(Found?) Re: Hard Hang with __alloc_pages: 0-order allocation failed (gfp=0x20/1) - Not out of memory

(Found?) Re: Hard Hang with __alloc_pages: 0-order allocation failed (gfp=0x20/1) - Not out of memory

[Roger Larsson]

Hi,

Since I read linux-kernel via achieves, this might be found already - but 
anyway... (CONFIG_SMP problem? Oh, noticed that this is 2.4.21...)

decode "(gfp=0x20/1)"  and you find that
	current->flags is PF_MEMALLOC 
	gfp is __GFP_HIGH (== GFP_ATOMIC)
and the backtrace says we are in_interrupt()

So I guess this is the key line...
	if (current->flags & PF_MEMALLOC && !in_interrupt()) {

* Less than (.min >> 2) of memory left.
* In interrupt
=> Not allowed to allocate anything more

Does the caller understand that repeating requests will not help?

In e1000_main.c
		skb = dev_alloc_skb(adapter->rx_buffer_len + reserve_len);

___skbuff.h___
static inline struct sk_buff *dev_alloc_skb(unsigned int length)
{
	return __dev_alloc_skb(length, GFP_ATOMIC);
}

static inline struct sk_buff *__dev_alloc_skb(unsigned int length,
					      int gfp_mask)
{
	struct sk_buff *skb;

	skb = alloc_skb(length+16, gfp_mask);
	if (skb)
		skb_reserve(skb,16);
	return skb;
}

And the rather big alloc_skb calls kmalloc

void * kmalloc (size_t size, int flags)
{
	cache_sizes_t *csizep = cache_sizes;

	for (; csizep->cs_size; csizep++) {
		if (size > csizep->cs_size)
			continue;
		return __kmem_cache_alloc(flags & GFP_DMA ?
			 csizep->cs_dmacachep : csizep->cs_cachep, flags);
	}
	return NULL;
}

- - - now take a close look at the try_again path - - -

static inline void * __kmem_cache_alloc (kmem_cache_t *cachep, int flags)
{
	unsigned long save_flags;
	void* objp;

	kmem_cache_alloc_head(cachep, flags);
try_again:
	local_irq_save(save_flags);

#ifdef CONFIG_SMP
	{
		cpucache_t *cc = cc_data(cachep);

		if (cc) {
			if (cc->avail) {
				STATS_INC_ALLOCHIT(cachep);
				objp = cc_entry(cc)[--cc->avail];
			} else {
				STATS_INC_ALLOCMISS(cachep);
				objp = kmem_cache_alloc_batch(cachep,cc,flags);
				if (!objp)
					goto alloc_new_slab_nolock;

- - -
alloc_new_slab_nolock:
#endif
	local_irq_restore(save_flags);
	if (kmem_cache_grow(cachep, flags))
		/* Someone may have stolen our objs.  Doesn't matter, we'll
		 * just come back here again.
		 */
		goto try_again;

But kmem_cache_grow will return failed...
	-> kmem_getpages ->  _get_free_pages -> alloc_pages

Or have I missed something?

/RogerL

-- 
Roger Larsson
Skellefteå
Sweden

Re: Hard Hang with __alloc_pages: 0-order allocation failed (gfp=0x20/1) - Not out of memory

[Roger Larsson]

Modifying the header so you find it. And write a better summary...

* When you have little memory left __alloc_pages wont allow
   allocations from interrupt context (even if PF_MEMALLOC)

	if (current->flags & PF_MEMALLOC && !in_interrupt()) {

* But in the callchain there is a loop in __kmem_cache_alloc when running SMP

> static inline void * __kmem_cache_alloc (kmem_cache_t *cachep, int flags)
> {
> 	unsigned long save_flags;
> 	void* objp;
>
> 	kmem_cache_alloc_head(cachep, flags);
> try_again:
> 	local_irq_save(save_flags);
>
> #ifdef CONFIG_SMP
> 	{
> 		cpucache_t *cc = cc_data(cachep);
>
> 		if (cc) {
> 			if (cc->avail) {
> 				STATS_INC_ALLOCHIT(cachep);
> 				objp = cc_entry(cc)[--cc->avail];
> 			} else {
> 				STATS_INC_ALLOCMISS(cachep);
> 				objp = kmem_cache_alloc_batch(cachep,cc,flags);
> 				if (!objp)
> 					goto alloc_new_slab_nolock;
>
> - - - snip - - -
> alloc_new_slab_nolock:
> #endif
> 	local_irq_restore(save_flags);
> 	if (kmem_cache_grow(cachep, flags))
> 		/* Someone may have stolen our objs.  Doesn't matter, we'll
> 		 * just come back here again.
> 		 */
> 		goto try_again;
>
> But kmem_cache_grow will return failed...
> 	-> kmem_getpages ->  _get_free_pages -> alloc_pages

But alloc_pages WILL fail every time... => infinite loop in interrupt
context... It probably should not try again in this case...

	if (!in_interrupt())
		goto try_again;

Or it should check if GFP_WAIT is set... A _busy_ wait is also a wait...

/RogerL

-- 
Roger Larsson
Skellefteå
Sweden

Re: PROBLEM: e100 or e1000 on SMP kernel freeze system (ipx+ncp)

[Roger Larsson]

Might be related to subject

"Re: Hard Hang with __alloc_pages: 0-order allocation failed (gfp=0x20/1) - 
Not out of memory"?

e1000 and SMP...

/RogerL

-- 
Roger Larsson
Skellefteå
Sweden

Re: PROBLEM: e100 or e1000 on SMP kernel freeze system (ipx+ncp)

[Andrew Shirrayev]

> Might be related to subject
>
> "Re: Hard Hang with __alloc_pages: 0-order allocation failed (gfp=0x20/1) -
> Not out of memory"?
Ok, thanks.

My problem very different, but maybe have common source.

My problem can repeat on simple system (some kernel
thread and one bash... 1Gb RAM) and very quiet... :-(

Yes, SMP and e10x required :-)

Infinit loop?

>
> e1000 and SMP...
>
> /RogerL