[lkml]   [2006]   [Nov]   [2]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectRe: New filesystem for Linux
> Mikulas Patocka <> writes:
>> new method to keep data consistent in case of crashes (instead
>> of journaling),
> What is that method?

Some tricks to avoid journal --- see

--- 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).


> -Andi
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to
More majordomo info at
Please read the FAQ at

 \ /
  Last update: 2006-11-03 02:47    [from the cache]
©2003-2014 Jasper Spaans. hosted at Digital Ocean