Implement APIs for in-place consumption of vmbus packets. Currently, each
packet is copied and processed one at a time and as part of processing
each packet we potentially may signal the host (if it is waiting for
room to produce a packet).
These APIs help batched in-place processing of vmbus packets.
We also optimize host signaling by having a separate API to signal
the end of in-place consumption. With netvsc using these APIs,
on an iperf run on average I see about 20X reduction in checks to
signal the host.
Signed-off-by: K. Y. Srinivasan <kys@microsoft.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
In preparation for implementing APIs for in-place consumption of VMBUS
packets, movve some ring buffer functionality into hyperv.h
Signed-off-by: K. Y. Srinivasan <kys@microsoft.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Use the virt_xx barriers that have been defined for use in virtual machines.
Signed-off-by: K. Y. Srinivasan <kys@microsoft.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Use the READ_ONCE macro to access variabes that can change asynchronously.
This is the recommended mechanism for dealing with "unsafe" compiler
optimizations.
Signed-off-by: K. Y. Srinivasan <kys@microsoft.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Introduce separate functions for estimating how much can be read from
and written to the ring buffer.
Signed-off-by: K. Y. Srinivasan <kys@microsoft.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
On the consumer side, we have interrupt driven flow management of the
producer. It is sufficient to base the signaling decision on the
amount of space that is available to write after the read is complete.
The current code samples the previous available space and uses this
in making the signaling decision. This state can be stale and is
unnecessary. Since the state can be stale, we end up not signaling
the host (when we should) and this can result in a hang. Fix this
problem by removing the unnecessary check. I would like to thank
Arseney Romanenko <arseneyr@microsoft.com> for pointing out this issue.
Also, issue a full memory barrier before making the signaling descision
to correctly deal with potential reordering of the write (read index)
followed by the read of pending_sz.
Signed-off-by: K. Y. Srinivasan <kys@microsoft.com>
Tested-by: Dexuan Cui <decui@microsoft.com>
Cc: <stable@vger.kernel.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
On the channel send side, many of the VMBUS
device drivers explicity serialize access to the
outgoing ring buffer. Give more control to the
VMBUS device drivers in terms how to serialize
accesss to the outgoing ring buffer.
The default behavior will be to aquire the
ring lock to preserve the current behavior.
Signed-off-by: K. Y. Srinivasan <kys@microsoft.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
The function hv_ringbuffer_read() is called always on a pre-assigned
CPU. Each chnnel is bound to a specific CPU and this function is
always called on the CPU the channel is bound. There is no need to
acquire the spin lock; get rid of this overhead.
Signed-off-by: K. Y. Srinivasan <kys@microsoft.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Currently, there is only one user for hv_ringbuffer_read()/
hv_ringbuffer_peak() functions and the usage of these functions is:
- insecure as we drop ring_lock between them, someone else (in theory
only) can acquire it in between;
- non-optimal as we do a number of things (acquire/release the above
mentioned lock, calculate available space on the ring, ...) twice and
this path is performance-critical.
Remove hv_ringbuffer_peek() moving the logic from __vmbus_recvpacket() to
hv_ringbuffer_read().
Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: K. Y. Srinivasan <kys@microsoft.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
hv_ringbuffer_peek() does the same as hv_ringbuffer_read() without
advancing the read index. The only functional change this patch brings
is moving hv_need_to_signal_on_read() call under the ring_lock but this
function is just a couple of comparisons.
Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: K. Y. Srinivasan <kys@microsoft.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Convert 6+-string comments repeating function names to normal kernel-style
comments and fix a couple of other comment style issues. No textual or
functional changes intended.
Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: K. Y. Srinivasan <kys@microsoft.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Fixes a bug where previously hv_ringbuffer_read would pass in the old
number of bytes available to read instead of the expected old read index
when calculating when to signal to the host that the ringbuffer is empty.
Since the previous write size is already saved, also changes the
hv_need_to_signal_on_read to use the previously read value rather than
recalculating it.
Signed-off-by: Christopher Oo <t-chriso@microsoft.com>
Signed-off-by: K. Y. Srinivasan <kys@microsoft.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
In win8 we have a feature that allows for interrupt driven flow management
for host/guest communication. For instance, if the host were blocked because
there was no space available in the ringbuffer, the host could request that the
guest send an interrupt when space becomes available in the ringbuffer (when
the guest drains the ringbuffer).
While this feature was implemented in the guest a while ago, we had not
advertised that the guest supported this feature. This patch advertises
the support to the host.
For pre-win8 hosts, this has no effect since the size of the ringbuffer
control structure has not changed and all changes have been backward
compatible - unused/reserved space has been used to implement this
feature.
In this version of the patch I have cleaned up the commit log based on
feedback from Greg KH.
Signed-off-by: K. Y. Srinivasan <kys@microsoft.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
The current channel code is using scatterlist abstraction to pass data to the
ringbuffer API on the send path. This causes unnecessary translations
between virtual and physical addresses. Fix this.
Signed-off-by: K. Y. Srinivasan <kys@microsoft.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
This patches add a read barriers to force the driver to check the interrupt mask
before read_index. Otherwise we may lost a kick to host.
Cc: K. Y. Srinivasan <kys@microsoft.com>
Cc: Haiyang Zhang <haiyangz@microsoft.com>
Signed-off-by: Jason Wang <jasowang@redhat.com>
Signed-off-by: K. Y. Srinivasan <kys@microsoft.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Even if guest were compiled without SMP support, it could not assume that host
wasn't. So switch to use mb() instead of smp_mb() to force memory barriers for
UP guest.
Signed-off-by: Jason Wang <jasowang@redhat.com>
Cc: Haiyang Zhang <haiyangz@microsoft.com>
Cc: stable <stable@vger.kernel.org>
Signed-off-by: K. Y. Srinivasan <kys@microsoft.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
As part of updating the vmbus protocol, the function hv_need_to_signal()
was introduced. This functions helps optimize signalling from guest to
host. The newly added memory barrier is needed to ensure that we correctly
decide when to signal the host.
Signed-off-by: K. Y. Srinivasan <kys@microsoft.com>
Reviewed-by: Haiyang Zhang <haiyangz@microsoft.com>
Reported-by: Olaf Hering <olh@suse.de>
Cc: Stable <stable@vger.kernel.org> (V3.8+)
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Implement flow management on the send side. When the sender is blocked, the reader
can potentially signal the sender to indicate there is now room to send.
Signed-off-by: K. Y. Srinivasan <kys@microsoft.com>
Reviewed-by: Haiyang Zhang <haiyangz@microsoft.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
This function is no longer used; get rid of it.
Signed-off-by: K. Y. Srinivasan <kys@microsoft.com>
Reviewed-by: Haiyang Zhang <haiyangz@microsoft.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
The host has already implemented the "read" side optimizations.
Leverage that to optimize "write" side signaling.
Signed-off-by: K. Y. Srinivasan <kys@microsoft.com>
Reviewed-by: Haiyang Zhang <haiyangz@microsoft.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Implement functions that will support read-side signaling optimization.
By having the reader indicate the start of the "read" operation and the
"end" of the read operation we can more efficiently handle the signaling
protocol: while the read is in progress, there is no need for the "writer"
to signal the "reader" as new items are put on the read queue.
Signed-off-by: K. Y. Srinivasan <kys@microsoft.com>
Reviewed-by: Haiyang Zhang <haiyangz@microsoft.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
In the existing code, we only stop queue when the ringbuffer is full,
so the current packet has to be dropped or retried from upper layer.
This patch stops the tx queue when available ringbuffer is below
the low watermark. So the ringbuffer still has small amount of space
available for the current packet. This will reduce the overhead of
retries on sending.
Signed-off-by: Haiyang Zhang <haiyangz@microsoft.com>
Reviewed-by: K. Y. Srinivasan <kys@microsoft.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
After many years wandering the desert, it is finally time for the
Microsoft HyperV code to move out of the staging directory. Or at least
the core hyperv bus code, and the utility driver, the rest still have
some review to get through by the various subsystem maintainers.
Signed-off-by: Greg Kroah-Hartman <gregkh@suse.de>
Signed-off-by: K. Y. Srinivasan <kys@microsoft.com>
K. Y. Srinivasan