Messages in this thread | | | Date | Tue, 14 Apr 2015 15:57:43 +0200 | From | Denys Vlasenko <> | Subject | Re: [PATCH] netns: deinline net_generic() |
| |
On 04/14/2015 03:19 PM, Eric Dumazet wrote: > On Tue, 2015-04-14 at 14:25 +0200, Denys Vlasenko wrote: >> On x86 allyesconfig build: >> The function compiles to 130 bytes of machine code. >> It has 493 callsites. >> Total reduction of vmlinux size: 27906 bytes. >> >> text data bss dec hex filename >> 82447071 22255384 20627456 125329911 77861f7 vmlinux4 >> 82419165 22255384 20627456 125302005 777f4f5 vmlinux5 > > This sounds a big hammer to me. > > These savings probably comes from the BUG_ON() that could simply be > removed. > The second one for sure has no purpose. First one looks defensive. > > For a typical (non allyesconfig) kernel, net_generic() would translate > to : > > return net->gen[id - 1]
My allyesconfig, with BUG_ON's commented out:
void *net_generic(const struct net *net, int id) { struct net_generic *ng; void *ptr;
rcu_read_lock(); ng = rcu_dereference(net->gen); // BUG_ON(id == 0 || id > ng->len); ptr = ng->ptr[id - 1]; rcu_read_unlock();
// BUG_ON(!ptr); return ptr; }
results in the following assembly:
net_generic: call __fentry__ pushq %rbp # movq %rsp, %rbp #, pushq %r12 # movl %esi, %r12d # id, id pushq %rbx # movq %rdi, %rbx # net, net call rcu_read_lock # movq 4784(%rbx), %rbx # MEM[(struct net_generic * const volatile *)net_2(D) + 4784B], _________p1 call debug_lockdep_rcu_enabled # testl %eax, %eax # D.48937 je .L93 #, cmpb $0, __warned.47396(%rip) #, __warned jne .L93 #, call rcu_read_lock_held # testl %eax, %eax # D.48944 jne .L93 #, movq $.LC6, %rdx #, movl $51, %esi #, movq $.LC0, %rdi #, movb $1, __warned.47396(%rip) #, __warned call lockdep_rcu_suspicious # .L93: decl %r12d # tmp68 movslq %r12d, %r12 # tmp68, tmp69 movq 24(%rbx,%r12,8), %rbx # _________p1_4->ptr, ptr call rcu_read_unlock # movq %rbx, %rax # ptr, popq %rbx # popq %r12 # popq %rbp # ret
This is still 112 bytes of code.
| |