Using atomic loads after NET_EPOCH_ENTER or atomic stores before/after
NET_EPOCH_WAIT are not needed due to the existing barriers inside of
epoch_*().
Details
- Reviewers
markj
Diff Detail
- Repository
- rS FreeBSD src repository - subversion
- Lint
Lint Passed - Unit
No Test Coverage - Build Status
Buildable 46682 Build 43571: arc lint + arc unit
Event Timeline
sys/dev/wg/if_wg.c | ||
---|---|---|
1048 | I think this is an example of a place where we want to use atomic_load: there's no synchronization with writers so the values being loaded are immediately stale. That's not a problem in itself but it should be made obvious to anyone reading the code. I wrote an implementation of plain atomic_load/store_bool(), will post it shortly. What exactly does the net epoch section do for us here? |
sys/dev/wg/if_wg.c | ||
---|---|---|
960 | Hmm, I guess this comment is the reason for the NET_EPOCH around the callout_reset you asked about. It is not clear to me that a per-peer mutex used for managing callouts would be significantly less efficient? It would be easier to understand I think. | |
1048 | I pointed to a comment above. It is abusing the epoch to serialize with the code that disables the peer. |
sys/dev/wg/if_wg.c | ||
---|---|---|
960 | Oh. That's weird. I'm not familiar enough with these timers to say whether a mutex will be less efficient per se, but certainly contention could be a problem if the callout_pending() checks are done under a mutex. | |
1048 | Ok. So I think the atomic_load/store_*s are not required for correctness, but would still serve as mildly helpful annotations. We do have atomic_load/store_bool in main now, BTW. | |
1159 | Here's a situation where atomics would be helpful. There is zero synchronization between writers and readers of p_need_another_keepalive. |
sys/dev/wg/if_wg.c | ||
---|---|---|
960 | With the mutex you wouldn't need the callout_pending() calls at all. You would protect p_enabled with the mutex and could just use callout_stop under the lock in this function. Maybe I will mock up a version that uses callout_init_mtx and upload it for you to see how it looks instead. |