601 vs head.S

[Franz Sirl]

Am Sun, 21 Feb 1999 schrieb Paul Mackerras:
>Brad Midgley <brad at pht.com> wrote:
>
>> my quest is for a kernel that boots both on the 7200 and the imac. help me
>
>Found it... had to take out the 2 nop's in hash_page this time.  Vger
>(cvs) and samba (rsync) have the fix, if you want a precompiled
>kernel, try ftp://ftp.linuxppc.org/linuxppc/kernel/powermac/test/vmlinux.

Ugh, a workaround again. As somebody will break that again for sure, I thought
a little bit about it and came up with the appended patch. An (possible?)
explanation is included. I tried it with 0-7 nop's at the old position after
the found_slot label and it always worked. I even did remove the SYNC here and
it still worked. For this final patch I left the SYNC in, cause I don't know
what it is good for. It is against Cort's 990216 vger snapshot, tested on my
7200/75.

Franz.


--- head.S-org	Wed Feb 17 06:59:32 1999
+++ head.S	Sun Feb 21 22:20:49 1999
@@ -1361,13 +1361,6 @@
 found_slot:
 	stw	r6,4(r3)
 	SYNC
-/*
- * These nop's seem to be necessary to avoid getting a machine
- * check on the rfi on 601 processors.
- */
-	nop
-	nop
-
 #else /* __SMP__ */
 /*
  * Between the tlbie above and updating the hash table entry below,
@@ -1441,7 +1434,22 @@
 	lwz	r21,GPR21(r21)
 	SYNC
 	rfi
-	
+
+/* 
+ * These 8 nop's are here to avoid lockup of the 601 on the previous rfi
+ * instruction. They make sure that rfi gets a different cache sector than
+ * the following hash_page_out code and thus the rfi instruction is not 
+ * prefetched by not-taken branch instructions in the hash_page code.
+ */
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+
 hash_page_out:
 #ifdef __SMP__
 	lis	r2,hash_table_lock at ha
-------------- next part --------------
A non-text attachment was scrubbed...
Name: linux-final.head.S.patch
Type: text/english
Size: 832 bytes
Desc: not available
URL: <http://lists.ozlabs.org/pipermail/linuxppc-dev/attachments/19990221/0d4d44a8/attachment.bin>