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 #include "llite_internal.h"
  42
  43 /* PG_inactive_clean is shorthand for rmap, we want free_high/low here.. */
  44 #ifdef PG_inactive_clean
  45 #include <linux/mm_inline.h>
  46 #endif
  47
  48 #define DEBUG_SUBSYSTEM S_LLITE
  49 #include <linux/lustre_lite.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 struct ll_writeback_pages {
  60         obd_count npgs, max;
  61         struct brw_page *pga;
  62 };
  63
  64 /*
  65  * check to see if we're racing with truncate and put the page in
  66  * the brw_page array.  returns 0 if there is more room and 1
  67  * if the array is full.
  68  */
  69 static int llwp_consume_page(struct ll_writeback_pages *llwp,
  70                              struct inode *inode, struct page *page)
  71 {
  72         obd_off off = ((obd_off)page->index) << PAGE_SHIFT;
  73         struct brw_page *pg;
  74
  75         /* we raced with truncate? */
  76         if ( off >= inode->i_size ) {
  77                 int rc;
  78                 rc = ll_clear_dirty_pages(ll_i2obdconn(inode),
  79                                           ll_i2info(inode)->lli_smd,
  80                                           page->index, page->index);
  81
  82                 LASSERT(rc == 0);
  83                 CDEBUG(D_CACHE, "offset "LPU64" (index %lu) > i_size %llu\n",
  84                        off, page->index, inode->i_size);
  85                 unlock_page(page);
  86                 return 0;
  87         }
  88
  89         page_cache_get(page);
  90         pg = &llwp->pga[llwp->npgs];
  91         llwp->npgs++;
  92         LASSERT(llwp->npgs <= llwp->max);
  93
  94         pg->pg = page;
  95         pg->off = off;
  96         pg->flag = OBD_BRW_CREATE|OBD_BRW_FROM_GRANT;
  97         pg->count = PAGE_CACHE_SIZE;
  98
  99         /* catch partial writes for files that end mid-page */
 100         if (pg->off + pg->count > inode->i_size)
 101                 pg->count = inode->i_size & ~PAGE_CACHE_MASK;
 102
 103         /*
 104          * matches ptlrpc_bulk_get assert that trickles down
 105          * from a 0 page length going through niobuf and into
 106          * the buffer regions being posted
 107          */
 108         LASSERT(pg->count >= 0);
 109
 110         CDEBUG(D_CACHE, "brw_page %p: off "LPU64" cnt %d, page %p: ind %ld"
 111                         " i_size: %llu\n", pg, pg->off, pg->count, page,
 112                         page->index, inode->i_size);
 113
 114         return llwp->npgs == llwp->max;
 115 }
 116
 117 /*
 118  * returns the number of pages that it added to the pgs array
 119  *
 120  * this duplicates filemap_fdatasync and gives us an opportunity to grab lots
 121  * of dirty pages..
 122  */
 123 static void ll_get_dirty_pages(struct inode *inode,
 124                                struct ll_writeback_pages *llwp)
 125 {
 126         struct address_space *mapping = inode->i_mapping;
 127         struct page *page;
 128         struct list_head *pos, *n;
 129         ENTRY;
 130
 131         PGCACHE_WRLOCK(mapping);
 132
 133         list_for_each_prev_safe(pos, n, &mapping->dirty_pages) {
 134                 page = list_entry(pos, struct page, list);
 135
 136                 if (TryLockPage(page))
 137                         continue;
 138
 139                 list_del(&page->list);
 140                 list_add(&page->list, &mapping->locked_pages);
 141
 142                 if ( ! PageDirty(page) ) {
 143                         unlock_page(page);
 144                         continue;
 145                 }
 146                 ClearPageDirty(page);
 147
 148                 if ( llwp_consume_page(llwp, inode, page) != 0)
 149                         break;
 150         }
 151
 152         PGCACHE_WRUNLOCK(mapping);
 153         EXIT;
 154 }
 155
 156 static void ll_writeback(struct inode *inode, struct ll_writeback_pages *llwp)
 157 {
 158         int rc, i;
 159         struct ptlrpc_request_set *set;
 160         ENTRY;
 161
 162         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),bytes=%u\n",
 163                inode->i_ino, inode->i_generation, inode,
 164                ((llwp->npgs-1) << PAGE_SHIFT) + llwp->pga[llwp->npgs-1].count);
 165
 166         set = ptlrpc_prep_set();
 167         if (set == NULL) {
 168                 CERROR ("Can't create request set\n");
 169                 rc = -ENOMEM;
 170         } else {
 171                 rc = obd_brw_async(OBD_BRW_WRITE, ll_i2obdconn(inode),
 172                                    ll_i2info(inode)->lli_smd, llwp->npgs,
 173                                    llwp->pga, set, NULL);
 174                 if (rc == 0)
 175                         rc = ptlrpc_set_wait (set);
 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         int rc = 0;
 282         ENTRY;
 283
 284         if (!should_writeback())
 285                 return 0;
 286
 287         old_flags = current->flags;
 288         current->flags |= PF_MEMALLOC;
 289
 290         spin_lock(&inode_lock);
 291
 292         /*
 293          * first we try and write back dirty pages from dirty inodes
 294          * until the VM thinkgs we're ok again..
 295          */
 296         do {
 297                 struct ll_writeback_pages llwp;
 298                 struct list_head *pos;
 299                 inode = NULL;
 300                 making_progress = 0;
 301
 302                 list_for_each_prev(pos, &sb->s_dirty) {
 303                         inode = list_entry(pos, struct inode, i_list);
 304
 305                         if (!(inode->i_state & I_DIRTY_PAGES)) {
 306                                 inode = NULL;
 307                                 continue;
 308                         }
 309                         break;
 310                 }
 311
 312                 if (inode == NULL)
 313                         break;
 314
 315                 /* duplicate __sync_one, *sigh* */
 316                 list_del(&inode->i_list);
 317                 list_add(&inode->i_list, &inode->i_sb->s_locked_inodes);
 318                 inode->i_state |= I_LOCK;
 319                 inode->i_state &= ~I_DIRTY_PAGES;
 320
 321                 spin_unlock(&inode_lock);
 322
 323                 rc = ll_alloc_brw(inode, &llwp);
 324                 if (rc != 0)
 325                         GOTO(cleanup, rc);
 326
 327                 do {
 328                         llwp.npgs = 0;
 329                         ll_get_dirty_pages(inode, &llwp);
 330                         if (llwp.npgs) {
 331                                lprocfs_counter_add(ll_i2sbi(inode)->ll_stats,
 332                                                    LPROC_LL_WB_PRESSURE,
 333                                                    llwp.npgs);
 334                                ll_writeback(inode, &llwp);
 335                                rc += llwp.npgs;
 336                                making_progress = 1;
 337                         }
 338                 } while (llwp.npgs && should_writeback());
 339
 340                 spin_lock(&inode_lock);
 341
 342                 if (!list_empty(&inode->i_mapping->dirty_pages))
 343                         inode->i_state |= I_DIRTY_PAGES;
 344
 345                 inode->i_state &= ~I_LOCK;
 346                 /*
 347                  * we are sneaky and leave the inode on the dirty list,
 348                  * even though it might not still be..
 349                  */
 350                 if (!(inode->i_state & I_FREEING)) {
 351                         list_del(&inode->i_list);
 352                         list_add(&inode->i_list, &inode->i_sb->s_dirty);
 353                 }
 354                 wake_up(&inode->i_wait);
 355                 kfree(llwp.pga);
 356         } while (making_progress && should_writeback());
 357
 358         /*
 359          * and if that didn't work, we sleep on any data that might
 360          * be under writeback..
 361          */
 362         while (should_writeback()) {
 363                 if (list_empty(&sb->s_locked_inodes))
 364                         break;
 365
 366                 inode = list_entry(sb->s_locked_inodes.next, struct inode,
 367                                    i_list);
 368
 369                 atomic_inc(&inode->i_count); /* XXX hack? */
 370                 spin_unlock(&inode_lock);
 371                 wait_event(inode->i_wait, !(inode->i_state & I_LOCK));
 372                 iput(inode);
 373                 spin_lock(&inode_lock);
 374         }
 375
 376         spin_unlock(&inode_lock);
 377
 378 cleanup:
 379         current->flags = old_flags;
 380
 381         RETURN(rc);
 382 }
 383 #endif /* linux 2.5 */
 384
 385 int ll_batch_writepage(struct inode *inode, struct page *page)
 386 {
 387         unsigned long old_flags; /* hack? */
 388         struct ll_writeback_pages llwp;
 389         int rc = 0;
 390         ENTRY;
 391
 392         old_flags = current->flags;
 393         current->flags |= PF_MEMALLOC;
 394         rc = ll_alloc_brw(inode, &llwp);
 395         if (rc != 0)
 396                 GOTO(restore_flags, rc);
 397
 398         if (llwp_consume_page(&llwp, inode, page) == 0)
 399                 ll_get_dirty_pages(inode, &llwp);
 400
 401         if (llwp.npgs) {
 402                 lprocfs_counter_add(ll_i2sbi(inode)->ll_stats,
 403                                     LPROC_LL_WB_WRITEPAGE, llwp.npgs);
 404                 ll_writeback(inode, &llwp);
 405         }
 406         kfree(llwp.pga);
 407
 408 restore_flags:
 409         current->flags = old_flags;
 410         RETURN(rc);
 411 }