Summary:
- Reason of both bugs:
1. For the very first frame, Unwinder doesn't check the validity
of Full UnwindPlan before creating StackFrame from it:
When 'process launch' command is run after setting a breakpoint
in inferior, the Unwinder runs and saves only Frame 0 (the frame
in which breakpoint was set) in thread's StackFrameList i.e.
m_curr_frames_sp. However, it doesn't check the validity of the
Full UnwindPlan for this frame by unwinding 2 more frames further.
2. Unwinder doesn't update the CFA value of Cursor when Full UnwindPlan
fails and FallBack UnwindPlan succeeds in providing valid CFA values
for frames:
Sometimes during unwinding of stack frames, the Full UnwindPlan
inside the RegisterContextLLDB object may fail to provide valid
CFA values for these frames. Then the Fallback UnwindPlan is used
to unwind the frames.
If the Fallback UnwindPlan succeeds, then it provides a valid new
CFA value. The RegisterContextLLDB::m_cfa field of Cursor object
is updated during the Fallback UnwindPlan execution. However,
UnwindLLDB misses the implementation to update the 'cfa' field
of this Cursor with this valid new CFA value.
- This patch fixes both these issues.
- Remove XFAIL in test files corresponding to these 2 Bugs
Change-Id: I932ea407545ceee2d628f946ecc61a4806d4cc86
Signed-off-by: Abhishek Aggarwal <abhishek.a.aggarwal@intel.com>
Reviewers: jingham, lldb-commits, jasonmolenda
Subscribers: lldb-commits, ovyalov, tberghammer
Differential Revision: http://reviews.llvm.org/D14226
llvm-svn: 253026
Previously we accepted a frame as correct result if the PC pointed
into an executable section of code. The isse with that approac is
that if we calculated PC correctly but messed up the value of CFA
then unwinding from the next fram will most likely fail.
With this change I modify the logic with keeping the requirement
for PC to point to an executable section and also check that we can
continue the unwind from the frame we calculated. If continuing from
the frame calculated with the primary unwind plan isn't working then
fall back to the fallback plan with the hope for a better frame (if
the fallback plan won't help then we acceot the frame from the
primary plan).
Differential revision: http://reviews.llvm.org/D10932
llvm-svn: 241434
This continues the effort to reduce header footprint and improve
build speed by removing clang and other unnecessary headers
from Target.h. In one case, some headers were included solely
for the purpose of declaring a nested class in Target, which was
not needed by anybody outside the class. In this case the
definition and implementation of the nested class were isolated
in the .cpp file so the header could be removed.
llvm-svn: 231107
a register value that is live in the stack frame 0 register context.
Fixes a problem where retrieving a register value on stack frame #n
would involved O(n!) stack frame checks. This could be very slow on
a deep stack when retrieving register values that had not been
modified/saved by any of the stack frames. Not common, but annoying
when it was hit.
<rdar://problem/19010211>
llvm-svn: 223843
UnwindLLDB::AddOneMoreFrame to try the fallback unwind plan on
that same stack frame before it tries the fallback unwind plan
on the "next" or callee frame.
In RegisterContextLLDB::TryFallbackUnwindPlan, when we're
trying the fallback unwind plan to see if it is valid, make
sure we change all of the object ivars that might be used in
the process of fetching the CFA & caller's saved pc value
and restore those if we decide not to use the fallback
unwindplan.
<rdar://problem/19035079>
llvm-svn: 222601
RegisterContextLLDB. I have core files of half a dozen tricky
unwind situations on x86/arm and they're all working pretty much
correctly at this point, but we'll need to keep an eye out for
unwinder regressions for a little while; it's tricky to get these
heuristics completely correct in all unwind situations.
<rdar://problem/18937193>
llvm-svn: 221866
let's let lldb try the arch default unwind every time but not destructively --
it doesn't permanently replace the main unwind method for that function from
now on.
This fix is for <rdar://problem/18683658>.
I tested it against Ryan Brown's go program test case and also a
collection of core files of tricky unwind scenarios
<rdar://problem/15664282> <rdar://problem/15835846>
<rdar://problem/15982682> <rdar://problem/16099440>
<rdar://problem/17364005> <rdar://problem/18556719>
that I've fixed over the last 6-9 months.
llvm-svn: 221238
is "invalid" -- it is past the end of the stack trace. Add a new
method IsCompletedStackWalk() so we can tell if an invalid stack
frame is from a complete backtrace or if it might be worth re-trying
the last unwind with a different method.
This fixes the unwinder problems Ryan Brown was having with go
programs. The unwinder can (under the right circumstances) still
destructively replace unwind plans permanently - I'll work on
that in a different patch.
<rdar://problem/18683658>
llvm-svn: 221229
This is a mechanical change addressing the various sign comparison warnings that
are identified by both clang and gcc. This helps cleanup some of the warning
spew that occurs during builds.
llvm-svn: 205390
class. If we try to unwind a stack frame to find a caller stack
frame, and we fail to get a valid-looking frame, AND if the UnwindPlan
we used is an assembly-inspection based UnwindPlan, then we should
throw away the assembly-inspection UnwindPlan and try unwinding with
the architectural default UnwindPlan.
This code path won't be taken if eh_frame unwind instructions are available -
lldb will always prefer those once it's off the zeroth frame.
The problem I'm trying to fix here is the class of unwind failures that
happen when we have hand-written assembly on the stack, with no eh_frame,
and lldb's assembly parser fails to understand the assembly. People usually
write their hand-written assembly to follow the frame-pointer-preserving
conventions of the platform so the architectural default UnwindPlan will
often work. We won't have the spill location for most of the non-volatile
registers if we fall back to this, but it's better than stopping the unwind
prematurely.
This is a bit of a tricky change that I believe is correct, but if we get
unwinds that go of into the weeds / unwind bogus frames at the end of the
stack, I'll need to revisit it.
<rdar://problem/16099440>
llvm-svn: 201839
specify a list of functions which should be treated as trap handlers.
This will be primarily useful to people working in non-user-level
process debugging - kernels and other standalone environments.
For most people, the trap handler functions provided by the Platform
plugin will be sufficient.
<rdar://problem/15835846>, <rdar://problem/15982682>
llvm-svn: 201386
aka asynchronous signal handlers, which subclasses should fill
in as appropriate. For most Unix user process environments,
the one entry in this list is _sigtramp. For bare-board and
kernel environments, there will be different sets of trap
handlers.
The unwinder needs to know when a frame is a trap handler
because the rules it enforces for the frame "above" the
trap handler is different from most middle-of-the-stack frames.
<rdar://problem/15835846>
llvm-svn: 201300
its stack frame is a constructed, fake thing that may not conform
correctly to these rules. This fixes a problem where lldb couldn't
backtrace past an asynchronous signal handler (_sigtramp) frame on
a stack on Mac OS X.
<rdar://problem/15035673>
llvm-svn: 198450
The original code was not completely correct, but a form of
this check is necessary to avoid an infinite recursion on
some unwind cases where a function unwinds to itself with the
same CFA. Ashok thought the recursion would be caught in
RegisterContextLLDB but this one isn't - we still need it here.
<rdar://problem/15664282>
llvm-svn: 197761
pure virtual base class and made StackFrame a subclass of that. As
I started to build on top of that arrangement today, I found that it
wasn't working out like I intended. Instead I'll try sticking with
the single StackFrame class -- there's too much code duplication to
make a more complicated class hierarchy sensible I think.
llvm-svn: 193983
defines a protocol that all subclasses will implement. StackFrame
is currently the only subclass and the methods that Frame vends are
nearly identical to StackFrame's old methods.
Update all callers to use Frame*/Frame& instead of pointers to
StackFrames.
This is almost entirely a mechanical change that touches a lot of
the code base so I'm committing it alone. No new functionality is
added with this patch, no new subclasses of Frame exist yet.
I'll probably need to tweak some of the separation, possibly moving
some of StackFrame's methods up in to Frame, but this is a good
starting point.
<rdar://problem/15314068>
llvm-svn: 193907
- Removes the block in UnwindLLDB::AddOneMoreFrame that tests for a bad stack setup,
since it is neither correct (tests the FP GPR), complete (doesn't consider multi-frame
cycles), nor reachable (the construction of RegisterContextLLDB will fail in the case
where either of the two (why just two?) previous frames have the same canonical frame
address as the frame that we propose adding to the stack).
llvm-svn: 191430
the link register save location being in the link register - in which case we
should iterate down the stack, not recursively try to find the lr in the current
frame over and over.
<rdar://problem/13932954>
llvm-svn: 183282
thread before UnwindLLDB::AddOneMoreFrame calls it quits. We have
a couple of reports of unending backtraces in the field and we
haven't been able to collect any information about what kind of
backtrace is causing this. We've found on Mac OS X that it's tricky
to get more than around 200k stack frames before a process exceeds
its stack space so we're starting with a hard limit of 300,000 frames.
<rdar://problem/13383069>
llvm-svn: 180995
LLDB is crashing when logging is enabled from lldb-perf-clang. This has to do with the global destructor chain as the process and its threads are being torn down.
All logging channels now make one and only one instance that is kept in a global pointer which is never freed. This guarantees that logging can correctly continue as the process tears itself down.
llvm-svn: 178191
Modify UnwindLLDB::SearchForSavedLocationForRegister so if the register
save locations for a register mid-stack is in another register (or in the
same register, indicating the reg wasn't modified in this frame), don't
return that as a found location. Keep iterating down the array of frames
until a concrete location/value for the register is found, or until we
get to frame 0 where the reg value can be used as-is.
If lldb was trying to backtrace a program that blew out its stack via
recursion and the unwind instructions had some kind of
this-reg-is-saved-in-that-reg instruction, lldb would revert to doing
a recursive search for a concrete value and blow out its own stack.
llvm-svn: 172887
allowed volatile registers to be returned up the stack. That leads
to unexpected/incorrect values provided to the user and we need to
avoid that.
<rdar://problem/12714247>
llvm-svn: 168123
Full UnwindPlan is trying to do an impossible unwind; in that case
invalidate the Full UnwindPlan and replace it with the architecture
default unwind plan.
This is a scenario that happens occasionally with arm unwinds in
particular; the instruction analysis based full unwindplan can
mis-parse the functions and the stack walk stops prematurely. Now
we can do a simpleminded frame-chain walk to find the caller frame
and continue the unwind. It's not ideal but given the complicated
nature of analyzing the arm functions, and the lack of eh_frame
information on iOS, it is a distinct improvement and fixes some
long-standing problems with the unwinder on that platform.
This is fixing <rdar://problem/12091421>. I may re-use this
invalidate feature in the future if I can identify other cases where
the full unwindplan's unwind information is clearly incorrect.
This checkin also includes some cleanup for the volatile register
definition in the arm ABI plugin for <rdar://problem/10652166>
although work remains to be done for that bug.
llvm-svn: 166757
I started work on being able to add symbol files after a debug session
had started with a new "target symfile add" command and quickly ran into
problems with stale Address objects in breakpoint locations that had
lldb_private::Section pointers into modules that had been removed or
replaced. This also let to grabbing stale modules from those sections.
So I needed to thread harded the Address, Section and related objects.
To do this I modified the ModuleChild class to now require a ModuleSP
on initialization so that a weak reference can created. I also changed
all places that were handing out "Section *" to have them hand out SectionSP.
All ObjectFile, SymbolFile and SymbolVendors were inheriting from ModuleChild
so all of the find plug-in, static creation function and constructors now
require ModuleSP references instead of Module *.
Address objects now have weak references to their sections which can
safely go stale when a module gets destructed.
This checkin doesn't complete the "target symfile add" command, but it
does get us a lot clioser to being able to do such things without a high
risk of crashing or memory corruption.
llvm-svn: 151336
objects for the backlink to the lldb_private::Process. The issues we were
running into before was someone was holding onto a shared pointer to a
lldb_private::Thread for too long, and the lldb_private::Process parent object
would get destroyed and the lldb_private::Thread had a "Process &m_process"
member which would just treat whatever memory that used to be a Process as a
valid Process. This was mostly happening for lldb_private::StackFrame objects
that had a member like "Thread &m_thread". So this completes the internal
strong/weak changes.
Documented the ExecutionContext and ExecutionContextRef classes so that our
LLDB developers can understand when and where to use ExecutionContext and
ExecutionContextRef objects.
llvm-svn: 151009
due to RTTI worries since llvm and clang don't use RTTI, but I was able to
switch back with no issues as far as I can tell. Once the RTTI issue wasn't
an issue, we were looking for a way to properly track weak pointers to objects
to solve some of the threading issues we have been running into which naturally
led us back to std::tr1::weak_ptr. We also wanted the ability to make a shared
pointer from just a pointer, which is also easily solved using the
std::tr1::enable_shared_from_this class.
The main reason for this move back is so we can start properly having weak
references to objects. Currently a lldb_private::Thread class has a refrence
to its parent lldb_private::Process. This doesn't work well when we now hand
out a SBThread object that contains a shared pointer to a lldb_private::Thread
as this SBThread can be held onto by external clients and if they end up
using one of these objects we can easily crash.
So the next task is to start adopting std::tr1::weak_ptr where ever it makes
sense which we can do with lldb_private::Debugger, lldb_private::Target,
lldb_private::Process, lldb_private::Thread, lldb_private::StackFrame, and
many more objects now that they are no longer using intrusive ref counted
pointer objects (you can't do std::tr1::weak_ptr functionality with intrusive
pointers).
llvm-svn: 149207
RegisterContextLLDBs it contains.
Previously RegisterContextLLDB objects had a pointer to their "next"
frame down the stack. e.g. stack starts at frame 0; frame 3 has a
pointer to frame 2. This is used to retreive callee saved register
values. When debugging an inferior that has blown out its own stack,
however, this could result in lldb blowing out its own stack while
recursing down to retrieve register values.
RegisterContextLLDB no longer has a pointer to its next frame; it
has a reference to the UnwindLLDB which contains it. When it needs
to retrieve a reg value, it asks the UnwindLLDB for that reg value
and UnwindLLDB iterates through the frames until it finds a location.
llvm-svn: 143423
virtual bool
ABI::StackUsesFrames () = 0;
Should return true if your ABI uses frames when doing stack backtraces. This
means a frame pointer is used that points to the previous stack frame in some
way or another.
virtual bool
ABI::CallFrameAddressIsValid (lldb::addr_t cfa) = 0;
Should take a look at a call frame address (CFA) which is just the stack
pointer value upon entry to a function. ABIs usually impose alignment
restrictions (4, 8 or 16 byte aligned), and zero is usually not allowed.
This function should return true if "cfa" is valid call frame address for
the ABI, and false otherwise. This is used by the generic stack frame unwinding
code to help determine when a stack ends.
virtual bool
ABI::CodeAddressIsValid (lldb::addr_t pc) = 0;
Validates a possible PC value and returns true if an opcode can be at "pc".
Some ABIs or architectures have fixed width instructions and must be aligned
to a 2 or 4 byte boundary. "pc" can be an opcode or a callable address which
means the load address might be decorated with extra bits (such as bit zero
to indicate a thumb function call for ARM targets), so take this into account
when returning true or false. The address should also be validated to ensure
it is a valid address for the address size of the inferior process. 32 bit
targets should make sure the address is less than UINT32_MAX.
Modified UnwindLLDB to use the new ABI functions to help it properly terminate
stacks.
Modified the mach-o function that extracts dependent files to not resolve the
path as the paths inside a binary might not match those on the current
host system.
llvm-svn: 132021
respective ABI plugins as they were plug-ins that supplied ABI specfic info.
Also hookep up the UnwindAssemblyInstEmulation so that it can generate the
unwind plans for ARM.
Changed the way ABI plug-ins are handed out when you get an instance from
the plug-in manager. They used to return pointers that would be mananged
individually by each client that requested them, but now they are handed out
as shared pointers since there is no state in the ABI objects, they can be
shared.
llvm-svn: 131193
ArchDefaultUnwindPlan plug-in interfaces are now cached per architecture
instead of being leaked for every frame.
Split the ArchDefaultUnwindPlan_x86 into ArchDefaultUnwindPlan_x86_64 and
ArchDefaultUnwindPlan_i386 interfaces.
There were sporadic crashes that were due to something leaking or being
destroyed when doing stack crawls. This patch should clear up these issues.
llvm-svn: 125541
bool RegisterContextLLDB::GetPC (addr_t& pc);
to:
bool RegisterContextLLDB::ReadPC (addr_t& pc);
To avoid confusion with the GetPC() function that is part of the
lldb_private::RegisterContext:
uint64_t RegisterContext::GetPC (uint64_t fail_value);
Bad things could happen if the two got intermixed and the wrong one got
called.
Fixed inifinite loop detection by watching for two frames where the
RegisterContextLLDB::CursorSP contains the same start_pc and cfa.
llvm-svn: 123673
a method:
void RegisterContext::InvalidateIfNeeded (bool force);
Each time this function is called, when "force" is false, it will only call
the pure virtual "virtual void RegisterContext::InvalideAllRegisters()" if
the register context's stop ID doesn't match that of the process. When the
stop ID doesn't match, or "force" is true, the base class will clear its
cached registers and the RegisterContext will update its stop ID to match
that of the process. This helps make it easier to correctly flush the register
context (possibly from multiple locations depending on when and where new
registers are availabe) without inadvertently clearing the register cache
when it doesn't need to be.
Modified the ProcessGDBRemote plug-in to be much more efficient when it comes
to:
- caching the expedited registers in the stop reply packets (we were ignoring
these before and it was causing us to read at least three registers every
time we stopped that were already supplied in the stop reply packet).
- When a thread has no stop reason, don't keep asking for the thread stopped
info. Prior to this fix we would continually send a qThreadStopInfo packet
over and over when any thread stop info was requested. We now note the stop
ID that the stop info was requested for and avoid multiple requests.
Cleaned up some of the expression code to not look for ClangExpressionVariable
objects up by name since they are now shared pointers and we can just look for
the exact pointer match and avoid possible errors.
Fixed an bug in the ValueObject code that would cause children to not be
displayed.
llvm-svn: 123127
frames, UnwindLLDB could create the incorrect RegisterContext for a given
stack frame because it was using the frame index (not the concrete frame
index). This was causing crashes when doing backtraces through the
SBFrame::GetFP() because a NULL register context was being returned for the
deepest stack frame.
llvm-svn: 123052
subclasses will automatically be able to take advantage of caching. The
cache line size is set to 512 by default.
This greatly speeds up stack backtraces on MacOSX when using the
ProcessGDBRemote process plug-in since only about 6300 packets per second
can be sent.
Initial speedups show:
Prior to caching: 10,000 stack frames took 5.2 seconds
After caching: 10,000 stack frames in 240 ms!
About a 20x speedup!
llvm-svn: 122996
an issue with the way the UnwindLLDB was handing out RegisterContexts: it
was making shared pointers to register contexts and then handing out just
the pointers (which would get put into shared pointers in the thread and
stack frame classes) and cause double free issues. MallocScribble helped to
find these issues after I did some other cleanup. To help avoid any
RegisterContext issue in the future, all code that deals with them now
returns shared pointers to the register contexts so we don't end up with
multiple deletions. Also now that the RegisterContext class doesn't require
a stack frame, we patched a memory leak where a StackFrame object was being
created and leaked.
Made the RegisterContext class not have a pointer to a StackFrame object as
one register context class can be used for N inlined stack frames so there is
not a 1 - 1 mapping. Updates the ExecutionContextScope part of the
RegisterContext class to never return a stack frame to indicate this when it
is asked to recreate the execution context. Now register contexts point to the
concrete frame using a concrete frame index. Concrete frames are all of the
frames that are actually formed on the stack of a thread. These concrete frames
can be turned into one or more user visible frames due to inlining. Each
inlined stack frame has the exact same register context (shared via shared
pointers) as any parent inlined stack frames all the way up to the concrete
frame itself.
So now the stack frames and the register contexts should behave much better.
llvm-svn: 122976
the frame count is requested or each frame is individually requested.
In practice this doesn't seem to help anything because we have
functions like StackFrameList::GetNumFrames() which is going to
request each frame anyway. And classes like ThreadPlanStepRange
and ThreadPlanStepOverRange get the stack depth in their ctor forcing
a full stack walk. But at least UnwindLLDB will delay doing a full
walk if it can.
llvm-svn: 118477
don't crash if we disable logging when some code already has a copy of the
logger. Prior to this fix, logs were handed out as pointers and if they were
held onto while a log got disabled, then it could cause a crash. Now all logs
are handed out as shared pointers so this problem shouldn't happen anymore.
We are also using our new shared pointers that put the shared pointer count
and the object into the same allocation for a tad better performance.
llvm-svn: 118319
fixed them. Added DISALLOW_COPY_AND_ASSIGN to classes that should
not be bitwise copied. Added default initializers for member
variables that weren't being initialized in the ctor. Fixed a few
shadowed local variable mistakes.
llvm-svn: 118240
RegisterContextLLDB holds a reference to the SymbolContext
in the vector of Cursors that UnwindLLDB maintains. Switch
UnwindLLDB to hold a vector of shared pointers of Cursors
so this reference doesn't become invalid.
Correctly falling back from the "fast" UnwindPlan to the
"full" UnwindPlan when additional registers need to be
retrieved.
llvm-svn: 118218
gets to be a problem if you have a unbounded stack walk.
Fix the CFA sanity checks. Add one to the arch default UnwindPlan run which was giving
one extra stack frame on the main thread. Fix a couple of logging lines that had their
argument order incorrect.
llvm-svn: 117350
prologue profiling.
Change the log print statements to elide the thread number, make
some of them only print when IsLogVerbose().
Add a couple of sanity checks for impossible CFA values so backtraces
don't go too far off into the weeds.
llvm-svn: 117343
Not yet enabled as the default unwinder but there are no known
backtrace problems with the code at this point.
Added 'log enable lldb unwind' to help diagnose backtrace problems;
this output needs a little refining but it's a good first step.
eh_frame information is currently read unconditionally - the code
is structured to allow this to be delayed until it's actually needed.
There is a performance hit when you have to parse the eh_frame
information for any largeish executable/library so it's necessary
to avoid if possible.
It's confusing having both the UnwindPlan::RegisterLocation struct
and the RegisterConextLLDB::RegisterLocation struct, I need to rename
one of them.
The writing of registers isn't done in the RegisterConextLLDB subclass
yet; neither is the running of complex DWARF expressions from eh_frame
(e.g. used for _sigtramp on Mac OS X).
llvm-svn: 117256
Abhishek Aggarwal