Re: [PATCH v3 0/5] MIPS: Add per-cpu IRQ stack

[Jason A. Donenfeld]

Hi James,

On Fri, Jan 13, 2017 at 10:49 AM, James Hogan <james.hogan@xxxxxxxxxx> wrote:
> Its quite a significant change/feature, especially in terms of potential
> for further breakage. I don't think its really stable material to be
> honest.  It sounds bad if the kernel stack requirement can be made
> arbitrarily large by stacking too many drivers.

Indeed I believe this is the case. If, say, a kthread is already using
a bit of stack, and then a softirq chain of stacked virtual network
drivers is called, the stack can be busted.

> Is there a simpler fix/workaround for the issue that would satisfy
> stable kernel users until they can upgrade to a kernel with irqstacks?

The simplest solution is probably just not stacking tons of network
drivers. For my own out-of-tree curve25519-donna code that's in
OpenWRT and uses a fair amount of stack, I just kmalloc on MIPS but
not on x86, so in terms of my own stuff there's already a workaround
in place. But this still doesn't solve things for users who have some
interesting networking requirements and stack a few drivers.

Unfortunately, most folks are only testing stuff on ARM and x86, which
already have the separate IRQ stacks, so they aren't hitting crashes.

So, in the end, I'm not quite sure. On the one hand, this fixes an
actual problem and it'd be nice to see stable kernels have the fix. On
the other hand, this is a rather big change. I don't know how to
assess it, but I've copied Greg on this email, who certainly has
better judgement about this than me.

Jason