Re: bpf: Run devmap xdp_prog on flush instead of bulk enqueue

[Maciej Fijalkowski]

On Thu, May 27, 2021 at 11:43:20AM +0100, Colin Ian King wrote:
> Hi,
> 
> Static analysis with Coverity on linux-next detected a minor issue that
> was introduced with the following commit:
> 
> commit cb261b594b4108668e00f565184c7c221efe0359
> Author: Jesper Dangaard Brouer <brouer@xxxxxxxxxx>
> Date:   Wed May 19 17:07:44 2021 +0800
> 
>     bpf: Run devmap xdp_prog on flush instead of bulk enqueue
> 
> The analysis is as follows:
> 
> 370static void bq_xmit_all(struct xdp_dev_bulk_queue *bq, u32 flags)
> 371{
> 372        struct net_device *dev = bq->dev;
> 373        int sent = 0, drops = 0, err = 0;
> 374        unsigned int cnt = bq->count;
> 375        int to_send = cnt;
> 376        int i;
> 377
> 378        if (unlikely(!cnt))
> 379                return;
> 380
> 381        for (i = 0; i < cnt; i++) {
> 382                struct xdp_frame *xdpf = bq->q[i];
> 383
> 384                prefetch(xdpf);
> 385        }
> 386
> 387        if (bq->xdp_prog) {
> 388                to_send = dev_map_bpf_prog_run(bq->xdp_prog, bq->q,
> cnt, dev);
> 389                if (!to_send)
> 390                        goto out;
> 391
>    Unused value (UNUSED_VALUE)
>    assigned_value: Assigning value from cnt - to_send to drops here, but
> that stored value is overwritten before it can be used.
> 
> 392                drops = cnt - to_send;
> 393        }
> 394
> 395        sent = dev->netdev_ops->ndo_xdp_xmit(dev, to_send, bq->q, flags);
> 396        if (sent < 0) {
> 397                /* If ndo_xdp_xmit fails with an errno, no frames have
> 398                 * been xmit'ed.
> 399                 */
> 400                err = sent;
> 401                sent = 0;
> 402        }
> 403
> 404        /* If not all frames have been transmitted, it is our
> 405         * responsibility to free them
> 406         */
> 407        for (i = sent; unlikely(i < to_send); i++)

FWIW at the time that I was suggesting a rewrite of bq_xmit_all we were
using the 'drops' above via:

		for (i = 0; i < cnt - drops; i++) {

So looks like now the calculation at line 392 is actually not needed.

> 408                xdp_return_frame_rx_napi(bq->q[i]);
> 409
> 410out:
> 
>    value_overwrite: Overwriting previous write to drops with value from
> cnt - sent.
> 
> 411        drops = cnt - sent;
> 412        bq->count = 0;
> 413        trace_xdp_devmap_xmit(bq->dev_rx, dev, sent, drops, err);
> 414}
> 
> drops is being calculated twice but the first value is not used. Not
> sure if that was intentional or an oversight.
> 
> Colin