Re: generic/224 failures on 4.5 - encrypted test case

[Theodore Ts'o <tytso@mit.edu>]

On Tue, Mar 15, 2016 at 05:37:27PM -0400, Eric Whitney wrote:
> The test fails relatively rarely - about one in ten trials.  The set of
> reported kernel errors varies considerably from run to run.  The one constant
> is an ENOSPC complaint from ext4_bio_write_page() which appears whether the
> test passes or fails.
>
> [   18.146614] ext4_bio_write_page: ret = -12
>

-12 is not ENOSPC; it's ENOMEM.   (ENOSPC is 28).

So basically we're hitting a case where generic/224 is submitting data
so quickly that we can't handle it fast enough.  I suspect the call
path is that we're inside the jbd2_commit(), and we're calling
journal_submit_data_buffers(), which ends up calling
ext4_writepages(), and this is returning the ENOMEM.  Unfortunately
we're not providing a better message there, and we just return
jbd2_journal_abort() if journal_submit_data_buffers() fails.

I suspect what we need to do is to pass down a flag through
ext4_bio_write_page() and ext4_encrypt() so that in the case where we
are doing a data integrity sync, that we have to use GFP_NOFAIL in our
data allocations in the fs/ext4/crypto.c.

							- Ted