Large stack usage in fs code (especially for PPC64)

Pekka Enberg penberg at cs.helsinki.fi
Tue Nov 18 08:25:30 EST 2008


On Mon, Nov 17, 2008 at 11:18 PM, Linus Torvalds
<torvalds at linux-foundation.org> wrote:
> I do wonder just _what_ it is that causes the stack frames to be so
> horrid. For example, you have
>
>         18)     8896     160   .kmem_cache_alloc+0xfc/0x140
>
> and I'm looking at my x86-64 compile, and it has a stack frame of just 8
> bytes (!) for local variables plus the save/restore area (which looks like
> three registers plus frame pointer plus return address). IOW, if I'm
> looking at the code right (so big caveat: I did _not_ do a real stack
> dump!) the x86-64 stack cost for that same function is on the order of 48
> bytes. Not 160.
>
> Where does that factor-of-three+ difference come from? From the numbers, I
> suspect ppc64 has a 32-byte stack alignment, which may be part of it, and
> I guess the compiler is more eager to use all those extra registers and
> will happily have many more callee-saved regs that are actually used.
>
> But that still a _lot_ of extra stack.
>
> Of course, you may have things like spinlock debugging etc enabled. Some
> of our debugging options do tend to blow things up.

Note that kmem_cache_alloc() is likely to contain lots of inlined
functions for both SLAB and SLUB. Perhaps that blows up stack usage on
ppc?



More information about the Linuxppc-dev mailing list