[PATCH -mm] ide_end_drive_cmd(): avoid instruction pipeline stall

[PATCH -mm] ide_end_drive_cmd(): avoid instruction pipeline stall

[Andreas Mohr]

Use an independently-formatted "unsigned int" for data instead of a
restrictive "u16" to avoid instruction fetch pipeline stalls
probably caused by the byte calculations later.


ide_end_drive_cmd() uses an u16 variable for the result of an INW()
which it then does some byte masking operations on.
On my P3/700, this results in a highly visible IFU_MEM_STALL oprofile blip
when doing a simple "load 30 larger GUI apps in parallel" benchmark
(which takes about 1:30 or so, BTW):
The ide_end_drive_cmd() IFU_MEM_STALL amounts to 0.59% of all IFU_MEM_STALL
events during the profiling, with this opcode line amounting to > 95%
IFU_MEM_STALL within the function itself.

Replacing the u16 by an architecture-independently formatted unsigned int
to ease the byte-masking operations:

	/* no u16 here: caused severe IFU_MEM_STALL! */
	unsigned int data                               = hwif->INW(IDE_DATA_REG);
	args->tfRegister[IDE_DATA_OFFSET]       = (data) & 0xFF;
	args->hobRegister[IDE_DATA_OFFSET]      = (data >> 8) & 0xFF;

completely puts ide_end_drive_cmd() off the IFU_MEM_STALL radar during
repeated profiling attempts (after a fresh reboot with the modified kernel),
as opposed to having been the *top* oprofile trace item before.

I suppose that this is something like a textbook example of why it's
sometimes not beneficial to not use native-sized (i.e., 32bit) variables,
right?

Run-tested on 2.6.17-mm4.

Signed-off-by: Andreas Mohr <andi@lisas.de>


diff -urN linux-2.6.17-mm4.orig/drivers/ide/ide-io.c linux-2.6.17-mm4.my/drivers/ide/ide-io.c
--- linux-2.6.17-mm4.orig/drivers/ide/ide-io.c	2006-06-29 11:57:12.000000000 +0200
+++ linux-2.6.17-mm4.my/drivers/ide/ide-io.c	2006-06-30 11:54:12.000000000 +0200
@@ -397,7 +397,8 @@
 			
 		if (args) {
 			if (args->tf_in_flags.b.data) {
-				u16 data				= hwif->INW(IDE_DATA_REG);
+				/* no u16 here: caused severe IFU_MEM_STALL! */
+				unsigned int data				= hwif->INW(IDE_DATA_REG);
 				args->tfRegister[IDE_DATA_OFFSET]	= (data) & 0xFF;
 				args->hobRegister[IDE_DATA_OFFSET]	= (data >> 8) & 0xFF;
 			}

Re: [PATCH -mm] ide_end_drive_cmd(): avoid instruction pipeline stall

[Alan Cox]

Ar Gwe, 2006-06-30 am 18:13 +0200, ysgrifennodd Andreas Mohr:
> Use an independently-formatted "unsigned int" for data instead of a
> restrictive "u16" to avoid instruction fetch pipeline stalls
> probably caused by the byte calculations later.

drivers/ide is on its way out. I'm also curious that this shows up given
that the inw() is going to cause a PCI sequence and stall the CPU
entirely for ages anyway.

NAK because
1. This is a gcc problem
2. Not everyone is using an intel x86-32 box which has such problems
3. IDE is in life-support mode and the relatives are already planning
the flowers.

Alan

Re: [PATCH -mm] ide_end_drive_cmd(): avoid instruction pipeline stall

[Andrew Morton]

On Fri, 30 Jun 2006 18:26:56 +0100
Alan Cox <alan@lxorguk.ukuu.org.uk> wrote:

> Ar Gwe, 2006-06-30 am 18:13 +0200, ysgrifennodd Andreas Mohr:
> > Use an independently-formatted "unsigned int" for data instead of a
> > restrictive "u16" to avoid instruction fetch pipeline stalls
> > probably caused by the byte calculations later.
> 
> drivers/ide is on its way out.

Like sound/oss ;)

> I'm also curious that this shows up given
> that the inw() is going to cause a PCI sequence and stall the CPU
> entirely for ages anyway.

I guess because he was profiling for IFU_MEM_STALL, not for wall-time.

> NAK because
> 1. This is a gcc problem
> 2. Not everyone is using an intel x86-32 box which has such problems
> 3. IDE is in life-support mode and the relatives are already planning
> the flowers.

Well.  If the patch breaks anything we can dine on hats for a month.  Seems
pretty inoffensive to me.

Re: [PATCH -mm] ide_end_drive_cmd(): avoid instruction pipeline stall

[Alan Cox]

Ar Gwe, 2006-06-30 am 11:00 -0700, ysgrifennodd Andrew Morton:
> I guess because he was profiling for IFU_MEM_STALL, not for wall-time.
> 
> > NAK because
> > 1. This is a gcc problem
> > 2. Not everyone is using an intel x86-32 box which has such problems
> > 3. IDE is in life-support mode and the relatives are already planning
> > the flowers.
> 
> Well.  If the patch breaks anything we can dine on hats for a month.  Seems
> pretty inoffensive to me.

Yeah its a do nothing change that has no value. We can equally not apply
it and see no difference. Its make-work that might slow down some
systems.

Send it to the gcc people see if its a case they could/should optimise

Alan