Message ID | 168330138101.5953.12575990094340826016.stgit@oracle-102.nfsv4bat.org (mailing list archive) |
---|---|
State | Rejected |
Headers | show |
Series | siw on tunnel devices | expand |
On Fri, May 05, 2023 at 11:43:11AM -0400, Chuck Lever wrote: > From: Chuck Lever <chuck.lever@oracle.com> > > In the past, LOOPBACK and NONE (tunnel) devices had all-zero MAC > addresses. siw_device_create() would fall back to copying the > device's name in those cases, because an all-zero MAC address breaks > the RDMA core IP-to-device lookup mechanism. Why not just make up a dummy address in SIW? It shouldn't need to leak out of it.. It is just some artifact of how the iWarp stuff has been designed Jason
> On May 5, 2023, at 3:58 PM, Jason Gunthorpe <jgg@ziepe.ca> wrote: > > On Fri, May 05, 2023 at 11:43:11AM -0400, Chuck Lever wrote: >> From: Chuck Lever <chuck.lever@oracle.com> >> >> In the past, LOOPBACK and NONE (tunnel) devices had all-zero MAC >> addresses. siw_device_create() would fall back to copying the >> device's name in those cases, because an all-zero MAC address breaks >> the RDMA core IP-to-device lookup mechanism. > > Why not just make up a dummy address in SIW? It shouldn't need to leak > out of it.. It is just some artifact of how the iWarp stuff has been > designed I've been trying that. Even though the siw0 device is now registered with a non-zero GID, cma_acquire_dev_by_src_ip() still comes up with a zero GID which matches no device. Address resolution then fails. I'm still looking into why. -- Chuck Lever
> On May 5, 2023, at 4:03 PM, Chuck Lever III <chuck.lever@oracle.com> wrote: > >> On May 5, 2023, at 3:58 PM, Jason Gunthorpe <jgg@ziepe.ca> wrote: >> >> On Fri, May 05, 2023 at 11:43:11AM -0400, Chuck Lever wrote: >>> From: Chuck Lever <chuck.lever@oracle.com> >>> >>> In the past, LOOPBACK and NONE (tunnel) devices had all-zero MAC >>> addresses. siw_device_create() would fall back to copying the >>> device's name in those cases, because an all-zero MAC address breaks >>> the RDMA core IP-to-device lookup mechanism. >> >> Why not just make up a dummy address in SIW? It shouldn't need to leak >> out of it.. It is just some artifact of how the iWarp stuff has been >> designed > > I've been trying that. > > Even though the siw0 device is now registered with a non-zero GID, > cma_acquire_dev_by_src_ip() still comes up with a zero GID which > matches no device. Address resolution then fails. > > I'm still looking into why. The tun0 device's flags are: UP|POINTOPOINT|NOARP|MULTICAST That flag combination turns addr_resolve_neigh() into a no-op, so that the returned GIDs and addresses are uninitialized. Cc'ing Parav because he's the last person who did significant work on this code path. I can hack this to make it work, but I have no idea what the proper solution would be. -- Chuck Lever
> On May 5, 2023, at 3:58 PM, Jason Gunthorpe <jgg@ziepe.ca> wrote: > > On Fri, May 05, 2023 at 11:43:11AM -0400, Chuck Lever wrote: >> From: Chuck Lever <chuck.lever@oracle.com> >> >> In the past, LOOPBACK and NONE (tunnel) devices had all-zero MAC >> addresses. siw_device_create() would fall back to copying the >> device's name in those cases, because an all-zero MAC address breaks >> the RDMA core IP-to-device lookup mechanism. > > Why not just make up a dummy address in SIW? It shouldn't need to leak > out of it.. It is just some artifact of how the iWarp stuff has been > designed So that approach is already being done in siw_device_create(), even though it is broken (the device name hasn't been initialized when the phony MAC is created, so it is all zeroes). I've fixed that and it still doesn't help. siw cannot modify the underlying net_device to add a made-up MAC address. The core address resolution code wants to find an L2 address for the egress device. The underlying ib_device, where a made-up GID might be stored, is not involved with address resolution AFAICT. tun devices have no L2 address. Neither do loopback devices, but address resolution makes an exception for LOOPBACK devices by redirecting to a local physical Ethernet device. Redirecting tun traffic to the local Ethernet device seems dodgy at best. I wasn't sure that an L2 address was required for siw before, but now I'm pretty confident that it is required by our implementation. -- Chuck Lever
On 5/23/2023 3:18 PM, Chuck Lever III wrote: > >> On May 5, 2023, at 3:58 PM, Jason Gunthorpe <jgg@ziepe.ca> wrote: >> >> On Fri, May 05, 2023 at 11:43:11AM -0400, Chuck Lever wrote: >>> From: Chuck Lever <chuck.lever@oracle.com> >>> >>> In the past, LOOPBACK and NONE (tunnel) devices had all-zero MAC >>> addresses. siw_device_create() would fall back to copying the >>> device's name in those cases, because an all-zero MAC address breaks >>> the RDMA core IP-to-device lookup mechanism. >> >> Why not just make up a dummy address in SIW? It shouldn't need to leak >> out of it.. It is just some artifact of how the iWarp stuff has been >> designed > > So that approach is already being done in siw_device_create(), > even though it is broken (the device name hasn't been initialized > when the phony MAC is created, so it is all zeroes). I've fixed > that and it still doesn't help. > > siw cannot modify the underlying net_device to add a made-up > MAC address. > > The core address resolution code wants to find an L2 address > for the egress device. The underlying ib_device, where a made-up > GID might be stored, is not involved with address resolution > AFAICT. > > tun devices have no L2 address. Neither do loopback devices, > but address resolution makes an exception for LOOPBACK devices > by redirecting to a local physical Ethernet device. > > Redirecting tun traffic to the local Ethernet device seems > dodgy at best. > > I wasn't sure that an L2 address was required for siw before, > but now I'm pretty confident that it is required by our > implementation. Does rxe work over tunnels? Seems like it would have the same issue. int rxe_register_device(struct rxe_dev *rxe, const char *ibdev_name) { ... addrconf_addr_eui48((unsigned char *)&dev->node_guid, rxe->ndev->dev_addr); static struct siw_device *siw_device_create(struct net_device *netdev) { ... addrconf_addr_eui48((unsigned char *)&base_dev->node_guid, netdev->dev_addr); Tom.
> On May 23, 2023, at 3:44 PM, Tom Talpey <tom@talpey.com> wrote: > > On 5/23/2023 3:18 PM, Chuck Lever III wrote: >>> On May 5, 2023, at 3:58 PM, Jason Gunthorpe <jgg@ziepe.ca> wrote: >>> >>> On Fri, May 05, 2023 at 11:43:11AM -0400, Chuck Lever wrote: >>>> From: Chuck Lever <chuck.lever@oracle.com> >>>> >>>> In the past, LOOPBACK and NONE (tunnel) devices had all-zero MAC >>>> addresses. siw_device_create() would fall back to copying the >>>> device's name in those cases, because an all-zero MAC address breaks >>>> the RDMA core IP-to-device lookup mechanism. >>> >>> Why not just make up a dummy address in SIW? It shouldn't need to leak >>> out of it.. It is just some artifact of how the iWarp stuff has been >>> designed >> So that approach is already being done in siw_device_create(), >> even though it is broken (the device name hasn't been initialized >> when the phony MAC is created, so it is all zeroes). I've fixed >> that and it still doesn't help. >> siw cannot modify the underlying net_device to add a made-up >> MAC address. >> The core address resolution code wants to find an L2 address >> for the egress device. The underlying ib_device, where a made-up >> GID might be stored, is not involved with address resolution >> AFAICT. >> tun devices have no L2 address. Neither do loopback devices, >> but address resolution makes an exception for LOOPBACK devices >> by redirecting to a local physical Ethernet device. >> Redirecting tun traffic to the local Ethernet device seems >> dodgy at best. >> I wasn't sure that an L2 address was required for siw before, >> but now I'm pretty confident that it is required by our >> implementation. > > Does rxe work over tunnels? (It's not tunnels per se, it's devices that don't have L2 addresses... and tun happens to be one instance of that class). My (brief) reading of the source code is that the use of devices that do not have L2 addresses is prohibited for rxe. > Seems like it would have the same issue. Agreed, if rxe did not prohibit them, it would have the same issue. To be clear: siw itself and the family of iWARP protocols shouldn't have any problem at all with such devices. The issue seems to be with the Linux implementation of address resolution. > int rxe_register_device(struct rxe_dev *rxe, const char *ibdev_name) > { > ... > addrconf_addr_eui48((unsigned char *)&dev->node_guid, > rxe->ndev->dev_addr); > > static struct siw_device *siw_device_create(struct net_device *netdev) > { > ... > addrconf_addr_eui48((unsigned char *)&base_dev->node_guid, > netdev->dev_addr); > > Tom. -- Chuck Lever
diff --git a/drivers/infiniband/sw/siw/siw_main.c b/drivers/infiniband/sw/siw/siw_main.c index 65b5cda5457b..2c31bf397993 100644 --- a/drivers/infiniband/sw/siw/siw_main.c +++ b/drivers/infiniband/sw/siw/siw_main.c @@ -304,10 +304,15 @@ static const struct ib_device_ops siw_device_ops = { static struct siw_device *siw_device_create(struct net_device *netdev) { + static const u8 zeromac[ETH_ALEN] = { 0 }; struct siw_device *sdev = NULL; struct ib_device *base_dev; int rv; + if ((netdev->addr_len != ETH_ALEN) || + (memcmp(netdev->dev_addr, zeromac, ETH_ALEN) == 0)) + return NULL; + sdev = ib_alloc_device(siw_device, base_dev); if (!sdev) return NULL; @@ -316,21 +321,8 @@ static struct siw_device *siw_device_create(struct net_device *netdev) sdev->netdev = netdev; - if (netdev->type != ARPHRD_LOOPBACK && netdev->type != ARPHRD_NONE) { - addrconf_addr_eui48((unsigned char *)&base_dev->node_guid, - netdev->dev_addr); - } else { - /* - * This device does not have a HW address, - * but connection mangagement lib expects gid != 0 - */ - size_t len = min_t(size_t, strlen(base_dev->name), 6); - char addr[6] = { }; - - memcpy(addr, base_dev->name, len); - addrconf_addr_eui48((unsigned char *)&base_dev->node_guid, - addr); - } + addrconf_addr_eui48((unsigned char *)&base_dev->node_guid, + netdev->dev_addr); base_dev->uverbs_cmd_mask |= BIT_ULL(IB_USER_VERBS_CMD_POST_SEND);