arch/x86/lib/checksum_32.S

   1 /*
   2  * INET         An implementation of the TCP/IP protocol suite for the LINUX
   3  *              operating system.  INET is implemented using the  BSD Socket
   4  *              interface as the means of communication with the user level.
   5  *
   6  *              IP/TCP/UDP checksumming routines
   7  *
   8  * Authors:     Jorge Cwik, <jorge@laser.satlink.net>
   9  *              Arnt Gulbrandsen, <agulbra@nvg.unit.no>
  10  *              Tom May, <ftom@netcom.com>
  11  *              Pentium Pro/II routines:
  12  *              Alexander Kjeldaas <astor@guardian.no>
  13  *              Finn Arne Gangstad <finnag@guardian.no>
  14  *              Lots of code moved from tcp.c and ip.c; see those files
  15  *              for more names.
  16  *
  17  * Changes:     Ingo Molnar, converted csum_partial_copy() to 2.1 exception
  18  *                           handling.
  19  *              Andi Kleen,  add zeroing on error
  20  *                   converted to pure assembler
  21  *
  22  *              This program is free software; you can redistribute it and/or
  23  *              modify it under the terms of the GNU General Public License
  24  *              as published by the Free Software Foundation; either version
  25  *              2 of the License, or (at your option) any later version.
  26  */
  27
  28 #include <linux/linkage.h>
  29 #include <asm/errno.h>
  30 #include <asm/asm.h>
  31 #include <asm/export.h>
  32 #include <asm/nospec-branch.h>
  33
  34 /*
  35  * computes a partial checksum, e.g. for TCP/UDP fragments
  36  */
  37
  38 /*
  39 unsigned int csum_partial(const unsigned char * buff, int len, unsigned int sum)
  40  */
  41
  42 .text
  43
  44 #ifndef CONFIG_X86_USE_PPRO_CHECKSUM
  45
  46           /*
  47            * Experiments with Ethernet and SLIP connections show that buff
  48            * is aligned on either a 2-byte or 4-byte boundary.  We get at
  49            * least a twofold speedup on 486 and Pentium if it is 4-byte aligned.
  50            * Fortunately, it is easy to convert 2-byte alignment to 4-byte
  51            * alignment for the unrolled loop.
  52            */
  53 ENTRY(csum_partial)
  54         pushl %esi
  55         pushl %ebx
  56         movl 20(%esp),%eax      # Function arg: unsigned int sum
  57         movl 16(%esp),%ecx      # Function arg: int len
  58         movl 12(%esp),%esi      # Function arg: unsigned char *buff
  59         testl $3, %esi          # Check alignment.
  60         jz 2f                   # Jump if alignment is ok.
  61         testl $1, %esi          # Check alignment.
  62         jz 10f                  # Jump if alignment is boundary of 2 bytes.
  63
  64         # buf is odd
  65         dec %ecx
  66         jl 8f
  67         movzbl (%esi), %ebx
  68         adcl %ebx, %eax
  69         roll $8, %eax
  70         inc %esi
  71         testl $2, %esi
  72         jz 2f
  73 10:
  74         subl $2, %ecx           # Alignment uses up two bytes.
  75         jae 1f                  # Jump if we had at least two bytes.
  76         addl $2, %ecx           # ecx was < 2.  Deal with it.
  77         jmp 4f
  78 1:      movw (%esi), %bx
  79         addl $2, %esi
  80         addw %bx, %ax
  81         adcl $0, %eax
  82 2:
  83         movl %ecx, %edx
  84         shrl $5, %ecx
  85         jz 2f
  86         testl %esi, %esi
  87 1:      movl (%esi), %ebx
  88         adcl %ebx, %eax
  89         movl 4(%esi), %ebx
  90         adcl %ebx, %eax
  91         movl 8(%esi), %ebx
  92         adcl %ebx, %eax
  93         movl 12(%esi), %ebx
  94         adcl %ebx, %eax
  95         movl 16(%esi), %ebx
  96         adcl %ebx, %eax
  97         movl 20(%esi), %ebx
  98         adcl %ebx, %eax
  99         movl 24(%esi), %ebx
 100         adcl %ebx, %eax
 101         movl 28(%esi), %ebx
 102         adcl %ebx, %eax
 103         lea 32(%esi), %esi
 104         dec %ecx
 105         jne 1b
 106         adcl $0, %eax
 107 2:      movl %edx, %ecx
 108         andl $0x1c, %edx
 109         je 4f
 110         shrl $2, %edx           # This clears CF
 111 3:      adcl (%esi), %eax
 112         lea 4(%esi), %esi
 113         dec %edx
 114         jne 3b
 115         adcl $0, %eax
 116 4:      andl $3, %ecx
 117         jz 7f
 118         cmpl $2, %ecx
 119         jb 5f
 120         movw (%esi),%cx
 121         leal 2(%esi),%esi
 122         je 6f
 123         shll $16,%ecx
 124 5:      movb (%esi),%cl
 125 6:      addl %ecx,%eax
 126         adcl $0, %eax
 127 7:
 128         testb $1, 12(%esp)
 129         jz 8f
 130         roll $8, %eax
 131 8:
 132         popl %ebx
 133         popl %esi
 134         ret
 135 ENDPROC(csum_partial)
 136
 137 #else
 138
 139 /* Version for PentiumII/PPro */
 140
 141 ENTRY(csum_partial)
 142         pushl %esi
 143         pushl %ebx
 144         movl 20(%esp),%eax      # Function arg: unsigned int sum
 145         movl 16(%esp),%ecx      # Function arg: int len
 146         movl 12(%esp),%esi      # Function arg: const unsigned char *buf
 147
 148         testl $3, %esi
 149         jnz 25f
 150 10:
 151         movl %ecx, %edx
 152         movl %ecx, %ebx
 153         andl $0x7c, %ebx
 154         shrl $7, %ecx
 155         addl %ebx,%esi
 156         shrl $2, %ebx
 157         negl %ebx
 158         lea 45f(%ebx,%ebx,2), %ebx
 159         testl %esi, %esi
 160         JMP_NOSPEC %ebx
 161
 162         # Handle 2-byte-aligned regions
 163 20:     addw (%esi), %ax
 164         lea 2(%esi), %esi
 165         adcl $0, %eax
 166         jmp 10b
 167 25:
 168         testl $1, %esi
 169         jz 30f
 170         # buf is odd
 171         dec %ecx
 172         jl 90f
 173         movzbl (%esi), %ebx
 174         addl %ebx, %eax
 175         adcl $0, %eax
 176         roll $8, %eax
 177         inc %esi
 178         testl $2, %esi
 179         jz 10b
 180
 181 30:     subl $2, %ecx
 182         ja 20b
 183         je 32f
 184         addl $2, %ecx
 185         jz 80f
 186         movzbl (%esi),%ebx      # csumming 1 byte, 2-aligned
 187         addl %ebx, %eax
 188         adcl $0, %eax
 189         jmp 80f
 190 32:
 191         addw (%esi), %ax        # csumming 2 bytes, 2-aligned
 192         adcl $0, %eax
 193         jmp 80f
 194
 195 40:
 196         addl -128(%esi), %eax
 197         adcl -124(%esi), %eax
 198         adcl -120(%esi), %eax
 199         adcl -116(%esi), %eax
 200         adcl -112(%esi), %eax
 201         adcl -108(%esi), %eax
 202         adcl -104(%esi), %eax
 203         adcl -100(%esi), %eax
 204         adcl -96(%esi), %eax
 205         adcl -92(%esi), %eax
 206         adcl -88(%esi), %eax
 207         adcl -84(%esi), %eax
 208         adcl -80(%esi), %eax
 209         adcl -76(%esi), %eax
 210         adcl -72(%esi), %eax
 211         adcl -68(%esi), %eax
 212         adcl -64(%esi), %eax
 213         adcl -60(%esi), %eax
 214         adcl -56(%esi), %eax
 215         adcl -52(%esi), %eax
 216         adcl -48(%esi), %eax
 217         adcl -44(%esi), %eax
 218         adcl -40(%esi), %eax
 219         adcl -36(%esi), %eax
 220         adcl -32(%esi), %eax
 221         adcl -28(%esi), %eax
 222         adcl -24(%esi), %eax
 223         adcl -20(%esi), %eax
 224         adcl -16(%esi), %eax
 225         adcl -12(%esi), %eax
 226         adcl -8(%esi), %eax
 227         adcl -4(%esi), %eax
 228 45:
 229         lea 128(%esi), %esi
 230         adcl $0, %eax
 231         dec %ecx
 232         jge 40b
 233         movl %edx, %ecx
 234 50:     andl $3, %ecx
 235         jz 80f
 236
 237         # Handle the last 1-3 bytes without jumping
 238         notl %ecx               # 1->2, 2->1, 3->0, higher bits are masked
 239         movl $0xffffff,%ebx     # by the shll and shrl instructions
 240         shll $3,%ecx
 241         shrl %cl,%ebx
 242         andl -128(%esi),%ebx    # esi is 4-aligned so should be ok
 243         addl %ebx,%eax
 244         adcl $0,%eax
 245 80:
 246         testb $1, 12(%esp)
 247         jz 90f
 248         roll $8, %eax
 249 90:
 250         popl %ebx
 251         popl %esi
 252         ret
 253 ENDPROC(csum_partial)
 254
 255 #endif
 256 EXPORT_SYMBOL(csum_partial)
 257
 258 /*
 259 unsigned int csum_partial_copy_generic (const char *src, char *dst,
 260                                   int len, int sum, int *src_err_ptr, int *dst_err_ptr)
 261  */
 262
 263 /*
 264  * Copy from ds while checksumming, otherwise like csum_partial
 265  *
 266  * The macros SRC and DST specify the type of access for the instruction.
 267  * thus we can call a custom exception handler for all access types.
 268  *
 269  * FIXME: could someone double-check whether I haven't mixed up some SRC and
 270  *        DST definitions? It's damn hard to trigger all cases.  I hope I got
 271  *        them all but there's no guarantee.
 272  */
 273
 274 #define SRC(y...)                       \
 275         9999: y;                        \
 276         _ASM_EXTABLE(9999b, 6001f)
 277
 278 #define DST(y...)                       \
 279         9999: y;                        \
 280         _ASM_EXTABLE(9999b, 6002f)
 281
 282 #ifndef CONFIG_X86_USE_PPRO_CHECKSUM
 283
 284 #define ARGBASE 16
 285 #define FP              12
 286
 287 ENTRY(csum_partial_copy_generic)
 288         subl  $4,%esp
 289         pushl %edi
 290         pushl %esi
 291         pushl %ebx
 292         movl ARGBASE+16(%esp),%eax      # sum
 293         movl ARGBASE+12(%esp),%ecx      # len
 294         movl ARGBASE+4(%esp),%esi       # src
 295         movl ARGBASE+8(%esp),%edi       # dst
 296
 297         testl $2, %edi                  # Check alignment.
 298         jz 2f                           # Jump if alignment is ok.
 299         subl $2, %ecx                   # Alignment uses up two bytes.
 300         jae 1f                          # Jump if we had at least two bytes.
 301         addl $2, %ecx                   # ecx was < 2.  Deal with it.
 302         jmp 4f
 303 SRC(1:  movw (%esi), %bx        )
 304         addl $2, %esi
 305 DST(    movw %bx, (%edi)        )
 306         addl $2, %edi
 307         addw %bx, %ax
 308         adcl $0, %eax
 309 2:
 310         movl %ecx, FP(%esp)
 311         shrl $5, %ecx
 312         jz 2f
 313         testl %esi, %esi
 314 SRC(1:  movl (%esi), %ebx       )
 315 SRC(    movl 4(%esi), %edx      )
 316         adcl %ebx, %eax
 317 DST(    movl %ebx, (%edi)       )
 318         adcl %edx, %eax
 319 DST(    movl %edx, 4(%edi)      )
 320
 321 SRC(    movl 8(%esi), %ebx      )
 322 SRC(    movl 12(%esi), %edx     )
 323         adcl %ebx, %eax
 324 DST(    movl %ebx, 8(%edi)      )
 325         adcl %edx, %eax
 326 DST(    movl %edx, 12(%edi)     )
 327
 328 SRC(    movl 16(%esi), %ebx     )
 329 SRC(    movl 20(%esi), %edx     )
 330         adcl %ebx, %eax
 331 DST(    movl %ebx, 16(%edi)     )
 332         adcl %edx, %eax
 333 DST(    movl %edx, 20(%edi)     )
 334
 335 SRC(    movl 24(%esi), %ebx     )
 336 SRC(    movl 28(%esi), %edx     )
 337         adcl %ebx, %eax
 338 DST(    movl %ebx, 24(%edi)     )
 339         adcl %edx, %eax
 340 DST(    movl %edx, 28(%edi)     )
 341
 342         lea 32(%esi), %esi
 343         lea 32(%edi), %edi
 344         dec %ecx
 345         jne 1b
 346         adcl $0, %eax
 347 2:      movl FP(%esp), %edx
 348         movl %edx, %ecx
 349         andl $0x1c, %edx
 350         je 4f
 351         shrl $2, %edx                   # This clears CF
 352 SRC(3:  movl (%esi), %ebx       )
 353         adcl %ebx, %eax
 354 DST(    movl %ebx, (%edi)       )
 355         lea 4(%esi), %esi
 356         lea 4(%edi), %edi
 357         dec %edx
 358         jne 3b
 359         adcl $0, %eax
 360 4:      andl $3, %ecx
 361         jz 7f
 362         cmpl $2, %ecx
 363         jb 5f
 364 SRC(    movw (%esi), %cx        )
 365         leal 2(%esi), %esi
 366 DST(    movw %cx, (%edi)        )
 367         leal 2(%edi), %edi
 368         je 6f
 369         shll $16,%ecx
 370 SRC(5:  movb (%esi), %cl        )
 371 DST(    movb %cl, (%edi)        )
 372 6:      addl %ecx, %eax
 373         adcl $0, %eax
 374 7:
 375 5000:
 376
 377 # Exception handler:
 378 .section .fixup, "ax"
 379
 380 6001:
 381         movl ARGBASE+20(%esp), %ebx     # src_err_ptr
 382         movl $-EFAULT, (%ebx)
 383
 384         # zero the complete destination - computing the rest
 385         # is too much work
 386         movl ARGBASE+8(%esp), %edi      # dst
 387         movl ARGBASE+12(%esp), %ecx     # len
 388         xorl %eax,%eax
 389         rep ; stosb
 390
 391         jmp 5000b
 392
 393 6002:
 394         movl ARGBASE+24(%esp), %ebx     # dst_err_ptr
 395         movl $-EFAULT,(%ebx)
 396         jmp 5000b
 397
 398 .previous
 399
 400         popl %ebx
 401         popl %esi
 402         popl %edi
 403         popl %ecx                       # equivalent to addl $4,%esp
 404         ret
 405 ENDPROC(csum_partial_copy_generic)
 406
 407 #else
 408
 409 /* Version for PentiumII/PPro */
 410
 411 #define ROUND1(x) \
 412         SRC(movl x(%esi), %ebx  )       ;       \
 413         addl %ebx, %eax                 ;       \
 414         DST(movl %ebx, x(%edi)  )       ;
 415
 416 #define ROUND(x) \
 417         SRC(movl x(%esi), %ebx  )       ;       \
 418         adcl %ebx, %eax                 ;       \
 419         DST(movl %ebx, x(%edi)  )       ;
 420
 421 #define ARGBASE 12
 422
 423 ENTRY(csum_partial_copy_generic)
 424         pushl %ebx
 425         pushl %edi
 426         pushl %esi
 427         movl ARGBASE+4(%esp),%esi       #src
 428         movl ARGBASE+8(%esp),%edi       #dst
 429         movl ARGBASE+12(%esp),%ecx      #len
 430         movl ARGBASE+16(%esp),%eax      #sum
 431 #       movl %ecx, %edx
 432         movl %ecx, %ebx
 433         movl %esi, %edx
 434         shrl $6, %ecx
 435         andl $0x3c, %ebx
 436         negl %ebx
 437         subl %ebx, %esi
 438         subl %ebx, %edi
 439         lea  -1(%esi),%edx
 440         andl $-32,%edx
 441         lea 3f(%ebx,%ebx), %ebx
 442         testl %esi, %esi
 443         JMP_NOSPEC %ebx
 444 1:      addl $64,%esi
 445         addl $64,%edi
 446         SRC(movb -32(%edx),%bl) ; SRC(movb (%edx),%bl)
 447         ROUND1(-64) ROUND(-60) ROUND(-56) ROUND(-52)
 448         ROUND (-48) ROUND(-44) ROUND(-40) ROUND(-36)
 449         ROUND (-32) ROUND(-28) ROUND(-24) ROUND(-20)
 450         ROUND (-16) ROUND(-12) ROUND(-8)  ROUND(-4)
 451 3:      adcl $0,%eax
 452         addl $64, %edx
 453         dec %ecx
 454         jge 1b
 455 4:      movl ARGBASE+12(%esp),%edx      #len
 456         andl $3, %edx
 457         jz 7f
 458         cmpl $2, %edx
 459         jb 5f
 460 SRC(    movw (%esi), %dx         )
 461         leal 2(%esi), %esi
 462 DST(    movw %dx, (%edi)         )
 463         leal 2(%edi), %edi
 464         je 6f
 465         shll $16,%edx
 466 5:
 467 SRC(    movb (%esi), %dl         )
 468 DST(    movb %dl, (%edi)         )
 469 6:      addl %edx, %eax
 470         adcl $0, %eax
 471 7:
 472 .section .fixup, "ax"
 473 6001:   movl    ARGBASE+20(%esp), %ebx  # src_err_ptr
 474         movl $-EFAULT, (%ebx)
 475         # zero the complete destination (computing the rest is too much work)
 476         movl ARGBASE+8(%esp),%edi       # dst
 477         movl ARGBASE+12(%esp),%ecx      # len
 478         xorl %eax,%eax
 479         rep; stosb
 480         jmp 7b
 481 6002:   movl ARGBASE+24(%esp), %ebx     # dst_err_ptr
 482         movl $-EFAULT, (%ebx)
 483         jmp  7b
 484 .previous
 485
 486         popl %esi
 487         popl %edi
 488         popl %ebx
 489         ret
 490 ENDPROC(csum_partial_copy_generic)
 491
 492 #undef ROUND
 493 #undef ROUND1
 494
 495 #endif
 496 EXPORT_SYMBOL(csum_partial_copy_generic)