* [PATCH 002/002] de2104x: support for systems lacking cache coherence
@ 2009-02-07 21:30 Risto Suominen
2009-02-10 1:01 ` Andrew Morton
0 siblings, 1 reply; 4+ messages in thread
From: Risto Suominen @ 2009-02-07 21:30 UTC (permalink / raw)
To: Jeff Garzik, lkml
[-- Attachment #1: Type: text/plain, Size: 2177 bytes --]
Add a configurable Descriptor Skip Length for systems that lack cache coherence.
Signed-off-by: Risto Suominen <Risto.Suominen@gmail.com>
---
The testing is done on kernel version 2.6.18.
--- drivers/net/tulip/Kconfig.org 2006-09-20 06:42:06.000000000 +0300
+++ drivers/net/tulip/Kconfig 2009-02-07 20:48:17.000000000 +0200
@@ -27,6 +27,18 @@ config DE2104X
To compile this driver as a module, choose M here. The module will
be called de2104x.
+config DE2104X_DSL
+ int "Descriptor Skip Length in 32 bit longwords"
+ depends on DE2104X
+ range 0 31
+ default 0
+ help
+ Setting this value allows to align ring buffer descriptors into their
+ own cache lines. Value of 4 corresponds to the typical 32 byte line
+ (the descriptor is 16 bytes). This is necessary on systems that lack
+ cache coherence, an example is PowerMac 5500. Otherwise 0 is safe.
+ Default is 0, and range is 0 to 31.
+
config TULIP
tristate "DECchip Tulip (dc2114x) PCI support"
depends on PCI
--- drivers/net/tulip/de2104x.c.org 2006-09-20 06:42:06.000000000 +0300
+++ drivers/net/tulip/de2104x.c 2009-02-07 15:04:04.000000000 +0200
@@ -82,6 +82,13 @@ MODULE_PARM_DESC (rx_copybreak, "de2104x
NETIF_MSG_RX_ERR | \
NETIF_MSG_TX_ERR)
+/* Descriptor skip length in 32 bit longwords. */
+#ifndef CONFIG_DE2104X_DSL
+#define DSL 0
+#else
+#define DSL CONFIG_DE2104X_DSL
+#endif
+
#define DE_RX_RING_SIZE 64
#define DE_TX_RING_SIZE 64
#define DE_RING_BYTES \
@@ -153,6 +160,7 @@ enum {
CmdReset = (1 << 0),
CacheAlign16 = 0x00008000,
BurstLen4 = 0x00000400,
+ DescSkipLen = (DSL << 2),
/* Rx/TxPoll bits */
NormalTxPoll = (1 << 0),
@@ -246,7 +254,7 @@ static const u32 de_intr_mask =
* Set the programmable burst length to 4 longwords for all:
* DMA errors result without these values. Cache align 16 long.
*/
-static const u32 de_bus_mode = CacheAlign16 | BurstLen4;
+static const u32 de_bus_mode = CacheAlign16 | BurstLen4 | DescSkipLen;
struct de_srom_media_block {
u8 opts;
@@ -266,6 +274,9 @@ struct de_desc {
__le32 opts2;
__le32 addr1;
__le32 addr2;
+#if DSL
+ __le32 skip[DSL];
+#endif
};
struct media_info {
[-- Attachment #2: de2104x-dsl.patch --]
[-- Type: text/x-diff, Size: 2182 bytes --]
Add a configurable Descriptor Skip Length for systems that lack cache coherence.
Signed-off-by: Risto Suominen <Risto.Suominen@gmail.com>
---
The testing is done on kernel version 2.6.18.
--- drivers/net/tulip/Kconfig.org 2006-09-20 06:42:06.000000000 +0300
+++ drivers/net/tulip/Kconfig 2009-02-07 20:48:17.000000000 +0200
@@ -27,6 +27,18 @@ config DE2104X
To compile this driver as a module, choose M here. The module will
be called de2104x.
+config DE2104X_DSL
+ int "Descriptor Skip Length in 32 bit longwords"
+ depends on DE2104X
+ range 0 31
+ default 0
+ help
+ Setting this value allows to align ring buffer descriptors into their
+ own cache lines. Value of 4 corresponds to the typical 32 byte line
+ (the descriptor is 16 bytes). This is necessary on systems that lack
+ cache coherence, an example is PowerMac 5500. Otherwise 0 is safe.
+ Default is 0, and range is 0 to 31.
+
config TULIP
tristate "DECchip Tulip (dc2114x) PCI support"
depends on PCI
--- drivers/net/tulip/de2104x.c.org 2006-09-20 06:42:06.000000000 +0300
+++ drivers/net/tulip/de2104x.c 2009-02-07 15:04:04.000000000 +0200
@@ -82,6 +82,13 @@ MODULE_PARM_DESC (rx_copybreak, "de2104x
NETIF_MSG_RX_ERR | \
NETIF_MSG_TX_ERR)
+/* Descriptor skip length in 32 bit longwords. */
+#ifndef CONFIG_DE2104X_DSL
+#define DSL 0
+#else
+#define DSL CONFIG_DE2104X_DSL
+#endif
+
#define DE_RX_RING_SIZE 64
#define DE_TX_RING_SIZE 64
#define DE_RING_BYTES \
@@ -153,6 +160,7 @@ enum {
CmdReset = (1 << 0),
CacheAlign16 = 0x00008000,
BurstLen4 = 0x00000400,
+ DescSkipLen = (DSL << 2),
/* Rx/TxPoll bits */
NormalTxPoll = (1 << 0),
@@ -246,7 +254,7 @@ static const u32 de_intr_mask =
* Set the programmable burst length to 4 longwords for all:
* DMA errors result without these values. Cache align 16 long.
*/
-static const u32 de_bus_mode = CacheAlign16 | BurstLen4;
+static const u32 de_bus_mode = CacheAlign16 | BurstLen4 | DescSkipLen;
struct de_srom_media_block {
u8 opts;
@@ -266,6 +274,9 @@ struct de_desc {
__le32 opts2;
__le32 addr1;
__le32 addr2;
+#if DSL
+ __le32 skip[DSL];
+#endif
};
struct media_info {
^ permalink raw reply [flat|nested] 4+ messages in thread* Re: [PATCH 002/002] de2104x: support for systems lacking cache coherence
2009-02-07 21:30 [PATCH 002/002] de2104x: support for systems lacking cache coherence Risto Suominen
@ 2009-02-10 1:01 ` Andrew Morton
2009-02-10 1:07 ` David Miller
2009-02-10 7:16 ` Risto Suominen
0 siblings, 2 replies; 4+ messages in thread
From: Andrew Morton @ 2009-02-10 1:01 UTC (permalink / raw)
To: Risto Suominen; +Cc: jgarzik, linux-kernel, netdev, Grant Grundler
(cc netdev, and maintainer)
On Sat, 7 Feb 2009 23:30:40 +0200
Risto Suominen <risto.suominen@gmail.com> wrote:
> Add a configurable Descriptor Skip Length for systems that lack cache coherence.
>
> Signed-off-by: Risto Suominen <Risto.Suominen@gmail.com>
> ---
> The testing is done on kernel version 2.6.18.
>
> --- drivers/net/tulip/Kconfig.org 2006-09-20 06:42:06.000000000 +0300
> +++ drivers/net/tulip/Kconfig 2009-02-07 20:48:17.000000000 +0200
Please prefer to prepare patches in `patch -p1' form:
--- a/drivers/net/tulip/Kconfig
+++ a/drivers/net/tulip/Kconfig
> @@ -27,6 +27,18 @@ config DE2104X
> To compile this driver as a module, choose M here. The module will
> be called de2104x.
>
> +config DE2104X_DSL
> + int "Descriptor Skip Length in 32 bit longwords"
> + depends on DE2104X
> + range 0 31
> + default 0
> + help
> + Setting this value allows to align ring buffer descriptors into their
> + own cache lines. Value of 4 corresponds to the typical 32 byte line
> + (the descriptor is 16 bytes). This is necessary on systems that lack
> + cache coherence, an example is PowerMac 5500. Otherwise 0 is safe.
> + Default is 0, and range is 0 to 31.
> +
> config TULIP
> tristate "DECchip Tulip (dc2114x) PCI support"
> depends on PCI
> --- drivers/net/tulip/de2104x.c.org 2006-09-20 06:42:06.000000000 +0300
> +++ drivers/net/tulip/de2104x.c 2009-02-07 15:04:04.000000000 +0200
> @@ -82,6 +82,13 @@ MODULE_PARM_DESC (rx_copybreak, "de2104x
> NETIF_MSG_RX_ERR | \
> NETIF_MSG_TX_ERR)
>
> +/* Descriptor skip length in 32 bit longwords. */
> +#ifndef CONFIG_DE2104X_DSL
> +#define DSL 0
> +#else
> +#define DSL CONFIG_DE2104X_DSL
> +#endif
I think CONFIG_DE2104X_DSL is always defined if this driver is being
compiled. So the Kconfig `default' should suffice here? In which case
we can do
#define DSL CONFIG_DE2104X_DSL
and leave it at that.
> #define DE_RX_RING_SIZE 64
> #define DE_TX_RING_SIZE 64
> #define DE_RING_BYTES \
> @@ -153,6 +160,7 @@ enum {
> CmdReset = (1 << 0),
> CacheAlign16 = 0x00008000,
> BurstLen4 = 0x00000400,
> + DescSkipLen = (DSL << 2),
In which case we can do away with DSL everywhere and just use
CONFIG_DE2104X_DSL here.
But really, we shouldn't do this at configuration time at all. It
would be much better to do it at runtime, via a module parameter.
And it would be much^2 better to do it automatically, based upon the
chip probing information or whatever. That's probably hard.
> /* Rx/TxPoll bits */
> NormalTxPoll = (1 << 0),
> @@ -246,7 +254,7 @@ static const u32 de_intr_mask =
> * Set the programmable burst length to 4 longwords for all:
> * DMA errors result without these values. Cache align 16 long.
> */
> -static const u32 de_bus_mode = CacheAlign16 | BurstLen4;
> +static const u32 de_bus_mode = CacheAlign16 | BurstLen4 | DescSkipLen;
>
> struct de_srom_media_block {
> u8 opts;
> @@ -266,6 +274,9 @@ struct de_desc {
> __le32 opts2;
> __le32 addr1;
> __le32 addr2;
> +#if DSL
> + __le32 skip[DSL];
> +#endif
> };
Can remove the ifdefs here. A zero-length array is OK, and will
consume zero space.
^ permalink raw reply [flat|nested] 4+ messages in thread* Re: [PATCH 002/002] de2104x: support for systems lacking cache coherence
2009-02-10 1:01 ` Andrew Morton
@ 2009-02-10 1:07 ` David Miller
2009-02-10 7:16 ` Risto Suominen
1 sibling, 0 replies; 4+ messages in thread
From: David Miller @ 2009-02-10 1:07 UTC (permalink / raw)
To: akpm; +Cc: risto.suominen, jgarzik, linux-kernel, netdev, grundler
From: Andrew Morton <akpm@linux-foundation.org>
Date: Mon, 9 Feb 2009 17:01:20 -0800
>
> (cc netdev, and maintainer)
We're already discussing this patch on netdev, he resent it
there.
^ permalink raw reply [flat|nested] 4+ messages in thread
* Re: [PATCH 002/002] de2104x: support for systems lacking cache coherence
2009-02-10 1:01 ` Andrew Morton
2009-02-10 1:07 ` David Miller
@ 2009-02-10 7:16 ` Risto Suominen
1 sibling, 0 replies; 4+ messages in thread
From: Risto Suominen @ 2009-02-10 7:16 UTC (permalink / raw)
To: Andrew Morton; +Cc: jgarzik, linux-kernel, netdev, Grant Grundler
2009/2/10 Andrew Morton <akpm@linux-foundation.org>:
>
> Please prefer to prepare patches in `patch -p1' form:
>
> --- a/drivers/net/tulip/Kconfig
> +++ a/drivers/net/tulip/Kconfig
>
I'll try to remember that.
> I think CONFIG_DE2104X_DSL is always defined if this driver is being
> compiled. So the Kconfig `default' should suffice here? In which case
> we can do
>
> #define DSL CONFIG_DE2104X_DSL
>
> and leave it at that.
>
> In which case we can do away with DSL everywhere and just use
> CONFIG_DE2104X_DSL here.
>
I was thinking of the case where someone just compiles without reconfiguring.
> But really, we shouldn't do this at configuration time at all. It
> would be much better to do it at runtime, via a module parameter.
>
Well, yeah, but that would probably cost cpu cycles. At least it would
mean lots of changes.
> And it would be much^2 better to do it automatically, based upon the
> chip probing information or whatever. That's probably hard.
>
I agree.
> Can remove the ifdefs here. A zero-length array is OK, and will
> consume zero space.
>
I was uncertain whether this applies to all compilers. I kind of
remember opposite cases in the past.
Risto
^ permalink raw reply [flat|nested] 4+ messages in thread
end of thread, other threads:[~2009-02-10 7:16 UTC | newest]
Thread overview: 4+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2009-02-07 21:30 [PATCH 002/002] de2104x: support for systems lacking cache coherence Risto Suominen
2009-02-10 1:01 ` Andrew Morton
2009-02-10 1:07 ` David Miller
2009-02-10 7:16 ` Risto Suominen
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox
all inboxes | Powered by JetHome®