Re: fallocate / posix_fallocate for new WAL file creation (etc...)

On 2013-05-30 07:48:51 -0400, Greg Smith wrote:
> On 5/30/13 7:17 AM, Andres Freund wrote:
> >That argument in contrast I find not very convincing though. What was
> >the last incidence of such a system call that did not just error out
> >with ENOTSUPP or such?
> 
> http://linux.die.net/man/2/posix_fadvise talks about  POSIX_FADV_NOREUSE and
> POSIX_FADV_WILLNEED being both buggy and quietly mapped to a no-op,
> depending on your version.  I know there were more examples than just that
> one that popped up during the testing of effective_io_concurrency.  My
> starting position has to assume that posix_fallocate can have the same sort
> of surprising behavior that showed up repeatedly when we were trying to use
> posix_fadvise more aggressively.

Uh. How is that a correctness problem? fadvise is a hint which is pretty
different from a fallocate where ignoring would have way much more
severe consequences.
I don't think that's a very meaningful comparison.

> The way O_SYNC was quietly mapped to O_DSYNC (which isn't the same thing)
> was a similar issue, and that's the first one that left me forever skeptical
> of Linux kernel claims in this area until they are explicitly validated:
> http://lwn.net/Articles/350225/

Yea, but that mistake is literally decades old...

Greetings,

Andres Freund

-- Andres Freund                       http://www.2ndQuadrant.com/PostgreSQL Development, 24x7 Support, Training &
Services