]> pilppa.com Git - linux-2.6-omap-h63xx.git/commit
[POWERPC] Make current preempt-safe
authorHugh Dickins <hugh@veritas.com>
Tue, 31 Oct 2006 18:39:31 +0000 (18:39 +0000)
committerPaul Mackerras <paulus@samba.org>
Wed, 1 Nov 2006 03:52:48 +0000 (14:52 +1100)
commit5fe8e8b88e68e517637e3f8287f1fee89e2d9252
tree5b7d1878e9d16107d7c8fcfa0d866d660372161d
parent302439d2167e0f1e01a6480ac40c06063f4e16a1
[POWERPC] Make current preempt-safe

Repeated -j20 kernel builds on a G5 Quad running an SMP PREEMPT kernel
would often collapse within a day, some exec failing with "Bad address".
In each case examined, load_elf_binary was doing a kernel_read, but
generic_file_aio_read's access_ok saw current->thread.fs.seg as USER_DS
instead of KERNEL_DS.

objdump of filemap.o shows gcc 4.1.0 emitting "mr r5,r13 ... ld r9,416(r5)"
here for get_paca()->__current, instead of the expected and much more usual
"ld r9,416(r13)"; I've seen other gcc4s do the same, but perhaps not gcc3s.

So, if the task is preempted and rescheduled on a different cpu in between
the mr and the ld, r5 will be looking at a different paca_struct from the
one it's now on, pick up the wrong __current, and perhaps the wrong seg.
Presumably much worse could happen elsewhere, though that split is rare.

Other architectures appear to be safe (x86_64's read_pda is more limiting
than get_paca), but ppc64 needs to force "current" into one instruction.

Signed-off-by: Hugh Dickins <hugh@veritas.com>
Signed-off-by: Paul Mackerras <paulus@samba.org>
include/asm-powerpc/current.h