Re: Section 4.2: wrong error reporting for pthread functions

["Paul E. McKenney" <paulmck@linux.vnet.ibm.com>]

On Wed, Jul 18, 2018 at 12:43:16AM +0900, Akira Yokosawa wrote:
> On 2018/07/16 09:39:24 -0700, Paul E. McKenney wrote:
> > On Tue, Jul 17, 2018 at 12:42:57AM +0900, Akira Yokosawa wrote:
> >> Hi Paul,
> >>
> >> See inline comments below for a few nits and suggestions.
> > 
> > I fixed the perror() calls straightforwardly, thank you!
> > Queued and pushed with both your and Elad's Reported-by.
> > 
> >> On 2018/07/14 16:33:13 -0700, Paul E. McKenney wrote:
> >>> On Sat, Jul 14, 2018 at 08:59:48AM -0400, Elad Lahav wrote:
> > 
> > [ . . . ]
> > 
> >> I see you have already updated most of the code samples under CodeSamples/,
> >> but let me suggest an alternative way not to increase line counts
> >> (or even to decrease line counts).
> >>
> >> "pthread_create(3)" man page gives you an example code.
> >>
> >> First, two helpers are defined as follows:
> >>
> >>        #define handle_error_en(en, msg) \
> >>                do { errno = en; perror(msg); exit(EXIT_FAILURE); } while (0)
> >>
> >>        #define handle_error(msg) \
> >>                do { perror(msg); exit(EXIT_FAILURE); } while (0)
> >>
> >> Then, one of the call sites looks as follows:
> >>
> >>                s = pthread_create(&tinfo[tnum].thread_id, &attr,
> >>                                   &thread_start, &tinfo[tnum]);
> >>                if (s != 0)
> >>                    handle_error_en(s, "pthread_create");
> >>
> >> If we employ this pattern, one of the hunks in your patch will look like:
> >>
> >> -	if (pthread_mutex_lock(pmlp) != 0) {
> >> -		perror("lock_reader:pthread_mutex_lock");
> >> -		exit(-1);
> >> -	}
> >> +	if ((en = pthread_mutex_lock(pmlp)) != 0)
> >> +		handle_error_en(en, "lock_reader:pthread_mutex_lock");
> >>
> >> Thoughts?
> >>
> >> I think these error cases are not our main topic, and to hide the
> >> details inside helpers sounds reasonable.
> > 
> > Does it make sense to pull the "if" into the handle_error_en() macro
> > as well, perhaps like this?
> > 
> > #define handle_error_en(en, msg) \
> > 	do { if (!en) break; errno = en; perror(msg); exit(EXIT_FAILURE); } while (0)
> > 
> > 	s = pthread_create(&tinfo[tnum].thread_id, &attr,
> > 			   &thread_start, &tinfo[tnum]);
> > 	handle_error_en(s, "pthread_create");
> > 
> 
> This version of handle_error_en() can return.
> As per Elad's suggestion, if we want to make fatal_en() not to return,
> we can't pull the "if".
> 
> It looks to me by keeping the "if" out of helper funcs, fatal-error
> conditions can be made more obvious.
> 
> 	s = pthread_create(&tinfo[tnum].thread_id, &attr,
> 			   &thread_start, &tinfo[tnum]);
> 	if (s != 0)
> 		fatal_en(s, "pthread_create");
> 
> I prefer this approach.

At this point, I believe that the right approach is for me to instead
make use of the existing api.h wrappers for the pthread_*() functions.
Smaller code and fewer places to change as opinions shift.  I was
thinking in terms of instead moving to the pthread_*() functions, but
this discussion has clearly shown the folly of that approach.  ;-)

The exceptions are of course the exposition of the pthread_*()
functions in the toolsoftrade chapter.

> >> Also, wouldn't it be a good idea to employ auto-numbering scheme as
> >> mentioned in Section D.3.1.1 of Style Guide when updating code snippets?
> >> This update will involve a lot of renumbering of line numbers otherwise.
> >>
> >> If you feel OK with this approach, I can prepare a patch series
> >> on behalf of you. (Can take a little while, though.)
> > 
> > This approach involves labeling lines that are referred to in the text?
> > If those labels could be introduced as comments in the original code,
> > that could be really nice!
> 
> By using "fancyvrb" package instead of "verbatimbox", this is mostly
> possible.  by "mostly possible", I mean label within comments can
> be made invisible in the output, but "/*", "*/", and "//" will remain
> in the output.
> 
> For example, original source code of Listing 4.1 (using fatal() helper)
> would look like:
> 
> 	pid = fork(); //%label[ln:toolsoftrade:fork:fork]
> 	if (pid == 0) { //%label[ln:toolsoftrade:fork:if]
> 		/* child */ //%label[ln:toolsoftrade:fork:child]
> 	} else if (pid < 0) { //%label[ln:toolsoftrade:fork:else]
> 		/* parent, upon error */ //%label[ln:toolsoftrade:fork:errora]
> 		fatal("fork"); //%label[ln:toolsoftrade:fork:errorb]
> 	} else {
> 		/* parent, pid == child ID */ //%label[ln:toolsoftrade:fork:parent]
> 	}
> 
> Corresponding source of the code snippet (after removal of "//") would
> look like:
> 
>   pid = fork();%label[ln:toolsoftrade:fork:fork]
>   if (pid == 0) {%label[ln:toolsoftrade:fork:if]
>     /* child */%label[ln:toolsoftrade:fork:child]
>   } else if (pid < 0) {%label[ln:toolsoftrade:fork:else]
>     /* parent, upon error */%label[ln:toolsoftrade:fork:errora]
>     fatal("fork");%label[ln:toolsoftrade:fork:errorb]
>   } else {
>     /* parent, pid == child ID */%label[ln:toolsoftrade:fork:parent]
>   }
> 
> Note that in this example, "%" represents "\" after escaping to LaTeX,
> "[" for "{", "]" for "}". These 3 escaping characters need to be
> chosen for each snippet so that they do not appear in the unescaped code.
> "[" and "]" can not be used for snippets that used array reference,
> for example.
> 
> So in theory, we can do what you want, but need somewhat ad-hoc
> manual tweaks. Still, it might be possible to write a script or two
> to do such tweaks in a semi-automated way.

I already use scripts to do the auto-numbering and auto-intenting
for the old-style listings, so why not?  ;-)

If I haven't already made these available, please let me know and
I can send them on.  They aren't exactly profound.

> If you'd like to see what the code snippet and reference to labels
> would be,  I can prepare a experimental branch which is relative to
> commit f2b9d37d3b95 ("count: Expand on gap between C11 atomics and
> the Linux kernel").
> 
> Thoughts?

Sounds worth a try, thank you!

							Thanx, Paul