Re: [PATCH 6/6] esp_scsi: Optimize PIO loops

[Michael Schmitz]

Hi Finn,

Am 13.10.2018 um 17:09 schrieb Finn Thain:
> On Sat, 13 Oct 2018, Michael Schmitz wrote:
>
> > Hi Finn,
> >
> > Am 13.10.2018 um 13:51 schrieb Finn Thain:
> > > Avoid function calls in the inner PIO loops. On a Centris 660av this
> > > improves throughput for sequential read transfers by about 40% and
> > > sequential write by about 10%.
> > >
> > > Unfortunately it is not possible to have method calls like esp_write8()
> > > placed inline so this is always going to be slow (even with LTO).
> > >
> > > Tested-by: Stan Johnson <userm57@xxxxxxxxx>
> > > Signed-off-by: Finn Thain <fthain@xxxxxxxxxxxxxxxxxxx>
> > > ---
> > >  drivers/scsi/esp_scsi.c | 14 +++++++-------
> > >  1 file changed, 7 insertions(+), 7 deletions(-)
> > >
> > > diff --git a/drivers/scsi/esp_scsi.c b/drivers/scsi/esp_scsi.c
> > > index 646701fc22a4..9f0e68cd0e99 100644
> > > --- a/drivers/scsi/esp_scsi.c
> > > +++ b/drivers/scsi/esp_scsi.c
> > > @@ -2788,7 +2788,7 @@ static inline unsigned int esp_wait_for_fifo(struct
> > > esp *esp)
> > >  		if (fbytes)
> > >  			return fbytes;
> > >
> > > -		udelay(2);
> > > +		udelay(1);
> > >  	} while (--i);
> > >
> > >  	pr_err("FIFO is empty (sreg %02x)\n", esp_read8(ESP_STATUS));
> > > @@ -2804,7 +2804,7 @@ static inline int esp_wait_for_intr(struct esp *esp)
> > >  		if (esp->sreg & ESP_STAT_INTR)
> > >  			return 0;
> > >
> > > -		udelay(2);
> > > +		udelay(1);
> > >  	} while (--i);
> > >
> > >  	pr_err("IRQ timeout (sreg %02x)\n", esp->sreg);
> > > @@ -2831,7 +2831,7 @@ void esp_send_pio_cmd(struct esp *esp, u32 addr, u32
> > > esp_count,
> > >  			if (!esp_wait_for_fifo(esp))
> > >  				break;
> > >
> > > -			*dst++ = esp_read8(ESP_FDATA);
> > > +			*dst++ = readb(esp->fifo_reg);
> > >  			--esp_count;
> > >
> > >  			if (!esp_count)
> > > @@ -2852,15 +2852,15 @@ void esp_send_pio_cmd(struct esp *esp, u32 addr, u32
> > > esp_count,
> > >  			}
> > >
> > >  			if (phase == ESP_MIP)
> > > -				scsi_esp_cmd(esp, ESP_CMD_MOK);
> > > +				esp_write8(ESP_CMD_MOK, ESP_CMD);
> >
> > You're no longer logging this command with this patch. (That'll be the reason
> > for the speedup you saw ...)
> >
> > > -			scsi_esp_cmd(esp, ESP_CMD_TI);
> > > +			esp_write8(ESP_CMD_TI, ESP_CMD);
> >
> > Same here..
> >
> > >  		}
> > >  	} else {
> > >  		unsigned int n = ESP_FIFO_SIZE;
> > >  		u8 *src = (u8 *)addr;
> > >
> > > -		scsi_esp_cmd(esp, ESP_CMD_FLUSH);
> > > +		esp_write8(ESP_CMD_FLUSH, ESP_CMD);
> >
> > here..
> >
> > >  		if (n > esp_count)
> > >  			n = esp_count;
> > > @@ -2894,7 +2894,7 @@ void esp_send_pio_cmd(struct esp *esp, u32 addr, u32
> > > esp_count,
> > >  			src += n;
> > >  			esp_count -= n;
> > >
> > > -			scsi_esp_cmd(esp, ESP_CMD_TI);
> > > +			esp_write8(ESP_CMD_TI, ESP_CMD);
> >
> > and here.
>
> Yes, it's deliberate.

I'm sure it was... and I wasn't objecting to that.

> > The burst of ESP_CMD_TI's in the log was quite useful to spot what went
> > wrong during PIO.
>
> I don't think it's as useful as you seem to think. Compare
> mac_esp_send_pdma_cmd().
>
> > Maybe mention in the changelog that commands during PIO are no longer
> > logged? Or introduce a new ESP_EVENT_PIO and log that at the start of
> > PIO?
>
> Yes, and I did leave a scsi_esp_cmd(esp, cmd) call at the start of PIO.

Which I missed from just looking at the patch, sorry.

> That should be sufficient, right?

It would indeed. Thanks for clarifying.

Cheers,

	Michael