-*-org-*-
- Each DSO can now ship an ltrace config file (called prototype library) that ltrace will open when that DSO is loaded to process image. See ltrace(1) for details.
- ltrace.conf is no longer part of installation tarball. Instead, we now ship libc.so.conf, libm.so.conf, libacl.so.conf, and syscalls.conf. Those are now istalled to usr/share/ltrace by default. /etc/ltrace.conf and $HOME.ltrace.conf are still loaded if present, and can contain arbitrary user configuration.
- The option -F was retrofitted to be a colon-separated list of prototype libraries, and directories to look for prototype libraries in. On Linux, ltrace looks into XDG_CONFIG_HOME, XDG_CONFIG_DIRS, and /usr/share/ltrace as well.
- Wide character strings are supported in prototypes. Use “string” lens as usual, but use array of integers as underlying type. libc.so.conf now contains prototypes of wide character functions.
- Sole void function parameter such as in the following example, is
now considered obsolete:
int fork(void); This use is still accepted, taken to mean “hide(int)”, but produces a warning, and will be removed in future.
- MIPS and MIPSel are now handled by the same backend.
- ARMv6, ARMv7 and ARMv8 (AArch64) are supported, including full fetch backend. ARMv8 backend doesn’t support tracing of 32-bit binaries, as currently there’s no 32-bit userspace available for ARM64 processors.
- Imagination Technologies Meta is now supported.
- On Linux, tracing of IFUNC symbols is supported. On i386, x86_64, ppc32 with secure PLT and ppc64, IRELATIVE PLT slots are traced as well.
The output format is similar to glibc’s backtrace_symbols, e.g.: > /bin/ls(_init+0x19be) [0x40398e] > /lib/x86_64-linux-gnu/libc.so.6(__libc_start_main+0xed) [0x7f50cbc3676d] > /bin/ls(_init+0x25fd) [0x4045cd]
This bug caused messages like:
| Making all in mips |
| /bin/sh: line 17: cd: mips: No such file or directory |
This bug caused runtime messages like:
| [0x4000000] — SIGSEGV (Segmentation fault) — |
| [0xffffffff] +++ killed by SIGSEGV +++ |
This bug caused messages like:
| Couldn’t determine base address of linux-vdso32.so.1 |
| Couldn’t load ELF object linux-vdso32.so.1: No such file or directory |
Ltrace was unable to find prototypes of static symbols in shared libraries. It changed the symbol name internally to the form name@libfoo.so. Thus mangled, symbols were not found in config files.
This fixes the ltrace abort:
| ./ltrace: proc.c: 755: breakpoint_for_symbol: Assertion `bp->libsym == ((void *)0)’ failed. |
This bug caused messages like:
| Couldn’t determine base address of linux-vdso32.so.1 |
| Couldn’t load ELF object linux-vdso32.so.1: No such file or directory |
More exactly, if there were typedefs X and XY (in this order), and a reference was made to X, ltrace would match XY instead.
__ctype_get_mb_cur_max, __endmntent, __getmntent_r, __setmntent, a64l, abort, abs, addmntent, getmntent_r, hasmntopt, inet_addr, inet_aton, inet_lnaof, inet_makeaddr, inet_netof, inet_network, inet_ntoa, kill, l64a, labs, memcpy, memmove, setlocale, sigaction, sigaddset, sigandset, sigdelset, sigemptyset, sigfillset, sigisemptyset, sigismember, signal, sigorset, sigpending, sigprocmask, sigsuspend.
Please read ltrace.conf(5) man page, chapter “recursive structures”, to learn about this new feature.
This allows displaying various data types as bit vectors. Please read ltrace.conf(5) to learn more.
For reasons of consistency with “hex”. “octal” is still valid and will be for forseeable future.
Ltrace now understands thread groups, and it stops all threads before manipulating breakpoints. The downside is that performance of multi-threaded processes is rather bad, because handling any event implies stopping the whole job. The upside is that individual threads don’t get random SIGILL’s and SIGSEGV’s and events are not missed.
-e and -x were extended to allow library selectors. See the man page for details.
This version brings more complete support for parameter passing, including passing structures in registers, passing double on i386, and other edge cases that were not handled correctly before. The following architectures now have implementation of fetch backend: i386, x86_64, ppc, ppc64, ia64, s390, s390x, m68k.
The deny_ptrace boolean denies all processes from being able to use ptrace to manipulate other processes. Ltrace now understands that this boolean exists, and recommends turning it off, if it is on and ltrace fails to attach to a process.
See the man page for details of the selector syntax. This changes the way -x and -l behave with respect to tracing libraries opened by dlopen:
- In 0.6.0, only those -x symbols that were unmatched in main binary were used to search through symbol table of libraries opened with dlopen. In 0.7.0, -x and -e are applied uniformly to each mapped binary: the main binary, dependent DSO’s, and any dlopened libraries.
- In 0.6.0, -l argument was a filename to open and inspect. In 0.7.0, -l is a glob expression matched against each mapped binary.
This option was introduced in 0.6.0 with the meaning identical to -L, which should be used instead.
Use –enable-valgrind to turn this on.
Parameter packs describe zero or more of actual arguments, whose type can be determined in runtime. The only parameter pack currently implemented is “format” for decoding printf-style format strings.
In future, it should be relatively straightforward to add more parameter packs for functions like execl, and others.
When used in array length expressions, it means “this array ends at the first element with zero value”. C strings are essentially array(char, zero)*.
Lenses are used similarly to parametrized types, e.g.:
| void func(lens(int)); |
“octal”, which used to be a separate type, is now lens, which can be used to render any underlying type in base 8. Plain “octal” is still valid and means the same thing as “octal(int)”.
Similarly, “hex” lens was introduced to format as base 16. “hide” was introduced to conceal given argument. “bool” lens was added to format objects as either true, or false.
“enum” became lens as well. Because enum already uses parentheses to denote the list of enumeration values, the underlying type is selected by square brackets:
| void func(enum[short](RED,GREEN,BLUE)); |
“string” was also turned to lens. The argument can be either a char*, or pointer to array of char, or array of char. The latter is useful in cases like the following:
| void func_struct_2(struct(string(array(char, 6)))); |
Pre-0.7 ltrace shipped a buggy version of ltrace.conf that contained the following prototype:
| itn acl_valid(addr); |
To support extant versions of ltrace.conf, that use is now considered valid, and “itn” is a synonym of “int”. It is however deprecated, and will produce a warning in future.
Functions that take no arguments shouldn’t pretend to take one parameter of void type. For example the following:
| int fork(void); |
… should be declared like this:
| int fork(); |
To support extant versions of ltrace.conf, that use is now considered valid. It is however deprecated, and will produce a warning in future.
Ltrace needs to know the exact underlying type to decide what the calling convention is. The use of void to mean “hide this argument”, such as the following example, is therefore obsolete:
| void func(void, array(int, arg1)); |
Instead, rewrite the prototype depending on the exact underlying type:
| void func(hide(int), array(int, arg1)); |
| void func(hide(long), array(int, arg1)); |
To support extant versions of ltrace.conf, this use is still accepted, and “void” is taken to mean “hide(int)”. It is however obsolete, produces a warning, and will be removed in future.
Earlier, a process that ltrace detached from would sometimes die of SIGSEGV, SIGTRAP, or SIGILL. These were caused by ltrace detaching from the process just after that process hit a breakpoint. Program counter would thus be left pointing mid-instruction, or signals would be left pending.
In cases where the system call shouldn’t be restarted, these are set to -1. In that case assume that the return is from the topmost syscall. This gets rid of some “unexpected breakpoint” messages on x86_64 and i386.
On ppc, the contents of PLT table change after the first call. Ltrace used to handle this by reinserting the (now overwritten) breakpoint after the function returns. This introduced a window where calls to the same function from the same binary (either a recursive calls, or calls from another thread) weren’t traced. This is fixed as of 0.7.0.
As a side effect, events requested via -e now only hit when a PLT call is made, which is consistent with other architectures.
ltrace might work on older kernels, but no attempt was made to fully support it. Newer kernels don’t support PTRACE_SINGLESTEP, which ltrace depends on. Before singlestepping is implemented in software, ARM is considered unsupported.
Pass –with-libunwind to configure to enable the feature. This requires libunwind.
These symbols are selected by -x.
Copyright (C) 2012-2014 Petr Machata <pmachata@redhat.com> This file is part of ltrace.
ltrace is free software; you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation; either version 2 of the License, or (at your option) any later version.
ltrace is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details.
You should have received a copy of the GNU General Public License along with this program. If not, see http://www.gnu.org/licenses/.