aic7xxx (O2 scsi) DMA coherency

aic7xxx (O2 scsi) DMA coherency

[Vivien Chappelier]

[-- Attachment #1: Type: TEXT/PLAIN, Size: 84 bytes --]

Hello,

	This tells the aic7xxx to use DMA safe memory for I/O.

Vivien Chappelier.

[-- Attachment #2: Type: TEXT/plain, Size: 606 bytes --]

diff -Naur linux/drivers/scsi/aic7xxx/aic7xxx_linux_host.h linux.patch/drivers/scsi/aic7xxx/aic7xxx_linux_host.h
--- linux/drivers/scsi/aic7xxx/aic7xxx_linux_host.h	Thu Dec 20 18:31:11 2001
+++ linux.patch/drivers/scsi/aic7xxx/aic7xxx_linux_host.h	Thu Dec 20 21:43:31 2001
@@ -87,7 +87,7 @@
 	sg_tablesize: 0,	/* max scatter-gather cmds    */\
 	cmd_per_lun: 2,		/* cmds per lun		      */\
 	present: 0,		/* number of 7xxx's present   */\
-	unchecked_isa_dma: 0,	/* no memory DMA restrictions */\
+	unchecked_isa_dma: 1,					\
 	use_clustering: ENABLE_CLUSTERING,			\
 	highmem_io: 1						\
 }

Re: aic7xxx (O2 scsi) DMA coherency

[Alan Cox]

> 	This tells the aic7xxx to use DMA safe memory for I/O.

That seems totally inappropriate. The unchecked dma option is for
ancient ISA DMA controllers that didnt do the 16Mb check. If you
find you need it debug your pci remapper

Re: aic7xxx (O2 scsi) DMA coherency

[Alan Cox]

> 	This tells the aic7xxx to use DMA safe memory for I/O.

That seems totally inappropriate. The unchecked dma option is for
ancient ISA DMA controllers that didnt do the 16Mb check. If you
find you need it debug your pci remapper

Re: aic7xxx (O2 scsi) DMA coherency

[Vivien Chappelier]

On Thu, 3 Jan 2002, Alan Cox wrote:

> > 	This tells the aic7xxx to use DMA safe memory for I/O.
> 
> That seems totally inappropriate. The unchecked dma option is for
> ancient ISA DMA controllers that didnt do the 16Mb check. If you
> find you need it debug your pci remapper

This is used when scaning for devices (drivers/scsi/scsi_scan.c) . When
this flag is not set, the code uses memory from the stack (unsigned char
scsi_result0[256]; in scan_scsis) instead of kmallocating it DMA safe as
it should on non-coherent systems. Maybe this is the thing to change?

regards,
Vivien Chappelier.

Re: aic7xxx (O2 scsi) DMA coherency

[Alan Cox]

> scsi_result0[256]; in scan_scsis) instead of kmallocating it DMA safe as
> it should on non-coherent systems. Maybe this is the thing to change?

Please fix that - thats a real bug. Actually you may find the PPC64 people
(Anton and co) already have a patch you can use

Re: aic7xxx (O2 scsi) DMA coherency

[Alan Cox]

> scsi_result0[256]; in scan_scsis) instead of kmallocating it DMA safe as
> it should on non-coherent systems. Maybe this is the thing to change?

Please fix that - thats a real bug. Actually you may find the PPC64 people
(Anton and co) already have a patch you can use

Re: aic7xxx (O2 scsi) DMA coherency

[Ralf Baechle]

On Thu, Jan 03, 2002 at 10:51:51PM +0100, Vivien Chappelier wrote:

> > > 	This tells the aic7xxx to use DMA safe memory for I/O.
> > 
> > That seems totally inappropriate. The unchecked dma option is for
> > ancient ISA DMA controllers that didnt do the 16Mb check. If you
> > find you need it debug your pci remapper
> 
> This is used when scaning for devices (drivers/scsi/scsi_scan.c) . When
> this flag is not set, the code uses memory from the stack (unsigned char
> scsi_result0[256]; in scan_scsis) instead of kmallocating it DMA safe as
> it should on non-coherent systems. Maybe this is the thing to change?

Indeed, it is.  I thought this one died ages ago.

  Ralf

Re: aic7xxx (O2 scsi) DMA coherency

[Vivien Chappelier]

[-- Attachment #1: Type: TEXT/PLAIN, Size: 899 bytes --]

On Thu, 3 Jan 2002, Ralf Baechle wrote:

> On Thu, Jan 03, 2002 at 10:51:51PM +0100, Vivien Chappelier wrote:
> 
> > > > 	This tells the aic7xxx to use DMA safe memory for I/O.
> > > 
> > > That seems totally inappropriate. The unchecked dma option is for
> > > ancient ISA DMA controllers that didnt do the 16Mb check. If you
> > > find you need it debug your pci remapper
> > 
> > This is used when scaning for devices (drivers/scsi/scsi_scan.c) . When
> > this flag is not set, the code uses memory from the stack (unsigned char
> > scsi_result0[256]; in scan_scsis) instead of kmallocating it DMA safe as
> > it should on non-coherent systems. Maybe this is the thing to change?
> 
> Indeed, it is.  I thought this one died ages ago.

Here is a patch to fix that then. It forces allocation of DMA safe
memory in any case. I've not looked at the PPC64 patch however.

regards,
Vivien Chappelier.

[-- Attachment #2: Type: TEXT/PLAIN, Size: 952 bytes --]

--- linux/drivers/scsi/scsi_scan.c	Thu Dec 20 18:31:09 2001
+++ linux.patch/drivers/scsi/scsi_scan.c	Fri Jan  4 13:17:31 2002
@@ -283,7 +283,6 @@
 	unsigned int lun;
 	unsigned int max_dev_lun;
 	unsigned char *scsi_result;
-	unsigned char scsi_result0[256];
 	Scsi_Device *SDpnt;
 	Scsi_Device *SDtail;
 	unsigned int sparse_lun;
@@ -305,8 +304,7 @@
 		scsi_initialize_queue(SDpnt, shpnt);
 		SDpnt->request_queue.queuedata = (void *) SDpnt;
 		/* Make sure we have something that is valid for DMA purposes */
-		scsi_result = ((!shpnt->unchecked_isa_dma)
-			       ? &scsi_result0[0] : kmalloc(512, GFP_DMA));
+		scsi_result = kmalloc(512, GFP_DMA);
 	}
 
 	if (scsi_result == NULL) {
@@ -463,7 +461,7 @@
 	}
 
 	/* If we allocated a buffer so we could do DMA, free it now */
-	if (scsi_result != &scsi_result0[0] && scsi_result != NULL) {
+	if (scsi_result != NULL) {
 		kfree(scsi_result);
 	} {
 		Scsi_Device *sdev;

Re: aic7xxx (O2 scsi) DMA coherency

[Vivien Chappelier]

[-- Attachment #1: Type: TEXT/PLAIN, Size: 567 bytes --]

Hello,

On Thu, 3 Jan 2002, Alan Cox wrote:

> > 	This tells the aic7xxx to use DMA safe memory for I/O.
> 
> That seems totally inappropriate. The unchecked dma option is for
> ancient ISA DMA controllers that didnt do the 16Mb check.
> 

Maybe this is more appropriate then :)
It tells the scsi scanner to always use DMA safe memory when doing its
INQUIRY commands...
On the O2, kernel is running in cacheable, non-coherent memory. kmalloc,
with GFP_DMA flag will return a piece of uncacheable memory which is safe
for I/O with devices.

regards,
Vivien Chappelier

[-- Attachment #2: Type: TEXT/plain, Size: 952 bytes --]

--- linux/drivers/scsi/scsi_scan.c	Thu Dec 20 18:31:09 2001
+++ linux.patch/drivers/scsi/scsi_scan.c	Fri Jan  4 13:17:31 2002
@@ -283,7 +283,6 @@
 	unsigned int lun;
 	unsigned int max_dev_lun;
 	unsigned char *scsi_result;
-	unsigned char scsi_result0[256];
 	Scsi_Device *SDpnt;
 	Scsi_Device *SDtail;
 	unsigned int sparse_lun;
@@ -305,8 +304,7 @@
 		scsi_initialize_queue(SDpnt, shpnt);
 		SDpnt->request_queue.queuedata = (void *) SDpnt;
 		/* Make sure we have something that is valid for DMA purposes */
-		scsi_result = ((!shpnt->unchecked_isa_dma)
-			       ? &scsi_result0[0] : kmalloc(512, GFP_DMA));
+		scsi_result = kmalloc(512, GFP_DMA);
 	}
 
 	if (scsi_result == NULL) {
@@ -463,7 +461,7 @@
 	}
 
 	/* If we allocated a buffer so we could do DMA, free it now */
-	if (scsi_result != &scsi_result0[0] && scsi_result != NULL) {
+	if (scsi_result != NULL) {
 		kfree(scsi_result);
 	} {
 		Scsi_Device *sdev;

Re: aic7xxx (O2 scsi) DMA coherency

[Vivien Chappelier]

On Fri, 4 Jan 2002, Ilya Volynets wrote:

> On Friday 04 January 2002 11:33 am, you wrote:
> This is true in theory. In practice you have to look at kmalloc implementation
> and how it uses GFP_DMA. I don't remember anything arch specific in there,
> but I never seriously dug in that part. I believe it doesn't do what we need,
> otherwice We'd be using it for pci_alloc_consistent.

Well, both pci_alloc_consistent and kmalloc will end up calling
__get_free_pages, with the GFP_DMA flag set. On setup, with the MIPS64
arch, DMA memory is set to start in non-cacheable space (don't know how).
So both pci_alloc_consistent and kmalloc(GFP_DMA,...) will return DMA safe
(i.e. non cacheable) memory.
see mm/slab.c (for kmalloc and how it finally calls __get_free_pages).

regards,
Vivien Chappelier.