March 2026 performance highlights: delivered robust IoUring-based buffer pool enhancements across folly and fbthrift, focusing on stability, deterministic threading, and scalable buffer handling to enable higher-throughput networking workloads. Key improvements span core buffer pool logic, kernel feature awareness, memory safety for shared buffers, and thread-safe sharing across worker threads and stress-test scenarios. Key features delivered: - IoUringZeroCopyBufferPool core enhancements: improved handling of io_uring_register return values during ZCRX_CTRL_EXPORT, and correct fd extraction from control data to ensure robust export paths. - Kernel feature awareness for zcrx: introduced SupportedFeatures tracking (importRing/exportRing/pageSize) and checkZcRxFeatures at construction time, gating import/export paths if the kernel lacks support. - Shared buffer metadata pre-initialization: pre-populated offset/length to support multiple CQEs per niov and concurrent metadata usage, improving throughput and predictability. - Clearer buffer terminology and capacity semantics: renamed numPages/pageSize to numBuffers/bufferSizeHint and set IOBuf capacity to actual CQE-reported length to preserve safety in shared memory regions. - Thread-safe buffer pool sharing: dispatches all io_uring operations to the owning thread via runInEventBaseThreadAndWait, and deterministic owner selection by zcRxQueueId to avoid cross-thread IO races. - Import configuration: introduced zcRxImport option so non-owner threads configure rings correctly for buffer pool import. - fbthrift-specific enhancements: IoUring Buffer Pool reliability and throughput improvements, including clearer parameter semantics, strengthened thread-safety, and integration with stress-test workflows for stability under load. Major bugs fixed: - IoUringZeroCopyBufferPool: Fix export handling by ensuring proper error checking and reading fd from ctrl.zc_export.zcrx_fd instead of using a potentially invalid return value. - Correct IOBuf capacity management by aligning capacity with CQE-length-reported data, preventing reads into kernel-modified regions. Overall impact and accomplishments: - Greater stability, predictability, and throughput for high-load scenarios requiring zero-copy IO; safer shared memory usage with correct capacity semantics; deterministic, thread-safe ring management reduces contention and race windows; improved test coverage and robustness for stress scenarios. Technologies/skills demonstrated: - io_uring, zcrx feature detection and gating, shared memory buffer management, C++ concurrency (event-base threading), memory safety, and integration testing with Thrift workload simulations.