[PATCH net V2] xen-netback: don't move event pointer in TX credit timeout callback

... otherwise the frontend will try to send TX event all the time, even
if no progress can be made. The pointer should only be advanced by the
routine that actually processes the ring (that is, xenvif_poll).

Reported-by: Jacek Konieczny <redacted>
Signed-off-by: Wei Liu <redacted>
Acked-by: Ian Campbell <redacted>
Cc: Paul Durrant <redacted>
---
 drivers/net/xen-netback/netback.c |    2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/drivers/net/xen-netback/netback.c b/drivers/net/xen-netback/netback.c
index 7666540..8e2cbeb 100644
--- a/drivers/net/xen-netback/netback.c
+++ b/drivers/net/xen-netback/netback.c

@@ -658,7 +658,7 @@ void xenvif_check_rx_xenvif(struct xenvif *vif)
 {
 	int more_to_do;
 
-	RING_FINAL_CHECK_FOR_REQUESTS(&vif->tx, more_to_do);
+	more_to_do = RING_HAS_UNCONSUMED_REQUESTS(&vif->tx);
 
 	if (more_to_do)
 		napi_schedule(&vif->napi);

-- 
1.7.10.4

On 05/15/14 13:59, Wei Liu wrote:

Unfortunately, this seems not enough to fix the problem I have reported
here:
http://lists.xenproject.org/archives/html/xen-devel/2014-05/msg01183.html

The dom0 network still stalls when using rate limiting on a VIF
interface after applying this patch to my 3.14.3 kernel (100% CPU#1
usage in the 'soft interrupts').

Greets,
	Jacek

On Thu, May 15, 2014 at 03:04:36PM +0200, Jacek Konieczny wrote:

On 05/15/14 13:59, Wei Liu wrote:

quoted

... otherwise the frontend will try to send TX event all the time, even
if no progress can be made. The pointer should only be advanced by the
routine that actually processes the ring (that is, xenvif_poll).

Reported-by: Jacek Konieczny <redacted>
Signed-off-by: Wei Liu <redacted>
Acked-by: Ian Campbell <redacted>
Cc: Paul Durrant <redacted>
---
 drivers/net/xen-netback/netback.c |    2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/drivers/net/xen-netback/netback.c b/drivers/net/xen-netback/netback.c
index 7666540..8e2cbeb 100644
--- a/drivers/net/xen-netback/netback.c
+++ b/drivers/net/xen-netback/netback.c

@@ -658,7 +658,7 @@ void xenvif_check_rx_xenvif(struct xenvif *vif)
 {
 	int more_to_do;
 
-	RING_FINAL_CHECK_FOR_REQUESTS(&vif->tx, more_to_do);
+	more_to_do = RING_HAS_UNCONSUMED_REQUESTS(&vif->tx);

Unfortunately, this seems not enough to fix the problem I have reported
here:
http://lists.xenproject.org/archives/html/xen-devel/2014-05/msg01183.html

The dom0 network still stalls when using rate limiting on a VIF
interface after applying this patch to my 3.14.3 kernel (100% CPU#1
usage in the 'soft interrupts').

Hmm... I will look into it further.

Wei.

Greets,
	Jacek

On 15/05/14 12:59, Wei Liu wrote:

... otherwise the frontend will try to send TX event all the time, even
if no progress can be made. The pointer should only be advanced by the
routine that actually processes the ring (that is, xenvif_poll).

No it does not.  RING_FINAL_CHECK_FOR_REQUESTS() only advances the event
index if the ring is empty.

This will also result in xenvif_up() failing to properly enable the event.

I think Jacek's bug may be that netback fails to call napi_complete()
when credit is exceeded and there still outstanding requests on the
from-guest ring and thus napi repeatedly polls.

David

On Thu, May 15, 2014 at 03:04:36PM +0200, Jacek Konieczny wrote:

On 05/15/14 13:59, Wei Liu wrote:

quoted

... otherwise the frontend will try to send TX event all the time, even
if no progress can be made. The pointer should only be advanced by the
routine that actually processes the ring (that is, xenvif_poll).

Reported-by: Jacek Konieczny <redacted>
Signed-off-by: Wei Liu <redacted>
Acked-by: Ian Campbell <redacted>
Cc: Paul Durrant <redacted>
---
 drivers/net/xen-netback/netback.c |    2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/drivers/net/xen-netback/netback.c b/drivers/net/xen-netback/netback.c
index 7666540..8e2cbeb 100644
--- a/drivers/net/xen-netback/netback.c
+++ b/drivers/net/xen-netback/netback.c

@@ -658,7 +658,7 @@ void xenvif_check_rx_xenvif(struct xenvif *vif)
 {
 	int more_to_do;
 
-	RING_FINAL_CHECK_FOR_REQUESTS(&vif->tx, more_to_do);
+	more_to_do = RING_HAS_UNCONSUMED_REQUESTS(&vif->tx);

Unfortunately, this seems not enough to fix the problem I have reported
here:
http://lists.xenproject.org/archives/html/xen-devel/2014-05/msg01183.html

The dom0 network still stalls when using rate limiting on a VIF
interface after applying this patch to my 3.14.3 kernel (100% CPU#1
usage in the 'soft interrupts').

Oh I was looking at the aggregated stats that's why the number looked
normal to me. :-/

Greets,
	Jacek

On Thu, May 15, 2014 at 02:40:58PM +0100, David Vrabel wrote:

On 15/05/14 12:59, Wei Liu wrote:

quoted

... otherwise the frontend will try to send TX event all the time, even
if no progress can be made. The pointer should only be advanced by the
routine that actually processes the ring (that is, xenvif_poll).

No it does not.  RING_FINAL_CHECK_FOR_REQUESTS() only advances the event
index if the ring is empty.

This will also result in xenvif_up() failing to properly enable the event.

I think Jacek's bug may be that netback fails to call napi_complete()
when credit is exceeded and there still outstanding requests on the
from-guest ring and thus napi repeatedly polls.

Correct. We should call napi_complete if this vif is rate limited.

Wei.

On Thu, May 15, 2014 at 03:04:36PM +0200, Jacek Konieczny wrote:

On 05/15/14 13:59, Wei Liu wrote:

quoted

... otherwise the frontend will try to send TX event all the time, even
if no progress can be made. The pointer should only be advanced by the
routine that actually processes the ring (that is, xenvif_poll).

Reported-by: Jacek Konieczny <redacted>
Signed-off-by: Wei Liu <redacted>
Acked-by: Ian Campbell <redacted>
Cc: Paul Durrant <redacted>
---
 drivers/net/xen-netback/netback.c |    2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/drivers/net/xen-netback/netback.c b/drivers/net/xen-netback/netback.c
index 7666540..8e2cbeb 100644
--- a/drivers/net/xen-netback/netback.c
+++ b/drivers/net/xen-netback/netback.c

@@ -658,7 +658,7 @@ void xenvif_check_rx_xenvif(struct xenvif *vif)
 {
 	int more_to_do;
 
-	RING_FINAL_CHECK_FOR_REQUESTS(&vif->tx, more_to_do);
+	more_to_do = RING_HAS_UNCONSUMED_REQUESTS(&vif->tx);

Unfortunately, this seems not enough to fix the problem I have reported
here:
http://lists.xenproject.org/archives/html/xen-devel/2014-05/msg01183.html

The dom0 network still stalls when using rate limiting on a VIF
interface after applying this patch to my 3.14.3 kernel (100% CPU#1
usage in the 'soft interrupts').

This is a patch for 3.14.4. I've tested it myself (and looking at the
right stats!) to confirm it works.

---8<---

From a4afed6c44027afff82d6fa7503faef83b01fffe Mon Sep 17 00:00:00 2001

From: Wei Liu <redacted>
Date: Thu, 15 May 2014 15:02:55 +0100
Subject: [PATCH] xen-netback: call napi_complete if vif is rate limited

Reported-by: Jacek Konieczny <redacted>
Signed-off-by: Wei Liu <redacted>
Cc: Ian Campbell <redacted>
Cc: Paul Durrant <redacted>
Cc: David Vrabel <redacted>
---
 drivers/net/xen-netback/common.h    |    2 +-
 drivers/net/xen-netback/interface.c |    5 +++--
 drivers/net/xen-netback/netback.c   |   12 ++++++++----
 3 files changed, 12 insertions(+), 7 deletions(-)

diff --git a/drivers/net/xen-netback/common.h b/drivers/net/xen-netback/common.h
index 4bf5b33..4c018de 100644
--- a/drivers/net/xen-netback/common.h
+++ b/drivers/net/xen-netback/common.h

@@ -219,7 +219,7 @@ void xenvif_check_rx_xenvif(struct xenvif *vif);
 /* Prevent the device from generating any further traffic. */
 void xenvif_carrier_off(struct xenvif *vif);
 
-int xenvif_tx_action(struct xenvif *vif, int budget);
+int xenvif_tx_action(struct xenvif *vif, int budget, bool *rate_limited);
 
 int xenvif_kthread(void *data);
 void xenvif_kick_thread(struct xenvif *vif);

diff --git a/drivers/net/xen-netback/interface.c b/drivers/net/xen-netback/interface.c
index 2e92d52..03cfbd6 100644
--- a/drivers/net/xen-netback/interface.c
+++ b/drivers/net/xen-netback/interface.c

@@ -61,6 +61,7 @@ static int xenvif_poll(struct napi_struct *napi, int budget)
 {
 	struct xenvif *vif = container_of(napi, struct xenvif, napi);
 	int work_done;
+	bool rate_limited;
 
 	/* This vif is rogue, we pretend we've there is nothing to do
 	 * for this vif to deschedule it from NAPI. But this interface

@@ -71,7 +72,7 @@ static int xenvif_poll(struct napi_struct *napi, int budget)
 		return 0;
 	}
 
-	work_done = xenvif_tx_action(vif, budget);
+	work_done = xenvif_tx_action(vif, budget, &rate_limited);
 
 	if (work_done < budget) {
 		int more_to_do = 0;

@@ -96,7 +97,7 @@ static int xenvif_poll(struct napi_struct *napi, int budget)
 		local_irq_save(flags);
 
 		RING_FINAL_CHECK_FOR_REQUESTS(&vif->tx, more_to_do);
-		if (!more_to_do)
+		if (!more_to_do || rate_limited)
 			__napi_complete(napi);
 
 		local_irq_restore(flags);

diff --git a/drivers/net/xen-netback/netback.c b/drivers/net/xen-netback/netback.c
index 97030c1..fa7d5da 100644
--- a/drivers/net/xen-netback/netback.c
+++ b/drivers/net/xen-netback/netback.c

@@ -1125,12 +1125,14 @@ static bool tx_credit_exceeded(struct xenvif *vif, unsigned size)
 	return false;
 }
 
-static unsigned xenvif_tx_build_gops(struct xenvif *vif, int budget)
+static unsigned xenvif_tx_build_gops(struct xenvif *vif, int budget,
+				     bool *rate_limited)
 {
 	struct gnttab_copy *gop = vif->tx_copy_ops, *request_gop;
 	struct sk_buff *skb;
 	int ret;
 
+	*rate_limited = false;
 	while ((nr_pending_reqs(vif) + XEN_NETBK_LEGACY_SLOTS_MAX
 		< MAX_PENDING_REQS) &&
 	       (skb_queue_len(&vif->tx_queue) < budget)) {

@@ -1165,8 +1167,10 @@ static unsigned xenvif_tx_build_gops(struct xenvif *vif, int budget)
 
 		/* Credit-based scheduling. */
 		if (txreq.size > vif->remaining_credit &&
-		    tx_credit_exceeded(vif, txreq.size))
+		    tx_credit_exceeded(vif, txreq.size)) {
+			*rate_limited = true;
 			break;
+		}
 
 		vif->remaining_credit -= txreq.size;
 

@@ -1382,7 +1386,7 @@ static int xenvif_tx_submit(struct xenvif *vif)
 }
 
 /* Called after netfront has transmitted */
-int xenvif_tx_action(struct xenvif *vif, int budget)
+int xenvif_tx_action(struct xenvif *vif, int budget, bool *rate_limited)
 {
 	unsigned nr_gops;
 	int work_done;

@@ -1390,7 +1394,7 @@ int xenvif_tx_action(struct xenvif *vif, int budget)
 	if (unlikely(!tx_work_todo(vif)))
 		return 0;
 
-	nr_gops = xenvif_tx_build_gops(vif, budget);
+	nr_gops = xenvif_tx_build_gops(vif, budget, rate_limited);
 
 	if (nr_gops == 0)
 		return 0;

-- 
1.7.10.4


> Greets,
> 	Jacek

On 15/05/14 15:13, Wei Liu wrote:

How about calling timer_pending(&vif->credit_timeout) instead?
Also, can this __napi_complete and the callback's napi_schedule race 
with each other? When napi_complete is between removing from the list 
and clearing the bit, and napi_schedule is just test&set the bit, the 
latter won't add the instance to the list again

On Thu, May 15, 2014 at 03:47:38PM +0100, Zoltan Kiss wrote:

On 15/05/14 15:13, Wei Liu wrote:

quoted

On Thu, May 15, 2014 at 03:04:36PM +0200, Jacek Konieczny wrote:

quoted

On 05/15/14 13:59, Wei Liu wrote:

quoted

... otherwise the frontend will try to send TX event all the time, even
if no progress can be made. The pointer should only be advanced by the
routine that actually processes the ring (that is, xenvif_poll).

Reported-by: Jacek Konieczny <redacted>
Signed-off-by: Wei Liu <redacted>
Acked-by: Ian Campbell <redacted>
Cc: Paul Durrant <redacted>
---
 drivers/net/xen-netback/netback.c |    2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/drivers/net/xen-netback/netback.c b/drivers/net/xen-netback/netback.c
index 7666540..8e2cbeb 100644
--- a/drivers/net/xen-netback/netback.c
+++ b/drivers/net/xen-netback/netback.c

@@ -658,7 +658,7 @@ void xenvif_check_rx_xenvif(struct xenvif *vif)
 {
 	int more_to_do;

-	RING_FINAL_CHECK_FOR_REQUESTS(&vif->tx, more_to_do);
+	more_to_do = RING_HAS_UNCONSUMED_REQUESTS(&vif->tx);

Unfortunately, this seems not enough to fix the problem I have reported
here:
http://lists.xenproject.org/archives/html/xen-devel/2014-05/msg01183.html

The dom0 network still stalls when using rate limiting on a VIF
interface after applying this patch to my 3.14.3 kernel (100% CPU#1
usage in the 'soft interrupts').

This is a patch for 3.14.4. I've tested it myself (and looking at the
right stats!) to confirm it works.

---8<---
From a4afed6c44027afff82d6fa7503faef83b01fffe Mon Sep 17 00:00:00 2001
From: Wei Liu <redacted>
Date: Thu, 15 May 2014 15:02:55 +0100
Subject: [PATCH] xen-netback: call napi_complete if vif is rate limited

Reported-by: Jacek Konieczny <redacted>
Signed-off-by: Wei Liu <redacted>
Cc: Ian Campbell <redacted>
Cc: Paul Durrant <redacted>
Cc: David Vrabel <redacted>
---
 drivers/net/xen-netback/common.h    |    2 +-
 drivers/net/xen-netback/interface.c |    5 +++--
 drivers/net/xen-netback/netback.c   |   12 ++++++++----
 3 files changed, 12 insertions(+), 7 deletions(-)

diff --git a/drivers/net/xen-netback/common.h b/drivers/net/xen-netback/common.h
index 4bf5b33..4c018de 100644
--- a/drivers/net/xen-netback/common.h
+++ b/drivers/net/xen-netback/common.h

@@ -219,7 +219,7 @@ void xenvif_check_rx_xenvif(struct xenvif *vif);
 /* Prevent the device from generating any further traffic. */
 void xenvif_carrier_off(struct xenvif *vif);

-int xenvif_tx_action(struct xenvif *vif, int budget);
+int xenvif_tx_action(struct xenvif *vif, int budget, bool *rate_limited);

 int xenvif_kthread(void *data);
 void xenvif_kick_thread(struct xenvif *vif);

diff --git a/drivers/net/xen-netback/interface.c b/drivers/net/xen-netback/interface.c
index 2e92d52..03cfbd6 100644
--- a/drivers/net/xen-netback/interface.c
+++ b/drivers/net/xen-netback/interface.c

@@ -61,6 +61,7 @@ static int xenvif_poll(struct napi_struct *napi, int budget)
 {
 	struct xenvif *vif = container_of(napi, struct xenvif, napi);
 	int work_done;
+	bool rate_limited;

 	/* This vif is rogue, we pretend we've there is nothing to do
 	 * for this vif to deschedule it from NAPI. But this interface

@@ -71,7 +72,7 @@ static int xenvif_poll(struct napi_struct *napi, int budget)
 		return 0;
 	}

-	work_done = xenvif_tx_action(vif, budget);
+	work_done = xenvif_tx_action(vif, budget, &rate_limited);

 	if (work_done < budget) {
 		int more_to_do = 0;

@@ -96,7 +97,7 @@ static int xenvif_poll(struct napi_struct *napi, int budget)
 		local_irq_save(flags);

 		RING_FINAL_CHECK_FOR_REQUESTS(&vif->tx, more_to_do);
-		if (!more_to_do)
+		if (!more_to_do || rate_limited)

How about calling timer_pending(&vif->credit_timeout) instead?

timer_pending(&vif->credit_timeout) covers only one of two senarios of
"credit exceeded", see tx_credit_exceeded.

Also, can this __napi_complete and the callback's napi_schedule race with
each other? When napi_complete is between removing from the list and
clearing the bit, and napi_schedule is just test&set the bit, the latter
won't add the instance to the list again

I think it should be fine. How is it different from what we already have
now? Is this something similar to what David once posted?

  [off-list ref]

Wei.

On 15/05/14 16:30, Wei Liu wrote:

On Thu, May 15, 2014 at 03:47:38PM +0100, Zoltan Kiss wrote:

quoted

On 15/05/14 15:13, Wei Liu wrote:

quoted

On Thu, May 15, 2014 at 03:04:36PM +0200, Jacek Konieczny wrote:

quoted

On 05/15/14 13:59, Wei Liu wrote:

quoted

... otherwise the frontend will try to send TX event all the time, even
if no progress can be made. The pointer should only be advanced by the
routine that actually processes the ring (that is, xenvif_poll).

Reported-by: Jacek Konieczny <redacted>
Signed-off-by: Wei Liu <redacted>
Acked-by: Ian Campbell <redacted>
Cc: Paul Durrant <redacted>
---
  drivers/net/xen-netback/netback.c |    2 +-
  1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/drivers/net/xen-netback/netback.c b/drivers/net/xen-netback/netback.c
index 7666540..8e2cbeb 100644
--- a/drivers/net/xen-netback/netback.c
+++ b/drivers/net/xen-netback/netback.c

@@ -658,7 +658,7 @@ void xenvif_check_rx_xenvif(struct xenvif *vif)
  {
  	int more_to_do;

-	RING_FINAL_CHECK_FOR_REQUESTS(&vif->tx, more_to_do);
+	more_to_do = RING_HAS_UNCONSUMED_REQUESTS(&vif->tx);

Unfortunately, this seems not enough to fix the problem I have reported
here:
http://lists.xenproject.org/archives/html/xen-devel/2014-05/msg01183.html

The dom0 network still stalls when using rate limiting on a VIF
interface after applying this patch to my 3.14.3 kernel (100% CPU#1
usage in the 'soft interrupts').

This is a patch for 3.14.4. I've tested it myself (and looking at the
right stats!) to confirm it works.

---8<---
 From a4afed6c44027afff82d6fa7503faef83b01fffe Mon Sep 17 00:00:00 2001
From: Wei Liu <redacted>
Date: Thu, 15 May 2014 15:02:55 +0100
Subject: [PATCH] xen-netback: call napi_complete if vif is rate limited

Reported-by: Jacek Konieczny <redacted>
Signed-off-by: Wei Liu <redacted>
Cc: Ian Campbell <redacted>
Cc: Paul Durrant <redacted>
Cc: David Vrabel <redacted>
---
  drivers/net/xen-netback/common.h    |    2 +-
  drivers/net/xen-netback/interface.c |    5 +++--
  drivers/net/xen-netback/netback.c   |   12 ++++++++----
  3 files changed, 12 insertions(+), 7 deletions(-)

diff --git a/drivers/net/xen-netback/common.h b/drivers/net/xen-netback/common.h
index 4bf5b33..4c018de 100644
--- a/drivers/net/xen-netback/common.h
+++ b/drivers/net/xen-netback/common.h

@@ -219,7 +219,7 @@ void xenvif_check_rx_xenvif(struct xenvif *vif);
  /* Prevent the device from generating any further traffic. */
  void xenvif_carrier_off(struct xenvif *vif);

-int xenvif_tx_action(struct xenvif *vif, int budget);
+int xenvif_tx_action(struct xenvif *vif, int budget, bool *rate_limited);

  int xenvif_kthread(void *data);
  void xenvif_kick_thread(struct xenvif *vif);

diff --git a/drivers/net/xen-netback/interface.c b/drivers/net/xen-netback/interface.c
index 2e92d52..03cfbd6 100644
--- a/drivers/net/xen-netback/interface.c
+++ b/drivers/net/xen-netback/interface.c

@@ -61,6 +61,7 @@ static int xenvif_poll(struct napi_struct *napi, int budget)
  {
  	struct xenvif *vif = container_of(napi, struct xenvif, napi);
  	int work_done;
+	bool rate_limited;

  	/* This vif is rogue, we pretend we've there is nothing to do
  	 * for this vif to deschedule it from NAPI. But this interface

@@ -71,7 +72,7 @@ static int xenvif_poll(struct napi_struct *napi, int budget)
  		return 0;
  	}

-	work_done = xenvif_tx_action(vif, budget);
+	work_done = xenvif_tx_action(vif, budget, &rate_limited);

  	if (work_done < budget) {
  		int more_to_do = 0;

@@ -96,7 +97,7 @@ static int xenvif_poll(struct napi_struct *napi, int budget)
  		local_irq_save(flags);

  		RING_FINAL_CHECK_FOR_REQUESTS(&vif->tx, more_to_do);
-		if (!more_to_do)
+		if (!more_to_do || rate_limited)

How about calling timer_pending(&vif->credit_timeout) instead?

timer_pending(&vif->credit_timeout) covers only one of two senarios of
"credit exceeded", see tx_credit_exceeded.

The other scenario is when the packet size exceeds the credit. There is 
no packet here actually, we just want to know if this vif ran out of 
credit and waiting for the timer to fire.

quoted

Also, can this __napi_complete and the callback's napi_schedule race with
each other? When napi_complete is between removing from the list and
clearing the bit, and napi_schedule is just test&set the bit, the latter
won't add the instance to the list again

I think it should be fine. How is it different from what we already have
now? Is this something similar to what David once posted?

   [off-list ref]

Unfortunately that discussion stalled, and my question were not 
answered, so I bumped it again. But that's different a bit: it was about 
racing between the NAPI instance (running in softirq context) and the 
interrupt. Here the danger is that the NAPI instance and the softirq can 
race. They both run in softirq context, and even if they were originally 
on the same CPU, I'm sure if the instance move somewhere else, the timer 
doesn't follow it.

Zoli

On Thu, May 15, 2014 at 05:34:09PM +0100, Zoltan Kiss wrote:
[...]

quoted

quoted

quoted

 		RING_FINAL_CHECK_FOR_REQUESTS(&vif->tx, more_to_do);
-		if (!more_to_do)
+		if (!more_to_do || rate_limited)

How about calling timer_pending(&vif->credit_timeout) instead?

timer_pending(&vif->credit_timeout) covers only one of two senarios of
"credit exceeded", see tx_credit_exceeded.

The other scenario is when the packet size exceeds the credit. There is no
packet here actually, we just want to know if this vif ran out of credit and
waiting for the timer to fire.

quoted

Which place are you referring to? There's packet in the ring, right? So
you're saying in xenvif_poll "more_to_do" is true and "timer_pending" is
also true when we come to xenvif_poll again?

quoted

quoted

Also, can this __napi_complete and the callback's napi_schedule race with
each other? When napi_complete is between removing from the list and
clearing the bit, and napi_schedule is just test&set the bit, the latter
won't add the instance to the list again

I think it should be fine. How is it different from what we already have
now? Is this something similar to what David once posted?

  [off-list ref]

Unfortunately that discussion stalled, and my question were not answered, so
I bumped it again. But that's different a bit: it was about racing between
the NAPI instance (running in softirq context) and the interrupt. Here the
danger is that the NAPI instance and the softirq can race. They both run in
softirq context, and even if they were originally on the same CPU, I'm sure
if the instance move somewhere else, the timer doesn't follow it.

This comes back to that original question, doesn't it? That's NAPI
running on CPU A and raised by CPU B.

Wei.

Zoli

On 15/05/14 17:53, Wei Liu wrote:

On Thu, May 15, 2014 at 05:34:09PM +0100, Zoltan Kiss wrote:
[...]

quoted

quoted

quoted

quoted

  		RING_FINAL_CHECK_FOR_REQUESTS(&vif->tx, more_to_do);
-		if (!more_to_do)
+		if (!more_to_do || rate_limited)

How about calling timer_pending(&vif->credit_timeout) instead?

timer_pending(&vif->credit_timeout) covers only one of two senarios of
"credit exceeded", see tx_credit_exceeded.

The other scenario is when the packet size exceeds the credit. There is no
packet here actually, we just want to know if this vif ran out of credit and
waiting for the timer to fire.

quoted

Which place are you referring to? There's packet in the ring, right? So
you're saying in xenvif_poll "more_to_do" is true and "timer_pending" is
also true when we come to xenvif_poll again?

The goal of this patch to deschedule NAPI if the vif ran out of credit. 
Either you can carry that information from build_gops via a bool, or you 
can check whether the timer is pending. That's what tx_credit_exceeded 
does as well, and then it checks if the actual packet fits in. But in 
xenvif_poll you don't want to know whether an actual packet fits in, you 
only need the information whether tx_credit_exceeded started the timer 
or not.
If it is, you can be sure there is no more credit. If not, you can keep 
the instance running.

quoted

quoted

quoted

Also, can this __napi_complete and the callback's napi_schedule race with
each other? When napi_complete is between removing from the list and
clearing the bit, and napi_schedule is just test&set the bit, the latter
won't add the instance to the list again

I think it should be fine. How is it different from what we already have
now? Is this something similar to what David once posted?

   [off-list ref]

Unfortunately that discussion stalled, and my question were not answered, so
I bumped it again. But that's different a bit: it was about racing between
the NAPI instance (running in softirq context) and the interrupt. Here the
danger is that the NAPI instance and the softirq can race. They both run in
softirq context, and even if they were originally on the same CPU, I'm sure
if the instance move somewhere else, the timer doesn't follow it.

This comes back to that original question, doesn't it? That's NAPI
running on CPU A and raised by CPU B.

Wei.

quoted

Zoli

On Thu, May 15, 2014 at 06:03:14PM +0100, Zoltan Kiss wrote:

On 15/05/14 17:53, Wei Liu wrote:

quoted

On Thu, May 15, 2014 at 05:34:09PM +0100, Zoltan Kiss wrote:
[...]

quoted

quoted

quoted

quoted

 		RING_FINAL_CHECK_FOR_REQUESTS(&vif->tx, more_to_do);
-		if (!more_to_do)
+		if (!more_to_do || rate_limited)

How about calling timer_pending(&vif->credit_timeout) instead?

timer_pending(&vif->credit_timeout) covers only one of two senarios of
"credit exceeded", see tx_credit_exceeded.

The other scenario is when the packet size exceeds the credit. There is no
packet here actually, we just want to know if this vif ran out of credit and
waiting for the timer to fire.

quoted

Which place are you referring to? There's packet in the ring, right? So
you're saying in xenvif_poll "more_to_do" is true and "timer_pending" is
also true when we come to xenvif_poll again?

The goal of this patch to deschedule NAPI if the vif ran out of credit.
Either you can carry that information from build_gops via a bool, or you can
check whether the timer is pending. That's what tx_credit_exceeded does as
well, and then it checks if the actual packet fits in. But in xenvif_poll
you don't want to know whether an actual packet fits in, you only need the
information whether tx_credit_exceeded started the timer or not.
If it is, you can be sure there is no more credit. If not, you can keep the
instance running.

quoted

OK, you convinced me. I will try to rework this patch with your
approach.

Wei.