| Message ID | 20210116012515.3152-3-tobias@waldekranz.com |
|---|---|
| State | New |
| Headers | show |
| Series | net: dsa: Sync local bridge FDB addresses to hardware | expand |
Hi Tobias, On Sat, Jan 16, 2021 at 02:25:10AM +0100, Tobias Waldekranz wrote: > Some switchdev drivers, notably DSA, ignore all dynamically learned > address notifications (!added_by_user) as these are autonomously added > by the switch. Previously, such a notification was indistinguishable > from a local address notification. Include a local bit in the > notification so that the two classes can be discriminated. > > This allows DSA-like devices to add local addresses to the hardware > FDB (with the CPU as the destination), thereby avoiding flows towards > the CPU being flooded by the switch as unknown unicast. > > Signed-off-by: Tobias Waldekranz <tobias@waldekranz.com> > --- In an ideal world, the BR_FDB_LOCAL bit of an FDB entry is what you would probably want to use as an indication that the packet must be delivered upstream by the hardware, considering that this is what the software data path does: br_handle_frame_finish: if (test_bit(BR_FDB_LOCAL, &dst->flags)) return br_pass_frame_up(skb); However, we are not in an ideal world, but in a cacophony of nonsensical flags that must be passed to the 'bridge fdb add' command. For example, I noticed this usage pattern in your patch 6/7: | br0 | / \ | swp0 dummy0 | | $ bridge fdb add 02:00:de:ad:00:01 dev dummy0 vlan 1 master Do you know that this command doesn't do what you think it does (assuming that 02:00:de:ad:00:01 is not the MAC address of dummy0)? The command you wrote will add a _local_ FDB entry on dummy0. I tried it on a DSA switch interface (swp0): $ bridge fdb add 02:00:de:ad:00:01 dev swp0 vlan 1 master [ 3162.165561] rtnl_fdb_add: addr 02:00:de:ad:00:01 vid 1 ndm_state NUD_NOARP|NUD_PERMANENT [ 3162.172487] fdb_add_entry: fdb 02:00:de:ad:00:01 state NUD_NOARP|NUD_PERMANENT, fdb_to_nud NUD_REACHABLE, flags 0x0 [ 3162.180515] mscc_felix 0000:00:00.5 swp0: local fdb: 02:00:de:ad:00:01 vid 1 So, after your patches, this command syntax will stop adding a front-facing FDB entry on swp0. It will add a CPU-facing FDB entry instead. You know why the bridge neighbour state is NUD_NOARP|NUD_PERMANENT in rtnl_fdb_add? Well, because iproute2 set it that way: /* Assume permanent */ if (!(req.ndm.ndm_state&(NUD_PERMANENT|NUD_REACHABLE))) req.ndm.ndm_state |= NUD_PERMANENT; See iproute2 commit 0849e60a10d0 ("bridge: manage VXLAN based forwarding tables"). I know so little about VXLAN's use of the bridge command, that I cannot tell why it was decided to "assume permanent" (which seems to have changed the default behavior for everybody). Otherwise said, even if not mentioned in the man page, the default FDB entry type is NUD_PERMANENT (which, in short, means a "local" entry, see a more detailed explanation at the end). The man page just says: bridge fdb add - add a new fdb entry This command creates a new fdb entry. LLADDR the Ethernet MAC address. dev DEV the interface to which this address is associated. local - is a local permanent fdb entry static - is a static (no arp) fdb entry which is utterly misleading and useless. It does not say: (a) what a local FDB entry is (b) that if neither "local" or "static"|"dynamic" is specified, "local" is default This already creates pretty bad premises. You would have needed to explicitly add "static" to your command. Not only you, but in fact also thousands of other people who already have switchdev deployments using the 'bridge fdb' command. So, in short, if everybody with switchdev hardware used the 'bridge fdb' command correctly so far, your series would have been great. But in fact, nobody did (me included). So we need to be more creative. For example, there's that annoying "self" flag. As far as I understand, there is zero reason for a DSA driver to use the "self" flag, since that means "bypass the bridge FDB and just call the .ndo_fdb_add of the device driver, which in the case of DSA is dsa_legacy_fdb_add". Instead you would just use the "master" flag, which makes the operation be propagated through br_fdb_add and the software FDB has a chance to be updated. Only that there's no one preventing you from using the 'self' and 'master' flags together. Which means that the FDB would be offloaded to the device twice: once through the SWITCHDEV_FDB_ADD_TO_DEVICE notification emitted by br_fdb_add, and once through dsa_legacy_fdb_add. Contradict me if I'm wrong, but I'm thinking that you may have missed this detail that bridge fdb addresses are implicitly 'local' because you also used some 'self master' commands, and the "to-CPU" address installed through switchdev was immediately overwritten by the correct one installed through dsa_legacy_fdb_add. So I think there is a very real issue in that the FDB entries with the is_local bit was never specified to switchdev thus far, and now suddenly is. I'm sorry, but what you're saying in the commit message, that "!added_by_user has so far been indistinguishable from is_local" is simply false. What I'm saying is that some of the is_local addresses should have been rejected by DSA from day one, like this one: bridge fdb add dev swp0 00:01:02:03:04:05 master but nonetheless DSA is happy to accept it anyway, because switchdev doesn't tell it it's local. Yay. It looks like Mellanox have been telling their users to explicitly use the "static" keyword when they mean a static FDB entry: https://github.com/Mellanox/mlxsw/wiki/Bridge#forwarding-database-configuration which, I mean, is great for them, but pretty much sucks for everybody else, because Documentation/networking/switchdev.rst just says: The switchdev driver should implement ndo_fdb_add, ndo_fdb_del and ndo_fdb_dump to support static FDB entries installed to the device. Static bridge FDB entries are installed, for example, using iproute2 bridge cmd:: bridge fdb add ADDR dev DEV [vlan VID] [self] which of course is completely bogus anyway (it indicates 'self', it doesn't indicate 'master' at all, it doesn't say anything about 'static'). Look, I'd be more than happy to accept that I'm the only idiot who misread the existing documentation on 'bridge fdb', but I fear that this is far from true. If we want to make progress with this patch, some user space breakage will be necessary - and I think I'm in favour of doing that. Appendix: The bridge FDB netlink UAPI is modeled as a neighbouring protocol for PF_BRIDGE, similar to what ARP is for IPv4 and ND is for IPv6. There is this forced correlation between generic neighbour states (NUD == "Network Unreachability Detection") and types of FDB entries: - NUD_NOARP: Normally this state is used to denote a neighbour that doesn't need any protocol to resolve the mapping between L3 address and L2 address. It seems to be mostly used when the neigh->dev does not implement header_ops (net device has no L2 header), and therefore no neighbor resolution is needed, because there is no L2 addressing on that device. For PF_BRIDGE, all FDB entries are NUD_NOARP, except for 'dynamic' ones (see below). - NUD_PERMANENT: Normally this state means that the L2 address of the neighbor has been statically configured by the user and therefore there is no need for a neighbour resolution. For PF_BRIDGE though, it means that an FDB entry is 'local', i.e. the L2 address belongs to a local interface. - NUD_REACHABLE: Normally this state means that the L2 address has been resolved by the neighbouring protocol. For PF_BRIDGE, this flag must be interpreted together with NUD_NOARP for full meaning. NUD_REACHABLE=true, NUD_NOARP=true: static FDB entry NUD_REACHABLE=true, NUD_NOARP=false: dynamic FDB entry
On Sun, Jan 17, 2021 at 21:30, Vladimir Oltean <olteanv@gmail.com> wrote: > Hi Tobias, > > On Sat, Jan 16, 2021 at 02:25:10AM +0100, Tobias Waldekranz wrote: >> Some switchdev drivers, notably DSA, ignore all dynamically learned >> address notifications (!added_by_user) as these are autonomously added >> by the switch. Previously, such a notification was indistinguishable >> from a local address notification. Include a local bit in the >> notification so that the two classes can be discriminated. >> >> This allows DSA-like devices to add local addresses to the hardware >> FDB (with the CPU as the destination), thereby avoiding flows towards >> the CPU being flooded by the switch as unknown unicast. >> >> Signed-off-by: Tobias Waldekranz <tobias@waldekranz.com> >> --- > > In an ideal world, the BR_FDB_LOCAL bit of an FDB entry is what you > would probably want to use as an indication that the packet must be > delivered upstream by the hardware, considering that this is what the > software data path does: > > br_handle_frame_finish: > if (test_bit(BR_FDB_LOCAL, &dst->flags)) > return br_pass_frame_up(skb); That was my thinking, yes. > However, we are not in an ideal world, but in a cacophony of nonsensical > flags that must be passed to the 'bridge fdb add' command. For example, > I noticed this usage pattern in your patch 6/7: > > | br0 > | / \ > | swp0 dummy0 > | > | $ bridge fdb add 02:00:de:ad:00:01 dev dummy0 vlan 1 master > > Do you know that this command doesn't do what you think it does > (assuming that 02:00:de:ad:00:01 is not the MAC address of dummy0)? > > The command you wrote will add a _local_ FDB entry on dummy0. > I tried it on a DSA switch interface (swp0): > > $ bridge fdb add 02:00:de:ad:00:01 dev swp0 vlan 1 master > [ 3162.165561] rtnl_fdb_add: addr 02:00:de:ad:00:01 vid 1 ndm_state NUD_NOARP|NUD_PERMANENT > [ 3162.172487] fdb_add_entry: fdb 02:00:de:ad:00:01 state NUD_NOARP|NUD_PERMANENT, fdb_to_nud NUD_REACHABLE, flags 0x0 > [ 3162.180515] mscc_felix 0000:00:00.5 swp0: local fdb: 02:00:de:ad:00:01 vid 1 > > So, after your patches, this command syntax will stop adding a > front-facing FDB entry on swp0. It will add a CPU-facing FDB entry > instead. Ah I see, no I was not aware of that. I just saw that the entry towards the CPU was added to the ATU, which it would in both cases. I.e. from the switch's POV, in this setup: br0 / \ (A) swp0 dummy0 (B) A "local" entry like (A), or a "static" entry like (B) means the same thing to the switch: "it is somewhere behind my CPU-port". > You know why the bridge neighbour state is NUD_NOARP|NUD_PERMANENT in > rtnl_fdb_add? Well, because iproute2 set it that way: > > /* Assume permanent */ > if (!(req.ndm.ndm_state&(NUD_PERMANENT|NUD_REACHABLE))) > req.ndm.ndm_state |= NUD_PERMANENT; > > See iproute2 commit 0849e60a10d0 ("bridge: manage VXLAN based forwarding > tables"). I know so little about VXLAN's use of the bridge command, that > I cannot tell why it was decided to "assume permanent" (which seems to > have changed the default behavior for everybody). > > Otherwise said, even if not mentioned in the man page, the default FDB > entry type is NUD_PERMANENT (which, in short, means a "local" entry, see > a more detailed explanation at the end). > > The man page just says: > > bridge fdb add - add a new fdb entry > This command creates a new fdb entry. > > LLADDR the Ethernet MAC address. > > dev DEV > the interface to which this address is associated. > > local - is a local permanent fdb entry > > static - is a static (no arp) fdb entry > > which is utterly misleading and useless. It does not say: > (a) what a local FDB entry is > (b) that if neither "local" or "static"|"dynamic" is specified, > "local" is default > > This already creates pretty bad premises. You would have needed to > explicitly add "static" to your command. Not only you, but in fact also > thousands of other people who already have switchdev deployments using > the 'bridge fdb' command. > > So, in short, if everybody with switchdev hardware used the 'bridge fdb' > command correctly so far, your series would have been great. But in > fact, nobody did (me included). So we need to be more creative. > > For example, there's that annoying "self" flag. > As far as I understand, there is zero reason for a DSA driver to use the > "self" flag, since that means "bypass the bridge FDB and just call the > .ndo_fdb_add of the device driver, which in the case of DSA is > dsa_legacy_fdb_add". Instead you would just use the "master" flag, which > makes the operation be propagated through br_fdb_add and the software > FDB has a chance to be updated. > > Only that there's no one preventing you from using the 'self' and > 'master' flags together. Which means that the FDB would be offloaded to > the device twice: once through the SWITCHDEV_FDB_ADD_TO_DEVICE > notification emitted by br_fdb_add, and once through dsa_legacy_fdb_add. > Contradict me if I'm wrong, but I'm thinking that you may have missed > this detail that bridge fdb addresses are implicitly 'local' because you > also used some 'self master' commands, and the "to-CPU" address > installed through switchdev was immediately overwritten by the correct > one installed through dsa_legacy_fdb_add. > > So I think there is a very real issue in that the FDB entries with the > is_local bit was never specified to switchdev thus far, and now suddenly > is. I'm sorry, but what you're saying in the commit message, that > "!added_by_user has so far been indistinguishable from is_local" is > simply false. Alright, so how do you do it? Here is the struct: struct switchdev_notifier_fdb_info { struct switchdev_notifier_info info; /* must be first */ const unsigned char *addr; u16 vid; u8 added_by_user:1, offloaded:1; }; Which field separates a local address on swp0 from a dynamically learned address on swp0? > What I'm saying is that some of the is_local addresses should have been > rejected by DSA from day one, like this one: > > bridge fdb add dev swp0 00:01:02:03:04:05 master > > but nonetheless DSA is happy to accept it anyway, because switchdev > doesn't tell it it's local. Yay. Yes that is a real problem, for sure. > It looks like Mellanox have been telling their users to explicitly use > the "static" keyword when they mean a static FDB entry: > https://github.com/Mellanox/mlxsw/wiki/Bridge#forwarding-database-configuration > which, I mean, is great for them, but pretty much sucks for everybody > else, because Documentation/networking/switchdev.rst just says: > > The switchdev driver should implement ndo_fdb_add, ndo_fdb_del and ndo_fdb_dump > to support static FDB entries installed to the device. Static bridge FDB > entries are installed, for example, using iproute2 bridge cmd:: > > bridge fdb add ADDR dev DEV [vlan VID] [self] > > which of course is completely bogus anyway (it indicates 'self', it > doesn't indicate 'master' at all, it doesn't say anything about 'static'). > > Look, I'd be more than happy to accept that I'm the only idiot who > misread the existing documentation on 'bridge fdb', but I fear that this > is far from true. If we want to make progress with this patch, some user > space breakage will be necessary - and I think I'm in favour of doing > that. Trust me, you are not. There is a running joke about being able to describe what "master" and "self" really means here at the office :) Ok, so just to see if I understand this correctly: The situation today it that `bridge fdb add ADDR dev DEV master` results in flows towards ADDR being sent to: 1. DEV if DEV belongs to a DSA switch. 2. To the host if DEV was a non-offloaded interface. With this series applied both would result in (2) which, while idiosyncratic, is as intended. But this of course runs the risk of breaking existing scripts which rely on the current behavior. Correct?
On Mon, Jan 18, 2021 at 07:58:59PM +0100, Tobias Waldekranz wrote: > Ah I see, no I was not aware of that. I just saw that the entry towards > the CPU was added to the ATU, which it would in both cases. I.e. from > the switch's POV, in this setup: > > br0 > / \ (A) > swp0 dummy0 > (B) > > A "local" entry like (A), or a "static" entry like (B) means the same > thing to the switch: "it is somewhere behind my CPU-port". Yes, except that if dummy0 was a real and non-switchdev interface, then the "local" entry would probably break your traffic if what you meant was "static". > > So I think there is a very real issue in that the FDB entries with the > > is_local bit was never specified to switchdev thus far, and now suddenly > > is. I'm sorry, but what you're saying in the commit message, that > > "!added_by_user has so far been indistinguishable from is_local" is > > simply false. > > Alright, so how do you do it? Here is the struct: > > struct switchdev_notifier_fdb_info { > struct switchdev_notifier_info info; /* must be first */ > const unsigned char *addr; > u16 vid; > u8 added_by_user:1, > offloaded:1; > }; > > Which field separates a local address on swp0 from a dynamically learned > address on swp0? None, that's the problem. Local addresses are already presented to switchdev without saying that they're local. Which is the entire reason that users are misled into thinking that the addresses are not local. I may have misread what you said, but to me, "!added_by_user has so far been indistinguishable from is_local" means that: - every struct switchdev_notifier_fdb_info with added_by_user == true also had an implicit is_local == false - every struct switchdev_notifier_fdb_info with added_by_user == false also had an implicit is_local == true It is _this_ that I deemed as clearly untrue. The is_local flag is not indistinguishable from !added_by_user, it is indistinguishable full stop. Which makes it hard to work with in a backwards-compatible way. > Ok, so just to see if I understand this correctly: > > The situation today it that `bridge fdb add ADDR dev DEV master` results > in flows towards ADDR being sent to: > > 1. DEV if DEV belongs to a DSA switch. > 2. To the host if DEV was a non-offloaded interface. Not quite. In the bridge software FDB, the entry is marked as is_local in both cases, doesn't matter if the interface is offloaded or not. Just that switchdev does not propagate the is_local flag, which makes the switchdev listeners think it is not local. The interpretation of that will probably vary among switchdev drivers. The subtlety is that for a non-offloading interface, the misconfiguration (when you mean static but use local) is easy to catch. Since only the entry from the software FDB will be hit, this means that the frame will never be forwarded, so traffic will break. But in the case of a switchdev offloading interface, the frames will hit the hardware FDB entry more often than the software FDB entry. So everything will work just fine and dandy even though it shouldn't. > With this series applied both would result in (2) which, while > idiosyncratic, is as intended. But this of course runs the risk of > breaking existing scripts which rely on the current behavior. Yes. My only hope is that we could just offload the entries pointing towards br0, and ignore the local ones. But for that I would need the bridge maintainers to clarify what is the difference between then, as I asked in your other patch.
On Sun, Jan 17, 2021 at 09:30:09PM +0200, Vladimir Oltean wrote: > Hi Tobias, > > On Sat, Jan 16, 2021 at 02:25:10AM +0100, Tobias Waldekranz wrote: > > Some switchdev drivers, notably DSA, ignore all dynamically learned > > address notifications (!added_by_user) as these are autonomously added > > by the switch. Previously, such a notification was indistinguishable > > from a local address notification. Include a local bit in the > > notification so that the two classes can be discriminated. > > > > This allows DSA-like devices to add local addresses to the hardware > > FDB (with the CPU as the destination), thereby avoiding flows towards > > the CPU being flooded by the switch as unknown unicast. > > > > Signed-off-by: Tobias Waldekranz <tobias@waldekranz.com> > > --- > > In an ideal world, the BR_FDB_LOCAL bit of an FDB entry is what you > would probably want to use as an indication that the packet must be > delivered upstream by the hardware, considering that this is what the > software data path does: > > br_handle_frame_finish: > if (test_bit(BR_FDB_LOCAL, &dst->flags)) > return br_pass_frame_up(skb); > > However, we are not in an ideal world, but in a cacophony of nonsensical > flags that must be passed to the 'bridge fdb add' command. For example, > I noticed this usage pattern in your patch 6/7: Thanks for adding me. Reflecting FDB flags is a very much needed change. I will take a look tomorrow or the day after.
On Mon, Jan 18, 2021 at 21:27, Vladimir Oltean <olteanv@gmail.com> wrote: > On Mon, Jan 18, 2021 at 07:58:59PM +0100, Tobias Waldekranz wrote: >> Ah I see, no I was not aware of that. I just saw that the entry towards >> the CPU was added to the ATU, which it would in both cases. I.e. from >> the switch's POV, in this setup: >> >> br0 >> / \ (A) >> swp0 dummy0 >> (B) >> >> A "local" entry like (A), or a "static" entry like (B) means the same >> thing to the switch: "it is somewhere behind my CPU-port". > > Yes, except that if dummy0 was a real and non-switchdev interface, then > the "local" entry would probably break your traffic if what you meant > was "static". Agreed. >> > So I think there is a very real issue in that the FDB entries with the >> > is_local bit was never specified to switchdev thus far, and now suddenly >> > is. I'm sorry, but what you're saying in the commit message, that >> > "!added_by_user has so far been indistinguishable from is_local" is >> > simply false. >> >> Alright, so how do you do it? Here is the struct: >> >> struct switchdev_notifier_fdb_info { >> struct switchdev_notifier_info info; /* must be first */ >> const unsigned char *addr; >> u16 vid; >> u8 added_by_user:1, >> offloaded:1; >> }; >> >> Which field separates a local address on swp0 from a dynamically learned >> address on swp0? > > None, that's the problem. Local addresses are already presented to > switchdev without saying that they're local. Which is the entire reason > that users are misled into thinking that the addresses are not local. > > I may have misread what you said, but to me, "!added_by_user has so far > been indistinguishable from is_local" means that: > - every struct switchdev_notifier_fdb_info with added_by_user == true > also had an implicit is_local == false > - every struct switchdev_notifier_fdb_info with added_by_user == false > also had an implicit is_local == true > It is _this_ that I deemed as clearly untrue. > > The is_local flag is not indistinguishable from !added_by_user, it is > indistinguishable full stop. Which makes it hard to work with in a > backwards-compatible way. This was probably a semantic mistake on my part, we meant the same thing. >> Ok, so just to see if I understand this correctly: >> >> The situation today it that `bridge fdb add ADDR dev DEV master` results >> in flows towards ADDR being sent to: >> >> 1. DEV if DEV belongs to a DSA switch. >> 2. To the host if DEV was a non-offloaded interface. > > Not quite. In the bridge software FDB, the entry is marked as is_local > in both cases, doesn't matter if the interface is offloaded or not. > Just that switchdev does not propagate the is_local flag, which makes > the switchdev listeners think it is not local. The interpretation of > that will probably vary among switchdev drivers. > > The subtlety is that for a non-offloading interface, the > misconfiguration (when you mean static but use local) is easy to catch. > Since only the entry from the software FDB will be hit, this means that > the frame will never be forwarded, so traffic will break. > But in the case of a switchdev offloading interface, the frames will hit > the hardware FDB entry more often than the software FDB entry. So > everything will work just fine and dandy even though it shouldn't. Quite right. >> With this series applied both would result in (2) which, while >> idiosyncratic, is as intended. But this of course runs the risk of >> breaking existing scripts which rely on the current behavior. > > Yes. > > My only hope is that we could just offload the entries pointing towards > br0, and ignore the local ones. But for that I would need the bridge That was my initial approach. Unfortunately that breaks down when the bridge inherits its address from a port, i.e. the default case. When the address is added to the bridge (fdb->dst == NULL), fdb_insert will find the previous local entry that is set on the port and bail out before sending a notification: if (fdb) { /* it is okay to have multiple ports with same * address, just use the first one. */ if (test_bit(BR_FDB_LOCAL, &fdb->flags)) return 0; br_warn(br, "adding interface %s with same address as a received packet (addr:%pM, vlan:%u)\n", source ? source->dev->name : br->dev->name, addr, vid); fdb_delete(br, fdb, true); } You could change this so that a notification always is sent out. Or you could give precedence to !fdb->dst and update the existing entry. > maintainers to clarify what is the difference between then, as I asked > in your other patch. I am pretty sure they mean the same thing, I believe that !fdb->dst implies is_local. It is just that "bridge fdb add ADDR dev br0 self" is a new(er) thing, and before that there was "local" entries on ports. Maybe I should try to get rid of the local flag in the bridge first, and then come back to this problem once that is done? Either way, I agree that 5/7 is all we want to add to DSA to get this working.
On Mon, Jan 18, 2021 at 09:19:11PM +0100, Tobias Waldekranz wrote: > > My only hope is that we could just offload the entries pointing towards > > br0, and ignore the local ones. > > That was my initial approach. Unfortunately that breaks down when the > bridge inherits its address from a port, i.e. the default case. > > When the address is added to the bridge (fdb->dst == NULL), fdb_insert > will find the previous local entry that is set on the port and bail out > before sending a notification: > > if (fdb) { > /* it is okay to have multiple ports with same > * address, just use the first one. > */ > if (test_bit(BR_FDB_LOCAL, &fdb->flags)) > return 0; > br_warn(br, "adding interface %s with same address as a received packet (addr:%pM, vlan:%u)\n", > source ? source->dev->name : br->dev->name, addr, vid); > fdb_delete(br, fdb, true); > } > > You could change this so that a notification always is sent out. Or you > could give precedence to !fdb->dst and update the existing entry. I'm afraid my competence ends here. IMO the problem is really the struct net_bridge_port *source argument of fdb_insert. The behavior we want is that all is_local FDB entries are coming from br0, and none from the brports (aka source == NULL, so the callers that had something non-NULL for source should be deleted). "You can't always get what you want" though. > > But for that I would need the bridge maintainers to clarify what is > > the difference between then, as I asked in your other patch. > > I am pretty sure they mean the same thing, I believe that !fdb->dst > implies is_local. It is just that "bridge fdb add ADDR dev br0 self" is > a new(er) thing, and before that there was "local" entries on ports. > Maybe I should try to get rid of the local flag in the bridge first, and > then come back to this problem once that is done? Either way, I agree > that 5/7 is all we want to add to DSA to get this working. Please expand on what you plan to do. The is_local bit is part of the bridge UAPI, how do you plan to get rid of it?
On 18/01/2021 22:19, Tobias Waldekranz wrote: > On Mon, Jan 18, 2021 at 21:27, Vladimir Oltean <olteanv@gmail.com> wrote: >> On Mon, Jan 18, 2021 at 07:58:59PM +0100, Tobias Waldekranz wrote: >>> Ah I see, no I was not aware of that. I just saw that the entry towards >>> the CPU was added to the ATU, which it would in both cases. I.e. from >>> the switch's POV, in this setup: >>> >>> br0 >>> / \ (A) >>> swp0 dummy0 >>> (B) >>> >>> A "local" entry like (A), or a "static" entry like (B) means the same >>> thing to the switch: "it is somewhere behind my CPU-port". >> >> Yes, except that if dummy0 was a real and non-switchdev interface, then >> the "local" entry would probably break your traffic if what you meant >> was "static". > > Agreed. > >>>> So I think there is a very real issue in that the FDB entries with the >>>> is_local bit was never specified to switchdev thus far, and now suddenly >>>> is. I'm sorry, but what you're saying in the commit message, that >>>> "!added_by_user has so far been indistinguishable from is_local" is >>>> simply false. >>> >>> Alright, so how do you do it? Here is the struct: >>> >>> struct switchdev_notifier_fdb_info { >>> struct switchdev_notifier_info info; /* must be first */ >>> const unsigned char *addr; >>> u16 vid; >>> u8 added_by_user:1, >>> offloaded:1; >>> }; >>> >>> Which field separates a local address on swp0 from a dynamically learned >>> address on swp0? >> >> None, that's the problem. Local addresses are already presented to >> switchdev without saying that they're local. Which is the entire reason >> that users are misled into thinking that the addresses are not local. >> >> I may have misread what you said, but to me, "!added_by_user has so far >> been indistinguishable from is_local" means that: >> - every struct switchdev_notifier_fdb_info with added_by_user == true >> also had an implicit is_local == false >> - every struct switchdev_notifier_fdb_info with added_by_user == false >> also had an implicit is_local == true >> It is _this_ that I deemed as clearly untrue. >> >> The is_local flag is not indistinguishable from !added_by_user, it is >> indistinguishable full stop. Which makes it hard to work with in a >> backwards-compatible way. > > This was probably a semantic mistake on my part, we meant the same > thing. > >>> Ok, so just to see if I understand this correctly: >>> >>> The situation today it that `bridge fdb add ADDR dev DEV master` results >>> in flows towards ADDR being sent to: >>> >>> 1. DEV if DEV belongs to a DSA switch. >>> 2. To the host if DEV was a non-offloaded interface. >> >> Not quite. In the bridge software FDB, the entry is marked as is_local >> in both cases, doesn't matter if the interface is offloaded or not. >> Just that switchdev does not propagate the is_local flag, which makes >> the switchdev listeners think it is not local. The interpretation of >> that will probably vary among switchdev drivers. >> >> The subtlety is that for a non-offloading interface, the >> misconfiguration (when you mean static but use local) is easy to catch. >> Since only the entry from the software FDB will be hit, this means that >> the frame will never be forwarded, so traffic will break. >> But in the case of a switchdev offloading interface, the frames will hit >> the hardware FDB entry more often than the software FDB entry. So >> everything will work just fine and dandy even though it shouldn't. > > Quite right. > >>> With this series applied both would result in (2) which, while >>> idiosyncratic, is as intended. But this of course runs the risk of >>> breaking existing scripts which rely on the current behavior. >> >> Yes. >> >> My only hope is that we could just offload the entries pointing towards >> br0, and ignore the local ones. But for that I would need the bridge > > That was my initial approach. Unfortunately that breaks down when the > bridge inherits its address from a port, i.e. the default case. > > When the address is added to the bridge (fdb->dst == NULL), fdb_insert > will find the previous local entry that is set on the port and bail out > before sending a notification: > > if (fdb) { > /* it is okay to have multiple ports with same > * address, just use the first one. > */ > if (test_bit(BR_FDB_LOCAL, &fdb->flags)) > return 0; > br_warn(br, "adding interface %s with same address as a received packet (addr:%pM, vlan:%u)\n", > source ? source->dev->name : br->dev->name, addr, vid); > fdb_delete(br, fdb, true); > } > > You could change this so that a notification always is sent out. Or you > could give precedence to !fdb->dst and update the existing entry. > >> maintainers to clarify what is the difference between then, as I asked >> in your other patch. > > I am pretty sure they mean the same thing, I believe that !fdb->dst > implies is_local. It is just that "bridge fdb add ADDR dev br0 self" is > a new(er) thing, and before that there was "local" entries on ports. > > Maybe I should try to get rid of the local flag in the bridge first, and > then come back to this problem once that is done? Either way, I agree > that 5/7 is all we want to add to DSA to get this working. > BR_FDB_LOCAL and !fdb->dst are not the same thing, check fdb_add_entry(). You cannot get rid of it, !fdb->dst implies BR_FDB_LOCAL, but it's not symmetrical. Cheers, Nik
On 18/01/2021 23:17, Nikolay Aleksandrov wrote: > On 18/01/2021 22:19, Tobias Waldekranz wrote: >> On Mon, Jan 18, 2021 at 21:27, Vladimir Oltean <olteanv@gmail.com> wrote: >>> On Mon, Jan 18, 2021 at 07:58:59PM +0100, Tobias Waldekranz wrote: >>>> Ah I see, no I was not aware of that. I just saw that the entry towards >>>> the CPU was added to the ATU, which it would in both cases. I.e. from >>>> the switch's POV, in this setup: >>>> >>>> br0 >>>> / \ (A) >>>> swp0 dummy0 >>>> (B) >>>> >>>> A "local" entry like (A), or a "static" entry like (B) means the same >>>> thing to the switch: "it is somewhere behind my CPU-port". >>> >>> Yes, except that if dummy0 was a real and non-switchdev interface, then >>> the "local" entry would probably break your traffic if what you meant >>> was "static". >> >> Agreed. >> >>>>> So I think there is a very real issue in that the FDB entries with the >>>>> is_local bit was never specified to switchdev thus far, and now suddenly >>>>> is. I'm sorry, but what you're saying in the commit message, that >>>>> "!added_by_user has so far been indistinguishable from is_local" is >>>>> simply false. >>>> >>>> Alright, so how do you do it? Here is the struct: >>>> >>>> struct switchdev_notifier_fdb_info { >>>> struct switchdev_notifier_info info; /* must be first */ >>>> const unsigned char *addr; >>>> u16 vid; >>>> u8 added_by_user:1, >>>> offloaded:1; >>>> }; >>>> >>>> Which field separates a local address on swp0 from a dynamically learned >>>> address on swp0? >>> >>> None, that's the problem. Local addresses are already presented to >>> switchdev without saying that they're local. Which is the entire reason >>> that users are misled into thinking that the addresses are not local. >>> >>> I may have misread what you said, but to me, "!added_by_user has so far >>> been indistinguishable from is_local" means that: >>> - every struct switchdev_notifier_fdb_info with added_by_user == true >>> also had an implicit is_local == false >>> - every struct switchdev_notifier_fdb_info with added_by_user == false >>> also had an implicit is_local == true >>> It is _this_ that I deemed as clearly untrue. >>> >>> The is_local flag is not indistinguishable from !added_by_user, it is >>> indistinguishable full stop. Which makes it hard to work with in a >>> backwards-compatible way. >> >> This was probably a semantic mistake on my part, we meant the same >> thing. >> >>>> Ok, so just to see if I understand this correctly: >>>> >>>> The situation today it that `bridge fdb add ADDR dev DEV master` results >>>> in flows towards ADDR being sent to: >>>> >>>> 1. DEV if DEV belongs to a DSA switch. >>>> 2. To the host if DEV was a non-offloaded interface. >>> >>> Not quite. In the bridge software FDB, the entry is marked as is_local >>> in both cases, doesn't matter if the interface is offloaded or not. >>> Just that switchdev does not propagate the is_local flag, which makes >>> the switchdev listeners think it is not local. The interpretation of >>> that will probably vary among switchdev drivers. >>> >>> The subtlety is that for a non-offloading interface, the >>> misconfiguration (when you mean static but use local) is easy to catch. >>> Since only the entry from the software FDB will be hit, this means that >>> the frame will never be forwarded, so traffic will break. >>> But in the case of a switchdev offloading interface, the frames will hit >>> the hardware FDB entry more often than the software FDB entry. So >>> everything will work just fine and dandy even though it shouldn't. >> >> Quite right. >> >>>> With this series applied both would result in (2) which, while >>>> idiosyncratic, is as intended. But this of course runs the risk of >>>> breaking existing scripts which rely on the current behavior. >>> >>> Yes. >>> >>> My only hope is that we could just offload the entries pointing towards >>> br0, and ignore the local ones. But for that I would need the bridge >> >> That was my initial approach. Unfortunately that breaks down when the >> bridge inherits its address from a port, i.e. the default case. >> >> When the address is added to the bridge (fdb->dst == NULL), fdb_insert >> will find the previous local entry that is set on the port and bail out >> before sending a notification: >> >> if (fdb) { >> /* it is okay to have multiple ports with same >> * address, just use the first one. >> */ >> if (test_bit(BR_FDB_LOCAL, &fdb->flags)) >> return 0; >> br_warn(br, "adding interface %s with same address as a received packet (addr:%pM, vlan:%u)\n", >> source ? source->dev->name : br->dev->name, addr, vid); >> fdb_delete(br, fdb, true); >> } >> >> You could change this so that a notification always is sent out. Or you >> could give precedence to !fdb->dst and update the existing entry. >> >>> maintainers to clarify what is the difference between then, as I asked >>> in your other patch. >> >> I am pretty sure they mean the same thing, I believe that !fdb->dst >> implies is_local. It is just that "bridge fdb add ADDR dev br0 self" is >> a new(er) thing, and before that there was "local" entries on ports. >> >> Maybe I should try to get rid of the local flag in the bridge first, and >> then come back to this problem once that is done? Either way, I agree >> that 5/7 is all we want to add to DSA to get this working. >> > > BR_FDB_LOCAL and !fdb->dst are not the same thing, check fdb_add_entry(). > You cannot get rid of it, !fdb->dst implies BR_FDB_LOCAL, but it's not > symmetrical. > Scratch that, I spoke too soon. You can get rid of it internally, just need to be careful not to break user-visible behaviour as Vladimir mentioned. > Cheers, > Nik >
On 18/01/2021 23:22, Nikolay Aleksandrov wrote: > On 18/01/2021 23:17, Nikolay Aleksandrov wrote: >> On 18/01/2021 22:19, Tobias Waldekranz wrote: >>> On Mon, Jan 18, 2021 at 21:27, Vladimir Oltean <olteanv@gmail.com> wrote: >>>> On Mon, Jan 18, 2021 at 07:58:59PM +0100, Tobias Waldekranz wrote: >>>>> Ah I see, no I was not aware of that. I just saw that the entry towards >>>>> the CPU was added to the ATU, which it would in both cases. I.e. from >>>>> the switch's POV, in this setup: >>>>> >>>>> br0 >>>>> / \ (A) >>>>> swp0 dummy0 >>>>> (B) >>>>> >>>>> A "local" entry like (A), or a "static" entry like (B) means the same >>>>> thing to the switch: "it is somewhere behind my CPU-port". >>>> >>>> Yes, except that if dummy0 was a real and non-switchdev interface, then >>>> the "local" entry would probably break your traffic if what you meant >>>> was "static". >>> >>> Agreed. >>> >>>>>> So I think there is a very real issue in that the FDB entries with the >>>>>> is_local bit was never specified to switchdev thus far, and now suddenly >>>>>> is. I'm sorry, but what you're saying in the commit message, that >>>>>> "!added_by_user has so far been indistinguishable from is_local" is >>>>>> simply false. >>>>> >>>>> Alright, so how do you do it? Here is the struct: >>>>> >>>>> struct switchdev_notifier_fdb_info { >>>>> struct switchdev_notifier_info info; /* must be first */ >>>>> const unsigned char *addr; >>>>> u16 vid; >>>>> u8 added_by_user:1, >>>>> offloaded:1; >>>>> }; >>>>> >>>>> Which field separates a local address on swp0 from a dynamically learned >>>>> address on swp0? >>>> >>>> None, that's the problem. Local addresses are already presented to >>>> switchdev without saying that they're local. Which is the entire reason >>>> that users are misled into thinking that the addresses are not local. >>>> >>>> I may have misread what you said, but to me, "!added_by_user has so far >>>> been indistinguishable from is_local" means that: >>>> - every struct switchdev_notifier_fdb_info with added_by_user == true >>>> also had an implicit is_local == false >>>> - every struct switchdev_notifier_fdb_info with added_by_user == false >>>> also had an implicit is_local == true >>>> It is _this_ that I deemed as clearly untrue. >>>> >>>> The is_local flag is not indistinguishable from !added_by_user, it is >>>> indistinguishable full stop. Which makes it hard to work with in a >>>> backwards-compatible way. >>> >>> This was probably a semantic mistake on my part, we meant the same >>> thing. >>> >>>>> Ok, so just to see if I understand this correctly: >>>>> >>>>> The situation today it that `bridge fdb add ADDR dev DEV master` results >>>>> in flows towards ADDR being sent to: >>>>> >>>>> 1. DEV if DEV belongs to a DSA switch. >>>>> 2. To the host if DEV was a non-offloaded interface. >>>> >>>> Not quite. In the bridge software FDB, the entry is marked as is_local >>>> in both cases, doesn't matter if the interface is offloaded or not. >>>> Just that switchdev does not propagate the is_local flag, which makes >>>> the switchdev listeners think it is not local. The interpretation of >>>> that will probably vary among switchdev drivers. >>>> >>>> The subtlety is that for a non-offloading interface, the >>>> misconfiguration (when you mean static but use local) is easy to catch. >>>> Since only the entry from the software FDB will be hit, this means that >>>> the frame will never be forwarded, so traffic will break. >>>> But in the case of a switchdev offloading interface, the frames will hit >>>> the hardware FDB entry more often than the software FDB entry. So >>>> everything will work just fine and dandy even though it shouldn't. >>> >>> Quite right. >>> >>>>> With this series applied both would result in (2) which, while >>>>> idiosyncratic, is as intended. But this of course runs the risk of >>>>> breaking existing scripts which rely on the current behavior. >>>> >>>> Yes. >>>> >>>> My only hope is that we could just offload the entries pointing towards >>>> br0, and ignore the local ones. But for that I would need the bridge >>> >>> That was my initial approach. Unfortunately that breaks down when the >>> bridge inherits its address from a port, i.e. the default case. >>> >>> When the address is added to the bridge (fdb->dst == NULL), fdb_insert >>> will find the previous local entry that is set on the port and bail out >>> before sending a notification: >>> >>> if (fdb) { >>> /* it is okay to have multiple ports with same >>> * address, just use the first one. >>> */ >>> if (test_bit(BR_FDB_LOCAL, &fdb->flags)) >>> return 0; >>> br_warn(br, "adding interface %s with same address as a received packet (addr:%pM, vlan:%u)\n", >>> source ? source->dev->name : br->dev->name, addr, vid); >>> fdb_delete(br, fdb, true); >>> } >>> >>> You could change this so that a notification always is sent out. Or you >>> could give precedence to !fdb->dst and update the existing entry. >>> >>>> maintainers to clarify what is the difference between then, as I asked >>>> in your other patch. >>> >>> I am pretty sure they mean the same thing, I believe that !fdb->dst >>> implies is_local. It is just that "bridge fdb add ADDR dev br0 self" is >>> a new(er) thing, and before that there was "local" entries on ports. >>> >>> Maybe I should try to get rid of the local flag in the bridge first, and >>> then come back to this problem once that is done? Either way, I agree >>> that 5/7 is all we want to add to DSA to get this working. >>> >> >> BR_FDB_LOCAL and !fdb->dst are not the same thing, check fdb_add_entry(). >> You cannot get rid of it, !fdb->dst implies BR_FDB_LOCAL, but it's not >> symmetrical. >> > > Scratch that, I spoke too soon. You can get rid of it internally, just need > to be careful not to break user-visible behaviour as Vladimir mentioned. > DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=nvidia.com; s=n1; t=1611005977; bh=bdfYBvMa8LNyHkRyEtaHStOZr794nuxZw02BF6Zfg5c=; h=ARC-Seal:ARC-Message-Signature:ARC-Authentication-Results: Authentication-Results:Subject:From:To:CC:References:Message-ID: Date:User-Agent:In-Reply-To:Content-Type:Content-Language: Content-Transfer-Encoding:X-Originating-IP:X-ClientProxiedBy: MIME-Version:X-MS-Exchange-MessageSentRepresentingType: X-MS-PublicTrafficType:X-MS-Office365-Filtering-Correlation-Id: X-MS-TrafficTypeDiagnostic:X-MS-Exchange-Transport-Forked: X-Microsoft-Antispam-PRVS:X-MS-Oob-TLC-OOBClassifiers: X-MS-Exchange-SenderADCheck:X-Microsoft-Antispam: X-Microsoft-Antispam-Message-Info:X-Forefront-Antispam-Report: X-MS-Exchange-AntiSpam-MessageData: X-MS-Exchange-CrossTenant-Network-Message-Id: X-MS-Exchange-CrossTenant-AuthSource: X-MS-Exchange-CrossTenant-AuthAs: X-MS-Exchange-CrossTenant-OriginalArrivalTime: X-MS-Exchange-CrossTenant-FromEntityHeader: X-MS-Exchange-CrossTenant-Id:X-MS-Exchange-CrossTenant-MailboxType: X-MS-Exchange-CrossTenant-UserPrincipalName: X-MS-Exchange-Transport-CrossTenantHeadersStamped:X-OriginatorOrg; b=lOH2ClNGfg5hhcLMJSlOtsM9K7JoVYkrv1v5FBtgFxxiiWBcOGYir2uqdIfXvMTD/ LUTYZ9tVzvZJ/tKymoPhlR+V27URN1nOwkEdz3k1u46QcB+3eQa1blAz8c8bU/HMAP joO4AKJ9BIuG/sc3ZTAX7jdOE3JUSOmCdfhqTCNc6sGmaVFBAwPrrhGPVth3niCkc7 JwfTXir+8JtBC0XV3Vw2DiYs8RCX22S/48evhzu6O3PNsmLTFaOZaDb0Ep76MquFPu rjCjXH2ZfG9J/D9YchY/hybtGMRK4aruos1La9mEVi6WzUeW+PhR0/FiXzfW/6fef7 cswqoYtddosLw== Apologies for the multiple emails, but wanted to leave an example: 00:11:22:33:44:55 dev ens16 master bridge permanent This must always exist and user-space must be able to create it, which might be against what you want to achieve (no BR_FDB_LOCAL entries with fdb->dst != NULL).
On Mon, Jan 18, 2021 at 11:39:27PM +0200, Nikolay Aleksandrov wrote: > Apologies for the multiple emails, but wanted to leave an example: > > 00:11:22:33:44:55 dev ens16 master bridge permanent > > This must always exist and user-space must be able to create it, which > might be against what you want to achieve (no BR_FDB_LOCAL entries with > fdb->dst != NULL). Can you give me an example of why it would matter that fdb->dst in this case is set to ens16?
On 18/01/2021 23:50, Vladimir Oltean wrote: > On Mon, Jan 18, 2021 at 11:39:27PM +0200, Nikolay Aleksandrov wrote: >> Apologies for the multiple emails, but wanted to leave an example: >> >> 00:11:22:33:44:55 dev ens16 master bridge permanent >> >> This must always exist and user-space must be able to create it, which >> might be against what you want to achieve (no BR_FDB_LOCAL entries with >> fdb->dst != NULL). > > Can you give me an example of why it would matter that fdb->dst in this > case is set to ens16? > Can you dump it as "dev ens16" without it? :) Or alternatively can you send a notification with "dev ens16" without it? I'm in favor of removing it, but it is risky since some script somewhere might be searching for it, or some user-space daemon might expect to see a notification for such entries and react on it.
On Mon, Jan 18, 2021 at 11:53:18PM +0200, Nikolay Aleksandrov wrote: > On 18/01/2021 23:50, Vladimir Oltean wrote: > > On Mon, Jan 18, 2021 at 11:39:27PM +0200, Nikolay Aleksandrov wrote: > >> Apologies for the multiple emails, but wanted to leave an example: > >> > >> 00:11:22:33:44:55 dev ens16 master bridge permanent > >> > >> This must always exist and user-space must be able to create it, which > >> might be against what you want to achieve (no BR_FDB_LOCAL entries with > >> fdb->dst != NULL). > > > > Can you give me an example of why it would matter that fdb->dst in this > > case is set to ens16? > > > > Can you dump it as "dev ens16" without it? :) > Or alternatively can you send a notification with "dev ens16" without it? > > I'm in favor of removing it, but it is risky since some script somewhere might > be searching for it, or some user-space daemon might expect to see a notification > for such entries and react on it. If "dev ens16" makes no difference to the forwarding and/or termination path of the bridge, just to user space reporting, then keeping appearances is a bit pointless. For example, DSA switch interfaces inherit by default the MAC address of the host interface. Having multiple net devices with the same MAC address works because either they are in different L2 domains (case in which the MAC addresses should still be unique per domain), or they are in the same L2 domain, under the same bridge (case in which it is the bridge who should do IP neighbour resolution etc). Having that said, let there be these commands: $ ip link add br0 type bridge $ ip link set swp0 master br0 $ ip link set swp1 master br0 $ ip link set swp2 master br0 $ ip link set swp3 master br0 $ bridge fdb | grep permanent 00:04:9f:05:de:0a dev swp0 vlan 1 master br0 permanent 00:04:9f:05:de:0a dev swp0 master br0 permanent And these: $ ip link add br0 type bridge $ ip link set swp3 master br0 $ ip link set swp2 master br0 $ ip link set swp1 master br0 $ ip link set swp0 master br0 $ bridge fdb | grep permanent 00:04:9f:05:de:0a dev swp0 vlan 1 master br0 permanent 00:04:9f:05:de:0a dev swp0 master br0 permanent 00:04:9f:05:de:0a dev swp3 vlan 1 master br0 permanent 00:04:9f:05:de:0a dev swp3 master br0 permanent Preserving the reporting for permanent/local FDB entries added by user is one thing. But do we need to also preserve this behavior (i.e. report the first unique MAC address of an interface that joins the bridge as a permanent/local address on that brport, but not on the others, and not on br0)? If yes, then I'm afraid there's nothing we can do.
On Tue, Jan 19, 2021 at 12:06:16AM +0200, Vladimir Oltean wrote: > And these: > > $ ip link add br0 type bridge > $ ip link set swp3 master br0 > $ ip link set swp2 master br0 > $ ip link set swp1 master br0 > $ ip link set swp0 master br0 > $ bridge fdb | grep permanent > 00:04:9f:05:de:0a dev swp0 vlan 1 master br0 permanent > 00:04:9f:05:de:0a dev swp0 master br0 permanent > 00:04:9f:05:de:0a dev swp3 vlan 1 master br0 permanent > 00:04:9f:05:de:0a dev swp3 master br0 permanent Ugh, I messed it up. The entries with swp0 are stray here. The permanent entries get reported only for swp3 in the second case.
On 19/01/2021 00:06, Vladimir Oltean wrote: > On Mon, Jan 18, 2021 at 11:53:18PM +0200, Nikolay Aleksandrov wrote: >> On 18/01/2021 23:50, Vladimir Oltean wrote: >>> On Mon, Jan 18, 2021 at 11:39:27PM +0200, Nikolay Aleksandrov wrote: >>>> Apologies for the multiple emails, but wanted to leave an example: >>>> >>>> 00:11:22:33:44:55 dev ens16 master bridge permanent >>>> >>>> This must always exist and user-space must be able to create it, which >>>> might be against what you want to achieve (no BR_FDB_LOCAL entries with >>>> fdb->dst != NULL). >>> >>> Can you give me an example of why it would matter that fdb->dst in this >>> case is set to ens16? >>> >> >> Can you dump it as "dev ens16" without it? :) >> Or alternatively can you send a notification with "dev ens16" without it? >> >> I'm in favor of removing it, but it is risky since some script somewhere might >> be searching for it, or some user-space daemon might expect to see a notification >> for such entries and react on it. > > If "dev ens16" makes no difference to the forwarding and/or termination > path of the bridge, just to user space reporting, then keeping > appearances is a bit pointless. > > For example, DSA switch interfaces inherit by default the MAC address of > the host interface. Having multiple net devices with the same MAC > address works because either they are in different L2 domains (case in > which the MAC addresses should still be unique per domain), or they are > in the same L2 domain, under the same bridge (case in which it is the > bridge who should do IP neighbour resolution etc). > Having that said, let there be these commands: > > $ ip link add br0 type bridge > $ ip link set swp0 master br0 > $ ip link set swp1 master br0 > $ ip link set swp2 master br0 > $ ip link set swp3 master br0 > $ bridge fdb | grep permanent > 00:04:9f:05:de:0a dev swp0 vlan 1 master br0 permanent > 00:04:9f:05:de:0a dev swp0 master br0 permanent > > And these: > > $ ip link add br0 type bridge > $ ip link set swp3 master br0 > $ ip link set swp2 master br0 > $ ip link set swp1 master br0 > $ ip link set swp0 master br0 > $ bridge fdb | grep permanent > 00:04:9f:05:de:0a dev swp0 vlan 1 master br0 permanent > 00:04:9f:05:de:0a dev swp0 master br0 permanent > 00:04:9f:05:de:0a dev swp3 vlan 1 master br0 permanent > 00:04:9f:05:de:0a dev swp3 master br0 permanent > > Preserving the reporting for permanent/local FDB entries added by user > is one thing. But do we need to also preserve this behavior (i.e. report > the first unique MAC address of an interface that joins the bridge as a > permanent/local address on that brport, but not on the others, and not > on br0)? If yes, then I'm afraid there's nothing we can do. > No, it shouldn't be a problem to change that. We should be careful about the way it's changed though because reporting it for all ports might become a scale issue with 4k vlans, and also today you can't add the same mac for multiple ports. Perhaps the best way is to report it for the bridge itself, while still allowing such entries to be added/deleted by user-space.
On Tue, Jan 19, 2021 at 12:42:04AM +0200, Nikolay Aleksandrov wrote: > No, it shouldn't be a problem to change that. We should be careful about the > way it's changed though because reporting it for all ports might become a scale > issue with 4k vlans, and also today you can't add the same mac for multiple ports. > Perhaps the best way is to report it for the bridge itself, while still allowing > such entries to be added/deleted by user-space. I think what Tobias is trying to achieve is: (a) offload the locally terminated FDB addresses through switchdev, in a way that is not "poisoned", i.e. the driver should not be forced to recognize these entries based on the is_local flag. This includes the ports MAC addresses which are currently notified as is_local and with fdb->dst = source brport (not NULL). (b) remain compatible with the mistakes of the past, i.e. DSA and probably other switchdev users will have to remain oblivious of the is_local flag. So we will still have to accept "bridge fdb add 00:01:02:03:04:05 dev swp0 master local", and it will have to keep incorrectly installing a front-facing static FDB entry on swp0 instead of a local/permanent one. In terms of implementation, this would mean that for added_by_user entries, we keep the existing notifications broken as they are. Whereas for !added_by_user, we replace them as much as possible with "fdb->dst == NULL" entries (i.e. for br0). I haven't looked closely at the code, and I hope that this will not happen, but maybe some of these addresses will inevitably have to be duplicated with is_local addresses that were previously notified. In that case I'm thinking there must be some hackery to always offload the addresses in this order: first the is_local address, then the br0 address, to allow the bad entry to be overwritten with the good one. Finally, we should modify the bridge manpage to say "we know that the local|permanent flag is added by default, but it's deprecated so pls don't use it anymore, just use fdb on br0". How does this sound?
On 19/01/2021 02:42, Vladimir Oltean wrote: > On Tue, Jan 19, 2021 at 12:42:04AM +0200, Nikolay Aleksandrov wrote: >> No, it shouldn't be a problem to change that. We should be careful about the >> way it's changed though because reporting it for all ports might become a scale >> issue with 4k vlans, and also today you can't add the same mac for multiple ports. >> Perhaps the best way is to report it for the bridge itself, while still allowing >> such entries to be added/deleted by user-space. > > I think what Tobias is trying to achieve is: > (a) offload the locally terminated FDB addresses through switchdev, in a > way that is not "poisoned", i.e. the driver should not be forced to > recognize these entries based on the is_local flag. This includes > the ports MAC addresses which are currently notified as is_local and > with fdb->dst = source brport (not NULL). > (b) remain compatible with the mistakes of the past, i.e. DSA and > probably other switchdev users will have to remain oblivious of the > is_local flag. So we will still have to accept "bridge fdb add > 00:01:02:03:04:05 dev swp0 master local", and it will have to keep > incorrectly installing a front-facing static FDB entry on swp0 > instead of a local/permanent one. > > In terms of implementation, this would mean that for added_by_user > entries, we keep the existing notifications broken as they are. > Whereas for !added_by_user, we replace them as much as possible with > "fdb->dst == NULL" entries (i.e. for br0). > > I haven't looked closely at the code, and I hope that this will not > happen, but maybe some of these addresses will inevitably have to be > duplicated with is_local addresses that were previously notified. In > that case I'm thinking there must be some hackery to always offload the > addresses in this order: first the is_local address, then the br0 > address, to allow the bad entry to be overwritten with the good one. > > Finally, we should modify the bridge manpage to say "we know that the > local|permanent flag is added by default, but it's deprecated so pls > don't use it anymore, just use fdb on br0". > > How does this sound? > We'll be supporting it forever, I don't see how it's being deprecated. :) Either way I'm ok with the above, but I'll be able to comment further when I see how exactly the code would change. We should be very careful not to break someone who uses these entries in a way we can't think of, for example we use permanent user-space added entries combined with ext_learn for mlag purposes which is a stranger use case, granted it won't be broken by the above. Perhaps we should consider making the new behaviour optional instead, then we can completely switch between the two modes and drop compatibility.
diff --git a/include/net/switchdev.h b/include/net/switchdev.h index 88fcac140966..43e4469a17b1 100644 --- a/include/net/switchdev.h +++ b/include/net/switchdev.h @@ -213,6 +213,7 @@ struct switchdev_notifier_fdb_info { const unsigned char *addr; u16 vid; u8 added_by_user:1, + local:1, offloaded:1; }; diff --git a/net/bridge/br_switchdev.c b/net/bridge/br_switchdev.c index ff470add8e52..1090bb3d4ee0 100644 --- a/net/bridge/br_switchdev.c +++ b/net/bridge/br_switchdev.c @@ -109,6 +109,7 @@ br_switchdev_fdb_notify(const struct net_bridge_fdb_entry *fdb, int type) .addr = fdb->key.addr.addr, .vid = fdb->key.vlan_id, .added_by_user = test_bit(BR_FDB_ADDED_BY_USER, &fdb->flags), + .local = test_bit(BR_FDB_LOCAL, &fdb->flags), .offloaded = test_bit(BR_FDB_OFFLOADED, &fdb->flags), };
Some switchdev drivers, notably DSA, ignore all dynamically learned address notifications (!added_by_user) as these are autonomously added by the switch. Previously, such a notification was indistinguishable from a local address notification. Include a local bit in the notification so that the two classes can be discriminated. This allows DSA-like devices to add local addresses to the hardware FDB (with the CPU as the destination), thereby avoiding flows towards the CPU being flooded by the switch as unknown unicast. Signed-off-by: Tobias Waldekranz <tobias@waldekranz.com> --- include/net/switchdev.h | 1 + net/bridge/br_switchdev.c | 1 + 2 files changed, 2 insertions(+)