On Sat, Nov 14, 2026 at 07:19:19PM +0100, Renzo Davoli wrote: > PTRACE_SET_SYSCALL_INFO is a generic ptrace API that complements > PTRACE_GET_SYSCALL_INFO by allowing a tracer to modify details of a > system call in which the tracee is currently blocked. > > This patch set extends PTRACE_SET_SYSCALL_INFO with support for > skipping a system call at a PTRACE_EVENT_SECCOMP stop and setting its > return value. To do this, the tracer changes 'op' from > PTRACE_SYSCALL_INFO_SECCOMP to PTRACE_SYSCALL_INFO_EXIT and fills in the > exit union fields (rval and is_error).
Hello, from a user point of view it is not clear to me why this would be done at the seccomp event when seccomp is not used. > > Michal pointed out in the review of v7 [1] that the register state alone > cannot tell whether a syscall has to be executed or skipped with a > preset return value, since on some architectures the return value > shares a register with the syscall number or the arguments, and that > the "already dispatched" information should be architecture > independent rather than an architecture-specific thread flag. On s390 the removal of such architecture-specific flag is proposed here: https://lore.kernel.org/linux-s390/[email protected]/T/#m04254224bb28b747314346153629207c5acca6f9 While that shows a way to address this problem not all affected architectures implement it. Thanks Michal > > v8 does exactly that: the tracer no longer sets the syscall number to -1 > and relies on syscall_get_nr() reading it back as negative. Instead, > patch 1 adds a per-task flag owned by seccomp, which the tracer sets > while the tracee is stopped. When the tracee resumes, seccomp skips the > system call just like for SECCOMP_RET_ERRNO, a path every architecture > already handles without touching the return value. This also fixes > skipping on s390, where -1 is read back as syscall 65535. > > On powerpc, a separate fix has been posted [2]: > syscall_set_return_value() currently sets TIF_SYSCALL_RET on the > *caller*, i.e. on the tracer when called from PTRACE_SET_SYSCALL_INFO, > and skipped system calls lose non-error return values. These problems > are already present in the source tree and break the existing > set_syscall_info selftest and seccomp_bpf user_notification_basic on > ppc64le independently of this series, which therefore no longer > depends on [2]. > > [1] https://lore.kernel.org/lkml/[email protected]/ > [2] https://lore.kernel.org/lkml/[email protected]/ > [3] > https://sashiko.dev/#/patchset/[email protected]?part=2 > > Testing (tools/testing/selftests/ptrace/set_syscall_info and > tools/testing/selftests/seccomp/seccomp_bpf): > > v11, QEMU, v7.4-rc2: > > x86_64 (KVM) > ptrace/set_syscall_info 2/2 > seccomp/seccomp_bpf 142/142 > > v9, QEMU, v7.3-rc6 (the v11 change is architecture independent): > > x86_64 (KVM) ppc64le (TCG) s390x (TCG) > ptrace/set_syscall_info 2/2 2/2 2/2 > seccomp/seccomp_bpf 111/111 111/111 110/111 (*) > > Build-tested also with CONFIG_SECCOMP=n and CONFIG_SECCOMP_FILTER=n. > > Skipping getpid() at a PTRACE_EVENT_SECCOMP stop, with a filter that > kills the process on unknown syscall numbers: > > op=SECCOMP, nr=-1 (before) error r3 left as is SIGSYS kill > op=SECCOMP->EXIT (this) -ESRCH -ESRCH -ESRCH > > (*) user_notification_wait_killable_after_reply timed out under TCG > while another VM was running; it passes 5/5 both with and without > this series on an idle x86_64 KVM guest. > > ppc64le results were obtained with an earlier version of the powerpc > fix [2] applied. > > Not tested: MIPS, arm64, riscv, loongarch, real hardware. > > Changes in v11: > * once the system call is skipped, PTRACE_GET_SYSCALL_INFO reports the > stop as a PTRACE_SYSCALL_INFO_EXIT one, as the syscall number and > arguments may have been overwritten by the return value (reported by > Sashiko [3]); a later PTRACE_SYSCALL_INFO_SECCOMP update is still > rejected, now by the generic check on the stop type > * selftests/ptrace/set_syscall_info.c: check PTRACE_GET_SYSCALL_INFO > after the system call is skipped > > Changes in v10: > * rebased to Linux 7.4-rc2 > * no longer depends on the powerpc fix [2], which addresses a bug > already present in the source tree > > Changes in v9: > * selftests/ptrace/set_syscall_info.c: check for seccomp filter support, > fix failure report > * rebased to Linux 7.3 > > Changes in v8: > * skip the system call via a seccomp per-task flag instead of > syscall_set_nr(-1), as suggested by Michal Suchánek > * new helper ptrace_set_syscall_info_seccomp_skip() instead of a bool > argument to ptrace_set_syscall_info_exit() (Michal Suchánek) > * once the system call is skipped, reject further > PTRACE_SYSCALL_INFO_SECCOMP updates during the same stop > * selftest: also test an error return value and the above > * dropped the Reviewed-by tags of Oleg and Dmitry, as patch 2 changed > significantly > > Changes in v7: > * rebased to Linux 7.3-rc1 > > Changes in v6: > * selftests/ptrace/set_syscall_info.c: improve messages in case of failure. > > Changes in v5: > * reworded the explanation for not supporting PTRACE_SYSCALL_INFO_ENTRY yet. > * selftests/ptrace: removed a redundant check. > > Changes in v4: > * Reworded the commit messages for clarity. > * Renamed the local variable child_op to op in ptrace_set_syscall_info() > * Clean up and improve the coding style of selftests/ptrace/set_syscall_info.c > (suggested by Dmitry V. Levin) > > Changes in v3: > * restrict the syscall skipping feature to PTRACE_SYSCALL_INFO_SECCOMP > > Changes in v2: > bugfix: _NONE -> _EXIT transition was erroneously permitted > > Davide Berardi (1): > seccomp: let the tracer skip a syscall independently of its number > > Renzo Davoli (2): > ptrace: add PTRACE_SET_SYSCALL_INFO syscall skipping support > selftests/ptrace: add a test case for PTRACE_SET_SYSCALL_INFO syscall > skipping > > include/linux/seccomp.h | 33 +++ > include/linux/seccomp_types.h | 4 + > kernel/ptrace.c | 49 +++- > kernel/seccomp.c | 11 +- > .../selftests/ptrace/set_syscall_info.c | 249 +++++++++++++++++- > 5 files changed, 342 insertions(+), 4 deletions(-) > > base-commit: 7e1392a42ea288d70a266f6553fe06f57f8647ea > -- > 2.55.0 >
