On Wed, Nov 23, 2022 at 12:20:44PM -0500, Xin Long wrote:
> On Tue, Nov 22, 2022 at 6:35 PM Marcelo Ricardo Leitner
> <marcelo.leitner@xxxxxxxxx> wrote:
> >
> > On Fri, Nov 18, 2022 at 10:15:50PM -0500, Xin Long wrote:
> > > On Fri, Nov 18, 2022 at 3:48 AM Zhengchao Shao <shaozhengchao@xxxxxxxxxx> wrote:
> > > >
> > > > When sctp_stream_outq_migrate() is called to release stream out resources,
> > > > the memory pointed to by prio_head in stream out is not released.
> > > >
> > > > The memory leak information is as follows:
> > > > unreferenced object 0xffff88801fe79f80 (size 64):
> > > > comm "sctp_repo", pid 7957, jiffies 4294951704 (age 36.480s)
> > > > hex dump (first 32 bytes):
> > > > 80 9f e7 1f 80 88 ff ff 80 9f e7 1f 80 88 ff ff ................
> > > > 90 9f e7 1f 80 88 ff ff 90 9f e7 1f 80 88 ff ff ................
> > > > backtrace:
> > > > [<ffffffff81b215c6>] kmalloc_trace+0x26/0x60
> > > > [<ffffffff88ae517c>] sctp_sched_prio_set+0x4cc/0x770
> > > > [<ffffffff88ad64f2>] sctp_stream_init_ext+0xd2/0x1b0
> > > > [<ffffffff88aa2604>] sctp_sendmsg_to_asoc+0x1614/0x1a30
> > > > [<ffffffff88ab7ff1>] sctp_sendmsg+0xda1/0x1ef0
> > > > [<ffffffff87f765ed>] inet_sendmsg+0x9d/0xe0
> > > > [<ffffffff8754b5b3>] sock_sendmsg+0xd3/0x120
> > > > [<ffffffff8755446a>] __sys_sendto+0x23a/0x340
> > > > [<ffffffff87554651>] __x64_sys_sendto+0xe1/0x1b0
> > > > [<ffffffff89978b49>] do_syscall_64+0x39/0xb0
> > > > [<ffffffff89a0008b>] entry_SYSCALL_64_after_hwframe+0x63/0xcd
> > > >
> > > > Fixes: 637784ade221 ("sctp: introduce priority based stream scheduler")
> > > > Reported-by: syzbot+29c402e56c4760763cc0@xxxxxxxxxxxxxxxxxxxxxxxxx
> > > > Signed-off-by: Zhengchao Shao <shaozhengchao@xxxxxxxxxx>
> > > > ---
> > > > net/sctp/stream.c | 6 ++++++
> > > > 1 file changed, 6 insertions(+)
> > > >
> > > > diff --git a/net/sctp/stream.c b/net/sctp/stream.c
> > > > index ef9fceadef8d..a17dc368876f 100644
> > > > --- a/net/sctp/stream.c
> > > > +++ b/net/sctp/stream.c
> > > > @@ -70,6 +70,9 @@ static void sctp_stream_outq_migrate(struct sctp_stream *stream,
> > > > * sctp_stream_update will swap ->out pointers.
> > > > */
> > > > for (i = 0; i < outcnt; i++) {
> > > > + if (SCTP_SO(new, i)->ext)
> > > > + kfree(SCTP_SO(new, i)->ext->prio_head);
> > > > +
> > > > kfree(SCTP_SO(new, i)->ext);
> > > > SCTP_SO(new, i)->ext = SCTP_SO(stream, i)->ext;
> > > > SCTP_SO(stream, i)->ext = NULL;
> > > > @@ -77,6 +80,9 @@ static void sctp_stream_outq_migrate(struct sctp_stream *stream,
> > > > }
> > > >
> > > > for (i = outcnt; i < stream->outcnt; i++) {
> > > > + if (SCTP_SO(stream, i)->ext)
> > > > + kfree(SCTP_SO(stream, i)->ext->prio_head);
> > > > +
> > > > kfree(SCTP_SO(stream, i)->ext);
> > > > SCTP_SO(stream, i)->ext = NULL;
> > > > }
> > > > --
> > > > 2.17.1
> > > >
> > > This is not a proper fix:
> > > 1. you shouldn't access "prio_head" outside stream_sched_prio.c.
> > > 2. the prio_head you freed might be used by other out streams, freeing
> > > it unconditionally would cause either a double free or use after free.
> > >
> > > I'm afraid we have to add a ".free_sid" in sctp_sched_ops, and
> > > implement it for sctp_sched_prio, like:
> > >
> > > +static void sctp_sched_prio_free_sid(struct sctp_stream *stream, __u16 sid)
> > > +{
> > > + struct sctp_stream_priorities *prio = SCTP_SO(stream,
> > > sid)->ext->prio_head;
> > > + int i;
> > > +
> > > + if (!prio)
> > > + return;
> > > +
> > > + SCTP_SO(stream, sid)->ext->prio_head = NULL;
> > > + for (i = 0; i < stream->outcnt; i++) {
> >
> > Instead of checking all streams, the for() can/should be replaced by
> > (from sctp_sched_prio_free):
> > if (!list_empty(&prio->prio_sched))
> > return;
> sctp_stream_outq_migrate() is called after unsched_all() for "stream",
> list_empty(prio_sched) is expected to be true.
Good point. Am I missing something or the 'prio_head == prio' below
would always be false then as well?
Anyhow, as this is moving to something that can potentially be called
from other places afterwards, keeping the check doesn't hurt.
>
> Note that kfree(SCTP_SO(new, i)->ext) shouldn't have the reported
> problem, as at that moment, the "new" stream hasn't been set
> stream_sched yet. It means there's only one place that needs to
> call free_sid in sctp_stream_outq_migrate().
> (Maybe Zhengchao can help us confirm this?)
That's the case in Tetsuo's patch (earlier today) as well. Yet, if we
have an official way to free a stream, if it's not error handling
during initialization, it should use it.
>
> >
> > > + if (SCTP_SO(stream, i)->ext &&
> > > + SCTP_SO(stream, i)->ext->prio_head == prio)
> > > + return;
> > > + }
> > > + kfree(prio);
> > > +}
> > > +
> > > static void sctp_sched_prio_free(struct sctp_stream *stream)
> > > {
> > > struct sctp_stream_priorities *prio, *n;
> > > @@ -323,6 +340,7 @@ static struct sctp_sched_ops sctp_sched_prio = {
> > > .get = sctp_sched_prio_get,
> > > .init = sctp_sched_prio_init,
> > > .init_sid = sctp_sched_prio_init_sid,
> > > + .free_sid = sctp_sched_prio_free_sid,
> > > .free = sctp_sched_prio_free,
> > > .enqueue = sctp_sched_prio_enqueue,
> > > .dequeue = sctp_sched_prio_dequeue,
> > >
> > > then call it in sctp_stream_outq_migrate(), like:
> > >
> > > +static void sctp_stream_free_ext(struct sctp_stream *stream, __u16 sid)
> > > +{
> > > + struct sctp_sched_ops *sched = sctp_sched_ops_from_stream(stream);
> > > +
> > > + sched->free_sid(stream, sid);
> > > + kfree(SCTP_SO(stream, sid)->ext);
> > > + SCTP_SO(stream, sid)->ext = NULL;
> > > +}
> > > +
> > > /* Migrates chunks from stream queues to new stream queues if needed,
> > > * but not across associations. Also, removes those chunks to streams
> > > * higher than the new max.
> > > @@ -70,16 +79,14 @@ static void sctp_stream_outq_migrate(struct
> > > sctp_stream *stream,
> > > * sctp_stream_update will swap ->out pointers.
> > > */
> > > for (i = 0; i < outcnt; i++) {
> > > - kfree(SCTP_SO(new, i)->ext);
> > > + sctp_stream_free_ext(new, i);
> > > SCTP_SO(new, i)->ext = SCTP_SO(stream, i)->ext;
> > > SCTP_SO(stream, i)->ext = NULL;
> > > }
> > > }
> > >
> > > - for (i = outcnt; i < stream->outcnt; i++) {
> > > - kfree(SCTP_SO(stream, i)->ext);
> > > - SCTP_SO(stream, i)->ext = NULL;
> > > - }
> > > + for (i = outcnt; i < stream->outcnt; i++)
> > > + sctp_stream_free_ext(new, i);
> > > }
> > >
> > > Marcelo, do you see a better solution?
> >
> > No. Your suggestion is the best I could think of too.
> >
> > Another approach would be to expose sched->free and do all the freeing
> > at once, like sctp_stream_free() does. But the above is looks cleaner
> > and makes it evident that freeing 'ext' is not trivial.
> >
> > With the proposal above, sctp_sched_prio_free() becomes an
> > optimization, if we can call it that. With the for/if replacement
> > above, not even that, and should be removed. Including sctp_sched_ops
> > 'free' pointer.
> Or we extract the common code to another function, like
> sctp_sched_prio_free_head(stream, prio), and pass prio as
> NULL in sctp_sched_prio_free() for freeing all.
>
> >
> > sctp_stream_free() then should be updated to use the new
> > sctp_stream_free_ext() instead, instead of mangling it directly.
> I thought about this, but there is ".free", which is more efficient
> to free all prio than calling ".free_sid" outcnt times.
How much more efficient, just by avoiding retpoline stuff on the
indirect functional call or something else?
>
> I may move free_sid() out of sctp_stream_free_ext(), then in
> sctp_stream_free() we can call sctp_stream_free_ext() without
> calling free_sid(), or just remove sctp_stream_free_ext().
It's easier to maintain it if we have symmetric paths for initializing
and for freeing it and less special cases. We already have
sctp_stream_init_ext(), so having sctp_stream_free_ext() is not off.
I'm happy to review any patch that also updates sctp_stream_free(),
one way or another.
>
> Thanks.
>
> >
> > Makes sense?
> >
> > Thanks,
> > Marcelo
