Date: Thu, 16 Jul 2026 18:24:13 +0000 From: Arthur Kiyanovski <akiyano@FreeBSD.org> To: src-committers@FreeBSD.org, dev-commits-src-all@FreeBSD.org, dev-commits-src-main@FreeBSD.org Cc: Gilad Ben Yakov <giladben@amazon.com> Subject: git: 74bcb1151bb9 - main - ena: Fix false 'missing TX completions' warnings due to timestamp race Message-ID: <6a5921cd.37d9c.4cd795c@gitrepo.freebsd.org>
index | next in thread | raw e-mail
The branch main has been updated by akiyano: URL: https://cgit.FreeBSD.org/src/commit/?id=74bcb1151bb94279a4269fc842aa2be00800545e commit 74bcb1151bb94279a4269fc842aa2be00800545e Author: Gilad Ben Yakov <giladben@amazon.com> AuthorDate: 2026-06-21 12:45:38 +0000 Commit: Arthur Kiyanovski <akiyano@FreeBSD.org> CommitDate: 2026-07-16 18:20:41 +0000 ena: Fix false 'missing TX completions' warnings due to timestamp race Sporadic 'Found a Tx that wasn't completed on time' warnings appear under sustained TX load, always reporting '1 msecs since last cleanup' despite the 5-second timeout threshold. The per-packet TX timestamp uses struct bintime (128 bits: two 64-bit fields sec and frac) which is read and written non-atomically. A race exists between the missing TX completion check (check_missing_comp_in_tx_queue reading the timestamp) and the TX submit path or cleanup path writing it on another CPU. Since the two fields are not updated atomically, the check can observe a partially written timestamp - one field from the old value and one from the new. This can produce a timestamp with {sec=0, frac=valid}, causing the check to compute a time offset equal to system uptime and falsely exceeding the 5-second timeout. Confirmed by instrumentation showing all occurrences had sec=0 with valid frac/mbuf, cleanup_running=0, and ticks==last_cleanup_ticks. Replace struct bintime with sbintime_t (a single 64-bit value) for tx_buf->timestamp. An aligned 64-bit store/load cannot be torn on 64-bit architectures. Additionally, snapshot the timestamp into a local variable in the check path to prevent a read-then-read race where the timestamp could be zeroed between the zero-check and the offset calculation. Testing: On m6i.large (FreeBSD 15.0-RELEASE-p6 amd64, 2 IO queues), two instances with MTU 1500. Ran iperf -P 20 -u -b 320kpps (CPU saturated at ~7 Gbps aggregate). Without the fix: 8 warnings in 6 hours (first at ~72 min). With the fix: 0 warnings after 20+ hours under identical conditions. Fixes: 9b8d05b8ac78 ("Add support for Amazon Elastic Network Adapter (ENA) NIC") Submitted by: Gilad Ben Yakov <giladben@amazon.com> MFC after: 2 weeks Sponsored by: Amazon, Inc. Reviewed by: cperciva Differential Revision: https://reviews.freebsd.org/D58241 --- sys/dev/ena/ena.c | 13 +++++++------ sys/dev/ena/ena.h | 2 +- sys/dev/ena/ena_datapath.c | 4 ++-- 3 files changed, 10 insertions(+), 9 deletions(-) diff --git a/sys/dev/ena/ena.c b/sys/dev/ena/ena.c index 79b9937991c1..060d9e26c1cf 100644 --- a/sys/dev/ena/ena.c +++ b/sys/dev/ena/ena.c @@ -3213,7 +3213,7 @@ check_missing_comp_in_tx_queue(struct ena_adapter *adapter, { uint32_t missed_tx = 0, new_missed_tx = 0; device_t pdev = adapter->pdev; - struct bintime curtime, time; + sbintime_t curtime; struct ena_tx_buffer *tx_buf; int time_since_last_cleanup; int missing_tx_comp_to; @@ -3222,17 +3222,18 @@ check_missing_comp_in_tx_queue(struct ena_adapter *adapter, enum ena_regs_reset_reason_types reset_reason = ENA_REGS_RESET_MISS_TX_CMPL; bool cleanup_scheduled, cleanup_running; - getbinuptime(&curtime); + curtime = getsbinuptime(); for (i = 0; i < tx_ring->ring_size; i++) { + sbintime_t ts; + tx_buf = &tx_ring->tx_buffer_info[i]; - if (bintime_isset(&tx_buf->timestamp) == 0) + ts = atomic_load_64(&tx_buf->timestamp); + if (ts == 0) continue; - time = curtime; - bintime_sub(&time, &tx_buf->timestamp); - time_offset = bttosbt(time); + time_offset = curtime - ts; if (unlikely(!atomic_load_8(&tx_ring->first_interrupt) && time_offset > 2 * adapter->missing_tx_timeout)) { diff --git a/sys/dev/ena/ena.h b/sys/dev/ena/ena.h index ca5abe494e04..77cb8730627f 100644 --- a/sys/dev/ena/ena.h +++ b/sys/dev/ena/ena.h @@ -263,7 +263,7 @@ struct ena_tx_buffer { bus_dmamap_t dmamap; /* Used to detect missing tx packets */ - struct bintime timestamp; + sbintime_t timestamp; bool print_once; #ifdef DEV_NETMAP diff --git a/sys/dev/ena/ena_datapath.c b/sys/dev/ena/ena_datapath.c index c11006cbc10f..b58779e51fc3 100644 --- a/sys/dev/ena/ena_datapath.c +++ b/sys/dev/ena/ena_datapath.c @@ -274,7 +274,7 @@ ena_tx_cleanup(struct ena_ring *tx_ring) mbuf = tx_info->mbuf; tx_info->mbuf = NULL; - bintime_clear(&tx_info->timestamp); + atomic_store_64(&tx_info->timestamp, 0); bus_dmamap_sync(adapter->tx_buf_tag, tx_info->dmamap, BUS_DMASYNC_POSTWRITE); @@ -1055,7 +1055,7 @@ ena_xmit_mbuf(struct ena_ring *tx_ring, struct mbuf **mbuf) counter_exit(); tx_info->tx_descs = nb_hw_desc; - getbinuptime(&tx_info->timestamp); + atomic_store_64(&tx_info->timestamp, getsbinuptime()); tx_info->print_once = true; tx_ring->next_to_use = ENA_TX_RING_IDX_NEXT(next_to_use,home | help
Want to link to this message? Use this
URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?6a5921cd.37d9c.4cd795c>