[Xenomai-core] [PATCH 1/2] nucleus: Add semaphore heap statistics

[Xenomai-core] [PATCH 1/2] nucleus: Add semaphore heap statistics

[Jan Kiszka]

This extends /proc/xenomai/heap with statistics about the global as well
as all per-process semaphore heaps. This is helpful to track down the
reason for ENOMEM (system or sem heap full?) and to find out that we
are leaking memory from the global heap on automatic native mutex,
queue, and heap deletion.

Signed-off-by: Jan Kiszka <jan.kiszka@domain.hid>
---

 include/nucleus/sys_ppd.h |   11 +++++++++++
 ksrc/nucleus/heap.c       |   30 ++++++++++++++++++++++++++++++
 ksrc/nucleus/shadow.c     |   23 +++++++++++++++++++++++
 3 files changed, 64 insertions(+), 0 deletions(-)

diff --git a/include/nucleus/sys_ppd.h b/include/nucleus/sys_ppd.h
index 8f8cb44..015af53 100644
--- a/include/nucleus/sys_ppd.h
+++ b/include/nucleus/sys_ppd.h
@@ -3,11 +3,17 @@
 
 #include <nucleus/ppd.h>
 #include <nucleus/heap.h>
+#include <nucleus/queue.h>
 
 struct xnsys_ppd {
 	xnshadow_ppd_t ppd;
 	xnheap_t sem_heap;
 
+#ifdef CONFIG_XENO_OPT_PERVASIVE
+	xnholder_t link;
+	pid_t pid;
+#endif
+
 #define ppd2sys(addr) container_of(addr, struct xnsys_ppd, ppd)
 };
 
@@ -15,6 +21,11 @@ extern struct xnsys_ppd __xnsys_global_ppd;
 
 #ifdef CONFIG_XENO_OPT_PERVASIVE
 
+#include <linux/spinlock.h>
+
+extern xnqueue_t xnsys_ppds;
+extern spinlock_t xnsys_ppd_lock;
+
 static inline struct xnsys_ppd *xnsys_ppd_get(int global)
 {
 	xnshadow_ppd_t *ppd;
diff --git a/ksrc/nucleus/heap.c b/ksrc/nucleus/heap.c
index 9ca2591..4a3abd0 100644
--- a/ksrc/nucleus/heap.c
+++ b/ksrc/nucleus/heap.c
@@ -67,6 +67,7 @@ HEAP {
 #include <nucleus/thread.h>
 #include <nucleus/heap.h>
 #include <nucleus/assert.h>
+#include <nucleus/sys_ppd.h>
 #include <asm/xenomai/bits/heap.h>
 
 xnheap_t kheap;		/* System heap */
@@ -1277,11 +1278,17 @@ int xnheap_destroy_mapped(xnheap_t *heap, void (*release)(struct xnheap *heap),
 #ifdef CONFIG_PROC_FS
 
 #include <linux/proc_fs.h>
+#include <nucleus/sys_ppd.h>
 
 static int heap_read_proc(char *page,
 			  char **start,
 			  off_t off, int count, int *eof, void *data)
 {
+#ifdef CONFIG_XENO_OPT_PERVASIVE
+	xnholder_t *entry;
+	struct xnsys_ppd *sys_ppd;
+#endif
+	xnheap_t *sem_heap;
 	int len;
 
 	if (!xnpod_active_p())
@@ -1298,6 +1305,29 @@ static int heap_read_proc(char *page,
 		       xnheap_used_mem(&kstacks),
 		       xnheap_page_size(&kstacks));
 #endif
+	sem_heap = &__xnsys_global_ppd.sem_heap;
+	if (sem_heap)
+		len += sprintf(page + len, "size=%lu:used=%lu:pagesz=%lu  "
+					   "(global sem heap)\n",
+			       xnheap_usable_mem(sem_heap),
+			       xnheap_used_mem(sem_heap),
+			       xnheap_page_size(sem_heap));
+#ifdef CONFIG_XENO_OPT_PERVASIVE
+	spin_lock(&xnsys_ppd_lock);
+	entry = getheadq(&xnsys_ppds);
+	while (entry) {
+		sys_ppd = container_of(entry, struct xnsys_ppd, link);
+		sem_heap = &sys_ppd->sem_heap;
+		len += sprintf(page + len, "size=%lu:used=%lu:pagesz=%lu  "
+					   "(private sem heap [%d])\n",
+			       xnheap_usable_mem(sem_heap),
+			       xnheap_used_mem(sem_heap),
+			       xnheap_page_size(sem_heap),
+			       sys_ppd->pid);
+		entry = nextq(&xnsys_ppds, entry);
+	}
+	spin_unlock(&xnsys_ppd_lock);
+#endif
 
 	len -= off;
 	if (len <= off + count)
diff --git a/ksrc/nucleus/shadow.c b/ksrc/nucleus/shadow.c
index d6d1203..ec69438 100644
--- a/ksrc/nucleus/shadow.c
+++ b/ksrc/nucleus/shadow.c
@@ -515,6 +515,9 @@ void xnshadow_rpi_check(void)
 
 #endif /* !CONFIG_XENO_OPT_RPIDISABLE */
 
+DEFINE_XNQUEUE(xnsys_ppds);
+DEFINE_SPINLOCK(xnsys_ppd_lock);
+
 static xnqueue_t *ppd_hash;
 #define PPD_HASH_SIZE 13
 
@@ -652,6 +655,20 @@ static inline void ppd_remove_mm(struct mm_struct *mm,
 	xnlock_put_irqrestore(&nklock, s);
 }
 
+static void xnsys_ppd_register(struct xnsys_ppd *sys_ppd)
+{
+	spin_lock(&xnsys_ppd_lock);
+	appendq(&xnsys_ppds, &sys_ppd->link);
+	spin_unlock(&xnsys_ppd_lock);
+}
+
+static void xnsys_ppd_unregister(struct xnsys_ppd *sys_ppd)
+{
+	spin_lock(&xnsys_ppd_lock);
+	removeq(&xnsys_ppds, &sys_ppd->link);
+	spin_unlock(&xnsys_ppd_lock);
+}
+
 static inline void request_syscall_restart(xnthread_t *thread,
 					   struct pt_regs *regs,
 					   int sysflags)
@@ -1892,10 +1909,16 @@ static void *xnshadow_sys_event(int event, void *data)
 			return ERR_PTR(err);
 		}
 
+		p->pid = current->pid;
+		xnsys_ppd_register(p);
+
 		return &p->ppd;
 
 	case XNSHADOW_CLIENT_DETACH:
 		p = ppd2sys((xnshadow_ppd_t *) data);
+
+		xnsys_ppd_unregister(p);
+
 		xnheap_destroy_mapped(&p->sem_heap, post_ppd_release, NULL);
 
 		return NULL;

[Xenomai-core] [PATCH 2/2] native: Release fastlock to the proper heap

[Jan Kiszka]

Don't assume rt_task_delete is only called for in-kernel users, it may
be triggered via auto-cleanup also on user space objects. So check for
the creator and release the fastlock to the correct heap.

Signed-off-by: Jan Kiszka <jan.kiszka@domain.hid>
---

 ksrc/skins/native/mutex.c |   14 ++++++++++++--
 1 files changed, 12 insertions(+), 2 deletions(-)

diff --git a/ksrc/skins/native/mutex.c b/ksrc/skins/native/mutex.c
index 20eb484..6cf7eb1 100644
--- a/ksrc/skins/native/mutex.c
+++ b/ksrc/skins/native/mutex.c
@@ -47,6 +47,7 @@
 #include <nucleus/pod.h>
 #include <nucleus/registry.h>
 #include <nucleus/heap.h>
+#include <nucleus/sys_ppd.h>
 #include <native/task.h>
 #include <native/mutex.h>
 
@@ -316,8 +317,17 @@ int rt_mutex_delete(RT_MUTEX *mutex)
 	err = rt_mutex_delete_inner(mutex);
 
 #ifdef CONFIG_XENO_FASTSYNCH
-	if (!err)
-		xnfree(mutex->synch_base.fastlock);
+	if (!err) {
+#ifdef CONFIG_XENO_OPT_PERVASIVE
+		if (mutex->cpid) {
+			int global = xnsynch_test_flags(&mutex->synch_base,
+							RT_MUTEX_EXPORTED);
+			xnheap_free(&xnsys_ppd_get(global)->sem_heap,
+				    mutex->synch_base.fastlock);
+		} else
+#endif /* CONFIG_XENO_OPT_PERVASIVE */
+			xnfree(mutex->synch_base.fastlock);
+	}
 #endif /* CONFIG_XENO_FASTSYNCH */
 
 	return err;

[Xenomai-core] [PATCH 0/2] Sem heap statistics & mutex auto-cleanup fixes

[Jan Kiszka]

This series fixes the native mutex auto-cleanup issues I reported in
[1]. It does not yet address the issues around heaps and queues as I
stumbled over more cleanup-related races, also in the nucleus, and want
to go through this with more care. Expect further patches to follow.

Also included in this pull request is a statistic enhancement for
/proc/xenomai/heap: It was lacking the sem heaps so far.

Please pull the series from

    git://xenomai.org/xenomai-jki.git for-upstream

if there are no concerns.

Jan Kiszka (2):
      nucleus: Add semaphore heap statistics
      native: Release fastlock to the proper heap

 include/nucleus/sys_ppd.h |   11 +++++++++++
 ksrc/nucleus/heap.c       |   30 ++++++++++++++++++++++++++++++
 ksrc/nucleus/shadow.c     |   23 +++++++++++++++++++++++
 ksrc/skins/native/mutex.c |   14 ++++++++++++--
 4 files changed, 76 insertions(+), 2 deletions(-)

[1] http://thread.gmane.org/gmane.linux.real-time.xenomai.devel/6559

Re: [Xenomai-core] [PATCH 1/2] nucleus: Add semaphore heap statistics

[Gilles Chanteperdrix]

Jan Kiszka wrote:
> This extends /proc/xenomai/heap with statistics about the global as well
> as all per-process semaphore heaps. This is helpful to track down the
> reason for ENOMEM (system or sem heap full?) and to find out that we
> are leaking memory from the global heap on automatic native mutex,
> queue, and heap deletion.

I'd rather see the xnholder_t in the xnheap structure. This way, we
would be able to add all heaps to /proc/xenomai/heaps.

-- 
                                          Gilles

Re: [Xenomai-core] [PATCH 1/2] nucleus: Add semaphore heap statistics

[Jan Kiszka]

Gilles Chanteperdrix wrote:
> Jan Kiszka wrote:
>> This extends /proc/xenomai/heap with statistics about the global as well
>> as all per-process semaphore heaps. This is helpful to track down the
>> reason for ENOMEM (system or sem heap full?) and to find out that we
>> are leaking memory from the global heap on automatic native mutex,
>> queue, and heap deletion.
> 
> I'd rather see the xnholder_t in the xnheap structure. This way, we
> would be able to add all heaps to /proc/xenomai/heaps.
> 

Makes sense, will rework this.

Jan

-- 
Siemens AG, Corporate Technology, CT SE 2
Corporate Competence Center Embedded Linux