Linux Kernel Journey
The article outlines six distinct kernel‑level bugs affecting the eBPF tailcall feature across multiple Linux versions, explains the underlying causes and the commits that fixed them, and introduces a detection tool to verify whether a running kernel is affected.
The article explains how the Earliest Eligible Virtual Deadline First (EEVDF) scheduler, introduced in Linux 6.6, replaces CFS by using virtual deadlines and a lag concept to achieve both fairness and interactivity, and walks through concrete formulas, parameter settings, and step‑by‑step scheduling examples.
The article explains how Linux's Completely Fair Scheduler (CFS) uses nice values and weight tables to compute virtual runtime, organizes tasks in a red‑black tree, and demonstrates through experiments how SCHED_NORMAL, SCHED_BATCH, and SCHED_IDLE differ in priority, preemption rules, and CPU share.
The author, with over two decades of development experience, argues that learning the Linux kernel requires understanding real-world scheduling problems rather than merely copying kernel code, and outlines key questions about task_struct management, priority, preemption, CPU placement, fairness, and universal scheduler design.
This article introduces tcpw, a small eBPF‑based utility that traces TCP, UDP, and Unix‑domain sockets to display the five‑tuple of commands like curl or telnet, explains its command‑line options, shows concrete usage examples, and details the underlying BPF and Go implementation, including connect, accept, and fork tracing.
This article systematically reviews GPU monitoring for large‑scale AI training, covering MFU/HFU definitions, key DCGM metrics, NVLink bandwidth, common failure codes such as Xid and SXid, experimental insights on T4 and H100 GPUs, and practical case studies for diagnosing and mitigating performance drops.
This tutorial shows how to use an eBPF program with the PERF_EVENT type to trace kernel activity, collect samples via performance counters, and pinpoint which processes and functions consume the most execution time, covering dynamic tracing concepts and overflow handling.
This article demonstrates how to write an eBPF program that attaches to Linux soft‑interrupt entry and exit points, records timestamps in eBPF maps, computes handling duration, updates counters and histograms, and exposes the data to user space for performance analysis.
The article explains why backtracing with eBPF fentry on arm64 is harder than on x86, details the stack layout differences, shows how recent commits changed register saving, and provides a practical detection routine to locate the frame pointer and retrieve the tracee's instruction pointer.
This article explains the kfunc mechanism behind Linux's sched_ext, covering instruction modification, verification, address fixup, and detailing common kfunc functions such as scx_bpf_dispatch, scx_bpf_consume, and queue management APIs.
