On 12/24/2015 11:42 PM, Arnd Bergmann wrote:
On Thursday 24 December 2015, Bamvor Zhang Jian wrote:
On 12/05/2015 12:44 AM, Arnd Bergmann wrote:
On Friday 04 December 2015 23:11:58 Bamvor Jian Zhang wrote:
diff --git a/drivers/char/lp.c b/drivers/char/lp.c index c4094c4..a207e0c 100644 --- a/drivers/char/lp.c +++ b/drivers/char/lp.c @@ -136,6 +136,14 @@ #include <asm/irq.h> #include <asm/uaccess.h>
- FIXME: It should be removed after COMPAT_USE_64BIT_TIME is accessible for
- 32bit architecture.
+#ifndef COMPAT_USE_64BIT_TIME +#define COMPAT_USE_64BIT_TIME (0) +#endif /* COMPAT_USE_64BIT_TIME */
Unfortunately, I think this can't work: the kernel has no idea whether user space uses a 32-bit or 64-bit time_t, and we need to be able to support both kinds of user space at the same time.
What we have to do instead is change the definition of LPSETTIMEOUT, like
#if __BITS_PER_SIZE_T >= __BITS_PER_LONG #define LPSETTIMEOUT 0x060f /* set parport timeout */ #else #define LPSETTIMEOUT _IOW(0x06, 0x0f, struct timeval) #endif
We need to decide what the correct #if condition should be, but the idea is that all current user space remains unchanged, while user space built with a 64-bit time_t sees the modified definition of LPSETTIMEOUT and passes that into the kernel when using the larger structure.
I am trying to follow your suggestion. But I do not understand why `__BITS_PER_SIZE_T` is relative to sizeof(time_t). And I also tried to use `__BITS_PER_TIME_T` according to the following table.
Sorry for the confusion, I meant to write __BITS_PER_TIME_T, not __BITS_PER_SIZE_T.
- userspace | kernel | `__BITS_PER_TIME_T` | `__BITS_PER_LONG` |
`__BITS_PER_TIME_T` <= `__BITS_PER_LONG`
32 | 32 | 32 | 32 | true
32 | 32 | 64 | 32 | false
32 | 64 | 32 | 32 | true
32 | 64 | 64 | 32 | false
64 | 64 | 64 | 64 | true
*/ #if __BITS_PER_TIME_T <= __BITS_PER_LONG #define LPSETTIMEOUT 0x060f /* set parport timeout */ #else #define LPSETTIMEOUT _IOW(0x06, 0x0f, struct timeval) #endif
As you know, there is no `__BITS_PER_TIME_T` as well. Finally, I thought I could make use of `__USE_TIME_BITS64` which mentioned in https://sourceware.org/glibc/wiki/Y2038ProofnessDesign
How about something like this?
#if defined(__USE_TIME_BITS64) && __BITS_PER_LONG == 32 #define LPSETTIMEOUT _IOW(0x06, 0x0f, struct timeval) #else #define LPSETTIMEOUT 0x060f /* set parport timeout */ #endif
Yes, this is what I meant. Unfortunately, we have not agreed on what we are going to call that macro, but this is roughly how it should work.
Got you, could I use __USE_TIME_BITS64 in my patch right now? Or should I need to wait for this agreement?
However, the definition you have here is only correct for user space, not for the kernel itself, which needs a slightly different definition when __KERNEL__ is defined and 'struct timeval' has been removed from the kernel.
Yeap, I am thinking it is header for uapi. I will define LPSETTIMEOUT64 in driver/char/lp.c: #define LPSETTIMEOUT64 _IOW(0x06, 0x0f, s64)
Is it make sense?