if there is any new tx need to clean up. do it in next napi poll.
collect tx related members to fe_tx_ring struct. for better
cache usage and more readable.
Signed-off-by: michael lee <igvtee@gmail.com>
SVN-Revision: 45895
when open device. first ready napi software rx.
then enable hardware interrupt.
final start software tx queue to send data.
Signed-off-by: michael lee <igvtee@gmail.com>
SVN-Revision: 45894
the old FE_INT_STATUS register becomes two registers.
FE_INT_STATUS and INT_STATUS. so the hw status almost full
must change to read from FE_INT_STATUS register.
tx/rx done read from INT_STATUS register.
mt7620 datasheet define CNT_GDM1_AF at BIT(29).
but after test it should be BIT(13). why?
Signed-off-by: michael lee <igvtee@gmail.com>
SVN-Revision: 44371
- budget is decremented with completed frames, so don't check if done is
smaller
- ACK the interrupt before processing further frames to fix a small race
condition.
Signed-off-by: Felix Fietkau <nbd@openwrt.org>
SVN-Revision: 44234
The mac address usually write in factory block. but sometime user erase this block , the mac address will change to ff:ff:ff:ff:ff:ff.
This patch is purpose to fix this issue.
Signed-off-by: wengbj <linux.c@foxmail.com>
SVN-Revision: 44166
* fix TSO features verify on mt7621 firewrt board
* improve tx clean up. no need to access uncached
memory. also use TX_DTX register instead of
read tx ring DONE bit
* mt7621 need napi weight 64 to get more performance
* remove netif_receive_skb, after kernel version
3.7 tcp4_gro_receive can handle tcp checksum.
on rt2880 use iperf tcp LAN to WAN throughput test.
with gro 135 Mbits/sec. without gro 80.4Mbits/sec.
Signed-off-by: michael lee <igvtee@gmail.com>
SVN-Revision: 44118
on mt7621 don't have tx vlan vid registers.
so set FE_REG_FE_DMA_VID_BASE to 0.
set rx vlan offload register to disable.
Signed-off-by: michael lee <igvtee@gmail.com>
SVN-Revision: 44117
the rx_buf_size now is 1534 when mtu is 1500.
the ethernet frame with vlan tag and FCS is 1522.
so the buffer is enough.
Signed-off-by: michael lee <igvtee@gmail.com>
SVN-Revision: 44116
* use default timeout value
* print more debug ring info
* move timeout reset function to workqueue
Signed-off-by: michael lee <igvtee@gmail.com>
SVN-Revision: 44043
hardware status and tx vlan offload support on all targets
except rt5350. so i modify the IS_ENABLE condition only for
mt7621.
support mt7621 hardware status reference by SDK. but i don't
have mt7621. if not work just set mt7621 FE_REG_FE_COUNTER_BASE
to 0 to let software count.
Signed-off-by: michael lee <igvtee@gmail.com>
SVN-Revision: 43303
hardware tx vlan offload only support max 16 vids
now use add/delete vlan interface to update vlan id table
when duplicate vlan id index detect.
disable hardware tx vlan offload support.
Signed-off-by: michael lee <igvtee@gmail.com>
SVN-Revision: 43301
the SDK does a bit of extra init that we did not do yet when using an external mt7530.
Signed-off-by: John Crispin <blogic@openwrt.org>
SVN-Revision: 43245
r43200 tries to detect if the fixup is needed or not. control the behaviour via
OF instead and disable unused ports.
Signed-off-by: John Crispin <blogic@openwrt.org>
SVN-Revision: 43201
The Kconfig identifier to enable debugging in the driver was different from the
actually used one. Fix that.
Signed-off-by: Paul Fertser <fercerpav@gmail.com>
Signed-off-by: Gabor Juhos <juhosg@openwrt.org>
SVN-Revision: 34332
This should fix the stalled irq problem seen by several people.
This is not the real fix, but rather moves the bug to the un/init patch of the driver.
The real bug still needs to be fixed, but this workaround should be suffcient to make
the ethernet stable.
SVN-Revision: 34177
Add missing andmask to ramips_esw register read for recv_good value.
Without the mask, recv_bad leaks into the recv_good packet count.
Didn't notice the bug before since you don't usually get bad
packets, so I only saw it when I was playing with overlength packets
earlier...
Signed-off-by: Tobias Diedrich <ranma+openwrt@tdiedrich.de>
SVN-Revision: 33322
Use doubletagging to disable ramips_esw vlan by default, it seems
more reliable.
Daniel Golle found an issue where sometimes (possibly only for
RT3352) the default vlan disable method (clearing en_vlan, untag,
doubletag and putting all ports into vlan 0) doesn't work and the
packets get sent out vlan-tagged with vlan 0.
Instead switch to using the doubletagging method (allow doubletagged
packets, put all ports into vlan 0 with untag enabled) by default.
Unless someone figures out a way to really globally disable vlan for
this switch, this seems like the best (most reliable) option.
I did some tests regarding maximum packet size and did not see any
difference between the two methods, both allow for slightly bigger packets
than the ramips_main.c ethernet driver (ping stops going through
above "ping -s 1472" (1514 bytes), on the switch packets are recv_good until
"ping -s 1490", or about 1532 bytes).
Signed-off-by: Tobias Diedrich <ranma+openwrt@tdiedrich.de>
SVN-Revision: 33321
