wg: Remove some atomics for kp pointers.

[jhb]

Summary

Atomics are not needed for accesses synchronized by a mutex.

There is one unlocked read of the pointer for which the atomic
load doesn't do anything (in terms of limiting races).

Atomics remain for accesses under the NET_EPOCH to provide
READ_ONCE semantics.

[emaste]

There is one unlocked read of the pointer for which the atomic load doesn't do anything (in terms of limiting races).

Does it make sense for this case to have a comment (i.e., explaining why the race is not important)?

[jhb]

sys/dev/wg/wg_noise.c

727

This is the unlocked read and it's just an optimization to avoid the lock below (and not an uncommon pattern, so I think it is obvious in context)

[emaste]

sys/dev/wg/wg_noise.c

727

Fair enough, I did not look at each case to find the unlocked read and when pointed out in context it's clear enough to me. [That said I don't think a comment wouldn't be out of place]

[jhb]

sys/dev/wg/wg_noise.c

586

Possibly these two should remain as atomics as they are used to provide READ_ONCE semantics.

702–703

Another READ_ONCE

860–861

Another READ_ONCE

889–890

Another READ_ONCE

[jhb]

• Keep atomics for READ_ONCE cases.

[jhb]

• Rebase post atomic(9).