Re: New filesystem for Linux

[Mikulas Patocka]

> Mikulas Patocka <mikulas@xxxxxxxxxxxxxxxxxxxxxxxx> writes:
>
> > new method to keep data consistent in case of crashes (instead
> > of journaling),
>
> What is that method?

Some tricks to avoid journal --- see 
http://artax.karlin.mff.cuni.cz/~mikulas/spadfs/download/INTERNALS

--- unlike journaling it survives only 65536 crashes :)

> > * There is a rw semaphore that is locked for read for nearly all
>
> Depending on the length of the critical section rw locks are often
> not faster than non rw locks because the read case has to bounce
> around the cache line of the lock anyways and they're actually
> a little more expensive.

This critical section is long --- i.e. any reads/writes to disk. Making it 
simple semaphore would effectively serialize all operations.

> > * This leads to another observation --- on i386 locking a semaphore is
> > 2 instructions, on x86_64 it is a call to two nested functions. Has it
>
> The second call should be a tail call, i.e. just a jump

It is down_write -> (tailcall) down_write_nested -> (normal call) 
spin_lock_irq and spin_unlock_irq.

> The first call isn't needed on a non debug kernel, but doing the
> two unconditional jumps should be basically free on a modern OOO CPU.

But it kills one cacheline.

> The actual implementation is spinlock based vs atomic based for i386.
> This was because at some point nobody could benchmark a difference
> between the two and the spinlock based version is much easier
> to verify and to understand. If you can demonstrate a difference
> that could be reevaluated.

Maybe one day I'll try it.

> > some reason or was it just implementator's laziness? Given the fact
> > that locked instruction takes 16 ticks on Opteron (and can overlap
> > about 2 ticks with other instructions), it would make sense to have
> > optimized semaphores too.
>
> In the last time we're going more for saved icache and better
> use of branch predictors (who are more happy with less branch locations
> to cache) I would expect the call/ret to be executed
> mostly in parallel with the other code anyways.

I see, but pushf, cli and popf in that spinlock hurt too (especially on 
Intel, it has them completely microcoded --- pushf/popf pair is 100 
ticks on Intel P4E and 12 ticks on Opteron).

Mikulas

> -Andi
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/