IF YOU WOULD LIKE TO GET AN ACCOUNT, please write an
email to Administrator. User accounts are meant only to access repo
and report issues and/or generate pull requests.
This is a purpose-specific Git hosting for
BaseALT
projects. Thank you for your understanding!
Только зарегистрированные пользователи имеют доступ к сервису!
Для получения аккаунта, обратитесь к администратору.
Some system calls require capturing the stack trace before they are
processed in kernel. Typical one is execve. Some system calls require
invalidating mmap cache after they are processed in kernel.
In current implementation these requirements are handled directly by
appropriate syscall handlers. However, it is difficult to keep the
source code maintainable using this approach to cover all system calls
which have such requirements.
A more generic way to implement this is to flag all syscalls that
require special processing, and handle these flags right in
trace_syscall_entering instead of changing syscall handlers.
This patch just defines new flags: STACKTRACE_INVALIDATE_CACHE and
STACKTRACE_CAPTURE_ON_ENTER.
The names of macros are suggested by Dmitry Levin.
Signed-off-by: Masatake YAMATO <yamato@redhat.com>
Here is the benchmark of the dwarf cache.
Target program:
#include <sched.h>
int main(void)
{
unsigned int max = 0x6fff, i;
for (i = 0; i < max; i++)
sched_yield();
return 0;
}
Command line:
./strace -o /dev/null -k a.out
With the dwarf cache:
real 0m12.081s
user 0m3.858s
sys 0m8.194s
Without the dwarf cache:
real 0m22.326s
user 0m5.218s
sys 0m16.952s
Signed-off-by: Masatake YAMATO <yamato@redhat.com>
When a file mmap'ed to the target process is unlink'ed, backtracing the
stack would fail. Current implementation reports it as
"backtracing_error". To avoid confusion, the message is changed to
"expected_backtracing_error".
Here is the reproducer:
$ cat ./p-deleted.c
#include <unistd.h>
int main(int argc, char **argv) {
return unlink(argv[0]) < 0;
}
$ strace -e unlink -k ./p-deleted
unlink("./p-deleted") = 0
> /usr/lib64/libc-2.18.so(unlink+0x7) [0xe7f17]
> /home/yamato/var/strace/t_unwind/p-deleted (deleted)(+0x0) [0x575]
> /usr/lib64/libc-2.18.so(__libc_start_main+0xf5) [0x21d65]
> backtracing_error [0x7ffff1365590]
+++ exited with 0 +++
p-deleted is deleted therefore backtracing_error is reported. This
patch records the deleted marker when making mmap cache and refers the
recorded information in the case "backtracing_error" to switch the
message.
Here is the output of this patch:
$ strace -e unlink -k ./p-deleted
unlink("./p-deleted") = 0
> /usr/lib64/libc-2.18.so(unlink+0x7) [0xe7f17]
> /home/yamato/var/strace/t_unwind/p-deleted (deleted)(+0x0) [0x575]
> /usr/lib64/libc-2.18.so(__libc_start_main+0xf5) [0x21d65]
> expected_backtracing_error [0x7ffff1365590]
+++ exited with 0 +++
This solution is not perfect: if a file is unlink'ed after making the
mmap cache and before unwinding, strace cannot have a chance to record
the deleted marker.
In this version of patch, hardcoded magic number used in comparing "(delete)"
string is replaced with strlen as suggested by Dmitry Levin.
In old version of patch, the deleted entry was thrown away from mmap
cache to avoid to report "backtracing_error". In this patch I keep it,
and just switch the error message.
Inspired by the review comment from Dmitry Levin.
Signed-off-by: Masatake YAMATO <yamato@redhat.com>
captured stacktrace is printed in unwind_tcb_fin if tcp->queue is not
empty. This should happen before printing detached message, so
unwind_tcb_fin is moved to the top of droptcb.
This is implicitly suggested by Dmitry Levin in patch review process.
Signed-off-by: Masatake YAMATO <yamato@redhat.com>
A mmap cache belonging to a tcb was updated when a system call which
changed the memory mapping was called. This implementation was assumed
the mapping was changed only by the tcb. However, this assumption is
incorrect if the target application is multi-threaded; more than two
tcbs can shared the same memory mapping and a tcb can modify it without
being noticed by the others.
This change introduces a global integer variable mmap_cache_generation,
and mmap_cache_generation field to struct tcb. The variable
is incremented each time a process enters a syscall that can modify its
memory mapping. Each tcb records the value of this variable at the
moment if building its mmap cache. Every mmap cache associated with
the given tcb can be validated by comparing its mmap_cache_generation
field with the variable mmap_cache_generation.
This implementation is inefficient. If strace attaches two processes
which don't share the memory mapping, rebuilding mmap cache of a tcb
triggered by another tcb's mmap system call is not necessary.
Signed-off-by: Masatake YAMATO <yamato@redhat.com>
This is the second step for splitting capturing from printing.
New `queue' field is added to tcb. Captured stacktrace is stored here.
The field is initialized/finalized at unwind_tcb_init/unwind_tcb_fin.
New API function unwind_capture_stacktrace is added. This function
captures the currest stack using stracktrace_walker and records it in
tcb. It's printing is delayed to the next call of
unwind_print_stacktrace.
unwind_print_stacktrace is extended. Now it checks queue field of
the given tcb at the start of function. If the function finds a
captured stack trace, the latter is printed using stracktrace_walker.
Currently unwind_capture_stacktrace invocations are added directly to
handlers of mmap, munmap, mprotect, and execve.
Here is the difference of output with/without patch:
(without patch)
execve("./test-fork", ["./test-fork"], [/* 56 vars */]) = 0
> /usr/lib64/ld-2.18.so(check_one_fd.part.0+0x82) [0x11f0]
(with patch)
execve("./test-fork", ["./test-fork"], [/* 54 vars */]) = 0
> /usr/lib64/libc-2.18.so(execve+0x7) [0xbcd27]
> /home/yamato/var/strace/strace(exec_or_die+0x10c) [0x26ac]
> /home/yamato/var/strace/strace(startup_child+0x346) [0x134f6]
> /home/yamato/var/strace/strace(init+0x89f) [0x13dff]
> /home/yamato/var/strace/strace(main+0xa) [0x26ca]
> /usr/lib64/libc-2.18.so(__libc_start_main+0xf5) [0x21d65]
> /home/yamato/var/strace/strace(_start+0x29) [0x2799]
In older version output lines of captured elements were built when
printing. In this version they are built when capturing the stack.
As result, unneeded dynamic memory allocations are avoided.
Suggested by Luca Clementi.
In older version the combination of snprintf and realloc were used.
In this version they are replaced with asprintf.
Suggested by Dmitry Levin.
Signed-off-by: Masatake YAMATO <yamato@redhat.com>
In current implementation, the stack trace is captured and printed at
the same time, in trace_syscall_exiting. This approach cannot
provide user expected information when a system call changes the
memory mapping. In such cases, the stack trace should be captured on
entering syscall and printed on exiting.
As the initial step for splitting capturing from printing, this change
introduces stacktrace_walker utility function. It can be used both for
capturing in trace_syscall_entering and printing in
trace_syscall_exiting.
Signed-off-by: Masatake YAMATO <yamato@redhat.com>
free_libunwind_ui is expected to release all unwind related resources
attached to tcp.
* strace.c (droptcb): Move delete_mmap_cache call ...
* unwind.c (free_libunwind_ui): ... to here.
Signed-off-by: Masatake YAMATO <yamato@redhat.com>
Print the stack trace of the traced process after each system call when
-k option is specified. It is implemented using libunwind to unwind the
stack and to obtain the function name pointed by the IP.
Based on the code that was originally taken from strace-plus
of Philip J. Guo.
* configure.ac: Add --with-libunwind option. Check libunwind support.
* Makefile.am: Add libunwind support.
* defs.h (struct tcb) [USE_LIBUNWIND]: Append libunwind specific fields.
[USE_LIBUNWIND] (stack_trace_enabled, alloc_mmap_cache,
delete_mmap_cache, print_stacktrace): New prototypes.
* mem.c (print_mmap, sys_munmap, sys_mprotect): Add libunwind support.
* process.c (sys_execve): Likewise.
* strace.c (usage, alloctcb, droptcb, init): Likewise.
* syscall.c (trace_syscall_exiting): Likewise.
* unwind.c: New file.
* strace.1: Document -k option.
* xlat/gen.sh: Define all xlat structs not declared in defs.h as static.
Some symbolic constants are not macros, extend #ifdef check to cover
symbolic constants checked by AC_CHECK_DECLS.
Handle complex symbolic constants in SYMBOL|... form.
Handle symbolic constants in 1<<SYMBOL form.
Handle numeric constants.
Implement #unconditional directive that turns off preprocessor checks.
Implement #unterminated directive that turns off adding XLAT_END.
Now that ./xlat/gen.sh has to be run before autoreconf,
replace all autoreconf calls with ./bootstrap call.
* bootstrap: Forward arguments to autoreconf.
* build_static_example.sh: Replace autoreconf call with bootstrap call.
* make-dist: Likewise.
* qemu_multiarch_testing/README: Likewise.
Time spent in system time is not useful where a syscall depends on some
non-CPU resource, eg. typically open() or stat() to a network drive.
This patch adds a new flag (-w) to produce a summary of the time
difference between beginning and end of the system call (ie. latency)
This functionality has been useful to profile slow processes that
are not CPU-bound.
Signed-off-by: Mark Hills <mark.hills@framestore.com>
* count.c (count_syscall): Add const qualifier to timeval argument and
rename it. Store the wall clock time spent while in syscall in separate
timeval variable.
* defs.h (count_syscall): Update prototype.
* syscall.c (trace_syscall_exiting): Update count_syscall invocation.
The argument passed to sethostname and setdomainname syscalls, as well
as the string returned by gethostname syscall, is not a pathname, so
printpathn is not the right method for its decoding.
* process.c (sys_sethostname, sys_setdomainname): Decode 1st argument
using printstr instead of printpathn.
[ALPHA] (sys_gethostname): Likewise.
When strace is built with large file support definitions in CFLAGS (as
may be provided by buildroot) the C library headers may expose a 64-bit
rlim_t even though the struct rlimit fields used by the system call
interface are only 32-bit. The SIZEOF_RLIM_T will then be 8 which
results in bad decoding of the getrlimit and setrlimit syscalls.
This is fixed by replacing unreliable SIZEOF_RLIM_T based checks with
checks for current_wordsize.
Signed-off-by: James Hogan <james.hogan@imgtec.com>
Signed-off-by: Dmitry V. Levin <ldv@altlinux.org>
I found that I use it quite often. Lets make it so that
after cut-n-pasting it into a file, there is no need
to edit the result (e.g. no need to remove C comment
chars from every line.
Signed-off-by: Denys Vlasenko <dvlasenk@redhat.com>
Without this patch, strace claims that parent_tidptr == tls, which is
clearly wrong. It is expected that parent_tidptr == child_tidptr.
* process.c [X86_64] (ARG_CTID, ARG_TLS): Take current
personality into account.
Signed-off-by: Elliott Hughes <enh@google.com>
Signed-off-by: Dmitry V. Levin <ldv@altlinux.org>
We need to handle this situation more like x86-64. 32-bit arm and i386
actually have a common struct stat64, except the arm one must not be
packed. Additionally, on aarch64 the 32-bit personality is personality 0.
Signed-off-by: Elliott Hughes <enh@google.com>
