Date:   Fri, 19 Jun 2020 01:34:50 +0300
From:   Stanimir Varbanov <stanimir.varbanov@...aro.org>
To:     Jason Baron <jbaron@...mai.com>, jim.cromie@...il.com
Cc:     Petr Mladek <pmladek@...e.com>,
        LKML <linux-kernel@...r.kernel.org>, akpm@...uxfoundation.org,
        Greg KH <gregkh@...uxfoundation.org>,
        Rasmus Villemoes <linux@...musvillemoes.dk>,
        Jonathan Corbet <corbet@....net>,
        Andrew Morton <akpm@...ux-foundation.org>,
        Will Deacon <will@...nel.org>,
        Orson Zhai <orson.zhai@...soc.com>,
        Linux Documentation List <linux-doc@...r.kernel.org>
Subject: Re: [PATCH v3 20/21] dyndbg: add user-flag, negating-flags, and
 filtering on flags

Hi Jason, Jim,

On 6/18/20 10:40 PM, Jason Baron wrote:
> 
> 
> On 6/18/20 3:11 PM, jim.cromie@...il.com wrote:
>> On Thu, Jun 18, 2020 at 12:17 PM Jason Baron <jbaron@...mai.com> wrote:
>>>
>>>
>>>
>>> On 6/18/20 1:40 PM, Petr Mladek wrote:
>>>> On Thu 2020-06-18 18:19:12, Petr Mladek wrote:
>>>>> On Wed 2020-06-17 10:25:35, Jim Cromie wrote:
>>>>>> 1. Add a user-flag [u] which works like the [pfmlt] flags, but has no
>>>>>> effect on callsite behavior; it allows incremental marking of
>>>>>> arbitrary sets of callsites.
>>>>>>
>>>>>> 2. Add [PFMLTU] flags, which negate their counterparts; P===!p etc.
>>>>>> And in ddebug_read_flags():
>>>>>>    current code does:       [pfmltu_] -> flags
>>>>>>    copy it to:              [PFMLTU_] -> mask
>>>>>>
>>>>>> also disallow both of a pair: ie no 'pP', no true & false.
>>>>>>
>>>>>> 3. Add filtering ops into ddebug_change(), right after all the
>>>>>> callsite-property selections are complete.  These filter on the
>>>>>> callsite's current flagstate before applying modflags.
>>>>>>
>>>>>> Why ?
>>>>>>
>>>>>> The u-flag & filter flags
>>>>>>
>>>>>> The 'u' flag lets the user assemble an arbitary set of callsites.
>>>>>> Then using filter flags, user can activate the 'u' callsite set.
>>>>>>
>>>>>>   #> echo 'file foo.c +u; file bar.c +u' > control   # and repeat
>>>>>>   #> echo 'u+p' > control
>>>>>>
>>>>>> Of course, you can continue to just activate your set without ever
>>>>>> marking it 1st, but you could trivially add the markup as you go, then
>>>>>> be able to use it as a constraint later, to undo or modify your set.
>>>>>>
>>>>>>   #> echo 'file foo.c +up' >control
>>>>>>   .. monitor, debug, finish ..
>>>>>>   #> echo 'u-p' >control
>>>>>>
>>>>>>   # then later resume
>>>>>>   #> echo 'u+p' >control
>>>>>>
>>>>>>   # disable some cluttering messages, and remove from u-set
>>>>>>   #> echo 'file noisy.c function:jabber_* u-pu' >control
>>>>>>
>>>>>>   # for doc, recollection
>>>>>>   grep =pu control > my-favorite-callsites
>>>>>>
>>>>>> Note:
>>>>>>
>>>>>> Your flagstate after boot is generally not all =_. -DDEBUG will arm
>>>>>> compiled callsites by default, $builtinmod.dyndbg=+p bootargs can
>>>>>> enable them early, and $module.dyndbg=+p bootargs will arm them when
>>>>>> the module is loaded.  But you could manage them with u-flags:
>>>>>>
>>>>>>   #> echo '-t' >control             # clear t-flag to use it as 2ndary markup
>>>>>>   #> echo 'p+ut' >control   # mark the boot-enabled set of callsites
>>>>>>   #> echo '-p' >control             # clean your dmesg -w stream
>>>>>>
>>>>>>   ... monitor, debug ..
>>>>>>   #> echo 'module of_interest $qterms +pu' >control # build your set of useful debugs
>>>>>>   #> echo 'module of_interest $qterms UT+pu' >control       # same, but dont alter ut marked set
>>>>>
>>>>> Does anyone requested this feature, please?
>>>>>
>>>>> For me, it is really hard to imagine people using these complex and hacky
>>>>> steps.
>>>>
>>>> I think that all this is motivated by adding support for module
>>>> specific groups.
>>>>
>>>> What about storing the group as yet another information for each
>>>> message? I mean the same way as we store module name, file, line,
>>>> function name.
>>>>
>>>> Then we could add API to define group for a given message:
>>>>
>>>>    pr_debug_group(group_id, fmt, ...);
>>>>
>>>> the interface for the control file might be via new keyword "group".
>>>> You could then do something like:
>>>>
>>>>    echo module=drm group=0x3 +p >control
>>>>
>>>> But more importantly you should add functions that might be called
>>>> when the drm.debug parameter is changes. I have already mentioned
>>>> it is another reply:
>>>>
>>>>     dd_enable_module_group(module_name, group_id);
>>>>     dd_disable_module_group(module_name, group_id);
>>>>
>>>>
>>>> It will _not_ need any new flag or flag filtering.
>>>>
>>>> Best Regards,
>>>> Petr
>>>>
>>>
>>> Yes, I'm wondering as well if people are really going to use the
>>> new flags and filter flags - I mentioned that here:
>>> https://urldefense.proofpoint.com/v2/url?u=https-3A__lkml.org_lkml_2020_6_12_732&d=DwIBaQ&c=96ZbZZcaMF4w0F4jpN6LZg&r=1fLh1mlLqbfetnnGsbwXfpwmGlG4m83mXgtV4vZ1B1A&m=vltk6sSzPDQIqO4gGkJeDY6jcEarG4xTztab2EHtPFY&s=6x1EHNoRxebA99Tu-C2i0s5dmdzyEF9bXVcv_cYoM_I&e= 
>>>
>>
>> yes, I saw, and replied there.
>> but since that was v1, and we're on v3, we should refresh.
>>
>> the central use-case is above, 1-liner version summarized here:
>>
>> 1- enable sites as you chase a problem with +up
>> 2- examine them with grep =pu
>> 3- change the set to suit, either by adding or subtracting callsites.
>> 4- continue debugging, and changing callsites to suit
>> 5- grep =pu control > ~/debugging-session-task1-callsites
>> 6- echo up-p >control   # disable for now, leave u-set for later
>> 7- do other stuff
>> 8 echo uP+p >control # reactivate useful debug-state and resume
>>
>>
>>> The grouping stuff is already being used by lots of modules so
>>> that seems useful.
>>
>> I now dont see the need.
>>
>> given N debug callsites, any group can be defined by <N queries,
>> probably a lot less
>> if module authors can use ddebug_exec_queries(), cuz its exported, (15/21)
>> then they can act (+p or -p) on those sets defined by <N queries.
>>
>> and now any callsite can be in any number of groups, not just one.
>> It would be prudent to evaluate such groupings case by case,
>> because the intersecting callsites are subject to "last manipulator wins"
>> but its unnecessary to insist that all sets are disjoint.
>> Unlike pr_debug_n, however its spelled.
>>
> 
> hmm - so I think you are saying there is then no need to change the
> calling functions themselves - its still 'pr_debug()'. You could even
> use the 'format' qualifier for example to implement your groups that
> way.
> 
> For example:
> 
> pr_debug("failure type1: blah");
> pr_debug("failure type2: blah blah");
> 
> and then do: ddebug_exec_queries("format type1 +p", module);
> 
> I would be curious to see what Stanimir thinks of this proposal
> and whether it would work for his venus driver, which is what
> prompted this module group discussion.

Hmm, we spin in a circle :)

Infact this was my first way of implementing the groups in Venus driver,
you can see it at [1].

 +#define VDBGL(fmt, args...)	pr_debug("VENUSL: " fmt, ##args)
 +#define VDBGM(fmt, args...)	pr_debug("VENUSM: " fmt, ##args)
 +#define VDBGH(fmt, args...)	pr_debug("VENUSH: " fmt, ##args)
 +#define VDBGFW(fmt, args...)	pr_debug("VENUSFW: " fmt, ##args)


[1] https://lkml.org/lkml/2020/5/21/668

-- 
regards,
Stan

Powered by blists - more mailing lists