prctl: avoid using mmap_sem for exe_file serialization
authorDavidlohr Bueso <dave@stgolabs.net>
Thu, 16 Apr 2015 19:47:59 +0000 (12:47 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Fri, 17 Apr 2015 13:04:07 +0000 (09:04 -0400)
Oleg cleverly suggested using xchg() to set the new mm->exe_file instead
of calling set_mm_exe_file() which requires some form of serialization --
mmap_sem in this case.  For archs that do not have atomic rmw instructions
we still fallback to a spinlock alternative, so this should always be
safe.  As such, we only need the mmap_sem for looking up the backing
vm_file, which can be done sharing the lock.  Naturally, this means we
need to manually deal with both the new and old file reference counting,
and we need not worry about the MMF_EXE_FILE_CHANGED bits, which can
probably be deleted in the future anyway.

Signed-off-by: Davidlohr Bueso <dbueso@suse.de>
Suggested-by: Oleg Nesterov <oleg@redhat.com>
Acked-by: Oleg Nesterov <oleg@redhat.com>
Reviewed-by: Konstantin Khlebnikov <khlebnikov@yandex-team.ru>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
fs/exec.c
kernel/fork.c
kernel/sys.c

index c7f9b733406dc22165647a8f415ce00324ae7daf..a5fef835ebc59fe1dab0311dab0f4fe427894b50 100644 (file)
--- a/fs/exec.c
+++ b/fs/exec.c
@@ -1078,7 +1078,13 @@ int flush_old_exec(struct linux_binprm * bprm)
        if (retval)
                goto out;
 
+       /*
+        * Must be called _before_ exec_mmap() as bprm->mm is
+        * not visibile until then. This also enables the update
+        * to be lockless.
+        */
        set_mm_exe_file(bprm->mm, bprm->file);
+
        /*
         * Release all of the old mmap stuff
         */
index 259202637531e9b10f3a985f41bd2490afd23c54..0d23e76a0c6155540009bfa3fe072973d6ca7530 100644 (file)
@@ -711,15 +711,22 @@ EXPORT_SYMBOL_GPL(mmput);
  *
  * This changes mm's executable file (shown as symlink /proc/[pid]/exe).
  *
- * Main users are mmput(), sys_execve() and sys_prctl(PR_SET_MM_MAP/EXE_FILE).
- * Callers prevent concurrent invocations: in mmput() nobody alive left,
- * in execve task is single-threaded, prctl holds mmap_sem exclusively.
+ * Main users are mmput() and sys_execve(). Callers prevent concurrent
+ * invocations: in mmput() nobody alive left, in execve task is single
+ * threaded. sys_prctl(PR_SET_MM_MAP/EXE_FILE) also needs to set the
+ * mm->exe_file, but does so without using set_mm_exe_file() in order
+ * to do avoid the need for any locks.
  */
 void set_mm_exe_file(struct mm_struct *mm, struct file *new_exe_file)
 {
-       struct file *old_exe_file = rcu_dereference_protected(mm->exe_file,
-                       !atomic_read(&mm->mm_users) || current->in_execve ||
-                       lockdep_is_held(&mm->mmap_sem));
+       struct file *old_exe_file;
+
+       /*
+        * It is safe to dereference the exe_file without RCU as
+        * this function is only called if nobody else can access
+        * this mm -- see comment above for justification.
+        */
+       old_exe_file = rcu_dereference_raw(mm->exe_file);
 
        if (new_exe_file)
                get_file(new_exe_file);
index 3be344902316e93d8b147d1c5367dc60fe5908b8..a4e372b798a5f29535f9120b32fea70b0489f603 100644 (file)
@@ -1649,14 +1649,13 @@ SYSCALL_DEFINE1(umask, int, mask)
        return mask;
 }
 
-static int prctl_set_mm_exe_file_locked(struct mm_struct *mm, unsigned int fd)
+static int prctl_set_mm_exe_file(struct mm_struct *mm, unsigned int fd)
 {
        struct fd exe;
+       struct file *old_exe, *exe_file;
        struct inode *inode;
        int err;
 
-       VM_BUG_ON_MM(!rwsem_is_locked(&mm->mmap_sem), mm);
-
        exe = fdget(fd);
        if (!exe.file)
                return -EBADF;
@@ -1680,15 +1679,22 @@ static int prctl_set_mm_exe_file_locked(struct mm_struct *mm, unsigned int fd)
        /*
         * Forbid mm->exe_file change if old file still mapped.
         */
+       exe_file = get_mm_exe_file(mm);
        err = -EBUSY;
-       if (mm->exe_file) {
+       if (exe_file) {
                struct vm_area_struct *vma;
 
-               for (vma = mm->mmap; vma; vma = vma->vm_next)
-                       if (vma->vm_file &&
-                           path_equal(&vma->vm_file->f_path,
-                                      &mm->exe_file->f_path))
-                               goto exit;
+               down_read(&mm->mmap_sem);
+               for (vma = mm->mmap; vma; vma = vma->vm_next) {
+                       if (!vma->vm_file)
+                               continue;
+                       if (path_equal(&vma->vm_file->f_path,
+                                      &exe_file->f_path))
+                               goto exit_err;
+               }
+
+               up_read(&mm->mmap_sem);
+               fput(exe_file);
        }
 
        /*
@@ -1702,10 +1708,18 @@ static int prctl_set_mm_exe_file_locked(struct mm_struct *mm, unsigned int fd)
                goto exit;
 
        err = 0;
-       set_mm_exe_file(mm, exe.file);  /* this grabs a reference to exe.file */
+       /* set the new file, lockless */
+       get_file(exe.file);
+       old_exe = xchg(&mm->exe_file, exe.file);
+       if (old_exe)
+               fput(old_exe);
 exit:
        fdput(exe);
        return err;
+exit_err:
+       up_read(&mm->mmap_sem);
+       fput(exe_file);
+       goto exit;
 }
 
 #ifdef CONFIG_CHECKPOINT_RESTORE
@@ -1840,10 +1854,9 @@ static int prctl_set_mm_map(int opt, const void __user *addr, unsigned long data
                user_auxv[AT_VECTOR_SIZE - 1] = AT_NULL;
        }
 
-       down_write(&mm->mmap_sem);
        if (prctl_map.exe_fd != (u32)-1)
-               error = prctl_set_mm_exe_file_locked(mm, prctl_map.exe_fd);
-       downgrade_write(&mm->mmap_sem);
+               error = prctl_set_mm_exe_file(mm, prctl_map.exe_fd);
+       down_read(&mm->mmap_sem);
        if (error)
                goto out;
 
@@ -1909,12 +1922,8 @@ static int prctl_set_mm(int opt, unsigned long addr,
        if (!capable(CAP_SYS_RESOURCE))
                return -EPERM;
 
-       if (opt == PR_SET_MM_EXE_FILE) {
-               down_write(&mm->mmap_sem);
-               error = prctl_set_mm_exe_file_locked(mm, (unsigned int)addr);
-               up_write(&mm->mmap_sem);
-               return error;
-       }
+       if (opt == PR_SET_MM_EXE_FILE)
+               return prctl_set_mm_exe_file(mm, (unsigned int)addr);
 
        if (addr >= TASK_SIZE || addr < mmap_min_addr)
                return -EINVAL;