Re: [PATCH RFC 1/1] futex: Create set_robust_list2

[André Almeida]

Hi Arnd,

Em 24/10/2024 12:57, Arnd Bergmann escreveu:
> On Thu, Oct 24, 2024, at 14:57, André Almeida wrote:
> > This new syscall allows to set multiple list to the same process. There
> > are two list types: 32 and 64 bit lists.
> >
> > It supports up to 10 lists per process (see ROBUST_LISTS_PER_TASK). The
> > lists are dynamically allocated on demand, as part of a linked list.
> > This is the proposed interface:
> >
> > 	long set_robust_list2(void *head, int index, unsigned int flags)
> >
> > Userspace can ask to set the head of a new list using (index = -1).
> > Kernel will allocate a new list, place in the linked list and return the
> > new index to userspace.
> >
> > Userspace can modify an existing head by using an index >= 0. If the
> > requested list doesn't exist, an error is returned.
> >
> > Userspace cannot remove a robust list.
> >
> > For now, flag is for the list type:
> >
> > 	enum robust_list_type {
> > 	 	ROBUST_LIST_32BIT,
> > 		ROBUST_LIST_64BIT,
> > 	 };
> >
> > Signed-off-by: André Almeida <andrealmeid@xxxxxxxxxx>
>
> Hi André,
>
> I have no opinion on the syscall itself, but I'll comment on
> the way you hook it up:
>
> >   arch/arm/tools/syscall.tbl             |  1 +
> >   arch/x86/entry/syscalls/syscall_64.tbl |  1 +
>
> If we agree on the number, this should be added to all
> architectures at the same time. In particular, when
> you add it to 32-bit arm, it also needs to be in the
> corresponding arch/arm64/tools/syscall_32.tbl for
> compat mode.

Ok

> >   include/uapi/asm-generic/unistd.h      |  5 +-
>
> This reminds me that I need to send the patch to remove this
> file, nothing should use it any more, though we still have
> the copy in tools/include/uapi/asm-generic/unistd.h that
> still gets referenced until the scripts are changed to
> use the syscall.tbl format.
>
> > +	if (unlikely(!list_empty(list2))) {
> > +		list_for_each_entry_safe(curr, n, list2, list) {
> > +			if (curr->head != NULL) {
> > +				if (curr->list_type == ROBUST_LIST_64BIT)
> > +					exit_robust_list(tsk, curr->head);
> > +				else if (curr->list_type == ROBUST_LIST_32BIT)
> > +					compat_exit_robust_list(tsk, curr->head);
> > +				curr->head = NULL;
> > +			}
>
> This looks like the behavior of a 32-bit task using
> ROBUST_LIST_64BIT is different on native 32-bit kernels
> compared to running on compat mode.
>
> Assuming we want them to behave the same way, did you intend
> ROBUST_LIST_64BIT to refer to 64-bit pointers on 32-bit
> tasks, or should they use normal word-size pointers?

Oh right, I haven't covered that indeed. I think I would need to have 
something like:

static void exit_robust_list_64()
static void exit_robust_list_32()

And then each function would use explicit sizes for pointers. Also, I 
would rewrite the conditions to make that every combination of 64/32bit 
kernel/app calls the appropriated function.

Alternatively, we could just disable 32bit kernel/app to use the 
ROBUST_LIST_64BIT option.

Thank you for your feedback!
	André