Hi,
On 9/14/2012 6:41 PM, Hugh Dickins wrote:
On Tue, 11 Sep 2012, Laura Abbott wrote:
When a buffer is added to the LRU list, a reference is taken which is not dropped until the buffer is evicted from the LRU list. This is the correct behavior, however this LRU reference will prevent the buffer from being dropped. This means that the buffer can't actually be dropped until it is selected for eviction. There's no bound on the time spent on the LRU list, which means that the buffer may be undroppable for very long periods of time. Given that migration involves dropping buffers, the associated page is now unmigratible for long periods of time as well.
Disclaimer: I'm no expert on buffer_heads, and haven't studied your patch. But it seems to me that this is an issue with the (unnamed) filesystem you use, rather than a problem to be solved in drop_buffers().
We are using ext4
extN, gfs2, ntfs, ocfs2 and xfs set .migratepage = buffer_migrate_page, and I cannot see that page migration involves drop_buffers() at all in that case: it transfers the buffer_heads from the old page to the new, whether they're busy or not, with no attempt to free them.
That's true for most of the address spaces EXCEPT for the journaled address space operations; ext4_ordered_aops, ext4_writeback_aops, ext4_da_aops all set migratepage but ext4_journalled_aops does not set migratepage at all. This seems to be true all the way back to when the migratepage was added for ext3.
Maybe your filesystem can be converted, with or without some extra help, to buffer_migrate_page() instead of the default fallback_migrate_page(): which indeed has to play safe, doing the try_to_release_page() you see. Maybe ask on the mailing list for your filesystem?
I could ask on the ext mailing list for the historical reasons why the journalled ops don't have migrate pages, but I'm still going to assert this is still a problem with fallback_migrate_page. It's still possible to have drop_buffers fail unnecessarily because the buffer is stuck on the LRU list and I don't see why the problem shouldn't be fixed there as well.
Hugh
Thanks, Laura