1 2 /*--------------------------------------------------------------------*/ 3 /*--- Handle system calls. syswrap-main.c ---*/ 4 /*--------------------------------------------------------------------*/ 5 6 /* 7 This file is part of Valgrind, a dynamic binary instrumentation 8 framework. 9 10 Copyright (C) 2000-2011 Julian Seward 11 jseward (at) acm.org 12 13 This program is free software; you can redistribute it and/or 14 modify it under the terms of the GNU General Public License as 15 published by the Free Software Foundation; either version 2 of the 16 License, or (at your option) any later version. 17 18 This program is distributed in the hope that it will be useful, but 19 WITHOUT ANY WARRANTY; without even the implied warranty of 20 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU 21 General Public License for more details. 22 23 You should have received a copy of the GNU General Public License 24 along with this program; if not, write to the Free Software 25 Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 26 02111-1307, USA. 27 28 The GNU General Public License is contained in the file COPYING. 29 */ 30 31 #include "libvex_guest_offsets.h" 32 #include "libvex_trc_values.h" 33 #include "pub_core_basics.h" 34 #include "pub_core_aspacemgr.h" 35 #include "pub_core_vki.h" 36 #include "pub_core_vkiscnums.h" 37 #include "pub_core_libcsetjmp.h" // to keep _threadstate.h happy 38 #include "pub_core_threadstate.h" 39 #include "pub_core_libcbase.h" 40 #include "pub_core_libcassert.h" 41 #include "pub_core_libcprint.h" 42 #include "pub_core_libcproc.h" // For VG_(getpid)() 43 #include "pub_core_libcsignal.h" 44 #include "pub_core_scheduler.h" // For VG_({acquire,release}_BigLock), 45 // and VG_(vg_yield) 46 #include "pub_core_stacktrace.h" // For VG_(get_and_pp_StackTrace)() 47 #include "pub_core_tooliface.h" 48 #include "pub_core_options.h" 49 #include "pub_core_signals.h" // For VG_SIGVGKILL, VG_(poll_signals) 50 #include "pub_core_syscall.h" 51 #include "pub_core_machine.h" 52 #include "pub_core_syswrap.h" 53 54 #include "priv_types_n_macros.h" 55 #include "priv_syswrap-main.h" 56 57 #if defined(VGO_darwin) 58 #include "priv_syswrap-darwin.h" 59 #endif 60 61 /* Useful info which needs to be recorded somewhere: 62 Use of registers in syscalls is: 63 64 NUM ARG1 ARG2 ARG3 ARG4 ARG5 ARG6 ARG7 ARG8 RESULT 65 LINUX: 66 x86 eax ebx ecx edx esi edi ebp n/a n/a eax (== NUM) 67 amd64 rax rdi rsi rdx r10 r8 r9 n/a n/a rax (== NUM) 68 ppc32 r0 r3 r4 r5 r6 r7 r8 n/a n/a r3+CR0.SO (== ARG1) 69 ppc64 r0 r3 r4 r5 r6 r7 r8 n/a n/a r3+CR0.SO (== ARG1) 70 arm r7 r0 r1 r2 r3 r4 r5 n/a n/a r0 (== ARG1) 71 72 On s390x the svc instruction is used for system calls. The system call 73 number is encoded in the instruction (8 bit immediate field). Since Linux 74 2.6 it is also allowed to use svc 0 with the system call number in r1. 75 This was introduced for system calls >255, but works for all. It is 76 also possible to see the svc 0 together with an EXecute instruction, that 77 fills in the immediate field. 78 s390x r1/SVC r2 r3 r4 r5 r6 r7 n/a n/a r2 (== ARG1) 79 80 DARWIN: 81 x86 eax +4 +8 +12 +16 +20 +24 +28 +32 edx:eax, eflags.c 82 amd64 rax rdi rsi rdx rcx r8 r9 +8 +16 rdx:rax, rflags.c 83 84 For x86-darwin, "+N" denotes "in memory at N(%esp)"; ditto 85 amd64-darwin. Apparently 0(%esp) is some kind of return address 86 (perhaps for syscalls done with "sysenter"?) I don't think it is 87 relevant for syscalls done with "int $0x80/1/2". 88 */ 89 90 /* This is the top level of the system-call handler module. All 91 system calls are channelled through here, doing two things: 92 93 * notify the tool of the events (mem/reg reads, writes) happening 94 95 * perform the syscall, usually by passing it along to the kernel 96 unmodified. 97 98 A magical piece of assembly code, do_syscall_for_client_WRK, in 99 syscall-$PLATFORM.S does the tricky bit of passing a syscall to the 100 kernel, whilst having the simulator retain control. 101 */ 102 103 /* The main function is VG_(client_syscall). The simulation calls it 104 whenever a client thread wants to do a syscall. The following is a 105 sketch of what it does. 106 107 * Ensures the root thread's stack is suitably mapped. Tedious and 108 arcane. See big big comment in VG_(client_syscall). 109 110 * First, it rounds up the syscall number and args (which is a 111 platform dependent activity) and puts them in a struct ("args") 112 and also a copy in "orig_args". 113 114 The pre/post wrappers refer to these structs and so no longer 115 need magic macros to access any specific registers. This struct 116 is stored in thread-specific storage. 117 118 119 * The pre-wrapper is called, passing it a pointer to struct 120 "args". 121 122 123 * The pre-wrapper examines the args and pokes the tool 124 appropriately. It may modify the args; this is why "orig_args" 125 is also stored. 126 127 The pre-wrapper may choose to 'do' the syscall itself, and 128 concludes one of three outcomes: 129 130 Success(N) -- syscall is already complete, with success; 131 result is N 132 133 Fail(N) -- syscall is already complete, with failure; 134 error code is N 135 136 HandToKernel -- (the usual case): this needs to be given to 137 the kernel to be done, using the values in 138 the possibly-modified "args" struct. 139 140 In addition, the pre-wrapper may set some flags: 141 142 MayBlock -- only applicable when outcome==HandToKernel 143 144 PostOnFail -- only applicable when outcome==HandToKernel or Fail 145 146 147 * If the pre-outcome is HandToKernel, the syscall is duly handed 148 off to the kernel (perhaps involving some thread switchery, but 149 that's not important). This reduces the possible set of outcomes 150 to either Success(N) or Fail(N). 151 152 153 * The outcome (Success(N) or Fail(N)) is written back to the guest 154 register(s). This is platform specific: 155 156 x86: Success(N) ==> eax = N 157 Fail(N) ==> eax = -N 158 159 ditto amd64 160 161 ppc32: Success(N) ==> r3 = N, CR0.SO = 0 162 Fail(N) ==> r3 = N, CR0.SO = 1 163 164 Darwin: 165 x86: Success(N) ==> edx:eax = N, cc = 0 166 Fail(N) ==> edx:eax = N, cc = 1 167 168 s390x: Success(N) ==> r2 = N 169 Fail(N) ==> r2 = -N 170 171 * The post wrapper is called if: 172 173 - it exists, and 174 - outcome==Success or (outcome==Fail and PostOnFail is set) 175 176 The post wrapper is passed the adulterated syscall args (struct 177 "args"), and the syscall outcome (viz, Success(N) or Fail(N)). 178 179 There are several other complications, primarily to do with 180 syscalls getting interrupted, explained in comments in the code. 181 */ 182 183 /* CAVEATS for writing wrappers. It is important to follow these! 184 185 The macros defined in priv_types_n_macros.h are designed to help 186 decouple the wrapper logic from the actual representation of 187 syscall args/results, since these wrappers are designed to work on 188 multiple platforms. 189 190 Sometimes a PRE wrapper will complete the syscall itself, without 191 handing it to the kernel. It will use one of SET_STATUS_Success, 192 SET_STATUS_Failure or SET_STATUS_from_SysRes to set the return 193 value. It is critical to appreciate that use of the macro does not 194 immediately cause the underlying guest state to be updated -- that 195 is done by the driver logic in this file, when the wrapper returns. 196 197 As a result, PRE wrappers of the following form will malfunction: 198 199 PRE(fooble) 200 { 201 ... do stuff ... 202 SET_STATUS_Somehow(...) 203 204 // do something that assumes guest state is up to date 205 } 206 207 In particular, direct or indirect calls to VG_(poll_signals) after 208 setting STATUS can cause the guest state to be read (in order to 209 build signal frames). Do not do this. If you want a signal poll 210 after the syscall goes through, do "*flags |= SfPollAfter" and the 211 driver logic will do it for you. 212 213 ----------- 214 215 Another critical requirement following introduction of new address 216 space manager (JRS, 20050923): 217 218 In a situation where the mappedness of memory has changed, aspacem 219 should be notified BEFORE the tool. Hence the following is 220 correct: 221 222 Bool d = VG_(am_notify_munmap)(s->start, s->end+1 - s->start); 223 VG_TRACK( die_mem_munmap, s->start, s->end+1 - s->start ); 224 if (d) 225 VG_(discard_translations)(s->start, s->end+1 - s->start); 226 227 whilst this is wrong: 228 229 VG_TRACK( die_mem_munmap, s->start, s->end+1 - s->start ); 230 Bool d = VG_(am_notify_munmap)(s->start, s->end+1 - s->start); 231 if (d) 232 VG_(discard_translations)(s->start, s->end+1 - s->start); 233 234 The reason is that the tool may itself ask aspacem for more shadow 235 memory as a result of the VG_TRACK call. In such a situation it is 236 critical that aspacem's segment array is up to date -- hence the 237 need to notify aspacem first. 238 239 ----------- 240 241 Also .. take care to call VG_(discard_translations) whenever 242 memory with execute permissions is unmapped. 243 */ 244 245 246 /* --------------------------------------------------------------------- 247 Do potentially blocking syscall for the client, and mess with 248 signal masks at the same time. 249 ------------------------------------------------------------------ */ 250 251 /* Perform a syscall on behalf of a client thread, using a specific 252 signal mask. On completion, the signal mask is set to restore_mask 253 (which presumably blocks almost everything). If a signal happens 254 during the syscall, the handler should call 255 VG_(fixup_guest_state_after_syscall_interrupted) to adjust the 256 thread's context to do the right thing. 257 258 The _WRK function is handwritten assembly, implemented per-platform 259 in coregrind/m_syswrap/syscall-$PLAT.S. It has some very magic 260 properties. See comments at the top of 261 VG_(fixup_guest_state_after_syscall_interrupted) below for details. 262 263 This function (these functions) are required to return zero in case 264 of success (even if the syscall itself failed), and nonzero if the 265 sigprocmask-swizzling calls failed. We don't actually care about 266 the failure values from sigprocmask, although most of the assembly 267 implementations do attempt to return that, using the convention 268 0 for success, or 0x8000 | error-code for failure. 269 */ 270 #if defined(VGO_linux) 271 extern 272 UWord ML_(do_syscall_for_client_WRK)( Word syscallno, 273 void* guest_state, 274 const vki_sigset_t *syscall_mask, 275 const vki_sigset_t *restore_mask, 276 Word sigsetSzB ); 277 #elif defined(VGO_darwin) 278 extern 279 UWord ML_(do_syscall_for_client_unix_WRK)( Word syscallno, 280 void* guest_state, 281 const vki_sigset_t *syscall_mask, 282 const vki_sigset_t *restore_mask, 283 Word sigsetSzB ); /* unused */ 284 extern 285 UWord ML_(do_syscall_for_client_mach_WRK)( Word syscallno, 286 void* guest_state, 287 const vki_sigset_t *syscall_mask, 288 const vki_sigset_t *restore_mask, 289 Word sigsetSzB ); /* unused */ 290 extern 291 UWord ML_(do_syscall_for_client_mdep_WRK)( Word syscallno, 292 void* guest_state, 293 const vki_sigset_t *syscall_mask, 294 const vki_sigset_t *restore_mask, 295 Word sigsetSzB ); /* unused */ 296 #else 297 # error "Unknown OS" 298 #endif 299 300 301 static 302 void do_syscall_for_client ( Int syscallno, 303 ThreadState* tst, 304 const vki_sigset_t* syscall_mask ) 305 { 306 vki_sigset_t saved; 307 UWord err; 308 # if defined(VGO_linux) 309 err = ML_(do_syscall_for_client_WRK)( 310 syscallno, &tst->arch.vex, 311 syscall_mask, &saved, sizeof(vki_sigset_t) 312 ); 313 # elif defined(VGO_darwin) 314 switch (VG_DARWIN_SYSNO_CLASS(syscallno)) { 315 case VG_DARWIN_SYSCALL_CLASS_UNIX: 316 err = ML_(do_syscall_for_client_unix_WRK)( 317 VG_DARWIN_SYSNO_FOR_KERNEL(syscallno), &tst->arch.vex, 318 syscall_mask, &saved, 0/*unused:sigsetSzB*/ 319 ); 320 break; 321 case VG_DARWIN_SYSCALL_CLASS_MACH: 322 err = ML_(do_syscall_for_client_mach_WRK)( 323 VG_DARWIN_SYSNO_FOR_KERNEL(syscallno), &tst->arch.vex, 324 syscall_mask, &saved, 0/*unused:sigsetSzB*/ 325 ); 326 break; 327 case VG_DARWIN_SYSCALL_CLASS_MDEP: 328 err = ML_(do_syscall_for_client_mdep_WRK)( 329 VG_DARWIN_SYSNO_FOR_KERNEL(syscallno), &tst->arch.vex, 330 syscall_mask, &saved, 0/*unused:sigsetSzB*/ 331 ); 332 break; 333 default: 334 vg_assert(0); 335 /*NOTREACHED*/ 336 break; 337 } 338 # else 339 # error "Unknown OS" 340 # endif 341 vg_assert2( 342 err == 0, 343 "ML_(do_syscall_for_client_WRK): sigprocmask error %d", 344 (Int)(err & 0xFFF) 345 ); 346 } 347 348 349 /* --------------------------------------------------------------------- 350 Impedance matchers and misc helpers 351 ------------------------------------------------------------------ */ 352 353 static 354 Bool eq_SyscallArgs ( SyscallArgs* a1, SyscallArgs* a2 ) 355 { 356 return a1->sysno == a2->sysno 357 && a1->arg1 == a2->arg1 358 && a1->arg2 == a2->arg2 359 && a1->arg3 == a2->arg3 360 && a1->arg4 == a2->arg4 361 && a1->arg5 == a2->arg5 362 && a1->arg6 == a2->arg6 363 && a1->arg7 == a2->arg7 364 && a1->arg8 == a2->arg8; 365 } 366 367 static 368 Bool eq_SyscallStatus ( SyscallStatus* s1, SyscallStatus* s2 ) 369 { 370 /* was: return s1->what == s2->what && sr_EQ( s1->sres, s2->sres ); */ 371 if (s1->what == s2->what && sr_EQ( s1->sres, s2->sres )) 372 return True; 373 # if defined(VGO_darwin) 374 /* Darwin-specific debugging guff */ 375 vg_assert(s1->what == s2->what); 376 VG_(printf)("eq_SyscallStatus:\n"); 377 VG_(printf)(" {%lu %lu %u}\n", s1->sres._wLO, s1->sres._wHI, s1->sres._mode); 378 VG_(printf)(" {%lu %lu %u}\n", s2->sres._wLO, s2->sres._wHI, s2->sres._mode); 379 vg_assert(0); 380 # endif 381 return False; 382 } 383 384 /* Convert between SysRes and SyscallStatus, to the extent possible. */ 385 386 static 387 SyscallStatus convert_SysRes_to_SyscallStatus ( SysRes res ) 388 { 389 SyscallStatus status; 390 status.what = SsComplete; 391 status.sres = res; 392 return status; 393 } 394 395 396 /* Impedance matchers. These convert syscall arg or result data from 397 the platform-specific in-guest-state format to the canonical 398 formats, and back. */ 399 400 static 401 void getSyscallArgsFromGuestState ( /*OUT*/SyscallArgs* canonical, 402 /*IN*/ VexGuestArchState* gst_vanilla, 403 /*IN*/ UInt trc ) 404 { 405 #if defined(VGP_x86_linux) 406 VexGuestX86State* gst = (VexGuestX86State*)gst_vanilla; 407 canonical->sysno = gst->guest_EAX; 408 canonical->arg1 = gst->guest_EBX; 409 canonical->arg2 = gst->guest_ECX; 410 canonical->arg3 = gst->guest_EDX; 411 canonical->arg4 = gst->guest_ESI; 412 canonical->arg5 = gst->guest_EDI; 413 canonical->arg6 = gst->guest_EBP; 414 canonical->arg7 = 0; 415 canonical->arg8 = 0; 416 417 #elif defined(VGP_amd64_linux) 418 VexGuestAMD64State* gst = (VexGuestAMD64State*)gst_vanilla; 419 canonical->sysno = gst->guest_RAX; 420 canonical->arg1 = gst->guest_RDI; 421 canonical->arg2 = gst->guest_RSI; 422 canonical->arg3 = gst->guest_RDX; 423 canonical->arg4 = gst->guest_R10; 424 canonical->arg5 = gst->guest_R8; 425 canonical->arg6 = gst->guest_R9; 426 canonical->arg7 = 0; 427 canonical->arg8 = 0; 428 429 #elif defined(VGP_ppc32_linux) 430 VexGuestPPC32State* gst = (VexGuestPPC32State*)gst_vanilla; 431 canonical->sysno = gst->guest_GPR0; 432 canonical->arg1 = gst->guest_GPR3; 433 canonical->arg2 = gst->guest_GPR4; 434 canonical->arg3 = gst->guest_GPR5; 435 canonical->arg4 = gst->guest_GPR6; 436 canonical->arg5 = gst->guest_GPR7; 437 canonical->arg6 = gst->guest_GPR8; 438 canonical->arg7 = 0; 439 canonical->arg8 = 0; 440 441 #elif defined(VGP_ppc64_linux) 442 VexGuestPPC64State* gst = (VexGuestPPC64State*)gst_vanilla; 443 canonical->sysno = gst->guest_GPR0; 444 canonical->arg1 = gst->guest_GPR3; 445 canonical->arg2 = gst->guest_GPR4; 446 canonical->arg3 = gst->guest_GPR5; 447 canonical->arg4 = gst->guest_GPR6; 448 canonical->arg5 = gst->guest_GPR7; 449 canonical->arg6 = gst->guest_GPR8; 450 canonical->arg7 = 0; 451 canonical->arg8 = 0; 452 453 #elif defined(VGP_arm_linux) 454 VexGuestARMState* gst = (VexGuestARMState*)gst_vanilla; 455 canonical->sysno = gst->guest_R7; 456 canonical->arg1 = gst->guest_R0; 457 canonical->arg2 = gst->guest_R1; 458 canonical->arg3 = gst->guest_R2; 459 canonical->arg4 = gst->guest_R3; 460 canonical->arg5 = gst->guest_R4; 461 canonical->arg6 = gst->guest_R5; 462 canonical->arg7 = 0; 463 canonical->arg8 = 0; 464 465 #elif defined(VGP_x86_darwin) 466 VexGuestX86State* gst = (VexGuestX86State*)gst_vanilla; 467 UWord *stack = (UWord *)gst->guest_ESP; 468 // GrP fixme hope syscalls aren't called with really shallow stacks... 469 canonical->sysno = gst->guest_EAX; 470 if (canonical->sysno != 0) { 471 // stack[0] is return address 472 canonical->arg1 = stack[1]; 473 canonical->arg2 = stack[2]; 474 canonical->arg3 = stack[3]; 475 canonical->arg4 = stack[4]; 476 canonical->arg5 = stack[5]; 477 canonical->arg6 = stack[6]; 478 canonical->arg7 = stack[7]; 479 canonical->arg8 = stack[8]; 480 } else { 481 // GrP fixme hack handle syscall() 482 // GrP fixme what about __syscall() ? 483 // stack[0] is return address 484 // DDD: the tool can't see that the params have been shifted! Can 485 // lead to incorrect checking, I think, because the PRRAn/PSARn 486 // macros will mention the pre-shifted args. 487 canonical->sysno = stack[1]; 488 vg_assert(canonical->sysno != 0); 489 canonical->arg1 = stack[2]; 490 canonical->arg2 = stack[3]; 491 canonical->arg3 = stack[4]; 492 canonical->arg4 = stack[5]; 493 canonical->arg5 = stack[6]; 494 canonical->arg6 = stack[7]; 495 canonical->arg7 = stack[8]; 496 canonical->arg8 = stack[9]; 497 498 PRINT("SYSCALL[%d,?](%s) syscall(%s, ...); please stand by...\n", 499 VG_(getpid)(), /*tid,*/ 500 VG_SYSNUM_STRING(0), VG_SYSNUM_STRING(canonical->sysno)); 501 } 502 503 // Here we determine what kind of syscall it was by looking at the 504 // interrupt kind, and then encode the syscall number using the 64-bit 505 // encoding for Valgrind's internal use. 506 // 507 // DDD: Would it be better to stash the JMP kind into the Darwin 508 // thread state rather than passing in the trc? 509 switch (trc) { 510 case VEX_TRC_JMP_SYS_INT128: 511 // int $0x80 = Unix, 64-bit result 512 vg_assert(canonical->sysno >= 0); 513 canonical->sysno = VG_DARWIN_SYSCALL_CONSTRUCT_UNIX(canonical->sysno); 514 break; 515 case VEX_TRC_JMP_SYS_SYSENTER: 516 // syscall = Unix, 32-bit result 517 // OR Mach, 32-bit result 518 if (canonical->sysno >= 0) { 519 // GrP fixme hack: 0xffff == I386_SYSCALL_NUMBER_MASK 520 canonical->sysno = VG_DARWIN_SYSCALL_CONSTRUCT_UNIX(canonical->sysno 521 & 0xffff); 522 } else { 523 canonical->sysno = VG_DARWIN_SYSCALL_CONSTRUCT_MACH(-canonical->sysno); 524 } 525 break; 526 case VEX_TRC_JMP_SYS_INT129: 527 // int $0x81 = Mach, 32-bit result 528 vg_assert(canonical->sysno < 0); 529 canonical->sysno = VG_DARWIN_SYSCALL_CONSTRUCT_MACH(-canonical->sysno); 530 break; 531 case VEX_TRC_JMP_SYS_INT130: 532 // int $0x82 = mdep, 32-bit result 533 vg_assert(canonical->sysno >= 0); 534 canonical->sysno = VG_DARWIN_SYSCALL_CONSTRUCT_MDEP(canonical->sysno); 535 break; 536 default: 537 vg_assert(0); 538 break; 539 } 540 541 #elif defined(VGP_amd64_darwin) 542 VexGuestAMD64State* gst = (VexGuestAMD64State*)gst_vanilla; 543 UWord *stack = (UWord *)gst->guest_RSP; 544 545 vg_assert(trc == VEX_TRC_JMP_SYS_SYSCALL); 546 547 // GrP fixme hope syscalls aren't called with really shallow stacks... 548 canonical->sysno = gst->guest_RAX; 549 if (canonical->sysno != __NR_syscall) { 550 // stack[0] is return address 551 canonical->arg1 = gst->guest_RDI; 552 canonical->arg2 = gst->guest_RSI; 553 canonical->arg3 = gst->guest_RDX; 554 canonical->arg4 = gst->guest_R10; // not rcx with syscall insn 555 canonical->arg5 = gst->guest_R8; 556 canonical->arg6 = gst->guest_R9; 557 canonical->arg7 = stack[1]; 558 canonical->arg8 = stack[2]; 559 } else { 560 // GrP fixme hack handle syscall() 561 // GrP fixme what about __syscall() ? 562 // stack[0] is return address 563 // DDD: the tool can't see that the params have been shifted! Can 564 // lead to incorrect checking, I think, because the PRRAn/PSARn 565 // macros will mention the pre-shifted args. 566 canonical->sysno = VG_DARWIN_SYSCALL_CONSTRUCT_UNIX(gst->guest_RDI); 567 vg_assert(canonical->sysno != __NR_syscall); 568 canonical->arg1 = gst->guest_RSI; 569 canonical->arg2 = gst->guest_RDX; 570 canonical->arg3 = gst->guest_R10; // not rcx with syscall insn 571 canonical->arg4 = gst->guest_R8; 572 canonical->arg5 = gst->guest_R9; 573 canonical->arg6 = stack[1]; 574 canonical->arg7 = stack[2]; 575 canonical->arg8 = stack[3]; 576 577 PRINT("SYSCALL[%d,?](%s) syscall(%s, ...); please stand by...\n", 578 VG_(getpid)(), /*tid,*/ 579 VG_SYSNUM_STRING(0), VG_SYSNUM_STRING(canonical->sysno)); 580 } 581 582 // no canonical->sysno adjustment needed 583 584 #elif defined(VGP_s390x_linux) 585 VexGuestS390XState* gst = (VexGuestS390XState*)gst_vanilla; 586 canonical->sysno = gst->guest_SYSNO; 587 canonical->arg1 = gst->guest_r2; 588 canonical->arg2 = gst->guest_r3; 589 canonical->arg3 = gst->guest_r4; 590 canonical->arg4 = gst->guest_r5; 591 canonical->arg5 = gst->guest_r6; 592 canonical->arg6 = gst->guest_r7; 593 canonical->arg7 = 0; 594 canonical->arg8 = 0; 595 #else 596 # error "getSyscallArgsFromGuestState: unknown arch" 597 #endif 598 } 599 600 static 601 void putSyscallArgsIntoGuestState ( /*IN*/ SyscallArgs* canonical, 602 /*OUT*/VexGuestArchState* gst_vanilla ) 603 { 604 #if defined(VGP_x86_linux) 605 VexGuestX86State* gst = (VexGuestX86State*)gst_vanilla; 606 gst->guest_EAX = canonical->sysno; 607 gst->guest_EBX = canonical->arg1; 608 gst->guest_ECX = canonical->arg2; 609 gst->guest_EDX = canonical->arg3; 610 gst->guest_ESI = canonical->arg4; 611 gst->guest_EDI = canonical->arg5; 612 gst->guest_EBP = canonical->arg6; 613 614 #elif defined(VGP_amd64_linux) 615 VexGuestAMD64State* gst = (VexGuestAMD64State*)gst_vanilla; 616 gst->guest_RAX = canonical->sysno; 617 gst->guest_RDI = canonical->arg1; 618 gst->guest_RSI = canonical->arg2; 619 gst->guest_RDX = canonical->arg3; 620 gst->guest_R10 = canonical->arg4; 621 gst->guest_R8 = canonical->arg5; 622 gst->guest_R9 = canonical->arg6; 623 624 #elif defined(VGP_ppc32_linux) 625 VexGuestPPC32State* gst = (VexGuestPPC32State*)gst_vanilla; 626 gst->guest_GPR0 = canonical->sysno; 627 gst->guest_GPR3 = canonical->arg1; 628 gst->guest_GPR4 = canonical->arg2; 629 gst->guest_GPR5 = canonical->arg3; 630 gst->guest_GPR6 = canonical->arg4; 631 gst->guest_GPR7 = canonical->arg5; 632 gst->guest_GPR8 = canonical->arg6; 633 634 #elif defined(VGP_ppc64_linux) 635 VexGuestPPC64State* gst = (VexGuestPPC64State*)gst_vanilla; 636 gst->guest_GPR0 = canonical->sysno; 637 gst->guest_GPR3 = canonical->arg1; 638 gst->guest_GPR4 = canonical->arg2; 639 gst->guest_GPR5 = canonical->arg3; 640 gst->guest_GPR6 = canonical->arg4; 641 gst->guest_GPR7 = canonical->arg5; 642 gst->guest_GPR8 = canonical->arg6; 643 644 #elif defined(VGP_arm_linux) 645 VexGuestARMState* gst = (VexGuestARMState*)gst_vanilla; 646 gst->guest_R7 = canonical->sysno; 647 gst->guest_R0 = canonical->arg1; 648 gst->guest_R1 = canonical->arg2; 649 gst->guest_R2 = canonical->arg3; 650 gst->guest_R3 = canonical->arg4; 651 gst->guest_R4 = canonical->arg5; 652 gst->guest_R5 = canonical->arg6; 653 654 #elif defined(VGP_x86_darwin) 655 VexGuestX86State* gst = (VexGuestX86State*)gst_vanilla; 656 UWord *stack = (UWord *)gst->guest_ESP; 657 658 gst->guest_EAX = VG_DARWIN_SYSNO_FOR_KERNEL(canonical->sysno); 659 660 // GrP fixme? gst->guest_TEMP_EFLAG_C = 0; 661 // stack[0] is return address 662 stack[1] = canonical->arg1; 663 stack[2] = canonical->arg2; 664 stack[3] = canonical->arg3; 665 stack[4] = canonical->arg4; 666 stack[5] = canonical->arg5; 667 stack[6] = canonical->arg6; 668 stack[7] = canonical->arg7; 669 stack[8] = canonical->arg8; 670 671 #elif defined(VGP_amd64_darwin) 672 VexGuestAMD64State* gst = (VexGuestAMD64State*)gst_vanilla; 673 UWord *stack = (UWord *)gst->guest_RSP; 674 675 gst->guest_RAX = VG_DARWIN_SYSNO_FOR_KERNEL(canonical->sysno); 676 // GrP fixme? gst->guest_TEMP_EFLAG_C = 0; 677 678 // stack[0] is return address 679 gst->guest_RDI = canonical->arg1; 680 gst->guest_RSI = canonical->arg2; 681 gst->guest_RDX = canonical->arg3; 682 gst->guest_RCX = canonical->arg4; 683 gst->guest_R8 = canonical->arg5; 684 gst->guest_R9 = canonical->arg6; 685 stack[1] = canonical->arg7; 686 stack[2] = canonical->arg8; 687 688 #elif defined(VGP_s390x_linux) 689 VexGuestS390XState* gst = (VexGuestS390XState*)gst_vanilla; 690 gst->guest_SYSNO = canonical->sysno; 691 gst->guest_r2 = canonical->arg1; 692 gst->guest_r3 = canonical->arg2; 693 gst->guest_r4 = canonical->arg3; 694 gst->guest_r5 = canonical->arg4; 695 gst->guest_r6 = canonical->arg5; 696 gst->guest_r7 = canonical->arg6; 697 698 #else 699 # error "putSyscallArgsIntoGuestState: unknown arch" 700 #endif 701 } 702 703 static 704 void getSyscallStatusFromGuestState ( /*OUT*/SyscallStatus* canonical, 705 /*IN*/ VexGuestArchState* gst_vanilla ) 706 { 707 # if defined(VGP_x86_linux) 708 VexGuestX86State* gst = (VexGuestX86State*)gst_vanilla; 709 canonical->sres = VG_(mk_SysRes_x86_linux)( gst->guest_EAX ); 710 canonical->what = SsComplete; 711 712 # elif defined(VGP_amd64_linux) 713 VexGuestAMD64State* gst = (VexGuestAMD64State*)gst_vanilla; 714 canonical->sres = VG_(mk_SysRes_amd64_linux)( gst->guest_RAX ); 715 canonical->what = SsComplete; 716 717 # elif defined(VGP_ppc32_linux) 718 VexGuestPPC32State* gst = (VexGuestPPC32State*)gst_vanilla; 719 UInt cr = LibVEX_GuestPPC32_get_CR( gst ); 720 UInt cr0so = (cr >> 28) & 1; 721 canonical->sres = VG_(mk_SysRes_ppc32_linux)( gst->guest_GPR3, cr0so ); 722 canonical->what = SsComplete; 723 724 # elif defined(VGP_ppc64_linux) 725 VexGuestPPC64State* gst = (VexGuestPPC64State*)gst_vanilla; 726 UInt cr = LibVEX_GuestPPC64_get_CR( gst ); 727 UInt cr0so = (cr >> 28) & 1; 728 canonical->sres = VG_(mk_SysRes_ppc64_linux)( gst->guest_GPR3, cr0so ); 729 canonical->what = SsComplete; 730 731 # elif defined(VGP_arm_linux) 732 VexGuestARMState* gst = (VexGuestARMState*)gst_vanilla; 733 canonical->sres = VG_(mk_SysRes_arm_linux)( gst->guest_R0 ); 734 canonical->what = SsComplete; 735 736 # elif defined(VGP_x86_darwin) 737 /* duplicates logic in m_signals.VG_UCONTEXT_SYSCALL_SYSRES */ 738 VexGuestX86State* gst = (VexGuestX86State*)gst_vanilla; 739 UInt carry = 1 & LibVEX_GuestX86_get_eflags(gst); 740 UInt err = 0; 741 UInt wLO = 0; 742 UInt wHI = 0; 743 switch (gst->guest_SC_CLASS) { 744 case VG_DARWIN_SYSCALL_CLASS_UNIX: 745 // int $0x80 = Unix, 64-bit result 746 err = carry; 747 wLO = gst->guest_EAX; 748 wHI = gst->guest_EDX; 749 break; 750 case VG_DARWIN_SYSCALL_CLASS_MACH: 751 // int $0x81 = Mach, 32-bit result 752 wLO = gst->guest_EAX; 753 break; 754 case VG_DARWIN_SYSCALL_CLASS_MDEP: 755 // int $0x82 = mdep, 32-bit result 756 wLO = gst->guest_EAX; 757 break; 758 default: 759 vg_assert(0); 760 break; 761 } 762 canonical->sres = VG_(mk_SysRes_x86_darwin)( 763 gst->guest_SC_CLASS, err ? True : False, 764 wHI, wLO 765 ); 766 canonical->what = SsComplete; 767 768 # elif defined(VGP_amd64_darwin) 769 /* duplicates logic in m_signals.VG_UCONTEXT_SYSCALL_SYSRES */ 770 VexGuestAMD64State* gst = (VexGuestAMD64State*)gst_vanilla; 771 ULong carry = 1 & LibVEX_GuestAMD64_get_rflags(gst); 772 ULong err = 0; 773 ULong wLO = 0; 774 ULong wHI = 0; 775 switch (gst->guest_SC_CLASS) { 776 case VG_DARWIN_SYSCALL_CLASS_UNIX: 777 // syscall = Unix, 128-bit result 778 err = carry; 779 wLO = gst->guest_RAX; 780 wHI = gst->guest_RDX; 781 break; 782 case VG_DARWIN_SYSCALL_CLASS_MACH: 783 // syscall = Mach, 64-bit result 784 wLO = gst->guest_RAX; 785 break; 786 case VG_DARWIN_SYSCALL_CLASS_MDEP: 787 // syscall = mdep, 64-bit result 788 wLO = gst->guest_RAX; 789 break; 790 default: 791 vg_assert(0); 792 break; 793 } 794 canonical->sres = VG_(mk_SysRes_amd64_darwin)( 795 gst->guest_SC_CLASS, err ? True : False, 796 wHI, wLO 797 ); 798 canonical->what = SsComplete; 799 800 # elif defined(VGP_s390x_linux) 801 VexGuestS390XState* gst = (VexGuestS390XState*)gst_vanilla; 802 canonical->sres = VG_(mk_SysRes_s390x_linux)( gst->guest_r2 ); 803 canonical->what = SsComplete; 804 805 # else 806 # error "getSyscallStatusFromGuestState: unknown arch" 807 # endif 808 } 809 810 static 811 void putSyscallStatusIntoGuestState ( /*IN*/ ThreadId tid, 812 /*IN*/ SyscallStatus* canonical, 813 /*OUT*/VexGuestArchState* gst_vanilla ) 814 { 815 # if defined(VGP_x86_linux) 816 VexGuestX86State* gst = (VexGuestX86State*)gst_vanilla; 817 vg_assert(canonical->what == SsComplete); 818 if (sr_isError(canonical->sres)) { 819 /* This isn't exactly right, in that really a Failure with res 820 not in the range 1 .. 4095 is unrepresentable in the 821 Linux-x86 scheme. Oh well. */ 822 gst->guest_EAX = - (Int)sr_Err(canonical->sres); 823 } else { 824 gst->guest_EAX = sr_Res(canonical->sres); 825 } 826 VG_TRACK( post_reg_write, Vg_CoreSysCall, tid, 827 OFFSET_x86_EAX, sizeof(UWord) ); 828 829 # elif defined(VGP_amd64_linux) 830 VexGuestAMD64State* gst = (VexGuestAMD64State*)gst_vanilla; 831 vg_assert(canonical->what == SsComplete); 832 if (sr_isError(canonical->sres)) { 833 /* This isn't exactly right, in that really a Failure with res 834 not in the range 1 .. 4095 is unrepresentable in the 835 Linux-amd64 scheme. Oh well. */ 836 gst->guest_RAX = - (Long)sr_Err(canonical->sres); 837 } else { 838 gst->guest_RAX = sr_Res(canonical->sres); 839 } 840 VG_TRACK( post_reg_write, Vg_CoreSysCall, tid, 841 OFFSET_amd64_RAX, sizeof(UWord) ); 842 843 # elif defined(VGP_ppc32_linux) 844 VexGuestPPC32State* gst = (VexGuestPPC32State*)gst_vanilla; 845 UInt old_cr = LibVEX_GuestPPC32_get_CR(gst); 846 vg_assert(canonical->what == SsComplete); 847 if (sr_isError(canonical->sres)) { 848 /* set CR0.SO */ 849 LibVEX_GuestPPC32_put_CR( old_cr | (1<<28), gst ); 850 gst->guest_GPR3 = sr_Err(canonical->sres); 851 } else { 852 /* clear CR0.SO */ 853 LibVEX_GuestPPC32_put_CR( old_cr & ~(1<<28), gst ); 854 gst->guest_GPR3 = sr_Res(canonical->sres); 855 } 856 VG_TRACK( post_reg_write, Vg_CoreSysCall, tid, 857 OFFSET_ppc32_GPR3, sizeof(UWord) ); 858 VG_TRACK( post_reg_write, Vg_CoreSysCall, tid, 859 OFFSET_ppc32_CR0_0, sizeof(UChar) ); 860 861 # elif defined(VGP_ppc64_linux) 862 VexGuestPPC64State* gst = (VexGuestPPC64State*)gst_vanilla; 863 UInt old_cr = LibVEX_GuestPPC64_get_CR(gst); 864 vg_assert(canonical->what == SsComplete); 865 if (sr_isError(canonical->sres)) { 866 /* set CR0.SO */ 867 LibVEX_GuestPPC64_put_CR( old_cr | (1<<28), gst ); 868 gst->guest_GPR3 = sr_Err(canonical->sres); 869 } else { 870 /* clear CR0.SO */ 871 LibVEX_GuestPPC64_put_CR( old_cr & ~(1<<28), gst ); 872 gst->guest_GPR3 = sr_Res(canonical->sres); 873 } 874 VG_TRACK( post_reg_write, Vg_CoreSysCall, tid, 875 OFFSET_ppc64_GPR3, sizeof(UWord) ); 876 VG_TRACK( post_reg_write, Vg_CoreSysCall, tid, 877 OFFSET_ppc64_CR0_0, sizeof(UChar) ); 878 879 # elif defined(VGP_arm_linux) 880 VexGuestARMState* gst = (VexGuestARMState*)gst_vanilla; 881 vg_assert(canonical->what == SsComplete); 882 if (sr_isError(canonical->sres)) { 883 /* This isn't exactly right, in that really a Failure with res 884 not in the range 1 .. 4095 is unrepresentable in the 885 Linux-arm scheme. Oh well. */ 886 gst->guest_R0 = - (Int)sr_Err(canonical->sres); 887 } else { 888 gst->guest_R0 = sr_Res(canonical->sres); 889 } 890 VG_TRACK( post_reg_write, Vg_CoreSysCall, tid, 891 OFFSET_arm_R0, sizeof(UWord) ); 892 893 #elif defined(VGP_x86_darwin) 894 VexGuestX86State* gst = (VexGuestX86State*)gst_vanilla; 895 SysRes sres = canonical->sres; 896 vg_assert(canonical->what == SsComplete); 897 /* Unfortunately here we have to break abstraction and look 898 directly inside 'res', in order to decide what to do. */ 899 switch (sres._mode) { 900 case SysRes_MACH: // int $0x81 = Mach, 32-bit result 901 case SysRes_MDEP: // int $0x82 = mdep, 32-bit result 902 gst->guest_EAX = sres._wLO; 903 VG_TRACK( post_reg_write, Vg_CoreSysCall, tid, 904 OFFSET_x86_EAX, sizeof(UInt) ); 905 break; 906 case SysRes_UNIX_OK: // int $0x80 = Unix, 64-bit result 907 case SysRes_UNIX_ERR: // int $0x80 = Unix, 64-bit error 908 gst->guest_EAX = sres._wLO; 909 VG_TRACK( post_reg_write, Vg_CoreSysCall, tid, 910 OFFSET_x86_EAX, sizeof(UInt) ); 911 gst->guest_EDX = sres._wHI; 912 VG_TRACK( post_reg_write, Vg_CoreSysCall, tid, 913 OFFSET_x86_EDX, sizeof(UInt) ); 914 LibVEX_GuestX86_put_eflag_c( sres._mode==SysRes_UNIX_ERR ? 1 : 0, 915 gst ); 916 // GrP fixme sets defined for entire eflags, not just bit c 917 // DDD: this breaks exp-ptrcheck. 918 VG_TRACK( post_reg_write, Vg_CoreSysCall, tid, 919 offsetof(VexGuestX86State, guest_CC_DEP1), sizeof(UInt) ); 920 break; 921 default: 922 vg_assert(0); 923 break; 924 } 925 926 #elif defined(VGP_amd64_darwin) 927 VexGuestAMD64State* gst = (VexGuestAMD64State*)gst_vanilla; 928 SysRes sres = canonical->sres; 929 vg_assert(canonical->what == SsComplete); 930 /* Unfortunately here we have to break abstraction and look 931 directly inside 'res', in order to decide what to do. */ 932 switch (sres._mode) { 933 case SysRes_MACH: // syscall = Mach, 64-bit result 934 case SysRes_MDEP: // syscall = mdep, 64-bit result 935 gst->guest_RAX = sres._wLO; 936 VG_TRACK( post_reg_write, Vg_CoreSysCall, tid, 937 OFFSET_amd64_RAX, sizeof(ULong) ); 938 break; 939 case SysRes_UNIX_OK: // syscall = Unix, 128-bit result 940 case SysRes_UNIX_ERR: // syscall = Unix, 128-bit error 941 gst->guest_RAX = sres._wLO; 942 VG_TRACK( post_reg_write, Vg_CoreSysCall, tid, 943 OFFSET_amd64_RAX, sizeof(ULong) ); 944 gst->guest_RDX = sres._wHI; 945 VG_TRACK( post_reg_write, Vg_CoreSysCall, tid, 946 OFFSET_amd64_RDX, sizeof(ULong) ); 947 LibVEX_GuestAMD64_put_rflag_c( sres._mode==SysRes_UNIX_ERR ? 1 : 0, 948 gst ); 949 // GrP fixme sets defined for entire rflags, not just bit c 950 // DDD: this breaks exp-ptrcheck. 951 VG_TRACK( post_reg_write, Vg_CoreSysCall, tid, 952 offsetof(VexGuestAMD64State, guest_CC_DEP1), sizeof(ULong) ); 953 break; 954 default: 955 vg_assert(0); 956 break; 957 } 958 959 # elif defined(VGP_s390x_linux) 960 VexGuestS390XState* gst = (VexGuestS390XState*)gst_vanilla; 961 vg_assert(canonical->what == SsComplete); 962 if (sr_isError(canonical->sres)) { 963 gst->guest_r2 = - (Long)sr_Err(canonical->sres); 964 } else { 965 gst->guest_r2 = sr_Res(canonical->sres); 966 } 967 968 # else 969 # error "putSyscallStatusIntoGuestState: unknown arch" 970 # endif 971 } 972 973 974 /* Tell me the offsets in the guest state of the syscall params, so 975 that the scalar argument checkers don't have to have this info 976 hardwired. */ 977 978 static 979 void getSyscallArgLayout ( /*OUT*/SyscallArgLayout* layout ) 980 { 981 #if defined(VGP_x86_linux) 982 layout->o_sysno = OFFSET_x86_EAX; 983 layout->o_arg1 = OFFSET_x86_EBX; 984 layout->o_arg2 = OFFSET_x86_ECX; 985 layout->o_arg3 = OFFSET_x86_EDX; 986 layout->o_arg4 = OFFSET_x86_ESI; 987 layout->o_arg5 = OFFSET_x86_EDI; 988 layout->o_arg6 = OFFSET_x86_EBP; 989 layout->uu_arg7 = -1; /* impossible value */ 990 layout->uu_arg8 = -1; /* impossible value */ 991 992 #elif defined(VGP_amd64_linux) 993 layout->o_sysno = OFFSET_amd64_RAX; 994 layout->o_arg1 = OFFSET_amd64_RDI; 995 layout->o_arg2 = OFFSET_amd64_RSI; 996 layout->o_arg3 = OFFSET_amd64_RDX; 997 layout->o_arg4 = OFFSET_amd64_R10; 998 layout->o_arg5 = OFFSET_amd64_R8; 999 layout->o_arg6 = OFFSET_amd64_R9; 1000 layout->uu_arg7 = -1; /* impossible value */ 1001 layout->uu_arg8 = -1; /* impossible value */ 1002 1003 #elif defined(VGP_ppc32_linux) 1004 layout->o_sysno = OFFSET_ppc32_GPR0; 1005 layout->o_arg1 = OFFSET_ppc32_GPR3; 1006 layout->o_arg2 = OFFSET_ppc32_GPR4; 1007 layout->o_arg3 = OFFSET_ppc32_GPR5; 1008 layout->o_arg4 = OFFSET_ppc32_GPR6; 1009 layout->o_arg5 = OFFSET_ppc32_GPR7; 1010 layout->o_arg6 = OFFSET_ppc32_GPR8; 1011 layout->uu_arg7 = -1; /* impossible value */ 1012 layout->uu_arg8 = -1; /* impossible value */ 1013 1014 #elif defined(VGP_ppc64_linux) 1015 layout->o_sysno = OFFSET_ppc64_GPR0; 1016 layout->o_arg1 = OFFSET_ppc64_GPR3; 1017 layout->o_arg2 = OFFSET_ppc64_GPR4; 1018 layout->o_arg3 = OFFSET_ppc64_GPR5; 1019 layout->o_arg4 = OFFSET_ppc64_GPR6; 1020 layout->o_arg5 = OFFSET_ppc64_GPR7; 1021 layout->o_arg6 = OFFSET_ppc64_GPR8; 1022 layout->uu_arg7 = -1; /* impossible value */ 1023 layout->uu_arg8 = -1; /* impossible value */ 1024 1025 #elif defined(VGP_arm_linux) 1026 layout->o_sysno = OFFSET_arm_R7; 1027 layout->o_arg1 = OFFSET_arm_R0; 1028 layout->o_arg2 = OFFSET_arm_R1; 1029 layout->o_arg3 = OFFSET_arm_R2; 1030 layout->o_arg4 = OFFSET_arm_R3; 1031 layout->o_arg5 = OFFSET_arm_R4; 1032 layout->o_arg6 = OFFSET_arm_R5; 1033 layout->uu_arg7 = -1; /* impossible value */ 1034 layout->uu_arg8 = -1; /* impossible value */ 1035 1036 #elif defined(VGP_x86_darwin) 1037 layout->o_sysno = OFFSET_x86_EAX; 1038 // syscall parameters are on stack in C convention 1039 layout->s_arg1 = sizeof(UWord) * 1; 1040 layout->s_arg2 = sizeof(UWord) * 2; 1041 layout->s_arg3 = sizeof(UWord) * 3; 1042 layout->s_arg4 = sizeof(UWord) * 4; 1043 layout->s_arg5 = sizeof(UWord) * 5; 1044 layout->s_arg6 = sizeof(UWord) * 6; 1045 layout->s_arg7 = sizeof(UWord) * 7; 1046 layout->s_arg8 = sizeof(UWord) * 8; 1047 1048 #elif defined(VGP_amd64_darwin) 1049 layout->o_sysno = OFFSET_amd64_RAX; 1050 layout->o_arg1 = OFFSET_amd64_RDI; 1051 layout->o_arg2 = OFFSET_amd64_RSI; 1052 layout->o_arg3 = OFFSET_amd64_RDX; 1053 layout->o_arg4 = OFFSET_amd64_RCX; 1054 layout->o_arg5 = OFFSET_amd64_R8; 1055 layout->o_arg6 = OFFSET_amd64_R9; 1056 layout->s_arg7 = sizeof(UWord) * 1; 1057 layout->s_arg8 = sizeof(UWord) * 2; 1058 1059 #elif defined(VGP_s390x_linux) 1060 layout->o_sysno = OFFSET_s390x_SYSNO; 1061 layout->o_arg1 = OFFSET_s390x_r2; 1062 layout->o_arg2 = OFFSET_s390x_r3; 1063 layout->o_arg3 = OFFSET_s390x_r4; 1064 layout->o_arg4 = OFFSET_s390x_r5; 1065 layout->o_arg5 = OFFSET_s390x_r6; 1066 layout->o_arg6 = OFFSET_s390x_r7; 1067 layout->uu_arg7 = -1; /* impossible value */ 1068 layout->uu_arg8 = -1; /* impossible value */ 1069 #else 1070 # error "getSyscallLayout: unknown arch" 1071 #endif 1072 } 1073 1074 1075 /* --------------------------------------------------------------------- 1076 The main driver logic 1077 ------------------------------------------------------------------ */ 1078 1079 /* Finding the handlers for a given syscall, or faking up one 1080 when no handler is found. */ 1081 1082 static 1083 void bad_before ( ThreadId tid, 1084 SyscallArgLayout* layout, 1085 /*MOD*/SyscallArgs* args, 1086 /*OUT*/SyscallStatus* status, 1087 /*OUT*/UWord* flags ) 1088 { 1089 VG_(dmsg)("WARNING: unhandled syscall: %s\n", 1090 VG_SYSNUM_STRING_EXTRA(args->sysno)); 1091 if (VG_(clo_verbosity) > 1) { 1092 VG_(get_and_pp_StackTrace)(tid, VG_(clo_backtrace_size)); 1093 } 1094 VG_(dmsg)("You may be able to write your own handler.\n"); 1095 VG_(dmsg)("Read the file README_MISSING_SYSCALL_OR_IOCTL.\n"); 1096 VG_(dmsg)("Nevertheless we consider this a bug. Please report\n"); 1097 VG_(dmsg)("it at http://valgrind.org/support/bug_reports.html.\n"); 1098 1099 SET_STATUS_Failure(VKI_ENOSYS); 1100 } 1101 1102 static SyscallTableEntry bad_sys = 1103 { bad_before, NULL }; 1104 1105 static const SyscallTableEntry* get_syscall_entry ( Int syscallno ) 1106 { 1107 const SyscallTableEntry* sys = NULL; 1108 1109 # if defined(VGO_linux) 1110 sys = ML_(get_linux_syscall_entry)( syscallno ); 1111 1112 # elif defined(VGO_darwin) 1113 Int idx = VG_DARWIN_SYSNO_INDEX(syscallno); 1114 1115 switch (VG_DARWIN_SYSNO_CLASS(syscallno)) { 1116 case VG_DARWIN_SYSCALL_CLASS_UNIX: 1117 if (idx >= 0 && idx < ML_(syscall_table_size) && 1118 ML_(syscall_table)[idx].before != NULL) 1119 sys = &ML_(syscall_table)[idx]; 1120 break; 1121 case VG_DARWIN_SYSCALL_CLASS_MACH: 1122 if (idx >= 0 && idx < ML_(mach_trap_table_size) && 1123 ML_(mach_trap_table)[idx].before != NULL) 1124 sys = &ML_(mach_trap_table)[idx]; 1125 break; 1126 case VG_DARWIN_SYSCALL_CLASS_MDEP: 1127 if (idx >= 0 && idx < ML_(mdep_trap_table_size) && 1128 ML_(mdep_trap_table)[idx].before != NULL) 1129 sys = &ML_(mdep_trap_table)[idx]; 1130 break; 1131 default: 1132 vg_assert(0); 1133 break; 1134 } 1135 1136 # else 1137 # error Unknown OS 1138 # endif 1139 1140 return sys == NULL ? &bad_sys : sys; 1141 } 1142 1143 1144 /* Add and remove signals from mask so that we end up telling the 1145 kernel the state we actually want rather than what the client 1146 wants. */ 1147 static void sanitize_client_sigmask(vki_sigset_t *mask) 1148 { 1149 VG_(sigdelset)(mask, VKI_SIGKILL); 1150 VG_(sigdelset)(mask, VKI_SIGSTOP); 1151 VG_(sigdelset)(mask, VG_SIGVGKILL); /* never block */ 1152 } 1153 1154 typedef 1155 struct { 1156 SyscallArgs orig_args; 1157 SyscallArgs args; 1158 SyscallStatus status; 1159 UWord flags; 1160 } 1161 SyscallInfo; 1162 1163 SyscallInfo syscallInfo[VG_N_THREADS]; 1164 1165 1166 /* The scheduler needs to be able to zero out these records after a 1167 fork, hence this is exported from m_syswrap. */ 1168 void VG_(clear_syscallInfo) ( Int tid ) 1169 { 1170 vg_assert(tid >= 0 && tid < VG_N_THREADS); 1171 VG_(memset)( & syscallInfo[tid], 0, sizeof( syscallInfo[tid] )); 1172 syscallInfo[tid].status.what = SsIdle; 1173 } 1174 1175 static void ensure_initialised ( void ) 1176 { 1177 Int i; 1178 static Bool init_done = False; 1179 if (init_done) 1180 return; 1181 init_done = True; 1182 for (i = 0; i < VG_N_THREADS; i++) { 1183 VG_(clear_syscallInfo)( i ); 1184 } 1185 } 1186 1187 /* --- This is the main function of this file. --- */ 1188 1189 void VG_(client_syscall) ( ThreadId tid, UInt trc ) 1190 { 1191 Word sysno; 1192 ThreadState* tst; 1193 const SyscallTableEntry* ent; 1194 SyscallArgLayout layout; 1195 SyscallInfo* sci; 1196 1197 ensure_initialised(); 1198 1199 vg_assert(VG_(is_valid_tid)(tid)); 1200 vg_assert(tid >= 1 && tid < VG_N_THREADS); 1201 vg_assert(VG_(is_running_thread)(tid)); 1202 1203 tst = VG_(get_ThreadState)(tid); 1204 1205 /* BEGIN ensure root thread's stack is suitably mapped */ 1206 /* In some rare circumstances, we may do the syscall without the 1207 bottom page of the stack being mapped, because the stack pointer 1208 was moved down just a few instructions before the syscall 1209 instruction, and there have been no memory references since 1210 then, that would cause a call to VG_(extend_stack) to have 1211 happened. 1212 1213 In native execution that's OK: the kernel automagically extends 1214 the stack's mapped area down to cover the stack pointer (or sp - 1215 redzone, really). In simulated normal execution that's OK too, 1216 since any signals we get from accessing below the mapped area of 1217 the (guest's) stack lead us to VG_(extend_stack), where we 1218 simulate the kernel's stack extension logic. But that leaves 1219 the problem of entering a syscall with the SP unmapped. Because 1220 the kernel doesn't know that the segment immediately above SP is 1221 supposed to be a grow-down segment, it causes the syscall to 1222 fail, and thereby causes a divergence between native behaviour 1223 (syscall succeeds) and simulated behaviour (syscall fails). 1224 1225 This is quite a rare failure mode. It has only been seen 1226 affecting calls to sys_readlink on amd64-linux, and even then it 1227 requires a certain code sequence around the syscall to trigger 1228 it. Here is one: 1229 1230 extern int my_readlink ( const char* path ); 1231 asm( 1232 ".text\n" 1233 ".globl my_readlink\n" 1234 "my_readlink:\n" 1235 "\tsubq $0x1008,%rsp\n" 1236 "\tmovq %rdi,%rdi\n" // path is in rdi 1237 "\tmovq %rsp,%rsi\n" // &buf[0] -> rsi 1238 "\tmovl $0x1000,%edx\n" // sizeof(buf) in rdx 1239 "\tmovl $"__NR_READLINK",%eax\n" // syscall number 1240 "\tsyscall\n" 1241 "\taddq $0x1008,%rsp\n" 1242 "\tret\n" 1243 ".previous\n" 1244 ); 1245 1246 For more details, see bug #156404 1247 (https://bugs.kde.org/show_bug.cgi?id=156404). 1248 1249 The fix is actually very simple. We simply need to call 1250 VG_(extend_stack) for this thread, handing it the lowest 1251 possible valid address for stack (sp - redzone), to ensure the 1252 pages all the way down to that address, are mapped. Because 1253 this is a potentially expensive and frequent operation, we 1254 filter in two ways: 1255 1256 First, only the main thread (tid=1) has a growdown stack. So 1257 ignore all others. It is conceivable, although highly unlikely, 1258 that the main thread exits, and later another thread is 1259 allocated tid=1, but that's harmless, I believe; 1260 VG_(extend_stack) will do nothing when applied to a non-root 1261 thread. 1262 1263 Secondly, first call VG_(am_find_nsegment) directly, to see if 1264 the page holding (sp - redzone) is mapped correctly. If so, do 1265 nothing. This is almost always the case. VG_(extend_stack) 1266 calls VG_(am_find_nsegment) twice, so this optimisation -- and 1267 that's all it is -- more or less halves the number of calls to 1268 VG_(am_find_nsegment) required. 1269 1270 TODO: the test "seg->kind == SkAnonC" is really inadequate, 1271 because although it tests whether the segment is mapped 1272 _somehow_, it doesn't check that it has the right permissions 1273 (r,w, maybe x) ? We could test that here, but it will also be 1274 necessary to fix the corresponding test in VG_(extend_stack). 1275 1276 All this guff is of course Linux-specific. Hence the ifdef. 1277 */ 1278 # if defined(VGO_linux) 1279 if (tid == 1/*ROOT THREAD*/) { 1280 Addr stackMin = VG_(get_SP)(tid) - VG_STACK_REDZONE_SZB; 1281 NSegment const* seg = VG_(am_find_nsegment)(stackMin); 1282 if (seg && seg->kind == SkAnonC) { 1283 /* stackMin is already mapped. Nothing to do. */ 1284 } else { 1285 (void)VG_(extend_stack)( stackMin, 1286 tst->client_stack_szB ); 1287 } 1288 } 1289 # endif 1290 /* END ensure root thread's stack is suitably mapped */ 1291 1292 /* First off, get the syscall args and number. This is a 1293 platform-dependent action. */ 1294 1295 sci = & syscallInfo[tid]; 1296 vg_assert(sci->status.what == SsIdle); 1297 1298 getSyscallArgsFromGuestState( &sci->orig_args, &tst->arch.vex, trc ); 1299 1300 /* Copy .orig_args to .args. The pre-handler may modify .args, but 1301 we want to keep the originals too, just in case. */ 1302 sci->args = sci->orig_args; 1303 1304 /* Save the syscall number in the thread state in case the syscall 1305 is interrupted by a signal. */ 1306 sysno = sci->orig_args.sysno; 1307 1308 /* It's sometimes useful, as a crude debugging hack, to get a 1309 stack trace at each (or selected) syscalls. */ 1310 if (0 && sysno == __NR_ioctl) { 1311 VG_(umsg)("\nioctl:\n"); 1312 VG_(get_and_pp_StackTrace)(tid, 10); 1313 VG_(umsg)("\n"); 1314 } 1315 1316 # if defined(VGO_darwin) 1317 /* Record syscall class. But why? Because the syscall might be 1318 interrupted by a signal, and in the signal handler (which will 1319 be m_signals.async_signalhandler) we will need to build a SysRes 1320 reflecting the syscall return result. In order to do that we 1321 need to know the syscall class. Hence stash it in the guest 1322 state of this thread. This madness is not needed on Linux 1323 because it only has a single syscall return convention and so 1324 there is no ambiguity involved in converting the post-signal 1325 machine state into a SysRes. */ 1326 tst->arch.vex.guest_SC_CLASS = VG_DARWIN_SYSNO_CLASS(sysno); 1327 # endif 1328 1329 /* The default what-to-do-next thing is hand the syscall to the 1330 kernel, so we pre-set that here. Set .sres to something 1331 harmless looking (is irrelevant because .what is not 1332 SsComplete.) */ 1333 sci->status.what = SsHandToKernel; 1334 sci->status.sres = VG_(mk_SysRes_Error)(0); 1335 sci->flags = 0; 1336 1337 /* Fetch the syscall's handlers. If no handlers exist for this 1338 syscall, we are given dummy handlers which force an immediate 1339 return with ENOSYS. */ 1340 ent = get_syscall_entry(sysno); 1341 1342 /* Fetch the layout information, which tells us where in the guest 1343 state the syscall args reside. This is a platform-dependent 1344 action. This info is needed so that the scalar syscall argument 1345 checks (PRE_REG_READ calls) know which bits of the guest state 1346 they need to inspect. */ 1347 getSyscallArgLayout( &layout ); 1348 1349 /* Make sure the tmp signal mask matches the real signal mask; 1350 sigsuspend may change this. */ 1351 vg_assert(VG_(iseqsigset)(&tst->sig_mask, &tst->tmp_sig_mask)); 1352 1353 /* Right, we're finally ready to Party. Call the pre-handler and 1354 see what we get back. At this point: 1355 1356 sci->status.what is Unset (we don't know yet). 1357 sci->orig_args contains the original args. 1358 sci->args is the same as sci->orig_args. 1359 sci->flags is zero. 1360 */ 1361 1362 PRINT("SYSCALL[%d,%d](%s) ", 1363 VG_(getpid)(), tid, VG_SYSNUM_STRING(sysno)); 1364 1365 /* Do any pre-syscall actions */ 1366 if (VG_(needs).syscall_wrapper) { 1367 UWord tmpv[8]; 1368 tmpv[0] = sci->orig_args.arg1; 1369 tmpv[1] = sci->orig_args.arg2; 1370 tmpv[2] = sci->orig_args.arg3; 1371 tmpv[3] = sci->orig_args.arg4; 1372 tmpv[4] = sci->orig_args.arg5; 1373 tmpv[5] = sci->orig_args.arg6; 1374 tmpv[6] = sci->orig_args.arg7; 1375 tmpv[7] = sci->orig_args.arg8; 1376 VG_TDICT_CALL(tool_pre_syscall, tid, sysno, 1377 &tmpv[0], sizeof(tmpv)/sizeof(tmpv[0])); 1378 } 1379 1380 vg_assert(ent); 1381 vg_assert(ent->before); 1382 (ent->before)( tid, 1383 &layout, 1384 &sci->args, &sci->status, &sci->flags ); 1385 1386 /* The pre-handler may have modified: 1387 sci->args 1388 sci->status 1389 sci->flags 1390 All else remains unchanged. 1391 Although the args may be modified, pre handlers are not allowed 1392 to change the syscall number. 1393 */ 1394 /* Now we proceed according to what the pre-handler decided. */ 1395 vg_assert(sci->status.what == SsHandToKernel 1396 || sci->status.what == SsComplete); 1397 vg_assert(sci->args.sysno == sci->orig_args.sysno); 1398 1399 if (sci->status.what == SsComplete && !sr_isError(sci->status.sres)) { 1400 /* The pre-handler completed the syscall itself, declaring 1401 success. */ 1402 if (sci->flags & SfNoWriteResult) { 1403 PRINT(" --> [pre-success] NoWriteResult"); 1404 } else { 1405 PRINT(" --> [pre-success] Success(0x%llx:0x%llx)", 1406 (ULong)sr_ResHI(sci->status.sres), 1407 (ULong)sr_Res(sci->status.sres)); 1408 } 1409 /* In this case the allowable flags are to ask for a signal-poll 1410 and/or a yield after the call. Changing the args isn't 1411 allowed. */ 1412 vg_assert(0 == (sci->flags 1413 & ~(SfPollAfter | SfYieldAfter | SfNoWriteResult))); 1414 vg_assert(eq_SyscallArgs(&sci->args, &sci->orig_args)); 1415 } 1416 1417 else 1418 if (sci->status.what == SsComplete && sr_isError(sci->status.sres)) { 1419 /* The pre-handler decided to fail syscall itself. */ 1420 PRINT(" --> [pre-fail] Failure(0x%llx)", (ULong)sr_Err(sci->status.sres)); 1421 /* In this case, the pre-handler is also allowed to ask for the 1422 post-handler to be run anyway. Changing the args is not 1423 allowed. */ 1424 vg_assert(0 == (sci->flags & ~(SfMayBlock | SfPostOnFail | SfPollAfter))); 1425 vg_assert(eq_SyscallArgs(&sci->args, &sci->orig_args)); 1426 } 1427 1428 else 1429 if (sci->status.what != SsHandToKernel) { 1430 /* huh?! */ 1431 vg_assert(0); 1432 } 1433 1434 else /* (sci->status.what == HandToKernel) */ { 1435 /* Ok, this is the usual case -- and the complicated one. There 1436 are two subcases: sync and async. async is the general case 1437 and is to be used when there is any possibility that the 1438 syscall might block [a fact that the pre-handler must tell us 1439 via the sci->flags field.] Because the tidying-away / 1440 context-switch overhead of the async case could be large, if 1441 we are sure that the syscall will not block, we fast-track it 1442 by doing it directly in this thread, which is a lot 1443 simpler. */ 1444 1445 /* Check that the given flags are allowable: MayBlock, PollAfter 1446 and PostOnFail are ok. */ 1447 vg_assert(0 == (sci->flags & ~(SfMayBlock | SfPostOnFail | SfPollAfter))); 1448 1449 if (sci->flags & SfMayBlock) { 1450 1451 /* Syscall may block, so run it asynchronously */ 1452 vki_sigset_t mask; 1453 1454 PRINT(" --> [async] ... \n"); 1455 1456 mask = tst->sig_mask; 1457 sanitize_client_sigmask(&mask); 1458 1459 /* Gack. More impedance matching. Copy the possibly 1460 modified syscall args back into the guest state. */ 1461 /* JRS 2009-Mar-16: if the syscall args are possibly modified, 1462 then this assertion is senseless: 1463 vg_assert(eq_SyscallArgs(&sci->args, &sci->orig_args)); 1464 The case that exposed it was sys_posix_spawn on Darwin, 1465 which heavily modifies its arguments but then lets the call 1466 go through anyway, with SfToBlock set, hence we end up here. */ 1467 putSyscallArgsIntoGuestState( &sci->args, &tst->arch.vex ); 1468 1469 /* Drop the bigLock */ 1470 VG_(release_BigLock)(tid, VgTs_WaitSys, "VG_(client_syscall)[async]"); 1471 /* Urr. We're now in a race against other threads trying to 1472 acquire the bigLock. I guess that doesn't matter provided 1473 that do_syscall_for_client only touches thread-local 1474 state. */ 1475 1476 /* Do the call, which operates directly on the guest state, 1477 not on our abstracted copies of the args/result. */ 1478 do_syscall_for_client(sysno, tst, &mask); 1479 1480 /* do_syscall_for_client may not return if the syscall was 1481 interrupted by a signal. In that case, flow of control is 1482 first to m_signals.async_sighandler, which calls 1483 VG_(fixup_guest_state_after_syscall_interrupted), which 1484 fixes up the guest state, and possibly calls 1485 VG_(post_syscall). Once that's done, control drops back 1486 to the scheduler. */ 1487 1488 /* Darwin: do_syscall_for_client may not return if the 1489 syscall was workq_ops(WQOPS_THREAD_RETURN) and the kernel 1490 responded by starting the thread at wqthread_hijack(reuse=1) 1491 (to run another workqueue item). In that case, wqthread_hijack 1492 calls ML_(wqthread_continue), which is similar to 1493 VG_(fixup_guest_state_after_syscall_interrupted). */ 1494 1495 /* Reacquire the lock */ 1496 VG_(acquire_BigLock)(tid, "VG_(client_syscall)[async]"); 1497 1498 /* Even more impedance matching. Extract the syscall status 1499 from the guest state. */ 1500 getSyscallStatusFromGuestState( &sci->status, &tst->arch.vex ); 1501 vg_assert(sci->status.what == SsComplete); 1502 1503 /* Be decorative, if required. */ 1504 if (VG_(clo_trace_syscalls)) { 1505 Bool failed = sr_isError(sci->status.sres); 1506 if (failed) { 1507 PRINT("SYSCALL[%d,%d](%s) ... [async] --> Failure(0x%llx)", 1508 VG_(getpid)(), tid, VG_SYSNUM_STRING(sysno), 1509 (ULong)sr_Err(sci->status.sres)); 1510 } else { 1511 PRINT("SYSCALL[%d,%d](%s) ... [async] --> " 1512 "Success(0x%llx:0x%llx)", 1513 VG_(getpid)(), tid, VG_SYSNUM_STRING(sysno), 1514 (ULong)sr_ResHI(sci->status.sres), 1515 (ULong)sr_Res(sci->status.sres) ); 1516 } 1517 } 1518 1519 } else { 1520 1521 /* run the syscall directly */ 1522 /* The pre-handler may have modified the syscall args, but 1523 since we're passing values in ->args directly to the 1524 kernel, there's no point in flushing them back to the 1525 guest state. Indeed doing so could be construed as 1526 incorrect. */ 1527 SysRes sres 1528 = VG_(do_syscall)(sysno, sci->args.arg1, sci->args.arg2, 1529 sci->args.arg3, sci->args.arg4, 1530 sci->args.arg5, sci->args.arg6, 1531 sci->args.arg7, sci->args.arg8 ); 1532 sci->status = convert_SysRes_to_SyscallStatus(sres); 1533 1534 /* Be decorative, if required. */ 1535 if (VG_(clo_trace_syscalls)) { 1536 Bool failed = sr_isError(sci->status.sres); 1537 if (failed) { 1538 PRINT("[sync] --> Failure(0x%llx)", 1539 (ULong)sr_Err(sci->status.sres) ); 1540 } else { 1541 PRINT("[sync] --> Success(0x%llx:0x%llx)", 1542 (ULong)sr_ResHI(sci->status.sres), 1543 (ULong)sr_Res(sci->status.sres) ); 1544 } 1545 } 1546 } 1547 } 1548 1549 vg_assert(sci->status.what == SsComplete); 1550 1551 vg_assert(VG_(is_running_thread)(tid)); 1552 1553 /* Dump the syscall result back in the guest state. This is 1554 a platform-specific action. */ 1555 if (!(sci->flags & SfNoWriteResult)) 1556 putSyscallStatusIntoGuestState( tid, &sci->status, &tst->arch.vex ); 1557 1558 /* Situation now: 1559 - the guest state is now correctly modified following the syscall 1560 - modified args, original args and syscall status are still 1561 available in the syscallInfo[] entry for this syscall. 1562 1563 Now go on to do the post-syscall actions (read on down ..) 1564 */ 1565 PRINT(" "); 1566 VG_(post_syscall)(tid); 1567 PRINT("\n"); 1568 } 1569 1570 1571 /* Perform post syscall actions. The expected state on entry is 1572 precisely as at the end of VG_(client_syscall), that is: 1573 1574 - guest state up to date following the syscall 1575 - modified args, original args and syscall status are still 1576 available in the syscallInfo[] entry for this syscall. 1577 - syscall status matches what's in the guest state. 1578 1579 There are two ways to get here: the normal way -- being called by 1580 VG_(client_syscall), and the unusual way, from 1581 VG_(fixup_guest_state_after_syscall_interrupted). 1582 Darwin: there's a third way, ML_(wqthread_continue). 1583 */ 1584 void VG_(post_syscall) (ThreadId tid) 1585 { 1586 SyscallInfo* sci; 1587 const SyscallTableEntry* ent; 1588 SyscallStatus test_status; 1589 ThreadState* tst; 1590 Word sysno; 1591 1592 /* Preliminaries */ 1593 vg_assert(VG_(is_valid_tid)(tid)); 1594 vg_assert(tid >= 1 && tid < VG_N_THREADS); 1595 vg_assert(VG_(is_running_thread)(tid)); 1596 1597 tst = VG_(get_ThreadState)(tid); 1598 sci = & syscallInfo[tid]; 1599 1600 /* m_signals.sigvgkill_handler might call here even when not in 1601 a syscall. */ 1602 if (sci->status.what == SsIdle || sci->status.what == SsHandToKernel) { 1603 sci->status.what = SsIdle; 1604 return; 1605 } 1606 1607 /* Validate current syscallInfo entry. In particular we require 1608 that the current .status matches what's actually in the guest 1609 state. At least in the normal case where we have actually 1610 previously written the result into the guest state. */ 1611 vg_assert(sci->status.what == SsComplete); 1612 1613 getSyscallStatusFromGuestState( &test_status, &tst->arch.vex ); 1614 if (!(sci->flags & SfNoWriteResult)) 1615 vg_assert(eq_SyscallStatus( &sci->status, &test_status )); 1616 /* Failure of the above assertion on Darwin can indicate a problem 1617 in the syscall wrappers that pre-fail or pre-succeed the 1618 syscall, by calling SET_STATUS_Success or SET_STATUS_Failure, 1619 when they really should call SET_STATUS_from_SysRes. The former 1620 create a UNIX-class syscall result on Darwin, which may not be 1621 correct for the syscall; if that's the case then this assertion 1622 fires. See PRE(thread_fast_set_cthread_self) for an example. On 1623 non-Darwin platforms this assertion is should never fail, and this 1624 comment is completely irrelevant. */ 1625 /* Ok, looks sane */ 1626 1627 /* Get the system call number. Because the pre-handler isn't 1628 allowed to mess with it, it should be the same for both the 1629 original and potentially-modified args. */ 1630 vg_assert(sci->args.sysno == sci->orig_args.sysno); 1631 sysno = sci->args.sysno; 1632 ent = get_syscall_entry(sysno); 1633 1634 /* pre: status == Complete (asserted above) */ 1635 /* Consider either success or failure. Now run the post handler if: 1636 - it exists, and 1637 - Success or (Failure and PostOnFail is set) 1638 */ 1639 if (ent->after 1640 && ((!sr_isError(sci->status.sres)) 1641 || (sr_isError(sci->status.sres) 1642 && (sci->flags & SfPostOnFail) ))) { 1643 1644 (ent->after)( tid, &sci->args, &sci->status ); 1645 } 1646 1647 /* Because the post handler might have changed the status (eg, the 1648 post-handler for sys_open can change the result from success to 1649 failure if the kernel supplied a fd that it doesn't like), once 1650 again dump the syscall result back in the guest state.*/ 1651 if (!(sci->flags & SfNoWriteResult)) 1652 putSyscallStatusIntoGuestState( tid, &sci->status, &tst->arch.vex ); 1653 1654 /* Do any post-syscall actions required by the tool. */ 1655 if (VG_(needs).syscall_wrapper) { 1656 UWord tmpv[8]; 1657 tmpv[0] = sci->orig_args.arg1; 1658 tmpv[1] = sci->orig_args.arg2; 1659 tmpv[2] = sci->orig_args.arg3; 1660 tmpv[3] = sci->orig_args.arg4; 1661 tmpv[4] = sci->orig_args.arg5; 1662 tmpv[5] = sci->orig_args.arg6; 1663 tmpv[6] = sci->orig_args.arg7; 1664 tmpv[7] = sci->orig_args.arg8; 1665 VG_TDICT_CALL(tool_post_syscall, tid, 1666 sysno, 1667 &tmpv[0], sizeof(tmpv)/sizeof(tmpv[0]), 1668 sci->status.sres); 1669 } 1670 1671 /* The syscall is done. */ 1672 vg_assert(sci->status.what == SsComplete); 1673 sci->status.what = SsIdle; 1674 1675 /* The pre/post wrappers may have concluded that pending signals 1676 might have been created, and will have set SfPollAfter to 1677 request a poll for them once the syscall is done. */ 1678 if (sci->flags & SfPollAfter) 1679 VG_(poll_signals)(tid); 1680 1681 /* Similarly, the wrappers might have asked for a yield 1682 afterwards. */ 1683 if (sci->flags & SfYieldAfter) 1684 VG_(vg_yield)(); 1685 } 1686 1687 1688 /* --------------------------------------------------------------------- 1689 Dealing with syscalls which get interrupted by a signal: 1690 VG_(fixup_guest_state_after_syscall_interrupted) 1691 ------------------------------------------------------------------ */ 1692 1693 /* Syscalls done on behalf of the client are finally handed off to the 1694 kernel in VG_(client_syscall) above, either by calling 1695 do_syscall_for_client (the async case), or by calling 1696 VG_(do_syscall6) (the sync case). 1697 1698 If the syscall is not interrupted by a signal (it may block and 1699 later unblock, but that's irrelevant here) then those functions 1700 eventually return and so control is passed to VG_(post_syscall). 1701 NB: not sure if the sync case can actually get interrupted, as it 1702 operates with all signals masked. 1703 1704 However, the syscall may get interrupted by an async-signal. In 1705 that case do_syscall_for_client/VG_(do_syscall6) do not 1706 return. Instead we wind up in m_signals.async_sighandler. We need 1707 to fix up the guest state to make it look like the syscall was 1708 interrupted for guest. So async_sighandler calls here, and this 1709 does the fixup. Note that from here we wind up calling 1710 VG_(post_syscall) too. 1711 */ 1712 1713 1714 /* These are addresses within ML_(do_syscall_for_client_WRK). See 1715 syscall-$PLAT.S for details. 1716 */ 1717 #if defined(VGO_linux) 1718 extern const Addr ML_(blksys_setup); 1719 extern const Addr ML_(blksys_restart); 1720 extern const Addr ML_(blksys_complete); 1721 extern const Addr ML_(blksys_committed); 1722 extern const Addr ML_(blksys_finished); 1723 #elif defined(VGO_darwin) 1724 /* Darwin requires extra uglyness */ 1725 extern const Addr ML_(blksys_setup_MACH); 1726 extern const Addr ML_(blksys_restart_MACH); 1727 extern const Addr ML_(blksys_complete_MACH); 1728 extern const Addr ML_(blksys_committed_MACH); 1729 extern const Addr ML_(blksys_finished_MACH); 1730 extern const Addr ML_(blksys_setup_MDEP); 1731 extern const Addr ML_(blksys_restart_MDEP); 1732 extern const Addr ML_(blksys_complete_MDEP); 1733 extern const Addr ML_(blksys_committed_MDEP); 1734 extern const Addr ML_(blksys_finished_MDEP); 1735 extern const Addr ML_(blksys_setup_UNIX); 1736 extern const Addr ML_(blksys_restart_UNIX); 1737 extern const Addr ML_(blksys_complete_UNIX); 1738 extern const Addr ML_(blksys_committed_UNIX); 1739 extern const Addr ML_(blksys_finished_UNIX); 1740 #else 1741 # error "Unknown OS" 1742 #endif 1743 1744 1745 /* Back up guest state to restart a system call. */ 1746 1747 void ML_(fixup_guest_state_to_restart_syscall) ( ThreadArchState* arch ) 1748 { 1749 #if defined(VGP_x86_linux) 1750 arch->vex.guest_EIP -= 2; // sizeof(int $0x80) 1751 1752 /* Make sure our caller is actually sane, and we're really backing 1753 back over a syscall. 1754 1755 int $0x80 == CD 80 1756 */ 1757 { 1758 UChar *p = (UChar *)arch->vex.guest_EIP; 1759 1760 if (p[0] != 0xcd || p[1] != 0x80) 1761 VG_(message)(Vg_DebugMsg, 1762 "?! restarting over syscall at %#x %02x %02x\n", 1763 arch->vex.guest_EIP, p[0], p[1]); 1764 1765 vg_assert(p[0] == 0xcd && p[1] == 0x80); 1766 } 1767 1768 #elif defined(VGP_amd64_linux) 1769 arch->vex.guest_RIP -= 2; // sizeof(syscall) 1770 1771 /* Make sure our caller is actually sane, and we're really backing 1772 back over a syscall. 1773 1774 syscall == 0F 05 1775 */ 1776 { 1777 UChar *p = (UChar *)arch->vex.guest_RIP; 1778 1779 if (p[0] != 0x0F || p[1] != 0x05) 1780 VG_(message)(Vg_DebugMsg, 1781 "?! restarting over syscall at %#llx %02x %02x\n", 1782 arch->vex.guest_RIP, p[0], p[1]); 1783 1784 vg_assert(p[0] == 0x0F && p[1] == 0x05); 1785 } 1786 1787 #elif defined(VGP_ppc32_linux) || defined(VGP_ppc64_linux) 1788 arch->vex.guest_CIA -= 4; // sizeof(ppc32 instr) 1789 1790 /* Make sure our caller is actually sane, and we're really backing 1791 back over a syscall. 1792 1793 sc == 44 00 00 02 1794 */ 1795 { 1796 UChar *p = (UChar *)arch->vex.guest_CIA; 1797 1798 if (p[0] != 0x44 || p[1] != 0x0 || p[2] != 0x0 || p[3] != 0x02) 1799 VG_(message)(Vg_DebugMsg, 1800 "?! restarting over syscall at %#llx %02x %02x %02x %02x\n", 1801 arch->vex.guest_CIA + 0ULL, p[0], p[1], p[2], p[3]); 1802 1803 vg_assert(p[0] == 0x44 && p[1] == 0x0 && p[2] == 0x0 && p[3] == 0x2); 1804 } 1805 1806 #elif defined(VGP_arm_linux) 1807 if (arch->vex.guest_R15T & 1) { 1808 // Thumb mode. SVC is a encoded as 1809 // 1101 1111 imm8 1810 // where imm8 is the SVC number, and we only accept 0. 1811 arch->vex.guest_R15T -= 2; // sizeof(thumb 16 bit insn) 1812 UChar* p = (UChar*)(arch->vex.guest_R15T - 1); 1813 Bool valid = p[0] == 0 && p[1] == 0xDF; 1814 if (!valid) { 1815 VG_(message)(Vg_DebugMsg, 1816 "?! restarting over (Thumb) syscall that is not syscall " 1817 "at %#llx %02x %02x\n", 1818 arch->vex.guest_R15T - 1ULL, p[0], p[1]); 1819 } 1820 vg_assert(valid); 1821 // FIXME: NOTE, this really isn't right. We need to back up 1822 // ITSTATE to what it was before the SVC instruction, but we 1823 // don't know what it was. At least assert that it is now 1824 // zero, because if it is nonzero then it must also have 1825 // been nonzero for the SVC itself, which means it was 1826 // conditional. Urk. 1827 vg_assert(arch->vex.guest_ITSTATE == 0); 1828 } else { 1829 // ARM mode. SVC is encoded as 1830 // cond 1111 imm24 1831 // where imm24 is the SVC number, and we only accept 0. 1832 arch->vex.guest_R15T -= 4; // sizeof(arm instr) 1833 UChar* p = (UChar*)arch->vex.guest_R15T; 1834 Bool valid = p[0] == 0 && p[1] == 0 && p[2] == 0 1835 && (p[3] & 0xF) == 0xF; 1836 if (!valid) { 1837 VG_(message)(Vg_DebugMsg, 1838 "?! restarting over (ARM) syscall that is not syscall " 1839 "at %#llx %02x %02x %02x %02x\n", 1840 arch->vex.guest_R15T + 0ULL, p[0], p[1], p[2], p[3]); 1841 } 1842 vg_assert(valid); 1843 } 1844 1845 #elif defined(VGP_x86_darwin) 1846 arch->vex.guest_EIP = arch->vex.guest_IP_AT_SYSCALL; 1847 1848 /* Make sure our caller is actually sane, and we're really backing 1849 back over a syscall. 1850 1851 int $0x80 == CD 80 1852 int $0x81 == CD 81 1853 int $0x82 == CD 82 1854 sysenter == 0F 34 1855 */ 1856 { 1857 UChar *p = (UChar *)arch->vex.guest_EIP; 1858 Bool ok = (p[0] == 0xCD && p[1] == 0x80) 1859 || (p[0] == 0xCD && p[1] == 0x81) 1860 || (p[0] == 0xCD && p[1] == 0x82) 1861 || (p[0] == 0x0F && p[1] == 0x34); 1862 if (!ok) 1863 VG_(message)(Vg_DebugMsg, 1864 "?! restarting over syscall at %#x %02x %02x\n", 1865 arch->vex.guest_EIP, p[0], p[1]); 1866 vg_assert(ok); 1867 } 1868 1869 #elif defined(VGP_amd64_darwin) 1870 // DDD: #warning GrP fixme amd64 restart unimplemented 1871 vg_assert(0); 1872 1873 #elif defined(VGP_s390x_linux) 1874 arch->vex.guest_IA -= 2; // sizeof(syscall) 1875 1876 /* Make sure our caller is actually sane, and we're really backing 1877 back over a syscall. 1878 1879 syscall == 0A <num> 1880 */ 1881 { 1882 UChar *p = (UChar *)arch->vex.guest_IA; 1883 if (p[0] != 0x0A) 1884 VG_(message)(Vg_DebugMsg, 1885 "?! restarting over syscall at %#llx %02x %02x\n", 1886 arch->vex.guest_IA, p[0], p[1]); 1887 1888 vg_assert(p[0] == 0x0A); 1889 } 1890 #else 1891 # error "ML_(fixup_guest_state_to_restart_syscall): unknown plat" 1892 #endif 1893 } 1894 1895 /* 1896 Fix up the guest state when a syscall is interrupted by a signal 1897 and so has been forced to return 'sysret'. 1898 1899 To do this, we determine the precise state of the syscall by 1900 looking at the (real) IP at the time the signal happened. The 1901 syscall sequence looks like: 1902 1903 1. unblock signals 1904 2. perform syscall 1905 3. save result to guest state (EAX, RAX, R3+CR0.SO) 1906 4. re-block signals 1907 1908 If a signal 1909 happens at Then Why? 1910 [1-2) restart nothing has happened (restart syscall) 1911 [2] restart syscall hasn't started, or kernel wants to restart 1912 [2-3) save syscall complete, but results not saved 1913 [3-4) syscall complete, results saved 1914 1915 Sometimes we never want to restart an interrupted syscall (because 1916 sigaction says not to), so we only restart if "restart" is True. 1917 1918 This will also call VG_(post_syscall) if the syscall has actually 1919 completed (either because it was interrupted, or because it 1920 actually finished). It will not call VG_(post_syscall) if the 1921 syscall is set up for restart, which means that the pre-wrapper may 1922 get called multiple times. 1923 */ 1924 1925 void 1926 VG_(fixup_guest_state_after_syscall_interrupted)( ThreadId tid, 1927 Addr ip, 1928 SysRes sres, 1929 Bool restart) 1930 { 1931 /* Note that we don't know the syscall number here, since (1) in 1932 general there's no reliable way to get hold of it short of 1933 stashing it in the guest state before the syscall, and (2) in 1934 any case we don't need to know it for the actions done by this 1935 routine. 1936 1937 Furthermore, 'sres' is only used in the case where the syscall 1938 is complete, but the result has not been committed to the guest 1939 state yet. In any other situation it will be meaningless and 1940 therefore ignored. */ 1941 1942 ThreadState* tst; 1943 SyscallStatus canonical; 1944 ThreadArchState* th_regs; 1945 SyscallInfo* sci; 1946 1947 /* Compute some Booleans indicating which range we're in. */ 1948 Bool outside_range, 1949 in_setup_to_restart, // [1,2) in the .S files 1950 at_restart, // [2] in the .S files 1951 in_complete_to_committed, // [3,4) in the .S files 1952 in_committed_to_finished; // [4,5) in the .S files 1953 1954 # if defined(VGO_linux) 1955 outside_range 1956 = ip < ML_(blksys_setup) || ip >= ML_(blksys_finished); 1957 in_setup_to_restart 1958 = ip >= ML_(blksys_setup) && ip < ML_(blksys_restart); 1959 at_restart 1960 = ip == ML_(blksys_restart); 1961 in_complete_to_committed 1962 = ip >= ML_(blksys_complete) && ip < ML_(blksys_committed); 1963 in_committed_to_finished 1964 = ip >= ML_(blksys_committed) && ip < ML_(blksys_finished); 1965 # elif defined(VGO_darwin) 1966 outside_range 1967 = (ip < ML_(blksys_setup_MACH) || ip >= ML_(blksys_finished_MACH)) 1968 && (ip < ML_(blksys_setup_MDEP) || ip >= ML_(blksys_finished_MDEP)) 1969 && (ip < ML_(blksys_setup_UNIX) || ip >= ML_(blksys_finished_UNIX)); 1970 in_setup_to_restart 1971 = (ip >= ML_(blksys_setup_MACH) && ip < ML_(blksys_restart_MACH)) 1972 || (ip >= ML_(blksys_setup_MDEP) && ip < ML_(blksys_restart_MDEP)) 1973 || (ip >= ML_(blksys_setup_UNIX) && ip < ML_(blksys_restart_UNIX)); 1974 at_restart 1975 = (ip == ML_(blksys_restart_MACH)) 1976 || (ip == ML_(blksys_restart_MDEP)) 1977 || (ip == ML_(blksys_restart_UNIX)); 1978 in_complete_to_committed 1979 = (ip >= ML_(blksys_complete_MACH) && ip < ML_(blksys_committed_MACH)) 1980 || (ip >= ML_(blksys_complete_MDEP) && ip < ML_(blksys_committed_MDEP)) 1981 || (ip >= ML_(blksys_complete_UNIX) && ip < ML_(blksys_committed_UNIX)); 1982 in_committed_to_finished 1983 = (ip >= ML_(blksys_committed_MACH) && ip < ML_(blksys_finished_MACH)) 1984 || (ip >= ML_(blksys_committed_MDEP) && ip < ML_(blksys_finished_MDEP)) 1985 || (ip >= ML_(blksys_committed_UNIX) && ip < ML_(blksys_finished_UNIX)); 1986 /* Wasn't that just So Much Fun? Does your head hurt yet? Mine does. */ 1987 # else 1988 # error "Unknown OS" 1989 # endif 1990 1991 if (VG_(clo_trace_signals)) 1992 VG_(message)( Vg_DebugMsg, 1993 "interrupted_syscall: tid=%d, ip=0x%llx, " 1994 "restart=%s, sres.isErr=%s, sres.val=%lld\n", 1995 (Int)tid, 1996 (ULong)ip, 1997 restart ? "True" : "False", 1998 sr_isError(sres) ? "True" : "False", 1999 (Long)(sr_isError(sres) ? sr_Err(sres) : sr_Res(sres)) ); 2000 2001 vg_assert(VG_(is_valid_tid)(tid)); 2002 vg_assert(tid >= 1 && tid < VG_N_THREADS); 2003 vg_assert(VG_(is_running_thread)(tid)); 2004 2005 tst = VG_(get_ThreadState)(tid); 2006 th_regs = &tst->arch; 2007 sci = & syscallInfo[tid]; 2008 2009 /* Figure out what the state of the syscall was by examining the 2010 (real) IP at the time of the signal, and act accordingly. */ 2011 if (outside_range) { 2012 if (VG_(clo_trace_signals)) 2013 VG_(message)( Vg_DebugMsg, 2014 " not in syscall at all: hmm, very suspicious\n" ); 2015 /* Looks like we weren't in a syscall at all. Hmm. */ 2016 vg_assert(sci->status.what != SsIdle); 2017 return; 2018 } 2019 2020 /* We should not be here unless this thread had first started up 2021 the machinery for a syscall by calling VG_(client_syscall). 2022 Hence: */ 2023 vg_assert(sci->status.what != SsIdle); 2024 2025 /* now, do one of four fixup actions, depending on where the IP has 2026 got to. */ 2027 2028 if (in_setup_to_restart) { 2029 /* syscall hasn't even started; go around again */ 2030 if (VG_(clo_trace_signals)) 2031 VG_(message)( Vg_DebugMsg, " not started: restarting\n"); 2032 vg_assert(sci->status.what == SsHandToKernel); 2033 ML_(fixup_guest_state_to_restart_syscall)(th_regs); 2034 } 2035 2036 else 2037 if (at_restart) { 2038 /* We're either about to run the syscall, or it was interrupted 2039 and the kernel restarted it. Restart if asked, otherwise 2040 EINTR it. */ 2041 if (restart) { 2042 if (VG_(clo_trace_signals)) 2043 VG_(message)( Vg_DebugMsg, " at syscall instr: restarting\n"); 2044 ML_(fixup_guest_state_to_restart_syscall)(th_regs); 2045 } else { 2046 if (VG_(clo_trace_signals)) 2047 VG_(message)( Vg_DebugMsg, " at syscall instr: returning EINTR\n"); 2048 canonical = convert_SysRes_to_SyscallStatus( 2049 VG_(mk_SysRes_Error)( VKI_EINTR ) 2050 ); 2051 if (!(sci->flags & SfNoWriteResult)) 2052 putSyscallStatusIntoGuestState( tid, &canonical, &th_regs->vex ); 2053 sci->status = canonical; 2054 VG_(post_syscall)(tid); 2055 } 2056 } 2057 2058 else 2059 if (in_complete_to_committed) { 2060 /* Syscall complete, but result hasn't been written back yet. 2061 Write the SysRes we were supplied with back to the guest 2062 state. */ 2063 if (VG_(clo_trace_signals)) 2064 VG_(message)( Vg_DebugMsg, 2065 " completed, but uncommitted: committing\n"); 2066 canonical = convert_SysRes_to_SyscallStatus( sres ); 2067 if (!(sci->flags & SfNoWriteResult)) 2068 putSyscallStatusIntoGuestState( tid, &canonical, &th_regs->vex ); 2069 sci->status = canonical; 2070 VG_(post_syscall)(tid); 2071 } 2072 2073 else 2074 if (in_committed_to_finished) { 2075 /* Result committed, but the signal mask has not been restored; 2076 we expect our caller (the signal handler) will have fixed 2077 this up. */ 2078 if (VG_(clo_trace_signals)) 2079 VG_(message)( Vg_DebugMsg, 2080 " completed and committed: nothing to do\n"); 2081 getSyscallStatusFromGuestState( &sci->status, &th_regs->vex ); 2082 vg_assert(sci->status.what == SsComplete); 2083 VG_(post_syscall)(tid); 2084 } 2085 2086 else 2087 VG_(core_panic)("?? strange syscall interrupt state?"); 2088 2089 /* In all cases, the syscall is now finished (even if we called 2090 ML_(fixup_guest_state_to_restart_syscall), since that just 2091 re-positions the guest's IP for another go at it). So we need 2092 to record that fact. */ 2093 sci->status.what = SsIdle; 2094 } 2095 2096 2097 #if defined(VGO_darwin) 2098 // Clean up after workq_ops(WQOPS_THREAD_RETURN) jumped to wqthread_hijack. 2099 // This is similar to VG_(fixup_guest_state_after_syscall_interrupted). 2100 // This longjmps back to the scheduler. 2101 void ML_(wqthread_continue_NORETURN)(ThreadId tid) 2102 { 2103 ThreadState* tst; 2104 SyscallInfo* sci; 2105 2106 VG_(acquire_BigLock)(tid, "wqthread_continue_NORETURN"); 2107 2108 PRINT("SYSCALL[%d,%d](%s) workq_ops() starting new workqueue item\n", 2109 VG_(getpid)(), tid, VG_SYSNUM_STRING(__NR_workq_ops)); 2110 2111 vg_assert(VG_(is_valid_tid)(tid)); 2112 vg_assert(tid >= 1 && tid < VG_N_THREADS); 2113 vg_assert(VG_(is_running_thread)(tid)); 2114 2115 tst = VG_(get_ThreadState)(tid); 2116 sci = & syscallInfo[tid]; 2117 vg_assert(sci->status.what != SsIdle); 2118 vg_assert(tst->os_state.wq_jmpbuf_valid); // check this BEFORE post_syscall 2119 2120 // Pretend the syscall completed normally, but don't touch the thread state. 2121 sci->status = convert_SysRes_to_SyscallStatus( VG_(mk_SysRes_Success)(0) ); 2122 sci->flags |= SfNoWriteResult; 2123 VG_(post_syscall)(tid); 2124 2125 sci->status.what = SsIdle; 2126 2127 vg_assert(tst->sched_jmpbuf_valid); 2128 VG_MINIMAL_LONGJMP(tst->sched_jmpbuf); 2129 2130 /* NOTREACHED */ 2131 vg_assert(0); 2132 } 2133 #endif 2134 2135 2136 /* --------------------------------------------------------------------- 2137 A place to store the where-to-call-when-really-done pointer 2138 ------------------------------------------------------------------ */ 2139 2140 // When the final thread is done, where shall I call to shutdown the 2141 // system cleanly? Is set once at startup (in m_main) and never 2142 // changes after that. Is basically a pointer to the exit 2143 // continuation. This is all just a nasty hack to avoid calling 2144 // directly from m_syswrap to m_main at exit, since that would cause 2145 // m_main to become part of a module cycle, which is silly. 2146 void (* VG_(address_of_m_main_shutdown_actions_NORETURN) ) 2147 (ThreadId,VgSchedReturnCode) 2148 = NULL; 2149 2150 /*--------------------------------------------------------------------*/ 2151 /*--- end ---*/ 2152 /*--------------------------------------------------------------------*/ 2153