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                 iobuf->dr_elapsed_valid = 0;
 151                 LASSERT(iobuf->dr_dev == d);
 152                 LASSERT(iobuf->dr_frags > 0);
 153                 lprocfs_oh_tally(&d->od_brw_stats.hist[BRW_R_DIO_FRAGS+rw],
 154                                  iobuf->dr_frags);
 155                 lprocfs_oh_tally_log2(&d->od_brw_stats.hist[BRW_R_IO_TIME+rw],
 156                                       ktime_to_ms(iobuf->dr_elapsed));
 157         }
 158 }
 159
 160 #ifdef HAVE_BIO_ENDIO_USES_ONE_ARG
 161 static void dio_complete_routine(struct bio *bio)
 162 {
 163         int error = blk_status_to_errno(bio->bi_status);
 164 #else
 165 static void dio_complete_routine(struct bio *bio, int error)
 166 {
 167 #endif
 168         struct osd_iobuf *iobuf = bio->bi_private;
 169         struct bio_vec *bvl;
 170
 171         /* CAVEAT EMPTOR: possibly in IRQ context
 172          * DO NOT record procfs stats here!!!
 173          */
 174
 175         if (unlikely(iobuf == NULL)) {
 176                 CERROR("***** bio->bi_private is NULL!  This should never happen.  Normally, I would crash here, but instead I will dump the bio contents to the console.  Please report this to <https://jira.whamcloud.com/> , along with any interesting messages leading up to this point (like SCSI errors, perhaps).  Because bi_private is NULL, I can't wake up the thread that initiated this IO - you will probably have to reboot this node.\n");
 177                 CERROR("bi_next: %p, bi_flags: %lx, " __stringify(bi_opf)
 178                        ": %x, bi_vcnt: %d, bi_idx: %d, bi->size: %d, bi_end_io: %p, bi_cnt: %d, bi_private: %p\n",
 179                        bio->bi_next, (unsigned long)bio->bi_flags,
 180                        (unsigned int)bio->bi_opf, bio->bi_vcnt, bio_idx(bio),
 181                        bio_sectors(bio) << 9, bio->bi_end_io,
 182                        atomic_read(&bio->__bi_cnt),
 183                        bio->bi_private);
 184                 return;
 185         }
 186
 187         /* the check is outside of the cycle for performance reason -bzzz */
 188         if (!bio_data_dir(bio)) {
 189                 DECLARE_BVEC_ITER_ALL(iter_all);
 190
 191                 bio_for_each_segment_all(bvl, bio, iter_all) {
 192                         if (likely(error == 0))
 193                                 SetPageUptodate(bvl_to_page(bvl));
 194                         LASSERT(PageLocked(bvl_to_page(bvl)));
 195                 }
 196                 atomic_dec(&iobuf->dr_dev->od_r_in_flight);
 197         } else {
 198                 atomic_dec(&iobuf->dr_dev->od_w_in_flight);
 199         }
 200
 201         /* any real error is good enough -bzzz */
 202         if (error != 0 && iobuf->dr_error == 0)
 203                 iobuf->dr_error = error;
 204
 205         /*
 206          * set dr_elapsed before dr_numreqs turns to 0, otherwise
 207          * it's possible that service thread will see dr_numreqs
 208          * is zero, but dr_elapsed is not set yet, leading to lost
 209          * data in this processing and an assertion in a subsequent
 210          * call to OSD.
 211          */
 212         if (atomic_read(&iobuf->dr_numreqs) == 1) {
 213                 ktime_t now = ktime_get();
 214
 215                 iobuf->dr_elapsed = ktime_sub(now, iobuf->dr_start_time);
 216                 iobuf->dr_elapsed_valid = 1;
 217         }
 218         if (atomic_dec_and_test(&iobuf->dr_numreqs))
 219                 wake_up(&iobuf->dr_wait);
 220
 221         /* Completed bios used to be chained off iobuf->dr_bios and freed in
 222          * filter_clear_dreq().  It was then possible to exhaust the biovec-256
 223          * mempool when serious on-disk fragmentation was encountered,
 224          * deadlocking the OST.  The bios are now released as soon as complete
 225          * so the pool cannot be exhausted while IOs are competing. b=10076
 226          */
 227         bio_put(bio);
 228 }
 229
 230 static void record_start_io(struct osd_iobuf *iobuf, int size)
 231 {
 232         struct osd_device    *osd = iobuf->dr_dev;
 233         struct obd_histogram *h = osd->od_brw_stats.hist;
 234
 235         iobuf->dr_frags++;
 236         atomic_inc(&iobuf->dr_numreqs);
 237
 238         if (iobuf->dr_rw == 0) {
 239                 atomic_inc(&osd->od_r_in_flight);
 240                 lprocfs_oh_tally(&h[BRW_R_RPC_HIST],
 241                                  atomic_read(&osd->od_r_in_flight));
 242                 lprocfs_oh_tally_log2(&h[BRW_R_DISK_IOSIZE], size);
 243         } else if (iobuf->dr_rw == 1) {
 244                 atomic_inc(&osd->od_w_in_flight);
 245                 lprocfs_oh_tally(&h[BRW_W_RPC_HIST],
 246                                  atomic_read(&osd->od_w_in_flight));
 247                 lprocfs_oh_tally_log2(&h[BRW_W_DISK_IOSIZE], size);
 248         } else {
 249                 LBUG();
 250         }
 251 }
 252
 253 static void osd_submit_bio(int rw, struct bio *bio)
 254 {
 255         LASSERTF(rw == 0 || rw == 1, "%x\n", rw);
 256 #ifdef HAVE_SUBMIT_BIO_2ARGS
 257         submit_bio(rw ? WRITE : READ, bio);
 258 #else
 259         bio->bi_opf |= rw;
 260         submit_bio(bio);
 261 #endif
 262 }
 263
 264 static int can_be_merged(struct bio *bio, sector_t sector)
 265 {
 266         if (bio == NULL)
 267                 return 0;
 268
 269         return bio_end_sector(bio) == sector ? 1 : 0;
 270 }
 271
 272 #if IS_ENABLED(CONFIG_BLK_DEV_INTEGRITY)
 273 /*
 274  * This function will change the data written, thus it should only be
 275  * used when checking data integrity feature
 276  */
 277 static void bio_integrity_fault_inject(struct bio *bio)
 278 {
 279         struct bio_vec *bvec;
 280         DECLARE_BVEC_ITER_ALL(iter_all);
 281         void *kaddr;
 282         char *addr;
 283
 284         bio_for_each_segment_all(bvec, bio, iter_all) {
 285                 struct page *page = bvec->bv_page;
 286
 287                 kaddr = kmap(page);
 288                 addr = kaddr;
 289                 *addr = ~(*addr);
 290                 kunmap(page);
 291                 break;
 292         }
 293 }
 294
 295 static int bio_dif_compare(__u16 *expected_guard_buf, void *bio_prot_buf,
 296                            unsigned int sectors, int tuple_size)
 297 {
 298         __u16 *expected_guard;
 299         __u16 *bio_guard;
 300         int i;
 301
 302         expected_guard = expected_guard_buf;
 303         for (i = 0; i < sectors; i++) {
 304                 bio_guard = (__u16 *)bio_prot_buf;
 305                 if (*bio_guard != *expected_guard) {
 306                         CERROR(
 307                                "unexpected guard tags on sector %d expected guard %u, bio guard %u, sectors %u, tuple size %d\n",
 308                                i, *expected_guard, *bio_guard, sectors,
 309                                tuple_size);
 310                         return -EIO;
 311                 }
 312                 expected_guard++;
 313                 bio_prot_buf += tuple_size;
 314         }
 315         return 0;
 316 }
 317
 318 static int osd_bio_integrity_compare(struct bio *bio, struct block_device *bdev,
 319                                      struct osd_iobuf *iobuf, int index)
 320 {
 321         struct blk_integrity *bi = bdev_get_integrity(bdev);
 322         struct bio_integrity_payload *bip = bio->bi_integrity;
 323         struct niobuf_local *lnb = NULL;
 324         unsigned short sector_size = blk_integrity_interval(bi);
 325         void *bio_prot_buf = page_address(bip->bip_vec->bv_page) +
 326                 bip->bip_vec->bv_offset;
 327         struct bio_vec *bv;
 328         sector_t sector = bio_start_sector(bio);
 329         unsigned int i, sectors, total;
 330         DECLARE_BVEC_ITER_ALL(iter_all);
 331         __u16 *expected_guard;
 332         int rc;
 333
 334         total = 0;
 335         bio_for_each_segment_all(bv, bio, iter_all) {
 336                 for (i = index; i < iobuf->dr_npages; i++) {
 337                         if (iobuf->dr_pages[i] == bv->bv_page) {
 338                                 lnb = iobuf->dr_lnbs[i];
 339                                 break;
 340                         }
 341                 }
 342                 if (!lnb)
 343                         continue;
 344                 expected_guard = lnb->lnb_guards;
 345                 sectors = bv->bv_len / sector_size;
 346                 if (lnb->lnb_guard_rpc) {
 347                         rc = bio_dif_compare(expected_guard, bio_prot_buf,
 348                                              sectors, bi->tuple_size);
 349                         if (rc)
 350                                 return rc;
 351                 }
 352
 353                 sector += sectors;
 354                 bio_prot_buf += sectors * bi->tuple_size;
 355                 total += sectors * bi->tuple_size;
 356                 LASSERT(total <= bip_size(bio->bi_integrity));
 357                 index++;
 358                 lnb = NULL;
 359         }
 360         return 0;
 361 }
 362
 363 static int osd_bio_integrity_handle(struct osd_device *osd, struct bio *bio,
 364                                     struct osd_iobuf *iobuf,
 365                                     int start_page_idx, bool fault_inject,
 366                                     bool integrity_enabled)
 367 {
 368         struct super_block *sb = osd_sb(osd);
 369         integrity_gen_fn *generate_fn = NULL;
 370         integrity_vrfy_fn *verify_fn = NULL;
 371         int rc;
 372
 373         ENTRY;
 374
 375         if (!integrity_enabled)
 376                 RETURN(0);
 377
 378         rc = osd_get_integrity_profile(osd, &generate_fn, &verify_fn);
 379         if (rc)
 380                 RETURN(rc);
 381
 382         rc = bio_integrity_prep_fn(bio, generate_fn, verify_fn);
 383         if (rc)
 384                 RETURN(rc);
 385
 386         /* Verify and inject fault only when writing */
 387         if (iobuf->dr_rw == 1) {
 388                 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_OST_INTEGRITY_CMP))) {
 389                         rc = osd_bio_integrity_compare(bio, sb->s_bdev, iobuf,
 390                                                        start_page_idx);
 391                         if (rc)
 392                                 RETURN(rc);
 393                 }
 394
 395                 if (unlikely(fault_inject))
 396                         bio_integrity_fault_inject(bio);
 397         }
 398
 399         RETURN(0);
 400 }
 401
 402 #ifdef HAVE_BIO_INTEGRITY_PREP_FN
 403 #  ifdef HAVE_BIO_ENDIO_USES_ONE_ARG
 404 static void dio_integrity_complete_routine(struct bio *bio)
 405 #  else
 406 static void dio_integrity_complete_routine(struct bio *bio, int error)
 407 #  endif
 408 {
 409         struct osd_bio_private *bio_private = bio->bi_private;
 410
 411         bio->bi_private = bio_private->obp_iobuf;
 412         osd_dio_complete_routine(bio, error);
 413
 414         OBD_FREE_PTR(bio_private);
 415 }
 416 #endif /* HAVE_BIO_INTEGRITY_PREP_FN */
 417 #else  /* !CONFIG_BLK_DEV_INTEGRITY */
 418 #define osd_bio_integrity_handle(osd, bio, iobuf, start_page_idx, \
 419                                  fault_inject, integrity_enabled) 0
 420 #endif /* CONFIG_BLK_DEV_INTEGRITY */
 421
 422 static int osd_bio_init(struct bio *bio, struct osd_iobuf *iobuf,
 423                         bool integrity_enabled, int start_page_idx,
 424                         struct osd_bio_private **pprivate)
 425 {
 426         ENTRY;
 427
 428         *pprivate = NULL;
 429
 430 #ifdef HAVE_BIO_INTEGRITY_PREP_FN
 431         if (integrity_enabled) {
 432                 struct osd_bio_private *bio_private = NULL;
 433
 434                 OBD_ALLOC_GFP(bio_private, sizeof(*bio_private), GFP_NOIO);
 435                 if (bio_private == NULL)
 436                         RETURN(-ENOMEM);
 437                 bio->bi_end_io = dio_integrity_complete_routine;
 438                 bio->bi_private = bio_private;
 439                 bio_private->obp_start_page_idx = start_page_idx;
 440                 bio_private->obp_iobuf = iobuf;
 441                 *pprivate = bio_private;
 442         } else
 443 #endif
 444         {
 445                 bio->bi_end_io = dio_complete_routine;
 446                 bio->bi_private = iobuf;
 447         }
 448
 449         RETURN(0);
 450 }
 451
 452 static void osd_mark_page_io_done(struct osd_iobuf *iobuf,
 453                                   struct inode *inode,
 454                                   sector_t start_blocks,
 455                                   sector_t count)
 456 {
 457         struct niobuf_local *lnb;
 458         int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
 459         pgoff_t pg_start, pg_end;
 460
 461         pg_start = start_blocks / blocks_per_page;
 462         if (start_blocks % blocks_per_page)
 463                 pg_start++;
 464         if (count >= blocks_per_page)
 465                 pg_end = (start_blocks + count -
 466                           blocks_per_page) / blocks_per_page;
 467         else
 468                 return; /* nothing to mark */
 469         for ( ; pg_start <= pg_end; pg_start++) {
 470                 lnb = iobuf->dr_lnbs[pg_start];
 471                 lnb->lnb_flags |= OBD_BRW_DONE;
 472         }
 473 }
 474
 475 static int osd_do_bio(struct osd_device *osd, struct inode *inode,
 476                       struct osd_iobuf *iobuf, sector_t start_blocks,
 477                       sector_t count)
 478 {
 479         int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
 480         struct page **pages = iobuf->dr_pages;
 481         int npages = iobuf->dr_npages;
 482         sector_t *blocks = iobuf->dr_blocks;
 483         struct super_block *sb = inode->i_sb;
 484         int sector_bits = sb->s_blocksize_bits - 9;
 485         unsigned int blocksize = sb->s_blocksize;
 486         struct block_device *bdev = sb->s_bdev;
 487         struct osd_bio_private *bio_private = NULL;
 488         struct bio *bio = NULL;
 489         int bio_start_page_idx;
 490         struct page *page;
 491         unsigned int page_offset;
 492         sector_t sector;
 493         int nblocks;
 494         int block_idx, block_idx_end;
 495         int page_idx, page_idx_start;
 496         int i;
 497         int rc = 0;
 498         bool fault_inject;
 499         bool integrity_enabled;
 500         struct blk_plug plug;
 501         int blocks_left_page;
 502
 503         ENTRY;
 504
 505         fault_inject = OBD_FAIL_CHECK(OBD_FAIL_OST_INTEGRITY_FAULT);
 506         LASSERT(iobuf->dr_npages == npages);
 507
 508         integrity_enabled = bdev_integrity_enabled(bdev, iobuf->dr_rw);
 509
 510         osd_brw_stats_update(osd, iobuf);
 511         iobuf->dr_start_time = ktime_get();
 512
 513         if (!count)
 514                 count = npages * blocks_per_page;
 515         block_idx_end = start_blocks + count;
 516
 517         blk_start_plug(&plug);
 518
 519         page_idx_start = start_blocks / blocks_per_page;
 520         for (page_idx = page_idx_start, block_idx = start_blocks;
 521              block_idx < block_idx_end; page_idx++,
 522              block_idx += blocks_left_page) {
 523                 page = pages[page_idx];
 524                 LASSERT(page_idx < iobuf->dr_npages);
 525
 526                 i = block_idx % blocks_per_page;
 527                 blocks_left_page = blocks_per_page - i;
 528                 for (page_offset = i * blocksize; i < blocks_left_page;
 529                      i += nblocks, page_offset += blocksize * nblocks) {
 530                         nblocks = 1;
 531
 532                         if (blocks[block_idx + i] == 0) {  /* hole */
 533                                 LASSERTF(iobuf->dr_rw == 0,
 534                                          "page_idx %u, block_idx %u, i %u,"
 535                                          "start_blocks: %llu, count: %llu, npages: %d\n",
 536                                          page_idx, block_idx, i,
 537                                          (unsigned long long)start_blocks,
 538                                          (unsigned long long)count, npages);
 539                                 memset(kmap(page) + page_offset, 0, blocksize);
 540                                 kunmap(page);
 541                                 continue;
 542                         }
 543
 544                         sector = (sector_t)blocks[block_idx + i] << sector_bits;
 545
 546                         /* Additional contiguous file blocks? */
 547                         while (i + nblocks < blocks_left_page &&
 548                                (sector + (nblocks << sector_bits)) ==
 549                                ((sector_t)blocks[block_idx + i + nblocks] <<
 550                                  sector_bits))
 551                                 nblocks++;
 552
 553                         if (bio && can_be_merged(bio, sector) &&
 554                             bio_add_page(bio, page, blocksize * nblocks,
 555                                          page_offset) != 0)
 556                                 continue;       /* added this frag OK */
 557
 558                         if (bio != NULL) {
 559                                 struct request_queue *q = bio_get_queue(bio);
 560                                 unsigned int bi_size = bio_sectors(bio) << 9;
 561
 562                                 /* Dang! I have to fragment this I/O */
 563                                 CDEBUG(D_INODE,
 564                                        "bio++ sz %d vcnt %d(%d) sectors %d(%d) psg %d(%d)\n",
 565                                        bi_size, bio->bi_vcnt, bio->bi_max_vecs,
 566                                        bio_sectors(bio),
 567                                        queue_max_sectors(q),
 568                                        osd_bio_nr_segs(bio),
 569                                        queue_max_segments(q));
 570                                 rc = osd_bio_integrity_handle(osd, bio,
 571                                         iobuf, bio_start_page_idx,
 572                                         fault_inject, integrity_enabled);
 573                                 if (rc) {
 574                                         bio_put(bio);
 575                                         goto out;
 576                                 }
 577
 578                                 record_start_io(iobuf, bi_size);
 579                                 osd_submit_bio(iobuf->dr_rw, bio);
 580                         }
 581
 582                         bio_start_page_idx = page_idx;
 583                         /* allocate new bio */
 584                         bio = bio_alloc(GFP_NOIO, min(BIO_MAX_PAGES,
 585                                         (block_idx_end - block_idx +
 586                                          blocks_left_page - 1)));
 587                         if (bio == NULL) {
 588                                 CERROR("Can't allocate bio %u pages\n",
 589                                        block_idx_end - block_idx +
 590                                        blocks_left_page - 1);
 591                                 rc = -ENOMEM;
 592                                 goto out;
 593                         }
 594
 595                         bio_set_dev(bio, bdev);
 596                         bio_set_sector(bio, sector);
 597                         bio->bi_opf = iobuf->dr_rw ? WRITE : READ;
 598                         rc = osd_bio_init(bio, iobuf, integrity_enabled,
 599                                           bio_start_page_idx, &bio_private);
 600                         if (rc) {
 601                                 bio_put(bio);
 602                                 goto out;
 603                         }
 604
 605                         rc = bio_add_page(bio, page,
 606                                           blocksize * nblocks, page_offset);
 607                         LASSERT(rc != 0);
 608                 }
 609         }
 610
 611         if (bio != NULL) {
 612                 rc = osd_bio_integrity_handle(osd, bio, iobuf,
 613                                               bio_start_page_idx,
 614                                               fault_inject,
 615                                               integrity_enabled);
 616                 if (rc) {
 617                         bio_put(bio);
 618                         goto out;
 619                 }
 620
 621                 record_start_io(iobuf, bio_sectors(bio) << 9);
 622                 osd_submit_bio(iobuf->dr_rw, bio);
 623                 rc = 0;
 624         }
 625
 626 out:
 627         blk_finish_plug(&plug);
 628
 629         /* in order to achieve better IO throughput, we don't wait for writes
 630          * completion here. instead we proceed with transaction commit in
 631          * parallel and wait for IO completion once transaction is stopped
 632          * see osd_trans_stop() for more details -bzzz
 633          */
 634         if (iobuf->dr_rw == 0 || fault_inject) {
 635                 wait_event(iobuf->dr_wait,
 636                            atomic_read(&iobuf->dr_numreqs) == 0);
 637                 osd_fini_iobuf(osd, iobuf);
 638         }
 639
 640         if (rc == 0) {
 641                 rc = iobuf->dr_error;
 642         } else {
 643                 if (bio_private)
 644                         OBD_FREE_PTR(bio_private);
 645         }
 646
 647         /* Write only now */
 648         if (rc == 0 && iobuf->dr_rw)
 649                 osd_mark_page_io_done(iobuf, inode,
 650                                       start_blocks, count);
 651
 652         RETURN(rc);
 653 }
 654
 655 static int osd_map_remote_to_local(loff_t offset, ssize_t len, int *nrpages,
 656                                    struct niobuf_local *lnb, int maxlnb)
 657 {
 658         int rc = 0;
 659         ENTRY;
 660
 661         *nrpages = 0;
 662
 663         while (len > 0) {
 664                 int poff = offset & (PAGE_SIZE - 1);
 665                 int plen = PAGE_SIZE - poff;
 666
 667                 if (*nrpages >= maxlnb) {
 668                         rc = -EOVERFLOW;
 669                         break;
 670                 }
 671
 672                 if (plen > len)
 673                         plen = len;
 674                 lnb->lnb_file_offset = offset;
 675                 lnb->lnb_page_offset = poff;
 676                 lnb->lnb_len = plen;
 677                 /* lnb->lnb_flags = rnb->rnb_flags; */
 678                 lnb->lnb_flags = 0;
 679                 lnb->lnb_page = NULL;
 680                 lnb->lnb_rc = 0;
 681                 lnb->lnb_guard_rpc = 0;
 682                 lnb->lnb_guard_disk = 0;
 683                 lnb->lnb_locked = 0;
 684
 685                 LASSERTF(plen <= len, "plen %u, len %lld\n", plen,
 686                          (long long) len);
 687                 offset += plen;
 688                 len -= plen;
 689                 lnb++;
 690                 (*nrpages)++;
 691         }
 692
 693         RETURN(rc);
 694 }
 695
 696 static struct page *osd_get_page(const struct lu_env *env, struct dt_object *dt,
 697                                  loff_t offset, gfp_t gfp_mask, bool cache)
 698 {
 699         struct osd_thread_info *oti = osd_oti_get(env);
 700         struct inode *inode = osd_dt_obj(dt)->oo_inode;
 701         struct osd_device *d = osd_obj2dev(osd_dt_obj(dt));
 702         struct page *page;
 703         int cur;
 704
 705         LASSERT(inode);
 706
 707         if (cache) {
 708                 page = find_or_create_page(inode->i_mapping,
 709                                            offset >> PAGE_SHIFT, gfp_mask);
 710
 711                 if (likely(page)) {
 712                         LASSERT(!PagePrivate2(page));
 713                         wait_on_page_writeback(page);
 714                 } else {
 715                         lprocfs_counter_add(d->od_stats, LPROC_OSD_NO_PAGE, 1);
 716                 }
 717
 718                 return page;
 719         }
 720
 721         if (inode->i_mapping->nrpages) {
 722                 /* consult with pagecache, but do not create new pages */
 723                 /* this is normally used once */
 724                 page = find_lock_page(inode->i_mapping, offset >> PAGE_SHIFT);
 725                 if (page) {
 726                         wait_on_page_writeback(page);
 727                         return page;
 728                 }
 729         }
 730
 731         LASSERT(oti->oti_dio_pages);
 732         cur = oti->oti_dio_pages_used;
 733         page = oti->oti_dio_pages[cur];
 734
 735         if (unlikely(!page)) {
 736                 LASSERT(cur < PTLRPC_MAX_BRW_PAGES);
 737                 page = alloc_page(gfp_mask);
 738                 if (!page)
 739                         return NULL;
 740                 oti->oti_dio_pages[cur] = page;
 741                 SetPagePrivate2(page);
 742                 lock_page(page);
 743         }
 744
 745         ClearPageUptodate(page);
 746         page->index = offset >> PAGE_SHIFT;
 747         oti->oti_dio_pages_used++;
 748
 749         return page;
 750 }
 751
 752 /*
 753  * there are following "locks":
 754  * journal_start
 755  * i_mutex
 756  * page lock
 757  *
 758  * osd write path:
 759  *  - lock page(s)
 760  *  - journal_start
 761  *  - truncate_sem
 762  *
 763  * ext4 vmtruncate:
 764  *  - lock pages, unlock
 765  *  - journal_start
 766  *  - lock partial page
 767  *  - i_data_sem
 768  *
 769  */
 770
 771 /**
 772  * Unlock and release pages loaded by osd_bufs_get()
 773  *
 774  * Unlock \a npages pages from \a lnb and drop the refcount on them.
 775  *
 776  * \param env           thread execution environment
 777  * \param dt            dt object undergoing IO (OSD object + methods)
 778  * \param lnb           array of pages undergoing IO
 779  * \param npages        number of pages in \a lnb
 780  *
 781  * \retval 0            always
 782  */
 783 static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt,
 784                         struct niobuf_local *lnb, int npages)
 785 {
 786         struct osd_thread_info *oti = osd_oti_get(env);
 787         struct pagevec pvec;
 788         int i;
 789
 790         ll_pagevec_init(&pvec, 0);
 791
 792         for (i = 0; i < npages; i++) {
 793                 struct page *page = lnb[i].lnb_page;
 794
 795                 if (page == NULL)
 796                         continue;
 797
 798                 /* if the page isn't cached, then reset uptodate
 799                  * to prevent reuse
 800                  */
 801                 if (PagePrivate2(page)) {
 802                         oti->oti_dio_pages_used--;
 803                 } else {
 804                         if (lnb[i].lnb_locked)
 805                                 unlock_page(page);
 806                         if (pagevec_add(&pvec, page) == 0)
 807                                 pagevec_release(&pvec);
 808                 }
 809
 810                 lnb[i].lnb_page = NULL;
 811         }
 812
 813         LASSERTF(oti->oti_dio_pages_used == 0, "%d\n", oti->oti_dio_pages_used);
 814
 815         /* Release any partial pagevec */
 816         pagevec_release(&pvec);
 817
 818         RETURN(0);
 819 }
 820
 821 /**
 822  * Load and lock pages undergoing IO
 823  *
 824  * Pages as described in the \a lnb array are fetched (from disk or cache)
 825  * and locked for IO by the caller.
 826  *
 827  * DLM locking protects us from write and truncate competing for same region,
 828  * but partial-page truncate can leave dirty pages in the cache for ldiskfs.
 829  * It's possible the writeout on a such a page is in progress when we access
 830  * it. It's also possible that during this writeout we put new (partial) data
 831  * into the page, but won't be able to proceed in filter_commitrw_write().
 832  * Therefore, just wait for writeout completion as it should be rare enough.
 833  *
 834  * \param env           thread execution environment
 835  * \param dt            dt object undergoing IO (OSD object + methods)
 836  * \param pos           byte offset of IO start
 837  * \param len           number of bytes of IO
 838  * \param lnb           array of extents undergoing IO
 839  * \param rw            read or write operation, and other flags
 840  * \param capa          capabilities
 841  *
 842  * \retval pages        (zero or more) loaded successfully
 843  * \retval -ENOMEM      on memory/page allocation error
 844  */
 845 static int osd_bufs_get(const struct lu_env *env, struct dt_object *dt,
 846                         loff_t pos, ssize_t len, struct niobuf_local *lnb,
 847                         int maxlnb, enum dt_bufs_type rw)
 848 {
 849         struct osd_thread_info *oti = osd_oti_get(env);
 850         struct osd_object *obj = osd_dt_obj(dt);
 851         struct osd_device *osd   = osd_obj2dev(obj);
 852         int npages, i, iosize, rc = 0;
 853         bool cache, write;
 854         loff_t fsize;
 855         gfp_t gfp_mask;
 856
 857         LASSERT(obj->oo_inode);
 858
 859         rc = osd_map_remote_to_local(pos, len, &npages, lnb, maxlnb);
 860         if (rc)
 861                 RETURN(rc);
 862
 863         write = rw & DT_BUFS_TYPE_WRITE;
 864
 865         fsize = lnb[npages - 1].lnb_file_offset + lnb[npages - 1].lnb_len;
 866         iosize = fsize - lnb[0].lnb_file_offset;
 867         fsize = max(fsize, i_size_read(obj->oo_inode));
 868
 869         cache = rw & DT_BUFS_TYPE_READAHEAD;
 870         if (cache)
 871                 goto bypass_checks;
 872
 873         cache = osd_use_page_cache(osd);
 874         while (cache) {
 875                 if (write) {
 876                         if (!osd->od_writethrough_cache) {
 877                                 cache = false;
 878                                 break;
 879                         }
 880                         if (iosize > osd->od_writethrough_max_iosize) {
 881                                 cache = false;
 882                                 break;
 883                         }
 884                 } else {
 885                         if (!osd->od_read_cache) {
 886                                 cache = false;
 887                                 break;
 888                         }
 889                         if (iosize > osd->od_readcache_max_iosize) {
 890                                 cache = false;
 891                                 break;
 892                         }
 893                 }
 894                 /* don't use cache on large files */
 895                 if (osd->od_readcache_max_filesize &&
 896                     fsize > osd->od_readcache_max_filesize)
 897                         cache = false;
 898                 break;
 899         }
 900
 901 bypass_checks:
 902         if (!cache && unlikely(!oti->oti_dio_pages)) {
 903                 OBD_ALLOC_PTR_ARRAY_LARGE(oti->oti_dio_pages,
 904                                           PTLRPC_MAX_BRW_PAGES);
 905                 if (!oti->oti_dio_pages)
 906                         return -ENOMEM;
 907         }
 908
 909         /* this could also try less hard for DT_BUFS_TYPE_READAHEAD pages */
 910         gfp_mask = rw & DT_BUFS_TYPE_LOCAL ? (GFP_NOFS | __GFP_HIGHMEM) :
 911                                              GFP_HIGHUSER;
 912         for (i = 0; i < npages; i++, lnb++) {
 913                 lnb->lnb_page = osd_get_page(env, dt, lnb->lnb_file_offset,
 914                                              gfp_mask, cache);
 915                 if (lnb->lnb_page == NULL)
 916                         GOTO(cleanup, rc = -ENOMEM);
 917
 918                 lnb->lnb_locked = 1;
 919                 if (cache)
 920                         mark_page_accessed(lnb->lnb_page);
 921         }
 922
 923 #if 0
 924         /* XXX: this version doesn't invalidate cached pages, but use them */
 925         if (!cache && write && obj->oo_inode->i_mapping->nrpages) {
 926                 /* do not allow data aliasing, invalidate pagecache */
 927                 /* XXX: can be quite expensive in mixed case */
 928                 invalidate_mapping_pages(obj->oo_inode->i_mapping,
 929                                 lnb[0].lnb_file_offset >> PAGE_SHIFT,
 930                                 lnb[npages - 1].lnb_file_offset >> PAGE_SHIFT);
 931         }
 932 #endif
 933
 934         RETURN(i);
 935
 936 cleanup:
 937         if (i > 0)
 938                 osd_bufs_put(env, dt, lnb - i, i);
 939         return rc;
 940 }
 941 /* Borrow @ext4_chunk_trans_blocks */
 942 static int osd_chunk_trans_blocks(struct inode *inode, int nrblocks)
 943 {
 944         ldiskfs_group_t groups;
 945         int gdpblocks;
 946         int idxblocks;
 947         int depth;
 948         int ret;
 949
 950         depth = ext_depth(inode);
 951         idxblocks = depth * 2;
 952
 953         /*
 954          * Now let's see how many group bitmaps and group descriptors need
 955          * to account.
 956          */
 957         groups = idxblocks + 1;
 958         gdpblocks = groups;
 959         if (groups > LDISKFS_SB(inode->i_sb)->s_groups_count)
 960                 groups = LDISKFS_SB(inode->i_sb)->s_groups_count;
 961         if (gdpblocks > LDISKFS_SB(inode->i_sb)->s_gdb_count)
 962                 gdpblocks = LDISKFS_SB(inode->i_sb)->s_gdb_count;
 963
 964         /* bitmaps and block group descriptor blocks */
 965         ret = idxblocks + groups + gdpblocks;
 966
 967         /* Blocks for super block, inode, quota and xattr blocks */
 968         ret += LDISKFS_META_TRANS_BLOCKS(inode->i_sb);
 969
 970         return ret;
 971 }
 972
 973 #ifdef HAVE_LDISKFS_JOURNAL_ENSURE_CREDITS
 974 static int osd_extend_restart_trans(handle_t *handle, int needed,
 975                                     struct inode *inode)
 976 {
 977         int rc;
 978
 979         rc = ldiskfs_journal_ensure_credits(handle, needed,
 980                 ldiskfs_trans_default_revoke_credits(inode->i_sb));
 981         /* this means journal has been restarted */
 982         if (rc > 0)
 983                 rc = 0;
 984
 985         return rc;
 986 }
 987 #else
 988 static int osd_extend_restart_trans(handle_t *handle, int needed,
 989                                     struct inode *inode)
 990 {
 991         int rc;
 992
 993         if (ldiskfs_handle_has_enough_credits(handle, needed))
 994                 return 0;
 995         rc = ldiskfs_journal_extend(handle,
 996                                 needed - handle->h_buffer_credits);
 997         if (rc <= 0)
 998                 return rc;
 999
1000         return ldiskfs_journal_restart(handle, needed);
1001 }
1002 #endif /* HAVE_LDISKFS_JOURNAL_ENSURE_CREDITS */
1003
1004 static int osd_ldiskfs_map_write(struct inode *inode, struct osd_iobuf *iobuf,
1005                                  struct osd_device *osd, sector_t start_blocks,
1006                                  sector_t count, loff_t *disk_size,
1007                                  __u64 user_size)
1008 {
1009         /* if file has grown, take user_size into account */
1010         if (user_size && *disk_size > user_size)
1011                 *disk_size = user_size;
1012
1013         spin_lock(&inode->i_lock);
1014         if (*disk_size > i_size_read(inode)) {
1015                 i_size_write(inode, *disk_size);
1016                 LDISKFS_I(inode)->i_disksize = *disk_size;
1017                 spin_unlock(&inode->i_lock);
1018                 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
1019         } else {
1020                 spin_unlock(&inode->i_lock);
1021         }
1022
1023         /*
1024          * We don't do stats here as in read path because
1025          * write is async: we'll do this in osd_put_bufs()
1026          */
1027         return osd_do_bio(osd, inode, iobuf, start_blocks, count);
1028 }
1029
1030 static unsigned int osd_extent_bytes(const struct osd_device *o)
1031 {
1032         unsigned int *extent_bytes_ptr =
1033                         raw_cpu_ptr(o->od_extent_bytes_percpu);
1034
1035         if (likely(*extent_bytes_ptr))
1036                 return *extent_bytes_ptr;
1037
1038         /* initialize on first access or CPU hotplug */
1039         if (!ldiskfs_has_feature_extents(osd_sb(o)))
1040                 *extent_bytes_ptr = 1 << osd_sb(o)->s_blocksize_bits;
1041         else
1042                 *extent_bytes_ptr = OSD_DEFAULT_EXTENT_BYTES;
1043
1044         return *extent_bytes_ptr;
1045 }
1046
1047 #define EXTENT_BYTES_DECAY 64
1048 static void osd_decay_extent_bytes(struct osd_device *osd,
1049                                    unsigned int new_bytes)
1050 {
1051         unsigned int old_bytes;
1052
1053         if (!ldiskfs_has_feature_extents(osd_sb(osd)))
1054                 return;
1055
1056         old_bytes = osd_extent_bytes(osd);
1057         *raw_cpu_ptr(osd->od_extent_bytes_percpu) =
1058                 (old_bytes * (EXTENT_BYTES_DECAY - 1) +
1059                  min(new_bytes, OSD_DEFAULT_EXTENT_BYTES) +
1060                  EXTENT_BYTES_DECAY - 1) / EXTENT_BYTES_DECAY;
1061 }
1062
1063 static int osd_ldiskfs_map_inode_pages(struct inode *inode,
1064                                        struct osd_iobuf *iobuf,
1065                                        struct osd_device *osd,
1066                                        int create, __u64 user_size,
1067                                        int check_credits,
1068                                        struct thandle *thandle)
1069 {
1070         int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
1071         int rc = 0, i = 0, mapped_index = 0;
1072         struct page *fp = NULL;
1073         int clen = 0;
1074         pgoff_t max_page_index;
1075         handle_t *handle = NULL;
1076         sector_t start_blocks = 0, count = 0;
1077         loff_t disk_size = 0;
1078         struct page **page = iobuf->dr_pages;
1079         int pages = iobuf->dr_npages;
1080         sector_t *blocks = iobuf->dr_blocks;
1081         struct niobuf_local *lnb1, *lnb2;
1082         loff_t size1, size2;
1083
1084         max_page_index = inode->i_sb->s_maxbytes >> PAGE_SHIFT;
1085
1086         CDEBUG(D_OTHER, "inode %lu: map %d pages from %lu\n",
1087                 inode->i_ino, pages, (*page)->index);
1088
1089         if (create) {
1090                 create = LDISKFS_GET_BLOCKS_CREATE;
1091                 handle = ldiskfs_journal_current_handle();
1092                 LASSERT(handle != NULL);
1093                 rc = osd_attach_jinode(inode);
1094                 if (rc)
1095                         return rc;
1096                 disk_size = i_size_read(inode);
1097                 /* if disk_size is already bigger than specified user_size,
1098                  * ignore user_size
1099                  */
1100                 if (disk_size > user_size)
1101                         user_size = 0;
1102         }
1103         /* pages are sorted already. so, we just have to find
1104          * contig. space and process them properly
1105          */
1106         while (i < pages) {
1107                 long blen, total = 0, previous_total = 0;
1108                 struct ldiskfs_map_blocks map = { 0 };
1109
1110                 if (fp == NULL) { /* start new extent */
1111                         fp = *page++;
1112                         clen = 1;
1113                         if (++i != pages)
1114                                 continue;
1115                 } else if (fp->index + clen == (*page)->index) {
1116                         /* continue the extent */
1117                         page++;
1118                         clen++;
1119                         if (++i != pages)
1120                                 continue;
1121                 }
1122                 if (fp->index + clen >= max_page_index)
1123                         GOTO(cleanup, rc = -EFBIG);
1124                 /* process found extent */
1125                 map.m_lblk = fp->index * blocks_per_page;
1126                 map.m_len = blen = clen * blocks_per_page;
1127 cont_map:
1128                 /**
1129                  * We might restart transaction for block allocations,
1130                  * in order to make sure data ordered mode, issue IO, disk
1131                  * size update and block allocations need be within same
1132                  * transaction to make sure consistency.
1133                  */
1134                 if (handle && check_credits) {
1135                         struct osd_thandle *oh;
1136
1137                         LASSERT(thandle != NULL);
1138                         oh = container_of(thandle, struct osd_thandle,
1139                                           ot_super);
1140                         /*
1141                          * only issue IO if restart transaction needed,
1142                          * as update disk size need hold inode lock, we
1143                          * want to avoid that as much as possible.
1144                          */
1145                         if (oh->oh_declared_ext <= 0) {
1146                                 rc = osd_ldiskfs_map_write(inode,
1147                                         iobuf, osd, start_blocks,
1148                                         count, &disk_size, user_size);
1149                                 if (rc)
1150                                         GOTO(cleanup, rc);
1151                                 thandle->th_restart_tran = 1;
1152                                 GOTO(cleanup, rc = -EAGAIN);
1153                         }
1154
1155                         if (OBD_FAIL_CHECK(OBD_FAIL_OST_RESTART_IO))
1156                                 oh->oh_declared_ext = 0;
1157                         else
1158                                 oh->oh_declared_ext--;
1159                 }
1160                 rc = ldiskfs_map_blocks(handle, inode, &map, create);
1161                 if (rc >= 0) {
1162                         int c = 0;
1163
1164                         for (; total < blen && c < map.m_len; c++, total++) {
1165                                 if (rc == 0) {
1166                                         *(blocks + total) = 0;
1167                                         total++;
1168                                         break;
1169                                 }
1170                                 if ((map.m_flags & LDISKFS_MAP_UNWRITTEN) &&
1171                                     !create) {
1172                                         /* don't try to read allocated, but
1173                                          * unwritten blocks, instead fill the
1174                                          * patches with zeros in osd_do_bio() */
1175                                         *(blocks + total) = 0;
1176                                         continue;
1177                                 }
1178                                 *(blocks + total) = map.m_pblk + c;
1179                                 /* unmap any possible underlying
1180                                  * metadata from the block device
1181                                  * mapping.  b=6998.
1182                                  */
1183                                 if ((map.m_flags & LDISKFS_MAP_NEW) &&
1184                                     create)
1185                                         clean_bdev_aliases(inode->i_sb->s_bdev,
1186                                                            map.m_pblk + c, 1);
1187                         }
1188                         rc = 0;
1189                 }
1190
1191                 if (rc == 0 && create) {
1192                         count += (total - previous_total);
1193                         mapped_index = (count + blocks_per_page -
1194                                         1) / blocks_per_page - 1;
1195                         lnb1 = iobuf->dr_lnbs[i - clen];
1196                         lnb2 = iobuf->dr_lnbs[mapped_index];
1197                         size1 = lnb1->lnb_file_offset -
1198                                 (lnb1->lnb_file_offset % PAGE_SIZE) +
1199                                 (total << inode->i_blkbits);
1200                         size2 = lnb2->lnb_file_offset + lnb2->lnb_len;
1201
1202                         if (size1 > size2)
1203                                 size1 = size2;
1204                         if (size1 > disk_size)
1205                                 disk_size = size1;
1206                 }
1207
1208                 if (rc == 0 && total < blen) {
1209                         /*
1210                          * decay extent blocks if we could not
1211                          * allocate extent once.
1212                          */
1213                         osd_decay_extent_bytes(osd,
1214                                 (total - previous_total) << inode->i_blkbits);
1215                         map.m_lblk = fp->index * blocks_per_page + total;
1216                         map.m_len = blen - total;
1217                         previous_total = total;
1218                         goto cont_map;
1219                 }
1220                 if (rc != 0)
1221                         GOTO(cleanup, rc);
1222                 /*
1223                  * decay extent blocks if we could allocate
1224                  * good large extent.
1225                  */
1226                 if (total - previous_total >=
1227                     osd_extent_bytes(osd) >> inode->i_blkbits)
1228                         osd_decay_extent_bytes(osd,
1229                                 (total - previous_total) << inode->i_blkbits);
1230                 /* look for next extent */
1231                 fp = NULL;
1232                 blocks += blocks_per_page * clen;
1233         }
1234 cleanup:
1235         if (rc == 0 && create &&
1236             start_blocks < pages * blocks_per_page) {
1237                 rc = osd_ldiskfs_map_write(inode, iobuf, osd, start_blocks,
1238                                            count, &disk_size, user_size);
1239                 LASSERT(start_blocks + count == pages * blocks_per_page);
1240         }
1241         return rc;
1242 }
1243
1244 static int osd_write_prep(const struct lu_env *env, struct dt_object *dt,
1245                           struct niobuf_local *lnb, int npages)
1246 {
1247         struct osd_thread_info *oti   = osd_oti_get(env);
1248         struct osd_iobuf       *iobuf = &oti->oti_iobuf;
1249         struct inode           *inode = osd_dt_obj(dt)->oo_inode;
1250         struct osd_device      *osd   = osd_obj2dev(osd_dt_obj(dt));
1251         ktime_t start, end;
1252         s64 timediff;
1253         ssize_t isize;
1254         __s64  maxidx;
1255         int i, rc = 0;
1256
1257         LASSERT(inode);
1258
1259         rc = osd_init_iobuf(osd, iobuf, 0, npages);
1260         if (unlikely(rc != 0))
1261                 RETURN(rc);
1262
1263         isize = i_size_read(inode);
1264         maxidx = ((isize + PAGE_SIZE - 1) >> PAGE_SHIFT) - 1;
1265
1266         start = ktime_get();
1267         for (i = 0; i < npages; i++) {
1268
1269                 /*
1270                  * till commit the content of the page is undefined
1271                  * we'll set it uptodate once bulk is done. otherwise
1272                  * subsequent reads can access non-stable data
1273                  */
1274                 ClearPageUptodate(lnb[i].lnb_page);
1275
1276                 if (lnb[i].lnb_len == PAGE_SIZE)
1277                         continue;
1278
1279                 if (maxidx >= lnb[i].lnb_page->index) {
1280                         osd_iobuf_add_page(iobuf, &lnb[i]);
1281                 } else {
1282                         long off;
1283                         char *p = kmap(lnb[i].lnb_page);
1284
1285                         off = lnb[i].lnb_page_offset;
1286                         if (off)
1287                                 memset(p, 0, off);
1288                         off = (lnb[i].lnb_page_offset + lnb[i].lnb_len) &
1289                               ~PAGE_MASK;
1290                         if (off)
1291                                 memset(p + off, 0, PAGE_SIZE - off);
1292                         kunmap(lnb[i].lnb_page);
1293                 }
1294         }
1295         end = ktime_get();
1296         timediff = ktime_us_delta(end, start);
1297         lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
1298
1299         if (iobuf->dr_npages) {
1300                 rc = osd_ldiskfs_map_inode_pages(inode, iobuf, osd, 0,
1301                                                  0, 0, NULL);
1302                 if (likely(rc == 0)) {
1303                         rc = osd_do_bio(osd, inode, iobuf, 0, 0);
1304                         /* do IO stats for preparation reads */
1305                         osd_fini_iobuf(osd, iobuf);
1306                 }
1307         }
1308         RETURN(rc);
1309 }
1310
1311 struct osd_fextent {
1312         sector_t        start;
1313         sector_t        end;
1314         __u32           flags;
1315         unsigned int    mapped:1;
1316 };
1317
1318 static int osd_is_mapped(struct dt_object *dt, __u64 offset,
1319                          struct osd_fextent *cached_extent)
1320 {
1321         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1322         sector_t block = offset >> inode->i_blkbits;
1323         sector_t start;
1324         struct fiemap_extent_info fei = { 0 };
1325         struct fiemap_extent fe = { 0 };
1326         int rc;
1327
1328         if (block >= cached_extent->start && block < cached_extent->end)
1329                 return cached_extent->mapped;
1330
1331         if (i_size_read(inode) == 0)
1332                 return 0;
1333
1334         /* Beyond EOF, must not be mapped */
1335         if (((i_size_read(inode) - 1) >> inode->i_blkbits) < block)
1336                 return 0;
1337
1338         fei.fi_extents_max = 1;
1339         fei.fi_extents_start = &fe;
1340
1341         rc = inode->i_op->fiemap(inode, &fei, offset, FIEMAP_MAX_OFFSET-offset);
1342         if (rc != 0)
1343                 return 0;
1344
1345         start = fe.fe_logical >> inode->i_blkbits;
1346         cached_extent->flags = fe.fe_flags;
1347         if (fei.fi_extents_mapped == 0) {
1348                 /* a special case - no extent found at this offset and forward.
1349                  * we can consider this as a hole to EOF. it's safe to cache
1350                  * as other threads can not allocate/punch blocks this thread
1351                  * is working on (LDLM). */
1352                 cached_extent->start = block;
1353                 cached_extent->end = i_size_read(inode) >> inode->i_blkbits;
1354                 cached_extent->mapped = 0;
1355                 return 0;
1356         }
1357
1358         if (start > block) {
1359                 cached_extent->start = block;
1360                 cached_extent->end = start;
1361                 cached_extent->mapped = 0;
1362         } else {
1363                 cached_extent->start = start;
1364                 cached_extent->end = (fe.fe_logical + fe.fe_length) >>
1365                                       inode->i_blkbits;
1366                 cached_extent->mapped = 1;
1367         }
1368
1369         return cached_extent->mapped;
1370 }
1371
1372 #define MAX_EXTENTS_PER_WRITE 100
1373 static int osd_declare_write_commit(const struct lu_env *env,
1374                                     struct dt_object *dt,
1375                                     struct niobuf_local *lnb, int npages,
1376                                     struct thandle *handle)
1377 {
1378         const struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1379         struct inode            *inode = osd_dt_obj(dt)->oo_inode;
1380         struct osd_thandle      *oh;
1381         int                     extents = 0, new_meta = 0;
1382         int                     depth, new_blocks = 0;
1383         int                     i;
1384         int                     dirty_groups = 0;
1385         int                     rc = 0;
1386         int                     credits = 0;
1387         long long               quota_space = 0;
1388         struct osd_fextent      mapped = { 0 }, extent = { 0 };
1389         enum osd_quota_local_flags local_flags = 0;
1390         enum osd_qid_declare_flags declare_flags = OSD_QID_BLK;
1391         unsigned int            extent_bytes;
1392         ENTRY;
1393
1394         LASSERT(handle != NULL);
1395         oh = container_of(handle, struct osd_thandle, ot_super);
1396         LASSERT(oh->ot_handle == NULL);
1397
1398         /*
1399          * We track a decaying average extent blocks per filesystem,
1400          * for most of time, it will be 1M, with filesystem becoming
1401          * heavily-fragmented, it will be reduced to 4K at the worst.
1402          */
1403         extent_bytes = osd_extent_bytes(osd);
1404         LASSERT(extent_bytes >= (1 << osd_sb(osd)->s_blocksize));
1405
1406         /* calculate number of extents (probably better to pass nb) */
1407         for (i = 0; i < npages; i++) {
1408                 /* ignore quota for the whole request if any page is from
1409                  * client cache or written by root.
1410                  *
1411                  * XXX once we drop the 1.8 client support, the checking
1412                  * for whether page is from cache can be simplified as:
1413                  * !(lnb[i].flags & OBD_BRW_SYNC)
1414                  *
1415                  * XXX we could handle this on per-lnb basis as done by
1416                  * grant.
1417                  */
1418                 if ((lnb[i].lnb_flags & OBD_BRW_NOQUOTA) ||
1419                     (lnb[i].lnb_flags & (OBD_BRW_FROM_GRANT | OBD_BRW_SYNC)) ==
1420                     OBD_BRW_FROM_GRANT)
1421                         declare_flags |= OSD_QID_FORCE;
1422
1423                 /*
1424                  * Convert unwritten extent might need split extents, could
1425                  * not skip it.
1426                  */
1427                 if (osd_is_mapped(dt, lnb[i].lnb_file_offset, &mapped) &&
1428                     !(mapped.flags & FIEMAP_EXTENT_UNWRITTEN)) {
1429                         lnb[i].lnb_flags |= OBD_BRW_MAPPED;
1430                         continue;
1431                 }
1432
1433                 if (lnb[i].lnb_flags & OBD_BRW_DONE) {
1434                         lnb[i].lnb_flags |= OBD_BRW_MAPPED;
1435                         continue;
1436                 }
1437
1438                 /* count only unmapped changes */
1439                 new_blocks++;
1440                 if (lnb[i].lnb_file_offset != extent.end || extent.end == 0) {
1441                         if (extent.end != 0)
1442                                 extents += (extent.end - extent.start +
1443                                             extent_bytes - 1) / extent_bytes;
1444                         extent.start = lnb[i].lnb_file_offset;
1445                         extent.end = lnb[i].lnb_file_offset + lnb[i].lnb_len;
1446                 } else {
1447                         extent.end += lnb[i].lnb_len;
1448                 }
1449
1450                 quota_space += PAGE_SIZE;
1451         }
1452
1453         credits++; /* inode */
1454         /*
1455          * overwrite case, no need to modify tree and
1456          * allocate blocks.
1457          */
1458         if (!extent.end)
1459                 goto out_declare;
1460
1461         extents += (extent.end - extent.start +
1462                     extent_bytes - 1) / extent_bytes;
1463         /**
1464          * with system space usage growing up, mballoc codes won't
1465          * try best to scan block group to align best free extent as
1466          * we can. So extent bytes per extent could be decayed to a
1467          * very small value, this could make us reserve too many credits.
1468          * We could be more optimistic in the credit reservations, even
1469          * in a case where the filesystem is nearly full, it is extremely
1470          * unlikely that the worst case would ever be hit.
1471          */
1472         if (extents > MAX_EXTENTS_PER_WRITE)
1473                 extents = MAX_EXTENTS_PER_WRITE;
1474
1475         /**
1476          * If we add a single extent, then in the worse case, each tree
1477          * level index/leaf need to be changed in case of the tree split.
1478          * If more extents are inserted, they could cause the whole tree
1479          * split more than once, but this is really rare.
1480          */
1481         if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL) {
1482                 /*
1483                  * many concurrent threads may grow tree by the time
1484                  * our transaction starts. so, consider 2 is a min depth.
1485                  */
1486                 depth = ext_depth(inode);
1487                 depth = min(max(depth, 1) + 1, LDISKFS_MAX_EXTENT_DEPTH);
1488                 if (extents <= 1) {
1489                         credits += depth * 2 * extents;
1490                         new_meta = depth;
1491                 } else {
1492                         credits += depth * 3 * extents;
1493                         new_meta = depth * 2 * extents;
1494                 }
1495         } else {
1496                 /*
1497                  * With N contiguous data blocks, we need at most
1498                  * N/EXT4_ADDR_PER_BLOCK(inode->i_sb) + 1 indirect blocks,
1499                  * 2 dindirect blocks, and 1 tindirect block
1500                  */
1501                 new_meta = DIV_ROUND_UP(new_blocks,
1502                                 LDISKFS_ADDR_PER_BLOCK(inode->i_sb)) + 4;
1503                 credits += new_meta;
1504         }
1505         dirty_groups += (extents + new_meta);
1506
1507         oh->oh_declared_ext = extents;
1508
1509         /* quota space for metadata blocks */
1510         quota_space += new_meta * LDISKFS_BLOCK_SIZE(osd_sb(osd));
1511
1512         /* quota space should be reported in 1K blocks */
1513         quota_space = toqb(quota_space);
1514
1515         /* each new block can go in different group (bitmap + gd) */
1516
1517         /* we can't dirty more bitmap blocks than exist */
1518         if (dirty_groups > LDISKFS_SB(osd_sb(osd))->s_groups_count)
1519                 credits += LDISKFS_SB(osd_sb(osd))->s_groups_count;
1520         else
1521                 credits += dirty_groups;
1522
1523         /* we can't dirty more gd blocks than exist */
1524         if (dirty_groups > LDISKFS_SB(osd_sb(osd))->s_gdb_count)
1525                 credits += LDISKFS_SB(osd_sb(osd))->s_gdb_count;
1526         else
1527                 credits += dirty_groups;
1528
1529         CDEBUG(D_INODE,
1530                "%s: inode #%lu extent_bytes %u extents %d credits %d\n",
1531                osd_ino2name(inode), inode->i_ino, extent_bytes, extents,
1532                credits);
1533
1534 out_declare:
1535         osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
1536
1537         /* make sure the over quota flags were not set */
1538         lnb[0].lnb_flags &= ~OBD_BRW_OVER_ALLQUOTA;
1539
1540         rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
1541                                    i_projid_read(inode), quota_space, oh,
1542                                    osd_dt_obj(dt), &local_flags, declare_flags);
1543
1544         /* we need only to store the overquota flags in the first lnb for
1545          * now, once we support multiple objects BRW, this code needs be
1546          * revised.
1547          */
1548         if (local_flags & QUOTA_FL_OVER_USRQUOTA)
1549                 lnb[0].lnb_flags |= OBD_BRW_OVER_USRQUOTA;
1550         if (local_flags & QUOTA_FL_OVER_GRPQUOTA)
1551                 lnb[0].lnb_flags |= OBD_BRW_OVER_GRPQUOTA;
1552         if (local_flags & QUOTA_FL_OVER_PRJQUOTA)
1553                 lnb[0].lnb_flags |= OBD_BRW_OVER_PRJQUOTA;
1554
1555         if (rc == 0)
1556                 rc = osd_trunc_lock(osd_dt_obj(dt), oh, true);
1557
1558         RETURN(rc);
1559 }
1560
1561 /* Check if a block is allocated or not */
1562 static int osd_write_commit(const struct lu_env *env, struct dt_object *dt,
1563                             struct niobuf_local *lnb, int npages,
1564                             struct thandle *thandle, __u64 user_size)
1565 {
1566         struct osd_thread_info *oti = osd_oti_get(env);
1567         struct osd_iobuf *iobuf = &oti->oti_iobuf;
1568         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1569         struct osd_device  *osd = osd_obj2dev(osd_dt_obj(dt));
1570         int rc = 0, i, check_credits = 0;
1571
1572         LASSERT(inode);
1573
1574         rc = osd_init_iobuf(osd, iobuf, 1, npages);
1575         if (unlikely(rc != 0))
1576                 RETURN(rc);
1577
1578         dquot_initialize(inode);
1579
1580         for (i = 0; i < npages; i++) {
1581                 if (lnb[i].lnb_rc == -ENOSPC &&
1582                     (lnb[i].lnb_flags & OBD_BRW_MAPPED)) {
1583                         /* Allow the write to proceed if overwriting an
1584                          * existing block
1585                          */
1586                         lnb[i].lnb_rc = 0;
1587                 }
1588
1589                 if (lnb[i].lnb_rc) { /* ENOSPC, network RPC error, etc. */
1590                         CDEBUG(D_INODE, "Skipping [%d] == %d\n", i,
1591                                lnb[i].lnb_rc);
1592                         LASSERT(lnb[i].lnb_page);
1593                         generic_error_remove_page(inode->i_mapping,
1594                                                   lnb[i].lnb_page);
1595                         continue;
1596                 }
1597
1598                 if (lnb[i].lnb_flags & OBD_BRW_DONE)
1599                         continue;
1600
1601                 if (!(lnb[i].lnb_flags & OBD_BRW_MAPPED))
1602                         check_credits = 1;
1603
1604                 LASSERT(PageLocked(lnb[i].lnb_page));
1605                 LASSERT(!PageWriteback(lnb[i].lnb_page));
1606
1607                 /*
1608                  * Since write and truncate are serialized by oo_sem, even
1609                  * partial-page truncate should not leave dirty pages in the
1610                  * page cache.
1611                  */
1612                 LASSERT(!PageDirty(lnb[i].lnb_page));
1613
1614                 SetPageUptodate(lnb[i].lnb_page);
1615
1616                 osd_iobuf_add_page(iobuf, &lnb[i]);
1617         }
1618
1619         osd_trans_exec_op(env, thandle, OSD_OT_WRITE);
1620
1621         if (OBD_FAIL_CHECK(OBD_FAIL_OST_MAPBLK_ENOSPC)) {
1622                 rc = -ENOSPC;
1623         } else if (iobuf->dr_npages > 0) {
1624                 rc = osd_ldiskfs_map_inode_pages(inode, iobuf, osd,
1625                                                  1, user_size,
1626                                                  check_credits,
1627                                                  thandle);
1628         } else {
1629                 /* no pages to write, no transno is needed */
1630                 thandle->th_local = 1;
1631         }
1632
1633         if (rc != 0 && !thandle->th_restart_tran)
1634                 osd_fini_iobuf(osd, iobuf);
1635
1636         osd_trans_exec_check(env, thandle, OSD_OT_WRITE);
1637
1638         if (unlikely(rc != 0 && !thandle->th_restart_tran)) {
1639                 /* if write fails, we should drop pages from the cache */
1640                 for (i = 0; i < npages; i++) {
1641                         if (lnb[i].lnb_page == NULL)
1642                                 continue;
1643                         if (!PagePrivate2(lnb[i].lnb_page)) {
1644                                 LASSERT(PageLocked(lnb[i].lnb_page));
1645                                 generic_error_remove_page(inode->i_mapping,
1646                                                           lnb[i].lnb_page);
1647                         }
1648                 }
1649         }
1650
1651         RETURN(rc);
1652 }
1653
1654 static int osd_read_prep(const struct lu_env *env, struct dt_object *dt,
1655                          struct niobuf_local *lnb, int npages)
1656 {
1657         struct osd_thread_info *oti = osd_oti_get(env);
1658         struct osd_iobuf *iobuf = &oti->oti_iobuf;
1659         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1660         struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1661         int rc = 0, i, cache_hits = 0, cache_misses = 0;
1662         ktime_t start, end;
1663         s64 timediff;
1664         loff_t isize;
1665
1666         LASSERT(inode);
1667
1668         rc = osd_init_iobuf(osd, iobuf, 0, npages);
1669         if (unlikely(rc != 0))
1670                 RETURN(rc);
1671
1672         isize = i_size_read(inode);
1673
1674         start = ktime_get();
1675         for (i = 0; i < npages; i++) {
1676
1677                 if (isize <= lnb[i].lnb_file_offset)
1678                         /* If there's no more data, abort early.
1679                          * lnb->lnb_rc == 0, so it's easy to detect later.
1680                          */
1681                         break;
1682
1683                 /* instead of looking if we go beyong isize, send complete
1684                  * pages all the time
1685                  */
1686                 lnb[i].lnb_rc = lnb[i].lnb_len;
1687
1688                 /* Bypass disk read if fail_loc is set properly */
1689                 if (OBD_FAIL_CHECK(OBD_FAIL_OST_FAKE_RW))
1690                         SetPageUptodate(lnb[i].lnb_page);
1691
1692                 if (PageUptodate(lnb[i].lnb_page)) {
1693                         cache_hits++;
1694                         unlock_page(lnb[i].lnb_page);
1695                 } else {
1696                         cache_misses++;
1697                         osd_iobuf_add_page(iobuf, &lnb[i]);
1698                 }
1699                 /* no need to unlock in osd_bufs_put(), the sooner page is
1700                  * unlocked, the earlier another client can access it.
1701                  * notice real unlock_page() can be called few lines
1702                  * below after osd_do_bio(). lnb is a per-thread, so it's
1703                  * fine to have PG_locked and lnb_locked inconsistent here
1704                  */
1705                 lnb[i].lnb_locked = 0;
1706         }
1707         end = ktime_get();
1708         timediff = ktime_us_delta(end, start);
1709         lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
1710
1711         if (cache_hits != 0)
1712                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_HIT,
1713                                     cache_hits);
1714         if (cache_misses != 0)
1715                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_MISS,
1716                                     cache_misses);
1717         if (cache_hits + cache_misses != 0)
1718                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_ACCESS,
1719                                     cache_hits + cache_misses);
1720
1721         if (iobuf->dr_npages) {
1722                 rc = osd_ldiskfs_map_inode_pages(inode, iobuf, osd, 0,
1723                                                  0, 0, NULL);
1724                 if (!rc)
1725                         rc = osd_do_bio(osd, inode, iobuf, 0, 0);
1726
1727                 /* IO stats will be done in osd_bufs_put() */
1728
1729                 /* early release to let others read data during the bulk */
1730                 for (i = 0; i < iobuf->dr_npages; i++) {
1731                         LASSERT(PageLocked(iobuf->dr_pages[i]));
1732                         if (!PagePrivate2(iobuf->dr_pages[i]))
1733                                 unlock_page(iobuf->dr_pages[i]);
1734                 }
1735         }
1736
1737         RETURN(rc);
1738 }
1739
1740 /*
1741  * XXX: Another layering violation for now.
1742  *
1743  * We don't want to use ->f_op->read methods, because generic file write
1744  *
1745  *         - serializes on ->i_sem, and
1746  *
1747  *         - does a lot of extra work like balance_dirty_pages(),
1748  *
1749  * which doesn't work for globally shared files like /last_rcvd.
1750  */
1751 static int osd_ldiskfs_readlink(struct inode *inode, char *buffer, int buflen)
1752 {
1753         struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
1754
1755         memcpy(buffer, (char *)ei->i_data, buflen);
1756
1757         return  buflen;
1758 }
1759
1760 int osd_ldiskfs_read(struct inode *inode, void *buf, int size, loff_t *offs)
1761 {
1762         struct buffer_head *bh;
1763         unsigned long block;
1764         int osize;
1765         int blocksize;
1766         int csize;
1767         int boffs;
1768
1769         /* prevent reading after eof */
1770         spin_lock(&inode->i_lock);
1771         if (i_size_read(inode) < *offs + size) {
1772                 loff_t diff = i_size_read(inode) - *offs;
1773
1774                 spin_unlock(&inode->i_lock);
1775                 if (diff < 0) {
1776                         CDEBUG(D_OTHER,
1777                                "size %llu is too short to read @%llu\n",
1778                                i_size_read(inode), *offs);
1779                         return -EBADR;
1780                 } else if (diff == 0) {
1781                         return 0;
1782                 } else {
1783                         size = diff;
1784                 }
1785         } else {
1786                 spin_unlock(&inode->i_lock);
1787         }
1788
1789         blocksize = 1 << inode->i_blkbits;
1790         osize = size;
1791         while (size > 0) {
1792                 block = *offs >> inode->i_blkbits;
1793                 boffs = *offs & (blocksize - 1);
1794                 csize = min(blocksize - boffs, size);
1795                 bh = __ldiskfs_bread(NULL, inode, block, 0);
1796                 if (IS_ERR(bh)) {
1797                         CERROR("%s: can't read %u@%llu on ino %lu: rc = %ld\n",
1798                                osd_ino2name(inode), csize, *offs, inode->i_ino,
1799                                PTR_ERR(bh));
1800                         return PTR_ERR(bh);
1801                 }
1802
1803                 if (bh != NULL) {
1804                         memcpy(buf, bh->b_data + boffs, csize);
1805                         brelse(bh);
1806                 } else {
1807                         memset(buf, 0, csize);
1808                 }
1809
1810                 *offs += csize;
1811                 buf += csize;
1812                 size -= csize;
1813         }
1814         return osize;
1815 }
1816
1817 static ssize_t osd_read(const struct lu_env *env, struct dt_object *dt,
1818                         struct lu_buf *buf, loff_t *pos)
1819 {
1820         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1821         int rc;
1822
1823         /* Read small symlink from inode body as we need to maintain correct
1824          * on-disk symlinks for ldiskfs.
1825          */
1826         if (S_ISLNK(dt->do_lu.lo_header->loh_attr)) {
1827                 loff_t size = i_size_read(inode);
1828
1829                 if (buf->lb_len < size)
1830                         return -EOVERFLOW;
1831
1832                 if (size < sizeof(LDISKFS_I(inode)->i_data))
1833                         rc = osd_ldiskfs_readlink(inode, buf->lb_buf, size);
1834                 else
1835                         rc = osd_ldiskfs_read(inode, buf->lb_buf, size, pos);
1836         } else {
1837                 rc = osd_ldiskfs_read(inode, buf->lb_buf, buf->lb_len, pos);
1838         }
1839
1840         return rc;
1841 }
1842
1843 static inline int osd_extents_enabled(struct super_block *sb,
1844                                       struct inode *inode)
1845 {
1846         if (inode != NULL) {
1847                 if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL)
1848                         return 1;
1849         } else if (ldiskfs_has_feature_extents(sb)) {
1850                 return 1;
1851         }
1852         return 0;
1853 }
1854
1855 int osd_calc_bkmap_credits(struct super_block *sb, struct inode *inode,
1856                            const loff_t size, const loff_t pos,
1857                            const int blocks)
1858 {
1859         int credits, bits, bs, i;
1860
1861         bits = sb->s_blocksize_bits;
1862         bs = 1 << bits;
1863
1864         /* legacy blockmap: 3 levels * 3 (bitmap,gd,itself)
1865          * we do not expect blockmaps on the large files,
1866          * so let's shrink it to 2 levels (4GB files)
1867          */
1868
1869         /* this is default reservation: 2 levels */
1870         credits = (blocks + 2) * 3;
1871
1872         /* actual offset is unknown, hard to optimize */
1873         if (pos == -1)
1874                 return credits;
1875
1876         /* now check for few specific cases to optimize */
1877         if (pos + size <= LDISKFS_NDIR_BLOCKS * bs) {
1878                 /* no indirects */
1879                 credits = blocks;
1880                 /* allocate if not allocated */
1881                 if (inode == NULL) {
1882                         credits += blocks * 2;
1883                         return credits;
1884                 }
1885                 for (i = (pos >> bits); i < (pos >> bits) + blocks; i++) {
1886                         LASSERT(i < LDISKFS_NDIR_BLOCKS);
1887                         if (LDISKFS_I(inode)->i_data[i] == 0)
1888                                 credits += 2;
1889                 }
1890         } else if (pos + size <= (LDISKFS_NDIR_BLOCKS + 1024) * bs) {
1891                 /* single indirect */
1892                 credits = blocks * 3;
1893                 if (inode == NULL ||
1894                     LDISKFS_I(inode)->i_data[LDISKFS_IND_BLOCK] == 0)
1895                         credits += 3;
1896                 else
1897                         /* The indirect block may be modified. */
1898                         credits += 1;
1899         }
1900
1901         return credits;
1902 }
1903
1904 static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt,
1905                                  const struct lu_buf *buf, loff_t _pos,
1906                                  struct thandle *handle)
1907 {
1908         struct osd_object  *obj  = osd_dt_obj(dt);
1909         struct inode       *inode = obj->oo_inode;
1910         struct super_block *sb = osd_sb(osd_obj2dev(obj));
1911         struct osd_thandle *oh;
1912         int                 rc = 0, est = 0, credits, blocks, allocated = 0;
1913         int                 bits, bs;
1914         int                 depth, size;
1915         loff_t              pos;
1916         ENTRY;
1917
1918         LASSERT(buf != NULL);
1919         LASSERT(handle != NULL);
1920
1921         oh = container_of(handle, struct osd_thandle, ot_super);
1922         LASSERT(oh->ot_handle == NULL);
1923
1924         size = buf->lb_len;
1925         bits = sb->s_blocksize_bits;
1926         bs = 1 << bits;
1927
1928         if (_pos == -1) {
1929                 /* if this is an append, then we
1930                  * should expect cross-block record
1931                  */
1932                 pos = 0;
1933         } else {
1934                 pos = _pos;
1935         }
1936
1937         /* blocks to modify */
1938         blocks = ((pos + size + bs - 1) >> bits) - (pos >> bits);
1939         LASSERT(blocks > 0);
1940
1941         if (inode != NULL && _pos != -1) {
1942                 /* object size in blocks */
1943                 est = (i_size_read(inode) + bs - 1) >> bits;
1944                 allocated = inode->i_blocks >> (bits - 9);
1945                 if (pos + size <= i_size_read(inode) && est <= allocated) {
1946                         /* looks like an overwrite, no need to modify tree */
1947                         credits = blocks;
1948                         /* no need to modify i_size */
1949                         goto out;
1950                 }
1951         }
1952
1953         if (osd_extents_enabled(sb, inode)) {
1954                 /*
1955                  * many concurrent threads may grow tree by the time
1956                  * our transaction starts. so, consider 2 is a min depth
1957                  * for every level we may need to allocate a new block
1958                  * and take some entries from the old one. so, 3 blocks
1959                  * to allocate (bitmap, gd, itself) + old block - 4 per
1960                  * level.
1961                  */
1962                 depth = inode != NULL ? ext_depth(inode) : 0;
1963                 depth = min(max(depth, 1) + 1, LDISKFS_MAX_EXTENT_DEPTH);
1964                 credits = depth;
1965                 /* if not append, then split may need to modify
1966                  * existing blocks moving entries into the new ones
1967                  */
1968                 if (_pos != -1)
1969                         credits += depth;
1970                 /* blocks to store data: bitmap,gd,itself */
1971                 credits += blocks * 3;
1972         } else {
1973                 credits = osd_calc_bkmap_credits(sb, inode, size, _pos, blocks);
1974         }
1975         /* if inode is created as part of the transaction,
1976          * then it's counted already by the creation method
1977          */
1978         if (inode != NULL)
1979                 credits++;
1980
1981 out:
1982
1983         osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
1984
1985         /* dt_declare_write() is usually called for system objects, such
1986          * as llog or last_rcvd files. We needn't enforce quota on those
1987          * objects, so always set the lqi_space as 0.
1988          */
1989         if (inode != NULL)
1990                 rc = osd_declare_inode_qid(env, i_uid_read(inode),
1991                                            i_gid_read(inode),
1992                                            i_projid_read(inode), 0,
1993                                            oh, obj, NULL, OSD_QID_BLK);
1994
1995         if (rc == 0)
1996                 rc = osd_trunc_lock(obj, oh, true);
1997
1998         RETURN(rc);
1999 }
2000
2001 static int osd_ldiskfs_writelink(struct inode *inode, char *buffer, int buflen)
2002 {
2003         /* LU-2634: clear the extent format for fast symlink */
2004         ldiskfs_clear_inode_flag(inode, LDISKFS_INODE_EXTENTS);
2005
2006         memcpy((char *)&LDISKFS_I(inode)->i_data, (char *)buffer, buflen);
2007         spin_lock(&inode->i_lock);
2008         LDISKFS_I(inode)->i_disksize = buflen;
2009         i_size_write(inode, buflen);
2010         spin_unlock(&inode->i_lock);
2011         osd_dirty_inode(inode, I_DIRTY_DATASYNC);
2012
2013         return 0;
2014 }
2015
2016 static int osd_ldiskfs_write_record(struct dt_object *dt, void *buf,
2017                                     int bufsize, int write_NUL, loff_t *offs,
2018                                     handle_t *handle)
2019 {
2020         struct inode *inode = osd_dt_obj(dt)->oo_inode;
2021         struct buffer_head *bh        = NULL;
2022         loff_t              offset    = *offs;
2023         loff_t              new_size  = i_size_read(inode);
2024         unsigned long       block;
2025         int                 blocksize = 1 << inode->i_blkbits;
2026         struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
2027         int                 err = 0;
2028         int                 size;
2029         int                 boffs;
2030         int                 dirty_inode = 0;
2031         bool create, sparse, sync = false;
2032
2033         if (write_NUL) {
2034                 /*
2035                  * long symlink write does not count the NUL terminator in
2036                  * bufsize, we write it, and the inode's file size does not
2037                  * count the NUL terminator as well.
2038                  */
2039                 ((char *)buf)[bufsize] = '\0';
2040                 ++bufsize;
2041         }
2042
2043         /* only the first flag-set matters */
2044         dirty_inode = !test_and_set_bit(LDISKFS_INODE_JOURNAL_DATA,
2045                                        &ei->i_flags);
2046
2047         /* sparse checking is racy, but sparse is very rare case, leave as is */
2048         sparse = (new_size > 0 && (inode->i_blocks >> (inode->i_blkbits - 9)) <
2049                   ((new_size - 1) >> inode->i_blkbits) + 1);
2050
2051         while (bufsize > 0) {
2052                 int credits = handle->h_buffer_credits;
2053                 unsigned long last_block = (new_size == 0) ? 0 :
2054                                            (new_size - 1) >> inode->i_blkbits;
2055
2056                 if (bh)
2057                         brelse(bh);
2058
2059                 block = offset >> inode->i_blkbits;
2060                 boffs = offset & (blocksize - 1);
2061                 size = min(blocksize - boffs, bufsize);
2062                 sync = (block > last_block || new_size == 0 || sparse);
2063
2064                 if (sync)
2065                         down(&ei->i_append_sem);
2066
2067                 bh = __ldiskfs_bread(handle, inode, block, 0);
2068
2069                 if (unlikely(IS_ERR_OR_NULL(bh) && !sync))
2070                         CWARN(
2071                               "%s: adding bh without locking off %llu (block %lu, size %d, offs %llu)\n",
2072                               osd_ino2name(inode),
2073                               offset, block, bufsize, *offs);
2074
2075                 if (IS_ERR_OR_NULL(bh)) {
2076                         struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
2077                         int flags = LDISKFS_GET_BLOCKS_CREATE;
2078
2079                         /* while the file system is being mounted, avoid
2080                          * preallocation otherwise mount can take a long
2081                          * time as mballoc cache is cold.
2082                          * XXX: this is a workaround until we have a proper
2083                          *      fix in mballoc
2084                          * XXX: works with extent-based files only */
2085                         if (!osd->od_cl_seq)
2086                                 flags |= LDISKFS_GET_BLOCKS_NO_NORMALIZE;
2087                         bh = __ldiskfs_bread(handle, inode, block, flags);
2088                         create = true;
2089                 } else {
2090                         if (sync) {
2091                                 up(&ei->i_append_sem);
2092                                 sync = false;
2093                         }
2094                         create = false;
2095                 }
2096                 if (IS_ERR_OR_NULL(bh)) {
2097                         if (bh == NULL) {
2098                                 err = -EIO;
2099                         } else {
2100                                 err = PTR_ERR(bh);
2101                                 bh = NULL;
2102                         }
2103
2104                         CERROR(
2105                                "%s: error reading offset %llu (block %lu, size %d, offs %llu), credits %d/%d: rc = %d\n",
2106                                osd_ino2name(inode), offset, block, bufsize,
2107                                *offs, credits, handle->h_buffer_credits, err);
2108                         break;
2109                 }
2110
2111                 err = ldiskfs_journal_get_write_access(handle, bh);
2112                 if (err) {
2113                         CERROR("journal_get_write_access() returned error %d\n",
2114                                err);
2115                         break;
2116                 }
2117                 LASSERTF(boffs + size <= bh->b_size,
2118                          "boffs %d size %d bh->b_size %lu\n",
2119                          boffs, size, (unsigned long)bh->b_size);
2120                 if (create) {
2121                         memset(bh->b_data, 0, bh->b_size);
2122                         if (sync) {
2123                                 up(&ei->i_append_sem);
2124                                 sync = false;
2125                         }
2126                 }
2127                 memcpy(bh->b_data + boffs, buf, size);
2128                 err = ldiskfs_handle_dirty_metadata(handle, NULL, bh);
2129                 if (err)
2130                         break;
2131
2132                 if (offset + size > new_size)
2133                         new_size = offset + size;
2134                 offset += size;
2135                 bufsize -= size;
2136                 buf += size;
2137         }
2138         if (sync)
2139                 up(&ei->i_append_sem);
2140
2141         if (bh)
2142                 brelse(bh);
2143
2144         if (write_NUL)
2145                 --new_size;
2146         /* correct in-core and on-disk sizes */
2147         if (new_size > i_size_read(inode)) {
2148                 spin_lock(&inode->i_lock);
2149                 if (new_size > i_size_read(inode))
2150                         i_size_write(inode, new_size);
2151                 if (i_size_read(inode) > ei->i_disksize) {
2152                         ei->i_disksize = i_size_read(inode);
2153                         dirty_inode = 1;
2154                 }
2155                 spin_unlock(&inode->i_lock);
2156         }
2157         if (dirty_inode)
2158                 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
2159
2160         if (err == 0)
2161                 *offs = offset;
2162         return err;
2163 }
2164
2165 static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt,
2166                          const struct lu_buf *buf, loff_t *pos,
2167                          struct thandle *handle)
2168 {
2169         struct inode            *inode = osd_dt_obj(dt)->oo_inode;
2170         struct osd_thandle      *oh;
2171         ssize_t                 result;
2172         int                     is_link;
2173
2174         LASSERT(dt_object_exists(dt));
2175
2176         LASSERT(handle != NULL);
2177         LASSERT(inode != NULL);
2178         dquot_initialize(inode);
2179
2180         /* XXX: don't check: one declared chunk can be used many times */
2181         /* osd_trans_exec_op(env, handle, OSD_OT_WRITE); */
2182
2183         oh = container_of(handle, struct osd_thandle, ot_super);
2184         LASSERT(oh->ot_handle->h_transaction != NULL);
2185         osd_trans_exec_op(env, handle, OSD_OT_WRITE);
2186
2187         /* Write small symlink to inode body as we need to maintain correct
2188          * on-disk symlinks for ldiskfs.
2189          * Note: the buf->lb_buf contains a NUL terminator while buf->lb_len
2190          * does not count it in.
2191          */
2192         is_link = S_ISLNK(dt->do_lu.lo_header->loh_attr);
2193         if (is_link && (buf->lb_len < sizeof(LDISKFS_I(inode)->i_data)))
2194                 result = osd_ldiskfs_writelink(inode, buf->lb_buf, buf->lb_len);
2195         else
2196                 result = osd_ldiskfs_write_record(dt, buf->lb_buf, buf->lb_len,
2197                                                   is_link, pos, oh->ot_handle);
2198         if (result == 0)
2199                 result = buf->lb_len;
2200
2201         osd_trans_exec_check(env, handle, OSD_OT_WRITE);
2202
2203         return result;
2204 }
2205
2206 static int osd_declare_fallocate(const struct lu_env *env,
2207                                  struct dt_object *dt, __u64 start, __u64 end,
2208                                  int mode, struct thandle *th)
2209 {
2210         struct osd_thandle *oh = container_of(th, struct osd_thandle, ot_super);
2211         struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
2212         struct inode *inode = osd_dt_obj(dt)->oo_inode;
2213         long long quota_space = 0;
2214         /* 5 is max tree depth. (inode + 4 index blocks) */
2215         int depth = 5;
2216         int rc;
2217
2218         ENTRY;
2219
2220         /*
2221          * mode == 0 (which is standard prealloc) and PUNCH is supported
2222          * Rest of mode options is not supported yet.
2223          */
2224         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2225                 RETURN(-EOPNOTSUPP);
2226
2227         /* disable fallocate completely */
2228         if (osd_dev(dt->do_lu.lo_dev)->od_fallocate_zero_blocks < 0)
2229                 RETURN(-EOPNOTSUPP);
2230
2231         LASSERT(th);
2232         LASSERT(inode);
2233
2234         if (mode & FALLOC_FL_PUNCH_HOLE) {
2235                 rc = osd_declare_inode_qid(env, i_uid_read(inode),
2236                                            i_gid_read(inode),
2237                                            i_projid_read(inode), 0, oh,
2238                                            osd_dt_obj(dt), NULL, OSD_QID_BLK);
2239                 if (rc == 0)
2240                         rc = osd_trunc_lock(osd_dt_obj(dt), oh, false);
2241                 RETURN(rc);
2242         }
2243
2244         /* quota space for metadata blocks
2245          * approximate metadata estimate should be good enough.
2246          */
2247         quota_space += PAGE_SIZE;
2248         quota_space += depth * LDISKFS_BLOCK_SIZE(osd_sb(osd));
2249
2250         /* quota space should be reported in 1K blocks */
2251         quota_space = toqb(quota_space) + toqb(end - start) +
2252                       LDISKFS_META_TRANS_BLOCKS(inode->i_sb);
2253
2254         /* We don't need to reserve credits for whole fallocate here.
2255          * We reserve space only for metadata. Fallocate credits are
2256          * extended as required
2257          */
2258         rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
2259                                    i_projid_read(inode), quota_space, oh,
2260                                    osd_dt_obj(dt), NULL, OSD_QID_BLK);
2261         RETURN(rc);
2262 }
2263
2264 static int osd_fallocate_preallocate(const struct lu_env *env,
2265                                      struct dt_object *dt,
2266                                      __u64 start, __u64 end, int mode,
2267                                      struct thandle *th)
2268 {
2269         struct osd_thandle *oh = container_of(th, struct osd_thandle, ot_super);
2270         handle_t *handle = ldiskfs_journal_current_handle();
2271         unsigned int save_credits = oh->ot_credits;
2272         struct osd_object *obj = osd_dt_obj(dt);
2273         struct inode *inode = obj->oo_inode;
2274         struct ldiskfs_map_blocks map;
2275         unsigned int credits;
2276         ldiskfs_lblk_t blen;
2277         ldiskfs_lblk_t boff;
2278         loff_t new_size = 0;
2279         int depth = 0;
2280         int flags;
2281         int rc = 0;
2282
2283         ENTRY;
2284
2285         LASSERT(dt_object_exists(dt));
2286         LASSERT(osd_invariant(obj));
2287         LASSERT(inode != NULL);
2288
2289         CDEBUG(D_INODE, "fallocate: inode #%lu: start %llu end %llu mode %d\n",
2290                inode->i_ino, start, end, mode);
2291
2292         dquot_initialize(inode);
2293
2294         LASSERT(th);
2295
2296         boff = start >> inode->i_blkbits;
2297         blen = (ALIGN(end, 1 << inode->i_blkbits) >> inode->i_blkbits) - boff;
2298
2299         /* Create and mark new extents as either zero or unwritten */
2300         flags = (osd_dev(dt->do_lu.lo_dev)->od_fallocate_zero_blocks ||
2301                  !ldiskfs_test_inode_flag(inode, LDISKFS_INODE_EXTENTS)) ?
2302                 LDISKFS_GET_BLOCKS_CREATE_ZERO :
2303                 LDISKFS_GET_BLOCKS_CREATE_UNWRIT_EXT;
2304 #ifndef HAVE_LDISKFS_GET_BLOCKS_KEEP_SIZE
2305         if (mode & FALLOC_FL_KEEP_SIZE)
2306                 flags |= LDISKFS_GET_BLOCKS_KEEP_SIZE;
2307 #endif
2308         inode_lock(inode);
2309
2310         if (!(mode & FALLOC_FL_KEEP_SIZE) && (end > i_size_read(inode) ||
2311             end > LDISKFS_I(inode)->i_disksize)) {
2312                 new_size = end;
2313                 rc = inode_newsize_ok(inode, new_size);
2314                 if (rc)
2315                         GOTO(out, rc);
2316         }
2317
2318         inode_dio_wait(inode);
2319
2320         map.m_lblk = boff;
2321         map.m_len = blen;
2322
2323         /* Don't normalize the request if it can fit in one extent so
2324          * that it doesn't get unnecessarily split into multiple extents.
2325          */
2326         if (blen <= EXT_UNWRITTEN_MAX_LEN)
2327                 flags |= LDISKFS_GET_BLOCKS_NO_NORMALIZE;
2328
2329         /*
2330          * credits to insert 1 extent into extent tree.
2331          */
2332         credits = osd_chunk_trans_blocks(inode, blen);
2333         depth = ext_depth(inode);
2334
2335         while (rc >= 0 && blen) {
2336                 loff_t epos;
2337
2338                 /*
2339                  * Recalculate credits when extent tree depth changes.
2340                  */
2341                 if (depth != ext_depth(inode)) {
2342                         credits = osd_chunk_trans_blocks(inode, blen);
2343                         depth = ext_depth(inode);
2344                 }
2345
2346                 /* TODO: quota check */
2347                 rc = osd_extend_restart_trans(handle, credits, inode);
2348                 if (rc)
2349                         break;
2350
2351                 rc = ldiskfs_map_blocks(handle, inode, &map, flags);
2352                 if (rc <= 0) {
2353                         CDEBUG(D_INODE,
2354                                "inode #%lu: block %u: len %u: ldiskfs_map_blocks returned %d\n",
2355                                inode->i_ino, map.m_lblk, map.m_len, rc);
2356                         ldiskfs_mark_inode_dirty(handle, inode);
2357                         break;
2358                 }
2359
2360                 map.m_lblk += rc;
2361                 map.m_len = blen = blen - rc;
2362                 epos = (loff_t)map.m_lblk << inode->i_blkbits;
2363                 inode->i_ctime = current_time(inode);
2364                 if (new_size) {
2365                         if (epos > end)
2366                                 epos = end;
2367                         if (ldiskfs_update_inode_size(inode, epos) & 0x1)
2368                                 inode->i_mtime = inode->i_ctime;
2369 #ifndef HAVE_LDISKFS_GET_BLOCKS_KEEP_SIZE
2370                 } else {
2371                         if (epos > inode->i_size)
2372                                 ldiskfs_set_inode_flag(inode,
2373                                                        LDISKFS_INODE_EOFBLOCKS);
2374 #endif
2375                 }
2376
2377                 ldiskfs_mark_inode_dirty(handle, inode);
2378         }
2379
2380 out:
2381         /* extand credits if needed for operations such as attribute set */
2382         if (rc >= 0)
2383                 rc = osd_extend_restart_trans(handle, save_credits, inode);
2384
2385         inode_unlock(inode);
2386
2387         RETURN(rc);
2388 }
2389
2390 static int osd_fallocate_punch(const struct lu_env *env, struct dt_object *dt,
2391                                __u64 start, __u64 end, int mode,
2392                                struct thandle *th)
2393 {
2394         struct osd_object *obj = osd_dt_obj(dt);
2395         struct inode *inode = obj->oo_inode;
2396         struct osd_access_lock *al;
2397         struct osd_thandle *oh;
2398         int rc = 0, found = 0;
2399
2400         ENTRY;
2401
2402         LASSERT(dt_object_exists(dt));
2403         LASSERT(osd_invariant(obj));
2404         LASSERT(inode != NULL);
2405
2406         dquot_initialize(inode);
2407
2408         LASSERT(th);
2409         oh = container_of(th, struct osd_thandle, ot_super);
2410         LASSERT(oh->ot_handle->h_transaction != NULL);
2411
2412         list_for_each_entry(al, &oh->ot_trunc_locks, tl_list) {
2413                 if (obj != al->tl_obj)
2414                         continue;
2415                 LASSERT(al->tl_shared == 0);
2416                 found = 1;
2417                 /* do actual punch in osd_trans_stop() */
2418                 al->tl_start = start;
2419                 al->tl_end = end;
2420                 al->tl_mode = mode;
2421                 al->tl_punch = true;
2422                 break;
2423         }
2424
2425         RETURN(rc);
2426 }
2427
2428 static int osd_fallocate(const struct lu_env *env, struct dt_object *dt,
2429                          __u64 start, __u64 end, int mode, struct thandle *th)
2430 {
2431         int rc;
2432
2433         ENTRY;
2434
2435         if (mode & FALLOC_FL_PUNCH_HOLE) {
2436                 /* punch */
2437                 rc = osd_fallocate_punch(env, dt, start, end, mode, th);
2438         } else {
2439                 /* standard preallocate */
2440                 rc = osd_fallocate_preallocate(env, dt, start, end, mode, th);
2441         }
2442         RETURN(rc);
2443 }
2444
2445 static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt,
2446                              __u64 start, __u64 end, struct thandle *th)
2447 {
2448         struct osd_thandle *oh;
2449         struct inode       *inode;
2450         int                 rc;
2451         ENTRY;
2452
2453         LASSERT(th);
2454         oh = container_of(th, struct osd_thandle, ot_super);
2455
2456         /*
2457          * we don't need to reserve credits for whole truncate
2458          * it's not possible as truncate may need to free too many
2459          * blocks and that won't fit a single transaction. instead
2460          * we reserve credits to change i_size and put inode onto
2461          * orphan list. if needed truncate will extend or restart
2462          * transaction
2463          */
2464         osd_trans_declare_op(env, oh, OSD_OT_PUNCH,
2465                              osd_dto_credits_noquota[DTO_ATTR_SET_BASE] + 3);
2466
2467         inode = osd_dt_obj(dt)->oo_inode;
2468         LASSERT(inode);
2469
2470         rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
2471                                    i_projid_read(inode), 0, oh, osd_dt_obj(dt),
2472                                    NULL, OSD_QID_BLK);
2473
2474         if (rc == 0)
2475                 rc = osd_trunc_lock(osd_dt_obj(dt), oh, false);
2476
2477         RETURN(rc);
2478 }
2479
2480 static int osd_punch(const struct lu_env *env, struct dt_object *dt,
2481                      __u64 start, __u64 end, struct thandle *th)
2482 {
2483         struct osd_object *obj = osd_dt_obj(dt);
2484         struct osd_device *osd = osd_obj2dev(obj);
2485         struct inode *inode = obj->oo_inode;
2486         struct osd_access_lock *al;
2487         struct osd_thandle *oh;
2488         int rc = 0, found = 0;
2489         bool grow = false;
2490         ENTRY;
2491
2492         LASSERT(dt_object_exists(dt));
2493         LASSERT(osd_invariant(obj));
2494         LASSERT(inode != NULL);
2495         dquot_initialize(inode);
2496
2497         LASSERT(th);
2498         oh = container_of(th, struct osd_thandle, ot_super);
2499         LASSERT(oh->ot_handle->h_transaction != NULL);
2500
2501         /* we used to skip truncate to current size to
2502          * optimize truncates on OST. with DoM we can
2503          * get attr_set to set specific size (MDS_REINT)
2504          * and then get truncate RPC which essentially
2505          * would be skipped. this is bad.. so, disable
2506          * this optimization on MDS till the client stop
2507          * to sent MDS_REINT (LU-11033) -bzzz
2508          */
2509         if (osd->od_is_ost && i_size_read(inode) == start)
2510                 RETURN(0);
2511
2512         osd_trans_exec_op(env, th, OSD_OT_PUNCH);
2513
2514         spin_lock(&inode->i_lock);
2515         if (i_size_read(inode) < start)
2516                 grow = true;
2517         i_size_write(inode, start);
2518         spin_unlock(&inode->i_lock);
2519         /* if object holds encrypted content, we need to make sure we truncate
2520          * on an encryption unit boundary, or subsequent reads will get
2521          * corrupted content
2522          */
2523         if (obj->oo_lma_flags & LUSTRE_ENCRYPT_FL &&
2524             start & ~LUSTRE_ENCRYPTION_MASK)
2525                 start = (start & LUSTRE_ENCRYPTION_MASK) +
2526                         LUSTRE_ENCRYPTION_UNIT_SIZE;
2527         ll_truncate_pagecache(inode, start);
2528
2529         /* optimize grow case */
2530         if (grow) {
2531                 osd_execute_truncate(obj);
2532                 GOTO(out, rc);
2533         }
2534
2535         inode_lock(inode);
2536         /* add to orphan list to ensure truncate completion
2537          * if this transaction succeed. ldiskfs_truncate()
2538          * will take the inode out of the list
2539          */
2540         rc = ldiskfs_orphan_add(oh->ot_handle, inode);
2541         inode_unlock(inode);
2542         if (rc != 0)
2543                 GOTO(out, rc);
2544
2545         list_for_each_entry(al, &oh->ot_trunc_locks, tl_list) {
2546                 if (obj != al->tl_obj)
2547                         continue;
2548                 LASSERT(al->tl_shared == 0);
2549                 found = 1;
2550                 /* do actual truncate in osd_trans_stop() */
2551                 al->tl_truncate = 1;
2552                 break;
2553         }
2554         LASSERT(found);
2555
2556 out:
2557         RETURN(rc);
2558 }
2559
2560 static int fiemap_check_ranges(struct inode *inode,
2561                                u64 start, u64 len, u64 *new_len)
2562 {
2563         loff_t maxbytes;
2564
2565         *new_len = len;
2566
2567         if (len == 0)
2568                 return -EINVAL;
2569
2570         if (ldiskfs_test_inode_flag(inode, LDISKFS_INODE_EXTENTS))
2571                 maxbytes = inode->i_sb->s_maxbytes;
2572         else
2573                 maxbytes = LDISKFS_SB(inode->i_sb)->s_bitmap_maxbytes;
2574
2575         if (start > maxbytes)
2576                 return -EFBIG;
2577
2578         /*
2579          * Shrink request scope to what the fs can actually handle.
2580          */
2581         if (len > maxbytes || (maxbytes - len) < start)
2582                 *new_len = maxbytes - start;
2583
2584         return 0;
2585 }
2586
2587 /* So that the fiemap access checks can't overflow on 32 bit machines. */
2588 #define FIEMAP_MAX_EXTENTS     (UINT_MAX / sizeof(struct fiemap_extent))
2589
2590 static int osd_fiemap_get(const struct lu_env *env, struct dt_object *dt,
2591                           struct fiemap *fm)
2592 {
2593         struct fiemap_extent_info fieinfo = {0, };
2594         struct inode *inode = osd_dt_obj(dt)->oo_inode;
2595         u64 len;
2596         int rc;
2597
2598         LASSERT(inode);
2599         if (inode->i_op->fiemap == NULL)
2600                 return -EOPNOTSUPP;
2601
2602         if (fm->fm_extent_count > FIEMAP_MAX_EXTENTS)
2603                 return -EINVAL;
2604
2605         rc = fiemap_check_ranges(inode, fm->fm_start, fm->fm_length, &len);
2606         if (rc)
2607                 return rc;
2608
2609         fieinfo.fi_flags = fm->fm_flags;
2610         fieinfo.fi_extents_max = fm->fm_extent_count;
2611         fieinfo.fi_extents_start = fm->fm_extents;
2612
2613         if (fieinfo.fi_flags & FIEMAP_FLAG_SYNC)
2614                 filemap_write_and_wait(inode->i_mapping);
2615
2616         rc = inode->i_op->fiemap(inode, &fieinfo, fm->fm_start, len);
2617         fm->fm_flags = fieinfo.fi_flags;
2618         fm->fm_mapped_extents = fieinfo.fi_extents_mapped;
2619
2620         return rc;
2621 }
2622
2623 static int osd_ladvise(const struct lu_env *env, struct dt_object *dt,
2624                        __u64 start, __u64 end, enum lu_ladvise_type advice)
2625 {
2626         struct osd_object *obj = osd_dt_obj(dt);
2627         int rc = 0;
2628         ENTRY;
2629
2630         switch (advice) {
2631         case LU_LADVISE_DONTNEED:
2632                 if (end)
2633                         invalidate_mapping_pages(obj->oo_inode->i_mapping,
2634                                                  start >> PAGE_SHIFT,
2635                                                  (end - 1) >> PAGE_SHIFT);
2636                 break;
2637         default:
2638                 rc = -ENOTSUPP;
2639                 break;
2640         }
2641
2642         RETURN(rc);
2643 }
2644
2645 static loff_t osd_lseek(const struct lu_env *env, struct dt_object *dt,
2646                         loff_t offset, int whence)
2647 {
2648         struct osd_object *obj = osd_dt_obj(dt);
2649         struct inode *inode = obj->oo_inode;
2650         struct file *file;
2651         loff_t result;
2652
2653         ENTRY;
2654
2655         LASSERT(dt_object_exists(dt));
2656         LASSERT(osd_invariant(obj));
2657         LASSERT(inode);
2658         LASSERT(offset >= 0);
2659
2660         file = osd_quasi_file(env, inode);
2661         result = file->f_op->llseek(file, offset, whence);
2662
2663         /*
2664          * If 'offset' is beyond end of object file then treat it as not error
2665          * but valid case for SEEK_HOLE and return 'offset' as result.
2666          * LOV will decide if it is beyond real end of file or not.
2667          */
2668         if (whence == SEEK_HOLE && result == -ENXIO)
2669                 result = offset;
2670
2671         CDEBUG(D_INFO, "seek %s from %lld: %lld\n", whence == SEEK_HOLE ?
2672                        "hole" : "data", offset, result);
2673         RETURN(result);
2674 }
2675
2676 /*
2677  * in some cases we may need declare methods for objects being created
2678  * e.g., when we create symlink
2679  */
2680 const struct dt_body_operations osd_body_ops_new = {
2681         .dbo_declare_write = osd_declare_write,
2682 };
2683
2684 const struct dt_body_operations osd_body_ops = {
2685         .dbo_read                       = osd_read,
2686         .dbo_declare_write              = osd_declare_write,
2687         .dbo_write                      = osd_write,
2688         .dbo_bufs_get                   = osd_bufs_get,
2689         .dbo_bufs_put                   = osd_bufs_put,
2690         .dbo_write_prep                 = osd_write_prep,
2691         .dbo_declare_write_commit       = osd_declare_write_commit,
2692         .dbo_write_commit               = osd_write_commit,
2693         .dbo_read_prep                  = osd_read_prep,
2694         .dbo_declare_punch              = osd_declare_punch,
2695         .dbo_punch                      = osd_punch,
2696         .dbo_fiemap_get                 = osd_fiemap_get,
2697         .dbo_ladvise                    = osd_ladvise,
2698         .dbo_declare_fallocate          = osd_declare_fallocate,
2699         .dbo_fallocate                  = osd_fallocate,
2700         .dbo_lseek                      = osd_lseek,
2701 };
2702
2703 /**
2704  * Get a truncate lock
2705  *
2706  * In order to take multi-transaction truncate out of main transaction we let
2707  * the caller grab a lock on the object passed. the lock can be shared (for
2708  * writes) and exclusive (for truncate). It's not allowed to mix truncate
2709  * and write in the same transaction handle (do not confuse with big ldiskfs
2710  * transaction containing lots of handles).
2711  * The lock must be taken at declaration.
2712  *
2713  * \param obj           object to lock
2714  * \oh                  transaction
2715  * \shared              shared or exclusive
2716  *
2717  * \retval 0            lock is granted
2718  * \retval -NOMEM       no memory to allocate lock
2719  */
2720 int osd_trunc_lock(struct osd_object *obj, struct osd_thandle *oh, bool shared)
2721 {
2722         struct osd_access_lock *al, *tmp;
2723
2724         LASSERT(obj);
2725         LASSERT(oh);
2726
2727         list_for_each_entry(tmp, &oh->ot_trunc_locks, tl_list) {
2728                 if (tmp->tl_obj != obj)
2729                         continue;
2730                 LASSERT(tmp->tl_shared == shared);
2731                 /* found same lock */
2732                 return 0;
2733         }
2734
2735         OBD_ALLOC_PTR(al);
2736         if (unlikely(al == NULL))
2737                 return -ENOMEM;
2738         al->tl_obj = obj;
2739         al->tl_truncate = false;
2740         if (shared)
2741                 down_read(&obj->oo_ext_idx_sem);
2742         else
2743                 down_write(&obj->oo_ext_idx_sem);
2744         al->tl_shared = shared;
2745         lu_object_get(&obj->oo_dt.do_lu);
2746
2747         list_add(&al->tl_list, &oh->ot_trunc_locks);
2748
2749         return 0;
2750 }
2751
2752 void osd_trunc_unlock_all(const struct lu_env *env, struct list_head *list)
2753 {
2754         struct osd_access_lock *al, *tmp;
2755
2756         list_for_each_entry_safe(al, tmp, list, tl_list) {
2757                 if (al->tl_shared)
2758                         up_read(&al->tl_obj->oo_ext_idx_sem);
2759                 else
2760                         up_write(&al->tl_obj->oo_ext_idx_sem);
2761                 osd_object_put(env, al->tl_obj);
2762                 list_del(&al->tl_list);
2763                 OBD_FREE_PTR(al);
2764         }
2765 }
2766
2767 /* For a partial-page punch, flush punch range to disk immediately */
2768 static void osd_partial_page_flush_punch(struct osd_device *d,
2769                                          struct inode *inode, loff_t start,
2770                                          loff_t end)
2771 {
2772         if (osd_use_page_cache(d)) {
2773                 filemap_fdatawrite_range(inode->i_mapping, start, end);
2774         } else {
2775                 /* Notice we use "wait" version to ensure I/O is complete */
2776                 filemap_write_and_wait_range(inode->i_mapping, start,
2777                                              end);
2778                 invalidate_mapping_pages(inode->i_mapping, start >> PAGE_SHIFT,
2779                                          end >> PAGE_SHIFT);
2780         }
2781 }
2782
2783 /*
2784  * For a partial-page truncate, flush the page to disk immediately to
2785  * avoid data corruption during direct disk write.  b=17397
2786  */
2787 static void osd_partial_page_flush(struct osd_device *d, struct inode *inode,
2788                                    loff_t offset)
2789 {
2790         if (!(offset & ~PAGE_MASK))
2791                 return;
2792
2793         if (osd_use_page_cache(d)) {
2794                 filemap_fdatawrite_range(inode->i_mapping, offset, offset + 1);
2795         } else {
2796                 /* Notice we use "wait" version to ensure I/O is complete */
2797                 filemap_write_and_wait_range(inode->i_mapping, offset,
2798                                              offset + 1);
2799                 invalidate_mapping_pages(inode->i_mapping, offset >> PAGE_SHIFT,
2800                                          offset >> PAGE_SHIFT);
2801         }
2802 }
2803
2804 void osd_execute_truncate(struct osd_object *obj)
2805 {
2806         struct osd_device *d = osd_obj2dev(obj);
2807         struct inode *inode = obj->oo_inode;
2808         __u64 size;
2809
2810         /* simulate crash before (in the middle) of delayed truncate */
2811         if (OBD_FAIL_CHECK(OBD_FAIL_OSD_FAIL_AT_TRUNCATE)) {
2812                 struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
2813                 struct ldiskfs_sb_info *sbi = LDISKFS_SB(inode->i_sb);
2814
2815                 mutex_lock(&sbi->s_orphan_lock);
2816                 list_del_init(&ei->i_orphan);
2817                 mutex_unlock(&sbi->s_orphan_lock);
2818                 return;
2819         }
2820
2821         size = i_size_read(inode);
2822         inode_lock(inode);
2823         /* if object holds encrypted content, we need to make sure we truncate
2824          * on an encryption unit boundary, or block content will get corrupted
2825          */
2826         if (obj->oo_lma_flags & LUSTRE_ENCRYPT_FL &&
2827             size & ~LUSTRE_ENCRYPTION_MASK)
2828                 inode->i_size = (size & LUSTRE_ENCRYPTION_MASK) +
2829                         LUSTRE_ENCRYPTION_UNIT_SIZE;
2830         ldiskfs_truncate(inode);
2831         inode_unlock(inode);
2832         if (inode->i_size != size) {
2833                 spin_lock(&inode->i_lock);
2834                 i_size_write(inode, size);
2835                 LDISKFS_I(inode)->i_disksize = size;
2836                 spin_unlock(&inode->i_lock);
2837                 osd_dirty_inode(inode, I_DIRTY_DATASYNC);
2838         }
2839         osd_partial_page_flush(d, inode, size);
2840 }
2841
2842 void osd_execute_punch(const struct lu_env *env, struct osd_object *obj,
2843                        loff_t start, loff_t end, int mode)
2844 {
2845         struct osd_device *d = osd_obj2dev(obj);
2846         struct inode *inode = obj->oo_inode;
2847         struct file *file = osd_quasi_file(env, inode);
2848
2849         file->f_op->fallocate(file, mode, start, end - start);
2850         osd_partial_page_flush_punch(d, inode, start, end - 1);
2851 }
2852
2853 void osd_process_truncates(const struct lu_env *env, struct list_head *list)
2854 {
2855         struct osd_access_lock *al;
2856
2857         LASSERT(journal_current_handle() == NULL);
2858
2859         list_for_each_entry(al, list, tl_list) {
2860                 if (al->tl_shared)
2861                         continue;
2862                 if (al->tl_truncate)
2863                         osd_execute_truncate(al->tl_obj);
2864                 else if (al->tl_punch)
2865                         osd_execute_punch(env, al->tl_obj, al->tl_start,
2866                                           al->tl_end, al->tl_mode);
2867         }
2868 }