lustre/llite/rw.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * Lustre Lite I/O Page Cache
   5  *
   6  *  Copyright (c) 2001-2003 Cluster File Systems, Inc.
   7  *
   8  *   This file is part of Lustre, http://www.lustre.org.
   9  *
  10  *   Lustre is free software; you can redistribute it and/or
  11  *   modify it under the terms of version 2 of the GNU General Public
  12  *   License as published by the Free Software Foundation.
  13  *
  14  *   Lustre is distributed in the hope that it will be useful,
  15  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  17  *   GNU General Public License for more details.
  18  *
  19  *   You should have received a copy of the GNU General Public License
  20  *   along with Lustre; if not, write to the Free Software
  21  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  22  */
  23
  24 #include <linux/config.h>
  25 #include <linux/kernel.h>
  26 #include <linux/mm.h>
  27 #include <linux/string.h>
  28 #include <linux/stat.h>
  29 #include <linux/errno.h>
  30 #include <linux/smp_lock.h>
  31 #include <linux/unistd.h>
  32 #include <linux/version.h>
  33 #include <asm/system.h>
  34 #include <asm/uaccess.h>
  35
  36
  37 #include <linux/fs.h>
  38 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
  39 #include <linux/buffer_head.h>
  40 #else
  41 #include <linux/iobuf.h>
  42 #endif
  43 #include <linux/stat.h>
  44 #include <asm/uaccess.h>
  45 #include <asm/segment.h>
  46 #include <linux/mm.h>
  47 #include <linux/pagemap.h>
  48 #include <linux/smp_lock.h>
  49
  50 #define DEBUG_SUBSYSTEM S_LLITE
  51
  52 #include <linux/lustre_mds.h>
  53 #include <linux/lustre_lite.h>
  54 #include <linux/lustre_lib.h>
  55 #include <linux/lustre_compat25.h>
  56
  57 /*
  58  * Remove page from dirty list
  59  */
  60 static void __set_page_clean(struct page *page)
  61 {
  62         struct address_space *mapping = page->mapping;
  63         struct inode *inode;
  64
  65         if (!mapping)
  66                 return;
  67
  68         PGCACHE_WRLOCK(mapping);
  69
  70         list_del(&page->list);
  71         list_add(&page->list, &mapping->clean_pages);
  72
  73         /* XXX doesn't inode_lock protect i_state ? */
  74         inode = mapping->host;
  75         if (list_empty(&mapping->dirty_pages)) {
  76                 CDEBUG(D_INODE, "inode clean\n");
  77                 inode->i_state &= ~I_DIRTY_PAGES;
  78         }
  79
  80         PGCACHE_WRUNLOCK(mapping);
  81         EXIT;
  82 }
  83
  84 void set_page_clean(struct page *page)
  85 {
  86         if (PageDirty(page)) {
  87                 ClearPageDirty(page);
  88                 __set_page_clean(page);
  89         }
  90 }
  91
  92 /* SYNCHRONOUS I/O to object storage for an inode */
  93 static int ll_brw(int cmd, struct inode *inode, struct page *page, int flags)
  94 {
  95         struct ll_inode_info *lli = ll_i2info(inode);
  96         struct lov_stripe_md *lsm = lli->lli_smd;
  97         struct brw_page pg;
  98         int rc;
  99         ENTRY;
 100
 101         pg.pg = page;
 102         pg.off = ((obd_off)page->index) << PAGE_SHIFT;
 103
 104         if (cmd == OBD_BRW_WRITE && (pg.off + PAGE_SIZE > inode->i_size))
 105                 pg.count = inode->i_size % PAGE_SIZE;
 106         else
 107                 pg.count = PAGE_SIZE;
 108
 109         CDEBUG(D_PAGE, "%s %d bytes ino %lu at "LPU64"/"LPX64"\n",
 110                cmd & OBD_BRW_WRITE ? "write" : "read", pg.count, inode->i_ino,
 111                pg.off, pg.off);
 112         if (pg.count == 0) {
 113                 CERROR("ZERO COUNT: ino %lu: size %p:%Lu(%p:%Lu) idx %lu off "
 114                        LPU64"\n",
 115                        inode->i_ino, inode, inode->i_size, page->mapping->host,
 116                        page->mapping->host->i_size, page->index, pg.off);
 117         }
 118
 119         pg.flag = flags;
 120
 121         rc = obd_brw(cmd, ll_i2obdconn(inode), lsm, 1, &pg, NULL);
 122         if (rc)
 123                 CERROR("error from obd_brw: rc = %d\n", rc);
 124
 125         RETURN(rc);
 126 }
 127
 128 /*
 129  * we were asked to read a single page but we're going to try and read a batch
 130  * of pages all at once.  this vaguely simulates 2.5's readpages.
 131  */
 132 static int ll_readpage(struct file *file, struct page *first_page)
 133 {
 134         struct inode *inode = first_page->mapping->host;
 135         struct ll_inode_info *lli = ll_i2info(inode);
 136         struct page *page = first_page;
 137         struct list_head *pos;
 138         struct brw_page *pgs;
 139         unsigned long end_index, extent_end = 0;
 140         struct ptlrpc_request_set *set;
 141         int npgs = 0, rc = 0, max_pages;
 142         ENTRY;
 143
 144         LASSERT(PageLocked(page));
 145         LASSERT(!PageUptodate(page));
 146         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),offset="LPX64"\n",
 147                inode->i_ino, inode->i_generation, inode,
 148                (((obd_off)page->index) << PAGE_SHIFT));
 149         LASSERT(atomic_read(&file->f_dentry->d_inode->i_count) > 0);
 150
 151         if (inode->i_size <= ((obd_off)page->index) << PAGE_SHIFT) {
 152                 CERROR("reading beyond EOF\n");
 153                 memset(kmap(page), 0, PAGE_SIZE);
 154                 kunmap(page);
 155                 SetPageUptodate(page);
 156                 unlock_page(page);
 157                 RETURN(rc);
 158         }
 159
 160         /* try to read the file's preferred block size in a one-er */
 161         end_index = first_page->index +
 162                 (inode->i_blksize >> PAGE_CACHE_SHIFT);
 163         if (end_index > (inode->i_size >> PAGE_CACHE_SHIFT))
 164                 end_index = inode->i_size >> PAGE_CACHE_SHIFT;
 165
 166         max_pages = ((end_index - first_page->index) << PAGE_CACHE_SHIFT) >>
 167                 PAGE_SHIFT;
 168         pgs = kmalloc(max_pages * sizeof(*pgs), GFP_USER);
 169         if (pgs == NULL)
 170                 RETURN(-ENOMEM);
 171
 172         /*
 173          * find how far we're allowed to read under the extent ll_file_read
 174          * is passing us..
 175          */
 176         spin_lock(&lli->lli_read_extent_lock);
 177         list_for_each(pos, &lli->lli_read_extents) {
 178                 struct ll_read_extent *rextent;
 179                 rextent = list_entry(pos, struct ll_read_extent, re_lli_item);
 180                 if (rextent->re_task != current)
 181                         continue;
 182
 183                 if (rextent->re_extent.end + PAGE_SIZE < rextent->re_extent.end)
 184                         /* extent wrapping */
 185                         extent_end = ~0;
 186                 else {
 187                         extent_end = (rextent->re_extent.end + PAGE_SIZE)
 188                                                         << PAGE_CACHE_SHIFT;
 189                         /* 32bit indexes, 64bit extents.. */
 190                         if (((u64)extent_end >> PAGE_CACHE_SHIFT) <
 191                                         rextent->re_extent.end)
 192                                 extent_end = ~0;
 193                 }
 194                 break;
 195         }
 196         spin_unlock(&lli->lli_read_extent_lock);
 197
 198         if (extent_end == 0) {
 199                 static long next_print;
 200                 if (time_after(jiffies, next_print)) {
 201                         next_print = jiffies + 30 * HZ;
 202                         CDEBUG(D_INODE, "mmap readpage - check locks\n");
 203                 }
 204                 end_index = page->index + 1;
 205         } else if (extent_end < end_index)
 206                 end_index = extent_end;
 207
 208         /* to balance the find_get_page ref the other pages get that is
 209          * decrefed on teardown.. */
 210         page_cache_get(page);
 211         do {
 212                 unsigned long index ;
 213
 214                 pgs[npgs].pg = page;
 215                 pgs[npgs].off = ((obd_off)page->index) << PAGE_CACHE_SHIFT;
 216                 pgs[npgs].flag = 0;
 217                 pgs[npgs].count = PAGE_SIZE;
 218                 /* XXX Workaround for BA OSTs returning short reads at EOF.
 219                  * The linux OST will return the full page, zero-filled at the
 220                  * end, which will just overwrite the data we set here.  Bug
 221                  * 593 relates to fixing this properly.
 222                  */
 223                 if (inode->i_size < pgs[npgs].off + PAGE_SIZE) {
 224                         int count = inode->i_size - pgs[npgs].off;
 225                         void *addr = kmap(page);
 226                         pgs[npgs].count = count;
 227                         //POISON(addr, 0x7c, count);
 228                         memset(addr + count, 0, PAGE_SIZE - count);
 229                         kunmap(page);
 230                 }
 231
 232                 npgs++;
 233                 if (npgs == max_pages)
 234                         break;
 235
 236                 /*
 237                  * find pages ahead of us that we can read in.
 238                  * grab_cache_page waits on pages that are locked so
 239                  * we first try find_get_page, which doesn't.  this stops
 240                  * the worst case behaviour of racing threads waiting on
 241                  * each other, but doesn't remove it entirely.
 242                  */
 243                 for (index = page->index + 1, page = NULL;
 244                      page == NULL && index < end_index; index++) {
 245
 246                         /* see if the page already exists and needs updating */
 247                         page = find_get_page(inode->i_mapping, index);
 248                         if (page) {
 249                                 if (Page_Uptodate(page) || TryLockPage(page))
 250                                         goto out_release;
 251                                 if (!page->mapping || Page_Uptodate(page))
 252                                         goto out_unlock;
 253                         } else {
 254                                 /* ok, we have to create it.. */
 255                                 page = grab_cache_page(inode->i_mapping, index);
 256                                 if (page == NULL)
 257                                         continue;
 258                                 if (Page_Uptodate(page))
 259                                         goto out_unlock;
 260                         }
 261
 262                         break;
 263
 264                 out_unlock:
 265                         unlock_page(page);
 266                 out_release:
 267                         page_cache_release(page);
 268                         page = NULL;
 269                 }
 270
 271         } while (page);
 272
 273         set = ptlrpc_prep_set();
 274         if (set == NULL) {
 275                 CERROR("ENOMEM allocing request set\n");
 276                 rc = -ENOMEM;
 277         } else {
 278                 rc = obd_brw_async(OBD_BRW_READ, ll_i2obdconn(inode),
 279                                    ll_i2info(inode)->lli_smd, npgs, pgs,
 280                                    set, NULL);
 281                 if (rc == 0)
 282                         rc = ptlrpc_set_wait(set);
 283                 ptlrpc_set_destroy(set);
 284                 if (rc && rc != -EIO)
 285                         CERROR("error from obd_brw_async: rc = %d\n", rc);
 286         }
 287
 288         while (npgs-- > 0) {
 289                 page = pgs[npgs].pg;
 290
 291                 if (rc == 0)
 292                         SetPageUptodate(page);
 293                 unlock_page(page);
 294                 page_cache_release(page);
 295         }
 296
 297         kfree(pgs);
 298         RETURN(rc);
 299 } /* ll_readpage */
 300
 301 /* this isn't where truncate starts.   roughly:
 302  * sys_truncate->ll_setattr_raw->vmtruncate->ll_truncate
 303  * we grab the lock back in setattr_raw to avoid races. */
 304 void ll_truncate(struct inode *inode)
 305 {
 306         struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
 307         struct obdo oa = {0};
 308         int err;
 309         ENTRY;
 310         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p)\n", inode->i_ino,
 311                inode->i_generation, inode);
 312
 313         if (!lsm) {
 314                 /* object not yet allocated */
 315                 inode->i_mtime = inode->i_ctime = CURRENT_TIME;
 316                 EXIT;
 317                 return;
 318         }
 319
 320         /* vmtruncate just threw away our dirty pages, make sure
 321          * we don't think they're still dirty, being careful to round
 322          * i_size to the first whole page that was tossed */
 323         ll_remove_dirty(inode,
 324                         (inode->i_size + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT,
 325                         ~0);
 326
 327         oa.o_id = lsm->lsm_object_id;
 328         oa.o_mode = inode->i_mode;
 329         oa.o_valid = OBD_MD_FLID | OBD_MD_FLMODE | OBD_MD_FLTYPE;
 330
 331         CDEBUG(D_INFO, "calling punch for "LPX64" (all bytes after %Lu)\n",
 332                oa.o_id, inode->i_size);
 333
 334         /* truncate == punch from new size to absolute end of file */
 335         err = obd_punch(ll_i2obdconn(inode), &oa, lsm, inode->i_size,
 336                         OBD_OBJECT_EOF, NULL);
 337         if (err)
 338                 CERROR("obd_truncate fails (%d) ino %lu\n", err, inode->i_ino);
 339         else
 340                 obdo_to_inode(inode, &oa, oa.o_valid);
 341
 342         EXIT;
 343         return;
 344 } /* ll_truncate */
 345
 346 //#if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
 347
 348 static int ll_prepare_write(struct file *file, struct page *page, unsigned from,
 349                             unsigned to)
 350 {
 351         struct inode *inode = page->mapping->host;
 352         struct ll_inode_info *lli = ll_i2info(inode);
 353         struct lov_stripe_md *lsm = lli->lli_smd;
 354         obd_off offset = ((obd_off)page->index) << PAGE_SHIFT;
 355         struct brw_page pg;
 356         int rc = 0;
 357         ENTRY;
 358
 359         if (!PageLocked(page))
 360                 LBUG();
 361
 362         if (PageUptodate(page))
 363                 RETURN(0);
 364
 365         //POISON(addr + from, 0xca, to - from);
 366
 367         /* Check to see if we should return -EIO right away */
 368         pg.pg = page;
 369         pg.off = offset;
 370         pg.count = PAGE_SIZE;
 371         pg.flag = 0;
 372         rc = obd_brw(OBD_BRW_CHECK, ll_i2obdconn(inode), lsm, 1, &pg, NULL);
 373         if (rc)
 374                 RETURN(rc);
 375
 376         /* We're completely overwriting an existing page, so _don't_ set it up
 377          * to date until commit_write */
 378         if (from == 0 && to == PAGE_SIZE)
 379                 RETURN(0);
 380
 381         /* If are writing to a new page, no need to read old data.
 382          * the extent locking and getattr procedures in ll_file_write have
 383          * guaranteed that i_size is stable enough for our zeroing needs */
 384         if (inode->i_size <= offset) {
 385                 memset(kmap(page), 0, PAGE_SIZE);
 386                 kunmap(page);
 387                 GOTO(prepare_done, rc = 0);
 388         }
 389
 390         rc = ll_brw(OBD_BRW_READ, inode, page, 0);
 391
 392         EXIT;
 393  prepare_done:
 394         if (rc == 0)
 395                 SetPageUptodate(page);
 396
 397         return rc;
 398 }
 399
 400 /*
 401  * background file writeback.  This is called regularly from kupdated to write
 402  * dirty data, from kswapd when memory is low, and from filemap_fdatasync when
 403  * super blocks or inodes are synced..
 404  *
 405  * obd_brw errors down in _batch_writepage are ignored, so pages are always
 406  * unlocked.  Also, there is nobody to return an error code to from here - the
 407  * application may not even be running anymore.
 408  *
 409  * this should be async so that things like kswapd can have a chance to
 410  * free some more pages that our allocating writeback may need, but it isn't
 411  * yet.
 412  */
 413 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
 414 static int ll_writepage(struct page *page)
 415 {
 416         struct inode *inode = page->mapping->host;
 417         ENTRY;
 418
 419         CDEBUG(D_CACHE, "page %p [lau %d] inode %p\n", page,
 420                         PageLaunder(page), inode);
 421         LASSERT(PageLocked(page));
 422
 423         /* XXX should obd_brw errors trickle up? */
 424         ll_batch_writepage(inode, page);
 425         RETURN(0);
 426 }
 427
 428 /*
 429  * we really don't want to start writeback here, we want to give callers some
 430  * time to further dirty the pages before we write them out.
 431  */
 432 static int ll_commit_write(struct file *file, struct page *page,
 433                            unsigned from, unsigned to)
 434 {
 435         struct inode *inode = page->mapping->host;
 436         loff_t size;
 437         ENTRY;
 438
 439         LASSERT(inode == file->f_dentry->d_inode);
 440         LASSERT(PageLocked(page));
 441
 442         CDEBUG(D_INODE, "inode %p is writing page %p from %d to %d at %lu\n",
 443                inode, page, from, to, page->index);
 444         /* to match full page case in prepare_write */
 445         SetPageUptodate(page);
 446         /* mark the page dirty, put it on mapping->dirty,
 447          * mark the inode PAGES_DIRTY, put it on sb->dirty */
 448         if (!PageDirty(page))
 449                 INODE_IO_STAT_ADD(inode, dirty_misses, 1);
 450         else
 451                 INODE_IO_STAT_ADD(inode, dirty_hits, 1);
 452
 453         size = (((obd_off)page->index) << PAGE_SHIFT) + to;
 454         if (size > inode->i_size)
 455                 inode->i_size = size;
 456
 457         /* XXX temporary, bug 1286 */
 458         {
 459                 struct ll_dirty_offsets *lldo = &ll_i2info(inode)->lli_dirty;
 460                 int rc;
 461                 if ((lldo->do_num_dirty * PAGE_CACHE_SIZE) > 10 * 1024 * 1024) {
 462                         rc = ll_batch_writepage(inode, page);
 463                         lock_page(page); /* caller expects to unlock */
 464                         RETURN(rc);
 465                 }
 466         }
 467
 468         set_page_dirty(page);
 469         ll_record_dirty(inode, page->index);
 470
 471         RETURN(0);
 472 } /* ll_commit_write */
 473 #else
 474 static int ll_writepage(struct page *page,
 475                         struct writeback_control *wbc)
 476 {
 477
 478         return 0;
 479 }
 480 static int ll_commit_write(struct file *file, struct page *page,
 481                            unsigned from, unsigned to)
 482 {
 483         return 0;
 484 }
 485 #endif
 486
 487 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
 488 static int ll_direct_IO(int rw, struct inode *inode, struct kiobuf *iobuf,
 489                         unsigned long blocknr, int blocksize)
 490 {
 491         struct ll_inode_info *lli = ll_i2info(inode);
 492         struct lov_stripe_md *lsm = lli->lli_smd;
 493         struct brw_page *pga;
 494         struct ptlrpc_request_set *set;
 495         int length, i, flags, rc = 0;
 496         loff_t offset;
 497         ENTRY;
 498
 499         if (!lsm || !lsm->lsm_object_id)
 500                 RETURN(-ENOMEM);
 501
 502         if ((iobuf->offset & (blocksize - 1)) ||
 503             (iobuf->length & (blocksize - 1)))
 504                 RETURN(-EINVAL);
 505
 506         set = ptlrpc_prep_set();
 507         if (set == NULL)
 508                 RETURN(-ENOMEM);
 509
 510         OBD_ALLOC(pga, sizeof(*pga) * iobuf->nr_pages);
 511         if (!pga) {
 512                 ptlrpc_set_destroy(set);
 513                 RETURN(-ENOMEM);
 514         }
 515
 516         flags = (rw == WRITE ? OBD_BRW_CREATE : 0) /* | OBD_BRW_DIRECTIO */;
 517         offset = ((obd_off)blocknr << inode->i_blkbits);
 518         length = iobuf->length;
 519
 520         for (i = 0, length = iobuf->length; length > 0;
 521              length -= pga[i].count, offset += pga[i].count, i++) { /*i last!*/
 522                 pga[i].pg = iobuf->maplist[i];
 523                 pga[i].off = offset;
 524                 /* To the end of the page, or the length, whatever is less */
 525                 pga[i].count = min_t(int, PAGE_SIZE - (offset & ~PAGE_MASK),
 526                                      length);
 527                 pga[i].flag = flags;
 528                 if (rw == READ) {
 529                         //POISON(kmap(iobuf->maplist[i]), 0xc5, PAGE_SIZE);
 530                         //kunmap(iobuf->maplist[i]);
 531                 }
 532         }
 533
 534         rc = obd_brw_async(rw == WRITE ? OBD_BRW_WRITE : OBD_BRW_READ,
 535                            ll_i2obdconn(inode), lsm, iobuf->nr_pages, pga, set,
 536                            NULL);
 537         if (rc) {
 538                 CDEBUG(rc == -ENOSPC ? D_INODE : D_ERROR,
 539                        "error from obd_brw_async: rc = %d\n", rc);
 540         } else {
 541                 rc = ptlrpc_set_wait(set);
 542                 if (rc)
 543                         CERROR("error from callback: rc = %d\n", rc);
 544         }
 545         ptlrpc_set_destroy(set);
 546         if (rc == 0)
 547                 rc = iobuf->length;
 548
 549         OBD_FREE(pga, sizeof(*pga) * iobuf->nr_pages);
 550         RETURN(rc);
 551 }
 552 #endif
 553
 554 //#endif
 555
 556 struct address_space_operations ll_aops = {
 557         readpage: ll_readpage,
 558 #if (LINUX_VERSION_CODE <= KERNEL_VERSION(2,5,0))
 559         direct_IO: ll_direct_IO,
 560 #endif
 561         writepage: ll_writepage,
 562         sync_page: block_sync_page,
 563         prepare_write: ll_prepare_write,
 564         commit_write: ll_commit_write,
 565         bmap: NULL
 566 //#endif
 567 };