lib/raid6/s390vx.uc

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * raid6_vx$#.c
   4  *
   5  * $#-way unrolled RAID6 gen/xor functions for s390
   6  * based on the vector facility
   7  *
   8  * Copyright IBM Corp. 2016
   9  * Author(s): Martin Schwidefsky <schwidefsky@de.ibm.com>
  10  *
  11  * This file is postprocessed using unroll.awk.
  12  */
  13
  14 #include <linux/raid/pq.h>
  15 #include <asm/fpu/api.h>
  16
  17 asm(".include \"asm/vx-insn.h\"\n");
  18
  19 #define NSIZE 16
  20
  21 static inline void LOAD_CONST(void)
  22 {
  23         asm volatile("VREPIB %v24,7");
  24         asm volatile("VREPIB %v25,0x1d");
  25 }
  26
  27 /*
  28  * The SHLBYTE() operation shifts each of the 16 bytes in
  29  * vector register y left by 1 bit and stores the result in
  30  * vector register x.
  31  */
  32 static inline void SHLBYTE(int x, int y)
  33 {
  34         asm volatile ("VAB %0,%1,%1" : : "i" (x), "i" (y));
  35 }
  36
  37 /*
  38  * For each of the 16 bytes in the vector register y the MASK()
  39  * operation returns 0xFF if the high bit of the byte is 1,
  40  * or 0x00 if the high bit is 0. The result is stored in vector
  41  * register x.
  42  */
  43 static inline void MASK(int x, int y)
  44 {
  45         asm volatile ("VESRAVB  %0,%1,24" : : "i" (x), "i" (y));
  46 }
  47
  48 static inline void AND(int x, int y, int z)
  49 {
  50         asm volatile ("VN %0,%1,%2" : : "i" (x), "i" (y), "i" (z));
  51 }
  52
  53 static inline void XOR(int x, int y, int z)
  54 {
  55         asm volatile ("VX %0,%1,%2" : : "i" (x), "i" (y), "i" (z));
  56 }
  57
  58 static inline void LOAD_DATA(int x, int n, u8 *ptr)
  59 {
  60         typedef struct { u8 _[16*n]; } addrtype;
  61         register addrtype *__ptr asm("1") = (addrtype *) ptr;
  62
  63         asm volatile ("VLM %2,%3,0,%r1"
  64                       : : "m" (*__ptr), "a" (__ptr), "i" (x), "i" (x + n - 1));
  65 }
  66
  67 static inline void STORE_DATA(int x, int n, u8 *ptr)
  68 {
  69         typedef struct { u8 _[16*n]; } addrtype;
  70         register addrtype *__ptr asm("1") = (addrtype *) ptr;
  71
  72         asm volatile ("VSTM %2,%3,0,1"
  73                       : "=m" (*__ptr) : "a" (__ptr), "i" (x), "i" (x + n - 1));
  74 }
  75
  76 static inline void COPY_VEC(int x, int y)
  77 {
  78         asm volatile ("VLR %0,%1" : : "i" (x), "i" (y));
  79 }
  80
  81 static void raid6_s390vx$#_gen_syndrome(int disks, size_t bytes, void **ptrs)
  82 {
  83         struct kernel_fpu vxstate;
  84         u8 **dptr, *p, *q;
  85         int d, z, z0;
  86
  87         kernel_fpu_begin(&vxstate, KERNEL_VXR);
  88         LOAD_CONST();
  89
  90         dptr = (u8 **) ptrs;
  91         z0 = disks - 3;         /* Highest data disk */
  92         p = dptr[z0 + 1];       /* XOR parity */
  93         q = dptr[z0 + 2];       /* RS syndrome */
  94
  95         for (d = 0; d < bytes; d += $#*NSIZE) {
  96                 LOAD_DATA(0,$#,&dptr[z0][d]);
  97                 COPY_VEC(8+$$,0+$$);
  98                 for (z = z0 - 1; z >= 0; z--) {
  99                         MASK(16+$$,8+$$);
 100                         AND(16+$$,16+$$,25);
 101                         SHLBYTE(8+$$,8+$$);
 102                         XOR(8+$$,8+$$,16+$$);
 103                         LOAD_DATA(16,$#,&dptr[z][d]);
 104                         XOR(0+$$,0+$$,16+$$);
 105                         XOR(8+$$,8+$$,16+$$);
 106                 }
 107                 STORE_DATA(0,$#,&p[d]);
 108                 STORE_DATA(8,$#,&q[d]);
 109         }
 110         kernel_fpu_end(&vxstate, KERNEL_VXR);
 111 }
 112
 113 static void raid6_s390vx$#_xor_syndrome(int disks, int start, int stop,
 114                                         size_t bytes, void **ptrs)
 115 {
 116         struct kernel_fpu vxstate;
 117         u8 **dptr, *p, *q;
 118         int d, z, z0;
 119
 120         dptr = (u8 **) ptrs;
 121         z0 = stop;              /* P/Q right side optimization */
 122         p = dptr[disks - 2];    /* XOR parity */
 123         q = dptr[disks - 1];    /* RS syndrome */
 124
 125         kernel_fpu_begin(&vxstate, KERNEL_VXR);
 126         LOAD_CONST();
 127
 128         for (d = 0; d < bytes; d += $#*NSIZE) {
 129                 /* P/Q data pages */
 130                 LOAD_DATA(0,$#,&dptr[z0][d]);
 131                 COPY_VEC(8+$$,0+$$);
 132                 for (z = z0 - 1; z >= start; z--) {
 133                         MASK(16+$$,8+$$);
 134                         AND(16+$$,16+$$,25);
 135                         SHLBYTE(8+$$,8+$$);
 136                         XOR(8+$$,8+$$,16+$$);
 137                         LOAD_DATA(16,$#,&dptr[z][d]);
 138                         XOR(0+$$,0+$$,16+$$);
 139                         XOR(8+$$,8+$$,16+$$);
 140                 }
 141                 /* P/Q left side optimization */
 142                 for (z = start - 1; z >= 0; z--) {
 143                         MASK(16+$$,8+$$);
 144                         AND(16+$$,16+$$,25);
 145                         SHLBYTE(8+$$,8+$$);
 146                         XOR(8+$$,8+$$,16+$$);
 147                 }
 148                 LOAD_DATA(16,$#,&p[d]);
 149                 XOR(16+$$,16+$$,0+$$);
 150                 STORE_DATA(16,$#,&p[d]);
 151                 LOAD_DATA(16,$#,&q[d]);
 152                 XOR(16+$$,16+$$,8+$$);
 153                 STORE_DATA(16,$#,&q[d]);
 154         }
 155         kernel_fpu_end(&vxstate, KERNEL_VXR);
 156 }
 157
 158 static int raid6_s390vx$#_valid(void)
 159 {
 160         return MACHINE_HAS_VX;
 161 }
 162
 163 const struct raid6_calls raid6_s390vx$# = {
 164         raid6_s390vx$#_gen_syndrome,
 165         raid6_s390vx$#_xor_syndrome,
 166         raid6_s390vx$#_valid,
 167         "vx128x$#",
 168         1
 169 };