PIPE_BUFFERS - was Re: What to expect after 0.99.8

PIPE_BUFFERS - was Re: What to expect after 0.99.8

[Luck, Tony]

On Tue, Oct 04, 2005 at 05:10:18PM -0700, Linus Torvalds wrote:
> Just for performance reasons, I ended up doing merging anyway, so in fact 
> regardless of how you write the current pipe buffer size is up to 16 
> pages.

s/git@vger.kernel.org/linux-kernel@vger.kernel.org/ in the Cc: list, this
isn't a GIT issue.

Should that:

#define PIPE_BUFFERS (16)

really be a function of PAGE_SIZE?  Many systems have a 4K page size
so they get 64K of buffering per pipe.  But ia64 defaults to 16K page
size, so pipes are suddenly 256K ... and with a 64K page size we
have 1MB per pipe.

So perhaps:

#define PIPE_BUFFERS (65536 / PAGE_SIZE)

But that would leave PIPE_BUFFERS as "1" on a 64K page size system, is
there any double-buffering benefit from having more than one page on the
pipe list?  If so, then something like:

#define PIPE_BUFFERS min(2, (65536 / PAGE_SIZE))

but then gcc is unhappy using the super-type safe "min" in the
places that use PIPE_BUFFERS ... so we'd end up with:

#define PIPE_BUFFERS ((65536 / PAGE_SIZE) > 1 ? (65536 / PAGE_SIZE) : 2)

-Tony