Skip to contentSkip to conversation

kni: fix possible alloc_q starvation when mbufs are exhausted

PATCH Superseded Yangchao Zhou 4 participants 10 messages Nov 9, 2022 – Mar 11, 2023

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 kni: fix possible alloc_q starvation when mbufs are exhausted

In some scenarios, mbufs returned by rte_kni_rx_burst are not freed immediately. So kni_allocate_mbufs may be failed, but we don't know.

Even worse, when alloc_q is completely exhausted, kni_net_tx in rte_kni.ko will drop all tx packets. kni_allocate_mbufs is never called again, even if the mbufs are eventually freed.

In this patch, we always try to allocate mbufs for alloc_q.

Don't worry about alloc_q being allocated too many mbufs, in fact, the old logic will gradually fill up alloc_q.
Also, the cost of more calls to kni_allocate_mbufs should be acceptable.

Fixes 3e12a98fe397 ("kni: optimize Rx burst")
---
 lib/kni/rte_kni.c | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)
lib/kni/rte_kni.c +3 -3
index 8ab6c47153..265c5262f7 100644
@@ -634,9 +634,9 @@ rte_kni_rx_burst(struct rte_kni *kni, struct rte_mbuf **mbufs, unsigned int num)
 {
 	unsigned int ret = kni_fifo_get(kni->tx_q, (void **)mbufs, num);
 
-	/* If buffers removed, allocate mbufs and then put them into alloc_q */
-	if (ret)
-		kni_allocate_mbufs(kni);
+	/* Always try to allocate mbufs for alloc_q to avoid starvation when
+	 * kni->pktmbuf_pool is exhausted. */
+	kni_allocate_mbufs(kni);
 
 	return ret;
 }
Signature · 1 line
2.25.1
#1 lore raw

PATCH v2 1/1 kni: fix possible alloc_q starvation when mbufs are exhausted

In some scenarios, mbufs returned by rte_kni_rx_burst are not freed immediately. So kni_allocate_mbufs may be failed, but we don't know.

Even worse, when alloc_q is completely exhausted, kni_net_tx in rte_kni.ko will drop all tx packets. kni_allocate_mbufs is never called again, even if the mbufs are eventually freed.

In this patch, we always try to allocate mbufs for alloc_q.

Don't worry about alloc_q being allocated too many mbufs, in fact, the old logic will gradually fill up alloc_q.
Also, the cost of more calls to kni_allocate_mbufs should be acceptable.

Fixes 3e12a98fe397 ("kni: optimize Rx burst")
---
 lib/kni/rte_kni.c | 8 +++++---
 1 file changed, 5 insertions(+), 3 deletions(-)
lib/kni/rte_kni.c +5 -3
index 8ab6c47153..ea9a507de3 100644
@@ -634,9 +634,11 @@ rte_kni_rx_burst(struct rte_kni *kni, struct rte_mbuf **mbufs, unsigned int num)
 {
 	unsigned int ret = kni_fifo_get(kni->tx_q, (void **)mbufs, num);
 
-	/* If buffers removed, allocate mbufs and then put them into alloc_q */
-	if (ret)
-		kni_allocate_mbufs(kni);
+	/*
+	 * Always try to allocate mbufs for alloc_q to avoid starvation when
+	 * kni->pktmbuf_pool is exhausted.
+	 */
+	kni_allocate_mbufs(kni);
 
 	return ret;
 }
Signature · 1 line
2.25.1
#2 lore raw
Quoted 13 lines · On Wed, 9 Nov 2022 14:04:34 +0800 Yangchao Zhou <[email protected]> wrote:

In some scenarios, mbufs returned by rte_kni_rx_burst are not freed immediately. So kni_allocate_mbufs may be failed, but we don't know.

Even worse, when alloc_q is completely exhausted, kni_net_tx in rte_kni.ko will drop all tx packets. kni_allocate_mbufs is never called again, even if the mbufs are eventually freed.

In this patch, we always try to allocate mbufs for alloc_q.

Don't worry about alloc_q being allocated too many mbufs, in fact, the old logic will gradually fill up alloc_q.
Also, the cost of more calls to kni_allocate_mbufs should be acceptable.

Quoted 5 lines
Fixes 3e12a98fe397 ("kni: optimize Rx burst")

Since fifo_get returning 0 (no buffers) is very common would this change impact performance.

If the problem is pool draining might be better to make the pool bigger.

author
#3 lore raw
Quoted 16 lines · Stephen Hemminger
Quoted 13 lines · On Wed, 9 Nov 2022 14:04:34 +0800 Yangchao Zhou <[email protected]> wrote:

In some scenarios, mbufs returned by rte_kni_rx_burst are not freed immediately. So kni_allocate_mbufs may be failed, but we don't know.

Even worse, when alloc_q is completely exhausted, kni_net_tx in rte_kni.ko will drop all tx packets. kni_allocate_mbufs is never called again, even if the mbufs are eventually freed.

In this patch, we always try to allocate mbufs for alloc_q.

Don't worry about alloc_q being allocated too many mbufs, in fact, the old logic will gradually fill up alloc_q.
Also, the cost of more calls to kni_allocate_mbufs should be acceptable.

Quoted 9 lines
Quoted 5 lines
Fixes 3e12a98fe397 ("kni: optimize Rx burst")

Since fifo_get returning 0 (no buffers) is very common would this change impact performance.

It does add a little cost, but there is no extra mbuf allocation and deallocation.

Quoted 4 lines

If the problem is pool draining might be better to make the pool bigger.

Yes, using a larger pool can avoid this problem. But this may lead to resource wastage and full resource calculation is a challenge for developers as it involves to mempool caching mechanism, IP fragment cache, ARP cache, NIC txq, other transit queue, etc.

The mbuf allocation failure may also occur on many NIC drivers, but if the mbuf allocation fails, the mbuf is not taken out so that it can be recovered after a retry later.
KNI currently does not have such a takedown and recovery mechanism. It is also possible to consider implementing something similar to the NIC driver, but with more changes and other overheads.

#4 lore raw
Quoted 26 lines · Matt

On Thu, Nov 10, 2022 at 12:39 AM Stephen Hemminger <[email protected] <mailto:[email protected]>> wrote:

On Wed,  9 Nov 2022 14:04:34 +0800
Yangchao Zhou <[email protected] <mailto:[email protected]>> wrote:

> In some scenarios, mbufs returned by rte_kni_rx_burst are not freed > immediately. So kni_allocate_mbufs may be failed, but we don't know. >
> Even worse, when alloc_q is completely exhausted, kni_net_tx in > rte_kni.ko will drop all tx packets. kni_allocate_mbufs is never > called again, even if the mbufs are eventually freed.
>
> In this patch, we always try to allocate mbufs for alloc_q. >
> Don't worry about alloc_q being allocated too many mbufs, in fact, > the old logic will gradually fill up alloc_q.
> Also, the cost of more calls to kni_allocate_mbufs should be acceptable.
>
> Fixes: 3e12a98fe397 ("kni: optimize Rx burst")
> Cc: [email protected] <mailto:[email protected]>
> Cc: [email protected] <mailto:[email protected]>
>

Quoted 9 lines

> Signed-off-by: Yangchao Zhou <[email protected] <mailto:[email protected]>>

Since fifo_get returning 0 (no buffers) is very common would this change impact performance.

It does add a little cost, but there is no extra mbuf allocation and deallocation.

It is common that 'rte_kni_rx_burst()' called in continuous loop, I expect to try to fill alloc_q in each call impacts performance, even no mbuf allocated at all.

Quoted 9 lines

If the problem is pool draining might be better to make the pool bigger.

Yes, using a larger pool can avoid this problem. But this may lead to resource wastage and full resource calculation is a challenge for developers as it involves to mempool caching mechanism, IP fragment cache, ARP cache, NIC txq, other transit queue, etc.

Quoted 6 lines

The mbuf allocation failure may also occur on many NIC drivers, but if the mbuf allocation fails, the mbuf is not taken out so that it can be recovered after a retry later.
KNI currently does not have such a takedown and recovery mechanism. It is also possible to consider implementing something similar to the NIC driver, but with more changes and other overheads.

I agree this can cause starvation and recovery mechanism is missing,

Following may be an option, but still concerned about performance impact and not sure about upstreaming it. Other solution can be to use larger pool as Stephen suggested.

rte_kni_rx_burst()

	ret = kni_fifo_get()

	if (ret & kni_fifo_count(alloc_q) == 0)
		kni_allocate_mbufs()
#5 lore raw

PATCH v3 1/1 kni: fix possible alloc_q starvation when mbufs are exhausted

In reply to Yangchao Zhou

In some scenarios, mbufs returned by rte_kni_rx_burst are not freed immediately. So kni_allocate_mbufs may be failed, but we don't know.

Even worse, when alloc_q is completely exhausted, kni_net_tx in rte_kni.ko will drop all tx packets. kni_allocate_mbufs is never called again, even if the mbufs are eventually freed.

In this patch, we try to allocate mbufs for alloc_q when it is empty.

According to historical experience, the performance bottleneck of KNI is offen the usleep_range of kni thread in rte_kni.ko.
The check of kni_fifo_count is trivial and the cost should be acceptable.

Fixes 3e12a98fe397 ("kni: optimize Rx burst")
---
 lib/kni/rte_kni.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
lib/kni/rte_kni.c +2 -2
index 8ab6c47153..bfa6a001ff 100644
@@ -634,8 +634,8 @@ rte_kni_rx_burst(struct rte_kni *kni, struct rte_mbuf **mbufs, unsigned int num)
 {
 	unsigned int ret = kni_fifo_get(kni->tx_q, (void **)mbufs, num);
 
-	/* If buffers removed, allocate mbufs and then put them into alloc_q */
-	if (ret)
+	/* If buffers removed or alloc_q is empty, allocate mbufs and then put them into alloc_q */
+	if (ret || (kni_fifo_count(kni->alloc_q) == 0))
 		kni_allocate_mbufs(kni);
 
 	return ret;
Signature · 1 line
2.25.1
#6 lore raw
Quoted 7 lines · Yangchao Zhou

In some scenarios, mbufs returned by rte_kni_rx_burst are not freed immediately. So kni_allocate_mbufs may be failed, but we don't know.

Even worse, when alloc_q is completely exhausted, kni_net_tx in rte_kni.ko will drop all tx packets. kni_allocate_mbufs is never called again, even if the mbufs are eventually freed.

Quoted 6 lines

In this patch, we try to allocate mbufs for alloc_q when it is empty.

According to historical experience, the performance bottleneck of KNI is offen the usleep_range of kni thread in rte_kni.ko.
The check of kni_fifo_count is trivial and the cost should be acceptable.

Hi Yangchao,

Are you observing any performance impact with this change in you use case?

Quoted 23 lines
Fixes 3e12a98fe397 ("kni: optimize Rx burst")
---
 lib/kni/rte_kni.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
lib/kni/rte_kni.c +2 -2
index 8ab6c47153..bfa6a001ff 100644
@@ -634,8 +634,8 @@ rte_kni_rx_burst(struct rte_kni *kni, struct rte_mbuf **mbufs, unsigned int num)
 {
 	unsigned int ret = kni_fifo_get(kni->tx_q, (void **)mbufs, num);
 
-	/* If buffers removed, allocate mbufs and then put them into alloc_q */
-	if (ret)
+	/* If buffers removed or alloc_q is empty, allocate mbufs and then put them into alloc_q */
+	if (ret || (kni_fifo_count(kni->alloc_q) == 0))
 		kni_allocate_mbufs(kni);
 
 	return ret;
author
#7 lore raw

Hi Ferruh,

In my case, the traffic is not large, so I can't see the impact. I also tested under high load(>2Mpps with 2 DPDK cores and 2 kernel threads) and found no significant difference in performance either.
I think the reason should be that it will not
run to 'kni_fifo_count(kni->alloc_q) == 0' under high load.

Quoted 48 lines · Ferruh Yigit
Quoted 7 lines · Yangchao Zhou

In some scenarios, mbufs returned by rte_kni_rx_burst are not freed immediately. So kni_allocate_mbufs may be failed, but we don't know.

Even worse, when alloc_q is completely exhausted, kni_net_tx in rte_kni.ko will drop all tx packets. kni_allocate_mbufs is never called again, even if the mbufs are eventually freed.

Quoted 6 lines

In this patch, we try to allocate mbufs for alloc_q when it is empty.

According to historical experience, the performance bottleneck of KNI is offen the usleep_range of kni thread in rte_kni.ko.
The check of kni_fifo_count is trivial and the cost should be acceptable.

Hi Yangchao,

Are you observing any performance impact with this change in you use case?

lib/kni/rte_kni.c rte_kni_rx_burst(struct rte_kni *kni, struct

Quoted 13 lines
Fixes 3e12a98fe397 ("kni: optimize Rx burst")
---
 lib/kni/rte_kni.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
lib/kni/rte_kni.c +0 -0
index 8ab6c47153..bfa6a001ff 100644
@@ -634,8 +634,8 @@ rte_kni_rx_burst(struct rte_kni *kni, struct

rte_mbuf **mbufs, unsigned int num)

lib/kni/rte_kni.c rte_kni_rx_burst(struct rte_kni *kni, struct rte_mbuf **mbufs, unsigned int num)

Quoted 4 lines
lib/kni/rte_kni.c +0 -1
 {
      unsigned int ret = kni_fifo_get(kni->tx_q, (void **)mbufs, num);
 
-     /* If buffers removed, allocate mbufs and then put them into

alloc_q */

lib/kni/rte_kni.c rte_kni_rx_burst(struct rte_kni *kni, struct · line 637

Quoted 2 lines
lib/kni/rte_kni.c +1 -1
-     if (ret)
+     /* If buffers removed or alloc_q is empty, allocate mbufs and then

put them into alloc_q */

Quoted 4 lines
lib/kni/rte_kni.c +1 -0
+     if (ret || (kni_fifo_count(kni->alloc_q) == 0))
              kni_allocate_mbufs(kni);
 
      return ret;
Acked-by
#8 lore raw
Quoted 8 lines · Matt

Hi Ferruh,

In my case, the traffic is not large, so I can't see the impact. I also tested under high load(>2Mpps with 2 DPDK cores and 2 kernel threads) and found no significant difference in performance either.
I think the reason should be that it will not
run to 'kni_fifo_count(kni->alloc_q) == 0' under high load.

I agree, additional check most likely hit on the low bandwidth, thanks for checking for performance impact.

Quoted 25 lines

On Tue, Jan 3, 2023 at 8:47 PM Ferruh Yigit <[email protected] <mailto:[email protected]>> wrote:

On 12/30/2022 4:23 AM, Yangchao Zhou wrote:
> In some scenarios, mbufs returned by rte_kni_rx_burst are not freed > immediately. So kni_allocate_mbufs may be failed, but we don't know. >
> Even worse, when alloc_q is completely exhausted, kni_net_tx in > rte_kni.ko will drop all tx packets. kni_allocate_mbufs is never > called again, even if the mbufs are eventually freed.
>
> In this patch, we try to allocate mbufs for alloc_q when it is empty. >
> According to historical experience, the performance bottleneck of KNI > is offen the usleep_range of kni thread in rte_kni.ko.
> The check of kni_fifo_count is trivial and the cost should be acceptable.
>

Hi Yangchao,

Are you observing any performance impact with this change in you use case?

Quoted 5 lines

> Fixes: 3e12a98fe397 ("kni: optimize Rx burst") > Cc: [email protected] <mailto:[email protected]> >
> Signed-off-by: Yangchao Zhou <[email protected] <mailto:[email protected]>>

#9 lore raw

04/01/2023 15:34, Ferruh Yigit:

Quoted 38 lines · Matt
Quoted 8 lines

Hi Ferruh,

In my case, the traffic is not large, so I can't see the impact. I also tested under high load(>2Mpps with 2 DPDK cores and 2 kernel threads) and found no significant difference in performance either.
I think the reason should be that it will not
run to 'kni_fifo_count(kni->alloc_q) == 0' under high load.

I agree, additional check most likely hit on the low bandwidth, thanks for checking for performance impact.

Quoted 25 lines

On Tue, Jan 3, 2023 at 8:47 PM Ferruh Yigit <[email protected] <mailto:[email protected]>> wrote:

On 12/30/2022 4:23 AM, Yangchao Zhou wrote:
> In some scenarios, mbufs returned by rte_kni_rx_burst are not freed > immediately. So kni_allocate_mbufs may be failed, but we don't know. >
> Even worse, when alloc_q is completely exhausted, kni_net_tx in > rte_kni.ko will drop all tx packets. kni_allocate_mbufs is never > called again, even if the mbufs are eventually freed.
>
> In this patch, we try to allocate mbufs for alloc_q when it is empty. >
> According to historical experience, the performance bottleneck of KNI > is offen the usleep_range of kni thread in rte_kni.ko.
> The check of kni_fifo_count is trivial and the cost should be acceptable.
>

Hi Yangchao,

Are you observing any performance impact with this change in you use case?

Quoted 7 lines
Quoted 5 lines
Quoted 4 lines
Fixes 3e12a98fe397 ("kni: optimize Rx burst")

<mailto:[email protected]>>

Applied, thanks.