On Tue 24-03-09 15:56:03, Peter Zijlstra wrote: > On Tue, 2009-03-24 at 15:47 +0100, Jan Kara wrote: > > > > Or we could implement ext3_mkwrite() to allocate buffers already when we > > make page writeable. But it costs some performace (we have to write page > > full of zeros when allocating those buffers, where previously we didn't > > have to do anything) and it's not trivial to make it work if pagesize > > > blocksize (we should not allocate buffers outside of i_size so if i_size > > = 1024, we create just one block in ext3_mkwrite() but then we need to > > allocate more when we extend the file). > > I think this is the best option, failing with SIGBUS when we fail to > allocate blocks seems consistent with other filesystems as well. I agree this looks attractive at the first sight. But there are drawbacks as I wrote - the problem with blocksize < pagesize, slight performance decrease due to additional write, page faults doing allocation can take a *long* time and overall fragmentation is going to be higher (previously writepage wrote pages for us in the right order, now we are going to allocate in the first-accessed order). So I'm not sure we really want to go this way. Hmm, maybe we could play a trick ala delayed allocation - i.e., reserve some space in mkwrite() but don't actually allocate it. That would be done in writepage(). This would solve all the problems I describe above. We could use PG_Checked flag to track that the page has a reservation and behave accordingly in writepage() / invalidatepage(). ext3 in data=journal mode already uses the flag but the use seems to be compatible with what I want to do now... So it may actually work. BTW: Note that there's a plenty of filesystems that don't implement mkwrite() (e.g. ext2, UDF, VFAT...) and thus have the same problem with ENOSPC. So I'd not speak too much about consistency ;). Honza -- Jan Kara <jack@xxxxxxx> SUSE Labs, CR -- To unsubscribe from this list: send the line "unsubscribe linux-ext4" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
