Re: [PATCH 15/23] io_uring: enable BPF to submit SQEs

[Alexei Starovoitov <alexei.starovoitov@gmail.com>]

On Wed, May 19, 2021 at 03:13:26PM +0100, Pavel Begunkov wrote:
>  
> +BPF_CALL_3(io_bpf_queue_sqe, struct io_bpf_ctx *,		bpf_ctx,
> +			     const struct io_uring_sqe *,	sqe,
> +			     u32,				sqe_len)
> +{
> +	struct io_ring_ctx *ctx = bpf_ctx->ctx;
> +	struct io_kiocb *req;
> +
> +	if (sqe_len != sizeof(struct io_uring_sqe))
> +		return -EINVAL;
> +
> +	req = io_alloc_req(ctx);

that is GFP_KERNEL allocation.
It's only allowed from sleepable bpf progs and further down
there is a correct check for it, so all good.
But submitting sqe is a fundemntal io_uring operation,
so what is the use case for non-sleepable?
In other words why bother? Allow sleepable only and simplify the code?

> +	if (unlikely(!req))
> +		return -ENOMEM;
> +	if (!percpu_ref_tryget_many(&ctx->refs, 1)) {
> +		kmem_cache_free(req_cachep, req);
> +		return -EAGAIN;
> +	}
> +	percpu_counter_add(&current->io_uring->inflight, 1);
> +	refcount_add(1, &current->usage);
> +
> +	/* returns number of submitted SQEs or an error */
> +	return !io_submit_sqe(ctx, req, sqe);

A buggy bpf prog will be able to pass junk sizeof(struct io_uring_sqe)
as 'sqe' here.
What kind of validation io_submit_sqe() does to avoid crashing the kernel?

General comments that apply to all patches:
- commit logs are way too terse. Pls expand with details.
- describe new bpf helpers in comments in bpf.h. Just adding them to an enum is not enough.
- selftest/bpf are mandatory for all new bpf features.
- consider bpf_link style of attaching bpf progs. We had enough issues with progs
  that get stuck due to application bugs. Auto-detach saves the day more often than not.