lustre/llite/iod.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  *  Copyright (c) 2002, 2003 Cluster File Systems, Inc.
   5  *
   6  *   This file is part of Lustre, http://www.lustre.org.
   7  *
   8  *   Lustre is free software; you can redistribute it and/or
   9  *   modify it under the terms of version 2 of the GNU General Public
  10  *   License as published by the Free Software Foundation.
  11  *
  12  *   Lustre is distributed in the hope that it will be useful,
  13  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  14  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  15  *   GNU General Public License for more details.
  16  *
  17  *   You should have received a copy of the GNU General Public License
  18  *   along with Lustre; if not, write to the Free Software
  19  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  20  *
  21  *  Copyright (C) 2002, 2003  Cluster File Systems, Inc
  22  *
  23  *  this started as an implementation of an io daemon that woke regularly
  24  *  to force writeback.. the throttling in prepare_write and kupdate's usual
  25  *  writeback pressure got rid of our thread, but the file name remains.
  26  */
  27
  28 #include <linux/version.h>
  29 #include <linux/config.h>
  30 #include <linux/module.h>
  31 #include <linux/fs.h>
  32 #include <linux/stat.h>
  33 #include <linux/sched.h>
  34 #include <linux/smp_lock.h>
  35 #include <linux/kmod.h>
  36 #include <linux/pagemap.h>
  37 #include <linux/mm.h>
  38 #include <linux/rbtree.h>
  39 #include <linux/seq_file.h>
  40 #include <linux/time.h>
  41
  42 /* PG_inactive_clean is shorthand for rmap, we want free_high/low here.. */
  43 #ifdef PG_inactive_clean
  44 #include <linux/mm_inline.h>
  45 #endif
  46
  47 #define DEBUG_SUBSYSTEM S_LLITE
  48 #include <linux/lustre_lite.h>
  49 #include "llite_internal.h"
  50
  51 #ifndef list_for_each_prev_safe
  52 #define list_for_each_prev_safe(pos, n, head) \
  53         for (pos = (head)->prev, n = pos->prev; pos != (head); \
  54                 pos = n, n = pos->prev )
  55 #endif
  56
  57 extern spinlock_t inode_lock;
  58
  59 /*
  60  * check to see if we're racing with truncate and put the page in
  61  * the brw_page array.  returns 0 if there is more room and 1
  62  * if the array is full.
  63  */
  64 static int llwp_consume_page(struct ll_writeback_pages *llwp,
  65                              struct inode *inode, struct page *page)
  66 {
  67         obd_off off = ((obd_off)page->index) << PAGE_SHIFT;
  68         struct brw_page *pg;
  69
  70         /* we raced with truncate? */
  71         if ( off >= inode->i_size ) {
  72                 int rc;
  73                 rc = ll_clear_dirty_pages(ll_i2obdconn(inode),
  74                                           ll_i2info(inode)->lli_smd,
  75                                           page->index, page->index);
  76
  77                 LASSERT(rc == 0);
  78                 CDEBUG(D_CACHE, "offset "LPU64" (index %lu) > i_size %llu\n",
  79                        off, page->index, inode->i_size);
  80                 unlock_page(page);
  81                 return 0;
  82         }
  83
  84         page_cache_get(page);
  85         pg = &llwp->pga[llwp->npgs];
  86         llwp->npgs++;
  87         LASSERT(llwp->npgs <= llwp->max);
  88
  89         pg->pg = page;
  90         pg->off = off;
  91         pg->flag = OBD_BRW_CREATE|OBD_BRW_FROM_GRANT;
  92         pg->count = PAGE_CACHE_SIZE;
  93
  94         /* catch partial writes for files that end mid-page */
  95         if (pg->off + pg->count > inode->i_size)
  96                 pg->count = inode->i_size & ~PAGE_CACHE_MASK;
  97
  98         /*
  99          * matches ptlrpc_bulk_get assert that trickles down
 100          * from a 0 page length going through niobuf and into
 101          * the buffer regions being posted
 102          */
 103         LASSERT(pg->count >= 0);
 104
 105         CDEBUG(D_CACHE, "brw_page %p: off "LPU64" cnt %d, page %p: ind %ld"
 106                         " i_size: %llu\n", pg, pg->off, pg->count, page,
 107                         page->index, inode->i_size);
 108
 109         return llwp->npgs == llwp->max;
 110 }
 111
 112 /*
 113  * returns the number of pages that it added to the pgs array
 114  *
 115  * this duplicates filemap_fdatasync and gives us an opportunity to grab lots
 116  * of dirty pages..
 117  */
 118 static void ll_get_dirty_pages(struct inode *inode,
 119                                struct ll_writeback_pages *llwp)
 120 {
 121         struct address_space *mapping = inode->i_mapping;
 122         struct page *page;
 123         struct list_head *pos, *n;
 124         ENTRY;
 125
 126         PGCACHE_WRLOCK(mapping);
 127
 128         list_for_each_prev_safe(pos, n, &mapping->dirty_pages) {
 129                 page = list_entry(pos, struct page, list);
 130
 131                 if (TryLockPage(page))
 132                         continue;
 133
 134                 list_del(&page->list);
 135                 list_add(&page->list, &mapping->locked_pages);
 136
 137                 if (!PageDirty(page)) {
 138                         unlock_page(page);
 139                         continue;
 140                 }
 141                 ClearPageDirty(page);
 142
 143                 if (llwp_consume_page(llwp, inode, page) != 0)
 144                         break;
 145         }
 146
 147         PGCACHE_WRUNLOCK(mapping);
 148         EXIT;
 149 }
 150
 151 static void ll_writeback(struct inode *inode, struct obdo *oa,
 152                          struct ll_writeback_pages *llwp)
 153 {
 154         struct ptlrpc_request_set *set;
 155         int rc, i;
 156         ENTRY;
 157
 158         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),bytes=%u\n",
 159                inode->i_ino, inode->i_generation, inode,
 160                ((llwp->npgs-1) << PAGE_SHIFT) + llwp->pga[llwp->npgs-1].count);
 161
 162         SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
 163         set = ptlrpc_prep_set();
 164         if (set == NULL) {
 165                 CERROR ("Can't create request set\n");
 166                 rc = -ENOMEM;
 167         } else {
 168                 rc = obd_brw_async(OBD_BRW_WRITE, ll_i2obdconn(inode), oa,
 169                                    ll_i2info(inode)->lli_smd, llwp->npgs,
 170                                    llwp->pga, set, NULL);
 171                 if (rc == 0)
 172                         rc = ptlrpc_set_wait(set);
 173                 if (rc == 0)
 174                         obdo_refresh_inode(inode, oa,
 175                                            oa->o_valid & ~OBD_MD_FLSIZE);
 176                 ptlrpc_set_destroy (set);
 177         }
 178         /*
 179          * b=1038, we need to pass _brw errors up so that writeback
 180          * doesn't get stuck in recovery leaving processes stuck in
 181          * D waiting for pages
 182          */
 183         if (rc) {
 184                 CERROR("error from obd_brw_async: rc = %d\n", rc);
 185                 lprocfs_counter_add(ll_i2sbi(inode)->ll_stats,
 186                                     LPROC_LL_WB_FAIL, llwp->npgs);
 187         } else {
 188                 lprocfs_counter_add(ll_i2sbi(inode)->ll_stats,
 189                                     LPROC_LL_WB_OK, (llwp->npgs));
 190         }
 191
 192         for (i = 0 ; i < llwp->npgs ; i++) {
 193                 struct page *page = llwp->pga[i].pg;
 194
 195                 CDEBUG(D_CACHE, "finished page %p at index %lu\n", page,
 196                        page->index);
 197                 LASSERT(PageLocked(page));
 198
 199                 rc = ll_clear_dirty_pages(ll_i2obdconn(inode),
 200                                           ll_i2info(inode)->lli_smd,
 201                                           page->index, page->index);
 202                 LASSERT(rc == 0);
 203                 unlock_page(page);
 204                 page_cache_release(page);
 205         }
 206
 207         EXIT;
 208 }
 209
 210 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
 211
 212 #ifndef PG_inactive_clean
 213 #ifdef CONFIG_DISCONTIGMEM
 214 #error "sorry, we don't support DISCONTIGMEM yet"
 215 #endif
 216
 217 /*
 218  * __alloc_pages marks a zone as needing balancing if an allocation is
 219  * performed when the zone has fewer free pages than its 'low' water
 220  * mark.  its cleared when try_to_free_pages makes progress.
 221  */
 222 static int zones_need_balancing(void)
 223 {
 224         pg_data_t * pgdat;
 225         zone_t *zone;
 226         int i;
 227
 228         for ( pgdat = pgdat_list ; pgdat != NULL ; pgdat = pgdat->node_next ) {
 229                 for ( i = pgdat->nr_zones-1 ; i >= 0 ; i-- ) {
 230                         zone = &pgdat->node_zones[i];
 231
 232                         if ( zone->need_balance )
 233                                 return 1;
 234                 }
 235         }
 236         return 0;
 237 }
 238 #endif
 239 /* 2.4 doesn't give us a way to find out how many pages we have
 240  * cached 'cause we're not using buffer_heads.  we are very
 241  * conservative here and flush the superblock of all dirty data
 242  * when the vm (rmap or stock) thinks that it is running low
 243  * and kswapd would have done work.  kupdated isn't good enough
 244  * because writers (dbench) can dirty _very quickly_, and we
 245  * allocate under writepage..
 246  *
 247  * 2.5 gets this right, see the {inc,dec}_page_state(nr_dirty, )
 248  */
 249 static int should_writeback(void)
 250 {
 251 #ifdef PG_inactive_clean
 252         if (free_high(ALL_ZONES) > 0 || free_low(ANY_ZONE) > 0)
 253 #else
 254         if (zones_need_balancing())
 255 #endif
 256                 return 1;
 257         return 0;
 258 }
 259
 260 static int ll_alloc_brw(struct inode *inode, struct ll_writeback_pages *llwp)
 261 {
 262         memset(llwp, 0, sizeof(struct ll_writeback_pages));
 263
 264         llwp->max = inode->i_blksize >> PAGE_CACHE_SHIFT;
 265         if (llwp->max == 0) {
 266                 CERROR("forcing llwp->max to 1.  blksize: %lu\n",
 267                        inode->i_blksize);
 268                 llwp->max = 1;
 269         }
 270         llwp->pga = kmalloc(llwp->max * sizeof(*llwp->pga), GFP_ATOMIC);
 271         if (llwp->pga == NULL)
 272                 RETURN(-ENOMEM);
 273         RETURN(0);
 274 }
 275
 276 int ll_check_dirty(struct super_block *sb)
 277 {
 278         unsigned long old_flags; /* hack? */
 279         int making_progress;
 280         struct inode *inode;
 281         struct obdo oa;
 282         int rc = 0;
 283         ENTRY;
 284
 285         if (!should_writeback())
 286                 return 0;
 287
 288         old_flags = current->flags;
 289         current->flags |= PF_MEMALLOC;
 290
 291         spin_lock(&inode_lock);
 292
 293         /*
 294          * first we try and write back dirty pages from dirty inodes
 295          * until the VM thinkgs we're ok again..
 296          */
 297         do {
 298                 struct ll_writeback_pages llwp;
 299                 struct list_head *pos;
 300                 inode = NULL;
 301                 making_progress = 0;
 302
 303                 list_for_each_prev(pos, &sb->s_dirty) {
 304                         inode = list_entry(pos, struct inode, i_list);
 305
 306                         if (!(inode->i_state & I_DIRTY_PAGES)) {
 307                                 inode = NULL;
 308                                 continue;
 309                         }
 310                         break;
 311                 }
 312
 313                 if (inode == NULL)
 314                         break;
 315
 316                 /* duplicate __sync_one, *sigh* */
 317                 list_del(&inode->i_list);
 318                 list_add(&inode->i_list, &inode->i_sb->s_locked_inodes);
 319                 inode->i_state |= I_LOCK;
 320                 inode->i_state &= ~I_DIRTY_PAGES;
 321
 322                 spin_unlock(&inode_lock);
 323
 324                 rc = ll_alloc_brw(inode, &llwp);
 325                 if (rc != 0)
 326                         GOTO(cleanup, rc);
 327
 328                 do {
 329                         llwp.npgs = 0;
 330                         ll_get_dirty_pages(inode, &llwp);
 331                         if (llwp.npgs) {
 332                                 oa.o_id =
 333                                       ll_i2info(inode)->lli_smd->lsm_object_id;
 334                                 oa.o_valid = OBD_MD_FLID;
 335                                 obdo_from_inode(&oa, inode,
 336                                                 OBD_MD_FLTYPE | OBD_MD_FLATIME|
 337                                                 OBD_MD_FLMTIME| OBD_MD_FLCTIME);
 338                                 lprocfs_counter_add(ll_i2sbi(inode)->ll_stats,
 339                                                     LPROC_LL_WB_PRESSURE,
 340                                                     llwp.npgs);
 341                                 ll_writeback(inode, &oa, &llwp);
 342                                 rc += llwp.npgs;
 343                                 making_progress = 1;
 344                         }
 345                 } while (llwp.npgs && should_writeback());
 346
 347                 spin_lock(&inode_lock);
 348
 349                 if (!list_empty(&inode->i_mapping->dirty_pages))
 350                         inode->i_state |= I_DIRTY_PAGES;
 351
 352                 inode->i_state &= ~I_LOCK;
 353                 /*
 354                  * we are sneaky and leave the inode on the dirty list,
 355                  * even though it might not still be..
 356                  */
 357                 if (!(inode->i_state & I_FREEING)) {
 358                         list_del(&inode->i_list);
 359                         list_add(&inode->i_list, &inode->i_sb->s_dirty);
 360                 }
 361                 wake_up(&inode->i_wait);
 362                 kfree(llwp.pga);
 363         } while (making_progress && should_writeback());
 364
 365         /*
 366          * and if that didn't work, we sleep on any data that might
 367          * be under writeback..
 368          */
 369         while (should_writeback()) {
 370                 if (list_empty(&sb->s_locked_inodes))
 371                         break;
 372
 373                 inode = list_entry(sb->s_locked_inodes.next, struct inode,
 374                                    i_list);
 375
 376                 atomic_inc(&inode->i_count); /* XXX hack? */
 377                 spin_unlock(&inode_lock);
 378                 wait_event(inode->i_wait, !(inode->i_state & I_LOCK));
 379                 iput(inode);
 380                 spin_lock(&inode_lock);
 381         }
 382
 383         spin_unlock(&inode_lock);
 384
 385 cleanup:
 386         current->flags = old_flags;
 387
 388         RETURN(rc);
 389 }
 390 #endif /* linux 2.5 */
 391
 392 int ll_batch_writepage(struct inode *inode, struct obdo *oa, struct page *page)
 393 {
 394         unsigned long old_flags; /* hack? */
 395         struct ll_writeback_pages llwp;
 396         int rc = 0;
 397         ENTRY;
 398
 399         SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
 400         old_flags = current->flags;
 401         current->flags |= PF_MEMALLOC;
 402         rc = ll_alloc_brw(inode, &llwp);
 403         if (rc != 0)
 404                 GOTO(restore_flags, rc);
 405
 406         if (llwp_consume_page(&llwp, inode, page) == 0)
 407                 ll_get_dirty_pages(inode, &llwp);
 408
 409         if (llwp.npgs) {
 410                 lprocfs_counter_add(ll_i2sbi(inode)->ll_stats,
 411                                     LPROC_LL_WB_WRITEPAGE, llwp.npgs);
 412                 ll_writeback(inode, oa, &llwp);
 413         }
 414         kfree(llwp.pga);
 415
 416 restore_flags:
 417         current->flags = old_flags;
 418         RETURN(rc);
 419 }