[Bug libc/11787] Program with large TLS segment fails aio_write

["law at redhat dot com" <sourceware-bugzilla at sourceware dot org>]

http://sourceware.org/bugzilla/show_bug.cgi?id=11787

law at redhat dot com changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |law at redhat dot com

--- Comment #24 from law at redhat dot com 2012-03-30 16:19:32 UTC ---
I think the real question here isn't the change in the size of the stack but
the increase in the # of pages and potential wastage.

If the TLS is small won't we end up burning an entire page for this small
amount of data?  For apps that use small stacks and lots of threads, that extra
page could turn out to be significant.

I'm not sure I really buy the argument that if a thread asks for 16k that they
get exactly 16k usable.

However, obviously, if the TLS is taking up a significant portion of the
thread's requested stack, then something needs to be done.

There's no perfect solution here.  I wonder if we could come up with a good
heuristic based on the relative sizes of the TLS and thread requested stack. 
If the TLS is sufficiently small relative to the size of the requested stack,
then just fold the TLS into the requested stack like we've always done. 
Otherwise, add the size of the TLS to the size of the requested stack (rounding
to a page of course).  Harder to document and implement, but seems like it'd
strike a better balance.

I don't know where the cutoff should be, 1%, 10%, 50%? Some experimentation may
guide us.

-- 
Configure bugmail: http://sourceware.org/bugzilla/userprefs.cgi?tab=email
------- You are receiving this mail because: -------
You are on the CC list for the bug.