Re: Silicon Image 3112A SATA trouble

[Jeff Garzik]

Jens Axboe wrote:
> On Sun, Nov 30 2003, Jeff Garzik wrote:
>
> > Jens Axboe wrote:
> >
> > > On Sun, Nov 30 2003, Jeff Garzik wrote:
> > >
> > > > fond of partial completions, as I feel they add complexity, particularly 
> > > > so in my case:  I can simply use the same error paths for both the 
> > > > single-sector taskfile and the "everything else" taskfile, regardless of 
> > > > which taskfile throws the error.
> > >
> > >
> > > It's just a questions of maintaining the proper request state so you
> > > know how much and what part of a request is pending. Requests have been
> > > handled this way ever since clustered requests, that is why
> > > current_nr_sectors differs from nr_sectors. And with hard_* duplicates,
> > > it's pretty easy to extend this a bit. I don't see this as something
> > > complex, and if the alternative you are suggesting (your implementation
> > > idea is not clear to me...) is to fork another request then I think it's
> > > a lot better.
> >
> > [snip howto]
> >
> > Yeah, I know how to do partial completions.  The increased complexity 
> > arises in my driver.  It's simply less code in my driver to treat each 
> > transaction as an "all or none" affair.
> >
> > For the vastly common case, it's less i-cache and less interrupts to do 
> > all-or-none.  In the future I'll probably want to put partial 
> > completions in the error path...
>
>
> Oh come one, i-cache? We're doing IO here, a cache line more or less in
> request handling is absolutely so much in the noise. 
>
> What are the "increased complexity" involved with doing partial
> completions? You don't even have to know it's a partial request in the
> error handling, it's "just the request" state. Honestly, I don't see a
> problem there. You'll have to expand on what exactly you see as added
> complexity. To me it still seems like the fastest and most elegant way
> to handle it. It requires no special attention on request buildup, it
> requires no extra request and ugly split-code in the request handling.
> And the partial-completions come for free with the block layer code.

libata, drivers/ide, and SCSI all must provide internal "submit this 
taskfile/cdb" API that is decoupled from struct request.  Therefore, 
submitting a transaction pair, or for ATAPI submitting the internal 
REQUEST SENSE, is quite simple and only a few lines of code.

Any extra diddling of the hardware, and struct request, to provide 
partial completions is extra code.  The hardware is currently set up to 
provide only "it's done" or "it failed" information.  Logically, then, 
partial completions must be more code than the current <none> ;-)


WRT error handling, according to ATA specs I can look at the error 
information to determine how much of the request, if any, completed 
successfully.  (dunno if this is also doable on ATAPI)  That's why 
partial completions in the error path make sense to me.


> > > > (thinking out loud)  Though best for simplicity, I am curious if a 
> > > > succession of "tiny/huge" transaction pairs are efficient?  I am hoping 
> > > > that the drive's cache, coupled with the fact that each pair of 
> > > > taskfiles is sequentially contiguous, will not hurt speed too much over 
> > > > a non-errata configuration...
> > >
> > >
> > > My gut would say rather two 64kb than a 124 and 4kb. But you should do
> > > the numbers, of course :). I'd be surprised if the former wouldn't be
> > > more efficient.
> >
> > That's why I was thinking out loud, and also why I CC'd Eric :)  We'll 
>
>
> Numbers are better than Eric :)

Agreed.


> > see.  I'll implement whichever is easier first, which will certainly be 
> > better than the current sledgehammer limit.  Any improvement over the 
>
>
> Definitely, the current static limit completely sucks...
>
>
> > current code will provide dramatic performance increases, and we can 
> > tune after that...
>
>
> A path needs to be chosen first, though.

The path has been chosen:  the "it works" solution first, then tune. :)

	Jeff



-
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/