePrivate2(page));
671                         wait_on_page_writeback(page);
672                 } else {
673                         lprocfs_counter_add(d->od_stats, LPROC_OSD_NO_PAGE, 1);
674                 }
675
676                 return page;
677         }
678
679         if (inode->i_mapping->nrpages) {
680                 /* consult with pagecache, but do not create new pages */
681                 /* this is normally used once */
682                 page = find_lock_page(inode->i_mapping, offset >> PAGE_SHIFT);
683                 if (page) {
684                         wait_on_page_writeback(page);
685                         return page;
686                 }
687         }
688
689         LASSERT(oti->oti_dio_pages);
690         cur = oti->oti_dio_pages_used;
691         page = oti->oti_dio_pages[cur];
692
693         if (unlikely(!page)) {
694                 LASSERT(cur < PTLRPC_MAX_BRW_PAGES);
695                 page = alloc_page(gfp_mask);
696                 if (!page)
697                         return NULL;
698                 oti->oti_dio_pages[cur] = page;
699                 SetPagePrivate2(page);
700                 lock_page(page);
701         }
702
703         ClearPageUptodate(page);
704         page->index = offset >> PAGE_SHIFT;
705         oti->oti_dio_pages_used++;
706
707         return page;
708 }
709
710 /*
711  * there are following "locks":
712  * journal_start
713  * i_mutex
714  * page lock
715  *
716  * osd write path:
717  *  - lock page(s)
718  *  - journal_start
719  *  - truncate_sem
720  *
721  * ext4 vmtruncate:
722  *  - lock pages, unlock
723  *  - journal_start
724  *  - lock partial page
725  *  - i_data_sem
726  *
727  */
728
729 /**
730  * Unlock and release pages loaded by osd_bufs_get()
731  *
732  * Unlock \a npages pages from \a lnb and drop the refcount on them.
733  *
734  * \param env           thread execution environment
735  * \param dt            dt object undergoing IO (OSD object + methods)
736  * \param lnb           array of pages undergoing IO
737  * \param npages        number of pages in \a lnb
738  *
739  * \retval 0            always
740  */
741 static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt,
742                         struct niobuf_local *lnb, int npages)
743 {
744         struct osd_thread_info *oti = osd_oti_get(env);
745         struct pagevec pvec;
746         int i;
747
748         ll_pagevec_init(&pvec, 0);
749
750         for (i = 0; i < npages; i++) {
751                 struct page *page = lnb[i].lnb_page;
752
753                 if (page == NULL)
754                         continue;
755
756                 /* if the page isn't cached, then reset uptodate
757                  * to prevent reuse */
758                 if (PagePrivate2(page)) {
759                         oti->oti_dio_pages_used--;
760                 } else {
761                         if (lnb[i].lnb_locked)
762                                 unlock_page(page);
763                         if (pagevec_add(&pvec, page) == 0)
764                                 pagevec_release(&pvec);
765                 }
766
767                 lnb[i].lnb_page = NULL;
768         }
769
770         LASSERTF(oti->oti_dio_pages_used == 0, "%d\n", oti->oti_dio_pages_used);
771
772         /* Release any partial pagevec */
773         pagevec_release(&pvec);
774
775         RETURN(0);
776 }
777
778 /**
779  * Load and lock pages undergoing IO
780  *
781  * Pages as described in the \a lnb array are fetched (from disk or cache)
782  * and locked for IO by the caller.
783  *
784  * DLM locking protects us from write and truncate competing for same region,
785  * but partial-page truncate can leave dirty pages in the cache for ldiskfs.
786  * It's possible the writeout on a such a page is in progress when we access
787  * it. It's also possible that during this writeout we put new (partial) data
788  * into the page, but won't be able to proceed in filter_commitrw_write().
789  * Therefore, just wait for writeout completion as it should be rare enough.
790  *
791  * \param env           thread execution environment
792  * \param dt            dt object undergoing IO (OSD object + methods)
793  * \param pos           byte offset of IO start
794  * \param len           number of bytes of IO
795  * \param lnb           array of extents undergoing IO
796  * \param rw            read or write operation, and other flags
797  * \param capa          capabilities
798  *
799  * \retval pages        (zero or more) loaded successfully
800  * \retval -ENOMEM      on memory/page allocation error
801  */
802 static int osd_bufs_get(const struct lu_env *env, struct dt_object *dt,
803                         loff_t pos, ssize_t len, struct niobuf_local *lnb,
804                         int maxlnb, enum dt_bufs_type rw)
805 {
806         struct osd_thread_info *oti = osd_oti_get(env);
807         struct osd_object *obj = osd_dt_obj(dt);
808         struct osd_device *osd   = osd_obj2dev(obj);
809         int npages, i, iosize, rc = 0;
810         bool cache, write;
811         loff_t fsize;
812         gfp_t gfp_mask;
813
814         LASSERT(obj->oo_inode);
815
816         rc = osd_map_remote_to_local(pos, len, &npages, lnb, maxlnb);
817         if (rc)
818                 RETURN(rc);
819
820         write = rw & DT_BUFS_TYPE_WRITE;
821
822         fsize = lnb[npages - 1].lnb_file_offset + lnb[npages - 1].lnb_len;
823         iosize = fsize - lnb[0].lnb_file_offset;
824         fsize = max(fsize, i_size_read(obj->oo_inode));
825
826         cache = rw & DT_BUFS_TYPE_READAHEAD;
827         if (cache)
828                 goto bypass_checks;
829
830         cache = osd_use_page_cache(osd);
831         while (cache) {
832                 if (write) {
833                         if (!osd->od_writethrough_cache) {
834                                 cache = false;
835                                 break;
836                         }
837                         if (iosize > osd->od_writethrough_max_iosize) {
838                                 cache = false;
839                                 break;
840                         }
841                 } else {
842                         if (!osd->od_read_cache) {
843                                 cache = false;
844                                 break;
845                         }
846                         if (iosize > osd->od_readcache_max_iosize) {
847                                 cache = false;
848                                 break;
849                         }
850                 }
851                 /* don't use cache on large files */
852                 if (osd->od_readcache_max_filesize &&
853                     fsize > osd->od_readcache_max_filesize)
854                         cache = false;
855                 break;
856         }
857
858 bypass_checks:
859         if (!cache && unlikely(!oti->oti_dio_pages)) {
860                 OBD_ALLOC_PTR_ARRAY(oti->oti_dio_pages, PTLRPC_MAX_BRW_PAGES);
861                 if (!oti->oti_dio_pages)
862                         return -ENOMEM;
863         }
864
865         /* this could also try less hard for DT_BUFS_TYPE_READAHEAD pages */
866         gfp_mask = rw & DT_BUFS_TYPE_LOCAL ? (GFP_NOFS | __GFP_HIGHMEM) :
867                                              GFP_HIGHUSER;
868         for (i = 0; i < npages; i++, lnb++) {
869                 lnb->lnb_page = osd_get_page(env, dt, lnb->lnb_file_offset,
870                                              gfp_mask, cache);
871                 if (lnb->lnb_page == NULL)
872                         GOTO(cleanup, rc = -ENOMEM);
873
874                 lnb->lnb_locked = 1;
875         }
876
877 #if 0
878         /* XXX: this version doesn't invalidate cached pages, but use them */
879         if (!cache && write && obj->oo_inode->i_mapping->nrpages) {
880                 /* do not allow data aliasing, invalidate pagecache */
881                 /* XXX: can be quite expensive in mixed case */
882                 invalidate_mapping_pages(obj->oo_inode->i_mapping,
883                                 lnb[0].lnb_file_offset >> PAGE_SHIFT,
884                                 lnb[npages - 1].lnb_file_offset >> PAGE_SHIFT);
885         }
886 #endif
887
888         RETURN(i);
889
890 cleanup:
891         if (i > 0)
892                 osd_bufs_put(env, dt, lnb - i, i);
893         return rc;
894 }
895
896 static int osd_ldiskfs_map_inode_pages(struct inode *inode, struct page **page,
897                                        int pages, sector_t *blocks,
898                                        int create)
899 {
900         int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
901         int rc = 0, i = 0;
902         struct page *fp = NULL;
903         int clen = 0;
904         pgoff_t max_page_index;
905         handle_t *handle = NULL;
906
907         max_page_index = inode->i_sb->s_maxbytes >> PAGE_SHIFT;
908
909         CDEBUG(D_OTHER, "inode %lu: map %d pages from %lu\n",
910                 inode->i_ino, pages, (*page)->index);
911
912         if (create) {
913                 create = LDISKFS_GET_BLOCKS_CREATE;
914                 handle = ldiskfs_journal_current_handle();
915                 LASSERT(handle != NULL);
916                 rc = osd_attach_jinode(inode);
917                 if (rc)
918                         return rc;
919         }
920         /* pages are sorted already. so, we just have to find
921          * contig. space and process them properly */
922         while (i < pages) {
923                 long blen, total = 0;
924                 struct ldiskfs_map_blocks map = { 0 };
925
926                 if (fp == NULL) { /* start new extent */
927                         fp = *page++;
928                         clen = 1;
929                         if (++i != pages)
930                                 continue;
931                 } else if (fp->index + clen == (*page)->index) {
932                         /* continue the extent */
933                         page++;
934                         clen++;
935                         if (++i != pages)
936                                 continue;
937                 }
938                 if (fp->index + clen >= max_page_index)
939                         GOTO(cleanup, rc = -EFBIG);
940                 /* process found extent */
941                 map.m_lblk = fp->index * blocks_per_page;
942                 map.m_len = blen = clen * blocks_per_page;
943 cont_map:
944                 rc = ldiskfs_map_blocks(handle, inode, &map, create);
945                 if (rc >= 0) {
946                         int c = 0;
947                         for (; total < blen && c < map.m_len; c++, total++) {
948                                 if (rc == 0) {
949                                         *(blocks + total) = 0;
950                                         total++;
951                                         break;
952                                 } else {
953                                         *(blocks + total) = map.m_pblk + c;
954                                         /* unmap any possible underlying
955                                          * metadata from the block device
956                                          * mapping.  bug 6998. */
957                                         if ((map.m_flags & LDISKFS_MAP_NEW) &&
958                                             create)
959                                                 clean_bdev_aliases(
960                                                         inode->i_sb->s_bdev,
961                                                         map.m_pblk + c, 1);
962                                 }
963                         }
964                         rc = 0;
965                 }
966                 if (rc == 0 && total < blen) {
967                         map.m_lblk = fp->index * blocks_per_page + total;
968                         map.m_len = blen - total;
969                         goto cont_map;
970                 }
971                 if (rc != 0)
972                         GOTO(cleanup, rc);
973
974                 /* look for next extent */
975                 fp = NULL;
976                 blocks += blocks_per_page * clen;
977         }
978 cleanup:
979         return rc;
980 }
981
982 static int osd_write_prep(const struct lu_env *env, struct dt_object *dt,
983                           struct niobuf_local *lnb, int npages)
984 {
985         struct osd_thread_info *oti   = osd_oti_get(env);
986         struct osd_iobuf       *iobuf = &oti->oti_iobuf;
987         struct inode           *inode = osd_dt_obj(dt)->oo_inode;
988         struct osd_device      *osd   = osd_obj2dev(osd_dt_obj(dt));
989         ktime_t start, end;
990         s64 timediff;
991         ssize_t isize;
992         __s64  maxidx;
993         int i, rc = 0;
994
995         LASSERT(inode);
996
997         rc = osd_init_iobuf(osd, iobuf, 0, npages);
998         if (unlikely(rc != 0))
999                 RETURN(rc);
1000
1001         isize = i_size_read(inode);
1002         maxidx = ((isize + PAGE_SIZE - 1) >> PAGE_SHIFT) - 1;
1003
1004         start = ktime_get();
1005         for (i = 0; i < npages; i++) {
1006
1007                 /*
1008                  * till commit the content of the page is undefined
1009                  * we'll set it uptodate once bulk is done. otherwise
1010                  * subsequent reads can access non-stable data
1011                  */
1012                 ClearPageUptodate(lnb[i].lnb_page);
1013
1014                 if (lnb[i].lnb_len == PAGE_SIZE)
1015                         continue;
1016
1017                 if (maxidx >= lnb[i].lnb_page->index) {
1018                         osd_iobuf_add_page(iobuf, &lnb[i]);
1019                 } else {
1020                         long off;
1021                         char *p = kmap(lnb[i].lnb_page);
1022
1023                         off = lnb[i].lnb_page_offset;
1024                         if (off)
1025                                 memset(p, 0, off);
1026                         off = (lnb[i].lnb_page_offset + lnb[i].lnb_len) &
1027                               ~PAGE_MASK;
1028                         if (off)
1029                                 memset(p + off, 0, PAGE_SIZE - off);
1030                         kunmap(lnb[i].lnb_page);
1031                 }
1032         }
1033         end = ktime_get();
1034         timediff = ktime_us_delta(end, start);
1035         lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
1036
1037         if (iobuf->dr_npages) {
1038                 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1039                                                  iobuf->dr_npages,
1040                                                  iobuf->dr_blocks, 0);
1041                 if (likely(rc == 0)) {
1042                         rc = osd_do_bio(osd, inode, iobuf);
1043                         /* do IO stats for preparation reads */
1044                         osd_fini_iobuf(osd, iobuf);
1045                 }
1046         }
1047         RETURN(rc);
1048 }
1049
1050 struct osd_fextent {
1051         sector_t        start;
1052         sector_t        end;
1053         unsigned int    mapped:1;
1054 };
1055
1056 static int osd_is_mapped(struct dt_object *dt, __u64 offset,
1057                          struct osd_fextent *cached_extent)
1058 {
1059         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1060         sector_t block = offset >> inode->i_blkbits;
1061         sector_t start;
1062         struct fiemap_extent_info fei = { 0 };
1063         struct fiemap_extent fe = { 0 };
1064         mm_segment_t saved_fs;
1065         int rc;
1066
1067         if (block >= cached_extent->start && block < cached_extent->end)
1068                 return cached_extent->mapped;
1069
1070         if (i_size_read(inode) == 0)
1071                 return 0;
1072
1073         /* Beyond EOF, must not be mapped */
1074         if (((i_size_read(inode) - 1) >> inode->i_blkbits) < block)
1075                 return 0;
1076
1077         fei.fi_extents_max = 1;
1078         fei.fi_extents_start = &fe;
1079
1080         saved_fs = get_fs();
1081         set_fs(KERNEL_DS);
1082         rc = inode->i_op->fiemap(inode, &fei, offset, FIEMAP_MAX_OFFSET-offset);
1083         set_fs(saved_fs);
1084         if (rc != 0)
1085                 return 0;
1086
1087         start = fe.fe_logical >> inode->i_blkbits;
1088
1089         if (start > block) {
1090                 cached_extent->start = block;
1091                 cached_extent->end = start;
1092                 cached_extent->mapped = 0;
1093         } else {
1094                 cached_extent->start = start;
1095                 cached_extent->end = (fe.fe_logical + fe.fe_length) >>
1096                                       inode->i_blkbits;
1097                 cached_extent->mapped = 1;
1098         }
1099
1100         return cached_extent->mapped;
1101 }
1102
1103 static int osd_declare_write_commit(const struct lu_env *env,
1104                                     struct dt_object *dt,
1105                                     struct niobuf_local *lnb, int npages,
1106                                     struct thandle *handle)
1107 {
1108         const struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1109         struct inode            *inode = osd_dt_obj(dt)->oo_inode;
1110         struct osd_thandle      *oh;
1111         int                     extents = 1;
1112         int                     depth;
1113         int                     i;
1114         int                     newblocks;
1115         int                     rc = 0;
1116         int                     credits = 0;
1117         long long               quota_space = 0;
1118         struct osd_fextent      extent = { 0 };
1119         enum osd_quota_local_flags local_flags = 0;
1120         enum osd_qid_declare_flags declare_flags = OSD_QID_BLK;
1121         ENTRY;
1122
1123         LASSERT(handle != NULL);
1124         oh = container_of(handle, struct osd_thandle, ot_super);
1125         LASSERT(oh->ot_handle == NULL);
1126
1127         newblocks = npages;
1128
1129         /* calculate number of extents (probably better to pass nb) */
1130         for (i = 0; i < npages; i++) {
1131                 if (i && lnb[i].lnb_file_offset !=
1132                     lnb[i - 1].lnb_file_offset + lnb[i - 1].lnb_len)
1133                         extents++;
1134
1135                 if (osd_is_mapped(dt, lnb[i].lnb_file_offset, &extent))
1136                         lnb[i].lnb_flags |= OBD_BRW_MAPPED;
1137                 else
1138                         quota_space += PAGE_SIZE;
1139
1140                 /* ignore quota for the whole request if any page is from
1141                  * client cache or written by root.
1142                  *
1143                  * XXX once we drop the 1.8 client support, the checking
1144                  * for whether page is from cache can be simplified as:
1145                  * !(lnb[i].flags & OBD_BRW_SYNC)
1146                  *
1147                  * XXX we could handle this on per-lnb basis as done by
1148                  * grant. */
1149                 if ((lnb[i].lnb_flags & OBD_BRW_NOQUOTA) ||
1150                     (lnb[i].lnb_flags & (OBD_BRW_FROM_GRANT | OBD_BRW_SYNC)) ==
1151                     OBD_BRW_FROM_GRANT)
1152                         declare_flags |= OSD_QID_FORCE;
1153         }
1154
1155         /*
1156          * each extent can go into new leaf causing a split
1157          * 5 is max tree depth: inode + 4 index blocks
1158          * with blockmaps, depth is 3 at most
1159          */
1160         if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL) {
1161                 /*
1162                  * many concurrent threads may grow tree by the time
1163                  * our transaction starts. so, consider 2 is a min depth
1164                  */
1165                 depth = ext_depth(inode);
1166                 depth = max(depth, 1) + 1;
1167                 newblocks += depth;
1168                 credits++; /* inode */
1169                 credits += depth * 2 * extents;
1170         } else {
1171                 depth = 3;
1172                 newblocks += depth;
1173                 credits++; /* inode */
1174                 credits += depth * extents;
1175         }
1176
1177         /* quota space for metadata blocks */
1178         quota_space += depth * extents * LDISKFS_BLOCK_SIZE(osd_sb(osd));
1179
1180         /* quota space should be reported in 1K blocks */
1181         quota_space = toqb(quota_space);
1182
1183         /* each new block can go in different group (bitmap + gd) */
1184
1185         /* we can't dirty more bitmap blocks than exist */
1186         if (newblocks > LDISKFS_SB(osd_sb(osd))->s_groups_count)
1187                 credits += LDISKFS_SB(osd_sb(osd))->s_groups_count;
1188         else
1189                 credits += newblocks;
1190
1191         /* we can't dirty more gd blocks than exist */
1192         if (newblocks > LDISKFS_SB(osd_sb(osd))->s_gdb_count)
1193                 credits += LDISKFS_SB(osd_sb(osd))->s_gdb_count;
1194         else
1195                 credits += newblocks;
1196
1197         osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
1198
1199         /* make sure the over quota flags were not set */
1200         lnb[0].lnb_flags &= ~OBD_BRW_OVER_ALLQUOTA;
1201
1202         rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
1203                                    i_projid_read(inode), quota_space, oh,
1204                                    osd_dt_obj(dt), &local_flags, declare_flags);
1205
1206         /* we need only to store the overquota flags in the first lnb for
1207          * now, once we support multiple objects BRW, this code needs be
1208          * revised. */
1209         if (local_flags & QUOTA_FL_OVER_USRQUOTA)
1210                 lnb[0].lnb_flags |= OBD_BRW_OVER_USRQUOTA;
1211         if (local_flags & QUOTA_FL_OVER_GRPQUOTA)
1212                 lnb[0].lnb_flags |= OBD_BRW_OVER_GRPQUOTA;
1213         if (local_flags & QUOTA_FL_OVER_PRJQUOTA)
1214                 lnb[0].lnb_flags |= OBD_BRW_OVER_PRJQUOTA;
1215
1216         if (rc == 0)
1217                 rc = osd_trunc_lock(osd_dt_obj(dt), oh, true);
1218
1219         RETURN(rc);
1220 }
1221
1222 /* Check if a block is allocated or not */
1223 static int osd_write_commit(const struct lu_env *env, struct dt_object *dt,
1224                             struct niobuf_local *lnb, int npages,
1225                             struct thandle *thandle, __u64 user_size)
1226 {
1227         struct osd_thread_info *oti = osd_oti_get(env);
1228         struct osd_iobuf *iobuf = &oti->oti_iobuf;
1229         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1230         struct osd_device  *osd = osd_obj2dev(osd_dt_obj(dt));
1231         loff_t disk_size;
1232         int rc = 0, i;
1233
1234         LASSERT(inode);
1235
1236         rc = osd_init_iobuf(osd, iobuf, 1, npages);
1237         if (unlikely(rc != 0))
1238                 RETURN(rc);
1239
1240         disk_size = i_size_read(inode);
1241         /* if disk_size is already bigger than specified user_size,
1242          * ignore user_size
1243          */
1244         if (disk_size > user_size)
1245                 user_size = 0;
1246         dquot_initialize(inode);
1247
1248         for (i = 0; i < npages; i++) {
1249                 if (lnb[i].lnb_rc == -ENOSPC &&
1250                     (lnb[i].lnb_flags & OBD_BRW_MAPPED)) {
1251                         /* Allow the write to proceed if overwriting an
1252                          * existing block */
1253                         lnb[i].lnb_rc = 0;
1254                 }
1255
1256                 if (lnb[i].lnb_rc) { /* ENOSPC, network RPC error, etc. */
1257                         CDEBUG(D_INODE, "Skipping [%d] == %d\n", i,
1258                                lnb[i].lnb_rc);
1259                         LASSERT(lnb[i].lnb_page);
1260                         generic_error_remove_page(inode->i_mapping,
1261                                                   lnb[i].lnb_page);
1262                         continue;
1263                 }
1264
1265                 LASSERT(PageLocked(lnb[i].lnb_page));
1266                 LASSERT(!PageWriteback(lnb[i].lnb_page));
1267
1268                 if (lnb[i].lnb_file_offset + lnb[i].lnb_len > disk_size)
1269                         disk_size = lnb[i].lnb_file_offset + lnb[i].lnb_len;
1270
1271                 /*
1272                  * Since write and truncate are serialized by oo_sem, even
1273                  * partial-page truncate should not leave dirty pages in the
1274                  * page cache.
1275                  */
1276                 LASSERT(!PageDirty(lnb[i].lnb_page));
1277
1278                 SetPageUptodate(lnb[i].lnb_page);
1279
1280                 osd_iobuf_add_page(iobuf, &lnb[i]);
1281         }
1282         /* if file has grown, take user_size into account */
1283         if (user_size && disk_size > user_size)
1284                 disk_size = user_size;
1285
1286         osd_trans_exec_op(env, thandle, OSD_OT_WRITE);
1287
1288         if (OBD_FAIL_CHECK(OBD_FAIL_OST_MAPBLK_ENOSPC)) {
1289                 rc = -ENOSPC;
1290         } else if (iobuf->dr_npages > 0) {
1291                 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1292                                                  iobuf->dr_npages,
1293                                                  iobuf->dr_blocks, 1);
1294         } else {
1295                 /* no pages to write, no transno is needed */
1296                 thandle->th_local = 1;
1297         }
1298
1299         if (likely(rc == 0)) {
1300                 spin_lock(&inode->i_lock);
1301                 if (disk_size > i_size_read(inode)) {
1302                         i_size_write(inode, disk_size);
1303                         LDISKFS_I(inode)->i_disksize = disk_size;
1304                         spin_unlock(&inode->i_lock);
1305                         osd_dirty_inode(inode, I_DIRTY_DATASYNC);
1306                 } else {
1307                         spin_unlock(&inode->i_lock);
1308                 }
1309
1310                 rc = osd_do_bio(osd, inode, iobuf);
1311                 /* we don't do stats here as in read path because
1312                  * write is async: we'll do this in osd_put_bufs() */
1313         } else {
1314                 osd_fini_iobuf(osd, iobuf);
1315         }
1316
1317         osd_trans_exec_check(env, thandle, OSD_OT_WRITE);
1318
1319         if (unlikely(rc != 0)) {
1320                 /* if write fails, we should drop pages from the cache */
1321                 for (i = 0; i < npages; i++) {
1322                         if (lnb[i].lnb_page == NULL)
1323                                 continue;
1324                         if (!PagePrivate2(lnb[i].lnb_page)) {
1325                                 LASSERT(PageLocked(lnb[i].lnb_page));
1326                                 generic_error_remove_page(inode->i_mapping,
1327                                                           lnb[i].lnb_page);
1328                         }
1329                 }
1330         }
1331
1332         RETURN(rc);
1333 }
1334
1335 static int osd_read_prep(const struct lu_env *env, struct dt_object *dt,
1336                          struct niobuf_local *lnb, int npages)
1337 {
1338         struct osd_thread_info *oti = osd_oti_get(env);
1339         struct osd_iobuf *iobuf = &oti->oti_iobuf;
1340         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1341         struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1342         int rc = 0, i, cache_hits = 0, cache_misses = 0;
1343         ktime_t start, end;
1344         s64 timediff;
1345         loff_t isize;
1346
1347         LASSERT(inode);
1348
1349         rc = osd_init_iobuf(osd, iobuf, 0, npages);
1350         if (unlikely(rc != 0))
1351                 RETURN(rc);
1352
1353         isize = i_size_read(inode);
1354
1355         start = ktime_get();
1356         for (i = 0; i < npages; i++) {
1357
1358                 if (isize <= lnb[i].lnb_file_offset)
1359                         /* If there's no more data, abort early.
1360                          * lnb->lnb_rc == 0, so it's easy to detect later. */
1361                         break;
1362
1363                 /* instead of looking if we go beyong isize, send complete
1364                  * pages all the time
1365                  */
1366                 lnb[i].lnb_rc = lnb[i].lnb_len;
1367
1368                 /* Bypass disk read if fail_loc is set properly */
1369                 if (OBD_FAIL_CHECK(OBD_FAIL_OST_FAKE_RW))
1370                         SetPageUptodate(lnb[i].lnb_page);
1371
1372                 if (PageUptodate(lnb[i].lnb_page)) {
1373                         cache_hits++;
1374                         unlock_page(lnb[i].lnb_page);
1375                 } else {
1376                         cache_misses++;
1377                         osd_iobuf_add_page(iobuf, &lnb[i]);
1378                 }
1379                 /* no need to unlock in osd_bufs_put(), the sooner page is
1380                  * unlocked, the earlier another client can access it.
1381                  * notice real unlock_page() can be called few lines
1382                  * below after osd_do_bio(). lnb is a per-thread, so it's
1383                  * fine to have PG_locked and lnb_locked inconsistent here */
1384                 lnb[i].lnb_locked = 0;
1385         }
1386         end = ktime_get();
1387         timediff = ktime_us_delta(end, start);
1388         lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
1389
1390         if (cache_hits != 0)
1391                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_HIT,
1392                                     cache_hits);
1393         if (cache_misses != 0)
1394                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_MISS,
1395                                     cache_misses);
1396         if (cache_hits + cache_misses != 0)
1397                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_ACCESS,
1398                                     cache_hits + cache_misses);
1399
1400         if (iobuf->dr_npages) {
1401                 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1402                                                  iobuf->dr_npages,
1403                                                  iobuf->dr_blocks, 0);
1404                 rc = osd_do_bio(osd, inode, iobuf);
1405
1406                 /* IO stats will be done in osd_bufs_put() */
1407
1408                 /* early release to let others read data during the bulk */
1409                 for (i = 0; i < iobuf->dr_npages; i++) {
1410                         LASSERT(PageLocked(iobuf->dr_pages[i]));
1411                         if (!PagePrivate2(iobuf->dr_pages[i]))
1412                                 unlock_page(iobuf->dr_pages[i]);
1413                 }
1414         }
1415
1416         RETURN(rc);
1417 }
1418
1419 /*
1420  * XXX: Another layering violation for now.
1421  *
1422  * We don't want to use ->f_op->read methods, because generic file write
1423  *
1424  *         - serializes on ->i_sem, and
1425  *
1426  *         - does a lot of extra work like balance_dirty_pages(),
1427  *
1428  * which doesn't work for globally shared files like /last_rcvd.
1429  */
1430 static int osd_ldiskfs_readlink(struct inode *inode, char *buffer, int buflen)
1431 {
1432         struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
1433
1434         memcpy(buffer, (char *)ei->i_data, buflen);
1435
1436         return  buflen;
1437 }
1438
1439 int osd_ldiskfs_read(struct inode *inode, void *buf, int size, loff_t *offs)
1440 {
1441         struct buffer_head *bh;
1442         unsigned long block;
1443         int osize;
1444         int blocksize;
1445         int csize;
1446         int boffs;
1447
1448         /* prevent reading after eof */
1449         spin_lock(&inode->i_lock);
1450         if (i_size_read(inode) < *offs + size) {
1451                 loff_t diff = i_size_read(inode) - *offs;
1452                 spin_unlock(&inode->i_lock);
1453                 if (diff < 0) {
1454                         CDEBUG(D_OTHER,
1455                                "size %llu is too short to read @%llu\n",
1456                                i_size_read(inode), *offs);
1457                         return -EBADR;
1458                 } else if (diff == 0) {
1459                         return 0;
1460                 } else {
1461                         size = diff;
1462                 }
1463         } else {
1464                 spin_unlock(&inode->i_lock);
1465         }
1466
1467         blocksize = 1 << inode->i_blkbits;
1468         osize = size;
1469         while (size > 0) {
1470                 block = *offs >> inode->i_blkbits;
1471                 boffs = *offs & (blocksize - 1);
1472                 csize = min(blocksize - boffs, size);
1473                 bh = __ldiskfs_bread(NULL, inode, block, 0);
1474                 if (IS_ERR(bh)) {
1475                         CERROR("%s: can't read %u@%llu on ino %lu: "
1476                                "rc = %ld\n", osd_ino2name(inode),
1477                                csize, *offs, inode->i_ino,
1478                                PTR_ERR(bh));
1479                         return PTR_ERR(bh);
1480                 }
1481
1482                 if (bh != NULL) {
1483                         memcpy(buf, bh->b_data + boffs, csize);
1484                         brelse(bh);
1485                 } else {
1486                         memset(buf, 0, csize);
1487                 }
1488
1489                 *offs += csize;
1490                 buf += csize;
1491                 size -= csize;
1492         }
1493         return osize;
1494 }
1495
1496 static ssize_t osd_read(const struct lu_env *env, struct dt_object *dt,
1497                         struct lu_buf *buf, loff_t *pos)
1498 {
1499         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1500         int rc;
1501
1502         /* Read small symlink from inode body as we need to maintain correct
1503          * on-disk symlinks for ldiskfs.
1504          */
1505         if (S_ISLNK(dt->do_lu.lo_header->loh_attr)) {
1506                 loff_t size = i_size_read(inode);
1507
1508                 if (buf->lb_len < size)
1509                         return -EOVERFLOW;
1510
1511                 if (size < sizeof(LDISKFS_I(inode)->i_data))
1512                         rc = osd_ldiskfs_readlink(inode, buf->lb_buf, size);
1513                 else
1514                         rc = osd_ldiskfs_read(inode, buf->lb_buf, size, pos);
1515         } else {
1516                 rc = osd_ldiskfs_read(inode, buf->lb_buf, buf->lb_len, pos);
1517         }
1518
1519         return rc;
1520 }
1521
1522 static inline int osd_extents_enabled(struct super_block *sb,
1523                                       struct inode *inode)
1524 {
1525         if (inode != NULL) {
1526                 if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL)
1527                         return 1;
1528         } else if (ldiskfs_has_feature_extents(sb)) {
1529                 return 1;
1530         }
1531         return 0;
1532 }
1533
1534 int osd_calc_bkmap_credits(struct super_block *sb, struct inode *inode,
1535                            const loff_t size, const loff_t pos,
1536                            const int blocks)
1537 {
1538         int credits, bits, bs, i;
1539
1540         bits = sb->s_blocksize_bits;
1541         bs = 1 << bits;
1542
1543         /* legacy blockmap: 3 levels * 3 (bitmap,gd,itself)
1544          * we do not expect blockmaps on the large files,
1545          * so let's shrink it to 2 levels (4GB files) */
1546
1547         /* this is default reservation: 2 levels */
1548         credits = (blocks + 2) * 3;
1549
1550         /* actual offset is unknown, hard to optimize */
1551         if (pos == -1)
1552                 return credits;
1553
1554         /* now check for few specific cases to optimize */
1555         if (pos + size <= LDISKFS_NDIR_BLOCKS * bs) {
1556                 /* no indirects */
1557                 credits = blocks;
1558                 /* allocate if not allocated */
1559                 if (inode == NULL) {
1560                         credits += blocks * 2;
1561                         return credits;
1562                 }
1563                 for (i = (pos >> bits); i < (pos >> bits) + blocks; i++) {
1564                         LASSERT(i < LDISKFS_NDIR_BLOCKS);
1565                         if (LDISKFS_I(inode)->i_data[i] == 0)
1566                                 credits += 2;
1567                 }
1568         } else if (pos + size <= (LDISKFS_NDIR_BLOCKS + 1024) * bs) {
1569                 /* single indirect */
1570                 credits = blocks * 3;
1571                 if (inode == NULL ||
1572                     LDISKFS_I(inode)->i_data[LDISKFS_IND_BLOCK] == 0)
1573                         credits += 3;
1574                 else
1575                         /* The indirect block may be modified. */
1576                         credits += 1;
1577         }
1578
1579         return credits;
1580 }
1581
1582 static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt,
1583                                  const struct lu_buf *buf, loff_t _pos,
1584                                  struct thandle *handle)
1585 {
1586         struct osd_object  *obj  = osd_dt_obj(dt);
1587         struct inode       *inode = obj->oo_inode;
1588         struct super_block *sb = osd_sb(osd_obj2dev(obj));
1589         struct osd_thandle *oh;
1590         int                 rc = 0, est = 0, credits, blocks, allocated = 0;
1591         int                 bits, bs;
1592         int                 depth, size;
1593         loff_t              pos;
1594         ENTRY;
1595
1596         LASSERT(buf != NULL);
1597         LASSERT(handle != NULL);
1598
1599         oh = container_of(handle, struct osd_thandle, ot_super);
1600         LASSERT(oh->ot_handle == NULL);
1601
1602         size = buf->lb_len;
1603         bits = sb->s_blocksize_bits;
1604         bs = 1 << bits;
1605
1606         if (_pos == -1) {
1607                 /* if this is an append, then we
1608                  * should expect cross-block record */
1609                 pos = 0;
1610         } else {
1611                 pos = _pos;
1612         }
1613
1614         /* blocks to modify */
1615         blocks = ((pos + size + bs - 1) >> bits) - (pos >> bits);
1616         LASSERT(blocks > 0);
1617
1618         if (inode != NULL && _pos != -1) {
1619                 /* object size in blocks */
1620                 est = (i_size_read(inode) + bs - 1) >> bits;
1621                 allocated = inode->i_blocks >> (bits - 9);
1622                 if (pos + size <= i_size_read(inode) && est <= allocated) {
1623                         /* looks like an overwrite, no need to modify tree */
1624                         credits = blocks;
1625                         /* no need to modify i_size */
1626                         goto out;
1627                 }
1628         }
1629
1630         if (osd_extents_enabled(sb, inode)) {
1631                 /*
1632                  * many concurrent threads may grow tree by the time
1633                  * our transaction starts. so, consider 2 is a min depth
1634                  * for every level we may need to allocate a new block
1635                  * and take some entries from the old one. so, 3 blocks
1636                  * to allocate (bitmap, gd, itself) + old block - 4 per
1637                  * level.
1638                  */
1639                 depth = inode != NULL ? ext_depth(inode) : 0;
1640                 depth = max(depth, 1) + 1;
1641                 credits = depth;
1642                 /* if not append, then split may need to modify
1643                  * existing blocks moving entries into the new ones */
1644                 if (_pos != -1)
1645                         credits += depth;
1646                 /* blocks to store data: bitmap,gd,itself */
1647                 credits += blocks * 3;
1648         } else {
1649                 credits = osd_calc_bkmap_credits(sb, inode, size, _pos, blocks);
1650         }
1651         /* if inode is created as part of the transaction,
1652          * then it's counted already by the creation method */
1653         if (inode != NULL)
1654                 credits++;
1655
1656 out:
1657
1658         osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
1659
1660         /* dt_declare_write() is usually called for system objects, such
1661          * as llog or last_rcvd files. We needn't enforce quota on those
1662          * objects, so always set the lqi_space as 0. */
1663         if (inode != NULL)
1664                 rc = osd_declare_inode_qid(env, i_uid_read(inode),
1665                                            i_gid_read(inode),
1666                                            i_projid_read(inode), 0,
1667                                            oh, obj, NULL, OSD_QID_BLK);
1668
1669         if (rc == 0)
1670                 rc = osd_trunc_lock(obj, oh, true);
1671
1672         RETURN(rc);
1673 }
1674
1675 static int osd_ldiskfs_writelink(struct inode *inode, char *buffer, int buflen)
1676 {
1677         /* LU-2634: clear the extent format for fast symlink */
1678         ldiskfs_clear_inode_flag(inode, LDISKFS_INODE_EXTENTS);
1679
1680         memcpy((char *)&LDISKFS_I(inode)->i_data, (char *)buffer, buflen);
1681         spin_lock(&inode->i_lock);
1682         LDISKFS_I(inode)->i_disksize = buflen;
1683         i_size_write(inode, buflen);
1684         spin_unlock(&inode->i_lock);
1685         osd_dirty_inode(inode, I_DIRTY_DATASYNC);
1686
1687         return 0;
1688 }
1689
1690 static int osd_ldiskfs_write_record(struct dt_object *dt, void *buf,
1691                                     int bufsize, int write_NUL, loff_t *offs,
1692                                     handle_t *handle)
1693 {
1694         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1695         struct buffer_head *bh        = NULL;
1696         loff_t              offset    = *offs;
1697         loff_t              new_size  = i_size_read(inode);
1698         unsigned long       block;
1699         int                 blocksize = 1 << inode->i_blkbits;
1700         struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
1701         int                 err = 0;
1702         int                 size;
1703         int                 boffs;
1704         int                 dirty_inode = 0;
1705         bool create, sparse, sync = false;
1706
1707         if (write_NUL) {
1708                 /*
1709                  * long symlink write does not count the NUL terminator in
1710                  * bufsize, we write it, and the inode's file size does not
1711                  * count the NUL terminator as well.
1712                  */
1713                 ((char *)buf)[bufsize] = '\0';
1714                 ++bufsize;
1715         }
1716
1717         dirty_inode = test_and_set_bit(LDISKFS_INODE_JOURNAL_DATA, &ei->i_flags);
1718
1719         /* sparse checking is racy, but sparse is very rare case, leave as is */
1720         sparse = (new_size > 0 && (inode->i_blocks >> (inode->i_blkbits - 9)) <
1721                   ((new_size - 1) >> inode->i_blkbits) + 1);
1722
1723         while (bufsize > 0) {
1724                 int credits = handle->h_buffer_credits;
1725                 unsigned long last_block = (new_size == 0) ? 0 :
1726                                            (new_size - 1) >> inode->i_blkbits;
1727
1728                 if (bh)
1729                         brelse(bh);
1730
1731                 block = offset >> inode->i_blkbits;
1732                 boffs = offset & (blocksize - 1);
1733                 size = min(blocksize - boffs, bufsize);
1734                 sync = (block > last_block || new_size == 0 || sparse);
1735
1736                 if (sync)
1737                         down(&ei->i_append_sem);
1738
1739                 bh = __ldiskfs_bread(handle, inode, block, 0);
1740
1741                 if (unlikely(IS_ERR_OR_NULL(bh) && !sync))
1742                         CWARN("%s: adding bh without locking off %llu (block %lu, "
1743                               "size %d, offs %llu)\n", inode->i_sb->s_id,
1744                               offset, block, bufsize, *offs);
1745
1746                 if (IS_ERR_OR_NULL(bh)) {
1747                         struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1748                         int flags = LDISKFS_GET_BLOCKS_CREATE;
1749
1750                         /* while the file system is being mounted, avoid
1751                          * preallocation otherwise mount can take a long
1752                          * time as mballoc cache is cold.
1753                          * XXX: this is a workaround until we have a proper
1754                          *      fix in mballoc
1755                          * XXX: works with extent-based files only */
1756                         if (!osd->od_cl_seq)
1757                                 flags |= LDISKFS_GET_BLOCKS_NO_NORMALIZE;
1758                         bh = __ldiskfs_bread(handle, inode, block, flags);
1759                         create = true;
1760                 } else {
1761                         if (sync) {
1762                                 up(&ei->i_append_sem);
1763                                 sync = false;
1764                         }
1765                         create = false;
1766                 }
1767                 if (IS_ERR_OR_NULL(bh)) {
1768                         if (bh == NULL) {
1769                                 err = -EIO;
1770                         } else {
1771                                 err = PTR_ERR(bh);
1772                                 bh = NULL;
1773                         }
1774
1775                         CERROR("%s: error reading offset %llu (block %lu, "
1776                                "size %d, offs %llu), credits %d/%d: rc = %d\n",
1777                                inode->i_sb->s_id, offset, block, bufsize, *offs,
1778                                credits, handle->h_buffer_credits, err);
1779                         break;
1780                 }
1781
1782                 err = ldiskfs_journal_get_write_access(handle, bh);
1783                 if (err) {
1784                         CERROR("journal_get_write_access() returned error %d\n",
1785                                err);
1786                         break;
1787                 }
1788                 LASSERTF(boffs + size <= bh->b_size,
1789                          "boffs %d size %d bh->b_size %lu\n",
1790                          boffs, size, (unsigned long)bh->b_size);
1791                 if (create) {
1792                         memset(bh->b_data, 0, bh->b_size);
1793                         if (sync) {
1794                                 up(&ei->i_append_sem);
1795                                 sync = false;
1796                         }
1797                 }
1798                 memcpy(bh->b_data + boffs, buf, size);
1799                 err = ldiskfs_handle_dirty_metadata(handle, NULL, bh);
1800                 if (err)
1801                         break;
1802
1803                 if (offset + size > new_size)
1804                         new_size = offset + size;
1805                 offset += size;
1806                 bufsize -= size;
1807                 buf += size;
1808         }
1809         if (sync)
1810                 up(&ei->i_append_sem);
1811
1812         if (bh)
1813                 brelse(bh);
1814
1815         if (write_NUL)
1816                 --new_size;
1817         /* correct in-core and on-disk sizes */
1818         if (new_size > i_size_read(inode)) {
1819                 spin_lock(&inode->i_lock);
1820                 if (new_size > i_size_read(inode))
1821                         i_size_write(inode, new_size);
1822                 if (i_size_read(inode) > ei->i_disksize) {
1823                         ei->i_disksize = i_size_read(inode);
1824                         dirty_inode = 1;
1825                 }
1826                 spin_unlock(&inode->i_lock);
1827         }
1828         if (dirty_inode)
1829                 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
1830
1831         if (err == 0)
1832                 *offs = offset;
1833         return err;
1834 }
1835
1836 static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt,
1837                          const struct lu_buf *buf, loff_t *pos,
1838                          struct thandle *handle)
1839 {
1840         struct inode            *inode = osd_dt_obj(dt)->oo_inode;
1841         struct osd_thandle      *oh;
1842         ssize_t                 result;
1843         int                     is_link;
1844
1845         LASSERT(dt_object_exists(dt));
1846
1847         LASSERT(handle != NULL);
1848         LASSERT(inode != NULL);
1849         dquot_initialize(inode);
1850
1851         /* XXX: don't check: one declared chunk can be used many times */
1852         /* osd_trans_exec_op(env, handle, OSD_OT_WRITE); */
1853
1854         oh = container_of(handle, struct osd_thandle, ot_super);
1855         LASSERT(oh->ot_handle->h_transaction != NULL);
1856         osd_trans_exec_op(env, handle, OSD_OT_WRITE);
1857
1858         /* Write small symlink to inode body as we need to maintain correct
1859          * on-disk symlinks for ldiskfs.
1860          * Note: the buf->lb_buf contains a NUL terminator while buf->lb_len
1861          * does not count it in.
1862          */
1863         is_link = S_ISLNK(dt->do_lu.lo_header->loh_attr);
1864         if (is_link && (buf->lb_len < sizeof(LDISKFS_I(inode)->i_data)))
1865                 result = osd_ldiskfs_writelink(inode, buf->lb_buf, buf->lb_len);
1866         else
1867                 result = osd_ldiskfs_write_record(dt, buf->lb_buf, buf->lb_len,
1868                                                   is_link, pos, oh->ot_handle);
1869         if (result == 0)
1870                 result = buf->lb_len;
1871
1872         osd_trans_exec_check(env, handle, OSD_OT_WRITE);
1873
1874         return result;
1875 }
1876
1877 static int osd_declare_fallocate(const struct lu_env *env,
1878                                  struct dt_object *dt, struct thandle *th)
1879 {
1880         struct osd_thandle *oh;
1881         struct inode *inode;
1882         int rc;
1883         ENTRY;
1884
1885         LASSERT(th);
1886         oh = container_of(th, struct osd_thandle, ot_super);
1887
1888         osd_trans_declare_op(env, oh, OSD_OT_PREALLOC,
1889                              osd_dto_credits_noquota[DTO_WRITE_BLOCK]);
1890         inode = osd_dt_obj(dt)->oo_inode;
1891         LASSERT(inode);
1892
1893         rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
1894                                    i_projid_read(inode), 0, oh, osd_dt_obj(dt),
1895                                    NULL, OSD_QID_BLK);
1896         RETURN(rc);
1897 }
1898
1899 static int osd_fallocate(const struct lu_env *env, struct dt_object *dt,
1900                          __u64 start, __u64 end, int mode, struct thandle *th)
1901 {
1902         struct osd_object *obj = osd_dt_obj(dt);
1903         struct inode *inode = obj->oo_inode;
1904         int rc = 0;
1905         struct osd_thread_info *info = osd_oti_get(env);
1906         struct dentry *dentry = &info->oti_obj_dentry;
1907         struct file *file = &info->oti_file;
1908
1909         ENTRY;
1910         /*
1911          * Only mode == 0 (which is standard prealloc) is supported now.
1912          * Rest of mode options is not supported yet.
1913          */
1914         if (mode & ~FALLOC_FL_KEEP_SIZE)
1915                 RETURN(-EOPNOTSUPP);
1916
1917         LASSERT(dt_object_exists(dt));
1918         LASSERT(osd_invariant(obj));
1919         LASSERT(inode != NULL);
1920         dquot_initialize(inode);
1921
1922         LASSERT(th);
1923
1924         osd_trans_exec_op(env, th, OSD_OT_PREALLOC);
1925
1926         /*
1927          * Because f_op->fallocate() does not have an inode arg
1928          */
1929         dentry->d_inode = inode;
1930         dentry->d_sb = inode->i_sb;
1931         file->f_path.dentry = dentry;
1932         file->f_mapping = inode->i_mapping;
1933         file->f_op = inode->i_fop;
1934         file->f_inode = inode;
1935         rc = file->f_op->fallocate(file, mode, start, end - start);
1936
1937         RETURN(rc);
1938 }
1939
1940 static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt,
1941                              __u64 start, __u64 end, struct thandle *th)
1942 {
1943         struct osd_thandle *oh;
1944         struct inode       *inode;
1945         int                 rc;
1946         ENTRY;
1947
1948         LASSERT(th);
1949         oh = container_of(th, struct osd_thandle, ot_super);
1950
1951         /*
1952          * we don't need to reserve credits for whole truncate
1953          * it's not possible as truncate may need to free too many
1954          * blocks and that won't fit a single transaction. instead
1955          * we reserve credits to change i_size and put inode onto
1956          * orphan list. if needed truncate will extend or restart
1957          * transaction
1958          */
1959         osd_trans_declare_op(env, oh, OSD_OT_PUNCH,
1960                              osd_dto_credits_noquota[DTO_ATTR_SET_BASE] + 3);
1961
1962         inode = osd_dt_obj(dt)->oo_inode;
1963         LASSERT(inode);
1964
1965         rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
1966                                    i_projid_read(inode), 0, oh, osd_dt_obj(dt),
1967                                    NULL, OSD_QID_BLK);
1968
1969         if (rc == 0)
1970                 rc = osd_trunc_lock(osd_dt_obj(dt), oh, false);
1971
1972         RETURN(rc);
1973 }
1974
1975 static int osd_punch(const struct lu_env *env, struct dt_object *dt,
1976                      __u64 start, __u64 end, struct thandle *th)
1977 {
1978         struct osd_object *obj = osd_dt_obj(dt);
1979         struct osd_device *osd = osd_obj2dev(obj);
1980         struct inode *inode = obj->oo_inode;
1981         struct osd_access_lock *al;
1982         struct osd_thandle *oh;
1983         int rc = 0, found = 0;
1984         bool grow = false;
1985         ENTRY;
1986
1987         LASSERT(dt_object_exists(dt));
1988         LASSERT(osd_invariant(obj));
1989         LASSERT(inode != NULL);
1990         dquot_initialize(inode);
1991
1992         LASSERT(th);
1993         oh = container_of(th, struct osd_thandle, ot_super);
1994         LASSERT(oh->ot_handle->h_transaction != NULL);
1995
1996         /* we used to skip truncate to current size to
1997          * optimize truncates on OST. with DoM we can
1998          * get attr_set to set specific size (MDS_REINT)
1999          * and then get truncate RPC which essentially
2000          * would be skipped. this is bad.. so, disable
2001          * this optimization on MDS till the client stop
2002          * to sent MDS_REINT (LU-11033) -bzzz */
2003         if (osd->od_is_ost && i_size_read(inode) == start)
2004                 RETURN(0);
2005
2006         osd_trans_exec_op(env, th, OSD_OT_PUNCH);
2007
2008         spin_lock(&inode->i_lock);
2009         if (i_size_read(inode) < start)
2010                 grow = true;
2011         i_size_write(inode, start);
2012         spin_unlock(&inode->i_lock);
2013         ll_truncate_pagecache(inode, start);
2014
2015         /* optimize grow case */
2016         if (grow) {
2017                 osd_execute_truncate(obj);
2018                 GOTO(out, rc);
2019         }
2020
2021         inode_lock(inode);
2022         /* add to orphan list to ensure truncate completion
2023          * if this transaction succeed. ldiskfs_truncate()
2024          * will take the inode out of the list */
2025         rc = ldiskfs_orphan_add(oh->ot_handle, inode);
2026         inode_unlock(inode);
2027         if (rc != 0)
2028                 GOTO(out, rc);
2029
2030         list_for_each_entry(al, &oh->ot_trunc_locks, tl_list) {
2031                 if (obj != al->tl_obj)
2032                         continue;
2033                 LASSERT(al->tl_shared == 0);
2034                 found = 1;
2035                 /* do actual truncate in osd_trans_stop() */
2036                 al->tl_truncate = 1;
2037                 break;
2038         }
2039         LASSERT(found);
2040
2041 out:
2042         RETURN(rc);
2043 }
2044
2045 static int fiemap_check_ranges(struct inode *inode,
2046                                u64 start, u64 len, u64 *new_len)
2047 {
2048         loff_t maxbytes;
2049
2050         *new_len = len;
2051
2052         if (len == 0)
2053                 return -EINVAL;
2054
2055         if (ldiskfs_test_inode_flag(inode, LDISKFS_INODE_EXTENTS))
2056                 maxbytes = inode->i_sb->s_maxbytes;
2057         else
2058                 maxbytes = LDISKFS_SB(inode->i_sb)->s_bitmap_maxbytes;
2059
2060         if (start > maxbytes)
2061                 return -EFBIG;
2062
2063         /*
2064          * Shrink request scope to what the fs can actually handle.
2065          */
2066         if (len > maxbytes || (maxbytes - len) < start)
2067                 *new_len = maxbytes - start;
2068
2069         return 0;
2070 }
2071
2072 /* So that the fiemap access checks can't overflow on 32 bit machines. */
2073 #define FIEMAP_MAX_EXTENTS     (UINT_MAX / sizeof(struct fiemap_extent))
2074
2075 static int osd_fiemap_get(const struct lu_env *env, struct dt_object *dt,
2076                           struct fiemap *fm)
2077 {
2078         struct fiemap_extent_info fieinfo = {0, };
2079         struct inode *inode = osd_dt_obj(dt)->oo_inode;
2080         u64 len;
2081         int rc;
2082         mm_segment_t cur_fs;
2083
2084         LASSERT(inode);
2085         if (inode->i_op->fiemap == NULL)
2086                 return -EOPNOTSUPP;
2087
2088         if (fm->fm_extent_count > FIEMAP_MAX_EXTENTS)
2089                 return -EINVAL;
2090
2091         rc = fiemap_check_ranges(inode, fm->fm_start, fm->fm_length, &len);
2092         if (rc)
2093                 return rc;
2094
2095         fieinfo.fi_flags = fm->fm_flags;
2096         fieinfo.fi_extents_max = fm->fm_extent_count;
2097         fieinfo.fi_extents_start = fm->fm_extents;
2098
2099         if (fieinfo.fi_flags & FIEMAP_FLAG_SYNC)
2100                 filemap_write_and_wait(inode->i_mapping);
2101
2102         /* Save previous value address limit */
2103         cur_fs = get_fs();
2104         /* Set the address limit of the kernel */
2105         set_fs(KERNEL_DS);
2106
2107         rc = inode->i_op->fiemap(inode, &fieinfo, fm->fm_start, len);
2108         fm->fm_flags = fieinfo.fi_flags;
2109         fm->fm_mapped_extents = fieinfo.fi_extents_mapped;
2110
2111         /* Restore the previous address limt */
2112         set_fs(cur_fs);
2113
2114         return rc;
2115 }
2116
2117 static int osd_ladvise(const struct lu_env *env, struct dt_object *dt,
2118                        __u64 start, __u64 end, enum lu_ladvise_type advice)
2119 {
2120         struct osd_object *obj = osd_dt_obj(dt);
2121         int rc = 0;
2122         ENTRY;
2123
2124         switch (advice) {
2125         case LU_LADVISE_DONTNEED:
2126                 if (end)
2127                         invalidate_mapping_pages(obj->oo_inode->i_mapping,
2128                                                  start >> PAGE_SHIFT,
2129                                                  (end - 1) >> PAGE_SHIFT);
2130                 break;
2131         default:
2132                 rc = -ENOTSUPP;
2133                 break;
2134         }
2135
2136         RETURN(rc);
2137 }
2138
2139 /*
2140  * in some cases we may need declare methods for objects being created
2141  * e.g., when we create symlink
2142  */
2143 const struct dt_body_operations osd_body_ops_new = {
2144         .dbo_declare_write = osd_declare_write,
2145 };
2146
2147 const struct dt_body_operations osd_body_ops = {
2148         .dbo_read                       = osd_read,
2149         .dbo_declare_write              = osd_declare_write,
2150         .dbo_write                      = osd_write,
2151         .dbo_bufs_get                   = osd_bufs_get,
2152         .dbo_bufs_put                   = osd_bufs_put,
2153         .dbo_write_prep                 = osd_write_prep,
2154         .dbo_declare_write_commit       = osd_declare_write_commit,
2155         .dbo_write_commit               = osd_write_commit,
2156         .dbo_read_prep                  = osd_read_prep,
2157         .dbo_declare_punch              = osd_declare_punch,
2158         .dbo_punch                      = osd_punch,
2159         .dbo_fiemap_get                 = osd_fiemap_get,
2160         .dbo_ladvise                    = osd_ladvise,
2161         .dbo_declare_fallocate          = osd_declare_fallocate,
2162         .dbo_fallocate                  = osd_fallocate,
2163 };
2164
2165 /**
2166  * Get a truncate lock
2167  *
2168  * In order to take multi-transaction truncate out of main transaction we let
2169  * the caller grab a lock on the object passed. the lock can be shared (for
2170  * writes) and exclusive (for truncate). It's not allowed to mix truncate
2171  * and write in the same transaction handle (do not confuse with big ldiskfs
2172  * transaction containing lots of handles).
2173  * The lock must be taken at declaration.
2174  *
2175  * \param obj           object to lock
2176  * \oh                  transaction
2177  * \shared              shared or exclusive
2178  *
2179  * \retval 0            lock is granted
2180  * \retval -NOMEM       no memory to allocate lock
2181  */
2182 int osd_trunc_lock(struct osd_object *obj, struct osd_thandle *oh, bool shared)
2183 {
2184         struct osd_access_lock *al, *tmp;
2185
2186         LASSERT(obj);
2187         LASSERT(oh);
2188
2189         list_for_each_entry(tmp, &oh->ot_trunc_locks, tl_list) {
2190                 if (tmp->tl_obj != obj)
2191                         continue;
2192                 LASSERT(tmp->tl_shared == shared);
2193                 /* found same lock */
2194                 return 0;
2195         }
2196
2197         OBD_ALLOC_PTR(al);
2198         if (unlikely(al == NULL))
2199                 return -ENOMEM;
2200         al->tl_obj = obj;
2201         al->tl_truncate = false;
2202         if (shared)
2203                 down_read(&obj->oo_ext_idx_sem);
2204         else
2205                 down_write(&obj->oo_ext_idx_sem);
2206         al->tl_shared = shared;
2207         lu_object_get(&obj->oo_dt.do_lu);
2208
2209         list_add(&al->tl_list, &oh->ot_trunc_locks);
2210
2211         return 0;
2212 }
2213
2214 void osd_trunc_unlock_all(const struct lu_env *env, struct list_head *list)
2215 {
2216         struct osd_access_lock *al, *tmp;
2217         list_for_each_entry_safe(al, tmp, list, tl_list) {
2218                 if (al->tl_shared)
2219                         up_read(&al->tl_obj->oo_ext_idx_sem);
2220                 else
2221                         up_write(&al->tl_obj->oo_ext_idx_sem);
2222                 osd_object_put(env, al->tl_obj);
2223                 list_del(&al->tl_list);
2224                 OBD_FREE_PTR(al);
2225         }
2226 }
2227
2228 void osd_execute_truncate(struct osd_object *obj)
2229 {
2230         struct osd_device *d = osd_obj2dev(obj);
2231         struct inode *inode = obj->oo_inode;
2232         __u64 size;
2233
2234         /* simulate crash before (in the middle) of delayed truncate */
2235         if (OBD_FAIL_CHECK(OBD_FAIL_OSD_FAIL_AT_TRUNCATE)) {
2236                 struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
2237                 struct ldiskfs_sb_info *sbi = LDISKFS_SB(inode->i_sb);
2238
2239                 mutex_lock(&sbi->s_orphan_lock);
2240                 list_del_init(&ei->i_orphan);
2241                 mutex_unlock(&sbi->s_orphan_lock);
2242                 return;
2243         }
2244
2245         inode_lock(inode);
2246         ldiskfs_truncate(inode);
2247         inode_unlock(inode);
2248
2249         /*
2250          * For a partial-page truncate, flush the page to disk immediately to
2251          * avoid data corruption during direct disk write.  b=17397
2252          */
2253         size = i_size_read(inode);
2254         if ((size & ~PAGE_MASK) == 0)
2255                 return;
2256         if (osd_use_page_cache(d)) {
2257                 filemap_fdatawrite_range(inode->i_mapping, size, size + 1);
2258         } else {
2259                 /* Notice we use "wait" version to ensure I/O is complete */
2260                 filemap_write_and_wait_range(inode->i_mapping, size, size + 1);
2261                 invalidate_mapping_pages(inode->i_mapping, size >> PAGE_SHIFT,
2262                                          size >> PAGE_SHIFT);
2263         }
2264 }
2265
2266 void osd_process_truncates(struct list_head *list)
2267 {
2268         struct osd_access_lock *al;
2269
2270         LASSERT(journal_current_handle() == NULL);
2271
2272         list_for_each_entry(al, list, tl_list) {
2273                 if (al->tl_shared)
2274                         continue;
2275                 if (!al->tl_truncate)
2276                         continue;
2277                 osd_execute_truncate(al->tl_obj);
2278         }
2279 }