On Sun, Mar 19, 2017 at 04:13:41PM -0700, Eric Dumazet wrote:
> On Mon, 2017-03-20 at 01:24 +0300, Alexey Dobriyan wrote:
> > Hash size can't negative so "unsigned int" is logically correct.
>
>
> > struct flow_cache_percpu *fcp = per_cpu_ptr(fc->percpu, cpu);
> > - size_t sz = sizeof(struct hlist_head) * flow_cache_hash_size(fc);
> > + unsigned int sz = sizeof(struct hlist_head) * flow_cache_hash_size(fc);
> >
> > if (!fcp->hash_table) {
> > fcp->hash_table = kzalloc_node(sz, GFP_KERNEL,
> > cpu_to_node(cpu));
> > if (!fcp->hash_table) {
> > - pr_err("NET: failed to allocate flow cache sz %zu\n",
> > sz);
> > + pr_err("NET: failed to allocate flow cache sz %u\n",
> > sz);
>
>
> I do not see any improvement here.
>
> What is wrong with size_t exactly ?
REX prefixes and sign extensions, lots of them.
Before:
ffffffff8505b1b5: 41 bd 01 00 00 00 mov r13d,0x1
ffffffff8505b1bb: 41 d3 e5 shl r13d,cl
ffffffff8505b1be: 4d 63 ed movsxd r13,r13d
ffffffff8505b1c1: 49 c1 e5 03 shl r13,0x3
ffffffff8505b1c5: e8 86 28 0a fc call __cpu_to_node
...
ffffffff8505b20b: 4c 89 ee mov rsi,r13
After:
ffffffff8505b1b5: 41 bd 08 00 00 00 mov r13d,0x8
ffffffff8505b1bb: 41 d3 e5 shl r13d,cl
ffffffff8505b1be: e8 8d 28 0a fc call __cpu_to_node
...
ffffffff8505b1c3: 44 89 ef mov edi,r13d
Basically, one can do s/size_t/unsigned int/g across whole networking
stack and nothing will change but the code becomes smaller (including
things like sendmsg() because VFS truncates lengths at INT_MAX).
