lustre/osd-ldiskfs/osd_io.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2012, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * lustre/osd/osd_io.c
  33  *
  34  * body operations
  35  *
  36  * Author: Nikita Danilov <nikita@clusterfs.com>
  37  * Author: Alex Zhuravlev <bzzz@whamcloud.com>
  38  *
  39  */
  40
  41 /* prerequisite for linux/xattr.h */
  42 #include <linux/types.h>
  43 /* prerequisite for linux/xattr.h */
  44 #include <linux/fs.h>
  45 #include <linux/mm.h>
  46 #include <linux/pagevec.h>
  47
  48 /*
  49  * struct OBD_{ALLOC,FREE}*()
  50  * OBD_FAIL_CHECK
  51  */
  52 #include <obd_support.h>
  53
  54 #include "osd_internal.h"
  55
  56 /* ext_depth() */
  57 #include <ldiskfs/ldiskfs_extents.h>
  58
  59 static inline bool osd_use_page_cache(struct osd_device *d)
  60 {
  61         /* do not use pagecache if write and read caching are disabled */
  62         if (d->od_writethrough_cache + d->od_read_cache == 0)
  63                 return false;
  64         /* use pagecache by default */
  65         return true;
  66 }
  67
  68 static int __osd_init_iobuf(struct osd_device *d, struct osd_iobuf *iobuf,
  69                             int rw, int line, int pages)
  70 {
  71         int blocks, i;
  72
  73         LASSERTF(iobuf->dr_elapsed_valid == 0,
  74                  "iobuf %p, reqs %d, rw %d, line %d\n", iobuf,
  75                  atomic_read(&iobuf->dr_numreqs), iobuf->dr_rw,
  76                  iobuf->dr_init_at);
  77         LASSERT(pages <= PTLRPC_MAX_BRW_PAGES);
  78
  79         init_waitqueue_head(&iobuf->dr_wait);
  80         atomic_set(&iobuf->dr_numreqs, 0);
  81         iobuf->dr_npages = 0;
  82         iobuf->dr_error = 0;
  83         iobuf->dr_dev = d;
  84         iobuf->dr_frags = 0;
  85         iobuf->dr_elapsed = ktime_set(0, 0);
  86         /* must be counted before, so assert */
  87         iobuf->dr_rw = rw;
  88         iobuf->dr_init_at = line;
  89
  90         blocks = pages * (PAGE_SIZE >> osd_sb(d)->s_blocksize_bits);
  91         if (iobuf->dr_bl_buf.lb_len >= blocks * sizeof(iobuf->dr_blocks[0])) {
  92                 LASSERT(iobuf->dr_pg_buf.lb_len >=
  93                         pages * sizeof(iobuf->dr_pages[0]));
  94                 return 0;
  95         }
  96
  97         /* start with 1MB for 4K blocks */
  98         i = 256;
  99         while (i <= PTLRPC_MAX_BRW_PAGES && i < pages)
 100                 i <<= 1;
 101
 102         CDEBUG(D_OTHER, "realloc %u for %u (%u) pages\n",
 103                (unsigned int)(pages * sizeof(iobuf->dr_pages[0])), i, pages);
 104         pages = i;
 105         blocks = pages * (PAGE_SIZE >> osd_sb(d)->s_blocksize_bits);
 106         iobuf->dr_max_pages = 0;
 107         CDEBUG(D_OTHER, "realloc %u for %u blocks\n",
 108                (unsigned int)(blocks * sizeof(iobuf->dr_blocks[0])), blocks);
 109
 110         lu_buf_realloc(&iobuf->dr_bl_buf, blocks * sizeof(iobuf->dr_blocks[0]));
 111         iobuf->dr_blocks = iobuf->dr_bl_buf.lb_buf;
 112         if (unlikely(iobuf->dr_blocks == NULL))
 113                 return -ENOMEM;
 114
 115         lu_buf_realloc(&iobuf->dr_pg_buf, pages * sizeof(iobuf->dr_pages[0]));
 116         iobuf->dr_pages = iobuf->dr_pg_buf.lb_buf;
 117         if (unlikely(iobuf->dr_pages == NULL))
 118                 return -ENOMEM;
 119
 120         lu_buf_realloc(&iobuf->dr_lnb_buf,
 121                        pages * sizeof(iobuf->dr_lnbs[0]));
 122         iobuf->dr_lnbs = iobuf->dr_lnb_buf.lb_buf;
 123         if (unlikely(iobuf->dr_lnbs == NULL))
 124                 return -ENOMEM;
 125
 126         iobuf->dr_max_pages = pages;
 127
 128         return 0;
 129 }
 130 #define osd_init_iobuf(dev, iobuf, rw, pages) \
 131         __osd_init_iobuf(dev, iobuf, rw, __LINE__, pages)
 132
 133 static void osd_iobuf_add_page(struct osd_iobuf *iobuf,
 134                                struct niobuf_local *lnb)
 135 {
 136         LASSERT(iobuf->dr_npages < iobuf->dr_max_pages);
 137         iobuf->dr_pages[iobuf->dr_npages] = lnb->lnb_page;
 138         iobuf->dr_lnbs[iobuf->dr_npages] = lnb;
 139         iobuf->dr_npages++;
 140 }
 141
 142 void osd_fini_iobuf(struct osd_device *d, struct osd_iobuf *iobuf)
 143 {
 144         int rw = iobuf->dr_rw;
 145
 146         if (iobuf->dr_elapsed_valid) {
 147                 iobuf->dr_elapsed_valid = 0;
 148                 LASSERT(iobuf->dr_dev == d);
 149                 LASSERT(iobuf->dr_frags > 0);
 150                 lprocfs_oh_tally(&d->od_brw_stats.hist[BRW_R_DIO_FRAGS+rw],
 151                                  iobuf->dr_frags);
 152                 lprocfs_oh_tally_log2(&d->od_brw_stats.hist[BRW_R_IO_TIME+rw],
 153                                       ktime_to_ms(iobuf->dr_elapsed));
 154         }
 155 }
 156
 157 #ifdef HAVE_BIO_ENDIO_USES_ONE_ARG
 158 static void dio_complete_routine(struct bio *bio)
 159 {
 160         int error = bio->bi_status;
 161 #else
 162 static void dio_complete_routine(struct bio *bio, int error)
 163 {
 164 #endif
 165         struct osd_iobuf *iobuf = bio->bi_private;
 166         struct bio_vec *bvl;
 167
 168         /* CAVEAT EMPTOR: possibly in IRQ context
 169          * DO NOT record procfs stats here!!!
 170          */
 171
 172         if (unlikely(iobuf == NULL)) {
 173                 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");
 174                 CERROR("bi_next: %p, bi_flags: %lx, " __stringify(bi_opf)
 175                        ": %x, bi_vcnt: %d, bi_idx: %d, bi->size: %d, bi_end_io: %p, bi_cnt: %d, bi_private: %p\n",
 176                        bio->bi_next, (unsigned long)bio->bi_flags,
 177                        (unsigned int)bio->bi_opf, bio->bi_vcnt, bio_idx(bio),
 178                        bio_sectors(bio) << 9, bio->bi_end_io,
 179                        atomic_read(&bio->__bi_cnt),
 180                        bio->bi_private);
 181                 return;
 182         }
 183
 184         /* the check is outside of the cycle for performance reason -bzzz */
 185         if (!bio_data_dir(bio)) {
 186                 DECLARE_BVEC_ITER_ALL(iter_all);
 187
 188                 bio_for_each_segment_all(bvl, bio, iter_all) {
 189                         if (likely(error == 0))
 190                                 SetPageUptodate(bvl_to_page(bvl));
 191                         LASSERT(PageLocked(bvl_to_page(bvl)));
 192                 }
 193                 atomic_dec(&iobuf->dr_dev->od_r_in_flight);
 194         } else {
 195                 atomic_dec(&iobuf->dr_dev->od_w_in_flight);
 196         }
 197
 198         /* any real error is good enough -bzzz */
 199         if (error != 0 && iobuf->dr_error == 0)
 200                 iobuf->dr_error = error;
 201
 202         /*
 203          * set dr_elapsed before dr_numreqs turns to 0, otherwise
 204          * it's possible that service thread will see dr_numreqs
 205          * is zero, but dr_elapsed is not set yet, leading to lost
 206          * data in this processing and an assertion in a subsequent
 207          * call to OSD.
 208          */
 209         if (atomic_read(&iobuf->dr_numreqs) == 1) {
 210                 ktime_t now = ktime_get();
 211
 212                 iobuf->dr_elapsed = ktime_sub(now, iobuf->dr_start_time);
 213                 iobuf->dr_elapsed_valid = 1;
 214         }
 215         if (atomic_dec_and_test(&iobuf->dr_numreqs))
 216                 wake_up(&iobuf->dr_wait);
 217
 218         /* Completed bios used to be chained off iobuf->dr_bios and freed in
 219          * filter_clear_dreq().  It was then possible to exhaust the biovec-256
 220          * mempool when serious on-disk fragmentation was encountered,
 221          * deadlocking the OST.  The bios are now released as soon as complete
 222          * so the pool cannot be exhausted while IOs are competing. b=10076
 223          */
 224         bio_put(bio);
 225 }
 226
 227 static void record_start_io(struct osd_iobuf *iobuf, int size)
 228 {
 229         struct osd_device    *osd = iobuf->dr_dev;
 230         struct obd_histogram *h = osd->od_brw_stats.hist;
 231
 232         iobuf->dr_frags++;
 233         atomic_inc(&iobuf->dr_numreqs);
 234
 235         if (iobuf->dr_rw == 0) {
 236                 atomic_inc(&osd->od_r_in_flight);
 237                 lprocfs_oh_tally(&h[BRW_R_RPC_HIST],
 238                                  atomic_read(&osd->od_r_in_flight));
 239                 lprocfs_oh_tally_log2(&h[BRW_R_DISK_IOSIZE], size);
 240         } else if (iobuf->dr_rw == 1) {
 241                 atomic_inc(&osd->od_w_in_flight);
 242                 lprocfs_oh_tally(&h[BRW_W_RPC_HIST],
 243                                  atomic_read(&osd->od_w_in_flight));
 244                 lprocfs_oh_tally_log2(&h[BRW_W_DISK_IOSIZE], size);
 245         } else {
 246                 LBUG();
 247         }
 248 }
 249
 250 static void osd_submit_bio(int rw, struct bio *bio)
 251 {
 252         LASSERTF(rw == 0 || rw == 1, "%x\n", rw);
 253 #ifdef HAVE_SUBMIT_BIO_2ARGS
 254         submit_bio(rw ? WRITE : READ, bio);
 255 #else
 256         bio->bi_opf |= rw;
 257         submit_bio(bio);
 258 #endif
 259 }
 260
 261 static int can_be_merged(struct bio *bio, sector_t sector)
 262 {
 263         if (bio == NULL)
 264                 return 0;
 265
 266         return bio_end_sector(bio) == sector ? 1 : 0;
 267 }
 268
 269 #if IS_ENABLED(CONFIG_BLK_DEV_INTEGRITY)
 270 /*
 271  * This function will change the data written, thus it should only be
 272  * used when checking data integrity feature
 273  */
 274 static void bio_integrity_fault_inject(struct bio *bio)
 275 {
 276         struct bio_vec *bvec;
 277         DECLARE_BVEC_ITER_ALL(iter_all);
 278         void *kaddr;
 279         char *addr;
 280
 281         bio_for_each_segment_all(bvec, bio, iter_all) {
 282                 struct page *page = bvec->bv_page;
 283
 284                 kaddr = kmap(page);
 285                 addr = kaddr;
 286                 *addr = ~(*addr);
 287                 kunmap(page);
 288                 break;
 289         }
 290 }
 291
 292 static int bio_dif_compare(__u16 *expected_guard_buf, void *bio_prot_buf,
 293                            unsigned int sectors, int tuple_size)
 294 {
 295         __u16 *expected_guard;
 296         __u16 *bio_guard;
 297         int i;
 298
 299         expected_guard = expected_guard_buf;
 300         for (i = 0; i < sectors; i++) {
 301                 bio_guard = (__u16 *)bio_prot_buf;
 302                 if (*bio_guard != *expected_guard) {
 303                         CERROR(
 304                                "unexpected guard tags on sector %d expected guard %u, bio guard %u, sectors %u, tuple size %d\n",
 305                                i, *expected_guard, *bio_guard, sectors,
 306                                tuple_size);
 307                         return -EIO;
 308                 }
 309                 expected_guard++;
 310                 bio_prot_buf += tuple_size;
 311         }
 312         return 0;
 313 }
 314
 315 static int osd_bio_integrity_compare(struct bio *bio, struct block_device *bdev,
 316                                      struct osd_iobuf *iobuf, int index)
 317 {
 318         struct blk_integrity *bi = bdev_get_integrity(bdev);
 319         struct bio_integrity_payload *bip = bio->bi_integrity;
 320         struct niobuf_local *lnb;
 321         unsigned short sector_size = blk_integrity_interval(bi);
 322         void *bio_prot_buf = page_address(bip->bip_vec->bv_page) +
 323                 bip->bip_vec->bv_offset;
 324         struct bio_vec *bv;
 325         sector_t sector = bio_start_sector(bio);
 326         unsigned int sectors, total;
 327         DECLARE_BVEC_ITER_ALL(iter_all);
 328         __u16 *expected_guard;
 329         int rc;
 330
 331         total = 0;
 332         bio_for_each_segment_all(bv, bio, iter_all) {
 333                 lnb = iobuf->dr_lnbs[index];
 334                 expected_guard = lnb->lnb_guards;
 335                 sectors = bv->bv_len / sector_size;
 336                 if (lnb->lnb_guard_rpc) {
 337                         rc = bio_dif_compare(expected_guard, bio_prot_buf,
 338                                              sectors, bi->tuple_size);
 339                         if (rc)
 340                                 return rc;
 341                 }
 342
 343                 sector += sectors;
 344                 bio_prot_buf += sectors * bi->tuple_size;
 345                 total += sectors * bi->tuple_size;
 346                 LASSERT(total <= bip_size(bio->bi_integrity));
 347                 index++;
 348         }
 349         return 0;
 350 }
 351
 352 static int osd_bio_integrity_handle(struct osd_device *osd, struct bio *bio,
 353                                     struct osd_iobuf *iobuf,
 354                                     int start_page_idx, bool fault_inject,
 355                                     bool integrity_enabled)
 356 {
 357         struct super_block *sb = osd_sb(osd);
 358         integrity_gen_fn *generate_fn = NULL;
 359         integrity_vrfy_fn *verify_fn = NULL;
 360         int rc;
 361
 362         ENTRY;
 363
 364         if (!integrity_enabled)
 365                 RETURN(0);
 366
 367         rc = osd_get_integrity_profile(osd, &generate_fn, &verify_fn);
 368         if (rc)
 369                 RETURN(rc);
 370
 371         rc = bio_integrity_prep_fn(bio, generate_fn, verify_fn);
 372         if (rc)
 373                 RETURN(rc);
 374
 375         /* Verify and inject fault only when writing */
 376         if (iobuf->dr_rw == 1) {
 377                 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_OST_INTEGRITY_CMP))) {
 378                         rc = osd_bio_integrity_compare(bio, sb->s_bdev, iobuf,
 379                                                        start_page_idx);
 380                         if (rc)
 381                                 RETURN(rc);
 382                 }
 383
 384                 if (unlikely(fault_inject))
 385                         bio_integrity_fault_inject(bio);
 386         }
 387
 388         RETURN(0);
 389 }
 390
 391 #ifdef HAVE_BIO_INTEGRITY_PREP_FN
 392 #  ifdef HAVE_BIO_ENDIO_USES_ONE_ARG
 393 static void dio_integrity_complete_routine(struct bio *bio)
 394 #  else
 395 static void dio_integrity_complete_routine(struct bio *bio, int error)
 396 #  endif
 397 {
 398         struct osd_bio_private *bio_private = bio->bi_private;
 399
 400         bio->bi_private = bio_private->obp_iobuf;
 401         osd_dio_complete_routine(bio, error);
 402
 403         OBD_FREE_PTR(bio_private);
 404 }
 405 #endif /* HAVE_BIO_INTEGRITY_PREP_FN */
 406 #else  /* !CONFIG_BLK_DEV_INTEGRITY */
 407 #define osd_bio_integrity_handle(osd, bio, iobuf, start_page_idx, \
 408                                  fault_inject, integrity_enabled) 0
 409 #endif /* CONFIG_BLK_DEV_INTEGRITY */
 410
 411 static int osd_bio_init(struct bio *bio, struct osd_iobuf *iobuf,
 412                         bool integrity_enabled, int start_page_idx,
 413                         struct osd_bio_private **pprivate)
 414 {
 415         ENTRY;
 416
 417         *pprivate = NULL;
 418
 419 #ifdef HAVE_BIO_INTEGRITY_PREP_FN
 420         if (integrity_enabled) {
 421                 struct osd_bio_private *bio_private = NULL;
 422
 423                 OBD_ALLOC_GFP(bio_private, sizeof(*bio_private), GFP_NOIO);
 424                 if (bio_private == NULL)
 425                         RETURN(-ENOMEM);
 426                 bio->bi_end_io = dio_integrity_complete_routine;
 427                 bio->bi_private = bio_private;
 428                 bio_private->obp_start_page_idx = start_page_idx;
 429                 bio_private->obp_iobuf = iobuf;
 430                 *pprivate = bio_private;
 431         } else
 432 #endif
 433         {
 434                 bio->bi_end_io = dio_complete_routine;
 435                 bio->bi_private = iobuf;
 436         }
 437
 438         RETURN(0);
 439 }
 440
 441 static int osd_do_bio(struct osd_device *osd, struct inode *inode,
 442                       struct osd_iobuf *iobuf)
 443 {
 444         int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
 445         struct page **pages = iobuf->dr_pages;
 446         int npages = iobuf->dr_npages;
 447         sector_t *blocks = iobuf->dr_blocks;
 448         int total_blocks = npages * blocks_per_page;
 449         struct super_block *sb = inode->i_sb;
 450         int sector_bits = sb->s_blocksize_bits - 9;
 451         unsigned int blocksize = sb->s_blocksize;
 452         struct block_device *bdev = sb->s_bdev;
 453         struct osd_bio_private *bio_private = NULL;
 454         struct bio *bio = NULL;
 455         int bio_start_page_idx;
 456         struct page *page;
 457         unsigned int page_offset;
 458         sector_t sector;
 459         int nblocks;
 460         int block_idx;
 461         int page_idx;
 462         int i;
 463         int rc = 0;
 464         bool fault_inject;
 465         bool integrity_enabled;
 466         struct blk_plug plug;
 467         ENTRY;
 468
 469         fault_inject = OBD_FAIL_CHECK(OBD_FAIL_OST_INTEGRITY_FAULT);
 470         LASSERT(iobuf->dr_npages == npages);
 471
 472         integrity_enabled = bdev_integrity_enabled(bdev, iobuf->dr_rw);
 473
 474         osd_brw_stats_update(osd, iobuf);
 475         iobuf->dr_start_time = ktime_get();
 476
 477         blk_start_plug(&plug);
 478         for (page_idx = 0, block_idx = 0;
 479              page_idx < npages;
 480              page_idx++, block_idx += blocks_per_page) {
 481                 page = pages[page_idx];
 482                 LASSERT(block_idx + blocks_per_page <= total_blocks);
 483
 484                 for (i = 0, page_offset = 0;
 485                      i < blocks_per_page;
 486                      i += nblocks, page_offset += blocksize * nblocks) {
 487                         nblocks = 1;
 488
 489                         if (blocks[block_idx + i] == 0) {  /* hole */
 490                                 LASSERTF(iobuf->dr_rw == 0,
 491                                          "page_idx %u, block_idx %u, i %u\n",
 492                                          page_idx, block_idx, i);
 493                                 memset(kmap(page) + page_offset, 0, blocksize);
 494                                 kunmap(page);
 495                                 continue;
 496                         }
 497
 498                         sector = (sector_t)blocks[block_idx + i] << sector_bits;
 499
 500                         /* Additional contiguous file blocks? */
 501                         while (i + nblocks < blocks_per_page &&
 502                                (sector + (nblocks << sector_bits)) ==
 503                                ((sector_t)blocks[block_idx + i + nblocks] <<
 504                                  sector_bits))
 505                                 nblocks++;
 506
 507                         if (bio && can_be_merged(bio, sector) &&
 508                             bio_add_page(bio, page, blocksize * nblocks,
 509                                          page_offset) != 0)
 510                                 continue;       /* added this frag OK */
 511
 512                         if (bio != NULL) {
 513                                 struct request_queue *q = bio_get_queue(bio);
 514                                 unsigned int bi_size = bio_sectors(bio) << 9;
 515
 516                                 /* Dang! I have to fragment this I/O */
 517                                 CDEBUG(D_INODE,
 518                                        "bio++ sz %d vcnt %d(%d) sectors %d(%d) psg %d(%d)\n",
 519                                        bi_size, bio->bi_vcnt, bio->bi_max_vecs,
 520                                        bio_sectors(bio),
 521                                        queue_max_sectors(q),
 522                                        osd_bio_nr_segs(bio),
 523                                        queue_max_segments(q));
 524                                 rc = osd_bio_integrity_handle(osd, bio,
 525                                         iobuf, bio_start_page_idx,
 526                                         fault_inject, integrity_enabled);
 527                                 if (rc) {
 528                                         bio_put(bio);
 529                                         goto out;
 530                                 }
 531
 532                                 record_start_io(iobuf, bi_size);
 533                                 osd_submit_bio(iobuf->dr_rw, bio);
 534                         }
 535
 536                         bio_start_page_idx = page_idx;
 537                         /* allocate new bio */
 538                         bio = bio_alloc(GFP_NOIO, min(BIO_MAX_PAGES,
 539                                                       (npages - page_idx) *
 540                                                       blocks_per_page));
 541                         if (bio == NULL) {
 542                                 CERROR("Can't allocate bio %u*%u = %u pages\n",
 543                                        (npages - page_idx), blocks_per_page,
 544                                        (npages - page_idx) * blocks_per_page);
 545                                 rc = -ENOMEM;
 546                                 goto out;
 547                         }
 548
 549                         bio_set_dev(bio, bdev);
 550                         bio_set_sector(bio, sector);
 551                         bio->bi_opf = iobuf->dr_rw ? WRITE : READ;
 552                         rc = osd_bio_init(bio, iobuf, integrity_enabled,
 553                                           bio_start_page_idx, &bio_private);
 554                         if (rc) {
 555                                 bio_put(bio);
 556                                 goto out;
 557                         }
 558
 559                         rc = bio_add_page(bio, page,
 560                                           blocksize * nblocks, page_offset);
 561                         LASSERT(rc != 0);
 562                 }
 563         }
 564
 565         if (bio != NULL) {
 566                 rc = osd_bio_integrity_handle(osd, bio, iobuf,
 567                                               bio_start_page_idx,
 568                                               fault_inject,
 569                                               integrity_enabled);
 570                 if (rc) {
 571                         bio_put(bio);
 572                         goto out;
 573                 }
 574
 575                 record_start_io(iobuf, bio_sectors(bio) << 9);
 576                 osd_submit_bio(iobuf->dr_rw, bio);
 577                 rc = 0;
 578         }
 579
 580 out:
 581         blk_finish_plug(&plug);
 582
 583         /* in order to achieve better IO throughput, we don't wait for writes
 584          * completion here. instead we proceed with transaction commit in
 585          * parallel and wait for IO completion once transaction is stopped
 586          * see osd_trans_stop() for more details -bzzz
 587          */
 588         if (iobuf->dr_rw == 0 || fault_inject) {
 589                 wait_event(iobuf->dr_wait,
 590                            atomic_read(&iobuf->dr_numreqs) == 0);
 591                 osd_fini_iobuf(osd, iobuf);
 592         }
 593
 594         if (rc == 0) {
 595                 rc = iobuf->dr_error;
 596         } else {
 597                 if (bio_private)
 598                         OBD_FREE_PTR(bio_private);
 599         }
 600
 601         RETURN(rc);
 602 }
 603
 604 static int osd_map_remote_to_local(loff_t offset, ssize_t len, int *nrpages,
 605                                    struct niobuf_local *lnb, int maxlnb)
 606 {
 607         int rc = 0;
 608         ENTRY;
 609
 610         *nrpages = 0;
 611
 612         while (len > 0) {
 613                 int poff = offset & (PAGE_SIZE - 1);
 614                 int plen = PAGE_SIZE - poff;
 615
 616                 if (*nrpages >= maxlnb) {
 617                         rc = -EOVERFLOW;
 618                         break;
 619                 }
 620
 621                 if (plen > len)
 622                         plen = len;
 623                 lnb->lnb_file_offset = offset;
 624                 lnb->lnb_page_offset = poff;
 625                 lnb->lnb_len = plen;
 626                 /* lnb->lnb_flags = rnb->rnb_flags; */
 627                 lnb->lnb_flags = 0;
 628                 lnb->lnb_page = NULL;
 629                 lnb->lnb_rc = 0;
 630                 lnb->lnb_guard_rpc = 0;
 631                 lnb->lnb_guard_disk = 0;
 632                 lnb->lnb_locked = 0;
 633
 634                 LASSERTF(plen <= len, "plen %u, len %lld\n", plen,
 635                          (long long) len);
 636                 offset += plen;
 637                 len -= plen;
 638                 lnb++;
 639                 (*nrpages)++;
 640         }
 641
 642         RETURN(rc);
 643 }
 644
 645 static struct page *osd_get_page(const struct lu_env *env, struct dt_object *dt,
 646                                  loff_t offset, gfp_t gfp_mask, bool cache)
 647 {
 648         struct osd_thread_info *oti = osd_oti_get(env);
 649         struct inode *inode = osd_dt_obj(dt)->oo_inode;
 650         struct osd_device *d = osd_obj2dev(osd_dt_obj(dt));
 651         struct page *page;
 652         int cur;
 653
 654         LASSERT(inode);
 655
 656         if (cache) {
 657                 page = find_or_create_page(inode->i_mapping,
 658                                            offset >> PAGE_SHIFT, gfp_mask);
 659
 660                 if (likely(page)) {
 661                         LASSERT(!PagePrivate2(page));
 662                         wait_on_page_writeback(page);
 663                 } else {
 664                         lprocfs_counter_add(d->od_stats, LPROC_OSD_NO_PAGE, 1);
 665                 }
 666
 667                 return page;
 668         }
 669
 670         if (inode->i_mapping->nrpages) {
 671                 /* consult with pagecache, but do not create new pages */
 672                 /* this is normally used once */
 673                 page = find_lock_page(inode->i_mapping, offset >> PAGE_SHIFT);
 674                 if (page) {
 675                         wait_on_page_writeback(page);
 676                         return page;
 677                 }
 678         }
 679
 680         LASSERT(oti->oti_dio_pages);
 681         cur = oti->oti_dio_pages_used;
 682         page = oti->oti_dio_pages[cur];
 683
 684         if (unlikely(!page)) {
 685                 LASSERT(cur < PTLRPC_MAX_BRW_PAGES);
 686                 page = alloc_page(gfp_mask);
 687                 if (!page)
 688                         return NULL;
 689                 oti->oti_dio_pages[cur] = page;
 690                 SetPagePrivate2(page);
 691                 lock_page(page);
 692         }
 693
 694         ClearPageUptodate(page);
 695         page->index = offset >> PAGE_SHIFT;
 696         oti->oti_dio_pages_used++;
 697
 698         return page;
 699 }
 700
 701 /*
 702  * there are following "locks":
 703  * journal_start
 704  * i_mutex
 705  * page lock
 706  *
 707  * osd write path:
 708  *  - lock page(s)
 709  *  - journal_start
 710  *  - truncate_sem
 711  *
 712  * ext4 vmtruncate:
 713  *  - lock pages, unlock
 714  *  - journal_start
 715  *  - lock partial page
 716  *  - i_data_sem
 717  *
 718  */
 719
 720 /**
 721  * Unlock and release pages loaded by osd_bufs_get()
 722  *
 723  * Unlock \a npages pages from \a lnb and drop the refcount on them.
 724  *
 725  * \param env           thread execution environment
 726  * \param dt            dt object undergoing IO (OSD object + methods)
 727  * \param lnb           array of pages undergoing IO
 728  * \param npages        number of pages in \a lnb
 729  *
 730  * \retval 0            always
 731  */
 732 static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt,
 733                         struct niobuf_local *lnb, int npages)
 734 {
 735         struct osd_thread_info *oti = osd_oti_get(env);
 736         struct pagevec pvec;
 737         int i;
 738
 739         ll_pagevec_init(&pvec, 0);
 740
 741         for (i = 0; i < npages; i++) {
 742                 struct page *page = lnb[i].lnb_page;
 743
 744                 if (page == NULL)
 745                         continue;
 746
 747                 /* if the page isn't cached, then reset uptodate
 748                  * to prevent reuse
 749                  */
 750                 if (PagePrivate2(page)) {
 751                         oti->oti_dio_pages_used--;
 752                 } else {
 753                         if (lnb[i].lnb_locked)
 754                                 unlock_page(page);
 755                         if (pagevec_add(&pvec, page) == 0)
 756                                 pagevec_release(&pvec);
 757                 }
 758
 759                 lnb[i].lnb_page = NULL;
 760         }
 761
 762         LASSERTF(oti->oti_dio_pages_used == 0, "%d\n", oti->oti_dio_pages_used);
 763
 764         /* Release any partial pagevec */
 765         pagevec_release(&pvec);
 766
 767         RETURN(0);
 768 }
 769
 770 /**
 771  * Load and lock pages undergoing IO
 772  *
 773  * Pages as described in the \a lnb array are fetched (from disk or cache)
 774  * and locked for IO by the caller.
 775  *
 776  * DLM locking protects us from write and truncate competing for same region,
 777  * but partial-page truncate can leave dirty pages in the cache for ldiskfs.
 778  * It's possible the writeout on a such a page is in progress when we access
 779  * it. It's also possible that during this writeout we put new (partial) data
 780  * into the page, but won't be able to proceed in filter_commitrw_write().
 781  * Therefore, just wait for writeout completion as it should be rare enough.
 782  *
 783  * \param env           thread execution environment
 784  * \param dt            dt object undergoing IO (OSD object + methods)
 785  * \param pos           byte offset of IO start
 786  * \param len           number of bytes of IO
 787  * \param lnb           array of extents undergoing IO
 788  * \param rw            read or write operation, and other flags
 789  * \param capa          capabilities
 790  *
 791  * \retval pages        (zero or more) loaded successfully
 792  * \retval -ENOMEM      on memory/page allocation error
 793  */
 794 static int osd_bufs_get(const struct lu_env *env, struct dt_object *dt,
 795                         loff_t pos, ssize_t len, struct niobuf_local *lnb,
 796                         int maxlnb, enum dt_bufs_type rw)
 797 {
 798         struct osd_thread_info *oti = osd_oti_get(env);
 799         struct osd_object *obj = osd_dt_obj(dt);
 800         struct osd_device *osd   = osd_obj2dev(obj);
 801         int npages, i, iosize, rc = 0;
 802         bool cache, write;
 803         loff_t fsize;
 804         gfp_t gfp_mask;
 805
 806         LASSERT(obj->oo_inode);
 807
 808         rc = osd_map_remote_to_local(pos, len, &npages, lnb, maxlnb);
 809         if (rc)
 810                 RETURN(rc);
 811
 812         write = rw & DT_BUFS_TYPE_WRITE;
 813
 814         fsize = lnb[npages - 1].lnb_file_offset + lnb[npages - 1].lnb_len;
 815         iosize = fsize - lnb[0].lnb_file_offset;
 816         fsize = max(fsize, i_size_read(obj->oo_inode));
 817
 818         cache = rw & DT_BUFS_TYPE_READAHEAD;
 819         if (cache)
 820                 goto bypass_checks;
 821
 822         cache = osd_use_page_cache(osd);
 823         while (cache) {
 824                 if (write) {
 825                         if (!osd->od_writethrough_cache) {
 826                                 cache = false;
 827                                 break;
 828                         }
 829                         if (iosize > osd->od_writethrough_max_iosize) {
 830                                 cache = false;
 831                                 break;
 832                         }
 833                 } else {
 834                         if (!osd->od_read_cache) {
 835                                 cache = false;
 836                                 break;
 837                         }
 838                         if (iosize > osd->od_readcache_max_iosize) {
 839                                 cache = false;
 840                                 break;
 841                         }
 842                 }
 843                 /* don't use cache on large files */
 844                 if (osd->od_readcache_max_filesize &&
 845                     fsize > osd->od_readcache_max_filesize)
 846                         cache = false;
 847                 break;
 848         }
 849
 850 bypass_checks:
 851         if (!cache && unlikely(!oti->oti_dio_pages)) {
 852                 OBD_ALLOC_PTR_ARRAY_LARGE(oti->oti_dio_pages,
 853                                           PTLRPC_MAX_BRW_PAGES);
 854                 if (!oti->oti_dio_pages)
 855                         return -ENOMEM;
 856         }
 857
 858         /* this could also try less hard for DT_BUFS_TYPE_READAHEAD pages */
 859         gfp_mask = rw & DT_BUFS_TYPE_LOCAL ? (GFP_NOFS | __GFP_HIGHMEM) :
 860                                              GFP_HIGHUSER;
 861         for (i = 0; i < npages; i++, lnb++) {
 862                 lnb->lnb_page = osd_get_page(env, dt, lnb->lnb_file_offset,
 863                                              gfp_mask, cache);
 864                 if (lnb->lnb_page == NULL)
 865                         GOTO(cleanup, rc = -ENOMEM);
 866
 867                 lnb->lnb_locked = 1;
 868         }
 869
 870 #if 0
 871         /* XXX: this version doesn't invalidate cached pages, but use them */
 872         if (!cache && write && obj->oo_inode->i_mapping->nrpages) {
 873                 /* do not allow data aliasing, invalidate pagecache */
 874                 /* XXX: can be quite expensive in mixed case */
 875                 invalidate_mapping_pages(obj->oo_inode->i_mapping,
 876                                 lnb[0].lnb_file_offset >> PAGE_SHIFT,
 877                                 lnb[npages - 1].lnb_file_offset >> PAGE_SHIFT);
 878         }
 879 #endif
 880
 881         RETURN(i);
 882
 883 cleanup:
 884         if (i > 0)
 885                 osd_bufs_put(env, dt, lnb - i, i);
 886         return rc;
 887 }
 888
 889 static int osd_ldiskfs_map_inode_pages(struct inode *inode, struct page **page,
 890                                        int pages, sector_t *blocks,
 891                                        int create)
 892 {
 893         int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
 894         int rc = 0, i = 0;
 895         struct page *fp = NULL;
 896         int clen = 0;
 897         pgoff_t max_page_index;
 898         handle_t *handle = NULL;
 899
 900         max_page_index = inode->i_sb->s_maxbytes >> PAGE_SHIFT;
 901
 902         CDEBUG(D_OTHER, "inode %lu: map %d pages from %lu\n",
 903                 inode->i_ino, pages, (*page)->index);
 904
 905         if (create) {
 906                 create = LDISKFS_GET_BLOCKS_CREATE;
 907                 handle = ldiskfs_journal_current_handle();
 908                 LASSERT(handle != NULL);
 909                 rc = osd_attach_jinode(inode);
 910                 if (rc)
 911                         return rc;
 912         }
 913         /* pages are sorted already. so, we just have to find
 914          * contig. space and process them properly
 915          */
 916         while (i < pages) {
 917                 long blen, total = 0;
 918                 struct ldiskfs_map_blocks map = { 0 };
 919
 920                 if (fp == NULL) { /* start new extent */
 921                         fp = *page++;
 922                         clen = 1;
 923                         if (++i != pages)
 924                                 continue;
 925                 } else if (fp->index + clen == (*page)->index) {
 926                         /* continue the extent */
 927                         page++;
 928                         clen++;
 929                         if (++i != pages)
 930                                 continue;
 931                 }
 932                 if (fp->index + clen >= max_page_index)
 933                         GOTO(cleanup, rc = -EFBIG);
 934                 /* process found extent */
 935                 map.m_lblk = fp->index * blocks_per_page;
 936                 map.m_len = blen = clen * blocks_per_page;
 937 cont_map:
 938                 rc = ldiskfs_map_blocks(handle, inode, &map, create);
 939                 if (rc >= 0) {
 940                         int c = 0;
 941
 942                         for (; total < blen && c < map.m_len; c++, total++) {
 943                                 if (rc == 0) {
 944                                         *(blocks + total) = 0;
 945                                         total++;
 946                                         break;
 947                                 }
 948                                 *(blocks + total) = map.m_pblk + c;
 949                                 /* unmap any possible underlying
 950                                  * metadata from the block device
 951                                  * mapping.  b=6998.
 952                                  */
 953                                 if ((map.m_flags & LDISKFS_MAP_NEW) &&
 954                                     create)
 955                                         clean_bdev_aliases(inode->i_sb->s_bdev,
 956                                                            map.m_pblk + c, 1);
 957                         }
 958                         rc = 0;
 959                 }
 960                 if (rc == 0 && total < blen) {
 961                         map.m_lblk = fp->index * blocks_per_page + total;
 962                         map.m_len = blen - total;
 963                         goto cont_map;
 964                 }
 965                 if (rc != 0)
 966                         GOTO(cleanup, rc);
 967
 968                 /* look for next extent */
 969                 fp = NULL;
 970                 blocks += blocks_per_page * clen;
 971         }
 972 cleanup:
 973         return rc;
 974 }
 975
 976 static int osd_write_prep(const struct lu_env *env, struct dt_object *dt,
 977                           struct niobuf_local *lnb, int npages)
 978 {
 979         struct osd_thread_info *oti   = osd_oti_get(env);
 980         struct osd_iobuf       *iobuf = &oti->oti_iobuf;
 981         struct inode           *inode = osd_dt_obj(dt)->oo_inode;
 982         struct osd_device      *osd   = osd_obj2dev(osd_dt_obj(dt));
 983         ktime_t start, end;
 984         s64 timediff;
 985         ssize_t isize;
 986         __s64  maxidx;
 987         int i, rc = 0;
 988
 989         LASSERT(inode);
 990
 991         rc = osd_init_iobuf(osd, iobuf, 0, npages);
 992         if (unlikely(rc != 0))
 993                 RETURN(rc);
 994
 995         isize = i_size_read(inode);
 996         maxidx = ((isize + PAGE_SIZE - 1) >> PAGE_SHIFT) - 1;
 997
 998         start = ktime_get();
 999         for (i = 0; i < npages; i++) {
1000
1001                 /*
1002                  * till commit the content of the page is undefined
1003                  * we'll set it uptodate once bulk is done. otherwise
1004                  * subsequent reads can access non-stable data
1005                  */
1006                 ClearPageUptodate(lnb[i].lnb_page);
1007
1008                 if (lnb[i].lnb_len == PAGE_SIZE)
1009                         continue;
1010
1011                 if (maxidx >= lnb[i].lnb_page->index) {
1012                         osd_iobuf_add_page(iobuf, &lnb[i]);
1013                 } else {
1014                         long off;
1015                         char *p = kmap(lnb[i].lnb_page);
1016
1017                         off = lnb[i].lnb_page_offset;
1018                         if (off)
1019                                 memset(p, 0, off);
1020                         off = (lnb[i].lnb_page_offset + lnb[i].lnb_len) &
1021                               ~PAGE_MASK;
1022                         if (off)
1023                                 memset(p + off, 0, PAGE_SIZE - off);
1024                         kunmap(lnb[i].lnb_page);
1025                 }
1026         }
1027         end = ktime_get();
1028         timediff = ktime_us_delta(end, start);
1029         lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
1030
1031         if (iobuf->dr_npages) {
1032                 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1033                                                  iobuf->dr_npages,
1034                                                  iobuf->dr_blocks, 0);
1035                 if (likely(rc == 0)) {
1036                         rc = osd_do_bio(osd, inode, iobuf);
1037                         /* do IO stats for preparation reads */
1038                         osd_fini_iobuf(osd, iobuf);
1039                 }
1040         }
1041         RETURN(rc);
1042 }
1043
1044 struct osd_fextent {
1045         sector_t        start;
1046         sector_t        end;
1047         unsigned int    mapped:1;
1048 };
1049
1050 static int osd_is_mapped(struct dt_object *dt, __u64 offset,
1051                          struct osd_fextent *cached_extent)
1052 {
1053         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1054         sector_t block = offset >> inode->i_blkbits;
1055         sector_t start;
1056         struct fiemap_extent_info fei = { 0 };
1057         struct fiemap_extent fe = { 0 };
1058         mm_segment_t saved_fs;
1059         int rc;
1060
1061         if (block >= cached_extent->start && block < cached_extent->end)
1062                 return cached_extent->mapped;
1063
1064         if (i_size_read(inode) == 0)
1065                 return 0;
1066
1067         /* Beyond EOF, must not be mapped */
1068         if (((i_size_read(inode) - 1) >> inode->i_blkbits) < block)
1069                 return 0;
1070
1071         fei.fi_extents_max = 1;
1072         fei.fi_extents_start = &fe;
1073
1074         saved_fs = get_fs();
1075         set_fs(KERNEL_DS);
1076         rc = inode->i_op->fiemap(inode, &fei, offset, FIEMAP_MAX_OFFSET-offset);
1077         set_fs(saved_fs);
1078         if (rc != 0)
1079                 return 0;
1080
1081         start = fe.fe_logical >> inode->i_blkbits;
1082
1083         if (start > block) {
1084                 cached_extent->start = block;
1085                 cached_extent->end = start;
1086                 cached_extent->mapped = 0;
1087         } else {
1088                 cached_extent->start = start;
1089                 cached_extent->end = (fe.fe_logical + fe.fe_length) >>
1090                                       inode->i_blkbits;
1091                 cached_extent->mapped = 1;
1092         }
1093
1094         return cached_extent->mapped;
1095 }
1096
1097 static int osd_declare_write_commit(const struct lu_env *env,
1098                                     struct dt_object *dt,
1099                                     struct niobuf_local *lnb, int npages,
1100                                     struct thandle *handle)
1101 {
1102         const struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1103         struct inode            *inode = osd_dt_obj(dt)->oo_inode;
1104         struct osd_thandle      *oh;
1105         int                     extents = 0;
1106         int                     depth;
1107         int                     i;
1108         int                     newblocks = 0;
1109         int                     rc = 0;
1110         int                     credits = 0;
1111         long long               quota_space = 0;
1112         struct osd_fextent      mapped = { 0 }, extent = { 0 };
1113         enum osd_quota_local_flags local_flags = 0;
1114         enum osd_qid_declare_flags declare_flags = OSD_QID_BLK;
1115         ENTRY;
1116
1117         LASSERT(handle != NULL);
1118         oh = container_of(handle, struct osd_thandle, ot_super);
1119         LASSERT(oh->ot_handle == NULL);
1120
1121         /* calculate number of extents (probably better to pass nb) */
1122         for (i = 0; i < npages; i++) {
1123                 /* ignore quota for the whole request if any page is from
1124                  * client cache or written by root.
1125                  *
1126                  * XXX once we drop the 1.8 client support, the checking
1127                  * for whether page is from cache can be simplified as:
1128                  * !(lnb[i].flags & OBD_BRW_SYNC)
1129                  *
1130                  * XXX we could handle this on per-lnb basis as done by
1131                  * grant.
1132                  */
1133                 if ((lnb[i].lnb_flags & OBD_BRW_NOQUOTA) ||
1134                     (lnb[i].lnb_flags & (OBD_BRW_FROM_GRANT | OBD_BRW_SYNC)) ==
1135                     OBD_BRW_FROM_GRANT)
1136                         declare_flags |= OSD_QID_FORCE;
1137
1138                 if (osd_is_mapped(dt, lnb[i].lnb_file_offset, &mapped)) {
1139                         lnb[i].lnb_flags |= OBD_BRW_MAPPED;
1140                         continue;
1141                 }
1142
1143                 /* count only unmapped changes */
1144                 newblocks++;
1145                 if (lnb[i].lnb_file_offset != extent.end || extent.end == 0) {
1146                         extents++;
1147                         extent.end = lnb[i].lnb_file_offset + lnb[i].lnb_len;
1148                 } else {
1149                         extent.end += lnb[i].lnb_len;
1150                 }
1151
1152                 quota_space += PAGE_SIZE;
1153         }
1154
1155         /*
1156          * each extent can go into new leaf causing a split
1157          * 5 is max tree depth: inode + 4 index blocks
1158          * with blockmaps, depth is 3 at most
1159          */
1160         if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL) {
1161                 /*
1162                  * many concurrent threads may grow tree by the time
1163                  * our transaction starts. so, consider 2 is a min depth
1164                  */
1165                 depth = ext_depth(inode);
1166                 depth = max(depth, 1) + 1;
1167                 newblocks += depth;
1168                 credits++; /* inode */
1169                 credits += depth * 2 * extents;
1170         } else {
1171                 depth = 3;
1172                 newblocks += depth;
1173                 credits++; /* inode */
1174                 credits += depth * extents;
1175         }
1176
1177         /* quota space for metadata blocks */
1178         quota_space += depth * extents * LDISKFS_BLOCK_SIZE(osd_sb(osd));
1179
1180         /* quota space should be reported in 1K blocks */
1181         quota_space = toqb(quota_space);
1182
1183         /* each new block can go in different group (bitmap + gd) */
1184
1185         /* we can't dirty more bitmap blocks than exist */
1186         if (newblocks > LDISKFS_SB(osd_sb(osd))->s_groups_count)
1187                 credits += LDISKFS_SB(osd_sb(osd))->s_groups_count;
1188         else
1189                 credits += newblocks;
1190
1191         /* we can't dirty more gd blocks than exist */
1192         if (newblocks > LDISKFS_SB(osd_sb(osd))->s_gdb_count)
1193                 credits += LDISKFS_SB(osd_sb(osd))->s_gdb_count;
1194         else
1195                 credits += newblocks;
1196
1197         osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
1198
1199         /* make sure the over quota flags were not set */
1200         lnb[0].lnb_flags &= ~OBD_BRW_OVER_ALLQUOTA;
1201
1202         rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
1203                                    i_projid_read(inode), quota_space, oh,
1204                                    osd_dt_obj(dt), &local_flags, declare_flags);
1205
1206         /* we need only to store the overquota flags in the first lnb for
1207          * now, once we support multiple objects BRW, this code needs be
1208          * revised.
1209          */
1210         if (local_flags & QUOTA_FL_OVER_USRQUOTA)
1211                 lnb[0].lnb_flags |= OBD_BRW_OVER_USRQUOTA;
1212         if (local_flags & QUOTA_FL_OVER_GRPQUOTA)
1213                 lnb[0].lnb_flags |= OBD_BRW_OVER_GRPQUOTA;
1214         if (local_flags & QUOTA_FL_OVER_PRJQUOTA)
1215                 lnb[0].lnb_flags |= OBD_BRW_OVER_PRJQUOTA;
1216
1217         if (rc == 0)
1218                 rc = osd_trunc_lock(osd_dt_obj(dt), oh, true);
1219
1220         RETURN(rc);
1221 }
1222
1223 /* Check if a block is allocated or not */
1224 static int osd_write_commit(const struct lu_env *env, struct dt_object *dt,
1225                             struct niobuf_local *lnb, int npages,
1226                             struct thandle *thandle, __u64 user_size)
1227 {
1228         struct osd_thread_info *oti = osd_oti_get(env);
1229         struct osd_iobuf *iobuf = &oti->oti_iobuf;
1230         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1231         struct osd_device  *osd = osd_obj2dev(osd_dt_obj(dt));
1232         loff_t disk_size;
1233         int rc = 0, i;
1234
1235         LASSERT(inode);
1236
1237         rc = osd_init_iobuf(osd, iobuf, 1, npages);
1238         if (unlikely(rc != 0))
1239                 RETURN(rc);
1240
1241         disk_size = i_size_read(inode);
1242         /* if disk_size is already bigger than specified user_size,
1243          * ignore user_size
1244          */
1245         if (disk_size > user_size)
1246                 user_size = 0;
1247         dquot_initialize(inode);
1248
1249         for (i = 0; i < npages; i++) {
1250                 if (lnb[i].lnb_rc == -ENOSPC &&
1251                     (lnb[i].lnb_flags & OBD_BRW_MAPPED)) {
1252                         /* Allow the write to proceed if overwriting an
1253                          * existing block
1254                          */
1255                         lnb[i].lnb_rc = 0;
1256                 }
1257
1258                 if (lnb[i].lnb_rc) { /* ENOSPC, network RPC error, etc. */
1259                         CDEBUG(D_INODE, "Skipping [%d] == %d\n", i,
1260                                lnb[i].lnb_rc);
1261                         LASSERT(lnb[i].lnb_page);
1262                         generic_error_remove_page(inode->i_mapping,
1263                                                   lnb[i].lnb_page);
1264                         continue;
1265                 }
1266
1267                 LASSERT(PageLocked(lnb[i].lnb_page));
1268                 LASSERT(!PageWriteback(lnb[i].lnb_page));
1269
1270                 if (lnb[i].lnb_file_offset + lnb[i].lnb_len > disk_size)
1271                         disk_size = lnb[i].lnb_file_offset + lnb[i].lnb_len;
1272
1273                 /*
1274                  * Since write and truncate are serialized by oo_sem, even
1275                  * partial-page truncate should not leave dirty pages in the
1276                  * page cache.
1277                  */
1278                 LASSERT(!PageDirty(lnb[i].lnb_page));
1279
1280                 SetPageUptodate(lnb[i].lnb_page);
1281
1282                 osd_iobuf_add_page(iobuf, &lnb[i]);
1283         }
1284
1285         /* if file has grown, take user_size into account */
1286         if (user_size && disk_size > user_size)
1287                 disk_size = user_size;
1288
1289         osd_trans_exec_op(env, thandle, OSD_OT_WRITE);
1290
1291         if (OBD_FAIL_CHECK(OBD_FAIL_OST_MAPBLK_ENOSPC)) {
1292                 rc = -ENOSPC;
1293         } else if (iobuf->dr_npages > 0) {
1294                 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1295                                                  iobuf->dr_npages,
1296                                                  iobuf->dr_blocks, 1);
1297         } else {
1298                 /* no pages to write, no transno is needed */
1299                 thandle->th_local = 1;
1300         }
1301
1302         if (likely(rc == 0)) {
1303                 spin_lock(&inode->i_lock);
1304                 if (disk_size > i_size_read(inode)) {
1305                         i_size_write(inode, disk_size);
1306                         LDISKFS_I(inode)->i_disksize = disk_size;
1307                         spin_unlock(&inode->i_lock);
1308                         osd_dirty_inode(inode, I_DIRTY_DATASYNC);
1309                 } else {
1310                         spin_unlock(&inode->i_lock);
1311                 }
1312
1313                 rc = osd_do_bio(osd, inode, iobuf);
1314                 /* we don't do stats here as in read path because
1315                  * write is async: we'll do this in osd_put_bufs()
1316                  */
1317         } else {
1318                 osd_fini_iobuf(osd, iobuf);
1319         }
1320
1321         osd_trans_exec_check(env, thandle, OSD_OT_WRITE);
1322
1323         if (unlikely(rc != 0)) {
1324                 /* if write fails, we should drop pages from the cache */
1325                 for (i = 0; i < npages; i++) {
1326                         if (lnb[i].lnb_page == NULL)
1327                                 continue;
1328                         if (!PagePrivate2(lnb[i].lnb_page)) {
1329                                 LASSERT(PageLocked(lnb[i].lnb_page));
1330                                 generic_error_remove_page(inode->i_mapping,
1331                                                           lnb[i].lnb_page);
1332                         }
1333                 }
1334         }
1335
1336         RETURN(rc);
1337 }
1338
1339 static int osd_read_prep(const struct lu_env *env, struct dt_object *dt,
1340                          struct niobuf_local *lnb, int npages)
1341 {
1342         struct osd_thread_info *oti = osd_oti_get(env);
1343         struct osd_iobuf *iobuf = &oti->oti_iobuf;
1344         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1345         struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1346         int rc = 0, i, cache_hits = 0, cache_misses = 0;
1347         ktime_t start, end;
1348         s64 timediff;
1349         loff_t isize;
1350
1351         LASSERT(inode);
1352
1353         rc = osd_init_iobuf(osd, iobuf, 0, npages);
1354         if (unlikely(rc != 0))
1355                 RETURN(rc);
1356
1357         isize = i_size_read(inode);
1358
1359         start = ktime_get();
1360         for (i = 0; i < npages; i++) {
1361
1362                 if (isize <= lnb[i].lnb_file_offset)
1363                         /* If there's no more data, abort early.
1364                          * lnb->lnb_rc == 0, so it's easy to detect later.
1365                          */
1366                         break;
1367
1368                 /* instead of looking if we go beyong isize, send complete
1369                  * pages all the time
1370                  */
1371                 lnb[i].lnb_rc = lnb[i].lnb_len;
1372
1373                 /* Bypass disk read if fail_loc is set properly */
1374                 if (OBD_FAIL_CHECK(OBD_FAIL_OST_FAKE_RW))
1375                         SetPageUptodate(lnb[i].lnb_page);
1376
1377                 if (PageUptodate(lnb[i].lnb_page)) {
1378                         cache_hits++;
1379                         unlock_page(lnb[i].lnb_page);
1380                 } else {
1381                         cache_misses++;
1382                         osd_iobuf_add_page(iobuf, &lnb[i]);
1383                 }
1384                 /* no need to unlock in osd_bufs_put(), the sooner page is
1385                  * unlocked, the earlier another client can access it.
1386                  * notice real unlock_page() can be called few lines
1387                  * below after osd_do_bio(). lnb is a per-thread, so it's
1388                  * fine to have PG_locked and lnb_locked inconsistent here
1389                  */
1390                 lnb[i].lnb_locked = 0;
1391         }
1392         end = ktime_get();
1393         timediff = ktime_us_delta(end, start);
1394         lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
1395
1396         if (cache_hits != 0)
1397                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_HIT,
1398                                     cache_hits);
1399         if (cache_misses != 0)
1400                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_MISS,
1401                                     cache_misses);
1402         if (cache_hits + cache_misses != 0)
1403                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_ACCESS,
1404                                     cache_hits + cache_misses);
1405
1406         if (iobuf->dr_npages) {
1407                 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1408                                                  iobuf->dr_npages,
1409                                                  iobuf->dr_blocks, 0);
1410                 rc = osd_do_bio(osd, inode, iobuf);
1411
1412                 /* IO stats will be done in osd_bufs_put() */
1413
1414                 /* early release to let others read data during the bulk */
1415                 for (i = 0; i < iobuf->dr_npages; i++) {
1416                         LASSERT(PageLocked(iobuf->dr_pages[i]));
1417                         if (!PagePrivate2(iobuf->dr_pages[i]))
1418                                 unlock_page(iobuf->dr_pages[i]);
1419                 }
1420         }
1421
1422         RETURN(rc);
1423 }
1424
1425 /*
1426  * XXX: Another layering violation for now.
1427  *
1428  * We don't want to use ->f_op->read methods, because generic file write
1429  *
1430  *         - serializes on ->i_sem, and
1431  *
1432  *         - does a lot of extra work like balance_dirty_pages(),
1433  *
1434  * which doesn't work for globally shared files like /last_rcvd.
1435  */
1436 static int osd_ldiskfs_readlink(struct inode *inode, char *buffer, int buflen)
1437 {
1438         struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
1439
1440         memcpy(buffer, (char *)ei->i_data, buflen);
1441
1442         return  buflen;
1443 }
1444
1445 int osd_ldiskfs_read(struct inode *inode, void *buf, int size, loff_t *offs)
1446 {
1447         struct buffer_head *bh;
1448         unsigned long block;
1449         int osize;
1450         int blocksize;
1451         int csize;
1452         int boffs;
1453
1454         /* prevent reading after eof */
1455         spin_lock(&inode->i_lock);
1456         if (i_size_read(inode) < *offs + size) {
1457                 loff_t diff = i_size_read(inode) - *offs;
1458
1459                 spin_unlock(&inode->i_lock);
1460                 if (diff < 0) {
1461                         CDEBUG(D_OTHER,
1462                                "size %llu is too short to read @%llu\n",
1463                                i_size_read(inode), *offs);
1464                         return -EBADR;
1465                 } else if (diff == 0) {
1466                         return 0;
1467                 } else {
1468                         size = diff;
1469                 }
1470         } else {
1471                 spin_unlock(&inode->i_lock);
1472         }
1473
1474         blocksize = 1 << inode->i_blkbits;
1475         osize = size;
1476         while (size > 0) {
1477                 block = *offs >> inode->i_blkbits;
1478                 boffs = *offs & (blocksize - 1);
1479                 csize = min(blocksize - boffs, size);
1480                 bh = __ldiskfs_bread(NULL, inode, block, 0);
1481                 if (IS_ERR(bh)) {
1482                         CERROR("%s: can't read %u@%llu on ino %lu: rc = %ld\n",
1483                                osd_ino2name(inode), csize, *offs, inode->i_ino,
1484                                PTR_ERR(bh));
1485                         return PTR_ERR(bh);
1486                 }
1487
1488                 if (bh != NULL) {
1489                         memcpy(buf, bh->b_data + boffs, csize);
1490                         brelse(bh);
1491                 } else {
1492                         memset(buf, 0, csize);
1493                 }
1494
1495                 *offs += csize;
1496                 buf += csize;
1497                 size -= csize;
1498         }
1499         return osize;
1500 }
1501
1502 static ssize_t osd_read(const struct lu_env *env, struct dt_object *dt,
1503                         struct lu_buf *buf, loff_t *pos)
1504 {
1505         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1506         int rc;
1507
1508         /* Read small symlink from inode body as we need to maintain correct
1509          * on-disk symlinks for ldiskfs.
1510          */
1511         if (S_ISLNK(dt->do_lu.lo_header->loh_attr)) {
1512                 loff_t size = i_size_read(inode);
1513
1514                 if (buf->lb_len < size)
1515                         return -EOVERFLOW;
1516
1517                 if (size < sizeof(LDISKFS_I(inode)->i_data))
1518                         rc = osd_ldiskfs_readlink(inode, buf->lb_buf, size);
1519                 else
1520                         rc = osd_ldiskfs_read(inode, buf->lb_buf, size, pos);
1521         } else {
1522                 rc = osd_ldiskfs_read(inode, buf->lb_buf, buf->lb_len, pos);
1523         }
1524
1525         return rc;
1526 }
1527
1528 static inline int osd_extents_enabled(struct super_block *sb,
1529                                       struct inode *inode)
1530 {
1531         if (inode != NULL) {
1532                 if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL)
1533                         return 1;
1534         } else if (ldiskfs_has_feature_extents(sb)) {
1535                 return 1;
1536         }
1537         return 0;
1538 }
1539
1540 int osd_calc_bkmap_credits(struct super_block *sb, struct inode *inode,
1541                            const loff_t size, const loff_t pos,
1542                            const int blocks)
1543 {
1544         int credits, bits, bs, i;
1545
1546         bits = sb->s_blocksize_bits;
1547         bs = 1 << bits;
1548
1549         /* legacy blockmap: 3 levels * 3 (bitmap,gd,itself)
1550          * we do not expect blockmaps on the large files,
1551          * so let's shrink it to 2 levels (4GB files)
1552          */
1553
1554         /* this is default reservation: 2 levels */
1555         credits = (blocks + 2) * 3;
1556
1557         /* actual offset is unknown, hard to optimize */
1558         if (pos == -1)
1559                 return credits;
1560
1561         /* now check for few specific cases to optimize */
1562         if (pos + size <= LDISKFS_NDIR_BLOCKS * bs) {
1563                 /* no indirects */
1564                 credits = blocks;
1565                 /* allocate if not allocated */
1566                 if (inode == NULL) {
1567                         credits += blocks * 2;
1568                         return credits;
1569                 }
1570                 for (i = (pos >> bits); i < (pos >> bits) + blocks; i++) {
1571                         LASSERT(i < LDISKFS_NDIR_BLOCKS);
1572                         if (LDISKFS_I(inode)->i_data[i] == 0)
1573                                 credits += 2;
1574                 }
1575         } else if (pos + size <= (LDISKFS_NDIR_BLOCKS + 1024) * bs) {
1576                 /* single indirect */
1577                 credits = blocks * 3;
1578                 if (inode == NULL ||
1579                     LDISKFS_I(inode)->i_data[LDISKFS_IND_BLOCK] == 0)
1580                         credits += 3;
1581                 else
1582                         /* The indirect block may be modified. */
1583                         credits += 1;
1584         }
1585
1586         return credits;
1587 }
1588
1589 static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt,
1590                                  const struct lu_buf *buf, loff_t _pos,
1591                                  struct thandle *handle)
1592 {
1593         struct osd_object  *obj  = osd_dt_obj(dt);
1594         struct inode       *inode = obj->oo_inode;
1595         struct super_block *sb = osd_sb(osd_obj2dev(obj));
1596         struct osd_thandle *oh;
1597         int                 rc = 0, est = 0, credits, blocks, allocated = 0;
1598         int                 bits, bs;
1599         int                 depth, size;
1600         loff_t              pos;
1601         ENTRY;
1602
1603         LASSERT(buf != NULL);
1604         LASSERT(handle != NULL);
1605
1606         oh = container_of(handle, struct osd_thandle, ot_super);
1607         LASSERT(oh->ot_handle == NULL);
1608
1609         size = buf->lb_len;
1610         bits = sb->s_blocksize_bits;
1611         bs = 1 << bits;
1612
1613         if (_pos == -1) {
1614                 /* if this is an append, then we
1615                  * should expect cross-block record
1616                  */
1617                 pos = 0;
1618         } else {
1619                 pos = _pos;
1620         }
1621
1622         /* blocks to modify */
1623         blocks = ((pos + size + bs - 1) >> bits) - (pos >> bits);
1624         LASSERT(blocks > 0);
1625
1626         if (inode != NULL && _pos != -1) {
1627                 /* object size in blocks */
1628                 est = (i_size_read(inode) + bs - 1) >> bits;
1629                 allocated = inode->i_blocks >> (bits - 9);
1630                 if (pos + size <= i_size_read(inode) && est <= allocated) {
1631                         /* looks like an overwrite, no need to modify tree */
1632                         credits = blocks;
1633                         /* no need to modify i_size */
1634                         goto out;
1635                 }
1636         }
1637
1638         if (osd_extents_enabled(sb, inode)) {
1639                 /*
1640                  * many concurrent threads may grow tree by the time
1641                  * our transaction starts. so, consider 2 is a min depth
1642                  * for every level we may need to allocate a new block
1643                  * and take some entries from the old one. so, 3 blocks
1644                  * to allocate (bitmap, gd, itself) + old block - 4 per
1645                  * level.
1646                  */
1647                 depth = inode != NULL ? ext_depth(inode) : 0;
1648                 depth = max(depth, 1) + 1;
1649                 credits = depth;
1650                 /* if not append, then split may need to modify
1651                  * existing blocks moving entries into the new ones
1652                  */
1653                 if (_pos != -1)
1654                         credits += depth;
1655                 /* blocks to store data: bitmap,gd,itself */
1656                 credits += blocks * 3;
1657         } else {
1658                 credits = osd_calc_bkmap_credits(sb, inode, size, _pos, blocks);
1659         }
1660         /* if inode is created as part of the transaction,
1661          * then it's counted already by the creation method
1662          */
1663         if (inode != NULL)
1664                 credits++;
1665
1666 out:
1667
1668         osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
1669
1670         /* dt_declare_write() is usually called for system objects, such
1671          * as llog or last_rcvd files. We needn't enforce quota on those
1672          * objects, so always set the lqi_space as 0.
1673          */
1674         if (inode != NULL)
1675                 rc = osd_declare_inode_qid(env, i_uid_read(inode),
1676                                            i_gid_read(inode),
1677                                            i_projid_read(inode), 0,
1678                                            oh, obj, NULL, OSD_QID_BLK);
1679
1680         if (rc == 0)
1681                 rc = osd_trunc_lock(obj, oh, true);
1682
1683         RETURN(rc);
1684 }
1685
1686 static int osd_ldiskfs_writelink(struct inode *inode, char *buffer, int buflen)
1687 {
1688         /* LU-2634: clear the extent format for fast symlink */
1689         ldiskfs_clear_inode_flag(inode, LDISKFS_INODE_EXTENTS);
1690
1691         memcpy((char *)&LDISKFS_I(inode)->i_data, (char *)buffer, buflen);
1692         spin_lock(&inode->i_lock);
1693         LDISKFS_I(inode)->i_disksize = buflen;
1694         i_size_write(inode, buflen);
1695         spin_unlock(&inode->i_lock);
1696         osd_dirty_inode(inode, I_DIRTY_DATASYNC);
1697
1698         return 0;
1699 }
1700
1701 static int osd_ldiskfs_write_record(struct dt_object *dt, void *buf,
1702                                     int bufsize, int write_NUL, loff_t *offs,
1703                                     handle_t *handle)
1704 {
1705         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1706         struct buffer_head *bh        = NULL;
1707         loff_t              offset    = *offs;
1708         loff_t              new_size  = i_size_read(inode);
1709         unsigned long       block;
1710         int                 blocksize = 1 << inode->i_blkbits;
1711         struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
1712         int                 err = 0;
1713         int                 size;
1714         int                 boffs;
1715         int                 dirty_inode = 0;
1716         bool create, sparse, sync = false;
1717
1718         if (write_NUL) {
1719                 /*
1720                  * long symlink write does not count the NUL terminator in
1721                  * bufsize, we write it, and the inode's file size does not
1722                  * count the NUL terminator as well.
1723                  */
1724                 ((char *)buf)[bufsize] = '\0';
1725                 ++bufsize;
1726         }
1727
1728         dirty_inode = test_and_set_bit(LDISKFS_INODE_JOURNAL_DATA,
1729                                        &ei->i_flags);
1730
1731         /* sparse checking is racy, but sparse is very rare case, leave as is */
1732         sparse = (new_size > 0 && (inode->i_blocks >> (inode->i_blkbits - 9)) <
1733                   ((new_size - 1) >> inode->i_blkbits) + 1);
1734
1735         while (bufsize > 0) {
1736                 int credits = handle->h_buffer_credits;
1737                 unsigned long last_block = (new_size == 0) ? 0 :
1738                                            (new_size - 1) >> inode->i_blkbits;
1739
1740                 if (bh)
1741                         brelse(bh);
1742
1743                 block = offset >> inode->i_blkbits;
1744                 boffs = offset & (blocksize - 1);
1745                 size = min(blocksize - boffs, bufsize);
1746                 sync = (block > last_block || new_size == 0 || sparse);
1747
1748                 if (sync)
1749                         down(&ei->i_append_sem);
1750
1751                 bh = __ldiskfs_bread(handle, inode, block, 0);
1752
1753                 if (unlikely(IS_ERR_OR_NULL(bh) && !sync))
1754                         CWARN(
1755                               "%s: adding bh without locking off %llu (block %lu, size %d, offs %llu)\n",
1756                               inode->i_sb->s_id,
1757                               offset, block, bufsize, *offs);
1758
1759                 if (IS_ERR_OR_NULL(bh)) {
1760                         struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1761                         int flags = LDISKFS_GET_BLOCKS_CREATE;
1762
1763                         /* while the file system is being mounted, avoid
1764                          * preallocation otherwise mount can take a long
1765                          * time as mballoc cache is cold.
1766                          * XXX: this is a workaround until we have a proper
1767                          *      fix in mballoc
1768                          * XXX: works with extent-based files only */
1769                         if (!osd->od_cl_seq)
1770                                 flags |= LDISKFS_GET_BLOCKS_NO_NORMALIZE;
1771                         bh = __ldiskfs_bread(handle, inode, block, flags);
1772                         create = true;
1773                 } else {
1774                         if (sync) {
1775                                 up(&ei->i_append_sem);
1776                                 sync = false;
1777                         }
1778                         create = false;
1779                 }
1780                 if (IS_ERR_OR_NULL(bh)) {
1781                         if (bh == NULL) {
1782                                 err = -EIO;
1783                         } else {
1784                                 err = PTR_ERR(bh);
1785                                 bh = NULL;
1786                         }
1787
1788                         CERROR(
1789                                "%s: error reading offset %llu (block %lu, size %d, offs %llu), credits %d/%d: rc = %d\n",
1790                                inode->i_sb->s_id, offset, block, bufsize, *offs,
1791                                credits, handle->h_buffer_credits, err);
1792                         break;
1793                 }
1794
1795                 err = ldiskfs_journal_get_write_access(handle, bh);
1796                 if (err) {
1797                         CERROR("journal_get_write_access() returned error %d\n",
1798                                err);
1799                         break;
1800                 }
1801                 LASSERTF(boffs + size <= bh->b_size,
1802                          "boffs %d size %d bh->b_size %lu\n",
1803                          boffs, size, (unsigned long)bh->b_size);
1804                 if (create) {
1805                         memset(bh->b_data, 0, bh->b_size);
1806                         if (sync) {
1807                                 up(&ei->i_append_sem);
1808                                 sync = false;
1809                         }
1810                 }
1811                 memcpy(bh->b_data + boffs, buf, size);
1812                 err = ldiskfs_handle_dirty_metadata(handle, NULL, bh);
1813                 if (err)
1814                         break;
1815
1816                 if (offset + size > new_size)
1817                         new_size = offset + size;
1818                 offset += size;
1819                 bufsize -= size;
1820                 buf += size;
1821         }
1822         if (sync)
1823                 up(&ei->i_append_sem);
1824
1825         if (bh)
1826                 brelse(bh);
1827
1828         if (write_NUL)
1829                 --new_size;
1830         /* correct in-core and on-disk sizes */
1831         if (new_size > i_size_read(inode)) {
1832                 spin_lock(&inode->i_lock);
1833                 if (new_size > i_size_read(inode))
1834                         i_size_write(inode, new_size);
1835                 if (i_size_read(inode) > ei->i_disksize) {
1836                         ei->i_disksize = i_size_read(inode);
1837                         dirty_inode = 1;
1838                 }
1839                 spin_unlock(&inode->i_lock);
1840         }
1841         if (dirty_inode)
1842                 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
1843
1844         if (err == 0)
1845                 *offs = offset;
1846         return err;
1847 }
1848
1849 static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt,
1850                          const struct lu_buf *buf, loff_t *pos,
1851                          struct thandle *handle)
1852 {
1853         struct inode            *inode = osd_dt_obj(dt)->oo_inode;
1854         struct osd_thandle      *oh;
1855         ssize_t                 result;
1856         int                     is_link;
1857
1858         LASSERT(dt_object_exists(dt));
1859
1860         LASSERT(handle != NULL);
1861         LASSERT(inode != NULL);
1862         dquot_initialize(inode);
1863
1864         /* XXX: don't check: one declared chunk can be used many times */
1865         /* osd_trans_exec_op(env, handle, OSD_OT_WRITE); */
1866
1867         oh = container_of(handle, struct osd_thandle, ot_super);
1868         LASSERT(oh->ot_handle->h_transaction != NULL);
1869         osd_trans_exec_op(env, handle, OSD_OT_WRITE);
1870
1871         /* Write small symlink to inode body as we need to maintain correct
1872          * on-disk symlinks for ldiskfs.
1873          * Note: the buf->lb_buf contains a NUL terminator while buf->lb_len
1874          * does not count it in.
1875          */
1876         is_link = S_ISLNK(dt->do_lu.lo_header->loh_attr);
1877         if (is_link && (buf->lb_len < sizeof(LDISKFS_I(inode)->i_data)))
1878                 result = osd_ldiskfs_writelink(inode, buf->lb_buf, buf->lb_len);
1879         else
1880                 result = osd_ldiskfs_write_record(dt, buf->lb_buf, buf->lb_len,
1881                                                   is_link, pos, oh->ot_handle);
1882         if (result == 0)
1883                 result = buf->lb_len;
1884
1885         osd_trans_exec_check(env, handle, OSD_OT_WRITE);
1886
1887         return result;
1888 }
1889
1890 static int osd_declare_fallocate(const struct lu_env *env,
1891                                  struct dt_object *dt, struct thandle *th)
1892 {
1893         struct osd_thandle *oh;
1894         struct inode *inode;
1895         int rc;
1896         ENTRY;
1897
1898         LASSERT(th);
1899         oh = container_of(th, struct osd_thandle, ot_super);
1900
1901         osd_trans_declare_op(env, oh, OSD_OT_PREALLOC,
1902                              osd_dto_credits_noquota[DTO_WRITE_BLOCK]);
1903         inode = osd_dt_obj(dt)->oo_inode;
1904         LASSERT(inode);
1905
1906         rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
1907                                    i_projid_read(inode), 0, oh, osd_dt_obj(dt),
1908                                    NULL, OSD_QID_BLK);
1909         RETURN(rc);
1910 }
1911
1912 static int osd_fallocate(const struct lu_env *env, struct dt_object *dt,
1913                          __u64 start, __u64 end, int mode, struct thandle *th)
1914 {
1915         struct osd_object *obj = osd_dt_obj(dt);
1916         struct inode *inode = obj->oo_inode;
1917         struct file *file;
1918         int rc = 0;
1919
1920         ENTRY;
1921         /*
1922          * Only mode == 0 (which is standard prealloc) is supported now.
1923          * Rest of mode options is not supported yet.
1924          */
1925         if (mode & ~FALLOC_FL_KEEP_SIZE)
1926                 RETURN(-EOPNOTSUPP);
1927
1928         LASSERT(dt_object_exists(dt));
1929         LASSERT(osd_invariant(obj));
1930         LASSERT(inode != NULL);
1931         dquot_initialize(inode);
1932
1933         LASSERT(th);
1934
1935         osd_trans_exec_op(env, th, OSD_OT_PREALLOC);
1936
1937         /*
1938          * Because f_op->fallocate() does not have an inode arg
1939          */
1940         file = osd_quasi_file(env, inode);
1941         rc = file->f_op->fallocate(file, mode, start, end - start);
1942
1943         RETURN(rc);
1944 }
1945
1946 static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt,
1947                              __u64 start, __u64 end, struct thandle *th)
1948 {
1949         struct osd_thandle *oh;
1950         struct inode       *inode;
1951         int                 rc;
1952         ENTRY;
1953
1954         LASSERT(th);
1955         oh = container_of(th, struct osd_thandle, ot_super);
1956
1957         /*
1958          * we don't need to reserve credits for whole truncate
1959          * it's not possible as truncate may need to free too many
1960          * blocks and that won't fit a single transaction. instead
1961          * we reserve credits to change i_size and put inode onto
1962          * orphan list. if needed truncate will extend or restart
1963          * transaction
1964          */
1965         osd_trans_declare_op(env, oh, OSD_OT_PUNCH,
1966                              osd_dto_credits_noquota[DTO_ATTR_SET_BASE] + 3);
1967
1968         inode = osd_dt_obj(dt)->oo_inode;
1969         LASSERT(inode);
1970
1971         rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
1972                                    i_projid_read(inode), 0, oh, osd_dt_obj(dt),
1973                                    NULL, OSD_QID_BLK);
1974
1975         if (rc == 0)
1976                 rc = osd_trunc_lock(osd_dt_obj(dt), oh, false);
1977
1978         RETURN(rc);
1979 }
1980
1981 static int osd_punch(const struct lu_env *env, struct dt_object *dt,
1982                      __u64 start, __u64 end, struct thandle *th)
1983 {
1984         struct osd_object *obj = osd_dt_obj(dt);
1985         struct osd_device *osd = osd_obj2dev(obj);
1986         struct inode *inode = obj->oo_inode;
1987         struct osd_access_lock *al;
1988         struct osd_thandle *oh;
1989         int rc = 0, found = 0;
1990         bool grow = false;
1991         ENTRY;
1992
1993         LASSERT(dt_object_exists(dt));
1994         LASSERT(osd_invariant(obj));
1995         LASSERT(inode != NULL);
1996         dquot_initialize(inode);
1997
1998         LASSERT(th);
1999         oh = container_of(th, struct osd_thandle, ot_super);
2000         LASSERT(oh->ot_handle->h_transaction != NULL);
2001
2002         /* we used to skip truncate to current size to
2003          * optimize truncates on OST. with DoM we can
2004          * get attr_set to set specific size (MDS_REINT)
2005          * and then get truncate RPC which essentially
2006          * would be skipped. this is bad.. so, disable
2007          * this optimization on MDS till the client stop
2008          * to sent MDS_REINT (LU-11033) -bzzz
2009          */
2010         if (osd->od_is_ost && i_size_read(inode) == start)
2011                 RETURN(0);
2012
2013         osd_trans_exec_op(env, th, OSD_OT_PUNCH);
2014
2015         spin_lock(&inode->i_lock);
2016         if (i_size_read(inode) < start)
2017                 grow = true;
2018         i_size_write(inode, start);
2019         spin_unlock(&inode->i_lock);
2020         /* if object holds encrypted content, we need to make sure we truncate
2021          * on an encryption unit boundary, or subsequent reads will get
2022          * corrupted content
2023          */
2024         if (obj->oo_lma_flags & LUSTRE_ENCRYPT_FL &&
2025             start & ~LUSTRE_ENCRYPTION_MASK)
2026                 start = (start & LUSTRE_ENCRYPTION_MASK) +
2027                         LUSTRE_ENCRYPTION_UNIT_SIZE;
2028         ll_truncate_pagecache(inode, start);
2029
2030         /* optimize grow case */
2031         if (grow) {
2032                 osd_execute_truncate(obj);
2033                 GOTO(out, rc);
2034         }
2035
2036         inode_lock(inode);
2037         /* add to orphan list to ensure truncate completion
2038          * if this transaction succeed. ldiskfs_truncate()
2039          * will take the inode out of the list
2040          */
2041         rc = ldiskfs_orphan_add(oh->ot_handle, inode);
2042         inode_unlock(inode);
2043         if (rc != 0)
2044                 GOTO(out, rc);
2045
2046         list_for_each_entry(al, &oh->ot_trunc_locks, tl_list) {
2047                 if (obj != al->tl_obj)
2048                         continue;
2049                 LASSERT(al->tl_shared == 0);
2050                 found = 1;
2051                 /* do actual truncate in osd_trans_stop() */
2052                 al->tl_truncate = 1;
2053                 break;
2054         }
2055         LASSERT(found);
2056
2057 out:
2058         RETURN(rc);
2059 }
2060
2061 static int fiemap_check_ranges(struct inode *inode,
2062                                u64 start, u64 len, u64 *new_len)
2063 {
2064         loff_t maxbytes;
2065
2066         *new_len = len;
2067
2068         if (len == 0)
2069                 return -EINVAL;
2070
2071         if (ldiskfs_test_inode_flag(inode, LDISKFS_INODE_EXTENTS))
2072                 maxbytes = inode->i_sb->s_maxbytes;
2073         else
2074                 maxbytes = LDISKFS_SB(inode->i_sb)->s_bitmap_maxbytes;
2075
2076         if (start > maxbytes)
2077                 return -EFBIG;
2078
2079         /*
2080          * Shrink request scope to what the fs can actually handle.
2081          */
2082         if (len > maxbytes || (maxbytes - len) < start)
2083                 *new_len = maxbytes - start;
2084
2085         return 0;
2086 }
2087
2088 /* So that the fiemap access checks can't overflow on 32 bit machines. */
2089 #define FIEMAP_MAX_EXTENTS     (UINT_MAX / sizeof(struct fiemap_extent))
2090
2091 static int osd_fiemap_get(const struct lu_env *env, struct dt_object *dt,
2092                           struct fiemap *fm)
2093 {
2094         struct fiemap_extent_info fieinfo = {0, };
2095         struct inode *inode = osd_dt_obj(dt)->oo_inode;
2096         u64 len;
2097         int rc;
2098         mm_segment_t cur_fs;
2099
2100         LASSERT(inode);
2101         if (inode->i_op->fiemap == NULL)
2102                 return -EOPNOTSUPP;
2103
2104         if (fm->fm_extent_count > FIEMAP_MAX_EXTENTS)
2105                 return -EINVAL;
2106
2107         rc = fiemap_check_ranges(inode, fm->fm_start, fm->fm_length, &len);
2108         if (rc)
2109                 return rc;
2110
2111         fieinfo.fi_flags = fm->fm_flags;
2112         fieinfo.fi_extents_max = fm->fm_extent_count;
2113         fieinfo.fi_extents_start = fm->fm_extents;
2114
2115         if (fieinfo.fi_flags & FIEMAP_FLAG_SYNC)
2116                 filemap_write_and_wait(inode->i_mapping);
2117
2118         /* Save previous value address limit */
2119         cur_fs = get_fs();
2120         /* Set the address limit of the kernel */
2121         set_fs(KERNEL_DS);
2122
2123         rc = inode->i_op->fiemap(inode, &fieinfo, fm->fm_start, len);
2124         fm->fm_flags = fieinfo.fi_flags;
2125         fm->fm_mapped_extents = fieinfo.fi_extents_mapped;
2126
2127         /* Restore the previous address limt */
2128         set_fs(cur_fs);
2129
2130         return rc;
2131 }
2132
2133 static int osd_ladvise(const struct lu_env *env, struct dt_object *dt,
2134                        __u64 start, __u64 end, enum lu_ladvise_type advice)
2135 {
2136         struct osd_object *obj = osd_dt_obj(dt);
2137         int rc = 0;
2138         ENTRY;
2139
2140         switch (advice) {
2141         case LU_LADVISE_DONTNEED:
2142                 if (end)
2143                         invalidate_mapping_pages(obj->oo_inode->i_mapping,
2144                                                  start >> PAGE_SHIFT,
2145                                                  (end - 1) >> PAGE_SHIFT);
2146                 break;
2147         default:
2148                 rc = -ENOTSUPP;
2149                 break;
2150         }
2151
2152         RETURN(rc);
2153 }
2154
2155 static loff_t osd_lseek(const struct lu_env *env, struct dt_object *dt,
2156                         loff_t offset, int whence)
2157 {
2158         struct osd_object *obj = osd_dt_obj(dt);
2159         struct inode *inode = obj->oo_inode;
2160         struct file *file;
2161         loff_t result;
2162
2163         ENTRY;
2164
2165         LASSERT(dt_object_exists(dt));
2166         LASSERT(osd_invariant(obj));
2167         LASSERT(inode);
2168
2169         file = osd_quasi_file(env, inode);
2170         result = file->f_op->llseek(file, offset, whence);
2171         /* when result is out of file range then it must be virtual hole
2172          * at the end of file, but this is not real file end, so return
2173          * just -ENXIO and LOV will merge all results
2174          */
2175         if (result == i_size_read(inode))
2176                 result = -ENXIO;
2177
2178         CDEBUG(D_INFO, "seek %s from %lld: %lld\n", whence == SEEK_HOLE ?
2179                        "hole" : "data", offset, result);
2180         RETURN(result);
2181 }
2182
2183 /*
2184  * in some cases we may need declare methods for objects being created
2185  * e.g., when we create symlink
2186  */
2187 const struct dt_body_operations osd_body_ops_new = {
2188         .dbo_declare_write = osd_declare_write,
2189 };
2190
2191 const struct dt_body_operations osd_body_ops = {
2192         .dbo_read                       = osd_read,
2193         .dbo_declare_write              = osd_declare_write,
2194         .dbo_write                      = osd_write,
2195         .dbo_bufs_get                   = osd_bufs_get,
2196         .dbo_bufs_put                   = osd_bufs_put,
2197         .dbo_write_prep                 = osd_write_prep,
2198         .dbo_declare_write_commit       = osd_declare_write_commit,
2199         .dbo_write_commit               = osd_write_commit,
2200         .dbo_read_prep                  = osd_read_prep,
2201         .dbo_declare_punch              = osd_declare_punch,
2202         .dbo_punch                      = osd_punch,
2203         .dbo_fiemap_get                 = osd_fiemap_get,
2204         .dbo_ladvise                    = osd_ladvise,
2205         .dbo_declare_fallocate          = osd_declare_fallocate,
2206         .dbo_fallocate                  = osd_fallocate,
2207         .dbo_lseek                      = osd_lseek,
2208 };
2209
2210 /**
2211  * Get a truncate lock
2212  *
2213  * In order to take multi-transaction truncate out of main transaction we let
2214  * the caller grab a lock on the object passed. the lock can be shared (for
2215  * writes) and exclusive (for truncate). It's not allowed to mix truncate
2216  * and write in the same transaction handle (do not confuse with big ldiskfs
2217  * transaction containing lots of handles).
2218  * The lock must be taken at declaration.
2219  *
2220  * \param obj           object to lock
2221  * \oh                  transaction
2222  * \shared              shared or exclusive
2223  *
2224  * \retval 0            lock is granted
2225  * \retval -NOMEM       no memory to allocate lock
2226  */
2227 int osd_trunc_lock(struct osd_object *obj, struct osd_thandle *oh, bool shared)
2228 {
2229         struct osd_access_lock *al, *tmp;
2230
2231         LASSERT(obj);
2232         LASSERT(oh);
2233
2234         list_for_each_entry(tmp, &oh->ot_trunc_locks, tl_list) {
2235                 if (tmp->tl_obj != obj)
2236                         continue;
2237                 LASSERT(tmp->tl_shared == shared);
2238                 /* found same lock */
2239                 return 0;
2240         }
2241
2242         OBD_ALLOC_PTR(al);
2243         if (unlikely(al == NULL))
2244                 return -ENOMEM;
2245         al->tl_obj = obj;
2246         al->tl_truncate = false;
2247         if (shared)
2248                 down_read(&obj->oo_ext_idx_sem);
2249         else
2250                 down_write(&obj->oo_ext_idx_sem);
2251         al->tl_shared = shared;
2252         lu_object_get(&obj->oo_dt.do_lu);
2253
2254         list_add(&al->tl_list, &oh->ot_trunc_locks);
2255
2256         return 0;
2257 }
2258
2259 void osd_trunc_unlock_all(const struct lu_env *env, struct list_head *list)
2260 {
2261         struct osd_access_lock *al, *tmp;
2262
2263         list_for_each_entry_safe(al, tmp, list, tl_list) {
2264                 if (al->tl_shared)
2265                         up_read(&al->tl_obj->oo_ext_idx_sem);
2266                 else
2267                         up_write(&al->tl_obj->oo_ext_idx_sem);
2268                 osd_object_put(env, al->tl_obj);
2269                 list_del(&al->tl_list);
2270                 OBD_FREE_PTR(al);
2271         }
2272 }
2273
2274 void osd_execute_truncate(struct osd_object *obj)
2275 {
2276         struct osd_device *d = osd_obj2dev(obj);
2277         struct inode *inode = obj->oo_inode;
2278         __u64 size;
2279
2280         /* simulate crash before (in the middle) of delayed truncate */
2281         if (OBD_FAIL_CHECK(OBD_FAIL_OSD_FAIL_AT_TRUNCATE)) {
2282                 struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
2283                 struct ldiskfs_sb_info *sbi = LDISKFS_SB(inode->i_sb);
2284
2285                 mutex_lock(&sbi->s_orphan_lock);
2286                 list_del_init(&ei->i_orphan);
2287                 mutex_unlock(&sbi->s_orphan_lock);
2288                 return;
2289         }
2290
2291         size = i_size_read(inode);
2292         inode_lock(inode);
2293         /* if object holds encrypted content, we need to make sure we truncate
2294          * on an encryption unit boundary, or block content will get corrupted
2295          */
2296         if (obj->oo_lma_flags & LUSTRE_ENCRYPT_FL &&
2297             size & ~LUSTRE_ENCRYPTION_MASK)
2298                 inode->i_size = (size & LUSTRE_ENCRYPTION_MASK) +
2299                         LUSTRE_ENCRYPTION_UNIT_SIZE;
2300         ldiskfs_truncate(inode);
2301         inode_unlock(inode);
2302         if (inode->i_size != size) {
2303                 spin_lock(&inode->i_lock);
2304                 i_size_write(inode, size);
2305                 LDISKFS_I(inode)->i_disksize = size;
2306                 spin_unlock(&inode->i_lock);
2307                 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
2308         }
2309
2310         /*
2311          * For a partial-page truncate, flush the page to disk immediately to
2312          * avoid data corruption during direct disk write.  b=17397
2313          */
2314         if ((size & ~PAGE_MASK) == 0)
2315                 return;
2316         if (osd_use_page_cache(d)) {
2317                 filemap_fdatawrite_range(inode->i_mapping, size, size + 1);
2318         } else {
2319                 /* Notice we use "wait" version to ensure I/O is complete */
2320                 filemap_write_and_wait_range(inode->i_mapping, size, size + 1);
2321                 invalidate_mapping_pages(inode->i_mapping, size >> PAGE_SHIFT,
2322                                          size >> PAGE_SHIFT);
2323         }
2324 }
2325
2326 void osd_process_truncates(struct list_head *list)
2327 {
2328         struct osd_access_lock *al;
2329
2330         LASSERT(journal_current_handle() == NULL);
2331
2332         list_for_each_entry(al, list, tl_list) {
2333                 if (al->tl_shared)
2334                         continue;
2335                 if (!al->tl_truncate)
2336                         continue;
2337                 osd_execute_truncate(al->tl_obj);
2338         }
2339 }