Ido Schimmel
2019-Feb-16 18:43 UTC
[Bridge] [PATCH RFC] net: bridge: don't flood known multicast traffic when snooping is enabled
On Sat, Feb 16, 2019 at 10:05:40AM +0200, Nikolay Aleksandrov wrote:> On 15/02/2019 19:13, Linus L?ssing wrote: > > On Fri, Feb 15, 2019 at 03:04:27PM +0200, Nikolay Aleksandrov wrote: > >> Every user would expect to have traffic forwarded only to the configured > >> mdb destination when snooping is enabled, instead now to get that one > >> needs to enable both snooping and querier. Enabling querier on all > >> switches could be problematic and is not a good solution, > > > > There is no need to set the querier on all snooping switches. > > br_multicast_querier_exists() checks if a querier exists on the > > link in general, not if this particular host/bridge is a querier. > > > > We need a generic solution for the case of existing mdst and no querier. > More below. > > > > >> for example as summarized by our multicast experts: > >> "every switch would send an IGMP query > > > > What? RFC3810, section 7.1 says: > > > > "If it is the case, a querier election mechanism (described in > > section 7.6.2) is used to elect a single multicast router to be > > in Querier state. [...] Nevertheless, it is only the [elected] Querier > > that sends periodical or triggered query messages on the subnet." > > >> for any random multicast traffic it > >> received across the entire domain and it would send it forever as long as a > >> host exists wanting that stream even if it has no downstream/directly > >> connected receivers" > > > > This was taken out of context and it's my bad, I think everyone is aware > of the election process, please nevermind the above statement. > > [snip]> > > > > Have you done some tests with this change yet, Nikolay? > > > > You've raised good questions, IPv6 indeed needs more work - we'll have to flood > link-local packets etc. but I wanted to have a discussion about no querier/existing mdst. > To simplify we can modify the patch and have traffic forwarded to the proper ports when an > mdst exists and there is no querier for both unsolicited report and user-added entry. > We can keep the current behaviour for unknown traffic with and without querier. > This would align it closer to what other vendors currently do as well IIRC. > What do you think ?The no querier condition is not currently reflected via switchdev, so the behavior you're proposing in your patch is what actually happens in the data plane. We already hit the problem Linus mentioned in commit b00589af3b04 ("bridge: disable snooping if there is no querier"). Namely, IPv6 ND broke because a port joined before the bridge was created. I introduced a workaround in commit 9d45deb04c59 ("mlxsw: spectrum: Treat IPv6 unregistered multicast as broadcast"). I'm interested to know what other vendors are doing. Can you elaborate? We can trap IPv6 ND packets at L2 (we'll eventually need to do for ND suppression) and let the bridge take care of flooding them correctly. I'm not sure it's good enough.
nikolay at cumulusnetworks.com
2019-Feb-16 19:15 UTC
[Bridge] [PATCH RFC] net: bridge: don't flood known multicast traffic when snooping is enabled
On 16 February 2019 20:43:53 EET, Ido Schimmel <idosch at idosch.org> wrote:>On Sat, Feb 16, 2019 at 10:05:40AM +0200, Nikolay Aleksandrov wrote: >> On 15/02/2019 19:13, Linus L?ssing wrote: >> > On Fri, Feb 15, 2019 at 03:04:27PM +0200, Nikolay Aleksandrov >wrote: >> >> Every user would expect to have traffic forwarded only to the >configured >> >> mdb destination when snooping is enabled, instead now to get that >one >> >> needs to enable both snooping and querier. Enabling querier on all >> >> switches could be problematic and is not a good solution, >> > >> > There is no need to set the querier on all snooping switches. >> > br_multicast_querier_exists() checks if a querier exists on the >> > link in general, not if this particular host/bridge is a querier. >> > >> >> We need a generic solution for the case of existing mdst and no >querier. >> More below. >> >> > >> >> for example as summarized by our multicast experts: >> >> "every switch would send an IGMP query >> > >> > What? RFC3810, section 7.1 says: >> > >> > "If it is the case, a querier election mechanism (described in >> > section 7.6.2) is used to elect a single multicast router to be >> > in Querier state. [...] Nevertheless, it is only the [elected] >Querier >> > that sends periodical or triggered query messages on the subnet." >> > >> for any random multicast traffic it >> >> received across the entire domain and it would send it forever as >long as a >> >> host exists wanting that stream even if it has no >downstream/directly >> >> connected receivers" >> > >> >> This was taken out of context and it's my bad, I think everyone is >aware >> of the election process, please nevermind the above statement. >> >> [snip]> >> > >> > Have you done some tests with this change yet, Nikolay? >> > >> >> You've raised good questions, IPv6 indeed needs more work - we'll >have to flood >> link-local packets etc. but I wanted to have a discussion about no >querier/existing mdst. >> To simplify we can modify the patch and have traffic forwarded to the >proper ports when an >> mdst exists and there is no querier for both unsolicited report and >user-added entry. >> We can keep the current behaviour for unknown traffic with and >without querier. >> This would align it closer to what other vendors currently do as well >IIRC. >> What do you think ? > >The no querier condition is not currently reflected via switchdev, so >the behavior you're proposing in your patch is what actually happens in >the data plane. > >We already hit the problem Linus mentioned in commit b00589af3b04 >("bridge: disable snooping if there is no querier"). Namely, IPv6 ND >broke because a port joined before the bridge was created. > >I introduced a workaround in commit 9d45deb04c59 ("mlxsw: spectrum: >Treat IPv6 unregistered multicast as broadcast"). I'm interested to >know >what other vendors are doing. Can you elaborate? >Exactly like your fix. :) Flood it, this patch unfortunately breaks it? because of mrouters flag, but we can retain the behaviour by forwarding only known mdsts to their ports and flooding unregistered mcast when there is no querier. I think that is what others do by default too, actually I think they flood with querier as well. Maybe unknown mcast flooding should be controlled by a flag when a querier is present because we've had this behaviour for a long time, personally I'd have it on by default. I am currently away and will be able to prepare a rfc patch after the weekend.>We can trap IPv6 ND packets at L2 (we'll eventually need to do for ND >suppression) and let the bridge take care of flooding them correctly. >I'm not sure it's good enough.