Re: [PATCH] virtio-blk: Convert QEMUBH callback to "bitops.h" API

["Philippe Mathieu-Daudé" <philmd@redhat.com>]

On 5/6/21 8:22 PM, Richard Henderson wrote:
> On 5/6/21 8:54 AM, Philippe Mathieu-Daudé wrote:
>>   static void notify_guest_bh(void *opaque)
>>   {
>>       VirtIOBlockDataPlane *s = opaque;
>> -    unsigned nvqs = s->conf->num_queues;
>> -    unsigned long bitmap[BITS_TO_LONGS(nvqs)];
>> -    unsigned j;
>>   -    memcpy(bitmap, s->batch_notify_vqs, sizeof(bitmap));
>> -    memset(s->batch_notify_vqs, 0, sizeof(bitmap));
>> -
>> -    for (j = 0; j < nvqs; j += BITS_PER_LONG) {
>> -        unsigned long bits = bitmap[j / BITS_PER_LONG];
>> -
>> -        while (bits != 0) {
>> -            unsigned i = j + ctzl(bits);
>> +    for (unsigned i = 0; i < s->conf->num_queues; i++) {
> 
> Is this bitmap dense enough that you want to iterate by index,

By 'iterate by index' do you mean the actual iteration with 'j'?

> or is it
> sparse enough to iterate via find_first_bit/find_next_bit?

I looked at find_first_bit/find_next_bit() but they seemed to do
a lot more than test_and_clear_bit(). As Stefan said this is hot
path, I thought this would be cheaper, but haven't profiled the
performance.

> In either case, leave the copy of  s->conf->num_queues to a local variable.

That is sensible to do :)