2 fs/ext3/inode.c | 116 ++++++++++++++++++++++
3 fs/ext3/super.c | 230 +++++++++++++++++++++++++++++++++++++++++++++
4 include/linux/ext3_fs.h | 5
5 include/linux/ext3_fs_sb.h | 10 +
6 5 files changed, 365 insertions(+)
8 Index: linux/fs/ext3/super.c
9 ===================================================================
10 --- linux.orig/fs/ext3/super.c Mon Feb 2 20:57:35 2004
11 +++ linux/fs/ext3/super.c Mon Feb 2 20:58:05 2004
16 +#ifdef EXT3_DELETE_THREAD
18 + * Delete inodes in a loop until there are no more to be deleted.
19 + * Normally, we run in the background doing the deletes and sleeping again,
20 + * and clients just add new inodes to be deleted onto the end of the list.
21 + * If someone is concerned about free space (e.g. block allocation or similar)
22 + * then they can sleep on s_delete_waiter_queue and be woken up when space
25 +int ext3_delete_thread(void *data)
27 + struct super_block *sb = data;
28 + struct ext3_sb_info *sbi = EXT3_SB(sb);
29 + struct task_struct *tsk = current;
31 + /* Almost like daemonize, but not quite */
36 + exit_files(current);
39 + sprintf(tsk->comm, "kdelext3-%s", kdevname(sb->s_dev));
40 + sigfillset(&tsk->blocked);
42 + /*tsk->flags |= PF_KERNTHREAD;*/
44 + INIT_LIST_HEAD(&sbi->s_delete_list);
45 + wake_up(&sbi->s_delete_waiter_queue);
46 + ext3_debug("delete thread on %s started\n", kdevname(sb->s_dev));
50 + wait_event_interruptible(sbi->s_delete_thread_queue,
51 + !list_empty(&sbi->s_delete_list) ||
52 + !test_opt(sb, ASYNCDEL));
53 + ext3_debug("%s woken up: %lu inodes, %lu blocks\n",
54 + tsk->comm,sbi->s_delete_inodes,sbi->s_delete_blocks);
56 + spin_lock(&sbi->s_delete_lock);
57 + if (list_empty(&sbi->s_delete_list)) {
58 + clear_opt(sbi->s_mount_opt, ASYNCDEL);
59 + memset(&sbi->s_delete_list, 0,
60 + sizeof(sbi->s_delete_list));
61 + spin_unlock(&sbi->s_delete_lock);
62 + ext3_debug("delete thread on %s exiting\n",
63 + kdevname(sb->s_dev));
64 + wake_up(&sbi->s_delete_waiter_queue);
68 + while (!list_empty(&sbi->s_delete_list)) {
69 + struct inode *inode=list_entry(sbi->s_delete_list.next,
70 + struct inode, i_dentry);
71 + unsigned long blocks = inode->i_blocks >>
72 + (inode->i_blkbits - 9);
74 + list_del_init(&inode->i_dentry);
75 + spin_unlock(&sbi->s_delete_lock);
76 + ext3_debug("%s delete ino %lu blk %lu\n",
77 + tsk->comm, inode->i_ino, blocks);
81 + spin_lock(&sbi->s_delete_lock);
82 + sbi->s_delete_blocks -= blocks;
83 + sbi->s_delete_inodes--;
85 + if (sbi->s_delete_blocks != 0 || sbi->s_delete_inodes != 0) {
86 + ext3_warning(sb, __FUNCTION__,
87 + "%lu blocks, %lu inodes on list?\n",
88 + sbi->s_delete_blocks,sbi->s_delete_inodes);
89 + sbi->s_delete_blocks = 0;
90 + sbi->s_delete_inodes = 0;
92 + spin_unlock(&sbi->s_delete_lock);
93 + wake_up(&sbi->s_delete_waiter_queue);
99 +static void ext3_start_delete_thread(struct super_block *sb)
101 + struct ext3_sb_info *sbi = EXT3_SB(sb);
104 + spin_lock_init(&sbi->s_delete_lock);
105 + init_waitqueue_head(&sbi->s_delete_thread_queue);
106 + init_waitqueue_head(&sbi->s_delete_waiter_queue);
108 + if (!test_opt(sb, ASYNCDEL))
111 + rc = kernel_thread(ext3_delete_thread, sb, CLONE_VM | CLONE_FILES);
113 + printk(KERN_ERR "EXT3-fs: cannot start delete thread: rc %d\n",
116 + wait_event(sbi->s_delete_waiter_queue, sbi->s_delete_list.next);
119 +static void ext3_stop_delete_thread(struct ext3_sb_info *sbi)
121 + if (sbi->s_delete_list.next == 0) /* thread never started */
124 + clear_opt(sbi->s_mount_opt, ASYNCDEL);
125 + wake_up(&sbi->s_delete_thread_queue);
126 + wait_event(sbi->s_delete_waiter_queue,
127 + sbi->s_delete_list.next == 0 && sbi->s_delete_inodes == 0);
130 +/* Instead of playing games with the inode flags, destruction, etc we just
131 + * create a new inode locally and put it on a list for the truncate thread.
132 + * We need large parts of the inode struct in order to complete the
133 + * truncate and unlink, so we may as well just have a real inode to do it.
135 + * If we have any problem deferring the delete, just delete it right away.
136 + * If we defer it, we also mark how many blocks it would free, so that we
137 + * can keep the statfs data correct, and we know if we should sleep on the
138 + * delete thread when we run out of space.
140 +static void ext3_delete_inode_thread(struct inode *old_inode)
142 + struct ext3_sb_info *sbi = EXT3_SB(old_inode->i_sb);
143 + struct ext3_inode_info *nei, *oei = EXT3_I(old_inode);
144 + struct inode *new_inode;
145 + unsigned long blocks = old_inode->i_blocks >> (old_inode->i_blkbits-9);
147 + if (is_bad_inode(old_inode)) {
148 + clear_inode(old_inode);
152 + if (!test_opt(old_inode->i_sb, ASYNCDEL) || !sbi->s_delete_list.next)
155 + /* We may want to delete the inode immediately and not defer it */
156 + if (IS_SYNC(old_inode) || blocks <= EXT3_NDIR_BLOCKS)
159 + /* We can't use the delete thread as-is during real orphan recovery,
160 + * as we add to the orphan list here, causing ext3_orphan_cleanup()
161 + * to loop endlessly. It would be nice to do so, but needs work.
163 + if (oei->i_state & EXT3_STATE_DELETE ||
164 + sbi->s_mount_state & EXT3_ORPHAN_FS) {
165 + ext3_debug("doing deferred inode %lu delete (%lu blocks)\n",
166 + old_inode->i_ino, blocks);
170 + /* We can iget this inode again here, because our caller has unhashed
171 + * old_inode, so new_inode will be in a different inode struct.
173 + * We need to ensure that the i_orphan pointers in the other inodes
174 + * point at the new inode copy instead of the old one so the orphan
175 + * list doesn't get corrupted when the old orphan inode is freed.
177 + down(&sbi->s_orphan_lock);
179 + sbi->s_mount_state |= EXT3_ORPHAN_FS;
180 + new_inode = iget(old_inode->i_sb, old_inode->i_ino);
181 + sbi->s_mount_state &= ~EXT3_ORPHAN_FS;
182 + if (is_bad_inode(new_inode)) {
183 + printk(KERN_WARNING "read bad inode %lu\n", old_inode->i_ino);
188 + up(&sbi->s_orphan_lock);
189 + ext3_debug("delete inode %lu directly (bad read)\n",
193 + J_ASSERT(new_inode != old_inode);
195 + J_ASSERT(!list_empty(&oei->i_orphan));
197 + nei = EXT3_I(new_inode);
198 + /* Ugh. We need to insert new_inode into the same spot on the list
199 + * as old_inode was, to ensure the in-memory orphan list is still
200 + * in the same order as the on-disk orphan list (badness otherwise).
202 + nei->i_orphan = oei->i_orphan;
203 + nei->i_orphan.next->prev = &nei->i_orphan;
204 + nei->i_orphan.prev->next = &nei->i_orphan;
205 + nei->i_state |= EXT3_STATE_DELETE;
206 + up(&sbi->s_orphan_lock);
208 + clear_inode(old_inode);
210 + spin_lock(&sbi->s_delete_lock);
211 + J_ASSERT(list_empty(&new_inode->i_dentry));
212 + list_add_tail(&new_inode->i_dentry, &sbi->s_delete_list);
213 + sbi->s_delete_blocks += blocks;
214 + sbi->s_delete_inodes++;
215 + spin_unlock(&sbi->s_delete_lock);
217 + ext3_debug("delete inode %lu (%lu blocks) by thread\n",
218 + new_inode->i_ino, blocks);
220 + wake_up(&sbi->s_delete_thread_queue);
224 + ext3_delete_inode(old_inode);
227 +#define ext3_start_delete_thread(sbi) do {} while(0)
228 +#define ext3_stop_delete_thread(sbi) do {} while(0)
229 +#endif /* EXT3_DELETE_THREAD */
231 void ext3_put_super (struct super_block * sb)
233 struct ext3_sb_info *sbi = EXT3_SB(sb);
235 kdev_t j_dev = sbi->s_journal->j_dev;
238 + J_ASSERT(sbi->s_delete_inodes == 0);
239 ext3_xattr_put_super(sb);
240 journal_destroy(sbi->s_journal);
241 if (!(sb->s_flags & MS_RDONLY)) {
243 write_inode: ext3_write_inode, /* BKL not held. Don't need */
244 dirty_inode: ext3_dirty_inode, /* BKL not held. We take it */
245 put_inode: ext3_put_inode, /* BKL not held. Don't need */
246 - delete_inode: ext3_delete_inode, /* BKL not held. We take it */
247 +#ifdef EXT3_DELETE_THREAD
248 + delete_inode: ext3_delete_inode_thread,/* BKL not held. We take it */
250 + delete_inode: ext3_delete_inode, /* BKL not held. We take it */
252 put_super: ext3_put_super, /* BKL held */
253 write_super: ext3_write_super, /* BKL held */
254 sync_fs: ext3_sync_fs,
256 clear_opt (*mount_options, XATTR_USER);
259 +#ifdef EXT3_DELETE_THREAD
260 + if (!strcmp(this_char, "asyncdel"))
261 + set_opt(*mount_options, ASYNCDEL);
262 + else if (!strcmp(this_char, "noasyncdel"))
263 + clear_opt(*mount_options, ASYNCDEL);
266 if (!strcmp (this_char, "bsddf"))
267 clear_opt (*mount_options, MINIX_DF);
268 else if (!strcmp (this_char, "nouid32")) {
269 @@ -1223,6 +1450,7 @@
272 ext3_setup_super (sb, es, sb->s_flags & MS_RDONLY);
273 + ext3_start_delete_thread(sb);
275 * akpm: core read_super() calls in here with the superblock locked.
276 * That deadlocks, because orphan cleanup needs to lock the superblock
277 @@ -1614,7 +1842,12 @@
278 static int ext3_sync_fs(struct super_block *sb)
283 + if (atomic_read(&sb->s_active) == 0) {
284 + /* fs is being umounted: time to stop delete thread */
285 + ext3_stop_delete_thread(EXT3_SB(sb));
289 target = log_start_commit(EXT3_SB(sb)->s_journal, NULL);
290 log_wait_commit(EXT3_SB(sb)->s_journal, target);
291 @@ -1678,6 +1911,9 @@
292 if (!parse_options(data, &tmp, sbi, &tmp, 1))
295 + if (!test_opt(sb, ASYNCDEL) || (*flags & MS_RDONLY))
296 + ext3_stop_delete_thread(sbi);
298 if (sbi->s_mount_opt & EXT3_MOUNT_ABORT)
299 ext3_abort(sb, __FUNCTION__, "Abort forced by user");
301 Index: linux/fs/ext3/inode.c
302 ===================================================================
303 --- linux.orig/fs/ext3/inode.c Mon Feb 2 20:57:35 2004
304 +++ linux/fs/ext3/inode.c Mon Feb 2 20:58:05 2004
305 @@ -2500,6 +2500,118 @@
309 +#ifdef EXT3_DELETE_THREAD
310 +/* Move blocks from to-be-truncated inode over to a new inode, and delete
311 + * that one from the delete thread instead. This avoids a lot of latency
312 + * when truncating large files.
314 + * If we have any problem deferring the truncate, just truncate it right away.
315 + * If we defer it, we also mark how many blocks it would free, so that we
316 + * can keep the statfs data correct, and we know if we should sleep on the
317 + * delete thread when we run out of space.
319 +void ext3_truncate_thread(struct inode *old_inode)
321 + struct ext3_sb_info *sbi = EXT3_SB(old_inode->i_sb);
322 + struct ext3_inode_info *nei, *oei = EXT3_I(old_inode);
323 + struct inode *new_inode;
325 + unsigned long blocks = old_inode->i_blocks >> (old_inode->i_blkbits-9);
327 + if (!test_opt(old_inode->i_sb, ASYNCDEL) || !sbi->s_delete_list.next)
330 + /* XXX This is a temporary limitation for code simplicity.
331 + * We could truncate to arbitrary sizes at some later time.
333 + if (old_inode->i_size != 0)
336 + /* We may want to truncate the inode immediately and not defer it */
337 + if (IS_SYNC(old_inode) || blocks <= EXT3_NDIR_BLOCKS ||
338 + old_inode->i_size > oei->i_disksize)
341 + /* We can't use the delete thread as-is during real orphan recovery,
342 + * as we add to the orphan list here, causing ext3_orphan_cleanup()
343 + * to loop endlessly. It would be nice to do so, but needs work.
345 + if (oei->i_state & EXT3_STATE_DELETE ||
346 + sbi->s_mount_state & EXT3_ORPHAN_FS) {
347 + ext3_debug("doing deferred inode %lu delete (%lu blocks)\n",
348 + old_inode->i_ino, blocks);
352 + ext3_discard_prealloc(old_inode);
355 + * new_inode = sb + GDT + ibitmap
356 + * orphan list = 1 inode/superblock for add, 2 inodes for del
357 + * quota files = 2 * EXT3_SINGLEDATA_TRANS_BLOCKS
359 + handle = ext3_journal_start(old_inode, 7);
360 + if (IS_ERR(handle))
363 + new_inode = ext3_new_inode(handle, old_inode, old_inode->i_mode);
364 + if (IS_ERR(new_inode)) {
365 + ext3_debug("truncate inode %lu directly (no new inodes)\n",
370 + nei = EXT3_I(new_inode);
372 + down_write(&oei->truncate_sem);
373 + new_inode->i_size = old_inode->i_size;
374 + new_inode->i_blocks = old_inode->i_blocks;
375 + new_inode->i_uid = old_inode->i_uid;
376 + new_inode->i_gid = old_inode->i_gid;
377 + new_inode->i_nlink = 0;
379 + /* FIXME when we do arbitrary truncates */
380 + old_inode->i_blocks = oei->i_file_acl ? old_inode->i_blksize / 512 : 0;
381 + old_inode->i_mtime = old_inode->i_ctime = CURRENT_TIME;
383 + memcpy(nei->i_data, oei->i_data, sizeof(nei->i_data));
384 + memset(oei->i_data, 0, sizeof(oei->i_data));
386 + nei->i_disksize = oei->i_disksize;
387 + nei->i_state |= EXT3_STATE_DELETE;
388 + up_write(&oei->truncate_sem);
390 + if (ext3_orphan_add(handle, new_inode) < 0)
393 + if (ext3_orphan_del(handle, old_inode) < 0) {
394 + ext3_orphan_del(handle, new_inode);
399 + ext3_journal_stop(handle, old_inode);
401 + spin_lock(&sbi->s_delete_lock);
402 + J_ASSERT(list_empty(&new_inode->i_dentry));
403 + list_add_tail(&new_inode->i_dentry, &sbi->s_delete_list);
404 + sbi->s_delete_blocks += blocks;
405 + sbi->s_delete_inodes++;
406 + spin_unlock(&sbi->s_delete_lock);
408 + ext3_debug("delete inode %lu (%lu blocks) by thread\n",
409 + new_inode->i_ino, blocks);
411 + wake_up(&sbi->s_delete_thread_queue);
415 + ext3_journal_stop(handle, old_inode);
417 + ext3_truncate(old_inode);
419 +#endif /* EXT3_DELETE_THREAD */
422 * On success, We end up with an outstanding reference count against
423 * iloc->bh. This _must_ be cleaned up later.
424 Index: linux/fs/ext3/file.c
425 ===================================================================
426 --- linux.orig/fs/ext3/file.c Mon Feb 2 20:57:34 2004
427 +++ linux/fs/ext3/file.c Mon Feb 2 20:58:05 2004
431 struct inode_operations ext3_file_inode_operations = {
432 +#ifdef EXT3_DELETE_THREAD
433 + truncate: ext3_truncate_thread, /* BKL held */
435 truncate: ext3_truncate, /* BKL held */
437 setattr: ext3_setattr, /* BKL held */
438 setxattr: ext3_setxattr, /* BKL held */
439 getxattr: ext3_getxattr, /* BKL held */
440 Index: linux/include/linux/ext3_fs.h
441 ===================================================================
442 --- linux.orig/include/linux/ext3_fs.h Mon Feb 2 20:57:35 2004
443 +++ linux/include/linux/ext3_fs.h Mon Feb 2 20:58:05 2004
446 #define EXT3_STATE_JDATA 0x00000001 /* journaled data exists */
447 #define EXT3_STATE_NEW 0x00000002 /* inode is newly created */
448 +#define EXT3_STATE_DELETE 0x00000010 /* deferred delete inode */
453 #define EXT3_MOUNT_UPDATE_JOURNAL 0x1000 /* Update the journal format */
454 #define EXT3_MOUNT_NO_UID32 0x2000 /* Disable 32-bit UIDs */
455 #define EXT3_MOUNT_XATTR_USER 0x4000 /* Extended user attributes */
456 +#define EXT3_MOUNT_ASYNCDEL 0x20000 /* Delayed deletion */
458 /* Compatibility, for having both ext2_fs.h and ext3_fs.h included at once */
459 #ifndef _LINUX_EXT2_FS_H
461 extern void ext3_dirty_inode(struct inode *);
462 extern int ext3_change_inode_journal_flag(struct inode *, int);
463 extern void ext3_truncate (struct inode *);
464 +#ifdef EXT3_DELETE_THREAD
465 +extern void ext3_truncate_thread(struct inode *inode);
469 extern int ext3_ioctl (struct inode *, struct file *, unsigned int,
470 Index: linux/include/linux/ext3_fs_sb.h
471 ===================================================================
472 --- linux.orig/include/linux/ext3_fs_sb.h Mon Feb 2 20:57:35 2004
473 +++ linux/include/linux/ext3_fs_sb.h Mon Feb 2 20:58:05 2004
476 #define EXT3_MAX_GROUP_LOADED 8
478 +#define EXT3_DELETE_THREAD
481 * third extended-fs super-block data in memory
484 struct timer_list turn_ro_timer; /* For turning read-only (crash simulation) */
485 wait_queue_head_t ro_wait_queue; /* For people waiting for the fs to go read-only */
487 +#ifdef EXT3_DELETE_THREAD
488 + spinlock_t s_delete_lock;
489 + struct list_head s_delete_list;
490 + unsigned long s_delete_blocks;
491 + unsigned long s_delete_inodes;
492 + wait_queue_head_t s_delete_thread_queue;
493 + wait_queue_head_t s_delete_waiter_queue;
497 #endif /* _LINUX_EXT3_FS_SB */