diff options
author | Chris Metcalf <cmetcalf@tilera.com> | 2013-09-26 13:24:53 -0400 |
---|---|---|
committer | Greg Kroah-Hartman <gregkh@linuxfoundation.org> | 2013-10-13 16:08:34 -0700 |
commit | bc30c3576d85f7d02f5620cbe8d03f4064c6a73a (patch) | |
tree | b9f7610b3391bc56629e74ae043df04ef1be9cea | |
parent | 22c748d6172867030c228d4737fdef0e9713a125 (diff) | |
download | lwn-bc30c3576d85f7d02f5620cbe8d03f4064c6a73a.tar.gz lwn-bc30c3576d85f7d02f5620cbe8d03f4064c6a73a.zip |
tile: use a more conservative __my_cpu_offset in CONFIG_PREEMPT
commit f862eefec0b68e099a9fa58d3761ffb10bad97e1 upstream.
It turns out the kernel relies on barrier() to force a reload of the
percpu offset value. Since we can't easily modify the definition of
barrier() to include "tp" as an output register, we instead provide a
definition of __my_cpu_offset as extended assembly that includes a fake
stack read to hazard against barrier(), forcing gcc to know that it
must reread "tp" and recompute anything based on "tp" after a barrier.
This fixes observed hangs in the slub allocator when we are looping
on a percpu cmpxchg_double.
A similar fix for ARMv7 was made in June in change 509eb76ebf97.
Signed-off-by: Chris Metcalf <cmetcalf@tilera.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
-rw-r--r-- | arch/tile/include/asm/percpu.h | 34 |
1 files changed, 31 insertions, 3 deletions
diff --git a/arch/tile/include/asm/percpu.h b/arch/tile/include/asm/percpu.h index 63294f5a8efb..4f7ae39fa202 100644 --- a/arch/tile/include/asm/percpu.h +++ b/arch/tile/include/asm/percpu.h @@ -15,9 +15,37 @@ #ifndef _ASM_TILE_PERCPU_H #define _ASM_TILE_PERCPU_H -register unsigned long __my_cpu_offset __asm__("tp"); -#define __my_cpu_offset __my_cpu_offset -#define set_my_cpu_offset(tp) (__my_cpu_offset = (tp)) +register unsigned long my_cpu_offset_reg asm("tp"); + +#ifdef CONFIG_PREEMPT +/* + * For full preemption, we can't just use the register variable + * directly, since we need barrier() to hazard against it, causing the + * compiler to reload anything computed from a previous "tp" value. + * But we also don't want to use volatile asm, since we'd like the + * compiler to be able to cache the value across multiple percpu reads. + * So we use a fake stack read as a hazard against barrier(). + * The 'U' constraint is like 'm' but disallows postincrement. + */ +static inline unsigned long __my_cpu_offset(void) +{ + unsigned long tp; + register unsigned long *sp asm("sp"); + asm("move %0, tp" : "=r" (tp) : "U" (*sp)); + return tp; +} +#define __my_cpu_offset __my_cpu_offset() +#else +/* + * We don't need to hazard against barrier() since "tp" doesn't ever + * change with PREEMPT_NONE, and with PREEMPT_VOLUNTARY it only + * changes at function call points, at which we are already re-reading + * the value of "tp" due to "my_cpu_offset_reg" being a global variable. + */ +#define __my_cpu_offset my_cpu_offset_reg +#endif + +#define set_my_cpu_offset(tp) (my_cpu_offset_reg = (tp)) #include <asm-generic/percpu.h> |