futex: Remove requirement for lock_page() in get_futex_key()

When dealing with key handling for shared futexes, we can drastically reduce
the usage/need of the page lock. 1) For anonymous pages, the associated futex
object is the mm_struct which does not require the page lock. 2) For inode
based, keys, we can check under RCU read lock if the page mapping is still
valid and take reference to the inode. This just leaves one rare race that
requires the page lock in the slow path when examining the swapcache.

Additionally realtime users currently have a problem with the page lock being
contended for unbounded periods of time during futex operations.

Task A
     get_futex_key()
     lock_page()
    ---> preempted

Now any other task trying to lock that page will have to wait until
task A gets scheduled back in, which is an unbound time.

With this patch, we pretty much have a lockless futex_get_key().

Experiments show that this patch can boost/speedup the hashing of shared
futexes with the perf futex benchmarks (which is good for measuring such
change) by up to 45% when there are high (> 100) thread counts on a 60 core
Westmere. Lower counts are pretty much in the noise range or less than 10%,
but mid range can be seen at over 30% overall throughput (hash ops/sec).
This makes anon-mem shared futexes much closer to its private counterpart.

Signed-off-by: Mel Gorman <mgorman@suse.de>
[ Ported on top of thp refcount rework, changelog, comments, fixes. ]
Signed-off-by: Davidlohr Bueso <dbueso@suse.de>
Reviewed-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Chris Mason <clm@fb.com>
Cc: Darren Hart <dvhart@linux.intel.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Cc: dave@stgolabs.net
Link: http://lkml.kernel.org/r/1455045314-8305-3-git-send-email-dave@stgolabs.net
Signed-off-by: Ingo Molnar <mingo@kernel.org>

1 files changed, 91 insertions, 8 deletions

diff --git a/kernel/futex.c b/kernel/futex.c
index 08ac7009488b..bae542e4b2e9 100644
--- a/kernel/futex.c
+++ b/kernel/futex.c
@@ -520,7 +520,20 @@ again:
         else
                 err = 0;
 
-        lock_page(page);
+        /*
+         * The treatment of mapping from this point on is critical. The page
+         * lock protects many things but in this context the page lock
+         * stabilizes mapping, prevents inode freeing in the shared
+         * file-backed region case and guards against movement to swap cache.
+         *
+         * Strictly speaking the page lock is not needed in all cases being
+         * considered here and page lock forces unnecessarily serialization
+         * From this point on, mapping will be re-verified if necessary and
+         * page lock will be acquired only if it is unavoidable
+         */
+        page = compound_head(page);
+        mapping = READ_ONCE(page->mapping);
+
         /*
          * If page->mapping is NULL, then it cannot be a PageAnon
          * page; but it might be the ZERO_PAGE or in the gate area or
@@ -536,19 +549,31 @@ again:
          * shmem_writepage move it from filecache to swapcache beneath us:
          * an unlikely race, but we do need to retry for page->mapping.
          */
-        mapping = compound_head(page)->mapping;
-        if (!mapping) {
-                int shmem_swizzled = PageSwapCache(page);
+        if (unlikely(!mapping)) {
+                int shmem_swizzled;
+
+                /*
+                 * Page lock is required to identify which special case above
+                 * applies. If this is really a shmem page then the page lock
+                 * will prevent unexpected transitions.
+                 */
+                lock_page(page);
+                shmem_swizzled = PageSwapCache(page) || page->mapping;
                 unlock_page(page);
                 put_page(page);
+
                 if (shmem_swizzled)
                         goto again;
+
                 return -EFAULT;
         }
 
         /*
          * Private mappings are handled in a simple way.
          *
+         * If the futex key is stored on an anonymous page, then the associated
+         * object is the mm which is implicitly pinned by the calling process.
+         *
          * NOTE: When userspace waits on a MAP_SHARED mapping, even if
          * it's a read-only handle, it's expected that futexes attach to
          * the object not the particular process.
@@ -566,16 +591,74 @@ again:
                 key->both.offset |= FUT_OFF_MMSHARED; /* ref taken on mm */
                 key->private.mm = mm;
                 key->private.address = address;
+
+                get_futex_key_refs(key); /* implies smp_mb(); (B) */
+
         } else {
+                struct inode *inode;
+
+                /*
+                 * The associated futex object in this case is the inode and
+                 * the page->mapping must be traversed. Ordinarily this should
+                 * be stabilised under page lock but it's not strictly
+                 * necessary in this case as we just want to pin the inode, not
+                 * update the radix tree or anything like that.
+                 *
+                 * The RCU read lock is taken as the inode is finally freed
+                 * under RCU. If the mapping still matches expectations then the
+                 * mapping->host can be safely accessed as being a valid inode.
+                 */
+                rcu_read_lock();
+
+                if (READ_ONCE(page->mapping) != mapping) {
+                        rcu_read_unlock();
+                        put_page(page);
+
+                        goto again;
+                }
+
+                inode = READ_ONCE(mapping->host);
+                if (!inode) {
+                        rcu_read_unlock();
+                        put_page(page);
+
+                        goto again;
+                }
+
+                /*
+                 * Take a reference unless it is about to be freed. Previously
+                 * this reference was taken by ihold under the page lock
+                 * pinning the inode in place so i_lock was unnecessary. The
+                 * only way for this check to fail is if the inode was
+                 * truncated in parallel so warn for now if this happens.
+                 *
+                 * We are not calling into get_futex_key_refs() in file-backed
+                 * cases, therefore a successful atomic_inc return below will
+                 * guarantee that get_futex_key() will still imply smp_mb(); (B).
+                 */
+                if (WARN_ON_ONCE(!atomic_inc_not_zero(&inode->i_count))) {
+                        rcu_read_unlock();
+                        put_page(page);
+
+                        goto again;
+                }
+
+                /* Should be impossible but lets be paranoid for now */
+                if (WARN_ON_ONCE(inode->i_mapping != mapping)) {
+                        err = -EFAULT;
+                        rcu_read_unlock();
+                        iput(inode);
+
+                        goto out;
+                }
+
                 key->both.offset |= FUT_OFF_INODE; /* inode-based key */
-                key->shared.inode = mapping->host;
+                key->shared.inode = inode;
                 key->shared.pgoff = basepage_index(page);
+                rcu_read_unlock();
         }
 
-        get_futex_key_refs(key); /* implies smp_mb(); (B) */
-
 out:
-        unlock_page(page);
         put_page(page);
         return err;
 }