xen: avoid allocation causing potential swap activity on the resume path
[safe/jmp/linux-2.6] / kernel / exit.c
index 6f50ef5..ceffc67 100644 (file)
 #include <asm/unistd.h>
 #include <asm/pgtable.h>
 #include <asm/mmu_context.h>
-#include "cred-internals.h"
 
 static void exit_mm(struct task_struct * tsk);
 
-static void __unhash_process(struct task_struct *p)
+static void __unhash_process(struct task_struct *p, bool group_dead)
 {
        nr_threads--;
        detach_pid(p, PIDTYPE_PID);
-       if (thread_group_leader(p)) {
+       if (group_dead) {
                detach_pid(p, PIDTYPE_PGID);
                detach_pid(p, PIDTYPE_SID);
 
                list_del_rcu(&p->tasks);
+               list_del_init(&p->sibling);
                __get_cpu_var(process_counts)--;
        }
        list_del_rcu(&p->thread_group);
-       list_del_init(&p->sibling);
 }
 
 /*
@@ -80,23 +79,26 @@ static void __unhash_process(struct task_struct *p)
 static void __exit_signal(struct task_struct *tsk)
 {
        struct signal_struct *sig = tsk->signal;
+       bool group_dead = thread_group_leader(tsk);
        struct sighand_struct *sighand;
+       struct tty_struct *uninitialized_var(tty);
 
-       BUG_ON(!sig);
-       BUG_ON(!atomic_read(&sig->count));
-
-       sighand = rcu_dereference(tsk->sighand);
+       sighand = rcu_dereference_check(tsk->sighand,
+                                       rcu_read_lock_held() ||
+                                       lockdep_tasklist_lock_is_held());
        spin_lock(&sighand->siglock);
 
        posix_cpu_timers_exit(tsk);
-       if (atomic_dec_and_test(&sig->count))
+       if (group_dead) {
                posix_cpu_timers_exit_group(tsk);
-       else {
+               tty = sig->tty;
+               sig->tty = NULL;
+       } else {
                /*
                 * If there is any task waiting for the group exit
                 * then notify it:
                 */
-               if (sig->group_exit_task && atomic_read(&sig->count) == sig->notify_count)
+               if (sig->notify_count > 0 && !--sig->notify_count)
                        wake_up_process(sig->group_exit_task);
 
                if (tsk == sig->curr_target)
@@ -122,32 +124,24 @@ static void __exit_signal(struct task_struct *tsk)
                sig->oublock += task_io_get_oublock(tsk);
                task_io_accounting_add(&sig->ioac, &tsk->ioac);
                sig->sum_sched_runtime += tsk->se.sum_exec_runtime;
-               sig = NULL; /* Marker for below. */
        }
 
-       __unhash_process(tsk);
+       sig->nr_threads--;
+       __unhash_process(tsk, group_dead);
 
        /*
         * Do this under ->siglock, we can race with another thread
         * doing sigqueue_free() if we have SIGQUEUE_PREALLOC signals.
         */
        flush_sigqueue(&tsk->pending);
-
-       tsk->signal = NULL;
        tsk->sighand = NULL;
        spin_unlock(&sighand->siglock);
 
        __cleanup_sighand(sighand);
        clear_tsk_thread_flag(tsk,TIF_SIGPENDING);
-       if (sig) {
+       if (group_dead) {
                flush_sigqueue(&sig->shared_pending);
-               taskstats_tgid_free(sig);
-               /*
-                * Make sure ->signal can't go away under rq->lock,
-                * see account_group_exec_runtime().
-                */
-               task_rq_unlock_wait(tsk);
-               __cleanup_signal(sig);
+               tty_kref_put(tty);
        }
 }
 
@@ -170,8 +164,10 @@ void release_task(struct task_struct * p)
 repeat:
        tracehook_prepare_release_task(p);
        /* don't need to get the RCU readlock here - the process is dead and
-        * can't be modifying its own credentials */
+        * can't be modifying its own credentials. But shut RCU-lockdep up */
+       rcu_read_lock();
        atomic_dec(&__task_cred(p)->user->processes);
+       rcu_read_unlock();
 
        proc_flush_task(p);
 
@@ -473,9 +469,11 @@ static void close_files(struct files_struct * files)
        /*
         * It is safe to dereference the fd table without RCU or
         * ->file_lock because this is the last reference to the
-        * files structure.
+        * files structure.  But use RCU to shut RCU-lockdep up.
         */
+       rcu_read_lock();
        fdt = files_fdtable(files);
+       rcu_read_unlock();
        for (;;) {
                unsigned long set;
                i = j * __NFDBITS;
@@ -521,10 +519,12 @@ void put_files_struct(struct files_struct *files)
                 * at the end of the RCU grace period. Otherwise,
                 * you can free files immediately.
                 */
+               rcu_read_lock();
                fdt = files_fdtable(files);
                if (fdt != &files->fdtab)
                        kmem_cache_free(files_cachep, files);
                free_fdtable(fdt);
+               rcu_read_unlock();
        }
 }
 
@@ -736,12 +736,9 @@ static struct task_struct *find_new_reaper(struct task_struct *father)
 /*
 * Any that need to be release_task'd are put on the @dead list.
  */
-static void reparent_thread(struct task_struct *father, struct task_struct *p,
+static void reparent_leader(struct task_struct *father, struct task_struct *p,
                                struct list_head *dead)
 {
-       if (p->pdeath_signal)
-               group_send_sig_info(p->pdeath_signal, SEND_SIG_NOINFO, p);
-
        list_move_tail(&p->sibling, &p->real_parent->children);
 
        if (task_detached(p))
@@ -780,12 +777,18 @@ static void forget_original_parent(struct task_struct *father)
        reaper = find_new_reaper(father);
 
        list_for_each_entry_safe(p, n, &father->children, sibling) {
-               p->real_parent = reaper;
-               if (p->parent == father) {
-                       BUG_ON(task_ptrace(p));
-                       p->parent = p->real_parent;
-               }
-               reparent_thread(father, p, &dead_children);
+               struct task_struct *t = p;
+               do {
+                       t->real_parent = reaper;
+                       if (t->parent == father) {
+                               BUG_ON(task_ptrace(t));
+                               t->parent = t->real_parent;
+                       }
+                       if (t->pdeath_signal)
+                               group_send_sig_info(t->pdeath_signal,
+                                                   SEND_SIG_NOINFO, t);
+               } while_each_thread(p, t);
+               reparent_leader(father, p, &dead_children);
        }
        write_unlock_irq(&tasklist_lock);
 
@@ -846,12 +849,9 @@ static void exit_notify(struct task_struct *tsk, int group_dead)
 
        tsk->exit_state = signal == DEATH_REAP ? EXIT_DEAD : EXIT_ZOMBIE;
 
-       /* mt-exec, de_thread() is waiting for us */
-       if (thread_group_leader(tsk) &&
-           tsk->signal->group_exit_task &&
-           tsk->signal->notify_count < 0)
+       /* mt-exec, de_thread() is waiting for group leader */
+       if (unlikely(tsk->signal->notify_count < 0))
                wake_up_process(tsk->signal->group_exit_task);
-
        write_unlock_irq(&tasklist_lock);
 
        tracehook_report_death(tsk, signal, cookie, group_dead);
@@ -933,7 +933,7 @@ NORET_TYPE void do_exit(long code)
         * an exiting task cleaning up the robust pi futexes.
         */
        smp_mb();
-       spin_unlock_wait(&tsk->pi_lock);
+       raw_spin_unlock_wait(&tsk->pi_lock);
 
        if (unlikely(in_atomic()))
                printk(KERN_INFO "note: %s[%d] exited with preempt_count %d\n",
@@ -941,7 +941,9 @@ NORET_TYPE void do_exit(long code)
                                preempt_count());
 
        acct_update_integrals(tsk);
-
+       /* sync mm's RSS info before statistics gathering */
+       if (tsk->mm)
+               sync_mm_rss(tsk, tsk->mm);
        group_dead = atomic_dec_and_test(&tsk->signal->live);
        if (group_dead) {
                hrtimer_cancel(&tsk->signal->real_timer);
@@ -990,8 +992,10 @@ NORET_TYPE void do_exit(long code)
 
        exit_notify(tsk, group_dead);
 #ifdef CONFIG_NUMA
+       task_lock(tsk);
        mpol_put(tsk->mempolicy);
        tsk->mempolicy = NULL;
+       task_unlock(tsk);
 #endif
 #ifdef CONFIG_FUTEX
        if (unlikely(current->pi_state_cache))
@@ -1177,7 +1181,7 @@ static int wait_task_zombie(struct wait_opts *wo, struct task_struct *p)
 
        if (unlikely(wo->wo_flags & WNOWAIT)) {
                int exit_code = p->exit_code;
-               int why, status;
+               int why;
 
                get_task_struct(p);
                read_unlock(&tasklist_lock);
@@ -1551,14 +1555,9 @@ static int do_wait_thread(struct wait_opts *wo, struct task_struct *tsk)
        struct task_struct *p;
 
        list_for_each_entry(p, &tsk->children, sibling) {
-               /*
-                * Do not consider detached threads.
-                */
-               if (!task_detached(p)) {
-                       int ret = wait_consider_task(wo, 0, p);
-                       if (ret)
-                               return ret;
-               }
+               int ret = wait_consider_task(wo, 0, p);
+               if (ret)
+                       return ret;
        }
 
        return 0;