Re: [PATCH] vmalloc: Don't use flush flag when no exec perm

[Edgecombe, Rick P]

On Thu, 2019-05-30 at 10:44 +0300, Meelis Roos wrote:
> > > The addition of VM_FLUSH_RESET_PERMS for BPF JIT allocations was
> > > bisected to prevent boot on an UltraSparc III machine. It was
> > > found
> > > that
> > > sometime shortly after the TLB flush this flag does on vfree of
> > > the
> > > BPF
> > > program, the machine hung. Further investigation showed that
> > > before
> > > any of
> > > the changes for this flag were introduced, with
> > > CONFIG_DEBUG_PAGEALLOC
> > > configured (which does a similar TLB flush of the vmalloc range
> > > on
> > > every vfree), this machine also hung shortly after the first
> > > vmalloc
> > > unmap/free.
> > > 
> > > So the evidence points to there being some existing issue with
> > > the
> > > vmalloc TLB flushes, but it's still unknown exactly why these
> > > hangs
> > > are
> > > happening on sparc. It is also unknown when someone with this
> > > hardware
> > > could resolve this, and in the meantime using this flag on it
> > > turns a
> > > lurking behavior into something that prevents boot.
> > 
> > The sparc TLB flush issue has been bisected and is being worked on
> > now,
> > so hopefully we won't need this patch:
> > https://marc.info/?l=linux-sparc&m=155915694304118&w=2
> 
> And the sparc64 patch that fixes CONFIG_DEBUG_PAGEALLOC also fixes
> booting
> of the latest git kernel on Sun V445 where my problem initially
> happened.
> 
Thanks Meelis. So the TLB flush on this platform will be fixed and we
won't need this patch.