Date: Wed, 3 Jan 2024 17:09:33 +0100
From: Jiri Pirko <jiri@...nulli.us>
To: Jamal Hadi Salim <jhs@...atatu.com>
Cc: Victor Nogueira <victor@...atatu.com>, davem@...emloft.net,
	edumazet@...gle.com, kuba@...nel.org, pabeni@...hat.com,
	xiyou.wangcong@...il.com, idosch@...sch.org, mleitner@...hat.com,
	vladbu@...dia.com, paulb@...dia.com, pctammela@...atatu.com,
	netdev@...r.kernel.org, kernel@...atatu.com,
	syzbot+84339b9e7330daae4d66@...kaller.appspotmail.com,
	syzbot+806b0572c8d06b66b234@...kaller.appspotmail.com,
	syzbot+0039110f932d438130f9@...kaller.appspotmail.com
Subject: Re: [PATCH net-next v2 1/1] net/sched: We should only add
 appropriate qdiscs blocks to ports' xarray

Wed, Jan 03, 2024 at 03:43:00PM CET, jhs@...atatu.com wrote:
>On Wed, Jan 3, 2024 at 9:26 AM Jiri Pirko <jiri@...nulli.us> wrote:
>>
>> Wed, Jan 03, 2024 at 03:09:14PM CET, jhs@...atatu.com wrote:
>> >On Wed, Jan 3, 2024 at 7:59 AM Jiri Pirko <jiri@...nulli.us> wrote:
>> >>
>> >> Tue, Jan 02, 2024 at 06:06:00PM CET, jhs@...atatu.com wrote:
>> >> >On Tue, Jan 2, 2024 at 10:54 AM Jiri Pirko <jiri@...nulli.us> wrote:
>> >> >>
>> >> >> Tue, Jan 02, 2024 at 03:52:01PM CET, jhs@...atatu.com wrote:
>> >> >> >On Tue, Jan 2, 2024 at 9:29 AM Jiri Pirko <jiri@...nulli.us> wrote:
>> >> >> >>
>> >> >> >> Tue, Jan 02, 2024 at 03:06:28PM CET, jhs@...atatu.com wrote:
>> >> >> >> >On Tue, Jan 2, 2024 at 4:59 AM Jiri Pirko <jiri@...nulli.us> wrote:
>> >> >> >> >>
>> >> >> >> >> The patch subject should briefly describe the nature of the change. Not
>> >> >> >> >> what "we" should or should not do.
>> >> >> >> >>
>> >> >> >> >>
>> >> >> >> >> Sun, Dec 31, 2023 at 06:23:20PM CET, victor@...atatu.com wrote:
>> >> >> >> >> >We should only add qdiscs to the blocks ports' xarray in ingress that
>> >> >> >> >> >support ingress_block_set/get or in egress that support
>> >> >> >> >> >egress_block_set/get.
>> >> >> >> >>
>> >> >> >> >> Tell the codebase what to do, be imperative. Please read again:
>> >> >> >> >> https://www.kernel.org/doc/html/v6.6/process/submitting-patches.html#describe-your-changes
>> >> >> >> >>
>> >> >> >> >
>> >> >> >> >We need another rule in the doc on nit-picking which states that we
>> >> >> >> >need to make progress at some point. We made many changes to this
>> >> >> >> >patchset based on your suggestions for no other reason other that we
>> >> >> >> >can progress the discussion. This is a patch that fixes a bug of which
>> >> >> >> >there are multiple syzbot reports and consumers of the API(last one
>> >> >> >> >just reported from the MTCP people). There's some sense of urgency to
>> >> >> >> >apply this patch before the original goes into net. More importantly:
>> >> >> >> >This patch fixes the issue and follows the same common check which was
>> >> >> >> >already being done in the committed patchset to check if the qdisc
>> >> >> >> >supports the block set/get operations.
>> >> >> >> >
>> >> >> >> >There are about 3 ways to do this check, you objected to the original,
>> >> >> >> >we picked something that works fine,  and now you are picking a
>> >> >> >> >different way with tcf_block. I dont see how tcf_block check would
>> >> >> >> >help or solve this problem at all given this is a qdisc issue not a
>> >> >> >> >class issue. What am I missing?
>> >> >> >>
>> >> >> >> Perhaps I got something wrong, but I thought that the issue is
>> >> >> >> cl_ops->tcf_block being null for some qdiscs, isn't it?
>> >> >> >>
>> >> >> >
>> >> >> >We attach these ports/netdevs only on capable qdiscs i.e ones that
>> >> >> >have  in/egress_block_set/get() - which happen to be ingress and
>> >> >> >clsact only.
>> >> >> >The problem was we were blindly assuming that presence of
>> >> >> >cl->tcf_block() implies presence of in/egress_block_set/get(). The
>> >> >> >earlier patches surrounded this code with attribute checks and so it
>> >> >> >worked there.
>> >> >>
>> >> >> Syskaller report says:
>> >> >>
>> >> >> KASAN: null-ptr-deref in range [0x0000000000000048-0x000000000000004f]
>> >> >> CPU: 1 PID: 5061 Comm: syz-executor323 Not tainted 6.7.0-rc6-syzkaller-01658-gc2b2ee36250d #0
>> >> >> Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS Google 11/17/2023
>> >> >> RIP: 0010:qdisc_block_add_dev net/sched/sch_api.c:1190 [inline]
>> >> >>
>> >> >> Line 1190 is:
>> >> >> block = cl_ops->tcf_block(sch, TC_H_MIN_INGRESS, NULL);
>> >> >>
>> >> >> So the cl_ops->tcf_block == NULL
>> >> >>
>> >> >> Why can't you just check it? Why do you want to check in/egress_block_set/get()
>> >> >> instead? I don't follow :/
>> >> >>
>> >> >
>> >> >Does it make sense to add to the port xarray just because we have
>> >> >cl_ops->tcf_block()? There are many qdiscs which have
>> >> >cl_ops->tcf_block() (example htb) but cant be used in the block add
>> >> >syntax (see question further below on tdc test).
>> >>
>> >> The whole block usage in qdiscs other than ingress and clsact seems odd
>> >> to me to be honest. What's the reason for that?.
>> >
>> >Well, you added that code so you tell me. Was the original idea to
>>
>> Well, I added it only for clsact and ingress. The rest is unrelated to
>> me.
>>
>
>Well git is blaming you..

Yeah, too long ago to remember I guess. It's no functional change,
just included new code into existing qdiscs, converting tcf_chain()
to tcf_block().


>
>> >allow grafting other qdiscs on a hierarchy? This is why i was asking
>> >for a sample use case to add to tdc.
>> >This was why our check is for "if (sch_ops->in/egress_block_get)"
>> >because the check for cl_ops->tcf_block() you suggested is not correct
>> >(it will match htb just fine for example) whereas this check will only
>> >catch cls_act and ingress.
>>
>> This code went off rails :/
>> The point is, mixing sch_ops->in/egress_block_get existence and cl_ops->tcf_block
>> looks awfully odd and inviting another bugs in the future.
>>
>
>What bugs? Be explicit so we can add tdc tests.
>
>> >> >--
>> >> >$sudo tc qdisc add dev lo egress_block 21 handle 1: root htb
>> >> >Error: Egress block sharing is not supported.
>> >> >---
>> >> >
>> >> >Did you look at the other syzbot reports?
>> >>
>> >> Yeah. The block usage in other qdiscs looks very odd.
>> >>
>> >
>> >And we have checks to catch it as you see.
>> >TBH, the idea of having cls_ops->tcf_block for a qdisc like htb is
>> >puzzling to me. It seems you are always creating a non-shared block
>> >for some but not all qdiscs regardless. What is that used for?
>>
>> No clue.
>>
>
>Well, if you cant remember a few years later then we'll look at
>removing it - it will require a lot more testing like i said.

It's actualy part of the conversion to introduce block processing. For
the existing other qdiscs, not functional change.


>
>> >
>> >>
>> >> >
>> >> >> Btw, the checks in __qdisc_destroy() do also look wrong.
>> >> >
>> >> >Now I am not following, please explain. The same code structure check
>> >> >is used in fill_qdisc
>> >> >(https://elixir.bootlin.com/linux/v6.7-rc8/source/net/sched/sch_api.c#L940)
>> >> >for example to pull the block info, is that wrong?
>> >>
>> >> There, you don't call tcf_block() at all, so how is that relevant?
>> >>
>> >
>> >Why do we need to call it? We just need it to retrieve the block id.
>>
>> Uff, that is my point. In the code you are pointing at, you don't use
>> tcf_block() at all, therefore it is not related to our discussion, is
>> it?
>>
>
>Huh? We are trying to check if it is legit to add a netdev to the
>xarray. The only way it is legit is if we have ingress or clsact.
>Those are the only two qdiscs with the set/get ops for blocks and the
>only potential ones which we can have valid shared blocks attached. It
>is related to the discussion.
>
>> >
>> >>
>> >>
>> >> >
>> >> >> >
>> >> >> >BTW: Do you have an example of a test case where we can test the class
>> >> >> >grafting (eg using htb with tcf_block)? It doesnt have any impact on
>> >> >> >this patcheset here but we want to add it as a regression checker on
>> >> >> >tdc in the future if someone makes a change.
>> >> >
>> >> >An answer to this will help.
>> >>
>> >> Answer is "no".
>> >
>> >Ok, so we cant test this or this is internal use only?
>> >
>> >I am going to repeat again here: you are holding back a bug fix (with
>> >many reports) with this discussion. We can have the discussion
>> >separately but let's make quick progress. If need be we can send fixes
>> >after.
>>
>> I don't mind. Code is a mess as it is already. One more crap won't
>> hurt...
>>
>
>Ok, so your code that you added a few years ago is crap then by that

True that.


>metric. You can't even remember why you added it.
>You have provided no legit argument for the approach you are
>suggesting and i see nothing wrong with what we did. Let's agree to
>disagree in order to make progress and get the bug resolved.
>Please ACK the patch and we can discuss if we need to remove the class
>ops separately.

I looked a bit deeper into the code, looks like the xarray insertion and
removal is done in inconvenient place, causing this bug. Moving it into
tcf_block_get/put_ext() seems much more suiting. See following untested
patch. It should:
1) fix the bug
2) make things nicer and easier to read

diff --git a/net/sched/cls_api.c b/net/sched/cls_api.c
index adf5de1ff773..253b26f2eddd 100644
--- a/net/sched/cls_api.c
+++ b/net/sched/cls_api.c
@@ -1428,6 +1428,7 @@ int tcf_block_get_ext(struct tcf_block **p_block, struct Qdisc *q,
 		      struct tcf_block_ext_info *ei,
 		      struct netlink_ext_ack *extack)
 {
+	struct net_device *dev = qdisc_dev(q);
 	struct net *net = qdisc_net(q);
 	struct tcf_block *block = NULL;
 	int err;
@@ -1461,9 +1462,18 @@ int tcf_block_get_ext(struct tcf_block **p_block, struct Qdisc *q,
 	if (err)
 		goto err_block_offload_bind;
 
+	if (tcf_block_shared(block)) {
+		err = xa_insert(&block->ports, dev->ifindex, dev, GFP_KERNEL);
+		if (err) {
+			NL_SET_ERR_MSG(extack, "block dev insert failed");
+			goto err_dev_insert;
+		}
+	}
+
 	*p_block = block;
 	return 0;
 
+err_dev_insert:
 err_block_offload_bind:
 	tcf_chain0_head_change_cb_del(block, ei);
 err_chain0_head_change_cb_add:
@@ -1502,8 +1512,12 @@ EXPORT_SYMBOL(tcf_block_get);
 void tcf_block_put_ext(struct tcf_block *block, struct Qdisc *q,
 		       struct tcf_block_ext_info *ei)
 {
+	struct net_device *dev = qdisc_dev(q);
+
 	if (!block)
 		return;
+	if (tcf_block_shared(block))
+		xa_erase(&block->ports, dev->ifindex);
 	tcf_chain0_head_change_cb_del(block, ei);
 	tcf_block_owner_del(block, q, ei->binder_type);
 
diff --git a/net/sched/sch_api.c b/net/sched/sch_api.c
index 299086bb6205..e9eaf637220e 100644
--- a/net/sched/sch_api.c
+++ b/net/sched/sch_api.c
@@ -1180,43 +1180,6 @@ static int qdisc_graft(struct net_device *dev, struct Qdisc *parent,
 	return 0;
 }
 
-static int qdisc_block_add_dev(struct Qdisc *sch, struct net_device *dev,
-			       struct netlink_ext_ack *extack)
-{
-	const struct Qdisc_class_ops *cl_ops = sch->ops->cl_ops;
-	struct tcf_block *block;
-	int err;
-
-	block = cl_ops->tcf_block(sch, TC_H_MIN_INGRESS, NULL);
-	if (block) {
-		err = xa_insert(&block->ports, dev->ifindex, dev, GFP_KERNEL);
-		if (err) {
-			NL_SET_ERR_MSG(extack,
-				       "ingress block dev insert failed");
-			return err;
-		}
-	}
-
-	block = cl_ops->tcf_block(sch, TC_H_MIN_EGRESS, NULL);
-	if (block) {
-		err = xa_insert(&block->ports, dev->ifindex, dev, GFP_KERNEL);
-		if (err) {
-			NL_SET_ERR_MSG(extack,
-				       "Egress block dev insert failed");
-			goto err_out;
-		}
-	}
-
-	return 0;
-
-err_out:
-	block = cl_ops->tcf_block(sch, TC_H_MIN_INGRESS, NULL);
-	if (block)
-		xa_erase(&block->ports, dev->ifindex);
-
-	return err;
-}
-
 static int qdisc_block_indexes_set(struct Qdisc *sch, struct nlattr **tca,
 				   struct netlink_ext_ack *extack)
 {
@@ -1387,10 +1350,6 @@ static struct Qdisc *qdisc_create(struct net_device *dev,
 	qdisc_hash_add(sch, false);
 	trace_qdisc_create(ops, dev, parent);
 
-	err = qdisc_block_add_dev(sch, dev, extack);
-	if (err)
-		goto err_out4;
-
 	return sch;
 
 err_out4:
diff --git a/net/sched/sch_generic.c b/net/sched/sch_generic.c
index e33568df97a5..9b3e9262040b 100644
--- a/net/sched/sch_generic.c
+++ b/net/sched/sch_generic.c
@@ -1052,8 +1052,6 @@ static void __qdisc_destroy(struct Qdisc *qdisc)
 {
 	const struct Qdisc_ops  *ops = qdisc->ops;
 	struct net_device *dev = qdisc_dev(qdisc);
-	const struct Qdisc_class_ops *cops;
-	struct tcf_block *block;
 
 #ifdef CONFIG_NET_SCHED
 	qdisc_hash_del(qdisc);
@@ -1064,18 +1062,6 @@ static void __qdisc_destroy(struct Qdisc *qdisc)
 
 	qdisc_reset(qdisc);
 
-	cops = ops->cl_ops;
-	if (ops->ingress_block_get) {
-		block = cops->tcf_block(qdisc, TC_H_MIN_INGRESS, NULL);
-		if (block)
-			xa_erase(&block->ports, dev->ifindex);
-	}
-
-	if (ops->egress_block_get) {
-		block = cops->tcf_block(qdisc, TC_H_MIN_EGRESS, NULL);
-		if (block)
-			xa_erase(&block->ports, dev->ifindex);
-	}
 
 	if (ops->destroy)
 		ops->destroy(qdisc);
diff --git a/net/sched/sch_ingress.c b/net/sched/sch_ingress.c
index 5fa9eaa79bfc..1770083052cd 100644
--- a/net/sched/sch_ingress.c
+++ b/net/sched/sch_ingress.c
@@ -284,7 +284,8 @@ static int clsact_init(struct Qdisc *sch, struct nlattr *opt,
 	q->egress_block_info.chain_head_change = clsact_chain_head_change;
 	q->egress_block_info.chain_head_change_priv = &q->miniqp_egress;
 
-	return tcf_block_get_ext(&q->egress_block, sch, &q->egress_block_info, extack);
+	return tcf_block_get_ext(&q->egress_block, sch, &q->egress_block_info,
+				 extack);
 }
 
 static void clsact_destroy(struct Qdisc *sch)


If you don't mind, I would test this and send it instead of your fix. If
you can test it too, that would be awesome.

Thanks!

Powered by blists - more mailing lists