On 2/19/25 9:27 PM, Steven Rostedt wrote:
On Wed, 19 Feb 2025 21:15:08 -0500
Waiman Long <llong@xxxxxxxxxx> wrote:
Writer - setting them:
WRITE_ONCE(lock)
smp_wmb()
WRITE_ONCE(type)
Clearing them:
WRITE_ONCE(type, 0)
smp_wmb()
WRITE_ONCE(lock, NULL)
Reader:
READ_ONCE(type)
again:
smp_rmb()
READ_ONCE(lock)
smp_rmb()
if (READ_ONCE(type) != type)
goto again
Do you really need the READ/WRITE_ONCE() with the memory barriers? From
what I understand, the compiler can't even assume what it read is the same
after passing a memory barrier like that. So there should be no reason it
can reread the memory location after a barrier.
You may be right. However, without using a READ_ONCE/WRITE_ONLY, a
compiler can potentially break up the read/write into multiple smaller
trunks resulting in partial data. So I will use them to be on the safe
side. In this particular scenario above, we may not need to use them on
type as we are going to reread it. I will keep them for lock though.
Cheers,
Longman
-- Steve
