#
✕ nick:arnd
#asahi · 2023-05-10 08:54
arndok, thanks!
#asahi · 2023-05-10 08:45
arndright, could be. I'm still running my own 6.2 kernel, plan to upgrade to 6.3 or 6.4-rc probably next week after I'm done with some other work, I might ask you again then if you have a patch to test
#asahi · 2023-05-10 08:41
arndi did not see anything odd using the same monitor with amdgpu, nouveau or the nvidia binary driver, so there is probably something in dcp as well that is odd
#asahi · 2023-05-10 08:40
arndjannau: I have an update on the monitor issues I kept seeing (config changes after display sleep requires me to click twice to discard xfce monitor dialog, sometimes black screen requires changing vt and back, screenlocker crashes, not acceping input). I still don't know what caused it, but after I ran it on a different monitor last week, I saw none of those problems, so it's clear that my particular monitor triggers something weird
#asahi · 2023-03-10 14:43
arndjonmason: see https://society.oftrolls.com/@AsahiLinux@treehouse.systems/109985145523510115 for the two major missing bits.
#asahi · 2023-02-26 22:12
arndbluetail: if you want to migrate your data to a 4k sector btrfs, this is fairly easy using 'btrfs send' and 'btrfs receive', storing a full image as a file on some other drive
#asahi · 2023-02-26 22:10
arndimplemented what? To be clear, I don't think that ext4 or btrfs have plans to support sectorsize>pagesize in the future
#asahi · 2023-02-26 22:08
arndthere is also a bit of overhead in the btrfs implementation for the sectorsize!=pagesize case
#asahi · 2023-02-26 22:08
arndaccessing large files is probably a bit faster with larger sectors, but I haven't done any measurements
#asahi · 2023-02-26 22:07
arndsame as RAM usage in page cache
#asahi · 2023-02-26 22:07
arnd16k sectors use a little more space than 4k sectors. For a kernel source tree this should be around 40% extra
#asahi · 2023-02-26 22:05
arndext4 has essentially the same problem, except that sectorsize < pagesize was always supported, and mkfs.ext4 defaults to 4k sectors on most distros
#asahi · 2023-02-26 22:04
arndthe problem is that 'mkfs.btrfs' defaults to sectorsize=pagesize, because that used to be the only supported setting. They added support for 4k sectors on 64k pages a while ago, and I think 5.19 added support for 4k sectors on 16k pages as well, but all other combinations are still broken
#asahi · 2023-02-26 22:02
arndbluetail: "mkfs.btrfs -s 4096" works fine to create a portable file system
#asahi · 2023-02-26 17:28
arndthis was for the arch/arm64/boot/dts pull requests
#asahi · 2023-02-26 17:27
arnd- The Apple T6000 (M1 Pro), T6001 (M1 Max) and T6002 (M1 Ultra) chips now have initial support. This is particularly nice as I am typing this on a T6002 Mac Studio with only a small number of driver patches."
#asahi · 2023-02-26 17:27
arnd"The most notable changes, by SoC platform, are:
#asahi · 2023-02-26 17:27
arndHe probably got it from my pull request message that I wrote for torvlds, which contains this paragraph:
#asahi · 2023-02-26 17:25
arndsjvn usually gets the facts right with his reporting, which probably contributed to the rest picking up that nonsense without fact checking
#asahi · 2023-01-14 23:00
arndbluetail: "smartctl -a /dev/nvme0n1" reports values for "Data Units Written" and "Percentage Used" that you can use to estimate the expected life. I'm mostly using an external drive for write intensive work, so the it's only at 3.24TB written on the internal 1TB drive, which is rounded to 0%.
#asahi · 2022-12-30 19:06
arndmerry: I have an USB-to-NVMe SSD based on JMicron JMS583 that has problems in UAS mode and with smartctl. It works mostly fine in usb-storage mode when you disable periodic smart statistics, just a little slower because of the lack of queuing and trim
#asahi · 2022-12-09 10:33
arndhttps://man.openbsd.org/bread.9 sounds like openbsd stores file data in a buffer cache with smaller than page size units
#asahi · 2022-12-09 10:31
arndI wonder if MacOS has the same sort of overhead from the page cache though, I suspect it does something more space efficient by caching smaller disk sectors rather than 16KB pages as the smallest unit
#asahi · 2022-12-09 10:27
arndtons of other source trees on my disk though, and those likely have very similar patterns, so even the 40% overhead would add up quickly
#asahi · 2022-12-09 10:26
arndas400: I would guess no more than 30 source trees, most of them completely obsolete. I tend to have only a handful of active source trees but a lot of separate object trees for testing, and those would be in tmpfs and not count towards disk usage though.
#asahi · 2022-12-09 09:37
arnd53920 files in the kernel (source+obj) are between 1 and 4096 bytes long, so storing those in 16KB units uses 4x the space, and storing them in 16KB units uses 16x the space. For the larger files, there is less overhead
#asahi · 2022-12-09 09:32
arndas400: absolutely. It's even worse if you just look at the source files. https://docs.google.com/spreadsheets/d/1_hNBVMZe29oAW3CKCm1OYMKSAmcXLARsD0KA6QjC4ic/edit?pli=1#gid=0 has my data if you are interested
#asahi · 2022-12-09 09:25
arndthe overhead grows to 400% when you compare 64KB pages to 4KB pages
#asahi · 2022-12-09 09:23
arndas400: there is actually one data point I have measured exactly: For my daily work (building a linux kernel from source), there is a 40% space storing all the source and object files (asahi defconfig) in 16KB units (sectors or pages, doesn't matter) compared to 4KB units
#asahi · 2022-12-09 09:19
arndI mainly cared about being able to access the same file system from both the 16KB page kernel I use as well as when mounting the drive on a different machine or when testing out a 4KB page kernel
#asahi · 2022-12-09 09:15
arndon the other hand, if you have a lot of files that are smaller than 4KB, it now has to transfer more data across the interface, which adds a little bit of latency on slower devices like USB
#asahi · 2022-12-09 09:14
arndI know that for a lot of SSD hardware, it takes the same time to access a 4KB sector as it takes to access a 16KB (or even 64KB) sector on disk since it can parallelize the access across channels
#asahi · 2022-12-09 09:12
arndfor comparing 4k sectors vs 16k sectors on a 16k page kernel, I would expect the usual workload specific tradeoffs to apply
#asahi · 2022-12-09 09:12
arndno idea. there should not be much overhead compared to a 4kb page kernel from the added indirection in the kernel code, that part looks fairly straightforward.
#asahi · 2022-12-09 09:10
arndgood to hear that you haven't had any issues, that gives me a little more confidence in putting important data on this partition
#asahi · 2022-12-09 09:09
arndI'm also using it now on 6.1-rc8
#asahi · 2022-12-09 09:05
arndchadmed: ok, I see. Apparently it's also been supported for longer than I thought: the btrfs support for 4K sectors already went into 5.19, just after I last tried (and failed) to use btrfs
#asahi · 2022-12-09 08:33
arndit previously already worked using 4k sectors and 64k pages, so they just had to use the same logic for both
#asahi · 2022-12-09 08:32
arndchadmed: in recent kernels btrfs supports 4k sectors in combination with 16k pages. There is still a warning about this being experimental, but the previous data corruption issue has been addressed
#asahi · 2022-08-18 15:45
arndas400[m]: it was AppliedMicro, which eventually into Ampere
#asahi · 2022-08-18 12:38
arndThere are also Cortex-A34, -A65AE and -R82 that are purely 64-bit, though I've never seen any chips with those
#asahi · 2022-08-18 12:37
arndas400[m]: I think most of of the third-party server cores had no aarch32 mode, but those also didn't make it into production. Fujitsu A64fx, Vulcan/ThunderX2 and Hisilicon Kunpeng have no aarch32 support IIRC, though Phytium, Xgene and anything based on Neoverse N1/V1 do.
#asahi · 2022-04-29 11:35
arndok
#asahi · 2022-04-29 11:34
arndNot sure if you can create an additional nvme 'namespace' on the internal drive as you could on a proper nvme disk. The internal drive is close to the nvme spec, but not quite the same
#asahi · 2022-04-29 11:32
arndat the moment, thunderbolt doesn't work in Linux, so it has to be a USB attached SSD
#asahi · 2022-04-29 11:31
arndvictorhooi[m]: last year I experimented with an external Thunderbolt SSD. At the time it was possible to install a bootable copy of MacOS on that, but I could not have two separate copies, or m1n1/Linux installed on it, because of the way the boot process is done. I don't know if this has changed in the meantime, I now use m1n1 on the internal disk to boot from an external USB-attached SSD
#asahi · 2022-04-10 13:02
arndI need to find one I can run, this is the only one I was running on my x86 machine, and it's packaged conveniently with a launcher tool here. The launcher itself seems to work fine fwiw, it only fails to run the actual Android app
#asahi · 2022-04-10 13:00
arndof course they shouldn't be, the question is whether that's just a bug in minecraft or in the android runtime environment
#asahi · 2022-04-10 12:57
arndI noticed that https://github.com/ChristopherHX/linux-packaging-scripts/releases/download/v0.3.3-683/Minecraft_Bedrock_Launcher-arm_aarch64-v0.3.3.683.AppImage works with 4K pages but fails with 16K pages. Is that a general problem with Android apps, or more likely a bug in the application itself?
#asahi · 2022-04-06 16:34
arndj`ey: 128GB DDR3-2400 vs 64GB LPDDR5

plus ›