Re: [PATCH] iomap: Handle I/O errors gracefully in page_mkwrite

[Matthew Wilcox]

On Fri, Jun 05, 2020 at 08:57:26AM +1000, Dave Chinner wrote:
> On Thu, Jun 04, 2020 at 01:23:40PM -0700, Matthew Wilcox wrote:
> > From: "Matthew Wilcox (Oracle)" <willy@xxxxxxxxxxxxx>
> > 
> > Test generic/019 often results in:
> > 
> > WARNING: at fs/iomap/buffered-io.c:1069 iomap_page_mkwrite_actor+0x57/0x70
> > 
> > Since this can happen due to a storage error, we should not WARN for it.
> > Just return -EIO, which will be converted to a SIGBUS for the hapless
> > task attempting to write to the page that we can't read.
> 
> Why didn't the "read" part of the fault which had the EIO error fail
> the page fault? i.e. why are we waiting until deep inside the write
> fault path to error out on a failed page read?

I have a hypothesis that I don't know how to verify.

First the task does a load from the page and we put a read-only PTE in
the page tables.  Then it writes to the page using write().  The page
gets written back, but hits an error in iomap_writepage_map()
which calls ClearPageUptodate().  Then the task with it mapped attempts
to store to it.

I haven't dug through what generic/019 does, so I don't know how plausible
this is.