Merge branch 'master' of /home/davem/src/GIT/linux-2.6/

Conflicts: drivers/firmware/iscsi_ibft.c
author: David S. Miller <davem@davemloft.net> 2010-02-28 19:23:06 -0800
committer: David S. Miller <davem@davemloft.net> 2010-02-28 19:23:06 -0800
commit: 47871889c601d8199c51a4086f77eebd77c29b0b (patch)
tree: 40cdcac3bff0ee40cc33dcca61d0577cdf965f77 /drivers/staging/octeon/ethernet-rx.c
parent: c16cc0b464b8876cfd57ce1c1dbcb6f9a6a0bce3 (diff)
parent: 30ff056c42c665b9ea535d8515890857ae382540 (diff)
1 files changed, 219 insertions, 165 deletions
diff --git a/drivers/staging/octeon/ethernet-rx.c b/drivers/staging/octeon/ethernet-rx.c
index 1b237b7e689d..cb38f9eb2cc0 100644
--- a/drivers/staging/octeon/ethernet-rx.c
+++ b/drivers/staging/octeon/ethernet-rx.c
@@ -4,7 +4,7 @@
  * Contact: support@caviumnetworks.com
  * This file is part of the OCTEON SDK
  *
- * Copyright (c) 2003-2007 Cavium Networks
+ * Copyright (c) 2003-2010 Cavium Networks
  *
  * This file is free software; you can redistribute it and/or modify
  * it under the terms of the GNU General Public License, Version 2, as
@@ -27,16 +27,14 @@
 #include <linux/module.h>
 #include <linux/kernel.h>
 #include <linux/cache.h>
+#include <linux/cpumask.h>
 #include <linux/netdevice.h>
 #include <linux/init.h>
 #include <linux/etherdevice.h>
 #include <linux/ip.h>
 #include <linux/string.h>
 #include <linux/prefetch.h>
-#include <linux/ethtool.h>
-#include <linux/mii.h>
-#include <linux/seq_file.h>
-#include <linux/proc_fs.h>
+#include <linux/smp.h>
 #include <net/dst.h>
 #ifdef CONFIG_XFRM
 #include <linux/xfrm.h>
@@ -48,8 +46,9 @@
 #include <asm/octeon/octeon.h>
 
 #include "ethernet-defines.h"
-#include "octeon-ethernet.h"
 #include "ethernet-mem.h"
+#include "ethernet-rx.h"
+#include "octeon-ethernet.h"
 #include "ethernet-util.h"
 
 #include "cvmx-helper.h"
@@ -61,62 +60,88 @@
 
 #include "cvmx-gmxx-defs.h"
 
-struct cvm_tasklet_wrapper {
-	struct tasklet_struct t;
-};
+struct cvm_napi_wrapper {
+	struct napi_struct napi;
+} ____cacheline_aligned_in_smp;
 
-/*
- * Aligning the tasklet_struct on cachline boundries seems to decrease
- * throughput even though in theory it would reduce contantion on the
- * cache lines containing the locks.
- */
+static struct cvm_napi_wrapper cvm_oct_napi[NR_CPUS] __cacheline_aligned_in_smp;
 
-static struct cvm_tasklet_wrapper cvm_oct_tasklet[NR_CPUS];
+struct cvm_oct_core_state {
+	int baseline_cores;
+	/*
+	 * The number of additional cores that could be processing
+	 * input packtes.
+	 */
+	atomic_t available_cores;
+	cpumask_t cpu_state;
+} ____cacheline_aligned_in_smp;
 
-/**
- * Interrupt handler. The interrupt occurs whenever the POW
- * transitions from 0->1 packets in our group.
- *
- * @cpl:
- * @dev_id:
- * @regs:
- * Returns
- */
-irqreturn_t cvm_oct_do_interrupt(int cpl, void *dev_id)
+static struct cvm_oct_core_state core_state __cacheline_aligned_in_smp;
+
+static void cvm_oct_enable_napi(void *_)
 {
-	/* Acknowledge the interrupt */
-	if (INTERRUPT_LIMIT)
-		cvmx_write_csr(CVMX_POW_WQ_INT, 1 << pow_receive_group);
-	else
-		cvmx_write_csr(CVMX_POW_WQ_INT, 0x10001 << pow_receive_group);
-	preempt_disable();
-	tasklet_schedule(&cvm_oct_tasklet[smp_processor_id()].t);
-	preempt_enable();
-	return IRQ_HANDLED;
+	int cpu = smp_processor_id();
+	napi_schedule(&cvm_oct_napi[cpu].napi);
+}
+
+static void cvm_oct_enable_one_cpu(void)
+{
+	int v;
+	int cpu;
+
+	/* Check to see if more CPUs are available for receive processing... */
+	v = atomic_sub_if_positive(1, &core_state.available_cores);
+	if (v < 0)
+		return;
+
+	/* ... if a CPU is available, Turn on NAPI polling for that CPU.  */
+	for_each_online_cpu(cpu) {
+		if (!cpu_test_and_set(cpu, core_state.cpu_state)) {
+			v = smp_call_function_single(cpu, cvm_oct_enable_napi,
+						     NULL, 0);
+			if (v)
+				panic("Can't enable NAPI.");
+			break;
+		}
+	}
+}
+
+static void cvm_oct_no_more_work(void)
+{
+	int cpu = smp_processor_id();
+
+	/*
+	 * CPU zero is special.  It always has the irq enabled when
+	 * waiting for incoming packets.
+	 */
+	if (cpu == 0) {
+		enable_irq(OCTEON_IRQ_WORKQ0 + pow_receive_group);
+		return;
+	}
+
+	cpu_clear(cpu, core_state.cpu_state);
+	atomic_add(1, &core_state.available_cores);
 }
 
-#ifdef CONFIG_NET_POLL_CONTROLLER
 /**
- * This is called when the kernel needs to manually poll the
- * device. For Octeon, this is simply calling the interrupt
- * handler. We actually poll all the devices, not just the
- * one supplied.
+ * cvm_oct_do_interrupt - interrupt handler.
+ *
+ * The interrupt occurs whenever the POW has packets in our group.
  *
- * @dev:    Device to poll. Unused
  */
-void cvm_oct_poll_controller(struct net_device *dev)
+static irqreturn_t cvm_oct_do_interrupt(int cpl, void *dev_id)
 {
-	preempt_disable();
-	tasklet_schedule(&cvm_oct_tasklet[smp_processor_id()].t);
-	preempt_enable();
+	/* Disable the IRQ and start napi_poll. */
+	disable_irq_nosync(OCTEON_IRQ_WORKQ0 + pow_receive_group);
+	cvm_oct_enable_napi(NULL);
+
+	return IRQ_HANDLED;
 }
-#endif
 
 /**
- * This is called on receive errors, and determines if the packet
- * can be dropped early-on in cvm_oct_tasklet_rx().
- *
+ * cvm_oct_check_rcv_error - process receive errors
  * @work: Work queue entry pointing to the packet.
+ *
  * Returns Non-zero if the packet can be dropped, zero otherwise.
  */
 static inline int cvm_oct_check_rcv_error(cvmx_wqe_t *work)
@@ -199,19 +224,20 @@ static inline int cvm_oct_check_rcv_error(cvmx_wqe_t *work)
 }
 
 /**
- * Tasklet function that is scheduled on a core when an interrupt occurs.
+ * cvm_oct_napi_poll - the NAPI poll function.
+ * @napi: The NAPI instance, or null if called from cvm_oct_poll_controller
+ * @budget: Maximum number of packets to receive.
  *
- * @unused:
+ * Returns the number of packets processed.
  */
-void cvm_oct_tasklet_rx(unsigned long unused)
+static int cvm_oct_napi_poll(struct napi_struct *napi, int budget)
 {
-	const int coreid = cvmx_get_core_num();
-	uint64_t old_group_mask;
-	uint64_t old_scratch;
-	int rx_count = 0;
-	int number_to_free;
-	int num_freed;
-	int packet_not_copied;
+	const int	coreid = cvmx_get_core_num();
+	uint64_t	old_group_mask;
+	uint64_t	old_scratch;
+	int		rx_count = 0;
+	int		did_work_request = 0;
+	int		packet_not_copied;
 
 	/* Prefetch cvm_oct_device since we know we need it soon */
 	prefetch(cvm_oct_device);
@@ -227,59 +253,63 @@ void cvm_oct_tasklet_rx(unsigned long unused)
 	cvmx_write_csr(CVMX_POW_PP_GRP_MSKX(coreid),
 		       (old_group_mask & ~0xFFFFull) | 1 << pow_receive_group);
 
-	if (USE_ASYNC_IOBDMA)
+	if (USE_ASYNC_IOBDMA) {
 		cvmx_pow_work_request_async(CVMX_SCR_SCRATCH, CVMX_POW_NO_WAIT);
+		did_work_request = 1;
+	}
 
-	while (1) {
+	while (rx_count < budget) {
 		struct sk_buff *skb = NULL;
+		struct sk_buff **pskb = NULL;
 		int skb_in_hw;
 		cvmx_wqe_t *work;
 
-		if (USE_ASYNC_IOBDMA) {
+		if (USE_ASYNC_IOBDMA && did_work_request)
 			work = cvmx_pow_work_response_async(CVMX_SCR_SCRATCH);
-		} else {
-			if ((INTERRUPT_LIMIT == 0)
-			    || likely(rx_count < MAX_RX_PACKETS))
-				work =
-				    cvmx_pow_work_request_sync
-				    (CVMX_POW_NO_WAIT);
-			else
-				work = NULL;
-		}
+		else
+			work = cvmx_pow_work_request_sync(CVMX_POW_NO_WAIT);
+
 		prefetch(work);
-		if (work == NULL)
+		did_work_request = 0;
+		if (work == NULL) {
+			union cvmx_pow_wq_int wq_int;
+			wq_int.u64 = 0;
+			wq_int.s.iq_dis = 1 << pow_receive_group;
+			wq_int.s.wq_int = 1 << pow_receive_group;
+			cvmx_write_csr(CVMX_POW_WQ_INT, wq_int.u64);
 			break;
+		}
+		pskb = (struct sk_buff **)(cvm_oct_get_buffer_ptr(work->packet_ptr) - sizeof(void *));
+		prefetch(pskb);
 
-		/*
-		 * Limit each core to processing MAX_RX_PACKETS
-		 * packets without a break.  This way the RX can't
-		 * starve the TX task.
-		 */
-		if (USE_ASYNC_IOBDMA) {
-
-			if ((INTERRUPT_LIMIT == 0)
-			    || likely(rx_count < MAX_RX_PACKETS))
-				cvmx_pow_work_request_async_nocheck
-				    (CVMX_SCR_SCRATCH, CVMX_POW_NO_WAIT);
-			else {
-				cvmx_scratch_write64(CVMX_SCR_SCRATCH,
-						     0x8000000000000000ull);
-				cvmx_pow_tag_sw_null_nocheck();
-			}
+		if (USE_ASYNC_IOBDMA && rx_count < (budget - 1)) {
+			cvmx_pow_work_request_async_nocheck(CVMX_SCR_SCRATCH, CVMX_POW_NO_WAIT);
+			did_work_request = 1;
+		}
+
+		if (rx_count == 0) {
+			/*
+			 * First time through, see if there is enough
+			 * work waiting to merit waking another
+			 * CPU.
+			 */
+			union cvmx_pow_wq_int_cntx counts;
+			int backlog;
+			int cores_in_use = core_state.baseline_cores - atomic_read(&core_state.available_cores);
+			counts.u64 = cvmx_read_csr(CVMX_POW_WQ_INT_CNTX(pow_receive_group));
+			backlog = counts.s.iq_cnt + counts.s.ds_cnt;
+			if (backlog > budget * cores_in_use && napi != NULL)
+				cvm_oct_enable_one_cpu();
 		}
 
 		skb_in_hw = USE_SKBUFFS_IN_HW && work->word2.s.bufs == 1;
 		if (likely(skb_in_hw)) {
-			skb =
-			    *(struct sk_buff
-			      **)(cvm_oct_get_buffer_ptr(work->packet_ptr) -
-				  sizeof(void *));
+			skb = *pskb;
 			prefetch(&skb->head);
 			prefetch(&skb->len);
 		}
 		prefetch(cvm_oct_device[work->ipprt]);
 
-		rx_count++;
 		/* Immediately throw away all packets with receive errors */
 		if (unlikely(work->word2.snoip.rcv_error)) {
 			if (cvm_oct_check_rcv_error(work))
@@ -292,39 +322,27 @@ void cvm_oct_tasklet_rx(unsigned long unused)
 		 * buffer.
 		 */
 		if (likely(skb_in_hw)) {
-			/*
-			 * This calculation was changed in case the
-			 * skb header is using a different address
-			 * aliasing type than the buffer. It doesn't
-			 * make any differnece now, but the new one is
-			 * more correct.
-			 */
-			skb->data =
-			    skb->head + work->packet_ptr.s.addr -
-			    cvmx_ptr_to_phys(skb->head);
+			skb->data = skb->head + work->packet_ptr.s.addr - cvmx_ptr_to_phys(skb->head);
 			prefetch(skb->data);
 			skb->len = work->len;
 			skb_set_tail_pointer(skb, skb->len);
 			packet_not_copied = 1;
 		} else {
-
 			/*
 			 * We have to copy the packet. First allocate
 			 * an skbuff for it.
 			 */
 			skb = dev_alloc_skb(work->len);
 			if (!skb) {
-				DEBUGPRINT("Port %d failed to allocate "
-					   "skbuff, packet dropped\n",
-				     work->ipprt);
+				DEBUGPRINT("Port %d failed to allocate skbuff, packet dropped\n",
+					   work->ipprt);
 				cvm_oct_free_work(work);
 				continue;
 			}
 
 			/*
 			 * Check if we've received a packet that was
-			 * entirely stored in the work entry. This is
-			 * untested.
+			 * entirely stored in the work entry.
 			 */
 			if (unlikely(work->word2.s.bufs == 0)) {
 				uint8_t *ptr = work->packet_data;
@@ -343,15 +361,13 @@ void cvm_oct_tasklet_rx(unsigned long unused)
 				/* No packet buffers to free */
 			} else {
 				int segments = work->word2.s.bufs;
-				union cvmx_buf_ptr segment_ptr =
-					work->packet_ptr;
+				union cvmx_buf_ptr segment_ptr = work->packet_ptr;
 				int len = work->len;
 
 				while (segments--) {
 					union cvmx_buf_ptr next_ptr =
-					    *(union cvmx_buf_ptr *)
-					    cvmx_phys_to_ptr(segment_ptr.s.
-							     addr - 8);
+					    *(union cvmx_buf_ptr *)cvmx_phys_to_ptr(segment_ptr.s.addr - 8);
+
 			/*
 			 * Octeon Errata PKI-100: The segment size is
 			 * wrong. Until it is fixed, calculate the
@@ -361,22 +377,18 @@ void cvm_oct_tasklet_rx(unsigned long unused)
 			 * one: int segment_size =
 			 * segment_ptr.s.size;
 			 */
-					int segment_size =
-					    CVMX_FPA_PACKET_POOL_SIZE -
-					    (segment_ptr.s.addr -
-					     (((segment_ptr.s.addr >> 7) -
-					       segment_ptr.s.back) << 7));
-					/* Don't copy more than what is left
-					   in the packet */
+					int segment_size = CVMX_FPA_PACKET_POOL_SIZE -
+						(segment_ptr.s.addr - (((segment_ptr.s.addr >> 7) - segment_ptr.s.back) << 7));
+					/*
+					 * Don't copy more than what
+					 * is left in the packet.
+					 */
 					if (segment_size > len)
 						segment_size = len;
 					/* Copy the data into the packet */
 					memcpy(skb_put(skb, segment_size),
-					       cvmx_phys_to_ptr(segment_ptr.s.
-								addr),
+					       cvmx_phys_to_ptr(segment_ptr.s.addr),
 					       segment_size);
-					/* Reduce the amount of bytes left
-					   to copy */
 					len -= segment_size;
 					segment_ptr = next_ptr;
 				}
@@ -389,16 +401,15 @@ void cvm_oct_tasklet_rx(unsigned long unused)
 			struct net_device *dev = cvm_oct_device[work->ipprt];
 			struct octeon_ethernet *priv = netdev_priv(dev);
 
-			/* Only accept packets for devices
-			   that are currently up */
+			/*
+			 * Only accept packets for devices that are
+			 * currently up.
+			 */
 			if (likely(dev->flags & IFF_UP)) {
 				skb->protocol = eth_type_trans(skb, dev);
 				skb->dev = dev;
 
-				if (unlikely
-				    (work->word2.s.not_IP
-				     || work->word2.s.IP_exc
-				     || work->word2.s.L4_error))
+				if (unlikely(work->word2.s.not_IP || work->word2.s.IP_exc || work->word2.s.L4_error))
 					skb->ip_summed = CHECKSUM_NONE;
 				else
 					skb->ip_summed = CHECKSUM_UNNECESSARY;
@@ -414,15 +425,13 @@ void cvm_oct_tasklet_rx(unsigned long unused)
 #endif
 				}
 				netif_receive_skb(skb);
+				rx_count++;
 			} else {
+				/* Drop any packet received for a device that isn't up */
 				/*
-				 * Drop any packet received for a
-				 * device that isn't up.
-				 */
-				/*
-				   DEBUGPRINT("%s: Device not up, packet dropped\n",
-				   dev->name);
-				 */
+				DEBUGPRINT("%s: Device not up, packet dropped\n",
+					   dev->name);
+				*/
 #ifdef CONFIG_64BIT
 				atomic64_add(1, (atomic64_t *)&priv->stats.rx_dropped);
 #else
@@ -435,9 +444,8 @@ void cvm_oct_tasklet_rx(unsigned long unused)
 			 * Drop any packet received for a device that
 			 * doesn't exist.
 			 */
-			DEBUGPRINT("Port %d not controlled by Linux, packet "
-				   "dropped\n",
-			     work->ipprt);
+			DEBUGPRINT("Port %d not controlled by Linux, packet dropped\n",
+				   work->ipprt);
 			dev_kfree_skb_irq(skb);
 		}
 		/*
@@ -459,47 +467,93 @@ void cvm_oct_tasklet_rx(unsigned long unused)
 			cvm_oct_free_work(work);
 		}
 	}
-
 	/* Restore the original POW group mask */
 	cvmx_write_csr(CVMX_POW_PP_GRP_MSKX(coreid), old_group_mask);
 	if (USE_ASYNC_IOBDMA) {
 		/* Restore the scratch area */
 		cvmx_scratch_write64(CVMX_SCR_SCRATCH, old_scratch);
 	}
+	cvm_oct_rx_refill_pool(0);
 
-	if (USE_SKBUFFS_IN_HW) {
-		/* Refill the packet buffer pool */
-		number_to_free =
-		    cvmx_fau_fetch_and_add32(FAU_NUM_PACKET_BUFFERS_TO_FREE, 0);
-
-		if (number_to_free > 0) {
-			cvmx_fau_atomic_add32(FAU_NUM_PACKET_BUFFERS_TO_FREE,
-					      -number_to_free);
-			num_freed =
-			    cvm_oct_mem_fill_fpa(CVMX_FPA_PACKET_POOL,
-						 CVMX_FPA_PACKET_POOL_SIZE,
-						 number_to_free);
-			if (num_freed != number_to_free) {
-				cvmx_fau_atomic_add32
-				    (FAU_NUM_PACKET_BUFFERS_TO_FREE,
-				     number_to_free - num_freed);
-			}
-		}
+	if (rx_count < budget && napi != NULL) {
+		/* No more work */
+		napi_complete(napi);
+		cvm_oct_no_more_work();
 	}
+	return rx_count;
 }
 
+#ifdef CONFIG_NET_POLL_CONTROLLER
+/**
+ * cvm_oct_poll_controller - poll for receive packets
+ * device.
+ *
+ * @dev:    Device to poll. Unused
+ */
+void cvm_oct_poll_controller(struct net_device *dev)
+{
+	cvm_oct_napi_poll(NULL, 16);
+}
+#endif
+
 void cvm_oct_rx_initialize(void)
 {
 	int i;
-	/* Initialize all of the tasklets */
-	for (i = 0; i < NR_CPUS; i++)
-		tasklet_init(&cvm_oct_tasklet[i].t, cvm_oct_tasklet_rx, 0);
+	struct net_device *dev_for_napi = NULL;
+	union cvmx_pow_wq_int_thrx int_thr;
+	union cvmx_pow_wq_int_pc int_pc;
+
+	for (i = 0; i < TOTAL_NUMBER_OF_PORTS; i++) {
+		if (cvm_oct_device[i]) {
+			dev_for_napi = cvm_oct_device[i];
+			break;
+		}
+	}
+
+	if (NULL == dev_for_napi)
+		panic("No net_devices were allocated.");
+
+	if (max_rx_cpus > 1  && max_rx_cpus < num_online_cpus())
+		atomic_set(&core_state.available_cores, max_rx_cpus);
+	else
+		atomic_set(&core_state.available_cores, num_online_cpus());
+	core_state.baseline_cores = atomic_read(&core_state.available_cores);
+
+	core_state.cpu_state = CPU_MASK_NONE;
+	for_each_possible_cpu(i) {
+		netif_napi_add(dev_for_napi, &cvm_oct_napi[i].napi,
+			       cvm_oct_napi_poll, rx_napi_weight);
+		napi_enable(&cvm_oct_napi[i].napi);
+	}
+	/* Register an IRQ hander for to receive POW interrupts */
+	i = request_irq(OCTEON_IRQ_WORKQ0 + pow_receive_group,
+			cvm_oct_do_interrupt, 0, "Ethernet", cvm_oct_device);
+
+	if (i)
+		panic("Could not acquire Ethernet IRQ %d\n",
+		      OCTEON_IRQ_WORKQ0 + pow_receive_group);
+
+	disable_irq_nosync(OCTEON_IRQ_WORKQ0 + pow_receive_group);
+
+	int_thr.u64 = 0;
+	int_thr.s.tc_en = 1;
+	int_thr.s.tc_thr = 1;
+	/* Enable POW interrupt when our port has at least one packet */
+	cvmx_write_csr(CVMX_POW_WQ_INT_THRX(pow_receive_group), int_thr.u64);
+
+	int_pc.u64 = 0;
+	int_pc.s.pc_thr = 5;
+	cvmx_write_csr(CVMX_POW_WQ_INT_PC, int_pc.u64);
+
+
+	/* Scheduld NAPI now.  This will indirectly enable interrupts. */
+	cvm_oct_enable_one_cpu();
 }
 
 void cvm_oct_rx_shutdown(void)
 {
 	int i;
-	/* Shutdown all of the tasklets */
-	for (i = 0; i < NR_CPUS; i++)
-		tasklet_kill(&cvm_oct_tasklet[i].t);
+	/* Shutdown all of the NAPIs */
+	for_each_possible_cpu(i)
+		netif_napi_del(&cvm_oct_napi[i].napi);
 }
author	David S. Miller <davem@davemloft.net>	2010-02-28 19:23:06 -0800
committer	David S. Miller <davem@davemloft.net>	2010-02-28 19:23:06 -0800
commit	47871889c601d8199c51a4086f77eebd77c29b0b (patch)
tree	40cdcac3bff0ee40cc33dcca61d0577cdf965f77 /drivers/staging/octeon/ethernet-rx.c
parent	c16cc0b464b8876cfd57ce1c1dbcb6f9a6a0bce3 (diff)
parent	30ff056c42c665b9ea535d8515890857ae382540 (diff)