]> bbs.cooldavid.org Git - net-next-2.6.git/commitdiff
FS-Cache: Fix lock misorder in fscache_write_op()
authorDavid Howells <dhowells@redhat.com>
Thu, 19 Nov 2009 18:11:25 +0000 (18:11 +0000)
committerDavid Howells <dhowells@redhat.com>
Thu, 19 Nov 2009 18:11:25 +0000 (18:11 +0000)
FS-Cache has two structs internally for keeping track of the internal state of
a cached file: the fscache_cookie struct, which represents the netfs's state,
and fscache_object struct, which represents the cache's state.  Each has a
pointer that points to the other (when both are in existence), and each has a
spinlock for pointer maintenance.

Since netfs operations approach these structures from the cookie side, they get
the cookie lock first, then the object lock.  Cache operations, on the other
hand, approach from the object side, and get the object lock first.  It is not
then permitted for a cache operation to get the cookie lock whilst it is
holding the object lock lest deadlock occur; instead, it must do one of two
things:

 (1) increment the cookie usage counter, drop the object lock and then get both
     locks in order, or

 (2) simply hold the object lock as certain parts of the cookie may not be
     altered whilst the object lock is held.

It is also not permitted to follow either pointer without holding the lock at
the end you start with.  To break the pointers between the cookie and the
object, both locks must be held.

fscache_write_op(), however, violates the locking rules: It attempts to get the
cookie lock without (a) checking that the cookie pointer is a valid pointer,
and (b) holding the object lock to protect the cookie pointer whilst it follows
it.  This is so that it can access the pending page store tree without
interference from __fscache_write_page().

This is fixed by splitting the cookie lock, such that the page store tracking
tree is protected by its own lock, and checking that the cookie pointer is
non-NULL before we attempt to follow it whilst holding the object lock.

The new lock is subordinate to both the cookie lock and the object lock, and so
should be taken after those.

Signed-off-by: David Howells <dhowells@redhat.com>
Documentation/filesystems/caching/fscache.txt
fs/fscache/cookie.c
fs/fscache/internal.h
fs/fscache/page.c
fs/fscache/stats.c
include/linux/fscache-cache.h

index 0a77868f4977f28d16129843b908f27765020864..9cf2cfbc81c98f95e211e6982e1ed123d0d5d4ef 100644 (file)
@@ -269,6 +269,9 @@ proc files.
                oom=N   Number of store reqs failed -ENOMEM
                ops=N   Number of store reqs submitted
                run=N   Number of store reqs granted CPU time
+               pgs=N   Number of pages given store req processing time
+               rxd=N   Number of store reqs deleted from tracking tree
+               olm=N   Number of store reqs over store limit
        Ops     pend=N  Number of times async ops added to pending queues
                run=N   Number of times async ops given CPU time
                enq=N   Number of times async ops queued for processing
index e6854f5222f5240c4e8a04d9c0dfbce98865f8b5..f979659c1b3fa6094e4901641b4485c94b014c71 100644 (file)
@@ -36,6 +36,7 @@ void fscache_cookie_init_once(void *_cookie)
 
        memset(cookie, 0, sizeof(*cookie));
        spin_lock_init(&cookie->lock);
+       spin_lock_init(&cookie->stores_lock);
        INIT_HLIST_HEAD(&cookie->backing_objects);
 }
 
index 50324ad2b1941ca1ad53d375673df87a6ff3dd9e..ba1853fa1ff93934fad9926055fa3326fb92d2cd 100644 (file)
@@ -17,6 +17,7 @@
  * - cache->object_list_lock
  * - object->lock
  * - object->parent->lock
+ * - cookie->stores_lock
  * - fscache_thread_lock
  *
  */
@@ -174,6 +175,9 @@ extern atomic_t fscache_n_stores_nobufs;
 extern atomic_t fscache_n_stores_oom;
 extern atomic_t fscache_n_store_ops;
 extern atomic_t fscache_n_store_calls;
+extern atomic_t fscache_n_store_pages;
+extern atomic_t fscache_n_store_radix_deletes;
+extern atomic_t fscache_n_store_pages_over_limit;
 
 extern atomic_t fscache_n_marks;
 extern atomic_t fscache_n_uncaches;
index e6f2e61133a140273edd4de253c65cec705caa09..3ea8897bc21762ae6f7ada8b6f0eed3b368fcd58 100644 (file)
@@ -45,16 +45,26 @@ EXPORT_SYMBOL(__fscache_wait_on_page_write);
 /*
  * note that a page has finished being written to the cache
  */
-static void fscache_end_page_write(struct fscache_cookie *cookie, struct page *page)
+static void fscache_end_page_write(struct fscache_object *object,
+                                  struct page *page)
 {
-       struct page *xpage;
+       struct fscache_cookie *cookie;
+       struct page *xpage = NULL;
 
-       spin_lock(&cookie->lock);
-       xpage = radix_tree_delete(&cookie->stores, page->index);
-       spin_unlock(&cookie->lock);
-       ASSERT(xpage != NULL);
-
-       wake_up_bit(&cookie->flags, 0);
+       spin_lock(&object->lock);
+       cookie = object->cookie;
+       if (cookie) {
+               /* delete the page from the tree if it is now no longer
+                * pending */
+               spin_lock(&cookie->stores_lock);
+               fscache_stat(&fscache_n_store_radix_deletes);
+               xpage = radix_tree_delete(&cookie->stores, page->index);
+               spin_unlock(&cookie->stores_lock);
+               wake_up_bit(&cookie->flags, 0);
+       }
+       spin_unlock(&object->lock);
+       if (xpage)
+               page_cache_release(xpage);
 }
 
 /*
@@ -591,7 +601,7 @@ static void fscache_write_op(struct fscache_operation *_op)
        struct fscache_storage *op =
                container_of(_op, struct fscache_storage, op);
        struct fscache_object *object = op->op.object;
-       struct fscache_cookie *cookie = object->cookie;
+       struct fscache_cookie *cookie;
        struct page *page;
        unsigned n;
        void *results[1];
@@ -601,16 +611,17 @@ static void fscache_write_op(struct fscache_operation *_op)
 
        fscache_set_op_state(&op->op, "GetPage");
 
-       spin_lock(&cookie->lock);
        spin_lock(&object->lock);
+       cookie = object->cookie;
 
-       if (!fscache_object_is_active(object)) {
+       if (!fscache_object_is_active(object) || !cookie) {
                spin_unlock(&object->lock);
-               spin_unlock(&cookie->lock);
                _leave("");
                return;
        }
 
+       spin_lock(&cookie->stores_lock);
+
        fscache_stat(&fscache_n_store_calls);
 
        /* find a page to store */
@@ -621,23 +632,25 @@ static void fscache_write_op(struct fscache_operation *_op)
                goto superseded;
        page = results[0];
        _debug("gang %d [%lx]", n, page->index);
-       if (page->index > op->store_limit)
+       if (page->index > op->store_limit) {
+               fscache_stat(&fscache_n_store_pages_over_limit);
                goto superseded;
+       }
 
        radix_tree_tag_clear(&cookie->stores, page->index,
                             FSCACHE_COOKIE_PENDING_TAG);
 
+       spin_unlock(&cookie->stores_lock);
        spin_unlock(&object->lock);
-       spin_unlock(&cookie->lock);
 
        if (page) {
                fscache_set_op_state(&op->op, "Store");
+               fscache_stat(&fscache_n_store_pages);
                fscache_stat(&fscache_n_cop_write_page);
                ret = object->cache->ops->write_page(op, page);
                fscache_stat_d(&fscache_n_cop_write_page);
                fscache_set_op_state(&op->op, "EndWrite");
-               fscache_end_page_write(cookie, page);
-               page_cache_release(page);
+               fscache_end_page_write(object, page);
                if (ret < 0) {
                        fscache_set_op_state(&op->op, "Abort");
                        fscache_abort_object(object);
@@ -653,9 +666,9 @@ superseded:
        /* this writer is going away and there aren't any more things to
         * write */
        _debug("cease");
+       spin_unlock(&cookie->stores_lock);
        clear_bit(FSCACHE_OBJECT_PENDING_WRITE, &object->flags);
        spin_unlock(&object->lock);
-       spin_unlock(&cookie->lock);
        _leave("");
 }
 
@@ -731,6 +744,7 @@ int __fscache_write_page(struct fscache_cookie *cookie,
        /* add the page to the pending-storage radix tree on the backing
         * object */
        spin_lock(&object->lock);
+       spin_lock(&cookie->stores_lock);
 
        _debug("store limit %llx", (unsigned long long) object->store_limit);
 
@@ -751,6 +765,7 @@ int __fscache_write_page(struct fscache_cookie *cookie,
        if (test_and_set_bit(FSCACHE_OBJECT_PENDING_WRITE, &object->flags))
                goto already_pending;
 
+       spin_unlock(&cookie->stores_lock);
        spin_unlock(&object->lock);
 
        op->op.debug_id = atomic_inc_return(&fscache_op_debug_id);
@@ -772,6 +787,7 @@ int __fscache_write_page(struct fscache_cookie *cookie,
 already_queued:
        fscache_stat(&fscache_n_stores_again);
 already_pending:
+       spin_unlock(&cookie->stores_lock);
        spin_unlock(&object->lock);
        spin_unlock(&cookie->lock);
        radix_tree_preload_end();
@@ -781,7 +797,9 @@ already_pending:
        return 0;
 
 submit_failed:
+       spin_lock(&cookie->stores_lock);
        radix_tree_delete(&cookie->stores, page->index);
+       spin_unlock(&cookie->stores_lock);
        page_cache_release(page);
        ret = -ENOBUFS;
        goto nobufs;
index 4c07439d13077bddbc7e595f3d01a3f610cb1f1c..1d53ea68409e437d29165b8c35c8214a2eedb6aa 100644 (file)
@@ -58,6 +58,9 @@ atomic_t fscache_n_stores_nobufs;
 atomic_t fscache_n_stores_oom;
 atomic_t fscache_n_store_ops;
 atomic_t fscache_n_store_calls;
+atomic_t fscache_n_store_pages;
+atomic_t fscache_n_store_radix_deletes;
+atomic_t fscache_n_store_pages_over_limit;
 
 atomic_t fscache_n_marks;
 atomic_t fscache_n_uncaches;
@@ -200,9 +203,12 @@ static int fscache_stats_show(struct seq_file *m, void *v)
                   atomic_read(&fscache_n_stores_again),
                   atomic_read(&fscache_n_stores_nobufs),
                   atomic_read(&fscache_n_stores_oom));
-       seq_printf(m, "Stores : ops=%u run=%u\n",
+       seq_printf(m, "Stores : ops=%u run=%u pgs=%u rxd=%u olm=%u\n",
                   atomic_read(&fscache_n_store_ops),
-                  atomic_read(&fscache_n_store_calls));
+                  atomic_read(&fscache_n_store_calls),
+                  atomic_read(&fscache_n_store_pages),
+                  atomic_read(&fscache_n_store_radix_deletes),
+                  atomic_read(&fscache_n_store_pages_over_limit));
 
        seq_printf(m, "Ops    : pend=%u run=%u enq=%u can=%u\n",
                   atomic_read(&fscache_n_op_pend),
index 184cbdfbcc992cc5e2dd234d8cc9a96162dcb68e..f3aa4bdafef6b946290013c62ab10f563f73c0e5 100644 (file)
@@ -310,6 +310,7 @@ struct fscache_cookie {
        atomic_t                        usage;          /* number of users of this cookie */
        atomic_t                        n_children;     /* number of children of this cookie */
        spinlock_t                      lock;
+       spinlock_t                      stores_lock;    /* lock on page store tree */
        struct hlist_head               backing_objects; /* object(s) backing this file/index */
        const struct fscache_cookie_def *def;           /* definition */
        struct fscache_cookie           *parent;        /* parent of this entry */