Re: [uml-user] slow ubd performance

[richard -r. w. <ric...@gm...>]

On Thu, Dec 8, 2011 at 12:38 PM, Riccardo Murri
<ric...@gm...> wrote:
> Hi *,
>
> the file-backed `ubd` devices in my UML instances are terribly slow,
> and generate a lot of "iowait" load on the host.  It looks like as if
> filesystems mounted from `ubd` devices were using the `sync` option.
>
> This happens with kernels downloaded from uml.devloop.org.uk, as well
> with kernels that I compiled myself (but using basically the same
> config as the devloop ones); I tried 2.6.38.x and 2.6.39.x (for some
> random values of "x").
>
> This does *not* happen with Debian/Ubuntu's kernel 2.6.32.5, which
> unfortunately exhibit a filesystem corruption bug.
>
> The host is in any case Ubuntu "natty", running 2.6.39-3-generic.
>
> Is this expected?

Not really ;)

The ubd got only a few patches since 2.6.32, bisecting should be an easy task.
Commit 4752690 fixes the data corruption issue.

[rw@inhelltoy tip]$ git log --oneline v2.6.32.. arch/um/drivers/ubd_*
8535639 um: fix ubd cow size
c506c0e um: take ubd_user.h to its users...
0a9e70b um: kill shared/mem_kern.h
dbddc51 um: don't include kern.h unless it's needed
078073a um: -include user.h for USER_OBJ, trim includes
22e6500 um: Replace deprecated spinlock initialization
5704e44 Merge branch 'config' of
git://git.kernel.org/pub/scm/linux/kernel/git/arnd/bkl
9a181c5 uml: kill big kernel lock
4752690 ubd: fix incorrect sector handling during request restart
6e9624b block: push down BKL into .open and .release
5a0e3ad include cleanup: Update gfp.h and slab.h includes to prepare
for breaking implicit slab.h inclusion from percpu.h
8a78362 block: Consolidate phys_segment and hw_segment limits
086fa5f block: Rename blk_queue_max_sectors to blk_queue_max_hw_sectors
6613c5e uml: convert to seq_file/proc_fops

-- 
Thanks,
//richard