lkml.org 
[lkml]   [2021]   [Jun]   [28]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH rdma-rc v2] RDMA/core: Simplify addition of restrack object
On Sun, Jun 27, 2021 at 08:15:28PM -0300, Jason Gunthorpe wrote:
> On Sun, Jun 27, 2021 at 11:07:33AM +0300, Leon Romanovsky wrote:
> > On Thu, Jun 24, 2021 at 02:48:41PM -0300, Jason Gunthorpe wrote:
> > > On Tue, Jun 08, 2021 at 08:23:48AM +0300, Leon Romanovsky wrote:
> > > > From: Leon Romanovsky <leonro@nvidia.com>
> > > >
> > > > Change location of rdma_restrack_add() callers to be near attachment
> > > > to device logic. Such improvement fixes the bug where task_struct was
> > > > acquired but not released, causing to resource leak.
> > > >
> > > > ucma_create_id() {
> > > > ucma_alloc_ctx();
> > > > rdma_create_user_id() {
> > > > rdma_restrack_new();
> > > > rdma_restrack_set_name() {
> > > > rdma_restrack_attach_task.part.0(); <--- task_struct was gotten
> > > > }
> > > > }
> > > > ucma_destroy_private_ctx() {
> > > > ucma_put_ctx();
> > > > rdma_destroy_id() {
> > > > _destroy_id() <--- id_priv was freed
> > > > }
> > > > }
> > > > }
> > >
> > > I still don't understand this patch
> > >
> > > > @@ -1852,6 +1849,7 @@ static void _destroy_id(struct rdma_id_private *id_priv,
> > > > {
> > > > cma_cancel_operation(id_priv, state);
> > > >
> > > > + rdma_restrack_del(&id_priv->res);
> > > > if (id_priv->cma_dev) {
> > > > if (rdma_cap_ib_cm(id_priv->id.device, 1)) {
> > > > if (id_priv->cm_id.ib)
> > > > @@ -1861,7 +1859,6 @@ static void _destroy_id(struct rdma_id_private *id_priv,
> > > > iw_destroy_cm_id(id_prgtiv->cm_id.iw);
> > > > }
> > > > cma_leave_mc_groups(id_priv);
> > > > - rdma_restrack_del(&id_priv->res);
> > > > cma_release_dev(id_priv);
> > >
> > > This seems to be the only hunk that is actually necessary, ensuring a
> > > non-added ID is always cleaned up is the necessary step to fixing the
> > > trace above.
> > >
> > > What is the rest of this doing?? It looks wrong:
> > >
> > > int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr)
> > > {
> > > [..]
> > > ret = cma_get_port(id_priv);
> > > if (ret)
> > > goto err2;
> > > err2:
> > > [..]
> > > if (!cma_any_addr(addr))
> > > rdma_restrack_del(&id_priv->res);
> > >
> > > Which means if rdma_bind_addr() fails then restrack will discard the
> > > task, even though the cm_id is still valid! The ucma is free to try
> > > bind again and keep using the ID.
> >
> > "Failure to bind" means that cma_attach_to_dev() needs to be unwind.
> >
> > It is the same if rdma_restrack_add() inside that function like in this
> > patch or in the line before rdma_bind_addr() returns as it was in
> > previous code.
>
> The previous code didn't call restrack_del. restrack_del undoes the
> restrack_set_name stuff, not just add - so it does not leave things
> back the way it found them

The previous code didn't call to restrack_add and this is why it didn't
call to restrack_del later. In old and new code, we are still calling to
acquire and release dev (cma_acquire_dev_by_src_ip/cma_release_dev) and
this is where the CM_ID is actually attached.

Thanks

>
> Jason

\
 
 \ /
  Last update: 2021-06-28 07:23    [W:0.070 / U:0.568 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site