[PATCH v2 0/2] net: thunderx: Performance enhancement changes

[PATCH v2 0/2] net: thunderx: Performance enhancement changes

[sunil.kovvuri]

From: Sunil Goutham <sgoutham@cavium.com>

Below patches attempts to improve performance by reducing
no of atomic operations while allocating new receive buffers
and reducing cache misses by adjusting nicvf structure elements.

Changes from v1:
 No changes, resubmitting a fresh as per David's suggestion.

Sunil Goutham (2):
  net: thunderx: Set recevie buffer page usage count in bulk
  net: thunderx: Adjust nicvf structure to reduce cache misses

 drivers/net/ethernet/cavium/thunder/nic.h          |   51 ++++++++++++--------
 drivers/net/ethernet/cavium/thunder/nicvf_queues.c |   31 +++++++++---
 2 files changed, 53 insertions(+), 29 deletions(-)

[PATCH v2 1/2] net: thunderx: Set recevie buffer page usage count in bulk

[sunil.kovvuri]

From: Sunil Goutham <sgoutham@cavium.com>

Instead of calling get_page() for every receive buffer carved out
of page, set page's usage count at the end, to reduce no of atomic
calls.

Signed-off-by: Sunil Goutham <sgoutham@cavium.com>
---
 drivers/net/ethernet/cavium/thunder/nic.h          |    1 +
 drivers/net/ethernet/cavium/thunder/nicvf_queues.c |   31 ++++++++++++++-----
 2 files changed, 24 insertions(+), 8 deletions(-)

diff --git a/drivers/net/ethernet/cavium/thunder/nic.h b/drivers/net/ethernet/cavium/thunder/nic.h
index 092f097..872b22d 100644
--- a/drivers/net/ethernet/cavium/thunder/nic.h
+++ b/drivers/net/ethernet/cavium/thunder/nic.h
@@ -294,6 +294,7 @@ struct nicvf {
 	u32			speed;
 	struct page		*rb_page;
 	u32			rb_page_offset;
+	u16			rb_pageref;
 	bool			rb_alloc_fail;
 	bool			rb_work_scheduled;
 	struct delayed_work	rbdr_work;
diff --git a/drivers/net/ethernet/cavium/thunder/nicvf_queues.c b/drivers/net/ethernet/cavium/thunder/nicvf_queues.c
index 0dd1abf..fa05e34 100644
--- a/drivers/net/ethernet/cavium/thunder/nicvf_queues.c
+++ b/drivers/net/ethernet/cavium/thunder/nicvf_queues.c
@@ -18,6 +18,15 @@
 #include "q_struct.h"
 #include "nicvf_queues.h"
 
+static void nicvf_get_page(struct nicvf *nic)
+{
+	if (!nic->rb_pageref || !nic->rb_page)
+		return;
+
+	atomic_add(nic->rb_pageref, &nic->rb_page->_count);
+	nic->rb_pageref = 0;
+}
+
 /* Poll a register for a specific value */
 static int nicvf_poll_reg(struct nicvf *nic, int qidx,
 			  u64 reg, int bit_pos, int bits, int val)
@@ -81,16 +90,15 @@ static inline int nicvf_alloc_rcv_buffer(struct nicvf *nic, gfp_t gfp,
 	int order = (PAGE_SIZE <= 4096) ?  PAGE_ALLOC_COSTLY_ORDER : 0;
 
 	/* Check if request can be accomodated in previous allocated page */
-	if (nic->rb_page) {
-		if ((nic->rb_page_offset + buf_len + buf_len) >
-		    (PAGE_SIZE << order)) {
-			nic->rb_page = NULL;
-		} else {
-			nic->rb_page_offset += buf_len;
-			get_page(nic->rb_page);
-		}
+	if (nic->rb_page &&
+	    ((nic->rb_page_offset + buf_len) < (PAGE_SIZE << order))) {
+		nic->rb_pageref++;
+		goto ret;
 	}
 
+	nicvf_get_page(nic);
+	nic->rb_page = NULL;
+
 	/* Allocate a new page */
 	if (!nic->rb_page) {
 		nic->rb_page = alloc_pages(gfp | __GFP_COMP | __GFP_NOWARN,
@@ -102,7 +110,9 @@ static inline int nicvf_alloc_rcv_buffer(struct nicvf *nic, gfp_t gfp,
 		nic->rb_page_offset = 0;
 	}
 
+ret:
 	*rbuf = (u64 *)((u64)page_address(nic->rb_page) + nic->rb_page_offset);
+	nic->rb_page_offset += buf_len;
 
 	return 0;
 }
@@ -158,6 +168,9 @@ static int  nicvf_init_rbdr(struct nicvf *nic, struct rbdr *rbdr,
 		desc = GET_RBDR_DESC(rbdr, idx);
 		desc->buf_addr = virt_to_phys(rbuf) >> NICVF_RCV_BUF_ALIGN;
 	}
+
+	nicvf_get_page(nic);
+
 	return 0;
 }
 
@@ -241,6 +254,8 @@ refill:
 		new_rb++;
 	}
 
+	nicvf_get_page(nic);
+
 	/* make sure all memory stores are done before ringing doorbell */
 	smp_wmb();
 
-- 
1.7.1

[PATCH v2 2/2] net: thunderx: Adjust nicvf structure to reduce cache misses

[sunil.kovvuri]

From: Sunil Goutham <sgoutham@cavium.com>

Adjusted nicvf structure such that all elements used in hot
path like napi, xmit e.t.c fall into same cache line. This reduced
no of cache misses and resulted in ~2% increase in no of packets
handled on a core.

Also modified elements with :1 notation to boolean, to be
consistent with other element definitions.

Signed-off-by: Sunil Goutham <sgoutham@cavium.com>
---
 drivers/net/ethernet/cavium/thunder/nic.h |   52 ++++++++++++++++------------
 1 files changed, 30 insertions(+), 22 deletions(-)

diff --git a/drivers/net/ethernet/cavium/thunder/nic.h b/drivers/net/ethernet/cavium/thunder/nic.h
index 872b22d..83025bb 100644
--- a/drivers/net/ethernet/cavium/thunder/nic.h
+++ b/drivers/net/ethernet/cavium/thunder/nic.h
@@ -272,46 +272,54 @@ struct nicvf {
 	struct nicvf		*pnicvf;
 	struct net_device	*netdev;
 	struct pci_dev		*pdev;
+	void __iomem		*reg_base;
+	struct queue_set	*qs;
+	struct nicvf_cq_poll	*napi[8];
 	u8			vf_id;
-	u8			node;
-	u8			tns_mode:1;
-	u8			sqs_mode:1;
-	u8			loopback_supported:1;
+	u8			sqs_id;
+	bool                    sqs_mode;
 	bool			hw_tso;
-	u16			mtu;
-	struct queue_set	*qs;
+
+	/* Receive buffer alloc */
+	u32			rb_page_offset;
+	u16			rb_pageref;
+	bool			rb_alloc_fail;
+	bool			rb_work_scheduled;
+	struct page		*rb_page;
+	struct delayed_work	rbdr_work;
+	struct tasklet_struct	rbdr_task;
+
+	/* Secondary Qset */
+	u8			sqs_count;
 #define	MAX_SQS_PER_VF_SINGLE_NODE		5
 #define	MAX_SQS_PER_VF				11
-	u8			sqs_id;
-	u8			sqs_count; /* Secondary Qset count */
 	struct nicvf		*snicvf[MAX_SQS_PER_VF];
+
+	/* Queue count */
 	u8			rx_queues;
 	u8			tx_queues;
 	u8			max_queues;
-	void __iomem		*reg_base;
+
+	u8			node;
+	u8			cpi_alg;
+	u16			mtu;
 	bool			link_up;
 	u8			duplex;
 	u32			speed;
-	struct page		*rb_page;
-	u32			rb_page_offset;
-	u16			rb_pageref;
-	bool			rb_alloc_fail;
-	bool			rb_work_scheduled;
-	struct delayed_work	rbdr_work;
-	struct tasklet_struct	rbdr_task;
-	struct tasklet_struct	qs_err_task;
-	struct tasklet_struct	cq_task;
-	struct nicvf_cq_poll	*napi[8];
+	bool			tns_mode;
+	bool			loopback_supported;
 	struct nicvf_rss_info	rss_info;
-	u8			cpi_alg;
+	struct tasklet_struct	qs_err_task;
+	struct work_struct	reset_task;
+
 	/* Interrupt coalescing settings */
 	u32			cq_coalesce_usecs;
-
 	u32			msg_enable;
+
+	/* Stats */
 	struct nicvf_hw_stats   hw_stats;
 	struct nicvf_drv_stats  drv_stats;
 	struct bgx_stats	bgx_stats;
-	struct work_struct	reset_task;
 
 	/* MSI-X  */
 	bool			msix_enabled;
-- 
1.7.1

Re: [PATCH v2 0/2] net: thunderx: Performance enhancement changes

[David Miller]

From: sunil.kovvuri@gmail.com
Date: Mon, 14 Mar 2016 16:36:13 +0530

> Below patches attempts to improve performance by reducing
> no of atomic operations while allocating new receive buffers
> and reducing cache misses by adjusting nicvf structure elements.
> 
> Changes from v1:
>  No changes, resubmitting a fresh as per David's suggestion.

Series applied, thanks.