RE: [PATCH V4] libxl: make libxl communicate with xenstored by socket or xenbus driver

RE: [PATCH V4] libxl: make libxl communicate with xenstored by socket or xenbus driver

[Jun Zhu (Intern)]

>> +    if (!xsh)
>> +        fprintf(stderr, "cannot connect to xenstore");
>
>This should use the logging functions provided, not just write to
>stderr.

Since libxl__xs_open does not ctx which LIBXL__LOG needs, so I choose to output log with the raw fprintf. 

libxl: make libxl communicate with xenstored by socket or xenbus driver. 
This version is remade from the newest staging version, since the libxl interface has been changed. 
 
signed-off-by: Jun Zhu <jun.zhu@citrix.com>

diff -r cca905a429aa tools/libxl/libxl.c
--- a/tools/libxl/libxl.c	Tue Sep 14 15:39:36 2010 +0100
+++ b/tools/libxl/libxl.c	Tue Sep 14 16:18:00 2010 +0100
@@ -53,12 +53,8 @@
         return ERROR_FAIL;
     }
 
-    ctx->xsh = xs_daemon_open();
-    if (!ctx->xsh)
-        ctx->xsh = xs_domain_open();
+    ctx->xsh = libxl__xs_open();
     if (!ctx->xsh) {
-        LIBXL__LOG_ERRNOVAL(ctx, LIBXL__LOG_ERROR, errno, 
-                        "cannot connect to xenstore");
         xc_interface_close(ctx->xch);
         return ERROR_FAIL;
     }
@@ -69,7 +65,7 @@
 {
     if (ctx->xch) xc_interface_close(ctx->xch);
     libxl_version_info_destroy(&ctx->version_info);
-    if (ctx->xsh) xs_daemon_close(ctx->xsh); 
+    if (ctx->xsh) libxl__xs_close(ctx->xsh);
     return 0;
 }
 
@@ -1395,21 +1391,22 @@
 {
     libxl_device_model_starting *starting = for_spawn;
     char *kvs[3];
-    int rc;
     struct xs_handle *xsh;
 
-    xsh = xs_daemon_open();
-    /* we mustn't use the parent's handle in the child */
-
     kvs[0] = "image/device-model-pid";
     if (asprintf(&kvs[1], "%d", innerchild) < 0)
         return;
     kvs[2] = NULL;
 
-    rc = xs_writev(xsh, XBT_NULL, starting->dom_path, kvs);
-    if (rc)
-        return;
-    xs_daemon_close(xsh);
+    /* we mustn't use the parent's handle in the child */
+    xsh = libxl__xs_open();
+    if (!xsh)
+        goto out;
+    xs_writev(xsh, XBT_NULL, starting->dom_path, kvs);
+    libxl__xs_close(xsh);
+
+out:
+    free(kvs[1]);
 }
 
 static int libxl_vfb_and_vkb_from_device_model_info(libxl_ctx *ctx,
diff -r cca905a429aa tools/libxl/libxl_device.c
--- a/tools/libxl/libxl_device.c	Tue Sep 14 15:39:36 2010 +0100
+++ b/tools/libxl/libxl_device.c	Tue Sep 14 16:18:00 2010 +0100
@@ -405,7 +405,10 @@
     unsigned int num;
     char **l = NULL;
 
-    xsh = xs_daemon_open();
+    xsh = libxl__xs_open();
+    if (!xsh)
+        return -1;
+
     path = libxl__sprintf(&gc, "/local/domain/0/device-model/%d/state", domid);
     xs_watch(xsh, path, path);
     tv.tv_sec = LIBXL_DEVICE_MODEL_START_TIMEOUT;
@@ -427,7 +430,7 @@
 
         free(p);
         xs_unwatch(xsh, path, path);
-        xs_daemon_close(xsh);
+        libxl__xs_close(xsh);
         libxl__free_all(&gc);
         return rc;
 again:
@@ -444,7 +447,7 @@
         }
     }
     xs_unwatch(xsh, path, path);
-    xs_daemon_close(xsh);
+    libxl__xs_close(xsh);
     LIBXL__LOG(ctx, LIBXL__LOG_ERROR, "Device Model not ready");
     libxl__free_all(&gc);
     return -1;
diff -r cca905a429aa tools/libxl/libxl_dom.c
--- a/tools/libxl/libxl_dom.c	Tue Sep 14 15:39:36 2010 +0100
+++ b/tools/libxl/libxl_dom.c	Tue Sep 14 16:18:00 2010 +0100
@@ -326,14 +326,16 @@
 
     snprintf(path, sizeof(path), "/local/domain/0/device-model/%u/logdirty/cmd", domid);
 
-    xsh = xs_daemon_open();
+    xsh = libxl__xs_open();
+    if (!xsh)
+        return;
 
     if (enable)
         xs_write(xsh, XBT_NULL, path, "enable", strlen("enable"));
     else
         xs_write(xsh, XBT_NULL, path, "disable", strlen("disable"));
 
-    xs_daemon_close(xsh);
+    libxl__xs_close(xsh);
 }
 
 static int libxl__domain_suspend_common_callback(void *data)
diff -r cca905a429aa tools/libxl/libxl_internal.h
--- a/tools/libxl/libxl_internal.h	Tue Sep 14 15:39:36 2010 +0100
+++ b/tools/libxl/libxl_internal.h	Tue Sep 14 16:18:00 2010 +0100
@@ -138,6 +138,8 @@
 _hidden char *libxl__xs_get_dompath(libxl__gc *gc, uint32_t domid); // logs errs
 _hidden char *libxl__xs_read(libxl__gc *gc, xs_transaction_t t, char *path);
 _hidden char **libxl__xs_directory(libxl__gc *gc, xs_transaction_t t, char *path, unsigned int *nb);
+_hidden struct xs_handle *libxl__xs_open(void);
+_hidden void libxl__xs_close(struct xs_handle *xsh);
 
 /* from xl_dom */
 _hidden int libxl__domain_is_hvm(libxl_ctx *ctx, uint32_t domid);
diff -r cca905a429aa tools/libxl/libxl_utils.c
--- a/tools/libxl/libxl_utils.c	Tue Sep 14 15:39:36 2010 +0100
+++ b/tools/libxl/libxl_utils.c	Tue Sep 14 16:18:00 2010 +0100
@@ -367,8 +367,9 @@
 
 int libxl_ctx_postfork(libxl_ctx *ctx) {
     if (ctx->xsh) xs_daemon_destroy_postfork(ctx->xsh);
-    ctx->xsh = xs_daemon_open();
-    if (!ctx->xsh) return ERROR_FAIL;
+    ctx->xsh = libxl__xs_open();
+    if (!ctx->xsh)
+        return ERROR_FAIL;
     return 0;
 }
 
diff -r cca905a429aa tools/libxl/libxl_xshelp.c
--- a/tools/libxl/libxl_xshelp.c	Tue Sep 14 15:39:36 2010 +0100
+++ b/tools/libxl/libxl_xshelp.c	Tue Sep 14 16:18:00 2010 +0100
@@ -141,3 +141,20 @@
     libxl__ptr_add(gc, ret);
     return ret;
 }
+
+struct xs_handle *libxl__xs_open()
+{
+    struct xs_handle *xsh = NULL;
+
+    xsh = xs_daemon_open();
+    if (!xsh)
+        xsh = xs_domain_open();
+    if (!xsh)
+        fprintf(stderr, "cannot connect to xenstore");
+    return xsh;
+}
+
+void libxl__xs_close(struct xs_handle *xsh)
+{
+    xs_daemon_close(xsh);
+}

Jun Zhu
Citrix Systems UK

RE: [PATCH V4] libxl: make libxl communicate with xenstored by socket or xenbus driver

[Ian Jackson]

Jun Zhu (Intern) writes ("RE: [Xen-devel] [PATCH V4] libxl: make libxl communicate with xenstored by socket or xenbus driver"):
> Since libxl__xs_open does not ctx which LIBXL__LOG needs, so I choose to output log with the raw fprintf. 

Then the right answer is to make it take a ctx parameter, surely ?
(And ensure that during initialisation, the logging is initialised
before the xenstore connection.)

Ian.

RE: [PATCH V4] libxl: make libxl communicate with xenstored by socket or xenbus driver

[Ian Campbell]

On Tue, 2010-09-14 at 17:27 +0100, Ian Jackson wrote:
> Jun Zhu (Intern) writes ("RE: [Xen-devel] [PATCH V4] libxl: make libxl communicate with xenstored by socket or xenbus driver"):
> > Since libxl__xs_open does not ctx which LIBXL__LOG needs, so I choose to output log with the raw fprintf. 
> 
> Then the right answer is to make it take a ctx parameter, surely ?

gc->owner contains the surrounding context, doesn't it?

Actually you probably want to use libxl__gc_owner().

> And ensure that during initialisation, the logging is initialised
> before the xenstore connection.)

This seems to be the case in libxl_ctx_init().

Ian.

RE: [PATCH V4] libxl: make libxl communicate with xenstored by socket or xenbus driver

[Ian Jackson]

Ian Campbell writes ("RE: [Xen-devel] [PATCH V4] libxl: make libxl communicate with xenstored by socket or xenbus driver"):
> On Tue, 2010-09-14 at 17:27 +0100, Ian Jackson wrote:
> > Then the right answer is to make it take a ctx parameter, surely ?
> 
> gc->owner contains the surrounding context, doesn't it?

Yes.  So passing a gc to libxl__xs_open would be just as good (better
probably, for consistency).  Jun Zhu's libxl__xs_open took neither.

> > And ensure that during initialisation, the logging is initialised
> > before the xenstore connection.)
> 
> This seems to be the case in libxl_ctx_init().

I thought it would be; I'm trying to help Jun out by pointing out the
things that need to be checked when preparing this patch.

Ian.

RE: [PATCH V4] libxl: make libxl communicate with xenstored by socket or xenbus driver

[Jun Zhu (Intern)]

Hi, 

Fine. I will pass a gc to libxl__xs_open and make a new patch.

Jun Zhu
Citrix Systems UK

RE: [PATCH V4] libxl: make libxl communicate with xenstored by socket or xenbus driver

[Jun Zhu (Intern)]

Hi, 

The functions, such as dm_xenstore_record_pid, do not have a ctx pointer in its function parameters. In these functions, if they invoke the libxl__xs_open, it does not have ctx pointer. Should we use the extern ctx directly? I find only the functions in xl_cmdimpl.c use ctx in this way, and the other functions under libxl get the ctx pointer from its function parameter.  

Jun Zhu
Citrix Systems UK
________________________________________

RE: [PATCH V4] libxl: make libxl communicate with xenstored by socket or xenbus driver

[Ian Campbell]

On Thu, 2010-09-16 at 20:21 +0100, Jun Zhu (Intern) wrote:
> The functions, such as dm_xenstore_record_pid, do not have a ctx
> pointer in its function parameters. In these functions, if they invoke
> the libxl__xs_open, it does not have ctx pointer. Should we use the
> extern ctx directly? I find only the functions in xl_cmdimpl.c use ctx
> in this way, and the other functions under libxl get the ctx pointer
> from its function parameter.  

Under no circumstances should libxl try and use a global ctx pointer
from the application using libxl.

dm_xenstore_record_pid runs in a new process and therefore there is no
existing context which can be used.

Maybe libxl__xs_open could, as a special case, accept a NULL gc and not
log anything in that case? dm_xenstore_record_pid doesn't log failure
today anyway and I presume something in the parent process will
eventually log the failure to start the DM.

Otherwise I think the choices are:

1) Continue to treat dm_xenstore_record_pid as a special case which uses
xs_*_open directly.

2) Allocate a new context in dm_xenstore_record_pid using
libxl_ctx_init. It's not clear what the right logger to use for this
would be, maybe some sort of /dev/null logger, in which case we might as
well just add a special case to libxl__xs_open which doesn't log...

3) Rework libxl__spawn_spawn so that the intermediate callback (i.e.
dm_xenstore_record_pid) is called in the context of the parent, and
hence can have the parent's context passed to it. I don't know if this
is even possible given the requirements of libxl__spawn_spawn but it
might for example involve setting up a pipe to pass the pid from the
intermediate process back to the parent, or something along those lines.

3 might be a good thing to do anyway but it's likely to be tricky to get
right and is probably more than you want to get into right now.

Ian.

RE: [PATCH V4] libxl: make libxl communicate with xenstored by socket or xenbus driver

[Ian Jackson]

Ian Campbell writes ("RE: [Xen-devel] [PATCH V4] libxl: make libxl communicate with xenstored by socket or xenbus driver"):
> On Thu, 2010-09-16 at 20:21 +0100, Jun Zhu (Intern) wrote:
> > The functions, such as dm_xenstore_record_pid, do not have a ctx
> > pointer in its function parameters. In these functions, if they invoke
> > the libxl__xs_open, it does not have ctx pointer. Should we use the
> > extern ctx directly? I find only the functions in xl_cmdimpl.c use ctx
> > in this way, and the other functions under libxl get the ctx pointer
> > from its function parameter.  
> 
> Under no circumstances should libxl try and use a global ctx pointer
> from the application using libxl.
> 
> dm_xenstore_record_pid runs in a new process and therefore there is no
> existing context which can be used.

I think it is fine to to reuse the ctx from the caller in
dm_xenstore_record_pid.  Your "new process" test seems to imply
problems for any other daemonic processes libxl may need to spawn.

If the caller needs to know when libxl forks, we should provide a
callback function pointer in the ctx to allow the caller to be told.

Ian.

RE: [PATCH V4] libxl: make libxl communicate with xenstored by socket or xenbus driver

[Ian Campbell]

On Fri, 2010-09-17 at 16:06 +0100, Ian Jackson wrote:
> Ian Campbell writes ("RE: [Xen-devel] [PATCH V4] libxl: make libxl communicate with xenstored by socket or xenbus driver"):
> > On Thu, 2010-09-16 at 20:21 +0100, Jun Zhu (Intern) wrote:
> > > The functions, such as dm_xenstore_record_pid, do not have a ctx
> > > pointer in its function parameters. In these functions, if they invoke
> > > the libxl__xs_open, it does not have ctx pointer. Should we use the
> > > extern ctx directly? I find only the functions in xl_cmdimpl.c use ctx
> > > in this way, and the other functions under libxl get the ctx pointer
> > > from its function parameter.  
> > 
> > Under no circumstances should libxl try and use a global ctx pointer
> > from the application using libxl.
> > 
> > dm_xenstore_record_pid runs in a new process and therefore there is no
> > existing context which can be used.
> 
> I think it is fine to to reuse the ctx from the caller in
> dm_xenstore_record_pid.  Your "new process" test seems to imply
> problems for any other daemonic processes libxl may need to spawn.

Hmm, It's not actually the same context any more since the fork but
perhaps that doesn't really matter.

I would have thought that reusing ctx->xsh over a fork would cause all
manner of mayhem but it looks like libxl_fork cleverly takes care of
that for us.

I'd similarly concerned about any other fds linked to the context, are
any fds used by the particular xentoollog_logger implementation used by
the context going to be alright for example?

Ian.

RE: [PATCH V4] libxl: make libxl communicate with xenstored by socket or xenbus driver

[Ian Jackson]

Ian Campbell writes ("RE: [Xen-devel] [PATCH V4] libxl: make libxl communicate with xenstored by socket or xenbus driver"):
> I'd similarly concerned about any other fds linked to the context, are
> any fds used by the particular xentoollog_logger implementation used by
> the context going to be alright for example?

All the current logger implementations will be fine.  If there were a
problem it would be no trouble for a logger to do getpid() before each
log message.

> I would have thought that reusing ctx->xsh over a fork would cause all
> manner of mayhem but it looks like libxl_fork cleverly takes care of
> that for us.

Yes, I did it like that to try to make it harder to accidentally
introduce xsh reuse problems.

Ian.