It would be nice to be able to report the actual errors from block devices instead of the default -EIO. In order to do that, we need to execute the BIO synchronously as the only way to get the error back to the caller is by returning it from readpage() -- we can't store it in the struct page. But we need to be able to respond to a fatal signal, as we do today with lock_page_killable(). This turns out to be quite hard. The solution I settled on is that the caller must pass in an alternate end_io to be called asynchronously if a fatal signal arrives. I believe the synchronize_rcu() call to be sufficient to ensure that the old bi_end_io() will not be called. If there are callers of bi_end_io() from BH-enabled regions, it may not be! Perhaps we could put a warning in bio_endio() to make sure that's true? Matthew Wilcox (Oracle) (2): block: Add submit_bio_killable fs: Make mpage_readpage synchronous block/bio.c | 87 +++++++++++++++++++++++++++++--------- fs/mpage.c | 25 +++++++++-- include/linux/bio.h | 1 + include/linux/completion.h | 1 + kernel/sched/completion.c | 9 ++-- 5 files changed, 97 insertions(+), 26 deletions(-) -- 2.28.0
