From: David Ahern <dsah...@gmail.com> Date: Mon, 16 Oct 2017 09:36:49 -0700
> Currently, exceeding the number of VRF instances or the number of router > interfaces either fails with a non-intuitive EBUSY: > $ ip li set swp1s1.6 vrf vrf-1s1-6 up > RTNETLINK answers: Device or resource busy > > or fails silently (IPv6) since the checks are done in a work queue. This > set adds support for the address validator notifier to spectrum which > allows ext-ack based messages to be returned on failure. > > To make that happen the IPv6 version needs to be converted from atomic > to blocking (patch 2), and then support for extack needs to be added > to the notifier (patch 3). Patch 1 reworks the locking in ipv6_add_addr > to work better in the atomic and non-atomic code paths. Patches 4 and 5 > add the validator notifier to spectrum and then plumb the extack argument > through spectrum_router. > > With this set, VRF overflows fail with: > $ ip li set swp1s1.6 vrf vrf-1s1-6 up > Error: spectrum: Exceeded number of supported VRF. > > and RIF overflows fail with: > $ ip addr add dev swp1s2.191 10.12.191.1/24 > Error: spectrum: Exceeded number of supported router interfaces. > > v1 -> v2 > - fix error path in ipv6_add_addr: reset rt to NULL (Ido comment) and > add in6_dev_put on ifa once the hold has been done > > RFC -> v1 > - addressed various comments from Ido > - refactored ipv6_add_addr to allow ifa's to be allocated with > GFP_KERNEL as requested by DaveM This doesn't apply cleanly to net-next, specifically the mlxsw driver changes. Please respin, thanks!
