#
✕ nick:null
#asahi · 2025-07-29 16:55
null1
#asahi · 2024-01-09 17:56
nullWait for this new fix from Szabolcs to land and we can go test and so on
#asahi · 2024-01-09 17:55
nullmaz: yep
#asahi · 2024-01-09 17:55
nullThe older issue was https://sourceware.org/bugzilla/show_bug.cgi?id=30076
#asahi · 2024-01-09 17:54
nullmaz: that particular issue got fixed (by virtue of delegating the issue to the linker), we *could* go enable it in the kernel now, modulo this new issue
#asahi · 2022-10-24 16:26
nullmarcan: tbh, I think this is one for the mailing lists, since there are others who care who don't sit on IRC, but I'll go over there now you've poked :)
#asahi · 2022-10-24 16:24
null(MMU off, definitely, unless it's going to promise to save/restore everything for uss)
#asahi · 2022-10-24 16:24
nullSure; that's roughly inline with what I was thinking
#asahi · 2022-10-24 16:06
nullWindowPain: I'd hoped that too, until I started getting messaged by badly written IRC bots
#asahi · 2022-10-24 16:02
nullMany nice properties of PSCI stem from having the exception boundary, so if we don't have that, we need to reconsider how everything's supposed to work
#asahi · 2022-10-24 16:02
nullgets reset, the state of CPUs at handover points, etc).
#asahi · 2022-10-24 16:01
nullPSCI and UEFI runtime calls have quite different behaviours that need to be considered (e.g. requisite TTBR management and concurrency/mutual-exclusion), and I'm really not keen on tying a "pretend" PSCI implementation behind UEFI calls as I suspect it's going to cause longer-term maintenance pain and cause us to overlook issues if we're lax on specifying how it's supposed to work (e.g. if/when/how CPU state
#asahi · 2022-10-24 15:59
null(which is why I said "ish")
#asahi · 2022-10-24 15:58
nullThere's no idle or CPU-off for spin-table today
#asahi · 2022-10-24 15:55
nullugh, sorry, tab completion for /win 17 went wrong
#asahi · 2022-10-24 15:55
nullWindowPain: 17
#asahi · 2022-10-24 15:23
nullI'd much rather we had a generic SMP spin table-ish mechanism that allows us to hand the CPU back, with a well defined CPU state
#asahi · 2022-10-24 15:22
nullTBH, I don't think we should try to have a non={HVC,SMC} PSCI here, since we *cannot* do most of the things PSCI functionally does, and we will need a very different contract for handing over CPU state
#asahi · 2022-02-08 18:30
nullFunny, that
#asahi · 2022-02-08 18:04
nullaxboe: cool. I reckon I need to go and investigate pidbench on a stable base to see what the breakdown of that cost is
#asahi · 2022-02-08 18:03
nullj`ey: thanks
#asahi · 2022-02-08 18:03
nullAh, I see
#asahi · 2022-02-08 18:03
null(so that I can go build up a setup fo comparison)
#asahi · 2022-02-08 18:02
nullaxboe: I take it this is with something close to v5.17-rc3?
#asahi · 2022-02-08 17:57
nullj`ey: it should practically only be a TBNZ, but logically it makes sense to avoid
#asahi · 2022-02-08 17:57
nullj`ey: agreed -- fancy sending a patch? :)
#asahi · 2022-02-08 17:54
nullhttps://paste.debian.net/1230046/ is the whole test; just calls getpid 2^24 times
#asahi · 2022-02-08 17:53
nullI don't have linux on an M1 atm, so testing on an entirely unrelated dev board
#asahi · 2022-02-08 17:53
null[mark@gravadlaks:~]% perf_4.17 stat --null ./pidbench
#asahi · 2022-02-08 17:53
nullyup
#asahi · 2022-02-08 17:52
nullI have "pidbench" locally
#asahi · 2022-02-08 17:40
nullaxboe: does that just run forever, or does it eventually summarize? Is the 'IOPS' bit roughly the nr of syscalls/sec (when those options are used) ?
#asahi · 2022-02-08 17:26
null(for syscall / entry, that is)
#asahi · 2022-02-08 17:26
nullaxboe: if there's a concrete benchmark where you think things are slow, I'm happy to take a look if you could send a mail to LKML or LAKML
#asahi · 2022-02-08 17:25
null(e.g. could just be config, could be more stuff that we save/restore like the PAC keys)
#asahi · 2022-02-08 17:25
nullI don't have an recent x86 box to compare with, and there could be a bunch of reasons why syscalls are slower, so difficult to say much without more info
#asahi · 2022-02-08 17:23
nullj`ey: sure; just trying to figure out what the question/context was
#asahi · 2022-02-08 17:22
nullj`ey: the "syscalls look more expensive" bit?
#asahi · 2021-09-22 17:22
nullSounds good -- my plan is to append something like (?) when it's the LR at an exception boundary
#asahi · 2021-09-22 12:27
nullx86 has an unwinder that just dumps addrs on the stack, but for arm64 we always walk the chain of frame records
#asahi · 2021-09-22 12:25
null... since using DWARF is considered out of the question within the kernel
#asahi · 2021-09-22 12:25
null... this is something I'd like to fix, but we either need toolchain help (restricting codegen or generating metadata) or some binary parsing to genetate some metadata
#asahi · 2021-09-22 12:24
nullmarcan: for the stack trace thing (for linux at least), it's only possible to reliably unwind at function-call boundaries, and we over-estimate at exception boundaries (always reporting the LR, even if that's bogus). Note that the LR can be used as a scratch register transisetnly (e.g. for ADR LR <func>; BLR LR), so there's no 100% reliable way to know whether it's valid without knowledge of the binary
#asahi · 2021-06-03 16:04
nullSure; I'm just providing context, and I'd generally agree that it's preferable to get the hierarchy in the DT if this isn't really a discrete unit
#asahi · 2021-06-03 16:03
nullI haven't been a DT maintainer for a while now, but the self-reference doesn't seem right to me
#asahi · 2021-06-03 16:02
nullThe two ways this is generally done is either to have discrete nodes or to have the internal relationships between those be inferred from the compatible stirng in the driver