bsd/dev/arm64/fbt_arm.c

   1 /*
   2  * Copyright (c) 2007-2018 Apple Inc. All rights reserved.
   3  */
   4 /*
   5  * CDDL HEADER START
   6  *
   7  * The contents of this file are subject to the terms of the
   8  * Common Development and Distribution License, Version 1.0 only
   9  * (the "License").  You may not use this file except in compliance
  10  * with the License.
  11  *
  12  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
  13  * or http://www.opensolaris.org/os/licensing.
  14  * See the License for the specific language governing permissions
  15  * and limitations under the License.
  16  *
  17  * When distributing Covered Code, include this CDDL HEADER in each
  18  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
  19  * If applicable, add the following below this CDDL HEADER, with the
  20  * fields enclosed by brackets "[]" replaced with your own identifying
  21  * information: Portions Copyright [yyyy] [name of copyright owner]
  22  *
  23  * CDDL HEADER END
  24  */
  25 /*
  26  * Copyright 2005 Sun Microsystems, Inc.  All rights reserved.
  27  * Use is subject to license terms.
  28  */
  29
  30 #include <kern/thread.h>
  31 #include <mach/thread_status.h>
  32 #include <arm/proc_reg.h>
  33 #include <arm/caches_internal.h>
  34
  35 #include <mach-o/loader.h>
  36 #include <mach-o/nlist.h>
  37 #include <libkern/kernel_mach_header.h>
  38
  39 #include <sys/param.h>
  40 #include <sys/systm.h>
  41 #include <sys/errno.h>
  42 #include <sys/stat.h>
  43 #include <sys/ioctl.h>
  44 #include <sys/conf.h>
  45 #include <sys/fcntl.h>
  46 #include <miscfs/devfs/devfs.h>
  47
  48 #include <sys/dtrace.h>
  49 #include <sys/dtrace_impl.h>
  50 #include <sys/fbt.h>
  51
  52 #include <sys/dtrace_glue.h>
  53
  54 #if __has_include(<ptrauth.h>)
  55 #include <ptrauth.h>
  56 #endif
  57
  58 #define DTRACE_INVOP_PUSH_FRAME 11
  59
  60 #define DTRACE_INVOP_NOP_SKIP           4
  61 #define DTRACE_INVOP_ADD_FP_SP_SKIP     4
  62
  63 #define DTRACE_INVOP_POP_PC_SKIP 2
  64
  65 /*
  66  * stp  fp, lr, [sp, #val]
  67  * stp  fp, lr, [sp, #val]!
  68  */
  69 #define FBT_IS_ARM64_FRAME_PUSH(x)      \
  70         (((x) & 0xffc07fff) == 0xa9007bfd || ((x) & 0xffc07fff) == 0xa9807bfd)
  71
  72 /*
  73  * stp  Xt1, Xt2, [sp, #val]
  74  * stp  Xt1, Xt2, [sp, #val]!
  75  */
  76 #define FBT_IS_ARM64_PUSH(x)            \
  77         (((x) & 0xffc003e0) == 0xa90003e0 || ((x) & 0xffc003e0) == 0xa98003e0)
  78
  79 /*
  80  * ldp  fp, lr, [sp,  #val]
  81  * ldp  fp, lr, [sp], #val
  82  */
  83 #define FBT_IS_ARM64_FRAME_POP(x)       \
  84         (((x) & 0xffc07fff) == 0xa9407bfd || ((x) & 0xffc07fff) == 0xa8c07bfd)
  85
  86 #define FBT_IS_ARM64_ADD_FP_SP(x)       (((x) & 0xffc003ff) == 0x910003fd)      /* add fp, sp, #val  (add fp, sp, #0 == mov fp, sp) */
  87 #define FBT_IS_ARM64_RET(x)             (((x) == 0xd65f03c0) || ((x) == 0xd65f0fff))                    /* ret, retab */
  88
  89
  90 #define FBT_B_MASK                      0xff000000
  91 #define FBT_B_IMM_MASK                  0x00ffffff
  92 #define FBT_B_INSTR                     0x14000000
  93
  94 #define FBT_IS_ARM64_B_INSTR(x)         ((x & FBT_B_MASK) == FBT_B_INSTR)
  95 #define FBT_GET_ARM64_B_IMM(x)          ((x & FBT_B_IMM_MASK) << 2)
  96
  97 #define FBT_PATCHVAL                    0xe7eeee7e
  98 #define FBT_AFRAMES_ENTRY               7
  99 #define FBT_AFRAMES_RETURN              7
 100
 101 #define FBT_ENTRY       "entry"
 102 #define FBT_RETURN      "return"
 103 #define FBT_ADDR2NDX(addr)      ((((uintptr_t)(addr)) >> 4) & fbt_probetab_mask)
 104
 105 extern dtrace_provider_id_t     fbt_id;
 106 extern fbt_probe_t               **fbt_probetab;
 107 extern int                      fbt_probetab_mask;
 108
 109 kern_return_t fbt_perfCallback(int, struct arm_saved_state *, __unused int, __unused int);
 110
 111 int
 112 fbt_invop(uintptr_t addr, uintptr_t * stack, uintptr_t rval)
 113 {
 114         fbt_probe_t    *fbt = fbt_probetab[FBT_ADDR2NDX(addr)];
 115
 116         for (; fbt != NULL; fbt = fbt->fbtp_hashnext) {
 117                 if ((uintptr_t) fbt->fbtp_patchpoint == addr) {
 118                         if (0 == CPU->cpu_dtrace_invop_underway) {
 119                                 CPU->cpu_dtrace_invop_underway = 1;     /* Race not possible on
 120                                                                         * this per-cpu state */
 121
 122                                 /*
 123                                  * Stack looks like this:
 124                                  *
 125                                  *      [Higher addresses]
 126                                  *
 127                                  *      Frame of caller
 128                                  *      Extra args for callee
 129                                  *      ------------------------
 130                                  *      fbt entry probe:
 131                                  *          Frame from traced function: <previous sp (e.g. 0x1000), return address>
 132                                  *      fbt return probe:
 133                                  *          Missing as the return probe has already popped the frame in the callee and
 134                                  *          traps with LR set to the return address in caller.
 135                                  *      ------------------------
 136                                  *      arm_context_t
 137                                  *      ------------------------
 138                                  *      Frame from trap handler:  <previous sp (e.g. 0x1000) , traced PC >
 139                                  *                                The traced function has either never pushed the frame
 140                                  *                                or already popped it.  So there is no frame in the
 141                                  *                                backtrace pointing to the frame on the stack containing
 142                                  *                                the LR in the caller.
 143                                  *     ------------------------
 144                                  *          |
 145                                  *          |
 146                                  *          |  stack grows this way
 147                                  *          |
 148                                  *          |
 149                                  *          v
 150                                  *     [Lower addresses]
 151                                  *
 152                                  * cpu_dtrace_caller compensates for fact that the LR is not stored on stack as explained
 153                                  * above.  When walking the stack, when we reach the frame where we extract a PC in the
 154                                  * patched function, we put the cpu_dtrace_caller in the backtrace instead.  The next
 155                                  * frame we extract will be in the caller's caller, so we output a backtrace starting
 156                                  * at the caller and going sequentially up the stack.
 157                                  */
 158                                 arm_saved_state_t *regs = (arm_saved_state_t *)(&((arm_context_t *)stack)->ss);
 159
 160                                 CPU->cpu_dtrace_caller = get_saved_state_lr(regs);
 161
 162                                 /* When fbt_roffset is non-zero, we know we are handling a return probe point. */
 163                                 if (fbt->fbtp_roffset == 0) {
 164                                         dtrace_probe(fbt->fbtp_id, get_saved_state_reg(regs, 0), get_saved_state_reg(regs, 1),
 165                                             get_saved_state_reg(regs, 2), get_saved_state_reg(regs, 3), get_saved_state_reg(regs, 4));
 166                                 } else {
 167                                         dtrace_probe(fbt->fbtp_id, fbt->fbtp_roffset, rval, 0, 0, 0);
 168                                 }
 169
 170                                 CPU->cpu_dtrace_caller = 0;
 171                                 CPU->cpu_dtrace_invop_underway = 0;
 172                         }
 173
 174                         /*
 175                          *       On other architectures, we return a DTRACE constant to let the callback function
 176                          *       know what was replaced. On the ARM, since the function prologue/epilogue machine code
 177                          *       can vary, we need the actual bytes of the instruction, so return the savedval instead.
 178                          */
 179                         return fbt->fbtp_savedval;
 180                 }
 181         }
 182
 183         return 0;
 184 }
 185
 186 #define IS_USER_TRAP(regs)   (PSR64_IS_USER(get_saved_state_cpsr(regs)))
 187 #define T_INVALID_OPCODE EXC_BAD_INSTRUCTION
 188 #define FBT_EXCEPTION_CODE T_INVALID_OPCODE
 189
 190 kern_return_t
 191 fbt_perfCallback(
 192         int trapno,
 193         struct arm_saved_state * regs,
 194         __unused int unused1,
 195         __unused int unused2)
 196 {
 197         kern_return_t   retval = KERN_FAILURE;
 198
 199         if (FBT_EXCEPTION_CODE == trapno && !IS_USER_TRAP(regs)) {
 200                 boolean_t oldlevel = 0;
 201                 machine_inst_t emul = 0;
 202                 uint64_t sp, lr;
 203                 uint32_t imm;
 204
 205                 oldlevel = ml_set_interrupts_enabled(FALSE);
 206
 207                 __asm__ volatile (
 208                          "Ldtrace_invop_callsite_pre_label:\n"
 209                          ".data\n"
 210                          ".private_extern _dtrace_invop_callsite_pre\n"
 211                          "_dtrace_invop_callsite_pre:\n"
 212                          "  .quad Ldtrace_invop_callsite_pre_label\n"
 213                          ".text\n"
 214                 );
 215
 216                 emul = dtrace_invop(get_saved_state_pc(regs), (uintptr_t*) regs, get_saved_state_reg(regs, 0));
 217
 218                 __asm__ volatile (
 219                          "Ldtrace_invop_callsite_post_label:\n"
 220                          ".data\n"
 221                          ".private_extern _dtrace_invop_callsite_post\n"
 222                          "_dtrace_invop_callsite_post:\n"
 223                          "  .quad Ldtrace_invop_callsite_post_label\n"
 224                          ".text\n"
 225                 );
 226
 227                 if (emul == DTRACE_INVOP_NOP) {
 228                         /*
 229                          * Skip over the patched NOP planted by sdt
 230                          */
 231                         add_saved_state_pc(regs, DTRACE_INVOP_NOP_SKIP);
 232                         retval = KERN_SUCCESS;
 233                 } else if (FBT_IS_ARM64_ADD_FP_SP(emul)) {
 234                         /* retrieve the value to add */
 235                         uint64_t val = (emul >> 10) & 0xfff;
 236                         assert(val < 4096);
 237
 238                         /* retrieve sp */
 239                         sp = get_saved_state_sp(regs);
 240
 241                         /*
 242                          * emulate the instruction:
 243                          *      add     fp, sp, #val
 244                          */
 245                         assert(sp < (UINT64_MAX - val));
 246                         set_saved_state_fp(regs, sp + val);
 247
 248                         /* skip over the bytes of the patched instruction */
 249                         add_saved_state_pc(regs, DTRACE_INVOP_ADD_FP_SP_SKIP);
 250
 251                         retval = KERN_SUCCESS;
 252                 } else if (FBT_IS_ARM64_RET(emul)) {
 253                         lr = get_saved_state_lr(regs);
 254 #if __has_feature(ptrauth_calls)
 255                         lr = (user_addr_t) ptrauth_strip((void *)lr, ptrauth_key_return_address);
 256 #endif
 257                         set_saved_state_pc(regs, lr);
 258                         retval = KERN_SUCCESS;
 259                 } else if (FBT_IS_ARM64_B_INSTR(emul)) {
 260                         imm = FBT_GET_ARM64_B_IMM(emul);
 261                         add_saved_state_pc(regs, imm);
 262                         retval = KERN_SUCCESS;
 263                 } else if (emul == FBT_PATCHVAL) {
 264                         /* Means we encountered an error but handled it, try same inst again */
 265                         retval = KERN_SUCCESS;
 266                 } else {
 267                         retval = KERN_FAILURE;
 268                 }
 269
 270                 ml_set_interrupts_enabled(oldlevel);
 271         }
 272
 273         return retval;
 274 }
 275
 276 void
 277 fbt_provide_probe(struct modctl *ctl, const char *modname, const char* symbolName, machine_inst_t* symbolStart, machine_inst_t *instrHigh)
 278 {
 279         int             doenable = 0;
 280         dtrace_id_t     thisid;
 281
 282         fbt_probe_t     *newfbt, *retfbt, *entryfbt;
 283         machine_inst_t *instr, *pushinstr = NULL, *limit, theInstr;
 284         int             foundPushLR, savedRegs;
 285
 286         /*
 287          * Guard against null and invalid symbols
 288          */
 289         if (!symbolStart || !instrHigh || instrHigh < symbolStart) {
 290                 kprintf("dtrace: %s has an invalid address\n", symbolName);
 291                 return;
 292         }
 293
 294         /*
 295          * Assume the compiler doesn't schedule instructions in the prologue.
 296          */
 297         foundPushLR = 0;
 298         savedRegs = -1;
 299         limit = (machine_inst_t *)instrHigh;
 300
 301         assert(sizeof(*instr) == 4);
 302
 303         for (instr = symbolStart, theInstr = 0; instr < instrHigh; instr++) {
 304                 /*
 305                  * Count the number of time we pushed something onto the stack
 306                  * before hitting a frame push. That will give us an estimation
 307                  * of how many stack pops we should expect when looking for the
 308                  * RET instruction.
 309                  */
 310                 theInstr = *instr;
 311                 if (FBT_IS_ARM64_FRAME_PUSH(theInstr)) {
 312                         foundPushLR = 1;
 313                         pushinstr = instr;
 314                 }
 315
 316                 if (foundPushLR && (FBT_IS_ARM64_ADD_FP_SP(theInstr))) {
 317                         /* Guard against a random setting of fp from sp, we make sure we found the push first */
 318                         break;
 319                 }
 320                 if (FBT_IS_ARM64_RET(theInstr)) { /* We've gone too far, bail. */
 321                         break;
 322                 }
 323                 if (FBT_IS_ARM64_FRAME_POP(theInstr)) { /* We've gone too far, bail. */
 324                         break;
 325                 }
 326         }
 327
 328         if (!(foundPushLR && (FBT_IS_ARM64_ADD_FP_SP(theInstr)))) {
 329                 return;
 330         }
 331
 332         thisid = dtrace_probe_lookup(fbt_id, modname, symbolName, FBT_ENTRY);
 333         newfbt = kmem_zalloc(sizeof(fbt_probe_t), KM_SLEEP);
 334         newfbt->fbtp_next = NULL;
 335         strlcpy((char *)&(newfbt->fbtp_name), symbolName, MAX_FBTP_NAME_CHARS );
 336
 337         if (thisid != 0) {
 338                 /*
 339                  * The dtrace_probe previously existed, so we have to hook
 340                  * the newfbt entry onto the end of the existing fbt's
 341                  * chain.
 342                  * If we find an fbt entry that was previously patched to
 343                  * fire, (as indicated by the current patched value), then
 344                  * we want to enable this newfbt on the spot.
 345                  */
 346                 entryfbt = dtrace_probe_arg(fbt_id, thisid);
 347                 ASSERT(entryfbt != NULL);
 348                 for (; entryfbt != NULL; entryfbt = entryfbt->fbtp_next) {
 349                         if (entryfbt->fbtp_currentval == entryfbt->fbtp_patchval) {
 350                                 doenable++;
 351                         }
 352
 353                         if (entryfbt->fbtp_next == NULL) {
 354                                 entryfbt->fbtp_next = newfbt;
 355                                 newfbt->fbtp_id = entryfbt->fbtp_id;
 356                                 break;
 357                         }
 358                 }
 359         } else {
 360                 /*
 361                  * The dtrace_probe did not previously exist, so we
 362                  * create it and hook in the newfbt.  Since the probe is
 363                  * new, we obviously do not need to enable it on the spot.
 364                  */
 365                 newfbt->fbtp_id = dtrace_probe_create(fbt_id, modname, symbolName, FBT_ENTRY, FBT_AFRAMES_ENTRY, newfbt);
 366                 doenable = 0;
 367         }
 368
 369         newfbt->fbtp_patchpoint = instr;
 370         newfbt->fbtp_ctl = ctl;
 371         newfbt->fbtp_loadcnt = ctl->mod_loadcnt;
 372         newfbt->fbtp_rval = DTRACE_INVOP_PUSH_FRAME;
 373         newfbt->fbtp_savedval = theInstr;
 374         newfbt->fbtp_patchval = FBT_PATCHVAL;
 375         newfbt->fbtp_currentval = 0;
 376         newfbt->fbtp_hashnext = fbt_probetab[FBT_ADDR2NDX(instr)];
 377         fbt_probetab[FBT_ADDR2NDX(instr)] = newfbt;
 378
 379         if (doenable) {
 380                 fbt_enable(NULL, newfbt->fbtp_id, newfbt);
 381         }
 382
 383         /*
 384          * The fbt entry chain is in place, one entry point per symbol.
 385          * The fbt return chain can have multiple return points per
 386          * symbol.
 387          * Here we find the end of the fbt return chain.
 388          */
 389
 390         doenable = 0;
 391
 392         thisid = dtrace_probe_lookup(fbt_id, modname, symbolName, FBT_RETURN);
 393
 394         if (thisid != 0) {
 395                 /* The dtrace_probe previously existed, so we have to
 396                  * find the end of the existing fbt chain.  If we find
 397                  * an fbt return that was previously patched to fire,
 398                  * (as indicated by the currrent patched value), then
 399                  * we want to enable any new fbts on the spot.
 400                  */
 401                 retfbt = dtrace_probe_arg(fbt_id, thisid);
 402                 ASSERT(retfbt != NULL);
 403                 for (; retfbt != NULL; retfbt =  retfbt->fbtp_next) {
 404                         if (retfbt->fbtp_currentval == retfbt->fbtp_patchval) {
 405                                 doenable++;
 406                         }
 407                         if (retfbt->fbtp_next == NULL) {
 408                                 break;
 409                         }
 410                 }
 411         } else {
 412                 doenable = 0;
 413                 retfbt = NULL;
 414         }
 415
 416         /*
 417          * Go back to the start of the function, in case
 418          * the compiler emitted pcrel data loads
 419          * before FP was adjusted.
 420          */
 421         instr = pushinstr + 1;
 422 again:
 423         if (instr >= limit) {
 424                 return;
 425         }
 426
 427         /* XXX FIXME ... extra jump table detection? */
 428
 429         /*
 430          * OK, it's an instruction.
 431          */
 432         theInstr = *instr;
 433
 434         /* Walked onto the start of the next routine? If so, bail out from this function */
 435         if (FBT_IS_ARM64_FRAME_PUSH(theInstr)) {
 436                 if (!retfbt) {
 437                         kprintf("dtrace: fbt: No return probe for %s, walked to next routine at 0x%016llx\n", symbolName, (uint64_t)instr);
 438                 }
 439                 return;
 440         }
 441
 442         /* XXX fancy detection of end of function using PC-relative loads */
 443
 444         /*
 445          * Look for:
 446          *      ldp fp, lr, [sp], #val
 447          *      ldp fp, lr, [sp,  #val]
 448          */
 449         if (!FBT_IS_ARM64_FRAME_POP(theInstr)) {
 450                 instr++;
 451                 goto again;
 452         }
 453
 454         /* go to the next instruction */
 455         instr++;
 456
 457         /* Scan ahead for a ret or a branch outside the function */
 458         for (; instr < limit; instr++) {
 459                 theInstr = *instr;
 460                 if (FBT_IS_ARM64_RET(theInstr)) {
 461                         break;
 462                 }
 463                 if (FBT_IS_ARM64_B_INSTR(theInstr)) {
 464                         machine_inst_t *dest = instr + FBT_GET_ARM64_B_IMM(theInstr);
 465                         /*
 466                          * Check whether the destination of the branch
 467                          * is outside of the function
 468                          */
 469                         if (dest >= limit || dest < symbolStart) {
 470                                 break;
 471                         }
 472                 }
 473         }
 474
 475         if (!FBT_IS_ARM64_RET(theInstr) && !FBT_IS_ARM64_B_INSTR(theInstr)) {
 476                 return;
 477         }
 478
 479         newfbt = kmem_zalloc(sizeof(fbt_probe_t), KM_SLEEP);
 480         newfbt->fbtp_next = NULL;
 481         strlcpy((char *)&(newfbt->fbtp_name), symbolName, MAX_FBTP_NAME_CHARS );
 482
 483         if (retfbt == NULL) {
 484                 newfbt->fbtp_id = dtrace_probe_create(fbt_id, modname,
 485                     symbolName, FBT_RETURN, FBT_AFRAMES_RETURN, newfbt);
 486         } else {
 487                 retfbt->fbtp_next = newfbt;
 488                 newfbt->fbtp_id = retfbt->fbtp_id;
 489         }
 490
 491         retfbt = newfbt;
 492         newfbt->fbtp_patchpoint = instr;
 493         newfbt->fbtp_ctl = ctl;
 494         newfbt->fbtp_loadcnt = ctl->mod_loadcnt;
 495
 496         ASSERT(FBT_IS_ARM64_RET(theInstr) || FBT_IS_ARM64_B_INSTR(theInstr));
 497         newfbt->fbtp_rval = DTRACE_INVOP_RET;
 498         newfbt->fbtp_roffset = (uintptr_t) ((uint8_t*) instr - (uint8_t *)symbolStart);
 499         newfbt->fbtp_savedval = theInstr;
 500         newfbt->fbtp_patchval = FBT_PATCHVAL;
 501         newfbt->fbtp_currentval = 0;
 502         newfbt->fbtp_hashnext = fbt_probetab[FBT_ADDR2NDX(instr)];
 503         fbt_probetab[FBT_ADDR2NDX(instr)] = newfbt;
 504
 505         if (doenable) {
 506                 fbt_enable(NULL, newfbt->fbtp_id, newfbt);
 507         }
 508
 509         instr++;
 510         goto again;
 511 }