xend: detect and report qemu-dm failure
authorKeir Fraser <keir.fraser@citrix.com>
Thu, 12 Jun 2008 16:01:31 +0000 (17:01 +0100)
committerKeir Fraser <keir.fraser@citrix.com>
Thu, 12 Jun 2008 16:01:31 +0000 (17:01 +0100)
Currently, when qemu-dm fails, typically its exit status is lost and
xend doesn't notice.  In the patch below I use a fifo (named pipe) to
detect qemu-dm's termination and report the exit status to the
logfile, if possible.  (If xend has been restarted since the domain
was created, this isn't possible but we can still know that it failed
and report that fact.)

It would be better to have a failure of qemu crash the domain, by
calling the SCHEDOP_shutdown hypercall with SHUTDOWN_crash.  However
if you have on_crash=restart and the configuration is broken in some
way that causes qemu-dm to bomb out straight away this causes xend to
spin endlessly restarting the doomed domain.  This is a general
problem with on_*=restart and ought to be fixed separately.  When it
is fixed, we can safely arrange for domains whose dm has crashed to be
themselves forcibly crashed.

Signed-off-by: Ian Jackson <ian.jackson@eu.citrix.com>
tools/python/xen/util/oshelp.py [new file with mode: 0644]
tools/python/xen/util/utils.py [new file with mode: 0644]
tools/python/xen/xend/XendDomain.py
tools/python/xen/xend/XendLogging.py
tools/python/xen/xend/image.py

diff --git a/tools/python/xen/util/oshelp.py b/tools/python/xen/util/oshelp.py
new file mode 100644 (file)
index 0000000..b5a0a2a
--- /dev/null
@@ -0,0 +1,20 @@
+import fcntl
+import os
+
+def fcntl_setfd_cloexec(file, bool):
+        f = fcntl.fcntl(file, fcntl.F_GETFD)
+        if bool: f |= fcntl.FD_CLOEXEC
+        else: f &= ~fcntl.FD_CLOEXEC
+        fcntl.fcntl(file, fcntl.F_SETFD)
+
+def waitstatus_description(st):
+        if os.WIFEXITED(st):
+                es = os.WEXITSTATUS(st)
+                if es: return "exited with nonzero status %i" % es
+                else: return "exited"
+        elif os.WIFSIGNALED(st):
+                s = "died due to signal %i" % os.WTERMSIG(st)
+                if os.WCOREDUMP(st): s += " (core dumped)"
+                return s
+        else:
+                return "failed with unexpected wait status %i" % st
diff --git a/tools/python/xen/util/utils.py b/tools/python/xen/util/utils.py
new file mode 100644 (file)
index 0000000..b6c7d2a
--- /dev/null
@@ -0,0 +1,6 @@
+import traceback
+import sys
+
+def exception_string(e):
+        (ty,v,tb) = sys.exc_info()
+        return traceback.format_exception_only(ty,v)
index d5fa1ba96dc74f144ea3d78f7d0fd58c50057041..e879216e2d6a185ef1cea61429701e3a4d974d2d 100644 (file)
@@ -34,7 +34,7 @@ import xen.lowlevel.xc
 
 from xen.xend import XendOptions, XendCheckpoint, XendDomainInfo
 from xen.xend.PrettyPrint import prettyprint
-from xen.xend import XendConfig
+from xen.xend import XendConfig, image
 from xen.xend.XendError import XendError, XendInvalidDomain, VmError
 from xen.xend.XendError import VMBadState
 from xen.xend.XendLogging import log
@@ -179,6 +179,8 @@ class XendDomain:
                         log.exception("Failed to create reference to running "
                                       "domain id: %d" % dom['domid'])
 
+            image.cleanup_stale_sentinel_fifos()
+
             # add all managed domains as dormant domains.
             for dom in managed:
                 dom_uuid = dom.get('uuid')
index 6fa58c02e35f9a0a48a20e1c4baacebda534cbd9..d95133b2c8f3b3a30c6dc14e34249fc7fcce249f 100644 (file)
@@ -25,10 +25,10 @@ import tempfile
 import types
 import logging
 import logging.handlers
-import fcntl
 
 from xen.util import mkdir
 from xen.xend.server import params
+from xen.util import oshelp
 
 
 __all__ = [ 'log', 'init', 'getLogFilename' ]
@@ -103,9 +103,7 @@ class XendRotatingFileHandler(logging.handlers.RotatingFileHandler):
     # entire FileHandler, StreamHandler & RotatingFileHandler classes which
     # is even worse
     def setCloseOnExec(self):
-        flags = fcntl.fcntl(self.stream.fileno(), fcntl.F_GETFD)
-        flags |= fcntl.FD_CLOEXEC
-        fcntl.fcntl(self.stream.fileno(), fcntl.F_SETFD, flags)
+        oshelp.fcntl_setfd_cloexec(self.stream, True)
         
 
 def init(filename, level):
index 8f46f86dd6315417d8396aaa8395385b341cfdf2..5ee9a8d95f9524243b01413c917e939423070d18 100644 (file)
@@ -22,6 +22,12 @@ import re
 import math
 import time
 import signal
+import thread
+import fcntl
+import sys
+import errno
+import glob
+import traceback
 
 import xen.lowlevel.xc
 from xen.xend.XendConstants import *
@@ -32,11 +38,23 @@ from xen.xend.xenstore.xstransact import xstransact
 from xen.xend.xenstore.xswatch import xswatch
 from xen.xend import arch
 from xen.xend import XendOptions
+from xen.util import oshelp
+from xen.util import utils
 
 xc = xen.lowlevel.xc.xc()
 
 MAX_GUEST_CMDLINE = 1024
 
+sentinel_path_prefix = '/var/run/xend/dm-'
+sentinel_fifos_inuse = { }
+
+def cleanup_stale_sentinel_fifos():
+    for path in glob.glob(sentinel_path_prefix + '*.fifo'):
+        if path in sentinel_fifos_inuse: continue
+        try: os.unlink(path)
+        except OSError, e:
+            log.warning('could not delete stale fifo %s: %s',
+                path, utils.exception_string(e))
 
 def create(vm, vmConfig):
     """Create an image handler for a vm.
@@ -324,6 +342,13 @@ class ImageHandler:
         args = args + self.dmargs
         return args
 
+    def _openSentinel(self, sentinel_path_fifo):
+        self.sentinel_fifo = file(sentinel_path_fifo, 'r')
+        self.sentinel_lock = thread.allocate_lock()
+        oshelp.fcntl_setfd_cloexec(self.sentinel_fifo, True)
+        sentinel_fifos_inuse[sentinel_path_fifo] = 1
+        self.sentinel_path_fifo = sentinel_path_fifo
+
     def createDeviceModel(self, restore = False):
         if self.device_model is None:
             return
@@ -339,21 +364,29 @@ class ImageHandler:
             env['XAUTHORITY'] = self.xauthority
         if self.vncconsole:
             args = args + ([ "-vncviewer" ])
+        unique_id = "%i-%i" % (self.vm.getDomid(), time.time())
+        sentinel_path = sentinel_path_prefix + unique_id
+        sentinel_path_fifo = sentinel_path + '.fifo'
+        os.mkfifo(sentinel_path_fifo, 0600)
+        sentinel_write = file(sentinel_path_fifo, 'r+')
+        self._openSentinel(sentinel_path_fifo)
+        self.vm.storeDom("image/device-model-fifo", sentinel_path_fifo)
         xstransact.Mkdir("/local/domain/0/device-model/%i" % self.vm.getDomid())
         xstransact.SetPermissions("/local/domain/0/device-model/%i" % self.vm.getDomid(),
                         { 'dom': self.vm.getDomid(), 'read': True, 'write': True })
         log.info("spawning device models: %s %s", self.device_model, args)
         # keep track of pid and spawned options to kill it later
 
-        logfile = "/var/log/xen/qemu-dm-%s.log" %  str(self.vm.info['name_label'])
-        if os.path.exists(logfile):
-            if os.path.exists(logfile + ".1"):
-                os.unlink(logfile + ".1")
-            os.rename(logfile, logfile + ".1")
+        self.logfile = "/var/log/xen/qemu-dm-%s.log" %  str(self.vm.info['name_label'])
+        if os.path.exists(self.logfile):
+            if os.path.exists(self.logfile + ".1"):
+                os.unlink(self.logfile + ".1")
+            os.rename(self.logfile, self.logfile + ".1")
 
         null = os.open("/dev/null", os.O_RDONLY)
-        logfd = os.open(logfile, os.O_WRONLY|os.O_CREAT|os.O_TRUNC)
+        logfd = os.open(self.logfile, os.O_WRONLY|os.O_CREAT|os.O_TRUNC|os.O_APPEND)
         
+        sys.stderr.flush()
         pid = os.fork()
         if pid == 0: #child
             try:
@@ -362,18 +395,26 @@ class ImageHandler:
                 os.dup2(logfd, 2)
                 os.close(null)
                 os.close(logfd)
+                self.sentinel_fifo.close()
                 try:
                     os.execve(self.device_model, args, env)
-                except:
-                    os._exit(127)
+                except Exception, e:
+                    print >>sys.stderr, (
+                        'failed to set up fds or execute dm %s: %s' %
+                        (self.device_model, utils.exception_string(e)))
+                    os._exit(126)
             except:
                 os._exit(127)
         else:
             self.pid = pid
             os.close(null)
             os.close(logfd)
+        sentinel_write.close()
         self.vm.storeDom("image/device-model-pid", self.pid)
         log.info("device model pid: %d", self.pid)
+        # we would very much prefer not to have a thread here and instead
+        #  have a callback but sadly we don't have Twisted in xend
+        self.sentinel_thread = thread.start_new_thread(self._sentinel_watch,())
 
     def signalDeviceModel(self, cmd, ret, par = None):
         if self.device_model is None:
@@ -419,46 +460,118 @@ class ImageHandler:
         xstransact.Store("/local/domain/0/device-model/%i"
                          % self.vm.getDomid(), ('command', 'continue'))
 
+    def _dmfailed(self, message):
+        log.warning("domain %s: %s", self.vm.getName(), message)
+        # ideally we would like to forcibly crash the domain with
+        # something like
+        #    xc.domain_shutdown(self.vm.getDomid(), DOMAIN_CRASH)
+        # but this can easily lead to very rapid restart loops against
+        # which we currently have no protection
+
     def recreate(self):
         if self.device_model is None:
             return
-        self.pid = self.vm.gatherDom(('image/device-model-pid', int))
+        name = self.vm.getName()
+        sentinel_path_fifo = self.vm.readDom('image/device-model-fifo')
+        fifo_fd = -1
+        log.debug("rediscovering %s", sentinel_path_fifo)
+        if sentinel_path_fifo is None:
+            log.debug("%s device model no sentinel, cannot rediscover", name)
+        else:
+            try:
+                # We open it O_WRONLY because that fails ENXIO if no-one
+                # has it open for reading (see SuSv3).  The dm process got
+                # a read/write descriptor from our earlier invocation.
+                fifo_fd = os.open(sentinel_path_fifo, os.O_WRONLY|os.O_NONBLOCK)
+            except OSError, e:
+                if e.errno == errno.ENXIO:
+                    self._dmfailed("%s device model no longer running"%name)
+                elif e.errno == errno.ENOENT:
+                    log.debug("%s device model sentinel %s absent!",
+                            name, sentinel_path_fifo)
+                else:
+                    raise
+        if fifo_fd >= 0:
+            self._openSentinel(sentinel_path_fifo)
+            os.close(fifo_fd)
+            self.pid = self.vm.gatherDom(('image/device-model-pid', int))
+            log.debug("%s device model rediscovered, pid %s sentinel fifo %s",
+                    name, self.pid, sentinel_path_fifo)
+            self.sentinel_thread = thread.start_new_thread(self._sentinel_watch,())
+
+    def _sentinel_watch(self):
+        log.info("waiting for sentinel_fifo")
+        try: self.sentinel_fifo.read(1)
+        except OSError, e: pass
+        self.sentinel_lock.acquire()
+        try:
+            if self.pid:
+                (p,st) = os.waitpid(self.pid, os.WNOHANG)
+                if p == self.pid:
+                    message = oshelp.waitstatus_description(st)
+                else:
+                    # obviously it is malfunctioning, kill it now
+                    try:
+                        os.kill(self.pid, signal.SIGKILL)
+                        message = "malfunctioning (closed sentinel), killed"
+                    except:
+                        message = "malfunctioning or died ?"
+                message = "pid %d: %s" % (self.pid, message)
+            else:
+                message = "no longer running"
+        except Exception, e:
+            message = "waitpid failed: %s" % utils.exception_string(e)
+        message = "device model failure: %s" % message
+        try: message += "; see %s " % self.logfile
+        except: pass
+        self._dmfailed(message)
+        self.pid = None
+        self.sentinel_lock.release()
 
     def destroyDeviceModel(self):
         if self.device_model is None:
             return
         if self.pid:
+            self.sentinel_lock.acquire()
             try:
-                os.kill(self.pid, signal.SIGHUP)
-            except OSError, exn:
-                log.exception(exn)
-            try:
-                # Try to reap the child every 100ms for 10s. Then SIGKILL it.
-                for i in xrange(100):
-                    (p, rv) = os.waitpid(self.pid, os.WNOHANG)
-                    if p == self.pid:
-                        break
-                    time.sleep(0.1)
-                else:
-                    log.warning("DeviceModel %d took more than 10s "
-                                "to terminate: sending SIGKILL" % self.pid)
+                try:
+                    os.kill(self.pid, signal.SIGHUP)
+                except OSError, exn:
+                    log.exception(exn)
+                try:
+                    # Try to reap the child every 100ms for 10s. Then SIGKILL it.
+                    for i in xrange(100):
+                        (p, rv) = os.waitpid(self.pid, os.WNOHANG)
+                        if p == self.pid:
+                            break
+                        time.sleep(0.1)
+                    else:
+                        log.warning("DeviceModel %d took more than 10s "
+                                    "to terminate: sending SIGKILL" % self.pid)
+                        os.kill(self.pid, signal.SIGKILL)
+                        os.waitpid(self.pid, 0)
+                except OSError, exn:
+                    # This is expected if Xend has been restarted within the
+                    # life of this domain.  In this case, we can kill the process,
+                    # but we can't wait for it because it's not our child.
+                    # We just make really sure it's going away (SIGKILL) first.
                     os.kill(self.pid, signal.SIGKILL)
-                    os.waitpid(self.pid, 0)
-            except OSError, exn:
-                # This is expected if Xend has been restarted within the
-                # life of this domain.  In this case, we can kill the process,
-                # but we can't wait for it because it's not our child.
-                # We just make really sure it's going away (SIGKILL) first.
-                os.kill(self.pid, signal.SIGKILL)
-            self.pid = None
-            state = xstransact.Remove("/local/domain/0/device-model/%i"
-                                      % self.vm.getDomid())
+                state = xstransact.Remove("/local/domain/0/device-model/%i"
+                                          % self.vm.getDomid())
+            finally:
+                self.pid = None
+                self.sentinel_lock.release()
             
             try:
                 os.unlink('/var/run/tap/qemu-read-%d' % self.vm.getDomid())
                 os.unlink('/var/run/tap/qemu-write-%d' % self.vm.getDomid())
             except:
                 pass
+            try:
+                del sentinel_fifos_inuse[self.sentinel_path_fifo]
+                os.unlink(self.sentinel_path_fifo)
+            except:
+                pass
 
     def setCpuid(self):
         xc.domain_set_policy_cpuid(self.vm.getDomid())