[PATCH] io_uring/cmd_net: drop unextractable TX_TIMESTAMP skbs instead of requeueing

[lollipopkit <a@lolli.tech>]

SOCKET_URING_OP_TX_TIMESTAMP selects error-queue skbs with
skb_has_tx_timestamp() but extracts the timestamp with
skb_get_tx_timestamp().  The two use different checks and can disagree:
a skb the walk accepted can be rejected with -ENOENT forever.  When
that happens the processing loop stops and skb_queue_splice() puts the
unprocessed skbs back onto the HEAD of sk_error_queue.  The rejected
skb is selected first on every following run, fails again, and blocks
every skb behind it: new timestamps keep arriving and keep re-running
the command, but nothing is ever delivered again.

The disagreement is reachable without a race.  With
SOF_TIMESTAMPING_BIND_PHC the walk never looks at sk_bind_phc, but
skb_get_tx_timestamp() converts the hardware timestamp through
ptp_convert_timestamp(), which returns 0 when no virtual clock matches
sk_bind_phc (for example when the vclock was destroyed after the
socket bound to it).  ktime_to_timespec64_cond() then fails and the
extractor returns -ENOENT permanently.  Any skb with a non-zero
hwtstamp is selected while SOF_TIMESTAMPING_RAW_HARDWARE is set, so a
single such skb blocks even the software timestamps queued behind it.

Measured on real hardware, an I219-V (e1000e), using
/sys/class/ptp/ptpN/n_vclocks to create and destroy the vclock the
socket was bound to: 22 error-queue skbs generated (16 software + 6
hardware).  Of the 21 extractor calls, 20 failed with -ENOENT and the
one that succeeded delivered 1 software timestamp; deliveries then
froze while 8 further sends kept re-running the command, and the
remaining 21 skbs (15 software + 6 hardware) sat in the error queue.
Both control arms (vclock kept alive; no BIND_PHC) delivered
everything.

With this patch applied on top of the CQ-full fix ("io_uring/cmd_net:
end TX_TIMESTAMP multishot when the CQ is full"), A/B tested in a KVM
guest from a single kernel binary, with the new behaviour behind a
test-only runtime switch (the switch is test scaffolding and is not
part of this patch).  The guest needs no special hardware: QEMU's igb
model emulates hardware TX timestamping (TSYNCTXCTL/TXSTMPL/TXSTMPH)
and the igb driver registers a PTP clock, so the guest reproduces the
same failure the same way.  The two arms are separate runs with their
own sockets and sends; the device holds one pending hardware
timestamp, so how many of the 16 sends get a hardware skb varies per
run (12 and 8 here).  Switch off: 28 skbs generated (16 software + 12
hardware), deliveries froze at 1 and the remaining 27 (15 software +
12 hardware) sat in the error queue.  Switch on: 24 skbs generated (16
software + 8 hardware); all 16 software timestamps were delivered, the
8 hardware skbs the extractor rejected with -ENOENT were dropped, and
the error queue drained.

recvmsg(MSG_ERRQUEUE) already treats such a skb as skippable:
sock_recv_errqueue() dequeues it, returns the error message without a
timestamp, and frees it.  Do the same here and keep the deliverable
timestamps moving: drop the skb the extractor rejected and continue
with the rest of the list.  The CQE-full case keeps its
requeue-and-end-multishot behaviour.

Fixes: 9e4ed359b8ef ("io_uring/netcmd: add tx timestamping cmd support")
Cc: stable@vger.kernel.org # needs the CQ-full fix above
Assisted-by: Claude:claude-opus-5-5
Signed-off-by: lollipopkit <a@lolli.tech>
---
This applies on top of axboe/for-next (3769bc31a5b7).  The CQ-full fix
is 3a3d93070c5e in io_uring-7.3; the two touch adjacent lines of the
same loop, hence the stable note.

Reproducer: a single-file liburing program that runs the three arms
(destroyed vclock, live vclock, no BIND_PHC) against any NIC with
hardware TX timestamps, plus a KVM guest setup using QEMU's igb model.
Happy to post it or turn it into a selftest if that is useful.

One behaviour change worth calling out: skb_has_tx_timestamp() and
skb_get_tx_timestamp() read sk_tsflags independently, so a concurrent
setsockopt(SO_TIMESTAMPING) can also make the extractor reject a skb
the walk selected.  Before this patch that skb was delivered on a later
run; now it is dropped, as recvmsg(MSG_ERRQUEUE) would.  With a thread
switching sk_tsflags about 800k times against 20000 sends on loopback,
this patch dropped 320 such timestamps.  Reading sk_tsflags once per
run and passing it to both helpers removes that case (0 rejections, 0
lost in the same test); it changes both helpers' signatures in
net/socket.c, so I left it out of this fix.  I can send it as a
preceding patch if you prefer the two together.

 io_uring/cmd_net.c | 15 +++++++++++++++
 1 file changed, 15 insertions(+)

diff --git a/io_uring/cmd_net.c b/io_uring/cmd_net.c
index 90d4ec7cc761..406fbe8b7d52 100644
--- a/io_uring/cmd_net.c
+++ b/io_uring/cmd_net.c
@@ -138,6 +138,21 @@ static int io_uring_cmd_timestamp(struct socket *sock,
 		if (!skb)
 			break;
 		ret = io_process_timestamp_skb(cmd, sk, skb, issue_flags);
+		if (ret && ret != -ENOBUFS) {
+			/*
+			 * The walk selected this skb, but skb_get_tx_timestamp()
+			 * cannot extract a timestamp from it: the two use
+			 * different checks.  Splicing it back would put it at
+			 * the head of sk_error_queue, where every later run
+			 * selects it first and fails on it again, so nothing
+			 * behind it is ever delivered.  recvmsg(MSG_ERRQUEUE)
+			 * frees such a skb without reporting a timestamp; drop
+			 * it here as well.
+			 */
+			__skb_dequeue(&list);
+			kfree_skb(skb);
+			continue;
+		}
 		if (ret)
 			break;
 		__skb_dequeue(&list);

base-commit: 3769bc31a5b79ec406df085592fd9dc14ce7df12
-- 
2.54.0