Re: [PATCH v3 1/4] accel/tcg: Invalidate translations when clearing PAGE_EXEC

[Ilya Leoshkevich]

On Wed, 2022-08-10 at 13:29 -0700, Richard Henderson wrote:
> On 8/8/22 10:10, Ilya Leoshkevich wrote:
> > After mprotect(addr, PROT_NONE), addr can still be executed if
> > there
> > are cached translations. Drop them.
> > 
> > Signed-off-by: Ilya Leoshkevich <iii@linux.ibm.com>
> > ---
> >   accel/tcg/translate-all.c | 17 ++++++++++++-----
> >   1 file changed, 12 insertions(+), 5 deletions(-)
> > 
> > diff --git a/accel/tcg/translate-all.c b/accel/tcg/translate-all.c
> > index ef62a199c7..32ea5f0adf 100644
> > --- a/accel/tcg/translate-all.c
> > +++ b/accel/tcg/translate-all.c
> > @@ -2295,12 +2295,19 @@ void page_set_flags(target_ulong start,
> > target_ulong end, int flags)
> >            len != 0;
> >            len -= TARGET_PAGE_SIZE, addr += TARGET_PAGE_SIZE) {
> >           PageDesc *p = page_find_alloc(addr >> TARGET_PAGE_BITS,
> > 1);
> > +        bool write_set, exec_cleared;
> >   
> > -        /* If the write protection bit is set, then we invalidate
> > -           the code inside.  */
> > -        if (!(p->flags & PAGE_WRITE) &&
> > -            (flags & PAGE_WRITE) &&
> > -            p->first_tb) {
> > +        /*
> > +         * If the write protection bit is set, then we invalidate
> > the code
> > +         * inside.
> > +         */
> > +        write_set = !(p->flags & PAGE_WRITE) && (flags &
> > PAGE_WRITE);
> > +        /*
> > +         * If PAGE_EXEC is cleared, we also need to invalidate the
> > code in
> > +         * order to force a fault when trying to run it.
> > +         */
> > +        exec_cleared = (p->flags & PAGE_EXEC) && !(flags &
> > PAGE_EXEC);
> > +        if ((write_set || exec_cleared) && p->first_tb) {
> 
> I believe the bug you're trying to fix is in get_page_addr_code,
> which for USER_ONLY is 
> currently a no-op.  It ought to be checking the page permissions
> there, as we do for softmmu.
> 
> I have a patch for get_page_addr_code in the works, because I was
> working on pther stuff 
> in the area.

How is qemu-user's get_page_addr_code() involved here?

I tried to experiment with it, and while I agree that it looks buggy,
it's called only from translation code paths. If we already have a
translation block, these code paths are not used.