Re: buildworld: lib/libc: install: short write to libc.so.7.debug: [_libinstall] Error code 71

On Sun, 22 Nov 2020 at 16:06, Kyle Evans <kevans_at_freebsd.org> wrote:
>
> On Sun, Nov 22, 2020 at 6:00 AM Dimitry Andric <dim_at_freebsd.org> wrote:
> >
> > On 22 Nov 2020, at 08:06, Graham Perrin <grahamperrin_at_gmail.com> wrote:
> > >
> > > On 20/11/2020 09:57, Graham Perrin wrote:
> > >> On 16/11/2020 09:27, Graham Perrin wrote:
> > >>> Attempting to build r367615 on Friday 13th:
> > >>>
> > >>> …
> > >>>
> > >>> ===> lib/libprocstat/zfs (install)
> > >>> install -U  -C -o root -g wheel -m 444 /usr/src/lib/libprocstat/libprocstat.h /usr/obj/usr/src/amd64.amd64/tmp/usr/include/
> > >>> ===> lib/libc (obj,all,install)
> > >>> install -U  -C -o root -g wheel -m 444   libc.a /usr/obj/usr/src/amd64.amd64/tmp/usr/lib/
> > >>> install -U  -s -o root -g wheel -m 444   -S  libc.so.7 /usr/obj/usr/src/amd64.amd64/tmp/lib/
> > >>> install -U  -o root -g wheel -m 444    libc.so.7.debug /usr/obj/usr/src/amd64.amd64/tmp/usr/lib/debug/lib/
> > >>> install: short write to /usr/obj/usr/src/amd64.amd64/tmp/usr/lib/debug/lib/libc.so.7.debug: 393216 bytes written, 7462472 bytes asked to write
> > >>> *** [_libinstall] Error code 71
> > >>>
> > >>> make[4]: stopped in /usr/src/lib/libc
> > >>> 1 error
> > >>>
> > >>> make[4]: stopped in /usr/src/lib/libc
> > >>> root_at_mowa219-gjp4-8570p:/usr/src #
> > >>
> > >> The same problem – short write to /usr/obj/usr/src/amd64.amd64/tmp/usr/lib/debug/lib/libc.so.7.debug – when attempting buildworld of r367847.
> > >>
> > >> If it's relevant: I'm using r367081 for these attempts.
> > >
> > > The same problem when attempting buildworld of r367923. Prior to this I performed a fresh checkout of /usr/src
> > >
> > > What might cause these failures?
> > >
> > > If it's relevant: I have compression set to gzip-9 for the ZFS dataset that mounts at /usr
> >
> > I'd guess it's an unintended side-effect of
> > https://svnweb.freebsd.org/base?view=revision&revision=366697
> > ("install(1): Avoid unncessary fstatfs() calls and use mmap() based on
> > size").
> >
>
> Almost certainly -- before, we would never attempt to mmap() on ZFS.
>
> Tossing arichardson_at_ into CC as the committer
> Tossing mmacy_at_ and freqlabs_at_ into CC as ZFS folks
>
> Looking through sys/contrib/openzfs, there's special handling for mmap
> on linux because it bypasses the page cache and relies on caching in
> ARC. AFAICT the FreeBSD side seems to handle write() to mmap'd
> regions, but doesn't do anything with VOP_MMAPPED which might be
> needed to sync the file when it's mmap'd for reading like this. My
> understanding of how this is supposed to actually work on FreeBSD is
> limited, though, so I defer...
>
I'm quite surprised by this, I simply re-used the logic that bin/cp also has.
However, it's possible that this is no longer used in cp due to copy_file_range?
To be honest I'm not sure whether mmap() even improves performance
compared to read/write since the all the MMU setup might be quite
expensive and the data might not even be available so we end up
copying anyway. Maybe we should change install to use read/write
unconditionally? It might also make sense to factor out the
copy_file_range+read/write fallback logic in bin/cp to a library that
can be used in usr.bin/install.

Alex