On Wed, 16 Sep 2026 21:11:02 +0800 Zhang Tengfei <[email protected]> wrote:
> v2: > - split L2 tunnel add-failure goto out into its own patch > - allocate FDIR object before installing the global mask > - add Fixes tag for the VF FDIR path > - set -EINVAL on remaining FDIR goto out paths with bad ret > - use struct assignment instead of rte_memcpy for filter copies > > 1/3 fix L2 tunnel error on flow create > 2/3 fix leak of filters on flow create > 3/3 fix flow create error codes > > Zhang Tengfei (3): > net/txgbe: fix L2 tunnel error on flow create > net/txgbe: fix leak of filters on flow create > net/txgbe: fix flow create error codes > > drivers/net/txgbe/txgbe_flow.c | 240 +++++++++++++++++---------------- > 1 file changed, 126 insertions(+), 114 deletions(-) > Looks good still some small items found by AI review: Review: [PATCH v2 0/3] net/txgbe: flow create fixes Author: Zhang Tengfei <[email protected]> Series applies cleanly to main; each commit builds with -Dwerror=true. Fixes: tags resolve (5c2352b9ece6 in 21.02, 7eef71080e16 in 25.11), so Cc: stable is correct. Patch 2/3 checked for early copies: ntuple, ethertype, SYN and L2 tunnel add helpers do not modify their input, so copying filter_info before programming is safe. FDIR (PF and VF) copies after programming. Element types match the old rte_memcpy sizes, struct assignment is equivalent. Patch 3/3: net/txgbe: fix flow create error codes Warning: mask-only FDIR rule fails but leaves global mask committed With b_mask set and b_spec clear on the first FDIR rule, txgbe_flow_create() programs the input mask via txgbe_fdir_set_input_mask(), sets fdir_info->mask_added = TRUE, then falls to the final path which now returns -EINVAL. The application gets a failed create, holds no handle, yet the global mask stays in hardware and mask_added stays set. Any later rule with a different mask is rejected with "only support one global mask" even though no FDIR flow exists. The b_spec failure path already clears mask_added when first_mask is set; this path does not. The parser reaches this state: txgbe_parse_fdir_filter_normal() sets b_mask on item->mask and b_spec only on item->spec. Since this patch now declares the path an error, reject it before touching hardware, e.g. right after the rte_zmalloc(): if (!fdir_rule.b_spec) { rte_free(fdir_rule_ptr); ret = -EINVAL; goto out; } and drop the trailing free/-EINVAL block. Info: commit message says memcmp stores a "positive result". memcmp() returns any nonzero value of either sign, so -ret could be a random positive or negative number. Reword to say the value is not an errno. Info: flex offset mismatch path returns -EINVAL with no log, unlike the mask mismatch path right above it. Add a PMD_DRV_LOG(ERR, ...) so the two rejections are distinguishable.
