Back to home page

OSCL-LXR

 
 

    


0001 /* -*- linux-c -*- ------------------------------------------------------- *
0002  *
0003  *   Copyright 2002-2004 H. Peter Anvin - All Rights Reserved
0004  *
0005  *   This program is free software; you can redistribute it and/or modify
0006  *   it under the terms of the GNU General Public License as published by
0007  *   the Free Software Foundation, Inc., 53 Temple Place Ste 330,
0008  *   Boston MA 02111-1307, USA; either version 2 of the License, or
0009  *   (at your option) any later version; incorporated herein by reference.
0010  *
0011  * ----------------------------------------------------------------------- */
0012 
0013 /*
0014  * int$#.c
0015  *
0016  * $#-way unrolled portable integer math RAID-6 instruction set
0017  *
0018  * This file is postprocessed using unroll.awk
0019  */
0020 
0021 #include <linux/raid/pq.h>
0022 
0023 /*
0024  * This is the C data type to use
0025  */
0026 
0027 /* Change this from BITS_PER_LONG if there is something better... */
0028 #if BITS_PER_LONG == 64
0029 # define NBYTES(x) ((x) * 0x0101010101010101UL)
0030 # define NSIZE  8
0031 # define NSHIFT 3
0032 # define NSTRING "64"
0033 typedef u64 unative_t;
0034 #else
0035 # define NBYTES(x) ((x) * 0x01010101U)
0036 # define NSIZE  4
0037 # define NSHIFT 2
0038 # define NSTRING "32"
0039 typedef u32 unative_t;
0040 #endif
0041 
0042 
0043 
0044 /*
0045  * IA-64 wants insane amounts of unrolling.  On other architectures that
0046  * is just a waste of space.
0047  */
0048 #if ($# <= 8) || defined(__ia64__)
0049 
0050 
0051 /*
0052  * These sub-operations are separate inlines since they can sometimes be
0053  * specially optimized using architecture-specific hacks.
0054  */
0055 
0056 /*
0057  * The SHLBYTE() operation shifts each byte left by 1, *not*
0058  * rolling over into the next byte
0059  */
0060 static inline __attribute_const__ unative_t SHLBYTE(unative_t v)
0061 {
0062         unative_t vv;
0063 
0064         vv = (v << 1) & NBYTES(0xfe);
0065         return vv;
0066 }
0067 
0068 /*
0069  * The MASK() operation returns 0xFF in any byte for which the high
0070  * bit is 1, 0x00 for any byte for which the high bit is 0.
0071  */
0072 static inline __attribute_const__ unative_t MASK(unative_t v)
0073 {
0074         unative_t vv;
0075 
0076         vv = v & NBYTES(0x80);
0077         vv = (vv << 1) - (vv >> 7); /* Overflow on the top bit is OK */
0078         return vv;
0079 }
0080 
0081 
0082 static void raid6_int$#_gen_syndrome(int disks, size_t bytes, void **ptrs)
0083 {
0084         u8 **dptr = (u8 **)ptrs;
0085         u8 *p, *q;
0086         int d, z, z0;
0087 
0088         unative_t wd$$, wq$$, wp$$, w1$$, w2$$;
0089 
0090         z0 = disks - 3;         /* Highest data disk */
0091         p = dptr[z0+1];         /* XOR parity */
0092         q = dptr[z0+2];         /* RS syndrome */
0093 
0094         for ( d = 0 ; d < bytes ; d += NSIZE*$# ) {
0095                 wq$$ = wp$$ = *(unative_t *)&dptr[z0][d+$$*NSIZE];
0096                 for ( z = z0-1 ; z >= 0 ; z-- ) {
0097                         wd$$ = *(unative_t *)&dptr[z][d+$$*NSIZE];
0098                         wp$$ ^= wd$$;
0099                         w2$$ = MASK(wq$$);
0100                         w1$$ = SHLBYTE(wq$$);
0101                         w2$$ &= NBYTES(0x1d);
0102                         w1$$ ^= w2$$;
0103                         wq$$ = w1$$ ^ wd$$;
0104                 }
0105                 *(unative_t *)&p[d+NSIZE*$$] = wp$$;
0106                 *(unative_t *)&q[d+NSIZE*$$] = wq$$;
0107         }
0108 }
0109 
0110 static void raid6_int$#_xor_syndrome(int disks, int start, int stop,
0111                                      size_t bytes, void **ptrs)
0112 {
0113         u8 **dptr = (u8 **)ptrs;
0114         u8 *p, *q;
0115         int d, z, z0;
0116 
0117         unative_t wd$$, wq$$, wp$$, w1$$, w2$$;
0118 
0119         z0 = stop;              /* P/Q right side optimization */
0120         p = dptr[disks-2];      /* XOR parity */
0121         q = dptr[disks-1];      /* RS syndrome */
0122 
0123         for ( d = 0 ; d < bytes ; d += NSIZE*$# ) {
0124                 /* P/Q data pages */
0125                 wq$$ = wp$$ = *(unative_t *)&dptr[z0][d+$$*NSIZE];
0126                 for ( z = z0-1 ; z >= start ; z-- ) {
0127                         wd$$ = *(unative_t *)&dptr[z][d+$$*NSIZE];
0128                         wp$$ ^= wd$$;
0129                         w2$$ = MASK(wq$$);
0130                         w1$$ = SHLBYTE(wq$$);
0131                         w2$$ &= NBYTES(0x1d);
0132                         w1$$ ^= w2$$;
0133                         wq$$ = w1$$ ^ wd$$;
0134                 }
0135                 /* P/Q left side optimization */
0136                 for ( z = start-1 ; z >= 0 ; z-- ) {
0137                         w2$$ = MASK(wq$$);
0138                         w1$$ = SHLBYTE(wq$$);
0139                         w2$$ &= NBYTES(0x1d);
0140                         wq$$ = w1$$ ^ w2$$;
0141                 }
0142                 *(unative_t *)&p[d+NSIZE*$$] ^= wp$$;
0143                 *(unative_t *)&q[d+NSIZE*$$] ^= wq$$;
0144         }
0145 
0146 }
0147 
0148 const struct raid6_calls raid6_intx$# = {
0149         raid6_int$#_gen_syndrome,
0150         raid6_int$#_xor_syndrome,
0151         NULL,                   /* always valid */
0152         "int" NSTRING "x$#",
0153         0
0154 };
0155 
0156 #endif