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  */
  16
  17 #include <linux/smp.h>
  18 #include <linux/ptrace.h>
  19 #include <linux/slab.h>
  20 #include <linux/thread_info.h>
  21 #include <linux/uaccess.h>
  22 #include <linux/mman.h>
  23 #include <linux/types.h>
  24 #include <linux/err.h>
  25 #include <linux/prctl.h>
  26 #include <asm/cacheflush.h>
  27 #include <asm/traps.h>
  28 #include <linux/uaccess.h>
  29 #include <asm/unaligned.h>
  30 #include <arch/abi.h>
  31 #include <arch/spr_def.h>
  32 #include <arch/opcode.h>
  33
  34
  35 #ifndef __tilegx__   /* Hardware support for single step unavailable. */
  36
  37 #define signExtend17(val) sign_extend((val), 17)
  38 #define TILE_X1_MASK (0xffffffffULL << 31)
  39
  40 enum mem_op {
  41         MEMOP_NONE,
  42         MEMOP_LOAD,
  43         MEMOP_STORE,
  44         MEMOP_LOAD_POSTINCR,
  45         MEMOP_STORE_POSTINCR
  46 };
  47
  48 static inline tilepro_bundle_bits set_BrOff_X1(tilepro_bundle_bits n,
  49         s32 offset)
  50 {
  51         tilepro_bundle_bits result;
  52
  53         /* mask out the old offset */
  54         tilepro_bundle_bits mask = create_BrOff_X1(-1);
  55         result = n & (~mask);
  56
  57         /* or in the new offset */
  58         result |= create_BrOff_X1(offset);
  59
  60         return result;
  61 }
  62
  63 static inline tilepro_bundle_bits move_X1(tilepro_bundle_bits n, int dest,
  64         int src)
  65 {
  66         tilepro_bundle_bits result;
  67         tilepro_bundle_bits op;
  68
  69         result = n & (~TILE_X1_MASK);
  70
  71         op = create_Opcode_X1(SPECIAL_0_OPCODE_X1) |
  72                 create_RRROpcodeExtension_X1(OR_SPECIAL_0_OPCODE_X1) |
  73                 create_Dest_X1(dest) |
  74                 create_SrcB_X1(TREG_ZERO) |
  75                 create_SrcA_X1(src) ;
  76
  77         result |= op;
  78         return result;
  79 }
  80
  81 static inline tilepro_bundle_bits nop_X1(tilepro_bundle_bits n)
  82 {
  83         return move_X1(n, TREG_ZERO, TREG_ZERO);
  84 }
  85
  86 static inline tilepro_bundle_bits addi_X1(
  87         tilepro_bundle_bits n, int dest, int src, int imm)
  88 {
  89         n &= ~TILE_X1_MASK;
  90
  91         n |=  (create_SrcA_X1(src) |
  92                create_Dest_X1(dest) |
  93                create_Imm8_X1(imm) |
  94                create_S_X1(0) |
  95                create_Opcode_X1(IMM_0_OPCODE_X1) |
  96                create_ImmOpcodeExtension_X1(ADDI_IMM_0_OPCODE_X1));
  97
  98         return n;
  99 }
 100
 101 static tilepro_bundle_bits rewrite_load_store_unaligned(
 102         struct single_step_state *state,
 103         tilepro_bundle_bits bundle,
 104         struct pt_regs *regs,
 105         enum mem_op mem_op,
 106         int size, int sign_ext)
 107 {
 108         unsigned char __user *addr;
 109         int val_reg, addr_reg, err, val;
 110         int align_ctl;
 111
 112         align_ctl = unaligned_fixup;
 113         switch (task_thread_info(current)->align_ctl) {
 114         case PR_UNALIGN_NOPRINT:
 115                 align_ctl = 1;
 116                 break;
 117         case PR_UNALIGN_SIGBUS:
 118                 align_ctl = 0;
 119                 break;
 120         }
 121
 122         /* Get address and value registers */
 123         if (bundle & TILEPRO_BUNDLE_Y_ENCODING_MASK) {
 124                 addr_reg = get_SrcA_Y2(bundle);
 125                 val_reg = get_SrcBDest_Y2(bundle);
 126         } else if (mem_op == MEMOP_LOAD || mem_op == MEMOP_LOAD_POSTINCR) {
 127                 addr_reg = get_SrcA_X1(bundle);
 128                 val_reg  = get_Dest_X1(bundle);
 129         } else {
 130                 addr_reg = get_SrcA_X1(bundle);
 131                 val_reg  = get_SrcB_X1(bundle);
 132         }
 133
 134         /*
 135          * If registers are not GPRs, don't try to handle it.
 136          *
 137          * FIXME: we could handle non-GPR loads by getting the real value
 138          * from memory, writing it to the single step buffer, using a
 139          * temp_reg to hold a pointer to that memory, then executing that
 140          * instruction and resetting temp_reg.  For non-GPR stores, it's a
 141          * little trickier; we could use the single step buffer for that
 142          * too, but we'd have to add some more state bits so that we could
 143          * call back in here to copy that value to the real target.  For
 144          * now, we just handle the simple case.
 145          */
 146         if ((val_reg >= PTREGS_NR_GPRS &&
 147              (val_reg != TREG_ZERO ||
 148               mem_op == MEMOP_LOAD ||
 149               mem_op == MEMOP_LOAD_POSTINCR)) ||
 150             addr_reg >= PTREGS_NR_GPRS)
 151                 return bundle;
 152
 153         /* If it's aligned, don't handle it specially */
 154         addr = (void __user *)regs->regs[addr_reg];
 155         if (((unsigned long)addr % size) == 0)
 156                 return bundle;
 157
 158         /*
 159          * Return SIGBUS with the unaligned address, if requested.
 160          * Note that we return SIGBUS even for completely invalid addresses
 161          * as long as they are in fact unaligned; this matches what the
 162          * tilepro hardware would be doing, if it could provide us with the
 163          * actual bad address in an SPR, which it doesn't.
 164          */
 165         if (align_ctl == 0) {
 166                 siginfo_t info;
 167
 168                 clear_siginfo(&info);
 169                 info.si_signo = SIGBUS;
 170                 info.si_code = BUS_ADRALN;
 171                 info.si_addr = addr;
 172
 173                 trace_unhandled_signal("unaligned trap", regs,
 174                                        (unsigned long)addr, SIGBUS);
 175                 force_sig_info(info.si_signo, &info, current);
 176                 return (tilepro_bundle_bits) 0;
 177         }
 178
 179         /* Handle unaligned load/store */
 180         if (mem_op == MEMOP_LOAD || mem_op == MEMOP_LOAD_POSTINCR) {
 181                 unsigned short val_16;
 182                 switch (size) {
 183                 case 2:
 184                         err = copy_from_user(&val_16, addr, sizeof(val_16));
 185                         val = sign_ext ? ((short)val_16) : val_16;
 186                         break;
 187                 case 4:
 188                         err = copy_from_user(&val, addr, sizeof(val));
 189                         break;
 190                 default:
 191                         BUG();
 192                 }
 193                 if (err == 0) {
 194                         state->update_reg = val_reg;
 195                         state->update_value = val;
 196                         state->update = 1;
 197                 }
 198         } else {
 199                 unsigned short val_16;
 200                 val = (val_reg == TREG_ZERO) ? 0 : regs->regs[val_reg];
 201                 switch (size) {
 202                 case 2:
 203                         val_16 = val;
 204                         err = copy_to_user(addr, &val_16, sizeof(val_16));
 205                         break;
 206                 case 4:
 207                         err = copy_to_user(addr, &val, sizeof(val));
 208                         break;
 209                 default:
 210                         BUG();
 211                 }
 212         }
 213
 214         if (err) {
 215                 siginfo_t info;
 216
 217                 clear_siginfo(&info);
 218                 info.si_signo = SIGBUS;
 219                 info.si_code = BUS_ADRALN;
 220                 info.si_addr = addr;
 221
 222                 trace_unhandled_signal("bad address for unaligned fixup", regs,
 223                                        (unsigned long)addr, SIGBUS);
 224                 force_sig_info(info.si_signo, &info, current);
 225                 return (tilepro_bundle_bits) 0;
 226         }
 227
 228         if (unaligned_printk || unaligned_fixup_count == 0) {
 229                 pr_info("Process %d/%s: PC %#lx: Fixup of unaligned %s at %#lx\n",
 230                         current->pid, current->comm, regs->pc,
 231                         mem_op == MEMOP_LOAD || mem_op == MEMOP_LOAD_POSTINCR ?
 232                         "load" : "store",
 233                         (unsigned long)addr);
 234                 if (!unaligned_printk) {
 235 #define P pr_info
 236 P("\n");
 237 P("Unaligned fixups in the kernel will slow your application considerably.\n");
 238 P("To find them, write a \"1\" to /proc/sys/tile/unaligned_fixup/printk,\n");
 239 P("which requests the kernel show all unaligned fixups, or write a \"0\"\n");
 240 P("to /proc/sys/tile/unaligned_fixup/enabled, in which case each unaligned\n");
 241 P("access will become a SIGBUS you can debug. No further warnings will be\n");
 242 P("shown so as to avoid additional slowdown, but you can track the number\n");
 243 P("of fixups performed via /proc/sys/tile/unaligned_fixup/count.\n");
 244 P("Use the tile-addr2line command (see \"info addr2line\") to decode PCs.\n");
 245 P("\n");
 246 #undef P
 247                 }
 248         }
 249         ++unaligned_fixup_count;
 250
 251         if (bundle & TILEPRO_BUNDLE_Y_ENCODING_MASK) {
 252                 /* Convert the Y2 instruction to a prefetch. */
 253                 bundle &= ~(create_SrcBDest_Y2(-1) |
 254                             create_Opcode_Y2(-1));
 255                 bundle |= (create_SrcBDest_Y2(TREG_ZERO) |
 256                            create_Opcode_Y2(LW_OPCODE_Y2));
 257         /* Replace the load postincr with an addi */
 258         } else if (mem_op == MEMOP_LOAD_POSTINCR) {
 259                 bundle = addi_X1(bundle, addr_reg, addr_reg,
 260                                  get_Imm8_X1(bundle));
 261         /* Replace the store postincr with an addi */
 262         } else if (mem_op == MEMOP_STORE_POSTINCR) {
 263                 bundle = addi_X1(bundle, addr_reg, addr_reg,
 264                                  get_Dest_Imm8_X1(bundle));
 265         } else {
 266                 /* Convert the X1 instruction to a nop. */
 267                 bundle &= ~(create_Opcode_X1(-1) |
 268                             create_UnShOpcodeExtension_X1(-1) |
 269                             create_UnOpcodeExtension_X1(-1));
 270                 bundle |= (create_Opcode_X1(SHUN_0_OPCODE_X1) |
 271                            create_UnShOpcodeExtension_X1(
 272                                    UN_0_SHUN_0_OPCODE_X1) |
 273                            create_UnOpcodeExtension_X1(
 274                                    NOP_UN_0_SHUN_0_OPCODE_X1));
 275         }
 276
 277         return bundle;
 278 }
 279
 280 /*
 281  * Called after execve() has started the new image.  This allows us
 282  * to reset the info state.  Note that the the mmap'ed memory, if there
 283  * was any, has already been unmapped by the exec.
 284  */
 285 void single_step_execve(void)
 286 {
 287         struct thread_info *ti = current_thread_info();
 288         kfree(ti->step_state);
 289         ti->step_state = NULL;
 290 }
 291
 292 /*
 293  * single_step_once() - entry point when single stepping has been triggered.
 294  * @regs: The machine register state
 295  *
 296  *  When we arrive at this routine via a trampoline, the single step
 297  *  engine copies the executing bundle to the single step buffer.
 298  *  If the instruction is a condition branch, then the target is
 299  *  reset to one past the next instruction. If the instruction
 300  *  sets the lr, then that is noted. If the instruction is a jump
 301  *  or call, then the new target pc is preserved and the current
 302  *  bundle instruction set to null.
 303  *
 304  *  The necessary post-single-step rewriting information is stored in
 305  *  single_step_state->  We use data segment values because the
 306  *  stack will be rewound when we run the rewritten single-stepped
 307  *  instruction.
 308  */
 309 void single_step_once(struct pt_regs *regs)
 310 {
 311         extern tilepro_bundle_bits __single_step_ill_insn;
 312         extern tilepro_bundle_bits __single_step_j_insn;
 313         extern tilepro_bundle_bits __single_step_addli_insn;
 314         extern tilepro_bundle_bits __single_step_auli_insn;
 315         struct thread_info *info = (void *)current_thread_info();
 316         struct single_step_state *state = info->step_state;
 317         int is_single_step = test_ti_thread_flag(info, TIF_SINGLESTEP);
 318         tilepro_bundle_bits __user *buffer, *pc;
 319         tilepro_bundle_bits bundle;
 320         int temp_reg;
 321         int target_reg = TREG_LR;
 322         int err;
 323         enum mem_op mem_op = MEMOP_NONE;
 324         int size = 0, sign_ext = 0;  /* happy compiler */
 325         int align_ctl;
 326
 327         align_ctl = unaligned_fixup;
 328         switch (task_thread_info(current)->align_ctl) {
 329         case PR_UNALIGN_NOPRINT:
 330                 align_ctl = 1;
 331                 break;
 332         case PR_UNALIGN_SIGBUS:
 333                 align_ctl = 0;
 334                 break;
 335         }
 336
 337         asm(
 338 "    .pushsection .rodata.single_step\n"
 339 "    .align 8\n"
 340 "    .globl    __single_step_ill_insn\n"
 341 "__single_step_ill_insn:\n"
 342 "    ill\n"
 343 "    .globl    __single_step_addli_insn\n"
 344 "__single_step_addli_insn:\n"
 345 "    { nop; addli r0, zero, 0 }\n"
 346 "    .globl    __single_step_auli_insn\n"
 347 "__single_step_auli_insn:\n"
 348 "    { nop; auli r0, r0, 0 }\n"
 349 "    .globl    __single_step_j_insn\n"
 350 "__single_step_j_insn:\n"
 351 "    j .\n"
 352 "    .popsection\n"
 353         );
 354
 355         /*
 356          * Enable interrupts here to allow touching userspace and the like.
 357          * The callers expect this: do_trap() already has interrupts
 358          * enabled, and do_work_pending() handles functions that enable
 359          * interrupts internally.
 360          */
 361         local_irq_enable();
 362
 363         if (state == NULL) {
 364                 /* allocate a page of writable, executable memory */
 365                 state = kmalloc(sizeof(struct single_step_state), GFP_KERNEL);
 366                 if (state == NULL) {
 367                         pr_err("Out of kernel memory trying to single-step\n");
 368                         return;
 369                 }
 370
 371                 /* allocate a cache line of writable, executable memory */
 372                 buffer = (void __user *) vm_mmap(NULL, 0, 64,
 373                                           PROT_EXEC | PROT_READ | PROT_WRITE,
 374                                           MAP_PRIVATE | MAP_ANONYMOUS,
 375                                           0);
 376
 377                 if (IS_ERR((void __force *)buffer)) {
 378                         kfree(state);
 379                         pr_err("Out of kernel pages trying to single-step\n");
 380                         return;
 381                 }
 382
 383                 state->buffer = buffer;
 384                 state->is_enabled = 0;
 385
 386                 info->step_state = state;
 387
 388                 /* Validate our stored instruction patterns */
 389                 BUG_ON(get_Opcode_X1(__single_step_addli_insn) !=
 390                        ADDLI_OPCODE_X1);
 391                 BUG_ON(get_Opcode_X1(__single_step_auli_insn) !=
 392                        AULI_OPCODE_X1);
 393                 BUG_ON(get_SrcA_X1(__single_step_addli_insn) != TREG_ZERO);
 394                 BUG_ON(get_Dest_X1(__single_step_addli_insn) != 0);
 395                 BUG_ON(get_JOffLong_X1(__single_step_j_insn) != 0);
 396         }
 397
 398         /*
 399          * If we are returning from a syscall, we still haven't hit the
 400          * "ill" for the swint1 instruction.  So back the PC up to be
 401          * pointing at the swint1, but we'll actually return directly
 402          * back to the "ill" so we come back in via SIGILL as if we
 403          * had "executed" the swint1 without ever being in kernel space.
 404          */
 405         if (regs->faultnum == INT_SWINT_1)
 406                 regs->pc -= 8;
 407
 408         pc = (tilepro_bundle_bits __user *)(regs->pc);
 409         if (get_user(bundle, pc) != 0) {
 410                 pr_err("Couldn't read instruction at %p trying to step\n", pc);
 411                 return;
 412         }
 413
 414         /* We'll follow the instruction with 2 ill op bundles */
 415         state->orig_pc = (unsigned long)pc;
 416         state->next_pc = (unsigned long)(pc + 1);
 417         state->branch_next_pc = 0;
 418         state->update = 0;
 419
 420         if (!(bundle & TILEPRO_BUNDLE_Y_ENCODING_MASK)) {
 421                 /* two wide, check for control flow */
 422                 int opcode = get_Opcode_X1(bundle);
 423
 424                 switch (opcode) {
 425                 /* branches */
 426                 case BRANCH_OPCODE_X1:
 427                 {
 428                         s32 offset = signExtend17(get_BrOff_X1(bundle));
 429
 430                         /*
 431                          * For branches, we use a rewriting trick to let the
 432                          * hardware evaluate whether the branch is taken or
 433                          * untaken.  We record the target offset and then
 434                          * rewrite the branch instruction to target 1 insn
 435                          * ahead if the branch is taken.  We then follow the
 436                          * rewritten branch with two bundles, each containing
 437                          * an "ill" instruction. The supervisor examines the
 438                          * pc after the single step code is executed, and if
 439                          * the pc is the first ill instruction, then the
 440                          * branch (if any) was not taken.  If the pc is the
 441                          * second ill instruction, then the branch was
 442                          * taken. The new pc is computed for these cases, and
 443                          * inserted into the registers for the thread.  If
 444                          * the pc is the start of the single step code, then
 445                          * an exception or interrupt was taken before the
 446                          * code started processing, and the same "original"
 447                          * pc is restored.  This change, different from the
 448                          * original implementation, has the advantage of
 449                          * executing a single user instruction.
 450                          */
 451                         state->branch_next_pc = (unsigned long)(pc + offset);
 452
 453                         /* rewrite branch offset to go forward one bundle */
 454                         bundle = set_BrOff_X1(bundle, 2);
 455                 }
 456                 break;
 457
 458                 /* jumps */
 459                 case JALB_OPCODE_X1:
 460                 case JALF_OPCODE_X1:
 461                         state->update = 1;
 462                         state->next_pc =
 463                                 (unsigned long) (pc + get_JOffLong_X1(bundle));
 464                         break;
 465
 466                 case JB_OPCODE_X1:
 467                 case JF_OPCODE_X1:
 468                         state->next_pc =
 469                                 (unsigned long) (pc + get_JOffLong_X1(bundle));
 470                         bundle = nop_X1(bundle);
 471                         break;
 472
 473                 case SPECIAL_0_OPCODE_X1:
 474                         switch (get_RRROpcodeExtension_X1(bundle)) {
 475                         /* jump-register */
 476                         case JALRP_SPECIAL_0_OPCODE_X1:
 477                         case JALR_SPECIAL_0_OPCODE_X1:
 478                                 state->update = 1;
 479                                 state->next_pc =
 480                                         regs->regs[get_SrcA_X1(bundle)];
 481                                 break;
 482
 483                         case JRP_SPECIAL_0_OPCODE_X1:
 484                         case JR_SPECIAL_0_OPCODE_X1:
 485                                 state->next_pc =
 486                                         regs->regs[get_SrcA_X1(bundle)];
 487                                 bundle = nop_X1(bundle);
 488                                 break;
 489
 490                         case LNK_SPECIAL_0_OPCODE_X1:
 491                                 state->update = 1;
 492                                 target_reg = get_Dest_X1(bundle);
 493                                 break;
 494
 495                         /* stores */
 496                         case SH_SPECIAL_0_OPCODE_X1:
 497                                 mem_op = MEMOP_STORE;
 498                                 size = 2;
 499                                 break;
 500
 501                         case SW_SPECIAL_0_OPCODE_X1:
 502                                 mem_op = MEMOP_STORE;
 503                                 size = 4;
 504                                 break;
 505                         }
 506                         break;
 507
 508                 /* loads and iret */
 509                 case SHUN_0_OPCODE_X1:
 510                         if (get_UnShOpcodeExtension_X1(bundle) ==
 511                             UN_0_SHUN_0_OPCODE_X1) {
 512                                 switch (get_UnOpcodeExtension_X1(bundle)) {
 513                                 case LH_UN_0_SHUN_0_OPCODE_X1:
 514                                         mem_op = MEMOP_LOAD;
 515                                         size = 2;
 516                                         sign_ext = 1;
 517                                         break;
 518
 519                                 case LH_U_UN_0_SHUN_0_OPCODE_X1:
 520                                         mem_op = MEMOP_LOAD;
 521                                         size = 2;
 522                                         sign_ext = 0;
 523                                         break;
 524
 525                                 case LW_UN_0_SHUN_0_OPCODE_X1:
 526                                         mem_op = MEMOP_LOAD;
 527                                         size = 4;
 528                                         break;
 529
 530                                 case IRET_UN_0_SHUN_0_OPCODE_X1:
 531                                 {
 532                                         unsigned long ex0_0 = __insn_mfspr(
 533                                                 SPR_EX_CONTEXT_0_0);
 534                                         unsigned long ex0_1 = __insn_mfspr(
 535                                                 SPR_EX_CONTEXT_0_1);
 536                                         /*
 537                                          * Special-case it if we're iret'ing
 538                                          * to PL0 again.  Otherwise just let
 539                                          * it run and it will generate SIGILL.
 540                                          */
 541                                         if (EX1_PL(ex0_1) == USER_PL) {
 542                                                 state->next_pc = ex0_0;
 543                                                 regs->ex1 = ex0_1;
 544                                                 bundle = nop_X1(bundle);
 545                                         }
 546                                 }
 547                                 }
 548                         }
 549                         break;
 550
 551                 /* postincrement operations */
 552                 case IMM_0_OPCODE_X1:
 553                         switch (get_ImmOpcodeExtension_X1(bundle)) {
 554                         case LWADD_IMM_0_OPCODE_X1:
 555                                 mem_op = MEMOP_LOAD_POSTINCR;
 556                                 size = 4;
 557                                 break;
 558
 559                         case LHADD_IMM_0_OPCODE_X1:
 560                                 mem_op = MEMOP_LOAD_POSTINCR;
 561                                 size = 2;
 562                                 sign_ext = 1;
 563                                 break;
 564
 565                         case LHADD_U_IMM_0_OPCODE_X1:
 566                                 mem_op = MEMOP_LOAD_POSTINCR;
 567                                 size = 2;
 568                                 sign_ext = 0;
 569                                 break;
 570
 571                         case SWADD_IMM_0_OPCODE_X1:
 572                                 mem_op = MEMOP_STORE_POSTINCR;
 573                                 size = 4;
 574                                 break;
 575
 576                         case SHADD_IMM_0_OPCODE_X1:
 577                                 mem_op = MEMOP_STORE_POSTINCR;
 578                                 size = 2;
 579                                 break;
 580
 581                         default:
 582                                 break;
 583                         }
 584                         break;
 585                 }
 586
 587                 if (state->update) {
 588                         /*
 589                          * Get an available register.  We start with a
 590                          * bitmask with 1's for available registers.
 591                          * We truncate to the low 32 registers since
 592                          * we are guaranteed to have set bits in the
 593                          * low 32 bits, then use ctz to pick the first.
 594                          */
 595                         u32 mask = (u32) ~((1ULL << get_Dest_X0(bundle)) |
 596                                            (1ULL << get_SrcA_X0(bundle)) |
 597                                            (1ULL << get_SrcB_X0(bundle)) |
 598                                            (1ULL << target_reg));
 599                         temp_reg = __builtin_ctz(mask);
 600                         state->update_reg = temp_reg;
 601                         state->update_value = regs->regs[temp_reg];
 602                         regs->regs[temp_reg] = (unsigned long) (pc+1);
 603                         regs->flags |= PT_FLAGS_RESTORE_REGS;
 604                         bundle = move_X1(bundle, target_reg, temp_reg);
 605                 }
 606         } else {
 607                 int opcode = get_Opcode_Y2(bundle);
 608
 609                 switch (opcode) {
 610                 /* loads */
 611                 case LH_OPCODE_Y2:
 612                         mem_op = MEMOP_LOAD;
 613                         size = 2;
 614                         sign_ext = 1;
 615                         break;
 616
 617                 case LH_U_OPCODE_Y2:
 618                         mem_op = MEMOP_LOAD;
 619                         size = 2;
 620                         sign_ext = 0;
 621                         break;
 622
 623                 case LW_OPCODE_Y2:
 624                         mem_op = MEMOP_LOAD;
 625                         size = 4;
 626                         break;
 627
 628                 /* stores */
 629                 case SH_OPCODE_Y2:
 630                         mem_op = MEMOP_STORE;
 631                         size = 2;
 632                         break;
 633
 634                 case SW_OPCODE_Y2:
 635                         mem_op = MEMOP_STORE;
 636                         size = 4;
 637                         break;
 638                 }
 639         }
 640
 641         /*
 642          * Check if we need to rewrite an unaligned load/store.
 643          * Returning zero is a special value meaning we generated a signal.
 644          */
 645         if (mem_op != MEMOP_NONE && align_ctl >= 0) {
 646                 bundle = rewrite_load_store_unaligned(state, bundle, regs,
 647                                                       mem_op, size, sign_ext);
 648                 if (bundle == 0)
 649                         return;
 650         }
 651
 652         /* write the bundle to our execution area */
 653         buffer = state->buffer;
 654         err = __put_user(bundle, buffer++);
 655
 656         /*
 657          * If we're really single-stepping, we take an INT_ILL after.
 658          * If we're just handling an unaligned access, we can just
 659          * jump directly back to where we were in user code.
 660          */
 661         if (is_single_step) {
 662                 err |= __put_user(__single_step_ill_insn, buffer++);
 663                 err |= __put_user(__single_step_ill_insn, buffer++);
 664         } else {
 665                 long delta;
 666
 667                 if (state->update) {
 668                         /* We have some state to update; do it inline */
 669                         int ha16;
 670                         bundle = __single_step_addli_insn;
 671                         bundle |= create_Dest_X1(state->update_reg);
 672                         bundle |= create_Imm16_X1(state->update_value);
 673                         err |= __put_user(bundle, buffer++);
 674                         bundle = __single_step_auli_insn;
 675                         bundle |= create_Dest_X1(state->update_reg);
 676                         bundle |= create_SrcA_X1(state->update_reg);
 677                         ha16 = (state->update_value + 0x8000) >> 16;
 678                         bundle |= create_Imm16_X1(ha16);
 679                         err |= __put_user(bundle, buffer++);
 680                         state->update = 0;
 681                 }
 682
 683                 /* End with a jump back to the next instruction */
 684                 delta = ((regs->pc + TILEPRO_BUNDLE_SIZE_IN_BYTES) -
 685                         (unsigned long)buffer) >>
 686                         TILEPRO_LOG2_BUNDLE_ALIGNMENT_IN_BYTES;
 687                 bundle = __single_step_j_insn;
 688                 bundle |= create_JOffLong_X1(delta);
 689                 err |= __put_user(bundle, buffer++);
 690         }
 691
 692         if (err) {
 693                 pr_err("Fault when writing to single-step buffer\n");
 694                 return;
 695         }
 696
 697         /*
 698          * Flush the buffer.
 699          * We do a local flush only, since this is a thread-specific buffer.
 700          */
 701         __flush_icache_range((unsigned long)state->buffer,
 702                              (unsigned long)buffer);
 703
 704         /* Indicate enabled */
 705         state->is_enabled = is_single_step;
 706         regs->pc = (unsigned long)state->buffer;
 707
 708         /* Fault immediately if we are coming back from a syscall. */
 709         if (regs->faultnum == INT_SWINT_1)
 710                 regs->pc += 8;
 711 }
 712
 713 #else
 714
 715 static DEFINE_PER_CPU(unsigned long, ss_saved_pc);
 716
 717
 718 /*
 719  * Called directly on the occasion of an interrupt.
 720  *
 721  * If the process doesn't have single step set, then we use this as an
 722  * opportunity to turn single step off.
 723  *
 724  * It has been mentioned that we could conditionally turn off single stepping
 725  * on each entry into the kernel and rely on single_step_once to turn it
 726  * on for the processes that matter (as we already do), but this
 727  * implementation is somewhat more efficient in that we muck with registers
 728  * once on a bum interrupt rather than on every entry into the kernel.
 729  *
 730  * If SINGLE_STEP_CONTROL_K has CANCELED set, then an interrupt occurred,
 731  * so we have to run through this process again before we can say that an
 732  * instruction has executed.
 733  *
 734  * swint will set CANCELED, but it's a legitimate instruction.  Fortunately
 735  * it changes the PC.  If it hasn't changed, then we know that the interrupt
 736  * wasn't generated by swint and we'll need to run this process again before
 737  * we can say an instruction has executed.
 738  *
 739  * If either CANCELED == 0 or the PC's changed, we send out SIGTRAPs and get
 740  * on with our lives.
 741  */
 742
 743 void gx_singlestep_handle(struct pt_regs *regs, int fault_num)
 744 {
 745         unsigned long *ss_pc = this_cpu_ptr(&ss_saved_pc);
 746         struct thread_info *info = (void *)current_thread_info();
 747         int is_single_step = test_ti_thread_flag(info, TIF_SINGLESTEP);
 748         unsigned long control = __insn_mfspr(SPR_SINGLE_STEP_CONTROL_K);
 749
 750         if (is_single_step == 0) {
 751                 __insn_mtspr(SPR_SINGLE_STEP_EN_K_K, 0);
 752
 753         } else if ((*ss_pc != regs->pc) ||
 754                    (!(control & SPR_SINGLE_STEP_CONTROL_1__CANCELED_MASK))) {
 755
 756                 control |= SPR_SINGLE_STEP_CONTROL_1__CANCELED_MASK;
 757                 control |= SPR_SINGLE_STEP_CONTROL_1__INHIBIT_MASK;
 758                 __insn_mtspr(SPR_SINGLE_STEP_CONTROL_K, control);
 759                 send_sigtrap(current, regs);
 760         }
 761 }
 762
 763
 764 /*
 765  * Called from need_singlestep.  Set up the control registers and the enable
 766  * register, then return back.
 767  */
 768
 769 void single_step_once(struct pt_regs *regs)
 770 {
 771         unsigned long *ss_pc = this_cpu_ptr(&ss_saved_pc);
 772         unsigned long control = __insn_mfspr(SPR_SINGLE_STEP_CONTROL_K);
 773
 774         *ss_pc = regs->pc;
 775         control |= SPR_SINGLE_STEP_CONTROL_1__CANCELED_MASK;
 776         control |= SPR_SINGLE_STEP_CONTROL_1__INHIBIT_MASK;
 777         __insn_mtspr(SPR_SINGLE_STEP_CONTROL_K, control);
 778         __insn_mtspr(SPR_SINGLE_STEP_EN_K_K, 1 << USER_PL);
 779 }
 780
 781 void single_step_execve(void)
 782 {
 783         /* Nothing */
 784 }
 785
 786 #endif /* !__tilegx__ */