Re: [PATCH 1/2 v2] tracing/events: provide string with undefinedsize support

From: Steven Rostedt
Date: Fri Apr 17 2009 - 06:11:16 EST



[ /me can't sleep, trying to get tired again ]


On Fri, 17 Apr 2009, Peter Zijlstra wrote:

> On Fri, 2009-04-17 at 10:59 +0200, Fr?d?ric Weisbecker wrote:
>
> > struct foo {
> > field1;
> > field2;
> > ...
> > };
> >
> > struct foo *f;
> >
> > event = ring_buffer_lock_reserve(sizeof(*f), ...);
> > f = ring_buffer_event_data(event);
> >
> > I can't use a kmalloc here. We are tracing a random event which can happen
> > at a random frequency, random context, etc...
>
> Can't you add a bit to that ring_buffer_lock_reserve() thing?
>
> The thing I do for perf_counters is I first iterate all the output, then
> make the reserve large enough to fit all the output in, then copy the
> bits into the output buffer.
>
> The result is that the output cannot be interpreted as a fixed offset
> struct, but that's not much of an issue anyway.
>
> Another possibility is using relative pointers for strings that point
> beyond the tail of the fixed offset struct.
>
> So something like:
>
> __field(int, foo);
> __string(bar);
> __field(int, foo2);
> __string(bar2);
> __field(int, foo3);
>
> would look like:
>
> struct plop {
> int foo;
> char *bar;
> int foo2;
> char *bar2;
> int foo3;
>
> char data[0];
> }
>
> and you'd do something like:
>
> size = sizeof(struct plop);
> size += strlen(bar) + 1;
> size += strlen(bar2) + 1;
>
> event = ring_buffer_lock_reserve(size);
> offset = sizeof(struct plop);
> my_plop.bar = (char *)offset;
> offset += strlen(bar) + 1;
> my_plop.bar2 = (char *)offset;
> memcpy(&event, &my_plop, sizeof(struct plop));
> memcpy(&event + my_plop.bar, bar, strlen(bar)+1);
> memcpy(&event + my_plop.bar2, bar2, strlen(bar2)+1);
> ring_buffer_unlock();
>
> Then on reading, you'd get a variable sized entry, with a fixed size
> fixed offset struct, that contains relative offset character pointers.

When I replied to Frederic, I thought I could come up with a way to do
something like you are proposing. Instead, I only ended up with the
variant that Frederic implemented.

I've done what you are suggesting several times in tracing. Logdev does
this in its tracing.

The problem that we have, is that we don't have actual code. We have a
TRACE_EVENT macro that is doing the work for us. This, unfortunately,
limits what we can do.

One response is to just free code it. Well, then all users of TRACE_EVENT
would end up doing a lot more work too.

The more power we put into TRACE_EVENT, the uglier and more complex it
ends up.

But that said, one could still do what you are suggesting. The "__ending"
wart still needs to be there, since it must happen at the end of the
struct.


TP_STRUCT__entry(
__field(int, str2loc)
__field(int, str3loc)
__string(str1)
__string(str2)
__ending_string(data, str3)
),
TP_fast_assign(
__entry->str2loc = strlen(str1);
__entry->str3loc = strlen(str2) + __entry->str2loc;
strcpy(__entry->data, str1);
strcpy(__entry->data + __entry->str2loc, str2)
strcpy(__entry->data + __entry->str3loc, str3)
),
TP_printk("str1:%s str2:%s str3:%s\n",
__entry->data,
__entry->data + __entry->str2loc,
__entry->date + __entry->str3loc)


For this to work, we need to just add a define for the __string macro for
the reservation:

#define __string(x) __str_size_ += strlen(x) + 1;

-- Steve

--
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/