> Thank you for your contribution! Sashiko AI review found 5 potential > issue(s) to consider. > > Critical: 0 · High: 0 · Medium: 1 · Low: 4 > > If you believe none of the issues are legitimate, please reply to each > issue of Medium severity or higher and explain why the AI is wrong. > If any of them are legitimate and you plan to respin, please let patchwork > know by including "pw-bot: cr" as a separate line at the end of your reply > (one such reply per series is enough). > > - [Medium] The commit message says a TX queue passed to > netdev_get_tx_queue() must be below dev->real_num_tx_queues while… > - [Low] The commit message's reason for the change is technically wrong. > - [Low] stmmac_xdp_get_tx_queue() now bounds the queue index by > priv->xmit_qdisc.num_tx_queues and assumes it equals… > - [Low] stmmac_xdp_get_tx_queue() (stmmac_main.c:5411) reads > priv->xmit_qdisc.num_tx_queues with a plain load from NAPI softirq and… > - [Low] For XDP_TX, the TX queue is computed twice per RX NAPI poll: - > stmmac_xdp_xmit_back() (stmmac_main.c:5436) calls… I will take care of them in v2. Regards, Lorenzo > > -- > > net: stmmac: fix num tx queues in stmmac_xdp_get_tx_queue() > > This commit makes stmmac_xdp_get_tx_queue() wrap the CPU index by > priv->xmit_qdisc.num_tx_queues instead of plat->tx_queues_to_use. The goal > is to keep XDP TX inside the real TX queue set when mqprio offload has > reduced real_num_tx_queues. > > > The TX queue returned by stmmac_xdp_get_tx_queue() is passed to > > netdev_get_tx_queue(), so it must be lower than dev->real_num_tx_queues. > > [Severity: Low] > Is this reasoning accurate? netdev_get_tx_queue() only requires the index > to be below num_tx_queues, not real_num_tx_queues: > > include/linux/netdevice.h:netdev_get_tx_queue() { > DEBUG_NET_WARN_ON_ONCE(index >= dev->num_tx_queues); > return &dev->_tx[index]; > } > > The old index always had a DMA ring behind it as well, because > tc_setup_dwmac510_mqprio() caps the mqprio queue range at > tx_queues_to_use: > > if (qopt->offset[i] + qopt->count[i] > > priv->plat->tx_queues_to_use) { > > As far as I can tell, what users see with the old code is XDP frames > bypassing the mqprio/FPE TC-to-queue mapping. In channel mode they can land > on unmapped queues, which tc_mqprio_config_queue_prio() sets to prio 0 > under strict priority. > > Could the commit message describe that impact instead, so stable > backporters can judge it? > > > diff --git a/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c b/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c > > index 25894771e9e9f..cf87afd273439 100644 > > --- a/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c > > +++ b/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c > > @@ -5408,13 +5408,14 @@ static int stmmac_xdp_xmit_xdpf(struct stmmac_priv *priv, int queue, > > static int stmmac_xdp_get_tx_queue(struct stmmac_priv *priv, > > int cpu) > > { > > + unsigned int num_tx_queues = priv->xmit_qdisc.num_tx_queues; > > [Severity: Low] > Can this value get out of sync with dev->real_num_tx_queues while the tc > mqprio configuration is being changed? > > In tc_setup_dwmac510_mqprio(), real_num_tx_queues is lowered (with > synchronize_net()) before the shadow copy is updated: > > err = netif_set_real_num_tx_queues(ndev, num_tx_queues); > ... > tc_mqprio_config_queue_prio(priv, qopt); > > priv->xmit_qdisc.num_tx_queues = num_tx_queues; > > No grace period follows the store. XDP_TX in NAPI and stmmac_xdp_xmit() > can therefore still pick a queue that was just removed from the real set, > and call netdev_get_tx_queue()/__netif_tx_lock() on it. > > stmmac_reset_tc_mqprio() does the updates in the opposite order, and it > ignores the return value of the grow: > > priv->xmit_qdisc.num_tx_queues = priv->plat->tx_queues_to_use; > priv->xmit_qdisc.enabled = false; > > netdev_reset_tc(ndev); > netif_set_real_num_tx_queues(ndev, priv->plat->tx_queues_to_use); > > If netif_set_real_num_tx_queues() fails here, the shadow stays above > real_num_tx_queues until the next stmmac_hw_setup(). That is the mismatch > this patch is trying to avoid. > > Would it be simpler to read READ_ONCE(priv->dev->real_num_tx_queues) > directly, since that follows the core's publish-then-synchronize_net() > ordering? > > [Severity: Low] > This is a plain load from NAPI softirq and ndo_xdp_xmit context. > tc_setup_dwmac510_mqprio() and stmmac_reset_tc_mqprio() write the field > with plain stores under RTNL while the device is up. Will KCSAN report > this as a data race? > > This path already uses READ_ONCE() for other values that can be tuned at > runtime, for example in stmmac_xdp_xmit_xdpf(): > > tx_coal = READ_ONCE(priv->tx_coal_frames[queue]); > > Without an annotation, the compiler may re-load the field inside the loop > below. Suppose a re-load during stmmac_reset_tc_mqprio() returns a larger > value than the one used in the loop test. Then index goes negative, and > the mixed int/unsigned comparison treats it as a very large value. > > I have no evidence that any compiler actually does this here. Would it > still be worth adding READ_ONCE() here and WRITE_ONCE() at both writers? > > Before this patch the divisor was plat->tx_queues_to_use. That value only > changes in stmmac_reinit_queues(), while the device is released and NAPI > is deleted. > > > int index = cpu; > > > > if (unlikely(index < 0)) > > index = 0; > > > > - while (index >= priv->plat->tx_queues_to_use) > > - index -= priv->plat->tx_queues_to_use; > > + while (index >= num_tx_queues) > > + index -= num_tx_queues; > > > > return index; > > } > > [Severity: Low] > The divisor can now change at runtime. Can the two calls to this helper > in one RX NAPI poll then return different queues for XDP_TX? > > stmmac_xdp_xmit_back() picks the queue for each frame, posts descriptors > and flushes the tail pointer. It does not arm the coalescing timer: > > queue = stmmac_xdp_get_tx_queue(priv, cpu); > nq = netdev_get_tx_queue(priv->dev, queue); > ... > stmmac_flush_tx_descriptors(priv, queue); > > stmmac_finalize_xdp_rx() then computes the queue again at the end of > stmmac_rx()/stmmac_rx_zc(), and arms the timer on that result: > > queue = stmmac_xdp_get_tx_queue(priv, cpu); > > if (xdp_status & STMMAC_XDP_TX) > stmmac_tx_timer_arm(priv, queue); > > If tc_setup_dwmac510_mqprio() changes xmit_qdisc.num_tx_queues between the > two calls, the frames sit on ring A while the timer is armed on ring B. > > stmmac_xdp_xmit_xdpf() only sets IC periodically: > > if (tx_coal && (tx_q->tx_count_frames % tx_coal) < num_frames) { > > So up to tx_coal - 1 descriptors on ring A may get neither a completion > interrupt nor a timer to schedule stmmac_tx_clean(). > > After a shrink, ring A can be outside the new real set and may not be used > again. In that case the page_pool pages and DMA buffers stay held until > ifdown, or until the queue is used again. > > stmmac_xdp_xmit() arms the timer on the same queue it flushes, so only > the split XDP_TX path appears to be affected. > > [Severity: Medium] > The commit message says the TX queue must be lower than > real_num_tx_queues. Does the AF_XDP zero-copy TX path still violate that? > > stmmac_tx_clean() calls stmmac_xdp_xmit_zc() for any queue whose > tx_q->xsk_pool is set, and stmmac_xdp_xmit_zc() uses the queue id > directly: > > struct netdev_queue *nq = netdev_get_tx_queue(priv->dev, queue); > struct stmmac_tx_queue *tx_q = &priv->dma_conf.tx_queue[queue]; > > The queue id is only checked against plat->tx_queues_to_use, in both > stmmac_xdp_enable_pool() and stmmac_xsk_wakeup(): > > if (queue >= priv->plat->rx_queues_to_use || > queue >= priv->plat->tx_queues_to_use) > return -EINVAL; > > The core's xsk_reg_pool_at_qid() accepts any queue_id below > max(real_num_rx_queues, real_num_tx_queues). stmmac_hw_setup() keeps > real_num_rx_queues at rx_queues_to_use. > > For example, take tx = rx = 4 and mqprio with 2 TCs on queues 0 and 1. > real_num_tx_queues is then 2. An AF_XDP zero-copy socket bound to queue 3 > still transmits on TX queue 3, outside the mqprio TC set. > > tc_setup_dwmac510_mqprio() does not check af_xdp_zc_qps either, so > shrinking the queue set after the bind gives the same result. > > Should stmmac_xdp_enable_pool(), stmmac_xsk_wakeup() and the mqprio setup > also check against xmit_qdisc.num_tx_queues? > > -- > Sashiko AI review · https://netdev-ai.bots.linux.dev/sashiko/#/patchset/20261117-stmmac-fix-xdp_get_tx_queue-v1-1-4cd74ee7a76c%40oss.qualcomm.com