Date:	Thu, 13 Nov 2008 13:18:50 -0800
From:	Andrew Morton <akpm@...ux-foundation.org>
To:	Yinghai Lu <yinghai@...nel.org>
Cc:	mingo@...e.hu, tglx@...utronix.de, hpa@...or.com,
	linux-kernel@...r.kernel.org, travis@....com
Subject: Re: [PATCH] sparse_irq aka dyn_irq v13

On Thu, 13 Nov 2008 12:16:56 -0800
Yinghai Lu <yinghai@...nel.org> wrote:

> From: Yinghai Lu <yinghai@...nel.org>
> Subject: sparseirq v13

My overall view on this is that it takes some of the kernel's most
fragile and most problem-dense code and makes it much more complex, and
by adding new configuration options it significantly worsens our
testing coverage.

The patch is HHHHHUUUUUUUUUUGGGGGEEE!  Did it really need to be a
single megapatch?

Other architectures want (or have) sparse interrupts.  Are those guys
paying attention here?

I don't have a clue what all this does.  I hope those who will work on
this code are sufficiently familiar with it all to be able to maintain
it when there are close to zero comments in some of our most tricky and
problem-prone code.

>
> ...
>
> +config SPARSE_IRQ
> +	bool "Support sparse irq numbering"
> +	depends on PCI_MSI || HT_IRQ
> +	default y
> +	help
> +	  This enables support for sparse irq, esp for msi/msi-x. the irq
> +	  number will be bus/dev/fn + 12bit. You may need if you have lots of
> +	  cards supports msi-x installed.
> +
> +	  If you don't know what to do here, say Y.
> +
> +config MOVE_IRQ_DESC
> +	bool "Move irq desc when changing irq smp_affinity"
> +	depends on SPARSE_IRQ && SMP
> +	default y
> +	help
> +	  This enables moving irq_desc to cpu/node that irq will use handled.
> +
> +	  If you don't know what to do here, say Y.

Do these reeeealy have to exist?  How are users to know which to
choose?  Which option will distros choose and why did we make them have
to decide?

>
> ...
>
> +{
> +	struct irq_pin_list *pin;
> +	int node;
> +
> +	if (cpu < 0)
> +		cpu = smp_processor_id();
> +	node = cpu_to_node(cpu);
> +
> +	pin = kzalloc_node(sizeof(*pin), GFP_KERNEL, node);

It's a bug to call smp_processor_id() from preemptible code and it's a
bug to use GFP_KERNEL in non-preemptible code.  How can this be?

> +	printk(KERN_DEBUG "  alloc irq_2_pin on cpu %d node %d\n", cpu, node);
> +
> +	return pin;
> +}
>
>
> ...
>
> -static struct irq_cfg *irq_cfg_alloc(unsigned int irq)
> +static struct irq_cfg *get_one_free_irq_cfg(int cpu)
>  {
> -	return irq_cfg(irq);
> +	struct irq_cfg *cfg;
> +	int node;
> +
> +	if (cpu < 0)
> +		cpu = smp_processor_id();
> +	node = cpu_to_node(cpu);
> +
> +	cfg = kzalloc_node(sizeof(*cfg), GFP_KERNEL, node);

See above.

> +	printk(KERN_DEBUG "  alloc irq_cfg on cpu %d node %d\n", cpu, node);
> +
> +	return cfg;
>  }

So all callers of this function must test the return value and if it is
NULL, take appropriate action.

That is something which should have been documented in this function's
interface description.  Only that doesn't exist.

The one caller which I checked (arch_init_copy_chip_data()) fails to
check for this and will oops.

>
> ...
>
> +static void set_extra_move_desc(struct irq_desc *desc, cpumask_t mask)
> +{
> +	struct irq_cfg *cfg = desc->chip_data;
> +
> +	if (!cfg->move_in_progress) {
> +		/* it means that domain is not changed */
> +		cpumask_t tmp;
> +
> +		cpus_and(tmp, desc->affinity, mask);
> +		if (cpus_empty(tmp))
> +			cfg->move_desc_in_progress_in_same_domain = 1;
> +	}

Aren't we trying to avoid on-stack cpumask_t's?

I'd have though that this one could be eliminated via the use of
cpus_intersects()?

>  }
> +#endif
>
> ...
>
>  static struct irq_chip lapic_chip __read_mostly = {
>  	.name		= "local-APIC",
>  	.mask		= mask_lapic_irq,
> @@ -2574,7 +2834,9 @@ int timer_through_8259 __initdata;
>   */
>  static inline void __init check_timer(void)

An inlined __init function makes little sense.

>
> ...
>
> @@ -3306,10 +3590,13 @@ static void dmar_msi_set_affinity(unsign
>  	if (cpus_empty(tmp))
>  		return;

`tmp' is always a bad choice of identifier.

>
> ...
>
> --- linux-2.6.orig/arch/x86/mm/init_32.c
> +++ linux-2.6/arch/x86/mm/init_32.c
> @@ -66,6 +66,7 @@ static unsigned long __meminitdata table
>  static unsigned long __meminitdata table_top;
>  
>  static int __initdata after_init_bootmem;
> +int after_bootmem;

This isn't a very well-chosen identifier for an x86-specific global.

>
> ...
>
> @@ -98,6 +126,7 @@ int __ht_create_irq(struct pci_dev *dev,
>  	int max_irq;
>  	int pos;
>  	int irq;
> +	unsigned int irq_want;
>  
>  	pos = pci_find_ht_capability(dev, HT_CAPTYPE_IRQ);
>  	if (!pos)
> @@ -125,7 +154,12 @@ int __ht_create_irq(struct pci_dev *dev,
>  	cfg->msg.address_lo = 0xffffffff;
>  	cfg->msg.address_hi = 0xffffffff;
>  
> +	irq_want = build_irq_for_pci_dev(dev);
> +#ifdef CONFIG_SPARSE_IRQ
> +	irq = create_irq_nr(irq_want + idx);
> +#else
>  	irq = create_irq();
> +#endif

irq_want is unused if CONFIG_SPARSE_IRQ=n.

>  	if (irq <= 0) {
>  		kfree(cfg);
> Index: linux-2.6/drivers/pci/intr_remapping.c
> ===================================================================
> --- linux-2.6.orig/drivers/pci/intr_remapping.c
> +++ linux-2.6/drivers/pci/intr_remapping.c
> @@ -19,17 +19,73 @@ struct irq_2_iommu {
>  	u8  irte_mask;
>  };
>  
> -static struct irq_2_iommu irq_2_iommuX[NR_IRQS];
> +#ifdef CONFIG_SPARSE_IRQ
> +static struct irq_2_iommu *get_one_free_irq_2_iommu(int cpu)
> +{
> +	struct irq_2_iommu *iommu;
> +	int node;
> +
> +	if (cpu < 0)
> +		cpu = smp_processor_id();
> +	node = cpu_to_node(cpu);
> +
> +	iommu = kzalloc_node(sizeof(*iommu), GFP_KERNEL, node);

See above.

> +	printk(KERN_DEBUG "alloc irq_2_iommu on cpu %d node %d\n", cpu, node);
> +
> +	return iommu;
> +}
>  
>
> ...
>
> --- linux-2.6.orig/fs/proc/stat.c
> +++ linux-2.6/fs/proc/stat.c
> @@ -27,6 +27,9 @@ static int show_stat(struct seq_file *p,
>  	u64 sum = 0;
>  	struct timespec boottime;
>  	unsigned int per_irq_sum;
> +#ifdef CONFIG_GENERIC_HARDIRQS
> +	struct irq_desc *desc;
> +#endif
>  
>  	user = nice = system = idle = iowait =
>  		irq = softirq = steal = cputime64_zero;
> @@ -44,10 +47,9 @@ static int show_stat(struct seq_file *p,
>  		softirq = cputime64_add(softirq, kstat_cpu(i).cpustat.softirq);
>  		steal = cputime64_add(steal, kstat_cpu(i).cpustat.steal);
>  		guest = cputime64_add(guest, kstat_cpu(i).cpustat.guest);
> -
> -		for_each_irq_nr(j)
> +		for_each_irq_desc(j, desc) {

This won't compile if CONFIG_GENERIC_HARDIRQS=n, I suspect.

>  			sum += kstat_irqs_cpu(j, i);
> -
> +		} end_for_each_irq_desc();
>  		sum += arch_irq_stat_cpu(i);
>  	}
>  	sum += arch_irq_stat();
> @@ -90,14 +92,17 @@ static int show_stat(struct seq_file *p,
>  	seq_printf(p, "intr %llu", (unsigned long long)sum);
>  
>  	/* sum again ? it could be updated? */
> -	for_each_irq_nr(j) {
> +	for_each_irq_desc(j, desc) {

ditto.

>
> ...
>
> +#define end_for_each_irq_desc()
> +#endif
> +
> +#define desc_chip_ack(irq, descp) desc->chip->ack(irq)
> +#define desc_chip_mask(irq, descp) desc->chip->mask(irq)
> +#define desc_chip_mask_ack(irq, descp) desc->chip->mask_ack(irq)
> +#define desc_chip_unmask(irq, descp) desc->chip->unmask(irq)
> +#define desc_chip_eoi(irq, descp) desc->chip->eoi(irq)

Was it necessary to implement these as macros?

>
> ...
>
> @@ -49,6 +56,299 @@ void handle_bad_irq(unsigned int irq, st
>  int nr_irqs = NR_IRQS;
>  EXPORT_SYMBOL_GPL(nr_irqs);
>  
> +void __init __attribute__((weak)) arch_early_irq_init_work(void)
> +{
> +}
> +
> +#ifdef CONFIG_SPARSE_IRQ
> +static struct irq_desc irq_desc_init = {
> +	.irq	    = -1U,

Plain old `-1' would be better here.  It works in all cases and the
reader doesn't need to go and check that this field really is an
unsigned int and it won't need editing if that field gets changed to
long.

> +	.status	    = IRQ_DISABLED,
> +	.chip	    = &no_irq_chip,
> +	.handle_irq = handle_bad_irq,
> +	.depth      = 1,
> +	.lock       = __SPIN_LOCK_UNLOCKED(irq_desc_init.lock),
> +#ifdef CONFIG_SMP
> +	.affinity   = CPU_MASK_ALL
> +#endif
> +};
> +
> +static void init_kstat_irqs(struct irq_desc *desc, int cpu, int nr)
> +{
> +	unsigned long bytes;
> +	char *ptr;
> +	int node;
> +
> +	/* Compute how many bytes we need per irq and allocate them */
> +	bytes = nr * sizeof(unsigned int);
> +
> +	if (cpu < 0)
> +		cpu = smp_processor_id();
> +
> +	node = cpu_to_node(cpu);
> +	ptr = kzalloc_node(bytes, GFP_KERNEL, node);

See above.

> +	printk(KERN_DEBUG "  alloc kstat_irqs on cpu %d node %d\n", cpu, node);
> +
> +	desc->kstat_irqs = (unsigned int *)ptr;
> +}
> +
>
> ...
>
> +#endif
> +/*
> + * Protect the sparse_irqs_free freelist:
> + */
> +static DEFINE_SPINLOCK(sparse_irq_lock);
> +LIST_HEAD(sparse_irqs_head);

It's strange that the list is global and is accessed from other .c
files, but the lock which protects it is static.

> +/*
> + * The sparse irqs are in a hash-table as well, for fast lookup:
> + */
> +#define SPARSEIRQHASH_BITS          (13 - 1)
> +#define SPARSEIRQHASH_SIZE          (1UL << SPARSEIRQHASH_BITS)
> +#define __sparseirqhashfn(key)      hash_long((unsigned long)key, SPARSEIRQHASH_BITS)
> +#define sparseirqhashentry(key)     (sparseirqhash_table + __sparseirqhashfn((key)))

Why implement these via macros?

> +static struct list_head sparseirqhash_table[SPARSEIRQHASH_SIZE];
> +
> +static struct irq_desc irq_desc_legacy[NR_IRQS_LEGACY] __cacheline_aligned_in_smp = {
> +	[0 ... NR_IRQS_LEGACY-1] = {
> +		.irq	    = -1U,
> +		.status	    = IRQ_DISABLED,
> +		.chip	    = &no_irq_chip,
> +		.handle_irq = handle_bad_irq,
> +		.depth	    = 1,
> +		.lock	    = __SPIN_LOCK_UNLOCKED(irq_desc_init.lock),
> +#ifdef CONFIG_SMP
> +		.affinity   = CPU_MASK_ALL
> +#endif
> +	}
> +};
> +
> +/* FIXME: use bootmem alloc ...*/
> +static unsigned int kstat_irqs_legacy[NR_IRQS_LEGACY][NR_CPUS];

Do these need to be 32-bit?  Maybe they'll fit in 16-bit, dunno.

> +void __init early_irq_init_work(void)

The use of "_work" implies that this function is invoked by
schedule_work().  But it isn't.

> +{
> +	struct irq_desc *desc;
> +	int legacy_count;
> +	int i;
> +
> +	/* init_work to init list for sparseirq */
> +	for (i = 0; i < SPARSEIRQHASH_SIZE; i++)
> +		INIT_LIST_HEAD(sparseirqhash_table + i);
> +
> +	desc = irq_desc_legacy;
> +	legacy_count = ARRAY_SIZE(irq_desc_legacy);
> +
> +	for (i = 0; i < legacy_count; i++) {
> +		struct list_head *hash_head;
> +
> +		hash_head = sparseirqhashentry(i);
> +		desc[i].irq = i;
> +		desc[i].kstat_irqs = kstat_irqs_legacy[i];
> +		list_add_tail(&desc[i].hash_entry, hash_head);
> +		list_add_tail(&desc[i].list, &sparse_irqs_head);
> +	}
> +
> +	arch_early_irq_init_work();
> +}
> +
>
> ...
>
> +struct irq_desc *irq_to_desc_alloc_cpu(unsigned int irq, int cpu)
> +{
> +	struct irq_desc *desc;
> +	struct list_head *hash_head;
> +	unsigned long flags;
> +	int node;
> +
> +	desc = irq_to_desc(irq);
> +	if (desc)
> +		return desc;
> +
> +	hash_head = sparseirqhashentry(irq);
> +
> +	spin_lock_irqsave(&sparse_irq_lock, flags);
> +
> +	/*
> +	 * We have to do the hash-walk again, to avoid races
> +	 * with another CPU:
> +	 */
> +	list_for_each_entry(desc, hash_head, hash_entry)
> +		if (desc->irq == irq)
> +			goto out_unlock;
> +
> +	if (cpu < 0)
> +		cpu = smp_processor_id();
> +
> +	node = cpu_to_node(cpu);
> +	desc = kzalloc_node(sizeof(*desc), GFP_KERNEL, node);

Oh for gawd's sake.  PLEASE read Documentation/SubmitChecklist. 
Carefully.  We've already discussed this.

You cannot do a GFP_KERNEL allocation under spin_lock_irqsave(). 

> +	printk(KERN_DEBUG "  alloc irq_desc for %d aka %#x on cpu %d node %d\n",
> +		 irq, irq, cpu, node);
> +	init_one_irq_desc(irq, desc, cpu);
> +
> +	/*
> +	 * We use RCU's safe list-add method to make
> +	 * parallel walking of the hash-list safe:
> +	 */
> +	list_add_tail_rcu(&desc->hash_entry, hash_head);
> +	/*
> +	 * Add it to the global list:
> +	 */
> +	list_add_tail_rcu(&desc->list, &sparse_irqs_head);
> +
> +out_unlock:
> +	spin_unlock_irqrestore(&sparse_irq_lock, flags);
> +
> +	return desc;
> +}
> +
> +struct irq_desc *irq_to_desc_alloc(unsigned int irq)
> +{
> +	return irq_to_desc_alloc_cpu(irq, -1);
> +}
> +
> +#ifdef CONFIG_MOVE_IRQ_DESC
> +static struct irq_desc *__real_move_irq_desc(struct irq_desc *old_desc,
> +						int cpu)
> +{
> +	struct irq_desc *desc;
> +	unsigned int irq;
> +	struct list_head *hash_head;
> +	unsigned long flags;
> +	int node;
> +
> +	irq = old_desc->irq;
> +
> +	hash_head = sparseirqhashentry(irq);
> +
> +	spin_lock_irqsave(&sparse_irq_lock, flags);
> +	/*
> +	 * We have to do the hash-walk again, to avoid races
> +	 * with another CPU:
> +	 */
> +	list_for_each_entry(desc, hash_head, hash_entry)
> +		if (desc->irq == irq && old_desc != desc)
> +			goto out_unlock;
> +
> +	if (cpu < 0)
> +		cpu = smp_processor_id();
> +
> +	node = cpu_to_node(cpu);
> +	desc = kzalloc_node(sizeof(*desc), GFP_KERNEL, node);

Ditto.

Also, the return value from the memory allocation attempt is not
checked.

> +	printk(KERN_DEBUG "  move irq_desc for %d aka %#x to cpu %d node %d\n",
> +		 irq, irq, cpu, node);
> +
> +	init_copy_one_irq_desc(irq, old_desc, desc, cpu);
> +
> +	list_replace_rcu(&old_desc->hash_entry, &desc->hash_entry);
> +	list_replace_rcu(&old_desc->list, &desc->list);
> +
> +	/* free the old one */
> +	free_one_irq_desc(old_desc);
> +	kfree(old_desc);
> +
> +out_unlock:
> +	spin_unlock_irqrestore(&sparse_irq_lock, flags);
> +
> +	return desc;
> +}
> +
>
> ...
>
> --- linux-2.6.orig/init/main.c
> +++ linux-2.6/init/main.c
> @@ -542,6 +542,15 @@ void __init __weak thread_info_cache_ini
>  {
>  }
>  
> +void __init __attribute__((weak)) arch_early_irq_init_work(void)
> +{
> +}
> +
> +void __init __attribute__((weak)) early_irq_init_work(void)
> +{
> +	arch_early_irq_init_work();
> +}

Please use __weak

>
> ...
>


--
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