Re: [PATCH/RFC 17/17] tracing/uprobes: Add @+file_offset fetch method

[Namhyung Kim <namhyung@kernel.org>]

Hi Oleg,

On Wed, 27 Nov 2013 19:55:46 +0100, Oleg Nesterov wrote:
> Hi Namhyung,
>
> I'll certainly try to read (and even apply ;) this series carefully.

Thanks in advance. :)

>
> But let me make a couple of nits right now, even if I do not understand
> this code yet.

Okay.

>
> On 11/27, Namhyung Kim wrote:
>>
>> +		} else if (arg[1] == '+') {
>> +			struct file_offset_fetch_param *foprm;
>> +
>> +			/* kprobes don't support file offsets */
>> +			if (is_kprobe)
>> +				return -EINVAL;
>> +
>> +			ret = kstrtol(arg + 2, 0, &offset);
>> +			if (ret)
>> +				break;
>> +
>> +			foprm = kzalloc(sizeof(*foprm), GFP_KERNEL);
>> +			if (!foprm)
>> +				return -ENOMEM;
>> +
>> +			foprm->tu = priv;
>> +			foprm->offset = offset;
>
> Hmm. I am not sure, but can't we simplify this?
>
> Why do we need this foprm at all? To pass tu/offset obviously. But
> why we need to store this info in fetch_param?
>
> translate_user_vaddr() needs to access utask->vaddr anyway. It seems
> to me it would be more clean to do the following:
>
> 	1. Add
> 		struct xxx {
> 			struct trace_uprobe *tu;
> 			unsigned long bp_addr;
> 		};
>
> 	   in trace_uprobe.c.
>
> 	2. Add
>
> 		struct xxx info = {
> 			.tu = tu,
> 			.bp_addr = instruction_pointer(regs);
> 		};
>
> 		current->utask->vaddr = (long)&info;
>
> 	   into uprobe_dispatcher() and uretprobe_dispatcher() (the latter
> 	   should obviously use func instead of instruction_pointer).
>
> 	 3. FETCH_FUNC_NAME(file_offset, type) can do
>
> 	 	struct xxx *info = (void*)current->utask->vaddr;
> 		void *addr = data + info->bp_addr - info->tu->offset;
>
> 	 	return FETCH_FUNC_NAME(memory, type)(regs, aaddr, dest);
>
> 	 4. Now, the only change we need in parse_probe_arg("@") is that
> 	    it should use either FETCH_MTD_memory or FETCH_MTD_file_offset
> 	    depending on arg[0] == '+'.
>
> 	    And we do not need to pass "void *prive" to parse_probe_arg().
>
> What do you think? One again, I can be easily wrong, I didn't read the
> code yet.

You are absolutely right.

I thought we need a fetch_param anyway if we will add support for
cross-fetch later.  But I won't insist it strongly, I can delay it to
later work and make current code simpler if you want. :)

>
>>  static int uprobe_dispatcher(struct uprobe_consumer *con, struct pt_regs *regs)
>>  {
>>  	struct trace_uprobe *tu;
>> +	struct uprobe_task *utask;
>>  	int ret = 0;
>>  
>>  	tu = container_of(con, struct trace_uprobe, consumer);
>>  	tu->nhit++;
>>  
>> +	utask = current->utask;
>> +	if (utask == NULL)
>> +		return UPROBE_HANDLER_REMOVE;
>
> Hmm, why? The previous change ensures ->utask is not NULL? If we hit
> NULL we have a bug, we should not remove this uprobe.

Yes, I just want to be defensive. :)

So do you suggest to add BUG_ON()?  And can I convert or remove a
similar check in uprobes.c:pre_ssout() too?

Thanks,
Namhyung