Re: [PATCH] fs : reorder some 'struct inode' fields to speedup i_size manipulations

[Eric Dumazet <dada1@cosmosbay.com>]

Andrew Morton a écrit :
> On Thu, 23 Nov 2006 11:57:29 +0100
> Eric Dumazet <dada1@cosmosbay.com> wrote:
> 
>> On 32bits SMP platforms, 64bits i_size is protected by a seqcount 
>> (i_size_seqcount).
>>
>> When i_size is read or written, i_size_seqcount is read/written as well, so it 
>> make sense to group these two fields together in the same cache line.
>>
>> Before this patch, accessing i_size needed 3 cache lines (2 for i_size, one 
>> for i_size_seqcount). After, only one cache line is needed/ (dirtied on a 
>> i_size change).
> 
> I didn't understand that paragraph at all, really, so I took it out.
> 
> At present an i_size change will dirty one, two or three cachelines, most
> likely one or two.
> 
> After your patch an i_size change will dirty one or two cachelines, most
> likely one.
> 
> yes?

nope

Before :
---------
  offsetof(i_size) = 0x3C

i_size is 8 bytes, so i_size spans 2 cache lines (if 64 or 32 bytes cache lines)

and offsetof(i_size_seqcount) = 0x160, so a read of i_size (coupled with a 
read of seqcount) needed 3 cache lines. A change of i_size dirtied 2 or 3 
cache lines.


After :
--------
offsetof(i_size) = 0x40
offsetof(i_size_seqcount) = 0x48

One cache line 'only', reading or writing.