arch/x86/lib/csum-partial_64.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * arch/x86_64/lib/csum-partial.c
   4  *
   5  * This file contains network checksum routines that are better done
   6  * in an architecture-specific manner due to speed.
   7  */
   8
   9 #include <linux/compiler.h>
  10 #include <linux/export.h>
  11 #include <asm/checksum.h>
  12 #include <asm/word-at-a-time.h>
  13
  14 static inline unsigned short from32to16(unsigned a)
  15 {
  16         unsigned short b = a >> 16;
  17         asm("addw %w2,%w0\n\t"
  18             "adcw $0,%w0\n"
  19             : "=r" (b)
  20             : "0" (b), "r" (a));
  21         return b;
  22 }
  23
  24 static inline __wsum csum_tail(u64 temp64, int odd)
  25 {
  26         unsigned int result;
  27
  28         result = add32_with_carry(temp64 >> 32, temp64 & 0xffffffff);
  29         if (unlikely(odd)) {
  30                 result = from32to16(result);
  31                 result = ((result >> 8) & 0xff) | ((result & 0xff) << 8);
  32         }
  33         return (__force __wsum)result;
  34 }
  35
  36 /*
  37  * Do a checksum on an arbitrary memory area.
  38  * Returns a 32bit checksum.
  39  *
  40  * This isn't as time critical as it used to be because many NICs
  41  * do hardware checksumming these days.
  42  *
  43  * Still, with CHECKSUM_COMPLETE this is called to compute
  44  * checksums on IPv6 headers (40 bytes) and other small parts.
  45  * it's best to have buff aligned on a 64-bit boundary
  46  */
  47 __wsum csum_partial(const void *buff, int len, __wsum sum)
  48 {
  49         u64 temp64 = (__force u64)sum;
  50         unsigned odd;
  51
  52         odd = 1 & (unsigned long) buff;
  53         if (unlikely(odd)) {
  54                 if (unlikely(len == 0))
  55                         return sum;
  56                 temp64 = ror32((__force u32)sum, 8);
  57                 temp64 += (*(unsigned char *)buff << 8);
  58                 len--;
  59                 buff++;
  60         }
  61
  62         /*
  63          * len == 40 is the hot case due to IPv6 headers, but annotating it likely()
  64          * has noticeable negative affect on codegen for all other cases with
  65          * minimal performance benefit here.
  66          */
  67         if (len == 40) {
  68                 asm("addq 0*8(%[src]),%[res]\n\t"
  69                     "adcq 1*8(%[src]),%[res]\n\t"
  70                     "adcq 2*8(%[src]),%[res]\n\t"
  71                     "adcq 3*8(%[src]),%[res]\n\t"
  72                     "adcq 4*8(%[src]),%[res]\n\t"
  73                     "adcq $0,%[res]"
  74                     : [res] "+r"(temp64)
  75                     : [src] "r"(buff), "m"(*(const char(*)[40])buff));
  76                 return csum_tail(temp64, odd);
  77         }
  78         if (unlikely(len >= 64)) {
  79                 /*
  80                  * Extra accumulators for better ILP in the loop.
  81                  */
  82                 u64 tmp_accum, tmp_carries;
  83
  84                 asm("xorl %k[tmp_accum],%k[tmp_accum]\n\t"
  85                     "xorl %k[tmp_carries],%k[tmp_carries]\n\t"
  86                     "subl $64, %[len]\n\t"
  87                     "1:\n\t"
  88                     "addq 0*8(%[src]),%[res]\n\t"
  89                     "adcq 1*8(%[src]),%[res]\n\t"
  90                     "adcq 2*8(%[src]),%[res]\n\t"
  91                     "adcq 3*8(%[src]),%[res]\n\t"
  92                     "adcl $0,%k[tmp_carries]\n\t"
  93                     "addq 4*8(%[src]),%[tmp_accum]\n\t"
  94                     "adcq 5*8(%[src]),%[tmp_accum]\n\t"
  95                     "adcq 6*8(%[src]),%[tmp_accum]\n\t"
  96                     "adcq 7*8(%[src]),%[tmp_accum]\n\t"
  97                     "adcl $0,%k[tmp_carries]\n\t"
  98                     "addq $64, %[src]\n\t"
  99                     "subl $64, %[len]\n\t"
 100                     "jge 1b\n\t"
 101                     "addq %[tmp_accum],%[res]\n\t"
 102                     "adcq %[tmp_carries],%[res]\n\t"
 103                     "adcq $0,%[res]"
 104                     : [tmp_accum] "=&r"(tmp_accum),
 105                       [tmp_carries] "=&r"(tmp_carries), [res] "+r"(temp64),
 106                       [len] "+r"(len), [src] "+r"(buff)
 107                     : "m"(*(const char *)buff));
 108         }
 109
 110         if (len & 32) {
 111                 asm("addq 0*8(%[src]),%[res]\n\t"
 112                     "adcq 1*8(%[src]),%[res]\n\t"
 113                     "adcq 2*8(%[src]),%[res]\n\t"
 114                     "adcq 3*8(%[src]),%[res]\n\t"
 115                     "adcq $0,%[res]"
 116                     : [res] "+r"(temp64)
 117                     : [src] "r"(buff), "m"(*(const char(*)[32])buff));
 118                 buff += 32;
 119         }
 120         if (len & 16) {
 121                 asm("addq 0*8(%[src]),%[res]\n\t"
 122                     "adcq 1*8(%[src]),%[res]\n\t"
 123                     "adcq $0,%[res]"
 124                     : [res] "+r"(temp64)
 125                     : [src] "r"(buff), "m"(*(const char(*)[16])buff));
 126                 buff += 16;
 127         }
 128         if (len & 8) {
 129                 asm("addq 0*8(%[src]),%[res]\n\t"
 130                     "adcq $0,%[res]"
 131                     : [res] "+r"(temp64)
 132                     : [src] "r"(buff), "m"(*(const char(*)[8])buff));
 133                 buff += 8;
 134         }
 135         if (len & 7) {
 136                 unsigned int shift = (-len << 3) & 63;
 137                 unsigned long trail;
 138
 139                 trail = (load_unaligned_zeropad(buff) << shift) >> shift;
 140
 141                 asm("addq %[trail],%[res]\n\t"
 142                     "adcq $0,%[res]"
 143                     : [res] "+r"(temp64)
 144                     : [trail] "r"(trail));
 145         }
 146         return csum_tail(temp64, odd);
 147 }
 148 EXPORT_SYMBOL(csum_partial);
 149
 150 /*
 151  * this routine is used for miscellaneous IP-like checksums, mainly
 152  * in icmp.c
 153  */
 154 __sum16 ip_compute_csum(const void *buff, int len)
 155 {
 156         return csum_fold(csum_partial(buff, len, 0));
 157 }
 158 EXPORT_SYMBOL(ip_compute_csum);