mm/mmu_notifier: allocate mmu_notifier in advance

While allocating mmu_notifier with parameter GFP_KERNEL, swap would start
to work in case of tight available memory.  Eventually, that would lead to
a deadlock while the swap deamon swaps anonymous pages.  It was caused by
commit e0f3c3f78da29b ("mm/mmu_notifier: init notifier if necessary").

  =================================
  [ INFO: inconsistent lock state ]
  3.7.0-rc1+ #518 Not tainted
  ---------------------------------
  inconsistent {RECLAIM_FS-ON-W} -> {IN-RECLAIM_FS-W} usage.
  kswapd0/35 [HC0[0]:SC0[0]:HE1:SE1] takes:
   (&mapping->i_mmap_mutex){+.+.?.}, at: page_referenced+0x9c/0x2e0
  {RECLAIM_FS-ON-W} state was registered at:
     mark_held_locks+0x86/0x150
     lockdep_trace_alloc+0x67/0xc0
     kmem_cache_alloc_trace+0x33/0x230
     do_mmu_notifier_register+0x87/0x180
     mmu_notifier_register+0x13/0x20
     kvm_dev_ioctl+0x428/0x510
     do_vfs_ioctl+0x98/0x570
     sys_ioctl+0x91/0xb0
     system_call_fastpath+0x16/0x1b
  irq event stamp: 825
  hardirqs last  enabled at (825): _raw_spin_unlock_irq+0x30/0x60
  hardirqs last disabled at (824): _raw_spin_lock_irq+0x19/0x80
  softirqs last  enabled at (0): copy_process+0x630/0x17c0
  softirqs last disabled at (0): (null)
  ...

Simply back out the above commit, which was a small performance
optimization.

Signed-off-by: Gavin Shan <shangw@linux.vnet.ibm.com>
Reported-by: Andrea Righi <andrea@betterlinux.com>
Tested-by: Andrea Righi <andrea@betterlinux.com>
Cc: Wanpeng Li <liwanp@linux.vnet.ibm.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Avi Kivity <avi@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com>
Cc: Sagi Grimberg <sagig@mellanox.co.il>
Cc: Haggai Eran <haggaie@mellanox.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

diff --git a/mm/mmu_notifier.c b/mm/mmu_notifier.c
index 479a1e751a73fb4781e1b856db60b2381909c9e6..8a5ac8c686b03480fae7d5b193a5c714c7c2397b 100644 (file)
--- a/mm/mmu_notifier.c
+++ b/mm/mmu_notifier.c
@@ -196,28 +196,28 @@ static int do_mmu_notifier_register(struct mmu_notifier *mn,
        BUG_ON(atomic_read(&mm->mm_users) <= 0);
 
        /*
-       * Verify that mmu_notifier_init() already run and the global srcu is
-       * initialized.
-       */
+        * Verify that mmu_notifier_init() already run and the global srcu is
+        * initialized.
+        */
        BUG_ON(!srcu.per_cpu_ref);
 
+       ret = -ENOMEM;
+       mmu_notifier_mm = kmalloc(sizeof(struct mmu_notifier_mm), GFP_KERNEL);
+       if (unlikely(!mmu_notifier_mm))
+               goto out;
+
        if (take_mmap_sem)
                down_write(&mm->mmap_sem);
        ret = mm_take_all_locks(mm);
        if (unlikely(ret))
-               goto out;
+               goto out_clean;
 
        if (!mm_has_notifiers(mm)) {
-               mmu_notifier_mm = kmalloc(sizeof(struct mmu_notifier_mm),
-                                       GFP_KERNEL);
-               if (unlikely(!mmu_notifier_mm)) {
-                       ret = -ENOMEM;
-                       goto out_of_mem;
-               }
                INIT_HLIST_HEAD(&mmu_notifier_mm->list);
                spin_lock_init(&mmu_notifier_mm->lock);
 
                mm->mmu_notifier_mm = mmu_notifier_mm;
+               mmu_notifier_mm = NULL;
        }
        atomic_inc(&mm->mm_count);
 
@@ -233,12 +233,12 @@ static int do_mmu_notifier_register(struct mmu_notifier *mn,
        hlist_add_head(&mn->hlist, &mm->mmu_notifier_mm->list);
        spin_unlock(&mm->mmu_notifier_mm->lock);
 
-out_of_mem:
        mm_drop_all_locks(mm);
-out:
+out_clean:
        if (take_mmap_sem)
                up_write(&mm->mmap_sem);
-
+       kfree(mmu_notifier_mm);
+out:
        BUG_ON(atomic_read(&mm->mm_users) <= 0);
        return ret;
 }