On Wed, Nov 04 2020 at 1:47am -0500,
JeffleXu <jefflexu@xxxxxxxxxxxxxxxxx> wrote:
>
> On 11/2/20 11:28 PM, Mike Snitzer wrote:
> >On Sun, Nov 01 2020 at 10:14pm -0500,
> >JeffleXu <jefflexu@xxxxxxxxxxxxxxxxx> wrote:
> >
> >>On 10/27/20 2:53 AM, Mike Snitzer wrote:
> >>>What you detailed there isn't properly modeling what it needs to.
> >>>A given dm_target_io could result in quite a few bios (e.g. for
> >>>dm-striped we clone each bio for each of N stripes). So the fan-out,
> >>>especially if then stacked on N layers of stacked devices, to all the
> >>>various hctx at the lowest layers is like herding cats.
> >>>
> >>>But the recursion in block core's submit_bio path makes that challenging
> >>>to say the least. So much so that any solution related to enabling
> >>>proper bio-based IO polling is going to need a pretty significant
> >>>investment in fixing block core (storing __submit_bio()'s cookie during
> >>>recursion, possibly storing to driver provided memory location,
> >>>e.g. DM initialized bio->submit_cookie pointer to a blk_qc_t within a DM
> >>>clone bio's per-bio-data).
> >>>
> >>>SO __submit_bio_noacct would become:
> >>>
> >>> retp = &ret;
> >>> if (bio->submit_cookie)
> >>> retp = bio->submit_cookie;
> >>> *retp = __submit_bio(bio);
> >>Sorry for the late reply. Exactly I missed this point before. IF you
> >>have not started working on this, I'd like to try to implement this as
> >>an RFC.
> >I did start on this line of development but it needs quite a bit more
> >work. Even the pseudo code I provided above isn't useful in the context
> >of DM clone bios that have their own per-bio-data to assist with this
> >implementation. Because the __submit_bio_noacct() recursive call
> >drivers/md/dm.c:__split_and_process_bio() makes is supplying the
> >original bio (modified to only point to remaining work).
>
> Yes I noticed this recently. Since the depth-first splitting
> introduced in commit 18a25da84354
>
> ("dm: ensure bio submission follows a depth-first tree walk"), one
> bio to dm device can be
>
> split into multiple bios to this dm device.
>
> ```
>
> one bio to dm device (dm0) = one dm_io (to nvme0) + one bio to this
> same dm device (dm0)
>
> ```
>
>
> In this case we need a mechanism to track all split sub-bios of the
> very beginning original bio.
Yes, splitting causes additional potential for sub-bios. There are
other cases that cause a 1-to-many bio generation (e.g. dm-striped) or
splitting cases where a DM target makes use of dm_accept_partial_bio
(e.g. dm-snapshot, dm-integrity, dm-writecache, etc).
> I'm doubted if this should be implemented in block layer like:
>
> ```
>
> struct bio {
>
> ...
>
> struct list_head cookies;
>
> };
>
> ```
>
> After all it's only used by bio-based queue, or more specifically
> only dm device currently.
I do think this line of work really should be handled in block core
because I cannot see any reason why MD or bcache or whatever bio-based
device wouldn't want the ability to better support io_uring (with IO
poll).
> Another design I can come up with is to maintain a global data
> structure for the very beginning
> original bio. Currently the blocking point is that now one original
> bio to the dm device (@bio of dm_submit()) can correspond to multiple
> dm_io and thus we have nowhere to place the @cookies list.
Yes, and that will always be the case. We need the design to handle an
arbitrary sprawl of splitting from a given bio. The graph of bios
resulting from that fan-out needs to be walked at various levels -- be
it the top-level original bio's submit_bio() returned cookie or some
intermediate point in the chain of bios.
The problem is the lifetime of the data structure created for a given
split bio versus layering boundaries (that come from block core's
simplistic recursion via bio using submit_bio).
> Now we have to maintain one data structure for every original bio,
> something like
>
> ```
>
> struct dm_poll_instance {
>
> ...
>
> struct list_head cookies;
>
> };
>
> ```
I do think we need a hybrid where at the point of recursion we're able
to make the associated data structure available across the recursion
boundary so that modeling the association in a chain of split bios is
possible. (e.g. struct dm_poll_data or dm_poll_instance as you named it,
_but_ that struct definition would live in block core, but would be part
of per-bio-data; so 'struct blk_poll_data' is more logical name when
elevated to block core).
It _might_ be worthwhile to see if a new BIO_ flag could be added to
allow augmenting the bio_split + bio_chain pattern to also track this
additional case of carrying additional data per-bio while creating
bio-chains. I may not be clear yet, said differently: augmenting
bio_chain to not only chain bios, but to _also_ thread/chain together
per-bio-data that lives within those chained bios. SO you have the
chain of bios _and_ the chain of potentially opaque void * that happens
to point to a list head for a list of 'struct blk_poll_data'.
Does that make sense?
> We can transfer this dm_poll_instance between split bios by
> bio->bi_private, like
>
> ```
>
> dm_submit_bio(...) {
>
> struct dm_poll_instance *ins;
>
> if (bio->bi_private)
>
> ins = bio->bi_private;
>
> else {
>
> ins = alloc_poll_instance();
>
> bio->bi_private = ins;
>
> }
>
> ...
>
> }
>
> ```
Sadly, we cannot (ab)use bi_private for this given its (ab)used via the
bio_chain() interface. It's almost like we need to add a new pointer in
the bio that isn't left for block core to hijack.
There is the well-worn pattern of saving off the original bi_private,
hooking a new endio method and then when that endio is called restoring
bi_private but we really want to avoid excessive indirect function calls
for this usecase. The entire point of implementing blk_poll support is
for performance after all.
Mike
--
dm-devel mailing list
dm-devel@xxxxxxxxxx
https://www.redhat.com/mailman/listinfo/dm-devel
