[PATCH V3] libxenstore: filter watch events in libxenstore when we unwatch

[PATCH V3] libxenstore: filter watch events in libxenstore when we unwatch

[Julien Grall]

XenStore puts in queued watch events via a thread and notifies the user.
Sometimes xs_unwatch is called before all related message is read. The use
case is non-threaded libevent, we have two event A and B:
    - Event A will destroy something and call xs_unwatch;
    - Event B is used to notify that a node has changed in XenStore.
As the event is called one by one, event A can be handled before event B.
So on next xs_watch_read the user could retrieve an unwatch token and
a segfault occured if the token store the pointer of the structure
(ie: "backend:0xcafe").

To avoid problem with previous application using libXenStore, this behaviour
will only be enabled if XS_UNWATCH_SAFE is give to xs_open.

Signed-off-by: Ian Jackson <ian.jackson@eu.citrix.com>
Signed-off-by: Julien Grall <julien.grall@citrix.com>
---
  Modification between V2 and V3:
    - Add XS_UNWATCH_SAFE;
    - Rename xs_clear_watch_pipe to xs_maybe_clear_watch_pipe.

  Modification between V1 and V2:
    - Add xs_clear_watch_pipe to avoid code duplication;
    - Modify commit message by Ian Jackson;
    - Rework list filtering.

 tools/xenstore/xenstore.h |    7 ++++
 tools/xenstore/xs.c       |   86 ++++++++++++++++++++++++++++++++++++++++-----
 2 files changed, 85 insertions(+), 8 deletions(-)

diff --git a/tools/xenstore/xenstore.h b/tools/xenstore/xenstore.h
index 7259e49..81ce8da 100644
--- a/tools/xenstore/xenstore.h
+++ b/tools/xenstore/xenstore.h
@@ -26,6 +26,13 @@
 
 #define XS_OPEN_READONLY	1UL<<0
 #define XS_OPEN_SOCKETONLY      1UL<<1
+/*
+ * Avoid race condition with unwatch - Not enabled by default
+ *
+ * If you want to enable it please be sure to watch a path (/foo) and
+ * a sub-path (/foo/bar) with the same token.
+ */
+#define XS_UNWATCH_SAFE     1UL<<2
 
 struct xs_handle;
 typedef uint32_t xs_transaction_t;
diff --git a/tools/xenstore/xs.c b/tools/xenstore/xs.c
index b951015..591fe79 100644
--- a/tools/xenstore/xs.c
+++ b/tools/xenstore/xs.c
@@ -67,6 +67,8 @@ struct xs_handle {
 
 	/* Clients can select() on this pipe to wait for a watch to fire. */
 	int watch_pipe[2];
+	/* Filtering watch event in unwatch function? */
+	bool unwatch_safe;
 
 	/*
          * A list of replies. Currently only one will ever be outstanding
@@ -125,6 +127,8 @@ struct xs_handle {
 	struct list_head watch_list;
 	/* Clients can select() on this pipe to wait for a watch to fire. */
 	int watch_pipe[2];
+	/* Filtering watch event in unwatch function? */
+	bool unwatch_safe;
 };
 
 #define mutex_lock(m)		((void)0)
@@ -247,6 +251,8 @@ static struct xs_handle *get_handle(const char *connect_to)
 	/* Watch pipe is allocated on demand in xs_fileno(). */
 	h->watch_pipe[0] = h->watch_pipe[1] = -1;
 
+	h->unwatch_safe = false;
+
 #ifdef USE_PTHREAD
 	pthread_mutex_init(&h->watch_mutex, NULL);
 	pthread_cond_init(&h->watch_condvar, NULL);
@@ -287,6 +293,9 @@ struct xs_handle *xs_open(unsigned long flags)
 	if (!xsh && !(flags & XS_OPEN_SOCKETONLY))
 		xsh = get_handle(xs_domain_dev());
 
+	if (xsh && (flags & XS_UNWATCH_SAFE))
+		xsh->unwatch_safe = true;
+
 	return xsh;
 }
 
@@ -753,6 +762,19 @@ bool xs_watch(struct xs_handle *h, const char *path, const char *token)
 				ARRAY_SIZE(iov), NULL));
 }
 
+
+/* Clear the pipe token if there are no more pending watchs.
+ * We suppose the watch_mutex is already taken.
+ */
+static void xs_maybe_clear_watch_pipe(struct xs_handle *h)
+{
+	char c;
+
+	if (list_empty(&h->watch_list) && (h->watch_pipe[0] != -1))
+		while (read(h->watch_pipe[0], &c, 1) != 1)
+			continue;
+}
+
 /* Find out what node change was on (will block if nothing pending).
  * Returns array of two pointers: path and token, or NULL.
  * Call free() after use.
@@ -761,7 +783,7 @@ static char **read_watch_internal(struct xs_handle *h, unsigned int *num,
 				  int nonblocking)
 {
 	struct xs_stored_msg *msg;
-	char **ret, *strings, c = 0;
+	char **ret, *strings;
 	unsigned int num_strings, i;
 
 	mutex_lock(&h->watch_mutex);
@@ -798,11 +820,7 @@ static char **read_watch_internal(struct xs_handle *h, unsigned int *num,
 	msg = list_top(&h->watch_list, struct xs_stored_msg, list);
 	list_del(&msg->list);
 
-	/* Clear the pipe token if there are no more pending watches. */
-	if (list_empty(&h->watch_list) && (h->watch_pipe[0] != -1))
-		while (read(h->watch_pipe[0], &c, 1) != 1)
-			continue;
-
+	xs_maybe_clear_watch_pipe(h);
 	mutex_unlock(&h->watch_mutex);
 
 	assert(msg->hdr.type == XS_WATCH_EVENT);
@@ -855,14 +873,66 @@ char **xs_read_watch(struct xs_handle *h, unsigned int *num)
 bool xs_unwatch(struct xs_handle *h, const char *path, const char *token)
 {
 	struct iovec iov[2];
+	struct xs_stored_msg *msg, *tmsg;
+	bool res;
+	char *s, *p;
+	unsigned int i;
+	char *l_token, *l_path;
 
 	iov[0].iov_base = (char *)path;
 	iov[0].iov_len = strlen(path) + 1;
 	iov[1].iov_base = (char *)token;
 	iov[1].iov_len = strlen(token) + 1;
 
-	return xs_bool(xs_talkv(h, XBT_NULL, XS_UNWATCH, iov,
-				ARRAY_SIZE(iov), NULL));
+	res = xs_bool(xs_talkv(h, XBT_NULL, XS_UNWATCH, iov,
+						   ARRAY_SIZE(iov), NULL));
+
+	if (!h->unwatch_safe) /* Don't filter the watch list */
+		return res;
+
+
+	/* Filter the watch list to remove potential message */
+	mutex_lock(&h->watch_mutex);
+
+	if (list_empty(&h->watch_list)) {
+	    mutex_unlock(&h->watch_mutex);
+	    return res;
+	}
+
+	list_for_each_entry_safe(msg, tmsg, &h->watch_list, list) {
+		assert(msg->hdr.type == XS_WATCH_EVENT);
+
+		s = msg->body;
+
+		l_token = NULL;
+		l_path = NULL;
+
+		for (p = s, i = 0; p < msg->body + msg->hdr.len; p++) {
+			if (*p == '\0')
+			{
+				if (i == XS_WATCH_TOKEN)
+					l_token = s;
+				else if (i == XS_WATCH_PATH)
+					l_path = s;
+				i++;
+				s = p + 1;
+			}
+		}
+
+		if (l_token && !strcmp(token, l_token)
+			/* Use strncmp because we can have a watch fired on sub-directory */
+			&& l_path && !strncmp(path, l_path, strlen(path))) {
+			fprintf (stderr, "DELETE\n");
+			list_del(&msg->list);
+			free(msg);
+		}
+	}
+
+	xs_maybe_clear_watch_pipe(h);
+
+	mutex_unlock(&h->watch_mutex);
+
+	return res;
 }
 
 /* Start a transaction: changes by others will not be seen during this
-- 
Julien Grall

[PATCH V3] libxenstore: filter watch events in libxenstore when we unwatch

[Ian Jackson]

Julien Grall writes ("[PATCH V3] libxenstore: filter watch events in libxenstore when we unwatch"):
> XenStore puts in queued watch events via a thread and notifies the user.
> Sometimes xs_unwatch is called before all related message is read. The use
> case is non-threaded libevent, we have two event A and B:
>     - Event A will destroy something and call xs_unwatch;
>     - Event B is used to notify that a node has changed in XenStore.
> As the event is called one by one, event A can be handled before event B.
> So on next xs_watch_read the user could retrieve an unwatch token and
> a segfault occured if the token store the pointer of the structure
> (ie: "backend:0xcafe").
> 
> To avoid problem with previous application using libXenStore, this behaviour
> will only be enabled if XS_UNWATCH_SAFE is give to xs_open.

Sorry I didn't reply to your previous email on this subject.

I think this is a reasonable approach but the option name needs to be
more descriptive and the documentation a bit better.

XS_UNWATCH_FILTER ?  XS_WATCH_TOKENS_UNIQUE ?

As for the documentation:

 /*
  * Setting XS_UNWATCH_FILTER arranges that after xs_unwatch, no
  * related watch events will be delivered via xs_read_watch.  But
  * this relies on all tokens provided by the application to
  * libxenstore being unique.  The choices are:
  *
  *   XS_UNWATCH_FILTER clear           XS_UNWATCH_FILTER set
  *
  *    Even after xs_unwatch, "stale"    After xs_unwatch returns, no
  *    instances of the watch event      watch events with the same
  *    may be delivered.                 token will be delivered.
  *
  *    Tokens need not be unique.        Tokens specified in xs_watch
  *                                      calls must be unique.
  */

With this specification it is not necessary to check the path when
filtering out the stale events.  Which is just as well because:

> +		if (l_token && !strcmp(token, l_token)
> +			/* Use strncmp because we can have a watch fired on sub-directory */
> +			&& l_path && !strncmp(path, l_path, strlen(path))) {

This isn't correct: the subpath relation is not the same as the prefix
relation.

Would this semantic be sufficient for your application ?

Ian.

[PATCH V3] libxenstore: filter watch events in libxenstore when we unwatch

[Ian Jackson]

Ian Jackson writes ("[PATCH V3] libxenstore: filter watch events in libxenstore when we unwatch"):
>   *    Tokens need not be unique.        Tokens specified in xs_watch
>   *                                      calls must be unique.

This is vague about lifetime.  Perhaps:

  A token specified in an xs_watch call may not be also specified in
  a second xs_watch call until a successful xs_unwatch for the first
  watch has returned.

?

Ian.

Re: [PATCH V3] libxenstore: filter watch events in libxenstore when we unwatch

[Julien Grall]

On 12/13/2012 06:35 PM, Ian Jackson wrote:

> Julien Grall writes ("[PATCH V3] libxenstore: filter watch events in libxenstore when we unwatch"):
>> XenStore puts in queued watch events via a thread and notifies the user.
>> Sometimes xs_unwatch is called before all related message is read. The use
>> case is non-threaded libevent, we have two event A and B:
>>     - Event A will destroy something and call xs_unwatch;
>>     - Event B is used to notify that a node has changed in XenStore.
>> As the event is called one by one, event A can be handled before event B.
>> So on next xs_watch_read the user could retrieve an unwatch token and
>> a segfault occured if the token store the pointer of the structure
>> (ie: "backend:0xcafe").
>>
>> To avoid problem with previous application using libXenStore, this behaviour
>> will only be enabled if XS_UNWATCH_SAFE is give to xs_open.
> 
> Sorry I didn't reply to your previous email on this subject.
> 
> I think this is a reasonable approach but the option name needs to be
> more descriptive and the documentation a bit better.
> 
> XS_UNWATCH_FILTER ?  XS_WATCH_TOKENS_UNIQUE ?


I think XS_UNWATCH_FILTER is better.

> 
> As for the documentation:


I think it's too restrictive for upstream. xs_unwatch could only filter
following the token and the subpath. I modified your documentation
proposal below to take into account this solution. What do you think?

>  /*
>   * Setting XS_UNWATCH_FILTER arranges that after xs_unwatch, no
>   * related watch events will be delivered via xs_read_watch.  But

    * this relies on the couple token, subpath is unique.

>   *
>   *   XS_UNWATCH_FILTER clear           XS_UNWATCH_FILTER set
>   *
>   *    Even after xs_unwatch, "stale"    After xs_unwatch returns, no
>   *    instances of the watch event      watch events with the same
>   *    may be delivered.                 token and with the same subpath

    *                                      will be delivered.
    *

    *    A path and a subpath can be       The application must avoid to

    *    register with the same token.     register a path (/foo/) and
    *                                      a subpath (/foo/bar) with the
    *                                      same path until a successful
    *                                      xs_unwatch for the first
    *                                      watch has returned.

>   *                                      
>   */
> 
> With this specification it is not necessary to check the path when
> filtering out the stale events.  Which is just as well because:
>

>> +		if (l_token && !strcmp(token, l_token)
>> +			/* Use strncmp because we can have a watch fired on sub-directory */
>> +			&& l_path && !strncmp(path, l_path, strlen(path))) {
> 
> This isn't correct: the subpath relation is not the same as the prefix
> relation.


I will fix the test by using xs_path_is_subpath.

Thanks,

Julien

Re: [PATCH V3] libxenstore: filter watch events in libxenstore when we unwatch

[Ian Jackson]

Julien Grall writes ("Re: [PATCH V3] libxenstore: filter watch events in libxenstore when we unwatch"):
> On 12/13/2012 06:35 PM, Ian Jackson wrote:
> > XS_UNWATCH_FILTER ?  XS_WATCH_TOKENS_UNIQUE ?
> 
> I think XS_UNWATCH_FILTER is better.

OK.

> > As for the documentation:
> 
> I think it's too restrictive for upstream. xs_unwatch could only filter
> following the token and the subpath. I modified your documentation
> proposal below to take into account this solution. What do you think?

I like your proposal.

      *    A path and a subpath can be       The application must avoid to
      *    register with the same token.     register a path (/foo/) and

A tiny grammar quibble: you should write "the application must avoid
registering".  "Avoid" needs the FOOing form of a verb rather than the
"to FOO" form.

> I will fix the test by using xs_path_is_subpath.

Great.

Thanks,
Ian.