mlock: downgrade mmap sem while populating mlocked regions

We need to hold the mmap_sem for write to initiatate mlock()/munlock()
because we may need to merge/split vmas.  However, this can lead to very
long lock hold times attempting to fault in a large memory region to mlock
it into memory.  This can hold off other faults against the mm
[multithreaded tasks] and other scans of the mm, such as via /proc.  To
alleviate this, downgrade the mmap_sem to read mode during the population
of the region for locking.  This is especially the case if we need to
reclaim memory to lock down the region.  We [probably?] don't need to do
this for unlocking as all of the pages should be resident--they're already
mlocked.

Now, the caller's of the mlock functions [mlock_fixup() and
mlock_vma_pages_range()] expect the mmap_sem to be returned in write mode.
 Changing all callers appears to be way too much effort at this point.
So, restore write mode before returning.  Note that this opens a window
where the mmap list could change in a multithreaded process.  So, at least
for mlock_fixup(), where we could be called in a loop over multiple vmas,
we check that a vma still exists at the start address and that vma still
covers the page range [start,end).  If not, we return an error, -EAGAIN,
and let the caller deal with it.

Return -EAGAIN from mlock_vma_pages_range() function and mlock_fixup() if
the vma at 'start' disappears or changes so that the page range
[start,end) is no longer contained in the vma.  Again, let the caller deal
with it.  Looks like only sys_remap_file_pages() [via mmap_region()]
should actually care.

With this patch, I no longer see processes like ps(1) blocked for seconds
or minutes at a time waiting for a large [multiple gigabyte] region to be
locked down.  However, I occassionally see delays while unlocking or
unmapping a large mlocked region.  Should we also downgrade the mmap_sem
for the unlock path?
Signed-off-by: Lee Schermerhorn <lee.schermerhorn@hp.com>
Signed-off-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm/mlock.c

@@ -318,6 +318,7 @@ static void __munlock_vma_pages_range(struct vm_area_struct *vma,
 int mlock_vma_pages_range(struct vm_area_struct *vma,
 			unsigned long start, unsigned long end)
 {
+	struct mm_struct *mm = vma->vm_mm;
 	int nr_pages = (end - start) / PAGE_SIZE;
 	BUG_ON(!(vma->vm_flags & VM_LOCKED));
 
@@ -329,8 +330,19 @@ int mlock_vma_pages_range(struct vm_area_struct *vma,
 
 	if (!((vma->vm_flags & (VM_DONTEXPAND | VM_RESERVED)) ||
 			is_vm_hugetlb_page(vma) ||
-			vma == get_gate_vma(current)))
-		return __mlock_vma_pages_range(vma, start, end);
+			vma == get_gate_vma(current))) {
+		downgrade_write(&mm->mmap_sem);
+		nr_pages = __mlock_vma_pages_range(vma, start, end);
+
+		up_read(&mm->mmap_sem);
+		/* vma can change or disappear */
+		down_write(&mm->mmap_sem);
+		vma = find_vma(mm, start);
+		/* non-NULL vma must contain @start, but need to check @end */
+		if (!vma ||  end > vma->vm_end)
+			return -EAGAIN;
+		return nr_pages;
+	}
 
 	/*
 	 * User mapped kernel pages or huge pages:
@@ -424,13 +436,41 @@ static int mlock_fixup(struct vm_area_struct *vma, struct vm_area_struct **prev,
 	vma->vm_flags = newflags;
 
 	if (lock) {
+		/*
+		 * mmap_sem is currently held for write.  Downgrade the write
+		 * lock to a read lock so that other faults, mmap scans, ...
+		 * while we fault in all pages.
+		 */
+		downgrade_write(&mm->mmap_sem);
+
 		ret = __mlock_vma_pages_range(vma, start, end);
 		if (ret > 0) {
 			mm->locked_vm -= ret;
 			ret = 0;
 		}
-	} else
+		/*
+		 * Need to reacquire mmap sem in write mode, as our callers
+		 * expect this.  We have no support for atomically upgrading
+		 * a sem to write, so we need to check for ranges while sem
+		 * is unlocked.
+		 */
+		up_read(&mm->mmap_sem);
+		/* vma can change or disappear */
+		down_write(&mm->mmap_sem);
+		*prev = find_vma(mm, start);
+		/* non-NULL *prev must contain @start, but need to check @end */
+		if (!(*prev) || end > (*prev)->vm_end)
+			ret = -EAGAIN;
+	} else {
+		/*
+		 * TODO:  for unlocking, pages will already be resident, so
+		 * we don't need to wait for allocations/reclaim/pagein, ...
+		 * However, unlocking a very large region can still take a
+		 * while.  Should we downgrade the semaphore for both lock
+		 * AND unlock ?
+		 */
 		__munlock_vma_pages_range(vma, start, end);
+	}
 
 out:
 	*prev = vma;