x86/mm/pat: Do a small optimization and fix in reserve_memtype

From: Pan Xinhui <[email protected]>

It's more reasonable to unlock memtype_lock right after rbt_memtype_check_insert. memtype_lock protects all data stored in rb-tree from multiple access. It's not cool to call kfree, pr_info, etc with this lock held. So move spin_unlock a little ahead.

If *new* succeed to be stored into the rb-tree, we might hit panic. Because we access *new* in dprintk "cattr_name(new->type)". Data stored in the rb-tree might be freed at any possbile time. It's abviously wrong to access such data without lock held. As new->type might be changed in rbt_memtype_check_insert, so save new->type to actual_type, then use actual_type in dprintk.

Signed-off-by Pan Xinhui <[email protected]>

---
change from v2:
update comments.
change from V1:
fix an access of *new* without memtype_lock held.

---
 arch/x86/mm/pat.c | 15 +++++++++------
 1 file changed, 9 insertions(+), 6 deletions(-)

arch/x86/mm/pat.c +0 -0

index 188e3e0..894a096 100644
@@ -538,22 +538,25 @@ int reserve_memtype(u64 start, u64 end, enum page_cache_mode req_type,
 	new->type	= actual_type;

arch/x86/mm/pat.c +3 -5

+	spin_unlock(&memtype_lock);
+
 	if (err) {
 		pr_info("x86/PAT: reserve_memtype failed [mem %#010Lx-%#010Lx], track %s, req %s\n",
 			start, end - 1,
 			cattr_name(new->type), cattr_name(req_type));
 		kfree(new);
-		spin_unlock(&memtype_lock);
-
 		return err;
 	}
 
-	spin_unlock(&memtype_lock);
-
 	dprintk("reserve_memtype added [mem %#010Lx-%#010Lx], track %s, req %s, ret %s\n",
-		start, end - 1, cattr_name(new->type), cattr_name(req_type),
+		start, end - 1, cattr_name(actual_type), cattr_name(req_type),
 		new_type ? cattr_name(*new_type) : "-");
 
 	return err;

Signature · 1 line

1.9.1

Quoted 15 lines · Pan Xinhui

From: Pan Xinhui <[email protected]>

It's more reasonable to unlock memtype_lock right after rbt_memtype_check_insert. memtype_lock protects all data stored in rb-tree from multiple access. It's not cool to call kfree, pr_info, etc with this lock held. So move spin_unlock a little ahead.

If *new* succeed to be stored into the rb-tree, we might hit panic. Because we access *new* in dprintk "cattr_name(new->type)". Data stored in the rb-tree might be freed at any possbile time. It's abviously wrong to access such data without lock held. As new->type might be changed in rbt_memtype_check_insert, so save new->type to actual_type, then use actual_type in dprintk.

Signed-off-by Pan Xinhui <[email protected]>

Quoted 8 lines

---
change from v2:
update comments.
change from V1:
fix an access of *new* without memtype_lock held.

---
 arch/x86/mm/pat.c | 15 +++++++++------
 1 file changed, 9 insertions(+), 6 deletions(-)

This patch still doesn't update the comments over memtype_lock.

Quoted 8 lines

arch/x86/mm/pat.c +0 -0

index 188e3e0..894a096 100644
@@ -538,22 +538,25 @@ int reserve_memtype(u64 start, u64 end, enum page_cache_mode req_type,
 	new->type	= actual_type;
 

arch/x86/mm/pat.c int reserve_memtype(u64 start, u64 end, enum page_cache_mode req_type, · line 547

Quoted 9 lines

arch/x86/mm/pat.c +6 -1

 	spin_lock(&memtype_lock);
-
 	err = rbt_memtype_check_insert(new, new_type);
+	/*
+	 * new->type might be changed in rbt_memtype_check_insert.
+	 * So save new->type to actual_type as dprintk uses it.
+	 * We are not allowed to touch new after unlocking memtype_lock.
+	 */
+	actual_type = new->type;

We already assign actual_type to new->type above. I think the dprintk needs actual_type and not what new->type has been changed to as that is in new_type.

Quoted 22 lines

arch/x86/mm/pat.c +3 -5

+	spin_unlock(&memtype_lock);
+
 	if (err) {
 		pr_info("x86/PAT: reserve_memtype failed [mem %#010Lx-%#010Lx], track %s, req %s\n",
 			start, end - 1,
 			cattr_name(new->type), cattr_name(req_type));
 		kfree(new);
-		spin_unlock(&memtype_lock);
-
 		return err;
 	}
 
-	spin_unlock(&memtype_lock);
-
 	dprintk("reserve_memtype added [mem %#010Lx-%#010Lx], track %s, req %s, ret %s\n",
-		start, end - 1, cattr_name(new->type), cattr_name(req_type),
+		start, end - 1, cattr_name(actual_type), cattr_name(req_type),
 		new_type ? cattr_name(*new_type) : "-");
 
 	return err;

Signature · 1 line

1.9.1

how about:
memtype_lock protects the rb-tree root and the rb-nodes which is a field of memtype from delete/add/lookup in a race.

Actually I have same questions. I find these output logs are added in
commit: 6997ab4982a29925e79f72c3a59823cf944c3529(x86: add PAT related debug prints) In the past, *new_type == actual_type == new->type on success. codes are below. author use actual_type there.

Quoted 2 lines · Pan Xinhui

how about:
memtype_lock protects the rb-tree root and the rb-nodes which is a field of memtype from delete/add/lookup in a race.

Use this:

"All pat_rbtree operations need to be performed while holding the memtype_lock."

Quoted 4 lines

Actually I have same questions. I find these output logs are added in
commit: 6997ab4982a29925e79f72c3a59823cf944c3529(x86: add PAT related debug prints) In the past, *new_type == actual_type == new->type on success. codes are below. author use actual_type there.

So this function is one bit PITA. So req_type is used to compute actual type a bit higher:

	actual_type = pat_x_mtrr_type(start, end, req_type);

and from then on actual_type is being used.

BUT!, in order to have *all* debugging information, the last dprintk()

call should dump actual_type and req_type because this way we show what

pat_x_mtrr_type() did too. And we don't need to dump new->type because this is the !err case and in that case we assigned new_type to it, which we dump already.

Ok?

Btw, you could also simplify this:

	if (is_range_ram == 1) {

		err = reserve_ram_pages_type(start, end, req_type, new_type);

		return err;
	}

to:

	if (is_range_ram == 1)
		return reserve_ram_pages_type(start, end, req_type, new_type);

while at it.

Thanks.