Re: [PATCH] ext3 barrier bits

Re: [PATCH] ext3 barrier bits

[Andrew Morton]

May as well cc lkml on this.  It's to do with the disk write barrier
implementation.


- How do I know that the barrier code is actually doing stuff?  It doesn't
  seem to affect benchmarks much, if at all.

- Does reiserfs support `mount -o remount,barrier=flush'? and "=none"?

- How do I test the "oh, barriers aren't working" fallback code in ext3?

- Does the kernel tell you if your disk doesn't supoprt barriers?  ie:
  how does the user know if it's working or not?

Re: [PATCH] ext3 barrier bits

[Arjan van de Ven]

[-- Attachment #1: Type: text/plain, Size: 237 bytes --]


> - Does the kernel tell you if your disk doesn't supoprt barriers?  ie:
>   how does the user know if it's working or not?

... and how do you know your disk isn't lying and ignoring the barriers?

"Storage is a Lie" -- Andre


[-- Attachment #2: This is a digitally signed message part --]
[-- Type: application/pgp-signature, Size: 189 bytes --]

Re: [PATCH] ext3 barrier bits

[Andrew Morton]

Arjan van de Ven <arjanv@redhat.com> wrote:
>
> 
> > - Does the kernel tell you if your disk doesn't supoprt barriers?  ie:
> >   how does the user know if it's working or not?
> 
> ... and how do you know your disk isn't lying and ignoring the barriers?
> 

If some benchmark slows down when barriers are enabled then that's a pretty
good indicator.  I don't know what benchmark that might be though.

Re: [PATCH] ext3 barrier bits

[Jens Axboe]

On Sat, May 22 2004, Andrew Morton wrote:
> 
> May as well cc lkml on this.  It's to do with the disk write barrier
> implementation.
> 
> 
> - How do I know that the barrier code is actually doing stuff?  It doesn't
>   seem to affect benchmarks much, if at all.

Usually not a lot of barriers would be generated. If you use reiser and
heavy fsync load, it should be pretty apparent of barriers are issued or
not (ie compare without barrier + wb cache, with barrier + wb cache, and
with/without - wb cache).

> - Does reiserfs support `mount -o remount,barrier=flush'? and "=none"?

Chris?

> - How do I test the "oh, barriers aren't working" fallback code in ext3?

A quick hack would be to replace 'err' with '1' in this line in
drivers/ide/ide-io.c:void ide_end_drive_cmd():

        if (blk_barrier_preflush(rq) || blk_barrier_postflush(rq))
                ide_complete_barrier(drive, rq, err);

that'll fail the first barrier, should trigger the fall back. Or just
try -o barrier=flush on a SCSI drive, should fail the barrier too.

> - Does the kernel tell you if your disk doesn't supoprt barriers?  ie:
>   how does the user know if it's working or not?

IDE will not tell you, the file system should flag the disabling of
barriers (like reiser currently does). Try SCSI case again.

-- 
Jens Axboe

Re: [PATCH] ext3 barrier bits

[Jens Axboe]

On Sat, May 22 2004, Arjan van de Ven wrote:
> 
> > - Does the kernel tell you if your disk doesn't supoprt barriers?  ie:
> >   how does the user know if it's working or not?
> 
> ... and how do you know your disk isn't lying and ignoring the barriers?

Easy to find out, time it. Or invent more imaginative cases where you
actually test if it's there.

-- 
Jens Axboe

Re: [PATCH] ext3 barrier bits

[Chris Mason]

On Sat, 2004-05-22 at 04:33, Andrew Morton wrote:
> Arjan van de Ven <arjanv@redhat.com> wrote:
> >
> > 
> > > - Does the kernel tell you if your disk doesn't supoprt barriers?  ie:
> > >   how does the user know if it's working or not?
> > 
> > ... and how do you know your disk isn't lying and ignoring the barriers?
> > 
> 
> If some benchmark slows down when barriers are enabled then that's a pretty
> good indicator.  I don't know what benchmark that might be though.

synctest -t 20 -n 1 -f dir, should show a performance hit when barriers
are on.

-chris

Re: [PATCH] ext3 barrier bits

[Chris Mason]

On Sat, 2004-05-22 at 04:11, Andrew Morton wrote:
> May as well cc lkml on this.  It's to do with the disk write barrier
> implementation.

> - Does reiserfs support `mount -o remount,barrier=flush'? and "=none"?
> 
Yes...there are no messages to clue the user in, but it does happen.

> - How do I test the "oh, barriers aren't working" fallback code in ext3?
> 
Jens suggested mounting with barriers on for scsi, that's what I've been
doing.  I also did some crash tests on scsi to make sure the io really
is happening when barriers fail.

-chris