On Fri, 23 Oct 2026 19:12:40 +0530
Prashant Gupta <[email protected]> wrote:
> This is the final DPAA2 upstreaming sub-series, applied on top of
> v7-S1 ("dpaa2: bus, DMA and mempool base fixes") and
> v14-S2 ("net/dpaa2 updates").
>
> The series reworks the DPAA2 datapath, flow steering, metering,
> eventdev integration and fslmc infrastructure. It consolidates
> device configuration handling, improves queue and portal management,
> and adds support for additional DPNI, eventdev and datapath
> capabilities.
>
> Highlights:
> - add DMA bypass translation support and table-driven devargs parsing;
> - optimize Tx queue, channel and dequeue processing paths;
> - centralize MC version, SoC detection and endpoint handling;
> - rework RSS, flow steering, QoS and default action programming;
> - add rte_mtr support and meter rte_flow actions;
> - improve eventdev scheduling, DPCI support and multi-port operation;
> - support runtime IEEE1588 timestamping;
> - rework xstats collection and MAC statistics handling;
> - unify enqueue/dequeue datapaths across net, crypto and eventdev;
> - add Tx fast paths and reduce confirmation overhead.
>
> Every commit builds cleanly with arm64 and -Werror and the
> series remains bisectable.
>
> Breeson Francis (1):
> net/dpaa2: rework xstats with page table and MAC support
>
> Hemant Agrawal (1):
> net/dpaa2: reject Rx queue deferred start
>
> Jun Yang (63):
> net/dpaa2: refactor and optimize Rx parser result decode
> net/dpaa2: generate devargs table and string from one list
> crypto/dpaa2_sec: use table-driven devargs parsing
> dma/dpaa2: use table-driven devargs parsing
> net/dpaa2: unify debug dump via log based helpers
> net/dpaa2: expose Rx L3/L4/L5 offsets via mbuf dynfield
> net/dpaa2: optimize dpaa2 Tx queue and channel mapping
> bus/fslmc: split bus-level device walk from VFIO primitives
> bus/fslmc: centralize MC version and SoC detection in the bus layer
> net/dpaa2: remove unused soft parser loader
> net/dpaa2: cache per-rxq DPNI queue cfg and share CGID per TC
> The dpaa2 PMD so far supported only two Tx buffer release schemes:
> direct release by hardware, and an absolute Tx confirmation mode
> (drv_tx_conf=1) where every transmitted buffer is returned on a Tx
> confirmation queue that software must poll to free the buffers.
> Absolute confirmation adds a confirmation descriptor and a poll cost
> for all packets, even those the hardware could free on its own.
> net/dpaa2: support runtime IEEE1588 timestamping
> bus/fslmc: consume dpaa2 DQRR entries in batches
> bus/fslmc: align dpaa2 event port depths with hardware rings
> net/dpaa2: support multi-DPNI FS redirect and QoS table v3
> net/dpaa2: unify xstats layout and add DPMAC counter struct
> net/dpaa2: reimplement xstats_get() on top of xstats_get_by_id()
> net/dpaa2: fetch only requested MAC xstats in one MC call
> net/dpaa2: rework flow with per-table profile and group types
> net/dpaa2: reimplement RSS distribution via rte_flow
> net/dpaa2: expand IP address extracts for RSS hashing
> net/dpaa2: add policer v2 MC command and L2/L3 byte units
> net/dpaa2: support rte_mtr and METER rte_flow action
> net/dpaa2: program flow table miss actions via rte_flow
> net/dpaa2: support DCB Rx mode via rte_flow
> bus/fslmc: support DPCI peer and dynamic queues
> bus/fslmc: add alloc/free DPIO API for non-affine portals
> event/dpaa2: rework event port model and DPCI handling
> event/dpaa2: fix Tx adapter enqueue for mixed destinations
> event/dpaa2: rework self test for dynamic link and balance check
> event/dpaa2: add DPCON based standalone Rx scheduler
> event/dpaa2: add push mode to DPCON Rx scheduler
> event/dpaa2: drain Rx scheduler queues on destroy
> event/dpaa2: track and detach attached Rx queues
> net/dpaa2: add multi-port Tx API
> event/dpaa2: limit attached Rx flows per event queue
> eventdev: fix port link/unlink callback return documentation
> event/dpaa2: allow event queue link to multiple ports
> event/dpaa2: report available DPIO count and fix create cleanup
> event/dpaa2: use event port DPIO for dequeue wait
> bus/fslmc: consolidate DPIO interrupt setup and bound coalescing
> event/dpaa2: honor configured dequeue timeout
> bus/fslmc: use EAL API for DPIO interrupt
> bus/fslmc: defer DPIO interrupt setup and bind portal to worker core
> bus/fslmc: move Tx SG mempool to fslmc bus
> net/dpaa2: move DPNI endpoint lookup out of the fslmc bus
> net/dpaa2: add APIs to query device info and set default action
> net/dpaa2: consolidate dequeue paths and optimize fast path
> net/dpaa2: bind per-queue DPIO for rx-queue interrupts
> net/dpaa2: add API to clean Tx confirmation queue
> event/dpaa2: fix DQRR held buffer accounting on enqueue
> bus/fslmc: store dequeue dpio in mbuf seq dynfield
> bus/fslmc: use explicit flags for mbuf seqn encoding
> net/dpaa2: unify tx burst path for ordered and atomic enqueue
> crypto/dpaa2_sec: unify enqueue burst path for ordered enqueue
> bus/fslmc: track atomic DCA context per dequeue dpio
> bus/fslmc: move enqueue response descriptors to per-queue
> event/dpaa2: add ordered schedule type support
> bus/fslmc: make DPAA2 ORP/ODP ordering per-queue instead of per-device
> event/dpaa2: support multi-port enqueue in selftest random event
> generator
> event/dpaa2: do not fail device creation on DPCI errors
> net/dpaa2: add Tx fast path for no-confirmation case
>
> doc/guides/nics/dpaa2.rst | 29 +-
> doc/guides/rel_notes/release_26_11.rst | 10 +
> drivers/bus/fslmc/bus_fslmc_driver.h | 46 +-
> drivers/bus/fslmc/fslmc_bus.c | 321 +-
> drivers/bus/fslmc/fslmc_logs.h | 2 +
> drivers/bus/fslmc/fslmc_vfio.c | 562 +-
> drivers/bus/fslmc/fslmc_vfio.h | 72 +-
> drivers/bus/fslmc/mc/dpci.c | 37 +-
> drivers/bus/fslmc/mc/fsl_dpci.h | 12 +-
> drivers/bus/fslmc/portal/dpaa2_hw_dpci.c | 304 +-
> drivers/bus/fslmc/portal/dpaa2_hw_dpcon.c | 25 +-
> drivers/bus/fslmc/portal/dpaa2_hw_dpio.c | 668 +-
> drivers/bus/fslmc/portal/dpaa2_hw_dpio.h | 37 +-
> drivers/bus/fslmc/portal/dpaa2_hw_dprc.c | 52 +-
> drivers/bus/fslmc/portal/dpaa2_hw_pvt.h | 305 +-
> drivers/bus/fslmc/private.h | 15 +-
> .../fslmc/qbman/include/fsl_qbman_portal.h | 24 +-
> drivers/bus/fslmc/qbman/qbman_portal.c | 100 +-
> drivers/bus/fslmc/qbman/qbman_portal.h | 11 +-
> drivers/bus/fslmc/qbman/qbman_sys.h | 4 +-
> drivers/common/dpaax/dpaax_devarg.h | 69 +
> drivers/crypto/dpaa2_sec/dpaa2_sec_dpseci.c | 559 +-
> drivers/crypto/dpaa2_sec/dpaa2_sec_priv.h | 11 +-
> drivers/dma/dpaa2/dpaa2_qdma.c | 92 +-
> drivers/dma/dpaa2/dpaa2_qdma.h | 15 +-
> drivers/event/dpaa2/dpaa2_eventdev.c | 1266 ++-
> drivers/event/dpaa2/dpaa2_eventdev.h | 61 +-
> drivers/event/dpaa2/dpaa2_eventdev_selftest.c | 370 +-
> drivers/event/dpaa2/dpaa2_scheduler.c | 673 ++
> drivers/event/dpaa2/meson.build | 3 +-
> drivers/net/dpaa2/base/dpaa2_hw_dpni.c | 544 --
> drivers/net/dpaa2/base/dpaa2_hw_dpni_annot.h | 6 +-
> drivers/net/dpaa2/dpaa2_ethdev.c | 3201 ++++---
> drivers/net/dpaa2/dpaa2_ethdev.h | 997 +-
> drivers/net/dpaa2/dpaa2_flow.c | 8426 ++++++++---------
> drivers/net/dpaa2/dpaa2_meter.c | 584 ++
> drivers/net/dpaa2/dpaa2_parse_dump.h | 707 +-
> drivers/net/dpaa2/dpaa2_parser_decode.h | 1148 +++
> drivers/net/dpaa2/dpaa2_pmd_logs.h | 34 +-
> drivers/net/dpaa2/dpaa2_ptp.c | 177 +-
> drivers/net/dpaa2/dpaa2_recycle.c | 18 +-
> drivers/net/dpaa2/dpaa2_rxtx.c | 2818 +++---
> drivers/net/dpaa2/dpaa2_sparser.c | 137 -
> drivers/net/dpaa2/dpaa2_sparser.h | 202 -
> drivers/net/dpaa2/mc/dpni.c | 127 +-
> drivers/net/dpaa2/mc/fsl_dpni.h | 378 +-
> drivers/net/dpaa2/mc/fsl_dpni_cmd.h | 30 +-
> drivers/net/dpaa2/mc/fsl_net.h | 4 +-
> drivers/net/dpaa2/meson.build | 3 +-
> drivers/net/dpaa2/rte_pmd_dpaa2.h | 238 +-
> lib/eventdev/eventdev_pmd.h | 9 +-
> 51 files changed, 14514 insertions(+), 11029 deletions(-)
> create mode 100644 drivers/common/dpaax/dpaax_devarg.h
> create mode 100644 drivers/event/dpaa2/dpaa2_scheduler.c
> delete mode 100644 drivers/net/dpaa2/base/dpaa2_hw_dpni.c
> create mode 100644 drivers/net/dpaa2/dpaa2_meter.c
> create mode 100644 drivers/net/dpaa2/dpaa2_parser_decode.h
> delete mode 100644 drivers/net/dpaa2/dpaa2_sparser.c
> delete mode 100644 drivers/net/dpaa2/dpaa2_sparser.h
>
Still has lots and lots of AI review warnings and errors.
Some maybe false positives but most likely not.
Review: [PATCH v4-S3 00/65] DPAA2 refactor (Jun Yang, Prashant Gupta)
Applied with git am --3way on dpdk-next-net for-main (52ebbf7d9de8),
all 65 apply. Per-commit builds: gcc 13.3, x86_64, -Dwerror=true,
net/event/crypto/dma/raw dpaa2 drivers, with and without
RTE_LIBRTE_IEEE1588. Findings checked against post-apply source.
Series-level
============
1. Bugs in released code are fixed silently inside refactors, or are
carried forward by them. These need standalone fixes at the head of
the series with Fixes: and Cc: [email protected] so they can be
backported:
- dpaa2_flow_item_convert() allocates num items, writes
new_pattern[num] (heap overflow on every flow create). 21 keeps it.
Fixes: 9ec293434e90 ("net/dpaa2: support tunnel inner protocol flow
matching")
- RAW item length not bounded against the key buffer. 21 keeps it.
Fixes: 3f881f8d6eb0 ("net/dpaa2: support raw flow classification")
- dpaa2_sec enqueue: "~(UINT64_C(1) << *dpaa2_seqn(m) & IDXMASK)",
<< binds tighter than &, shift by >= 2^31 is UB. 56 keeps it.
Fixes: 4562de326d30 ("crypto/dpaa2_sec: support ordered queue")
- Cross-driver seqn encoding mismatch: dpaa2_sec stores
"DCA | idx", net/event decode "seqn - 1" (56).
- eventdev enqueue clears DQRR_HELD/DQRR_SIZE before the enqueue
succeeds (54).
Fixes: b75061260c34 ("event/dpaa2: add enqueue and dequeue functionality")
- dpaa2_eth_eventq_attach(): "LOWEST / queue_conf->ev.priority"
divides by zero for priority 0 (highest). 34 replaces it with a
new divide by zero for num_priorities == 1.
Fixes: 3835cc228f32 ("event/dpaa2: set priority as per DPCON device")
- priv->ep_dev_type/ep_name are never set in main, so
rte_pmd_dpaa2_ep_name() returns "" and the speed capability MAC
check is dead. 49 fixes it without saying so.
Fixes: a0f8ddc41218 ("net/dpaa2: add API to get endpoint name")
Fixes: c78fa95fa9b0 ("net/dpaa2: set speed capabilities based on actual
MAC")
- nb_tx_queues = num_tx_tcs * num_channels can exceed tx_vq[16] (07).
Fixes: 72100f0dee21 ("net/dpaa2: support level 2 in traffic management")
- dpaax_ptp.h compares be16 udp->dst_port to host order 319/320; UDP
PTP never matches on LE. 13 makes dpaa2 depend on it.
Fixes: 6cc1246b3b1a ("net/dpaa: support IEEE 1588 by timesync API")
- qbman_swp_interrupt_set_inhibit() writes 0xffffffff (44).
Fixes: 293c0ca94c36 ("bus/fslmc: support memory backed portals with QBMAN
5.0")
- eventdev create leaks DPCON/DPCI on setup_dpci failure, and
info_get counts DPIOs already in use (42).
Fixes: 8a5d7a8ec74b ("event/dpaa2: initialize device")
Fixes: 342f299568ea ("event/dpaa2: return correct number of event ports")
- tx_burst returns short on rte_vlan_insert() failure (57 keeps
it).
2. Build failures on for-main (gcc 13, -Werror):
- 08..65: fslmc_vfio.h adds #include <rte_vfio.h>, which no longer
exists after the VFIO rework in main (now dev_vfio.h). Every
patch from 08 on fails. Rebase on current for-main.
- 21..65: -Warray-bounds in dpaa2_flow_extract_support():
uint8_t mask[64], i;
rte_memcpy(mask, mask_support, size); /* uint8_t size */
Bound size (if (size > sizeof(mask)) return -EINVAL;) and use
memcpy.
- 42/43: rte_atomic16_read() is gone from for-main; implicit
declaration. Replaced in 44.
- 12: DPAA2_TX_CONF_ENABLE removed but still used under
RTE_LIBRTE_IEEE1588 until 13.
With the vfio include and the mask bound patched locally, every
other commit builds and the tip links.
Functional bisect breaks:
- 29/30: enqueue checks priv->evq_info[q].event_port which nothing
sets until 31; no event can be enqueued.
- 31: dequeue_wait still uses DPAA2_PER_LCORE_DPIO which dequeue no
longer affines; NULL deref until 43/44.
- 57/58: dqrr_held[] stores ~(1 << idx) then does HELD &= ~held,
corrupting held state until 59.
- 08: MC firmware version check removed, restored in 09.
3. Squash or reorder: 12+13, 17..20 (xstats rewritten four times, 17
has wrong name/id mapping), 23 before 22, 34+35, 44+47+52 (DPIO intr
API reworked three times), 49 before 17 (all DPAA2_MAC checks in
17..20 are dead until 49), 59 into 57/58.
4. Too large to review or bisect: 01 (~1800 lines), 12 (~1400 lines
in rxtx.c), 21 (10k lines), 31 (4 drivers), 51, 62. Split refactor
(no behaviour change) from behaviour changes and from new features.
Several commit logs do not describe the diff (01, 12, 18, 21, 37,
47, 63).
5. Eventdev API problems: enqueue only accepted to queues linked to
the calling port (29, 41; main crashes instead, still wrong);
port_link replaces links instead of
adding (31); ORDERED dropped (29) and re-added (61); Rx adapter
queue_id overwritten with the ethdev rxq id (31); 10 ms delay in
enqueue on unlinked ports (31). The selftest (33) was rewritten
around these restrictions. Run dpdk-test-eventdev pipeline/perf
against the series.
6. 47 and 52 write "performance" to cpufreq scaling_governor and
rewrite /proc/irq/N/smp_affinity_list from the event and Rx fast
paths, using rte_lcore_id() as a CPU number. Host power policy is
not a PMD decision. Drop it.
7. getenv() for device config: new DPAA2_SCHEDULE_RX_PREFETCH (34);
44 extends the existing DPAA2_PORTAL_INTR_* env vars to event and
rx-intr DPIOs. Use devargs or API arguments. (Removing env vars in
01, 21, 62 is welcome; note it in release notes.)
8. New PMD-specific API surface: rte_pmd_dpaa2_rx_get_offset,
flow_table_query, rte_dpaa2_flow_group_set_miss_actions,
rte_dpaa2_scheduler_*, rte_dpaa2_dev_tx_multi_ports,
rx_queue_info_get, dev_info_get, configure_default_action,
clean_tx_conf, and group type bits encoded into rte_flow
attr->group. None have doxygen or release notes. Mixed rte_dpaa2_
and rte_pmd_dpaa2_ prefixes. Justify on the list why these cannot
be done with generic ethdev/rte_flow/eventdev API.
9. Commit 12 has no subject; the first body paragraph became the
Subject header.
10. 21 rewrites the flow engine on top of the current for-main
dpaa2_flow.c and drops three fixes just merged there:
- 5384de9604 MPLS key width 4 bytes: back to 3, every extract
after MPLS is one byte off.
- a441f8345c minimum key size only for DPNI >= 8.7: version
check dropped, older DPNI get 24-byte entries.
- fa6bed304e default miss queue = last queue of the TC: back to
the first queue; its 26.11 release note is now wrong.
Rebase the rework onto these.
Patch findings
==============
[01/65] net/dpaa2: refactor and optimize Rx parser result decode
Error: VLAN ptype wrong, QinQ never detected
dpaa2_dev_rx_parse_vlan() gets the summary after
"frc_u.u16 &= ~DPAA2_PKT_TYPE_VLAN", so "sum->vlan > 1" is never
true. RTE_PTYPE_L2_* is an enumerated field: ETHER(1) | VLAN(6) == 7
== QINQ, so every single-tagged frame is reported as QinQ. Same for
VLAN|ARP in parse_frc.
Fix: pass the unmasked summary;
*pt = (*pt & ~RTE_PTYPE_L2_MASK) |
(sum->vlan > 1 ? RTE_PTYPE_L2_ETHER_QINQ
: RTE_PTYPE_L2_ETHER_VLAN);
Error: IEEE1588 Rx timestamp no longer recorded
The RTE_LIBRTE_IEEE1588 block in dpaa2_dev_prefetch_rx_common() is
deleted and parse_frc() never sets RX_IEEE1588_TMST.
Fix: keep the block; set PTP|TMST from fas->ptp in parse_frc().
Error: rejected Tx SG FD freed with the Rx SG converter
dpaa2_dev_free_eqresp_buf() calls dpaa2_eth_sg_fd_to_mbuf() on a Tx
FD. Tx SGEs hold data addresses, not buffer starts, so
DPAA2_INLINE_MBUF_FROM_BUF() yields a bogus mbuf that is written and
freed; inline SGT is freed twice.
Fix: Tx-specific free (recover the mbuf stored for the FD).
Warning: !by_channel guard on the error queue drain removed from the
prefetch path; reintroduces the portal conflict the comment in
main describes. Restore it.
Warning: dpaa2_dump_print() logs to RTE_LOGTYPE_USER1. Use
dpaa2_logtype_pmd. 05 claims driver-logtype filtering as its goal.
Warning: eCPRI type 4 dump memcpy()s wire bytes into a bitfield over a
be32 word; rma_id/ele_id are swapped. Main assembled bytes
explicitly.
[02/65] net/dpaa2: generate devargs table and string from one list
Error: secondary process clears primary's priv->flags
"priv->flags = 0; dpaax_get_devargs(...)" now runs before the
secondary early return. priv is shared dev_private.
Fix: parse/reset flags in the primary only.
Warning: devarg semantics changed (also 03, 04). Old handler accepted
only "1"; now atoi(): "=0" clears, "=yes" clears silently, key-only
args pass value == NULL to handlers that atoi() it (03, 06 crash).
Fix: check pair->value != NULL once in dpaax_get_devargs(), parse
with strtoul + end check, document the change.
[03/65] crypto/dpaa2_sec: use table-driven devargs parsing
Error: drv_strict_order can never take effect
dpaax_get_devargs() now runs at the top of dpaa2_sec_dev_init(); the
primary path later does "internals->en_loose_ordered = true;".
Fix: set the default before parsing, drop the later assignment.
Error: atoi(pair->value) with value NULL on "drv_strict_order" with no
"=" crashes probe. Same in 06 (drv_parse_protocol_offset).
Warning: dump_level and en_loose_ordered are written by secondary
probe into shared dev_private. Primary only.
[06/65] net/dpaa2: expose Rx L3/L4/L5 offsets via mbuf dynfield
Warning: rte_pmd_dpaa2_rx_get_offset() returns stale dynfield data
for eventdev and loopback Rx paths (never filled there). Per-packet
DPAA2_PMD_ERR. No doxygen, no doc for drv_parse_protocol_offset.
Warning: dpaa2_dev_rx_parse_offset() runs after
dpaa2_eth_sg_fd_to_mbuf() and reads the annotation at
DPAA2_GET_FD_ADDR(fd); for S/G that SGT buffer was already freed
with rte_pktmbuf_free_seg(). Parse inside the fd_to_mbuf helpers
before the free. 13 copies the pattern in print_parser_result().
[07/65] net/dpaa2: optimize dpaa2 Tx queue and channel mapping
Error: Tx queue to TC/channel mapping changed, not in log
"channel_id = tc_id < num_channels ? tx_channels[tc_id] :
tx_channels[num_channels - 1]": with 2 channels and 8 TCs, TC1..7
share ch1 and channels 2..N are unused. txq i moves from TC (i % tcs)
to TC (i / q_per_tc). dpaa2_tm still advertises
num_tx_tc * num_channels leaves.
Fix: keep the per-channel mapping or document and update dpaa2_tm.
Warning: the tx_vq[]/rx_vq[] caps fix a released overflow (see
series item 1). Split out with Fixes/stable.
[08/65] bus/fslmc: split bus-level device walk from VFIO primitives
Error: #include <rte_vfio.h> in fslmc_vfio.h; header no longer exists
on for-main (dev_vfio.h). Build fails from here to 65.
Warning: MC firmware version check removed here, restored in 09.
Leave it in place in 08.
[09/65] bus/fslmc: centralize MC version and SoC detection
Warning: fslmc_vfio_core_cluster_sdest() divides by
dpaa2_cluster_size before checking it for zero. Check first.
[10/65] net/dpaa2: remove unused soft parser loader
Warning: ss_offset/ss_iova/ss_param_iova are now unused; remove.
The /dev/mem O_RDWR probe moved to bus scan should be O_RDONLY /
PROT_READ and log errno, not fd.
[11/65] net/dpaa2: cache per-rxq DPNI queue cfg and share CGID per TC
Error: cgid_in_use[] shrunk to MAX_TCS (8), still indexed up to
max_cgs = attr.num_cgs (up to 255).
Fix: priv->max_cgs = RTE_MIN(attr.num_cgs, MAX_TCS).
Error: eventq_detach no longer detaches
Clears DPNI_QUEUE_OPT_DEST from options and then calls
dpni_set_queue(); MC ignores the destination, FQ stays on the DPCON.
Still so at 65. cfg is NULL after rx_queue_release or close with no
re-setup and is dereferenced unchecked.
Fix: if (!cfg) return -EINVAL; set DEST_NONE and pass
DPNI_QUEUE_OPT_DEST.
Error: rx_queue_setup err_quit frees a DPCON owned by an earlier setup
(dpcon_allocated removed). Restore the local flag.
Warning: cgid_in_use[tc]++ not undone on setup failure; CG taildrop
never programmed for that TC again.
[12/65] (no subject)
Error: subject line missing (see series item 9). Log claims Rx
FD-to-mbuf helpers are reused, omits IEEE1588 and TM taildrop
removal.
Error: bisect break with RTE_LIBRTE_IEEE1588 (series item 2).
Error: indirect mbufs sent with parent's data_off/data_len
"mi = rte_mbuf_from_indirect(mbuf); ... mbuf = mi;" then the FD is
built from mi. A clone after adj/trim (ip_frag) sends wrong bytes.
Fix: build FD from rte_pktmbuf_iova(m)/m->data_len of the indirect.
Error: indirect refcount handling is use-after-free
simple_fd drops a ref on the clone without taking one on mi; sge
helpers rte_pktmbuf_init()+free_seg() live indirects; conf path
writes "mi->next = cur_seg->next" into a shared parent.
Fix: never reinit a live mbuf; take a ref on mi or copy.
Error: NO_CONF single-segment path lets HW free shared buffers
Shared direct mbufs get mempool_to_bpid(); HW releases a buffer
others hold. Main set IVP and dropped one ref.
Fix: if (rte_mbuf_refcnt_read(m) > 1) { DPAA2_SET_FD_IVP(fd);
rte_mbuf_refcnt_update(m, -1); }
Error: inline SGT buffer gets a BPID in both FD and SGE[0]; released
twice. Set DPAA2_SET_FLE_IVP for cur_seg == sg_mbuf.
Error: conf S/G with cloned head: sg_mbuf->next points at the freed
head; double free on confirmation.
Error: per-packet helper failure returns short after the helper
already freed/relinked the mbuf; caller frees it again. Free the
packet, count a drop, continue.
Error: Tx-conf VDQ left outstanding when storage is freed; QBMan can
DMA into freed memory. Drain and dpaa2_clear_queue_active_dps()
before dpaa2_queue_storage_free().
Warning: TM software taildrop (tm_sw_td) silently removed; dpaa2_tm
still sets it.
Warning: dpaa2_dev_free_eqresp_buf() rebuilds mbuf via
rte_dpaa2_bpid_info[256] for conf FDs (BPID MAX_BPID); loopback FDs
with drv_tx_conf=1 have stash config in FLC, deref'd as mbuf.
Warning: dynamic confirm writes FAEAD at buf_addr + 0x58 with no
headroom check for foreign/extbuf mbufs.
Warning: tx_ordered "void *mq[nb_pkts]" VLA; multi_txq tx_again never
taken; tx_conf pulled once per packet.
[13/65] net/dpaa2: support runtime IEEE1588 timestamping
Error: read_tx_timestamp tests txq->ts_to_cnfd on the Tx queue; the
counter lives on tx_conf_queue. Never waits; returns stale value.
53 rewrites the line and keeps the bug.
Fix: txq->tx_conf_queue->ts_to_cnfd, bounded wait, -EAGAIN.
Error: dynfield written at offset 0 when unregistered
rx_ts_offset is registered only with RX_OFFLOAD_TIMESTAMP; after
timesync_enable() without it every Rx mbuf's buf_addr is
overwritten.
Fix: rx_ts_offset = -1 at init; write only when >= 0.
Error: tx_burst issues dpni_set_single_step_cfg() (global MC lock, up
to 1 s spin) and races on ptp_correction_offset across lcores.
Configure one-step from the control path only.
Error: S/G frames get a stale timestamp
dpaa2_dev_rx_read_timestamp() reads m->buf_addr + PTA after
dpaa2_eth_sg_fd_to_mbuf(); for S/G the annotation is in the SGT
buffer (FD addr), already freed. Main read fd_addr + PTA. Read it
inside the fd_to_mbuf helpers.
Warning: priv->rx_timestamp stored for every packet from every lcore,
not only FAS_PTP frames. Eventdev Rx paths lose timestamps.
Warning: ts_to_cnfd incremented after enqueue (conf may decrement
first). Once the wrong-counter bug is fixed, read_tx_timestamp
drains conf from the control thread with q_storage[rte_lcore_id()]
(LCORE_ID_ANY out of bounds).
Warning: per-packet WARN/INFO in Tx path. PTP Tx now needs an
undocumented drv_tx_conf/drv_tx_dyn_conf.
[14/65] bus/fslmc: consume dpaa2 DQRR entries in batches
Warning: released atomic contexts stay held until ci_flush_th is
reached; if the loop exits on nb_events they block those flows on
all ports until this lcore polls again. Flush after the release
loop.
[17/65] net/dpaa2: rework xstats with page table and MAC support
Error: xstats_get(NULL, 0) returns 0 instead of the count, so the
driver xstats disappear from size queries (telemetry). 19 keeps it;
n > count returns -EINVAL on non-DPMAC ports.
Fix: count = dpaa2_xstats_count(dev);
if (xstats == NULL || n < count) return count;
Warning: stats_get memsets stats (wipes ethdev rx_nombuf), puts HW
nobuffer in rx_nombuf, counts discards in both ierrors and imissed.
Keep main's mapping.
Warning: policer page fetched for all 8 TCs ("tc < MAX_TCS" always
true); MC error on a missing TC fails xstats. Use num_rx_tc.
Info: rte_le_to_cpu_64() dropped on MAC counters read from MC.
[18/65] net/dpaa2: unify xstats layout and add DPMAC counter struct
Warning: log says no behaviour change; all MAC/policer xstat names
change and DPAA2_MAC_NUM_STATS 46->62 requests new counters from MC.
User visible: document it.
Warning: ".pg.field = (uint64_t)RTE_STR(field)" stores pointers in
uint64_t; not a constant expression on 32-bit. Use a separate
const char * table.
[20/65] net/dpaa2: fetch only requested MAC xstats in one MC call
Error: repeated MAC ids overrun mac_val[62] on stack and the DMA
buffer; "uint16_t i" vs "uint32_t n" loops forever for n > 65535.
Fix: bound mac_num, flush in chunks; uint32_t i.
Warning: MAC DMA buffers allocated on every call and every port; an
alloc failure fails stats_get on non-MAC ports.
[21/65] net/dpaa2: rework flow with per-table profile and group types
Error: -Warray-bounds build failure (series item 2)
dpaa2_flow_extract_support(): uint8_t mask[64] filled with
rte_memcpy(mask, mask_support, size) with no bound on size.
Error: reverts for-main 5384de9604, a441f8345c, fa6bed304e (series
item 10). MPLS field_size "sizeof(uint8_t) * 3" must be 4;
dpaa2_flow_entry_size() must keep the dpaa2_dev_cmp_dpni_ver(priv,
8, 7) check; default miss queue must stay the last queue of the TC.
Error: QoS-only flows accept DROP/PORT_ID/REPRESENTED_PORT, but QoS
entries cannot drop or redirect; traffic is steered to the TC and
create returns 0. Allow only QUEUE/JUMP for QoS-only flows.
Error: s_dpaa2_flow_proto_handle[64] indexed by user item type; enum
already goes to 70 and PMD items are negative. Bounds check.
Error: actions_update[4] on stack, check_actions_support() writes one
entry per matching action plus END with no bound.
Error: dpaa2_flow_actions_update() passes flow->qos_flow to the FS
update; NULL for FS-only flows. Pass flow->fs_flow.
Error: dpaa2_flow_rss_action_update() tests items[i - 1].type != END
after the loop: rejects every non-empty pattern, reads items[-1] for
[END].
Error: RAW extract writes dpkg.extracts[raw_idx] in place and bumps
num_extracts; overwrites the IP address extract. Use
dpaa2_dpkg_insert_extract().
Error: group type bits not stripped
attr->group = id | type << 8 is compared with rxq->tc_index and used
in "group * qos_entries_per_tc + priority". QUEUE/RSS on QoS/MIX
groups always rejected; with drv_qos_entries_per_tc, QoS/MIX flows
(including the PMD's DCB flows from 27) fail with -ENOSPC.
Fix: tc = RTE_DPAA2_FLOW_GROUP_ID_GET(attr->group) and use it.
Error: IPv6 item sets flow->ip_src/ip_dst = NET_PROT_IPV4; first
destroy shrinks IPv6 extracts to 4 bytes.
Error: dpaa2_flow_create_*_flow() calls dpaa2_flow_clean() which
frees application-held flows (UAF on later rte_flow_destroy()),
before the new flow is even parsed. Return -EEXIST instead.
Error: [JUMP, END] also programs an FS entry with JUMP as its action;
log says JUMP is QoS-only.
Error: failed create leaves the shared key profile modified (holes
inserted in every flow's key) with HW not reprogrammed.
Error: PPPoE handler uses DPAA2_PARSER_MPLS_ID, fields_num 1
(session_id never extracted).
Error: IPv4 fragment_offset: any mask sets the IP_FRAG FAF bit to 1,
so "spec 0 / mask 0x3fff" (non-fragments) matches fragments only.
Error: remove_invalid_extract() returns 0 when only IP extracts are
removed; table not reprogrammed, RSS destroy returns -EINVAL and
leaks.
Error: flow_table_query derefs flow->qos_flow/fs_flow without NULL
check; crashes on QoS-only/FS-only/RSS flows.
Error: new symbols tagged with old versions
RTE_EXPORT_EXPERIMENTAL_SYMBOL(rte_dpaa2_flow_group_set_miss_actions, 23.11)
RTE_EXPORT_EXPERIMENTAL_SYMBOL(rte_pmd_dpaa2_flow_table_query, 24.11)
Neither exists in main. Use 26.11.
Error: single-TC ports now always create a QoS entry, but QoS table
config returns early for num_rx_tc <= 1.
Warning: ip_addr_offset not adjusted when a preceding non-IP extract
is removed; second IP offset uses ip_addr[1].max_size instead of
ip_addr[0].max_size.
Warning: flow_destroy does LIST_REMOVE then frees only on full
success; "ret |= -errno" ORs errnos; error never set.
Warning: query does a table dump, discards it, returns 0 for any
action. Return -ENOTSUP.
Warning: validate ignores unsupported actions that create rejects;
out-of-range group silently becomes QoS group 0.
Warning: fprintf(stderr, ...) in dpaa2_extract_ip_addr_add(); ~350
lines of static inline flow helpers in dpaa2_ethdev.h. Use
DPAA2_PMD_ERR, move helpers into dpaa2_flow.c.
Warning: dev_init returns 0 on extract_param/entry_map alloc failure.
Warning: drv_qos_entries_per_tc undocumented, atoi() into uint16_t.
Warning: commit log does not match code (env var vs devarg, flag
names, query, JUMP). Split: data model, handler table, devargs,
actions_update, miss actions, query, RSS.
[22/65] net/dpaa2: reimplement RSS distribution via rte_flow
Error: tbl_profile->rss_flow is the internal dpaa2_generic_flow, not
the rte_flow handle; rte_flow_destroy() on it does LIST_REMOVE on the
wrong type. rss_hash_update/reta_update always -EINVAL once RSS is
active.
Error: rte_pmd_dpaa2.h now includes <ethdev_driver.h> and the inline
derefs rte_eth_devices[]; installed header no longer usable by apps.
The inline also leaks the flow handle and passes queue = NULL with
queue_num > 0.
Warning: stable symbol rte_pmd_dpaa2_set_custom_hash removed (now an
experimental inline). 26.11 is an ABI window, but the removal needs
a deprecation notice in a prior release.
Warning: creating any FS flow destroys the PMD's RSS flow; after that
rss_hash_update silently returns 0.
Warning: rss_hash_update returns 0 with nothing programmed when no
RSS flow exists.
Warning: PMD-created RSS/DCB/miss flows live on the user's flow list;
rte_flow_flush() removes them and leaves rss_flow/dcb_flow[] stale
(also 26, 27). Use internal helpers.
[25/65] net/dpaa2: support rte_mtr and METER rte_flow action
Error: METER unreachable
} else if (meter_action->type == RTE_FLOW_ACTION_TYPE_METER_MARK)
repeats the first test. Compare with RTE_FLOW_ACTION_TYPE_METER.
Error: rte_mtr rates are bytes/s uint64_t; policer cir/eir are
uint32_t Kbps. Programs 125x too high and truncates.
Fix: cir_kbps = cir * 8 / 1000; reject > UINT32_MAX at profile_add.
Warning: profile/policy delete frees meters still in use (no -EBUSY),
duplicate ids accepted, meter_destroy ignores bound flows; objects
leak on close; rte_zmalloc for CPU-only lists.
Warning: rte_spinlock_init(&priv->meter_lock) in mtr_ops_get, which
runs on every rte_mtr call. Control ops are not thread safe by
convention; drop the lock.
Warning: packet_mode > 1 used as a hidden unit selector; meter flow
accepts any pattern; DO_NOT_RESET_COUNTERS sent to old MC.
Warning: dpaa2_flow_create_rss_flow() destroys user meter flows on
reconfigure; later rte_flow_destroy() is use-after-free.
Warning: meter_profile_update/meter_policy_update only store the new
id; the DPNI policer is never reprogrammed, so updating a bound meter
has no effect. Reprogram or return -ENOTSUP while in use.
[26/65] net/dpaa2: program flow table miss actions via rte_flow
Warning: miss-action errors only logged in dev_configure; QoS call
made when qos_entries == 0 logs ERR on every configure. Also ~170
lines of unrelated code motion and errno changes not in the log.
[27/65] net/dpaa2: support DCB Rx mode via rte_flow
Error: dcb_tc[] is indexed by priority and holds the TC; code treats
it the other way round.
Fix: for (prio = 0; prio < RTE_ETH_DCB_NUM_USER_PRIORITIES; prio++)
dpaa2_setup_flow_dcb_dist(dev, prio, dcb_tc[prio]);
Error: dcb_flow[] cleared only at close; stop/configure fails with
-ENODATA on the second configure.
Warning: dcb_info reports identity mapping, not the configured one.
[29/65] bus/fslmc: support DPCI peer and dynamic queues
Error: "uint16_t expected = 0;" moved out of TAILQ_FOREACH in
rte_dpaa2_alloc_dpci_dev(); a failed CAS leaves expected == 1, so
the next in-use DPCI is handed out again. Same in 30 for DPIO.
Fix: declare expected inside the loop.
Error: dpci_get_peer_attributes() decodes with dpci_rsp_get_attr;
use dpci_rsp_get_peer_attr (num_of_priorities at byte 8).
Error: DPCI Rx DQ storage alloc removed; raw/dpaa2_cmdif still does
rxq->q_storage[0]->dq_storage[0] -> NULL deref.
Error: enqueue validation failures "goto send_partial" without
frames_to_send = loop; garbage descriptors pushed to QBMan (29-30,
fixed in 31). rte_errno never set (still at 65).
Error: dpaa2_eventdev_create() do/while over free DPCON/DPCI pairs
has no bound; more than 16 pairs write past priv->evq_info[16].
Fixed in 41/42; keep the bounded for loop here.
Warning: ORDERED and NONSEQ_MODE dropped while QUEUE_ALL_TYPES still
advertised; not in log. 61 re-adds them; do not drop them here.
Warning: free does fetch_sub on a 0/1 flag; no detach resets cb/dest,
so a freed DPCI returns -EBUSY on reuse.
[31/65] event/dpaa2: rework event port model and DPCI handling
Error: DCA consumed on the wrong portal
Dequeue records held DQRR on portal->dpio_dev; net Tx, Tx adapter
and dpaa2_sec still use DPAA2_PER_LCORE_PORTAL/DQRR_*.
Error: port_link: "ret = rte_event_port_unlink(...); if (ret) return
ret;" unlink returns the count unlinked; second link returns it
without linking (31-40). Linking must add, not replace.
Error: dequeue_wait still uses DPAA2_PER_LCORE_DPIO; dev_stop drains
with dequeue_burst(.., 1, 1) from the control thread, so
rte_event_dev_stop() can crash even with timeout 0 (until 43).
Error: dequeue counts an unwritten event when rxq has no cb.
Error: enqueue livelock: after validation error or env_pool empty,
"while (nb_events)" restarts at the same event forever. Return
num_tx with rte_errno.
Error: dpaa2_ethq->ev.queue_id = eth_rx_queue_id overrides the Rx
adapter's event queue id; flow_id forced regardless of
FLOW_ID_VALID.
Warning: rte_delay_us(10000) in enqueue/dequeue on unlinked ports,
under port_lock.
Warning: Rx adapter queue_del made a no-op until 37.
Warning: RTE_DPAA2_EVENT_PORT_CFG_ATOMIC uses bit 31 of the library
owned event_port_cfg, declared in the net header.
Warning: 1024-entry mempool with 512 cache per DPCI Rx queue at probe
(~1 MB hugepage each), whether used or not; caps in-flight events at
1024 while max_num_events reports INT32_MAX - 1.
[32/65] event/dpaa2: fix Tx adapter enqueue for mixed destinations
Warning: SAME_DEST guarantees one port/queue; the old code was right.
Dropping rte_eth_tx_burst() breaks non-dpaa2 ethdevs, which caps
report as INTERNAL_PORT and get cast to struct dpaa2_queue. Drop or
restrict caps to dpaa2 ports.
[34/65] event/dpaa2: add DPCON based standalone Rx scheduler
Error: pull size not clamped: nb_pkts passed to the pull descriptor
instead of pull_size; storage holds dpaa2_dqrr_size entries.
Error: "uint8_t priority_step = 256 / num_priorities" is 0 for
num_priorities == 1; then divides by it. Use unsigned int.
Warning: getenv("DPAA2_SCHEDULE_RX_PREFETCH"). Use the mode argument
added in 35.
Warning: destroy leaks DQ storage; q_storage[rte_lcore_id()] with no
LCORE_ID_ANY check; scheduler API declared in the net header, no
doxygen; max_prio returns num_priorities not num_priorities - 1.
[35/65] event/dpaa2: add push mode to DPCON Rx scheduler
Error: first rx on each lcore issues MC commands (attach, static
dequeue channel) and writes shared port_queue[].configured from
concurrent workers. Attach on the control path.
Error: rte_dpaa2_scheduler_max_prio() derefs dpcon_dev[0], NULL in
push mode until lcore 0 receives.
Warning: push mode allocates an unused DPCI per lcore.
Warning: push mode leaves queue_conf zeroed; sched_type 0 is ORDERED,
so every rxq gets the ordered cb and an ORP. Set PARALLEL for both
modes in scheduler_add().
[36/65] event/dpaa2: drain Rx scheduler queues on destroy
Warning: drains other lcores' portals from the control thread;
prefetch pulls left by workers never consumed; rte_delay_ms(1) per
drained packet.
[37/65] event/dpaa2: track and detach attached Rx queues
Error: push detach loops attach_num (lcore count) over port_queue[];
NULL eth_dev or rxqs left bound to a freed DPCON.
Error: queue_del(-1) detaches every ethdev's rxqs, not this port's.
Error: Rx adapter stop drains all event ports from the control thread
while workers run, then detaches; start is a no-op, so stop/start
loses Rx. Log does not match.
Error: evq_info[].dpni_rxqs[] keep rxq pointers past ethdev close;
eventdev close detaches freed memory with priv->hw NULL.
Warning: attach not idempotent (re-add double counts); close skipped
when never started (leaks DPIOs).
[38/65] net/dpaa2: add multi-port Tx API
Warning: "struct dpaa2_queue *txq[nb_pkts]" VLA; no port/txq_id
validation; no doxygen; per-packet failure returns short; copy-path
mbufs freed before enqueue are reported unsent (double free).
[39/65] event/dpaa2: limit attached Rx flows per event queue
Error: max_event_queue_flows lowered to 1024 while queue_def_conf
still sets nb_atomic_flows 2048; rte_event_queue_setup() with
defaults fails. Use a separate attach bound.
Warning: dpni_rxq_num widened to uint16_t but eth_queue_del() loops
with uint8_t j, k; more than 255 rxqs never terminates.
[40/65] eventdev: fix port link/unlink callback return documentation
Warning: library change in a vendor series; send standalone to the
eventdev maintainer with Fixes:. "Number of unlinks successfully
established" reads wrong; drop the NXP copyright on a Cavium header
for a comment fix.
[41/65] event/dpaa2: allow event queue link to multiple ports
Warning: port_unlink indexes pos[16]/evq_infos[16] by nb_unlinks,
which can exceed 16 with repeated queue ids. Stack overflow.
Warning: enqueue still rejects queues not linked to the port
(series item 5). Main crashes there instead; both violate the API.
[42/65] event/dpaa2: report available DPIO count and fix create cleanup
Error: rte_atomic16_read() no longer exists in for-main; 42 and 43
fail to build (implicit declaration).
Fix: rte_atomic_load_explicit(&dpio_dev->ref_count,
rte_memory_order_relaxed)
Warning: failed create still calls event_dev_probing_finish() and
never releases the eventdev (also 64). The create leak and info_get
count fixes need Fixes/stable (series item 1).
Warning: DPAA2_EVENTQ_LINK_CONF_MAX rewrite and unused
DPAA2_EVENT_MAX_PORTS not in log.
[44/65] bus/fslmc: consolidate DPIO interrupt setup
Warning: existing getenv() thresholds now applied in the bus
allocator to every interrupt DPIO; sscanf return unchecked. The
set_inhibit change is a fix (series item 1).
Warning: ethrx portals now armed at alloc; net rx_queue_intr_enable
then returns early and gets holdoff 0xff. 52 reverts it.
[45/65] event/dpaa2: honor configured dequeue timeout
Warning: timeout_ticks honoured without PER_DEQUEUE_TIMEOUT, and in
global mode timeout_ticks == 0 still waits dequeue_timeout_ns;
64-bit divide on every burst; ns * event_hz overflows; ms truncated
to int for epoll_wait.
[46/65] bus/fslmc: use EAL API for DPIO interrupt
Warning: event driver logs with DPAA2_PMD_DP_DEBUG (net logtype,
reached through an accidental include chain); dpaa2_logtype_pmd is
not exported, so shared builds with DP debug fail to link. Use the
eventdev log macro.
[47/65] bus/fslmc: defer DPIO interrupt setup and bind portal to worker
Error: driver writes cpufreq scaling_governor and /proc/irq affinity
from dpaa2_event_portal_io_prepare() (fast path). Drop the governor
code; do affinity/stashing at port setup.
Error: "cpu = rte_lcore_id()" used as physical CPU for IRQ affinity,
governor and stashing. Use dpaa2_get_core_id().
Warning: eventfd leaked per intr init/deinit cycle.
Warning: log mentions an "unlikely" hunk that does not exist; three
helpers exported then made static in 52.
[48/65] bus/fslmc: move Tx SG mempool to fslmc bus
Error: Rx SG walk bound removed
- while (!DPAA2_SG_IS_FINAL(sge) && i < DPAA2_MAX_SGS) {
+ while (!DPAA2_SG_IS_FINAL(sge)) {
Log claims it is now bounded. Restore the bound.
Warning: SG pool created unconditionally at bus probe, also for
crypto-only and secondary; NULL result ignored.
[49/65] net/dpaa2: move DPNI endpoint lookup out of the fslmc bus
Warning: fixes ep_dev_type/ep_name never being set in main (series
item 1). Say so, add Fixes, move before 17.
[50/65] net/dpaa2: add APIs to query device info and set default action
Error: new symbols tagged 25.11
RTE_EXPORT_EXPERIMENTAL_SYMBOL(rte_pmd_dpaa2_rx_queue_info_get, 25.11)
RTE_EXPORT_EXPERIMENTAL_SYMBOL(rte_pmd_dpaa2_dev_info_get, 25.11)
Use 26.11. No doxygen or release notes for the three functions.
[51/65] net/dpaa2: consolidate dequeue paths and optimize fast path
Error: scheduler prefetch priming inverted
"if (likely(q_storage->active_dqs))" never primes on first call,
then derefs NULL dq_storage. Use if (unlikely(!...)).
Warning: tx_conf "if (refcnt_read > 1) refcnt_update(-1)" races with
the other owner and skips chained segments. Use rte_pktmbuf_free().
Warning: error queue dump uses q_storage[0] from all lcores.
[52/65] net/dpaa2: bind per-queue DPIO for rx-queue interrupts
Error: Rx burst now sets IRQ affinity, stashing and governor on first
call (see 47). Releasing a queue or port never polled passes
cpu_affine -1 to the governor restore, which restores every CPU,
including ones other queues still use.
Warning: intr_enable returns early when napi_armed, CDAN is one-shot;
dev_started check removed. Holdoff default changed from 0 to 0xff.
[53/65] net/dpaa2: add API to clean Tx confirmation queue
Error: rte_pmd_dpaa2_clean_tx_conf() uses tx_queues[txq_id] without
checking txq_id < nb_tx_queues or NULL.
Warning: no doxygen; must run on the lcore owning the Tx queue (races
with tx_burst).
[54/65] event/dpaa2: fix DQRR held buffer accounting on enqueue
Warning: bug is in released code. Add
Fixes: b75061260c34 ("event/dpaa2: add enqueue and dequeue functionality")
Cc: [email protected]
and send ahead of the refactor.
[55/65] bus/fslmc: store dequeue dpio in mbuf seq dynfield
Warning: seqn dynfield grows to 16 bytes, 8-aligned, registered at
bus probe for every user. Store a DPIO index, not a pointer.
[56/65] bus/fslmc: use explicit flags for mbuf seqn encoding
Error: rewrites the UB shift in dpaa2_sec and keeps it (series 1).
Fix: ~(UINT64_C(1) << (*dpaa2_seqn(m) & QBMAN_EQCR_DCA_IDXMASK))
Warning: cross-driver encoding mismatch is a real bug; log calls it
"fragile". dpaa2_sec is even inconsistent with itself (ordered path
decodes seqn - 1, dequeue stores DCA | idx). Describe it,
Fixes/stable.
[57/65] net/dpaa2: unify tx burst path for ordered and atomic enqueue
Error: dqrr_held[loop] = ~(1 << idx) then HELD &= ~dqrr_held[i];
clears all other held bits. Fixed in 59; fold it here.
Error: "uint32_t flags[MAX_TX_RING_SLOTS]" uninitialized, restored
into seqn of every unsent mbuf. Zero per slot, restore only if set.
Error: enqueue retry exhaustion: break with ret == 0, then
"if (nb_pkts > 0 && !ret && !order) goto tx_again" skips the unsent
prepared frames, which the copy path already freed. Return value is
no longer a prefix. Same in 65.
Info: rte_vlan_insert() failure still returns short (pre-existing,
series item 1).
[58/65] crypto/dpaa2_sec: unify enqueue burst path for ordered enqueue
Error: "int32_t ret, order = false;" then "if (ret >= 0 && order)"
reads ret uninitialized when ops[0] is ordered. int32_t ret = 0.
[59/65] bus/fslmc: track atomic DCA context per dequeue dpio
Error: on DPIO mismatch the mbuf keeps DCA|idx and the origin portal
keeps HELD and held_bufs.mbuf[idx]; origin later writes through a
freed/reused mbuf. Clear seqn on mismatch, release by index only.
Warning: statement-expression macros test &eqdesc[loop]/&flags[loop]
for NULL and assign unparenthesized args. Make them static inline
functions.
[60/65] bus/fslmc: move enqueue response descriptors to per-queue
Error: every ORP frame uses priv->tx_vq[0]->eq_resp from all lcores
without a lock; the old ring was per DPIO.
Error: responses recorded on tx_vq[0] but only the caller's queues
are drained; Tx on queue != 0 stops after 127 ORP frames.
Error: slots reserved at build time never complete for frames not
enqueued; queue runs out and dpaa2_queue_eqresp_free() spins forever.
Error: dpaa2_sec_queue_pair_release() frees fle_pool, then
dpaa2_queue_eqresp_free() puts FLEs back into it. Swap the order.
Warning: eqresp pool leaked on tx_queue_setup re-setup;
DPAA2_ORDER_DEBUG dead code and USER1 logger in bus header; add
RTE_BUILD_BUG_ON(!rte_is_power_of_2(DPAA2_EQRESP_MAX_NUM)).
[61/65] event/dpaa2: add ordered schedule type support
Warning: ORP restore keyed on event->sched_type (the next queue's
type); ordered -> atomic skips restoration, ordered -> parallel or
direct Tx leaks the ORP flag. Decide on the seqn flag.
Warning: ORDERED accepted but RTE_EVENT_DEV_CAP_ORDERED not set and
ordered_scheduling not in features/dpaa2.ini.
Warning: dpaa2_dq_order_save() writes the mbuf dynfield for non-mbuf
events.
[62/65] bus/fslmc: make DPAA2 ORP/ODP ordering per-queue
Error: strict ORP on a DPCI queue uses a NULL eq_resp pool
The strict bit now comes from the mbuf. An event from a strict-ORP
DPNI or dpaa2_sec queue forwarded to an ORDERED event queue reaches
dpaa2_eq_set_order_desc() with a DPCI tx queue, which never got
dpaa2_queue_eqresp_create(): pool_len 0, mask 0xffffffff, and QBMan
DMAs the response to IOVA(NULL + n * 64). Nothing drains it.
Fix: no response (loose) on the eventdev path; return -EINVAL when
eqresp_pool is NULL.
Warning: odpid not masked; bits 13..15 land in STRICT/ORP/DCA.
((id) & DPAA2_EQCR_OPRID_MASK) << DPAA2_EQCR_OPRID_SHIFT
Warning: bus, net, crypto, event and a devarg in one patch. Removing
DPAA2_STRICT_ORDERING_ENABLE silently drops strict ordering for
existing users (the log claims it is still honoured); new devarg
name differs from dpaa2_sec's drv_strict_order. Split, reuse the
name, release note.
[64/65] event/dpaa2: do not fail device creation on DPCI errors
Warning: -ENODEV path leaves a registered, probed eventdev. Release
before event_dev_probing_finish(). Partially attached queues still
accept all sched types in queue_setup.
[65/65] net/dpaa2: add Tx fast path for no-confirmation case
Error: dpaa2_dev_tx_fast() ignores DPAA2_IEEE1588_TX_TS_FLAG; one-step
PTP stops working on the default NO_CONF config. Take the fast path
only when the flag is clear.
Error: retry exhaustion hole duplicated from 57.
Warning: "struct qbman_fd fd_arr[MAX_TX_RING_SLOTS] = {0};" zeroes
1 KiB per burst before dispatching to the fast path. Move it.
