On 08/14/2011 01:33 PM, Tom Lane wrote:
> (The changes in question were purely for performance, and involved a
> conversion from write barriers in the block layer to flush+fua, whatever
> that is.) Furthermore, this affected every filesystem not only ext4,
> so it really entirely fails to support Greg's argument.
>
FUA is "Force Unit Access", shorthand for forcing write-through. If
you're staring at this from the perspective I am, where I assume that
every line of new filesystem code always takes a while from release
until it's stable, it still concerns me. I suspect that both of these
code paths--the write barrier one and the flush+FUA one--are fine on
XFS. One of the reasons XFS was so terrible in its early years was
because it took a long time to get those parts in particular correct,
which was complicated by how badly drives lied about everything back then.
If these changes shift around which section of the ext4 write path code
are exercised, it could easily be the case that it moves fsync related
work (which has been one of the buggiest parts of that filesystem) from
code that has been tested heavily to parts that haven't been hit as hard
yet by users. My paranoid raving about this particular bug isn't
strongly supported; I'll concede that with the additional data you've
provided. But I don't think it's completely unfounded either.
--
Greg Smith 2ndQuadrant US greg@2ndQuadrant.com Baltimore, MD
PostgreSQL Training, Services, and 24x7 Support www.2ndQuadrant.us