Conversation

Notices

1. Embed this notice
   Haelwenn /элвэн/ :triskell: (lanodan@queer.hacktivis.me)'s status on Wednesday, 29-Nov-2023 22:14:15 JST Haelwenn /элвэн/ :triskell:
   @marcan What the fuck, why would you even link sector sizes with memory pages.
   • Embed this notice
     Neal Gompa (ニール・ゴンパ) :fedora: (conan_kudo@fosstodon.org)'s status on Wednesday, 29-Nov-2023 22:27:02 JST Neal Gompa (ニール・ゴンパ) :fedora:

     in reply to

     @lanodan @marcan This is a Linux thing. A lot of filesystems have had this problem, some still do, and some are in even worse position than Btrfs. For example, F2FS does this, doesn't support subpage at all, and it used to crash when you tried to use it on non-4K kernels (that was recently fixed).

     Haelwenn /элвэн/ :triskell: likes this.
   • Embed this notice
     feld (feld@bikeshed.party)'s status on Thursday, 30-Nov-2023 01:05:51 JST feld

     in reply to
     @lanodan IO performance. It's really important to have these sizes aligned. And then the write to disk happens in a single SCSI WRITE command -- doesn't get split.
   • Embed this notice
     Haelwenn /элвэн/ :triskell: (lanodan@queer.hacktivis.me)'s status on Thursday, 30-Nov-2023 01:13:47 JST Haelwenn /элвэн/ :triskell:

     in reply to

     • feld
     @feld I'm aware of problems like write amplification (it's one of the reason why I hate the lack of detailed spec sheets on product pages), which is exactly why you shouldn't be too willy-nilly with defaults.
     And AFAIK alignment for RAM isn't per-page but more like per-word.
   • Embed this notice
     feld (feld@bikeshed.party)'s status on Thursday, 30-Nov-2023 01:16:15 JST feld

     in reply to
     @lanodan I also don't know about Linux as much here but in FreeBSD land the kernel devs assured me the alignment guarantees no copies/modifications from memory to storage if these are in alignment which will punch your IOPS up a lot, especially if you're running a database (also using correct sizes for its own storage)
   • Embed this notice
     Marcus Müller (funkylab@mastodon.social)'s status on Friday, 01-Dec-2023 01:43:24 JST Marcus Müller

     in reply to

     @lanodan @marcan the MMU / mmap() can only map pages between kernel space and userland, so it *does* make a lot of sense

   • Embed this notice
     Neal Gompa (ニール・ゴンパ) :fedora: (conan_kudo@fosstodon.org)'s status on Friday, 01-Dec-2023 01:43:24 JST Neal Gompa (ニール・ゴンパ) :fedora:

     in reply to

     • Marcus Müller

     @funkylab @lanodan @marcan No. It's a terrible design choice, but we're stuck with it until everything moves to folios.

     Haelwenn /элвэн/ :triskell: likes this.
   • Embed this notice
     Neal Gompa (ニール・ゴンパ) :fedora: (conan_kudo@fosstodon.org)'s status on Friday, 01-Dec-2023 03:14:56 JST Neal Gompa (ニール・ゴンパ) :fedora:

     in reply to

     • Marcus Müller

     @funkylab @lanodan @marcan I don't know if that's something separate from the folios work that is going on for all the filesystems, but if it isn't, then yes it's being adopted slowly everywhere.

     Haelwenn /элвэн/ :triskell: likes this.
   • Embed this notice
     Marcus Müller (funkylab@mastodon.social)'s status on Friday, 01-Dec-2023 03:14:58 JST Marcus Müller

     in reply to

     • Neal Gompa (ニール・ゴンパ) :fedora:

     @Conan_Kudo @lanodan @marcan somewhat relatedly, I haven't been following things, but did memory folios actually get a lot of traction?

   • Embed this notice
     Marcus Müller (funkylab@mastodon.social)'s status on Friday, 01-Dec-2023 03:14:59 JST Marcus Müller

     in reply to

     • Neal Gompa (ニール・ゴンパ) :fedora:

     @Conan_Kudo @lanodan @marcan ah sorry, didn't mean to imply it was a *good* idea, just that if you're designing something to be easy to implement, going sector=page is an easy choice.