Re: CACHE_LINE_SIZE too small, so struct vpglocks size alignment doesn't work

From: Svatopluk Kraus <onwahe_at_gmail.com>
Date: Tue, 5 Oct 2010 14:35:38 +0200
On Fri, Oct 1, 2010 at 4:01 PM, Matthew Fleming <mdf356_at_gmail.com> wrote:
> On Fri, Oct 1, 2010 at 1:00 PM, Svatopluk Kraus <onwahe_at_gmail.com> wrote:
>> Hallo,
>>
>>  a size of 'struct vpglocks' is padded to CACHE_LINE_SIZE size in
>> 'sys/vm/vm_page.h'
>> header file. I work on a 'coldfire' port where CACHE_LINE_SIZE is 16 bytes and
>> sizeof(struct mtx) is 20 bytes thus size alignment doesn't work.
>>
>>  I solved it somehow, but I like to learn how to solve it in spirit
>> of FreeBSD.
>> There are a couple of possibilities:
>>
>> A1. Do nothing for small CACHE_LINE_SIZE.
>> A2. Pad to multiple of CACHE_LINE_SIZE.
>>
>> B1. use #if with CACHE_LINE_SIZE
>> B2. use #if with __coldfire__
>>
>> When I use B1 solution I need to known sizeof(struct mtx) value in
>> preprocessing time.
>> So, is it correct to use something like 'assym.s' magic
>> (sys/i386/i386/genassym.c)
>> in MI code? Or has someone another suggestion?
>
> What about padding to CACHE_LINE_SIZE - (sizeof(struct vpglocks) &
> (CACHE_LINE_SIZE-1)) ?
>
> Some compilers will complain about 0-sized arrays, but gcc isn't one of them.
>
> Cheers,
> matthew

Thanks for your nice suggestion. Probably, CACHE_LINE_SIZE size is a
multiple of 2 at all times, so your solution is applicable.

BTW, I use gcc 4.5.1 and when a structure is aligned to a value, then
a sizeof of the structure is padded to the value too implicitly. So no
explicit padding is needed.

Cheers, Svata
Received on Tue Oct 05 2010 - 10:35:39 UTC

This archive was generated by hypermail 2.4.0 : Wed May 19 2021 - 11:40:08 UTC