Skip to contentSkip to conversation

About invalidate cache of SH4

PATCH Yoshihiro Shimoda 2 participants 2 messages Jan 8, 2008 SuperH

lore
Share & tools
lore

Yoshihiro Shimoda

author
#0 lore raw

Hi, Paul,

When start or size of __flush_invalidate_region() is not cache alignment, enlarge the range to match it. However, the data is lost if there are the data which should write-back in there.

For example, struct scsi_cmnd has sense_buffer[]. sense_buffer is not cache alignment. When there is invalidated sense_buffer, struct request *request of struct scsi_cmnd may be invalidated. Connection of the USB memory sometimes fails and cannnot mount USB-CDROM(oops occured) for this processing.

I made a patch, but this implementation may not be good...

Thanks,
Yoshihiro Shimoda

---

arch/sh/mm/cache-sh4.c +22 -0
index 43d7ff6..cc1751d 100644
@@ -164,15 +164,37 @@ void __flush_invalidate_region(void *start, int size)
 {
 	unsigned long v;
 	unsigned long begin, end;
+	int no_align = 0;
 
 	begin = (unsigned long)start & ~(L1_CACHE_BYTES-1);
 	end = ((unsigned long)start + size + L1_CACHE_BYTES-1)
 		& ~(L1_CACHE_BYTES-1);
+
+	if ((unsigned long)start & (L1_CACHE_BYTES - 1) ||
+	    size & (L1_CACHE_BYTES - 1)) {
+		if (size <= (L1_CACHE_BYTES << 1)) {
+			__flush_purge_region(start, size);
+			return;
+		} else {
+			no_align = 1;
+			begin += L1_CACHE_BYTES;
+			end -= L1_CACHE_BYTES;
+		}
+	}
+
+	if (no_align)
+		__flush_purge_region((void *)(begin - L1_CACHE_BYTES),
+				     L1_CACHE_BYTES);
+
 	for (v = begin; v < end; v+=L1_CACHE_BYTES) {
 		asm volatile("ocbi	%0"
 			     : /* no output */
 			     : "m" (__m(v)));
 	}
+
+	if (no_align)
+		__flush_purge_region((void *)(end + L1_CACHE_BYTES),
+				     L1_CACHE_BYTES);
 }
 
 /*
#1 lore raw
Quoted 6 lines · Yoshihiro Shimoda

For example, struct scsi_cmnd has sense_buffer[]. sense_buffer is not cache alignment. When there is invalidated sense_buffer, struct request *request of struct scsi_cmnd may be invalidated. Connection of the USB memory sometimes fails and cannnot mount USB-CDROM(oops occured) for this processing.

This is a known issue, and is currently being debated on l-k:

http://lkml.org/lkml/2007/12/20/660

Once there's a resolution, the driver will either be fixed, or we set up the minimum alignment to a cacheline boundary.