selftests/powerpc: Test preservation of FPU and VMX regs across preemption

[mirror_ubuntu-zesty-kernel.git] / tools / testing / selftests / powerpc / math / vmx_asm.S

diff --git a/tools/testing/selftests/powerpc/math/vmx_asm.S b/tools/testing/selftests/powerpc/math/vmx_asm.S
index 4ce64212d5e0c43fbbfa4522775a9a7c22006992..1b8c248b3ac18770a23a30c3946da27d92aa708a 100644 (file)
--- a/tools/testing/selftests/powerpc/math/vmx_asm.S
+++ b/tools/testing/selftests/powerpc/math/vmx_asm.S
@@ -9,6 +9,7 @@
 
 #include "../basic_asm.h"
 
+# POS MUST BE 16 ALIGNED!
 #define PUSH_VMX(pos,reg) \
        li      reg,pos; \
        stvx    v20,reg,sp; \
@@ -35,6 +36,7 @@
        addi    reg,reg,16; \
        stvx    v31,reg,sp;
 
+# POS MUST BE 16 ALIGNED!
 #define POP_VMX(pos,reg) \
        li      reg,pos; \
        lvx     v20,reg,sp; \
@@ -93,7 +95,7 @@ FUNC_END(load_vmx)
 
 # Should be safe from C, only touches r4, r5 and v0,v1,v2
 FUNC_START(check_vmx)
-       PUSH_BASIC_STACK(16)
+       PUSH_BASIC_STACK(32)
        mr r4,r3
        li      r3,1 # assume a bad result
        li      r5,0
@@ -162,7 +164,7 @@ FUNC_START(check_vmx)
        cmpdi   r0,0xffffffffffffffff
        bne     1f
        li      r3,0
-1:     POP_BASIC_STACK(16)
+1:     POP_BASIC_STACK(32)
        blr
 FUNC_END(check_vmx)
 
@@ -193,3 +195,41 @@ FUNC_START(test_vmx)
        POP_BASIC_STACK(512)
        blr
 FUNC_END(test_vmx)
+
+# int preempt_vmx(vector int *varray, int *threads_starting, int *running)
+# On starting will (atomically) decrement threads_starting as a signal that
+# the VMX have been loaded with varray. Will proceed to check the validity of
+# the VMX registers while running is not zero.
+FUNC_START(preempt_vmx)
+       PUSH_BASIC_STACK(512)
+       std r3,STACK_FRAME_PARAM(0)(sp) # vector int *varray
+       std r4,STACK_FRAME_PARAM(1)(sp) # int *threads_starting
+       std r5,STACK_FRAME_PARAM(2)(sp) # int *running
+       # VMX need to write to 16 byte aligned addresses, skip STACK_FRAME_LOCAL(3,0)
+       PUSH_VMX(STACK_FRAME_LOCAL(4,0),r4)
+
+       bl load_vmx
+       nop
+
+       sync
+       # Atomic DEC
+       ld r3,STACK_FRAME_PARAM(1)(sp)
+1:     lwarx r4,0,r3
+       addi r4,r4,-1
+       stwcx. r4,0,r3
+       bne- 1b
+
+2:     ld r3,STACK_FRAME_PARAM(0)(sp)
+       bl check_vmx
+       nop
+       cmpdi r3,0
+       bne 3f
+       ld r4,STACK_FRAME_PARAM(2)(sp)
+       ld r5,0(r4)
+       cmpwi r5,0
+       bne 2b
+
+3:     POP_VMX(STACK_FRAME_LOCAL(4,0),r4)
+       POP_BASIC_STACK(512)
+       blr
+FUNC_END(preempt_vmx)