Re: ftruncate-mmap: pages are lost after writing to mmaped file.

[Linus Torvalds <torvalds@linux-foundation.org>]

On Thu, 19 Mar 2009, Jan Kara wrote:
> > 
> > Ahh, so you re-created it? On ext2 only, or is it visible on ext3 as well? 
> > I've not even tested - I assumed that I would have to boot into less 
> > memory and downgrade my filesystem to ext2, which made me hope somebody 
> > else would pick it up first ;)
>
>   In thread http://lkml.org/lkml/2009/3/4/179 I've reported similar problem
> - write lost. I'm able to reproduce under UML linux at will. ext3 takes
> with 1KB blocksize about 20 minutes to hit the corruption, ext2 with 1 KB
> blocksize about an hour, ext2 with 4KB blocksize several hours...

Hmm. I can't seem to recreate it with Ying Han's testprog, at least. 
That's with the fs/buffer.c patch applied, but that shouldn't matter since 
Nick reports that his (roughly equivalent) patch didn't help.

I'll continue to run it for a while, but it's been going for about an hour 
now, with vmstat reporting bo/bi at roughly 5-10MB/s pretty continuosly.

Of course, that's with my SSD's and an insanely fast Nehalem box, so my 
timings are likely rather different from most other peoples.

			Linus