On 6/15/23 12:17, Alexei Starovoitov wrote:
On Wed, Jun 14, 2023 at 1:32 PM Barret Rhoden <brho@xxxxxxxxxx> wrote:
Hi -
Would it be possible to add logic to the verifier to handle calling
functions within my program (subprograms?) while holding a bpf_spin_lock?
Some of my functions are large enough that the compiler won't inline
them, so I'll get a BPF_CALL to PC + offset (relative call within my
program). Whenever this pops up, I force the compiler to inline the
function, but that's brittle. I'd rather just have the ability to call
a function.
always_inline works as a workaround, right?
And it's guaranteed to work, no?
I'm not sure why you're saying it's brittle.
yeah, it works. the brittleness comes when i don't mark a function
always_inline, but i don't notice since the compiler was inlining it
anyways. but eventually i make a change and the compiler decides to
not-inline it. e.g. i call the same function again somewhere else in my
program, and now it's worth it to make it a separate function.
it's not super urgent - and i've been hit by it enough times that i can
usually find the problem if it pops up.
It probably generates less performant code,
so it would be good to add such support.
It wasn't done earlier, because spin_lock-ed section
supposed to be short. So the restriction was kinda forcing
program authors to minimize the lock time.
Could you please share the example code where you want to use it?
stuff like this:
https://github.com/google/ghost-userspace/blob/main/third_party/bpf/biff_flux.bpf.c#L115
similar to that one, i have an "AVL tree insert" function that the
compiler didn't want to inline - especially if i called it twice. (the
AVL code hasn't hit our opensource ghost repo yet).
Just to make sure we're talking about calling bpf subprograms only
and you're not requesting to call arbitrary helpers and kfuncs
while holding the lock.
Some of the kfuncs can be allowed under lock if there is a real need.
i was talking about bpf subprogs. though one helper that would be nice
to call while holding a lock is bpf_loop. i've got some loops that i'd
turn into bpf-loops, but can't due to the spinlock.
thanks,
barret
