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

[PATCH 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.

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 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 00cc915..5628aea 100644
--- a/drivers/net/ethernet/cavium/thunder/nic.h
+++ b/drivers/net/ethernet/cavium/thunder/nic.h
@@ -285,6 +285,7 @@ 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

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

[David Miller]

From: sunil.kovvuri@gmail.com
Date: Mon,  7 Mar 2016 13:05:56 +0530

> 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 00cc915..5628aea 100644
> --- a/drivers/net/ethernet/cavium/thunder/nic.h
> +++ b/drivers/net/ethernet/cavium/thunder/nic.h
> @@ -285,6 +285,7 @@ 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;
>  	}

I do not see how this can sanely work.

By deferring the atomic increment of the page count, you create a window of
time during which the consumer can release the page and prematurely free it.

I'm not applying this, as it looks extremely buggy.  Sorry.

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

[Sunil Kovvuri]

Hi David,

>> you create a window of time during which the consumer
>> can release the page and prematurely free it.
Okay, but here the consumer i.e HW is notified only after page count
is incremented.
For example if you check 'nicvf_refill_rbdr' fn() only after receive
buffer ring is
refilled with buffers, doorbell is issued to hardware to start using
the new buffers.

So
@@ -241,6 +254,8 @@ refill:
                new_rb++;
        }

+       nicvf_get_page(nic);
+
calling 'nicvf_get_page' before issuing doorbell ensures page ref
count is properly
set before hw/consumer can use the buffers.

Thanks,
Sunil.

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

[David Miller]

From: Sunil Kovvuri <sunil.kovvuri@gmail.com>
Date: Mon, 7 Mar 2016 22:28:39 +0530

> Hi David,
> 
>>> you create a window of time during which the consumer
>>> can release the page and prematurely free it.
> Okay, but here the consumer i.e HW is notified only after page count
> is incremented.
> For example if you check 'nicvf_refill_rbdr' fn() only after receive
> buffer ring is
> refilled with buffers, doorbell is issued to hardware to start using
> the new buffers.
> 
> So
> @@ -241,6 +254,8 @@ refill:
>                 new_rb++;
>         }
> 
> +       nicvf_get_page(nic);
> +
> calling 'nicvf_get_page' before issuing doorbell ensures page ref
> count is properly
> set before hw/consumer can use the buffers.

So if you know ahead of time how the page will be split up, just
calculate that when you get the page and increment the page count
appropriately.

That's what we do in the NIU driver.

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

[Sunil Kovvuri]

On Mon, Mar 7, 2016 at 10:34 PM, David Miller <davem@davemloft.net> wrote:
> From: Sunil Kovvuri <sunil.kovvuri@gmail.com>
> Date: Mon, 7 Mar 2016 22:28:39 +0530
>
>> Hi David,
>>
>>>> you create a window of time during which the consumer
>>>> can release the page and prematurely free it.
>> Okay, but here the consumer i.e HW is notified only after page count
>> is incremented.
>> For example if you check 'nicvf_refill_rbdr' fn() only after receive
>> buffer ring is
>> refilled with buffers, doorbell is issued to hardware to start using
>> the new buffers.
>>
>> So
>> @@ -241,6 +254,8 @@ refill:
>>                 new_rb++;
>>         }
>>
>> +       nicvf_get_page(nic);
>> +
>> calling 'nicvf_get_page' before issuing doorbell ensures page ref
>> count is properly
>> set before hw/consumer can use the buffers.
>
> So if you know ahead of time how the page will be split up, just
> calculate that when you get the page and increment the page count
> appropriately.
>
> That's what we do in the NIU driver.

Thanks for the suggestion, will check and get back.

Regards,
Sunil.

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

[Sunil Kovvuri]

Hi David,


>> So if you know ahead of time how the page will be split up, just
>> calculate that when you get the page and increment the page count
>> appropriately.
>>
>> That's what we do in the NIU driver.
>
> Thanks for the suggestion, will check and get back.
>

I looked at the NIU driver and in fn() niu_rbr_refill()
static void niu_rbr_refill(struct niu *np, struct rx_ring_info *rp, gfp_t mask)
{
        int index = rp->rbr_index;

        rp->rbr_pending++;
        if ((rp->rbr_pending % rp->rbr_blocks_per_page) == 0) {

Here it's been checked whether rbr_pending is a exact multiple of page
split count.
And hence updating page count based on fixed calculation is right.

On my platform driver receives a interrupt when free buffer count
falls below a threshold
and by the time SW reads count of buffers to be refilled it can be any
number i.e
may or may not be a exact multiple of page split count.

When interface is being brought down and while freeing unused receive
buffers one
page's count will go wrong and hence will disappear completely. In the
patch i am
updating page->count before issuing doorbell to HW irrespective of
full page utilised
or not. So there is no mismatch while doing put_page(unused buffer).

Thanks,
Sunil.

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

[David Miller]

From: Sunil Kovvuri <sunil.kovvuri@gmail.com>
Date: Thu, 10 Mar 2016 16:13:28 +0530

> Hi David,
> 
> 
>>> So if you know ahead of time how the page will be split up, just
>>> calculate that when you get the page and increment the page count
>>> appropriately.
>>>
>>> That's what we do in the NIU driver.
>>
>> Thanks for the suggestion, will check and get back.
>>
> 
> I looked at the NIU driver and in fn() niu_rbr_refill()
> static void niu_rbr_refill(struct niu *np, struct rx_ring_info *rp, gfp_t mask)
> {
>         int index = rp->rbr_index;
> 
>         rp->rbr_pending++;
>         if ((rp->rbr_pending % rp->rbr_blocks_per_page) == 0) {
> 
> Here it's been checked whether rbr_pending is a exact multiple of page
> split count.
> And hence updating page count based on fixed calculation is right.
> 
> On my platform driver receives a interrupt when free buffer count
> falls below a threshold
> and by the time SW reads count of buffers to be refilled it can be any
> number i.e
> may or may not be a exact multiple of page split count.

So calculate the modulus on the page split count and optimize the
increment ahead of time when possible, and for the sub page split
pieces do it one at a time.

I don't understand what the problem is.

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

[Sunil Kovvuri]

>
> So calculate the modulus on the page split count and optimize the
> increment ahead of time when possible, and for the sub page split
> pieces do it one at a time.
>
Patch does almost the same with a negligible overhead of a counter
for page->_count increment at a later time but still before HW starts
using buffers.
Difference between NIU driver and this patch is there it's
calculate split count, increment page count and then divide page into
buffers. Here it's divide page into buffers, have a counter which increments
at every split and then at the end do a atomic increment of page->_count.

Any issue with this approach ?

Thanks,
Sunil.

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

[David Miller]

From: Sunil Kovvuri <sunil.kovvuri@gmail.com>
Date: Thu, 10 Mar 2016 23:57:48 +0530

> Difference between NIU driver and this patch is there it's
> calculate split count, increment page count and then divide page into
> buffers. Here it's divide page into buffers, have a counter which increments
> at every split and then at the end do a atomic increment of page->_count.
> 
> Any issue with this approach ?

I guess not.

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

[Sunil Kovvuri]

On Fri, Mar 11, 2016 at 1:18 AM, David Miller <davem@davemloft.net> wrote:
> From: Sunil Kovvuri <sunil.kovvuri@gmail.com>
> Date: Thu, 10 Mar 2016 23:57:48 +0530
>
>> Difference between NIU driver and this patch is there it's
>> calculate split count, increment page count and then divide page into
>> buffers. Here it's divide page into buffers, have a counter which increments
>> at every split and then at the end do a atomic increment of page->_count.
>>
>> Any issue with this approach ?
>
> I guess not.

Okay, so can i assume you will consider the patch for merging if no
other comments
are received. Or do you want me to resubmit patches a fresh. ?

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

[David Miller]

From: Sunil Kovvuri <sunil.kovvuri@gmail.com>
Date: Fri, 11 Mar 2016 22:34:19 +0530

> On Fri, Mar 11, 2016 at 1:18 AM, David Miller <davem@davemloft.net> wrote:
>> From: Sunil Kovvuri <sunil.kovvuri@gmail.com>
>> Date: Thu, 10 Mar 2016 23:57:48 +0530
>>
>>> Difference between NIU driver and this patch is there it's
>>> calculate split count, increment page count and then divide page into
>>> buffers. Here it's divide page into buffers, have a counter which increments
>>> at every split and then at the end do a atomic increment of page->_count.
>>>
>>> Any issue with this approach ?
>>
>> I guess not.
> 
> Okay, so can i assume you will consider the patch for merging if no
> other comments
> are received. Or do you want me to resubmit patches a fresh. ?

Please resubmit, thanks a lot.

[PATCH 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 5628aea..c063d92 100644
--- a/drivers/net/ethernet/cavium/thunder/nic.h
+++ b/drivers/net/ethernet/cavium/thunder/nic.h
@@ -263,46 +263,54 @@ 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