summaryrefslogtreecommitdiff
path: root/net
diff options
context:
space:
mode:
authorMiklos Szeredi <mszeredi@suse.cz>2007-07-11 14:22:39 -0700
committerDavid S. Miller <davem@davemloft.net>2007-07-11 14:22:39 -0700
commit1fd05ba5a2f2aa8e7b9b52ef55df850e2e7d54c9 (patch)
tree3403194403ab25f1f7a360a002adf63be2b4e9b0 /net
parent99d24edeb6abc6ca3a0d0fbdb83c664c04403c8c (diff)
[AF_UNIX]: Rewrite garbage collector, fixes race.
Throw out the old mark & sweep garbage collector and put in a refcounting cycle detecting one. The old one had a race with recvmsg, that resulted in false positives and hence data loss. The old algorithm operated on all unix sockets in the system, so any additional locking would have meant performance problems for all users of these. The new algorithm instead only operates on "in flight" sockets, which are very rare, and the additional locking for these doesn't negatively impact the vast majority of users. In fact it's probable, that there weren't *any* heavy senders of sockets over sockets, otherwise the above race would have been discovered long ago. The patch works OK with the app that exposed the race with the old code. The garbage collection has also been verified to work in a few simple cases. Signed-off-by: Miklos Szeredi <mszeredi@suse.cz> Signed-off-by: David S. Miller <davem@davemloft.net>
Diffstat (limited to 'net')
-rw-r--r--net/unix/af_unix.c6
-rw-r--r--net/unix/garbage.c325
2 files changed, 188 insertions, 143 deletions
diff --git a/net/unix/af_unix.c b/net/unix/af_unix.c
index 3654b647ac79..65ebccc0a698 100644
--- a/net/unix/af_unix.c
+++ b/net/unix/af_unix.c
@@ -592,7 +592,8 @@ static struct sock * unix_create1(struct socket *sock)
u->dentry = NULL;
u->mnt = NULL;
spin_lock_init(&u->lock);
- atomic_set(&u->inflight, sock ? 0 : -1);
+ atomic_set(&u->inflight, 0);
+ INIT_LIST_HEAD(&u->link);
mutex_init(&u->readlock); /* single task reading lock */
init_waitqueue_head(&u->peer_wait);
unix_insert_socket(unix_sockets_unbound, sk);
@@ -1134,9 +1135,6 @@ restart:
/* take ten and and send info to listening sock */
spin_lock(&other->sk_receive_queue.lock);
__skb_queue_tail(&other->sk_receive_queue, skb);
- /* Undo artificially decreased inflight after embrion
- * is installed to listening socket. */
- atomic_inc(&newu->inflight);
spin_unlock(&other->sk_receive_queue.lock);
unix_state_unlock(other);
other->sk_data_ready(other, 0);
diff --git a/net/unix/garbage.c b/net/unix/garbage.c
index f20b7ea7c555..406b6433e467 100644
--- a/net/unix/garbage.c
+++ b/net/unix/garbage.c
@@ -62,6 +62,10 @@
* AV 1 Mar 1999
* Damn. Added missing check for ->dead in listen queues scanning.
*
+ * Miklos Szeredi 25 Jun 2007
+ * Reimplement with a cycle collecting algorithm. This should
+ * solve several problems with the previous code, like being racy
+ * wrt receive and holding up unrelated socket operations.
*/
#include <linux/kernel.h>
@@ -84,10 +88,9 @@
/* Internal data structures and random procedures: */
-#define GC_HEAD ((struct sock *)(-1))
-#define GC_ORPHAN ((struct sock *)(-3))
-
-static struct sock *gc_current = GC_HEAD; /* stack of objects to mark */
+static LIST_HEAD(gc_inflight_list);
+static LIST_HEAD(gc_candidates);
+static DEFINE_SPINLOCK(unix_gc_lock);
atomic_t unix_tot_inflight = ATOMIC_INIT(0);
@@ -122,8 +125,16 @@ void unix_inflight(struct file *fp)
{
struct sock *s = unix_get_socket(fp);
if(s) {
- atomic_inc(&unix_sk(s)->inflight);
+ struct unix_sock *u = unix_sk(s);
+ spin_lock(&unix_gc_lock);
+ if (atomic_inc_return(&u->inflight) == 1) {
+ BUG_ON(!list_empty(&u->link));
+ list_add_tail(&u->link, &gc_inflight_list);
+ } else {
+ BUG_ON(list_empty(&u->link));
+ }
atomic_inc(&unix_tot_inflight);
+ spin_unlock(&unix_gc_lock);
}
}
@@ -131,182 +142,218 @@ void unix_notinflight(struct file *fp)
{
struct sock *s = unix_get_socket(fp);
if(s) {
- atomic_dec(&unix_sk(s)->inflight);
+ struct unix_sock *u = unix_sk(s);
+ spin_lock(&unix_gc_lock);
+ BUG_ON(list_empty(&u->link));
+ if (atomic_dec_and_test(&u->inflight))
+ list_del_init(&u->link);
atomic_dec(&unix_tot_inflight);
+ spin_unlock(&unix_gc_lock);
}
}
+static inline struct sk_buff *sock_queue_head(struct sock *sk)
+{
+ return (struct sk_buff *) &sk->sk_receive_queue;
+}
-/*
- * Garbage Collector Support Functions
- */
+#define receive_queue_for_each_skb(sk, next, skb) \
+ for (skb = sock_queue_head(sk)->next, next = skb->next; \
+ skb != sock_queue_head(sk); skb = next, next = skb->next)
-static inline struct sock *pop_stack(void)
+static void scan_inflight(struct sock *x, void (*func)(struct sock *),
+ struct sk_buff_head *hitlist)
{
- struct sock *p = gc_current;
- gc_current = unix_sk(p)->gc_tree;
- return p;
+ struct sk_buff *skb;
+ struct sk_buff *next;
+
+ spin_lock(&x->sk_receive_queue.lock);
+ receive_queue_for_each_skb(x, next, skb) {
+ /*
+ * Do we have file descriptors ?
+ */
+ if (UNIXCB(skb).fp) {
+ bool hit = false;
+ /*
+ * Process the descriptors of this socket
+ */
+ int nfd = UNIXCB(skb).fp->count;
+ struct file **fp = UNIXCB(skb).fp->fp;
+ while (nfd--) {
+ /*
+ * Get the socket the fd matches
+ * if it indeed does so
+ */
+ struct sock *sk = unix_get_socket(*fp++);
+ if(sk) {
+ hit = true;
+ func(sk);
+ }
+ }
+ if (hit && hitlist != NULL) {
+ __skb_unlink(skb, &x->sk_receive_queue);
+ __skb_queue_tail(hitlist, skb);
+ }
+ }
+ }
+ spin_unlock(&x->sk_receive_queue.lock);
}
-static inline int empty_stack(void)
+static void scan_children(struct sock *x, void (*func)(struct sock *),
+ struct sk_buff_head *hitlist)
{
- return gc_current == GC_HEAD;
+ if (x->sk_state != TCP_LISTEN)
+ scan_inflight(x, func, hitlist);
+ else {
+ struct sk_buff *skb;
+ struct sk_buff *next;
+ struct unix_sock *u;
+ LIST_HEAD(embryos);
+
+ /*
+ * For a listening socket collect the queued embryos
+ * and perform a scan on them as well.
+ */
+ spin_lock(&x->sk_receive_queue.lock);
+ receive_queue_for_each_skb(x, next, skb) {
+ u = unix_sk(skb->sk);
+
+ /*
+ * An embryo cannot be in-flight, so it's safe
+ * to use the list link.
+ */
+ BUG_ON(!list_empty(&u->link));
+ list_add_tail(&u->link, &embryos);
+ }
+ spin_unlock(&x->sk_receive_queue.lock);
+
+ while (!list_empty(&embryos)) {
+ u = list_entry(embryos.next, struct unix_sock, link);
+ scan_inflight(&u->sk, func, hitlist);
+ list_del_init(&u->link);
+ }
+ }
}
-static void maybe_unmark_and_push(struct sock *x)
+static void dec_inflight(struct sock *sk)
{
- struct unix_sock *u = unix_sk(x);
+ atomic_dec(&unix_sk(sk)->inflight);
+}
- if (u->gc_tree != GC_ORPHAN)
- return;
- sock_hold(x);
- u->gc_tree = gc_current;
- gc_current = x;
+static void inc_inflight(struct sock *sk)
+{
+ atomic_inc(&unix_sk(sk)->inflight);
}
+static void inc_inflight_move_tail(struct sock *sk)
+{
+ struct unix_sock *u = unix_sk(sk);
+
+ atomic_inc(&u->inflight);
+ /*
+ * If this is still a candidate, move it to the end of the
+ * list, so that it's checked even if it was already passed
+ * over
+ */
+ if (u->gc_candidate)
+ list_move_tail(&u->link, &gc_candidates);
+}
/* The external entry point: unix_gc() */
void unix_gc(void)
{
- static DEFINE_MUTEX(unix_gc_sem);
- int i;
- struct sock *s;
- struct sk_buff_head hitlist;
- struct sk_buff *skb;
+ static bool gc_in_progress = false;
- /*
- * Avoid a recursive GC.
- */
+ struct unix_sock *u;
+ struct unix_sock *next;
+ struct sk_buff_head hitlist;
+ struct list_head cursor;
- if (!mutex_trylock(&unix_gc_sem))
- return;
+ spin_lock(&unix_gc_lock);
- spin_lock(&unix_table_lock);
+ /* Avoid a recursive GC. */
+ if (gc_in_progress)
+ goto out;
- forall_unix_sockets(i, s)
- {
- unix_sk(s)->gc_tree = GC_ORPHAN;
- }
+ gc_in_progress = true;
/*
- * Everything is now marked
- */
-
- /* Invariant to be maintained:
- - everything unmarked is either:
- -- (a) on the stack, or
- -- (b) has all of its children unmarked
- - everything on the stack is always unmarked
- - nothing is ever pushed onto the stack twice, because:
- -- nothing previously unmarked is ever pushed on the stack
+ * First, select candidates for garbage collection. Only
+ * in-flight sockets are considered, and from those only ones
+ * which don't have any external reference.
+ *
+ * Holding unix_gc_lock will protect these candidates from
+ * being detached, and hence from gaining an external
+ * reference. This also means, that since there are no
+ * possible receivers, the receive queues of these sockets are
+ * static during the GC, even though the dequeue is done
+ * before the detach without atomicity guarantees.
*/
+ list_for_each_entry_safe(u, next, &gc_inflight_list, link) {
+ int total_refs;
+ int inflight_refs;
+
+ total_refs = file_count(u->sk.sk_socket->file);
+ inflight_refs = atomic_read(&u->inflight);
+
+ BUG_ON(inflight_refs < 1);
+ BUG_ON(total_refs < inflight_refs);
+ if (total_refs == inflight_refs) {
+ list_move_tail(&u->link, &gc_candidates);
+ u->gc_candidate = 1;
+ }
+ }
/*
- * Push root set
+ * Now remove all internal in-flight reference to children of
+ * the candidates.
*/
-
- forall_unix_sockets(i, s)
- {
- int open_count = 0;
-
- /*
- * If all instances of the descriptor are not
- * in flight we are in use.
- *
- * Special case: when socket s is embrion, it may be
- * hashed but still not in queue of listening socket.
- * In this case (see unix_create1()) we set artificial
- * negative inflight counter to close race window.
- * It is trick of course and dirty one.
- */
- if (s->sk_socket && s->sk_socket->file)
- open_count = file_count(s->sk_socket->file);
- if (open_count > atomic_read(&unix_sk(s)->inflight))
- maybe_unmark_and_push(s);
- }
+ list_for_each_entry(u, &gc_candidates, link)
+ scan_children(&u->sk, dec_inflight, NULL);
/*
- * Mark phase
+ * Restore the references for children of all candidates,
+ * which have remaining references. Do this recursively, so
+ * only those remain, which form cyclic references.
+ *
+ * Use a "cursor" link, to make the list traversal safe, even
+ * though elements might be moved about.
*/
+ list_add(&cursor, &gc_candidates);
+ while (cursor.next != &gc_candidates) {
+ u = list_entry(cursor.next, struct unix_sock, link);
- while (!empty_stack())
- {
- struct sock *x = pop_stack();
- struct sock *sk;
-
- spin_lock(&x->sk_receive_queue.lock);
- skb = skb_peek(&x->sk_receive_queue);
-
- /*
- * Loop through all but first born
- */
+ /* Move cursor to after the current position. */
+ list_move(&cursor, &u->link);
- while (skb && skb != (struct sk_buff *)&x->sk_receive_queue) {
- /*
- * Do we have file descriptors ?
- */
- if(UNIXCB(skb).fp)
- {
- /*
- * Process the descriptors of this socket
- */
- int nfd=UNIXCB(skb).fp->count;
- struct file **fp = UNIXCB(skb).fp->fp;
- while(nfd--)
- {
- /*
- * Get the socket the fd matches if
- * it indeed does so
- */
- if((sk=unix_get_socket(*fp++))!=NULL)
- {
- maybe_unmark_and_push(sk);
- }
- }
- }
- /* We have to scan not-yet-accepted ones too */
- if (x->sk_state == TCP_LISTEN)
- maybe_unmark_and_push(skb->sk);
- skb=skb->next;
+ if (atomic_read(&u->inflight) > 0) {
+ list_move_tail(&u->link, &gc_inflight_list);
+ u->gc_candidate = 0;
+ scan_children(&u->sk, inc_inflight_move_tail, NULL);
}
- spin_unlock(&x->sk_receive_queue.lock);
- sock_put(x);
}
+ list_del(&cursor);
+ /*
+ * Now gc_candidates contains only garbage. Restore original
+ * inflight counters for these as well, and remove the skbuffs
+ * which are creating the cycle(s).
+ */
skb_queue_head_init(&hitlist);
+ list_for_each_entry(u, &gc_candidates, link)
+ scan_children(&u->sk, inc_inflight, &hitlist);
- forall_unix_sockets(i, s)
- {
- struct unix_sock *u = unix_sk(s);
+ spin_unlock(&unix_gc_lock);
- if (u->gc_tree == GC_ORPHAN) {
- struct sk_buff *nextsk;
+ /* Here we are. Hitlist is filled. Die. */
+ __skb_queue_purge(&hitlist);
- spin_lock(&s->sk_receive_queue.lock);
- skb = skb_peek(&s->sk_receive_queue);
- while (skb &&
- skb != (struct sk_buff *)&s->sk_receive_queue) {
- nextsk = skb->next;
- /*
- * Do we have file descriptors ?
- */
- if (UNIXCB(skb).fp) {
- __skb_unlink(skb,
- &s->sk_receive_queue);
- __skb_queue_tail(&hitlist, skb);
- }
- skb = nextsk;
- }
- spin_unlock(&s->sk_receive_queue.lock);
- }
- u->gc_tree = GC_ORPHAN;
- }
- spin_unlock(&unix_table_lock);
+ spin_lock(&unix_gc_lock);
- /*
- * Here we are. Hitlist is filled. Die.
- */
+ /* All candidates should have been detached by now. */
+ BUG_ON(!list_empty(&gc_candidates));
+ gc_in_progress = false;
- __skb_queue_purge(&hitlist);
- mutex_unlock(&unix_gc_sem);
+ out:
+ spin_unlock(&unix_gc_lock);
}