[Qemu-devel] [PATCH v5 0/7] dataplane: switch to N:M devices-per-thread model

[Qemu-devel] [PATCH v5 0/7] dataplane: switch to N:M devices-per-thread model

[Stefan Hajnoczi]

v5:
 * Use Igor's get_pointer() malloc string patch [Igor]
 * Add object_get_canonical_basename() and use it for iothread_get_id() [Igor]

v4:
 * Rename qdev property to virtio-blk-pci,x-iothread= since we will switch to
   QOM links once object_property_add_link() is fixed [bonzini]

v3:
 * Fixed "Reliquinish" typo [fam]
 * Rebased onto qemu.git/master which now has Igor's -object improvements

v2:
 * Based off Igor's "-object/object-add support custom location and 2nd stage
   initialization" series
 * Dropped dedicated -iothread option in favor of -object
 * Avoid re-acquiring rfifo in iothread_run() [mdroth]

This series moves the event loop thread out of dataplane code.  It makes
-object iothread,id=foo a separate concept so several devices can be bound to
same iothread.

Syntax:

  qemu -object iothread,id=iothread0 \
       -device virtio-blk-pci,x-iothread=iothread0,x-data-plane=on,...

For backwards-compatibility the iothread= parameter can be omitted.  A
per-device IOThread will be created behind the scenes (just like the old 1:1
threading model).

This series includes the aio_context_acquire/release API which makes it easy to
synchronize access to AioContext across threads.

After this series I will send separate patches for a "query-iothreads" command
that returns thread IDs similar to "query-cpus".  This will allow binding
dataplane threads to host CPUs.

Igor Mammedov (1):
  qdev: make get_pointer() handle temporary strings

Stefan Hajnoczi (6):
  object: add object_get_canonical_basename()
  rfifolock: add recursive FIFO lock
  aio: add aio_context_acquire() and aio_context_release()
  iothread: add I/O thread object
  iothread: add "iothread" qdev property type
  dataplane: replace internal thread with IOThread

 Makefile.objs                    |   1 +
 async.c                          |  18 ++++++
 hw/block/dataplane/virtio-blk.c  |  96 +++++++++++++++++--------------
 hw/core/qdev-properties-system.c |  70 ++++++++++++++++++++---
 include/block/aio.h              |  18 ++++++
 include/hw/qdev-properties.h     |   3 +
 include/hw/virtio/virtio-blk.h   |   8 ++-
 include/qemu/rfifolock.h         |  54 ++++++++++++++++++
 include/qom/object.h             |   8 +++
 include/sysemu/iothread.h        |  30 ++++++++++
 iothread.c                       | 119 +++++++++++++++++++++++++++++++++++++++
 qom/object.c                     |  53 +++++++++--------
 tests/Makefile                   |   2 +
 tests/test-aio.c                 |  58 +++++++++++++++++++
 tests/test-rfifolock.c           |  90 +++++++++++++++++++++++++++++
 util/Makefile.objs               |   1 +
 util/rfifolock.c                 |  78 +++++++++++++++++++++++++
 17 files changed, 633 insertions(+), 74 deletions(-)
 create mode 100644 include/qemu/rfifolock.h
 create mode 100644 include/sysemu/iothread.h
 create mode 100644 iothread.c
 create mode 100644 tests/test-rfifolock.c
 create mode 100644 util/rfifolock.c

-- 
1.8.5.3

[Qemu-devel] [PATCH v5 1/7] object: add object_get_canonical_basename()

[Stefan Hajnoczi]

It is often useful to find an object's child property name.  Also use
this new function to simplify the implementation of
object_get_canonical_path().

Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
---
 include/qom/object.h |  8 ++++++++
 qom/object.c         | 53 ++++++++++++++++++++++++++++++----------------------
 2 files changed, 39 insertions(+), 22 deletions(-)

diff --git a/include/qom/object.h b/include/qom/object.h
index 9c7c361..8c6db7c 100644
--- a/include/qom/object.h
+++ b/include/qom/object.h
@@ -974,6 +974,14 @@ const char *object_property_get_type(Object *obj, const char *name,
 Object *object_get_root(void);
 
 /**
+ * object_get_canonical_basename:
+ *
+ * Returns: The final component in the object's canonical path.  The canonical
+ * path is the path within the composition tree starting from the root.
+ */
+gchar *object_get_canonical_basename(Object *obj);
+
+/**
  * object_get_canonical_path:
  *
  * Returns: The canonical path for a object.  This is the path within the
diff --git a/qom/object.c b/qom/object.c
index 660859c..0cdc319 100644
--- a/qom/object.c
+++ b/qom/object.c
@@ -1102,39 +1102,48 @@ void object_property_add_link(Object *obj, const char *name,
     g_free(full_type);
 }
 
+gchar *object_get_canonical_basename(Object *obj)
+{
+    ObjectProperty *prop = NULL;
+
+    g_assert(obj->parent != NULL);
+
+    QTAILQ_FOREACH(prop, &obj->parent->properties, node) {
+        if (!object_property_is_child(prop)) {
+            continue;
+        }
+
+        if (prop->opaque == obj) {
+            return g_strdup(prop->name);
+        }
+    }
+
+    /* obj had a parent but was not a child, should never happen */
+    g_assert_not_reached();
+    return NULL;
+}
+
 gchar *object_get_canonical_path(Object *obj)
 {
     Object *root = object_get_root();
-    char *newpath = NULL, *path = NULL;
+    char *newpath, *path = NULL;
 
     while (obj != root) {
-        ObjectProperty *prop = NULL;
-
-        g_assert(obj->parent != NULL);
-
-        QTAILQ_FOREACH(prop, &obj->parent->properties, node) {
-            if (!object_property_is_child(prop)) {
-                continue;
-            }
+        char *component = object_get_canonical_basename(obj);
 
-            if (prop->opaque == obj) {
-                if (path) {
-                    newpath = g_strdup_printf("%s/%s", prop->name, path);
-                    g_free(path);
-                    path = newpath;
-                } else {
-                    path = g_strdup(prop->name);
-                }
-                break;
-            }
+        if (path) {
+            newpath = g_strdup_printf("%s/%s", component, path);
+            g_free(component);
+            g_free(path);
+            path = newpath;
+        } else {
+            path = component;
         }
 
-        g_assert(prop != NULL);
-
         obj = obj->parent;
     }
 
-    newpath = g_strdup_printf("/%s", path);
+    newpath = g_strdup_printf("/%s", path ? path : "");
     g_free(path);
 
     return newpath;
-- 
1.8.5.3

Re: [Qemu-devel] [PATCH v5 1/7] object: add object_get_canonical_basename()

[Andreas Färber]

Am 28.02.2014 16:18, schrieb Stefan Hajnoczi:
> It is often useful to find an object's child property name.  Also use
> this new function to simplify the implementation of
> object_get_canonical_path().
> 
> Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
> ---
>  include/qom/object.h |  8 ++++++++
>  qom/object.c         | 53 ++++++++++++++++++++++++++++++----------------------
>  2 files changed, 39 insertions(+), 22 deletions(-)
> 
> diff --git a/include/qom/object.h b/include/qom/object.h
> index 9c7c361..8c6db7c 100644
> --- a/include/qom/object.h
> +++ b/include/qom/object.h
> @@ -974,6 +974,14 @@ const char *object_property_get_type(Object *obj, const char *name,
>  Object *object_get_root(void);
>  
>  /**
> + * object_get_canonical_basename:
> + *
> + * Returns: The final component in the object's canonical path.  The canonical
> + * path is the path within the composition tree starting from the root.
> + */
> +gchar *object_get_canonical_basename(Object *obj);

I find this name very confusing. There is no such thing as a canonical
base name, ..._canonical_path_component would make its purpose much more
obvious.

An underlying issue here probably is that Anthony didn't want a public
API to access the Object::parent. But the only other user is
iothread_get_id() in 4/7, so we might just loop over its known path
prefix there to discover the right child<>. On the other hand, couldn't
device IDs in /machine/peripheral benefit from this today, too?

In any way I would've liked to get CC'ed on this QOM API proposal please.

> +
> +/**
>   * object_get_canonical_path:
>   *
>   * Returns: The canonical path for a object.  This is the path within the
> diff --git a/qom/object.c b/qom/object.c
> index 660859c..0cdc319 100644
> --- a/qom/object.c
> +++ b/qom/object.c
> @@ -1102,39 +1102,48 @@ void object_property_add_link(Object *obj, const char *name,
>      g_free(full_type);
>  }
>  
> +gchar *object_get_canonical_basename(Object *obj)
> +{
> +    ObjectProperty *prop = NULL;
> +
> +    g_assert(obj->parent != NULL);

It might make sense to assert obj first? Accessing ->parent would not be
much different from ->parent->properties otherwise. But applies to the
original code as well and the movement looks OK.

Regards,
Andreas

> +
> +    QTAILQ_FOREACH(prop, &obj->parent->properties, node) {
> +        if (!object_property_is_child(prop)) {
> +            continue;
> +        }
> +
> +        if (prop->opaque == obj) {
> +            return g_strdup(prop->name);
> +        }
> +    }
> +
> +    /* obj had a parent but was not a child, should never happen */
> +    g_assert_not_reached();
> +    return NULL;
> +}
> +
>  gchar *object_get_canonical_path(Object *obj)
>  {
>      Object *root = object_get_root();
> -    char *newpath = NULL, *path = NULL;
> +    char *newpath, *path = NULL;
>  
>      while (obj != root) {
> -        ObjectProperty *prop = NULL;
> -
> -        g_assert(obj->parent != NULL);
> -
> -        QTAILQ_FOREACH(prop, &obj->parent->properties, node) {
> -            if (!object_property_is_child(prop)) {
> -                continue;
> -            }
> +        char *component = object_get_canonical_basename(obj);
>  
> -            if (prop->opaque == obj) {
> -                if (path) {
> -                    newpath = g_strdup_printf("%s/%s", prop->name, path);
> -                    g_free(path);
> -                    path = newpath;
> -                } else {
> -                    path = g_strdup(prop->name);
> -                }
> -                break;
> -            }
> +        if (path) {
> +            newpath = g_strdup_printf("%s/%s", component, path);
> +            g_free(component);
> +            g_free(path);
> +            path = newpath;
> +        } else {
> +            path = component;
>          }
>  
> -        g_assert(prop != NULL);
> -
>          obj = obj->parent;
>      }
>  
> -    newpath = g_strdup_printf("/%s", path);
> +    newpath = g_strdup_printf("/%s", path ? path : "");
>      g_free(path);
>  
>      return newpath;

-- 
SUSE LINUX Products GmbH, Maxfeldstr. 5, 90409 Nürnberg, Germany
GF: Jeff Hawn, Jennifer Guild, Felix Imendörffer; HRB 16746 AG Nürnberg

Re: [Qemu-devel] [PATCH v5 1/7] object: add object_get_canonical_basename()

[Igor Mammedov]

On Fri, 28 Feb 2014 18:15:55 +0100
Andreas Färber <afaerber@suse.de> wrote:

> Am 28.02.2014 16:18, schrieb Stefan Hajnoczi:
> > It is often useful to find an object's child property name.  Also use
> > this new function to simplify the implementation of
> > object_get_canonical_path().
> > 
> > Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
> > ---
> >  include/qom/object.h |  8 ++++++++
> >  qom/object.c         | 53 ++++++++++++++++++++++++++++++----------------------
> >  2 files changed, 39 insertions(+), 22 deletions(-)
> > 
> > diff --git a/include/qom/object.h b/include/qom/object.h
> > index 9c7c361..8c6db7c 100644
> > --- a/include/qom/object.h
> > +++ b/include/qom/object.h
> > @@ -974,6 +974,14 @@ const char *object_property_get_type(Object *obj, const char *name,
> >  Object *object_get_root(void);
> >  
> >  /**
> > + * object_get_canonical_basename:
> > + *
> > + * Returns: The final component in the object's canonical path.  The canonical
> > + * path is the path within the composition tree starting from the root.
> > + */
> > +gchar *object_get_canonical_basename(Object *obj);
> 
> I find this name very confusing. There is no such thing as a canonical
> base name, ..._canonical_path_component would make its purpose much more
> obvious.
> 
> An underlying issue here probably is that Anthony didn't want a public
> API to access the Object::parent. But the only other user is
> iothread_get_id() in 4/7, so we might just loop over its known path
> prefix there to discover the right child<>. On the other hand, couldn't
> device IDs in /machine/peripheral benefit from this today, too?
there is another user for this, I need similar routine for memdev object.

alternatively it seems that Object that has parent is always name so we could
add 'name' property to Object and replace with it device's 'id'

> 
> In any way I would've liked to get CC'ed on this QOM API proposal please.
> 
> > +
> > +/**
> >   * object_get_canonical_path:
> >   *
> >   * Returns: The canonical path for a object.  This is the path within the
> > diff --git a/qom/object.c b/qom/object.c
> > index 660859c..0cdc319 100644
> > --- a/qom/object.c
> > +++ b/qom/object.c
> > @@ -1102,39 +1102,48 @@ void object_property_add_link(Object *obj, const char *name,
> >      g_free(full_type);
> >  }
> >  
> > +gchar *object_get_canonical_basename(Object *obj)
> > +{
> > +    ObjectProperty *prop = NULL;
> > +
> > +    g_assert(obj->parent != NULL);
> 
> It might make sense to assert obj first? Accessing ->parent would not be
> much different from ->parent->properties otherwise. But applies to the
> original code as well and the movement looks OK.
> 
> Regards,
> Andreas
> 
> > +
> > +    QTAILQ_FOREACH(prop, &obj->parent->properties, node) {
> > +        if (!object_property_is_child(prop)) {
> > +            continue;
> > +        }
> > +
> > +        if (prop->opaque == obj) {
> > +            return g_strdup(prop->name);
> > +        }
> > +    }
> > +
> > +    /* obj had a parent but was not a child, should never happen */
> > +    g_assert_not_reached();
> > +    return NULL;
> > +}
> > +
> >  gchar *object_get_canonical_path(Object *obj)
> >  {
> >      Object *root = object_get_root();
> > -    char *newpath = NULL, *path = NULL;
> > +    char *newpath, *path = NULL;
> >  
> >      while (obj != root) {
> > -        ObjectProperty *prop = NULL;
> > -
> > -        g_assert(obj->parent != NULL);
> > -
> > -        QTAILQ_FOREACH(prop, &obj->parent->properties, node) {
> > -            if (!object_property_is_child(prop)) {
> > -                continue;
> > -            }
> > +        char *component = object_get_canonical_basename(obj);
> >  
> > -            if (prop->opaque == obj) {
> > -                if (path) {
> > -                    newpath = g_strdup_printf("%s/%s", prop->name, path);
> > -                    g_free(path);
> > -                    path = newpath;
> > -                } else {
> > -                    path = g_strdup(prop->name);
> > -                }
> > -                break;
> > -            }
> > +        if (path) {
> > +            newpath = g_strdup_printf("%s/%s", component, path);
> > +            g_free(component);
> > +            g_free(path);
> > +            path = newpath;
> > +        } else {
> > +            path = component;
> >          }
> >  
> > -        g_assert(prop != NULL);
> > -
> >          obj = obj->parent;
> >      }
> >  
> > -    newpath = g_strdup_printf("/%s", path);
> > +    newpath = g_strdup_printf("/%s", path ? path : "");
> >      g_free(path);
> >  
> >      return newpath;
> 

Re: [Qemu-devel] [PATCH v5 1/7] object: add object_get_canonical_basename()

[Stefan Hajnoczi]

On Fri, Feb 28, 2014 at 06:15:55PM +0100, Andreas Färber wrote:
> Am 28.02.2014 16:18, schrieb Stefan Hajnoczi:
> > It is often useful to find an object's child property name.  Also use
> > this new function to simplify the implementation of
> > object_get_canonical_path().
> > 
> > Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
> > ---
> >  include/qom/object.h |  8 ++++++++
> >  qom/object.c         | 53 ++++++++++++++++++++++++++++++----------------------
> >  2 files changed, 39 insertions(+), 22 deletions(-)
> > 
> > diff --git a/include/qom/object.h b/include/qom/object.h
> > index 9c7c361..8c6db7c 100644
> > --- a/include/qom/object.h
> > +++ b/include/qom/object.h
> > @@ -974,6 +974,14 @@ const char *object_property_get_type(Object *obj, const char *name,
> >  Object *object_get_root(void);
> >  
> >  /**
> > + * object_get_canonical_basename:
> > + *
> > + * Returns: The final component in the object's canonical path.  The canonical
> > + * path is the path within the composition tree starting from the root.
> > + */
> > +gchar *object_get_canonical_basename(Object *obj);
> 
> I find this name very confusing. There is no such thing as a canonical
> base name, ..._canonical_path_component would make its purpose much more
> obvious.

If it's confusing then we need to choose a different name.  I will
rename it to object_get_canonical_path_component() like you suggested.

That said, I still think the name I chose makes sense.  We already have
object_get_canonical_path() and there are POSIX dirname(3)/basename(3)
APIs which split paths.  So this is basically
basename(object_get_canonical_path()), hence
object_get_canonical_basename().  And it is canonical because, while
there may be link properties with different names that point to this
object, only this property name comes from the canonical path.

> An underlying issue here probably is that Anthony didn't want a public
> API to access the Object::parent. But the only other user is
> iothread_get_id() in 4/7, so we might just loop over its known path
> prefix there to discover the right child<>. On the other hand, couldn't
> device IDs in /machine/peripheral benefit from this today, too?

Looping over properties of a well-known object is an alternative but
also requires a change:

  int object_child_foreach(Object *obj,
                           int (*fn)(Object *child, void *opaque),
                           void *opaque);

Notice this API does not provide the name of the child to fn()!

So either way, we need to extend the QOM API.

> In any way I would've liked to get CC'ed on this QOM API proposal please.

Will do in the future.

> > @@ -1102,39 +1102,48 @@ void object_property_add_link(Object *obj, const char *name,
> >      g_free(full_type);
> >  }
> >  
> > +gchar *object_get_canonical_basename(Object *obj)
> > +{
> > +    ObjectProperty *prop = NULL;
> > +
> > +    g_assert(obj->parent != NULL);
> 
> It might make sense to assert obj first? Accessing ->parent would not be
> much different from ->parent->properties otherwise. But applies to the
> original code as well and the movement looks OK.

Will fix.

Stefan

[Qemu-devel] [PATCH v5 2/7] rfifolock: add recursive FIFO lock

[Stefan Hajnoczi]

QemuMutex does not guarantee fairness and cannot be acquired
recursively:

Fairness means each locker gets a turn and the scheduler cannot cause
starvation.

Recursive locking is useful for composition, it allows a sequence of
locking operations to be invoked atomically by acquiring the lock around
them.

This patch adds RFifoLock, a recursive lock that guarantees FIFO order.
Its first user is added in the next patch.

RFifoLock has one additional feature: it can be initialized with an
optional contention callback.  The callback is invoked whenever a thread
must wait for the lock.  For example, it can be used to poke the current
owner so that they release the lock soon.

Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
---
 include/qemu/rfifolock.h | 54 +++++++++++++++++++++++++++++
 tests/Makefile           |  2 ++
 tests/test-rfifolock.c   | 90 ++++++++++++++++++++++++++++++++++++++++++++++++
 util/Makefile.objs       |  1 +
 util/rfifolock.c         | 78 +++++++++++++++++++++++++++++++++++++++++
 5 files changed, 225 insertions(+)
 create mode 100644 include/qemu/rfifolock.h
 create mode 100644 tests/test-rfifolock.c
 create mode 100644 util/rfifolock.c

diff --git a/include/qemu/rfifolock.h b/include/qemu/rfifolock.h
new file mode 100644
index 0000000..b23ab53
--- /dev/null
+++ b/include/qemu/rfifolock.h
@@ -0,0 +1,54 @@
+/*
+ * Recursive FIFO lock
+ *
+ * Copyright Red Hat, Inc. 2013
+ *
+ * Authors:
+ *  Stefan Hajnoczi   <stefanha@redhat.com>
+ *
+ * This work is licensed under the terms of the GNU GPL, version 2 or later.
+ * See the COPYING file in the top-level directory.
+ *
+ */
+
+#ifndef QEMU_RFIFOLOCK_H
+#define QEMU_RFIFOLOCK_H
+
+#include "qemu/thread.h"
+
+/* Recursive FIFO lock
+ *
+ * This lock provides more features than a plain mutex:
+ *
+ * 1. Fairness - enforces FIFO order.
+ * 2. Nesting - can be taken recursively.
+ * 3. Contention callback - optional, called when thread must wait.
+ *
+ * The recursive FIFO lock is heavyweight so prefer other synchronization
+ * primitives if you do not need its features.
+ */
+typedef struct {
+    QemuMutex lock;             /* protects all fields */
+
+    /* FIFO order */
+    unsigned int head;          /* active ticket number */
+    unsigned int tail;          /* waiting ticket number */
+    QemuCond cond;              /* used to wait for our ticket number */
+
+    /* Nesting */
+    QemuThread owner_thread;    /* thread that currently has ownership */
+    unsigned int nesting;       /* amount of nesting levels */
+
+    /* Contention callback */
+    void (*cb)(void *);         /* called when thread must wait, with ->lock
+                                 * held so it may not recursively lock/unlock
+                                 */
+    void *cb_opaque;
+} RFifoLock;
+
+void rfifolock_init(RFifoLock *r, void (*cb)(void *), void *opaque);
+void rfifolock_destroy(RFifoLock *r);
+void rfifolock_lock(RFifoLock *r);
+void rfifolock_unlock(RFifoLock *r);
+
+#endif /* QEMU_RFIFOLOCK_H */
diff --git a/tests/Makefile b/tests/Makefile
index b17d41e..78e9803 100644
--- a/tests/Makefile
+++ b/tests/Makefile
@@ -35,6 +35,7 @@ check-unit-y += tests/test-visitor-serialization$(EXESUF)
 check-unit-y += tests/test-iov$(EXESUF)
 gcov-files-test-iov-y = util/iov.c
 check-unit-y += tests/test-aio$(EXESUF)
+check-unit-y += tests/test-rfifolock$(EXESUF)
 check-unit-y += tests/test-throttle$(EXESUF)
 gcov-files-test-aio-$(CONFIG_WIN32) = aio-win32.c
 gcov-files-test-aio-$(CONFIG_POSIX) = aio-posix.c
@@ -172,6 +173,7 @@ tests/check-qjson$(EXESUF): tests/check-qjson.o libqemuutil.a libqemustub.a
 tests/check-qom-interface$(EXESUF): tests/check-qom-interface.o $(qom-core-obj) libqemuutil.a libqemustub.a
 tests/test-coroutine$(EXESUF): tests/test-coroutine.o $(block-obj-y) libqemuutil.a libqemustub.a
 tests/test-aio$(EXESUF): tests/test-aio.o $(block-obj-y) libqemuutil.a libqemustub.a
+tests/test-rfifolock$(EXESUF): tests/test-rfifolock.o libqemuutil.a libqemustub.a
 tests/test-throttle$(EXESUF): tests/test-throttle.o $(block-obj-y) libqemuutil.a libqemustub.a
 tests/test-thread-pool$(EXESUF): tests/test-thread-pool.o $(block-obj-y) libqemuutil.a libqemustub.a
 tests/test-iov$(EXESUF): tests/test-iov.o libqemuutil.a
diff --git a/tests/test-rfifolock.c b/tests/test-rfifolock.c
new file mode 100644
index 0000000..440dbcb
--- /dev/null
+++ b/tests/test-rfifolock.c
@@ -0,0 +1,90 @@
+/*
+ * RFifoLock tests
+ *
+ * Copyright Red Hat, Inc. 2013
+ *
+ * Authors:
+ *  Stefan Hajnoczi    <stefanha@redhat.com>
+ *
+ * This work is licensed under the terms of the GNU LGPL, version 2 or later.
+ * See the COPYING.LIB file in the top-level directory.
+ */
+
+#include <glib.h>
+#include "qemu-common.h"
+#include "qemu/rfifolock.h"
+
+static void test_nesting(void)
+{
+    RFifoLock lock;
+
+    /* Trivial test, ensure the lock is recursive */
+    rfifolock_init(&lock, NULL, NULL);
+    rfifolock_lock(&lock);
+    rfifolock_lock(&lock);
+    rfifolock_lock(&lock);
+    rfifolock_unlock(&lock);
+    rfifolock_unlock(&lock);
+    rfifolock_unlock(&lock);
+    rfifolock_destroy(&lock);
+}
+
+typedef struct {
+    RFifoLock lock;
+    int fd[2];
+} CallbackTestData;
+
+static void rfifolock_cb(void *opaque)
+{
+    CallbackTestData *data = opaque;
+    int ret;
+    char c = 0;
+
+    ret = write(data->fd[1], &c, sizeof(c));
+    g_assert(ret == 1);
+}
+
+static void *callback_thread(void *opaque)
+{
+    CallbackTestData *data = opaque;
+
+    /* The other thread holds the lock so the contention callback will be
+     * invoked...
+     */
+    rfifolock_lock(&data->lock);
+    rfifolock_unlock(&data->lock);
+    return NULL;
+}
+
+static void test_callback(void)
+{
+    CallbackTestData data;
+    QemuThread thread;
+    int ret;
+    char c;
+
+    rfifolock_init(&data.lock, rfifolock_cb, &data);
+    ret = qemu_pipe(data.fd);
+    g_assert(ret == 0);
+
+    /* Hold lock but allow the callback to kick us by writing to the pipe */
+    rfifolock_lock(&data.lock);
+    qemu_thread_create(&thread, callback_thread, &data, QEMU_THREAD_JOINABLE);
+    ret = read(data.fd[0], &c, sizeof(c));
+    g_assert(ret == 1);
+    rfifolock_unlock(&data.lock);
+    /* If we got here then the callback was invoked, as expected */
+
+    qemu_thread_join(&thread);
+    close(data.fd[0]);
+    close(data.fd[1]);
+    rfifolock_destroy(&data.lock);
+}
+
+int main(int argc, char **argv)
+{
+    g_test_init(&argc, &argv, NULL);
+    g_test_add_func("/nesting", test_nesting);
+    g_test_add_func("/callback", test_callback);
+    return g_test_run();
+}
diff --git a/util/Makefile.objs b/util/Makefile.objs
index 937376b..df83b62 100644
--- a/util/Makefile.objs
+++ b/util/Makefile.objs
@@ -14,3 +14,4 @@ util-obj-y += crc32c.o
 util-obj-y += throttle.o
 util-obj-y += getauxval.o
 util-obj-y += readline.o
+util-obj-y += rfifolock.o
diff --git a/util/rfifolock.c b/util/rfifolock.c
new file mode 100644
index 0000000..afbf748
--- /dev/null
+++ b/util/rfifolock.c
@@ -0,0 +1,78 @@
+/*
+ * Recursive FIFO lock
+ *
+ * Copyright Red Hat, Inc. 2013
+ *
+ * Authors:
+ *  Stefan Hajnoczi   <stefanha@redhat.com>
+ *
+ * This work is licensed under the terms of the GNU LGPL, version 2 or later.
+ * See the COPYING.LIB file in the top-level directory.
+ *
+ */
+
+#include <assert.h>
+#include "qemu/rfifolock.h"
+
+void rfifolock_init(RFifoLock *r, void (*cb)(void *), void *opaque)
+{
+    qemu_mutex_init(&r->lock);
+    r->head = 0;
+    r->tail = 0;
+    qemu_cond_init(&r->cond);
+    r->nesting = 0;
+    r->cb = cb;
+    r->cb_opaque = opaque;
+}
+
+void rfifolock_destroy(RFifoLock *r)
+{
+    qemu_cond_destroy(&r->cond);
+    qemu_mutex_destroy(&r->lock);
+}
+
+/*
+ * Theory of operation:
+ *
+ * In order to ensure FIFO ordering, implement a ticketlock.  Threads acquiring
+ * the lock enqueue themselves by incrementing the tail index.  When the lock
+ * is unlocked, the head is incremented and waiting threads are notified.
+ *
+ * Recursive locking does not take a ticket since the head is only incremented
+ * when the outermost recursive caller unlocks.
+ */
+void rfifolock_lock(RFifoLock *r)
+{
+    qemu_mutex_lock(&r->lock);
+
+    /* Take a ticket */
+    unsigned int ticket = r->tail++;
+
+    if (r->nesting > 0 && qemu_thread_is_self(&r->owner_thread)) {
+        r->tail--; /* put ticket back, we're nesting */
+    } else {
+        while (ticket != r->head) {
+            /* Invoke optional contention callback */
+            if (r->cb) {
+                r->cb(r->cb_opaque);
+            }
+            qemu_cond_wait(&r->cond, &r->lock);
+        }
+    }
+
+    qemu_thread_get_self(&r->owner_thread);
+    r->nesting++;
+    qemu_mutex_unlock(&r->lock);
+}
+
+void rfifolock_unlock(RFifoLock *r)
+{
+    qemu_mutex_lock(&r->lock);
+    assert(r->nesting > 0);
+    assert(qemu_thread_is_self(&r->owner_thread));
+    if (--r->nesting == 0) {
+        r->head++;
+        qemu_cond_broadcast(&r->cond);
+    }
+    qemu_mutex_unlock(&r->lock);
+}
-- 
1.8.5.3

[Qemu-devel] [PATCH v5 3/7] aio: add aio_context_acquire() and aio_context_release()

[Stefan Hajnoczi]

It can be useful to run an AioContext from a thread which normally does
not "own" the AioContext.  For example, request draining can be
implemented by acquiring the AioContext and looping aio_poll() until all
requests have been completed.

The following pattern should work:

  /* Event loop thread */
  while (running) {
      aio_context_acquire(ctx);
      aio_poll(ctx, true);
      aio_context_release(ctx);
  }

  /* Another thread */
  aio_context_acquire(ctx);
  bdrv_read(bs, 0x1000, buf, 1);
  aio_context_release(ctx);

This patch implements aio_context_acquire() and aio_context_release().

Note that existing aio_poll() callers do not need to worry about
acquiring and releasing - it is only needed when multiple threads will
call aio_poll() on the same AioContext.

Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
---
 async.c             | 18 +++++++++++++++++
 include/block/aio.h | 18 +++++++++++++++++
 tests/test-aio.c    | 58 +++++++++++++++++++++++++++++++++++++++++++++++++++++
 3 files changed, 94 insertions(+)

diff --git a/async.c b/async.c
index 5fb3fa6..6930185 100644
--- a/async.c
+++ b/async.c
@@ -214,6 +214,7 @@ aio_ctx_finalize(GSource     *source)
     thread_pool_free(ctx->thread_pool);
     aio_set_event_notifier(ctx, &ctx->notifier, NULL);
     event_notifier_cleanup(&ctx->notifier);
+    rfifolock_destroy(&ctx->lock);
     qemu_mutex_destroy(&ctx->bh_lock);
     g_array_free(ctx->pollfds, TRUE);
     timerlistgroup_deinit(&ctx->tlg);
@@ -250,6 +251,12 @@ static void aio_timerlist_notify(void *opaque)
     aio_notify(opaque);
 }
 
+static void aio_rfifolock_cb(void *opaque)
+{
+    /* Kick owner thread in case they are blocked in aio_poll() */
+    aio_notify(opaque);
+}
+
 AioContext *aio_context_new(void)
 {
     AioContext *ctx;
@@ -257,6 +264,7 @@ AioContext *aio_context_new(void)
     ctx->pollfds = g_array_new(FALSE, FALSE, sizeof(GPollFD));
     ctx->thread_pool = NULL;
     qemu_mutex_init(&ctx->bh_lock);
+    rfifolock_init(&ctx->lock, aio_rfifolock_cb, ctx);
     event_notifier_init(&ctx->notifier, false);
     aio_set_event_notifier(ctx, &ctx->notifier, 
                            (EventNotifierHandler *)
@@ -275,3 +283,13 @@ void aio_context_unref(AioContext *ctx)
 {
     g_source_unref(&ctx->source);
 }
+
+void aio_context_acquire(AioContext *ctx)
+{
+    rfifolock_lock(&ctx->lock);
+}
+
+void aio_context_release(AioContext *ctx)
+{
+    rfifolock_unlock(&ctx->lock);
+}
diff --git a/include/block/aio.h b/include/block/aio.h
index 2efdf41..a92511b 100644
--- a/include/block/aio.h
+++ b/include/block/aio.h
@@ -19,6 +19,7 @@
 #include "qemu/queue.h"
 #include "qemu/event_notifier.h"
 #include "qemu/thread.h"
+#include "qemu/rfifolock.h"
 #include "qemu/timer.h"
 
 typedef struct BlockDriverAIOCB BlockDriverAIOCB;
@@ -47,6 +48,9 @@ typedef void IOHandler(void *opaque);
 struct AioContext {
     GSource source;
 
+    /* Protects all fields from multi-threaded access */
+    RFifoLock lock;
+
     /* The list of registered AIO handlers */
     QLIST_HEAD(, AioHandler) aio_handlers;
 
@@ -104,6 +108,20 @@ void aio_context_ref(AioContext *ctx);
  */
 void aio_context_unref(AioContext *ctx);
 
+/* Take ownership of the AioContext.  If the AioContext will be shared between
+ * threads, a thread must have ownership when calling aio_poll().
+ *
+ * Note that multiple threads calling aio_poll() means timers, BHs, and
+ * callbacks may be invoked from a different thread than they were registered
+ * from.  Therefore, code must use AioContext acquire/release or use
+ * fine-grained synchronization to protect shared state if other threads will
+ * be accessing it simultaneously.
+ */
+void aio_context_acquire(AioContext *ctx);
+
+/* Relinquish ownership of the AioContext. */
+void aio_context_release(AioContext *ctx);
+
 /**
  * aio_bh_new: Allocate a new bottom half structure.
  *
diff --git a/tests/test-aio.c b/tests/test-aio.c
index 592721e..d384b0b 100644
--- a/tests/test-aio.c
+++ b/tests/test-aio.c
@@ -112,6 +112,63 @@ static void test_notify(void)
     g_assert(!aio_poll(ctx, false));
 }
 
+typedef struct {
+    QemuMutex start_lock;
+    bool thread_acquired;
+} AcquireTestData;
+
+static void *test_acquire_thread(void *opaque)
+{
+    AcquireTestData *data = opaque;
+
+    /* Wait for other thread to let us start */
+    qemu_mutex_lock(&data->start_lock);
+    qemu_mutex_unlock(&data->start_lock);
+
+    aio_context_acquire(ctx);
+    aio_context_release(ctx);
+
+    data->thread_acquired = true; /* success, we got here */
+
+    return NULL;
+}
+
+static void dummy_notifier_read(EventNotifier *unused)
+{
+    g_assert(false); /* should never be invoked */
+}
+
+static void test_acquire(void)
+{
+    QemuThread thread;
+    EventNotifier notifier;
+    AcquireTestData data;
+
+    /* Dummy event notifier ensures aio_poll() will block */
+    event_notifier_init(&notifier, false);
+    aio_set_event_notifier(ctx, &notifier, dummy_notifier_read);
+    g_assert(!aio_poll(ctx, false)); /* consume aio_notify() */
+
+    qemu_mutex_init(&data.start_lock);
+    qemu_mutex_lock(&data.start_lock);
+    data.thread_acquired = false;
+
+    qemu_thread_create(&thread, test_acquire_thread,
+                       &data, QEMU_THREAD_JOINABLE);
+
+    /* Block in aio_poll(), let other thread kick us and acquire context */
+    aio_context_acquire(ctx);
+    qemu_mutex_unlock(&data.start_lock); /* let the thread run */
+    g_assert(!aio_poll(ctx, true));
+    aio_context_release(ctx);
+
+    qemu_thread_join(&thread);
+    aio_set_event_notifier(ctx, &notifier, NULL);
+    event_notifier_cleanup(&notifier);
+
+    g_assert(data.thread_acquired);
+}
+
 static void test_bh_schedule(void)
 {
     BHTestData data = { .n = 0 };
@@ -775,6 +832,7 @@ int main(int argc, char **argv)
 
     g_test_init(&argc, &argv, NULL);
     g_test_add_func("/aio/notify",                  test_notify);
+    g_test_add_func("/aio/acquire",                 test_acquire);
     g_test_add_func("/aio/bh/schedule",             test_bh_schedule);
     g_test_add_func("/aio/bh/schedule10",           test_bh_schedule10);
     g_test_add_func("/aio/bh/cancel",               test_bh_cancel);
-- 
1.8.5.3

[Qemu-devel] [PATCH v5 4/7] iothread: add I/O thread object

[Stefan Hajnoczi]

This is a stand-in for Michael Roth's QContext.  I expect this to be
replaced once QContext is completed.

The IOThread object is an AioContext event loop thread.  This patch adds
the concept of multiple event loop threads, allowing users to define
them.

When SMP guests run on SMP hosts it makes sense to instantiate multiple
IOThreads.  This spreads event loop processing across multiple cores.
Note that additional patches are required to actually bind a device to
an IOThread.

Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
---
 Makefile.objs             |   1 +
 include/sysemu/iothread.h |  30 ++++++++++++
 iothread.c                | 119 ++++++++++++++++++++++++++++++++++++++++++++++
 3 files changed, 150 insertions(+)
 create mode 100644 include/sysemu/iothread.h
 create mode 100644 iothread.c

diff --git a/Makefile.objs b/Makefile.objs
index 4a62913..846ed66 100644
--- a/Makefile.objs
+++ b/Makefile.objs
@@ -44,6 +44,7 @@ libcacard-y += libcacard/vcardt.o
 
 ifeq ($(CONFIG_SOFTMMU),y)
 common-obj-y = blockdev.o blockdev-nbd.o block/
+common-obj-y += iothread.o
 common-obj-y += net/
 common-obj-y += qdev-monitor.o device-hotplug.o
 common-obj-$(CONFIG_WIN32) += os-win32.o
diff --git a/include/sysemu/iothread.h b/include/sysemu/iothread.h
new file mode 100644
index 0000000..a32214a
--- /dev/null
+++ b/include/sysemu/iothread.h
@@ -0,0 +1,30 @@
+/*
+ * Event loop thread
+ *
+ * Copyright Red Hat Inc., 2013
+ *
+ * Authors:
+ *  Stefan Hajnoczi   <stefanha@redhat.com>
+ *
+ * This work is licensed under the terms of the GNU GPL, version 2 or later.
+ * See the COPYING file in the top-level directory.
+ *
+ */
+
+#ifndef IOTHREAD_H
+#define IOTHREAD_H
+
+#include "block/aio.h"
+
+#define TYPE_IOTHREAD "iothread"
+
+typedef struct IOThread IOThread;
+
+#define IOTHREAD(obj) \
+   OBJECT_CHECK(IOThread, obj, TYPE_IOTHREAD)
+
+IOThread *iothread_find(const char *id);
+char *iothread_get_id(IOThread *iothread);
+AioContext *iothread_get_aio_context(IOThread *iothread);
+
+#endif /* IOTHREAD_H */
diff --git a/iothread.c b/iothread.c
new file mode 100644
index 0000000..95ff330
--- /dev/null
+++ b/iothread.c
@@ -0,0 +1,119 @@
+/*
+ * Event loop thread
+ *
+ * Copyright Red Hat Inc., 2013
+ *
+ * Authors:
+ *  Stefan Hajnoczi   <stefanha@redhat.com>
+ *
+ * This work is licensed under the terms of the GNU GPL, version 2 or later.
+ * See the COPYING file in the top-level directory.
+ *
+ */
+
+#include "qom/object.h"
+#include "qom/object_interfaces.h"
+#include "qemu/module.h"
+#include "qemu/thread.h"
+#include "block/aio.h"
+#include "sysemu/iothread.h"
+
+#define IOTHREADS_PATH "/objects"
+
+typedef ObjectClass IOThreadClass;
+struct IOThread {
+    Object parent;
+    QemuThread thread;
+    AioContext *ctx;
+    bool stopping;
+};
+
+#define IOTHREAD_GET_CLASS(obj) \
+   OBJECT_GET_CLASS(IOThreadClass, obj, TYPE_IOTHREAD)
+#define IOTHREAD_CLASS(klass) \
+   OBJECT_CLASS_CHECK(IOThreadClass, klass, TYPE_IOTHREAD)
+
+static void *iothread_run(void *opaque)
+{
+    IOThread *iothread = opaque;
+
+    while (!iothread->stopping) {
+        aio_context_acquire(iothread->ctx);
+        while (!iothread->stopping && aio_poll(iothread->ctx, true)) {
+            /* Progress was made, keep going */
+        }
+        aio_context_release(iothread->ctx);
+    }
+    return NULL;
+}
+
+static void iothread_instance_finalize(Object *obj)
+{
+    IOThread *iothread = IOTHREAD(obj);
+
+    iothread->stopping = true;
+    aio_notify(iothread->ctx);
+    qemu_thread_join(&iothread->thread);
+    aio_context_unref(iothread->ctx);
+}
+
+static void iothread_complete(UserCreatable *obj, Error **errp)
+{
+    IOThread *iothread = IOTHREAD(obj);
+
+    iothread->stopping = false;
+    iothread->ctx = aio_context_new();
+
+    /* This assumes we are called from a thread with useful CPU affinity for us
+     * to inherit.
+     */
+    qemu_thread_create(&iothread->thread, iothread_run,
+                       iothread, QEMU_THREAD_JOINABLE);
+}
+
+static void iothread_class_init(ObjectClass *klass, void *class_data)
+{
+    UserCreatableClass *ucc = USER_CREATABLE_CLASS(klass);
+    ucc->complete = iothread_complete;
+}
+
+static const TypeInfo iothread_info = {
+    .name = TYPE_IOTHREAD,
+    .parent = TYPE_OBJECT,
+    .class_init = iothread_class_init,
+    .instance_size = sizeof(IOThread),
+    .instance_finalize = iothread_instance_finalize,
+    .interfaces = (InterfaceInfo[]) {
+        {TYPE_USER_CREATABLE},
+        {}
+    },
+};
+
+static void iothread_register_types(void)
+{
+    type_register_static(&iothread_info);
+}
+
+type_init(iothread_register_types)
+
+IOThread *iothread_find(const char *id)
+{
+    Object *container = container_get(object_get_root(), IOTHREADS_PATH);
+    Object *child;
+
+    child = object_property_get_link(container, id, NULL);
+    if (!child) {
+        return NULL;
+    }
+    return (IOThread *)object_dynamic_cast(child, TYPE_IOTHREAD);
+}
+
+char *iothread_get_id(IOThread *iothread)
+{
+    return object_get_canonical_basename(OBJECT(iothread));
+}
+
+AioContext *iothread_get_aio_context(IOThread *iothread)
+{
+    return iothread->ctx;
+}
-- 
1.8.5.3

[Qemu-devel] [PATCH v5 5/7] qdev: make get_pointer() handle temporary strings

[Stefan Hajnoczi]

From: Igor Mammedov <imammedo@redhat.com>

get_pointer()'s print() callback might return a heap allocated
string, to avoid adding dedicated get_pointer_foo for this case
convert current print() callbacks to return temporary heap
allocated string and make get_pointer() free it.

Signed-off-by: Igor Mammedov <imammedo@redhat.com>
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
---
 hw/core/qdev-properties-system.c | 19 +++++++++++--------
 1 file changed, 11 insertions(+), 8 deletions(-)

diff --git a/hw/core/qdev-properties-system.c b/hw/core/qdev-properties-system.c
index 5f5957e..878c3bc 100644
--- a/hw/core/qdev-properties-system.c
+++ b/hw/core/qdev-properties-system.c
@@ -20,15 +20,16 @@
 #include "sysemu/char.h"
 
 static void get_pointer(Object *obj, Visitor *v, Property *prop,
-                        const char *(*print)(void *ptr),
+                        char *(*print)(void *ptr),
                         const char *name, Error **errp)
 {
     DeviceState *dev = DEVICE(obj);
     void **ptr = qdev_get_prop_ptr(dev, prop);
     char *p;
 
-    p = (char *) (*ptr ? print(*ptr) : "");
+    p = *ptr ? print(*ptr) : g_strdup("");
     visit_type_str(v, &p, name, errp);
+    g_free(p);
 }
 
 static void set_pointer(Object *obj, Visitor *v, Property *prop,
@@ -91,9 +92,9 @@ static void release_drive(Object *obj, const char *name, void *opaque)
     }
 }
 
-static const char *print_drive(void *ptr)
+static char *print_drive(void *ptr)
 {
-    return bdrv_get_device_name(ptr);
+    return g_strdup(bdrv_get_device_name(ptr));
 }
 
 static void get_drive(Object *obj, Visitor *v, void *opaque,
@@ -145,11 +146,12 @@ static void release_chr(Object *obj, const char *name, void *opaque)
 }
 
 
-static const char *print_chr(void *ptr)
+static char *print_chr(void *ptr)
 {
     CharDriverState *chr = ptr;
+    const char *val = chr->label ? chr->label : "";
 
-    return chr->label ? chr->label : "";
+    return g_strdup(val);
 }
 
 static void get_chr(Object *obj, Visitor *v, void *opaque,
@@ -224,11 +226,12 @@ err:
     return ret;
 }
 
-static const char *print_netdev(void *ptr)
+static char *print_netdev(void *ptr)
 {
     NetClientState *netdev = ptr;
+    const char *val = netdev->name ? netdev->name : "";
 
-    return netdev->name ? netdev->name : "";
+    return g_strdup(val);
 }
 
 static void get_netdev(Object *obj, Visitor *v, void *opaque,
-- 
1.8.5.3

Re: [Qemu-devel] [PATCH v5 5/7] qdev: make get_pointer() handle temporary strings

[Andreas Färber]

Am 28.02.2014 16:18, schrieb Stefan Hajnoczi:
> From: Igor Mammedov <imammedo@redhat.com>
> 
> get_pointer()'s print() callback might return a heap allocated
> string, to avoid adding dedicated get_pointer_foo for this case
> convert current print() callbacks to return temporary heap
> allocated string and make get_pointer() free it.
> 
> Signed-off-by: Igor Mammedov <imammedo@redhat.com>
> Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
> ---
>  hw/core/qdev-properties-system.c | 19 +++++++++++--------
>  1 file changed, 11 insertions(+), 8 deletions(-)

Not aware of the history of this patch (not CC'ed?) but looks okay,

Reviewed-by: Andreas Färber <afaerber@suse.de>

Do we have any users outside of this series?

Regards,
Andreas

-- 
SUSE LINUX Products GmbH, Maxfeldstr. 5, 90409 Nürnberg, Germany
GF: Jeff Hawn, Jennifer Guild, Felix Imendörffer; HRB 16746 AG Nürnberg

Re: [Qemu-devel] [PATCH v5 5/7] qdev: make get_pointer() handle temporary strings

[Stefan Hajnoczi]

On Fri, Feb 28, 2014 at 05:49:09PM +0100, Andreas Färber wrote:
> Am 28.02.2014 16:18, schrieb Stefan Hajnoczi:
> > From: Igor Mammedov <imammedo@redhat.com>
> > 
> > get_pointer()'s print() callback might return a heap allocated
> > string, to avoid adding dedicated get_pointer_foo for this case
> > convert current print() callbacks to return temporary heap
> > allocated string and make get_pointer() free it.
> > 
> > Signed-off-by: Igor Mammedov <imammedo@redhat.com>
> > Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
> > ---
> >  hw/core/qdev-properties-system.c | 19 +++++++++++--------
> >  1 file changed, 11 insertions(+), 8 deletions(-)
> 
> Not aware of the history of this patch (not CC'ed?) but looks okay,
> 
> Reviewed-by: Andreas Färber <afaerber@suse.de>
> 
> Do we have any users outside of this series?

The patch comes from Igor originally so maybe he has future patches that
will use it.

Stefan

Re: [Qemu-devel] [PATCH v5 5/7] qdev: make get_pointer() handle temporary strings

[Igor Mammedov]

On Mon, 3 Mar 2014 10:41:30 +0100
Stefan Hajnoczi <stefanha@redhat.com> wrote:

> On Fri, Feb 28, 2014 at 05:49:09PM +0100, Andreas Färber wrote:
> > Am 28.02.2014 16:18, schrieb Stefan Hajnoczi:
> > > From: Igor Mammedov <imammedo@redhat.com>
> > > 
> > > get_pointer()'s print() callback might return a heap allocated
> > > string, to avoid adding dedicated get_pointer_foo for this case
> > > convert current print() callbacks to return temporary heap
> > > allocated string and make get_pointer() free it.
> > > 
> > > Signed-off-by: Igor Mammedov <imammedo@redhat.com>
> > > Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
> > > ---
> > >  hw/core/qdev-properties-system.c | 19 +++++++++++--------
> > >  1 file changed, 11 insertions(+), 8 deletions(-)
> > 
> > Not aware of the history of this patch (not CC'ed?) but looks okay,
> > 
> > Reviewed-by: Andreas Färber <afaerber@suse.de>
> > 
> > Do we have any users outside of this series?
> 
> The patch comes from Igor originally so maybe he has future patches that
> will use it.
>
> Stefan

I don't have other users/plans to use it.
Patch was suggested as alternative to adding a new API:
 http://lists.gnu.org/archive/html/qemu-devel/2014-02/msg04704.html

[Qemu-devel] [PATCH v5 6/7] iothread: add "iothread" qdev property type

[Stefan Hajnoczi]

Add a "iothread" qdev property type so devices can be hooked up to an
IOThread from the comand-line:

  qemu -object iothread,id=iothread0 \
       -device some-device,x-iothread=iothread0

Note that Paolo Bonzini <pbonzini@redhat.com> has suggested using QOM
links instead.  This way the relationship between the objects is
reflected in QOM.  There are currently shortcomings of
object_property_add_link() which prevent this use case.  I will attempt
to fix them and move to QOM links in a separate series.

Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
---
 hw/core/qdev-properties-system.c | 51 ++++++++++++++++++++++++++++++++++++++++
 include/hw/qdev-properties.h     |  3 +++
 2 files changed, 54 insertions(+)

diff --git a/hw/core/qdev-properties-system.c b/hw/core/qdev-properties-system.c
index 878c3bc..de83561 100644
--- a/hw/core/qdev-properties-system.c
+++ b/hw/core/qdev-properties-system.c
@@ -18,6 +18,7 @@
 #include "net/hub.h"
 #include "qapi/visitor.h"
 #include "sysemu/char.h"
+#include "sysemu/iothread.h"
 
 static void get_pointer(Object *obj, Visitor *v, Property *prop,
                         char *(*print)(void *ptr),
@@ -385,6 +386,56 @@ void qdev_set_nic_properties(DeviceState *dev, NICInfo *nd)
     nd->instantiated = 1;
 }
 
+/* --- iothread --- */
+
+static char *print_iothread(void *ptr)
+{
+    return iothread_get_id(ptr);
+}
+
+static int parse_iothread(DeviceState *dev, const char *str, void **ptr)
+{
+    IOThread *iothread;
+
+    iothread = iothread_find(str);
+    if (!iothread) {
+        return -ENOENT;
+    }
+    object_ref(OBJECT(iothread));
+    *ptr = iothread;
+    return 0;
+}
+
+static void get_iothread(Object *obj, struct Visitor *v, void *opaque,
+                         const char *name, Error **errp)
+{
+    get_pointer(obj, v, opaque, print_iothread, name, errp);
+}
+
+static void set_iothread(Object *obj, struct Visitor *v, void *opaque,
+                         const char *name, Error **errp)
+{
+    set_pointer(obj, v, opaque, parse_iothread, name, errp);
+}
+
+static void release_iothread(Object *obj, const char *name, void *opaque)
+{
+    DeviceState *dev = DEVICE(obj);
+    Property *prop = opaque;
+    IOThread **ptr = qdev_get_prop_ptr(dev, prop);
+
+    if (*ptr) {
+        object_unref(OBJECT(*ptr));
+    }
+}
+
+PropertyInfo qdev_prop_iothread = {
+    .name = "iothread",
+    .get = get_iothread,
+    .set = set_iothread,
+    .release = release_iothread,
+};
+
 static int qdev_add_one_global(QemuOpts *opts, void *opaque)
 {
     GlobalProperty *g;
diff --git a/include/hw/qdev-properties.h b/include/hw/qdev-properties.h
index 0c0babf..3c000ee 100644
--- a/include/hw/qdev-properties.h
+++ b/include/hw/qdev-properties.h
@@ -22,6 +22,7 @@ extern PropertyInfo qdev_prop_bios_chs_trans;
 extern PropertyInfo qdev_prop_drive;
 extern PropertyInfo qdev_prop_netdev;
 extern PropertyInfo qdev_prop_vlan;
+extern PropertyInfo qdev_prop_iothread;
 extern PropertyInfo qdev_prop_pci_devfn;
 extern PropertyInfo qdev_prop_blocksize;
 extern PropertyInfo qdev_prop_pci_host_devaddr;
@@ -142,6 +143,8 @@ extern PropertyInfo qdev_prop_arraylen;
     DEFINE_PROP(_n, _s, _f, qdev_prop_vlan, NICPeers)
 #define DEFINE_PROP_DRIVE(_n, _s, _f) \
     DEFINE_PROP(_n, _s, _f, qdev_prop_drive, BlockDriverState *)
+#define DEFINE_PROP_IOTHREAD(_n, _s, _f)             \
+    DEFINE_PROP(_n, _s, _f, qdev_prop_iothread, IOThread *)
 #define DEFINE_PROP_MACADDR(_n, _s, _f)         \
     DEFINE_PROP(_n, _s, _f, qdev_prop_macaddr, MACAddr)
 #define DEFINE_PROP_LOSTTICKPOLICY(_n, _s, _f, _d) \
-- 
1.8.5.3

[Qemu-devel] [PATCH v5 7/7] dataplane: replace internal thread with IOThread

[Stefan Hajnoczi]

Today virtio-blk dataplane uses a 1:1 device-per-thread model.  Now that
IOThreads have been introduced we can generalize this to N:M devices per
threads.

This patch drops thread code from dataplane in favor of running inside
an IOThread AioContext.

As a bonus we solve the case where a guest keeps submitting I/O requests
while dataplane is trying to stop.  Previously the dataplane thread
would continue to process requests until the request gave it a break.
Now we can shut down in bounded time thanks to
aio_context_acquire/release.

Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
---
 hw/block/dataplane/virtio-blk.c | 96 +++++++++++++++++++++++------------------
 include/hw/virtio/virtio-blk.h  |  8 +++-
 2 files changed, 60 insertions(+), 44 deletions(-)

diff --git a/hw/block/dataplane/virtio-blk.c b/hw/block/dataplane/virtio-blk.c
index 2237edb..a5afc21 100644
--- a/hw/block/dataplane/virtio-blk.c
+++ b/hw/block/dataplane/virtio-blk.c
@@ -23,6 +23,7 @@
 #include "virtio-blk.h"
 #include "block/aio.h"
 #include "hw/virtio/virtio-bus.h"
+#include "monitor/monitor.h" /* for object_add() */
 
 enum {
     SEG_MAX = 126,                  /* maximum number of I/O segments */
@@ -44,8 +45,6 @@ struct VirtIOBlockDataPlane {
     bool started;
     bool starting;
     bool stopping;
-    QEMUBH *start_bh;
-    QemuThread thread;
 
     VirtIOBlkConf *blk;
     int fd;                         /* image file descriptor */
@@ -59,12 +58,14 @@ struct VirtIOBlockDataPlane {
      * (because you don't own the file descriptor or handle; you just
      * use it).
      */
+    IOThread *iothread;
+    bool internal_iothread;
     AioContext *ctx;
     EventNotifier io_notifier;      /* Linux AIO completion */
     EventNotifier host_notifier;    /* doorbell */
 
     IOQueue ioqueue;                /* Linux AIO queue (should really be per
-                                       dataplane thread) */
+                                       IOThread) */
     VirtIOBlockRequest requests[REQ_MAX]; /* pool of requests, managed by the
                                              queue */
 
@@ -342,26 +343,7 @@ static void handle_io(EventNotifier *e)
     }
 }
 
-static void *data_plane_thread(void *opaque)
-{
-    VirtIOBlockDataPlane *s = opaque;
-
-    while (!s->stopping || s->num_reqs > 0) {
-        aio_poll(s->ctx, true);
-    }
-    return NULL;
-}
-
-static void start_data_plane_bh(void *opaque)
-{
-    VirtIOBlockDataPlane *s = opaque;
-
-    qemu_bh_delete(s->start_bh);
-    s->start_bh = NULL;
-    qemu_thread_create(&s->thread, data_plane_thread,
-                       s, QEMU_THREAD_JOINABLE);
-}
-
+/* Context: QEMU global mutex held */
 void virtio_blk_data_plane_create(VirtIODevice *vdev, VirtIOBlkConf *blk,
                                   VirtIOBlockDataPlane **dataplane,
                                   Error **errp)
@@ -408,12 +390,33 @@ void virtio_blk_data_plane_create(VirtIODevice *vdev, VirtIOBlkConf *blk,
     s->fd = fd;
     s->blk = blk;
 
+    if (blk->iothread) {
+        s->internal_iothread = false;
+        s->iothread = blk->iothread;
+        object_ref(OBJECT(s->iothread));
+    } else {
+        /* Create per-device IOThread if none specified */
+        Error *local_err = NULL;
+
+        s->internal_iothread = true;
+        object_add(TYPE_IOTHREAD, vdev->name, NULL, NULL, &local_err);
+        if (error_is_set(&local_err)) {
+            error_propagate(errp, local_err);
+            g_free(s);
+            return;
+        }
+        s->iothread = iothread_find(vdev->name);
+        assert(s->iothread);
+    }
+    s->ctx = iothread_get_aio_context(s->iothread);
+
     /* Prevent block operations that conflict with data plane thread */
     bdrv_set_in_use(blk->conf.bs, 1);
 
     *dataplane = s;
 }
 
+/* Context: QEMU global mutex held */
 void virtio_blk_data_plane_destroy(VirtIOBlockDataPlane *s)
 {
     if (!s) {
@@ -422,9 +425,14 @@ void virtio_blk_data_plane_destroy(VirtIOBlockDataPlane *s)
 
     virtio_blk_data_plane_stop(s);
     bdrv_set_in_use(s->blk->conf.bs, 0);
+    object_unref(OBJECT(s->iothread));
+    if (s->internal_iothread) {
+        object_unparent(OBJECT(s->iothread));
+    }
     g_free(s);
 }
 
+/* Context: QEMU global mutex held */
 void virtio_blk_data_plane_start(VirtIOBlockDataPlane *s)
 {
     BusState *qbus = BUS(qdev_get_parent_bus(DEVICE(s->vdev)));
@@ -448,8 +456,6 @@ void virtio_blk_data_plane_start(VirtIOBlockDataPlane *s)
         return;
     }
 
-    s->ctx = aio_context_new();
-
     /* Set up guest notifier (irq) */
     if (k->set_guest_notifiers(qbus->parent, 1, true) != 0) {
         fprintf(stderr, "virtio-blk failed to set guest notifier, "
@@ -464,7 +470,6 @@ void virtio_blk_data_plane_start(VirtIOBlockDataPlane *s)
         exit(1);
     }
     s->host_notifier = *virtio_queue_get_host_notifier(vq);
-    aio_set_event_notifier(s->ctx, &s->host_notifier, handle_notify);
 
     /* Set up ioqueue */
     ioq_init(&s->ioqueue, s->fd, REQ_MAX);
@@ -472,7 +477,6 @@ void virtio_blk_data_plane_start(VirtIOBlockDataPlane *s)
         ioq_put_iocb(&s->ioqueue, &s->requests[i].iocb);
     }
     s->io_notifier = *ioq_get_notifier(&s->ioqueue);
-    aio_set_event_notifier(s->ctx, &s->io_notifier, handle_io);
 
     s->starting = false;
     s->started = true;
@@ -481,11 +485,14 @@ void virtio_blk_data_plane_start(VirtIOBlockDataPlane *s)
     /* Kick right away to begin processing requests already in vring */
     event_notifier_set(virtio_queue_get_host_notifier(vq));
 
-    /* Spawn thread in BH so it inherits iothread cpusets */
-    s->start_bh = qemu_bh_new(start_data_plane_bh, s);
-    qemu_bh_schedule(s->start_bh);
+    /* Get this show started by hooking up our callbacks */
+    aio_context_acquire(s->ctx);
+    aio_set_event_notifier(s->ctx, &s->host_notifier, handle_notify);
+    aio_set_event_notifier(s->ctx, &s->io_notifier, handle_io);
+    aio_context_release(s->ctx);
 }
 
+/* Context: QEMU global mutex held */
 void virtio_blk_data_plane_stop(VirtIOBlockDataPlane *s)
 {
     BusState *qbus = BUS(qdev_get_parent_bus(DEVICE(s->vdev)));
@@ -496,27 +503,32 @@ void virtio_blk_data_plane_stop(VirtIOBlockDataPlane *s)
     s->stopping = true;
     trace_virtio_blk_data_plane_stop(s);
 
-    /* Stop thread or cancel pending thread creation BH */
-    if (s->start_bh) {
-        qemu_bh_delete(s->start_bh);
-        s->start_bh = NULL;
-    } else {
-        aio_notify(s->ctx);
-        qemu_thread_join(&s->thread);
+    aio_context_acquire(s->ctx);
+
+    /* Stop notifications for new requests from guest */
+    aio_set_event_notifier(s->ctx, &s->host_notifier, NULL);
+
+    /* Complete pending requests */
+    while (s->num_reqs > 0) {
+        aio_poll(s->ctx, true);
     }
 
+    /* Stop ioq callbacks (there are no pending requests left) */
     aio_set_event_notifier(s->ctx, &s->io_notifier, NULL);
-    ioq_cleanup(&s->ioqueue);
 
-    aio_set_event_notifier(s->ctx, &s->host_notifier, NULL);
-    k->set_host_notifier(qbus->parent, 0, false);
+    aio_context_release(s->ctx);
 
-    aio_context_unref(s->ctx);
+    /* Sync vring state back to virtqueue so that non-dataplane request
+     * processing can continue when we disable the host notifier below.
+     */
+    vring_teardown(&s->vring, s->vdev, 0);
+
+    ioq_cleanup(&s->ioqueue);
+    k->set_host_notifier(qbus->parent, 0, false);
 
     /* Clean up guest notifier (irq) */
     k->set_guest_notifiers(qbus->parent, 1, false);
 
-    vring_teardown(&s->vring, s->vdev, 0);
     s->started = false;
     s->stopping = false;
 }
diff --git a/include/hw/virtio/virtio-blk.h b/include/hw/virtio/virtio-blk.h
index 41885da..e4c41ff 100644
--- a/include/hw/virtio/virtio-blk.h
+++ b/include/hw/virtio/virtio-blk.h
@@ -16,6 +16,7 @@
 
 #include "hw/virtio/virtio.h"
 #include "hw/block/block.h"
+#include "sysemu/iothread.h"
 
 #define TYPE_VIRTIO_BLK "virtio-blk-device"
 #define VIRTIO_BLK(obj) \
@@ -106,6 +107,7 @@ struct virtio_scsi_inhdr
 struct VirtIOBlkConf
 {
     BlockConf conf;
+    IOThread *iothread;
     char *serial;
     uint32_t scsi;
     uint32_t config_wce;
@@ -140,13 +142,15 @@ typedef struct VirtIOBlock {
         DEFINE_BLOCK_CHS_PROPERTIES(_state, _field.conf),                     \
         DEFINE_PROP_STRING("serial", _state, _field.serial),                  \
         DEFINE_PROP_BIT("config-wce", _state, _field.config_wce, 0, true),    \
-        DEFINE_PROP_BIT("scsi", _state, _field.scsi, 0, true)
+        DEFINE_PROP_BIT("scsi", _state, _field.scsi, 0, true),                \
+        DEFINE_PROP_IOTHREAD("x-iothread", _state, _field.iothread)
 #else
 #define DEFINE_VIRTIO_BLK_PROPERTIES(_state, _field)                          \
         DEFINE_BLOCK_PROPERTIES(_state, _field.conf),                         \
         DEFINE_BLOCK_CHS_PROPERTIES(_state, _field.conf),                     \
         DEFINE_PROP_STRING("serial", _state, _field.serial),                  \
-        DEFINE_PROP_BIT("config-wce", _state, _field.config_wce, 0, true)
+        DEFINE_PROP_BIT("config-wce", _state, _field.config_wce, 0, true),    \
+        DEFINE_PROP_IOTHREAD("x-iothread", _state, _field.iothread)
 #endif /* __linux__ */
 
 void virtio_blk_set_conf(DeviceState *dev, VirtIOBlkConf *blk);
-- 
1.8.5.3