lustre/llite/rw.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * Lustre Lite I/O page cache routines shared by different kernel revs
   5  *
   6  *  Copyright (c) 2001-2003 Cluster File Systems, Inc.
   7  *
   8  *   This file is part of Lustre, http://www.lustre.org.
   9  *
  10  *   Lustre is free software; you can redistribute it and/or
  11  *   modify it under the terms of version 2 of the GNU General Public
  12  *   License as published by the Free Software Foundation.
  13  *
  14  *   Lustre is distributed in the hope that it will be useful,
  15  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  17  *   GNU General Public License for more details.
  18  *
  19  *   You should have received a copy of the GNU General Public License
  20  *   along with Lustre; if not, write to the Free Software
  21  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  22  */
  23
  24 #include <linux/config.h>
  25 #include <linux/kernel.h>
  26 #include <linux/mm.h>
  27 #include <linux/string.h>
  28 #include <linux/stat.h>
  29 #include <linux/errno.h>
  30 #include <linux/smp_lock.h>
  31 #include <linux/unistd.h>
  32 #include <linux/version.h>
  33 #include <asm/system.h>
  34 #include <asm/uaccess.h>
  35
  36 #include <linux/fs.h>
  37 #include <linux/stat.h>
  38 #include <asm/uaccess.h>
  39 #include <asm/segment.h>
  40 #include <linux/mm.h>
  41 #include <linux/pagemap.h>
  42 #include <linux/smp_lock.h>
  43
  44 #define DEBUG_SUBSYSTEM S_LLITE
  45
  46 #include <linux/lustre_mds.h>
  47 #include <linux/lustre_lite.h>
  48 #include "llite_internal.h"
  49 #include <linux/lustre_compat25.h>
  50
  51 #ifndef list_for_each_prev_safe
  52 #define list_for_each_prev_safe(pos, n, head) \
  53         for (pos = (head)->prev, n = pos->prev; pos != (head); \
  54                 pos = n, n = pos->prev )
  55 #endif
  56
  57 /* SYNCHRONOUS I/O to object storage for an inode */
  58 static int ll_brw(int cmd, struct inode *inode, struct obdo *oa,
  59                   struct page *page, int flags)
  60 {
  61         struct ll_inode_info *lli = ll_i2info(inode);
  62         struct lov_stripe_md *lsm = lli->lli_smd;
  63         struct timeval start;
  64         struct brw_page pg;
  65         int rc;
  66         ENTRY;
  67
  68         do_gettimeofday(&start);
  69
  70         pg.pg = page;
  71         pg.disk_offset = pg.page_offset = ((obd_off)page->index) << PAGE_SHIFT;
  72
  73         if (cmd == OBD_BRW_WRITE &&
  74             (pg.disk_offset + PAGE_SIZE > inode->i_size))
  75                 pg.count = inode->i_size % PAGE_SIZE;
  76         else
  77                 pg.count = PAGE_SIZE;
  78
  79         CDEBUG(D_PAGE, "%s %d bytes ino %lu at "LPU64"/"LPX64"\n",
  80                cmd & OBD_BRW_WRITE ? "write" : "read", pg.count, inode->i_ino,
  81                pg.disk_offset, pg.disk_offset);
  82         if (pg.count == 0) {
  83                 CERROR("ZERO COUNT: ino %lu: size %p:%Lu(%p:%Lu) idx %lu off "
  84                        LPU64"\n", inode->i_ino, inode, inode->i_size,
  85                        page->mapping->host, page->mapping->host->i_size,
  86                        page->index, pg.disk_offset);
  87         }
  88
  89         pg.flag = flags;
  90
  91         if (cmd == OBD_BRW_WRITE)
  92                 lprocfs_counter_add(ll_i2sbi(inode)->ll_stats,
  93                                     LPROC_LL_BRW_WRITE, pg.count);
  94         else
  95                 lprocfs_counter_add(ll_i2sbi(inode)->ll_stats,
  96                                     LPROC_LL_BRW_READ, pg.count);
  97         rc = obd_brw(cmd, ll_i2dtexp(inode), oa, lsm, 1, &pg, NULL);
  98         if (rc == 0)
  99                 obdo_to_inode(inode, oa, OBD_MD_FLBLOCKS);
 100         else if (rc != -EIO)
 101                 CERROR("error from obd_brw: rc = %d\n", rc);
 102         ll_stime_record(ll_i2sbi(inode), &start,
 103                         &ll_i2sbi(inode)->ll_brw_stime);
 104         RETURN(rc);
 105 }
 106
 107 __u64 lov_merge_size(struct lov_stripe_md *lsm, int kms);
 108
 109 /*
 110  * this isn't where truncate starts.   roughly:
 111  * sys_truncate->ll_setattr_raw->vmtruncate->ll_truncate
 112  * we grab the lock back in setattr_raw to avoid races.
 113  *
 114  * must be called with lli_size_sem held.
 115  */
 116 void ll_truncate(struct inode *inode)
 117 {
 118         struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
 119         struct ll_inode_info *lli = ll_i2info(inode);
 120         struct obdo *oa = NULL;
 121         int rc;
 122         ENTRY;
 123
 124         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p) to %llu\n", inode->i_ino,
 125                inode->i_generation, inode, inode->i_size);
 126
 127         if (!lsm) {
 128                 CDEBUG(D_INODE, "truncate on inode %lu with no objects\n",
 129                        inode->i_ino);
 130                 GOTO(out_unlock, 0);
 131         }
 132
 133         if (lov_merge_size(lsm, 0) == inode->i_size) {
 134                 CDEBUG(D_VFSTRACE, "skipping punch for "LPX64" (size = %llu)\n",
 135                        lsm->lsm_object_id, inode->i_size);
 136                 GOTO(out_unlock, 0);
 137         }
 138
 139         CDEBUG(D_INFO, "calling punch for "LPX64" (new size %llu)\n",
 140                lsm->lsm_object_id, inode->i_size);
 141
 142         oa = obdo_alloc();
 143         if (oa == NULL) {
 144                 CERROR("cannot alloc oa, error %d\n",
 145                        -ENOMEM);
 146                 EXIT;
 147                 return;
 148         }
 149
 150         oa->o_id = lsm->lsm_object_id;
 151         oa->o_gr = lsm->lsm_object_gr;
 152         oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
 153         obdo_from_inode(oa, inode, OBD_MD_FLTYPE | OBD_MD_FLMODE |
 154                         OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME);
 155
 156         obd_adjust_kms(ll_i2dtexp(inode), lsm, inode->i_size, 1);
 157
 158         LASSERT(atomic_read(&lli->lli_size_sem.count) <= 0);
 159         up(&lli->lli_size_sem);
 160
 161         rc = obd_punch(ll_i2dtexp(inode), oa, lsm, inode->i_size,
 162                        OBD_OBJECT_EOF, NULL);
 163         if (rc)
 164                 CERROR("obd_truncate fails (%d) ino %lu\n", rc, inode->i_ino);
 165         else
 166                 obdo_to_inode(inode, oa, OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
 167                               OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME);
 168
 169         obdo_free(oa);
 170
 171         EXIT;
 172         return;
 173
 174 out_unlock:
 175         LASSERT(atomic_read(&lli->lli_size_sem.count) <= 0);
 176         up(&lli->lli_size_sem);
 177 } /* ll_truncate */
 178
 179 int ll_prepare_write(struct file *file, struct page *page, unsigned from,
 180                      unsigned to)
 181 {
 182         struct inode *inode = page->mapping->host;
 183         struct ll_inode_info *lli = ll_i2info(inode);
 184         struct lov_stripe_md *lsm = lli->lli_smd;
 185         obd_off offset = ((obd_off)page->index) << PAGE_SHIFT;
 186         struct obdo *oa = NULL;
 187         struct brw_page pga;
 188         __u64 kms;
 189         int rc = 0;
 190         ENTRY;
 191
 192         LASSERT(PageLocked(page));
 193         (void)llap_cast_private(page); /* assertion */
 194
 195         /* Check to see if we should return -EIO right away */
 196         pga.pg = page;
 197         pga.disk_offset = pga.page_offset = offset;
 198         pga.count = PAGE_SIZE;
 199         pga.flag = 0;
 200
 201         oa = obdo_alloc();
 202         if (oa == NULL)
 203                 RETURN(-ENOMEM);
 204
 205         oa->o_id = lsm->lsm_object_id;
 206         oa->o_gr = lsm->lsm_object_gr;
 207         oa->o_mode = inode->i_mode;
 208         oa->o_valid = OBD_MD_FLID | OBD_MD_FLMODE |
 209                 OBD_MD_FLTYPE | OBD_MD_FLGROUP;
 210
 211         rc = obd_brw(OBD_BRW_CHECK, ll_i2dtexp(inode), oa, lsm,
 212                      1, &pga, NULL);
 213         if (rc)
 214                 GOTO(out_free_oa, rc);
 215
 216         if (PageUptodate(page))
 217                 GOTO(out_free_oa, 0);
 218
 219         /* We're completely overwriting an existing page, so _don't_ set it up
 220          * to date until commit_write */
 221         if (from == 0 && to == PAGE_SIZE) {
 222                 POISON_PAGE(page, 0x11);
 223                 GOTO(out_free_oa, 0);
 224         }
 225
 226         /* If are writing to a new page, no need to read old data.  The extent
 227          * locking will have updated the KMS, and for our purposes here we can
 228          * treat it like i_size. */
 229         down(&lli->lli_size_sem);
 230         kms = lov_merge_size(lsm, 1);
 231         up(&lli->lli_size_sem);
 232         if (kms <= offset) {
 233                 memset(kmap(page), 0, PAGE_SIZE);
 234                 kunmap(page);
 235                 GOTO(prepare_done, rc = 0);
 236         }
 237
 238         /* XXX could be an async ocp read.. read-ahead? */
 239         rc = ll_brw(OBD_BRW_READ, inode, oa, page, 0);
 240         if (rc == 0) {
 241                 /* bug 1598: don't clobber blksize */
 242                 oa->o_valid &= ~(OBD_MD_FLSIZE | OBD_MD_FLBLKSZ);
 243                 obdo_refresh_inode(inode, oa, oa->o_valid);
 244         }
 245
 246         EXIT;
 247 prepare_done:
 248         if (rc == 0)
 249                 SetPageUptodate(page);
 250 out_free_oa:
 251         obdo_free(oa);
 252         return rc;
 253 }
 254
 255 static int ll_ap_make_ready(void *data, int cmd)
 256 {
 257         struct ll_async_page *llap;
 258         struct page *page;
 259         ENTRY;
 260
 261         llap = LLAP_FROM_COOKIE(data);
 262         page = llap->llap_page;
 263
 264         LASSERT(cmd != OBD_BRW_READ);
 265
 266         /* we're trying to write, but the page is locked.. come back later */
 267         if (TryLockPage(page))
 268                 RETURN(-EAGAIN);
 269
 270         LL_CDEBUG_PAGE(D_PAGE, page, "made ready\n");
 271         page_cache_get(page);
 272
 273         /* if we left PageDirty we might get another writepage call
 274          * in the future.  list walkers are bright enough
 275          * to check page dirty so we can leave it on whatever list
 276          * its on.  XXX also, we're called with the cli list so if
 277          * we got the page cache list we'd create a lock inversion
 278          * with the removepage path which gets the page lock then the
 279          * cli lock */
 280         clear_page_dirty(page);
 281         RETURN(0);
 282 }
 283
 284 /* We have two reasons for giving llite the opportunity to change the
 285  * write length of a given queued page as it builds the RPC containing
 286  * the page:
 287  *
 288  * 1) Further extending writes may have landed in the page cache
 289  *    since a partial write first queued this page requiring us
 290  *    to write more from the page cache.
 291  * 2) We might have raced with truncate and want to avoid performing
 292  *    write RPCs that are just going to be thrown away by the
 293  *    truncate's punch on the storage targets.
 294  *
 295  * The kms serves these purposes as it is set at both truncate and extending
 296  * writes.
 297  */
 298 static int ll_ap_refresh_count(void *data, int cmd)
 299 {
 300         struct ll_async_page *llap;
 301         struct lov_stripe_md *lsm;
 302         struct page *page;
 303         __u64 kms;
 304         ENTRY;
 305
 306         /* readpage queues with _COUNT_STABLE, shouldn't get here. */
 307         LASSERT(cmd != OBD_BRW_READ);
 308
 309         llap = LLAP_FROM_COOKIE(data);
 310         page = llap->llap_page;
 311         lsm = ll_i2info(page->mapping->host)->lli_smd;
 312         kms = lov_merge_size(lsm, 1);
 313
 314         /* catch race with truncate */
 315         if (((__u64)page->index << PAGE_SHIFT) >= kms)
 316                 return 0;
 317
 318         /* catch sub-page write at end of file */
 319         if (((__u64)page->index << PAGE_SHIFT) + PAGE_SIZE > kms)
 320                 return kms % PAGE_SIZE;
 321
 322         return PAGE_SIZE;
 323 }
 324
 325 void ll_inode_fill_obdo(struct inode *inode, int cmd, struct obdo *oa)
 326 {
 327         struct lov_stripe_md *lsm;
 328         obd_valid valid_flags;
 329
 330         lsm = ll_i2info(inode)->lli_smd;
 331
 332         oa->o_id = lsm->lsm_object_id;
 333         oa->o_gr = lsm->lsm_object_gr;
 334         oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
 335         valid_flags = OBD_MD_FLTYPE | OBD_MD_FLATIME;
 336         if (cmd == OBD_BRW_WRITE) {
 337                 oa->o_valid |= OBD_MD_FLIFID | OBD_MD_FLEPOCH;
 338                 mdc_pack_id(obdo_id(oa), inode->i_ino, 0, inode->i_mode,
 339                             id_group(&ll_i2info(inode)->lli_id),
 340                             id_fid(&ll_i2info(inode)->lli_id));
 341
 342                 oa->o_easize = ll_i2info(inode)->lli_io_epoch;
 343                 valid_flags |= OBD_MD_FLMTIME | OBD_MD_FLCTIME;
 344         }
 345
 346         obdo_from_inode(oa, inode, valid_flags);
 347 }
 348
 349 static void ll_ap_fill_obdo(void *data, int cmd, struct obdo *oa)
 350 {
 351         struct ll_async_page *llap;
 352         ENTRY;
 353
 354         llap = LLAP_FROM_COOKIE(data);
 355         ll_inode_fill_obdo(llap->llap_page->mapping->host, cmd, oa);
 356         EXIT;
 357 }
 358
 359 static struct obd_async_page_ops ll_async_page_ops = {
 360         .ap_make_ready =        ll_ap_make_ready,
 361         .ap_refresh_count =     ll_ap_refresh_count,
 362         .ap_fill_obdo =         ll_ap_fill_obdo,
 363         .ap_completion =        ll_ap_completion,
 364 };
 365
 366
 367 struct ll_async_page *llap_cast_private(struct page *page)
 368 {
 369         struct ll_async_page *llap = (struct ll_async_page *)page->private;
 370
 371         LASSERTF(llap == NULL || llap->llap_magic == LLAP_MAGIC,
 372                  "page %p private %lu gave magic %d which != %d\n",
 373                  page, page->private, llap->llap_magic, LLAP_MAGIC);
 374
 375         return llap;
 376 }
 377
 378 /* XXX have the exp be an argument? */
 379 struct ll_async_page *llap_from_page(struct page *page, unsigned origin)
 380 {
 381         struct ll_async_page *llap;
 382         struct obd_export *exp;
 383         struct inode *inode = page->mapping->host;
 384         struct ll_sb_info *sbi = ll_i2sbi(inode);
 385         int rc;
 386         ENTRY;
 387
 388         LASSERTF(origin < LLAP__ORIGIN_MAX, "%u\n", origin);
 389
 390         llap = llap_cast_private(page);
 391         if (llap != NULL)
 392                 GOTO(out, llap);
 393
 394         exp = ll_i2dtexp(page->mapping->host);
 395         if (exp == NULL)
 396                 RETURN(ERR_PTR(-EINVAL));
 397
 398         OBD_ALLOC(llap, sizeof(*llap));
 399         if (llap == NULL)
 400                 RETURN(ERR_PTR(-ENOMEM));
 401         llap->llap_magic = LLAP_MAGIC;
 402         rc = obd_prep_async_page(exp, ll_i2info(inode)->lli_smd, NULL, page,
 403                                  (obd_off)page->index << PAGE_SHIFT,
 404                                  &ll_async_page_ops, llap, &llap->llap_cookie);
 405         if (rc) {
 406                 OBD_FREE(llap, sizeof(*llap));
 407                 RETURN(ERR_PTR(rc));
 408         }
 409
 410         CDEBUG(D_CACHE, "llap %p page %p cookie %p obj off "LPU64"\n", llap,
 411                page, llap->llap_cookie, (obd_off)page->index << PAGE_SHIFT);
 412         /* also zeroing the PRIVBITS low order bitflags */
 413         __set_page_ll_data(page, llap);
 414         llap->llap_page = page;
 415
 416         spin_lock(&sbi->ll_lock);
 417         sbi->ll_pglist_gen++;
 418         list_add_tail(&llap->llap_proc_item, &sbi->ll_pglist);
 419         spin_unlock(&sbi->ll_lock);
 420
 421 out:
 422         llap->llap_origin = origin;
 423         RETURN(llap);
 424 }
 425
 426 static int queue_or_sync_write(struct obd_export *exp,
 427                                struct lov_stripe_md *lsm,
 428                                struct ll_async_page *llap,
 429                                unsigned to,
 430                                obd_flags async_flags)
 431 {
 432         struct obd_io_group *oig;
 433         int rc;
 434         ENTRY;
 435
 436         /* _make_ready only sees llap once we've unlocked the page */
 437         llap->llap_write_queued = 1;
 438         rc = obd_queue_async_io(exp, lsm, NULL, llap->llap_cookie,
 439                                 OBD_BRW_WRITE, 0, 0, 0, async_flags);
 440         if (rc == 0) {
 441                 LL_CDEBUG_PAGE(D_PAGE, llap->llap_page, "write queued\n");
 442                 //llap_write_pending(inode, llap);
 443                 GOTO(out, 0);
 444         }
 445
 446         llap->llap_write_queued = 0;
 447
 448         rc = oig_init(&oig);
 449         if (rc)
 450                 GOTO(out, rc);
 451
 452         rc = obd_queue_group_io(exp, lsm, NULL, oig, llap->llap_cookie,
 453                                 OBD_BRW_WRITE, 0, to, 0, ASYNC_READY |
 454                                 ASYNC_URGENT | ASYNC_COUNT_STABLE |
 455                                 ASYNC_GROUP_SYNC);
 456         if (rc)
 457                 GOTO(free_oig, rc);
 458
 459         rc = obd_trigger_group_io(exp, lsm, NULL, oig);
 460         if (rc)
 461                 GOTO(free_oig, rc);
 462
 463         rc = oig_wait(oig);
 464
 465         if (!rc && async_flags & ASYNC_READY)
 466                 unlock_page(llap->llap_page);
 467
 468         LL_CDEBUG_PAGE(D_PAGE, llap->llap_page,
 469                        "sync write returned %d\n", rc);
 470
 471         EXIT;
 472 free_oig:
 473         oig_release(oig);
 474 out:
 475         return rc;
 476 }
 477
 478 /* be careful not to return success without setting the page Uptodate or
 479  * the next pass through prepare_write will read in stale data from disk. */
 480 int ll_commit_write(struct file *file, struct page *page, unsigned from,
 481                     unsigned to)
 482 {
 483         struct inode *inode = page->mapping->host;
 484         struct ll_inode_info *lli = ll_i2info(inode);
 485         struct lov_stripe_md *lsm = lli->lli_smd;
 486         struct obd_export *exp = NULL;
 487         struct ll_async_page *llap;
 488         loff_t size;
 489         int rc = 0;
 490         ENTRY;
 491
 492         SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
 493         LASSERT(inode == file->f_dentry->d_inode);
 494         LASSERT(PageLocked(page));
 495
 496         CDEBUG(D_INODE, "inode %p is writing page %p from %d to %d at %lu\n",
 497                inode, page, from, to, page->index);
 498
 499         llap = llap_from_page(page, LLAP_ORIGIN_COMMIT_WRITE);
 500         if (IS_ERR(llap))
 501                 RETURN(PTR_ERR(llap));
 502
 503         exp = ll_i2dtexp(inode);
 504         if (exp == NULL)
 505                 RETURN(-EINVAL);
 506
 507         /* queue a write for some time in the future the first time we
 508          * dirty the page */
 509         if (!PageDirty(page)) {
 510                 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats,
 511                                      LPROC_LL_DIRTY_MISSES);
 512
 513                 rc = queue_or_sync_write(exp, ll_i2info(inode)->lli_smd,
 514                                          llap, to, 0);
 515                 if (rc)
 516                         GOTO(out, rc);
 517         } else {
 518                 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats,
 519                                      LPROC_LL_DIRTY_HITS);
 520         }
 521
 522         /* put the page in the page cache, from now on ll_removepage is
 523          * responsible for cleaning up the llap.
 524          * don't dirty the page if it has been write out in q_o_s_w */
 525         if (llap->llap_write_queued)
 526                 set_page_dirty(page);
 527         EXIT;
 528 out:
 529         size = (((obd_off)page->index) << PAGE_SHIFT) + to;
 530         down(&lli->lli_size_sem);
 531         if (rc == 0) {
 532                 obd_adjust_kms(exp, lsm, size, 0);
 533                 if (size > inode->i_size)
 534                         inode->i_size = size;
 535                 SetPageUptodate(page);
 536         } else if (size > inode->i_size) {
 537                 /* this page beyond the pales of i_size, so it can't be
 538                  * truncated in ll_p_r_e during lock revoking. we must
 539                  * teardown our book-keeping here. */
 540                 ll_removepage(page);
 541         }
 542         up(&lli->lli_size_sem);
 543         return rc;
 544 }
 545
 546 static unsigned long ll_ra_count_get(struct ll_sb_info *sbi, unsigned long len)
 547 {
 548         struct ll_ra_info *ra = &sbi->ll_ra_info;
 549         unsigned long ret;
 550         ENTRY;
 551
 552         spin_lock(&sbi->ll_lock);
 553         ret = min(ra->ra_max_pages - ra->ra_cur_pages, len);
 554         ra->ra_cur_pages += ret;
 555         spin_unlock(&sbi->ll_lock);
 556
 557         RETURN(ret);
 558 }
 559
 560 static void ll_ra_count_put(struct ll_sb_info *sbi, unsigned long len)
 561 {
 562         struct ll_ra_info *ra = &sbi->ll_ra_info;
 563         spin_lock(&sbi->ll_lock);
 564         LASSERTF(ra->ra_cur_pages >= len, "r_c_p %lu len %lu\n",
 565                  ra->ra_cur_pages, len);
 566         ra->ra_cur_pages -= len;
 567         spin_unlock(&sbi->ll_lock);
 568 }
 569
 570 int ll_writepage(struct page *page)
 571 {
 572         struct inode *inode = page->mapping->host;
 573         struct obd_export *exp;
 574         struct ll_async_page *llap;
 575         int rc = 0;
 576         ENTRY;
 577
 578         LASSERT(!PageDirty(page));
 579         LASSERT(PageLocked(page));
 580
 581         exp = ll_i2dtexp(inode);
 582         if (exp == NULL)
 583                 GOTO(out, rc = -EINVAL);
 584
 585         llap = llap_from_page(page, LLAP_ORIGIN_WRITEPAGE);
 586         if (IS_ERR(llap))
 587                 GOTO(out, rc = PTR_ERR(llap));
 588
 589         page_cache_get(page);
 590         if (llap->llap_write_queued) {
 591                 LL_CDEBUG_PAGE(D_PAGE, page, "marking urgent\n");
 592                 rc = obd_set_async_flags(exp, ll_i2info(inode)->lli_smd, NULL,
 593                                          llap->llap_cookie,
 594                                          ASYNC_READY | ASYNC_URGENT);
 595         } else {
 596                 rc = queue_or_sync_write(exp, ll_i2info(inode)->lli_smd, llap,
 597                                          PAGE_SIZE, ASYNC_READY |
 598                                          ASYNC_URGENT);
 599         }
 600         if (rc)
 601                 page_cache_release(page);
 602         EXIT;
 603 out:
 604         if (rc)
 605                 unlock_page(page);
 606         return rc;
 607 }
 608
 609 /* called for each page in a completed rpc.*/
 610 void ll_ap_completion(void *data, int cmd, struct obdo *oa, int rc)
 611 {
 612         struct ll_async_page *llap;
 613         struct page *page;
 614         ENTRY;
 615
 616         llap = LLAP_FROM_COOKIE(data);
 617         page = llap->llap_page;
 618         LASSERT(PageLocked(page));
 619
 620         LL_CDEBUG_PAGE(D_PAGE, page, "completing cmd %d with %d\n", cmd, rc);
 621
 622         if (cmd == OBD_BRW_READ && llap->llap_defer_uptodate)
 623                 ll_ra_count_put(ll_i2sbi(page->mapping->host), 1);
 624
 625         if (rc == 0)  {
 626                 if (cmd == OBD_BRW_READ) {
 627                         if (!llap->llap_defer_uptodate)
 628                                 SetPageUptodate(page);
 629                 } else {
 630                         llap->llap_write_queued = 0;
 631                 }
 632                 ClearPageError(page);
 633         } else {
 634                 if (cmd == OBD_BRW_READ)
 635                         llap->llap_defer_uptodate = 0;
 636                 SetPageError(page);
 637         }
 638
 639         unlock_page(page);
 640
 641         if (0 && cmd == OBD_BRW_WRITE) {
 642                 llap_write_complete(page->mapping->host, llap);
 643                 ll_try_done_writing(page->mapping->host);
 644         }
 645
 646         if (PageWriteback(page)) {
 647                 end_page_writeback(page);
 648         }
 649         page_cache_release(page);
 650         EXIT;
 651 }
 652
 653 /* the kernel calls us here when a page is unhashed from the page cache.
 654  * the page will be locked and the kernel is holding a spinlock, so
 655  * we need to be careful.  we're just tearing down our book-keeping
 656  * here. */
 657 void ll_removepage(struct page *page)
 658 {
 659         struct inode *inode = page->mapping->host;
 660         struct obd_export *exp;
 661         struct ll_async_page *llap;
 662         struct ll_sb_info *sbi = ll_i2sbi(inode);
 663         int rc;
 664         ENTRY;
 665
 666         LASSERT(!in_interrupt());
 667
 668         /* sync pages or failed read pages can leave pages in the page
 669          * cache that don't have our data associated with them anymore */
 670         if (page->private == 0) {
 671                 EXIT;
 672                 return;
 673         }
 674
 675         LL_CDEBUG_PAGE(D_PAGE, page, "being evicted\n");
 676
 677         exp = ll_i2dtexp(inode);
 678         if (exp == NULL) {
 679                 CERROR("page %p ind %lu gave null export\n", page, page->index);
 680                 EXIT;
 681                 return;
 682         }
 683
 684         llap = llap_from_page(page, 0);
 685         if (IS_ERR(llap)) {
 686                 CERROR("page %p ind %lu couldn't find llap: %ld\n", page,
 687                        page->index, PTR_ERR(llap));
 688                 EXIT;
 689                 return;
 690         }
 691
 692         //llap_write_complete(inode, llap);
 693         rc = obd_teardown_async_page(exp, ll_i2info(inode)->lli_smd, NULL,
 694                                      llap->llap_cookie);
 695         if (rc != 0)
 696                 CERROR("page %p ind %lu failed: %d\n", page, page->index, rc);
 697
 698         /* this unconditional free is only safe because the page lock
 699          * is providing exclusivity to memory pressure/truncate/writeback..*/
 700         __clear_page_ll_data(page);
 701
 702         spin_lock(&sbi->ll_lock);
 703         if (!list_empty(&llap->llap_proc_item))
 704                 list_del_init(&llap->llap_proc_item);
 705         sbi->ll_pglist_gen++;
 706         spin_unlock(&sbi->ll_lock);
 707         OBD_FREE(llap, sizeof(*llap));
 708         EXIT;
 709 }
 710
 711 static int ll_page_matches(struct page *page, int fd_flags)
 712 {
 713         struct lustre_handle match_lockh = {0};
 714         struct inode *inode = page->mapping->host;
 715         ldlm_policy_data_t page_extent;
 716         int flags, matches;
 717         ENTRY;
 718
 719         if (fd_flags & LL_FILE_GROUP_LOCKED)
 720                 RETURN(1);
 721
 722         page_extent.l_extent.start = (__u64)page->index << PAGE_CACHE_SHIFT;
 723         page_extent.l_extent.end =
 724                 page_extent.l_extent.start + PAGE_CACHE_SIZE - 1;
 725         flags = LDLM_FL_CBPENDING | LDLM_FL_BLOCK_GRANTED | LDLM_FL_TEST_LOCK;
 726         matches = obd_match(ll_i2sbi(inode)->ll_dt_exp,
 727                             ll_i2info(inode)->lli_smd, LDLM_EXTENT,
 728                             &page_extent, LCK_PR | LCK_PW, &flags, inode,
 729                             &match_lockh);
 730         RETURN(matches);
 731 }
 732
 733 static int ll_issue_page_read(struct obd_export *exp,
 734                               struct ll_async_page *llap,
 735                               struct obd_io_group *oig, int defer)
 736 {
 737         struct page *page = llap->llap_page;
 738         int rc;
 739
 740         page_cache_get(page);
 741         llap->llap_defer_uptodate = defer;
 742         llap->llap_ra_used = 0;
 743         rc = obd_queue_group_io(exp, ll_i2info(page->mapping->host)->lli_smd,
 744                                 NULL, oig, llap->llap_cookie, OBD_BRW_READ, 0,
 745                                 PAGE_SIZE, 0, ASYNC_COUNT_STABLE | ASYNC_READY
 746                                               | ASYNC_URGENT);
 747         if (rc) {
 748                 LL_CDEBUG_PAGE(D_ERROR, page, "read queue failed: rc %d\n", rc);
 749                 page_cache_release(page);
 750         }
 751         RETURN(rc);
 752 }
 753
 754 static void ll_ra_stats_inc_unlocked(struct ll_ra_info *ra, enum ra_stat which)
 755 {
 756         LASSERTF(which >= 0 && which < _NR_RA_STAT, "which: %u\n", which);
 757         ra->ra_stats[which]++;
 758 }
 759
 760 static void ll_ra_stats_inc(struct address_space *mapping, enum ra_stat which)
 761 {
 762         struct ll_sb_info *sbi = ll_i2sbi(mapping->host);
 763         struct ll_ra_info *ra = &ll_i2sbi(mapping->host)->ll_ra_info;
 764
 765         spin_lock(&sbi->ll_lock);
 766         ll_ra_stats_inc_unlocked(ra, which);
 767         spin_unlock(&sbi->ll_lock);
 768 }
 769
 770 void ll_ra_accounting(struct page *page, struct address_space *mapping)
 771 {
 772         struct ll_async_page *llap;
 773
 774         llap = llap_from_page(page, LLAP_ORIGIN_WRITEPAGE);
 775         if (IS_ERR(llap))
 776                 return;
 777
 778         if (!llap->llap_defer_uptodate || llap->llap_ra_used)
 779                 return;
 780
 781         ll_ra_stats_inc(mapping, RA_STAT_DISCARDED);
 782 }
 783
 784 #define RAS_CDEBUG(ras) \
 785         CDEBUG(D_READA, "lrp %lu c %lu ws %lu wl %lu nra %lu\n",        \
 786                ras->ras_last_readpage, ras->ras_consecutive,            \
 787                ras->ras_window_start, ras->ras_window_len,              \
 788                ras->ras_next_readahead);
 789
 790 static int index_in_window(unsigned long index, unsigned long point,
 791                            unsigned long before, unsigned long after)
 792 {
 793         unsigned long start = point - before, end = point + after;
 794
 795         if (start > point)
 796                start = 0;
 797         if (end < point)
 798                end = ~0;
 799
 800         return start <= index && index <= end;
 801 }
 802
 803 static int ll_readahead(struct ll_readahead_state *ras,
 804                          struct obd_export *exp, struct address_space *mapping,
 805                          struct obd_io_group *oig, int flags)
 806 {
 807         unsigned long i, start = 0, end = 0, reserved;
 808         struct ll_async_page *llap;
 809         struct page *page;
 810         int rc, ret = 0, match_failed = 0;
 811         __u64 kms;
 812         ENTRY;
 813
 814         kms = lov_merge_size(ll_i2info(mapping->host)->lli_smd, 1);
 815         if (kms == 0) {
 816                 ll_ra_stats_inc(mapping, RA_STAT_ZERO_LEN);
 817                 RETURN(0);
 818         }
 819         spin_lock(&ras->ras_lock);
 820
 821         /* reserve a part of the read-ahead window that we'll be issuing */
 822         if (ras->ras_window_len) {
 823                 start = ras->ras_next_readahead;
 824                 end = ras->ras_window_start + ras->ras_window_len - 1;
 825                 end = min(end, (unsigned long)(kms >> PAGE_CACHE_SHIFT));
 826                 ras->ras_next_readahead = max(end, end + 1);
 827
 828                 RAS_CDEBUG(ras);
 829         }
 830
 831         spin_unlock(&ras->ras_lock);
 832
 833         if (end == 0) {
 834                 ll_ra_stats_inc(mapping, RA_STAT_ZERO_WINDOW);
 835                 RETURN(0);
 836         }
 837
 838         reserved = ll_ra_count_get(ll_i2sbi(mapping->host), end - start + 1);
 839         if (reserved < end - start + 1)
 840                 ll_ra_stats_inc(mapping, RA_STAT_MAX_IN_FLIGHT);
 841
 842         for (i = start; reserved > 0 && !match_failed && i <= end; i++) {
 843                 /* skip locked pages from previous readpage calls */
 844                 page = grab_cache_page_nowait(mapping, i);
 845                 if (page == NULL) {
 846                         CDEBUG(D_READA, "g_c_p_n failed\n");
 847                         continue;
 848                 }
 849
 850                 /* we do this first so that we can see the page in the /proc
 851                  * accounting */
 852                 llap = llap_from_page(page, LLAP_ORIGIN_READAHEAD);
 853                 if (IS_ERR(llap) || llap->llap_defer_uptodate)
 854                         goto next_page;
 855
 856                 /* skip completed pages */
 857                 if (Page_Uptodate(page))
 858                         goto next_page;
 859
 860                 /* bail when we hit the end of the lock. */
 861                 if ((rc = ll_page_matches(page, flags)) <= 0) {
 862                         LL_CDEBUG_PAGE(D_READA | D_PAGE, page,
 863                                        "lock match failed: rc %d\n", rc);
 864                         ll_ra_stats_inc(mapping, RA_STAT_FAILED_MATCH);
 865                         match_failed = 1;
 866                         goto next_page;
 867                 }
 868
 869                 rc = ll_issue_page_read(exp, llap, oig, 1);
 870                 if (rc == 0) {
 871                         reserved--;
 872                         ret++;
 873                         LL_CDEBUG_PAGE(D_READA| D_PAGE, page,
 874                                        "started read-ahead\n");
 875                 }
 876                 if (rc) {
 877         next_page:
 878                         LL_CDEBUG_PAGE(D_READA | D_PAGE, page,
 879                                        "skipping read-ahead\n");
 880
 881                         unlock_page(page);
 882                 }
 883                 page_cache_release(page);
 884         }
 885
 886         LASSERTF(reserved >= 0, "reserved %lu\n", reserved);
 887         if (reserved != 0)
 888                 ll_ra_count_put(ll_i2sbi(mapping->host), reserved);
 889
 890         if (i == end + 1 && end == (kms >> PAGE_CACHE_SHIFT))
 891                 ll_ra_stats_inc(mapping, RA_STAT_EOF);
 892
 893         /* if we didn't get to the end of the region we reserved from
 894          * the ras we need to go back and update the ras so that the
 895          * next read-ahead tries from where we left off.  we only do so
 896          * if the region we failed to issue read-ahead on is still ahead
 897          * of the app and behind the next index to start read-ahead from */
 898         if (i != end + 1) {
 899                 spin_lock(&ras->ras_lock);
 900                 if (i < ras->ras_next_readahead &&
 901                     index_in_window(i, ras->ras_window_start, 0,
 902                                     ras->ras_window_len)) {
 903                         ras->ras_next_readahead = i;
 904                         RAS_CDEBUG(ras);
 905                 }
 906                 spin_unlock(&ras->ras_lock);
 907         }
 908
 909         RETURN(ret);
 910 }
 911
 912 static void ras_set_start(struct ll_readahead_state *ras, unsigned long index)
 913 {
 914         ras->ras_window_start = index & (~(PTLRPC_MAX_BRW_PAGES - 1));
 915 }
 916
 917 /* called with the ras_lock held or from places where it doesn't matter */
 918 static void ras_reset(struct ll_readahead_state *ras, unsigned long index)
 919 {
 920         ras->ras_last_readpage = index;
 921         ras->ras_consecutive = 1;
 922         ras->ras_window_len = 0;
 923         ras_set_start(ras, index);
 924         ras->ras_next_readahead = ras->ras_window_start;
 925
 926         RAS_CDEBUG(ras);
 927 }
 928
 929 void ll_readahead_init(struct inode *inode, struct ll_readahead_state *ras)
 930 {
 931         spin_lock_init(&ras->ras_lock);
 932         ras_reset(ras, 0);
 933 }
 934
 935 static void ras_update(struct ll_sb_info *sbi, struct ll_readahead_state *ras,
 936                        unsigned long index, unsigned hit)
 937 {
 938         struct ll_ra_info *ra = &sbi->ll_ra_info;
 939         int zero = 0;
 940         ENTRY;
 941
 942         spin_lock(&sbi->ll_lock);
 943         spin_lock(&ras->ras_lock);
 944
 945         ll_ra_stats_inc_unlocked(ra, hit ? RA_STAT_HIT : RA_STAT_MISS);
 946
 947         /* reset the read-ahead window in two cases.  First when the app seeks
 948          * or reads to some other part of the file.  Secondly if we get a
 949          * read-ahead miss that we think we've previously issued.  This can
 950          * be a symptom of there being so many read-ahead pages that the VM is
 951          * reclaiming it before we get to it. */
 952         if (!index_in_window(index, ras->ras_last_readpage, 8, 8)) {
 953                 zero = 1;
 954                 ll_ra_stats_inc_unlocked(ra, RA_STAT_DISTANT_READPAGE);
 955         } else if (!hit && ras->ras_window_len &&
 956                    index < ras->ras_next_readahead &&
 957                    index_in_window(index, ras->ras_window_start, 0,
 958                                    ras->ras_window_len)) {
 959                 zero = 1;
 960                 ll_ra_stats_inc_unlocked(ra, RA_STAT_MISS_IN_WINDOW);
 961         }
 962
 963         if (zero) {
 964                 ras_reset(ras, index);
 965                 GOTO(out_unlock, 0);
 966         }
 967
 968         ras->ras_last_readpage = index;
 969         ras->ras_consecutive++;
 970         ras_set_start(ras, index);
 971         ras->ras_next_readahead = max(ras->ras_window_start,
 972                                       ras->ras_next_readahead);
 973
 974         /* wait for a few pages to arrive before issuing readahead to avoid
 975          * the worst overutilization */
 976         if (ras->ras_consecutive == 3) {
 977                 ras->ras_window_len = PTLRPC_MAX_BRW_PAGES;
 978                 GOTO(out_unlock, 0);
 979         }
 980
 981         /* we need to increase the window sometimes.  we'll arbitrarily
 982          * do it half-way through the pages in an rpc */
 983         if ((index & (PTLRPC_MAX_BRW_PAGES - 1)) ==
 984             (PTLRPC_MAX_BRW_PAGES >> 1)) {
 985                 ras->ras_window_len += PTLRPC_MAX_BRW_PAGES;
 986                 ras->ras_window_len = min(ras->ras_window_len,
 987                                           ra->ra_max_pages);
 988
 989         }
 990
 991         EXIT;
 992 out_unlock:
 993         RAS_CDEBUG(ras);
 994         spin_unlock(&ras->ras_lock);
 995         spin_unlock(&sbi->ll_lock);
 996 }
 997
 998 /*
 999  * for now we do our readpage the same on both 2.4 and 2.5.  The kernel's
1000  * read-ahead assumes it is valid to issue readpage all the way up to
1001  * i_size, but our dlm locks make that not the case.  We disable the
1002  * kernel's read-ahead and do our own by walking ahead in the page cache
1003  * checking for dlm lock coverage.  the main difference between 2.4 and
1004  * 2.6 is how read-ahead gets batched and issued, but we're using our own,
1005  * so they look the same.
1006  */
1007 int ll_readpage(struct file *filp, struct page *page)
1008 {
1009         struct ll_file_data *fd = filp->private_data;
1010         struct inode *inode = page->mapping->host;
1011         struct obd_export *exp;
1012         struct ll_async_page *llap;
1013         struct obd_io_group *oig = NULL;
1014         int rc;
1015         ENTRY;
1016
1017         LASSERT(PageLocked(page));
1018         LASSERT(!PageUptodate(page));
1019         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),offset="LPX64"\n",
1020                inode->i_ino, inode->i_generation, inode,
1021                (((obd_off)page->index) << PAGE_SHIFT));
1022         LASSERT(atomic_read(&filp->f_dentry->d_inode->i_count) > 0);
1023
1024         rc = oig_init(&oig);
1025         if (rc < 0)
1026                 GOTO(out, rc);
1027
1028         exp = ll_i2dtexp(inode);
1029         if (exp == NULL)
1030                 GOTO(out, rc = -EINVAL);
1031
1032         llap = llap_from_page(page, LLAP_ORIGIN_READPAGE);
1033         if (IS_ERR(llap))
1034                 GOTO(out, rc = PTR_ERR(llap));
1035
1036         if (ll_i2sbi(inode)->ll_flags & LL_SBI_READAHEAD)
1037                 ras_update(ll_i2sbi(inode), &fd->fd_ras, page->index,
1038                            llap->llap_defer_uptodate);
1039
1040         if (llap->llap_defer_uptodate) {
1041                 llap->llap_ra_used = 1;
1042                 rc = ll_readahead(&fd->fd_ras, exp, page->mapping, oig,
1043                                   fd->fd_flags);
1044                 if (rc > 0)
1045                         obd_trigger_group_io(exp, ll_i2info(inode)->lli_smd,
1046                                              NULL, oig);
1047                 LL_CDEBUG_PAGE(D_PAGE, page, "marking uptodate from defer\n");
1048                 SetPageUptodate(page);
1049                 unlock_page(page);
1050                 GOTO(out_oig, rc = 0);
1051         }
1052
1053         rc = ll_page_matches(page, fd->fd_flags);
1054         if (rc < 0) {
1055                 LL_CDEBUG_PAGE(D_ERROR, page, "lock match failed: rc %d\n", rc);
1056                 GOTO(out, rc);
1057         }
1058
1059         if (rc == 0) {
1060                 CWARN("ino %lu page %lu (%llu) not covered by "
1061                       "a lock (mmap?).  check debug logs.\n",
1062                       inode->i_ino, page->index,
1063                       (long long)page->index << PAGE_CACHE_SHIFT);
1064         }
1065
1066         rc = ll_issue_page_read(exp, llap, oig, 0);
1067         if (rc)
1068                 GOTO(out, rc);
1069
1070         LL_CDEBUG_PAGE(D_PAGE, page, "queued readpage\n");
1071         if (ll_i2sbi(inode)->ll_flags & LL_SBI_READAHEAD)
1072                 ll_readahead(&fd->fd_ras, exp, page->mapping, oig,
1073                              fd->fd_flags);
1074
1075         rc = obd_trigger_group_io(exp, ll_i2info(inode)->lli_smd, NULL, oig);
1076         EXIT;
1077 out:
1078         if (rc)
1079                 unlock_page(page);
1080 out_oig:
1081         if (oig != NULL)
1082                 oig_release(oig);
1083         return rc;
1084 }