updates vendored version of x86asm, adds a symbol lookup function to
pass to the disassembler.
This will show global symbol names in the disassembly like go tool
objdump does.
Registers XMM1 and XMM2 get sometimes clobbered between the time we set
them and the panic. There is no guarantee that they won't in the go
spec so we shouldn't expect any register to keep its value. However
since this seems to only affect 1 and 2 let's try to use 9 and 10
instead.
Every time we read an empty string we accidentally issue a read for 0
bytes at address 0, this is fine for real memory but the core file
reader doesn't like it.
Fixes an issue reported on the mailing list.
If a breakpoint is hit close to process death on a thread that isn't
the group leader the process could die while we are trying to stop it.
This can be easily reproduced by having the goroutine that's executing
main.main (which will almost always run on the thread group leader)
wait for a second goroutine before exiting, then setting a breakpoint
on the second goroutine and stepping through it (see TestIssue1101 in
proc_test.go).
When stepping over the return instruction of main.f the deferred
wg.Done() call will be executed which will cause the main goroutine to
resume and proceed to exit. Both the temporary breakpoint on wg.Done
and the temporary breakpoint on the return address of main.f will be in
close proximity to main.main calling os.Exit() and causing the death of
the thread group leader.
Under these circumstances the call to native.(*Thread).waitFast in
native.(*Thread).halt can hang forever due to a bug similar to
https://sourceware.org/bugzilla/show_bug.cgi?id=12702 (see comment in
native.(*Thread).wait for an explanation).
Replacing waitFast with a normal wait work in most circumstances,
however, besides the performance hit, it looks like in this
circumstances trapWait sometimes receives a spurious SIGTRAP on the
dying group leader which would cause the subsequent call to wait in
halt to accidentally reap the process without noting that it did exit.
Instead this patch removes the call to wait from halt and instead calls
trapWait in a loop in setCurrentBreakpoints until all threads are set
to running=false. This is also a better fix than the workaround to
ESRCH error while setting current breakpoints implemented in 94b50d.
Fixes#1101
If the last entry of the package path contains a '.' the corresponding
DIEs for its types will replace the '.' character with '%2e'. We must
do the same when resolving the package path of the concrete type of an
interface variable.
Fixes#1137
debug_info entries can use DW_AT_abstract_origin to inherit the
attributes of another entry, supporting this attribute is necessary to
support DW_TAG_inlined_subroutine.
Go, starting with 1.10, emits DW_TAG_inlined_subroutine entries when
inlining is enabled.
* Handle race between fork and task_for_pid
On macOS a call to fork and a subsequent call to task_for_pid will race each other. This is because the macOS kernel assigns a new proc_t structure early but the new task, thread and uthread come much later. The function exec_mach_imgact in the XNU sources contains this logic.
In a system under load or one with delays in fork processing (i.e. various security software), task_for_pid as currently called by Delve often returns the parent task. This can be seen by printing out the task number around line 86. In a normal system we would see three calls:
-> ~/go/bin/dlv --listen=localhost:59115 --headless=true --api-version=2 --backend=native exec ./___main_go --
Task: 9731
Task: 9731
Task: 9731
API server listening at: 127.0.0.1:59115
This is the result on a system where the race is lost:
-> ~/go/bin/dlv --listen=localhost:59115 --headless=true --api-version=2 --backend=native exec ./___main_go --
Task: 8707
Task: 10499
Task: 10499
could not launch process: could not get thread count
In this latter case, task 8707 is the parent task. The child task of 10499 was desired and hence the error.
This code change checks to make sure the returned task is not that of the parent. If it is, it retries. It's possible other macOS reported Delve issues are the result of this failed race.
* proc: correct formatting
Much like the bug in issue #1031 and commit
f6f6f0bf13e4c708cb501202b83a6327a0f00e31 pointers can also escape to
the heap and then have a zero address (and no children) when we
autodereference.
1. Mark autodereferenced escaped variables with a 0 address as
unreadable.
2. Add guards to the pretty printers for unsafe.Pointer and pointers.
Fixes#1075
Depending on how the runtime schedules our goroutines we can get
unlucky and have the first call to runtime.newstack we intercept be for
a different goroutine (usually the garbage collector).
Only check stacktraces that happen on the same goroutine that executed
main.main.
The runtime calls into g0 in many places, not necessarily using
runtime.systemstack or runtime.asmcgocall.
One example of this is the call to runtime.newstack inside
runtime.morestack.
If we stop the process while one goroutine is executing
runtime.newstack we would be unable to fully scan its stack because we
don't know that we have to switch back to the goroutine stack after
runtime.newstack.
Instead of tracking down every possible way that the runtime switches
to g0 we switch to the goroutine stack immediately after the top of the
stack, unless cgo is being executed on the systemstack.
Fixes#1066
Adds a configuration option (show-location-expr) that when activated
will cause the whatis command to also print the DWARF location
expression for a variable.
On macOS, externally linked programs will have an abbrev for
DW_TAG_subprogram without the haschildren flag set. We should handle
this case instead of expecting all DW_TAG_subprogram entries to have
list of children.
Fixes#1034
When creating a stack trace we should switch between the goroutine
stack and the system stack (where cgo code is executed) as appropriate
to reconstruct the logical stacktrace.
Goroutines that are currently executing on the system stack will have
the SystemStack flag set, frames of the goroutine stack will have a
negative FrameOffset (like always) and frames of the system stack will
have a positive FrameOffset (which is actually just the CFA value for
the frame).
Updates #935
Either the CPU or the kernel may not support the calls we do when
retrieving floating point registers, this isn't an error we should
propagate.
Also improve the error reporint of pkg/proc/native.fpRegisters.
Fixes#1022
StepBreakpoints are set on CALL instructions, when they are hit we
disassemble the current instruction, figure out the destination address
and set a breakpoint after the prologue of the called function.
In order to disassemble the current instruction we disassemble the area
of memory starting from PC and going to PC+15 (because 15 bytes is the
maximum length of one instruction on AMD64). This means that we won't
just disassemble one instruction but also a few instructions following
it ending with one truncated instruction.
This usually works fine but sometimes the disassembler will panic with
an array out of bounds error when trying to disassemble a truncated
instruction. To avoid this problem this commit changes the funciton
disassemble to take one extra parameter, singleInstr, when singleInstr
is set disassemble will quit after disassembling a single instruction.
Replace the unsafe.Pointer type of the buf field of channels with the
appropriate array type, allow expressions accessing member field of the
channel struct.
Fixes#962
Conditional breakpoints with unmet conditions would cause next and step
to skip the line.
This breakpoint changes the Kind field of proc.Breakpoint from a single
value to a bit field, each breakpoint object can represent
simultaneously a user breakpoint and one internal breakpoint (of which
we have several different kinds).
The breakpoint condition for internal breakpoints is stored in the new
internalCond field of proc.Breakpoint so that it will not conflict with
user specified conditions.
The breakpoint setting code is changed to allow overlapping one
internal breakpoint on a user breakpoint, or a user breakpoint on an
existing internal breakpoint. All other combinations are rejected. The
breakpoint clearing code is changed to clear the UserBreakpoint bit and
only remove the phisical breakpoint if no other bits are set in the
Kind field. ClearInternalBreakpoints does the same thing but clearing
all bits that aren't the UserBreakpoint bit.
Fixes#844
Move some duplicate code, related to breakpoints, that was in both
backends into a single place.
This is in preparation to solve issue #844 (conditional breakpoints
make step and next fail) which will make this common breakpoint code
more complicated.
Instead of only tracking a few cherrypicked registers in stack.go track
all DWARF registers.
This is needed for cgo code and for the locationlists emitted by go in
1.10:
* The debug_frame sections emitted by C compilers can not be used
without tracking all registers
* the loclists emitted by go1.10 need all registers of a frame to be
interpreted.
A thread could terminate between the point when we stop for a
breakpoint and the point where we send a stop signal to all threads, if
this happens setCurrentBreakpoints will fail with an error.
We should tolerate this.
For some reason this happens very frequently when running delve on
processes with the race detector enabed.
Adds test command line flag to compile target fixtures using the -race flag.
Multiple tests will fail because of https://github.com/golang/go/issues/22600
but eventually this should work.
1. Use a slice instead of a map to access standard and extended opcodes
(reduces BenchmarkStateMachine from ~12ms/op to ~7ms/op)
2. Cache StateMachine values for the entry point of functions.
gosymtab and gopclntab only contain informations about go code, linked
C code isn't there, we should use debug_line instead to also cover C.
Updates #935
* string to []rune
* string to []byte
* []rune to string
* []byte to string
* any pointer to uintptr
The string, []rune, []byte conversion pairs aligns this to the go
language.
The pointer -> uintptr conversion pair is symmetric to the uintptr ->
pointer that we already have.
Also lets the user specify any size for byte array types instead of
just the ones already used by the program, this can be used to read
arbitrary memory.
Fixes#548, #867
Make 'next' skip deferred functions unless they are called via a panic.
Call to a deferred function through 'return' are predictable, if the
user wants to step into them 'step' can be used but without this change
there is no way to avoid stepping into them.
Implements #956
On macOS we can also stop when we receive a signal,
propagate this reason upwards to the client.
Also clear internal breakpoints after an unrecovered-panic since they
can not be reached anymore.
Fixes#872
debugserver doesn't support qXfer:exec-file:read, and it doesn't return
the executable path in the response to qProcessInfoPID, however we can
find out the executable path by using jGetLoadedDynamicLibrariesInfos.
While we are waiting for the process to exit in native.(*Process).Kill
we could receive queued exception events, those must be continued or
the wait will never finish.
Before go1.9 embedded struct fields had name == "" in runtime and ==
type name in DWARF. After go1.9 both runtime and DWARF use a simplified
version of the type as name.
Embedded structs are distinguished from normal fields by setting a flag
in the runtime.structfield, for runtime, and by adding a custom
attribute in DWARF.
Splits out type parsing and go-specific Type hierarchy from
x/debug/dwarf, replace x/debug/dwarf with debug/dwarf everywhere,
remove x/debug/dwarf from vendoring.
Instead of panicing for sending on a closed channel, detect that the
process has exited and return a proper error message.
This patch also cleans up some spots where the Pid is omitted from the
error.
Fixes#920
* proc/native: make sure debugged executable can be deleted on windows
Delve opens debugged executable to read binary info it
contains, but it never closes the file. Windows will not
let you delete file that is opened. So close Process.bi
in Process.postExit, and actually call Process.postExit
from windows Process.Kill.
Also Windows sends some debugging events
(EXIT_PROCESS_DEBUG_EVENT event in particular) after Delve
calls TerminateProcess. The events need to be consumed by
debugger before debugged process will be released by
Windows. So call Process.waitForDebugEvent after
TerminateProcess in Process.Kill.
Fixes#398
* cmd/dlv: make TestIssue398 pass on darwin
* cmd/dlv: add comment for TestIssue398
* proc/native: wait for debuggee to exit before returning from windows Process.Kill
* proc/native: close process handle before returning from windows killProcess
* proc/native: remove not used Process.Process
When stepping through runtime sometimes the current goroutine will
change. It is impossible to handle this in Next, Step and StepOut but
StepInstruction can reset the current goroutine correctly.
Next will add internal breakpoints with nil condition if it can't find
the current goroutine (possibly because there isn't a current goroutine
because the runtime hasn't been initialized yet).
onNextGoroutine should skip breakpoints with nil condition, otherwise
we'll end up with an internal debugger error trying to walk a nil
expression.
Updates #893
When there's a error reading the stack trace the call stack itself
could be corrupted and we should return the partial stacktrace that we
have.
Fixes#868
Variable lookup is slow because it requires a full scan of debug_info
to check for package variables, this doesn't matter much in interactive
use but can slow down evaluation of breakpoint conditions
significantly.
Providing benchmark proof for this is hard since this effect doesn't
show for small programs with small debug_info sections.
* proc: fix interaction of RequestManualStop and conditional breakpoints
A conditional breakpoint that is hit but has the condition evaluate to
false can block a RequestManualStop from working. If the conditional
breakpoint is set on an instruction that is executed very frequently by
multiple goroutines (or many conditional breakpoints are set) it could
prevent all calls to RequestManualStop from working.
This commit fixes the problem by changing proc.Continue to exit
unconditionally after a RequestManualStop is called.
* proc/gdbserial: fix ContinueOnce getting stuck on macOS
Fixes#902
The concrete value of an interface is always stored as a pointer inside
an interface variable. So far we have followed the memory layout and
reported the type of the 'data' attribute of interfaces as a pointer,
however this makes it impossible to distinguish interfaces with
concrete value of type 'A' from interfaces of concrete value of type
'*A'.
With this changeset when we autodereference pointers when the concrete
type of an interface is not a pointer.
* Fix various issues detected by megacheck
I've ran honnef.co/go/tools/cmd/megacheck and fixed a few of the
things that came up there.
* Cleanup using Gogland
When a Go program is externally linked, the external linker is
responsible for picking the TLS offset. It records its decision in the
runtime.tlsg symbol. Read the offset from that rather than guessing -16.
This implementation causes a regression: 1.4 and earlier will no longer
work.
Implementing proc.Process.Running in a thread safe way is complicated
and nothing actually uses it besides tests, so we are better off
rewriting the tests without Running and removing it.
In particular:
* The call to d.target.Running() in service/debugger/debugger.go
(Restart) can never return true because that line executes while
holding processMutex and all continue operations are also executed
while holding processMutex.
* The call to dbp.Running() pkg/proc/native/proc.go (Detach) can never
return true, because it's only called from
debugger.(*Debugger).detach() which is also always called while
holding processMutex.
Since some tests are hard to write correctly without Process.Running a
simpler interface, Process.NotifyResumed, is introduced.
Fixes#830
RequestManualStop will run concurrently with trapWait, since one writes
dbp.halt and the other reads it dbp.halt should be protected by a
mutex.
Updates #830
type M struct was never used (as far as I know).
type VariableEval interface was used for a brief period of time during
the refactoring, now both its methods are functions.
The authorization prompt on macOS can take a long time to be
acknowledged by the user, we should keep waiting for a connection as
long as the debugserver instance we launched remains alive.
Before this commit our temp breakpoints only checked that we would stay
on the same goroutine.
However this isn't enough for recursive functions we must check that we
stay on the same goroutine AND on the same stack frame (or, in the case
of the StepOut breakpoint, the previous stack frame).
This commit:
1. adds a new synthetic variable runtime.frameoff that returns the
offset of the current frame from the base of the call stack.
This is similar to runtime.curg
2. Changes the condition used for breakpoints on the lines of the
current function to check that runtime.frameoff hasn't changed.
3. Changes the condition used for breakpoints on the return address to
check that runtime.frameoff corresponds to the previous frame in the
stack.
4. All other temporary breakpoints (the step-into breakpoints and defer
breakpoints) remain unchanged.
Fixes#828
Debugserver does not work as documented, "--" needs to be specified to
pass arguments to the target process (but only if it's an argument that
starts with a dash).
Fixes#839
While implementing the gdbserial backend everything was changed to call
Detach to "close" a process so that gdbserial could do its clean up in
a single place. However the native implementation of Detach does not
actually kill processes we launched.
Fixes#821
Lldb-server does not currently support windows/amd64 (it's in
development). And delve's native backend seems to work well on windows.
Disable the lldb backend on windows for now.
Fixes#817
- moved target.Interface into proc as proc.Process
- rename proc.IThread to proc.Thread
- replaced interfaces DisassembleInfo, Continuable and
EvalScopeConvertible with Process.
- removed superfluous Gdbserver prefix from types in the gdbserial
backend.
- removed superfluous Core prefix from types in the core backend.
Core files contain a variety of memory mappings either to files or
anonymous regions stored in the core file. These regions can overlap, so
figuring out what exactly to read can be tricky. This commit contains
a data structure, SplicedMemory, which accumulates mappings and reads
from the correct sources.
* pkg/proc: use golang.org/x/arch/x86/x86asm
instead of rsc.io/x86/x86asm
* pkg/dwarf: migrate to github.com/pkg/profile
from github.com/davecheney/profile
* scripts: keep script go files from being considered for the build
scripts/gen-*.go files are scripts for generating documentation
files and don't follow the typical Go package layout. Expected
usage is like
go run scripts/gen-cli-docs.go
* vendor: update vendored packages
There were many changes in delve, and go tool chains since last
vendored package update. I just rerun godpes from scratch.
$ rm vendor/*
$ rm Godeps/Godeps.json
$ go list ./... | grep -v /vendor/ | grep -v /scripts/ | go get -u -t
$ go get -u github.com/mattn/go-colorable
$ go get -u github.com/mattn/go-isatty
$ go list ./... | grep -v /vendor/ | grep -v /scripts/ | godeps save
* proc: Refactor stackIterator to use memoryReadWriter and BinaryInfo
* proc: refactor EvalScope to use memoryReadWriter and BinaryInfo
* proc: refactor Disassemble to use memoryReadWriter and BinaryInfo
* proc: refactor BinaryInfo part of proc.Process to own type
The data structures and associated code used by proc.Process
to implement target.BinaryInfo will also be useful to support a
backend for examining core dumps, split this part of proc.Process
to a different type.
* proc: compile support for all executable formats unconditionally
So far we only compiled in support for loading the executable format
supported by the host operating system.
Once support for core files is introduced it is however useful to
support loading in all executable formats, there is no reason why it
shouldn't be possible to examine a linux coredump on windows, or
viceversa.
* proc: bugfix: do not resume threads on detach if killing
* Replace BinaryInfo interface with BinInfo() method returning proc.BinaryInfo
Detach did not work for processes we attach to via PID.
Linux: we were only detaching from the main thread, all threads are
detached independently
Windows: we must resume all threads before detaching.
macOS: still broken.
Updates #772
Stack barriers were removed in Go 1.9, and thus code that
expected various stack-barrier-related symbols to exist
does not find them. Check for their absence and do not
crash when they are missing. Disable stack-barrier-handling
test for 1.9 and beyond.
Fixes#754.
The rest of the code uses Detach to "close" a proc.Process, the tests
should do the same.
Any cleanup that proc.Process needs to do can then be put inside Detach
and the tests will run it.
It was possible for TestStepParked to pick a runtime goroutine and
attempt to step it. Nothing guarantees that a goroutine other than the
ones we are using to run the code would actually ever resume before the
end of the program.
This makes the test more discerning in its choice of goroutines.
On Windows we can sometimes encounter threads stopped in locations for
which we do not have entries in debug_frame.
These cases seem to be due to calls to Windows API in the go runtime,
we can still produce a (partial) stack trace in this circumstance by
following frame pointers (starting with BP).
We still prefer debug_frame entries when available since go functions
do not have frame pointers before go1.8.
Sometimes windows will send us events about breakpoints we have
already removed from the code despite the fact that we go to great
lengths to avoid this already.
Change waitForDebugEvent to check that when we receive a breakpoint
event the corresponding memory actually contains an INT 3
instruction, if it doesn't ignore the event and restart the thread.
Changes to the GC in 1.8 make this test fail because the GC isn't
inserting stack barriers into our test program anymore. This isn't the
same thing as being unable to print stacktraces in presence of stack
barriers so the test shouldn't fail.
The PC we have is relative to the first instruction after the CALL
instruction currently being executed.
Anyone watching a disassembly will understand what's happening if we
report the return PC, but reporting the first PC of the current line is
useless and confusing.
