tcp: add tcp_min_snd_mss sysctl
[pandora-kernel.git] / ipc / shm.c
index 27884ad..df1d608 100644 (file)
--- a/ipc/shm.c
+++ b/ipc/shm.c
@@ -74,6 +74,7 @@ void shm_init_ns(struct ipc_namespace *ns)
        ns->shm_ctlmax = SHMMAX;
        ns->shm_ctlall = SHMALL;
        ns->shm_ctlmni = SHMMNI;
+       ns->shm_rmid_forced = 0;
        ns->shm_tot = 0;
        ipc_init_ids(&shm_ids(ns));
 }
@@ -104,9 +105,16 @@ void shm_exit_ns(struct ipc_namespace *ns)
 }
 #endif
 
-void __init shm_init (void)
+static int __init ipc_ns_init(void)
 {
        shm_init_ns(&init_ipc_ns);
+       return 0;
+}
+
+pure_initcall(ipc_ns_init);
+
+void __init shm_init (void)
+{
        ipc_init_proc_interface("sysvipc/shm",
 #if BITS_PER_LONG <= 32
                                "       key      shmid perms       size  cpid  lpid nattch   uid   gid  cuid  cgid      atime      dtime      ctime        rss       swap\n",
@@ -130,6 +138,12 @@ static inline struct shmid_kernel *shm_lock(struct ipc_namespace *ns, int id)
        return container_of(ipcp, struct shmid_kernel, shm_perm);
 }
 
+static inline void shm_lock_by_ptr(struct shmid_kernel *ipcp)
+{
+       rcu_read_lock();
+       spin_lock(&ipcp->shm_perm.lock);
+}
+
 static inline struct shmid_kernel *shm_lock_check(struct ipc_namespace *ns,
                                                int id)
 {
@@ -186,6 +200,23 @@ static void shm_destroy(struct ipc_namespace *ns, struct shmid_kernel *shp)
        ipc_rcu_putref(shp);
 }
 
+/*
+ * shm_may_destroy - identifies whether shm segment should be destroyed now
+ *
+ * Returns true if and only if there are no active users of the segment and
+ * one of the following is true:
+ *
+ * 1) shmctl(id, IPC_RMID, NULL) was called for this shp
+ *
+ * 2) sysctl kernel.shm_rmid_forced is set to 1.
+ */
+static bool shm_may_destroy(struct ipc_namespace *ns, struct shmid_kernel *shp)
+{
+       return (shp->shm_nattch == 0) &&
+              (ns->shm_rmid_forced ||
+               (shp->shm_perm.mode & SHM_DEST));
+}
+
 /*
  * remove the attach descriptor vma.
  * free memory for segment if it is marked destroyed.
@@ -206,14 +237,90 @@ static void shm_close(struct vm_area_struct *vma)
        shp->shm_lprid = task_tgid_vnr(current);
        shp->shm_dtim = get_seconds();
        shp->shm_nattch--;
-       if(shp->shm_nattch == 0 &&
-          shp->shm_perm.mode & SHM_DEST)
+       if (shm_may_destroy(ns, shp))
                shm_destroy(ns, shp);
        else
                shm_unlock(shp);
        up_write(&shm_ids(ns).rw_mutex);
 }
 
+/* Called with ns->shm_ids(ns).rw_mutex locked */
+static int shm_try_destroy_current(int id, void *p, void *data)
+{
+       struct ipc_namespace *ns = data;
+       struct kern_ipc_perm *ipcp = p;
+       struct shmid_kernel *shp = container_of(ipcp, struct shmid_kernel, shm_perm);
+
+       if (shp->shm_creator != current)
+               return 0;
+
+       /*
+        * Mark it as orphaned to destroy the segment when
+        * kernel.shm_rmid_forced is changed.
+        * It is noop if the following shm_may_destroy() returns true.
+        */
+       shp->shm_creator = NULL;
+
+       /*
+        * Don't even try to destroy it.  If shm_rmid_forced=0 and IPC_RMID
+        * is not set, it shouldn't be deleted here.
+        */
+       if (!ns->shm_rmid_forced)
+               return 0;
+
+       if (shm_may_destroy(ns, shp)) {
+               shm_lock_by_ptr(shp);
+               shm_destroy(ns, shp);
+       }
+       return 0;
+}
+
+/* Called with ns->shm_ids(ns).rw_mutex locked */
+static int shm_try_destroy_orphaned(int id, void *p, void *data)
+{
+       struct ipc_namespace *ns = data;
+       struct kern_ipc_perm *ipcp = p;
+       struct shmid_kernel *shp = container_of(ipcp, struct shmid_kernel, shm_perm);
+
+       /*
+        * We want to destroy segments without users and with already
+        * exit'ed originating process.
+        *
+        * As shp->* are changed under rw_mutex, it's safe to skip shp locking.
+        */
+       if (shp->shm_creator != NULL)
+               return 0;
+
+       if (shm_may_destroy(ns, shp)) {
+               shm_lock_by_ptr(shp);
+               shm_destroy(ns, shp);
+       }
+       return 0;
+}
+
+void shm_destroy_orphaned(struct ipc_namespace *ns)
+{
+       down_write(&shm_ids(ns).rw_mutex);
+       if (shm_ids(ns).in_use)
+               idr_for_each(&shm_ids(ns).ipcs_idr, &shm_try_destroy_orphaned, ns);
+       up_write(&shm_ids(ns).rw_mutex);
+}
+
+
+void exit_shm(struct task_struct *task)
+{
+       struct ipc_namespace *ns = task->nsproxy->ipc_ns;
+
+       if (shm_ids(ns).in_use == 0)
+               return;
+
+       /* Destroy all already created segments, but not mapped yet */
+       down_write(&shm_ids(ns).rw_mutex);
+       if (shm_ids(ns).in_use)
+               idr_for_each(&shm_ids(ns).ipcs_idr, &shm_try_destroy_current, ns);
+       up_write(&shm_ids(ns).rw_mutex);
+}
+
 static int shm_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
 {
        struct file *file = vma->vm_file;
@@ -343,7 +450,7 @@ static int newseg(struct ipc_namespace *ns, struct ipc_params *params)
        size_t size = params->u.size;
        int error;
        struct shmid_kernel *shp;
-       int numpages = (size + PAGE_SIZE -1) >> PAGE_SHIFT;
+       size_t numpages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
        struct file * file;
        char name[13];
        int id;
@@ -391,12 +498,6 @@ static int newseg(struct ipc_namespace *ns, struct ipc_params *params)
        if (IS_ERR(file))
                goto no_file;
 
-       id = ipc_addid(&shm_ids(ns), &shp->shm_perm, ns->shm_ctlmni);
-       if (id < 0) {
-               error = id;
-               goto no_id;
-       }
-
        shp->shm_cprid = task_tgid_vnr(current);
        shp->shm_lprid = 0;
        shp->shm_atim = shp->shm_dtim = 0;
@@ -404,6 +505,14 @@ static int newseg(struct ipc_namespace *ns, struct ipc_params *params)
        shp->shm_segsz = size;
        shp->shm_nattch = 0;
        shp->shm_file = file;
+       shp->shm_creator = current;
+
+       id = ipc_addid(&shm_ids(ns), &shp->shm_perm, ns->shm_ctlmni);
+       if (id < 0) {
+               error = id;
+               goto no_id;
+       }
+
        /*
         * shmid gets reported as "inode#" in /proc/pid/maps.
         * proc-ps tools use this. Changing this will break them.
@@ -762,9 +871,7 @@ SYSCALL_DEFINE3(shmctl, int, shmid, int, cmd, struct shmid_ds __user *, buf)
        case SHM_LOCK:
        case SHM_UNLOCK:
        {
-               struct file *uninitialized_var(shm_file);
-
-               lru_add_drain_all();  /* drain pagevecs to lru lists */
+               struct file *shm_file;
 
                shp = shm_lock_check(ns, shmid);
                if (IS_ERR(shp)) {
@@ -787,22 +894,31 @@ SYSCALL_DEFINE3(shmctl, int, shmid, int, cmd, struct shmid_ds __user *, buf)
                err = security_shm_shmctl(shp, cmd);
                if (err)
                        goto out_unlock;
-               
-               if(cmd==SHM_LOCK) {
+
+               shm_file = shp->shm_file;
+               if (is_file_hugepages(shm_file))
+                       goto out_unlock;
+
+               if (cmd == SHM_LOCK) {
                        struct user_struct *user = current_user();
-                       if (!is_file_hugepages(shp->shm_file)) {
-                               err = shmem_lock(shp->shm_file, 1, user);
-                               if (!err && !(shp->shm_perm.mode & SHM_LOCKED)){
-                                       shp->shm_perm.mode |= SHM_LOCKED;
-                                       shp->mlock_user = user;
-                               }
+                       err = shmem_lock(shm_file, 1, user);
+                       if (!err && !(shp->shm_perm.mode & SHM_LOCKED)) {
+                               shp->shm_perm.mode |= SHM_LOCKED;
+                               shp->mlock_user = user;
                        }
-               } else if (!is_file_hugepages(shp->shm_file)) {
-                       shmem_lock(shp->shm_file, 0, shp->mlock_user);
-                       shp->shm_perm.mode &= ~SHM_LOCKED;
-                       shp->mlock_user = NULL;
+                       goto out_unlock;
                }
+
+               /* SHM_UNLOCK */
+               if (!(shp->shm_perm.mode & SHM_LOCKED))
+                       goto out_unlock;
+               shmem_lock(shm_file, 0, shp->mlock_user);
+               shp->shm_perm.mode &= ~SHM_LOCKED;
+               shp->mlock_user = NULL;
+               get_file(shm_file);
                shm_unlock(shp);
+               shmem_unlock_mapping(shm_file->f_mapping);
+               fput(shm_file);
                goto out;
        }
        case IPC_RMID:
@@ -847,8 +963,13 @@ long do_shmat(int shmid, char __user *shmaddr, int shmflg, ulong *raddr)
                goto out;
        else if ((addr = (ulong)shmaddr)) {
                if (addr & (SHMLBA-1)) {
-                       if (shmflg & SHM_RND)
-                               addr &= ~(SHMLBA-1);       /* round down */
+                       /*
+                        * Round down to the nearest multiple of shmlba.
+                        * For sane do_mmap_pgoff() parameters, avoid
+                        * round downs that trigger nil-page and MAP_FIXED.
+                        */
+                       if ((shmflg & SHM_RND) && addr >= SHMLBA)
+                               addr &= ~(SHMLBA - 1);
                        else
 #ifndef __ARCH_FORCE_SHMLBA
                                if (addr & ~PAGE_MASK)
@@ -950,8 +1071,7 @@ out_nattch:
        shp = shm_lock(ns, shmid);
        BUG_ON(IS_ERR(shp));
        shp->shm_nattch--;
-       if(shp->shm_nattch == 0 &&
-          shp->shm_perm.mode & SHM_DEST)
+       if (shm_may_destroy(ns, shp))
                shm_destroy(ns, shp);
        else
                shm_unlock(shp);