lustre/osd-ldiskfs/osd_io.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2012, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * lustre/osd/osd_io.c
  33  *
  34  * body operations
  35  *
  36  * Author: Nikita Danilov <nikita@clusterfs.com>
  37  * Author: Alex Zhuravlev <bzzz@whamcloud.com>
  38  *
  39  */
  40
  41 /* prerequisite for linux/xattr.h */
  42 #include <linux/types.h>
  43 /* prerequisite for linux/xattr.h */
  44 #include <linux/fs.h>
  45 #include <linux/mm.h>
  46 #include <linux/pagevec.h>
  47
  48 /*
  49  * struct OBD_{ALLOC,FREE}*()
  50  * OBD_FAIL_CHECK
  51  */
  52 #include <obd_support.h>
  53
  54 #include "osd_internal.h"
  55
  56 /* ext_depth() */
  57 #include <ldiskfs/ldiskfs_extents.h>
  58
  59 static inline bool osd_use_page_cache(struct osd_device *d)
  60 {
  61         /* do not use pagecache if write and read caching are disabled */
  62         if (d->od_writethrough_cache + d->od_read_cache == 0)
  63                 return false;
  64         /* use pagecache by default */
  65         return true;
  66 }
  67
  68 static int __osd_init_iobuf(struct osd_device *d, struct osd_iobuf *iobuf,
  69                             int rw, int line, int pages)
  70 {
  71         int blocks, i;
  72
  73         LASSERTF(iobuf->dr_elapsed_valid == 0,
  74                  "iobuf %p, reqs %d, rw %d, line %d\n", iobuf,
  75                  atomic_read(&iobuf->dr_numreqs), iobuf->dr_rw,
  76                  iobuf->dr_init_at);
  77         LASSERT(pages <= PTLRPC_MAX_BRW_PAGES);
  78
  79         init_waitqueue_head(&iobuf->dr_wait);
  80         atomic_set(&iobuf->dr_numreqs, 0);
  81         iobuf->dr_npages = 0;
  82         iobuf->dr_error = 0;
  83         iobuf->dr_dev = d;
  84         iobuf->dr_frags = 0;
  85         iobuf->dr_elapsed = ktime_set(0, 0);
  86         /* must be counted before, so assert */
  87         iobuf->dr_rw = rw;
  88         iobuf->dr_init_at = line;
  89
  90         blocks = pages * (PAGE_SIZE >> osd_sb(d)->s_blocksize_bits);
  91         if (iobuf->dr_bl_buf.lb_len >= blocks * sizeof(iobuf->dr_blocks[0])) {
  92                 LASSERT(iobuf->dr_pg_buf.lb_len >=
  93                         pages * sizeof(iobuf->dr_pages[0]));
  94                 return 0;
  95         }
  96
  97         /* start with 1MB for 4K blocks */
  98         i = 256;
  99         while (i <= PTLRPC_MAX_BRW_PAGES && i < pages)
 100                 i <<= 1;
 101
 102         CDEBUG(D_OTHER, "realloc %u for %u (%u) pages\n",
 103                (unsigned)(pages * sizeof(iobuf->dr_pages[0])), i, pages);
 104         pages = i;
 105         blocks = pages * (PAGE_SIZE >> osd_sb(d)->s_blocksize_bits);
 106         iobuf->dr_max_pages = 0;
 107         CDEBUG(D_OTHER, "realloc %u for %u blocks\n",
 108                (unsigned)(blocks * sizeof(iobuf->dr_blocks[0])), blocks);
 109
 110         lu_buf_realloc(&iobuf->dr_bl_buf, blocks * sizeof(iobuf->dr_blocks[0]));
 111         iobuf->dr_blocks = iobuf->dr_bl_buf.lb_buf;
 112         if (unlikely(iobuf->dr_blocks == NULL))
 113                 return -ENOMEM;
 114
 115         lu_buf_realloc(&iobuf->dr_pg_buf, pages * sizeof(iobuf->dr_pages[0]));
 116         iobuf->dr_pages = iobuf->dr_pg_buf.lb_buf;
 117         if (unlikely(iobuf->dr_pages == NULL))
 118                 return -ENOMEM;
 119
 120         lu_buf_realloc(&iobuf->dr_lnb_buf,
 121                        pages * sizeof(iobuf->dr_lnbs[0]));
 122         iobuf->dr_lnbs = iobuf->dr_lnb_buf.lb_buf;
 123         if (unlikely(iobuf->dr_lnbs == NULL))
 124                 return -ENOMEM;
 125
 126         iobuf->dr_max_pages = pages;
 127
 128         return 0;
 129 }
 130 #define osd_init_iobuf(dev, iobuf, rw, pages) \
 131         __osd_init_iobuf(dev, iobuf, rw, __LINE__, pages)
 132
 133 static void osd_iobuf_add_page(struct osd_iobuf *iobuf,
 134                                struct niobuf_local *lnb)
 135 {
 136         LASSERT(iobuf->dr_npages < iobuf->dr_max_pages);
 137         iobuf->dr_pages[iobuf->dr_npages] = lnb->lnb_page;
 138         iobuf->dr_lnbs[iobuf->dr_npages] = lnb;
 139         iobuf->dr_npages++;
 140 }
 141
 142 void osd_fini_iobuf(struct osd_device *d, struct osd_iobuf *iobuf)
 143 {
 144         int rw = iobuf->dr_rw;
 145
 146         if (iobuf->dr_elapsed_valid) {
 147                 iobuf->dr_elapsed_valid = 0;
 148                 LASSERT(iobuf->dr_dev == d);
 149                 LASSERT(iobuf->dr_frags > 0);
 150                 lprocfs_oh_tally(&d->od_brw_stats.
 151                                  hist[BRW_R_DIO_FRAGS+rw],
 152                                  iobuf->dr_frags);
 153                 lprocfs_oh_tally_log2(&d->od_brw_stats.hist[BRW_R_IO_TIME+rw],
 154                                       ktime_to_ms(iobuf->dr_elapsed));
 155         }
 156 }
 157
 158 #ifdef HAVE_BIO_ENDIO_USES_ONE_ARG
 159 static void dio_complete_routine(struct bio *bio)
 160 {
 161         int error = bio->bi_status;
 162 #else
 163 static void dio_complete_routine(struct bio *bio, int error)
 164 {
 165 #endif
 166         struct osd_iobuf *iobuf = bio->bi_private;
 167         int iter;
 168         struct bio_vec *bvl;
 169
 170         /* CAVEAT EMPTOR: possibly in IRQ context
 171          * DO NOT record procfs stats here!!! */
 172
 173         if (unlikely(iobuf == NULL)) {
 174                 CERROR("***** bio->bi_private is NULL!  This should never "
 175                        "happen.  Normally, I would crash here, but instead I "
 176                        "will dump the bio contents to the console.  Please "
 177                        "report this to <https://jira.whamcloud.com/> , along "
 178                        "with any interesting messages leading up to this point "
 179                        "(like SCSI errors, perhaps).  Because bi_private is "
 180                        "NULL, I can't wake up the thread that initiated this "
 181                        "IO - you will probably have to reboot this node.\n");
 182                 CERROR("bi_next: %p, bi_flags: %lx, " __stringify(bi_opf)
 183                        ": %x, bi_vcnt: %d, bi_idx: %d, bi->size: %d, bi_end_io: %p, bi_cnt: %d, bi_private: %p\n",
 184                        bio->bi_next, (unsigned long)bio->bi_flags,
 185                        (unsigned int)bio->bi_opf, bio->bi_vcnt, bio_idx(bio),
 186                        bio_sectors(bio) << 9, bio->bi_end_io,
 187                        atomic_read(&bio->__bi_cnt),
 188                        bio->bi_private);
 189                 return;
 190         }
 191
 192         /* the check is outside of the cycle for performance reason -bzzz */
 193         if (!bio_data_dir(bio)) {
 194                 bio_for_each_segment_all(bvl, bio, iter) {
 195                         if (likely(error == 0))
 196                                 SetPageUptodate(bvl_to_page(bvl));
 197                         LASSERT(PageLocked(bvl_to_page(bvl)));
 198                 }
 199                 atomic_dec(&iobuf->dr_dev->od_r_in_flight);
 200         } else {
 201                 atomic_dec(&iobuf->dr_dev->od_w_in_flight);
 202         }
 203
 204         /* any real error is good enough -bzzz */
 205         if (error != 0 && iobuf->dr_error == 0)
 206                 iobuf->dr_error = error;
 207
 208         /*
 209          * set dr_elapsed before dr_numreqs turns to 0, otherwise
 210          * it's possible that service thread will see dr_numreqs
 211          * is zero, but dr_elapsed is not set yet, leading to lost
 212          * data in this processing and an assertion in a subsequent
 213          * call to OSD.
 214          */
 215         if (atomic_read(&iobuf->dr_numreqs) == 1) {
 216                 ktime_t now = ktime_get();
 217
 218                 iobuf->dr_elapsed = ktime_sub(now, iobuf->dr_start_time);
 219                 iobuf->dr_elapsed_valid = 1;
 220         }
 221         if (atomic_dec_and_test(&iobuf->dr_numreqs))
 222                 wake_up(&iobuf->dr_wait);
 223
 224         /* Completed bios used to be chained off iobuf->dr_bios and freed in
 225          * filter_clear_dreq().  It was then possible to exhaust the biovec-256
 226          * mempool when serious on-disk fragmentation was encountered,
 227          * deadlocking the OST.  The bios are now released as soon as complete
 228          * so the pool cannot be exhausted while IOs are competing. bug 10076 */
 229         bio_put(bio);
 230 }
 231
 232 static void record_start_io(struct osd_iobuf *iobuf, int size)
 233 {
 234         struct osd_device    *osd = iobuf->dr_dev;
 235         struct obd_histogram *h = osd->od_brw_stats.hist;
 236
 237         iobuf->dr_frags++;
 238         atomic_inc(&iobuf->dr_numreqs);
 239
 240         if (iobuf->dr_rw == 0) {
 241                 atomic_inc(&osd->od_r_in_flight);
 242                 lprocfs_oh_tally(&h[BRW_R_RPC_HIST],
 243                                  atomic_read(&osd->od_r_in_flight));
 244                 lprocfs_oh_tally_log2(&h[BRW_R_DISK_IOSIZE], size);
 245         } else if (iobuf->dr_rw == 1) {
 246                 atomic_inc(&osd->od_w_in_flight);
 247                 lprocfs_oh_tally(&h[BRW_W_RPC_HIST],
 248                                  atomic_read(&osd->od_w_in_flight));
 249                 lprocfs_oh_tally_log2(&h[BRW_W_DISK_IOSIZE], size);
 250         } else {
 251                 LBUG();
 252         }
 253 }
 254
 255 static void osd_submit_bio(int rw, struct bio *bio)
 256 {
 257         LASSERTF(rw == 0 || rw == 1, "%x\n", rw);
 258 #ifdef HAVE_SUBMIT_BIO_2ARGS
 259         submit_bio(rw ? WRITE : READ, bio);
 260 #else
 261         bio->bi_opf |= rw;
 262         submit_bio(bio);
 263 #endif
 264 }
 265
 266 static int can_be_merged(struct bio *bio, sector_t sector)
 267 {
 268         if (bio == NULL)
 269                 return 0;
 270
 271         return bio_end_sector(bio) == sector ? 1 : 0;
 272 }
 273
 274 #if IS_ENABLED(CONFIG_BLK_DEV_INTEGRITY)
 275 /*
 276  * This function will change the data written, thus it should only be
 277  * used when checking data integrity feature
 278  */
 279 static void bio_integrity_fault_inject(struct bio *bio)
 280 {
 281         struct bio_vec *bvec;
 282         int i;
 283         void *kaddr;
 284         char *addr;
 285
 286         bio_for_each_segment_all(bvec, bio, i) {
 287                 struct page *page = bvec->bv_page;
 288
 289                 kaddr = kmap(page);
 290                 addr = kaddr;
 291                 *addr = ~(*addr);
 292                 kunmap(page);
 293                 break;
 294         }
 295 }
 296
 297 static int bio_dif_compare(__u16 *expected_guard_buf, void *bio_prot_buf,
 298                            unsigned int sectors, int tuple_size)
 299 {
 300         __u16 *expected_guard;
 301         __u16 *bio_guard;
 302         int i;
 303
 304         expected_guard = expected_guard_buf;
 305         for (i = 0; i < sectors; i++) {
 306                 bio_guard = (__u16 *)bio_prot_buf;
 307                 if (*bio_guard != *expected_guard) {
 308                         CERROR("unexpected guard tags on sector %d "
 309                                "expected guard %u, bio guard "
 310                                "%u, sectors %u, tuple size %d\n",
 311                                i, *expected_guard, *bio_guard, sectors,
 312                                tuple_size);
 313                         return -EIO;
 314                 }
 315                 expected_guard++;
 316                 bio_prot_buf += tuple_size;
 317         }
 318         return 0;
 319 }
 320
 321 static int osd_bio_integrity_compare(struct bio *bio, struct block_device *bdev,
 322                                      struct osd_iobuf *iobuf, int index)
 323 {
 324         struct blk_integrity *bi = bdev_get_integrity(bdev);
 325         struct bio_integrity_payload *bip = bio->bi_integrity;
 326         struct niobuf_local *lnb;
 327         unsigned short sector_size = blk_integrity_interval(bi);
 328         void *bio_prot_buf = page_address(bip->bip_vec->bv_page) +
 329                 bip->bip_vec->bv_offset;
 330         struct bio_vec *bv;
 331         sector_t sector = bio_start_sector(bio);
 332         unsigned int i, sectors, total;
 333         __u16 *expected_guard;
 334         int rc;
 335
 336         total = 0;
 337         bio_for_each_segment_all(bv, bio, i) {
 338                 lnb = iobuf->dr_lnbs[index];
 339                 expected_guard = lnb->lnb_guards;
 340                 sectors = bv->bv_len / sector_size;
 341                 if (lnb->lnb_guard_rpc) {
 342                         rc = bio_dif_compare(expected_guard, bio_prot_buf,
 343                                              sectors, bi->tuple_size);
 344                         if (rc)
 345                                 return rc;
 346                 }
 347
 348                 sector += sectors;
 349                 bio_prot_buf += sectors * bi->tuple_size;
 350                 total += sectors * bi->tuple_size;
 351                 LASSERT(total <= bip_size(bio->bi_integrity));
 352                 index++;
 353         }
 354         return 0;
 355 }
 356
 357 static int osd_bio_integrity_handle(struct osd_device *osd, struct bio *bio,
 358                                     struct osd_iobuf *iobuf,
 359                                     int start_page_idx, bool fault_inject,
 360                                     bool integrity_enabled)
 361 {
 362         struct super_block *sb = osd_sb(osd);
 363         integrity_gen_fn *generate_fn = NULL;
 364         integrity_vrfy_fn *verify_fn = NULL;
 365         int rc;
 366
 367         ENTRY;
 368
 369         if (!integrity_enabled)
 370                 RETURN(0);
 371
 372         rc = osd_get_integrity_profile(osd, &generate_fn, &verify_fn);
 373         if (rc)
 374                 RETURN(rc);
 375
 376         rc = bio_integrity_prep_fn(bio, generate_fn, verify_fn);
 377         if (rc)
 378                 RETURN(rc);
 379
 380         /* Verify and inject fault only when writing */
 381         if (iobuf->dr_rw == 1) {
 382                 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_OST_INTEGRITY_CMP))) {
 383                         rc = osd_bio_integrity_compare(bio, sb->s_bdev, iobuf,
 384                                                        start_page_idx);
 385                         if (rc)
 386                                 RETURN(rc);
 387                 }
 388
 389                 if (unlikely(fault_inject))
 390                         bio_integrity_fault_inject(bio);
 391         }
 392
 393         RETURN(0);
 394 }
 395
 396 #ifdef HAVE_BIO_INTEGRITY_PREP_FN
 397 #  ifdef HAVE_BIO_ENDIO_USES_ONE_ARG
 398 static void dio_integrity_complete_routine(struct bio *bio)
 399 #  else
 400 static void dio_integrity_complete_routine(struct bio *bio, int error)
 401 #  endif
 402 {
 403         struct osd_bio_private *bio_private = bio->bi_private;
 404
 405         bio->bi_private = bio_private->obp_iobuf;
 406         osd_dio_complete_routine(bio, error);
 407
 408         OBD_FREE_PTR(bio_private);
 409 }
 410 #endif /* HAVE_BIO_INTEGRITY_PREP_FN */
 411 #else  /* !CONFIG_BLK_DEV_INTEGRITY */
 412 #define osd_bio_integrity_handle(osd, bio, iobuf, start_page_idx, \
 413                                  fault_inject, integrity_enabled) 0
 414 #endif /* CONFIG_BLK_DEV_INTEGRITY */
 415
 416 static int osd_bio_init(struct bio *bio, struct osd_iobuf *iobuf,
 417                         bool integrity_enabled, int start_page_idx,
 418                         struct osd_bio_private **pprivate)
 419 {
 420         ENTRY;
 421
 422         *pprivate = NULL;
 423
 424 #ifdef HAVE_BIO_INTEGRITY_PREP_FN
 425         if (integrity_enabled) {
 426                 struct osd_bio_private *bio_private = NULL;
 427
 428                 OBD_ALLOC_GFP(bio_private, sizeof(*bio_private), GFP_NOIO);
 429                 if (bio_private == NULL)
 430                         RETURN(-ENOMEM);
 431                 bio->bi_end_io = dio_integrity_complete_routine;
 432                 bio->bi_private = bio_private;
 433                 bio_private->obp_start_page_idx = start_page_idx;
 434                 bio_private->obp_iobuf = iobuf;
 435                 *pprivate = bio_private;
 436         } else
 437 #endif
 438         {
 439                 bio->bi_end_io = dio_complete_routine;
 440                 bio->bi_private = iobuf;
 441         }
 442
 443         RETURN(0);
 444 }
 445
 446 static int osd_do_bio(struct osd_device *osd, struct inode *inode,
 447                       struct osd_iobuf *iobuf)
 448 {
 449         int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
 450         struct page **pages = iobuf->dr_pages;
 451         int npages = iobuf->dr_npages;
 452         sector_t *blocks = iobuf->dr_blocks;
 453         int total_blocks = npages * blocks_per_page;
 454         struct super_block *sb = inode->i_sb;
 455         int sector_bits = sb->s_blocksize_bits - 9;
 456         unsigned int blocksize = sb->s_blocksize;
 457         struct block_device *bdev = sb->s_bdev;
 458         struct osd_bio_private *bio_private = NULL;
 459         struct bio *bio = NULL;
 460         int bio_start_page_idx;
 461         struct page *page;
 462         unsigned int page_offset;
 463         sector_t sector;
 464         int nblocks;
 465         int block_idx;
 466         int page_idx;
 467         int i;
 468         int rc = 0;
 469         bool fault_inject;
 470         bool integrity_enabled;
 471         DECLARE_PLUG(plug);
 472         ENTRY;
 473
 474         fault_inject = OBD_FAIL_CHECK(OBD_FAIL_OST_INTEGRITY_FAULT);
 475         LASSERT(iobuf->dr_npages == npages);
 476
 477         integrity_enabled = bdev_integrity_enabled(bdev, iobuf->dr_rw);
 478
 479         osd_brw_stats_update(osd, iobuf);
 480         iobuf->dr_start_time = ktime_get();
 481
 482         blk_start_plug(&plug);
 483         for (page_idx = 0, block_idx = 0;
 484              page_idx < npages;
 485              page_idx++, block_idx += blocks_per_page) {
 486
 487                 page = pages[page_idx];
 488                 LASSERT(block_idx + blocks_per_page <= total_blocks);
 489
 490                 for (i = 0, page_offset = 0;
 491                      i < blocks_per_page;
 492                      i += nblocks, page_offset += blocksize * nblocks) {
 493
 494                         nblocks = 1;
 495
 496                         if (blocks[block_idx + i] == 0) {  /* hole */
 497                                 LASSERTF(iobuf->dr_rw == 0,
 498                                          "page_idx %u, block_idx %u, i %u\n",
 499                                          page_idx, block_idx, i);
 500                                 memset(kmap(page) + page_offset, 0, blocksize);
 501                                 kunmap(page);
 502                                 continue;
 503                         }
 504
 505                         sector = (sector_t)blocks[block_idx + i] << sector_bits;
 506
 507                         /* Additional contiguous file blocks? */
 508                         while (i + nblocks < blocks_per_page &&
 509                                (sector + (nblocks << sector_bits)) ==
 510                                ((sector_t)blocks[block_idx + i + nblocks] <<
 511                                 sector_bits))
 512                                 nblocks++;
 513
 514                         if (bio != NULL &&
 515                             can_be_merged(bio, sector) &&
 516                             bio_add_page(bio, page,
 517                                          blocksize * nblocks, page_offset) != 0)
 518                                 continue;       /* added this frag OK */
 519
 520                         if (bio != NULL) {
 521                                 struct request_queue *q = bio_get_queue(bio);
 522                                 unsigned int bi_size = bio_sectors(bio) << 9;
 523
 524                                 /* Dang! I have to fragment this I/O */
 525                                 CDEBUG(D_INODE, "bio++ sz %d vcnt %d(%d) "
 526                                        "sectors %d(%d) psg %d(%d) hsg %d(%d)\n",
 527                                        bi_size, bio->bi_vcnt, bio->bi_max_vecs,
 528                                        bio_sectors(bio),
 529                                        queue_max_sectors(q),
 530                                        bio->bi_phys_segments,
 531                                        queue_max_phys_segments(q),
 532                                        0, queue_max_hw_segments(q));
 533                                 rc = osd_bio_integrity_handle(osd, bio,
 534                                         iobuf, bio_start_page_idx,
 535                                         fault_inject, integrity_enabled);
 536                                 if (rc) {
 537                                         bio_put(bio);
 538                                         goto out;
 539                                 }
 540
 541                                 record_start_io(iobuf, bi_size);
 542                                 osd_submit_bio(iobuf->dr_rw, bio);
 543                         }
 544
 545                         bio_start_page_idx = page_idx;
 546                         /* allocate new bio */
 547                         bio = bio_alloc(GFP_NOIO, min(BIO_MAX_PAGES,
 548                                                       (npages - page_idx) *
 549                                                       blocks_per_page));
 550                         if (bio == NULL) {
 551                                 CERROR("Can't allocate bio %u*%u = %u pages\n",
 552                                        (npages - page_idx), blocks_per_page,
 553                                        (npages - page_idx) * blocks_per_page);
 554                                 rc = -ENOMEM;
 555                                 goto out;
 556                         }
 557
 558                         bio_set_dev(bio, bdev);
 559                         bio_set_sector(bio, sector);
 560                         bio->bi_opf = iobuf->dr_rw ? WRITE : READ;
 561                         rc = osd_bio_init(bio, iobuf, integrity_enabled,
 562                                           bio_start_page_idx, &bio_private);
 563                         if (rc) {
 564                                 bio_put(bio);
 565                                 goto out;
 566                         }
 567
 568                         rc = bio_add_page(bio, page,
 569                                           blocksize * nblocks, page_offset);
 570                         LASSERT(rc != 0);
 571                 }
 572         }
 573
 574         if (bio != NULL) {
 575                 rc = osd_bio_integrity_handle(osd, bio, iobuf,
 576                                               bio_start_page_idx,
 577                                               fault_inject,
 578                                               integrity_enabled);
 579                 if (rc) {
 580                         bio_put(bio);
 581                         goto out;
 582                 }
 583
 584                 record_start_io(iobuf, bio_sectors(bio) << 9);
 585                 osd_submit_bio(iobuf->dr_rw, bio);
 586                 rc = 0;
 587         }
 588
 589 out:
 590         blk_finish_plug(&plug);
 591
 592         /* in order to achieve better IO throughput, we don't wait for writes
 593          * completion here. instead we proceed with transaction commit in
 594          * parallel and wait for IO completion once transaction is stopped
 595          * see osd_trans_stop() for more details -bzzz */
 596         if (iobuf->dr_rw == 0 || fault_inject) {
 597                 wait_event(iobuf->dr_wait,
 598                            atomic_read(&iobuf->dr_numreqs) == 0);
 599                 osd_fini_iobuf(osd, iobuf);
 600         }
 601
 602         if (rc == 0) {
 603                 rc = iobuf->dr_error;
 604         } else {
 605                 if (bio_private)
 606                         OBD_FREE_PTR(bio_private);
 607         }
 608
 609         RETURN(rc);
 610 }
 611
 612 static int osd_map_remote_to_local(loff_t offset, ssize_t len, int *nrpages,
 613                                    struct niobuf_local *lnb)
 614 {
 615         ENTRY;
 616
 617         *nrpages = 0;
 618
 619         while (len > 0) {
 620                 int poff = offset & (PAGE_SIZE - 1);
 621                 int plen = PAGE_SIZE - poff;
 622
 623                 if (plen > len)
 624                         plen = len;
 625                 lnb->lnb_file_offset = offset;
 626                 lnb->lnb_page_offset = poff;
 627                 lnb->lnb_len = plen;
 628                 /* lnb->lnb_flags = rnb->rnb_flags; */
 629                 lnb->lnb_flags = 0;
 630                 lnb->lnb_page = NULL;
 631                 lnb->lnb_rc = 0;
 632                 lnb->lnb_guard_rpc = 0;
 633                 lnb->lnb_guard_disk = 0;
 634
 635                 LASSERTF(plen <= len, "plen %u, len %lld\n", plen,
 636                          (long long) len);
 637                 offset += plen;
 638                 len -= plen;
 639                 lnb++;
 640                 (*nrpages)++;
 641         }
 642
 643         RETURN(0);
 644 }
 645
 646 static struct page *osd_get_page(const struct lu_env *env, struct dt_object *dt,
 647                                  loff_t offset, gfp_t gfp_mask)
 648 {
 649         struct osd_thread_info *oti = osd_oti_get(env);
 650         struct inode *inode = osd_dt_obj(dt)->oo_inode;
 651         struct osd_device *d = osd_obj2dev(osd_dt_obj(dt));
 652         struct page *page;
 653         int cur = oti->oti_dio_pages_used;
 654
 655         LASSERT(inode);
 656
 657         if (osd_use_page_cache(d)) {
 658                 page = find_or_create_page(inode->i_mapping,
 659                                            offset >> PAGE_SHIFT,
 660                                            gfp_mask);
 661
 662                 if (likely(page))
 663                         LASSERT(!test_bit(PG_private_2, &page->flags));
 664                 else
 665                         lprocfs_counter_add(d->od_stats, LPROC_OSD_NO_PAGE, 1);
 666         } else {
 667
 668                 LASSERT(oti->oti_dio_pages);
 669
 670                 if (unlikely(!oti->oti_dio_pages[cur])) {
 671                         LASSERT(cur < PTLRPC_MAX_BRW_PAGES);
 672                         page = alloc_page(gfp_mask);
 673                         if (!page)
 674                                 return NULL;
 675                         oti->oti_dio_pages[cur] = page;
 676                 }
 677
 678                 page = oti->oti_dio_pages[cur];
 679                 LASSERT(!test_bit(PG_private_2, &page->flags));
 680                 set_bit(PG_private_2, &page->flags);
 681                 oti->oti_dio_pages_used++;
 682
 683                 LASSERT(!PageLocked(page));
 684                 lock_page(page);
 685
 686                 LASSERT(!page->mapping);
 687                 LASSERT(!PageWriteback(page));
 688                 ClearPageUptodate(page);
 689
 690                 page->index = offset >> PAGE_SHIFT;
 691         }
 692
 693         return page;
 694 }
 695
 696 /*
 697  * there are following "locks":
 698  * journal_start
 699  * i_mutex
 700  * page lock
 701  *
 702  * osd write path:
 703  *  - lock page(s)
 704  *  - journal_start
 705  *  - truncate_sem
 706  *
 707  * ext4 vmtruncate:
 708  *  - lock pages, unlock
 709  *  - journal_start
 710  *  - lock partial page
 711  *  - i_data_sem
 712  *
 713  */
 714
 715 /**
 716  * Unlock and release pages loaded by osd_bufs_get()
 717  *
 718  * Unlock \a npages pages from \a lnb and drop the refcount on them.
 719  *
 720  * \param env           thread execution environment
 721  * \param dt            dt object undergoing IO (OSD object + methods)
 722  * \param lnb           array of pages undergoing IO
 723  * \param npages        number of pages in \a lnb
 724  *
 725  * \retval 0            always
 726  */
 727 static int osd_bufs_put(const struct lu_env *env, struct dt_object *dt,
 728                         struct niobuf_local *lnb, int npages)
 729 {
 730         struct osd_thread_info *oti = osd_oti_get(env);
 731         struct pagevec pvec;
 732         int i;
 733
 734         ll_pagevec_init(&pvec, 0);
 735
 736         for (i = 0; i < npages; i++) {
 737                 struct page *page = lnb[i].lnb_page;
 738
 739                 if (page == NULL)
 740                         continue;
 741                 LASSERT(PageLocked(page));
 742
 743                 /* if the page isn't cached, then reset uptodate
 744                  * to prevent reuse */
 745                 if (test_bit(PG_private_2, &page->flags)) {
 746                         clear_bit(PG_private_2, &page->flags);
 747                         ClearPageUptodate(page);
 748                         unlock_page(page);
 749                         oti->oti_dio_pages_used--;
 750                 } else {
 751                         unlock_page(page);
 752                         if (pagevec_add(&pvec, page) == 0)
 753                                 pagevec_release(&pvec);
 754                 }
 755                 dt_object_put(env, dt);
 756
 757                 lnb[i].lnb_page = NULL;
 758         }
 759
 760         LASSERTF(oti->oti_dio_pages_used == 0, "%d\n", oti->oti_dio_pages_used);
 761
 762         /* Release any partial pagevec */
 763         pagevec_release(&pvec);
 764
 765         RETURN(0);
 766 }
 767
 768 /**
 769  * Load and lock pages undergoing IO
 770  *
 771  * Pages as described in the \a lnb array are fetched (from disk or cache)
 772  * and locked for IO by the caller.
 773  *
 774  * DLM locking protects us from write and truncate competing for same region,
 775  * but partial-page truncate can leave dirty pages in the cache for ldiskfs.
 776  * It's possible the writeout on a such a page is in progress when we access
 777  * it. It's also possible that during this writeout we put new (partial) data
 778  * into the page, but won't be able to proceed in filter_commitrw_write().
 779  * Therefore, just wait for writeout completion as it should be rare enough.
 780  *
 781  * \param env           thread execution environment
 782  * \param dt            dt object undergoing IO (OSD object + methods)
 783  * \param pos           byte offset of IO start
 784  * \param len           number of bytes of IO
 785  * \param lnb           array of extents undergoing IO
 786  * \param rw            read or write operation, and other flags
 787  * \param capa          capabilities
 788  *
 789  * \retval pages        (zero or more) loaded successfully
 790  * \retval -ENOMEM      on memory/page allocation error
 791  */
 792 static int osd_bufs_get(const struct lu_env *env, struct dt_object *dt,
 793                         loff_t pos, ssize_t len, struct niobuf_local *lnb,
 794                         enum dt_bufs_type rw)
 795 {
 796         struct osd_thread_info *oti = osd_oti_get(env);
 797         struct osd_object *obj = osd_dt_obj(dt);
 798         int npages, i, rc = 0;
 799         gfp_t gfp_mask;
 800
 801         LASSERT(obj->oo_inode);
 802
 803         if (!osd_use_page_cache(osd_obj2dev(obj))) {
 804                 if (unlikely(!oti->oti_dio_pages)) {
 805                         OBD_ALLOC(oti->oti_dio_pages,
 806                                   sizeof(struct page *) * PTLRPC_MAX_BRW_PAGES);
 807                         if (!oti->oti_dio_pages)
 808                                 return -ENOMEM;
 809                 }
 810         }
 811
 812         osd_map_remote_to_local(pos, len, &npages, lnb);
 813
 814         /* this could also try less hard for DT_BUFS_TYPE_READAHEAD pages */
 815         gfp_mask = rw & DT_BUFS_TYPE_LOCAL ? (GFP_NOFS | __GFP_HIGHMEM) :
 816                                              GFP_HIGHUSER;
 817         for (i = 0; i < npages; i++, lnb++) {
 818                 lnb->lnb_page = osd_get_page(env, dt, lnb->lnb_file_offset,
 819                                              gfp_mask);
 820                 if (lnb->lnb_page == NULL)
 821                         GOTO(cleanup, rc = -ENOMEM);
 822
 823                 wait_on_page_writeback(lnb->lnb_page);
 824                 BUG_ON(PageWriteback(lnb->lnb_page));
 825
 826                 lu_object_get(&dt->do_lu);
 827         }
 828
 829         RETURN(i);
 830
 831 cleanup:
 832         if (i > 0)
 833                 osd_bufs_put(env, dt, lnb - i, i);
 834         return rc;
 835 }
 836
 837 #ifndef HAVE_LDISKFS_MAP_BLOCKS
 838
 839 #ifdef HAVE_EXT_PBLOCK /* Name changed to ext4_ext_pblock for kernel 2.6.35 */
 840 #define ldiskfs_ext_pblock(ex) ext_pblock((ex))
 841 #endif
 842
 843 struct bpointers {
 844         sector_t *blocks;
 845         unsigned long start;
 846         int num;
 847         int init_num;
 848         int create;
 849 };
 850
 851 static long ldiskfs_ext_find_goal(struct inode *inode,
 852                                   struct ldiskfs_ext_path *path,
 853                                   unsigned long block, int *aflags)
 854 {
 855         struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
 856         unsigned long bg_start;
 857         unsigned long colour;
 858         int depth;
 859
 860         if (path) {
 861                 struct ldiskfs_extent *ex;
 862                 depth = path->p_depth;
 863
 864                 /* try to predict block placement */
 865                 if ((ex = path[depth].p_ext))
 866                         return ldiskfs_ext_pblock(ex) +
 867                                 (block - le32_to_cpu(ex->ee_block));
 868
 869                 /* it looks index is empty
 870                  * try to find starting from index itself */
 871                 if (path[depth].p_bh)
 872                         return path[depth].p_bh->b_blocknr;
 873         }
 874
 875         /* OK. use inode's group */
 876         bg_start = (ei->i_block_group * LDISKFS_BLOCKS_PER_GROUP(inode->i_sb)) +
 877                 le32_to_cpu(LDISKFS_SB(inode->i_sb)->s_es->s_first_data_block);
 878         colour = (current->pid % 16) *
 879                 (LDISKFS_BLOCKS_PER_GROUP(inode->i_sb) / 16);
 880         return bg_start + colour + block;
 881 }
 882
 883 static unsigned long new_blocks(handle_t *handle, struct inode *inode,
 884                                 struct ldiskfs_ext_path *path,
 885                                 unsigned long block, unsigned long *count,
 886                                 int *err)
 887 {
 888         struct ldiskfs_allocation_request ar;
 889         unsigned long pblock;
 890         int aflags;
 891
 892         /* find neighbour allocated blocks */
 893         ar.lleft = block;
 894         *err = ldiskfs_ext_search_left(inode, path, &ar.lleft, &ar.pleft);
 895         if (*err)
 896                 return 0;
 897         ar.lright = block;
 898         *err = ldiskfs_ext_search_right(inode, path, &ar.lright, &ar.pright);
 899         if (*err)
 900                 return 0;
 901
 902         /* allocate new block */
 903         ar.goal = ldiskfs_ext_find_goal(inode, path, block, &aflags);
 904         ar.inode = inode;
 905         ar.logical = block;
 906         ar.len = *count;
 907         ar.flags = LDISKFS_MB_HINT_DATA;
 908         pblock = ldiskfs_mb_new_blocks(handle, &ar, err);
 909         *count = ar.len;
 910         return pblock;
 911 }
 912
 913 static int ldiskfs_ext_new_extent_cb(struct inode *inode,
 914                                      struct ldiskfs_ext_path *path,
 915                                      struct ldiskfs_ext_cache *cex,
 916 #ifdef HAVE_EXT_PREPARE_CB_EXTENT
 917                                      struct ldiskfs_extent *ex,
 918 #endif
 919                                      void *cbdata)
 920 {
 921         struct bpointers *bp = cbdata;
 922         struct ldiskfs_extent nex;
 923         unsigned long pblock = 0;
 924         unsigned long tgen;
 925         int err, i;
 926         unsigned long count;
 927         handle_t *handle;
 928
 929 #ifdef LDISKFS_EXT_CACHE_EXTENT /* until kernel 2.6.37 */
 930         if (cex->ec_type == LDISKFS_EXT_CACHE_EXTENT) {
 931 #else
 932         if ((cex->ec_len != 0) && (cex->ec_start != 0)) {
 933 #endif
 934                 err = EXT_CONTINUE;
 935                 goto map;
 936         }
 937
 938         if (bp->create == 0) {
 939                 i = 0;
 940                 if (cex->ec_block < bp->start)
 941                         i = bp->start - cex->ec_block;
 942                 if (i >= cex->ec_len)
 943                         CERROR("nothing to do?! i = %d, e_num = %u\n",
 944                                         i, cex->ec_len);
 945                 for (; i < cex->ec_len && bp->num; i++) {
 946                         *(bp->blocks) = 0;
 947                         bp->blocks++;
 948                         bp->num--;
 949                         bp->start++;
 950                 }
 951
 952                 return EXT_CONTINUE;
 953         }
 954
 955         tgen = LDISKFS_I(inode)->i_ext_generation;
 956         count = ldiskfs_ext_calc_credits_for_insert(inode, path);
 957
 958         handle = osd_journal_start(inode, LDISKFS_HT_MISC,
 959                                    count + LDISKFS_ALLOC_NEEDED + 1);
 960         if (IS_ERR(handle)) {
 961                 return PTR_ERR(handle);
 962         }
 963
 964         if (tgen != LDISKFS_I(inode)->i_ext_generation) {
 965                 /* the tree has changed. so path can be invalid at moment */
 966                 ldiskfs_journal_stop(handle);
 967                 return EXT_REPEAT;
 968         }
 969
 970         /* In 2.6.32 kernel, ldiskfs_ext_walk_space()'s callback func is not
 971          * protected by i_data_sem as whole. so we patch it to store
 972          * generation to path and now verify the tree hasn't changed */
 973         down_write((&LDISKFS_I(inode)->i_data_sem));
 974
 975         /* validate extent, make sure the extent tree does not changed */
 976         if (LDISKFS_I(inode)->i_ext_generation != path[0].p_generation) {
 977                 /* cex is invalid, try again */
 978                 up_write(&LDISKFS_I(inode)->i_data_sem);
 979                 ldiskfs_journal_stop(handle);
 980                 return EXT_REPEAT;
 981         }
 982
 983         count = cex->ec_len;
 984         pblock = new_blocks(handle, inode, path, cex->ec_block, &count, &err);
 985         if (!pblock)
 986                 goto out;
 987         BUG_ON(count > cex->ec_len);
 988
 989         /* insert new extent */
 990         nex.ee_block = cpu_to_le32(cex->ec_block);
 991         ldiskfs_ext_store_pblock(&nex, pblock);
 992         nex.ee_len = cpu_to_le16(count);
 993         err = ldiskfs_ext_insert_extent(handle, inode, path, &nex, 0);
 994         if (err) {
 995                 /* free data blocks we just allocated */
 996                 /* not a good idea to call discard here directly,
 997                  * but otherwise we'd need to call it every free() */
 998                 ldiskfs_discard_preallocations(inode);
 999 #ifdef HAVE_EXT_FREE_BLOCK_WITH_BUFFER_HEAD /* Introduced in 2.6.32-rc7 */
1000                 ldiskfs_free_blocks(handle, inode, NULL,
1001                                     ldiskfs_ext_pblock(&nex),
1002                                     le16_to_cpu(nex.ee_len), 0);
1003 #else
1004                 ldiskfs_free_blocks(handle, inode, ldiskfs_ext_pblock(&nex),
1005                                     le16_to_cpu(nex.ee_len), 0);
1006 #endif
1007                 goto out;
1008         }
1009
1010         /*
1011          * Putting len of the actual extent we just inserted,
1012          * we are asking ldiskfs_ext_walk_space() to continue
1013          * scaning after that block
1014          */
1015         cex->ec_len = le16_to_cpu(nex.ee_len);
1016         cex->ec_start = ldiskfs_ext_pblock(&nex);
1017         BUG_ON(le16_to_cpu(nex.ee_len) == 0);
1018         BUG_ON(le32_to_cpu(nex.ee_block) != cex->ec_block);
1019
1020 out:
1021         up_write((&LDISKFS_I(inode)->i_data_sem));
1022         ldiskfs_journal_stop(handle);
1023 map:
1024         if (err >= 0) {
1025                 /* map blocks */
1026                 if (bp->num == 0) {
1027                         CERROR("hmm. why do we find this extent?\n");
1028                         CERROR("initial space: %lu:%u\n",
1029                                 bp->start, bp->init_num);
1030 #ifdef LDISKFS_EXT_CACHE_EXTENT /* until kernel 2.6.37 */
1031                         CERROR("current extent: %u/%u/%llu %d\n",
1032                                 cex->ec_block, cex->ec_len,
1033                                 (unsigned long long)cex->ec_start,
1034                                 cex->ec_type);
1035 #else
1036                         CERROR("current extent: %u/%u/%llu\n",
1037                                 cex->ec_block, cex->ec_len,
1038                                 (unsigned long long)cex->ec_start);
1039 #endif
1040                 }
1041                 i = 0;
1042                 if (cex->ec_block < bp->start)
1043                         i = bp->start - cex->ec_block;
1044                 if (i >= cex->ec_len)
1045                         CERROR("nothing to do?! i = %d, e_num = %u\n",
1046                                         i, cex->ec_len);
1047                 for (; i < cex->ec_len && bp->num; i++) {
1048                         *(bp->blocks) = cex->ec_start + i;
1049                         /* unmap any underlying metadata from
1050                          * the block device mapping.  b=6998.
1051                          */
1052                         if (pblock != 0)
1053                                 clean_bdev_aliases(inode->i_sb->s_bdev,
1054                                                    *(bp->blocks), 1);
1055                         bp->blocks++;
1056                         bp->num--;
1057                         bp->start++;
1058                 }
1059         }
1060         return err;
1061 }
1062
1063 static int osd_ldiskfs_map_nblocks(struct inode *inode, unsigned long index,
1064                                    int clen, sector_t *blocks, int create)
1065 {
1066         int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
1067         struct bpointers bp;
1068         int err;
1069
1070         if (index + clen >= inode->i_sb->s_maxbytes >> PAGE_SHIFT)
1071                 return -EFBIG;
1072
1073         bp.blocks = blocks;
1074         bp.start = index * blocks_per_page;
1075         bp.init_num = bp.num = clen * blocks_per_page;
1076         bp.create = create;
1077
1078         CDEBUG(D_OTHER, "blocks %lu-%lu requested for inode %u\n",
1079                bp.start, bp.start + bp.num - 1, (unsigned)inode->i_ino);
1080
1081         err = ldiskfs_ext_walk_space(inode, bp.start, bp.num,
1082                                      ldiskfs_ext_new_extent_cb, &bp);
1083         ldiskfs_ext_invalidate_cache(inode);
1084
1085         return err;
1086 }
1087
1088 static int osd_ldiskfs_map_bm_inode_pages(struct inode *inode,
1089                                           struct page **page, int pages,
1090                                           sector_t *blocks, int create)
1091 {
1092         int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
1093         pgoff_t bitmap_max_page_index;
1094         sector_t *b;
1095         int rc = 0, i;
1096
1097         bitmap_max_page_index = LDISKFS_SB(inode->i_sb)->s_bitmap_maxbytes >>
1098                                 PAGE_SHIFT;
1099         for (i = 0, b = blocks; i < pages; i++, page++) {
1100                 if ((*page)->index + 1 >= bitmap_max_page_index) {
1101                         rc = -EFBIG;
1102                         break;
1103                 }
1104                 rc = ldiskfs_map_inode_page(inode, *page, b, create);
1105                 if (rc) {
1106                         CERROR("ino %lu, blk %llu create %d: rc %d\n",
1107                                inode->i_ino,
1108                                (unsigned long long)*b, create, rc);
1109                         break;
1110                 }
1111                 b += blocks_per_page;
1112         }
1113         return rc;
1114 }
1115
1116 static int osd_ldiskfs_map_ext_inode_pages(struct inode *inode,
1117                                            struct page **page,
1118                                            int pages, sector_t *blocks,
1119                                            int create)
1120 {
1121         int rc = 0, i = 0, clen = 0;
1122         struct page *fp = NULL;
1123
1124         CDEBUG(D_OTHER, "inode %lu: map %d pages from %lu\n",
1125                 inode->i_ino, pages, (*page)->index);
1126
1127         /* pages are sorted already. so, we just have to find
1128          * contig. space and process them properly */
1129         while (i < pages) {
1130                 if (fp == NULL) {
1131                         /* start new extent */
1132                         fp = *page++;
1133                         clen = 1;
1134                         i++;
1135                         continue;
1136                 } else if (fp->index + clen == (*page)->index) {
1137                         /* continue the extent */
1138                         page++;
1139                         clen++;
1140                         i++;
1141                         continue;
1142                 }
1143
1144                 /* process found extent */
1145                 rc = osd_ldiskfs_map_nblocks(inode, fp->index, clen,
1146                                              blocks, create);
1147                 if (rc)
1148                         GOTO(cleanup, rc);
1149
1150                 /* look for next extent */
1151                 fp = NULL;
1152                 blocks += clen * (PAGE_SIZE >> inode->i_blkbits);
1153         }
1154
1155         if (fp)
1156                 rc = osd_ldiskfs_map_nblocks(inode, fp->index, clen,
1157                                              blocks, create);
1158
1159 cleanup:
1160         return rc;
1161 }
1162
1163 static int osd_ldiskfs_map_inode_pages(struct inode *inode, struct page **page,
1164                                        int pages, sector_t *blocks,
1165                                        int create)
1166 {
1167         int rc;
1168
1169         if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL) {
1170                 rc = osd_ldiskfs_map_ext_inode_pages(inode, page, pages,
1171                                                      blocks, create);
1172                 return rc;
1173         }
1174         rc = osd_ldiskfs_map_bm_inode_pages(inode, page, pages, blocks, create);
1175
1176         return rc;
1177 }
1178 #else
1179 static int osd_ldiskfs_map_inode_pages(struct inode *inode, struct page **page,
1180                                        int pages, sector_t *blocks,
1181                                        int create)
1182 {
1183         int blocks_per_page = PAGE_SIZE >> inode->i_blkbits;
1184         int rc = 0, i = 0;
1185         struct page *fp = NULL;
1186         int clen = 0;
1187         pgoff_t max_page_index;
1188         handle_t *handle = NULL;
1189
1190         max_page_index = inode->i_sb->s_maxbytes >> PAGE_SHIFT;
1191
1192         CDEBUG(D_OTHER, "inode %lu: map %d pages from %lu\n",
1193                 inode->i_ino, pages, (*page)->index);
1194
1195         if (create) {
1196                 create = LDISKFS_GET_BLOCKS_CREATE;
1197                 handle = ldiskfs_journal_current_handle();
1198                 LASSERT(handle != NULL);
1199                 rc = osd_attach_jinode(inode);
1200                 if (rc)
1201                         return rc;
1202         }
1203         /* pages are sorted already. so, we just have to find
1204          * contig. space and process them properly */
1205         while (i < pages) {
1206                 long blen, total = 0;
1207                 struct ldiskfs_map_blocks map = { 0 };
1208
1209                 if (fp == NULL) { /* start new extent */
1210                         fp = *page++;
1211                         clen = 1;
1212                         if (++i != pages)
1213                                 continue;
1214                 } else if (fp->index + clen == (*page)->index) {
1215                         /* continue the extent */
1216                         page++;
1217                         clen++;
1218                         if (++i != pages)
1219                                 continue;
1220                 }
1221                 if (fp->index + clen >= max_page_index)
1222                         GOTO(cleanup, rc = -EFBIG);
1223                 /* process found extent */
1224                 map.m_lblk = fp->index * blocks_per_page;
1225                 map.m_len = blen = clen * blocks_per_page;
1226 cont_map:
1227                 rc = ldiskfs_map_blocks(handle, inode, &map, create);
1228                 if (rc >= 0) {
1229                         int c = 0;
1230                         for (; total < blen && c < map.m_len; c++, total++) {
1231                                 if (rc == 0) {
1232                                         *(blocks + total) = 0;
1233                                         total++;
1234                                         break;
1235                                 } else {
1236                                         *(blocks + total) = map.m_pblk + c;
1237                                         /* unmap any possible underlying
1238                                          * metadata from the block device
1239                                          * mapping.  bug 6998. */
1240                                         if ((map.m_flags & LDISKFS_MAP_NEW) &&
1241                                             create)
1242                                                 clean_bdev_aliases(
1243                                                         inode->i_sb->s_bdev,
1244                                                         map.m_pblk + c, 1);
1245                                 }
1246                         }
1247                         rc = 0;
1248                 }
1249                 if (rc == 0 && total < blen) {
1250                         map.m_lblk = fp->index * blocks_per_page + total;
1251                         map.m_len = blen - total;
1252                         goto cont_map;
1253                 }
1254                 if (rc != 0)
1255                         GOTO(cleanup, rc);
1256
1257                 /* look for next extent */
1258                 fp = NULL;
1259                 blocks += blocks_per_page * clen;
1260         }
1261 cleanup:
1262         return rc;
1263 }
1264 #endif /* HAVE_LDISKFS_MAP_BLOCKS */
1265
1266 static int osd_write_prep(const struct lu_env *env, struct dt_object *dt,
1267                           struct niobuf_local *lnb, int npages)
1268 {
1269         struct osd_thread_info *oti   = osd_oti_get(env);
1270         struct osd_iobuf       *iobuf = &oti->oti_iobuf;
1271         struct inode           *inode = osd_dt_obj(dt)->oo_inode;
1272         struct osd_device      *osd   = osd_obj2dev(osd_dt_obj(dt));
1273         ktime_t start;
1274         ktime_t end;
1275         s64 timediff;
1276         ssize_t                 isize;
1277         __s64                   maxidx;
1278         int                     rc = 0;
1279         int                     i;
1280         int                     cache = 0;
1281
1282         LASSERT(inode);
1283
1284         rc = osd_init_iobuf(osd, iobuf, 0, npages);
1285         if (unlikely(rc != 0))
1286                 RETURN(rc);
1287
1288         isize = i_size_read(inode);
1289         maxidx = ((isize + PAGE_SIZE - 1) >> PAGE_SHIFT) - 1;
1290
1291         if (osd->od_writethrough_cache)
1292                 cache = 1;
1293         if (isize > osd->od_readcache_max_filesize)
1294                 cache = 0;
1295
1296         start = ktime_get();
1297         for (i = 0; i < npages; i++) {
1298
1299                 if (cache == 0)
1300                         generic_error_remove_page(inode->i_mapping,
1301                                                   lnb[i].lnb_page);
1302
1303                 /*
1304                  * till commit the content of the page is undefined
1305                  * we'll set it uptodate once bulk is done. otherwise
1306                  * subsequent reads can access non-stable data
1307                  */
1308                 ClearPageUptodate(lnb[i].lnb_page);
1309
1310                 if (lnb[i].lnb_len == PAGE_SIZE)
1311                         continue;
1312
1313                 if (maxidx >= lnb[i].lnb_page->index) {
1314                         osd_iobuf_add_page(iobuf, &lnb[i]);
1315                 } else {
1316                         long off;
1317                         char *p = kmap(lnb[i].lnb_page);
1318
1319                         off = lnb[i].lnb_page_offset;
1320                         if (off)
1321                                 memset(p, 0, off);
1322                         off = (lnb[i].lnb_page_offset + lnb[i].lnb_len) &
1323                               ~PAGE_MASK;
1324                         if (off)
1325                                 memset(p + off, 0, PAGE_SIZE - off);
1326                         kunmap(lnb[i].lnb_page);
1327                 }
1328         }
1329         end = ktime_get();
1330         timediff = ktime_us_delta(end, start);
1331         lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
1332
1333         if (iobuf->dr_npages) {
1334                 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1335                                                  iobuf->dr_npages,
1336                                                  iobuf->dr_blocks, 0);
1337                 if (likely(rc == 0)) {
1338                         rc = osd_do_bio(osd, inode, iobuf);
1339                         /* do IO stats for preparation reads */
1340                         osd_fini_iobuf(osd, iobuf);
1341                 }
1342         }
1343         RETURN(rc);
1344 }
1345
1346 struct osd_fextent {
1347         sector_t        start;
1348         sector_t        end;
1349         unsigned int    mapped:1;
1350 };
1351
1352 static int osd_is_mapped(struct dt_object *dt, __u64 offset,
1353                          struct osd_fextent *cached_extent)
1354 {
1355         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1356         sector_t block = offset >> inode->i_blkbits;
1357         sector_t start;
1358         struct fiemap_extent_info fei = { 0 };
1359         struct fiemap_extent fe = { 0 };
1360         mm_segment_t saved_fs;
1361         int rc;
1362
1363         if (block >= cached_extent->start && block < cached_extent->end)
1364                 return cached_extent->mapped;
1365
1366         if (i_size_read(inode) == 0)
1367                 return 0;
1368
1369         /* Beyond EOF, must not be mapped */
1370         if (((i_size_read(inode) - 1) >> inode->i_blkbits) < block)
1371                 return 0;
1372
1373         fei.fi_extents_max = 1;
1374         fei.fi_extents_start = &fe;
1375
1376         saved_fs = get_fs();
1377         set_fs(KERNEL_DS);
1378         rc = inode->i_op->fiemap(inode, &fei, offset, FIEMAP_MAX_OFFSET-offset);
1379         set_fs(saved_fs);
1380         if (rc != 0)
1381                 return 0;
1382
1383         start = fe.fe_logical >> inode->i_blkbits;
1384
1385         if (start > block) {
1386                 cached_extent->start = block;
1387                 cached_extent->end = start;
1388                 cached_extent->mapped = 0;
1389         } else {
1390                 cached_extent->start = start;
1391                 cached_extent->end = (fe.fe_logical + fe.fe_length) >>
1392                                       inode->i_blkbits;
1393                 cached_extent->mapped = 1;
1394         }
1395
1396         return cached_extent->mapped;
1397 }
1398
1399 static int osd_declare_write_commit(const struct lu_env *env,
1400                                     struct dt_object *dt,
1401                                     struct niobuf_local *lnb, int npages,
1402                                     struct thandle *handle)
1403 {
1404         const struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1405         struct inode            *inode = osd_dt_obj(dt)->oo_inode;
1406         struct osd_thandle      *oh;
1407         int                     extents = 1;
1408         int                     depth;
1409         int                     i;
1410         int                     newblocks;
1411         int                     rc = 0;
1412         int                     flags = 0;
1413         int                     credits = 0;
1414         long long               quota_space = 0;
1415         struct osd_fextent      extent = { 0 };
1416         enum osd_qid_declare_flags declare_flags = OSD_QID_BLK;
1417         ENTRY;
1418
1419         LASSERT(handle != NULL);
1420         oh = container_of0(handle, struct osd_thandle, ot_super);
1421         LASSERT(oh->ot_handle == NULL);
1422
1423         newblocks = npages;
1424
1425         /* calculate number of extents (probably better to pass nb) */
1426         for (i = 0; i < npages; i++) {
1427                 if (i && lnb[i].lnb_file_offset !=
1428                     lnb[i - 1].lnb_file_offset + lnb[i - 1].lnb_len)
1429                         extents++;
1430
1431                 if (osd_is_mapped(dt, lnb[i].lnb_file_offset, &extent))
1432                         lnb[i].lnb_flags |= OBD_BRW_MAPPED;
1433                 else
1434                         quota_space += PAGE_SIZE;
1435
1436                 /* ignore quota for the whole request if any page is from
1437                  * client cache or written by root.
1438                  *
1439                  * XXX once we drop the 1.8 client support, the checking
1440                  * for whether page is from cache can be simplified as:
1441                  * !(lnb[i].flags & OBD_BRW_SYNC)
1442                  *
1443                  * XXX we could handle this on per-lnb basis as done by
1444                  * grant. */
1445                 if ((lnb[i].lnb_flags & OBD_BRW_NOQUOTA) ||
1446                     (lnb[i].lnb_flags & (OBD_BRW_FROM_GRANT | OBD_BRW_SYNC)) ==
1447                     OBD_BRW_FROM_GRANT)
1448                         declare_flags |= OSD_QID_FORCE;
1449         }
1450
1451         /*
1452          * each extent can go into new leaf causing a split
1453          * 5 is max tree depth: inode + 4 index blocks
1454          * with blockmaps, depth is 3 at most
1455          */
1456         if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL) {
1457                 /*
1458                  * many concurrent threads may grow tree by the time
1459                  * our transaction starts. so, consider 2 is a min depth
1460                  */
1461                 depth = ext_depth(inode);
1462                 depth = max(depth, 1) + 1;
1463                 newblocks += depth;
1464                 credits++; /* inode */
1465                 credits += depth * 2 * extents;
1466         } else {
1467                 depth = 3;
1468                 newblocks += depth;
1469                 credits++; /* inode */
1470                 credits += depth * extents;
1471         }
1472
1473         /* quota space for metadata blocks */
1474         quota_space += depth * extents * LDISKFS_BLOCK_SIZE(osd_sb(osd));
1475
1476         /* quota space should be reported in 1K blocks */
1477         quota_space = toqb(quota_space);
1478
1479         /* each new block can go in different group (bitmap + gd) */
1480
1481         /* we can't dirty more bitmap blocks than exist */
1482         if (newblocks > LDISKFS_SB(osd_sb(osd))->s_groups_count)
1483                 credits += LDISKFS_SB(osd_sb(osd))->s_groups_count;
1484         else
1485                 credits += newblocks;
1486
1487         /* we can't dirty more gd blocks than exist */
1488         if (newblocks > LDISKFS_SB(osd_sb(osd))->s_gdb_count)
1489                 credits += LDISKFS_SB(osd_sb(osd))->s_gdb_count;
1490         else
1491                 credits += newblocks;
1492
1493         osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
1494
1495         /* make sure the over quota flags were not set */
1496         lnb[0].lnb_flags &= ~OBD_BRW_OVER_ALLQUOTA;
1497
1498         rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
1499                                    i_projid_read(inode), quota_space, oh,
1500                                    osd_dt_obj(dt), &flags, declare_flags);
1501
1502         /* we need only to store the overquota flags in the first lnb for
1503          * now, once we support multiple objects BRW, this code needs be
1504          * revised. */
1505         if (flags & QUOTA_FL_OVER_USRQUOTA)
1506                 lnb[0].lnb_flags |= OBD_BRW_OVER_USRQUOTA;
1507         if (flags & QUOTA_FL_OVER_GRPQUOTA)
1508                 lnb[0].lnb_flags |= OBD_BRW_OVER_GRPQUOTA;
1509         if (flags & QUOTA_FL_OVER_PRJQUOTA)
1510                 lnb[0].lnb_flags |= OBD_BRW_OVER_PRJQUOTA;
1511
1512         if (rc == 0)
1513                 rc = osd_trunc_lock(osd_dt_obj(dt), oh, true);
1514
1515         RETURN(rc);
1516 }
1517
1518 /* Check if a block is allocated or not */
1519 static int osd_write_commit(const struct lu_env *env, struct dt_object *dt,
1520                             struct niobuf_local *lnb, int npages,
1521                             struct thandle *thandle)
1522 {
1523         struct osd_thread_info *oti = osd_oti_get(env);
1524         struct osd_iobuf *iobuf = &oti->oti_iobuf;
1525         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1526         struct osd_device  *osd = osd_obj2dev(osd_dt_obj(dt));
1527         loff_t isize;
1528         int rc = 0, i;
1529
1530         LASSERT(inode);
1531
1532         rc = osd_init_iobuf(osd, iobuf, 1, npages);
1533         if (unlikely(rc != 0))
1534                 RETURN(rc);
1535
1536         isize = i_size_read(inode);
1537         ll_vfs_dq_init(inode);
1538
1539         for (i = 0; i < npages; i++) {
1540                 if (lnb[i].lnb_rc == -ENOSPC &&
1541                     (lnb[i].lnb_flags & OBD_BRW_MAPPED)) {
1542                         /* Allow the write to proceed if overwriting an
1543                          * existing block */
1544                         lnb[i].lnb_rc = 0;
1545                 }
1546
1547                 if (lnb[i].lnb_rc) { /* ENOSPC, network RPC error, etc. */
1548                         CDEBUG(D_INODE, "Skipping [%d] == %d\n", i,
1549                                lnb[i].lnb_rc);
1550                         LASSERT(lnb[i].lnb_page);
1551                         generic_error_remove_page(inode->i_mapping,
1552                                                   lnb[i].lnb_page);
1553                         continue;
1554                 }
1555
1556                 LASSERT(PageLocked(lnb[i].lnb_page));
1557                 LASSERT(!PageWriteback(lnb[i].lnb_page));
1558
1559                 if (lnb[i].lnb_file_offset + lnb[i].lnb_len > isize)
1560                         isize = lnb[i].lnb_file_offset + lnb[i].lnb_len;
1561
1562                 /*
1563                  * Since write and truncate are serialized by oo_sem, even
1564                  * partial-page truncate should not leave dirty pages in the
1565                  * page cache.
1566                  */
1567                 LASSERT(!PageDirty(lnb[i].lnb_page));
1568
1569                 SetPageUptodate(lnb[i].lnb_page);
1570
1571                 osd_iobuf_add_page(iobuf, &lnb[i]);
1572         }
1573
1574         osd_trans_exec_op(env, thandle, OSD_OT_WRITE);
1575
1576         if (OBD_FAIL_CHECK(OBD_FAIL_OST_MAPBLK_ENOSPC)) {
1577                 rc = -ENOSPC;
1578         } else if (iobuf->dr_npages > 0) {
1579                 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1580                                                  iobuf->dr_npages,
1581                                                  iobuf->dr_blocks, 1);
1582         } else {
1583                 /* no pages to write, no transno is needed */
1584                 thandle->th_local = 1;
1585         }
1586
1587         if (likely(rc == 0)) {
1588                 spin_lock(&inode->i_lock);
1589                 if (isize > i_size_read(inode)) {
1590                         i_size_write(inode, isize);
1591                         LDISKFS_I(inode)->i_disksize = isize;
1592                         spin_unlock(&inode->i_lock);
1593                         ll_dirty_inode(inode, I_DIRTY_DATASYNC);
1594                 } else {
1595                         spin_unlock(&inode->i_lock);
1596                 }
1597
1598                 rc = osd_do_bio(osd, inode, iobuf);
1599                 /* we don't do stats here as in read path because
1600                  * write is async: we'll do this in osd_put_bufs() */
1601         } else {
1602                 osd_fini_iobuf(osd, iobuf);
1603         }
1604
1605         osd_trans_exec_check(env, thandle, OSD_OT_WRITE);
1606
1607         if (unlikely(rc != 0)) {
1608                 /* if write fails, we should drop pages from the cache */
1609                 for (i = 0; i < npages; i++) {
1610                         if (lnb[i].lnb_page == NULL)
1611                                 continue;
1612                         LASSERT(PageLocked(lnb[i].lnb_page));
1613                         generic_error_remove_page(inode->i_mapping,
1614                                                   lnb[i].lnb_page);
1615                 }
1616         }
1617
1618         RETURN(rc);
1619 }
1620
1621 static int osd_read_prep(const struct lu_env *env, struct dt_object *dt,
1622                          struct niobuf_local *lnb, int npages)
1623 {
1624         struct osd_thread_info *oti = osd_oti_get(env);
1625         struct osd_iobuf *iobuf = &oti->oti_iobuf;
1626         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1627         struct osd_device *osd = osd_obj2dev(osd_dt_obj(dt));
1628         int rc = 0, i, cache = 0, cache_hits = 0, cache_misses = 0;
1629         ktime_t start, end;
1630         s64 timediff;
1631         loff_t isize;
1632
1633         LASSERT(inode);
1634
1635         rc = osd_init_iobuf(osd, iobuf, 0, npages);
1636         if (unlikely(rc != 0))
1637                 RETURN(rc);
1638
1639         isize = i_size_read(inode);
1640
1641         if (osd->od_read_cache)
1642                 cache = 1;
1643         if (isize > osd->od_readcache_max_filesize)
1644                 cache = 0;
1645
1646         start = ktime_get();
1647         for (i = 0; i < npages; i++) {
1648
1649                 if (isize <= lnb[i].lnb_file_offset)
1650                         /* If there's no more data, abort early.
1651                          * lnb->lnb_rc == 0, so it's easy to detect later. */
1652                         break;
1653
1654                 if (isize < lnb[i].lnb_file_offset + lnb[i].lnb_len)
1655                         lnb[i].lnb_rc = isize - lnb[i].lnb_file_offset;
1656                 else
1657                         lnb[i].lnb_rc = lnb[i].lnb_len;
1658
1659                 /* Bypass disk read if fail_loc is set properly */
1660                 if (OBD_FAIL_CHECK(OBD_FAIL_OST_FAKE_RW))
1661                         SetPageUptodate(lnb[i].lnb_page);
1662
1663                 if (PageUptodate(lnb[i].lnb_page)) {
1664                         cache_hits++;
1665                 } else {
1666                         cache_misses++;
1667                         osd_iobuf_add_page(iobuf, &lnb[i]);
1668                 }
1669
1670                 if (cache == 0)
1671                         generic_error_remove_page(inode->i_mapping,
1672                                                   lnb[i].lnb_page);
1673         }
1674         end = ktime_get();
1675         timediff = ktime_us_delta(end, start);
1676         lprocfs_counter_add(osd->od_stats, LPROC_OSD_GET_PAGE, timediff);
1677
1678         if (cache_hits != 0)
1679                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_HIT,
1680                                     cache_hits);
1681         if (cache_misses != 0)
1682                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_MISS,
1683                                     cache_misses);
1684         if (cache_hits + cache_misses != 0)
1685                 lprocfs_counter_add(osd->od_stats, LPROC_OSD_CACHE_ACCESS,
1686                                     cache_hits + cache_misses);
1687
1688         if (iobuf->dr_npages) {
1689                 rc = osd_ldiskfs_map_inode_pages(inode, iobuf->dr_pages,
1690                                                  iobuf->dr_npages,
1691                                                  iobuf->dr_blocks, 0);
1692                 rc = osd_do_bio(osd, inode, iobuf);
1693
1694                 /* IO stats will be done in osd_bufs_put() */
1695         }
1696
1697         RETURN(rc);
1698 }
1699
1700 /*
1701  * XXX: Another layering violation for now.
1702  *
1703  * We don't want to use ->f_op->read methods, because generic file write
1704  *
1705  *         - serializes on ->i_sem, and
1706  *
1707  *         - does a lot of extra work like balance_dirty_pages(),
1708  *
1709  * which doesn't work for globally shared files like /last_rcvd.
1710  */
1711 static int osd_ldiskfs_readlink(struct inode *inode, char *buffer, int buflen)
1712 {
1713         struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
1714
1715         memcpy(buffer, (char *)ei->i_data, buflen);
1716
1717         return  buflen;
1718 }
1719
1720 int osd_ldiskfs_read(struct inode *inode, void *buf, int size, loff_t *offs)
1721 {
1722         struct buffer_head *bh;
1723         unsigned long block;
1724         int osize;
1725         int blocksize;
1726         int csize;
1727         int boffs;
1728
1729         /* prevent reading after eof */
1730         spin_lock(&inode->i_lock);
1731         if (i_size_read(inode) < *offs + size) {
1732                 loff_t diff = i_size_read(inode) - *offs;
1733                 spin_unlock(&inode->i_lock);
1734                 if (diff < 0) {
1735                         CDEBUG(D_OTHER,
1736                                "size %llu is too short to read @%llu\n",
1737                                i_size_read(inode), *offs);
1738                         return -EBADR;
1739                 } else if (diff == 0) {
1740                         return 0;
1741                 } else {
1742                         size = diff;
1743                 }
1744         } else {
1745                 spin_unlock(&inode->i_lock);
1746         }
1747
1748         blocksize = 1 << inode->i_blkbits;
1749         osize = size;
1750         while (size > 0) {
1751                 block = *offs >> inode->i_blkbits;
1752                 boffs = *offs & (blocksize - 1);
1753                 csize = min(blocksize - boffs, size);
1754                 bh = __ldiskfs_bread(NULL, inode, block, 0);
1755                 if (IS_ERR(bh)) {
1756                         CERROR("%s: can't read %u@%llu on ino %lu: "
1757                                "rc = %ld\n", osd_ino2name(inode),
1758                                csize, *offs, inode->i_ino,
1759                                PTR_ERR(bh));
1760                         return PTR_ERR(bh);
1761                 }
1762
1763                 if (bh != NULL) {
1764                         memcpy(buf, bh->b_data + boffs, csize);
1765                         brelse(bh);
1766                 } else {
1767                         memset(buf, 0, csize);
1768                 }
1769
1770                 *offs += csize;
1771                 buf += csize;
1772                 size -= csize;
1773         }
1774         return osize;
1775 }
1776
1777 static ssize_t osd_read(const struct lu_env *env, struct dt_object *dt,
1778                         struct lu_buf *buf, loff_t *pos)
1779 {
1780         struct inode *inode = osd_dt_obj(dt)->oo_inode;
1781         int           rc;
1782
1783         /* Read small symlink from inode body as we need to maintain correct
1784          * on-disk symlinks for ldiskfs.
1785          */
1786         if (S_ISLNK(dt->do_lu.lo_header->loh_attr) &&
1787             (buf->lb_len < sizeof(LDISKFS_I(inode)->i_data)))
1788                 rc = osd_ldiskfs_readlink(inode, buf->lb_buf, buf->lb_len);
1789         else
1790                 rc = osd_ldiskfs_read(inode, buf->lb_buf, buf->lb_len, pos);
1791
1792         return rc;
1793 }
1794
1795 static inline int osd_extents_enabled(struct super_block *sb,
1796                                       struct inode *inode)
1797 {
1798         if (inode != NULL) {
1799                 if (LDISKFS_I(inode)->i_flags & LDISKFS_EXTENTS_FL)
1800                         return 1;
1801         } else if (ldiskfs_has_feature_extents(sb)) {
1802                 return 1;
1803         }
1804         return 0;
1805 }
1806
1807 int osd_calc_bkmap_credits(struct super_block *sb, struct inode *inode,
1808                            const loff_t size, const loff_t pos,
1809                            const int blocks)
1810 {
1811         int credits, bits, bs, i;
1812
1813         bits = sb->s_blocksize_bits;
1814         bs = 1 << bits;
1815
1816         /* legacy blockmap: 3 levels * 3 (bitmap,gd,itself)
1817          * we do not expect blockmaps on the large files,
1818          * so let's shrink it to 2 levels (4GB files) */
1819
1820         /* this is default reservation: 2 levels */
1821         credits = (blocks + 2) * 3;
1822
1823         /* actual offset is unknown, hard to optimize */
1824         if (pos == -1)
1825                 return credits;
1826
1827         /* now check for few specific cases to optimize */
1828         if (pos + size <= LDISKFS_NDIR_BLOCKS * bs) {
1829                 /* no indirects */
1830                 credits = blocks;
1831                 /* allocate if not allocated */
1832                 if (inode == NULL) {
1833                         credits += blocks * 2;
1834                         return credits;
1835                 }
1836                 for (i = (pos >> bits); i < (pos >> bits) + blocks; i++) {
1837                         LASSERT(i < LDISKFS_NDIR_BLOCKS);
1838                         if (LDISKFS_I(inode)->i_data[i] == 0)
1839                                 credits += 2;
1840                 }
1841         } else if (pos + size <= (LDISKFS_NDIR_BLOCKS + 1024) * bs) {
1842                 /* single indirect */
1843                 credits = blocks * 3;
1844                 if (inode == NULL ||
1845                     LDISKFS_I(inode)->i_data[LDISKFS_IND_BLOCK] == 0)
1846                         credits += 3;
1847                 else
1848                         /* The indirect block may be modified. */
1849                         credits += 1;
1850         }
1851
1852         return credits;
1853 }
1854
1855 static ssize_t osd_declare_write(const struct lu_env *env, struct dt_object *dt,
1856                                  const struct lu_buf *buf, loff_t _pos,
1857                                  struct thandle *handle)
1858 {
1859         struct osd_object  *obj  = osd_dt_obj(dt);
1860         struct inode       *inode = obj->oo_inode;
1861         struct super_block *sb = osd_sb(osd_obj2dev(obj));
1862         struct osd_thandle *oh;
1863         int                 rc = 0, est = 0, credits, blocks, allocated = 0;
1864         int                 bits, bs;
1865         int                 depth, size;
1866         loff_t              pos;
1867         ENTRY;
1868
1869         LASSERT(buf != NULL);
1870         LASSERT(handle != NULL);
1871
1872         oh = container_of0(handle, struct osd_thandle, ot_super);
1873         LASSERT(oh->ot_handle == NULL);
1874
1875         size = buf->lb_len;
1876         bits = sb->s_blocksize_bits;
1877         bs = 1 << bits;
1878
1879         if (_pos == -1) {
1880                 /* if this is an append, then we
1881                  * should expect cross-block record */
1882                 pos = 0;
1883         } else {
1884                 pos = _pos;
1885         }
1886
1887         /* blocks to modify */
1888         blocks = ((pos + size + bs - 1) >> bits) - (pos >> bits);
1889         LASSERT(blocks > 0);
1890
1891         if (inode != NULL && _pos != -1) {
1892                 /* object size in blocks */
1893                 est = (i_size_read(inode) + bs - 1) >> bits;
1894                 allocated = inode->i_blocks >> (bits - 9);
1895                 if (pos + size <= i_size_read(inode) && est <= allocated) {
1896                         /* looks like an overwrite, no need to modify tree */
1897                         credits = blocks;
1898                         /* no need to modify i_size */
1899                         goto out;
1900                 }
1901         }
1902
1903         if (osd_extents_enabled(sb, inode)) {
1904                 /*
1905                  * many concurrent threads may grow tree by the time
1906                  * our transaction starts. so, consider 2 is a min depth
1907                  * for every level we may need to allocate a new block
1908                  * and take some entries from the old one. so, 3 blocks
1909                  * to allocate (bitmap, gd, itself) + old block - 4 per
1910                  * level.
1911                  */
1912                 depth = inode != NULL ? ext_depth(inode) : 0;
1913                 depth = max(depth, 1) + 1;
1914                 credits = depth;
1915                 /* if not append, then split may need to modify
1916                  * existing blocks moving entries into the new ones */
1917                 if (_pos != -1)
1918                         credits += depth;
1919                 /* blocks to store data: bitmap,gd,itself */
1920                 credits += blocks * 3;
1921         } else {
1922                 credits = osd_calc_bkmap_credits(sb, inode, size, _pos, blocks);
1923         }
1924         /* if inode is created as part of the transaction,
1925          * then it's counted already by the creation method */
1926         if (inode != NULL)
1927                 credits++;
1928
1929 out:
1930
1931         osd_trans_declare_op(env, oh, OSD_OT_WRITE, credits);
1932
1933         /* dt_declare_write() is usually called for system objects, such
1934          * as llog or last_rcvd files. We needn't enforce quota on those
1935          * objects, so always set the lqi_space as 0. */
1936         if (inode != NULL)
1937                 rc = osd_declare_inode_qid(env, i_uid_read(inode),
1938                                            i_gid_read(inode),
1939                                            i_projid_read(inode), 0,
1940                                            oh, obj, NULL, OSD_QID_BLK);
1941
1942         if (rc == 0)
1943                 rc = osd_trunc_lock(obj, oh, true);
1944
1945         RETURN(rc);
1946 }
1947
1948 static int osd_ldiskfs_writelink(struct inode *inode, char *buffer, int buflen)
1949 {
1950         /* LU-2634: clear the extent format for fast symlink */
1951         ldiskfs_clear_inode_flag(inode, LDISKFS_INODE_EXTENTS);
1952
1953         memcpy((char *)&LDISKFS_I(inode)->i_data, (char *)buffer, buflen);
1954         spin_lock(&inode->i_lock);
1955         LDISKFS_I(inode)->i_disksize = buflen;
1956         i_size_write(inode, buflen);
1957         spin_unlock(&inode->i_lock);
1958         ll_dirty_inode(inode, I_DIRTY_DATASYNC);
1959
1960         return 0;
1961 }
1962
1963 int osd_ldiskfs_write_record(struct inode *inode, void *buf, int bufsize,
1964                              int write_NUL, loff_t *offs, handle_t *handle)
1965 {
1966         struct buffer_head *bh        = NULL;
1967         loff_t              offset    = *offs;
1968         loff_t              new_size  = i_size_read(inode);
1969         unsigned long       block;
1970         int                 blocksize = 1 << inode->i_blkbits;
1971         int                 err = 0;
1972         int                 size;
1973         int                 boffs;
1974         int                 dirty_inode = 0;
1975
1976         if (write_NUL) {
1977                 /*
1978                  * long symlink write does not count the NUL terminator in
1979                  * bufsize, we write it, and the inode's file size does not
1980                  * count the NUL terminator as well.
1981                  */
1982                 ((char *)buf)[bufsize] = '\0';
1983                 ++bufsize;
1984         }
1985
1986         while (bufsize > 0) {
1987                 int credits = handle->h_buffer_credits;
1988
1989                 if (bh)
1990                         brelse(bh);
1991
1992                 block = offset >> inode->i_blkbits;
1993                 boffs = offset & (blocksize - 1);
1994                 size = min(blocksize - boffs, bufsize);
1995                 bh = __ldiskfs_bread(handle, inode, block, 1);
1996                 if (IS_ERR_OR_NULL(bh)) {
1997                         if (bh == NULL) {
1998                                 err = -EIO;
1999                         } else {
2000                                 err = PTR_ERR(bh);
2001                                 bh = NULL;
2002                         }
2003
2004                         CERROR("%s: error reading offset %llu (block %lu, "
2005                                "size %d, offs %llu), credits %d/%d: rc = %d\n",
2006                                inode->i_sb->s_id, offset, block, bufsize, *offs,
2007                                credits, handle->h_buffer_credits, err);
2008                         break;
2009                 }
2010
2011                 err = ldiskfs_journal_get_write_access(handle, bh);
2012                 if (err) {
2013                         CERROR("journal_get_write_access() returned error %d\n",
2014                                err);
2015                         break;
2016                 }
2017                 LASSERTF(boffs + size <= bh->b_size,
2018                          "boffs %d size %d bh->b_size %lu\n",
2019                          boffs, size, (unsigned long)bh->b_size);
2020                 memcpy(bh->b_data + boffs, buf, size);
2021                 err = ldiskfs_handle_dirty_metadata(handle, NULL, bh);
2022                 if (err)
2023                         break;
2024
2025                 if (offset + size > new_size)
2026                         new_size = offset + size;
2027                 offset += size;
2028                 bufsize -= size;
2029                 buf += size;
2030         }
2031         if (bh)
2032                 brelse(bh);
2033
2034         if (write_NUL)
2035                 --new_size;
2036         /* correct in-core and on-disk sizes */
2037         if (new_size > i_size_read(inode)) {
2038                 spin_lock(&inode->i_lock);
2039                 if (new_size > i_size_read(inode))
2040                         i_size_write(inode, new_size);
2041                 if (i_size_read(inode) > LDISKFS_I(inode)->i_disksize) {
2042                         LDISKFS_I(inode)->i_disksize = i_size_read(inode);
2043                         dirty_inode = 1;
2044                 }
2045                 spin_unlock(&inode->i_lock);
2046                 if (dirty_inode)
2047                         ll_dirty_inode(inode, I_DIRTY_DATASYNC);
2048         }
2049
2050         if (err == 0)
2051                 *offs = offset;
2052         return err;
2053 }
2054
2055 static ssize_t osd_write(const struct lu_env *env, struct dt_object *dt,
2056                          const struct lu_buf *buf, loff_t *pos,
2057                          struct thandle *handle)
2058 {
2059         struct inode            *inode = osd_dt_obj(dt)->oo_inode;
2060         struct osd_thandle      *oh;
2061         ssize_t                 result;
2062         int                     is_link;
2063
2064         LASSERT(dt_object_exists(dt));
2065
2066         LASSERT(handle != NULL);
2067         LASSERT(inode != NULL);
2068         ll_vfs_dq_init(inode);
2069
2070         /* XXX: don't check: one declared chunk can be used many times */
2071         /* osd_trans_exec_op(env, handle, OSD_OT_WRITE); */
2072
2073         oh = container_of(handle, struct osd_thandle, ot_super);
2074         LASSERT(oh->ot_handle->h_transaction != NULL);
2075         osd_trans_exec_op(env, handle, OSD_OT_WRITE);
2076
2077         /* Write small symlink to inode body as we need to maintain correct
2078          * on-disk symlinks for ldiskfs.
2079          * Note: the buf->lb_buf contains a NUL terminator while buf->lb_len
2080          * does not count it in.
2081          */
2082         is_link = S_ISLNK(dt->do_lu.lo_header->loh_attr);
2083         if (is_link && (buf->lb_len < sizeof(LDISKFS_I(inode)->i_data)))
2084                 result = osd_ldiskfs_writelink(inode, buf->lb_buf, buf->lb_len);
2085         else
2086                 result = osd_ldiskfs_write_record(inode, buf->lb_buf,
2087                                                   buf->lb_len, is_link, pos,
2088                                                   oh->ot_handle);
2089         if (result == 0)
2090                 result = buf->lb_len;
2091
2092         osd_trans_exec_check(env, handle, OSD_OT_WRITE);
2093
2094         return result;
2095 }
2096
2097 static int osd_declare_punch(const struct lu_env *env, struct dt_object *dt,
2098                              __u64 start, __u64 end, struct thandle *th)
2099 {
2100         struct osd_thandle *oh;
2101         struct inode       *inode;
2102         int                 rc;
2103         ENTRY;
2104
2105         LASSERT(th);
2106         oh = container_of(th, struct osd_thandle, ot_super);
2107
2108         /*
2109          * we don't need to reserve credits for whole truncate
2110          * it's not possible as truncate may need to free too many
2111          * blocks and that won't fit a single transaction. instead
2112          * we reserve credits to change i_size and put inode onto
2113          * orphan list. if needed truncate will extend or restart
2114          * transaction
2115          */
2116         osd_trans_declare_op(env, oh, OSD_OT_PUNCH,
2117                              osd_dto_credits_noquota[DTO_ATTR_SET_BASE] + 3);
2118
2119         inode = osd_dt_obj(dt)->oo_inode;
2120         LASSERT(inode);
2121
2122         rc = osd_declare_inode_qid(env, i_uid_read(inode), i_gid_read(inode),
2123                                    i_projid_read(inode), 0, oh, osd_dt_obj(dt),
2124                                    NULL, OSD_QID_BLK);
2125
2126         if (rc == 0)
2127                 rc = osd_trunc_lock(osd_dt_obj(dt), oh, false);
2128
2129         RETURN(rc);
2130 }
2131
2132 static int osd_punch(const struct lu_env *env, struct dt_object *dt,
2133                      __u64 start, __u64 end, struct thandle *th)
2134 {
2135         struct osd_object *obj = osd_dt_obj(dt);
2136         struct osd_device *osd = osd_obj2dev(obj);
2137         struct inode *inode = obj->oo_inode;
2138         struct osd_access_lock *al;
2139         struct osd_thandle *oh;
2140         int rc = 0, found = 0;
2141         bool grow = false;
2142         ENTRY;
2143
2144         LASSERT(end == OBD_OBJECT_EOF);
2145         LASSERT(dt_object_exists(dt));
2146         LASSERT(osd_invariant(obj));
2147         LASSERT(inode != NULL);
2148         ll_vfs_dq_init(inode);
2149
2150         LASSERT(th);
2151         oh = container_of(th, struct osd_thandle, ot_super);
2152         LASSERT(oh->ot_handle->h_transaction != NULL);
2153
2154         /* we used to skip truncate to current size to
2155          * optimize truncates on OST. with DoM we can
2156          * get attr_set to set specific size (MDS_REINT)
2157          * and then get truncate RPC which essentially
2158          * would be skipped. this is bad.. so, disable
2159          * this optimization on MDS till the client stop
2160          * to sent MDS_REINT (LU-11033) -bzzz */
2161         if (osd->od_is_ost && i_size_read(inode) == start)
2162                 RETURN(0);
2163
2164         osd_trans_exec_op(env, th, OSD_OT_PUNCH);
2165
2166         spin_lock(&inode->i_lock);
2167         if (i_size_read(inode) < start)
2168                 grow = true;
2169         i_size_write(inode, start);
2170         spin_unlock(&inode->i_lock);
2171         ll_truncate_pagecache(inode, start);
2172
2173         /* optimize grow case */
2174         if (grow) {
2175                 osd_execute_truncate(obj);
2176                 GOTO(out, rc);
2177         }
2178
2179         /* add to orphan list to ensure truncate completion
2180          * if this transaction succeed. ldiskfs_truncate()
2181          * will take the inode out of the list */
2182         rc = ldiskfs_orphan_add(oh->ot_handle, inode);
2183         if (rc != 0)
2184                 GOTO(out, rc);
2185
2186         list_for_each_entry(al, &oh->ot_trunc_locks, tl_list) {
2187                 if (obj != al->tl_obj)
2188                         continue;
2189                 LASSERT(al->tl_shared == 0);
2190                 found = 1;
2191                 /* do actual truncate in osd_trans_stop() */
2192                 al->tl_truncate = 1;
2193                 break;
2194         }
2195         LASSERT(found);
2196
2197 out:
2198         RETURN(rc);
2199 }
2200
2201 static int fiemap_check_ranges(struct inode *inode,
2202                                u64 start, u64 len, u64 *new_len)
2203 {
2204         loff_t maxbytes;
2205
2206         *new_len = len;
2207
2208         if (len == 0)
2209                 return -EINVAL;
2210
2211         if (ldiskfs_test_inode_flag(inode, LDISKFS_INODE_EXTENTS))
2212                 maxbytes = inode->i_sb->s_maxbytes;
2213         else
2214                 maxbytes = LDISKFS_SB(inode->i_sb)->s_bitmap_maxbytes;
2215
2216         if (start > maxbytes)
2217                 return -EFBIG;
2218
2219         /*
2220          * Shrink request scope to what the fs can actually handle.
2221          */
2222         if (len > maxbytes || (maxbytes - len) < start)
2223                 *new_len = maxbytes - start;
2224
2225         return 0;
2226 }
2227
2228 /* So that the fiemap access checks can't overflow on 32 bit machines. */
2229 #define FIEMAP_MAX_EXTENTS     (UINT_MAX / sizeof(struct fiemap_extent))
2230
2231 static int osd_fiemap_get(const struct lu_env *env, struct dt_object *dt,
2232                           struct fiemap *fm)
2233 {
2234         struct fiemap_extent_info fieinfo = {0, };
2235         struct inode *inode = osd_dt_obj(dt)->oo_inode;
2236         u64 len;
2237         int rc;
2238         mm_segment_t cur_fs;
2239
2240         LASSERT(inode);
2241         if (inode->i_op->fiemap == NULL)
2242                 return -EOPNOTSUPP;
2243
2244         if (fm->fm_extent_count > FIEMAP_MAX_EXTENTS)
2245                 return -EINVAL;
2246
2247         rc = fiemap_check_ranges(inode, fm->fm_start, fm->fm_length, &len);
2248         if (rc)
2249                 return rc;
2250
2251         fieinfo.fi_flags = fm->fm_flags;
2252         fieinfo.fi_extents_max = fm->fm_extent_count;
2253         fieinfo.fi_extents_start = fm->fm_extents;
2254
2255         if (fieinfo.fi_flags & FIEMAP_FLAG_SYNC)
2256                 filemap_write_and_wait(inode->i_mapping);
2257
2258         /* Save previous value address limit */
2259         cur_fs = get_fs();
2260         /* Set the address limit of the kernel */
2261         set_fs(KERNEL_DS);
2262
2263         rc = inode->i_op->fiemap(inode, &fieinfo, fm->fm_start, len);
2264         fm->fm_flags = fieinfo.fi_flags;
2265         fm->fm_mapped_extents = fieinfo.fi_extents_mapped;
2266
2267         /* Restore the previous address limt */
2268         set_fs(cur_fs);
2269
2270         return rc;
2271 }
2272
2273 static int osd_ladvise(const struct lu_env *env, struct dt_object *dt,
2274                        __u64 start, __u64 end, enum lu_ladvise_type advice)
2275 {
2276         int              rc = 0;
2277         struct inode    *inode = osd_dt_obj(dt)->oo_inode;
2278         ENTRY;
2279
2280         switch (advice) {
2281         case LU_LADVISE_DONTNEED:
2282                 if (end == 0)
2283                         break;
2284                 invalidate_mapping_pages(inode->i_mapping,
2285                                          start >> PAGE_SHIFT,
2286                                          (end - 1) >> PAGE_SHIFT);
2287                 break;
2288         default:
2289                 rc = -ENOTSUPP;
2290                 break;
2291         }
2292
2293         RETURN(rc);
2294 }
2295
2296 /*
2297  * in some cases we may need declare methods for objects being created
2298  * e.g., when we create symlink
2299  */
2300 const struct dt_body_operations osd_body_ops_new = {
2301         .dbo_declare_write = osd_declare_write,
2302 };
2303
2304 const struct dt_body_operations osd_body_ops = {
2305         .dbo_read                       = osd_read,
2306         .dbo_declare_write              = osd_declare_write,
2307         .dbo_write                      = osd_write,
2308         .dbo_bufs_get                   = osd_bufs_get,
2309         .dbo_bufs_put                   = osd_bufs_put,
2310         .dbo_write_prep                 = osd_write_prep,
2311         .dbo_declare_write_commit       = osd_declare_write_commit,
2312         .dbo_write_commit               = osd_write_commit,
2313         .dbo_read_prep                  = osd_read_prep,
2314         .dbo_declare_punch              = osd_declare_punch,
2315         .dbo_punch                      = osd_punch,
2316         .dbo_fiemap_get                 = osd_fiemap_get,
2317         .dbo_ladvise                    = osd_ladvise,
2318 };
2319
2320 /**
2321  * Get a truncate lock
2322  *
2323  * In order to take multi-transaction truncate out of main transaction we let
2324  * the caller grab a lock on the object passed. the lock can be shared (for
2325  * writes) and exclusive (for truncate). It's not allowed to mix truncate
2326  * and write in the same transaction handle (do not confuse with big ldiskfs
2327  * transaction containing lots of handles).
2328  * The lock must be taken at declaration.
2329  *
2330  * \param obj           object to lock
2331  * \oh                  transaction
2332  * \shared              shared or exclusive
2333  *
2334  * \retval 0            lock is granted
2335  * \retval -NOMEM       no memory to allocate lock
2336  */
2337 int osd_trunc_lock(struct osd_object *obj, struct osd_thandle *oh, bool shared)
2338 {
2339         struct osd_access_lock *al, *tmp;
2340
2341         LASSERT(obj);
2342         LASSERT(oh);
2343
2344         list_for_each_entry(tmp, &oh->ot_trunc_locks, tl_list) {
2345                 if (tmp->tl_obj != obj)
2346                         continue;
2347                 LASSERT(tmp->tl_shared == shared);
2348                 /* found same lock */
2349                 return 0;
2350         }
2351
2352         OBD_ALLOC_PTR(al);
2353         if (unlikely(al == NULL))
2354                 return -ENOMEM;
2355         al->tl_obj = obj;
2356         al->tl_truncate = false;
2357         if (shared)
2358                 down_read(&obj->oo_ext_idx_sem);
2359         else
2360                 down_write(&obj->oo_ext_idx_sem);
2361         al->tl_shared = shared;
2362
2363         list_add(&al->tl_list, &oh->ot_trunc_locks);
2364
2365         return 0;
2366 }
2367
2368 void osd_trunc_unlock_all(struct list_head *list)
2369 {
2370         struct osd_access_lock *al, *tmp;
2371         list_for_each_entry_safe(al, tmp, list, tl_list) {
2372                 if (al->tl_shared)
2373                         up_read(&al->tl_obj->oo_ext_idx_sem);
2374                 else
2375                         up_write(&al->tl_obj->oo_ext_idx_sem);
2376                 list_del(&al->tl_list);
2377                 OBD_FREE_PTR(al);
2378         }
2379 }
2380
2381 void osd_execute_truncate(struct osd_object *obj)
2382 {
2383         struct osd_device *d = osd_obj2dev(obj);
2384         struct inode *inode = obj->oo_inode;
2385         __u64 size;
2386
2387         /* simulate crash before (in the middle) of delayed truncate */
2388         if (OBD_FAIL_CHECK(OBD_FAIL_OSD_FAIL_AT_TRUNCATE)) {
2389                 struct ldiskfs_inode_info *ei = LDISKFS_I(inode);
2390                 struct ldiskfs_sb_info *sbi = LDISKFS_SB(inode->i_sb);
2391
2392                 mutex_lock(&sbi->s_orphan_lock);
2393                 list_del_init(&ei->i_orphan);
2394                 mutex_unlock(&sbi->s_orphan_lock);
2395                 return;
2396         }
2397
2398         ldiskfs_truncate(inode);
2399
2400         /*
2401          * For a partial-page truncate, flush the page to disk immediately to
2402          * avoid data corruption during direct disk write.  b=17397
2403          */
2404         size = i_size_read(inode);
2405         if ((size & ~PAGE_MASK) == 0)
2406                 return;
2407         if (osd_use_page_cache(d)) {
2408                 filemap_fdatawrite_range(inode->i_mapping, size, size + 1);
2409         } else {
2410                 /* Notice we use "wait" version to ensure I/O is complete */
2411                 filemap_write_and_wait_range(inode->i_mapping, size, size + 1);
2412                 invalidate_mapping_pages(inode->i_mapping, size >> PAGE_SHIFT,
2413                                          size >> PAGE_SHIFT);
2414         }
2415 }
2416
2417 void osd_process_truncates(struct list_head *list)
2418 {
2419         struct osd_access_lock *al;
2420
2421         LASSERT(journal_current_handle() == NULL);
2422
2423         list_for_each_entry(al, list, tl_list) {
2424                 if (al->tl_shared)
2425                         continue;
2426                 if (!al->tl_truncate)
2427                         continue;
2428                 osd_execute_truncate(al->tl_obj);
2429         }
2430 }