Am 21.11.2010 um 22:53 schrieb Alexander Graf:
On 21.11.2010, at 22:40, Andreas Färber andreas.faerber@web.de wrote:
Am 21.11.2010 um 22:22 schrieb Alexander Graf:
On 21.11.2010, at 21:08, Andreas Färber wrote:
Clean up use of SDR1 fields.
IBM's "The Programming Environments for 32-Bit Microprocessors": The HTABMASK field in SDR1 contains a mask value that determines how many bits from the hash are used in the page table index. This mask must be of the form 0b00...011...1; that is, a string of 0 bits followed by a string of 1bits. The 1 bits determine how many additional bits (at least 10) from the hash are used in the index; HTABORG must have this same number of low-order bits equal to 0.
IBM's "The Programming Environments Manual for 64-bit Microprocessors" 3.0: The HTABSIZE field in SDR1 contains an integer value that determines how many bits from the hash are used in the page table index. This number must not exceed 28. HTABSIZE is used to generate a mask of the form 0b00...011...1; that is, a string of 0 bits followed by a string of 1-bits. The 1-bits determine how many additional bits (beyond the minimum of 11) from the hash are used in the index. The HTABORG must have this same number of low-order bits equal to 0.
Adjust alignment mask and hash size for ppc64.
Note that the page table does not yet scale with the memory size, and the secondary hash is not yet used.
v4:
- Don't redefine hash_page_{32,64} but add TODO comments for
renaming.
- Keep and simplify hash_page().
v3:
- Fix shifts. Thanks to Segher and Alex.
- Fix HTABSIZE for 32-bit ppc64: ((2 << 15) - 1) >> 16 happened
to be 0, so that it worked as expected. Avoid underflow, spotted by Alex. Make this fix available for legacy 64-bit support as well.
- Drop get_hash_size(). It was only used in hash_page_32() and
caused unnecessary conversions between HTABMASK and size.
- Introduce {mfsdr1,mtsdr1} as suggested by Alex.
Note that this changes the volatility for hash_page_64().
- Conditionalize 64-bit support, drop 32-bit support for ppc64.
- Adjust hash size for ppc64.
v2:
- Use HTABSIZE for ppc64, fix HTABMASK usage for ppc.
Error spotted by Segher.
Cc: Alexander Graf agraf@suse.de Cc: Segher Boessenkool segher@kernel.crashing.org Signed-off-by: Andreas Färber andreas.faerber@web.de
arch/ppc/qemu/ofmem.c | 69 +++++++++++++++++++++++++++ +------------- include/arch/ppc/processor.h | 12 +++++++ 2 files changed, 59 insertions(+), 22 deletions(-)
diff --git a/arch/ppc/qemu/ofmem.c b/arch/ppc/qemu/ofmem.c index 72694b3..6f3079a 100644 --- a/arch/ppc/qemu/ofmem.c +++ b/arch/ppc/qemu/ofmem.c @@ -48,7 +48,12 @@ extern void setup_mmu( unsigned long code_base ); #define OF_CODE_START 0xfff00000UL #define IO_BASE 0x80000000
-#define HASH_SIZE (2 << 15) +#ifdef __powerpc64__ +#define HASH_BITS 18 +#else +#define HASH_BITS 15 +#endif +#define HASH_SIZE (2 << HASH_BITS) #define OFMEM_SIZE (2 * 1024 * 1024)
#define SEGR_USER BIT(2) @@ -57,21 +62,11 @@ extern void setup_mmu( unsigned long code_base ); static inline unsigned long get_hash_base( void ) {
- unsigned long sdr1;
- asm volatile("mfsdr1 %0" : "=r" (sdr1) );
- return (sdr1 & 0xffff0000);
-}
-static inline unsigned long -get_hash_size( void ) -{
- unsigned long sdr1;
- asm volatile("mfsdr1 %0" : "=r" (sdr1) );
- return ((sdr1 << 16) | 0x0000ffff) + 1;
+#ifdef __powerpc64__
- return (mfsdr1() & 0x3FFFFFFFFFFC0000UL);
Just define these as costants and keep the actual code clean :).
OK.
+#else
- return (mfsdr1() & 0xffff0000UL);
+#endif }
static inline unsigned long @@ -228,6 +223,8 @@ ea_to_phys( ucell ea, ucell *mode ) return phys; }
+#if defined(__powerpc64__) || defined(CONFIG_PPC_64BITSUPPORT)
static void hash_page_64( ucell ea, ucell phys, ucell mode ) { @@ -242,7 +239,7 @@ hash_page_64( ucell ea, ucell phys, ucell mode ) vsid = (ea >> 28) + SEGR_BASE; vsid_sh = 7; vsid_mask = 0x00003FFFFFFFFF80ULL;
- asm ( "mfsdr1 %0" : "=r" (sdr) );
- sdr = mfsdr1(); sdr_sh = 18; sdr_mask = 0x3FF80; page_mask = 0x0FFFFFFF; // XXX correct?
@@ -294,6 +291,12 @@ hash_page_64( ucell ea, ucell phys, ucell mode ) asm volatile( "tlbie %0" :: "r"(ea) ); }
+#endif /* ppc64 */ +#ifndef __powerpc64__ +/* XXX Turn this into an #else when legacy 64-bit support is gone
and rename hash_page{32,64} to hash_page.Just keep them around. They shouldn't hurt. The code will be optimized away by the compiler anyways if CONFIG_PPC_64BITSUPPORT is not set, as static functions that never get caled (if on a constant 0 is never get called) just don't get compiled in.
They do hurt in that not calling them leads to an error.
Nope.
static void x(void) { }
This breaks:
#if 0 x(); #endif
This works:
if (0) x();
So if you keep the call to the 32 bit stuff in compiled code that gets optimized out because it can never be called, everybody is happy :).
I do and did understand that. Make that everybody excluding me.
- */
static void hash_page_32( ucell ea, ucell phys, ucell mode ) { @@ -307,7 +310,7 @@ hash_page_32( ucell ea, ucell phys, ucell mode )
hash1 = vsid; hash1 ^= (ea >> 12) & 0xffff;
- hash1 &= (get_hash_size() - 1) >> 6;
hash1 &= (((mfsdr1() & 0x1ff) << 16) | 0xffff) >> 6;
pp = (mPTE_t*)(get_hash_base() + (hash1 << 6)); upte = (unsigned long*)pp;
@@ -334,6 +337,9 @@ hash_page_32( ucell ea, ucell phys, ucell mode ) asm volatile( "tlbie %0" :: "r"(ea) ); }
+#endif /* ppc */ +#ifdef CONFIG_PPC_64BITSUPPORT
static int is_ppc64(void) { unsigned int pvr; @@ -342,12 +348,21 @@ static int is_ppc64(void) return ((pvr >= 0x330000) && (pvr < 0x70330000)); }
+#endif /* CONFIG_PPC_64BITSUPPORT */
static int is_ppc64(void) { #if defined(__powerpc64__) return 1; #elif defined(CONFIG_PPC_64BITSUPPORT) unsigned int pvr = mfpvr(); return ((pvr >= 0x330000) && (pvr < 0x70330000)); #else return 0; #endif }
+/* XXX Drop this function when legacy 64-bit support is dropped. */ static void hash_page( unsigned long ea, unsigned long phys, ucell mode ) { +#ifdef __powerpc64__
- hash_page_64(ea, phys, mode);
+#else +#ifdef CONFIG_PPC_64BITSUPPORT if ( is_ppc64() ) hash_page_64(ea, phys, mode); else +#endif hash_page_32(ea, phys, mode); +#endif
if (is_ppc64) hash_page_64(ea, phys, mode); else hash_page_32(ea, phys, mode);
}
void @@ -387,17 +402,27 @@ void setup_mmu( unsigned long ramsize ) { ofmem_t *ofmem;
- unsigned long sdr1, sr_base;
- unsigned long sr_base; unsigned long hash_base;
- unsigned long hash_mask = 0xfff00000; /* alignment for ppc64
*/
- unsigned long hash_mask = ~0x000fffffUL; /* alignment for
ppc64 */ int i;
/* SDR1: Storage Description Register 1 */
hash_base = (ramsize - 0x00100000 - HASH_SIZE) & hash_mask; memset((void *)hash_base, 0, HASH_SIZE);
- sdr1 = hash_base | ((HASH_SIZE-1) >> 16);
- asm volatile("mtsdr1 %0" :: "r" (sdr1) );
+#if defined(__powerpc64__) || defined(CONFIG_PPC_64BITSUPPORT) +#ifdef CONFIG_PPC_64BITSUPPORT
- if (is_ppc64())
+#endif
mtsdr1(hash_base | MAX(HASH_BITS - 18, 0));+#ifdef CONFIG_PPC_64BITSUPPORT
- else
+#endif +#endif +#ifndef __powerpc64__
mtsdr1(hash_base | ((HASH_SIZE - 1) >> 16));+#endif
if (is_ppc64()) mtsdr1(hash_base | MAX(HASH_BITS - 18, 0)); else mtsdr1(hash_base | (HASH_SIZE - 1) >> 16, 0));
That's a lot easier to read, no?
So is
#ifdef __powerpc64__ mtsdr1(...); #else mtsdr1(...); #endif
which I'm trying to reach.
In the SLB patch for comparison the final
#ifdef __powerpc64__ ... #else ... #endif
code is so much nicer than if (is_ppc64()) ... else ...
which requires a lot of reindentation to remain readable - reindentation that would later be reverted again when legacy support is dropped. Trying to avoid that.
For in total 10 lines of code?
Why are you trying so hard to get rid of legacy ppc64 compat? It doesn't hurt to have it around as long as you keep the #ifdefs to a reasonable minimum :). By then it's actually nice to have for code check reasons: compile ppc64 and be reasonably sure you don't break ppc32.
Nope. ppc code uses unsigned long and expects it to be 32 bits. All instances of such code will get compiled with sizes it was not intended for and it causes to loose time fiddling with code that I don't need to touch. This hit us before (e.g., get_hash_size) and will do so in the future. If code is designed for one bitness, don't force unnecessarily compiling it for another, even if the linker might optimize it out *after* successful compilation.
Compiling ppc64 code allows no definite conclusions about ppc.
Rebasing it now and will look for a solution.
Andreas