> From: linux-rdma-owner@xxxxxxxxxxxxxxx <linux-rdma-
> owner@xxxxxxxxxxxxxxx> On Behalf Of Michal Kalderon
>
> > From: Gal Pressman <galpress@xxxxxxxxxx>
> > Sent: Thursday, August 22, 2019 11:35 AM
> >
> > On 20/08/2019 15:18, Michal Kalderon wrote:
>
> > >
> > > +void rdma_user_mmap_entry_free(struct kref *kref) {
> > > + struct rdma_user_mmap_entry *entry =
> > > + container_of(kref, struct rdma_user_mmap_entry, ref);
> > > + unsigned long i, npages = (u32)DIV_ROUND_UP(entry->length,
> > PAGE_SIZE);
> > > + struct ib_ucontext *ucontext = entry->ucontext;
> > > +
> > > + /* need to erase all entries occupied... */
> > > + for (i = 0; i < npages; i++) {
> > > + xa_erase(&ucontext->mmap_xa, entry->mmap_page + i);
> > > +
> > > + ibdev_dbg(ucontext->device,
> > > + "mmap: obj[0x%p] key[%#llx] addr[%#llx] len[%#llx]
> > npages[%#lx] removed\n",
> > > + entry->obj, rdma_user_mmap_get_key(entry),
> > > + entry->address, entry->length, npages);
> > > +
> > > + if (ucontext->device->ops.mmap_free)
> > > + ucontext->device->ops.mmap_free(entry);
> > > + }
> >
> > Should this loop be surrounded with a lock? What happens with
> > concurrent insertions?
> This is a good point, thanks for catching it.
Gal, after revisiting this, I think there is no need for a lock, the xa_erase is safe
When run concurrently with the insertions under xa_lock, and if an entry is erased
During the insert there is no harm.
Thanks,
Michal
>
> >
