Re: [PATCH 1/2] powerpc: Restore all registers after a traced syscall again

[Shrikanth Hegde <sshegde@linux.ibm.com>]

Hi Michal,

On 10/12/26 11:08 PM, Michal Suchanek wrote:
> From: Bernhard M. Wiedemann <bwiedemann@suse.de>
> 
> syscall_exit_prepare() asks the exit assembly to reload all GPRs, CTR
> and XER from pt_regs only if the thread flags contain any of the flags
> in _TIF_SYSCALL_DOTRACE. With the new entry code based GENERIC_ENTRY the
> flags are no longer set.
> 
> This breaks tracers that modify a stopped syscall, for example ones that
> re-execute the syscall instruction by moving the NIP back by 4 at the
> syscall-exit stop with the syscall number put back into r0: the sc then
> runs with r0 = 0, which is __NR_restart_syscall. dettrace fails like
> this with "No filter rule for system call: restart_syscall". A tracer
> that writes r4 at the syscall-exit stop of an sc sees the tracee
> continue with r4 = 0 instead; with a 7.1 kernel both work.
> 
> Check the equivalent syscall_work bits instead, and drop the now unused
> _TIF_SYSCALL_DOTRACE.
> 
> Fixes: bee25f97ad24 ("powerpc: Enable GENERIC_ENTRY feature")
> Cc: stable@vger.kernel.org # 7.2.x: c7585b8e99ad: powerpc: Don't drop _TIF_RESTOREALL on syscall restart
> Cc: stable@vger.kernel.org # 7.2.x
> Assisted-by: LLM
> Signed-off-by: Bernhard M. Wiedemann <bwiedemann@suse.de>
> [ms: reword commit message]
> Signed-off-by: Michal Suchánek <msuchanek@suse.de>
> ---
> A backport also needs commit c7585b8e99ad ("powerpc: Don't drop
> _TIF_RESTOREALL on syscall restart"), without which the flag can still
> be lost.
> 
> It has been pointed out that the newly defined
> PPC_SYSCALL_WORK_RESTOREALL is almost the same as SYSCALL_WORK_ENTER and
> the existing define might be used instead. However, it is not clear if
> this is expected or if this is merely coincidence, and not guaranteed.
> 

They seems to semantically different. So for now, we can keep it separate.

> Testing: a small tracer (pokes r4 at the syscall-exit stop of an sc,
> and replays a syscall by moving NIP back with a new r0) shows both
> writes lost on a ppc64le and both working on 7.1.8.
> ------------[ cut here ]------------
> // SPDX-License-Identifier: GPL-2.0
> /*
>   * Does a register a tracer writes at the syscall-exit stop of an "sc"
>   * reach the tracee? Checks r4, and a replay: NIP moved back onto the sc
>   * with r0 set to another syscall number.
>   * Build: gcc -O2 -o ppc-exitregs ppc-exitregs.c (powerpc64 only)
>   */
> 
> /* word offsets in struct pt_regs, arch/powerpc/include/uapi/asm/ptrace.h */
> 
> static long do_sc(long nr, long *r4io)
> {
> 	register long r0 __asm__("r0") = nr;
> 	register long r3 __asm__("r3") = 0;
> 	register long r4 __asm__("r4") = *r4io;
> 
> 	__asm__ volatile("sc"
> 			 : "+r"(r0), "+r"(r3), "+r"(r4)
> 			 :
> 			 : "r5", "r6", "r7", "r8", "r9", "r10", "r11", "r12",
> 			   "cr0", "ctr", "xer", "memory");
> 	*r4io = r4;
> 	return r3;
> }
> 
> static long peek(pid_t pid, int w)
> {
> 	errno = 0;
> 	long v = ptrace(PTRACE_PEEKUSER, pid, (void *)(w * sizeof(long)), 0);
> 	if (errno) {
> 		perror("PEEKUSER");
> 		exit(2);
> 	}
> 	return v;
> }
> 
> static void poke(pid_t pid, int w, long v)
> {
> 	if (ptrace(PTRACE_POKEUSER, pid, (void *)(w * sizeof(long)), (void *)v)) {
> 		perror("POKEUSER");
> 		exit(2);
> 	}
> }
> 
> /* Resume to the next syscall stop, which must be one. */
> static void next_stop(pid_t pid)
> {
> 	int st;
> 
> 	ptrace(PTRACE_SYSCALL, pid, 0, 0);
> 	if (waitpid(pid, &st, 0) != pid || !WIFSTOPPED(st) ||
> 	    WSTOPSIG(st) != (SIGTRAP | 0x80)) {
> 		fprintf(stderr, "unexpected wait status %#x\n", st);
> 		exit(2);
> 	}
> }
> 
> int main(void)
> {
> 	pid_t pid = fork();
> 	int st, fail = 0;
> 
> 	if (pid == 0) {
> 		long r4 = 0x1111, r4a, ret;
> 
> 		ptrace(PTRACE_TRACEME, 0, 0, 0);
> 		raise(SIGSTOP);
> 		do_sc(__NR_getpid, &r4);
> 		r4a = r4;
> 		ret = do_sc(__NR_getpid, &r4);
> 		/* bit 0: the tracer's r4 was lost; bit 1: the replay of
> 		 * the second getpid as getppid returned something else */
> 		_exit((r4a != 0x2222) | (ret != getppid()) << 1);
> 	}
> 
> 	waitpid(pid, &st, 0);
> 	ptrace(PTRACE_SETOPTIONS, pid, 0, (void *)PTRACE_O_TRACESYSGOOD);
> 
> 	/* 1: write r4 at the exit stop of getpid */
> 	next_stop(pid);				/* entry */
> 	next_stop(pid);				/* exit */
> 	poke(pid, W_R4, 0x2222);
> 	next_stop(pid);				/* entry of the second getpid */
> 
> 	/* 2: replay the second getpid as getppid */
> 	next_stop(pid);				/* exit */
> 	poke(pid, W_NIP, peek(pid, W_NIP) - 4);
> 	poke(pid, W_R0, __NR_getppid);
> 	next_stop(pid);				/* entry of the replay */
> 	long nr = peek(pid, W_R0);
> 	printf("replayed syscall arrives as nr %ld (%s)\n", nr,
> 	       nr == __NR_getppid ? "getppid, OK" :
> 	       nr == __NR_restart_syscall ? "restart_syscall, BUG" : "?");
> 	if (nr != __NR_getppid) {
> 		fail = 1;
> 		poke(pid, W_R0, __NR_getppid);	/* entry writes do work */
> 	}
> 	next_stop(pid);				/* exit */
> 	ptrace(PTRACE_CONT, pid, 0, 0);
> 	waitpid(pid, &st, 0);
> 	if (!WIFEXITED(st)) {
> 		printf("child did not exit normally: %#x\n", st);
> 		return 2;
> 	}
> 	printf("r4 written at the exit stop: %s\n",
> 	       WEXITSTATUS(st) & 1 ? "lost, BUG" : "kept, OK");
> 	printf("replay result: %s\n",
> 	       WEXITSTATUS(st) & 2 ? "wrong" : "getppid(), OK");
> 	return fail || WEXITSTATUS(st);
> }
> ---
>   arch/powerpc/include/asm/thread_info.h |  3 ---
>   arch/powerpc/kernel/interrupt.c        | 16 ++++++++++++++--
>   2 files changed, 14 insertions(+), 5 deletions(-)
> 
> diff --git a/arch/powerpc/include/asm/thread_info.h b/arch/powerpc/include/asm/thread_info.h
> index 1e069a2e7ce87..0e8dc71c8d3fe 100644
> --- a/arch/powerpc/include/asm/thread_info.h
> +++ b/arch/powerpc/include/asm/thread_info.h
> @@ -142,9 +142,6 @@ void arch_setup_new_exec(void);
>   #define _TIF_SYSCALL_TRACEPOINT	(1<<TIF_SYSCALL_TRACEPOINT)
>   #define _TIF_EMULATE_STACK_STORE	(1<<TIF_EMULATE_STACK_STORE)
>   #define _TIF_SYSCALL_EMU	(1<<TIF_SYSCALL_EMU)
> -#define _TIF_SYSCALL_DOTRACE	(_TIF_SYSCALL_TRACE | _TIF_SYSCALL_AUDIT | \
> -				 _TIF_SECCOMP | _TIF_SYSCALL_TRACEPOINT | \
> -				 _TIF_SYSCALL_EMU)
>   
>   #define _TIF_USER_WORK_MASK	(_TIF_SIGPENDING | _TIF_NEED_RESCHED | \
>   				 _TIF_NEED_RESCHED_LAZY | _TIF_NOTIFY_RESUME | \
> diff --git a/arch/powerpc/kernel/interrupt.c b/arch/powerpc/kernel/interrupt.c
> index 55f9c0c9922ac..524c51104e4c9 100644
> --- a/arch/powerpc/kernel/interrupt.c
> +++ b/arch/powerpc/kernel/interrupt.c
> @@ -75,6 +75,18 @@ static notrace __always_inline bool prep_irq_for_enabled_exit(bool restartable)
>   	return true;
>   }
>   
> +/*
> + * Syscall work after which the exit must reload all registers from
> + * pt_regs: a tracer may have changed any of them at a syscall or seccomp
> + * stop, or under PTRACE_SYSEMU. Audit and the tracepoint are included to
> + * keep the pre-GENERIC_ENTRY behaviour.
> + */
> +#define PPC_SYSCALL_WORK_RESTOREALL	(SYSCALL_WORK_SECCOMP |		\
> +					 SYSCALL_WORK_SYSCALL_TRACEPOINT | \
> +					 SYSCALL_WORK_SYSCALL_TRACE |	\
> +					 SYSCALL_WORK_SYSCALL_EMU |	\
> +					 SYSCALL_WORK_SYSCALL_AUDIT)
> +

nit: Apart from mukesh's comment,
can we keep the name as SYSCALL_WORK_RESTOREALL instead? It is already in PPC folder.

>   /*
>    * This should be called after a syscall returns, with r3 the return value
>    * from the syscall. If this function returns non-zero, the system call
> @@ -117,9 +129,9 @@ notrace unsigned long syscall_exit_prepare(unsigned long r3,
>   		regs->gpr[3] = r3;
>   	}
>   
> -	if (unlikely(ti_flags & _TIF_SYSCALL_DOTRACE)) {
> +	if (unlikely(READ_ONCE(current_thread_info()->syscall_work) &
> +		     PPC_SYSCALL_WORK_RESTOREALL))
>   		ret |= _TIF_RESTOREALL;
> -	}
>   
>   	syscall_exit_to_user_mode(regs);
>