On Fri, Dec 07, 2018 at 12:20:36PM +0100, Michal Hocko wrote:

On Fri 07-12-18 08:16:15, Michal Hocko wrote:
[...]

quoted

Memcg v1 indeed doesn't have any dirty IO throttling and this is a
poor's man workaround. We still do not have that AFAIK and I do not know
of an elegant way around that. Fortunatelly we shouldn't have that many
GFP_KERNEL | __GFP_ACCOUNT allocations under page lock and we can work
around this specific one quite easily. I haven't tested this yet but the
following should work

diff --git a/mm/memory.c b/mm/memory.c
index 4ad2d293ddc2..59c98eeb0260 100644
--- a/mm/memory.c
+++ b/mm/memory.c

@@ -2993,6 +2993,16 @@ static vm_fault_t __do_fault(struct vm_fault *vmf)
 	struct vm_area_struct *vma = vmf->vma;
 	vm_fault_t ret;
 
+	/*
+	 * Preallocate pte before we take page_lock because this might lead to
+	 * deadlocks for memcg reclaim which waits for pages under writeback.
+	 */
+	if (!vmf->prealloc_pte) {
+		vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm>mm, vmf->address);
+		if (!vmf->prealloc_pte)
+			return VM_FAULT_OOM;
+	}
+
 	ret = vma->vm_ops->fault(vmf);
 	if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY |
 			    VM_FAULT_DONE_COW)))

This is too eager to allocate pte even when it is not really needed.
Jack has also pointed out that I am missing a write barrier. So here we
go with an updated patch. This is essentially what fault around code
does.

Makes sense to me, unfortunately we don't have a local reproducer to verify it
and we've disabled CONFIG_MEMCG_KMEM to workaround the problem.  Given the stack
I put, the patch should address the deadlock at least.

thanks,
-liubo