On Tuesday 17 February 2009, Sergei Shtylyov wrote:
> Hello.
>
> Bartlomiej Zolnierkiewicz wrote:
> > On Monday 16 February 2009, Sergei Shtylyov wrote:
> >
> >> Hello, I wrote:
> >>
> >>
> >>>>> There should be no functional changes caused by this patch.
> >>>>> Signed-off-by: Bartlomiej Zolnierkiewicz <bzolnier@xxxxxxxxx>
> >>>>> Index: b/drivers/ide/ide-iops.c
> >>>>> ===================================================================
> >>>>> --- a/drivers/ide/ide-iops.c
> >>>>> +++ b/drivers/ide/ide-iops.c
> >>>>> @@ -88,11 +88,15 @@ void SELECT_DRIVE (ide_drive_t *drive)
> >>>>> {
> >>>>> ide_hwif_t *hwif = drive->hwif;
> >>>>> const struct ide_port_ops *port_ops = hwif->port_ops;
> >>>>> + ide_task_t task;
> >>>>>
> >>>>> if (port_ops && port_ops->selectproc)
> >>>>> port_ops->selectproc(drive);
> >>>>>
> >>>>> - hwif->OUTB(drive->select.all, hwif->io_ports.device_addr);
> >>>>> + memset(&task, 0, sizeof(task));
> >>>>> + task.tf_flags = IDE_TFLAG_OUT_DEVICE;
> >>>>> +
> >>>>> + drive->hwif->tf_load(drive, &task);
> >>>>>
> >>>> This actually doesn't seem like a bright idea to me, considering
> >>>> that this gets called when starting every request. How will you look
> >>>> at me adding the transport method for writing this register? :-)
> >>>>
> >
> > Please check profiles first -- it might not be worth it. [1]
> >
> >
> >>> Convert SELECT_DRIVE() to use ->tf_load instead of ->OUTB.
> >>>
> >>> OTOH, adding such a "backdoor" to the taskfile doesn't seem very
> >>> consistent... well, I'm not excited about the whole idea conversion to
> >>> tf_{load|read}() -- it's not clear what exactly this bought us.
> >>>
> >
> > This was explained some months ago already, so just to recall -- it was
> > a part of a bigger work removing duplicated code and allowing abstraction
> > of the ATA logic.
> >
> > Anyway this is not set in a stone so if you have proposal of a better
> > approach please come forward with it.
> >
>
> Er... I think that the previous IN()/OUT() methods were better. Note
> that we ended up using the local version of them in the dafault
> ide_tf_{load}read}() anyway -- as Alan has pointed out it might be worth
During ide_tf_{load,read}() addition I was a bit too optimistic about
the possibility of the quick io{read,write}* conversion later...
> splitting those into I/O and memory space versions... although given
> general slowness of the I/O accesses, this is probably not going to win
> much speed-wise.
Maybe it would be worth to add ->tf_{inb,outb} to struct ide_tp_ops
and convert default tp_ops to use them... OTOH we should reinvestigate
the io{read,write}*() way first (maybe things have improved there)...
> >> We at least could have saved on memset() -- tf_load() method ignores
> >> fields other than tf_flags anyway...
> >>
> >
> > Unless it is huge performance win (unlikely) this is not a good idea as it would be a maintainance nightmare.
> >
> > ->tf_load does only use cmd->tf_flags today but it might change one day
> > and nobody will remember to audit all users that they pass a valid cmd...
> >
>
> It's just quite unbearable to see (especially for a long time
> assembly coder) how a single register write is turning into *that*.
> So, it still seems worth risking... :-)
I see your point here. If SELECT_DRIVE() is performance sensitive we
may just add another struct ide_tp_ops method for it...
Thanks,
Bart
--
To unsubscribe from this list: send the line "unsubscribe linux-ide" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
