arch/tile/kernel/single_step.c

   1 /*
   2  * Copyright 2010 Tilera Corporation. All Rights Reserved.
   3  *
   4  *   This program is free software; you can redistribute it and/or
   5  *   modify it under the terms of the GNU General Public License
   6  *   as published by the Free Software Foundation, version 2.
   7  *
   8  *   This program is distributed in the hope that it will be useful, but
   9  *   WITHOUT ANY WARRANTY; without even the implied warranty of
  10  *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
  11  *   NON INFRINGEMENT.  See the GNU General Public License for
  12  *   more details.
  13  *
  14  * A code-rewriter that enables instruction single-stepping.
  15  * Derived from iLib's single-stepping code.
  16  */
  17
  18 #ifndef __tilegx__   /* Hardware support for single step unavailable. */
  19
  20 /* These functions are only used on the TILE platform */
  21 #include <linux/slab.h>
  22 #include <linux/thread_info.h>
  23 #include <linux/uaccess.h>
  24 #include <linux/mman.h>
  25 #include <linux/types.h>
  26 #include <linux/err.h>
  27 #include <asm/cacheflush.h>
  28 #include <asm/opcode-tile.h>
  29 #include <asm/opcode_constants.h>
  30 #include <arch/abi.h>
  31
  32 #define signExtend17(val) sign_extend((val), 17)
  33 #define TILE_X1_MASK (0xffffffffULL << 31)
  34
  35 int unaligned_printk;
  36
  37 static int __init setup_unaligned_printk(char *str)
  38 {
  39         long val;
  40         if (strict_strtol(str, 0, &val) != 0)
  41                 return 0;
  42         unaligned_printk = val;
  43         pr_info("Printk for each unaligned data accesses is %s\n",
  44                 unaligned_printk ? "enabled" : "disabled");
  45         return 1;
  46 }
  47 __setup("unaligned_printk=", setup_unaligned_printk);
  48
  49 unsigned int unaligned_fixup_count;
  50
  51 enum mem_op {
  52         MEMOP_NONE,
  53         MEMOP_LOAD,
  54         MEMOP_STORE,
  55         MEMOP_LOAD_POSTINCR,
  56         MEMOP_STORE_POSTINCR
  57 };
  58
  59 static inline tile_bundle_bits set_BrOff_X1(tile_bundle_bits n, s32 offset)
  60 {
  61         tile_bundle_bits result;
  62
  63         /* mask out the old offset */
  64         tile_bundle_bits mask = create_BrOff_X1(-1);
  65         result = n & (~mask);
  66
  67         /* or in the new offset */
  68         result |= create_BrOff_X1(offset);
  69
  70         return result;
  71 }
  72
  73 static inline tile_bundle_bits move_X1(tile_bundle_bits n, int dest, int src)
  74 {
  75         tile_bundle_bits result;
  76         tile_bundle_bits op;
  77
  78         result = n & (~TILE_X1_MASK);
  79
  80         op = create_Opcode_X1(SPECIAL_0_OPCODE_X1) |
  81                 create_RRROpcodeExtension_X1(OR_SPECIAL_0_OPCODE_X1) |
  82                 create_Dest_X1(dest) |
  83                 create_SrcB_X1(TREG_ZERO) |
  84                 create_SrcA_X1(src) ;
  85
  86         result |= op;
  87         return result;
  88 }
  89
  90 static inline tile_bundle_bits nop_X1(tile_bundle_bits n)
  91 {
  92         return move_X1(n, TREG_ZERO, TREG_ZERO);
  93 }
  94
  95 static inline tile_bundle_bits addi_X1(
  96         tile_bundle_bits n, int dest, int src, int imm)
  97 {
  98         n &= ~TILE_X1_MASK;
  99
 100         n |=  (create_SrcA_X1(src) |
 101                create_Dest_X1(dest) |
 102                create_Imm8_X1(imm) |
 103                create_S_X1(0) |
 104                create_Opcode_X1(IMM_0_OPCODE_X1) |
 105                create_ImmOpcodeExtension_X1(ADDI_IMM_0_OPCODE_X1));
 106
 107         return n;
 108 }
 109
 110 static tile_bundle_bits rewrite_load_store_unaligned(
 111         struct single_step_state *state,
 112         tile_bundle_bits bundle,
 113         struct pt_regs *regs,
 114         enum mem_op mem_op,
 115         int size, int sign_ext)
 116 {
 117         unsigned char __user *addr;
 118         int val_reg, addr_reg, err, val;
 119
 120         /* Get address and value registers */
 121         if (bundle & TILE_BUNDLE_Y_ENCODING_MASK) {
 122                 addr_reg = get_SrcA_Y2(bundle);
 123                 val_reg = get_SrcBDest_Y2(bundle);
 124         } else if (mem_op == MEMOP_LOAD || mem_op == MEMOP_LOAD_POSTINCR) {
 125                 addr_reg = get_SrcA_X1(bundle);
 126                 val_reg  = get_Dest_X1(bundle);
 127         } else {
 128                 addr_reg = get_SrcA_X1(bundle);
 129                 val_reg  = get_SrcB_X1(bundle);
 130         }
 131
 132         /*
 133          * If registers are not GPRs, don't try to handle it.
 134          *
 135          * FIXME: we could handle non-GPR loads by getting the real value
 136          * from memory, writing it to the single step buffer, using a
 137          * temp_reg to hold a pointer to that memory, then executing that
 138          * instruction and resetting temp_reg.  For non-GPR stores, it's a
 139          * little trickier; we could use the single step buffer for that
 140          * too, but we'd have to add some more state bits so that we could
 141          * call back in here to copy that value to the real target.  For
 142          * now, we just handle the simple case.
 143          */
 144         if ((val_reg >= PTREGS_NR_GPRS &&
 145              (val_reg != TREG_ZERO ||
 146               mem_op == MEMOP_LOAD ||
 147               mem_op == MEMOP_LOAD_POSTINCR)) ||
 148             addr_reg >= PTREGS_NR_GPRS)
 149                 return bundle;
 150
 151         /* If it's aligned, don't handle it specially */
 152         addr = (void __user *)regs->regs[addr_reg];
 153         if (((unsigned long)addr % size) == 0)
 154                 return bundle;
 155
 156 #ifndef __LITTLE_ENDIAN
 157 # error We assume little-endian representation with copy_xx_user size 2 here
 158 #endif
 159         /* Handle unaligned load/store */
 160         if (mem_op == MEMOP_LOAD || mem_op == MEMOP_LOAD_POSTINCR) {
 161                 unsigned short val_16;
 162                 switch (size) {
 163                 case 2:
 164                         err = copy_from_user(&val_16, addr, sizeof(val_16));
 165                         val = sign_ext ? ((short)val_16) : val_16;
 166                         break;
 167                 case 4:
 168                         err = copy_from_user(&val, addr, sizeof(val));
 169                         break;
 170                 default:
 171                         BUG();
 172                 }
 173                 if (err == 0) {
 174                         state->update_reg = val_reg;
 175                         state->update_value = val;
 176                         state->update = 1;
 177                 }
 178         } else {
 179                 val = (val_reg == TREG_ZERO) ? 0 : regs->regs[val_reg];
 180                 err = copy_to_user(addr, &val, size);
 181         }
 182
 183         if (err) {
 184                 siginfo_t info = {
 185                         .si_signo = SIGSEGV,
 186                         .si_code = SEGV_MAPERR,
 187                         .si_addr = addr
 188                 };
 189                 force_sig_info(info.si_signo, &info, current);
 190                 return (tile_bundle_bits) 0;
 191         }
 192
 193         if (unaligned_fixup == 0) {
 194                 siginfo_t info = {
 195                         .si_signo = SIGBUS,
 196                         .si_code = BUS_ADRALN,
 197                         .si_addr = addr
 198                 };
 199                 force_sig_info(info.si_signo, &info, current);
 200                 return (tile_bundle_bits) 0;
 201         }
 202
 203         if (unaligned_printk || unaligned_fixup_count == 0) {
 204                 pr_info("Process %d/%s: PC %#lx: Fixup of"
 205                         " unaligned %s at %#lx.\n",
 206                         current->pid, current->comm, regs->pc,
 207                         (mem_op == MEMOP_LOAD ||
 208                          mem_op == MEMOP_LOAD_POSTINCR) ?
 209                         "load" : "store",
 210                         (unsigned long)addr);
 211                 if (!unaligned_printk) {
 212 #define P pr_info
 213 P("\n");
 214 P("Unaligned fixups in the kernel will slow your application considerably.\n");
 215 P("To find them, write a \"1\" to /proc/sys/tile/unaligned_fixup/printk,\n");
 216 P("which requests the kernel show all unaligned fixups, or write a \"0\"\n");
 217 P("to /proc/sys/tile/unaligned_fixup/enabled, in which case each unaligned\n");
 218 P("access will become a SIGBUS you can debug. No further warnings will be\n");
 219 P("shown so as to avoid additional slowdown, but you can track the number\n");
 220 P("of fixups performed via /proc/sys/tile/unaligned_fixup/count.\n");
 221 P("Use the tile-addr2line command (see \"info addr2line\") to decode PCs.\n");
 222 P("\n");
 223 #undef P
 224                 }
 225         }
 226         ++unaligned_fixup_count;
 227
 228         if (bundle & TILE_BUNDLE_Y_ENCODING_MASK) {
 229                 /* Convert the Y2 instruction to a prefetch. */
 230                 bundle &= ~(create_SrcBDest_Y2(-1) |
 231                             create_Opcode_Y2(-1));
 232                 bundle |= (create_SrcBDest_Y2(TREG_ZERO) |
 233                            create_Opcode_Y2(LW_OPCODE_Y2));
 234         /* Replace the load postincr with an addi */
 235         } else if (mem_op == MEMOP_LOAD_POSTINCR) {
 236                 bundle = addi_X1(bundle, addr_reg, addr_reg,
 237                                  get_Imm8_X1(bundle));
 238         /* Replace the store postincr with an addi */
 239         } else if (mem_op == MEMOP_STORE_POSTINCR) {
 240                 bundle = addi_X1(bundle, addr_reg, addr_reg,
 241                                  get_Dest_Imm8_X1(bundle));
 242         } else {
 243                 /* Convert the X1 instruction to a nop. */
 244                 bundle &= ~(create_Opcode_X1(-1) |
 245                             create_UnShOpcodeExtension_X1(-1) |
 246                             create_UnOpcodeExtension_X1(-1));
 247                 bundle |= (create_Opcode_X1(SHUN_0_OPCODE_X1) |
 248                            create_UnShOpcodeExtension_X1(
 249                                    UN_0_SHUN_0_OPCODE_X1) |
 250                            create_UnOpcodeExtension_X1(
 251                                    NOP_UN_0_SHUN_0_OPCODE_X1));
 252         }
 253
 254         return bundle;
 255 }
 256
 257 /*
 258  * Called after execve() has started the new image.  This allows us
 259  * to reset the info state.  Note that the the mmap'ed memory, if there
 260  * was any, has already been unmapped by the exec.
 261  */
 262 void single_step_execve(void)
 263 {
 264         struct thread_info *ti = current_thread_info();
 265         kfree(ti->step_state);
 266         ti->step_state = NULL;
 267 }
 268
 269 /**
 270  * single_step_once() - entry point when single stepping has been triggered.
 271  * @regs: The machine register state
 272  *
 273  *  When we arrive at this routine via a trampoline, the single step
 274  *  engine copies the executing bundle to the single step buffer.
 275  *  If the instruction is a condition branch, then the target is
 276  *  reset to one past the next instruction. If the instruction
 277  *  sets the lr, then that is noted. If the instruction is a jump
 278  *  or call, then the new target pc is preserved and the current
 279  *  bundle instruction set to null.
 280  *
 281  *  The necessary post-single-step rewriting information is stored in
 282  *  single_step_state->  We use data segment values because the
 283  *  stack will be rewound when we run the rewritten single-stepped
 284  *  instruction.
 285  */
 286 void single_step_once(struct pt_regs *regs)
 287 {
 288         extern tile_bundle_bits __single_step_ill_insn;
 289         extern tile_bundle_bits __single_step_j_insn;
 290         extern tile_bundle_bits __single_step_addli_insn;
 291         extern tile_bundle_bits __single_step_auli_insn;
 292         struct thread_info *info = (void *)current_thread_info();
 293         struct single_step_state *state = info->step_state;
 294         int is_single_step = test_ti_thread_flag(info, TIF_SINGLESTEP);
 295         tile_bundle_bits __user *buffer, *pc;
 296         tile_bundle_bits bundle;
 297         int temp_reg;
 298         int target_reg = TREG_LR;
 299         int err;
 300         enum mem_op mem_op = MEMOP_NONE;
 301         int size = 0, sign_ext = 0;  /* happy compiler */
 302
 303         asm(
 304 "    .pushsection .rodata.single_step\n"
 305 "    .align 8\n"
 306 "    .globl    __single_step_ill_insn\n"
 307 "__single_step_ill_insn:\n"
 308 "    ill\n"
 309 "    .globl    __single_step_addli_insn\n"
 310 "__single_step_addli_insn:\n"
 311 "    { nop; addli r0, zero, 0 }\n"
 312 "    .globl    __single_step_auli_insn\n"
 313 "__single_step_auli_insn:\n"
 314 "    { nop; auli r0, r0, 0 }\n"
 315 "    .globl    __single_step_j_insn\n"
 316 "__single_step_j_insn:\n"
 317 "    j .\n"
 318 "    .popsection\n"
 319         );
 320
 321         if (state == NULL) {
 322                 /* allocate a page of writable, executable memory */
 323                 state = kmalloc(sizeof(struct single_step_state), GFP_KERNEL);
 324                 if (state == NULL) {
 325                         pr_err("Out of kernel memory trying to single-step\n");
 326                         return;
 327                 }
 328
 329                 /* allocate a cache line of writable, executable memory */
 330                 down_write(&current->mm->mmap_sem);
 331                 buffer = (void __user *) do_mmap(NULL, 0, 64,
 332                                           PROT_EXEC | PROT_READ | PROT_WRITE,
 333                                           MAP_PRIVATE | MAP_ANONYMOUS,
 334                                           0);
 335                 up_write(&current->mm->mmap_sem);
 336
 337                 if (IS_ERR((void __force *)buffer)) {
 338                         kfree(state);
 339                         pr_err("Out of kernel pages trying to single-step\n");
 340                         return;
 341                 }
 342
 343                 state->buffer = buffer;
 344                 state->is_enabled = 0;
 345
 346                 info->step_state = state;
 347
 348                 /* Validate our stored instruction patterns */
 349                 BUG_ON(get_Opcode_X1(__single_step_addli_insn) !=
 350                        ADDLI_OPCODE_X1);
 351                 BUG_ON(get_Opcode_X1(__single_step_auli_insn) !=
 352                        AULI_OPCODE_X1);
 353                 BUG_ON(get_SrcA_X1(__single_step_addli_insn) != TREG_ZERO);
 354                 BUG_ON(get_Dest_X1(__single_step_addli_insn) != 0);
 355                 BUG_ON(get_JOffLong_X1(__single_step_j_insn) != 0);
 356         }
 357
 358         /*
 359          * If we are returning from a syscall, we still haven't hit the
 360          * "ill" for the swint1 instruction.  So back the PC up to be
 361          * pointing at the swint1, but we'll actually return directly
 362          * back to the "ill" so we come back in via SIGILL as if we
 363          * had "executed" the swint1 without ever being in kernel space.
 364          */
 365         if (regs->faultnum == INT_SWINT_1)
 366                 regs->pc -= 8;
 367
 368         pc = (tile_bundle_bits __user *)(regs->pc);
 369         if (get_user(bundle, pc) != 0) {
 370                 pr_err("Couldn't read instruction at %p trying to step\n", pc);
 371                 return;
 372         }
 373
 374         /* We'll follow the instruction with 2 ill op bundles */
 375         state->orig_pc = (unsigned long)pc;
 376         state->next_pc = (unsigned long)(pc + 1);
 377         state->branch_next_pc = 0;
 378         state->update = 0;
 379
 380         if (!(bundle & TILE_BUNDLE_Y_ENCODING_MASK)) {
 381                 /* two wide, check for control flow */
 382                 int opcode = get_Opcode_X1(bundle);
 383
 384                 switch (opcode) {
 385                 /* branches */
 386                 case BRANCH_OPCODE_X1:
 387                 {
 388                         s32 offset = signExtend17(get_BrOff_X1(bundle));
 389
 390                         /*
 391                          * For branches, we use a rewriting trick to let the
 392                          * hardware evaluate whether the branch is taken or
 393                          * untaken.  We record the target offset and then
 394                          * rewrite the branch instruction to target 1 insn
 395                          * ahead if the branch is taken.  We then follow the
 396                          * rewritten branch with two bundles, each containing
 397                          * an "ill" instruction. The supervisor examines the
 398                          * pc after the single step code is executed, and if
 399                          * the pc is the first ill instruction, then the
 400                          * branch (if any) was not taken.  If the pc is the
 401                          * second ill instruction, then the branch was
 402                          * taken. The new pc is computed for these cases, and
 403                          * inserted into the registers for the thread.  If
 404                          * the pc is the start of the single step code, then
 405                          * an exception or interrupt was taken before the
 406                          * code started processing, and the same "original"
 407                          * pc is restored.  This change, different from the
 408                          * original implementation, has the advantage of
 409                          * executing a single user instruction.
 410                          */
 411                         state->branch_next_pc = (unsigned long)(pc + offset);
 412
 413                         /* rewrite branch offset to go forward one bundle */
 414                         bundle = set_BrOff_X1(bundle, 2);
 415                 }
 416                 break;
 417
 418                 /* jumps */
 419                 case JALB_OPCODE_X1:
 420                 case JALF_OPCODE_X1:
 421                         state->update = 1;
 422                         state->next_pc =
 423                                 (unsigned long) (pc + get_JOffLong_X1(bundle));
 424                         break;
 425
 426                 case JB_OPCODE_X1:
 427                 case JF_OPCODE_X1:
 428                         state->next_pc =
 429                                 (unsigned long) (pc + get_JOffLong_X1(bundle));
 430                         bundle = nop_X1(bundle);
 431                         break;
 432
 433                 case SPECIAL_0_OPCODE_X1:
 434                         switch (get_RRROpcodeExtension_X1(bundle)) {
 435                         /* jump-register */
 436                         case JALRP_SPECIAL_0_OPCODE_X1:
 437                         case JALR_SPECIAL_0_OPCODE_X1:
 438                                 state->update = 1;
 439                                 state->next_pc =
 440                                         regs->regs[get_SrcA_X1(bundle)];
 441                                 break;
 442
 443                         case JRP_SPECIAL_0_OPCODE_X1:
 444                         case JR_SPECIAL_0_OPCODE_X1:
 445                                 state->next_pc =
 446                                         regs->regs[get_SrcA_X1(bundle)];
 447                                 bundle = nop_X1(bundle);
 448                                 break;
 449
 450                         case LNK_SPECIAL_0_OPCODE_X1:
 451                                 state->update = 1;
 452                                 target_reg = get_Dest_X1(bundle);
 453                                 break;
 454
 455                         /* stores */
 456                         case SH_SPECIAL_0_OPCODE_X1:
 457                                 mem_op = MEMOP_STORE;
 458                                 size = 2;
 459                                 break;
 460
 461                         case SW_SPECIAL_0_OPCODE_X1:
 462                                 mem_op = MEMOP_STORE;
 463                                 size = 4;
 464                                 break;
 465                         }
 466                         break;
 467
 468                 /* loads and iret */
 469                 case SHUN_0_OPCODE_X1:
 470                         if (get_UnShOpcodeExtension_X1(bundle) ==
 471                             UN_0_SHUN_0_OPCODE_X1) {
 472                                 switch (get_UnOpcodeExtension_X1(bundle)) {
 473                                 case LH_UN_0_SHUN_0_OPCODE_X1:
 474                                         mem_op = MEMOP_LOAD;
 475                                         size = 2;
 476                                         sign_ext = 1;
 477                                         break;
 478
 479                                 case LH_U_UN_0_SHUN_0_OPCODE_X1:
 480                                         mem_op = MEMOP_LOAD;
 481                                         size = 2;
 482                                         sign_ext = 0;
 483                                         break;
 484
 485                                 case LW_UN_0_SHUN_0_OPCODE_X1:
 486                                         mem_op = MEMOP_LOAD;
 487                                         size = 4;
 488                                         break;
 489
 490                                 case IRET_UN_0_SHUN_0_OPCODE_X1:
 491                                 {
 492                                         unsigned long ex0_0 = __insn_mfspr(
 493                                                 SPR_EX_CONTEXT_0_0);
 494                                         unsigned long ex0_1 = __insn_mfspr(
 495                                                 SPR_EX_CONTEXT_0_1);
 496                                         /*
 497                                          * Special-case it if we're iret'ing
 498                                          * to PL0 again.  Otherwise just let
 499                                          * it run and it will generate SIGILL.
 500                                          */
 501                                         if (EX1_PL(ex0_1) == USER_PL) {
 502                                                 state->next_pc = ex0_0;
 503                                                 regs->ex1 = ex0_1;
 504                                                 bundle = nop_X1(bundle);
 505                                         }
 506                                 }
 507                                 }
 508                         }
 509                         break;
 510
 511 #if CHIP_HAS_WH64()
 512                 /* postincrement operations */
 513                 case IMM_0_OPCODE_X1:
 514                         switch (get_ImmOpcodeExtension_X1(bundle)) {
 515                         case LWADD_IMM_0_OPCODE_X1:
 516                                 mem_op = MEMOP_LOAD_POSTINCR;
 517                                 size = 4;
 518                                 break;
 519
 520                         case LHADD_IMM_0_OPCODE_X1:
 521                                 mem_op = MEMOP_LOAD_POSTINCR;
 522                                 size = 2;
 523                                 sign_ext = 1;
 524                                 break;
 525
 526                         case LHADD_U_IMM_0_OPCODE_X1:
 527                                 mem_op = MEMOP_LOAD_POSTINCR;
 528                                 size = 2;
 529                                 sign_ext = 0;
 530                                 break;
 531
 532                         case SWADD_IMM_0_OPCODE_X1:
 533                                 mem_op = MEMOP_STORE_POSTINCR;
 534                                 size = 4;
 535                                 break;
 536
 537                         case SHADD_IMM_0_OPCODE_X1:
 538                                 mem_op = MEMOP_STORE_POSTINCR;
 539                                 size = 2;
 540                                 break;
 541
 542                         default:
 543                                 break;
 544                         }
 545                         break;
 546 #endif /* CHIP_HAS_WH64() */
 547                 }
 548
 549                 if (state->update) {
 550                         /*
 551                          * Get an available register.  We start with a
 552                          * bitmask with 1's for available registers.
 553                          * We truncate to the low 32 registers since
 554                          * we are guaranteed to have set bits in the
 555                          * low 32 bits, then use ctz to pick the first.
 556                          */
 557                         u32 mask = (u32) ~((1ULL << get_Dest_X0(bundle)) |
 558                                            (1ULL << get_SrcA_X0(bundle)) |
 559                                            (1ULL << get_SrcB_X0(bundle)) |
 560                                            (1ULL << target_reg));
 561                         temp_reg = __builtin_ctz(mask);
 562                         state->update_reg = temp_reg;
 563                         state->update_value = regs->regs[temp_reg];
 564                         regs->regs[temp_reg] = (unsigned long) (pc+1);
 565                         regs->flags |= PT_FLAGS_RESTORE_REGS;
 566                         bundle = move_X1(bundle, target_reg, temp_reg);
 567                 }
 568         } else {
 569                 int opcode = get_Opcode_Y2(bundle);
 570
 571                 switch (opcode) {
 572                 /* loads */
 573                 case LH_OPCODE_Y2:
 574                         mem_op = MEMOP_LOAD;
 575                         size = 2;
 576                         sign_ext = 1;
 577                         break;
 578
 579                 case LH_U_OPCODE_Y2:
 580                         mem_op = MEMOP_LOAD;
 581                         size = 2;
 582                         sign_ext = 0;
 583                         break;
 584
 585                 case LW_OPCODE_Y2:
 586                         mem_op = MEMOP_LOAD;
 587                         size = 4;
 588                         break;
 589
 590                 /* stores */
 591                 case SH_OPCODE_Y2:
 592                         mem_op = MEMOP_STORE;
 593                         size = 2;
 594                         break;
 595
 596                 case SW_OPCODE_Y2:
 597                         mem_op = MEMOP_STORE;
 598                         size = 4;
 599                         break;
 600                 }
 601         }
 602
 603         /*
 604          * Check if we need to rewrite an unaligned load/store.
 605          * Returning zero is a special value meaning we need to SIGSEGV.
 606          */
 607         if (mem_op != MEMOP_NONE && unaligned_fixup >= 0) {
 608                 bundle = rewrite_load_store_unaligned(state, bundle, regs,
 609                                                       mem_op, size, sign_ext);
 610                 if (bundle == 0)
 611                         return;
 612         }
 613
 614         /* write the bundle to our execution area */
 615         buffer = state->buffer;
 616         err = __put_user(bundle, buffer++);
 617
 618         /*
 619          * If we're really single-stepping, we take an INT_ILL after.
 620          * If we're just handling an unaligned access, we can just
 621          * jump directly back to where we were in user code.
 622          */
 623         if (is_single_step) {
 624                 err |= __put_user(__single_step_ill_insn, buffer++);
 625                 err |= __put_user(__single_step_ill_insn, buffer++);
 626         } else {
 627                 long delta;
 628
 629                 if (state->update) {
 630                         /* We have some state to update; do it inline */
 631                         int ha16;
 632                         bundle = __single_step_addli_insn;
 633                         bundle |= create_Dest_X1(state->update_reg);
 634                         bundle |= create_Imm16_X1(state->update_value);
 635                         err |= __put_user(bundle, buffer++);
 636                         bundle = __single_step_auli_insn;
 637                         bundle |= create_Dest_X1(state->update_reg);
 638                         bundle |= create_SrcA_X1(state->update_reg);
 639                         ha16 = (state->update_value + 0x8000) >> 16;
 640                         bundle |= create_Imm16_X1(ha16);
 641                         err |= __put_user(bundle, buffer++);
 642                         state->update = 0;
 643                 }
 644
 645                 /* End with a jump back to the next instruction */
 646                 delta = ((regs->pc + TILE_BUNDLE_SIZE_IN_BYTES) -
 647                         (unsigned long)buffer) >>
 648                         TILE_LOG2_BUNDLE_ALIGNMENT_IN_BYTES;
 649                 bundle = __single_step_j_insn;
 650                 bundle |= create_JOffLong_X1(delta);
 651                 err |= __put_user(bundle, buffer++);
 652         }
 653
 654         if (err) {
 655                 pr_err("Fault when writing to single-step buffer\n");
 656                 return;
 657         }
 658
 659         /*
 660          * Flush the buffer.
 661          * We do a local flush only, since this is a thread-specific buffer.
 662          */
 663         __flush_icache_range((unsigned long)state->buffer,
 664                              (unsigned long)buffer);
 665
 666         /* Indicate enabled */
 667         state->is_enabled = is_single_step;
 668         regs->pc = (unsigned long)state->buffer;
 669
 670         /* Fault immediately if we are coming back from a syscall. */
 671         if (regs->faultnum == INT_SWINT_1)
 672                 regs->pc += 8;
 673 }
 674
 675 #else
 676 #include <linux/smp.h>
 677 #include <linux/ptrace.h>
 678 #include <arch/spr_def.h>
 679
 680 static DEFINE_PER_CPU(unsigned long, ss_saved_pc);
 681
 682
 683 /*
 684  * Called directly on the occasion of an interrupt.
 685  *
 686  * If the process doesn't have single step set, then we use this as an
 687  * opportunity to turn single step off.
 688  *
 689  * It has been mentioned that we could conditionally turn off single stepping
 690  * on each entry into the kernel and rely on single_step_once to turn it
 691  * on for the processes that matter (as we already do), but this
 692  * implementation is somewhat more efficient in that we muck with registers
 693  * once on a bum interrupt rather than on every entry into the kernel.
 694  *
 695  * If SINGLE_STEP_CONTROL_K has CANCELED set, then an interrupt occurred,
 696  * so we have to run through this process again before we can say that an
 697  * instruction has executed.
 698  *
 699  * swint will set CANCELED, but it's a legitimate instruction.  Fortunately
 700  * it changes the PC.  If it hasn't changed, then we know that the interrupt
 701  * wasn't generated by swint and we'll need to run this process again before
 702  * we can say an instruction has executed.
 703  *
 704  * If either CANCELED == 0 or the PC's changed, we send out SIGTRAPs and get
 705  * on with our lives.
 706  */
 707
 708 void gx_singlestep_handle(struct pt_regs *regs, int fault_num)
 709 {
 710         unsigned long *ss_pc = &__get_cpu_var(ss_saved_pc);
 711         struct thread_info *info = (void *)current_thread_info();
 712         int is_single_step = test_ti_thread_flag(info, TIF_SINGLESTEP);
 713         unsigned long control = __insn_mfspr(SPR_SINGLE_STEP_CONTROL_K);
 714
 715         if (is_single_step == 0) {
 716                 __insn_mtspr(SPR_SINGLE_STEP_EN_K_K, 0);
 717
 718         } else if ((*ss_pc != regs->pc) ||
 719                    (!(control & SPR_SINGLE_STEP_CONTROL_1__CANCELED_MASK))) {
 720
 721                 ptrace_notify(SIGTRAP);
 722                 control |= SPR_SINGLE_STEP_CONTROL_1__CANCELED_MASK;
 723                 control |= SPR_SINGLE_STEP_CONTROL_1__INHIBIT_MASK;
 724                 __insn_mtspr(SPR_SINGLE_STEP_CONTROL_K, control);
 725         }
 726 }
 727
 728
 729 /*
 730  * Called from need_singlestep.  Set up the control registers and the enable
 731  * register, then return back.
 732  */
 733
 734 void single_step_once(struct pt_regs *regs)
 735 {
 736         unsigned long *ss_pc = &__get_cpu_var(ss_saved_pc);
 737         unsigned long control = __insn_mfspr(SPR_SINGLE_STEP_CONTROL_K);
 738
 739         *ss_pc = regs->pc;
 740         control |= SPR_SINGLE_STEP_CONTROL_1__CANCELED_MASK;
 741         control |= SPR_SINGLE_STEP_CONTROL_1__INHIBIT_MASK;
 742         __insn_mtspr(SPR_SINGLE_STEP_CONTROL_K, control);
 743         __insn_mtspr(SPR_SINGLE_STEP_EN_K_K, 1 << USER_PL);
 744 }
 745
 746 void single_step_execve(void)
 747 {
 748         /* Nothing */
 749 }
 750
 751 #endif /* !__tilegx__ */