Neal Cardwell found that recent kernels were having RWIN limited
issues, even when net.ipv4.tcp_rmem[2] was set to a very big value like
512MB.
He suspected that tcp_stream default buffer size (64KB) was triggering
heuristic added in ea33537d8292 ("tcp: add receive queue awareness
in tcp_rcv_space_adjust()").
After more testing, it turns out the bug was added earlier
with commit 65c5287892e9 ("tcp: fix sk_rcvbuf overshoot").
I forgot once again that DRS has one RTT latency.
MPTCP also got the same issue.
This series :
- Prevent calling tcp_rcvbuf_grow() on some MPTCP subflows.
- adds rcv_ssthresh, window_clamp and rcv_wnd to trace_tcp_rcvbuf_grow().
- Refactors code in a patch with no functional changes.
- Fixes the issue in the final patch.
Signed-off-by: Eric Dumazet <[email protected]>
[ Added patch 1/4. ]
Signed-off-by: Matthieu Baerts (NGI0) <[email protected]>
---
Changes in v3:
- Fix warnings at build time by moving 'oldval' declaration (Matthieu)
- Prevent possible divide by zero issue in mptcp_rcv_space_adjust() (Paolo)
- Note: this v3 is not being sent by Eric because he is unavailable.
- Link to v2:
https://patch.msgid.link/[email protected]
Changes in v2:
- Rebased to net tree
- Changed mptcp_rcvbuf_grow() to read/write msk->rcvq_space.space (Paolo)
- Link to v1:
https://patch.msgid.link/[email protected]
---
Eric Dumazet (3):
trace: tcp: add three metrics to trace_tcp_rcvbuf_grow()
tcp: add newval parameter to tcp_rcvbuf_grow()
tcp: fix too slow tcp_rcvbuf_grow() action
Paolo Abeni (1):
mptcp: fix subflow rcvbuf adjust
include/net/tcp.h | 2 +-
include/trace/events/tcp.h | 9 +++++++++
net/ipv4/tcp_input.c | 21 ++++++++++++++-------
net/mptcp/protocol.c | 26 +++++++++++++++++---------
4 files changed, 41 insertions(+), 17 deletions(-)
---
base-commit: 210b35d6a7ea415494ce75490c4b43b4e717d935
change-id: 20251028-net-tcp-recv-autotune-5876d6d85d8a
Best regards,
--
Matthieu Baerts (NGI0) <[email protected]>
