powerpc/pseries/vas: Hold mmap_mutex after mmap lock during window close

[ Upstream commit b59c9dc ]

Commit 8ef7b9e ("powerpc/pseries/vas: Close windows with DLPAR
core removal") unmaps the window paste address and issues HCALL to
close window in the hypervisor for migration or DLPAR core removal
events. So holds mmap_mutex and then mmap lock before unmap the
paste address. But if the user space issue mmap paste address at
the same time with the migration event, coproc_mmap() is called
after holding the mmap lock which can trigger deadlock when trying
to acquire mmap_mutex in coproc_mmap().

t1: mmap() call to mmap              t2: Migration event
    window paste address

do_mmap2()                           migration_store()
 ksys_mmap_pgoff()                    pseries_migrate_partition()
  vm_mmap_pgoff()                      vas_migration_handler()
    Acquire mmap lock                   reconfig_close_windows()
    do_mmap()                             lock mmap_mutex
     mmap_region()                        Acquire mmap lock
      call_mmap()                         //Wait for mmap lock
       coproc_mmap()                        unmap vma
         lock mmap_mutex                    update window status
         //wait for mmap_mutex            Release mmap lock
          mmap vma                        unlock mmap_mutex
          update window status
         unlock mmap_mutex
    ...
    Release mmap lock

Fix this deadlock issue by holding mmap lock first before mmap_mutex
in reconfig_close_windows().

Fixes: 8ef7b9e ("powerpc/pseries/vas: Close windows with DLPAR core removal")
Signed-off-by: Haren Myneni <haren@linux.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://msgid.link/20230716100506.7833-1-haren@linux.ibm.com
Signed-off-by: Sasha Levin <sashal@kernel.org>

arch/powerpc/platforms/pseries/vas.c

@@ -744,6 +744,12 @@ static int reconfig_close_windows(struct vas_caps *vcap, int excess_creds,
 		}
 
 		task_ref = &win->vas_win.task_ref;
+		/*
+		 * VAS mmap (coproc_mmap()) and its fault handler
+		 * (vas_mmap_fault()) are called after holding mmap lock.
+		 * So hold mmap mutex after mmap_lock to avoid deadlock.
+		 */
+		mmap_write_lock(task_ref->mm);
 		mutex_lock(&task_ref->mmap_mutex);
 		vma = task_ref->vma;
 		/*
@@ -752,7 +758,6 @@ static int reconfig_close_windows(struct vas_caps *vcap, int excess_creds,
 		 */
 		win->vas_win.status |= flag;
 
-		mmap_write_lock(task_ref->mm);
 		/*
 		 * vma is set in the original mapping. But this mapping
 		 * is done with mmap() after the window is opened with ioctl.
@@ -763,8 +768,8 @@ static int reconfig_close_windows(struct vas_caps *vcap, int excess_creds,
 			zap_page_range(vma, vma->vm_start,
 					vma->vm_end - vma->vm_start);
 
-		mmap_write_unlock(task_ref->mm);
 		mutex_unlock(&task_ref->mmap_mutex);
+		mmap_write_unlock(task_ref->mm);
 		/*
 		 * Close VAS window in the hypervisor, but do not
 		 * free vas_window struct since it may be reused