1 // Copyright (c) 2010 Google Inc.
2 // All rights reserved.
3 //
4 // Redistribution and use in source and binary forms, with or without
5 // modification, are permitted provided that the following conditions are
6 // met:
7 //
8 // * Redistributions of source code must retain the above copyright
9 // notice, this list of conditions and the following disclaimer.
10 // * Redistributions in binary form must reproduce the above
11 // copyright notice, this list of conditions and the following disclaimer
12 // in the documentation and/or other materials provided with the
13 // distribution.
14 // * Neither the name of Google Inc. nor the names of its
15 // contributors may be used to endorse or promote products derived from
16 // this software without specific prior written permission.
17 //
18 // THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
19 // "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
20 // LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
21 // A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
22 // OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
23 // SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
24 // LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
25 // DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
26 // THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
27 // (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
28 // OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
29
30 // stackwalker_x86.cc: x86-specific stackwalker.
31 //
32 // See stackwalker_x86.h for documentation.
33 //
34 // Author: Mark Mentovai
35
36 #include <assert.h>
37 #include <string>
38
39 #include "common/scoped_ptr.h"
40 #include "google_breakpad/processor/call_stack.h"
41 #include "google_breakpad/processor/code_modules.h"
42 #include "google_breakpad/processor/memory_region.h"
43 #include "google_breakpad/processor/source_line_resolver_interface.h"
44 #include "google_breakpad/processor/stack_frame_cpu.h"
45 #include "processor/logging.h"
46 #include "processor/postfix_evaluator-inl.h"
47 #include "processor/stackwalker_x86.h"
48 #include "processor/windows_frame_info.h"
49 #include "processor/cfi_frame_info.h"
50
51 namespace google_breakpad {
52
53 // Max reasonable size for a single x86 frame is 128 KB. This value is used in
54 // a heuristic for recovering of the EBP chain after a scan for return address.
55 // This value is based on a stack frame size histogram built for a set of
56 // popular third party libraries which suggests that 99.5% of all frames are
57 // smaller than 128 KB.
58 static const uint32_t kMaxReasonableGapBetweenFrames = 128 * 1024;
59
60 const StackwalkerX86::CFIWalker::RegisterSet
61 StackwalkerX86::cfi_register_map_[] = {
62 // It may seem like $eip and $esp are callee-saves, because (with Unix or
63 // cdecl calling conventions) the callee is responsible for having them
64 // restored upon return. But the callee_saves flags here really means
65 // that the walker should assume they're unchanged if the CFI doesn't
66 // mention them, which is clearly wrong for $eip and $esp.
67 { "$eip", ".ra", false,
68 StackFrameX86::CONTEXT_VALID_EIP, &MDRawContextX86::eip },
69 { "$esp", ".cfa", false,
70 StackFrameX86::CONTEXT_VALID_ESP, &MDRawContextX86::esp },
71 { "$ebp", NULL, true,
72 StackFrameX86::CONTEXT_VALID_EBP, &MDRawContextX86::ebp },
73 { "$eax", NULL, false,
74 StackFrameX86::CONTEXT_VALID_EAX, &MDRawContextX86::eax },
75 { "$ebx", NULL, true,
76 StackFrameX86::CONTEXT_VALID_EBX, &MDRawContextX86::ebx },
77 { "$ecx", NULL, false,
78 StackFrameX86::CONTEXT_VALID_ECX, &MDRawContextX86::ecx },
79 { "$edx", NULL, false,
80 StackFrameX86::CONTEXT_VALID_EDX, &MDRawContextX86::edx },
81 { "$esi", NULL, true,
82 StackFrameX86::CONTEXT_VALID_ESI, &MDRawContextX86::esi },
83 { "$edi", NULL, true,
84 StackFrameX86::CONTEXT_VALID_EDI, &MDRawContextX86::edi },
85 };
86
StackwalkerX86(const SystemInfo * system_info,const MDRawContextX86 * context,MemoryRegion * memory,const CodeModules * modules,StackFrameSymbolizer * resolver_helper)87 StackwalkerX86::StackwalkerX86(const SystemInfo* system_info,
88 const MDRawContextX86* context,
89 MemoryRegion* memory,
90 const CodeModules* modules,
91 StackFrameSymbolizer* resolver_helper)
92 : Stackwalker(system_info, memory, modules, resolver_helper),
93 context_(context),
94 cfi_walker_(cfi_register_map_,
95 (sizeof(cfi_register_map_) / sizeof(cfi_register_map_[0]))) {
96 if (memory_ && memory_->GetBase() + memory_->GetSize() - 1 > 0xffffffff) {
97 // The x86 is a 32-bit CPU, the limits of the supplied stack are invalid.
98 // Mark memory_ = NULL, which will cause stackwalking to fail.
99 BPLOG(ERROR) << "Memory out of range for stackwalking: " <<
100 HexString(memory_->GetBase()) << "+" <<
101 HexString(memory_->GetSize());
102 memory_ = NULL;
103 }
104 }
105
~StackFrameX86()106 StackFrameX86::~StackFrameX86() {
107 if (windows_frame_info)
108 delete windows_frame_info;
109 windows_frame_info = NULL;
110 if (cfi_frame_info)
111 delete cfi_frame_info;
112 cfi_frame_info = NULL;
113 }
114
ReturnAddress() const115 uint64_t StackFrameX86::ReturnAddress() const {
116 assert(context_validity & StackFrameX86::CONTEXT_VALID_EIP);
117 return context.eip;
118 }
119
GetContextFrame()120 StackFrame* StackwalkerX86::GetContextFrame() {
121 if (!context_) {
122 BPLOG(ERROR) << "Can't get context frame without context";
123 return NULL;
124 }
125
126 StackFrameX86* frame = new StackFrameX86();
127
128 // The instruction pointer is stored directly in a register, so pull it
129 // straight out of the CPU context structure.
130 frame->context = *context_;
131 frame->context_validity = StackFrameX86::CONTEXT_VALID_ALL;
132 frame->trust = StackFrame::FRAME_TRUST_CONTEXT;
133 frame->instruction = frame->context.eip;
134
135 return frame;
136 }
137
GetCallerByWindowsFrameInfo(const vector<StackFrame * > & frames,WindowsFrameInfo * last_frame_info,bool stack_scan_allowed)138 StackFrameX86* StackwalkerX86::GetCallerByWindowsFrameInfo(
139 const vector<StackFrame*> &frames,
140 WindowsFrameInfo* last_frame_info,
141 bool stack_scan_allowed) {
142 StackFrame::FrameTrust trust = StackFrame::FRAME_TRUST_NONE;
143
144 StackFrameX86* last_frame = static_cast<StackFrameX86*>(frames.back());
145
146 // Save the stack walking info we found, in case we need it later to
147 // find the callee of the frame we're constructing now.
148 last_frame->windows_frame_info = last_frame_info;
149
150 // This function only covers the full STACK WIN case. If
151 // last_frame_info is VALID_PARAMETER_SIZE-only, then we should
152 // assume the traditional frame format or use some other strategy.
153 if (last_frame_info->valid != WindowsFrameInfo::VALID_ALL)
154 return NULL;
155
156 // This stackwalker sets each frame's %esp to its value immediately prior
157 // to the CALL into the callee. This means that %esp points to the last
158 // callee argument pushed onto the stack, which may not be where %esp points
159 // after the callee returns. Specifically, the value is correct for the
160 // cdecl calling convention, but not other conventions. The cdecl
161 // convention requires a caller to pop its callee's arguments from the
162 // stack after the callee returns. This is usually accomplished by adding
163 // the known size of the arguments to %esp. Other calling conventions,
164 // including stdcall, thiscall, and fastcall, require the callee to pop any
165 // parameters stored on the stack before returning. This is usually
166 // accomplished by using the RET n instruction, which pops n bytes off
167 // the stack after popping the return address.
168 //
169 // Because each frame's %esp will point to a location on the stack after
170 // callee arguments have been PUSHed, when locating things in a stack frame
171 // relative to %esp, the size of the arguments to the callee need to be
172 // taken into account. This seems a little bit unclean, but it's better
173 // than the alternative, which would need to take these same things into
174 // account, but only for cdecl functions. With this implementation, we get
175 // to be agnostic about each function's calling convention. Furthermore,
176 // this is how Windows debugging tools work, so it means that the %esp
177 // values produced by this stackwalker directly correspond to the %esp
178 // values you'll see there.
179 //
180 // If the last frame has no callee (because it's the context frame), just
181 // set the callee parameter size to 0: the stack pointer can't point to
182 // callee arguments because there's no callee. This is correct as long
183 // as the context wasn't captured while arguments were being pushed for
184 // a function call. Note that there may be functions whose parameter sizes
185 // are unknown, 0 is also used in that case. When that happens, it should
186 // be possible to walk to the next frame without reference to %esp.
187
188 uint32_t last_frame_callee_parameter_size = 0;
189 int frames_already_walked = frames.size();
190 if (frames_already_walked >= 2) {
191 const StackFrameX86* last_frame_callee
192 = static_cast<StackFrameX86*>(frames[frames_already_walked - 2]);
193 WindowsFrameInfo* last_frame_callee_info
194 = last_frame_callee->windows_frame_info;
195 if (last_frame_callee_info &&
196 (last_frame_callee_info->valid
197 & WindowsFrameInfo::VALID_PARAMETER_SIZE)) {
198 last_frame_callee_parameter_size =
199 last_frame_callee_info->parameter_size;
200 }
201 }
202
203 // Set up the dictionary for the PostfixEvaluator. %ebp and %esp are used
204 // in each program string, and their previous values are known, so set them
205 // here.
206 PostfixEvaluator<uint32_t>::DictionaryType dictionary;
207 // Provide the current register values.
208 dictionary["$ebp"] = last_frame->context.ebp;
209 dictionary["$esp"] = last_frame->context.esp;
210 // Provide constants from the debug info for last_frame and its callee.
211 // .cbCalleeParams is a Breakpad extension that allows us to use the
212 // PostfixEvaluator engine when certain types of debugging information
213 // are present without having to write the constants into the program
214 // string as literals.
215 dictionary[".cbCalleeParams"] = last_frame_callee_parameter_size;
216 dictionary[".cbSavedRegs"] = last_frame_info->saved_register_size;
217 dictionary[".cbLocals"] = last_frame_info->local_size;
218
219 uint32_t raSearchStart = last_frame->context.esp +
220 last_frame_callee_parameter_size +
221 last_frame_info->local_size +
222 last_frame_info->saved_register_size;
223
224 uint32_t raSearchStartOld = raSearchStart;
225 uint32_t found = 0; // dummy value
226 // Scan up to three words above the calculated search value, in case
227 // the stack was aligned to a quadword boundary.
228 //
229 // TODO(ivan.penkov): Consider cleaning up the scan for return address that
230 // follows. The purpose of this scan is to adjust the .raSearchStart
231 // calculation (which is based on register %esp) in the cases where register
232 // %esp may have been aligned (up to a quadword). There are two problems
233 // with this approach:
234 // 1) In practice, 64 byte boundary alignment is seen which clearly can not
235 // be handled by a three word scan.
236 // 2) A search for a return address is "guesswork" by definition because
237 // the results will be different depending on what is left on the stack
238 // from previous executions.
239 // So, basically, the results from this scan should be ignored if other means
240 // for calculation of the value of .raSearchStart are available.
241 if (ScanForReturnAddress(raSearchStart, &raSearchStart, &found, 3) &&
242 last_frame->trust == StackFrame::FRAME_TRUST_CONTEXT &&
243 last_frame->windows_frame_info != NULL &&
244 last_frame_info->type_ == WindowsFrameInfo::STACK_INFO_FPO &&
245 raSearchStartOld == raSearchStart &&
246 found == last_frame->context.eip) {
247 // The context frame represents an FPO-optimized Windows system call.
248 // On the top of the stack we have a pointer to the current instruction.
249 // This means that the callee has returned but the return address is still
250 // on the top of the stack which is very atypical situaltion.
251 // Skip one slot from the stack and do another scan in order to get the
252 // actual return address.
253 raSearchStart += 4;
254 ScanForReturnAddress(raSearchStart, &raSearchStart, &found, 3);
255 }
256
257 dictionary[".cbParams"] = last_frame_info->parameter_size;
258
259 // Decide what type of program string to use. The program string is in
260 // postfix notation and will be passed to PostfixEvaluator::Evaluate.
261 // Given the dictionary and the program string, it is possible to compute
262 // the return address and the values of other registers in the calling
263 // function. Because of bugs described below, the stack may need to be
264 // scanned for these values. The results of program string evaluation
265 // will be used to determine whether to scan for better values.
266 string program_string;
267 bool recover_ebp = true;
268
269 trust = StackFrame::FRAME_TRUST_CFI;
270 if (!last_frame_info->program_string.empty()) {
271 // The FPO data has its own program string, which will tell us how to
272 // get to the caller frame, and may even fill in the values of
273 // nonvolatile registers and provide pointers to local variables and
274 // parameters. In some cases, particularly with program strings that use
275 // .raSearchStart, the stack may need to be scanned afterward.
276 program_string = last_frame_info->program_string;
277 } else if (last_frame_info->allocates_base_pointer) {
278 // The function corresponding to the last frame doesn't use the frame
279 // pointer for conventional purposes, but it does allocate a new
280 // frame pointer and use it for its own purposes. Its callee's
281 // information is still accessed relative to %esp, and the previous
282 // value of %ebp can be recovered from a location in its stack frame,
283 // within the saved-register area.
284 //
285 // Functions that fall into this category use the %ebp register for
286 // a purpose other than the frame pointer. They restore the caller's
287 // %ebp before returning. These functions create their stack frame
288 // after a CALL by decrementing the stack pointer in an amount
289 // sufficient to store local variables, and then PUSHing saved
290 // registers onto the stack. Arguments to a callee function, if any,
291 // are PUSHed after that. Walking up to the caller, therefore,
292 // can be done solely with calculations relative to the stack pointer
293 // (%esp). The return address is recovered from the memory location
294 // above the known sizes of the callee's parameters, saved registers,
295 // and locals. The caller's stack pointer (the value of %esp when
296 // the caller executed CALL) is the location immediately above the
297 // saved return address. The saved value of %ebp to be restored for
298 // the caller is at a known location in the saved-register area of
299 // the stack frame.
300 //
301 // For this type of frame, MSVC 14 (from Visual Studio 8/2005) in
302 // link-time code generation mode (/LTCG and /GL) can generate erroneous
303 // debugging data. The reported size of saved registers can be 0,
304 // which is clearly an error because these frames must, at the very
305 // least, save %ebp. For this reason, in addition to those given above
306 // about the use of .raSearchStart, the stack may need to be scanned
307 // for a better return address and a better frame pointer after the
308 // program string is evaluated.
309 //
310 // %eip_new = *(%esp_old + callee_params + saved_regs + locals)
311 // %ebp_new = *(%esp_old + callee_params + saved_regs - 8)
312 // %esp_new = %esp_old + callee_params + saved_regs + locals + 4
313 program_string = "$eip .raSearchStart ^ = "
314 "$ebp $esp .cbCalleeParams + .cbSavedRegs + 8 - ^ = "
315 "$esp .raSearchStart 4 + =";
316 } else {
317 // The function corresponding to the last frame doesn't use %ebp at
318 // all. The callee frame is located relative to %esp.
319 //
320 // The called procedure's instruction pointer and stack pointer are
321 // recovered in the same way as the case above, except that no
322 // frame pointer (%ebp) is used at all, so it is not saved anywhere
323 // in the callee's stack frame and does not need to be recovered.
324 // Because %ebp wasn't used in the callee, whatever value it has
325 // is the value that it had in the caller, so it can be carried
326 // straight through without bringing its validity into question.
327 //
328 // Because of the use of .raSearchStart, the stack will possibly be
329 // examined to locate a better return address after program string
330 // evaluation. The stack will not be examined to locate a saved
331 // %ebp value, because these frames do not save (or use) %ebp.
332 //
333 // %eip_new = *(%esp_old + callee_params + saved_regs + locals)
334 // %esp_new = %esp_old + callee_params + saved_regs + locals + 4
335 // %ebp_new = %ebp_old
336 program_string = "$eip .raSearchStart ^ = "
337 "$esp .raSearchStart 4 + =";
338 recover_ebp = false;
339 }
340
341 // Check for alignment operators in the program string. If alignment
342 // operators are found, then current %ebp must be valid and it is the only
343 // reliable data point that can be used for getting to the previous frame.
344 // E.g. the .raSearchStart calculation (above) is based on %esp and since
345 // %esp was aligned in the current frame (which is a lossy operation) the
346 // calculated value of .raSearchStart cannot be correct and should not be
347 // used. Instead .raSearchStart must be calculated based on %ebp.
348 // The code that follows assumes that .raSearchStart is supposed to point
349 // at the saved return address (ebp + 4).
350 // For some more details on this topic, take a look at the following thread:
351 // https://groups.google.com/forum/#!topic/google-breakpad-dev/ZP1FA9B1JjM
352 if ((StackFrameX86::CONTEXT_VALID_EBP & last_frame->context_validity) != 0 &&
353 program_string.find('@') != string::npos) {
354 raSearchStart = last_frame->context.ebp + 4;
355 }
356
357 // The difference between raSearch and raSearchStart is unknown,
358 // but making them the same seems to work well in practice.
359 dictionary[".raSearchStart"] = raSearchStart;
360 dictionary[".raSearch"] = raSearchStart;
361
362 // Now crank it out, making sure that the program string set at least the
363 // two required variables.
364 PostfixEvaluator<uint32_t> evaluator =
365 PostfixEvaluator<uint32_t>(&dictionary, memory_);
366 PostfixEvaluator<uint32_t>::DictionaryValidityType dictionary_validity;
367 if (!evaluator.Evaluate(program_string, &dictionary_validity) ||
368 dictionary_validity.find("$eip") == dictionary_validity.end() ||
369 dictionary_validity.find("$esp") == dictionary_validity.end()) {
370 // Program string evaluation failed. It may be that %eip is not somewhere
371 // with stack frame info, and %ebp is pointing to non-stack memory, so
372 // our evaluation couldn't succeed. We'll scan the stack for a return
373 // address. This can happen if the stack is in a module for which
374 // we don't have symbols, and that module is compiled without a
375 // frame pointer.
376 uint32_t location_start = last_frame->context.esp;
377 uint32_t location, eip;
378 if (!stack_scan_allowed
379 || !ScanForReturnAddress(location_start, &location, &eip,
380 frames.size() == 1 /* is_context_frame */)) {
381 // if we can't find an instruction pointer even with stack scanning,
382 // give up.
383 return NULL;
384 }
385
386 // This seems like a reasonable return address. Since program string
387 // evaluation failed, use it and set %esp to the location above the
388 // one where the return address was found.
389 dictionary["$eip"] = eip;
390 dictionary["$esp"] = location + 4;
391 trust = StackFrame::FRAME_TRUST_SCAN;
392 }
393
394 // Since this stack frame did not use %ebp in a traditional way,
395 // locating the return address isn't entirely deterministic. In that
396 // case, the stack can be scanned to locate the return address.
397 //
398 // However, if program string evaluation resulted in both %eip and
399 // %ebp values of 0, trust that the end of the stack has been
400 // reached and don't scan for anything else.
401 if (dictionary["$eip"] != 0 || dictionary["$ebp"] != 0) {
402 int offset = 0;
403
404 // This scan can only be done if a CodeModules object is available, to
405 // check that candidate return addresses are in fact inside a module.
406 //
407 // TODO(mmentovai): This ignores dynamically-generated code. One possible
408 // solution is to check the minidump's memory map to see if the candidate
409 // %eip value comes from a mapped executable page, although this would
410 // require dumps that contain MINIDUMP_MEMORY_INFO, which the Breakpad
411 // client doesn't currently write (it would need to call MiniDumpWriteDump
412 // with the MiniDumpWithFullMemoryInfo type bit set). Even given this
413 // ability, older OSes (pre-XP SP2) and CPUs (pre-P4) don't enforce
414 // an independent execute privilege on memory pages.
415
416 uint32_t eip = dictionary["$eip"];
417 if (modules_ && !modules_->GetModuleForAddress(eip)) {
418 // The instruction pointer at .raSearchStart was invalid, so start
419 // looking one 32-bit word above that location.
420 uint32_t location_start = dictionary[".raSearchStart"] + 4;
421 uint32_t location;
422 if (stack_scan_allowed
423 && ScanForReturnAddress(location_start, &location, &eip,
424 frames.size() == 1 /* is_context_frame */)) {
425 // This is a better return address that what program string
426 // evaluation found. Use it, and set %esp to the location above the
427 // one where the return address was found.
428 dictionary["$eip"] = eip;
429 dictionary["$esp"] = location + 4;
430 offset = location - location_start;
431 trust = StackFrame::FRAME_TRUST_CFI_SCAN;
432 }
433 }
434
435 if (recover_ebp) {
436 // When trying to recover the previous value of the frame pointer (%ebp),
437 // start looking at the lowest possible address in the saved-register
438 // area, and look at the entire saved register area, increased by the
439 // size of |offset| to account for additional data that may be on the
440 // stack. The scan is performed from the highest possible address to
441 // the lowest, because the expectation is that the function's prolog
442 // would have saved %ebp early.
443 uint32_t ebp = dictionary["$ebp"];
444
445 // When a scan for return address is used, it is possible to skip one or
446 // more frames (when return address is not in a known module). One
447 // indication for skipped frames is when the value of %ebp is lower than
448 // the location of the return address on the stack
449 bool has_skipped_frames =
450 (trust != StackFrame::FRAME_TRUST_CFI && ebp <= raSearchStart + offset);
451
452 uint32_t value; // throwaway variable to check pointer validity
453 if (has_skipped_frames || !memory_->GetMemoryAtAddress(ebp, &value)) {
454 int fp_search_bytes = last_frame_info->saved_register_size + offset;
455 uint32_t location_end = last_frame->context.esp +
456 last_frame_callee_parameter_size;
457
458 for (uint32_t location = location_end + fp_search_bytes;
459 location >= location_end;
460 location -= 4) {
461 if (!memory_->GetMemoryAtAddress(location, &ebp))
462 break;
463
464 if (memory_->GetMemoryAtAddress(ebp, &value)) {
465 // The candidate value is a pointer to the same memory region
466 // (the stack). Prefer it as a recovered %ebp result.
467 dictionary["$ebp"] = ebp;
468 break;
469 }
470 }
471 }
472 }
473 }
474
475 // Create a new stack frame (ownership will be transferred to the caller)
476 // and fill it in.
477 StackFrameX86* frame = new StackFrameX86();
478
479 frame->trust = trust;
480 frame->context = last_frame->context;
481 frame->context.eip = dictionary["$eip"];
482 frame->context.esp = dictionary["$esp"];
483 frame->context.ebp = dictionary["$ebp"];
484 frame->context_validity = StackFrameX86::CONTEXT_VALID_EIP |
485 StackFrameX86::CONTEXT_VALID_ESP |
486 StackFrameX86::CONTEXT_VALID_EBP;
487
488 // These are nonvolatile (callee-save) registers, and the program string
489 // may have filled them in.
490 if (dictionary_validity.find("$ebx") != dictionary_validity.end()) {
491 frame->context.ebx = dictionary["$ebx"];
492 frame->context_validity |= StackFrameX86::CONTEXT_VALID_EBX;
493 }
494 if (dictionary_validity.find("$esi") != dictionary_validity.end()) {
495 frame->context.esi = dictionary["$esi"];
496 frame->context_validity |= StackFrameX86::CONTEXT_VALID_ESI;
497 }
498 if (dictionary_validity.find("$edi") != dictionary_validity.end()) {
499 frame->context.edi = dictionary["$edi"];
500 frame->context_validity |= StackFrameX86::CONTEXT_VALID_EDI;
501 }
502
503 return frame;
504 }
505
GetCallerByCFIFrameInfo(const vector<StackFrame * > & frames,CFIFrameInfo * cfi_frame_info)506 StackFrameX86* StackwalkerX86::GetCallerByCFIFrameInfo(
507 const vector<StackFrame*> &frames,
508 CFIFrameInfo* cfi_frame_info) {
509 StackFrameX86* last_frame = static_cast<StackFrameX86*>(frames.back());
510 last_frame->cfi_frame_info = cfi_frame_info;
511
512 scoped_ptr<StackFrameX86> frame(new StackFrameX86());
513 if (!cfi_walker_
514 .FindCallerRegisters(*memory_, *cfi_frame_info,
515 last_frame->context, last_frame->context_validity,
516 &frame->context, &frame->context_validity))
517 return NULL;
518
519 // Make sure we recovered all the essentials.
520 static const int essentials = (StackFrameX86::CONTEXT_VALID_EIP
521 | StackFrameX86::CONTEXT_VALID_ESP
522 | StackFrameX86::CONTEXT_VALID_EBP);
523 if ((frame->context_validity & essentials) != essentials)
524 return NULL;
525
526 frame->trust = StackFrame::FRAME_TRUST_CFI;
527
528 return frame.release();
529 }
530
GetCallerByEBPAtBase(const vector<StackFrame * > & frames,bool stack_scan_allowed)531 StackFrameX86* StackwalkerX86::GetCallerByEBPAtBase(
532 const vector<StackFrame*> &frames,
533 bool stack_scan_allowed) {
534 StackFrame::FrameTrust trust;
535 StackFrameX86* last_frame = static_cast<StackFrameX86*>(frames.back());
536 uint32_t last_esp = last_frame->context.esp;
537 uint32_t last_ebp = last_frame->context.ebp;
538
539 // Assume that the standard %ebp-using x86 calling convention is in
540 // use.
541 //
542 // The typical x86 calling convention, when frame pointers are present,
543 // is for the calling procedure to use CALL, which pushes the return
544 // address onto the stack and sets the instruction pointer (%eip) to
545 // the entry point of the called routine. The called routine then
546 // PUSHes the calling routine's frame pointer (%ebp) onto the stack
547 // before copying the stack pointer (%esp) to the frame pointer (%ebp).
548 // Therefore, the calling procedure's frame pointer is always available
549 // by dereferencing the called procedure's frame pointer, and the return
550 // address is always available at the memory location immediately above
551 // the address pointed to by the called procedure's frame pointer. The
552 // calling procedure's stack pointer (%esp) is 8 higher than the value
553 // of the called procedure's frame pointer at the time the calling
554 // procedure made the CALL: 4 bytes for the return address pushed by the
555 // CALL itself, and 4 bytes for the callee's PUSH of the caller's frame
556 // pointer.
557 //
558 // %eip_new = *(%ebp_old + 4)
559 // %esp_new = %ebp_old + 8
560 // %ebp_new = *(%ebp_old)
561
562 uint32_t caller_eip, caller_esp, caller_ebp;
563
564 if (memory_->GetMemoryAtAddress(last_ebp + 4, &caller_eip) &&
565 memory_->GetMemoryAtAddress(last_ebp, &caller_ebp)) {
566 caller_esp = last_ebp + 8;
567 trust = StackFrame::FRAME_TRUST_FP;
568 } else {
569 // We couldn't read the memory %ebp refers to. It may be that %ebp
570 // is pointing to non-stack memory. We'll scan the stack for a
571 // return address. This can happen if last_frame is executing code
572 // for a module for which we don't have symbols, and that module
573 // is compiled without a frame pointer.
574 if (!stack_scan_allowed
575 || !ScanForReturnAddress(last_esp, &caller_esp, &caller_eip,
576 frames.size() == 1 /* is_context_frame */)) {
577 // if we can't find an instruction pointer even with stack scanning,
578 // give up.
579 return NULL;
580 }
581
582 // ScanForReturnAddress found a reasonable return address. Advance %esp to
583 // the location immediately above the one where the return address was
584 // found.
585 caller_esp += 4;
586 // Try to restore the %ebp chain. The caller %ebp should be stored at a
587 // location immediately below the one where the return address was found.
588 // A valid caller %ebp must be greater than the address where it is stored
589 // and the gap between the two adjacent frames should be reasonable.
590 uint32_t restored_ebp_chain = caller_esp - 8;
591 if (!memory_->GetMemoryAtAddress(restored_ebp_chain, &caller_ebp) ||
592 caller_ebp <= restored_ebp_chain ||
593 caller_ebp - restored_ebp_chain > kMaxReasonableGapBetweenFrames) {
594 // The restored %ebp chain doesn't appear to be valid.
595 // Assume that %ebp is unchanged.
596 caller_ebp = last_ebp;
597 }
598
599 trust = StackFrame::FRAME_TRUST_SCAN;
600 }
601
602 // Create a new stack frame (ownership will be transferred to the caller)
603 // and fill it in.
604 StackFrameX86* frame = new StackFrameX86();
605
606 frame->trust = trust;
607 frame->context = last_frame->context;
608 frame->context.eip = caller_eip;
609 frame->context.esp = caller_esp;
610 frame->context.ebp = caller_ebp;
611 frame->context_validity = StackFrameX86::CONTEXT_VALID_EIP |
612 StackFrameX86::CONTEXT_VALID_ESP |
613 StackFrameX86::CONTEXT_VALID_EBP;
614
615 return frame;
616 }
617
GetCallerFrame(const CallStack * stack,bool stack_scan_allowed)618 StackFrame* StackwalkerX86::GetCallerFrame(const CallStack* stack,
619 bool stack_scan_allowed) {
620 if (!memory_ || !stack) {
621 BPLOG(ERROR) << "Can't get caller frame without memory or stack";
622 return NULL;
623 }
624
625 const vector<StackFrame*> &frames = *stack->frames();
626 StackFrameX86* last_frame = static_cast<StackFrameX86*>(frames.back());
627 scoped_ptr<StackFrameX86> new_frame;
628
629 // If the resolver has Windows stack walking information, use that.
630 WindowsFrameInfo* windows_frame_info
631 = frame_symbolizer_->FindWindowsFrameInfo(last_frame);
632 if (windows_frame_info)
633 new_frame.reset(GetCallerByWindowsFrameInfo(frames, windows_frame_info,
634 stack_scan_allowed));
635
636 // If the resolver has DWARF CFI information, use that.
637 if (!new_frame.get()) {
638 CFIFrameInfo* cfi_frame_info =
639 frame_symbolizer_->FindCFIFrameInfo(last_frame);
640 if (cfi_frame_info)
641 new_frame.reset(GetCallerByCFIFrameInfo(frames, cfi_frame_info));
642 }
643
644 // Otherwise, hope that the program was using a traditional frame structure.
645 if (!new_frame.get())
646 new_frame.reset(GetCallerByEBPAtBase(frames, stack_scan_allowed));
647
648 // If nothing worked, tell the caller.
649 if (!new_frame.get())
650 return NULL;
651
652 // Treat an instruction address of 0 as end-of-stack.
653 if (new_frame->context.eip == 0)
654 return NULL;
655
656 // If the new stack pointer is at a lower address than the old, then
657 // that's clearly incorrect. Treat this as end-of-stack to enforce
658 // progress and avoid infinite loops.
659 if (new_frame->context.esp <= last_frame->context.esp)
660 return NULL;
661
662 // new_frame->context.eip is the return address, which is the instruction
663 // after the CALL that caused us to arrive at the callee. Set
664 // new_frame->instruction to one less than that, so it points within the
665 // CALL instruction. See StackFrame::instruction for details, and
666 // StackFrameAMD64::ReturnAddress.
667 new_frame->instruction = new_frame->context.eip - 1;
668
669 return new_frame.release();
670 }
671
672 } // namespace google_breakpad
673