[resending because kernel.org seems to have mangled my SMTP
credentials.  I wonder if this is a common problem.]

On Thu, May 11, 2017 at 4:44 PM, Linus Torvalds
[off-list ref] wrote:

On Thu, May 11, 2017 at 4:17 PM, Thomas Garnier [off-list ref] wrote:

quoted

Ingo: Do you want the change as-is? Would you like it to be optional?
What do you think?

I'm not ingo, but I don't like that patch. It's in the wrong place -
that system call return code is too timing-critical to add address
limit checks.

Now what I think you *could* do is:

 - make "set_fs()" actually set a work flag in the current thread flags

 - do the test in the slow-path (syscall_return_slowpath).

Yes, yes, that ends up being architecture-specific, but it's fairly simple.

And it only slows down the system calls that actually use "set_fs()".
Sure, it will slow those down a fair amount, but they are hopefully a
small subset of all cases.

How does that sound to people?  Thats' where we currently do that

        if (IS_ENABLED(CONFIG_PROVE_LOCKING) &&
            WARN(irqs_disabled(), "syscall %ld left IRQs disabled",
regs->orig_ax))
                local_irq_enable();

check too, which is a fairly similar issue.

I like this.  It wouldn't help the problem that I suspect is a major
part of the motivation for this patch: a stack overflow could
overwrite addr_limit.  But we fixed that for real already.

Slightly off-topic: I would *love* to see syscall_return_slowpath() or
similar moved or at least mostly moved into generic code.  Aside from
the fact that it used to be written in asm, there's nothing
fundamentally arch-specific about it.

And it only slows down the system calls that actually use "set_fs()".
Sure, it will slow those down a fair amount, but they are hopefully a
small subset of all cases.

It won't even slow them down that much.  The slow path is reasonably
fast these days.