Re: [BUG] 2.6.37-rc5 Memory leak in net/ipv4/udp.c

[Eric Dumazet <eric.dumazet@gmail.com>]

Le vendredi 17 décembre 2010 à 12:11 +0100, Lothar Waßmann a écrit :
> Hi,
> 
> Eric Dumazet writes:
> > Le vendredi 17 décembre 2010 à 11:18 +0100, Lothar Waßmann a écrit :
> > > The offending code in net/ipv4/udp.c is:
> > > |void __init udp_table_init(struct udp_table *table, const char *name)
> > > |{
> > > |	unsigned int i;
> > > |
> > > |	if (!CONFIG_BASE_SMALL)
> > > |		table->hash = alloc_large_system_hash(name,
> > > |			2 * sizeof(struct udp_hslot),
> > > |			uhash_entries,
> > > |			21, /* one slot per 2 MB */
> > > |			0,
> > > |			&table->log,
> > > |			&table->mask,
> > > |			64 * 1024);
> > > |	/*
> > > |	 * Make sure hash table has the minimum size
> > > |	 */
> > > |	if (CONFIG_BASE_SMALL || table->mask < UDP_HTABLE_SIZE_MIN - 1) {
> > > |		table->hash = kmalloc(UDP_HTABLE_SIZE_MIN *
> > > |				      2 * sizeof(struct udp_hslot), GFP_KERNEL);
> > > In case of !CONFIG_BASE_SMALL and 'table->mask < UDP_HTABLE_SIZE_MIN - 1)'
> > > the memory allocated in the previous if clause becomes inacessible!
> > > 
> > > Shouldn't this be:
> > > |	if (!CONFIG_BASE_SMALL && table->mask >= UDP_HTABLE_SIZE_MIN - 1) {
> > > |		table->hash = alloc_large_system_hash(name,
> > > |			2 * sizeof(struct udp_hslot),
> > > |			uhash_entries,
> > > |			21, /* one slot per 2 MB */
> > > |			0,
> > > |			&table->log,
> > > |			&table->mask,
> > > |			64 * 1024);
> > > |	} else {
> > > |		table->hash = kmalloc(UDP_HTABLE_SIZE_MIN *
> > > |				      2 * sizeof(struct udp_hslot), GFP_KERNEL);
> > > [...]
> > > 
> > 
> > Nothing we can do about it, there is no API to reverse the
> > alloc_large_system_hash() effect. We could call kmemleak api to at least
> > avoid this false alarm.
> > 
> Do you have to call it at all in case of table->mask < UDP_HTABLE_SIZE_MIN - 1?
> 

We call alloc_large_system_hash() asking it to size the table _itself_.
We give some hints : 

- How many slots per MB of avail memory.
- An upper limit (64*1024 slots because we only handle 65536 udp ports)
- but not a lower limit (not available in the API)

Problem is in your case, alloc_large_system_hash() allocates a very
small area. Then we catch the problem, seeing table->mask is too small
for our needs. We prefer to 'lost' this too small memory than crashing
kernel later.


> > We really want a minimum size for the UDP hash table, because our algos
> > depend on this.
> > 
> I can't see why this could not be achieved by doing _either_
> alloc_large_system_hash() _OR_ kmalloc() as stated above, but not
> both.

We definitly want alloc_large_system_hash() for the general case
(nice NUMA spread, while kmalloc() would allocate the hash table on a
single memory node. Not so nice)

One way to handle this problem would be to add a new parameter to
alloc_large_system_hash() to specify a lower limit.
alloc_large_system_hash() would not even try to allocate a too small
array.

Is your system so small ?