On 10/01/2013 05:54 AM, Ralf Baechle wrote:
On Mon, Sep 30, 2013 at 02:29:02PM -0700, Leonid Yegoshin wrote:
On 09/24/2013 10:27 PM, Ralf Baechle wrote:
Commit 006a851b10a395955c153a145ad8241494d43688 adds 74K support in c-r4k.c:
+static inline void alias_74k_erratum(struct cpuinfo_mips *c)
+{
+ /*
+ * Early versions of the 74K do not update the cache tags on a
+ * vtag miss/ptag hit which can occur in the case of KSEG0/KUSEG
+ * aliases. In this case it is better to treat the cache as always
+ * having aliases.
+ */
+ if ((c->processor_id & 0xff) <= PRID_REV_ENCODE_332(2, 4, 0))
+ c->dcache.flags |= MIPS_CACHE_VTAG;
+ if ((c->processor_id & 0xff) == PRID_REV_ENCODE_332(2, 4, 0))
+ write_c0_config6(read_c0_config6() | MIPS_CONF6_SYND);
+ if (((c->processor_id & 0xff00) == PRID_IMP_1074K) &&
+ ((c->processor_id & 0xff) <= PRID_REV_ENCODE_332(1, 1, 0))) {
+ c->dcache.flags |= MIPS_CACHE_VTAG;
+ write_c0_config6(read_c0_config6() | MIPS_CONF6_SYND);
+ }
+}
But MIPS D-caches are never virtually tagged, so there is nothing in
the kernel that ever tests the MIPS_CACHE_VTAG flag in a D-cache
descriptor.
Cargo cult programming at its finest? Or was MIPS_CACHE_ALIASES what
really was meant to be set?
Ralf
There is a problem on early versions of 74K/1074K which can be effectively cured by setting MIPS_CACHE_VTAG.
It enforces the needed cache handling.
I hope it will go away as customers replace RTL for newer versions.
But I prefer the patch version from Maciej W. Rozycki, it is more clear.
There is nothing in the kernel that _reads_ that flag if it's set in an
I-cache descriptor. See:
arch/mips/include/asm/cpu-features.h:#define cpu_has_vtag_icache (cpu_data[0].icache.flags & MIPS_CACHE_VTAG)
arch/mips/include/asm/cpu-info.h:#define MIPS_CACHE_VTAG 0x00000002 /* Virtually tagged cache */
arch/mips/mm/c-octeon.c: c->icache.flags |= MIPS_CACHE_VTAG;
arch/mips/mm/c-octeon.c: c->icache.flags |= MIPS_CACHE_VTAG;
arch/mips/mm/c-octeon.c: c->icache.flags |= MIPS_CACHE_VTAG;
arch/mips/mm/c-r4k.c: c->dcache.flags |= MIPS_CACHE_VTAG;
arch/mips/mm/c-r4k.c: c->dcache.flags |= MIPS_CACHE_VTAG;
arch/mips/mm/c-r4k.c: c->icache.flags |= MIPS_CACHE_VTAG;
arch/mips/mm/c-r4k.c: c->icache.flags |= MIPS_CACHE_VTAG;
arch/mips/mm/c-r4k.c: c->icache.flags & MIPS_CACHE_VTAG ? "VIVT" : "VIPT",
There simply is not support whatsoever for virtually tagged D-caches in
the kernel, so setting that flag may feel good - but achieves nothing.
See, the sole test for the MIPS_CACHE_VTAG above tests it in a D-cache
descriptor. So I'm not sure what the intention is. Maybe there is some
code that has not yet been merged?
Ralf
Damn, you are right - it requires a patch
http://patchwork.linux-mips.org/patch/5939/
to work. There is cpu_has_vtag_dcache definition and an appropriate code
in it.
It seems to happen during unfortunate internal patch merge/split,
initially all that stuff was in separate patch from HIGHMEM.
- Leonid.
