Re: [RFC PATCH v2 0/9] Block/XFS: Support alternative mirror device retry

[Dave Chinner <david@fromorbit.com>]

On Thu, Feb 28, 2019 at 04:28:53PM -0700, Andreas Dilger wrote:
> On Feb 28, 2019, at 7:22 AM, Bob Liu <bob.liu@oracle.com> wrote:
> > 
> > On 2/19/19 5:31 AM, Dave Chinner wrote:
> >> On Wed, Feb 13, 2019 at 05:50:35PM +0800, Bob Liu wrote:
> >>> Motivation:
> >>> When fs data/metadata checksum mismatch, lower block devices may have other
> >>> correct copies. e.g. If XFS successfully reads a metadata buffer off a raid1 but
> >>> decides that the metadata is garbage, today it will shut down the entire
> >>> filesystem without trying any of the other mirrors.  This is a severe
> >>> loss of service, and we propose these patches to have XFS try harder to
> >>> avoid failure.
> >>> 
> >>> This patch prototype this mirror retry idea by:
> >>> * Adding @nr_mirrors to struct request_queue which is similar as
> >>>  blk_queue_nonrot(), filesystem can grab device request queue and check max
> >>>  mirrors this block device has.
> >>>  Helper functions were also added to get/set the nr_mirrors.
> >>> 
> >>> * Introducing bi_rd_hint just like bi_write_hint, but bi_rd_hint is a long bitmap
> >>> in order to support stacked layer case.
> >>> 
> >>> * Modify md/raid1 to support this retry feature.
> >>> 
> >>> * Adapter xfs to use this feature.
> >>>  If the read verify fails, we loop over the available mirrors and retry the read.
> >> 
> >> Why does the filesystem have to iterate every single posible
> >> combination of devices that are underneath it?
> 
> Even if the filesystem isn't doing this iteration, there needs to be
> some way to track which devices or combinations of devices have been
> tried for the bio, which likely still means something inside the bio.

I don't beleive it needs to be "in the bio". The thing that does
the iteration (i.e. the layer with multiple copies or rebuild
capability) is the one that captures the IO completion state, runs
the verifier it is supplied with and re-issues the read if the
verifier or initial IO fails.

i.e. it moves the iteration down to the thing that knows what can be
iterated, and so there's no state needed in the bio itself.

> >> Wouldn't it be much simpler to be able to attach a verifier
> >> function to the bio, and have each layer that gets called iterate
> >> over all it's copies internally until the verfier function passes
> >> or all copies are exhausted?
> >> 
> >> This works for stacked mirrors - it can pass the higher layer
> >> verifier down as far as necessary. It can work for RAID5/6, too, by
> >> having that layer supply it's own verifier for reads that verifies
> >> parity and can reconstruct of failure, then when it's reconstructed
> >> a valid stripe it can run the verifier that was supplied to it from
> >> above, etc.
> >> 
> >> i.e. I dont see why only filesystems should drive retries or have to
> >> be aware of the underlying storage stacking. ISTM that each
> >> layer of the storage stack should be able to verify what has been
> >> returned to it is valid independently of the higher layer
> >> requirements. The only difference from a caller point of view should
> >> be submit_bio(bio); vs submit_bio_verify(bio, verifier_cb_func);
> 
> I don't think the filesystem should be aware of the stacking (nor are
> they in the proposed implementation).  That said, the filesystem-level
> checksums should, IMHO, be checked at the filesystem level, and this
> proposal allows the filesystem to tell the lower layer "this read was
> bad, try something else".

After the fact, yes. I want the verification during the IO while
the layer that knows about iteration and recovery can do this
easily.

i.e. all the complexity right now is because we back out of the
layer that can do iteration before we can run the verification, and
so we have to carry some state up to a higher level and then pass it
back down in a completely separate IO context. That's where all this
"need to carry satate in the bio" stuff comes from, and that's what
I'm trying to get rid of.

> One option, instead of having a bitmap, with one bit for every possible
> device/combination in the system, would be to have a counter instead.
> This is much denser, and even the existing "__u16 bio_write_hint" field
> would be enough for 2^16 different devices/combinations of devices to
> be tried.  The main difference would be that the retry layers in the
> device layer would need to have a deterministic iterator for the bio.

The problem there is stacked layers - each layer needs a unique
ID for it's iterator function, as this complexity:

> For stacked devices it would need to use the same API to determine how
> many possible combinations are below it, and do a modulus to pass down
> the per-device iteration number.  The easiest would be to iterate in
> numeric order, but it would also be possible to use something like a
> PRNG seeded by e.g. the block number to change the order on a per-bio
> basis to even out the load, if that is desirable.
> 
> > For a two layer stacked md case like:
> >                              /dev/md0
> >             /                  |                  \
> >      /dev/md1-a             /dev/md1-b          /dev/sdf
> >   /        \           /       |        \
> > /dev/sda /dev/sdb  /dev/sdc /dev/sdd  /dev/sde
> 
> In this case, the top-level md0 would call blk_queue_get_copies() on each
> sub-devices to determine how many sub-devices/combinations they have,
> and pick the maximum (3 in this case), multiplied by the number of
> top-level devices (also 3 in this case).  That means the top-level device
> would return blk_queue_get_copies() == 9 combinations, but the same
> could be done recursively for more/non-uniform layers if needed.
>
> The top-level device maps md1-a = [0-2], md1-b = [3-5], md1-c = [6-8],
> and can easily map an incoming bio_read_hint to the next device, either
> by simple increment or by predetermining a device ordering and following
> that (e.g. 0, 3, 6, 1, 4, 7, 2, 5, 8), or any other deterministic order
> that hits all of the devices exactly once).  During submission bio_read_hint
> is set to the modulus of the value (so that each layer in the stack sees
> only values in the range [0, copies), and when the bio completes the top-level
> device will set bio_read_hint to be the next sub-device to try (like the
> original proposal was splitting and combining the bitmaps).  If a sub-device
> gets a bad index (e.g. md1-a sees bio_read_hint == 2, or sdf sees anything
> other than 0) it is a no-op and returns e.g. -EAGAIN to the upper device
> so that it moves to the next device without returning to the caller.

.... clearly demonstrates.

I'd much prefer stacking completions and running them on demand as
it should work for all stacked types and not require magic to
iterate constructs like the above.

Passing the verifier down also allows the underlying layer to repair
itself. i.e. if it gets a verifier failure, then retries and gets
success, it knows immediately which part of the mirror contains bad
data and can repair it. It can also trigger a region scrub, knowing
which device might be bad and which is likely to contain good data.
i.e. we can start to think about automated block device self-repair
if we can supply a data verifier with submit_bio()...

> >> I suspect there's a more important issue to worry about: we run the
> >> XFS read verifiers in an async work queue context after collecting
> >> the IO completion status from the bio, rather than running directly
> >> in bio->bi_end_io() call chain.
> 
> In this proposal, XFS would just have to save the __u16 bio_read_hint
> field from the previous bio completion and set it in the retried bio,
> so that it could start at the next device/combination.  Obviously,
> this would mean that the internal device iterator couldn't have any
> hidden state for the bio so that just setting bio_read_hint would be
> the same as resubmitting the original bio again, but that is already
> a given or this whole problem wouldn't exist in the first place.

It still requires code in the filesystem to iterate and retry N
times, instead of never. And we still have to re-write the data we
read to fix the underlying device issue (which the device should
already know about and have fixed by this point!) i.e. we either get
verified data returned on bio completion or we get an error to say
the data was corrupt and unrecoverable. If someone wants "fail fast"
semantics, then they simply don't provide a verifier....

Cheers,

Dave.
-- 
Dave Chinner
david@fromorbit.com