Re: [ovs-dev] [PATCH v5 00/27] dpif-netdev: Parallel offload processing
On Wed, Jan 26, 2022 at 3:26 AM Ilya Maximets wrote:
>
> On 1/23/22 17:28, Sriharsha Basavapatna wrote:
> > Hi Ilya,
> >
> > On Wed, Jan 19, 2022 at 7:24 AM Ilya Maximets wrote:
> >
> >>
> >> I also spotted one bug where the flow stays offloaded if it gets added
> >> and removed shortly after that. But it's an old existing race condition,
> >> not introduced by this patch set. We basically have to enqueue the
> >> flow deletion regardless of the flow->mark being set.
> >> I'll send a patch for that issue, probably, in the next couple of days.
> >> Or if you want to work on that, that's fine for me too.
> >
> > We are seeing a similar issue with OVS-2.16. While adding and deleting
> > flows in a loop, after a few iterations, rte_flow_destroy() is not
> > seen by the PMD for some of the offloaded flows. And those flows stay
> > offloaded in HW. I tried the following change in 2.16 like you
> > suggested above and it resolved the issue.
> >
> > diff --git a/lib/dpif-netdev.c b/lib/dpif-netdev.c
> > index d6bee2a5a..8cca57f1f 100644
> > --- a/lib/dpif-netdev.c
> > +++ b/lib/dpif-netdev.c
> > @@ -2758,9 +2758,7 @@ dp_netdev_pmd_remove_flow(struct
> > dp_netdev_pmd_thread *pmd,
> > ovs_assert(cls != NULL);
> > dpcls_remove(cls, >cr);
> > cmap_remove(>flow_table, node, dp_netdev_flow_hash(>ufid));
> > -if (flow->mark != INVALID_FLOW_MARK) {
> > -queue_netdev_flow_del(pmd, flow);
> > -}
> > +queue_netdev_flow_del(pmd, flow);
> > flow->dead = true;
> >
> > dp_netdev_flow_unref(flow);
>
> Yeah, I was thinking about something similar. Would you mind sending an
> official patch?
>
> Best regards, Ilya Maximets.
Sure, I will send it.
Thanks,
-Harsha
--
This electronic communication and the information and any files transmitted
with it, or attached to it, are confidential and are intended solely for
the use of the individual or entity to whom it is addressed and may contain
information that is confidential, legally privileged, protected by privacy
laws, or otherwise restricted from disclosure to anyone else. If you are
not the intended recipient or the person responsible for delivering the
e-mail to the intended recipient, you are hereby notified that any use,
copying, distributing, dissemination, forwarding, printing, or copying of
this e-mail is strictly prohibited. If you received this e-mail in error,
please return the e-mail to the sender, delete it from your computer, and
destroy any printed copy of it.
___
dev mailing list
d...@openvswitch.org
https://mail.openvswitch.org/mailman/listinfo/ovs-dev
Re: [ovs-dev] [PATCH v5 00/27] dpif-netdev: Parallel offload processing
On 1/23/22 17:28, Sriharsha Basavapatna wrote:
> Hi Ilya,
>
> On Wed, Jan 19, 2022 at 7:24 AM Ilya Maximets wrote:
>
>>
>> I also spotted one bug where the flow stays offloaded if it gets added
>> and removed shortly after that. But it's an old existing race condition,
>> not introduced by this patch set. We basically have to enqueue the
>> flow deletion regardless of the flow->mark being set.
>> I'll send a patch for that issue, probably, in the next couple of days.
>> Or if you want to work on that, that's fine for me too.
>
> We are seeing a similar issue with OVS-2.16. While adding and deleting
> flows in a loop, after a few iterations, rte_flow_destroy() is not
> seen by the PMD for some of the offloaded flows. And those flows stay
> offloaded in HW. I tried the following change in 2.16 like you
> suggested above and it resolved the issue.
>
> diff --git a/lib/dpif-netdev.c b/lib/dpif-netdev.c
> index d6bee2a5a..8cca57f1f 100644
> --- a/lib/dpif-netdev.c
> +++ b/lib/dpif-netdev.c
> @@ -2758,9 +2758,7 @@ dp_netdev_pmd_remove_flow(struct
> dp_netdev_pmd_thread *pmd,
> ovs_assert(cls != NULL);
> dpcls_remove(cls, >cr);
> cmap_remove(>flow_table, node, dp_netdev_flow_hash(>ufid));
> -if (flow->mark != INVALID_FLOW_MARK) {
> -queue_netdev_flow_del(pmd, flow);
> -}
> +queue_netdev_flow_del(pmd, flow);
> flow->dead = true;
>
> dp_netdev_flow_unref(flow);
Yeah, I was thinking about something similar. Would you mind sending an
official patch?
Best regards, Ilya Maximets.
___
dev mailing list
d...@openvswitch.org
https://mail.openvswitch.org/mailman/listinfo/ovs-dev
Re: [ovs-dev] [PATCH v5 00/27] dpif-netdev: Parallel offload processing
Hi Ilya, On Wed, Jan 19, 2022 at 7:24 AM Ilya Maximets wrote: > > On 9/15/21 15:45, Maxime Coquelin wrote: > > Hi, > > > > On 9/8/21 11:47 AM, Gaetan Rivet wrote: > >> This patch series aims to improve the performance of the management > >> of hw-offloads in dpif-netdev. In the current version, some setup > >> will experience high memory usage and poor latency between a flow > >> decision and its execution regarding hardware offloading. > >> > >> This series starts by measuring key metrics regarding both issues > >> Those patches are introduced first to compare the current status > >> with each improvements introduced. > >> Offloads enqueued and inserted, as well as the latency > >> from queue insertion to hardware insertion is measured. A new > >> command 'ovs-appctl dpctl/offload-stats-show' is introduced > >> to show the current measure. > >> > >> In my current performance test setup I am measuring an > >> average latency hovering between 1~2 seconds. > >> After the optimizations, it is reduced to 500~900 ms. > >> Finally when using multiple threads and with proper driver > >> support[1], it is measured in the order of 1 ms. > >> > >> A few modules are introduced: > >> > >> * An ID pool with reduced capabilities, simplifying its > >> operations and allowing better performances in both > >> single and multi-thread setup. > >> > >> * A lockless queue between PMDs / revalidators and > >> offload thread(s). As the number of PMDs increases, > >> contention can be high on the shared queue. > >> This queue is designed to serve as message queue > >> between threads. > >> > >> * A bounded lockless MPMC ring and some helpers for > >> calculating moving averages. > >> > >> * A moving average module for Cumulative and Exponential > >> moving averages. > >> > >> The netdev-offload-dpdk module is made thread-safe. > >> Internal maps are made per-netdev instead, and locks are > >> taken for shorter critical sections within the module. > >> > >> CI result: https://github.com/grivet/ovs/actions/runs/554918929 > >> > >> [1]: The rte_flow API was made thread-safe in the 20.11 DPDK > >> release. Drivers that do not implement those operations > >> concurrently are protected by a lock. Others will > >> allow better concurrency, that improve the result > >> of this series. > >> > >> v2: > >> > >> * Improved the MPSC queue API to simplify usage. > >> > >> * Moved flush operation from initiator thread to offload > >> thread(s). This ensures offload metadata are shared only > >> among the offload thread pool. > >> > >> * Flush operation needs additional thread synchronization. > >> The ovs_barrier currently triggers a UAF. Add a unit-test to > >> validate its operations and a fix for the UAF. > >> > >> CI result: https://github.com/grivet/ovs/actions/runs/741430135 > >>The error comes from a failure to download 'automake' on > >>osx, unrelated to any change in this series. > >> > >> v3: > >> > >> * Re-ordered commits so fixes are first. No conflict seen currently, > >> but it might prevent them if some requested changes to the series > >> were to move code in the same parts. > >> > >> * Modified the reduced quiescing the thread to use ovsrcu_quiesce(), > >> and base next_rcu on the current time value (after quiescing happened, > >> however long it takes). > >> > >> * Added Reviewed-by tags to the relevant commits. > >> > >> CI result: https://github.com/grivet/ovs/actions/runs/782655601 > >> > >> v4: > >> > >> * Modified the seq-pool to use batches of IDs with a spinlock > >> instead of lockless rings. > >> > >> * The llring structure is removed. > >> > >> * Due to the length of the changes to the structure, some > >> acked-by or reviewed-by were not ported to the id-fpool patch. > >> > >> CI result: https://github.com/grivet/ovs/actions/runs/921095015 > >> > >> v5: > >> > >> * Rebase on master. > >> Conflicts were seen related to the vxlan-decap and pmd rebalance > >> series. > >> > >> * Fix typo in xchg patch spotted by Maxime Coquelin. > >> > >> * Added Reviewed-by Maxime Coquelin on 4 patches. > > > > > > I went through the changes between v4 and v5, and would like to confirm > > these are are minor and OK to me. > > > > Regards, > > Maxime > > > >> CI result: https://github.com/grivet/ovs/actions/runs/1212804378 > >> > >> Gaetan Rivet (27): > >> ovs-thread: Fix barrier use-after-free > >> dpif-netdev: Rename flow offload thread > >> tests: Add ovs-barrier unit test > >> netdev: Add flow API uninit function > >> netdev-offload-dpdk: Use per-netdev offload metadata > >> netdev-offload-dpdk: Implement hw-offload statistics read > >> dpctl: Add function to read hardware offload statistics > >> dpif-netdev: Rename offload thread structure > >> mov-avg: Add a moving average helper structure > >> dpif-netdev: Implement hardware offloads stats query > >>
Re: [ovs-dev] [PATCH v5 00/27] dpif-netdev: Parallel offload processing
On 9/15/21 15:45, Maxime Coquelin wrote: > Hi, > > On 9/8/21 11:47 AM, Gaetan Rivet wrote: >> This patch series aims to improve the performance of the management >> of hw-offloads in dpif-netdev. In the current version, some setup >> will experience high memory usage and poor latency between a flow >> decision and its execution regarding hardware offloading. >> >> This series starts by measuring key metrics regarding both issues >> Those patches are introduced first to compare the current status >> with each improvements introduced. >> Offloads enqueued and inserted, as well as the latency >> from queue insertion to hardware insertion is measured. A new >> command 'ovs-appctl dpctl/offload-stats-show' is introduced >> to show the current measure. >> >> In my current performance test setup I am measuring an >> average latency hovering between 1~2 seconds. >> After the optimizations, it is reduced to 500~900 ms. >> Finally when using multiple threads and with proper driver >> support[1], it is measured in the order of 1 ms. >> >> A few modules are introduced: >> >> * An ID pool with reduced capabilities, simplifying its >> operations and allowing better performances in both >> single and multi-thread setup. >> >> * A lockless queue between PMDs / revalidators and >> offload thread(s). As the number of PMDs increases, >> contention can be high on the shared queue. >> This queue is designed to serve as message queue >> between threads. >> >> * A bounded lockless MPMC ring and some helpers for >> calculating moving averages. >> >> * A moving average module for Cumulative and Exponential >> moving averages. >> >> The netdev-offload-dpdk module is made thread-safe. >> Internal maps are made per-netdev instead, and locks are >> taken for shorter critical sections within the module. >> >> CI result: https://github.com/grivet/ovs/actions/runs/554918929 >> >> [1]: The rte_flow API was made thread-safe in the 20.11 DPDK >> release. Drivers that do not implement those operations >> concurrently are protected by a lock. Others will >> allow better concurrency, that improve the result >> of this series. >> >> v2: >> >> * Improved the MPSC queue API to simplify usage. >> >> * Moved flush operation from initiator thread to offload >> thread(s). This ensures offload metadata are shared only >> among the offload thread pool. >> >> * Flush operation needs additional thread synchronization. >> The ovs_barrier currently triggers a UAF. Add a unit-test to >> validate its operations and a fix for the UAF. >> >> CI result: https://github.com/grivet/ovs/actions/runs/741430135 >>The error comes from a failure to download 'automake' on >>osx, unrelated to any change in this series. >> >> v3: >> >> * Re-ordered commits so fixes are first. No conflict seen currently, >> but it might prevent them if some requested changes to the series >> were to move code in the same parts. >> >> * Modified the reduced quiescing the thread to use ovsrcu_quiesce(), >> and base next_rcu on the current time value (after quiescing happened, >> however long it takes). >> >> * Added Reviewed-by tags to the relevant commits. >> >> CI result: https://github.com/grivet/ovs/actions/runs/782655601 >> >> v4: >> >> * Modified the seq-pool to use batches of IDs with a spinlock >> instead of lockless rings. >> >> * The llring structure is removed. >> >> * Due to the length of the changes to the structure, some >> acked-by or reviewed-by were not ported to the id-fpool patch. >> >> CI result: https://github.com/grivet/ovs/actions/runs/921095015 >> >> v5: >> >> * Rebase on master. >> Conflicts were seen related to the vxlan-decap and pmd rebalance >> series. >> >> * Fix typo in xchg patch spotted by Maxime Coquelin. >> >> * Added Reviewed-by Maxime Coquelin on 4 patches. > > > I went through the changes between v4 and v5, and would like to confirm > these are are minor and OK to me. > > Regards, > Maxime > >> CI result: https://github.com/grivet/ovs/actions/runs/1212804378 >> >> Gaetan Rivet (27): >> ovs-thread: Fix barrier use-after-free >> dpif-netdev: Rename flow offload thread >> tests: Add ovs-barrier unit test >> netdev: Add flow API uninit function >> netdev-offload-dpdk: Use per-netdev offload metadata >> netdev-offload-dpdk: Implement hw-offload statistics read >> dpctl: Add function to read hardware offload statistics >> dpif-netdev: Rename offload thread structure >> mov-avg: Add a moving average helper structure >> dpif-netdev: Implement hardware offloads stats query >> ovs-atomic: Expose atomic exchange operation >> mpsc-queue: Module for lock-free message passing >> id-fpool: Module for fast ID generation >> netdev-offload: Add multi-thread API >> dpif-netdev: Quiesce offload thread periodically >> dpif-netdev: Postpone flow offload item freeing >> dpif-netdev:
Re: [ovs-dev] [PATCH v5 00/27] dpif-netdev: Parallel offload processing
Hi, On 9/8/21 11:47 AM, Gaetan Rivet wrote: > This patch series aims to improve the performance of the management > of hw-offloads in dpif-netdev. In the current version, some setup > will experience high memory usage and poor latency between a flow > decision and its execution regarding hardware offloading. > > This series starts by measuring key metrics regarding both issues > Those patches are introduced first to compare the current status > with each improvements introduced. > Offloads enqueued and inserted, as well as the latency > from queue insertion to hardware insertion is measured. A new > command 'ovs-appctl dpctl/offload-stats-show' is introduced > to show the current measure. > > In my current performance test setup I am measuring an > average latency hovering between 1~2 seconds. > After the optimizations, it is reduced to 500~900 ms. > Finally when using multiple threads and with proper driver > support[1], it is measured in the order of 1 ms. > > A few modules are introduced: > > * An ID pool with reduced capabilities, simplifying its > operations and allowing better performances in both > single and multi-thread setup. > > * A lockless queue between PMDs / revalidators and > offload thread(s). As the number of PMDs increases, > contention can be high on the shared queue. > This queue is designed to serve as message queue > between threads. > > * A bounded lockless MPMC ring and some helpers for > calculating moving averages. > > * A moving average module for Cumulative and Exponential > moving averages. > > The netdev-offload-dpdk module is made thread-safe. > Internal maps are made per-netdev instead, and locks are > taken for shorter critical sections within the module. > > CI result: https://github.com/grivet/ovs/actions/runs/554918929 > > [1]: The rte_flow API was made thread-safe in the 20.11 DPDK > release. Drivers that do not implement those operations > concurrently are protected by a lock. Others will > allow better concurrency, that improve the result > of this series. > > v2: > > * Improved the MPSC queue API to simplify usage. > > * Moved flush operation from initiator thread to offload > thread(s). This ensures offload metadata are shared only > among the offload thread pool. > > * Flush operation needs additional thread synchronization. > The ovs_barrier currently triggers a UAF. Add a unit-test to > validate its operations and a fix for the UAF. > > CI result: https://github.com/grivet/ovs/actions/runs/741430135 >The error comes from a failure to download 'automake' on >osx, unrelated to any change in this series. > > v3: > > * Re-ordered commits so fixes are first. No conflict seen currently, > but it might prevent them if some requested changes to the series > were to move code in the same parts. > > * Modified the reduced quiescing the thread to use ovsrcu_quiesce(), > and base next_rcu on the current time value (after quiescing happened, > however long it takes). > > * Added Reviewed-by tags to the relevant commits. > > CI result: https://github.com/grivet/ovs/actions/runs/782655601 > > v4: > > * Modified the seq-pool to use batches of IDs with a spinlock > instead of lockless rings. > > * The llring structure is removed. > > * Due to the length of the changes to the structure, some > acked-by or reviewed-by were not ported to the id-fpool patch. > > CI result: https://github.com/grivet/ovs/actions/runs/921095015 > > v5: > > * Rebase on master. > Conflicts were seen related to the vxlan-decap and pmd rebalance > series. > > * Fix typo in xchg patch spotted by Maxime Coquelin. > > * Added Reviewed-by Maxime Coquelin on 4 patches. I went through the changes between v4 and v5, and would like to confirm these are are minor and OK to me. Regards, Maxime > CI result: https://github.com/grivet/ovs/actions/runs/1212804378 > > Gaetan Rivet (27): > ovs-thread: Fix barrier use-after-free > dpif-netdev: Rename flow offload thread > tests: Add ovs-barrier unit test > netdev: Add flow API uninit function > netdev-offload-dpdk: Use per-netdev offload metadata > netdev-offload-dpdk: Implement hw-offload statistics read > dpctl: Add function to read hardware offload statistics > dpif-netdev: Rename offload thread structure > mov-avg: Add a moving average helper structure > dpif-netdev: Implement hardware offloads stats query > ovs-atomic: Expose atomic exchange operation > mpsc-queue: Module for lock-free message passing > id-fpool: Module for fast ID generation > netdev-offload: Add multi-thread API > dpif-netdev: Quiesce offload thread periodically > dpif-netdev: Postpone flow offload item freeing > dpif-netdev: Use id-fpool for mark allocation > dpif-netdev: Introduce tagged union of offload requests > dpif-netdev: Execute flush from offload thread >
[ovs-dev] [PATCH v5 00/27] dpif-netdev: Parallel offload processing
This patch series aims to improve the performance of the management of hw-offloads in dpif-netdev. In the current version, some setup will experience high memory usage and poor latency between a flow decision and its execution regarding hardware offloading. This series starts by measuring key metrics regarding both issues Those patches are introduced first to compare the current status with each improvements introduced. Offloads enqueued and inserted, as well as the latency from queue insertion to hardware insertion is measured. A new command 'ovs-appctl dpctl/offload-stats-show' is introduced to show the current measure. In my current performance test setup I am measuring an average latency hovering between 1~2 seconds. After the optimizations, it is reduced to 500~900 ms. Finally when using multiple threads and with proper driver support[1], it is measured in the order of 1 ms. A few modules are introduced: * An ID pool with reduced capabilities, simplifying its operations and allowing better performances in both single and multi-thread setup. * A lockless queue between PMDs / revalidators and offload thread(s). As the number of PMDs increases, contention can be high on the shared queue. This queue is designed to serve as message queue between threads. * A bounded lockless MPMC ring and some helpers for calculating moving averages. * A moving average module for Cumulative and Exponential moving averages. The netdev-offload-dpdk module is made thread-safe. Internal maps are made per-netdev instead, and locks are taken for shorter critical sections within the module. CI result: https://github.com/grivet/ovs/actions/runs/554918929 [1]: The rte_flow API was made thread-safe in the 20.11 DPDK release. Drivers that do not implement those operations concurrently are protected by a lock. Others will allow better concurrency, that improve the result of this series. v2: * Improved the MPSC queue API to simplify usage. * Moved flush operation from initiator thread to offload thread(s). This ensures offload metadata are shared only among the offload thread pool. * Flush operation needs additional thread synchronization. The ovs_barrier currently triggers a UAF. Add a unit-test to validate its operations and a fix for the UAF. CI result: https://github.com/grivet/ovs/actions/runs/741430135 The error comes from a failure to download 'automake' on osx, unrelated to any change in this series. v3: * Re-ordered commits so fixes are first. No conflict seen currently, but it might prevent them if some requested changes to the series were to move code in the same parts. * Modified the reduced quiescing the thread to use ovsrcu_quiesce(), and base next_rcu on the current time value (after quiescing happened, however long it takes). * Added Reviewed-by tags to the relevant commits. CI result: https://github.com/grivet/ovs/actions/runs/782655601 v4: * Modified the seq-pool to use batches of IDs with a spinlock instead of lockless rings. * The llring structure is removed. * Due to the length of the changes to the structure, some acked-by or reviewed-by were not ported to the id-fpool patch. CI result: https://github.com/grivet/ovs/actions/runs/921095015 v5: * Rebase on master. Conflicts were seen related to the vxlan-decap and pmd rebalance series. * Fix typo in xchg patch spotted by Maxime Coquelin. * Added Reviewed-by Maxime Coquelin on 4 patches. CI result: https://github.com/grivet/ovs/actions/runs/1212804378 Gaetan Rivet (27): ovs-thread: Fix barrier use-after-free dpif-netdev: Rename flow offload thread tests: Add ovs-barrier unit test netdev: Add flow API uninit function netdev-offload-dpdk: Use per-netdev offload metadata netdev-offload-dpdk: Implement hw-offload statistics read dpctl: Add function to read hardware offload statistics dpif-netdev: Rename offload thread structure mov-avg: Add a moving average helper structure dpif-netdev: Implement hardware offloads stats query ovs-atomic: Expose atomic exchange operation mpsc-queue: Module for lock-free message passing id-fpool: Module for fast ID generation netdev-offload: Add multi-thread API dpif-netdev: Quiesce offload thread periodically dpif-netdev: Postpone flow offload item freeing dpif-netdev: Use id-fpool for mark allocation dpif-netdev: Introduce tagged union of offload requests dpif-netdev: Execute flush from offload thread netdev-offload-dpdk: Use per-thread HW offload stats netdev-offload-dpdk: Lock rte_flow map access netdev-offload-dpdk: Protect concurrent offload destroy/query dpif-netdev: Use lockless queue to manage offloads dpif-netdev: Make megaflow and mark mappings thread objects dpif-netdev: Replace port mutex by rwlock dpif-netdev: Use one or more offload threads netdev-dpdk: Remove rte-flow API access locks
