Re: [patch 14/22] pollfs: pollable futex

[Eric Dumazet <dada1@cosmosbay.com>]

Davi Arnaut a écrit :
> Eric Dumazet wrote:
>> Davi Arnaut a écrit :
>>> Asynchronously wait for FUTEX_WAKE operation on a futex if it still contains
>>> a given value. There can be only one futex wait per file descriptor. However,
>>> it can be rearmed (possibly at a different address) anytime.
>>>
>>> The pollable futex approach is far superior (send and receive events from
>>> userspace or kernel) to eventfd and fixes (supercedes) FUTEX_FD at the same time.
>>>
>>> Building block for pollable semaphores and user-defined events.
>>>
>>> Signed-off-by: Davi E. M. Arnaut <davi@haxent.com.br>
>>>
>>> ---
>>>  fs/pollfs/Makefile |    1 
>>>  fs/pollfs/futex.c  |  154 +++++++++++++++++++++++++++++++++++++++++++++++++++++
>>>  init/Kconfig       |    7 ++
>>>  3 files changed, 162 insertions(+)
>>>
>>> Index: linux-2.6/fs/pollfs/Makefile
>>> ===================================================================
>>> --- linux-2.6.orig/fs/pollfs/Makefile
>>> +++ linux-2.6/fs/pollfs/Makefile
>>> @@ -3,3 +3,4 @@ pollfs-y := file.o
>>>  
>>>  pollfs-$(CONFIG_POLLFS_SIGNAL) += signal.o
>>>  pollfs-$(CONFIG_POLLFS_TIMER) += timer.o
>>> +pollfs-$(CONFIG_POLLFS_FUTEX) += futex.o
>>> Index: linux-2.6/fs/pollfs/futex.c
>>> ===================================================================
>>> --- /dev/null
>>> +++ linux-2.6/fs/pollfs/futex.c
>>> @@ -0,0 +1,154 @@
>>> +/*
>>> + * pollable futex
>>> + *
>>> + * Copyright (C) 2007 Davi E. M. Arnaut
>>> + *
>>> + * Licensed under the GNU GPL. See the file COPYING for details.
>>> + */
>>> +
>>> +#include <linux/kernel.h>
>>> +#include <linux/sched.h>
>>> +#include <linux/module.h>
>>> +#include <linux/slab.h>
>>> +#include <linux/err.h>
>>> +#include <linux/wait.h>
>>> +#include <linux/poll.h>
>>> +#include <linux/pollfs_fs.h>
>>> +#include <linux/futex.h>
>>> +
>>> +struct futex_event {
>>> +	union {
>>> +		void __user *addr;
>>> +		u64 padding;
>>> +	};
>>> +	int val;
>>> +};
>> Hum... Here we might have a problem with 64 bit futexes, or private futexes
>>
>> So I believe this interface is not well defined and not expandable: in case of 
>> future additions to futexes, an old application compiled with an old pollable 
>> futex_event type might fail.
>>
> 
> Hmm, how about:
> 
> struct futex_event {
> 	union {
> 		void __user *addr;
> 		u64 padding;
> 	};
> 	union {
> 		int val;
> 		s64 val64;
> 	};
> 	/* whatever room is necessary for future improvements */
> };
> 
> I haven't been keeping up with 64 bit or private futexes. What else
> could probably go wrong?

Well, that's the point : This interface is like an ioctl() one : pretty bad if 
not properly designed :)

You probably need to stick one field containing one command or version number, 
something like that.


struct futex_event {
	int     type;
	union {
		void __user *addr;
		u64 padding;
	};
	union {
		int val;
		s64 val64;
	};
};

#define  FUTEX_EVENT_SHARED32  1
#define  FUTEX_EVENT_SHARED64  2
#define  FUTEX_EVENT_PRIVATE32 (128|1)
#define  FUTEX_EVENT_PRIVATE64 (128|2)

...

Also, you should take care of alignements constraints (a 32bit user program 
might run on a 64bit kernel)