On Thu, Feb 04, 2021 at 03:36:15PM +0000, Will Deacon wrote:

On Mon, Feb 01, 2021 at 11:40:11AM -0800, Andrei Vagin wrote:

quoted

We have some ABI weirdness in the way that we handle syscall
exit stops because we indicate whether or not the stop has been
signalled from syscall entry or syscall exit by clobbering a general
purpose register (ip/r12 for AArch32, x7 for AArch64) in the tracee
and restoring its old value after the stop.

This behavior was inherited from ARM and it isn't common for other
architectures. Now, we have PTRACE_GET_SYSCALL_INFO that gives all
required information about system calls, so the hack with clobbering
registers isn't needed anymore.

This change adds the new ptrace option PTRACE_O_ARM64_RAW_REGS.  If it
is set, PTRACE_GETREGSET returns values of all registers without
clobbering r12 or x7 and PTRACE_SETREGSE sets all registers even if a
process has been stopped in syscall-enter or syscall-exit.

Signed-off-by: Andrei Vagin <redacted>
---
 arch/arm64/include/uapi/asm/ptrace.h |  4 ++
 arch/arm64/kernel/ptrace.c           | 70 ++++++++++++++++------------
 include/linux/ptrace.h               |  1 +
 include/uapi/linux/ptrace.h          |  9 +++-
 4 files changed, 52 insertions(+), 32 deletions(-)

Please split this up so that the arm64-specific changes are separate to
the core changes.

ok

quoted

diff --git a/include/uapi/linux/ptrace.h b/include/uapi/linux/ptrace.h
index 83ee45fa634b..bcc8c362ddd9 100644
--- a/include/uapi/linux/ptrace.h
+++ b/include/uapi/linux/ptrace.h

@@ -7,6 +7,7 @@
 /* has the defines to get at the registers. */
 
 #include <linux/types.h>
+#include <asm/ptrace.h>
 
 #define PTRACE_TRACEME		   0
 #define PTRACE_PEEKTEXT		   1

@@ -137,8 +138,14 @@ struct ptrace_syscall_info {
 #define PTRACE_O_EXITKILL		(1 << 20)
 #define PTRACE_O_SUSPEND_SECCOMP	(1 << 21)
 
+/* (1<<28) is reserved for arch specific options. */
+#ifndef _PTRACE_O_ARCH_OPTIONS
+#define _PTRACE_O_ARCH_OPTIONS 0
+#endif

It seems a bit fragile to rely on a comment here to define the user ABI;
why not define _PTRACE_O_ARCH_OPTIONS to the right value unconditionally?

We don't want to allow setting options that are not supported.
_PTRACE_O_ARCH_OPTIONS is added to PTRACE_O_MASK and then
ptrace_setoptions checks whether all specified options is supported or
not.

Also, it seems as though we immediately burn this bit on arm64, so if we
ever wanted another option we'd have to come back here and allocate another
bit. Could we do better, e.g. by calling into an arch hook
(arch_ptrace_setoptions()) and passing the 'addr' parameter?

I am not sure that I understand the idea. Do you suggest to have
PTRACE_O_ARCH_OPTION and pass arch-specific options in addr? In this
case, I think it could be more cleaner to introduce a new ptrace
command. If this is the idea, I think it worth doing this only if we
expect to have more than one,two,three options.

As for my solution, we need to come back to allocate a new bit
to be sure that we don't intersect with non-arch specific options.
And those who add a non-arch option should see that they don't use bits
of arch-specific options.

Let's decide what interface we want to use to solve the problem and then
if we will stop on a ptrace option I will figure out how to improve
this code.

How do other architectures manage this sort of thing? I'm wondering whether
a separate regset containing just "real x7" and orig_x0 would be preferable
after all...

Yeah, it might be a good idea. We will need to do one extra ptrace
system call, but in comparison with ptrace context-switches, this is
nothing.

Dave, Keno, what do you think about this?

Will