Messages in this thread | | | Date | Tue, 12 Dec 2023 11:16:01 +0000 | From | Simon Horman <> | Subject | Re: [net-next PATCH] octeontx2-af: Fix multicast/mirror group lock/unlock issue |
| |
On Tue, Dec 12, 2023 at 02:45:58PM +0530, Suman Ghosh wrote: > As per the existing implementation, there exists a race between finding > a multicast/mirror group entry and deleting that entry. The group lock > was taken and released independently by rvu_nix_mcast_find_grp_elem() > function. Which is incorrect and group lock should be taken during the > entire operation of group updation/deletion. This patch fixes the same. > > Fixes: 51b2804c19cd ("octeontx2-af: Add new mbox to support multicast/mirror offload") > Signed-off-by: Suman Ghosh <sumang@marvell.com>
...
> @@ -6306,6 +6310,13 @@ int rvu_mbox_handler_nix_mcast_grp_destroy(struct rvu *rvu, > return err; > > mcast_grp = &nix_hw->mcast_grp; > + > + /* If AF is requesting for the deletion, > + * then AF is already taking the lock > + */ > + if (!req->is_af) > + mutex_lock(&mcast_grp->mcast_grp_lock); > + > elem = rvu_nix_mcast_find_grp_elem(mcast_grp, req->mcast_grp_idx); > if (!elem)
Hi Suman,
Does mcast_grp_lock need to be released here? If so, I would suggest a goto label, say unlock_grp.
> return NIX_AF_ERR_INVALID_MCAST_GRP; > @@ -6333,12 +6344,6 @@ int rvu_mbox_handler_nix_mcast_grp_destroy(struct rvu *rvu, > mutex_unlock(&mcast->mce_lock); > > delete_grp: > - /* If AF is requesting for the deletion, > - * then AF is already taking the lock > - */ > - if (!req->is_af) > - mutex_lock(&mcast_grp->mcast_grp_lock); > - > list_del(&elem->list); > kfree(elem); > mcast_grp->count--; > @@ -6370,9 +6375,20 @@ int rvu_mbox_handler_nix_mcast_grp_update(struct rvu *rvu, > return err; > > mcast_grp = &nix_hw->mcast_grp; > + > + /* If AF is requesting for the updation, > + * then AF is already taking the lock > + */ > + if (!req->is_af) > + mutex_lock(&mcast_grp->mcast_grp_lock); > + > elem = rvu_nix_mcast_find_grp_elem(mcast_grp, req->mcast_grp_idx); > - if (!elem) > + if (!elem) { > + if (!req->is_af) > + mutex_unlock(&mcast_grp->mcast_grp_lock); > + > return NIX_AF_ERR_INVALID_MCAST_GRP; > + } > > /* If any pcifunc matches the group's pcifunc, then we can > * delete the entire group. > @@ -6383,8 +6399,11 @@ int rvu_mbox_handler_nix_mcast_grp_update(struct rvu *rvu, > /* Delete group */ > dreq.hdr.pcifunc = elem->pcifunc; > dreq.mcast_grp_idx = elem->mcast_grp_idx; > - dreq.is_af = req->is_af; > + dreq.is_af = 1; > rvu_mbox_handler_nix_mcast_grp_destroy(rvu, &dreq, NULL); > + if (!req->is_af) > + mutex_unlock(&mcast_grp->mcast_grp_lock); > + > return 0; > } > } > @@ -6467,5 +6486,8 @@ int rvu_mbox_handler_nix_mcast_grp_update(struct rvu *rvu, > > done:
I think it would be good to rename this label, say unlock_mce;
> mutex_unlock(&mcast->mce_lock);
Add a new label here, say unlock_grp; And jump to this label whenever there is a need for the mutex_unlock() below.
> + if (!req->is_af) > + mutex_unlock(&mcast_grp->mcast_grp_lock); > + > return ret; > } > -- > 2.25.1 >
| |