Under driver verifier with low resources simulation the
attempt to access NIC configuration may fail, important
fields of the context are not initialized correctly
causing driver verifier BSOD on attempt to allocate
zero bytes from memory. Fail initialization if the NIC
configuration is not accessible.
Signed-off-by: Yuri Benditovich <[email protected]>
Global value of max RX buffers, shared among queues, is
bad idea. It is especially bad when memory allocation
during queue initialization may fail, for example in case
of driver running under driver verifier. Changing the
global field to per-queue field.
Signed-off-by: Yuri Benditovich <[email protected]>
Currently the driver always assume the maximim packet size
is ~64K as it is compiled with RSC support. The initial
reason was that if the RSC is disabled by admin (registry
value of the driver), it can be turned on by command, i.e.
without disable-enable. However this depends also on host
features. If the host does not support the GUEST_TSOx,
RSC will be reported to the OS as not supported but the
buffers are still huge without any reason.
(In such case the powershell reports IPv4 and IPv6 operational
state as Off even if the RSC keywords are enabled).
The commit moves the decision to initialization time - if the
RSC is not supported, the buffers are sized accoring to MTU.
The memory consumption is lower 16 times.
Signed-off-by: Yuri Benditovich <[email protected]>
https://issues.redhat.com/browse/RHEL-19814
Do not allocate redundant memory, use full 4K+ pages when
needed, tail of the data (usually small) allocate together
with header/indirect block and make this block smaller than
4K when possible. This reduces 10% of all the shared memory
allocated for RX buffers. For non-RSC configuration it is
possible to reduce much more (for future).
Signed-off-by: Yuri Benditovich <[email protected]>
Split the decisions about size of shared buffers for RXQ and
allocation of shared memories. Create optimal layout of RX shared
buffers before creation of queues; allocate the memories according
the layout. Initially optimized layout is desabled and we create
the "legacy" layout, fully compatible with what was done before.
This commit contains only creation of the layout, it is not yet
applied when allocating RX buffer's memories.
Signed-off-by: Yuri Benditovich <[email protected]>
Fixes: https://issues.redhat.com/browse/RHEL-32200
Supporting poll mode flows is very complicated in
case the driver acts as protocol driver above VF
device, so the driver will ignore the poll mode setting
in case it works in standby mode. If the VF driver does
not use the poll mode everything will work the same way
as in Win10. If the VF driver uses poll mode it reports
incoming packets from its poll mode handler already and
the NetKVM just indicates the packets up.
Signed-off-by: Yuri Benditovich <[email protected]>
This provides simple way to build the driver with minimal
modification and enable poll mode on any NDIS 6.85+,
for engineering testing purposes.
Signed-off-by: Yuri Benditovich <[email protected]>
This procedure will be called from poll handler, it retrieves
packets from RX virtqueue (if any) and gets NBLs from RSS queue
instead of indicate them to NDIS.
Signed-off-by: Yuri Benditovich <[email protected]>
In existing build the (un)registration is implemented as
a stub procedure. To start function we need the define
PARANDIS_SUPPORT_POLL.
Signed-off-by: Yuri Benditovich <[email protected]>
Implemented a workaround for rx packet loss encountered with TCP_NODELAY. By default,
this feature is turned off to minimize the impact on CPU performance and network bandwidth.
To activate the feature, go to the network card properties and set the MinRxBufferPercent
to a value above 0. A good starting point is to set MinRxBufferPercent to 25% and then monitor
for any packet loss. If packet loss continues, incrementally adjust the value to 50%, then 75%,
and if necessary, up to the maximum of 100%.
In extreme cases, it may be required to not only adjust the MinRxBufferPercent but also to increase
the number of queues and the size of each queue (queue_size).
It is important to note that this feature is only applicable to rx packet loss.
For more details, see [this thread](https://github.com/virtio-win/kvm-guest-drivers-windows/issues/1012).
Signed-off-by: junjiehua <[email protected]>
This commit introduces the MAX_FRAGMENTS_IN_ONE_NB macro, which
defines the maximum number of fragments in a single NET_BUFFER. It
is used to constrain the SGTableCapacity value, ensuring it does
not exceed the meaningful limit set by MAX_FRAGMENTS_IN_ONE_NB.
Signed-off-by: Zhang JianMing <[email protected]>
https://issues.redhat.com/browse/RHEL-20575
Currently the driver get RX descriptor from the rx virtQ,
analyses it (L2/L3), calculates hash and delivers to the
proper RSS queue. Later it extract the packet from the RSS
queue and checks whether it passes RX filters.
Now we move the filtering to the earliest stage - just after
basic analysis.
Note that in case of vhost there are no receive filters
configured by the host and typically the host does not work
in promiscuous mode, so it is redundant to calculate the hash
and trigger the RSS queue for packets that should be dropped.
Signed-off-by: Yuri Benditovich <[email protected]>
Adding statistics for failed TX packets
(ones that can't be mapped using available indirect area) and
minimal available TX descriptors (if one or more TX virtqueues
were completely filled by packets this field will be read as 0)
These statistics fields are printed on disable and also available
at run-time via WMI.
Signed-off-by: Yuri Benditovich <[email protected]>
resolve:https://github.com/virtio-win/kvm-guest-drivers-windows/issues/830
The problem, RxCapacity of value 16 is in the inf file. But the minimum
of defaultConfiguration is 32, meaning that setting RxCapacity to 16 in
the device manager doesn't take effect.
Fix: change the minimum of defaultConfiguration to 16
Signed-off-by: Kfir Manor <[email protected]>
Prepare the code to support NDIS 6.83 if such define will
be provided in a build (currently it is not).
Signed-off-by: Yuri Benditovich <[email protected]>
Win2022 HLK (NICStrictPropertyValidation) requires NumRssQueues
parameter default value to be 16 for 10Gbps adapters with
MSI-X support.
Signed-off-by: Viktor Prutyanov <[email protected]>
NetKVM treats VLAN ID = 4095 as well as IDs from 1-4094, that is frames
with ID = 4095 pass and others don't. According to IEEE 802.1Q, 4095
can't be valid VLAN ID. Also, Win2022 HLK assumes that NDIS accepts
ID <= 4094.
Signed-off-by: Viktor Prutyanov <[email protected]>
Rename IsSendPossible to IsTxRxPossible and use fCurrentLinkState
instead of bConnected. Reuse this procedure in Rx path also, the
condition is the same for both directions.
Signed-off-by: Yuri Benditovich <[email protected]>
Reuse ReadLinkState to obtain the link status from the device
and set properly fCurrentLinkState. When reporting the link
status on driver initialization we can't rely on bConnected
anymore as there is a possibility that we should hide the
link until VF is attached.
Signed-off-by: Yuri Benditovich <[email protected]>
If the device reports standby feature but the protocol is not
installed yet or VF not present suppress the link-up indication
until the VF comes first time. After that the link-up follows
link status of the device until the driver is disabled.
Set packet filter (effectively enable the RX from the VF) after
the link-up is indicated.
Signed-off-by: Yuri Benditovich <[email protected]>
Instead of calling destructor of CGuestAnnouncePackets
out of the blue, introduce Clear procedure and use it
from the destructor of the adapter context.
Signed-off-by: Yuri Benditovich <[email protected]>
Allocate RW lock in the constructor, if failed we will fail
the context initialization.
Free allocated CPU maps in the destructor, if allocated.
No need in "IsRSSinitialized" anymore.
Signed-off-by: Yuri Benditovich <[email protected]>
Automatic call to destructor frees resources allocated
by the object (memory block for control commands).
If virtqueue was created it is also deleted by this call or
earlier implicitly by Shutdown().
Signed-off-by: Yuri Benditovich <[email protected]>
Delete all the explicit initialization of receive queues
and the variable that indicates whether they are initialized.
Signed-off-by: Yuri Benditovich <[email protected]>
Change creation/deletion from alloc/free to new/destroy.
Remove some of explicit calls to constructors/destructors
of embedded objects.
Signed-off-by: Yuri Benditovich <[email protected]>
*JamboPacket provides the same control over MTU.
The difference is that *JumboPacket includes ethernet header
size while the previous MTU does not.
Signed-off-by: Yuri Benditovich <[email protected]>
NdisMAllocateSharedMemory maps memory as non cacheable on Arm Windows,
Given that Windows on Arm does not support unaligned accesses on uncached memory,
this behavior does result in crashes.
As such, use MmAllocateContiguousMemorySpecifyCache for the Arm64 architecture.
This choice is safe.
It's also to be noted that NdisMAllocateSharedMemory
ignores the isCached attribute anyway, and always
maps as cached on amd64.
This also allows us to remove the need for the Rx bounce buffer.
With this fix, networking w/ netkvm for Windows on Arm works fine.
(tested with QEMU emulator version 5.0.50 (v5.0.0-1065-g9f1f264edb))
Signed-off-by: Mohamed Mediouni <[email protected]>
SDV is modern mode fails without any information if there
are source files with '-' sign in the name. This is
preparation change for further work with modern SDV.
Signed-off-by: Yuri Benditovich <[email protected]>
Yuri Benditovich