Unverified Commit 78f16256 authored by Tobias Koppers's avatar Tobias Koppers Committed by GitHub

turbo-tasks-backend: improve print_cache_item_size instrumentation (#91742)

### What?

Fixes a compilation hang that occurred when the `print_cache_item_size`
debug feature was enabled,
and cleans up the surrounding instrumentation code.

**Root cause of the hang (double lock / deadlock):**

During `persist_snapshot`, the storage iterates over all modified tasks.
For each task it calls:

```rust
let inner = self.shard.storage.map.get(&task_id).unwrap();
//          ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
//          acquires a READ lock on the DashMap shard for task_id
let item = (self.shard.process)(task_id, &inner, &mut self.buffer);
```

The `process` closure receives `&TaskStorage` while that **read lock is
still held**. Inside the
closure, the old code called:

```rust
self.get_task_name(task_id, turbo_tasks)
```

`get_task_name` creates an `ExecuteContext`, then calls
`ctx.task(task_id, TaskDataCategory::Data)`,
which calls `storage.access_mut(task_id)`, which calls
`self.map.entry(key)`. That `entry()` call
tries to acquire a **write lock on the same DashMap shard** — which is
already read-locked on the
same thread. `parking_lot::RwLock` is not reentrant, so the thread
blocks forever waiting for its
own read lock to be released.

**Fix:** the `&TaskStorage` reference (`inner`) is already in scope
inside the closure. Call
`inner.get_persistent_task_type()` directly instead of going through
`get_task_name`, which avoids
any additional lock acquisition.

**Additional changes in this PR:**

1. **Make compressed-size reporting opt-in** — split
`print_cache_item_size` into two Cargo features.
The base feature now shows uncompressed sizes only (no lz4 overhead). A
new
`print_cache_item_size_with_compressed` feature re-enables
compressed-size reporting for when
   you need it.

2. **Extract helpers to eliminate duplication** — the output block
repeated `#[cfg]`-gated
`FormatSizes { … }` struct literals six times and duplicated the
`task_name` computation twice.
   Extract:
- `FormatSizes` struct + `impl Display` (with `#[cfg]` isolated inside)
- `TaskCacheStats::task_name(storage)` — single source for the grouping
key
- `TaskCacheStats::sort_key()` — encapsulates the `#[cfg]`-switched
primary sort field
- `TaskCacheStats::format_total/data/avg_data/meta/avg_meta()` — each
returning `FormatSizes`

### Why?

The compilation was hanging and unusable. The lock ordering bug was
non-obvious because the re-entrant
acquisition happened across an abstraction boundary (`get_task_name`
hiding the `access_mut` call).

### How?

Avoid the redundant lock by using the `&TaskStorage` already provided to
the closure.
Pure refactor for everything else — no behaviour change for any given
feature combination.

---------
Co-authored-by: 's avatarTobias Koppers <sokra@users.noreply.github.com>
Co-authored-by: 's avatarClaude <noreply@anthropic.com>
parent b5cb015c
...@@ -14,7 +14,8 @@ workspace = true ...@@ -14,7 +14,8 @@ workspace = true
[features] [features]
default = [] default = []
print_cache_item_size = ["dep:lzzzz"] print_cache_item_size_with_compressed = ["print_cache_item_size", "dep:lzzzz"]
print_cache_item_size = []
no_fast_stale = [] no_fast_stale = []
aggregation_update_no_batch = [] aggregation_update_no_batch = []
verify_serialization = [] verify_serialization = []
......
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment