Changes in kernel/arch/arm32/include/barrier.h [9d58539:c124dce3] in mainline

File:

• kernel/arch/arm32/include/barrier.h (modified) (1 diff)

kernel/arch/arm32/include/barrier.h

@@ -37 +37 @@
 #define KERN_arm32_BARRIER_H_
 
-/*
- * TODO: implement true ARM memory barriers for macros below.
- */
+#ifdef KERNEL
+#include <arch/cp15.h>
+#else
+#include <libarch/cp15.h>
+#endif
+
 #define CS_ENTER_BARRIER()  asm volatile ("" ::: "memory")
 #define CS_LEAVE_BARRIER()  asm volatile ("" ::: "memory")
 
+#if defined PROCESSOR_ARCH_armv7_a
+/* ARMv7 uses instructions for memory barriers see ARM Architecture reference
+ * manual for details:
+ * DMB: ch. A8.8.43 page A8-376
+ * DSB: ch. A8.8.44 page A8-378
+ * See ch. A3.8.3 page A3-148 for details about memory barrier implementation
+ * and functionality on armv7 architecture.
+ */
+#define memory_barrier()  asm volatile ("dmb" ::: "memory")
+#define read_barrier()    asm volatile ("dsb" ::: "memory")
+#define write_barrier()   asm volatile ("dsb st" ::: "memory")
+#define inst_barrier()    asm volatile ("isb" ::: "memory")
+#elif defined PROCESSOR_ARCH_armv6 | defined KERNEL
+/*
+ * ARMv6 introduced user access of the following commands:
+ * - Prefetch flush
+ * - Data synchronization barrier
+ * - Data memory barrier
+ * - Clean and prefetch range operations.
+ * ARM Architecture Reference Manual version I ch. B.3.2.1 p. B3-4
+ */
+/* ARMv6- use system control coprocessor (CP15) for memory barrier instructions.
+ * Although at least mcr p15, 0, r0, c7, c10, 4 is mentioned in earlier archs,
+ * CP15 implementation is mandatory only for armv6+.
+ */
+#define memory_barrier()  CP15DMB_write(0)
+#define read_barrier()    CP15DSB_write(0)
+#define write_barrier()   read_barrier()
+#define inst_barrier()    CP15ISB_write(0)
+#else
+/* Older manuals mention syscalls as a way to implement cache coherency and
+ * barriers. See for example ARM Architecture Reference Manual Version D
+ * chapter 2.7.4 Prefetching and self-modifying code (p. A2-28)
+ */
+// TODO implement on per PROCESSOR basis or via syscalls
 #define memory_barrier()  asm volatile ("" ::: "memory")
 #define read_barrier()    asm volatile ("" ::: "memory")
 #define write_barrier()   asm volatile ("" ::: "memory")
+#define inst_barrier()    asm volatile ("" ::: "memory")
+#endif
 
+/*
+ * There are multiple ways ICache can be implemented on ARM machines. Namely
+ * PIPT, VIPT, and ASID and VMID tagged VIVT (see ARM Architecture Reference
+ * Manual B3.11.2 (p. 1383).  However, CortexA8 Manual states: "For maximum
+ * compatibility across processors, ARM recommends that operating systems target
+ * the ARMv7 base architecture that uses ASID-tagged VIVT instruction caches,
+ * and do not assume the presence of the IVIPT extension. Software that relies
+ * on the IVIPT extension might fail in an unpredictable way on an ARMv7
+ * implementation that does not include the IVIPT extension." (7.2.6 p. 245).
+ * Only PIPT invalidates cache for all VA aliases if one block is invalidated.
+ *
+ * @note: Supporting ASID and VMID tagged VIVT may need to add ICache
+ * maintenance to other places than just smc.
+ */
+
+#if defined PROCESSOR_ARCH_armv7_a | defined PROCESSOR_ARCH_armv6 | defined KERNEL
+/* Available on all supported arms,
+ * invalidates entire ICache so the written value does not matter. */
+//TODO might be PL1 only on armv5-
+#define smc_coherence(a) \
+do { \
+        DCCMVAU_write((uint32_t)(a));  /* Flush changed memory */\
+        write_barrier();               /* Wait for completion */\
+        ICIALLU_write(0);              /* Flush ICache */\
+        inst_barrier();                /* Wait for Inst refetch */\
+} while (0)
+/* @note: Cache type register is not available in uspace. We would need
+ * to export the cache line value, or use syscall for uspace smc_coherence */
+#define smc_coherence_block(a, l) \
+do { \
+        for (uintptr_t addr = (uintptr_t)a; addr < (uintptr_t)a + l; addr += 4)\
+                smc_coherence(addr); \
+} while (0)
+#else
 #define smc_coherence(a)
 #define smc_coherence_block(a, l)
+#endif
+
 
 #endif