Message-ID: <20080427005816.GB21687@linux.vnet.ibm.com>
Date:	Sat, 26 Apr 2008 17:58:16 -0700
From:	"Paul E. McKenney" <paulmck@...ux.vnet.ibm.com>
To:	Andrew Morton <akpm@...ux-foundation.org>
Cc:	Jens Axboe <jens.axboe@...cle.com>, linux-arch@...r.kernel.org,
	linux-kernel@...r.kernel.org, npiggin@...e.de,
	torvalds@...ux-foundation.org, peterz@...radead.org,
	sam@...nborg.org
Subject: Re: [PATCH 1/11] Add generic helpers for arch IPI function calls

On Fri, Apr 25, 2008 at 11:11:00PM -0700, Andrew Morton wrote:
> > On Tue, 22 Apr 2008 20:50:17 +0200 Jens Axboe <jens.axboe@...cle.com> wrote:
> > This adds kernel/smp.c which contains helpers for IPI function calls. In
> > addition to supporting the existing smp_call_function() in a more efficient
> > manner, it also adds a more scalable variant called smp_call_function_single()
> > for calling a given function on a single CPU only.

Where is the new smp_call_function() defined?  I see only the declaration.

Probably just blind...

See comments interspersed below.

> > The core of this is based on the x86-64 patch from Nick Piggin, lots of
> > changes since then. "Alan D. Brunelle" <Alan.Brunelle@...com> has
> > contributed lots of fixes and suggestions as well.
> > 
> > Acked-by: Ingo Molnar <mingo@...e.hu>
> > Signed-off-by: Jens Axboe <jens.axboe@...cle.com>
> > ---
> >  arch/Kconfig        |    3 +
> >  include/linux/smp.h |   27 ++++-
> >  init/main.c         |    3 +
> >  kernel/Makefile     |    1 +
> >  kernel/smp.c        |  366 +++++++++++++++++++++++++++++++++++++++++++++++++++
> >  5 files changed, 398 insertions(+), 2 deletions(-)
> >  create mode 100644 kernel/smp.c
> > 
> > diff --git a/arch/Kconfig b/arch/Kconfig
> > index 694c9af..a5a0184 100644
> > --- a/arch/Kconfig
> > +++ b/arch/Kconfig
> > @@ -36,3 +36,6 @@ config HAVE_KPROBES
> >  
> >  config HAVE_KRETPROBES
> >  	def_bool n
> > +
> > +config USE_GENERIC_SMP_HELPERS
> > +	def_bool n
> > diff --git a/include/linux/smp.h b/include/linux/smp.h
> > index 55232cc..4a5418b 100644
> > --- a/include/linux/smp.h
> > +++ b/include/linux/smp.h
> > @@ -7,9 +7,19 @@
> >   */
> >  
> >  #include <linux/errno.h>
> > +#include <linux/list.h>
> > +#include <linux/spinlock.h>
> > +#include <linux/cpumask.h>
> >  
> >  extern void cpu_idle(void);
> >  
> > +struct call_single_data {
> > +	struct list_head list;
> > +	void (*func) (void *info);
> > +	void *info;
> > +	unsigned int flags;
> > +};
> > +
> >  #ifdef CONFIG_SMP
> >  
> >  #include <linux/preempt.h>
> > @@ -53,9 +63,23 @@ extern void smp_cpus_done(unsigned int max_cpus);
> >   * Call a function on all other processors
> >   */
> >  int smp_call_function(void(*func)(void *info), void *info, int retry, int wait);
> > -
> > +int smp_call_function_mask(cpumask_t mask, void(*func)(void *info), void *info,
> > +				int wait);
> >  int smp_call_function_single(int cpuid, void (*func) (void *info), void *info,
> >  				int retry, int wait);
> > +void __smp_call_function_single(int cpuid, struct call_single_data *data);
> > +
> > +/*
> > + * Generic and arch helpers
> > + */
> > +#ifdef CONFIG_USE_GENERIC_SMP_HELPERS
> > +void generic_smp_call_function_single_interrupt(void);
> > +void generic_smp_call_function_interrupt(void);
> > +void init_call_single_data(void);
> > +void arch_send_call_function_single_ipi(int cpu);
> > +void arch_send_call_function_ipi(cpumask_t mask);
> > +extern spinlock_t call_function_lock;
> > +#endif
> 
> Add a #else here ...
> 
> > diff --git a/init/main.c b/init/main.c
> > index 833a67d..0b7578c 100644
> > --- a/init/main.c
> > +++ b/init/main.c
> > @@ -773,6 +773,9 @@ static void __init do_pre_smp_initcalls(void)
> >  {
> >  	extern int spawn_ksoftirqd(void);
> >  
> > +#ifdef CONFIG_USE_GENERIC_SMP_HELPERS
> > +	init_call_single_data();
> > +#endif
> 
> and we can lose this ifdef here.
> 
> Also, init_call_single_data() is __cpuinit but isn't declared that way. 
> There have been rare occasions (FRV was one) where this matters - iirc the
> compiler was emitting a short-relative-addressing-form instruction which
> which wasn't able to branch far enough once things were linked.  We hae
> this problem in eight zillion other places, of course.  And it's a pita to
> go adding __cpunit etc to the declaration because the compiler usually
> won't tell us when it gets out of sync with reality.  So we could leave he
> code as-is and wait for stuff to break :(
> 
> 
> >  	migration_init();
> >  	spawn_ksoftirqd();
> >  	if (!nosoftlockup)
> > diff --git a/kernel/Makefile b/kernel/Makefile
> > index 6c5f081..7e275d4 100644
> > --- a/kernel/Makefile
> > +++ b/kernel/Makefile
> > @@ -28,6 +28,7 @@ obj-$(CONFIG_DEBUG_RT_MUTEXES) += rtmutex-debug.o
> >  obj-$(CONFIG_RT_MUTEX_TESTER) += rtmutex-tester.o
> >  obj-$(CONFIG_GENERIC_ISA_DMA) += dma.o
> >  obj-$(CONFIG_SMP) += cpu.o spinlock.o
> > +obj-$(CONFIG_USE_GENERIC_SMP_HELPERS) += smp.o
> >  obj-$(CONFIG_DEBUG_SPINLOCK) += spinlock.o
> >  obj-$(CONFIG_PROVE_LOCKING) += spinlock.o
> >  obj-$(CONFIG_UID16) += uid16.o
> > diff --git a/kernel/smp.c b/kernel/smp.c
> > new file mode 100644
> > index 0000000..a177a0d
> > --- /dev/null
> > +++ b/kernel/smp.c
> > @@ -0,0 +1,366 @@
> > +/*
> > + * Generic helpers for smp ipi calls
> > + *
> > + * (C) Jens Axboe <jens.axboe@...cle.com> 2008
> > + *
> > + */
> > +#include <linux/init.h>
> > +#include <linux/module.h>
> > +#include <linux/percpu.h>
> > +#include <linux/rcupdate.h>
> > +#include <linux/smp.h>
> > +
> > +static DEFINE_PER_CPU(struct call_single_queue, call_single_queue);
> > +static LIST_HEAD(call_function_queue);
> > +__cacheline_aligned_in_smp DEFINE_SPINLOCK(call_function_lock);
> > +
> > +enum {
> > +	CSD_FLAG_WAIT		= 0x01,
> > +	CSD_FLAG_ALLOC		= 0x02,
> > +	CSD_FLAG_FALLBACK	= 0x04,
> > +};
> > +
> > +struct call_function_data {
> > +	struct call_single_data csd;
> > +	spinlock_t lock;
> > +	unsigned int refs;
> > +	cpumask_t cpumask;
> > +	struct rcu_head rcu_head;
> > +};
> > +
> > +struct call_single_queue {
> > +	struct list_head list;
> > +	spinlock_t lock;
> > +};
> > +
> > +static struct call_function_data cfd_fallback;
> > +static unsigned long cfd_fallback_used;
> > +
> > +void __cpuinit init_call_single_data(void)
> > +{
> > +	int i;
> > +
> > +	for_each_cpu_mask(i, cpu_possible_map) {
> 
> for_each_possible_cpu()?
> 
> Do we _have_ to consider all possible CPUs here?  That can be much larger
> than num_online_cpus.
> 
> > +		struct call_single_queue *q = &per_cpu(call_single_queue, i);
> > +
> > +		spin_lock_init(&q->lock);
> > +		INIT_LIST_HEAD(&q->list);
> > +	}
> > +}
> > +
> > +static inline void csd_flag_wait(struct call_single_data *data)
> > +{
> > +	/* Wait for response */
> > +	do {
> > +		/*
> > +		 * We need to see the flags store in the IPI handler
> > +		 */
> > +		smp_mb();

But don't we need the smp_mb() -after- the CSD_FLAG_WAIT check?

We want the caller to see the effects of the on-other-CPU-called
function upon return, correct?

> > +		if (!(data->flags & CSD_FLAG_WAIT))
> > +			break;
> > +		cpu_relax();
> > +	} while (1);
> > +}
> 
> -ETOOLAGETOINLINE?
> 
> > +/*
> > + * Insert a previously allocated call_single_data element for execution
> > + * on the given CPU. data must already have ->func, ->info, and ->flags set.
> > + */
> > +static void generic_exec_single(int cpu, struct call_single_data *data)
> > +{
> > +	struct call_single_queue *dst = &per_cpu(call_single_queue, cpu);
> > +	int wait = data->flags & CSD_FLAG_WAIT, ipi;
> > +	unsigned long flags;
> > +
> > +	spin_lock_irqsave(&dst->lock, flags);
> > +	ipi = list_empty(&dst->list);
> > +	list_add_tail(&data->list, &dst->list);
> > +	spin_unlock_irqrestore(&dst->lock, flags);
> > +
> > +	if (ipi)
> > +		arch_send_call_function_single_ipi(cpu);
> > +
> > +	if (wait)
> > +		csd_flag_wait(data);
> > +}
> 
> Is this a must-be-called-with-local-interrupts-enable function?  (It
> doesn't say so in the covering comment) If so, a runtime check for that
> would be needed - we screw that up regularly.  Ditto any other places where
> this applies.
> 
> > +static void rcu_free_call_data(struct rcu_head *head)
> > +{
> > +	struct call_function_data *cfd;
> > +
> > +	cfd = container_of(head, struct call_function_data, rcu_head);
> > +	kfree(cfd);
> > +}
> > +
> > +static void call_func_data_free(struct call_function_data *data)
> > +{
> > +	if (data->csd.flags & CSD_FLAG_ALLOC)
> > +		call_rcu(&data->rcu_head, rcu_free_call_data);
> > +	else
> > +		clear_bit_unlock(0, &cfd_fallback_used);
> > +}
> 
> Let's cc Mr RCU - he reviews well and often finds problems.

Well, one question...  Is it really OK to clear this bit immediately?
Why don't you need to wait for a grace period to elapse before freeing it?

(OK, I eventually saw why you can't wait for a grace period, but you
still seem to have a bug...  See interspersed below...)

> > +/*
> > + * Invoked by arch to handle an IPI for call function. Must be called with
> > + * interrupts disabled.
> 
> A runtime check would be nice.  The get_cpu() will give us partial coverage
> but won't detect irqs-on-inside-spinlock state.
> 
> > + */
> > +void generic_smp_call_function_interrupt(void)
> > +{
> > +	struct list_head *pos;
> > +	int cpu = get_cpu();
> > +
> > +	/*
> > +	 * It's ok to use list_for_each_rcu() here even though we may delete
> > +	 * 'pos', since list_del_rcu() doesn't clear ->next
> > +	 */
> > +	rcu_read_lock();
> 
> Sigh.  irqs are disabled, so this is a waste of CPU cycles.  With some of
> our many RCU flavours, at least.

To safely omit the rcu_read_lock(), Jens would need to use the in-flight
call_rcu_sched() instead of call_rcu().  In addition, if we ever get
threaded interrupts, this interrupt handler would need IRQ_NODELAY.

So keeping the rcu_read_lock() woud be more future-proof...

> > +	list_for_each_rcu(pos, &call_function_queue) {
> > +		struct call_function_data *data;
> > +		int refs;
> > +
> > +		data = list_entry(pos, struct call_function_data, csd.list);

How about using list_for_each_entry_rcu() instead?

	list_for_each_entry_rcu(pos, &call_function_queue, csd.list) {
		int refs;

And then move the definition of "data" up to replace that of "pos".

A couple fewer lines and one fewer local variable.  ;-)

> > +		if (!cpu_isset(cpu, data->cpumask))
> > +			continue;

Need an smp_mb() here, at least in the case where you are using the
fallback queue element.  See my comments in smp_call_function_mask()
for the justification -- the key point is that if the CPU sees itself
in the mask, it also needs to see the data element all filled out.

See below for a scenario requiring this.

Such a memory barrier is -not- needed for the normally allocated
queue elements.

But memory barriers are a bit ugly...  How about instead acquiring the
data->lock around the cpu_isset() check above?

> > +
> > +		data->csd.func(data->csd.info);
> > +
> > +		spin_lock(&data->lock);
> > +		cpu_clear(cpu, data->cpumask);
> > +		WARN_ON(data->refs == 0);
> > +		data->refs--;
> > +		refs = data->refs;
> > +		spin_unlock(&data->lock);
> > +
> > +		if (refs)
> > +			continue;
> > +
> > +		WARN_ON(cpus_weight(data->cpumask));
> 
> !cpus_empty()
> 
> > +		spin_lock(&call_function_lock);
> > +		list_del_rcu(&data->csd.list);
> > +		spin_unlock(&call_function_lock);
> > +
> > +		if (data->csd.flags & CSD_FLAG_WAIT) {
> > +			smp_wmb();
> 
> It's nice to comment open-coded barriers.
> 
> > +			data->csd.flags &= ~CSD_FLAG_WAIT;
> > +		} else
> > +			call_func_data_free(data);

And the above statements are one reason that I believe you need to wait
for a grace period even when using the static fallback element.  Here
is the sequence of events that I am worried about:

o	CPU 0 is the last CPU to invoke a given queue element, and is just
	about to list_del_rcu() it.

o	CPU 1 gets the interrupt (but already did this block last time).
	It nevertheless finds the queue element, because CPU 0 has not
	yet list_del_rcu() it.

	CPU 1 is just about to execute the: 
	
		"if (!cpu_isset(cpu, data->cpumask))"

o	CPU 0 now does the list_del_rcu() that CPU 1 already has a
	reference to, and eventually invokes the call_func_data_free().

o	call_func_data_free() sees the CSD_FLAG_ALLOC flag set, and
	therefore simply clears the bit, without waiting for a grace
	period.

o	The queue element is eligible to be reused.  CPU 0 therefore
	starts filling in data.  Note that both the compiler and the
	CPU are free to fully scramble the order of the queue element's
	initialization.

o	CPU 1 resumes execution, sees the half-filled-in queue element,
	which might have CPU 1's bit set but garbage otherwise.

	Life gets hard.  For example, CPU 1 might end up calling the
	wrong function, or the right function with the wrong argument,
	and so on.

Or am I missing something here?

Well, one thing that I was missing is that you are spinning waiting for
the element to become free, so any such grace period would never terminate.

Blech.  Some alternatives:

o	Use memory barriers.  See my comments in smp_call_function_mask()
	and the comment above.

o	Move the check of the bitmask under the per-element lock.
	Then the initialization sequence must also acquire the lock,
	which means that the fallback element must have its lock
	initialized at boot-up.

Other thoughts?

> > +	}
> > +	rcu_read_unlock();
> > +
> > +	put_cpu();
> > +}
> > +
> > +/*
> > + * Invoked by arch to handle an IPI for call function single. Must be called
> > + * from the arch with interrupts disabled.
> 
> runtime check?
> 
> > + */
> > +void generic_smp_call_function_single_interrupt(void)
> > +{
> > +	struct call_single_queue *q = &__get_cpu_var(call_single_queue);
> > +	LIST_HEAD(list);
> > +
> > +	smp_mb();
> 
> Unclear why this is here - comment?

Got me on that one...  Forcing ordering between the initialization of
"q" and the list_empty() check?  Don't immediately see why that is
needed.

> > +	while (!list_empty(&q->list)) {
> > +		unsigned int data_flags;
> > +
> > +		spin_lock(&q->lock);
> > +		list_replace_init(&q->list, &list);
> > +		spin_unlock(&q->lock);
> > +
> > +		while (!list_empty(&list)) {
> > +			struct call_single_data *data;
> > +
> > +			data = list_entry(list.next, struct call_single_data,
> > +						list);
> > +			list_del(&data->list);
> > +
> > +			/*
> > +			 * 'data' can be invalid after this call if
> > +			 * flags == 0 (when called through
> > +			 * generic_exec_single(), so save them away before
> > +			 * making the call.
> > +			 */
> > +			data_flags = data->flags;
> > +
> > +			data->func(data->info);
> > +
> > +			if (data_flags & CSD_FLAG_WAIT) {
> > +				smp_wmb();

This one, as well.  Looks like you want to make sure that the above
function call's writes are seen before the following flag-clearing.

> > +				data->flags &= ~CSD_FLAG_WAIT;
> > +			} else if (data_flags & CSD_FLAG_ALLOC)
> > +				kfree(data);
> > +			else if (data_flags & CSD_FLAG_FALLBACK)
> > +				clear_bit_unlock(0, &cfd_fallback_used);
> > +		}
> > +		smp_mb();

Can't say that I understand what the above memory barrier is doing...

> > +	}
> > +}
> > +
> > +/*
> > + * smp_call_function_single - Run a function on a specific CPU
> > + * @func: The function to run. This must be fast and non-blocking.
> > + * @info: An arbitrary pointer to pass to the function.
> > + * @retry: Unused
> > + * @wait: If true, wait until function has completed on other CPUs.
> > + *
> > + * Returns 0 on success, else a negative status code.
> > + *
> 
> stray line here.
> 
> > + */
> 
> Some of the exported functions have kerneldoc, others do not.
> 
> > +int smp_call_function_single(int cpu, void (*func) (void *info), void *info,
> > +			     int retry, int wait)
> > +{
> > +	unsigned long flags;
> > +	/* prevent preemption and reschedule on another processor */
> > +	int me = get_cpu();
> > +
> > +	/* Can deadlock when called with interrupts disabled */
> > +	WARN_ON(wait && irqs_disabled());
> 
> Perhaps that addresses above comments?
> 
> > +	if (cpu == me) {
> > +		local_irq_save(flags);
> > +		func(info);
> > +		local_irq_restore(flags);
> > +	} else {
> > +		struct call_single_data *data;
> > +
> > +		if (wait) {
> > +			struct call_single_data d;
> > +
> > +			data = &d;
> > +			data->flags = CSD_FLAG_WAIT;
> > +		} else {
> > +			data = kmalloc(sizeof(*data), GFP_ATOMIC);
> > +			if (data)
> > +				data->flags = CSD_FLAG_ALLOC;
> > +			else {
> > +				while (test_and_set_bit_lock(0,
> > +							&cfd_fallback_used))
> 
> cfd_fallback_used is rather mysterious - it could do with a comment at its
> definition site.  I'm wondering if we should/could be doing bit_spin_lock
> on it.  But if we did that, it could just be a spinlock.
> 
> > +					cpu_relax();
> > +
> > +				data = &cfd_fallback.csd;
> > +				data->flags = CSD_FLAG_FALLBACK;
> > +			}
> > +		}
> > +
> > +		data->func = func;
> > +		data->info = info;
> > +		generic_exec_single(cpu, data);
> > +	}
> > +
> > +	put_cpu();
> > +	return 0;
> > +}
> > +EXPORT_SYMBOL(smp_call_function_single);
> > +
> > +/**
> > + * __smp_call_function_single(): Run a function on another CPU
> > + * @cpu: The CPU to run on.
> > + * @data: Pre-allocated and setup data structure
> > + *
> > + * Like smp_call_function_single(), but allow caller to pass in a pre-allocated
> > + * data structure. Useful for embedding @data inside other structures, for
> > + * instance.
> > + *
> > + */
> > +void __smp_call_function_single(int cpu, struct call_single_data *data)
> > +{
> > +	generic_exec_single(cpu, data);
> > +}
> > +
> > +/**
> > + * smp_call_function_mask(): Run a function on a set of other CPUs.
> > + * @mask: The set of cpus to run on.
> > + * @func: The function to run. This must be fast and non-blocking.
> > + * @info: An arbitrary pointer to pass to the function.
> > + * @wait: If true, wait (atomically) until function has completed on other CPUs.
> > + *
> > + * Returns 0 on success, else a negative status code.
> > + *
> > + * If @wait is true, then returns once @func has returned.
> > + *
> > + * You must not call this function with disabled interrupts or from a
> > + * hardware interrupt handler or from a bottom half handler.
> > + */
> > +int smp_call_function_mask(cpumask_t mask, void (*func)(void *), void *info,
> > +			   int wait)
> > +{
> > +	struct call_function_data *data;
> > +	cpumask_t allbutself;
> > +	unsigned long flags;
> > +	int num_cpus;
> > +
> > +	/* Can deadlock when called with interrupts disabled */
> > +	WARN_ON(wait && irqs_disabled());
> 
> Ditto
> 
> > +	allbutself = cpu_online_map;
> > +	cpu_clear(smp_processor_id(), allbutself);
> > +	cpus_and(mask, mask, allbutself);
> > +	num_cpus = cpus_weight(mask);
> > +
> > +	if (!num_cpus)
> > +		return 0;
> > +
> > +	if (wait) {
> > +		struct call_function_data d;
> > +
> > +		data = &d;
> > +		data->csd.flags = CSD_FLAG_WAIT;
> > +	} else {
> > +		data = kmalloc(sizeof(*data), GFP_ATOMIC);
> > +		if (data)
> > +			data->csd.flags = CSD_FLAG_ALLOC;
> > +		else {
> > +			while (test_and_set_bit_lock(0, &cfd_fallback_used))
> > +				cpu_relax();
> > +
> > +			data = &cfd_fallback;
> > +			data->csd.flags = CSD_FLAG_FALLBACK;
> > +		}
> > +	}
> > +
> > +	spin_lock_init(&data->lock);
> > +	data->csd.func = func;
> > +	data->csd.info = info;
> > +	data->refs = num_cpus;

You need an smp_wmb() here.  Otherwise, a CPU that picked up a reference
to the fallback element after having processed it on the first pass could
see the next use of the fallback element half-initialized.

See 

> > +	data->cpumask = mask;
> > +
> > +	spin_lock_irqsave(&call_function_lock, flags);
> > +	list_add_tail_rcu(&data->csd.list, &call_function_queue);
> > +	spin_unlock_irqrestore(&call_function_lock, flags);
> > +
> > +	/* Send a message to all CPUs in the map */
> > +	arch_send_call_function_ipi(mask);
> > +
> > +	/* optionally wait for the CPUs to complete */
> > +	if (wait)
> > +		csd_flag_wait(&data->csd);
> > +
> > +	return 0;
> > +}
> > +EXPORT_SYMBOL(smp_call_function_mask);
> 
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists