s390: fadvise hint values.
PATCH 1/1 s390: fadvise hint values.
Hi Andrew,
an ugly one. The fadvise hint values for POSIX_FADV_DONTNEED and POSIX_FADV_NOREUSE in the kernel and the glibc differ for s390-64 (and worse the values for s390-31 and s390-64 differ as well ..). The glibc always had 6 and 7 instead of 4 and 5 for these two values for s390-64. My first reaction was to correct the values in the glibc headers but as Ulrich Drepper pointed out that has some unwanted consequences:
1) the applications build against the wrong values will get -EINVAL and the advice gets ignored, and
2) if the values 6 and 7 are ever used for some new advice then these applications might show erratic behaviour.
I can't say which and how many applications use fadvise so it might be a better idea to fix this in the kernel.
Patch is attached, what do you think ?
blue skies,
Martin.
---
[patch] s390: fadvise hint values.
From: Martin Schwidefsky <[email protected]>
Add special case for the POSIX_FADV_DONTNEED and POSIX_FADV_NOREUSE hint values for s390-64. The user space values in the s390-64 glibc headers for these two defines have always been 6 and 7 instead of 4 and 5. All 64 bit applications therefore use the "wrong" values. To get these applications working without recompiling the kernel needs to accept the "wrong" values. Since the values for s390-31 are 4 and 5 the compat wrapper for fadvise64 and fadvise64_64 need to rewrite the values for 31 bit system calls.
diffstat:
arch/s390/kernel/compat_linux.c | 38 ++++++++++++++++++++++++++++++++++++++ arch/s390/kernel/compat_wrapper.S | 4 ++-- include/linux/fadvise.h | 10 ++++++++++ 3 files changed, 50 insertions(+), 2 deletions(-)
arch/s390/kernel/compat_linux.c +38 -0
@@ -58,6 +58,7 @@#include <linux/compat.h> #include <linux/vfs.h> #include <linux/ptrace.h> +#include <linux/fadvise.h> #include <asm/types.h> #include <asm/ipc.h>
@@ -1043,3 +1044,40 @@ sys32_timer_create(clockid_t which_clockreturn ret; } + +/* +* 31 bit emulation wrapper functions for sys_fadvise64/fadvise64_64. +* These need to rewrite the advise values for POSIX_FADV_{DONTNEED,NOREUSE} +* because the 31 bit values differ from the 64 bit values. +*/ + +asmlinkage long +sys32_fadvise64 (int fd,loff_t offset,size_t len,int advise) +{ +if (advise ==4 ) + advise =POSIX_FADV_DONTNEED ; +else if (advise ==5 ) + advise =POSIX_FADV_NOREUSE ; +return sys_fadvise64 (fd, offset, len, advise); +} + +struct fadvise64_64_args { +int fd; +long long offset; +long long len; +int advice; +}; + +asmlinkage long +sys32_fadvise64_64 (struct fadvise64_64_args __user *args) +{ +struct fadvise64_64_args a; + +if (copy_from_user (&a, args,sizeof (a)) ) +return -EFAULT ; +if (a.advice ==4 ) + a.advice =POSIX_FADV_DONTNEED ; +else if (a.advice ==5 ) + a.advice =POSIX_FADV_NOREUSE ; +return sys_fadvise64_64 (a.fd, a.offset, a.len, a.advice); +}
arch/s390/kernel/compat_wrapper.S +2 -2
@@ -1251,12 +1251,12 @@ sys32_fadvise64_wrapper: or %r3,%r4 # get low word of 64bit loff_t llgfr %r4,%r5 # size_t (unsigned long) lgfr %r5,%r6 # int - jg sys_fadvise64 + jg sys32_fadvise64 .globl sys32_fadvise64_64_wrapper sys32_fadvise64_64_wrapper: llgtr %r2,%r2 # struct fadvise64_64_args * - jg s390_fadvise64_64 + jg sys32_fadvise64_64 .globl sys32_clock_settime_wrapper sys32_clock_settime_wrapper:
include/linux/fadvise.h +10 -0
@@ -5,7 +5,17 @@#define POSIX_FADV_RANDOM 1/* Expect random page references. */ #define POSIX_FADV_SEQUENTIAL 2/* Expect sequential page references. */ #define POSIX_FADV_WILLNEED 3/* Will need these pages. */ + +/* +* The advise values for POSIX_FADV_DONTNEED and POSIX_ADV_NOREUSE +* for s390-64 differ from the values for the rest of the world. +*/ +#if defined(__s390x__) +#define POSIX_FADV_DONTNEED 6/* Don't need these pages. */ +#define POSIX_FADV_NOREUSE 7/* Data will be accessed once. */ +#else #define POSIX_FADV_DONTNEED 4/* Don't need these pages. */ #define POSIX_FADV_NOREUSE 5/* Data will be accessed once. */ +#endif #endif /* FADVISE_H_INCLUDED */
Quoted 15 lines · Martin Schwidefsky
Hi Andrew,
an ugly one. The fadvise hint values forPOSIX_FADV_DONTNEEDandPOSIX_FADV_NOREUSEin the kernel and the glibc differ for s390-64 (and worse the values for s390-31 and s390-64 differ as well ..). The glibc always had 6 and 7 instead of 4 and 5 for these two values for s390-64. My first reaction was to correct the values in the glibc headers but as Ulrich Drepper pointed out that has some unwanted consequences:
1) the applications build against the wrong values will get-EINVALand the advice gets ignored, and
2) if the values 6 and 7 are ever used for some new advice then these applications might show erratic behaviour.
I can't say which and how many applications use fadvise so it might be a better idea to fix this in the kernel.
Quoted 1 line
Patch is attached, what do you think ?
I'd rather fix the kernel and do some symbol versioning magic in glibc. After all it's their stupidity that caused all these problems.
Quoted 4 lines
Quoted 1 line
Patch is attached, what do you think ?
I'd rather fix the kernel and do some symbol versioning magic in glibc. After all it's their stupidity that caused all these problems.
Well I was my stupidity that I copied the wrong fcntl.h file for s390-64. The fcntl.h glibc header for alpha contained the "wrong" values and since day 1 of the s390-64 port it has been 6 and 7 for us as well.
If we decide to fix the kernel we don't need to do anything in the glibc.
blue skies,
Martin
Martin Schwidefsky
Linux for zSeries Development & Services
IBM Deutschland Entwicklung GmbH