[ Upstream commit f99c9cd9c4 ]
The mt7530_{r,w}32 operation over MDIO uses 3 mdiobus operations and
does not hold a lock, which causes a race condition when multiple
threads try to access a register, they may get unexpected results.
To avoid this, handle the MDIO lock manually, and use the unlocked
__mdiobus_{read,write} in the critical section.
This fixes the "Ghost VLAN" artifact[1] in MT7530/7621 when the VLAN
operation and the swconfig LED link status poll race between each other.
[1] https://forum.openwrt.org/t/mysterious-vlan-ids-on-mt7621-device/64495
Signed-off-by: DENG Qingfang <dqfext@gmail.com>
(cherry picked from commit f99c9cd9c4)
Signed-off-by: Tianling Shen <cnsztl@immortalwrt.org>
The mt7530_{r,w}32 operation over MDIO uses 3 mdiobus operations and
does not hold a lock, which causes a race condition when multiple
threads try to access a register, they may get unexpected results.
To avoid this, handle the MDIO lock manually, and use the unlocked
__mdiobus_{read,write} in the critical section.
This fixes the "Ghost VLAN" artifact[1] in MT7530/7621 when the VLAN
operation and the swconfig LED link status poll race between each other.
[1] https://forum.openwrt.org/t/mysterious-vlan-ids-on-mt7621-device/64495
Signed-off-by: DENG Qingfang <dqfext@gmail.com>
Stability of this Ethernet driver has been a long-standing issue, with
many people reporting frequent "transmit queue timeouts" and even
occasional crashes.
Disabling TSO in the driver helps with stability, although it is likely a
workaround and might not fix the issue completely.
There is a slight slowdown in forwarding performance for TCP packets
(75 kpps vs. 80 kpps with comparable CPU utilization), but this is still
enough to forward close to 1 Gbit/s of full-sized packets across multiple
flows.
Master is using a different ethernet driver, so this is not a backport.
Because of this different driver, the upcoming 21.02 release does not seem
to be affected by these stability issues.
Thanks to mrakotiq for the initial patch.
Fixes: FS#2628
Signed-off-by: Baptiste Jonglez <git@bitsofnetworks.org>
Signed-off-by: Tianling Shen <cnsztl@project-openwrt.eu.org>
The current code acknowledged interrupts *after* polling.
This is the wrong way around, and could cause an interrupt to
be missed.
This is not likely to be fatal as another packet, and so another
interrupt, should come along soon. But maybe it is causing
problems, so let's fix it anyway.
Signed-off-by: NeilBrown <neil@brown.name>
(Note that this matches the upstream driver.)
Signed-off-by: Rosen Penev <rosenp@gmail.com>
* revert: ramips: gsw_mt7621: disable PORT 5 MAC RX/TX flow control
revert: ramips: gsw_mt7621: disable PORT 5 MAC RX/TX flow control by default
This revert c8f8e59
The TX/RX flow control is not the cause of the TX timeouts issue
Signed-off-by: Chen Minqiang <ptpt52@gmail.com>
* ramips: net/mediatek fix logical error
ramips: net/mediatek fix logical error
fe_empty_txd() should return `tx_ring_size - 1` on ring empty, and
return 0 on ring full.
* ramips: net/mediatek disable eee
ramips: net/mediatek disable eee
This disable eee for mt7530 ports, it causes the link down/up
issue, which happens when connecting to 100Mbit switch
Fixes: FS#1449
Signed-off-by: Chen Minqiang <ptpt52@gmail.com>
Since commit f8c55dc ("MIPS: use generic dma noncoherent ops for
simple noncoherent platforms") changed MIPS dma handling, the mmc
driver fails because it doesn't have a dma mask is set.
So set the correct dma mask.
Signed-off-by: NeilBrown <neil@brown.name>
Signed-off-by: DENG Qingfang <dengqf6@mail2.sysu.edu.cn>
Co-authored-by: DENG Qingfang <dengqf6@mail2.sysu.edu.cn>
This commit sync ramips source code from openwrt master.
1. ramips: gsw_mt7621: disable PORT 5 MAC RX/TX flow control by default
2. ramips: allow to set switchdev by board in ramips_set_preinit_iface
3. kernel: fix typos in KernelPackage description
