Re: [patch] forcedeth: add support for interrupt mitigation

[Jeff Garzik <jgarzik@xxxxxxxxx>]

Manfred Spraul wrote:
> Hi,
>
> The current forcedeth driver doesn't support interrupt mitigation, this 
> can result in an incredible number of interrupts/sec for gigabit links. 
> The attached patch adds a throughput mode that enables an interrupt 
> mitigation scheme.
>
> --
>    Manfred
>
>
> ------------------------------------------------------------------------
>
> --- 2.6/drivers/net/forcedeth.c 2005-10-20 23:17:24.000000000 +0200
> +++ build-2.6/drivers/net/forcedeth.c   2005-10-19 21:01:55.000000000 +0200
> @@ -98,6 +98,7 @@
>   *     0.43: 10 Aug 2005: Add support for tx checksum.
>   *     0.44: 20 Aug 2005: Add support for scatter gather and segmentation.
>   *     0.45: 18 Sep 2005: Remove nv_stop/start_rx from every link check
> + *     0.46: 20 Aug 2005: Add irq optimization modes.
>   *
>   * Known bugs:
>   * We suspect that on some hardware no TX done interrupts are generated.
> @@ -109,7 +110,7 @@
>   * DEV_NEED_TIMERIRQ will not harm you on sane hardware, only generating a few
>   * superfluous timer interrupts from the nic.
>   */
> -#define FORCEDETH_VERSION              "0.45"
> +#define FORCEDETH_VERSION              "0.46"
>  #define DRV_NAME                       "forcedeth"
>  
>  #include <linux/module.h>
> @@ -164,7 +165,8 @@
>  #define NVREG_IRQ_LINK                 0x0040
>  #define NVREG_IRQ_TX_ERROR             0x0080
>  #define NVREG_IRQ_TX1                  0x0100
> -#define NVREG_IRQMASK_WANTED           0x00df
> +#define NVREG_IRQMASK_THROUGHPUT       0x00df
> +#define NVREG_IRQMASK_CPU              0x0040
>  
>  #define NVREG_IRQ_UNKNOWN	(~(NVREG_IRQ_RX_ERROR|NVREG_IRQ_RX|NVREG_IRQ_RX_NOBUF|NVREG_IRQ_TX_ERR| \
>                                         
> NVREG_IRQ_TX_OK|NVREG_IRQ_TIMER|NVREG_IRQ_LINK|NVREG_IRQ_TX_ERROR| \
> @@ -178,7 +180,8 @@
>   * NVREG_POLL_DEFAULT=97 would result in an interval length of 1 ms
>   */
>         NvRegPollingInterval = 0x00c,
> -#define NVREG_POLL_DEFAULT     970
> +#define NVREG_POLL_DEFAULT_THROUGHPUT  970
> +#define NVREG_POLL_DEFAULT_CPU 13
>         NvRegMisc1 = 0x080,
>  #define NVREG_MISC1_HD         0x02
>  #define NVREG_MISC1_FORCE      0x3b0f3c
> @@ -539,6 +542,18 @@
>   */
>  static int max_interrupt_work = 5;
>  
> +/*
> + * Optimization can be either throuput mode or cpu mode
> + */
> +#define NV_OPTIMIZATION_MODE_THROUGHPUT 0
> +#define NV_OPTIMIZATION_MODE_CPU        1
> +static int optimization_mode = NV_OPTIMIZATION_MODE_THROUGHPUT;
> +
> +/*
> + * Poll interval for timer irq
> + */
> +static int poll_interval = -1;
The code changes themselves seem correct, but this patch suffers from a 
few overall problems.
* "throughput or cpu" doesn't tell the user very much -- or me, for that 
matter.  Does "cpu mode" mean low latency, high interrupt count?  If so, 
just allow the user to choose between throughput and latency.
* making this a static decision at module load time is sub-optimal.  99% 
of users will simply use the default.  the option should be 
per-interface, ideally.  controlled by ethtool?
* there is zero information on how to use poll interval.  again, 99% of 
users will simply use the default.  since the poll interval is written 
directly to a hardware register, you should give the user some idea of 
the unit of measure (ms? ticks? bus cycles? ns?), and some idea of min/max.