Using the NAPI or netdev frag cache along with other drivers can lead to
32 KiB pages being held for a long time, despite only being used for
very few page fragment.
This can happen if the ethernet driver grabs one or two fragments for rx
ring refill, while other drivers use (and free up) the remaining
fragments. The 32 KiB higher-order page can only be freed once all users
have freed their fragments, which only happens after the rings of all
drivers holding the fragments have wrapped around.
Depending on the traffic patterns, this can waste a lot of memory and
look a lot like a memory leak
Signed-off-by: Felix Fietkau <nbd@nbd.name>
GRO stores packets as fraglist. If they are routed back to the ethernet
device, they need to be re-segmented if the driver does not support
sending fraglists.
Add the missing support for that, along with a missing feature flag that
allows full routed GRO->TSO offload.
Considerably reduces CPU utilization for routing
Signed-off-by: Felix Fietkau <nbd@nbd.name>
Since kernel 4.10 commit 61e84623ace3 ("net: centralize net_device
min/max MTU checking"), the range of mtu is [min_mtu, max_mtu], which
is [68, 1500] by default.
It's necessary to set a max_mtu if a mtu > 1500 is supported.
Signed-off-by: Mathias Kresin <dev@kresin.me>
This reverts commit 8a570921b5.
This seems to have been accidentally reverted. This fixes mt7620 and
mt7628.
Signed-off-by: Rosen Penev <rosenp@gmail.com>
Acked-by: Daniel Golle <daniel@makrotopia.org>
Mostly whitespace cleanups. Some unneeded code was removed.
MMC init was also moved to the probe function as in
6069bdd
The cleanup commits are over 100, making it hard to do them individually.
Tested on GnuBee PC1 with an SD card being used as swap.
Signed-off-by: Rosen Penev <rosenp@gmail.com>
v2: Fixed compile issue with mt7620
In case of error, the function devm_ioremap_resource() returns ERR_PTR()
and never returns NULL. The NULL test in the return value check should
be replaced with IS_ERR().
Fixes: f079b6406348 ("staging: mt7621-eth: add gigabit switch driver (GSW)")
Signed-off-by: Wei Yongjun <weiyongjun1@huawei.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
In case of error, the function of_phy_connect() returns NULL pointer not
ERR_PTR(). The IS_ERR() test in the return value check should be
replaced with NULL test.
Fixes: e3cbf478f846 ("staging: mt7621-eth: add the drivers core files")
Signed-off-by: Wei Yongjun <weiyongjun1@huawei.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
In case of error, the function devm_ioremap_resource() returns ERR_PTR()
and never returns NULL. The NULL test in the return value check should
be replaced with IS_ERR().
Fixes: e3cbf478f846 ("staging: mt7621-eth: add the drivers core files")
Signed-off-by: Wei Yongjun <weiyongjun1@huawei.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
This is an amagalmation of two upstream commits dealing with whitespace
and dead code removal. I'm synching instead of having two separate commits
as they go out of order compared to previous commits here.
Tested on GnuBee PC1.
Signed-off-by: Rosen Penev <rosenp@gmail.com>
Mostly whitespace cleanups. Some unneeded code was removed.
MMC init was also moved to the probe function as in
6069bdd087
The cleanup commits are over 100, making it hard to do them individually.
Tested on GnuBee PC1 with an SD card being used as swap.
Signed-off-by: Rosen Penev <rosenp@gmail.com>
These resources are extracted from devicetree, so they aren't
needed here.
Signed-off-by: NeilBrown <neil@brown.name>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Code currently defines:
and then compiles code only if they are defined.
We might want to disable some of these via devicetree one
day, but for now just remove the #defines and the
conditions - all the code for different ports is
easy to identify.
Signed-off-by: NeilBrown <neil@brown.name>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
As the Interrupts for the PCI adapters are listed in
devicetree we shouldn't need to have them explicit in the code.
The simplest way to do this is to use of_irq_parse_and_map_pci()
and specify an interrupt-map which identifies the different
PCI hosts by bus/slot numbers.
This has the advantage that the hwirq number are mapped to virq
numbers for us, so the ugly hack can go.
Signed-off-by: NeilBrown <neil@brown.name>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Remove the owner assignment form the platform driver as
platform_driver_register() already initializes the owner.
Found using coccinelle.
Signed-off-by: Christian Lütke-Stetzkamp <christian@lkamp.de>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
The stats update needs to protect against being interrupted by a tasklet
The u64 stats seqlock needs to be initialized
Signed-off-by: Felix Fietkau <nbd@nbd.name>
Supports IPv4 flow offloading on MT7621 for Routing, SNAT and DNAT
Supported are regular ethernet->ethernet connections, including one
802.1q VLAN and/or PPPoE encapsulation
Signed-off-by: John Crispin <john@phrozen.org>
Signed-off-by: Felix Fietkau <nbd@nbd.name>
of_dma_configure() sets a default ->dma_mask of
DMA_BIT_MASK(32), claiming devices can DMA from
the full 32bit address space.
The mtk-mmc driver does not support access to
highmem pages, so it is really limited to the
bottom 512M (actually 448M due to 64M of IO space).
Setting ->dma_mask to NULL causes mmc_setup_queue()
to fall-back to using BLK_BOUNCE_HIGH to tell the
block layer to use a bounce-buffer for any highmem
pages requiring IO.
Signed-off-by: NeilBrown <neil@brown.name>
Signed-off-by: Felix Fietkau <nbd@nbd.name>
Our .dts files only have one device defined and this is unlikely to
change, so statically initialize host->id to 0.
Signed-off-by: Felix Fietkau <nbd@nbd.name>
In some boards port 4 of mt7620 switch is powered down in bootloader
configuration as lan/wan package leak mitigation.
Reset ephy port 4 to default settings to power up the port.
Signed-off-by: Pawel Dembicki <paweldembicki@gmail.com>
[clarify issue in commit message]
Signed-off-by: Mathias Kresin <dev@kresin.me>
The hardware emits some interrupts while initializing and handling them
can mess up the state or cause infinite loops.
Fix this by disabling IRQs during init and re-enabling them afterwards
Signed-off-by: Felix Fietkau <nbd@nbd.name>
- removed upstreamed patches
- 0901-spansion_nand_id_fix.patch is disabled, not clear if it's needed
Signed-off-by: Roman Yeryomin <roman@advem.lv>
Signed-off-by: John Crispin <john@phrozen.org>
