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 brw_page pg;
  64         int rc;
  65         ENTRY;
  66
  67         pg.pg = page;
  68         pg.off = ((obd_off)page->index) << PAGE_SHIFT;
  69
  70         if (cmd == OBD_BRW_WRITE &&
  71             (pg.off + PAGE_SIZE > inode->i_size))
  72                 pg.count = inode->i_size % PAGE_SIZE;
  73         else
  74                 pg.count = PAGE_SIZE;
  75
  76         CDEBUG(D_PAGE, "%s %d bytes ino %lu at "LPU64"/"LPX64"\n",
  77                cmd & OBD_BRW_WRITE ? "write" : "read", pg.count, inode->i_ino,
  78                pg.off, pg.off);
  79         if (pg.count == 0) {
  80                 CERROR("ZERO COUNT: ino %lu: size %p:%Lu(%p:%Lu) idx %lu off "
  81                        LPU64"\n", inode->i_ino, inode, inode->i_size,
  82                        page->mapping->host, page->mapping->host->i_size,
  83                        page->index, pg.off);
  84         }
  85
  86         pg.flag = flags;
  87
  88         if (cmd == OBD_BRW_WRITE)
  89                 lprocfs_counter_add(ll_i2sbi(inode)->ll_stats,
  90                                     LPROC_LL_BRW_WRITE, pg.count);
  91         else
  92                 lprocfs_counter_add(ll_i2sbi(inode)->ll_stats,
  93                                     LPROC_LL_BRW_READ, pg.count);
  94         rc = obd_brw(cmd, ll_i2obdexp(inode), oa, lsm, 1, &pg, NULL);
  95         if (rc == 0)
  96                 obdo_to_inode(inode, oa, OBD_MD_FLBLOCKS);
  97         else if (rc != -EIO)
  98                 CERROR("error from obd_brw: rc = %d\n", rc);
  99         RETURN(rc);
 100 }
 101
 102 /* this isn't where truncate starts.   roughly:
 103  * sys_truncate->ll_setattr_raw->vmtruncate->ll_truncate
 104  * we grab the lock back in setattr_raw to avoid races. */
 105 void ll_truncate(struct inode *inode)
 106 {
 107         struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
 108         struct obdo oa;
 109         int rc;
 110         ENTRY;
 111         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p)\n", inode->i_ino,
 112                inode->i_generation, inode);
 113
 114         if (!lsm) {
 115                 CDEBUG(D_INODE, "truncate on inode %lu with no objects\n",
 116                        inode->i_ino);
 117                 EXIT;
 118                 return;
 119         }
 120
 121         oa.o_id = lsm->lsm_object_id;
 122         oa.o_gr = lsm->lsm_object_gr;
 123         oa.o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
 124         obdo_from_inode(&oa, inode, OBD_MD_FLTYPE|OBD_MD_FLMODE|OBD_MD_FLATIME|
 125                                     OBD_MD_FLMTIME | OBD_MD_FLCTIME);
 126
 127         CDEBUG(D_INFO, "calling punch for "LPX64" (all bytes after %Lu)\n",
 128                oa.o_id, inode->i_size);
 129
 130         /* truncate == punch from new size to absolute end of file */
 131         /* NB: obd_punch must be called with i_sem held!  It updates the kms! */
 132         rc = obd_punch(ll_i2obdexp(inode), &oa, lsm, inode->i_size,
 133                        OBD_OBJECT_EOF, NULL);
 134         if (rc)
 135                 CERROR("obd_truncate fails (%d) ino %lu\n", rc, inode->i_ino);
 136         else
 137                 obdo_to_inode(inode, &oa, OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
 138                                           OBD_MD_FLATIME | OBD_MD_FLMTIME |
 139                                           OBD_MD_FLCTIME);
 140
 141         EXIT;
 142         return;
 143 } /* ll_truncate */
 144
 145 __u64 lov_merge_size(struct lov_stripe_md *lsm, int kms);
 146 int ll_prepare_write(struct file *file, struct page *page, unsigned from,
 147                      unsigned to)
 148 {
 149         struct inode *inode = page->mapping->host;
 150         struct ll_inode_info *lli = ll_i2info(inode);
 151         struct lov_stripe_md *lsm = lli->lli_smd;
 152         obd_off offset = ((obd_off)page->index) << PAGE_SHIFT;
 153         struct brw_page pga;
 154         struct obdo oa;
 155         __u64 kms;
 156         int rc = 0;
 157         ENTRY;
 158
 159         LASSERT(PageLocked(page));
 160         (void)llap_cast_private(page); /* assertion */
 161
 162         /* Check to see if we should return -EIO right away */
 163         pga.pg = page;
 164         pga.off = offset;
 165         pga.count = PAGE_SIZE;
 166         pga.flag = 0;
 167
 168         oa.o_id = lsm->lsm_object_id;
 169         oa.o_gr = lsm->lsm_object_gr;
 170         oa.o_mode = inode->i_mode;
 171         oa.o_valid = OBD_MD_FLID | OBD_MD_FLMODE
 172                                 | OBD_MD_FLTYPE | OBD_MD_FLGROUP;
 173
 174         rc = obd_brw(OBD_BRW_CHECK, ll_i2obdexp(inode), &oa, lsm, 1, &pga,
 175                      NULL);
 176         if (rc)
 177                 RETURN(rc);
 178
 179         if (PageUptodate(page))
 180                 RETURN(0);
 181
 182         /* We're completely overwriting an existing page, so _don't_ set it up
 183          * to date until commit_write */
 184         if (from == 0 && to == PAGE_SIZE) {
 185                 POISON_PAGE(page, 0x11);
 186                 RETURN(0);
 187         }
 188
 189         /* If are writing to a new page, no need to read old data.  The extent
 190          * locking will have updated the KMS, and for our purposes here we can
 191          * treat it like i_size. */
 192         kms = lov_merge_size(lsm, 1);
 193         if (kms <= offset) {
 194                 memset(kmap(page), 0, PAGE_SIZE);
 195                 kunmap(page);
 196                 GOTO(prepare_done, rc = 0);
 197         }
 198
 199         /* XXX could be an async ocp read.. read-ahead? */
 200         rc = ll_brw(OBD_BRW_READ, inode, &oa, page, 0);
 201         if (rc == 0) {
 202                 /* bug 1598: don't clobber blksize */
 203                 oa.o_valid &= ~(OBD_MD_FLSIZE | OBD_MD_FLBLKSZ);
 204                 obdo_refresh_inode(inode, &oa, oa.o_valid);
 205         }
 206
 207         EXIT;
 208  prepare_done:
 209         if (rc == 0)
 210                 SetPageUptodate(page);
 211
 212         return rc;
 213 }
 214
 215 struct ll_async_page *llap_from_cookie(void *cookie)
 216 {
 217         struct ll_async_page *llap = cookie;
 218         if (llap->llap_magic != LLAP_MAGIC)
 219                 return ERR_PTR(-EINVAL);
 220         return llap;
 221 };
 222
 223 static int ll_ap_make_ready(void *data, int cmd)
 224 {
 225         struct ll_async_page *llap;
 226         struct page *page;
 227         ENTRY;
 228
 229         llap = llap_from_cookie(data);
 230         if (IS_ERR(llap))
 231                 RETURN(-EINVAL);
 232
 233         page = llap->llap_page;
 234
 235         LASSERT(cmd != OBD_BRW_READ);
 236
 237         /* we're trying to write, but the page is locked.. come back later */
 238         if (TryLockPage(page))
 239                 RETURN(-EAGAIN);
 240
 241         LL_CDEBUG_PAGE(D_PAGE, page, "made ready\n");
 242         page_cache_get(page);
 243
 244         /* if we left PageDirty we might get another writepage call
 245          * in the future.  list walkers are bright enough
 246          * to check page dirty so we can leave it on whatever list
 247          * its on.  XXX also, we're called with the cli list so if
 248          * we got the page cache list we'd create a lock inversion
 249          * with the removepage path which gets the page lock then the
 250          * cli lock */
 251         clear_page_dirty(page);
 252         RETURN(0);
 253 }
 254
 255 /* We have two reasons for giving llite the opportunity to change the
 256  * write length of a given queued page as it builds the RPC containing
 257  * the page:
 258  *
 259  * 1) Further extending writes may have landed in the page cache
 260  *    since a partial write first queued this page requiring us
 261  *    to write more from the page cache.
 262  * 2) We might have raced with truncate and want to avoid performing
 263  *    write RPCs that are just going to be thrown away by the
 264  *    truncate's punch on the storage targets.
 265  *
 266  * The kms serves these purposes as it is set at both truncate and extending
 267  * writes.
 268  */
 269 static int ll_ap_refresh_count(void *data, int cmd)
 270 {
 271         struct ll_async_page *llap;
 272         struct lov_stripe_md *lsm;
 273         struct page *page;
 274         __u64 kms;
 275         ENTRY;
 276
 277         /* readpage queues with _COUNT_STABLE, shouldn't get here. */
 278         LASSERT(cmd != OBD_BRW_READ);
 279
 280         llap = llap_from_cookie(data);
 281         if (IS_ERR(llap))
 282                 RETURN(PTR_ERR(llap));
 283
 284         page = llap->llap_page;
 285         lsm = ll_i2info(page->mapping->host)->lli_smd;
 286         kms = lov_merge_size(lsm, 1);
 287
 288         /* catch race with truncate */
 289         if (((__u64)page->index << PAGE_SHIFT) >= kms)
 290                 return 0;
 291
 292         /* catch sub-page write at end of file */
 293         if (((__u64)page->index << PAGE_SHIFT) + PAGE_SIZE > kms)
 294                 return kms % PAGE_SIZE;
 295
 296         return PAGE_SIZE;
 297 }
 298
 299 void ll_inode_fill_obdo(struct inode *inode, int cmd, struct obdo *oa)
 300 {
 301         struct lov_stripe_md *lsm;
 302         obd_flag valid_flags;
 303
 304         lsm = ll_i2info(inode)->lli_smd;
 305
 306         oa->o_id = lsm->lsm_object_id;
 307         oa->o_gr = lsm->lsm_object_gr;
 308         oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
 309         valid_flags = OBD_MD_FLTYPE | OBD_MD_FLATIME;
 310         if (cmd == OBD_BRW_WRITE) {
 311                 oa->o_valid |= OBD_MD_FLIFID | OBD_MD_FLEPOCH;
 312                 mdc_pack_fid(obdo_fid(oa), inode->i_ino, 0, inode->i_mode);
 313                 obdo_fid(oa)->mds = ll_i2info(inode)->lli_mds;
 314                 oa->o_easize = ll_i2info(inode)->lli_io_epoch;
 315
 316                 valid_flags |= OBD_MD_FLMTIME | OBD_MD_FLCTIME;
 317         }
 318
 319         obdo_from_inode(oa, inode, valid_flags);
 320 }
 321
 322 static void ll_ap_fill_obdo(void *data, int cmd, struct obdo *oa)
 323 {
 324         struct ll_async_page *llap;
 325         ENTRY;
 326
 327         llap = llap_from_cookie(data);
 328         if (IS_ERR(llap)) {
 329                 EXIT;
 330                 return;
 331         }
 332
 333         ll_inode_fill_obdo(llap->llap_page->mapping->host, cmd, oa);
 334         EXIT;
 335 }
 336
 337 static struct obd_async_page_ops ll_async_page_ops = {
 338         .ap_make_ready =        ll_ap_make_ready,
 339         .ap_refresh_count =     ll_ap_refresh_count,
 340         .ap_fill_obdo =         ll_ap_fill_obdo,
 341         .ap_completion =        ll_ap_completion,
 342 };
 343
 344 struct ll_async_page *llap_cast_private(struct page *page)
 345 {
 346         struct ll_async_page *llap = (struct ll_async_page *)page->private;
 347
 348         LASSERTF(llap == NULL || llap->llap_magic == LLAP_MAGIC,
 349                  "page %p private %lu gave magic %d which != %d\n",
 350                  page, page->private, llap->llap_magic, LLAP_MAGIC);
 351
 352         return llap;
 353 }
 354
 355 /* XXX have the exp be an argument? */
 356 struct ll_async_page *llap_from_page(struct page *page)
 357 {
 358         struct ll_async_page *llap;
 359         struct obd_export *exp;
 360         struct inode *inode = page->mapping->host;
 361         struct ll_sb_info *sbi = ll_i2sbi(inode);
 362         int rc;
 363         ENTRY;
 364
 365         llap = llap_cast_private(page);
 366         if (llap != NULL)
 367                 RETURN(llap);
 368
 369         exp = ll_i2obdexp(page->mapping->host);
 370         if (exp == NULL)
 371                 RETURN(ERR_PTR(-EINVAL));
 372
 373         OBD_ALLOC(llap, sizeof(*llap));
 374         if (llap == NULL)
 375                 RETURN(ERR_PTR(-ENOMEM));
 376         llap->llap_magic = LLAP_MAGIC;
 377         rc = obd_prep_async_page(exp, ll_i2info(inode)->lli_smd, NULL, page,
 378                                  (obd_off)page->index << PAGE_SHIFT,
 379                                  &ll_async_page_ops, llap, &llap->llap_cookie);
 380         if (rc) {
 381                 OBD_FREE(llap, sizeof(*llap));
 382                 RETURN(ERR_PTR(rc));
 383         }
 384
 385         CDEBUG(D_CACHE, "llap %p page %p cookie %p obj off "LPU64"\n", llap,
 386                page, llap->llap_cookie, (obd_off)page->index << PAGE_SHIFT);
 387         /* also zeroing the PRIVBITS low order bitflags */
 388         page->private = (unsigned long)llap;
 389         llap->llap_page = page;
 390
 391         spin_lock(&sbi->ll_lock);
 392         sbi->ll_pglist_gen++;
 393         list_add_tail(&llap->llap_proc_item, &sbi->ll_pglist);
 394         spin_unlock(&sbi->ll_lock);
 395
 396         RETURN(llap);
 397 }
 398
 399 void lov_increase_kms(struct obd_export *exp, struct lov_stripe_md *lsm,
 400                       obd_off size);
 401 /* update our write count to account for i_size increases that may have
 402  * happened since we've queued the page for io. */
 403
 404 /* be careful not to return success without setting the page Uptodate or
 405  * the next pass through prepare_write will read in stale data from disk. */
 406 int ll_commit_write(struct file *file, struct page *page, unsigned from,
 407                     unsigned to)
 408 {
 409         struct inode *inode = page->mapping->host;
 410         struct ll_inode_info *lli = ll_i2info(inode);
 411         struct lov_stripe_md *lsm = lli->lli_smd;
 412         struct obd_export *exp = NULL;
 413         struct ll_async_page *llap;
 414         loff_t size;
 415         int rc = 0;
 416         ENTRY;
 417
 418         SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
 419         LASSERT(inode == file->f_dentry->d_inode);
 420         LASSERT(PageLocked(page));
 421
 422         CDEBUG(D_INODE, "inode %p is writing page %p from %d to %d at %lu\n",
 423                inode, page, from, to, page->index);
 424
 425         llap = llap_from_page(page);
 426         if (IS_ERR(llap))
 427                 RETURN(PTR_ERR(llap));
 428
 429         /* queue a write for some time in the future the first time we
 430          * dirty the page */
 431         if (!PageDirty(page)) {
 432                 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats,
 433                                      LPROC_LL_DIRTY_MISSES);
 434
 435                 exp = ll_i2obdexp(inode);
 436                 if (exp == NULL)
 437                         RETURN(-EINVAL);
 438
 439                 /* _make_ready only sees llap once we've unlocked the page */
 440                 llap->llap_write_queued = 1;
 441                 rc = obd_queue_async_io(exp, lsm, NULL, llap->llap_cookie,
 442                                         OBD_BRW_WRITE, 0, 0, 0, 0);
 443                 if (rc != 0) { /* async failed, try sync.. */
 444                         struct obd_io_group *oig;
 445                         rc = oig_init(&oig);
 446                         if (rc)
 447                                 GOTO(out, rc);
 448
 449                         llap->llap_write_queued = 0;
 450                         rc = obd_queue_group_io(exp, lsm, NULL, oig,
 451                                                 llap->llap_cookie,
 452                                                 OBD_BRW_WRITE, 0, to, 0,
 453                                                 ASYNC_READY | ASYNC_URGENT |
 454                                                 ASYNC_COUNT_STABLE |
 455                                                 ASYNC_GROUP_SYNC);
 456
 457                         if (rc)
 458                                 GOTO(free_oig, rc);
 459
 460                         rc = obd_trigger_group_io(exp, lsm, NULL, oig);
 461                         if (rc)
 462                                 GOTO(free_oig, rc);
 463
 464                         rc = oig_wait(oig);
 465 free_oig:
 466                         oig_release(oig);
 467                         GOTO(out, rc);
 468                 }
 469                 LL_CDEBUG_PAGE(D_PAGE, page, "write queued\n");
 470                 //llap_write_pending(inode, llap);
 471         } else {
 472                 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats,
 473                                      LPROC_LL_DIRTY_HITS);
 474         }
 475
 476         /* put the page in the page cache, from now on ll_removepage is
 477          * responsible for cleaning up the llap */
 478         set_page_dirty(page);
 479
 480 out:
 481         if (rc == 0) {
 482                 size = (((obd_off)page->index) << PAGE_SHIFT) + to;
 483                 lov_increase_kms(exp, lsm, size);
 484                 if (size > inode->i_size)
 485                         inode->i_size = size;
 486                 SetPageUptodate(page);
 487         }
 488         RETURN(rc);
 489 }
 490
 491 static unsigned long ll_ra_count_get(struct ll_sb_info *sbi, unsigned long len)
 492 {
 493         unsigned long ret;
 494         ENTRY;
 495
 496         spin_lock(&sbi->ll_lock);
 497         ret = min(sbi->ll_max_read_ahead_pages - sbi->ll_read_ahead_pages,
 498                   len);
 499         sbi->ll_read_ahead_pages += ret;
 500         spin_unlock(&sbi->ll_lock);
 501
 502         RETURN(ret);
 503 }
 504
 505 static void ll_ra_count_put(struct ll_sb_info *sbi, unsigned long len)
 506 {
 507         spin_lock(&sbi->ll_lock);
 508         LASSERTF(sbi->ll_read_ahead_pages >= len, "r_a_p %lu len %lu\n",
 509                  sbi->ll_read_ahead_pages, len);
 510         sbi->ll_read_ahead_pages -= len;
 511         spin_unlock(&sbi->ll_lock);
 512 }
 513
 514 /* called for each page in a completed rpc.*/
 515 void ll_ap_completion(void *data, int cmd, struct obdo *oa, int rc)
 516 {
 517         struct ll_async_page *llap;
 518         struct page *page;
 519         ENTRY;
 520
 521         llap = llap_from_cookie(data);
 522         if (IS_ERR(llap)) {
 523                 EXIT;
 524                 return;
 525         }
 526
 527         page = llap->llap_page;
 528         LASSERT(PageLocked(page));
 529
 530         LL_CDEBUG_PAGE(D_PAGE, page, "completing cmd %d with %d\n", cmd, rc);
 531
 532         if (cmd == OBD_BRW_READ && llap->llap_defer_uptodate)
 533                 ll_ra_count_put(ll_i2sbi(page->mapping->host), 1);
 534
 535         if (rc == 0)  {
 536                 if (cmd == OBD_BRW_READ) {
 537                         if (!llap->llap_defer_uptodate)
 538                                 SetPageUptodate(page);
 539                 } else {
 540                         llap->llap_write_queued = 0;
 541                 }
 542                 ClearPageError(page);
 543         } else {
 544                 if (cmd == OBD_BRW_READ)
 545                         llap->llap_defer_uptodate = 0;
 546                 SetPageError(page);
 547         }
 548
 549         unlock_page(page);
 550
 551         if (0 && cmd == OBD_BRW_WRITE) {
 552                 llap_write_complete(page->mapping->host, llap);
 553                 ll_try_done_writing(page->mapping->host);
 554         }
 555
 556         page_cache_release(page);
 557         EXIT;
 558 }
 559
 560 /* the kernel calls us here when a page is unhashed from the page cache.
 561  * the page will be locked and the kernel is holding a spinlock, so
 562  * we need to be careful.  we're just tearing down our book-keeping
 563  * here. */
 564 void ll_removepage(struct page *page)
 565 {
 566         struct inode *inode = page->mapping->host;
 567         struct obd_export *exp;
 568         struct ll_async_page *llap;
 569         struct ll_sb_info *sbi = ll_i2sbi(inode);
 570         int rc;
 571         ENTRY;
 572
 573         LASSERT(!in_interrupt());
 574
 575         /* sync pages or failed read pages can leave pages in the page
 576          * cache that don't have our data associated with them anymore */
 577         if (page->private == 0) {
 578                 EXIT;
 579                 return;
 580         }
 581
 582         LL_CDEBUG_PAGE(D_PAGE, page, "being evicted\n");
 583
 584         exp = ll_i2obdexp(inode);
 585         if (exp == NULL) {
 586                 CERROR("page %p ind %lu gave null export\n", page, page->index);
 587                 EXIT;
 588                 return;
 589         }
 590
 591         llap = llap_from_page(page);
 592         if (IS_ERR(llap)) {
 593                 CERROR("page %p ind %lu couldn't find llap: %ld\n", page,
 594                        page->index, PTR_ERR(llap));
 595                 EXIT;
 596                 return;
 597         }
 598
 599         //llap_write_complete(inode, llap);
 600         rc = obd_teardown_async_page(exp, ll_i2info(inode)->lli_smd, NULL,
 601                                      llap->llap_cookie);
 602         if (rc != 0)
 603                 CERROR("page %p ind %lu failed: %d\n", page, page->index, rc);
 604
 605         /* this unconditional free is only safe because the page lock
 606          * is providing exclusivity to memory pressure/truncate/writeback..*/
 607         page->private = 0;
 608
 609         spin_lock(&sbi->ll_lock);
 610         if (!list_empty(&llap->llap_proc_item))
 611                 list_del_init(&llap->llap_proc_item);
 612         sbi->ll_pglist_gen++;
 613         spin_unlock(&sbi->ll_lock);
 614         OBD_FREE(llap, sizeof(*llap));
 615         EXIT;
 616 }
 617
 618 static int ll_page_matches(struct page *page, int fd_flags)
 619 {
 620         struct lustre_handle match_lockh = {0};
 621         struct inode *inode = page->mapping->host;
 622         ldlm_policy_data_t page_extent;
 623         int flags, matches;
 624         ENTRY;
 625
 626         if (fd_flags & LL_FILE_GROUP_LOCKED)
 627                 RETURN(1);
 628
 629         page_extent.l_extent.start = (__u64)page->index << PAGE_CACHE_SHIFT;
 630         page_extent.l_extent.end =
 631                 page_extent.l_extent.start + PAGE_CACHE_SIZE - 1;
 632         flags = LDLM_FL_CBPENDING | LDLM_FL_BLOCK_GRANTED | LDLM_FL_TEST_LOCK;
 633         matches = obd_match(ll_i2sbi(inode)->ll_osc_exp,
 634                             ll_i2info(inode)->lli_smd, LDLM_EXTENT,
 635                             &page_extent, LCK_PR | LCK_PW, &flags, inode,
 636                             &match_lockh);
 637         RETURN(matches);
 638 }
 639
 640 static int ll_issue_page_read(struct obd_export *exp,
 641                               struct ll_async_page *llap,
 642                               struct obd_io_group *oig, int defer)
 643 {
 644         struct page *page = llap->llap_page;
 645         int rc;
 646
 647         page_cache_get(page);
 648         llap->llap_defer_uptodate = defer;
 649         rc = obd_queue_group_io(exp, ll_i2info(page->mapping->host)->lli_smd,
 650                                 NULL, oig, llap->llap_cookie, OBD_BRW_READ, 0,
 651                                 PAGE_SIZE, 0, ASYNC_COUNT_STABLE | ASYNC_READY
 652                                               | ASYNC_URGENT);
 653         if (rc) {
 654                 LL_CDEBUG_PAGE(D_ERROR, page, "read queue failed: rc %d\n", rc);
 655                 page_cache_release(page);
 656         }
 657         RETURN(rc);
 658 }
 659
 660 #define RAS_CDEBUG(ras) \
 661         CDEBUG(D_READA, "lrp %lu c %lu ws %lu wl %lu nra %lu\n",        \
 662                ras->ras_last_readpage, ras->ras_consecutive,            \
 663                ras->ras_window_start, ras->ras_window_len,              \
 664                ras->ras_next_readahead);
 665
 666 static int ll_readahead(struct ll_readahead_state *ras,
 667                          struct obd_export *exp, struct address_space *mapping,
 668                          struct obd_io_group *oig, int flags)
 669 {
 670         unsigned long i, start = 0, end = 0, reserved;
 671         struct ll_async_page *llap;
 672         struct page *page;
 673         int rc, ret = 0;
 674         __u64 kms;
 675         ENTRY;
 676
 677         kms = lov_merge_size(ll_i2info(mapping->host)->lli_smd, 1);
 678         if (kms == 0)
 679                 RETURN(0);
 680
 681         spin_lock(&ras->ras_lock);
 682
 683         if (ras->ras_window_len) {
 684                 start = ras->ras_next_readahead;
 685                 end = ras->ras_window_start + ras->ras_window_len - 1;
 686                 end = min(end, (unsigned long)(kms >> PAGE_CACHE_SHIFT));
 687                 ras->ras_next_readahead = max(end, end + 1);
 688
 689                 RAS_CDEBUG(ras);
 690         }
 691
 692         spin_unlock(&ras->ras_lock);
 693
 694         if (end == 0)
 695                 RETURN(0);
 696
 697         reserved = ll_ra_count_get(ll_i2sbi(mapping->host), end - start + 1);
 698
 699         for (i = start; reserved > 0 && i <= end; i++) {
 700                 /* skip locked pages from previous readpage calls */
 701                 page = grab_cache_page_nowait(mapping, i);
 702                 if (page == NULL) {
 703                         CDEBUG(D_READA, "g_c_p_n failed\n");
 704                         continue;
 705                 }
 706
 707                 /* we do this first so that we can see the page in the /proc
 708                  * accounting */
 709                 llap = llap_from_page(page);
 710                 if (IS_ERR(llap) || llap->llap_defer_uptodate)
 711                         goto next_page;
 712
 713                 /* skip completed pages */
 714                 if (Page_Uptodate(page))
 715                         goto next_page;
 716
 717                 /* bail when we hit the end of the lock. */
 718                 if ((rc = ll_page_matches(page, flags)) <= 0) {
 719                         LL_CDEBUG_PAGE(D_READA | D_PAGE, page,
 720                                        "lock match failed: rc %d\n", rc);
 721                         i = end;
 722                         goto next_page;
 723                 }
 724
 725                 rc = ll_issue_page_read(exp, llap, oig, 1);
 726                 if (rc == 0) {
 727                         reserved--;
 728                         ret++;
 729                         LL_CDEBUG_PAGE(D_READA| D_PAGE, page,
 730                                        "started read-ahead\n");
 731                 }
 732                 if (rc) {
 733         next_page:
 734                         LL_CDEBUG_PAGE(D_READA | D_PAGE, page,
 735                                        "skipping read-ahead\n");
 736
 737                         unlock_page(page);
 738                 }
 739                 page_cache_release(page);
 740         }
 741
 742         LASSERTF(reserved >= 0, "reserved %lu\n", reserved);
 743         if (reserved != 0)
 744                 ll_ra_count_put(ll_i2sbi(mapping->host), reserved);
 745         RETURN(ret);
 746 }
 747
 748 static void ras_set_start(struct ll_readahead_state *ras,
 749                                unsigned long index)
 750 {
 751         ras->ras_window_start = index & (~(PTLRPC_MAX_BRW_PAGES - 1));
 752         ras->ras_next_readahead = max(ras->ras_window_start,
 753                                       ras->ras_next_readahead);
 754 }
 755
 756 /* called with the ras_lock held or from places where it doesn't matter */
 757 static void ras_reset(struct ll_readahead_state *ras,
 758                       unsigned long index)
 759 {
 760         ras->ras_last_readpage = index;
 761         ras->ras_consecutive = 1;
 762         ras->ras_window_len = 0;
 763         ras_set_start(ras, index);
 764         ras->ras_next_readahead = ras->ras_window_start;
 765
 766         RAS_CDEBUG(ras);
 767 }
 768
 769 void ll_readahead_init(struct inode *inode, struct ll_readahead_state *ras)
 770 {
 771         spin_lock_init(&ras->ras_lock);
 772         ras_reset(ras, 0);
 773 }
 774
 775 static void ras_update(struct ll_readahead_state *ras,
 776                        unsigned long index, unsigned long max)
 777 {
 778         ENTRY;
 779
 780         spin_lock(&ras->ras_lock);
 781
 782         if (index != ras->ras_last_readpage + 1) {
 783                 ras_reset(ras, index);
 784                 GOTO(out_unlock, 0);
 785         }
 786
 787         ras->ras_last_readpage = index;
 788         ras->ras_consecutive++;
 789         ras_set_start(ras, index);
 790
 791         if (ras->ras_consecutive == 2) {
 792                 ras->ras_window_len = PTLRPC_MAX_BRW_PAGES;
 793                 GOTO(out_unlock, 0);
 794         }
 795
 796         /* we need to increase the window sometimes.  we'll arbitrarily
 797          * do it half-way through the pages in an rpc */
 798         if ((index & (PTLRPC_MAX_BRW_PAGES - 1)) ==
 799             (PTLRPC_MAX_BRW_PAGES >> 1)) {
 800                 ras->ras_window_len += PTLRPC_MAX_BRW_PAGES;
 801                 ras->ras_window_len = min(ras->ras_window_len, max);
 802         }
 803
 804         EXIT;
 805 out_unlock:
 806         RAS_CDEBUG(ras);
 807         spin_unlock(&ras->ras_lock);
 808         return;
 809 }
 810
 811 /*
 812  * for now we do our readpage the same on both 2.4 and 2.5.  The kernel's
 813  * read-ahead assumes it is valid to issue readpage all the way up to
 814  * i_size, but our dlm locks make that not the case.  We disable the
 815  * kernel's read-ahead and do our own by walking ahead in the page cache
 816  * checking for dlm lock coverage.  the main difference between 2.4 and
 817  * 2.6 is how read-ahead gets batched and issued, but we're using our own,
 818  * so they look the same.
 819  */
 820 int ll_readpage(struct file *filp, struct page *page)
 821 {
 822         struct ll_file_data *fd = filp->private_data;
 823         struct inode *inode = page->mapping->host;
 824         struct obd_export *exp;
 825         struct ll_async_page *llap;
 826         struct obd_io_group *oig = NULL;
 827         int rc;
 828         ENTRY;
 829
 830         LASSERT(PageLocked(page));
 831         LASSERT(!PageUptodate(page));
 832         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),offset="LPX64"\n",
 833                inode->i_ino, inode->i_generation, inode,
 834                (((obd_off)page->index) << PAGE_SHIFT));
 835         LASSERT(atomic_read(&filp->f_dentry->d_inode->i_count) > 0);
 836
 837         rc = oig_init(&oig);
 838         if (rc < 0)
 839                 GOTO(out, rc);
 840
 841         exp = ll_i2obdexp(inode);
 842         if (exp == NULL)
 843                 GOTO(out, rc = -EINVAL);
 844
 845         llap = llap_from_page(page);
 846         if (IS_ERR(llap))
 847                 GOTO(out, rc = PTR_ERR(llap));
 848
 849         if (ll_i2sbi(inode)->ll_flags & LL_SBI_READAHEAD)
 850                 ras_update(&fd->fd_ras, page->index,
 851                            ll_i2sbi(inode)->ll_max_read_ahead_pages);
 852
 853         if (llap->llap_defer_uptodate) {
 854                 rc = ll_readahead(&fd->fd_ras, exp, page->mapping, oig,
 855                                   fd->fd_flags);
 856                 if (rc > 0)
 857                         obd_trigger_group_io(exp, ll_i2info(inode)->lli_smd,
 858                                              NULL, oig);
 859                 LL_CDEBUG_PAGE(D_PAGE, page, "marking uptodate from defer\n");
 860                 SetPageUptodate(page);
 861                 unlock_page(page);
 862                 GOTO(out_oig, rc = 0);
 863         }
 864
 865         rc = ll_page_matches(page, fd->fd_flags);
 866         if (rc < 0) {
 867                 LL_CDEBUG_PAGE(D_ERROR, page, "lock match failed: rc %d\n", rc);
 868                 GOTO(out, rc);
 869         }
 870
 871         if (rc == 0) {
 872                 static unsigned long next_print;
 873                 CDEBUG(D_INODE, "ino %lu page %lu (%llu) didn't match a lock\n",
 874                        inode->i_ino, page->index,
 875                        (long long)page->index << PAGE_CACHE_SHIFT);
 876                 if (time_after(jiffies, next_print)) {
 877                         CWARN("ino %lu page %lu (%llu) not covered by "
 878                                "a lock (mmap?).  check debug logs.\n",
 879                                inode->i_ino, page->index,
 880                                (long long)page->index << PAGE_CACHE_SHIFT);
 881                         next_print = jiffies + 30 * HZ;
 882                 }
 883         }
 884
 885         rc = ll_issue_page_read(exp, llap, oig, 0);
 886         if (rc)
 887                 GOTO(out, rc);
 888
 889         LL_CDEBUG_PAGE(D_PAGE, page, "queued readpage\n");
 890         if (ll_i2sbi(inode)->ll_flags & LL_SBI_READAHEAD)
 891                 ll_readahead(&fd->fd_ras, exp, page->mapping, oig,
 892                              fd->fd_flags);
 893
 894         rc = obd_trigger_group_io(exp, ll_i2info(inode)->lli_smd, NULL, oig);
 895
 896 out:
 897         if (rc)
 898                 unlock_page(page);
 899 out_oig:
 900         if (oig != NULL)
 901                 oig_release(oig);
 902         RETURN(rc);
 903 }