make ppc64 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: Adrian Bunk <bunk@stusta.de>

diff --git a/include/asm-powerpc/current.h b/include/asm-powerpc/current.h
index 1938d6abd255056f46711b2786c3720123d90445..b8708aedf925327c13e727d094ad6398b3156850 100644 (file)
--- a/include/asm-powerpc/current.h
+++ b/include/asm-powerpc/current.h
@@ -14,7 +14,17 @@ struct task_struct;
 #ifdef __powerpc64__
 #include <asm/paca.h>
 
-#define current                (get_paca()->__current)
+static inline struct task_struct *get_current(void)
+{
+       struct task_struct *task;
+
+       __asm__ __volatile__("ld %0,%1(13)"
+       : "=r" (task)
+       : "i" (offsetof(struct paca_struct, __current)));
+
+       return task;
+}
+#define current        get_current()
 
 #else