The first commit moves m_requested_hb_cmpctblocks and m_provides_cmpctblocks from CNodeState to Peer and also makes the two fields atomic.
The second commit moves fPreferredDownload to Peer and makes it atomic. Since m_num_preferred_download_peers is a counter based on this field, it can also be atomic. This also lets us drop cs_main usage in a few places. I found this commit harder to reason about, specifically whether the asserts in FinalizeNode were still safe for m_num_preferred_download_peers if another thread could modify the value before/after the m_num_preferred_download_peers -= peer->fPreferredDownload line and trigger the assert. Afaict, these asserts are still safe because ProcessMessage cannot run for the same peer as the peer being cleaned up in FinalizeNode. I am also happy to drop this second commit or explain my reasoning in more detail.
For reference:
m_requested_hb_cmpctblocks can be accessed from:
- NewPoWValidBlock (read) (b-msghand, http, ipc? threads)
- ProcessMessage (write) (b-msghand)
- SendMessages (read) (b-msghand)
m_provides_cmpctblocks can be accessed from:
- MaybeSetPeerAsAnnouncingHeaderAndIDs (read) (b-msghand, http, ipc?)
- HeadersDirectFetchBlocks (read) (b-msghand)
- ProcessMessage (write) (b-msghand)
fPreferredDownload can be accessed from:
- ProcessMessage (read/write) (b-msghand)
- SendMessages (read) (b-msghand)
- FinalizeNode (read) (bitcoind, b-net)
m_num_preferred_download_peers can be accessed from:
- ProcessMessage (write) (b-msghand)
- SendMessages (read) (b-msghand)
- FinalizeNode (read/write) (bitcoind, b-net)
