Re: [PATCH v9 02/17] drivers/net/b44: Align pwol_mask to unsigned long for better performance

[Paolo Bonzini]

On 24/06/19 17:12, David Laight wrote:
> From: Fenghua Yu
>> Sent: 18 June 2019 23:41
>> From: Peter Zijlstra <peterz@xxxxxxxxxxxxx>
>>
>> A bit in pwol_mask is set in b44_magic_pattern() by atomic set_bit().
>> But since pwol_mask is local and never exposed to concurrency, there is
>> no need to set bit in pwol_mask atomically.
>>
>> set_bit() sets the bit in a single unsigned long location. Because
>> pwol_mask may not be aligned to unsigned long, the location may cross two
>> cache lines. On x86, accessing two cache lines in locked instruction in
>> set_bit() is called split locked access and can cause overall performance
>> degradation.
>>
>> So use non atomic __set_bit() to set pwol_mask bits. __set_bit() won't hit
>> split lock issue on x86.
>>
>> Signed-off-by: Peter Zijlstra <peterz@xxxxxxxxxxxxx>
>> Signed-off-by: Fenghua Yu <fenghua.yu@xxxxxxxxx>
>> ---
>>  drivers/net/ethernet/broadcom/b44.c | 4 ++--
>>  1 file changed, 2 insertions(+), 2 deletions(-)
>>
>> diff --git a/drivers/net/ethernet/broadcom/b44.c b/drivers/net/ethernet/broadcom/b44.c
>> index 97ab0dd25552..5738ab963dfb 100644
>> --- a/drivers/net/ethernet/broadcom/b44.c
>> +++ b/drivers/net/ethernet/broadcom/b44.c
>> @@ -1520,7 +1520,7 @@ static int b44_magic_pattern(u8 *macaddr, u8 *ppattern, u8 *pmask, int offset)
>>
>>  	memset(ppattern + offset, 0xff, magicsync);
>>  	for (j = 0; j < magicsync; j++)
>> -		set_bit(len++, (unsigned long *) pmask);
>> +		__set_bit(len++, (unsigned long *)pmask);
>>
>>  	for (j = 0; j < B44_MAX_PATTERNS; j++) {
>>  		if ((B44_PATTERN_SIZE - len) >= ETH_ALEN)
>> @@ -1532,7 +1532,7 @@ static int b44_magic_pattern(u8 *macaddr, u8 *ppattern, u8 *pmask, int offset)
>>  		for (k = 0; k< ethaddr_bytes; k++) {
>>  			ppattern[offset + magicsync +
>>  				(j * ETH_ALEN) + k] = macaddr[k];
>> -			set_bit(len++, (unsigned long *) pmask);
>> +			__set_bit(len++, (unsigned long *)pmask);
> 
> Is this code expected to do anything sensible on BE systems?

Probably not, but it's not wrong in different ways before/after the patch.

Paolo

> Casting the bitmask[] argument to any of the set_bit() functions is dubious at best.