General Session - Workshop Summary (Daniel Borkmann, Martin Lau & all other track leads)

### Overview This general session marked the conclusion of the annual LSF/MM/BPF (Linux Storage, Filesystem, Memory Management, and BPF) conference. Track leads gathered to provide high-level readouts summarizing the primary discussions, tentative consensus points, and planned patches across four core tracks: Storage, Filesystems (FS), Memory Management (MM), and BPF. Following the technical recaps, attendees and organizers conducted an open retrospective discussing conference logistics, session scheduling, remote participation, recordings, and social coordination.

### Key Topics - **Storage Track**: Topics included atomic/torn writes, Flexible Data Placement (FDP) for write amplification reduction on NVMe, userland PI pass-through improvements, copy offload progress, Rust support in the block layer, replacing biovecs with folios, device probing renumbering bugs, and NVMe multipathing over asymmetric or flaky links. - **Filesystems Track**: Focus areas included APIs for atomic writes, NFS change cookie fixes, a shared `fs-next` branch to unify testing, Mount API enhancements (error reporting and mount notifications), evaluating new file system proposals (such as `thatfs` vs. extending FUSE), `fanotify` hooks for page faults, test runner coordination, and proposals for anonymous temporary file renaming and cache dropping. - **Memory Management Track**: Discussions covered memory tiering, cleanup and technical debt reduction in `hugetlbfs`, improving large page allocation reliability, and exploratory talks on CXL architectures. - **BPF Track**: Key items included struct_ops and BPF qdisc, GCC and LLVM compiler alignment, BPF memory model quirks (ARMv8 and PowerPC), BPF LSM and kfunc extensions, Android loader ecosystem integration, segmented stacks, verifier value tracking, static keys and jump tables, faster uprobes via a new system call, user stack trace capturing, pahole BTF improvements, netkit, open-coded iterators, IETF BPF standardization, and eBPF Foundation updates.

### Technical Details - **Storage & Block Layer**: The storage group noted that copy offload is finally on the verge of being merged after a ten-year development effort. In addition, block layer Rust support is ready for merging. A discussion on replacing `bio_vec` structures with folios and modern MM primitives proved productive enough to be split across multiple sessions. Addressing a device-probing issue where `dd` commands could trigger node reshuffling in `/dev` was also prioritized. - **Filesystems Infrastructure**: To streamline continuous integration across fragmented file system trees, participants agreed to create an `fs-next` branch merging major trees. For mount management, consensus leaned toward improving error reporting and mount notifications, though specific developer ownership remains open. A proposal for `thatfs` raised questions about whether requirements could instead be satisfied by extending FUSE. In memory file systems, Al Viro outlined cleanup efforts for persistent dcache handling across virtual file systems like debugfs. - **BPF Verification and Tooling**: Consensus was reached on APIs for static keys and jump tables within BPF. Work on faster uprobes demonstrated potential 2x improvements using a dedicated system call. Stack trace evolution focused on asynchronous user stack collection to reduce overhead, while compiler discussions centered on harmonizing GCC's BPF backend with LLVM.

### Takeaways & Logistics Discussion - **Logistics and Scheduling**: Participants commended the co-location of hotels and conference rooms for spontaneous hallway discussions. However, starting sessions promptly was identified as an issue, leading to proposals to reinstate a physical warning bell or buffer time between talks. Track scheduling challenges between overlapping filesystem and MM topics were also noted. - **Remote Participation & Recordings**: Track leads agreed that remote attendance serves as a vital fallback for absent developers, though remote audiences exceeding 20 participants risk logistics degradation. Regarding recordings, some attendees favored recording talks for LWN coverage and historical reference, while others expressed privacy concerns; organizers noted speakers can request sessions remain unrecorded or excluded from public platforms. - **Community Communication**: Attendees requested the restoration of an attendee mailing list (handling GDPR compliance via conference registration opt-ins) and advocated for coordinated evening venues to foster cross-track technical exchange.