Re: [PATCH v4 00/48] block: do not drain while holding the graph lock

[Kevin Wolf <kwolf@redhat.com>]

Am 30.05.2025 um 17:10 hat Fiona Ebner geschrieben:
> Previous discussions:
> v3: [0]
> v2: [1]
> v1: [2]
> 
> Changes in v4:
> * Document requirement to drain all block nodes for affected
>   functions.
> * Also cover the generated bdrv_co_unref_child().
> * Remove now superfluous drain_bs variable in bdrv_set_backing_hd().
> * Mark bdrv_graph_wrlock_drained() wrapper as GRAPH_UNLOCKED.
> * Unify bdrv_set_backing_hd() with its drained_variant.
> * Mark more functions up the call-stack as GRAPH_UNLOCKED. This is
>   almost all of the new patches in the latter half of the series, most
>   do not require substantial changes, but there are a few where
>   something needed to be done. I did not mark functions outside the
>   block layer like qemu_cleanup(), save_snapshot(), qmp_xyz(), etc.
>   and also not functions that explicitly do a rdunlock_main_loop()
>   before calling a function that is GRAPH_UNLOCKED.
> 
> There were no changes for patches 01/48-09/48 and 17/48-23/48, endpoints
> inclusive. All patches starting from 25/48 are new in v4.

This is starting to become a little unmanageable. :-)

I'm sure we could keep adding more and more cleanup patches with each
version of the series, but we don't really have to let the first fixes
wait for all cleanups and fix the whole world in a single series.

So I decided to start with a prefix of this series and applied
patches 1-22, which all had received review before and were easy to
compare against the previous version. If you don't object, I'd send a
pull request for those without waiting for the rest.

If the other patches still need some changes, you can start a new patch
series at v1 containing the remaining patches. (No need to resend them
now, though, I can review them in this series.)

Does this make sense to you?

Kevin