This patch allows the Breakpad minidump processor to use data from
STACK CFI records to generate stack traces for the ARM processor.
In the symbol dumper, we need a table mapping DWARF CFI register
numbers to their names: STACK CFI records refer to registers by name.
In the processor, we expand StackwalkerARM::GetCallerFrame to see if
there are STACK CFI records covering the callee, and then use those to
recover the caller's register values.
There's no good reason the ARM walker couldn't use the SimpleCFIWalker
interface declared in cfi_frame_info.h. Unfortunately, that interface
assumes that one can map register names to member pointers of the raw
context type, while MDRawContextARM uses an array to hold the
registers' values: C++ pointer-to-member types can't refer to elements
of member arrays. So we have to write out SimpleCFIWalker::FindCallerRegisters
in StackwalkerARM::GetCallerFrame.
We define enum MDARMRegisterNumbers in minidump_cpu_arm.h, for
convenience in referring to certain ARM registers with dedicated
purposes, like the stack pointer and the PC.
We define validity flags in StackFrameARM for all the registers, since
CFI could theoretically recover any of them. In the same vein, we
expand minidump_stackwalk.cc to print the values of all valid
callee-saves registers in the context --- and use the proper names for
special-purpose registers.
We provide unit tests that give full code and branch coverage (with
minor exceptions). We add a testing interface to StackwalkerARM that
allows us to create context frames that lack some register values.
a=jimblandy, r=mmentovai
git-svn-id: http://google-breakpad.googlecode.com/svn/trunk@553 4c0a9323-5329-0410-9bdc-e9ce6186880e
Extend google_breakpad::CFISection with the ability to produce
.eh_frame data. Entry headers have a different format, and pointers
can be encoded in new and fascinating ways.
Extend dwarf2reader::CallFrameInfo to be able to parse either DWARF
CFI or .eh_frame data, as determined by an argument to the
constructor. Cope with variations in header formats, encoded pointers,
and additional data in 'z' augmentation data blocks. Extend the unit
tests appropriately.
Extend dump_syms to look for a .eh_frame section, and if it is
present, find the necessary base addresess and parse its contents.
There's no need for DwarfCFIToModule to check the version numbers; if
CallFrameInfo can parse it, DwarfCFIToModule should be able to handle
it. Adjust tests accordingly.
a=jimblandy, r=nealsid
git-svn-id: http://google-breakpad.googlecode.com/svn/trunk@552 4c0a9323-5329-0410-9bdc-e9ce6186880e
The Linux C++ exception handling data format (.eh_frame) can specify a
number of different encodings for the addresses it contains. This
patch extends dwarf2reader::ByteReader to read pointers encoded in
these ways.
a=jimblandy, r=nealsid
git-svn-id: http://google-breakpad.googlecode.com/svn/trunk@551 4c0a9323-5329-0410-9bdc-e9ce6186880e
Define a new DWARF parser class, dwarf2reader::CallFrameInfo.
Extend google_breakpad::Module to store and write out 'STACK CFI' records.
Define a new google_breakpad::DwarfCFIToModule class, to accept DWARF
CFI data from the parser and populate a Module with the equivalent
STACK CFI records.
Extend the Linux symbol dumping tool, dump_syms, to use
dwarf2reader::CallFrameInfo, google_breakpad::DwarfCFIToModule, and
google_breakpad::Module to extract DWARF CFI from the executable or
shared library files and write it to the Breakpad symbol file.
Define CFISection, a new class derived from TestAssembler::Section,
for use in creating DWARF CFI data for test cases.
a=jimblandy, r=nealsid
git-svn-id: http://google-breakpad.googlecode.com/svn/trunk@550 4c0a9323-5329-0410-9bdc-e9ce6186880e
Add a CFIFrameInfo class (named for symmetry with WindowsFrameInfo) to
represent the set of STACK CFI rules in effect at a given instruction,
and apply them to a set of register values. Provide a SimpleCFIWalker
class template, to allow the essential CFI code to be shared amongst
the different architectures.
Teach BasicSourceLineResolver to partially parse 'STACK CFI' records,
and produce the set of rules in effect at a given instruction on
demand, by combining the initial rule set and the appropriate rule
deltas in a CFIFrameInfo object.
Adapt StackwalkerX86 and StackFrameX86 to retrieve, store, and apply
CFI stack walking information.
Add validity flags for all the general-purpose registers to
StackFrameX86::ContextValidity.
a=jimblandy, r=mmentovai
git-svn-id: http://google-breakpad.googlecode.com/svn/trunk@549 4c0a9323-5329-0410-9bdc-e9ce6186880e
Issue 53001 (http://breakpad.appspot.com/53001) defines the
TestAssembler classes; those, along with a new set of mock classes
defined in stackwalker_unittest_utils.h, make it possible for us to
actually do proper unit testing of a stack walker. These tests get us
full code coverage for stackwalker_x86.cc.
a=jimblandy, r=mmentovai
git-svn-id: http://google-breakpad.googlecode.com/svn/trunk@548 4c0a9323-5329-0410-9bdc-e9ce6186880e
This also adds two new test utility class groups, TestAssembler and
SynthMinidump. These are overkill for what I'm doing with them here
(and may simply be overkill, period), but they make it easy to write
unit tests for code that works on binary files or raw memory contents
in a cross-platform way. I'm planning to use them for the DWARF CFI
unwinding tests and the DWARF CFI parser tests.
a=jimblandy, r=nealsid
git-svn-id: http://google-breakpad.googlecode.com/svn/trunk@547 4c0a9323-5329-0410-9bdc-e9ce6186880e
Programs compiled with -ffunction-sections -Wl,--gc-sections may have
SO entries for the start of the compilation unit whose addresses are
zero, even when the compilation unit contains non-omitted functions at
non-zero addresses. The breakpad dumper should not assume that the
compilation unit starting address is always non-zero.
a=jimblandy, r=nealsid
git-svn-id: http://google-breakpad.googlecode.com/svn/trunk@542 4c0a9323-5329-0410-9bdc-e9ce6186880e
As explained in the code:
Given the right options, the GNU toolchain will omit unreferenced
functions from the final executable. Unfortunately, when it does so,
it does not remove the associated portions of the line number program;
instead, it lets the symbol references in the DW_LNE_set_address
instructions pointing to the now-deleted code resolve to zero. Given
this input, the DWARF line parser will call AddLine with a series of
lines starting at address zero.
Rather than collecting series of lines describing code that is not
there, we should drop them. Since the linker doesn't explicitly
distinguish references to dropped sections from genuine references to
zero, we must use a heuristic. We have chosen:
- If a line starts at address zero, omit it. (On the platforms
breakpad targets, it is extremely unlikely that there will be code
at address zero.)
- If a line starts immediately after an omitted line, omit it too.
a=jimblandy, r=nealsid
git-svn-id: http://google-breakpad.googlecode.com/svn/trunk@538 4c0a9323-5329-0410-9bdc-e9ce6186880e
Some versions of the libstdc++, the GNU standard C++ library, have
stream extractors for unsigned integer values that permit a leading
'-' sign (6.0.13); others do not (6.0.9). Regardless of the behavior
of the extractors, Breakpad postfix expressions should support
negative literals.
a=jimblandy, r=nealsid
git-svn-id: http://google-breakpad.googlecode.com/svn/trunk@537 4c0a9323-5329-0410-9bdc-e9ce6186880e
Some of the error messages that could be generated in the process of
parsing DWARF debugging information lack terminating newlines.
a=jimblandly, r=nealsid
git-svn-id: http://google-breakpad.googlecode.com/svn/trunk@536 4c0a9323-5329-0410-9bdc-e9ce6186880e
Any DIE with an DW_AT_inline attribute can be cited by
DW_AT_abstract_origin attributes --- even if the value of the
DW_AT_inline attribute is DW_INL_not_inlined. Thus, we need to set the
inline_ flag on all such DIEs, regardless of the attribute's value.
This allows us to find names in situations like this:
<1><30cf>: Abbrev Number: 57 (DW_TAG_subprogram)
<30d0> DW_AT_specification: <0x3013>
<30d4> DW_AT_decl_file : 1
<30d5> DW_AT_decl_line : 92
<30d6> DW_AT_inline : 0 (not inlined)
<30d7> DW_AT_sibling : <0x30f0>
...
<1><30f5>: Abbrev Number: 59 (DW_TAG_subprogram)
<30f6> DW_AT_abstract_origin: <0x30cf>
<30fa> DW_AT_low_pc : 0x13bc
<30fe> DW_AT_high_pc : 0x13ec
<3102> DW_AT_frame_base : 0x2c (location list)
<3106> DW_AT_sibling : <0x3113>
a=jimblandy, r=nealsid,dmuir
git-svn-id: http://google-breakpad.googlecode.com/svn/trunk@526 4c0a9323-5329-0410-9bdc-e9ce6186880e
FindSectionByName will return the first section whose name starts with
NAME, because strncmp stops the comparison once NAME's characters have
been found to match. The comparison stops before the terminating '\0'.
For example, if we search for the section named ".eh_frame", we may
get the section named ".eh_frame_hdr".
Instead, check that the section name section has enough space to store
the complete name with its terminating '\0', and then use strcmp,
which will never examine more than strlen(NAME) + 1 bytes from the
section name section, regardless of its contents, and will require the
terminating '\0' to match as well.
a=jimblandy, r=nealsid
git-svn-id: http://google-breakpad.googlecode.com/svn/trunk@525 4c0a9323-5329-0410-9bdc-e9ce6186880e
This is preparation for adding support for reading Linux C++ exception
handling data's encoded pointers. The change should have no user-visible
effect; it simply expands the comments for dwarf2reader::ByteReader, and
regroups the member functions.
a=jimblandy, r=nealsid
git-svn-id: http://google-breakpad.googlecode.com/svn/trunk@522 4c0a9323-5329-0410-9bdc-e9ce6186880e
Change configure.ac to note that Breakpad requires automake version
1.11.1 or later. This will cause older versions of automake to refuse
to process the Makefile.am file.
Earlier versions of automake generate 'make dist' rules that have a
security flaw; see:
http://web.nvd.nist.gov/view/vuln/detail?vulnId=CVE-2009-4029
However, note that that notice itself has a flaw: the bug is *fixed
in* automake 1.11.1, not present. See:
http://lists.gnu.org/archive/html/automake/2009-12/msg00012.html
(The change to Makefile.in is a consequence of my having neglected to
rebuild Makefile.in after landing r517.)
a=jimblandy, r=mmentovai
git-svn-id: http://google-breakpad.googlecode.com/svn/trunk@521 4c0a9323-5329-0410-9bdc-e9ce6186880e
Without this patch, debugging information like the following will produce
FUNC records with no names, because the dumper (correctly) ignores the
DW_TAG_subprogram DIEs that lack DW_AT_low_pc/DW_AT_high_pc attributes, but
won't follow the DW_AT_abstract_origin link from the DIE that does have
code addresses to find its name.
<1><168>: Abbrev Number: 5 (DW_TAG_class_type)
<169> DW_AT_name : Foo
<2><183>: Abbrev Number: 7 (DW_TAG_subprogram)
<185> DW_AT_name : Foo
<18b> DW_AT_declaration : 1
<1><1b7>: Abbrev Number: 12 (DW_TAG_subprogram)
<1b8> DW_AT_specification: <0x183>
<1bc> DW_AT_inline : 2 (declared as inline but ignored)
<1><1dc>: Abbrev Number: 16 (DW_TAG_subprogram)
<1dd> DW_AT_abstract_origin: <0x1b7>
<1e1> DW_AT_low_pc : 0x8048578
<1e5> DW_AT_high_pc : 0x8048588
a=dmuir, r=jimblandy
git-svn-id: http://google-breakpad.googlecode.com/svn/trunk@520 4c0a9323-5329-0410-9bdc-e9ce6186880e
Yes, classes are useful. But that doesn't mean that every function has
to gratuitously become a member function. The Google C++ Style Guide
does not require this silliness, since the function is in the
google_breakpad namespace anyway.
a=jimblandy, r=nealsid
git-svn-id: http://google-breakpad.googlecode.com/svn/trunk@519 4c0a9323-5329-0410-9bdc-e9ce6186880e
This also includes some comments I promised Cary Coutant I'd write
about the appropriateness of processing attributes in EndAttributes
calls.
The Google C++ Style Guide requires each file to have an author notice
and a comment explaining the file's general purpose. For the record, I
don't think putting an author notice on the files is a good idea; it's
odd to have the original author retain prominence even if the file has
been heavily edited by others; the version control system answers this
question more accurately. This is only for Style Guide compliance. The
Apache group decided to discourage author annotations, partially for
these reasons:
http://mail-archives.apache.org/mod_mbox/jakarta-jmeter-dev/200402.mbox/%3C4039F65E.7020406@atg.com%3E
a=jimblandy, r=nealsid
git-svn-id: http://google-breakpad.googlecode.com/svn/trunk@518 4c0a9323-5329-0410-9bdc-e9ce6186880e
We've gotten mixed advice from the lawyery types about whether this
matters. But it's easy enough to do.
a=jimblandy, r=nealsid
git-svn-id: http://google-breakpad.googlecode.com/svn/trunk@517 4c0a9323-5329-0410-9bdc-e9ce6186880e
This looks a little odd right now, since ParseStackInfo has only one
alternative to handle, but I think breaking this out should make the
subsequent addition of STACK CFI record support easier to review.
a=jimblandy, r=mmentovai
git-svn-id: http://google-breakpad.googlecode.com/svn/trunk@514 4c0a9323-5329-0410-9bdc-e9ce6186880e
Rename BasicSourceLineResolver::Module::StackInfoTypes to
WindowsFrameInfoTypes. This enum really describes the forms of
Windows-specific stack unwinding data (STACK WIN records), and its
name should reflect that, especially since we'll be adding support for
other kinds of stack walking information.
The 'stack' -> 'frame' shift matches the naming of the
WindowsFrameInfo type.
Similarly, rename BasicSourceLineResolver::Module::stack_info_ to
windows_frame_info_.
Do similar renamings in basic_source_line_resolver_unittest.cc.
a=jimblandy, r=mmentovai
git-svn-id: http://google-breakpad.googlecode.com/svn/trunk@513 4c0a9323-5329-0410-9bdc-e9ce6186880e
This patch moves the code for finding caller frames using STACK WIN
data and the code to do so using the traditional frame layout (%ebp
points at saved %ebp, pushed just after return address) into their own
functions. In addition to making things a little clearer, this is
preparation for adding support for STACK CFI records into the mix.
a=jimblandy, r=mmentovai
git-svn-id: http://google-breakpad.googlecode.com/svn/trunk@512 4c0a9323-5329-0410-9bdc-e9ce6186880e
This adds an EvaluateForValue member function to PostfixEvaluator, and
along with appropriate unit tests.
a=jimblandy, r=nealsid
git-svn-id: http://google-breakpad.googlecode.com/svn/trunk@511 4c0a9323-5329-0410-9bdc-e9ce6186880e
The Google C++ Style Guide requires all parameters passed by reference
to be labeled 'const', and says that pointers should be used for
output arguments. This patch brings google_breakpad::StackwalkerX86
into line.
a=jimblandy, r=mmentovai
git-svn-id: http://google-breakpad.googlecode.com/svn/trunk@510 4c0a9323-5329-0410-9bdc-e9ce6186880e
In order to be able to treat any MemoryRegion as const, the accessor
functions need to be declared this-const, which means annotations on
all the subclasses, etc. etc.
Since MinidumpMemoryRegion fills its memory_ member on demand, that
member needs to be marked 'mutable', but this is exactly the sort of
situation the 'mutable' keyword was intended for, so that seems all
right.
a=jimblandy, r=nealsid
git-svn-id: http://google-breakpad.googlecode.com/svn/trunk@509 4c0a9323-5329-0410-9bdc-e9ce6186880e
At the moment, StackwalkerX86::GetCallerFrame doesn't save the
WindowsFrameInfo that it finds for a frame unless it successfully
constructs the caller frame. This means that the windows_frame_info
field of the last frame on the stack is left unset, even when that
frame does have windows unwinding information.
This is not user-visible behavior, so it doesn't matter, but it is a
blemish on the interface, and unit tests (added in a later patch)
expect it.
This patch saves the information in the frame as soon as we find it.
a=jimblandy, r=mmentovai
git-svn-id: http://google-breakpad.googlecode.com/svn/trunk@508 4c0a9323-5329-0410-9bdc-e9ce6186880e
This extends the ElfArchitecture function to recognize the
architectures it seemed to me that breakpad was most likely to see.
Also: the dumper has historically not provided very helpful error
messages. This patch adds a few that were convenient, but we should do
an audit for this.
a=jimblandy, r=ted.mielczarek
git-svn-id: http://google-breakpad.googlecode.com/svn/trunk@507 4c0a9323-5329-0410-9bdc-e9ce6186880e
The Google C++ Style Guide says that members of structures needn't
have names ending in underscores. The structure types in
google_breakpad::Module don't follow this rule.
a=jimblandy, r=nealsid
git-svn-id: http://google-breakpad.googlecode.com/svn/trunk@505 4c0a9323-5329-0410-9bdc-e9ce6186880e
This patch moves the ReadInitialFunction from dwarf2reader.cc, where
it was a static function, to being a member function of
google_breakpad::ByteReader.
a=jimblandy, r=nealsid
git-svn-id: http://google-breakpad.googlecode.com/svn/trunk@504 4c0a9323-5329-0410-9bdc-e9ce6186880e
In r480, I botched the change to make the comparisons that decide
whether an address falls within a function's range safe from overflow.
The original code said:
address >= function_base && address < function_base + function_size
which is fine unless the function abuts the end of the address space,
in which case the addition overflows and you get a false negative.
My change subtracted function_size from both sides of the latter
comparison, which is meaning-preserving in true math, and gets you:
address >= function_base && address - function_size < function_base
This not only reads strangely, but also still overflows if
function_size is greater than address. That's rare, but I've added a
case to the unit tests that checks it.
My intent had been to replace the addition which could overflow with a
subtraction that was known not to overflow, namely:
address >= function_base && address - function_base < function_size
This is equivalent to the original in true math, and because of the
first comparison, we know the subtraction won't underflow in MemAddr
math.
The patch includes similar fixes to the public symbol lookup code, and
to FindWindowsFrameInfo, which was the only other function affected by
r480.
a=jimblandy, r=mmentovai
git-svn-id: http://google-breakpad.googlecode.com/svn/trunk@503 4c0a9323-5329-0410-9bdc-e9ce6186880e
Having NDEBUG be the default has wasted my time more often than I'm
proud to admit. There are no expensive asserts in the Linux symbol
dumper.
a=jimblandy, r=nealsid
git-svn-id: http://google-breakpad.googlecode.com/svn/trunk@502 4c0a9323-5329-0410-9bdc-e9ce6186880e
RangeMaps use the range's upper end as the key in the underlying map,
but RetrieveNearestRange was treating the key as the lower end.
a=jimblandy, r=mmentovai
git-svn-id: http://google-breakpad.googlecode.com/svn/trunk@501 4c0a9323-5329-0410-9bdc-e9ce6186880e
This adds DWARF support to the Breakpad Linux dumper. This is
implemented as two handler classes: google_breakpad::DwarfCUToModule
accepts data from dwarf2reader::CompilationUnit, and
google_breakpad::DwarfLineToModule accepts data from a
dwarf2reader::LineInfo, each populating a google_breakpad::Module with
the results. Behaviors specific to particular source languages are
handled by instances of a new class, google_breakpad::Language.
An input executable may contain both STABS and DWARF debugging
information: the dumper automatically recognizes what sorts of
information are available, and integrates the data into a single
output file.
All classes have unit tests, providing line and branch coverage of all
interesting code. Unit tests are written using the Google C++ Testing
Framework, and the Google C++ Mocking Framework where appropriate.
a=jimblandy, r=ccoutant
git-svn-id: http://google-breakpad.googlecode.com/svn/trunk@497 4c0a9323-5329-0410-9bdc-e9ce6186880e
dwarf2reader::CompilationUnit is a simple and direct parser for DWARF
data, but its handler interface is not convenient to use. In
particular, the same handler object receives data about all DIEs
processed. One can't use distinct classes to separate the information
needed to handle different kinds of data.
This patch defines a new adapter type, dwarf2reader::DIEHandler, which
implements the existing DWARF parser's handler interface, given a
handler written to a more comfortable, object-orient interface. The
comments in dwarf2diehandler.h provide more detail.
a=jimblandy, r=ccoutant
git-svn-id: http://google-breakpad.googlecode.com/svn/trunk@495 4c0a9323-5329-0410-9bdc-e9ce6186880e
Breakpad's DWARF line number info parser provides a code address,
file, and line number for each code/source pairing, but doesn't
provide the length of the machine code. This makes that change, as
discussed in the following thread:
http://groups.google.com/group/google-breakpad-dev/browse_thread/thread/ed8d2fde79319368p
This patch also makes the corresponding changes to the functioninfo.cc
module, used by the Mac dumper. This patch has no effect on the Mac
dumper's output.
a=jimblandy, r=ccoutant
git-svn-id: http://google-breakpad.googlecode.com/svn/trunk@494 4c0a9323-5329-0410-9bdc-e9ce6186880e
The DWARF specification specifices which names the sections containing
DWARF information should have. OSX uses slightly different names. This
patch changes the DWARF reader to look for the sections under both
sets of names.
a=jimblandy, r=ccoutant
git-svn-id: http://google-breakpad.googlecode.com/svn/trunk@493 4c0a9323-5329-0410-9bdc-e9ce6186880e
At the moment, the StackWalker GetCallerFrame member function expects
a vector of WindowsFrameInfo structures, even though WindowsFrameInfo
is only used or useful on one one implementation (StackWalkerX86).
This patch changes StackWalker::GetCallerFrame to no longer expect the
WindowsFrameInfo structures, and changes all implementations to match.
In particular, StackWalkerX86 is changed to find the WindowsFrameInfo
data itself, and store a pointer to whatever it got in the StackFrame
object itself (which is really a StackFrameX86).
To allow GetCallerFrame implementations to look up stack walking data,
StackWalker::resolver_ needs to be made protected, not private.
a=jimblandy, r=mmentovai
git-svn-id: http://google-breakpad.googlecode.com/svn/trunk@491 4c0a9323-5329-0410-9bdc-e9ce6186880e
The stabs reading code in google-breakpad incorrectly assumes that the
stabs data is a single compilation unit. Specifically, it ignores
N_UNDF stabs and assumes that all string indices are relative to the
beginning of the .stabstr section.
This is true when linking with the GNU linker by default, because the
GNU linker optimizes stabs debug info. The gold linker does not do
this optimization. It can be disabled when using the GNU linker with
the --traditional-format command line option.
For more details of the problem, see:
http://sourceware.org/bugzilla/show_bug.cgi?id=10338http://code.google.com/p/google-breakpad/issues/detail?id=359
This patch adds unit tests that reproduce the failure, and fixes the
stabs parser.
a=jimblandy, r=nealsid
git-svn-id: http://google-breakpad.googlecode.com/svn/trunk@490 4c0a9323-5329-0410-9bdc-e9ce6186880e
This adds a new variable, TEST_WRAPPER, to src/tools/linux/dump_syms.
Comments in the patch provide details.
This patch also moves the public variable section to sit after the
public phony targets.
a=jimblandy, r=nealsid
git-svn-id: http://google-breakpad.googlecode.com/svn/trunk@486 4c0a9323-5329-0410-9bdc-e9ce6186880e
