On Mon, May 11, 2020 at 11:44:11PM +0100, Peter Grandi wrote:
> >>> With lock / unlock, I get around 1.2MB/sec per device
> >>> component, with ~13% CPU load. Wihtout lock / unlock, I get
> >>> around 15.5MB/sec per device component, with ~30% CPU load.
>
> >> [...] we still need to avoid race conditions. [...]
>
> Not all race conditions are equally bad in this situation.
>
> > 1. Per your previous reply, only call raid6check when array is
> > RO, then we don't need the lock.
> > 2. Investigate if it is possible that acquire stripe_lock in
> > suspend_lo/hi_store [...]
>
> Some other ways could be considered:
>
> * Read a stripe without locking and check it; if it checks good,
> no problem, else either it was modified during the read, or it
> was faulty, so acquire a W lock, reread and recheck it (it
> could have become good in the meantime).
>
> The assumption here is that there is a modest write load from
> applications on the RAID set, so the check will almost always
> succeed, and it is worth rereading the stripe in very rare
> cases of "collisions" or faults.
>
> * Variants, like acquiring a W lock (if possible) on the stripe
> solely while reading it ("atomic" read, which may be possible
> in other ways without locking) and then if check fails we know
> it was faulty, so optionally acquire a new W lock and reread
> and recheck it (it could have become good in the meantime).
>
> The assumption here is that the write load is less modest, but
> there are a lot more reads than writes, so a W lock only
> during read will eliminate the rereads and rechecks from
> relatively rare "collisions".
The locking method was suggested by Neil,
I'm not aware of other methods.
About the check -> maybe lock -> re-check,
it is a possible workaround, but I find it
a bit extreme.
In any case, we should keep it in mind.
bye,
pg
> The case where there is at the same time a large application
> write load on the RAID set and checking at the same time is hard
> to improve and probably eliminating rereads and rechecks by just
> acquiring the stripe W lock for the whole duration of read and
> check.
--
piergiorgio
