On Tue, 15 Sep 2026 17:04:46 +0530
Prashant Gupta <[email protected]> wrote:
> This is the second of four series upstreaming the missing NXP dpaa2
> driver changes. It does not depend on series 1 and can be applied
> independently. It adds the net/dpaa2 flow, metering and parser features:
>
> - fix an integer overflow in the CCSR region mapping,
> - set Tx confirmation on device init and support a larger burst size,
> - support MPLS and PPPoE flow distribution, meter and policing, and the
> flow drop action, with a per-device default flow miss action,
> - identify Rx mbuf hash information by FLC and add minimum key size
> support,
> - restructure the parser processing, parse tunnel and fragmented packet
> types, remove the unused soft parser driver and rename the Rx queue
> flags.
>
> Every commit builds cleanly (including the aarch64 DPAA cross build with
> -Werror) and the series is bisectable.
AI review with Opus found lots of issues that still need to be fixed.
Series: [PATCH v3-S2 00/13] net/dpaa2 flow, meter, parser rework
Patch 10 claims no functional change but rewrites the Rx path and
introduces mbuf memory corruption on every received packet. Patches 5
and 7 have functional problems of their own. Not ready.
All 13 commits apply to main and build with -Dwerror=true (default
config, RTE_LIBRTE_IEEE1588 off).
----------------------------------------------------------------------
Patch 01/13 net/dpaa2: fix integer overflow in CCSR region mapping
Warning
PAGE_SIZE and PAGE_MASK (dpaa2_recycle.c:32-35) have no users after
this patch. Remove them.
Info (pre-existing, not introduced here)
lsx_ccsr_map_region() rounds len down with page_mask and does not
add offset to the mapped length, so a region crossing a page
boundary is only partially mapped. The mapping is never unmapped.
----------------------------------------------------------------------
Patch 03/13 net/dpaa2: support larger burst size
Warning
The non-LX2160A limit drops from DPAA2_BURST_MAX (65536) to 0xF7FF.
The old value was broken anyway: max_burst_size was uint16_t, so
65536 truncated to 0. Say both in the commit message; the
truncation is a stable fix on its own.
Warning
return (dpaa2_svr_family == SVR_LX2160A) ? 229375 : 0xF7FF;
The header comment documents this as 0x37FFF. Use hex in both
places, and name the 8.7 version like the existing
DPNI_GET_MAC_SUPPORTED_IFS_VER_* defines.
Info
DPNI_BURST_LO/HI describe the MC command layout and belong in
fsl_dpni_cmd.h rather than fsl_dpni.h.
----------------------------------------------------------------------
Patch 04/13 net/dpaa2: support MPLS and PPPoE flow distribution
Warning
mask = pattern->mask ? pattern->mask : &dpaa2_flow_item_mpls_mask;
The driver default is { 0xff, 0xff, 0xff }, covering TC and the
bottom-of-stack bit. rte_flow defines a NULL mask as the item
default (label only). A spec with only the label set does not
match single-label packets (S=1).
Warning
Every MPLS item extracts NH_FLD_MPLS_MPLSL_1. For
eth / mpls / mpls the second item overwrites the first label's key
data instead of matching label 2. Reject a second MPLS item or map
it to MPLSL_2 / MPLSL_N.
Info
Extra blank line added after the #endif of the mask block.
----------------------------------------------------------------------
Patch 05/13 net/dpaa2: support meter and policing
Error
The meter ops only maintain software lists. Nothing calls
dpni_set_rx_tc_policing(), and dpaa2_flow.c has no
RTE_FLOW_ACTION_TYPE_METER handling. Creating a meter has no effect
on traffic, while capabilities, dpaa2.rst, dpaa2.ini and the
release notes advertise metering and policing.
Error
dpni_set_rx_tc_policing_v1() is declared in fsl_dpni.h and defined
nowhere. DPNI_POLICER_OPT_DO_NOT_RESET_COUNTERS and
DPNI_POLICER_UNIT_BYTES_L2_WITHOUT_FCS are unused.
Warning
s_dpaa2_mtr_capa is file-scope and written under priv->meter_lock.
The lock is per port, so two ports race on the shared struct. Fill
the caller's capa directly.
Warning
static char s_err_msg[128] is shared across ports and threads and
handed back through rte_mtr_error. Concurrent failures overwrite
each other's message.
Warning
dpaa2_mtr_profile_add() and dpaa2_mtr_policy_add() accept a
duplicate id; the new entry is appended and lookups return the
first.
Warning
Profiles, policies and meters are never freed on dev_close.
priv->profiles, priv->policies and priv->meters are referenced only
in dpaa2_meter.c.
Warning
rte_zmalloc() for control-path bookkeeping. Use malloc/calloc.
Warning
The release note hunk adds "Tx queue based flow control and
confirmation queue handling" and "software parser based packet
dump", neither of which is in this patch; patch 12 removes the soft
parser. Each patch should carry only its own note.
----------------------------------------------------------------------
Patch 07/13 net/dpaa2: set default flow miss action per device
Error
priv->default_flow = RTE_MIN(priv->fs_entries,
priv->dist_queues) - 1;
fs_miss_flow_id is a queue id within the TC. fs_entries is the FS
table size and has nothing to do with it.
The miss queue also moves from 0 to the last hardware queue of the
TC. dev_info reports max_rx_queues = priv->nb_rx_queues, so an
application that configures fewer queues never sets up or polls
that queue and loses every unmatched frame. Queue 0 is always
configured.
With fs_entries == 0 the value becomes 0xffff, which is
DPNI_FS_MISS_DROP.
Warning
The commit message calls this the "lowest priority flow". The field
is a queue id, not a priority.
----------------------------------------------------------------------
Patch 08/13 net/dpaa2: identify Rx mbuf hash information by FLC
Warning
rte_mbuf_sched_set(m, flow, tc, 0) together with
RTE_MBUF_F_RX_FDIR. hash.sched is the Tx scheduler field; RX_FDIR
tells applications to read hash.fdir from the same union. Use
hash.fdir.hi with RTE_MBUF_F_RX_FDIR_ID, or implement
RTE_FLOW_ACTION_TYPE_MARK.
Warning
The FS action sets DPNI_FS_OPT_SET_FLC |
DPNI_FS_OPT_SET_STASH_CONTROL with data stashing on every SoC.
dpaa2_dev_rx_queue_setup() skips FLC/stashing on SVR_LS2080A; the
FS path must follow the same rule.
Info
The QUEUE action index is checked against MAX_RX_QUEUES and
rx_vq[] != NULL, not against dev->data->nb_rx_queues, so an
unconfigured hardware queue is accepted.
----------------------------------------------------------------------
Patch 09/13 net/dpaa2: add minimum key size support
Warning
Removes "Current MC only support fixed entry size(56)" and starts
using 24-byte entries with no MC/DPNI version check. State the
minimum firmware in the commit message and gate it with
dpaa2_dev_cmp_dpni_ver(), or older MC rejects the table setup.
----------------------------------------------------------------------
Patch 10/13 net/dpaa2: restructure dpaa2 parser processing
Error
if (priv->psr_dynfield_offset >= 0)
dpaa2_dev_rx_parse_offset(priv, bufs[num_rx], fd);
psr_dynfield_offset is never assigned anywhere in the driver. With
zeroed dev_private the test is true for every packet, and
dpaa2_dev_rx_parse_offset() writes the l3/l4/l5 offsets at
(uint8_t *)mbuf + 0, i.e. into buf_addr. The RTE_ASSERT guarding it
is compiled out in release builds. Affects
dpaa2_dev_prefetch_rx_common(), dpaa2_dev_rx_common() and
dump_err_pkts().
Error
dpaa2_timestamp_dynfield() is removed, but dpaa2_dev_rx_common()
still calls it under #if defined(RTE_LIBRTE_IEEE1588). The build
breaks with IEEE1588 enabled.
Error
Rx timestamp handling is removed from dpaa2_dev_rx_parse() and
dpaa2_dev_rx_parse_new(). dpaa2_enable_ts[] is still set in
dpaa2_dev_configure() but never read, so
RTE_ETH_RX_OFFLOAD_TIMESTAMP silently stops working. The IEEE1588
priv->rx_timestamp update in dpaa2_dev_prefetch_rx_common() is also
gone.
Error
Checksum flag handling is removed from dpaa2_dev_rx_parse() and now
exists only in dpaa2_dev_rx_parse_slow(). On non-LX2160A SoCs the
fast-path return (word3 test false) no longer reports
RX_IP_CKSUM_* or RX_L4_CKSUM_*.
Error
The "if (!by_channel)" guard around dump_err_pkts() is removed from
dpaa2_dev_prefetch_rx_common(). The removed comment explains that
the channel/interrupt path owns the ethrx portal, and
dpaa2_dev_rx_common() still keeps the guard. This reverts that fix.
Error
The commit message says no functional change is intended. Also
changed:
- DPAA2_PKT_TYPE_IPV4_EXT 0x0001 -> 0x0040,
IPV6_EXT 0x0021 -> 0x0050
- VLAN_1/VLAN_2 0x0160/0x0260 -> 0x0100/0x0200
- new VLAN TCI reporting (RTE_MBUF_F_RX_VLAN)
- event and eqresp paths now handle SG frames
- IPv4/IPv6 EXT, SCTP and ICMP dropped from the LX2 fast switch
- parser result no longer printed for error-queue frames
- hash.sched.color now set from DPAA2_GET_FD_DROPP() on FS frames
Split into the header move plus separate functional patches, each
with its own justification.
Warning
dpaa2_dev_rx_print_parser_result() loops i < MAX_RX_QUEUES and
dereferences priv->rx_vq[i]->tc_index; entries past
priv->nb_rx_queues are NULL. The function is unreachable today
because DPAA2_RX_PRINT_PSR_RESULT_FLAG is never set, while
dpaa2_dev_init() still sets dpaa2_print_parser_result from getenv()
and nothing reads it. The parser debug print is broken either way.
Warning
Leftovers: rx_ts_offset, rx_ts_flag and
L3/L4/L5_OFFSET_OF_MBUF_DYN are unused. sp_protocol is never set,
so the fafe2/fafe3 GENEVE branch in dpaa2_dev_rx_parse_new() is
dead. dpaa2_dev_rx_parse_new() does RTE_SET_USED(priv) and then
uses priv.
----------------------------------------------------------------------
Patch 11/13 net/dpaa2: parse tunnel and fragmented packet types
Warning
The commit message says the new types are mapped "in the slow-path
parser". They are added only to the LX2160A FRC switch in
dpaa2_dev_rx_parse_new(); dpaa2_dev_rx_parse() for other SoCs is
unchanged.
Warning
DPAA2_PKT_TYPE_IPV4_FRAG and IPV6_FRAG evaluate to 0x0001 and
0x0021, the values the driver reported as RTE_PTYPE_L3_IPV4_EXT and
L3_IPV6_EXT before patch 10. Fragments were misreported as IP
options. That is a bug fix for released code and should be a
separate patch with Fixes: and Cc: [email protected].
----------------------------------------------------------------------
Patch 12/13 net/dpaa2: remove unused soft parser driver
Warning
dpaa2_soft_parser_loaded() is still defined in dpaa2_flow.c and
declared in dpaa2_ethdev.h with no callers, and it still opens
/dev/mem. Remove it together with dpaa2_sp_loaded.
Warning
The commit message is one sentence. Explain what users lose and why
runtime loading is unsupported.
----------------------------------------------------------------------
Patch 13/13 net/dpaa2: rename Rx queue flags
Info
The commit message refers to "the Tx confirmation rework that
follows", but this is the last patch in the series.
Info
The rename should come before patch 10, which adds
DPAA2_RX_PRINT_PSR_RESULT_FLAG next to the old DPAAX_ name.
