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 = blk_status_to_errno(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 void osd_mark_page_io_done(struct osd_iobuf *iobuf,
 444                                   struct inode *inode,
 445                                   sector_t start_blocks,
 446                                   sector_t count)
 447 {
 448         struct niobuf_local *lnb;
 449         int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
 450         pgoff_t pg_start, pg_end;
 451
 452         pg_start = start_blocks / blocks_per_page;
 453         if (start_blocks % blocks_per_page)
 454                 pg_start++;
 455         if (count >= blocks_per_page)
 456                 pg_end = (start_blocks + count -
 457                           blocks_per_page) / blocks_per_page;
 458         else
 459                 return; /* nothing to mark */
 460         for ( ; pg_start <= pg_end; pg_start++) {
 461                 lnb = iobuf->dr_lnbs[pg_start];
 462                 lnb->lnb_flags |= OBD_BRW_DONE;
 463         }
 464 }
 465
 466 static int osd_do_bio(struct osd_device *osd, struct inode *inode,
 467                       struct osd_iobuf *iobuf, sector_t start_blocks,
 468                       sector_t count)
 469 {
 470         int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
 471         struct page **pages = iobuf->dr_pages;
 472         int npages = iobuf->dr_npages;
 473         sector_t *blocks = iobuf->dr_blocks;
 474         struct super_block *sb = inode->i_sb;
 475         int sector_bits = sb->s_blocksize_bits - 9;
 476         unsigned int blocksize = sb->s_blocksize;
 477         struct block_device *bdev = sb->s_bdev;
 478         struct osd_bio_private *bio_private = NULL;
 479         struct bio *bio = NULL;
 480         int bio_start_page_idx;
 481         struct page *page;
 482         unsigned int page_offset;
 483         sector_t sector;
 484         int nblocks;
 485         int block_idx, block_idx_end;
 486         int page_idx, page_idx_start;
 487         int i;
 488         int rc = 0;
 489         bool fault_inject;
 490         bool integrity_enabled;
 491         struct blk_plug plug;
 492         int blocks_left_page;
 493
 494         ENTRY;
 495
 496         fault_inject = OBD_FAIL_CHECK(OBD_FAIL_OST_INTEGRITY_FAULT);
 497         LASSERT(iobuf->dr_npages == npages);
 498
 499         integrity_enabled = bdev_integrity_enabled(bdev, iobuf->dr_rw);
 500
 501         osd_brw_stats_update(osd, iobuf);
 502         iobuf->dr_start_time = ktime_get();
 503
 504         if (!count)
 505                 count = npages * blocks_per_page;
 506         block_idx_end = start_blocks + count;
 507
 508         blk_start_plug(&plug);
 509
 510         page_idx_start = start_blocks / blocks_per_page;
 511         for (page_idx = page_idx_start, block_idx = start_blocks;
 512              block_idx < block_idx_end; page_idx++,
 513              block_idx += blocks_left_page) {
 514                 page = pages[page_idx];
 515                 LASSERT(page_idx < iobuf->dr_npages);
 516
 517                 i = block_idx % blocks_per_page;
 518                 blocks_left_page = blocks_per_page - i;
 519                 for (page_offset = i * blocksize; i < blocks_left_page;
 520                      i += nblocks, page_offset += blocksize * nblocks) {
 521                         nblocks = 1;
 522
 523                         if (blocks[block_idx + i] == 0) {  /* hole */
 524                                 LASSERTF(iobuf->dr_rw == 0,
 525                                          "page_idx %u, block_idx %u, i %u,"
 526                                          "start_blocks: %llu, count: %llu, npages: %d\n",
 527                                          page_idx, block_idx, i,
 528                                          (unsigned long long)start_blocks,
 529                                          (unsigned long long)count, npages);
 530                                 memset(kmap(page) + page_offset, 0, blocksize);
 531                                 kunmap(page);
 532                                 continue;
 533                         }
 534
 535                         sector = (sector_t)blocks[block_idx + i] << sector_bits;
 536
 537                         /* Additional contiguous file blocks? */
 538                         while (i + nblocks < blocks_left_page &&
 539                                (sector + (nblocks << sector_bits)) ==
 540                                ((sector_t)blocks[block_idx + i + nblocks] <<
 541                                  sector_bits))
 542                                 nblocks++;
 543
 544                         if (bio && can_be_merged(bio, sector) &&
 545                             bio_add_page(bio, page, blocksize * nblocks,
 546                                          page_offset) != 0)
 547                                 continue;       /* added this frag OK */
 548
 549                         if (bio != NULL) {
 550                                 struct request_queue *q = bio_get_queue(bio);
 551                                 unsigned int bi_size = bio_sectors(bio) << 9;
 552
 553                                 /* Dang! I have to fragment this I/O */
 554                                 CDEBUG(D_INODE,
 555                                        "bio++ sz %d vcnt %d(%d) sectors %d(%d) psg %d(%d)\n",
 556                                        bi_size, bio->bi_vcnt, bio->bi_max_vecs,
 557                                        bio_sectors(bio),
 558                                        queue_max_sectors(q),
 559                                        osd_bio_nr_segs(bio),
 560                                        queue_max_segments(q));
 561                                 rc = osd_bio_integrity_handle(osd, bio,
 562                                         iobuf, bio_start_page_idx,
 563                                         fault_inject, integrity_enabled);
 564                                 if (rc) {
 565                                         bio_put(bio);
 566                                         goto out;
 567                                 }
 568
 569                                 record_start_io(iobuf, bi_size);
 570                                 osd_submit_bio(iobuf->dr_rw, bio);
 571                         }
 572
 573                         bio_start_page_idx = page_idx;
 574                         /* allocate new bio */
 575                         bio = bio_alloc(GFP_NOIO, min(BIO_MAX_PAGES,
 576                                         (block_idx_end - block_idx +
 577                                          blocks_left_page - 1)));
 578                         if (bio == NULL) {
 579                                 CERROR("Can't allocate bio %u pages\n",
 580                                        block_idx_end - block_idx +
 581                                        blocks_left_page - 1);
 582                                 rc = -ENOMEM;
 583                                 goto out;
 584                         }
 585
 586                         bio_set_dev(bio, bdev);
 587                         bio_set_sector(bio, sector);
 588                         bio->bi_opf = iobuf->dr_rw ? WRITE : READ;
 589                         rc = osd_bio_init(bio, iobuf, integrity_enabled,
 590                                           bio_start_page_idx, &bio_private);
 591                         if (rc) {
 592                                 bio_put(bio);
 593                                 goto out;
 594                         }
 595
 596                         rc = bio_add_page(bio, page,
 597                                           blocksize * nblocks, page_offset);
 598                         LASSERT(rc != 0);
 599                 }
 600         }
 601
 602         if (bio != NULL) {
 603                 rc = osd_bio_integrity_handle(osd, bio, iobuf,
 604                                               bio_start_page_idx,
 605                                               fault_inject,
 606                                               integrity_enabled);
 607                 if (rc) {
 608                         bio_put(bio);
 609                         goto out;
 610                 }
 611
 612                 record_start_io(iobuf, bio_sectors(bio) << 9);
 613                 osd_submit_bio(iobuf->dr_rw, bio);
 614                 rc = 0;
 615         }
 616
 617 out:
 618         blk_finish_plug(&plug);
 619
 620         /* in order to achieve better IO throughput, we don't wait for writes
 621          * completion here. instead we proceed with transaction commit in
 622          * parallel and wait for IO completion once transaction is stopped
 623          * see osd_trans_stop() for more details -bzzz
 624          */
 625         if (iobuf->dr_rw == 0 || fault_inject) {
 626                 wait_event(iobuf->dr_wait,
 627                            atomic_read(&iobuf->dr_numreqs) == 0);
 628                 osd_fini_iobuf(osd, iobuf);
 629         }
 630
 631         if (rc == 0) {
 632                 rc = iobuf->dr_error;
 633         } else {
 634                 if (bio_private)
 635                         OBD_FREE_PTR(bio_private);
 636         }
 637
 638         /* Write only now */
 639         if (rc == 0 && iobuf->dr_rw)
 640                 osd_mark_page_io_done(iobuf, inode,
 641                                       start_blocks, count);
 642
 643         RETURN(rc);
 644 }
 645
 646 static int osd_map_remote_to_local(loff_t offset, ssize_t len, int *nrpages,
 647                                    struct niobuf_local *lnb, int maxlnb)
 648 {
 649         int rc = 0;
 650         ENTRY;
 651
 652         *nrpages = 0;
 653
 654         while (len > 0) {
 655                 int poff = offset & (PAGE_SIZE - 1);
 656                 int plen = PAGE_SIZE - poff;
 657
 658                 if (*nrpages >= maxlnb) {
 659                         rc = -EOVERFLOW;
 660                         break;
 661                 }
 662
 663                 if (plen > len)
 664                         plen = len;
 665                 lnb->lnb_file_offset = offset;
 666                 lnb->lnb_page_offset = poff;
 667                 lnb->lnb_len = plen;
 668                 /* lnb->lnb_flags = rnb->rnb_flags; */
 669                 lnb->lnb_flags = 0;
 670                 lnb->lnb_page = NULL;
 671                 lnb->lnb_rc = 0;
 672                 lnb->lnb_guard_rpc = 0;
 673                 lnb->lnb_guard_disk = 0;
 674                 lnb->lnb_locked = 0;
 675
 676                 LASSERTF(plen <= len, "plen %u, len %lld\n", plen,
 677                          (long long) len);
 678                 offset += plen;
 679                 len -= plen;
 680                 lnb++;
 681                 (*nrpages)++;
 682         }
 683
 684         RETURN(rc);
 685 }
 686
 687 static struct page *osd_get_page(const struct lu_env *env, struct dt_object *dt,
 688                                  loff_t offset, gfp_t gfp_mask, bool cache)
 689 {
 690         struct osd_thread_info *oti = osd_oti_get(env);
 691         struct inode *inode = osd_dt_obj(dt)->oo_inode;
 692         struct osd_device *d = osd_obj2dev(osd_dt_obj(dt));
 693         struct page *page;
 694         int cur;
 695
 696         LASSERT(inode);
 697
 698         if (cache) {
 699                 page = find_or_create_page(inode->i_mapping,
 700                                            offset >> PAGE_SHIFT, gfp_mask);
 701
 702                 if (likely(page)) {
 703                         LASSERT(!PagePrivate2(page));
 704                         wait_on_page_writeback(page);
 705                 } else {
 706                         lprocfs_counter_add(d->od_stats, LPROC_OSD_NO_PAGE, 1);
 707                 }
 708
 709                 return page;
 710         }
 711
 712         if (inode->i_mapping->nrpages) {
 713                 /* consult with pagecache, but do not create new pages */
 714                 /* this is normally used once */
 715                 page = find_lock_page(inode->i_mapping, offset >> PAGE_SHIFT);
 716                 if (page) {
 717                         wait_on_page_writeback(page);
 718                         return page;
 719                 }
 720         }
 721
 722         LASSERT(oti->oti_dio_pages);
 723         cur = oti->oti_dio_pages_used;
 724         page = oti->oti_dio_pages[cur];
 725
 726         if (unlikely(!page)) {
 727                 LASSERT(cur < PTLRPC_MAX_BRW_PAGES);
 728                 page = alloc_page(gfp_mask);
 729                 if (!page)
 730                         return NULL;
 731                 oti->oti_dio_pages[cur] = page;
 732                 SetPagePrivate2(page);
 733                 lock_page(page);
 734         }
 735
 736         ClearPageUptodate(page);
 737         page->index = offset >> PAGE_SHIFT;
 738         oti->oti_dio_pages_used++;
 739
 740         return page;
 741 }
 742
 743 /*
 744  * there are following "locks":
 745  * journal_start
 746  * i_mutex
 747  * page lock
 748  *
 749  * osd write path:
 750  *  - lock page(s)
 751  *  - journal_start
 752  *  - truncate_sem
 753  *
 754  * ext4 vmtruncate:
 755  *  - lock pages, unlock
 756  *  - journal_start
 757  *  - lock partial page
 758  *  - i_data_sem
 759  *
 760  */
 761
 762 /**
 763  * Unlock and release pages loaded by osd_bufs_get()
 764  *
 765  * Unlock \a npages pages from \a lnb and drop the refcount on them.
 766  *
 767  * \param env           thread execution environment
 768  * \param dt            dt object undergoing IO (OSD object + methods)
 769  * \param lnb           array of pages undergoing IO
 770  * \param npages        number of pages in \a lnb
 771  *
 772  * \retval 0            always
 773  */
 774 static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt,
 775                         struct niobuf_local *lnb, int npages)
 776 {
 777         struct osd_thread_info *oti = osd_oti_get(env);
 778         struct pagevec pvec;
 779         int i;
 780
 781         ll_pagevec_init(&pvec, 0);
 782
 783         for (i = 0; i < npages; i++) {
 784                 struct page *page = lnb[i].lnb_page;
 785
 786                 if (page == NULL)
 787                         continue;
 788
 789                 /* if the page isn't cached, then reset uptodate
 790                  * to prevent reuse
 791                  */
 792                 if (PagePrivate2(page)) {
 793                         oti->oti_dio_pages_used--;
 794                 } else {
 795                         if (lnb[i].lnb_locked)
 796                                 unlock_page(page);
 797                         if (pagevec_add(&pvec, page) == 0)
 798                                 pagevec_release(&pvec);
 799                 }
 800
 801                 lnb[i].lnb_page = NULL;
 802         }
 803
 804         LASSERTF(oti->oti_dio_pages_used == 0, "%d\n", oti->oti_dio_pages_used);
 805
 806         /* Release any partial pagevec */
 807         pagevec_release(&pvec);
 808
 809         RETURN(0);
 810 }
 811
 812 /**
 813  * Load and lock pages undergoing IO
 814  *
 815  * Pages as described in the \a lnb array are fetched (from disk or cache)
 816  * and locked for IO by the caller.
 817  *
 818  * DLM locking protects us from write and truncate competing for same region,
 819  * but partial-page truncate can leave dirty pages in the cache for ldiskfs.
 820  * It's possible the writeout on a such a page is in progress when we access
 821  * it. It's also possible that during this writeout we put new (partial) data
 822  * into the page, but won't be able to proceed in filter_commitrw_write().
 823  * Therefore, just wait for writeout completion as it should be rare enough.
 824  *
 825  * \param env           thread execution environment
 826  * \param dt            dt object undergoing IO (OSD object + methods)
 827  * \param pos           byte offset of IO start
 828  * \param len           number of bytes of IO
 829  * \param lnb           array of extents undergoing IO
 830  * \param rw            read or write operation, and other flags
 831  * \param capa          capabilities
 832  *
 833  * \retval pages        (zero or more) loaded successfully
 834  * \retval -ENOMEM      on memory/page allocation error
 835  */
 836 static int osd_bufs_get(const struct lu_env *env, struct dt_object *dt,
 837                         loff_t pos, ssize_t len, struct niobuf_local *lnb,
 838                         int maxlnb, enum dt_bufs_type rw)
 839 {
 840         struct osd_thread_info *oti = osd_oti_get(env);
 841         struct osd_object *obj = osd_dt_obj(dt);
 842         struct osd_device *osd   = osd_obj2dev(obj);
 843         int npages, i, iosize, rc = 0;
 844         bool cache, write;
 845         loff_t fsize;
 846         gfp_t gfp_mask;
 847
 848         LASSERT(obj->oo_inode);
 849
 850         rc = osd_map_remote_to_local(pos, len, &npages, lnb, maxlnb);
 851         if (rc)
 852                 RETURN(rc);
 853
 854         write = rw & DT_BUFS_TYPE_WRITE;
 855
 856         fsize = lnb[npages - 1].lnb_file_offset + lnb[npages - 1].lnb_len;
 857         iosize = fsize - lnb[0].lnb_file_offset;
 858         fsize = max(fsize, i_size_read(obj->oo_inode));
 859
 860         cache = rw & DT_BUFS_TYPE_READAHEAD;
 861         if (cache)
 862                 goto bypass_checks;
 863
 864         cache = osd_use_page_cache(osd);
 865         while (cache) {
 866                 if (write) {
 867                         if (!osd->od_writethrough_cache) {
 868                                 cache = false;
 869                                 break;
 870                         }
 871                         if (iosize > osd->od_writethrough_max_iosize) {
 872                                 cache = false;
 873                                 break;
 874                         }
 875                 } else {
 876                         if (!osd->od_read_cache) {
 877                                 cache = false;
 878                                 break;
 879                         }
 880                         if (iosize > osd->od_readcache_max_iosize) {
 881                                 cache = false;
 882                                 break;
 883                         }
 884                 }
 885                 /* don't use cache on large files */
 886                 if (osd->od_readcache_max_filesize &&
 887                     fsize > osd->od_readcache_max_filesize)
 888                         cache = false;
 889                 break;
 890         }
 891
 892 bypass_checks:
 893         if (!cache && unlikely(!oti->oti_dio_pages)) {
 894                 OBD_ALLOC_PTR_ARRAY_LARGE(oti->oti_dio_pages,
 895                                           PTLRPC_MAX_BRW_PAGES);
 896                 if (!oti->oti_dio_pages)
 897                         return -ENOMEM;
 898         }
 899
 900         /* this could also try less hard for DT_BUFS_TYPE_READAHEAD pages */
 901         gfp_mask = rw & DT_BUFS_TYPE_LOCAL ? (GFP_NOFS | __GFP_HIGHMEM) :
 902                                              GFP_HIGHUSER;
 903         for (i = 0; i < npages; i++, lnb++) {
 904                 lnb->lnb_page = osd_get_page(env, dt, lnb->lnb_file_offset,
 905                                              gfp_mask, cache);
 906                 if (lnb->lnb_page == NULL)
 907                         GOTO(cleanup, rc = -ENOMEM);
 908
 909                 lnb->lnb_locked = 1;
 910         }
 911
 912 #if 0
 913         /* XXX: this version doesn't invalidate cached pages, but use them */
 914         if (!cache && write && obj->oo_inode->i_mapping->nrpages) {
 915                 /* do not allow data aliasing, invalidate pagecache */
 916                 /* XXX: can be quite expensive in mixed case */
 917                 invalidate_mapping_pages(obj->oo_inode->i_mapping,
 918                                 lnb[0].lnb_file_offset >> PAGE_SHIFT,
 919                                 lnb[npages - 1].lnb_file_offset >> PAGE_SHIFT);
 920         }
 921 #endif
 922
 923         RETURN(i);
 924
 925 cleanup:
 926         if (i > 0)
 927                 osd_bufs_put(env, dt, lnb - i, i);
 928         return rc;
 929 }
 930 /* Borrow @ext4_chunk_trans_blocks */
 931 static int osd_chunk_trans_blocks(struct inode *inode, int nrblocks)
 932 {
 933         ldiskfs_group_t groups;
 934         int gdpblocks;
 935         int idxblocks;
 936         int depth;
 937         int ret;
 938
 939         depth = ext_depth(inode);
 940         idxblocks = depth * 2;
 941
 942         /*
 943          * Now let's see how many group bitmaps and group descriptors need
 944          * to account.
 945          */
 946         groups = idxblocks + 1;
 947         gdpblocks = groups;
 948         if (groups > LDISKFS_SB(inode->i_sb)->s_groups_count)
 949                 groups = LDISKFS_SB(inode->i_sb)->s_groups_count;
 950         if (gdpblocks > LDISKFS_SB(inode->i_sb)->s_gdb_count)
 951                 gdpblocks = LDISKFS_SB(inode->i_sb)->s_gdb_count;
 952
 953         /* bitmaps and block group descriptor blocks */
 954         ret = idxblocks + groups + gdpblocks;
 955
 956         /* Blocks for super block, inode, quota and xattr blocks */
 957         ret += LDISKFS_META_TRANS_BLOCKS(inode->i_sb);
 958
 959         return ret;
 960 }
 961
 962 #ifdef HAVE_LDISKFS_JOURNAL_ENSURE_CREDITS
 963 static int osd_extend_restart_trans(handle_t *handle, int needed,
 964                                     struct inode *inode)
 965 {
 966         int rc;
 967
 968         rc = ldiskfs_journal_ensure_credits(handle, needed,
 969                 ldiskfs_trans_default_revoke_credits(inode->i_sb));
 970         /* this means journal has been restarted */
 971         if (rc > 0)
 972                 rc = 0;
 973
 974         return rc;
 975 }
 976 #else
 977 static int osd_extend_restart_trans(handle_t *handle, int needed,
 978                                     struct inode *inode)
 979 {
 980         int rc;
 981
 982         if (ldiskfs_handle_has_enough_credits(handle, needed))
 983                 return 0;
 984         rc = ldiskfs_journal_extend(handle,
 985                                 needed - handle->h_buffer_credits);
 986         if (rc <= 0)
 987                 return rc;
 988
 989         return ldiskfs_journal_restart(handle, needed);
 990 }
 991 #endif /* HAVE_LDISKFS_JOURNAL_ENSURE_CREDITS */
 992
 993 static int osd_ldiskfs_map_write(struct inode *inode, struct osd_iobuf *iobuf,
 994                                  struct osd_device *osd, sector_t start_blocks,
 995                                  sector_t count, loff_t *disk_size,
 996                                  __u64 user_size)
 997 {
 998         /* if file has grown, take user_size into account */
 999         if (user_size && *disk_size > user_size)
1000                 *disk_size = user_size;
1001
1002         spin_lock(&inode->i_lock);
1003         if (*disk_size > i_size_read(inode)) {
1004                 i_size_write(inode, *disk_size);
1005                 LDISKFS_I(inode)->i_disksize = *disk_size;
1006                 spin_unlock(&inode->i_lock);
1007                 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
1008         } else {
1009                 spin_unlock(&inode->i_lock);
1010         }
1011
1012         /*
1013          * We don't do stats here as in read path because
1014          * write is async: we'll do this in osd_put_bufs()
1015          */
1016         return osd_do_bio(osd, inode, iobuf, start_blocks, count);
1017 }
1018
1019 static unsigned int osd_extent_bytes(const struct osd_device *o)
1020 {
1021         unsigned int *extent_bytes_ptr =
1022                         raw_cpu_ptr(o->od_extent_bytes_percpu);
1023
1024         if (likely(*extent_bytes_ptr))
1025                 return *extent_bytes_ptr;
1026
1027         /* initialize on first access or CPU hotplug */
1028         if (!ldiskfs_has_feature_extents(osd_sb(o)))
1029                 *extent_bytes_ptr = 1 << osd_sb(o)->s_blocksize_bits;
1030         else
1031                 *extent_bytes_ptr = OSD_DEFAULT_EXTENT_BYTES;
1032
1033         return *extent_bytes_ptr;
1034 }
1035
1036 #define EXTENT_BYTES_DECAY 64
1037 static void osd_decay_extent_bytes(struct osd_device *osd,
1038                                    unsigned int new_bytes)
1039 {
1040         unsigned int old_bytes;
1041
1042         if (!ldiskfs_has_feature_extents(osd_sb(osd)))
1043                 return;
1044
1045         old_bytes = osd_extent_bytes(osd);
1046         *raw_cpu_ptr(osd->od_extent_bytes_percpu) =
1047                 (old_bytes * (EXTENT_BYTES_DECAY - 1) +
1048                  min(new_bytes, OSD_DEFAULT_EXTENT_BYTES) +
1049                  EXTENT_BYTES_DECAY - 1) / EXTENT_BYTES_DECAY;
1050 }
1051
1052 static int osd_ldiskfs_map_inode_pages(struct inode *inode,
1053                                        struct osd_iobuf *iobuf,
1054                                        struct osd_device *osd,
1055                                        int create, __u64 user_size,
1056                                        int check_credits,
1057                                        struct thandle *thandle)
1058 {
1059         int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
1060         int rc = 0, i = 0, mapped_index = 0;
1061         struct page *fp = NULL;
1062         int clen = 0;
1063         pgoff_t max_page_index;
1064         handle_t *handle = NULL;
1065         sector_t start_blocks = 0, count = 0;
1066         loff_t disk_size = 0;
1067         struct page **page = iobuf->dr_pages;
1068         int pages = iobuf->dr_npages;
1069         sector_t *blocks = iobuf->dr_blocks;
1070         struct niobuf_local *lnb1, *lnb2;
1071         loff_t size1, size2;
1072
1073         max_page_index = inode->i_sb->s_maxbytes >> PAGE_SHIFT;
1074
1075         CDEBUG(D_OTHER, "inode %lu: map %d pages from %lu\n",
1076                 inode->i_ino, pages, (*page)->index);
1077
1078         if (create) {
1079                 create = LDISKFS_GET_BLOCKS_CREATE;
1080                 handle = ldiskfs_journal_current_handle();
1081                 LASSERT(handle != NULL);
1082                 rc = osd_attach_jinode(inode);
1083                 if (rc)
1084                         return rc;
1085                 disk_size = i_size_read(inode);
1086                 /* if disk_size is already bigger than specified user_size,
1087                  * ignore user_size
1088                  */
1089                 if (disk_size > user_size)
1090                         user_size = 0;
1091         }
1092         /* pages are sorted already. so, we just have to find
1093          * contig. space and process them properly
1094          */
1095         while (i < pages) {
1096                 long blen, total = 0, previous_total = 0;
1097                 struct ldiskfs_map_blocks map = { 0 };
1098
1099                 if (fp == NULL) { /* start new extent */
1100                         fp = *page++;
1101                         clen = 1;
1102                         if (++i != pages)
1103                                 continue;
1104                 } else if (fp->index + clen == (*page)->index) {
1105                         /* continue the extent */
1106                         page++;
1107                         clen++;
1108                         if (++i != pages)
1109                                 continue;
1110                 }
1111                 if (fp->index + clen >= max_page_index)
1112                         GOTO(cleanup, rc = -EFBIG);
1113                 /* process found extent */
1114                 map.m_lblk = fp->index * blocks_per_page;
1115                 map.m_len = blen = clen * blocks_per_page;
1116 cont_map:
1117                 /**
1118                  * We might restart transaction for block allocations,
1119                  * in order to make sure data ordered mode, issue IO, disk
1120                  * size update and block allocations need be within same
1121                  * transaction to make sure consistency.
1122                  */
1123                 if (handle && check_credits) {
1124                         struct osd_thandle *oh;
1125
1126                         LASSERT(thandle != NULL);
1127                         oh = container_of(thandle, struct osd_thandle,
1128                                           ot_super);
1129                         /*
1130                          * only issue IO if restart transaction needed,
1131                          * as update disk size need hold inode lock, we
1132                          * want to avoid that as much as possible.
1133                          */
1134                         if (oh->oh_declared_ext <= 0) {
1135                                 rc = osd_ldiskfs_map_write(inode,
1136                                         iobuf, osd, start_blocks,
1137                                         count, &disk_size, user_size);
1138                                 if (rc)
1139                                         GOTO(cleanup, rc);
1140                                 thandle->th_restart_tran = 1;
1141                                 GOTO(cleanup, rc = -EAGAIN);
1142                         }
1143
1144                         if (OBD_FAIL_CHECK(OBD_FAIL_OST_RESTART_IO))
1145                                 oh->oh_declared_ext = 0;
1146                         else
1147                                 oh->oh_declared_ext--;
1148                 }
1149                 rc = ldiskfs_map_blocks(handle, inode, &map, create);
1150                 if (rc >= 0) {
1151                         int c = 0;
1152
1153                         for (; total < blen && c < map.m_len; c++, total++) {
1154                                 if (rc == 0) {
1155                                         *(blocks + total) = 0;
1156                                         total++;
1157                                         break;
1158                                 }
1159                                 if ((map.m_flags & LDISKFS_MAP_UNWRITTEN) &&
1160                                     !create) {
1161                                         /* don't try to read allocated, but
1162                                          * unwritten blocks, instead fill the
1163                                          * patches with zeros in osd_do_bio() */
1164                                         *(blocks + total) = 0;
1165                                         continue;
1166                                 }
1167                                 *(blocks + total) = map.m_pblk + c;
1168                                 /* unmap any possible underlying
1169                                  * metadata from the block device
1170                                  * mapping.  b=6998.
1171                                  */
1172                                 if ((map.m_flags & LDISKFS_MAP_NEW) &&
1173                                     create)
1174                                         clean_bdev_aliases(inode->i_sb->s_bdev,
1175                                                            map.m_pblk + c, 1);
1176                         }
1177                         rc = 0;
1178                 }
1179
1180                 if (rc == 0 && create) {
1181                         count += (total - previous_total);
1182                         mapped_index = (count + blocks_per_page -
1183                                         1) / blocks_per_page - 1;
1184                         lnb1 = iobuf->dr_lnbs[i - clen];
1185                         lnb2 = iobuf->dr_lnbs[mapped_index];
1186                         size1 = lnb1->lnb_file_offset -
1187                                 (lnb1->lnb_file_offset % PAGE_SIZE) +
1188                                 (total << inode->i_blkbits);
1189                         size2 = lnb2->lnb_file_offset + lnb2->lnb_len;
1190
1191                         if (size1 > size2)
1192                                 size1 = size2;
1193                         if (size1 > disk_size)
1194                                 disk_size = size1;
1195                 }
1196
1197                 if (rc == 0 && total < blen) {
1198                         /*
1199                          * decay extent blocks if we could not
1200                          * allocate extent once.
1201                          */
1202                         osd_decay_extent_bytes(osd,
1203                                 (total - previous_total) << inode->i_blkbits);
1204                         map.m_lblk = fp->index * blocks_per_page + total;
1205                         map.m_len = blen - total;
1206                         previous_total = total;
1207                         goto cont_map;
1208                 }
1209                 if (rc != 0)
1210                         GOTO(cleanup, rc);
1211                 /*
1212                  * decay extent blocks if we could allocate
1213                  * good large(1M) extent.
1214                  */
1215                 if (previous_total == 0 &&
1216                     total >= OSD_DEFAULT_EXTENT_BYTES >> inode->i_blkbits)
1217                         osd_decay_extent_bytes(osd,
1218                                                total << inode->i_blkbits);
1219                 /* look for next extent */
1220                 fp = NULL;
1221                 blocks += blocks_per_page * clen;
1222         }
1223 cleanup:
1224         if (rc == 0 && create &&
1225             start_blocks < pages * blocks_per_page) {
1226                 rc = osd_ldiskfs_map_write(inode, iobuf, osd, start_blocks,
1227                                            count, &disk_size, user_size);
1228                 LASSERT(start_blocks + count == pages * blocks_per_page);
1229         }
1230         return rc;
1231 }
1232
1233 static int osd_write_prep(const struct lu_env *env, struct dt_object *dt,
1234                           struct niobuf_local *lnb, int npages)
1235 {
1236         struct osd_thread_info *oti   = osd_oti_get(env);
1237         struct osd_iobuf       *iobuf = &oti->oti_iobuf;
1238         struct inode           *inode = osd_dt_obj(dt)->oo_inode;
1239         struct osd_device      *osd   = osd_obj2dev(osd_dt_obj(dt));
1240         ktime_t start, end;
1241         s64 timediff;
1242         ssize_t isize;
1243         __s64  maxidx;
1244         int i, rc = 0;
1245
1246         LASSERT(inode);
1247
1248         rc = osd_init_iobuf(osd, iobuf, 0, npages);
1249         if (unlikely(rc != 0))
1250                 RETURN(rc);
1251
1252         isize = i_size_read(inode);
1253         maxidx = ((isize + PAGE_SIZE - 1) >> PAGE_SHIFT) - 1;
1254
1255         start = ktime_get();
1256         for (i = 0; i < npages; i++) {
1257
1258                 /*
1259                  * till commit the content of the page is undefined
1260                  * we'll set it uptodate once bulk is done. otherwise
1261                  * subsequent reads can access non-stable data
1262                  */
1263                 ClearPageUptodate(lnb[i].lnb_page);
1264
1265                 if (lnb[i].lnb_len == PAGE_SIZE)
1266                         continue;
1267
1268                 if (maxidx >= lnb[i].lnb_page->index) {
1269                         osd_iobuf_add_page(iobuf, &lnb[i]);
1270                 } else {
1271                         long off;
1272                         char *p = kmap(lnb[i].lnb_page);
1273
1274                         off = lnb[i].lnb_page_offset;
1275                         if (off)
1276                                 memset(p, 0, off);
1277                         off = (lnb[i].lnb_page_offset + lnb[i].lnb_len) &
1278                               ~PAGE_MASK;
1279                         if (off)
1280                                 memset(p + off, 0, PAGE_SIZE - off);
1281                         kunmap(lnb[i].lnb_page);
1282                 }
1283         }
1284         end = ktime_get();
1285         timediff = ktime_us_delta(end, start);
1286         lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
1287
1288         if (iobuf->dr_npages) {
1289                 rc = osd_ldiskfs_map_inode_pages(inode, iobuf, osd, 0,
1290                                                  0, 0, NULL);
1291                 if (likely(rc == 0)) {
1292                         rc = osd_do_bio(osd, inode, iobuf, 0, 0);
1293                         /* do IO stats for preparation reads */
1294                         osd_fini_iobuf(osd, iobuf);
1295                 }
1296         }
1297         RETURN(rc);
1298 }
1299
1300 struct osd_fextent {
1301         sector_t        start;
1302         sector_t        end;
1303         unsigned int    mapped:1;
1304 };
1305
1306 static int osd_is_mapped(struct dt_object *dt, __u64 offset,
1307                          struct osd_fextent *cached_extent)
1308 {
1309         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1310         sector_t block = offset >> inode->i_blkbits;
1311         sector_t start;
1312         struct fiemap_extent_info fei = { 0 };
1313         struct fiemap_extent fe = { 0 };
1314         int rc;
1315
1316         if (block >= cached_extent->start && block < cached_extent->end)
1317                 return cached_extent->mapped;
1318
1319         if (i_size_read(inode) == 0)
1320                 return 0;
1321
1322         /* Beyond EOF, must not be mapped */
1323         if (((i_size_read(inode) - 1) >> inode->i_blkbits) < block)
1324                 return 0;
1325
1326         fei.fi_extents_max = 1;
1327         fei.fi_extents_start = &fe;
1328
1329         rc = inode->i_op->fiemap(inode, &fei, offset, FIEMAP_MAX_OFFSET-offset);
1330         if (rc != 0)
1331                 return 0;
1332
1333         start = fe.fe_logical >> inode->i_blkbits;
1334         if (fei.fi_extents_mapped == 0) {
1335                 /* a special case - no extent found at this offset and forward.
1336                  * we can consider this as a hole to EOF. it's safe to cache
1337                  * as other threads can not allocate/punch blocks this thread
1338                  * is working on (LDLM). */
1339                 cached_extent->start = block;
1340                 cached_extent->end = i_size_read(inode) >> inode->i_blkbits;
1341                 cached_extent->mapped = 0;
1342                 return 0;
1343         }
1344
1345         if (start > block) {
1346                 cached_extent->start = block;
1347                 cached_extent->end = start;
1348                 cached_extent->mapped = 0;
1349         } else {
1350                 cached_extent->start = start;
1351                 cached_extent->end = (fe.fe_logical + fe.fe_length) >>
1352                                       inode->i_blkbits;
1353                 cached_extent->mapped = 1;
1354         }
1355
1356         return cached_extent->mapped;
1357 }
1358
1359 static int osd_declare_write_commit(const struct lu_env *env,
1360                                     struct dt_object *dt,
1361                                     struct niobuf_local *lnb, int npages,
1362                                     struct thandle *handle)
1363 {
1364         const struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1365         struct inode            *inode = osd_dt_obj(dt)->oo_inode;
1366         struct osd_thandle      *oh;
1367         int                     extents = 0;
1368         int                     depth;
1369         int                     i;
1370         int                     newblocks = 0;
1371         int                     rc = 0;
1372         int                     credits = 0;
1373         long long               quota_space = 0;
1374         struct osd_fextent      mapped = { 0 }, extent = { 0 };
1375         enum osd_quota_local_flags local_flags = 0;
1376         enum osd_qid_declare_flags declare_flags = OSD_QID_BLK;
1377         unsigned int            extent_bytes;
1378         ENTRY;
1379
1380         LASSERT(handle != NULL);
1381         oh = container_of(handle, struct osd_thandle, ot_super);
1382         LASSERT(oh->ot_handle == NULL);
1383
1384         /*
1385          * We track a decaying average extent blocks per filesystem,
1386          * for most of time, it will be 1M, with filesystem becoming
1387          * heavily-fragmented, it will be reduced to 4K at the worst.
1388          */
1389         extent_bytes = osd_extent_bytes(osd);
1390         LASSERT(extent_bytes >= (1 << osd_sb(osd)->s_blocksize));
1391
1392         /* calculate number of extents (probably better to pass nb) */
1393         for (i = 0; i < npages; i++) {
1394                 /* ignore quota for the whole request if any page is from
1395                  * client cache or written by root.
1396                  *
1397                  * XXX once we drop the 1.8 client support, the checking
1398                  * for whether page is from cache can be simplified as:
1399                  * !(lnb[i].flags & OBD_BRW_SYNC)
1400                  *
1401                  * XXX we could handle this on per-lnb basis as done by
1402                  * grant.
1403                  */
1404                 if ((lnb[i].lnb_flags & OBD_BRW_NOQUOTA) ||
1405                     (lnb[i].lnb_flags & (OBD_BRW_FROM_GRANT | OBD_BRW_SYNC)) ==
1406                     OBD_BRW_FROM_GRANT)
1407                         declare_flags |= OSD_QID_FORCE;
1408
1409                 if (osd_is_mapped(dt, lnb[i].lnb_file_offset, &mapped)) {
1410                         lnb[i].lnb_flags |= OBD_BRW_MAPPED;
1411                         continue;
1412                 }
1413
1414                 if (lnb[i].lnb_flags & OBD_BRW_DONE) {
1415                         lnb[i].lnb_flags |= OBD_BRW_MAPPED;
1416                         continue;
1417                 }
1418
1419                 /* count only unmapped changes */
1420                 newblocks++;
1421                 if (lnb[i].lnb_file_offset != extent.end || extent.end == 0) {
1422                         if (extent.end != 0)
1423                                 extents += (extent.end - extent.start +
1424                                         extent_bytes - 1) / extent_bytes;
1425                         extent.start = lnb[i].lnb_file_offset;
1426                         extent.end = lnb[i].lnb_file_offset + lnb[i].lnb_len;
1427                 } else {
1428                         extent.end += lnb[i].lnb_len;
1429                 }
1430
1431                 quota_space += PAGE_SIZE;
1432         }
1433
1434         credits++; /* inode */
1435         /*
1436          * overwrite case, no need to modify tree and
1437          * allocate blocks.
1438          */
1439         if (!newblocks)
1440                 goto out_declare;
1441
1442         extents += (extent.end - extent.start +
1443                     extent_bytes - 1) / extent_bytes;
1444         /*
1445          * each extent can go into new leaf causing a split
1446          * 5 is max tree depth: inode + 4 index blocks
1447          * with blockmaps, depth is 3 at most
1448          */
1449         if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL) {
1450                 /*
1451                  * many concurrent threads may grow tree by the time
1452                  * our transaction starts. so, consider 2 is a min depth
1453                  */
1454                 depth = ext_depth(inode);
1455                 depth = max(depth, 1) + 1;
1456                 newblocks += depth;
1457                 credits += depth * 2 * extents;
1458         } else {
1459                 depth = 3;
1460                 newblocks += depth;
1461                 credits += depth * extents;
1462         }
1463
1464         oh->oh_declared_ext = extents;
1465
1466         /* quota space for metadata blocks */
1467         quota_space += depth * extents * LDISKFS_BLOCK_SIZE(osd_sb(osd));
1468
1469         /* quota space should be reported in 1K blocks */
1470         quota_space = toqb(quota_space);
1471
1472         /* each new block can go in different group (bitmap + gd) */
1473
1474         /* we can't dirty more bitmap blocks than exist */
1475         if (extents > LDISKFS_SB(osd_sb(osd))->s_groups_count)
1476                 credits += LDISKFS_SB(osd_sb(osd))->s_groups_count;
1477         else
1478                 credits += extents;
1479
1480         /* we can't dirty more gd blocks than exist */
1481         if (extents > LDISKFS_SB(osd_sb(osd))->s_gdb_count)
1482                 credits += LDISKFS_SB(osd_sb(osd))->s_gdb_count;
1483         else
1484                 credits += extents;
1485
1486 out_declare:
1487         osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
1488
1489         /* make sure the over quota flags were not set */
1490         lnb[0].lnb_flags &= ~OBD_BRW_OVER_ALLQUOTA;
1491
1492         rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
1493                                    i_projid_read(inode), quota_space, oh,
1494                                    osd_dt_obj(dt), &local_flags, declare_flags);
1495
1496         /* we need only to store the overquota flags in the first lnb for
1497          * now, once we support multiple objects BRW, this code needs be
1498          * revised.
1499          */
1500         if (local_flags & QUOTA_FL_OVER_USRQUOTA)
1501                 lnb[0].lnb_flags |= OBD_BRW_OVER_USRQUOTA;
1502         if (local_flags & QUOTA_FL_OVER_GRPQUOTA)
1503                 lnb[0].lnb_flags |= OBD_BRW_OVER_GRPQUOTA;
1504         if (local_flags & QUOTA_FL_OVER_PRJQUOTA)
1505                 lnb[0].lnb_flags |= OBD_BRW_OVER_PRJQUOTA;
1506
1507         if (rc == 0)
1508                 rc = osd_trunc_lock(osd_dt_obj(dt), oh, true);
1509
1510         RETURN(rc);
1511 }
1512
1513 /* Check if a block is allocated or not */
1514 static int osd_write_commit(const struct lu_env *env, struct dt_object *dt,
1515                             struct niobuf_local *lnb, int npages,
1516                             struct thandle *thandle, __u64 user_size)
1517 {
1518         struct osd_thread_info *oti = osd_oti_get(env);
1519         struct osd_iobuf *iobuf = &oti->oti_iobuf;
1520         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1521         struct osd_device  *osd = osd_obj2dev(osd_dt_obj(dt));
1522         int rc = 0, i, check_credits = 0;
1523
1524         LASSERT(inode);
1525
1526         rc = osd_init_iobuf(osd, iobuf, 1, npages);
1527         if (unlikely(rc != 0))
1528                 RETURN(rc);
1529
1530         dquot_initialize(inode);
1531
1532         for (i = 0; i < npages; i++) {
1533                 if (lnb[i].lnb_rc == -ENOSPC &&
1534                     (lnb[i].lnb_flags & OBD_BRW_MAPPED)) {
1535                         /* Allow the write to proceed if overwriting an
1536                          * existing block
1537                          */
1538                         lnb[i].lnb_rc = 0;
1539                 }
1540
1541                 if (lnb[i].lnb_rc) { /* ENOSPC, network RPC error, etc. */
1542                         CDEBUG(D_INODE, "Skipping [%d] == %d\n", i,
1543                                lnb[i].lnb_rc);
1544                         LASSERT(lnb[i].lnb_page);
1545                         generic_error_remove_page(inode->i_mapping,
1546                                                   lnb[i].lnb_page);
1547                         continue;
1548                 }
1549
1550                 if (lnb[i].lnb_flags & OBD_BRW_DONE)
1551                         continue;
1552
1553                 if (!(lnb[i].lnb_flags & OBD_BRW_MAPPED))
1554                         check_credits = 1;
1555
1556                 LASSERT(PageLocked(lnb[i].lnb_page));
1557                 LASSERT(!PageWriteback(lnb[i].lnb_page));
1558
1559                 /*
1560                  * Since write and truncate are serialized by oo_sem, even
1561                  * partial-page truncate should not leave dirty pages in the
1562                  * page cache.
1563                  */
1564                 LASSERT(!PageDirty(lnb[i].lnb_page));
1565
1566                 SetPageUptodate(lnb[i].lnb_page);
1567
1568                 osd_iobuf_add_page(iobuf, &lnb[i]);
1569         }
1570
1571         osd_trans_exec_op(env, thandle, OSD_OT_WRITE);
1572
1573         if (OBD_FAIL_CHECK(OBD_FAIL_OST_MAPBLK_ENOSPC)) {
1574                 rc = -ENOSPC;
1575         } else if (iobuf->dr_npages > 0) {
1576                 rc = osd_ldiskfs_map_inode_pages(inode, iobuf, osd,
1577                                                  1, user_size,
1578                                                  check_credits,
1579                                                  thandle);
1580         } else {
1581                 /* no pages to write, no transno is needed */
1582                 thandle->th_local = 1;
1583         }
1584
1585         if (rc != 0 && !thandle->th_restart_tran)
1586                 osd_fini_iobuf(osd, iobuf);
1587
1588         osd_trans_exec_check(env, thandle, OSD_OT_WRITE);
1589
1590         if (unlikely(rc != 0 && !thandle->th_restart_tran)) {
1591                 /* if write fails, we should drop pages from the cache */
1592                 for (i = 0; i < npages; i++) {
1593                         if (lnb[i].lnb_page == NULL)
1594                                 continue;
1595                         if (!PagePrivate2(lnb[i].lnb_page)) {
1596                                 LASSERT(PageLocked(lnb[i].lnb_page));
1597                                 generic_error_remove_page(inode->i_mapping,
1598                                                           lnb[i].lnb_page);
1599                         }
1600                 }
1601         }
1602
1603         RETURN(rc);
1604 }
1605
1606 static int osd_read_prep(const struct lu_env *env, struct dt_object *dt,
1607                          struct niobuf_local *lnb, int npages)
1608 {
1609         struct osd_thread_info *oti = osd_oti_get(env);
1610         struct osd_iobuf *iobuf = &oti->oti_iobuf;
1611         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1612         struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1613         int rc = 0, i, cache_hits = 0, cache_misses = 0;
1614         ktime_t start, end;
1615         s64 timediff;
1616         loff_t isize;
1617
1618         LASSERT(inode);
1619
1620         rc = osd_init_iobuf(osd, iobuf, 0, npages);
1621         if (unlikely(rc != 0))
1622                 RETURN(rc);
1623
1624         isize = i_size_read(inode);
1625
1626         start = ktime_get();
1627         for (i = 0; i < npages; i++) {
1628
1629                 if (isize <= lnb[i].lnb_file_offset)
1630                         /* If there's no more data, abort early.
1631                          * lnb->lnb_rc == 0, so it's easy to detect later.
1632                          */
1633                         break;
1634
1635                 /* instead of looking if we go beyong isize, send complete
1636                  * pages all the time
1637                  */
1638                 lnb[i].lnb_rc = lnb[i].lnb_len;
1639
1640                 /* Bypass disk read if fail_loc is set properly */
1641                 if (OBD_FAIL_CHECK(OBD_FAIL_OST_FAKE_RW))
1642                         SetPageUptodate(lnb[i].lnb_page);
1643
1644                 if (PageUptodate(lnb[i].lnb_page)) {
1645                         cache_hits++;
1646                         unlock_page(lnb[i].lnb_page);
1647                 } else {
1648                         cache_misses++;
1649                         osd_iobuf_add_page(iobuf, &lnb[i]);
1650                 }
1651                 /* no need to unlock in osd_bufs_put(), the sooner page is
1652                  * unlocked, the earlier another client can access it.
1653                  * notice real unlock_page() can be called few lines
1654                  * below after osd_do_bio(). lnb is a per-thread, so it's
1655                  * fine to have PG_locked and lnb_locked inconsistent here
1656                  */
1657                 lnb[i].lnb_locked = 0;
1658         }
1659         end = ktime_get();
1660         timediff = ktime_us_delta(end, start);
1661         lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
1662
1663         if (cache_hits != 0)
1664                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_HIT,
1665                                     cache_hits);
1666         if (cache_misses != 0)
1667                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_MISS,
1668                                     cache_misses);
1669         if (cache_hits + cache_misses != 0)
1670                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_ACCESS,
1671                                     cache_hits + cache_misses);
1672
1673         if (iobuf->dr_npages) {
1674                 rc = osd_ldiskfs_map_inode_pages(inode, iobuf, osd, 0,
1675                                                  0, 0, NULL);
1676                 if (!rc)
1677                         rc = osd_do_bio(osd, inode, iobuf, 0, 0);
1678
1679                 /* IO stats will be done in osd_bufs_put() */
1680
1681                 /* early release to let others read data during the bulk */
1682                 for (i = 0; i < iobuf->dr_npages; i++) {
1683                         LASSERT(PageLocked(iobuf->dr_pages[i]));
1684                         if (!PagePrivate2(iobuf->dr_pages[i]))
1685                                 unlock_page(iobuf->dr_pages[i]);
1686                 }
1687         }
1688
1689         RETURN(rc);
1690 }
1691
1692 /*
1693  * XXX: Another layering violation for now.
1694  *
1695  * We don't want to use ->f_op->read methods, because generic file write
1696  *
1697  *         - serializes on ->i_sem, and
1698  *
1699  *         - does a lot of extra work like balance_dirty_pages(),
1700  *
1701  * which doesn't work for globally shared files like /last_rcvd.
1702  */
1703 static int osd_ldiskfs_readlink(struct inode *inode, char *buffer, int buflen)
1704 {
1705         struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
1706
1707         memcpy(buffer, (char *)ei->i_data, buflen);
1708
1709         return  buflen;
1710 }
1711
1712 int osd_ldiskfs_read(struct inode *inode, void *buf, int size, loff_t *offs)
1713 {
1714         struct buffer_head *bh;
1715         unsigned long block;
1716         int osize;
1717         int blocksize;
1718         int csize;
1719         int boffs;
1720
1721         /* prevent reading after eof */
1722         spin_lock(&inode->i_lock);
1723         if (i_size_read(inode) < *offs + size) {
1724                 loff_t diff = i_size_read(inode) - *offs;
1725
1726                 spin_unlock(&inode->i_lock);
1727                 if (diff < 0) {
1728                         CDEBUG(D_OTHER,
1729                                "size %llu is too short to read @%llu\n",
1730                                i_size_read(inode), *offs);
1731                         return -EBADR;
1732                 } else if (diff == 0) {
1733                         return 0;
1734                 } else {
1735                         size = diff;
1736                 }
1737         } else {
1738                 spin_unlock(&inode->i_lock);
1739         }
1740
1741         blocksize = 1 << inode->i_blkbits;
1742         osize = size;
1743         while (size > 0) {
1744                 block = *offs >> inode->i_blkbits;
1745                 boffs = *offs & (blocksize - 1);
1746                 csize = min(blocksize - boffs, size);
1747                 bh = __ldiskfs_bread(NULL, inode, block, 0);
1748                 if (IS_ERR(bh)) {
1749                         CERROR("%s: can't read %u@%llu on ino %lu: rc = %ld\n",
1750                                osd_ino2name(inode), csize, *offs, inode->i_ino,
1751                                PTR_ERR(bh));
1752                         return PTR_ERR(bh);
1753                 }
1754
1755                 if (bh != NULL) {
1756                         memcpy(buf, bh->b_data + boffs, csize);
1757                         brelse(bh);
1758                 } else {
1759                         memset(buf, 0, csize);
1760                 }
1761
1762                 *offs += csize;
1763                 buf += csize;
1764                 size -= csize;
1765         }
1766         return osize;
1767 }
1768
1769 static ssize_t osd_read(const struct lu_env *env, struct dt_object *dt,
1770                         struct lu_buf *buf, loff_t *pos)
1771 {
1772         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1773         int rc;
1774
1775         /* Read small symlink from inode body as we need to maintain correct
1776          * on-disk symlinks for ldiskfs.
1777          */
1778         if (S_ISLNK(dt->do_lu.lo_header->loh_attr)) {
1779                 loff_t size = i_size_read(inode);
1780
1781                 if (buf->lb_len < size)
1782                         return -EOVERFLOW;
1783
1784                 if (size < sizeof(LDISKFS_I(inode)->i_data))
1785                         rc = osd_ldiskfs_readlink(inode, buf->lb_buf, size);
1786                 else
1787                         rc = osd_ldiskfs_read(inode, buf->lb_buf, size, pos);
1788         } else {
1789                 rc = osd_ldiskfs_read(inode, buf->lb_buf, buf->lb_len, pos);
1790         }
1791
1792         return rc;
1793 }
1794
1795 static inline int osd_extents_enabled(struct super_block *sb,
1796                                       struct inode *inode)
1797 {
1798         if (inode != NULL) {
1799                 if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL)
1800                         return 1;
1801         } else if (ldiskfs_has_feature_extents(sb)) {
1802                 return 1;
1803         }
1804         return 0;
1805 }
1806
1807 int osd_calc_bkmap_credits(struct super_block *sb, struct inode *inode,
1808                            const loff_t size, const loff_t pos,
1809                            const int blocks)
1810 {
1811         int credits, bits, bs, i;
1812
1813         bits = sb->s_blocksize_bits;
1814         bs = 1 << bits;
1815
1816         /* legacy blockmap: 3 levels * 3 (bitmap,gd,itself)
1817          * we do not expect blockmaps on the large files,
1818          * so let's shrink it to 2 levels (4GB files)
1819          */
1820
1821         /* this is default reservation: 2 levels */
1822         credits = (blocks + 2) * 3;
1823
1824         /* actual offset is unknown, hard to optimize */
1825         if (pos == -1)
1826                 return credits;
1827
1828         /* now check for few specific cases to optimize */
1829         if (pos + size <= LDISKFS_NDIR_BLOCKS * bs) {
1830                 /* no indirects */
1831                 credits = blocks;
1832                 /* allocate if not allocated */
1833                 if (inode == NULL) {
1834                         credits += blocks * 2;
1835                         return credits;
1836                 }
1837                 for (i = (pos >> bits); i < (pos >> bits) + blocks; i++) {
1838                         LASSERT(i < LDISKFS_NDIR_BLOCKS);
1839                         if (LDISKFS_I(inode)->i_data[i] == 0)
1840                                 credits += 2;
1841                 }
1842         } else if (pos + size <= (LDISKFS_NDIR_BLOCKS + 1024) * bs) {
1843                 /* single indirect */
1844                 credits = blocks * 3;
1845                 if (inode == NULL ||
1846                     LDISKFS_I(inode)->i_data[LDISKFS_IND_BLOCK] == 0)
1847                         credits += 3;
1848                 else
1849                         /* The indirect block may be modified. */
1850                         credits += 1;
1851         }
1852
1853         return credits;
1854 }
1855
1856 static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt,
1857                                  const struct lu_buf *buf, loff_t _pos,
1858                                  struct thandle *handle)
1859 {
1860         struct osd_object  *obj  = osd_dt_obj(dt);
1861         struct inode       *inode = obj->oo_inode;
1862         struct super_block *sb = osd_sb(osd_obj2dev(obj));
1863         struct osd_thandle *oh;
1864         int                 rc = 0, est = 0, credits, blocks, allocated = 0;
1865         int                 bits, bs;
1866         int                 depth, size;
1867         loff_t              pos;
1868         ENTRY;
1869
1870         LASSERT(buf != NULL);
1871         LASSERT(handle != NULL);
1872
1873         oh = container_of(handle, struct osd_thandle, ot_super);
1874         LASSERT(oh->ot_handle == NULL);
1875
1876         size = buf->lb_len;
1877         bits = sb->s_blocksize_bits;
1878         bs = 1 << bits;
1879
1880         if (_pos == -1) {
1881                 /* if this is an append, then we
1882                  * should expect cross-block record
1883                  */
1884                 pos = 0;
1885         } else {
1886                 pos = _pos;
1887         }
1888
1889         /* blocks to modify */
1890         blocks = ((pos + size + bs - 1) >> bits) - (pos >> bits);
1891         LASSERT(blocks > 0);
1892
1893         if (inode != NULL && _pos != -1) {
1894                 /* object size in blocks */
1895                 est = (i_size_read(inode) + bs - 1) >> bits;
1896                 allocated = inode->i_blocks >> (bits - 9);
1897                 if (pos + size <= i_size_read(inode) && est <= allocated) {
1898                         /* looks like an overwrite, no need to modify tree */
1899                         credits = blocks;
1900                         /* no need to modify i_size */
1901                         goto out;
1902                 }
1903         }
1904
1905         if (osd_extents_enabled(sb, inode)) {
1906                 /*
1907                  * many concurrent threads may grow tree by the time
1908                  * our transaction starts. so, consider 2 is a min depth
1909                  * for every level we may need to allocate a new block
1910                  * and take some entries from the old one. so, 3 blocks
1911                  * to allocate (bitmap, gd, itself) + old block - 4 per
1912                  * level.
1913                  */
1914                 depth = inode != NULL ? ext_depth(inode) : 0;
1915                 depth = max(depth, 1) + 1;
1916                 credits = depth;
1917                 /* if not append, then split may need to modify
1918                  * existing blocks moving entries into the new ones
1919                  */
1920                 if (_pos != -1)
1921                         credits += depth;
1922                 /* blocks to store data: bitmap,gd,itself */
1923                 credits += blocks * 3;
1924         } else {
1925                 credits = osd_calc_bkmap_credits(sb, inode, size, _pos, blocks);
1926         }
1927         /* if inode is created as part of the transaction,
1928          * then it's counted already by the creation method
1929          */
1930         if (inode != NULL)
1931                 credits++;
1932
1933 out:
1934
1935         osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
1936
1937         /* dt_declare_write() is usually called for system objects, such
1938          * as llog or last_rcvd files. We needn't enforce quota on those
1939          * objects, so always set the lqi_space as 0.
1940          */
1941         if (inode != NULL)
1942                 rc = osd_declare_inode_qid(env, i_uid_read(inode),
1943                                            i_gid_read(inode),
1944                                            i_projid_read(inode), 0,
1945                                            oh, obj, NULL, OSD_QID_BLK);
1946
1947         if (rc == 0)
1948                 rc = osd_trunc_lock(obj, oh, true);
1949
1950         RETURN(rc);
1951 }
1952
1953 static int osd_ldiskfs_writelink(struct inode *inode, char *buffer, int buflen)
1954 {
1955         /* LU-2634: clear the extent format for fast symlink */
1956         ldiskfs_clear_inode_flag(inode, LDISKFS_INODE_EXTENTS);
1957
1958         memcpy((char *)&LDISKFS_I(inode)->i_data, (char *)buffer, buflen);
1959         spin_lock(&inode->i_lock);
1960         LDISKFS_I(inode)->i_disksize = buflen;
1961         i_size_write(inode, buflen);
1962         spin_unlock(&inode->i_lock);
1963         osd_dirty_inode(inode, I_DIRTY_DATASYNC);
1964
1965         return 0;
1966 }
1967
1968 static int osd_ldiskfs_write_record(struct dt_object *dt, void *buf,
1969                                     int bufsize, int write_NUL, loff_t *offs,
1970                                     handle_t *handle)
1971 {
1972         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1973         struct buffer_head *bh        = NULL;
1974         loff_t              offset    = *offs;
1975         loff_t              new_size  = i_size_read(inode);
1976         unsigned long       block;
1977         int                 blocksize = 1 << inode->i_blkbits;
1978         struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
1979         int                 err = 0;
1980         int                 size;
1981         int                 boffs;
1982         int                 dirty_inode = 0;
1983         bool create, sparse, sync = false;
1984
1985         if (write_NUL) {
1986                 /*
1987                  * long symlink write does not count the NUL terminator in
1988                  * bufsize, we write it, and the inode's file size does not
1989                  * count the NUL terminator as well.
1990                  */
1991                 ((char *)buf)[bufsize] = '\0';
1992                 ++bufsize;
1993         }
1994
1995         /* only the first flag-set matters */
1996         dirty_inode = !test_and_set_bit(LDISKFS_INODE_JOURNAL_DATA,
1997                                        &ei->i_flags);
1998
1999         /* sparse checking is racy, but sparse is very rare case, leave as is */
2000         sparse = (new_size > 0 && (inode->i_blocks >> (inode->i_blkbits - 9)) <
2001                   ((new_size - 1) >> inode->i_blkbits) + 1);
2002
2003         while (bufsize > 0) {
2004                 int credits = handle->h_buffer_credits;
2005                 unsigned long last_block = (new_size == 0) ? 0 :
2006                                            (new_size - 1) >> inode->i_blkbits;
2007
2008                 if (bh)
2009                         brelse(bh);
2010
2011                 block = offset >> inode->i_blkbits;
2012                 boffs = offset & (blocksize - 1);
2013                 size = min(blocksize - boffs, bufsize);
2014                 sync = (block > last_block || new_size == 0 || sparse);
2015
2016                 if (sync)
2017                         down(&ei->i_append_sem);
2018
2019                 bh = __ldiskfs_bread(handle, inode, block, 0);
2020
2021                 if (unlikely(IS_ERR_OR_NULL(bh) && !sync))
2022                         CWARN(
2023                               "%s: adding bh without locking off %llu (block %lu, size %d, offs %llu)\n",
2024                               osd_ino2name(inode),
2025                               offset, block, bufsize, *offs);
2026
2027                 if (IS_ERR_OR_NULL(bh)) {
2028                         struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
2029                         int flags = LDISKFS_GET_BLOCKS_CREATE;
2030
2031                         /* while the file system is being mounted, avoid
2032                          * preallocation otherwise mount can take a long
2033                          * time as mballoc cache is cold.
2034                          * XXX: this is a workaround until we have a proper
2035                          *      fix in mballoc
2036                          * XXX: works with extent-based files only */
2037                         if (!osd->od_cl_seq)
2038                                 flags |= LDISKFS_GET_BLOCKS_NO_NORMALIZE;
2039                         bh = __ldiskfs_bread(handle, inode, block, flags);
2040                         create = true;
2041                 } else {
2042                         if (sync) {
2043                                 up(&ei->i_append_sem);
2044                                 sync = false;
2045                         }
2046                         create = false;
2047                 }
2048                 if (IS_ERR_OR_NULL(bh)) {
2049                         if (bh == NULL) {
2050                                 err = -EIO;
2051                         } else {
2052                                 err = PTR_ERR(bh);
2053                                 bh = NULL;
2054                         }
2055
2056                         CERROR(
2057                                "%s: error reading offset %llu (block %lu, size %d, offs %llu), credits %d/%d: rc = %d\n",
2058                                osd_ino2name(inode), offset, block, bufsize,
2059                                *offs, credits, handle->h_buffer_credits, err);
2060                         break;
2061                 }
2062
2063                 err = ldiskfs_journal_get_write_access(handle, bh);
2064                 if (err) {
2065                         CERROR("journal_get_write_access() returned error %d\n",
2066                                err);
2067                         break;
2068                 }
2069                 LASSERTF(boffs + size <= bh->b_size,
2070                          "boffs %d size %d bh->b_size %lu\n",
2071                          boffs, size, (unsigned long)bh->b_size);
2072                 if (create) {
2073                         memset(bh->b_data, 0, bh->b_size);
2074                         if (sync) {
2075                                 up(&ei->i_append_sem);
2076                                 sync = false;
2077                         }
2078                 }
2079                 memcpy(bh->b_data + boffs, buf, size);
2080                 err = ldiskfs_handle_dirty_metadata(handle, NULL, bh);
2081                 if (err)
2082                         break;
2083
2084                 if (offset + size > new_size)
2085                         new_size = offset + size;
2086                 offset += size;
2087                 bufsize -= size;
2088                 buf += size;
2089         }
2090         if (sync)
2091                 up(&ei->i_append_sem);
2092
2093         if (bh)
2094                 brelse(bh);
2095
2096         if (write_NUL)
2097                 --new_size;
2098         /* correct in-core and on-disk sizes */
2099         if (new_size > i_size_read(inode)) {
2100                 spin_lock(&inode->i_lock);
2101                 if (new_size > i_size_read(inode))
2102                         i_size_write(inode, new_size);
2103                 if (i_size_read(inode) > ei->i_disksize) {
2104                         ei->i_disksize = i_size_read(inode);
2105                         dirty_inode = 1;
2106                 }
2107                 spin_unlock(&inode->i_lock);
2108         }
2109         if (dirty_inode)
2110                 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
2111
2112         if (err == 0)
2113                 *offs = offset;
2114         return err;
2115 }
2116
2117 static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt,
2118                          const struct lu_buf *buf, loff_t *pos,
2119                          struct thandle *handle)
2120 {
2121         struct inode            *inode = osd_dt_obj(dt)->oo_inode;
2122         struct osd_thandle      *oh;
2123         ssize_t                 result;
2124         int                     is_link;
2125
2126         LASSERT(dt_object_exists(dt));
2127
2128         LASSERT(handle != NULL);
2129         LASSERT(inode != NULL);
2130         dquot_initialize(inode);
2131
2132         /* XXX: don't check: one declared chunk can be used many times */
2133         /* osd_trans_exec_op(env, handle, OSD_OT_WRITE); */
2134
2135         oh = container_of(handle, struct osd_thandle, ot_super);
2136         LASSERT(oh->ot_handle->h_transaction != NULL);
2137         osd_trans_exec_op(env, handle, OSD_OT_WRITE);
2138
2139         /* Write small symlink to inode body as we need to maintain correct
2140          * on-disk symlinks for ldiskfs.
2141          * Note: the buf->lb_buf contains a NUL terminator while buf->lb_len
2142          * does not count it in.
2143          */
2144         is_link = S_ISLNK(dt->do_lu.lo_header->loh_attr);
2145         if (is_link && (buf->lb_len < sizeof(LDISKFS_I(inode)->i_data)))
2146                 result = osd_ldiskfs_writelink(inode, buf->lb_buf, buf->lb_len);
2147         else
2148                 result = osd_ldiskfs_write_record(dt, buf->lb_buf, buf->lb_len,
2149                                                   is_link, pos, oh->ot_handle);
2150         if (result == 0)
2151                 result = buf->lb_len;
2152
2153         osd_trans_exec_check(env, handle, OSD_OT_WRITE);
2154
2155         return result;
2156 }
2157
2158 static int osd_declare_fallocate(const struct lu_env *env,
2159                                  struct dt_object *dt, __u64 start, __u64 end,
2160                                  int mode, struct thandle *th)
2161 {
2162         struct osd_thandle *oh = container_of(th, struct osd_thandle, ot_super);
2163         struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
2164         struct inode *inode = osd_dt_obj(dt)->oo_inode;
2165         long long quota_space = 0;
2166         /* 5 is max tree depth. (inode + 4 index blocks) */
2167         int depth = 5;
2168         int rc;
2169
2170         ENTRY;
2171
2172         /*
2173          * mode == 0 (which is standard prealloc) and PUNCH is supported
2174          * Rest of mode options is not supported yet.
2175          */
2176         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2177                 RETURN(-EOPNOTSUPP);
2178
2179         /* disable fallocate completely */
2180         if (osd_dev(dt->do_lu.lo_dev)->od_fallocate_zero_blocks < 0)
2181                 RETURN(-EOPNOTSUPP);
2182
2183         LASSERT(th);
2184         LASSERT(inode);
2185
2186         if (mode & FALLOC_FL_PUNCH_HOLE) {
2187                 rc = osd_declare_inode_qid(env, i_uid_read(inode),
2188                                            i_gid_read(inode),
2189                                            i_projid_read(inode), 0, oh,
2190                                            osd_dt_obj(dt), NULL, OSD_QID_BLK);
2191                 if (rc == 0)
2192                         rc = osd_trunc_lock(osd_dt_obj(dt), oh, false);
2193                 RETURN(rc);
2194         }
2195
2196         /* quota space for metadata blocks
2197          * approximate metadata estimate should be good enough.
2198          */
2199         quota_space += PAGE_SIZE;
2200         quota_space += depth * LDISKFS_BLOCK_SIZE(osd_sb(osd));
2201
2202         /* quota space should be reported in 1K blocks */
2203         quota_space = toqb(quota_space) + toqb(end - start) +
2204                       LDISKFS_META_TRANS_BLOCKS(inode->i_sb);
2205
2206         /* We don't need to reserve credits for whole fallocate here.
2207          * We reserve space only for metadata. Fallocate credits are
2208          * extended as required
2209          */
2210         rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
2211                                    i_projid_read(inode), quota_space, oh,
2212                                    osd_dt_obj(dt), NULL, OSD_QID_BLK);
2213         RETURN(rc);
2214 }
2215
2216 static int osd_fallocate_preallocate(const struct lu_env *env,
2217                                      struct dt_object *dt,
2218                                      __u64 start, __u64 end, int mode,
2219                                      struct thandle *th)
2220 {
2221         struct osd_thandle *oh = container_of(th, struct osd_thandle, ot_super);
2222         handle_t *handle = ldiskfs_journal_current_handle();
2223         unsigned int save_credits = oh->ot_credits;
2224         struct osd_object *obj = osd_dt_obj(dt);
2225         struct inode *inode = obj->oo_inode;
2226         struct ldiskfs_map_blocks map;
2227         unsigned int credits;
2228         ldiskfs_lblk_t blen;
2229         ldiskfs_lblk_t boff;
2230         loff_t new_size = 0;
2231         int depth = 0;
2232         int flags;
2233         int rc = 0;
2234
2235         ENTRY;
2236
2237         LASSERT(dt_object_exists(dt));
2238         LASSERT(osd_invariant(obj));
2239         LASSERT(inode != NULL);
2240
2241         CDEBUG(D_INODE, "fallocate: inode #%lu: start %llu end %llu mode %d\n",
2242                inode->i_ino, start, end, mode);
2243
2244         dquot_initialize(inode);
2245
2246         LASSERT(th);
2247
2248         boff = start >> inode->i_blkbits;
2249         blen = (ALIGN(end, 1 << inode->i_blkbits) >> inode->i_blkbits) - boff;
2250
2251         /* Create and mark new extents as either zero or unwritten */
2252         flags = osd_dev(dt->do_lu.lo_dev)->od_fallocate_zero_blocks ?
2253                 LDISKFS_GET_BLOCKS_CREATE_ZERO :
2254                 LDISKFS_GET_BLOCKS_CREATE_UNWRIT_EXT;
2255         if (mode & FALLOC_FL_KEEP_SIZE)
2256                 flags |= LDISKFS_GET_BLOCKS_KEEP_SIZE;
2257
2258         inode_lock(inode);
2259
2260         /*
2261          * We only support preallocation for extent-based file only.
2262          */
2263         if (!(ldiskfs_test_inode_flag(inode, LDISKFS_INODE_EXTENTS)))
2264                 GOTO(out, rc = -EOPNOTSUPP);
2265
2266         if (!(mode & FALLOC_FL_KEEP_SIZE) && (end > i_size_read(inode) ||
2267             end > LDISKFS_I(inode)->i_disksize)) {
2268                 new_size = end;
2269                 rc = inode_newsize_ok(inode, new_size);
2270                 if (rc)
2271                         GOTO(out, rc);
2272         }
2273
2274         inode_dio_wait(inode);
2275
2276         map.m_lblk = boff;
2277         map.m_len = blen;
2278
2279         /* Don't normalize the request if it can fit in one extent so
2280          * that it doesn't get unnecessarily split into multiple extents.
2281          */
2282         if (blen <= EXT_UNWRITTEN_MAX_LEN)
2283                 flags |= LDISKFS_GET_BLOCKS_NO_NORMALIZE;
2284
2285         /*
2286          * credits to insert 1 extent into extent tree.
2287          */
2288         credits = osd_chunk_trans_blocks(inode, blen);
2289         depth = ext_depth(inode);
2290
2291         while (rc >= 0 && blen) {
2292                 loff_t epos;
2293
2294                 /*
2295                  * Recalculate credits when extent tree depth changes.
2296                  */
2297                 if (depth != ext_depth(inode)) {
2298                         credits = osd_chunk_trans_blocks(inode, blen);
2299                         depth = ext_depth(inode);
2300                 }
2301
2302                 /* TODO: quota check */
2303                 rc = osd_extend_restart_trans(handle, credits, inode);
2304                 if (rc)
2305                         break;
2306
2307                 rc = ldiskfs_map_blocks(handle, inode, &map, flags);
2308                 if (rc <= 0) {
2309                         CDEBUG(D_INODE,
2310                                "inode #%lu: block %u: len %u: ldiskfs_map_blocks returned %d\n",
2311                                inode->i_ino, map.m_lblk, map.m_len, rc);
2312                         ldiskfs_mark_inode_dirty(handle, inode);
2313                         break;
2314                 }
2315
2316                 map.m_lblk += rc;
2317                 map.m_len = blen = blen - rc;
2318                 epos = (loff_t)map.m_lblk << inode->i_blkbits;
2319                 inode->i_ctime = current_time(inode);
2320                 if (new_size) {
2321                         if (epos > end)
2322                                 epos = end;
2323                         if (ldiskfs_update_inode_size(inode, epos) & 0x1)
2324                                 inode->i_mtime = inode->i_ctime;
2325                 } else {
2326                         if (epos > inode->i_size)
2327                                 ldiskfs_set_inode_flag(inode,
2328                                                        LDISKFS_INODE_EOFBLOCKS);
2329                 }
2330
2331                 ldiskfs_mark_inode_dirty(handle, inode);
2332         }
2333
2334 out:
2335         /* extand credits if needed for operations such as attribute set */
2336         if (rc >= 0)
2337                 rc = osd_extend_restart_trans(handle, save_credits, inode);
2338
2339         inode_unlock(inode);
2340
2341         RETURN(rc);
2342 }
2343
2344 static int osd_fallocate_punch(const struct lu_env *env, struct dt_object *dt,
2345                                __u64 start, __u64 end, int mode,
2346                                struct thandle *th)
2347 {
2348         struct osd_object *obj = osd_dt_obj(dt);
2349         struct inode *inode = obj->oo_inode;
2350         struct osd_access_lock *al;
2351         struct osd_thandle *oh;
2352         int rc = 0, found = 0;
2353
2354         ENTRY;
2355
2356         LASSERT(dt_object_exists(dt));
2357         LASSERT(osd_invariant(obj));
2358         LASSERT(inode != NULL);
2359
2360         dquot_initialize(inode);
2361
2362         LASSERT(th);
2363         oh = container_of(th, struct osd_thandle, ot_super);
2364         LASSERT(oh->ot_handle->h_transaction != NULL);
2365
2366         list_for_each_entry(al, &oh->ot_trunc_locks, tl_list) {
2367                 if (obj != al->tl_obj)
2368                         continue;
2369                 LASSERT(al->tl_shared == 0);
2370                 found = 1;
2371                 /* do actual punch in osd_trans_stop() */
2372                 al->tl_start = start;
2373                 al->tl_end = end;
2374                 al->tl_mode = mode;
2375                 al->tl_punch = true;
2376                 break;
2377         }
2378
2379         RETURN(rc);
2380 }
2381
2382 static int osd_fallocate(const struct lu_env *env, struct dt_object *dt,
2383                          __u64 start, __u64 end, int mode, struct thandle *th)
2384 {
2385         int rc;
2386
2387         ENTRY;
2388
2389         if (mode & FALLOC_FL_PUNCH_HOLE) {
2390                 /* punch */
2391                 rc = osd_fallocate_punch(env, dt, start, end, mode, th);
2392         } else {
2393                 /* standard preallocate */
2394                 rc = osd_fallocate_preallocate(env, dt, start, end, mode, th);
2395         }
2396         RETURN(rc);
2397 }
2398
2399 static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt,
2400                              __u64 start, __u64 end, struct thandle *th)
2401 {
2402         struct osd_thandle *oh;
2403         struct inode       *inode;
2404         int                 rc;
2405         ENTRY;
2406
2407         LASSERT(th);
2408         oh = container_of(th, struct osd_thandle, ot_super);
2409
2410         /*
2411          * we don't need to reserve credits for whole truncate
2412          * it's not possible as truncate may need to free too many
2413          * blocks and that won't fit a single transaction. instead
2414          * we reserve credits to change i_size and put inode onto
2415          * orphan list. if needed truncate will extend or restart
2416          * transaction
2417          */
2418         osd_trans_declare_op(env, oh, OSD_OT_PUNCH,
2419                              osd_dto_credits_noquota[DTO_ATTR_SET_BASE] + 3);
2420
2421         inode = osd_dt_obj(dt)->oo_inode;
2422         LASSERT(inode);
2423
2424         rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
2425                                    i_projid_read(inode), 0, oh, osd_dt_obj(dt),
2426                                    NULL, OSD_QID_BLK);
2427
2428         if (rc == 0)
2429                 rc = osd_trunc_lock(osd_dt_obj(dt), oh, false);
2430
2431         RETURN(rc);
2432 }
2433
2434 static int osd_punch(const struct lu_env *env, struct dt_object *dt,
2435                      __u64 start, __u64 end, struct thandle *th)
2436 {
2437         struct osd_object *obj = osd_dt_obj(dt);
2438         struct osd_device *osd = osd_obj2dev(obj);
2439         struct inode *inode = obj->oo_inode;
2440         struct osd_access_lock *al;
2441         struct osd_thandle *oh;
2442         int rc = 0, found = 0;
2443         bool grow = false;
2444         ENTRY;
2445
2446         LASSERT(dt_object_exists(dt));
2447         LASSERT(osd_invariant(obj));
2448         LASSERT(inode != NULL);
2449         dquot_initialize(inode);
2450
2451         LASSERT(th);
2452         oh = container_of(th, struct osd_thandle, ot_super);
2453         LASSERT(oh->ot_handle->h_transaction != NULL);
2454
2455         /* we used to skip truncate to current size to
2456          * optimize truncates on OST. with DoM we can
2457          * get attr_set to set specific size (MDS_REINT)
2458          * and then get truncate RPC which essentially
2459          * would be skipped. this is bad.. so, disable
2460          * this optimization on MDS till the client stop
2461          * to sent MDS_REINT (LU-11033) -bzzz
2462          */
2463         if (osd->od_is_ost && i_size_read(inode) == start)
2464                 RETURN(0);
2465
2466         osd_trans_exec_op(env, th, OSD_OT_PUNCH);
2467
2468         spin_lock(&inode->i_lock);
2469         if (i_size_read(inode) < start)
2470                 grow = true;
2471         i_size_write(inode, start);
2472         spin_unlock(&inode->i_lock);
2473         /* if object holds encrypted content, we need to make sure we truncate
2474          * on an encryption unit boundary, or subsequent reads will get
2475          * corrupted content
2476          */
2477         if (obj->oo_lma_flags & LUSTRE_ENCRYPT_FL &&
2478             start & ~LUSTRE_ENCRYPTION_MASK)
2479                 start = (start & LUSTRE_ENCRYPTION_MASK) +
2480                         LUSTRE_ENCRYPTION_UNIT_SIZE;
2481         ll_truncate_pagecache(inode, start);
2482
2483         /* optimize grow case */
2484         if (grow) {
2485                 osd_execute_truncate(obj);
2486                 GOTO(out, rc);
2487         }
2488
2489         inode_lock(inode);
2490         /* add to orphan list to ensure truncate completion
2491          * if this transaction succeed. ldiskfs_truncate()
2492          * will take the inode out of the list
2493          */
2494         rc = ldiskfs_orphan_add(oh->ot_handle, inode);
2495         inode_unlock(inode);
2496         if (rc != 0)
2497                 GOTO(out, rc);
2498
2499         list_for_each_entry(al, &oh->ot_trunc_locks, tl_list) {
2500                 if (obj != al->tl_obj)
2501                         continue;
2502                 LASSERT(al->tl_shared == 0);
2503                 found = 1;
2504                 /* do actual truncate in osd_trans_stop() */
2505                 al->tl_truncate = 1;
2506                 break;
2507         }
2508         LASSERT(found);
2509
2510 out:
2511         RETURN(rc);
2512 }
2513
2514 static int fiemap_check_ranges(struct inode *inode,
2515                                u64 start, u64 len, u64 *new_len)
2516 {
2517         loff_t maxbytes;
2518
2519         *new_len = len;
2520
2521         if (len == 0)
2522                 return -EINVAL;
2523
2524         if (ldiskfs_test_inode_flag(inode, LDISKFS_INODE_EXTENTS))
2525                 maxbytes = inode->i_sb->s_maxbytes;
2526         else
2527                 maxbytes = LDISKFS_SB(inode->i_sb)->s_bitmap_maxbytes;
2528
2529         if (start > maxbytes)
2530                 return -EFBIG;
2531
2532         /*
2533          * Shrink request scope to what the fs can actually handle.
2534          */
2535         if (len > maxbytes || (maxbytes - len) < start)
2536                 *new_len = maxbytes - start;
2537
2538         return 0;
2539 }
2540
2541 /* So that the fiemap access checks can't overflow on 32 bit machines. */
2542 #define FIEMAP_MAX_EXTENTS     (UINT_MAX / sizeof(struct fiemap_extent))
2543
2544 static int osd_fiemap_get(const struct lu_env *env, struct dt_object *dt,
2545                           struct fiemap *fm)
2546 {
2547         struct fiemap_extent_info fieinfo = {0, };
2548         struct inode *inode = osd_dt_obj(dt)->oo_inode;
2549         u64 len;
2550         int rc;
2551
2552         LASSERT(inode);
2553         if (inode->i_op->fiemap == NULL)
2554                 return -EOPNOTSUPP;
2555
2556         if (fm->fm_extent_count > FIEMAP_MAX_EXTENTS)
2557                 return -EINVAL;
2558
2559         rc = fiemap_check_ranges(inode, fm->fm_start, fm->fm_length, &len);
2560         if (rc)
2561                 return rc;
2562
2563         fieinfo.fi_flags = fm->fm_flags;
2564         fieinfo.fi_extents_max = fm->fm_extent_count;
2565         fieinfo.fi_extents_start = fm->fm_extents;
2566
2567         if (fieinfo.fi_flags & FIEMAP_FLAG_SYNC)
2568                 filemap_write_and_wait(inode->i_mapping);
2569
2570         rc = inode->i_op->fiemap(inode, &fieinfo, fm->fm_start, len);
2571         fm->fm_flags = fieinfo.fi_flags;
2572         fm->fm_mapped_extents = fieinfo.fi_extents_mapped;
2573
2574         return rc;
2575 }
2576
2577 static int osd_ladvise(const struct lu_env *env, struct dt_object *dt,
2578                        __u64 start, __u64 end, enum lu_ladvise_type advice)
2579 {
2580         struct osd_object *obj = osd_dt_obj(dt);
2581         int rc = 0;
2582         ENTRY;
2583
2584         switch (advice) {
2585         case LU_LADVISE_DONTNEED:
2586                 if (end)
2587                         invalidate_mapping_pages(obj->oo_inode->i_mapping,
2588                                                  start >> PAGE_SHIFT,
2589                                                  (end - 1) >> PAGE_SHIFT);
2590                 break;
2591         default:
2592                 rc = -ENOTSUPP;
2593                 break;
2594         }
2595
2596         RETURN(rc);
2597 }
2598
2599 static loff_t osd_lseek(const struct lu_env *env, struct dt_object *dt,
2600                         loff_t offset, int whence)
2601 {
2602         struct osd_object *obj = osd_dt_obj(dt);
2603         struct inode *inode = obj->oo_inode;
2604         struct file *file;
2605         loff_t result;
2606
2607         ENTRY;
2608
2609         LASSERT(dt_object_exists(dt));
2610         LASSERT(osd_invariant(obj));
2611         LASSERT(inode);
2612         LASSERT(offset >= 0);
2613
2614         file = osd_quasi_file(env, inode);
2615         result = file->f_op->llseek(file, offset, whence);
2616
2617         /*
2618          * If 'offset' is beyond end of object file then treat it as not error
2619          * but valid case for SEEK_HOLE and return 'offset' as result.
2620          * LOV will decide if it is beyond real end of file or not.
2621          */
2622         if (whence == SEEK_HOLE && result == -ENXIO)
2623                 result = offset;
2624
2625         CDEBUG(D_INFO, "seek %s from %lld: %lld\n", whence == SEEK_HOLE ?
2626                        "hole" : "data", offset, result);
2627         RETURN(result);
2628 }
2629
2630 /*
2631  * in some cases we may need declare methods for objects being created
2632  * e.g., when we create symlink
2633  */
2634 const struct dt_body_operations osd_body_ops_new = {
2635         .dbo_declare_write = osd_declare_write,
2636 };
2637
2638 const struct dt_body_operations osd_body_ops = {
2639         .dbo_read                       = osd_read,
2640         .dbo_declare_write              = osd_declare_write,
2641         .dbo_write                      = osd_write,
2642         .dbo_bufs_get                   = osd_bufs_get,
2643         .dbo_bufs_put                   = osd_bufs_put,
2644         .dbo_write_prep                 = osd_write_prep,
2645         .dbo_declare_write_commit       = osd_declare_write_commit,
2646         .dbo_write_commit               = osd_write_commit,
2647         .dbo_read_prep                  = osd_read_prep,
2648         .dbo_declare_punch              = osd_declare_punch,
2649         .dbo_punch                      = osd_punch,
2650         .dbo_fiemap_get                 = osd_fiemap_get,
2651         .dbo_ladvise                    = osd_ladvise,
2652         .dbo_declare_fallocate          = osd_declare_fallocate,
2653         .dbo_fallocate                  = osd_fallocate,
2654         .dbo_lseek                      = osd_lseek,
2655 };
2656
2657 /**
2658  * Get a truncate lock
2659  *
2660  * In order to take multi-transaction truncate out of main transaction we let
2661  * the caller grab a lock on the object passed. the lock can be shared (for
2662  * writes) and exclusive (for truncate). It's not allowed to mix truncate
2663  * and write in the same transaction handle (do not confuse with big ldiskfs
2664  * transaction containing lots of handles).
2665  * The lock must be taken at declaration.
2666  *
2667  * \param obj           object to lock
2668  * \oh                  transaction
2669  * \shared              shared or exclusive
2670  *
2671  * \retval 0            lock is granted
2672  * \retval -NOMEM       no memory to allocate lock
2673  */
2674 int osd_trunc_lock(struct osd_object *obj, struct osd_thandle *oh, bool shared)
2675 {
2676         struct osd_access_lock *al, *tmp;
2677
2678         LASSERT(obj);
2679         LASSERT(oh);
2680
2681         list_for_each_entry(tmp, &oh->ot_trunc_locks, tl_list) {
2682                 if (tmp->tl_obj != obj)
2683                         continue;
2684                 LASSERT(tmp->tl_shared == shared);
2685                 /* found same lock */
2686                 return 0;
2687         }
2688
2689         OBD_ALLOC_PTR(al);
2690         if (unlikely(al == NULL))
2691                 return -ENOMEM;
2692         al->tl_obj = obj;
2693         al->tl_truncate = false;
2694         if (shared)
2695                 down_read(&obj->oo_ext_idx_sem);
2696         else
2697                 down_write(&obj->oo_ext_idx_sem);
2698         al->tl_shared = shared;
2699         lu_object_get(&obj->oo_dt.do_lu);
2700
2701         list_add(&al->tl_list, &oh->ot_trunc_locks);
2702
2703         return 0;
2704 }
2705
2706 void osd_trunc_unlock_all(const struct lu_env *env, struct list_head *list)
2707 {
2708         struct osd_access_lock *al, *tmp;
2709
2710         list_for_each_entry_safe(al, tmp, list, tl_list) {
2711                 if (al->tl_shared)
2712                         up_read(&al->tl_obj->oo_ext_idx_sem);
2713                 else
2714                         up_write(&al->tl_obj->oo_ext_idx_sem);
2715                 osd_object_put(env, al->tl_obj);
2716                 list_del(&al->tl_list);
2717                 OBD_FREE_PTR(al);
2718         }
2719 }
2720
2721 /*
2722  * For a partial-page truncate, flush the page to disk immediately to
2723  * avoid data corruption during direct disk write.  b=17397
2724  */
2725 static void osd_partial_page_flush(struct osd_device *d, struct inode *inode,
2726                                    loff_t offset)
2727 {
2728         if (!(offset & ~PAGE_MASK))
2729                 return;
2730
2731         if (osd_use_page_cache(d)) {
2732                 filemap_fdatawrite_range(inode->i_mapping, offset, offset + 1);
2733         } else {
2734                 /* Notice we use "wait" version to ensure I/O is complete */
2735                 filemap_write_and_wait_range(inode->i_mapping, offset,
2736                                              offset + 1);
2737                 invalidate_mapping_pages(inode->i_mapping, offset >> PAGE_SHIFT,
2738                                          offset >> PAGE_SHIFT);
2739         }
2740 }
2741
2742 void osd_execute_truncate(struct osd_object *obj)
2743 {
2744         struct osd_device *d = osd_obj2dev(obj);
2745         struct inode *inode = obj->oo_inode;
2746         __u64 size;
2747
2748         /* simulate crash before (in the middle) of delayed truncate */
2749         if (OBD_FAIL_CHECK(OBD_FAIL_OSD_FAIL_AT_TRUNCATE)) {
2750                 struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
2751                 struct ldiskfs_sb_info *sbi = LDISKFS_SB(inode->i_sb);
2752
2753                 mutex_lock(&sbi->s_orphan_lock);
2754                 list_del_init(&ei->i_orphan);
2755                 mutex_unlock(&sbi->s_orphan_lock);
2756                 return;
2757         }
2758
2759         size = i_size_read(inode);
2760         inode_lock(inode);
2761         /* if object holds encrypted content, we need to make sure we truncate
2762          * on an encryption unit boundary, or block content will get corrupted
2763          */
2764         if (obj->oo_lma_flags & LUSTRE_ENCRYPT_FL &&
2765             size & ~LUSTRE_ENCRYPTION_MASK)
2766                 inode->i_size = (size & LUSTRE_ENCRYPTION_MASK) +
2767                         LUSTRE_ENCRYPTION_UNIT_SIZE;
2768         ldiskfs_truncate(inode);
2769         inode_unlock(inode);
2770         if (inode->i_size != size) {
2771                 spin_lock(&inode->i_lock);
2772                 i_size_write(inode, size);
2773                 LDISKFS_I(inode)->i_disksize = size;
2774                 spin_unlock(&inode->i_lock);
2775                 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
2776         }
2777         osd_partial_page_flush(d, inode, size);
2778 }
2779
2780 void osd_execute_punch(const struct lu_env *env, struct osd_object *obj,
2781                        loff_t start, loff_t end, int mode)
2782 {
2783         struct osd_device *d = osd_obj2dev(obj);
2784         struct inode *inode = obj->oo_inode;
2785         struct file *file = osd_quasi_file(env, inode);
2786
2787         file->f_op->fallocate(file, mode, start, end - start);
2788         osd_partial_page_flush(d, inode, start);
2789         osd_partial_page_flush(d, inode, end - 1);
2790 }
2791
2792 void osd_process_truncates(const struct lu_env *env, struct list_head *list)
2793 {
2794         struct osd_access_lock *al;
2795
2796         LASSERT(journal_current_handle() == NULL);
2797
2798         list_for_each_entry(al, list, tl_list) {
2799                 if (al->tl_shared)
2800                         continue;
2801                 if (al->tl_truncate)
2802                         osd_execute_truncate(al->tl_obj);
2803                 else if (al->tl_punch)
2804                         osd_execute_punch(env, al->tl_obj, al->tl_start,
2805                                           al->tl_end, al->tl_mode);
2806         }
2807 }