rework shadow CPU affinity
RFC PATCH 1/2 rework shadow CPU affinity
Picking up the earlier discussion of CPU affinity of shadow threads, this patch now opens the possibility to assign a CPU mask on xnshadow_map. It cleans up the related nucleus code and updates all callers of xnshadow_map appropriately.
Be warned, it's only tested on (q)emulated SMP. Comments and test reports on real hardware are welcome.
Jan
PS: I would consider these patches for 2.4.
--- include/asm-generic/system.h | 1 - include/nucleus/pod.h | 3 +++ include/nucleus/shadow.h | 4 +++- ksrc/nucleus/shadow.c | 43 ++++++++++++++++++------------------------- ksrc/skins/native/syscall.c | 3 ++- ksrc/skins/posix/syscall.c | 4 ++-- ksrc/skins/psos+/syscall.c | 3 ++- ksrc/skins/vrtx/syscall.c | 3 ++- ksrc/skins/vxworks/syscall.c | 3 ++- 9 files changed, 34 insertions(+), 33 deletions(-)
include/asm-generic/system.h +0 -1
@@ -38,7 +38,6 @@#include <asm/ptrace.h> #include <asm/xenomai/hal.h> #include <asm/xenomai/atomic.h> -#include <nucleus/shadow.h> /* Tracer interface */ #define xnarch_trace_max_begin (v) rthal_trace_max_begin(v)
include/nucleus/shadow.h +3 -1
@@ -22,6 +22,7 @@#include <asm/xenomai/atomic.h> #include <asm/xenomai/syscall.h> +#include <asm/xenomai/system.h> #ifdef CONFIG_PROC_FS #include <linux/proc_fs.h> #endif /* CONFIG_PROC_FS */
@@ -65,7 +66,8 @@ void xnshadow_grab_events(void);void xnshadow_release_events (void );int xnshadow_map (struct xnthread *thread, -xncompletion_t __user *u_completion); +xncompletion_t __user *u_completion, +xnarch_cpumask_t affinity);void xnshadow_unmap (struct xnthread *thread);
ksrc/nucleus/shadow.c +18 -25
87 lines not on this page · raw message
include/nucleus/pod.h +3 -0
@@ -33,6 +33,9 @@#include <nucleus/thread.h> #include <nucleus/intr.h> +#if defined(__KERNEL__) && defined(CONFIG_XENO_OPT_PERVASIVE ) +#include <nucleus/shadow.h> +#endif /* __KERNEL__ && CONFIG_XENO_OPT_PERVASIVE */ /* Creation flags */ #define XNREUSE 0x00000001/* Reuse pod with identical properties */
ksrc/skins/native/syscall.c +2 -1
@@ -173,7 +173,8 @@ static int __rt_task_create(struct task_ ph.opaque2 = bulk.a5;/* hidden pthread_t identifier. */ __xn_copy_to_user (curr, (void __user *)bulk.a1, &ph,sizeof (ph)); - err =xnshadow_map (&task->thread_base, u_completion); + err =xnshadow_map (&task->thread_base, u_completion, + task->affinity); }else {xnfree (task);/* Unblock and pass back error code. */
ksrc/skins/posix/syscall.c +2 -2
@@ -174,7 +174,7 @@ static int __pthread_create(struct task_if (err)return -err;/* Conventionally, our error codes are negative. */ - err =xnshadow_map (&k_tid->threadbase,NULL ); + err =xnshadow_map (&k_tid->threadbase,NULL ,XNPOD_ALL_CPUS );if (!err && !__pthread_hash (&hkey, k_tid)) err = -ENOMEM ;
@@ -205,7 +205,7 @@ static pthread_t __pthread_shadow(structif (err)return ERR_PTR (-err); - err =xnshadow_map (&k_tid->threadbase,NULL ); + err =xnshadow_map (&k_tid->threadbase,NULL ,XNPOD_ALL_CPUS );if (!err && !__pthread_hash (hkey, k_tid)) err = -EAGAIN ;
ksrc/skins/psos+/syscall.c +2 -1
@@ -98,7 +98,8 @@ static int __t_create(struct task_struct tid =xnthread_handle (&task->threadbase);__xn_copy_to_user (curr, (void __user *)__xn_reg_arg4 (regs), &tid,sizeof (tid)); - err =xnshadow_map (&task->threadbase, u_completion); + err =xnshadow_map (&task->threadbase, u_completion, +XNPOD_ALL_CPUS ); }else {/* Unblock and pass back error code. */
ksrc/skins/vrtx/syscall.c +2 -1
@@ -95,7 +95,8 @@ static int __sc_tecreate(struct task_str }else {__xn_copy_to_user (curr, (void __user *)__xn_reg_arg2 (regs), &tid,sizeof (tid)); - err =xnshadow_map (&task->threadbase, u_completion); + err =xnshadow_map (&task->threadbase, u_completion, +XNPOD_ALL_CPUS ); }if (err)
ksrc/skins/vxworks/syscall.c +2 -1
@@ -122,7 +122,8 @@ static int __wind_task_init(struct task_ ph.handle =xnthread_handle (&task->threadbase);__xn_copy_to_user (curr, (void __user *)__xn_reg_arg2 (regs), &ph,sizeof (ph)); - err =xnshadow_map (&task->threadbase, u_completion); + err =xnshadow_map (&task->threadbase, u_completion, +XNPOD_ALL_CPUS ); }else {/* Unblock and pass back error code. */
Quoted 7 lines · Jan Kiszka
Picking up the earlier discussion of CPU affinity of shadow threads, this patch now opens the possibility to assign a CPU mask on
xnshadow_map. It cleans up the related nucleus code and updates all callers ofxnshadow_mapappropriately.Be warned, it's only tested on (q)emulated SMP. Comments and test reports on real hardware are welcome.
It is only possible to set the affinity of a user-space thread, why adding another way to do it ?
Signature · 1 line
Gilles Chanteperdrix
Quoted 6 lines · Gilles Chanteperdrix
Quoted 5 lines · Jan Kiszka
Picking up the earlier discussion of CPU affinity of shadow threads, this patch now opens the possibility to assign a CPU mask on
xnshadow_map. It cleans up the related nucleus code and updates all callers ofxnshadow_mapappropriately.
Quoted 6 lines
Quoted 2 lines
Be warned, it's only tested on (q)emulated SMP. Comments and test reports on real hardware are welcome.
It is only possible to set the affinity of a user-space thread, why adding another way to do it ?
I'm not yet completely understanding what you mean.
So far the affinity was indirectly assigned to some xnthread by the CPU the caller of xnshadow_map happened to run on. Now this is done explicitly, thus in a more robust way (there was an undefined path in the old code) and without duplicating code (/wrt skins).
Jan
Quoted 17 lines · Jan Kiszka
Quoted 7 lines · Gilles Chanteperdrix
Quoted 5 lines · Jan Kiszka
Picking up the earlier discussion of CPU affinity of shadow threads, this patch now opens the possibility to assign a CPU mask on
xnshadow_map. It cleans up the related nucleus code and updates all callers ofxnshadow_mapappropriately.Quoted 6 lines
Quoted 2 lines
Be warned, it's only tested on (q)emulated SMP. Comments and test reports on real hardware are welcome.
It is only possible to set the affinity of a user-space thread, why adding another way to do it ?
Quoted 6 lines
I'm not yet completely understanding what you mean.
So far the affinity was indirectly assigned to some xnthread by the CPU the caller of
xnshadow_maphappened to run on. Now this is done explicitly, thus in a more robust way (there was an undefined path in the old code) and without duplicating code (/wrt skins).
I meant "already" instead of "only", it is already possible to set the affinity of a user-space thread with sched_setaffinity.
Signature · 1 line
Gilles Chanteperdrix
Quoted 18 lines · Gilles Chanteperdrix
Quoted 17 lines · Jan Kiszka
Quoted 7 lines · Gilles Chanteperdrix
Quoted 5 lines · Jan Kiszka
Picking up the earlier discussion of CPU affinity of shadow threads, this patch now opens the possibility to assign a CPU mask on
xnshadow_map. It cleans up the related nucleus code and updates all callers ofxnshadow_mapappropriately.Quoted 5 lines
Quoted 2 lines
Be warned, it's only tested on (q)emulated SMP. Comments and test reports on real hardware are welcome.
It is only possible to set the affinity of a user-space thread, why adding another way to do it ?
I'm not yet completely understanding what you mean.
Quoted 8 lines
Quoted 4 lines
So far the affinity was indirectly assigned to some xnthread by the CPU the caller of
xnshadow_maphappened to run on. Now this is done explicitly, thus in a more robust way (there was an undefined path in the old code) and without duplicating code (/wrt skins).I meant "already" instead of "only", it is already possible to set the affinity of a user-space thread with
sched_setaffinity.
OK. So you are suggesting to read out the affinity mask from task_struct instead of passing it as an additional argument?
Quoted 3 lines · Jan Kiszka
OK. So you are suggesting to read out the affinity mask from
task_structinstead of passing it as an additional argument?
Unless I am wrong, it is the way it currently works.
Signature · 1 line
Gilles Chanteperdrix
Quoted 7 lines · Gilles Chanteperdrix
Quoted 3 lines · Jan Kiszka
OK. So you are suggesting to read out the affinity mask from
task_structinstead of passing it as an additional argument?Unless I am wrong, it is the way it currently works.
Yes, for the posix skin. But native is different due to the spilt-up create/start and its own affinity API. And that caused problems when applying nkaffinity.
Nevertheless, I should try to combine the cleanup of xnshadow_map with passing the mask ALWAYS in task_struct. Let's see how this works out, most likely with less patching.
Quoted 8 lines · Jan Kiszka
Quoted 6 lines · Gilles Chanteperdrix
Quoted 3 lines · Jan Kiszka
OK. So you are suggesting to read out the affinity mask from
task_structinstead of passing it as an additional argument?Unless I am wrong, it is the way it currently works.
Quoted 7 lines
Yes, for the posix skin. But native is different due to the spilt-up create/start and its own affinity API. And that caused problems when applying nkaffinity.
Nevertheless, I should try to combine the cleanup of
xnshadow_mapwith passing the mask ALWAYS intask_struct. Let's see how this works out, most likely with less patching.
sched_setaffinity is 2.6 only. So we would have to push the mask application into kernel space for the native skin anyway. Hmm.
What are posix skin users supposed to do over 2.4? Upgrade to 2.6? BTW, do you plan to support pthread_attr_setaffinity_np somehow?
Jan
Quoted 22 lines · Jan Kiszka
Quoted 11 lines · Jan Kiszka
Quoted 8 lines · Gilles Chanteperdrix
Quoted 3 lines · Jan Kiszka
OK. So you are suggesting to read out the affinity mask from
task_structinstead of passing it as an additional argument?Unless I am wrong, it is the way it currently works.
Quoted 7 lines
Yes, for the posix skin. But native is different due to the spilt-up create/start and its own affinity API. And that caused problems when applying nkaffinity.
Nevertheless, I should try to combine the cleanup of
xnshadow_mapwith passing the mask ALWAYS intask_struct. Let's see how this works out, most likely with less patching.
Quoted 4 lines
sched_setaffinityis 2.6 only. So we would have to push the mask application into kernel space for the native skin anyway. Hmm.What are posix skin users supposed to do over 2.4?
Right, I did not think about the 2.4 kernel, but I am not sure there are many people using the 2.4 kernel on SMP machines anyway.
As for the native skin, it already allows to pass the affinity to rt_task_create, using T_CPU.
Quoted 1 line
Upgrade to 2.6? BTW, do you plan to support
pthread_attr_setaffinity_npsomehow?
pthread_attr_setaffinity_np is implemented in kernel space, and the pthread_attr_t is passed unchanged by __wrap_pthread_create to __real_pthread_create. So pthread_attr_setaffinity_np should be already working provided that you use a libc that supports it.
Quoted 3 lines
Jan
Signature · 1 line
Gilles Chanteperdrix
Second try of this patch, now consequently basing the affinity API for xnshadow_map on current->cpus_allowed (affects the native skin). Just like the current code does, this patch restricts a shadow thread to a single (the first) CPU of its affinity mask.
Jan
--- ksrc/nucleus/shadow.c | 35 +++++++++++++---------------------- ksrc/skins/native/syscall.c | 3 +++ 2 files changed, 16 insertions(+), 22 deletions(-)
ksrc/nucleus/shadow.c +13 -22
@@ -409,18 +409,6 @@ static void lostage_handler(void *cookiexnltt_log_event (xeno_ev_lohandler, reqnum, p->comm, p->pid);switch (rq->req[reqnum].type) { -case LO_START_REQ : - -#ifdef CONFIG_SMP -if (xnshadow_thread (p)) -/* Set up the initial task affinity using the -information passed to xnpod_start_thread(). */ -set_cpus_allowed (p, -xnshadow_thread (p)->affinity); -#endif /* CONFIG_SMP */ - - goto do_wakeup; -case LO_WAKEUP_REQ :#ifdef CONFIG_SMP
@@ -443,7 +431,9 @@ static void lostage_handler(void *cookieif (!xnshadow_thread (current))xnpod_renice_root (XNPOD_ROOT_PRIO_BASE );#endif /* CONFIG_XENO_OPT_RPIDISABLE */ - do_wakeup: + +/* fall through */ +case LO_START_REQ :#ifdef CONFIG_XENO_OPT_ISHIELD if (xnshadow_thread(p) &&
@@ -853,22 +843,23 @@ int xnshadow_map(xnthread_t *thread, xncxnshadow_thrptd (current) = thread;xnpod_suspend_thread (thread,XNRELAX ,XN_INFINITE ,NULL ); +/* Restrict affinity to a single CPU of the current set */ + affinity =xnarch_cpumask_of_cpu ( +xnarch_first_cpu (current->cpus_allowed)); +set_cpus_allowed (current, affinity); +if (u_completion) {xnshadow_signal_completion (u_completion,0 );return 0 ; } -/* Nobody waits for us, so we may start the shadow immediately -after having forced the CPU affinity to the current -processor. Note that we don't use smp_processor_id() to prevent -kernel debug stuff to yell at us for calling it in a preemptible -section of code. */ - - affinity =xnarch_cpumask_of_cpu (rthal_processor_id ()); -set_cpus_allowed (current, affinity); +/* Nobody waits for us, so we may start the shadow immediately. */ mode = thread->rrperiod !=XN_INFINITE ?XNRRB :0 ; -xnpod_start_thread (thread, mode,0 , affinity,NULL ,NULL ); + err =xnpod_start_thread (thread, mode,0 , affinity,NULL ,NULL ); + +if (err) +return err; err =xnshadow_harden ();
ksrc/skins/native/syscall.c +3 -0
@@ -168,6 +168,9 @@ static int __rt_task_create(struct task_ err =rt_task_create (task, name,0 , prio,XNFPU |XNSHADOW | mode);if (err ==0 ) { +/* Apply CPU affinity */ +set_cpus_allowed (current, task->affinity); +/* Copy back the registry handle to the ph struct. */ ph.opaque =xnthread_handle (&task->thread_base); ph.opaque2 = bulk.a5;/* hidden pthread_t identifier. */
Quoted 23 lines · Gilles Chanteperdrix
Quoted 22 lines · Jan Kiszka
Quoted 18 lines · Jan Kiszka
Quoted 8 lines · Gilles Chanteperdrix
Quoted 3 lines · Jan Kiszka
OK. So you are suggesting to read out the affinity mask from
task_structinstead of passing it as an additional argument?Unless I am wrong, it is the way it currently works.
Yes, for the posix skin. But native is different due to the spilt-up create/start and its own affinity API. And that caused problems when applying nkaffinity.
Nevertheless, I should try to combine the cleanup of
xnshadow_mapwith passing the mask ALWAYS intask_struct. Let's see how this works out, most likely with less patching.
Quoted 8 lines
Quoted 4 lines
sched_setaffinityis 2.6 only. So we would have to push the mask application into kernel space for the native skin anyway. Hmm.What are posix skin users supposed to do over 2.4?
Right, I did not think about the 2.4 kernel, but I am not sure there are many people using the 2.4 kernel on SMP machines anyway.
Ack. 2.4 backports are provided to help people who do not want to double jump from <some-rtos> to Xenomai _and_ from 2.4 to 2.6 in the same move for upgrading some embedded setup; they can move to Xenomai keeping their good old system software if they wish, then move to 2.6 if they wish so later, once the dust has settled. People interested by multi-cores and/or big irons of some sort should obviously go for 2.6.
Quoted 17 lines
As for the native skin, it already allows to pass the affinity to
rt_task_create, usingT_CPU.Quoted 1 line
Upgrade to 2.6? BTW, do you plan to support
pthread_attr_setaffinity_npsomehow?
pthread_attr_setaffinity_npis implemented in kernel space, and thepthread_attr_tis passed unchanged by__wrap_pthread_createto__real_pthread_create. Sopthread_attr_setaffinity_npshould be already working provided that you use a libc that supports it.Quoted 3 lines
Jan
Signature · 1 line
Philippe.