Re: [Dle-develop] [RFC][PATCH v2 -next 2/2] Adding lock operations to kmsg_dump()/pstore_dump()

[Luck, T. <ton...@in...>]

> It ain't pretty but it moves things towards a more reliable message dump.
> The odds of us needing to bust the spinlocks are really small.  Most of
> the time no one reads the pstore filesystem.

Does it really change the odds much?  As you say, the common case is that
pstore front-end doesn't have the lock held - so that case is unchanged.
We can get the lock anyway, we don't need to bust it.

Looking at the uncommon case where the lock is held - that means that
pstore was in the middle of some back-end operation. Busting the lock
means that the back-end will be surprised by being called again when the
first operation had not yet completed. In the case of a state machine
driven back end like ERST, I don't think this has a high probability of
working out well.

So you might be moving the needle from 99.999% chance of saving to pstore
with 0.001% chance of hanging on the spin lock. to 99.9991% chance of
saving, and 0.0009% chance of something highly weird happening in the
back-end driver because you busted the lock and called it anyway.

> I would love to figure out a prettier solution for this locking mess, but
> I can't think of anything.  We have customers who want to utilize this
> technology, so I am trying to make sure it is stable and robust for now.
> A little selfish I suppose.  But we are open to ideas?

If a prettier solution is needed - it will have to involve the back-end.
Perhaps a whole separate write/panic path (with separate buffer). Then
a sufficiently smart back end could do the right thing. I have little
confidence that ERST could be made smart in this way, because almost all
of the heavy lifting is done by the BIOS - so Linux has no way to influence
the flow of execution.

-Tony