lustre/llite/rw.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright  2008 Sun Microsystems, Inc. All rights reserved
  30  * Use is subject to license terms.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/llite/rw.c
  37  *
  38  * Lustre Lite I/O page cache routines shared by different kernel revs
  39  */
  40
  41 #ifndef AUTOCONF_INCLUDED
  42 #include <linux/config.h>
  43 #endif
  44 #include <linux/kernel.h>
  45 #include <linux/mm.h>
  46 #include <linux/string.h>
  47 #include <linux/stat.h>
  48 #include <linux/errno.h>
  49 #include <linux/smp_lock.h>
  50 #include <linux/unistd.h>
  51 #include <linux/version.h>
  52 #include <asm/system.h>
  53 #include <asm/uaccess.h>
  54
  55 #include <linux/fs.h>
  56 #include <linux/stat.h>
  57 #include <asm/uaccess.h>
  58 #include <linux/mm.h>
  59 #include <linux/pagemap.h>
  60 #include <linux/smp_lock.h>
  61
  62 #define DEBUG_SUBSYSTEM S_LLITE
  63
  64 #include <linux/page-flags.h>
  65
  66 #include <lustre_lite.h>
  67 #include "llite_internal.h"
  68 #include <linux/lustre_compat25.h>
  69
  70 #ifndef list_for_each_prev_safe
  71 #define list_for_each_prev_safe(pos, n, head) \
  72         for (pos = (head)->prev, n = pos->prev; pos != (head); \
  73                 pos = n, n = pos->prev )
  74 #endif
  75
  76 cfs_mem_cache_t *ll_async_page_slab = NULL;
  77 size_t ll_async_page_slab_size = 0;
  78
  79 /* SYNCHRONOUS I/O to object storage for an inode */
  80 static int ll_brw(int cmd, struct inode *inode, struct obdo *oa,
  81                   struct page *page, int flags)
  82 {
  83         struct ll_inode_info *lli = ll_i2info(inode);
  84         struct lov_stripe_md *lsm = lli->lli_smd;
  85         struct obd_info oinfo = { { { 0 } } };
  86         struct brw_page pg;
  87         int rc;
  88         ENTRY;
  89
  90         pg.pg = page;
  91         pg.off = ((obd_off)page->index) << CFS_PAGE_SHIFT;
  92
  93         if ((cmd & OBD_BRW_WRITE) && (pg.off+CFS_PAGE_SIZE>i_size_read(inode)))
  94                 pg.count = i_size_read(inode) % CFS_PAGE_SIZE;
  95         else
  96                 pg.count = CFS_PAGE_SIZE;
  97
  98         LL_CDEBUG_PAGE(D_PAGE, page, "%s %d bytes ino %lu at "LPU64"/"LPX64"\n",
  99                        cmd & OBD_BRW_WRITE ? "write" : "read", pg.count,
 100                        inode->i_ino, pg.off, pg.off);
 101         if (pg.count == 0) {
 102                 CERROR("ZERO COUNT: ino %lu: size %p:%Lu(%p:%Lu) idx %lu off "
 103                        LPU64"\n", inode->i_ino, inode, i_size_read(inode),
 104                        page->mapping->host, i_size_read(page->mapping->host),
 105                        page->index, pg.off);
 106         }
 107
 108         pg.flag = flags;
 109
 110         if (cmd & OBD_BRW_WRITE)
 111                 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_BRW_WRITE,
 112                                    pg.count);
 113         else
 114                 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_BRW_READ,
 115                            pg.count);
 116         oinfo.oi_oa = oa;
 117         oinfo.oi_md = lsm;
 118         rc = obd_brw(cmd, ll_i2obdexp(inode), &oinfo, 1, &pg, NULL);
 119         if (rc == 0)
 120                 obdo_to_inode(inode, oa, OBD_MD_FLBLOCKS);
 121         else if (rc != -EIO)
 122                 CERROR("error from obd_brw: rc = %d\n", rc);
 123         RETURN(rc);
 124 }
 125
 126 int ll_file_punch(struct inode * inode, loff_t new_size, int srvlock)
 127 {
 128         struct ll_inode_info *lli = ll_i2info(inode);
 129         struct obd_info oinfo = { { { 0 } } };
 130         struct obdo oa = { 0 };
 131         obd_valid valid;
 132         int rc;
 133
 134         ENTRY;
 135         CDEBUG(D_INFO, "calling punch for "LPX64" (new size %Lu=%#Lx)\n",
 136                lli->lli_smd->lsm_object_id, new_size, new_size);
 137
 138         oinfo.oi_md = lli->lli_smd;
 139         oinfo.oi_policy.l_extent.start = new_size;
 140         oinfo.oi_policy.l_extent.end = OBD_OBJECT_EOF;
 141         oinfo.oi_oa = &oa;
 142         oa.o_id = lli->lli_smd->lsm_object_id;
 143         oa.o_gr = lli->lli_smd->lsm_object_gr;
 144         oa.o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
 145
 146         valid = OBD_MD_FLTYPE | OBD_MD_FLMODE |OBD_MD_FLFID |
 147                 OBD_MD_FLATIME | OBD_MD_FLUID | OBD_MD_FLGID | OBD_MD_FLGENER |
 148                 OBD_MD_FLBLOCKS;
 149         if (srvlock) {
 150                 /* set OBD_MD_FLFLAGS in o_valid, only if we
 151                  * set OBD_FL_TRUNCLOCK, otherwise ost_punch
 152                  * and filter_setattr get confused, see the comment
 153                  * in ost_punch */
 154                 oa.o_flags = OBD_FL_TRUNCLOCK;
 155                 oa.o_valid |= OBD_MD_FLFLAGS;
 156
 157                 /* lockless truncate
 158                  *
 159                  * 1. do not use inode's timestamps because concurrent
 160                  * stat might fill the inode with out-of-date times,
 161                  * send current instead
 162                  *
 163                  * 2.do no update lsm, as long as stat (via
 164                  * ll_glimpse_size) will bring attributes from osts
 165                  * anyway */
 166                 oa.o_mtime = oa.o_ctime = LTIME_S(CURRENT_TIME);
 167                 oa.o_valid |= OBD_MD_FLMTIME | OBD_MD_FLCTIME;
 168         } else {
 169                 struct ost_lvb *xtimes;
 170                 /* truncate under locks
 171                  *
 172                  * 1. update inode's mtime and ctime as long as
 173                  * concurrent stat (via ll_glimpse_size) might bring
 174                  * out-of-date ones
 175                  *
 176                  * 2. update lsm so that next stat (via
 177                  * ll_glimpse_size) could get correct values in lsm */
 178                 OBD_ALLOC_PTR(xtimes);
 179                 if (NULL == xtimes)
 180                         RETURN(-ENOMEM);
 181
 182                 lov_stripe_lock(lli->lli_smd);
 183                 LTIME_S(inode->i_mtime) = LTIME_S(CURRENT_TIME);
 184                 LTIME_S(inode->i_ctime) = LTIME_S(CURRENT_TIME);
 185                 xtimes->lvb_mtime = LTIME_S(inode->i_mtime);
 186                 xtimes->lvb_ctime = LTIME_S(inode->i_ctime);
 187                 obd_update_lvb(ll_i2obdexp(inode), lli->lli_smd, xtimes,
 188                                OBD_MD_FLMTIME | OBD_MD_FLCTIME);
 189                 lov_stripe_unlock(lli->lli_smd);
 190                 OBD_FREE_PTR(xtimes);
 191
 192                 valid |= OBD_MD_FLMTIME | OBD_MD_FLCTIME;
 193         }
 194         obdo_from_inode(&oa, inode, valid);
 195
 196         rc = obd_punch_rqset(ll_i2obdexp(inode), &oinfo, NULL);
 197         if (rc) {
 198                 CERROR("obd_truncate fails (%d) ino %lu\n", rc, inode->i_ino);
 199                 RETURN(rc);
 200         }
 201         obdo_to_inode(inode, &oa, OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
 202                       OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME);
 203         RETURN(0);
 204 }
 205 /* this isn't where truncate starts.   roughly:
 206  * sys_truncate->ll_setattr_raw->vmtruncate->ll_truncate. setattr_raw grabs
 207  * DLM lock on [size, EOF], i_mutex, ->lli_size_sem, and WRITE_I_ALLOC_SEM to
 208  * avoid races.
 209  *
 210  * must be called under ->lli_size_sem */
 211 void ll_truncate(struct inode *inode)
 212 {
 213         struct ll_inode_info *lli = ll_i2info(inode);
 214         int srvlock = test_bit(LLI_F_SRVLOCK, &lli->lli_flags);
 215         loff_t new_size;
 216         ENTRY;
 217         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p) to %Lu=%#Lx\n",
 218                inode->i_ino, inode->i_generation, inode, i_size_read(inode),
 219                i_size_read(inode));
 220
 221         ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_TRUNC, 1);
 222         if (lli->lli_size_sem_owner != current) {
 223                 EXIT;
 224                 return;
 225         }
 226
 227         if (!lli->lli_smd) {
 228                 CDEBUG(D_INODE, "truncate on inode %lu with no objects\n",
 229                        inode->i_ino);
 230                 GOTO(out_unlock, 0);
 231         }
 232
 233         LASSERT(SEM_COUNT(&lli->lli_size_sem) <= 0);
 234
 235         if (!srvlock) {
 236                 struct ost_lvb lvb;
 237                 int rc;
 238
 239                 /* XXX I'm pretty sure this is a hack to paper over a more
 240                  * fundamental race condition. */
 241                 lov_stripe_lock(lli->lli_smd);
 242                 inode_init_lvb(inode, &lvb);
 243                 rc = obd_merge_lvb(ll_i2obdexp(inode), lli->lli_smd, &lvb, 0);
 244                 inode->i_blocks = lvb.lvb_blocks;
 245                 if (lvb.lvb_size == i_size_read(inode) && rc == 0) {
 246                         CDEBUG(D_VFSTRACE, "skipping punch for obj "LPX64
 247                                ", %Lu=%#Lx\n",
 248                                lli->lli_smd->lsm_object_id, i_size_read(inode),
 249                                i_size_read(inode));
 250                         lov_stripe_unlock(lli->lli_smd);
 251                         GOTO(out_unlock, 0);
 252                 }
 253
 254                 obd_adjust_kms(ll_i2obdexp(inode), lli->lli_smd,
 255                                i_size_read(inode), 1);
 256                 lov_stripe_unlock(lli->lli_smd);
 257         }
 258
 259         if (unlikely((ll_i2sbi(inode)->ll_flags & LL_SBI_LLITE_CHECKSUM) &&
 260                      (i_size_read(inode) & ~CFS_PAGE_MASK))) {
 261                 /* If the truncate leaves a partial page, update its checksum */
 262                 struct page *page = find_get_page(inode->i_mapping,
 263                                                   i_size_read(inode) >>
 264                                                   CFS_PAGE_SHIFT);
 265                 if (page != NULL) {
 266                         struct ll_async_page *llap = llap_cast_private(page);
 267                         if (llap != NULL) {
 268                                 char *kaddr = kmap_atomic(page, KM_USER0);
 269                                 llap->llap_checksum =
 270                                         init_checksum(OSC_DEFAULT_CKSUM);
 271                                 llap->llap_checksum =
 272                                         compute_checksum(llap->llap_checksum,
 273                                                          kaddr, CFS_PAGE_SIZE,
 274                                                          OSC_DEFAULT_CKSUM);
 275                                 kunmap_atomic(kaddr, KM_USER0);
 276                         }
 277                         page_cache_release(page);
 278                 }
 279         }
 280
 281         new_size = i_size_read(inode);
 282         ll_inode_size_unlock(inode, 0);
 283         if (!srvlock)
 284                 ll_file_punch(inode, new_size, 0);
 285         else
 286                 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_LOCKLESS_TRUNC, 1);
 287
 288         EXIT;
 289         return;
 290
 291  out_unlock:
 292         ll_inode_size_unlock(inode, 0);
 293 } /* ll_truncate */
 294
 295 int ll_prepare_write(struct file *file, struct page *page, unsigned from,
 296                      unsigned to)
 297 {
 298         struct inode *inode = page->mapping->host;
 299         struct ll_inode_info *lli = ll_i2info(inode);
 300         struct lov_stripe_md *lsm = lli->lli_smd;
 301         obd_off offset = ((obd_off)page->index) << CFS_PAGE_SHIFT;
 302         struct obd_info oinfo = { { { 0 } } };
 303         struct brw_page pga;
 304         struct obdo oa = { 0 };
 305         struct ost_lvb lvb;
 306         int rc = 0;
 307         ENTRY;
 308
 309         LASSERT(PageLocked(page));
 310         (void)llap_cast_private(page); /* assertion */
 311
 312         /* Check to see if we should return -EIO right away */
 313         pga.pg = page;
 314         pga.off = offset;
 315         pga.count = CFS_PAGE_SIZE;
 316         pga.flag = 0;
 317
 318         oa.o_mode = inode->i_mode;
 319         oa.o_id = lsm->lsm_object_id;
 320         oa.o_gr = lsm->lsm_object_gr;
 321         oa.o_valid = OBD_MD_FLID   | OBD_MD_FLMODE |
 322                      OBD_MD_FLTYPE | OBD_MD_FLGROUP;
 323         obdo_from_inode(&oa, inode, OBD_MD_FLFID | OBD_MD_FLGENER);
 324
 325         oinfo.oi_oa = &oa;
 326         oinfo.oi_md = lsm;
 327         rc = obd_brw(OBD_BRW_CHECK, ll_i2obdexp(inode), &oinfo, 1, &pga, NULL);
 328         if (rc)
 329                 RETURN(rc);
 330
 331         if (PageUptodate(page)) {
 332                 LL_CDEBUG_PAGE(D_PAGE, page, "uptodate\n");
 333                 RETURN(0);
 334         }
 335
 336         /* We're completely overwriting an existing page, so _don't_ set it up
 337          * to date until commit_write */
 338         if (from == 0 && to == CFS_PAGE_SIZE) {
 339                 LL_CDEBUG_PAGE(D_PAGE, page, "full page write\n");
 340                 POISON_PAGE(page, 0x11);
 341                 RETURN(0);
 342         }
 343
 344         /* If are writing to a new page, no need to read old data.  The extent
 345          * locking will have updated the KMS, and for our purposes here we can
 346          * treat it like i_size. */
 347         lov_stripe_lock(lsm);
 348         inode_init_lvb(inode, &lvb);
 349         obd_merge_lvb(ll_i2obdexp(inode), lsm, &lvb, 1);
 350         lov_stripe_unlock(lsm);
 351         if (lvb.lvb_size <= offset) {
 352                 char *kaddr = kmap_atomic(page, KM_USER0);
 353                 LL_CDEBUG_PAGE(D_PAGE, page, "kms "LPU64" <= offset "LPU64"\n",
 354                                lvb.lvb_size, offset);
 355                 memset(kaddr, 0, CFS_PAGE_SIZE);
 356                 kunmap_atomic(kaddr, KM_USER0);
 357                 GOTO(prepare_done, rc = 0);
 358         }
 359
 360         /* XXX could be an async ocp read.. read-ahead? */
 361         rc = ll_brw(OBD_BRW_READ, inode, &oa, page, 0);
 362         if (rc == 0) {
 363                 /* bug 1598: don't clobber blksize */
 364                 oa.o_valid &= ~(OBD_MD_FLSIZE | OBD_MD_FLBLKSZ);
 365                 obdo_refresh_inode(inode, &oa, oa.o_valid);
 366         }
 367
 368         EXIT;
 369  prepare_done:
 370         if (rc == 0)
 371                 SetPageUptodate(page);
 372
 373         return rc;
 374 }
 375
 376 /**
 377  * make page ready for ASYNC write
 378  * \param data - pointer to llap cookie
 379  * \param cmd - is OBD_BRW_* macroses
 380  *
 381  * \retval 0 is page successfully prepared to send
 382  * \retval -EAGAIN is page not need to send
 383  */
 384 static int ll_ap_make_ready(void *data, int cmd)
 385 {
 386         struct ll_async_page *llap;
 387         struct page *page;
 388         ENTRY;
 389
 390         llap = LLAP_FROM_COOKIE(data);
 391         page = llap->llap_page;
 392
 393         /* we're trying to write, but the page is locked.. come back later */
 394         if (TryLockPage(page))
 395                 RETURN(-EAGAIN);
 396
 397         LASSERTF(!(cmd & OBD_BRW_READ) || !PageWriteback(page),
 398                 "cmd %x page %p ino %lu index %lu fl %lx\n", cmd, page,
 399                  page->mapping->host->i_ino, page->index, page->flags);
 400
 401         /* if we left PageDirty we might get another writepage call
 402          * in the future.  list walkers are bright enough
 403          * to check page dirty so we can leave it on whatever list
 404          * its on.  XXX also, we're called with the cli list so if
 405          * we got the page cache list we'd create a lock inversion
 406          * with the removepage path which gets the page lock then the
 407          * cli lock */
 408         if(!clear_page_dirty_for_io(page)) {
 409                 unlock_page(page);
 410                 RETURN(-EAGAIN);
 411         }
 412
 413         /* This actually clears the dirty bit in the radix tree.*/
 414         set_page_writeback(page);
 415
 416         LL_CDEBUG_PAGE(D_PAGE, page, "made ready\n");
 417         page_cache_get(page);
 418
 419         RETURN(0);
 420 }
 421
 422 /* We have two reasons for giving llite the opportunity to change the
 423  * write length of a given queued page as it builds the RPC containing
 424  * the page:
 425  *
 426  * 1) Further extending writes may have landed in the page cache
 427  *    since a partial write first queued this page requiring us
 428  *    to write more from the page cache.  (No further races are possible,
 429  *    since by the time this is called, the page is locked.)
 430  * 2) We might have raced with truncate and want to avoid performing
 431  *    write RPCs that are just going to be thrown away by the
 432  *    truncate's punch on the storage targets.
 433  *
 434  * The kms serves these purposes as it is set at both truncate and extending
 435  * writes.
 436  */
 437 static int ll_ap_refresh_count(void *data, int cmd)
 438 {
 439         struct ll_inode_info *lli;
 440         struct ll_async_page *llap;
 441         struct lov_stripe_md *lsm;
 442         struct page *page;
 443         struct inode *inode;
 444         struct ost_lvb lvb;
 445         __u64 kms;
 446         ENTRY;
 447
 448         /* readpage queues with _COUNT_STABLE, shouldn't get here. */
 449         LASSERT(cmd != OBD_BRW_READ);
 450
 451         llap = LLAP_FROM_COOKIE(data);
 452         page = llap->llap_page;
 453         inode = page->mapping->host;
 454         lli = ll_i2info(inode);
 455         lsm = lli->lli_smd;
 456
 457         lov_stripe_lock(lsm);
 458         inode_init_lvb(inode, &lvb);
 459         obd_merge_lvb(ll_i2obdexp(inode), lsm, &lvb, 1);
 460         kms = lvb.lvb_size;
 461         lov_stripe_unlock(lsm);
 462
 463         /* catch race with truncate */
 464         if (((__u64)page->index << CFS_PAGE_SHIFT) >= kms)
 465                 return 0;
 466
 467         /* catch sub-page write at end of file */
 468         if (((__u64)page->index << CFS_PAGE_SHIFT) + CFS_PAGE_SIZE > kms)
 469                 return kms % CFS_PAGE_SIZE;
 470
 471         return CFS_PAGE_SIZE;
 472 }
 473
 474 void ll_inode_fill_obdo(struct inode *inode, int cmd, struct obdo *oa)
 475 {
 476         struct lov_stripe_md *lsm;
 477         obd_flag valid_flags;
 478
 479         lsm = ll_i2info(inode)->lli_smd;
 480
 481         oa->o_id = lsm->lsm_object_id;
 482         oa->o_gr = lsm->lsm_object_gr;
 483         oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP;
 484         valid_flags = OBD_MD_FLTYPE | OBD_MD_FLATIME;
 485         if (cmd & OBD_BRW_WRITE) {
 486                 oa->o_valid |= OBD_MD_FLEPOCH;
 487                 oa->o_easize = ll_i2info(inode)->lli_io_epoch;
 488
 489                 valid_flags |= OBD_MD_FLMTIME | OBD_MD_FLCTIME |
 490                         OBD_MD_FLUID | OBD_MD_FLGID |
 491                         OBD_MD_FLFID | OBD_MD_FLGENER;
 492         }
 493
 494         obdo_from_inode(oa, inode, valid_flags);
 495 }
 496
 497 static void ll_ap_fill_obdo(void *data, int cmd, struct obdo *oa)
 498 {
 499         struct ll_async_page *llap;
 500         ENTRY;
 501
 502         llap = LLAP_FROM_COOKIE(data);
 503         ll_inode_fill_obdo(llap->llap_page->mapping->host, cmd, oa);
 504
 505         EXIT;
 506 }
 507
 508 static void ll_ap_update_obdo(void *data, int cmd, struct obdo *oa,
 509                               obd_valid valid)
 510 {
 511         struct ll_async_page *llap;
 512         ENTRY;
 513
 514         llap = LLAP_FROM_COOKIE(data);
 515         obdo_from_inode(oa, llap->llap_page->mapping->host, valid);
 516
 517         EXIT;
 518 }
 519
 520 static struct obd_async_page_ops ll_async_page_ops = {
 521         .ap_make_ready =        ll_ap_make_ready,
 522         .ap_refresh_count =     ll_ap_refresh_count,
 523         .ap_fill_obdo =         ll_ap_fill_obdo,
 524         .ap_update_obdo =       ll_ap_update_obdo,
 525         .ap_completion =        ll_ap_completion,
 526 };
 527
 528 struct ll_async_page *llap_cast_private(struct page *page)
 529 {
 530         struct ll_async_page *llap = (struct ll_async_page *)page_private(page);
 531
 532         LASSERTF(llap == NULL || llap->llap_magic == LLAP_MAGIC,
 533                  "page %p private %lu gave magic %d which != %d\n",
 534                  page, page_private(page), llap->llap_magic, LLAP_MAGIC);
 535
 536         return llap;
 537 }
 538
 539 /* Try to reap @target pages in the specific @cpu's async page list.
 540  *
 541  * There is an llap attached onto every page in lustre, linked off @sbi.
 542  * We add an llap to the list so we don't lose our place during list walking.
 543  * If llaps in the list are being moved they will only move to the end
 544  * of the LRU, and we aren't terribly interested in those pages here (we
 545  * start at the beginning of the list where the least-used llaps are. */
 546 static inline int llap_shrink_cache_internal(struct ll_sb_info *sbi,
 547         int cpu, int target)
 548 {
 549         struct ll_async_page *llap, dummy_llap = { .llap_magic = 0xd11ad11a };
 550         struct ll_pglist_data *pd;
 551         struct list_head *head;
 552         int count = 0;
 553
 554         pd = ll_pglist_cpu_lock(sbi, cpu);
 555         head = &pd->llpd_list;
 556         list_add(&dummy_llap.llap_pglist_item, head);
 557         while (count < target) {
 558                 struct page *page;
 559                 int keep;
 560
 561                 if (unlikely(need_resched())) {
 562                         list_del(&dummy_llap.llap_pglist_item);
 563                         ll_pglist_cpu_unlock(sbi, cpu);
 564                         /* vmscan::shrink_slab() have own schedule() */
 565                         return count;
 566                 }
 567
 568                 llap = llite_pglist_next_llap(head,
 569                         &dummy_llap.llap_pglist_item);
 570                 list_del_init(&dummy_llap.llap_pglist_item);
 571                 if (llap == NULL)
 572                         break;
 573
 574                 page = llap->llap_page;
 575                 LASSERT(page != NULL);
 576
 577                 list_add(&dummy_llap.llap_pglist_item, &llap->llap_pglist_item);
 578
 579                 /* Page needs/undergoing IO */
 580                 if (TryLockPage(page)) {
 581                         LL_CDEBUG_PAGE(D_PAGE, page, "can't lock\n");
 582                         continue;
 583                 }
 584
 585                keep = (llap->llap_write_queued || PageDirty(page) ||
 586                       PageWriteback(page) || (!PageUptodate(page) &&
 587                       llap->llap_origin != LLAP_ORIGIN_READAHEAD));
 588
 589                 LL_CDEBUG_PAGE(D_PAGE, page,
 590                                "%s LRU page: %s%s%s%s%s origin %s\n",
 591                                keep ? "keep" : "drop",
 592                                llap->llap_write_queued ? "wq " : "",
 593                                PageDirty(page) ? "pd " : "",
 594                                PageUptodate(page) ? "" : "!pu ",
 595                                PageWriteback(page) ? "wb" : "",
 596                                llap->llap_defer_uptodate ? "" : "!du",
 597                                llap_origins[llap->llap_origin]);
 598
 599                 /* If page is dirty or undergoing IO don't discard it */
 600                 if (keep) {
 601                         unlock_page(page);
 602                         continue;
 603                 }
 604
 605                 page_cache_get(page);
 606                 ll_pglist_cpu_unlock(sbi, cpu);
 607
 608                 if (page->mapping != NULL) {
 609                         ll_teardown_mmaps(page->mapping,
 610                                          (__u64)page->index << CFS_PAGE_SHIFT,
 611                                          ((__u64)page->index << CFS_PAGE_SHIFT)|
 612                                           ~CFS_PAGE_MASK);
 613                         if (!PageDirty(page) && !page_mapped(page)) {
 614                                 ll_ra_accounting(llap, page->mapping);
 615                                 ll_truncate_complete_page(page);
 616                                 ++count;
 617                         } else {
 618                                 LL_CDEBUG_PAGE(D_PAGE, page,
 619                                                "Not dropping page because it is"
 620                                                " %s\n", PageDirty(page) ?
 621                                                "dirty" : "mapped");
 622                         }
 623                 }
 624                 unlock_page(page);
 625                 page_cache_release(page);
 626
 627                 ll_pglist_cpu_lock(sbi, cpu);
 628         }
 629         list_del(&dummy_llap.llap_pglist_item);
 630         ll_pglist_cpu_unlock(sbi, cpu);
 631
 632         CDEBUG(D_CACHE, "shrank %d, expected %d however. \n", count, target);
 633         return count;
 634 }
 635
 636
 637 /* Try to shrink the page cache for the @sbi filesystem by 1/@shrink_fraction.
 638  *
 639  * At first, this code calculates total pages wanted by @shrink_fraction, then
 640  * it deduces how many pages should be reaped from each cpu in proportion as
 641  * their own # of page count(llpd_count).
 642  */
 643 int llap_shrink_cache(struct ll_sb_info *sbi, int shrink_fraction)
 644 {
 645         unsigned long total, want, percpu_want, count = 0;
 646         int cpu, nr_cpus;
 647
 648         total = lcounter_read_positive(&sbi->ll_async_page_count);
 649         if (total == 0)
 650                 return 0;
 651
 652 #ifdef HAVE_SHRINKER_CACHE
 653         want = shrink_fraction;
 654         if (want == 0)
 655                 return total;
 656 #else
 657         /* There can be a large number of llaps (600k or more in a large
 658          * memory machine) so the VM 1/6 shrink ratio is likely too much.
 659          * Since we are freeing pages also, we don't necessarily want to
 660          * shrink so much.  Limit to 40MB of pages + llaps per call. */
 661         if (shrink_fraction <= 0)
 662                 want = total - sbi->ll_async_page_max + 32*num_online_cpus();
 663         else
 664                 want = (total + shrink_fraction - 1) / shrink_fraction;
 665 #endif
 666
 667         if (want > 40 << (20 - CFS_PAGE_SHIFT))
 668                 want = 40 << (20 - CFS_PAGE_SHIFT);
 669
 670         CDEBUG(D_CACHE, "shrinking %lu of %lu pages (1/%d)\n",
 671                want, total, shrink_fraction);
 672
 673         nr_cpus = num_possible_cpus();
 674         cpu = sbi->ll_async_page_clock_hand;
 675         /* we at most do one round */
 676         do {
 677                 int c;
 678
 679                 cpu = (cpu + 1) % nr_cpus;
 680                 c = LL_PGLIST_DATA_CPU(sbi, cpu)->llpd_count;
 681                 if (!cpu_online(cpu))
 682                         percpu_want = c;
 683                 else
 684                         percpu_want = want / ((total / (c + 1)) + 1);
 685                 if (percpu_want == 0)
 686                         continue;
 687
 688                 count += llap_shrink_cache_internal(sbi, cpu, percpu_want);
 689                 if (count >= want)
 690                         sbi->ll_async_page_clock_hand = cpu;
 691         } while (cpu != sbi->ll_async_page_clock_hand);
 692
 693         CDEBUG(D_CACHE, "shrank %lu/%lu and left %lu unscanned\n",
 694                count, want, total);
 695
 696 #ifdef HAVE_SHRINKER_CACHE
 697         return lcounter_read_positive(&sbi->ll_async_page_count);
 698 #else
 699         return count;
 700 #endif
 701 }
 702
 703 /* Rebalance the async page queue len for each cpu. We hope that the cpu
 704  * which do much IO job has a relative longer queue len.
 705  * This function should be called with preempt disabled.
 706  */
 707 static inline int llap_async_cache_rebalance(struct ll_sb_info *sbi)
 708 {
 709         unsigned long sample = 0, *cpu_sample, bias, slice;
 710         struct ll_pglist_data *pd;
 711         cpumask_t mask;
 712         int cpu, surplus;
 713         int w1 = 7, w2 = 3, base = (w1 + w2); /* weight value */
 714         atomic_t *pcnt;
 715
 716         if (!spin_trylock(&sbi->ll_async_page_reblnc_lock)) {
 717                 /* someone else is doing the job */
 718                 return 1;
 719         }
 720
 721         pcnt = &LL_PGLIST_DATA(sbi)->llpd_sample_count;
 722         if (!atomic_read(pcnt)) {
 723                 /* rare case, somebody else has gotten this job done */
 724                 spin_unlock(&sbi->ll_async_page_reblnc_lock);
 725                 return 1;
 726         }
 727
 728         sbi->ll_async_page_reblnc_count++;
 729         cpu_sample = sbi->ll_async_page_sample;
 730         memset(cpu_sample, 0, num_possible_cpus() * sizeof(unsigned long));
 731         for_each_online_cpu(cpu) {
 732                 pcnt = &LL_PGLIST_DATA_CPU(sbi, cpu)->llpd_sample_count;
 733                 cpu_sample[cpu] = atomic_read(pcnt);
 734                 atomic_set(pcnt, 0);
 735                 sample += cpu_sample[cpu];
 736         }
 737
 738         cpus_clear(mask);
 739         surplus = sbi->ll_async_page_max;
 740         slice = surplus / sample + 1;
 741         sample /= num_online_cpus();
 742         bias = sample >> 4;
 743         for_each_online_cpu(cpu) {
 744                 pd = LL_PGLIST_DATA_CPU(sbi, cpu);
 745                 if (labs((long int)sample - cpu_sample[cpu]) > bias) {
 746                         unsigned long budget = pd->llpd_budget;
 747                         /* weighted original queue length and expected queue
 748                          * length to avoid thrashing. */
 749                         pd->llpd_budget = (budget * w1) / base +
 750                                         (slice * cpu_sample[cpu]) * w2 / base;
 751                         cpu_set(cpu, mask);
 752                 }
 753                 surplus -= min_t(int, pd->llpd_budget, surplus);
 754         }
 755         surplus /= cpus_weight(mask) ?: 1;
 756         for_each_cpu_mask(cpu, mask)
 757                 LL_PGLIST_DATA_CPU(sbi, cpu)->llpd_budget += surplus;
 758         spin_unlock(&sbi->ll_async_page_reblnc_lock);
 759
 760         /* We need to call llap_shrink_cache_internal() for every cpu to
 761          * ensure the sbi->ll_async_page_max limit is enforced. */
 762         for_each_cpu_mask(cpu, mask) {
 763                 pd = LL_PGLIST_DATA_CPU(sbi, cpu);
 764                 llap_shrink_cache_internal(sbi, cpu, max_t(int, pd->llpd_count-
 765                                            pd->llpd_budget, 0) + 32);
 766         }
 767
 768         return 0;
 769 }
 770
 771 static struct ll_async_page *llap_from_page_with_lockh(struct page *page,
 772                                                        unsigned origin,
 773                                                        struct lustre_handle *lockh)
 774 {
 775         struct ll_async_page *llap;
 776         struct obd_export *exp;
 777         struct inode *inode = page->mapping->host;
 778         struct ll_sb_info *sbi;
 779         struct ll_pglist_data *pd;
 780         int rc, cpu, target;
 781         ENTRY;
 782
 783         if (!inode) {
 784                 static int triggered;
 785
 786                 if (!triggered) {
 787                         LL_CDEBUG_PAGE(D_ERROR, page, "Bug 10047. Wrong anon "
 788                                        "page received\n");
 789                         libcfs_debug_dumpstack(NULL);
 790                         triggered = 1;
 791                 }
 792                 RETURN(ERR_PTR(-EINVAL));
 793         }
 794         sbi = ll_i2sbi(inode);
 795         LASSERT(ll_async_page_slab);
 796         LASSERTF(origin < LLAP__ORIGIN_MAX, "%u\n", origin);
 797
 798         llap = llap_cast_private(page);
 799         if (llap != NULL) {
 800                 /* move to end of LRU list, except when page is just about to
 801                  * die */
 802                 if (origin != LLAP_ORIGIN_REMOVEPAGE) {
 803                         int old_cpu = llap->llap_pglist_cpu;
 804                         struct ll_pglist_data *old_pd;
 805
 806                         pd = ll_pglist_double_lock(sbi, old_cpu, &old_pd);
 807                         pd->llpd_hit++;
 808                         while (old_cpu != llap->llap_pglist_cpu) {
 809                                 /* rarely case, someone else is touching this
 810                                  * page too. */
 811                                 ll_pglist_double_unlock(sbi, old_cpu);
 812                                 old_cpu = llap->llap_pglist_cpu;
 813                                 pd=ll_pglist_double_lock(sbi, old_cpu, &old_pd);
 814                         }
 815
 816                         list_move(&llap->llap_pglist_item,
 817                                   &pd->llpd_list);
 818                         old_pd->llpd_gen++;
 819                         if (pd->llpd_cpu != old_cpu) {
 820                                 pd->llpd_count++;
 821                                 old_pd->llpd_count--;
 822                                 old_pd->llpd_gen++;
 823                                 llap->llap_pglist_cpu = pd->llpd_cpu;
 824                                 pd->llpd_cross++;
 825                         }
 826                         ll_pglist_double_unlock(sbi, old_cpu);
 827                 }
 828                 GOTO(out, llap);
 829         }
 830
 831         exp = ll_i2obdexp(page->mapping->host);
 832         if (exp == NULL)
 833                 RETURN(ERR_PTR(-EINVAL));
 834
 835         /* limit the number of lustre-cached pages */
 836         cpu = cfs_get_cpu();
 837         pd = LL_PGLIST_DATA(sbi);
 838         target = pd->llpd_count - pd->llpd_budget;
 839         if (target > 0) {
 840                 rc = 0;
 841                 atomic_inc(&pd->llpd_sample_count);
 842                 if (atomic_read(&pd->llpd_sample_count) >
 843                     sbi->ll_async_page_sample_max) {
 844                         pd->llpd_reblnc_count++;
 845                         rc = llap_async_cache_rebalance(sbi);
 846                         if (rc == 0)
 847                                 target = pd->llpd_count - pd->llpd_budget;
 848                 }
 849                 /* if rc equals 1, it means other cpu is doing the rebalance
 850                  * job, and our budget # would be modified when we read it.
 851                  * Furthermore, it is much likely being increased because
 852                  * we have already reached the rebalance threshold. In this
 853                  * case, we skip to shrink cache here. */
 854                 if ((rc == 0) && target > 0)
 855                         llap_shrink_cache_internal(sbi, cpu, target + 32);
 856         }
 857         cfs_put_cpu();
 858
 859         OBD_SLAB_ALLOC(llap, ll_async_page_slab, CFS_ALLOC_STD,
 860                        ll_async_page_slab_size);
 861         if (llap == NULL)
 862                 RETURN(ERR_PTR(-ENOMEM));
 863         llap->llap_magic = LLAP_MAGIC;
 864         llap->llap_cookie = (void *)llap + size_round(sizeof(*llap));
 865
 866         /* XXX: for bug 11270 - check for lockless origin here! */
 867         if (origin == LLAP_ORIGIN_LOCKLESS_IO)
 868                 llap->llap_nocache = 1;
 869
 870         rc = obd_prep_async_page(exp, ll_i2info(inode)->lli_smd, NULL, page,
 871                                  (obd_off)page->index << CFS_PAGE_SHIFT,
 872                                  &ll_async_page_ops, llap, &llap->llap_cookie,
 873                                  llap->llap_nocache, lockh);
 874         if (rc) {
 875                 OBD_SLAB_FREE(llap, ll_async_page_slab,
 876                               ll_async_page_slab_size);
 877                 RETURN(ERR_PTR(rc));
 878         }
 879
 880         CDEBUG(D_CACHE, "llap %p page %p cookie %p obj off "LPU64"\n", llap,
 881                page, llap->llap_cookie, (obd_off)page->index << CFS_PAGE_SHIFT);
 882         /* also zeroing the PRIVBITS low order bitflags */
 883         __set_page_ll_data(page, llap);
 884         llap->llap_page = page;
 885
 886         lcounter_inc(&sbi->ll_async_page_count);
 887         pd = ll_pglist_lock(sbi);
 888         list_add_tail(&llap->llap_pglist_item, &pd->llpd_list);
 889         pd->llpd_count++;
 890         pd->llpd_gen++;
 891         pd->llpd_miss++;
 892         llap->llap_pglist_cpu = pd->llpd_cpu;
 893         ll_pglist_unlock(sbi);
 894
 895  out:
 896         if (unlikely(sbi->ll_flags & LL_SBI_LLITE_CHECKSUM)) {
 897                 __u32 csum;
 898                 char *kaddr = kmap_atomic(page, KM_USER0);
 899                 csum = init_checksum(OSC_DEFAULT_CKSUM);
 900                 csum = compute_checksum(csum, kaddr, CFS_PAGE_SIZE,
 901                                         OSC_DEFAULT_CKSUM);
 902                 kunmap_atomic(kaddr, KM_USER0);
 903                 if (origin == LLAP_ORIGIN_READAHEAD ||
 904                     origin == LLAP_ORIGIN_READPAGE ||
 905                     origin == LLAP_ORIGIN_LOCKLESS_IO) {
 906                         llap->llap_checksum = 0;
 907                 } else if (origin == LLAP_ORIGIN_COMMIT_WRITE ||
 908                            llap->llap_checksum == 0) {
 909                         llap->llap_checksum = csum;
 910                         CDEBUG(D_PAGE, "page %p cksum %x\n", page, csum);
 911                 } else if (llap->llap_checksum == csum) {
 912                         /* origin == LLAP_ORIGIN_WRITEPAGE */
 913                         CDEBUG(D_PAGE, "page %p cksum %x confirmed\n",
 914                                page, csum);
 915                 } else {
 916                         /* origin == LLAP_ORIGIN_WRITEPAGE */
 917                         LL_CDEBUG_PAGE(D_ERROR, page, "old cksum %x != new "
 918                                        "%x!\n", llap->llap_checksum, csum);
 919                 }
 920         }
 921
 922         llap->llap_origin = origin;
 923         RETURN(llap);
 924 }
 925
 926 static inline struct ll_async_page *llap_from_page(struct page *page,
 927                                                    unsigned origin)
 928 {
 929         return llap_from_page_with_lockh(page, origin, NULL);
 930 }
 931
 932 static int queue_or_sync_write(struct obd_export *exp, struct inode *inode,
 933                                struct ll_async_page *llap,
 934                                unsigned to, obd_flag async_flags)
 935 {
 936         unsigned long size_index = i_size_read(inode) >> CFS_PAGE_SHIFT;
 937         struct obd_io_group *oig;
 938         struct ll_sb_info *sbi = ll_i2sbi(inode);
 939         int rc, noquot = llap->llap_ignore_quota ? OBD_BRW_NOQUOTA : 0;
 940         int brwflags = OBD_BRW_ASYNC;
 941         ENTRY;
 942
 943         /* _make_ready only sees llap once we've unlocked the page */
 944         llap->llap_write_queued = 1;
 945         rc = obd_queue_async_io(exp, ll_i2info(inode)->lli_smd, NULL,
 946                                 llap->llap_cookie, OBD_BRW_WRITE | noquot,
 947                                 0, 0, brwflags, async_flags);
 948         if (rc == 0) {
 949                 LL_CDEBUG_PAGE(D_PAGE, llap->llap_page, "write queued\n");
 950                 llap_write_pending(inode, llap);
 951                 GOTO(out, 0);
 952         }
 953
 954         llap->llap_write_queued = 0;
 955
 956         rc = oig_init(&oig);
 957         if (rc)
 958                 GOTO(out, rc);
 959
 960         /* make full-page requests if we are not at EOF (bug 4410) */
 961         if (to != CFS_PAGE_SIZE && llap->llap_page->index < size_index) {
 962                 LL_CDEBUG_PAGE(D_PAGE, llap->llap_page,
 963                                "sync write before EOF: size_index %lu, to %d\n",
 964                                size_index, to);
 965                 to = CFS_PAGE_SIZE;
 966         } else if (to != CFS_PAGE_SIZE && llap->llap_page->index == size_index){
 967                 int size_to = i_size_read(inode) & ~CFS_PAGE_MASK;
 968                 LL_CDEBUG_PAGE(D_PAGE, llap->llap_page,
 969                                "sync write at EOF: size_index %lu, to %d/%d\n",
 970                                size_index, to, size_to);
 971                 if (to < size_to)
 972                         to = size_to;
 973         }
 974
 975         /* compare the checksum once before the page leaves llite */
 976         if (unlikely((sbi->ll_flags & LL_SBI_LLITE_CHECKSUM) &&
 977                      llap->llap_checksum != 0)) {
 978                 __u32 csum;
 979                 struct page *page = llap->llap_page;
 980                 char *kaddr = kmap_atomic(page, KM_USER0);
 981                 csum = init_checksum(OSC_DEFAULT_CKSUM);
 982                 csum = compute_checksum(csum, kaddr, CFS_PAGE_SIZE,
 983                                         OSC_DEFAULT_CKSUM);
 984                 kunmap_atomic(kaddr, KM_USER0);
 985                 if (llap->llap_checksum == csum) {
 986                         CDEBUG(D_PAGE, "page %p cksum %x confirmed\n",
 987                                page, csum);
 988                 } else {
 989                         CERROR("page %p old cksum %x != new cksum %x!\n",
 990                                page, llap->llap_checksum, csum);
 991                 }
 992         }
 993
 994         rc = obd_queue_group_io(exp, ll_i2info(inode)->lli_smd, NULL, oig,
 995                                 llap->llap_cookie, OBD_BRW_WRITE | noquot,
 996                                 0, to, 0, ASYNC_READY | ASYNC_URGENT |
 997                                 ASYNC_COUNT_STABLE | ASYNC_GROUP_SYNC);
 998         if (rc)
 999                 GOTO(free_oig, rc);
1000
1001         rc = obd_trigger_group_io(exp, ll_i2info(inode)->lli_smd, NULL, oig);
1002         if (rc)
1003                 GOTO(free_oig, rc);
1004
1005         rc = oig_wait(oig);
1006
1007         if (!rc && async_flags & ASYNC_READY) {
1008                 unlock_page(llap->llap_page);
1009                 if (PageWriteback(llap->llap_page))
1010                         end_page_writeback(llap->llap_page);
1011         }
1012
1013         LL_CDEBUG_PAGE(D_PAGE, llap->llap_page, "sync write returned %d\n", rc);
1014
1015 free_oig:
1016         oig_release(oig);
1017 out:
1018         RETURN(rc);
1019 }
1020
1021 /* update our write count to account for i_size increases that may have
1022  * happened since we've queued the page for io. */
1023
1024 /* be careful not to return success without setting the page Uptodate or
1025  * the next pass through prepare_write will read in stale data from disk. */
1026 int ll_commit_write(struct file *file, struct page *page, unsigned from,
1027                     unsigned to)
1028 {
1029         struct ll_file_data *fd = LUSTRE_FPRIVATE(file);
1030         struct inode *inode = page->mapping->host;
1031         struct ll_inode_info *lli = ll_i2info(inode);
1032         struct lov_stripe_md *lsm = lli->lli_smd;
1033         struct obd_export *exp;
1034         struct ll_async_page *llap;
1035         loff_t size;
1036         struct lustre_handle *lockh = NULL;
1037         int rc = 0;
1038         ENTRY;
1039
1040         SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
1041         LASSERT(inode == file->f_dentry->d_inode);
1042         LASSERT(PageLocked(page));
1043
1044         CDEBUG(D_INODE, "inode %p is writing page %p from %d to %d at %lu\n",
1045                inode, page, from, to, page->index);
1046
1047         if (fd->fd_flags & LL_FILE_GROUP_LOCKED)
1048                 lockh = &fd->fd_cwlockh;
1049
1050         llap = llap_from_page_with_lockh(page, LLAP_ORIGIN_COMMIT_WRITE, lockh);
1051         if (IS_ERR(llap))
1052                 RETURN(PTR_ERR(llap));
1053
1054         exp = ll_i2obdexp(inode);
1055         if (exp == NULL)
1056                 RETURN(-EINVAL);
1057
1058         llap->llap_ignore_quota = cfs_capable(CFS_CAP_SYS_RESOURCE);
1059
1060         /* queue a write for some time in the future the first time we
1061          * dirty the page */
1062         if (!PageDirty(page)) {
1063                 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_DIRTY_MISSES, 1);
1064
1065                 rc = queue_or_sync_write(exp, inode, llap, to, 0);
1066                 if (rc)
1067                         GOTO(out, rc);
1068         } else {
1069                 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_DIRTY_HITS, 1);
1070         }
1071
1072         /* put the page in the page cache, from now on ll_removepage is
1073          * responsible for cleaning up the llap.
1074          * only set page dirty when it's queued to be write out */
1075         if (llap->llap_write_queued)
1076                 set_page_dirty(page);
1077
1078 out:
1079         size = (((obd_off)page->index) << CFS_PAGE_SHIFT) + to;
1080         ll_inode_size_lock(inode, 0);
1081         if (rc == 0) {
1082                 lov_stripe_lock(lsm);
1083                 obd_adjust_kms(exp, lsm, size, 0);
1084                 lov_stripe_unlock(lsm);
1085                 if (size > i_size_read(inode))
1086                         i_size_write(inode, size);
1087                 SetPageUptodate(page);
1088         } else if (size > i_size_read(inode)) {
1089                 /* this page beyond the pales of i_size, so it can't be
1090                  * truncated in ll_p_r_e during lock revoking. we must
1091                  * teardown our book-keeping here. */
1092                 ll_removepage(page);
1093         }
1094         ll_inode_size_unlock(inode, 0);
1095         RETURN(rc);
1096 }
1097
1098 static void ll_ra_stats_inc_sbi(struct ll_sb_info *sbi, enum ra_stat which);
1099
1100 /* WARNING: This algorithm is used to reduce the contention on
1101  * sbi->ll_lock. It should work well if the ra_max_pages is much
1102  * greater than the single file's read-ahead window.
1103  *
1104  * TODO: There may exist a `global sync problem' in this implementation.
1105  * Considering the global ra window is 100M, and each file's ra window is 10M,
1106  * there are over 10 files trying to get its ra budget and reach
1107  * ll_ra_count_get at the exactly same time. All of them will get a zero ra
1108  * window, although the global window is 100M. -jay
1109  */
1110 static unsigned long ll_ra_count_get(struct ll_sb_info *sbi, unsigned long len)
1111 {
1112         struct ll_ra_info *ra = &sbi->ll_ra_info;
1113         unsigned long ret = 0;
1114         ENTRY;
1115
1116         /**
1117          * If read-ahead pages left are less than 1M, do not do read-ahead,
1118          * otherwise it will form small read RPC(< 1M), which hurt server
1119          * performance a lot.
1120          */
1121         ret = min(ra->ra_max_pages - atomic_read(&ra->ra_cur_pages), len);
1122         if ((int)ret < min((unsigned long)PTLRPC_MAX_BRW_PAGES, len))
1123                 GOTO(out, ret = 0);
1124
1125         if (atomic_add_return(ret, &ra->ra_cur_pages) > ra->ra_max_pages) {
1126                 atomic_sub(ret, &ra->ra_cur_pages);
1127                 ret = 0;
1128         }
1129 out:
1130         RETURN(ret);
1131 }
1132
1133 static void ll_ra_count_put(struct ll_sb_info *sbi, unsigned long len)
1134 {
1135         struct ll_ra_info *ra = &sbi->ll_ra_info;
1136         atomic_sub(len, &ra->ra_cur_pages);
1137 }
1138
1139 /* called for each page in a completed rpc.*/
1140 int ll_ap_completion(void *data, int cmd, struct obdo *oa, int rc)
1141 {
1142         struct ll_async_page *llap;
1143         struct page *page;
1144         int ret = 0;
1145         ENTRY;
1146
1147         llap = LLAP_FROM_COOKIE(data);
1148         page = llap->llap_page;
1149         LASSERT(PageLocked(page));
1150         LASSERT(CheckWriteback(page,cmd));
1151
1152         LL_CDEBUG_PAGE(D_PAGE, page, "completing cmd %d with %d\n", cmd, rc);
1153
1154         if (cmd & OBD_BRW_READ && llap->llap_defer_uptodate)
1155                 ll_ra_count_put(ll_i2sbi(page->mapping->host), 1);
1156
1157         if (rc == 0)  {
1158                 if (cmd & OBD_BRW_READ) {
1159                         if (!llap->llap_defer_uptodate)
1160                                 SetPageUptodate(page);
1161                 } else {
1162                         llap->llap_write_queued = 0;
1163                 }
1164                 ClearPageError(page);
1165         } else {
1166                 if (cmd & OBD_BRW_READ) {
1167                         llap->llap_defer_uptodate = 0;
1168                 }
1169                 SetPageError(page);
1170                 if (rc == -ENOSPC)
1171                         set_bit(AS_ENOSPC, &page->mapping->flags);
1172                 else
1173                         set_bit(AS_EIO, &page->mapping->flags);
1174         }
1175
1176         /* be carefull about clear WB.
1177          * if WB will cleared after page lock is released - paralel IO can be
1178          * started before ap_make_ready is finished - so we will be have page
1179          * with PG_Writeback set from ->writepage() and completed READ which
1180          * clear this flag */
1181         if ((cmd & OBD_BRW_WRITE) && PageWriteback(page))
1182                 end_page_writeback(page);
1183
1184         unlock_page(page);
1185
1186         if (cmd & OBD_BRW_WRITE) {
1187                 llap_write_complete(page->mapping->host, llap);
1188                 ll_try_done_writing(page->mapping->host);
1189         }
1190
1191         page_cache_release(page);
1192
1193         RETURN(ret);
1194 }
1195
1196 static void __ll_put_llap(struct page *page)
1197 {
1198         struct inode *inode = page->mapping->host;
1199         struct obd_export *exp;
1200         struct ll_async_page *llap;
1201         struct ll_sb_info *sbi = ll_i2sbi(inode);
1202         struct ll_pglist_data *pd;
1203         int rc, cpu;
1204         ENTRY;
1205
1206         exp = ll_i2obdexp(inode);
1207         if (exp == NULL) {
1208                 CERROR("page %p ind %lu gave null export\n", page, page->index);
1209                 EXIT;
1210                 return;
1211         }
1212
1213         llap = llap_from_page(page, LLAP_ORIGIN_REMOVEPAGE);
1214         if (IS_ERR(llap)) {
1215                 CERROR("page %p ind %lu couldn't find llap: %ld\n", page,
1216                        page->index, PTR_ERR(llap));
1217                 EXIT;
1218                 return;
1219         }
1220
1221         //llap_write_complete(inode, llap);
1222         rc = obd_teardown_async_page(exp, ll_i2info(inode)->lli_smd, NULL,
1223                                      llap->llap_cookie);
1224         if (rc != 0)
1225                 CERROR("page %p ind %lu failed: %d\n", page, page->index, rc);
1226
1227         /* this unconditional free is only safe because the page lock
1228          * is providing exclusivity to memory pressure/truncate/writeback..*/
1229         __clear_page_ll_data(page);
1230
1231         lcounter_dec(&sbi->ll_async_page_count);
1232         cpu = llap->llap_pglist_cpu;
1233         pd = ll_pglist_cpu_lock(sbi, cpu);
1234         pd->llpd_gen++;
1235         pd->llpd_count--;
1236         if (!list_empty(&llap->llap_pglist_item))
1237                 list_del_init(&llap->llap_pglist_item);
1238         ll_pglist_cpu_unlock(sbi, cpu);
1239         OBD_SLAB_FREE(llap, ll_async_page_slab, ll_async_page_slab_size);
1240
1241         EXIT;
1242 }
1243
1244 /* the kernel calls us here when a page is unhashed from the page cache.
1245  * the page will be locked and the kernel is holding a spinlock, so
1246  * we need to be careful.  we're just tearing down our book-keeping
1247  * here. */
1248 void ll_removepage(struct page *page)
1249 {
1250         struct ll_async_page *llap = llap_cast_private(page);
1251         ENTRY;
1252
1253         LASSERT(!in_interrupt());
1254
1255         /* sync pages or failed read pages can leave pages in the page
1256          * cache that don't have our data associated with them anymore */
1257         if (page_private(page) == 0) {
1258                 EXIT;
1259                 return;
1260         }
1261
1262         LASSERT(!llap->llap_lockless_io_page);
1263         LASSERT(!llap->llap_nocache);
1264
1265         LL_CDEBUG_PAGE(D_PAGE, page, "being evicted\n");
1266         __ll_put_llap(page);
1267
1268         EXIT;
1269 }
1270
1271 static int ll_issue_page_read(struct obd_export *exp,
1272                               struct ll_async_page *llap,
1273                               struct obd_io_group *oig, int defer)
1274 {
1275         struct page *page = llap->llap_page;
1276         int rc;
1277
1278         page_cache_get(page);
1279         llap->llap_defer_uptodate = defer;
1280         llap->llap_ra_used = 0;
1281         rc = obd_queue_group_io(exp, ll_i2info(page->mapping->host)->lli_smd,
1282                                 NULL, oig, llap->llap_cookie, OBD_BRW_READ, 0,
1283                                 CFS_PAGE_SIZE, 0, ASYNC_COUNT_STABLE |
1284                                 ASYNC_READY | ASYNC_URGENT);
1285         if (rc) {
1286                 LL_CDEBUG_PAGE(D_ERROR, page, "read queue failed: rc %d\n", rc);
1287                 page_cache_release(page);
1288         }
1289         RETURN(rc);
1290 }
1291
1292 static void ll_ra_stats_inc_sbi(struct ll_sb_info *sbi, enum ra_stat which)
1293 {
1294         LASSERTF(which >= 0 && which < _NR_RA_STAT, "which: %u\n", which);
1295         lprocfs_counter_incr(sbi->ll_ra_stats, which);
1296 }
1297
1298 static void ll_ra_stats_inc(struct address_space *mapping, enum ra_stat which)
1299 {
1300         struct ll_sb_info *sbi = ll_i2sbi(mapping->host);
1301         ll_ra_stats_inc_sbi(sbi, which);
1302 }
1303
1304 void ll_ra_accounting(struct ll_async_page *llap, struct address_space *mapping)
1305 {
1306         if (!llap->llap_defer_uptodate || llap->llap_ra_used)
1307                 return;
1308
1309         ll_ra_stats_inc(mapping, RA_STAT_DISCARDED);
1310 }
1311
1312 #define RAS_CDEBUG(ras) \
1313         CDEBUG(D_READA,                                                      \
1314                "lrp %lu cr %lu cp %lu ws %lu wl %lu nra %lu r %lu ri %lu"    \
1315                "csr %lu sf %lu sp %lu sl %lu \n",                            \
1316                ras->ras_last_readpage, ras->ras_consecutive_requests,        \
1317                ras->ras_consecutive_pages, ras->ras_window_start,            \
1318                ras->ras_window_len, ras->ras_next_readahead,                 \
1319                ras->ras_requests, ras->ras_request_index,                    \
1320                ras->ras_consecutive_stride_requests, ras->ras_stride_offset, \
1321                ras->ras_stride_pages, ras->ras_stride_length)
1322
1323 static int index_in_window(unsigned long index, unsigned long point,
1324                            unsigned long before, unsigned long after)
1325 {
1326         unsigned long start = point - before, end = point + after;
1327
1328         if (start > point)
1329                start = 0;
1330         if (end < point)
1331                end = ~0;
1332
1333         return start <= index && index <= end;
1334 }
1335
1336 static struct ll_readahead_state *ll_ras_get(struct file *f)
1337 {
1338         struct ll_file_data       *fd;
1339
1340         fd = LUSTRE_FPRIVATE(f);
1341         return &fd->fd_ras;
1342 }
1343
1344 void ll_ra_read_init(struct file *f, struct ll_ra_read *rar,
1345                      loff_t offset, size_t count)
1346 {
1347         struct ll_readahead_state *ras;
1348
1349         ras = ll_ras_get(f);
1350
1351         rar->lrr_start = offset >> CFS_PAGE_SHIFT;
1352         rar->lrr_count = (count + CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT;
1353
1354         spin_lock(&ras->ras_lock);
1355         ras->ras_requests++;
1356         ras->ras_request_index = 0;
1357         ras->ras_consecutive_requests++;
1358         rar->lrr_reader = current;
1359
1360         list_add(&rar->lrr_linkage, &ras->ras_read_beads);
1361         spin_unlock(&ras->ras_lock);
1362 }
1363
1364 void ll_ra_read_ex(struct file *f, struct ll_ra_read *rar)
1365 {
1366         struct ll_readahead_state *ras;
1367
1368         ras = ll_ras_get(f);
1369
1370         spin_lock(&ras->ras_lock);
1371         list_del_init(&rar->lrr_linkage);
1372         spin_unlock(&ras->ras_lock);
1373 }
1374
1375 static struct ll_ra_read *ll_ra_read_get_locked(struct ll_readahead_state *ras)
1376 {
1377         struct ll_ra_read *scan;
1378
1379         list_for_each_entry(scan, &ras->ras_read_beads, lrr_linkage) {
1380                 if (scan->lrr_reader == current)
1381                         return scan;
1382         }
1383         return NULL;
1384 }
1385
1386 struct ll_ra_read *ll_ra_read_get(struct file *f)
1387 {
1388         struct ll_readahead_state *ras;
1389         struct ll_ra_read         *bead;
1390
1391         ras = ll_ras_get(f);
1392
1393         spin_lock(&ras->ras_lock);
1394         bead = ll_ra_read_get_locked(ras);
1395         spin_unlock(&ras->ras_lock);
1396         return bead;
1397 }
1398
1399 static int ll_read_ahead_page(struct obd_export *exp, struct obd_io_group *oig,
1400                               pgoff_t index, struct address_space *mapping)
1401 {
1402         struct ll_async_page *llap;
1403         struct page *page;
1404         unsigned int gfp_mask = 0;
1405         int rc = 0;
1406
1407         gfp_mask = GFP_HIGHUSER & ~__GFP_WAIT;
1408 #ifdef __GFP_NOWARN
1409         gfp_mask |= __GFP_NOWARN;
1410 #endif
1411         page = grab_cache_page_nowait_gfp(mapping, index, gfp_mask);
1412         if (page == NULL) {
1413                 ll_ra_stats_inc(mapping, RA_STAT_FAILED_GRAB_PAGE);
1414                 CDEBUG(D_READA, "g_c_p_n failed\n");
1415                 return 0;
1416         }
1417
1418         /* Check if page was truncated or reclaimed */
1419         if (page->mapping != mapping) {
1420                 ll_ra_stats_inc(mapping, RA_STAT_WRONG_GRAB_PAGE);
1421                 CDEBUG(D_READA, "g_c_p_n returned invalid page\n");
1422                 GOTO(unlock_page, rc = 0);
1423         }
1424
1425         /* we do this first so that we can see the page in the /proc
1426          * accounting */
1427         llap = llap_from_page(page, LLAP_ORIGIN_READAHEAD);
1428         if (IS_ERR(llap) || llap->llap_defer_uptodate) {
1429                 if (PTR_ERR(llap) == -ENOLCK) {
1430                         ll_ra_stats_inc(mapping, RA_STAT_FAILED_MATCH);
1431                         CDEBUG(D_READA | D_PAGE,
1432                                "Adding page to cache failed index "
1433                                 "%ld\n", index);
1434                                 CDEBUG(D_READA, "nolock page\n");
1435                                 GOTO(unlock_page, rc = -ENOLCK);
1436                 }
1437                 CDEBUG(D_READA, "read-ahead page\n");
1438                 GOTO(unlock_page, rc = 0);
1439         }
1440
1441         /* skip completed pages */
1442         if (Page_Uptodate(page))
1443                 GOTO(unlock_page, rc = 0);
1444
1445         /* bail out when we hit the end of the lock. */
1446         rc = ll_issue_page_read(exp, llap, oig, 1);
1447         if (rc == 0) {
1448                 LL_CDEBUG_PAGE(D_READA | D_PAGE, page, "started read-ahead\n");
1449                 rc = 1;
1450         } else {
1451 unlock_page:
1452                 unlock_page(page);
1453                 LL_CDEBUG_PAGE(D_READA | D_PAGE, page, "skipping read-ahead\n");
1454         }
1455         page_cache_release(page);
1456         return rc;
1457 }
1458
1459 /* ra_io_arg will be filled in the beginning of ll_readahead with
1460  * ras_lock, then the following ll_read_ahead_pages will read RA
1461  * pages according to this arg, all the items in this structure are
1462  * counted by page index.
1463  */
1464 struct ra_io_arg {
1465         unsigned long ria_start;  /* start offset of read-ahead*/
1466         unsigned long ria_end;    /* end offset of read-ahead*/
1467         /* If stride read pattern is detected, ria_stoff means where
1468          * stride read is started. Note: for normal read-ahead, the
1469          * value here is meaningless, and also it will not be accessed*/
1470         pgoff_t ria_stoff;
1471         /* ria_length and ria_pages are the length and pages length in the
1472          * stride I/O mode. And they will also be used to check whether
1473          * it is stride I/O read-ahead in the read-ahead pages*/
1474         unsigned long ria_length;
1475         unsigned long ria_pages;
1476 };
1477
1478 #define RIA_DEBUG(ria)                                                \
1479         CDEBUG(D_READA, "rs %lu re %lu ro %lu rl %lu rp %lu\n",       \
1480         ria->ria_start, ria->ria_end, ria->ria_stoff, ria->ria_length,\
1481         ria->ria_pages)
1482
1483 #define INIT_RAS_WINDOW_PAGES PTLRPC_MAX_BRW_PAGES
1484
1485 static inline int stride_io_mode(struct ll_readahead_state *ras)
1486 {
1487         return ras->ras_consecutive_stride_requests > 1;
1488 }
1489
1490 /* The function calculates how much pages will be read in
1491  * [off, off + length], which will be read by stride I/O mode,
1492  * stride_offset = st_off, stride_lengh = st_len,
1493  * stride_pages = st_pgs
1494  */
1495 static unsigned long
1496 stride_pg_count(pgoff_t st_off, unsigned long st_len, unsigned long st_pgs,
1497                 unsigned long off, unsigned length)
1498 {
1499         unsigned long cont_len = st_off > off ?  st_off - off : 0;
1500         __u64 stride_len = length + off > st_off ?
1501                            length + off + 1 - st_off : 0;
1502         unsigned long left, pg_count;
1503
1504         if (st_len == 0 || length == 0)
1505                 return length;
1506
1507         left = do_div(stride_len, st_len);
1508         left = min(left, st_pgs);
1509
1510         pg_count = left + stride_len * st_pgs + cont_len;
1511
1512         LASSERT(pg_count >= left);
1513
1514         CDEBUG(D_READA, "st_off %lu, st_len %lu st_pgs %lu off %lu length %u"
1515                "pgcount %lu\n", st_off, st_len, st_pgs, off, length, pg_count);
1516
1517         return pg_count;
1518 }
1519
1520 static int ria_page_count(struct ra_io_arg *ria)
1521 {
1522         __u64 length = ria->ria_end >= ria->ria_start ?
1523                        ria->ria_end - ria->ria_start + 1 : 0;
1524
1525         return stride_pg_count(ria->ria_stoff, ria->ria_length,
1526                                ria->ria_pages, ria->ria_start,
1527                                length);
1528 }
1529
1530 /*Check whether the index is in the defined ra-window */
1531 static int ras_inside_ra_window(unsigned long idx, struct ra_io_arg *ria)
1532 {
1533         /* If ria_length == ria_pages, it means non-stride I/O mode,
1534          * idx should always inside read-ahead window in this case
1535          * For stride I/O mode, just check whether the idx is inside
1536          * the ria_pages. */
1537         return ria->ria_length == 0 || ria->ria_length == ria->ria_pages ||
1538                (idx - ria->ria_stoff) % ria->ria_length < ria->ria_pages;
1539 }
1540
1541 static int ll_read_ahead_pages(struct obd_export *exp,
1542                                struct obd_io_group *oig,
1543                                struct ra_io_arg *ria,
1544                                unsigned long *reserved_pages,
1545                                struct address_space *mapping,
1546                                unsigned long *ra_end)
1547 {
1548         int rc, count = 0, stride_ria;
1549         unsigned long page_idx;
1550
1551         LASSERT(ria != NULL);
1552         RIA_DEBUG(ria);
1553
1554         stride_ria = ria->ria_length > ria->ria_pages && ria->ria_pages > 0;
1555         for (page_idx = ria->ria_start; page_idx <= ria->ria_end &&
1556                         *reserved_pages > 0; page_idx++) {
1557                 if (ras_inside_ra_window(page_idx, ria)) {
1558                         /* If the page is inside the read-ahead window*/
1559                         rc = ll_read_ahead_page(exp, oig, page_idx, mapping);
1560                         if (rc == 1) {
1561                                 (*reserved_pages)--;
1562                                 count ++;
1563                         } else if (rc == -ENOLCK)
1564                                 break;
1565                 } else if (stride_ria) {
1566                         /* If it is not in the read-ahead window, and it is
1567                          * read-ahead mode, then check whether it should skip
1568                          * the stride gap */
1569                         pgoff_t offset;
1570                         /* FIXME: This assertion only is valid when it is for
1571                          * forward read-ahead, it will be fixed when backward
1572                          * read-ahead is implemented */
1573                         LASSERTF(page_idx > ria->ria_stoff, "since %lu in the"
1574                                 " gap of ra window,it should bigger than stride"
1575                                 " offset %lu \n", page_idx, ria->ria_stoff);
1576
1577                         offset = page_idx - ria->ria_stoff;
1578                         offset = offset % (ria->ria_length);
1579                         if (offset > ria->ria_pages) {
1580                                 page_idx += ria->ria_length - offset;
1581                                 CDEBUG(D_READA, "i %lu skip %lu \n", page_idx,
1582                                        ria->ria_length - offset);
1583                                 continue;
1584                         }
1585                 }
1586         }
1587         *ra_end = page_idx;
1588         return count;
1589 }
1590
1591 static int ll_readahead(struct ll_readahead_state *ras,
1592                          struct obd_export *exp, struct address_space *mapping,
1593                          struct obd_io_group *oig, int flags)
1594 {
1595         unsigned long start = 0, end = 0, reserved;
1596         unsigned long ra_end, len;
1597         struct inode *inode;
1598         struct lov_stripe_md *lsm;
1599         struct ll_ra_read *bead;
1600         struct ost_lvb lvb;
1601         struct ra_io_arg ria = { 0 };
1602         int ret = 0;
1603         __u64 kms;
1604         ENTRY;
1605
1606         inode = mapping->host;
1607         lsm = ll_i2info(inode)->lli_smd;
1608
1609         lov_stripe_lock(lsm);
1610         inode_init_lvb(inode, &lvb);
1611         obd_merge_lvb(ll_i2obdexp(inode), lsm, &lvb, 1);
1612         kms = lvb.lvb_size;
1613         lov_stripe_unlock(lsm);
1614         if (kms == 0) {
1615                 ll_ra_stats_inc(mapping, RA_STAT_ZERO_LEN);
1616                 RETURN(0);
1617         }
1618
1619         spin_lock(&ras->ras_lock);
1620         bead = ll_ra_read_get_locked(ras);
1621         /* Enlarge the RA window to encompass the full read */
1622         if (bead != NULL && ras->ras_window_start + ras->ras_window_len <
1623             bead->lrr_start + bead->lrr_count) {
1624                 obd_off read_end = (bead->lrr_start + bead->lrr_count) <<
1625                                     CFS_PAGE_SHIFT;
1626                 obd_extent_calc(exp, lsm, OBD_CALC_STRIPE_RPC_END_ALIGN,
1627                                 &read_end);
1628                 ras->ras_window_len = ((read_end + 1) >> CFS_PAGE_SHIFT) -
1629                                       ras->ras_window_start;
1630         }
1631         /* Reserve a part of the read-ahead window that we'll be issuing */
1632         if (ras->ras_window_len) {
1633                 start = ras->ras_next_readahead;
1634                 end = ras->ras_window_start + ras->ras_window_len - 1;
1635         }
1636         if (end != 0) {
1637                 /* Truncate RA window to end of file */
1638                 end = min(end, (unsigned long)((kms - 1) >> CFS_PAGE_SHIFT));
1639                 ras->ras_next_readahead = max(end, end + 1);
1640                 RAS_CDEBUG(ras);
1641         }
1642         ria.ria_start = start;
1643         ria.ria_end = end;
1644         /* If stride I/O mode is detected, get stride window*/
1645         if (stride_io_mode(ras)) {
1646                 ria.ria_stoff = ras->ras_stride_offset;
1647                 ria.ria_length = ras->ras_stride_length;
1648                 ria.ria_pages = ras->ras_stride_pages;
1649         }
1650         spin_unlock(&ras->ras_lock);
1651
1652         if (end == 0) {
1653                 ll_ra_stats_inc(mapping, RA_STAT_ZERO_WINDOW);
1654                 RETURN(0);
1655         }
1656
1657         len = ria_page_count(&ria);
1658         if (len == 0)
1659                 RETURN(0);
1660
1661         reserved = ll_ra_count_get(ll_i2sbi(inode), len);
1662         if (reserved < len)
1663                 ll_ra_stats_inc(mapping, RA_STAT_MAX_IN_FLIGHT);
1664
1665         CDEBUG(D_READA, "reserved page %lu \n", reserved);
1666
1667         ret = ll_read_ahead_pages(exp, oig, &ria, &reserved, mapping, &ra_end);
1668
1669         LASSERTF(reserved >= 0, "reserved %lu\n", reserved);
1670         if (reserved != 0)
1671                 ll_ra_count_put(ll_i2sbi(inode), reserved);
1672
1673         if (ra_end == end + 1 && ra_end == (kms >> CFS_PAGE_SHIFT))
1674                 ll_ra_stats_inc(mapping, RA_STAT_EOF);
1675
1676         /* if we didn't get to the end of the region we reserved from
1677          * the ras we need to go back and update the ras so that the
1678          * next read-ahead tries from where we left off.  we only do so
1679          * if the region we failed to issue read-ahead on is still ahead
1680          * of the app and behind the next index to start read-ahead from */
1681         CDEBUG(D_READA, "ra_end %lu end %lu stride end %lu \n",
1682                ra_end, end, ria.ria_end);
1683
1684         if (ra_end != (end + 1)) {
1685                 spin_lock(&ras->ras_lock);
1686                 if (ra_end < ras->ras_next_readahead &&
1687                     index_in_window(ra_end, ras->ras_window_start, 0,
1688                                     ras->ras_window_len)) {
1689                         ras->ras_next_readahead = ra_end;
1690                         RAS_CDEBUG(ras);
1691                 }
1692                 spin_unlock(&ras->ras_lock);
1693         }
1694
1695         RETURN(ret);
1696 }
1697
1698 static void ras_set_start(struct ll_readahead_state *ras, unsigned long index)
1699 {
1700         ras->ras_window_start = index & (~(INIT_RAS_WINDOW_PAGES - 1));
1701 }
1702
1703 /* called with the ras_lock held or from places where it doesn't matter */
1704 static void ras_reset(struct ll_readahead_state *ras, unsigned long index)
1705 {
1706         ras->ras_last_readpage = index;
1707         ras->ras_consecutive_requests = 0;
1708         ras->ras_consecutive_pages = 0;
1709         ras->ras_window_len = 0;
1710         ras_set_start(ras, index);
1711         ras->ras_next_readahead = max(ras->ras_window_start, index);
1712
1713         RAS_CDEBUG(ras);
1714 }
1715
1716 /* called with the ras_lock held or from places where it doesn't matter */
1717 static void ras_stride_reset(struct ll_readahead_state *ras)
1718 {
1719         ras->ras_consecutive_stride_requests = 0;
1720         ras->ras_stride_length = 0;
1721         ras->ras_stride_pages = 0;
1722         RAS_CDEBUG(ras);
1723 }
1724
1725 void ll_readahead_init(struct inode *inode, struct ll_readahead_state *ras)
1726 {
1727         spin_lock_init(&ras->ras_lock);
1728         ras_reset(ras, 0);
1729         ras->ras_requests = 0;
1730         INIT_LIST_HEAD(&ras->ras_read_beads);
1731 }
1732
1733 /*
1734  * Check whether the read request is in the stride window.
1735  * If it is in the stride window, return 1, otherwise return 0.
1736  */
1737 static int index_in_stride_window(unsigned long index,
1738                                   struct ll_readahead_state *ras,
1739                                   struct inode *inode)
1740 {
1741         unsigned long stride_gap = index - ras->ras_last_readpage - 1;
1742
1743         if (ras->ras_stride_length == 0 || ras->ras_stride_pages == 0)
1744                 return 0;
1745
1746         /* If it is contiguous read */
1747         if (stride_gap == 0)
1748                 return ras->ras_consecutive_pages + 1 <= ras->ras_stride_pages;
1749
1750         /*Otherwise check the stride by itself */
1751         return (ras->ras_stride_length - ras->ras_stride_pages) == stride_gap &&
1752              ras->ras_consecutive_pages == ras->ras_stride_pages;
1753 }
1754
1755 static void ras_update_stride_detector(struct ll_readahead_state *ras,
1756                                        unsigned long index)
1757 {
1758         unsigned long stride_gap = index - ras->ras_last_readpage - 1;
1759
1760         if (!stride_io_mode(ras) && (stride_gap != 0 ||
1761              ras->ras_consecutive_stride_requests == 0)) {
1762                 ras->ras_stride_pages = ras->ras_consecutive_pages;
1763                 ras->ras_stride_length = stride_gap +ras->ras_consecutive_pages;
1764         }
1765         RAS_CDEBUG(ras);
1766 }
1767
1768 static unsigned long
1769 stride_page_count(struct ll_readahead_state *ras, unsigned long len)
1770 {
1771         return stride_pg_count(ras->ras_stride_offset, ras->ras_stride_length,
1772                                ras->ras_stride_pages, ras->ras_stride_offset,
1773                                len);
1774 }
1775
1776 /* Stride Read-ahead window will be increased inc_len according to
1777  * stride I/O pattern */
1778 static void ras_stride_increase_window(struct ll_readahead_state *ras,
1779                                        struct ll_ra_info *ra,
1780                                        unsigned long inc_len)
1781 {
1782         unsigned long left, step, window_len;
1783         unsigned long stride_len;
1784
1785         LASSERT(ras->ras_stride_length > 0);
1786         LASSERTF(ras->ras_window_start + ras->ras_window_len
1787                  >= ras->ras_stride_offset, "window_start %lu, window_len %lu"
1788                  " stride_offset %lu\n", ras->ras_window_start,
1789                  ras->ras_window_len, ras->ras_stride_offset);
1790
1791         stride_len = ras->ras_window_start + ras->ras_window_len -
1792                      ras->ras_stride_offset;
1793
1794         left = stride_len % ras->ras_stride_length;
1795         window_len = ras->ras_window_len - left;
1796
1797         if (left < ras->ras_stride_pages)
1798                 left += inc_len;
1799         else
1800                 left = ras->ras_stride_pages + inc_len;
1801
1802         LASSERT(ras->ras_stride_pages != 0);
1803
1804         step = left / ras->ras_stride_pages;
1805         left %= ras->ras_stride_pages;
1806
1807         window_len += step * ras->ras_stride_length + left;
1808
1809         if (stride_page_count(ras, window_len) <= ra->ra_max_pages_per_file)
1810                 ras->ras_window_len = window_len;
1811
1812         RAS_CDEBUG(ras);
1813 }
1814
1815 /* Set stride I/O read-ahead window start offset */
1816 static void ras_set_stride_offset(struct ll_readahead_state *ras)
1817 {
1818         unsigned long window_len = ras->ras_next_readahead -
1819                                    ras->ras_window_start;
1820         unsigned long left;
1821
1822         LASSERT(ras->ras_stride_length != 0);
1823
1824         left = window_len % ras->ras_stride_length;
1825
1826         ras->ras_stride_offset = ras->ras_next_readahead - left;
1827
1828         RAS_CDEBUG(ras);
1829 }
1830
1831 static void ras_increase_window(struct ll_readahead_state *ras,
1832                                 struct ll_ra_info *ra, struct inode *inode)
1833 {
1834         __u64 step;
1835         __u32 size;
1836         int rc;
1837
1838         step = ((loff_t)(ras->ras_window_start +
1839                          ras->ras_window_len)) << CFS_PAGE_SHIFT;
1840         size = sizeof(step);
1841         /*Get rpc_size for this offset (step) */
1842         rc = obd_get_info(ll_i2obdexp(inode), sizeof(KEY_OFF_RPCSIZE),
1843                           KEY_OFF_RPCSIZE, &size, &step,
1844                           ll_i2info(inode)->lli_smd);
1845         if (rc)
1846                 step = INIT_RAS_WINDOW_PAGES;
1847
1848         if (stride_io_mode(ras))
1849                 ras_stride_increase_window(ras, ra, (unsigned long)step);
1850         else
1851                 ras->ras_window_len = min(ras->ras_window_len +
1852                                           (unsigned long)step,
1853                                           ra->ra_max_pages);
1854 }
1855
1856 static void ras_update(struct ll_sb_info *sbi, struct inode *inode,
1857                        struct ll_readahead_state *ras, unsigned long index,
1858                        unsigned hit)
1859 {
1860         struct ll_ra_info *ra = &sbi->ll_ra_info;
1861         int zero = 0, stride_detect = 0, ra_miss = 0;
1862         ENTRY;
1863
1864         spin_lock(&ras->ras_lock);
1865
1866         ll_ra_stats_inc_sbi(sbi, hit ? RA_STAT_HIT : RA_STAT_MISS);
1867
1868         /* reset the read-ahead window in two cases.  First when the app seeks
1869          * or reads to some other part of the file.  Secondly if we get a
1870          * read-ahead miss that we think we've previously issued.  This can
1871          * be a symptom of there being so many read-ahead pages that the VM is
1872          * reclaiming it before we get to it. */
1873         if (!index_in_window(index, ras->ras_last_readpage, 8, 8)) {
1874                 zero = 1;
1875                 ll_ra_stats_inc_sbi(sbi, RA_STAT_DISTANT_READPAGE);
1876         } else if (!hit && ras->ras_window_len &&
1877                    index < ras->ras_next_readahead &&
1878                    index_in_window(index, ras->ras_window_start, 0,
1879                                    ras->ras_window_len)) {
1880                 ra_miss = 1;
1881                 ll_ra_stats_inc_sbi(sbi, RA_STAT_MISS_IN_WINDOW);
1882         }
1883
1884         /* On the second access to a file smaller than the tunable
1885          * ra_max_read_ahead_whole_pages trigger RA on all pages in the
1886          * file up to ra_max_pages_per_file.  This is simply a best effort
1887          * and only occurs once per open file.  Normal RA behavior is reverted
1888          * to for subsequent IO.  The mmap case does not increment
1889          * ras_requests and thus can never trigger this behavior. */
1890         if (ras->ras_requests == 2 && !ras->ras_request_index) {
1891                 __u64 kms_pages;
1892
1893                 kms_pages = (i_size_read(inode) + CFS_PAGE_SIZE - 1) >>
1894                             CFS_PAGE_SHIFT;
1895
1896                 CDEBUG(D_READA, "kmsp "LPU64" mwp %lu mp %lu\n", kms_pages,
1897                        ra->ra_max_read_ahead_whole_pages,
1898                        ra->ra_max_pages_per_file);
1899
1900                 if (kms_pages &&
1901                     kms_pages <= ra->ra_max_read_ahead_whole_pages) {
1902                         ras->ras_window_start = 0;
1903                         ras->ras_last_readpage = 0;
1904                         ras->ras_next_readahead = 0;
1905                         ras->ras_window_len = min(ra->ra_max_pages_per_file,
1906                                 ra->ra_max_read_ahead_whole_pages);
1907                         GOTO(out_unlock, 0);
1908                 }
1909         }
1910         if (zero) {
1911                 /* check whether it is in stride I/O mode*/
1912                 if (!index_in_stride_window(index, ras, inode)) {
1913                         ras_reset(ras, index);
1914                         ras->ras_consecutive_pages++;
1915                         ras_stride_reset(ras);
1916                         GOTO(out_unlock, 0);
1917                 } else {
1918                         ras->ras_consecutive_requests = 0;
1919                         if (++ras->ras_consecutive_stride_requests > 1)
1920                                 stride_detect = 1;
1921                         RAS_CDEBUG(ras);
1922                 }
1923         } else {
1924                 if (ra_miss) {
1925                         if (index_in_stride_window(index, ras, inode) &&
1926                             stride_io_mode(ras)) {
1927                                 /*If stride-RA hit cache miss, the stride dector
1928                                  *will not be reset to avoid the overhead of
1929                                  *redetecting read-ahead mode */
1930                                 if (index != ras->ras_last_readpage + 1)
1931                                        ras->ras_consecutive_pages = 0;
1932                                 RAS_CDEBUG(ras);
1933                         } else {
1934                                 /* Reset both stride window and normal RA
1935                                  * window */
1936                                 ras_reset(ras, index);
1937                                 ras->ras_consecutive_pages++;
1938                                 ras_stride_reset(ras);
1939                                 GOTO(out_unlock, 0);
1940                         }
1941                 } else if (stride_io_mode(ras)) {
1942                         /* If this is contiguous read but in stride I/O mode
1943                          * currently, check whether stride step still is valid,
1944                          * if invalid, it will reset the stride ra window*/
1945                         if (!index_in_stride_window(index, ras, inode)) {
1946                                 /* Shrink stride read-ahead window to be zero */
1947                                 ras_stride_reset(ras);
1948                                 ras->ras_window_len = 0;
1949                                 ras->ras_next_readahead = index;
1950                         }
1951                 }
1952         }
1953         ras->ras_consecutive_pages++;
1954         ras_update_stride_detector(ras, index);
1955         ras->ras_last_readpage = index;
1956         ras_set_start(ras, index);
1957         ras->ras_next_readahead = max(ras->ras_window_start,
1958                                       ras->ras_next_readahead);
1959         RAS_CDEBUG(ras);
1960
1961         /* Trigger RA in the mmap case where ras_consecutive_requests
1962          * is not incremented and thus can't be used to trigger RA */
1963         if (!ras->ras_window_len && ras->ras_consecutive_pages == 4) {
1964                 ras->ras_window_len = INIT_RAS_WINDOW_PAGES;
1965                 GOTO(out_unlock, 0);
1966         }
1967
1968         /* Initially reset the stride window offset to next_readahead*/
1969         if (ras->ras_consecutive_stride_requests == 2 && stride_detect)
1970                 ras_set_stride_offset(ras);
1971
1972         /* The initial ras_window_len is set to the request size.  To avoid
1973          * uselessly reading and discarding pages for random IO the window is
1974          * only increased once per consecutive request received. */
1975         if ((ras->ras_consecutive_requests > 1 &&
1976             !ras->ras_request_index) || stride_detect)
1977                 ras_increase_window(ras, ra, inode);
1978         EXIT;
1979 out_unlock:
1980         RAS_CDEBUG(ras);
1981         ras->ras_request_index++;
1982         spin_unlock(&ras->ras_lock);
1983         return;
1984 }
1985
1986 int ll_writepage(struct page *page)
1987 {
1988         struct inode *inode = page->mapping->host;
1989         struct ll_inode_info *lli = ll_i2info(inode);
1990         struct obd_export *exp;
1991         struct ll_async_page *llap;
1992         int rc = 0;
1993         ENTRY;
1994
1995         LASSERT(PageLocked(page));
1996
1997         exp = ll_i2obdexp(inode);
1998         if (exp == NULL)
1999                 GOTO(out, rc = -EINVAL);
2000
2001         llap = llap_from_page(page, LLAP_ORIGIN_WRITEPAGE);
2002         if (IS_ERR(llap))
2003                 GOTO(out, rc = PTR_ERR(llap));
2004
2005         LASSERT(!llap->llap_nocache);
2006         LASSERT(!PageWriteback(page));
2007         set_page_writeback(page);
2008
2009         page_cache_get(page);
2010         if (llap->llap_write_queued) {
2011                 LL_CDEBUG_PAGE(D_PAGE, page, "marking urgent\n");
2012                 rc = obd_set_async_flags(exp, lli->lli_smd, NULL,
2013                                          llap->llap_cookie,
2014                                          ASYNC_READY | ASYNC_URGENT);
2015         } else {
2016                 rc = queue_or_sync_write(exp, inode, llap, CFS_PAGE_SIZE,
2017                                          ASYNC_READY | ASYNC_URGENT);
2018         }
2019         if (rc) {
2020                 /* re-dirty page on error so it retries write */
2021                 if (PageWriteback(page))
2022                         end_page_writeback(page);
2023
2024                 /* resend page only for not started IO*/
2025                 if (!PageError(page))
2026                         ll_redirty_page(page);
2027
2028                 page_cache_release(page);
2029         }
2030 out:
2031         if (rc) {
2032                 if (!lli->lli_async_rc)
2033                         lli->lli_async_rc = rc;
2034                 /* resend page only for not started IO*/
2035                 unlock_page(page);
2036         }
2037         RETURN(rc);
2038 }
2039
2040 /*
2041  * for now we do our readpage the same on both 2.4 and 2.5.  The kernel's
2042  * read-ahead assumes it is valid to issue readpage all the way up to
2043  * i_size, but our dlm locks make that not the case.  We disable the
2044  * kernel's read-ahead and do our own by walking ahead in the page cache
2045  * checking for dlm lock coverage.  the main difference between 2.4 and
2046  * 2.6 is how read-ahead gets batched and issued, but we're using our own,
2047  * so they look the same.
2048  */
2049 int ll_readpage(struct file *filp, struct page *page)
2050 {
2051         struct ll_file_data *fd = LUSTRE_FPRIVATE(filp);
2052         struct inode *inode = page->mapping->host;
2053         struct obd_export *exp;
2054         struct ll_async_page *llap;
2055         struct obd_io_group *oig = NULL;
2056         struct lustre_handle *lockh = NULL;
2057         int rc;
2058         ENTRY;
2059
2060         LASSERT(PageLocked(page));
2061         LASSERT(!PageUptodate(page));
2062         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),offset=%Lu=%#Lx\n",
2063                inode->i_ino, inode->i_generation, inode,
2064                (((loff_t)page->index) << CFS_PAGE_SHIFT),
2065                (((loff_t)page->index) << CFS_PAGE_SHIFT));
2066         LASSERT(atomic_read(&filp->f_dentry->d_inode->i_count) > 0);
2067
2068         if (!ll_i2info(inode)->lli_smd) {
2069                 /* File with no objects - one big hole */
2070                 /* We use this just for remove_from_page_cache that is not
2071                  * exported, we'd make page back up to date. */
2072                 ll_truncate_complete_page(page);
2073                 clear_page(kmap(page));
2074                 kunmap(page);
2075                 SetPageUptodate(page);
2076                 unlock_page(page);
2077                 RETURN(0);
2078         }
2079
2080         rc = oig_init(&oig);
2081         if (rc < 0)
2082                 GOTO(out, rc);
2083
2084         exp = ll_i2obdexp(inode);
2085         if (exp == NULL)
2086                 GOTO(out, rc = -EINVAL);
2087
2088         if (fd->fd_flags & LL_FILE_GROUP_LOCKED)
2089                 lockh = &fd->fd_cwlockh;
2090
2091         llap = llap_from_page_with_lockh(page, LLAP_ORIGIN_READPAGE, lockh);
2092         if (IS_ERR(llap)) {
2093                 if (PTR_ERR(llap) == -ENOLCK) {
2094                         CWARN("ino %lu page %lu (%llu) not covered by "
2095                               "a lock (mmap?).  check debug logs.\n",
2096                               inode->i_ino, page->index,
2097                               (long long)page->index << PAGE_CACHE_SHIFT);
2098                 }
2099                 GOTO(out, rc = PTR_ERR(llap));
2100         }
2101
2102         if (ll_i2sbi(inode)->ll_ra_info.ra_max_pages_per_file)
2103                 ras_update(ll_i2sbi(inode), inode, &fd->fd_ras, page->index,
2104                            llap->llap_defer_uptodate);
2105
2106
2107         if (llap->llap_defer_uptodate) {
2108                 /* This is the callpath if we got the page from a readahead */
2109                 llap->llap_ra_used = 1;
2110                 rc = ll_readahead(&fd->fd_ras, exp, page->mapping, oig,
2111                                   fd->fd_flags);
2112                 if (rc > 0)
2113                         obd_trigger_group_io(exp, ll_i2info(inode)->lli_smd,
2114                                              NULL, oig);
2115                 LL_CDEBUG_PAGE(D_PAGE, page, "marking uptodate from defer\n");
2116                 SetPageUptodate(page);
2117                 unlock_page(page);
2118                 GOTO(out_oig, rc = 0);
2119         }
2120
2121         rc = ll_issue_page_read(exp, llap, oig, 0);
2122         if (rc)
2123                 GOTO(out, rc);
2124
2125         LL_CDEBUG_PAGE(D_PAGE, page, "queued readpage\n");
2126         /* We have just requested the actual page we want, see if we can tack
2127          * on some readahead to that page's RPC before it is sent. */
2128         if (ll_i2sbi(inode)->ll_ra_info.ra_max_pages_per_file)
2129                 ll_readahead(&fd->fd_ras, exp, page->mapping, oig,
2130                              fd->fd_flags);
2131
2132         rc = obd_trigger_group_io(exp, ll_i2info(inode)->lli_smd, NULL, oig);
2133
2134 out:
2135         if (rc)
2136                 unlock_page(page);
2137 out_oig:
2138         if (oig != NULL)
2139                 oig_release(oig);
2140         RETURN(rc);
2141 }
2142
2143 static void ll_file_put_pages(struct page **pages, int numpages)
2144 {
2145         int i;
2146         struct page **pp;
2147         ENTRY;
2148
2149         for (i = 0, pp = pages; i < numpages; i++, pp++) {
2150                 if (*pp) {
2151                         LL_CDEBUG_PAGE(D_PAGE, (*pp), "free\n");
2152                         __ll_put_llap(*pp);
2153                         if (page_private(*pp))
2154                                 CERROR("the llap wasn't freed\n");
2155                         (*pp)->mapping = NULL;
2156                         if (page_count(*pp) != 1)
2157                                 CERROR("page %p, flags %#lx, count %i, "
2158                                        "private %p\n", (*pp),
2159                                        (unsigned long)(*pp)->flags,
2160                                        page_count(*pp),
2161                                        (void*)page_private(*pp));
2162                         __free_pages(*pp, 0);
2163                 }
2164         }
2165         OBD_FREE(pages, numpages * sizeof(struct page*));
2166         EXIT;
2167 }
2168
2169 static struct page **ll_file_prepare_pages(int numpages, struct inode *inode,
2170                                            unsigned long first)
2171 {
2172         struct page **pages;
2173         int i;
2174         int rc = 0;
2175         ENTRY;
2176
2177         OBD_ALLOC(pages, sizeof(struct page *) * numpages);
2178         if (pages == NULL)
2179                 RETURN(ERR_PTR(-ENOMEM));
2180         for (i = 0; i < numpages; i++) {
2181                 struct page *page;
2182                 struct ll_async_page *llap;
2183
2184                 page = alloc_pages(GFP_HIGHUSER, 0);
2185                 if (page == NULL)
2186                         GOTO(err, rc = -ENOMEM);
2187                 pages[i] = page;
2188                 /* llap_from_page needs page index and mapping to be set */
2189                 page->index = first++;
2190                 page->mapping = inode->i_mapping;
2191                 llap = llap_from_page(page, LLAP_ORIGIN_LOCKLESS_IO);
2192                 if (IS_ERR(llap))
2193                         GOTO(err, rc = PTR_ERR(llap));
2194                 llap->llap_lockless_io_page = 1;
2195         }
2196         RETURN(pages);
2197 err:
2198         ll_file_put_pages(pages, numpages);
2199         RETURN(ERR_PTR(rc));
2200  }
2201
2202 static ssize_t ll_file_copy_pages(struct page **pages, int numpages,
2203                                   const struct iovec *iov, unsigned long nsegs,
2204                                   ssize_t iov_offset, loff_t pos, size_t count,
2205                                   int rw)
2206 {
2207         ssize_t amount = 0;
2208         int i;
2209         int updatechecksum = ll_i2sbi(pages[0]->mapping->host)->ll_flags &
2210                              LL_SBI_LLITE_CHECKSUM;
2211         ENTRY;
2212
2213         for (i = 0; i < numpages; i++) {
2214                 unsigned offset, bytes, left = 0;
2215                 char *vaddr;
2216
2217                 vaddr = kmap(pages[i]);
2218                 offset = pos & (CFS_PAGE_SIZE - 1);
2219                 bytes = min_t(unsigned, CFS_PAGE_SIZE - offset, count);
2220                 LL_CDEBUG_PAGE(D_PAGE, pages[i], "op = %s, addr = %p, "
2221                                "bytes = %u\n",
2222                                (rw == WRITE) ? "CFU" : "CTU",
2223                                vaddr + offset, bytes);
2224                 while (bytes > 0 && !left && nsegs) {
2225                         unsigned copy = min_t(ssize_t, bytes,
2226                                                iov->iov_len - iov_offset);
2227                         if (rw == WRITE) {
2228                                 left = copy_from_user(vaddr + offset,
2229                                                       iov->iov_base +iov_offset,
2230                                                       copy);
2231                                 if (updatechecksum) {
2232                                         struct ll_async_page *llap;
2233
2234                                         llap = llap_cast_private(pages[i]);
2235                                         llap->llap_checksum =
2236                                                 init_checksum(OSC_DEFAULT_CKSUM);
2237                                         llap->llap_checksum =
2238                                            compute_checksum(llap->llap_checksum,
2239                                                             vaddr,CFS_PAGE_SIZE,
2240                                                             OSC_DEFAULT_CKSUM);
2241                                 }
2242                         } else {
2243                                 left = copy_to_user(iov->iov_base + iov_offset,
2244                                                     vaddr + offset, copy);
2245                         }
2246
2247                         amount += copy;
2248                         count -= copy;
2249                         pos += copy;
2250                         iov_offset += copy;
2251                         bytes -= copy;
2252                         if (iov_offset == iov->iov_len) {
2253                                 iov_offset = 0;
2254                                 iov++;
2255                                 nsegs--;
2256                         }
2257                 }
2258                 kunmap(pages[i]);
2259                 if (left) {
2260                         amount -= left;
2261                         break;
2262                 }
2263         }
2264         if (amount == 0)
2265                 RETURN(-EFAULT);
2266         RETURN(amount);
2267 }
2268
2269 static int ll_file_oig_pages(struct inode * inode, struct page **pages,
2270                              int numpages, loff_t pos, size_t count, int rw)
2271 {
2272         struct obd_io_group *oig;
2273         struct ll_inode_info *lli = ll_i2info(inode);
2274         struct obd_export *exp;
2275         loff_t org_pos = pos;
2276         obd_flag brw_flags;
2277         int rc;
2278         int i;
2279         ENTRY;
2280
2281         exp = ll_i2obdexp(inode);
2282         if (exp == NULL)
2283                 RETURN(-EINVAL);
2284         rc = oig_init(&oig);
2285         if (rc)
2286                 RETURN(rc);
2287         brw_flags = OBD_BRW_SRVLOCK;
2288         if (cfs_capable(CFS_CAP_SYS_RESOURCE))
2289                 brw_flags |= OBD_BRW_NOQUOTA;
2290
2291         for (i = 0; i < numpages; i++) {
2292                 struct ll_async_page *llap;
2293                 unsigned from, bytes;
2294
2295                 from = pos & (CFS_PAGE_SIZE - 1);
2296                 bytes = min_t(unsigned, CFS_PAGE_SIZE - from,
2297                               count - pos + org_pos);
2298                 llap = llap_cast_private(pages[i]);
2299                 LASSERT(llap);
2300
2301                 lock_page(pages[i]);
2302
2303                 LL_CDEBUG_PAGE(D_PAGE, pages[i], "offset "LPU64","
2304                                " from %u, bytes = %u\n",
2305                                (__u64)pos, from, bytes);
2306                 LASSERTF(pos >> CFS_PAGE_SHIFT == pages[i]->index,
2307                          "wrong page index %lu (%lu)\n",
2308                          pages[i]->index,
2309                          (unsigned long)(pos >> CFS_PAGE_SHIFT));
2310                 rc = obd_queue_group_io(exp, lli->lli_smd, NULL, oig,
2311                                         llap->llap_cookie,
2312                                         (rw == WRITE) ?
2313                                         OBD_BRW_WRITE:OBD_BRW_READ,
2314                                         from, bytes, brw_flags,
2315                                         ASYNC_READY | ASYNC_URGENT |
2316                                         ASYNC_COUNT_STABLE | ASYNC_GROUP_SYNC);
2317                 if (rc) {
2318                         i++;
2319                         GOTO(out, rc);
2320                 }
2321                 pos += bytes;
2322         }
2323         rc = obd_trigger_group_io(exp, lli->lli_smd, NULL, oig);
2324         if (rc)
2325                 GOTO(out, rc);
2326         rc = oig_wait(oig);
2327 out:
2328         while(--i >= 0)
2329                 unlock_page(pages[i]);
2330         oig_release(oig);
2331         RETURN(rc);
2332 }
2333
2334 /* Advance through passed iov, adjust iov pointer as necessary and return
2335  * starting offset in individual entry we are pointing at. Also reduce
2336  * nr_segs as needed */
2337 static ssize_t ll_iov_advance(const struct iovec **iov, unsigned long *nr_segs,
2338                               ssize_t offset)
2339 {
2340         while (*nr_segs > 0) {
2341                 if ((*iov)->iov_len > offset)
2342                         return ((*iov)->iov_len - offset);
2343                 offset -= (*iov)->iov_len;
2344                 (*iov)++;
2345                 (*nr_segs)--;
2346         }
2347         return 0;
2348 }
2349
2350 ssize_t ll_file_lockless_io(struct file *file, const struct iovec *iov,
2351                             unsigned long nr_segs,
2352                             loff_t *ppos, int rw, ssize_t count)
2353 {
2354         loff_t pos;
2355         struct inode *inode = file->f_dentry->d_inode;
2356         ssize_t rc = 0;
2357         int max_pages;
2358         size_t amount = 0;
2359         unsigned long first, last;
2360         const struct iovec *iv = &iov[0];
2361         unsigned long nsegs = nr_segs;
2362         unsigned long offset = 0;
2363         ENTRY;
2364
2365         if (rw == READ) {
2366                 loff_t isize;
2367
2368                 ll_inode_size_lock(inode, 0);
2369                 isize = i_size_read(inode);
2370                 ll_inode_size_unlock(inode, 0);
2371                 if (*ppos >= isize)
2372                         GOTO(out, rc = 0);
2373                 if (*ppos + count >= isize)
2374                         count -= *ppos + count - isize;
2375                 if (count == 0)
2376                         GOTO(out, rc);
2377         } else {
2378                 rc = generic_write_checks(file, ppos, &count, 0);
2379                 if (rc)
2380                         GOTO(out, rc);
2381                 rc = ll_remove_suid(file, file->f_vfsmnt);
2382                 if (rc)
2383                         GOTO(out, rc);
2384         }
2385
2386         pos = *ppos;
2387         first = pos >> CFS_PAGE_SHIFT;
2388         last = (pos + count - 1) >> CFS_PAGE_SHIFT;
2389         max_pages = PTLRPC_MAX_BRW_PAGES *
2390                 ll_i2info(inode)->lli_smd->lsm_stripe_count;
2391         CDEBUG(D_INFO, "%u, stripe_count = %u\n",
2392                PTLRPC_MAX_BRW_PAGES /* max_pages_per_rpc */,
2393                ll_i2info(inode)->lli_smd->lsm_stripe_count);
2394
2395         while (first <= last && rc >= 0) {
2396                 int pages_for_io;
2397                 struct page **pages;
2398                 size_t bytes = count - amount;
2399
2400                 pages_for_io = min_t(int, last - first + 1, max_pages);
2401                 pages = ll_file_prepare_pages(pages_for_io, inode, first);
2402                 if (IS_ERR(pages)) {
2403                         rc = PTR_ERR(pages);
2404                         break;
2405                 }
2406                 if (rw == WRITE) {
2407                         rc = ll_file_copy_pages(pages, pages_for_io, iv, nsegs,
2408                                                 offset, pos + amount, bytes,
2409                                                 rw);
2410                         if (rc < 0)
2411                                 GOTO(put_pages, rc);
2412                         offset = ll_iov_advance(&iv, &nsegs, offset + rc);
2413                         bytes = rc;
2414                 }
2415                 rc = ll_file_oig_pages(inode, pages, pages_for_io,
2416                                        pos + amount, bytes, rw);
2417                 if (rc)
2418                         GOTO(put_pages, rc);
2419                 if (rw == READ) {
2420                         rc = ll_file_copy_pages(pages, pages_for_io, iv, nsegs,
2421                                                 offset, pos + amount, bytes, rw);
2422                         if (rc < 0)
2423                                 GOTO(put_pages, rc);
2424                         offset = ll_iov_advance(&iv, &nsegs, offset + rc);
2425                         bytes = rc;
2426                 }
2427                 amount += bytes;
2428 put_pages:
2429                 ll_file_put_pages(pages, pages_for_io);
2430                 first += pages_for_io;
2431                 /* a short read/write check */
2432                 if (pos + amount < ((loff_t)first << CFS_PAGE_SHIFT))
2433                         break;
2434                 /* Check if we are out of userspace buffers. (how that could
2435                    happen?) */
2436                 if (nsegs == 0)
2437                         break;
2438         }
2439         /* NOTE: don't update i_size and KMS in absence of LDLM locks even
2440          * write makes the file large */
2441         file_accessed(file);
2442         if (rw == READ && amount < count && rc == 0) {
2443                 unsigned long not_cleared;
2444
2445                 while (nsegs > 0) {
2446                         ssize_t to_clear = min_t(ssize_t, count - amount,
2447                                                  iv->iov_len - offset);
2448                         not_cleared = clear_user(iv->iov_base + offset,
2449                                                  to_clear);
2450                         amount += to_clear - not_cleared;
2451                         if (not_cleared) {
2452                                 rc = -EFAULT;
2453                                 break;
2454                         }
2455                         offset = 0;
2456                         iv++;
2457                         nsegs--;
2458                 }
2459         }
2460         if (amount > 0) {
2461                 lprocfs_counter_add(ll_i2sbi(inode)->ll_stats,
2462                                     (rw == WRITE) ?
2463                                     LPROC_LL_LOCKLESS_WRITE :
2464                                     LPROC_LL_LOCKLESS_READ,
2465                                     (long)amount);
2466                 *ppos += amount;
2467                 RETURN(amount);
2468         }
2469 out:
2470         RETURN(rc);
2471 }