Re: [PATCH 4/5] mremap04: Convert to new API

Hi!
> > > +	 * Pass old_size of one page while the segment is two pages: the
> > > +	 * mismatch is intentional and must be preserved.
> > >  	 */
> > 
> > Wouldn't this be enough for the mremap() to fail even with regular
> > mmap()? I do not see what the test needs the sysv shm for.
> 
> actually yes, but I think the original idea was to mix SysV with mremap()
> syscall.

I would keep this test simple.

And we can add another test with a gap as well. If we map 3 pages, then
unmap page in the middle we will get:

| page 0 | hole | page 2 |

Now remap without MAYMOVE on page0 that spans into page2 will fail too.

> > Also the interaction between SysV shm and mremap() seems to be more
> > complicated and may warrant more tests.
> 
> mmmh, any idea?

Looking stack overflow mremap() should be strictly local to the process.
That means that that SysV mapping or MAP_SHARED mappings shouldn't
change the mapping length in other processes sharing the mapping.
Strangely there does not seem to be much documentation for this. So it's
a corner case worth exploring (raised more than 10 years ago
https://lwn.net/Articles/591985/).

Also while reading the manual I found another possible test for mremap()
+ userfaultfd(). You should get a fault event for expanded mappings if
you access the expanded region.

-- 
Cyril Hrubis
[email protected]

-- 
Mailing list info: https://lists.linux.it/listinfo/ltp