The 500ms Slumber: How a One-Line Patch Unleashed 10x Throughput in QEMU

Occasionally, the most profound escalations in computational performance manifest almost as a whimsical irony. Jens Axboe, the esteemed Linux architect and progenitor of io_uring, recently pinpointed the catalyst for anomalous latency during QEMU operations with AHCI and SCSI disks. He subsequently proposed a remediation that, in “vacant” system environments, yields a staggering ten-fold throughput increase for the io_uring subsystem.

The predicament surfaced during the execution of io_uring regression tests within a virtualized milieu. On AHCI-interfaced devices, the tests would sporadically encounter “hangs” or expire due to timeouts, whereas identical scenarios on virtio-blk or NVMe would conclude within a solitary second. Axboe discerned that within a specific QEMU event loop, the ppoll() invocation could fall into a dormant state for as long as 500 milliseconds—even when the I/O queue possessed pending data for transmission. To the application, this manifested as an inexplicable, half-second stagnation arising from nowhere.

The allure of this discovery lies in its brevity: the fundamental correction is distilled into a singular line of source code. By integrating a notification to the primary loop following an SQE submission, the system is roused instantaneously, rather than languishing for nearly 500 milliseconds before dispatching requests to io_uring. The discourse surrounding the patch explicitly states that without this notification, ppoll() might “slumber” during the submission phase—an oversight particularly egregious for AHCI and SATA due to MMIO constraints and the absence of the ioeventfd mechanism that typically awakens the virtio loop.

Axboe notably enlisted the assistance of Claude to unravel this intricate causal chain; according to his account, the AI proved instrumental in elucidating the interplay between disparate event loops. During the diagnostic process, the AI inadvertently “corrupted” the test machine’s virtual disk but successfully facilitated its recovery. As of February 18, 2026, both remediations have been merged into the QEMU block branch, signifying their impending integration into the master project.