static buffer for timer-bheap
PATCH 1/1 static buffer for timer-bheap
Hi,
this patch moves the heap for scalable timer queues into the xnpod_t structure instead of allocating it dynamically. This simplifies the pod initialisation and cleanup, which was not fully robust in this regard anyway. If a pod is now considered too large (but we are discussion kbytes here), allocating its buffer via xnarch_sysalloc would be an option.
In case this patch is acceptable, I would suggest merging it before 2.2 due to the contained (minor) fix.
Jan
nucleus/pod.c +3 -25
@@ -433,22 +433,11 @@ int xnpod_init(xnpod_t *pod, int minpri,#ifdef CONFIG_XENO_OPT_TIMING_PERIODIC unsigned n; -for (n =0 ; n <XNTIMER_WHEELSIZE ; n++) { - err = -xntlist_init (&pod->sched[cpu].timerwheel[n]); - -if (err) { -xnheap_destroy (&kheap, &xnpod_flush_heap,NULL ); - goto fail; - } - } +for (n =0 ; n <XNTIMER_WHEELSIZE ; n++) +xntlist_init (&pod->sched[cpu].timerwheel[n]);#endif /* CONFIG_XENO_OPT_TIMING_PERIODIC */ - err = -xntimerq_init (&pod->sched[cpu].timerqueue); - -if (err) { -xnheap_destroy (&kheap, &xnpod_flush_heap,NULL ); - goto fail; - } +xntimerq_init (&pod->sched[cpu].timerqueue); }for (cpu =0 ; cpu < nr_cpus; ++cpu) {
@@ -547,7 +536,6 @@ void xnpod_shutdown(int xtype) {xnholder_t *holder, *nholder;xnthread_t *thread; -unsigned cpu;spl_t s;xnlock_get_irqsave (&nklock, s);
@@ -585,16 +573,6 @@ void xnpod_shutdown(int xtype)__setbits (nkpod->status,XNPIDLE ); -for (cpu =0 ; cpu <xnarch_num_online_cpus (); cpu++) { -#ifdef CONFIG_XENO_OPT_TIMING_PERIODIC -unsigned n; - -for (n =0 ; n <XNTIMER_WHEELSIZE ; n++) -xntlist_destroy (&nkpod->sched[cpu].timerwheel[n]); -#endif /* CONFIG_XENO_OPT_TIMING_PERIODIC */ -xntimerq_destroy (&nkpod->sched[cpu].timerqueue); - } -xnlock_put_irqrestore (&nklock, s);#ifdef CONFIG_XENO_OPT_REGISTRY
nucleus/bheap.h +12 -22
@@ -43,9 +43,15 @@ typedef struct bheaph {typedef struct bheap {unsigned sz;unsigned last; -bheaph_t **elems; +bheaph_t *elems[1 ];/* only padding, indexing starts at 1 */ }bheap_t ; +#define DECLARE_BHEAP_CONTAINER (name, sz) \ + struct { \ + bheap_t bheap; \ + bheaph_t elems[sz]; \ + } name +static inline bheaph_t *bheap_gethead (bheap_t *heap) {if (heap->last ==1 )
@@ -68,26 +74,10 @@ static inline bheaph_t *bheaph_child(bhereturn likely (pos < heap->last) ? heap->elems[pos] :NULL ; } -static inline int bheap_init (bheap_t *heap,unsigned sz) +static inline void bheap_init (bheap_t *heap,unsigned sz) { heap->sz = sz; heap->last =1 ; - heap->elems = (bheaph_t **)xnarch_sysalloc (sz *sizeof (void *)); - -if (!heap->elems) -return ENOMEM ; - -/* start indexing at 1. */ - heap->elems -=1 ; - -return 0 ; -} - -static inline void bheap_destroy (bheap_t *heap) -{ -xnarch_sysfree (heap->elems +1 , heap->sz *sizeof (void *)); - heap->last =0 ; - heap->sz =0 ; }static inline void bheap_swap (bheap_t *heap,bheaph_t *h1,bheaph_t *h2)
@@ -115,7 +105,7 @@ static inline void bheap_down(bheap_t *hfor (;;) { left =bheaph_child (heap, holder,0 ); right =bheaph_child (heap, holder,1 ); - +if (left && right) minchild =bheaph_lt (left, right) ? left : right;else
@@ -143,10 +133,10 @@ static inline int bheap_insert(bheap_t *static inline int bheap_delete (bheap_t *heap,bheaph_t *holder) {bheaph_t *lasth; - +if (heap->last ==1 )return EINVAL ; - + --heap->last;if (heap->last >1 ) { lasth = heap->elems[heap->last];
@@ -154,7 +144,7 @@ static inline int bheap_delete(bheap_t *bheaph_pos (lasth) =bheaph_pos (holder);bheap_down (heap, lasth); } - +return 0 ; }
nucleus/timer.h +14 -13
@@ -58,11 +58,10 @@ typedef struct { ((xntlholder_t *)(((char *)laddr) - offsetof(xntlholder_t, link))) } xntlholder_t; -#define xntlholder_date (h) ((h)->key) -#define xntlholder_prio (h) ((h)->prio) -#define xntlholder_init (h) inith(&(h)->link) -#define xntlist_init (q) (initq(q),0) -#define xntlist_destroy (q) do { } while (0) +#define xntlholder_date (h) ((h)->key) +#define xntlholder_prio (h) ((h)->prio) +#define xntlholder_init (h) inith(&(h)->link) +#define xntlist_init (q) initq(q)#define xntlist_head (q) \ ({ xnholder_t *_h = getheadq(q); \ !_h ? NULL : link2tlholder(_h); \
@@ -71,7 +70,7 @@ typedef struct {static inline void xntlist_insert (xnqueue_t *q,xntlholder_t *holder) {xnholder_t *p; - +/* Insert the new timer at the proper place in the single queue managed when running in aperiodic mode. O(N) here, but users of the aperiodic mode need to pay a price for
@@ -82,7 +81,7 @@ static inline void xntlist_insert(xnqueu (holder->key ==link2tlholder (p)->key && holder->prio <= link2tlholder(p)->prio)) break; - +insertq (q,p->next,&holder->link); }
@@ -94,12 +93,14 @@ typedef bheaph_t xntimerh_t;#define xntimerh_date (h) bheaph_key(h)#define xntimerh_prio (h) bheaph_prio(h)#define xntimerh_init (h) bheaph_init(h) -typedef bheap_t xntimerq_t ; -#define xntimerq_init (q) bheap_init((q), CONFIG_XENO_OPT_TIMER_HEAP_CAPACITY) -#define xntimerq_destroy (q) bheap_destroy(q) -#define xntimerq_head (q) bheap_gethead(q) -#define xntimerq_insert (q, h) bheap_insert((q),(h)) -#define xntimerq_remove (q, h) bheap_delete((q),(h)) +typedef struct { +bheap_t bheap; +bheaph_t __buffer[CONFIG_XENO_OPT_TIMER_HEAP_CAPACITY ]; +}xntimerq_t ; +#define xntimerq_init (q) bheap_init(&(q)->bheap, CONFIG_XENO_OPT_TIMER_HEAP_CAPACITY) +#define xntimerq_head (q) bheap_gethead(&(q)->bheap) +#define xntimerq_insert (q, h) bheap_insert(&(q)->bheap,(h)) +#define xntimerq_remove (q, h) bheap_delete(&(q)->bheap,(h))#else /* CONFIG_XENO_OPT_TIMER_LIST */ typedef xntlholder_t xntimerh_t ;
ChangeLog +7 -1
@@ -1,3 +1,9 @@ +2006-06-01 Jan Kiszka <[email protected]> + + * include/nucleus/{bheap.h,timer.h}, ksrc/nucleus/pod.c: + Move timer heap into xnpod_t structure, simplify pod init and + cleanup. + 2006-05-23 Gilles Chanteperdrix <[email protected]> * ksrc/arch/i386/nmi.c: Fix alignement for gcc-4.1.
@@ -14,7 +20,7 @@ * ksrc/arch/arm/patches: Upgrade to 2.6.1{4,5}-1.3-04. -2006-05-19 Jan Kiszka <[email protected]> +2006-05-19 Jan Kiszka <[email protected]> * src/testsuite/latency/latency.c: Add pid to registered names allowing multiple instances of latency to run. Add -c switch to
Quoted 10 lines · Jan Kiszka
Hi,
this patch moves the heap for scalable timer queues into the
xnpod_tstructure instead of allocating it dynamically. This simplifies the pod initialisation and cleanup, which was not fully robust in this regard anyway. If a pod is now considered too large (but we are discussion kbytes here), allocating its buffer viaxnarch_sysallocwould be an option.In case this patch is acceptable, I would suggest merging it before 2.2 due to the contained (minor) fix.
Since this memory is specific to the data structure, its allocation seems better by the data structure functions. It also make the binary heap structure reusable for other purposes.
Could you explain why the current scheme is not robust ?
Signature · 1 line
Gilles Chanteperdrix.
Quoted 9 lines · Gilles Chanteperdrix
Quoted 8 lines · Jan Kiszka
Hi,
this patch moves the heap for scalable timer queues into the
xnpod_tstructure instead of allocating it dynamically. This simplifies the pod initialisation and cleanup, which was not fully robust in this regard anyway. If a pod is now considered too large (but we are discussion kbytes here), allocating its buffer viaxnarch_sysallocwould be an option.
Quoted 6 lines
Quoted 2 lines
In case this patch is acceptable, I would suggest merging it before 2.2 due to the contained (minor) fix.
Since this memory is specific to the data structure, its allocation seems better by the data structure functions. It also make the binary heap structure reusable for other purposes.
That's what I added the DECLARE_BHEAP_CONTAINER macro for. Sorry, did not mention this.
The idea is to leave the allocation policy up to the user instead of enforcing it in the bheap layer. The current approach also implicitly excludes its usage from RT contexts, BTW.
Quoted 3 lines
Could you explain why the current scheme is not robust ?
Given a SMP system where the allocation fails right in the middle of the per-CPU loop [1], cleanup will not be performed for already allocated heaps.
Jan
[1]http://www.rts.uni-hannover.de/xenomai/lxr/source/ksrc/nucleus/pod.c?v=SVN-trunk#L432
Quoted 25 lines · Jan Kiszka
Quoted 9 lines · Gilles Chanteperdrix
Quoted 8 lines · Jan Kiszka
Hi,
this patch moves the heap for scalable timer queues into the
xnpod_tstructure instead of allocating it dynamically. This simplifies the pod initialisation and cleanup, which was not fully robust in this regard anyway. If a pod is now considered too large (but we are discussion kbytes here), allocating its buffer viaxnarch_sysallocwould be an option.Quoted 6 lines
Quoted 2 lines
In case this patch is acceptable, I would suggest merging it before 2.2 due to the contained (minor) fix.
Since this memory is specific to the data structure, its allocation seems better by the data structure functions. It also make the binary heap structure reusable for other purposes.
That's what I added the
DECLARE_BHEAP_CONTAINERmacro for. Sorry, did not mention this.The idea is to leave the allocation policy up to the user instead of enforcing it in the bheap layer. The current approach also implicitly excludes its usage from RT contexts, BTW.
Quoted 1 line
Quoted 5 lines
Quoted 2 lines
Could you explain why the current scheme is not robust ?
Given a SMP system where the allocation fails right in the middle of the per-CPU loop [1], cleanup will not be performed for already allocated heaps.
You are right.
Signature · 1 line
Gilles Chanteperdrix.
nucleus/bheap.h typedef struct bheaph { · line 54
Quoted 7 lines · Jan Kiszka
(...)
nucleus/bheap.h +6
-0+#define DECLARE_BHEAP_CONTAINER (name, sz) \ + struct { \ + bheap_t bheap; \ + bheaph_t elems[sz]; \ + } name +
Don't you mean bheaph_t *elems[sz] ?
nucleus/timer.h typedef bheaph_t xntimerh_t; · line 99
Quoted 5 lines
(...)
nucleus/timer.h +4
-0+typedef struct { +bheap_t bheap; +bheaph_t __buffer[CONFIG_XENO_OPT_TIMER_HEAP_CAPACITY ]; +}xntimerq_t ;
Same remark. And why not using the macro here ?
I would also prefer passing the bheaph_t** storage to bheap_init, and conserve bheap_destroy (with a callback called with the bheap_t** storage) in case the storage was dynamically allocated by the caller.
Signature · 1 line
Gilles Chanteperdrix.
Quoted 10 lines · Gilles Chanteperdrix
Review
Quoted 7 lines · Jan Kiszka
(...)
(quoted diff) +6
-0+#define DECLARE_BHEAP_CONTAINER(name, sz) \ + struct { \ + bheap_t bheap; \ + bheaph_t elems[sz]; \ + } name +Don't you mean
bheaph_t*elems[sz] ?
Yep.
Quoted 8 lines
Review
Quoted 6 lines
(...)
(quoted diff) +4
-0+typedef struct { + bheap_t bheap; + bheaph_t __buffer[CONFIG_XENO_OPT_TIMER_HEAP_CAPACITY]; +} xntimerq_t;Same remark. And why not using the macro here ?
Indeed. Ugly late-night hack, improved version attached.
Quoted 4 lines
I would also prefer passing the bheaph_t** storage to
bheap_init, and conservebheap_destroy(with a callback called with the bheap_t** storage) in case the storage was dynamically allocated by the caller.
Do you have a concrete usage scenario in mind where this would be required? I would rather bet that potential callers of bheap_destroy know very well when some buffer is to be released. Looks at bit like overkill unless someone has the real need to mix dynamically with statically allocated bheaps.
Jan
nucleus/pod.c +3 -25
@@ -433,22 +433,11 @@ int xnpod_init(xnpod_t *pod, int minpri,#ifdef CONFIG_XENO_OPT_TIMING_PERIODIC unsigned n; -for (n =0 ; n <XNTIMER_WHEELSIZE ; n++) { - err = -xntlist_init (&pod->sched[cpu].timerwheel[n]); - -if (err) { -xnheap_destroy (&kheap, &xnpod_flush_heap,NULL ); - goto fail; - } - } +for (n =0 ; n <XNTIMER_WHEELSIZE ; n++) +xntlist_init (&pod->sched[cpu].timerwheel[n]);#endif /* CONFIG_XENO_OPT_TIMING_PERIODIC */ - err = -xntimerq_init (&pod->sched[cpu].timerqueue); - -if (err) { -xnheap_destroy (&kheap, &xnpod_flush_heap,NULL ); - goto fail; - } +xntimerq_init (&pod->sched[cpu].timerqueue); }for (cpu =0 ; cpu < nr_cpus; ++cpu) {
@@ -547,7 +536,6 @@ void xnpod_shutdown(int xtype) {xnholder_t *holder, *nholder;xnthread_t *thread; -unsigned cpu;spl_t s;xnlock_get_irqsave (&nklock, s);
@@ -585,16 +573,6 @@ void xnpod_shutdown(int xtype)__setbits (nkpod->status,XNPIDLE ); -for (cpu =0 ; cpu <xnarch_num_online_cpus (); cpu++) { -#ifdef CONFIG_XENO_OPT_TIMING_PERIODIC -unsigned n; - -for (n =0 ; n <XNTIMER_WHEELSIZE ; n++) -xntlist_destroy (&nkpod->sched[cpu].timerwheel[n]); -#endif /* CONFIG_XENO_OPT_TIMING_PERIODIC */ -xntimerq_destroy (&nkpod->sched[cpu].timerqueue); - } -xnlock_put_irqrestore (&nklock, s);#ifdef CONFIG_XENO_OPT_REGISTRY
nucleus/bheap.h +12 -22
@@ -43,9 +43,15 @@ typedef struct bheaph {typedef struct bheap {unsigned sz;unsigned last; -bheaph_t **elems; +bheaph_t *elems[1 ];/* only padding, indexing starts at 1 */ }bheap_t ; +#define DECLARE_BHEAP_CONTAINER (name, sz) \ + struct { \ + bheap_t bheap; \ + bheaph_t *elems[sz]; \ + } name +static inline bheaph_t *bheap_gethead (bheap_t *heap) {if (heap->last ==1 )
@@ -68,26 +74,10 @@ static inline bheaph_t *bheaph_child(bhereturn likely (pos < heap->last) ? heap->elems[pos] :NULL ; } -static inline int bheap_init (bheap_t *heap,unsigned sz) +static inline void bheap_init (bheap_t *heap,unsigned sz) { heap->sz = sz; heap->last =1 ; - heap->elems = (bheaph_t **)xnarch_sysalloc (sz *sizeof (void *)); - -if (!heap->elems) -return ENOMEM ; - -/* start indexing at 1. */ - heap->elems -=1 ; - -return 0 ; -} - -static inline void bheap_destroy (bheap_t *heap) -{ -xnarch_sysfree (heap->elems +1 , heap->sz *sizeof (void *)); - heap->last =0 ; - heap->sz =0 ; }static inline void bheap_swap (bheap_t *heap,bheaph_t *h1,bheaph_t *h2)
@@ -115,7 +105,7 @@ static inline void bheap_down(bheap_t *hfor (;;) { left =bheaph_child (heap, holder,0 ); right =bheaph_child (heap, holder,1 ); - +if (left && right) minchild =bheaph_lt (left, right) ? left : right;else
@@ -143,10 +133,10 @@ static inline int bheap_insert(bheap_t *static inline int bheap_delete (bheap_t *heap,bheaph_t *holder) {bheaph_t *lasth; - +if (heap->last ==1 )return EINVAL ; - + --heap->last;if (heap->last >1 ) { lasth = heap->elems[heap->last];
@@ -154,7 +144,7 @@ static inline int bheap_delete(bheap_t *bheaph_pos (lasth) =bheaph_pos (holder);bheap_down (heap, lasth); } - +return 0 ; }
nucleus/timer.h +11 -13
@@ -58,11 +58,10 @@ typedef struct { ((xntlholder_t *)(((char *)laddr) - offsetof(xntlholder_t, link))) } xntlholder_t; -#define xntlholder_date (h) ((h)->key) -#define xntlholder_prio (h) ((h)->prio) -#define xntlholder_init (h) inith(&(h)->link) -#define xntlist_init (q) (initq(q),0) -#define xntlist_destroy (q) do { } while (0) +#define xntlholder_date (h) ((h)->key) +#define xntlholder_prio (h) ((h)->prio) +#define xntlholder_init (h) inith(&(h)->link) +#define xntlist_init (q) initq(q)#define xntlist_head (q) \ ({ xnholder_t *_h = getheadq(q); \ !_h ? NULL : link2tlholder(_h); \
@@ -71,7 +70,7 @@ typedef struct {static inline void xntlist_insert (xnqueue_t *q,xntlholder_t *holder) {xnholder_t *p; - +/* Insert the new timer at the proper place in the single queue managed when running in aperiodic mode. O(N) here, but users of the aperiodic mode need to pay a price for
@@ -82,7 +81,7 @@ static inline void xntlist_insert(xnqueu (holder->key ==link2tlholder (p)->key && holder->prio <= link2tlholder(p)->prio)) break; - +insertq (q,p->next,&holder->link); }
@@ -94,12 +93,11 @@ typedef bheaph_t xntimerh_t;#define xntimerh_date (h) bheaph_key(h)#define xntimerh_prio (h) bheaph_prio(h)#define xntimerh_init (h) bheaph_init(h) -typedef bheap_t xntimerq_t ; -#define xntimerq_init (q) bheap_init((q), CONFIG_XENO_OPT_TIMER_HEAP_CAPACITY) -#define xntimerq_destroy (q) bheap_destroy(q) -#define xntimerq_head (q) bheap_gethead(q) -#define xntimerq_insert (q, h) bheap_insert((q),(h)) -#define xntimerq_remove (q, h) bheap_delete((q),(h)) +typedef DECLARE_BHEAP_CONTAINER(xntimerq_t, CONFIG_XENO_OPT_TIMER_HEAP_CAPACITY); +#define xntimerq_init (q) bheap_init(&(q)->bheap, CONFIG_XENO_OPT_TIMER_HEAP_CAPACITY) +#define xntimerq_head (q) bheap_gethead(&(q)->bheap) +#define xntimerq_insert (q, h) bheap_insert(&(q)->bheap,(h)) +#define xntimerq_remove (q, h) bheap_delete(&(q)->bheap,(h))#else /* CONFIG_XENO_OPT_TIMER_LIST */ typedef xntlholder_t xntimerh_t ;
ChangeLog +7 -1
@@ -1,3 +1,9 @@ +2006-06-02 Jan Kiszka <[email protected]> + + * include/nucleus/{bheap.h,timer.h}, ksrc/nucleus/pod.c: + Move timer heap into xnpod_t structure, simplify pod init and + cleanup. + 2006-05-23 Gilles Chanteperdrix <[email protected]> * ksrc/arch/i386/nmi.c: Fix alignement for gcc-4.1.
@@ -14,7 +20,7 @@ * ksrc/arch/arm/patches: Upgrade to 2.6.1{4,5}-1.3-04. -2006-05-19 Jan Kiszka <[email protected]> +2006-05-19 Jan Kiszka <[email protected]> * src/testsuite/latency/latency.c: Add pid to registered names allowing multiple instances of latency to run. Add -c switch to
Quoted 10 lines · Jan Kiszka
Quoted 3 lines · Gilles Chanteperdrix
I would also prefer passing the bheaph_t** storage to
bheap_init, and conservebheap_destroy(with a callback called with the bheap_t** storage) in case the storage was dynamically allocated by the caller.Do you have a concrete usage scenario in mind where this would be required? I would rather bet that potential callers of
bheap_destroyknow very well when some buffer is to be released. Looks at bit like overkill unless someone has the real need to mix dynamically with statically allocated bheaps.
Requesting a bheaph_t ** to be passed to bheap_init is type-safe and would have caught the kind of mistake you have done. bheap_destroy allow setting the bheap_t structure to an invalid value which, in turn, allow helping upper layers in catching invalid uses of the bheap after its destruction.
It is a low price to pay to make the interface a bit safer.
Signature · 1 line
Gilles Chanteperdrix.
Quoted 6 lines · Gilles Chanteperdrix
Quoted 5 lines · Jan Kiszka
Quoted 3 lines · Gilles Chanteperdrix
I would also prefer passing the bheaph_t** storage to
bheap_init, and conservebheap_destroy(with a callback called with the bheap_t** storage) in case the storage was dynamically allocated by the caller.
Quoted 8 lines
Quoted 5 lines
Do you have a concrete usage scenario in mind where this would be required? I would rather bet that potential callers of
bheap_destroyknow very well when some buffer is to be released. Looks at bit like overkill unless someone has the real need to mix dynamically with statically allocated bheaps.Requesting a
bheaph_t** to be passed tobheap_initis type-safe and would have caught the kind of mistake you have done.
But I also did not use the macro, which avoids such mistakes now.
Quoted 4 lines
bheap_destroyallow
setting thebheap_tstructure to an invalid value which, in turn, allow helping upper layers in catching invalid uses of the bheap after its destruction.
Ok, we could make bheap_destroy pop up under XENO_OPT_DEBUG. We may then also add assertions to the bheap functions themselves to check for invalid usage.
Quoted 3 lines
It is a low price to pay to make the interface a bit safer.
The price is also larger code in every heap runtime function due to the additional indirection. I still vote for the combined structure and to give its allocation completely in user hand.
Jan
Quoted 9 lines · Jan Kiszka
Quoted 4 lines · Gilles Chanteperdrix
bheap_destroyallow
setting thebheap_tstructure to an invalid value which, in turn, allow helping upper layers in catching invalid uses of the bheap after its destruction.Ok, we could make
bheap_destroypop up underXENO_OPT_DEBUG. We may then also add assertions to the bheap functions themselves to check for invalid usage.
The checkings are already there, if you do not remove bheap_destroy. They are unconditionnal, because I find such checkings way too important to be optimized out.
Signature · 1 line
Gilles Chanteperdrix.
Quoted 7 lines · Gilles Chanteperdrix
Quoted 6 lines · Jan Kiszka
Quoted 4 lines · Gilles Chanteperdrix
bheap_destroyallow
setting thebheap_tstructure to an invalid value which, in turn, allow helping upper layers in catching invalid uses of the bheap after its destruction.
Quoted 7 lines
Quoted 3 lines
Ok, we could make
bheap_destroypop up underXENO_OPT_DEBUG. We may then also add assertions to the bheap functions themselves to check for invalid usage.The checkings are already there, if you do not remove
bheap_destroy. They are unconditionnal, because I find such checkings way too important to be optimized out.
Sorry, I might be blind, but how do bheap_gethead, bheap_insert, or bheap_delete detect that heap->last or heap->sz became 0 after bheap_destroy?
Given that this is code to be inlined and heavily used, we should really take care for size and efficiency on production systems, even if it's about a few bytes here. But I also agree that verbose(!) checking (XENO_ASSERT/XENO_BUGON) is very useful during development.
Jan
Quoted 16 lines · Jan Kiszka
Quoted 7 lines · Gilles Chanteperdrix
Quoted 6 lines · Jan Kiszka
Quoted 4 lines · Gilles Chanteperdrix
bheap_destroyallow
setting thebheap_tstructure to an invalid value which, in turn, allow helping upper layers in catching invalid uses of the bheap after its destruction.Quoted 7 lines
Quoted 3 lines
Ok, we could make
bheap_destroypop up underXENO_OPT_DEBUG. We may then also add assertions to the bheap functions themselves to check for invalid usage.The checkings are already there, if you do not remove
bheap_destroy. They are unconditionnal, because I find such checkings way too important to be optimized out.
Quoted 8 lines
Sorry, I might be blind, but how do
bheap_gethead,bheap_insert, orbheap_deletedetect thatheap->lastorheap->szbecame 0 afterbheap_destroy?Given that this is code to be inlined and heavily used, we should really take care for size and efficiency on production systems, even if it's about a few bytes here. But I also agree that verbose(!) checking (XENO_ASSERT/XENO_BUGON) is very useful during development.
Ok, I understood that consistency checks can come for zero price when fixing a tiny bug in bheap_destroy. This is done in the attached revision of my patch, and I kept bheap_destroy therefore.
I also added XENO_BUGON to bheap_gethead, bheap_insert, and bheap_delete in case CONFIG_XENO_OPT_DEBUG_BHEAP is set (which is the case now for CONFIG_XENO_OPT_DEBUG - it's a cheap test).
Be warned, this patch is only compile-tested, I'm in a hurry.
Jan
nucleus/Kconfig +4 -0
@@ -145,6 +145,10 @@ config XENO_OPT_DEBUG_QUEUES operations of the Xenomai core. It adds even more runtime overhead then CONFIG_XENO_OPT_DEBUG, use with care. +config XENO_OPT_DEBUG_BHEAP + bool + default y if XENO_OPT_DEBUG + config XENO_OPT_WATCHDOG bool "Watchdog support" default n
nucleus/pod.c +4 -22
@@ -433,22 +433,11 @@ int xnpod_init(xnpod_t *pod, int minpri,#ifdef CONFIG_XENO_OPT_TIMING_PERIODIC unsigned n; -for (n =0 ; n <XNTIMER_WHEELSIZE ; n++) { - err = -xntlist_init (&pod->sched[cpu].timerwheel[n]); - -if (err) { -xnheap_destroy (&kheap, &xnpod_flush_heap,NULL ); - goto fail; - } - } +for (n =0 ; n <XNTIMER_WHEELSIZE ; n++) +xntlist_init (&pod->sched[cpu].timerwheel[n]);#endif /* CONFIG_XENO_OPT_TIMING_PERIODIC */ - err = -xntimerq_init (&pod->sched[cpu].timerqueue); - -if (err) { -xnheap_destroy (&kheap, &xnpod_flush_heap,NULL ); - goto fail; - } +xntimerq_init (&pod->sched[cpu].timerqueue); }for (cpu =0 ; cpu < nr_cpus; ++cpu) {
@@ -585,15 +574,8 @@ void xnpod_shutdown(int xtype)__setbits (nkpod->status,XNPIDLE ); -for (cpu =0 ; cpu <xnarch_num_online_cpus (); cpu++) { -#ifdef CONFIG_XENO_OPT_TIMING_PERIODIC -unsigned n; - -for (n =0 ; n <XNTIMER_WHEELSIZE ; n++) -xntlist_destroy (&nkpod->sched[cpu].timerwheel[n]); -#endif /* CONFIG_XENO_OPT_TIMING_PERIODIC */ +for (cpu =0 ; cpu <xnarch_num_online_cpus (); cpu++)xntimerq_destroy (&nkpod->sched[cpu].timerqueue); - }xnlock_put_irqrestore (&nklock, s);
nucleus/Config.in +1 -0
@@ -32,6 +32,7 @@ if [ "$CONFIG_XENO_OPT_NUCLEUS" != "n" ] bool 'Debug support' CONFIG_XENO_OPT_DEBUG if [ "$CONFIG_XENO_OPT_DEBUG" = "y" ]; then bool 'Queue Debugging support' CONFIG_XENO_OPT_DEBUG_QUEUES + define_bool CONFIG_XENO_OPT_DEBUG_BHEAP y fi bool 'Watchdog support' CONFIG_XENO_OPT_WATCHDOG
nucleus/bheap.h +40 -24
117 lines not on this page · raw message
nucleus/timer.h +13 -14
@@ -58,11 +58,10 @@ typedef struct { ((xntlholder_t *)(((char *)laddr) - offsetof(xntlholder_t, link))) } xntlholder_t; -#define xntlholder_date (h) ((h)->key) -#define xntlholder_prio (h) ((h)->prio) -#define xntlholder_init (h) inith(&(h)->link) -#define xntlist_init (q) (initq(q),0) -#define xntlist_destroy (q) do { } while (0) +#define xntlholder_date (h) ((h)->key) +#define xntlholder_prio (h) ((h)->prio) +#define xntlholder_init (h) inith(&(h)->link) +#define xntlist_init (q) initq(q)#define xntlist_head (q) \ ({ xnholder_t *_h = getheadq(q); \ !_h ? NULL : link2tlholder(_h); \
@@ -71,7 +70,7 @@ typedef struct {static inline void xntlist_insert (xnqueue_t *q,xntlholder_t *holder) {xnholder_t *p; - +/* Insert the new timer at the proper place in the single queue managed when running in aperiodic mode. O(N) here, but users of the aperiodic mode need to pay a price for
@@ -82,7 +81,7 @@ static inline void xntlist_insert(xnqueu (holder->key ==link2tlholder (p)->key && holder->prio <= link2tlholder(p)->prio)) break; - +insertq (q,p->next,&holder->link); }
@@ -94,12 +93,12 @@ typedef bheaph_t xntimerh_t;#define xntimerh_date (h) bheaph_key(h)#define xntimerh_prio (h) bheaph_prio(h)#define xntimerh_init (h) bheaph_init(h) -typedef bheap_t xntimerq_t ; -#define xntimerq_init (q) bheap_init((q), CONFIG_XENO_OPT_TIMER_HEAP_CAPACITY) -#define xntimerq_destroy (q) bheap_destroy(q) -#define xntimerq_head (q) bheap_gethead(q) -#define xntimerq_insert (q, h) bheap_insert((q),(h)) -#define xntimerq_remove (q, h) bheap_delete((q),(h)) +typedef DECLARE_BHEAP_CONTAINER(xntimerq_t, CONFIG_XENO_OPT_TIMER_HEAP_CAPACITY); +#define xntimerq_init (q) bheap_init(&(q)->bheap, CONFIG_XENO_OPT_TIMER_HEAP_CAPACITY) +#define xntimerq_destroy (q) bheap_destroy(&(q)->bheap) +#define xntimerq_head (q) bheap_gethead(&(q)->bheap) +#define xntimerq_insert (q, h) bheap_insert(&(q)->bheap,(h)) +#define xntimerq_remove (q, h) bheap_delete(&(q)->bheap,(h))#else /* CONFIG_XENO_OPT_TIMER_LIST */ typedef xntlholder_t xntimerh_t ;
@@ -108,7 +107,7 @@ typedef xntlholder_t xntimerh_t;#define xntimerh_init (h) xntlholder_init(h)typedef xnqueue_t xntimerq_t ;#define xntimerq_init (q) xntlist_init(q) -#define xntimerq_destroy (q) xntlist_destroy(q) +#define xntimerq_destroy (q) do { } while (0)#define xntimerq_head (q) xntlist_head(q)#define xntimerq_insert (q,h) xntlist_insert((q),(h))#define xntimerq_remove (q, h) xntlist_remove((q),(h))
ChangeLog +8 -1
@@ -1,3 +1,10 @@ +2006-06-02 Jan Kiszka <[email protected]> + + * include/nucleus/{bheap.h,timer.h}, ksrc/nucleus/pod.c: + Combine bheap head with element buffer, fix consistency checks + and additionally make use of XENO_BUGON. Simplify pod init and + cleanup of timer structures. + 2006-05-23 Gilles Chanteperdrix <[email protected]> * ksrc/arch/i386/nmi.c: Fix alignement for gcc-4.1.
@@ -14,7 +21,7 @@ * ksrc/arch/arm/patches: Upgrade to 2.6.1{4,5}-1.3-04. -2006-05-19 Jan Kiszka <[email protected]> +2006-05-19 Jan Kiszka <[email protected]> * src/testsuite/latency/latency.c: Add pid to registered names allowing multiple instances of latency to run. Add -c switch to
Quoted 1 line · Jan Kiszka
Be warned, this patch is only compile-tested, I'm in a hurry.
I tested the lightly modified version attached and it works. Do you agree with my changes ?
Signature · 1 line
Gilles Chanteperdrix.
nucleus/Kconfig +4 -0
@@ -145,6 +145,10 @@ operations of the Xenomai core. It adds even more runtime overhead then CONFIG_XENO_OPT_DEBUG, use with care. +config XENO_OPT_DEBUG_BHEAP + bool + default y if XENO_OPT_DEBUG + config XENO_OPT_WATCHDOG bool "Watchdog support" default n
nucleus/pod.c +4 -22
@@ -433,22 +433,11 @@#ifdef CONFIG_XENO_OPT_TIMING_PERIODIC unsigned n; -for (n =0 ; n <XNTIMER_WHEELSIZE ; n++) { - err = -xntlist_init (&pod->sched[cpu].timerwheel[n]); - -if (err) { -xnheap_destroy (&kheap, &xnpod_flush_heap,NULL ); - goto fail; - } - } +for (n =0 ; n <XNTIMER_WHEELSIZE ; n++) +xntlist_init (&pod->sched[cpu].timerwheel[n]);#endif /* CONFIG_XENO_OPT_TIMING_PERIODIC */ - err = -xntimerq_init (&pod->sched[cpu].timerqueue); - -if (err) { -xnheap_destroy (&kheap, &xnpod_flush_heap,NULL ); - goto fail; - } +xntimerq_init (&pod->sched[cpu].timerqueue); }for (cpu =0 ; cpu < nr_cpus; ++cpu) {
@@ -585,15 +574,8 @@__setbits (nkpod->status,XNPIDLE ); -for (cpu =0 ; cpu <xnarch_num_online_cpus (); cpu++) { -#ifdef CONFIG_XENO_OPT_TIMING_PERIODIC -unsigned n; - -for (n =0 ; n <XNTIMER_WHEELSIZE ; n++) -xntlist_destroy (&nkpod->sched[cpu].timerwheel[n]); -#endif /* CONFIG_XENO_OPT_TIMING_PERIODIC */ +for (cpu =0 ; cpu <xnarch_num_online_cpus (); cpu++)xntimerq_destroy (&nkpod->sched[cpu].timerqueue); - }xnlock_put_irqrestore (&nklock, s);
nucleus/Config.in +1 -0
@@ -32,6 +32,7 @@ bool 'Debug support' CONFIG_XENO_OPT_DEBUG if [ "$CONFIG_XENO_OPT_DEBUG" = "y" ]; then bool 'Queue Debugging support' CONFIG_XENO_OPT_DEBUG_QUEUES + define_bool CONFIG_XENO_OPT_DEBUG_BHEAP y fi bool 'Watchdog support' CONFIG_XENO_OPT_WATCHDOG
nucleus/bheap.h +62 -27
150 lines not on this page · raw message
nucleus/timer.h +8 -9
@@ -58,11 +58,10 @@ ((xntlholder_t *)(((char *)laddr) - offsetof(xntlholder_t, link))) } xntlholder_t; -#define xntlholder_date (h) ((h)->key) -#define xntlholder_prio (h) ((h)->prio) -#define xntlholder_init (h) inith(&(h)->link) -#define xntlist_init (q) (initq(q),0) -#define xntlist_destroy (q) do { } while (0) +#define xntlholder_date (h) ((h)->key) +#define xntlholder_prio (h) ((h)->prio) +#define xntlholder_init (h) inith(&(h)->link) +#define xntlist_init (q) initq(q)#define xntlist_head (q) \ ({ xnholder_t *_h = getheadq(q); \ !_h ? NULL : link2tlholder(_h); \
@@ -71,7 +70,7 @@static inline void xntlist_insert (xnqueue_t *q,xntlholder_t *holder) {xnholder_t *p; - +/* Insert the new timer at the proper place in the single queue managed when running in aperiodic mode. O(N) here, but users of the aperiodic mode need to pay a price for
@@ -82,7 +81,7 @@ (holder->key ==link2tlholder (p)->key && holder->prio <= link2tlholder(p)->prio)) break; - +insertq (q,p->next,&holder->link); }
@@ -94,7 +93,7 @@#define xntimerh_date (h) bheaph_key(h)#define xntimerh_prio (h) bheaph_prio(h)#define xntimerh_init (h) bheaph_init(h) -typedef bheap_t xntimerq_t ; +typedef DECLARE_BHEAP_CONTAINER(xntimerq_t, CONFIG_XENO_OPT_TIMER_HEAP_CAPACITY);#define xntimerq_init (q) bheap_init((q), CONFIG_XENO_OPT_TIMER_HEAP_CAPACITY)#define xntimerq_destroy (q) bheap_destroy(q)#define xntimerq_head (q) bheap_gethead(q)
@@ -108,7 +107,7 @@#define xntimerh_init (h) xntlholder_init(h)typedef xnqueue_t xntimerq_t ;#define xntimerq_init (q) xntlist_init(q) -#define xntimerq_destroy (q) xntlist_destroy(q) +#define xntimerq_destroy (q) do { } while (0)#define xntimerq_head (q) xntlist_head(q)#define xntimerq_insert (q,h) xntlist_insert((q),(h))#define xntimerq_remove (q, h) xntlist_remove((q),(h))
ChangeLog +8 -1
@@ -1,3 +1,10 @@ +2006-06-02 Jan Kiszka <[email protected]> + + * include/nucleus/{bheap.h,timer.h}, ksrc/nucleus/pod.c: + Combine bheap head with element buffer, fix consistency checks + and additionally make use of XENO_BUGON. Simplify pod init and + cleanup of timer structures. + 2006-05-23 Gilles Chanteperdrix <[email protected]> * ksrc/arch/i386/nmi.c: Fix alignement for gcc-4.1.
@@ -14,7 +21,7 @@ * ksrc/arch/arm/patches: Upgrade to 2.6.1{4,5}-1.3-04. -2006-05-19 Jan Kiszka <[email protected]> +2006-05-19 Jan Kiszka <[email protected]> * src/testsuite/latency/latency.c: Add pid to registered names allowing multiple instances of latency to run. Add -c switch to
Quoted 5 lines · Gilles Chanteperdrix
Quoted 1 line · Jan Kiszka
Be warned, this patch is only compile-tested, I'm in a hurry.
I tested the lightly modified version attached and it works. Do you agree with my changes ?
Yes, good point. Makes the container concept more consistent.
Jan