Messages in this thread |  | | Date | Sat, 22 Dec 2007 02:06:11 -0800 | From | Andrew Morton <> | Subject | Re: Possible fix for lockup in drop_caches |
| |
On Mon, 17 Dec 2007 12:13:22 +0000 richard <richard@rsk.demon.co.uk> wrote:
> fix lockup in when calling drop_caches > > calling /proc/sys/vm/drop_caches can hang due to a AB/BA lock dependency > between j_list_lock and the inode_lock. This patch moves the redirtying of the buffer head out > from under the j_list_lock. > > based on a suggestion by Andrew Morton. >
Oh boy. Do we really want to add all this stuff to JBD just for drop_caches which is a silly root-only broken-in-22-other-ways thing?
Michael, might your convert-inode-lists-to-tree patches eliminate the need for taking inode_lock in drop_pagecache_sb()? Probably not, as it uses an rbtree. It would have been possible if it was using a radix-tree, I suspect..
> -void __journal_unfile_buffer(struct journal_head *jh) > +void __journal_unfile_buffer(struct journal_head *jh, > + struct buffer_head **dirty_bh) > { > - __journal_temp_unlink_buffer(jh); > + __journal_temp_unlink_buffer(jh, dirty_bh); > jh->b_transaction = NULL; > }
I suspect the code would end up simpler if __journal_unfile_buffer() were to take an additional ref on the bh which it placed at *dirty_bh.
Callers of __journal_unfile_buffer() could then call
void handle_dirty_bh(struct buffer_head *bh) { if (bh) { jbd_mark_buffer_dirty(bh); put_bh(bh); } }
?
|  |