Skip to contentSkip to conversation

net/udp_gso: Allow TX timestamp with UDP GSO

PATCH In review Fred Klassen 3 participants 3 messages Jun 17, 2019 – Jun 19, 2019

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH v4 1/1 net/udp_gso: Allow TX timestamp with UDP GSO

Fixes an issue where TX Timestamps are not arriving on the error queue when UDP_SEGMENT CMSG type is combined with CMSG type SO_TIMESTAMPING. This can be illustrated with an updated updgso_bench_tx program which includes the '-T' option to test for this condition. It also introduces the '-P' option which will call poll() before reading the error queue.

./udpgso_bench_tx -4ucTPv -S 1472 -l2 -D 172.16.120.18 poll timeout
udp tx: 0 MB/s 1 calls/s 1 msg/s

The "poll timeout" message above indicates that TX timestamp never arrived.

This patch preserves tx_flags for the first UDP GSO segment. Only the first segment is timestamped, even though in some cases there may be benefital in timestamping both the first and last segment.

Factors in deciding on first segment timestamp only:

- Timestamping both first and last segmented is not feasible. Hardware can only have one outstanding TS request at a time.

- Timestamping last segment may under report network latency of the previous segments. Even though the doorbell is suppressed, the ring producer counter has been incremented.

- Timestamping the first segment has the upside in that it reports timestamps from the application's view, e.g. RTT.

- Timestamping the first segment has the downside that it may underreport tx host network latency. It appears that we have to pick one or the other. And possibly follow-up with a config flag to choose behavior.

v2: Remove tests as noted by Willem de Bruijn <[email protected]>
Moving tests from net to net-next

v3: Update only relevant tx_flag bits as per
Willem de Bruijn <[email protected]>

v4: Update comments and commit message as per
Willem de Bruijn <[email protected]>

Fixes ee80d1ebe5ba ("udp: add udp gso")
---
 net/ipv4/udp_offload.c | 5 +++++
 1 file changed, 5 insertions(+)
net/ipv4/udp_offload.c +5 -0
index 06b3e2c1fcdc..9763464a75d7 100644
@@ -224,6 +224,11 @@ struct sk_buff *__udp_gso_segment(struct sk_buff *gso_skb,
 	seg = segs;
 	uh = udp_hdr(seg);
 
+	/* preserve TX timestamp flags and TS key for first segment */
+	skb_shinfo(seg)->tskey = skb_shinfo(gso_skb)->tskey;
+	skb_shinfo(seg)->tx_flags |=
+			(skb_shinfo(gso_skb)->tx_flags & SKBTX_ANY_TSTAMP);
+
 	/* compute checksum adjustment based on old length versus new */
 	newlen = htons(sizeof(*uh) + mss);
 	check = csum16_add(csum16_sub(uh->check, uh->len), newlen);
Signature · 1 line
2.11.0
Acked-by on v4 1/1
#1 lore raw
Quoted 38 lines · Fred Klassen

Fixes an issue where TX Timestamps are not arriving on the error queue when UDP_SEGMENT CMSG type is combined with CMSG type SO_TIMESTAMPING. This can be illustrated with an updated updgso_bench_tx program which includes the '-T' option to test for this condition. It also introduces the '-P' option which will call poll() before reading the error queue.

./udpgso_bench_tx -4ucTPv -S 1472 -l2 -D 172.16.120.18 poll timeout
udp tx: 0 MB/s 1 calls/s 1 msg/s

The "poll timeout" message above indicates that TX timestamp never arrived.

This patch preserves tx_flags for the first UDP GSO segment. Only the first segment is timestamped, even though in some cases there may be benefital in timestamping both the first and last segment.

Factors in deciding on first segment timestamp only:

- Timestamping both first and last segmented is not feasible. Hardware can only have one outstanding TS request at a time.

- Timestamping last segment may under report network latency of the previous segments. Even though the doorbell is suppressed, the ring producer counter has been incremented.

- Timestamping the first segment has the upside in that it reports timestamps from the application's view, e.g. RTT.

- Timestamping the first segment has the downside that it may underreport tx host network latency. It appears that we have to pick one or the other. And possibly follow-up with a config flag to choose behavior.

v2: Remove tests as noted by Willem de Bruijn <[email protected]>
Moving tests from net to net-next

Quoted 8 lines

v3: Update only relevant tx_flag bits as per
Willem de Bruijn <[email protected]>

v4: Update comments and commit message as per
Willem de Bruijn <[email protected]>

Fixes ee80d1ebe5ba ("udp: add udp gso")
maintainer
#2 lore raw
In reply to Fred Klassen

From: Fred Klassen <[email protected]>
Date: Mon, 17 Jun 2019 12:05:07 -0700

Quoted 5 lines

Fixes an issue where TX Timestamps are not arriving on the error queue when UDP_SEGMENT CMSG type is combined with CMSG type SO_TIMESTAMPING. This can be illustrated with an updated updgso_bench_tx program which includes the '-T' option to test for this condition. It also introduces the '-P' option which will call poll() before reading the error queue.

...

Quoted 23 lines

The "poll timeout" message above indicates that TX timestamp never arrived.

This patch preserves tx_flags for the first UDP GSO segment. Only the first segment is timestamped, even though in some cases there may be benefital in timestamping both the first and last segment.

Factors in deciding on first segment timestamp only:

- Timestamping both first and last segmented is not feasible. Hardware can only have one outstanding TS request at a time.

- Timestamping last segment may under report network latency of the previous segments. Even though the doorbell is suppressed, the ring producer counter has been incremented.

- Timestamping the first segment has the upside in that it reports timestamps from the application's view, e.g. RTT.

- Timestamping the first segment has the downside that it may underreport tx host network latency. It appears that we have to pick one or the other. And possibly follow-up with a config flag to choose behavior.

...

Quoted 2 lines
Fixes ee80d1ebe5ba ("udp: add udp gso")

Applied and queued up for -stable, thanks.