Re: [PATCH net v3 07/10] net: au1000: restore RX buffer addresses when restarting the MAC
From: Florian Fainelli
Date: Fri Nov 20 2026 - 19:22:43 EST
On 11/20/26 02:32, netdev-bot+sashiko@xxxxxxxxxx wrote:
Thank you for your contribution! Sashiko AI review found 1 potential
issue(s) to consider.
Critical: 0 · High: 0 · Medium: 1 · Low: 0
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 patch puts each RX descriptor's buffer address back to
rx_db_inuse[i]->dma_addr only in au1000_init()…
--
net: au1000: restore RX buffer addresses when restarting the MAC
au1000_init() now rewrites each RX descriptor with its buffer's start
address from rx_db_inuse[] before setting RX_DMA_ENABLE. Before this,
it only set the enable bit on whatever address the DMA engine had left
behind. That could let a receive after close/open start partway through
its buffer.
diff --git a/drivers/net/ethernet/amd/au1000_eth.c b/drivers/net/ethernet/amd/au1000_eth.c
index 0e9226ea05e18..cb2f3333ba9fd 100644
--- a/drivers/net/ethernet/amd/au1000_eth.c
+++ b/drivers/net/ethernet/amd/au1000_eth.c
@@ -724,7 +724,8 @@ static int au1000_init(struct net_device *dev)
for (i = 0; i < NUM_RX_DMA; i++)
- aup->rx_dma_ring[i]->buff_stat |= RX_DMA_ENABLE;
+ writel(lower_32_bits(aup->rx_db_inuse[i]->dma_addr) |
+ RX_DMA_ENABLE, &aup->rx_dma_ring[i]->buff_stat);
[Severity: Medium]
Is au1000_init() the only place the receiver gets restarted?
au1000_adjust_link() also stops and restarts RX on every duplex change,
and it does not go through au1000_init() or au1000_reset_mac_unlocked():
drivers/net/ethernet/amd/au1000_eth.c:au1000_adjust_link() {
...
if (phydev->link && (aup->old_duplex != phydev->duplex)) {
/* duplex mode changed */
/* switching duplex mode requires to disable rx and tx! */
au1000_hard_stop(dev);
...
au1000_enable_rx_tx(dev);
...
}
On link down, old_duplex is set to -1, so this path runs on every link-up
after a link-down. au1000_hard_stop() only clears MAC_RX_ENABLE and
MAC_TX_ENABLE in mac->control. It leaves rx_dma_ring[i]->buff_stat alone.
Suppose clearing RE stops a frame partway through. The descriptor would
keep RX_DMA_ENABLE with RX_T_DONE still clear. au1000_rx() only reloads
the address inside its while (buff_stat & RX_T_DONE) loop, so it would
not reset this descriptor:
prxd->buff_stat = lower_32_bits(pDB->dma_addr) | RX_DMA_ENABLE;
Could the next frame then start at the advanced address, partway into
the MAX_BUF_SIZE buffer? The free list is LIFO, so rx_db_inuse[0] at
probe is the last slot of the dma_alloc_coherent() region. An overrun
from that slot would land outside the allocation.
A later patch in this series, "net: au1000: keep the MAC stopped until
the PHY link is ready", adds another au1000_hard_stop() on link-down in
au1000_adjust_link(). Each link flap would then stop RX twice, and the
restart on link-up still skips the reload.
This depends on what the MAC does when RE is cleared mid-frame. If it
always finishes the frame and sets RX_T_DONE, au1000_rx() covers it.
Nothing in the driver documents that behaviour.
Would it make sense to put the descriptor reload in a helper that both
au1000_init() and the au1000_adjust_link() stop/start sequence call?
wmb(); /* drain writebuffer */
Thanks for pointing out the additional restart path. I checked it on MTX-1 using a temporary instrumented kernel, switching duplex modes under incoming traffic and recording RX descriptor state before clearing RE and after the existing 10 ms wait in au1000_hard_stop().
Across 11 stops and 44 descriptor snapshots, nine descriptors had advanced addresses with EN set and DN clear before stopping. All nine had DN set after the wait. No descriptor retained an advanced address with DN clear afterward. Those completed descriptors can therefore be processed and rearmed by au1000_rx().
The Au1500 databook, section 7.4.4.1, page 134 (https://assets.sourcengine.com/datasheets/e176e7fe-d2b8-4143-8002-16670131c1ce.pdf#page=134) also specifies that an enabled RX address register must not be rewritten until DN is set. Moving an unconditional reload into the link-change path would need a demonstrated safe DMA quiesce point.
I have not reproduced the proposed failure on the tested hardware, so I am leaving this patch scoped to restoring addresses after a full MAC reset. This finite test does not establish the behavior of every silicon revision or every link-loss timing, but it provides evidence that the in-flight receive completes during the existing stop delay on MTX-1.
pw-bot: cr
--
Florian