[Top] [All Lists]

Re: [PATCH] xfs: improve xfs_bitmap_empty()

To: Jeff Liu <jeff.liu@xxxxxxxxxx>
Subject: Re: [PATCH] xfs: improve xfs_bitmap_empty()
From: Dave Chinner <david@xxxxxxxxxxxxx>
Date: Mon, 3 Feb 2014 08:52:31 +1100
Cc: Mark Tinguely <tinguely@xxxxxxx>, Eric Sandeen <sandeen@xxxxxxxxxxx>, "xfs@xxxxxxxxxxx" <xfs@xxxxxxxxxxx>
Delivered-to: xfs@xxxxxxxxxxx
In-reply-to: <52EC6EA0.9000501@xxxxxxxxxx>
References: <52EBAF91.10608@xxxxxxxxxx> <52EBBC3D.5030507@xxxxxxxxxxx> <52EBC103.5050006@xxxxxxxxxx> <52EBC1AF.6020000@xxxxxxxxxxx> <52EBC67B.7020806@xxxxxxxxxx> <52EBCF12.1020403@xxxxxxx> <52EC6EA0.9000501@xxxxxxxxxx>
User-agent: Mutt/1.5.21 (2010-09-15)
On Sat, Feb 01, 2014 at 11:48:48AM +0800, Jeff Liu wrote:
> On 02/01 2014 00:28 AM, Mark Tinguely wrote:
> > On 01/31/14 09:51, Jeff Liu wrote:
> >> On 01/31 2014 23:30 PM, Eric Sandeen wrote:
> >>> On 1/31/14, 9:28 AM, Jeff Liu wrote:
> >>>> Well, when I looking through our bitmap source, I once thought if
> >>>> we can replace the current code with the generic bitmap library.
> >>>> However, our map is uint rather than unsigned long...
> >>>
> >>> Technically the unsigned long (pointer) is just the bitmap address,
> >>> I think.
> >>
> >> Yeah, so this might worth to try on long terms.
> > 
> > The blf_data_map[] is int aligned, not long aligned.
> > You could reflect the alignment difference in the offset or
> > change the alignment in the structure.
> For now, I think we can not simply turn to generic bitmap just because
> of the alignment difference on 64-bits OS.

The bitmaps end up on disk (in the log), so replacing the
implementation with a generic implementation is something we need to
be very careful about.

IMO, we should be getting rid of the bitmaps from the
xfs_buf_log_item first (by moving to a low byte/high byte offset
range), then we only have to worry about bitmaps when doing log
recovery after a kernel upgrade on a filesystem with a dirty log.

Getting rid of the bitmaps also solves a scalability problem with
large block sizes tracking all the changes in buffer - we burn a
huge amount of CPU walking bits when logging 64k directory buffers:

+  21.19%  [kernel]  [k] xfs_dir3_leaf_check_int
+  12.20%  [kernel]  [k] memcpy
+   9.29%  [kernel]  [k] xfs_next_bit
+   5.04%  [kernel]  [k] xfs_buf_offset
+   3.63%  [kernel]  [k] xfs_buf_item_format
+   3.59%  [kernel]  [k] xfs_buf_item_size_segment

The logging of xfs_buf_log_items there is consuming >30% of the CPU
being used under this workload (xfs_dir3_leaf_check_int() is high
because this is from a debug kernel.)

IOWs, we should work to remove the bitmap code from general
operations first, then replace the remaining legacy log recovery
code with the generic bitmap implemention....


Dave Chinner

<Prev in Thread] Current Thread [Next in Thread>