[bukka]

The callers expect the input brigade to be empty when the filter returns, so buckets left on it leaked. Buckets put back with PSFS_FEED_ME to be seen again with the next data are now kept by the user filter and prepended to its input on the next call, which also fixes them aliasing the read chunk buffer, being re-filtered by preceding filters in a chain and being lost on the write path. Leftover buckets on any other status, or on the closing call, are released with the existing warning.

This is alternative to #20058

[Truta446]

Two things I noticed while testing this on PHP-8.4 with tests from master: ext/standard/tests/filters/stream_filter_register_mock_class_filter.phpt and stream_filter_register_class_coerce_consumed_by_ref_param.phpt (master only, --XFAIL-- This leaks memory) pass with this patch, so their XFAIL sections should be dropped when merging up. And the reentrant case from #23440 (an error handler that appends a bucket to $in while handling the "Unprocessed filter buckets" warning) still leaks 2 allocations with this patch, on debug and ASan.

[bukka]

Two things I noticed while testing this on PHP-8.4 with tests from master: ext/standard/tests/filters/stream_filter_register_mock_class_filter.phpt and stream_filter_register_class_coerce_consumed_by_ref_param.phpt (master only, --XFAIL-- This leaks memory) pass with this patch, so their XFAIL sections should be dropped when merging up. And the reentrant case from #23440 (an error handler that appends a bucket to $in while handling the "Unprocessed filter buckets" warning) still leaks 2 allocations with this patch, on debug and ASan.

Good points. All should be fixed now.