Re: Cache coherency bug: stale reads on /dev/sda1

[Joshua Hudson]

"Whole disk and all partitions have page caches of their own."

That's so bad.

I can think of numerous cases where this will cause problems; including
some I encountered last year and did not understand at the time. Manipulating
EFI partitions through the whole disk device makes sense because FAT filesystems
*know their offset on the disk*, and some of the existing tools really
don't like being
given a partition device.

There's also the astounding: write stuff to disk, umount everything,
copy one disk to
another using the whole disk device doesn't work because reading the whole disk
yields a stale cache (sometimes).

On the other hand, I can think of very few cases where the file vs
disk buffer pool
matters, because the loop device is unaffected (writing to a loop
block device is
coherent with the file).

On Sat, Aug 26, 2023 at 7:06 PM Al Viro <viro@xxxxxxxxxxxxxxxxxx> wrote:
>
> On Sat, Aug 26, 2023 at 06:25:58PM -0700, Joshua Hudson wrote:
> > I am presuming you mean "Are you sure this not a hardware issue?"
> >
> > I am sure it is not, for two reasons.
> >
> > 1) If it were a hardware issue I would still expect the two device
> > nodes (whole disk and partition) to report the *same* data.
> >
> > 2) I have since developed a workaround involving BLKFLSDEV. The
> > workaround is really ugly.
>
> Whole disk and all partitions have page caches of their own.
> There is no coherency between those; write through one, then
> read through another and you are not guaranteed that read won't
> see the stale cached data.
>
> Doctor, it hurts when I do it...
>
> Incidentally, read from device/write to the place you've read via
> file/read from device again also has no coherency warranties, for
> exact same reason.
>
> IOW, not a bug.