| Message ID | 20191212113024.336702-1-leon@kernel.org (mailing list archive) |
|---|---|
| Headers | show |
| Series | Let IB core distribute cache update events | expand |
On Thu, Dec 12, 2019 at 01:30:20PM +0200, Leon Romanovsky wrote: > Parav Pandit (4): > IB/mlx5: Do reverse sequence during device removal > IB/core: Let IB core distribute cache update events > IB/core: Cut down single member ib_cache structure > IB/core: Prefix qp to event_handler_lock I used qp_open_list_lock in the last patch, and I'm still interested if/why globally serializing the qp handlers is required, or if that could be rw spinlock too. Otherwise applied to for-next Thanks, Jason
On 1/8/2020 5:58 AM, Jason Gunthorpe wrote: > On Thu, Dec 12, 2019 at 01:30:20PM +0200, Leon Romanovsky wrote: > >> Parav Pandit (4): >> IB/mlx5: Do reverse sequence during device removal >> IB/core: Let IB core distribute cache update events >> IB/core: Cut down single member ib_cache structure >> IB/core: Prefix qp to event_handler_lock > > I used qp_open_list_lock in the last patch, and I'm still interested > if/why globally serializing the qp handlers is required, or if that > could be rw spinlock too. > My understanding is as in email of patch-2, its open_list_lock. probably there isn't too much contention, but yes it can be changed to rw spinlock. > Otherwise applied to for-next > Thanks.
From: Leon Romanovsky <leonro@mellanox.com> Changelog: --- v0->v1: - Address Jason's comment to split qp event handler lock with IB device event - Added patch that add qp_ prefix to reflect QP event operation Note: Patch #1 can go to the -rc too, and is sent here because "fix" is needed only if we are using those cache patches. ------------------------------------------------------------------------- From Parav, Currently when low level driver notifies Pkey, GID, port change events, they are notified to the registered handlers in the order they are registered. IB core and other ULPs such as IPoIB are interested in GID, LID, Pkey change events. Since all GID query done by ULPs is serviced by IB core, IB core is yet to update the GID cache when IPoIB queries the GID, resulting into not updating IPoIB address. Hence, all events which require cache update are handled first by the IB core. Once cache update work is completed, IB core distributes the event to subscriber clients. This is tested with opensm's /etc/rdma/opensm.conf subnet_prefix configuration update where before the update $ ip link show dev ib0 ib0: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 2044 qdisc pfifo_fast state UP mode DEFAULT group default qlen 256 link/infiniband 80:00:01:07:fe:80:00:00:00:00:00:00:24:8a:07:03:00:b3:d1:12 brd 00:ff:ff:ff:ff:12:40:1b:ff:ff:00:00:00:00:00:00:ff:ff:ff:ff And after the subnet prefix update: ib0: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 2044 qdisc pfifo_fast state UP mode DEFAULT group default qlen 256 link/infiniband 80:00:01:07:fe:80:00:00:00:00:00:02:24:8a:07:03:00:b3:d1:12 brd 00:ff:ff:ff:ff:12:40:1b:ff:ff:00:00:00:00:00:00:ff:ff:ff:ff Thanks Parav Pandit (4): IB/mlx5: Do reverse sequence during device removal IB/core: Let IB core distribute cache update events IB/core: Cut down single member ib_cache structure IB/core: Prefix qp to event_handler_lock drivers/infiniband/core/cache.c | 151 +++++++++++++++++----------- drivers/infiniband/core/core_priv.h | 1 + drivers/infiniband/core/device.c | 35 ++----- drivers/infiniband/core/verbs.c | 12 +-- drivers/infiniband/hw/mlx5/main.c | 2 + include/rdma/ib_verbs.h | 16 +-- 6 files changed, 118 insertions(+), 99 deletions(-) -- 2.20.1