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 /* this isn't where truncate starts.   roughly:
 110  * sys_truncate->ll_setattr_raw->vmtruncate->ll_truncate
 111  * we grab the lock back in setattr_raw to avoid races. */
 112 void ll_truncate(struct inode *inode)
 113 {
 114         struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
 115         struct obdo *oa = NULL;
 116         int rc;
 117         ENTRY;
 118
 119         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p)\n", inode->i_ino,
 120                inode->i_generation, inode);
 121
 122         if (!lsm) {
 123                 CDEBUG(D_INODE, "truncate on inode %lu with no objects\n",
 124                        inode->i_ino);
 125                 EXIT;
 126                 return;
 127         }
 128
 129         if (lov_merge_size(lsm, 0) == inode->i_size) {
 130                 CDEBUG(D_VFSTRACE, "skipping punch for "LPX64" (size = %llu)\n",
 131                        lsm->lsm_object_id, inode->i_size);
 132         } else {
 133                 CDEBUG(D_INFO, "calling punch for "LPX64" (new size %llu)\n",
 134                        lsm->lsm_object_id, inode->i_size);
 135
 136                 oa = obdo_alloc();
 137                 if (oa == NULL) {
 138                         CERROR("cannot alloc oa, error %d\n",
 139                                 -ENOMEM);
 140                         EXIT;
 141                         return;
 142                 }
 143
 144                 oa->o_id = lsm->lsm_object_id;
 145                 oa->o_gr = lsm->lsm_object_gr;
 146                 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
 147                 obdo_from_inode(oa, inode, OBD_MD_FLTYPE | OBD_MD_FLMODE |
 148                                 OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME);
 149
 150                 /* truncate == punch from new size to absolute end of file */
 151                 /* NB: obd_punch must be called with i_sem held!  It updates the kms! */
 152                 rc = obd_punch(ll_i2dtexp(inode), oa, lsm, inode->i_size,
 153                                OBD_OBJECT_EOF, NULL);
 154                 if (rc)
 155                         CERROR("obd_truncate fails (%d) ino %lu\n", rc, inode->i_ino);
 156                 else
 157                         obdo_to_inode(inode, oa, OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
 158                                       OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME);
 159
 160                 obdo_free(oa);
 161         }
 162
 163         EXIT;
 164         return;
 165 } /* ll_truncate */
 166
 167 int ll_prepare_write(struct file *file, struct page *page, unsigned from,
 168                      unsigned to)
 169 {
 170         struct inode *inode = page->mapping->host;
 171         struct ll_inode_info *lli = ll_i2info(inode);
 172         struct lov_stripe_md *lsm = lli->lli_smd;
 173         obd_off offset = ((obd_off)page->index) << PAGE_SHIFT;
 174         struct obdo *oa = NULL;
 175         struct brw_page pga;
 176         __u64 kms;
 177         int rc = 0;
 178         ENTRY;
 179
 180         LASSERT(PageLocked(page));
 181         (void)llap_cast_private(page); /* assertion */
 182
 183         /* Check to see if we should return -EIO right away */
 184         pga.pg = page;
 185         pga.disk_offset = pga.page_offset = offset;
 186         pga.count = PAGE_SIZE;
 187         pga.flag = 0;
 188
 189         oa = obdo_alloc();
 190         if (oa == NULL)
 191                 RETURN(-ENOMEM);
 192
 193         oa->o_id = lsm->lsm_object_id;
 194         oa->o_gr = lsm->lsm_object_gr;
 195         oa->o_mode = inode->i_mode;
 196         oa->o_valid = OBD_MD_FLID | OBD_MD_FLMODE |
 197                 OBD_MD_FLTYPE | OBD_MD_FLGROUP;
 198
 199         rc = obd_brw(OBD_BRW_CHECK, ll_i2dtexp(inode), oa, lsm,
 200                      1, &pga, NULL);
 201         if (rc)
 202                 GOTO(out_free_oa, rc);
 203
 204         if (PageUptodate(page))
 205                 GOTO(out_free_oa, 0);
 206
 207         /* We're completely overwriting an existing page, so _don't_ set it up
 208          * to date until commit_write */
 209         if (from == 0 && to == PAGE_SIZE) {
 210                 POISON_PAGE(page, 0x11);
 211                 GOTO(out_free_oa, 0);
 212         }
 213
 214         /* If are writing to a new page, no need to read old data.  The extent
 215          * locking will have updated the KMS, and for our purposes here we can
 216          * treat it like i_size. */
 217         kms = lov_merge_size(lsm, 1);
 218         if (kms <= offset) {
 219                 memset(kmap(page), 0, PAGE_SIZE);
 220                 kunmap(page);
 221                 GOTO(prepare_done, rc = 0);
 222         }
 223
 224         /* XXX could be an async ocp read.. read-ahead? */
 225         rc = ll_brw(OBD_BRW_READ, inode, oa, page, 0);
 226         if (rc == 0) {
 227                 /* bug 1598: don't clobber blksize */
 228                 oa->o_valid &= ~(OBD_MD_FLSIZE | OBD_MD_FLBLKSZ);
 229                 obdo_refresh_inode(inode, oa, oa->o_valid);
 230         }
 231
 232         EXIT;
 233 prepare_done:
 234         if (rc == 0)
 235                 SetPageUptodate(page);
 236 out_free_oa:
 237         obdo_free(oa);
 238         return rc;
 239 }
 240
 241 struct ll_async_page *llap_from_cookie(void *cookie)
 242 {
 243         struct ll_async_page *llap = cookie;
 244         if (llap->llap_magic != LLAP_MAGIC)
 245                 return ERR_PTR(-EINVAL);
 246         return llap;
 247 };
 248
 249 static int ll_ap_make_ready(void *data, int cmd)
 250 {
 251         struct ll_async_page *llap;
 252         struct page *page;
 253         ENTRY;
 254
 255         llap = llap_from_cookie(data);
 256         if (IS_ERR(llap))
 257                 RETURN(-EINVAL);
 258
 259         page = llap->llap_page;
 260
 261         LASSERT(cmd != OBD_BRW_READ);
 262
 263         /* we're trying to write, but the page is locked.. come back later */
 264         if (TryLockPage(page))
 265                 RETURN(-EAGAIN);
 266
 267         LL_CDEBUG_PAGE(D_PAGE, page, "made ready\n");
 268         page_cache_get(page);
 269
 270         /* if we left PageDirty we might get another writepage call
 271          * in the future.  list walkers are bright enough
 272          * to check page dirty so we can leave it on whatever list
 273          * its on.  XXX also, we're called with the cli list so if
 274          * we got the page cache list we'd create a lock inversion
 275          * with the removepage path which gets the page lock then the
 276          * cli lock */
 277         clear_page_dirty(page);
 278         RETURN(0);
 279 }
 280
 281 /* We have two reasons for giving llite the opportunity to change the
 282  * write length of a given queued page as it builds the RPC containing
 283  * the page:
 284  *
 285  * 1) Further extending writes may have landed in the page cache
 286  *    since a partial write first queued this page requiring us
 287  *    to write more from the page cache.
 288  * 2) We might have raced with truncate and want to avoid performing
 289  *    write RPCs that are just going to be thrown away by the
 290  *    truncate's punch on the storage targets.
 291  *
 292  * The kms serves these purposes as it is set at both truncate and extending
 293  * writes.
 294  */
 295 static int ll_ap_refresh_count(void *data, int cmd)
 296 {
 297         struct ll_async_page *llap;
 298         struct lov_stripe_md *lsm;
 299         struct page *page;
 300         __u64 kms;
 301         ENTRY;
 302
 303         /* readpage queues with _COUNT_STABLE, shouldn't get here. */
 304         LASSERT(cmd != OBD_BRW_READ);
 305
 306         llap = llap_from_cookie(data);
 307         if (IS_ERR(llap))
 308                 RETURN(PTR_ERR(llap));
 309
 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         if (IS_ERR(llap)) {
 356                 EXIT;
 357                 return;
 358         }
 359
 360         ll_inode_fill_obdo(llap->llap_page->mapping->host, cmd, oa);
 361         EXIT;
 362 }
 363
 364 static struct obd_async_page_ops ll_async_page_ops = {
 365         .ap_make_ready =        ll_ap_make_ready,
 366         .ap_refresh_count =     ll_ap_refresh_count,
 367         .ap_fill_obdo =         ll_ap_fill_obdo,
 368         .ap_completion =        ll_ap_completion,
 369 };
 370
 371 struct ll_async_page *llap_cast_private(struct page *page)
 372 {
 373         struct ll_async_page *llap = (struct ll_async_page *)page->private;
 374
 375         LASSERTF(llap == NULL || llap->llap_magic == LLAP_MAGIC,
 376                  "page %p private %lu gave magic %d which != %d\n",
 377                  page, page->private, llap->llap_magic, LLAP_MAGIC);
 378
 379         return llap;
 380 }
 381
 382 /* XXX have the exp be an argument? */
 383 struct ll_async_page *llap_from_page(struct page *page)
 384 {
 385         struct ll_async_page *llap;
 386         struct obd_export *exp;
 387         struct inode *inode = page->mapping->host;
 388         struct ll_sb_info *sbi = ll_i2sbi(inode);
 389         int rc;
 390         ENTRY;
 391
 392         llap = llap_cast_private(page);
 393         if (llap != NULL)
 394                 RETURN(llap);
 395
 396         exp = ll_i2dtexp(page->mapping->host);
 397         if (exp == NULL)
 398                 RETURN(ERR_PTR(-EINVAL));
 399
 400         OBD_ALLOC(llap, sizeof(*llap));
 401         if (llap == NULL)
 402                 RETURN(ERR_PTR(-ENOMEM));
 403         llap->llap_magic = LLAP_MAGIC;
 404         rc = obd_prep_async_page(exp, ll_i2info(inode)->lli_smd, NULL, page,
 405                                  (obd_off)page->index << PAGE_SHIFT,
 406                                  &ll_async_page_ops, llap, &llap->llap_cookie);
 407         if (rc) {
 408                 OBD_FREE(llap, sizeof(*llap));
 409                 RETURN(ERR_PTR(rc));
 410         }
 411
 412         CDEBUG(D_CACHE, "llap %p page %p cookie %p obj off "LPU64"\n", llap,
 413                page, llap->llap_cookie, (obd_off)page->index << PAGE_SHIFT);
 414         /* also zeroing the PRIVBITS low order bitflags */
 415         __set_page_ll_data(page, llap);
 416         llap->llap_page = page;
 417
 418         spin_lock(&sbi->ll_lock);
 419         sbi->ll_pglist_gen++;
 420         list_add_tail(&llap->llap_proc_item, &sbi->ll_pglist);
 421         spin_unlock(&sbi->ll_lock);
 422
 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 void lov_increase_kms(struct obd_export *exp, struct lov_stripe_md *lsm,
 479                       obd_off size);
 480
 481 /* be careful not to return success without setting the page Uptodate or
 482  * the next pass through prepare_write will read in stale data from disk. */
 483 int ll_commit_write(struct file *file, struct page *page, unsigned from,
 484                     unsigned to)
 485 {
 486         struct inode *inode = page->mapping->host;
 487         struct ll_inode_info *lli = ll_i2info(inode);
 488         struct lov_stripe_md *lsm = lli->lli_smd;
 489         struct obd_export *exp = NULL;
 490         struct ll_async_page *llap;
 491         loff_t size;
 492         int rc = 0;
 493         ENTRY;
 494
 495         SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
 496         LASSERT(inode == file->f_dentry->d_inode);
 497         LASSERT(PageLocked(page));
 498
 499         CDEBUG(D_INODE, "inode %p is writing page %p from %d to %d at %lu\n",
 500                inode, page, from, to, page->index);
 501
 502         llap = llap_from_page(page);
 503         if (IS_ERR(llap))
 504                 RETURN(PTR_ERR(llap));
 505
 506         /* queue a write for some time in the future the first time we
 507          * dirty the page */
 508         if (!PageDirty(page)) {
 509                 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats,
 510                                      LPROC_LL_DIRTY_MISSES);
 511
 512                 exp = ll_i2dtexp(inode);
 513                 if (exp == NULL)
 514                         RETURN(-EINVAL);
 515
 516                 rc = queue_or_sync_write(exp, ll_i2info(inode)->lli_smd, llap,
 517                                          to, 0);
 518                 if (rc)
 519                         GOTO(out, rc);
 520         } else {
 521                 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats,
 522                                      LPROC_LL_DIRTY_HITS);
 523         }
 524
 525         /* put the page in the page cache, from now on ll_removepage is
 526          * responsible for cleaning up the llap.
 527          * don't dirty the page if it has been write out in q_o_s_w */
 528         if (llap->llap_write_queued)
 529                 set_page_dirty(page);
 530         EXIT;
 531 out:
 532         if (rc == 0) {
 533                 size = (((obd_off)page->index) << PAGE_SHIFT) + to;
 534                 lov_increase_kms(exp, lsm, size);
 535                 if (size > inode->i_size)
 536                         inode->i_size = size;
 537                 SetPageUptodate(page);
 538         }
 539         return rc;
 540 }
 541
 542 int ll_writepage(struct page *page)
 543 {
 544         struct inode *inode = page->mapping->host;
 545         struct obd_export *exp;
 546         struct ll_async_page *llap;
 547         int rc = 0;
 548         ENTRY;
 549
 550         LASSERT(!PageDirty(page));
 551         LASSERT(PageLocked(page));
 552
 553         exp = ll_i2dtexp(inode);
 554         if (exp == NULL)
 555                 GOTO(out, rc = -EINVAL);
 556
 557         llap = llap_from_page(page);
 558         if (IS_ERR(llap))
 559                 GOTO(out, rc = PTR_ERR(llap));
 560
 561         page_cache_get(page);
 562         if (llap->llap_write_queued) {
 563                 LL_CDEBUG_PAGE(D_PAGE, page, "marking urgent\n");
 564                 rc = obd_set_async_flags(exp, ll_i2info(inode)->lli_smd, NULL,
 565                                          llap->llap_cookie,
 566                                          ASYNC_READY | ASYNC_URGENT);
 567         } else {
 568                 rc = queue_or_sync_write(exp, ll_i2info(inode)->lli_smd, llap,
 569                                          PAGE_SIZE, ASYNC_READY |
 570                                          ASYNC_URGENT);
 571         }
 572         if (rc)
 573                 page_cache_release(page);
 574         EXIT;
 575 out:
 576         if (rc)
 577                 unlock_page(page);
 578         return rc;
 579 }
 580
 581 static unsigned long
 582 ll_ra_count_get(struct ll_sb_info *sbi, unsigned long len)
 583 {
 584         struct ll_ra_info *ra = &sbi->ll_ra_info;
 585         unsigned long ret;
 586         ENTRY;
 587
 588         spin_lock(&sbi->ll_lock);
 589         ret = min(ra->ra_max_pages - ra->ra_cur_pages, len);
 590         ra->ra_cur_pages += ret;
 591         spin_unlock(&sbi->ll_lock);
 592
 593         RETURN(ret);
 594 }
 595
 596 static void ll_ra_count_put(struct ll_sb_info *sbi, unsigned long len)
 597 {
 598         struct ll_ra_info *ra = &sbi->ll_ra_info;
 599         spin_lock(&sbi->ll_lock);
 600         LASSERTF(ra->ra_cur_pages >= len, "r_c_p %lu len %lu\n",
 601                  ra->ra_cur_pages, len);
 602         ra->ra_cur_pages -= len;
 603         spin_unlock(&sbi->ll_lock);
 604 }
 605
 606 /* called for each page in a completed rpc.*/
 607 void ll_ap_completion(void *data, int cmd, struct obdo *oa, int rc)
 608 {
 609         struct ll_async_page *llap;
 610         struct page *page;
 611         ENTRY;
 612
 613         llap = llap_from_cookie(data);
 614         if (IS_ERR(llap)) {
 615                 EXIT;
 616                 return;
 617         }
 618
 619         page = llap->llap_page;
 620         LASSERT(PageLocked(page));
 621
 622         LL_CDEBUG_PAGE(D_PAGE, page, "completing cmd %d with %d\n", cmd, rc);
 623
 624         if (cmd == OBD_BRW_READ && llap->llap_defer_uptodate)
 625                 ll_ra_count_put(ll_i2sbi(page->mapping->host), 1);
 626
 627         if (rc == 0)  {
 628                 if (cmd == OBD_BRW_READ) {
 629                         if (!llap->llap_defer_uptodate)
 630                                 SetPageUptodate(page);
 631                 } else {
 632                         llap->llap_write_queued = 0;
 633                 }
 634                 ClearPageError(page);
 635         } else {
 636                 if (cmd == OBD_BRW_READ)
 637                         llap->llap_defer_uptodate = 0;
 638                 SetPageError(page);
 639         }
 640
 641         unlock_page(page);
 642
 643         if (0 && cmd == OBD_BRW_WRITE) {
 644                 llap_write_complete(page->mapping->host, llap);
 645                 ll_try_done_writing(page->mapping->host);
 646         }
 647
 648         if (PageWriteback(page)) {
 649                 end_page_writeback(page);
 650         }
 651         page_cache_release(page);
 652         EXIT;
 653 }
 654
 655 /* the kernel calls us here when a page is unhashed from the page cache.
 656  * the page will be locked and the kernel is holding a spinlock, so
 657  * we need to be careful.  we're just tearing down our book-keeping
 658  * here. */
 659 void ll_removepage(struct page *page)
 660 {
 661         struct inode *inode = page->mapping->host;
 662         struct obd_export *exp;
 663         struct ll_async_page *llap;
 664         struct ll_sb_info *sbi = ll_i2sbi(inode);
 665         int rc;
 666         ENTRY;
 667
 668         LASSERT(!in_interrupt());
 669
 670         /* sync pages or failed read pages can leave pages in the page
 671          * cache that don't have our data associated with them anymore */
 672         if (page->private == 0) {
 673                 EXIT;
 674                 return;
 675         }
 676
 677         LL_CDEBUG_PAGE(D_PAGE, page, "being evicted\n");
 678
 679         exp = ll_i2dtexp(inode);
 680         if (exp == NULL) {
 681                 CERROR("page %p ind %lu gave null export\n", page, page->index);
 682                 EXIT;
 683                 return;
 684         }
 685
 686         llap = llap_from_page(page);
 687         if (IS_ERR(llap)) {
 688                 CERROR("page %p ind %lu couldn't find llap: %ld\n", page,
 689                        page->index, PTR_ERR(llap));
 690                 EXIT;
 691                 return;
 692         }
 693
 694         //llap_write_complete(inode, llap);
 695         rc = obd_teardown_async_page(exp, ll_i2info(inode)->lli_smd, NULL,
 696                                      llap->llap_cookie);
 697         if (rc != 0)
 698                 CERROR("page %p ind %lu failed: %d\n", page, page->index, rc);
 699
 700         /* this unconditional free is only safe because the page lock
 701          * is providing exclusivity to memory pressure/truncate/writeback..*/
 702         __clear_page_ll_data(page);
 703
 704         spin_lock(&sbi->ll_lock);
 705         if (!list_empty(&llap->llap_proc_item))
 706                 list_del_init(&llap->llap_proc_item);
 707         sbi->ll_pglist_gen++;
 708         spin_unlock(&sbi->ll_lock);
 709         OBD_FREE(llap, sizeof(*llap));
 710         EXIT;
 711 }
 712
 713 static int ll_page_matches(struct page *page, int fd_flags)
 714 {
 715         struct lustre_handle match_lockh = {0};
 716         struct inode *inode = page->mapping->host;
 717         ldlm_policy_data_t page_extent;
 718         int flags, matches;
 719         ENTRY;
 720
 721         if (fd_flags & LL_FILE_GROUP_LOCKED)
 722                 RETURN(1);
 723
 724         page_extent.l_extent.start = (__u64)page->index << PAGE_CACHE_SHIFT;
 725         page_extent.l_extent.end =
 726                 page_extent.l_extent.start + PAGE_CACHE_SIZE - 1;
 727         flags = LDLM_FL_CBPENDING | LDLM_FL_BLOCK_GRANTED | LDLM_FL_TEST_LOCK;
 728         matches = obd_match(ll_i2sbi(inode)->ll_dt_exp,
 729                             ll_i2info(inode)->lli_smd, LDLM_EXTENT,
 730                             &page_extent, LCK_PR | LCK_PW, &flags, inode,
 731                             &match_lockh);
 732         RETURN(matches);
 733 }
 734
 735 static int ll_issue_page_read(struct obd_export *exp,
 736                               struct ll_async_page *llap,
 737                               struct obd_io_group *oig, int defer)
 738 {
 739         struct page *page = llap->llap_page;
 740         int rc;
 741
 742         page_cache_get(page);
 743         llap->llap_defer_uptodate = defer;
 744         llap->llap_ra_used = 0;
 745         rc = obd_queue_group_io(exp, ll_i2info(page->mapping->host)->lli_smd,
 746                                 NULL, oig, llap->llap_cookie, OBD_BRW_READ, 0,
 747                                 PAGE_SIZE, 0, ASYNC_COUNT_STABLE | ASYNC_READY
 748                                               | ASYNC_URGENT);
 749         if (rc) {
 750                 LL_CDEBUG_PAGE(D_ERROR, page, "read queue failed: rc %d\n", rc);
 751                 page_cache_release(page);
 752         }
 753         RETURN(rc);
 754 }
 755
 756 static void ll_ra_stats_inc_unlocked(struct ll_ra_info *ra, enum ra_stat which)
 757 {
 758         LASSERTF(which >= 0 && which < _NR_RA_STAT, "which: %u\n", which);
 759         ra->ra_stats[which]++;
 760 }
 761
 762 static void ll_ra_stats_inc(struct address_space *mapping, enum ra_stat which)
 763 {
 764         struct ll_sb_info *sbi = ll_i2sbi(mapping->host);
 765         struct ll_ra_info *ra = &ll_i2sbi(mapping->host)->ll_ra_info;
 766
 767         spin_lock(&sbi->ll_lock);
 768         ll_ra_stats_inc_unlocked(ra, which);
 769         spin_unlock(&sbi->ll_lock);
 770 }
 771
 772 void ll_ra_accounting(struct page *page, struct address_space *mapping)
 773 {
 774         struct ll_async_page *llap;
 775
 776         llap = llap_from_page(page);
 777         if (IS_ERR(llap))
 778                 return;
 779
 780         if (!llap->llap_defer_uptodate || llap->llap_ra_used)
 781                 return;
 782
 783         ll_ra_stats_inc(mapping, RA_STAT_DISCARDED);
 784 }
 785
 786 #define RAS_CDEBUG(ras) \
 787         CDEBUG(D_READA, "lrp %lu c %lu ws %lu wl %lu nra %lu\n",        \
 788                ras->ras_last_readpage, ras->ras_consecutive,            \
 789                ras->ras_window_start, ras->ras_window_len,              \
 790                ras->ras_next_readahead);
 791
 792 static int index_in_window(unsigned long index, unsigned long point,
 793                            unsigned long before, unsigned long after)
 794 {
 795         unsigned long start = point - before, end = point + after;
 796
 797         if (start > point)
 798                start = 0;
 799         if (end < point)
 800                end = ~0;
 801
 802         return start <= index && index <= end;
 803 }
 804
 805 static int ll_readahead(struct ll_readahead_state *ras,
 806                          struct obd_export *exp, struct address_space *mapping,
 807                          struct obd_io_group *oig, int flags)
 808 {
 809         unsigned long i, start = 0, end = 0, reserved;
 810         struct ll_async_page *llap;
 811         struct page *page;
 812         int rc, ret = 0, match_failed = 0;
 813         __u64 kms;
 814         ENTRY;
 815
 816         kms = lov_merge_size(ll_i2info(mapping->host)->lli_smd, 1);
 817         if (kms == 0) {
 818                 ll_ra_stats_inc(mapping, RA_STAT_ZERO_LEN);
 819                 RETURN(0);
 820         }
 821         spin_lock(&ras->ras_lock);
 822
 823         /* reserve a part of the read-ahead window that we'll be issuing */
 824         if (ras->ras_window_len) {
 825                 start = ras->ras_next_readahead;
 826                 end = ras->ras_window_start + ras->ras_window_len - 1;
 827                 end = min(end, (unsigned long)(kms >> PAGE_CACHE_SHIFT));
 828                 ras->ras_next_readahead = max(end, end + 1);
 829
 830                 RAS_CDEBUG(ras);
 831         }
 832
 833         spin_unlock(&ras->ras_lock);
 834
 835         if (end == 0) {
 836                 ll_ra_stats_inc(mapping, RA_STAT_ZERO_WINDOW);
 837                 RETURN(0);
 838         }
 839
 840         reserved = ll_ra_count_get(ll_i2sbi(mapping->host), end - start + 1);
 841         if (reserved < end - start + 1)
 842                 ll_ra_stats_inc(mapping, RA_STAT_MAX_IN_FLIGHT);
 843
 844         for (i = start; reserved > 0 && !match_failed && i <= end; i++) {
 845                 /* skip locked pages from previous readpage calls */
 846                 page = grab_cache_page_nowait(mapping, i);
 847                 if (page == NULL) {
 848                         CDEBUG(D_READA, "g_c_p_n failed\n");
 849                         continue;
 850                 }
 851
 852                 /* we do this first so that we can see the page in the /proc
 853                  * accounting */
 854                 llap = llap_from_page(page);
 855                 if (IS_ERR(llap) || llap->llap_defer_uptodate)
 856                         goto next_page;
 857
 858                 /* skip completed pages */
 859                 if (Page_Uptodate(page))
 860                         goto next_page;
 861
 862                 /* bail when we hit the end of the lock. */
 863                 if ((rc = ll_page_matches(page, flags)) <= 0) {
 864                         LL_CDEBUG_PAGE(D_READA | D_PAGE, page,
 865                                        "lock match failed: rc %d\n", rc);
 866                         ll_ra_stats_inc(mapping, RA_STAT_FAILED_MATCH);
 867                         match_failed = 1;
 868                         goto next_page;
 869                 }
 870
 871                 rc = ll_issue_page_read(exp, llap, oig, 1);
 872                 if (rc == 0) {
 873                         reserved--;
 874                         ret++;
 875                         LL_CDEBUG_PAGE(D_READA| D_PAGE, page,
 876                                        "started read-ahead\n");
 877                 }
 878                 if (rc) {
 879         next_page:
 880                         LL_CDEBUG_PAGE(D_READA | D_PAGE, page,
 881                                        "skipping read-ahead\n");
 882
 883                         unlock_page(page);
 884                 }
 885                 page_cache_release(page);
 886         }
 887
 888         LASSERTF(reserved >= 0, "reserved %lu\n", reserved);
 889         if (reserved != 0)
 890                 ll_ra_count_put(ll_i2sbi(mapping->host), reserved);
 891
 892         if (i == end + 1 && end == (kms >> PAGE_CACHE_SHIFT))
 893                 ll_ra_stats_inc(mapping, RA_STAT_EOF);
 894
 895         /* if we didn't get to the end of the region we reserved from
 896          * the ras we need to go back and update the ras so that the
 897          * next read-ahead tries from where we left off.  we only do so
 898          * if the region we failed to issue read-ahead on is still ahead
 899          * of the app and behind the next index to start read-ahead from */
 900         if (i != end + 1) {
 901                 spin_lock(&ras->ras_lock);
 902                 if (i < ras->ras_next_readahead &&
 903                     index_in_window(i, ras->ras_window_start, 0,
 904                                     ras->ras_window_len)) {
 905                         ras->ras_next_readahead = i;
 906                         RAS_CDEBUG(ras);
 907                 }
 908                 spin_unlock(&ras->ras_lock);
 909         }
 910
 911         RETURN(ret);
 912 }
 913
 914 static void ras_set_start(struct ll_readahead_state *ras, unsigned long index)
 915 {
 916         ras->ras_window_start = index & (~(PTLRPC_MAX_BRW_PAGES - 1));
 917 }
 918
 919 /* called with the ras_lock held or from places where it doesn't matter */
 920 static void ras_reset(struct ll_readahead_state *ras, unsigned long index)
 921 {
 922         ras->ras_last_readpage = index;
 923         ras->ras_consecutive = 1;
 924         ras->ras_window_len = 0;
 925         ras_set_start(ras, index);
 926         ras->ras_next_readahead = ras->ras_window_start;
 927
 928         RAS_CDEBUG(ras);
 929 }
 930
 931 void ll_readahead_init(struct inode *inode, struct ll_readahead_state *ras)
 932 {
 933         spin_lock_init(&ras->ras_lock);
 934         ras_reset(ras, 0);
 935 }
 936
 937 static void ras_update(struct ll_sb_info *sbi, struct ll_readahead_state *ras,
 938                        unsigned long index, unsigned hit)
 939 {
 940         struct ll_ra_info *ra = &sbi->ll_ra_info;
 941         int zero = 0;
 942         ENTRY;
 943
 944         spin_lock(&sbi->ll_lock);
 945         spin_lock(&ras->ras_lock);
 946
 947         ll_ra_stats_inc_unlocked(ra, hit ? RA_STAT_HIT : RA_STAT_MISS);
 948
 949         /* reset the read-ahead window in two cases.  First when the app seeks
 950          * or reads to some other part of the file.  Secondly if we get a
 951          * read-ahead miss that we think we've previously issued.  This can
 952          * be a symptom of there being so many read-ahead pages that the VM is
 953          * reclaiming it before we get to it. */
 954         if (!index_in_window(index, ras->ras_last_readpage, 8, 8)) {
 955                 zero = 1;
 956                 ll_ra_stats_inc_unlocked(ra, RA_STAT_DISTANT_READPAGE);
 957         } else if (!hit && ras->ras_window_len &&
 958                    index < ras->ras_next_readahead &&
 959                    index_in_window(index, ras->ras_window_start, 0,
 960                                    ras->ras_window_len)) {
 961                 zero = 1;
 962                 ll_ra_stats_inc_unlocked(ra, RA_STAT_MISS_IN_WINDOW);
 963         }
 964
 965         if (zero) {
 966                 ras_reset(ras, index);
 967                 GOTO(out_unlock, 0);
 968         }
 969
 970         ras->ras_last_readpage = index;
 971         ras->ras_consecutive++;
 972         ras_set_start(ras, index);
 973         ras->ras_next_readahead = max(ras->ras_window_start,
 974                                       ras->ras_next_readahead);
 975
 976         /* wait for a few pages to arrive before issuing readahead to avoid
 977          * the worst overutilization */
 978         if (ras->ras_consecutive == 3) {
 979                 ras->ras_window_len = PTLRPC_MAX_BRW_PAGES;
 980                 GOTO(out_unlock, 0);
 981         }
 982
 983         /* we need to increase the window sometimes.  we'll arbitrarily
 984          * do it half-way through the pages in an rpc */
 985         if ((index & (PTLRPC_MAX_BRW_PAGES - 1)) ==
 986             (PTLRPC_MAX_BRW_PAGES >> 1)) {
 987                 ras->ras_window_len += PTLRPC_MAX_BRW_PAGES;
 988                 ras->ras_window_len = min(ras->ras_window_len,
 989                                           ra->ra_max_pages);
 990
 991         }
 992
 993         EXIT;
 994 out_unlock:
 995         RAS_CDEBUG(ras);
 996         spin_unlock(&ras->ras_lock);
 997         spin_unlock(&sbi->ll_lock);
 998 }
 999
1000 /*
1001  * for now we do our readpage the same on both 2.4 and 2.5.  The kernel's
1002  * read-ahead assumes it is valid to issue readpage all the way up to
1003  * i_size, but our dlm locks make that not the case.  We disable the
1004  * kernel's read-ahead and do our own by walking ahead in the page cache
1005  * checking for dlm lock coverage.  the main difference between 2.4 and
1006  * 2.6 is how read-ahead gets batched and issued, but we're using our own,
1007  * so they look the same.
1008  */
1009 int ll_readpage(struct file *filp, struct page *page)
1010 {
1011         struct ll_file_data *fd = filp->private_data;
1012         struct inode *inode = page->mapping->host;
1013         struct obd_export *exp;
1014         struct ll_async_page *llap;
1015         struct obd_io_group *oig = NULL;
1016         int rc;
1017         ENTRY;
1018
1019         LASSERT(PageLocked(page));
1020         LASSERT(!PageUptodate(page));
1021         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),offset="LPX64"\n",
1022                inode->i_ino, inode->i_generation, inode,
1023                (((obd_off)page->index) << PAGE_SHIFT));
1024         LASSERT(atomic_read(&filp->f_dentry->d_inode->i_count) > 0);
1025
1026         rc = oig_init(&oig);
1027         if (rc < 0)
1028                 GOTO(out, rc);
1029
1030         exp = ll_i2dtexp(inode);
1031         if (exp == NULL)
1032                 GOTO(out, rc = -EINVAL);
1033
1034         llap = llap_from_page(page);
1035         if (IS_ERR(llap))
1036                 GOTO(out, rc = PTR_ERR(llap));
1037
1038         if (ll_i2sbi(inode)->ll_flags & LL_SBI_READAHEAD)
1039                 ras_update(ll_i2sbi(inode), &fd->fd_ras, page->index,
1040                            llap->llap_defer_uptodate);
1041
1042         if (llap->llap_defer_uptodate) {
1043                 llap->llap_ra_used = 1;
1044                 rc = ll_readahead(&fd->fd_ras, exp, page->mapping, oig,
1045                                   fd->fd_flags);
1046                 if (rc > 0)
1047                         obd_trigger_group_io(exp, ll_i2info(inode)->lli_smd,
1048                                              NULL, oig);
1049                 LL_CDEBUG_PAGE(D_PAGE, page, "marking uptodate from defer\n");
1050                 SetPageUptodate(page);
1051                 unlock_page(page);
1052                 GOTO(out_oig, rc = 0);
1053         }
1054
1055         rc = ll_page_matches(page, fd->fd_flags);
1056         if (rc < 0) {
1057                 LL_CDEBUG_PAGE(D_ERROR, page, "lock match failed: rc %d\n", rc);
1058                 GOTO(out, rc);
1059         }
1060
1061         if (rc == 0) {
1062                 CWARN("ino %lu page %lu (%llu) not covered by "
1063                       "a lock (mmap?).  check debug logs.\n",
1064                       inode->i_ino, page->index,
1065                       (long long)page->index << PAGE_CACHE_SHIFT);
1066         }
1067
1068         rc = ll_issue_page_read(exp, llap, oig, 0);
1069         if (rc)
1070                 GOTO(out, rc);
1071
1072         LL_CDEBUG_PAGE(D_PAGE, page, "queued readpage\n");
1073         if (ll_i2sbi(inode)->ll_flags & LL_SBI_READAHEAD)
1074                 ll_readahead(&fd->fd_ras, exp, page->mapping, oig,
1075                              fd->fd_flags);
1076
1077         rc = obd_trigger_group_io(exp, ll_i2info(inode)->lli_smd, NULL, oig);
1078         EXIT;
1079 out:
1080         if (rc)
1081                 unlock_page(page);
1082 out_oig:
1083         if (oig != NULL)
1084                 oig_release(oig);
1085         return rc;
1086 }