Lmst

#LinuxPlumbers

Throwback: @mtaht's "State of the #Bufferbloat" address at @linuxplumbersconf conference 2024:

https://www.youtube.com/watch?v=nU38C0uvmOM

#LibreQoS #latency #QoE #WiFi #LinuxPlumbers #OSSummit #broadband #ISP #WISP #FISP #DaveTaht #RFC8290 #QoS #InternetServiceProvider #FLOSS #OpenSource #Linux #Starlink
#BandwidthIsALIE #Kernel #LinuxKernel #QualityOfExperience #FQ_CoDel
#QualityOfService #sch_CAKE #FQCoDel #schCAKE

An official video is also uploaded on YouTube: https://youtu.be/mRU1ZeNB9WY?feature=shared

#linux #kernel #damon #linuxplumbers

RE: https://social.kernel.org/objects/f6b89ea2-239f-40c8-b910-3eb3ef48c66b

Many dozens of recordings from this years @linuxplumbersconf are now available on YouTube. You can find them in the list of videos (https://www.youtube.com/@LinuxPlumbersConference/videos ) on in a dedicated playlist (https://www.youtube.com/playlist?list=PLVsQ_xZBEyN0XmmLaDFQEnuEu5HKSobpy ).

Abstracts as well as slides for most of the talks are available through the #LinuxPlumbers2024's schedule page: https://lpc.events/event/18/timetable/?view=lpc

#Linux #kernel #LinuxKernel #LinuxPlumbers

The slides for my #LinuxPlumbers talk on the formal verification of the #eBPF verifier are available at: https://pchaigno.github.io/assets/Linux%20Plumbers%202024%20Agni.pdf. I explained how Hari Vishwanathan et al. sped up the verification to a few minutes and how we're now able to continuously verify the verifier's range analysis in CI.

Slide showing the Agni runtime for various kernels. The plot shows the runtime got down to a few minutes and went up again in v6.8. The divide-and-conquer could be used here again to reduce the runtime, by verifying the subfunctions of reg_bounds_sync separately.

Everyone wants packet metadata in Linux, and everyone ends up (over-)using skb->mark. At #LinuxPlumbers, Arthur Fabre and Jakub Sitnicki (Cloudflare) discussed a new, larger metadata mechanism that could be accessed by all #eBPF programs along the stack.

Slides: https://lpc.events/event/18/contributions/1935/attachments/1669/3453/LPC%202024%20Metadata-4.pdf
Recording: https://www.youtube.com/live/rQtQ-Nd1ZmA?t=20977s

Slide showing one use case for packet metadata, to identify a packet in tracing. In this case, an ID needs to be preserved across all points of the stack, including through NAT, decapsulation, etc.

日本に戻りました。いろんな人と話せてよかったです。
#OSSummit #linuxplumbers #kvmforum

Luis Gerhorst (FAU) and his colleagues found that many existing #eBPF programs would be rejected due to the verifier's Spectre mitigations if they were loaded as unprivileged. They proposed to add lfences to cover those cases. #LinuxPlumbers

Slides: https://lpc.events/event/18/contributions/1954/attachments/1561/3249/presentation.pdf
Recording: https://www.youtube.com/live/rQtQ-Nd1ZmA?feature=shared&t=11011
Paper: https://arxiv.org/pdf/2405.00078
Code: https://gitlab.cs.fau.de/un65esoq/linux/-/commits/bpf-spectre

Slide illustrating a Spectre-PHT vulnerability in a program, with the control flow graph of the program and dotted lines for speculative execution branches. The example shows a type confusion vulnerability under speculative execution, with a scalar speculatively accessed as a pointer.

So much is now happening in #eBPF that tracing kernel functions isn't enough to debug the network stack. You need to trace inside BPF programs! At #LinuxPlumbers, my colleagues @brb and Gray Liang explained how they extended pwru to trace BPF programs and why it wasn't so simple.

Slides: https://lpc.events/event/18/contributions/1942/attachments/1659/3444/LPC'24_%20pwru-1.pdf
Recording: https://www.youtube.com/live/rQtQ-Nd1ZmA?t=6156s
Code: https://github.com/cilium/pwru

Slide showing the challenge of tracing BPF-based datapath with only a view of kernel functions. The top of the slide shows the output of the old pwru, where only Linux functions from the network stack are printed. The bottom shows the huge control-flow graph of the BPF code, highlighting that it's also a complex piece that needs tracing.

On the last day of #LinuxPlumbers, @kkdwivedi (EPFL) took a look at how we could redesign the network stack in a backward compatible way to improve its efficiency without losing isolation. Basically the holy grail of network stack research (dataplane OSes, kernel bypasses, kernel offloads). He thinks #eBPF could help implement a new Fibers abstraction to solve this problem.

Slides: https://lpc.events/event/18/contributions/1948/attachments/1671/3457/Nous%20LPC%202024.pdf
Recording: https://www.youtube.com/live/rQtQ-Nd1ZmA?feature=shared&t=23356

Slide presenting the what the new abstraction would need to provide, with an example of its application to the existing RX stack.

I found this picture at the hotel I stayed for #OSSummit and #linuxplumbers. Google says it is Commodore PET 8032. Rarely seen in Japan.

I really enjoyed my first #LinuxPlumbers, in Vienna. There were so many in-depth & engaging talks, which sparked plenty of impromptu discussions and hallway track huddles. I particularly liked Julia Lawall's talk on formal verification, meeting up with the KernelCI folks, and the chance to see more of my ELISA comrades in real life. Thanks to the organisers and all the other participants!

I gave a talk about my new CPU scheduler, named LAVD. While the recorded audio is not great, I hope it is still understandable.

"Using sched_ext to improve frame rates on the SteamDeck: Ideas behind the LAVD scheduler" in the sched_ext MC at the #linuxplumbers 2024, Vienna.

https://youtu.be/FmrRfD51cgc?t=6154

My Kernel Memory Management LPC Micro Conference talk titled "DAMON: Long-term Plans for Kernel That {Just Works,Extensible}" was successful.
The video[1] and slides[2] are available. I was obviously nervous since I forgot contents of the slides due to the great talks from two previous speakers (I highly recommend you to watch those). I like it though, since it reminds me how humble I am ;)

[1] https://www.youtube.com/live/CTWQ-d7pj5s?feature=shared&t=20182
[2] https://lpc.events/event/18/contributions/1768/

#linux #kernel #damon #linuxplumbers

And so we end a week of intense knowledge exchange! Thanks everyone who attended our talks and visited our booth at #OSSummit EU, #LinuxPlumbers and #LFEuropeMemberSummit. Big thank you to the sponsors and to the @linuxfoundation for making these events possible.👏 #OpenSource

Collection of six pictures taken during 2024 OSSummit Europe, LPC and LF Europe Member Summit in Vienna, Austria. From left to right, top to bottom order:
1. Audience of Changwoo Min's talk at LPC 2024, sched_ext MC.
2. Christian Gmeiner on the stage for Embedded Linux Conference in OSSummit Europe 2024.
3. Manuel Rego presenting at LF Europe Member Summit 2024.
4. Visitors of Igalia's booth at OSSummit Europe 2024.
5. André Almeida on the stage in the Graphics & DRM MC of LPC 2024.
6. Thadeu Cascardo presenting at LPC 2024 in the Real-time MC.

Double CPU scheduling (at host and at guest) causes inefficiencies in virtualized hosts. @joel_linux (Google) believes that eBPF could help the guest and host cooperate to improve performance. At #LinuxPlumbers now!

Slides: https://lpc.events/event/18/contributions/1733/attachments/1654/3418/pvsched_lpc24.pdf
Stream: https://www.youtube.com/live/lQXmHdefoXE

Overview of the architecture of Paravirt with the host-guest cooperation.

Optimizing memory management is becoming critical, especially with CXL and the hardware limits of TLB scaling.
Kaiyang Zhao (CMU) believes making memory management programmable with #eBPF could help. #LinuxPlumbers

Slides: https://lpc.events/event/18/contributions/1932/attachments/1646/3414/Towards%20Programmable%20Memory%20Management%20with%20eBPF%20(LPC%202024).pdf
Recording: https://www.youtube.com/live/rQtQ-Nd1ZmA?t=12426

Slide stating the basic idea/need for programmable memory management, with some ideas of how it could help. The slide mentions deciding where allocated pages should go, how many huge pages to give to processes, etc.

When troubleshooting with bpftrace, you don't want to have to troubleshoot bpftrace itself! bpftrace needs to be really reliable. At #LinuxPlumbers, Daniel Xu (Meta) explained why and how he built a Nix-based CI to address this challenge.

Slides: https://lpc.events/event/18/contributions/1938/attachments/1594/3360/Improving%20bpftrace%20reliability%20(1).pdf
Recording: https://www.youtube.com/live/mfMbtQYasB0?t=20158

Slide showing how the Nix-based CI is used and what it does.

Next at #LinuxPlumbers, to improve the #eBPF verifier, @kkdwivedi (EPFL) proposes KFlex, a better tradeoff between runtime sandboxing (ex. SFI) and static analysis (the eBPF verifiers). Out of all the approaches presented to improve/replace the verifier, this is IMO the most likely to make it upstream!

Slides: https://lpc.events/event/18/contributions/1943/attachments/1656/3426/KFlex%20LPC%202024.pdf
Recording: https://www.youtube.com/live/rQtQ-Nd1ZmA?t=3963
Code: https://github.com/rs3lab/KFlex