Re: [PATCH] another reader/writer lock for btrfs metadata

[Chris Mason <chris.mason@oracle.com>]

Excerpts from Arne Jansen's message of 2011-07-27 04:13:44 -0400:
> On 27.07.2011 00:00, Chris Mason wrote:
> > Excerpts from Chris Mason's message of 2011-07-25 21:28:30 -0400:
> >> Excerpts from Chris Mason's message of 2011-07-25 14:34:49 -0400:
> >>> Hi everyone,
> >>>
> >>> I've updated the integration-test branch to use this code instead.  It
> >>> is a shiny new reader/writer lock built around rw spinlocks.  I've
> >>> removed all the adaptive spinning and started trusting the hints btrfs
> >>> already has about when blocks should block or spin.
> >>
> >> I tested with lockdep and looks like I've got a bug in btrfs_next_leaf's
> >> lockdep handling.  So, please don't run this code with
> >> CONFIG_DEBUG_LOCK_ALLOC turned on.  The bug is only in lockdep mode, so
> >> we're fine with it off.
> >>
> >> I'll fix it up in the morning.
> >
> > Ok, I've rebased the btrfs integration-test branch to include fixes for
> > the lockdep problems.  I've also adapted Tejun's lockdep class patch to
> > the new code.  With this setup I'm not getting lockdep warnings, but I'm
> > always looking for more bug reports.
> >
> 
> My fs_mark test look strange. As I can't make sense of it and don't have
> much time today to investigate, I just paste them here:
> 
> fs_mark with b249e55006c87f (pre-r/w-lock):
> FSUse%        Count         Size    Files/sec     App Overhead
>      16        65536        51200       9603.3          2064173
>      27       131072        51200       7995.1          1710337
>      37       196608        51200       8662.7          2046263
>      47       262144        51200       8024.2          1876913
>      58       327680        51200       7564.5          1593306
> 
> # btrfs fi df /mnt/fsm
> Data: total=4.21GB, used=3.70GB
> System: total=4.00MB, used=4.00KB
> Metadata: total=520.00MB, used=65.24MB
> 
> fs_mark with integration-test:
> FSUse%        Count         Size    Files/sec     App Overhead
>      18        65536        51200       7617.4           882409
>      34       131072        51200       6447.1           935868
>      51       196608        51200       5697.7           938026
>      63       262144        51200       7462.9           933696
>      76       327680        51200       6807.7           921403
> 
> # btrfs fi df /mnt/fsm
> Data: total=5.61GB, used=4.98GB
> System: total=4.00MB, used=4.00KB
> Metadata: total=520.00MB, used=77.04MB
> 
> Please note the different fs usage after a single fs_mark run with
> the same parameters (fs_mark  -d  /mnt/fsm  -D  512  -t  16  -n  4096
> -s  51200  -L5  -S0).
> Maybe fs_mark just have some problems with accounting when running
> in multiple threads.

Hmmm, I've always had fs_mark give me the same output (fsuse % numbers)
for each run.  I'd say you had a file in the FS at the start of the run.

It looks like you're not CPU bound.  We've got a few different tweaks in
the integration-test branch right now, one of them must be flushing more
often.  You're also filling up the FS pretty quickly, so these results
are probably bound by writeback.

-chris