diff mbox series

[v2] vhost: fix mbuf alloc failure

Message ID 20200504171118.93782-1-Sivaprasad.Tummala@intel.com (mailing list archive)
State Superseded, archived
Delegated to: Maxime Coquelin
Headers show
Series [v2] vhost: fix mbuf alloc failure | expand

Checks

Context Check Description
ci/iol-testing success Testing PASS
ci/iol-mellanox-Performance success Performance Testing PASS
ci/Intel-compilation fail Compilation issues
ci/travis-robot success Travis build: passed
ci/iol-nxp-Performance success Performance Testing PASS
ci/iol-intel-Performance success Performance Testing PASS
ci/checkpatch success coding style OK

Commit Message

Sivaprasad Tummala May 4, 2020, 5:11 p.m. UTC
vhost buffer allocation is successful for packets that fit
into a linear buffer. If it fails, vhost library is expected
to drop the current packet and skip to the next.

The patch fixes the error scenario by skipping to next packet.
Note: Drop counters are not currently supported.

Fixes: c3ff0ac70acb ("vhost: improve performance by supporting large buffer")
Cc: stable@dpdk.org
Cc: fbl@sysclose.org

---
v2:
 * fixed review comments - Maxime Coquelin
 * fixed mbuf alloc errors for packed virtqueues - Maxime Coquelin
 * fixed mbuf copy errors - Flavio Leitner

Signed-off-by: Sivaprasad Tummala <Sivaprasad.Tummala@intel.com>
---
 lib/librte_vhost/virtio_net.c | 50 ++++++++++++++++++++++++++---------
 1 file changed, 37 insertions(+), 13 deletions(-)

Comments

Flavio Leitner May 4, 2020, 7:32 p.m. UTC | #1
On Mon, May 04, 2020 at 10:41:17PM +0530, Sivaprasad Tummala wrote:
> vhost buffer allocation is successful for packets that fit
> into a linear buffer. If it fails, vhost library is expected
> to drop the current packet and skip to the next.
> 
> The patch fixes the error scenario by skipping to next packet.
> Note: Drop counters are not currently supported.
> 
> Fixes: c3ff0ac70acb ("vhost: improve performance by supporting large buffer")
> Cc: stable@dpdk.org
> Cc: fbl@sysclose.org
> 
> ---
> v2:
>  * fixed review comments - Maxime Coquelin
>  * fixed mbuf alloc errors for packed virtqueues - Maxime Coquelin
>  * fixed mbuf copy errors - Flavio Leitner
> 
> Signed-off-by: Sivaprasad Tummala <Sivaprasad.Tummala@intel.com>
> ---
>  lib/librte_vhost/virtio_net.c | 50 ++++++++++++++++++++++++++---------
>  1 file changed, 37 insertions(+), 13 deletions(-)
> 
> diff --git a/lib/librte_vhost/virtio_net.c b/lib/librte_vhost/virtio_net.c
> index 1fc30c681..764c514fd 100644
> --- a/lib/librte_vhost/virtio_net.c
> +++ b/lib/librte_vhost/virtio_net.c
> @@ -1674,6 +1674,7 @@ virtio_dev_tx_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
>  {
>  	uint16_t i;
>  	uint16_t free_entries;
> +	uint16_t dropped = 0;
>  
>  	if (unlikely(dev->dequeue_zero_copy)) {
>  		struct zcopy_mbuf *zmbuf, *next;
> @@ -1737,13 +1738,31 @@ virtio_dev_tx_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
>  			update_shadow_used_ring_split(vq, head_idx, 0);
>  
>  		pkts[i] = virtio_dev_pktmbuf_alloc(dev, mbuf_pool, buf_len);
> -		if (unlikely(pkts[i] == NULL))
> +		if (unlikely(pkts[i] == NULL)) {
> +			/*
> +			 * mbuf allocation fails for jumbo packets when external
> +			 * buffer allocation is not allowed and linear buffer
> +			 * is required. Drop this packet.
> +			 */
> +#ifdef RTE_LIBRTE_VHOST_DEBUG
> +			VHOST_LOG_DATA(ERR,
> +				"Failed to allocate memory for mbuf. Packet dropped!\n");
> +#endif

That message is useful to spot that missing packets that happens
once in a while, so we should be able to see it even in production
without debug enabled. However, we can't let it flood the log.

I am not sure if librte eal has this functionality, but if not you
could limit by using a static bool:

static bool allocerr_warned = false;

if (allocerr_warned) {
    VHOST_LOG_DATA(ERR,
    "Failed to allocate memory for mbuf. Packet dropped!\n");
    allocerr_warned = true;
}



> +			dropped += 1;
> +			i++;
>  			break;
> +		}
>  
>  		err = copy_desc_to_mbuf(dev, vq, buf_vec, nr_vec, pkts[i],
>  				mbuf_pool);
>  		if (unlikely(err)) {
>  			rte_pktmbuf_free(pkts[i]);
> +#ifdef RTE_LIBRTE_VHOST_DEBUG
> +			VHOST_LOG_DATA(ERR,
> +				"Failed to copy desc to mbuf. Packet dropped!\n");
> +#endif

Same here.


> +			dropped += 1;
> +			i++;
>  			break;
>  		}
>  
> @@ -1753,6 +1772,8 @@ virtio_dev_tx_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
>  			zmbuf = get_zmbuf(vq);
>  			if (!zmbuf) {
>  				rte_pktmbuf_free(pkts[i]);
> +				dropped += 1;
> +				i++;
>  				break;
>  			}
>  			zmbuf->mbuf = pkts[i];
> @@ -1782,7 +1803,7 @@ virtio_dev_tx_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
>  		}
>  	}
>  
> -	return i;
> +	return (i - dropped);
>  }
>  
>  static __rte_always_inline int
> @@ -1946,21 +1967,24 @@ virtio_dev_tx_single_packed(struct virtio_net *dev,
>  			    struct rte_mbuf **pkts)
>  {
>  
> -	uint16_t buf_id, desc_count;
> +	uint16_t buf_id, desc_count = 0;
> +	int ret;
>  
> -	if (vhost_dequeue_single_packed(dev, vq, mbuf_pool, pkts, &buf_id,
> -					&desc_count))
> -		return -1;
> +	ret = vhost_dequeue_single_packed(dev, vq, mbuf_pool, pkts, &buf_id,
> +					&desc_count);
>  
> -	if (virtio_net_is_inorder(dev))
> -		vhost_shadow_dequeue_single_packed_inorder(vq, buf_id,
> -							   desc_count);
> -	else
> -		vhost_shadow_dequeue_single_packed(vq, buf_id, desc_count);
> +	if (likely(desc_count > 0)) {


The vhost_dequeue_single_packed() could return -1 with desc_count > 0
and this change doesn't handle that.

Thanks,
fbl


> +		if (virtio_net_is_inorder(dev))
> +			vhost_shadow_dequeue_single_packed_inorder(vq, buf_id,
> +								   desc_count);
> +		else
> +			vhost_shadow_dequeue_single_packed(vq, buf_id,
> +					desc_count);
>  
> -	vq_inc_last_avail_packed(vq, desc_count);
> +		vq_inc_last_avail_packed(vq, desc_count);
> +	}
>  
> -	return 0;
> +	return ret;
>  }
>  
>  static __rte_always_inline int
> -- 
> 2.17.1
>
Sivaprasad Tummala May 5, 2020, 5:48 a.m. UTC | #2
Hi Flavio,



Thanks for your comments.



SNIPPED



>                      pkts[i] = virtio_dev_pktmbuf_alloc(dev, mbuf_pool, buf_len);

> -                   if (unlikely(pkts[i] == NULL))

> +                  if (unlikely(pkts[i] == NULL)) {

> +                              /*

> +                              * mbuf allocation fails for jumbo packets when external

> +                              * buffer allocation is not allowed and linear buffer

> +                              * is required. Drop this packet.

> +                              */

> +#ifdef RTE_LIBRTE_VHOST_DEBUG

> +                              VHOST_LOG_DATA(ERR,

> +                                          "Failed to allocate memory for mbuf. Packet dropped!\n"); #endif



That message is useful to spot that missing packets that happens once in a while, so we should be able to see it even in production without debug enabled. However, we can't let it flood the log.

Agreed, but VHOST_LOG wrapper does not have rate limit functionality.





I am not sure if librte eal has this functionality, but if not you could limit by using a static bool:



static bool allocerr_warned = false;



if (allocerr_warned) {

    VHOST_LOG_DATA(ERR,

    "Failed to allocate memory for mbuf. Packet dropped!\n");

    allocerr_warned = true;

}



This is good idea, but having a static variable makes it file scope making it to  entire VHOST devices. Hence if the intention is to implement device specific

log rate limit, should not we resort to `dev->allocerr_warn` counter mechanism, which resets after n failures `#define LOG_ALLOCFAIL 32`.



SNIPPED



>  static __rte_always_inline int

> @@ -1946,21 +1967,24 @@ virtio_dev_tx_single_packed(struct virtio_net *dev,

>                                      struct rte_mbuf **pkts)

>  {

>

> -       uint16_t buf_id, desc_count;

> +      uint16_t buf_id, desc_count = 0;

> +      int ret;

>

> -       if (vhost_dequeue_single_packed(dev, vq, mbuf_pool, pkts, &buf_id,

> -                                                       &desc_count))

> -                   return -1;

> +      ret = vhost_dequeue_single_packed(dev, vq, mbuf_pool, pkts, &buf_id,

> +                                                      &desc_count);

>

> -       if (virtio_net_is_inorder(dev))

> -                   vhost_shadow_dequeue_single_packed_inorder(vq, buf_id,

> -                                                                                  desc_count);

> -       else

> -                   vhost_shadow_dequeue_single_packed(vq, buf_id, desc_count);

> +      if (likely(desc_count > 0)) {





The vhost_dequeue_single_packed() could return -1 with desc_count > 0 and this change doesn't handle that.


Yes, as per my current understanding in either success or failure we need to flush the descriptors `desc_count` to handle this issue.
Is there an expectation  for partial or incomplete packet where `num_desc` is greater than 0, we need to preserve it.


SNIPPED



Thanks & Regards,

Sivaprasad
Maxime Coquelin May 5, 2020, 8:20 a.m. UTC | #3
(Please try to avoid HTML for the replies, it makes it hard to follow)

See my replies below:

On 5/5/20 7:48 AM, Tummala, Sivaprasad wrote:
> Hi Flavio,
> 
>  
> 
> Thanks for your comments.
> 
>  
> 
> SNIPPED
> 
>  
> 
>>                      pkts[i] = virtio_dev_pktmbuf_alloc(dev,
> mbuf_pool, buf_len);
> 
>> -                   if (unlikely(pkts[i] == NULL))
> 
>> +                  if (unlikely(pkts[i] == NULL)) {
> 
>> +                              /*
> 
>> +                              * mbuf allocation fails for jumbo
> packets when external
> 
>> +                              * buffer allocation is not allowed and
> linear buffer
> 
>> +                              * is required. Drop this packet.
> 
>> +                              */
> 
>> +#ifdef RTE_LIBRTE_VHOST_DEBUG
> 
>> +                              VHOST_LOG_DATA(ERR,
> 
>> +                                          "Failed to allocate memory
> for mbuf. Packet dropped!\n"); #endif
> 
>  
> 
> That message is useful to spot that missing packets that happens once in
> a while, so we should be able to see it even in production without debug
> enabled. However, we can't let it flood the log.
> 
> Agreed, but VHOST_LOG wrapper does not have rate limit functionality.
> 
>  
> 
>  
> 
> I am not sure if librte eal has this functionality, but if not you could
> limit by using a static bool:
> 
>  
> 
> static bool allocerr_warned = false;
> 
>  
> 
> if (allocerr_warned) {
> 
>     VHOST_LOG_DATA(ERR,
> 
>     "Failed to allocate memory for mbuf. Packet dropped!\n");
> 
>     allocerr_warned = true;
> 
> }
> 
>  
> 
> This is good idea, but having a static variable makes it file scope
> making it to  entire VHOST devices. Hence if the intention is to
> implement device specific
> 
> log rate limit, should not we resort to `dev->allocerr_warn` counter
> mechanism, which resets after n failures `#define LOG_ALLOCFAIL 32`.

I prefer Flavio's proposal, it would have less performance impact than
increasing struct virtio_net size. As soon as we can see the error
popping once in the log message, it gives some clues on what to
investigate. Maybe providing more details on the failure could help,
like printing the pool name and the requested length.

Maxime
Sivaprasad Tummala May 5, 2020, 11:56 a.m. UTC | #4
Hi Maxime, 

Thanks for your comments. 

SNIPPED

if (allocerr_warned) {
> 
>     VHOST_LOG_DATA(ERR,
> 
>     "Failed to allocate memory for mbuf. Packet dropped!\n");
> 
>     allocerr_warned = true;
> 
> }
> 
>  
> 
> This is good idea, but having a static variable makes it file scope 
> making it to  entire VHOST devices. Hence if the intention is to 
> implement device specific
> 
> log rate limit, should not we resort to `dev->allocerr_warn` counter 
> mechanism, which resets after n failures `#define LOG_ALLOCFAIL 32`.

I prefer Flavio's proposal, it would have less performance impact than increasing struct virtio_net size. As soon as we can see the error popping once in the log message, it gives some clues on what to investigate. Maybe providing more details on the failure could help, like printing the pool name and the requested length.

Agreed. Change in the next patch, sample format as `VHOST_DATA : Failed mbuf alloc of size 2054 from mbuf_pool_socket_0 on /tmp/vhost1.`

SNIPPED
diff mbox series

Patch

diff --git a/lib/librte_vhost/virtio_net.c b/lib/librte_vhost/virtio_net.c
index 1fc30c681..764c514fd 100644
--- a/lib/librte_vhost/virtio_net.c
+++ b/lib/librte_vhost/virtio_net.c
@@ -1674,6 +1674,7 @@  virtio_dev_tx_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
 {
 	uint16_t i;
 	uint16_t free_entries;
+	uint16_t dropped = 0;
 
 	if (unlikely(dev->dequeue_zero_copy)) {
 		struct zcopy_mbuf *zmbuf, *next;
@@ -1737,13 +1738,31 @@  virtio_dev_tx_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
 			update_shadow_used_ring_split(vq, head_idx, 0);
 
 		pkts[i] = virtio_dev_pktmbuf_alloc(dev, mbuf_pool, buf_len);
-		if (unlikely(pkts[i] == NULL))
+		if (unlikely(pkts[i] == NULL)) {
+			/*
+			 * mbuf allocation fails for jumbo packets when external
+			 * buffer allocation is not allowed and linear buffer
+			 * is required. Drop this packet.
+			 */
+#ifdef RTE_LIBRTE_VHOST_DEBUG
+			VHOST_LOG_DATA(ERR,
+				"Failed to allocate memory for mbuf. Packet dropped!\n");
+#endif
+			dropped += 1;
+			i++;
 			break;
+		}
 
 		err = copy_desc_to_mbuf(dev, vq, buf_vec, nr_vec, pkts[i],
 				mbuf_pool);
 		if (unlikely(err)) {
 			rte_pktmbuf_free(pkts[i]);
+#ifdef RTE_LIBRTE_VHOST_DEBUG
+			VHOST_LOG_DATA(ERR,
+				"Failed to copy desc to mbuf. Packet dropped!\n");
+#endif
+			dropped += 1;
+			i++;
 			break;
 		}
 
@@ -1753,6 +1772,8 @@  virtio_dev_tx_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
 			zmbuf = get_zmbuf(vq);
 			if (!zmbuf) {
 				rte_pktmbuf_free(pkts[i]);
+				dropped += 1;
+				i++;
 				break;
 			}
 			zmbuf->mbuf = pkts[i];
@@ -1782,7 +1803,7 @@  virtio_dev_tx_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
 		}
 	}
 
-	return i;
+	return (i - dropped);
 }
 
 static __rte_always_inline int
@@ -1946,21 +1967,24 @@  virtio_dev_tx_single_packed(struct virtio_net *dev,
 			    struct rte_mbuf **pkts)
 {
 
-	uint16_t buf_id, desc_count;
+	uint16_t buf_id, desc_count = 0;
+	int ret;
 
-	if (vhost_dequeue_single_packed(dev, vq, mbuf_pool, pkts, &buf_id,
-					&desc_count))
-		return -1;
+	ret = vhost_dequeue_single_packed(dev, vq, mbuf_pool, pkts, &buf_id,
+					&desc_count);
 
-	if (virtio_net_is_inorder(dev))
-		vhost_shadow_dequeue_single_packed_inorder(vq, buf_id,
-							   desc_count);
-	else
-		vhost_shadow_dequeue_single_packed(vq, buf_id, desc_count);
+	if (likely(desc_count > 0)) {
+		if (virtio_net_is_inorder(dev))
+			vhost_shadow_dequeue_single_packed_inorder(vq, buf_id,
+								   desc_count);
+		else
+			vhost_shadow_dequeue_single_packed(vq, buf_id,
+					desc_count);
 
-	vq_inc_last_avail_packed(vq, desc_count);
+		vq_inc_last_avail_packed(vq, desc_count);
+	}
 
-	return 0;
+	return ret;
 }
 
 static __rte_always_inline int