Re: mlockall(MCL_CURRENT) blocking infinitely

[Josef Bacik <josef@toxicpanda.com>]

On Wed, Nov 06, 2019 at 04:05:24PM +0100, Jan Kara wrote:
> On Wed 06-11-19 09:56:09, Josef Bacik wrote:
> > On Wed, Nov 06, 2019 at 02:45:43PM +0100, Robert Stupp wrote:
> > > On Wed, 2019-11-06 at 13:03 +0100, Jan Kara wrote:
> > > > On Tue 05-11-19 13:22:11, Johannes Weiner wrote:
> > > > > What I don't quite understand yet is why the fault path doesn't
> > > > > make
> > > > > progress eventually. We must drop the mmap_sem without changing the
> > > > > state in any way. How can we keep looping on the same page?
> > > >
> > > > That may be a slight suboptimality with Josef's patches. If the page
> > > > is marked as PageReadahead, we always drop mmap_sem if we can and
> > > > start
> > > > readahead without checking whether that makes sense or not in
> > > > do_async_mmap_readahead(). OTOH page_cache_async_readahead() then
> > > > clears
> > > > PageReadahead so the only way how I can see we could loop like this
> > > > is when
> > > > file->ra->ra_pages is 0. Not sure if that's what's happening through.
> > > > We'd
> > > > need to find which of the paths in filemap_fault() calls
> > > > maybe_unlock_mmap_for_io() to tell more.
> > > 
> > > Yes, ra_pages==0
> > > Attached the dmesg + smaps outputs
> > > 
> > > 
> > 
> > Ah ok I see what's happening, __get_user_pages() returns 0 if we get an EBUSY
> > from faultin_page, and then __mm_populate does nend = nstart + ret * PAGE_SIZE,
> > which just leaves us where we are.
> > 
> > We need to handle the non-blocking and the locking separately in __mm_populate
> > so we know what's going on.  Jan's fix for the readahead thing is definitely
> > valid as well, but this will keep us from looping forever in other retry cases.
> 
> I don't think this will work. AFAICS faultin_page() just checks whether
> 'nonblocking' is != NULL but doesn't ever look at its value... Honestly the
> whole interface is rather weird like lots of things around gup().
> 

Oh what the hell, yeah this is super bonkers.  The whole fault path probably
should be cleaned up to handle retry better.  This will do the trick I think?

Josef

diff --git a/mm/gup.c b/mm/gup.c
index 8f236a335ae9..2468789298e6 100644
--- a/mm/gup.c
+++ b/mm/gup.c
@@ -628,7 +628,7 @@ static int faultin_page(struct task_struct *tsk, struct vm_area_struct *vma,
 		fault_flags |= FAULT_FLAG_WRITE;
 	if (*flags & FOLL_REMOTE)
 		fault_flags |= FAULT_FLAG_REMOTE;
-	if (nonblocking)
+	if (nonblocking && *nonblocking != 0)
 		fault_flags |= FAULT_FLAG_ALLOW_RETRY;
 	if (*flags & FOLL_NOWAIT)
 		fault_flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT;
@@ -1237,6 +1237,7 @@ int __mm_populate(unsigned long start, unsigned long len, int ignore_errors)
 	unsigned long end, nstart, nend;
 	struct vm_area_struct *vma = NULL;
 	int locked = 0;
+	int nonblocking = 1;
 	long ret = 0;
 
 	end = start + len;
@@ -1268,7 +1269,7 @@ int __mm_populate(unsigned long start, unsigned long len, int ignore_errors)
 		 * double checks the vma flags, so that it won't mlock pages
 		 * if the vma was already munlocked.
 		 */
-		ret = populate_vma_page_range(vma, nstart, nend, &locked);
+		ret = populate_vma_page_range(vma, nstart, nend, &nonblocking);
 		if (ret < 0) {
 			if (ignore_errors) {
 				ret = 0;
@@ -1276,6 +1277,14 @@ int __mm_populate(unsigned long start, unsigned long len, int ignore_errors)
 			}
 			break;
 		}
+
+		/*
+		 * We dropped the mmap_sem, so we need to re-lock, and the next
+		 * loop around we won't drop because nonblocking is now 0.
+		 */
+		if (!nonblocking)
+			locked = 0;
+
 		nend = nstart + ret * PAGE_SIZE;
 		ret = 0;
 	}