Patchset for task https://virtuozzo.atlassian.net/browse/VSTOR-153491
mptcp_join.sh test 021 never completes in CI: a link fails during a
bidirectional transfer and the transfer stalls. The series backports
mainline fixes for the two problems behind it:
- Endless retransmission with csum enabled: A retransmission can send
data beyond snd_nxt, the acks for that data are dropped, and the same
data is retransmitted again and again. Patch 1/11 fixes it.
- Receive stalls under memory pressure: Incoming data over the receive
buffer limit waits in the subflow queue, and may get stuck, blocking
further progress. Patch 8 drops it instead and patch 9 prunes the
out-of-order queue, both relying on MPTCP-level retransmission.
- Patches 10/11 and 11/11 are mainline fixes for patch 9/11.
The other patches are prerequisites:
- Patches 2/11 and 3/11 are cleanups the later patches are based on.
- Patch 4/11 lets one MPTCP retransmit timeout resend more than one data
fragment, so the data dropped by 8/11 and 9/11 retransmits sooner.
- Patches 5/11 and 6/11 keep the window announced to the peer within
receive buffer, so that a peer that behaves correctly does not send
data that patch 8/11 would even drop.
- Patch 7/11 builds the dummy mapping of a fallback socket from the TCP
sequence number instead of msk->ack_seq. Without it, data dropped by
patch 8/11 and then a fallback to TCP would give the following bytes
the DSN of the dropped ones, and the reader would get a stream with
bytes missing.
Left out of the mainline series, between patches 8/11 and 9/11:
b1224c4b40f6 ("mptcp: enforce hard limit on backlog flushing"), as this
tree has no backlog, and 996643574cc8 ("mptcp: avoid code duplication
in __mptcp_move_skb()"), a cleanup. Mainline 0e24d17bd966 ("tcp:
implement RFC 7323 window retraction receiver requirements"), which
patch 6/11 sits on in mainline, is not taken either, and 6/11 rebases
the same way it does in upstream stable 6.12.y branch.
Tested in QEMU (4 vCPUs): test 021 with a failing link have not failed,
against 3 failures on unpatched kernel, and the net/mptcp selftest suite
passes with no warnings. The v1 also did fix the failures described here
but was arguably a less coherent series of patches.
Changes in v3:
- Add "ms/mptcp: cleanup fallback dummy mapping generation" before
"ms/mptcp: explicitly drop over memory limits".
- 6/11: A rebase note on the commit, based on 6.12.y tree observations.
- Expand cover letter.
- Rebase.
Link to v2:
http://lore.virtuozzo.com/kernel/179484894400.861498.16940306217017869314.b4-review@b4/T/#t
Changes in v2:
- Add "ms/mptcp: close TOCTOU race while computing rcv_wnd" and
"ms/mptcp: allow subflow rcv wnd to shrink before the drop over memory
limits".
- 3/10: reword the rebase note.
- 8/10: name the two skipped mainline commits in the rebase note.
- Rebase.
Link to v1:
http://lore.virtuozzo.com/kernel/179450441149.4060562.12778470791653323060.b4-review@b4/T/#t
Paolo Abeni (11):
ms/mptcp: fix retransmission loop when csum is enabled
ms/mptcp: move the retrans loop to a separate helper
ms/mptcp: move the stale logic out of retrans scheduler
ms/mptcp: let the retrans scheduler do its job
ms/mptcp: close TOCTOU race while computing rcv_wnd
ms/mptcp: allow subflow rcv wnd to shrink
ms/mptcp: cleanup fallback dummy mapping generation
ms/mptcp: explicitly drop over memory limits
ms/mptcp: implemented OoO queue pruning
ms/mptcp: being below memory limit is a likely() condition
ms/mptcp: avoid pruning for OoW data
net/mptcp/mib.c | 1 +
net/mptcp/mib.h | 1 +
net/mptcp/options.c | 43 ++++----
net/mptcp/pm.c | 43 +++++---
net/mptcp/protocol.c | 226 ++++++++++++++++++++++++++++++++++---------
net/mptcp/protocol.h | 11 ++-
net/mptcp/subflow.c | 8 +-
7 files changed, 249 insertions(+), 84 deletions(-)
--
2.56.1
_______________________________________________
Devel mailing list
[email protected]
https://lists.openvz.org/mailman/listinfo/devel