Since we no longer localize channel/CPU affiliation within one NUMA
node, num_online_cpus() is used as the number of channel cap, instead of
the number of processors in a NUMA node.
This patch allows a bigger range for tuning the number of channels.
Signed-off-by: Haiyang Zhang <haiyangz@microsoft.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
If the transmit queue is known full, then don't keep aggregating
data. And the cp_partial flag which indicates that the current
aggregation buffer is full can be folded in to avoid more
conditionals.
Signed-off-by: Stephen Hemminger <sthemmin@microsoft.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
There is only ever a single instance of network device object
referencing the internal rndis object. Therefore the open_cnt atomic
is not necessary.
Signed-off-by: Stephen Hemminger <sthemmin@microsoft.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The netvsc_receive_callback function was using RCU to find the
appropriate underlying netvsc_device. Since calling function already
had that pointer, this was unnecessary.
Signed-off-by: Stephen Hemminger <sthemmin@microsoft.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The caller (netvsc_receive) already has the net device pointer,
and should just pass that to functions rather than the hyperv device.
This eliminates several impossible error paths in the process.
Signed-off-by: Stephen Hemminger <sthemmin@microsoft.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
When skb can not be allocated, update ethtool statisitics
rather than rx_dropped which is intended for netif_receive.
Signed-off-by: Stephen Hemminger <sthemmin@microsoft.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Since only caller does not care about return value.
Signed-off-by: Stephen Hemminger <sthemmin@microsoft.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The values were not computed correctly. There are no significant
visible impact, though.
The intended size of RX buffer is 16 MB, and the default slot size is 1728.
So, NETVSC_DEFAULT_RX should be 16*1024*1024 / 1728 = 9709.
The intended size of TX buffer is 1 MB, and the slot size is 6144.
So, NETVSC_DEFAULT_TX should be 1024*1024 / 6144 = 170.
The patch puts the formula directly into the macro, and moves them to
hyperv_net.h, together with related macros.
Fixes: 5023a6db73 ("netvsc: increase default receive buffer size")
Signed-off-by: Haiyang Zhang <haiyangz@microsoft.com>
Signed-off-by: Stephen Hemminger <sthemmin@microsoft.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The max should be 31 MB on host with NVSP version > 2.
On legacy hosts (NVSP version <=2) only 15 MB receive buffer is allowed,
otherwise the buffer request will be rejected by the host, resulting
vNIC not coming up.
The NVSP version is only available after negotiation. So, we add the
limit checking for legacy hosts in netvsc_init_buf().
Fixes: 5023a6db73 ("netvsc: increase default receive buffer size")
Signed-off-by: Haiyang Zhang <haiyangz@microsoft.com>
Signed-off-by: Stephen Hemminger <sthemmin@microsoft.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The memset of the whole maximum possible RNDIS header is unnecessary.
For the main part of the header use a structure assignment.
No need to memset the whole per packet info. Instead rely on caller to
set what it wants. Also get rid of cast to void and signed/unsigned
conversion. Now return pointer to per packet data (rather than the
header) which simplifies use by code setting up the packet data.
Signed-off-by: Stephen Hemminger <sthemmin@microsoft.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Every packet sent checks the available ring space. The calculation
can be sped up by using reciprocal divide which is multiplication.
Since ring_size can only be configured by module parameter, so it doesn't
have to be passed around everywhere. Also it should be unsigned
since it is number of pages.
Signed-off-by: Stephen Hemminger <sthemmin@microsoft.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Packet alignment is always a power of 2 therefore modulus can
be replaced with a faster and operation
Signed-off-by: Stephen Hemminger <sthemmin@microsoft.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Since skb is always non-NULL in the copy portion of netvsc_send
do not need local variable.
Signed-off-by: Stephen Hemminger <sthemmin@microsoft.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
rndis_filter_device_add() is called both from netvsc_probe() when we
initially create the device and from set channels/mtu/ringparam
routines where we basically remove the device and add it back.
hw_features is reset in rndis_filter_device_add() and filled with
host data. However, we lose all additional flags which are set outside
of the driver, e.g. register_netdevice() adds NETIF_F_SOFT_FEATURES and
many others.
Unfortunately, calls to rndis_{query_hwcaps(), _set_offload_params()}
calls cannot be avoided on every RNDIS reset: host expects us to set
required features explicitly. Moreover, in theory hardware capabilities
can change and we need to reflect the change in hw_features.
Reset net->hw_features bits according to host data in
rndis_netdev_set_hwcaps(), clear corresponding feature bits
from net->features in case some features went missing (will never happen
in real life I guess but let's be consistent).
Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Reviewed-by: Haiyang Zhang <haiyangz@microsoft.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Hyper-V hosts are known to send RNDIS messages even after we halt the
device in rndis_filter_halt_device(). Remove user visible messages
as they are not really useful.
Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
It was found that in some cases host refuses to teardown GPADL for send/
receive buffers (probably when some work with these buffere is scheduled or
ongoing). Change the teardown logic to be:
1) Send NVSP_MSG1_TYPE_REVOKE_* messages
2) Close the channel
3) Teardown GPADLs.
This seems to work reliably.
Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
In some cases, like internal vSwitch, the host doesn't provide
send indirection table updates. This patch sets the table to be
equal weight after subchannels are all open. Otherwise, all workload
will be on one TX channel.
As tested, this patch has largely increased the throughput over
internal vSwitch.
Signed-off-by: Haiyang Zhang <haiyangz@microsoft.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
tx_table is part of the private data of kernel net_device. It is only
zero-ed out when allocating net_device.
We may recreate netvsc_device w/o recreating net_device, so the private
netdev data, including tx_table, are not zeroed. It may contain channel
numbers for the older netvsc_device.
This patch adds initialization of tx_table each time we recreate
netvsc_device.
Signed-off-by: Haiyang Zhang <haiyangz@microsoft.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Rename this variable because it is the Receive indirection
table.
Signed-off-by: Haiyang Zhang <haiyangz@microsoft.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The patch supports the options to switch TCP hash level between
L3 and L4 by ethtool command. TCP over IPv4 and v6 can be set
differently. The default hash level is L4. We currently only
allow switching TX hash level from within the guests.
For example, for TCP over IPv4 on eth0:
To include TCP port numbers in hashing:
ethtool -N eth0 rx-flow-hash tcp4 sdfn
To exclude TCP port numbers in hashing:
ethtool -N eth0 rx-flow-hash tcp4 sd
To show TCP hash level:
ethtool -n eth0 rx-flow-hash tcp4
Signed-off-by: Haiyang Zhang <haiyangz@microsoft.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This simplifies the logic and make it easier to add more
options.
Signed-off-by: Haiyang Zhang <haiyangz@microsoft.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Add extack arg to netdev_upper_dev_link and netdev_master_upper_dev_link
Signed-off-by: David Ahern <dsahern@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Report the numbers of events for stop_queue and wake_queue in
ethtool stats.
Example:
ethtool -S eth0
NIC statistics:
...
stop_queue: 7
wake_queue: 7
...
Signed-off-by: Simon Xiao <sixiao@microsoft.com>
Reviewed-by: Haiyang Zhang <haiyangz@microsoft.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
For older hosts without multi-channel (vRSS) support, and some error
cases, we still need to set the real number of queues to one.
This patch adds this missing setting.
Fixes: 8195b1396e ("hv_netvsc: fix deadlock on hotplug")
Signed-off-by: Haiyang Zhang <haiyangz@microsoft.com>
Reviewed-by: Stephen Hemminger <sthemmin@microsoft.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Don't populate const array ver_list on the stack, instead make it
static. Makes the object code smaller by over 400 bytes:
Before:
text data bss dec hex filename
18444 3168 320 21932 55ac drivers/net/hyperv/netvsc.o
After:
text data bss dec hex filename
17950 3224 320 21494 53f6 drivers/net/hyperv/netvsc.o
(gcc 6.3.0, x86-64)
Signed-off-by: Colin Ian King <colin.king@canonical.com>
Reviewed-by: Haiyang Zhang <haiyangz@microsoft.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
If MTU is changed the host would reject the send buffer change.
This problem is result of recent change to allow changing send
buffer size.
Every time we change the MTU, we store the previous net_device section
count before destroying the buffer, but we don’t store the previous
section size. When we reinitialize the buffer, its size is calculated
by multiplying the previous count and previous size. Since we
continuously increase the MTU, the host returns us a decreasing count
value while the section size is reinitialized to 1728 bytes every
time.
This eventually leads to a condition where the calculated buf_size is
so small that the host rejects it.
Fixes: 8b5327975a ("netvsc: allow controlling send/recv buffer size")
Signed-off-by: Alex Ng <alexng@microsoft.com>
Signed-off-by: Stephen Hemminger <sthemmin@microsoft.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The default receive buffer size was reduced by recent change
to a value which was appropriate for 10G and Windows Server 2016.
But the value is too small for full performance with 40G on Azure.
Increase the default back to maximum supported by host.
Fixes: 8b5327975a ("netvsc: allow controlling send/recv buffer size")
Signed-off-by: Stephen Hemminger <sthemmin@microsoft.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Only need to wakeup the initiator after all sub-channels
are opened.
Signed-off-by: Stephen Hemminger <sthemmin@microsoft.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
When a virtual device is added dynamically (via host console), then
the vmbus sends an offer message for the primary channel. The processing
of this message for networking causes the network device to then
initialize the sub channels.
The problem is that setting up the sub channels needs to wait until
the subsequent subchannel offers have been processed. These offers
come in on the same ring buffer and work queue as where the primary
offer is being processed; leading to a deadlock.
This did not happen in older kernels, because the sub channel waiting
logic was broken (it wasn't really waiting).
The solution is to do the sub channel setup in its own work queue
context that is scheduled by the primary channel setup; and then
happens later.
Fixes: 732e49850c ("netvsc: fix race on sub channel creation")
Reported-by: Dexuan Cui <decui@microsoft.com>
Signed-off-by: Stephen Hemminger <sthemmin@microsoft.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The limit of setting receive indirection table value should be
the current number of channels, not the VRSS_CHANNEL_MAX.
Signed-off-by: Haiyang Zhang <haiyangz@microsoft.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Because of the following code, net->num_tx_queues equals to
VRSS_CHANNEL_MAX, and max_chn is less than or equals to VRSS_CHANNEL_MAX.
netvsc_drv.c:
alloc_etherdev_mq(sizeof(struct net_device_context),
VRSS_CHANNEL_MAX);
rndis_filter.c:
net_device->max_chn = min_t(u32, VRSS_CHANNEL_MAX, num_possible_rss_qs);
So this patch removes the unnecessary limit check before comparing
with "max_chn".
Signed-off-by: Haiyang Zhang <haiyangz@microsoft.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The minus one and assignment to a local variable is not necessary.
This patch simplifies it.
Signed-off-by: Haiyang Zhang <haiyangz@microsoft.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch removes the parameter, num_queue in
rndis_filter_set_rss_param(), which is no longer in use.
Signed-off-by: Haiyang Zhang <haiyangz@microsoft.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
If VF is attached then can still allow netvsc driver module to
be removed. Just have to make sure and do the cleanup.
Also, avoid extra rtnl round trip when calling unregister.
Signed-off-by: Stephen Hemminger <sthemmin@microsoft.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Use one routine for datapath up/down. Don't need to reopen
the rndis layer.
Signed-off-by: Stephen Hemminger <sthemmin@microsoft.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
There is a deadlock possible when canceling the link status
delayed work queue. The removal process is run with RTNL held,
and the link status callback is acquring RTNL.
Resolve the issue by using trylock and rescheduling.
If cancel is in process, that block it from happening.
Fixes: 122a5f6410 ("staging: hv: use delayed_work for netvsc_send_garp()")
Signed-off-by: Stephen Hemminger <sthemmin@microsoft.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
We now remove rndis filter before unregister_netdev(), which calls
device close. It involves closing rndis filter already removed.
This patch fixes this error.
Signed-off-by: Haiyang Zhang <haiyangz@microsoft.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The patch add the functions to switch UDP hash level between
L3 and L4 by ethtool command. UDP over IPv4 and v6 can be set
differently. The default hash level is L4. We currently only
allow switching TX hash level from within the guests.
On Azure, fragmented UDP packets have high loss rate with L4
hashing. Using L3 hashing is recommended in this case.
For example, for UDP over IPv4 on eth0:
To include UDP port numbers in hasing:
ethtool -N eth0 rx-flow-hash udp4 sdfn
To exclude UDP port numbers in hasing:
ethtool -N eth0 rx-flow-hash udp4 sd
To show UDP hash level:
ethtool -n eth0 rx-flow-hash udp4
Signed-off-by: Haiyang Zhang <haiyangz@microsoft.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The only usage of vmbus_sendpacket_ctl was by vmbus_sendpacket.
Signed-off-by: Stephen Hemminger <sthemmin@microsoft.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The function vmbus_sendpacket_pagebuffer_ctl was never used directly.
Just have vmbus_send_pagebuffer
Signed-off-by: Stephen Hemminger <sthemmin@microsoft.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Add ethtool statistics for case where send chimmeny buffer is
exhausted and driver has to fall back to doing scatter/gather
send. Also, add statistic for case where ring buffer is full and
receive completions are delayed.
Signed-off-by: Stephen Hemminger <sthemmin@microsoft.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Control the size of the buffer areas via ethtool ring settings.
They aren't really traditional hardware rings, but host API breaks
receive and send buffer into chunks. The final size of the chunks are
controlled by the host.
The default value of send and receive buffer area for host DMA
is much larger than it needs to be. Experimentation shows that
4M receive and 1M send is sufficient.
Signed-off-by: Stephen Hemminger <sthemmin@microsoft.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The function init_page_array is always called with a valid pointer
to RNDIS header. No check for NULL is needed.
Signed-off-by: Stephen Hemminger <sthemmin@microsoft.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Assignment to a typed pointer is sufficient in C.
No cast is needed.
Signed-off-by: Stephen Hemminger <sthemmin@microsoft.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Haiyang Zhang