(Apologies for multiple emails. Was having trouble with an extension,
cause of which emails did not get delivered to the mailing list.
Resolved now.)
> Somehow nvme-rdma works:
I think that's because the callchain during the nvme_rdma_init_module
initialization stops at "nvmf_register_transport()". Here only the
"struct nvmf_transport_ops nvme_rdma_transport" is registered, which
contains the function "nvme_rdma_create_ctrl()". I tested this in my
local setup and during kernel boot, that's the extent of the
callchain.
The ".create_ctrl"; which now points to "nvme_rdma_create_ctrl()" is
called later from "nvmf_dev_write()". I am not sure when this is
called, probably when the "discover" happens from the client side or
during the server config. I am trying to test this to confirm, will
send more details once I am done.
Am I missing something here?
> If the rdma_create_id() is not on a callchain from module_init then you don't have a problem.
I am a little confused. I thought the problem occurs from a call to
either "rdma_resolve_addr()" which calls "rdma_bind_addr()",
or a direct call to "rdma_bind_addr()" as in rtrs case.
In both the cases, a call to "rdma_create_id()" is needed before this.
> Similarly they are supposed to be created from the client attachment.
I am a beginner in terms of concepts here. Did you mean when a client
tries to establish the first connection to an rdma server?
On Thu, Jun 18, 2020 at 12:56 AM Jason Gunthorpe <jgg@xxxxxxxx> wrote:
>
> On Wed, Jun 17, 2020 at 10:07:56PM +0300, Leon Romanovsky wrote:
> > On Wed, Jun 17, 2020 at 03:20:46PM -0300, Jason Gunthorpe wrote:
> > > On Wed, Jun 17, 2020 at 02:28:11PM +0300, Leon Romanovsky wrote:
> > > > On Wed, Jun 17, 2020 at 04:07:32PM +0530, haris.iqbal@xxxxxxxxxxxxxxx wrote:
> > > > > From: Md Haris Iqbal <haris.iqbal@xxxxxxxxxxxxxxx>
> > > > >
> > > > > Fixes: 2de6c8de192b ("block/rnbd: server: main functionality")
> > > > > Reported-by: kernel test robot <rong.a.chen@xxxxxxxxx>
> > > > > Signed-off-by: Md Haris Iqbal <haris.iqbal@xxxxxxxxxxxxxxx>
> > > > >
> > > > > The rnbd_server module's communication manager initialization depends on the
> > > > > registration of the "network namespace subsystem" of the RDMA CM agent module.
> > > > > As such, when the kernel is configured to load the rnbd_server and the RDMA
> > > > > cma module during initialization; and if the rnbd_server module is initialized
> > > > > before RDMA cma module, a null ptr dereference occurs during the RDMA bind
> > > > > operation.
> > > > > This patch delays the initialization of the rnbd_server module to the
> > > > > late_initcall level, since RDMA cma module uses module_init which puts it into
> > > > > the device_initcall level.
> > > > > drivers/block/rnbd/rnbd-srv.c | 2 +-
> > > > > 1 file changed, 1 insertion(+), 1 deletion(-)
> > > > >
> > > > > diff --git a/drivers/block/rnbd/rnbd-srv.c b/drivers/block/rnbd/rnbd-srv.c
> > > > > index 86e61523907b..213df05e5994 100644
> > > > > +++ b/drivers/block/rnbd/rnbd-srv.c
> > > > > @@ -840,5 +840,5 @@ static void __exit rnbd_srv_cleanup_module(void)
> > > > > rnbd_srv_destroy_sysfs_files();
> > > > > }
> > > > >
> > > > > -module_init(rnbd_srv_init_module);
> > > > > +late_initcall(rnbd_srv_init_module);
> > > >
> > > > I don't think that this is correct change. Somehow nvme-rdma works:
> > > > module_init(nvme_rdma_init_module);
> > > > -> nvme_rdma_init_module
> > > > -> nvmf_register_transport(&nvme_rdma_transport);
> > > > -> nvme_rdma_create_ctrl
> > > > -> nvme_rdma_setup_ctrl
> > > > -> nvme_rdma_configure_admin_queue
> > > > -> nvme_rdma_alloc_queue
> > > > -> rdma_create_id
> > >
> > > If it does work, it is by luck.
> >
> > I didn't check every ULP, but it seems that all ULPs use the same
> > module_init.
> >
> > >
> > > Keep in mind all this only matters for kernels without modules.
> >
> > Can it be related to the fact that other ULPs call to ib_register_client()
> > before calling to rdma-cm? RNBD does not have such call.
>
> If the rdma_create_id() is not on a callchain from module_init then
> you don't have a problem.
>
> nvme has a bug here, IIRC. It is not OK to create RDMA CM IDs outside
> a client - CM IDs are supposed to be cleaned up when the client is
> removed.
>
> Similarly they are supposed to be created from the client attachment.
>
> Though listening CM IDs unbound to any device may change that
> slightly, I think it is probably best practice to create the listening
> ID only if a client is bound.
>
> Most probably that is the best way to fix rnbd
>
> > I'm not proposing this, but just loudly wondering, do we really need rdma-cm
> > as a separate module? Can we bring it to be part of ib_core?
>
> No idea.. It doesn't help this situation at least
>
> Jason
--
Regards
-Haris
