buf_ring: Keep the full head and tail values

[andrew]

Summary

If a thread reads the head but then sleeps for long enough that
another thread fills the ring and leaves the new head with the
expected value then the cmpset can pass when it should have failed.

To work around this keep the full head and tail value and use the
upper bits as a generation count.

Sponsored by: Arm Ltd

[markj]

sys/sys/buf_ring.h

85

If you define i in the loop header, the ifdef becomes a bit less ugly.

91

Is there some reason to load the mask in the critical section? It also duplicates the load in the debug code above.

[kib]

Am I right that this revision does not fix the issue, but covers it with more bits from the unused space in the head index? It is still possible that head wraps around, just less likely now that full 2^32 iterations needs to occur under it?

[andrew]

In D46151#1052139, @kib wrote:

Am I right that this revision does not fix the issue, but covers it with more bits from the unused space in the head index? It is still possible that head wraps around, just less likely now that full 2^32 iterations needs to occur under it?

It's still possible, just very unlikely. Even if the value wrapped it would need to be the correct value for the cmpset to succeed.

[kib]

In D46151#1052141, @andrew wrote:

In D46151#1052139, @kib wrote:

Am I right that this revision does not fix the issue, but covers it with more bits from the unused space in the head index? It is still possible that head wraps around, just less likely now that full 2^32 iterations needs to occur under it?

It's still possible, just very unlikely. Even if the value wrapped it would need to be the correct value for the cmpset to succeed.

I suggest adding a comment, to save the future reader from realizing this on his own,

[andrew]

Update based on feedback

[andrew]

ping