Re: [PATCH] fs: inode per-cpu last_ino allocator

[Nick Piggin <npiggin@kernel.dk>]

On Thu, Sep 30, 2010 at 07:28:05PM +0200, Eric Dumazet wrote:
> Le jeudi 30 septembre 2010 à 09:45 -0700, Andrew Morton a écrit :
> 
> > Could eliminate `p' I guess, but that would involve using
> > __get_cpu_var() as an lval, which looks vile and might generate worse
> > code.
> > 
> 
> Hmm, I see, please check this new patch, using the most modern stuff ;)
> 
> > Readers of this code won't know why last_ino_get() was marked noinline.
> > It looks wrong, really.
> 
> Oops sorry, this was a temporary hack of mine to ease disassembly
> analysis. Good catch !
> 
> Here is the new generated code on i686 (with the noinline) : 
> pretty good ;)
> 
> c02e5930 <last_ino_get>:
> c02e5930:	55                   	push   %ebp
> c02e5931:	89 e5                	mov    %esp,%ebp
> c02e5933:	64 a1 44 29 7d c0    	mov    %fs:0xc07d2944,%eax
> c02e5939:	a9 ff 03 00 00       	test   $0x3ff,%eax
> c02e593e:	74 09                	je     c02e5949 <last_ino_get+0x19>
> c02e5940:	40                   	inc    %eax
> c02e5941:	64 a3 44 29 7d c0    	mov    %eax,%fs:0xc07d2944
> c02e5947:	c9                   	leave  
> c02e5948:	c3                   	ret    
> c02e5949:	b8 00 04 00 00       	mov    $0x400,%eax
> c02e594e:	f0 0f c1 05 80 c8 92 c0	lock xadd %eax,0xc092c880
> c02e5956:	eb e8                	jmp    c02e5940 <last_ino_get+0x10>
> 
> 
> Thanks
> 
> [PATCH] fs: inode per-cpu last_ino allocator

Thanks Eric, this looks good. You didn't seem to add a comment about
preempt safety that Andrew wanted, but I'll add it.

> 
> new_inode() dirties a contended cache line to get increasing
> inode numbers.
> 
> Solve this problem by providing to each cpu a per_cpu variable,
> feeded by the shared last_ino, but once every 1024 allocations.
> This reduces contention on the shared last_ino, and give same
> spreading ino numbers than before (i.e. same wraparound after 2^32
> allocations).
> 
> Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
> Signed-off-by: Nick Piggin <npiggin@suse.de>
> Signed-off-by: Dave Chinner <dchinner@redhat.com>
> ---
>  fs/inode.c |   47 ++++++++++++++++++++++++++++++++++++++++-------
>  1 file changed, 40 insertions(+), 7 deletions(-)
> 
> diff --git a/fs/inode.c b/fs/inode.c
> index 8646433..5c233f0 100644
> --- a/fs/inode.c
> +++ b/fs/inode.c
> @@ -624,6 +624,45 @@ void inode_add_to_lists(struct super_block *sb, struct inode *inode)
>  }
>  EXPORT_SYMBOL_GPL(inode_add_to_lists);
>  
> +#define LAST_INO_BATCH 1024
> +
> +/*
> + * Each cpu owns a range of LAST_INO_BATCH numbers.
> + * 'shared_last_ino' is dirtied only once out of LAST_INO_BATCH allocations,
> + * to renew the exhausted range.
> + *
> + * This does not significantly increase overflow rate because every CPU can
> + * consume at most LAST_INO_BATCH-1 unused inode numbers. So there is
> + * NR_CPUS*(LAST_INO_BATCH-1) wastage. At 4096 and 1024, this is ~0.1% of the
> + * 2^32 range, and is a worst-case. Even a 50% wastage would only increase
> + * overflow rate by 2x, which does not seem too significant.
> + *
> + * On a 32bit, non LFS stat() call, glibc will generate an EOVERFLOW
> + * error if st_ino won't fit in target struct field. Use 32bit counter
> + * here to attempt to avoid that.
> + */
> +static DEFINE_PER_CPU(unsigned int, last_ino);
> +
> +static unsigned int last_ino_get(void)
> +{
> +	unsigned int res;
> +
> +	get_cpu();
> +	res = __this_cpu_read(last_ino);
> +#ifdef CONFIG_SMP
> +	if (unlikely((res & (LAST_INO_BATCH - 1)) == 0)) {
> +		static atomic_t shared_last_ino;
> +		int next = atomic_add_return(LAST_INO_BATCH, &shared_last_ino);
> +
> +		res = next - LAST_INO_BATCH;
> +	}
> +#endif
> +	res++;
> +	__this_cpu_write(last_ino, res);
> +	put_cpu();
> +	return res;
> +}
> +
>  /**
>   *	new_inode 	- obtain an inode
>   *	@sb: superblock
> @@ -638,12 +677,6 @@ EXPORT_SYMBOL_GPL(inode_add_to_lists);
>   */
>  struct inode *new_inode(struct super_block *sb)
>  {
> -	/*
> -	 * On a 32bit, non LFS stat() call, glibc will generate an EOVERFLOW
> -	 * error if st_ino won't fit in target struct field. Use 32bit counter
> -	 * here to attempt to avoid that.
> -	 */
> -	static unsigned int last_ino;
>  	struct inode *inode;
>  
>  	spin_lock_prefetch(&inode_lock);
> @@ -652,7 +685,7 @@ struct inode *new_inode(struct super_block *sb)
>  	if (inode) {
>  		spin_lock(&inode_lock);
>  		__inode_add_to_lists(sb, NULL, inode);
> -		inode->i_ino = ++last_ino;
> +		inode->i_ino = last_ino_get();
>  		inode->i_state = 0;
>  		spin_unlock(&inode_lock);
>  	}
> 
> 
> --
> To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html