copyuser_64.S (11151B)
1/* SPDX-License-Identifier: GPL-2.0-or-later */ 2/* 3 * Copyright (C) 2002 Paul Mackerras, IBM Corp. 4 */ 5#include <asm/processor.h> 6#include <asm/ppc_asm.h> 7#include <asm/export.h> 8#include <asm/asm-compat.h> 9#include <asm/feature-fixups.h> 10 11#ifndef SELFTEST_CASE 12/* 0 == most CPUs, 1 == POWER6, 2 == Cell */ 13#define SELFTEST_CASE 0 14#endif 15 16#ifdef __BIG_ENDIAN__ 17#define sLd sld /* Shift towards low-numbered address. */ 18#define sHd srd /* Shift towards high-numbered address. */ 19#else 20#define sLd srd /* Shift towards low-numbered address. */ 21#define sHd sld /* Shift towards high-numbered address. */ 22#endif 23 24/* 25 * These macros are used to generate exception table entries. 26 * The exception handlers below use the original arguments 27 * (stored on the stack) and the point where we're up to in 28 * the destination buffer, i.e. the address of the first 29 * unmodified byte. Generally r3 points into the destination 30 * buffer, but the first unmodified byte is at a variable 31 * offset from r3. In the code below, the symbol r3_offset 32 * is set to indicate the current offset at each point in 33 * the code. This offset is then used as a negative offset 34 * from the exception handler code, and those instructions 35 * before the exception handlers are addi instructions that 36 * adjust r3 to point to the correct place. 37 */ 38 .macro lex /* exception handler for load */ 39100: EX_TABLE(100b, .Lld_exc - r3_offset) 40 .endm 41 42 .macro stex /* exception handler for store */ 43100: EX_TABLE(100b, .Lst_exc - r3_offset) 44 .endm 45 46 .align 7 47_GLOBAL_TOC(__copy_tofrom_user) 48#ifdef CONFIG_PPC_BOOK3S_64 49BEGIN_FTR_SECTION 50 nop 51FTR_SECTION_ELSE 52 b __copy_tofrom_user_power7 53ALT_FTR_SECTION_END_IFCLR(CPU_FTR_VMX_COPY) 54#endif 55_GLOBAL(__copy_tofrom_user_base) 56 /* first check for a 4kB copy on a 4kB boundary */ 57 cmpldi cr1,r5,16 58 cmpdi cr6,r5,4096 59 or r0,r3,r4 60 neg r6,r3 /* LS 3 bits = # bytes to 8-byte dest bdry */ 61 andi. r0,r0,4095 62 std r3,-24(r1) 63 crand cr0*4+2,cr0*4+2,cr6*4+2 64 std r4,-16(r1) 65 std r5,-8(r1) 66 dcbt 0,r4 67 beq .Lcopy_page_4K 68 andi. r6,r6,7 69 PPC_MTOCRF(0x01,r5) 70 blt cr1,.Lshort_copy 71/* Below we want to nop out the bne if we're on a CPU that has the 72 * CPU_FTR_UNALIGNED_LD_STD bit set and the CPU_FTR_CP_USE_DCBTZ bit 73 * cleared. 74 * At the time of writing the only CPU that has this combination of bits 75 * set is Power6. 76 */ 77test_feature = (SELFTEST_CASE == 1) 78BEGIN_FTR_SECTION 79 nop 80FTR_SECTION_ELSE 81 bne .Ldst_unaligned 82ALT_FTR_SECTION_END(CPU_FTR_UNALIGNED_LD_STD | CPU_FTR_CP_USE_DCBTZ, \ 83 CPU_FTR_UNALIGNED_LD_STD) 84.Ldst_aligned: 85 addi r3,r3,-16 86r3_offset = 16 87test_feature = (SELFTEST_CASE == 0) 88BEGIN_FTR_SECTION 89 andi. r0,r4,7 90 bne .Lsrc_unaligned 91END_FTR_SECTION_IFCLR(CPU_FTR_UNALIGNED_LD_STD) 92 blt cr1,.Ldo_tail /* if < 16 bytes to copy */ 93 srdi r0,r5,5 94 cmpdi cr1,r0,0 95lex; ld r7,0(r4) 96lex; ld r6,8(r4) 97 addi r4,r4,16 98 mtctr r0 99 andi. r0,r5,0x10 100 beq 22f 101 addi r3,r3,16 102r3_offset = 0 103 addi r4,r4,-16 104 mr r9,r7 105 mr r8,r6 106 beq cr1,72f 10721: 108lex; ld r7,16(r4) 109lex; ld r6,24(r4) 110 addi r4,r4,32 111stex; std r9,0(r3) 112r3_offset = 8 113stex; std r8,8(r3) 114r3_offset = 16 11522: 116lex; ld r9,0(r4) 117lex; ld r8,8(r4) 118stex; std r7,16(r3) 119r3_offset = 24 120stex; std r6,24(r3) 121 addi r3,r3,32 122r3_offset = 0 123 bdnz 21b 12472: 125stex; std r9,0(r3) 126r3_offset = 8 127stex; std r8,8(r3) 128r3_offset = 16 129 andi. r5,r5,0xf 130 beq+ 3f 131 addi r4,r4,16 132.Ldo_tail: 133 addi r3,r3,16 134r3_offset = 0 135 bf cr7*4+0,246f 136lex; ld r9,0(r4) 137 addi r4,r4,8 138stex; std r9,0(r3) 139 addi r3,r3,8 140246: bf cr7*4+1,1f 141lex; lwz r9,0(r4) 142 addi r4,r4,4 143stex; stw r9,0(r3) 144 addi r3,r3,4 1451: bf cr7*4+2,2f 146lex; lhz r9,0(r4) 147 addi r4,r4,2 148stex; sth r9,0(r3) 149 addi r3,r3,2 1502: bf cr7*4+3,3f 151lex; lbz r9,0(r4) 152stex; stb r9,0(r3) 1533: li r3,0 154 blr 155 156.Lsrc_unaligned: 157r3_offset = 16 158 srdi r6,r5,3 159 addi r5,r5,-16 160 subf r4,r0,r4 161 srdi r7,r5,4 162 sldi r10,r0,3 163 cmpldi cr6,r6,3 164 andi. r5,r5,7 165 mtctr r7 166 subfic r11,r10,64 167 add r5,r5,r0 168 bt cr7*4+0,28f 169 170lex; ld r9,0(r4) /* 3+2n loads, 2+2n stores */ 171lex; ld r0,8(r4) 172 sLd r6,r9,r10 173lex; ldu r9,16(r4) 174 sHd r7,r0,r11 175 sLd r8,r0,r10 176 or r7,r7,r6 177 blt cr6,79f 178lex; ld r0,8(r4) 179 b 2f 180 18128: 182lex; ld r0,0(r4) /* 4+2n loads, 3+2n stores */ 183lex; ldu r9,8(r4) 184 sLd r8,r0,r10 185 addi r3,r3,-8 186r3_offset = 24 187 blt cr6,5f 188lex; ld r0,8(r4) 189 sHd r12,r9,r11 190 sLd r6,r9,r10 191lex; ldu r9,16(r4) 192 or r12,r8,r12 193 sHd r7,r0,r11 194 sLd r8,r0,r10 195 addi r3,r3,16 196r3_offset = 8 197 beq cr6,78f 198 1991: or r7,r7,r6 200lex; ld r0,8(r4) 201stex; std r12,8(r3) 202r3_offset = 16 2032: sHd r12,r9,r11 204 sLd r6,r9,r10 205lex; ldu r9,16(r4) 206 or r12,r8,r12 207stex; stdu r7,16(r3) 208r3_offset = 8 209 sHd r7,r0,r11 210 sLd r8,r0,r10 211 bdnz 1b 212 21378: 214stex; std r12,8(r3) 215r3_offset = 16 216 or r7,r7,r6 21779: 218stex; std r7,16(r3) 219r3_offset = 24 2205: sHd r12,r9,r11 221 or r12,r8,r12 222stex; std r12,24(r3) 223r3_offset = 32 224 bne 6f 225 li r3,0 226 blr 2276: cmpwi cr1,r5,8 228 addi r3,r3,32 229r3_offset = 0 230 sLd r9,r9,r10 231 ble cr1,7f 232lex; ld r0,8(r4) 233 sHd r7,r0,r11 234 or r9,r7,r9 2357: 236 bf cr7*4+1,1f 237#ifdef __BIG_ENDIAN__ 238 rotldi r9,r9,32 239#endif 240stex; stw r9,0(r3) 241#ifdef __LITTLE_ENDIAN__ 242 rotrdi r9,r9,32 243#endif 244 addi r3,r3,4 2451: bf cr7*4+2,2f 246#ifdef __BIG_ENDIAN__ 247 rotldi r9,r9,16 248#endif 249stex; sth r9,0(r3) 250#ifdef __LITTLE_ENDIAN__ 251 rotrdi r9,r9,16 252#endif 253 addi r3,r3,2 2542: bf cr7*4+3,3f 255#ifdef __BIG_ENDIAN__ 256 rotldi r9,r9,8 257#endif 258stex; stb r9,0(r3) 259#ifdef __LITTLE_ENDIAN__ 260 rotrdi r9,r9,8 261#endif 2623: li r3,0 263 blr 264 265.Ldst_unaligned: 266r3_offset = 0 267 PPC_MTOCRF(0x01,r6) /* put #bytes to 8B bdry into cr7 */ 268 subf r5,r6,r5 269 li r7,0 270 cmpldi cr1,r5,16 271 bf cr7*4+3,1f 272100: EX_TABLE(100b, .Lld_exc_r7) 273 lbz r0,0(r4) 274100: EX_TABLE(100b, .Lst_exc_r7) 275 stb r0,0(r3) 276 addi r7,r7,1 2771: bf cr7*4+2,2f 278100: EX_TABLE(100b, .Lld_exc_r7) 279 lhzx r0,r7,r4 280100: EX_TABLE(100b, .Lst_exc_r7) 281 sthx r0,r7,r3 282 addi r7,r7,2 2832: bf cr7*4+1,3f 284100: EX_TABLE(100b, .Lld_exc_r7) 285 lwzx r0,r7,r4 286100: EX_TABLE(100b, .Lst_exc_r7) 287 stwx r0,r7,r3 2883: PPC_MTOCRF(0x01,r5) 289 add r4,r6,r4 290 add r3,r6,r3 291 b .Ldst_aligned 292 293.Lshort_copy: 294r3_offset = 0 295 bf cr7*4+0,1f 296lex; lwz r0,0(r4) 297lex; lwz r9,4(r4) 298 addi r4,r4,8 299stex; stw r0,0(r3) 300stex; stw r9,4(r3) 301 addi r3,r3,8 3021: bf cr7*4+1,2f 303lex; lwz r0,0(r4) 304 addi r4,r4,4 305stex; stw r0,0(r3) 306 addi r3,r3,4 3072: bf cr7*4+2,3f 308lex; lhz r0,0(r4) 309 addi r4,r4,2 310stex; sth r0,0(r3) 311 addi r3,r3,2 3123: bf cr7*4+3,4f 313lex; lbz r0,0(r4) 314stex; stb r0,0(r3) 3154: li r3,0 316 blr 317 318/* 319 * exception handlers follow 320 * we have to return the number of bytes not copied 321 * for an exception on a load, we set the rest of the destination to 0 322 * Note that the number of bytes of instructions for adjusting r3 needs 323 * to equal the amount of the adjustment, due to the trick of using 324 * .Lld_exc - r3_offset as the handler address. 325 */ 326 327.Lld_exc_r7: 328 add r3,r3,r7 329 b .Lld_exc 330 331 /* adjust by 24 */ 332 addi r3,r3,8 333 nop 334 /* adjust by 16 */ 335 addi r3,r3,8 336 nop 337 /* adjust by 8 */ 338 addi r3,r3,8 339 nop 340 341/* 342 * Here we have had a fault on a load and r3 points to the first 343 * unmodified byte of the destination. We use the original arguments 344 * and r3 to work out how much wasn't copied. Since we load some 345 * distance ahead of the stores, we continue copying byte-by-byte until 346 * we hit the load fault again in order to copy as much as possible. 347 */ 348.Lld_exc: 349 ld r6,-24(r1) 350 ld r4,-16(r1) 351 ld r5,-8(r1) 352 subf r6,r6,r3 353 add r4,r4,r6 354 subf r5,r6,r5 /* #bytes left to go */ 355 356/* 357 * first see if we can copy any more bytes before hitting another exception 358 */ 359 mtctr r5 360r3_offset = 0 361100: EX_TABLE(100b, .Ldone) 36243: lbz r0,0(r4) 363 addi r4,r4,1 364stex; stb r0,0(r3) 365 addi r3,r3,1 366 bdnz 43b 367 li r3,0 /* huh? all copied successfully this time? */ 368 blr 369 370/* 371 * here we have trapped again, amount remaining is in ctr. 372 */ 373.Ldone: 374 mfctr r3 375 blr 376 377/* 378 * exception handlers for stores: we need to work out how many bytes 379 * weren't copied, and we may need to copy some more. 380 * Note that the number of bytes of instructions for adjusting r3 needs 381 * to equal the amount of the adjustment, due to the trick of using 382 * .Lst_exc - r3_offset as the handler address. 383 */ 384.Lst_exc_r7: 385 add r3,r3,r7 386 b .Lst_exc 387 388 /* adjust by 24 */ 389 addi r3,r3,8 390 nop 391 /* adjust by 16 */ 392 addi r3,r3,8 393 nop 394 /* adjust by 8 */ 395 addi r3,r3,4 396 /* adjust by 4 */ 397 addi r3,r3,4 398.Lst_exc: 399 ld r6,-24(r1) /* original destination pointer */ 400 ld r4,-16(r1) /* original source pointer */ 401 ld r5,-8(r1) /* original number of bytes */ 402 add r7,r6,r5 403 /* 404 * If the destination pointer isn't 8-byte aligned, 405 * we may have got the exception as a result of a 406 * store that overlapped a page boundary, so we may be 407 * able to copy a few more bytes. 408 */ 40917: andi. r0,r3,7 410 beq 19f 411 subf r8,r6,r3 /* #bytes copied */ 412100: EX_TABLE(100b,19f) 413 lbzx r0,r8,r4 414100: EX_TABLE(100b,19f) 415 stb r0,0(r3) 416 addi r3,r3,1 417 cmpld r3,r7 418 blt 17b 41919: subf r3,r3,r7 /* #bytes not copied in r3 */ 420 blr 421 422/* 423 * Routine to copy a whole page of data, optimized for POWER4. 424 * On POWER4 it is more than 50% faster than the simple loop 425 * above (following the .Ldst_aligned label). 426 */ 427 .macro exc 428100: EX_TABLE(100b, .Labort) 429 .endm 430.Lcopy_page_4K: 431 std r31,-32(1) 432 std r30,-40(1) 433 std r29,-48(1) 434 std r28,-56(1) 435 std r27,-64(1) 436 std r26,-72(1) 437 std r25,-80(1) 438 std r24,-88(1) 439 std r23,-96(1) 440 std r22,-104(1) 441 std r21,-112(1) 442 std r20,-120(1) 443 li r5,4096/32 - 1 444 addi r3,r3,-8 445 li r0,5 4460: addi r5,r5,-24 447 mtctr r0 448exc; ld r22,640(4) 449exc; ld r21,512(4) 450exc; ld r20,384(4) 451exc; ld r11,256(4) 452exc; ld r9,128(4) 453exc; ld r7,0(4) 454exc; ld r25,648(4) 455exc; ld r24,520(4) 456exc; ld r23,392(4) 457exc; ld r10,264(4) 458exc; ld r8,136(4) 459exc; ldu r6,8(4) 460 cmpwi r5,24 4611: 462exc; std r22,648(3) 463exc; std r21,520(3) 464exc; std r20,392(3) 465exc; std r11,264(3) 466exc; std r9,136(3) 467exc; std r7,8(3) 468exc; ld r28,648(4) 469exc; ld r27,520(4) 470exc; ld r26,392(4) 471exc; ld r31,264(4) 472exc; ld r30,136(4) 473exc; ld r29,8(4) 474exc; std r25,656(3) 475exc; std r24,528(3) 476exc; std r23,400(3) 477exc; std r10,272(3) 478exc; std r8,144(3) 479exc; std r6,16(3) 480exc; ld r22,656(4) 481exc; ld r21,528(4) 482exc; ld r20,400(4) 483exc; ld r11,272(4) 484exc; ld r9,144(4) 485exc; ld r7,16(4) 486exc; std r28,664(3) 487exc; std r27,536(3) 488exc; std r26,408(3) 489exc; std r31,280(3) 490exc; std r30,152(3) 491exc; stdu r29,24(3) 492exc; ld r25,664(4) 493exc; ld r24,536(4) 494exc; ld r23,408(4) 495exc; ld r10,280(4) 496exc; ld r8,152(4) 497exc; ldu r6,24(4) 498 bdnz 1b 499exc; std r22,648(3) 500exc; std r21,520(3) 501exc; std r20,392(3) 502exc; std r11,264(3) 503exc; std r9,136(3) 504exc; std r7,8(3) 505 addi r4,r4,640 506 addi r3,r3,648 507 bge 0b 508 mtctr r5 509exc; ld r7,0(4) 510exc; ld r8,8(4) 511exc; ldu r9,16(4) 5123: 513exc; ld r10,8(4) 514exc; std r7,8(3) 515exc; ld r7,16(4) 516exc; std r8,16(3) 517exc; ld r8,24(4) 518exc; std r9,24(3) 519exc; ldu r9,32(4) 520exc; stdu r10,32(3) 521 bdnz 3b 5224: 523exc; ld r10,8(4) 524exc; std r7,8(3) 525exc; std r8,16(3) 526exc; std r9,24(3) 527exc; std r10,32(3) 5289: ld r20,-120(1) 529 ld r21,-112(1) 530 ld r22,-104(1) 531 ld r23,-96(1) 532 ld r24,-88(1) 533 ld r25,-80(1) 534 ld r26,-72(1) 535 ld r27,-64(1) 536 ld r28,-56(1) 537 ld r29,-48(1) 538 ld r30,-40(1) 539 ld r31,-32(1) 540 li r3,0 541 blr 542 543/* 544 * on an exception, reset to the beginning and jump back into the 545 * standard __copy_tofrom_user 546 */ 547.Labort: 548 ld r20,-120(1) 549 ld r21,-112(1) 550 ld r22,-104(1) 551 ld r23,-96(1) 552 ld r24,-88(1) 553 ld r25,-80(1) 554 ld r26,-72(1) 555 ld r27,-64(1) 556 ld r28,-56(1) 557 ld r29,-48(1) 558 ld r30,-40(1) 559 ld r31,-32(1) 560 ld r3,-24(r1) 561 ld r4,-16(r1) 562 li r5,4096 563 b .Ldst_aligned 564EXPORT_SYMBOL(__copy_tofrom_user)