FUTEX: new PRIVATE futexes
authorEric Dumazet <dada1@cosmosbay.com>
Wed, 9 May 2007 09:35:04 +0000 (02:35 -0700)
committerLinus Torvalds <torvalds@woody.linux-foundation.org>
Wed, 9 May 2007 19:30:55 +0000 (12:30 -0700)
  Analysis of current linux futex code :
  --------------------------------------

A central hash table futex_queues[] holds all contexts (futex_q) of waiting
threads.

Each futex_wait()/futex_wait() has to obtain a spinlock on a hash slot to
perform lookups or insert/deletion of a futex_q.

When a futex_wait() is done, calling thread has to :

1) - Obtain a read lock on mmap_sem to be able to validate the user pointer
     (calling find_vma()). This validation tells us if the futex uses
     an inode based store (mapped file), or mm based store (anonymous mem)

2) - compute a hash key

3) - Atomic increment of reference counter on an inode or a mm_struct

4) - lock part of futex_queues[] hash table

5) - perform the test on value of futex.
(rollback is value != expected_value, returns EWOULDBLOCK)
(various loops if test triggers mm faults)

6) queue the context into hash table, release the lock got in 4)

7) - release the read_lock on mmap_sem

   <block>

8) Eventually unqueue the context (but rarely, as this part  may be done
   by the futex_wake())

Futexes were designed to improve scalability but current implementation has
various problems :

- Central hashtable :

  This means scalability problems if many processes/threads want to use
  futexes at the same time.
  This means NUMA unbalance because this hashtable is located on one node.

- Using mmap_sem on every futex() syscall :

  Even if mmap_sem is a rw_semaphore, up_read()/down_read() are doing atomic
  ops on mmap_sem, dirtying cache line :
    - lot of cache line ping pongs on SMP configurations.

  mmap_sem is also extensively used by mm code (page faults, mmap()/munmap())
  Highly threaded processes might suffer from mmap_sem contention.

  mmap_sem is also used by oprofile code. Enabling oprofile hurts threaded
  programs because of contention on the mmap_sem cache line.

- Using an atomic_inc()/atomic_dec() on inode ref counter or mm ref counter:
  It's also a cache line ping pong on SMP. It also increases mmap_sem hold time
  because of cache misses.

Most of these scalability problems come from the fact that futexes are in
one global namespace.  As we use a central hash table, we must make sure
they are all using the same reference (given by the mm subsystem).  We
chose to force all futexes be 'shared'.  This has a cost.

But fact is POSIX defined PRIVATE and SHARED, allowing clear separation,
and optimal performance if carefuly implemented.  Time has come for linux
to have better threading performance.

The goal is to permit new futex commands to avoid :
 - Taking the mmap_sem semaphore, conflicting with other subsystems.
 - Modifying a ref_count on mm or an inode, still conflicting with mm or fs.

This is possible because, for one process using PTHREAD_PROCESS_PRIVATE
futexes, we only need to distinguish futexes by their virtual address, no
matter the underlying mm storage is.

If glibc wants to exploit this new infrastructure, it should use new
_PRIVATE futex subcommands for PTHREAD_PROCESS_PRIVATE futexes.  And be
prepared to fallback on old subcommands for old kernels.  Using one global
variable with the FUTEX_PRIVATE_FLAG or 0 value should be OK.

PTHREAD_PROCESS_SHARED futexes should still use the old subcommands.

Compatibility with old applications is preserved, they still hit the
scalability problems, but new applications can fly :)

Note : the same SHARED futex (mapped on a file) can be used by old binaries
*and* new binaries, because both binaries will use the old subcommands.

Note : Vast majority of futexes should be using PROCESS_PRIVATE semantic,
as this is the default semantic. Almost all applications should benefit
of this changes (new kernel and updated libc)

Some bench results on a Pentium M 1.6 GHz (SMP kernel on a UP machine)

/* calling futex_wait(addr, value) with value != *addr */
433 cycles per futex(FUTEX_WAIT) call (mixing 2 futexes)
424 cycles per futex(FUTEX_WAIT) call (using one futex)
334 cycles per futex(FUTEX_WAIT_PRIVATE) call (mixing 2 futexes)
334 cycles per futex(FUTEX_WAIT_PRIVATE) call (using one futex)
For reference :
187 cycles per getppid() call
188 cycles per umask() call
181 cycles per ni_syscall() call

Signed-off-by: Eric Dumazet <dada1@cosmosbay.com>
Pierre Peiffer <pierre.peiffer@bull.net>
Cc: "Ulrich Drepper" <drepper@gmail.com>
Cc: "Nick Piggin" <nickpiggin@yahoo.com.au>
Cc: "Ingo Molnar" <mingo@elte.hu>
Cc: Rusty Russell <rusty@rustcorp.com.au>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
include/linux/futex.h
kernel/futex.c

index 1bd8dfcb037b5958abcb07d0dcb578517db82004..899fc7f20edd1b761023a58cbb883b12ba3e36ef 100644 (file)
@@ -19,6 +19,18 @@ union ktime;
 #define FUTEX_TRYLOCK_PI       8
 #define FUTEX_CMP_REQUEUE_PI   9
 
+#define FUTEX_PRIVATE_FLAG     128
+#define FUTEX_CMD_MASK         ~FUTEX_PRIVATE_FLAG
+
+#define FUTEX_WAIT_PRIVATE     (FUTEX_WAIT | FUTEX_PRIVATE_FLAG)
+#define FUTEX_WAKE_PRIVATE     (FUTEX_WAKE | FUTEX_PRIVATE_FLAG)
+#define FUTEX_REQUEUE_PRIVATE  (FUTEX_REQUEUE | FUTEX_PRIVATE_FLAG)
+#define FUTEX_CMP_REQUEUE_PRIVATE (FUTEX_CMP_REQUEUE | FUTEX_PRIVATE_FLAG)
+#define FUTEX_WAKE_OP_PRIVATE  (FUTEX_WAKE_OP | FUTEX_PRIVATE_FLAG)
+#define FUTEX_LOCK_PI_PRIVATE  (FUTEX_LOCK_PI | FUTEX_PRIVATE_FLAG)
+#define FUTEX_UNLOCK_PI_PRIVATE        (FUTEX_UNLOCK_PI | FUTEX_PRIVATE_FLAG)
+#define FUTEX_TRYLOCK_PI_PRIVATE (FUTEX_TRYLOCK_PI | FUTEX_PRIVATE_FLAG)
+
 /*
  * Support for robust futexes: the kernel cleans up held futexes at
  * thread exit time.
@@ -114,8 +126,18 @@ handle_futex_death(u32 __user *uaddr, struct task_struct *curr, int pi);
  * Don't rearrange members without looking at hash_futex().
  *
  * offset is aligned to a multiple of sizeof(u32) (== 4) by definition.
- * We set bit 0 to indicate if it's an inode-based key.
- */
+ * We use the two low order bits of offset to tell what is the kind of key :
+ *  00 : Private process futex (PTHREAD_PROCESS_PRIVATE)
+ *       (no reference on an inode or mm)
+ *  01 : Shared futex (PTHREAD_PROCESS_SHARED)
+ *     mapped on a file (reference on the underlying inode)
+ *  10 : Shared futex (PTHREAD_PROCESS_SHARED)
+ *       (but private mapping on an mm, and reference taken on it)
+*/
+
+#define FUT_OFF_INODE    1 /* We set bit 0 if key has a reference on inode */
+#define FUT_OFF_MMSHARED 2 /* We set bit 1 if key has a reference on mm */
+
 union futex_key {
        u32 __user *uaddr;
        struct {
@@ -134,7 +156,8 @@ union futex_key {
                int offset;
        } both;
 };
-int get_futex_key(u32 __user *uaddr, union futex_key *key);
+int get_futex_key(u32 __user *uaddr, struct rw_semaphore *shared,
+                 union futex_key *key);
 void get_futex_key_refs(union futex_key *key);
 void drop_futex_key_refs(union futex_key *key);
 
index 4a60ef55dab4e7bccc8cf0111445aa9a0e9b952b..b7ce15c67e324b468d13599d47df7423adcd69d4 100644 (file)
@@ -16,6 +16,9 @@
  *  Copyright (C) 2006 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
  *  Copyright (C) 2006 Timesys Corp., Thomas Gleixner <tglx@timesys.com>
  *
+ *  PRIVATE futexes by Eric Dumazet
+ *  Copyright (C) 2007 Eric Dumazet <dada1@cosmosbay.com>
+ *
  *  Thanks to Ben LaHaise for yelling "hashed waitqueues" loudly
  *  enough at me, Linus for the original (flawed) idea, Matthew
  *  Kirkwood for proof-of-concept implementation.
@@ -150,19 +153,26 @@ static inline int match_futex(union futex_key *key1, union futex_key *key2)
                && key1->both.offset == key2->both.offset);
 }
 
-/*
- * Get parameters which are the keys for a futex.
+/**
+ * get_futex_key - Get parameters which are the keys for a futex.
+ * @uaddr: virtual address of the futex
+ * @shared: NULL for a PROCESS_PRIVATE futex,
+ *     &current->mm->mmap_sem for a PROCESS_SHARED futex
+ * @key: address where result is stored.
+ *
+ * Returns a negative error code or 0
+ * The key words are stored in *key on success.
  *
  * For shared mappings, it's (page->index, vma->vm_file->f_path.dentry->d_inode,
  * offset_within_page).  For private mappings, it's (uaddr, current->mm).
  * We can usually work out the index without swapping in the page.
  *
- * Returns: 0, or negative error code.
- * The key words are stored in *key on success.
- *
- * Should be called with &current->mm->mmap_sem but NOT any spinlocks.
+ * fshared is NULL for PROCESS_PRIVATE futexes
+ * For other futexes, it points to &current->mm->mmap_sem and
+ * caller must have taken the reader lock. but NOT any spinlocks.
  */
-int get_futex_key(u32 __user *uaddr, union futex_key *key)
+int get_futex_key(u32 __user *uaddr, struct rw_semaphore *fshared,
+                 union futex_key *key)
 {
        unsigned long address = (unsigned long)uaddr;
        struct mm_struct *mm = current->mm;
@@ -174,10 +184,24 @@ int get_futex_key(u32 __user *uaddr, union futex_key *key)
         * The futex address must be "naturally" aligned.
         */
        key->both.offset = address % PAGE_SIZE;
-       if (unlikely((key->both.offset % sizeof(u32)) != 0))
+       if (unlikely((address % sizeof(u32)) != 0))
                return -EINVAL;
        address -= key->both.offset;
 
+       /*
+        * PROCESS_PRIVATE futexes are fast.
+        * As the mm cannot disappear under us and the 'key' only needs
+        * virtual address, we dont even have to find the underlying vma.
+        * Note : We do have to check 'uaddr' is a valid user address,
+        *        but access_ok() should be faster than find_vma()
+        */
+       if (!fshared) {
+               if (unlikely(!access_ok(VERIFY_WRITE, uaddr, sizeof(u32))))
+                       return -EFAULT;
+               key->private.mm = mm;
+               key->private.address = address;
+               return 0;
+       }
        /*
         * The futex is hashed differently depending on whether
         * it's in a shared or private mapping.  So check vma first.
@@ -205,6 +229,7 @@ int get_futex_key(u32 __user *uaddr, union futex_key *key)
         * mappings of _writable_ handles.
         */
        if (likely(!(vma->vm_flags & VM_MAYSHARE))) {
+               key->both.offset |= FUT_OFF_MMSHARED; /* reference taken on mm */
                key->private.mm = mm;
                key->private.address = address;
                return 0;
@@ -214,7 +239,7 @@ int get_futex_key(u32 __user *uaddr, union futex_key *key)
         * Linear file mappings are also simple.
         */
        key->shared.inode = vma->vm_file->f_path.dentry->d_inode;
-       key->both.offset++; /* Bit 0 of offset indicates inode-based key. */
+       key->both.offset |= FUT_OFF_INODE; /* inode-based key. */
        if (likely(!(vma->vm_flags & VM_NONLINEAR))) {
                key->shared.pgoff = (((address - vma->vm_start) >> PAGE_SHIFT)
                                     + vma->vm_pgoff);
@@ -242,16 +267,18 @@ EXPORT_SYMBOL_GPL(get_futex_key);
  * Take a reference to the resource addressed by a key.
  * Can be called while holding spinlocks.
  *
- * NOTE: mmap_sem MUST be held between get_futex_key() and calling this
- * function, if it is called at all.  mmap_sem keeps key->shared.inode valid.
  */
 inline void get_futex_key_refs(union futex_key *key)
 {
-       if (key->both.ptr != 0) {
-               if (key->both.offset & 1)
+       if (key->both.ptr == 0)
+               return;
+       switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
+               case FUT_OFF_INODE:
                        atomic_inc(&key->shared.inode->i_count);
-               else
+                       break;
+               case FUT_OFF_MMSHARED:
                        atomic_inc(&key->private.mm->mm_count);
+                       break;
        }
 }
 EXPORT_SYMBOL_GPL(get_futex_key_refs);
@@ -262,11 +289,15 @@ EXPORT_SYMBOL_GPL(get_futex_key_refs);
  */
 void drop_futex_key_refs(union futex_key *key)
 {
-       if (key->both.ptr != 0) {
-               if (key->both.offset & 1)
+       if (key->both.ptr == 0)
+               return;
+       switch (key->both.offset & (FUT_OFF_INODE|FUT_OFF_MMSHARED)) {
+               case FUT_OFF_INODE:
                        iput(key->shared.inode);
-               else
+                       break;
+               case FUT_OFF_MMSHARED:
                        mmdrop(key->private.mm);
+                       break;
        }
 }
 EXPORT_SYMBOL_GPL(drop_futex_key_refs);
@@ -283,28 +314,38 @@ static inline int get_futex_value_locked(u32 *dest, u32 __user *from)
 }
 
 /*
- * Fault handling. Called with current->mm->mmap_sem held.
+ * Fault handling.
+ * if fshared is non NULL, current->mm->mmap_sem is already held
  */
-static int futex_handle_fault(unsigned long address, int attempt)
+static int futex_handle_fault(unsigned long address,
+                             struct rw_semaphore *fshared, int attempt)
 {
        struct vm_area_struct * vma;
        struct mm_struct *mm = current->mm;
+       int ret = -EFAULT;
 
-       if (attempt > 2 || !(vma = find_vma(mm, address)) ||
-           vma->vm_start > address || !(vma->vm_flags & VM_WRITE))
-               return -EFAULT;
+       if (attempt > 2)
+               return ret;
 
-       switch (handle_mm_fault(mm, vma, address, 1)) {
-       case VM_FAULT_MINOR:
-               current->min_flt++;
-               break;
-       case VM_FAULT_MAJOR:
-               current->maj_flt++;
-               break;
-       default:
-               return -EFAULT;
+       if (!fshared)
+               down_read(&mm->mmap_sem);
+       vma = find_vma(mm, address);
+       if (vma && address >= vma->vm_start &&
+           (vma->vm_flags & VM_WRITE)) {
+               switch (handle_mm_fault(mm, vma, address, 1)) {
+               case VM_FAULT_MINOR:
+                       ret = 0;
+                       current->min_flt++;
+                       break;
+               case VM_FAULT_MAJOR:
+                       ret = 0;
+                       current->maj_flt++;
+                       break;
+               }
        }
-       return 0;
+       if (!fshared)
+               up_read(&mm->mmap_sem);
+       return ret;
 }
 
 /*
@@ -647,7 +688,8 @@ double_lock_hb(struct futex_hash_bucket *hb1, struct futex_hash_bucket *hb2)
  * Wake up all waiters hashed on the physical page that is mapped
  * to this virtual address:
  */
-static int futex_wake(u32 __user *uaddr, int nr_wake)
+static int futex_wake(u32 __user *uaddr, struct rw_semaphore *fshared,
+                     int nr_wake)
 {
        struct futex_hash_bucket *hb;
        struct futex_q *this, *next;
@@ -655,9 +697,10 @@ static int futex_wake(u32 __user *uaddr, int nr_wake)
        union futex_key key;
        int ret;
 
-       down_read(&current->mm->mmap_sem);
+       if (fshared)
+               down_read(fshared);
 
-       ret = get_futex_key(uaddr, &key);
+       ret = get_futex_key(uaddr, fshared, &key);
        if (unlikely(ret != 0))
                goto out;
 
@@ -679,7 +722,8 @@ static int futex_wake(u32 __user *uaddr, int nr_wake)
 
        spin_unlock(&hb->lock);
 out:
-       up_read(&current->mm->mmap_sem);
+       if (fshared)
+               up_read(fshared);
        return ret;
 }
 
@@ -746,7 +790,9 @@ retry:
  * and requeue the next nr_requeue waiters following hashed on
  * one physical page to another physical page (PI-futex uaddr2)
  */
-static int futex_requeue_pi(u32 __user *uaddr1, u32 __user *uaddr2,
+static int futex_requeue_pi(u32 __user *uaddr1,
+                           struct rw_semaphore *fshared,
+                           u32 __user *uaddr2,
                            int nr_wake, int nr_requeue, u32 *cmpval)
 {
        union futex_key key1, key2;
@@ -765,12 +811,13 @@ retry:
        /*
         * First take all the futex related locks:
         */
-       down_read(&current->mm->mmap_sem);
+       if (fshared)
+               down_read(fshared);
 
-       ret = get_futex_key(uaddr1, &key1);
+       ret = get_futex_key(uaddr1, fshared, &key1);
        if (unlikely(ret != 0))
                goto out;
-       ret = get_futex_key(uaddr2, &key2);
+       ret = get_futex_key(uaddr2, fshared, &key2);
        if (unlikely(ret != 0))
                goto out;
 
@@ -793,7 +840,8 @@ retry:
                         * If we would have faulted, release mmap_sem, fault
                         * it in and start all over again.
                         */
-                       up_read(&current->mm->mmap_sem);
+                       if (fshared)
+                               up_read(fshared);
 
                        ret = get_user(curval, uaddr1);
 
@@ -927,7 +975,8 @@ out_unlock:
                drop_futex_key_refs(&key1);
 
 out:
-       up_read(&current->mm->mmap_sem);
+       if (fshared)
+               up_read(fshared);
        return ret;
 }
 
@@ -936,7 +985,8 @@ out:
  * to this virtual address:
  */
 static int
-futex_wake_op(u32 __user *uaddr1, u32 __user *uaddr2,
+futex_wake_op(u32 __user *uaddr1, struct rw_semaphore *fshared,
+             u32 __user *uaddr2,
              int nr_wake, int nr_wake2, int op)
 {
        union futex_key key1, key2;
@@ -946,12 +996,13 @@ futex_wake_op(u32 __user *uaddr1, u32 __user *uaddr2,
        int ret, op_ret, attempt = 0;
 
 retryfull:
-       down_read(&current->mm->mmap_sem);
+       if (fshared)
+               down_read(fshared);
 
-       ret = get_futex_key(uaddr1, &key1);
+       ret = get_futex_key(uaddr1, fshared, &key1);
        if (unlikely(ret != 0))
                goto out;
-       ret = get_futex_key(uaddr2, &key2);
+       ret = get_futex_key(uaddr2, fshared, &key2);
        if (unlikely(ret != 0))
                goto out;
 
@@ -991,11 +1042,10 @@ retry:
                 * still holding the mmap_sem.
                 */
                if (attempt++) {
-                       if (futex_handle_fault((unsigned long)uaddr2,
-                                               attempt)) {
-                               ret = -EFAULT;
+                       ret = futex_handle_fault((unsigned long)uaddr2,
+                                               fshared, attempt);
+                       if (ret)
                                goto out;
-                       }
                        goto retry;
                }
 
@@ -1003,7 +1053,8 @@ retry:
                 * If we would have faulted, release mmap_sem,
                 * fault it in and start all over again.
                 */
-               up_read(&current->mm->mmap_sem);
+               if (fshared)
+                       up_read(fshared);
 
                ret = get_user(dummy, uaddr2);
                if (ret)
@@ -1040,7 +1091,8 @@ retry:
        if (hb1 != hb2)
                spin_unlock(&hb2->lock);
 out:
-       up_read(&current->mm->mmap_sem);
+       if (fshared)
+               up_read(fshared);
        return ret;
 }
 
@@ -1048,7 +1100,8 @@ out:
  * Requeue all waiters hashed on one physical page to another
  * physical page.
  */
-static int futex_requeue(u32 __user *uaddr1, u32 __user *uaddr2,
+static int futex_requeue(u32 __user *uaddr1, struct rw_semaphore *fshared,
+                        u32 __user *uaddr2,
                         int nr_wake, int nr_requeue, u32 *cmpval)
 {
        union futex_key key1, key2;
@@ -1058,12 +1111,13 @@ static int futex_requeue(u32 __user *uaddr1, u32 __user *uaddr2,
        int ret, drop_count = 0;
 
  retry:
-       down_read(&current->mm->mmap_sem);
+       if (fshared)
+               down_read(fshared);
 
-       ret = get_futex_key(uaddr1, &key1);
+       ret = get_futex_key(uaddr1, fshared, &key1);
        if (unlikely(ret != 0))
                goto out;
-       ret = get_futex_key(uaddr2, &key2);
+       ret = get_futex_key(uaddr2, fshared, &key2);
        if (unlikely(ret != 0))
                goto out;
 
@@ -1086,7 +1140,8 @@ static int futex_requeue(u32 __user *uaddr1, u32 __user *uaddr2,
                         * If we would have faulted, release mmap_sem, fault
                         * it in and start all over again.
                         */
-                       up_read(&current->mm->mmap_sem);
+                       if (fshared)
+                               up_read(fshared);
 
                        ret = get_user(curval, uaddr1);
 
@@ -1139,7 +1194,8 @@ out_unlock:
                drop_futex_key_refs(&key1);
 
 out:
-       up_read(&current->mm->mmap_sem);
+       if (fshared)
+               up_read(fshared);
        return ret;
 }
 
@@ -1273,7 +1329,8 @@ static void unqueue_me_pi(struct futex_q *q)
  * The cur->mm semaphore must be  held, it is released at return of this
  * function.
  */
-static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
+static int fixup_pi_state_owner(u32 __user *uaddr, struct rw_semaphore *fshared,
+                               struct futex_q *q,
                                struct futex_hash_bucket *hb,
                                struct task_struct *curr)
 {
@@ -1300,7 +1357,8 @@ static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
 
        /* Unqueue and drop the lock */
        unqueue_me_pi(q);
-       up_read(&curr->mm->mmap_sem);
+       if (fshared)
+               up_read(fshared);
        /*
         * We own it, so we have to replace the pending owner
         * TID. This must be atomic as we have preserve the
@@ -1321,8 +1379,15 @@ static int fixup_pi_state_owner(u32 __user *uaddr, struct futex_q *q,
        return ret;
 }
 
+/*
+ * In case we must use restart_block to restart a futex_wait,
+ * we encode in the 'arg3' shared capability
+ */
+#define ARG3_SHARED  1
+
 static long futex_wait_restart(struct restart_block *restart);
-static int futex_wait(u32 __user *uaddr, u32 val, ktime_t *abs_time)
+static int futex_wait(u32 __user *uaddr, struct rw_semaphore *fshared,
+                     u32 val, ktime_t *abs_time)
 {
        struct task_struct *curr = current;
        DECLARE_WAITQUEUE(wait, curr);
@@ -1335,9 +1400,10 @@ static int futex_wait(u32 __user *uaddr, u32 val, ktime_t *abs_time)
 
        q.pi_state = NULL;
  retry:
-       down_read(&curr->mm->mmap_sem);
+       if (fshared)
+               down_read(fshared);
 
-       ret = get_futex_key(uaddr, &q.key);
+       ret = get_futex_key(uaddr, fshared, &q.key);
        if (unlikely(ret != 0))
                goto out_release_sem;
 
@@ -1360,8 +1426,8 @@ static int futex_wait(u32 __user *uaddr, u32 val, ktime_t *abs_time)
         * a wakeup when *uaddr != val on entry to the syscall.  This is
         * rare, but normal.
         *
-        * We hold the mmap semaphore, so the mapping cannot have changed
-        * since we looked it up in get_futex_key.
+        * for shared futexes, we hold the mmap semaphore, so the mapping
+        * cannot have changed since we looked it up in get_futex_key.
         */
        ret = get_futex_value_locked(&uval, uaddr);
 
@@ -1372,7 +1438,8 @@ static int futex_wait(u32 __user *uaddr, u32 val, ktime_t *abs_time)
                 * If we would have faulted, release mmap_sem, fault it in and
                 * start all over again.
                 */
-               up_read(&curr->mm->mmap_sem);
+               if (fshared)
+                       up_read(fshared);
 
                ret = get_user(uval, uaddr);
 
@@ -1399,7 +1466,8 @@ static int futex_wait(u32 __user *uaddr, u32 val, ktime_t *abs_time)
         * Now the futex is queued and we have checked the data, we
         * don't want to hold mmap_sem while we sleep.
         */
-       up_read(&curr->mm->mmap_sem);
+       if (fshared)
+               up_read(fshared);
 
        /*
         * There might have been scheduling since the queue_me(), as we
@@ -1469,7 +1537,8 @@ static int futex_wait(u32 __user *uaddr, u32 val, ktime_t *abs_time)
                else
                        ret = rt_mutex_timed_lock(lock, to, 1);
 
-               down_read(&curr->mm->mmap_sem);
+               if (fshared)
+                       down_read(fshared);
                spin_lock(q.lock_ptr);
 
                /*
@@ -1486,7 +1555,8 @@ static int futex_wait(u32 __user *uaddr, u32 val, ktime_t *abs_time)
 
                        /* mmap_sem and hash_bucket lock are unlocked at
                           return of this function */
-                       ret = fixup_pi_state_owner(uaddr, &q, hb, curr);
+                       ret = fixup_pi_state_owner(uaddr, fshared,
+                                                  &q, hb, curr);
                } else {
                        /*
                         * Catch the rare case, where the lock was released
@@ -1499,7 +1569,8 @@ static int futex_wait(u32 __user *uaddr, u32 val, ktime_t *abs_time)
                        }
                        /* Unqueue and drop the lock */
                        unqueue_me_pi(&q);
-                       up_read(&curr->mm->mmap_sem);
+                       if (fshared)
+                               up_read(fshared);
                }
 
                debug_rt_mutex_free_waiter(&q.waiter);
@@ -1528,6 +1599,9 @@ static int futex_wait(u32 __user *uaddr, u32 val, ktime_t *abs_time)
                restart->arg0 = (unsigned long)uaddr;
                restart->arg1 = (unsigned long)val;
                restart->arg2 = (unsigned long)abs_time;
+               restart->arg3 = 0;
+               if (fshared)
+                       restart->arg3 |= ARG3_SHARED;
                return -ERESTART_RESTARTBLOCK;
        }
 
@@ -1535,7 +1609,8 @@ static int futex_wait(u32 __user *uaddr, u32 val, ktime_t *abs_time)
        queue_unlock(&q, hb);
 
  out_release_sem:
-       up_read(&curr->mm->mmap_sem);
+       if (fshared)
+               up_read(fshared);
        return ret;
 }
 
@@ -1545,9 +1620,12 @@ static long futex_wait_restart(struct restart_block *restart)
        u32 __user *uaddr = (u32 __user *)restart->arg0;
        u32 val = (u32)restart->arg1;
        ktime_t *abs_time = (ktime_t *)restart->arg2;
+       struct rw_semaphore *fshared = NULL;
 
        restart->fn = do_no_restart_syscall;
-       return (long)futex_wait(uaddr, val, abs_time);
+       if (restart->arg3 & ARG3_SHARED)
+               fshared = &current->mm->mmap_sem;
+       return (long)futex_wait(uaddr, fshared, val, abs_time);
 }
 
 
@@ -1602,8 +1680,8 @@ static void set_pi_futex_owner(struct futex_hash_bucket *hb,
  * if there are waiters then it will block, it does PI, etc. (Due to
  * races the kernel might see a 0 value of the futex too.)
  */
-static int futex_lock_pi(u32 __user *uaddr, int detect, ktime_t *time,
-                        int trylock)
+static int futex_lock_pi(u32 __user *uaddr, struct rw_semaphore *fshared,
+                        int detect, ktime_t *time, int trylock)
 {
        struct hrtimer_sleeper timeout, *to = NULL;
        struct task_struct *curr = current;
@@ -1624,9 +1702,10 @@ static int futex_lock_pi(u32 __user *uaddr, int detect, ktime_t *time,
 
        q.pi_state = NULL;
  retry:
-       down_read(&curr->mm->mmap_sem);
+       if (fshared)
+               down_read(fshared);
 
-       ret = get_futex_key(uaddr, &q.key);
+       ret = get_futex_key(uaddr, fshared, &q.key);
        if (unlikely(ret != 0))
                goto out_release_sem;
 
@@ -1747,7 +1826,8 @@ static int futex_lock_pi(u32 __user *uaddr, int detect, ktime_t *time,
         * Now the futex is queued and we have checked the data, we
         * don't want to hold mmap_sem while we sleep.
         */
-       up_read(&curr->mm->mmap_sem);
+       if (fshared)
+               up_read(fshared);
 
        WARN_ON(!q.pi_state);
        /*
@@ -1761,7 +1841,8 @@ static int futex_lock_pi(u32 __user *uaddr, int detect, ktime_t *time,
                ret = ret ? 0 : -EWOULDBLOCK;
        }
 
-       down_read(&curr->mm->mmap_sem);
+       if (fshared)
+               down_read(fshared);
        spin_lock(q.lock_ptr);
 
        /*
@@ -1770,7 +1851,7 @@ static int futex_lock_pi(u32 __user *uaddr, int detect, ktime_t *time,
         */
        if (!ret && q.pi_state->owner != curr)
                /* mmap_sem is unlocked at return of this function */
-               ret = fixup_pi_state_owner(uaddr, &q, hb, curr);
+               ret = fixup_pi_state_owner(uaddr, fshared, &q, hb, curr);
        else {
                /*
                 * Catch the rare case, where the lock was released
@@ -1783,7 +1864,8 @@ static int futex_lock_pi(u32 __user *uaddr, int detect, ktime_t *time,
                }
                /* Unqueue and drop the lock */
                unqueue_me_pi(&q);
-               up_read(&curr->mm->mmap_sem);
+               if (fshared)
+                       up_read(fshared);
        }
 
        if (!detect && ret == -EDEADLK && 0)
@@ -1795,7 +1877,8 @@ static int futex_lock_pi(u32 __user *uaddr, int detect, ktime_t *time,
        queue_unlock(&q, hb);
 
  out_release_sem:
-       up_read(&curr->mm->mmap_sem);
+       if (fshared)
+               up_read(fshared);
        return ret;
 
  uaddr_faulted:
@@ -1806,15 +1889,16 @@ static int futex_lock_pi(u32 __user *uaddr, int detect, ktime_t *time,
         * still holding the mmap_sem.
         */
        if (attempt++) {
-               if (futex_handle_fault((unsigned long)uaddr, attempt)) {
-                       ret = -EFAULT;
+               ret = futex_handle_fault((unsigned long)uaddr, fshared,
+                                        attempt);
+               if (ret)
                        goto out_unlock_release_sem;
-               }
                goto retry_locked;
        }
 
        queue_unlock(&q, hb);
-       up_read(&curr->mm->mmap_sem);
+       if (fshared)
+               up_read(fshared);
 
        ret = get_user(uval, uaddr);
        if (!ret && (uval != -EFAULT))
@@ -1828,7 +1912,7 @@ static int futex_lock_pi(u32 __user *uaddr, int detect, ktime_t *time,
  * This is the in-kernel slowpath: we look up the PI state (if any),
  * and do the rt-mutex unlock.
  */
-static int futex_unlock_pi(u32 __user *uaddr)
+static int futex_unlock_pi(u32 __user *uaddr, struct rw_semaphore *fshared)
 {
        struct futex_hash_bucket *hb;
        struct futex_q *this, *next;
@@ -1848,9 +1932,10 @@ retry:
        /*
         * First take all the futex related locks:
         */
-       down_read(&current->mm->mmap_sem);
+       if (fshared)
+               down_read(fshared);
 
-       ret = get_futex_key(uaddr, &key);
+       ret = get_futex_key(uaddr, fshared, &key);
        if (unlikely(ret != 0))
                goto out;
 
@@ -1909,7 +1994,8 @@ retry_locked:
 out_unlock:
        spin_unlock(&hb->lock);
 out:
-       up_read(&current->mm->mmap_sem);
+       if (fshared)
+               up_read(fshared);
 
        return ret;
 
@@ -1921,15 +2007,16 @@ pi_faulted:
         * still holding the mmap_sem.
         */
        if (attempt++) {
-               if (futex_handle_fault((unsigned long)uaddr, attempt)) {
-                       ret = -EFAULT;
+               ret = futex_handle_fault((unsigned long)uaddr, fshared,
+                                        attempt);
+               if (ret)
                        goto out_unlock;
-               }
                goto retry_locked;
        }
 
        spin_unlock(&hb->lock);
-       up_read(&current->mm->mmap_sem);
+       if (fshared)
+               up_read(fshared);
 
        ret = get_user(uval, uaddr);
        if (!ret && (uval != -EFAULT))
@@ -1981,6 +2068,7 @@ static int futex_fd(u32 __user *uaddr, int signal)
        struct futex_q *q;
        struct file *filp;
        int ret, err;
+       struct rw_semaphore *fshared;
        static unsigned long printk_interval;
 
        if (printk_timed_ratelimit(&printk_interval, 60 * 60 * 1000)) {
@@ -2022,11 +2110,12 @@ static int futex_fd(u32 __user *uaddr, int signal)
        }
        q->pi_state = NULL;
 
-       down_read(&current->mm->mmap_sem);
-       err = get_futex_key(uaddr, &q->key);
+       fshared = &current->mm->mmap_sem;
+       down_read(fshared);
+       err = get_futex_key(uaddr, fshared, &q->key);
 
        if (unlikely(err != 0)) {
-               up_read(&current->mm->mmap_sem);
+               up_read(fshared);
                kfree(q);
                goto error;
        }
@@ -2038,7 +2127,7 @@ static int futex_fd(u32 __user *uaddr, int signal)
        filp->private_data = q;
 
        queue_me(q, ret, filp);
-       up_read(&current->mm->mmap_sem);
+       up_read(fshared);
 
        /* Now we map fd to filp, so userspace can access it */
        fd_install(ret, filp);
@@ -2167,7 +2256,7 @@ retry:
                 */
                if (!pi) {
                        if (uval & FUTEX_WAITERS)
-                               futex_wake(uaddr, 1);
+                               futex_wake(uaddr, &curr->mm->mmap_sem, 1);
                }
        }
        return 0;
@@ -2223,7 +2312,8 @@ void exit_robust_list(struct task_struct *curr)
                return;
 
        if (pending)
-               handle_futex_death((void __user *)pending + futex_offset, curr, pip);
+               handle_futex_death((void __user *)pending + futex_offset,
+                                  curr, pip);
 
        while (entry != &head->list) {
                /*
@@ -2253,38 +2343,43 @@ long do_futex(u32 __user *uaddr, int op, u32 val, ktime_t *timeout,
                u32 __user *uaddr2, u32 val2, u32 val3)
 {
        int ret;
+       int cmd = op & FUTEX_CMD_MASK;
+       struct rw_semaphore *fshared = NULL;
+
+       if (!(op & FUTEX_PRIVATE_FLAG))
+               fshared = &current->mm->mmap_sem;
 
-       switch (op) {
+       switch (cmd) {
        case FUTEX_WAIT:
-               ret = futex_wait(uaddr, val, timeout);
+               ret = futex_wait(uaddr, fshared, val, timeout);
                break;
        case FUTEX_WAKE:
-               ret = futex_wake(uaddr, val);
+               ret = futex_wake(uaddr, fshared, val);
                break;
        case FUTEX_FD:
                /* non-zero val means F_SETOWN(getpid()) & F_SETSIG(val) */
                ret = futex_fd(uaddr, val);
                break;
        case FUTEX_REQUEUE:
-               ret = futex_requeue(uaddr, uaddr2, val, val2, NULL);
+               ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, NULL);
                break;
        case FUTEX_CMP_REQUEUE:
-               ret = futex_requeue(uaddr, uaddr2, val, val2, &val3);
+               ret = futex_requeue(uaddr, fshared, uaddr2, val, val2, &val3);
                break;
        case FUTEX_WAKE_OP:
-               ret = futex_wake_op(uaddr, uaddr2, val, val2, val3);
+               ret = futex_wake_op(uaddr, fshared, uaddr2, val, val2, val3);
                break;
        case FUTEX_LOCK_PI:
-               ret = futex_lock_pi(uaddr, val, timeout, 0);
+               ret = futex_lock_pi(uaddr, fshared, val, timeout, 0);
                break;
        case FUTEX_UNLOCK_PI:
-               ret = futex_unlock_pi(uaddr);
+               ret = futex_unlock_pi(uaddr, fshared);
                break;
        case FUTEX_TRYLOCK_PI:
-               ret = futex_lock_pi(uaddr, 0, timeout, 1);
+               ret = futex_lock_pi(uaddr, fshared, 0, timeout, 1);
                break;
        case FUTEX_CMP_REQUEUE_PI:
-               ret = futex_requeue_pi(uaddr, uaddr2, val, val2, &val3);
+               ret = futex_requeue_pi(uaddr, fshared, uaddr2, val, val2, &val3);
                break;
        default:
                ret = -ENOSYS;
@@ -2300,23 +2395,24 @@ asmlinkage long sys_futex(u32 __user *uaddr, int op, u32 val,
        struct timespec ts;
        ktime_t t, *tp = NULL;
        u32 val2 = 0;
+       int cmd = op & FUTEX_CMD_MASK;
 
-       if (utime && (op == FUTEX_WAIT || op == FUTEX_LOCK_PI)) {
+       if (utime && (cmd == FUTEX_WAIT || cmd == FUTEX_LOCK_PI)) {
                if (copy_from_user(&ts, utime, sizeof(ts)) != 0)
                        return -EFAULT;
                if (!timespec_valid(&ts))
                        return -EINVAL;
 
                t = timespec_to_ktime(ts);
-               if (op == FUTEX_WAIT)
+               if (cmd == FUTEX_WAIT)
                        t = ktime_add(ktime_get(), t);
                tp = &t;
        }
        /*
-        * requeue parameter in 'utime' if op == FUTEX_REQUEUE.
+        * requeue parameter in 'utime' if cmd == FUTEX_REQUEUE.
         */
-       if (op == FUTEX_REQUEUE || op == FUTEX_CMP_REQUEUE
-           || op == FUTEX_CMP_REQUEUE_PI)
+       if (cmd == FUTEX_REQUEUE || cmd == FUTEX_CMP_REQUEUE
+           || cmd == FUTEX_CMP_REQUEUE_PI)
                val2 = (u32) (unsigned long) utime;
 
        return do_futex(uaddr, op, val, tp, uaddr2, val2, val3);