Re: [RFC PATCH v1] mm/vmalloc: fix incorrect __vmap_pages_range_noflush() if vm_area_alloc_pages() from high order fallback to order0

[Barry Song]

On Thu, Jul 25, 2024 at 8:02 AM Matthew Wilcox <willy@xxxxxxxxxxxxx> wrote:
>
> On Thu, Jul 25, 2024 at 02:28:27AM +0800, Hailong.Liu wrote:
> > >     if (!IS_ENABLED(CONFIG_HAVE_ARCH_HUGE_VMALLOC) ||
> > > -                   page_shift == PAGE_SHIFT)
> > > -           return vmap_small_pages_range_noflush(addr, end, prot, pages);
> > > +                   page_shift == PAGE_SHIFT ||
> > > +                   page_private(pages[0]) == VM_AREA_ALLOC_PAGES_FALLBACK) {
> > > +           int ret = vmap_small_pages_range_noflush(addr, end, prot, pages);
> > > +
> > > +           set_page_private(pages[0], 0);
> > > +           return ret;
> > > +   }
> > >
> > >     for (i = 0; i < nr; i += 1U << (page_shift - PAGE_SHIFT)) {
> > >             int err;
> > > @@ -3583,6 +3590,7 @@ vm_area_alloc_pages(gfp_t gfp, int nid,
> > >
> > >                     /* fall back to the zero order allocations */
> > >                     alloc_gfp |= __GFP_NOFAIL;
> > > +                   fallback = true;
> > Sry for my mistake, I forget define fallback here.
> > BTW, This is not the optimal solution. Does anyone have a better idea? Glad to
> > hear:)
>
> Yeah, I really don't like this approach.  You could return a small
> struct indicating both nr_allocated and whether you had to fall back.
> Or you could pass a bool * parameter.  They're both pretty nasty.

Yes, I feel returning a bool won't work very well.

the result could be a mixture of PMD and PTE if the allocated pages are
larger than a PMD.

For example, if we allocate 8MB, it might result in the first 4MB being
2* PMD, and the remaining 4MB being PTE order-0 pages.

I am also curious what will happen if we allocate 3MB(1PMD + some PTEs),
is the below doing the correct mapping?

        do {
                ret = vmap_pages_range(addr, addr + size, prot, area->pages,
                        page_shift);
                if (nofail && (ret < 0))
                        schedule_timeout_uninterruptible(1);
        } while (nofail && (ret < 0));

Is it possible we have only mapped the first 2MB if page_shift is PMD?

Thanks
Barry