diff options
author | Julian Wiedmann | 2017-12-01 10:14:50 +0100 |
---|---|---|
committer | David S. Miller | 2017-12-03 03:35:21 +0100 |
commit | 6d69b1f1eb7a2edf8a3547f361c61f2538e054bb (patch) | |
tree | 11944d807ae8a95faa6a3f77491a52cbe0bd1378 /drivers/s390/net/qeth_core_main.c | |
parent | s390/qeth: fix thinko in IPv4 multicast address tracking (diff) | |
download | kernel-qcow2-linux-6d69b1f1eb7a2edf8a3547f361c61f2538e054bb.tar.gz kernel-qcow2-linux-6d69b1f1eb7a2edf8a3547f361c61f2538e054bb.tar.xz kernel-qcow2-linux-6d69b1f1eb7a2edf8a3547f361c61f2538e054bb.zip |
s390/qeth: fix GSO throughput regression
Using GSO with small MTUs currently results in a substantial throughput
regression - which is caused by how qeth needs to map non-linear skbs
into its IO buffer elements:
compared to a linear skb, each GSO-segmented skb effectively consumes
twice as many buffer elements (ie two instead of one) due to the
additional header-only part. This causes the Output Queue to be
congested with low-utilized IO buffers.
Fix this as follows:
If the MSS is low enough so that a non-SG GSO segmentation produces
order-0 skbs (currently ~3500 byte), opt out from NETIF_F_SG. This is
where we anticipate the biggest savings, since an SG-enabled
GSO segmentation produces skbs that always consume at least two
buffer elements.
Larger MSS values continue to get a SG-enabled GSO segmentation, since
1) the relative overhead of the additional header-only buffer element
becomes less noticeable, and
2) the linearization overhead increases.
With the throughput regression fixed, re-enable NETIF_F_SG by default to
reap the significant CPU savings of GSO.
Fixes: 5722963a8e83 ("qeth: do not turn on SG per default")
Reported-by: Nils Hoppmann <niho@de.ibm.com>
Signed-off-by: Julian Wiedmann <jwi@linux.vnet.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Diffstat (limited to 'drivers/s390/net/qeth_core_main.c')
-rw-r--r-- | drivers/s390/net/qeth_core_main.c | 31 |
1 files changed, 31 insertions, 0 deletions
diff --git a/drivers/s390/net/qeth_core_main.c b/drivers/s390/net/qeth_core_main.c index 49b9efeba1bd..d9b0e07d4fa7 100644 --- a/drivers/s390/net/qeth_core_main.c +++ b/drivers/s390/net/qeth_core_main.c @@ -19,6 +19,11 @@ #include <linux/mii.h> #include <linux/kthread.h> #include <linux/slab.h> +#include <linux/if_vlan.h> +#include <linux/netdevice.h> +#include <linux/netdev_features.h> +#include <linux/skbuff.h> + #include <net/iucv/af_iucv.h> #include <net/dsfield.h> @@ -6438,6 +6443,32 @@ netdev_features_t qeth_fix_features(struct net_device *dev, } EXPORT_SYMBOL_GPL(qeth_fix_features); +netdev_features_t qeth_features_check(struct sk_buff *skb, + struct net_device *dev, + netdev_features_t features) +{ + /* GSO segmentation builds skbs with + * a (small) linear part for the headers, and + * page frags for the data. + * Compared to a linear skb, the header-only part consumes an + * additional buffer element. This reduces buffer utilization, and + * hurts throughput. So compress small segments into one element. + */ + if (netif_needs_gso(skb, features)) { + /* match skb_segment(): */ + unsigned int doffset = skb->data - skb_mac_header(skb); + unsigned int hsize = skb_shinfo(skb)->gso_size; + unsigned int hroom = skb_headroom(skb); + + /* linearize only if resulting skb allocations are order-0: */ + if (SKB_DATA_ALIGN(hroom + doffset + hsize) <= SKB_MAX_HEAD(0)) + features &= ~NETIF_F_SG; + } + + return vlan_features_check(skb, features); +} +EXPORT_SYMBOL_GPL(qeth_features_check); + static int __init qeth_core_init(void) { int rc; |