Drivers: hv: Optimize the signaling on the write path

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>

3 files changed, 49 insertions, 10 deletions

diff --git a/drivers/hv/channel.c b/drivers/hv/channel.c
index 773a2f25a8f0..727c5f1d6acf 100644
--- a/drivers/hv/channel.c
+++ b/drivers/hv/channel.c
@@ -564,6 +564,7 @@ int vmbus_sendpacket(struct vmbus_channel *channel, const void *buffer,
         struct scatterlist bufferlist[3];
         u64 aligned_data = 0;
         int ret;
+        bool signal = false;
 
 
         /* Setup the descriptor */
@@ -580,9 +581,9 @@ int vmbus_sendpacket(struct vmbus_channel *channel, const void *buffer,
         sg_set_buf(&bufferlist[2], &aligned_data,
                    packetlen_aligned - packetlen);
 
-        ret = hv_ringbuffer_write(&channel->outbound, bufferlist, 3);
+        ret = hv_ringbuffer_write(&channel->outbound, bufferlist, 3, &signal);
 
-        if (ret == 0 && !hv_get_ringbuffer_interrupt_mask(&channel->outbound))
+        if (ret == 0 && signal)
                 vmbus_setevent(channel);
 
         return ret;
@@ -606,6 +607,7 @@ int vmbus_sendpacket_pagebuffer(struct vmbus_channel *channel,
         u32 packetlen_aligned;
         struct scatterlist bufferlist[3];
         u64 aligned_data = 0;
+        bool signal = false;
 
         if (pagecount > MAX_PAGE_BUFFER_COUNT)
                 return -EINVAL;
@@ -641,9 +643,9 @@ int vmbus_sendpacket_pagebuffer(struct vmbus_channel *channel,
         sg_set_buf(&bufferlist[2], &aligned_data,
                 packetlen_aligned - packetlen);
 
-        ret = hv_ringbuffer_write(&channel->outbound, bufferlist, 3);
+        ret = hv_ringbuffer_write(&channel->outbound, bufferlist, 3, &signal);
 
-        if (ret == 0 && !hv_get_ringbuffer_interrupt_mask(&channel->outbound))
+        if (ret == 0 && signal)
                 vmbus_setevent(channel);
 
         return ret;
@@ -665,6 +667,7 @@ int vmbus_sendpacket_multipagebuffer(struct vmbus_channel *channel,
         u32 packetlen_aligned;
         struct scatterlist bufferlist[3];
         u64 aligned_data = 0;
+        bool signal = false;
         u32 pfncount = NUM_PAGES_SPANNED(multi_pagebuffer->offset,
                                          multi_pagebuffer->len);
 
@@ -703,9 +706,9 @@ int vmbus_sendpacket_multipagebuffer(struct vmbus_channel *channel,
         sg_set_buf(&bufferlist[2], &aligned_data,
                 packetlen_aligned - packetlen);
 
-        ret = hv_ringbuffer_write(&channel->outbound, bufferlist, 3);
+        ret = hv_ringbuffer_write(&channel->outbound, bufferlist, 3, &signal);
 
-        if (ret == 0 && !hv_get_ringbuffer_interrupt_mask(&channel->outbound))
+        if (ret == 0 && signal)
                 vmbus_setevent(channel);
 
         return ret;
diff --git a/drivers/hv/hyperv_vmbus.h b/drivers/hv/hyperv_vmbus.h
index 3184f6ff4e74..895c898812bd 100644
--- a/drivers/hv/hyperv_vmbus.h
+++ b/drivers/hv/hyperv_vmbus.h
@@ -555,7 +555,7 @@ void hv_ringbuffer_cleanup(struct hv_ring_buffer_info *ring_info);
 
 int hv_ringbuffer_write(struct hv_ring_buffer_info *ring_info,
                     struct scatterlist *sglist,
-                    u32 sgcount);
+                    u32 sgcount, bool *signal);
 
 int hv_ringbuffer_peek(struct hv_ring_buffer_info *ring_info, void *buffer,
                    u32 buflen);
diff --git a/drivers/hv/ring_buffer.c b/drivers/hv/ring_buffer.c
index 001079287709..9bc55f0dcf47 100644
--- a/drivers/hv/ring_buffer.c
+++ b/drivers/hv/ring_buffer.c
@@ -53,6 +53,37 @@ u32 hv_end_read(struct hv_ring_buffer_info *rbi)
         return read;
 }
 
+/*
+ * When we write to the ring buffer, check if the host needs to
+ * be signaled. Here is the details of this protocol:
+ *
+ *      1. The host guarantees that while it is draining the
+ *         ring buffer, it will set the interrupt_mask to
+ *         indicate it does not need to be interrupted when
+ *         new data is placed.
+ *
+ *      2. The host guarantees that it will completely drain
+ *         the ring buffer before exiting the read loop. Further,
+ *         once the ring buffer is empty, it will clear the
+ *         interrupt_mask and re-check to see if new data has
+ *         arrived.
+ */
+
+static bool hv_need_to_signal(u32 old_write, struct hv_ring_buffer_info *rbi)
+{
+        if (rbi->ring_buffer->interrupt_mask)
+                return false;
+
+        /*
+         * This is the only case we need to signal when the
+         * ring transitions from being empty to non-empty.
+         */
+        if (old_write == rbi->ring_buffer->read_index)
+                return true;
+
+        return false;
+}
+
 
 /*
  * hv_get_next_write_location()
@@ -322,7 +353,7 @@ void hv_ringbuffer_cleanup(struct hv_ring_buffer_info *ring_info)
  *
  */
 int hv_ringbuffer_write(struct hv_ring_buffer_info *outring_info,
-                    struct scatterlist *sglist, u32 sgcount)
+                    struct scatterlist *sglist, u32 sgcount, bool *signal)
 {
         int i = 0;
         u32 bytes_avail_towrite;
@@ -331,6 +362,7 @@ int hv_ringbuffer_write(struct hv_ring_buffer_info *outring_info,
 
         struct scatterlist *sg;
         u32 next_write_location;
+        u32 old_write;
         u64 prev_indices = 0;
         unsigned long flags;
 
@@ -359,6 +391,8 @@ int hv_ringbuffer_write(struct hv_ring_buffer_info *outring_info,
         /* Write to the ring buffer */
         next_write_location = hv_get_next_write_location(outring_info);
 
+        old_write = next_write_location;
+
         for_each_sg(sglist, sg, sgcount, i)
         {
                 next_write_location = hv_copyto_ringbuffer(outring_info,
@@ -375,14 +409,16 @@ int hv_ringbuffer_write(struct hv_ring_buffer_info *outring_info,
                                              &prev_indices,
                                              sizeof(u64));
 
-        /* Make sure we flush all writes before updating the writeIndex */
-        smp_wmb();
+        /* Issue a full memory barrier before updating the write index */
+        smp_mb();
 
         /* Now, update the write location */
         hv_set_next_write_location(outring_info, next_write_location);
 
 
         spin_unlock_irqrestore(&outring_info->ring_lock, flags);
+
+        *signal = hv_need_to_signal(old_write, outring_info);
         return 0;
 }