Dmitry V. Levin [Wed, 24 Aug 2016 00:29:47 +0000 (00:29 +0000)]
Revert "tests/xstatx.c: fix stat syscall tests on mips64"
Previous commit effectively changed types of st_atime, st_ctime, and
st_mtime members of struct stat and struct stat64 to signed integers,
making the mips64 workaround obsolete.
Dmitry V. Levin [Wed, 24 Aug 2016 00:29:40 +0000 (00:29 +0000)]
Mpersify parsers of struct stat and struct stat64
On many architectures that support multiple personalities,
struct stat differ between personalities. While old code could handle
these differences, there are some architectures, e.g. sparc64, that also
have different struct stat64. Rewrite parsers using mpers functionality
to fix these issues.
Dmitry V. Levin [Tue, 23 Aug 2016 23:39:45 +0000 (23:39 +0000)]
Move redefinition of stat types to asm_stat.h
* file.c: Move redefinition of types that might be used
to define struct stat ...
* linux/asm_stat.h: ... here.
* tests/xstatx.c: Remove redefinition of stat types.
Dmitry V. Levin [Tue, 23 Aug 2016 14:49:20 +0000 (14:49 +0000)]
x86_64: provide a replacement of <asm/stat.h> for x32 personality
For the reason described in commit v4.10-517-gcfde1e3, a correct
definition of struct stat for x32 personality is necessary to enable
"asm_stat.h" with -mx32 on x86_64.
* linux/x32/asm_stat.h: Rename to ...
* linux/x86_64/asm_stat.h: ... new file.
* Makefile.am (strace_SOURCES): Add it.
* linux/x32/asm_stat.h: New file, include "x86_64/asm_stat.h".
Dmitry V. Levin [Mon, 22 Aug 2016 09:34:54 +0000 (09:34 +0000)]
sparc64: fix decoding of struct stat64 related syscalls
For some reason, struct stat and struct stat64 are different on sparc64.
This change fixes decoding of struct stat64 related syscalls for sparc64
personality, sparc32 personality on sparc64 needs more work.
* file.c (printstat64) [SPARC64]: Do not use printstat.
(SYS_FUNC(newfstatat)): Likewise.
tests: fill old_value argument in timer{,fd}_xettime tests
This is needed in order to differentiate it from the value returned
by the call. As a consequence, it enables revealing possible bugs
in syscall parsers, for example, when the value read on syscall
entering and not on syscall exiting, as it was the case with
timerfd_settime parser.
* tests/timer_xettime.c (main): Fill old.its field with value different
from the expected one upon call return.
* tests/timerfd_xettime.c: Likewise.
Fix old_value argument retrieval in timerfd_settime parser
This is done similar to timer_settime syscall parser.
* time.c (SYS_FUNC(timerfd_settime)): Retrieve old_value argument
on exiting and not on entering. Return 0 instead of RVAL_DECODED
since the call hasn't been decoded in full on entering.
Move SH-specific argument number calculation to getllval
This change prevents scattering of ll-related hacks and simplifies
pread/pwrite syscalls parsers' logic a bit.
* util.c (getllval): Add fixup for arg_no for SuperH when argument
number is equal to 3.
* io.c (PREAD_OFFSET_ARG): Remove.
(SYS_FUNC(pread)): Always use argument number 3 for "count" argument
printing.
(SYS_FUNC(pwrite)): Likewise.
Dmitry V. Levin [Sun, 21 Aug 2016 22:02:09 +0000 (22:02 +0000)]
sparc64: fix sigreturn decoding for sparc personality
* linux/sparc/arch_sigreturn.c (arch_sigreturn): Parametrize
member types of struct signal_frame.
* linux/sparc64/arch_sigreturn.c (sparc64_arch_sigreturn,
sparc32_arch_sigreturn): New functions.
(arch_sigreturn): Use them.
Dmitry V. Levin [Mon, 15 Aug 2016 18:33:47 +0000 (18:33 +0000)]
Drop support of dummy members of struct stat
As st_flags, st_fstype, and st_gen members of struct stat are not filled
by the kernel, there is no use supporting them.
* configure.ac (AC_CHECK_MEMBERS): Remove struct stat.st_flags,
struct stat.st_fstype, and struct stat.st_gen.
* file.c [STAT32_PERSONALITY, HAVE_STRUCT_STAT64]: Do not undefine
HAVE_STRUCT_STAT_ST_FLAGS, HAVE_STRUCT_STAT_ST_FSTYPE,
and HAVE_STRUCT_STAT_ST_GEN.
* printstat.h (DO_PRINTSTAT): Do not check for
HAVE_STRUCT_STAT_ST_FLAGS, HAVE_STRUCT_STAT_ST_FSTYPE,
and HAVE_STRUCT_STAT_ST_GEN.
Dmitry V. Levin [Mon, 15 Aug 2016 17:33:25 +0000 (17:33 +0000)]
Assume that struct stat contains st_blksize, st_blocks, and st_rdev
Our test suite already assumes that struct stat contains st_blksize,
st_blocks, and st_rdev members, and there haven't been any complaints.
* configure.ac (AC_CHECK_MEMBERS): Remove struct stat.st_blksize,
struct stat.st_blocks, and struct stat.st_rdev.
* printstat.h (DO_PRINTSTAT): Do not check for
HAVE_STRUCT_STAT_ST_BLKSIZE, HAVE_STRUCT_STAT_ST_BLOCKS,
and HAVE_STRUCT_STAT_ST_RDEV.
Dmitry V. Levin [Sun, 14 Aug 2016 00:02:55 +0000 (00:02 +0000)]
s390x, x32: remove redundant definitions of ARCH_PC_REG
* linux/s390x/arch_regs.c (ARCH_PC_REG): Remove, it is already defined
in just included linux/s390/arch_regs.c.
* linux/x32/arch_regs.c (ARCH_PC_REG): Remove, it is already defined
in just included linux/x86_64/arch_regs.c.
Dmitry V. Levin [Sat, 13 Aug 2016 22:05:20 +0000 (22:05 +0000)]
tests/init.sh: print exit code of failed commands
* init.sh (run_prog, run_prog_skip_if_failed): When the program fails,
add its exit code to the diagnostic message.
(run_strace): When strace fails, add its exit code to the diagnostic
message.
(run_strace_merge): When strace-log-merge fails, add its exit code
to the diagnostic message.
Dmitry V. Levin [Sat, 13 Aug 2016 22:04:59 +0000 (22:04 +0000)]
Fix leakage of placeholder descriptors to tracees
As a side effect of commit v4.11-211-g0736d4e, strace used to leak
placeholders for standard descriptors to tracees thus affecting their
behaviour. Fix this by setting close-on-exec flag on placeholder
descriptors.
* strace.c (open_dummy_desc): Set close-on-exec flag on the descriptor
that is going to be returned to the caller.
(fd_is_placeholder): New array.
(ensure_standard_fds_opened, redirect_standard_fds): New functions.
(startup_child): Use redirect_standard_fds.
(init): Use ensure_standard_fds_opened.
Dmitry V. Levin [Thu, 11 Aug 2016 22:31:08 +0000 (22:31 +0000)]
tests: fix *stat64 tests on alpha
On some architectures including alpha, <asm/stat.h> provides a
definition of struct stat that has no st_atime_nsec, st_mtime_nsec, and
st_ctime_nsec fields. At the same time, struct stat64 always has these
fields. Fix tests to take this difference into account.
* tests/fstat64.c (STRUCT_STAT_IS_STAT64): New macro, defined to 1.
* tests/lstat64.c (STRUCT_STAT_IS_STAT64): Likewise.
* tests/stat64.c (STRUCT_STAT_IS_STAT64): Likewise.
* tests/xstatx.c [!STRUCT_STAT] (STRUCT_STAT_IS_STAT64): New macro,
defined to 0.
[USE_ASM_STAT && STRUCT_STAT_IS_STAT64]:
(HAVE_STRUCT_STAT_ST_ATIME_NSEC, HAVE_STRUCT_STAT_ST_CTIME_NSEC,
HAVE_STRUCT_STAT_ST_MTIME_NSEC): Redefine to 1.
James Cowgill [Thu, 11 Aug 2016 16:33:02 +0000 (16:33 +0000)]
tests/nsyscalls.test: only trace the "syscall" syscall on mips o32
The "syscall" syscall only exists on o32 and causes strace to error out on
64-bit mips ABIs. Pass MIPS_ABI from the configure script through to
nsyscalls.test so the MIPS ABI can be checked.
* configure.ac (MIPS_ABI): Substitute into output files.
* tests/Makefile.am (MIPS_ABI): Export via AM_TEST_LOG_FLAGS.
* tests/nsyscalls.test: Restrict special mips handling to mips o32.
James Cowgill [Thu, 11 Aug 2016 16:33:01 +0000 (16:33 +0000)]
tests/xstatx.c: fix stat syscall tests on mips64
For historical reasons the kernel struct stat represents times as unsigned
32-bit integers on mips64. Therefore, while it's possible to give a file a
timestamp before 1970 with futimens, reading the same timestamp through
struct stat will give a positive time (around 2106).
Workaround by using positive timestamps for testing on mips64.
* tests/xstatx.c (create_sample): Use positive timestamps on mips64.
Dmitry V. Levin [Tue, 9 Aug 2016 00:07:53 +0000 (00:07 +0000)]
Fix decoding of invalid syscalls mapped to indirect subcalls
When the syscall number returned by arch_get_scno is a mapped indirect
subcall (i.e. mapped subcall of socketcall or ipc syscall), do not
mistakenly treat it as a valid indirect subcall.
* defs.h (SCNO_IS_VALID): Treat scno with TRACE_INDIRECT_SUBCALL flag
as invalid.
* syscall.c (syscall_name): Do no shuffle scno.
(trace_syscall_entering, trace_syscall_exiting): Use
tcp->s_ent->sys_name instead of syscall_name.
(get_scno): In case of invalid syscall, allocate a dynamic struct sysent
containing an appropriate .sys_name.
* tests/nsyscalls.c (main) [SYS_socket_subcall]: Check decoding
of direct syscall number SYS_socket_subcall+1.
(main) [SYS_ipc_subcall]: Check decoding of direct syscall number
SYS_ipc_subcall+1.
Dmitry V. Levin [Mon, 8 Aug 2016 21:29:58 +0000 (21:29 +0000)]
sparc64: swap personality numbers
Fix inconsistency between syscall and ioctl entries on sparc64.
Make layout of personalities on sparc64 the same as on other
architectures that support two personalities.
* defs.h [SPARC64] (PERSONALITY0_WORDSIZE): Change to 8.
[SPARC64] (PERSONALITY1_WORDSIZE): Change to 4.
* linux/sparc64/get_scno.c (arch_get_scno): Swap personality numbers.
* linux/sparc64/ioctls_arch1.h: Rename to ioctls_arch0.h.
* linux/sparc64/ioctls_arch0.h: Rename to ioctls_arch1.h.
* linux/sparc64/ioctls_inc0.h: Rename to ioctls_inc1.h.
* linux/sparc64/ioctls_inc1.h: Rename to ioctls_inc0.h.
Dmitry V. Levin [Mon, 8 Aug 2016 21:11:47 +0000 (21:11 +0000)]
defs.h: cleanup personality specific macro definitions
Move the code that defines PERSONALITY[12]_INCLUDE_FUNCS,
PERSONALITY[12]_INCLUDE_PRINTERS_DECLS,
PERSONALITY[12]_INCLUDE_PRINTERS_DEFS, and MPERS_{m,mx}32_IOCTL_MACROS
macros outside arch specific ifdefs.
This commit is an attempt to unify usage of include guards (in top-level
headers, at least). As a side note, different files with *.h extension
have different semantics: for example, printargs.h is included multiple
times in order to generate slightly varying code depending on values of
macro definitions - maybe it's better to change extension of such files
to something like *.inc.
* defs.h: Add #include guard.
* flock.h: Likewise.
* ipc_defs.h: Likewise.
* mpers_type.h: Likewise.
* printsiginfo.h: Likewise.
* ptrace.h: Likewise.
* regs.h: Likewise.
* seccomp_fprog.h: Likewise.
* gcc_compat.h: Rename the macro used for #include guard.
* msghdr.h: Likewise.
* sigevent.h: Likewise.
* kernel_types.h: Comment the trailing part of #include guard.
* xlat.h: Add missing macro definition for #include guard.
Dmitry V. Levin [Thu, 4 Aug 2016 00:40:33 +0000 (00:40 +0000)]
travis: use a suitable musl revision
There seems to be no readily available binary packages of musl that are
sufficiently up to date to be used to build and test strace, so prepare
a suitable musl from source.
* travis-install.sh (musl-gcc): Do not add an extra repository,
do not install musl-tools and linux-musl-dev packages.
Download, build, and install a suitable musl revision instead.
Dmitry V. Levin [Wed, 3 Aug 2016 14:05:39 +0000 (14:05 +0000)]
Fix printing of mode_t, umode_t, and umask types
Print numeric umode_t type using %#03ho format.
Print return value of umask syscall using %#03lo format.
When printing symbolic mode_t type, always print lower 9 bits,
and print the numeric part using %#03o format.
* defs.h (sprintmode): Remove.
(print_symbolic_mode_t, print_numeric_umode_t,
print_numeric_long_umask): New prototypes.
* printmode.c (sprintmode): Remove.
(print_symbolic_mode_t, print_numeric_umode_t,
print_numeric_long_umask): New functions.
* chmod.c (decode_chmod): Use print_numeric_umode_t.
* ipc_msg.c (SYS_FUNC(msgget)): Likewise.
* ipc_msgctl.c (print_msqid_ds): Likewise.
* ipc_sem.c (SYS_FUNC(semget)): Likewise.
* ipc_shm.c (SYS_FUNC(shmget)): Likewise.
* ipc_shmctl.c (print_shmid_ds): Likewise.
* mq.c (SYS_FUNC(mq_open)): Likewise.
* open.c (decode_open, SYS_FUNC(creat)): Likewise.
* umask.c (SYS_FUNC(umask)): Likewise.
* mknod.c (decode_mknod): Use print_symbolic_mode_t.
* printstat.h (DO_PRINTSTAT): Likewise.
* syscall.c (trace_syscall_exiting): Use print_numeric_long_umask.
* tests/umode_t.c: New file.
* tests/Makefile.am (EXTRA_DIST): Add it.
* tests/creat.c: Rewrite as a thin wrapper around umode_t.c
* tests/mkdir.c: Likewise.
* tests/mkdirat.c: Likewise.
* tests/mknod.c: Extend test coverage of mknod syscall.
* tests/mknodat.c: Extend test coverage of mknodat syscall.
* tests/umask.c: Extend test coverage of umask syscall.
* tests/creat.test: Update the value specified for strace -a parameter.
* tests/mkdir.test: Likewise.
* tests/mkdirat.test: Likewise.
* tests/mknodat.test: Likewise.
Dmitry V. Levin [Tue, 2 Aug 2016 16:47:34 +0000 (16:47 +0000)]
tests: simplify attach-f-p.test
* tests/attach-f-p.c Simplify communications between threads
by replacing per-thread timers with pipes.
* tests/attach-f-p.test: Do not ignore signals.
signalfd4.test: rewrite from match_grep to match_diff
* tests/signalfd4.c (get_sigset_size): New function.
(main): Use it. Print expected output.
* tests/signalfd4.test: Use run_strace_match_diff.
* tests/signalfd4.expected: Remove.
* tests/Makefile.am (EXTRA_DIST): Remove it.
Report the name of ptrace command when ptrace_attach_or_seize fails.
* strace.c (ptrace_attach_cmd): New variable.
(ptrace_attach_or_seize): Save last ptrace command's name
to ptrace_attach_cmd.
(attach_tcb, startup_child): Use it in error diagnostics.
This script has been created by commit v4.8-180-g0ed617b
for the one-shot job of automatic convertion of xlat structures
from *.c files to xlat/*.in files.
Consistently use extended regular expressions where appropriate
When grep or sed is used with basic regular expressions containing
'(', ')', '{', '}', '|', '?', and '+' special characters, convert them
to extended regular expressions for better portability and readability.
* xlat/ipccalls.in: New file.
* ipc.c: New file.
* Makefile.am (libstrace_a_SOURCES): Add it.
* linux/dummy.h (sys_ipc): Remove stub alias.
* syscall.c (decode_ipc_subcall): Treat 1st argument of ipc syscall
as "unsigned int".
[S390 || S390X]: Skip ipc cubcalls that have non-zero version.
[SPARC64]: Likewise, for the native personality.
Save ipc cubcall version for later use by specific ipc parsers.
* ipc_msg.c (SYS_FUNC(msgrcv)): Handle non-zero ipc subcall version.
[SPARC64]: Handle non-ipc_kludge case for the native personality.
* linux/subcall.h (msgrcv): Change nargs from 4 to 5.
* linux/s390/syscallent.h (ipc): Change nargs from 6 to 5.
* linux/s390x/syscallent.h (ipc): Likewise.
* tests/netlink_protocol.c: Check that HAVE_SYS_XATTR_H is defined.
Include <stdlib.h> and <sys/xattr.h>.
(main): Check that system.sockprotoname attribute is supported.
x32, mips n32: make printargs print full 64-bit syscall arguments
As most of x32 syscalls are regular x86_64 syscalls, they should be
printed by default as 64-bit syscalls, and syscall parsers at their
discretion might print syscalls differently. This is already
implemented in 64-bit strace tracing x32 personality processes,
and now x32 strace will also behave this way.
Similar reasoning applies to mips n32.
* util.c (printargs) [HAVE_STRUCT_TCB_EXT_ARG]: Print syscall arguments
from ext_arg array using %llx format.