Re: [Qemu-devel] [PATCHv3 2/9] cutils: add a function to find non-zero content in a buffer

[Peter Lieven]

Am 21.03.2013 um 19:12 schrieb Eric Blake <address@hidden>:

> On 03/21/2013 09:57 AM, Peter Lieven wrote:
>> this adds buffer_find_nonzero_offset() which is a SSE2/Altives
> 
> s/Altives/Altivec/
> 
>> optimized function that searches for non-zero content in a
>> buffer.
>> 
>> due to the optimizations used in the function there are restrictions
>> on buffer address and search length. the function
>> can_use_buffer_find_nonzero_content() can be used to check if
>> the function can be used safely.
>> 
>> Signed-off-by: Peter Lieven <address@hidden>
>> ---
>> include/qemu-common.h |    3 +++
>> util/cutils.c         |   50 
>> +++++++++++++++++++++++++++++++++++++++++++++++++
>> 2 files changed, 53 insertions(+)
> 
>> +inline bool can_use_buffer_find_nonzero_offset(const void *buf, size_t len);
>> +inline size_t buffer_find_nonzero_offset(const void *buf, size_t len);
> 
> Ouch.  It is okay to add a 'static inline' function, but then the
> implementation must live in this header.  Otherwise, the function must
> not be inline, or you risk linker errors.

Sorry, i was not aware. 

> 
>> +++ b/util/cutils.c
>> @@ -143,6 +143,56 @@ int qemu_fdatasync(int fd)
>> }
>> 
>> /*
>> + * Searches for an area with non-zero content in a buffer
>> + *
>> + * Attention! The len must be a multiple of 8 * sizeof(VECTYPE) 
> 
> Should we call out BUFFER_FIND_NONZERO_OFFSET_UNROLL_FACTOR instead of a
> magic number here?  But I'm okay with leaving it as-is.
> 
>> + * and addr must be a multiple of sizeof(VECTYPE) due to 
> 
> Trailing whitespace (here, and on several other lines).  Please run your
> series through scripts/checkpatch.pl before submitting v4.

thanks for the pointer.

> 
>> + * restriction of optimizations in this function.
>> + * 
>> + * can_use_buffer_find_nonzero_offset() can be used to check
>> + * these requirements.
>> + * 
>> + * The return value is the offset of the non-zero area rounded
>> + * down to 8 * sizeof(VECTYPE). If the buffer is all zero 
> 
> Same comment on this use of '8'.
> 
>> + * the return value is equal to len.
>> + */
>> +
>> +inline size_t buffer_find_nonzero_offset(const void *buf, size_t len)
> 
> s/inline// (or move it to a 'static inline' definition in the .h)

I would move at least the can_use_buffer_find_nonzero_offset() function
completely into the header and rely the compiler inlineing
buffer_find_nonzero_offset().

> 
>> +{
>> +    VECTYPE *p = (VECTYPE *)buf;
>> +    VECTYPE zero = ZERO_SPLAT;
>> +    size_t i;
>> +    
> 
> You copied the 'Attention! ...' message from buffer_is_zero, which
> currently asserts that its condition is held.  Therefore, consistency
> would argue that you should assert your preconditions here, even if it
> adds more to the code size.  But this is something where a maintainer
> might have a better opinion on whether to keep the code robust with an
> assert(), or whether the faster operation without sanity checking is
> more appropriate (in which case a followup to remove the assert from
> buffer_is_zero would make sense).

I would appreciate feedback on this from the maintainers. My concern
with buffer_find_nonzero_offset() is that it is used for checking millions
of 4KByte pages. buffer_is_zero is used for larger objects in the bdrv
context, afaik.



> 
>>  * Checks if a buffer is all zeroes
>>  *
>>  * Attention! The len must be a multiple of 4 * sizeof(long) due to
>> 
> 
> Cleaning up whitespace is trivial; but the incorrect use of 'inline'
> requires a v4.
> 
> -- 
> Eric Blake   eblake redhat com    +1-919-301-3266
> Libvirt virtualization library http://libvirt.org
>