Re: [patch] voluntary-preempt-2.6.9-rc1-bk4-R1

[Ingo Molnar <mingo@elte.hu>]

* Mark_H_Johnson@raytheon.com <Mark_H_Johnson@raytheon.com> wrote:

> Not quite sure where to add touch_preempt_timing() calls - somewhere in the
> loop in ide_outsl and ide_insl? [so we keep resetting the start /end
> times?]

> 00000001 0.002ms (+0.000ms): ata_output_data (taskfile_output_data)
> 00000001 0.003ms (+0.542ms): ide_outsl (ata_output_data)
> 00000001 0.545ms (+0.000ms): kunmap_atomic (ide_multwrite)

yeah, i'd add it to ide_outsl. Something like the patch below. (it
compiles but is untested otherwise.)

another possibility besides DMA starvation is starvation between CPUs. 
The only way to exclude DMA as a source of CPU starvation would be to
try a maxcpus=1 test using the SMP kernel. You dont even have to run the
RT-latency tester for this - just the disk read/write workload should
trigger the latency traces!

	Ingo

--- linux/drivers/ide/ide-iops.c.orig	
+++ linux/drivers/ide/ide-iops.c	
@@ -56,7 +56,15 @@ static u32 ide_inl (unsigned long port)
 
 static void ide_insl (unsigned long port, void *addr, u32 count)
 {
-	insl(port, addr, count);
+	unsigned int chunk, offset = 0;
+
+	while (count) {
+		chunk = min(128U, count);
+		insl(port, addr + offset, chunk);
+		count -= chunk;
+		offset += chunk;
+		touch_preempt_timing();
+	}
 }
 
 static void ide_outb (u8 val, unsigned long port)
@@ -86,7 +94,15 @@ static void ide_outl (u32 val, unsigned 
 
 static void ide_outsl (unsigned long port, void *addr, u32 count)
 {
-	outsl(port, addr, count);
+	unsigned int chunk, offset = 0;
+
+	while (count) {
+		chunk = min(128U, count);
+		outsl(port, addr + offset, chunk);
+		count -= chunk;
+		offset += chunk;
+		touch_preempt_timing();
+	}
 }
 
 void default_hwif_iops (ide_hwif_t *hwif)