Re: [PATCH 10/13] tracing/uprobes: Fetch args before reserving a ring buffer

[zhangwei(Jovi)]

On Sat, Aug 10, 2013 at 9:41 AM, zhangwei(Jovi) <jovi.zhangwei@xxxxxxxxx> wrote:
> On Sat, Aug 10, 2013 at 9:26 AM, zhangwei(Jovi) <jovi.zhangwei@xxxxxxxxx> wrote:
>> On Sat, Aug 10, 2013 at 12:20 AM, Oleg Nesterov <oleg@xxxxxxxxxx> wrote:
>>>
>>> Sorry, I didn't read this series yet. Not that I think this needs my
>>> help, but I'll try to do this a later...
>>>
>>> On 08/09, Masami Hiramatsu wrote:
>>> >
>>> > I just concern using kmalloc() in the event handler.
>>>
>>> GFP_KERNEL should be fine for uprobe handler.
>>>
>>> However, iirc this conflicts with the patches from Jovi,
>>> "Support ftrace_event_file base multibuffer" adds rcu_read_lock()
>>> around uprobe_trace_print().
>>
>> (Sorry about html text rejected by kernel.org, send again with plain text.)
>>
>> Then we might need to call kmalloc before rcu_read_lock, also call kfree
>> after rcu_read_unlock.
>>
>> And it's not needed to call kmalloc for each instances in multi-buffer
>> case, just
>> kmalloc once is enough.
>>
>> I also have same concern about use kmalloc in uprobe handler, use kmalloc
>> in uprobe handler seems have a little overhead, why not pre-allocate one page
>> static memory for temp buffer(perhaps trace_uprobe based)? one page size
>> would be enough for all uprobe args storage, then we don't need to call
>> kmalloc in that "fast path".
>>
> forgotten to say, that pre-allocated buffer would need to be per-cpu, to prevent
> buffer corruption.
>
> It's a memory space vs. performance trade-off problem. :)
>
Oops, I missed the per-cpu buffer operation still need
preempt_disable, ignore this part
of my comments, now I agree kmalloc in uprobe handler is needed.

jovi.
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/