Re: [PATCH 3/3] sctp: fix association hangs due to reassembly/orderinglogic

From: Vlad Yasevich
Date: Wed Feb 20 2013 - 15:14:06 EST

On 02/20/2013 02:24 PM, Roberts, Lee A. wrote:
Vlad,

-----Original Message-----
From: Vlad Yasevich [mailto:vyasevich@xxxxxxxxx]
Sent: Wednesday, February 20, 2013 11:06 AM
To: Roberts, Lee A.
Cc: linux-sctp@xxxxxxxxxxxxxxx; netdev@xxxxxxxxxxxxxxx; linux-
kernel@xxxxxxxxxxxxxxx
Subject: Re: [PATCH 3/3] sctp: fix association hangs due to
reassembly/ordering logic

Hi Lee

On 02/20/2013 10:56 AM, Roberts, Lee A. wrote:
From: Lee A. Roberts <lee.roberts@xxxxxx>

Resolve SCTP association hangs observed during SCTP stress
testing. Observable symptoms include communications hangs
with data being held in the association reassembly and/or lobby
(ordering) queues. Close examination of reassembly queue shows
missing packets.

As a general note for this patch series, you could really benefit
from a cover letter that describes the symptoms and all the different
issues you found.

Also, please title your patches based on the context of the patch.
Giving them all the same title is very confusing and at quick glance
makes appear that the same patch was applied 3 times.


In sctp_eat_data(), enter partial delivery mode only if the
data on the head of the reassembly queue is at or before the
cumulative TSN ACK point.

In sctp_ulpq_retrieve_partial() and sctp_ulpq_retrieve_first(),
correct message reassembly logic for SCTP partial delivery.
Change logic to ensure that as much data as possible is sent
with the initial partial delivery and that following partial
deliveries contain all available data.

In sctp_ulpq_renege(), adjust logic to enter partial delivery
only if the incoming chunk remains on the reassembly queue
after processing by sctp_ulpq_tail_data(). Remove call to
sctp_tsnmap_mark(), as this is handled correctly in call to
sctp_ulpq_tail_data().

Patch applies to linux-3.8 kernel.

Signed-off-by: Lee A. Roberts <lee.roberts@xxxxxx>
---
net/sctp/sm_statefuns.c | 12 ++++++++++--
net/sctp/ulpqueue.c | 33 ++++++++++++++++++++++++++-------
2 files changed, 36 insertions(+), 9 deletions(-)

diff -uprN -X linux-3.8-vanilla/Documentation/dontdiff linux-3.8-SCTP
+2/net/sctp/sm_statefuns.c linux-3.8-SCTP+3/net/sctp/sm_statefuns.c
--- linux-3.8-SCTP+2/net/sctp/sm_statefuns.c 2013-02-18
16:58:34.000000000 -0700
+++ linux-3.8-SCTP+3/net/sctp/sm_statefuns.c 2013-02-20
08:31:51.092132884 -0700
@@ -6090,7 +6090,8 @@ static int sctp_eat_data(const struct sc
size_t datalen;
sctp_verb_t deliver;
int tmp;
- __u32 tsn;
+ __u32 tsn, ctsn;
+ struct sk_buff *skb;
struct sctp_tsnmap *map = (struct sctp_tsnmap *)&asoc-
peer.tsn_map;
struct sock *sk = asoc->base.sk;
struct net *net = sock_net(sk);
@@ -6160,7 +6161,14 @@ static int sctp_eat_data(const struct sc
/* Even if we don't accept this chunk there is
* memory pressure.
*/
- sctp_add_cmd_sf(commands, SCTP_CMD_PART_DELIVER,
SCTP_NULL());
+ skb = skb_peek(&asoc->ulpq.reasm);
+ if (skb != NULL) {
+ ctsn = sctp_skb2event(skb)->tsn;
+ if (TSN_lte(ctsn,
+ sctp_tsnmap_get_ctsn(&asoc->peer.tsn_map)))
+ sctp_add_cmd_sf(commands,
+ SCTP_CMD_PART_DELIVER, SCTP_NULL());
+ }

What if the tsn you are currently processing will advance the
CUM TSN? You may still need to eneter PD. This is why
sctp_eat_data() is not the right place to place this check.

The more I look at the SCTP_CMD_PART_DELIVERY the more I am thinking
that it has to come after the current TSN has been delivered to the
queue. That way, if we are currently processing the first fragment,
we'll queue it, then enter PD and pull it off the queue properly.
If we are processing the middle or last, it will get queued first, and
then PD will be entered to fetch anything that we can fetch.

In fact, the above is what happens when we issue a RENEGE command, but
the order is reversed is RENEGE is skipped for some reason. I am still
trying to figure out if it's possible to enter PD without RENEGE, but I
did notice the above aberration.


The current code enters partial delivery by calling sctp_ulpq_partial_delivery()
in two locations. In sctp_cmd_interpreter() [located in ../net/sctp/sm_sideeffect.c],
as a result of the code in sctp_eat_data() [located in ../net/sctp/sctp_statefuns.c]:

1675 case SCTP_CMD_PART_DELIVER:
1676 sctp_ulpq_partial_delivery(&asoc->ulpq, GFP_ATOMIC) ;
1677 break;

and in sctp_ulpq_renege() [located in ../net/sctp/ulpqueue.c]


1055 /* If able to free enough room, accept this chunk. */
1056 if (chunk && (freed >= needed)) {
1057 __u32 tsn;
1058 tsn = ntohl(chunk->subh.data_hdr->tsn);
1059 sctp_tsnmap_mark(&asoc->peer.tsn_map, tsn, chunk->transport );
1060 sctp_ulpq_tail_data(ulpq, chunk, gfp);
1061
1062 sctp_ulpq_partial_delivery(ulpq, gfp);
1063 }

Neither location checks to see whether the tsn on the head of the reassembly queue
is less than or equal to the cumulative TSN ACK point. Perhaps a better solution
is to put this check in the beginning of sctp_ulpq_partial_delivery() and abort the
partial delivery attempt if this condition isn't met. I'll try this.

I don't really think you need do any of that. What you needs to happen is that sctp_ulpq_retrieve_first() and sctp_ulpq_retrieve_partial() need code in them to make sure we don't use a fragment that has
TSN > CUM_TSN.

PD is not really entered if sctp_ulpq_retrieve_first() returns NULL so we are ok.


The code in sctp_eat_data() involving SCTP_CMD_PART_DELIVER may be better if invoked
after the current packet is handled, but that is probably an optimization. As is,
the partial delivery may not start until another packet arrives.


What I was saying is that have the possibilities of the following 2 sequences of commands:

SCTP_CMD_PART_DELIVER - calls sctp_ulpq_partial_delivery
SCTP_CMD_RENEGE - calls sctp_ulpq_tail_data followed by sctp_ulpq_partial_delivery

or
SCTP_CMD_PART_DELIVER - calls sctp_ulpq_partial_delivery
SCTP_CMD_CHUNK_ULP - calls sctp_ulpq_tail_data


What I am saying is that in the first sequence, we call sctp_ulpq_partial_delivery twice which is more then useless. It is actually wrong. We may Partial Deliver first and may not even need to reneg at all after that.

In the second case, we have the order reversed. We should put the data of the uplq first and then if needed attempt PD since now we have a higher chance of it succeeding.


}

/* Spill over rwnd a little bit. Note: While allowed, this spill
over
diff -uprN -X linux-3.8-vanilla/Documentation/dontdiff linux-3.8-SCTP
+2/net/sctp/ulpqueue.c linux-3.8-SCTP+3/net/sctp/ulpqueue.c
--- linux-3.8-SCTP+2/net/sctp/ulpqueue.c 2013-02-20
08:17:53.679233365
-0700
+++ linux-3.8-SCTP+3/net/sctp/ulpqueue.c 2013-02-20
08:27:02.785042744
-0700
@@ -540,14 +540,19 @@ static struct sctp_ulpevent *sctp_ulpq_r
ctsn = cevent->tsn;

switch (cevent->msg_flags & SCTP_DATA_FRAG_MASK) {
+ case SCTP_DATA_FIRST_FRAG:
+ if (!first_frag)
+ return NULL;
+ goto done;
case SCTP_DATA_MIDDLE_FRAG:
if (!first_frag) {
first_frag = pos;
next_tsn = ctsn + 1;
last_frag = pos;
- } else if (next_tsn == ctsn)
+ } else if (next_tsn == ctsn) {
next_tsn++;
- else
+ last_frag = pos;
+ } else
goto done;
break;
case SCTP_DATA_LAST_FRAG:

This may still allow you to skip over a gap if the first middle
fragment
in the queue starts after the gap. We need to make sure that
TSN of the current chunk is less then equal to
sctp_tsnmap_get_ctsn(map).



594 if (!ulpq->pd_mode)
595 retval = sctp_ulpq_retrieve_reassembled(ulpq);
596 else {
597 __u32 ctsn, ctsnap;
598
599 /* Do not even bother unless this is the next tsn to
600 * be delivered.
601 */
602 ctsn = event->tsn;
603 ctsnap = sctp_tsnmap_get_ctsn(&ulpq->asoc->peer.tsn_map);
604 if (TSN_lte(ctsn, ctsnap))
605 retval = sctp_ulpq_retrieve_partial(ulpq);
606 }

This is the only place we call sctp_ulpq_retrieve_partial().

Ok, you are right. We are seem to be ok here, but I am not really crazy about doing the checks ouside these functions. the functions themselves should be smart enough to return the right data.


@@ -651,6 +656,14 @@ static struct sctp_ulpevent *sctp_ulpq_r
} else
goto done;
break;
+
+ case SCTP_DATA_LAST_FRAG:
+ if (!first_frag)
+ return NULL;
+ else
+ goto done;
+ break;
+
default:
return NULL;

Same thing here. Both, FIRST and MIDDLE fragments need to be validated
against sctp_tsnmap_get_ctsn(), otherwise you may be stepping over a
gap.


If we check that the head of the reassembly queue is at or below the cumulative tsn
before entering partial delivery, we should be OK.

But you have to do it everywhere you try to enter PD instead of just in once place that actually iterates and checks tsns.


In sctp_ulpq_retrieve_first(), we expect the first packet on the reassembly queue
will be a FIRST. (If not, something went wrong somewhere else.)

Not necessarily wrong. It may have just been still lost on the network. It is perfectly ok that we may still be waiting for this FIRST fragment.

Also, what if the FIRST fragment you have here is bigger then cum_tsn_ack_point? I guess this is why you check before calling into partial delivery functions, but you call before fully processing the TSN sometimes. It is more correct to have these functions do the validation and return proper results that have the callers do validation. It is simpler to miss call sights and you duplicate code.


If we find another
FIRST, we know that we've gone too far. If the second packet is a MIDDLE, we check
that it has the correct tsn. We keep processing MIDDLE packets as long as the tsn
values are sequential. If the packet is a MIDDLE and the tsn isn't right, we've
gone too far. In sctp_ulpq_retreive_first(), we don't return a LAST, by definition,
since that wouldn't be a partial delivery.

In sctp_ulpq_retrieve_partial(), we expect that the first packet on the reassembly queue
is not a FIRST. We're looking for MIDDLE or LAST fragments to complete the message.
Since we know that the head of the reassembly queue is at or before the cumulative TSN,
we shouldn't have a gap. We want to pick up MIDDLE packets until we see a tsn gap
or until we find the right LAST packet. When we find the right LAST packet, we set
MSG_EOR and exit from partial delivery.

}
@@ -1054,6 +1067,7 @@ void sctp_ulpq_renege(struct sctp_ulpq *
gfp_t gfp)
{
struct sctp_association *asoc;
+ struct sk_buff *skb;
__u16 needed, freed;

asoc = ulpq->asoc;
@@ -1074,12 +1088,17 @@ void sctp_ulpq_renege(struct sctp_ulpq *
}
/* If able to free enough room, accept this chunk. */
if (chunk && (freed >= needed)) {
- __u32 tsn;
+ __u32 tsn, ctsn;
tsn = ntohl(chunk->subh.data_hdr->tsn);
- sctp_tsnmap_mark(&asoc->peer.tsn_map, tsn, chunk-
transport);
- sctp_ulpq_tail_data(ulpq, chunk, gfp);
-
- sctp_ulpq_partial_delivery(ulpq, gfp);
+ if (sctp_ulpq_tail_data(ulpq, chunk, gfp) == 0) {
+ skb = skb_peek(&ulpq->reasm);
+ if (skb != NULL) {
+ ctsn = sctp_skb2event(skb)->tsn;
+ if (TSN_lte(ctsn, tsn))
+ sctp_ulpq_partial_delivery(ulpq, chunk,
+ gfp);
+ }
+ }
}

I am not sure this hunk is really needed.

You are trying to use this code make sure that you start PD with
something to deliver, but the PD code already takes care of that.
You also get some basic cum_tsn checking for the current chunk because
of how renege is called, but you still don't do the checks for
subsequent chunks in the queue as I stated above, so you are still
subject to possible hangs.

-vlad


If our incoming packet completes a message, the call to sctp_ulpq_tail_data()
should cause it to be delivered. If so, the whole message is gone, so we shouldn't
need to enter partial delivery. We don't want to enter partial delivery
unnecessarily, since it can block delivery on other streams.

OK. Then it might be nice to have some status from sctp_ulpq_tail_data() of whether EOR has been reached or not. If not EOR , then do partial deliver. If EOR, then you may want to consider calling sctp_ulpq_reasm_drain() to see you can pull get more fully reassembled events.

-vlad




sk_mem_reclaim(asoc->base.sk);

NïïïïïrïïyïïïbïXïïÇvï^ï)Þ{.nï+ïïïï{ïïïiï{ayïÊÚï,j
ïïfïïïhïïïzïïwïïï

ïïïj:+vïïïwïjïmïïïï
ïïïïzZ+ïïÝj"ïï!tml=



--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/