xfs
[Top] [All Lists]

Re: [PATCH] xfs: free the EFI entries from AIL on forced shutdown

To: Dave Chinner <david@xxxxxxxxxxxxx>
Subject: Re: [PATCH] xfs: free the EFI entries from AIL on forced shutdown
From: Mark Tinguely <tinguely@xxxxxxx>
Date: Thu, 20 Mar 2014 14:47:45 -0500
Cc: xfs@xxxxxxxxxxx
Delivered-to: xfs@xxxxxxxxxxx
In-reply-to: <20140317052951.GE7072@dastard>
References: <20140314163723.916178776@xxxxxxx> <20140317052951.GE7072@dastard>
User-agent: Mozilla/5.0 (X11; FreeBSD amd64; rv:9.0) Gecko/20120122 Thunderbird/9.0
On 03/17/14 00:29, Dave Chinner wrote:
On Fri, Mar 14, 2014 at 11:37:01AM -0500, Mark Tinguely wrote:
The extent free intention (EFI) and extent free done (EFD)
log items are in separate transactions. It is possible that
the EFI can be pushed to the AIL before a forced shutdown
where it gets stuck for following reasons:

  No EFD. If freeing the extent fails in xfs_bmap_finish() or
  xfs_recover_process_efi(), then the corresponding extent
  free done (EFD) entry will not be created.

You don't handle the xfs_bmap_finish() case properly - the patch
only addresses the case where the EFD was created and contains a
reference to the EFI. The only time we can have an EFI in the AIL
without an EFD pointing to it is if xfs_trans_reserve() call in
xfs_bmap_finish() fails due to a shutdown. That failure case
leaks the EFI reference that is supposed to be passed to the EFD...

  EFD IOP Abort processing. If xfs_trans_cancel() is called with
  an abort flag, or if the xfs_trans_commit() is called when the
  file system is in forced shutdown or if the log buffer write fails,
  then the EFD iop commands will not remove the EFI from the AIL.

Which they should, because after xfs_trans_get_efd(), the EFD owns
the reference to the EFI. Hence aborting the EFD should release the
EFI reference it owns.

Nod.


Index: b/fs/xfs/xfs_extfree_item.c
===================================================================
--- a/fs/xfs/xfs_extfree_item.c
+++ b/fs/xfs/xfs_extfree_item.c
@@ -420,8 +420,15 @@ STATIC void
  xfs_efd_item_unlock(
        struct xfs_log_item     *lip)
  {
-       if (lip->li_flags&  XFS_LI_ABORTED)
+       /*
+        * Clear the EFI if on AIL when aborting xfs_bmap_finish.
+        * The forced shutdown will force the log, so other EFDs
+        * should not be processed from the CIL.
+        */
+       if (lip->li_flags&  XFS_LI_ABORTED) {
+               xfs_efi_clear_ail(lip->li_ailp);
                xfs_efd_item_free(EFD_ITEM(lip));
+       }
  }

So, we abort one EFI, so we kill every EFI in the filesystem? We
want to be able to cancel and rollback transactions eventually, and
this will prevent us from being able to do that as it will kill EFIs
unrelated to the EFD being aborted.

AFAICT, all this needs to do is:

        if (!(lip->li_flags&  XFS_LI_ABORTED))
                return;

        xfs_efi_release(efdp->efd_efip, efdp->efd_format.efd_nextents);
        xfs_efd_item_free(efdp);

i.e. before freeing the EFD, release the reference to the EFI
that it was passed.

Nod.

@@ -439,10 +446,15 @@ xfs_efd_item_committed(
        struct xfs_efd_log_item *efdp = EFD_ITEM(lip);

        /*
-        * If we got a log I/O error, it's always the case that the LR with the
-        * EFI got unpinned and freed before the EFD got aborted.
+        * If we got a log I/O error and the EFI is also in this buffer, it
+        * will be unpinned and freed before the EFD got aborted. But the EFI
+        * is in an earlier transaction and could be on the AIL when the log
+        * I/O error happened for this EFD. In that case, manually remove the
+        * remaining EFIs from the AIL.
         */
-       if (!(lip->li_flags&  XFS_LI_ABORTED))
+       if (lip->li_flags&  XFS_LI_ABORTED)
+               xfs_efi_clear_ail(lip->li_ailp);
+       else
                xfs_efi_release(efdp->efd_efip, efdp->efd_format.efd_nextents);

Same here - we can simply always call xfs_efi_release() and we'll
end up doing the right thing w.r.t. the EFi reference count.

And if we do the right thing with the EFI reference count when
aborting the EFI (i.e. call __xfs_efi_release() rather than freeing
it) the aborting of transactions will correctly free all the
references to the EFI and remove them from the AIL.


This is fine if the EFI made it to the AIL before the log abort - which it may or may not have done.

Log errors will not force the log. The call to xfs_trans_committed_bulk() with the abort flag set will not place the EFI on the AIL (takes the iop_unpin/continue path).

Your proposal of changing xfs_efi_committed() from freeing the EFI to a call to __xfs_efi_release() and changing xfs_efd_committed() to xfs_efi_release() will do the right thing for the counters and removing both the EFI/EFD entries, but it will also want to remove the non-existent EFI entry from the AIL. I do not like adding the EFI to the AIL in the abort path.

The only case we
then have to handle specially is the xfs_trans_reserve failure in
xfs_bmap_finish(), where we need to release the EFI directly in the
error path.

any reason to not move the efd creation earlier to not special case it?

That gets rid of the "big hammer" error handling for normal runtime
shutdown that xfs_efi_clear_ail(). i.e. we need to fix the reference
count handling, not work around it.


Index: b/fs/xfs/xfs_log_recover.c
===================================================================
--- a/fs/xfs/xfs_log_recover.c
+++ b/fs/xfs/xfs_log_recover.c
@@ -3634,20 +3634,23 @@ xlog_recover_process_data(
  /*
   * Process an extent free intent item that was recovered from
   * the log.  We need to free the extents that it describes.
+ * The caller will free all EFI entries on error.
   */
  STATIC int
  xlog_recover_process_efi(
-       xfs_mount_t             *mp,
-       xfs_efi_log_item_t      *efip)
+       struct xfs_mount        *mp,
+       struct xfs_efi_log_item *efip)
  {
-       xfs_efd_log_item_t      *efdp;
-       xfs_trans_t             *tp;
+       struct xfs_efd_log_item *efdp;
+       struct xfs_trans        *tp;
        int                     i;
        int                     error = 0;
        xfs_extent_t            *extp;
        xfs_fsblock_t           startblock_fsb;

        ASSERT(!test_bit(XFS_EFI_RECOVERED,&efip->efi_flags));
+       /* All paths need the XFS_EFI_RECOVERED flag set. Do it here. */
+       set_bit(XFS_EFI_RECOVERED,&efip->efi_flags);

        /*
         * First check the validity of the extents described by the
@@ -3662,12 +3665,6 @@ xlog_recover_process_efi(
                    (extp->ext_len == 0) ||
                    (startblock_fsb>= mp->m_sb.sb_dblocks) ||
                    (extp->ext_len>= mp->m_sb.sb_agblocks)) {
-                       /*
-                        * This will pull the EFI from the AIL and
-                        * free the memory associated with it.
-                        */
-                       set_bit(XFS_EFI_RECOVERED,&efip->efi_flags);
-                       xfs_efi_release(efip, efip->efi_format.efi_nextents);
                        return XFS_ERROR(EIO);
                }
        }
@@ -3687,7 +3684,6 @@ xlog_recover_process_efi(
                                         extp->ext_len);
        }

-       set_bit(XFS_EFI_RECOVERED,&efip->efi_flags);
        error = xfs_trans_commit(tp, 0);
        return error;

This is basically saying "xfs_efi_release() should drop both the AIL
and the EFD reference as we really only only have one reference". I
think we should simply remove the XFS_EFI_RECOVERED bit from the
reference counting code, and simply subtract one of the EFI
references in xlog_recover_efi_pass2() where we insert the EFI into
the AIL, so that it behaves exactly like the runtime case when
later processing the EFDs.

That is, in xlog_recover_efd_pass2() when we find a matching EFD we
simply call xfs_efi_release() and that does all the freeing because
we're removing the EFD reference which is the only remaining
reference.

Then in xlog_recover_process_efi(), we simply call xfs_efi_release()
for the error cases to drop the last reference to the EFI, otherwise
the commit of the EFD will free it because it calls
xfs_efi_release() appropriately.

Works for me.



@@ -3718,8 +3714,8 @@ STATIC int
  xlog_recover_process_efis(
        struct xlog     *log)
  {
-       xfs_log_item_t          *lip;
-       xfs_efi_log_item_t      *efip;
+       struct xfs_log_item     *lip;
+       struct xfs_efi_log_item *efip;
        int                     error = 0;
        struct xfs_ail_cursor   cur;
        struct xfs_ail          *ailp;
@@ -3753,12 +3749,13 @@ xlog_recover_process_efis(
                error = xlog_recover_process_efi(log->l_mp, efip);
                spin_lock(&ailp->xa_lock);
                if (error)
-                       goto out;
+                       break;
                lip = xfs_trans_ail_cursor_next(ailp,&cur);
        }
-out:
        xfs_trans_ail_cursor_done(ailp,&cur);
        spin_unlock(&ailp->xa_lock);
+       if (error)
+               xfs_efi_clear_ail(ailp);
        return error;
  }

If we fix the reference counting, then I think all we need here is:

        int     error = 0;

        while (ail cursor next) {
                ....
                if (!error)
                        error = xlog_recover_process_efi()
                else
                        xfs_efi_release(efip)
                ....
        }

So that any error will trigger freeing of all the other unprocessed
EFIs on the AIL that are pending recovery without needing any more
special looping....


Nod.

--Mark.

<Prev in Thread] Current Thread [Next in Thread>