RFC: Page size on PPC/PPC64 builders

[David Woodhouse]

The PPC64 build hosts (which also build ppc32 packages) recently
switched to RHEL5, and hence have 64KiB page size.

This is permitted by the PPC ABI, just as it is by the x86_64 ABI, and
works perfectly well for almost everything.

Occasionally, however, it shows up some bugs in packages. This week, it
found bug #435337, which is a generic bug and not really ppc-specific --
although I won't enter the debate on whether it's a glibc bug or a bug
with the programs which 'naively' expect pthread_attr_setstacksize() to
actually set the stack size to the number they provide.

However, despite the fact that the majority of build failures on PPC
actually demonstrate _generic_ bugs and not arch-specific bugs at all,
some packagers don't bother to look -- they prefer to just exclude the
'offending' architecture and move on without checking to see what the
problem is. And mutter to themselves that PPC is 'painful', despite the
fact that it's helping to improve the quality of Fedora overall by
finding their generic bugs. Primarily for that reason, I'm tempted to
switch the PPC builders back to 4KiB pages. 

I'm not sure it's the right thing to do though -- there are arguments on
either side.

For 4KiB pages:
 - Makes life 'easier' for PPC builds, to stop people whining about it
   finding their bugs for them.
 - Consistent with what we actually ship in Fedora.

For 64KiB pages:
 - Finds real bugs
 - Consistent with what will be shipped for RHEL.

Note that although some of us might not care so much about RHEL, the
point about consistency in that case is stronger -- because programs
_built_ with 64KiB pages are almost always going to work when run with
4KiB pages, but the other way round is not so frequently true.

I'm concerned that switching back to 4KiB pages is just papering over
real bugs to make life easier for PPC folks. I suspect that what I
should _actually_ do is keep it 64KiB, brazen it out and LART people who
just exclude ppc builds without actually looking at the problem for
themselves. But I'm lazy too... maybe we should switch the x86_64
builders to 64KiB instead? :)

Opinions?

-- 
dwmw2