From 04864808029e59ea1bf075c756a0f35c8398fc11 Mon Sep 17 00:00:00 2001 From: Hendrik Brueckner Date: Wed, 18 Feb 2015 14:46:00 +0100 Subject: [PATCH] s390/vx: add support functions for in-kernel FPU use Introduce the kernel_fpu_begin() and kernel_fpu_end() function to enclose any in-kernel use of FPU instructions and registers. In enclosed sections, you can perform floating-point or vector (SIMD) computations. The functions take care of saving and restoring FPU register contents and controls. For usage details, see the guidelines in arch/s390/include/asm/fpu/api.h Signed-off-by: Hendrik Brueckner Signed-off-by: Martin Schwidefsky --- arch/s390/include/asm/fpu/api.h | 75 +++++++++ arch/s390/include/asm/fpu/types.h | 10 ++ arch/s390/kernel/Makefile | 2 +- arch/s390/kernel/fpu.c | 249 ++++++++++++++++++++++++++++++ 4 files changed, 335 insertions(+), 1 deletion(-) create mode 100644 arch/s390/kernel/fpu.c diff --git a/arch/s390/include/asm/fpu/api.h b/arch/s390/include/asm/fpu/api.h index 5e04f3cbd320..78ba3ddb9e18 100644 --- a/arch/s390/include/asm/fpu/api.h +++ b/arch/s390/include/asm/fpu/api.h @@ -1,6 +1,41 @@ /* * In-kernel FPU support functions * + * + * Consider these guidelines before using in-kernel FPU functions: + * + * 1. Use kernel_fpu_begin() and kernel_fpu_end() to enclose all in-kernel + * use of floating-point or vector registers and instructions. + * + * 2. For kernel_fpu_begin(), specify the vector register range you want to + * use with the KERNEL_VXR_* constants. Consider these usage guidelines: + * + * a) If your function typically runs in process-context, use the lower + * half of the vector registers, for example, specify KERNEL_VXR_LOW. + * b) If your function typically runs in soft-irq or hard-irq context, + * prefer using the upper half of the vector registers, for example, + * specify KERNEL_VXR_HIGH. + * + * If you adhere to these guidelines, an interrupted process context + * does not require to save and restore vector registers because of + * disjoint register ranges. + * + * Also note that the __kernel_fpu_begin()/__kernel_fpu_end() functions + * includes logic to save and restore up to 16 vector registers at once. + * + * 3. You can nest kernel_fpu_begin()/kernel_fpu_end() by using different + * struct kernel_fpu states. Vector registers that are in use by outer + * levels are saved and restored. You can minimize the save and restore + * effort by choosing disjoint vector register ranges. + * + * 5. To use vector floating-point instructions, specify the KERNEL_FPC + * flag to save and restore floating-point controls in addition to any + * vector register range. + * + * 6. To use floating-point registers and instructions only, specify the + * KERNEL_FPR flag. This flag triggers a save and restore of vector + * registers V0 to V15 and floating-point controls. + * * Copyright IBM Corp. 2015 * Author(s): Hendrik Brueckner */ @@ -8,6 +43,8 @@ #ifndef _ASM_S390_FPU_API_H #define _ASM_S390_FPU_API_H +#include + void save_fpu_regs(void); static inline int test_fp_ctl(u32 fpc) @@ -27,4 +64,42 @@ static inline int test_fp_ctl(u32 fpc) return rc; } +#define KERNEL_VXR_V0V7 1 +#define KERNEL_VXR_V8V15 2 +#define KERNEL_VXR_V16V23 4 +#define KERNEL_VXR_V24V31 8 +#define KERNEL_FPR 16 +#define KERNEL_FPC 256 + +#define KERNEL_VXR_LOW (KERNEL_VXR_V0V7|KERNEL_VXR_V8V15) +#define KERNEL_VXR_MID (KERNEL_VXR_V8V15|KERNEL_VXR_V16V23) +#define KERNEL_VXR_HIGH (KERNEL_VXR_V16V23|KERNEL_VXR_V24V31) + +#define KERNEL_FPU_MASK (KERNEL_VXR_LOW|KERNEL_VXR_HIGH|KERNEL_FPR) + +struct kernel_fpu; + +/* + * Note the functions below must be called with preemption disabled. + * Do not enable preemption before calling __kernel_fpu_end() to prevent + * an corruption of an existing kernel FPU state. + * + * Prefer using the kernel_fpu_begin()/kernel_fpu_end() pair of functions. + */ +void __kernel_fpu_begin(struct kernel_fpu *state, u32 flags); +void __kernel_fpu_end(struct kernel_fpu *state); + + +static inline void kernel_fpu_begin(struct kernel_fpu *state, u32 flags) +{ + preempt_disable(); + __kernel_fpu_begin(state, flags); +} + +static inline void kernel_fpu_end(struct kernel_fpu *state) +{ + __kernel_fpu_end(state); + preempt_enable(); +} + #endif /* _ASM_S390_FPU_API_H */ diff --git a/arch/s390/include/asm/fpu/types.h b/arch/s390/include/asm/fpu/types.h index fe937c9b6471..bce255ead72b 100644 --- a/arch/s390/include/asm/fpu/types.h +++ b/arch/s390/include/asm/fpu/types.h @@ -24,4 +24,14 @@ struct fpu { /* VX array structure for address operand constraints in inline assemblies */ struct vx_array { __vector128 _[__NUM_VXRS]; }; +/* In-kernel FPU state structure */ +struct kernel_fpu { + u32 mask; + u32 fpc; + union { + freg_t fprs[__NUM_FPRS]; + __vector128 vxrs[__NUM_VXRS]; + }; +}; + #endif /* _ASM_S390_FPU_TYPES_H */ diff --git a/arch/s390/kernel/Makefile b/arch/s390/kernel/Makefile index 2f5586ab8a6a..8d1419120bb7 100644 --- a/arch/s390/kernel/Makefile +++ b/arch/s390/kernel/Makefile @@ -45,7 +45,7 @@ obj-y := traps.o time.o process.o base.o early.o setup.o idle.o vtime.o obj-y += processor.o sys_s390.o ptrace.o signal.o cpcmd.o ebcdic.o nmi.o obj-y += debug.o irq.o ipl.o dis.o diag.o sclp.o vdso.o obj-y += sysinfo.o jump_label.o lgr.o os_info.o machine_kexec.o pgm_check.o -obj-y += runtime_instr.o cache.o dumpstack.o +obj-y += runtime_instr.o cache.o fpu.o dumpstack.o obj-y += entry.o reipl.o relocate_kernel.o extra-y += head.o head64.o vmlinux.lds diff --git a/arch/s390/kernel/fpu.c b/arch/s390/kernel/fpu.c new file mode 100644 index 000000000000..81d1d1887507 --- /dev/null +++ b/arch/s390/kernel/fpu.c @@ -0,0 +1,249 @@ +/* + * In-kernel vector facility support functions + * + * Copyright IBM Corp. 2015 + * Author(s): Hendrik Brueckner + */ +#include +#include +#include +#include +#include + +/* + * 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) + + +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. + */ + if (!(state->mask & KERNEL_FPU_STATE_MASK)) + return; + + /* + * 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])); + 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 */ + + /* + * 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. + */ + " tmll %[m],6\n" /* KERNEL_VXR_MID */ + " jo 17f\n" /* -> save V8..V23 */ + + /* Test and save 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 -> save V0..V15 */ + " brc 4,3f\n" /* 01 -> save V0..V7 */ + " brc 2,4f\n" /* 10 -> save V8..V15 */ + + /* 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:" + : [vxrs] "=Q" (*(struct vx_array *) &state->vxrs) + : [m] "d" (state->mask) + : "1", "cc"); +} +EXPORT_SYMBOL(__kernel_fpu_begin); + +void __kernel_fpu_end(struct kernel_fpu *state) +{ + /* 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. + */ + 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; + } + + /* Test and restore floating-point controls */ + if (state->mask & KERNEL_FPC) + asm volatile("lfpc %0" : : "Q" (state->fpc)); + + /* Test and restore (load) vector registers */ + asm volatile ( + /* + * Test if any vector registers 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 */ + + /* + * Below are the vstm combinations to load multiple vector + * registers at once. + */ + "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) + : "1", "cc"); + +update_fpu_state: + /* Update current kernel VX state */ + __this_cpu_write(kernel_fpu_state, state->mask); +} +EXPORT_SYMBOL(__kernel_fpu_end); -- 2.39.5