On Fri, Mar 25, 2022 at 09:54:21AM +0100, Greg KH wrote:
> On Fri, Mar 25, 2022 at 09:46:46AM +0100, Karel Zak wrote:
> > On Thu, Mar 24, 2022 at 09:44:38AM +0100, Miklos Szeredi wrote:
> > > > If so, have you benchmarked lsof using this new interface?
> > >
> > > Not yet. Looked yesterday at both lsof and procps source code, and
> > > both are pretty complex and not easy to plug in a new interface. But
> > > I've not yet given up...
> >
> > I can imagine something like getvalues(2) in lsblk (based on /sys) or
> > in lsfd (based on /proc; lsof replacement). The tools have defined set
> > of information to read from kernel, so gather all the requests to the
> > one syscall for each process or block device makes sense and it will
> > dramatically reduce number of open+read+close syscalls.
>
> And do those open+read+close syscalls actually show up in measurements?
>
> Again, I tried to find a real-world application that turning those 3
> into 1 would matter, and I couldn't. procps had no decreased system
> load that I could notice. I'll mess with lsof but that's really just a
> stress-test, not anything that is run all the time, right?
Right, the speed of ps(1) or lsof(1) is not important. IMHO the current
discussion about getvalues() goes in wrong direction :-)
I guess the primary motivation is not to replace open+read+close, but
provide to userspace something usable to get information from mount
table, because the current /proc/#/mountinfo and notification by
poll() is horrible.
Don't forget that the previous attempt was fsinfo() from David Howells
(unfortunately, it was too complex and rejected by Linus).
> And as others have said, using io_uring() would also solve the 3 syscall
> issue, but no one seems to want to convert these tools to use that,
> which implies that it's not really an issue for anyone :)
OK, I'll think about it :-)
Karel
--
Karel Zak <kzak@xxxxxxxxxx>
http://karelzak.blogspot.com
