Re: [PATCH 0/4] cleancache: remove limit on the number of cleancache enabled filesystems

[Vladimir Davydov <vdavydov@parallels.com>]

On Mon, Feb 23, 2015 at 11:12:22AM -0500, Konrad Rzeszutek Wilk wrote:
> Thank you for posting these patches. I was wondering if you had
> run through some of the different combinations that you can
> load the filesystems/tmem drivers in random order? The #4 patch
> deleted a nice chunk of documentation that outlines the different
> combinations.

Yeah, I admit the synchronization between cleancache_register_ops and
cleancache_init_fs is far not obvious. I should have updated the comment
instead of merely dropping it, sorry. What about the following patch
proving correctness of register_ops-vs-init_fs synchronization? It is
meant to be applied incrementally on top of patch #4.
---
diff --git a/mm/cleancache.c b/mm/cleancache.c
index fbdaf9c77d7a..8fc50811119b 100644
--- a/mm/cleancache.c
+++ b/mm/cleancache.c
@@ -54,6 +54,57 @@ int cleancache_register_ops(struct cleancache_ops *ops)
 	if (cmpxchg(&cleancache_ops, NULL, ops))
 		return -EBUSY;
 
+	/*
+	 * A cleancache backend can be built as a module and hence loaded after
+	 * a cleancache enabled filesystem has called cleancache_init_fs. To
+	 * handle such a scenario, here we call ->init_fs or ->init_shared_fs
+	 * for each active super block. To differentiate between local and
+	 * shared filesystems, we temporarily initialize sb->cleancache_poolid
+	 * to CLEANCACHE_NO_BACKEND or CLEANCACHE_NO_BACKEND_SHARED
+	 * respectively in case there is no backend registered at the time
+	 * cleancache_init_fs or cleancache_init_shared_fs is called.
+	 *
+	 * Since filesystems can be mounted concurrently with cleancache
+	 * backend registration, we have to be careful to guarantee that all
+	 * cleancache enabled filesystems that has been mounted by the time
+	 * cleancache_register_ops is called has got and all mounted later will
+	 * get cleancache_poolid. This is assured by the following statements
+	 * tied together:
+	 *
+	 * a) iterate_supers skips only those super blocks that has started
+	 *    ->kill_sb
+	 *
+	 * b) if iterate_supers encounters a super block that has not finished
+	 *    ->mount yet, it waits until it is finished
+	 *
+	 * c) cleancache_init_fs is called from ->mount and
+	 *    cleancache_invalidate_fs is called from ->kill_sb
+	 *
+	 * d) we call iterate_supers after cleancache_ops has been set
+	 *
+	 * From a) it follows that if iterate_supers skips a super block, then
+	 * either the super block is already dead, in which case we do not need
+	 * to bother initializing cleancache for it, or it was mounted after we
+	 * initiated iterate_supers. In the latter case, it must have seen
+	 * cleancache_ops set according to d) and initialized cleancache from
+	 * ->mount by itself according to c). This proves that we call
+	 * ->init_fs at least once for each active super block.
+	 *
+	 * From b) and c) it follows that if iterate_supers encounters a super
+	 * block that has already started ->init_fs, it will wait until ->mount
+	 * and hence ->init_fs has finished, then check cleancache_poolid, see
+	 * that it has already been set and therefore do nothing. This proves
+	 * that we call ->init_fs no more than once for each super block.
+	 *
+	 * Combined together, the last two paragraphs prove the function
+	 * correctness.
+	 *
+	 * Note that various cleancache callbacks may proceed before this
+	 * function is called or even concurrently with it, but since
+	 * CLEANCACHE_NO_BACKEND is negative, they will all result in a noop
+	 * until the corresponding ->init_fs has been actually called and
+	 * cleancache_ops has been set.
+	 */
 	iterate_supers(cleancache_register_ops_sb, NULL);
 	return 0;
 }

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>