linux-kernel - Re: [PATCH 6/6] esp_scsi: Optimize PIO loops

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Date:   Sat, 13 Oct 2018 17:02:09 +1300
From:   Michael Schmitz <schmitzmic@...il.com>
To:     Finn Thain <fthain@...egraphics.com.au>,
        "James E.J. Bottomley" <jejb@...ux.vnet.ibm.com>,
        "Martin K. Petersen" <martin.petersen@...cle.com>
Cc:     Hannes Reinecke <hare@...e.de>, linux-scsi@...r.kernel.org,
        linux-m68k@...ts.linux-m68k.org, linux-kernel@...r.kernel.org
Subject: Re: [PATCH 6/6] esp_scsi: Optimize PIO loops

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@...oo.com>
> Signed-off-by: Finn Thain <fthain@...egraphics.com.au>
> ---
>  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.

The burst of ESP_CMD_TI's in the log was quite useful to spot what went 
wrong during PIO. 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?

Cheers,

	Michael



>  		}
>  	}
>
>