On 08/29/2012 03:34 PM, Jim Quinlan wrote:
For non MIPSr2 processors, such as the BMIPS 5000, calls to
arch_local_irq_disable() and others may be preempted, and in doing
so a stale value may be restored to c0_status. This fix disables
preemption for such processors prior to the call and enables it
after the call.
This bug was observed in a BMIPS 5000, occuring once every few hours
in a continuous reboot test. It was traced to the write_lock_irq()
function which was being invoked in release_task() in exit.c.
By placing a number of "nops" inbetween the mfc0/mtc0 pair in
arch_local_irq_disable(), which is called by write_lock_irq(), we
were able to greatly increase the occurance of this bug. Similarly,
the application of this commit silenced the bug.
It is better to use the preemption functions declared in <linux/preempt.h>
rather than defining new ones as is done in this commit. However,
including that file from irqflags effected many compiler errors.
Signed-off-by: Jim Quinlan <email@example.com>
arch/mips/include/asm/irqflags.h | 81 ++++++++++++++++++++++++++++++++++++++
1 files changed, 81 insertions(+), 0 deletions(-)
diff --git a/arch/mips/include/asm/irqflags.h b/arch/mips/include/asm/irqflags.h
index 309cbcd..d6e71ed 100644
@@ -16,6 +16,71 @@
+#if defined(__GENERATING_BOUNDS_H) || defined(__GENERATING_OFFSETS_S)
+#define __TI_PRE_COUNT (-1)
+#define __TI_PRE_COUNT TI_PRE_COUNT
+ * Non-mipsr2 processors executing functions such as arch_local_irq_disable()
+ * are not preempt-safe: if preemption occurs between the mfc0 and the mtc0,
+ * a stale status value may be stored. To prevent this, we define
+ * here arch_local_preempt_disable() and arch_local_preempt_enable(), which
+ * are called before the mfc0 and after the mtc0, respectively. A better
+ * solution would "#include <linux/preempt.h> and use its declared routines,
+ * but that is not viable due to numerous compile errors.
I'm with Ralf's idea from the other branch of the thread. Put all this
non-mipsr2 stuff out of line (perhaps creating lib/mips-atomic.c).
+ * MipsR2 processors with atomic interrupt enable/disable instructions
+ * (ei/di) do not have this issue.
For mipsr2, we leave them alone so they can be inlined.
This way you shouldn't need the ugly #include hackery.