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 #ifdef HAVE_KERNEL_CONFIG_H
  24 #include <linux/config.h>
  25 #endif
  26 #include <linux/kernel.h>
  27 #include <linux/mm.h>
  28 #include <linux/string.h>
  29 #include <linux/stat.h>
  30 #include <linux/errno.h>
  31 #include <linux/smp_lock.h>
  32 #include <linux/unistd.h>
  33 #include <linux/version.h>
  34 #include <asm/system.h>
  35 #include <asm/uaccess.h>
  36
  37 #include <linux/fs.h>
  38 #include <linux/stat.h>
  39 #include <asm/uaccess.h>
  40 #include <asm/segment.h>
  41 #include <linux/mm.h>
  42 #include <linux/pagemap.h>
  43 #include <linux/smp_lock.h>
  44
  45 #define DEBUG_SUBSYSTEM S_LLITE
  46
  47 #include <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 kmem_cache_t *ll_async_page_slab = NULL;
  58 size_t ll_async_page_slab_size = 0;
  59
  60 /* SYNCHRONOUS I/O to object storage for an inode */
  61 static int ll_brw(int cmd, struct inode *inode, struct obdo *oa,
  62                   struct page *page, int flags)
  63 {
  64         struct ll_inode_info *lli = ll_i2info(inode);
  65         struct lov_stripe_md *lsm = lli->lli_smd;
  66         struct obd_info oinfo = { { { 0 } } };
  67         struct brw_page pg;
  68         int rc;
  69         ENTRY;
  70
  71         pg.pg = page;
  72         pg.off = ((obd_off)page->index) << CFS_PAGE_SHIFT;
  73
  74         if ((cmd & OBD_BRW_WRITE) && (pg.off + CFS_PAGE_SIZE > inode->i_size))
  75                 pg.count = inode->i_size % CFS_PAGE_SIZE;
  76         else
  77                 pg.count = CFS_PAGE_SIZE;
  78
  79         LL_CDEBUG_PAGE(D_PAGE, page, "%s %d bytes ino %lu at "LPU64"/"LPX64"\n",
  80                        cmd & OBD_BRW_WRITE ? "write" : "read", pg.count,
  81                        inode->i_ino, pg.off, pg.off);
  82         if (pg.count == 0) {
  83                 CERROR("ZERO COUNT: ino %lu: size %p:%Lu(%p:%Lu) idx %lu off "
  84                        LPU64"\n",
  85                        inode->i_ino, inode, inode->i_size, page->mapping->host,
  86                        page->mapping->host->i_size, page->index, pg.off);
  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         oinfo.oi_oa = oa;
  98         oinfo.oi_md = lsm;
  99         rc = obd_brw(cmd, ll_i2obdexp(inode), &oinfo, 1, &pg, NULL);
 100         if (rc == 0)
 101                 obdo_to_inode(inode, oa, OBD_MD_FLBLOCKS);
 102         else if (rc != -EIO)
 103                 CERROR("error from obd_brw: rc = %d\n", rc);
 104         RETURN(rc);
 105 }
 106
 107 /* this isn't where truncate starts.   roughly:
 108  * sys_truncate->ll_setattr_raw->vmtruncate->ll_truncate. setattr_raw grabs
 109  * DLM lock on [size, EOF], i_mutex, ->lli_size_sem, and WRITE_I_ALLOC_SEM to
 110  * avoid races.
 111  *
 112  * must be called under ->lli_size_sem */
 113 void ll_truncate(struct inode *inode)
 114 {
 115         struct ll_inode_info *lli = ll_i2info(inode);
 116         struct obd_info oinfo = { { { 0 } } };
 117         struct ost_lvb lvb;
 118         struct obdo oa;
 119         int rc;
 120         ENTRY;
 121         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p) to %Lu=%#Lx\n",inode->i_ino,
 122                inode->i_generation, inode, inode->i_size, inode->i_size);
 123
 124         ll_vfs_ops_tally(ll_i2sbi(inode), VFS_OPS_TRUNCATE);
 125         if (lli->lli_size_sem_owner != current) {
 126                 EXIT;
 127                 return;
 128         }
 129
 130         if (!lli->lli_smd) {
 131                 CDEBUG(D_INODE, "truncate on inode %lu with no objects\n",
 132                        inode->i_ino);
 133                 GOTO(out_unlock, 0);
 134         }
 135
 136         LASSERT(atomic_read(&lli->lli_size_sem.count) <= 0);
 137
 138         /* XXX I'm pretty sure this is a hack to paper over a more fundamental
 139          * race condition. */
 140         lov_stripe_lock(lli->lli_smd);
 141         inode_init_lvb(inode, &lvb);
 142         obd_merge_lvb(ll_i2obdexp(inode), lli->lli_smd, &lvb, 0);
 143         if (lvb.lvb_size == inode->i_size) {
 144                 CDEBUG(D_VFSTRACE, "skipping punch for obj "LPX64", %Lu=%#Lx\n",
 145                        lli->lli_smd->lsm_object_id,inode->i_size,inode->i_size);
 146                 lov_stripe_unlock(lli->lli_smd);
 147                 GOTO(out_unlock, 0);
 148         }
 149
 150         obd_adjust_kms(ll_i2obdexp(inode), lli->lli_smd, inode->i_size, 1);
 151         lov_stripe_unlock(lli->lli_smd);
 152
 153         if (unlikely((ll_i2sbi(inode)->ll_flags & LL_SBI_CHECKSUM) &&
 154                      (inode->i_size & ~CFS_PAGE_MASK))) {
 155                 /* If the truncate leaves behind a partial page, update its
 156                  * checksum. */
 157                 struct page *page = find_get_page(inode->i_mapping,
 158                                                   inode->i_size >> CFS_PAGE_SHIFT);
 159                 if (page != NULL) {
 160                         struct ll_async_page *llap = llap_cast_private(page);
 161                         if (llap != NULL) {
 162                                 llap->llap_checksum =
 163                                         crc32_le(0, kmap(page), CFS_PAGE_SIZE);
 164                                 kunmap(page);
 165                         }
 166                         page_cache_release(page);
 167                 }
 168         }
 169
 170         CDEBUG(D_INFO, "calling punch for "LPX64" (new size %Lu=%#Lx)\n",
 171                lli->lli_smd->lsm_object_id, inode->i_size, inode->i_size);
 172
 173         oinfo.oi_md = lli->lli_smd;
 174         oinfo.oi_policy.l_extent.start = inode->i_size;
 175         oinfo.oi_policy.l_extent.end = OBD_OBJECT_EOF;
 176         oinfo.oi_oa = &oa;
 177         oa.o_id = lli->lli_smd->lsm_object_id;
 178         oa.o_valid = OBD_MD_FLID;
 179
 180         obdo_from_inode(&oa, inode, OBD_MD_FLTYPE | OBD_MD_FLMODE |OBD_MD_FLFID|
 181                         OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME |
 182                         OBD_MD_FLUID | OBD_MD_FLGID | OBD_MD_FLGENER |
 183                         OBD_MD_FLBLOCKS);
 184
 185         ll_inode_size_unlock(inode, 0);
 186
 187         rc = obd_punch_rqset(ll_i2obdexp(inode), &oinfo, NULL);
 188         if (rc)
 189                 CERROR("obd_truncate fails (%d) ino %lu\n", rc, inode->i_ino);
 190         else
 191                 obdo_to_inode(inode, &oa, OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
 192                               OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME);
 193         EXIT;
 194         return;
 195
 196  out_unlock:
 197         ll_inode_size_unlock(inode, 0);
 198 } /* ll_truncate */
 199
 200 int ll_prepare_write(struct file *file, struct page *page, unsigned from,
 201                      unsigned to)
 202 {
 203         struct inode *inode = page->mapping->host;
 204         struct ll_inode_info *lli = ll_i2info(inode);
 205         struct lov_stripe_md *lsm = lli->lli_smd;
 206         obd_off offset = ((obd_off)page->index) << CFS_PAGE_SHIFT;
 207         struct obd_info oinfo = { { { 0 } } };
 208         struct brw_page pga;
 209         struct obdo oa;
 210         struct ost_lvb lvb;
 211         int rc = 0;
 212         ENTRY;
 213
 214         LASSERT(PageLocked(page));
 215         (void)llap_cast_private(page); /* assertion */
 216
 217         /* Check to see if we should return -EIO right away */
 218         pga.pg = page;
 219         pga.off = offset;
 220         pga.count = CFS_PAGE_SIZE;
 221         pga.flag = 0;
 222
 223         oa.o_mode = inode->i_mode;
 224         oa.o_id = lsm->lsm_object_id;
 225         oa.o_valid = OBD_MD_FLID | OBD_MD_FLMODE | OBD_MD_FLTYPE;
 226         obdo_from_inode(&oa, inode, OBD_MD_FLFID | OBD_MD_FLGENER);
 227
 228         oinfo.oi_oa = &oa;
 229         oinfo.oi_md = lsm;
 230         rc = obd_brw(OBD_BRW_CHECK, ll_i2obdexp(inode), &oinfo, 1, &pga, NULL);
 231         if (rc)
 232                 RETURN(rc);
 233
 234         if (PageUptodate(page)) {
 235                 LL_CDEBUG_PAGE(D_PAGE, page, "uptodate\n");
 236                 RETURN(0);
 237         }
 238
 239         /* We're completely overwriting an existing page, so _don't_ set it up
 240          * to date until commit_write */
 241         if (from == 0 && to == CFS_PAGE_SIZE) {
 242                 LL_CDEBUG_PAGE(D_PAGE, page, "full page write\n");
 243                 POISON_PAGE(page, 0x11);
 244                 RETURN(0);
 245         }
 246
 247         /* If are writing to a new page, no need to read old data.  The extent
 248          * locking will have updated the KMS, and for our purposes here we can
 249          * treat it like i_size. */
 250         lov_stripe_lock(lsm);
 251         inode_init_lvb(inode, &lvb);
 252         obd_merge_lvb(ll_i2obdexp(inode), lsm, &lvb, 1);
 253         lov_stripe_unlock(lsm);
 254         if (lvb.lvb_size <= offset) {
 255                 LL_CDEBUG_PAGE(D_PAGE, page, "kms "LPU64" <= offset "LPU64"\n",
 256                                lvb.lvb_size, offset);
 257                 memset(kmap(page), 0, CFS_PAGE_SIZE);
 258                 kunmap(page);
 259                 GOTO(prepare_done, rc = 0);
 260         }
 261
 262         /* XXX could be an async ocp read.. read-ahead? */
 263         rc = ll_brw(OBD_BRW_READ, inode, &oa, page, 0);
 264         if (rc == 0) {
 265                 /* bug 1598: don't clobber blksize */
 266                 oa.o_valid &= ~(OBD_MD_FLSIZE | OBD_MD_FLBLKSZ);
 267                 obdo_refresh_inode(inode, &oa, oa.o_valid);
 268         }
 269
 270         EXIT;
 271  prepare_done:
 272         if (rc == 0)
 273                 SetPageUptodate(page);
 274
 275         return rc;
 276 }
 277
 278 static int ll_ap_make_ready(void *data, int cmd)
 279 {
 280         struct ll_async_page *llap;
 281         struct page *page;
 282         ENTRY;
 283
 284         llap = LLAP_FROM_COOKIE(data);
 285         page = llap->llap_page;
 286
 287         LASSERTF(!(cmd & OBD_BRW_READ), "cmd %x page %p ino %lu index %lu\n", cmd, page,
 288                  page->mapping->host->i_ino, page->index);
 289
 290         /* we're trying to write, but the page is locked.. come back later */
 291         if (TryLockPage(page))
 292                 RETURN(-EAGAIN);
 293
 294         LASSERT(!PageWriteback(page));
 295
 296         /* if we left PageDirty we might get another writepage call
 297          * in the future.  list walkers are bright enough
 298          * to check page dirty so we can leave it on whatever list
 299          * its on.  XXX also, we're called with the cli list so if
 300          * we got the page cache list we'd create a lock inversion
 301          * with the removepage path which gets the page lock then the
 302          * cli lock */
 303 #if LINUX_VERSION_CODE < KERNEL_VERSION(2,6,0)
 304         clear_page_dirty(page);
 305 #else
 306         LASSERTF(!PageWriteback(page),"cmd %x page %p ino %lu index %lu\n", cmd, page,
 307                  page->mapping->host->i_ino, page->index);
 308         clear_page_dirty_for_io(page);
 309
 310         /* This actually clears the dirty bit in the radix tree.*/
 311         set_page_writeback(page);
 312 #endif
 313
 314         LL_CDEBUG_PAGE(D_PAGE, page, "made ready\n");
 315         page_cache_get(page);
 316
 317         RETURN(0);
 318 }
 319
 320 /* We have two reasons for giving llite the opportunity to change the
 321  * write length of a given queued page as it builds the RPC containing
 322  * the page:
 323  *
 324  * 1) Further extending writes may have landed in the page cache
 325  *    since a partial write first queued this page requiring us
 326  *    to write more from the page cache.  (No further races are possible, since
 327  *    by the time this is called, the page is locked.)
 328  * 2) We might have raced with truncate and want to avoid performing
 329  *    write RPCs that are just going to be thrown away by the
 330  *    truncate's punch on the storage targets.
 331  *
 332  * The kms serves these purposes as it is set at both truncate and extending
 333  * writes.
 334  */
 335 static int ll_ap_refresh_count(void *data, int cmd)
 336 {
 337         struct ll_inode_info *lli;
 338         struct ll_async_page *llap;
 339         struct lov_stripe_md *lsm;
 340         struct page *page;
 341         struct inode *inode;
 342         struct ost_lvb lvb;
 343         __u64 kms;
 344         ENTRY;
 345
 346         /* readpage queues with _COUNT_STABLE, shouldn't get here. */
 347         LASSERT(cmd != OBD_BRW_READ);
 348
 349         llap = LLAP_FROM_COOKIE(data);
 350         page = llap->llap_page;
 351         inode = page->mapping->host;
 352         lli = ll_i2info(inode);
 353         lsm = lli->lli_smd;
 354
 355         lov_stripe_lock(lsm);
 356         inode_init_lvb(inode, &lvb);
 357         obd_merge_lvb(ll_i2obdexp(inode), lsm, &lvb, 1);
 358         kms = lvb.lvb_size;
 359         lov_stripe_unlock(lsm);
 360
 361         /* catch race with truncate */
 362         if (((__u64)page->index << CFS_PAGE_SHIFT) >= kms)
 363                 return 0;
 364
 365         /* catch sub-page write at end of file */
 366         if (((__u64)page->index << CFS_PAGE_SHIFT) + CFS_PAGE_SIZE > kms)
 367                 return kms % CFS_PAGE_SIZE;
 368
 369         return CFS_PAGE_SIZE;
 370 }
 371
 372 void ll_inode_fill_obdo(struct inode *inode, int cmd, struct obdo *oa)
 373 {
 374         struct lov_stripe_md *lsm;
 375         obd_flag valid_flags;
 376
 377         lsm = ll_i2info(inode)->lli_smd;
 378
 379         oa->o_id = lsm->lsm_object_id;
 380         oa->o_valid = OBD_MD_FLID;
 381         valid_flags = OBD_MD_FLTYPE | OBD_MD_FLATIME;
 382         if (cmd & OBD_BRW_WRITE) {
 383                 oa->o_valid |= OBD_MD_FLEPOCH;
 384                 oa->o_easize = ll_i2info(inode)->lli_io_epoch;
 385
 386                 valid_flags |= OBD_MD_FLMTIME | OBD_MD_FLCTIME |
 387                         OBD_MD_FLUID | OBD_MD_FLGID |
 388                         OBD_MD_FLFID | OBD_MD_FLGENER;
 389         }
 390
 391         obdo_from_inode(oa, inode, valid_flags);
 392 }
 393
 394 static void ll_ap_fill_obdo(void *data, int cmd, struct obdo *oa)
 395 {
 396         struct ll_async_page *llap;
 397         ENTRY;
 398
 399         llap = LLAP_FROM_COOKIE(data);
 400         ll_inode_fill_obdo(llap->llap_page->mapping->host, cmd, oa);
 401
 402         EXIT;
 403 }
 404
 405 static void ll_ap_update_obdo(void *data, int cmd, struct obdo *oa,
 406                               obd_valid valid)
 407 {
 408         struct ll_async_page *llap;
 409         ENTRY;
 410
 411         llap = LLAP_FROM_COOKIE(data);
 412         obdo_from_inode(oa, llap->llap_page->mapping->host, valid);
 413
 414         EXIT;
 415 }
 416
 417 static struct obd_async_page_ops ll_async_page_ops = {
 418         .ap_make_ready =        ll_ap_make_ready,
 419         .ap_refresh_count =     ll_ap_refresh_count,
 420         .ap_fill_obdo =         ll_ap_fill_obdo,
 421         .ap_update_obdo =       ll_ap_update_obdo,
 422         .ap_completion =        ll_ap_completion,
 423 };
 424
 425 struct ll_async_page *llap_cast_private(struct page *page)
 426 {
 427         struct ll_async_page *llap = (struct ll_async_page *)page_private(page);
 428
 429         LASSERTF(llap == NULL || llap->llap_magic == LLAP_MAGIC,
 430                  "page %p private %lu gave magic %d which != %d\n",
 431                  page, page_private(page), llap->llap_magic, LLAP_MAGIC);
 432
 433         return llap;
 434 }
 435
 436 /* Try to shrink the page cache for the @sbi filesystem by 1/@shrink_fraction.
 437  *
 438  * There is an llap attached onto every page in lustre, linked off @sbi.
 439  * We add an llap to the list so we don't lose our place during list walking.
 440  * If llaps in the list are being moved they will only move to the end
 441  * of the LRU, and we aren't terribly interested in those pages here (we
 442  * start at the beginning of the list where the least-used llaps are.
 443  */
 444 int llap_shrink_cache(struct ll_sb_info *sbi, int shrink_fraction)
 445 {
 446         struct ll_async_page *llap, dummy_llap = { .llap_magic = 0xd11ad11a };
 447         unsigned long total, want, count = 0;
 448
 449         total = sbi->ll_async_page_count;
 450
 451         /* There can be a large number of llaps (600k or more in a large
 452          * memory machine) so the VM 1/6 shrink ratio is likely too much.
 453          * Since we are freeing pages also, we don't necessarily want to
 454          * shrink so much.  Limit to 40MB of pages + llaps per call. */
 455         if (shrink_fraction == 0)
 456                 want = sbi->ll_async_page_count - sbi->ll_async_page_max + 32;
 457         else
 458                 want = (total + shrink_fraction - 1) / shrink_fraction;
 459
 460         if (want > 40 << (20 - CFS_PAGE_SHIFT))
 461                 want = 40 << (20 - CFS_PAGE_SHIFT);
 462
 463         CDEBUG(D_CACHE, "shrinking %lu of %lu pages (1/%d)\n",
 464                want, total, shrink_fraction);
 465
 466         spin_lock(&sbi->ll_lock);
 467         list_add(&dummy_llap.llap_pglist_item, &sbi->ll_pglist);
 468
 469         while (--total >= 0 && count < want) {
 470                 struct page *page;
 471                 int keep;
 472
 473                 if (unlikely(need_resched())) {
 474                         spin_unlock(&sbi->ll_lock);
 475                         cond_resched();
 476                         spin_lock(&sbi->ll_lock);
 477                 }
 478
 479                 llap = llite_pglist_next_llap(sbi,&dummy_llap.llap_pglist_item);
 480                 list_del_init(&dummy_llap.llap_pglist_item);
 481                 if (llap == NULL)
 482                         break;
 483
 484                 page = llap->llap_page;
 485                 LASSERT(page != NULL);
 486
 487                 list_add(&dummy_llap.llap_pglist_item, &llap->llap_pglist_item);
 488
 489                 /* Page needs/undergoing IO */
 490                 if (TryLockPage(page)) {
 491                         LL_CDEBUG_PAGE(D_PAGE, page, "can't lock\n");
 492                         continue;
 493                 }
 494
 495                 if (llap->llap_write_queued || PageDirty(page) ||
 496                     (!PageUptodate(page) &&
 497                      llap->llap_origin != LLAP_ORIGIN_READAHEAD))
 498                         keep = 1;
 499                 else
 500                         keep = 0;
 501
 502                 LL_CDEBUG_PAGE(D_PAGE, page,"%s LRU page: %s%s%s%s origin %s\n",
 503                                keep ? "keep" : "drop",
 504                                llap->llap_write_queued ? "wq " : "",
 505                                PageDirty(page) ? "pd " : "",
 506                                PageUptodate(page) ? "" : "!pu ",
 507                                llap->llap_defer_uptodate ? "" : "!du",
 508                                llap_origins[llap->llap_origin]);
 509
 510                 /* If page is dirty or undergoing IO don't discard it */
 511                 if (keep) {
 512                         unlock_page(page);
 513                         continue;
 514                 }
 515
 516                 page_cache_get(page);
 517                 spin_unlock(&sbi->ll_lock);
 518
 519                 if (page->mapping != NULL) {
 520                         ll_teardown_mmaps(page->mapping,
 521                                          (__u64)page->index << CFS_PAGE_SHIFT,
 522                                          ((__u64)page->index << CFS_PAGE_SHIFT)|
 523                                           ~CFS_PAGE_MASK);
 524                         if (!PageDirty(page) && !page_mapped(page)) {
 525                                 ll_ra_accounting(llap, page->mapping);
 526                                 ll_truncate_complete_page(page);
 527                                 ++count;
 528                         } else {
 529                                 LL_CDEBUG_PAGE(D_PAGE, page, "Not dropping page"
 530                                                              " because it is "
 531                                                              "%s\n",
 532                                                               PageDirty(page)?
 533                                                               "dirty":"mapped");
 534                         }
 535                 }
 536                 unlock_page(page);
 537                 page_cache_release(page);
 538
 539                 spin_lock(&sbi->ll_lock);
 540         }
 541         list_del(&dummy_llap.llap_pglist_item);
 542         spin_unlock(&sbi->ll_lock);
 543
 544         CDEBUG(D_CACHE, "shrank %lu/%lu and left %lu unscanned\n",
 545                count, want, total);
 546
 547         return count;
 548 }
 549
 550 static struct ll_async_page *llap_from_page(struct page *page, unsigned origin)
 551 {
 552         struct ll_async_page *llap;
 553         struct obd_export *exp;
 554         struct inode *inode = page->mapping->host;
 555         struct ll_sb_info *sbi;
 556         int rc;
 557         ENTRY;
 558
 559         if (!inode) {
 560                 static int triggered;
 561
 562                 if (!triggered) {
 563                         LL_CDEBUG_PAGE(D_ERROR, page, "Bug 10047. Wrong anon "
 564                                        "page received\n");
 565                         libcfs_debug_dumpstack(NULL);
 566                         triggered = 1;
 567                 }
 568                 RETURN(ERR_PTR(-EINVAL));
 569         }
 570         sbi = ll_i2sbi(inode);
 571         LASSERT(ll_async_page_slab);
 572         LASSERTF(origin < LLAP__ORIGIN_MAX, "%u\n", origin);
 573
 574         llap = llap_cast_private(page);
 575         if (llap != NULL) {
 576                 /* move to end of LRU list, except when page is just about to
 577                  * die */
 578                 if (origin != LLAP_ORIGIN_REMOVEPAGE) {
 579                         spin_lock(&sbi->ll_lock);
 580                         sbi->ll_pglist_gen++;
 581                         list_del_init(&llap->llap_pglist_item);
 582                         list_add_tail(&llap->llap_pglist_item, &sbi->ll_pglist);
 583                         spin_unlock(&sbi->ll_lock);
 584                 }
 585                 GOTO(out, llap);
 586         }
 587
 588         exp = ll_i2obdexp(page->mapping->host);
 589         if (exp == NULL)
 590                 RETURN(ERR_PTR(-EINVAL));
 591
 592         /* limit the number of lustre-cached pages */
 593         if (sbi->ll_async_page_count >= sbi->ll_async_page_max)
 594                 llap_shrink_cache(sbi, 0);
 595
 596         OBD_SLAB_ALLOC(llap, ll_async_page_slab, SLAB_KERNEL,
 597                        ll_async_page_slab_size);
 598         if (llap == NULL)
 599                 RETURN(ERR_PTR(-ENOMEM));
 600         llap->llap_magic = LLAP_MAGIC;
 601         llap->llap_cookie = (void *)llap + size_round(sizeof(*llap));
 602
 603         rc = obd_prep_async_page(exp, ll_i2info(inode)->lli_smd, NULL, page,
 604                                  (obd_off)page->index << CFS_PAGE_SHIFT,
 605                                  &ll_async_page_ops, llap, &llap->llap_cookie);
 606         if (rc) {
 607                 OBD_SLAB_FREE(llap, ll_async_page_slab,
 608                               ll_async_page_slab_size);
 609                 RETURN(ERR_PTR(rc));
 610         }
 611
 612         CDEBUG(D_CACHE, "llap %p page %p cookie %p obj off "LPU64"\n", llap,
 613                page, llap->llap_cookie, (obd_off)page->index << CFS_PAGE_SHIFT);
 614         /* also zeroing the PRIVBITS low order bitflags */
 615         __set_page_ll_data(page, llap);
 616         llap->llap_page = page;
 617
 618         spin_lock(&sbi->ll_lock);
 619         sbi->ll_pglist_gen++;
 620         sbi->ll_async_page_count++;
 621         list_add_tail(&llap->llap_pglist_item, &sbi->ll_pglist);
 622         spin_unlock(&sbi->ll_lock);
 623
 624  out:
 625         if (unlikely(sbi->ll_flags & LL_SBI_CHECKSUM)) {
 626                 __u32 csum = 0;
 627                 csum = crc32_le(csum, kmap(page), CFS_PAGE_SIZE);
 628                 kunmap(page);
 629                 if (origin == LLAP_ORIGIN_READAHEAD ||
 630                     origin == LLAP_ORIGIN_READPAGE) {
 631                         llap->llap_checksum = 0;
 632                 } else if (origin == LLAP_ORIGIN_COMMIT_WRITE ||
 633                            llap->llap_checksum == 0) {
 634                         llap->llap_checksum = csum;
 635                         CDEBUG(D_PAGE, "page %p cksum %x\n", page, csum);
 636                 } else if (llap->llap_checksum == csum) {
 637                         /* origin == LLAP_ORIGIN_WRITEPAGE */
 638                         CDEBUG(D_PAGE, "page %p cksum %x confirmed\n",
 639                                page, csum);
 640                 } else {
 641                         /* origin == LLAP_ORIGIN_WRITEPAGE */
 642                         LL_CDEBUG_PAGE(D_ERROR, page, "old cksum %x != new "
 643                                        "%x!\n", llap->llap_checksum, csum);
 644                 }
 645         }
 646
 647         llap->llap_origin = origin;
 648         RETURN(llap);
 649 }
 650
 651 static int queue_or_sync_write(struct obd_export *exp, struct inode *inode,
 652                                struct ll_async_page *llap,
 653                                unsigned to, obd_flag async_flags)
 654 {
 655         unsigned long size_index = inode->i_size >> CFS_PAGE_SHIFT;
 656         struct obd_io_group *oig;
 657         struct ll_sb_info *sbi = ll_i2sbi(inode);
 658         int rc, noquot = llap->llap_ignore_quota ? OBD_BRW_NOQUOTA : 0;
 659         ENTRY;
 660
 661         /* _make_ready only sees llap once we've unlocked the page */
 662         llap->llap_write_queued = 1;
 663         rc = obd_queue_async_io(exp, ll_i2info(inode)->lli_smd, NULL,
 664                                 llap->llap_cookie, OBD_BRW_WRITE | noquot,
 665                                 0, 0, 0, async_flags);
 666         if (rc == 0) {
 667                 LL_CDEBUG_PAGE(D_PAGE, llap->llap_page, "write queued\n");
 668                 //llap_write_pending(inode, llap);
 669                 GOTO(out, 0);
 670         }
 671
 672         llap->llap_write_queued = 0;
 673
 674         rc = oig_init(&oig);
 675         if (rc)
 676                 GOTO(out, rc);
 677
 678         /* make full-page requests if we are not at EOF (bug 4410) */
 679         if (to != CFS_PAGE_SIZE && llap->llap_page->index < size_index) {
 680                 LL_CDEBUG_PAGE(D_PAGE, llap->llap_page,
 681                                "sync write before EOF: size_index %lu, to %d\n",
 682                                size_index, to);
 683                 to = CFS_PAGE_SIZE;
 684         } else if (to != CFS_PAGE_SIZE && llap->llap_page->index == size_index) {
 685                 int size_to = inode->i_size & ~CFS_PAGE_MASK;
 686                 LL_CDEBUG_PAGE(D_PAGE, llap->llap_page,
 687                                "sync write at EOF: size_index %lu, to %d/%d\n",
 688                                size_index, to, size_to);
 689                 if (to < size_to)
 690                         to = size_to;
 691         }
 692
 693         /* compare the checksum once before the page leaves llite */
 694         if (unlikely((sbi->ll_flags & LL_SBI_CHECKSUM) &&
 695                      llap->llap_checksum != 0)) {
 696                 __u32 csum = 0;
 697                 struct page *page = llap->llap_page;
 698                 csum = crc32_le(csum, kmap(page), CFS_PAGE_SIZE);
 699                 kunmap(page);
 700                 if (llap->llap_checksum == csum) {
 701                         CDEBUG(D_PAGE, "page %p cksum %x confirmed\n",
 702                                page, csum);
 703                 } else {
 704                         CERROR("page %p old cksum %x != new cksum %x!\n",
 705                                page, llap->llap_checksum, csum);
 706                 }
 707         }
 708
 709         rc = obd_queue_group_io(exp, ll_i2info(inode)->lli_smd, NULL, oig,
 710                                 llap->llap_cookie, OBD_BRW_WRITE | noquot,
 711                                 0, to, 0, ASYNC_READY | ASYNC_URGENT |
 712                                 ASYNC_COUNT_STABLE | ASYNC_GROUP_SYNC);
 713         if (rc)
 714                 GOTO(free_oig, rc);
 715
 716         rc = obd_trigger_group_io(exp, ll_i2info(inode)->lli_smd, NULL, oig);
 717         if (rc)
 718                 GOTO(free_oig, rc);
 719
 720         rc = oig_wait(oig);
 721
 722         if (!rc && async_flags & ASYNC_READY) {
 723                 unlock_page(llap->llap_page);
 724                 if (PageWriteback(llap->llap_page)) {
 725                         end_page_writeback(llap->llap_page);
 726                 }
 727         }
 728
 729         LL_CDEBUG_PAGE(D_PAGE, llap->llap_page, "sync write returned %d\n", rc);
 730
 731 free_oig:
 732         oig_release(oig);
 733 out:
 734         RETURN(rc);
 735 }
 736
 737 /* update our write count to account for i_size increases that may have
 738  * happened since we've queued the page for io. */
 739
 740 /* be careful not to return success without setting the page Uptodate or
 741  * the next pass through prepare_write will read in stale data from disk. */
 742 int ll_commit_write(struct file *file, struct page *page, unsigned from,
 743                     unsigned to)
 744 {
 745         struct inode *inode = page->mapping->host;
 746         struct ll_inode_info *lli = ll_i2info(inode);
 747         struct lov_stripe_md *lsm = lli->lli_smd;
 748         struct obd_export *exp;
 749         struct ll_async_page *llap;
 750         loff_t size;
 751         int rc = 0;
 752         ENTRY;
 753
 754         SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
 755         LASSERT(inode == file->f_dentry->d_inode);
 756         LASSERT(PageLocked(page));
 757
 758         CDEBUG(D_INODE, "inode %p is writing page %p from %d to %d at %lu\n",
 759                inode, page, from, to, page->index);
 760
 761         llap = llap_from_page(page, LLAP_ORIGIN_COMMIT_WRITE);
 762         if (IS_ERR(llap))
 763                 RETURN(PTR_ERR(llap));
 764
 765         exp = ll_i2obdexp(inode);
 766         if (exp == NULL)
 767                 RETURN(-EINVAL);
 768
 769         llap->llap_ignore_quota = capable(CAP_SYS_RESOURCE);
 770
 771         /* queue a write for some time in the future the first time we
 772          * dirty the page */
 773         if (!PageDirty(page)) {
 774                 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats,
 775                                      LPROC_LL_DIRTY_MISSES);
 776
 777                 rc = queue_or_sync_write(exp, inode, llap, to, 0);
 778                 if (rc)
 779                         GOTO(out, rc);
 780         } else {
 781                 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats,
 782                                      LPROC_LL_DIRTY_HITS);
 783         }
 784
 785         /* put the page in the page cache, from now on ll_removepage is
 786          * responsible for cleaning up the llap.
 787          * only set page dirty when it's queued to be write out */
 788         if (llap->llap_write_queued)
 789                 set_page_dirty(page);
 790
 791 out:
 792         size = (((obd_off)page->index) << CFS_PAGE_SHIFT) + to;
 793         ll_inode_size_lock(inode, 0);
 794         if (rc == 0) {
 795                 lov_stripe_lock(lsm);
 796                 obd_adjust_kms(exp, lsm, size, 0);
 797                 lov_stripe_unlock(lsm);
 798                 if (size > inode->i_size)
 799                         inode->i_size = size;
 800                 SetPageUptodate(page);
 801         } else if (size > inode->i_size) {
 802                 /* this page beyond the pales of i_size, so it can't be
 803                  * truncated in ll_p_r_e during lock revoking. we must
 804                  * teardown our book-keeping here. */
 805                 ll_removepage(page);
 806         }
 807         ll_inode_size_unlock(inode, 0);
 808         RETURN(rc);
 809 }
 810
 811 static unsigned long ll_ra_count_get(struct ll_sb_info *sbi, unsigned long len)
 812 {
 813         struct ll_ra_info *ra = &sbi->ll_ra_info;
 814         unsigned long ret;
 815         ENTRY;
 816
 817         spin_lock(&sbi->ll_lock);
 818         ret = min(ra->ra_max_pages - ra->ra_cur_pages, len);
 819         ra->ra_cur_pages += ret;
 820         spin_unlock(&sbi->ll_lock);
 821
 822         RETURN(ret);
 823 }
 824
 825 static void ll_ra_count_put(struct ll_sb_info *sbi, unsigned long len)
 826 {
 827         struct ll_ra_info *ra = &sbi->ll_ra_info;
 828         spin_lock(&sbi->ll_lock);
 829         LASSERTF(ra->ra_cur_pages >= len, "r_c_p %lu len %lu\n",
 830                  ra->ra_cur_pages, len);
 831         ra->ra_cur_pages -= len;
 832         spin_unlock(&sbi->ll_lock);
 833 }
 834
 835 /* called for each page in a completed rpc.*/
 836 int ll_ap_completion(void *data, int cmd, struct obdo *oa, int rc)
 837 {
 838         struct ll_async_page *llap;
 839         struct page *page;
 840         int ret = 0;
 841         ENTRY;
 842
 843         llap = LLAP_FROM_COOKIE(data);
 844         page = llap->llap_page;
 845         LASSERT(PageLocked(page));
 846         LASSERT(CheckWriteback(page,cmd));
 847
 848         LL_CDEBUG_PAGE(D_PAGE, page, "completing cmd %d with %d\n", cmd, rc);
 849
 850         if (cmd & OBD_BRW_READ && llap->llap_defer_uptodate)
 851                 ll_ra_count_put(ll_i2sbi(page->mapping->host), 1);
 852
 853         if (rc == 0)  {
 854                 if (cmd & OBD_BRW_READ) {
 855                         if (!llap->llap_defer_uptodate)
 856                                 SetPageUptodate(page);
 857                 } else {
 858                         llap->llap_write_queued = 0;
 859                 }
 860                 ClearPageError(page);
 861         } else {
 862                 if (cmd & OBD_BRW_READ) {
 863                         llap->llap_defer_uptodate = 0;
 864                 } else {
 865                         ll_redirty_page(page);
 866                         ret = 1;
 867                 }
 868                 SetPageError(page);
 869         }
 870
 871         unlock_page(page);
 872
 873         if (0 && cmd & OBD_BRW_WRITE) {
 874                 llap_write_complete(page->mapping->host, llap);
 875                 ll_try_done_writing(page->mapping->host);
 876         }
 877
 878         if (PageWriteback(page)) {
 879                 end_page_writeback(page);
 880         }
 881         page_cache_release(page);
 882
 883         RETURN(ret);
 884 }
 885
 886 /* the kernel calls us here when a page is unhashed from the page cache.
 887  * the page will be locked and the kernel is holding a spinlock, so
 888  * we need to be careful.  we're just tearing down our book-keeping
 889  * here. */
 890 void ll_removepage(struct page *page)
 891 {
 892         struct inode *inode = page->mapping->host;
 893         struct obd_export *exp;
 894         struct ll_async_page *llap;
 895         struct ll_sb_info *sbi = ll_i2sbi(inode);
 896         int rc;
 897         ENTRY;
 898
 899         LASSERT(!in_interrupt());
 900
 901         /* sync pages or failed read pages can leave pages in the page
 902          * cache that don't have our data associated with them anymore */
 903         if (page_private(page) == 0) {
 904                 EXIT;
 905                 return;
 906         }
 907
 908         LL_CDEBUG_PAGE(D_PAGE, page, "being evicted\n");
 909
 910         exp = ll_i2obdexp(inode);
 911         if (exp == NULL) {
 912                 CERROR("page %p ind %lu gave null export\n", page, page->index);
 913                 EXIT;
 914                 return;
 915         }
 916
 917         llap = llap_from_page(page, 0);
 918         if (IS_ERR(llap)) {
 919                 CERROR("page %p ind %lu couldn't find llap: %ld\n", page,
 920                        page->index, PTR_ERR(llap));
 921                 EXIT;
 922                 return;
 923         }
 924
 925         //llap_write_complete(inode, llap);
 926         rc = obd_teardown_async_page(exp, ll_i2info(inode)->lli_smd, NULL,
 927                                      llap->llap_cookie);
 928         if (rc != 0)
 929                 CERROR("page %p ind %lu failed: %d\n", page, page->index, rc);
 930
 931         /* this unconditional free is only safe because the page lock
 932          * is providing exclusivity to memory pressure/truncate/writeback..*/
 933         __clear_page_ll_data(page);
 934
 935         spin_lock(&sbi->ll_lock);
 936         if (!list_empty(&llap->llap_pglist_item))
 937                 list_del_init(&llap->llap_pglist_item);
 938         sbi->ll_pglist_gen++;
 939         sbi->ll_async_page_count--;
 940         spin_unlock(&sbi->ll_lock);
 941         OBD_SLAB_FREE(llap, ll_async_page_slab, ll_async_page_slab_size);
 942         EXIT;
 943 }
 944
 945 static int ll_page_matches(struct page *page, int fd_flags)
 946 {
 947         struct lustre_handle match_lockh = {0};
 948         struct inode *inode = page->mapping->host;
 949         ldlm_policy_data_t page_extent;
 950         int flags, matches;
 951         ENTRY;
 952
 953         if (unlikely(fd_flags & LL_FILE_GROUP_LOCKED))
 954                 RETURN(1);
 955
 956         page_extent.l_extent.start = (__u64)page->index << CFS_PAGE_SHIFT;
 957         page_extent.l_extent.end =
 958                 page_extent.l_extent.start + CFS_PAGE_SIZE - 1;
 959         flags = LDLM_FL_TEST_LOCK | LDLM_FL_BLOCK_GRANTED;
 960         if (!(fd_flags & LL_FILE_READAHEAD))
 961                 flags |= LDLM_FL_CBPENDING;
 962         matches = obd_match(ll_i2sbi(inode)->ll_osc_exp,
 963                             ll_i2info(inode)->lli_smd, LDLM_EXTENT,
 964                             &page_extent, LCK_PR | LCK_PW, &flags, inode,
 965                             &match_lockh);
 966         RETURN(matches);
 967 }
 968
 969 static int ll_issue_page_read(struct obd_export *exp,
 970                               struct ll_async_page *llap,
 971                               struct obd_io_group *oig, int defer)
 972 {
 973         struct page *page = llap->llap_page;
 974         int rc;
 975
 976         page_cache_get(page);
 977         llap->llap_defer_uptodate = defer;
 978         llap->llap_ra_used = 0;
 979         rc = obd_queue_group_io(exp, ll_i2info(page->mapping->host)->lli_smd,
 980                                 NULL, oig, llap->llap_cookie, OBD_BRW_READ, 0,
 981                                 CFS_PAGE_SIZE, 0, ASYNC_COUNT_STABLE | ASYNC_READY |
 982                                               ASYNC_URGENT);
 983         if (rc) {
 984                 LL_CDEBUG_PAGE(D_ERROR, page, "read queue failed: rc %d\n", rc);
 985                 page_cache_release(page);
 986         }
 987         RETURN(rc);
 988 }
 989
 990 static void ll_ra_stats_inc_unlocked(struct ll_ra_info *ra, enum ra_stat which)
 991 {
 992         LASSERTF(which >= 0 && which < _NR_RA_STAT, "which: %u\n", which);
 993         ra->ra_stats[which]++;
 994 }
 995
 996 static void ll_ra_stats_inc(struct address_space *mapping, enum ra_stat which)
 997 {
 998         struct ll_sb_info *sbi = ll_i2sbi(mapping->host);
 999         struct ll_ra_info *ra = &ll_i2sbi(mapping->host)->ll_ra_info;
1000
1001         spin_lock(&sbi->ll_lock);
1002         ll_ra_stats_inc_unlocked(ra, which);
1003         spin_unlock(&sbi->ll_lock);
1004 }
1005
1006 void ll_ra_accounting(struct ll_async_page *llap, struct address_space *mapping)
1007 {
1008         if (!llap->llap_defer_uptodate || llap->llap_ra_used)
1009                 return;
1010
1011         ll_ra_stats_inc(mapping, RA_STAT_DISCARDED);
1012 }
1013
1014 #define RAS_CDEBUG(ras) \
1015         CDEBUG(D_READA,                                                      \
1016                "lrp %lu cr %lu cp %lu ws %lu wl %lu nra %lu r %lu ri %lu\n", \
1017                ras->ras_last_readpage, ras->ras_consecutive_requests,        \
1018                ras->ras_consecutive_pages, ras->ras_window_start,            \
1019                ras->ras_window_len, ras->ras_next_readahead,                 \
1020                ras->ras_requests, ras->ras_request_index);
1021
1022 static int index_in_window(unsigned long index, unsigned long point,
1023                            unsigned long before, unsigned long after)
1024 {
1025         unsigned long start = point - before, end = point + after;
1026
1027         if (start > point)
1028                start = 0;
1029         if (end < point)
1030                end = ~0;
1031
1032         return start <= index && index <= end;
1033 }
1034
1035 static struct ll_readahead_state *ll_ras_get(struct file *f)
1036 {
1037         struct ll_file_data       *fd;
1038
1039         fd = LUSTRE_FPRIVATE(f);
1040         return &fd->fd_ras;
1041 }
1042
1043 void ll_ra_read_in(struct file *f, struct ll_ra_read *rar)
1044 {
1045         struct ll_readahead_state *ras;
1046
1047         ras = ll_ras_get(f);
1048
1049         spin_lock(&ras->ras_lock);
1050         ras->ras_requests++;
1051         ras->ras_request_index = 0;
1052         ras->ras_consecutive_requests++;
1053         rar->lrr_reader = current;
1054
1055         list_add(&rar->lrr_linkage, &ras->ras_read_beads);
1056         spin_unlock(&ras->ras_lock);
1057 }
1058
1059 void ll_ra_read_ex(struct file *f, struct ll_ra_read *rar)
1060 {
1061         struct ll_readahead_state *ras;
1062
1063         ras = ll_ras_get(f);
1064
1065         spin_lock(&ras->ras_lock);
1066         list_del_init(&rar->lrr_linkage);
1067         spin_unlock(&ras->ras_lock);
1068 }
1069
1070 static struct ll_ra_read *ll_ra_read_get_locked(struct ll_readahead_state *ras)
1071 {
1072         struct ll_ra_read *scan;
1073
1074         list_for_each_entry(scan, &ras->ras_read_beads, lrr_linkage) {
1075                 if (scan->lrr_reader == current)
1076                         return scan;
1077         }
1078         return NULL;
1079 }
1080
1081 struct ll_ra_read *ll_ra_read_get(struct file *f)
1082 {
1083         struct ll_readahead_state *ras;
1084         struct ll_ra_read         *bead;
1085
1086         ras = ll_ras_get(f);
1087
1088         spin_lock(&ras->ras_lock);
1089         bead = ll_ra_read_get_locked(ras);
1090         spin_unlock(&ras->ras_lock);
1091         return bead;
1092 }
1093
1094 static int ll_readahead(struct ll_readahead_state *ras,
1095                          struct obd_export *exp, struct address_space *mapping,
1096                          struct obd_io_group *oig, int flags)
1097 {
1098         unsigned long i, start = 0, end = 0, reserved;
1099         struct ll_async_page *llap;
1100         struct page *page;
1101         int rc, ret = 0, match_failed = 0;
1102         __u64 kms;
1103         unsigned int gfp_mask;
1104         struct inode *inode;
1105         struct lov_stripe_md *lsm;
1106         struct ll_ra_read *bead;
1107         struct ost_lvb lvb;
1108         ENTRY;
1109
1110         inode = mapping->host;
1111         lsm = ll_i2info(inode)->lli_smd;
1112
1113         lov_stripe_lock(lsm);
1114         inode_init_lvb(inode, &lvb);
1115         obd_merge_lvb(ll_i2obdexp(inode), lsm, &lvb, 1);
1116         kms = lvb.lvb_size;
1117         lov_stripe_unlock(lsm);
1118         if (kms == 0) {
1119                 ll_ra_stats_inc(mapping, RA_STAT_ZERO_LEN);
1120                 RETURN(0);
1121         }
1122
1123         spin_lock(&ras->ras_lock);
1124         bead = ll_ra_read_get_locked(ras);
1125         /* Enlarge the RA window to encompass the full read */
1126         if (bead != NULL && ras->ras_window_start + ras->ras_window_len <
1127             bead->lrr_start + bead->lrr_count) {
1128                 ras->ras_window_len = bead->lrr_start + bead->lrr_count -
1129                                       ras->ras_window_start;
1130         }
1131         /* Reserve a part of the read-ahead window that we'll be issuing */
1132         if (ras->ras_window_len) {
1133                 start = ras->ras_next_readahead;
1134                 end = ras->ras_window_start + ras->ras_window_len - 1;
1135         }
1136         if (end != 0) {
1137                 /* Truncate RA window to end of file */
1138                 end = min(end, (unsigned long)((kms - 1) >> CFS_PAGE_SHIFT));
1139                 ras->ras_next_readahead = max(end, end + 1);
1140                 RAS_CDEBUG(ras);
1141         }
1142         spin_unlock(&ras->ras_lock);
1143
1144         if (end == 0) {
1145                 ll_ra_stats_inc(mapping, RA_STAT_ZERO_WINDOW);
1146                 RETURN(0);
1147         }
1148
1149         reserved = ll_ra_count_get(ll_i2sbi(inode), end - start + 1);
1150         if (reserved < end - start + 1)
1151                 ll_ra_stats_inc(mapping, RA_STAT_MAX_IN_FLIGHT);
1152
1153         gfp_mask = GFP_HIGHUSER & ~__GFP_WAIT;
1154 #ifdef __GFP_NOWARN
1155         gfp_mask |= __GFP_NOWARN;
1156 #endif
1157
1158         for (i = start; reserved > 0 && !match_failed && i <= end; i++) {
1159                 /* skip locked pages from previous readpage calls */
1160                 page = grab_cache_page_nowait_gfp(mapping, i, gfp_mask);
1161                 if (page == NULL) {
1162                         ll_ra_stats_inc(mapping, RA_STAT_FAILED_GRAB_PAGE);
1163                         CDEBUG(D_READA, "g_c_p_n failed\n");
1164                         continue;
1165                 }
1166
1167                 /* Check if page was truncated or reclaimed */
1168                 if (page->mapping != mapping) {
1169                         ll_ra_stats_inc(mapping, RA_STAT_WRONG_GRAB_PAGE);
1170                         CDEBUG(D_READA, "g_c_p_n returned invalid page\n");
1171                         goto next_page;
1172                 }
1173
1174                 /* we do this first so that we can see the page in the /proc
1175                  * accounting */
1176                 llap = llap_from_page(page, LLAP_ORIGIN_READAHEAD);
1177                 if (IS_ERR(llap) || llap->llap_defer_uptodate)
1178                         goto next_page;
1179
1180                 /* skip completed pages */
1181                 if (Page_Uptodate(page))
1182                         goto next_page;
1183
1184                 /* bail when we hit the end of the lock. */
1185                 if ((rc = ll_page_matches(page, flags|LL_FILE_READAHEAD)) <= 0){
1186                         LL_CDEBUG_PAGE(D_READA | D_PAGE, page,
1187                                        "lock match failed: rc %d\n", rc);
1188                         ll_ra_stats_inc(mapping, RA_STAT_FAILED_MATCH);
1189                         match_failed = 1;
1190                         goto next_page;
1191                 }
1192
1193                 rc = ll_issue_page_read(exp, llap, oig, 1);
1194                 if (rc == 0) {
1195                         reserved--;
1196                         ret++;
1197                         LL_CDEBUG_PAGE(D_READA| D_PAGE, page,
1198                                        "started read-ahead\n");
1199                 } else {
1200         next_page:
1201                         LL_CDEBUG_PAGE(D_READA | D_PAGE, page,
1202                                        "skipping read-ahead\n");
1203
1204                         unlock_page(page);
1205                 }
1206                 page_cache_release(page);
1207         }
1208
1209         LASSERTF(reserved >= 0, "reserved %lu\n", reserved);
1210         if (reserved != 0)
1211                 ll_ra_count_put(ll_i2sbi(inode), reserved);
1212         if (i == end + 1 && end == (kms >> CFS_PAGE_SHIFT))
1213                 ll_ra_stats_inc(mapping, RA_STAT_EOF);
1214
1215         /* if we didn't get to the end of the region we reserved from
1216          * the ras we need to go back and update the ras so that the
1217          * next read-ahead tries from where we left off.  we only do so
1218          * if the region we failed to issue read-ahead on is still ahead
1219          * of the app and behind the next index to start read-ahead from */
1220         if (i != end + 1) {
1221                 spin_lock(&ras->ras_lock);
1222                 if (i < ras->ras_next_readahead &&
1223                     index_in_window(i, ras->ras_window_start, 0,
1224                                     ras->ras_window_len)) {
1225                         ras->ras_next_readahead = i;
1226                         RAS_CDEBUG(ras);
1227                 }
1228                 spin_unlock(&ras->ras_lock);
1229         }
1230
1231         RETURN(ret);
1232 }
1233
1234 static void ras_set_start(struct ll_readahead_state *ras, unsigned long index)
1235 {
1236         ras->ras_window_start = index & (~((1024 * 1024 >> CFS_PAGE_SHIFT) - 1));
1237 }
1238
1239 /* called with the ras_lock held or from places where it doesn't matter */
1240 static void ras_reset(struct ll_readahead_state *ras, unsigned long index)
1241 {
1242         ras->ras_last_readpage = index;
1243         ras->ras_consecutive_requests = 0;
1244         ras->ras_consecutive_pages = 0;
1245         ras->ras_window_len = 0;
1246         ras_set_start(ras, index);
1247         ras->ras_next_readahead = max(ras->ras_window_start, index);
1248
1249         RAS_CDEBUG(ras);
1250 }
1251
1252 void ll_readahead_init(struct inode *inode, struct ll_readahead_state *ras)
1253 {
1254         spin_lock_init(&ras->ras_lock);
1255         ras_reset(ras, 0);
1256         ras->ras_requests = 0;
1257         INIT_LIST_HEAD(&ras->ras_read_beads);
1258 }
1259
1260 static void ras_update(struct ll_sb_info *sbi, struct inode *inode,
1261                        struct ll_readahead_state *ras, unsigned long index,
1262                        unsigned hit)
1263 {
1264         struct ll_ra_info *ra = &sbi->ll_ra_info;
1265         int zero = 0;
1266         ENTRY;
1267
1268         spin_lock(&sbi->ll_lock);
1269         spin_lock(&ras->ras_lock);
1270
1271         ll_ra_stats_inc_unlocked(ra, hit ? RA_STAT_HIT : RA_STAT_MISS);
1272
1273         /* reset the read-ahead window in two cases.  First when the app seeks
1274          * or reads to some other part of the file.  Secondly if we get a
1275          * read-ahead miss that we think we've previously issued.  This can
1276          * be a symptom of there being so many read-ahead pages that the VM is
1277          * reclaiming it before we get to it. */
1278         if (!index_in_window(index, ras->ras_last_readpage, 8, 8)) {
1279                 zero = 1;
1280                 ll_ra_stats_inc_unlocked(ra, RA_STAT_DISTANT_READPAGE);
1281         } else if (!hit && ras->ras_window_len &&
1282                    index < ras->ras_next_readahead &&
1283                    index_in_window(index, ras->ras_window_start, 0,
1284                                    ras->ras_window_len)) {
1285                 zero = 1;
1286                 ll_ra_stats_inc_unlocked(ra, RA_STAT_MISS_IN_WINDOW);
1287         }
1288
1289         /* On the second access to a file smaller than the tunable
1290          * ra_max_read_ahead_whole_pages trigger RA on all pages in the
1291          * file up to ra_max_pages.  This is simply a best effort and
1292          * only occurs once per open file.  Normal RA behavior is reverted
1293          * to for subsequent IO.  The mmap case does not increment
1294          * ras_requests and thus can never trigger this behavior. */
1295         if (ras->ras_requests == 2 && !ras->ras_request_index) {
1296                 __u64 kms_pages;
1297
1298                 kms_pages = (inode->i_size + CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT;
1299
1300                 CDEBUG(D_READA, "kmsp "LPU64" mwp %lu mp %lu\n", kms_pages,
1301                        ra->ra_max_read_ahead_whole_pages, ra->ra_max_pages);
1302
1303                 if (kms_pages &&
1304                     kms_pages <= ra->ra_max_read_ahead_whole_pages) {
1305                         ras->ras_window_start = 0;
1306                         ras->ras_last_readpage = 0;
1307                         ras->ras_next_readahead = 0;
1308                         ras->ras_window_len = min(ra->ra_max_pages,
1309                                 ra->ra_max_read_ahead_whole_pages);
1310                         GOTO(out_unlock, 0);
1311                 }
1312         }
1313
1314         if (zero) {
1315                 ras_reset(ras, index);
1316                 GOTO(out_unlock, 0);
1317         }
1318
1319         ras->ras_last_readpage = index;
1320         ras->ras_consecutive_pages++;
1321         ras_set_start(ras, index);
1322         ras->ras_next_readahead = max(ras->ras_window_start,
1323                                       ras->ras_next_readahead);
1324
1325         /* Trigger RA in the mmap case where ras_consecutive_requests
1326          * is not incremented and thus can't be used to trigger RA */
1327         if (!ras->ras_window_len && ras->ras_consecutive_pages == 3) {
1328                 ras->ras_window_len = 1024 * 1024 >> CFS_PAGE_SHIFT;
1329                 GOTO(out_unlock, 0);
1330         }
1331
1332         /* The initial ras_window_len is set to the request size.  To avoid
1333          * uselessly reading and discarding pages for random IO the window is
1334          * only increased once per consecutive request received. */
1335         if (ras->ras_consecutive_requests > 1 && !ras->ras_request_index) {
1336                 ras->ras_window_len = min(ras->ras_window_len +
1337                                           (1024 * 1024 >> CFS_PAGE_SHIFT),
1338                                           ra->ra_max_pages);
1339         }
1340
1341         EXIT;
1342 out_unlock:
1343         RAS_CDEBUG(ras);
1344         ras->ras_request_index++;
1345         spin_unlock(&ras->ras_lock);
1346         spin_unlock(&sbi->ll_lock);
1347         return;
1348 }
1349
1350 int ll_writepage(struct page *page)
1351 {
1352         struct inode *inode = page->mapping->host;
1353         struct ll_inode_info *lli = ll_i2info(inode);
1354         struct obd_export *exp;
1355         struct ll_async_page *llap;
1356         int rc = 0;
1357         ENTRY;
1358
1359         LASSERT(!PageDirty(page));
1360         LASSERT(PageLocked(page));
1361
1362         exp = ll_i2obdexp(inode);
1363         if (exp == NULL)
1364                 GOTO(out, rc = -EINVAL);
1365
1366         llap = llap_from_page(page, LLAP_ORIGIN_WRITEPAGE);
1367         if (IS_ERR(llap))
1368                 GOTO(out, rc = PTR_ERR(llap));
1369
1370         LASSERT(!PageWriteback(page));
1371         set_page_writeback(page);
1372
1373         page_cache_get(page);
1374         if (llap->llap_write_queued) {
1375                 LL_CDEBUG_PAGE(D_PAGE, page, "marking urgent\n");
1376                 rc = obd_set_async_flags(exp, lli->lli_smd, NULL,
1377                                          llap->llap_cookie,
1378                                          ASYNC_READY | ASYNC_URGENT);
1379         } else {
1380                 rc = queue_or_sync_write(exp, inode, llap, CFS_PAGE_SIZE,
1381                                          ASYNC_READY | ASYNC_URGENT);
1382         }
1383         if (rc)
1384                 page_cache_release(page);
1385 out:
1386         if (rc) {
1387                 if (!lli->lli_async_rc)
1388                         lli->lli_async_rc = rc;
1389                 /* re-dirty page on error so it retries write */
1390                 if (PageWriteback(page)) {
1391                         end_page_writeback(page);
1392                 }
1393                 ll_redirty_page(page);
1394                 unlock_page(page);
1395         }
1396         RETURN(rc);
1397 }
1398
1399 /*
1400  * for now we do our readpage the same on both 2.4 and 2.5.  The kernel's
1401  * read-ahead assumes it is valid to issue readpage all the way up to
1402  * i_size, but our dlm locks make that not the case.  We disable the
1403  * kernel's read-ahead and do our own by walking ahead in the page cache
1404  * checking for dlm lock coverage.  the main difference between 2.4 and
1405  * 2.6 is how read-ahead gets batched and issued, but we're using our own,
1406  * so they look the same.
1407  */
1408 int ll_readpage(struct file *filp, struct page *page)
1409 {
1410         struct ll_file_data *fd = LUSTRE_FPRIVATE(filp);
1411         struct inode *inode = page->mapping->host;
1412         struct obd_export *exp;
1413         struct ll_async_page *llap;
1414         struct obd_io_group *oig = NULL;
1415         int rc;
1416         ENTRY;
1417
1418         LASSERT(PageLocked(page));
1419         LASSERT(!PageUptodate(page));
1420         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),offset=%Lu=%#Lx\n",
1421                inode->i_ino, inode->i_generation, inode,
1422                (((loff_t)page->index) << CFS_PAGE_SHIFT),
1423                (((loff_t)page->index) << CFS_PAGE_SHIFT));
1424         LASSERT(atomic_read(&filp->f_dentry->d_inode->i_count) > 0);
1425
1426         if (!ll_i2info(inode)->lli_smd) {
1427                 /* File with no objects - one big hole */
1428                 /* We use this just for remove_from_page_cache that is not
1429                  * exported, we'd make page back up to date. */
1430                 ll_truncate_complete_page(page);
1431                 clear_page(kmap(page));
1432                 kunmap(page);
1433                 SetPageUptodate(page);
1434                 unlock_page(page);
1435                 RETURN(0);
1436         }
1437
1438         rc = oig_init(&oig);
1439         if (rc < 0)
1440                 GOTO(out, rc);
1441
1442         exp = ll_i2obdexp(inode);
1443         if (exp == NULL)
1444                 GOTO(out, rc = -EINVAL);
1445
1446         llap = llap_from_page(page, LLAP_ORIGIN_READPAGE);
1447         if (IS_ERR(llap))
1448                 GOTO(out, rc = PTR_ERR(llap));
1449
1450         if (ll_i2sbi(inode)->ll_ra_info.ra_max_pages)
1451                 ras_update(ll_i2sbi(inode), inode, &fd->fd_ras, page->index,
1452                            llap->llap_defer_uptodate);
1453
1454         if (llap->llap_defer_uptodate) {
1455                 llap->llap_ra_used = 1;
1456                 rc = ll_readahead(&fd->fd_ras, exp, page->mapping, oig,
1457                                   fd->fd_flags);
1458                 if (rc > 0)
1459                         obd_trigger_group_io(exp, ll_i2info(inode)->lli_smd,
1460                                              NULL, oig);
1461                 LL_CDEBUG_PAGE(D_PAGE, page, "marking uptodate from defer\n");
1462                 SetPageUptodate(page);
1463                 unlock_page(page);
1464                 GOTO(out_oig, rc = 0);
1465         }
1466
1467         if (likely((fd->fd_flags & LL_FILE_IGNORE_LOCK) == 0)) {
1468                 rc = ll_page_matches(page, fd->fd_flags);
1469                 if (rc < 0) {
1470                         LL_CDEBUG_PAGE(D_ERROR, page, "lock match failed: rc %d\n", rc);
1471                         GOTO(out, rc);
1472                 }
1473
1474                 if (rc == 0) {
1475                         CWARN("ino %lu page %lu (%llu) not covered by "
1476                               "a lock (mmap?).  check debug logs.\n",
1477                               inode->i_ino, page->index,
1478                               (long long)page->index << CFS_PAGE_SHIFT);
1479                 }
1480         }
1481
1482         rc = ll_issue_page_read(exp, llap, oig, 0);
1483         if (rc)
1484                 GOTO(out, rc);
1485
1486         LL_CDEBUG_PAGE(D_PAGE, page, "queued readpage\n");
1487         if (ll_i2sbi(inode)->ll_ra_info.ra_max_pages)
1488                 ll_readahead(&fd->fd_ras, exp, page->mapping, oig,
1489                              fd->fd_flags);
1490
1491         rc = obd_trigger_group_io(exp, ll_i2info(inode)->lli_smd, NULL, oig);
1492
1493 out:
1494         if (rc)
1495                 unlock_page(page);
1496 out_oig:
1497         if (oig != NULL)
1498                 oig_release(oig);
1499         RETURN(rc);
1500 }