Hi xfs developers.
These days when I studied the xfs code, I found that
if a process allocats one inode extents, while another process
frees its extents which belong to another inode at the meantime,
these two processes maybe deadlock. Following is my analysis:
when allocating the extents, the process may loop over allocation groups
and lock their allocation agf blocks. (the code 2380-2430 in xfs_alloc.c)
At the same time, if another process free its extents, then the deadlock of
these two process will happen, for it will also lock some allocation agf
block,
and furthermore the lock sequences of agf in these two process is apparently
different.(the code 4040-4056 in fs/xfs/xfs_bmap.c)
I really want to figure out it.
Your responses is highly appreciated.
Best regards.
|