On Tue, 2018-05-15 at 16:05 +0200, Joerg Roedel wrote:

On Mon, Apr 30, 2018 at 11:59:24AM -0600, Toshi Kani wrote:

quoted

 int pud_free_pmd_page(pud_t *pud, unsigned long addr)
 {
-	pmd_t *pmd;
+	pmd_t *pmd, *pmd_sv;
+	pte_t *pte;
 	int i;
 
 	if (pud_none(*pud))
 		return 1;
 
 	pmd = (pmd_t *)pud_page_vaddr(*pud);
+	pmd_sv = (pmd_t *)__get_free_page(GFP_KERNEL);

So you need to allocate a page to free a page? It is better to put the
pages into a list with a list_head on the stack.

The code should have checked if pmd_sv is NULL...  I will update the
patch.

For performance, I do not think this page alloc is a problem.  Unlike
pmd_free_pte_page(), pud_free_pmd_page() covers an extremely rare case. 
  Since pud requires 1GB-alignment, pud and pmd/pte mappings do not
share the same ranges within the vmalloc space.  I had to instrument the
kernel to force them share the same ranges in order to test this patch.

I am still on favour of just reverting the broken commit and do a
correct and working fix for the/a merge window.

I will reorder the patch series, and change patch 3/3 to 1/3 so that we
can take it first to fix the BUG_ON on PAE.  This revert will disable
2MB ioremap on PAE in some cases, but I do not think it's important on
PAE anyway.

I do not think revert on x86/64 is necessary and I am more worried about
disabling 2MB ioremap in some cases, which can be seen as degradation. 
Patch 2/3 fixes a possible page-directory cache issue that I cannot hit
even though I put ioremap/iounmap with various sizes into a tight loop
for a day.

Thanks,
-Toshi