[RFC 0/5] "Target set" of command

[Yao Qi]

On 04/02/2013 09:40 PM, Yao Qi wrote:
> Hi,
> This patch series introduce a concept "target set" in GDB.  "Target
> set" is described in HPD standard [1].  In "1.2.2 Process/Thread Sets",
>
> "The concept of the target p/t set is used to restrict a debugger
> command so that is applies to one, many, or all threads of control."
> The command can be applied to processes and threads within the "target
> p/t set".
>
> Something similar is mentioned in [2], "Unlike a serial debugger in which
> each command clearly applies to the only executing thread, TotalView can
> control and monitor many threads with their PCs at many different locations.
> The P/T set indicates the groups, processes, and threads that are the target
> of the CLI command.  No limitation exists on the number of groups, processes,
> and threads in a set".
>
> GDB does process/thread specification now, but in a piecemeal and
> disorganized way.  Some of GDB commands have done something to control
> the scope of inferiors and threads.  For example,
>
>   - "thread apply THREADS COMMAND", for example, "thread apply 1-2 bt"
>     it will apply "bt" to thread 1 and 2.
>
>   - detach/kill/remove inferiors, like "detach inferior 1 3"
>     it will apply "detach" command to inferior 1 and 3.
>
>   - break FOO thread 1
>     it will apply "break" command to thread 1.
>
> In HPD, the "target set" can be defined explicitly by p/t set prefix
> of command or implicitly by current p/t set.  We propose to define
> "target set" explicitly by command "scope" (we'll explain this decision below).
> The equivalents of commands above are:
>
>   - scope t1-2 bt
>
>   - scope i1,i3 detach
>
>   - scope t1 break FOO
>
> As we can see, the command syntax becomes:
>
>    [scope PTSET] COMMAND
>
> Here are some underlying rationales of this design
>
>   1. "Target set" can only be defined by "scope PTSET" as a prefix
>       of command.  Usually, there are four options to set "target set",
>       and the first three are mentioned in HPD [1].
>
>       a) "[all]> [t1] bt".  Current p/t set is "all", and use "prefix
> 	p/t set" to set "target set".  We don't choose it because it
> 	increases the complexity of command parsing.
>       b) "[t1]> bt".  Current p/t set is t1, and use current p/t set to define
> 	"target set".  Without command output aggregation, this choice
> 	is not user friendly.  For example, if current set is "all",
> 	and there are 1000 threads, user types command "bt", GDB will
> 	show backtrace of all 1000 threads.
>       c) "[all]> scope t1 bt".  Set the current set to t1 temporarily,
> 	and "target set" is defined by it.  We choose this option,
> 	because user has to explicitly type "scope" it he/she wants
> 	to use "target set", otherwise, GDB still behaves in the
> 	traditional way.
>       d) The debugger manual clearly lists how "prefix p/t set" and "current
> 	p/t set" affect "target p/t set".  For example, in PGDBG, "the
> 	command set is divided into thread disjoint subsets according to
> 	how each command reacts to the current p/t-set" [3]. Commands are
> 	divided to:
>          "Process level commands" Parallel by current p/t set or prefix p/t
> 	set, for example, "cont", "next" and "step".
> 	"Thread level" Parallel by prefix p/t set only, for example, "print"
> 	and "bt".
> 	"Global commands" No-parallel commands, for example, "enable" and
> 	"disable".
> 	We don't choose this way in GDB because user has to remember which
> 	group each command is in.
>
>    2. Introduce "trigger set" to breakpoint, without "stop set".
>       The "target set" of break means the set of threads break command
>       can apply, in other words, it means the set of threads can trigger
>       the breakpoint.  The "target set" of break command can be called
>       "trigger set" of the breakpoint created by this command.  In HPD,
>       both "trigger set" and "stop set" are described, and also in
>       Pedro's initial ITSET implementation, "trigger set" and "stop set"
>       were supported.  "Trigger set" means which threads are able to
>       trigger breakpoint, while "stop sets" means which threads stop
>       when the breakpoint is triggered.
>
>       We don't support "stop set" because:
>       a) it needs a lot of work in GDB on stopping/suspending threads,
>       b) "stop set" and "trigger set" are quite independent.  Each one
>       can work without the other one.  The "stop set" in TotalView is
>       quite simple, it is about "stop the thread hits breakpoint",
>       "stop the threads of the process in which a thread hits breakpoint",
>       or "stop all processes in a group when execution reaches the
>       breakpoint" [4].  PGDBG Debugger [3] doesn't support "stop set"
>       either.
>       so, we propose to implement "trigger set" first in breakpoint.
>
>    3. The syntax of PTSET is still incomplete.
>       It is not the focus of this patch series to get a consensus on
>       PTSET syntax, and its discussion should go to a separate thread.
>       The PTSET syntax used in this patch is a subset of PTSET syntax,
>       which everyone would agree with, and is sufficient for "target
>       set".  It leaves the open space to complete the syntax step by
>       step.
>
> The patches are stolen from Pedro's initial implementation, with some
> revise and modification.  After my last ITSET proposal was rejected in
> May last year, I kept thinking how to get it, or part of it,
> acceptable.  The major reason that patches are unacceptable is that
> we want to address too many issues in one patch series.  These issues
> are:
>
>    1) Get an complete PTSET syntax,
>    2) Improve GDB on stopping/suspending arbitrary threads,
>    3) Support "trigger set" and "stop set" in breakpoint,
>    4) Support execution control regards of PTSET, like "cont" or
> "step".
>
> Each of them is not a small piece of work, in terms of code and
> discussion.  We'd better to address them one by one.  This patch
> series get touch on part of #1 and #3.  IMO, it is a reasonable set
> to go in, and unfreeze the P/T SET work.  Patch 1 is about the P/T SET,
> patch 2 is to add a command "scope".  Patch 3 and 4 are to teach
> two commands "break" and "bt" be aware of "target set".  Patch 5
> is about test case.  The doc and NEWS entry is not included in
> the patch, and I'll add them if this approach is acceptable.
>
> Once this series go in, we can do something more:
>
>    1.  Transform "thread apply THREADS COMMAND" to
> "scope PTSET COMMAND".
>    2.  Transform "detach/kill/remove inferior INFERIORS" to
> "scope PTSET detach/kill/remove inferior".
>    3.  Teach more commands, such as "gcore" and "kill" to be aware of
> PTSET.
>
> What do you think?
>
> [1] HPD Version 1 Standard.http://web.engr.oregonstate.edu/~pancake/ptools/hpdf/draft/
> [2] TotalView User Guide.
> [3] PGDBG Debugger Guide.http://www.pgroup.com/doc/pgdbgug.pdf
> [4] TotalView Reference Guide.

No comments?

-- 
Yao (齐尧)