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.
the instruction after the call for a normal frame. libunwind uses
IP-1 to lookup unwind information. However, this is not necessary for
interrupted frames such as signal frames (or interrupt frames) in
the kernel context.
This patch handles both cases correctly.
Based on work by Mark Wielaard <mwielaard@redhat.com>
This is rather on the obvious side.
While doing strace on an executable using libunwind, I noticed a
lot of:
msync(0, 1, MS_SYNC) = -1 ENOMEM (Cannot allocate memory)
Since we know that the first page isn't mapped (or at least doesn't
contain the data we are looking for), we can eliminate all such
msync calls.
Tested on Linux/x86_64 with no regressions.
Greetings,
We use libunwind just for stack traces (I suspect many others do as well).
The use pattern is:
GetStackTrace(void** result, int max_depth)
{
...
unw_getcontext(&uc);
unw_init_local(&cursor, &uc);
while (n < max_depth) {
if (unw_get_reg(&cursor, UNW_REG_IP, (unw_word_t *) &ip) < 0) {
break;
}
result[n++] = ip;
if (unw_step(&cursor) <= 0) {
break;
}
}
Given this usage, it is quite convenient for us to block signals (or
prevent signal handlers from re-entering libunwind by other means) at the
"top level", which makes most of the sigprocmask calls performed by
libunwind itself unneccessary.
The second patch in this series adds a configure option which removes most
of the sigprocmask calls.
Attached patch is a preliminary for it -- consolidating all of the
"sigprocmask; mutex_lock;" sequences into lock_acquire and "mutex_unlock;
sigprocmask;" sequences into lock_release.
Thanks,
--
Paul Pluzhnikov
commit 402d15b123d54a7669db7cf17a76dd315094e472
Author: Paul Pluzhnikov <ppluzhnikov@google.com>
Date: Mon Sep 21 10:18:28 2009 -0700
Replace "sigprocmask + mutext_lock" with a single lock_acquire.
Likewise, replace "mutext_unlock + sigprocmask" with lock_release.
getcontext in libc.
Also cleanup the namespace (check-name-space passes on x86_64 now).
Replace uses of offsets.h with ucontext_i.h.
Rename _x86_64_setcontext to _Ux86_64_setcontext.
TBD: Add CFI annotations for get/setcontext.
Signed-off-by: Paul Pluzhnikov <ppluzhnikov@google.com>
Signed-off-by: Arun Sharma <arun.sharma@google.com>
bad/missing unwind information, which could result in libunwind
dereferencing bad pointers. This mechanism is based on msync(2) system
call and significantly reduces the chances of a bad pointer
dereference in libunwind.
The original idea was to turn this mechanism on only when necessary
i.e. libunwind didn't find proper unwind information for a IP.
There are a couple of problems in the current implementation.
* The flag is global and is modified without locking
* The flag isn't reset when starting a new unwind
The attached patch makes ->validate a per-thread setting by moving it
into struct cursor from unw_local_addr_space and resets it to false
when starting a new unwind. As a result, cursor->as_arg points to the
cursor itself instead of the ucontext (for the local case).
This was found to reduce the number of msync() system calls from an
application using libunwind significantly.
Signed-off-by: Paul Pluzhnikov <ppluzhnikov@google.com>
Signed-off-by: Arun Sharma <arun.sharma@google.com>
The current pattern is too restrictive and doesn't work well on
modern glibcs.
Signed-off-by: Paul Pluzhnikov <ppluzhnikov@google.com>
Signed-off-by: Arun Sharma <arun.sharma@google.com>
On some systems executable stacks are denied. Since libunwind and the
tests don't actually need executable stacks this patch marks all
assembly files as not needing it.
The original patch comes from frysk:
2007-04-05 Jan Kratochvil <jan.kratochvil@redhat.com>
* src/hppa/getcontext.S, src/hppa/setcontext.S, src/hppa/siglongjmp.S,
src/ia64/Ginstall_cursor.S, src/ia64/Linstall_cursor.S,
src/ia64/dyn_info_list.S, src/ia64/getcontext.S, src/ia64/longjmp.S,
src/ia64/setjmp.S, src/ia64/siglongjmp.S, src/ia64/sigsetjmp.S,
src/ppc64/longjmp.S, src/ppc64/siglongjmp.S, src/x86/longjmp.S,
src/x86/siglongjmp.S, src/x86_64/longjmp.S, src/x86_64/setcontext.S,
src/x86_64/siglongjmp.S: Stack should be non-executable, for SELinux.
I added a couple more markers for new files in current libunwind.
Before this patch you would get the following on selinux enabled
systems without allow_exec_stack: error while loading shared
libraries:
libunwind.so.7: cannot enable executable stack as shared object
requires: Permission denied
After the patch that error disappears and all test results are similar
to the results on systems without executable stack protection.
routine and add address-space argument. This is needed because on
PPC64, a the function-name symbol refers to a function descriptor
(unlike, for example, on ia64, where the @fptr() operator is needed to
refer to a function descriptor). Thus, in order to look up the name
of a function, we need to dereference the function descriptor. To
make matters more "interesting", the function descriptors are normally
resolved by the dynamic linker, so we can't get their values from the
ELF file. Instead, we have to read them from the running image, hence
the need for the address-space argument.
This is so that the source file gets distributed and _ucontext_i.h is
generated properly on the target machine.
Signed-off-by: Arun Sharma <arun.sharma@google.com>
When libunwind is linked with a C++ program that throws exceptions,
the exception that's thrown is passed in %rax. However, libc's
setcontext clears %rax, causing problems.
This patch implements a setcontext that doesn't clobber rax.
TBD: Add dwarf CFI annotations
Signed-off-by: Arun Sharma <arun.sharma@google.com>
CFA should be incremented by 16 in the hope that the previous frame
may have valid unwind info.
Also increase the default frame pointer recognition heuristic from 4k
to 16k.
Signed-off-by: Andrey Veskov <Andrey.Veskov@intel.com>
Signed-off-by: Arun Sharma <arun.sharma@google.com>
a signal-trampoline, assume that it's a PLT stub.
If non-DWARF stepping fails to change IP and CFA, declare
it a bad frame.
2004/11/23 16:59:56-08:00 mostang.com!davidm
(unw_step): Also print IP as part of the function-trace.
2004/11/23 16:17:37-08:00 mostang.com!davidm
(unw_step): When dwarf_step() fails on a signal-frame, fill in all
the known locations because dwarf_step() fails on older
kernels which don't export the kernel vDSO even though every-
thing else may be providing proper DWARF unwind-info.
2004/10/25 17:43:57+02:00 homeip.net!davidm
Add Debug statement for return-value.
(Logical change 1.290)
