Merge bitcoin/bitcoin#34165: coins: don't mutate main cache when connecting block

cae6d895f8 fuzz: add target for CoinsViewOverlay (Andrew Toth)
86eda88c8e fuzz: move backend mutating block to end of coins_view (Andrew Toth)
89824fb27b fuzz: pass coins_view_cache to TestCoinsView in coins_view (Andrew Toth)
73e99a5966 coins: don't mutate main cache when connecting block (Andrew Toth)
67c0d1798e coins: introduce CoinsViewOverlay (Andrew Toth)
69b01af0eb coins: add PeekCoin() (Andrew Toth)

Pull request description:

  This is a slightly modified version of the first few commits of #31132, which can be merged as an independent change. It has a small benefit on its own, but will help in moving the parent PR forward.

  When accessing coins via the `CCoinsViewCache`, methods like `GetCoin` can call `FetchCoin` which actually mutate `cacheCoins` internally to cache entries when they are pulled from the backing db. This is generally a performance improvement for single threaded access patterns, but it precludes us from accessing entries in a `CCoinsViewCache` from multiple threads without a lock.

  Another aspect is that when we use the resettable `CCoinsViewCache` view backed by the main cache for use in `ConnectBlock()`, we will insert entries into the main cache even if the block is determined to be invalid. This is not the biggest concern, since an invalid block requires proof-of-work. But, an attacker could craft multiple invalid blocks to fill the main cache. This would make us `Flush` the cache more often than necessary. Obviously this would be very expensive to do on mainnet.

  Introduce `CoinsViewOverlay`, a `CCoinsViewCache` subclass that reads coins without mutating the underlying cache via `FetchCoin()`.

  Add `PeekCoin()` to look up a Coin through a stack of `CCoinsViewCache` layers without populating parent caches. This prevents the main cache from caching inputs pulled from disk for a block that has not yet been fully validated. Once `Flush()` is called on the view, these inputs will be added as spent to `coinsCache` in the main cache via `BatchWrite()`.

  This is the foundation for async input fetching, where worker threads must not mutate shared state.

ACKs for top commit:
  l0rinc:
    ACK cae6d895f8
  sipa:
    reACK cae6d895f8
  sedited:
    Re-ACK cae6d895f8
  willcl-ark:
    ACK cae6d895f8
  vasild:
    Cursory ACK cae6d895f8
  ryanofsky:
    Code review ACK cae6d895f8. PR is basically back to the form I had acked the first time, implementing `PeekCoin()` by calling `GetCoin()`. This is not ideal because `PeekCoin()` is not supposed to modify caches and `GetCoin()` does that, but it at least avoids problems of the subsequent approach tried where `GetCoin()` calls `PeekCoin` and would result in bugs when subclasses implement `GetCoin` forgetting to override `PeekCoin`. Hopefully #34124 can clean all of this by making relevant methods pure virtual.

Tree-SHA512: a81a98e60ca9e47454933ad879840cc226cb3b841bc36a4b746c34b350e07c546cdb5ddc55ec1ff66cf65d1ec503d22201d3dc12d4e82a8f4d386ccc52ba6441
This commit is contained in:
Ryan Ofsky
2026-02-19 21:39:14 -05:00
10 changed files with 386 additions and 38 deletions

View File

@@ -10,6 +10,7 @@
#include <attributes.h>
#include <chain.h>
#include <checkqueue.h>
#include <coins.h>
#include <consensus/amount.h>
#include <cuckoocache.h>
#include <deploymentstatus.h>
@@ -489,9 +490,9 @@ public:
//! can fit per the dbcache setting.
std::unique_ptr<CCoinsViewCache> m_cacheview GUARDED_BY(cs_main);
//! Temporary CCoinsViewCache layered on top of m_cacheview and passed to ConnectBlock().
//! Reused CoinsViewOverlay layered on top of m_cacheview and passed to ConnectBlock().
//! Reset between calls and flushed only on success, so invalid blocks don't pollute the underlying cache.
std::unique_ptr<CCoinsViewCache> m_connect_block_view GUARDED_BY(cs_main);
std::unique_ptr<CoinsViewOverlay> m_connect_block_view GUARDED_BY(cs_main);
//! This constructor initializes CCoinsViewDB and CCoinsViewErrorCatcher instances, but it
//! *does not* create a CCoinsViewCache instance by default. This is done separately because the