Re: [PATCH RFC 1/3] mm: When CONFIG_ZONE_DMA32 is set, use DMA32 for SLAB_CACHE_DMA

[Nicolas Boichat]

On Fri, Nov 9, 2018 at 6:43 PM Vlastimil Babka <vbabka@xxxxxxx> wrote:
>
> On 11/9/18 9:24 AM, Nicolas Boichat wrote:
> > Some callers, namely iommu/io-pgtable-arm-v7s, expect the physical
> > address returned by kmem_cache_alloc with GFP_DMA parameter to be
> > a 32-bit address.
> >
> > Instead of adding a separate SLAB_CACHE_DMA32 (and then audit
> > all the calls to check if they require memory from DMA or DMA32
> > zone), we simply allocate SLAB_CACHE_DMA cache in DMA32 region,
> > if CONFIG_ZONE_DMA32 is set.
> >
> > Fixes: ad67f5a6545f ("arm64: replace ZONE_DMA with ZONE_DMA32")
> > Signed-off-by: Nicolas Boichat <drinkcat@xxxxxxxxxxxx>
> > ---
> >  include/linux/slab.h | 13 ++++++++++++-
> >  mm/slab.c            |  2 +-
> >  mm/slub.c            |  2 +-
> >  3 files changed, 14 insertions(+), 3 deletions(-)
> >
> > diff --git a/include/linux/slab.h b/include/linux/slab.h
> > index 918f374e7156f4..390afe90c5dec0 100644
> > --- a/include/linux/slab.h
> > +++ b/include/linux/slab.h
> > @@ -30,7 +30,7 @@
> >  #define SLAB_POISON          ((slab_flags_t __force)0x00000800U)
> >  /* Align objs on cache lines */
> >  #define SLAB_HWCACHE_ALIGN   ((slab_flags_t __force)0x00002000U)
> > -/* Use GFP_DMA memory */
> > +/* Use GFP_DMA or GFP_DMA32 memory */
> >  #define SLAB_CACHE_DMA               ((slab_flags_t __force)0x00004000U)
> >  /* DEBUG: Store the last owner for bug hunting */
> >  #define SLAB_STORE_USER              ((slab_flags_t __force)0x00010000U)
> > @@ -126,6 +126,17 @@
> >  #define ZERO_OR_NULL_PTR(x) ((unsigned long)(x) <= \
> >                               (unsigned long)ZERO_SIZE_PTR)
> >
> > +/*
> > + * When ZONE_DMA32 is defined, have SLAB_CACHE_DMA allocate memory with
> > + * GFP_DMA32 instead of GFP_DMA, as this is what some of the callers
> > + * require (instead of duplicating cache for DMA and DMA32 zones).
> > + */
> > +#ifdef CONFIG_ZONE_DMA32
> > +#define SLAB_CACHE_DMA_GFP GFP_DMA32
> > +#else
> > +#define SLAB_CACHE_DMA_GFP GFP_DMA
> > +#endif
>
> AFAICS this will break e.g. x86 which can have both ZONE_DMA and
> ZONE_DMA32, and now you would make kmalloc(__GFP_DMA) return objects
> from ZONE_DMA32 instead of __ZONE_DMA, which can break something.

Oh, I was not aware that both ZONE_DMA and ZONE_DMA32 can be defined
at the same time. I guess the test should be inverted, something like
this (can be simplified...):
#ifdef CONFIG_ZONE_DMA
#define SLAB_CACHE_DMA_GFP GFP_DMA
#elif defined(CONFIG_ZONE_DMA32)
#define SLAB_CACHE_DMA_GFP GFP_DMA32
#else
#define SLAB_CACHE_DMA_GFP GFP_DMA // ?
#endif

> Also I'm probably missing the point of this all. In patch 3 you use
> __get_dma32_pages() thus __get_free_pages(__GFP_DMA32), which uses
> alloc_pages, thus the page allocator directly, and there's no slab
> caches involved.

__get_dma32_pages fixes level 1 page allocations in the patch 3.

This change fixes level 2 page allocations
(kmem_cache_zalloc(data->l2_tables, gfp | GFP_DMA)), by transparently
remapping GFP_DMA to an underlying ZONE_DMA32.

The alternative would be to create a new SLAB_CACHE_DMA32 when
CONFIG_ZONE_DMA32 is defined, but then I'm concerned that the callers
would need to choose between the 2 (GFP_DMA or GFP_DMA32...), and also
need to use some ifdefs (but maybe that's not a valid concern?).

> It makes little sense to involve slab for page table
> allocations anyway, as those tend to be aligned to a page size (or
> high-order page size). So what am I missing?

Level 2 tables are ARM_V7S_TABLE_SIZE(2) => 1kb, so we'd waste 3kb if
we allocated a full page.

Thanks,