[rtems commit] Fix interrupt epilogue for ARMv7-AR and PowerPC

[Sebastian Huber]

Module:    rtems
Branch:    4.11
Commit:    0c9bf40b89cd6763f9ec5d913d440c8b0074a092
Changeset: http://git.rtems.org/rtems/commit/?id=0c9bf40b89cd6763f9ec5d913d440c8b0074a092

Author:    Sebastian Huber <sebastian.huber at embedded-brains.de>
Date:      Wed Nov 11 11:49:45 2015 +0100

Fix interrupt epilogue for ARMv7-AR and PowerPC

Close #2470.

---

 .../bspsupport/ppc_exc_async_normal.S              | 51 ++++++++++++++++------
 cpukit/score/cpu/arm/arm_exc_interrupt.S           | 27 ++++++++++++
 2 files changed, 65 insertions(+), 13 deletions(-)

diff --git a/c/src/lib/libcpu/powerpc/new-exceptions/bspsupport/ppc_exc_async_normal.S b/c/src/lib/libcpu/powerpc/new-exceptions/bspsupport/ppc_exc_async_normal.S
index 59e621f..0e71dad 100644
--- a/c/src/lib/libcpu/powerpc/new-exceptions/bspsupport/ppc_exc_async_normal.S
+++ b/c/src/lib/libcpu/powerpc/new-exceptions/bspsupport/ppc_exc_async_normal.S
@@ -29,29 +29,29 @@
 #define SCRATCH_5_REGISTER r12
 #define FRAME_REGISTER r14
 
-#define VECTOR_OFFSET(reg) GPR4_OFFSET(reg)
-#define SELF_CPU_OFFSET(reg) GPR5_OFFSET(reg)
-#define ISR_NEST_OFFSET(reg) GPR6_OFFSET(reg)
-#define DISPATCH_LEVEL_OFFSET(reg) GPR7_OFFSET(reg)
-#define HANDLER_OFFSET(reg) GPR8_OFFSET(reg)
-#define SCRATCH_0_OFFSET(reg) GPR0_OFFSET(reg)
-#define SCRATCH_1_OFFSET(reg) GPR3_OFFSET(reg)
-#define SCRATCH_2_OFFSET(reg) GPR9_OFFSET(reg)
-#define SCRATCH_3_OFFSET(reg) GPR10_OFFSET(reg)
-#define SCRATCH_4_OFFSET(reg) GPR11_OFFSET(reg)
-#define SCRATCH_5_OFFSET(reg) GPR12_OFFSET(reg)
+#define VECTOR_OFFSET GPR4_OFFSET
+#define SELF_CPU_OFFSET GPR5_OFFSET
+#define ISR_NEST_OFFSET GPR6_OFFSET
+#define DISPATCH_LEVEL_OFFSET GPR7_OFFSET
+#define HANDLER_OFFSET GPR8_OFFSET
+#define SCRATCH_0_OFFSET GPR0_OFFSET
+#define SCRATCH_1_OFFSET GPR3_OFFSET
+#define SCRATCH_2_OFFSET GPR9_OFFSET
+#define SCRATCH_3_OFFSET GPR10_OFFSET
+#define SCRATCH_4_OFFSET GPR11_OFFSET
+#define SCRATCH_5_OFFSET GPR12_OFFSET
 
 /*
  * The register 2 slot is free, since this is the read-only small data anchor.
  */
-#define FRAME_OFFSET(reg) GPR2_OFFSET(reg)
+#define FRAME_OFFSET GPR2_OFFSET
 
 #ifdef RTEMS_PROFILING
 /*
  * The PPC_EXC_MINIMAL_FRAME_SIZE is enough to store this additional register.
  */
 #define ENTRY_INSTANT_REGISTER r15
-#define ENTRY_INSTANT_OFFSET(reg) GPR13_OFFSET(reg)
+#define ENTRY_INSTANT_OFFSET GPR13_OFFSET
 
 .macro GET_TIME_BASE REG
 #ifdef ppc8540
@@ -399,6 +399,31 @@ thread_dispatching_done:
 	evldd	HANDLER_REGISTER, PPC_EXC_ACC_OFFSET(r1)
 #endif
 
+	/*
+	 * We must clear reservations here, since otherwise compare-and-swap
+	 * atomic operations with interrupts enabled may yield wrong results.
+	 * A compare-and-swap atomic operation is generated by the compiler
+	 * like this:
+	 *
+	 *   .L1:
+	 *     lwarx  r9, r0, r3
+	 *     cmpw   r9, r4
+	 *     bne-   .L2
+	 *     stwcx. r5, r0, r3
+	 *     bne-   .L1
+	 *   .L2:
+	 *
+	 * Consider the following scenario.  A thread is interrupted right
+	 * before the stwcx.  The interrupt updates the value using a
+	 * compare-and-swap sequence.  Everything is fine up to this point.
+	 * The interrupt performs now a compare-and-swap sequence which fails
+	 * with a branch to .L2.  The current processor has now a reservation.
+	 * The interrupt returns without further stwcx.  The thread updates the
+	 * value using the unrelated reservation of the interrupt.
+	 */
+	li	SCRATCH_0_REGISTER, FRAME_OFFSET
+	stwcx.	SCRATCH_0_REGISTER, r1, SCRATCH_0_REGISTER
+
 	/* Load SRR0, SRR1, CR, CTR, XER, and LR */
 	lwz	SCRATCH_0_REGISTER, SRR0_FRAME_OFFSET(r1)
 	lwz	SCRATCH_1_REGISTER, SRR1_FRAME_OFFSET(r1)
diff --git a/cpukit/score/cpu/arm/arm_exc_interrupt.S b/cpukit/score/cpu/arm/arm_exc_interrupt.S
index 7930c32..fcb1510 100644
--- a/cpukit/score/cpu/arm/arm_exc_interrupt.S
+++ b/cpukit/score/cpu/arm/arm_exc_interrupt.S
@@ -209,6 +209,33 @@ thread_dispatch_done:
 	/* Restore EXCHANGE_LR and EXCHANGE_SPSR registers from exchange area */
 	ldmia	sp!, {EXCHANGE_LR, EXCHANGE_SPSR}
 
+#ifdef ARM_MULTILIB_HAS_LOAD_STORE_EXCLUSIVE
+	/*
+	 * We must clear reservations here, since otherwise compare-and-swap
+	 * atomic operations with interrupts enabled may yield wrong results.
+	 * A compare-and-swap atomic operation is generated by the compiler
+	 * like this:
+	 *
+	 *   .L1:
+	 *     ldrex r1, [r0]
+	 *     cmp   r1, r3
+	 *     bne   .L2
+	 *     strex r3, r2, [r0]
+	 *     cmp   r3, #0
+	 *     bne   .L1
+	 *   .L2:
+	 *
+	 * Consider the following scenario.  A thread is interrupted right
+	 * before the strex.  The interrupt updates the value using a
+	 * compare-and-swap sequence.  Everything is fine up to this point.
+	 * The interrupt performs now a compare-and-swap sequence which fails
+	 * with a branch to .L2.  The current processor has now a reservation.
+	 * The interrupt returns without further strex.  The thread updates the
+	 * value using the unrelated reservation of the interrupt.
+	 */
+	clrex
+#endif
+
 	/* Return from interrupt */
 	subs	pc, lr, #4