When adjusting the stack for a DW_CFA_arg_size adjustment, ensure that
we use the target dependent register name as the generic name does not
necessarily map to the same register. For example, on x86, ESP maps to
the eip register, which results in the wrong stack adjustment being
applied.
Currently setcontext for x86_64 restores the signal mask, even
though it is never saved anywhere. This means the signal mask
is often garbage after an unw_resume.
(changed in commit f8a15e9679)
It looks like this was a fix for the Gtest-resume-sig function -
testing if signal masks are restored across signal frames. The
root issue looks like that x86_64 only uses sigreturn for the exact
signal frame, and not for any decedant frames as well (as i64 does).
Instead, modify Gresume to use sigreturn if *any* frame on the stack
is a signal frame, so that we correct fixup the signal mask and any
sigaltstacks. The sigreturn os-specific functions are changed slightly
to copy in the saved ucontext structure if we are jumping farther
up the stack.
This should fix sigprocmask reported issues such as
https://github.com/dropbox/pyston/blob/master/libunwind_patches/0002-pyston-stop-x86_64-setcontext-restoring-uninitialize.patch
Tests pass on freebsd, linux
When resuming execution, DW_CFA_GNU_args_size from the current frame
must be added back to the stack pointer. Clang now generates these frequently
at -O3. A simple repro for x86_64, that will crash with clang ~3.9 or newer:
void f(int, int,int,int,int,int,int,int,int);
int main() {
try {
f(0,1,2,3,4,5,6,7,8);
} catch (int) {
return 0;
}
return 1;
}
Where f is something that throws an int, but in a different translation unit to
prevent optimization.
This results in cfi instructions before the call:
.cfi_escape 0x2e, 0x20
Grabbing the args_size means fully parsing the cfi in the current frame, which
is unfortunate because it means nearly twice the work at each step. The logic
to grab args_size can be in unw_step or get_proc_info (since this is always
called before resuming in stack unwinding). Putting it in get_proc_info allows
the more common unw_step code to remain fast.
It would potentially fit in nicely with a proc info cache (as mentioned in the
if0 comment block)
GCC versions 4.9~current will often generate stack alignment prologues like:
lea 0x8(%rsp),%r10
and $0xfffffffffffffff0,%rsp
...
push %rbp
mov %rsp, %rbp
push %r10
resulting in dwarf expressions:
DW_CFA_def_cfa_expression (DW_OP_breg6: -8; DW_OP_deref)
DW_CFA_expression: r6 (rbp) (DW_OP_breg6: 0)
These prologues seem to be generated for SSE/AVX code, but sometimes
other times as well.
tdep_trace fastpath currently falls back to the slow dwarf parsing path
if it encounters any cfa_expressions. Unfortunately this is happening
often enough in our codebase to cause perf issues. We could also fix the
fallback path (make the rs cache bigger, lock-free instead of locking, etc),
but that seems like a separate issue, and it will ever be as fast as the tracing
code. Our binaries each have at least ~100 functions in them like this.
This patch teaches the tdep_trace about the two specific cfa_expressions,
which really just result in a single extra memory dereference of the stack
at a fixed offset from rbp.
unw_is_signal_frame() returns <= 0 if not a signal frame. Several places in
code were only checking for a "if (unw_is_signal_frame())", or
"if (!unw_is_signal_frame())".
The detection logic introduced in 28f33c8ce0 is
broken, because it tests mincore using an address that is almost certainly not
page-aligned. straces confirms that msync is used all the time.
This patch fixes the logic by page-aligning the test pointer. strace now shows
that mincore is actually used. Furthermore, the return value of mincore is not
sufficient to assess whether the address can be safely dereferenced: we should
also check that the pages are mapped, through the passed mvec array. This patch
also adds this verification.
Tested on a system where unwinding a stack across a JNI boundary would cause
sporadic segfaults; no more crashes were observed after the patch.
By default, the start_ip_offset in libunwind's table_entry struct is
relative to the unw_dyn_info_t's segbase. This presents a problem
for us in conjunction with using LLVM's MCJIT because it likes to
spread text sections and the corresponding eh_frame sections quite
far apart. This represents my attempt to support this use case in the
simplest manner that is backwards compatible, by adding a new format
kind (UNW_INFO_FORMAT_REMOTE_TABLE2) that indicates that the
`start_ip_offset` should be interpreted as relative to `start_ip`
rather than segbase.
We've just traced a large memory increase to that patch (Google ref:
b/18069427).
It appears that labs() was there for a good reason.
Sorry about that :-(
For the curious:
unsigned long u1 = ~0UL;
unsigned long u2 = labs(u1);
assert(u1 != u2); // labs on unsigned *may* have an effect, despite
what Clang says.
Attached patch suppresses the Clang warning, while still keeping the
original behavior (which I believe to be correct).
Thanks,
--
Paul Pluzhnikov
Greetings,
Roman reports that Clang warns on unnecessary calls to labs():
taking the absolute value of unsigned type 'unw_word_t' (aka
'unsigned long') has no effect [-Werror,-Wabsolute-value]
Since rs->reg[...].val is unw_word_t and unsigned on all platforms,
this patch removes the unnecessary calls to labs().
Tested on Linux x86_64, no regressions.
Thanks,
--
Paul Pluzhnikov
Mark frames which are unwound with the frame-chain walker or
syscall frame code, as non-interrupted. The return PC in the frame
points to the instruction after the call.
When JITs generate code without unwind information, it may be possible
to continue unwinding via RBP chaining. However, we currently disallow
RBP==RSP condition even though we can make forward progress.
Relax the check a bit in the code where we switch from one type of
unwinding to another to handle this situation. JIT authors
are encouraged to use the dynamic unwind info registration API when
the underlying platform supports it.
Signed-off-by: Kevin Modzelewski <kmod@dropbox.com>
Check the endianness earlier in unw_create_addr_space() on all
architectures to avoid leaking the dynamically allocated address space
struct.
This was already fixed for ARM in commit cf6a998796 ("Fix memory leak
in ARM unw_create_addr_space()"). Move the endianness check also on ARM
to avoid malloc() & free() in the error case.
This reverts commit c9c5a40be1.
dwarf_get() returns 0 on success. We should continue unwinding
in that case.
TBD: investigate test-ptrace failure on some platforms.
If dwarf_get returns 0 (indicating the end of call chain), we should
not override the return value to 1. This may result in the caller
continuing to unwind and getting spurious errors.
In the commit d04dc94cc2, the check for
dwarf.ip == 0 was removed from non-dwarf walker in x86_64 version of
unw_step(). Apparently this broke the detection of the end of frame
chain when NULL %rbp is specified, because the case just marked
dwarf.ip as 0. Explicitly set ret to 0 to indicate the end of
iteration.
There is a window of time between the munmap and the tls_cache being
marked as destroyed, where there could be a bad access to memory that
has been unmapped/freed. Reorder the code a bit to close the window.
Signed-off-by: Paul Pluzhnikov <ppluzhnikov@google.com>
The crashes were tracked down to f->rpb_cfa_offset being incorrect.
The problem is that {rsp,rbp}_cfa_offset only have 15 bits, but for
SIGRETURN frame they are filled with:
// src/x86_64/Gstash_frame.c
f->cfa_reg_offset = d->cfa - c->sigcontext_addr;
f->rbp_cfa_offset = DWARF_GET_LOC(d->loc[RBP]) - d->cfa;
f->rsp_cfa_offset = DWARF_GET_LOC(d->loc[RSP]) - d->cfa;
The problem is that the delta here can be arbitrarily large when
sigaltstack is used, and can easily overflow the 15 and 30-bit fields.
When signal handler starts running, the stack layout is:
... higher addresses ...
ucontext
CFA->
__restore_rt (== pretcode in rt_sigframe from
linux-2.6/arch/x86/include/asm/sigframe.h)
SP ->
... sighandler runs on this stack.
... lower addresses ...
This makes it very convenient to find ucontext from the CFA.
Attached patch re-tested on Linux/x86_64, no new failures.
Signed-off-by: Paul Pluzhnikov <ppluzhnikov@google.com>
Reviwed-by: Lassi Tuura <lat@cern.ch>
Glibc calls thread-specific dtors in the order in which the keys were added,
so the first dtor is the trace_cache_free() one. Then thread-specific
data for some other key is free()d, which calls into unw_backtrace(),
which uses dangling cache and munmapped cache->frames.
[ Minor rename + compiler warning fix: asharma@fb.com ]
Signed-off-by: Paul Pluzhnikov <ppluzhnikov@google.com>
On FreeBSD, as well as on the Solaris < 10, weak pthread_once stub is
always exported from libc. But it does nothing, which means that if
threaded library is not loaded, then pthread_once() call do not actually
call the initializer finction. The construct
if (likely (pthread_once != 0))
{
pthread_once(&trace_cache_once, &trace_cache_init_once);
then fails to initialize the trace cache on x86_64.
Work around by checking that the initializer was indeed called.
Note that this can break if libthr is loaded dynamically, but my belief
is that there is no platforms which allow dynamic loading of the threading
library.
Insert static branch prediction predicates in useful places and avoid
unnecessary code in the hottest paths. Bypass unnecessary indirect
calls, in particular to access_mem(), when known to be safe.
Since the fast unwinding code path doesn't need the full context,
a faster target dependent getcontext is implemented.
Signed-off-by: Lassi Tuura <lat@cern.ch>
Dropping the extra frame for unw_backtrace itself using unw_step is
approximately 15% slower than skipping the frame in tdep_trace. So
drop the frame in the latter, and make the function a private
implementation detail for libunwind, not an exported interface.
Also moves unw_getcontext call back into unw_backtrace to avoid an
extra call frame in case slow_backtrace does not get inlined into
unw_backtrace.
Adds new function to perform a pure stack walk without unwinding,
functionally similar to backtrace() but accelerated by an address
attribute cache the caller maintains across calls.
