Re: [PATCH 1/3] osdep: Add qemu_mkdir_with_parents()

[Peter Xu]

On Mon, Dec 16, 2024 at 04:56:33PM +0000, Peter Maydell wrote:
> On Mon, 16 Dec 2024 at 16:14, Peter Xu <peterx@redhat.com> wrote:
> >
> > QEMU uses g_mkdir_with_parents() a lot, especially in the case where the
> > failure case is ignored so an abort is expected when happened.
> >
> > Provide a helper qemu_mkdir_with_parents() to do that, and use it in the
> > two cases in qga/.  To be used in more places later.
> >
> > Signed-off-by: Peter Xu <peterx@redhat.com>
> > ---
> >  include/qemu/osdep.h     | 7 +++++++
> >  qga/commands-posix-ssh.c | 8 ++------
> >  util/osdep.c             | 6 ++++++
> >  3 files changed, 15 insertions(+), 6 deletions(-)
> >
> > diff --git a/include/qemu/osdep.h b/include/qemu/osdep.h
> > index fdff07fd99..dc67fb2e5e 100644
> > --- a/include/qemu/osdep.h
> > +++ b/include/qemu/osdep.h
> > @@ -828,6 +828,13 @@ static inline int 
> > platform_does_not_support_system(const char *command)
> >  }
> >  #endif /* !HAVE_SYSTEM_FUNCTION */
> >
> > +/**
> > + * qemu_mkdir_with_parents:
> > + *
> > + * Create directories with parents.  Abort on failures.
> > + */
> > +void qemu_mkdir_with_parents(const char *dir, int mode);
> 
> Don't put new function prototypes into osdep.h, please.
> It is included by every single C file in the codebase.
> There is always somewhere better to put things.
> 
> QEMU shouldn't abort on things that are kind of expected
> OS errors like "couldn't create a directory", so I'm
> a bit dubious about this function.

That's what qga/ is doing right now, rather than a decision made in this
series, though.

> 
> The two use cases in this commit seem to be test code,
> so asserting is reasonable. But a "for test code only"
> function should go in a header file that's only included
> by test cases and the comment should be clear about that,
> and it shouldn't have a function name that implies
> "this is the normal way any code in QEMU might want
> to create directories".
> 
> For the qtest tests, I currently ignore Coverity
> reports in our test code unless it seems particularly
> worthwhile to fix them. This is especially true for
> complaints about unchecked return values and the like.

OK.

> 
> Even in a test case it is still not great to call
> g_assert(), because this makes the test binary crash,
> rather than reporting an error. The surrounding TAP
> protocol parsing code then doesn't report the test
> failure the way you might like.

Hmm, I normally always think crash is better especially in tests to keep
everything around when an error happens as general rule.

TAP parsing especially on errors is more useful to me when we constantly
expect failures, IIUC that's not the case for QEMU tests?  Because I do
expect the CI to pass all the tests always.  But I also admit I don't know
the whole picture of QEMU tests..

If we don't care about retval checks in tests, we can definitely drop patch
1-2 here in all cases.

Thanks,

-- 
Peter Xu