X print wcfs.py py-level traceback on SIGBUS (e.g. wcfs.go aborting due to bug/panic)

Needs https://github.com/cython/cython/pull/2997

wcfs/internal/wcfs_test.pyx

@@ -21,9 +21,68 @@
 
 """Module wcfs_test complements wcfs_test.py with things that are impossible to implement in Python"""
 
+from posix.signal cimport sigaction, sigaction_t, siginfo_t, SA_SIGINFO
+from libc.signal cimport SIGBUS
+from libc.stdlib cimport abort
+from libc.string cimport strlen
+from posix.unistd cimport write, sleep
+
+from cpython.exc cimport PyErr_SetFromErrno
+from cpython.pystate cimport PyGILState_Ensure, PyGILState_Release, PyGILState_STATE
+
+# XXX -> cpython.lifecycle? .run ?
+cdef extern from "Python.h":
+    int PyRun_SimpleString(const char *)
+
 # read0_nogil accesses mem[0] with GIL released and returns first byte read from there.
 def read0_nogil(const unsigned char[::1] mem not None) -> int:
     cdef unsigned char b
     with nogil:
         b = mem[0]
     return b
+
+
+# ---- signal handling ----
+# XXX -> golang.signal ?
+
+# install_sigbus_trap installs SIGBUS handler that prints python-level
+# traceback before aborting.
+#
+# Such handler is useful, because when wcfs.go bugs/panics while handling file
+# access from wcfs.py, wcfs.py receives SIGBUS signal and by default aborts.
+def install_sigbus_trap():
+    cdef sigaction_t act
+    act.sa_sigaction = on_sigbus
+    act.sa_flags     = SA_SIGINFO
+
+    cdef int err = sigaction(SIGBUS, &act, NULL)
+    if err:
+        PyErr_SetFromErrno(OSError)
+
+cdef void on_sigbus(int sig, siginfo_t *si, void *_uc):
+    # - wait a bit to give time for other threads to complete their exception dumps
+    #   (e.g. getting "Transport endpoint is not connected" after wcfs.go dying)
+    # - dump py-level traceback and abort.
+    # TODO turn SIGBUS into python-level exception? (see sigpanic in Go how to do).
+
+    writeerr("\nC: SIGBUS received; giving time to other threads" +
+             "to dump their exceptions (if any) ...\n")
+    cdef PyGILState_STATE gstate = PyGILState_Ensure()
+    PyGILState_Release(gstate)
+    sleep(1)
+
+    writeerr("\nC: SIGBUS thread traceback:\n")
+    PyGILState_Ensure()
+    PyRun_SimpleString("import traceback; traceback.print_stack()")
+    writeerr("-> SIGBUS\n");
+    abort()
+
+# writeerr writes msg to stderr without depending on stdio buffering and locking.
+cdef void writeerr(const char *msg):
+    cdef ssize_t n, left = strlen(msg)
+    while left > 0:
+        n = write(2, msg, left)
+        if n == -1:
+            return  # nothing we can do under crash?
+        left -= n
+        msg  += n

wcfs/wcfs_test.py

@@ -42,7 +42,7 @@ from zodbtools.util import ashex as h, fromhex
 from pytest import raises
 from six import reraise
 from .internal import mm
-from .internal.wcfs_test import read0_nogil
+from .internal.wcfs_test import read0_nogil, install_sigbus_trap
 
 # setup:
 # - create test database, compute zurl and mountpoint for wcfs
@@ -52,6 +52,10 @@ testdb = None
 testzurl = None     # URL of testdb
 testmntpt = None    # wcfs is mounted here
 def setup_module():
+    # if wcfs.py receives SIGBUS because wcfs.go panics, we want to see
+    # python-level traceback instead of being killed.
+    install_sigbus_trap()
+
     global testdb, testzurl, testmntpt
     testdb = getTestDB()
     testdb.setup()
@@ -1172,7 +1176,7 @@ def test_wcfs():
 
     # >>> XXX commit data to not yet accessed f part - nothing happens
 
-    """
+#   """
     # >>> invalidation protocol
     print('\n\n inv. protocol \n\n')
 
@@ -1238,7 +1242,7 @@ def test_wcfs():
             for at in revv[1:]:
                 wl.watch(zf, at)
             wl.close()
-    """
+#   """
 
     # XXX move before setup watch?
     print('\n\n\n\nWATCH+COMMIT\n\n')