> -----Original Message----- > From: Jason Gunthorpe > Sent: Friday, May 3, 2019 7:45 PM > To: Leon Romanovsky <leon@xxxxxxxxxx> > Cc: Doug Ledford <dledford@xxxxxxxxxx>; Leon Romanovsky > <leonro@xxxxxxxxxxxx>; RDMA mailing list <linux-rdma@xxxxxxxxxxxxxxx>; > Huy Nguyen <huyn@xxxxxxxxxxxx>; Martin Wilck <mwilck@xxxxxxxx>; > Parav Pandit <parav@xxxxxxxxxxxx> > Subject: Re: [PATCH rdma-next v2 0/7] Allow RoCE GID attribute netdev > detachment > > On Thu, May 02, 2019 at 10:48:00AM +0300, Leon Romanovsky wrote: > > From: Leon Romanovsky <leonro@xxxxxxxxxxxx> > > > > Changelog: > > v1 -> v2: > > * Resent > > v0 -> v1: > > * Fixed wrong RCU pointer access in patch "RDMA/core: Allow detaching > > gid attribute netdevice for RoCE" > > > > ---------------------------------------------------------------------- > > - > > > > >From Parav, > > > > This series covers following changes. > > > > 1. A fix in RXE to consider right reserved space of the netdev. > > 2. ib_cm to avoid accessing netdev of GID attribute. > > 3. Several RoCE drivers and net/smc needs to know the mac and vlan of > > the GID entry. > > > > Instead of open coded accessing netdev fields, we introduce an API to > > get such fields filled up using new API rdma_read_gid_l2_fields(). > > > > 4. When there is active traffic through a GID, a QP/AH holds reference > > to this GID entry. GID entry holds reference to its attached netdevice. > > Due to this when netdevice (such as vlan netdev) is deleted by admin > > user, its refcount is not dropped. > > > > Therefore, use netdev under rcu lock so that netdev reference can be > > dropped when netdev and associated RoCE GID entry is deleted. This is > > facilitated by existing API rdma_read_gid_attr_ndev_rcu. > > > > Thanks > > > > > > Parav Pandit (7): > > RDMA/rxe: Consider skb reserve space based on netdev of GID > > IB/cm: Reduce dependency on gid attribute ndev check > > RDMA: Introduce and use GID attr helper to read RoCE L2 fields > > RDMA/cma: Use rdma_read_gid_attr_ndev_rcu to access netdev > > RDMA/rxe: Use rdma_read_gid_attr_ndev_rcu to access netdev > > net/smc: Use rdma_read_gid_l2_fields to L2 fields > > RDMA/core: Allow detaching gid attribute netdevice for RoCE > > > > drivers/infiniband/core/addr.c | 1 + > > drivers/infiniband/core/cache.c | 117 +++++++++++++++++++-- > > drivers/infiniband/core/cm.c | 5 +- > > drivers/infiniband/core/cma.c | 12 ++- > > drivers/infiniband/core/sysfs.c | 13 ++- > > drivers/infiniband/hw/bnxt_re/ib_verbs.c | 18 ++-- > > drivers/infiniband/hw/hns/hns_roce_ah.c | 14 +-- > > drivers/infiniband/hw/hns/hns_roce_hw_v2.c | 7 +- > > drivers/infiniband/hw/mlx4/ah.c | 8 +- > > drivers/infiniband/hw/mlx4/qp.c | 6 +- > > drivers/infiniband/hw/mlx5/main.c | 42 ++------ > > drivers/infiniband/hw/ocrdma/ocrdma_ah.c | 9 +- > > drivers/infiniband/hw/ocrdma/ocrdma_hw.c | 7 +- > > drivers/infiniband/hw/qedr/qedr_roce_cm.c | 11 +- > > drivers/infiniband/hw/qedr/verbs.c | 5 +- > > drivers/infiniband/sw/rxe/rxe_net.c | 18 +++- > > include/rdma/ib_cache.h | 4 + > > include/rdma/ib_verbs.h | 2 +- > > net/smc/smc_ib.c | 16 +-- > > 19 files changed, 221 insertions(+), 94 deletions(-) > > Now that we have the RCU pointer in the gid_attr it is really ugly that the > onstack version and the pointer version are the same type, this needs a > cleanup to add some kind of gid_attr_init structure instead > Hmm ok. Other option I was considering is to pass ndev pointer as additional argument in add_gid code flow. But it was too much changes in roce_gid_mgmt.c. We can possibly keep the scope limited of gid_init_attr within cache.c and git_mgmt.c. > But otherwise applied to for-next > Thanks > Thanks, > Jason
