RDMA/umem: Get rid of per_mm->notifier_count
authorJason Gunthorpe <jgg@mellanox.com>
Sun, 16 Sep 2018 17:48:09 +0000 (20:48 +0300)
committerDoug Ledford <dledford@redhat.com>
Fri, 21 Sep 2018 15:58:36 +0000 (11:58 -0400)
This is intrinsically racy and the scheme is simply unnecessary. New MR
registration can wait for any on going invalidation to fully complete.

      CPU0                              CPU1
                                  if (atomic_read())
 if (atomic_dec_and_test() &&
     !list_empty())
  { /* not taken */ }
                                       list_add()

Putting the new UMEM into some kind of purgatory until another invalidate
rolls through..

Instead hold the read side of the umem_rwsem across the pair'd start/end
and get rid of the racy 'deferred add' approach.

Since all umem's in the rbt are always ready to go, also get rid of the
mn_counters_active stuff.

Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
drivers/infiniband/core/umem_odp.c
include/rdma/ib_umem_odp.h

index 0577f9ff600f6b57536f8beeefac72e8a304824b..1c0c4a431218603f885f7511bef6edb218c4eeb3 100644 (file)
@@ -80,83 +80,29 @@ INTERVAL_TREE_DEFINE(struct umem_odp_node, rb, u64, __subtree_last,
 static void ib_umem_notifier_start_account(struct ib_umem_odp *umem_odp)
 {
        mutex_lock(&umem_odp->umem_mutex);
-
-       /* Only update private counters for this umem if it has them.
-        * Otherwise skip it. All page faults will be delayed for this umem. */
-       if (umem_odp->mn_counters_active) {
-               int notifiers_count = umem_odp->notifiers_count++;
-
-               if (notifiers_count == 0)
-                       /* Initialize the completion object for waiting on
-                        * notifiers. Since notifier_count is zero, no one
-                        * should be waiting right now. */
-                       reinit_completion(&umem_odp->notifier_completion);
-       }
+       if (umem_odp->notifiers_count++ == 0)
+               /*
+                * Initialize the completion object for waiting on
+                * notifiers. Since notifier_count is zero, no one should be
+                * waiting right now.
+                */
+               reinit_completion(&umem_odp->notifier_completion);
        mutex_unlock(&umem_odp->umem_mutex);
 }
 
 static void ib_umem_notifier_end_account(struct ib_umem_odp *umem_odp)
 {
        mutex_lock(&umem_odp->umem_mutex);
-
-       /* Only update private counters for this umem if it has them.
-        * Otherwise skip it. All page faults will be delayed for this umem. */
-       if (umem_odp->mn_counters_active) {
-               /*
-                * This sequence increase will notify the QP page fault that
-                * the page that is going to be mapped in the spte could have
-                * been freed.
-                */
-               ++umem_odp->notifiers_seq;
-               if (--umem_odp->notifiers_count == 0)
-                       complete_all(&umem_odp->notifier_completion);
-       }
+       /*
+        * This sequence increase will notify the QP page fault that the page
+        * that is going to be mapped in the spte could have been freed.
+        */
+       ++umem_odp->notifiers_seq;
+       if (--umem_odp->notifiers_count == 0)
+               complete_all(&umem_odp->notifier_completion);
        mutex_unlock(&umem_odp->umem_mutex);
 }
 
-/* Account for a new mmu notifier in an ib_ucontext. */
-static void
-ib_ucontext_notifier_start_account(struct ib_ucontext_per_mm *per_mm)
-{
-       atomic_inc(&per_mm->notifier_count);
-}
-
-/* Account for a terminating mmu notifier in an ib_ucontext.
- *
- * Must be called with the ib_ucontext->umem_rwsem semaphore unlocked, since
- * the function takes the semaphore itself. */
-static void ib_ucontext_notifier_end_account(struct ib_ucontext_per_mm *per_mm)
-{
-       int zero_notifiers = atomic_dec_and_test(&per_mm->notifier_count);
-
-       if (zero_notifiers &&
-           !list_empty(&per_mm->no_private_counters)) {
-               /* No currently running mmu notifiers. Now is the chance to
-                * add private accounting to all previously added umems. */
-               struct ib_umem_odp *odp_data, *next;
-
-               /* Prevent concurrent mmu notifiers from working on the
-                * no_private_counters list. */
-               down_write(&per_mm->umem_rwsem);
-
-               /* Read the notifier_count again, with the umem_rwsem
-                * semaphore taken for write. */
-               if (!atomic_read(&per_mm->notifier_count)) {
-                       list_for_each_entry_safe(odp_data, next,
-                                                &per_mm->no_private_counters,
-                                                no_private_counters) {
-                               mutex_lock(&odp_data->umem_mutex);
-                               odp_data->mn_counters_active = true;
-                               list_del(&odp_data->no_private_counters);
-                               complete_all(&odp_data->notifier_completion);
-                               mutex_unlock(&odp_data->umem_mutex);
-                       }
-               }
-
-               up_write(&per_mm->umem_rwsem);
-       }
-}
-
 static int ib_umem_notifier_release_trampoline(struct ib_umem_odp *umem_odp,
                                               u64 start, u64 end, void *cookie)
 {
@@ -186,7 +132,6 @@ static void ib_umem_notifier_release(struct mmu_notifier *mn,
        if (!per_mm->context->invalidate_range)
                return;
 
-       ib_ucontext_notifier_start_account(per_mm);
        down_read(&per_mm->umem_rwsem);
        rbt_ib_umem_for_each_in_range(&per_mm->umem_tree, 0,
                                      ULLONG_MAX,
@@ -231,14 +176,9 @@ static int ib_umem_notifier_invalidate_range_start(struct mmu_notifier *mn,
        else if (!down_read_trylock(&per_mm->umem_rwsem))
                return -EAGAIN;
 
-       ib_ucontext_notifier_start_account(per_mm);
-       ret = rbt_ib_umem_for_each_in_range(&per_mm->umem_tree, start,
-                                     end,
-                                     invalidate_range_start_trampoline,
-                                     blockable, NULL);
-       up_read(&per_mm->umem_rwsem);
-
-       return ret;
+       return rbt_ib_umem_for_each_in_range(&per_mm->umem_tree, start, end,
+                                            invalidate_range_start_trampoline,
+                                            blockable, NULL);
 }
 
 static int invalidate_range_end_trampoline(struct ib_umem_odp *item, u64 start,
@@ -259,17 +199,10 @@ static void ib_umem_notifier_invalidate_range_end(struct mmu_notifier *mn,
        if (!per_mm->context->invalidate_range)
                return;
 
-       /*
-        * TODO: we currently bail out if there is any sleepable work to be done
-        * in ib_umem_notifier_invalidate_range_start so we shouldn't really block
-        * here. But this is ugly and fragile.
-        */
-       down_read(&per_mm->umem_rwsem);
        rbt_ib_umem_for_each_in_range(&per_mm->umem_tree, start,
                                      end,
                                      invalidate_range_end_trampoline, true, NULL);
        up_read(&per_mm->umem_rwsem);
-       ib_ucontext_notifier_end_account(per_mm);
 }
 
 static const struct mmu_notifier_ops ib_umem_notifiers = {
@@ -287,12 +220,6 @@ static void add_umem_to_per_mm(struct ib_umem_odp *umem_odp)
        if (likely(ib_umem_start(umem) != ib_umem_end(umem)))
                rbt_ib_umem_insert(&umem_odp->interval_tree,
                                   &per_mm->umem_tree);
-
-       if (likely(!atomic_read(&per_mm->notifier_count)))
-               umem_odp->mn_counters_active = true;
-       else
-               list_add(&umem_odp->no_private_counters,
-                        &per_mm->no_private_counters);
        up_write(&per_mm->umem_rwsem);
 }
 
@@ -305,10 +232,7 @@ static void remove_umem_from_per_mm(struct ib_umem_odp *umem_odp)
        if (likely(ib_umem_start(umem) != ib_umem_end(umem)))
                rbt_ib_umem_remove(&umem_odp->interval_tree,
                                   &per_mm->umem_tree);
-       if (!umem_odp->mn_counters_active) {
-               list_del(&umem_odp->no_private_counters);
-               complete_all(&umem_odp->notifier_completion);
-       }
+       complete_all(&umem_odp->notifier_completion);
 
        up_write(&per_mm->umem_rwsem);
 }
@@ -327,7 +251,6 @@ static struct ib_ucontext_per_mm *alloc_per_mm(struct ib_ucontext *ctx,
        per_mm->mm = mm;
        per_mm->umem_tree = RB_ROOT_CACHED;
        init_rwsem(&per_mm->umem_rwsem);
-       INIT_LIST_HEAD(&per_mm->no_private_counters);
 
        rcu_read_lock();
        per_mm->tgid = get_task_pid(current->group_leader, PIDTYPE_PID);
index 259eb08dfc9e5193d942ba1f358594b721403458..ce950254590330a44911e086dd6f184173514d58 100644 (file)
@@ -67,15 +67,9 @@ struct ib_umem_odp {
        struct mutex            umem_mutex;
        void                    *private; /* for the HW driver to use. */
 
-       /* When false, use the notifier counter in the ucontext struct. */
-       bool mn_counters_active;
        int notifiers_seq;
        int notifiers_count;
 
-       /* A linked list of umems that don't have private mmu notifier
-        * counters yet. */
-       struct list_head no_private_counters;
-
        /* Tree tracking */
        struct umem_odp_node    interval_tree;
 
@@ -99,11 +93,8 @@ struct ib_ucontext_per_mm {
        struct rb_root_cached umem_tree;
        /* Protects umem_tree */
        struct rw_semaphore umem_rwsem;
-       atomic_t notifier_count;
 
        struct mmu_notifier mn;
-       /* A list of umems that don't have private mmu notifier counters yet. */
-       struct list_head no_private_counters;
        unsigned int odp_mrs_count;
 
        struct list_head ucontext_list;
@@ -162,12 +153,6 @@ static inline int ib_umem_mmu_notifier_retry(struct ib_umem_odp *umem_odp,
         * and the ucontext umem_mutex semaphore locked for read).
         */
 
-       /* Do not allow page faults while the new ib_umem hasn't seen a state
-        * with zero notifiers yet, and doesn't have its own valid set of
-        * private counters. */
-       if (!umem_odp->mn_counters_active)
-               return 1;
-
        if (unlikely(umem_odp->notifiers_count))
                return 1;
        if (umem_odp->notifiers_seq != mmu_seq)