Re: [PATCH v14 01/13] sk_run_filter: add BPF_S_ANC_SECCOMP_LD_W

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

Le mercredi 14 mars 2012 à 06:12 +0100, Indan Zupancic a écrit :

> Problem is that 'old_size' can be up to 32kB in size and it would be nice
> if that memory could be released. If it isn't, then using JIT increases
> memory usage, while also not accounting it to the socket.
> 

It is accounted for, since jit size is in relation with standard filter
size. Check sock_kmalloc()

Fact we can have a litle underestimation was already the case without
jit, since kmalloc() does a roundup to next power of two.

I dont think this discussion has anything to do with SECCOMP anyway.

These accounting dont need to be 100% precise, we only want a limit to
prevent rogue users from using all kernel memory.

> >
> > If it did, this kind of 'optimization' can actually be not good, because
> > sizeof(*fp) is small enough (less than half cache line size) to trigger
> > a possible false sharing issue. (other part of the cache line could be
> > used to hold a often dirtied object)
> 
> It could avoid this by allocating at least a cache size. But this is a
> problem for all small kmalloc's, isn't it?

Its a problem that was already met on several critical paths :

# find net|xargs grep -n L1_CACHE_BYTES
net/core/dev_addr_lists.c:51:	if (alloc_size < L1_CACHE_BYTES)
net/core/dev_addr_lists.c:52:		alloc_size = L1_CACHE_BYTES;
net/core/net-sysfs.c:586:	    RPS_MAP_SIZE(cpumask_weight(mask)), L1_CACHE_BYTES),
net/core/net-sysfs.c:1111:	    XPS_DEV_MAPS_SIZE, L1_CACHE_BYTES), GFP_KERNEL);
net/ipv6/ip6_fib.c:1612:	size = max_t(size_t, size, L1_CACHE_BYTES);
net/ipv4/fib_frontend.c:1049:	size = max_t(size_t, size, L1_CACHE_BYTES);