sfc: Add option to use a separate channel for TX completions

In a bidirectional forwarding test, we find that the best performance
is achieved by sending the TX completion interrupts from one NIC to a
CPU which shares an L2 cache with RX completion interrupts from the
other NIC.  To facilitate this, add an option (through a module
parameter) to create separate channels for RX and TX completion with
separate IRQs when MSI-X is available.

Signed-off-by: Ben Hutchings <bhutchings@solarflare.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

2 files changed, 28 insertions, 15 deletions

diff --git a/drivers/net/sfc/efx.c b/drivers/net/sfc/efx.c
index 15ae2fcf6416..2faaa965afca 100644
--- a/drivers/net/sfc/efx.c
+++ b/drivers/net/sfc/efx.c
@@ -64,13 +64,15 @@ MODULE_PARM_DESC(lro, "Large receive offload acceleration");
 /*
  * Use separate channels for TX and RX events
  *
- * Set this to 1 to use separate channels for TX and RX. It allows us to
- * apply a higher level of interrupt moderation to TX events.
+ * Set this to 1 to use separate channels for TX and RX. It allows us
+ * to control interrupt affinity separately for TX and RX.
  *
- * This is forced to 0 for MSI interrupt mode as the interrupt vector
- * is not written
+ * This is only used in MSI-X interrupt mode
  */
-static unsigned int separate_tx_and_rx_channels = true;
+static unsigned int separate_tx_channels;
+module_param(separate_tx_channels, uint, 0644);
+MODULE_PARM_DESC(separate_tx_channels,
+                 "Use separate channels for TX and RX");
 
 /* This is the weight assigned to each of the (per-channel) virtual
  * NAPI devices.
@@ -846,26 +848,33 @@ static void efx_probe_interrupts(struct efx_nic *efx)
         if (efx->interrupt_mode == EFX_INT_MODE_MSIX) {
                 struct msix_entry xentries[EFX_MAX_CHANNELS];
                 int wanted_ints;
+                int rx_queues;
 
                 /* We want one RX queue and interrupt per CPU package
                  * (or as specified by the rss_cpus module parameter).
                  * We will need one channel per interrupt.
                  */
-                wanted_ints = rss_cpus ? rss_cpus : efx_wanted_rx_queues();
-                efx->n_rx_queues = min(wanted_ints, max_channels);
+                rx_queues = rss_cpus ? rss_cpus : efx_wanted_rx_queues();
+                wanted_ints = rx_queues + (separate_tx_channels ? 1 : 0);
+                wanted_ints = min(wanted_ints, max_channels);
 
-                for (i = 0; i < efx->n_rx_queues; i++)
+                for (i = 0; i < wanted_ints; i++)
                         xentries[i].entry = i;
-                rc = pci_enable_msix(efx->pci_dev, xentries, efx->n_rx_queues);
+                rc = pci_enable_msix(efx->pci_dev, xentries, wanted_ints);
                 if (rc > 0) {
-                        EFX_BUG_ON_PARANOID(rc >= efx->n_rx_queues);
-                        efx->n_rx_queues = rc;
+                        EFX_ERR(efx, "WARNING: Insufficient MSI-X vectors"
+                                " available (%d < %d).\n", rc, wanted_ints);
+                        EFX_ERR(efx, "WARNING: Performance may be reduced.\n");
+                        EFX_BUG_ON_PARANOID(rc >= wanted_ints);
+                        wanted_ints = rc;
                         rc = pci_enable_msix(efx->pci_dev, xentries,
-                                             efx->n_rx_queues);
+                                             wanted_ints);
                 }
 
                 if (rc == 0) {
-                        for (i = 0; i < efx->n_rx_queues; i++)
+                        efx->n_rx_queues = min(rx_queues, wanted_ints);
+                        efx->n_channels = wanted_ints;
+                        for (i = 0; i < wanted_ints; i++)
                                 efx->channel[i].irq = xentries[i].vector;
                 } else {
                         /* Fall back to single channel MSI */
@@ -877,6 +886,7 @@ static void efx_probe_interrupts(struct efx_nic *efx)
         /* Try single interrupt MSI */
         if (efx->interrupt_mode == EFX_INT_MODE_MSI) {
                 efx->n_rx_queues = 1;
+                efx->n_channels = 1;
                 rc = pci_enable_msi(efx->pci_dev);
                 if (rc == 0) {
                         efx->channel[0].irq = efx->pci_dev->irq;
@@ -889,6 +899,7 @@ static void efx_probe_interrupts(struct efx_nic *efx)
         /* Assume legacy interrupts */
         if (efx->interrupt_mode == EFX_INT_MODE_LEGACY) {
                 efx->n_rx_queues = 1;
+                efx->n_channels = 1 + (separate_tx_channels ? 1 : 0);
                 efx->legacy_irq = efx->pci_dev->irq;
         }
 }
@@ -913,8 +924,8 @@ static void efx_set_channels(struct efx_nic *efx)
         struct efx_rx_queue *rx_queue;
 
         efx_for_each_tx_queue(tx_queue, efx) {
-                if (!EFX_INT_MODE_USE_MSI(efx) && separate_tx_and_rx_channels)
-                        tx_queue->channel = &efx->channel[1];
+                if (separate_tx_channels)
+                        tx_queue->channel = &efx->channel[efx->n_channels-1];
                 else
                         tx_queue->channel = &efx->channel[0];
                 tx_queue->channel->used_flags |= EFX_USED_BY_TX;
diff --git a/drivers/net/sfc/net_driver.h b/drivers/net/sfc/net_driver.h
index f91060963475..2c5b5fabac1e 100644
--- a/drivers/net/sfc/net_driver.h
+++ b/drivers/net/sfc/net_driver.h
@@ -649,6 +649,7 @@ union efx_multicast_hash {
  * @rx_queue: RX DMA queues
  * @channel: Channels
  * @n_rx_queues: Number of RX queues
+ * @n_channels: Number of channels in use
  * @rx_buffer_len: RX buffer length
  * @rx_buffer_order: Order (log2) of number of pages for each RX buffer
  * @irq_status: Interrupt status buffer
@@ -728,6 +729,7 @@ struct efx_nic {
         struct efx_channel channel[EFX_MAX_CHANNELS];
 
         int n_rx_queues;
+        int n_channels;
         unsigned int rx_buffer_len;
         unsigned int rx_buffer_order;