Make SysV IPC decoders on 32-bit architectures 64-bit time_t compatible
When libc moves to 64-bit time_t on 32-bit architectures, IPC
definitions provided by sys/ipc.h, sys/msg.h, sys/sem.h, and sys/shm.h
header files become unsuitable for SysV IPC decoders.
Prepare for this change by using Linux kernel uapi headers instead of
libc headers as the preferred source of SysV IPC definitions.
Fall back to libc headers if Linux kernel uapi headers have known bugs
in their SysV IPC definitions.
* configure.ac (AC_CHECK_SIZEOF): Check for struct msqid64_ds in
<linux/msg.h>.
* m4/mpers.m4 (st_MPERS): Likewise.
* ipc_defs.h: Prioritize HAVE_LINUX_IPC_H over HAVE_SYS_IPC_H.
[HAVE_LINUX_IPC_H && !IN_MPERS]: Undefine HAVE_LINUX_IPC_H if
ARCH_SIZEOF_STRUCT_MSQID64_DS != SIZEOF_STRUCT_MSQID64_DS.
[HAVE_LINUX_IPC_H && MPERS_IS_m32]: Undefine HAVE_LINUX_IPC_H if
ARCH_M32_SIZEOF_STRUCT_MSQID64_DS != M32_SIZEOF_STRUCT_MSQID64_DS.
[HAVE_LINUX_IPC_H && MPERS_IS_mx32]: Undefine HAVE_LINUX_IPC_H if
ARCH_MX32_SIZEOF_STRUCT_MSQID64_DS != MX32_SIZEOF_STRUCT_MSQID64_DS.
* linux/sparc/arch_defs_.h (ARCH_SIZEOF_STRUCT_MSQID64_DS): New macro.
* linux/sparc64/arch_defs_.h (ARCH_M32_SIZEOF_STRUCT_MSQID64_DS):
Likewise.
* linux/x32/arch_defs_.h (ARCH_SIZEOF_STRUCT_MSQID64_DS): Likewise.
* linux/x86_64/arch_defs_.h (ARCH_MX32_SIZEOF_STRUCT_MSQID64_DS):
Likewise.
Reported-by: Khem Raj <raj.khem@gmail.com>
Reported-by: Rich Felker <dalias@aerifal.cx>
Resolves: https://github.com/strace/strace/issues/116
Consistently choose SysV IPC definitions between libc and uapi providers
strace decoders should use SysV IPC definitions provided by either libc
or Linux kernel uapi headers, not a random mix of them.
* configure.ac (AC_CHECK_HEADERS): Remove linux/msg.h, linux/sem.h,
linux/shm.h, sys/msg.h, sys/sem.h, and sys/shm.h.
* ipc_defs.h (MSG_H_PROVIDER, SEM_H_PROVIDER, SHM_H_PROVIDER,
NAME_OF_STRUCT_MSQID_DS, NAME_OF_STRUCT_SHMID_DS,
NAME_OF_STRUCT_IPC_PERM_KEY): New macros.
* ipc_msg.c: Include MSG_H_PROVIDER instead of <sys/msg.h> or
<linux/msg.h>.
* ipc_msgctl.c: Likewise.
(msqid_ds_t): Redefine to struct NAME_OF_STRUCT_MSQID_DS.
(print_msqid_ds): Replace __key with NAME_OF_STRUCT_IPC_PERM_KEY.
* ipc_sem.c: Include SEM_H_PROVIDER instead of <sys/sem.h> or
<linux/sem.h>.
(print_sembuf): Define unconditionally.
(tprint_sembuf_array): Use print_sembuf unconditionally.
* ipc_shm.c: Include SHM_H_PROVIDER instead of <sys/shm.h> or
<linux/shm.h>.
* ipc_shmctl.c: Likewise.
(shmid_ds_t): Redefine to struct NAME_OF_STRUCT_SHMID_DS.
(print_shmid_ds): Replace __key with NAME_OF_STRUCT_IPC_PERM_KEY.
* print_msgbuf.c: Include MSG_H_PROVIDER instead of <sys/msg.h>.
Include "ipc_defs.h" in all decoder files that also include
one of system ipc header files.
* ipc_shm.c: Include "ipc_defs.h".
* print_msgbuf.c: Likewise.
Before this change, the test expected that the syscall always returns
a value of type int, which is not the case when the Linux kernel
is built with CONFIG_MODIFY_LDT_SYSCALL disabled.
* tests/modify_ldt.c (printrc): Do not mangle errno if rc is not in
the suitable range.
Resolves: https://github.com/strace/strace/issues/118
Fix build with glibc-2.31 and Linux kernel headers < 5.3
PTRACE_GET_SYSCALL_INFO support was added in strace since version 4.26
released in December 2018, in Linux kernel headers since version 5.3
released in September 2019, and in glibc since version 2.31 released
in February 2020. strace is expected to support all reasonable
combinations of Linux kernel header versions and glibc versions, in
particular, in cases when PTRACE_GET_SYSCALL_INFO definitions are not
available in system headers or available only in Linux kernel headers
or glibc. Due to the longer release cycle of glibc compared to Linux
kernel, it's more likely to have PTRACE_GET_SYSCALL_INFO definitions
provided by <linux/ptrace.h> rather than <sys/ptrace.h>.
This change fixes build in a less likely environment when
PTRACE_GET_SYSCALL_INFO definitions are provided by <sys/ptrace.h> only,
e.g. with glibc >= 2.31 and Linux kernel headers < 5.3.
* configure.ac (AC_CHECK_TYPES): Check for struct __ptrace_syscall_info
in <sys/ptrace.h>, check for struct ptrace_syscall_info in
<linux/ptrace.h>.
* ptrace.h (struct ptrace_syscall_info): Define only if
!HAVE_STRUCT_PTRACE_SYSCALL_INFO && !HAVE_STRUCT___PTRACE_SYSCALL_INFO.
(struct_ptrace_syscall_info): New typedef.
* get_personality.h: Replace struct ptrace_syscall_info with
struct_ptrace_syscall_info.
* linux/aarch64/arch_get_personality.c: Likewise.
* linux/powerpc64/arch_get_personality.c: Likewise.
* linux/s390x/arch_get_personality.c: Likewise.
* linux/sparc64/arch_get_personality.c: Likewise.
* linux/tile/arch_get_personality.c: Likewise.
* linux/x86_64/arch_get_personality.c: Likewise.
* ptrace_syscall_info.c: Likewise.
* syscall.c: Likewise.
* tests/ptrace_syscall_info.c: Likewise.
* NEWS: Mention this fix.
Reported-by: Vineet Gupta <Vineet.Gupta1@synopsys.com>
Fixes: v4.26~51 "Add definitions for PTRACE_GET_SYSCALL_INFO API"
ci: re-enable testing with the latest mainline kernel headers
There should be no issues with v5.5-rc6.
This reverts commit 1aa46bd14e.
* .gitlab-ci.yml (variables): Remove KBRANCH.
* .travis.yml (env:global): Likewise.
Move the declaration of filter_generators out of HAVE_FORK ifdef
to fix build for no-MMU targets:
filter_seccomp.c: In function ‘check_seccomp_filter_properties’:
filter_seccomp.c:608:42: error: ‘filter_generators’ undeclared (first use in this function); did you mean ‘linear_filter_generator’?
for (unsigned int i = 0; i < ARRAY_SIZE(filter_generators); ++i) {
^~~~~~~~~~~~~~~~~
* filter_seccomp.c (linear_filter_generator,
binary_match_filter_generator): Declare unconditionally.
(filter_generator_t, filter_generators, filters, bpf_prog): Define
unconditionally.
* NEWS: Mention this fix.
Fixes: v5.4~57 "filter_seccomp: list of seccomp filter generation strategies"
clone: implement decoding of new fields in struct clone_args
* configure.ac (AC_CHECK_TYPES): Remove struct clone_args.
(AC_CHECK_MEMBERS): Check for struct clone_args.set_tid_size
in <linux/sched.h>.
* clone.c (struct strace_clone_args): Add set_tid and set_tid_size
fields.
(SYS_FUNC(clone3)): Decode set_tid and set_tid_size fields of struct
clone_args introduced by Linux kernel commit v5.5-rc1~180^2~5.
* NEWS: Mention this change.
* tests/clone3.c: Check HAVE_STRUCT_CLONE_ARGS_SET_TID_SIZE instead of
HAVE_STRUCT_CLONE_ARGS.
(struct test_clone_args): Add set_tid and set_tid_size fields.
(MAX_SET_TID_SIZE): New macro.
(print_set_tid): New function.
(print_clone3): Use it.
(main): Check decoding of set_tid and set_tid_size fields.
clone: add support for clone3-specific CLONE_CLEAR_SIGHAND flag decoding
Introduced by Linux commit v5.5-rc1~180^2~8.
* xlat/clone3_flags.in: New file.
* clone.c: Include "xlat/clone3_flags.h".
(SYS_FUNC(clone3)): Use both clone_flags and clone3_flags for flag field
decoding.
* gcc_compat.h [!GNUC_PREREQ(3, 0)] (FAIL_BUILD_ON_ZERO): Stub
definition.
* tests/clone3.c: Include "xlat/clone3_flags.h".
(ERR): New macro.
(do_clone3_): Accept possible_errors bitmask instead of should_Fail
flag.
(main): Add CLONE_CLEAR_SIGHAND-specific checks, update flag-related
checks.
* tests/clone3.c: Include <stdint.h> and "xlat/clone_flags.h"
unconditionally. Do not include <linux/types.h>.
(struct clone_args): Rename to struct test_clone_args, define
unconditionally.
(struct_clone_args): New typedef.
(do_clone3_, print_clone3, main): Switch all users of
struct test_clone_args to struct_clone_args.
(main): Use TAIL_ALLOC_OBJECT_CONST_PTR. Move printing of "clone3("
prefix ...
(print_clone3): ... here to fix errno clobbering.
s390: workaround Linux kernel bug in syscall_get_nr
Apparently, arch/s390/include/asm/syscall.h:syscall_get_nr() returns 0
for out-of-range syscall numbers.
This kernel bug is exposed via PTRACE_GET_SYSCALL_INFO interface.
Workaround it by falling back to get_regs() for the zero syscall number.
* linux/s390/check_scno.c: New file.
* linux/s390x/check_scno.c: Likewise.
* Makefile.am (EXTRA_DIST): Add them.
* syscall.c (get_scno): Update comment.
tests: skip stack tracing tests when /proc/self/maps is invalid
Apparently, some container orchestration systems are capable of messing
up with /proc/self/maps. For example, one can face with the following
garbage in Travis CI:
$ cat /proc/self/maps
ee72dcc000-ee72dcd000 r-xp 00000000 00:43 16 /var/snap/lxd/common/lxd/storage-pools/instances/containers/travis-job-strace-strace-634261782/rootfs/bin/cat
ee72dcd000-ee72dce000 r--p 00000000 00:43 16 /var/snap/lxd/common/lxd/storage-pools/instances/containers/travis-job-strace-strace-634261782/rootfs/bin/cat
ee72dce000-ee72dcf000 rw-p 00010000 00:43 16 /var/snap/lxd/common/lxd/storage-pools/instances/containers/travis-job-strace-strace-634261782/rootfs/bin/cat
...
Our stack tracing cannot work properly when /proc/self/maps is damaged
this way.
* tests/strace-k.test: Check that the path to executable recorded
in /proc/self/maps is valid and skip the test if it's not the case.
* filter_seccomp.c (traced_by_seccomp): Trace all syscalls that have
MEMORY_MAPPING_CHANGE flag set when stack tracing is enabled.
* NEWS: Mention this fix.
clone: simplify signal printing in legacy clone handler a bit
printsignal() does the right thing for quite a while (since v4.26~66
"Stop using SIG_0") so the signal printing code can be simplified.
This change is covered by a check in clone-flags test.
* clone.c (SYS_FUNC(clone)): Simplify signal printing in absence
of flags.
* xlat/kvm_cap.in (KVM_CAP_PPC_GUEST_DEBUG_SSTEP): New constant
introduced by Linux kernel commit v5.5-rc1~47^2~5^2~14.
(KVM_CAP_ARM_NISV_TO_USER): New constant introduced by Linux kernel
commits v5.5-rc1~47^2~16^2~3 and v5.5-rc1~47^2~16.
(KVM_CAP_ARM_INJECT_EXT_DABT): New constant introduced by Linux kernel
commits v5.5-rc1~47^2~16^2~2 and v5.5-rc1~47^2~16.
* xlat/kvm_exit_reason.in (KVM_EXIT_ARM_NISV): New constant introduced
by Linux kernel commit v5.5-rc1~47^2~16^2~3.
NEWS: mention Fedora bug number for the libdwfl cache issue
Also, mention the reporter in the git commit message.
* NEWS: Mention Fedora bug number for the libdwfl cache issue.
Reported-by: Pavel Zhukov <pzhukov@redhat.com>
Complements: v5.4-18-g69b2c33 "unwind-libdw: fix initialization of libdwfl cache"
This fixes stack trace printing for early syscalls that precede
the first syscall from memory mapping or execve families.
* unwind-libdw.c (tcb_init): Set struct ctx.last_proc_updating
to a value different from mapping_generation so that libdwfl cache
is properly initialized before the first use.
* NEWS: Mention this fix.
Resolves: https://bugzilla.redhat.com/show_bug.cgi?id=1788636
tests: Do not dereference SCM_TIMESTAMP* control messages in-place
* tests/sockopt-timestamp.c (print_timestamp_old,
print_timestampns_old): Do not dereference SCM_TIMESTAMP* control
messages in-place to avoid potential alignment issues in the future.
* tests/msg_control.c (test_scm_timestamp_old, test_scm_timestampns_old,
test_scm_timestamping_old, test_scm_timestamp_new,
test_scm_timestampns_new, test_scm_timestamping_new): Likewise.
(main): Use "tail_alloc(1024) + 1024" instead of "tail_alloc(1) + 1"
for obtaining a tail_alloc'ed page to avoid false -Warray-bounds
warnings from some versions of gcc.
* print_timespec.h (PRINT_TIMESPEC_ARRAY_DATA_SIZE): Use loop initial
declaration. Do not treat potentially unaligned pointers as pointers
to TIMESPEC_T.
Fix potentially unaligned access in SO_TIMESTAMP*_NEW decoder and tests
* print_timespec.h [PRINT_TIMESPEC_DATA_SIZE ||
PRINT_TIMESPEC_ARRAY_DATA_SIZE] (print_unaligned_timespec_t): New
function.
(PRINT_TIMESPEC_DATA_SIZE, PRINT_TIMESPEC_ARRAY_DATA_SIZE): Use it
instead of print_timespec_t.
* tests/sockopt-timestamp.c: Include <string.h>.
(print_timestamp_new, print_timestampns_new): Fix potentially unaligned
access of CMSG_DATA.
* NEWS: Mention this fix.
Fixes: v5.1~21 "Implement decoding of SO_TIMESTAMP*_NEW control messages"
tests: increase the child stack size in some tests that invoke clone
Increase the child stack size from the half page size to the full page
size in those few tests that invoke clone syscall.
The former size appears to be insufficient in some build setups.
For example, clang 7 on Ubuntu bionic x86_64 makes ioctl_nsfs test fail
due to child process segfaults.
* tests/clone_parent.c (child_stack_size, clone): Remove.
(do_clone): New macro.
(main): Use it.
* tests/clone_ptrace.c: Likewise.
* tests/ioctl_nsfs.c (child_stack_size, clone): Remove.
(do_clone): New macro.
(test_user_namespace): Use it.
tests: insert a delay into bpf-obj_get_info_by_fd tests
According to Quentin Monnet [1], there is a delay when the locked memory
is being reclaimed after a BPF program or map is removed.
Privileged tools like iproute2 and bpftool workaround this Linux kernel
limitation by raising RLIMIT_MEMLOCK to infinity prior to creating BPF
objects.
Our BPF tests are expected to be invoked without extra privileges
and therefore do not have this option.
Instead, we serialize all BPF tests invocations and insert a delay
long enough to let the locked memory be reclaimed.
[1] https://lists.strace.io/pipermail/strace-devel/2019-December/009274.html
* tests/bpf-obj_get_info_by_fd.c (main): Add a comment and a sleep
invocation.
* tests/pure_executables.list: Remove bpf-obj_get_info_by_fd,
bpf-obj_get_info_by_fd-prog, bpf-obj_get_info_by_fd-prog-v,
and bpf-obj_get_info_by_fd-v because they call sleep function
that may be implemented on top of clock_nanosleep* syscalls.
Thanks-to: Quentin Monnet <quentin.monnet@netronome.com>
References: https://github.com/strace/strace/issues/74
* xlat/bpf_attach_type.in (BPF_TRACE_RAW_TP): New constant introduced in
Linux commit v5.5-rc1~174^2~249^2~4^2~1.
(BPF_TRACE_FENTRY, BPF_TRACE_FEXIT): New constants introduced in Linux
commit v5.5-rc1~174^2~49^2~12^2~16.
* xlat/bpf_map_flags.in (BPF_F_MMAPABLE): New constant introduced in Linux
commit v5.5-rc1~174^2~49^2~11^2~2.
* xlat/bpf_prog_types.in (BPF_PROG_TYPE_TRACING): New constant
introduced in Linux commit v5.5-rc1~174^2~249^2~4^2~1.
* NEWS: Mention this.
* tests/bpf.c (BPF_MAP_CREATE_checks, BPF_PROG_LOAD_checks,
BPF_PROG_QUERY_checks): Update expected outputs.
* tests/kernel_version.c (print_bpf_attr, main): Likewise.
Signed-off-by: Paul Chaignon <paul.chaignon@gmail.com>
Signed-off-by: Jamal Hadi Salim <jhs@mojatatu.com>
Before this fix NDA_LLADDR attributes used to be printed
as inet addresses, not as mac addresses.
* rtnl_neigh.c (decode_neigh_lladdr): New function.
(ndmsg_nla_decoders): Use it for NDA_LLADDR decoding.
* NEWS: Mention this fix.
Co-authored-by: Dmitry V. Levin <ldv@altlinux.org>
Fixes: v4.19~33 "rtnl_neigh: decode ndmsg netlink attributes"
configure.ac: do not copy CFLAGS to CFLAGS_FOR_{M32,MX32}
As result, CFLAGS is no longer passed to mpers.sh,
reintroducing the traditional behaviour changed by commit v5.4~32.
The copying behaviour in it's current form was problematic because some
of compiler options may break mpers.sh; for example, if CFLAGS includes
-flto, then mpers.sh breaks, since intermediate *.o files created by
mpers.sh do not contain any real code.
* configure.ac [$arch != aarch64]: Initialize CFLAGS_FOR_M32 and
CFLAGS_FOR_MX32 to empty strings by default.
* NEWS: Mention this fix.
Reported-by: Andreas Schwab <schwab@linux-m68k.org>
Reported-by: Thomas Deutschmann <whissi@gentoo.org>
Resolves: https://github.com/strace/strace/issues/114
Fixes: v5.4~32 "Add support for per-mpers compiler settings"
tests: tweak restart_syscall test to workaround nanosleep changes
nanosleep function in glibc used to be implemented as a simple syscall
wrapper around __NR_nanosleep, but starting with glibc commit
glibc-2.30.9000-234-g3537ecb49c, it is implemented on top of
__clock_nanosleep function which in turn is implemented on top of
__clock_nanosleep_time64 function.
These functions do not forward the last argument written by the kernel
when the underlying syscall succeeds, breaking the expected output.
Besides that, these proxy functions pass garbage in the upper 32-bits
of tv_nsec on 32-bit architectures .
Rather than changing restart_syscall test to use __NR_nanosleep syscall
directly, relax the expected output with regards to the remaining time
written by the kernel as this not what is being tested by
restart_syscall test.
* tests/restart_syscall.c: Stop including <assert.h>.
(main): Do not use assert. Change the initial value of rem variable.
Use regexps instead of specific values of rem.tv_sec and rem.tv_nsec
in the expected output.
* NEWS: Mention this change.
Indent the C preprocessor directives to reflect their nesting
using the following script:
$ cppi -l $(git grep -El '^[[:space:]]*#[[:space:]]*(if|ifdef|ifndef|elif|else|endif|define|pragma)[[:space:]]' |grep -v '\.sh$') |while read f; do
cppi < "$f" > "$f".cppi; mv "$f".cppi "$f"
done
tests: tweak restart_syscall test to break in a more obvious way in case of faulty libc
If libc implements nanosleep on top of clock_nanosleep, there is
a chance of struct timespec translation bugs in libc breaking this test.
Tweak the test to make this breakage more obvious.
* tests/restart_syscall.c (main): Initialize rem variable with invalid
timespec data.
Dmitry V. Levin
Eugene Syromyatnikov
Baruch Siach
Andrey Zax
Vladimir D. Seleznev
Paul Chaignon
Jamal Hadi Salim