| Message ID | 1559228799-84473-1-git-send-email-liucheng32@huawei.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | Stacktrace in ARM32 architecture has jumped the first 2 layers, which may ignore the display of save_stack_trace_tsk. | expand |
On Thu, May 30, 2019 at 11:06:39PM +0800, l00383200 wrote: > Without optimization, both save_stack_trace_tsk and __save_stack_trace > will have stacktrace information in ARM32. > > In this situation, "data.skip += 2" operation will skip the first two layers, > which may make the stacktrace strange and different from other architectures. > > A simple example is as follows: > In ARM32 architecture: > [<ffffff80083cb3f8>] proc_pid_stack+0xac/0x12c > [<ffffff80083c7c70>] proc_single_show+0x5c/0xa8 > [<ffffff800838aca8>] seq_read+0x130/0x420 > [<ffffff8008365c54>] __vfs_read+0x60/0x11c > [<ffffff80083665dc>] vfs_read+0x8c/0x140 > [<ffffff800836717c>] SyS_read+0x6c/0xcc > [<ffffff8008202cb8>] __sys_trace_return+0x0/0x4 > [<ffffffffffffffff>] 0xffffffffffffffff > > In some other architectures(ARM64): > [<ffffff8008209be0>] save_stack_trace_tsk+0x0/0xf0 > [<ffffff80083cb3f8>] proc_pid_stack+0xac/0x12c > [<ffffff80083c7c70>] proc_single_show+0x5c/0xa8 > [<ffffff800838aca8>] seq_read+0x130/0x420 > [<ffffff8008365c54>] __vfs_read+0x60/0x11c > [<ffffff80083665dc>] vfs_read+0x8c/0x140 > [<ffffff800836717c>] SyS_read+0x6c/0xcc > [<ffffff8008202cb8>] __sys_trace_return+0x0/0x4 > [<ffffffffffffffff>] 0xffffffffffffffff > > Therefore, we'd better just jump only one layer to ensure accuracy and consistency. Why do we want to log the function we called to save the stack trace _in_ the stack trace? What useful purpose does it serve? I've always taken the attitude that if we want a stack trace from a certain point in the function, then that's the point that the stack trace should start. It's entirely sensible. > > Signed-off-by: liucheng <liucheng32@huawei.com> > --- > arch/arm/kernel/stacktrace.c | 2 +- > 1 file changed, 1 insertion(+), 1 deletion(-) > > diff --git a/arch/arm/kernel/stacktrace.c b/arch/arm/kernel/stacktrace.c > index 71778bb..bb3da38 100644 > --- a/arch/arm/kernel/stacktrace.c > +++ b/arch/arm/kernel/stacktrace.c > @@ -125,7 +125,7 @@ static noinline void __save_stack_trace(struct task_struct *tsk, > #endif > } else { > /* We don't want this function nor the caller */ > - data.skip += 2; > + data.skip += 1; > frame.fp = (unsigned long)__builtin_frame_address(0); > frame.sp = current_stack_pointer; > frame.lr = (unsigned long)__builtin_return_address(0); > -- > 1.8.5.6 > >
On Thu, May 30, 2019 at 05:22:19PM +0100, Russell King - ARM Linux admin wrote: > On Thu, May 30, 2019 at 11:06:39PM +0800, l00383200 wrote: > > Without optimization, both save_stack_trace_tsk and __save_stack_trace > > will have stacktrace information in ARM32. > > > > In this situation, "data.skip += 2" operation will skip the first two layers, > > which may make the stacktrace strange and different from other architectures. > > > > A simple example is as follows: > > In ARM32 architecture: > > [<ffffff80083cb3f8>] proc_pid_stack+0xac/0x12c > > [<ffffff80083c7c70>] proc_single_show+0x5c/0xa8 > > [<ffffff800838aca8>] seq_read+0x130/0x420 > > [<ffffff8008365c54>] __vfs_read+0x60/0x11c > > [<ffffff80083665dc>] vfs_read+0x8c/0x140 > > [<ffffff800836717c>] SyS_read+0x6c/0xcc > > [<ffffff8008202cb8>] __sys_trace_return+0x0/0x4 > > [<ffffffffffffffff>] 0xffffffffffffffff > > > > In some other architectures(ARM64): > > [<ffffff8008209be0>] save_stack_trace_tsk+0x0/0xf0 > > [<ffffff80083cb3f8>] proc_pid_stack+0xac/0x12c > > [<ffffff80083c7c70>] proc_single_show+0x5c/0xa8 > > [<ffffff800838aca8>] seq_read+0x130/0x420 > > [<ffffff8008365c54>] __vfs_read+0x60/0x11c > > [<ffffff80083665dc>] vfs_read+0x8c/0x140 > > [<ffffff800836717c>] SyS_read+0x6c/0xcc > > [<ffffff8008202cb8>] __sys_trace_return+0x0/0x4 > > [<ffffffffffffffff>] 0xffffffffffffffff > > > > Therefore, we'd better just jump only one layer to ensure accuracy and consistency. > > Why do we want to log the function we called to save the stack trace > _in_ the stack trace? What useful purpose does it serve? > > I've always taken the attitude that if we want a stack trace from a > certain point in the function, then that's the point that the stack > trace should start. It's entirely sensible. Agreed, also the .skip interface sucks and is slated for replacement (whenever we get around to it).
diff --git a/arch/arm/kernel/stacktrace.c b/arch/arm/kernel/stacktrace.c index 71778bb..bb3da38 100644 --- a/arch/arm/kernel/stacktrace.c +++ b/arch/arm/kernel/stacktrace.c @@ -125,7 +125,7 @@ static noinline void __save_stack_trace(struct task_struct *tsk, #endif } else { /* We don't want this function nor the caller */ - data.skip += 2; + data.skip += 1; frame.fp = (unsigned long)__builtin_frame_address(0); frame.sp = current_stack_pointer; frame.lr = (unsigned long)__builtin_return_address(0);
Without optimization, both save_stack_trace_tsk and __save_stack_trace will have stacktrace information in ARM32. In this situation, "data.skip += 2" operation will skip the first two layers, which may make the stacktrace strange and different from other architectures. A simple example is as follows: In ARM32 architecture: [<ffffff80083cb3f8>] proc_pid_stack+0xac/0x12c [<ffffff80083c7c70>] proc_single_show+0x5c/0xa8 [<ffffff800838aca8>] seq_read+0x130/0x420 [<ffffff8008365c54>] __vfs_read+0x60/0x11c [<ffffff80083665dc>] vfs_read+0x8c/0x140 [<ffffff800836717c>] SyS_read+0x6c/0xcc [<ffffff8008202cb8>] __sys_trace_return+0x0/0x4 [<ffffffffffffffff>] 0xffffffffffffffff In some other architectures(ARM64): [<ffffff8008209be0>] save_stack_trace_tsk+0x0/0xf0 [<ffffff80083cb3f8>] proc_pid_stack+0xac/0x12c [<ffffff80083c7c70>] proc_single_show+0x5c/0xa8 [<ffffff800838aca8>] seq_read+0x130/0x420 [<ffffff8008365c54>] __vfs_read+0x60/0x11c [<ffffff80083665dc>] vfs_read+0x8c/0x140 [<ffffff800836717c>] SyS_read+0x6c/0xcc [<ffffff8008202cb8>] __sys_trace_return+0x0/0x4 [<ffffffffffffffff>] 0xffffffffffffffff Therefore, we'd better just jump only one layer to ensure accuracy and consistency. Signed-off-by: liucheng <liucheng32@huawei.com> --- arch/arm/kernel/stacktrace.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)