On Wed, Mar 04, 2020 at 10:01:33AM +0800, Ian Kent wrote:
> On Tue, 2020-03-03 at 14:03 +0100, Greg Kroah-Hartman wrote:
> > Actually, I like this idea (the syscall, not just the unlimited
> > beers).
> > Maybe this could make a lot of sense, I'll write some actual tests
> > for
> > it now that syscalls are getting "heavy" again due to CPU vendors
> > finally paying the price for their madness...
>
> The problem isn't with open->read->close but with the mount info.
> changing between reads (ie. seq file read takes and drops the
> needed lock between reads at least once).
readfile() is not reaction to mountinfo.
The motivation is that we have many places with trivial
open->read->close for very small text files due to /sys and /proc. The
current way how kernel delivers these small strings to userspace seems
pretty inefficient if we can do the same by one syscall.
Karel
$ strace -e openat,read,close -c ps aux
...
% time seconds usecs/call calls errors syscall
------ ----------- ----------- --------- --------- ----------------
43.32 0.004190 4 987 read
31.42 0.003039 3 844 4 openat
25.26 0.002443 2 842 close
------ ----------- ----------- --------- --------- ----------------
100.00 0.009672 2673 4 total
$ strace -e openat,read,close -c lsns
...
% time seconds usecs/call calls errors syscall
------ ----------- ----------- --------- --------- ----------------
39.95 0.001567 2 593 openat
30.93 0.001213 2 597 close
29.12 0.001142 3 365 read
------ ----------- ----------- --------- --------- ----------------
100.00 0.003922 1555 total
$ strace -e openat,read,close -c lscpu
...
% time seconds usecs/call calls errors syscall
------ ----------- ----------- --------- --------- ----------------
44.67 0.001480 7 189 52 openat
34.77 0.001152 6 180 read
20.56 0.000681 4 140 close
------ ----------- ----------- --------- --------- ----------------
100.00 0.003313 509 52 total
--
Karel Zak <kzak@xxxxxxxxxx>
http://karelzak.blogspot.com
