### What does this PR do?
Resume work on https://github.com/oven-sh/bun/pull/21898
### How did you verify your code works?
Manually tested on MacOS, Windows 11 and Ubuntu 25.04. CI changes are
needed for the tests
---------
Co-authored-by: Claude Bot <claude-bot@bun.sh>
Co-authored-by: Claude <noreply@anthropic.com>
Co-authored-by: autofix-ci[bot] <114827586+autofix-ci[bot]@users.noreply.github.com>
Co-authored-by: Jarred Sumner <jarred@jarredsumner.com>
### What does this PR do?
On Linux, AbortSignal.timeout created a file descriptor for each timeout
and did not keep the event loop alive when a timer was active. This is
fixed.
### How did you verify your code works?
Fewer flaky tests
---------
Co-authored-by: autofix-ci[bot] <114827586+autofix-ci[bot]@users.noreply.github.com>
Co-authored-by: Claude <claude@anthropic.ai>
### What does this PR do?
dropAllLocks causes Thread::yield several times which means more system
calls which means more thread switches which means slower
### How did you verify your code works?
### What does this PR do?
Releasing heap access causes all the heap helper threads to wake up and
lock and then unlock futexes, but it's important to do that to ensure
finalizers run quickly.
That means releasing heap access is a balance between:
1. CPU usage
2. Memory usage
Not releasing heap access causes benchmarks like
https://github.com/oven-sh/bun/pull/14885 to regress due to finalizers
not being called quickly enough.
Releasing heap access too often causes high idle CPU usage.
For the following code:
```
setTimeout(() => {}, 10 * 1000)
```
command time -v when with defaultRemainingRunsUntilSkipReleaseAccess =
0:
>
> Involuntary context switches: 605
>
command time -v when with defaultRemainingRunsUntilSkipReleaseAccess =
5:
>
> Involuntary context switches: 350
>
command time -v when with defaultRemainingRunsUntilSkipReleaseAccess =
10:
>
> Involuntary context switches: 241
>
Also comapre the #14885 benchmark with different values.
The idea here is if you entered JS "recently", running any
finalizers that might've been waiting to be run is a good idea.
But if you haven't, like if the process is just waiting on I/O
then don't bother.
### How did you verify your code works?
## Summary
This PR optimizes the uSockets sweep timer to only run when there are
active connections that need timeout checking, rather than running
continuously even when no connections exist.
**Problem**: The sweep timer was running every 4 seconds
(LIBUS_TIMEOUT_GRANULARITY) regardless of whether there were any active
connections, causing unnecessary CPU usage when Bun applications are
idle.
**Solution**: Implement reference counting for active sockets so the
timer is only enabled when needed.
## Changes
- **Add sweep_timer_count field** to both C and Zig loop data structures
- **Implement helper functions** `us_internal_enable_sweep_timer()` and
`us_internal_disable_sweep_timer()`
- **Timer lifecycle management**:
- Timer starts disabled when loop is initialized
- Timer enables when first socket is linked (count 0→1)
- Timer disables when last socket is unlinked (count 1→0)
- **Socket coverage**: Applied to both regular sockets and connecting
sockets that need timeout sweeping
- **Listen socket exclusion**: Listen sockets don't increment the
counter as they don't need timeout checking
## Files Modified
- `packages/bun-usockets/src/internal/loop_data.h` - Added
sweep_timer_count field
- `src/deps/uws/InternalLoopData.zig` - Updated Zig struct to match C
struct
- `packages/bun-usockets/src/loop.c` - Helper functions and
initialization
- `packages/bun-usockets/src/internal/internal.h` - Function
declarations
- `packages/bun-usockets/src/context.c` - Socket link/unlink
modifications
## Test Results
Verified the optimization works correctly:
1. **✅ No timer during idle**: 5-second idle test showed no sweep timer
activity
2. **✅ Timer activates with connections**: Timer enables when server
starts listening
3. **✅ Timer runs periodically**: Sweep timer callbacks occur every ~4
seconds when connections are active
4. **✅ Timer deactivates**: Timer disables when connections are closed
## Performance Impact
This change significantly reduces CPU usage for idle Bun applications
with no active HTTP connections by eliminating unnecessary timer
callbacks. The optimization maintains all existing timeout functionality
while only running the sweep when actually needed.
## Test plan
- [x] Verify no timer activity during idle periods
- [x] Verify timer enables when connections are created
- [x] Verify timer runs at expected intervals when active
- [x] Verify timer disables when connections are closed
- [x] Test with HTTP server scenarios
- [ ] Run existing HTTP server test suite to ensure no regressions
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-authored-by: Claude Bot <claude-bot@bun.sh>
Co-authored-by: Claude <noreply@anthropic.com>
