highperfocused/bitcoin
1
0
Fork 0
mirror of https://github.com/bitcoin/bitcoin.git synced 2026-01-19 14:53:43 +01:00
Code Issues Packages Projects Releases Wiki Activity

Merge bitcoin/bitcoin#32465: thread-safety: fix annotations with REVERSE_LOCK

Browse Source 
a201a99f8c thread-safety: fix annotations with REVERSE_LOCK (Cory Fields)
aeea5f0ec1 thread-safety: add missing lock annotation (Cory Fields)
832c57a534 thread-safety: modernize thread safety macros (Cory Fields)

Pull request description:

  This is one of several PRs to cleanup/modernize our threading primitives.

  While replacing the old critical section locks in the mining code with a `REVERSE_LOCK`, I noticed that our thread-safety annotations weren't hooked up to it. This PR gets `REVERSE_LOCK` working properly.

  Firstly it modernizes the attributes as-recommended by the [clang docs](https://clang.llvm.org/docs/ThreadSafetyAnalysis.html) (ctrl+f for `USE_LOCK_STYLE_THREAD_SAFETY_ATTRIBUTES`). There's a subtle difference between the old `unlock_function` and new `release_capability`, where our `reverse_lock` only works with the latter. I believe this is an upstream bug. I've [reported and attempted a fix here](https://github.com/llvm/llvm-project/pull/139343), but either way it makes sense to me to modernize.

  The second adds a missing annotation pointed out by a fixed `REVERSE_LOCK`. Because clang's thread-safety annotations aren't passed through a reference to `UniqueLock` as one may assume (see [here](https://clang.llvm.org/docs/ThreadSafetyAnalysis.html#no-alias-analysis) for more details), `cs_main` has to be listed explicitly as a requirement.

  The last commit actually fixes the `reverse_lock` by making it a `SCOPED_LOCK` and using the pattern [found in a clang test](https://github.com/llvm/llvm-project/blob/main/clang/test/SemaCXX/warn-thread-safety-analysis.cpp#L3126). Though the docs don't describe how to accomplish it, the functionality was added [in this commit](6a68efc959). Due to aliasing issues (see link above), in order to work correctly, the original mutex has to be passed along with the lock, so all existing `REVERSE_LOCK`s have been updated. To ensure that the mutexes actually match, a runtime assertion is added.

ACKs for top commit:
  fjahr:
    re-ACK a201a99f8c
  davidgumberg:
    reACK a201a99f8c
  theuni:
    Ok, done. Those last pushes can be ignored. ACKs on a201a99 are still fresh.
  ryanofsky:
    Code review ACK a201a99f8c. Just dropping 0065b9673db5da2994b0b07c1d50ebfb19af39d0 and fixing incorrect `reverse_lock::lockname` initialization since last review.
  TheCharlatan:
    Re-ACK a201a99f8c

Tree-SHA512: 2755fae0c41021976a1a633014a86d927f104ccbc8014c01c06dae89af363f92e5bc5d4276ad6d759302ac4679fe02a543758124d48318074db1c370989af7a7
This commit is contained in:
Ryan Ofsky
2025-06-17 13:40:14 -04:00
parent 1be688f575 a201a99f8c
commit 5e6dbfd14e
6 changed files with 33 additions and 23 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
src/node/interfaces.cpp
View File

@@ -431,7 +431,7 @@ public:
};
// NOLINTNEXTLINE(misc-no-recursion)
bool FillBlock(const CBlockIndex* index, const FoundBlock& block, UniqueLock<RecursiveMutex>& lock, const CChain& active, const BlockManager& blockman)
bool FillBlock(const CBlockIndex* index, const FoundBlock& block, UniqueLock<RecursiveMutex>& lock, const CChain& active, const BlockManager& blockman) EXCLUSIVE_LOCKS_REQUIRED(cs_main)
{
if (!index) return false;
if (block.m_hash) *block.m_hash = index->GetBlockHash();
@@ -443,7 +443,7 @@ bool FillBlock(const CBlockIndex* index, const FoundBlock& block, UniqueLock<Rec
if (block.m_locator) { *block.m_locator = GetLocator(index); }
if (block.m_next_block) FillBlock(active[index->nHeight] == index ? active[index->nHeight + 1] : nullptr, *block.m_next_block, lock, active, blockman);
if (block.m_data) {
REVERSE_LOCK(lock);
REVERSE_LOCK(lock, cs_main);
if (!blockman.ReadBlock(*block.m_data, *index)) block.m_data->SetNull();
}
block.found = true;