[PATCH v2] tracepoint: add new trace command "printf" and agent expression "printf" [0]

[Stan Shebs]

On 2/24/11 12:36 AM, Hui Zhu wrote:
> Hi guys,
>
> The prev version patches were reverted because I misunderstood the
> Tom's comments and checked in the patch.
>
> So I send a new mail for review.
>
> I add a new patch add new trace command "printf" and agent expression
> "printf" in gdb, gdbserver and doc.

This one is a tough choice.  As Doug said a while back, it doesn't 
really fit into the overall tracepoint structure - the basic idea of 
tracepoints is to accumulate data into a buffer, while this change adds 
a tracepoint action that prints back to the host, and doesn't save 
anything at all.  This can actually lose massively, because in the 
absence of GDB roundtrips, a tracepoint can easily rack up a million 
hits in a second (I've done it) - but a million printfs coming back to 
the host is going to choke both the network and GDB.  In fact it reduces 
debugging performance considerably; while still faster than a breakpoint 
with a command list, most of the tracepoint performance win is lost to 
the I/O back-and-forth.

The flip side is that it does seem like a useful pragmatic addition.  
When a tracepoint experiment runs, there is a bit of guessing at how 
it's going.  You can do tstatus commands, and maybe tfind will work, but 
neither has the live update advantage of the familiar printf.  So I like 
the ability for users to have a sort of live monitoring of their trace 
run, even at the risk of massive spew back up to GDB.

Now if these are the only considerations, I would say to put this in, 
give it a spin, see how it works for users.  But this is not an isolated 
addition.

The first other consideration is that new bytecodes amount to permanent 
commitments.  We now have a handful of targets that think they know what 
the valid bytecodes are, but if we're adding new bytecodes to some but 
not others, then we need to start doing some new infrastructure for the 
target to tell the host about its supported bytecodes.

The second consideration is that we have proposals for target-side 
breakpoint handling on the table.  Mentor now has a contract to do some 
work on this, and we're going to be kicking off public design discussion 
within the next few weeks.  So if we can do random breakpoint actions on 
the target, with no GDB involvement, I think it removes a large part of 
the motivation for this patch.

So overall I think we should put this on the back burner for now, and 
focus on target-side breakpoint handling instead.  There are some 
specific issues that would have to addressed to make this particular 
patch acceptable (the pervasive 100-byte limitation is hard to justify 
for instance), but I'd rather put our brain power into achieving the 
real goal, rather than piggybacking on tracepoints.

Stan

> This printf is same with the simple gdb command, it can do formatted
> output.  But it will happen in gdbserver part when it break by a
> tracepoint.
> Then the user can get the format value that he want in tracepint.  It
> will be more easy and clear to handle the bug sometimes.
>
> About why I add printf to the tracepoint, I have 2 reasons:
> 1. The gdb and gdbserver connect through a low speed net.  Sometimes,
> the debug target that I use is in the other side of the earth.
> The breakpoint commands "printf" is too slow for that issue, because
> each time the inferior is break by the breakpoint, gdbserver need send
> the rsp package to gdb, and gdb will get the data that "printf" need
> though low speed net from gdbsever.  And sometime, it will disconnect.
> But if through tracepoint, I will not have this trouble.  I can "set
> disconnected-tracing on" to handle the network disconnect issue.  I
> still need to get the value from inferior through tfind and other
> commands.  It is still be affect by the low speed network.  So I make
> the tracepoint "printf" to handle it.
>
> 2.  KGTP(https://code.google.com/p/kgtp/) just support the gdb
> tracepoint rsp commands.  For not stop the Linux the Kernel.  It
> doesn't support the breakpoint.
> So if it want directly show the Kernel val value, it need "printf".
> This printf will be very powerful that can set most part of Kernel and
> we can set condition for it.
> And in https://code.google.com/p/kgtp/wiki/HOWTO#Offline_debug,  we
> can dump the gdbrsp package to a file and send to Kernel.  Then kernel
> can be debug without a local gdb or a remote connect.   But user still
> need copy the trace file to pc that have GDB.  But if support
> tracepoint "printf", we will not need do that.
> BTW, the kgtp have supported the agent expression "printf".
>
> About the command part, I use the "printf" instead add a new commands
> because the behavior of  this command is same with printf. They will
> use the same function string_printf(update from ui_printf) to parse
> the command arg.
>
> To support the printf command, I add a new agent expression 0x31
> printf, the format for it is:
> 0x31(op_printf) + arg(1/0) + format string with end by 0x0.
> The arg is the number of argument of printf.  It will only be 1 (one
> argument) or 0 (no argument).  I make it cannot have more than one
> argument because I cannot found a good way to handle va_list that send
> arguments to vprintf.
> The format string with end by 0x0 is the simple format string.  It end
> by 0x0 then the gdbserver can use it directly.
>
> Example:
> (gdb) trace 16
> During symbol reading, DW_AT_name missing from DW_TAG_base_type.
> Tracepoint 1 at 0x4004c3: file 1.c, line 16.
> (gdb) tvariable $c
> Trace state variable $c created, with initial value 0.
> (gdb) actions
> Enter actions for tracepoint 1, one per line.
> End with a line saying just "end".
>> printf "%d 0x%lx %d\n",$c=$c+1,$rax,argc
>> end
> (gdb) target remote localhost:1234
> (gdb) tstart
> (gdb) c
>
> gdbserver/gdbserver  localhost:1234 ./a.out
> Process ./a.out created; pid = 25804
> Listening on port 1234
> Remote debugging from host 127.0.0.1
> 1 0x7f2cb8711ec8 1
> 2 0x7f2cb8711ec8 2
> 3 0x7f2cb8711ec8 3
> 4 0x7f2cb8711ec8 4
> 5 0x7f2cb8711ec8 5
> 6 0x7f2cb8711ec8 6
> 7 0x7f2cb8711ec8 7
>
> Please help me review the patches.
>
> Thanks,
> Hui
>