[RFC,2/2] riscv: selftests: Add a ptrace test to check a0 of restarted syscall

From: Quan Zhou <zhouquan@iscas.ac.cn>

This test creates two processes: a tracer and a tracee. The tracer actively
sends a SIGUSR1 signal in user mode to interrupt the read syscall being
executed by the tracee. We will reset a0/orig_a0 and then observe the value
of a0 held by the restarted read syscall.

Compared to the test program, a more common scenario is the use of the
exece syscall, which sends a signal in the kernel path to restart
the syscall.

Signed-off-by: Quan Zhou <zhouquan@iscas.ac.cn>
---
 tools/testing/selftests/riscv/Makefile        |   2 +-
 tools/testing/selftests/riscv/abi/.gitignore  |   1 +
 tools/testing/selftests/riscv/abi/Makefile    |  12 ++
 .../riscv/abi/ptrace_restart_syscall.c        | 148 ++++++++++++++++++
 4 files changed, 162 insertions(+), 1 deletion(-)
 create mode 100644 tools/testing/selftests/riscv/abi/.gitignore
 create mode 100644 tools/testing/selftests/riscv/abi/Makefile
 create mode 100644 tools/testing/selftests/riscv/abi/ptrace_restart_syscall.c

On 2024/6/20 10:55, Charlie Jenkins wrote:
> On Wed, Jun 19, 2024 at 10:01:47AM +0800, zhouquan@iscas.ac.cn wrote:
>> From: Quan Zhou <zhouquan@iscas.ac.cn>
>>
>> This test creates two processes: a tracer and a tracee. The tracer actively
>> sends a SIGUSR1 signal in user mode to interrupt the read syscall being
>> executed by the tracee. We will reset a0/orig_a0 and then observe the value
>> of a0 held by the restarted read syscall.
> 
> I don't quite follow what the goal of this test is. With orig_a0 being
> added to the previous patch for ptrace, a more constrained test could
> ensure that this value is respected.
> 

Sry, I may not have described the patch clearly enough. This patch 
provides a channel for modifying a0 in user-space ptrace via orig_a0. 
Here, I will try to outline the whole situation:

1、When the tracer calls ptrace to modify regs->a0, can the tracee's a0 
be correctly modified?

Through testing, if the user only modifies regs->a0, it doesn't work. Why?

The execution flow of the tracee in the test program is as follows.Prior 
to this explanation:

- PTRACE_SYSCALL can make the tracee block before and after executing
   a syscall.
- The tracer sends SIGUSR1 to interrupt read, and the kernel will
   restart it.
- Please note the point marked with (*), which I believe is the cause
   of the issue.

user     kernel
     |
     |
     |
      read
         | +-> regs->orig_a0 = regs->a0; //(*1)
         |                                       <=tracer:PTRACE_SYSCALL
         | +-> syscall_enter_from_user_mode
               +-> ptrace_report_syscall_entry
                   +-> ptrace_stop
         | //stopped
         |                                       <= tracer:SIGUSR1
         |
         | //resume                              <= tracer:PTRACE_SYSCALL
         | syscall_handler...
         |
         | +-> syscall_exit_to_user_mode
               +-> syscall_exit_to_user_mode_prepare
                   +-> ptrace_report_syscall_exit
                       +-> ptrace_stop
     | //stopped
     |
     | /* Change a0/orig_a0 here and observe the restarted syscall */
     | regs->{a0/orig_a0} = fd_zero; //(*2)
     | ptrace(PTRACE_SETREGSET, ...);
         |                                      <= tracer:PTRACE_SYSCALL
         | //restarting..., skip SIGUSR1
         |
         | +-> exit_to_user_mode_loop
               +-> arch_do_signal_or_restart
                   +-> /* Settings for syscall restart */
                       regs->a0 = regs->orig_a0; //(*3)
     | //stopped
     | //and block before the syscall again, get current regs->a0
     | *result = regs->a0;
     |
     | /* Now, Check regs->a0 of restarted syscall */
     | EXPECT_NE(0x5, result); //for PTRACE_SETREGSSET a0, failed
     | EXPECT_EQ(0x5, result); //for PTRACE_SETREGSSET orig_a0, succeed

If I'm wrong, please let me know. 🙂

2、Actually, I discovered the issue while using the execve function.
When I tried to modify the first parameter of execve in the tracer,
I found it didn't work.

As for why not use execve for testing, there are two reasons:

1) The root cause of this issue is that when a syscall is interrupted
and then resumed, it restarts with orig_a0 instead of a0, so modifying
a0 doesn't work. I want to focus the test on the "restarted syscall".

2) Compared to the current test scenario, execve is terminated by ptrace
earlier, so I chose a later point. In fact, setting regs->a0 in the path
between (*1) and (*3) is ineffective because it will eventually be 
overwritten by orig_a0, correct?

The current test may not intuitively reflect the issue. If possible, I 
will provide a more comprehensive test based on everyone's suggestions.

Thanks,
Quan

>>
>> Compared to the test program, a more common scenario is the use of the
>> exece syscall, which sends a signal in the kernel path to restart
>> the syscall.
>>
>> Signed-off-by: Quan Zhou <zhouquan@iscas.ac.cn>
>> ---
>>   tools/testing/selftests/riscv/Makefile        |   2 +-
>>   tools/testing/selftests/riscv/abi/.gitignore  |   1 +
>>   tools/testing/selftests/riscv/abi/Makefile    |  12 ++
>>   .../riscv/abi/ptrace_restart_syscall.c        | 148 ++++++++++++++++++
>>   4 files changed, 162 insertions(+), 1 deletion(-)
>>   create mode 100644 tools/testing/selftests/riscv/abi/.gitignore
>>   create mode 100644 tools/testing/selftests/riscv/abi/Makefile
>>   create mode 100644 tools/testing/selftests/riscv/abi/ptrace_restart_syscall.c
>>
>> diff --git a/tools/testing/selftests/riscv/Makefile b/tools/testing/selftests/riscv/Makefile
>> index 7ce03d832b64..98541dc2f164 100644
>> --- a/tools/testing/selftests/riscv/Makefile
>> +++ b/tools/testing/selftests/riscv/Makefile
>> @@ -5,7 +5,7 @@
>>   ARCH ?= $(shell uname -m 2>/dev/null || echo not)
>>   
>>   ifneq (,$(filter $(ARCH),riscv))
>> -RISCV_SUBTARGETS ?= hwprobe vector mm sigreturn
>> +RISCV_SUBTARGETS ?= hwprobe vector mm sigreturn abi
>>   else
>>   RISCV_SUBTARGETS :=
>>   endif
>> diff --git a/tools/testing/selftests/riscv/abi/.gitignore b/tools/testing/selftests/riscv/abi/.gitignore
>> new file mode 100644
>> index 000000000000..e1e00ffb9db9
>> --- /dev/null
>> +++ b/tools/testing/selftests/riscv/abi/.gitignore
>> @@ -0,0 +1 @@
>> +abi
> 
> The gitignore should contain a list of all of the generated binaries
> that should be ignored. Can you put ptrace_restart_syscall in here
> instead of abi?
> 

...yeah, I will fix it later.

>> diff --git a/tools/testing/selftests/riscv/abi/Makefile b/tools/testing/selftests/riscv/abi/Makefile
>> new file mode 100644
>> index 000000000000..634fa7de74e6
>> --- /dev/null
>> +++ b/tools/testing/selftests/riscv/abi/Makefile
>> @@ -0,0 +1,12 @@
>> +# SPDX-License-Identifier: GPL-2.0
>> +# Copyright (C) 2021 ARM Limited
>> +# Originally tools/testing/arm64/abi/Makefile
>> +
>> +CFLAGS += -I$(top_srcdir)/tools/include
>> +
>> +TEST_GEN_PROGS := ptrace_restart_syscall
>> +
>> +include ../../lib.mk
>> +
>> +$(OUTPUT)/ptrace_restart_syscall: ptrace_restart_syscall.c
>> +	$(CC) -static -o$@ $(CFLAGS) $(LDFLAGS) $^
>> diff --git a/tools/testing/selftests/riscv/abi/ptrace_restart_syscall.c b/tools/testing/selftests/riscv/abi/ptrace_restart_syscall.c
>> new file mode 100644
>> index 000000000000..3e25548cb95e
>> --- /dev/null
>> +++ b/tools/testing/selftests/riscv/abi/ptrace_restart_syscall.c
>> @@ -0,0 +1,148 @@
>> +// SPDX-License-Identifier: GPL-2.0-only
>> +#include <stdio.h>
>> +#include <stdlib.h>
>> +#include <string.h>
>> +#include <unistd.h>
>> +#include <fcntl.h>
>> +#include <signal.h>
>> +#include <errno.h>
>> +#include <sys/types.h>
>> +#include <sys/ptrace.h>
>> +#include <sys/stat.h>
>> +#include <sys/user.h>
>> +#include <sys/wait.h>
>> +#include <sys/uio.h>
>> +#include <linux/elf.h>
>> +#include <linux/unistd.h>
>> +#include <asm/ptrace.h>
>> +
>> +#include "../../kselftest_harness.h"
>> +
>> +#define ORIG_A0_AFTER_MODIFIED  0x5
>> +#define MODIFY_A0               0x01
>> +#define MODIFY_ORIG_A0          0x02
>> +
>> +#define perr_and_exit(fmt, ...) do {			\
>> +	char buf[256];					\
>> +	snprintf(buf, sizeof(buf), "%s:%d: " fmt ": %m\n",	\
>> +			__func__, __LINE__, ##__VA_ARGS__);	\
>> +	perror(buf);						\
>> +	exit(-1);						\
>> +} while (0)
>> +
>> +static inline void resume_and_wait_tracee(pid_t pid, int flag)
>> +{
>> +	int status;
>> +
>> +	if (ptrace(flag, pid, 0, 0))
>> +		perr_and_exit("failed to resume the tracee %d", pid);
>> +
>> +	if (waitpid(pid, &status, 0) != pid)
>> +		perr_and_exit("failed to wait for the tracee %d", pid);
>> +}
>> +
>> +static void ptrace_restart_syscall(int opt, int *result)
>> +{
>> +	int status;
>> +	int p[2], fd_zero;
>> +	pid_t pid;
>> +
>> +	struct user_regs_struct regs;
>> +	struct iovec iov = {
>> +		.iov_base = &regs,
>> +		.iov_len = sizeof(regs),
>> +	};
>> +
>> +	if (pipe(p))
>> +		perr_and_exit("failed to create a pipe");
>> +
>> +	fd_zero = open("/dev/zero", O_RDONLY);
>> +	if (fd_zero < 0)
>> +		perr_and_exit("failed to open /dev/zero");
>> +
>> +	pid = fork();
>> +	if (pid == 0) {
>> +		char c;
>> +
>> +		/* Mark oneself being traced */
>> +		if (ptrace(PTRACE_TRACEME, 0, 0, 0))
>> +			perr_and_exit("failed to request for tracer to trace me");
>> +
>> +		kill(getpid(), SIGSTOP);
>> +
>> +		if (read(p[0], &c, 1) != 1)
>> +			exit(1);
>> +
>> +		exit(0);
>> +	} else if (pid < 0)
>> +		exit(1);
>> +
>> +	if (waitpid(pid, &status, 0) != pid)
>> +		perr_and_exit("failed to wait for the tracee %d\n", pid);
>> +
>> +	/* Resume the tracee until the next syscall */
>> +	resume_and_wait_tracee(pid, PTRACE_SYSCALL);
>> +
>> +	/* Deliver a signal to interrupt the syscall */
>> +	kill(pid, SIGUSR1);
>> +
>> +	/* The tracee stops at syscall exit */
>> +	resume_and_wait_tracee(pid, PTRACE_SYSCALL);
>> +
>> +	/* Check tracee orig_a0 before syscall restart */
>> +	if (ptrace(PTRACE_GETREGSET, pid, NT_PRSTATUS, &iov))
>> +		perr_and_exit("failed to get tracee registers");
>> +	if (regs.orig_a0 != p[0])
>> +		perr_and_exit("unexpected a0");
>> +
>> +	/* Modify a0/orig_a0 for the restarted syscall */
>> +	switch (opt) {
>> +	case MODIFY_A0:
>> +		regs.a0 = fd_zero;
>> +		break;
>> +	case MODIFY_ORIG_A0:
>> +		regs.orig_a0 = fd_zero;
>> +		break;
>> +	}
>> +
>> +	if (ptrace(PTRACE_SETREGSET, pid, NT_PRSTATUS, &iov))
>> +		perr_and_exit("failed to set tracee registers");
>> +
>> +	/* Ignore SIGUSR1 signal */
>> +	resume_and_wait_tracee(pid, PTRACE_SYSCALL);
>> +
>> +	/* Stop at the entry point of the restarted syscall */
>> +	resume_and_wait_tracee(pid, PTRACE_SYSCALL);
>> +
>> +	/* Now, check regs.a0 of the restarted syscall */
>> +	if (ptrace(PTRACE_GETREGSET, pid, NT_PRSTATUS, &iov))
>> +		perr_and_exit("failed to get tracee registers");
>> +	*result = regs.a0;
>> +
>> +	/* Resume the tracee */
>> +	ptrace(PTRACE_CONT, pid, 0, 0);
>> +	if (waitpid(pid, &status, 0) != pid)
>> +		perr_and_exit("failed to wait for the tracee");
>> +}
>> +
>> +TEST(ptrace_modify_a0)
>> +{
>> +	int result;
>> +
>> +	ptrace_restart_syscall(MODIFY_A0, &result);
>> +
>> +	/* The tracer's modification of a0 cannot affect the restarted tracee */
>> +	EXPECT_NE(ORIG_A0_AFTER_MODIFIED, result);
>> +}
>> +
>> +TEST(ptrace_modify_orig_a0)
>> +{
>> +	int result;
>> +
>> +	ptrace_restart_syscall(MODIFY_ORIG_A0, &result);
>> +
>> +	/* The tracer must modify orig_a0 to actually change the tracee's a0 */
>> +	EXPECT_EQ(ORIG_A0_AFTER_MODIFIED, result);
> 
> How does the value end up being 5?
> 
> - Charlie
> 

The tracer ultimately sets `fd_zero` to the restarted syscall.

Since 0, 1, and 2 are standard input, output, and error, the file 
descriptors will be allocated in this order: `p[0] -> p[1] -> fd_zero`. 
Thus, fd_zero will be 5.

>> +}
>> +
>> +TEST_HARNESS_MAIN
>> -- 
>> 2.34.1
>>

On Fri, Jun 21, 2024 at 02:29:07PM +0800, Quan Zhou wrote:
> 
> 
> On 2024/6/20 10:55, Charlie Jenkins wrote:
> > On Wed, Jun 19, 2024 at 10:01:47AM +0800, zhouquan@iscas.ac.cn wrote:
> > > From: Quan Zhou <zhouquan@iscas.ac.cn>
> > > 
> > > This test creates two processes: a tracer and a tracee. The tracer actively
> > > sends a SIGUSR1 signal in user mode to interrupt the read syscall being
> > > executed by the tracee. We will reset a0/orig_a0 and then observe the value
> > > of a0 held by the restarted read syscall.
> > 
> > I don't quite follow what the goal of this test is. With orig_a0 being
> > added to the previous patch for ptrace, a more constrained test could
> > ensure that this value is respected.
> > 
> 
> Sry, I may not have described the patch clearly enough. This patch provides
> a channel for modifying a0 in user-space ptrace via orig_a0. Here, I will
> try to outline the whole situation:
> 
> 1、When the tracer calls ptrace to modify regs->a0, can the tracee's a0 be
> correctly modified?
> 
> Through testing, if the user only modifies regs->a0, it doesn't work. Why?
> 
> The execution flow of the tracee in the test program is as follows.Prior to
> this explanation:
> 
> - PTRACE_SYSCALL can make the tracee block before and after executing
>   a syscall.
> - The tracer sends SIGUSR1 to interrupt read, and the kernel will
>   restart it.
> - Please note the point marked with (*), which I believe is the cause
>   of the issue.
> 
> user     kernel
>     |
>     |
>     |
>      read
>         | +-> regs->orig_a0 = regs->a0; //(*1)
>         |                                       <=tracer:PTRACE_SYSCALL
>         | +-> syscall_enter_from_user_mode
>               +-> ptrace_report_syscall_entry
>                   +-> ptrace_stop
>         | //stopped
>         |                                       <= tracer:SIGUSR1
>         |
>         | //resume                              <= tracer:PTRACE_SYSCALL
>         | syscall_handler...
>         |
>         | +-> syscall_exit_to_user_mode
>               +-> syscall_exit_to_user_mode_prepare
>                   +-> ptrace_report_syscall_exit
>                       +-> ptrace_stop
>     | //stopped
>     |
>     | /* Change a0/orig_a0 here and observe the restarted syscall */
>     | regs->{a0/orig_a0} = fd_zero; //(*2)
>     | ptrace(PTRACE_SETREGSET, ...);
>         |                                      <= tracer:PTRACE_SYSCALL
>         | //restarting..., skip SIGUSR1
>         |
>         | +-> exit_to_user_mode_loop
>               +-> arch_do_signal_or_restart
>                   +-> /* Settings for syscall restart */
>                       regs->a0 = regs->orig_a0; //(*3)
>     | //stopped
>     | //and block before the syscall again, get current regs->a0
>     | *result = regs->a0;
>     |
>     | /* Now, Check regs->a0 of restarted syscall */
>     | EXPECT_NE(0x5, result); //for PTRACE_SETREGSSET a0, failed
>     | EXPECT_EQ(0x5, result); //for PTRACE_SETREGSSET orig_a0, succeed
> 
> If I'm wrong, please let me know. 🙂
> 
> 2、Actually, I discovered the issue while using the execve function.
> When I tried to modify the first parameter of execve in the tracer,
> I found it didn't work.
> 
> As for why not use execve for testing, there are two reasons:
> 
> 1) The root cause of this issue is that when a syscall is interrupted
> and then resumed, it restarts with orig_a0 instead of a0, so modifying
> a0 doesn't work. I want to focus the test on the "restarted syscall".
> 
> 2) Compared to the current test scenario, execve is terminated by ptrace
> earlier, so I chose a later point. In fact, setting regs->a0 in the path
> between (*1) and (*3) is ineffective because it will eventually be
> overwritten by orig_a0, correct?

Thank you for the thorough explanation! I feel like a test case like the
following achieves the same goal but without needing the pipes and the
file. What do you think?

On 2024/6/22 04:20, Charlie Jenkins wrote:
> On Fri, Jun 21, 2024 at 02:29:07PM +0800, Quan Zhou wrote:
>>
>>
>> On 2024/6/20 10:55, Charlie Jenkins wrote:
>>> On Wed, Jun 19, 2024 at 10:01:47AM +0800, zhouquan@iscas.ac.cn wrote:
>>>> From: Quan Zhou <zhouquan@iscas.ac.cn>
>>>>
>>>> This test creates two processes: a tracer and a tracee. The tracer actively
>>>> sends a SIGUSR1 signal in user mode to interrupt the read syscall being
>>>> executed by the tracee. We will reset a0/orig_a0 and then observe the value
>>>> of a0 held by the restarted read syscall.
>>>
>>> I don't quite follow what the goal of this test is. With orig_a0 being
>>> added to the previous patch for ptrace, a more constrained test could
>>> ensure that this value is respected.
>>>
>>
>> Sry, I may not have described the patch clearly enough. This patch provides
>> a channel for modifying a0 in user-space ptrace via orig_a0. Here, I will
>> try to outline the whole situation:
>>
>> 1、When the tracer calls ptrace to modify regs->a0, can the tracee's a0 be
>> correctly modified?
>>
>> Through testing, if the user only modifies regs->a0, it doesn't work. Why?
>>
>> The execution flow of the tracee in the test program is as follows.Prior to
>> this explanation:
>>
>> - PTRACE_SYSCALL can make the tracee block before and after executing
>>    a syscall.
>> - The tracer sends SIGUSR1 to interrupt read, and the kernel will
>>    restart it.
>> - Please note the point marked with (*), which I believe is the cause
>>    of the issue.
>>
>> user     kernel
>>      |
>>      |
>>      |
>>       read
>>          | +-> regs->orig_a0 = regs->a0; //(*1)
>>          |                                       <=tracer:PTRACE_SYSCALL
>>          | +-> syscall_enter_from_user_mode
>>                +-> ptrace_report_syscall_entry
>>                    +-> ptrace_stop
>>          | //stopped
>>          |                                       <= tracer:SIGUSR1
>>          |
>>          | //resume                              <= tracer:PTRACE_SYSCALL
>>          | syscall_handler...
>>          |
>>          | +-> syscall_exit_to_user_mode
>>                +-> syscall_exit_to_user_mode_prepare
>>                    +-> ptrace_report_syscall_exit
>>                        +-> ptrace_stop
>>      | //stopped
>>      |
>>      | /* Change a0/orig_a0 here and observe the restarted syscall */
>>      | regs->{a0/orig_a0} = fd_zero; //(*2)
>>      | ptrace(PTRACE_SETREGSET, ...);
>>          |                                      <= tracer:PTRACE_SYSCALL
>>          | //restarting..., skip SIGUSR1
>>          |
>>          | +-> exit_to_user_mode_loop
>>                +-> arch_do_signal_or_restart
>>                    +-> /* Settings for syscall restart */
>>                        regs->a0 = regs->orig_a0; //(*3)
>>      | //stopped
>>      | //and block before the syscall again, get current regs->a0
>>      | *result = regs->a0;
>>      |
>>      | /* Now, Check regs->a0 of restarted syscall */
>>      | EXPECT_NE(0x5, result); //for PTRACE_SETREGSSET a0, failed
>>      | EXPECT_EQ(0x5, result); //for PTRACE_SETREGSSET orig_a0, succeed
>>
>> If I'm wrong, please let me know. 🙂
>>
>> 2、Actually, I discovered the issue while using the execve function.
>> When I tried to modify the first parameter of execve in the tracer,
>> I found it didn't work.
>>
>> As for why not use execve for testing, there are two reasons:
>>
>> 1) The root cause of this issue is that when a syscall is interrupted
>> and then resumed, it restarts with orig_a0 instead of a0, so modifying
>> a0 doesn't work. I want to focus the test on the "restarted syscall".
>>
>> 2) Compared to the current test scenario, execve is terminated by ptrace
>> earlier, so I chose a later point. In fact, setting regs->a0 in the path
>> between (*1) and (*3) is ineffective because it will eventually be
>> overwritten by orig_a0, correct?
> 
> Thank you for the thorough explanation! I feel like a test case like the
> following achieves the same goal but without needing the pipes and the
> file. What do you think?
> 
>>From 5f13cdf8f7312b2b3cc98fbfbb3c95fcef62c0f0 Mon Sep 17 00:00:00 2001
> From: Charlie Jenkins <charlie@rivosinc.com>
> Date: Fri, 21 Jun 2024 12:58:29 -0700
> Subject: [PATCH] riscv: selftests: Add a ptrace test to check a0 of restarted
>   syscall
> 
> Add a riscv selftest that checks that a0 of syscalls are able to be
> replaced. When entering a syscall, a0 contains the first argument to the
> syscall and upon exiting, a0 contains the return value. To replace the
> a0 argument instead of the a0 return value with ptrace after halting the
> program with PTRACE_SYSCALL, orig_a0 must be used. This test checks that
> orig_a0 allows a syscall argument to be modified, and that changing a0
> does not change the syscall argument.
> 
> Signed-off-by: Charlie Jenkins <charlie@rivosinc.com>
> ---
>   .../riscv/abi/ptrace_restart_syscall.c        | 123 ++++++++++++++++++
>   1 file changed, 123 insertions(+)
>   create mode 100644 tools/testing/selftests/riscv/abi/ptrace_restart_syscall.c
> 
> diff --git a/tools/testing/selftests/riscv/abi/ptrace_restart_syscall.c b/tools/testing/selftests/riscv/abi/ptrace_restart_syscall.c
> new file mode 100644
> index 000000000000..e74ae02c6850
> --- /dev/null
> +++ b/tools/testing/selftests/riscv/abi/ptrace_restart_syscall.c
> @@ -0,0 +1,123 @@
> +// SPDX-License-Identifier: GPL-2.0-only
> +#include <stdio.h>
> +#include <stdlib.h>
> +#include <string.h>
> +#include <unistd.h>
> +#include <fcntl.h>
> +#include <signal.h>
> +#include <errno.h>
> +#include <sys/types.h>
> +#include <sys/ptrace.h>
> +#include <sys/stat.h>
> +#include <sys/user.h>
> +#include <sys/wait.h>
> +#include <sys/uio.h>
> +#include <linux/elf.h>
> +#include <linux/unistd.h>
> +#include <asm/ptrace.h>
> +
> +#include "../../kselftest_harness.h"
> +
> +#define DEFAULT_A0		0x6
> +
> +#define ORIG_A0_AFTER_MODIFIED  0x5
> +#define MODIFY_A0               0x01
> +#define MODIFY_ORIG_A0          0x02
> +
> +#define perr_and_exit(fmt, ...) do {			\
> +	char buf[256];					\
> +	snprintf(buf, sizeof(buf), "%s:%d: " fmt ": %m\n",	\
> +			__func__, __LINE__, ##__VA_ARGS__);	\
> +	perror(buf);						\
> +	exit(-1);						\
> +} while (0)
> +
> +static inline void resume_and_wait_tracee(pid_t pid, int flag)
> +{
> +	int status;
> +
> +	if (ptrace(flag, pid, 0, 0))
> +		perr_and_exit("failed to resume the tracee %d\n", pid);
> +
> +	if (waitpid(pid, &status, 0) != pid)
> +		perr_and_exit("failed to wait for the tracee %d\n", pid);
> +}
> +
> +static void ptrace_restart_syscall(int opt, int *result)
> +{
> +	int status;
> +	pid_t pid;
> +
> +	struct user_regs_struct regs;
> +	struct iovec iov = {
> +		.iov_base = &regs,
> +		.iov_len = sizeof(regs),
> +	};
> +
> +	pid = fork();
> +	if (pid == 0) {
> +		/* Mark oneself being traced */
> +		long val = ptrace(PTRACE_TRACEME, 0, 0, 0);
> +		if (val)
> +			perr_and_exit("failed to request for tracer to trace me: %ld\n", val);
> +
> +		kill(getpid(), SIGSTOP);
> +
> +		/* Perform exit syscall that will be intercepted */
> +		exit(DEFAULT_A0);
> +	} else if (pid < 0) {
> +		exit(1);
> +	}
> +
> +	if (waitpid(pid, &status, 0) != pid)
> +		perr_and_exit("failed to wait for the tracee %d\n", pid);
> +
> +	/* Stop at the entry point of the restarted syscall */
> +	resume_and_wait_tracee(pid, PTRACE_SYSCALL);
> +
> +	/* Now, check regs.a0 of the restarted syscall */
> +	if (ptrace(PTRACE_GETREGSET, pid, NT_PRSTATUS, &iov))
> +		perr_and_exit("failed to get tracee registers\n");
> +
> +	/* Modify a0/orig_a0 for the restarted syscall */
> +	switch (opt) {
> +	case MODIFY_A0:
> +		regs.a0 = ORIG_A0_AFTER_MODIFIED;
> +		break;
> +	case MODIFY_ORIG_A0:
> +		regs.orig_a0 = ORIG_A0_AFTER_MODIFIED;
> +		break;
> +	}
> +
> +	if (ptrace(PTRACE_SETREGSET, pid, NT_PRSTATUS, &iov))
> +		perr_and_exit("failed to set tracee registers\n");
> +
> +	/* Resume the tracee */
> +	ptrace(PTRACE_CONT, pid, 0, 0);
> +	if (waitpid(pid, &status, 0) != pid)
> +		perr_and_exit("failed to wait for the tracee\n");
> +
> +	*result = WEXITSTATUS(status);
> +}
> +
> +TEST(ptrace_modify_a0)
> +{
> +	int result;
> +
> +	ptrace_restart_syscall(MODIFY_A0, &result);
> +
> +	/* The tracer's modification of a0 cannot affect the restarted tracee */
> +	EXPECT_EQ(DEFAULT_A0, result);
> +}
> +
> +TEST(ptrace_modify_orig_a0)
> +{
> +	int result;
> +
> +	ptrace_restart_syscall(MODIFY_ORIG_A0, &result);
> +
> +	/* The tracer must modify orig_a0 to actually change the tracee's a0 */
> +	EXPECT_EQ(ORIG_A0_AFTER_MODIFIED, result);
> +}
> +
> +TEST_HARNESS_MAIN

Great, this test can reflect issues more concisely compared to the 
previous one. I will update it in the next PATCH.

Thanks a lot!