On Sat, Jan 18, 2025 at 03:13:42PM +0100, Oleg Nesterov wrote:
> On 01/17, Dmitry V. Levin wrote:
[...]
> > For example, on x86_64 sizeof(struct ptrace_syscall_info) is currently 88,
> > while on x86 it is 84.
>
> Not good, but too late to complain...
Actually, I don't think it's too late to add an extra __u32 padding
there since it wouldn't affect PTRACE_GET_SYSCALL_INFO.
I can add an explicit padding to the structure if you say
you like it better this way.
[...]
> Thats said... Can't resist,
>
> > An absolutely artificial example: let's say we're adding an optional
> > 64-bit field "artificial" to ptrace_syscall_info.seccomp, this means
> > sizeof(ptrace_syscall_info) grows by 8 bytes. When userspace wants
> > to set this optional field, it sets a bit in ptrace_syscall_info.flags,
> > this tells the kernel to look into this new "artificial" field.
> > When userspace is not interested in setting new optional fields,
> > it just keeps ptrace_syscall_info.flags == 0. Remember, however, that
> > by adding the new optional field sizeof(ptrace_syscall_info) grew by 8 bytes.
> >
> > What we need is to make sure that an older kernel that has no idea of this
> > new field would still accept the bigger size, so that userspace would be
> > able to continue doing its
> > ptrace(PTRACE_SET_SYSCALL_INFO, pid, sizeof(info), &info)
> > despite of potential growth of sizeof(info) until it actually starts using
> > new optional fields.
>
> This is clear, but personally I don't really like this pattern... Consider
>
> void set_syscall_info(int unlikely_condition)
> {
> struct ptrace_syscall_info info;
>
> fill_info(&info);
> if (unlikely_condition) {
> info.flags = USE_ARTIFICIAL;
> info.artificial = 1;
> }
>
> assert(ptrace(PTRACE_SET_SYSCALL_INFO, sizeof(info), &info) == 0);
> }
>
> Now this application (running on the older kernel) can fail or not, depending
> on "unlikely_condition". To me it would be better to always fail in this case.
In practice, user-space programs rarely have the luxury to assume that
some new kernel API is available. For example, strace still performs a
runtime check for PTRACE_GET_SYSCALL_INFO (introduced more than 5 years
ago) and falls back to pre-PTRACE_GET_SYSCALL_INFO interfaces when the
kernel lacks support. Consequently, user-space programs would have to
keep track of PTRACE_SET_SYSCALL_INFO interfaces supported by the kernel,
so ...
> That is why I tried to suggest to use "user_size" as a version number.
> Currently we have PTRACE_SYSCALL_INFO_SIZE_VER0, when we add the new
> "artificial" member we will have PTRACE_SYSCALL_INFO_SIZE_VER1. Granted,
> this way set_syscall_info() can't use sizeof(info), it should do
>
> ptrace(PTRACE_SET_SYSCALL_INFO, PTRACE_SYSCALL_INFO_SIZE_VER1, info);
>
> and the kernel needs more checks, but this is what I had in mind when I said
> that the 1st version can just require "user_size == PTRACE_SYSCALL_INFO_SIZE_VER0".
... it wouldn't be a big deal for user-space to specify also an
appropriate "user_size", e.g. PTRACE_SYSCALL_INFO_SIZE_VER1 when it starts
using the interface available since VER1, but it wouldn't help user-space
programs either as they would have to update "op" and/or "flags" anyway,
and "user_size" would become just yet another detail they have to care
about.
At the same time, "flags" is needed anyway because the most likely
extension of PTRACE_SET_SYSCALL_INFO would be support of setting some
fields that are present in the structure already, e.g.
instruction_pointer.
--
ldv
