• Bob Peterson's avatar
    gfs2: Rework how rgrp buffer_heads are managed · b3422cac
    Bob Peterson authored
    Before this patch, the rgrp code had a serious problem related to
    how it managed buffer_heads for resource groups. The problem caused
    file system corruption, especially in cases of journal replay.
    
    When an rgrp glock was demoted to transfer ownership to a
    different cluster node, do_xmote() first calls rgrp_go_sync and then
    rgrp_go_inval, as expected. When it calls rgrp_go_sync, that called
    gfs2_rgrp_brelse() that dropped the buffer_head reference count.
    In most cases, the reference count went to zero, which is right.
    However, there were other places where the buffers are handled
    differently.
    
    After rgrp_go_sync, do_xmote called rgrp_go_inval which called
    gfs2_rgrp_brelse a second time, then rgrp_go_inval's call to
    truncate_inode_pages_range would get rid of the pages in memory,
    but only if the reference count drops to 0.
    
    Unfortunately, gfs2_rgrp_brelse was setting bi->bi_bh = NULL.
    So when rgrp_go_sync called gfs2_rgrp_brelse, it lost the pointer
    to the buffer_heads in cases where the reference count was still 1.
    Therefore, when rgrp_go_inval called gfs2_rgrp_brelse a second time,
    it failed the check for "if (bi->bi_bh)" and thus failed to call
    brelse a second time. Because of that, the reference count on those
    buffers sometimes failed to drop from 1 to 0. And that caused
    function truncate_inode_pages_range to keep the pages in page cache
    rather than freeing them.
    
    The next time the rgrp glock was acquired, the metadata read of
    the rgrp buffers re-used the pages in memory, which were now
    wrong because they were likely modified by the other node who
    acquired the glock in EX (which is why we demoted the glock).
    This re-use of the page cache caused corruption because changes
    made by the other nodes were never seen, so the bitmaps were
    inaccurate.
    
    For some reason, the problem became most apparent when journal
    replay forced the replay of rgrps in memory, which caused newer
    rgrp data to be overwritten by the older in-core pages.
    
    A big part of the problem was that the rgrp buffer were released
    in multiple places: The go_unlock function would release them when
    the glock was released rather than when the glock is demoted,
    which is clearly wrong because our intent was to cache them until
    the glock is demoted from SH or EX.
    
    This patch attempts to clean up the mess and make one consistent
    and centralized mechanism for managing the rgrp buffer_heads by
    implementing several changes:
    
    1. It eliminates the call to gfs2_rgrp_brelse() from rgrp_go_sync.
       We don't want to release the buffers or zero the pointers when
       syncing for the reasons stated above. It only makes sense to
       release them when the glock is actually invalidated (go_inval).
       And when we do, then we set the bh pointers to NULL.
    2. The go_unlock function (which was only used for rgrps) is
       eliminated, as we've talked about doing many times before.
       The go_unlock function was called too early in the glock dq
       process, and should not happen until the glock is invalidated.
    3. It also eliminates the call to rgrp_brelse in gfs2_clear_rgrpd.
       That will now happen automatically when the rgrp glocks are
       demoted, and shouldn't happen any sooner or later than that.
       Instead, function gfs2_clear_rgrpd has been modified to demote
       the rgrp glocks, and therefore, free those pages, before the
       remaining glocks are culled by gfs2_gl_hash_clear. This
       prevents the gl_object from hanging around when the glocks are
       culled.
    Signed-off-by: default avatarBob Peterson <rpeterso@redhat.com>
    Reviewed-by: default avatarAndreas Gruenbacher <agruenba@redhat.com>
    b3422cac
glops.c 16.4 KB