Re: [PATCH v3 3/7] mm/vmalloc.c: allow vread() to read out vm_map_ram areas

[Baoquan He <bhe@redhat.com>]

On 01/16/23 at 04:08pm, Dan Carpenter wrote:
> On Sun, Jan 15, 2023 at 10:08:55PM +0800, Baoquan He wrote:
> > > f181234a5a21fd0 Chen Wandun             2021-09-02  3650  	if ((unsigned long)addr + count <= va->va_start)
> > > f181234a5a21fd0 Chen Wandun             2021-09-02  3651  		goto finished;
> > > f181234a5a21fd0 Chen Wandun             2021-09-02  3652  
> > > f608788cd2d6cae Serapheim Dimitropoulos 2021-04-29  3653  	list_for_each_entry_from(va, &vmap_area_list, list) {
> > > e81ce85f960c2e2 Joonsoo Kim             2013-04-29  3654  		if (!count)
> > > e81ce85f960c2e2 Joonsoo Kim             2013-04-29  3655  			break;
> > > e81ce85f960c2e2 Joonsoo Kim             2013-04-29  3656  
> > > 129dbdf298d7383 Baoquan He              2023-01-13  3657  		vm = va->vm;
> > > 129dbdf298d7383 Baoquan He              2023-01-13  3658  		flags = va->flags & VMAP_FLAGS_MASK;
> > > 129dbdf298d7383 Baoquan He              2023-01-13  3659  
> > > 129dbdf298d7383 Baoquan He              2023-01-13  3660  		if (!vm && !flags)
> > >                                                                             ^^^
> > > vm can be NULL if a flag in VMAP_FLAGS_MASK is set.
> > > 
> > > e81ce85f960c2e2 Joonsoo Kim             2013-04-29  3661  			continue;
> > 
> > Right, after the 'continue;' line, only two cases could happen when it
> > comes here. (vm != null) or (vm->flags & VMAP_RAM) is true.
> >
> 
> You're saying VMAP_RAM, but strictly speaking the code is checking
> VMAP_FLAGS_MASK and not VMAP_RAM.
> 
> +#define VMAP_RAM               0x1 /* indicates vm_map_ram area*/
> +#define VMAP_BLOCK             0x2 /* mark out the vmap_block sub-type*/
> +#define VMAP_FLAGS_MASK                0x3
> 
> If we assume that vm is NULL, VMAP_BLOCK is set and VMAP_RAM is clear
> then it would lead to a NULL dereference.  There might be reasons why
> that combination is impossible outside the function but we can't tell
> from the information we have here.

VMAP_BLOCK has no chance to be set alone. It has to be set together with
VMAP_RAM if needed.

> 
> Which is fine, outside information is a common reason for false
> positives with this check.  But I was just concerned about the mix of
> VMAP_FLAGS_MASK and VMAP_RAM.

Thanks, I see your point now, will consider how to improve it.