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 #define DEBUG_SUBSYSTEM S_OSD
  42
  43 /* prerequisite for linux/xattr.h */
  44 #include <linux/types.h>
  45 /* prerequisite for linux/xattr.h */
  46 #include <linux/fs.h>
  47 #include <linux/mm.h>
  48 #include <linux/pagevec.h>
  49
  50 /*
  51  * struct OBD_{ALLOC,FREE}*()
  52  * OBD_FAIL_CHECK
  53  */
  54 #include <obd_support.h>
  55
  56 #include "osd_internal.h"
  57
  58 /* ext_depth() */
  59 #include <ldiskfs/ldiskfs_extents.h>
  60
  61 static inline bool osd_use_page_cache(struct osd_device *d)
  62 {
  63         /* do not use pagecache if write and read caching are disabled */
  64         if (d->od_writethrough_cache + d->od_read_cache == 0)
  65                 return false;
  66         /* use pagecache by default */
  67         return true;
  68 }
  69
  70 static int __osd_init_iobuf(struct osd_device *d, struct osd_iobuf *iobuf,
  71                             int rw, int line, int pages)
  72 {
  73         int blocks, i;
  74
  75         LASSERTF(iobuf->dr_elapsed_valid == 0,
  76                  "iobuf %p, reqs %d, rw %d, line %d\n", iobuf,
  77                  atomic_read(&iobuf->dr_numreqs), iobuf->dr_rw,
  78                  iobuf->dr_init_at);
  79         LASSERT(pages <= PTLRPC_MAX_BRW_PAGES);
  80
  81         init_waitqueue_head(&iobuf->dr_wait);
  82         atomic_set(&iobuf->dr_numreqs, 0);
  83         iobuf->dr_npages = 0;
  84         iobuf->dr_error = 0;
  85         iobuf->dr_dev = d;
  86         iobuf->dr_frags = 0;
  87         iobuf->dr_elapsed = ktime_set(0, 0);
  88         /* must be counted before, so assert */
  89         iobuf->dr_rw = rw;
  90         iobuf->dr_init_at = line;
  91
  92         blocks = pages * (PAGE_SIZE >> osd_sb(d)->s_blocksize_bits);
  93         if (iobuf->dr_bl_buf.lb_len >= blocks * sizeof(iobuf->dr_blocks[0])) {
  94                 LASSERT(iobuf->dr_pg_buf.lb_len >=
  95                         pages * sizeof(iobuf->dr_pages[0]));
  96                 return 0;
  97         }
  98
  99         /* start with 1MB for 4K blocks */
 100         i = 256;
 101         while (i <= PTLRPC_MAX_BRW_PAGES && i < pages)
 102                 i <<= 1;
 103
 104         CDEBUG(D_OTHER, "realloc %u for %u (%u) pages\n",
 105                (unsigned int)(pages * sizeof(iobuf->dr_pages[0])), i, pages);
 106         pages = i;
 107         blocks = pages * (PAGE_SIZE >> osd_sb(d)->s_blocksize_bits);
 108         iobuf->dr_max_pages = 0;
 109         CDEBUG(D_OTHER, "realloc %u for %u blocks\n",
 110                (unsigned int)(blocks * sizeof(iobuf->dr_blocks[0])), blocks);
 111
 112         lu_buf_realloc(&iobuf->dr_bl_buf, blocks * sizeof(iobuf->dr_blocks[0]));
 113         iobuf->dr_blocks = iobuf->dr_bl_buf.lb_buf;
 114         if (unlikely(iobuf->dr_blocks == NULL))
 115                 return -ENOMEM;
 116
 117         lu_buf_realloc(&iobuf->dr_pg_buf, pages * sizeof(iobuf->dr_pages[0]));
 118         iobuf->dr_pages = iobuf->dr_pg_buf.lb_buf;
 119         if (unlikely(iobuf->dr_pages == NULL))
 120                 return -ENOMEM;
 121
 122         lu_buf_realloc(&iobuf->dr_lnb_buf,
 123                        pages * sizeof(iobuf->dr_lnbs[0]));
 124         iobuf->dr_lnbs = iobuf->dr_lnb_buf.lb_buf;
 125         if (unlikely(iobuf->dr_lnbs == NULL))
 126                 return -ENOMEM;
 127
 128         iobuf->dr_max_pages = pages;
 129
 130         return 0;
 131 }
 132 #define osd_init_iobuf(dev, iobuf, rw, pages) \
 133         __osd_init_iobuf(dev, iobuf, rw, __LINE__, pages)
 134
 135 static void osd_iobuf_add_page(struct osd_iobuf *iobuf,
 136                                struct niobuf_local *lnb)
 137 {
 138         LASSERT(iobuf->dr_npages < iobuf->dr_max_pages);
 139         iobuf->dr_pages[iobuf->dr_npages] = lnb->lnb_page;
 140         iobuf->dr_lnbs[iobuf->dr_npages] = lnb;
 141         iobuf->dr_npages++;
 142 }
 143
 144 void osd_fini_iobuf(struct osd_device *d, struct osd_iobuf *iobuf)
 145 {
 146         int rw = iobuf->dr_rw;
 147
 148         if (iobuf->dr_elapsed_valid) {
 149                 iobuf->dr_elapsed_valid = 0;
 150                 LASSERT(iobuf->dr_dev == d);
 151                 LASSERT(iobuf->dr_frags > 0);
 152                 lprocfs_oh_tally(&d->od_brw_stats.hist[BRW_R_DIO_FRAGS+rw],
 153                                  iobuf->dr_frags);
 154                 lprocfs_oh_tally_log2(&d->od_brw_stats.hist[BRW_R_IO_TIME+rw],
 155                                       ktime_to_ms(iobuf->dr_elapsed));
 156         }
 157 }
 158
 159 #ifdef HAVE_BIO_ENDIO_USES_ONE_ARG
 160 static void dio_complete_routine(struct bio *bio)
 161 {
 162         int error = bio->bi_status;
 163 #else
 164 static void dio_complete_routine(struct bio *bio, int error)
 165 {
 166 #endif
 167         struct osd_iobuf *iobuf = bio->bi_private;
 168         struct bio_vec *bvl;
 169
 170         /* CAVEAT EMPTOR: possibly in IRQ context
 171          * DO NOT record procfs stats here!!!
 172          */
 173
 174         if (unlikely(iobuf == NULL)) {
 175                 CERROR("***** bio->bi_private is NULL!  This should never happen.  Normally, I would crash here, but instead I will dump the bio contents to the console.  Please report this to <https://jira.whamcloud.com/> , along with any interesting messages leading up to this point (like SCSI errors, perhaps).  Because bi_private is NULL, I can't wake up the thread that initiated this IO - you will probably have to reboot this node.\n");
 176                 CERROR("bi_next: %p, bi_flags: %lx, " __stringify(bi_opf)
 177                        ": %x, bi_vcnt: %d, bi_idx: %d, bi->size: %d, bi_end_io: %p, bi_cnt: %d, bi_private: %p\n",
 178                        bio->bi_next, (unsigned long)bio->bi_flags,
 179                        (unsigned int)bio->bi_opf, bio->bi_vcnt, bio_idx(bio),
 180                        bio_sectors(bio) << 9, bio->bi_end_io,
 181                        atomic_read(&bio->__bi_cnt),
 182                        bio->bi_private);
 183                 return;
 184         }
 185
 186         /* the check is outside of the cycle for performance reason -bzzz */
 187         if (!bio_data_dir(bio)) {
 188                 DECLARE_BVEC_ITER_ALL(iter_all);
 189
 190                 bio_for_each_segment_all(bvl, bio, iter_all) {
 191                         if (likely(error == 0))
 192                                 SetPageUptodate(bvl_to_page(bvl));
 193                         LASSERT(PageLocked(bvl_to_page(bvl)));
 194                 }
 195                 atomic_dec(&iobuf->dr_dev->od_r_in_flight);
 196         } else {
 197                 atomic_dec(&iobuf->dr_dev->od_w_in_flight);
 198         }
 199
 200         /* any real error is good enough -bzzz */
 201         if (error != 0 && iobuf->dr_error == 0)
 202                 iobuf->dr_error = error;
 203
 204         /*
 205          * set dr_elapsed before dr_numreqs turns to 0, otherwise
 206          * it's possible that service thread will see dr_numreqs
 207          * is zero, but dr_elapsed is not set yet, leading to lost
 208          * data in this processing and an assertion in a subsequent
 209          * call to OSD.
 210          */
 211         if (atomic_read(&iobuf->dr_numreqs) == 1) {
 212                 ktime_t now = ktime_get();
 213
 214                 iobuf->dr_elapsed = ktime_sub(now, iobuf->dr_start_time);
 215                 iobuf->dr_elapsed_valid = 1;
 216         }
 217         if (atomic_dec_and_test(&iobuf->dr_numreqs))
 218                 wake_up(&iobuf->dr_wait);
 219
 220         /* Completed bios used to be chained off iobuf->dr_bios and freed in
 221          * filter_clear_dreq().  It was then possible to exhaust the biovec-256
 222          * mempool when serious on-disk fragmentation was encountered,
 223          * deadlocking the OST.  The bios are now released as soon as complete
 224          * so the pool cannot be exhausted while IOs are competing. b=10076
 225          */
 226         bio_put(bio);
 227 }
 228
 229 static void record_start_io(struct osd_iobuf *iobuf, int size)
 230 {
 231         struct osd_device    *osd = iobuf->dr_dev;
 232         struct obd_histogram *h = osd->od_brw_stats.hist;
 233
 234         iobuf->dr_frags++;
 235         atomic_inc(&iobuf->dr_numreqs);
 236
 237         if (iobuf->dr_rw == 0) {
 238                 atomic_inc(&osd->od_r_in_flight);
 239                 lprocfs_oh_tally(&h[BRW_R_RPC_HIST],
 240                                  atomic_read(&osd->od_r_in_flight));
 241                 lprocfs_oh_tally_log2(&h[BRW_R_DISK_IOSIZE], size);
 242         } else if (iobuf->dr_rw == 1) {
 243                 atomic_inc(&osd->od_w_in_flight);
 244                 lprocfs_oh_tally(&h[BRW_W_RPC_HIST],
 245                                  atomic_read(&osd->od_w_in_flight));
 246                 lprocfs_oh_tally_log2(&h[BRW_W_DISK_IOSIZE], size);
 247         } else {
 248                 LBUG();
 249         }
 250 }
 251
 252 static void osd_submit_bio(int rw, struct bio *bio)
 253 {
 254         LASSERTF(rw == 0 || rw == 1, "%x\n", rw);
 255 #ifdef HAVE_SUBMIT_BIO_2ARGS
 256         submit_bio(rw ? WRITE : READ, bio);
 257 #else
 258         bio->bi_opf |= rw;
 259         submit_bio(bio);
 260 #endif
 261 }
 262
 263 static int can_be_merged(struct bio *bio, sector_t sector)
 264 {
 265         if (bio == NULL)
 266                 return 0;
 267
 268         return bio_end_sector(bio) == sector ? 1 : 0;
 269 }
 270
 271 #if IS_ENABLED(CONFIG_BLK_DEV_INTEGRITY)
 272 /*
 273  * This function will change the data written, thus it should only be
 274  * used when checking data integrity feature
 275  */
 276 static void bio_integrity_fault_inject(struct bio *bio)
 277 {
 278         struct bio_vec *bvec;
 279         DECLARE_BVEC_ITER_ALL(iter_all);
 280         void *kaddr;
 281         char *addr;
 282
 283         bio_for_each_segment_all(bvec, bio, iter_all) {
 284                 struct page *page = bvec->bv_page;
 285
 286                 kaddr = kmap(page);
 287                 addr = kaddr;
 288                 *addr = ~(*addr);
 289                 kunmap(page);
 290                 break;
 291         }
 292 }
 293
 294 static int bio_dif_compare(__u16 *expected_guard_buf, void *bio_prot_buf,
 295                            unsigned int sectors, int tuple_size)
 296 {
 297         __u16 *expected_guard;
 298         __u16 *bio_guard;
 299         int i;
 300
 301         expected_guard = expected_guard_buf;
 302         for (i = 0; i < sectors; i++) {
 303                 bio_guard = (__u16 *)bio_prot_buf;
 304                 if (*bio_guard != *expected_guard) {
 305                         CERROR(
 306                                "unexpected guard tags on sector %d expected guard %u, bio guard %u, sectors %u, tuple size %d\n",
 307                                i, *expected_guard, *bio_guard, sectors,
 308                                tuple_size);
 309                         return -EIO;
 310                 }
 311                 expected_guard++;
 312                 bio_prot_buf += tuple_size;
 313         }
 314         return 0;
 315 }
 316
 317 static int osd_bio_integrity_compare(struct bio *bio, struct block_device *bdev,
 318                                      struct osd_iobuf *iobuf, int index)
 319 {
 320         struct blk_integrity *bi = bdev_get_integrity(bdev);
 321         struct bio_integrity_payload *bip = bio->bi_integrity;
 322         struct niobuf_local *lnb;
 323         unsigned short sector_size = blk_integrity_interval(bi);
 324         void *bio_prot_buf = page_address(bip->bip_vec->bv_page) +
 325                 bip->bip_vec->bv_offset;
 326         struct bio_vec *bv;
 327         sector_t sector = bio_start_sector(bio);
 328         unsigned int sectors, total;
 329         DECLARE_BVEC_ITER_ALL(iter_all);
 330         __u16 *expected_guard;
 331         int rc;
 332
 333         total = 0;
 334         bio_for_each_segment_all(bv, bio, iter_all) {
 335                 lnb = iobuf->dr_lnbs[index];
 336                 expected_guard = lnb->lnb_guards;
 337                 sectors = bv->bv_len / sector_size;
 338                 if (lnb->lnb_guard_rpc) {
 339                         rc = bio_dif_compare(expected_guard, bio_prot_buf,
 340                                              sectors, bi->tuple_size);
 341                         if (rc)
 342                                 return rc;
 343                 }
 344
 345                 sector += sectors;
 346                 bio_prot_buf += sectors * bi->tuple_size;
 347                 total += sectors * bi->tuple_size;
 348                 LASSERT(total <= bip_size(bio->bi_integrity));
 349                 index++;
 350         }
 351         return 0;
 352 }
 353
 354 static int osd_bio_integrity_handle(struct osd_device *osd, struct bio *bio,
 355                                     struct osd_iobuf *iobuf,
 356                                     int start_page_idx, bool fault_inject,
 357                                     bool integrity_enabled)
 358 {
 359         struct super_block *sb = osd_sb(osd);
 360         integrity_gen_fn *generate_fn = NULL;
 361         integrity_vrfy_fn *verify_fn = NULL;
 362         int rc;
 363
 364         ENTRY;
 365
 366         if (!integrity_enabled)
 367                 RETURN(0);
 368
 369         rc = osd_get_integrity_profile(osd, &generate_fn, &verify_fn);
 370         if (rc)
 371                 RETURN(rc);
 372
 373         rc = bio_integrity_prep_fn(bio, generate_fn, verify_fn);
 374         if (rc)
 375                 RETURN(rc);
 376
 377         /* Verify and inject fault only when writing */
 378         if (iobuf->dr_rw == 1) {
 379                 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_OST_INTEGRITY_CMP))) {
 380                         rc = osd_bio_integrity_compare(bio, sb->s_bdev, iobuf,
 381                                                        start_page_idx);
 382                         if (rc)
 383                                 RETURN(rc);
 384                 }
 385
 386                 if (unlikely(fault_inject))
 387                         bio_integrity_fault_inject(bio);
 388         }
 389
 390         RETURN(0);
 391 }
 392
 393 #ifdef HAVE_BIO_INTEGRITY_PREP_FN
 394 #  ifdef HAVE_BIO_ENDIO_USES_ONE_ARG
 395 static void dio_integrity_complete_routine(struct bio *bio)
 396 #  else
 397 static void dio_integrity_complete_routine(struct bio *bio, int error)
 398 #  endif
 399 {
 400         struct osd_bio_private *bio_private = bio->bi_private;
 401
 402         bio->bi_private = bio_private->obp_iobuf;
 403         osd_dio_complete_routine(bio, error);
 404
 405         OBD_FREE_PTR(bio_private);
 406 }
 407 #endif /* HAVE_BIO_INTEGRITY_PREP_FN */
 408 #else  /* !CONFIG_BLK_DEV_INTEGRITY */
 409 #define osd_bio_integrity_handle(osd, bio, iobuf, start_page_idx, \
 410                                  fault_inject, integrity_enabled) 0
 411 #endif /* CONFIG_BLK_DEV_INTEGRITY */
 412
 413 static int osd_bio_init(struct bio *bio, struct osd_iobuf *iobuf,
 414                         bool integrity_enabled, int start_page_idx,
 415                         struct osd_bio_private **pprivate)
 416 {
 417         ENTRY;
 418
 419         *pprivate = NULL;
 420
 421 #ifdef HAVE_BIO_INTEGRITY_PREP_FN
 422         if (integrity_enabled) {
 423                 struct osd_bio_private *bio_private = NULL;
 424
 425                 OBD_ALLOC_GFP(bio_private, sizeof(*bio_private), GFP_NOIO);
 426                 if (bio_private == NULL)
 427                         RETURN(-ENOMEM);
 428                 bio->bi_end_io = dio_integrity_complete_routine;
 429                 bio->bi_private = bio_private;
 430                 bio_private->obp_start_page_idx = start_page_idx;
 431                 bio_private->obp_iobuf = iobuf;
 432                 *pprivate = bio_private;
 433         } else
 434 #endif
 435         {
 436                 bio->bi_end_io = dio_complete_routine;
 437                 bio->bi_private = iobuf;
 438         }
 439
 440         RETURN(0);
 441 }
 442
 443 static int osd_do_bio(struct osd_device *osd, struct inode *inode,
 444                       struct osd_iobuf *iobuf, sector_t start_blocks,
 445                       sector_t count)
 446 {
 447         int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
 448         struct page **pages = iobuf->dr_pages;
 449         int npages = iobuf->dr_npages;
 450         sector_t *blocks = iobuf->dr_blocks;
 451         struct super_block *sb = inode->i_sb;
 452         int sector_bits = sb->s_blocksize_bits - 9;
 453         unsigned int blocksize = sb->s_blocksize;
 454         struct block_device *bdev = sb->s_bdev;
 455         struct osd_bio_private *bio_private = NULL;
 456         struct bio *bio = NULL;
 457         int bio_start_page_idx;
 458         struct page *page;
 459         unsigned int page_offset;
 460         sector_t sector;
 461         int nblocks;
 462         int block_idx, block_idx_end;
 463         int page_idx, page_idx_start;
 464         int i;
 465         int rc = 0;
 466         bool fault_inject;
 467         bool integrity_enabled;
 468         struct blk_plug plug;
 469         int blocks_left_page;
 470
 471         ENTRY;
 472
 473         fault_inject = OBD_FAIL_CHECK(OBD_FAIL_OST_INTEGRITY_FAULT);
 474         LASSERT(iobuf->dr_npages == npages);
 475
 476         integrity_enabled = bdev_integrity_enabled(bdev, iobuf->dr_rw);
 477
 478         osd_brw_stats_update(osd, iobuf);
 479         iobuf->dr_start_time = ktime_get();
 480
 481         if (!count)
 482                 count = npages * blocks_per_page;
 483         block_idx_end = start_blocks + count;
 484
 485         blk_start_plug(&plug);
 486
 487         page_idx_start = start_blocks / blocks_per_page;
 488         for (page_idx = page_idx_start, block_idx = start_blocks;
 489              block_idx < block_idx_end; page_idx++,
 490              block_idx += blocks_left_page) {
 491                 page = pages[page_idx];
 492                 LASSERT(page_idx < iobuf->dr_npages);
 493
 494                 i = block_idx % blocks_per_page;
 495                 blocks_left_page = blocks_per_page - i;
 496                 for (page_offset = i * blocksize; i < blocks_left_page;
 497                      i += nblocks, page_offset += blocksize * nblocks) {
 498                         nblocks = 1;
 499
 500                         if (blocks[block_idx + i] == 0) {  /* hole */
 501                                 LASSERTF(iobuf->dr_rw == 0,
 502                                          "page_idx %u, block_idx %u, i %u,"
 503                                          "start_blocks: %llu, count: %llu, npages: %d\n",
 504                                          page_idx, block_idx, i,
 505                                          (unsigned long long)start_blocks,
 506                                          (unsigned long long)count, npages);
 507                                 memset(kmap(page) + page_offset, 0, blocksize);
 508                                 kunmap(page);
 509                                 continue;
 510                         }
 511
 512                         sector = (sector_t)blocks[block_idx + i] << sector_bits;
 513
 514                         /* Additional contiguous file blocks? */
 515                         while (i + nblocks < blocks_left_page &&
 516                                (sector + (nblocks << sector_bits)) ==
 517                                ((sector_t)blocks[block_idx + i + nblocks] <<
 518                                  sector_bits))
 519                                 nblocks++;
 520
 521                         if (bio && can_be_merged(bio, sector) &&
 522                             bio_add_page(bio, page, blocksize * nblocks,
 523                                          page_offset) != 0)
 524                                 continue;       /* added this frag OK */
 525
 526                         if (bio != NULL) {
 527                                 struct request_queue *q = bio_get_queue(bio);
 528                                 unsigned int bi_size = bio_sectors(bio) << 9;
 529
 530                                 /* Dang! I have to fragment this I/O */
 531                                 CDEBUG(D_INODE,
 532                                        "bio++ sz %d vcnt %d(%d) sectors %d(%d) psg %d(%d)\n",
 533                                        bi_size, bio->bi_vcnt, bio->bi_max_vecs,
 534                                        bio_sectors(bio),
 535                                        queue_max_sectors(q),
 536                                        osd_bio_nr_segs(bio),
 537                                        queue_max_segments(q));
 538                                 rc = osd_bio_integrity_handle(osd, bio,
 539                                         iobuf, bio_start_page_idx,
 540                                         fault_inject, integrity_enabled);
 541                                 if (rc) {
 542                                         bio_put(bio);
 543                                         goto out;
 544                                 }
 545
 546                                 record_start_io(iobuf, bi_size);
 547                                 osd_submit_bio(iobuf->dr_rw, bio);
 548                         }
 549
 550                         bio_start_page_idx = page_idx;
 551                         /* allocate new bio */
 552                         bio = bio_alloc(GFP_NOIO, min(BIO_MAX_PAGES,
 553                                         (block_idx_end - block_idx +
 554                                          blocks_left_page - 1)));
 555                         if (bio == NULL) {
 556                                 CERROR("Can't allocate bio %u pages\n",
 557                                        block_idx_end - block_idx +
 558                                        blocks_left_page - 1);
 559                                 rc = -ENOMEM;
 560                                 goto out;
 561                         }
 562
 563                         bio_set_dev(bio, bdev);
 564                         bio_set_sector(bio, sector);
 565                         bio->bi_opf = iobuf->dr_rw ? WRITE : READ;
 566                         rc = osd_bio_init(bio, iobuf, integrity_enabled,
 567                                           bio_start_page_idx, &bio_private);
 568                         if (rc) {
 569                                 bio_put(bio);
 570                                 goto out;
 571                         }
 572
 573                         rc = bio_add_page(bio, page,
 574                                           blocksize * nblocks, page_offset);
 575                         LASSERT(rc != 0);
 576                 }
 577         }
 578
 579         if (bio != NULL) {
 580                 rc = osd_bio_integrity_handle(osd, bio, iobuf,
 581                                               bio_start_page_idx,
 582                                               fault_inject,
 583                                               integrity_enabled);
 584                 if (rc) {
 585                         bio_put(bio);
 586                         goto out;
 587                 }
 588
 589                 record_start_io(iobuf, bio_sectors(bio) << 9);
 590                 osd_submit_bio(iobuf->dr_rw, bio);
 591                 rc = 0;
 592         }
 593
 594 out:
 595         blk_finish_plug(&plug);
 596
 597         /* in order to achieve better IO throughput, we don't wait for writes
 598          * completion here. instead we proceed with transaction commit in
 599          * parallel and wait for IO completion once transaction is stopped
 600          * see osd_trans_stop() for more details -bzzz
 601          */
 602         if (iobuf->dr_rw == 0 || fault_inject) {
 603                 wait_event(iobuf->dr_wait,
 604                            atomic_read(&iobuf->dr_numreqs) == 0);
 605                 osd_fini_iobuf(osd, iobuf);
 606         }
 607
 608         if (rc == 0) {
 609                 rc = iobuf->dr_error;
 610         } else {
 611                 if (bio_private)
 612                         OBD_FREE_PTR(bio_private);
 613         }
 614
 615         RETURN(rc);
 616 }
 617
 618 static int osd_map_remote_to_local(loff_t offset, ssize_t len, int *nrpages,
 619                                    struct niobuf_local *lnb, int maxlnb)
 620 {
 621         int rc = 0;
 622         ENTRY;
 623
 624         *nrpages = 0;
 625
 626         while (len > 0) {
 627                 int poff = offset & (PAGE_SIZE - 1);
 628                 int plen = PAGE_SIZE - poff;
 629
 630                 if (*nrpages >= maxlnb) {
 631                         rc = -EOVERFLOW;
 632                         break;
 633                 }
 634
 635                 if (plen > len)
 636                         plen = len;
 637                 lnb->lnb_file_offset = offset;
 638                 lnb->lnb_page_offset = poff;
 639                 lnb->lnb_len = plen;
 640                 /* lnb->lnb_flags = rnb->rnb_flags; */
 641                 lnb->lnb_flags = 0;
 642                 lnb->lnb_page = NULL;
 643                 lnb->lnb_rc = 0;
 644                 lnb->lnb_guard_rpc = 0;
 645                 lnb->lnb_guard_disk = 0;
 646                 lnb->lnb_locked = 0;
 647
 648                 LASSERTF(plen <= len, "plen %u, len %lld\n", plen,
 649                          (long long) len);
 650                 offset += plen;
 651                 len -= plen;
 652                 lnb++;
 653                 (*nrpages)++;
 654         }
 655
 656         RETURN(rc);
 657 }
 658
 659 static struct page *osd_get_page(const struct lu_env *env, struct dt_object *dt,
 660                                  loff_t offset, gfp_t gfp_mask, bool cache)
 661 {
 662         struct osd_thread_info *oti = osd_oti_get(env);
 663         struct inode *inode = osd_dt_obj(dt)->oo_inode;
 664         struct osd_device *d = osd_obj2dev(osd_dt_obj(dt));
 665         struct page *page;
 666         int cur;
 667
 668         LASSERT(inode);
 669
 670         if (cache) {
 671                 page = find_or_create_page(inode->i_mapping,
 672                                            offset >> PAGE_SHIFT, gfp_mask);
 673
 674                 if (likely(page)) {
 675                         LASSERT(!PagePrivate2(page));
 676                         wait_on_page_writeback(page);
 677                 } else {
 678                         lprocfs_counter_add(d->od_stats, LPROC_OSD_NO_PAGE, 1);
 679                 }
 680
 681                 return page;
 682         }
 683
 684         if (inode->i_mapping->nrpages) {
 685                 /* consult with pagecache, but do not create new pages */
 686                 /* this is normally used once */
 687                 page = find_lock_page(inode->i_mapping, offset >> PAGE_SHIFT);
 688                 if (page) {
 689                         wait_on_page_writeback(page);
 690                         return page;
 691                 }
 692         }
 693
 694         LASSERT(oti->oti_dio_pages);
 695         cur = oti->oti_dio_pages_used;
 696         page = oti->oti_dio_pages[cur];
 697
 698         if (unlikely(!page)) {
 699                 LASSERT(cur < PTLRPC_MAX_BRW_PAGES);
 700                 page = alloc_page(gfp_mask);
 701                 if (!page)
 702                         return NULL;
 703                 oti->oti_dio_pages[cur] = page;
 704                 SetPagePrivate2(page);
 705                 lock_page(page);
 706         }
 707
 708         ClearPageUptodate(page);
 709         page->index = offset >> PAGE_SHIFT;
 710         oti->oti_dio_pages_used++;
 711
 712         return page;
 713 }
 714
 715 /*
 716  * there are following "locks":
 717  * journal_start
 718  * i_mutex
 719  * page lock
 720  *
 721  * osd write path:
 722  *  - lock page(s)
 723  *  - journal_start
 724  *  - truncate_sem
 725  *
 726  * ext4 vmtruncate:
 727  *  - lock pages, unlock
 728  *  - journal_start
 729  *  - lock partial page
 730  *  - i_data_sem
 731  *
 732  */
 733
 734 /**
 735  * Unlock and release pages loaded by osd_bufs_get()
 736  *
 737  * Unlock \a npages pages from \a lnb and drop the refcount on them.
 738  *
 739  * \param env           thread execution environment
 740  * \param dt            dt object undergoing IO (OSD object + methods)
 741  * \param lnb           array of pages undergoing IO
 742  * \param npages        number of pages in \a lnb
 743  *
 744  * \retval 0            always
 745  */
 746 static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt,
 747                         struct niobuf_local *lnb, int npages)
 748 {
 749         struct osd_thread_info *oti = osd_oti_get(env);
 750         struct pagevec pvec;
 751         int i;
 752
 753         ll_pagevec_init(&pvec, 0);
 754
 755         for (i = 0; i < npages; i++) {
 756                 struct page *page = lnb[i].lnb_page;
 757
 758                 if (page == NULL)
 759                         continue;
 760
 761                 /* if the page isn't cached, then reset uptodate
 762                  * to prevent reuse
 763                  */
 764                 if (PagePrivate2(page)) {
 765                         oti->oti_dio_pages_used--;
 766                 } else {
 767                         if (lnb[i].lnb_locked)
 768                                 unlock_page(page);
 769                         if (pagevec_add(&pvec, page) == 0)
 770                                 pagevec_release(&pvec);
 771                 }
 772
 773                 lnb[i].lnb_page = NULL;
 774         }
 775
 776         LASSERTF(oti->oti_dio_pages_used == 0, "%d\n", oti->oti_dio_pages_used);
 777
 778         /* Release any partial pagevec */
 779         pagevec_release(&pvec);
 780
 781         RETURN(0);
 782 }
 783
 784 /**
 785  * Load and lock pages undergoing IO
 786  *
 787  * Pages as described in the \a lnb array are fetched (from disk or cache)
 788  * and locked for IO by the caller.
 789  *
 790  * DLM locking protects us from write and truncate competing for same region,
 791  * but partial-page truncate can leave dirty pages in the cache for ldiskfs.
 792  * It's possible the writeout on a such a page is in progress when we access
 793  * it. It's also possible that during this writeout we put new (partial) data
 794  * into the page, but won't be able to proceed in filter_commitrw_write().
 795  * Therefore, just wait for writeout completion as it should be rare enough.
 796  *
 797  * \param env           thread execution environment
 798  * \param dt            dt object undergoing IO (OSD object + methods)
 799  * \param pos           byte offset of IO start
 800  * \param len           number of bytes of IO
 801  * \param lnb           array of extents undergoing IO
 802  * \param rw            read or write operation, and other flags
 803  * \param capa          capabilities
 804  *
 805  * \retval pages        (zero or more) loaded successfully
 806  * \retval -ENOMEM      on memory/page allocation error
 807  */
 808 static int osd_bufs_get(const struct lu_env *env, struct dt_object *dt,
 809                         loff_t pos, ssize_t len, struct niobuf_local *lnb,
 810                         int maxlnb, enum dt_bufs_type rw)
 811 {
 812         struct osd_thread_info *oti = osd_oti_get(env);
 813         struct osd_object *obj = osd_dt_obj(dt);
 814         struct osd_device *osd   = osd_obj2dev(obj);
 815         int npages, i, iosize, rc = 0;
 816         bool cache, write;
 817         loff_t fsize;
 818         gfp_t gfp_mask;
 819
 820         LASSERT(obj->oo_inode);
 821
 822         rc = osd_map_remote_to_local(pos, len, &npages, lnb, maxlnb);
 823         if (rc)
 824                 RETURN(rc);
 825
 826         write = rw & DT_BUFS_TYPE_WRITE;
 827
 828         fsize = lnb[npages - 1].lnb_file_offset + lnb[npages - 1].lnb_len;
 829         iosize = fsize - lnb[0].lnb_file_offset;
 830         fsize = max(fsize, i_size_read(obj->oo_inode));
 831
 832         cache = rw & DT_BUFS_TYPE_READAHEAD;
 833         if (cache)
 834                 goto bypass_checks;
 835
 836         cache = osd_use_page_cache(osd);
 837         while (cache) {
 838                 if (write) {
 839                         if (!osd->od_writethrough_cache) {
 840                                 cache = false;
 841                                 break;
 842                         }
 843                         if (iosize > osd->od_writethrough_max_iosize) {
 844                                 cache = false;
 845                                 break;
 846                         }
 847                 } else {
 848                         if (!osd->od_read_cache) {
 849                                 cache = false;
 850                                 break;
 851                         }
 852                         if (iosize > osd->od_readcache_max_iosize) {
 853                                 cache = false;
 854                                 break;
 855                         }
 856                 }
 857                 /* don't use cache on large files */
 858                 if (osd->od_readcache_max_filesize &&
 859                     fsize > osd->od_readcache_max_filesize)
 860                         cache = false;
 861                 break;
 862         }
 863
 864 bypass_checks:
 865         if (!cache && unlikely(!oti->oti_dio_pages)) {
 866                 OBD_ALLOC_PTR_ARRAY_LARGE(oti->oti_dio_pages,
 867                                           PTLRPC_MAX_BRW_PAGES);
 868                 if (!oti->oti_dio_pages)
 869                         return -ENOMEM;
 870         }
 871
 872         /* this could also try less hard for DT_BUFS_TYPE_READAHEAD pages */
 873         gfp_mask = rw & DT_BUFS_TYPE_LOCAL ? (GFP_NOFS | __GFP_HIGHMEM) :
 874                                              GFP_HIGHUSER;
 875         for (i = 0; i < npages; i++, lnb++) {
 876                 lnb->lnb_page = osd_get_page(env, dt, lnb->lnb_file_offset,
 877                                              gfp_mask, cache);
 878                 if (lnb->lnb_page == NULL)
 879                         GOTO(cleanup, rc = -ENOMEM);
 880
 881                 lnb->lnb_locked = 1;
 882         }
 883
 884 #if 0
 885         /* XXX: this version doesn't invalidate cached pages, but use them */
 886         if (!cache && write && obj->oo_inode->i_mapping->nrpages) {
 887                 /* do not allow data aliasing, invalidate pagecache */
 888                 /* XXX: can be quite expensive in mixed case */
 889                 invalidate_mapping_pages(obj->oo_inode->i_mapping,
 890                                 lnb[0].lnb_file_offset >> PAGE_SHIFT,
 891                                 lnb[npages - 1].lnb_file_offset >> PAGE_SHIFT);
 892         }
 893 #endif
 894
 895         RETURN(i);
 896
 897 cleanup:
 898         if (i > 0)
 899                 osd_bufs_put(env, dt, lnb - i, i);
 900         return rc;
 901 }
 902 /* Borrow @ext4_chunk_trans_blocks */
 903 static int osd_chunk_trans_blocks(struct inode *inode, int nrblocks)
 904 {
 905         ldiskfs_group_t groups;
 906         int gdpblocks;
 907         int idxblocks;
 908         int depth;
 909         int ret;
 910
 911         depth = ext_depth(inode);
 912         idxblocks = depth * 2;
 913
 914         /*
 915          * Now let's see how many group bitmaps and group descriptors need
 916          * to account.
 917          */
 918         groups = idxblocks + 1;
 919         gdpblocks = groups;
 920         if (groups > LDISKFS_SB(inode->i_sb)->s_groups_count)
 921                 groups = LDISKFS_SB(inode->i_sb)->s_groups_count;
 922         if (gdpblocks > LDISKFS_SB(inode->i_sb)->s_gdb_count)
 923                 gdpblocks = LDISKFS_SB(inode->i_sb)->s_gdb_count;
 924
 925         /* bitmaps and block group descriptor blocks */
 926         ret = idxblocks + groups + gdpblocks;
 927
 928         /* Blocks for super block, inode, quota and xattr blocks */
 929         ret += LDISKFS_META_TRANS_BLOCKS(inode->i_sb);
 930
 931         return ret;
 932 }
 933
 934 static int osd_extend_trans(handle_t *handle, int needed)
 935 {
 936         if (ldiskfs_handle_has_enough_credits(handle, needed))
 937                 return 0;
 938
 939         return ldiskfs_journal_extend(handle,
 940                                       needed - handle->h_buffer_credits);
 941 }
 942
 943 static int osd_extend_restart_trans(handle_t *handle, int needed)
 944 {
 945
 946         int rc = osd_extend_trans(handle, needed);
 947
 948         if (rc <= 0)
 949                 return rc;
 950
 951         return ldiskfs_journal_restart(handle, needed);
 952 }
 953
 954 static int osd_ldiskfs_map_write(struct inode *inode, struct osd_iobuf *iobuf,
 955                                  struct osd_device *osd, sector_t start_blocks,
 956                                  sector_t count, loff_t *disk_size,
 957                                  __u64 user_size)
 958 {
 959         /* if file has grown, take user_size into account */
 960         if (user_size && *disk_size > user_size)
 961                 *disk_size = user_size;
 962
 963         spin_lock(&inode->i_lock);
 964         if (*disk_size > i_size_read(inode)) {
 965                 i_size_write(inode, *disk_size);
 966                 LDISKFS_I(inode)->i_disksize = *disk_size;
 967                 spin_unlock(&inode->i_lock);
 968                 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
 969         } else {
 970                 spin_unlock(&inode->i_lock);
 971         }
 972
 973         /*
 974          * We don't do stats here as in read path because
 975          * write is async: we'll do this in osd_put_bufs()
 976          */
 977         return osd_do_bio(osd, inode, iobuf, start_blocks, count);
 978 }
 979
 980
 981 static int osd_ldiskfs_map_inode_pages(struct inode *inode,
 982                                        struct osd_iobuf *iobuf,
 983                                        struct osd_device *osd,
 984                                        int create, __u64 user_size,
 985                                        int check_credits)
 986 {
 987         int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
 988         int rc = 0, i = 0, mapped_index = 0;
 989         struct page *fp = NULL;
 990         int clen = 0;
 991         pgoff_t max_page_index;
 992         handle_t *handle = NULL;
 993         int credits;
 994         sector_t start_blocks = 0, count = 0;
 995         loff_t disk_size = 0;
 996         struct page **page = iobuf->dr_pages;
 997         int pages = iobuf->dr_npages;
 998         sector_t *blocks = iobuf->dr_blocks;
 999         struct niobuf_local *lnb1, *lnb2;
1000         loff_t size1, size2;
1001
1002         max_page_index = inode->i_sb->s_maxbytes >> PAGE_SHIFT;
1003
1004         CDEBUG(D_OTHER, "inode %lu: map %d pages from %lu\n",
1005                 inode->i_ino, pages, (*page)->index);
1006
1007         if (create) {
1008                 create = LDISKFS_GET_BLOCKS_CREATE;
1009                 handle = ldiskfs_journal_current_handle();
1010                 LASSERT(handle != NULL);
1011                 rc = osd_attach_jinode(inode);
1012                 if (rc)
1013                         return rc;
1014                 disk_size = i_size_read(inode);
1015                 /* if disk_size is already bigger than specified user_size,
1016                  * ignore user_size
1017                  */
1018                 if (disk_size > user_size)
1019                         user_size = 0;
1020         }
1021         /* pages are sorted already. so, we just have to find
1022          * contig. space and process them properly
1023          */
1024         while (i < pages) {
1025                 long blen, total = 0, previous_total = 0;
1026                 struct ldiskfs_map_blocks map = { 0 };
1027
1028                 if (fp == NULL) { /* start new extent */
1029                         fp = *page++;
1030                         clen = 1;
1031                         if (++i != pages)
1032                                 continue;
1033                 } else if (fp->index + clen == (*page)->index) {
1034                         /* continue the extent */
1035                         page++;
1036                         clen++;
1037                         if (++i != pages)
1038                                 continue;
1039                 }
1040                 if (fp->index + clen >= max_page_index)
1041                         GOTO(cleanup, rc = -EFBIG);
1042                 /* process found extent */
1043                 map.m_lblk = fp->index * blocks_per_page;
1044                 map.m_len = blen = clen * blocks_per_page;
1045 cont_map:
1046                 /**
1047                  * We might restart transaction for block allocations,
1048                  * in order to make sure data ordered mode, issue IO, disk
1049                  * size update and block allocations need be within same
1050                  * transaction to make sure consistency.
1051                  */
1052                 if (handle && check_credits) {
1053                         /*
1054                          * credits to insert 1 extent into extent tree.
1055                          */
1056                         credits = osd_chunk_trans_blocks(inode, blen);
1057                         rc = osd_extend_trans(handle, credits);
1058                         if (rc < 0)
1059                                 GOTO(cleanup, rc);
1060                         /*
1061                          * only issue IO if restart transaction needed,
1062                          * as update disk size need hold inode lock, we
1063                          * want to avoid that as much as possible.
1064                          */
1065                         if (rc > 0) {
1066                                 WARN_ON_ONCE(start_blocks == 0);
1067                                 rc = osd_ldiskfs_map_write(inode,
1068                                         iobuf, osd, start_blocks,
1069                                         count, &disk_size, user_size);
1070                                 if (rc)
1071                                         GOTO(cleanup, rc);
1072                                 rc = ldiskfs_journal_restart(handle, credits);
1073                                 if (rc)
1074                                         GOTO(cleanup, rc);
1075                                 start_blocks += count;
1076                                 /* reset IO block count */
1077                                 count = 0;
1078                         }
1079                 }
1080                 rc = ldiskfs_map_blocks(handle, inode, &map, create);
1081                 if (rc >= 0) {
1082                         int c = 0;
1083
1084                         for (; total < blen && c < map.m_len; c++, total++) {
1085                                 if (rc == 0) {
1086                                         *(blocks + total) = 0;
1087                                         total++;
1088                                         break;
1089                                 }
1090                                 *(blocks + total) = map.m_pblk + c;
1091                                 /* unmap any possible underlying
1092                                  * metadata from the block device
1093                                  * mapping.  b=6998.
1094                                  */
1095                                 if ((map.m_flags & LDISKFS_MAP_NEW) &&
1096                                     create)
1097                                         clean_bdev_aliases(inode->i_sb->s_bdev,
1098                                                            map.m_pblk + c, 1);
1099                         }
1100                         rc = 0;
1101                 }
1102
1103                 if (rc == 0 && create) {
1104                         count += (total - previous_total);
1105                         mapped_index = (count + blocks_per_page -
1106                                         1) / blocks_per_page - 1;
1107                         lnb1 = iobuf->dr_lnbs[i - clen];
1108                         lnb2 = iobuf->dr_lnbs[mapped_index];
1109                         size1 = lnb1->lnb_file_offset -
1110                                 (lnb1->lnb_file_offset % PAGE_SIZE) +
1111                                 (total << inode->i_blkbits);
1112                         size2 = lnb2->lnb_file_offset + lnb2->lnb_len;
1113
1114                         if (size1 > size2)
1115                                 size1 = size2;
1116                         if (size1 > disk_size)
1117                                 disk_size = size1;
1118                 }
1119
1120                 if (rc == 0 && total < blen) {
1121                         map.m_lblk = fp->index * blocks_per_page + total;
1122                         map.m_len = blen - total;
1123                         previous_total = total;
1124                         goto cont_map;
1125                 }
1126                 if (rc != 0)
1127                         GOTO(cleanup, rc);
1128
1129                 /* look for next extent */
1130                 fp = NULL;
1131                 blocks += blocks_per_page * clen;
1132         }
1133 cleanup:
1134         if (rc == 0 && create &&
1135             start_blocks < pages * blocks_per_page) {
1136                 rc = osd_ldiskfs_map_write(inode, iobuf, osd, start_blocks,
1137                                            count, &disk_size, user_size);
1138                 LASSERT(start_blocks + count == pages * blocks_per_page);
1139         }
1140         return rc;
1141 }
1142
1143 static int osd_write_prep(const struct lu_env *env, struct dt_object *dt,
1144                           struct niobuf_local *lnb, int npages)
1145 {
1146         struct osd_thread_info *oti   = osd_oti_get(env);
1147         struct osd_iobuf       *iobuf = &oti->oti_iobuf;
1148         struct inode           *inode = osd_dt_obj(dt)->oo_inode;
1149         struct osd_device      *osd   = osd_obj2dev(osd_dt_obj(dt));
1150         ktime_t start, end;
1151         s64 timediff;
1152         ssize_t isize;
1153         __s64  maxidx;
1154         int i, rc = 0;
1155
1156         LASSERT(inode);
1157
1158         rc = osd_init_iobuf(osd, iobuf, 0, npages);
1159         if (unlikely(rc != 0))
1160                 RETURN(rc);
1161
1162         isize = i_size_read(inode);
1163         maxidx = ((isize + PAGE_SIZE - 1) >> PAGE_SHIFT) - 1;
1164
1165         start = ktime_get();
1166         for (i = 0; i < npages; i++) {
1167
1168                 /*
1169                  * till commit the content of the page is undefined
1170                  * we'll set it uptodate once bulk is done. otherwise
1171                  * subsequent reads can access non-stable data
1172                  */
1173                 ClearPageUptodate(lnb[i].lnb_page);
1174
1175                 if (lnb[i].lnb_len == PAGE_SIZE)
1176                         continue;
1177
1178                 if (maxidx >= lnb[i].lnb_page->index) {
1179                         osd_iobuf_add_page(iobuf, &lnb[i]);
1180                 } else {
1181                         long off;
1182                         char *p = kmap(lnb[i].lnb_page);
1183
1184                         off = lnb[i].lnb_page_offset;
1185                         if (off)
1186                                 memset(p, 0, off);
1187                         off = (lnb[i].lnb_page_offset + lnb[i].lnb_len) &
1188                               ~PAGE_MASK;
1189                         if (off)
1190                                 memset(p + off, 0, PAGE_SIZE - off);
1191                         kunmap(lnb[i].lnb_page);
1192                 }
1193         }
1194         end = ktime_get();
1195         timediff = ktime_us_delta(end, start);
1196         lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
1197
1198         if (iobuf->dr_npages) {
1199                 rc = osd_ldiskfs_map_inode_pages(inode, iobuf, osd, 0,
1200                                                  0, 0);
1201                 if (likely(rc == 0)) {
1202                         rc = osd_do_bio(osd, inode, iobuf, 0, 0);
1203                         /* do IO stats for preparation reads */
1204                         osd_fini_iobuf(osd, iobuf);
1205                 }
1206         }
1207         RETURN(rc);
1208 }
1209
1210 struct osd_fextent {
1211         sector_t        start;
1212         sector_t        end;
1213         unsigned int    mapped:1;
1214 };
1215
1216 static int osd_is_mapped(struct dt_object *dt, __u64 offset,
1217                          struct osd_fextent *cached_extent)
1218 {
1219         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1220         sector_t block = offset >> inode->i_blkbits;
1221         sector_t start;
1222         struct fiemap_extent_info fei = { 0 };
1223         struct fiemap_extent fe = { 0 };
1224         mm_segment_t saved_fs;
1225         int rc;
1226
1227         if (block >= cached_extent->start && block < cached_extent->end)
1228                 return cached_extent->mapped;
1229
1230         if (i_size_read(inode) == 0)
1231                 return 0;
1232
1233         /* Beyond EOF, must not be mapped */
1234         if (((i_size_read(inode) - 1) >> inode->i_blkbits) < block)
1235                 return 0;
1236
1237         fei.fi_extents_max = 1;
1238         fei.fi_extents_start = &fe;
1239
1240         saved_fs = get_fs();
1241         set_fs(KERNEL_DS);
1242         rc = inode->i_op->fiemap(inode, &fei, offset, FIEMAP_MAX_OFFSET-offset);
1243         set_fs(saved_fs);
1244         if (rc != 0)
1245                 return 0;
1246
1247         start = fe.fe_logical >> inode->i_blkbits;
1248
1249         if (start > block) {
1250                 cached_extent->start = block;
1251                 cached_extent->end = start;
1252                 cached_extent->mapped = 0;
1253         } else {
1254                 cached_extent->start = start;
1255                 cached_extent->end = (fe.fe_logical + fe.fe_length) >>
1256                                       inode->i_blkbits;
1257                 cached_extent->mapped = 1;
1258         }
1259
1260         return cached_extent->mapped;
1261 }
1262
1263 static int osd_declare_write_commit(const struct lu_env *env,
1264                                     struct dt_object *dt,
1265                                     struct niobuf_local *lnb, int npages,
1266                                     struct thandle *handle)
1267 {
1268         const struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1269         struct inode            *inode = osd_dt_obj(dt)->oo_inode;
1270         struct osd_thandle      *oh;
1271         int                     extents = 0;
1272         int                     depth;
1273         int                     i;
1274         int                     newblocks = 0;
1275         int                     rc = 0;
1276         int                     credits = 0;
1277         long long               quota_space = 0;
1278         struct osd_fextent      mapped = { 0 }, extent = { 0 };
1279         enum osd_quota_local_flags local_flags = 0;
1280         enum osd_qid_declare_flags declare_flags = OSD_QID_BLK;
1281         ENTRY;
1282
1283         LASSERT(handle != NULL);
1284         oh = container_of(handle, struct osd_thandle, ot_super);
1285         LASSERT(oh->ot_handle == NULL);
1286
1287         /* calculate number of extents (probably better to pass nb) */
1288         for (i = 0; i < npages; i++) {
1289                 /* ignore quota for the whole request if any page is from
1290                  * client cache or written by root.
1291                  *
1292                  * XXX once we drop the 1.8 client support, the checking
1293                  * for whether page is from cache can be simplified as:
1294                  * !(lnb[i].flags & OBD_BRW_SYNC)
1295                  *
1296                  * XXX we could handle this on per-lnb basis as done by
1297                  * grant.
1298                  */
1299                 if ((lnb[i].lnb_flags & OBD_BRW_NOQUOTA) ||
1300                     (lnb[i].lnb_flags & (OBD_BRW_FROM_GRANT | OBD_BRW_SYNC)) ==
1301                     OBD_BRW_FROM_GRANT)
1302                         declare_flags |= OSD_QID_FORCE;
1303
1304                 if (osd_is_mapped(dt, lnb[i].lnb_file_offset, &mapped)) {
1305                         lnb[i].lnb_flags |= OBD_BRW_MAPPED;
1306                         continue;
1307                 }
1308
1309                 /* count only unmapped changes */
1310                 newblocks++;
1311                 if (lnb[i].lnb_file_offset != extent.end || extent.end == 0) {
1312                         extents++;
1313                         extent.end = lnb[i].lnb_file_offset + lnb[i].lnb_len;
1314                 } else {
1315                         extent.end += lnb[i].lnb_len;
1316                 }
1317
1318                 quota_space += PAGE_SIZE;
1319         }
1320
1321         credits++; /* inode */
1322         /*
1323          * overwrite case, no need to modify tree and
1324          * allocate blocks.
1325          */
1326         if (!newblocks)
1327                 goto out_declare;
1328         /*
1329          * each extent can go into new leaf causing a split
1330          * 5 is max tree depth: inode + 4 index blocks
1331          * with blockmaps, depth is 3 at most
1332          */
1333         if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL) {
1334                 /*
1335                  * many concurrent threads may grow tree by the time
1336                  * our transaction starts. so, consider 2 is a min depth
1337                  */
1338                 depth = ext_depth(inode);
1339                 depth = max(depth, 1) + 1;
1340                 newblocks += depth;
1341                 credits += depth * 2 * extents;
1342         } else {
1343                 depth = 3;
1344                 newblocks += depth;
1345                 credits += depth * extents;
1346         }
1347
1348         /*
1349          * try a bit more extents to avoid restart
1350          * as much as possible in normal case.
1351          */
1352         if (npages > 1 && extents)
1353                 extents <<= 1;
1354
1355         /* quota space for metadata blocks */
1356         quota_space += depth * extents * LDISKFS_BLOCK_SIZE(osd_sb(osd));
1357
1358         /* quota space should be reported in 1K blocks */
1359         quota_space = toqb(quota_space);
1360
1361         /* each new block can go in different group (bitmap + gd) */
1362
1363         /* we can't dirty more bitmap blocks than exist */
1364         if (extents > LDISKFS_SB(osd_sb(osd))->s_groups_count)
1365                 credits += LDISKFS_SB(osd_sb(osd))->s_groups_count;
1366         else
1367                 credits += extents;
1368
1369         /* we can't dirty more gd blocks than exist */
1370         if (extents > LDISKFS_SB(osd_sb(osd))->s_gdb_count)
1371                 credits += LDISKFS_SB(osd_sb(osd))->s_gdb_count;
1372         else
1373                 credits += extents;
1374
1375 out_declare:
1376         osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
1377
1378         /* make sure the over quota flags were not set */
1379         lnb[0].lnb_flags &= ~OBD_BRW_OVER_ALLQUOTA;
1380
1381         rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
1382                                    i_projid_read(inode), quota_space, oh,
1383                                    osd_dt_obj(dt), &local_flags, declare_flags);
1384
1385         /* we need only to store the overquota flags in the first lnb for
1386          * now, once we support multiple objects BRW, this code needs be
1387          * revised.
1388          */
1389         if (local_flags & QUOTA_FL_OVER_USRQUOTA)
1390                 lnb[0].lnb_flags |= OBD_BRW_OVER_USRQUOTA;
1391         if (local_flags & QUOTA_FL_OVER_GRPQUOTA)
1392                 lnb[0].lnb_flags |= OBD_BRW_OVER_GRPQUOTA;
1393         if (local_flags & QUOTA_FL_OVER_PRJQUOTA)
1394                 lnb[0].lnb_flags |= OBD_BRW_OVER_PRJQUOTA;
1395
1396         if (rc == 0)
1397                 rc = osd_trunc_lock(osd_dt_obj(dt), oh, true);
1398
1399         RETURN(rc);
1400 }
1401
1402 /* Check if a block is allocated or not */
1403 static int osd_write_commit(const struct lu_env *env, struct dt_object *dt,
1404                             struct niobuf_local *lnb, int npages,
1405                             struct thandle *thandle, __u64 user_size)
1406 {
1407         struct osd_thread_info *oti = osd_oti_get(env);
1408         struct osd_iobuf *iobuf = &oti->oti_iobuf;
1409         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1410         struct osd_device  *osd = osd_obj2dev(osd_dt_obj(dt));
1411         int rc = 0, i, check_credits = 0;
1412         struct osd_thandle *oh = container_of(thandle,
1413                                               struct osd_thandle, ot_super);
1414         unsigned int save_credits = oh->ot_credits;
1415
1416         LASSERT(inode);
1417
1418         rc = osd_init_iobuf(osd, iobuf, 1, npages);
1419         if (unlikely(rc != 0))
1420                 RETURN(rc);
1421
1422         dquot_initialize(inode);
1423
1424         for (i = 0; i < npages; i++) {
1425                 if (lnb[i].lnb_rc == -ENOSPC &&
1426                     (lnb[i].lnb_flags & OBD_BRW_MAPPED)) {
1427                         /* Allow the write to proceed if overwriting an
1428                          * existing block
1429                          */
1430                         lnb[i].lnb_rc = 0;
1431                 }
1432
1433                 if (lnb[i].lnb_rc) { /* ENOSPC, network RPC error, etc. */
1434                         CDEBUG(D_INODE, "Skipping [%d] == %d\n", i,
1435                                lnb[i].lnb_rc);
1436                         LASSERT(lnb[i].lnb_page);
1437                         generic_error_remove_page(inode->i_mapping,
1438                                                   lnb[i].lnb_page);
1439                         continue;
1440                 }
1441
1442                 if (!(lnb[i].lnb_flags & OBD_BRW_MAPPED))
1443                         check_credits = 1;
1444
1445                 LASSERT(PageLocked(lnb[i].lnb_page));
1446                 LASSERT(!PageWriteback(lnb[i].lnb_page));
1447
1448                 /*
1449                  * Since write and truncate are serialized by oo_sem, even
1450                  * partial-page truncate should not leave dirty pages in the
1451                  * page cache.
1452                  */
1453                 LASSERT(!PageDirty(lnb[i].lnb_page));
1454
1455                 SetPageUptodate(lnb[i].lnb_page);
1456
1457                 osd_iobuf_add_page(iobuf, &lnb[i]);
1458         }
1459
1460         osd_trans_exec_op(env, thandle, OSD_OT_WRITE);
1461
1462         if (OBD_FAIL_CHECK(OBD_FAIL_OST_MAPBLK_ENOSPC)) {
1463                 rc = -ENOSPC;
1464         } else if (iobuf->dr_npages > 0) {
1465                 rc = osd_ldiskfs_map_inode_pages(inode, iobuf, osd,
1466                                                  1, user_size,
1467                                                  check_credits);
1468                 /*
1469                  * Write might restart transaction, extend credits
1470                  * if needed for operations such as attribute set.
1471                  */
1472                 if (rc == 0) {
1473                         handle_t *handle = ldiskfs_journal_current_handle();
1474
1475                         LASSERT(handle != NULL);
1476                         rc = osd_extend_restart_trans(handle, save_credits);
1477                 }
1478         } else {
1479                 /* no pages to write, no transno is needed */
1480                 thandle->th_local = 1;
1481         }
1482
1483         if (rc != 0)
1484                 osd_fini_iobuf(osd, iobuf);
1485
1486         osd_trans_exec_check(env, thandle, OSD_OT_WRITE);
1487
1488         if (unlikely(rc != 0)) {
1489                 /* if write fails, we should drop pages from the cache */
1490                 for (i = 0; i < npages; i++) {
1491                         if (lnb[i].lnb_page == NULL)
1492                                 continue;
1493                         if (!PagePrivate2(lnb[i].lnb_page)) {
1494                                 LASSERT(PageLocked(lnb[i].lnb_page));
1495                                 generic_error_remove_page(inode->i_mapping,
1496                                                           lnb[i].lnb_page);
1497                         }
1498                 }
1499         }
1500
1501         RETURN(rc);
1502 }
1503
1504 static int osd_read_prep(const struct lu_env *env, struct dt_object *dt,
1505                          struct niobuf_local *lnb, int npages)
1506 {
1507         struct osd_thread_info *oti = osd_oti_get(env);
1508         struct osd_iobuf *iobuf = &oti->oti_iobuf;
1509         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1510         struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1511         int rc = 0, i, cache_hits = 0, cache_misses = 0;
1512         ktime_t start, end;
1513         s64 timediff;
1514         loff_t isize;
1515
1516         LASSERT(inode);
1517
1518         rc = osd_init_iobuf(osd, iobuf, 0, npages);
1519         if (unlikely(rc != 0))
1520                 RETURN(rc);
1521
1522         isize = i_size_read(inode);
1523
1524         start = ktime_get();
1525         for (i = 0; i < npages; i++) {
1526
1527                 if (isize <= lnb[i].lnb_file_offset)
1528                         /* If there's no more data, abort early.
1529                          * lnb->lnb_rc == 0, so it's easy to detect later.
1530                          */
1531                         break;
1532
1533                 /* instead of looking if we go beyong isize, send complete
1534                  * pages all the time
1535                  */
1536                 lnb[i].lnb_rc = lnb[i].lnb_len;
1537
1538                 /* Bypass disk read if fail_loc is set properly */
1539                 if (OBD_FAIL_CHECK(OBD_FAIL_OST_FAKE_RW))
1540                         SetPageUptodate(lnb[i].lnb_page);
1541
1542                 if (PageUptodate(lnb[i].lnb_page)) {
1543                         cache_hits++;
1544                         unlock_page(lnb[i].lnb_page);
1545                 } else {
1546                         cache_misses++;
1547                         osd_iobuf_add_page(iobuf, &lnb[i]);
1548                 }
1549                 /* no need to unlock in osd_bufs_put(), the sooner page is
1550                  * unlocked, the earlier another client can access it.
1551                  * notice real unlock_page() can be called few lines
1552                  * below after osd_do_bio(). lnb is a per-thread, so it's
1553                  * fine to have PG_locked and lnb_locked inconsistent here
1554                  */
1555                 lnb[i].lnb_locked = 0;
1556         }
1557         end = ktime_get();
1558         timediff = ktime_us_delta(end, start);
1559         lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
1560
1561         if (cache_hits != 0)
1562                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_HIT,
1563                                     cache_hits);
1564         if (cache_misses != 0)
1565                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_MISS,
1566                                     cache_misses);
1567         if (cache_hits + cache_misses != 0)
1568                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_ACCESS,
1569                                     cache_hits + cache_misses);
1570
1571         if (iobuf->dr_npages) {
1572                 rc = osd_ldiskfs_map_inode_pages(inode, iobuf, osd, 0,
1573                                                  0, 0);
1574                 if (!rc)
1575                         rc = osd_do_bio(osd, inode, iobuf, 0, 0);
1576
1577                 /* IO stats will be done in osd_bufs_put() */
1578
1579                 /* early release to let others read data during the bulk */
1580                 for (i = 0; i < iobuf->dr_npages; i++) {
1581                         LASSERT(PageLocked(iobuf->dr_pages[i]));
1582                         if (!PagePrivate2(iobuf->dr_pages[i]))
1583                                 unlock_page(iobuf->dr_pages[i]);
1584                 }
1585         }
1586
1587         RETURN(rc);
1588 }
1589
1590 /*
1591  * XXX: Another layering violation for now.
1592  *
1593  * We don't want to use ->f_op->read methods, because generic file write
1594  *
1595  *         - serializes on ->i_sem, and
1596  *
1597  *         - does a lot of extra work like balance_dirty_pages(),
1598  *
1599  * which doesn't work for globally shared files like /last_rcvd.
1600  */
1601 static int osd_ldiskfs_readlink(struct inode *inode, char *buffer, int buflen)
1602 {
1603         struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
1604
1605         memcpy(buffer, (char *)ei->i_data, buflen);
1606
1607         return  buflen;
1608 }
1609
1610 int osd_ldiskfs_read(struct inode *inode, void *buf, int size, loff_t *offs)
1611 {
1612         struct buffer_head *bh;
1613         unsigned long block;
1614         int osize;
1615         int blocksize;
1616         int csize;
1617         int boffs;
1618
1619         /* prevent reading after eof */
1620         spin_lock(&inode->i_lock);
1621         if (i_size_read(inode) < *offs + size) {
1622                 loff_t diff = i_size_read(inode) - *offs;
1623
1624                 spin_unlock(&inode->i_lock);
1625                 if (diff < 0) {
1626                         CDEBUG(D_OTHER,
1627                                "size %llu is too short to read @%llu\n",
1628                                i_size_read(inode), *offs);
1629                         return -EBADR;
1630                 } else if (diff == 0) {
1631                         return 0;
1632                 } else {
1633                         size = diff;
1634                 }
1635         } else {
1636                 spin_unlock(&inode->i_lock);
1637         }
1638
1639         blocksize = 1 << inode->i_blkbits;
1640         osize = size;
1641         while (size > 0) {
1642                 block = *offs >> inode->i_blkbits;
1643                 boffs = *offs & (blocksize - 1);
1644                 csize = min(blocksize - boffs, size);
1645                 bh = __ldiskfs_bread(NULL, inode, block, 0);
1646                 if (IS_ERR(bh)) {
1647                         CERROR("%s: can't read %u@%llu on ino %lu: rc = %ld\n",
1648                                osd_ino2name(inode), csize, *offs, inode->i_ino,
1649                                PTR_ERR(bh));
1650                         return PTR_ERR(bh);
1651                 }
1652
1653                 if (bh != NULL) {
1654                         memcpy(buf, bh->b_data + boffs, csize);
1655                         brelse(bh);
1656                 } else {
1657                         memset(buf, 0, csize);
1658                 }
1659
1660                 *offs += csize;
1661                 buf += csize;
1662                 size -= csize;
1663         }
1664         return osize;
1665 }
1666
1667 static ssize_t osd_read(const struct lu_env *env, struct dt_object *dt,
1668                         struct lu_buf *buf, loff_t *pos)
1669 {
1670         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1671         int rc;
1672
1673         /* Read small symlink from inode body as we need to maintain correct
1674          * on-disk symlinks for ldiskfs.
1675          */
1676         if (S_ISLNK(dt->do_lu.lo_header->loh_attr)) {
1677                 loff_t size = i_size_read(inode);
1678
1679                 if (buf->lb_len < size)
1680                         return -EOVERFLOW;
1681
1682                 if (size < sizeof(LDISKFS_I(inode)->i_data))
1683                         rc = osd_ldiskfs_readlink(inode, buf->lb_buf, size);
1684                 else
1685                         rc = osd_ldiskfs_read(inode, buf->lb_buf, size, pos);
1686         } else {
1687                 rc = osd_ldiskfs_read(inode, buf->lb_buf, buf->lb_len, pos);
1688         }
1689
1690         return rc;
1691 }
1692
1693 static inline int osd_extents_enabled(struct super_block *sb,
1694                                       struct inode *inode)
1695 {
1696         if (inode != NULL) {
1697                 if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL)
1698                         return 1;
1699         } else if (ldiskfs_has_feature_extents(sb)) {
1700                 return 1;
1701         }
1702         return 0;
1703 }
1704
1705 int osd_calc_bkmap_credits(struct super_block *sb, struct inode *inode,
1706                            const loff_t size, const loff_t pos,
1707                            const int blocks)
1708 {
1709         int credits, bits, bs, i;
1710
1711         bits = sb->s_blocksize_bits;
1712         bs = 1 << bits;
1713
1714         /* legacy blockmap: 3 levels * 3 (bitmap,gd,itself)
1715          * we do not expect blockmaps on the large files,
1716          * so let's shrink it to 2 levels (4GB files)
1717          */
1718
1719         /* this is default reservation: 2 levels */
1720         credits = (blocks + 2) * 3;
1721
1722         /* actual offset is unknown, hard to optimize */
1723         if (pos == -1)
1724                 return credits;
1725
1726         /* now check for few specific cases to optimize */
1727         if (pos + size <= LDISKFS_NDIR_BLOCKS * bs) {
1728                 /* no indirects */
1729                 credits = blocks;
1730                 /* allocate if not allocated */
1731                 if (inode == NULL) {
1732                         credits += blocks * 2;
1733                         return credits;
1734                 }
1735                 for (i = (pos >> bits); i < (pos >> bits) + blocks; i++) {
1736                         LASSERT(i < LDISKFS_NDIR_BLOCKS);
1737                         if (LDISKFS_I(inode)->i_data[i] == 0)
1738                                 credits += 2;
1739                 }
1740         } else if (pos + size <= (LDISKFS_NDIR_BLOCKS + 1024) * bs) {
1741                 /* single indirect */
1742                 credits = blocks * 3;
1743                 if (inode == NULL ||
1744                     LDISKFS_I(inode)->i_data[LDISKFS_IND_BLOCK] == 0)
1745                         credits += 3;
1746                 else
1747                         /* The indirect block may be modified. */
1748                         credits += 1;
1749         }
1750
1751         return credits;
1752 }
1753
1754 static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt,
1755                                  const struct lu_buf *buf, loff_t _pos,
1756                                  struct thandle *handle)
1757 {
1758         struct osd_object  *obj  = osd_dt_obj(dt);
1759         struct inode       *inode = obj->oo_inode;
1760         struct super_block *sb = osd_sb(osd_obj2dev(obj));
1761         struct osd_thandle *oh;
1762         int                 rc = 0, est = 0, credits, blocks, allocated = 0;
1763         int                 bits, bs;
1764         int                 depth, size;
1765         loff_t              pos;
1766         ENTRY;
1767
1768         LASSERT(buf != NULL);
1769         LASSERT(handle != NULL);
1770
1771         oh = container_of(handle, struct osd_thandle, ot_super);
1772         LASSERT(oh->ot_handle == NULL);
1773
1774         size = buf->lb_len;
1775         bits = sb->s_blocksize_bits;
1776         bs = 1 << bits;
1777
1778         if (_pos == -1) {
1779                 /* if this is an append, then we
1780                  * should expect cross-block record
1781                  */
1782                 pos = 0;
1783         } else {
1784                 pos = _pos;
1785         }
1786
1787         /* blocks to modify */
1788         blocks = ((pos + size + bs - 1) >> bits) - (pos >> bits);
1789         LASSERT(blocks > 0);
1790
1791         if (inode != NULL && _pos != -1) {
1792                 /* object size in blocks */
1793                 est = (i_size_read(inode) + bs - 1) >> bits;
1794                 allocated = inode->i_blocks >> (bits - 9);
1795                 if (pos + size <= i_size_read(inode) && est <= allocated) {
1796                         /* looks like an overwrite, no need to modify tree */
1797                         credits = blocks;
1798                         /* no need to modify i_size */
1799                         goto out;
1800                 }
1801         }
1802
1803         if (osd_extents_enabled(sb, inode)) {
1804                 /*
1805                  * many concurrent threads may grow tree by the time
1806                  * our transaction starts. so, consider 2 is a min depth
1807                  * for every level we may need to allocate a new block
1808                  * and take some entries from the old one. so, 3 blocks
1809                  * to allocate (bitmap, gd, itself) + old block - 4 per
1810                  * level.
1811                  */
1812                 depth = inode != NULL ? ext_depth(inode) : 0;
1813                 depth = max(depth, 1) + 1;
1814                 credits = depth;
1815                 /* if not append, then split may need to modify
1816                  * existing blocks moving entries into the new ones
1817                  */
1818                 if (_pos != -1)
1819                         credits += depth;
1820                 /* blocks to store data: bitmap,gd,itself */
1821                 credits += blocks * 3;
1822         } else {
1823                 credits = osd_calc_bkmap_credits(sb, inode, size, _pos, blocks);
1824         }
1825         /* if inode is created as part of the transaction,
1826          * then it's counted already by the creation method
1827          */
1828         if (inode != NULL)
1829                 credits++;
1830
1831 out:
1832
1833         osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
1834
1835         /* dt_declare_write() is usually called for system objects, such
1836          * as llog or last_rcvd files. We needn't enforce quota on those
1837          * objects, so always set the lqi_space as 0.
1838          */
1839         if (inode != NULL)
1840                 rc = osd_declare_inode_qid(env, i_uid_read(inode),
1841                                            i_gid_read(inode),
1842                                            i_projid_read(inode), 0,
1843                                            oh, obj, NULL, OSD_QID_BLK);
1844
1845         if (rc == 0)
1846                 rc = osd_trunc_lock(obj, oh, true);
1847
1848         RETURN(rc);
1849 }
1850
1851 static int osd_ldiskfs_writelink(struct inode *inode, char *buffer, int buflen)
1852 {
1853         /* LU-2634: clear the extent format for fast symlink */
1854         ldiskfs_clear_inode_flag(inode, LDISKFS_INODE_EXTENTS);
1855
1856         memcpy((char *)&LDISKFS_I(inode)->i_data, (char *)buffer, buflen);
1857         spin_lock(&inode->i_lock);
1858         LDISKFS_I(inode)->i_disksize = buflen;
1859         i_size_write(inode, buflen);
1860         spin_unlock(&inode->i_lock);
1861         osd_dirty_inode(inode, I_DIRTY_DATASYNC);
1862
1863         return 0;
1864 }
1865
1866 static int osd_ldiskfs_write_record(struct dt_object *dt, void *buf,
1867                                     int bufsize, int write_NUL, loff_t *offs,
1868                                     handle_t *handle)
1869 {
1870         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1871         struct buffer_head *bh        = NULL;
1872         loff_t              offset    = *offs;
1873         loff_t              new_size  = i_size_read(inode);
1874         unsigned long       block;
1875         int                 blocksize = 1 << inode->i_blkbits;
1876         struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
1877         int                 err = 0;
1878         int                 size;
1879         int                 boffs;
1880         int                 dirty_inode = 0;
1881         bool create, sparse, sync = false;
1882
1883         if (write_NUL) {
1884                 /*
1885                  * long symlink write does not count the NUL terminator in
1886                  * bufsize, we write it, and the inode's file size does not
1887                  * count the NUL terminator as well.
1888                  */
1889                 ((char *)buf)[bufsize] = '\0';
1890                 ++bufsize;
1891         }
1892
1893         dirty_inode = test_and_set_bit(LDISKFS_INODE_JOURNAL_DATA,
1894                                        &ei->i_flags);
1895
1896         /* sparse checking is racy, but sparse is very rare case, leave as is */
1897         sparse = (new_size > 0 && (inode->i_blocks >> (inode->i_blkbits - 9)) <
1898                   ((new_size - 1) >> inode->i_blkbits) + 1);
1899
1900         while (bufsize > 0) {
1901                 int credits = handle->h_buffer_credits;
1902                 unsigned long last_block = (new_size == 0) ? 0 :
1903                                            (new_size - 1) >> inode->i_blkbits;
1904
1905                 if (bh)
1906                         brelse(bh);
1907
1908                 block = offset >> inode->i_blkbits;
1909                 boffs = offset & (blocksize - 1);
1910                 size = min(blocksize - boffs, bufsize);
1911                 sync = (block > last_block || new_size == 0 || sparse);
1912
1913                 if (sync)
1914                         down(&ei->i_append_sem);
1915
1916                 bh = __ldiskfs_bread(handle, inode, block, 0);
1917
1918                 if (unlikely(IS_ERR_OR_NULL(bh) && !sync))
1919                         CWARN(
1920                               "%s: adding bh without locking off %llu (block %lu, size %d, offs %llu)\n",
1921                               inode->i_sb->s_id,
1922                               offset, block, bufsize, *offs);
1923
1924                 if (IS_ERR_OR_NULL(bh)) {
1925                         struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1926                         int flags = LDISKFS_GET_BLOCKS_CREATE;
1927
1928                         /* while the file system is being mounted, avoid
1929                          * preallocation otherwise mount can take a long
1930                          * time as mballoc cache is cold.
1931                          * XXX: this is a workaround until we have a proper
1932                          *      fix in mballoc
1933                          * XXX: works with extent-based files only */
1934                         if (!osd->od_cl_seq)
1935                                 flags |= LDISKFS_GET_BLOCKS_NO_NORMALIZE;
1936                         bh = __ldiskfs_bread(handle, inode, block, flags);
1937                         create = true;
1938                 } else {
1939                         if (sync) {
1940                                 up(&ei->i_append_sem);
1941                                 sync = false;
1942                         }
1943                         create = false;
1944                 }
1945                 if (IS_ERR_OR_NULL(bh)) {
1946                         if (bh == NULL) {
1947                                 err = -EIO;
1948                         } else {
1949                                 err = PTR_ERR(bh);
1950                                 bh = NULL;
1951                         }
1952
1953                         CERROR(
1954                                "%s: error reading offset %llu (block %lu, size %d, offs %llu), credits %d/%d: rc = %d\n",
1955                                inode->i_sb->s_id, offset, block, bufsize, *offs,
1956                                credits, handle->h_buffer_credits, err);
1957                         break;
1958                 }
1959
1960                 err = ldiskfs_journal_get_write_access(handle, bh);
1961                 if (err) {
1962                         CERROR("journal_get_write_access() returned error %d\n",
1963                                err);
1964                         break;
1965                 }
1966                 LASSERTF(boffs + size <= bh->b_size,
1967                          "boffs %d size %d bh->b_size %lu\n",
1968                          boffs, size, (unsigned long)bh->b_size);
1969                 if (create) {
1970                         memset(bh->b_data, 0, bh->b_size);
1971                         if (sync) {
1972                                 up(&ei->i_append_sem);
1973                                 sync = false;
1974                         }
1975                 }
1976                 memcpy(bh->b_data + boffs, buf, size);
1977                 err = ldiskfs_handle_dirty_metadata(handle, NULL, bh);
1978                 if (err)
1979                         break;
1980
1981                 if (offset + size > new_size)
1982                         new_size = offset + size;
1983                 offset += size;
1984                 bufsize -= size;
1985                 buf += size;
1986         }
1987         if (sync)
1988                 up(&ei->i_append_sem);
1989
1990         if (bh)
1991                 brelse(bh);
1992
1993         if (write_NUL)
1994                 --new_size;
1995         /* correct in-core and on-disk sizes */
1996         if (new_size > i_size_read(inode)) {
1997                 spin_lock(&inode->i_lock);
1998                 if (new_size > i_size_read(inode))
1999                         i_size_write(inode, new_size);
2000                 if (i_size_read(inode) > ei->i_disksize) {
2001                         ei->i_disksize = i_size_read(inode);
2002                         dirty_inode = 1;
2003                 }
2004                 spin_unlock(&inode->i_lock);
2005         }
2006         if (dirty_inode)
2007                 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
2008
2009         if (err == 0)
2010                 *offs = offset;
2011         return err;
2012 }
2013
2014 static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt,
2015                          const struct lu_buf *buf, loff_t *pos,
2016                          struct thandle *handle)
2017 {
2018         struct inode            *inode = osd_dt_obj(dt)->oo_inode;
2019         struct osd_thandle      *oh;
2020         ssize_t                 result;
2021         int                     is_link;
2022
2023         LASSERT(dt_object_exists(dt));
2024
2025         LASSERT(handle != NULL);
2026         LASSERT(inode != NULL);
2027         dquot_initialize(inode);
2028
2029         /* XXX: don't check: one declared chunk can be used many times */
2030         /* osd_trans_exec_op(env, handle, OSD_OT_WRITE); */
2031
2032         oh = container_of(handle, struct osd_thandle, ot_super);
2033         LASSERT(oh->ot_handle->h_transaction != NULL);
2034         osd_trans_exec_op(env, handle, OSD_OT_WRITE);
2035
2036         /* Write small symlink to inode body as we need to maintain correct
2037          * on-disk symlinks for ldiskfs.
2038          * Note: the buf->lb_buf contains a NUL terminator while buf->lb_len
2039          * does not count it in.
2040          */
2041         is_link = S_ISLNK(dt->do_lu.lo_header->loh_attr);
2042         if (is_link && (buf->lb_len < sizeof(LDISKFS_I(inode)->i_data)))
2043                 result = osd_ldiskfs_writelink(inode, buf->lb_buf, buf->lb_len);
2044         else
2045                 result = osd_ldiskfs_write_record(dt, buf->lb_buf, buf->lb_len,
2046                                                   is_link, pos, oh->ot_handle);
2047         if (result == 0)
2048                 result = buf->lb_len;
2049
2050         osd_trans_exec_check(env, handle, OSD_OT_WRITE);
2051
2052         return result;
2053 }
2054
2055 static int osd_declare_fallocate(const struct lu_env *env,
2056                                  struct dt_object *dt, __u64 start, __u64 end,
2057                                  int mode, struct thandle *th)
2058 {
2059         struct osd_thandle *oh = container_of(th, struct osd_thandle, ot_super);
2060         struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
2061         struct inode *inode = osd_dt_obj(dt)->oo_inode;
2062         long long quota_space = 0;
2063         /* 5 is max tree depth. (inode + 4 index blocks) */
2064         int depth = 5;
2065         int rc;
2066
2067         ENTRY;
2068
2069         /*
2070          * Only mode == 0 (which is standard prealloc) is supported now.
2071          * Rest of mode options is not supported yet.
2072          */
2073         if (mode & ~FALLOC_FL_KEEP_SIZE)
2074                 RETURN(-EOPNOTSUPP);
2075
2076         LASSERT(th);
2077         LASSERT(inode);
2078
2079         /* quota space for metadata blocks
2080          * approximate metadata estimate should be good enough.
2081          */
2082         quota_space += PAGE_SIZE;
2083         quota_space += depth * LDISKFS_BLOCK_SIZE(osd_sb(osd));
2084
2085         /* quota space should be reported in 1K blocks */
2086         quota_space = toqb(quota_space) + toqb(end - start) +
2087                       LDISKFS_META_TRANS_BLOCKS(inode->i_sb);
2088
2089         /* We don't need to reserve credits for whole fallocate here.
2090          * We reserve space only for metadata. Fallocate credits are
2091          * extended as required
2092          */
2093         rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
2094                                    i_projid_read(inode), quota_space, oh,
2095                                    osd_dt_obj(dt), NULL, OSD_QID_BLK);
2096         RETURN(rc);
2097 }
2098
2099 static int osd_fallocate(const struct lu_env *env, struct dt_object *dt,
2100                          __u64 start, __u64 end, int mode, struct thandle *th)
2101 {
2102         struct osd_thandle *oh = container_of(th, struct osd_thandle, ot_super);
2103         handle_t *handle = ldiskfs_journal_current_handle();
2104         unsigned int save_credits = oh->ot_credits;
2105         struct osd_object *obj = osd_dt_obj(dt);
2106         struct inode *inode = obj->oo_inode;
2107         struct ldiskfs_map_blocks map;
2108         unsigned int credits;
2109         ldiskfs_lblk_t blen;
2110         ldiskfs_lblk_t boff;
2111         loff_t new_size = 0;
2112         int depth = 0;
2113         int flags;
2114         int rc = 0;
2115
2116         ENTRY;
2117
2118         LASSERT(dt_object_exists(dt));
2119         LASSERT(osd_invariant(obj));
2120         LASSERT(inode != NULL);
2121
2122         CDEBUG(D_INODE, "fallocate: inode #%lu: start %llu end %llu mode %d\n",
2123                inode->i_ino, start, end, mode);
2124
2125         dquot_initialize(inode);
2126
2127         LASSERT(th);
2128
2129         boff = start >> inode->i_blkbits;
2130         blen = (ALIGN(end, 1 << inode->i_blkbits) >> inode->i_blkbits) - boff;
2131
2132         flags = LDISKFS_GET_BLOCKS_CREATE;
2133         if (mode & FALLOC_FL_KEEP_SIZE)
2134                 flags |= LDISKFS_GET_BLOCKS_KEEP_SIZE;
2135
2136         inode_lock(inode);
2137
2138         /*
2139          * We only support preallocation for extent-based file only.
2140          */
2141         if (!(ldiskfs_test_inode_flag(inode, LDISKFS_INODE_EXTENTS)))
2142                 GOTO(out, rc = -EOPNOTSUPP);
2143
2144         if (!(mode & FALLOC_FL_KEEP_SIZE) && (end > i_size_read(inode) ||
2145             end > LDISKFS_I(inode)->i_disksize)) {
2146                 new_size = end;
2147                 rc = inode_newsize_ok(inode, new_size);
2148                 if (rc)
2149                         GOTO(out, rc);
2150         }
2151
2152         inode_dio_wait(inode);
2153
2154         map.m_lblk = boff;
2155         map.m_len = blen;
2156
2157         /* Don't normalize the request if it can fit in one extent so
2158          * that it doesn't get unnecessarily split into multiple extents.
2159          */
2160         if (blen <= EXT_UNWRITTEN_MAX_LEN)
2161                 flags |= LDISKFS_GET_BLOCKS_NO_NORMALIZE;
2162
2163         /*
2164          * credits to insert 1 extent into extent tree.
2165          */
2166         credits = osd_chunk_trans_blocks(inode, blen);
2167         depth = ext_depth(inode);
2168
2169         while (rc >= 0 && blen) {
2170                 loff_t epos;
2171
2172                 /*
2173                  * Recalculate credits when extent tree depth changes.
2174                  */
2175                 if (depth != ext_depth(inode)) {
2176                         credits = osd_chunk_trans_blocks(inode, blen);
2177                         depth = ext_depth(inode);
2178                 }
2179
2180                 /* TODO: quota check */
2181                 rc = osd_extend_restart_trans(handle, credits);
2182                 if (rc)
2183                         break;
2184
2185                 rc = ldiskfs_map_blocks(handle, inode, &map, flags);
2186                 if (rc <= 0) {
2187                         CDEBUG(D_INODE,
2188                                "inode #%lu: block %u: len %u: ldiskfs_map_blocks returned %d\n",
2189                                inode->i_ino, map.m_lblk, map.m_len, rc);
2190                         ldiskfs_mark_inode_dirty(handle, inode);
2191                         break;
2192                 }
2193
2194                 map.m_lblk += rc;
2195                 map.m_len = blen = blen - rc;
2196                 epos = (loff_t)map.m_lblk << inode->i_blkbits;
2197                 inode->i_ctime = current_time(inode);
2198                 if (new_size) {
2199                         if (epos > end)
2200                                 epos = end;
2201                         if (ldiskfs_update_inode_size(inode, epos) & 0x1)
2202                                 inode->i_mtime = inode->i_ctime;
2203                 } else {
2204                         if (epos > inode->i_size)
2205                                 ldiskfs_set_inode_flag(inode,
2206                                                        LDISKFS_INODE_EOFBLOCKS);
2207                 }
2208
2209                 ldiskfs_mark_inode_dirty(handle, inode);
2210         }
2211
2212 out:
2213         inode_unlock(inode);
2214
2215         /* extand credits if needed for operations such as attribute set */
2216         if (rc >= 0)
2217                 rc = osd_extend_restart_trans(handle, save_credits);
2218
2219         RETURN(rc);
2220 }
2221
2222 static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt,
2223                              __u64 start, __u64 end, struct thandle *th)
2224 {
2225         struct osd_thandle *oh;
2226         struct inode       *inode;
2227         int                 rc;
2228         ENTRY;
2229
2230         LASSERT(th);
2231         oh = container_of(th, struct osd_thandle, ot_super);
2232
2233         /*
2234          * we don't need to reserve credits for whole truncate
2235          * it's not possible as truncate may need to free too many
2236          * blocks and that won't fit a single transaction. instead
2237          * we reserve credits to change i_size and put inode onto
2238          * orphan list. if needed truncate will extend or restart
2239          * transaction
2240          */
2241         osd_trans_declare_op(env, oh, OSD_OT_PUNCH,
2242                              osd_dto_credits_noquota[DTO_ATTR_SET_BASE] + 3);
2243
2244         inode = osd_dt_obj(dt)->oo_inode;
2245         LASSERT(inode);
2246
2247         rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
2248                                    i_projid_read(inode), 0, oh, osd_dt_obj(dt),
2249                                    NULL, OSD_QID_BLK);
2250
2251         if (rc == 0)
2252                 rc = osd_trunc_lock(osd_dt_obj(dt), oh, false);
2253
2254         RETURN(rc);
2255 }
2256
2257 static int osd_punch(const struct lu_env *env, struct dt_object *dt,
2258                      __u64 start, __u64 end, struct thandle *th)
2259 {
2260         struct osd_object *obj = osd_dt_obj(dt);
2261         struct osd_device *osd = osd_obj2dev(obj);
2262         struct inode *inode = obj->oo_inode;
2263         struct osd_access_lock *al;
2264         struct osd_thandle *oh;
2265         int rc = 0, found = 0;
2266         bool grow = false;
2267         ENTRY;
2268
2269         LASSERT(dt_object_exists(dt));
2270         LASSERT(osd_invariant(obj));
2271         LASSERT(inode != NULL);
2272         dquot_initialize(inode);
2273
2274         LASSERT(th);
2275         oh = container_of(th, struct osd_thandle, ot_super);
2276         LASSERT(oh->ot_handle->h_transaction != NULL);
2277
2278         /* we used to skip truncate to current size to
2279          * optimize truncates on OST. with DoM we can
2280          * get attr_set to set specific size (MDS_REINT)
2281          * and then get truncate RPC which essentially
2282          * would be skipped. this is bad.. so, disable
2283          * this optimization on MDS till the client stop
2284          * to sent MDS_REINT (LU-11033) -bzzz
2285          */
2286         if (osd->od_is_ost && i_size_read(inode) == start)
2287                 RETURN(0);
2288
2289         osd_trans_exec_op(env, th, OSD_OT_PUNCH);
2290
2291         spin_lock(&inode->i_lock);
2292         if (i_size_read(inode) < start)
2293                 grow = true;
2294         i_size_write(inode, start);
2295         spin_unlock(&inode->i_lock);
2296         /* if object holds encrypted content, we need to make sure we truncate
2297          * on an encryption unit boundary, or subsequent reads will get
2298          * corrupted content
2299          */
2300         if (obj->oo_lma_flags & LUSTRE_ENCRYPT_FL &&
2301             start & ~LUSTRE_ENCRYPTION_MASK)
2302                 start = (start & LUSTRE_ENCRYPTION_MASK) +
2303                         LUSTRE_ENCRYPTION_UNIT_SIZE;
2304         ll_truncate_pagecache(inode, start);
2305
2306         /* optimize grow case */
2307         if (grow) {
2308                 osd_execute_truncate(obj);
2309                 GOTO(out, rc);
2310         }
2311
2312         inode_lock(inode);
2313         /* add to orphan list to ensure truncate completion
2314          * if this transaction succeed. ldiskfs_truncate()
2315          * will take the inode out of the list
2316          */
2317         rc = ldiskfs_orphan_add(oh->ot_handle, inode);
2318         inode_unlock(inode);
2319         if (rc != 0)
2320                 GOTO(out, rc);
2321
2322         list_for_each_entry(al, &oh->ot_trunc_locks, tl_list) {
2323                 if (obj != al->tl_obj)
2324                         continue;
2325                 LASSERT(al->tl_shared == 0);
2326                 found = 1;
2327                 /* do actual truncate in osd_trans_stop() */
2328                 al->tl_truncate = 1;
2329                 break;
2330         }
2331         LASSERT(found);
2332
2333 out:
2334         RETURN(rc);
2335 }
2336
2337 static int fiemap_check_ranges(struct inode *inode,
2338                                u64 start, u64 len, u64 *new_len)
2339 {
2340         loff_t maxbytes;
2341
2342         *new_len = len;
2343
2344         if (len == 0)
2345                 return -EINVAL;
2346
2347         if (ldiskfs_test_inode_flag(inode, LDISKFS_INODE_EXTENTS))
2348                 maxbytes = inode->i_sb->s_maxbytes;
2349         else
2350                 maxbytes = LDISKFS_SB(inode->i_sb)->s_bitmap_maxbytes;
2351
2352         if (start > maxbytes)
2353                 return -EFBIG;
2354
2355         /*
2356          * Shrink request scope to what the fs can actually handle.
2357          */
2358         if (len > maxbytes || (maxbytes - len) < start)
2359                 *new_len = maxbytes - start;
2360
2361         return 0;
2362 }
2363
2364 /* So that the fiemap access checks can't overflow on 32 bit machines. */
2365 #define FIEMAP_MAX_EXTENTS     (UINT_MAX / sizeof(struct fiemap_extent))
2366
2367 static int osd_fiemap_get(const struct lu_env *env, struct dt_object *dt,
2368                           struct fiemap *fm)
2369 {
2370         struct fiemap_extent_info fieinfo = {0, };
2371         struct inode *inode = osd_dt_obj(dt)->oo_inode;
2372         u64 len;
2373         int rc;
2374         mm_segment_t cur_fs;
2375
2376         LASSERT(inode);
2377         if (inode->i_op->fiemap == NULL)
2378                 return -EOPNOTSUPP;
2379
2380         if (fm->fm_extent_count > FIEMAP_MAX_EXTENTS)
2381                 return -EINVAL;
2382
2383         rc = fiemap_check_ranges(inode, fm->fm_start, fm->fm_length, &len);
2384         if (rc)
2385                 return rc;
2386
2387         fieinfo.fi_flags = fm->fm_flags;
2388         fieinfo.fi_extents_max = fm->fm_extent_count;
2389         fieinfo.fi_extents_start = fm->fm_extents;
2390
2391         if (fieinfo.fi_flags & FIEMAP_FLAG_SYNC)
2392                 filemap_write_and_wait(inode->i_mapping);
2393
2394         /* Save previous value address limit */
2395         cur_fs = get_fs();
2396         /* Set the address limit of the kernel */
2397         set_fs(KERNEL_DS);
2398
2399         rc = inode->i_op->fiemap(inode, &fieinfo, fm->fm_start, len);
2400         fm->fm_flags = fieinfo.fi_flags;
2401         fm->fm_mapped_extents = fieinfo.fi_extents_mapped;
2402
2403         /* Restore the previous address limt */
2404         set_fs(cur_fs);
2405
2406         return rc;
2407 }
2408
2409 static int osd_ladvise(const struct lu_env *env, struct dt_object *dt,
2410                        __u64 start, __u64 end, enum lu_ladvise_type advice)
2411 {
2412         struct osd_object *obj = osd_dt_obj(dt);
2413         int rc = 0;
2414         ENTRY;
2415
2416         switch (advice) {
2417         case LU_LADVISE_DONTNEED:
2418                 if (end)
2419                         invalidate_mapping_pages(obj->oo_inode->i_mapping,
2420                                                  start >> PAGE_SHIFT,
2421                                                  (end - 1) >> PAGE_SHIFT);
2422                 break;
2423         default:
2424                 rc = -ENOTSUPP;
2425                 break;
2426         }
2427
2428         RETURN(rc);
2429 }
2430
2431 static loff_t osd_lseek(const struct lu_env *env, struct dt_object *dt,
2432                         loff_t offset, int whence)
2433 {
2434         struct osd_object *obj = osd_dt_obj(dt);
2435         struct inode *inode = obj->oo_inode;
2436         struct file *file;
2437         loff_t result;
2438
2439         ENTRY;
2440
2441         LASSERT(dt_object_exists(dt));
2442         LASSERT(osd_invariant(obj));
2443         LASSERT(inode);
2444         LASSERT(offset >= 0);
2445
2446         file = osd_quasi_file(env, inode);
2447         result = file->f_op->llseek(file, offset, whence);
2448
2449         /*
2450          * If 'offset' is beyond end of object file then treat it as not error
2451          * but valid case for SEEK_HOLE and return 'offset' as result.
2452          * LOV will decide if it is beyond real end of file or not.
2453          */
2454         if (whence == SEEK_HOLE && result == -ENXIO)
2455                 result = offset;
2456
2457         CDEBUG(D_INFO, "seek %s from %lld: %lld\n", whence == SEEK_HOLE ?
2458                        "hole" : "data", offset, result);
2459         RETURN(result);
2460 }
2461
2462 /*
2463  * in some cases we may need declare methods for objects being created
2464  * e.g., when we create symlink
2465  */
2466 const struct dt_body_operations osd_body_ops_new = {
2467         .dbo_declare_write = osd_declare_write,
2468 };
2469
2470 const struct dt_body_operations osd_body_ops = {
2471         .dbo_read                       = osd_read,
2472         .dbo_declare_write              = osd_declare_write,
2473         .dbo_write                      = osd_write,
2474         .dbo_bufs_get                   = osd_bufs_get,
2475         .dbo_bufs_put                   = osd_bufs_put,
2476         .dbo_write_prep                 = osd_write_prep,
2477         .dbo_declare_write_commit       = osd_declare_write_commit,
2478         .dbo_write_commit               = osd_write_commit,
2479         .dbo_read_prep                  = osd_read_prep,
2480         .dbo_declare_punch              = osd_declare_punch,
2481         .dbo_punch                      = osd_punch,
2482         .dbo_fiemap_get                 = osd_fiemap_get,
2483         .dbo_ladvise                    = osd_ladvise,
2484         .dbo_declare_fallocate          = osd_declare_fallocate,
2485         .dbo_fallocate                  = osd_fallocate,
2486         .dbo_lseek                      = osd_lseek,
2487 };
2488
2489 /**
2490  * Get a truncate lock
2491  *
2492  * In order to take multi-transaction truncate out of main transaction we let
2493  * the caller grab a lock on the object passed. the lock can be shared (for
2494  * writes) and exclusive (for truncate). It's not allowed to mix truncate
2495  * and write in the same transaction handle (do not confuse with big ldiskfs
2496  * transaction containing lots of handles).
2497  * The lock must be taken at declaration.
2498  *
2499  * \param obj           object to lock
2500  * \oh                  transaction
2501  * \shared              shared or exclusive
2502  *
2503  * \retval 0            lock is granted
2504  * \retval -NOMEM       no memory to allocate lock
2505  */
2506 int osd_trunc_lock(struct osd_object *obj, struct osd_thandle *oh, bool shared)
2507 {
2508         struct osd_access_lock *al, *tmp;
2509
2510         LASSERT(obj);
2511         LASSERT(oh);
2512
2513         list_for_each_entry(tmp, &oh->ot_trunc_locks, tl_list) {
2514                 if (tmp->tl_obj != obj)
2515                         continue;
2516                 LASSERT(tmp->tl_shared == shared);
2517                 /* found same lock */
2518                 return 0;
2519         }
2520
2521         OBD_ALLOC_PTR(al);
2522         if (unlikely(al == NULL))
2523                 return -ENOMEM;
2524         al->tl_obj = obj;
2525         al->tl_truncate = false;
2526         if (shared)
2527                 down_read(&obj->oo_ext_idx_sem);
2528         else
2529                 down_write(&obj->oo_ext_idx_sem);
2530         al->tl_shared = shared;
2531         lu_object_get(&obj->oo_dt.do_lu);
2532
2533         list_add(&al->tl_list, &oh->ot_trunc_locks);
2534
2535         return 0;
2536 }
2537
2538 void osd_trunc_unlock_all(const struct lu_env *env, struct list_head *list)
2539 {
2540         struct osd_access_lock *al, *tmp;
2541
2542         list_for_each_entry_safe(al, tmp, list, tl_list) {
2543                 if (al->tl_shared)
2544                         up_read(&al->tl_obj->oo_ext_idx_sem);
2545                 else
2546                         up_write(&al->tl_obj->oo_ext_idx_sem);
2547                 osd_object_put(env, al->tl_obj);
2548                 list_del(&al->tl_list);
2549                 OBD_FREE_PTR(al);
2550         }
2551 }
2552
2553 void osd_execute_truncate(struct osd_object *obj)
2554 {
2555         struct osd_device *d = osd_obj2dev(obj);
2556         struct inode *inode = obj->oo_inode;
2557         __u64 size;
2558
2559         /* simulate crash before (in the middle) of delayed truncate */
2560         if (OBD_FAIL_CHECK(OBD_FAIL_OSD_FAIL_AT_TRUNCATE)) {
2561                 struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
2562                 struct ldiskfs_sb_info *sbi = LDISKFS_SB(inode->i_sb);
2563
2564                 mutex_lock(&sbi->s_orphan_lock);
2565                 list_del_init(&ei->i_orphan);
2566                 mutex_unlock(&sbi->s_orphan_lock);
2567                 return;
2568         }
2569
2570         size = i_size_read(inode);
2571         inode_lock(inode);
2572         /* if object holds encrypted content, we need to make sure we truncate
2573          * on an encryption unit boundary, or block content will get corrupted
2574          */
2575         if (obj->oo_lma_flags & LUSTRE_ENCRYPT_FL &&
2576             size & ~LUSTRE_ENCRYPTION_MASK)
2577                 inode->i_size = (size & LUSTRE_ENCRYPTION_MASK) +
2578                         LUSTRE_ENCRYPTION_UNIT_SIZE;
2579         ldiskfs_truncate(inode);
2580         inode_unlock(inode);
2581         if (inode->i_size != size) {
2582                 spin_lock(&inode->i_lock);
2583                 i_size_write(inode, size);
2584                 LDISKFS_I(inode)->i_disksize = size;
2585                 spin_unlock(&inode->i_lock);
2586                 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
2587         }
2588
2589         /*
2590          * For a partial-page truncate, flush the page to disk immediately to
2591          * avoid data corruption during direct disk write.  b=17397
2592          */
2593         if ((size & ~PAGE_MASK) == 0)
2594                 return;
2595         if (osd_use_page_cache(d)) {
2596                 filemap_fdatawrite_range(inode->i_mapping, size, size + 1);
2597         } else {
2598                 /* Notice we use "wait" version to ensure I/O is complete */
2599                 filemap_write_and_wait_range(inode->i_mapping, size, size + 1);
2600                 invalidate_mapping_pages(inode->i_mapping, size >> PAGE_SHIFT,
2601                                          size >> PAGE_SHIFT);
2602         }
2603 }
2604
2605 void osd_process_truncates(struct list_head *list)
2606 {
2607         struct osd_access_lock *al;
2608
2609         LASSERT(journal_current_handle() == NULL);
2610
2611         list_for_each_entry(al, list, tl_list) {
2612                 if (al->tl_shared)
2613                         continue;
2614                 if (!al->tl_truncate)
2615                         continue;
2616                 osd_execute_truncate(al->tl_obj);
2617         }
2618 }