s390/fpu: improve kernel_fpu_[begin|end]

In case of nested user of the FPU or vector registers in the kernel
the current code uses the mask of the FPU/vector registers of the
previous contexts to decide which registers to save and restore.
E.g. if the previous context used KERNEL_VXR_V0V7 and the next
context wants to use KERNEL_VXR_V24V31 the first 8 vector registers
are stored to the FPU state structure. But this is not necessary
as the next context does not use these registers.

Rework the FPU/vector register save and restore code. The new code
does a few things differently:
1) A lowcore field is used instead of a per-cpu variable.
2) The kernel_fpu_end function now has two parameters just like
   kernel_fpu_begin. The register flags are required by both
   functions to save / restore the minimal register set.
3) The inline functions kernel_fpu_begin/kernel_fpu_end now do the
   update of the register masks. If the user space FPU registers
   have already been stored neither save_fpu_regs nor the
   __kernel_fpu_begin/__kernel_fpu_end functions have to be called
   for the first context. In this case kernel_fpu_begin adds 7
   instructions and kernel_fpu_end adds 4 instructions.
3) The inline assemblies in __kernel_fpu_begin / __kernel_fpu_end
   to save / restore the vector registers are simplified a bit.

Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com>

1 files changed, 122 insertions, 195 deletions

diff --git a/arch/s390/kernel/fpu.c b/arch/s390/kernel/fpu.c
index 81d1d1887507..1235b9438df4 100644
--- a/arch/s390/kernel/fpu.c
+++ b/arch/s390/kernel/fpu.c
@@ -10,240 +10,167 @@
 #include <asm/fpu/types.h>
 #include <asm/fpu/api.h>
 
-/*
- * Per-CPU variable to maintain FPU register ranges that are in use
- * by the kernel.
- */
-static DEFINE_PER_CPU(u32, kernel_fpu_state);
-
-#define KERNEL_FPU_STATE_MASK	(KERNEL_FPU_MASK|KERNEL_FPC)
-
+asm(".include \"asm/vx-insn.h\"\n");
 
 void __kernel_fpu_begin(struct kernel_fpu *state, u32 flags)
 {
-	if (!__this_cpu_read(kernel_fpu_state)) {
-		/*
-		 * Save user space FPU state and register contents.  Multiple
-		 * calls because of interruptions do not matter and return
-		 * immediately.  This also sets CIF_FPU to lazy restore FP/VX
-		 * register contents when returning to user space.
-		 */
-		save_fpu_regs();
-	}
-
-	/* Update flags to use the vector facility for KERNEL_FPR */
-	if (MACHINE_HAS_VX && (state->mask & KERNEL_FPR)) {
-		flags |= KERNEL_VXR_LOW | KERNEL_FPC;
-		flags &= ~KERNEL_FPR;
-	}
-
-	/* Save and update current kernel VX state */
-	state->mask = __this_cpu_read(kernel_fpu_state);
-	__this_cpu_or(kernel_fpu_state, flags & KERNEL_FPU_STATE_MASK);
-
 	/*
-	 * If this is the first call to __kernel_fpu_begin(), no additional
-	 * work is required.
+	 * Limit the save to the FPU/vector registers already
+	 * in use by the previous context
 	 */
-	if (!(state->mask & KERNEL_FPU_STATE_MASK))
-		return;
+	flags &= state->mask;
 
-	/*
-	 * If KERNEL_FPR is still set, the vector facility is not available
-	 * and, thus, save floating-point control and registers only.
-	 */
-	if (state->mask & KERNEL_FPR) {
-		asm volatile("stfpc %0" : "=Q" (state->fpc));
-		asm volatile("std 0,%0" : "=Q" (state->fprs[0]));
-		asm volatile("std 1,%0" : "=Q" (state->fprs[1]));
-		asm volatile("std 2,%0" : "=Q" (state->fprs[2]));
-		asm volatile("std 3,%0" : "=Q" (state->fprs[3]));
-		asm volatile("std 4,%0" : "=Q" (state->fprs[4]));
-		asm volatile("std 5,%0" : "=Q" (state->fprs[5]));
-		asm volatile("std 6,%0" : "=Q" (state->fprs[6]));
-		asm volatile("std 7,%0" : "=Q" (state->fprs[7]));
-		asm volatile("std 8,%0" : "=Q" (state->fprs[8]));
-		asm volatile("std 9,%0" : "=Q" (state->fprs[9]));
-		asm volatile("std 10,%0" : "=Q" (state->fprs[10]));
-		asm volatile("std 11,%0" : "=Q" (state->fprs[11]));
-		asm volatile("std 12,%0" : "=Q" (state->fprs[12]));
-		asm volatile("std 13,%0" : "=Q" (state->fprs[13]));
-		asm volatile("std 14,%0" : "=Q" (state->fprs[14]));
-		asm volatile("std 15,%0" : "=Q" (state->fprs[15]));
+	if (flags & KERNEL_FPC)
+		/* Save floating point control */
+		asm volatile("stfpc %0" : "=m" (state->fpc));
+
+	if (!MACHINE_HAS_VX) {
+		if (flags & KERNEL_VXR_V0V7) {
+			/* Save floating-point registers */
+			asm volatile("std 0,%0" : "=Q" (state->fprs[0]));
+			asm volatile("std 1,%0" : "=Q" (state->fprs[1]));
+			asm volatile("std 2,%0" : "=Q" (state->fprs[2]));
+			asm volatile("std 3,%0" : "=Q" (state->fprs[3]));
+			asm volatile("std 4,%0" : "=Q" (state->fprs[4]));
+			asm volatile("std 5,%0" : "=Q" (state->fprs[5]));
+			asm volatile("std 6,%0" : "=Q" (state->fprs[6]));
+			asm volatile("std 7,%0" : "=Q" (state->fprs[7]));
+			asm volatile("std 8,%0" : "=Q" (state->fprs[8]));
+			asm volatile("std 9,%0" : "=Q" (state->fprs[9]));
+			asm volatile("std 10,%0" : "=Q" (state->fprs[10]));
+			asm volatile("std 11,%0" : "=Q" (state->fprs[11]));
+			asm volatile("std 12,%0" : "=Q" (state->fprs[12]));
+			asm volatile("std 13,%0" : "=Q" (state->fprs[13]));
+			asm volatile("std 14,%0" : "=Q" (state->fprs[14]));
+			asm volatile("std 15,%0" : "=Q" (state->fprs[15]));
+		}
 		return;
 	}
 
-	/*
-	 * If this is a nested call to __kernel_fpu_begin(), check the saved
-	 * state mask to save and later restore the vector registers that
-	 * are already in use.	Let's start with checking floating-point
-	 * controls.
-	 */
-	if (state->mask & KERNEL_FPC)
-		asm volatile("stfpc %0" : "=m" (state->fpc));
-
 	/* Test and save vector registers */
 	asm volatile (
 		/*
 		 * Test if any vector register must be saved and, if so,
 		 * test if all register can be saved.
 		 */
-		"	tmll	%[m],15\n"	/* KERNEL_VXR_MASK */
-		"	jz	20f\n"		/* no work -> done */
 		"	la	1,%[vxrs]\n"	/* load save area */
-		"	jo	18f\n"		/* -> save V0..V31 */
-
+		"	tmll	%[m],30\n"	/* KERNEL_VXR */
+		"	jz	7f\n"		/* no work -> done */
+		"	jo	5f\n"		/* -> save V0..V31 */
 		/*
-		 * Test if V8..V23 can be saved at once... this speeds up
-		 * for KERNEL_fpu_MID only. Otherwise continue to split the
-		 * range of vector registers into two halves and test them
-		 * separately.
+		 * Test for special case KERNEL_FPU_MID only. In this
+		 * case a vstm V8..V23 is the best instruction
 		 */
-		"	tmll	%[m],6\n"	/* KERNEL_VXR_MID */
-		"	jo	17f\n"		/* -> save V8..V23 */
-
+		"	chi	%[m],12\n"	/* KERNEL_VXR_MID */
+		"	jne	0f\n"		/* -> save V8..V23 */
+		"	VSTM	8,23,128,1\n"	/* vstm %v8,%v23,128(%r1) */
+		"	j	7f\n"
 		/* Test and save the first half of 16 vector registers */
-		"1:	tmll	%[m],3\n"	/* KERNEL_VXR_LOW */
-		"	jz	10f\n"		/* -> KERNEL_VXR_HIGH */
+		"0:	tmll	%[m],6\n"	/* KERNEL_VXR_LOW */
+		"	jz	3f\n"		/* -> KERNEL_VXR_HIGH */
 		"	jo	2f\n"		/* 11 -> save V0..V15 */
-		"	brc	4,3f\n"		/* 01 -> save V0..V7  */
-		"	brc	2,4f\n"		/* 10 -> save V8..V15 */
-
+		"	brc	2,1f\n"		/* 10 -> save V8..V15 */
+		"	VSTM	0,7,0,1\n"	/* vstm %v0,%v7,0(%r1) */
+		"	j	3f\n"
+		"1:	VSTM	8,15,128,1\n"	/* vstm %v8,%v15,128(%r1) */
+		"	j	3f\n"
+		"2:	VSTM	0,15,0,1\n"	/* vstm %v0,%v15,0(%r1) */
 		/* Test and save the second half of 16 vector registers */
-		"10:	tmll	%[m],12\n"	/* KERNEL_VXR_HIGH */
-		"	jo	19f\n"		/* 11 -> save V16..V31 */
-		"	brc	4,11f\n"	/* 01 -> save V16..V23	*/
-		"	brc	2,12f\n"	/* 10 -> save V24..V31 */
-		"	j	20f\n"		/* 00 -> done */
-
-		/*
-		 * Below are the vstm combinations to save multiple vector
-		 * registers at once.
-		 */
-		"2:	.word	0xe70f,0x1000,0x003e\n"	/* vstm 0,15,0(1) */
-		"	j	10b\n"			/* -> VXR_HIGH */
-		"3:	.word	0xe707,0x1000,0x003e\n" /* vstm 0,7,0(1) */
-		"	j	10b\n"			/* -> VXR_HIGH */
-		"4:	.word	0xe78f,0x1080,0x003e\n" /* vstm 8,15,128(1) */
-		"	j	10b\n"			/* -> VXR_HIGH */
-		"\n"
-		"11:	.word	0xe707,0x1100,0x0c3e\n"	/* vstm 16,23,256(1) */
-		"	j	20f\n"			/* -> done */
-		"12:	.word	0xe78f,0x1180,0x0c3e\n" /* vstm 24,31,384(1) */
-		"	j	20f\n"			/* -> done */
-		"\n"
-		"17:	.word	0xe787,0x1080,0x043e\n"	/* vstm 8,23,128(1) */
-		"	nill	%[m],249\n"		/* m &= ~VXR_MID    */
-		"	j	1b\n"			/* -> VXR_LOW */
-		"\n"
-		"18:	.word	0xe70f,0x1000,0x003e\n"	/* vstm 0,15,0(1) */
-		"19:	.word	0xe70f,0x1100,0x0c3e\n"	/* vstm 16,31,256(1) */
-		"20:"
+		"3:	tmll	%[m],24\n"	/* KERNEL_VXR_HIGH */
+		"	jz	7f\n"
+		"	jo	6f\n"		/* 11 -> save V16..V31 */
+		"	brc	2,4f\n"		/* 10 -> save V24..V31 */
+		"	VSTM	16,23,256,1\n"	/* vstm %v16,%v23,256(%r1) */
+		"	j	7f\n"
+		"4:	VSTM	24,31,384,1\n"	/* vstm %v24,%v31,384(%r1) */
+		"	j	7f\n"
+		"5:	VSTM	0,15,0,1\n"	/* vstm %v0,%v15,0(%r1) */
+		"6:	VSTM	16,31,256,1\n"	/* vstm %v16,%v31,256(%r1) */
+		"7:"
 		: [vxrs] "=Q" (*(struct vx_array *) &state->vxrs)
-		: [m] "d" (state->mask)
+		: [m] "d" (flags)
 		: "1", "cc");
 }
 EXPORT_SYMBOL(__kernel_fpu_begin);
 
-void __kernel_fpu_end(struct kernel_fpu *state)
+void __kernel_fpu_end(struct kernel_fpu *state, u32 flags)
 {
-	/* Just update the per-CPU state if there is nothing to restore */
-	if (!(state->mask & KERNEL_FPU_STATE_MASK))
-		goto update_fpu_state;
-
 	/*
-	 * If KERNEL_FPR is specified, the vector facility is not available
-	 * and, thus, restore floating-point control and registers only.
+	 * Limit the restore to the FPU/vector registers of the
+	 * previous context that have been overwritte by the
+	 * current context
 	 */
-	if (state->mask & KERNEL_FPR) {
-		asm volatile("lfpc %0" : : "Q" (state->fpc));
-		asm volatile("ld 0,%0" : : "Q" (state->fprs[0]));
-		asm volatile("ld 1,%0" : : "Q" (state->fprs[1]));
-		asm volatile("ld 2,%0" : : "Q" (state->fprs[2]));
-		asm volatile("ld 3,%0" : : "Q" (state->fprs[3]));
-		asm volatile("ld 4,%0" : : "Q" (state->fprs[4]));
-		asm volatile("ld 5,%0" : : "Q" (state->fprs[5]));
-		asm volatile("ld 6,%0" : : "Q" (state->fprs[6]));
-		asm volatile("ld 7,%0" : : "Q" (state->fprs[7]));
-		asm volatile("ld 8,%0" : : "Q" (state->fprs[8]));
-		asm volatile("ld 9,%0" : : "Q" (state->fprs[9]));
-		asm volatile("ld 10,%0" : : "Q" (state->fprs[10]));
-		asm volatile("ld 11,%0" : : "Q" (state->fprs[11]));
-		asm volatile("ld 12,%0" : : "Q" (state->fprs[12]));
-		asm volatile("ld 13,%0" : : "Q" (state->fprs[13]));
-		asm volatile("ld 14,%0" : : "Q" (state->fprs[14]));
-		asm volatile("ld 15,%0" : : "Q" (state->fprs[15]));
-		goto update_fpu_state;
-	}
+	flags &= state->mask;
 
-	/* Test and restore floating-point controls */
-	if (state->mask & KERNEL_FPC)
+	if (flags & KERNEL_FPC)
+		/* Restore floating-point controls */
 		asm volatile("lfpc %0" : : "Q" (state->fpc));
 
+	if (!MACHINE_HAS_VX) {
+		if (flags & KERNEL_VXR_V0V7) {
+			/* Restore floating-point registers */
+			asm volatile("ld 0,%0" : : "Q" (state->fprs[0]));
+			asm volatile("ld 1,%0" : : "Q" (state->fprs[1]));
+			asm volatile("ld 2,%0" : : "Q" (state->fprs[2]));
+			asm volatile("ld 3,%0" : : "Q" (state->fprs[3]));
+			asm volatile("ld 4,%0" : : "Q" (state->fprs[4]));
+			asm volatile("ld 5,%0" : : "Q" (state->fprs[5]));
+			asm volatile("ld 6,%0" : : "Q" (state->fprs[6]));
+			asm volatile("ld 7,%0" : : "Q" (state->fprs[7]));
+			asm volatile("ld 8,%0" : : "Q" (state->fprs[8]));
+			asm volatile("ld 9,%0" : : "Q" (state->fprs[9]));
+			asm volatile("ld 10,%0" : : "Q" (state->fprs[10]));
+			asm volatile("ld 11,%0" : : "Q" (state->fprs[11]));
+			asm volatile("ld 12,%0" : : "Q" (state->fprs[12]));
+			asm volatile("ld 13,%0" : : "Q" (state->fprs[13]));
+			asm volatile("ld 14,%0" : : "Q" (state->fprs[14]));
+			asm volatile("ld 15,%0" : : "Q" (state->fprs[15]));
+		}
+		return;
+	}
+
 	/* Test and restore (load) vector registers */
 	asm volatile (
 		/*
-		 * Test if any vector registers must be loaded and, if so,
+		 * Test if any vector register must be loaded and, if so,
 		 * test if all registers can be loaded at once.
 		 */
-		"	tmll	%[m],15\n"	/* KERNEL_VXR_MASK */
-		"	jz	20f\n"		/* no work -> done */
-		"	la	1,%[vxrs]\n"	/* load load area */
-		"	jo	18f\n"		/* -> load V0..V31 */
-
-		/*
-		 * Test if V8..V23 can be restored at once... this speeds up
-		 * for KERNEL_VXR_MID only. Otherwise continue to split the
-		 * range of vector registers into two halves and test them
-		 * separately.
-		 */
-		"	tmll	%[m],6\n"	/* KERNEL_VXR_MID */
-		"	jo	17f\n"		/* -> load V8..V23 */
-
-		/* Test and load the first half of 16 vector registers */
-		"1:	tmll	%[m],3\n"	/* KERNEL_VXR_LOW */
-		"	jz	10f\n"		/* -> KERNEL_VXR_HIGH */
-		"	jo	2f\n"		/* 11 -> load V0..V15 */
-		"	brc	4,3f\n"		/* 01 -> load V0..V7  */
-		"	brc	2,4f\n"		/* 10 -> load V8..V15 */
-
-		/* Test and load the second half of 16 vector registers */
-		"10:	tmll	%[m],12\n"	/* KERNEL_VXR_HIGH */
-		"	jo	19f\n"		/* 11 -> load V16..V31 */
-		"	brc	4,11f\n"	/* 01 -> load V16..V23	*/
-		"	brc	2,12f\n"	/* 10 -> load V24..V31 */
-		"	j	20f\n"		/* 00 -> done */
-
+		"	la	1,%[vxrs]\n"	/* load restore area */
+		"	tmll	%[m],30\n"	/* KERNEL_VXR */
+		"	jz	7f\n"		/* no work -> done */
+		"	jo	5f\n"		/* -> restore V0..V31 */
 		/*
-		 * Below are the vstm combinations to load multiple vector
-		 * registers at once.
+		 * Test for special case KERNEL_FPU_MID only. In this
+		 * case a vlm V8..V23 is the best instruction
 		 */
-		"2:	.word	0xe70f,0x1000,0x0036\n"	/* vlm 0,15,0(1) */
-		"	j	10b\n"			/* -> VXR_HIGH */
-		"3:	.word	0xe707,0x1000,0x0036\n" /* vlm 0,7,0(1) */
-		"	j	10b\n"			/* -> VXR_HIGH */
-		"4:	.word	0xe78f,0x1080,0x0036\n" /* vlm 8,15,128(1) */
-		"	j	10b\n"			/* -> VXR_HIGH */
-		"\n"
-		"11:	.word	0xe707,0x1100,0x0c36\n"	/* vlm 16,23,256(1) */
-		"	j	20f\n"			/* -> done */
-		"12:	.word	0xe78f,0x1180,0x0c36\n" /* vlm 24,31,384(1) */
-		"	j	20f\n"			/* -> done */
-		"\n"
-		"17:	.word	0xe787,0x1080,0x0436\n"	/* vlm 8,23,128(1) */
-		"	nill	%[m],249\n"		/* m &= ~VXR_MID    */
-		"	j	1b\n"			/* -> VXR_LOW */
-		"\n"
-		"18:	.word	0xe70f,0x1000,0x0036\n"	/* vlm 0,15,0(1) */
-		"19:	.word	0xe70f,0x1100,0x0c36\n"	/* vlm 16,31,256(1) */
-		"20:"
-		:
-		: [vxrs] "Q" (*(struct vx_array *) &state->vxrs),
-		  [m] "d" (state->mask)
+		"	chi	%[m],12\n"	/* KERNEL_VXR_MID */
+		"	jne	0f\n"		/* -> restore V8..V23 */
+		"	VLM	8,23,128,1\n"	/* vlm %v8,%v23,128(%r1) */
+		"	j	7f\n"
+		/* Test and restore the first half of 16 vector registers */
+		"0:	tmll	%[m],6\n"	/* KERNEL_VXR_LOW */
+		"	jz	3f\n"		/* -> KERNEL_VXR_HIGH */
+		"	jo	2f\n"		/* 11 -> restore V0..V15 */
+		"	brc	2,1f\n"		/* 10 -> restore V8..V15 */
+		"	VLM	0,7,0,1\n"	/* vlm %v0,%v7,0(%r1) */
+		"	j	3f\n"
+		"1:	VLM	8,15,128,1\n"	/* vlm %v8,%v15,128(%r1) */
+		"	j	3f\n"
+		"2:	VLM	0,15,0,1\n"	/* vlm %v0,%v15,0(%r1) */
+		/* Test and restore the second half of 16 vector registers */
+		"3:	tmll	%[m],24\n"	/* KERNEL_VXR_HIGH */
+		"	jz	7f\n"
+		"	jo	6f\n"		/* 11 -> restore V16..V31 */
+		"	brc	2,4f\n"		/* 10 -> restore V24..V31 */
+		"	VLM	16,23,256,1\n"	/* vlm %v16,%v23,256(%r1) */
+		"	j	7f\n"
+		"4:	VLM	24,31,384,1\n"	/* vlm %v24,%v31,384(%r1) */
+		"	j	7f\n"
+		"5:	VLM	0,15,0,1\n"	/* vlm %v0,%v15,0(%r1) */
+		"6:	VLM	16,31,256,1\n"	/* vlm %v16,%v31,256(%r1) */
+		"7:"
+		: [vxrs] "=Q" (*(struct vx_array *) &state->vxrs)
+		: [m] "d" (flags)
 		: "1", "cc");
-
-update_fpu_state:
-	/* Update current kernel VX state */
-	__this_cpu_write(kernel_fpu_state, state->mask);
 }
 EXPORT_SYMBOL(__kernel_fpu_end);