> From: Gal Pressman <galpress@xxxxxxxxxx> > Sent: Sunday, August 25, 2019 1:40 PM > > On 25/08/2019 11:36, Michal Kalderon wrote: > >>> diff --git a/drivers/infiniband/core/rdma_core.c > >>> b/drivers/infiniband/core/rdma_core.c > >>> index ccf4d069c25c..7166741834c8 100644 > >>> --- a/drivers/infiniband/core/rdma_core.c > >>> +++ b/drivers/infiniband/core/rdma_core.c > >>> @@ -817,6 +817,7 @@ static void ufile_destroy_ucontext(struct > >> ib_uverbs_file *ufile, > >>> rdma_restrack_del(&ucontext->res); > >>> > >>> ib_dev->ops.dealloc_ucontext(ucontext); > >>> + rdma_user_mmap_entries_remove_free(ucontext); > >> > >> Why did you switch the order again? > > To enable drivers to remove the entries from the mmap_xa otherwise > > entires_remove_free Will run into a mmap_xa that is not empty. I should > have mentioned this in the cover letter. > > I don't understand. > Do you expect drivers to explicitly drain the mmap xarray during > dealloc_ucontext? I didn't see that in the EFA patch. > > I still think the xarray should be cleared prior to calling the driver's > dealloc_ucontext. Drivers don't need to explicitly drain the xarray, but in qedr the driver inserts An entry during alloc_ucontext, and therefore should remove the entry on dealloc Ucontext. We want to reacc the remove_free function with no entries. When going Over this code again I think I can completely remove the function and just add a WARN_ON on the dealloc_ucontext function. Having an entry at this point means there Is a leak.
