On 10/21/2026 11:59 AM, Bruce Richardson wrote:
On Tue, Oct 20, 2026 at 01:11:09PM +0100, Anatoly Burakov wrote:
Flow engines manage rule memory independently, with parsers returning
either temporary metadata or shared scratch storage. This spreads
ownership and cleanup across parsing, creation, destruction and
uninitialization across the flow infrastructure, which is difficult to
verify, reason about, and may have hidden bugs.
Make parsers allocate the final rule and transfer ownership to the
generic flow layer on success. Free rules centrally after validation,
failed creation, successful destruction and uninitialization, retaining
them when destruction fails. Parsers clean up their own allocations on
failure.
Some engines, in addition to the rule itself, will also allocate additional
data for the flow (e.g. RAW pattern buffer in FDIR engine). We can make the
flow memory management centralized in generic flow, but this additional
data is engine-specific, so use the `free` callbacks to release only said
additional flow data. Remove the RSS and ACL callbacks, the FDIR and ACL
scratch storage, and the redundant allocations and copies during rule
creation, and use stack storage where appropriate.
Signed-off-by: Anatoly Burakov <[email protected]>
Acked-by: Bruce Richardson <[email protected]>
After this cleanup, it looks like the flow parameter to each engine can be
removed completely, since all four engines mark it as __rte_unused now.
Yes, true. In ice, flows are basically two (or more) allocations - the
flow header itself, and the actual rule it encodes (with whatever nested
allocations the rule has - switch and FDIR RAW do that), but all of the
latter happens inside the `flow->rule` pointer, so the `flow` itself is
not needed. Arguably, `meta` also needs renaming into `rule`, because
that's what it does now: it is meant to be a pointer to `flow->rule`.
--
Thanks,
Anatoly
