ARM: entry: efficiency cleanups

Make the "fast" syscall return path fast again.  The addition of IRQ
tracing and context tracking has made this path grossly inefficient.
We can do much better if these options are enabled if we save the
syscall return code on the stack - we then don't need to save a bunch
of registers around every single callout to C code.

Acked-by: Will Deacon <will.deacon@arm.com>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>

1 files changed, 45 insertions, 16 deletions

diff --git a/arch/arm/kernel/entry-common.S b/arch/arm/kernel/entry-common.S
index 92828a1dec80..dd3721d1185e 100644
--- a/arch/arm/kernel/entry-common.S
+++ b/arch/arm/kernel/entry-common.S
@@ -24,35 +24,55 @@
 
 
 	.align	5
+#if !(IS_ENABLED(CONFIG_TRACE_IRQFLAGS) || IS_ENABLED(CONFIG_CONTEXT_TRACKING))
 /*
- * This is the fast syscall return path.  We do as little as
- * possible here, and this includes saving r0 back into the SVC
- * stack.
+ * This is the fast syscall return path.  We do as little as possible here,
+ * such as avoiding writing r0 to the stack.  We only use this path if we
+ * have tracing and context tracking disabled - the overheads from those
+ * features make this path too inefficient.
  */
 ret_fast_syscall:
  UNWIND(.fnstart	)
  UNWIND(.cantunwind	)
-	disable_irq				@ disable interrupts
+	disable_irq_notrace			@ disable interrupts
 	ldr	r1, [tsk, #TI_FLAGS]		@ re-check for syscall tracing
-	tst	r1, #_TIF_SYSCALL_WORK
-	bne	__sys_trace_return
-	tst	r1, #_TIF_WORK_MASK
+	tst	r1, #_TIF_SYSCALL_WORK | _TIF_WORK_MASK
 	bne	fast_work_pending
-	asm_trace_hardirqs_on
 
 	/* perform architecture specific actions before user return */
 	arch_ret_to_user r1, lr
-	ct_user_enter
 
 	restore_user_regs fast = 1, offset = S_OFF
  UNWIND(.fnend		)
+ENDPROC(ret_fast_syscall)
 
-/*
- * Ok, we need to do extra processing, enter the slow path.
- */
+	/* Ok, we need to do extra processing, enter the slow path. */
 fast_work_pending:
 	str	r0, [sp, #S_R0+S_OFF]!		@ returned r0
-work_pending:
+	/* fall through to work_pending */
+#else
+/*
+ * The "replacement" ret_fast_syscall for when tracing or context tracking
+ * is enabled.  As we will need to call out to some C functions, we save
+ * r0 first to avoid needing to save registers around each C function call.
+ */
+ret_fast_syscall:
+ UNWIND(.fnstart	)
+ UNWIND(.cantunwind	)
+	str	r0, [sp, #S_R0 + S_OFF]!	@ save returned r0
+	disable_irq_notrace			@ disable interrupts
+	ldr	r1, [tsk, #TI_FLAGS]		@ re-check for syscall tracing
+	tst	r1, #_TIF_SYSCALL_WORK | _TIF_WORK_MASK
+	beq	no_work_pending
+ UNWIND(.fnend		)
+ENDPROC(ret_fast_syscall)
+
+	/* Slower path - fall through to work_pending */
+#endif
+
+	tst	r1, #_TIF_SYSCALL_WORK
+	bne	__sys_trace_return_nosave
+slow_work_pending:
 	mov	r0, sp				@ 'regs'
 	mov	r2, why				@ 'syscall'
 	bl	do_work_pending
@@ -64,16 +84,19 @@ work_pending:
 
 /*
  * "slow" syscall return path.  "why" tells us if this was a real syscall.
+ * IRQs may be enabled here, so always disable them.  Note that we use the
+ * "notrace" version to avoid calling into the tracing code unnecessarily.
+ * do_work_pending() will update this state if necessary.
  */
 ENTRY(ret_to_user)
 ret_slow_syscall:
-	disable_irq				@ disable interrupts
+	disable_irq_notrace			@ disable interrupts
 ENTRY(ret_to_user_from_irq)
 	ldr	r1, [tsk, #TI_FLAGS]
 	tst	r1, #_TIF_WORK_MASK
-	bne	work_pending
+	bne	slow_work_pending
 no_work_pending:
-	asm_trace_hardirqs_on
+	asm_trace_hardirqs_on save = 0
 
 	/* perform architecture specific actions before user return */
 	arch_ret_to_user r1, lr
@@ -251,6 +274,12 @@ __sys_trace_return:
 	bl	syscall_trace_exit
 	b	ret_slow_syscall
 
+__sys_trace_return_nosave:
+	asm_trace_hardirqs_off save=0
+	mov	r0, sp
+	bl	syscall_trace_exit
+	b	ret_slow_syscall
+
 	.align	5
 #ifdef CONFIG_ALIGNMENT_TRAP
 	.type	__cr_alignment, #object