Skip to contentSkip to conversation

Is this a possible race condition.

Lksctp-developers Vlad Yasevich 3 participants 5 messages Apr 21, 2008 – Apr 22, 2008 SCTP protocol

lore
Share & tools
lore
#0 lore raw
Quoted 86 lines · Neil Horman
Quoted 70 lines · Vlad Yasevich

Hi All

Background:

During the course of some stress testing we found an interesting scenario when
setting the socket send buffer fairly low.

The applications we use sets the send buffer low, and them performs a bunch of
short sends followed by a single receive. The test runs for about 10-12 hours
with the application run a tight loop on each interface of the system.

Once in while this application ends up sitting in sendmsg(). Looking at the kernel
stack dumps (SysRq-T), the task in scheduled out sleeping in sctp_wait_for_sndbuf().
Looking at the memory consumption, everything is 0, i.e. all the skbs have been freed
and there is nothing on any queues.

This got me thinking that maybe wake-ups aren't getting to the socket. I am working on
instrumenting a module to verify this, but looking through the wake-up logic, I think
we have a interesting race.

Race:

We charge every DATA chunk twice against the socket memory accounting, once in sctp_sendmsg()
and once in sctp_packet_transmit() when we allocated the skb that will hold all the bundled
chunks. Now, a very important note is that we only perform the second charge against the
sk_wmem_alloc, so for short span of time socket wmem accounting does not match association
wmem accounting. This is a separate bug, but it shows that the behavior described in the
Background section does not occur when sendbuf_policy is changed.

What I believe is happening is this. The interface queue becomes full and thus a packet
that was charged during sctp_packet_transmit() is still sitting on this queue, while the
sendmsg() call completes successfully. The actually draining of the queue happens later
and freeing up of the buffer happen when the interface is cleaning its tx rings. This
can happen asynchronously from subsequent sendmsg() calls and also without any locking.

So, my supposition is that sock_rfree() and thus sctp_write_space() call happens without a lock and in parallel with another sendmsg() call. It appears that
the race is between checking to see if the asoc->wait queue is active and the activation
of this wait queue by the sctp_wait_for_sndbuf(). Thus, if by some very odd coincidence,
this happens to the last message that the application sent, the application we be scheduled
away and never wake up because the wake up logic already ran.

As you can see, the timing of this is very very narrow, and for us it takes about 10-12 hours
to reproduce. The problem also happens on one or two interfaces out of 13 that are on the
systems and it's never the same interfaces every time.

Quoted 5 lines

Did I miss anything? I was using e1000 as an example of how and when the transmitted skb is
freed and thus above is based on that.

-vlad

Vlad, I'm still digesting all this, and it seems reasonable to me, but I was under the impression that the task waitqueue scheme was specifically written to be race tolerant. the prepare_to_wait_exclusive function enqueues the task to a waitqueue prior to that task calling schedule or schedule_timeout. It does this specifically so that, in the event that a wakeup occurs prior to the schedule call, the task is immediately returned to the run queue, specifically so that this sort of race doesn't occur.

Quoted 7 lines

Of course, I could be missing something, like I said, I'm still digesting this. Either way, it should be easy to test. If you set your socket timeout for your test app to be small (something like 10 seconds), instead of the default unlmited timeout, you should be able to avoid this lockup, if the above reasoning is correct.

I'll provide more feedback as I continue to go through this.

Hey Neil

We just ran a 90 hour test with sndbuf_policy = 1 and it passed :).

The only difference between the 2 is that second accounting done by sctp_packet_transmit(). It changes sk_wmem_alloc, but not assoc.sndbuf_used. Thus in one case, there is an extra change against the send buffer that causes us to exceed it temporarily and thus get scheduled out. Since we never wake up, I am now starting to think that the wake-up code is called while the application is still running. So the race is between a wake-up and a call to schedule_timeout.

The reason this is not an issue for TCP is that TCP never drops the socket lock when going to sleep so the socket is still locked when the buffer cleanup happens. That's not a case for SCTP.

-vlad

#1 lore raw
Quoted 93 lines · Vlad Yasevich
Quoted 92 lines · Neil Horman
Quoted 70 lines · Vlad Yasevich

Hi All

Background:

During the course of some stress testing we found an interesting scenario when
setting the socket send buffer fairly low.

The applications we use sets the send buffer low, and them performs a bunch of
short sends followed by a single receive. The test runs for about 10-12 hours
with the application run a tight loop on each interface of the system.

Once in while this application ends up sitting in sendmsg(). Looking at the kernel
stack dumps (SysRq-T), the task in scheduled out sleeping in sctp_wait_for_sndbuf().
Looking at the memory consumption, everything is 0, i.e. all the skbs have been freed
and there is nothing on any queues.

This got me thinking that maybe wake-ups aren't getting to the socket. I am working on
instrumenting a module to verify this, but looking through the wake-up logic, I think
we have a interesting race.

Race:

We charge every DATA chunk twice against the socket memory accounting, once in sctp_sendmsg()
and once in sctp_packet_transmit() when we allocated the skb that will hold all the bundled
chunks. Now, a very important note is that we only perform the second charge against the
sk_wmem_alloc, so for short span of time socket wmem accounting does not match association
wmem accounting. This is a separate bug, but it shows that the behavior described in the
Background section does not occur when sendbuf_policy is changed.

What I believe is happening is this. The interface queue becomes full and thus a packet
that was charged during sctp_packet_transmit() is still sitting on this queue, while the
sendmsg() call completes successfully. The actually draining of the queue happens later
and freeing up of the buffer happen when the interface is cleaning its tx rings. This
can happen asynchronously from subsequent sendmsg() calls and also without any locking.

So, my supposition is that sock_rfree() and thus sctp_write_space() call happens without a lock and in parallel with another sendmsg() call. It appears that
the race is between checking to see if the asoc->wait queue is active and the activation
of this wait queue by the sctp_wait_for_sndbuf(). Thus, if by some very odd coincidence,
this happens to the last message that the application sent, the application we be scheduled
away and never wake up because the wake up logic already ran.

As you can see, the timing of this is very very narrow, and for us it takes about 10-12 hours
to reproduce. The problem also happens on one or two interfaces out of 13 that are on the
systems and it's never the same interfaces every time.

Quoted 5 lines

Did I miss anything? I was using e1000 as an example of how and when the transmitted skb is
freed and thus above is based on that.

-vlad

Vlad, I'm still digesting all this, and it seems reasonable to me, but I was under the impression that the task waitqueue scheme was specifically written to be race tolerant. the prepare_to_wait_exclusive function enqueues the task to a waitqueue prior to that task calling schedule or schedule_timeout. It does this specifically so that, in the event that a wakeup occurs prior to the schedule call, the task is immediately returned to the run queue, specifically so that this sort of race doesn't occur.

Of course, I could be missing something, like I said, I'm still digesting this. Either way, it should be easy to test. If you set your socket timeout for your test app to be small (something like 10 seconds), instead of the default unlmited timeout, you should be able to avoid this lockup, if the above reasoning is correct.

Quoted 6 lines
Quoted 1 line

I'll provide more feedback as I continue to go through this.

Hey Neil

We just ran a 90 hour test with sndbuf_policy = 1 and it passed :).

I believe you. I completely agree with your observation that there is a small window during which our two accounting values are mismatches, and as a result we might call schedule_timeout without actually having to. My assertion (which may well be wrong) is that the way the waitqueue api is constructed makes that irrelevant (or at least it should).

Quoted 8 lines

The only difference between the 2 is that second accounting done by sctp_packet_transmit(). It changes sk_wmem_alloc, but not assoc.sndbuf_used. Thus in one case, there is an extra change against the send buffer that causes us to exceed it temporarily and thus get scheduled out. Since we never wake up, I am now starting to think that the wake-up code is called while the application is still running. So the race is between a wake-up and a call to schedule_timeout.

Yes, I think this is exactly what you described above, that we enter sctp_wait_for_sndbuf because we detect that we are out of sendbuffer space, and that this happens in parallel with a call to sctp_write_space, and its use of wake_up[_interruptible]. My assertion is that that should be, generally speaking, ok. The pthread library has a very simmilar problem in userspace, in which they do in fact use a mutex to protect against this exact race (see pthread_cond_wait and pthread_cond_signal). The kernel however, has no mechanism to atomically manipulate a spinlock or mutex, and remove oneself from the scheduling queue). So in the kernel we use queues. My assertion (or perhaps assumption), is that prepare_to_wait_exclusive (which is called from sctp_wait_for_sndbuf) adds the task in question to the wait queue, and sets its state to TASK_UNITERRUPTIBLE, or some other 'sleeping' state. At this point the wait mechanism is 'armed' for lack of a better term. Even if the call to wake_up happens between this point and the call to schedule[_timeout], all that will (or should) happen is that you will be removed from the wait_queue that it was added to, its task state will be reset to TASK_RUNNING, and it will be rescheduled, without needing to wait for the timeout. Thats why the call to prepare_to_wait_exclusive is made in sctp_wait_for_sndbuf prior to us checking association state, write space, etc. Its safe to do there because, even if any of those data values are changed right after we read them, the schedule_timeout will effectively just fall through, and we'll break out of the loop on the next iteration. Thats why I suggested that you experiment with setting the sndtimeo to a low value, that would break you out of the loop, and confirm thats where you're deadlocking, and why.

So In summary, I'm saying that I completely agree with you that the way we charge our skbs in the transmit path can lead to us calling wake_up on a given associations wait_queue. I just don't think that doing so is a problem, because of the way the kernels wait_queues guard against the possiblility of a given task scheduling itself away forever. Please take a look and correct me if I'm wrong, but thats how I read the wake up code.

That all being said, I wonder if we're not seeing a race in which the return value of sctp_wspace toggles to a value that might cause a wake_up, and then back to a value that causes a subsequent sleep. You mentioned that your test application runs on each interface of the system. Is it multithreaded, or does it form multiple processes and share a socket by any chance? What if the following were to happen:

1) process A and process B share socket 1
2) process A and B start sending sctp packets over socket 1
3) process A sends a packet, and while in sctp_sendmsg, discovers that its out of send buffer space, it calls sctp_wait_for_sndbuf
4) process A adds itself to the wait queue, check to make sure that no extra space has been returned, and winds up going to sleep.
5) A packet gets sent of the transmit queue and is reclaimed. sctp_write_space is called and updates the buffer accounting, and then wakes up process A. The packet transmit queue is now empty
6) process B tries to send a message over socket 1, get through sctp_sendmsg before process A gets back on the cpu, and re-allocates the sndbuf space
7) process A loops back around, finds that the write space it thought it could get was stolen back (by process B in (6)) and waits again. But since the transmit queue is empty, it will wait forever. Or more precisely, until process B transmits its chunk, and the corresponding data gets free. If something else is blocking process B, its deadlock.

Thats just pie in the sky though, adn makes lots of assumptions.

But, FWIW, my $0.02. Feel free to poke holes in all of this. I do think the wait queue code is hardened against the scenario you are proposing though.

Thanks & Regards
Neil

Quoted 12 lines

The reason this is not an issue for TCP is that TCP never drops the socket lock when going to sleep so the socket is still locked when the buffer cleanup happens. That's not a case for SCTP.

-vlad

--
To unsubscribe from this list: send the line "unsubscribe linux-sctp" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

Signature · 4 lines
/****************************************************
 * Neil Horman <[email protected]>
 * Software Engineer, Red Hat
 ****************************************************/
#2 lore raw
In reply to Vlad Yasevich
Quoted 31 lines · Neil Horman
Quoted 1 line

I believe you. I completely agree with your observation that there is a small window during which our two accounting values are mismatches, and as a result we might call schedule_timeout without actually having to. My assertion (which may well be wrong) is that the way the waitqueue api is constructed makes that irrelevant (or at least it should).

Quoted 8 lines

The only difference between the 2 is that second accounting done by sctp_packet_transmit(). It changes sk_wmem_alloc, but not assoc.sndbuf_used. Thus in one case, there is an extra change against the send buffer that causes us to exceed it temporarily and thus get scheduled out. Since we never wake up, I am now starting to think that the wake-up code is called while the application is still running. So the race is between a wake-up and a call to schedule_timeout.

Yes, I think this is exactly what you described above, that we enter sctp_wait_for_sndbuf because we detect that we are out of sendbuffer space, and that this happens in parallel with a call to sctp_write_space, and its use of wake_up[_interruptible]. My assertion is that that should be, generally speaking, ok. The pthread library has a very simmilar problem in userspace, in which they do in fact use a mutex to protect against this exact race (see pthread_cond_wait and pthread_cond_signal). The kernel however, has no mechanism to atomically manipulate a spinlock or mutex, and remove oneself from the scheduling queue). So in the kernel we use queues. My assertion (or perhaps assumption), is that prepare_to_wait_exclusive (which is called from sctp_wait_for_sndbuf) adds the task in question to the wait queue, and sets its state to TASK_UNITERRUPTIBLE, or some other 'sleeping' state. At this point the wait mechanism is 'armed' for lack of a better term. Even if the call to wake_up happens between this point and the call to schedule[_timeout], all that will (or should) happen is that you will be removed from the wait_queue that it was added to, its task state will be reset to TASK_RUNNING, and it will be

Quoted 8 lines

rescheduled, without needing to wait for the timeout. Thats why the call to prepare_to_wait_exclusive is made in sctp_wait_for_sndbuf prior to us checking association state, write space, etc. Its safe to do there because, even if any of those data values are changed right after we read them, the schedule_timeout will effectively just fall through, and we'll break out of the loop on the next iteration. Thats why I suggested that you experiment with setting the sndtimeo to a low value, that would break you out of the loop, and confirm thats where you're deadlocking, and why.

Will try it.

Quoted 7 lines

So In summary, I'm saying that I completely agree with you that the way we charge our skbs in the transmit path can lead to us calling wake_up on a given associations wait_queue. I just don't think that doing so is a problem, because of the way the kernels wait_queues guard against the possiblility of a given task scheduling itself away forever. Please take a look and correct me if I'm wrong, but thats how I read the wake up code.

Ok, here is a possible race:

CPU1 CPU2
prepare_to_wait_exclusive()

    wait->flags |= WQ_FLAG_EXCLUSIVE;
    spin_lock_irqsave(&q->lock, flags);
    if (list_empty(&wait->task_list))
        __add_wait_queue_tail(q, wait);
	 					   __sctp_write_space()
                                                        waitqueue_active()
                                                        wake_up_interruptible();

    set_current_state().

This one should be possible. The waitqueue_active() checks the 'q' list without a lock. So, it's possible for it to succeed and attempt wake-up before the process state has changed.

Quoted 7 lines

That all being said, I wonder if we're not seeing a race in which the return value of sctp_wspace toggles to a value that might cause a wake_up, and then back to a value that causes a subsequent sleep. You mentioned that your test application runs on each interface of the system. Is it multithreaded, or does it form multiple processes and share a socket by any chance? What if the following were to happen:

There is no socket sharing. It's a stand alone app that just gets run multiple times.

-vlad

#3 lore raw
In reply to Vlad Yasevich

From: Vlad Yasevich <[email protected]>
Date: Mon, 21 Apr 2008 11:27:03 -0400

Quoted 11 lines

The only difference between the 2 is that second accounting done by sctp_packet_transmit(). It changes sk_wmem_alloc, but not assoc.sndbuf_used. Thus in one case, there is an extra change against the send buffer that causes us to exceed it temporarily and thus get scheduled out. Since we never wake up, I am now starting to think that the wake-up code is called while the application is still running. So the race is between a wake-up and a call to schedule_timeout.

The reason this is not an issue for TCP is that TCP never drops the socket lock when going to sleep so the socket is still locked when the buffer cleanup happens. That's not a case for SCTP.

You need to provide more precise context when making such broad statements of fact about TCP :-)))

For example, if TCP hits it's send buffer limits when trying to send, it will release the socket lock while sleeping (via sk_stream_wait_memory() --> sk_wait_event()). Otherwise ACKs would not be processed because a locked socket forces all incoming packets to get queued to the backlog, to be processed at the next release_sock() call.

I imagine you're talking about the case where the memory allocation on transmit itself sleeps.

#4 lore raw
In reply to Vlad Yasevich
Quoted 65 lines · Vlad Yasevich
Quoted 31 lines · Neil Horman
Quoted 1 line

I believe you. I completely agree with your observation that there is a small window during which our two accounting values are mismatches, and as a result we might call schedule_timeout without actually having to. My assertion (which may well be wrong) is that the way the waitqueue api is constructed makes that irrelevant (or at least it should).

Quoted 8 lines

The only difference between the 2 is that second accounting done by sctp_packet_transmit(). It changes sk_wmem_alloc, but not assoc.sndbuf_used. Thus in one case, there is an extra change against the send buffer that causes us to exceed it temporarily and thus get scheduled out. Since we never wake up, I am now starting to think that the wake-up code is called while the application is still running. So the race is between a wake-up and a call to schedule_timeout.

Yes, I think this is exactly what you described above, that we enter sctp_wait_for_sndbuf because we detect that we are out of sendbuffer space, and that this happens in parallel with a call to sctp_write_space, and its use of wake_up[_interruptible]. My assertion is that that should be, generally speaking, ok. The pthread library has a very simmilar problem in userspace, in which they do in fact use a mutex to protect against this exact race (see pthread_cond_wait and pthread_cond_signal). The kernel however, has no mechanism to atomically manipulate a spinlock or mutex, and remove oneself from the scheduling queue). So in the kernel we use queues. My assertion (or perhaps assumption), is that prepare_to_wait_exclusive (which is called from sctp_wait_for_sndbuf) adds the task in question to the wait queue, and sets its state to TASK_UNITERRUPTIBLE, or some other 'sleeping' state. At this point the wait mechanism is 'armed' for lack of a better term. Even if the call to wake_up happens between this point and the call to schedule[_timeout], all that will (or should) happen is that you will be removed from the wait_queue that it was added to, its task state will be reset to TASK_RUNNING, and it will be

Quoted 8 lines

rescheduled, without needing to wait for the timeout. Thats why the call to prepare_to_wait_exclusive is made in sctp_wait_for_sndbuf prior to us checking association state, write space, etc. Its safe to do there because, even if any of those data values are changed right after we read them, the schedule_timeout will effectively just fall through, and we'll break out of the loop on the next iteration. Thats why I suggested that you experiment with setting the sndtimeo to a low value, that would break you out of the loop, and confirm thats where you're deadlocking, and why.

Will try it.

Quoted 7 lines

So In summary, I'm saying that I completely agree with you that the way we charge our skbs in the transmit path can lead to us calling wake_up on a given associations wait_queue. I just don't think that doing so is a problem, because of the way the kernels wait_queues guard against the possiblility of a given task scheduling itself away forever. Please take a look and correct me if I'm wrong, but thats how I read the wake up code.

Ok, here is a possible race:

CPU1 CPU2
prepare_to_wait_exclusive()

   wait->flags |= WQ_FLAG_EXCLUSIVE;
   spin_lock_irqsave(&q->lock, flags);
   if (list_empty(&wait->task_list))
       __add_wait_queue_tail(q, wait);
	 					   __sctp_write_space()
                                                       waitqueue_active()
                                                       wake_up_interruptible();

   set_current_state().
Quoted 4 lines

This one should be possible. The waitqueue_active() checks the 'q' list without a lock. So, it's possible for it to succeed and attempt wake-up before the process state has changed.

Definately, that looks bad. I think that should definately be protected with the q->lock.

Quoted 10 lines
Quoted 7 lines

That all being said, I wonder if we're not seeing a race in which the return value of sctp_wspace toggles to a value that might cause a wake_up, and then back to a value that causes a subsequent sleep. You mentioned that your test application runs on each interface of the system. Is it multithreaded, or does it form multiple processes and share a socket by any chance? What if the following were to happen:

There is no socket sharing. It's a stand alone app that just gets run multiple times.

Ok, that scraps my scenario then :)
Neil

Quoted 1 line

-vlad

Signature · 4 lines
/****************************************************
 * Neil Horman <[email protected]>
 * Software Engineer, Red Hat
 ****************************************************/