Enable TSN pacing for E830 NIC#1635
Merged
Merged
Conversation
Sakoram
requested review from
DawidWesierski4,
awilczyns,
moleksy and
soopel
as code owners
June 29, 2026 08:43
Sakoram
force-pushed
the
tx-time-new-patches2
branch
from
June 29, 2026 08:52
a49d07d to
65d3daf
Compare
DawidWesierski4
requested changes
Jun 29, 2026
Sakoram
force-pushed
the
tx-time-new-patches2
branch
6 times, most recently
from
July 8, 2026 10:59
619ae11 to
4b53c97
Compare
DawidWesierski4
left a comment
DawidWesierski4
left a comment
Collaborator
There was a problem hiding this comment.
Weak review to NOCTX tests and UNIT tests
Rest looks oki?
Sakoram
force-pushed
the
tx-time-new-patches2
branch
3 times, most recently
from
July 15, 2026 10:59
04afc92 to
1bd916a
Compare
DawidWesierski4
approved these changes
Jul 15, 2026
Promote pacing structures (audio, video, ancillary, fastmetadata) and their pacing cursors such as trs, frame_time, tsc_time_cursor, and ptp_time_cursor to long double. Ensure pacing calculations are carried out inside high-precision types and utilize nextafterl() instead of double-precision variants to prevent precision loss. Modify logging and print patterns to use %Lf for the updated members. Signed-off-by: Kasiewicz, Marek <marek.kasiewicz@intel.com>
The out-of-tree ICE driver's single hardware TX timestamp slot can delay t3 availability or latch older timestamps if they're not drained. This forces the async polling path to be resilient against stale timestamps and racing completion. Stiffened the async polling path: - Introduced a lock-free claim step (rte_atomic32_t) to ensure exactly-once delta result processing when t3 or t4 asynchronously complete. - Set the polling retry interval to 200us (up from 50us) to significantly reduce the PMD call frequency during heavy TX loads, lowering logging pressure when the hardware slot is busy. - Added strict sequence-id check to verify t3 belongs to the current measurement cycle, plus an egress guard ensuring the reading exceeds the live t2 value to ignore stale buffered logs. - Downgraded the recoverable single-cycle sync timeout status log level from error to warn. - Corrected unit tests in tests/unit/ptp/ to accurately emulate and validate the correct clock synchronization behavior, ensuring robust and non-overfitted test execution. Signed-off-by: Kasiewicz, Marek <marek.kasiewicz@intel.com>
Enable hardware launch-time (TxPP) pacing on E830 NICs used by the TSN pacing path. Register the TX timestamp dynflag/dynfield via the public mbuf dynfield API and enable RTE_ETH_TX_OFFLOAD_SEND_ON_TIMESTAMP on both port and per-queue config. Drop the PTP-locked guard in the launch-time tasklet so pacing no longer falls back to TSC, and widen timesync retry count for slower lock. Reject ST21_TX_PACING_WAY_TSN at pacing-select time on ports that don't advertise the offload, instead of silently falling through to another pacing way. Carries five ICE driver patches: TxPP timer association, PHC read-clock, 19-bit launch-time encoding, direct MMIO PHC update, and gating the send-on-timestamp offload to E830 only. Also carries the PHC-owner-init patch: on multi-port ICE devices only the PCI function reporting func_caps.ts_func_info.src_tmr_owned may start the shared PHC counter, but the PMD only does so lazily inside ice_timesync_enable() on the port the application calls it on -- if timesync is enabled only on a non-owning port, the owner's PHC never starts and rte_eth_timesync_read_time() on that port always returns 0. Bundle all six patches together with a single DPDK bump to 26.03.92_mtl so the version-gated install check in build_dpdk.sh cannot silently skip applying a subset of them. Signed-off-by: Kasiewicz, Marek <marek.kasiewicz@intel.com>
A PTP step forward that lands the clock beyond max_onward_epochs was handled inconsistently between the onward and late branches of calc_frame_count_since_epoch()/tx_ancillary_calc_epoch(): - The onward branch resynced to current_epoch only when required_tai was already zero, so a caller-supplied user timestamp kept the stale next_free_epoch/next_free_frame_slot forever, never catching up to the stepped clock. - The late branch unconditionally overwrote a valid required_tai with frame_count_tai, discarding the caller's timestamp. - A required_tai that rounds down to epoch/frame_count 0 collided with the sentinel used elsewhere to mean "no user timestamp", silently falling back to whatever stale epoch was already in flight. - EXACT_USER_PACING with a zero timestamp went unflagged in stat_error_user_timestamp for callers only ever driving the frame-level tv_pacing_required_tai()/tx_ancillary_pacing_required_tai() path (RTP-level pacing calls calc_frame_count_since_epoch() directly with a non-zero required_tai per packet, so flagging it there would have produced false positives every packet). Guard the onward-range resync with the onward-gap check regardless of required_tai, preserve required_tai through both branches unless it is absent, fall back to real time when required_tai rounds to zero, and move the zero-timestamp stat increment to the frame-level helpers. Signed-off-by: Kasiewicz, Marek <marek.kasiewicz@intel.com>
Adds a white-box harness that includes st_tx_video_session.c directly so calc_frame_count_since_epoch(), tv_sync_pacing(), and tv_pacing_required_tai() can be exercised without DPDK EAL, mbuf pools, or packet I/O. mt_get_tsc() is mocked with a preprocessor seam local to the harness: mt_main.h is included first under its real name so the genuine definition compiles normally, then mt_get_tsc is redefined to the mock before including the production .c file, redirecting only the call sites written inside it. mt_get_ptp_time() reuses the existing production ptp_get_time_fn extensibility point instead. 14 epoch-recovery tests and 5 pacing tests cover the onward/late resync paths, the epoch-0 sentinel collision, required_tai preservation across resync, and the EXACT_USER_PACING zero-timestamp stat. Signed-off-by: Kasiewicz, Marek <marek.kasiewicz@intel.com>
Simulates a PHC seeded ahead of the true grandmaster and later corrected by the servo, reproducing the epoch/pacing stuck-recovery bug fixed in st_tx_video_session.c. Parametrized over every st21_tx_pacing_way, since calc_frame_count_since_epoch() runs identically regardless of pacing way. Signed-off-by: Kasiewicz, Marek <marek.kasiewicz@intel.com>
Most NoCtxTest cases only need a single TX/RX port pair; only the redundant-session tests need all 4. Requiring 4 ports up front made it impossible to run the suite (including TSN/launch-time-pacing tests, which need a PF rather than a VF) on hosts with just 2 ports bound. run.sh keeps requiring all 4 TEST_PORT_1..4 slots, since silently building --port_list from whichever ports happen to be set left the redundant-session tests failing individually and ambiguously instead of the script refusing to run up front. Add a dedicated run_pf.sh for the PF-only TSN/launch-time-pacing case: it requires just TEST_PF_PORT_1/2 and filters to NoCtxTest cases named with a `_pf_` infix, so the two port sets are never conflated. Rename the TSN pacing test to st20p_tx_epoch_onward_recovers_after_ptp_step_pf_tsn_pacing to match. TSN launch-time pacing is currently only validated on Intel E830 (PCI device 12d2); note this in noctx.md rather than gating on device ID in the test itself. Mirror the same split in the MCP server: run_noctx_tests() keeps the strict 4-port requirement and excludes `_pf_` tests, and a new run_noctx_pf_tests() tool auto-discovers 2 PF ports (absence of a physfn symlink) and runs the `_pf_` tests against them. Its gtest_filter enumerates by the caller's substring first and requires `_pf_` as a separate post-filter, instead of a single glob that assumed `_pf_` always precedes the caller's filter text -- the naming convention puts the descriptive prefix first (e.g. `st20p_..._pf_tsn_pacing`), so that glob ordering could never match a filter on that prefix. Signed-off-by: Kasiewicz, Marek <marek.kasiewicz@intel.com>
TSN launch-time pacing compares each packet's launch timestamp against the NIC's hardware clock (PHC). Selecting TSN pacing without the built-in PTP service enabled, or with it forced to the TSC source, leaves the timestamp in an unrelated clock domain, so the NIC releases packets at the wrong time. Reject TSN pacing selection in that case at session init instead. Document the full set of TSN launch-time pacing requirements (E830-only capability, PF-only, ptp-synced phc) in design.md. Signed-off-by: Kasiewicz, Marek <marek.kasiewicz@intel.com>
The built-in PTP service disciplines the hardware clock of the primary port only. If MTL doesn't open/use that primary port (for example, only a secondary port is passed to mtl_init), --ptp has no effect on the secondary port's clock. Signed-off-by: Kasiewicz, Marek <marek.kasiewicz@intel.com>
Parametrize every RxTxApp tx pacing_way (auto, rl, tsc, tsc_narrow, ptp, be, tsn) against both VF and PF interfaces. VF-tsn is skipped: tsn launch-time pacing has no HW PHC to compare against on a VF and is rejected at mtl_init() by dev_if_init_pacing(), so it has no positive case here. Signed-off-by: Kasiewicz, Marek <marek.kasiewicz@intel.com>
E810_DEVICE_ID lived inline in test_pacing_way.py even though NIC device IDs are cross-test facts, not test-specific state (dma.py follows the same shared-constant pattern in mtl_engine). Move it to const.py alongside new E830/E835 IDs so other tests can reuse them. Signed-off-by: Kasiewicz, Marek <marek.kasiewicz@intel.com>
RxTxApp is currently the only backend that takes pacing_way, so add the same application axis test_interlace.py/test_pacing.py use, with ffmpeg skipped, instead of hardcoding "rxtxapp" in the test body. Signed-off-by: Kasiewicz, Marek <marek.kasiewicz@intel.com>
Keep ST20/ST40 user timing authoritative across epoch recovery. Validate exact targets, unwrap ST40 media clocks, and preserve wait boundaries and completion metadata. Add deterministic pacing/stat tests. Fixes: dad214f ("Fix: Prevent stuck epoch recovery on ST20/ST40 TX pacing") Signed-off-by: Kasiewicz, Marek <marek.kasiewicz@intel.com>
ST40_TX_FLAG_USER_PACING accepted ST10_TIMESTAMP_FMT_MEDIA_CLK and converted it to TAI by reconstructing the wrap era from the current media clock, an ambiguous operation whenever scheduling jitter shifts the target across a 32-bit wrap boundary. ST20 already rejects media clock for pacing; ST40 now does the same, falling back to the default epoch-based pacing and counting a user-timestamp error. Media clock is still accepted for ST40_TX_FLAG_USER_TIMESTAMP, which only affects the reported RTP timestamp and never TAI conversion. st10_media_clk_to_ns_u64() is no longer needed and is removed along with its declaration. Signed-off-by: Kasiewicz, Marek <marek.kasiewicz@intel.com>
ST20_TX_FLAG_USER_PACING, ST22_TX_FLAG_USER_PACING, and ST40_TX_FLAG_USER_PACING (and their pipeline-level equivalents) only honor ST10_TIMESTAMP_FMT_TAI; ST10_TIMESTAMP_FMT_MEDIA_CLK falls back to default pacing. Document this on the flags and the tfmt fields of the associated frame meta structs. ST30_TX_FLAG_USER_PACING is left unchanged since audio genuinely supports media clock for pacing. Signed-off-by: Kasiewicz, Marek <marek.kasiewicz@intel.com>
tx_audio_pacing_required_tai() converted a media-clock timestamp to nanoseconds with a bare ticks/rate multiplication, treating the tick count as elapsed time since zero instead of an absolute TAI instant. Since real TAI is on the order of 10^18 ns and a 32-bit tick count at audio sampling rates converts to at most ~10^14 ns, the result was always deep in the past relative to the current PTP time, so pacing silently collapsed to "send ASAP" on every frame instead of honoring the requested time. Reject ST10_TIMESTAMP_FMT_MEDIA_CLK for ST30_TX_FLAG_USER_PACING and fall back to default epoch-based pacing, matching ST20 and ST40. Documents the TAI-only restriction on the flag and the tfmt field of st30_tx_frame_meta. Signed-off-by: Kasiewicz, Marek <marek.kasiewicz@intel.com>
st20p_harness.c and st30p_harness.c each define their own st20_rx_put_framebuff()/st30_rx_put_framebuff() (and get/reset stats) stubs to avoid touching the real transport layer, but session/ st20_harness.c and session/st30_harness.c pull in the real definitions of the same symbols by #include-ing the production st_rx_video_session.c / st_rx_audio_session.c into the same UnitTest binary. -Wl,--allow-multiple-definition silently resolves the clash by link order instead of erroring, and it picked the real, HW-backed symbol over the harness stub. The pipeline put_frame() path then called the real st20_rx_put_framebuff()/st30_rx_put_framebuff() with the harness's fake ctx->transport = 0x1 handle. MT_HANDLE_GUARD dereferenced that fake handle's type field, reading address 0x9 and crashing with SIGSEGV under ASan. Redirect the production pipeline .c's calls to distinctly named local stub functions via a #define around the #include, so the two definitions can never collide as global symbols regardless of link order. Signed-off-by: Kasiewicz, Marek <marek.kasiewicz@intel.com>
dev_if_init_pacing() rejects ST21_TX_PACING_WAY_TSN unless the built-in ptp service is enabled (MTL_FLAG_PTP_ENABLE), since the launch-time timestamp is only meaningful in the phc-synced clock domain. The st20p_tx_epoch_onward_recovers_after_ptp_step_pf_tsn_pacing case never set this flag, so mtl_init() failed with -EINVAL before the repro could even run. Set the flag for the TSN pacing_way only; TSC pacing does not compare against the phc and needs no ptp service. Signed-off-by: Kasiewicz, Marek <marek.kasiewicz@intel.com>
Adds st20p_tx_packets_are_spread_over_frame_pf_tsn_pacing, which uses E830 RX hardware timestamps (st20_rx_tp_meta::ipt_avg via ST20P_RX_FLAG_TIMING_PARSER_META) to confirm TSN launch-time pacing actually spreads packets across the frame interval on the wire, instead of only completing session setup like the existing epoch recovery test. CollectPacingSpanRatios waits up to 10s for the first complete RX frame before starting a 6s steady-state sampling window, since real PF hardware RX session establishment latency varies independently of steady-state frame arrival. Signed-off-by: Kasiewicz, Marek <marek.kasiewicz@intel.com>
Sakoram
force-pushed
the
tx-time-new-patches2
branch
from
July 15, 2026 13:40
1bd916a to
73c6049
Compare
DawidWesierski4
approved these changes
Jul 15, 2026
Sakoram
enabled auto-merge (rebase)
July 15, 2026 13:51
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Adds hardware launch-time (TxPP) pacing support for Intel E830 NICs as
an opt-in TSN pacing mode. Hardens the PTP delay-request timestamp path
and TX epoch recovery against a PTP clock step, and raises pacing math
precision to long double to eliminate accumulation error.
Changes
Enables HW launch-time pacing on E830: registers the TX
timestamp dynflag/dynfield via the mbuf dynfield API and enables
RTE_ETH_TX_OFFLOAD_SEND_ON_TIMESTAMP on port + per-queue config. Drops
the PTP-locked guard in the launch-time tasklet (no TSC fallback) and
widens the timesync retry for slower lock.
Bundles four ICE driver patches (TxPP timer association, PHC
read-clock, launch-time encoding, MMIO PHC update) plus a fifth for the
PHC owner on non-owning ports; bumps DPDK to 26.03.91_mtl.
Rejected at session init on a port without the launch-time
offload, or without PTP synced to the NIC's hardware clock (PHC).
Documents requirements (E830-only, PF-only,
--ptp) in design.md; addsa pacing_way pytest matrix across VF/PF.
Lock-free exactly-once claim for t3/t4 delta processing
against the ICE driver's single timestamp slot, plus stricter
sequence-id and stale-timestamp checks. Fixes a stuck epoch-recovery
stall where a PTP clock step left next_free_epoch stale. Adds PTP,
epoch/pacing, and NoCtx regression tests.
Pacing structures and cursors (trs, frame_time,
tsc_time_cursor, ptp_time_cursor) promoted to long double with
nextafterl() to prevent precision loss.