Re: [RFC net-next 0/2] prevent sync issues with hw offload of flower

[John Hurley <john.hurley@netronome.com>]

On Thu, Oct 3, 2019 at 6:19 PM Vlad Buslov <vladbu@mellanox.com> wrote:
>
>
> On Thu 03 Oct 2019 at 19:59, John Hurley <john.hurley@netronome.com> wrote:
> > On Thu, Oct 3, 2019 at 5:26 PM Vlad Buslov <vladbu@mellanox.com> wrote:
> >>
> >>
> >> On Thu 03 Oct 2019 at 02:14, John Hurley <john.hurley@netronome.com> wrote:
> >> > Hi,
> >> >
> >> > Putting this out an RFC built on net-next. It fixes some issues
> >> > discovered in testing when using the TC API of OvS to generate flower
> >> > rules and subsequently offloading them to HW. Rules seen contain the same
> >> > match fields or may be rule modifications run as a delete plus an add.
> >> > We're seeing race conditions whereby the rules present in kernel flower
> >> > are out of sync with those offloaded. Note that there are some issues
> >> > that will need fixed in the RFC before it becomes a patch such as
> >> > potential races between releasing locks and re-taking them. However, I'm
> >> > putting this out for comments or potential alternative solutions.
> >> >
> >> > The main cause of the races seem to be in the chain table of cls_api. If
> >> > a tcf_proto is destroyed then it is removed from its chain. If a new
> >> > filter is then added to the same chain with the same priority and protocol
> >> > a new tcf_proto will be created - this may happen before the first is
> >> > fully removed and the hw offload message sent to the driver. In cls_flower
> >> > this means that the fl_ht_insert_unique() function can pass as its
> >> > hashtable is associated with the tcf_proto. We are then in a position
> >> > where the 'delete' and the 'add' are in a race to get offloaded. We also
> >> > noticed that doing an offload add, then checking if a tcf_proto is
> >> > concurrently deleting, then remove the offload if it is, can extend the
> >> > out of order messages. Drivers do not expect to get duplicate rules.
> >> > However, the kernel TC datapath they are not duplicates so we can get out
> >> > of sync here.
> >> >
> >> > The RFC fixes this by adding a pre_destroy hook to cls_api that is called
> >> > when a tcf_proto is signaled to be destroyed but before it is removed from
> >> > its chain (which is essentially the lock for allowing duplicates in
> >> > flower). Flower then uses this new hook to send the hw delete messages
> >> > from tcf_proto destroys, preventing them racing with duplicate adds. It
> >> > also moves the check for 'deleting' to before the sending the hw add
> >> > message.
> >> >
> >> > John Hurley (2):
> >> >   net: sched: add tp_op for pre_destroy
> >> >   net: sched: fix tp destroy race conditions in flower
> >> >
> >> >  include/net/sch_generic.h |  3 +++
> >> >  net/sched/cls_api.c       | 29 ++++++++++++++++++++++++-
> >> >  net/sched/cls_flower.c    | 55 ++++++++++++++++++++++++++---------------------
> >> >  3 files changed, 61 insertions(+), 26 deletions(-)
> >>
> >> Hi John,
> >>
> >> Thanks for working on this!
> >>
> >> Are there any other sources for race conditions described in this
> >> letter? When you describe tcf_proto deletion you say "main cause" but
> >> don't provide any others. If tcf_proto is the only problematic part,
> >
> > Hi Vlad,
> > Thanks for the input.
> > The tcf_proto deletion was the cause from the tests we ran. That's not
> > to say there are not more I wasn't seeing in my analysis.
> >
> >> then it might be worth to look into alternative ways to force concurrent
> >> users to wait for proto deletion/destruction to be properly finished.
> >> Maybe having some table that maps chain id + prio to completion would be
> >> simpler approach? With such infra tcf_proto_create() can wait for
> >> previous proto with same prio and chain to be fully destroyed (including
> >> offloads) before creating a new one.
> >
> > I think a problem with this is that the chain removal functions call
> > tcf_proto_put() (which calls destroy when ref is 0) so, if other
> > concurrent processes (like a dump) have references to the tcf_proto
> > then we may not get the hw offload even by the time the chain deletion
> > function has finished. We would need to make sure this was tracked -
> > say after the tcf_proto_destroy function has completed.
> > How would you suggest doing the wait? With a replay flag as happens in
> > some other places?
> >
> > To me it seems the main problem is that the tcf_proto being in a chain
> > almost acts like the lock to prevent duplicates filters getting to the
> > driver. We need some mechanism to ensure a delete has made it to HW
> > before we release this 'lock'.
>
> Maybe something like:

Ok, I'll need to give this more thought.
Initially it does sound like overkill.

>
> 1. Extend block with hash table with key being chain id and prio
> combined and value is some structure that contains struct completion
> (completed in tcf_proto_destroy() where we sure that all rules were
> removed from hw) and a reference counter.
>

Maybe it could live in each chain rather than block to be more fine grained?
Or would this potentially cause a similar issue on deletion of chains?

> 2. When cls API wants to delete proto instance
> (tcf_chain_tp_delete_empty(), chain flush, etc.), new member is added to
> table from 1. with chain+prio of proto that is being deleted (atomically
> with detaching of proto from chain).
>
> 3. When inserting new proto, verify that there are no corresponding
> entry in hash table with same chain+prio. If there is, increment
> reference counter and wait for completion. Release reference counter
> when completed.

How would the 'wait' work? Loop back via replay flag?
It feels a bit like we are adding a lot more complexity to this and
almost hacking something in to work around a (relatively) newly
introduced problem.

>
> >
> >>
> >> Regards,
> >> Vlad