Hi Kirill,

Here is the result from the user:"This patch does appear fix the issue."

Thanks,

Alex (Bin) Xie

Hi Kirill,


Thanks for the patch. I have sent the patch to the user asking whether he can give it a try.


Regards,

Alex (Bin) Xie



On 2017-08-04 09:49 AM, Kirill A. Shutemov wrote:

On Thu, Aug 03, 2017 at 03:39:02PM -0700, Andrew Morton wrote:

(cc Kirill)

On Thu, 3 Aug 2017 12:35:28 -0400 axie <axie@amd.com> wrote:

Hi Andrew,


I got a report yesterday with "BUG: sleeping function called from
invalid context at kernel/locking/mutex.c"

I checked the relevant functions for the issue. Function
page_vma_mapped_walk did acquire spinlock. Later, in MMU notifier,
amdgpu_mn_invalidate_page called function mutex_lock, which triggered
the "bug".

Function page_vma_mapped_walk was introduced recently by you in commit
c7ab0d2fdc840266b39db94538f74207ec2afbf6 and
ace71a19cec5eb430207c3269d8a2683f0574306.

Would you advise how to proceed with this bug? Change
page_vma_mapped_walk not to use spinlock? Or change
amdgpu_mn_invalidate_page to use spinlock to meet the change, or
something else?

hm, as far as I can tell this was an unintended side-effect of
c7ab0d2fd ("mm: convert try_to_unmap_one() to use
page_vma_mapped_walk()").A Before that patch,
mmu_notifier_invalidate_page() was not called under page_table_lock.
After that patch, mmu_notifier_invalidate_page() is called under
page_table_lock.

Perhaps Kirill can suggest a fix?

Sorry for this.

What about the patch below?

A From f48dbcdd0ed83dee9a157062b7ca1e2915172678 Mon Sep 17 00:00:00 2001
From: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Date: Fri, 4 Aug 2017 16:37:26 +0300
Subject: [PATCH] rmap: do not call mmu_notifier_invalidate_page() under ptl

MMU notifiers can sleep, but in page_mkclean_one() we call
mmu_notifier_invalidate_page() under page table lock.

Let's instead use mmu_notifier_invalidate_range() outside
page_vma_mapped_walk() loop.

Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Fixes: c7ab0d2fdc84 ("mm: convert try_to_unmap_one() to use page_vma_mapped_walk()")
---
A mm/rmap.c | 21 +++++++++++++--------
A 1 file changed, 13 insertions(+), 8 deletions(-)

diff --git a/mm/rmap.c b/mm/rmap.c
index ced14f1af6dc..b4b711a82c01 100644
--- a/mm/rmap.c
+++ b/mm/rmap.c
@@ -852,10 +852,10 @@ static bool page_mkclean_one(struct page *page, struct vm_area_struct *vma,
A A A A A A A A A .flags = PVMW_SYNC,
A A A A A };
A A A A A int *cleaned = arg;
+A A A bool invalidation_needed = false;
A A A A A A while (page_vma_mapped_walk(&pvmw)) {
A A A A A A A A A int ret = 0;
-A A A A A A A address = pvmw.address;
A A A A A A A A A if (pvmw.pte) {
A A A A A A A A A A A A A pte_t entry;
A A A A A A A A A A A A A pte_t *pte = pvmw.pte;
@@ -863,11 +863,11 @@ static bool page_mkclean_one(struct page *page, struct vm_area_struct *vma,
A A A A A A A A A A A A A if (!pte_dirty(*pte) && !pte_write(*pte))
A A A A A A A A A A A A A A A A A continue;
A -A A A A A A A A A A A flush_cache_page(vma, address, pte_pfn(*pte));
-A A A A A A A A A A A entry = ptep_clear_flush(vma, address, pte);
+A A A A A A A A A A A flush_cache_page(vma, pvmw.address, pte_pfn(*pte));
+A A A A A A A A A A A entry = ptep_clear_flush(vma, pvmw.address, pte);
A A A A A A A A A A A A A entry = pte_wrprotect(entry);
A A A A A A A A A A A A A entry = pte_mkclean(entry);
-A A A A A A A A A A A set_pte_at(vma->vm_mm, address, pte, entry);
+A A A A A A A A A A A set_pte_at(vma->vm_mm, pvmw.address, pte, entry);
A A A A A A A A A A A A A ret = 1;
A A A A A A A A A } else {
A #ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
@@ -877,11 +877,11 @@ static bool page_mkclean_one(struct page *page, struct vm_area_struct *vma,
A A A A A A A A A A A A A if (!pmd_dirty(*pmd) && !pmd_write(*pmd))
A A A A A A A A A A A A A A A A A continue;
A -A A A A A A A A A A A flush_cache_page(vma, address, page_to_pfn(page));
-A A A A A A A A A A A entry = pmdp_huge_clear_flush(vma, address, pmd);
+A A A A A A A A A A A flush_cache_page(vma, pvmw.address, page_to_pfn(page));
+A A A A A A A A A A A entry = pmdp_huge_clear_flush(vma, pvmw.address, pmd);
A A A A A A A A A A A A A entry = pmd_wrprotect(entry);
A A A A A A A A A A A A A entry = pmd_mkclean(entry);
-A A A A A A A A A A A set_pmd_at(vma->vm_mm, address, pmd, entry);
+A A A A A A A A A A A set_pmd_at(vma->vm_mm, pvmw.address, pmd, entry);
A A A A A A A A A A A A A ret = 1;
A #else
A A A A A A A A A A A A A /* unexpected pmd-mapped page? */
@@ -890,11 +890,16 @@ static bool page_mkclean_one(struct page *page, struct vm_area_struct *vma,
A A A A A A A A A }
A A A A A A A A A A if (ret) {
-A A A A A A A A A A A mmu_notifier_invalidate_page(vma->vm_mm, address);
A A A A A A A A A A A A A (*cleaned)++;
+A A A A A A A A A A A invalidation_needed = true;
A A A A A A A A A }
A A A A A }
A +A A A if (invalidation_needed) {
+A A A A A A A mmu_notifier_invalidate_range(vma->vm_mm, address,
+A A A A A A A A A A A A A A A address + (1UL << compound_order(page)));
+A A A }
+
A A A A A return true;
A }
A