Hi, On Wed, 2012-03-28 at 13:54 +0200, Jan Kara wrote: > Hi, > > On Wed 28-03-12 10:04:15, Steven Whitehouse wrote: > > On Wed, 2012-03-28 at 03:38 +0100, Al Viro wrote: > > > On Tue, Mar 27, 2012 at 11:08:58PM +0200, Jan Kara wrote: > > > > Hello, > > > > > > > > maybe the name of this topic could be "How hard should be life of > > > > filesystems?" but that's kind of broad topic and suggests too much of > > > > bikeshedding. I'd like to concentrate on concrete possible pain points > > > > between filesystems & VFS (possibly writeback or even generally MM). > > > > Lately, I've myself came across the two issues in $SUBJECT: > > > > 1) dropping of last file reference can happen from munmap() and in that > > > > case mmap_sem will be held when ->release() is called. Even more it > > > > could be held when ->evict_inode() is called to delete inode because > > > > inode was unlinked. > > > > > > Yes, it can. > > > > > > > 2) since flusher thread takes inode reference when writing inode out, the > > > > last inode reference can be dropped from flusher thread. Thus inode may > > > > get deleted in the flusher thread context. This does not seem that > > > > problematic on its own but if we realize progress of memory reclaim > > > > depends (at least from a longterm perspective) on flusher thread making > > > > progress, things start looking a bit uncertain. Even more so when we > > > > would like avoid ->writepage() calls from reclaim and let flusher thread > > > > do the work instead. That would then require filesystems to carefully > > > > design their ->evict_inode() routines so that things are not > > > > deadlockable. > > > > > > You mean "use GFP_NOIO for allocations when holding fs-internal locks"? > > > > > > > Both these issues should be avoidable (we can postpone fput() after we > > > > drop mmap_sem; we can tweak inode refcounting to avoid last iput() from > > > > flusher thread) but obviously there's some cost in the complexity of generic > > > > layer. So the question is, is it worth it? > > > > > > I don't thing it is. ->i_mutex in ->release() is never needed; existing > > > cases are racy and dropping preallocation that way is simply wrong. And > > > ->evict_inode() is a non-issue, since it has no reason whatsoever to take > > > *any* locks in mutex - the damn thing is called when nobody has references > > > to struct inode anymore. Deadlocks with flusher... that's what NOIO and > > > NOFS are for. > > > > > For cluster filesystems, we have to take locks (cluster wide) in > > ->evict_inode() in order to establish for certain whether we are the > > last opener of the inode. Just because there are no references on the > > local node, doesn't mean that a remote node doesn't hold the file open > > still. > > > > We do always use GFP_NOFS when allocating memory while holding such > > locks, so I'm not quite sure from the above whether or not that will be > > an issue, > Yeah, but you have to use networking to communicate with other nodes > about locks and this creates another interesting dependecy. > > Currently, everything seems to work out just fine and I don't say I know > about a particular deadlock. I just say that the dependencies are so > complex that I don't know whether things will work OK e.g. if we change > page reclaim to offload more to flusher thread. And that's what I feel > uneasy about. > > Honza Yes, I agree. I've certainly seen some issues with this code path in GFS2 in the past though, so making it more robust in this way seems to be a good plan to me, Steve. -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxx. For more info on Linux MM, see: http://www.linux-mm.org/ . Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/ Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>