Re: [libvirt] [PATCH 2/2] Cancel migration/dump if user presses Ctrl-C when migration/dump is in progress

27 Dec 2010


      On Thu, Dec 23, 2010 at 09:05:58AM -0700, Eric Blake wrote:
...
On 12/23/2010 02:07 AM, Hu Tao wrote:
...
---
No commit message beyond the title?
...
@@ -208,6 +211,7 @@ typedef struct __vshCmd {
 typedef struct __vshControl {
     char *name;                 /* connection name */
     virConnectPtr conn;         /* connection to hypervisor (MAY BE NULL) */
+    virDomainPtr dom;
No comment as to it's purpose?
Will update it if I update the patch.
...
...
vshCmd *cmd;                /* the current command */
     char *cmdstr;               /* string with command */
     int imode;                  /* interactive mode? */
@@ -221,6 +225,9 @@ typedef struct __vshControl {
     int log_fd;                 /* log file descriptor */
     char *historydir;           /* readline history directory name */
     char *historyfile;          /* readline history file name */
+
+    virMutex mutex;
+    virThreadPoolPtr threadPool;
Likewise.
Will update it if I update the patch.
...
...
@@ -509,6 +518,44 @@ static void vshCatchDisconnect(int sig, siginfo_t * siginfo,
     if ((sig == SIGPIPE) || (siginfo->si_signo == SIGPIPE))
         disconnected++;
 }
+#endif
+
+#ifdef SIGINT
Unnecessary ifdef - SIGINT is available on all platforms (it is required
by C).
...
+static void vshCatchInt(int sig, siginfo_t *siginfo,
+                        void *context ATTRIBUTE_UNUSED) {
+    vshControl *ctl = &_ctl;
+    vshCmd *cmds, *c;
+    char *cmdStr = NULL;
+
+    virMutexLock(&ctl->mutex);
+    if (!ctl->dom)
+        goto error;
+
+    if (virAsprintf(&cmdStr, "domjobabort %s", ctl->dom->name) < 0)
_BIG_ no-no.  virAsprintf calls malloc(), which is NOT signal-safe (that
is, if the signal arrived while some other thread was in the middle of a
malloc(), you've just caused deadlock).  The only safe thing to do in a
signal handler is to set some state variables that are then checked in
the main processing loop, when you are back in a safe state to act on a
(minimally-delayed) basis on the signal.
See how daemon/libvirtd.c uses qemudDispatchSignalEvent (although that's
being rewritten by danpb's factorization into a cleaner rpc
server/client model).
...
+        goto error;
+
+    if ((sig == SIGINT) || (siginfo->si_signo == SIGINT)) {
+        if (!vshConnectionUsability (ctl, ctl->conn))
+            goto error;
+
+        vshCommandStringParse(ctl, cmdStr);
+        cmds = ctl->cmd;
+        ctl->cmd = NULL;
+        virMutexUnlock(&ctl->mutex);
+        do {
+            c = cmds;
+            cmds = cmds->next;
+            c->next = NULL;
+            ignore_value(virThreadPoolSendJob(ctl->threadPool, c, true));
Continuing in the vein of bad practice - blocking on another thread to
complete inside a signal handler is crazy.  But setting up a threadpool
I did a copy&paste and forgot to change true to false :)
...
to handle signals in a separate thread means that your signal-handling
thread is already independently available to cancel the job that is
underway in the primary thread.
The intention of setting up a threadpool is to submit another command in
a seprate thread while there is already a command being processed. If we
do it in one thread, then we will end up with deaklock eventually. Take
dump as example:

  1. submit command dump 
  2. polling for reply from RPC server
  3. SIGINT(poll is interrupted)
  4. submit command domjobabort form signal handler
  5. (in remoteIO())there is a priv->waitDispatch, so go to
     sleep(vifCondWait). The impl of remote driver is thread-aware, and
     this thread (processing command domjobabort) is supposed to be
     waken up by the thread that is having priv->waitDispatch(processing
     command dump). But we are the same thread, so deadlock.


Well, I have to say that the patch is nasty. Do you (or anyone) have a
good idea to achieve the goal to cancel the active job if user presses
Ctrl-C? Thanks.
...
...
+        } while (cmds);
+    }
+    VIR_FREE(cmdStr);
+    return;
+error:
+    virMutexUnlock(&ctl->mutex);
+    VIR_FREE(cmdStr);
+}
+#endif
/*
  * vshSetupSignals:
@@ -520,17 +567,20 @@ static void
 vshSetupSignals(void) {
     struct sigaction sig_action;
-    sig_action.sa_sigaction = vshCatchDisconnect;
     sig_action.sa_flags = SA_SIGINFO;
     sigemptyset(&sig_action.sa_mask);
+#ifdef SIGPIPE
Might be unnecessary - gnulib guarantees that a working replacement for
SIGPIPE is available on mingw, which is the only platform that lacks it
natively.  Oh, but that module is currently LGPLv3+, I'll have to ask on
the gnulib list if people are willing to relax it to LGPLv2+ so we can
use it.
-- 
Eric Blake   eblake@redhat.com    +1-801-349-2682
Libvirt virtualization library http://libvirt.org
-- 
Thanks,
Hu Tao

    