mm/sharepool: Fix add group failed with errno 28

ascend inclusion
category: bugfix
bugzilla: https://gitee.com/openeuler/kernel/issues/I612UG


CVE: NA

--------------------------------

We increase task->mm->mm_users by one when we add the task to a
sharepool group. Correspondingly we should drop the mm_users count when
the task exits. Currently we hijack the mmput function and make it
return early and decrease mm->mm_users by one (just as mmput would do)
if it is not called from a task's exiting process, or we decrease
mm->mm_users by the group number the task was added to. This has two
problems:
1. It makes mmput and sp_group_exit hard to understand.
2. The process of judging if the task (also its mm) is exiting and
   decrease its mm_users count is not atomic. We use this condition:
     mm->mm_users == master->count + MM_WOULD_FREE(1)
   If someone else change the mm->mm_users during those two steps, the
   mm->mm_users would be wrong and mm_struct cannot be released anymore.

Suppose the following process:

        proc1                                        proc2

1)      mmput
          |
          V
2)  enter sp_group_exit and
    'mm->mm_users == master->count + 1' is true
3)        |                                         mmget
          V
4)  decrease mm->mm_users by master->count
          |
          V
5)  enter __mmput and release mm_struct
    if mm->mm_users == 1
6)                                                  mmput

The statistical structure who has the same id of the task would get leaked
together with mm_struct, so the next time we try to create the statistical
structure of the same id, we get a failure.

We fix this by moving sp_group_exit to do_exit() actually where the task is
exiting. We don't need to judge if the task is exiting when someone
calling mmput so there is no chance to change mm_users wrongly.

Signed-off-by: Xu Qiang <xuqiang36@huawei.com>
Signed-off-by: Wang Wensheng <wangwensheng4@huawei.com>