Re: [PATCH 0/7] Uprobes: Support SDT markers having reference count (semaphore)
From: Ravi Bangoria
Date: Thu Jun 07 2018 - 22:29:58 EST
Hi Masami,
On 06/08/2018 06:40 AM, Masami Hiramatsu wrote:
> On Wed, 6 Jun 2018 14:03:37 +0530
> Ravi Bangoria <ravi.bangoria@xxxxxxxxxxxxx> wrote:
>
>> Why RFC again:
>>
>> This series is different from earlier versions[1]. Earlier series
>> implemented this feature in trace_uprobe while this has implemented
>> the logic in core uprobe. Few reasons for this:
>> 1. One of the major reason was the deadlock between uprobe_lock and
>> mm->mmap inside trace_uprobe_mmap(). That deadlock was not easy to fix
>> because mm->mmap is not in control of trace_uprobe_mmap() and it has
>> to take uprobe_lock to loop over trace_uprobe list. More details can
>> be found at[2]. With this new approach, there are no deadlocks found
>> so far.
>> 2. Many of the core uprobe function and data-structures needs to be
>> exported to make earlier implementation simple. With this new approach,
>> reference counter logic is been implemented in core uprobe and thus
>> no need to export anything.
>
> I agree with you. Moreover, since uprobe_register/unregister() are
> exported to modules, this enablement would better be implemented
> inside uprobe so that all uprobe users benefit from this.
Sorry, I think you got me wrong. I meant, I don't need to expose all core
uprobe _static_ functions to tarce_uprobe.
Now, about kernel modules, basically uprobe_register() takes three parameters:
inode, offset and consumer.
There is no scope for the reference counter there. So I've created one more
function: uprobe_register_refctr(). But this function is not exported as ABI
to kernel module. i.e. kernel modules still does not have a way to create
uprobe with reference counter. So for kernel modules,
is it fine to change current ABI from
uprobe_register(inode, offset, consumer)
to
uprobe_register(inode, offset, ref_ctr_offset, consumer)
Or I should introduce new function for this:
uprobe_register_refctr(inode, offset, ref_ctr_offset, consumer)
and export it to kernel module?
What's your suggestion?
[...]
>>
>> - This patches still has one issue. If there are multiple instances of
>> same application running and user wants to trace any particular
>> instance, trace_uprobe is updating reference counter in all instances.
>> This is not a problem on user side because instruction is not replaced
>> with trap/int3 and thus user will only see samples from his interested
>> process. But still this is more of a correctness issue. I'm working on
>> a fix for this.
>
> Hmm, it sounds like not a correctness issue, but there maybe a performace
> tradeoff. Tracing one particulear instance, other instances also will get
> a performance loss
Right, but it's temporary. I mean, putting everything in to this series was making
it complex. So this is the initial one and I'll send followup patches which will
optimize the reference counter update.
> (Only if the parameter preparation block is heavy,
> because the heaviest part of probing - trap/int3 and recording data - isn't
> executed.)
>> BTW, why this happens? I thought the refcounter part is just a data which
> is not shared among processes...
>
This happens because we are not calling consumer_filter function. consumer_filter
is the one who decides whether to change the instruction to trap or not in a given
mm. We also need to call it before updating reference counter.
Let me know your thoughts.
Thanks,
Ravi