Re: [PATCH] kretprobe: avoid re-registration of the same kretprobe earlier

[Masami Hiramatsu <mhiramat@kernel.org>]

On Mon, 21 Dec 2020 21:31:42 +0800
"Wangshaobo (bobo)" <bobo.shaobowang@huawei.com> wrote:

> Hi steven, Masami,
> We have encountered a problem, when we attempted to use steven's suggestion as following,
> 
> >>> If you call this here, you must make sure kprobe_addr() is called on rp->kp.
> >>> But if kretprobe_blacklist_size == 0, kprobe_addr() is not called before
> >>> this check. So it should be in between kprobe_on_func_entry() and
> >>> kretprobe_blacklist_size check, like this
> >>>
> >>> 	if (!kprobe_on_func_entry(rp->kp.addr, rp->kp.symbol_name, rp->kp.offset))
> >>> 		return -EINVAL;
> >>>
> >>> 	addr = kprobe_addr(&rp->kp);
> >>> 	if (IS_ERR(addr))
> >>> 		return PTR_ERR(addr);
> >>> 	rp->kp.addr = addr;
> 
> //there exists no-atomic operation risk, we should not modify any rp->kp's information, not all arch ensure atomic operation here.
> 
> >>>
> >>> 	ret = check_kprobe_rereg(&rp->kp);
> >>> 	if (WARN_ON(ret))
> >>> 		return ret;
> >>>
> >>>           if (kretprobe_blacklist_size) {
> >>> 		for (i = 0; > > +	ret = check_kprobe_rereg(&rp->kp);
> 
> it returns failure from register_kprobe() end called by register_kretprobe() when
> we registered a kretprobe through .symbol_name at first time(through .addr is OK),
> kprobe_addr() called at the begaining of register_kprobe() will recheck and
> failed at following place because at this time we symbol_name is not NULL and addr is also.

Good catch! Yes, it will reject if both kp->addr and kp->symbol are set.

> 
>    static kprobe_opcode_t *_kprobe_addr(const char *symbol_name,
>                           unsigned int offset)
>     {
>           if ((symbol_name && addr) || (!symbol_name && !addr))  //we failed here
> 
> 
> So we attempted to move this sentence rp->kp.addr = addr to __get_valid_kprobe() like this to
> avoid explict usage of rp->kp.addr = addr in register_kretprobe().
> 
> diff --git a/kernel/kprobes.c b/kernel/kprobes.c
> index dd5821f753e6..ea014779edfe 100644
> --- a/kernel/kprobes.c
> +++ b/kernel/kprobes.c
> @@ -1502,10 +1502,15 @@ static kprobe_opcode_t *kprobe_addr(struct kprobe *p)
>   static struct kprobe *__get_valid_kprobe(struct kprobe *p)
>   {
>          struct kprobe *ap, *list_p;
> +       void *addr;
> 
>          lockdep_assert_held(&kprobe_mutex);
> 
> -       ap = get_kprobe(p->addr);
> +       addr = kprobe_addr(p);
> +       if (IS_ERR(addr))
> +               return NULL;
> +
> +       ap = get_kprobe(addr);
>          if (unlikely(!ap))
>                  return NULL;
> 
> But it also failed when we second time attempted to register a same kretprobe, it is also
> becasue symbol_name and addr is not NULL when we used __get_valid_kprobe().

What the "second time" means? If you reuse the kretprobe (and kprobe) you must
reset (cleanup) the kp->addr or kp->symbol_name. That is the initial state.
I think the API should not allow users to enter inconsistent information.

> 
> So it seems has no idea expect for modifying _kprobe_addr() like following this, the reason is that
> the patch 0bd476e6c671 ("kallsyms: unexport kallsyms_lookup_name() and kallsyms_on_each_symbol()")
> has telled us we'd better use symbol name to register but not address anymore.
> 
> -static kprobe_opcode_t *_kprobe_addr(kprobe_opcode_t *addr,
> -                       const char *symbol_name, unsigned int offset)
> +static kprobe_opcode_t *_kprobe_addr(const char *symbol_name,
> +                       unsigned int offset)
>   {
> -       if ((symbol_name && addr) || (!symbol_name && !addr))
> +       kprobe_opcode_t *addr;
> +       if (!symbol_name)
>                  goto invalid;

No, there are cases that the user will set only kp->addr, but no kp->symbol_name.

> 
> For us, this modification has not caused a big impact on other modules, only expects a little
> influence on bpf from calling trace_kprobe_on_func_entry(), it can not use addr to fill in
> rp.kp in struct trace_event_call anymore.
> 
> So i want to know your views, and i will resend this patch soon.

OK, I think it is simpler to check the rp->kp.addr && rp->kp.symbol_name
because it is not allowed (it can lead inconsistent setting).

How about this code? Is this work for you?

diff --git a/kernel/kprobes.c b/kernel/kprobes.c
index 41fdbb7953c6..73500be564be 100644
--- a/kernel/kprobes.c
+++ b/kernel/kprobes.c
@@ -2103,6 +2103,14 @@ int register_kretprobe(struct kretprobe *rp)
        int i;
        void *addr;
 
+       /* It is not allowed to specify addr and symbol_name at the same time */
+       if (rp->kp.addr && rp->kp.symbol_name)
+               return -EINVAL;
+
+       /* If only rp->kp.addr is specified, check reregistering kprobes */
+       if (rp->kp.addr && check_kprobe_rereg(&rp->kp))
+               return -EINVAL;
+
        if (!kprobe_on_func_entry(rp->kp.addr, rp->kp.symbol_name, rp->kp.offset))
                return -EINVAL;
 

Thank you,

-- 
Masami Hiramatsu <mhiramat@kernel.org>