Re: libata FUA revisited

[Ric Wheeler <ric@emc.com>]

Tejun Heo wrote:
> Jens Axboe wrote:
>> On Wed, Feb 21 2007, Tejun Heo wrote:
>>> [cc'ing Ric, Hannes and Dongjun, Hello.  Feel free to drag other people in.]
>>>
>>> Robert Hancock wrote:
>>>> Jens Axboe wrote:
>>>>> But we can't really change that, since you need the cache flushed before
>>>>> issuing the FUA write. I've been advocating for an ordered bit for
>>>>> years, so that we could just do:
>>>>>
>>>>> 3. w/FUA+ORDERED
>>>>>
>>>>> normal operation -> barrier issued -> write barrier FUA+ORDERED
>>>>>  -> normal operation resumes
>>>>>
>>>>> So we don't have to serialize everything both at the block and device
>>>>> level. I would have made FUA imply this already, but apparently it's not
>>>>> what MS wanted FUA for, so... The current implementations take the FUA
>>>>> bit (or WRITE FUA) as a hint to boost it to head of queue, so you are
>>>>> almost certainly going to jump ahead of already queued writes. Which we
>>>>> of course really do not.
>>> Yeah, I think if we have tagged write command and flush tagged (or
>>> barrier tagged) things can be pretty efficient.  Again, I'm much more
>>> comfortable with separate opcodes for those rather than bits changing
>>> the behavior.
>> ORDERED+FUA NCQ would still be preferable to an NCQ enabled flush
>> command, though.
> 
> I think we're talking about two different things here.
> 
> 1. The barrier write (FUA write) combined with flush.  I think it would
> help improving the performance but I think issuing two commands
> shouldn't be too slower than issuing one combined command unless it
> causes extra physical activity (moving head, etc...).
> 
> 2. FLUSH currently flushes all writes.  If we can mark certain commands
> requiring ordering, we can selectively flush or order necessary writes.
>  (No need to flush 16M buffer all over the disk when only journal needs
> barriering)

We can certainly (given time to play in the lab!) try to measure this in 
with a micro-benchmark (with an analyzer or with block trace?).

A normal flush command in my old tests seemed to be in the 20 ms range 
(mixed in with and occasional "freebie" cache flush which returns in 50 
usecs or so - cache must be empty).


>>> Another idea Dongjun talked about while drinking in LSF was ranged
>>> flush.  Not as flexible/efficient as the previous option but much less
>>> intrusive and should help quite a bit, I think.
>> But that requires extensive tracking, I'm not so sure the implementation
>> of that for barriers would be very clean. It'd probably be good for
>> fsync, though.
> 
> I was mostly thinking about journal area.  Using it for other purposes
> would incur a lot of complexity.  :-(
>