The patch discussion seems got lost due to a delay it took us to get the numbers. We believe that a 24% improvement in VM's network performance (and probably the better improvement the more guests are running on a host) is worth commenting and getting to some conclusion.
> Before your patch : > > sizeof(struct skb_shared_info)=0x140 > offsetof(struct skb_shared_info, frags[1])=0x40 > > SKB_DATA_ALIGN(sizeof(struct skb_shared_info)) -> 0x140 > > After your patch : > > sizeof(struct skb_shared_info)=0x148 > offsetof(struct skb_shared_info, frags[1])=0x48 > > SKB_DATA_ALIGN(sizeof(struct skb_shared_info)) -> 0x180 > > Thats a serious bump, because it increases all skb truesizes, and > typical skb with one fragment will use 2 cache lines instead of one in > struct skb_shared_info, so this adds memory pressure in fast path. > > So while this patch might increase performance for some workloads, > it generally decreases performance on many others. Would moving the parent fragment pointer from skb_shared_info to skbuff solve this issue? Regards, -Igor On Sun, May 25, 2014 at 1:54 PM, Igor Royzis <[email protected]> wrote: >> If true, I'd like to see some performance numbers please. > > The numbers have been obtained by running iperf between 2 QEMU Win2012 > VMs, 4 vCPU/ 4GB RAM each. > iperf parameters: -w 256K -l 256K -t 300 > Original kernel 3.15.0-rc5: 34.4 Gbytes transferred, 984 > Mbits/sec bandwidth. > Kernel 3.15.0-rc5 with our patch: 42.5 Gbytes transferred, 1.22 > Gbits/sec bandwidth. > > Overall improvement is about 24%. > Below are raw iperf outputs. > > kernel 3.15.0-rc5: > C:\iperf>iperf -c 192.168.11.2 -w 256K -l 256K -t 300 > ------------------------------------------------------------ > Client connecting to 192.168.11.2, TCP port 5001 > TCP window size: 256 KByte > ------------------------------------------------------------ > [ 3] local 192.168.11.1 port 49167 connected with 192.168.11.2 port 5001 > [ ID] Interval Transfer Bandwidth > [ 3] 0.0-300.7 sec 34.4 GBytes 984 Mbits/sec > > kernel 3.15.0-rc5-patched: > C:\iperf>iperf -c 192.168.11.2 -w 256K -l 256K -t 300 > ------------------------------------------------------------ > Client connecting to 192.168.11.2, TCP port 5001 > TCP window size: 256 KByte > ------------------------------------------------------------ > [ 3] local 192.168.11.1 port 49167 connected with 192.168.11.2 port 5001 > [ ID] Interval Transfer Bandwidth > [ 3] 0.0-300.7 sec 42.5 GBytes 1.22 Gbits/sec > > On Tue, May 20, 2014 at 2:50 PM, Michael S. Tsirkin <[email protected]> wrote: >> >> On Tue, May 20, 2014 at 02:24:21PM +0300, Igor Royzis wrote: >> > Fix accessing GSO fragments memory (and a possible corruption therefore) >> > after >> > reporting completion in a zero copy callback. The previous fix in the >> > commit 1fd819ec >> > orphaned frags which eliminates zero copy advantages. The fix makes the >> > completion >> > called after all the fragments were processed avoiding unnecessary >> > orphaning/copying >> > from userspace. >> > >> > The GSO fragments corruption issue was observed in a typical QEMU/KVM VM >> > setup that >> > hosts a Windows guest (since QEMU virtio-net Windows driver doesn't >> > support GRO). >> > The fix has been verified by running the HCK OffloadLSO test. >> > >> > Signed-off-by: Igor Royzis <[email protected]> >> > Signed-off-by: Anton Nayshtut <[email protected]> >> >> OK but with 1fd819ec there's no corruption, correct? >> So this patch is in fact an optimization? >> If true, I'd like to see some performance numbers please. >> >> Thanks! >> >> > --- >> > include/linux/skbuff.h | 1 + >> > net/core/skbuff.c | 18 +++++++++++++----- >> > 2 files changed, 14 insertions(+), 5 deletions(-) >> > >> > diff --git a/include/linux/skbuff.h b/include/linux/skbuff.h >> > index 08074a8..8c49edc 100644 >> > --- a/include/linux/skbuff.h >> > +++ b/include/linux/skbuff.h >> > @@ -287,6 +287,7 @@ struct skb_shared_info { >> > struct sk_buff *frag_list; >> > struct skb_shared_hwtstamps hwtstamps; >> > __be32 ip6_frag_id; >> > + struct sk_buff *zcopy_src; >> > >> > /* >> > * Warning : all fields before dataref are cleared in __alloc_skb() >> > diff --git a/net/core/skbuff.c b/net/core/skbuff.c >> > index 1b62343..6fa6342 100644 >> > --- a/net/core/skbuff.c >> > +++ b/net/core/skbuff.c >> > @@ -610,14 +610,18 @@ EXPORT_SYMBOL(__kfree_skb); >> > */ >> > void kfree_skb(struct sk_buff *skb) >> > { >> > + struct sk_buff *zcopy_src; >> > if (unlikely(!skb)) >> > return; >> > if (likely(atomic_read(&skb->users) == 1)) >> > smp_rmb(); >> > else if (likely(!atomic_dec_and_test(&skb->users))) >> > return; >> > + zcopy_src = skb_shinfo(skb)->zcopy_src; >> > trace_kfree_skb(skb, __builtin_return_address(0)); >> > __kfree_skb(skb); >> > + if (unlikely(zcopy_src)) >> > + kfree_skb(zcopy_src); >> > } >> > EXPORT_SYMBOL(kfree_skb); >> > >> > @@ -662,14 +666,18 @@ EXPORT_SYMBOL(skb_tx_error); >> > */ >> > void consume_skb(struct sk_buff *skb) >> > { >> > + struct sk_buff *zcopy_src; >> > if (unlikely(!skb)) >> > return; >> > if (likely(atomic_read(&skb->users) == 1)) >> > smp_rmb(); >> > else if (likely(!atomic_dec_and_test(&skb->users))) >> > return; >> > + zcopy_src = skb_shinfo(skb)->zcopy_src; >> > trace_consume_skb(skb); >> > __kfree_skb(skb); >> > + if (unlikely(zcopy_src)) >> > + consume_skb(zcopy_src); >> > } >> > EXPORT_SYMBOL(consume_skb); >> > >> > @@ -2867,7 +2875,6 @@ struct sk_buff *skb_segment(struct sk_buff *head_skb, >> > skb_frag_t *frag = skb_shinfo(head_skb)->frags; >> > unsigned int mss = skb_shinfo(head_skb)->gso_size; >> > unsigned int doffset = head_skb->data - skb_mac_header(head_skb); >> > - struct sk_buff *frag_skb = head_skb; >> > unsigned int offset = doffset; >> > unsigned int tnl_hlen = skb_tnl_header_len(head_skb); >> > unsigned int headroom; >> > @@ -2913,7 +2920,6 @@ struct sk_buff *skb_segment(struct sk_buff *head_skb, >> > i = 0; >> > nfrags = skb_shinfo(list_skb)->nr_frags; >> > frag = skb_shinfo(list_skb)->frags; >> > - frag_skb = list_skb; >> > pos += skb_headlen(list_skb); >> > >> > while (pos < offset + len) { >> > @@ -2975,6 +2981,11 @@ struct sk_buff *skb_segment(struct sk_buff >> > *head_skb, >> > nskb->data - tnl_hlen, >> > doffset + tnl_hlen); >> > >> > + if (skb_shinfo(head_skb)->tx_flags & SKBTX_DEV_ZEROCOPY) { >> > + skb_shinfo(nskb)->zcopy_src = head_skb; >> > + atomic_inc(&head_skb->users); >> > + } >> > + >> > if (nskb->len == len + doffset) >> > goto perform_csum_check; >> > >> > @@ -3001,7 +3012,6 @@ struct sk_buff *skb_segment(struct sk_buff *head_skb, >> > i = 0; >> > nfrags = skb_shinfo(list_skb)->nr_frags; >> > frag = skb_shinfo(list_skb)->frags; >> > - frag_skb = list_skb; >> > >> > BUG_ON(!nfrags); >> > >> > @@ -3016,8 +3026,6 @@ struct sk_buff *skb_segment(struct sk_buff *head_skb, >> > goto err; >> > } >> > >> > - if (unlikely(skb_orphan_frags(frag_skb, GFP_ATOMIC))) >> > - goto err; >> > >> > *nskb_frag = *frag; >> > __skb_frag_ref(nskb_frag); >> > -- >> > 1.7.9.5 -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [email protected] More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
