8ae002
commit b3a810d0d3d5c6ce7ddfb61321cd7971808ca703
8ae002
Author: Stefan Liebler <stli@linux.vnet.ibm.com>
8ae002
Date:   Tue May 17 10:45:48 2016 +0200
8ae002
8ae002
    Fix tst-cancel17/tst-cancelx17, which sometimes segfaults while exiting.
8ae002
    
8ae002
    The testcase tst-cancel[x]17 ends sometimes with a segmentation fault.
8ae002
    This happens in one of 10000 cases. Then the real testcase has already
8ae002
    exited with success and returned from do_test(). The segmentation fault
8ae002
    occurs after returning from main in _dl_fini().
8ae002
    
8ae002
    In those cases, the aio_read(&a) was not canceled because the read
8ae002
    request was already in progress. In the meanwhile aio_write(ap) wrote
8ae002
    something to the pipe and the read request is able to read the
8ae002
    requested byte.
8ae002
    The read request hasn't finished before returning from do_test().
8ae002
    After it finishes, it writes the return value and error code from the
8ae002
    read syscall to the struct aiocb a, which lies on the stack of do_test.
8ae002
    The stack of the subsequent function call of _dl_fini or _dl_sort_fini,
8ae002
    which is inlined in _dl_fini is corrupted.
8ae002
    
8ae002
    In case of S390, it reads a zero and decrements it by 1:
8ae002
    unsigned int k = nmaps - 1;
8ae002
    struct link_map **runp = maps[k]->l_initfini;
8ae002
    The load from unmapped memory leads to the segmentation fault.
8ae002
    The stack corruption also happens on other architectures.
8ae002
    I saw them e.g. on x86 and ppc, too.
8ae002
    
8ae002
    This patch adds an aio_suspend call to ensure, that the read request
8ae002
    is finished before returning from do_test().
8ae002
    
8ae002
    ChangeLog:
8ae002
    
8ae002
    	* nptl/tst-cancel17.c (do_test): Wait for finishing aio_read(&a).
8ae002
8ae002
diff --git a/nptl/tst-cancel17.c b/nptl/tst-cancel17.c
8ae002
index fb89292..eedd28e 100644
8ae002
--- a/nptl/tst-cancel17.c
8ae002
+++ b/nptl/tst-cancel17.c
8ae002
@@ -333,6 +333,22 @@ do_test (void)
8ae002
 
8ae002
   puts ("early cancellation succeeded");
8ae002
 
8ae002
+  if (ap == &a2)
8ae002
+    {
8ae002
+      /* The aio_read(&a) was not canceled because the read request was
8ae002
+	 already in progress. In the meanwhile aio_write(ap) wrote something
8ae002
+	 to the pipe and the read request either has already been finished or
8ae002
+	 is able to read the requested byte.
8ae002
+	 Wait for the read request before returning from this function because
8ae002
+	 the return value and error code from the read syscall will be written
8ae002
+	 to the struct aiocb a, which lies on the stack of this function.
8ae002
+	 Otherwise the stack from subsequent function calls - e.g. _dl_fini -
8ae002
+	 will be corrupted, which can lead to undefined behaviour like a
8ae002
+	 segmentation fault.  */
8ae002
+      const struct aiocb *l[1] = { &a };
8ae002
+      TEMP_FAILURE_RETRY (aio_suspend(l, 1, NULL));
8ae002
+    }
8ae002
+
8ae002
   return 0;
8ae002
 }
8ae002