On Mon, Jan 21, 2019 at 02:43:41PM -0700, Jason Gunthorpe wrote: > On Fri, Jan 11, 2019 at 07:31:22PM -0700, Jason Gunthorpe wrote: > > From: Jason Gunthorpe <jgg@xxxxxxxxxxxx> > > > > Work needs to hold a reference on the pointers that it plans to use otherwise > > they can become free'd or unregistered. Expose the existing registration lock > > that netlink uses for this purpose. > > > > Jason Gunthorpe (2): > > RDMA/device: Expose ib_device_try_get(() > > IB/mlx5: Fix how advise_mr() launches async work > > I've applied both of these to for-rc, with the change in comment that > Parav gave. > > We can revisit if get_device/put_device is the best way to handle this > when the work queue issue is resolved, most probably we can drop the > get/put in the driver once the work queue is flushed prior to > deallocing the device. Didn't we agree that the right fix is to delete advise_mr and sync system queue on exit? It will be much more clean to do it right from the beginning instead of adding new get_device/put_device and removing them immediately in following patch. Thanks > > Jason
Attachment:
signature.asc
Description: PGP signature
