]> err.no Git - linux-2.6/blobdiff - fs/exec.c
coredump: move mm->core_waiters into struct core_state
[linux-2.6] / fs / exec.c
index 190ed1f927740fd1dc9a3437a6081dd213be4801..50de3aaff4d008921a62fb54ea4cc809babeb2d7 100644 (file)
--- a/fs/exec.c
+++ b/fs/exec.c
 #include <linux/slab.h>
 #include <linux/file.h>
 #include <linux/fdtable.h>
-#include <linux/mman.h>
+#include <linux/mm.h>
 #include <linux/stat.h>
 #include <linux/fcntl.h>
 #include <linux/smp_lock.h>
+#include <linux/swap.h>
 #include <linux/string.h>
 #include <linux/init.h>
-#include <linux/pagemap.h>
 #include <linux/highmem.h>
 #include <linux/spinlock.h>
 #include <linux/key.h>
 #include <linux/personality.h>
 #include <linux/binfmts.h>
-#include <linux/swap.h>
 #include <linux/utsname.h>
 #include <linux/pid_namespace.h>
 #include <linux/module.h>
@@ -47,7 +46,6 @@
 #include <linux/mount.h>
 #include <linux/security.h>
 #include <linux/syscalls.h>
-#include <linux/rmap.h>
 #include <linux/tsacct_kern.h>
 #include <linux/cn_proc.h>
 #include <linux/audit.h>
@@ -724,12 +722,10 @@ static int exec_mmap(struct mm_struct *mm)
                 * Make sure that if there is a core dump in progress
                 * for the old mm, we get out and die instead of going
                 * through with the exec.  We must hold mmap_sem around
-                * checking core_waiters and changing tsk->mm.  The
-                * core-inducing thread will increment core_waiters for
-                * each thread whose ->mm == old_mm.
+                * checking core_state and changing tsk->mm.
                 */
                down_read(&old_mm->mmap_sem);
-               if (unlikely(old_mm->core_waiters)) {
+               if (unlikely(old_mm->core_state)) {
                        up_read(&old_mm->mmap_sem);
                        return -EINTR;
                }
@@ -1328,6 +1324,7 @@ int do_execve(char * filename,
        if (retval < 0)
                goto out;
 
+       current->flags &= ~PF_KTHREAD;
        retval = search_binary_handler(bprm,regs);
        if (retval >= 0) {
                /* execve success */
@@ -1515,11 +1512,11 @@ static void zap_process(struct task_struct *start)
        t = start;
        do {
                if (t != current && t->mm) {
-                       t->mm->core_waiters++;
+                       t->mm->core_state->nr_threads++;
                        sigaddset(&t->pending.signal, SIGKILL);
                        signal_wake_up(t, 1);
                }
-       } while ((t = next_thread(t)) != start);
+       } while_each_thread(start, t);
 }
 
 static inline int zap_threads(struct task_struct *tsk, struct mm_struct *mm,
@@ -1539,48 +1536,77 @@ static inline int zap_threads(struct task_struct *tsk, struct mm_struct *mm,
        if (err)
                return err;
 
-       if (atomic_read(&mm->mm_users) == mm->core_waiters + 1)
+       if (atomic_read(&mm->mm_users) == mm->core_state->nr_threads + 1)
                goto done;
-
+       /*
+        * We should find and kill all tasks which use this mm, and we should
+        * count them correctly into ->nr_threads. We don't take tasklist
+        * lock, but this is safe wrt:
+        *
+        * fork:
+        *      None of sub-threads can fork after zap_process(leader). All
+        *      processes which were created before this point should be
+        *      visible to zap_threads() because copy_process() adds the new
+        *      process to the tail of init_task.tasks list, and lock/unlock
+        *      of ->siglock provides a memory barrier.
+        *
+        * do_exit:
+        *      The caller holds mm->mmap_sem. This means that the task which
+        *      uses this mm can't pass exit_mm(), so it can't exit or clear
+        *      its ->mm.
+        *
+        * de_thread:
+        *      It does list_replace_rcu(&leader->tasks, &current->tasks),
+        *      we must see either old or new leader, this does not matter.
+        *      However, it can change p->sighand, so lock_task_sighand(p)
+        *      must be used. Since p->mm != NULL and we hold ->mmap_sem
+        *      it can't fail.
+        *
+        *      Note also that "g" can be the old leader with ->mm == NULL
+        *      and already unhashed and thus removed from ->thread_group.
+        *      This is OK, __unhash_process()->list_del_rcu() does not
+        *      clear the ->next pointer, we will find the new leader via
+        *      next_thread().
+        */
        rcu_read_lock();
        for_each_process(g) {
                if (g == tsk->group_leader)
                        continue;
-
+               if (g->flags & PF_KTHREAD)
+                       continue;
                p = g;
                do {
                        if (p->mm) {
-                               if (p->mm == mm) {
-                                       /*
-                                        * p->sighand can't disappear, but
-                                        * may be changed by de_thread()
-                                        */
+                               if (unlikely(p->mm == mm)) {
                                        lock_task_sighand(p, &flags);
                                        zap_process(p);
                                        unlock_task_sighand(p, &flags);
                                }
                                break;
                        }
-               } while ((p = next_thread(p)) != g);
+               } while_each_thread(g, p);
        }
        rcu_read_unlock();
 done:
-       return mm->core_waiters;
+       return mm->core_state->nr_threads;
 }
 
 static int coredump_wait(int exit_code)
 {
        struct task_struct *tsk = current;
        struct mm_struct *mm = tsk->mm;
-       struct completion startup_done;
+       struct core_state core_state;
        struct completion *vfork_done;
        int core_waiters;
 
        init_completion(&mm->core_done);
-       init_completion(&startup_done);
-       mm->core_startup_done = &startup_done;
+       init_completion(&core_state.startup);
+       core_state.nr_threads = 0;
+       mm->core_state = &core_state;
 
        core_waiters = zap_threads(tsk, mm, exit_code);
+       if (core_waiters < 0)
+               mm->core_state = NULL;
        up_write(&mm->mmap_sem);
 
        if (unlikely(core_waiters < 0))
@@ -1597,9 +1623,9 @@ static int coredump_wait(int exit_code)
        }
 
        if (core_waiters)
-               wait_for_completion(&startup_done);
+               wait_for_completion(&core_state.startup);
+       mm->core_state = NULL;
 fail:
-       BUG_ON(mm->core_waiters);
        return core_waiters;
 }
 
@@ -1677,7 +1703,7 @@ int do_coredump(long signr, int exit_code, struct pt_regs * regs)
        /*
         * If another thread got here first, or we are not dumpable, bail out.
         */
-       if (mm->core_waiters || !get_dumpable(mm)) {
+       if (mm->core_state || !get_dumpable(mm)) {
                up_write(&mm->mmap_sem);
                goto fail;
        }