Re: [PATCH RFC 0/2] KVM: s390: avoid having to enable vm.alloc_pgste

[Martin Schwidefsky <schwidefsky@de.ibm.com>]

On Thu, 8 Jun 2017 08:25:31 +0200
Heiko Carstens <heiko.carstens@de.ibm.com> wrote:

> On Thu, Jun 08, 2017 at 07:35:28AM +0200, Martin Schwidefsky wrote:
> > On Wed, 7 Jun 2017 22:47:56 +0200
> > Heiko Carstens <heiko.carstens@de.ibm.com> wrote:  
> > > On Wed, Jun 07, 2017 at 02:34:40PM +0200, Martin Schwidefsky wrote:  
> > > > +#define arch_elf_pt_proc(ehdr, phdr, elf, interp, state)	\
> > > > +({								\
> > > > +	struct elf64_hdr *_ehdr = (void *) ehdr;		\
> > > > +	struct elf64_phdr *_phdr = (void *) phdr;		\
> > > > +	int _rc = 0;						\
> > > > +	if (_ehdr->e_ident[EI_CLASS] == ELFCLASS64 &&		\
> > > > +	    _phdr->p_type == PT_S390_REQUEST_PGSTE &&		\
> > > > +	    !page_table_allocate_pgste &&			\
> > > > +	    !test_thread_flag(TIF_REQUEST_PGSTE)) {		\
> > > > +		set_thread_flag(TIF_REQUEST_PGSTE);		\
> > > > +		set_pt_regs_flag(task_pt_regs(current),		\
> > > > +				 PIF_SYSCALL_RESTART);		\
> > > > +		_rc = -EAGAIN;					\
> > > > +	}							\
> > > > +	_rc;							\
> > > > +})    
> > > 
> > > I'm wondering if this should simply fail, if a PT_S390_REQUEST_PGSTE type
> > > segment exists, but it is not ELFCLASS64?
> > > It will fail later anyway on s390_enable_sie(), but...  
> > 
> > Does it matter if it fails for a 32-bit ELF file? Just makes the code more
> > complex without benefit, no?  
> 
> It would be more consistent, since right now a 32-bit ELF file with
> PT_S390_REQUEST_PGSTE will be exectuted, but the page tables won't have any
> pgstes. That's sort of odd, isn't it? And that later on it won't be able to
> create a virtual machine because our current implementation doesn't allow
> that for compat tasks is sort of unrelated.
> But anyway, I'll leave that up to you, it doesn't really matter.

Actually the code will be less complex if we add PT_S390_REQUEST_PGSTE for
32-bit ELF files as well. It does not make sense to define the segment for
a compat process as KVM won't work but you get what you ask for..

This looks like this:

#define arch_elf_pt_proc(ehdr, phdr, elf, interp, state)        \
({                                                              \
        int _rc = 0;                                            \
        if (phdr->p_type == PT_S390_REQUEST_PGSTE &&            \
            !page_table_allocate_pgste &&                       \
            !test_thread_flag(TIF_REQUEST_PGSTE)) {             \
                set_thread_flag(TIF_REQUEST_PGSTE);             \
                set_pt_regs_flag(task_pt_regs(current),         \
                                 PIF_SYSCALL_RESTART);          \
                _rc = -EAGAIN;                                  \
        }                                                       \
        _rc;                                                    \
})

phdr is a (struct elf_phd *) which is either define to a a (struct elf64_phdr *)
or a (struct elf32_phdr *). The check works in both cases.
 
> >   
> > > > diff --git a/arch/s390/include/asm/mmu_context.h b/arch/s390/include/asm/mmu_context.h
> > > > index c119d564d8f2..1201b18e817d 100644
> > > > --- a/arch/s390/include/asm/mmu_context.h
> > > > +++ b/arch/s390/include/asm/mmu_context.h
> > > > @@ -25,7 +25,8 @@ static inline int init_new_context(struct task_struct *tsk,
> > > >  	mm->context.gmap_asce = 0;
> > > >  	mm->context.flush_mm = 0;
> > > >  #ifdef CONFIG_PGSTE
> > > > -	mm->context.alloc_pgste = page_table_allocate_pgste;
> > > > +	mm->context.alloc_pgste = page_table_allocate_pgste ||
> > > > +		test_thread_flag(TIF_REQUEST_PGSTE);    
> > > 
> > > I think the alloc_pgste flag should be inherited on fork, no?  
> > 
> > Yes, that makes it more consistent. I'll add it.  
> 
> By the way, what prevents with the _current_ code a scenario like:
> 
> - set allocate_pgste sysctl to 1
> - create kvm guest
> - s390_enable_sie
> - run vcpu
> - set allocate_pgste sysctl to 0
> - clone(... CLONE_FILES ...) (that is: new mm without pgstes, but shared fds)
> - [child] run vcpu
> 
> Is there anything that makes sure we cannot execute the sie instruction in
> the child process?

Yes, that looks like a loop-hole.

-- 
blue skies,
   Martin.

"Reality continues to ruin my life." - Calvin.