On Fri, Sep 08, 2023 at 04:42:33PM +0200, Phil Sutter wrote: > On Fri, Sep 08, 2023 at 04:01:02PM +0200, Pablo Neira Ayuso wrote: > > Hi Phil, > > > > On Fri, Sep 08, 2023 at 10:10:33AM +0200, Phil Sutter wrote: > > > The value in idx and the number of rules handled in that particular > > > __nf_tables_dump_rules() call is not identical. The former is a cursor > > > to pick up from if multiple netlink messages are needed, so its value is > > > ever increasing. > > > > My (buggy) intention was to display this audit log once per chain, at > > the end of the chain dump. > > Ah, I wasn't aware you did that on purpose. > > > > Fixing this is not just a matter of subtracting s_idx > > > from it, though: When resetting rules in multiple chains, > > > __nf_tables_dump_rules() is called for each and cb->args is not > > > adjusted in between. > > > > > > The audit notification in __nf_tables_dump_rules() had another problem: > > > If nf_tables_fill_rule_info() failed (e.g. due to buffer exhaustion), no > > > notification was sent - despite the rules having been reset already. > > > > Hm. that should not happen, when nf_tables_fill_rule_info() fails, > > that means buffer is full and userspace will invoke recvmsg() again. > > The next buffer resumes from the last entry that could not fit into > > the buffer. > > I didn't explicitly test for this case, but __nf_tables_dump_rules() > calls nf_tables_fill_rule_info() in a loop for reach rule. If it fails, > the function immediately returns 1 without calling > audit_log_rule_reset(). So while the last (failing) rule dump/reset will > be repeated after the detour to user space, the preceding rules > successfully dumped/reset slip through. I see, note that "failing" in this case means, "dump is still in progress" and "userspace will invoke recvmsg() again to resume from where we stopped. > > > To catch all the above and return to a single (if possible) notification > > > per table again, move audit logging back into the caller but into the > > > table loop instead of past it to avoid the potential null-pointer > > > deref. > > > > > > This requires to trigger the notification in two spots. Care has to be > > > taken in the second case as cb->args is also not updated in between > > > tables. This requires a helper variable as either it is the first table > > > (with potential non-zero cb->args cursor) or a consecutive one (with > > > idx holding the current cursor already). > > > > Your intention is to trigger one single audit log per table, right? > > Did you test a chain with a large ruleset that needs several buffers > > to be delivered to userspace in the netlink dump? > > Yes, see the last part in the proposed kselftest: Resetting rules in > a chain with 503 rules causes three notifications to be sent, for 189, > 188 and 126 rules each. > > > I would be inclined to do this once per-chain, so this can be extended > > later on to display the chain. Yes, that means this will send one > > audit log per chain, but this is where follow up updates will go? > > If you prefer that, no problem. I'll prepare a v3 then. If patch is smaller and we agree that chains are useful to be in the audit log (in follow up updates), then I'd suggest to go for a v3, yes. Thanks.