Re: git: 969f6380eb66 - main - kdump: nicer printing of kill(2) PID argument
Date: Thu, 05 Jun 2025 17:59:14 UTC
On 6/5/25 11:54, Warner Losh wrote: > On Thu, Jun 5, 2025 at 6:14 AM John Baldwin <jhb@freebsd.org> wrote: >> >> On 6/4/25 19:03, Kyle Evans wrote: >>> On 6/4/25 17:55, Gleb Smirnoff wrote: >>>> On Wed, Jun 04, 2025 at 01:51:14AM +0000, Kyle Evans wrote: >>>> K> The branch main has been updated by kevans: >>>> K> >>>> K> URL: https://cgit.FreeBSD.org/src/commit/?id=969f6380eb66f809eed3e5c38b6021824a4cc2bf >>>> K> >>>> K> commit 969f6380eb66f809eed3e5c38b6021824a4cc2bf >>>> K> Author: Kyle Evans <kevans@FreeBSD.org> >>>> K> AuthorDate: 2025-06-04 01:51:06 +0000 >>>> K> Commit: Kyle Evans <kevans@FreeBSD.org> >>>> K> CommitDate: 2025-06-04 01:51:06 +0000 >>>> K> >>>> K> kdump: nicer printing of kill(2) PID argument >>>> K> >>>> K> Similar to wait*(), kill(2) operates on a pid that currently gets output >>>> K> as hex. Output it in decimal to make it a little easier to eyeball the >>>> K> pid we're signalling. >>>> K> >>>> K> Reviewed by: markj >>>> K> Differential Revision: https://reviews.freebsd.org/D50508 >>>> >>>> I didn't review if PIDs are always printed as decimals or not, but for >>>> the file descriptors it is a mix of hex and decimals. :( Usually I go >>>> with a sed script over kdump output to make it consistent. >>>> >>> >>> To be fair, I'd like to fix that, too- I noticed close() the other day >>> for fd > 0, but paused when I: >>> >>> 1.) couldn't tell where we even output close args >> >> close is probably handled by the default case where where all of the >> arguments are just output as hex values. Note that for kdump, most >> syscalls fall into this case including syscalls with pointer arguments. >> You'd probably be a bit sad with 64-bit pointers printed as decimal >> for many system calls. >> >> truss is different as truss stores some rudimentary type information about >> system call arguments and then defaults to printing certain types like >> file descriptors and ids as decimal. truss also prints NULL for null >> pointers IIRC. >> >> A useful project perhaps would be to move the table describing system >> call argument types out of syscalls.c in truss and into libsysdecode so >> that kdump could also reuse it. Probably the API you would want is >> something that returns an individual `struct syscall_decode` given >> (ABI, number) input arguments. > > Yes. All this generation is why we did the lua project to make the > master system call table parsing into a library. We need it also for > qemu-bsd-user long-term: writing new system calls by hand is a pain > and error-prone. > >> I don't think you can generate this table automatically from makesyscalls.lua >> as many of the "types" truss uses are synthetic types that aren't visible >> in C, e.g. "OpenFlags" meaning O_* flags passed to open(2). > > I think that we should, in the fullness of time, flag these so we can > do that. If we do add additional notations, we can not only have > better truss/trace output, as well as being able to generate the right > flag translations for running FreeBSD binaries on Linux (which people > are doing by hand right now...). > >> This would allow kdump's hand-written per-syscall-number rules to instead >> be closer to truss where you instead just iterate over types. It would >> also mean only having to update a single table in libsysdecode when adding >> a new system call to add both truss and kdump support. Only when a new >> argument type is added would one have to actually touch truss or kdump >> directly. > > I'd love for this to be generated as well... It was certainly one of > the use cases that we had in mind for the GSOC project that did this. > If we don't currently have a GSoC project idea around the above thoughts about truss -> libsysdecode and how we might be able to leverage makesyscalls and/or other annotations for it, maybe that'd be good to fish for in the 2026 program. I'm afraid I don't really have time for more than the sniping of random cases that make my life a tiny bit harder for the foreseeable future. Thanks, Kyle Evans