Re: [patch] cache flush bug in mm/filemap.c (all kernels >= 2.5.30(at least))

[Lothar Wassmann]

Russell King writes:
> Lothar - can you confirm that your problem vanishes when you turn off
> write allocation on the caches please?  (cachepolicy=writeback)
> 
No, its still there. But I seem to be unable to turn writealloc ON
anyway. I get:
|CPU: XScale-PXA250 [69052904] revision 4 (ARMv5TE)
|CPU: D undefined 5 cache
|CPU: I cache: 32768 bytes, associativity 32, 32 byte lines, 32 sets
|CPU: D cache: 32768 bytes, associativity 32, 32 byte lines, 32 sets
|Machine: KARO electronics PXA25x module
|Memory policy: ECC disabled, Data cache write back
no matter whether I specify 'cachepolicy=writeback', '...writealloc'
or no cachepolicy at all.
(This is because ARMv5TE is not recognized as an ARMv5 architecture and
 PMD_SECT_WBWA is turned into PMD_SECT_WB in build_mem_type_table()
 ('arch/arm/mm/mm-armv.c' line 304).
 Another bug?)
Shouldn't it be:
@@ -295,12 +295,13 @@
 	/*
 	 * ARMv5 can use ECC memory.
 	 */
-	if (cpu_arch == CPU_ARCH_ARMv5) {
+	if (cpu_arch == CPU_ARCH_ARMv5 || cpu_arch == CPU_ARCH_ARMv5T ||
+		cpu_arch == CPU_ARCH_ARMv5TE) {
 		mem_types[MT_VECTORS].prot_l1 |= ecc_mask;
 		mem_types[MT_MEMORY].prot_sect |= ecc_mask;
 	} else {
 		mem_types[MT_MINICLEAN].prot_sect &= ~PMD_SECT_TEX(1);
 		if (cachepolicy == PMD_SECT_WBWA)
 			cachepolicy = PMD_SECT_WB;
 		ecc_mask = 0;
 	}

Only 'cachepolicy=writethrough' makes the problem disappear.

Lothar