IB/core: Properly handle registration of on-demand paging MRs after dereg

When the last on-demand paging MR is released the notifier count is
left non-zero so that concurrent page faults will have to abort. If a
new MR is then registered, the counter is reset. However, the decision
is made to put the new MR in the list waiting for the notifier count
to reach zero, before the counter is reset. An invalidation or another
MR registration can release the MR to handle page faults, but without
such an event the MR can wait forever.

The patch fixes this issue by adding a check whether the MR is the
first on-demand paging MR when deciding whether it is ready to handle
page faults. If it is the first MR, we know that there are no mmu
notifiers running in parallel to the registration.

Fixes: 882214e2b128 ("IB/core: Implement support for MMU notifiers regarding on demand paging regions")
Signed-off-by: Haggai Eran <haggaie@mellanox.com>
Signed-off-by: Shachar Raindel <raindel@mellanox.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>

diff --git a/drivers/infiniband/core/umem_odp.c b/drivers/infiniband/core/umem_odp.c
index 6095872549e79fb0dd9c0b3702c4eb01e974fb3f..8b8cc6fa0ab0c1ebf966b2ace4932807d9181bf1 100644 (file)
--- a/drivers/infiniband/core/umem_odp.c
+++ b/drivers/infiniband/core/umem_odp.c
@@ -294,7 +294,8 @@ int ib_umem_odp_get(struct ib_ucontext *context, struct ib_umem *umem)
        if (likely(ib_umem_start(umem) != ib_umem_end(umem)))
                rbt_ib_umem_insert(&umem->odp_data->interval_tree,
                                   &context->umem_tree);
-       if (likely(!atomic_read(&context->notifier_count)))
+       if (likely(!atomic_read(&context->notifier_count)) ||
+           context->odp_mrs_count == 1)
                umem->odp_data->mn_counters_active = true;
        else
                list_add(&umem->odp_data->no_private_counters,