Re: [nft PATCH 3/5] mnl: Allow for updating devices on existing inet ingress hook chains

[Phil Sutter <phil@nwl.cc>]

On Mon, Sep 08, 2025 at 01:01:00PM +0200, Pablo Neira Ayuso wrote:
> On Mon, Sep 08, 2025 at 12:29:03PM +0200, Pablo Neira Ayuso wrote:
> > On Fri, Aug 29, 2025 at 04:25:11PM +0200, Phil Sutter wrote:
> > > Complete commit a66b5ad9540dd ("src: allow for updating devices on
> > > existing netdev chain") in supporting inet family ingress hook chains as
> > > well. The kernel does already but nft has to add a proper hooknum
> > > attribute to pass the checks.
> > > 
> > > The hook.num field has to be initialized from hook.name using
> > > str2hooknum(), which is part of chain evaluation. Calling
> > > chain_evaluate() just for that purpose is a bit over the top, but the
> > > hook name lookup may fail and performing chain evaluation for delete
> > > command as well fits more into the code layout than duplicating parts of
> > > it in mnl_nft_chain_del() or elsewhere. Just avoid the
> > > chain_cache_find() call as its assert() triggers when deleting by
> > > handle and also don't add to be deleted chains to cache.
> > > 
> > > Signed-off-by: Phil Sutter <phil@nwl.cc>
> > > ---
> > >  src/evaluate.c | 6 ++++--
> > >  src/mnl.c      | 2 ++
> > >  2 files changed, 6 insertions(+), 2 deletions(-)
> > > 
> > > diff --git a/src/evaluate.c b/src/evaluate.c
> > > index b7e4f71fdfbc9..db4ac18f1dc9f 100644
> > > --- a/src/evaluate.c
> > > +++ b/src/evaluate.c
> > > @@ -5758,7 +5758,9 @@ static int chain_evaluate(struct eval_ctx *ctx, struct chain *chain)
> > >  		return table_not_found(ctx);
> > >  
> > >  	if (chain == NULL) {
> > > -		if (!chain_cache_find(table, ctx->cmd->handle.chain.name)) {
> > > +		if (ctx->cmd->op != CMD_DELETE &&
> > > +		    ctx->cmd->op != CMD_DESTROY &&
> > > +		    !chain_cache_find(table, ctx->cmd->handle.chain.name)) {
> > >  			chain = chain_alloc();
> > >  			handle_merge(&chain->handle, &ctx->cmd->handle);
> > >  			chain_cache_add(chain, table);
> > > @@ -6070,7 +6072,7 @@ static int cmd_evaluate_delete(struct eval_ctx *ctx, struct cmd *cmd)
> > >  		return 0;
> > >  	case CMD_OBJ_CHAIN:
> > >  		chain_del_cache(ctx, cmd);
> > > -		return 0;
> > > +		return chain_evaluate(ctx, cmd->chain);
> > 
> > Maybe fix this to perform chain_del_cache() after chain_evaluate()?

I agree, side-effects of reusing chain_evaluate() for deletion are not
worth it.

> > ie.
> > 
> >                 if (chain_evaluate(ctx, cmd->chain) < 0)
> >                         return -1;
> > 
> >                 chain_del_cache(ctx, cmd);
> 
> My suggestion won't work.
> 
> Maybe add a specific chain_del_evaluate(), see untested patch attached.

Since we only need a proper value in chain->hook.num, a more minimal
version is fine:

diff --git a/src/evaluate.c b/src/evaluate.c
index b7e4f71fdfbc9..8cecbe09de01c 100644
--- a/src/evaluate.c
+++ b/src/evaluate.c
@@ -5992,6 +5992,22 @@ static void chain_del_cache(struct eval_ctx *ctx, struct cmd *cmd)
        chain_free(chain);
 }
 
+static int chain_del_evaluate(struct eval_ctx *ctx, struct cmd *cmd)
+{
+       struct chain *chain = cmd->chain;
+
+       if (chain && chain->flags & CHAIN_F_BASECHAIN && chain->hook.name) {
+               chain->hook.num = str2hooknum(chain->handle.family,
+                                             chain->hook.name);
+               if (chain->hook.num == NF_INET_NUMHOOKS)
+                       return __stmt_binary_error(ctx, &chain->hook.loc, NULL,
+                                                  "The %s family does not support this hook",
+                                                  family2str(chain->handle.family));
+       }
+       chain_del_cache(ctx, cmd);
+       return 0;
+}
+
 static void set_del_cache(struct eval_ctx *ctx, struct cmd *cmd)
 {
        struct table *table;
@@ -6069,8 +6085,7 @@ static int cmd_evaluate_delete(struct eval_ctx *ctx, struct cmd *cmd)
        case CMD_OBJ_RULE:
                return 0;
        case CMD_OBJ_CHAIN:
-               chain_del_cache(ctx, cmd);
-               return 0;
+               return chain_del_evaluate(ctx, cmd);
        case CMD_OBJ_TABLE:
                table_del_cache(ctx, cmd);
                return 0;

Fine with you?

Cheers, Phil