Re: Regression after commit 19809c2da28a ("mm, vmalloc: use __GFP_HIGHMEM implicitly")

[Michal Hocko]

[I am crawling over a large backlog after vacation so I will get to
 other emails in this thread later. Let's just fix the regression
 first. The patch with the full changelog is at the end of this email.
 CC Andrew - the original report is http://lkml.kernel.org/r/627DA40A-D0F6-41C1-BB5A-55830FBC9800@xxxxxxxxxxxxx]

On Sun 11-02-18 03:28:08, Matthew Wilcox wrote:
> On Sun, Feb 11, 2018 at 10:26:52AM +0100, Michal Hocko wrote:
> > On Thu 08-02-18 15:20:04, Matthew Wilcox wrote:
> > > ... nevertheless, 19809c2da28a does in fact break vmalloc_32 on 32-bit.  Look:
> > > 
> > > #if defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA32)
> > > #define GFP_VMALLOC32 GFP_DMA32 | GFP_KERNEL
> > > #elif defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA)
> > > #define GFP_VMALLOC32 GFP_DMA | GFP_KERNEL
> > > #else
> > > #define GFP_VMALLOC32 GFP_KERNEL
> > > #endif
> > > 
> > > So we pass in GFP_KERNEL to __vmalloc_node, which calls __vmalloc_node_range
> > > which calls __vmalloc_area_node, which ORs in __GFP_HIGHMEM.
> > 
> > Dohh. I have missed this. I was convinced that we always add GFP_DMA32
> > when doing vmalloc_32. Sorry about that. The above definition looks
> > quite weird to be honest. First of all do we have any 64b system without
> > both DMA and DMA32 zones? If yes, what is the actual semantic of
> > vmalloc_32? Or is there any magic forcing GFP_KERNEL into low 32b?
> 
> mmzone.h has the following, which may be inaccurate / out of date:
> 
>          * parisc, ia64, sparc  <4G
>          * s390                 <2G
>          * arm                  Various
>          * alpha                Unlimited or 0-16MB.
>          *
>          * i386, x86_64 and multiple other arches
>          *                      <16M.
> 
> It claims ZONE_DMA32 is x86-64 only, which is incorrect; it's now used
> by arm64, ia64, mips, powerpc, tile.

yes, nobody seem to keep this one in sync.

> > Also I would expect that __GFP_DMA32 should do the right thing on 32b
> > systems. So something like the below should do the trick
> 
> Oh, I see.  Because we have:
> 
> #ifdef CONFIG_ZONE_DMA32
> #define OPT_ZONE_DMA32 ZONE_DMA32
> #else
> #define OPT_ZONE_DMA32 ZONE_NORMAL
> #endif
> 
> we'll end up allocating from ZONE_NORMAL if a non-DMA32 architecture asks
> for GFP_DMA32 memory.  Thanks; I missed that.

yep

> I'd recommend this instead then:
> 
> #if defined(CONFIG_64BIT) && !defined(CONFIG_ZONE_DMA32)
> #define GFP_VMALLOC32 GFP_DMA | GFP_KERNEL
> #else
> #define GFP_VMALLOC32 GFP_DMA32 | GFP_KERNEL
> #endif
> 
> I think it's clearer than the three-way #if.

I do not have a strong opinion here. I just wanted the change to be
obvious without meddling with the 64b ifdefs much. Follow up cleanups
are certainly possible.

> Now, longer-term, perhaps we should do the following:
> 
> #ifdef CONFIG_ZONE_DMA32
> #define OPT_ZONE_DMA32	ZONE_DMA32
> #elif defined(CONFIG_64BIT)
> #define OPT_ZONE_DMA	OPT_ZONE_DMA
> #else
> #define OPT_ZONE_DMA32 ZONE_NORMAL
> #endif
> 
> Then we wouldn't need the ifdef here and could always use GFP_DMA32
> | GFP_KERNEL.  Would need to audit current users and make sure they
> wouldn't be broken by such a change.

I am pretty sure improvements are possible.

> I noticed a mistake in 704b862f9efd;
> 
> -               pages = __vmalloc_node(array_size, 1, nested_gfp|__GFP_HIGHMEM,
> +               pages = __vmalloc_node(array_size, 1, nested_gfp|highmem_mask,
> 
> We should unconditionally use __GFP_HIGHMEM here instead of highmem_mask
> because this is where we allocate the array to hold the struct page
> pointers.  This can be allocated from highmem, and does not need to be
> allocated from ZONE_NORMAL.

You seem to be right. nested_gfp doesn't include zone modifiers. Care to
send a patch?

> Similarly,
> 
> -               if (gfpflags_allow_blocking(gfp_mask))
> +               if (gfpflags_allow_blocking(gfp_mask|highmem_mask))
> 
> is not needed (it's not *wrong*, it was just an unnecessary change).

yes. highmem_mask has no influence on the blocking behavior.

The fix for the regressions should be