lustre/llite/rw.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * Lustre Lite I/O page cache routines shared by different kernel revs
   5  *
   6  *  Copyright (c) 2001-2003 Cluster File Systems, Inc.
   7  *
   8  *   This file is part of Lustre, http://www.lustre.org.
   9  *
  10  *   Lustre is free software; you can redistribute it and/or
  11  *   modify it under the terms of version 2 of the GNU General Public
  12  *   License as published by the Free Software Foundation.
  13  *
  14  *   Lustre is distributed in the hope that it will be useful,
  15  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  17  *   GNU General Public License for more details.
  18  *
  19  *   You should have received a copy of the GNU General Public License
  20  *   along with Lustre; if not, write to the Free Software
  21  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  22  */
  23
  24 #include <linux/config.h>
  25 #include <linux/kernel.h>
  26 #include <linux/mm.h>
  27 #include <linux/string.h>
  28 #include <linux/stat.h>
  29 #include <linux/errno.h>
  30 #include <linux/smp_lock.h>
  31 #include <linux/unistd.h>
  32 #include <linux/version.h>
  33 #include <asm/system.h>
  34 #include <asm/uaccess.h>
  35
  36 #include <linux/fs.h>
  37 #include <linux/stat.h>
  38 #include <asm/uaccess.h>
  39 #include <asm/segment.h>
  40 #include <linux/mm.h>
  41 #include <linux/pagemap.h>
  42 #include <linux/smp_lock.h>
  43
  44 #define DEBUG_SUBSYSTEM S_LLITE
  45
  46 #include <linux/lustre_mds.h>
  47 #include <linux/lustre_lite.h>
  48 #include "llite_internal.h"
  49 #include <linux/lustre_compat25.h>
  50
  51 #ifndef list_for_each_prev_safe
  52 #define list_for_each_prev_safe(pos, n, head) \
  53         for (pos = (head)->prev, n = pos->prev; pos != (head); \
  54                 pos = n, n = pos->prev )
  55 #endif
  56
  57 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 brw_page pg;
  67         int rc;
  68         ENTRY;
  69
  70         pg.pg = page;
  71         pg.off = ((obd_off)page->index) << PAGE_SHIFT;
  72
  73         if (cmd == OBD_BRW_WRITE && (pg.off + PAGE_SIZE > inode->i_size))
  74                 pg.count = inode->i_size % PAGE_SIZE;
  75         else
  76                 pg.count = PAGE_SIZE;
  77
  78         LL_CDEBUG_PAGE(D_PAGE, page, "%s %d bytes ino %lu at "LPU64"/"LPX64"\n",
  79                        cmd & OBD_BRW_WRITE ? "write" : "read", pg.count,
  80                        inode->i_ino, pg.off, pg.off);
  81         if (pg.count == 0) {
  82                 CERROR("ZERO COUNT: ino %lu: size %p:%Lu(%p:%Lu) idx %lu off "
  83                        LPU64"\n",
  84                        inode->i_ino, inode, inode->i_size, page->mapping->host,
  85                        page->mapping->host->i_size, page->index, pg.off);
  86         }
  87
  88         pg.flag = flags;
  89
  90         if (cmd == OBD_BRW_WRITE)
  91                 lprocfs_counter_add(ll_i2sbi(inode)->ll_stats,
  92                                     LPROC_LL_BRW_WRITE, pg.count);
  93         else
  94                 lprocfs_counter_add(ll_i2sbi(inode)->ll_stats,
  95                                     LPROC_LL_BRW_READ, pg.count);
  96         rc = obd_brw(cmd, ll_i2obdexp(inode), oa, lsm, 1, &pg, NULL);
  97         if (rc == 0)
  98                 obdo_to_inode(inode, oa, OBD_MD_FLBLOCKS);
  99         else if (rc != -EIO)
 100                 CERROR("error from obd_brw: rc = %d\n", rc);
 101         RETURN(rc);
 102 }
 103
 104 __u64 lov_merge_size(struct lov_stripe_md *lsm, int kms);
 105
 106 /* this isn't where truncate starts.   roughly:
 107  * sys_truncate->ll_setattr_raw->vmtruncate->ll_truncate
 108  * we grab the lock back in setattr_raw to avoid races. */
 109 void ll_truncate(struct inode *inode)
 110 {
 111         struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
 112         struct obdo oa;
 113         int rc;
 114         ENTRY;
 115         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p) to %llu\n", inode->i_ino,
 116                inode->i_generation, inode, inode->i_size);
 117
 118         if (!lsm) {
 119                 CDEBUG(D_INODE, "truncate on inode %lu with no objects\n",
 120                        inode->i_ino);
 121                 EXIT;
 122                 return;
 123         }
 124
 125         if (lov_merge_size(lsm, 0) == inode->i_size) {
 126                 CDEBUG(D_VFSTRACE, "skipping punch for "LPX64" (size = %llu)\n",
 127                        lsm->lsm_object_id, inode->i_size);
 128         } else {
 129                 CDEBUG(D_INFO, "calling punch for "LPX64" (new size %llu)\n",
 130                        lsm->lsm_object_id, inode->i_size);
 131
 132                 oa.o_id = lsm->lsm_object_id;
 133                 oa.o_valid = OBD_MD_FLID;
 134                 obdo_from_inode(&oa, inode, OBD_MD_FLTYPE | OBD_MD_FLMODE |
 135                                 OBD_MD_FLATIME |OBD_MD_FLMTIME |OBD_MD_FLCTIME);
 136
 137                 /* truncate == punch from new size to absolute end of file */
 138                 /* NB: must call obd_punch with i_sem held!  It updates kms! */
 139                 rc = obd_punch(ll_i2obdexp(inode), &oa, lsm, inode->i_size,
 140                                OBD_OBJECT_EOF, NULL);
 141                 if (rc)
 142                         CERROR("obd_truncate fails (%d) ino %lu\n", rc,
 143                                inode->i_ino);
 144                 else
 145                         obdo_to_inode(inode, &oa, OBD_MD_FLSIZE|OBD_MD_FLBLOCKS|
 146                                       OBD_MD_FLATIME | OBD_MD_FLMTIME |
 147                                       OBD_MD_FLCTIME);
 148         }
 149
 150         EXIT;
 151         return;
 152 } /* ll_truncate */
 153
 154 __u64 lov_merge_size(struct lov_stripe_md *lsm, int kms);
 155 int ll_prepare_write(struct file *file, struct page *page, unsigned from,
 156                      unsigned to)
 157 {
 158         struct inode *inode = page->mapping->host;
 159         struct ll_inode_info *lli = ll_i2info(inode);
 160         struct lov_stripe_md *lsm = lli->lli_smd;
 161         obd_off offset = ((obd_off)page->index) << PAGE_SHIFT;
 162         struct brw_page pga;
 163         struct obdo oa;
 164         __u64 kms;
 165         int rc = 0;
 166         ENTRY;
 167
 168         LASSERT(PageLocked(page));
 169         (void)llap_cast_private(page); /* assertion */
 170
 171         /* Check to see if we should return -EIO right away */
 172         pga.pg = page;
 173         pga.off = offset;
 174         pga.count = PAGE_SIZE;
 175         pga.flag = 0;
 176
 177         oa.o_id = lsm->lsm_object_id;
 178         oa.o_mode = inode->i_mode;
 179         oa.o_valid = OBD_MD_FLID | OBD_MD_FLMODE | OBD_MD_FLTYPE;
 180
 181         rc = obd_brw(OBD_BRW_CHECK, ll_i2obdexp(inode), &oa, lsm, 1, &pga,
 182                      NULL);
 183         if (rc)
 184                 RETURN(rc);
 185
 186         if (PageUptodate(page)) {
 187                 LL_CDEBUG_PAGE(D_PAGE, page, "uptodate\n");
 188                 RETURN(0);
 189         }
 190
 191         /* We're completely overwriting an existing page, so _don't_ set it up
 192          * to date until commit_write */
 193         if (from == 0 && to == PAGE_SIZE) {
 194                 LL_CDEBUG_PAGE(D_PAGE, page, "full page write\n");
 195                 POISON_PAGE(page, 0x11);
 196                 RETURN(0);
 197         }
 198
 199         /* If are writing to a new page, no need to read old data.  The extent
 200          * locking will have updated the KMS, and for our purposes here we can
 201          * treat it like i_size. */
 202         kms = lov_merge_size(lsm, 1);
 203         if (kms <= offset) {
 204                 LL_CDEBUG_PAGE(D_PAGE, page, "kms "LPU64" <= offset "LPU64"\n",
 205                                kms, offset);
 206                 memset(kmap(page), 0, PAGE_SIZE);
 207                 kunmap(page);
 208                 GOTO(prepare_done, rc = 0);
 209         }
 210
 211         /* XXX could be an async ocp read.. read-ahead? */
 212         rc = ll_brw(OBD_BRW_READ, inode, &oa, page, 0);
 213         if (rc == 0) {
 214                 /* bug 1598: don't clobber blksize */
 215                 oa.o_valid &= ~(OBD_MD_FLSIZE | OBD_MD_FLBLKSZ);
 216                 obdo_refresh_inode(inode, &oa, oa.o_valid);
 217         }
 218
 219         EXIT;
 220  prepare_done:
 221         if (rc == 0)
 222                 SetPageUptodate(page);
 223
 224         return rc;
 225 }
 226
 227 struct ll_async_page *llap_from_cookie(void *cookie)
 228 {
 229         struct ll_async_page *llap = cookie;
 230         if (llap->llap_magic != LLAP_MAGIC)
 231                 return ERR_PTR(-EINVAL);
 232         return llap;
 233 };
 234
 235 static int ll_ap_make_ready(void *data, int cmd)
 236 {
 237         struct ll_async_page *llap;
 238         struct page *page;
 239         ENTRY;
 240
 241         llap = llap_from_cookie(data);
 242         if (IS_ERR(llap))
 243                 RETURN(-EINVAL);
 244
 245         page = llap->llap_page;
 246
 247         LASSERT(cmd != OBD_BRW_READ);
 248
 249         /* we're trying to write, but the page is locked.. come back later */
 250         if (TryLockPage(page))
 251                 RETURN(-EAGAIN);
 252
 253         LL_CDEBUG_PAGE(D_PAGE, page, "made ready\n");
 254         page_cache_get(page);
 255
 256         /* if we left PageDirty we might get another writepage call
 257          * in the future.  list walkers are bright enough
 258          * to check page dirty so we can leave it on whatever list
 259          * its on.  XXX also, we're called with the cli list so if
 260          * we got the page cache list we'd create a lock inversion
 261          * with the removepage path which gets the page lock then the
 262          * cli lock */
 263         clear_page_dirty(page);
 264         RETURN(0);
 265 }
 266
 267 /* We have two reasons for giving llite the opportunity to change the
 268  * write length of a given queued page as it builds the RPC containing
 269  * the page:
 270  *
 271  * 1) Further extending writes may have landed in the page cache
 272  *    since a partial write first queued this page requiring us
 273  *    to write more from the page cache.
 274  * 2) We might have raced with truncate and want to avoid performing
 275  *    write RPCs that are just going to be thrown away by the
 276  *    truncate's punch on the storage targets.
 277  *
 278  * The kms serves these purposes as it is set at both truncate and extending
 279  * writes.
 280  */
 281 static int ll_ap_refresh_count(void *data, int cmd)
 282 {
 283         struct ll_async_page *llap;
 284         struct lov_stripe_md *lsm;
 285         struct page *page;
 286         __u64 kms;
 287         ENTRY;
 288
 289         /* readpage queues with _COUNT_STABLE, shouldn't get here. */
 290         LASSERT(cmd != OBD_BRW_READ);
 291
 292         llap = llap_from_cookie(data);
 293         if (IS_ERR(llap))
 294                 RETURN(PTR_ERR(llap));
 295
 296         page = llap->llap_page;
 297         lsm = ll_i2info(page->mapping->host)->lli_smd;
 298         kms = lov_merge_size(lsm, 1);
 299
 300         /* catch race with truncate */
 301         if (((__u64)page->index << PAGE_SHIFT) >= kms)
 302                 return 0;
 303
 304         /* catch sub-page write at end of file */
 305         if (((__u64)page->index << PAGE_SHIFT) + PAGE_SIZE > kms)
 306                 return kms % PAGE_SIZE;
 307
 308         return PAGE_SIZE;
 309 }
 310
 311 void ll_inode_fill_obdo(struct inode *inode, int cmd, struct obdo *oa)
 312 {
 313         struct lov_stripe_md *lsm;
 314         obd_flag valid_flags;
 315
 316         lsm = ll_i2info(inode)->lli_smd;
 317
 318         oa->o_id = lsm->lsm_object_id;
 319         oa->o_valid = OBD_MD_FLID;
 320         valid_flags = OBD_MD_FLTYPE | OBD_MD_FLATIME;
 321         if (cmd == OBD_BRW_WRITE) {
 322                 oa->o_valid |= OBD_MD_FLIFID | OBD_MD_FLEPOCH;
 323                 mdc_pack_fid(obdo_fid(oa), inode->i_ino, 0, inode->i_mode);
 324                 oa->o_easize = ll_i2info(inode)->lli_io_epoch;
 325
 326                 valid_flags |= OBD_MD_FLMTIME | OBD_MD_FLCTIME;
 327         }
 328
 329         obdo_from_inode(oa, inode, valid_flags);
 330 }
 331
 332 static void ll_ap_fill_obdo(void *data, int cmd, struct obdo *oa)
 333 {
 334         struct ll_async_page *llap;
 335         ENTRY;
 336
 337         llap = llap_from_cookie(data);
 338         if (IS_ERR(llap)) {
 339                 EXIT;
 340                 return;
 341         }
 342
 343         ll_inode_fill_obdo(llap->llap_page->mapping->host, cmd, oa);
 344         EXIT;
 345 }
 346
 347 static struct obd_async_page_ops ll_async_page_ops = {
 348         .ap_make_ready =        ll_ap_make_ready,
 349         .ap_refresh_count =     ll_ap_refresh_count,
 350         .ap_fill_obdo =         ll_ap_fill_obdo,
 351         .ap_completion =        ll_ap_completion,
 352 };
 353
 354 struct ll_async_page *llap_cast_private(struct page *page)
 355 {
 356         struct ll_async_page *llap = (struct ll_async_page *)page->private;
 357
 358         LASSERTF(llap == NULL || llap->llap_magic == LLAP_MAGIC,
 359                  "page %p private %lu gave magic %d which != %d\n",
 360                  page, page->private, llap->llap_magic, LLAP_MAGIC);
 361
 362         return llap;
 363 }
 364
 365 /* Try to shrink the page cache for the @sbi filesystem by 1/@shrink_fraction.
 366  *
 367  * There is an llap attached onto every page in lustre, linked off @sbi.
 368  * We add an llap to the list so we don't lose our place during list walking.
 369  * If llaps in the list are being moved they will only move to the end
 370  * of the LRU, and we aren't terribly interested in those pages here (we
 371  * start at the beginning of the list where the least-used llaps are.
 372  */
 373 int llap_shrink_cache(struct ll_sb_info *sbi, int shrink_fraction)
 374 {
 375         struct ll_async_page *llap, dummy_llap = { .llap_magic = 0xd11ad11a };
 376         unsigned long total, want, count = 0;
 377
 378         total = sbi->ll_async_page_count;
 379
 380         /* There can be a large number of llaps (600k or more in a large
 381          * memory machine) so the VM 1/6 shrink ratio is likely too much.
 382          * Since we are freeing pages also, we don't necessarily want to
 383          * shrink so much.  Limit to 40MB of pages + llaps per call. */
 384         if (shrink_fraction == 0)
 385                 want = sbi->ll_async_page_count - sbi->ll_async_page_max + 32;
 386         else
 387                 want = (total + shrink_fraction - 1) / shrink_fraction;
 388
 389         if (want > 40 << (20 - PAGE_CACHE_SHIFT))
 390                 want = 40 << (20 - PAGE_CACHE_SHIFT);
 391
 392         CDEBUG(D_CACHE, "shrinking %lu of %lu pages (1/%d)\n",
 393                want, total, shrink_fraction);
 394
 395         spin_lock(&sbi->ll_lock);
 396         list_add(&dummy_llap.llap_pglist_item, &sbi->ll_pglist);
 397
 398         while (--total >= 0 && count < want) {
 399                 struct page *page;
 400
 401                 if (unlikely(need_resched())) {
 402                         spin_unlock(&sbi->ll_lock);
 403                         cond_resched();
 404                         spin_lock(&sbi->ll_lock);
 405                 }
 406
 407                 llap = llite_pglist_next_llap(sbi,&dummy_llap.llap_pglist_item);
 408                 list_del_init(&dummy_llap.llap_pglist_item);
 409                 if (llap == NULL)
 410                         break;
 411
 412                 page = llap->llap_page;
 413                 LASSERT(page != NULL);
 414
 415                 list_add(&dummy_llap.llap_pglist_item, &llap->llap_pglist_item);
 416
 417                 /* Page needs/undergoing IO */
 418                 if (TryLockPage(page)) {
 419                         LL_CDEBUG_PAGE(D_PAGE, page, "can't lock\n");
 420                         continue;
 421                 }
 422
 423                 /* If page is dirty or undergoing IO don't discard it */
 424                 if (llap->llap_write_queued || PageDirty(page) ||
 425                     (!PageUptodate(page) &&
 426                      llap->llap_origin != LLAP_ORIGIN_READAHEAD)) {
 427                         unlock_page(page);
 428                         LL_CDEBUG_PAGE(D_PAGE, page, "can't drop from cache: "
 429                                        "%s%s%s%s origin %s\n",
 430                                        llap->llap_write_queued ? "wq " : "",
 431                                        PageDirty(page) ? "pd " : "",
 432                                        PageUptodate(page) ? "" : "!pu ",
 433                                        llap->llap_defer_uptodate ? "" : "!du",
 434                                        llap_origins[llap->llap_origin]);
 435                         continue;
 436                 }
 437
 438                 page_cache_get(page);
 439                 spin_unlock(&sbi->ll_lock);
 440
 441                 ++count;
 442                 LL_CDEBUG_PAGE(D_PAGE, page, "drop from cache %lu/%lu\n",
 443                                count, want);
 444                 if (page->mapping != NULL) {
 445                         ll_ra_accounting(page, page->mapping);
 446                         ll_truncate_complete_page(page);
 447                 }
 448                 unlock_page(page);
 449                 page_cache_release(page);
 450
 451                 spin_lock(&sbi->ll_lock);
 452         }
 453         list_del(&dummy_llap.llap_pglist_item);
 454         spin_unlock(&sbi->ll_lock);
 455
 456         CDEBUG(D_CACHE, "shrank %lu/%lu and left %lu unscanned\n",
 457                count, want, total);
 458
 459         return count;
 460 }
 461
 462 struct ll_async_page *llap_from_page(struct page *page, unsigned origin)
 463 {
 464         struct ll_async_page *llap;
 465         struct obd_export *exp;
 466         struct inode *inode = page->mapping->host;
 467         struct ll_sb_info *sbi = ll_i2sbi(inode);
 468         int rc;
 469         ENTRY;
 470
 471         LASSERT(ll_async_page_slab);
 472         LASSERTF(origin < LLAP__ORIGIN_MAX, "%u\n", origin);
 473
 474         llap = llap_cast_private(page);
 475         if (llap != NULL) {
 476                 /* move to end of LRU list */
 477                 spin_lock(&sbi->ll_lock);
 478                 sbi->ll_pglist_gen++;
 479                 list_del_init(&llap->llap_pglist_item);
 480                 list_add_tail(&llap->llap_pglist_item, &sbi->ll_pglist);
 481                 spin_unlock(&sbi->ll_lock);
 482                 GOTO(out, llap);
 483         }
 484
 485         exp = ll_i2obdexp(page->mapping->host);
 486         if (exp == NULL)
 487                 RETURN(ERR_PTR(-EINVAL));
 488
 489         /* limit the number of lustre-cached pages */
 490         if (sbi->ll_async_page_count >= sbi->ll_async_page_max)
 491                 llap_shrink_cache(sbi, 0);
 492
 493         OBD_SLAB_ALLOC(llap, ll_async_page_slab, SLAB_KERNEL,
 494                        ll_async_page_slab_size);
 495         if (llap == NULL)
 496                 RETURN(ERR_PTR(-ENOMEM));
 497         llap->llap_magic = LLAP_MAGIC;
 498         llap->llap_cookie = (void *)llap + size_round(sizeof(*llap));
 499         rc = obd_prep_async_page(exp, ll_i2info(inode)->lli_smd, NULL, page,
 500                                  (obd_off)page->index << PAGE_SHIFT,
 501                                  &ll_async_page_ops, llap, &llap->llap_cookie);
 502         if (rc) {
 503                 OBD_SLAB_FREE(llap, ll_async_page_slab,
 504                               ll_async_page_slab_size);
 505                 RETURN(ERR_PTR(rc));
 506         }
 507
 508         CDEBUG(D_CACHE, "llap %p page %p cookie %p obj off "LPU64"\n", llap,
 509                page, llap->llap_cookie, (obd_off)page->index << PAGE_SHIFT);
 510         /* also zeroing the PRIVBITS low order bitflags */
 511         __set_page_ll_data(page, llap);
 512         llap->llap_page = page;
 513
 514         spin_lock(&sbi->ll_lock);
 515         sbi->ll_pglist_gen++;
 516         sbi->ll_async_page_count++;
 517         list_add_tail(&llap->llap_pglist_item, &sbi->ll_pglist);
 518         spin_unlock(&sbi->ll_lock);
 519
 520 out:
 521         llap->llap_origin = origin;
 522         RETURN(llap);
 523 }
 524
 525 static int queue_or_sync_write(struct obd_export *exp, struct inode *inode,
 526                                struct ll_async_page *llap,
 527                                unsigned to, obd_flag async_flags)
 528 {
 529         unsigned long size_index = inode->i_size >> PAGE_SHIFT;
 530         struct obd_io_group *oig;
 531         int rc;
 532         ENTRY;
 533
 534         /* _make_ready only sees llap once we've unlocked the page */
 535         llap->llap_write_queued = 1;
 536         rc = obd_queue_async_io(exp, ll_i2info(inode)->lli_smd, NULL,
 537                                 llap->llap_cookie, OBD_BRW_WRITE, 0, 0, 0,
 538                                 async_flags);
 539         if (rc == 0) {
 540                 LL_CDEBUG_PAGE(D_PAGE, llap->llap_page, "write queued\n");
 541                 //llap_write_pending(inode, llap);
 542                 GOTO(out, 0);
 543         }
 544
 545         llap->llap_write_queued = 0;
 546
 547         rc = oig_init(&oig);
 548         if (rc)
 549                 GOTO(out, rc);
 550
 551         /* make full-page requests if we are not at EOF (bug 4410) */
 552         if (to != PAGE_SIZE && llap->llap_page->index < size_index) {
 553                 LL_CDEBUG_PAGE(D_PAGE, llap->llap_page,
 554                                "sync write before EOF: size_index %lu, to %d\n",
 555                                size_index, to);
 556                 to = PAGE_SIZE;
 557         } else if (to != PAGE_SIZE && llap->llap_page->index == size_index) {
 558                 int size_to = inode->i_size & ~PAGE_MASK;
 559                 LL_CDEBUG_PAGE(D_PAGE, llap->llap_page,
 560                                "sync write at EOF: size_index %lu, to %d/%d\n",
 561                                size_index, to, size_to);
 562                 if (to < size_to)
 563                         to = size_to;
 564         }
 565
 566         rc = obd_queue_group_io(exp, ll_i2info(inode)->lli_smd, NULL, oig,
 567                                 llap->llap_cookie, OBD_BRW_WRITE, 0, to, 0,
 568                                 ASYNC_READY | ASYNC_URGENT |
 569                                 ASYNC_COUNT_STABLE | ASYNC_GROUP_SYNC);
 570         if (rc)
 571                 GOTO(free_oig, rc);
 572
 573         rc = obd_trigger_group_io(exp, ll_i2info(inode)->lli_smd, NULL, oig);
 574         if (rc)
 575                 GOTO(free_oig, rc);
 576
 577         rc = oig_wait(oig);
 578
 579         if (!rc && async_flags & ASYNC_READY)
 580                 unlock_page(llap->llap_page);
 581
 582         LL_CDEBUG_PAGE(D_PAGE, llap->llap_page, "sync write returned %d\n", rc);
 583
 584 free_oig:
 585         oig_release(oig);
 586 out:
 587         RETURN(rc);
 588 }
 589
 590 /* update our write count to account for i_size increases that may have
 591  * happened since we've queued the page for io. */
 592
 593 /* be careful not to return success without setting the page Uptodate or
 594  * the next pass through prepare_write will read in stale data from disk. */
 595 int ll_commit_write(struct file *file, struct page *page, unsigned from,
 596                     unsigned to)
 597 {
 598         struct inode *inode = page->mapping->host;
 599         struct ll_inode_info *lli = ll_i2info(inode);
 600         struct lov_stripe_md *lsm = lli->lli_smd;
 601         struct obd_export *exp;
 602         struct ll_async_page *llap;
 603         loff_t size;
 604         int rc = 0;
 605         ENTRY;
 606
 607         SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
 608         LASSERT(inode == file->f_dentry->d_inode);
 609         LASSERT(PageLocked(page));
 610
 611         CDEBUG(D_INODE, "inode %p is writing page %p from %d to %d at %lu\n",
 612                inode, page, from, to, page->index);
 613
 614         llap = llap_from_page(page, LLAP_ORIGIN_COMMIT_WRITE);
 615         if (IS_ERR(llap))
 616                 RETURN(PTR_ERR(llap));
 617
 618         exp = ll_i2obdexp(inode);
 619         if (exp == NULL)
 620                 RETURN(-EINVAL);
 621
 622         /* queue a write for some time in the future the first time we
 623          * dirty the page */
 624         if (!PageDirty(page)) {
 625                 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats,
 626                                      LPROC_LL_DIRTY_MISSES);
 627
 628                 rc = queue_or_sync_write(exp, inode, llap, to, 0);
 629                 if (rc)
 630                         GOTO(out, rc);
 631         } else {
 632                 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats,
 633                                      LPROC_LL_DIRTY_HITS);
 634         }
 635
 636         /* put the page in the page cache, from now on ll_removepage is
 637          * responsible for cleaning up the llap.
 638          * only set page dirty when it's queued to be write out */
 639         if (llap->llap_write_queued)
 640                 set_page_dirty(page);
 641
 642 out:
 643         size = (((obd_off)page->index) << PAGE_SHIFT) + to;
 644         if (rc == 0) {
 645                 spin_lock(&lli->lli_lock);
 646                 obd_increase_kms(exp, lsm, size);
 647                 spin_unlock(&lli->lli_lock);
 648                 if (size > inode->i_size)
 649                         inode->i_size = size;
 650                 SetPageUptodate(page);
 651         } else if (size > inode->i_size) {
 652                 /* this page beyond the pales of i_size, so it can't be
 653                  * truncated in ll_p_r_e during lock revoking. we must
 654                  * teardown our book-keeping here. */
 655                 ll_removepage(page);
 656         }
 657         RETURN(rc);
 658 }
 659
 660 static unsigned long ll_ra_count_get(struct ll_sb_info *sbi, unsigned long len)
 661 {
 662         struct ll_ra_info *ra = &sbi->ll_ra_info;
 663         unsigned long ret;
 664         ENTRY;
 665
 666         spin_lock(&sbi->ll_lock);
 667         ret = min(ra->ra_max_pages - ra->ra_cur_pages, len);
 668         ra->ra_cur_pages += ret;
 669         spin_unlock(&sbi->ll_lock);
 670
 671         RETURN(ret);
 672 }
 673
 674 static void ll_ra_count_put(struct ll_sb_info *sbi, unsigned long len)
 675 {
 676         struct ll_ra_info *ra = &sbi->ll_ra_info;
 677         spin_lock(&sbi->ll_lock);
 678         LASSERTF(ra->ra_cur_pages >= len, "r_c_p %lu len %lu\n",
 679                  ra->ra_cur_pages, len);
 680         ra->ra_cur_pages -= len;
 681         spin_unlock(&sbi->ll_lock);
 682 }
 683
 684 /* called for each page in a completed rpc.*/
 685 void ll_ap_completion(void *data, int cmd, struct obdo *oa, int rc)
 686 {
 687         struct ll_async_page *llap;
 688         struct page *page;
 689         ENTRY;
 690
 691         llap = llap_from_cookie(data);
 692         if (IS_ERR(llap)) {
 693                 EXIT;
 694                 return;
 695         }
 696
 697         page = llap->llap_page;
 698         LASSERT(PageLocked(page));
 699
 700         LL_CDEBUG_PAGE(D_PAGE, page, "completing cmd %d with %d\n", cmd, rc);
 701
 702         if (cmd == OBD_BRW_READ && llap->llap_defer_uptodate)
 703                 ll_ra_count_put(ll_i2sbi(page->mapping->host), 1);
 704
 705         if (rc == 0)  {
 706                 if (cmd == OBD_BRW_READ) {
 707                         if (!llap->llap_defer_uptodate)
 708                                 SetPageUptodate(page);
 709                 } else {
 710                         llap->llap_write_queued = 0;
 711                 }
 712                 ClearPageError(page);
 713         } else {
 714                 if (cmd == OBD_BRW_READ) {
 715                         llap->llap_defer_uptodate = 0;
 716                 } else {
 717                         SetPageDirty(page);
 718                         ClearPageLaunder(page);
 719                 }
 720                 SetPageError(page);
 721         }
 722
 723         unlock_page(page);
 724
 725         if (0 && cmd == OBD_BRW_WRITE) {
 726                 llap_write_complete(page->mapping->host, llap);
 727                 ll_try_done_writing(page->mapping->host);
 728         }
 729
 730         if (PageWriteback(page)) {
 731                 end_page_writeback(page);
 732         }
 733         page_cache_release(page);
 734         EXIT;
 735 }
 736
 737 /* the kernel calls us here when a page is unhashed from the page cache.
 738  * the page will be locked and the kernel is holding a spinlock, so
 739  * we need to be careful.  we're just tearing down our book-keeping
 740  * here. */
 741 void ll_removepage(struct page *page)
 742 {
 743         struct inode *inode = page->mapping->host;
 744         struct obd_export *exp;
 745         struct ll_async_page *llap;
 746         struct ll_sb_info *sbi = ll_i2sbi(inode);
 747         int rc;
 748         ENTRY;
 749
 750         LASSERT(!in_interrupt());
 751
 752         /* sync pages or failed read pages can leave pages in the page
 753          * cache that don't have our data associated with them anymore */
 754         if (page->private == 0) {
 755                 EXIT;
 756                 return;
 757         }
 758
 759         LL_CDEBUG_PAGE(D_PAGE, page, "being evicted\n");
 760
 761         exp = ll_i2obdexp(inode);
 762         if (exp == NULL) {
 763                 CERROR("page %p ind %lu gave null export\n", page, page->index);
 764                 EXIT;
 765                 return;
 766         }
 767
 768         llap = llap_from_page(page, 0);
 769         if (IS_ERR(llap)) {
 770                 CERROR("page %p ind %lu couldn't find llap: %ld\n", page,
 771                        page->index, PTR_ERR(llap));
 772                 EXIT;
 773                 return;
 774         }
 775
 776         //llap_write_complete(inode, llap);
 777         rc = obd_teardown_async_page(exp, ll_i2info(inode)->lli_smd, NULL,
 778                                      llap->llap_cookie);
 779         if (rc != 0)
 780                 CERROR("page %p ind %lu failed: %d\n", page, page->index, rc);
 781
 782         /* this unconditional free is only safe because the page lock
 783          * is providing exclusivity to memory pressure/truncate/writeback..*/
 784         __clear_page_ll_data(page);
 785
 786         spin_lock(&sbi->ll_lock);
 787         if (!list_empty(&llap->llap_pglist_item))
 788                 list_del_init(&llap->llap_pglist_item);
 789         sbi->ll_pglist_gen++;
 790         sbi->ll_async_page_count--;
 791         spin_unlock(&sbi->ll_lock);
 792         OBD_SLAB_FREE(llap, ll_async_page_slab, ll_async_page_slab_size);
 793         EXIT;
 794 }
 795
 796 static int ll_page_matches(struct page *page)
 797 {
 798         struct lustre_handle match_lockh = {0};
 799         struct inode *inode = page->mapping->host;
 800         ldlm_policy_data_t page_extent;
 801         int flags, matches;
 802         ENTRY;
 803
 804         page_extent.l_extent.start = (__u64)page->index << PAGE_CACHE_SHIFT;
 805         page_extent.l_extent.end =
 806                 page_extent.l_extent.start + PAGE_CACHE_SIZE - 1;
 807         flags = LDLM_FL_CBPENDING | LDLM_FL_BLOCK_GRANTED | LDLM_FL_TEST_LOCK;
 808         matches = obd_match(ll_i2sbi(inode)->ll_osc_exp,
 809                             ll_i2info(inode)->lli_smd, LDLM_EXTENT,
 810                             &page_extent, LCK_PR | LCK_PW, &flags, inode,
 811                             &match_lockh);
 812         RETURN(matches);
 813 }
 814
 815 static int ll_issue_page_read(struct obd_export *exp,
 816                               struct ll_async_page *llap,
 817                               struct obd_io_group *oig, int defer)
 818 {
 819         struct page *page = llap->llap_page;
 820         int rc;
 821
 822         page_cache_get(page);
 823         llap->llap_defer_uptodate = defer;
 824         llap->llap_ra_used = 0;
 825         rc = obd_queue_group_io(exp, ll_i2info(page->mapping->host)->lli_smd,
 826                                 NULL, oig, llap->llap_cookie, OBD_BRW_READ, 0,
 827                                 PAGE_SIZE, 0, ASYNC_COUNT_STABLE | ASYNC_READY
 828                                               | ASYNC_URGENT);
 829         if (rc) {
 830                 LL_CDEBUG_PAGE(D_ERROR, page, "read queue failed: rc %d\n", rc);
 831                 page_cache_release(page);
 832         }
 833         RETURN(rc);
 834 }
 835
 836 static void ll_ra_stats_inc_unlocked(struct ll_ra_info *ra, enum ra_stat which)
 837 {
 838         LASSERTF(which >= 0 && which < _NR_RA_STAT, "which: %u\n", which);
 839         ra->ra_stats[which]++;
 840 }
 841
 842 static void ll_ra_stats_inc(struct address_space *mapping, enum ra_stat which)
 843 {
 844         struct ll_sb_info *sbi = ll_i2sbi(mapping->host);
 845         struct ll_ra_info *ra = &ll_i2sbi(mapping->host)->ll_ra_info;
 846
 847         spin_lock(&sbi->ll_lock);
 848         ll_ra_stats_inc_unlocked(ra, which);
 849         spin_unlock(&sbi->ll_lock);
 850 }
 851
 852 void ll_ra_accounting(struct page *page, struct address_space *mapping)
 853 {
 854         struct ll_async_page *llap;
 855
 856         llap = llap_from_page(page, LLAP_ORIGIN_WRITEPAGE);
 857         if (IS_ERR(llap))
 858                 return;
 859
 860         if (!llap->llap_defer_uptodate || llap->llap_ra_used)
 861                 return;
 862
 863         ll_ra_stats_inc(mapping, RA_STAT_DISCARDED);
 864 }
 865
 866 #define RAS_CDEBUG(ras) \
 867         CDEBUG(D_READA, "lrp %lu c %lu ws %lu wl %lu nra %lu\n",        \
 868                ras->ras_last_readpage, ras->ras_consecutive,            \
 869                ras->ras_window_start, ras->ras_window_len,              \
 870                ras->ras_next_readahead);
 871
 872 static int index_in_window(unsigned long index, unsigned long point,
 873                            unsigned long before, unsigned long after)
 874 {
 875         unsigned long start = point - before, end = point + after;
 876
 877         if (start > point)
 878                start = 0;
 879         if (end < point)
 880                end = ~0;
 881
 882         return start <= index && index <= end;
 883 }
 884
 885 static int ll_readahead(struct ll_readahead_state *ras,
 886                          struct obd_export *exp, struct address_space *mapping,
 887                          struct obd_io_group *oig, int flags)
 888 {
 889         unsigned long i, start = 0, end = 0, reserved;
 890         struct ll_async_page *llap;
 891         struct page *page;
 892         int rc, ret = 0, match_failed = 0;
 893         __u64 kms;
 894         unsigned int gfp_mask;
 895         ENTRY;
 896
 897         kms = lov_merge_size(ll_i2info(mapping->host)->lli_smd, 1);
 898         if (kms == 0) {
 899                 ll_ra_stats_inc(mapping, RA_STAT_ZERO_LEN);
 900                 RETURN(0);
 901         }
 902
 903         spin_lock(&ras->ras_lock);
 904         /* reserve a part of the read-ahead window that we'll be issuing */
 905         if (ras->ras_window_len) {
 906                 start = ras->ras_next_readahead;
 907                 end = ras->ras_window_start + ras->ras_window_len - 1;
 908                 end = min(end, (unsigned long)((kms - 1) >> PAGE_CACHE_SHIFT));
 909                 ras->ras_next_readahead = max(end, end + 1);
 910
 911                 RAS_CDEBUG(ras);
 912         }
 913         spin_unlock(&ras->ras_lock);
 914
 915         if (end == 0) {
 916                 ll_ra_stats_inc(mapping, RA_STAT_ZERO_WINDOW);
 917                 RETURN(0);
 918         }
 919
 920         reserved = ll_ra_count_get(ll_i2sbi(mapping->host), end - start + 1);
 921         if (reserved < end - start + 1)
 922                 ll_ra_stats_inc(mapping, RA_STAT_MAX_IN_FLIGHT);
 923
 924         gfp_mask = GFP_HIGHUSER & ~__GFP_WAIT;
 925 #ifdef __GFP_NOWARN
 926         gfp_mask |= __GFP_NOWARN;
 927 #endif
 928
 929         for (i = start; reserved > 0 && !match_failed && i <= end; i++) {
 930                 /* skip locked pages from previous readpage calls */
 931                 page = grab_cache_page_nowait_gfp(mapping, i, gfp_mask);
 932                 if (page == NULL) {
 933                         CDEBUG(D_READA, "g_c_p_n failed\n");
 934                         continue;
 935                 }
 936
 937                 /* we do this first so that we can see the page in the /proc
 938                  * accounting */
 939                 llap = llap_from_page(page, LLAP_ORIGIN_READAHEAD);
 940                 if (IS_ERR(llap) || llap->llap_defer_uptodate)
 941                         goto next_page;
 942
 943                 /* skip completed pages */
 944                 if (Page_Uptodate(page))
 945                         goto next_page;
 946
 947                 /* bail when we hit the end of the lock. */
 948                 if ((rc = ll_page_matches(page)) <= 0) {
 949                         LL_CDEBUG_PAGE(D_READA | D_PAGE, page,
 950                                        "lock match failed: rc %d\n", rc);
 951                         ll_ra_stats_inc(mapping, RA_STAT_FAILED_MATCH);
 952                         match_failed = 1;
 953                         goto next_page;
 954                 }
 955
 956                 rc = ll_issue_page_read(exp, llap, oig, 1);
 957                 if (rc == 0) {
 958                         reserved--;
 959                         ret++;
 960                         LL_CDEBUG_PAGE(D_READA| D_PAGE, page,
 961                                        "started read-ahead\n");
 962                 }
 963                 if (rc) {
 964         next_page:
 965                         LL_CDEBUG_PAGE(D_READA | D_PAGE, page,
 966                                        "skipping read-ahead\n");
 967
 968                         unlock_page(page);
 969                 }
 970                 page_cache_release(page);
 971         }
 972
 973         LASSERTF(reserved >= 0, "reserved %lu\n", reserved);
 974         if (reserved != 0)
 975                 ll_ra_count_put(ll_i2sbi(mapping->host), reserved);
 976         if (i == end + 1 && end == (kms >> PAGE_CACHE_SHIFT))
 977                 ll_ra_stats_inc(mapping, RA_STAT_EOF);
 978
 979         /* if we didn't get to the end of the region we reserved from
 980          * the ras we need to go back and update the ras so that the
 981          * next read-ahead tries from where we left off.  we only do so
 982          * if the region we failed to issue read-ahead on is still ahead
 983          * of the app and behind the next index to start read-ahead from */
 984         if (i != end + 1) {
 985                 spin_lock(&ras->ras_lock);
 986                 if (i < ras->ras_next_readahead &&
 987                     index_in_window(i, ras->ras_window_start, 0,
 988                                     ras->ras_window_len)) {
 989                         ras->ras_next_readahead = i;
 990                         RAS_CDEBUG(ras);
 991                 }
 992                 spin_unlock(&ras->ras_lock);
 993         }
 994
 995         RETURN(ret);
 996 }
 997
 998 static void ras_set_start(struct ll_readahead_state *ras, unsigned long index)
 999 {
1000         ras->ras_window_start = index & (~(PTLRPC_MAX_BRW_PAGES - 1));
1001 }
1002
1003 /* called with the ras_lock held or from places where it doesn't matter */
1004 static void ras_reset(struct ll_readahead_state *ras, unsigned long index)
1005 {
1006         ras->ras_last_readpage = index;
1007         ras->ras_consecutive = 1;
1008         ras->ras_window_len = 0;
1009         ras_set_start(ras, index);
1010         ras->ras_next_readahead = ras->ras_window_start;
1011
1012         RAS_CDEBUG(ras);
1013 }
1014
1015 void ll_readahead_init(struct inode *inode, struct ll_readahead_state *ras)
1016 {
1017         spin_lock_init(&ras->ras_lock);
1018         ras_reset(ras, 0);
1019 }
1020
1021 static void ras_update(struct ll_sb_info *sbi, struct ll_readahead_state *ras,
1022                        unsigned long index, unsigned hit)
1023 {
1024         struct ll_ra_info *ra = &sbi->ll_ra_info;
1025         int zero = 0;
1026         ENTRY;
1027
1028         spin_lock(&sbi->ll_lock);
1029         spin_lock(&ras->ras_lock);
1030
1031         ll_ra_stats_inc_unlocked(ra, hit ? RA_STAT_HIT : RA_STAT_MISS);
1032
1033         /* reset the read-ahead window in two cases.  First when the app seeks
1034          * or reads to some other part of the file.  Secondly if we get a
1035          * read-ahead miss that we think we've previously issued.  This can
1036          * be a symptom of there being so many read-ahead pages that the VM is
1037          * reclaiming it before we get to it. */
1038         if (!index_in_window(index, ras->ras_last_readpage, 8, 8)) {
1039                 zero = 1;
1040                 ll_ra_stats_inc_unlocked(ra, RA_STAT_DISTANT_READPAGE);
1041         } else if (!hit && ras->ras_window_len &&
1042                    index < ras->ras_next_readahead &&
1043                    index_in_window(index, ras->ras_window_start, 0,
1044                                    ras->ras_window_len)) {
1045                 zero = 1;
1046                 ll_ra_stats_inc_unlocked(ra, RA_STAT_MISS_IN_WINDOW);
1047         }
1048
1049         if (zero) {
1050                 ras_reset(ras, index);
1051                 GOTO(out_unlock, 0);
1052         }
1053
1054         ras->ras_last_readpage = index;
1055         ras->ras_consecutive++;
1056         ras_set_start(ras, index);
1057         ras->ras_next_readahead = max(ras->ras_window_start,
1058                                       ras->ras_next_readahead);
1059
1060         /* wait for a few pages to arrive before issuing readahead to avoid
1061          * the worst overutilization */
1062         if (ras->ras_consecutive == 3) {
1063                 ras->ras_window_len = PTLRPC_MAX_BRW_PAGES;
1064                 GOTO(out_unlock, 0);
1065         }
1066
1067         /* we need to increase the window sometimes.  we'll arbitrarily
1068          * do it half-way through the pages in an rpc */
1069         if ((index & (PTLRPC_MAX_BRW_PAGES - 1)) ==
1070             (PTLRPC_MAX_BRW_PAGES >> 1)) {
1071                 ras->ras_window_len += PTLRPC_MAX_BRW_PAGES;
1072                 ras->ras_window_len = min(ras->ras_window_len,
1073                                           ra->ra_max_pages);
1074         }
1075
1076         EXIT;
1077 out_unlock:
1078         RAS_CDEBUG(ras);
1079         spin_unlock(&ras->ras_lock);
1080         spin_unlock(&sbi->ll_lock);
1081         return;
1082 }
1083
1084 /*
1085  * for now we do our readpage the same on both 2.4 and 2.5.  The kernel's
1086  * read-ahead assumes it is valid to issue readpage all the way up to
1087  * i_size, but our dlm locks make that not the case.  We disable the
1088  * kernel's read-ahead and do our own by walking ahead in the page cache
1089  * checking for dlm lock coverage.  the main difference between 2.4 and
1090  * 2.6 is how read-ahead gets batched and issued, but we're using our own,
1091  * so they look the same.
1092  */
1093 int ll_readpage(struct file *filp, struct page *page)
1094 {
1095         struct ll_file_data *fd = filp->private_data;
1096         struct inode *inode = page->mapping->host;
1097         struct obd_export *exp;
1098         struct ll_async_page *llap;
1099         struct obd_io_group *oig = NULL;
1100         int rc;
1101         ENTRY;
1102
1103         LASSERT(PageLocked(page));
1104         LASSERT(!PageUptodate(page));
1105         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),offset="LPX64"\n",
1106                inode->i_ino, inode->i_generation, inode,
1107                (((obd_off)page->index) << PAGE_SHIFT));
1108         LASSERT(atomic_read(&filp->f_dentry->d_inode->i_count) > 0);
1109
1110         rc = oig_init(&oig);
1111         if (rc < 0)
1112                 GOTO(out, rc);
1113
1114         exp = ll_i2obdexp(inode);
1115         if (exp == NULL)
1116                 GOTO(out, rc = -EINVAL);
1117
1118         llap = llap_from_page(page, LLAP_ORIGIN_READPAGE);
1119         if (IS_ERR(llap))
1120                 GOTO(out, rc = PTR_ERR(llap));
1121
1122         if (ll_i2sbi(inode)->ll_ra_info.ra_max_pages)
1123                 ras_update(ll_i2sbi(inode), &fd->fd_ras, page->index,
1124                            llap->llap_defer_uptodate);
1125
1126         if (llap->llap_defer_uptodate) {
1127                 llap->llap_ra_used = 1;
1128                 rc = ll_readahead(&fd->fd_ras, exp, page->mapping, oig,
1129                                   fd->fd_flags);
1130                 if (rc > 0)
1131                         obd_trigger_group_io(exp, ll_i2info(inode)->lli_smd,
1132                                              NULL, oig);
1133                 LL_CDEBUG_PAGE(D_PAGE, page, "marking uptodate from defer\n");
1134                 SetPageUptodate(page);
1135                 unlock_page(page);
1136                 GOTO(out_oig, rc = 0);
1137         }
1138
1139         rc = ll_page_matches(page);
1140         if (rc < 0) {
1141                 LL_CDEBUG_PAGE(D_ERROR, page, "lock match failed: rc %d\n", rc);
1142                 GOTO(out, rc);
1143         }
1144
1145         if (rc == 0) {
1146 #if 0
1147                 CWARN("ino %lu page %lu (%llu) not covered by "
1148                       "a lock (mmap?).  check debug logs.\n",
1149                       inode->i_ino, page->index,
1150                       (long long)page->index << PAGE_CACHE_SHIFT);
1151 #endif
1152         }
1153
1154         rc = ll_issue_page_read(exp, llap, oig, 0);
1155         if (rc)
1156                 GOTO(out, rc);
1157
1158         LL_CDEBUG_PAGE(D_PAGE, page, "queued readpage\n");
1159         if (ll_i2sbi(inode)->ll_ra_info.ra_max_pages)
1160                 ll_readahead(&fd->fd_ras, exp, page->mapping, oig,
1161                              fd->fd_flags);
1162
1163         rc = obd_trigger_group_io(exp, ll_i2info(inode)->lli_smd, NULL, oig);
1164
1165 out:
1166         if (rc)
1167                 unlock_page(page);
1168 out_oig:
1169         if (oig != NULL)
1170                 oig_release(oig);
1171         RETURN(rc);
1172 }