Re: [PATCH nf 0/2] nft_set_pipapo: Fix crash due to dangling entries in mapping table

[Pablo Neira Ayuso <pablo@netfilter.org>]

[-- Attachment #1: Type: text/plain, Size: 2581 bytes --]

On Tue, Feb 25, 2020 at 09:58:47PM +0100, Pablo Neira Ayuso wrote:
> On Tue, Feb 25, 2020 at 09:38:15PM +0100, Stefano Brivio wrote:
> > On Tue, 25 Feb 2020 21:21:43 +0100
> > Pablo Neira Ayuso <pablo@netfilter.org> wrote:
> > 
> > > Hi Stefano,
> > > 
> > > On Tue, Feb 25, 2020 at 03:34:35PM +0100, Stefano Brivio wrote:
> > > [...]
> > > > This is the problem Phil reported:  
> > > [...]
> > > > Or also simply with:
> > > > 
> > > > # nft add element t s '{ 20-30 . 40 }'
> > > > # nft add element t s '{ 25-35 . 40 }'
> > > > 
> > > > the second element is silently ignored. I'm returning -EEXIST from
> > > > nft_pipapo_insert(), but nft_add_set_elem() clears it because NLM_F_EXCL
> > > > is not set.
> > > > 
> > > > Are you suggesting that this is consistent and therefore not a problem?  
> > > 
> > >                         NLM_F_EXCL      !NLM_F_EXCL
> > >         exact match       EEXIST             0 [*]
> > >         partial match     EEXIST           EEXIST
> > > 
> > > The [*] case would allow for element timeout/expiration updates from
> > > the control plane for exact matches.
> > 
> > A-ha. I didn't even consider that.
> > 
> > > Note that element updates are not
> > > supported yet, so this check for !NLM_F_EXCL is a stub. I don't think
> > > we should allow for updates on partial matches
> > > 
> > > I think what it is missing is a error to report "partial match" from
> > > pipapo. Then, the core translates this "partial match" error to EEXIST
> > > whether NLM_F_EXCL is set or not.
> > 
> > Yes, given what you explained, I also think it's the case.
> > 
> > > Would this work for you?
> > 
> > It would. I need to write a few more lines in nft_pipapo_insert(),
> > because right now I don't have a special case for "entirely
> > overlapping". Something on the lines of:
> > 
> > 	dup = pipapo_get(net, set, start, genmask);
> > 	if (PTR_ERR(dup) == -ENOENT) {
> > 
> > -->		compare start and end key for this entry with
> > 		start and end key from 'ext'
> > 
> > Let me know if you want me to post a patch with a placeholder for
> > whatever you have in mind, or if I can help implementing this, etc.
> 
> Please, go ahead with the placeholder, it might be faster. I'll jump
> on it.

Sorry, I just realized I can be probably quicker on the core side.
Here is my proposal.

I'm attaching a patch for the core. This is handling -ENOTEMPTY which
is (ab)used to report the partial element matching.

if NLM_F_EXCL is set off, then -EEXIST becomes 0.
                          then -ENOTEMPTY becomes -EEXIST.

Would this work for you?

Thanks.

[-- Attachment #2: x.patch --]
[-- Type: text/x-diff, Size: 1164 bytes --]

diff --git a/net/netfilter/nf_tables_api.c b/net/netfilter/nf_tables_api.c
index d1318bdf49ca..0bbe36e731b8 100644
--- a/net/netfilter/nf_tables_api.c
+++ b/net/netfilter/nf_tables_api.c
@@ -5059,7 +5059,8 @@ static int nft_add_set_elem(struct nft_ctx *ctx, struct nft_set *set,
 	ext->genmask = nft_genmask_cur(ctx->net) | NFT_SET_ELEM_BUSY_MASK;
 	err = set->ops->insert(ctx->net, set, &elem, &ext2);
 	if (err) {
-		if (err == -EEXIST) {
+		/* ENOTEMPTY reports partial element matching. */
+		if (err == -EEXIST || err == -ENOTEMPTY) {
 			if (nft_set_ext_exists(ext, NFT_SET_EXT_DATA) ^
 			    nft_set_ext_exists(ext2, NFT_SET_EXT_DATA) ||
 			    nft_set_ext_exists(ext, NFT_SET_EXT_OBJREF) ^
@@ -5075,8 +5076,12 @@ static int nft_add_set_elem(struct nft_ctx *ctx, struct nft_set *set,
 			     nft_set_ext_exists(ext2, NFT_SET_EXT_OBJREF) &&
 			     *nft_set_ext_obj(ext) != *nft_set_ext_obj(ext2)))
 				err = -EBUSY;
-			else if (!(nlmsg_flags & NLM_F_EXCL))
-				err = 0;
+			else if (!(nlmsg_flags & NLM_F_EXCL)) {
+				if (err == -EEXIST)
+					err = 0;
+				else if (err == -ENOTEMPTY)
+					err = -EEXIST;
+			}
 		}
 		goto err_element_clash;
 	}