Re: [PATCH 0/2] eventfd: new EFD_STATE flag

["Michael S. Tsirkin" <mst@redhat.com>]

On Thu, Aug 27, 2009 at 11:05:30AM +0200, Paolo Bonzini wrote:
> > Ok, so why not using the eventfd counter as state?
> > On the device side:
> > 
> > void write_state(int sfd, int state) {
> > 	u64 cnt;
> > 
> > 	/* Clear the current state, sfd is in non-blocking mode */
> > 	read(sfd,&cnt, sizeof(cnt));
> > 	/* Writes new state */
> > 	cnt = 1 + !!state;
> > 	write(sfd,&cnt, sizeof(cnt));
> > }
> 
> It's interesting [no sarcasm intended, mind] that EFD_SEMAPHORE was
> added exactly to avoid a read+write combination for the case of
> decrementing a value.  Here it's the same, just it's about the case of
> writing a *given* value.  What about having EFD_STATE simply mean "do
> not use a counter, just write the value" without affecting the way
> read works, and use
> 
> 	/* Writes new state */
> 	cnt = 1 + !!state;
> 	write(sfd,&cnt, sizeof(cnt));

That would work for kvm.

> See below?
> 
> Paolo
> 
> > On the hypervisor side:
> > 
> > int read_state(int sfd) {
> > 	u64 cnt;
> > 
> > 	read(sfd,&cnt, sizeof(cnt));
> > 	return state - 1;
> > }
> 
> 
> ------------- 8<-- ---------------
> Subject: [PATCH] eventfd: new EFD_ABSOLUTE flag
> 
> This implements a new EFD_ABSOLUTE flag for eventfd.
> This changes eventfd behaviour so that write simply
> stores the value written, and is always non-blocking.
> 
> Untested (I just modified Michael's patch, and given
> the simpler code I'm not sure it's now worthwhile
> introducing the inline functions), but otherwise
> 
> Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
> ---
>  fs/eventfd.c            |   13 ++++++++-----
>  include/linux/eventfd.h |    4 ++--
>  2 files changed, 10 insertions(+), 7 deletions(-)
> 
> diff --git a/fs/eventfd.c b/fs/eventfd.c
> index 347a0e0..7b279e3 100644
> --- a/fs/eventfd.c
> +++ b/fs/eventfd.c
> @@ -31,8 +31,6 @@
>  
>  static inline int eventfd_writeable(struct eventfd_ctx *ctx, u64 n)
>  {
> -	return ULLONG_MAX - n > ctx->count;
> -	return (ctx->flags & EFD_ABSOLUTE) || (ULLONG_MAX - n > ctx->count);
>  }
>  
>  static inline int eventfd_overflow(struct eventfd_ctx *ctx, u64 cnt)
> @@ -42,10 +40,14 @@
>  
>  static inline void eventfd_dowrite(struct eventfd_ctx *ctx, u64 ucnt)
>  {
> -	if (eventfd_writeable(ctx, ucnt))
> -		ucnt = ULLONG_MAX - ctx->count;
> +	if (ctx->flags & EFD_ABSOLUTE)
> +		ctx->count = ucnt;
> +	else {
> +		if (ULLONG_MAX - ctx->count < ucnt)
> +			ucnt = ULLONG_MAX - ctx->count;
>  
> -	ctx->count += ucnt;
> +		ctx->count += ucnt;
> +	}
>  }
>  
>  static inline u64 eventfd_doread(struct eventfd_ctx *ctx)
> diff --git a/include/linux/eventfd.h b/include/linux/eventfd.h
> index 3b85ba6..78ff649 100644
> --- a/include/linux/eventfd.h
> +++ b/include/linux/eventfd.h
> @@ -19,11 +19,12 @@
>   * shared O_* flags.
>   */
>  #define EFD_SEMAPHORE (1 << 0)
> +#define EFD_ABSOLUTE (1 << 1)
>  #define EFD_CLOEXEC O_CLOEXEC
>  #define EFD_NONBLOCK O_NONBLOCK
>  
>  #define EFD_SHARED_FCNTL_FLAGS (O_CLOEXEC | O_NONBLOCK)
> -#define EFD_FLAGS_SET (EFD_SHARED_FCNTL_FLAGS | EFD_SEMAPHORE)
> +#define EFD_FLAGS_SET (EFD_SHARED_FCNTL_FLAGS | EFD_SEMAPHORE | EFD_ABSOLUTE)
>  
>  #ifdef CONFIG_EVENTFD
> --
> 1.6.2.5