About pmap_mapdev() & pmap_unmapdev()

[Konstantin Belousov]

On Sat, Oct 04, 2014 at 05:00:36PM +0900, Kohji Okuno wrote:
> Hi, Konstantin,
> 
> Thank you for your comment.
> And, your change is better than mine.
At the end of the mail is commit candidate.  I did not even compiled this.
Can you test and report back, please ?

> 
> > Do you mean that this panic is related to missed pmap_remove() ?
> > I doubt it, since pmap_mapdev() does not establish managed mappings.
> 
> Yes, pmap_mapdev() does not establish managed mappings. But, if
> kernel_pmap.pm_stats.resident_count is zero, then any managed pages
> (for example pipe_map, exec_map, or etc.) are not able to change
> unmanaged status, because pmap_remove() returns without calling
> pmap_remove_pte().
> 
> In this result, I encounterd the panic. Could you refer the following?
Yes, kmem_back() indeed uses managed mapping.

Index: amd64/amd64/pmap.c
===================================================================
--- amd64/amd64/pmap.c	(revision 272506)
+++ amd64/amd64/pmap.c	(working copy)
@@ -5040,6 +5040,9 @@ pmap_mapdev_attr(vm_paddr_t pa, vm_size_t size, in
 	pa = trunc_page(pa);
 	for (tmpsize = 0; tmpsize < size; tmpsize += PAGE_SIZE)
 		pmap_kenter_attr(va + tmpsize, pa + tmpsize, mode);
+	PMAP_LOCK(kernel_pmap);
+	kernel_pmap.pm_stats.resident_count += OFF_TO_IDX(size);
+	PMAP_UNLOCK(kernel_pmap);
 	pmap_invalidate_range(kernel_pmap, va, va + tmpsize);
 	pmap_invalidate_cache_range(va, va + tmpsize);
 	return ((void *)(va + offset));
Index: i386/i386/pmap.c
===================================================================
--- i386/i386/pmap.c	(revision 272506)
+++ i386/i386/pmap.c	(working copy)
@@ -5066,10 +5066,14 @@ pmap_mapdev_attr(vm_paddr_t pa, vm_size_t size, in
 	size = roundup(offset + size, PAGE_SIZE);
 	pa = pa & PG_FRAME;
 
-	if (pa < KERNLOAD && pa + size <= KERNLOAD)
+	if (pa < KERNLOAD && pa + size <= KERNLOAD) {
 		va = KERNBASE + pa;
-	else
+	} else {
 		va = kmem_alloc_nofault(kernel_map, size);
+		PMAP_LOCK(kernel_pmap);
+		kernel_pmap.pm_stats.resident_count += OFF_TO_IDX(size);
+		PMAP_UNLOCK(kernel_pmap);
+	}
 	if (!va)
 		panic("pmap_mapdev: Couldn't alloc kernel virtual memory");