[PATCH 3/3] Drivers: hv: vmbus: Implement a mechanism to tag the channel for low latency

[kys]

From: K. Y. Srinivasan <kys@xxxxxxxxxxxxx>

On Hyper-V, performance critical channels use the monitor
mechanism to signal the host when the guest posts mesages
for the host. This mechanism minimizes the hypervisor intercepts
and also makes the host more efficient in that each time the
host is woken up, it processes a batch of messages as opposed to
just one. The goal here is improve the throughput and this is at
the expense of increased latency.
Implement a mechanism to let the client driver decide if latency
is important.

Signed-off-by: K. Y. Srinivasan <kys@xxxxxxxxxxxxx>
---
 drivers/hv/channel.c   |    7 ++++++-
 include/linux/hyperv.h |   35 +++++++++++++++++++++++++++++++++++
 2 files changed, 41 insertions(+), 1 deletions(-)

diff --git a/drivers/hv/channel.c b/drivers/hv/channel.c
index da022d3..e3a0048 100644
--- a/drivers/hv/channel.c
+++ b/drivers/hv/channel.c
@@ -43,7 +43,12 @@ static void vmbus_setevent(struct vmbus_channel *channel)
 {
 	struct hv_monitor_page *monitorpage;
 
-	if (channel->offermsg.monitor_allocated) {
+	/*
+	 * For channels marked as in "low latency" mode
+	 * bypass the monitor page mechanism.
+	 */
+	if ((channel->offermsg.monitor_allocated) &&
+	    (!channel->low_latency)) {
 		/* Each u32 represents 32 channels */
 		sync_set_bit(channel->offermsg.child_relid & 31,
 			(unsigned long *) vmbus_connection.send_int_page +
diff --git a/include/linux/hyperv.h b/include/linux/hyperv.h
index b10954a..362acf0 100644
--- a/include/linux/hyperv.h
+++ b/include/linux/hyperv.h
@@ -850,6 +850,31 @@ struct vmbus_channel {
 	 * ring lock to preserve the current behavior.
 	 */
 	bool acquire_ring_lock;
+	/*
+	 * For performance critical channels (storage, networking
+	 * etc,), Hyper-V has a mechanism to enhance the throughput
+	 * at the expense of latency:
+	 * When the host is to be signaled, we just set a bit in a shared page
+	 * and this bit will be inspected by the hypervisor within a certain
+	 * window and if the bit is set, the host will be signaled. The window
+	 * of time is the monitor latency - currently around 100 usecs. This
+	 * mechanism improves throughput by:
+	 *
+	 * A) Making the host more efficient - each time it wakes up,
+	 *    potentially it will process morev number of packets. The
+	 *    monitor latency allows a batch to build up.
+	 * B) By deferring the hypercall to signal, we will also minimize
+	 *    the interrupts.
+	 *
+	 * Clearly, these optimizations improve throughput at the expense of
+	 * latency. Furthermore, since the channel is shared for both
+	 * control and data messages, control messages currently suffer
+	 * unnecessary latency adversley impacting performance and boot
+	 * time. To fix this issue, permit tagging the channel as being
+	 * in "low latency" mode. In this mode, we will bypass the monitor
+	 * mechanism.
+	 */
+	bool low_latency;
 
 };
 
@@ -891,6 +916,16 @@ static inline void set_channel_pending_send_size(struct vmbus_channel *c,
 	c->outbound.ring_buffer->pending_send_sz = size;
 }
 
+static inline void set_low_latency_mode(struct vmbus_channel *c)
+{
+	c->low_latency = true;
+}
+
+static inline void clear_low_latency_mode(struct vmbus_channel *c)
+{
+	c->low_latency = false;
+}
+
 void vmbus_onmessage(void *context);
 
 int vmbus_request_offers(void);
-- 
1.7.4.1