[intelxl] Choose to operate in non-PXE mode

The physical function defaults to operating in "PXE mode" after a
power-on reset.  In this mode, receive descriptors are fetched and
written back as single descriptors.  In normal (non-PXE mode)
operation, receive descriptors are fetched and written back only as
complete cachelines unless an interrupt is raised.

There is no way to return to PXE mode from non-PXE mode, and there is
no way for the virtual function driver to operate in PXE mode.

Choose to operate in non-PXE mode.  This requires us to trick the
hardware into believing that it is raising an interrupt, so that it
will not defer writing back receive descriptors until a complete
cacheline (i.e. four packets) have been consumed.  We do so by
configuring the hardware to use MSI-X with a dummy target location in
place of the usual APIC register.

Signed-off-by: Michael Brown <mcb30@ipxe.org>

src/drivers/net/intelxl.c

@@ -116,6 +116,56 @@ static int intelxl_fetch_mac ( struct intelxl_nic *intelxl,
 	return 0;
 }
 
+/******************************************************************************
+ *
+ * MSI-X interrupts
+ *
+ ******************************************************************************
+ */
+
+/**
+ * Enable MSI-X dummy interrupt
+ *
+ * @v intelxl		Intel device
+ * @v pci		PCI device
+ * @ret rc		Return status code
+ */
+int intelxl_msix_enable ( struct intelxl_nic *intelxl,
+			  struct pci_device *pci ) {
+	int rc;
+
+	/* Enable MSI-X capability */
+	if ( ( rc = pci_msix_enable ( pci, &intelxl->msix ) ) != 0 ) {
+		DBGC ( intelxl, "INTELXL %p could not enable MSI-X: %s\n",
+		       intelxl, strerror ( rc ) );
+		return rc;
+	}
+
+	/* Configure interrupt zero to write to dummy location */
+	pci_msix_map ( &intelxl->msix, 0, virt_to_bus ( &intelxl->msg ), 0 );
+
+	/* Enable dummy interrupt zero */
+	pci_msix_unmask ( &intelxl->msix, 0 );
+
+	return 0;
+}
+
+/**
+ * Disable MSI-X dummy interrupt
+ *
+ * @v intelxl		Intel device
+ * @v pci		PCI device
+ */
+void intelxl_msix_disable ( struct intelxl_nic *intelxl,
+			    struct pci_device *pci ) {
+
+	/* Disable dummy interrupt zero */
+	pci_msix_mask ( &intelxl->msix, 0 );
+
+	/* Disable MSI-X capability */
+	pci_msix_disable ( pci, &intelxl->msix );
+}
+
 /******************************************************************************
  *
  * Admin queue
@@ -480,6 +530,39 @@ static int intelxl_admin_shutdown ( struct intelxl_nic *intelxl ) {
 	return 0;
 }
 
+/**
+ * Clear PXE mode
+ *
+ * @v intelxl		Intel device
+ * @ret rc		Return status code
+ */
+static int intelxl_admin_clear_pxe ( struct intelxl_nic *intelxl ) {
+	struct intelxl_admin_descriptor *cmd;
+	struct intelxl_admin_clear_pxe_params *pxe;
+	uint32_t gllan_rctl_0;
+	int rc;
+
+	/* Do nothing if device is already out of PXE mode */
+	gllan_rctl_0 = readl ( intelxl->regs + INTELXL_GLLAN_RCTL_0 );
+	if ( ! ( gllan_rctl_0 & INTELXL_GLLAN_RCTL_0_PXE_MODE ) ) {
+		DBGC2 ( intelxl, "INTELXL %p already in non-PXE mode\n",
+			intelxl );
+		return 0;
+	}
+
+	/* Populate descriptor */
+	cmd = intelxl_admin_command_descriptor ( intelxl );
+	cmd->opcode = cpu_to_le16 ( INTELXL_ADMIN_CLEAR_PXE );
+	pxe = &cmd->params.pxe;
+	pxe->magic = INTELXL_ADMIN_CLEAR_PXE_MAGIC;
+
+	/* Issue command */
+	if ( ( rc = intelxl_admin_command ( intelxl ) ) != 0 )
+		return rc;
+
+	return 0;
+}
+
 /**
  * Get switch configuration
  *
@@ -1504,13 +1587,6 @@ static void intelxl_poll_rx ( struct net_device *netdev ) {
 void intelxl_poll ( struct net_device *netdev ) {
 	struct intelxl_nic *intelxl = netdev->priv;
 
-	/* Acknowledge interrupts, if applicable */
-	if ( netdev_irq_enabled ( netdev ) ) {
-		writel ( ( INTELXL_INT_DYN_CTL_CLEARPBA |
-			   INTELXL_INT_DYN_CTL_INTENA_MASK ),
-			 ( intelxl->regs + intelxl->intr ) );
-	}
-
 	/* Poll for completed packets */
 	intelxl_poll_tx ( netdev );
 
@@ -1522,19 +1598,23 @@ void intelxl_poll ( struct net_device *netdev ) {
 
 	/* Refill RX ring */
 	intelxl_refill_rx ( intelxl );
-}
-
-/**
- * Enable or disable interrupts
- *
- * @v netdev		Network device
- * @v enable		Interrupts should be enabled
- */
-static void intelxl_irq ( struct net_device *netdev, int enable ) {
-	struct intelxl_nic *intelxl = netdev->priv;
 
-	writel ( ( enable ? INTELXL_INT_DYN_CTL_INTENA : 0 ),
-		 ( intelxl->regs + intelxl->intr ) );
+	/* Rearm interrupt, since otherwise receive descriptors will
+	 * be written back only after a complete cacheline (four
+	 * packets) have been received.
+	 *
+	 * There is unfortunately no efficient way to determine
+	 * whether or not rearming the interrupt is necessary.  If we
+	 * are running inside a hypervisor (e.g. using a VF or PF as a
+	 * passed-through PCI device), then the MSI-X write is
+	 * redirected by the hypervisor to the real host APIC and the
+	 * host ISR then raises an interrupt within the guest.  We
+	 * therefore cannot poll the nominal MSI-X target location to
+	 * watch for the value being written.  We could read from the
+	 * INT_DYN_CTL register, but this is even less efficient than
+	 * just unconditionally rearming the interrupt.
+	 */
+	writel ( INTELXL_INT_DYN_CTL_INTENA, intelxl->regs + intelxl->intr );
 }
 
 /** Network device operations */
@@ -1543,7 +1623,6 @@ static struct net_device_operations intelxl_operations = {
 	.close		= intelxl_close,
 	.transmit	= intelxl_transmit,
 	.poll		= intelxl_poll,
-	.irq		= intelxl_irq,
 };
 
 /******************************************************************************
@@ -1617,10 +1696,18 @@ static int intelxl_probe ( struct pci_device *pci ) {
 	if ( ( rc = intelxl_fetch_mac ( intelxl, netdev ) ) != 0 )
 		goto err_fetch_mac;
 
+	/* Enable MSI-X dummy interrupt */
+	if ( ( rc = intelxl_msix_enable ( intelxl, pci ) ) != 0 )
+		goto err_msix;
+
 	/* Open admin queues */
 	if ( ( rc = intelxl_open_admin ( intelxl ) ) != 0 )
 		goto err_open_admin;
 
+	/* Clear PXE mode */
+	if ( ( rc = intelxl_admin_clear_pxe ( intelxl ) ) != 0 )
+		goto err_admin_clear_pxe;
+
 	/* Get switch configuration */
 	if ( ( rc = intelxl_admin_switch ( intelxl ) ) != 0 )
 		goto err_admin_switch;
@@ -1667,8 +1754,11 @@ static int intelxl_probe ( struct pci_device *pci ) {
  err_admin_promisc:
  err_admin_vsi:
  err_admin_switch:
+ err_admin_clear_pxe:
 	intelxl_close_admin ( intelxl );
  err_open_admin:
+	intelxl_msix_disable ( intelxl, pci );
+ err_msix:
  err_fetch_mac:
 	intelxl_reset ( intelxl );
  err_reset:
@@ -1695,6 +1785,9 @@ static void intelxl_remove ( struct pci_device *pci ) {
 	/* Close admin queues */
 	intelxl_close_admin ( intelxl );
 
+	/* Disable MSI-X dummy interrupt */
+	intelxl_msix_disable ( intelxl, pci );
+
 	/* Reset the NIC */
 	intelxl_reset ( intelxl );
 

src/drivers/net/intelxl.h

@@ -11,6 +11,7 @@ FILE_LICENCE ( GPL2_OR_LATER_OR_UBDL );
 
 #include <stdint.h>
 #include <ipxe/if_ether.h>
+#include <ipxe/pcimsix.h>
 
 struct intelxl_nic;
 
@@ -143,6 +144,20 @@ struct intelxl_admin_shutdown_params {
 /** Driver is unloading */
 #define INTELXL_ADMIN_SHUTDOWN_UNLOADING 0x01
 
+/** Admin queue Clear PXE Mode command */
+#define INTELXL_ADMIN_CLEAR_PXE 0x0110
+
+/** Admin queue Clear PXE Mode command parameters */
+struct intelxl_admin_clear_pxe_params {
+	/** Magic value */
+	uint8_t magic;
+	/** Reserved */
+	uint8_t reserved[15];
+} __attribute__ (( packed ));
+
+/** Clear PXE Mode magic value */
+#define INTELXL_ADMIN_CLEAR_PXE_MAGIC 0x02
+
 /** Admin queue Get Switch Configuration command */
 #define INTELXL_ADMIN_SWITCH 0x0200
 
@@ -305,6 +320,8 @@ union intelxl_admin_params {
 	struct intelxl_admin_driver_params driver;
 	/** Shutdown command parameters */
 	struct intelxl_admin_shutdown_params shutdown;
+	/** Clear PXE Mode command parameters */
+	struct intelxl_admin_clear_pxe_params pxe;
 	/** Get Switch Configuration command parameters */
 	struct intelxl_admin_switch_params sw;
 	/** Get VSI Parameters command parameters */
@@ -563,6 +580,10 @@ struct intelxl_context_rx {
 /** Queue Tail Pointer Register (offset) */
 #define INTELXL_QXX_TAIL 0x8000
 
+/** Global RLAN Control 0 register */
+#define INTELXL_GLLAN_RCTL_0 0x12a500
+#define INTELXL_GLLAN_RCTL_0_PXE_MODE	0x00000001UL	/**< PXE mode */
+
 /** Transmit data descriptor */
 struct intelxl_tx_data_descriptor {
 	/** Buffer address */
@@ -709,22 +730,27 @@ intelxl_init_ring ( struct intelxl_ring *ring, unsigned int count, size_t len,
 	ring->context = context;
 }
 
-/** Number of transmit descriptors */
-#define INTELXL_TX_NUM_DESC 16
+/** Number of transmit descriptors
+ *
+ * Chosen to exceed the receive ring fill level, in order to avoid
+ * running out of transmit descriptors when sending TCP ACKs.
+ */
+#define INTELXL_TX_NUM_DESC 64
 
 /** Transmit descriptor ring maximum fill level */
 #define INTELXL_TX_FILL ( INTELXL_TX_NUM_DESC - 1 )
 
 /** Number of receive descriptors
  *
- * In PXE mode (i.e. able to post single receive descriptors), 8
- * descriptors is the only permitted value covering all possible
- * numbers of PFs.
+ * Must be a multiple of 32.
  */
-#define INTELXL_RX_NUM_DESC 8
+#define INTELXL_RX_NUM_DESC 32
 
-/** Receive descriptor ring fill level */
-#define INTELXL_RX_FILL ( INTELXL_RX_NUM_DESC - 1 )
+/** Receive descriptor ring fill level
+ *
+ * Must be a multiple of 8 and greater than 8.
+ */
+#define INTELXL_RX_FILL 16
 
 /******************************************************************************
  *
@@ -837,6 +863,10 @@ struct intelxl_nic {
 	unsigned int qset;
 	/** Interrupt control register */
 	unsigned int intr;
+	/** MSI-X capability */
+	struct pci_msix msix;
+	/** MSI-X dummy interrupt target */
+	uint32_t msg;
 
 	/** Admin command queue */
 	struct intelxl_admin command;
@@ -851,6 +881,10 @@ struct intelxl_nic {
 	struct io_buffer *rx_iobuf[INTELXL_RX_NUM_DESC];
 };
 
+extern int intelxl_msix_enable ( struct intelxl_nic *intelxl,
+				 struct pci_device *pci );
+extern void intelxl_msix_disable ( struct intelxl_nic *intelxl,
+				   struct pci_device *pci );
 extern struct intelxl_admin_descriptor *
 intelxl_admin_command_descriptor ( struct intelxl_nic *intelxl );
 extern union intelxl_admin_buffer *