ipc/sem.c: rewrite undo list locking
authorManfred Spraul <manfred@colorfullife.com>
Fri, 25 Jul 2008 08:48:06 +0000 (01:48 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Fri, 25 Jul 2008 17:53:42 +0000 (10:53 -0700)
The attached patch:
- reverses the locking order of ulp->lock and sem_lock:
  Previously, it was first ulp->lock, then inside sem_lock.
  Now it's the other way around.
- converts the undo structure to rcu.

Benefits:
- With the old locking order, IPC_RMID could not kfree the undo structures.
  The stale entries remained in the linked lists and were released later.
- The patch fixes a a race in semtimedop(): if both IPC_RMID and a semget() that
  recreates exactly the same id happen between find_alloc_undo() and sem_lock,
  then semtimedop() would access already kfree'd memory.

[akpm@linux-foundation.org: coding-style fixes]
Signed-off-by: Manfred Spraul <manfred@colorfullife.com>
Reviewed-by: Nadia Derbey <Nadia.Derbey@bull.net>
Cc: Pierre Peiffer <peifferp@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
include/linux/sem.h
ipc/sem.c

index d425993..1b191c1 100644 (file)
@@ -78,6 +78,7 @@ struct  seminfo {
 
 #ifdef __KERNEL__
 #include <asm/atomic.h>
+#include <linux/rcupdate.h>
 
 struct task_struct;
 
@@ -114,7 +115,10 @@ struct sem_queue {
  * when the process exits.
  */
 struct sem_undo {
-       struct list_head        list_proc;      /* per-process list: all undos from one process */
+       struct list_head        list_proc;      /* per-process list: all undos from one process. */
+                                               /* rcu protected */
+       struct rcu_head         rcu;            /* rcu struct for sem_undo() */
+       struct sem_undo_list    *ulp;           /* sem_undo_list for the process */
        struct list_head        list_id;        /* per semaphore array list: all undos for one array */
        int                     semid;          /* semaphore set identifier */
        short *                 semadj;         /* array of adjustments, one per semaphore */
index 3ca2327..bf1bc36 100644 (file)
--- a/ipc/sem.c
+++ b/ipc/sem.c
@@ -504,27 +504,35 @@ static int count_semzcnt (struct sem_array * sma, ushort semnum)
        return semzcnt;
 }
 
+void free_un(struct rcu_head *head)
+{
+       struct sem_undo *un = container_of(head, struct sem_undo, rcu);
+       kfree(un);
+}
+
 /* Free a semaphore set. freeary() is called with sem_ids.rw_mutex locked
  * as a writer and the spinlock for this semaphore set hold. sem_ids.rw_mutex
  * remains locked on exit.
  */
 static void freeary(struct ipc_namespace *ns, struct kern_ipc_perm *ipcp)
 {
-       struct sem_undo *un;
-       struct sem_queue *q, *t;
+       struct sem_undo *un, *tu;
+       struct sem_queue *q, *tq;
        struct sem_array *sma = container_of(ipcp, struct sem_array, sem_perm);
 
-       /* Invalidate the existing undo structures for this semaphore set.
-        * (They will be freed without any further action in exit_sem()
-        * or during the next semop.)
-        */
+       /* Free the existing undo structures for this semaphore set.  */
        assert_spin_locked(&sma->sem_perm.lock);
-       list_for_each_entry(un, &sma->list_id, list_id)
+       list_for_each_entry_safe(un, tu, &sma->list_id, list_id) {
+               list_del(&un->list_id);
+               spin_lock(&un->ulp->lock);
                un->semid = -1;
+               list_del_rcu(&un->list_proc);
+               spin_unlock(&un->ulp->lock);
+               call_rcu(&un->rcu, free_un);
+       }
 
        /* Wake up all pending processes and let them fail with EIDRM. */
-
-       list_for_each_entry_safe(q, t, &sma->sem_pending, list) {
+       list_for_each_entry_safe(q, tq, &sma->sem_pending, list) {
                list_del(&q->list);
 
                q->status = IN_WAKEUP;
@@ -948,16 +956,11 @@ static inline int get_undo_list(struct sem_undo_list **undo_listp)
 
 static struct sem_undo *lookup_undo(struct sem_undo_list *ulp, int semid)
 {
-       struct sem_undo *walk, *tmp;
+       struct sem_undo *walk;
 
-       assert_spin_locked(&ulp->lock);
-       list_for_each_entry_safe(walk, tmp, &ulp->list_proc, list_proc) {
+       list_for_each_entry_rcu(walk, &ulp->list_proc, list_proc) {
                if (walk->semid == semid)
                        return walk;
-               if (walk->semid == -1) {
-                       list_del(&walk->list_proc);
-                       kfree(walk);
-               }
        }
        return NULL;
 }
@@ -970,6 +973,8 @@ static struct sem_undo *lookup_undo(struct sem_undo_list *ulp, int semid)
  * The function looks up (and if not present creates) the undo structure.
  * The size of the undo structure depends on the size of the semaphore
  * array, thus the alloc path is not that straightforward.
+ * Lifetime-rules: sem_undo is rcu-protected, on success, the function
+ * performs a rcu_read_lock().
  */
 static struct sem_undo *find_alloc_undo(struct ipc_namespace *ns, int semid)
 {
@@ -983,11 +988,13 @@ static struct sem_undo *find_alloc_undo(struct ipc_namespace *ns, int semid)
        if (error)
                return ERR_PTR(error);
 
+       rcu_read_lock();
        spin_lock(&ulp->lock);
        un = lookup_undo(ulp, semid);
        spin_unlock(&ulp->lock);
        if (likely(un!=NULL))
                goto out;
+       rcu_read_unlock();
 
        /* no undo structure around - allocate one. */
        /* step 1: figure out the size of the semaphore array */
@@ -1005,38 +1012,38 @@ static struct sem_undo *find_alloc_undo(struct ipc_namespace *ns, int semid)
                return ERR_PTR(-ENOMEM);
        }
 
-       /* step 3: Acquire the lock on the undo list pointer */
-       spin_lock(&ulp->lock);
-
-       /* step 4: check for races: someone else allocated the undo struct,
-        *         semaphore array was destroyed.
-        */
-       un = lookup_undo(ulp, semid);
-       if (un) {
-               spin_unlock(&ulp->lock);
-               kfree(new);
-               sem_putref(sma);
-               goto out;
-       }
+       /* step 3: Acquire the lock on semaphore array */
        sem_lock_and_putref(sma);
        if (sma->sem_perm.deleted) {
                sem_unlock(sma);
-               spin_unlock(&ulp->lock);
                kfree(new);
                un = ERR_PTR(-EIDRM);
                goto out;
        }
+       spin_lock(&ulp->lock);
+
+       /*
+        * step 4: check for races: did someone else allocate the undo struct?
+        */
+       un = lookup_undo(ulp, semid);
+       if (un) {
+               kfree(new);
+               goto success;
+       }
        /* step 5: initialize & link new undo structure */
        new->semadj = (short *) &new[1];
+       new->ulp = ulp;
        new->semid = semid;
        assert_spin_locked(&ulp->lock);
-       list_add(&new->list_proc, &ulp->list_proc);
+       list_add_rcu(&new->list_proc, &ulp->list_proc);
        assert_spin_locked(&sma->sem_perm.lock);
        list_add(&new->list_id, &sma->list_id);
+       un = new;
 
-       sem_unlock(sma);
+success:
        spin_unlock(&ulp->lock);
-       un = new;
+       rcu_read_lock();
+       sem_unlock(sma);
 out:
        return un;
 }
@@ -1103,6 +1110,8 @@ asmlinkage long sys_semtimedop(int semid, struct sembuf __user *tsops,
 
        sma = sem_lock_check(ns, semid);
        if (IS_ERR(sma)) {
+               if (un)
+                       rcu_read_unlock();
                error = PTR_ERR(sma);
                goto out_free;
        }
@@ -1111,10 +1120,26 @@ asmlinkage long sys_semtimedop(int semid, struct sembuf __user *tsops,
         * semid identifiers are not unique - find_alloc_undo may have
         * allocated an undo structure, it was invalidated by an RMID
         * and now a new array with received the same id. Check and fail.
+        * This case can be detected checking un->semid. The existance of
+        * "un" itself is guaranteed by rcu.
         */
        error = -EIDRM;
-       if (un && un->semid == -1)
-               goto out_unlock_free;
+       if (un) {
+               if (un->semid == -1) {
+                       rcu_read_unlock();
+                       goto out_unlock_free;
+               } else {
+                       /*
+                        * rcu lock can be released, "un" cannot disappear:
+                        * - sem_lock is acquired, thus IPC_RMID is
+                        *   impossible.
+                        * - exit_sem is impossible, it always operates on
+                        *   current (or a dead task).
+                        */
+
+                       rcu_read_unlock();
+               }
+       }
 
        error = -EFBIG;
        if (max >= sma->sem_nsems)
@@ -1242,7 +1267,6 @@ int copy_semundo(unsigned long clone_flags, struct task_struct *tsk)
 void exit_sem(struct task_struct *tsk)
 {
        struct sem_undo_list *ulp;
-       struct sem_undo *un, *tmp;
 
        ulp = tsk->sysvsem.undo_list;
        if (!ulp)
@@ -1252,28 +1276,47 @@ void exit_sem(struct task_struct *tsk)
        if (!atomic_dec_and_test(&ulp->refcnt))
                return;
 
-       spin_lock(&ulp->lock);
-
-       list_for_each_entry_safe(un, tmp, &ulp->list_proc, list_proc) {
+       for (;;) {
                struct sem_array *sma;
+               struct sem_undo *un;
+               int semid;
                int i;
 
-               if (un->semid == -1)
-                       goto free;
+               rcu_read_lock();
+               un = list_entry(rcu_dereference(ulp->list_proc.next),
+                                       struct sem_undo, list_proc);
+               if (&un->list_proc == &ulp->list_proc)
+                       semid = -1;
+                else
+                       semid = un->semid;
+               rcu_read_unlock();
 
-               sma = sem_lock(tsk->nsproxy->ipc_ns, un->semid);
-               if (IS_ERR(sma))
-                       goto free;
+               if (semid == -1)
+                       break;
 
-               if (un->semid == -1)
-                       goto unlock_free;
+               sma = sem_lock_check(tsk->nsproxy->ipc_ns, un->semid);
 
-               BUG_ON(sem_checkid(sma, un->semid));
+               /* exit_sem raced with IPC_RMID, nothing to do */
+               if (IS_ERR(sma))
+                       continue;
 
-               /* remove un from sma->list_id */
+               un = lookup_undo(ulp, semid);
+               if (un == NULL) {
+                       /* exit_sem raced with IPC_RMID+semget() that created
+                        * exactly the same semid. Nothing to do.
+                        */
+                       sem_unlock(sma);
+                       continue;
+               }
+
+               /* remove un from the linked lists */
                assert_spin_locked(&sma->sem_perm.lock);
                list_del(&un->list_id);
 
+               spin_lock(&ulp->lock);
+               list_del_rcu(&un->list_proc);
+               spin_unlock(&ulp->lock);
+
                /* perform adjustments registered in un */
                for (i = 0; i < sma->sem_nsems; i++) {
                        struct sem * semaphore = &sma->sem_base[i];
@@ -1302,14 +1345,10 @@ void exit_sem(struct task_struct *tsk)
                sma->sem_otime = get_seconds();
                /* maybe some queued-up processes were waiting for this */
                update_queue(sma);
-unlock_free:
                sem_unlock(sma);
-free:
-               assert_spin_locked(&ulp->lock);
-               list_del(&un->list_proc);
-               kfree(un);
+
+               call_rcu(&un->rcu, free_un);
        }
-       spin_unlock(&ulp->lock);
        kfree(ulp);
 }