Re: [PATCH v6 1/2] mm,hwpoison: fix race with hugetlb page allocation

[Naoya Horiguchi <naoya.horiguchi@linux.dev>]

On Fri, Aug 13, 2021 at 03:07:20PM +0000, Luck, Tony wrote:
> I'm running the default case from my einj_mem_uc test. That just:
> 
> 1) allocates a page using:
> 
> 	mmap(NULL, pagesize, PROT_READ|PROT_WRITE, MAP_SHARED|MAP_ANON, -1, 0);
> 
> 2) fills the page with random data (to make sure it has been allocated, and that the kernel can't
> do KSM tricks to share this physical page with some other user).
> 
> 3) injects the error at a 1KB offset within the page.
> 
> 4) does a memory read of the poison address.
> 
> 
> >                            action_result(pfn, MF_MSG_UNKNOWN, MF_IGNORED);
> >   +                       dump_page(p, "hwpoison unknown page");
> >                            res = -EBUSY;
> >                            goto unlock_mutex;
> >                    }
> 
> I added that patch against upstream (v5.14-rc5).  Here's the dump. The "pfn" matches the physical address where I injected,
> and it has the hwpoison flag bit that was set early in memory_failure() --- so this is the right page.
> 
> [   79.368212] Memory failure: 0x623889: recovery action for unknown page: Ignored
> [   79.375525] page:0000000065ad9479 refcount:3 mapcount:1 mapping:00000000a4ac843b index:0x0 pfn:0x623889
> [   79.384909] memcg:ff40a569f2966000
> [   79.388313] aops:shmem_aops ino:4c00 dentry name:"dev/zero"
> [   79.393896] flags: 0x17ffffc088000c(uptodate|dirty|swapbacked|hwpoison|node=0|zone=2|lastcpupid=0x1fffff)
> [   79.403455] raw: 0017ffffc088000c 0000000000000000 dead000000000122 ff40a569f45a7160
> [   79.411191] raw: 0000000000000000 0000000000000000 0000000300000000 ff40a569f2966000
> [   79.418931] page dumped because: hwpoison unknown page

Thank you for your help.

This dump indicates that HWPoisonHandlable() returned false due to
the lack of PG_lru flag.  In older code before 5.13, get_any_page() does
retry with shake_page(), but does not since 5.13, which seems to me
the root cause of the issue. So my suggestion is to call shake_page()
when HWPoisonHandlable() is false.

Could you try checking that the following diff fixes the issue?
I could still have better fix (like inserting shake_page() to other
retry paths in get_any_page()), but the below is the minimum one.

diff --git a/mm/memory-failure.c b/mm/memory-failure.c
index 76cc53b2999a..3e770e4f259e 100644
--- a/mm/memory-failure.c
+++ b/mm/memory-failure.c
@@ -1146,7 +1146,7 @@ static int __get_hwpoison_page(struct page *page)
         * unexpected races caused by taking a page refcount.
         */
        if (!HWPoisonHandlable(head))
-               return 0;
+               return -EBUSY;

        if (PageTransHuge(head)) {
                /*
@@ -1199,9 +1199,14 @@ static int get_any_page(struct page *p, unsigned long flags)
                        }
                        goto out;
                } else if (ret == -EBUSY) {
-                       /* We raced with freeing huge page to buddy, retry. */
-                       if (pass++ < 3)
+                       /*
+                        * We raced with (possibly temporary) unhandlable
+                        * page, retry.
+                        */
+                       if (pass++ < 3) {
+                               shake_page(p, 1);
                                goto try_again;
+                       }
                        goto out;
                }
        }


Thanks,
Naoya Horiguchi