* .travis.yml (matrix): Switch all gcc-8 builds except arm64 ones
to gcc-9. gcc-8 arm64 builds have to remain until gcc-9 arm64 builds
are fixed, currently they error out because of missing
gcc-9-multilib-arm-linux-gnueabihf package.
travis: switch some stack trace enabled builds from libunwind to libdw
Switch to Bionic allows to use libdw instead of libunwind.
* .travis.yml (matrix): Change STACKTRACE from libunwind to libdw
in most of Bionic-based builds.
Ideally, all builds have to be switched, but coverage build,
unfortunately, errors out by Travis CI with a build timeout,
musl-gcc x86 build fails with
"undefined reference to `__stack_chk_fail_local'" error, and gcc-8 arm64
builds error out because of missing gcc-8-multilib-arm-linux-gnueabihf
package.
* .travis.yml (dist): Change to bionic.
(matrix): Add "dist: trusty" to coverage, musl-gcc x86, and gcc-8 arm64
builds.
Travis CI now supports aarch64 builds, launched with the new "arch: arm64"
tag. These builds are running inside LXD only for the moment.
The aarch64 builds tend to be a lot slower than their x86_64 counterparts,
so several timeout values need to be increased.
Some tests occasionally take more than 10 minutes, hitting Travis CI's
timeout on output inactivity. Using "travis_wait 40" disables that
timeout; it delays the start time for the output inactivity timeout by 40
minutes, bringing it to the maximum build timeout of 50 minutes.
* .travis.yml (matrix): Add aarch64 builds and mark previous builds as
amd64.
* ci/install-dependencies.sh: Install packages for aarch64 builds.
* ci/run-build-and-tests.sh: Check mpers for aarch64.
Signed-off-by: Paul Chaignon <paul.chaignon@gmail.com>
debian/control: add build dependencies for aarch64 mpers
* debian/control (Build-Depends): Add gcc-arm-linux-gnueabi,
libc6-dev-armel-cross, linux-libc-dev-armel-cross as additional
dependencies on arm64 in order to build proper mpers support.
configure.ac: use a separate compiler for ARM EABI on AArch64
As gcc on AArch64 doesn't support compat ABI target, we have to find
and use a separate compiler in order to get proper compat personality
support on AArch64.
* configure.ac (arch == aarch64): Try to find a compiler for ARM EABI
based on known names for it in various distributions.
As some very special architectures require a separate compiler
in order to compile binaries for additional supported personalities,
we have to introduce such support.
* Makefile.am (mpers_CPPFLAGS): Move $(strace_CPPFLAGS) ...
(mpers_sh_opts): ... here.
(mpers-m%.stamp:): Use $(mpers_CC) instead of $(CC) for compiler,
pass $(mpers_CC_FLAGS) in quotes as it might be empty or contain spaces.
($(mpers_m32_targets)): Add @CFLAGS_FOR_M32@ to mpers_CC_FLAGS, set
mpers_CPPFLAGS and mpers_CC.
($(mpers_mx32_targets)): Add @CFLAGS_FOR_MX32@ to mpers_CC_FLAGS, set
mpers_CPPFLAGS and mpers_CC.
* bootstrap (m_upper): Derive uppercase personality name.
Set CC for mpers Makefile.am, add @CFLAGS_FOR_{M32,MX32}@ to
MPERS_CC_FLAGS definition.
* configure.ac: Derive default values for
{CC,CPP,CFLAGS,CPPFLAGS}_FOR_{M32,MX32} from the respective CC, CPP,
CFLAGS, and CPPFLAGS values.
* m4/mpers.m4 (AC_DEFUN([st_MPERS])): Use
{CC,CPP,CFLAGS,CPPFLAGS}_FOR_<$1> as CC, CPP, CFLAGS, and CPPFLAGS,
respectively. Enclose MPERS_CFLAGS in quotes in mpers.sh call. Provide
information about compiler in the AC_CACHE_CHECK message.
* tests/Makefile.am (CC): New variable.
Co-Authored-by: Dmitry V. Levin <ldv@altlinux.org>
tests: fix invalid copyright years range in a few tests
* maint/update_copyright_years.sh (process_file): When updating
copyright years, if there is existing notice and its starting year is
earlier than the year of the first commit, the former takes precedence.
* tests/filter_seccomp-flag.c: Fix copyright years range.
* tests/filter_seccomp-perf.c: Likewise.
* tests/filter_seccomp-perf.test: Likewise.
* tests/filter_seccomp.sh: Likewise.
* tests/status-none-f.c: Likewise.
* xlat/ptp_flags_options.in: Rename to ...
* xlat/ptp_extts_flags.in: ... new file, add fallback definitions
for constants.
* xlat/ptp_perout_flags.in: New file.
* ptp.c: Include "xlat/ptp_extts_flags.h" and "xlat/ptp_perout_flags.h"
instead of "xlat/ptp_flags_options.h".
(ptp_ioctl): Use ptp_extts_flags and ptp_perout_flags instead of
ptp_flags_options.
* tests/ioctl_ptp.c: Include "xlat/ptp_extts_flags.h" and
"xlat/ptp_perout_flags.h" instead of "xlat/ptp_flags_options.h".
(test_no_device): Use ptp_extts_flags and ptp_perout_flags instead of
ptp_flags_options.
* xlat/kvm_cap.in (KVM_CAP_ARM_IRQ_LINE_LAYOUT_2): New constant
introduced by Linux kernel commit v5.4-rc1~138^2~20^2.
(KVM_CAP_HYPERV_DIRECT_TLBFLUSH): New constant introduced by Linux
kernel commit v5.4-rc1~26^2~53.
* NEWS: Mention this.
* xlat/madvise_cmds.in (MADV_COLD): New constant introduced by Linux
kernel commit v5.4-rc1~46^2~13.
(MADV_PAGEOUT): New constant introduced by Linux kernel commit
v5.4-rc1~46^2~11.
* NEWS: Mention this.
filter_seccomp: fix jump offset overflow in binary match generator
binary_match_filter_generator() is missing a check for jump offset
overflows which might result in incorrect behavior if the binary match
strategy is selected and overflows.
I have only been able to reproduce the bug on mips after forcing strace to
use the binary match generator. Due to the large number of syscalls on
mips, the binary match algorithm is suboptimal and the linear one is
selected. This bug could however be triggered inadvertently if tracing a
very large set of syscalls not grouped together; in that case, the linear
strategy might have a jump offset overflow itself and strace would
fallback to the binary match one.
* filter_seccomp (binary_match_filter_generator): Check for jump offset
overflows.
Signed-off-by: Paul Chaignon <paul.chaignon@gmail.com>
tests: tolerate zero process group ID and session ID in strace-DDD.test
Paul Chaignon reported [1] that in some of Travis CI build environments
process group ID and session ID may be equal to zero.
Change the affected strace-DDD.test to tolerate these odd values as long
as they don't interfere with the purpose of the test.
[1] https://travis-ci.community/t/unexpected-zero-values-in-proc-pid-stat/5774
* tests/strace-DDD.test: Tolerate zero process group ID and session ID.
Reported-by: Paul Chaignon <paul.chaignon@gmail.com>
tests: robustify umount2.test against filtered umount2 syscall
Do not assume that umount2 syscall invoked with invalid arguments always
fail with EINVAL as in some environments it can be filtered out e.g.
using seccomp-bpf.
* tests/umount2.c (main): Use sprintrc() to prepare expected output.
Signed-off-by: Paul Chaignon <paul.chaignon@gmail.com>
Co-authored-by: Dmitry V. Levin <ldv@altlinux.org>
tests: enhance termination signal forwarding check
Check that termination signal forwarding works properly when strace
is invoked with the corresponding termination signal blocked.
* tests/block_reset_run.c: New file.
* tests/Makefile.am (check_PROGRAMS): Add block_reset_run.
* tests/.gitignore: Likewise.
* tests/termsig.test: Extend the check to cover the case when
strace is invoked with the termination signal blocked.
Use __gcov_dump declaration from <gcov.h> if available
The function declaration provided by <gcov.h> is fine,
use it when <gcov.h> is available.
* configure.ac (AC_CHECK_HEADERS): Add gcov.h.
* defs.h [ENABLE_COVERAGE_GCOV && HAVE_GCOV_H]: Include <gcov.h>.
Use __gcov_dump instead of __gcov_flush to dump profile information
__gcov_flush is essentially __gcov_dump followed by __gcov_reset.
The way we use __gcov_flush so far does not need any __gcov_reset,
so use __gcov_dump instead.
* defs.h [ENABLE_COVERAGE_GCOV] (__gcov_flush): Replace with
__gcov_dump.
(GCOV_FLUSH): Replace with GCOV_DUMP.
* filter_seccomp.c (check_seccomp_order_do_child): Replace GCOV_FLUSH
with GCOV_DUMP.
* strace.c (terminate): Likewise.
Introduce a new macro GCOV_FLUSH and use it instead of __gcov_flush.
* defs.h (GCOV_FLUSH): New macro.
[ENABLE_COVERAGE_GCOV] (__gcov_flush): New prototype.
* filter_seccomp.c (__gcov_flush): Remove.
(check_seccomp_order_do_child): Call GCOV_FLUSH instead of __gcov_flush.
* strace.c (__gcov_flush): Remove.
(terminate): Call GCOV_FLUSH instead of __gcov_flush.
This commit introduces a new BPF program generation strategy. Traced
syscalls are encoded in 32-bit bit arrays in the BPF program. Syscalls
are then matched against bit arrays at runtime with two ALU operations:
a division to select the appropriate bit array to compare with,
and a shift to select the appropriate bit in the bit array.
Since there is no way to implement a jump table in BPF (jumps have fixed
offsets), we have to iterate over all bit arrays to select the appropriate
bit array. The division and modulo are also converted into a shift and
a bitwise AND, to improve performance and because seccomp-bpf disallows
modulos in BPF.
Compared to the linear generation strategy, this strategy generates
programs of near constant size. There is a single optimization that
depends on traced syscalls: if a bit array is all-0 or all-1, we don't
need to do a jset against it, we can simply jump to either RET_ALLOW
or RET_TRACE.
* filter_seccomp.c (JMP_PLACEHOLDER_ALLOW): New constant.
(binary_match_filter_generator): New prototype.
(filter_generators): Add binary_match_filter_generator.
(replace_jmp_placeholders): Handle JMP_PLACEHOLDER_ALLOW case.
(linear_filter_generator): New argument for replace_jmp_placeholders.
(bpf_syscalls_match, binary_match_filter_generator): New functions.
(dump_seccomp_bpf): Handle ldwimm, jset, rsh, lsh, and, tax, and txa
instructions.
Signed-off-by: Paul Chaignon <paul.chaignon@gmail.com>
filter_seccomp: list of seccomp filter generation strategies
Iterate over a list of BPF program generators, to select the shortest
program. This commit doesn't introduce new generation strategies.
* filter_seccomp.c (filter_generator_t): New typedef.
(linear_filter_generator): New prototype.
(filter_generators, filters): New variables.
(seccomp_filter): Remove.
(init_sock_filter): Rename to linear_filter_generator, signal overflow
via function argument instead of erroring out.
(check_seccomp_filter): Find the shortest program generated by all
generation strategies, error in case of overflow in all generated
programs.
Signed-off-by: Paul Chaignon <paul.chaignon@gmail.com>
tests: skip a few tests on systems with nproc == 1
Several tests require some concurrency between threads,
skip these tests if nproc reports 1.
* tests/init.sh (require_min_nproc): New function.
* tests/status-none-threads.test: Use it.
* tests/threads-execve.test: Likewise.
* xlat/v4l2_sliced_flags.in (V4L2_SLICED_TELETEXT_B, V4L2_SLICED_VPS,
V4L2_SLICED_CAPTION_525, V4L2_SLICED_WSS_625): Add fallback definitions.
(V4L2_SLICED_VBI_525): Comment out as it has the same value as
V4L2_SLICED_CAPTION_525.
(V4L2_SLICED_VBI_625): Add fallback definition, move up.
For some reason, it also used to include V4L2_VBI_ITU_[56]25_F[12]_START
constants that are definitely unrelated to V4L2_VBI_* flags.
* xlat/v4l2_vbi_flags.in (V4L2_VBI_UNSYNC, V4L2_VBI_INTERLACED): Add
fallback definitions.
(V4L2_VBI_ITU_525_F1_START, V4L2_VBI_ITU_525_F2_START,
V4L2_VBI_ITU_625_F1_START, V4L2_VBI_ITU_625_F2_START): Remove.
xlat: provide fallback definitions for V4L2_TUNER_* constants
Also, since V4L2_TUNER_*_SAP have the same values as V4L2_TUNER_*_LANG2,
comment out the former.
* xlat/v4l2_tuner_audmodes.in: Add fallback definitions for constants,
add #value_indexed.
(V4L2_TUNER_MODE_SAP): Comment out.
* xlat/v4l2_tuner_capabilities.in: Add fallback definitions for
constants.
(V4L2_TUNER_CAP_SAP): Comment out.
* xlat/v4l2_tuner_rxsubchanses.in: Add fallback definitions for
constants.
(V4L2_TUNER_SUB_SAP): Comment out.
When a structure is initialized using designated initializers,
clang >= 7, apparently, does not initialize trailing padding
of the structure.
* tests/bpf-obj_get_info_by_fd.c [CHECK_OBJ_PROG]
(bpf_prog_load_attr_sizes): Replace the size of struct
BPF_PROG_LOAD_struct with BPF_PROG_LOAD_struct_size.
filter_seccomp: use init_sock_filter to check the number of BPF instructions
With this commit, the number of instructions in the seccomp BPF program
is checked by directly constructing the program. The BPF program is
therefore created during the check for seccomp availability instead of
when seccomp filtering is initialized. The BPF program is saved
as a global variable between the two operations.
* filter_seccomp.c (seccomp_filter, bpf_prog): New variables.
(init_sock_filter): Disable seccomp-filter in case of overflow, move dump
of BPF program in debug mode...
(init_seccomp_filter): ...here.
(check_bpf_program_size): Remove.
(dump_seccomp_bpf): Remove arguments as they are now global variables.
Signed-off-by: Paul Chaignon <paul.chaignon@gmail.com>
Dmitry V. Levin
Paul Chaignon
Eugene Syromyatnikov
Gleb Fotengauer-Malinovskiy