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  *
  31  * lustre/osd/osd_io.c
  32  *
  33  * body operations
  34  *
  35  * Author: Nikita Danilov <nikita@clusterfs.com>
  36  * Author: Alex Zhuravlev <bzzz@whamcloud.com>
  37  *
  38  */
  39
  40 #define DEBUG_SUBSYSTEM S_OSD
  41
  42 /* prerequisite for linux/xattr.h */
  43 #include <linux/types.h>
  44 /* prerequisite for linux/xattr.h */
  45 #include <linux/fs.h>
  46 #include <linux/mm.h>
  47 #include <linux/swap.h>
  48 #include <linux/pagevec.h>
  49
  50 /*
  51  * struct OBD_{ALLOC,FREE}*()
  52  * OBD_FAIL_CHECK
  53  */
  54 #include <obd_support.h>
  55
  56 #include "osd_internal.h"
  57
  58 /* ext_depth() */
  59 #include <ldiskfs/ldiskfs_extents.h>
  60 #include <ldiskfs/ldiskfs.h>
  61
  62 static inline bool osd_use_page_cache(struct osd_device *d)
  63 {
  64         /* do not use pagecache if write and read caching are disabled */
  65         if (d->od_writethrough_cache + d->od_read_cache == 0)
  66                 return false;
  67         /* use pagecache by default */
  68         return true;
  69 }
  70
  71 static int __osd_init_iobuf(struct osd_device *d, struct osd_iobuf *iobuf,
  72                             int rw, int line, int pages)
  73 {
  74         int blocks, i;
  75
  76         LASSERTF(iobuf->dr_elapsed_valid == 0,
  77                  "iobuf %p, reqs %d, rw %d, line %d\n", iobuf,
  78                  atomic_read(&iobuf->dr_numreqs), iobuf->dr_rw,
  79                  iobuf->dr_init_at);
  80         LASSERT(pages <= PTLRPC_MAX_BRW_PAGES);
  81
  82         init_waitqueue_head(&iobuf->dr_wait);
  83         atomic_set(&iobuf->dr_numreqs, 0);
  84         iobuf->dr_npages = 0;
  85         iobuf->dr_error = 0;
  86         iobuf->dr_dev = d;
  87         iobuf->dr_frags = 0;
  88         iobuf->dr_elapsed = ktime_set(0, 0);
  89         /* must be counted before, so assert */
  90         iobuf->dr_rw = rw;
  91         iobuf->dr_init_at = line;
  92
  93         blocks = pages * (PAGE_SIZE >> osd_sb(d)->s_blocksize_bits);
  94         if (iobuf->dr_bl_buf.lb_len >= blocks * sizeof(iobuf->dr_blocks[0])) {
  95                 LASSERT(iobuf->dr_pg_buf.lb_len >=
  96                         pages * sizeof(iobuf->dr_pages[0]));
  97                 return 0;
  98         }
  99
 100         /* start with 1MB for 4K blocks */
 101         i = 256;
 102         while (i <= PTLRPC_MAX_BRW_PAGES && i < pages)
 103                 i <<= 1;
 104
 105         CDEBUG(D_OTHER, "realloc %u for %u (%u) pages\n",
 106                (unsigned int)(pages * sizeof(iobuf->dr_pages[0])), i, pages);
 107         pages = i;
 108         blocks = pages * (PAGE_SIZE >> osd_sb(d)->s_blocksize_bits);
 109         iobuf->dr_max_pages = 0;
 110         CDEBUG(D_OTHER, "realloc %u for %u blocks\n",
 111                (unsigned int)(blocks * sizeof(iobuf->dr_blocks[0])), blocks);
 112
 113         lu_buf_realloc(&iobuf->dr_bl_buf, blocks * sizeof(iobuf->dr_blocks[0]));
 114         iobuf->dr_blocks = iobuf->dr_bl_buf.lb_buf;
 115         if (unlikely(iobuf->dr_blocks == NULL))
 116                 return -ENOMEM;
 117
 118         lu_buf_realloc(&iobuf->dr_pg_buf, pages * sizeof(iobuf->dr_pages[0]));
 119         iobuf->dr_pages = iobuf->dr_pg_buf.lb_buf;
 120         if (unlikely(iobuf->dr_pages == NULL))
 121                 return -ENOMEM;
 122
 123         lu_buf_realloc(&iobuf->dr_lnb_buf,
 124                        pages * sizeof(iobuf->dr_lnbs[0]));
 125         iobuf->dr_lnbs = iobuf->dr_lnb_buf.lb_buf;
 126         if (unlikely(iobuf->dr_lnbs == NULL))
 127                 return -ENOMEM;
 128
 129         iobuf->dr_max_pages = pages;
 130
 131         return 0;
 132 }
 133 #define osd_init_iobuf(dev, iobuf, rw, pages) \
 134         __osd_init_iobuf(dev, iobuf, rw, __LINE__, pages)
 135
 136 static void osd_iobuf_add_page(struct osd_iobuf *iobuf,
 137                                struct niobuf_local *lnb)
 138 {
 139         LASSERT(iobuf->dr_npages < iobuf->dr_max_pages);
 140         iobuf->dr_pages[iobuf->dr_npages] = lnb->lnb_page;
 141         iobuf->dr_lnbs[iobuf->dr_npages] = lnb;
 142         iobuf->dr_npages++;
 143 }
 144
 145 void osd_fini_iobuf(struct osd_device *d, struct osd_iobuf *iobuf)
 146 {
 147         int rw = iobuf->dr_rw;
 148
 149         if (iobuf->dr_elapsed_valid) {
 150                 struct brw_stats *h = &d->od_brw_stats;
 151
 152                 iobuf->dr_elapsed_valid = 0;
 153                 LASSERT(iobuf->dr_dev == d);
 154                 LASSERT(iobuf->dr_frags > 0);
 155                 lprocfs_oh_tally_pcpu(&h->bs_hist[BRW_R_DIO_FRAGS+rw],
 156                                       iobuf->dr_frags);
 157                 lprocfs_oh_tally_log2_pcpu(&h->bs_hist[BRW_R_IO_TIME+rw],
 158                                            ktime_to_ms(iobuf->dr_elapsed));
 159         }
 160 }
 161
 162 #ifdef HAVE_BIO_ENDIO_USES_ONE_ARG
 163 static void dio_complete_routine(struct bio *bio)
 164 {
 165         int error = blk_status_to_errno(bio->bi_status);
 166 #else
 167 static void dio_complete_routine(struct bio *bio, int error)
 168 {
 169 #endif
 170         struct osd_iobuf *iobuf = bio->bi_private;
 171         struct bio_vec *bvl;
 172
 173         /* CAVEAT EMPTOR: possibly in IRQ context
 174          * DO NOT record procfs stats here!!!
 175          */
 176
 177         if (unlikely(iobuf == NULL)) {
 178                 CERROR("***** bio->bi_private is NULL! Dump the bio contents to the console. Please report this to <https://jira.whamcloud.com/>, and probably have to reboot this node.\n");
 179                 CERROR("bi_next: %p, bi_flags: %lx, " __stringify(bi_opf)
 180                        ": %x, bi_vcnt: %d, bi_idx: %d, bi->size: %d, bi_end_io: %p, bi_cnt: %d, bi_private: %p\n",
 181                        bio->bi_next, (unsigned long)bio->bi_flags,
 182                        (unsigned int)bio->bi_opf, bio->bi_vcnt, bio_idx(bio),
 183                        bio_sectors(bio) << 9, bio->bi_end_io,
 184                        atomic_read(&bio->__bi_cnt),
 185                        bio->bi_private);
 186                 return;
 187         }
 188
 189         /* the check is outside of the cycle for performance reason -bzzz */
 190         if (!bio_data_dir(bio)) {
 191                 DECLARE_BVEC_ITER_ALL(iter_all);
 192
 193                 bio_for_each_segment_all(bvl, bio, iter_all) {
 194                         if (likely(error == 0))
 195                                 SetPageUptodate(bvl_to_page(bvl));
 196                         LASSERT(PageLocked(bvl_to_page(bvl)));
 197                 }
 198                 atomic_dec(&iobuf->dr_dev->od_r_in_flight);
 199         } else {
 200                 atomic_dec(&iobuf->dr_dev->od_w_in_flight);
 201         }
 202
 203         /* any real error is good enough -bzzz */
 204         if (error != 0 && iobuf->dr_error == 0)
 205                 iobuf->dr_error = error;
 206
 207         /*
 208          * set dr_elapsed before dr_numreqs turns to 0, otherwise
 209          * it's possible that service thread will see dr_numreqs
 210          * is zero, but dr_elapsed is not set yet, leading to lost
 211          * data in this processing and an assertion in a subsequent
 212          * call to OSD.
 213          */
 214         if (atomic_read(&iobuf->dr_numreqs) == 1) {
 215                 ktime_t now = ktime_get();
 216
 217                 iobuf->dr_elapsed = ktime_sub(now, iobuf->dr_start_time);
 218                 iobuf->dr_elapsed_valid = 1;
 219         }
 220         if (atomic_dec_and_test(&iobuf->dr_numreqs))
 221                 wake_up(&iobuf->dr_wait);
 222
 223         /* Completed bios used to be chained off iobuf->dr_bios and freed in
 224          * filter_clear_dreq().  It was then possible to exhaust the biovec-256
 225          * mempool when serious on-disk fragmentation was encountered,
 226          * deadlocking the OST.  The bios are now released as soon as complete
 227          * so the pool cannot be exhausted while IOs are competing. b=10076
 228          */
 229         bio_put(bio);
 230 }
 231
 232 static void record_start_io(struct osd_iobuf *iobuf, int size)
 233 {
 234         struct osd_device *osd = iobuf->dr_dev;
 235         struct brw_stats *h = &osd->od_brw_stats;
 236
 237         iobuf->dr_frags++;
 238         atomic_inc(&iobuf->dr_numreqs);
 239
 240         if (iobuf->dr_rw == 0) {
 241                 atomic_inc(&osd->od_r_in_flight);
 242                 lprocfs_oh_tally_pcpu(&h->bs_hist[BRW_R_RPC_HIST],
 243                                  atomic_read(&osd->od_r_in_flight));
 244                 lprocfs_oh_tally_log2_pcpu(&h->bs_hist[BRW_R_DISK_IOSIZE],
 245                                            size);
 246         } else if (iobuf->dr_rw == 1) {
 247                 atomic_inc(&osd->od_w_in_flight);
 248                 lprocfs_oh_tally_pcpu(&h->bs_hist[BRW_W_RPC_HIST],
 249                                  atomic_read(&osd->od_w_in_flight));
 250                 lprocfs_oh_tally_log2_pcpu(&h->bs_hist[BRW_W_DISK_IOSIZE],
 251                                            size);
 252         } else {
 253                 LBUG();
 254         }
 255 }
 256
 257 static void osd_submit_bio(int rw, struct bio *bio)
 258 {
 259         LASSERTF(rw == 0 || rw == 1, "%x\n", rw);
 260 #ifdef HAVE_SUBMIT_BIO_2ARGS
 261         submit_bio(rw ? WRITE : READ, bio);
 262 #else
 263         bio->bi_opf |= rw;
 264         submit_bio(bio);
 265 #endif
 266 }
 267
 268 static int can_be_merged(struct bio *bio, sector_t sector)
 269 {
 270         if (bio == NULL)
 271                 return 0;
 272
 273         return bio_end_sector(bio) == sector ? 1 : 0;
 274 }
 275
 276 #if IS_ENABLED(CONFIG_BLK_DEV_INTEGRITY)
 277 /*
 278  * This function will change the data written, thus it should only be
 279  * used when checking data integrity feature
 280  */
 281 static void bio_integrity_fault_inject(struct bio *bio)
 282 {
 283         struct bio_vec *bvec;
 284         DECLARE_BVEC_ITER_ALL(iter_all);
 285         void *kaddr;
 286         char *addr;
 287
 288         bio_for_each_segment_all(bvec, bio, iter_all) {
 289                 struct page *page = bvec->bv_page;
 290
 291                 kaddr = kmap(page);
 292                 addr = kaddr;
 293                 *addr = ~(*addr);
 294                 kunmap(page);
 295                 break;
 296         }
 297 }
 298
 299 static int bio_dif_compare(__u16 *expected_guard_buf, void *bio_prot_buf,
 300                            unsigned int sectors, int tuple_size)
 301 {
 302         __u16 *expected_guard;
 303         __u16 *bio_guard;
 304         int i;
 305
 306         expected_guard = expected_guard_buf;
 307         for (i = 0; i < sectors; i++) {
 308                 bio_guard = (__u16 *)bio_prot_buf;
 309                 if (*bio_guard != *expected_guard) {
 310                         CERROR(
 311                                "unexpected guard tags on sector %d expected guard %u, bio guard %u, sectors %u, tuple size %d\n",
 312                                i, *expected_guard, *bio_guard, sectors,
 313                                tuple_size);
 314                         return -EIO;
 315                 }
 316                 expected_guard++;
 317                 bio_prot_buf += tuple_size;
 318         }
 319         return 0;
 320 }
 321
 322 static int osd_bio_integrity_compare(struct bio *bio, struct block_device *bdev,
 323                                      struct osd_iobuf *iobuf, int index)
 324 {
 325         struct blk_integrity *bi = bdev_get_integrity(bdev);
 326         struct bio_integrity_payload *bip = bio->bi_integrity;
 327         struct niobuf_local *lnb = NULL;
 328         unsigned short sector_size = blk_integrity_interval(bi);
 329         void *bio_prot_buf = page_address(bip->bip_vec->bv_page) +
 330                 bip->bip_vec->bv_offset;
 331         struct bio_vec *bv;
 332         sector_t sector = bio_start_sector(bio);
 333         unsigned int i, sectors, total;
 334         DECLARE_BVEC_ITER_ALL(iter_all);
 335         __u16 *expected_guard;
 336         int rc;
 337
 338         total = 0;
 339         bio_for_each_segment_all(bv, bio, iter_all) {
 340                 for (i = index; i < iobuf->dr_npages; i++) {
 341                         if (iobuf->dr_pages[i] == bv->bv_page) {
 342                                 lnb = iobuf->dr_lnbs[i];
 343                                 break;
 344                         }
 345                 }
 346                 if (!lnb)
 347                         continue;
 348                 expected_guard = lnb->lnb_guards;
 349                 sectors = bv->bv_len / sector_size;
 350                 if (lnb->lnb_guard_rpc) {
 351                         rc = bio_dif_compare(expected_guard, bio_prot_buf,
 352                                              sectors, bi->tuple_size);
 353                         if (rc)
 354                                 return rc;
 355                 }
 356
 357                 sector += sectors;
 358                 bio_prot_buf += sectors * bi->tuple_size;
 359                 total += sectors * bi->tuple_size;
 360                 LASSERT(total <= bip_size(bio->bi_integrity));
 361                 index++;
 362                 lnb = NULL;
 363         }
 364         return 0;
 365 }
 366
 367 static int osd_bio_integrity_handle(struct osd_device *osd, struct bio *bio,
 368                                     struct osd_iobuf *iobuf,
 369                                     int start_page_idx, bool fault_inject,
 370                                     bool integrity_enabled)
 371 {
 372         struct super_block *sb = osd_sb(osd);
 373         integrity_gen_fn *generate_fn = NULL;
 374         integrity_vrfy_fn *verify_fn = NULL;
 375         int rc;
 376
 377         ENTRY;
 378
 379         if (!integrity_enabled)
 380                 RETURN(0);
 381
 382         rc = osd_get_integrity_profile(osd, &generate_fn, &verify_fn);
 383         if (rc)
 384                 RETURN(rc);
 385
 386         rc = bio_integrity_prep_fn(bio, generate_fn, verify_fn);
 387         if (rc)
 388                 RETURN(rc);
 389
 390         /* Verify and inject fault only when writing */
 391         if (iobuf->dr_rw == 1) {
 392                 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_OST_INTEGRITY_CMP))) {
 393                         rc = osd_bio_integrity_compare(bio, sb->s_bdev, iobuf,
 394                                                        start_page_idx);
 395                         if (rc)
 396                                 RETURN(rc);
 397                 }
 398
 399                 if (unlikely(fault_inject))
 400                         bio_integrity_fault_inject(bio);
 401         }
 402
 403         RETURN(0);
 404 }
 405
 406 #ifdef HAVE_BIO_INTEGRITY_PREP_FN
 407 #  ifdef HAVE_BIO_ENDIO_USES_ONE_ARG
 408 static void dio_integrity_complete_routine(struct bio *bio)
 409 #  else
 410 static void dio_integrity_complete_routine(struct bio *bio, int error)
 411 #  endif
 412 {
 413         struct osd_bio_private *bio_private = bio->bi_private;
 414
 415         bio->bi_private = bio_private->obp_iobuf;
 416         osd_dio_complete_routine(bio, error);
 417
 418         OBD_FREE_PTR(bio_private);
 419 }
 420 #endif /* HAVE_BIO_INTEGRITY_PREP_FN */
 421 #else  /* !CONFIG_BLK_DEV_INTEGRITY */
 422 #define osd_bio_integrity_handle(osd, bio, iobuf, start_page_idx, \
 423                                  fault_inject, integrity_enabled) 0
 424 #endif /* CONFIG_BLK_DEV_INTEGRITY */
 425
 426 static int osd_bio_init(struct bio *bio, struct osd_iobuf *iobuf,
 427                         bool integrity_enabled, int start_page_idx,
 428                         struct osd_bio_private **pprivate)
 429 {
 430         ENTRY;
 431
 432         *pprivate = NULL;
 433
 434 #ifdef HAVE_BIO_INTEGRITY_PREP_FN
 435         if (integrity_enabled) {
 436                 struct osd_bio_private *bio_private = NULL;
 437
 438                 OBD_ALLOC_GFP(bio_private, sizeof(*bio_private), GFP_NOIO);
 439                 if (bio_private == NULL)
 440                         RETURN(-ENOMEM);
 441                 bio->bi_end_io = dio_integrity_complete_routine;
 442                 bio->bi_private = bio_private;
 443                 bio_private->obp_start_page_idx = start_page_idx;
 444                 bio_private->obp_iobuf = iobuf;
 445                 *pprivate = bio_private;
 446         } else
 447 #endif
 448         {
 449                 bio->bi_end_io = dio_complete_routine;
 450                 bio->bi_private = iobuf;
 451         }
 452
 453         RETURN(0);
 454 }
 455
 456 static void osd_mark_page_io_done(struct osd_iobuf *iobuf,
 457                                   struct inode *inode,
 458                                   sector_t start_blocks,
 459                                   sector_t count)
 460 {
 461         struct niobuf_local *lnb;
 462         int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
 463         pgoff_t pg_start, pg_end;
 464
 465         pg_start = start_blocks / blocks_per_page;
 466         if (start_blocks % blocks_per_page)
 467                 pg_start++;
 468         if (count >= blocks_per_page)
 469                 pg_end = (start_blocks + count -
 470                           blocks_per_page) / blocks_per_page;
 471         else
 472                 return; /* nothing to mark */
 473         for ( ; pg_start <= pg_end; pg_start++) {
 474                 lnb = iobuf->dr_lnbs[pg_start];
 475                 lnb->lnb_flags |= OBD_BRW_DONE;
 476         }
 477 }
 478
 479 /*
 480  * Linux v5.12-rc1-20-ga8affc03a9b3
 481  *  block: rename BIO_MAX_PAGES to BIO_MAX_VECS
 482  */
 483 #ifndef BIO_MAX_VECS
 484 #define BIO_MAX_VECS    BIO_MAX_PAGES
 485 #endif
 486
 487 static int osd_do_bio(struct osd_device *osd, struct inode *inode,
 488                       struct osd_iobuf *iobuf, sector_t start_blocks,
 489                       sector_t count)
 490 {
 491         int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
 492         struct page **pages = iobuf->dr_pages;
 493         int npages = iobuf->dr_npages;
 494         sector_t *blocks = iobuf->dr_blocks;
 495         struct super_block *sb = inode->i_sb;
 496         int sector_bits = sb->s_blocksize_bits - 9;
 497         unsigned int blocksize = sb->s_blocksize;
 498         struct block_device *bdev = sb->s_bdev;
 499         struct osd_bio_private *bio_private = NULL;
 500         struct bio *bio = NULL;
 501         int bio_start_page_idx;
 502         struct page *page;
 503         unsigned int page_offset;
 504         sector_t sector;
 505         int nblocks;
 506         int block_idx, block_idx_end;
 507         int page_idx, page_idx_start;
 508         int i;
 509         int rc = 0;
 510         bool fault_inject;
 511         bool integrity_enabled;
 512         struct blk_plug plug;
 513         int blocks_left_page;
 514
 515         ENTRY;
 516
 517         fault_inject = OBD_FAIL_CHECK(OBD_FAIL_OST_INTEGRITY_FAULT);
 518         LASSERT(iobuf->dr_npages == npages);
 519
 520         integrity_enabled = bdev_integrity_enabled(bdev, iobuf->dr_rw);
 521
 522         osd_brw_stats_update(osd, iobuf);
 523         iobuf->dr_start_time = ktime_get();
 524
 525         if (!count)
 526                 count = npages * blocks_per_page;
 527         block_idx_end = start_blocks + count;
 528
 529         blk_start_plug(&plug);
 530
 531         page_idx_start = start_blocks / blocks_per_page;
 532         for (page_idx = page_idx_start, block_idx = start_blocks;
 533              block_idx < block_idx_end; page_idx++,
 534              block_idx += blocks_left_page) {
 535                 /* For cases where the filesystems blocksize is not the
 536                  * same as PAGE_SIZE (e.g. ARM with PAGE_SIZE=64KB and
 537                  * blocksize=4KB), there will be multiple blocks to
 538                  * read/write per page. Also, the start and end block may
 539                  * not be aligned to the start and end of the page, so the
 540                  * first page may skip some blocks at the start ("i != 0",
 541                  * "blocks_left_page" is reduced), and the last page may
 542                  * skip some blocks at the end (limited by "count").
 543                  */
 544                 page = pages[page_idx];
 545                 LASSERT(page_idx < iobuf->dr_npages);
 546
 547                 i = block_idx % blocks_per_page;
 548                 blocks_left_page = blocks_per_page - i;
 549                 if (block_idx + blocks_left_page > block_idx_end)
 550                         blocks_left_page = block_idx_end - block_idx;
 551                 page_offset = i * blocksize;
 552                 for (i = 0; i < blocks_left_page;
 553                      i += nblocks, page_offset += blocksize * nblocks) {
 554                         nblocks = 1;
 555
 556                         if (blocks[block_idx + i] == 0) {  /* hole */
 557                                 LASSERTF(iobuf->dr_rw == 0,
 558                                          "page_idx %u, block_idx %u, i %u,"
 559                                          "start_blocks: %llu, count: %llu, npages: %d\n",
 560                                          page_idx, block_idx, i,
 561                                          (unsigned long long)start_blocks,
 562                                          (unsigned long long)count, npages);
 563                                 memset(kmap(page) + page_offset, 0, blocksize);
 564                                 kunmap(page);
 565                                 continue;
 566                         }
 567
 568                         sector = (sector_t)blocks[block_idx + i] << sector_bits;
 569
 570                         /* Additional contiguous file blocks? */
 571                         while (i + nblocks < blocks_left_page &&
 572                                (sector + (nblocks << sector_bits)) ==
 573                                ((sector_t)blocks[block_idx + i + nblocks] <<
 574                                  sector_bits))
 575                                 nblocks++;
 576
 577                         if (bio && can_be_merged(bio, sector) &&
 578                             bio_add_page(bio, page, blocksize * nblocks,
 579                                          page_offset) != 0)
 580                                 continue;       /* added this frag OK */
 581
 582                         if (bio != NULL) {
 583                                 struct request_queue *q = bio_get_queue(bio);
 584                                 unsigned int bi_size = bio_sectors(bio) << 9;
 585
 586                                 /* Dang! I have to fragment this I/O */
 587                                 CDEBUG(D_INODE,
 588                                        "bio++ sz %d vcnt %d(%d) sectors %d(%d) psg %d(%d)\n",
 589                                        bi_size, bio->bi_vcnt, bio->bi_max_vecs,
 590                                        bio_sectors(bio),
 591                                        queue_max_sectors(q),
 592                                        osd_bio_nr_segs(bio),
 593                                        queue_max_segments(q));
 594                                 rc = osd_bio_integrity_handle(osd, bio,
 595                                         iobuf, bio_start_page_idx,
 596                                         fault_inject, integrity_enabled);
 597                                 if (rc) {
 598                                         bio_put(bio);
 599                                         goto out;
 600                                 }
 601
 602                                 record_start_io(iobuf, bi_size);
 603                                 osd_submit_bio(iobuf->dr_rw, bio);
 604                         }
 605
 606                         bio_start_page_idx = page_idx;
 607                         /* allocate new bio */
 608                         bio = bio_alloc(GFP_NOIO, min_t(unsigned short,
 609                                         BIO_MAX_VECS,
 610                                         (block_idx_end - block_idx +
 611                                          blocks_left_page - 1)));
 612                         if (bio == NULL) {
 613                                 CERROR("Can't allocate bio %u pages\n",
 614                                        block_idx_end - block_idx +
 615                                        blocks_left_page - 1);
 616                                 rc = -ENOMEM;
 617                                 goto out;
 618                         }
 619
 620                         bio_set_dev(bio, bdev);
 621                         bio_set_sector(bio, sector);
 622                         bio->bi_opf = iobuf->dr_rw ? WRITE : READ;
 623                         rc = osd_bio_init(bio, iobuf, integrity_enabled,
 624                                           bio_start_page_idx, &bio_private);
 625                         if (rc) {
 626                                 bio_put(bio);
 627                                 goto out;
 628                         }
 629
 630                         rc = bio_add_page(bio, page,
 631                                           blocksize * nblocks, page_offset);
 632                         LASSERT(rc != 0);
 633                 }
 634         }
 635
 636         if (bio != NULL) {
 637                 rc = osd_bio_integrity_handle(osd, bio, iobuf,
 638                                               bio_start_page_idx,
 639                                               fault_inject,
 640                                               integrity_enabled);
 641                 if (rc) {
 642                         bio_put(bio);
 643                         goto out;
 644                 }
 645
 646                 record_start_io(iobuf, bio_sectors(bio) << 9);
 647                 osd_submit_bio(iobuf->dr_rw, bio);
 648                 rc = 0;
 649         }
 650
 651 out:
 652         blk_finish_plug(&plug);
 653
 654         /* in order to achieve better IO throughput, we don't wait for writes
 655          * completion here. instead we proceed with transaction commit in
 656          * parallel and wait for IO completion once transaction is stopped
 657          * see osd_trans_stop() for more details -bzzz
 658          */
 659         if (iobuf->dr_rw == 0 || fault_inject) {
 660                 wait_event(iobuf->dr_wait,
 661                            atomic_read(&iobuf->dr_numreqs) == 0);
 662                 osd_fini_iobuf(osd, iobuf);
 663         }
 664
 665         if (rc == 0) {
 666                 rc = iobuf->dr_error;
 667         } else {
 668                 if (bio_private)
 669                         OBD_FREE_PTR(bio_private);
 670         }
 671
 672         /* Write only now */
 673         if (rc == 0 && iobuf->dr_rw)
 674                 osd_mark_page_io_done(iobuf, inode,
 675                                       start_blocks, count);
 676
 677         RETURN(rc);
 678 }
 679
 680 static int osd_map_remote_to_local(loff_t offset, ssize_t len, int *nrpages,
 681                                    struct niobuf_local *lnb, int maxlnb)
 682 {
 683         int rc = 0;
 684         ENTRY;
 685
 686         *nrpages = 0;
 687
 688         while (len > 0) {
 689                 int poff = offset & (PAGE_SIZE - 1);
 690                 int plen = PAGE_SIZE - poff;
 691
 692                 if (*nrpages >= maxlnb) {
 693                         rc = -EOVERFLOW;
 694                         break;
 695                 }
 696
 697                 if (plen > len)
 698                         plen = len;
 699                 lnb->lnb_file_offset = offset;
 700                 lnb->lnb_page_offset = poff;
 701                 lnb->lnb_len = plen;
 702                 /* lnb->lnb_flags = rnb->rnb_flags; */
 703                 lnb->lnb_flags = 0;
 704                 lnb->lnb_page = NULL;
 705                 lnb->lnb_rc = 0;
 706                 lnb->lnb_guard_rpc = 0;
 707                 lnb->lnb_guard_disk = 0;
 708                 lnb->lnb_locked = 0;
 709
 710                 LASSERTF(plen <= len, "plen %u, len %lld\n", plen,
 711                          (long long) len);
 712                 offset += plen;
 713                 len -= plen;
 714                 lnb++;
 715                 (*nrpages)++;
 716         }
 717
 718         RETURN(rc);
 719 }
 720
 721 static struct page *osd_get_page(const struct lu_env *env, struct dt_object *dt,
 722                                  loff_t offset, gfp_t gfp_mask, bool cache)
 723 {
 724         struct osd_thread_info *oti = osd_oti_get(env);
 725         struct inode *inode = osd_dt_obj(dt)->oo_inode;
 726         struct osd_device *d = osd_obj2dev(osd_dt_obj(dt));
 727         struct page *page;
 728         int cur;
 729
 730         LASSERT(inode);
 731
 732         if (cache) {
 733                 page = find_or_create_page(inode->i_mapping,
 734                                            offset >> PAGE_SHIFT, gfp_mask);
 735
 736                 if (likely(page)) {
 737                         LASSERT(!PagePrivate2(page));
 738                         wait_on_page_writeback(page);
 739                 } else {
 740                         lprocfs_counter_add(d->od_stats, LPROC_OSD_NO_PAGE, 1);
 741                 }
 742
 743                 return page;
 744         }
 745
 746         if (inode->i_mapping->nrpages) {
 747                 /* consult with pagecache, but do not create new pages */
 748                 /* this is normally used once */
 749                 page = find_lock_page(inode->i_mapping, offset >> PAGE_SHIFT);
 750                 if (page) {
 751                         wait_on_page_writeback(page);
 752                         return page;
 753                 }
 754         }
 755
 756         LASSERT(oti->oti_dio_pages);
 757         cur = oti->oti_dio_pages_used;
 758         page = oti->oti_dio_pages[cur];
 759
 760         if (unlikely(!page)) {
 761                 LASSERT(cur < PTLRPC_MAX_BRW_PAGES);
 762                 page = alloc_page(gfp_mask);
 763                 if (!page)
 764                         return NULL;
 765                 oti->oti_dio_pages[cur] = page;
 766                 SetPagePrivate2(page);
 767                 lock_page(page);
 768         }
 769
 770         ClearPageUptodate(page);
 771         page->index = offset >> PAGE_SHIFT;
 772         oti->oti_dio_pages_used++;
 773
 774         return page;
 775 }
 776
 777 /*
 778  * there are following "locks":
 779  * journal_start
 780  * i_mutex
 781  * page lock
 782  *
 783  * osd write path:
 784  *  - lock page(s)
 785  *  - journal_start
 786  *  - truncate_sem
 787  *
 788  * ext4 vmtruncate:
 789  *  - lock pages, unlock
 790  *  - journal_start
 791  *  - lock partial page
 792  *  - i_data_sem
 793  *
 794  */
 795
 796 /**
 797  * Unlock and release pages loaded by osd_bufs_get()
 798  *
 799  * Unlock \a npages pages from \a lnb and drop the refcount on them.
 800  *
 801  * \param env           thread execution environment
 802  * \param dt            dt object undergoing IO (OSD object + methods)
 803  * \param lnb           array of pages undergoing IO
 804  * \param npages        number of pages in \a lnb
 805  *
 806  * \retval 0            always
 807  */
 808 static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt,
 809                         struct niobuf_local *lnb, int npages)
 810 {
 811         struct osd_thread_info *oti = osd_oti_get(env);
 812         struct pagevec pvec;
 813         int i;
 814
 815         ll_pagevec_init(&pvec, 0);
 816
 817         for (i = 0; i < npages; i++) {
 818                 struct page *page = lnb[i].lnb_page;
 819
 820                 if (page == NULL)
 821                         continue;
 822
 823                 /* if the page isn't cached, then reset uptodate
 824                  * to prevent reuse
 825                  */
 826                 if (PagePrivate2(page)) {
 827                         oti->oti_dio_pages_used--;
 828                 } else {
 829                         if (lnb[i].lnb_locked)
 830                                 unlock_page(page);
 831                         if (pagevec_add(&pvec, page) == 0)
 832                                 pagevec_release(&pvec);
 833                 }
 834
 835                 lnb[i].lnb_page = NULL;
 836         }
 837
 838         LASSERTF(oti->oti_dio_pages_used == 0, "%d\n", oti->oti_dio_pages_used);
 839
 840         /* Release any partial pagevec */
 841         pagevec_release(&pvec);
 842
 843         RETURN(0);
 844 }
 845
 846 /**
 847  * Load and lock pages undergoing IO
 848  *
 849  * Pages as described in the \a lnb array are fetched (from disk or cache)
 850  * and locked for IO by the caller.
 851  *
 852  * DLM locking protects us from write and truncate competing for same region,
 853  * but partial-page truncate can leave dirty pages in the cache for ldiskfs.
 854  * It's possible the writeout on a such a page is in progress when we access
 855  * it. It's also possible that during this writeout we put new (partial) data
 856  * into the page, but won't be able to proceed in filter_commitrw_write().
 857  * Therefore, just wait for writeout completion as it should be rare enough.
 858  *
 859  * \param env           thread execution environment
 860  * \param dt            dt object undergoing IO (OSD object + methods)
 861  * \param pos           byte offset of IO start
 862  * \param len           number of bytes of IO
 863  * \param lnb           array of extents undergoing IO
 864  * \param rw            read or write operation, and other flags
 865  * \param capa          capabilities
 866  *
 867  * \retval pages        (zero or more) loaded successfully
 868  * \retval -ENOMEM      on memory/page allocation error
 869  */
 870 static int osd_bufs_get(const struct lu_env *env, struct dt_object *dt,
 871                         loff_t pos, ssize_t len, struct niobuf_local *lnb,
 872                         int maxlnb, enum dt_bufs_type rw)
 873 {
 874         struct osd_thread_info *oti = osd_oti_get(env);
 875         struct osd_object *obj = osd_dt_obj(dt);
 876         struct osd_device *osd   = osd_obj2dev(obj);
 877         int npages, i, iosize, rc = 0;
 878         bool cache, write;
 879         loff_t fsize;
 880         gfp_t gfp_mask;
 881
 882         LASSERT(obj->oo_inode);
 883
 884         if (unlikely(obj->oo_destroyed))
 885                 RETURN(-ENOENT);
 886
 887         rc = osd_map_remote_to_local(pos, len, &npages, lnb, maxlnb);
 888         if (rc)
 889                 RETURN(rc);
 890
 891         write = rw & DT_BUFS_TYPE_WRITE;
 892
 893         fsize = lnb[npages - 1].lnb_file_offset + lnb[npages - 1].lnb_len;
 894         iosize = fsize - lnb[0].lnb_file_offset;
 895         fsize = max(fsize, i_size_read(obj->oo_inode));
 896
 897         cache = rw & DT_BUFS_TYPE_READAHEAD;
 898         if (cache)
 899                 goto bypass_checks;
 900
 901         cache = osd_use_page_cache(osd);
 902         while (cache) {
 903                 if (write) {
 904                         if (!osd->od_writethrough_cache) {
 905                                 cache = false;
 906                                 break;
 907                         }
 908                         if (iosize > osd->od_writethrough_max_iosize) {
 909                                 cache = false;
 910                                 break;
 911                         }
 912                 } else {
 913                         if (!osd->od_read_cache) {
 914                                 cache = false;
 915                                 break;
 916                         }
 917                         if (iosize > osd->od_readcache_max_iosize) {
 918                                 cache = false;
 919                                 break;
 920                         }
 921                 }
 922                 /* don't use cache on large files */
 923                 if (osd->od_readcache_max_filesize &&
 924                     fsize > osd->od_readcache_max_filesize)
 925                         cache = false;
 926                 break;
 927         }
 928
 929 bypass_checks:
 930         if (!cache && unlikely(!oti->oti_dio_pages)) {
 931                 OBD_ALLOC_PTR_ARRAY_LARGE(oti->oti_dio_pages,
 932                                           PTLRPC_MAX_BRW_PAGES);
 933                 if (!oti->oti_dio_pages)
 934                         return -ENOMEM;
 935         }
 936
 937         /* this could also try less hard for DT_BUFS_TYPE_READAHEAD pages */
 938         gfp_mask = rw & DT_BUFS_TYPE_LOCAL ? (GFP_NOFS | __GFP_HIGHMEM) :
 939                                              GFP_HIGHUSER;
 940         for (i = 0; i < npages; i++, lnb++) {
 941                 lnb->lnb_page = osd_get_page(env, dt, lnb->lnb_file_offset,
 942                                              gfp_mask, cache);
 943                 if (lnb->lnb_page == NULL)
 944                         GOTO(cleanup, rc = -ENOMEM);
 945
 946                 lnb->lnb_locked = 1;
 947                 if (cache)
 948                         mark_page_accessed(lnb->lnb_page);
 949         }
 950
 951 #if 0
 952         /* XXX: this version doesn't invalidate cached pages, but use them */
 953         if (!cache && write && obj->oo_inode->i_mapping->nrpages) {
 954                 /* do not allow data aliasing, invalidate pagecache */
 955                 /* XXX: can be quite expensive in mixed case */
 956                 invalidate_mapping_pages(obj->oo_inode->i_mapping,
 957                                 lnb[0].lnb_file_offset >> PAGE_SHIFT,
 958                                 lnb[npages - 1].lnb_file_offset >> PAGE_SHIFT);
 959         }
 960 #endif
 961
 962         RETURN(i);
 963
 964 cleanup:
 965         if (i > 0)
 966                 osd_bufs_put(env, dt, lnb - i, i);
 967         return rc;
 968 }
 969 /* Borrow @ext4_chunk_trans_blocks */
 970 static int osd_chunk_trans_blocks(struct inode *inode, int nrblocks)
 971 {
 972         ldiskfs_group_t groups;
 973         int gdpblocks;
 974         int idxblocks;
 975         int depth;
 976         int ret;
 977
 978         depth = ext_depth(inode);
 979         idxblocks = depth * 2;
 980
 981         /*
 982          * Now let's see how many group bitmaps and group descriptors need
 983          * to account.
 984          */
 985         groups = idxblocks + 1;
 986         gdpblocks = groups;
 987         if (groups > LDISKFS_SB(inode->i_sb)->s_groups_count)
 988                 groups = LDISKFS_SB(inode->i_sb)->s_groups_count;
 989         if (gdpblocks > LDISKFS_SB(inode->i_sb)->s_gdb_count)
 990                 gdpblocks = LDISKFS_SB(inode->i_sb)->s_gdb_count;
 991
 992         /* bitmaps and block group descriptor blocks */
 993         ret = idxblocks + groups + gdpblocks;
 994
 995         /* Blocks for super block, inode, quota and xattr blocks */
 996         ret += LDISKFS_META_TRANS_BLOCKS(inode->i_sb);
 997
 998         return ret;
 999 }
1000
1001 #ifdef HAVE_LDISKFS_JOURNAL_ENSURE_CREDITS
1002 static int osd_extend_restart_trans(handle_t *handle, int needed,
1003                                     struct inode *inode)
1004 {
1005         int rc;
1006
1007         rc = ldiskfs_journal_ensure_credits(handle, needed,
1008                 ldiskfs_trans_default_revoke_credits(inode->i_sb));
1009         /* this means journal has been restarted */
1010         if (rc > 0)
1011                 rc = 0;
1012
1013         return rc;
1014 }
1015 #else
1016 static int osd_extend_restart_trans(handle_t *handle, int needed,
1017                                     struct inode *inode)
1018 {
1019         int rc;
1020
1021         if (ldiskfs_handle_has_enough_credits(handle, needed))
1022                 return 0;
1023         rc = ldiskfs_journal_extend(handle,
1024                                 needed - handle->h_buffer_credits);
1025         if (rc <= 0)
1026                 return rc;
1027
1028         return ldiskfs_journal_restart(handle, needed);
1029 }
1030 #endif /* HAVE_LDISKFS_JOURNAL_ENSURE_CREDITS */
1031
1032 static int osd_ldiskfs_map_write(struct inode *inode, struct osd_iobuf *iobuf,
1033                                  struct osd_device *osd, sector_t start_blocks,
1034                                  sector_t count, loff_t *disk_size,
1035                                  __u64 user_size)
1036 {
1037         /* if file has grown, take user_size into account */
1038         if (user_size && *disk_size > user_size)
1039                 *disk_size = user_size;
1040
1041         spin_lock(&inode->i_lock);
1042         if (*disk_size > i_size_read(inode)) {
1043                 i_size_write(inode, *disk_size);
1044                 LDISKFS_I(inode)->i_disksize = *disk_size;
1045                 spin_unlock(&inode->i_lock);
1046                 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
1047         } else {
1048                 spin_unlock(&inode->i_lock);
1049         }
1050
1051         /*
1052          * We don't do stats here as in read path because
1053          * write is async: we'll do this in osd_put_bufs()
1054          */
1055         return osd_do_bio(osd, inode, iobuf, start_blocks, count);
1056 }
1057
1058 static unsigned int osd_extent_bytes(const struct osd_device *o)
1059 {
1060         unsigned int *extent_bytes_ptr =
1061                         raw_cpu_ptr(o->od_extent_bytes_percpu);
1062
1063         if (likely(*extent_bytes_ptr))
1064                 return *extent_bytes_ptr;
1065
1066         /* initialize on first access or CPU hotplug */
1067         if (!ldiskfs_has_feature_extents(osd_sb(o)))
1068                 *extent_bytes_ptr = 1 << osd_sb(o)->s_blocksize_bits;
1069         else
1070                 *extent_bytes_ptr = OSD_DEFAULT_EXTENT_BYTES;
1071
1072         return *extent_bytes_ptr;
1073 }
1074
1075 #define EXTENT_BYTES_DECAY 64
1076 static void osd_decay_extent_bytes(struct osd_device *osd,
1077                                    unsigned int new_bytes)
1078 {
1079         unsigned int old_bytes;
1080
1081         if (!ldiskfs_has_feature_extents(osd_sb(osd)))
1082                 return;
1083
1084         old_bytes = osd_extent_bytes(osd);
1085         *raw_cpu_ptr(osd->od_extent_bytes_percpu) =
1086                 (old_bytes * (EXTENT_BYTES_DECAY - 1) +
1087                  min(new_bytes, OSD_DEFAULT_EXTENT_BYTES) +
1088                  EXTENT_BYTES_DECAY - 1) / EXTENT_BYTES_DECAY;
1089 }
1090
1091 static int osd_ldiskfs_map_inode_pages(struct inode *inode,
1092                                        struct osd_iobuf *iobuf,
1093                                        struct osd_device *osd,
1094                                        int create, __u64 user_size,
1095                                        int check_credits,
1096                                        struct thandle *thandle)
1097 {
1098         int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
1099         int blocksize = 1 << inode->i_blkbits;
1100         int rc = 0, i = 0, mapped_index = 0;
1101         struct page *fp = NULL;
1102         int clen = 0;
1103         pgoff_t max_page_index;
1104         handle_t *handle = NULL;
1105         sector_t start_blocks = 0, count = 0;
1106         loff_t disk_size = 0;
1107         struct page **page = iobuf->dr_pages;
1108         int pages = iobuf->dr_npages;
1109         sector_t *blocks = iobuf->dr_blocks;
1110         struct niobuf_local *lnb1, *lnb2;
1111         loff_t size1, size2;
1112
1113         max_page_index = inode->i_sb->s_maxbytes >> PAGE_SHIFT;
1114
1115         CDEBUG(D_OTHER, "inode %lu: map %d pages from %lu\n",
1116                 inode->i_ino, pages, (*page)->index);
1117
1118         if (create) {
1119                 create = LDISKFS_GET_BLOCKS_CREATE;
1120                 handle = ldiskfs_journal_current_handle();
1121                 LASSERT(handle != NULL);
1122                 rc = osd_attach_jinode(inode);
1123                 if (rc)
1124                         return rc;
1125                 disk_size = i_size_read(inode);
1126                 /* if disk_size is already bigger than specified user_size,
1127                  * ignore user_size
1128                  */
1129                 if (disk_size > user_size)
1130                         user_size = 0;
1131         }
1132         /* pages are sorted already. so, we just have to find
1133          * contig. space and process them properly
1134          */
1135         while (i < pages) {
1136                 long blen, total = 0, previous_total = 0;
1137                 struct ldiskfs_map_blocks map = { 0 };
1138                 ktime_t time;
1139
1140                 if (fp == NULL) { /* start new extent */
1141                         fp = *page++;
1142                         clen = 1;
1143                         if (++i != pages)
1144                                 continue;
1145                 } else if (fp->index + clen == (*page)->index) {
1146                         /* continue the extent */
1147                         page++;
1148                         clen++;
1149                         if (++i != pages)
1150                                 continue;
1151                 }
1152                 if (fp->index + clen >= max_page_index)
1153                         GOTO(cleanup, rc = -EFBIG);
1154                 /* process found extent */
1155                 map.m_lblk = fp->index * blocks_per_page;
1156                 map.m_len = blen = clen * blocks_per_page;
1157
1158                 /*
1159                  * For PAGE_SIZE > blocksize block allocation mapping, the
1160                  * ldiskfs_map_blocks() aims at looking up already mapped
1161                  * blocks, recording them to iobuf->dr_blocks and fixing up
1162                  * m_lblk, m_len for un-allocated blocks to be created/mapped
1163                  * in the second ldiskfs_map_blocks().
1164                  *
1165                  * M_lblk should be the first un-allocated block if m_lblk
1166                  * points at an already allocated block when create = 1,
1167                  * ldiskfs_map_blocks() will just return with already
1168                  * allocated blocks and without allocating any requested
1169                  * new blocks for the extent. For PAGE_SIZE = blocksize
1170                  * case, if m_lblk points at an already allocated block it
1171                  * will point at an un-allocated block in next restart
1172                  * transaction, because the already mapped block/page will
1173                  * be filtered out in next restart transaction via flag
1174                  * OBD_BRW_DONE in osd_declare_write_commit().
1175                  */
1176                 if (create && PAGE_SIZE > blocksize) {
1177                         /* With flags=0 just for already mapped blocks lookup */
1178                         rc = ldiskfs_map_blocks(handle, inode, &map, 0);
1179                         if (rc > 0 && map.m_flags & LDISKFS_MAP_MAPPED) {
1180                                 for (; total < blen && total < map.m_len;
1181                                                 total++)
1182                                         *(blocks + total) = map.m_pblk + total;
1183
1184                                 /* The extent is already full mapped */
1185                                 if (total == blen) {
1186                                         rc = 0;
1187                                         goto ext_already_mapped;
1188                                 }
1189                         }
1190                         /*
1191                          * Fixup or reset m_lblk and m_len for un-mapped blocks.
1192                          * The second ldiskfs_map_blocks() will create and map
1193                          * them.
1194                          */
1195                         map.m_lblk = fp->index * blocks_per_page + total;
1196                         map.m_len = blen - total;
1197                 }
1198
1199 cont_map:
1200                 /**
1201                  * We might restart transaction for block allocations,
1202                  * in order to make sure data ordered mode, issue IO, disk
1203                  * size update and block allocations need be within same
1204                  * transaction to make sure consistency.
1205                  */
1206                 if (handle && check_credits) {
1207                         struct osd_thandle *oh;
1208
1209                         LASSERT(thandle != NULL);
1210                         oh = container_of(thandle, struct osd_thandle,
1211                                           ot_super);
1212                         /*
1213                          * only issue IO if restart transaction needed,
1214                          * as update disk size need hold inode lock, we
1215                          * want to avoid that as much as possible.
1216                          */
1217                         if (oh->oh_declared_ext <= 0) {
1218                                 rc = osd_ldiskfs_map_write(inode,
1219                                         iobuf, osd, start_blocks,
1220                                         count, &disk_size, user_size);
1221                                 if (rc)
1222                                         GOTO(cleanup, rc);
1223                                 thandle->th_restart_tran = 1;
1224                                 GOTO(cleanup, rc = -EAGAIN);
1225                         }
1226
1227                         if (OBD_FAIL_CHECK(OBD_FAIL_OST_RESTART_IO))
1228                                 oh->oh_declared_ext = 0;
1229                         else
1230                                 oh->oh_declared_ext--;
1231                 }
1232
1233                 time = ktime_get();
1234                 rc = ldiskfs_map_blocks(handle, inode, &map, create);
1235                 time = ktime_sub(ktime_get(), time);
1236
1237                 if (rc >= 0) {
1238                         struct brw_stats *h = &osd->od_brw_stats;
1239                         int idx, c = 0;
1240
1241                         idx = map.m_flags & LDISKFS_MAP_NEW ?
1242                                 BRW_ALLOC_TIME : BRW_MAP_TIME;
1243                         lprocfs_oh_tally_log2_pcpu(&h->bs_hist[idx],
1244                                                    ktime_to_ms(time));
1245
1246                         for (; total < blen && c < map.m_len; c++, total++) {
1247                                 if (rc == 0) {
1248                                         *(blocks + total) = 0;
1249                                         total++;
1250                                         break;
1251                                 }
1252                                 if ((map.m_flags & LDISKFS_MAP_UNWRITTEN) &&
1253                                     !create) {
1254                                         /* don't try to read allocated, but
1255                                          * unwritten blocks, instead fill the
1256                                          * patches with zeros in osd_do_bio() */
1257                                         *(blocks + total) = 0;
1258                                         continue;
1259                                 }
1260                                 *(blocks + total) = map.m_pblk + c;
1261                                 /* unmap any possible underlying
1262                                  * metadata from the block device
1263                                  * mapping.  b=6998.
1264                                  */
1265                                 if ((map.m_flags & LDISKFS_MAP_NEW) &&
1266                                     create)
1267                                         clean_bdev_aliases(inode->i_sb->s_bdev,
1268                                                            map.m_pblk + c, 1);
1269                         }
1270                         rc = 0;
1271                 }
1272
1273 ext_already_mapped:
1274                 if (rc == 0 && create) {
1275                         count += (total - previous_total);
1276                         mapped_index = (count + blocks_per_page -
1277                                         1) / blocks_per_page - 1;
1278                         lnb1 = iobuf->dr_lnbs[i - clen];
1279                         lnb2 = iobuf->dr_lnbs[mapped_index];
1280                         size1 = lnb1->lnb_file_offset -
1281                                 (lnb1->lnb_file_offset % PAGE_SIZE) +
1282                                 (total << inode->i_blkbits);
1283                         size2 = lnb2->lnb_file_offset + lnb2->lnb_len;
1284
1285                         if (size1 > size2)
1286                                 size1 = size2;
1287                         if (size1 > disk_size)
1288                                 disk_size = size1;
1289                 }
1290
1291                 if (rc == 0 && total < blen) {
1292                         /*
1293                          * decay extent blocks if we could not
1294                          * allocate extent once.
1295                          */
1296                         osd_decay_extent_bytes(osd,
1297                                 (total - previous_total) << inode->i_blkbits);
1298                         map.m_lblk = fp->index * blocks_per_page + total;
1299                         map.m_len = blen - total;
1300                         previous_total = total;
1301                         goto cont_map;
1302                 }
1303                 if (rc != 0)
1304                         GOTO(cleanup, rc);
1305                 /*
1306                  * decay extent blocks if we could allocate
1307                  * good large extent.
1308                  */
1309                 if (total - previous_total >=
1310                     osd_extent_bytes(osd) >> inode->i_blkbits)
1311                         osd_decay_extent_bytes(osd,
1312                                 (total - previous_total) << inode->i_blkbits);
1313                 /* look for next extent */
1314                 fp = NULL;
1315                 blocks += blocks_per_page * clen;
1316         }
1317 cleanup:
1318         if (rc == 0 && create &&
1319             start_blocks < pages * blocks_per_page) {
1320                 rc = osd_ldiskfs_map_write(inode, iobuf, osd, start_blocks,
1321                                            count, &disk_size, user_size);
1322                 LASSERT(start_blocks + count == pages * blocks_per_page);
1323         }
1324         return rc;
1325 }
1326
1327 static int osd_write_prep(const struct lu_env *env, struct dt_object *dt,
1328                           struct niobuf_local *lnb, int npages)
1329 {
1330         struct osd_thread_info *oti   = osd_oti_get(env);
1331         struct osd_iobuf       *iobuf = &oti->oti_iobuf;
1332         struct inode           *inode = osd_dt_obj(dt)->oo_inode;
1333         struct osd_device      *osd   = osd_obj2dev(osd_dt_obj(dt));
1334         ktime_t start, end;
1335         s64 timediff;
1336         ssize_t isize;
1337         __s64  maxidx;
1338         int i, rc = 0;
1339
1340         LASSERT(inode);
1341
1342         rc = osd_init_iobuf(osd, iobuf, 0, npages);
1343         if (unlikely(rc != 0))
1344                 RETURN(rc);
1345
1346         isize = i_size_read(inode);
1347         maxidx = ((isize + PAGE_SIZE - 1) >> PAGE_SHIFT) - 1;
1348
1349         start = ktime_get();
1350         for (i = 0; i < npages; i++) {
1351
1352                 /*
1353                  * till commit the content of the page is undefined
1354                  * we'll set it uptodate once bulk is done. otherwise
1355                  * subsequent reads can access non-stable data
1356                  */
1357                 ClearPageUptodate(lnb[i].lnb_page);
1358
1359                 if (lnb[i].lnb_len == PAGE_SIZE)
1360                         continue;
1361
1362                 if (maxidx >= lnb[i].lnb_page->index) {
1363                         osd_iobuf_add_page(iobuf, &lnb[i]);
1364                 } else {
1365                         long off;
1366                         char *p = kmap(lnb[i].lnb_page);
1367
1368                         off = lnb[i].lnb_page_offset;
1369                         if (off)
1370                                 memset(p, 0, off);
1371                         off = (lnb[i].lnb_page_offset + lnb[i].lnb_len) &
1372                               ~PAGE_MASK;
1373                         if (off)
1374                                 memset(p + off, 0, PAGE_SIZE - off);
1375                         kunmap(lnb[i].lnb_page);
1376                 }
1377         }
1378         end = ktime_get();
1379         timediff = ktime_us_delta(end, start);
1380         lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
1381
1382         if (iobuf->dr_npages) {
1383                 rc = osd_ldiskfs_map_inode_pages(inode, iobuf, osd, 0,
1384                                                  0, 0, NULL);
1385                 if (likely(rc == 0)) {
1386                         rc = osd_do_bio(osd, inode, iobuf, 0, 0);
1387                         /* do IO stats for preparation reads */
1388                         osd_fini_iobuf(osd, iobuf);
1389                 }
1390         }
1391         RETURN(rc);
1392 }
1393
1394 struct osd_fextent {
1395         sector_t        start;
1396         sector_t        end;
1397         __u32           flags;
1398         unsigned int    mapped:1;
1399 };
1400
1401 static int osd_is_mapped(struct dt_object *dt, __u64 offset,
1402                          struct osd_fextent *cached_extent)
1403 {
1404         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1405         sector_t block = offset >> inode->i_blkbits;
1406         sector_t start;
1407         struct fiemap_extent_info fei = { 0 };
1408         struct fiemap_extent fe = { 0 };
1409         int rc;
1410
1411         if (block >= cached_extent->start && block < cached_extent->end)
1412                 return cached_extent->mapped;
1413
1414         if (i_size_read(inode) == 0)
1415                 return 0;
1416
1417         /* Beyond EOF, must not be mapped */
1418         if (((i_size_read(inode) - 1) >> inode->i_blkbits) < block)
1419                 return 0;
1420
1421         fei.fi_extents_max = 1;
1422         fei.fi_extents_start = &fe;
1423
1424         rc = inode->i_op->fiemap(inode, &fei, offset, FIEMAP_MAX_OFFSET-offset);
1425         if (rc != 0)
1426                 return 0;
1427
1428         start = fe.fe_logical >> inode->i_blkbits;
1429         cached_extent->flags = fe.fe_flags;
1430         if (fei.fi_extents_mapped == 0) {
1431                 /* a special case - no extent found at this offset and forward.
1432                  * we can consider this as a hole to EOF. it's safe to cache
1433                  * as other threads can not allocate/punch blocks this thread
1434                  * is working on (LDLM). */
1435                 cached_extent->start = block;
1436                 cached_extent->end = i_size_read(inode) >> inode->i_blkbits;
1437                 cached_extent->mapped = 0;
1438                 return 0;
1439         }
1440
1441         if (start > block) {
1442                 cached_extent->start = block;
1443                 cached_extent->end = start;
1444                 cached_extent->mapped = 0;
1445         } else {
1446                 cached_extent->start = start;
1447                 cached_extent->end = (fe.fe_logical + fe.fe_length) >>
1448                                       inode->i_blkbits;
1449                 cached_extent->mapped = 1;
1450         }
1451
1452         return cached_extent->mapped;
1453 }
1454
1455 #define MAX_EXTENTS_PER_WRITE 100
1456 static int osd_declare_write_commit(const struct lu_env *env,
1457                                     struct dt_object *dt,
1458                                     struct niobuf_local *lnb, int npages,
1459                                     struct thandle *handle)
1460 {
1461         const struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1462         struct inode            *inode = osd_dt_obj(dt)->oo_inode;
1463         struct osd_thandle      *oh;
1464         int                     extents = 0, new_meta = 0;
1465         int                     depth, new_blocks = 0;
1466         int                     i;
1467         int                     dirty_groups = 0;
1468         int                     rc = 0;
1469         int                     credits = 0;
1470         long long               quota_space = 0;
1471         struct osd_fextent      mapped = { 0 }, extent = { 0 };
1472         enum osd_quota_local_flags local_flags = 0;
1473         enum osd_qid_declare_flags declare_flags = OSD_QID_BLK;
1474         unsigned int            extent_bytes;
1475         ENTRY;
1476
1477         LASSERT(handle != NULL);
1478         oh = container_of(handle, struct osd_thandle, ot_super);
1479         LASSERT(oh->ot_handle == NULL);
1480
1481         /*
1482          * We track a decaying average extent blocks per filesystem,
1483          * for most of time, it will be 1M, with filesystem becoming
1484          * heavily-fragmented, it will be reduced to 4K at the worst.
1485          */
1486         extent_bytes = osd_extent_bytes(osd);
1487         LASSERT(extent_bytes >= osd_sb(osd)->s_blocksize);
1488
1489         /* calculate number of extents (probably better to pass nb) */
1490         for (i = 0; i < npages; i++) {
1491                 /* ignore quota for the whole request if any page is from
1492                  * client cache or written by root.
1493                  *
1494                  * XXX we could handle this on per-lnb basis as done by
1495                  * grant.
1496                  */
1497                 if ((lnb[i].lnb_flags & OBD_BRW_NOQUOTA) ||
1498                     (lnb[i].lnb_flags & OBD_BRW_SYS_RESOURCE) ||
1499                     !(lnb[i].lnb_flags & OBD_BRW_SYNC))
1500                         declare_flags |= OSD_QID_FORCE;
1501
1502                 /*
1503                  * Convert unwritten extent might need split extents, could
1504                  * not skip it.
1505                  */
1506                 if (osd_is_mapped(dt, lnb[i].lnb_file_offset, &mapped) &&
1507                     !(mapped.flags & FIEMAP_EXTENT_UNWRITTEN)) {
1508                         lnb[i].lnb_flags |= OBD_BRW_MAPPED;
1509                         continue;
1510                 }
1511
1512                 if (lnb[i].lnb_flags & OBD_BRW_DONE) {
1513                         lnb[i].lnb_flags |= OBD_BRW_MAPPED;
1514                         continue;
1515                 }
1516
1517                 /* count only unmapped changes */
1518                 new_blocks++;
1519                 if (lnb[i].lnb_file_offset != extent.end || extent.end == 0) {
1520                         if (extent.end != 0)
1521                                 extents += (extent.end - extent.start +
1522                                             extent_bytes - 1) / extent_bytes;
1523                         extent.start = lnb[i].lnb_file_offset;
1524                         extent.end = lnb[i].lnb_file_offset + lnb[i].lnb_len;
1525                 } else {
1526                         extent.end += lnb[i].lnb_len;
1527                 }
1528
1529                 quota_space += PAGE_SIZE;
1530         }
1531
1532         credits++; /* inode */
1533         /*
1534          * overwrite case, no need to modify tree and
1535          * allocate blocks.
1536          */
1537         if (!extent.end)
1538                 goto out_declare;
1539
1540         extents += (extent.end - extent.start +
1541                     extent_bytes - 1) / extent_bytes;
1542         /**
1543          * with system space usage growing up, mballoc codes won't
1544          * try best to scan block group to align best free extent as
1545          * we can. So extent bytes per extent could be decayed to a
1546          * very small value, this could make us reserve too many credits.
1547          * We could be more optimistic in the credit reservations, even
1548          * in a case where the filesystem is nearly full, it is extremely
1549          * unlikely that the worst case would ever be hit.
1550          */
1551         if (extents > MAX_EXTENTS_PER_WRITE)
1552                 extents = MAX_EXTENTS_PER_WRITE;
1553
1554         /**
1555          * If we add a single extent, then in the worse case, each tree
1556          * level index/leaf need to be changed in case of the tree split.
1557          * If more extents are inserted, they could cause the whole tree
1558          * split more than once, but this is really rare.
1559          */
1560         if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL) {
1561                 /*
1562                  * many concurrent threads may grow tree by the time
1563                  * our transaction starts. so, consider 2 is a min depth.
1564                  */
1565                 depth = ext_depth(inode);
1566                 depth = min(max(depth, 1) + 1, LDISKFS_MAX_EXTENT_DEPTH);
1567                 if (extents <= 1) {
1568                         credits += depth * 2 * extents;
1569                         new_meta = depth;
1570                 } else {
1571                         credits += depth * 3 * extents;
1572                         new_meta = depth * 2 * extents;
1573                 }
1574         } else {
1575                 /*
1576                  * With N contiguous data blocks, we need at most
1577                  * N/EXT4_ADDR_PER_BLOCK(inode->i_sb) + 1 indirect blocks,
1578                  * 2 dindirect blocks, and 1 tindirect block
1579                  */
1580                 new_meta = DIV_ROUND_UP(new_blocks,
1581                                 LDISKFS_ADDR_PER_BLOCK(inode->i_sb)) + 4;
1582                 credits += new_meta;
1583         }
1584         dirty_groups += (extents + new_meta);
1585
1586         oh->oh_declared_ext = extents;
1587
1588         /* quota space for metadata blocks */
1589         quota_space += new_meta * LDISKFS_BLOCK_SIZE(osd_sb(osd));
1590
1591         /* quota space should be reported in 1K blocks */
1592         quota_space = toqb(quota_space);
1593
1594         /* each new block can go in different group (bitmap + gd) */
1595
1596         /* we can't dirty more bitmap blocks than exist */
1597         if (dirty_groups > LDISKFS_SB(osd_sb(osd))->s_groups_count)
1598                 credits += LDISKFS_SB(osd_sb(osd))->s_groups_count;
1599         else
1600                 credits += dirty_groups;
1601
1602         /* we can't dirty more gd blocks than exist */
1603         if (dirty_groups > LDISKFS_SB(osd_sb(osd))->s_gdb_count)
1604                 credits += LDISKFS_SB(osd_sb(osd))->s_gdb_count;
1605         else
1606                 credits += dirty_groups;
1607
1608         CDEBUG(D_INODE,
1609                "%s: inode #%lu extent_bytes %u extents %d credits %d\n",
1610                osd_ino2name(inode), inode->i_ino, extent_bytes, extents,
1611                credits);
1612
1613 out_declare:
1614         osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
1615
1616         /* make sure the over quota flags were not set */
1617         lnb[0].lnb_flags &= ~OBD_BRW_OVER_ALLQUOTA;
1618
1619         rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
1620                                    i_projid_read(inode), quota_space, oh,
1621                                    osd_dt_obj(dt), &local_flags, declare_flags);
1622
1623         /* we need only to store the overquota flags in the first lnb for
1624          * now, once we support multiple objects BRW, this code needs be
1625          * revised.
1626          */
1627         if (local_flags & QUOTA_FL_OVER_USRQUOTA)
1628                 lnb[0].lnb_flags |= OBD_BRW_OVER_USRQUOTA;
1629         if (local_flags & QUOTA_FL_OVER_GRPQUOTA)
1630                 lnb[0].lnb_flags |= OBD_BRW_OVER_GRPQUOTA;
1631         if (local_flags & QUOTA_FL_OVER_PRJQUOTA)
1632                 lnb[0].lnb_flags |= OBD_BRW_OVER_PRJQUOTA;
1633
1634         if (rc == 0)
1635                 rc = osd_trunc_lock(osd_dt_obj(dt), oh, true);
1636
1637         RETURN(rc);
1638 }
1639
1640 /* Check if a block is allocated or not */
1641 static int osd_write_commit(const struct lu_env *env, struct dt_object *dt,
1642                             struct niobuf_local *lnb, int npages,
1643                             struct thandle *thandle, __u64 user_size)
1644 {
1645         struct osd_thread_info *oti = osd_oti_get(env);
1646         struct osd_iobuf *iobuf = &oti->oti_iobuf;
1647         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1648         struct osd_device  *osd = osd_obj2dev(osd_dt_obj(dt));
1649         int rc = 0, i, check_credits = 0;
1650
1651         LASSERT(inode);
1652
1653         rc = osd_init_iobuf(osd, iobuf, 1, npages);
1654         if (unlikely(rc != 0))
1655                 RETURN(rc);
1656
1657         dquot_initialize(inode);
1658
1659         for (i = 0; i < npages; i++) {
1660                 if (lnb[i].lnb_rc == -ENOSPC &&
1661                     (lnb[i].lnb_flags & OBD_BRW_MAPPED)) {
1662                         /* Allow the write to proceed if overwriting an
1663                          * existing block
1664                          */
1665                         lnb[i].lnb_rc = 0;
1666                 }
1667
1668                 if (lnb[i].lnb_rc) { /* ENOSPC, network RPC error, etc. */
1669                         CDEBUG(D_INODE, "Skipping [%d] == %d\n", i,
1670                                lnb[i].lnb_rc);
1671                         LASSERT(lnb[i].lnb_page);
1672                         generic_error_remove_page(inode->i_mapping,
1673                                                   lnb[i].lnb_page);
1674                         continue;
1675                 }
1676
1677                 if (lnb[i].lnb_flags & OBD_BRW_DONE)
1678                         continue;
1679
1680                 if (!(lnb[i].lnb_flags & OBD_BRW_MAPPED))
1681                         check_credits = 1;
1682
1683                 LASSERT(PageLocked(lnb[i].lnb_page));
1684                 LASSERT(!PageWriteback(lnb[i].lnb_page));
1685
1686                 /*
1687                  * Since write and truncate are serialized by oo_sem, even
1688                  * partial-page truncate should not leave dirty pages in the
1689                  * page cache.
1690                  */
1691                 LASSERT(!PageDirty(lnb[i].lnb_page));
1692
1693                 SetPageUptodate(lnb[i].lnb_page);
1694
1695                 osd_iobuf_add_page(iobuf, &lnb[i]);
1696         }
1697
1698         osd_trans_exec_op(env, thandle, OSD_OT_WRITE);
1699
1700         if (OBD_FAIL_CHECK(OBD_FAIL_OST_MAPBLK_ENOSPC)) {
1701                 rc = -ENOSPC;
1702         } else if (iobuf->dr_npages > 0) {
1703                 rc = osd_ldiskfs_map_inode_pages(inode, iobuf, osd,
1704                                                  1, user_size,
1705                                                  check_credits,
1706                                                  thandle);
1707         } else {
1708                 /* no pages to write, no transno is needed */
1709                 thandle->th_local = 1;
1710         }
1711
1712         if (rc != 0 && !thandle->th_restart_tran)
1713                 osd_fini_iobuf(osd, iobuf);
1714
1715         osd_trans_exec_check(env, thandle, OSD_OT_WRITE);
1716
1717         if (unlikely(rc != 0 && !thandle->th_restart_tran)) {
1718                 /* if write fails, we should drop pages from the cache */
1719                 for (i = 0; i < npages; i++) {
1720                         if (lnb[i].lnb_page == NULL)
1721                                 continue;
1722                         if (!PagePrivate2(lnb[i].lnb_page)) {
1723                                 LASSERT(PageLocked(lnb[i].lnb_page));
1724                                 generic_error_remove_page(inode->i_mapping,
1725                                                           lnb[i].lnb_page);
1726                         }
1727                 }
1728         }
1729
1730         RETURN(rc);
1731 }
1732
1733 static int osd_read_prep(const struct lu_env *env, struct dt_object *dt,
1734                          struct niobuf_local *lnb, int npages)
1735 {
1736         struct osd_thread_info *oti = osd_oti_get(env);
1737         struct osd_iobuf *iobuf = &oti->oti_iobuf;
1738         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1739         struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1740         int rc = 0, i, cache_hits = 0, cache_misses = 0;
1741         ktime_t start, end;
1742         s64 timediff;
1743         loff_t isize;
1744
1745         LASSERT(inode);
1746
1747         rc = osd_init_iobuf(osd, iobuf, 0, npages);
1748         if (unlikely(rc != 0))
1749                 RETURN(rc);
1750
1751         isize = i_size_read(inode);
1752
1753         start = ktime_get();
1754         for (i = 0; i < npages; i++) {
1755
1756                 if (isize <= lnb[i].lnb_file_offset)
1757                         /* If there's no more data, abort early.
1758                          * lnb->lnb_rc == 0, so it's easy to detect later.
1759                          */
1760                         break;
1761
1762                 /* instead of looking if we go beyong isize, send complete
1763                  * pages all the time
1764                  */
1765                 lnb[i].lnb_rc = lnb[i].lnb_len;
1766
1767                 /* Bypass disk read if fail_loc is set properly */
1768                 if (OBD_FAIL_CHECK_QUIET(OBD_FAIL_OST_FAKE_RW))
1769                         SetPageUptodate(lnb[i].lnb_page);
1770
1771                 if (PageUptodate(lnb[i].lnb_page)) {
1772                         cache_hits++;
1773                         unlock_page(lnb[i].lnb_page);
1774                 } else {
1775                         cache_misses++;
1776                         osd_iobuf_add_page(iobuf, &lnb[i]);
1777                 }
1778                 /* no need to unlock in osd_bufs_put(), the sooner page is
1779                  * unlocked, the earlier another client can access it.
1780                  * notice real unlock_page() can be called few lines
1781                  * below after osd_do_bio(). lnb is a per-thread, so it's
1782                  * fine to have PG_locked and lnb_locked inconsistent here
1783                  */
1784                 lnb[i].lnb_locked = 0;
1785         }
1786         end = ktime_get();
1787         timediff = ktime_us_delta(end, start);
1788         lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
1789
1790         if (cache_hits != 0)
1791                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_HIT,
1792                                     cache_hits);
1793         if (cache_misses != 0)
1794                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_MISS,
1795                                     cache_misses);
1796         if (cache_hits + cache_misses != 0)
1797                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_ACCESS,
1798                                     cache_hits + cache_misses);
1799
1800         if (iobuf->dr_npages) {
1801                 rc = osd_ldiskfs_map_inode_pages(inode, iobuf, osd, 0,
1802                                                  0, 0, NULL);
1803                 if (!rc)
1804                         rc = osd_do_bio(osd, inode, iobuf, 0, 0);
1805
1806                 /* IO stats will be done in osd_bufs_put() */
1807
1808                 /* early release to let others read data during the bulk */
1809                 for (i = 0; i < iobuf->dr_npages; i++) {
1810                         LASSERT(PageLocked(iobuf->dr_pages[i]));
1811                         if (!PagePrivate2(iobuf->dr_pages[i]))
1812                                 unlock_page(iobuf->dr_pages[i]);
1813                 }
1814         }
1815
1816         RETURN(rc);
1817 }
1818
1819 /*
1820  * XXX: Another layering violation for now.
1821  *
1822  * We don't want to use ->f_op->read methods, because generic file write
1823  *
1824  *         - serializes on ->i_sem, and
1825  *
1826  *         - does a lot of extra work like balance_dirty_pages(),
1827  *
1828  * which doesn't work for globally shared files like /last_rcvd.
1829  */
1830 static int osd_ldiskfs_readlink(struct inode *inode, char *buffer, int buflen)
1831 {
1832         struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
1833
1834         memcpy(buffer, (char *)ei->i_data, buflen);
1835
1836         return  buflen;
1837 }
1838
1839 int osd_ldiskfs_read(struct inode *inode, void *buf, int size, loff_t *offs)
1840 {
1841         struct buffer_head *bh;
1842         unsigned long block;
1843         int osize;
1844         int blocksize;
1845         int csize;
1846         int boffs;
1847
1848         /* prevent reading after eof */
1849         spin_lock(&inode->i_lock);
1850         if (i_size_read(inode) < *offs + size) {
1851                 loff_t diff = i_size_read(inode) - *offs;
1852
1853                 spin_unlock(&inode->i_lock);
1854                 if (diff < 0) {
1855                         CDEBUG(D_OTHER,
1856                                "size %llu is too short to read @%llu\n",
1857                                i_size_read(inode), *offs);
1858                         return -EBADR;
1859                 } else if (diff == 0) {
1860                         return 0;
1861                 } else {
1862                         size = diff;
1863                 }
1864         } else {
1865                 spin_unlock(&inode->i_lock);
1866         }
1867
1868         blocksize = 1 << inode->i_blkbits;
1869         osize = size;
1870         while (size > 0) {
1871                 block = *offs >> inode->i_blkbits;
1872                 boffs = *offs & (blocksize - 1);
1873                 csize = min(blocksize - boffs, size);
1874                 bh = __ldiskfs_bread(NULL, inode, block, 0);
1875                 if (IS_ERR(bh)) {
1876                         CERROR("%s: can't read %u@%llu on ino %lu: rc = %ld\n",
1877                                osd_ino2name(inode), csize, *offs, inode->i_ino,
1878                                PTR_ERR(bh));
1879                         return PTR_ERR(bh);
1880                 }
1881
1882                 if (bh != NULL) {
1883                         memcpy(buf, bh->b_data + boffs, csize);
1884                         brelse(bh);
1885                 } else {
1886                         memset(buf, 0, csize);
1887                 }
1888
1889                 *offs += csize;
1890                 buf += csize;
1891                 size -= csize;
1892         }
1893         return osize;
1894 }
1895
1896 static ssize_t osd_read(const struct lu_env *env, struct dt_object *dt,
1897                         struct lu_buf *buf, loff_t *pos)
1898 {
1899         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1900         int rc;
1901
1902         /* Read small symlink from inode body as we need to maintain correct
1903          * on-disk symlinks for ldiskfs.
1904          */
1905         if (S_ISLNK(dt->do_lu.lo_header->loh_attr)) {
1906                 loff_t size = i_size_read(inode);
1907
1908                 if (buf->lb_len < size)
1909                         return -EOVERFLOW;
1910
1911                 if (size < sizeof(LDISKFS_I(inode)->i_data))
1912                         rc = osd_ldiskfs_readlink(inode, buf->lb_buf, size);
1913                 else
1914                         rc = osd_ldiskfs_read(inode, buf->lb_buf, size, pos);
1915         } else {
1916                 rc = osd_ldiskfs_read(inode, buf->lb_buf, buf->lb_len, pos);
1917         }
1918
1919         return rc;
1920 }
1921
1922 static inline int osd_extents_enabled(struct super_block *sb,
1923                                       struct inode *inode)
1924 {
1925         if (inode != NULL) {
1926                 if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL)
1927                         return 1;
1928         } else if (ldiskfs_has_feature_extents(sb)) {
1929                 return 1;
1930         }
1931         return 0;
1932 }
1933
1934 int osd_calc_bkmap_credits(struct super_block *sb, struct inode *inode,
1935                            const loff_t size, const loff_t pos,
1936                            const int blocks)
1937 {
1938         int credits, bits, bs, i;
1939
1940         bits = sb->s_blocksize_bits;
1941         bs = 1 << bits;
1942
1943         /* legacy blockmap: 3 levels * 3 (bitmap,gd,itself)
1944          * we do not expect blockmaps on the large files,
1945          * so let's shrink it to 2 levels (4GB files)
1946          */
1947
1948         /* this is default reservation: 2 levels */
1949         credits = (blocks + 2) * 3;
1950
1951         /* actual offset is unknown, hard to optimize */
1952         if (pos == -1)
1953                 return credits;
1954
1955         /* now check for few specific cases to optimize */
1956         if (pos + size <= LDISKFS_NDIR_BLOCKS * bs) {
1957                 /* no indirects */
1958                 credits = blocks;
1959                 /* allocate if not allocated */
1960                 if (inode == NULL) {
1961                         credits += blocks * 2;
1962                         return credits;
1963                 }
1964                 for (i = (pos >> bits); i < (pos >> bits) + blocks; i++) {
1965                         LASSERT(i < LDISKFS_NDIR_BLOCKS);
1966                         if (LDISKFS_I(inode)->i_data[i] == 0)
1967                                 credits += 2;
1968                 }
1969         } else if (pos + size <= (LDISKFS_NDIR_BLOCKS + 1024) * bs) {
1970                 /* single indirect */
1971                 credits = blocks * 3;
1972                 if (inode == NULL ||
1973                     LDISKFS_I(inode)->i_data[LDISKFS_IND_BLOCK] == 0)
1974                         credits += 3;
1975                 else
1976                         /* The indirect block may be modified. */
1977                         credits += 1;
1978         }
1979
1980         return credits;
1981 }
1982
1983 static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt,
1984                                  const struct lu_buf *buf, loff_t _pos,
1985                                  struct thandle *handle)
1986 {
1987         struct osd_object  *obj  = osd_dt_obj(dt);
1988         struct inode       *inode = obj->oo_inode;
1989         struct super_block *sb = osd_sb(osd_obj2dev(obj));
1990         struct osd_thandle *oh;
1991         int                 rc = 0, est = 0, credits, blocks, allocated = 0;
1992         int                 bits, bs;
1993         int                 depth, size;
1994         loff_t              pos;
1995         ENTRY;
1996
1997         LASSERT(buf != NULL);
1998         LASSERT(handle != NULL);
1999
2000         oh = container_of(handle, struct osd_thandle, ot_super);
2001         LASSERT(oh->ot_handle == NULL);
2002
2003         size = buf->lb_len;
2004         bits = sb->s_blocksize_bits;
2005         bs = 1 << bits;
2006
2007         if (_pos == -1) {
2008                 /* if this is an append, then we
2009                  * should expect cross-block record
2010                  */
2011                 pos = 0;
2012         } else {
2013                 pos = _pos;
2014         }
2015
2016         /* blocks to modify */
2017         blocks = ((pos + size + bs - 1) >> bits) - (pos >> bits);
2018         LASSERT(blocks > 0);
2019
2020         if (inode != NULL && _pos != -1) {
2021                 /* object size in blocks */
2022                 est = (i_size_read(inode) + bs - 1) >> bits;
2023                 allocated = inode->i_blocks >> (bits - 9);
2024                 if (pos + size <= i_size_read(inode) && est <= allocated) {
2025                         /* looks like an overwrite, no need to modify tree */
2026                         credits = blocks;
2027                         /* no need to modify i_size */
2028                         goto out;
2029                 }
2030         }
2031
2032         if (osd_extents_enabled(sb, inode)) {
2033                 /*
2034                  * many concurrent threads may grow tree by the time
2035                  * our transaction starts. so, consider 2 is a min depth
2036                  * for every level we may need to allocate a new block
2037                  * and take some entries from the old one. so, 3 blocks
2038                  * to allocate (bitmap, gd, itself) + old block - 4 per
2039                  * level.
2040                  */
2041                 depth = inode != NULL ? ext_depth(inode) : 0;
2042                 depth = min(max(depth, 1) + 3, LDISKFS_MAX_EXTENT_DEPTH);
2043                 credits = depth;
2044                 /* if not append, then split may need to modify
2045                  * existing blocks moving entries into the new ones
2046                  */
2047                 if (_pos != -1)
2048                         credits += depth;
2049                 /* blocks to store data: bitmap,gd,itself */
2050                 credits += blocks * 3;
2051         } else {
2052                 credits = osd_calc_bkmap_credits(sb, inode, size, _pos, blocks);
2053         }
2054         /* if inode is created as part of the transaction,
2055          * then it's counted already by the creation method
2056          */
2057         if (inode != NULL)
2058                 credits++;
2059
2060 out:
2061
2062         osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
2063
2064         /* dt_declare_write() is usually called for system objects, such
2065          * as llog or last_rcvd files. We needn't enforce quota on those
2066          * objects, so always set the lqi_space as 0.
2067          */
2068         if (inode != NULL)
2069                 rc = osd_declare_inode_qid(env, i_uid_read(inode),
2070                                            i_gid_read(inode),
2071                                            i_projid_read(inode), 0,
2072                                            oh, obj, NULL, OSD_QID_BLK);
2073
2074         if (rc == 0)
2075                 rc = osd_trunc_lock(obj, oh, true);
2076
2077         RETURN(rc);
2078 }
2079
2080 static int osd_ldiskfs_writelink(struct inode *inode, char *buffer, int buflen)
2081 {
2082         /* LU-2634: clear the extent format for fast symlink */
2083         ldiskfs_clear_inode_flag(inode, LDISKFS_INODE_EXTENTS);
2084
2085         /* Copying the NUL byte terminating the link target as well */
2086         memcpy((char *)&LDISKFS_I(inode)->i_data, (char *)buffer, buflen + 1);
2087         spin_lock(&inode->i_lock);
2088         LDISKFS_I(inode)->i_disksize = buflen;
2089         i_size_write(inode, buflen);
2090         spin_unlock(&inode->i_lock);
2091         osd_dirty_inode(inode, I_DIRTY_DATASYNC);
2092
2093         return 0;
2094 }
2095
2096 static int osd_ldiskfs_write_record(struct dt_object *dt, void *buf,
2097                                     int bufsize, int write_NUL, loff_t *offs,
2098                                     handle_t *handle)
2099 {
2100         struct inode *inode = osd_dt_obj(dt)->oo_inode;
2101         struct buffer_head *bh        = NULL;
2102         loff_t              offset    = *offs;
2103         loff_t              new_size  = i_size_read(inode);
2104         unsigned long       block;
2105         int                 blocksize = 1 << inode->i_blkbits;
2106         struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
2107         int                 err = 0;
2108         int                 size;
2109         int                 boffs;
2110         int                 dirty_inode = 0;
2111         bool create, sparse, sync = false;
2112
2113         if (write_NUL) {
2114                 /*
2115                  * long symlink write does not count the NUL terminator in
2116                  * bufsize, we write it, and the inode's file size does not
2117                  * count the NUL terminator as well.
2118                  */
2119                 ((char *)buf)[bufsize] = '\0';
2120                 ++bufsize;
2121         }
2122
2123         /* only the first flag-set matters */
2124         dirty_inode = !test_and_set_bit(LDISKFS_INODE_JOURNAL_DATA,
2125                                        &ei->i_flags);
2126
2127         /* sparse checking is racy, but sparse is very rare case, leave as is */
2128         sparse = (new_size > 0 && (inode->i_blocks >> (inode->i_blkbits - 9)) <
2129                   ((new_size - 1) >> inode->i_blkbits) + 1);
2130
2131         while (bufsize > 0) {
2132                 int credits = handle->h_buffer_credits;
2133                 unsigned long last_block = (new_size == 0) ? 0 :
2134                                            (new_size - 1) >> inode->i_blkbits;
2135
2136                 if (bh)
2137                         brelse(bh);
2138
2139                 block = offset >> inode->i_blkbits;
2140                 boffs = offset & (blocksize - 1);
2141                 size = min(blocksize - boffs, bufsize);
2142                 sync = (block > last_block || new_size == 0 || sparse);
2143
2144                 if (sync)
2145                         down(&ei->i_append_sem);
2146
2147                 bh = __ldiskfs_bread(handle, inode, block, 0);
2148
2149                 if (unlikely(IS_ERR_OR_NULL(bh) && !sync))
2150                         CWARN(
2151                               "%s: adding bh without locking off %llu (block %lu, size %d, offs %llu)\n",
2152                               osd_ino2name(inode),
2153                               offset, block, bufsize, *offs);
2154
2155                 if (IS_ERR_OR_NULL(bh)) {
2156                         struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
2157                         int flags = LDISKFS_GET_BLOCKS_CREATE;
2158
2159                         /* while the file system is being mounted, avoid
2160                          * preallocation otherwise mount can take a long
2161                          * time as mballoc cache is cold.
2162                          * XXX: this is a workaround until we have a proper
2163                          *      fix in mballoc
2164                          * XXX: works with extent-based files only */
2165                         if (!osd->od_cl_seq)
2166                                 flags |= LDISKFS_GET_BLOCKS_NO_NORMALIZE;
2167                         bh = __ldiskfs_bread(handle, inode, block, flags);
2168                         create = true;
2169                 } else {
2170                         if (sync) {
2171                                 up(&ei->i_append_sem);
2172                                 sync = false;
2173                         }
2174                         create = false;
2175                 }
2176                 if (IS_ERR_OR_NULL(bh)) {
2177                         if (bh == NULL) {
2178                                 err = -EIO;
2179                         } else {
2180                                 err = PTR_ERR(bh);
2181                                 bh = NULL;
2182                         }
2183
2184                         CERROR(
2185                                "%s: error reading offset %llu (block %lu, size %d, offs %llu), credits %d/%d: rc = %d\n",
2186                                osd_ino2name(inode), offset, block, bufsize,
2187                                *offs, credits, handle->h_buffer_credits, err);
2188                         break;
2189                 }
2190
2191                 err = osd_ldiskfs_journal_get_write_access(handle, inode->i_sb,
2192                                                            bh,
2193                                                            LDISKFS_JTR_NONE);
2194                 if (err) {
2195                         CERROR("journal_get_write_access() returned error %d\n",
2196                                err);
2197                         break;
2198                 }
2199                 LASSERTF(boffs + size <= bh->b_size,
2200                          "boffs %d size %d bh->b_size %lu\n",
2201                          boffs, size, (unsigned long)bh->b_size);
2202                 if (create) {
2203                         memset(bh->b_data, 0, bh->b_size);
2204                         if (sync) {
2205                                 up(&ei->i_append_sem);
2206                                 sync = false;
2207                         }
2208                 }
2209                 memcpy(bh->b_data + boffs, buf, size);
2210                 err = ldiskfs_handle_dirty_metadata(handle, NULL, bh);
2211                 if (err)
2212                         break;
2213
2214                 if (offset + size > new_size)
2215                         new_size = offset + size;
2216                 offset += size;
2217                 bufsize -= size;
2218                 buf += size;
2219         }
2220         if (sync)
2221                 up(&ei->i_append_sem);
2222
2223         if (bh)
2224                 brelse(bh);
2225
2226         if (write_NUL)
2227                 --new_size;
2228         /* correct in-core and on-disk sizes */
2229         if (new_size > i_size_read(inode)) {
2230                 spin_lock(&inode->i_lock);
2231                 if (new_size > i_size_read(inode))
2232                         i_size_write(inode, new_size);
2233                 if (i_size_read(inode) > ei->i_disksize) {
2234                         ei->i_disksize = i_size_read(inode);
2235                         dirty_inode = 1;
2236                 }
2237                 spin_unlock(&inode->i_lock);
2238         }
2239         if (dirty_inode)
2240                 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
2241
2242         if (err == 0)
2243                 *offs = offset;
2244         return err;
2245 }
2246
2247 static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt,
2248                          const struct lu_buf *buf, loff_t *pos,
2249                          struct thandle *handle)
2250 {
2251         struct inode            *inode = osd_dt_obj(dt)->oo_inode;
2252         struct osd_thandle      *oh;
2253         ssize_t                 result;
2254         int                     is_link;
2255
2256         LASSERT(dt_object_exists(dt));
2257
2258         LASSERT(handle != NULL);
2259         LASSERT(inode != NULL);
2260         dquot_initialize(inode);
2261
2262         /* XXX: don't check: one declared chunk can be used many times */
2263         /* osd_trans_exec_op(env, handle, OSD_OT_WRITE); */
2264
2265         oh = container_of(handle, struct osd_thandle, ot_super);
2266         LASSERT(oh->ot_handle->h_transaction != NULL);
2267         osd_trans_exec_op(env, handle, OSD_OT_WRITE);
2268
2269         /* Write small symlink to inode body as we need to maintain correct
2270          * on-disk symlinks for ldiskfs.
2271          * Note: the buf->lb_buf contains a NUL terminator while buf->lb_len
2272          * does not count it in.
2273          */
2274         is_link = S_ISLNK(dt->do_lu.lo_header->loh_attr);
2275         if (is_link && (buf->lb_len < sizeof(LDISKFS_I(inode)->i_data)))
2276                 result = osd_ldiskfs_writelink(inode, buf->lb_buf, buf->lb_len);
2277         else
2278                 result = osd_ldiskfs_write_record(dt, buf->lb_buf, buf->lb_len,
2279                                                   is_link, pos, oh->ot_handle);
2280         if (result == 0)
2281                 result = buf->lb_len;
2282
2283         osd_trans_exec_check(env, handle, OSD_OT_WRITE);
2284
2285         return result;
2286 }
2287
2288 static int osd_declare_fallocate(const struct lu_env *env,
2289                                  struct dt_object *dt, __u64 start, __u64 end,
2290                                  int mode, struct thandle *th)
2291 {
2292         struct osd_thandle *oh = container_of(th, struct osd_thandle, ot_super);
2293         struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
2294         struct inode *inode = osd_dt_obj(dt)->oo_inode;
2295         long long quota_space = 0;
2296         /* 5 is max tree depth. (inode + 4 index blocks) */
2297         int depth = 5;
2298         int rc;
2299
2300         ENTRY;
2301
2302         /*
2303          * mode == 0 (which is standard prealloc) and PUNCH is supported
2304          * Rest of mode options is not supported yet.
2305          */
2306         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2307                 RETURN(-EOPNOTSUPP);
2308
2309         /* disable fallocate completely */
2310         if (osd_dev(dt->do_lu.lo_dev)->od_fallocate_zero_blocks < 0)
2311                 RETURN(-EOPNOTSUPP);
2312
2313         LASSERT(th);
2314         LASSERT(inode);
2315
2316         if (mode & FALLOC_FL_PUNCH_HOLE) {
2317                 rc = osd_declare_inode_qid(env, i_uid_read(inode),
2318                                            i_gid_read(inode),
2319                                            i_projid_read(inode), 0, oh,
2320                                            osd_dt_obj(dt), NULL, OSD_QID_BLK);
2321                 if (rc == 0)
2322                         rc = osd_trunc_lock(osd_dt_obj(dt), oh, false);
2323                 RETURN(rc);
2324         }
2325
2326         /* quota space for metadata blocks
2327          * approximate metadata estimate should be good enough.
2328          */
2329         quota_space += PAGE_SIZE;
2330         quota_space += depth * LDISKFS_BLOCK_SIZE(osd_sb(osd));
2331
2332         /* quota space should be reported in 1K blocks */
2333         quota_space = toqb(quota_space) + toqb(end - start) +
2334                       LDISKFS_META_TRANS_BLOCKS(inode->i_sb);
2335
2336         /* We don't need to reserve credits for whole fallocate here.
2337          * We reserve space only for metadata. Fallocate credits are
2338          * extended as required
2339          */
2340         rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
2341                                    i_projid_read(inode), quota_space, oh,
2342                                    osd_dt_obj(dt), NULL, OSD_QID_BLK);
2343         RETURN(rc);
2344 }
2345
2346 static int osd_fallocate_preallocate(const struct lu_env *env,
2347                                      struct dt_object *dt,
2348                                      __u64 start, __u64 end, int mode,
2349                                      struct thandle *th)
2350 {
2351         struct osd_thandle *oh = container_of(th, struct osd_thandle, ot_super);
2352         handle_t *handle = ldiskfs_journal_current_handle();
2353         unsigned int save_credits = oh->ot_credits;
2354         struct osd_object *obj = osd_dt_obj(dt);
2355         struct inode *inode = obj->oo_inode;
2356         struct ldiskfs_map_blocks map;
2357         unsigned int credits;
2358         ldiskfs_lblk_t blen;
2359         ldiskfs_lblk_t boff;
2360         loff_t new_size = 0;
2361         int depth = 0;
2362         int flags;
2363         int rc = 0;
2364
2365         ENTRY;
2366
2367         LASSERT(dt_object_exists(dt));
2368         LASSERT(osd_invariant(obj));
2369         LASSERT(inode != NULL);
2370
2371         CDEBUG(D_INODE, "fallocate: inode #%lu: start %llu end %llu mode %d\n",
2372                inode->i_ino, start, end, mode);
2373
2374         dquot_initialize(inode);
2375
2376         LASSERT(th);
2377
2378         boff = start >> inode->i_blkbits;
2379         blen = (ALIGN(end, 1 << inode->i_blkbits) >> inode->i_blkbits) - boff;
2380
2381         /* Create and mark new extents as either zero or unwritten */
2382         flags = (osd_dev(dt->do_lu.lo_dev)->od_fallocate_zero_blocks ||
2383                  !ldiskfs_test_inode_flag(inode, LDISKFS_INODE_EXTENTS)) ?
2384                 LDISKFS_GET_BLOCKS_CREATE_ZERO :
2385                 LDISKFS_GET_BLOCKS_CREATE_UNWRIT_EXT;
2386 #ifdef LDISKFS_GET_BLOCKS_KEEP_SIZE
2387         if (mode & FALLOC_FL_KEEP_SIZE)
2388                 flags |= LDISKFS_GET_BLOCKS_KEEP_SIZE;
2389 #endif
2390         inode_lock(inode);
2391
2392         if (!(mode & FALLOC_FL_KEEP_SIZE) && (end > i_size_read(inode) ||
2393             end > LDISKFS_I(inode)->i_disksize)) {
2394                 new_size = end;
2395                 rc = inode_newsize_ok(inode, new_size);
2396                 if (rc)
2397                         GOTO(out, rc);
2398         }
2399
2400         inode_dio_wait(inode);
2401
2402         map.m_lblk = boff;
2403         map.m_len = blen;
2404
2405         /* Don't normalize the request if it can fit in one extent so
2406          * that it doesn't get unnecessarily split into multiple extents.
2407          */
2408         if (blen <= EXT_UNWRITTEN_MAX_LEN)
2409                 flags |= LDISKFS_GET_BLOCKS_NO_NORMALIZE;
2410
2411         /*
2412          * credits to insert 1 extent into extent tree.
2413          */
2414         credits = osd_chunk_trans_blocks(inode, blen);
2415         depth = ext_depth(inode);
2416
2417         while (rc >= 0 && blen) {
2418                 loff_t epos;
2419
2420                 /*
2421                  * Recalculate credits when extent tree depth changes.
2422                  */
2423                 if (depth != ext_depth(inode)) {
2424                         credits = osd_chunk_trans_blocks(inode, blen);
2425                         depth = ext_depth(inode);
2426                 }
2427
2428                 /* TODO: quota check */
2429                 rc = osd_extend_restart_trans(handle, credits, inode);
2430                 if (rc)
2431                         break;
2432
2433                 rc = ldiskfs_map_blocks(handle, inode, &map, flags);
2434                 if (rc <= 0) {
2435                         CDEBUG(D_INODE,
2436                                "inode #%lu: block %u: len %u: ldiskfs_map_blocks returned %d\n",
2437                                inode->i_ino, map.m_lblk, map.m_len, rc);
2438                         ldiskfs_mark_inode_dirty(handle, inode);
2439                         break;
2440                 }
2441
2442                 map.m_lblk += rc;
2443                 map.m_len = blen = blen - rc;
2444                 epos = (loff_t)map.m_lblk << inode->i_blkbits;
2445                 inode->i_ctime = current_time(inode);
2446                 if (new_size) {
2447                         if (epos > end)
2448                                 epos = end;
2449                         if (ldiskfs_update_inode_size(inode, epos) & 0x1)
2450                                 inode->i_mtime = inode->i_ctime;
2451 #ifdef LDISKFS_EOFBLOCKS_FL
2452                 } else {
2453                         if (epos > inode->i_size)
2454                                 ldiskfs_set_inode_flag(inode,
2455                                                        LDISKFS_INODE_EOFBLOCKS);
2456 #endif
2457                 }
2458
2459                 ldiskfs_mark_inode_dirty(handle, inode);
2460         }
2461
2462 out:
2463         /* extand credits if needed for operations such as attribute set */
2464         if (rc >= 0)
2465                 rc = osd_extend_restart_trans(handle, save_credits, inode);
2466
2467         inode_unlock(inode);
2468
2469         RETURN(rc);
2470 }
2471
2472 static int osd_fallocate_punch(const struct lu_env *env, struct dt_object *dt,
2473                                __u64 start, __u64 end, int mode,
2474                                struct thandle *th)
2475 {
2476         struct osd_object *obj = osd_dt_obj(dt);
2477         struct inode *inode = obj->oo_inode;
2478         struct osd_access_lock *al;
2479         struct osd_thandle *oh;
2480         int rc = 0, found = 0;
2481
2482         ENTRY;
2483
2484         LASSERT(dt_object_exists(dt));
2485         LASSERT(osd_invariant(obj));
2486         LASSERT(inode != NULL);
2487
2488         dquot_initialize(inode);
2489
2490         LASSERT(th);
2491         oh = container_of(th, struct osd_thandle, ot_super);
2492         LASSERT(oh->ot_handle->h_transaction != NULL);
2493
2494         list_for_each_entry(al, &oh->ot_trunc_locks, tl_list) {
2495                 if (obj != al->tl_obj)
2496                         continue;
2497                 LASSERT(al->tl_shared == 0);
2498                 found = 1;
2499                 /* do actual punch in osd_trans_stop() */
2500                 al->tl_start = start;
2501                 al->tl_end = end;
2502                 al->tl_mode = mode;
2503                 al->tl_punch = true;
2504                 break;
2505         }
2506
2507         RETURN(rc);
2508 }
2509
2510 static int osd_fallocate(const struct lu_env *env, struct dt_object *dt,
2511                          __u64 start, __u64 end, int mode, struct thandle *th)
2512 {
2513         int rc;
2514
2515         ENTRY;
2516
2517         if (mode & FALLOC_FL_PUNCH_HOLE) {
2518                 /* punch */
2519                 rc = osd_fallocate_punch(env, dt, start, end, mode, th);
2520         } else {
2521                 /* standard preallocate */
2522                 rc = osd_fallocate_preallocate(env, dt, start, end, mode, th);
2523         }
2524         RETURN(rc);
2525 }
2526
2527 static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt,
2528                              __u64 start, __u64 end, struct thandle *th)
2529 {
2530         struct osd_thandle *oh;
2531         struct osd_object  *obj = osd_dt_obj(dt);
2532         struct inode       *inode;
2533         int                 rc;
2534         ENTRY;
2535
2536         LASSERT(th);
2537         oh = container_of(th, struct osd_thandle, ot_super);
2538
2539         /*
2540          * we don't need to reserve credits for whole truncate
2541          * it's not possible as truncate may need to free too many
2542          * blocks and that won't fit a single transaction. instead
2543          * we reserve credits to change i_size and put inode onto
2544          * orphan list. if needed truncate will extend or restart
2545          * transaction
2546          */
2547         osd_trans_declare_op(env, oh, OSD_OT_PUNCH,
2548                              osd_dto_credits_noquota[DTO_ATTR_SET_BASE] + 3);
2549
2550         inode = obj->oo_inode;
2551         LASSERT(inode);
2552
2553         rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
2554                                    i_projid_read(inode), 0, oh, obj,
2555                                    NULL, OSD_QID_BLK);
2556
2557         /* if object holds encrypted content, we need to make sure we truncate
2558          * on an encryption unit boundary, or subsequent reads will get
2559          * corrupted content
2560          */
2561         if (rc == 0) {
2562                 if (obj->oo_lma_flags & LUSTRE_ENCRYPT_FL &&
2563                     start & ~LUSTRE_ENCRYPTION_MASK)
2564                         start = (start & LUSTRE_ENCRYPTION_MASK) +
2565                                 LUSTRE_ENCRYPTION_UNIT_SIZE;
2566                 ll_truncate_pagecache(inode, start);
2567                 rc = osd_trunc_lock(obj, oh, false);
2568         }
2569
2570         RETURN(rc);
2571 }
2572
2573 static int osd_punch(const struct lu_env *env, struct dt_object *dt,
2574                      __u64 start, __u64 end, struct thandle *th)
2575 {
2576         struct osd_object *obj = osd_dt_obj(dt);
2577         struct osd_device *osd = osd_obj2dev(obj);
2578         struct inode *inode = obj->oo_inode;
2579         struct osd_access_lock *al;
2580         struct osd_thandle *oh;
2581         int rc = 0, found = 0;
2582         bool grow = false;
2583         ENTRY;
2584
2585         LASSERT(dt_object_exists(dt));
2586         LASSERT(osd_invariant(obj));
2587         LASSERT(inode != NULL);
2588         dquot_initialize(inode);
2589
2590         LASSERT(th);
2591         oh = container_of(th, struct osd_thandle, ot_super);
2592         LASSERT(oh->ot_handle->h_transaction != NULL);
2593
2594         /* we used to skip truncate to current size to
2595          * optimize truncates on OST. with DoM we can
2596          * get attr_set to set specific size (MDS_REINT)
2597          * and then get truncate RPC which essentially
2598          * would be skipped. this is bad.. so, disable
2599          * this optimization on MDS till the client stop
2600          * to sent MDS_REINT (LU-11033) -bzzz
2601          */
2602         if (osd->od_is_ost && i_size_read(inode) == start)
2603                 RETURN(0);
2604
2605         osd_trans_exec_op(env, th, OSD_OT_PUNCH);
2606
2607         spin_lock(&inode->i_lock);
2608         if (i_size_read(inode) < start)
2609                 grow = true;
2610         i_size_write(inode, start);
2611         spin_unlock(&inode->i_lock);
2612
2613         /* optimize grow case */
2614         if (grow) {
2615                 osd_execute_truncate(obj);
2616                 GOTO(out, rc);
2617         }
2618
2619         inode_lock(inode);
2620         /* add to orphan list to ensure truncate completion
2621          * if this transaction succeed. ldiskfs_truncate()
2622          * will take the inode out of the list
2623          */
2624         rc = ldiskfs_orphan_add(oh->ot_handle, inode);
2625         inode_unlock(inode);
2626         if (rc != 0)
2627                 GOTO(out, rc);
2628
2629         list_for_each_entry(al, &oh->ot_trunc_locks, tl_list) {
2630                 if (obj != al->tl_obj)
2631                         continue;
2632                 LASSERT(al->tl_shared == 0);
2633                 found = 1;
2634                 /* do actual truncate in osd_trans_stop() */
2635                 al->tl_truncate = 1;
2636                 break;
2637         }
2638         LASSERT(found);
2639
2640 out:
2641         RETURN(rc);
2642 }
2643
2644 static int fiemap_check_ranges(struct inode *inode,
2645                                u64 start, u64 len, u64 *new_len)
2646 {
2647         loff_t maxbytes;
2648
2649         *new_len = len;
2650
2651         if (len == 0)
2652                 return -EINVAL;
2653
2654         if (ldiskfs_test_inode_flag(inode, LDISKFS_INODE_EXTENTS))
2655                 maxbytes = inode->i_sb->s_maxbytes;
2656         else
2657                 maxbytes = LDISKFS_SB(inode->i_sb)->s_bitmap_maxbytes;
2658
2659         if (start > maxbytes)
2660                 return -EFBIG;
2661
2662         /*
2663          * Shrink request scope to what the fs can actually handle.
2664          */
2665         if (len > maxbytes || (maxbytes - len) < start)
2666                 *new_len = maxbytes - start;
2667
2668         return 0;
2669 }
2670
2671 /* So that the fiemap access checks can't overflow on 32 bit machines. */
2672 #define FIEMAP_MAX_EXTENTS     (UINT_MAX / sizeof(struct fiemap_extent))
2673
2674 static int osd_fiemap_get(const struct lu_env *env, struct dt_object *dt,
2675                           struct fiemap *fm)
2676 {
2677         struct fiemap_extent_info fieinfo = {0, };
2678         struct inode *inode = osd_dt_obj(dt)->oo_inode;
2679         u64 len;
2680         int rc;
2681
2682         LASSERT(inode);
2683         if (inode->i_op->fiemap == NULL)
2684                 return -EOPNOTSUPP;
2685
2686         if (fm->fm_extent_count > FIEMAP_MAX_EXTENTS)
2687                 return -EINVAL;
2688
2689         rc = fiemap_check_ranges(inode, fm->fm_start, fm->fm_length, &len);
2690         if (rc)
2691                 return rc;
2692
2693         fieinfo.fi_flags = fm->fm_flags;
2694         fieinfo.fi_extents_max = fm->fm_extent_count;
2695         fieinfo.fi_extents_start = fm->fm_extents;
2696
2697         if (fieinfo.fi_flags & FIEMAP_FLAG_SYNC)
2698                 filemap_write_and_wait(inode->i_mapping);
2699
2700         rc = inode->i_op->fiemap(inode, &fieinfo, fm->fm_start, len);
2701         fm->fm_flags = fieinfo.fi_flags;
2702         fm->fm_mapped_extents = fieinfo.fi_extents_mapped;
2703
2704         return rc;
2705 }
2706
2707 static int osd_ladvise(const struct lu_env *env, struct dt_object *dt,
2708                        __u64 start, __u64 end, enum lu_ladvise_type advice)
2709 {
2710         struct osd_object *obj = osd_dt_obj(dt);
2711         int rc = 0;
2712         ENTRY;
2713
2714         switch (advice) {
2715         case LU_LADVISE_DONTNEED:
2716                 if (end)
2717                         invalidate_mapping_pages(obj->oo_inode->i_mapping,
2718                                                  start >> PAGE_SHIFT,
2719                                                  (end - 1) >> PAGE_SHIFT);
2720                 break;
2721         default:
2722                 rc = -ENOTSUPP;
2723                 break;
2724         }
2725
2726         RETURN(rc);
2727 }
2728
2729 static loff_t osd_lseek(const struct lu_env *env, struct dt_object *dt,
2730                         loff_t offset, int whence)
2731 {
2732         struct osd_object *obj = osd_dt_obj(dt);
2733         struct osd_device *dev = osd_obj2dev(obj);
2734         struct inode *inode = obj->oo_inode;
2735         struct file *file;
2736         loff_t result;
2737
2738         ENTRY;
2739         LASSERT(dt_object_exists(dt));
2740         LASSERT(osd_invariant(obj));
2741         LASSERT(inode);
2742         LASSERT(offset >= 0);
2743
2744         file = alloc_file_pseudo(inode, dev->od_mnt, "/", O_NOATIME,
2745                                  inode->i_fop);
2746         if (IS_ERR(file))
2747                 RETURN(PTR_ERR(file));
2748
2749         file->f_mode |= FMODE_64BITHASH;
2750         result = file->f_op->llseek(file, offset, whence);
2751         ihold(inode);
2752         fput(file);
2753         /*
2754          * If 'offset' is beyond end of object file then treat it as not error
2755          * but valid case for SEEK_HOLE and return 'offset' as result.
2756          * LOV will decide if it is beyond real end of file or not.
2757          */
2758         if (whence == SEEK_HOLE && result == -ENXIO)
2759                 result = offset;
2760
2761         CDEBUG(D_INFO, "seek %s from %lld: %lld\n", whence == SEEK_HOLE ?
2762                        "hole" : "data", offset, result);
2763         RETURN(result);
2764 }
2765
2766 /*
2767  * in some cases we may need declare methods for objects being created
2768  * e.g., when we create symlink
2769  */
2770 const struct dt_body_operations osd_body_ops_new = {
2771         .dbo_declare_write = osd_declare_write,
2772 };
2773
2774 const struct dt_body_operations osd_body_ops = {
2775         .dbo_read                       = osd_read,
2776         .dbo_declare_write              = osd_declare_write,
2777         .dbo_write                      = osd_write,
2778         .dbo_bufs_get                   = osd_bufs_get,
2779         .dbo_bufs_put                   = osd_bufs_put,
2780         .dbo_write_prep                 = osd_write_prep,
2781         .dbo_declare_write_commit       = osd_declare_write_commit,
2782         .dbo_write_commit               = osd_write_commit,
2783         .dbo_read_prep                  = osd_read_prep,
2784         .dbo_declare_punch              = osd_declare_punch,
2785         .dbo_punch                      = osd_punch,
2786         .dbo_fiemap_get                 = osd_fiemap_get,
2787         .dbo_ladvise                    = osd_ladvise,
2788         .dbo_declare_fallocate          = osd_declare_fallocate,
2789         .dbo_fallocate                  = osd_fallocate,
2790         .dbo_lseek                      = osd_lseek,
2791 };
2792
2793 /**
2794  * Get a truncate lock
2795  *
2796  * In order to take multi-transaction truncate out of main transaction we let
2797  * the caller grab a lock on the object passed. the lock can be shared (for
2798  * writes) and exclusive (for truncate). It's not allowed to mix truncate
2799  * and write in the same transaction handle (do not confuse with big ldiskfs
2800  * transaction containing lots of handles).
2801  * The lock must be taken at declaration.
2802  *
2803  * \param obj           object to lock
2804  * \oh                  transaction
2805  * \shared              shared or exclusive
2806  *
2807  * \retval 0            lock is granted
2808  * \retval -NOMEM       no memory to allocate lock
2809  */
2810 int osd_trunc_lock(struct osd_object *obj, struct osd_thandle *oh, bool shared)
2811 {
2812         struct osd_access_lock *al, *tmp;
2813
2814         LASSERT(obj);
2815         LASSERT(oh);
2816
2817         list_for_each_entry(tmp, &oh->ot_trunc_locks, tl_list) {
2818                 if (tmp->tl_obj != obj)
2819                         continue;
2820                 LASSERT(tmp->tl_shared == shared);
2821                 /* found same lock */
2822                 return 0;
2823         }
2824
2825         OBD_ALLOC_PTR(al);
2826         if (unlikely(al == NULL))
2827                 return -ENOMEM;
2828         al->tl_obj = obj;
2829         al->tl_truncate = false;
2830         if (shared)
2831                 down_read(&obj->oo_ext_idx_sem);
2832         else
2833                 down_write(&obj->oo_ext_idx_sem);
2834         al->tl_shared = shared;
2835         lu_object_get(&obj->oo_dt.do_lu);
2836
2837         list_add(&al->tl_list, &oh->ot_trunc_locks);
2838
2839         return 0;
2840 }
2841
2842 void osd_trunc_unlock_all(const struct lu_env *env, struct list_head *list)
2843 {
2844         struct osd_access_lock *al, *tmp;
2845
2846         list_for_each_entry_safe(al, tmp, list, tl_list) {
2847                 if (al->tl_shared)
2848                         up_read(&al->tl_obj->oo_ext_idx_sem);
2849                 else
2850                         up_write(&al->tl_obj->oo_ext_idx_sem);
2851                 osd_object_put(env, al->tl_obj);
2852                 list_del(&al->tl_list);
2853                 OBD_FREE_PTR(al);
2854         }
2855 }
2856
2857 /* For a partial-page punch, flush punch range to disk immediately */
2858 static void osd_partial_page_flush_punch(struct osd_device *d,
2859                                          struct inode *inode, loff_t start,
2860                                          loff_t end)
2861 {
2862         if (osd_use_page_cache(d)) {
2863                 filemap_fdatawrite_range(inode->i_mapping, start, end);
2864         } else {
2865                 /* Notice we use "wait" version to ensure I/O is complete */
2866                 filemap_write_and_wait_range(inode->i_mapping, start,
2867                                              end);
2868                 invalidate_mapping_pages(inode->i_mapping, start >> PAGE_SHIFT,
2869                                          end >> PAGE_SHIFT);
2870         }
2871 }
2872
2873 /*
2874  * For a partial-page truncate, flush the page to disk immediately to
2875  * avoid data corruption during direct disk write.  b=17397
2876  */
2877 static void osd_partial_page_flush(struct osd_device *d, struct inode *inode,
2878                                    loff_t offset)
2879 {
2880         if (!(offset & ~PAGE_MASK))
2881                 return;
2882
2883         if (osd_use_page_cache(d)) {
2884                 filemap_fdatawrite_range(inode->i_mapping, offset, offset + 1);
2885         } else {
2886                 /* Notice we use "wait" version to ensure I/O is complete */
2887                 filemap_write_and_wait_range(inode->i_mapping, offset,
2888                                              offset + 1);
2889                 invalidate_mapping_pages(inode->i_mapping, offset >> PAGE_SHIFT,
2890                                          offset >> PAGE_SHIFT);
2891         }
2892 }
2893
2894 void osd_execute_truncate(struct osd_object *obj)
2895 {
2896         struct osd_device *d = osd_obj2dev(obj);
2897         struct inode *inode = obj->oo_inode;
2898         __u64 size;
2899
2900         /* simulate crash before (in the middle) of delayed truncate */
2901         if (OBD_FAIL_CHECK(OBD_FAIL_OSD_FAIL_AT_TRUNCATE)) {
2902                 struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
2903                 struct ldiskfs_sb_info *sbi = LDISKFS_SB(inode->i_sb);
2904
2905                 mutex_lock(&sbi->s_orphan_lock);
2906                 list_del_init(&ei->i_orphan);
2907                 mutex_unlock(&sbi->s_orphan_lock);
2908                 return;
2909         }
2910
2911         size = i_size_read(inode);
2912         inode_lock(inode);
2913         /* if object holds encrypted content, we need to make sure we truncate
2914          * on an encryption unit boundary, or block content will get corrupted
2915          */
2916         if (obj->oo_lma_flags & LUSTRE_ENCRYPT_FL &&
2917             size & ~LUSTRE_ENCRYPTION_MASK)
2918                 inode->i_size = (size & LUSTRE_ENCRYPTION_MASK) +
2919                         LUSTRE_ENCRYPTION_UNIT_SIZE;
2920         ldiskfs_truncate(inode);
2921         inode_unlock(inode);
2922         if (inode->i_size != size) {
2923                 spin_lock(&inode->i_lock);
2924                 i_size_write(inode, size);
2925                 LDISKFS_I(inode)->i_disksize = size;
2926                 spin_unlock(&inode->i_lock);
2927                 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
2928         }
2929         osd_partial_page_flush(d, inode, size);
2930 }
2931
2932 static int osd_execute_punch(const struct lu_env *env, struct osd_object *obj,
2933                              loff_t start, loff_t end, int mode)
2934 {
2935         struct osd_device *d = osd_obj2dev(obj);
2936         struct inode *inode = obj->oo_inode;
2937         struct file *file;
2938         int rc;
2939
2940         file = alloc_file_pseudo(inode, d->od_mnt, "/", O_NOATIME,
2941                                  inode->i_fop);
2942         if (IS_ERR(file))
2943                 RETURN(PTR_ERR(file));
2944
2945         file->f_mode |= FMODE_64BITHASH;
2946         rc = file->f_op->fallocate(file, mode, start, end - start);
2947         ihold(inode);
2948         fput(file);
2949         if (rc == 0)
2950                 osd_partial_page_flush_punch(d, inode, start, end - 1);
2951         return rc;
2952 }
2953
2954 int osd_process_truncates(const struct lu_env *env, struct list_head *list)
2955 {
2956         struct osd_access_lock *al;
2957         int rc = 0;
2958
2959         LASSERT(!journal_current_handle());
2960
2961         list_for_each_entry(al, list, tl_list) {
2962                 if (al->tl_shared)
2963                         continue;
2964                 if (al->tl_truncate)
2965                         osd_execute_truncate(al->tl_obj);
2966                 else if (al->tl_punch)
2967                         rc = osd_execute_punch(env, al->tl_obj, al->tl_start,
2968                                                al->tl_end, al->tl_mode);
2969         }
2970
2971         return rc;
2972 }