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 #ifndef AUTOCONF_INCLUDED
  25 #include <linux/config.h>
  26 #endif
  27 #include <linux/kernel.h>
  28 #include <linux/mm.h>
  29 #include <linux/string.h>
  30 #include <linux/stat.h>
  31 #include <linux/errno.h>
  32 #include <linux/smp_lock.h>
  33 #include <linux/unistd.h>
  34 #include <linux/version.h>
  35 #include <asm/system.h>
  36 #include <asm/uaccess.h>
  37
  38 #include <linux/fs.h>
  39 #include <linux/stat.h>
  40 #include <asm/uaccess.h>
  41 #include <asm/segment.h>
  42 #include <linux/mm.h>
  43 #include <linux/pagemap.h>
  44 #include <linux/smp_lock.h>
  45
  46 #define DEBUG_SUBSYSTEM S_LLITE
  47
  48 //#include <lustre_mdc.h>
  49 #include <lustre_lite.h>
  50 #include "llite_internal.h"
  51 #include <linux/lustre_compat25.h>
  52
  53 #ifndef list_for_each_prev_safe
  54 #define list_for_each_prev_safe(pos, n, head) \
  55         for (pos = (head)->prev, n = pos->prev; pos != (head); \
  56                 pos = n, n = pos->prev )
  57 #endif
  58
  59 cfs_mem_cache_t *ll_async_page_slab = NULL;
  60 size_t ll_async_page_slab_size = 0;
  61
  62 /* SYNCHRONOUS I/O to object storage for an inode */
  63 static int ll_brw(int cmd, struct inode *inode, struct obdo *oa,
  64                   struct page *page, int flags)
  65 {
  66         struct ll_inode_info *lli = ll_i2info(inode);
  67         struct lov_stripe_md *lsm = lli->lli_smd;
  68         struct obd_info oinfo = { { { 0 } } };
  69         struct brw_page pg;
  70         int opc, rc;
  71         ENTRY;
  72
  73         pg.pg = page;
  74         pg.off = ((obd_off)page->index) << CFS_PAGE_SHIFT;
  75
  76         if ((cmd & OBD_BRW_WRITE) && (pg.off+CFS_PAGE_SIZE>i_size_read(inode)))
  77                 pg.count = i_size_read(inode) % CFS_PAGE_SIZE;
  78         else
  79                 pg.count = CFS_PAGE_SIZE;
  80
  81         LL_CDEBUG_PAGE(D_PAGE, page, "%s %d bytes ino %lu at "LPU64"/"LPX64"\n",
  82                        cmd & OBD_BRW_WRITE ? "write" : "read", pg.count,
  83                        inode->i_ino, pg.off, pg.off);
  84         if (pg.count == 0) {
  85                 CERROR("ZERO COUNT: ino %lu: size %p:%Lu(%p:%Lu) idx %lu off "
  86                        LPU64"\n", inode->i_ino, inode, i_size_read(inode),
  87                        page->mapping->host, i_size_read(page->mapping->host),
  88                        page->index, pg.off);
  89         }
  90
  91         pg.flag = flags;
  92
  93         if (cmd & OBD_BRW_WRITE)
  94                 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_BRW_WRITE,
  95                                    pg.count);
  96         else
  97                 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_BRW_READ,
  98                                    pg.count);
  99         oinfo.oi_oa = oa;
 100         oinfo.oi_md = lsm;
 101         /* NB partial write, so we might not have CAPA_OPC_OSS_READ capa */
 102         opc = cmd & OBD_BRW_WRITE ? CAPA_OPC_OSS_WRITE : CAPA_OPC_OSS_RW;
 103         oinfo.oi_capa = ll_osscapa_get(inode, opc);
 104         rc = obd_brw(cmd, ll_i2dtexp(inode), &oinfo, 1, &pg, NULL);
 105         capa_put(oinfo.oi_capa);
 106         if (rc == 0)
 107                 obdo_to_inode(inode, oa, OBD_MD_FLBLOCKS);
 108         else if (rc != -EIO)
 109                 CERROR("error from obd_brw: rc = %d\n", rc);
 110         RETURN(rc);
 111 }
 112
 113 /* this isn't where truncate starts.   roughly:
 114  * sys_truncate->ll_setattr_raw->vmtruncate->ll_truncate. setattr_raw grabs
 115  * DLM lock on [size, EOF], i_mutex, ->lli_size_sem, and WRITE_I_ALLOC_SEM to
 116  * avoid races.
 117  *
 118  * must be called under ->lli_size_sem */
 119 void ll_truncate(struct inode *inode)
 120 {
 121         struct ll_inode_info *lli = ll_i2info(inode);
 122         struct obd_info oinfo = { { { 0 } } };
 123         struct ost_lvb lvb;
 124         struct obdo oa;
 125         int rc;
 126         ENTRY;
 127         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p) to %Lu=%#Lx\n",inode->i_ino,
 128                inode->i_generation, inode, i_size_read(inode),
 129                i_size_read(inode));
 130
 131         ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_TRUNC, 1);
 132         if (lli->lli_size_sem_owner != current) {
 133                 EXIT;
 134                 return;
 135         }
 136
 137         if (!lli->lli_smd) {
 138                 CDEBUG(D_INODE, "truncate on inode %lu with no objects\n",
 139                        inode->i_ino);
 140                 GOTO(out_unlock, 0);
 141         }
 142
 143         LASSERT(atomic_read(&lli->lli_size_sem.count) <= 0);
 144
 145         /* XXX I'm pretty sure this is a hack to paper over a more fundamental
 146          * race condition. */
 147         lov_stripe_lock(lli->lli_smd);
 148         inode_init_lvb(inode, &lvb);
 149         obd_merge_lvb(ll_i2dtexp(inode), lli->lli_smd, &lvb, 0);
 150         if (lvb.lvb_size == i_size_read(inode)) {
 151                 CDEBUG(D_VFSTRACE, "skipping punch for obj "LPX64", %Lu=%#Lx\n",
 152                        lli->lli_smd->lsm_object_id, i_size_read(inode),
 153                        i_size_read(inode));
 154                 lov_stripe_unlock(lli->lli_smd);
 155                 GOTO(out_unlock, 0);
 156         }
 157
 158         obd_adjust_kms(ll_i2dtexp(inode), lli->lli_smd, i_size_read(inode), 1);
 159         lov_stripe_unlock(lli->lli_smd);
 160
 161         if (unlikely((ll_i2sbi(inode)->ll_flags & LL_SBI_CHECKSUM) &&
 162                      (i_size_read(inode) & ~CFS_PAGE_MASK))) {
 163                 /* If the truncate leaves behind a partial page, update its
 164                  * checksum. */
 165                 struct page *page = find_get_page(inode->i_mapping,
 166                                                   i_size_read(inode) >>
 167                                                   CFS_PAGE_SHIFT);
 168                 if (page != NULL) {
 169                         struct ll_async_page *llap = llap_cast_private(page);
 170                         if (llap != NULL) {
 171                                 llap->llap_checksum =
 172                                         crc32_le(0, kmap(page), CFS_PAGE_SIZE);
 173                                 kunmap(page);
 174                         }
 175                         page_cache_release(page);
 176                 }
 177         }
 178
 179         CDEBUG(D_INFO, "calling punch for "LPX64" (new size %Lu=%#Lx)\n",
 180                lli->lli_smd->lsm_object_id, i_size_read(inode), i_size_read(inode));
 181
 182         oinfo.oi_md = lli->lli_smd;
 183         oinfo.oi_policy.l_extent.start = i_size_read(inode);
 184         oinfo.oi_policy.l_extent.end = OBD_OBJECT_EOF;
 185         oinfo.oi_oa = &oa;
 186         oa.o_id = lli->lli_smd->lsm_object_id;
 187         oa.o_gr = lli->lli_smd->lsm_object_gr;
 188         oa.o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
 189
 190         obdo_from_inode(&oa, inode, OBD_MD_FLTYPE | OBD_MD_FLMODE |
 191                         OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME |
 192                         OBD_MD_FLFID | OBD_MD_FLGENER);
 193
 194         ll_inode_size_unlock(inode, 0);
 195
 196         oinfo.oi_capa = ll_osscapa_get(inode, CAPA_OPC_OSS_TRUNC);
 197         rc = obd_punch_rqset(ll_i2dtexp(inode), &oinfo, NULL);
 198         ll_truncate_free_capa(oinfo.oi_capa);
 199         if (rc)
 200                 CERROR("obd_truncate fails (%d) ino %lu\n", rc, inode->i_ino);
 201         else
 202                 obdo_to_inode(inode, &oa, OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
 203                               OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME);
 204         EXIT;
 205         return;
 206
 207  out_unlock:
 208         ll_inode_size_unlock(inode, 0);
 209 } /* ll_truncate */
 210
 211 int ll_prepare_write(struct file *file, struct page *page, unsigned from,
 212                      unsigned to)
 213 {
 214         struct inode *inode = page->mapping->host;
 215         struct ll_inode_info *lli = ll_i2info(inode);
 216         struct lov_stripe_md *lsm = lli->lli_smd;
 217         obd_off offset = ((obd_off)page->index) << CFS_PAGE_SHIFT;
 218         struct obd_info oinfo = { { { 0 } } };
 219         struct brw_page pga;
 220         struct obdo oa;
 221         struct ost_lvb lvb;
 222         int rc = 0;
 223         ENTRY;
 224
 225         LASSERT(PageLocked(page));
 226         (void)llap_cast_private(page); /* assertion */
 227
 228         /* Check to see if we should return -EIO right away */
 229         pga.pg = page;
 230         pga.off = offset;
 231         pga.count = CFS_PAGE_SIZE;
 232         pga.flag = 0;
 233
 234         oa.o_mode = inode->i_mode;
 235         oa.o_id = lsm->lsm_object_id;
 236         oa.o_gr = lsm->lsm_object_gr;
 237         oa.o_valid = OBD_MD_FLID | OBD_MD_FLMODE |
 238                      OBD_MD_FLTYPE | OBD_MD_FLGROUP;
 239         obdo_from_inode(&oa, inode, OBD_MD_FLFID | OBD_MD_FLGENER);
 240
 241         oinfo.oi_oa = &oa;
 242         oinfo.oi_md = lsm;
 243         rc = obd_brw(OBD_BRW_CHECK, ll_i2dtexp(inode), &oinfo, 1, &pga, NULL);
 244         if (rc)
 245                 RETURN(rc);
 246
 247         if (PageUptodate(page)) {
 248                 LL_CDEBUG_PAGE(D_PAGE, page, "uptodate\n");
 249                 RETURN(0);
 250         }
 251
 252         /* We're completely overwriting an existing page, so _don't_ set it up
 253          * to date until commit_write */
 254         if (from == 0 && to == CFS_PAGE_SIZE) {
 255                 LL_CDEBUG_PAGE(D_PAGE, page, "full page write\n");
 256                 POISON_PAGE(page, 0x11);
 257                 RETURN(0);
 258         }
 259
 260         /* If are writing to a new page, no need to read old data.  The extent
 261          * locking will have updated the KMS, and for our purposes here we can
 262          * treat it like i_size. */
 263         lov_stripe_lock(lsm);
 264         inode_init_lvb(inode, &lvb);
 265         obd_merge_lvb(ll_i2dtexp(inode), lsm, &lvb, 1);
 266         lov_stripe_unlock(lsm);
 267         if (lvb.lvb_size <= offset) {
 268                 LL_CDEBUG_PAGE(D_PAGE, page, "kms "LPU64" <= offset "LPU64"\n",
 269                                lvb.lvb_size, offset);
 270                 memset(kmap(page), 0, CFS_PAGE_SIZE);
 271                 kunmap(page);
 272                 GOTO(prepare_done, rc = 0);
 273         }
 274
 275         /* XXX could be an async ocp read.. read-ahead? */
 276         rc = ll_brw(OBD_BRW_READ, inode, &oa, page, 0);
 277         if (rc == 0) {
 278                 /* bug 1598: don't clobber blksize */
 279                 oa.o_valid &= ~(OBD_MD_FLSIZE | OBD_MD_FLBLKSZ);
 280                 obdo_refresh_inode(inode, &oa, oa.o_valid);
 281         }
 282
 283         EXIT;
 284  prepare_done:
 285         if (rc == 0)
 286                 SetPageUptodate(page);
 287
 288         return rc;
 289 }
 290
 291 static int ll_ap_make_ready(void *data, int cmd)
 292 {
 293         struct ll_async_page *llap;
 294         struct page *page;
 295         ENTRY;
 296
 297         llap = LLAP_FROM_COOKIE(data);
 298         page = llap->llap_page;
 299
 300         LASSERTF(!(cmd & OBD_BRW_READ), "cmd %x page %p ino %lu index %lu\n", cmd, page,
 301                  page->mapping->host->i_ino, page->index);
 302
 303         /* we're trying to write, but the page is locked.. come back later */
 304         if (TryLockPage(page))
 305                 RETURN(-EAGAIN);
 306
 307         LASSERT(!PageWriteback(page));
 308
 309         /* if we left PageDirty we might get another writepage call
 310          * in the future.  list walkers are bright enough
 311          * to check page dirty so we can leave it on whatever list
 312          * its on.  XXX also, we're called with the cli list so if
 313          * we got the page cache list we'd create a lock inversion
 314          * with the removepage path which gets the page lock then the
 315          * cli lock */
 316         LASSERTF(!PageWriteback(page),"cmd %x page %p ino %lu index %lu\n", cmd, page,
 317                  page->mapping->host->i_ino, page->index);
 318         clear_page_dirty_for_io(page);
 319
 320         /* This actually clears the dirty bit in the radix tree.*/
 321         set_page_writeback(page);
 322
 323         LL_CDEBUG_PAGE(D_PAGE, page, "made ready\n");
 324         page_cache_get(page);
 325
 326         RETURN(0);
 327 }
 328
 329 /* We have two reasons for giving llite the opportunity to change the
 330  * write length of a given queued page as it builds the RPC containing
 331  * the page:
 332  *
 333  * 1) Further extending writes may have landed in the page cache
 334  *    since a partial write first queued this page requiring us
 335  *    to write more from the page cache.  (No further races are possible, since
 336  *    by the time this is called, the page is locked.)
 337  * 2) We might have raced with truncate and want to avoid performing
 338  *    write RPCs that are just going to be thrown away by the
 339  *    truncate's punch on the storage targets.
 340  *
 341  * The kms serves these purposes as it is set at both truncate and extending
 342  * writes.
 343  */
 344 static int ll_ap_refresh_count(void *data, int cmd)
 345 {
 346         struct ll_inode_info *lli;
 347         struct ll_async_page *llap;
 348         struct lov_stripe_md *lsm;
 349         struct page *page;
 350         struct inode *inode;
 351         struct ost_lvb lvb;
 352         __u64 kms;
 353         ENTRY;
 354
 355         /* readpage queues with _COUNT_STABLE, shouldn't get here. */
 356         LASSERT(cmd != OBD_BRW_READ);
 357
 358         llap = LLAP_FROM_COOKIE(data);
 359         page = llap->llap_page;
 360         inode = page->mapping->host;
 361         lli = ll_i2info(inode);
 362         lsm = lli->lli_smd;
 363
 364         lov_stripe_lock(lsm);
 365         inode_init_lvb(inode, &lvb);
 366         obd_merge_lvb(ll_i2dtexp(inode), lsm, &lvb, 1);
 367         kms = lvb.lvb_size;
 368         lov_stripe_unlock(lsm);
 369
 370         /* catch race with truncate */
 371         if (((__u64)page->index << CFS_PAGE_SHIFT) >= kms)
 372                 return 0;
 373
 374         /* catch sub-page write at end of file */
 375         if (((__u64)page->index << CFS_PAGE_SHIFT) + CFS_PAGE_SIZE > kms)
 376                 return kms % CFS_PAGE_SIZE;
 377
 378         return CFS_PAGE_SIZE;
 379 }
 380
 381 void ll_inode_fill_obdo(struct inode *inode, int cmd, struct obdo *oa)
 382 {
 383         struct lov_stripe_md *lsm;
 384         obd_flag valid_flags;
 385
 386         lsm = ll_i2info(inode)->lli_smd;
 387
 388         oa->o_id = lsm->lsm_object_id;
 389         oa->o_gr = lsm->lsm_object_gr;
 390         oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
 391         valid_flags = OBD_MD_FLTYPE | OBD_MD_FLATIME;
 392         if (cmd & OBD_BRW_WRITE) {
 393                 oa->o_valid |= OBD_MD_FLEPOCH;
 394                 oa->o_easize = ll_i2info(inode)->lli_ioepoch;
 395
 396                 valid_flags |= OBD_MD_FLMTIME | OBD_MD_FLCTIME |
 397                         OBD_MD_FLUID | OBD_MD_FLGID |
 398                         OBD_MD_FLFID | OBD_MD_FLGENER;
 399         }
 400
 401         obdo_from_inode(oa, inode, valid_flags);
 402 }
 403
 404 static void ll_ap_fill_obdo(void *data, int cmd, struct obdo *oa)
 405 {
 406         struct ll_async_page *llap;
 407         ENTRY;
 408
 409         llap = LLAP_FROM_COOKIE(data);
 410         ll_inode_fill_obdo(llap->llap_page->mapping->host, cmd, oa);
 411
 412         EXIT;
 413 }
 414
 415 static void ll_ap_update_obdo(void *data, int cmd, struct obdo *oa,
 416                               obd_valid valid)
 417 {
 418         struct ll_async_page *llap;
 419         ENTRY;
 420
 421         llap = LLAP_FROM_COOKIE(data);
 422         obdo_from_inode(oa, llap->llap_page->mapping->host, valid);
 423
 424         EXIT;
 425 }
 426
 427 static struct obd_capa *ll_ap_lookup_capa(void *data, int cmd)
 428 {
 429         struct ll_async_page *llap = LLAP_FROM_COOKIE(data);
 430         int opc = cmd & OBD_BRW_WRITE ? CAPA_OPC_OSS_WRITE : CAPA_OPC_OSS_RW;
 431
 432         return ll_osscapa_get(llap->llap_page->mapping->host, opc);
 433 }
 434
 435 static struct obd_async_page_ops ll_async_page_ops = {
 436         .ap_make_ready =        ll_ap_make_ready,
 437         .ap_refresh_count =     ll_ap_refresh_count,
 438         .ap_fill_obdo =         ll_ap_fill_obdo,
 439         .ap_update_obdo =       ll_ap_update_obdo,
 440         .ap_completion =        ll_ap_completion,
 441         .ap_lookup_capa =       ll_ap_lookup_capa,
 442 };
 443
 444 struct ll_async_page *llap_cast_private(struct page *page)
 445 {
 446         struct ll_async_page *llap = (struct ll_async_page *)page_private(page);
 447
 448         LASSERTF(llap == NULL || llap->llap_magic == LLAP_MAGIC,
 449                  "page %p private %lu gave magic %d which != %d\n",
 450                  page, page_private(page), llap->llap_magic, LLAP_MAGIC);
 451
 452         return llap;
 453 }
 454
 455 /* Try to shrink the page cache for the @sbi filesystem by 1/@shrink_fraction.
 456  *
 457  * There is an llap attached onto every page in lustre, linked off @sbi.
 458  * We add an llap to the list so we don't lose our place during list walking.
 459  * If llaps in the list are being moved they will only move to the end
 460  * of the LRU, and we aren't terribly interested in those pages here (we
 461  * start at the beginning of the list where the least-used llaps are.
 462  */
 463 int llap_shrink_cache(struct ll_sb_info *sbi, int shrink_fraction)
 464 {
 465         struct ll_async_page *llap, dummy_llap = { .llap_magic = 0xd11ad11a };
 466         unsigned long total, want, count = 0;
 467
 468         total = sbi->ll_async_page_count;
 469
 470         /* There can be a large number of llaps (600k or more in a large
 471          * memory machine) so the VM 1/6 shrink ratio is likely too much.
 472          * Since we are freeing pages also, we don't necessarily want to
 473          * shrink so much.  Limit to 40MB of pages + llaps per call. */
 474         if (shrink_fraction == 0)
 475                 want = sbi->ll_async_page_count - sbi->ll_async_page_max + 32;
 476         else
 477                 want = (total + shrink_fraction - 1) / shrink_fraction;
 478
 479         if (want > 40 << (20 - CFS_PAGE_SHIFT))
 480                 want = 40 << (20 - CFS_PAGE_SHIFT);
 481
 482         CDEBUG(D_CACHE, "shrinking %lu of %lu pages (1/%d)\n",
 483                want, total, shrink_fraction);
 484
 485         spin_lock(&sbi->ll_lock);
 486         list_add(&dummy_llap.llap_pglist_item, &sbi->ll_pglist);
 487
 488         while (--total >= 0 && count < want) {
 489                 struct page *page;
 490                 int keep;
 491
 492                 if (unlikely(need_resched())) {
 493                         spin_unlock(&sbi->ll_lock);
 494                         cond_resched();
 495                         spin_lock(&sbi->ll_lock);
 496                 }
 497
 498                 llap = llite_pglist_next_llap(sbi,&dummy_llap.llap_pglist_item);
 499                 list_del_init(&dummy_llap.llap_pglist_item);
 500                 if (llap == NULL)
 501                         break;
 502
 503                 page = llap->llap_page;
 504                 LASSERT(page != NULL);
 505
 506                 list_add(&dummy_llap.llap_pglist_item, &llap->llap_pglist_item);
 507
 508                 /* Page needs/undergoing IO */
 509                 if (TryLockPage(page)) {
 510                         LL_CDEBUG_PAGE(D_PAGE, page, "can't lock\n");
 511                         continue;
 512                 }
 513
 514                keep = (llap->llap_write_queued || PageDirty(page) ||
 515                       PageWriteback(page) || (!PageUptodate(page) &&
 516                       llap->llap_origin != LLAP_ORIGIN_READAHEAD));
 517
 518                 LL_CDEBUG_PAGE(D_PAGE, page,"%s LRU page: %s%s%s%s%s origin %s\n",
 519                                keep ? "keep" : "drop",
 520                                llap->llap_write_queued ? "wq " : "",
 521                                PageDirty(page) ? "pd " : "",
 522                                PageUptodate(page) ? "" : "!pu ",
 523                                PageWriteback(page) ? "wb" : "",
 524                                llap->llap_defer_uptodate ? "" : "!du",
 525                                llap_origins[llap->llap_origin]);
 526
 527                 /* If page is dirty or undergoing IO don't discard it */
 528                 if (keep) {
 529                         unlock_page(page);
 530                         continue;
 531                 }
 532
 533                 page_cache_get(page);
 534                 spin_unlock(&sbi->ll_lock);
 535
 536                 if (page->mapping != NULL) {
 537                         ll_teardown_mmaps(page->mapping,
 538                                          (__u64)page->index << CFS_PAGE_SHIFT,
 539                                          ((__u64)page->index << CFS_PAGE_SHIFT)|
 540                                           ~CFS_PAGE_MASK);
 541                         if (!PageDirty(page) && !page_mapped(page)) {
 542                                 ll_ra_accounting(llap, page->mapping);
 543                                 ll_truncate_complete_page(page);
 544                                 ++count;
 545                         } else {
 546                                 LL_CDEBUG_PAGE(D_PAGE, page, "Not dropping page"
 547                                                              " because it is "
 548                                                              "%s\n",
 549                                                               PageDirty(page)?
 550                                                               "dirty":"mapped");
 551                         }
 552                 }
 553                 unlock_page(page);
 554                 page_cache_release(page);
 555
 556                 spin_lock(&sbi->ll_lock);
 557         }
 558         list_del(&dummy_llap.llap_pglist_item);
 559         spin_unlock(&sbi->ll_lock);
 560
 561         CDEBUG(D_CACHE, "shrank %lu/%lu and left %lu unscanned\n",
 562                count, want, total);
 563
 564         return count;
 565 }
 566
 567 struct ll_async_page *llap_from_page(struct page *page, unsigned origin)
 568 {
 569         struct ll_async_page *llap;
 570         struct obd_export *exp;
 571         struct inode *inode = page->mapping->host;
 572         struct ll_sb_info *sbi;
 573         int rc;
 574         ENTRY;
 575
 576         if (!inode) {
 577                 static int triggered;
 578
 579                 if (!triggered) {
 580                         LL_CDEBUG_PAGE(D_ERROR, page, "Bug 10047. Wrong anon "
 581                                        "page received\n");
 582                         libcfs_debug_dumpstack(NULL);
 583                         triggered = 1;
 584                 }
 585                 RETURN(ERR_PTR(-EINVAL));
 586         }
 587         sbi = ll_i2sbi(inode);
 588         LASSERT(ll_async_page_slab);
 589         LASSERTF(origin < LLAP__ORIGIN_MAX, "%u\n", origin);
 590
 591         llap = llap_cast_private(page);
 592         if (llap != NULL) {
 593                 /* move to end of LRU list, except when page is just about to
 594                  * die */
 595                 if (origin != LLAP_ORIGIN_REMOVEPAGE) {
 596                         spin_lock(&sbi->ll_lock);
 597                         sbi->ll_pglist_gen++;
 598                         list_del_init(&llap->llap_pglist_item);
 599                         list_add_tail(&llap->llap_pglist_item, &sbi->ll_pglist);
 600                         spin_unlock(&sbi->ll_lock);
 601                 }
 602                 GOTO(out, llap);
 603         }
 604
 605         exp = ll_i2dtexp(page->mapping->host);
 606         if (exp == NULL)
 607                 RETURN(ERR_PTR(-EINVAL));
 608
 609         /* limit the number of lustre-cached pages */
 610         if (sbi->ll_async_page_count >= sbi->ll_async_page_max)
 611                 llap_shrink_cache(sbi, 0);
 612
 613         OBD_SLAB_ALLOC(llap, ll_async_page_slab, CFS_ALLOC_STD,
 614                        ll_async_page_slab_size);
 615         if (llap == NULL)
 616                 RETURN(ERR_PTR(-ENOMEM));
 617         llap->llap_magic = LLAP_MAGIC;
 618         llap->llap_cookie = (void *)llap + size_round(sizeof(*llap));
 619
 620         rc = obd_prep_async_page(exp, ll_i2info(inode)->lli_smd, NULL, page,
 621                                  (obd_off)page->index << CFS_PAGE_SHIFT,
 622                                  &ll_async_page_ops, llap, &llap->llap_cookie);
 623         if (rc) {
 624                 OBD_SLAB_FREE(llap, ll_async_page_slab,
 625                               ll_async_page_slab_size);
 626                 RETURN(ERR_PTR(rc));
 627         }
 628
 629         CDEBUG(D_CACHE, "llap %p page %p cookie %p obj off "LPU64"\n", llap,
 630                page, llap->llap_cookie, (obd_off)page->index << CFS_PAGE_SHIFT);
 631         /* also zeroing the PRIVBITS low order bitflags */
 632         __set_page_ll_data(page, llap);
 633         llap->llap_page = page;
 634         spin_lock(&sbi->ll_lock);
 635         sbi->ll_pglist_gen++;
 636         sbi->ll_async_page_count++;
 637         list_add_tail(&llap->llap_pglist_item, &sbi->ll_pglist);
 638         INIT_LIST_HEAD(&llap->llap_pending_write);
 639         spin_unlock(&sbi->ll_lock);
 640
 641  out:
 642         if (unlikely(sbi->ll_flags & LL_SBI_CHECKSUM)) {
 643                 __u32 csum = 0;
 644                 csum = crc32_le(csum, kmap(page), CFS_PAGE_SIZE);
 645                 kunmap(page);
 646                 if (origin == LLAP_ORIGIN_READAHEAD ||
 647                     origin == LLAP_ORIGIN_READPAGE) {
 648                         llap->llap_checksum = 0;
 649                 } else if (origin == LLAP_ORIGIN_COMMIT_WRITE ||
 650                            llap->llap_checksum == 0) {
 651                         llap->llap_checksum = csum;
 652                         CDEBUG(D_PAGE, "page %p cksum %x\n", page, csum);
 653                 } else if (llap->llap_checksum == csum) {
 654                         /* origin == LLAP_ORIGIN_WRITEPAGE */
 655                         CDEBUG(D_PAGE, "page %p cksum %x confirmed\n",
 656                                page, csum);
 657                 } else {
 658                         /* origin == LLAP_ORIGIN_WRITEPAGE */
 659                         LL_CDEBUG_PAGE(D_ERROR, page, "old cksum %x != new "
 660                                        "%x!\n", llap->llap_checksum, csum);
 661                 }
 662         }
 663
 664         llap->llap_origin = origin;
 665         RETURN(llap);
 666 }
 667
 668 static int queue_or_sync_write(struct obd_export *exp, struct inode *inode,
 669                                struct ll_async_page *llap,
 670                                unsigned to, obd_flag async_flags)
 671 {
 672         unsigned long size_index = i_size_read(inode) >> CFS_PAGE_SHIFT;
 673         struct obd_io_group *oig;
 674         struct ll_sb_info *sbi = ll_i2sbi(inode);
 675         int rc, noquot = llap->llap_ignore_quota ? OBD_BRW_NOQUOTA : 0;
 676         ENTRY;
 677
 678         /* _make_ready only sees llap once we've unlocked the page */
 679         llap->llap_write_queued = 1;
 680         rc = obd_queue_async_io(exp, ll_i2info(inode)->lli_smd, NULL,
 681                                 llap->llap_cookie, OBD_BRW_WRITE | noquot,
 682                                 0, 0, 0, async_flags);
 683         if (rc == 0) {
 684                 LL_CDEBUG_PAGE(D_PAGE, llap->llap_page, "write queued\n");
 685                 GOTO(out, 0);
 686         }
 687
 688         llap->llap_write_queued = 0;
 689         /* Do not pass llap here as it is sync write. */
 690         llap_write_pending(inode, NULL);
 691
 692         rc = oig_init(&oig);
 693         if (rc)
 694                 GOTO(out, rc);
 695
 696         /* make full-page requests if we are not at EOF (bug 4410) */
 697         if (to != CFS_PAGE_SIZE && llap->llap_page->index < size_index) {
 698                 LL_CDEBUG_PAGE(D_PAGE, llap->llap_page,
 699                                "sync write before EOF: size_index %lu, to %d\n",
 700                                size_index, to);
 701                 to = CFS_PAGE_SIZE;
 702         } else if (to != CFS_PAGE_SIZE && llap->llap_page->index == size_index) {
 703                 int size_to = i_size_read(inode) & ~CFS_PAGE_MASK;
 704                 LL_CDEBUG_PAGE(D_PAGE, llap->llap_page,
 705                                "sync write at EOF: size_index %lu, to %d/%d\n",
 706                                size_index, to, size_to);
 707                 if (to < size_to)
 708                         to = size_to;
 709         }
 710
 711         /* compare the checksum once before the page leaves llite */
 712         if (unlikely((sbi->ll_flags & LL_SBI_CHECKSUM) &&
 713                      llap->llap_checksum != 0)) {
 714                 __u32 csum = 0;
 715                 struct page *page = llap->llap_page;
 716                 csum = crc32_le(csum, kmap(page), CFS_PAGE_SIZE);
 717                 kunmap(page);
 718                 if (llap->llap_checksum == csum) {
 719                         CDEBUG(D_PAGE, "page %p cksum %x confirmed\n",
 720                                page, csum);
 721                 } else {
 722                         CERROR("page %p old cksum %x != new cksum %x!\n",
 723                                page, llap->llap_checksum, csum);
 724                 }
 725         }
 726
 727         rc = obd_queue_group_io(exp, ll_i2info(inode)->lli_smd, NULL, oig,
 728                                 llap->llap_cookie, OBD_BRW_WRITE | noquot,
 729                                 0, to, 0, ASYNC_READY | ASYNC_URGENT |
 730                                 ASYNC_COUNT_STABLE | ASYNC_GROUP_SYNC);
 731         if (rc)
 732                 GOTO(free_oig, rc);
 733
 734         rc = obd_trigger_group_io(exp, ll_i2info(inode)->lli_smd, NULL, oig);
 735         if (rc)
 736                 GOTO(free_oig, rc);
 737
 738         rc = oig_wait(oig);
 739
 740         if (!rc && async_flags & ASYNC_READY) {
 741                 unlock_page(llap->llap_page);
 742                 if (PageWriteback(llap->llap_page)) {
 743                         end_page_writeback(llap->llap_page);
 744                 }
 745         }
 746
 747         if (rc == 0 && llap_write_complete(inode, llap))
 748                 ll_queue_done_writing(inode, 0);
 749
 750         LL_CDEBUG_PAGE(D_PAGE, llap->llap_page, "sync write returned %d\n", rc);
 751
 752 free_oig:
 753         oig_release(oig);
 754 out:
 755         RETURN(rc);
 756 }
 757
 758 /* update our write count to account for i_size increases that may have
 759  * happened since we've queued the page for io. */
 760
 761 /* be careful not to return success without setting the page Uptodate or
 762  * the next pass through prepare_write will read in stale data from disk. */
 763 int ll_commit_write(struct file *file, struct page *page, unsigned from,
 764                     unsigned to)
 765 {
 766         struct inode *inode = page->mapping->host;
 767         struct ll_inode_info *lli = ll_i2info(inode);
 768         struct lov_stripe_md *lsm = lli->lli_smd;
 769         struct obd_export *exp;
 770         struct ll_async_page *llap;
 771         loff_t size;
 772         int rc = 0;
 773         ENTRY;
 774
 775         SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
 776         LASSERT(inode == file->f_dentry->d_inode);
 777         LASSERT(PageLocked(page));
 778
 779         CDEBUG(D_INODE, "inode %p is writing page %p from %d to %d at %lu\n",
 780                inode, page, from, to, page->index);
 781
 782         llap = llap_from_page(page, LLAP_ORIGIN_COMMIT_WRITE);
 783         if (IS_ERR(llap))
 784                 RETURN(PTR_ERR(llap));
 785
 786         exp = ll_i2dtexp(inode);
 787         if (exp == NULL)
 788                 RETURN(-EINVAL);
 789
 790         llap->llap_ignore_quota = capable(CAP_SYS_RESOURCE);
 791
 792         /*
 793          * queue a write for some time in the future the first time we
 794          * dirty the page.
 795          *
 796          * This is different from what other file systems do: they usually
 797          * just mark page (and some of its buffers) dirty and rely on
 798          * balance_dirty_pages() to start a write-back. Lustre wants write-back
 799          * to be started earlier for the following reasons:
 800          *
 801          *     (1) with a large number of clients we need to limit the amount
 802          *     of cached data on the clients a lot;
 803          *
 804          *     (2) large compute jobs generally want compute-only then io-only
 805          *     and the IO should complete as quickly as possible;
 806          *
 807          *     (3) IO is batched up to the RPC size and is async until the
 808          *     client max cache is hit
 809          *     (/proc/fs/lustre/osc/OSC.../max_dirty_mb)
 810          *
 811          */
 812         if (!PageDirty(page)) {
 813                 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_DIRTY_MISSES, 1);
 814
 815                 rc = queue_or_sync_write(exp, inode, llap, to, 0);
 816                 if (rc)
 817                         GOTO(out, rc);
 818         } else {
 819                 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_DIRTY_HITS, 1);
 820         }
 821
 822         /* put the page in the page cache, from now on ll_removepage is
 823          * responsible for cleaning up the llap.
 824          * only set page dirty when it's queued to be write out */
 825         if (llap->llap_write_queued)
 826                 set_page_dirty(page);
 827
 828 out:
 829         size = (((obd_off)page->index) << CFS_PAGE_SHIFT) + to;
 830         ll_inode_size_lock(inode, 0);
 831         if (rc == 0) {
 832                 lov_stripe_lock(lsm);
 833                 obd_adjust_kms(exp, lsm, size, 0);
 834                 lov_stripe_unlock(lsm);
 835                 if (size > i_size_read(inode))
 836                         i_size_write(inode, size);
 837                 SetPageUptodate(page);
 838         } else if (size > i_size_read(inode)) {
 839                 /* this page beyond the pales of i_size, so it can't be
 840                  * truncated in ll_p_r_e during lock revoking. we must
 841                  * teardown our book-keeping here. */
 842                 ll_removepage(page);
 843         }
 844         ll_inode_size_unlock(inode, 0);
 845         RETURN(rc);
 846 }
 847
 848 static unsigned long ll_ra_count_get(struct ll_sb_info *sbi, unsigned long len)
 849 {
 850         struct ll_ra_info *ra = &sbi->ll_ra_info;
 851         unsigned long ret;
 852         ENTRY;
 853
 854         spin_lock(&sbi->ll_lock);
 855         ret = min(ra->ra_max_pages - ra->ra_cur_pages, len);
 856         ra->ra_cur_pages += ret;
 857         spin_unlock(&sbi->ll_lock);
 858
 859         RETURN(ret);
 860 }
 861
 862 static void ll_ra_count_put(struct ll_sb_info *sbi, unsigned long len)
 863 {
 864         struct ll_ra_info *ra = &sbi->ll_ra_info;
 865         spin_lock(&sbi->ll_lock);
 866         LASSERTF(ra->ra_cur_pages >= len, "r_c_p %lu len %lu\n",
 867                  ra->ra_cur_pages, len);
 868         ra->ra_cur_pages -= len;
 869         spin_unlock(&sbi->ll_lock);
 870 }
 871
 872 /* called for each page in a completed rpc.*/
 873 int ll_ap_completion(void *data, int cmd, struct obdo *oa, int rc)
 874 {
 875         struct ll_async_page *llap;
 876         struct page *page;
 877         int ret = 0;
 878         ENTRY;
 879
 880         llap = LLAP_FROM_COOKIE(data);
 881         page = llap->llap_page;
 882         LASSERT(PageLocked(page));
 883         LASSERT(CheckWriteback(page,cmd));
 884
 885         LL_CDEBUG_PAGE(D_PAGE, page, "completing cmd %d with %d\n", cmd, rc);
 886
 887         if (cmd & OBD_BRW_READ && llap->llap_defer_uptodate)
 888                 ll_ra_count_put(ll_i2sbi(page->mapping->host), 1);
 889
 890         if (rc == 0)  {
 891                 if (cmd & OBD_BRW_READ) {
 892                         if (!llap->llap_defer_uptodate)
 893                                 SetPageUptodate(page);
 894                 } else {
 895                         llap->llap_write_queued = 0;
 896                 }
 897                 ClearPageError(page);
 898         } else {
 899                 if (cmd & OBD_BRW_READ) {
 900                         llap->llap_defer_uptodate = 0;
 901                 }
 902                 SetPageError(page);
 903                 if (rc == -ENOSPC)
 904                         set_bit(AS_ENOSPC, &page->mapping->flags);
 905                 else
 906                         set_bit(AS_EIO, &page->mapping->flags);
 907         }
 908
 909         unlock_page(page);
 910
 911         if (cmd & OBD_BRW_WRITE) {
 912                 /* Only rc == 0, write succeed, then this page could be deleted
 913                  * from the pending_writing list
 914                  */
 915                 if (rc == 0 && llap_write_complete(page->mapping->host, llap))
 916                         ll_queue_done_writing(page->mapping->host, 0);
 917         }
 918
 919         if (PageWriteback(page)) {
 920                 end_page_writeback(page);
 921         }
 922         page_cache_release(page);
 923
 924         RETURN(ret);
 925 }
 926
 927 /* the kernel calls us here when a page is unhashed from the page cache.
 928  * the page will be locked and the kernel is holding a spinlock, so
 929  * we need to be careful.  we're just tearing down our book-keeping
 930  * here. */
 931 void ll_removepage(struct page *page)
 932 {
 933         struct inode *inode = page->mapping->host;
 934         struct obd_export *exp;
 935         struct ll_async_page *llap;
 936         struct ll_sb_info *sbi = ll_i2sbi(inode);
 937         int rc;
 938         ENTRY;
 939
 940         LASSERT(!in_interrupt());
 941
 942         /* sync pages or failed read pages can leave pages in the page
 943          * cache that don't have our data associated with them anymore */
 944         if (page_private(page) == 0) {
 945                 EXIT;
 946                 return;
 947         }
 948
 949         LL_CDEBUG_PAGE(D_PAGE, page, "being evicted\n");
 950
 951         exp = ll_i2dtexp(inode);
 952         if (exp == NULL) {
 953                 CERROR("page %p ind %lu gave null export\n", page, page->index);
 954                 EXIT;
 955                 return;
 956         }
 957
 958         llap = llap_from_page(page, LLAP_ORIGIN_REMOVEPAGE);
 959         if (IS_ERR(llap)) {
 960                 CERROR("page %p ind %lu couldn't find llap: %ld\n", page,
 961                        page->index, PTR_ERR(llap));
 962                 EXIT;
 963                 return;
 964         }
 965
 966         if (llap_write_complete(inode, llap))
 967                 ll_queue_done_writing(inode, 0);
 968
 969         rc = obd_teardown_async_page(exp, ll_i2info(inode)->lli_smd, NULL,
 970                                      llap->llap_cookie);
 971         if (rc != 0)
 972                 CERROR("page %p ind %lu failed: %d\n", page, page->index, rc);
 973
 974         /* this unconditional free is only safe because the page lock
 975          * is providing exclusivity to memory pressure/truncate/writeback..*/
 976         __clear_page_ll_data(page);
 977
 978         spin_lock(&sbi->ll_lock);
 979         if (!list_empty(&llap->llap_pglist_item))
 980                 list_del_init(&llap->llap_pglist_item);
 981         sbi->ll_pglist_gen++;
 982         sbi->ll_async_page_count--;
 983         spin_unlock(&sbi->ll_lock);
 984         OBD_SLAB_FREE(llap, ll_async_page_slab, ll_async_page_slab_size);
 985         EXIT;
 986 }
 987
 988 static int ll_page_matches(struct page *page, int fd_flags)
 989 {
 990         struct lustre_handle match_lockh = {0};
 991         struct inode *inode = page->mapping->host;
 992         ldlm_policy_data_t page_extent;
 993         int flags, matches;
 994         ENTRY;
 995
 996         if (unlikely(fd_flags & LL_FILE_GROUP_LOCKED))
 997                 RETURN(1);
 998
 999         page_extent.l_extent.start = (__u64)page->index << CFS_PAGE_SHIFT;
1000         page_extent.l_extent.end =
1001                 page_extent.l_extent.start + CFS_PAGE_SIZE - 1;
1002         flags = LDLM_FL_TEST_LOCK | LDLM_FL_BLOCK_GRANTED;
1003         if (!(fd_flags & LL_FILE_READAHEAD))
1004                 flags |= LDLM_FL_CBPENDING;
1005         matches = obd_match(ll_i2sbi(inode)->ll_dt_exp,
1006                             ll_i2info(inode)->lli_smd, LDLM_EXTENT,
1007                             &page_extent, LCK_PR | LCK_PW, &flags, inode,
1008                             &match_lockh);
1009         RETURN(matches);
1010 }
1011
1012 static int ll_issue_page_read(struct obd_export *exp,
1013                               struct ll_async_page *llap,
1014                               struct obd_io_group *oig, int defer)
1015 {
1016         struct page *page = llap->llap_page;
1017         int rc;
1018
1019         page_cache_get(page);
1020         llap->llap_defer_uptodate = defer;
1021         llap->llap_ra_used = 0;
1022         rc = obd_queue_group_io(exp, ll_i2info(page->mapping->host)->lli_smd,
1023                                 NULL, oig, llap->llap_cookie, OBD_BRW_READ, 0,
1024                                 CFS_PAGE_SIZE, 0, ASYNC_COUNT_STABLE |
1025                                                   ASYNC_READY | ASYNC_URGENT);
1026         if (rc) {
1027                 LL_CDEBUG_PAGE(D_ERROR, page, "read queue failed: rc %d\n", rc);
1028                 page_cache_release(page);
1029         }
1030         RETURN(rc);
1031 }
1032
1033 static void ll_ra_stats_inc_unlocked(struct ll_ra_info *ra, enum ra_stat which)
1034 {
1035         LASSERTF(which >= 0 && which < _NR_RA_STAT, "which: %u\n", which);
1036         ra->ra_stats[which]++;
1037 }
1038
1039 static void ll_ra_stats_inc(struct address_space *mapping, enum ra_stat which)
1040 {
1041         struct ll_sb_info *sbi = ll_i2sbi(mapping->host);
1042         struct ll_ra_info *ra = &ll_i2sbi(mapping->host)->ll_ra_info;
1043
1044         spin_lock(&sbi->ll_lock);
1045         ll_ra_stats_inc_unlocked(ra, which);
1046         spin_unlock(&sbi->ll_lock);
1047 }
1048
1049 void ll_ra_accounting(struct ll_async_page *llap, struct address_space *mapping)
1050 {
1051         if (!llap->llap_defer_uptodate || llap->llap_ra_used)
1052                 return;
1053
1054         ll_ra_stats_inc(mapping, RA_STAT_DISCARDED);
1055 }
1056
1057 #define RAS_CDEBUG(ras) \
1058         CDEBUG(D_READA,                                                      \
1059                "lrp %lu cr %lu cp %lu ws %lu wl %lu nra %lu r %lu ri %lu\n", \
1060                ras->ras_last_readpage, ras->ras_consecutive_requests,        \
1061                ras->ras_consecutive_pages, ras->ras_window_start,            \
1062                ras->ras_window_len, ras->ras_next_readahead,                 \
1063                ras->ras_requests, ras->ras_request_index);
1064
1065 static int index_in_window(unsigned long index, unsigned long point,
1066                            unsigned long before, unsigned long after)
1067 {
1068         unsigned long start = point - before, end = point + after;
1069
1070         if (start > point)
1071                start = 0;
1072         if (end < point)
1073                end = ~0;
1074
1075         return start <= index && index <= end;
1076 }
1077
1078 static struct ll_readahead_state *ll_ras_get(struct file *f)
1079 {
1080         struct ll_file_data       *fd;
1081
1082         fd = LUSTRE_FPRIVATE(f);
1083         return &fd->fd_ras;
1084 }
1085
1086 void ll_ra_read_in(struct file *f, struct ll_ra_read *rar)
1087 {
1088         struct ll_readahead_state *ras;
1089
1090         ras = ll_ras_get(f);
1091
1092         spin_lock(&ras->ras_lock);
1093         ras->ras_requests++;
1094         ras->ras_request_index = 0;
1095         ras->ras_consecutive_requests++;
1096         rar->lrr_reader = current;
1097
1098         list_add(&rar->lrr_linkage, &ras->ras_read_beads);
1099         spin_unlock(&ras->ras_lock);
1100 }
1101
1102 void ll_ra_read_ex(struct file *f, struct ll_ra_read *rar)
1103 {
1104         struct ll_readahead_state *ras;
1105
1106         ras = ll_ras_get(f);
1107
1108         spin_lock(&ras->ras_lock);
1109         list_del_init(&rar->lrr_linkage);
1110         spin_unlock(&ras->ras_lock);
1111 }
1112
1113 static struct ll_ra_read *ll_ra_read_get_locked(struct ll_readahead_state *ras)
1114 {
1115         struct ll_ra_read *scan;
1116
1117         list_for_each_entry(scan, &ras->ras_read_beads, lrr_linkage) {
1118                 if (scan->lrr_reader == current)
1119                         return scan;
1120         }
1121         return NULL;
1122 }
1123
1124 struct ll_ra_read *ll_ra_read_get(struct file *f)
1125 {
1126         struct ll_readahead_state *ras;
1127         struct ll_ra_read         *bead;
1128
1129         ras = ll_ras_get(f);
1130
1131         spin_lock(&ras->ras_lock);
1132         bead = ll_ra_read_get_locked(ras);
1133         spin_unlock(&ras->ras_lock);
1134         return bead;
1135 }
1136
1137 static int ll_readahead(struct ll_readahead_state *ras,
1138                          struct obd_export *exp, struct address_space *mapping,
1139                          struct obd_io_group *oig, int flags)
1140 {
1141         unsigned long i, start = 0, end = 0, reserved;
1142         struct ll_async_page *llap;
1143         struct page *page;
1144         int rc, ret = 0, match_failed = 0;
1145         __u64 kms;
1146         unsigned int gfp_mask;
1147         struct inode *inode;
1148         struct lov_stripe_md *lsm;
1149         struct ll_ra_read *bead;
1150         struct ost_lvb lvb;
1151         ENTRY;
1152
1153         inode = mapping->host;
1154         lsm = ll_i2info(inode)->lli_smd;
1155
1156         lov_stripe_lock(lsm);
1157         inode_init_lvb(inode, &lvb);
1158         obd_merge_lvb(ll_i2dtexp(inode), lsm, &lvb, 1);
1159         kms = lvb.lvb_size;
1160         lov_stripe_unlock(lsm);
1161         if (kms == 0) {
1162                 ll_ra_stats_inc(mapping, RA_STAT_ZERO_LEN);
1163                 RETURN(0);
1164         }
1165
1166         spin_lock(&ras->ras_lock);
1167         bead = ll_ra_read_get_locked(ras);
1168         /* Enlarge the RA window to encompass the full read */
1169         if (bead != NULL && ras->ras_window_start + ras->ras_window_len <
1170             bead->lrr_start + bead->lrr_count) {
1171                 ras->ras_window_len = bead->lrr_start + bead->lrr_count -
1172                                       ras->ras_window_start;
1173         }
1174         /* Reserve a part of the read-ahead window that we'll be issuing */
1175         if (ras->ras_window_len) {
1176                 start = ras->ras_next_readahead;
1177                 end = ras->ras_window_start + ras->ras_window_len - 1;
1178         }
1179         if (end != 0) {
1180                 /* Truncate RA window to end of file */
1181                 end = min(end, (unsigned long)((kms - 1) >> CFS_PAGE_SHIFT));
1182                 ras->ras_next_readahead = max(end, end + 1);
1183                 RAS_CDEBUG(ras);
1184         }
1185         spin_unlock(&ras->ras_lock);
1186
1187         if (end == 0) {
1188                 ll_ra_stats_inc(mapping, RA_STAT_ZERO_WINDOW);
1189                 RETURN(0);
1190         }
1191
1192         reserved = ll_ra_count_get(ll_i2sbi(inode), end - start + 1);
1193         if (reserved < end - start + 1)
1194                 ll_ra_stats_inc(mapping, RA_STAT_MAX_IN_FLIGHT);
1195
1196         gfp_mask = GFP_HIGHUSER & ~__GFP_WAIT;
1197 #ifdef __GFP_NOWARN
1198         gfp_mask |= __GFP_NOWARN;
1199 #endif
1200
1201         for (i = start; reserved > 0 && !match_failed && i <= end; i++) {
1202                 /* skip locked pages from previous readpage calls */
1203                 page = grab_cache_page_nowait_gfp(mapping, i, gfp_mask);
1204                 if (page == NULL) {
1205                         ll_ra_stats_inc(mapping, RA_STAT_FAILED_GRAB_PAGE);
1206                         CDEBUG(D_READA, "g_c_p_n failed\n");
1207                         continue;
1208                 }
1209
1210                 /* Check if page was truncated or reclaimed */
1211                 if (page->mapping != mapping) {
1212                         ll_ra_stats_inc(mapping, RA_STAT_WRONG_GRAB_PAGE);
1213                         CDEBUG(D_READA, "g_c_p_n returned invalid page\n");
1214                         goto next_page;
1215                 }
1216
1217                 /* we do this first so that we can see the page in the /proc
1218                  * accounting */
1219                 llap = llap_from_page(page, LLAP_ORIGIN_READAHEAD);
1220                 if (IS_ERR(llap) || llap->llap_defer_uptodate)
1221                         goto next_page;
1222
1223                 /* skip completed pages */
1224                 if (Page_Uptodate(page))
1225                         goto next_page;
1226
1227                 /* bail when we hit the end of the lock. */
1228                 if ((rc = ll_page_matches(page, flags|LL_FILE_READAHEAD)) <= 0){
1229                         LL_CDEBUG_PAGE(D_READA | D_PAGE, page,
1230                                        "lock match failed: rc %d\n", rc);
1231                         ll_ra_stats_inc(mapping, RA_STAT_FAILED_MATCH);
1232                         match_failed = 1;
1233                         goto next_page;
1234                 }
1235
1236                 rc = ll_issue_page_read(exp, llap, oig, 1);
1237                 if (rc == 0) {
1238                         reserved--;
1239                         ret++;
1240                         LL_CDEBUG_PAGE(D_READA| D_PAGE, page,
1241                                        "started read-ahead\n");
1242                 } else {
1243         next_page:
1244                         LL_CDEBUG_PAGE(D_READA | D_PAGE, page,
1245                                        "skipping read-ahead\n");
1246
1247                         unlock_page(page);
1248                 }
1249                 page_cache_release(page);
1250         }
1251
1252         LASSERTF(reserved >= 0, "reserved %lu\n", reserved);
1253         if (reserved != 0)
1254                 ll_ra_count_put(ll_i2sbi(inode), reserved);
1255         if (i == end + 1 && end == (kms >> CFS_PAGE_SHIFT))
1256                 ll_ra_stats_inc(mapping, RA_STAT_EOF);
1257
1258         /* if we didn't get to the end of the region we reserved from
1259          * the ras we need to go back and update the ras so that the
1260          * next read-ahead tries from where we left off.  we only do so
1261          * if the region we failed to issue read-ahead on is still ahead
1262          * of the app and behind the next index to start read-ahead from */
1263         if (i != end + 1) {
1264                 spin_lock(&ras->ras_lock);
1265                 if (i < ras->ras_next_readahead &&
1266                     index_in_window(i, ras->ras_window_start, 0,
1267                                     ras->ras_window_len)) {
1268                         ras->ras_next_readahead = i;
1269                         RAS_CDEBUG(ras);
1270                 }
1271                 spin_unlock(&ras->ras_lock);
1272         }
1273
1274         RETURN(ret);
1275 }
1276
1277 static void ras_set_start(struct ll_readahead_state *ras, unsigned long index)
1278 {
1279         ras->ras_window_start = index & (~((1024 * 1024 >> CFS_PAGE_SHIFT) - 1));
1280 }
1281
1282 /* called with the ras_lock held or from places where it doesn't matter */
1283 static void ras_reset(struct ll_readahead_state *ras, unsigned long index)
1284 {
1285         ras->ras_last_readpage = index;
1286         ras->ras_consecutive_requests = 0;
1287         ras->ras_consecutive_pages = 0;
1288         ras->ras_window_len = 0;
1289         ras_set_start(ras, index);
1290         ras->ras_next_readahead = max(ras->ras_window_start, index);
1291
1292         RAS_CDEBUG(ras);
1293 }
1294
1295 void ll_readahead_init(struct inode *inode, struct ll_readahead_state *ras)
1296 {
1297         spin_lock_init(&ras->ras_lock);
1298         ras_reset(ras, 0);
1299         ras->ras_requests = 0;
1300         INIT_LIST_HEAD(&ras->ras_read_beads);
1301 }
1302
1303 static void ras_update(struct ll_sb_info *sbi, struct inode *inode,
1304                        struct ll_readahead_state *ras, unsigned long index,
1305                        unsigned hit)
1306 {
1307         struct ll_ra_info *ra = &sbi->ll_ra_info;
1308         int zero = 0;
1309         ENTRY;
1310
1311         spin_lock(&sbi->ll_lock);
1312         spin_lock(&ras->ras_lock);
1313
1314         ll_ra_stats_inc_unlocked(ra, hit ? RA_STAT_HIT : RA_STAT_MISS);
1315
1316         /* reset the read-ahead window in two cases.  First when the app seeks
1317          * or reads to some other part of the file.  Secondly if we get a
1318          * read-ahead miss that we think we've previously issued.  This can
1319          * be a symptom of there being so many read-ahead pages that the VM is
1320          * reclaiming it before we get to it. */
1321         if (!index_in_window(index, ras->ras_last_readpage, 8, 8)) {
1322                 zero = 1;
1323                 ll_ra_stats_inc_unlocked(ra, RA_STAT_DISTANT_READPAGE);
1324         } else if (!hit && ras->ras_window_len &&
1325                    index < ras->ras_next_readahead &&
1326                    index_in_window(index, ras->ras_window_start, 0,
1327                                    ras->ras_window_len)) {
1328                 zero = 1;
1329                 ll_ra_stats_inc_unlocked(ra, RA_STAT_MISS_IN_WINDOW);
1330         }
1331
1332         /* On the second access to a file smaller than the tunable
1333          * ra_max_read_ahead_whole_pages trigger RA on all pages in the
1334          * file up to ra_max_pages.  This is simply a best effort and
1335          * only occurs once per open file.  Normal RA behavior is reverted
1336          * to for subsequent IO.  The mmap case does not increment
1337          * ras_requests and thus can never trigger this behavior. */
1338         if (ras->ras_requests == 2 && !ras->ras_request_index) {
1339                 __u64 kms_pages;
1340
1341                 kms_pages = (i_size_read(inode) + CFS_PAGE_SIZE - 1) >>
1342                             CFS_PAGE_SHIFT;
1343
1344                 CDEBUG(D_READA, "kmsp "LPU64" mwp %lu mp %lu\n", kms_pages,
1345                        ra->ra_max_read_ahead_whole_pages, ra->ra_max_pages);
1346
1347                 if (kms_pages &&
1348                     kms_pages <= ra->ra_max_read_ahead_whole_pages) {
1349                         ras->ras_window_start = 0;
1350                         ras->ras_last_readpage = 0;
1351                         ras->ras_next_readahead = 0;
1352                         ras->ras_window_len = min(ra->ra_max_pages,
1353                                 ra->ra_max_read_ahead_whole_pages);
1354                         GOTO(out_unlock, 0);
1355                 }
1356         }
1357
1358         if (zero) {
1359                 ras_reset(ras, index);
1360                 GOTO(out_unlock, 0);
1361         }
1362
1363         ras->ras_last_readpage = index;
1364         ras->ras_consecutive_pages++;
1365         ras_set_start(ras, index);
1366         ras->ras_next_readahead = max(ras->ras_window_start,
1367                                       ras->ras_next_readahead);
1368
1369         /* Trigger RA in the mmap case where ras_consecutive_requests
1370          * is not incremented and thus can't be used to trigger RA */
1371         if (!ras->ras_window_len && ras->ras_consecutive_pages == 3) {
1372                 ras->ras_window_len = 1024 * 1024 >> CFS_PAGE_SHIFT;
1373                 GOTO(out_unlock, 0);
1374         }
1375
1376         /* The initial ras_window_len is set to the request size.  To avoid
1377          * uselessly reading and discarding pages for random IO the window is
1378          * only increased once per consecutive request received. */
1379         if (ras->ras_consecutive_requests > 1 && !ras->ras_request_index) {
1380                 ras->ras_window_len = min(ras->ras_window_len +
1381                                           (1024 * 1024 >> CFS_PAGE_SHIFT),
1382                                           ra->ra_max_pages);
1383         }
1384
1385         EXIT;
1386 out_unlock:
1387         RAS_CDEBUG(ras);
1388         ras->ras_request_index++;
1389         spin_unlock(&ras->ras_lock);
1390         spin_unlock(&sbi->ll_lock);
1391         return;
1392 }
1393
1394 int ll_writepage(struct page *page)
1395 {
1396         struct inode *inode = page->mapping->host;
1397         struct ll_inode_info *lli = ll_i2info(inode);
1398         struct obd_export *exp;
1399         struct ll_async_page *llap;
1400         int rc = 0;
1401         ENTRY;
1402
1403         LASSERT(!PageDirty(page));
1404         LASSERT(PageLocked(page));
1405
1406         exp = ll_i2dtexp(inode);
1407         if (exp == NULL)
1408                 GOTO(out, rc = -EINVAL);
1409
1410         llap = llap_from_page(page, LLAP_ORIGIN_WRITEPAGE);
1411         if (IS_ERR(llap))
1412                 GOTO(out, rc = PTR_ERR(llap));
1413
1414         LASSERT(!PageWriteback(page));
1415         set_page_writeback(page);
1416
1417         page_cache_get(page);
1418         if (llap->llap_write_queued) {
1419                 LL_CDEBUG_PAGE(D_PAGE, page, "marking urgent\n");
1420                 rc = obd_set_async_flags(exp, lli->lli_smd, NULL,
1421                                          llap->llap_cookie,
1422                                          ASYNC_READY | ASYNC_URGENT);
1423         } else {
1424                 rc = queue_or_sync_write(exp, inode, llap, CFS_PAGE_SIZE,
1425                                          ASYNC_READY | ASYNC_URGENT);
1426         }
1427         if (rc)
1428                 page_cache_release(page);
1429 out:
1430         if (rc) {
1431                 if (!lli->lli_async_rc)
1432                         lli->lli_async_rc = rc;
1433                 /* re-dirty page on error so it retries write */
1434                 if (PageWriteback(page)) {
1435                         end_page_writeback(page);
1436                 }
1437                 /* resend page only for not started IO*/
1438                 if (!PageError(page))
1439                         ll_redirty_page(page);
1440                 unlock_page(page);
1441         }
1442         RETURN(rc);
1443 }
1444
1445 /*
1446  * for now we do our readpage the same on both 2.4 and 2.5.  The kernel's
1447  * read-ahead assumes it is valid to issue readpage all the way up to
1448  * i_size, but our dlm locks make that not the case.  We disable the
1449  * kernel's read-ahead and do our own by walking ahead in the page cache
1450  * checking for dlm lock coverage.  the main difference between 2.4 and
1451  * 2.6 is how read-ahead gets batched and issued, but we're using our own,
1452  * so they look the same.
1453  */
1454 int ll_readpage(struct file *filp, struct page *page)
1455 {
1456         struct ll_file_data *fd = LUSTRE_FPRIVATE(filp);
1457         struct inode *inode = page->mapping->host;
1458         struct obd_export *exp;
1459         struct ll_async_page *llap;
1460         struct obd_io_group *oig = NULL;
1461         int rc;
1462         ENTRY;
1463
1464         LASSERT(PageLocked(page));
1465         LASSERT(!PageUptodate(page));
1466         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),offset=%Lu=%#Lx\n",
1467                inode->i_ino, inode->i_generation, inode,
1468                (((loff_t)page->index) << CFS_PAGE_SHIFT),
1469                (((loff_t)page->index) << CFS_PAGE_SHIFT));
1470         LASSERT(atomic_read(&filp->f_dentry->d_inode->i_count) > 0);
1471
1472         if (!ll_i2info(inode)->lli_smd) {
1473                 /* File with no objects - one big hole */
1474                 /* We use this just for remove_from_page_cache that is not
1475                  * exported, we'd make page back up to date. */
1476                 ll_truncate_complete_page(page);
1477                 clear_page(kmap(page));
1478                 kunmap(page);
1479                 SetPageUptodate(page);
1480                 unlock_page(page);
1481                 RETURN(0);
1482         }
1483
1484         rc = oig_init(&oig);
1485         if (rc < 0)
1486                 GOTO(out, rc);
1487
1488         exp = ll_i2dtexp(inode);
1489         if (exp == NULL)
1490                 GOTO(out, rc = -EINVAL);
1491
1492         llap = llap_from_page(page, LLAP_ORIGIN_READPAGE);
1493         if (IS_ERR(llap))
1494                 GOTO(out, rc = PTR_ERR(llap));
1495
1496         if (ll_i2sbi(inode)->ll_ra_info.ra_max_pages)
1497                 ras_update(ll_i2sbi(inode), inode, &fd->fd_ras, page->index,
1498                            llap->llap_defer_uptodate);
1499
1500         if (llap->llap_defer_uptodate) {
1501                 llap->llap_ra_used = 1;
1502                 rc = ll_readahead(&fd->fd_ras, exp, page->mapping, oig,
1503                                   fd->fd_flags);
1504                 if (rc > 0)
1505                         obd_trigger_group_io(exp, ll_i2info(inode)->lli_smd,
1506                                              NULL, oig);
1507                 LL_CDEBUG_PAGE(D_PAGE, page, "marking uptodate from defer\n");
1508                 SetPageUptodate(page);
1509                 unlock_page(page);
1510                 GOTO(out_oig, rc = 0);
1511         }
1512
1513         if (likely((fd->fd_flags & LL_FILE_IGNORE_LOCK) == 0)) {
1514                 rc = ll_page_matches(page, fd->fd_flags);
1515                 if (rc < 0) {
1516                         LL_CDEBUG_PAGE(D_ERROR, page, "lock match failed: rc %d\n", rc);
1517                         GOTO(out, rc);
1518                 }
1519
1520                 if (rc == 0) {
1521                         CWARN("ino %lu page %lu (%llu) not covered by "
1522                               "a lock (mmap?).  check debug logs.\n",
1523                               inode->i_ino, page->index,
1524                               (long long)page->index << CFS_PAGE_SHIFT);
1525                 }
1526         }
1527
1528         rc = ll_issue_page_read(exp, llap, oig, 0);
1529         if (rc)
1530                 GOTO(out, rc);
1531
1532         LL_CDEBUG_PAGE(D_PAGE, page, "queued readpage\n");
1533         if (ll_i2sbi(inode)->ll_ra_info.ra_max_pages)
1534                 ll_readahead(&fd->fd_ras, exp, page->mapping, oig,
1535                              fd->fd_flags);
1536
1537         rc = obd_trigger_group_io(exp, ll_i2info(inode)->lli_smd, NULL, oig);
1538
1539 out:
1540         if (rc)
1541                 unlock_page(page);
1542 out_oig:
1543         if (oig != NULL)
1544                 oig_release(oig);
1545         RETURN(rc);
1546 }