From: Daniel P. Berrangé <berrange@redhat.com>

The main XML parser code global initializer historically had a mutex
protecting it, and more recently uses a pthread_once. The RelaxNG
code, however, relies on two other global initializers that are
not thread safe, just relying on setting an integer "initialized"
flag.

Calling the relevant initializers from libvirt in a protected global
initializer will protect libvirt's own concurrent usage, however, it
cannot protect against other libraries loaded in process that might
be using libxml2's schema code. Fortunately:

 * The chances of other loaded non-libvirt code using libxml is
   relatively low
 * The chances of other loaded non-libvirt code using the schema
   validation / catalog functionality inside libxml is even
   lower
 * The chances of both libvirt and the non-libvirt usage having
   their *1st* usage of libxml2 be concurrent is tiny

IOW, in practice, although our solution doesn't fully fix the thread
safety, it is good enough.

libxml2 should none the less still be fixed to make its global
initializers be thread safe without special actions by its API
consumers[1].

Resolves: https://gitlab.com/libvirt/libvirt/-/issues/788
[1] https://gitlab.gnome.org/GNOME/libxml2/-/merge_requests/326
Signed-off-by: Daniel P. Berrangé <berrange@redhat.com>
---

Changed in v3:

 - Drop xmlInitializeCatalog - I misread the code wrt
   thread safety - it has a sufficiently protective mutex
 - Cope with return type change for xmlSchemaInitTypes
   in libxml >= 2.11.0


 src/util/virxml.c | 28 ++++++++++++++++++++++++++++
 1 file changed, 28 insertions(+)

diff --git a/src/util/virxml.c b/src/util/virxml.c
index 9d46e5f32f..c851d6f407 100644
--- a/src/util/virxml.c
+++ b/src/util/virxml.c
@@ -26,6 +26,8 @@
 
 #include <libxml/xmlsave.h>
 #include <libxml/xpathInternals.h>
+#include <libxml/xmlschemastypes.h>
+#include <libxml/catalog.h>
 
 #include "virerror.h"
 #include "virxml.h"
@@ -35,6 +37,7 @@
 #include "virstring.h"
 #include "virutil.h"
 #include "viruuid.h"
+#include "virthread.h"
 #include "configmake.h"
 
 #define VIR_FROM_THIS VIR_FROM_XML
@@ -50,6 +53,28 @@ struct virParserData {
 };
 
 
+static int
+virXMLSchemaOnceInit(void)
+{
+#if LIBXML_VERSION >= 21100
+    if (xmlSchemaInitTypes() < 0) {
+        virReportError(VIR_ERR_INTERNAL_ERROR,
+                       _("Unable to initialize libxml2 schema types"));
+        return -1;
+    }
+#else
+    xmlSchemaInitTypes();
+#endif
+    if (xmlRelaxNGInitTypes() < 0) {
+        virReportError(VIR_ERR_INTERNAL_ERROR,
+                       _("Unable to initialize libxml2 RelaxNG data"));
+        return -1;
+    }
+    return 0;
+}
+
+VIR_ONCE_GLOBAL_INIT(virXMLSchema);
+
 static xmlXPathContextPtr
 virXMLXPathContextNew(xmlDocPtr xml)
 {
@@ -1603,6 +1628,9 @@ virXMLValidatorInit(const char *schemafile)
 {
     g_autoptr(virXMLValidator) validator = NULL;
 
+    if (virXMLSchemaInitialize() < 0)
+        return NULL;
+
     validator = g_new0(virXMLValidator, 1);
 
     validator->schemafile = g_strdup(schemafile);
-- 
2.49.0

On Mon, Jun 23, 2025 at 17:17:06 +0100, Daniel P. Berrangé via Devel wrote:
> From: Daniel P. Berrangé <berrange@redhat.com>
> 
> The main XML parser code global initializer historically had a mutex
> protecting it, and more recently uses a pthread_once. The RelaxNG
> code, however, relies on two other global initializers that are
> not thread safe, just relying on setting an integer "initialized"
> flag.
> 
> Calling the relevant initializers from libvirt in a protected global
> initializer will protect libvirt's own concurrent usage, however, it
> cannot protect against other libraries loaded in process that might
> be using libxml2's schema code. Fortunately:
> 
>  * The chances of other loaded non-libvirt code using libxml is
>    relatively low
>  * The chances of other loaded non-libvirt code using the schema
>    validation / catalog functionality inside libxml is even
>    lower
>  * The chances of both libvirt and the non-libvirt usage having
>    their *1st* usage of libxml2 be concurrent is tiny
> 
> IOW, in practice, although our solution doesn't fully fix the thread
> safety, it is good enough.
> 
> libxml2 should none the less still be fixed to make its global
> initializers be thread safe without special actions by its API
> consumers[1].
> 
> Resolves: https://gitlab.com/libvirt/libvirt/-/issues/788
> [1] https://gitlab.gnome.org/GNOME/libxml2/-/merge_requests/326
> Signed-off-by: Daniel P. Berrangé <berrange@redhat.com>
> ---
> 
> Changed in v3:
> 
>  - Drop xmlInitializeCatalog - I misread the code wrt
>    thread safety - it has a sufficiently protective mutex
>  - Cope with return type change for xmlSchemaInitTypes
>    in libxml >= 2.11.0

Reviewed-by: Peter Krempa <pkrempa@redhat.com>