usb: gadget: phonet: Add Phonet over ACM for RMC
[linux-3.10.git] / net / unix / garbage.c
index 411802b..9bc73f8 100644 (file)
  *     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>
-#include <linux/sched.h>
 #include <linux/string.h>
 #include <linux/socket.h>
 #include <linux/un.h>
 #include <linux/net.h>
 #include <linux/fs.h>
-#include <linux/slab.h>
 #include <linux/skbuff.h>
 #include <linux/netdevice.h>
 #include <linux/file.h>
 #include <linux/proc_fs.h>
+#include <linux/mutex.h>
+#include <linux/wait.h>
 
 #include <net/sock.h>
 #include <net/af_unix.h>
 
 /* 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);
+static DECLARE_WAIT_QUEUE_HEAD(unix_gc_wait);
 
-atomic_t unix_tot_inflight = ATOMIC_INIT(0);
+unsigned int unix_tot_inflight;
 
 
-static struct sock *unix_get_socket(struct file *filp)
+struct sock *unix_get_socket(struct file *filp)
 {
        struct sock *u_sock = NULL;
-       struct inode *inode = filp->f_dentry->d_inode;
+       struct inode *inode = file_inode(filp);
 
        /*
         *      Socket ?
         */
-       if (S_ISSOCK(inode->i_mode)) {
-               struct socket * sock = SOCKET_I(inode);
-               struct sock * s = sock->sk;
+       if (S_ISSOCK(inode->i_mode) && !(filp->f_mode & FMODE_PATH)) {
+               struct socket *sock = SOCKET_I(inode);
+               struct sock *s = sock->sk;
 
                /*
                 *      PF_UNIX ?
@@ -117,196 +121,266 @@ static struct sock *unix_get_socket(struct file *filp)
  *     Keep the number of times in flight count for the file
  *     descriptor if it is for an AF_UNIX socket.
  */
+
 void unix_inflight(struct file *fp)
 {
        struct sock *s = unix_get_socket(fp);
-       if(s) {
-               atomic_inc(&unix_sk(s)->inflight);
-               atomic_inc(&unix_tot_inflight);
+       if (s) {
+               struct unix_sock *u = unix_sk(s);
+               spin_lock(&unix_gc_lock);
+               if (atomic_long_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));
+               }
+               unix_tot_inflight++;
+               spin_unlock(&unix_gc_lock);
        }
 }
 
 void unix_notinflight(struct file *fp)
 {
        struct sock *s = unix_get_socket(fp);
-       if(s) {
-               atomic_dec(&unix_sk(s)->inflight);
-               atomic_dec(&unix_tot_inflight);
+       if (s) {
+               struct unix_sock *u = unix_sk(s);
+               spin_lock(&unix_gc_lock);
+               BUG_ON(list_empty(&u->link));
+               if (atomic_long_dec_and_test(&u->inflight))
+                       list_del_init(&u->link);
+               unix_tot_inflight--;
+               spin_unlock(&unix_gc_lock);
        }
 }
 
+static void scan_inflight(struct sock *x, void (*func)(struct unix_sock *),
+                         struct sk_buff_head *hitlist)
+{
+       struct sk_buff *skb;
+       struct sk_buff *next;
 
-/*
- *     Garbage Collector Support Functions
- */
+       spin_lock(&x->sk_receive_queue.lock);
+       skb_queue_walk_safe(&x->sk_receive_queue, skb, next) {
+               /*
+                *      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) {
+                                       struct unix_sock *u = unix_sk(sk);
+
+                                       /*
+                                        * Ignore non-candidates, they could
+                                        * have been added to the queues after
+                                        * starting the garbage collection
+                                        */
+                                       if (test_bit(UNIX_GC_CANDIDATE, &u->gc_flags)) {
+                                               hit = true;
+                                               func(u);
+                                       }
+                               }
+                       }
+                       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 struct sock *pop_stack(void)
+static void scan_children(struct sock *x, void (*func)(struct unix_sock *),
+                         struct sk_buff_head *hitlist)
 {
-       struct sock *p = gc_current;
-       gc_current = unix_sk(p)->gc_tree;
-       return p;
+       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);
+               skb_queue_walk_safe(&x->sk_receive_queue, skb, next) {
+                       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 inline int empty_stack(void)
+static void dec_inflight(struct unix_sock *usk)
 {
-       return gc_current == GC_HEAD;
+       atomic_long_dec(&usk->inflight);
 }
 
-static void maybe_unmark_and_push(struct sock *x)
+static void inc_inflight(struct unix_sock *usk)
 {
-       struct unix_sock *u = unix_sk(x);
+       atomic_long_inc(&usk->inflight);
+}
 
-       if (u->gc_tree != GC_ORPHAN)
-               return;
-       sock_hold(x);
-       u->gc_tree = gc_current;
-       gc_current = x;
+static void inc_inflight_move_tail(struct unix_sock *u)
+{
+       atomic_long_inc(&u->inflight);
+       /*
+        * If this still might be part of a cycle, move it to the end
+        * of the list, so that it's checked even if it was already
+        * passed over
+        */
+       if (test_bit(UNIX_GC_MAYBE_CYCLE, &u->gc_flags))
+               list_move_tail(&u->link, &gc_candidates);
 }
 
+static bool gc_in_progress = false;
+#define UNIX_INFLIGHT_TRIGGER_GC 16000
 
-/* The external entry point: unix_gc() */
+void wait_for_unix_gc(void)
+{
+       /*
+        * If number of inflight sockets is insane,
+        * force a garbage collect right now.
+        */
+       if (unix_tot_inflight > UNIX_INFLIGHT_TRIGGER_GC && !gc_in_progress)
+               unix_gc();
+       wait_event(unix_gc_wait, gc_in_progress == false);
+}
 
+/* The external entry point: unix_gc() */
 void unix_gc(void)
 {
-       static DECLARE_MUTEX(unix_gc_sem);
-       int i;
-       struct sock *s;
+       struct unix_sock *u;
+       struct unix_sock *next;
        struct sk_buff_head hitlist;
-       struct sk_buff *skb;
+       struct list_head cursor;
+       LIST_HEAD(not_cycle_list);
 
-       /*
-        *      Avoid a recursive GC.
-        */
-
-       if (down_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 
+        * 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.  Since there are no possible receivers, all
+        * buffers currently on the candidates' queues stay there
+        * during the garbage collection.
+        *
+        * We also know that no new candidate can be added onto the
+        * receive queues.  Other, non candidate sockets _can_ be
+        * added to queue, so we must make sure only to touch
+        * candidates.
         */
+       list_for_each_entry_safe(u, next, &gc_inflight_list, link) {
+               long total_refs;
+               long inflight_refs;
+
+               total_refs = file_count(u->sk.sk_socket->file);
+               inflight_refs = atomic_long_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);
+                       __set_bit(UNIX_GC_CANDIDATE, &u->gc_flags);
+                       __set_bit(UNIX_GC_MAYBE_CYCLE, &u->gc_flags);
+               }
+       }
 
-       /* 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
+       /*
+        * Now remove all internal in-flight reference to children of
+        * the candidates.
         */
+       list_for_each_entry(u, &gc_candidates, link)
+               scan_children(&u->sk, dec_inflight, NULL);
 
        /*
-        *      Push root set
+        * 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);
 
-       forall_unix_sockets(i, s)
-       {
-               int open_count = 0;
+               /* Move cursor to after the current position. */
+               list_move(&cursor, &u->link);
 
-               /*
-                *      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);
+               if (atomic_long_read(&u->inflight) > 0) {
+                       list_move_tail(&u->link, &not_cycle_list);
+                       __clear_bit(UNIX_GC_MAYBE_CYCLE, &u->gc_flags);
+                       scan_children(&u->sk, inc_inflight_move_tail, NULL);
+               }
        }
+       list_del(&cursor);
 
        /*
-        *      Mark phase 
+        * not_cycle_list contains those sockets which do not make up a
+        * cycle.  Restore these to the inflight list.
         */
-
-       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 
-                */
-               
-               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;
-               }
-               spin_unlock(&x->sk_receive_queue.lock);
-               sock_put(x);
+       while (!list_empty(&not_cycle_list)) {
+               u = list_entry(not_cycle_list.next, struct unix_sock, link);
+               __clear_bit(UNIX_GC_CANDIDATE, &u->gc_flags);
+               list_move_tail(&u->link, &gc_inflight_list);
        }
 
+       /*
+        * 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;
+       wake_up(&unix_gc_wait);
 
-       __skb_queue_purge(&hitlist);
-       up(&unix_gc_sem);
+ out:
+       spin_unlock(&unix_gc_lock);
 }