64bit DMA in i2o_block

64bit DMA in i2o_block

[Vasily Averin]

Markus,

I've found that your driver does not work correctly with following options:
CONFIG_I2O_EXT_ADAPTEC=y
# CONFIG_I2O_EXT_ADAPTEC_DMA64 is not set

I've checked it on 2.6.18 kernel on x86_64 node with 4Gb memory.

I would note that I've not noticed any bugs on x86 kernels, but x86_64 does not
work at all.

The behaviors of x86_64 kernels can differ:
sometimes driver can not find partition table:
i2o/hda: unknown partition table

sometimes it find not all partition:
i2o/hda: i2o/hda1 i2o/hda2 i2o/hda3 i2o/hda4 < >

in other cases it detects devices well but corrupt the memory and filesystem:
Sep 21 16:46:09 ts10 block-osm: registered device at major 80
Sep 21 16:46:09 ts10  i2o/hda: i2o/hda1 i2o/hda2 i2o/hda3 i2o/hda4 < i2o/hda5 >
Sep 21 16:46:09 ts10 block-osm: device added (TID: 207): i2o/hda
Sep 21 16:46:09 ts10 kjournald starting.  Commit interval 5 seconds
Sep 21 16:46:09 ts10 EXT3-fs: mounted filesystem with ordered data mode.
Sep 21 16:46:10 ts10 Freeing unused kernel memory: 236k freed
Sep 21 16:46:10 ts10 hotplug[961]: segfault at 0000000000000008 rip
00000032411088cf rsp 00007fff5b189050 error 4

When I enabled CONFIG_I2O_EXT_ADAPTEC_DMA64 -- all the troubles went away.

I've tried to investigate this issue, but I've not found and documentation
described format of Adaptec's private messages.

I would like to ask you, is this behavior specific for my hardware (Adaptec
ASR-2010S I2O Zero Channel)? If not, I believe it makes sense to enable
 CONFIG_I2O_EXT_ADAPTEC_DMA64 in Kconfig by default for 64-bit kernels.

Thank you,
	Vasily Averin

Re: 64bit DMA in i2o_block

[Markus Lidel]

[-- Attachment #1: Type: text/plain, Size: 2152 bytes --]

Hello,

Vasily Averin wrote:
> I've found that your driver does not work correctly with following options:
> CONFIG_I2O_EXT_ADAPTEC=y
> # CONFIG_I2O_EXT_ADAPTEC_DMA64 is not set
> I've checked it on 2.6.18 kernel on x86_64 node with 4Gb memory.

That could be the reason for the problems. I've only tried it on x86_64 
machines with < 4 GB memory...

> I would note that I've not noticed any bugs on x86 kernels, but x86_64 does not
> work at all.
> The behaviors of x86_64 kernels can differ:
> sometimes driver can not find partition table:
> i2o/hda: unknown partition table
> sometimes it find not all partition:
> i2o/hda: i2o/hda1 i2o/hda2 i2o/hda3 i2o/hda4 < >
> in other cases it detects devices well but corrupt the memory and filesystem:
> Sep 21 16:46:09 ts10 block-osm: registered device at major 80
> Sep 21 16:46:09 ts10  i2o/hda: i2o/hda1 i2o/hda2 i2o/hda3 i2o/hda4 < i2o/hda5 >
> Sep 21 16:46:09 ts10 block-osm: device added (TID: 207): i2o/hda
> Sep 21 16:46:09 ts10 kjournald starting.  Commit interval 5 seconds
> Sep 21 16:46:09 ts10 EXT3-fs: mounted filesystem with ordered data mode.
> Sep 21 16:46:10 ts10 Freeing unused kernel memory: 236k freed
> Sep 21 16:46:10 ts10 hotplug[961]: segfault at 0000000000000008 rip
> 00000032411088cf rsp 00007fff5b189050 error 4
> When I enabled CONFIG_I2O_EXT_ADAPTEC_DMA64 -- all the troubles went away.
> I've tried to investigate this issue, but I've not found and documentation
> described format of Adaptec's private messages.
> I would like to ask you, is this behavior specific for my hardware (Adaptec
> ASR-2010S I2O Zero Channel)? If not, I believe it makes sense to enable
>  CONFIG_I2O_EXT_ADAPTEC_DMA64 in Kconfig by default for 64-bit kernels.

I believe it's a driver issue. Do you have a testing-system where you 
could try out the appended patch?

Thank you very much.


Best regards,


Markus Lidel
------------------------------------------
Markus Lidel (Senior IT Consultant)

Shadow Connect GmbH
Carl-Reisch-Weg 12
D-86381 Krumbach
Germany

Phone:  +49 82 82/99 51-0
Fax:    +49 82 82/99 51-11

E-Mail: Markus.Lidel@shadowconnect.com
URL:    http://www.shadowconnect.com

[-- Attachment #2: i2o-test-32bit-dma.patch --]
[-- Type: text/x-patch, Size: 606 bytes --]

--- include/linux/i2o.h	2006-09-22 12:07:29.286557446 +0200
+++ include/linux/i2o.h	2006-09-22 12:08:01.904303452 +0200
@@ -776,6 +776,8 @@
 #ifdef CONFIG_I2O_EXT_ADAPTEC_DMA64
 		if ((sizeof(dma_addr_t) > 4) && c->pae_support)
 			*mptr++ = cpu_to_le32(i2o_dma_high(dma_addr));
+#else
+		BUG_ON((u32) ((u64) dma_addr >> 32));
 #endif
 		*sg_ptr = mptr;
 	}
@@ -835,6 +837,8 @@
 #ifdef CONFIG_I2O_EXT_ADAPTEC_DMA64
 		if ((sizeof(dma_addr_t) > 4) && c->pae_support)
 			*mptr++ = cpu_to_le32(i2o_dma_high(sg_dma_address(sg)));
+#else
+		BUG_ON((u32) ((u64) sg_dma_address(sg) >> 32));
 #endif
 		sg++;
 	}

Re: 64bit DMA in i2o_block

[Andi Kleen]

Markus Lidel <Markus.Lidel@shadowconnect.com> writes:
> 
> I believe it's a driver issue. Do you have a testing-system where you
> could try out the appended patch?

So it would bug instead? 

Can you please just get rid of that CONFIG? Such things shouldn't 
be CONFIGs, but have a suitable default. In particular drivers
running under 64bit should always support 64bit DMA.

-Andi

Re: 64bit DMA in i2o_block

[Vasily Averin]

Markus Lidel wrote:
> Hello,
> 
> Vasily Averin wrote:
>> I've found that your driver does not work correctly with following
>> options:
>> CONFIG_I2O_EXT_ADAPTEC=y
>> # CONFIG_I2O_EXT_ADAPTEC_DMA64 is not set
>> I've checked it on 2.6.18 kernel on x86_64 node with 4Gb memory.
> 
> I believe it's a driver issue. Do you have a testing-system where you
> could try out the appended patch?

Done. troubles are present, BUG_ON did not trigger.

thank you,
	Vasily Averin

Re: 64bit DMA in i2o_block

[Markus Lidel]

Hello,

Andi Kleen wrote:
> Markus Lidel <Markus.Lidel@shadowconnect.com> writes:
>> I believe it's a driver issue. Do you have a testing-system where you
>> could try out the appended patch?
> So it would bug instead? 

Yes, because ATM i don't know why it doesn't work. It's just for debugging.

> Can you please just get rid of that CONFIG? Such things shouldn't 
> be CONFIGs, but have a suitable default. In particular drivers
> running under 64bit should always support 64bit DMA.

The 64-bit DMA is only available on Adaptec controllers, all other I2O 
controllers only work with 32-bit DMA. Although the 64-bit DMA is (as far 
as i can remember) slower then the 32-bit DMA and is therefore only 
useful on systems which run out of memory.


Best regards,


Markus Lidel
------------------------------------------
Markus Lidel (Senior IT Consultant)

Shadow Connect GmbH
Carl-Reisch-Weg 12
D-86381 Krumbach
Germany

Phone:  +49 82 82/99 51-0
Fax:    +49 82 82/99 51-11

E-Mail: Markus.Lidel@shadowconnect.com
URL:    http://www.shadowconnect.com