Re: [PATCH RESEND v6 1/9] pagemap: Introduce ->memory_failure()

On Tue, Aug 17, 2021 at 11:08:40PM -0700, Jane Chu wrote:

On 8/17/2021 10:43 PM, Jane Chu wrote:

quoted

More information -

On 8/16/2021 10:20 AM, Jane Chu wrote:

quoted

Hi, ShiYang,

So I applied the v6 patch series to my 5.14-rc3 as it's what you
indicated is what v6 was based at, and injected a hardware poison.

I'm seeing the same problem that was reported a while ago after the
poison was consumed - in the SIGBUS payload, the si_addr is missing:

** SIGBUS(7): canjmp=1, whichstep=0, **
** si_addr(0x(nil)), si_lsb(0xC), si_code(0x4, BUS_MCEERR_AR) **

The si_addr ought to be 0x7f6568000000 - the vaddr of the first page
in this case.

The failure came from here :

[PATCH RESEND v6 6/9] xfs: Implement ->notify_failure() for XFS

+static int
+xfs_dax_notify_failure(
...
+    if (!xfs_sb_version_hasrmapbt(&mp->m_sb)) {
+        xfs_warn(mp, "notify_failure() needs rmapbt enabled!");
+        return -EOPNOTSUPP;
+    }

I am not familiar with XFS, but I have a few questions I hope to get
answers -

1) What does it take and cost to make
    xfs_sb_version_hasrmapbt(&mp->m_sb) to return true?

mkfs.xfs -m rmapbt=1

quoted

2) For a running environment that fails the above check, is it
    okay to leave the poison handle in limbo and why?

3) If the above regression is not acceptable, any potential remedy?

How about moving the check to prior to the notifier registration?
And register only if the check is passed?  This seems better
than an alternative which is to fall back to the legacy memory_failure
handling in case the filesystem returns -EOPNOTSUPP.

"return -EOPNOTSUPP;" is the branching point where a future patch could
probe the (DRAM) buffer cache to bwrite the contents to restore the pmem
contents.  Right now the focus should be on landing the core code
changes without drawing any more NAKs from Dan.

--D

thanks,
-jane

quoted

thanks!
-jane

quoted

Something is not right...

thanks,
-jane


On 8/5/2021 6:17 PM, Jane Chu wrote:

quoted

The filesystem part of the pmem failure handling is at minimum built
on PAGE_SIZE granularity - an inheritance from general
memory_failure handling.  However, with Intel's DCPMEM
technology, the error blast
radius is no more than 256bytes, and might get smaller with future
hardware generation, also advanced atomic 64B write to clear the poison.
But I don't see any of that could be incorporated in, given that the
filesystem is notified a corruption with pfn, rather than an exact
address.

So I guess this question is also for Dan: how to avoid unnecessarily
repairing a PMD range for a 256B corrupt range going forward?

thanks,
-jane


On 7/30/2021 3:01 AM, Shiyang Ruan wrote:

quoted

When memory-failure occurs, we call this function which is implemented
by each kind of devices.  For the fsdax case, pmem device driver
implements it.  Pmem device driver will find out the
filesystem in which
the corrupted page located in.  And finally call filesystem handler to
deal with this error.

The filesystem will try to recover the corrupted data if necessary.