On 12/1/23 12:06, Tejun Heo wrote:
Hello,
On Wed, Nov 29, 2023 at 11:01:04AM -0500, Waiman Long wrote:
...
Depending on how the cpumask operators are implemented, we may not have a
guarantee that testing CPU 2, for instance, will always return true. That is
Can you please elaborate this part a bit? I'm having a difficult time
imagining the sequence of operations where this would matter but that could
easily be me not being familiar with the details.
I may be a bit paranoid about incorrect result due to racing as I had been
burned before. Just testing a bit in the bitmask may probably be OK. I don't
Setting and clearing a bit is as atomic as it gets, right?
Yes, I think so.
think it will be a problem for x86, but I am less certain about other more
exotic architectures like arm64 or PPC which I am less familiar about. I add
a seqcount for synchronization just for the peace of mind. I can take the
seqcount out if you don't it is necessary.
I just can't think of a case where this would be broken. The data being read
and written is atomic. There's no way to break a bit operation into multiple
pieces. It is possible to write a really bone-headed bitmask operations
(like, if you shift the bits into place or sth) to make the bits go through
unintended changes but that'd just be a flat-out broken implementation. Even
for a bitmask where write accesses are synchronized through a spinlock, we
should still be able to use test_bit() without holding the lock. This seems
like a pretty basic assumption.
Adding unnecessary synchronization confuses the readers. If we don't need
it, we shouldn't have it.
OK, I will send a simplified v2 patch.
Cheers,
Longman
