Hi all, Here's a stab at a fix for the issue demonstrated by the previously posted xfstest [1]. This basically just filters out error returns to write_cache_pages() and returns the first encountered error from the higher level ->writepage[s]() call. AFAICT the end result should be the same as today in the event of a persistent writeback error. I'm not totally sure if this is the best solution, but it's the simplest that comes to mind. FWIW, this patch changes error return behavior unconditionally. A slightly modified approach might be to only filter errors as such on integrity writeback so background writeback will presumably behave as it always has, while allowing fsync/unmount to process the address space in one writeback job. If there's a reason we want to preserve the current writeback behavior in general, I suppose we could consider more involved options like tag any inode that preempts writeback due to error and then make sure we process/scan/clear such inodes of delalloc blocks before we unmount (and/or on fsync). Either way, we may also want to consider rate limiting the "page discard" errors if it's more likely we dump a bunch at once. Thoughts? Brian [1] https://marc.info/?l=linux-xfs&m=154031860122440&w=2 Brian Foster (2): xfs: add writepage map error tag xfs: don't preempt writeback sequence on single page wb error fs/xfs/libxfs/xfs_errortag.h | 4 +++- fs/xfs/xfs_aops.c | 19 ++++++++++++++----- fs/xfs/xfs_error.c | 3 +++ 3 files changed, 20 insertions(+), 6 deletions(-) -- 2.17.2
