From: Florian Fainelli <[email protected]>
Date: Tue, 13 Mar 2018 14:45:07 -0700
> There is no need for complex checking between the last consumed index
> and current consumed index, a simple subtraction will do.
>
> This also eliminates the possibility of a permanent transmit queue stall
> under the following conditions:
>
> - one CPU bursts ring->size worth of traffic (up to 256 buffers), to the
> point where we run out of free descriptors, so we stop the transmit
> queue at the end of bcm_sysport_xmit()
>
> - because of our locking, we have the transmit process disable
> interrupts which means we can be blocking the TX reclamation process
>
> - when TX reclamation finally runs, we will be computing the difference
> between ring->c_index (last consumed index by SW) and what the HW
> reports through its register
>
> - this register is masked with (ring->size - 1) = 0xff, which will lead
> to stripping the upper bits of the index (register is 16-bits wide)
>
> - we will be computing last_tx_cn as 0, which means there is no work to
> be done, and we never wake-up the transmit queue, leaving it
> permanently disabled
>
> A practical example is e.g: ring->c_index aka last_c_index = 12, we
> pushed 256 entries, HW consumer index = 268, we mask it with 0xff = 12,
> so last_tx_cn == 0, nothing happens.
>
> Fixes: 80105befdb4b ("net: systemport: add Broadcom SYSTEMPORT Ethernet MAC
> driver")
> Signed-off-by: Florian Fainelli <[email protected]>
Indeed if you mask the indexes too early this can happen.
Applied and queued up for -stable, thanks Florian.
