[PATCH net] net: sched: fix packet stuck problem for lockless qdisc
Lockless qdisc has below concurrent problem: cpu0 cpu1 . . q->enqueue . . . qdisc_run_begin() . . . dequeue_skb() . . . sch_direct_xmit() . . . . q->enqueue . qdisc_run_begin() . return and do nothing . . qdisc_run_end() . cpu1 enqueue a skb without calling __qdisc_run() because cpu0 has not released the lock yet and spin_trylock() return false for cpu1 in qdisc_run_begin(), and cpu0 do not see the skb enqueued by cpu1 when calling dequeue_skb() because cpu1 may enqueue the skb after cpu0 calling dequeue_skb() and before cpu0 calling qdisc_run_end(). Lockless qdisc has another concurrent problem when tx_action is involved: cpu0(serving tx_action) cpu1 cpu2 . . . . q->enqueue . . qdisc_run_begin() . . dequeue_skb() . . . q->enqueue . . . . sch_direct_xmit() . . . qdisc_run_begin() . . return and do nothing . . . clear __QDISC_STATE_SCHED . . qdisc_run_begin() . . return and do nothing . . . . . . qdisc_run_begin() . This patch fixes the above data race by: 1. Set a flag after spin_trylock() return false. 2. Retry a spin_trylock() in case other CPU may not see the new flag after it releases the lock. 3. reschedule if the flag is set after the lock is released at the end of qdisc_run_end(). For tx_action case, the flags is also set when cpu1 is at the end if qdisc_run_begin(), so tx_action will be rescheduled again to dequeue the skb enqueued by cpu2. Also clear the flag before dequeuing in order to reduce the overhead of the above process, and aviod doing the heavy test_and_clear_bit() at the end of qdisc_run_end(). Fixes: 6b3ba9146fe6 ("net: sched: allow qdiscs to handle locking") Signed-off-by: Yunsheng Lin <linyunsheng@huawei.com> --- For those who has not been following the qdsic scheduling discussion, there is packet stuck problem for lockless qdisc, see [1], and I has done some cleanup and added some enhanced features too, see [2] [3]. While I was doing the optimization for lockless qdisc, it accurred to me that these optimization is useless if there is still basic bug in lockless qdisc, even the bug is not easily reproducible. So look through [1] again, I found that the data race for tx action mentioned by Michael, and thought deep about it and came up with this patch trying to fix it. So I am really appreciated some who still has the reproducer can try this patch and report back. 1. https://lore.kernel.org/netdev/d102074f-7489-e35a-98cf-e2cad7efd8a2@netround... 2. https://patchwork.kernel.org/project/netdevbpf/patch/1615777818-13969-1-git-... 3. https://patchwork.kernel.org/project/netdevbpf/patch/1615800610-34700-1-git-... --- include/net/sch_generic.h | 23 ++++++++++++++++++++--- net/sched/sch_generic.c | 1 + 2 files changed, 21 insertions(+), 3 deletions(-) diff --git a/include/net/sch_generic.h b/include/net/sch_generic.h index f7a6e14..4220eab 100644 --- a/include/net/sch_generic.h +++ b/include/net/sch_generic.h @@ -36,6 +36,7 @@ struct qdisc_rate_table { enum qdisc_state_t { __QDISC_STATE_SCHED, __QDISC_STATE_DEACTIVATED, + __QDISC_STATE_NEED_RESCHEDULE, }; struct qdisc_size_table { @@ -159,8 +160,17 @@ static inline bool qdisc_is_empty(const struct Qdisc *qdisc) static inline bool qdisc_run_begin(struct Qdisc *qdisc) { if (qdisc->flags & TCQ_F_NOLOCK) { - if (!spin_trylock(&qdisc->seqlock)) - return false; + if (!spin_trylock(&qdisc->seqlock)) { + set_bit(__QDISC_STATE_NEED_RESCHEDULE, + &qdisc->state); + + /* Retry again in case other CPU may not see the + * new flags after it releases the lock at the + * end of qdisc_run_end(). + */ + if (!spin_trylock(&qdisc->seqlock)) + return false; + } WRITE_ONCE(qdisc->empty, false); } else if (qdisc_is_running(qdisc)) { return false; @@ -176,8 +186,15 @@ static inline bool qdisc_run_begin(struct Qdisc *qdisc) static inline void qdisc_run_end(struct Qdisc *qdisc) { write_seqcount_end(&qdisc->running); - if (qdisc->flags & TCQ_F_NOLOCK) + if (qdisc->flags & TCQ_F_NOLOCK) { spin_unlock(&qdisc->seqlock); + + if (unlikely(test_bit(__QDISC_STATE_NEED_RESCHEDULE, + &qdisc->state) && + !test_bit(__QDISC_STATE_DEACTIVATED, + &qdisc->state))) + __netif_schedule(qdisc); + } } static inline bool qdisc_may_bulk(const struct Qdisc *qdisc) diff --git a/net/sched/sch_generic.c b/net/sched/sch_generic.c index 44991ea..25d75d8 100644 --- a/net/sched/sch_generic.c +++ b/net/sched/sch_generic.c @@ -205,6 +205,7 @@ static struct sk_buff *dequeue_skb(struct Qdisc *q, bool *validate, const struct netdev_queue *txq = q->dev_queue; struct sk_buff *skb = NULL; + clear_bit(__QDISC_STATE_NEED_RESCHEDULE, &q->state); *packets = 1; if (unlikely(!skb_queue_empty(&q->gso_skb))) { spinlock_t *lock = NULL; -- 2.7.4
On 2021/3/18 14:53, Yunsheng Lin wrote:
Lockless qdisc has below concurrent problem: cpu0 cpu1 . . q->enqueue . . . qdisc_run_begin() . . . dequeue_skb() . . . sch_direct_xmit() . . . . q->enqueue . qdisc_run_begin() . return and do nothing . . qdisc_run_end() .
cpu1 enqueue a skb without calling __qdisc_run() because cpu0 has not released the lock yet and spin_trylock() return false for cpu1 in qdisc_run_begin(), and cpu0 do not see the skb enqueued by cpu1 when calling dequeue_skb() because cpu1 may enqueue the skb after cpu0 calling dequeue_skb() and before cpu0 calling qdisc_run_end().
Lockless qdisc has another concurrent problem when tx_action is involved:
cpu0(serving tx_action) cpu1 cpu2 . . . . q->enqueue . . qdisc_run_begin() . . dequeue_skb() . . . q->enqueue . . . . sch_direct_xmit() . . . qdisc_run_begin() . . return and do nothing . . . clear __QDISC_STATE_SCHED . . qdisc_run_begin() . . return and do nothing . . . . . . qdisc_run_begin() .
This patch fixes the above data race by: 1. Set a flag after spin_trylock() return false. 2. Retry a spin_trylock() in case other CPU may not see the new flag after it releases the lock. 3. reschedule if the flag is set after the lock is released at the end of qdisc_run_end().
For tx_action case, the flags is also set when cpu1 is at the end if qdisc_run_begin(), so tx_action will be rescheduled again to dequeue the skb enqueued by cpu2.
Also clear the flag before dequeuing in order to reduce the overhead of the above process, and aviod doing the heavy test_and_clear_bit() at the end of qdisc_run_end().
Fixes: 6b3ba9146fe6 ("net: sched: allow qdiscs to handle locking") Signed-off-by: Yunsheng Lin <linyunsheng@huawei.com> --- For those who has not been following the qdsic scheduling discussion, there is packet stuck problem for lockless qdisc, see [1], and I has done some cleanup and added some enhanced features too, see [2] [3]. While I was doing the optimization for lockless qdisc, it accurred to me that these optimization is useless if there is still basic bug in lockless qdisc, even the bug is not easily reproducible. So look through [1] again, I found that the data race for tx action mentioned by Michael, and thought deep about it and came up with this patch trying to fix it.
So I am really appreciated some who still has the reproducer can try this patch and report back.
I had done some performance test to see if there is value to fix the packet stuck problem and support lockless qdisc bypass, here is some result using pktgen in 'queue_xmit' mode on a dummy device as Paolo Abeni had done in [1], and using pfifo_fast qdisc: threads vanilla locked-qdisc vanilla+this_patch 1 2.6Mpps 2.9Mpps 2.5Mpps 2 3.9Mpps 4.8Mpps 3.6Mpps 4 5.6Mpps 3.0Mpps 4.7Mpps 8 2.7Mpps 1.6Mpps 2.8Mpps 16 2.2Mpps 1.3Mpps 2.3Mpps locked-qdisc: test by removing the "TCQ_F_NOLOCK | TCQ_F_CPUSTATS". And add the lockless qdisc bypatch and other optimization upon this patch: threads patch_set_1 patch_set_2 patch_set_3 1 2.5Mpps 3.0Mpps 3.0Mpps 2 3.6Mpps 4.1Mpps 5.3Mpps 4 4.7Mpps 4.6Mpps 5.1Mpps 8 2.8Mpps 2.6Mpps 2.7Mpps 16 2.3Mpps 2.2Mpps 2.2Mpps patch_set_1: vanilla + this_patch patch_set_2: vanilla + this_patch + lockless_qdisc_bypass_patch patch_set_3: vanilla + this_patch + lockless_qdisc_bypass_patch + remove_seq_operation_for_lockless_qdisc_optimization + check_rc_before_calling_qdisc_run()_optimization + spin_trylock()_retry_optimization. So all the fix and optimization added together, the lockless qdisc has better performance than vanilla except for the 4 threads case, which has about 9% performance degradation than vanilla one, but still better than the locked-qdisc.
1. https://lore.kernel.org/netdev/d102074f-7489-e35a-98cf-e2cad7efd8a2@netround... 2. https://patchwork.kernel.org/project/netdevbpf/patch/1615777818-13969-1-git-... 3. https://patchwork.kernel.org/project/netdevbpf/patch/1615800610-34700-1-git-...
On Fri, Mar 19, 2021 at 2:25 AM Yunsheng Lin <linyunsheng@huawei.com> wrote:
I had done some performance test to see if there is value to fix the packet stuck problem and support lockless qdisc bypass, here is some result using pktgen in 'queue_xmit' mode on a dummy device as Paolo Abeni had done in [1], and using pfifo_fast qdisc:
threads vanilla locked-qdisc vanilla+this_patch 1 2.6Mpps 2.9Mpps 2.5Mpps 2 3.9Mpps 4.8Mpps 3.6Mpps 4 5.6Mpps 3.0Mpps 4.7Mpps 8 2.7Mpps 1.6Mpps 2.8Mpps 16 2.2Mpps 1.3Mpps 2.3Mpps
locked-qdisc: test by removing the "TCQ_F_NOLOCK | TCQ_F_CPUSTATS".
I read this as this patch introduces somehow a performance regression for -net, as the lockless bypass patch you submitted is for -net-next. Thanks.
On 2021/3/20 3:45, Cong Wang wrote:
On Fri, Mar 19, 2021 at 2:25 AM Yunsheng Lin <linyunsheng@huawei.com> wrote:
I had done some performance test to see if there is value to fix the packet stuck problem and support lockless qdisc bypass, here is some result using pktgen in 'queue_xmit' mode on a dummy device as Paolo Abeni had done in [1], and using pfifo_fast qdisc:
threads vanilla locked-qdisc vanilla+this_patch 1 2.6Mpps 2.9Mpps 2.5Mpps 2 3.9Mpps 4.8Mpps 3.6Mpps 4 5.6Mpps 3.0Mpps 4.7Mpps 8 2.7Mpps 1.6Mpps 2.8Mpps 16 2.2Mpps 1.3Mpps 2.3Mpps
locked-qdisc: test by removing the "TCQ_F_NOLOCK | TCQ_F_CPUSTATS".
I read this as this patch introduces somehow a performance regression for -net, as the lockless bypass patch you submitted is for -net-next.
Yes, right now there is performance regression for fixing this bug, but the problem is that if we can not fix the above data race without any performance regression, do you prefer to send this patch to -net, or to -net-next with the lockless bypass patch? Any idea to fix this with less performance regression?
Thanks.
.
On Wed, Mar 17, 2021 at 11:52 PM Yunsheng Lin <linyunsheng@huawei.com> wrote:
Lockless qdisc has below concurrent problem: cpu0 cpu1 . . q->enqueue . . . qdisc_run_begin() . . . dequeue_skb() . . . sch_direct_xmit() . . . . q->enqueue . qdisc_run_begin() . return and do nothing . . qdisc_run_end() .
cpu1 enqueue a skb without calling __qdisc_run() because cpu0 has not released the lock yet and spin_trylock() return false for cpu1 in qdisc_run_begin(), and cpu0 do not see the skb enqueued by cpu1 when calling dequeue_skb() because cpu1 may enqueue the skb after cpu0 calling dequeue_skb() and before cpu0 calling qdisc_run_end().
Lockless qdisc has another concurrent problem when tx_action is involved:
cpu0(serving tx_action) cpu1 cpu2 . . . . q->enqueue . . qdisc_run_begin() . . dequeue_skb() . . . q->enqueue . . . . sch_direct_xmit() . . . qdisc_run_begin() . . return and do nothing . . . clear __QDISC_STATE_SCHED . . qdisc_run_begin() . . return and do nothing . . . . . . qdisc_run_begin() .
This patch fixes the above data race by: 1. Set a flag after spin_trylock() return false. 2. Retry a spin_trylock() in case other CPU may not see the new flag after it releases the lock. 3. reschedule if the flag is set after the lock is released at the end of qdisc_run_end().
For tx_action case, the flags is also set when cpu1 is at the end if qdisc_run_begin(), so tx_action will be rescheduled again to dequeue the skb enqueued by cpu2.
Also clear the flag before dequeuing in order to reduce the overhead of the above process, and aviod doing the heavy test_and_clear_bit() at the end of qdisc_run_end().
Fixes: 6b3ba9146fe6 ("net: sched: allow qdiscs to handle locking") Signed-off-by: Yunsheng Lin <linyunsheng@huawei.com> --- For those who has not been following the qdsic scheduling discussion, there is packet stuck problem for lockless qdisc, see [1], and I has done some cleanup and added some enhanced features too, see [2] [3]. While I was doing the optimization for lockless qdisc, it accurred to me that these optimization is useless if there is still basic bug in lockless qdisc, even the bug is not easily reproducible. So look through [1] again, I found that the data race for tx action mentioned by Michael, and thought deep about it and came up with this patch trying to fix it.
So I am really appreciated some who still has the reproducer can try this patch and report back.
1. https://lore.kernel.org/netdev/d102074f-7489-e35a-98cf-e2cad7efd8a2@netround... 2. https://patchwork.kernel.org/project/netdevbpf/patch/1615777818-13969-1-git-... 3. https://patchwork.kernel.org/project/netdevbpf/patch/1615800610-34700-1-git-... --- include/net/sch_generic.h | 23 ++++++++++++++++++++--- net/sched/sch_generic.c | 1 + 2 files changed, 21 insertions(+), 3 deletions(-)
diff --git a/include/net/sch_generic.h b/include/net/sch_generic.h index f7a6e14..4220eab 100644 --- a/include/net/sch_generic.h +++ b/include/net/sch_generic.h @@ -36,6 +36,7 @@ struct qdisc_rate_table { enum qdisc_state_t { __QDISC_STATE_SCHED, __QDISC_STATE_DEACTIVATED, + __QDISC_STATE_NEED_RESCHEDULE, };
struct qdisc_size_table { @@ -159,8 +160,17 @@ static inline bool qdisc_is_empty(const struct Qdisc *qdisc) static inline bool qdisc_run_begin(struct Qdisc *qdisc) { if (qdisc->flags & TCQ_F_NOLOCK) { - if (!spin_trylock(&qdisc->seqlock)) - return false; + if (!spin_trylock(&qdisc->seqlock)) { + set_bit(__QDISC_STATE_NEED_RESCHEDULE, + &qdisc->state);
Why do we need another bit? I mean why not just call __netif_schedule()?
+ + /* Retry again in case other CPU may not see the + * new flags after it releases the lock at the + * end of qdisc_run_end(). + */ + if (!spin_trylock(&qdisc->seqlock)) + return false; + } WRITE_ONCE(qdisc->empty, false); } else if (qdisc_is_running(qdisc)) { return false; @@ -176,8 +186,15 @@ static inline bool qdisc_run_begin(struct Qdisc *qdisc) static inline void qdisc_run_end(struct Qdisc *qdisc) { write_seqcount_end(&qdisc->running); - if (qdisc->flags & TCQ_F_NOLOCK) + if (qdisc->flags & TCQ_F_NOLOCK) { spin_unlock(&qdisc->seqlock); + + if (unlikely(test_bit(__QDISC_STATE_NEED_RESCHEDULE, + &qdisc->state) && + !test_bit(__QDISC_STATE_DEACTIVATED, + &qdisc->state)))
Testing two bits one by one is not atomic...
+ __netif_schedule(qdisc); + } }
static inline bool qdisc_may_bulk(const struct Qdisc *qdisc) diff --git a/net/sched/sch_generic.c b/net/sched/sch_generic.c index 44991ea..25d75d8 100644 --- a/net/sched/sch_generic.c +++ b/net/sched/sch_generic.c @@ -205,6 +205,7 @@ static struct sk_buff *dequeue_skb(struct Qdisc *q, bool *validate, const struct netdev_queue *txq = q->dev_queue; struct sk_buff *skb = NULL;
+ clear_bit(__QDISC_STATE_NEED_RESCHEDULE, &q->state); *packets = 1; if (unlikely(!skb_queue_empty(&q->gso_skb))) { spinlock_t *lock = NULL; -- 2.7.4
On 2021/3/20 3:40, Cong Wang wrote:
On Wed, Mar 17, 2021 at 11:52 PM Yunsheng Lin <linyunsheng@huawei.com> wrote:
Lockless qdisc has below concurrent problem: cpu0 cpu1 . . q->enqueue . . . qdisc_run_begin() . . . dequeue_skb() . . . sch_direct_xmit() . . . . q->enqueue . qdisc_run_begin() . return and do nothing . . qdisc_run_end() .
cpu1 enqueue a skb without calling __qdisc_run() because cpu0 has not released the lock yet and spin_trylock() return false for cpu1 in qdisc_run_begin(), and cpu0 do not see the skb enqueued by cpu1 when calling dequeue_skb() because cpu1 may enqueue the skb after cpu0 calling dequeue_skb() and before cpu0 calling qdisc_run_end().
Lockless qdisc has another concurrent problem when tx_action is involved:
cpu0(serving tx_action) cpu1 cpu2 . . . . q->enqueue . . qdisc_run_begin() . . dequeue_skb() . . . q->enqueue . . . . sch_direct_xmit() . . . qdisc_run_begin() . . return and do nothing . . . clear __QDISC_STATE_SCHED . . qdisc_run_begin() . . return and do nothing . . . . . . qdisc_run_begin() .
This patch fixes the above data race by: 1. Set a flag after spin_trylock() return false. 2. Retry a spin_trylock() in case other CPU may not see the new flag after it releases the lock. 3. reschedule if the flag is set after the lock is released at the end of qdisc_run_end().
For tx_action case, the flags is also set when cpu1 is at the end if qdisc_run_begin(), so tx_action will be rescheduled again to dequeue the skb enqueued by cpu2.
Also clear the flag before dequeuing in order to reduce the overhead of the above process, and aviod doing the heavy test_and_clear_bit() at the end of qdisc_run_end().
Fixes: 6b3ba9146fe6 ("net: sched: allow qdiscs to handle locking") Signed-off-by: Yunsheng Lin <linyunsheng@huawei.com> --- For those who has not been following the qdsic scheduling discussion, there is packet stuck problem for lockless qdisc, see [1], and I has done some cleanup and added some enhanced features too, see [2] [3]. While I was doing the optimization for lockless qdisc, it accurred to me that these optimization is useless if there is still basic bug in lockless qdisc, even the bug is not easily reproducible. So look through [1] again, I found that the data race for tx action mentioned by Michael, and thought deep about it and came up with this patch trying to fix it.
So I am really appreciated some who still has the reproducer can try this patch and report back.
1. https://lore.kernel.org/netdev/d102074f-7489-e35a-98cf-e2cad7efd8a2@netround... 2. https://patchwork.kernel.org/project/netdevbpf/patch/1615777818-13969-1-git-... 3. https://patchwork.kernel.org/project/netdevbpf/patch/1615800610-34700-1-git-... --- include/net/sch_generic.h | 23 ++++++++++++++++++++--- net/sched/sch_generic.c | 1 + 2 files changed, 21 insertions(+), 3 deletions(-)
diff --git a/include/net/sch_generic.h b/include/net/sch_generic.h index f7a6e14..4220eab 100644 --- a/include/net/sch_generic.h +++ b/include/net/sch_generic.h @@ -36,6 +36,7 @@ struct qdisc_rate_table { enum qdisc_state_t { __QDISC_STATE_SCHED, __QDISC_STATE_DEACTIVATED, + __QDISC_STATE_NEED_RESCHEDULE, };
struct qdisc_size_table { @@ -159,8 +160,17 @@ static inline bool qdisc_is_empty(const struct Qdisc *qdisc) static inline bool qdisc_run_begin(struct Qdisc *qdisc) { if (qdisc->flags & TCQ_F_NOLOCK) { - if (!spin_trylock(&qdisc->seqlock)) - return false; + if (!spin_trylock(&qdisc->seqlock)) { + set_bit(__QDISC_STATE_NEED_RESCHEDULE, + &qdisc->state);
Why do we need another bit? I mean why not just call __netif_schedule()?
I think that was your proposal in [1], the only difference is that it also handle the tx_action case when __netif_schedule() is called here. So yes, it can also fix the two data race described in this patch, but with a bigger performance degradation, quoting performance data in the [1]: pktgen threads vanilla patched delta nr kpps kpps % 1 3240 3240 0 2 3910 2710 -30.5 4 5140 4920 -4 performance data with this patch: threads vanilla vanilla+this_patch delta 1 2.6Mpps 2.5Mpps -3% 2 3.9Mpps 3.6Mpps -7% 4 5.6Mpps 4.7Mpps -16% So the performance is why I does not call __netif_schedule() directly here. 1. https://lore.kernel.org/netdev/d102074f-7489-e35a-98cf-e2cad7efd8a2@netround...
+ + /* Retry again in case other CPU may not see the + * new flags after it releases the lock at the + * end of qdisc_run_end(). + */ + if (!spin_trylock(&qdisc->seqlock)) + return false; + } WRITE_ONCE(qdisc->empty, false); } else if (qdisc_is_running(qdisc)) { return false; @@ -176,8 +186,15 @@ static inline bool qdisc_run_begin(struct Qdisc *qdisc) static inline void qdisc_run_end(struct Qdisc *qdisc) { write_seqcount_end(&qdisc->running); - if (qdisc->flags & TCQ_F_NOLOCK) + if (qdisc->flags & TCQ_F_NOLOCK) { spin_unlock(&qdisc->seqlock); + + if (unlikely(test_bit(__QDISC_STATE_NEED_RESCHEDULE, + &qdisc->state) && + !test_bit(__QDISC_STATE_DEACTIVATED, + &qdisc->state)))
Testing two bits one by one is not atomic...
For non-tx_action case, actually it is atomic because the above two bits testing is within the rcu protection, and qdisc reset will do a synchronize_net() after setting __QDISC_STATE_DEACTIVATED. For tx_action case, I think we need a rcu protection explicitly in net_tx_action() too, at least for PREEMPT_RCU: https://stackoverflow.com/questions/21287932/is-it-necessary-invoke-rcu-read...
+ __netif_schedule(qdisc); + } }
static inline bool qdisc_may_bulk(const struct Qdisc *qdisc) diff --git a/net/sched/sch_generic.c b/net/sched/sch_generic.c index 44991ea..25d75d8 100644 --- a/net/sched/sch_generic.c +++ b/net/sched/sch_generic.c @@ -205,6 +205,7 @@ static struct sk_buff *dequeue_skb(struct Qdisc *q, bool *validate, const struct netdev_queue *txq = q->dev_queue; struct sk_buff *skb = NULL;
+ clear_bit(__QDISC_STATE_NEED_RESCHEDULE, &q->state); *packets = 1; if (unlikely(!skb_queue_empty(&q->gso_skb))) { spinlock_t *lock = NULL; -- 2.7.4
.
On 2021/3/22 9:31, Yunsheng Lin wrote:
On 2021/3/20 3:40, Cong Wang wrote:
On Wed, Mar 17, 2021 at 11:52 PM Yunsheng Lin <linyunsheng@huawei.com> wrote:
Lockless qdisc has below concurrent problem: cpu0 cpu1 . . q->enqueue . . . qdisc_run_begin() . . . dequeue_skb() . . . sch_direct_xmit() . . . . q->enqueue . qdisc_run_begin() . return and do nothing . . qdisc_run_end() .
cpu1 enqueue a skb without calling __qdisc_run() because cpu0 has not released the lock yet and spin_trylock() return false for cpu1 in qdisc_run_begin(), and cpu0 do not see the skb enqueued by cpu1 when calling dequeue_skb() because cpu1 may enqueue the skb after cpu0 calling dequeue_skb() and before cpu0 calling qdisc_run_end().
Lockless qdisc has another concurrent problem when tx_action is involved:
cpu0(serving tx_action) cpu1 cpu2 . . . . q->enqueue . . qdisc_run_begin() . . dequeue_skb() . . . q->enqueue . . . . sch_direct_xmit() . . . qdisc_run_begin() . . return and do nothing . . . clear __QDISC_STATE_SCHED . . qdisc_run_begin() . . return and do nothing . . . . . . qdisc_run_begin() .
This patch fixes the above data race by: 1. Set a flag after spin_trylock() return false. 2. Retry a spin_trylock() in case other CPU may not see the new flag after it releases the lock. 3. reschedule if the flag is set after the lock is released at the end of qdisc_run_end().
For tx_action case, the flags is also set when cpu1 is at the end if qdisc_run_begin(), so tx_action will be rescheduled again to dequeue the skb enqueued by cpu2.
Also clear the flag before dequeuing in order to reduce the overhead of the above process, and aviod doing the heavy test_and_clear_bit() at the end of qdisc_run_end().
Fixes: 6b3ba9146fe6 ("net: sched: allow qdiscs to handle locking") Signed-off-by: Yunsheng Lin <linyunsheng@huawei.com> --- For those who has not been following the qdsic scheduling discussion, there is packet stuck problem for lockless qdisc, see [1], and I has done some cleanup and added some enhanced features too, see [2] [3]. While I was doing the optimization for lockless qdisc, it accurred to me that these optimization is useless if there is still basic bug in lockless qdisc, even the bug is not easily reproducible. So look through [1] again, I found that the data race for tx action mentioned by Michael, and thought deep about it and came up with this patch trying to fix it.
So I am really appreciated some who still has the reproducer can try this patch and report back.
1. https://lore.kernel.org/netdev/d102074f-7489-e35a-98cf-e2cad7efd8a2@netround... 2. https://patchwork.kernel.org/project/netdevbpf/patch/1615777818-13969-1-git-... 3. https://patchwork.kernel.org/project/netdevbpf/patch/1615800610-34700-1-git-... --- include/net/sch_generic.h | 23 ++++++++++++++++++++--- net/sched/sch_generic.c | 1 + 2 files changed, 21 insertions(+), 3 deletions(-)
diff --git a/include/net/sch_generic.h b/include/net/sch_generic.h index f7a6e14..4220eab 100644 --- a/include/net/sch_generic.h +++ b/include/net/sch_generic.h @@ -36,6 +36,7 @@ struct qdisc_rate_table { enum qdisc_state_t { __QDISC_STATE_SCHED, __QDISC_STATE_DEACTIVATED, + __QDISC_STATE_NEED_RESCHEDULE, };
struct qdisc_size_table { @@ -159,8 +160,17 @@ static inline bool qdisc_is_empty(const struct Qdisc *qdisc) static inline bool qdisc_run_begin(struct Qdisc *qdisc) { if (qdisc->flags & TCQ_F_NOLOCK) { - if (!spin_trylock(&qdisc->seqlock)) - return false; + if (!spin_trylock(&qdisc->seqlock)) { + set_bit(__QDISC_STATE_NEED_RESCHEDULE, + &qdisc->state);
Why do we need another bit? I mean why not just call __netif_schedule()?
I think that was your proposal in [1], the only difference is that it also handle the tx_action case when __netif_schedule() is called here.
So yes, it can also fix the two data race described in this patch, but with a bigger performance degradation, quoting performance data in the [1]:
pktgen threads vanilla patched delta nr kpps kpps %
1 3240 3240 0 2 3910 2710 -30.5 4 5140 4920 -4
performance data with this patch:
threads vanilla vanilla+this_patch delta 1 2.6Mpps 2.5Mpps -3% 2 3.9Mpps 3.6Mpps -7% 4 5.6Mpps 4.7Mpps -16%
So the performance is why I does not call __netif_schedule() directly here.
1. https://lore.kernel.org/netdev/d102074f-7489-e35a-98cf-e2cad7efd8a2@netround...
When spin_trylock() and __netif_schedule() is avoided as much as possible，we have performance closer to vanilla kernel now. Performance data and additional optimization based on this patch as below: threads vanilla vanilla+this_patch+opt delta 1 2.6Mpps 2.48Mpps -4% 2 3.9Mpps 3.77Mpps -3% 4 5.6Mpps 5.51Mpps -1% diff --git a/include/net/sch_generic.h b/include/net/sch_generic.h index 9afe231..dad6f8f 100644 --- a/include/net/sch_generic.h +++ b/include/net/sch_generic.h @@ -160,21 +160,30 @@ static inline bool qdisc_is_empty(const struct Qdisc *qdisc) static inline bool qdisc_run_begin(struct Qdisc *qdisc) { if (qdisc->flags & TCQ_F_NOLOCK) { - if (!spin_trylock(&qdisc->seqlock)) { - set_bit(__QDISC_STATE_NEED_RESCHEDULE, - &qdisc->state); - - /* Retry again in case other CPU may not see the - * new flags after it releases the lock at the - * end of qdisc_run_end(). - */ - if (!spin_trylock(&qdisc->seqlock)) - return false; - } + bool dont_retry = test_bit(__QDISC_STATE_NEED_RESCHEDULE, + &qdisc->state); + + if (spin_trylock(&qdisc->seqlock)) + goto out; + + if (dont_retry) + return false; + + set_bit(__QDISC_STATE_NEED_RESCHEDULE, + &qdisc->state); + + /* Retry again in case other CPU may not see the new flag + * after it releases the lock at the end of qdisc_run_end(). + */ + if (!spin_trylock(&qdisc->seqlock)) + return false; + WRITE_ONCE(qdisc->empty, false); } else if (qdisc_is_running(qdisc)) { return false; } + +out: /* Variant of write_seqcount_begin() telling lockdep a trylock * was attempted. */ diff --git a/net/sched/sch_generic.c b/net/sched/sch_generic.c index 0ab551c..3099247 100644 --- a/net/sched/sch_generic.c +++ b/net/sched/sch_generic.c @@ -205,7 +205,6 @@ static struct sk_buff *dequeue_skb(struct Qdisc *q, bool *validate, const struct netdev_queue *txq = q->dev_queue; struct sk_buff *skb = NULL; - clear_bit(__QDISC_STATE_NEED_RESCHEDULE, &q->state); *packets = 1; if (unlikely(!skb_queue_empty(&q->gso_skb))) { spinlock_t *lock = NULL; @@ -641,8 +640,10 @@ static struct sk_buff *pfifo_fast_dequeue(struct Qdisc *qdisc) { struct pfifo_fast_priv *priv = qdisc_priv(qdisc); struct sk_buff *skb = NULL; + bool need_retry = true; int band; +retry: for (band = 0; band < PFIFO_FAST_BANDS && !skb; band++) { struct skb_array *q = band2list(priv, band); @@ -653,6 +654,12 @@ static struct sk_buff *pfifo_fast_dequeue(struct Qdisc *qdisc) } if (likely(skb)) { qdisc_update_stats_at_dequeue(qdisc, skb); + } else if (need_retry) { + clear_bit(__QDISC_STATE_NEED_RESCHEDULE, &qdisc->state); + smp_mb__after_atomic(); + need_retry = false; + + goto retry; } else { WRITE_ONCE(qdisc->empty, true); }
Currently pfifo_fast has both TCQ_F_CAN_BYPASS and TCQ_F_NOLOCK flag set, but queue discipline by-pass does not work for lockless qdisc because skb is always enqueued to qdisc even when the qdisc is empty, see __dev_xmit_skb(). This patch calls sch_direct_xmit() to transmit the skb directly to the driver for empty lockless qdisc too, which aviod enqueuing and dequeuing operation. qdisc->empty is set to false whenever a skb is enqueued, see pfifo_fast_enqueue(), and is set to true when skb dequeuing return NULL, see pfifo_fast_dequeue(). There is a data race between enqueue/dequeue and qdisc->empty setting, qdisc->empty is only used as a hint, so we need to call sch_may_need_requeuing() to see if the queue is really empty and if there is requeued skb, which has higher priority than the current skb. The performance for ip_forward test increases about 10% with this patch. Signed-off-by: Yunsheng Lin <linyunsheng@huawei.com> --- Hi, Vladimir and Ahmad Please give it a test to see if there is any out of order packet for this patch, which has removed the priv->lock added in RFC v2. There is a data race as below: CPU1 CPU2 qdisc_run_begin(q) . . q->enqueue() sch_may_need_requeuing() . return true . . . . . q->enqueue() . When above happen, the skb enqueued by CPU1 is dequeued after the skb enqueued by CPU2 because sch_may_need_requeuing() return true. If there is not qdisc bypass, the CPU1 has better chance to queue the skb quicker than CPU2. This patch does not take care of the above data race, because I view this as similar as below: Even at the same time CPU1 and CPU2 write the skb to two socket which both heading to the same qdisc, there is no guarantee that which skb will hit the qdisc first, becuase there is a lot of factor like interrupt/softirq/cache miss/scheduling afffecting that. So I hope the above data race will not cause problem for Vladimir and Ahmad. --- include/net/pkt_sched.h | 1 + include/net/sch_generic.h | 1 - net/core/dev.c | 22 ++++++++++++++++++++++ net/sched/sch_generic.c | 11 +++++++++++ 4 files changed, 34 insertions(+), 1 deletion(-) diff --git a/include/net/pkt_sched.h b/include/net/pkt_sched.h index f5c1bee..5715ddf 100644 --- a/include/net/pkt_sched.h +++ b/include/net/pkt_sched.h @@ -122,6 +122,7 @@ void qdisc_warn_nonwc(const char *txt, struct Qdisc *qdisc); bool sch_direct_xmit(struct sk_buff *skb, struct Qdisc *q, struct net_device *dev, struct netdev_queue *txq, spinlock_t *root_lock, bool validate); +bool sch_may_need_requeuing(struct Qdisc *q); void __qdisc_run(struct Qdisc *q); diff --git a/include/net/sch_generic.h b/include/net/sch_generic.h index f7a6e14..e08cc77 100644 --- a/include/net/sch_generic.h +++ b/include/net/sch_generic.h @@ -161,7 +161,6 @@ static inline bool qdisc_run_begin(struct Qdisc *qdisc) if (qdisc->flags & TCQ_F_NOLOCK) { if (!spin_trylock(&qdisc->seqlock)) return false; - WRITE_ONCE(qdisc->empty, false); } else if (qdisc_is_running(qdisc)) { return false; } diff --git a/net/core/dev.c b/net/core/dev.c index be941ed..317180a 100644 --- a/net/core/dev.c +++ b/net/core/dev.c @@ -3796,9 +3796,31 @@ static inline int __dev_xmit_skb(struct sk_buff *skb, struct Qdisc *q, qdisc_calculate_pkt_len(skb, q); if (q->flags & TCQ_F_NOLOCK) { + if (q->flags & TCQ_F_CAN_BYPASS && READ_ONCE(q->empty) && + qdisc_run_begin(q)) { + if (sch_may_need_requeuing(q)) { + rc = q->enqueue(skb, q, &to_free) & NET_XMIT_MASK; + __qdisc_run(q); + qdisc_run_end(q); + + goto no_lock_out; + } + + qdisc_bstats_cpu_update(q, skb); + + if (sch_direct_xmit(skb, q, dev, txq, NULL, true) && + !READ_ONCE(q->empty)) + __qdisc_run(q); + + qdisc_run_end(q); + return NET_XMIT_SUCCESS; + } + rc = q->enqueue(skb, q, &to_free) & NET_XMIT_MASK; + WRITE_ONCE(q->empty, false); qdisc_run(q); +no_lock_out: if (unlikely(to_free)) kfree_skb_list(to_free); return rc; diff --git a/net/sched/sch_generic.c b/net/sched/sch_generic.c index 44991ea..2145fdad 100644 --- a/net/sched/sch_generic.c +++ b/net/sched/sch_generic.c @@ -146,6 +146,8 @@ static inline void dev_requeue_skb(struct sk_buff *skb, struct Qdisc *q) } if (lock) spin_unlock(lock); + + WRITE_ONCE(q->empty, false); __netif_schedule(q); } @@ -273,6 +275,15 @@ static struct sk_buff *dequeue_skb(struct Qdisc *q, bool *validate, return skb; } +bool sch_may_need_requeuing(struct Qdisc *q) +{ + if (likely(skb_queue_empty(&q->gso_skb) && + !q->ops->peek(q))) + return false; + + return true; +} + /* * Transmit possibly several skbs, and handle the return status as * required. Owning running seqcount bit guarantees that -- 2.7.4
Hi Yunsheng, On Mon, Mar 22, 2021 at 05:09:16PM +0800, Yunsheng Lin wrote:
Currently pfifo_fast has both TCQ_F_CAN_BYPASS and TCQ_F_NOLOCK flag set, but queue discipline by-pass does not work for lockless qdisc because skb is always enqueued to qdisc even when the qdisc is empty, see __dev_xmit_skb().
This patch calls sch_direct_xmit() to transmit the skb directly to the driver for empty lockless qdisc too, which aviod enqueuing and dequeuing operation. qdisc->empty is set to false whenever a skb is enqueued, see pfifo_fast_enqueue(), and is set to true when skb dequeuing return NULL, see pfifo_fast_dequeue().
There is a data race between enqueue/dequeue and qdisc->empty setting, qdisc->empty is only used as a hint, so we need to call sch_may_need_requeuing() to see if the queue is really empty and if there is requeued skb, which has higher priority than the current skb.
The performance for ip_forward test increases about 10% with this patch.
Signed-off-by: Yunsheng Lin <linyunsheng@huawei.com> --- Hi, Vladimir and Ahmad Please give it a test to see if there is any out of order packet for this patch, which has removed the priv->lock added in RFC v2.
There is a data race as below:
CPU1 CPU2 qdisc_run_begin(q) . . q->enqueue() sch_may_need_requeuing() . return true . . . . . q->enqueue() .
When above happen, the skb enqueued by CPU1 is dequeued after the skb enqueued by CPU2 because sch_may_need_requeuing() return true. If there is not qdisc bypass, the CPU1 has better chance to queue the skb quicker than CPU2.
This patch does not take care of the above data race, because I view this as similar as below:
Even at the same time CPU1 and CPU2 write the skb to two socket which both heading to the same qdisc, there is no guarantee that which skb will hit the qdisc first, becuase there is a lot of factor like interrupt/softirq/cache miss/scheduling afffecting that.
So I hope the above data race will not cause problem for Vladimir and Ahmad. ---
Preliminary results on my test setup look fine, but please allow me to run the canfdtest overnight, since as you say, races are still theoretically possible.
On Mon, Mar 22, 2021 at 10:00:33PM +0200, Vladimir Oltean wrote:
Hi Yunsheng,
On Mon, Mar 22, 2021 at 05:09:16PM +0800, Yunsheng Lin wrote:
Currently pfifo_fast has both TCQ_F_CAN_BYPASS and TCQ_F_NOLOCK flag set, but queue discipline by-pass does not work for lockless qdisc because skb is always enqueued to qdisc even when the qdisc is empty, see __dev_xmit_skb().
This patch calls sch_direct_xmit() to transmit the skb directly to the driver for empty lockless qdisc too, which aviod enqueuing and dequeuing operation. qdisc->empty is set to false whenever a skb is enqueued, see pfifo_fast_enqueue(), and is set to true when skb dequeuing return NULL, see pfifo_fast_dequeue().
There is a data race between enqueue/dequeue and qdisc->empty setting, qdisc->empty is only used as a hint, so we need to call sch_may_need_requeuing() to see if the queue is really empty and if there is requeued skb, which has higher priority than the current skb.
The performance for ip_forward test increases about 10% with this patch.
Signed-off-by: Yunsheng Lin <linyunsheng@huawei.com> --- Hi, Vladimir and Ahmad Please give it a test to see if there is any out of order packet for this patch, which has removed the priv->lock added in RFC v2.
There is a data race as below:
CPU1 CPU2 qdisc_run_begin(q) . . q->enqueue() sch_may_need_requeuing() . return true . . . . . q->enqueue() .
When above happen, the skb enqueued by CPU1 is dequeued after the skb enqueued by CPU2 because sch_may_need_requeuing() return true. If there is not qdisc bypass, the CPU1 has better chance to queue the skb quicker than CPU2.
This patch does not take care of the above data race, because I view this as similar as below:
Even at the same time CPU1 and CPU2 write the skb to two socket which both heading to the same qdisc, there is no guarantee that which skb will hit the qdisc first, becuase there is a lot of factor like interrupt/softirq/cache miss/scheduling afffecting that.
So I hope the above data race will not cause problem for Vladimir and Ahmad. ---
Preliminary results on my test setup look fine, but please allow me to run the canfdtest overnight, since as you say, races are still theoretically possible.
I haven't found any issues during the overnight test and until now. Tested-by: Vladimir Oltean <vladimir.oltean@nxp.com> # flexcan
Hi, On 22.03.21 10:09, Yunsheng Lin wrote:
Currently pfifo_fast has both TCQ_F_CAN_BYPASS and TCQ_F_NOLOCK flag set, but queue discipline by-pass does not work for lockless qdisc because skb is always enqueued to qdisc even when the qdisc is empty, see __dev_xmit_skb().
This patch calls sch_direct_xmit() to transmit the skb directly to the driver for empty lockless qdisc too, which aviod enqueuing and dequeuing operation. qdisc->empty is set to false whenever a skb is enqueued, see pfifo_fast_enqueue(), and is set to true when skb dequeuing return NULL, see pfifo_fast_dequeue().
There is a data race between enqueue/dequeue and qdisc->empty setting, qdisc->empty is only used as a hint, so we need to call sch_may_need_requeuing() to see if the queue is really empty and if there is requeued skb, which has higher priority than the current skb.
The performance for ip_forward test increases about 10% with this patch.
Signed-off-by: Yunsheng Lin <linyunsheng@huawei.com> --- Hi, Vladimir and Ahmad Please give it a test to see if there is any out of order packet for this patch, which has removed the priv->lock added in RFC v2.
Overnight test (10h, 64 mil frames) didn't see any out-of-order frames between 2 FlexCANs on a dual core machine: Tested-by: Ahmad Fatoum <a.fatoum@pengutronix.de> No performance measurements taken.
There is a data race as below:
CPU1 CPU2 qdisc_run_begin(q) . . q->enqueue() sch_may_need_requeuing() . return true . . . . . q->enqueue() .
When above happen, the skb enqueued by CPU1 is dequeued after the skb enqueued by CPU2 because sch_may_need_requeuing() return true. If there is not qdisc bypass, the CPU1 has better chance to queue the skb quicker than CPU2.
This patch does not take care of the above data race, because I view this as similar as below:
Even at the same time CPU1 and CPU2 write the skb to two socket which both heading to the same qdisc, there is no guarantee that which skb will hit the qdisc first, becuase there is a lot of factor like interrupt/softirq/cache miss/scheduling afffecting that.
So I hope the above data race will not cause problem for Vladimir and Ahmad. --- include/net/pkt_sched.h | 1 + include/net/sch_generic.h | 1 - net/core/dev.c | 22 ++++++++++++++++++++++ net/sched/sch_generic.c | 11 +++++++++++ 4 files changed, 34 insertions(+), 1 deletion(-)
diff --git a/include/net/pkt_sched.h b/include/net/pkt_sched.h index f5c1bee..5715ddf 100644 --- a/include/net/pkt_sched.h +++ b/include/net/pkt_sched.h @@ -122,6 +122,7 @@ void qdisc_warn_nonwc(const char *txt, struct Qdisc *qdisc); bool sch_direct_xmit(struct sk_buff *skb, struct Qdisc *q, struct net_device *dev, struct netdev_queue *txq, spinlock_t *root_lock, bool validate); +bool sch_may_need_requeuing(struct Qdisc *q);
void __qdisc_run(struct Qdisc *q);
diff --git a/include/net/sch_generic.h b/include/net/sch_generic.h index f7a6e14..e08cc77 100644 --- a/include/net/sch_generic.h +++ b/include/net/sch_generic.h @@ -161,7 +161,6 @@ static inline bool qdisc_run_begin(struct Qdisc *qdisc) if (qdisc->flags & TCQ_F_NOLOCK) { if (!spin_trylock(&qdisc->seqlock)) return false; - WRITE_ONCE(qdisc->empty, false); } else if (qdisc_is_running(qdisc)) { return false; } diff --git a/net/core/dev.c b/net/core/dev.c index be941ed..317180a 100644 --- a/net/core/dev.c +++ b/net/core/dev.c @@ -3796,9 +3796,31 @@ static inline int __dev_xmit_skb(struct sk_buff *skb, struct Qdisc *q, qdisc_calculate_pkt_len(skb, q);
if (q->flags & TCQ_F_NOLOCK) { + if (q->flags & TCQ_F_CAN_BYPASS && READ_ONCE(q->empty) && + qdisc_run_begin(q)) { + if (sch_may_need_requeuing(q)) { + rc = q->enqueue(skb, q, &to_free) & NET_XMIT_MASK; + __qdisc_run(q); + qdisc_run_end(q); + + goto no_lock_out; + } + + qdisc_bstats_cpu_update(q, skb); + + if (sch_direct_xmit(skb, q, dev, txq, NULL, true) && + !READ_ONCE(q->empty)) + __qdisc_run(q); + + qdisc_run_end(q); + return NET_XMIT_SUCCESS; + } + rc = q->enqueue(skb, q, &to_free) & NET_XMIT_MASK; + WRITE_ONCE(q->empty, false); qdisc_run(q);
+no_lock_out: if (unlikely(to_free)) kfree_skb_list(to_free); return rc; diff --git a/net/sched/sch_generic.c b/net/sched/sch_generic.c index 44991ea..2145fdad 100644 --- a/net/sched/sch_generic.c +++ b/net/sched/sch_generic.c @@ -146,6 +146,8 @@ static inline void dev_requeue_skb(struct sk_buff *skb, struct Qdisc *q) } if (lock) spin_unlock(lock); + + WRITE_ONCE(q->empty, false); __netif_schedule(q); }
@@ -273,6 +275,15 @@ static struct sk_buff *dequeue_skb(struct Qdisc *q, bool *validate, return skb; }
+bool sch_may_need_requeuing(struct Qdisc *q) +{ + if (likely(skb_queue_empty(&q->gso_skb) && + !q->ops->peek(q))) + return false; + + return true; +} + /* * Transmit possibly several skbs, and handle the return status as * required. Owning running seqcount bit guarantees that
-- Pengutronix e.K. | | Steuerwalder Str. 21 | http://www.pengutronix.de/ | 31137 Hildesheim, Germany | Phone: +49-5121-206917-0 | Amtsgericht Hildesheim, HRA 2686 | Fax: +49-5121-206917-5555 |
On 2021/3/23 14:37, Ahmad Fatoum wrote:
Hi,
On 22.03.21 10:09, Yunsheng Lin wrote:
Currently pfifo_fast has both TCQ_F_CAN_BYPASS and TCQ_F_NOLOCK flag set, but queue discipline by-pass does not work for lockless qdisc because skb is always enqueued to qdisc even when the qdisc is empty, see __dev_xmit_skb().
This patch calls sch_direct_xmit() to transmit the skb directly to the driver for empty lockless qdisc too, which aviod enqueuing and dequeuing operation. qdisc->empty is set to false whenever a skb is enqueued, see pfifo_fast_enqueue(), and is set to true when skb dequeuing return NULL, see pfifo_fast_dequeue().
There is a data race between enqueue/dequeue and qdisc->empty setting, qdisc->empty is only used as a hint, so we need to call sch_may_need_requeuing() to see if the queue is really empty and if there is requeued skb, which has higher priority than the current skb.
The performance for ip_forward test increases about 10% with this patch.
Signed-off-by: Yunsheng Lin <linyunsheng@huawei.com> --- Hi, Vladimir and Ahmad Please give it a test to see if there is any out of order packet for this patch, which has removed the priv->lock added in RFC v2.
Overnight test (10h, 64 mil frames) didn't see any out-of-order frames between 2 FlexCANs on a dual core machine:
Tested-by: Ahmad Fatoum <a.fatoum@pengutronix.de>
No performance measurements taken.
Thanks for the testing. And I has done the performance measurement. L3 forward testing improves from 1.09Mpps to 1.21Mpps, still about 10% improvement. pktgen + dummy netdev: threads without+this_patch with+this_patch delta 1 2.56Mpps 3.11Mpps +21% 2 3.76Mpps 4.31Mpps +14% 4 5.51Mpps 5.53Mpps +0.3% 8 2.81Mpps 2.72Mpps -3% 16 2.24Mpps 2.22Mpps -0.8%
participants (4)
-
Ahmad Fatoum -
Cong Wang -
Vladimir Oltean -
Yunsheng Lin