On Tue, Apr 19, 2022 at 9:32 AM Alexei Starovoitov
<alexei.starovoitov@xxxxxxxxx> wrote:
>
> On Tue, Apr 19, 2022 at 9:20 AM Alexei Starovoitov
> <alexei.starovoitov@xxxxxxxxx> wrote:
> >
> > On Tue, Apr 19, 2022 at 8:42 AM <sdf@xxxxxxxxxx> wrote:
> > >
> > > On 04/18, Alexei Starovoitov wrote:
> > > > On Mon, Apr 18, 2022 at 9:50 AM <sdf@xxxxxxxxxx> wrote:
> > > > >
> > > > > On 04/16, Alexei Starovoitov wrote:
> > > > > > On Thu, Apr 14, 2022 at 9:12 AM Stanislav Fomichev <sdf@xxxxxxxxxx>
> > > > wrote:
> > > > > > > +static int
> > > > > > > +bpf_prog_run_array_cg_flags(const struct cgroup_bpf *cgrp,
> > > > > > > + enum cgroup_bpf_attach_type atype,
> > > > > > > + const void *ctx, bpf_prog_run_fn
> > > > run_prog,
> > > > > > > + int retval, u32 *ret_flags)
> > > > > > > +{
> > > > > > > + const struct bpf_prog_array_item *item;
> > > > > > > + const struct bpf_prog *prog;
> > > > > > > + const struct bpf_prog_array *array;
> > > > > > > + struct bpf_run_ctx *old_run_ctx;
> > > > > > > + struct bpf_cg_run_ctx run_ctx;
> > > > > > > + u32 func_ret;
> > > > > > > +
> > > > > > > + run_ctx.retval = retval;
> > > > > > > + migrate_disable();
> > > > > > > + rcu_read_lock();
> > > > > > > + array = rcu_dereference(cgrp->effective[atype]);
> > > > > > > + item = &array->items[0];
> > > > > > > + old_run_ctx = bpf_set_run_ctx(&run_ctx.run_ctx);
> > > > > > > + while ((prog = READ_ONCE(item->prog))) {
> > > > > > > + run_ctx.prog_item = item;
> > > > > > > + func_ret = run_prog(prog, ctx);
> > > > > > ...
> > > > > > > + ret = bpf_prog_run_array_cg(&cgrp->bpf, CGROUP_GETSOCKOPT,
> > > > > > > &ctx, bpf_prog_run, retval);
> > > > >
> > > > > > Did you check the asm that bpf_prog_run gets inlined
> > > > > > after being passed as a pointer to a function?
> > > > > > Crossing fingers... I suspect not every compiler can do that :(
> > > > > > De-virtualization optimization used to be tricky.
> > > > >
> > > > > No, I didn't, but looking at it right now, both gcc and clang
> > > > > seem to be doing inlining all way up to bpf_dispatcher_nop_func.
> > > > >
> > > > > clang:
> > > > >
> > > > > 0000000000001750 <__cgroup_bpf_run_filter_sock_addr>:
> > > > > __cgroup_bpf_run_filter_sock_addr():
> > > > > ./kernel/bpf/cgroup.c:1226
> > > > > int __cgroup_bpf_run_filter_sock_addr(struct sock *sk,
> > > > > struct sockaddr *uaddr,
> > > > > enum cgroup_bpf_attach_type atype,
> > > > > void *t_ctx,
> > > > > u32 *flags)
> > > > > {
> > > > >
> > > > > ...
> > > > >
> > > > > ./include/linux/filter.h:628
> > > > > ret = dfunc(ctx, prog->insnsi, prog->bpf_func);
> > > > > 1980: 49 8d 75 48 lea 0x48(%r13),%rsi
> > > > > bpf_dispatcher_nop_func():
> > > > > ./include/linux/bpf.h:804
> > > > > return bpf_func(ctx, insnsi);
> > > > > 1984: 4c 89 f7 mov %r14,%rdi
> > > > > 1987: 41 ff 55 30 call *0x30(%r13)
> > > > > 198b: 89 c3 mov %eax,%ebx
> > > > >
> > > > > gcc (w/retpoline):
> > > > >
> > > > > 0000000000001110 <__cgroup_bpf_run_filter_sock_addr>:
> > > > > __cgroup_bpf_run_filter_sock_addr():
> > > > > kernel/bpf/cgroup.c:1226
> > > > > {
> > > > >
> > > > > ...
> > > > >
> > > > > ./include/linux/filter.h:628
> > > > > ret = dfunc(ctx, prog->insnsi, prog->bpf_func);
> > > > > 11c5: 49 8d 75 48 lea 0x48(%r13),%rsi
> > > > > bpf_dispatcher_nop_func():
> > > > > ./include/linux/bpf.h:804
> > > > > 11c9: 48 8d 7c 24 10 lea 0x10(%rsp),%rdi
> > > > > 11ce: e8 00 00 00 00 call 11d3
> > > > > <__cgroup_bpf_run_filter_sock_addr+0xc3>
> > > > > 11cf: R_X86_64_PLT32
> > > > __x86_indirect_thunk_rax-0x4
> > > > > 11d3: 89 c3 mov %eax,%ebx
> > >
> > > > Hmm. I'm not sure how you've got this asm.
> > > > Here is what I see with gcc 8 and gcc 10:
> > > > bpf_prog_run_array_cg:
> > > > ...
> > > > movq %rcx, %r12 # run_prog, run_prog
> > > > ...
> > > > # ../kernel/bpf/cgroup.c:77: run_ctx.prog_item = item;
> > > > movq %rbx, (%rsp) # item, run_ctx.prog_item
> > > > # ../kernel/bpf/cgroup.c:78: if (!run_prog(prog, ctx) &&
> > > > !IS_ERR_VALUE((long)run_ctx.retval))
> > > > movq %rbp, %rsi # ctx,
> > > > call *%r12 # run_prog
> > >
> > > > __cgroup_bpf_run_filter_sk:
> > > > movq $bpf_prog_run, %rcx #,
> > > > # ../kernel/bpf/cgroup.c:1202: return
> > > > bpf_prog_run_array_cg(&cgrp->bpf, atype, sk, bpf_prog_run, 0);
> > > > leaq 1520(%rax), %rdi #, tmp92
> > > > # ../kernel/bpf/cgroup.c:1202: return
> > > > bpf_prog_run_array_cg(&cgrp->bpf, atype, sk, bpf_prog_run, 0);
> > > > jmp bpf_prog_run_array_cg #
> > >
> > > > This is without kasan, lockdep and all debug configs are off.
> > >
> > > > So the generated code is pretty bad as I predicted :(
> > >
> > > > So I'm afraid this approach is no go.
> > >
> > > I've retested again and it still unrolls it for me on gcc 11 :-/
> > > Anyway, I guess we have two options:
> > >
> > > 1. Go back to defines.
> > > 2. Don't pass a ptr to func, but pass an enum which indicates whether
> > > to use bpf_prog_run or __bpf_prog_run_save_cb. Seems like in this
> > > case the compiler shouldn't have any trouble unwrapping it?
> > >
> > > I'll prototype and send (2). If it won't work out we can always get back
> > > to (1).
> >
> > Going back to defines is probably not necessary.
> > Could you try moving bpf_prog_run_array_cg*() back to .h
> > and use static __always_inline ?
>
> Actually below was enough for gcc 8 and 10:
> -static int
> +static __always_inline int
> bpf_prog_run_array_cg_flags(const struct cgroup_bpf *cgrp,
> enum cgroup_bpf_attach_type atype,
> const void *ctx, bpf_prog_run_fn run_prog,
> @@ -55,7 +55,7 @@ bpf_prog_run_array_cg_flags(const struct cgroup_bpf *cgrp,
> return run_ctx.retval;
> }
>
> -static int
> +static __always_inline int
> bpf_prog_run_array_cg(const struct cgroup_bpf *cgrp,
>
> we can keep them in .c and generated code looks good.
>
> I can apply it with the above change.
> wdyt?
Sure, let's go with that if it works! On my side, I managed to get the
same bad results on gcc-8; moving them to bpf-cgroup.h with
__always_inline seems to fix it. But if we can keep them in .c, that
looks even better.
