Hello, First I'd like to briefly remind the bug I was fixing recently (or you can read the original report in: http://marc.theaimsgroup.com/?l=linux_fsdevel&m=114513520128759&w=2). The original problem is that when we are processing BJ_Forget list in journal_commit_transaction() we can process bitmap buffer before some buffer that was deleted by the committing transaction. As soon as bitmap buffer is processed we can reallocate all the blocks freed by the comitting transaction. Hence if there are still some unprocessed deleted buffers on the BJ_Forget list, we will BUG as we think they should not have been reallocated. The fix (currently in -mm) was to not BUG in such a case but file the buffer to BJ_Metadata list of the running transaction. But that is not actually correct at all times. If the buffer is not jbddirty, it should rather be filed into BJ_Reserved list (non-dirty buffer in BJ_Metadata list is a bug). In theory there could be even cases when the buffer should be filed in BJ_Forget list of the running transaction (if the buffer was reallocated and freed before we managed to process the BJ_Forget list of the comitting transaction). To make it short I think this gets rather hairy and some better solution is really needed. Maybe we could allow do_get_write_access and similar functions to remove buffer from BJ_Forget list of the comitting transaction. That would complicate a bit the locking of BJ_Forget list but it should be doable. Does anyone else have some idea? Honza - To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
