Re: [PATCH] mm: simplify the VM_BUG_ON condition in pmdp_huge_clear_flush()

From: Andrew Morton
Date: Mon Feb 01 2021 - 18:34:43 EST


On Mon, 1 Feb 2021 06:43:19 -0500 Miaohe Lin <linmiaohe@xxxxxxxxxx> wrote:

> The condition (A && !C && !D) || !A is equivalent to !A || (A && !C && !D)
> and can be further simplified to !A || (!C && !D).
>
> ..
>
> --- a/mm/pgtable-generic.c
> +++ b/mm/pgtable-generic.c
> @@ -135,8 +135,8 @@ pmd_t pmdp_huge_clear_flush(struct vm_area_struct *vma, unsigned long address,
> {
> pmd_t pmd;
> VM_BUG_ON(address & ~HPAGE_PMD_MASK);
> - VM_BUG_ON((pmd_present(*pmdp) && !pmd_trans_huge(*pmdp) &&
> - !pmd_devmap(*pmdp)) || !pmd_present(*pmdp));
> + VM_BUG_ON(!pmd_present(*pmdp) || (!pmd_trans_huge(*pmdp) &&
> + !pmd_devmap(*pmdp)));
> pmd = pmdp_huge_get_and_clear(vma->vm_mm, address, pmdp);
> flush_pmd_tlb_range(vma, address, address + HPAGE_PMD_SIZE);
> return pmd;

True, and the resulting code is still readable enough.

But a problem with such a complex expression is that the developer will
have trouble figuring out why the BUG actually triggered.

If we had a VM_BUG_ON_PMD() then we could print the pmd's value and
permit diagnosis from that. But we don't have such a thing.

So I suggest that it would be better to have

VM_BUG_ON((pmd_present(*pmdp) && !pmd_trans_huge(*pmdp) &&
!pmd_devmap(*pmdp)));
VM_BUG_ON(!pmd_present(*pmdp));

This way, the BUG()'s file-n-line output will tell us more about why the
kernel went splat.


I suppose maybe this could be optimized the same way, as

VM_BUG_ON(!pmd_present(*pmdp));
/* Below assumes pmd_present() is true */
VM_BUG_ON(!pmd_trans_huge(*pmdp) && !pmd_devmap(*pmdp));

Which works because VM_BUG_ON is, depending up Kconfig, either a no-op
or a noreturn-if-it-triggered. I'm not sure if I like this trick much though.