On Fri, Dec 15, 2023 at 10:53:44PM +0100, Phil Sutter wrote: > This is the former RFC turned into a complete implementation including > Florian's suggested improvements. > > Patch 1 is fallout, took me a while debugging the segfaulting test case > until I noticed it wasn't my fault! :) > > Patch 2 is the same as in the RFC. > > Patch 3 separates the type value checking from patch 2 and drops > expressions' default switch cases where all possible values are handled. > > Patch 4 is prep work for patch 5. > > Patch 5 adds the new struct expr_ops field and defines policies for all > expressions. > > Patch 6 then enables policy checking. > > Some remarks for consideration: > > * This adds kernel-internal knowledge to libnftnl, namely in max name > lengths. Maybe not ideal, but I found it more sensible than Florian's > suggested alternative of using 65528 to just not exceed netlink > limits. > > * nftnl_expr_set_u*() setters start failing when they would happily > overstep boundaries before. This is intentional, but getting the > policy values right (at first I thought 'sizeof(enum nft_registers)' > was a good idea) showed how hard to diagnose bugs in that area are. I > think we should make the setters return success/fail like > nftnl_expr_set_str does already, even if that breaks ABI (does it?). > nftables probably benefits from setter wrappers which call > netlink_abi_error() if the setter fails. > > Phil Sutter (6): > tests: Fix objref test case > expr: Repurpose struct expr_ops::max_attr field > expr: Call expr_ops::set with legal types only > include: Sync nf_log.h with kernel headers > expr: Introduce struct expr_ops::attr_policy > expr: Enforce attr_policy compliance in nftnl_expr_set() Series applied after checking theres no effect on nftables' py testsuite results in different releases (v0.9.9, v1.0.6, v1.0.9).
