diff options
author | Martin Schwidefsky <schwidefsky@de.ibm.com> | 2006-09-28 16:55:39 +0200 |
---|---|---|
committer | Martin Schwidefsky <schwidefsky@de.ibm.com> | 2006-09-28 16:55:39 +0200 |
commit | d9f7a745d55527d0d41684b22506a86c4381f7f1 (patch) | |
tree | ea8870ef06c3723ad59b78aac97bfe8152894c72 /arch/s390/lib | |
parent | 1fce518e8e7de62597c823d6d795cafc694e7910 (diff) | |
download | kernel_samsung_smdk4412-d9f7a745d55527d0d41684b22506a86c4381f7f1.zip kernel_samsung_smdk4412-d9f7a745d55527d0d41684b22506a86c4381f7f1.tar.gz kernel_samsung_smdk4412-d9f7a745d55527d0d41684b22506a86c4381f7f1.tar.bz2 |
[S390] __div64_32 for 31 bit.
The clocksource infrastructure introduced with commit
ad596171ed635c51a9eef829187af100cbf8dcf7 broke 31 bit s390.
The reason is that the do_div() primitive for 31 bit always
had a restriction: it could only divide an unsigned 64 bit
integer by an unsigned 31 bit integer. The clocksource code
now uses do_div() with a base value that has the most
significant bit set. The result is that clock->cycle_interval
has a funny value which causes the linux time to jump around
like mad.
The solution is "obvious": implement a proper __div64_32
function for 31 bit s390.
Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com>
Diffstat (limited to 'arch/s390/lib')
-rw-r--r-- | arch/s390/lib/Makefile | 1 | ||||
-rw-r--r-- | arch/s390/lib/div64.c | 151 |
2 files changed, 152 insertions, 0 deletions
diff --git a/arch/s390/lib/Makefile b/arch/s390/lib/Makefile index c42ffed..b0cfa6c 100644 --- a/arch/s390/lib/Makefile +++ b/arch/s390/lib/Makefile @@ -5,5 +5,6 @@ EXTRA_AFLAGS := -traditional lib-y += delay.o string.o uaccess_std.o +lib-$(CONFIG_32BIT) += div64.o lib-$(CONFIG_64BIT) += uaccess_mvcos.o lib-$(CONFIG_SMP) += spinlock.o diff --git a/arch/s390/lib/div64.c b/arch/s390/lib/div64.c new file mode 100644 index 0000000..0481f34 --- /dev/null +++ b/arch/s390/lib/div64.c @@ -0,0 +1,151 @@ +/* + * arch/s390/lib/div64.c + * + * __div64_32 implementation for 31 bit. + * + * Copyright (C) IBM Corp. 2006 + * Author(s): Martin Schwidefsky (schwidefsky@de.ibm.com), + */ + +#include <linux/types.h> +#include <linux/module.h> + +#ifdef CONFIG_MARCH_G5 + +/* + * Function to divide an unsigned 64 bit integer by an unsigned + * 31 bit integer using signed 64/32 bit division. + */ +static uint32_t __div64_31(uint64_t *n, uint32_t base) +{ + register uint32_t reg2 asm("2"); + register uint32_t reg3 asm("3"); + uint32_t *words = (uint32_t *) n; + uint32_t tmp; + + /* Special case base==1, remainder = 0, quotient = n */ + if (base == 1) + return 0; + /* + * Special case base==0 will cause a fixed point divide exception + * on the dr instruction and may not happen anyway. For the + * following calculation we can assume base > 1. The first + * signed 64 / 32 bit division with an upper half of 0 will + * give the correct upper half of the 64 bit quotient. + */ + reg2 = 0UL; + reg3 = words[0]; + asm volatile( + " dr %0,%2\n" + : "+d" (reg2), "+d" (reg3) : "d" (base) : "cc" ); + words[0] = reg3; + reg3 = words[1]; + /* + * To get the lower half of the 64 bit quotient and the 32 bit + * remainder we have to use a little trick. Since we only have + * a signed division the quotient can get too big. To avoid this + * the 64 bit dividend is halved, then the signed division will + * work. Afterwards the quotient and the remainder are doubled. + * If the last bit of the dividend has been one the remainder + * is increased by one then checked against the base. If the + * remainder has overflown subtract base and increase the + * quotient. Simple, no ? + */ + asm volatile( + " nr %2,%1\n" + " srdl %0,1\n" + " dr %0,%3\n" + " alr %0,%0\n" + " alr %1,%1\n" + " alr %0,%2\n" + " clr %0,%3\n" + " jl 0f\n" + " slr %0,%3\n" + " alr %1,%2\n" + "0:\n" + : "+d" (reg2), "+d" (reg3), "=d" (tmp) + : "d" (base), "2" (1UL) : "cc" ); + words[1] = reg3; + return reg2; +} + +/* + * Function to divide an unsigned 64 bit integer by an unsigned + * 32 bit integer using the unsigned 64/31 bit division. + */ +uint32_t __div64_32(uint64_t *n, uint32_t base) +{ + uint32_t r; + + /* + * If the most significant bit of base is set, divide n by + * (base/2). That allows to use 64/31 bit division and gives a + * good approximation of the result: n = (base/2)*q + r. The + * result needs to be corrected with two simple transformations. + * If base is already < 2^31-1 __div64_31 can be used directly. + */ + r = __div64_31(n, ((signed) base < 0) ? (base/2) : base); + if ((signed) base < 0) { + uint64_t q = *n; + /* + * First transformation: + * n = (base/2)*q + r + * = ((base/2)*2)*(q/2) + ((q&1) ? (base/2) : 0) + r + * Since r < (base/2), r + (base/2) < base. + * With q1 = (q/2) and r1 = r + ((q&1) ? (base/2) : 0) + * n = ((base/2)*2)*q1 + r1 with r1 < base. + */ + if (q & 1) + r += base/2; + q >>= 1; + /* + * Second transformation. ((base/2)*2) could have lost the + * last bit. + * n = ((base/2)*2)*q1 + r1 + * = base*q1 - ((base&1) ? q1 : 0) + r1 + */ + if (base & 1) { + int64_t rx = r - q; + /* + * base is >= 2^31. The worst case for the while + * loop is n=2^64-1 base=2^31+1. That gives a + * maximum for q=(2^64-1)/2^31 = 0x1ffffffff. Since + * base >= 2^31 the loop is finished after a maximum + * of three iterations. + */ + while (rx < 0) { + rx += base; + q--; + } + r = rx; + } + *n = q; + } + return r; +} + +#else /* MARCH_G5 */ + +uint32_t __div64_32(uint64_t *n, uint32_t base) +{ + register uint32_t reg2 asm("2"); + register uint32_t reg3 asm("3"); + uint32_t *words = (uint32_t *) n; + + reg2 = 0UL; + reg3 = words[0]; + asm volatile( + " dlr %0,%2\n" + : "+d" (reg2), "+d" (reg3) : "d" (base) : "cc" ); + words[0] = reg3; + reg3 = words[1]; + asm volatile( + " dlr %0,%2\n" + : "+d" (reg2), "+d" (reg3) : "d" (base) : "cc" ); + words[1] = reg3; + return reg2; +} + +#endif /* MARCH_G5 */ + +EXPORT_SYMBOL(__div64_32); |