Re: [PATCH] mm/vmalloc: add missing __GFP_HIGHMEM flag for vmalloc

[Matthew Wilcox]

On Mon, Jul 26, 2021 at 12:05:44PM +0800, Chen Wandun wrote:
> 在 2021/7/26 11:25, Matthew Wilcox 写道:
> > On Mon, Jul 26, 2021 at 11:23:33AM +0800, Chen Wandun wrote:
> > > struct page array can also be allocated in highmem during vmalloc,
> > > that will ease the low memory stress in 32bit system.
> > Huh?  Where does it get kmapped in order to access it?

(i misread your patch, apologies)

> The struct page array contain numbers of pointer of struct page, it is used to save
> 
> pages that allocated for vmalloc mapping in vmap_pages_range, it does't need to kmap.
> 
> 
> The main idea of this patch is come from:
> 
> https://lore.kernel.org/lkml/20170307141020.29107-1-mhocko@xxxxxxxxxx/

you've misunderstood that patch.  think about it:

static void *__vmalloc_area_node(struct vm_struct *area, gfp_t gfp_mask,
                                 pgprot_t prot, unsigned int page_shift,
                                 int node)
{
	const gfp_t nested_gfp = (gfp_mask & GFP_RECLAIM_MASK) | __GFP_ZERO;
(now nested_gfp has neither DMA nor DMA32 set)

        gfp_mask |= __GFP_NOWARN;
        if (!(gfp_mask & (GFP_DMA | GFP_DMA32)))
                gfp_mask |= __GFP_HIGHMEM;
(we set HIGHMEM if DMA and DMA32 are both not set)

                area->pages = __vmalloc_node(array_size, 1, nested_gfp, node,
                                        area->caller);
(we don't need to set HIGHMEM here because it will be set for us in
the second call to __vmalloc_area_node)