Re: [PATCH] mm/cma: cma_declare_contiguous: correct err handling

[Catalin Marinas <catalin.marinas@arm.com>]

On Tue, Feb 19, 2019 at 07:46:11PM +0200, Mike Rapoport wrote:
> On Tue, Feb 19, 2019 at 05:55:33PM +0100, Vlastimil Babka wrote:
> > On 2/14/19 9:38 PM, Andrew Morton wrote:
> > > On Thu, 14 Feb 2019 12:45:51 +0000 Peng Fan <peng.fan@nxp.com> wrote:
> > > 
> > >> In case cma_init_reserved_mem failed, need to free the memblock allocated
> > >> by memblock_reserve or memblock_alloc_range.
> > >>
> > >> ...
> > >>
> > >> --- a/mm/cma.c
> > >> +++ b/mm/cma.c
> > >> @@ -353,12 +353,14 @@ int __init cma_declare_contiguous(phys_addr_t base,
> > >>  
> > >>  	ret = cma_init_reserved_mem(base, size, order_per_bit, name, res_cma);
> > >>  	if (ret)
> > >> -		goto err;
> > >> +		goto free_mem;
> > >>  
> > >>  	pr_info("Reserved %ld MiB at %pa\n", (unsigned long)size / SZ_1M,
> > >>  		&base);
> > >>  	return 0;
> > >>  
> > >> +free_mem:
> > >> +	memblock_free(base, size);
> > >>  err:
> > >>  	pr_err("Failed to reserve %ld MiB\n", (unsigned long)size / SZ_1M);
> > >>  	return ret;
> > > 
> > > This doesn't look right to me.  In the `fixed==true' case we didn't
> > > actually allocate anything and in the `fixed==false' case, the
> > > allocated memory is at `addr', not at `base'.
> > 
> > I think it's ok as the fixed==true path has "memblock_reserve()", but
> > better leave this to the memblock maintainer :)
> 
> As Peng Fan noted in the other e-mail, fixed==true has memblock_reserve()
> and fixed==false resets base = addr, so this is Ok.
>  
> > There's also 'kmemleak_ignore_phys(addr)' which should probably be
> > undone (or not called at all) in the failure case. But it seems to be
> > missing from the fixed==true path?
> 
> Well, memblock and kmemleak interaction does not seem to have clear
> semantics anyway. memblock_free() calls kmemleak_free_part_phys() which
> does not seem to care about ignored objects.
> As for the fixed==true path, memblock_reserve() does not register the area
> with kmemleak, so there would be no object to free in memblock_free().
> AFAIU, kmemleak simply ignores this.

Kmemleak is supposed to work with the memblock_{alloc,free} pair and it
ignores the memblock_reserve() as a memblock_alloc() implementation
detail. It is, however, tolerant to memblock_free() being called on a
sub-range or just a different range from a previous memblock_alloc(). So
the original patch looks fine to me. FWIW:

Reviewed-by: Catalin Marinas <catalin.marinas@arm.com>