[ANN] netdev development stats for 7.3
From: Jakub Kicinski <kuba@kernel.org>
Date: 2026-08-19 18:47:00
Preface ------- As is tradition here are the development statistics based on mailing list traffic on netdev@vger. These stats are somewhat like LWN stats: https://lwn.net/Articles/1088776/ but more focused on mailing list participation. In particular "review score" tries to capture the balance between reviewing other people's code vs posting patches. Previous 3 reports: - for 7.0: https://lore.kernel.org/20260212124208.187e53ae@kernel.org (local) - for 7.1: https://lore.kernel.org/20260414182653.40d84ccc@kernel.org (local) - for 7.2 https://lore.kernel.org/20260617115319.43a5942d@kernel.org (local) General stats ------------- 7.3 release cycle spanned summer vacation season and netdev.conf. The direct comparisons to the previous release are a little misleading. The number of commits by maintainers dropped by 24%. back to 22 commits/day. Which used to be the usual rate before AI. Messages to the list dropped only by 7%, so less than the drop in commits. We were far less efficient at accepting code. Everyone has probably experienced the huge patch backlog so this is no surprise. Patch review has recovered to 64% in total and 54% across company (up 14% in both cases). Best result in a while. I think it's partially due to our use of the SUIE ranking (see the netdev.conf talk by Paolo). Since we now always have >300 patches awaiting review looking thru one's inbox or even patchwork is impractical. SUIE shows all patchwork data, sorted by a combination of "readiness" and "author's reviewer score". The readiness means CI signals + did anyone review the patch already. This means that maintainers apply reviewed patches first - potentially having a positive effect on overall review ratio. The only volume metric which did not go down in this release was growth in the number of individuals that emailed the list. We have seen 4.5% more people in this release cycle. Of course, these are mostly LLM-aided "pure authors" who never comment on other peoples posting. Posting efficiency ------------------ Given the overall drop in our efficiency in converting postings into commits I decided to dig a little deeper into that side of the data. First, I fixed how we count "revisions" in the scripts - they used to only count postings, so if versioning started before the release cycle we'd miss-count (e.g. if v22 was merged, but script only saw postings from v19 on - we'd count it as the fourth revision). With that fix, our "average revisions" numbers are at 1.7 for a single patch and a round 3.0 for a patch series. Here is the ranking of people who need fewest and most patch postings for a commit to appear in the tree: Lowest postings/commit (min 5): Highest postings/commit (min 2 commits): 1 [13@ 1.00] Ciprian Regus 1 [ 3@39.67] Christian Marangi (Ansuel) --2-[56@ 1.00]-Florian Westphal 2 [ 5@19.20] javen 3 [ 7@ 1.00] Ido Schimmel 3 [ 8@14.38] Théo Lebrun 4 [ 8@ 1.00] Krzysztof Kozlowski 4 [ 3@14.00] Chuck Lever --5-[39@ 1.00]-Marc Kleine-Budde 5 [ 2@12.00] Jakub Sitnicki --6-[24@ 1.04]-Steffen Klassert 6 [ 2@11.50] Pengpeng Hou 7 [15@ 1.07] Ioana Ciornei 7 [ 4@11.25] Willem de Bruijn 8 [13@ 1.08] Arnd Bergmann 8 [ 5@10.80] Xin Xie 9 [12@ 1.08] Ilya Maximets 9 [ 4@10.25] Muhammad Nazim Amirul 10 [15@ 1.13] Mohsin Bashir 10 [ 6@ 9.00] Mahe Tardy -11-[20@ 1.20]-Antonio Quartulli 11 [12@ 8.83] Ratheesh Kannoth 12 [ 5@ 1.20] Avi Weiss 12 [ 6@ 8.50] Michael Chan 13 [ 5@ 1.20] Dawei Feng 13 [11@ 8.36] Daniel Golle 14 [ 8@ 1.25] Joe Damato 14 [ 2@ 8.00] Yehyeong Lee -15-[55@ 1.27]-Simon Wunderlich 15 [ 2@ 8.00] Weimin Xiong ^ ^ number of postings per commits 1 commit I marked with "-" our co-maintainers, who send us already reviewed patches / PRs. These are of course more likely to be accepted on the first attempt. And here is similar breakdown for companies (well, plus people who's employer we don't know): Lowest postings/commit (min 10): Highest postings/commit (min 5): 1 [ 14@ 1.00] Analog Devices 1 [ 5@19.20] Realsil 2 [ 41@ 1.15] Pengutronix 2 [ 5@10.80] Xin Xie 3 [ 20@ 1.20] (OpenVPN) 3 [11@ 8.36] Daniel Golle 4 [ 55@ 1.27] (batman-adv) 4 [16@ 8.19] Microsoft 5 [ 91@ 1.34] (netfilter) 5 [14@ 7.64] SUSE 6 [ 38@ 1.39] (mptcp) 6 [13@ 7.00] Broadcom 7 [ 24@ 1.46] (xfrm) 7 [10@ 6.80] Alvin Sun 8 [ 28@ 1.50] Linaro 8 [24@ 5.83] Marvell 9 [ 30@ 1.83] Arizona State 9 [ 8@ 5.62] Weiming Shi 10 [ 26@ 2.04] Gary Guo 10 [40@ 5.25] Bootlin 11 [ 12@ 2.33] Mojatatu 11 [28@ 5.18] IBM 12 [155@ 2.38] Meta 12 [10@ 5.10] Trustnetic 13 [ 25@ 2.48] Tsinghua University 13 [16@ 5.00] Oracle 14 [112@ 2.63] RedHat 14 [12@ 5.00] Code Construct 15 [107@ 2.75] nVidia 15 [51@ 4.88] Qualcomm I'm curious what sets the two sides of the comparison aside. I could venture two guesses. One - smaller patch sets are easier to merge. nVidia/Tariq is putting in the effort to break the submissions up which is probably part of the reason why they only need 2.75 revision vs (eg.) Broadcom's 7. With AI reviewers semi-randomly flagging issues the chances of a "clean" 15 patch submission are pretty much zero. Two - I suspect companies like nVidia (2.75 avg rev) and Meta (2.38 revs) do much more in-house AI review of their code than Broadcom (7 revs), Marvell (5.83 revs), or Qualcomm (4.88 revs). So tl;dr - break up your patch sets, run AI reviews locally first. Here (for no good reason) is a list of people / companies who made the most posts without getting a _single one_ of them merged: Nothing merged, people (min 30): Nothing merged, companies (min 30): 1 (***) [ 106] Louis-Alexis Eyraud 1 (***) [ 106] Collabora 2 (***) [ 86] Birger Koblitz 2 (***) [ 86] Birger Koblitz 3 (***) [ 83] Mingming Cao 3 (***) [ 68] Nebula Matrix 4 (***) [ 75] Long Li 4 (***) [ 64] James Hilliard 5 (***) [ 68] Illusion Wang 5 (***) [ 53] Luke Howard 6 (***) [ 64] James Hilliard 6 (***) [ 48] Onsemi 7 (***) [ 53] Luke Howard 7 (***) [ 44] Jonas Jelonek 8 (***) [ 48] Selvamani Rajagopal 8 (***) [ 40] SK Hynix 9 (***) [ 44] Jonas Jelonek 9 (***) [ 32] Akari Tsuyukusa 10 (***) [ 40] Byungchul Park 10 (***) [ 32] Rong Zhang 11 (***) [ 32] Akari Tsuyukusa 12 (***) [ 32] Rong Zhang 13 (***) [ 30] Amery Hung Testing and CI -------------- Customary top ranking of selftest contributors: Contributions to selftests: 1 [ 15] Mohsin Bashir 2 [ 11] Jakub Kicinski 3 [ 10] Matthieu Baerts 4 [ 8] Joe Damato 5 [ 6] Marcelo Mendes Spessoto Junior 6 [ 3] Minxi Hou 7 [ 3] Breno Leitao 8 [ 3] Willem de Bruijn 9 [ 3] Bobby Eshleman 10 [ 3] Maciej Fijalkowski 11 [ 3] Nirmoy Das 12 [ 3] Tushar Vyavahare The main infra side changes were around HW testing. The NICs we have in NIPA and the results reported by external runners are reaching stability. Hopefully we'll be soon reporting the HW test results to patchwork. Smaller changes: - wireless tree is now using netdev foundation machines - UI color highlight for TAP results - lots of stability / error handling fixes all over the place as all external services we talk to are periodically overloaded by AI bots Developer rankings ------------------ Top reviewers (cs): Top reviewers (msg): 1 ( ) [46] Jakub Kicinski 1 ( ) [90] Jakub Kicinski 2 ( ) [28] Simon Horman 2 ( ) [59] Andrew Lunn 3 ( ) [21] Andrew Lunn 3 ( ) [43] Simon Horman 4 ( ) [17] Paolo Abeni 4 ( ) [27] Paolo Abeni 5 ( +1) [ 8] Ido Schimmel 5 ( +1) [15] Ido Schimmel 6 ( +7) [ 7] Maxime Chevallier 6 ( +7) [15] Maxime Chevallier 7 ( +2) [ 6] Kuniyuki Iwashima 7 ( +3) [12] Kuniyuki Iwashima 8 ( -3) [ 6] Eric Dumazet 8 ( +8) [10] Krzysztof Kozlowski 9 ( +1) [ 5] Aleksandr Loktionov 9 ( +3) [10] Willem de Bruijn 10 (+12) [ 5] Vadim Fedorenko 10 (+30) [10] Vadim Fedorenko 11 (+16) [ 5] Breno Leitao 11 ( -3) [ 8] Aleksandr Loktionov 12 ( +3) [ 4] Willem de Bruijn 12 ( -7) [ 8] Eric Dumazet 13 ( -1) [ 4] Krzysztof Kozlowski 13 (+31) [ 8] Breno Leitao 14 (***) [ 4] Xin Long 14 ( -3) [ 7] Stanislav Fomichev 15 ( +5) [ 3] Stanislav Fomichev 15 ( +4) [ 6] Michael S. Tsirkin 16 (***) [ 3] Tung Nguyen Ido has been handling a lot of AI-generated fixes, primarily to the IP layer but not only. Similarly Willem has been handling AI fixes for AF_PACKET, tap, UDP etc. Xin Long is quite busy reviewing all the SCTP fixes. Tung Nguyen handles all TIPC fixes. Maxime has stepped up further in phylink reviews and has taken on stmmac reviews, which is the busiest driver, and certainly one which has the most contributors. We would have completely fallen apart without these people so huge thanks! Top authors (cs): Top authors (msg): 1 ( ) [5] Eric Dumazet 1 (+23) [18] Tony Nguyen 2 (***) [5] Xiang Mei (Microsoft) 2 ( ) [17] Tariq Toukan 3 ( -1) [4] Jakub Kicinski 3 ( +2) [15] Pablo Neira Ayuso 4 (***) [4] Ren Wei 4 (***) [14] Christian Marangi (Ansuel) 5 ( -2) [4] Tariq Toukan 5 ( +3) [14] Ratheesh Kannoth 6 (***) [3] Lorenzo Bianconi 6 (***) [14] Théo Lebrun 7 (***) [3] Xuanqiang Luo 7 (+12) [13] Markus Stockhausen 8 (***) [2] David Lee 8 ( -1) [13] Kuniyuki Iwashima 9 (***) [2] Doruk Tan Ozturk 9 (***) [12] Louis-Alexis Eyraud 10 (***) [2] Pengpeng Hou 10 ( ) [12] javen 11 (***) [2] Zhiling Zou 11 (+47) [12] David Howells 12 ( +4) [2] Yizhou Zhao 12 (+14) [12] Daniel Golle 13 (+49) [2] Ratheesh Kannoth 13 ( -9) [12] Wei Fang 14 ( -8) [2] Weiming Shi 14 (***) [11] Lorenzo Bianconi 15 (+23) [2] Runyu Xiao 15 (-14) [11] Eric Dumazet Eyeballing the left - "top author (change sets)" side - 10 of the names belong to AI fixers. On the right side we have the vendor maintainers at the top, and a large number of embedded folks who like to keep re-submitting giant series. Top scores (positive): Top scores (negative): 1 ( ) [684] Jakub Kicinski 1 (+26) [62] Tony Nguyen 2 ( ) [390] Simon Horman 2 (+47) [55] Christian Marangi (Ansuel) 3 ( ) [380] Andrew Lunn 3 ( +8) [52] Markus Stockhausen 4 ( ) [239] Paolo Abeni 4 ( -3) [51] Tariq Toukan 5 ( ) [126] Ido Schimmel 5 (***) [50] Louis-Alexis Eyraud 6 ( +7) [ 91] Maxime Chevallier 6 ( -2) [50] javen 7 (+18) [ 80] Vadim Fedorenko 7 ( -4) [50] Ratheesh Kannoth 8 ( ) [ 70] Aleksandr Loktionov 8 (***) [48] David Howells 9 ( -2) [ 64] Krzysztof Kozlowski 9 (+17) [48] Daniel Golle 10 (+21) [ 51] Tung Nguyen 10 ( -8) [44] Wei Fang 11 ( -1) [ 47] Willem de Bruijn 11 (***) [43] Théo Lebrun 12 (+45) [ 43] Kuniyuki Iwashima 12 (***) [42] Birger Koblitz 13 ( +4) [ 43] Michael S. Tsirkin 13 (***) [40] Mingming Cao 14 (***) [ 38] Xin Long 14 (***) [39] Ren Wei 15 ( +4) [ 37] Sasha Levin 15 (+26) [36] Long Li Company rankings ---------------- Top reviewers (cs): Top reviewers (msg): 1 ( ) [58] Meta 1 ( ) [125] Meta 2 ( ) [57] RedHat 2 ( ) [108] RedHat 3 ( ) [21] Andrew Lunn 3 ( ) [ 59] Andrew Lunn 4 ( ) [19] Intel 4 ( ) [ 40] Intel 5 ( ) [18] Google 5 ( ) [ 39] Google 6 ( ) [15] nVidia 6 ( ) [ 31] nVidia 7 ( +1) [ 9] Bootlin 7 ( +3) [ 18] Bootlin 8 ( +1) [ 6] Linaro 8 ( +1) [ 16] Linaro 9 (+14) [ 5] Qualcomm 9 (+12) [ 11] Qualcomm Top authors (cs): Top authors (msg): 1 ( ) [12] Google 1 ( +1) [52] Google 2 ( ) [10] Meta 2 ( -1) [46] Meta 3 (***) [ 7] Nebula Sec 3 ( ) [43] Intel 4 (+11) [ 7] Kylin Software 4 ( +1) [39] nVidia 5 ( ) [ 7] nVidia 5 ( -1) [37] RedHat 6 ( -2) [ 6] Intel 6 ( +2) [31] Qualcomm 7 ( -4) [ 5] RedHat 7 ( +8) [26] Bootlin 8 (+33) [ 5] Arizona State 8 ( +6) [20] AMD 9 (+16) [ 5] Tsinghua University 9 ( ) [19] Microsoft Top scores (positive): Top scores (negative): 1 ( +1) [768] Meta 1 (***) [72] Nebula Sec 2 ( -1) [734] RedHat 2 ( +5) [70] AMD 3 ( ) [380] Andrew Lunn 3 ( -1) [59] Marvell 4 ( +8) [124] Intel 4 (+30) [55] Christian Marangi (Ansuel) 5 ( ) [100] nVidia 5 ( +8) [52] Markus Stockhausen 6 (+10) [ 88] Google 6 ( -2) [50] Realsil 7 ( ) [ 79] Linaro 7 (***) [49] Collabora 8 (***) [ 63] Stable team 8 (+13) [48] Daniel Golle 9 ( +6) [ 51] Tung Nguyen 9 ( -1) [46] Microsoft AI reviews ---------- Last time I was complaining that we lack access to frontier models. This has been fixed a couple of weeks ago. We have Opus 5 + GPT 5.6 integrated into our instance of Sashiko. The switch from Opus 4.7 to Opus 5 led to a major increase in accuracy and number of reported problems. It also, unfortunately, doubled the cost of the reviews. Luckily Meta is still willing to foot the bill. Special shout out to Chris Mason for the help getting us the GPT access. With the cross reviews in place I'm really hoping we'll have enough confidence in the reviews to send them out to the list automatically. -- Code: https://github.com/kuba-moo/ml-stat Raw output: https://netdev.bots.linux.dev/ml-stats/stats-7.3