diff options
Diffstat (limited to 'community/gsoc/2013/nlightnfotis.mdwn')
| -rw-r--r-- | community/gsoc/2013/nlightnfotis.mdwn | 2587 |
1 files changed, 2587 insertions, 0 deletions
diff --git a/community/gsoc/2013/nlightnfotis.mdwn b/community/gsoc/2013/nlightnfotis.mdwn index 43f9b14c..a9176f51 100644 --- a/community/gsoc/2013/nlightnfotis.mdwn +++ b/community/gsoc/2013/nlightnfotis.mdwn @@ -448,3 +448,2590 @@ License|/fdl]]."]]"""]] <tschwinge> nlightnfotis: OK, so probably waiting at the FSF office to be processed. Let's allow for some more time. After all, this is not critical for your progress. + + +# IRC, freenode, #hurd, 2013-07-10 + + <nlightnfotis> tschwinge: I have run the diff of the GCC repo on the Hurd + against the one on my host linux os, and there was nothing relevant to + fixcontext and initcontext that are the ones that fail the + compilation. In any case I did recheck out the branch, and I have + attempted a build with it. It fails at the same point. Now I am + attempting a build with the -w (inhibit warnings) flag enabled + <tschwinge> nlightnfotis: Have there been any differences in the diff? + There should be none at all. + <nlightnfotis> tschwinge: there were some small changes due to the repo's + being checked out at different times. It was a large diff however. I + inspected it and didn't find anythign that was of much use. Here it is in + case you might want to see it: + https://www.dropbox.com/s/ilgc3skmhst7lpv/diffs_in_git.txt + <tschwinge> nlightnfotis: Well, the idea of this exercise precisely was to + use the same Git revisions on both sides of the diff -- to show that + there are no spurious differences -- which can't be shown from your + 124486 lines diff. (Even though indeed there is no difference in + libgo/configure that would explain the mis-match, but who knows what else + might be relevant for that. + <tschwinge> Would you please repeat that? + <nlightnfotis> tschwinge: I will do so. It was wrong from me to not diff + against the same revisions, but going through the diff results grepping + for the problematic code didn't yield any results, so I thought that + might not be the issue. + <nlightnfotis> I will perform the diff again tomorrow morning and report on + the results. + <tschwinge> nlightnfotis: Anyway, if you checked out again, the latest + revision, and it still fails in exactly the same way, there is something + wrong. + <tschwinge> nlightnfotis: And -w won't help, as there is a hard error + involved. + <tschwinge> nlightnfotis: Are yous till working on GSoC things today? + <nlightnfotis> tschwinge: yeah I am here. I decided to do the diff today + instead of tomorrow. + <nlightnfotis> It finished now btw + <nlightnfotis> let me tell you + <nlightnfotis> ah and this time, the gits were checked out at the same time + <nlightnfotis> from the same source + <nlightnfotis> and are at the same branch + <tschwinge> nlightnfotis: Coulod you upload the + gccbuild/i686-unknown-gnu0.3/libgo/config.log of the build that failed? + <nlightnfotis> tschwinge: sure. give me a minute + <nlightnfotis> tschwinge: there is something strange going on. The two + repos are at the exact same state (or at least should be, and the logs + indicate them to be) but still the diff output is 4.4 mb + <nlightnfotis> but no presence of initcontext of fixcontext + <nlightnfotis> tschwinge: the config.log file --> + http://pastebin.com/bSCW1JfF + <nlightnfotis> wow! I can see several errors in the config.log file + <nlightnfotis> but I am not so sure about their fatality. Config returns 0 + at the end of the log + <tschwinge> nlightnfotis: As the configure scripts probe for all kings of + features on all kings of strange systems, it's to be expected that some + of these fail on GNU/Hurd. + <tschwinge> What is not expected, however, is: + <tschwinge> configure:15046: checking whether setcontext clobbers TLS + variables + <tschwinge> [...] + <tschwinge> configure:15172: ./conftest + <tschwinge> /root/gcc_new/gcc/libgo/configure: line 1740: 1015 Aborted + ./conftest$ac_exeext + <tschwinge> Hmm. apt-cache policy libc0.3 + <tschwinge> nlightnfotis: ^ + <nlightnfotis> tschwinge: Installed 2.13-39+hurd.3 + <nlightnfotis> Candidate: 2.1-6 + <nlightnfotis> *2.17 + <tschwinge> Bummer. + <tschwinge> nlightnfotis: As indicated in + <http://news.gmane.org/find-root.php?message_id=%3C87li6cvjnl.fsf%40kepler.schwinge.homeip.net%3E> + and thereabouts, you need 2.17-3+hurd.4 or later... + <tschwinge> Well. + <tschwinge> At least that now explains what is going on. + <nlightnfotis> tschwinge: i see. I am in the process of updating my hurd + vm. I saw that libc has also been updated to 2.17 + <nlightnfotis> I will confirm when updating is done + <tschwinge> nlightnfotis: Anyway, is the diff between the two repositories + empty now or are there still differences? + <nlightnfotis> there are differences + <nlightnfotis> and they were checked out at the same time + <nlightnfotis> from the same source + <nlightnfotis> (the official git mirror) + <nlightnfotis> and they are both at the same branch + <nlightnfotis> and still diff output is 4.4 MB + <nlightnfotis> but quick grepping into it and there is not mention of + initcontext or fixcontext + <tschwinge> That's... unexpected. + <nlightnfotis> may be a mistake I am making + <nlightnfotis> but considering that diff run for some time before + completing + <tschwinge> In both Git repositories, »git rev-parse HEAD« shows the same + thing? + <tschwinge> Could you please upload the diff again? + <nlightnfotis> tschwinge: confirmed. libc is now version 2.17-1 + <nlightnfotis> tschwinge: http://pastebin.com/bSCW1JfF + <nlightnfotis> for the rev-parse give me a second + <tschwinge> nlightnfotis: Where is libc0.3 2.17-1 coming from? You need + 2.17-3+hurd.4 or later. + <nlightnfotis> it is 2.17-7+hurd.1 + <tschwinge> OK, good. + <tschwinge> The URL you just have is the config.log file, not the diff. + <tschwinge> s%have%gave + <nlightnfotis> oh my mistake + <nlightnfotis> wait a minute + <nlightnfotis> the two repos have different output to rev-parse + <tschwinge> Phew. + <tschwinge> That explains. + <tschwinge> So the Git branches are at different revisions. + <nlightnfotis> that confused me... when I run git pull -a the branches that + were changed were all updated to the same revision + <nlightnfotis> unless... there were some automatic merges in the *host* GCC + repo required during some pulls + <nlightnfotis> but that was some time ago + <nlightnfotis> would it have messed my local history that much? + <nlightnfotis> that's the only thing that may be different between the two + repos + <nlightnfotis> they checkout from the same source + <tschwinge> nlightnfotis: At which revisions are the two + repositories/branches? + <tschwinge> I have never used »put pull -a«. What does that do? + <nlightnfotis> tschwinge: from what I know it does an automatic git fetch + followed by git merge. The -a flag must signal to pull all branches (I + think it's possible to pull only one branch) + <tschwinge> That's the --all option. -a is something different (that I + don't understand off-hand). + <tschwinge> Well, --all means to pull all remotes. + <tschwinge> But you just want the GCC upstream, I guess. + <tschwinge> I always use git fetch and git merge manually. + <nlightnfotis> oh my god! You are write. -a is equivallent to --append + <nlightnfotis> + https://www.kernel.org/pub/software/scm/git/docs/git-pull.html + <nlightnfotis> git pull must be safe though + <nlightnfotis> + http://stackoverflow.com/questions/292357/whats-the-difference-between-git-pull-and-git-fetch + <nlightnfotis> without the -a + <nlightnfotis> *right + <nlightnfotis> why did I even write "right" as "write" above I don't + even... + <nlightnfotis> what did I write in the sentence above + <nlightnfotis> oh my god... + <nlightnfotis> tschwinge: they are indeed on different revisions: The host + repo's last commit was made by me apparently, to merge master into + tschwinge/t/hurd/go, whereas the last commit of the Hurd repo was by you + and it reverted commit 2eb51ea + <nlightnfotis> and that should also explain the large diff file + <nlightnfotis> with master merged into the tschwinge/t/hurd/go branch + <nlightnfotis> I will purge the debian repo and redownload it + <nlightnfotis> *reclone it + <nlightnfotis> that should bring it to a safe state I suppose. + + +# IRC, freenode, #hurd, 2013-07-11 + + <teythoon> nlightnfotis: how's your build going? + <nlightnfotis> I tried one earlier and it seemed to build without any + issues, something that was...strange. I am repeating the build now, but I + am saving the compilation output this time to study it. + <teythoon> it was strange that the build succeeded? that sounds sad :/ + <nlightnfotis> teythoon: considering that 3 weeks now I failed to build it + without errors, it sure seems weird that it builds without errors now :) + <braunr> what did you change ? + <nlightnfotis> braunr: not many things apparently. To be honest the change + that seemed to do the trick was (under thomas' guidance) update of libc + from 2.13 to 2.17 + <braunr> well that can explain + <nlightnfotis> tschwinge: Big update! GCC-go not compiles without errors + under the Hurd. I have done 2 compilations so far, none of which had + issues. Time needed for full build (without bootstrap) is 45 minutes +- 1 + minute. I also run the test suite, and I can confirm your results + <pinotree> s/not/now/, perhaps? + <nlightnfotis> pinotree yeah. I don't know how it came up with not there. I + meant now + <nlightnfotis> tschwinge: link for the go.sum is here --> + https://www.dropbox.com/s/7qze9znhv96t1wj/go.sum + + +# IRC, freenode, #hurd, 2013-07-12 + + <tschwinge> nlightnfotis: Great! So you finally reproduced my results. + :-) + <nlightnfotis> tschwinge: Yep! I am now building a blog, so that I can move + my reports there, so that they are more detailed, to allow for greater + transparency of my actions + <tschwinge> nlightnfotis: Did you recently (in email, I think?) indicate + that there is another Go testsuite, for libgo? + <tschwinge> nlightnfotis: As you prefer. + <nlightnfotis> tschwinge: there seemed to be one, at least in linux. I + think I saw one in the Hurd too. + <tschwinge> Oh indeed there is a libgo testsuite, too. + <nlightnfotis> as a matter of fact, make check-go + <nlightnfotis> did check for the lib + <nlightnfotis> but lib was failing + <nlightnfotis> yeah + <tschwinge> So please have a look at that testsuite's results, too, and + compare to the GNU/Linux ones. + <nlightnfotis> sure. I can do that now. + <tschwinge> And for the go.sum you posted, please have a look at the tests + that do not pass (»grep -v ^PASS: < go.sum«), assuming they do pass on + GNU/Linux. + <tschwinge> I suggest you add a list of the differences between GNU/Linux + and GNU/Hurd testresults to the wiki page, + <http://darnassus.sceen.net/~hurd-web/open_issues/gccgo/>, at the end of + the Part I section. + <nlightnfotis> I'm on it. + <tschwinge> For now, please ignore any failing tests that have »select« in + their name -- that is, do file them, but do not spend a lot of time + figuring out what might be wrong there. + <tschwinge> The Hurd's select implementation is a bit of a beast, and I + don't want you -- at this time -- spend a lot of time on that. We + already know there are some deficiencies, so we should postpone that to + later. + <nlightnfotis> tschwinge: noted. + <tschwinge> So what I would like at the moment, is a list of the testresult + differences to GNU/Linux, then from the go.log file any useful + information about the failing test (which perhaps already explains) + what's going wrong, and then a analysis of the failure. + <tschwinge> nlightnfotis: I assume you must be really happy that you + finally got it build fine, and reproduced my results. :-) + <nlightnfotis> tschwinge: yeah! I can not hide from you the fact that + failing all those builds made me really nervous about me missing my + schedule. Having finally built that and revisiting my application I can + see I am on schedule, but I have to intensify my work to compensate for + any potential unforeseen obstacles + <nlightnfotis> , in the futute + <nlightnfotis> *future + + +# IRC, freenode, #hurd, 2013-07-15 + + <youpi> nlightnfotis: btw, do you have a weekly progress report? + <nlightnfotis> youpi: not yet. Will write it shortly and post it here. I + made a new blog to keep track of my progress. + <nlightnfotis> Will report much more frequently now via my blog + <youpi> did you add your blog url to the hurd iwki? + <nlightnfotis> currently I am running gcc tests on both gcc go and libgo to + see what the differences are with Linux + <nlightnfotis> I believe I have done so, let me see + <nlightnfotis> youpi: gccgo passes most of its tests (it fails a small + number, and I am looking into those tests) but libgo fails 130/131 tests + (on the Hurd that is) + <youpi> ok + + <nlightnfotis> guys I wrote my report. This time I made it available on my + personal blog. You can find it here: + www.fotiskoutoulakis.com/blog/2013/07/15/gsoc-week-4-report/ As always, + open to (and encouraging) criticism, suggestions, anything that might + help me. + <nlightnfotis> I also have to mention that now that my personal website is + online, I will report much more frequently, to the scale of reporting day + by day, or every 2-3 days. + <youpi> nlightnfotis: without spending time on select, it'd be good to have + an idea of what is going wrong + <braunr> eh, go having trouble with select + <youpi> select is a beast, but we do have fixed things lately and we don't + currently know any issue still pending + <nlightnfotis> youpi: are you suggesting to not skip the select tests too? + <braunr> select is kind of critical .. + <braunr> as youpi said, if you can determine what's wrong, at the interface + level (not the implementation), it would be a good thing to do + <youpi> so we know what's wrong + <youpi> we're not asking to fix it, though + <nlightnfotis> braunr: youpi: noted. Thanks for the feedback. Is there + something else you might want me to improve? Something with the report + itself? Something you were expecting to see but I failed to provide? + <braunr> no it's ok + <braunr> it's short, readable, and readily answers the questions i might + have had so it's good + <braunr> as you say, now you have to work on the core of your task :) + <youpi> note: the "select" word in the testsuite is not strictly bound to + the C "select" + <youpi> so it is probably really worth digging a bit at least on the go + side + <braunr> but it's really worth doing in the end, as it will probably reveal + some nasty bugs on the way + <nlightnfotis> I appreciate your input. I will start working on it asap + (today) and will report on Wednesday perhaps (or Thursday at worst). + + +# IRC, freenode, #hurd, 2013-07-18 + + <nlightnfotis> braunr: I found out what was causing the fails in the tests + <nlightnfotis> in both libgo and gccgo + <nlightnfotis> it's a assertion: mach_port_t ktid = __mach_thread_self (); + int ok = thread->kernel_thread == ktid; __mach_port_deallocate + ((__mach_task_self_ + 0), ktid); ok; }) + <braunr> is all that the assertion ? + <nlightnfotis> yes + <braunr> please paste the code somewhere + <braunr> or is it in libpthread ? + <nlightnfotis> http://pastebin.com/G2w9d474 + nonblock.x: ./pthread/pt-create.c:167: __pthread_create_internal: Assertion `({ mach_port_t ktid = __mach_thread_self (); int ok = thread->kernel_thread == ktid; __mach_port_deallocate ((__mach_task_self_ + 0), ktid); ok; })' failed. + 9 FAIL: go.test/test/chan/nonblock.go execution, -O2 -g + <braunr> yes + <braunr> that's related to my current work on thread destruction + +[[open_issues/libpthread/t/fix_have_kernel_resources]]. + + <braunr> thread resources recycling is buggy + <braunr> i suggest you make your own thread pool if you can + <nlightnfotis> I will look into it further and let you know. Thanks for + that. + + +# IRC, freenode, #hurd, 2013-07-22 + + <nlightnfotis> tschwinge, I have found what is failing both libgo and gccgo + tests, but for the life of me, I can not really find the offending code + on any repository. + <nlightnfotis> not even the eglibc-source debian package. it's driving me + insane. + <tschwinge> nlightnfotis: If this is driving you insane, we should quickly + have a look at that! + <nlightnfotis> thanks tschwinge: I have found that the offending code is an + assertion: { mach_port_t ktid = __mach_thread_self (); int ok = + thread->kernel_th read == ktid; __mach_port_deallocate ((__mach_task_s + elf_ + 0), ktid); ok; } on a file called pt-create.c under the + libpthread on line 167 + <nlightnfotis> but for the life of me, I can not find that piece of code + anywhere. And when I mean anywhere, I mean anywhere. I have looked for it + on all of the branches of glibc, libpthread and the source code of + eglibc. + <nlightnfotis> that's why if you don't mind I would like to write my report + in a day or two, when (hopefully) I will have more progress to report on. + <youpi> nlightnfotis: isn't that libpthread/sysdeps/mach/pt-thread-start.c + ? + <youpi> or rather, ./sysdeps/mach/hurd/pt-sysdep.h + <nlightnfotis> youpi: let me check this out. If that's it I'm gonna cry. + <youpi> which unfortunately is inlined in a lot of places + <youpi> nlightnfotis: does the assertion not tell you the file & line? + <nlightnfotis> youpi: holy smokes! That's the code I was looking for! Oh + boy. Yeah the logs do tell me, but it was very misleading. So misleading, + taht I was actually looking at the wrong place. All logs suggest that + this piece of code is at libpthread/pthread/pt-create.c in line 167 + <youpi> what is that line in your tree? + <youpi> a call to _pthread_self(), isn't it? + <youpi> then it's not actually misleading, this is indeed where the + pt-sysdep.h definition gets inlined + <nlightnfotis> it seems so, yeah. it's err = __pthread_sigstate + (_pthread_self (), 0, 0, &sigset, 0); + <youpi> nlightnfotis: and what is the backtrace? + <nlightnfotis> youpi: _pthread_create_internal: Assertion failed. + <nlightnfotis> The assertion is the one above + <youpi> nlightnfotis: sure, but what is the backtrace? + <nlightnfotis> I don't have the full backtrace. These are the logs from the + compiler. All I can get is: reports like this: nonblock.x: + ./pthread/pt-create.c:167: __pthread_create_internal: Assertion `({ + mach_port_t ktid = __mach_thread_self (); int ok = thread->kernel_thread + == ktid; __mach_port_deallocate ((__mach_task_self_ + 0), ktid); + ok; })' failed. + <youpi> nlightnfotis: you should probably have a look at running the tests + by hand + <youpi> so you can run them in a debugger, and get backtraces etc. + <braunr> nlightnfotis: did i answer that ? + <nlightnfotis> braunr: which one? + <braunr> the problems you're seeing are the pthread resources leaks i've + been trying to fix lately + <braunr> they're not only leaks + <braunr> creation and destruction are buggy + <nlightnfotis> I have read so in + http://www.gnu.org/software/hurd/libpthread.html. I believe it's under + Thread's Death right? + <braunr> nlightnfotis: yes but it's buggy + <braunr> and the description doesn't describe the bugs + <nlightnfotis> so we will either have to find a temporary workaround, or + better yet work on a fix, right? + <braunr> nlightnfotis: i also told you the work around + <braunr> nlightnfotis: create a thread pool + <nlightnfotis> braunr: since thread creation is also buggy, wouldn't the + thread pool be buggy too? + <braunr> nlightnfotis: creation *and* destruction is buggy + <braunr> nlightnfotis: i.e. recycling is buggy + <braunr> nlightnfotis: the hurd servers aren't affected much because the + worker threads are actually never destroyed on debian (because of a + debian specific patch) + + <teythoon> youpi, nlightnfotis, hacklu_: btw, what about the copyright + assignment process + <tschwinge> nlightnfotis just got his on file, so there is progress. + <tschwinge> I have email from Donald R Robertson III + <copyright-clerk@fsf.org> about that -- but it is not yet present in the + FSF copyright.list file... + <tschwinge> I think I received that email because I was CCed on + nlightnfotis' submission. + <nlightnfotis> tschwinge: I have got the papers, and they were signed by + the FSF. They stated delivery date 11 of July, but the documents were + signed on the 10th of July :P + <tschwinge> Ah, no, I received it via hurd-maintainers@gnu.org -- and the + strange thing is that not all assignments that got processed got sent + there... + <tschwinge> At the recent GNU Tools Cauldron we also discussed this in the + GCC context; and their experience was the very same. Emails get lost, + and/or take ages to be processed, etc. + <tschwinge> It seems the FSF is undermanned. + + +# IRC, freenode, #hurd, 2013-07-27 + + <nlightnfotis> I have one question about the Mach sources: I can see it + uses its own scheduler (more like, initializes) and also does the same + for the linux scheduler. Which one does it use? + <youpi> it doesn't use the linux scheduler + <youpi> the linux glue just glues linux scheduling concepts onto the mach + scheduler + <nlightnfotis> ohh I see now. Thanks for that youpi. + + +# IRC, freenode, #hurd, 2013-07-28 + + <nlightnfotis> In the mach kernel source code, does the (void) before a + function call have a semantic meaning, or is it just remnants of the past + (or even documentation) + <pinotree> for example? + <nlightnfotis> pinotree: (void) thread_create (kernel_task, + &startup_thread); + <nlightnfotis> I read on stack overflow that there is only one case where + it has a semantic meaning, most of the times it doesn't + <nlightnfotis> + http://stackoverflow.com/questions/13954517/use-of-void-before-a-function-call + <pinotree> most probably thread_create has a non-void return value, and + this way you're explicitly suppressing its return value (usually because + you don't want/need to care about it) + <nlightnfotis> isn't the value discarded if the (void) is not there? + <pinotree> yes, but depending on extra attributes and/or compiler warning + flags the compiler might warn that the return value is not used while it + ought to + <pinotree> the cast to void should suppress that + <nlightnfotis> oh, okay, thanks for that pinotree + <nlightnfotis> and yes you are right that thread_create actually does + return something + <pinotree> even if there would be no compiler message about that, adding + the explicit cast could mean "yes, i know the function does return + something, but i don't care about it" + <pinotree> ... as hint to other code readers + <nlightnfotis> as a form of documentation then + <pinotree> also + + <nlightnfotis> oh well, I am gonna ask and I hope someone will answer it: + In the Mach's dmesg (/var/log/dmesg) I can see that the version string + along with initial memory mapping information are printed twice, when in + fact they are supposed to be called only once. Is this a bug, or some + buffering error, or are they actually called twice for some reason? + + +# IRC, freenode, #hurd, 2013-07-29 + + <nlightnfotis> guys is the evaluation today? + <hacklu_> yes + <teythoon> right + <nlightnfotis> where can we find the evaluation papers on melange? + <hacklu_> wait untill 12pm UTC. + <nlightnfotis> yeah, I just noticed thanks hacklu_ + <hacklu_> nlightnfotis:) + + <NlightNFotis> tschwinge: I only have one question regarding my project. If + I make some changes to libpthread, what's the best way to test them in + the hurd? Rebuild glibc with the updated libpthread? + <tschwinge> NlightNFotis: Yes, you'll have to rebuild glibc. I have a + cheat sheet for that: + http://darnassus.sceen.net/~hurd-web/open_issues/glibc/debian/ + <tschwinge> It may be that the »Run debian/rules patch to apply patches« + step is no longer encessary with the 2.17 glibc packages. + <NlightNFotis> thanks for that tschwinge. :) + <tschwinge> NlightNFotis: Sure. :-) + + <tschwinge> NlightNFotis: Where's your weekly status? + <NlightNFotis> I will write it today at the noon. I have written all the + other ones, and they are available at www.fotiskoutoulakis.com + <NlightNFotis> the next one will be available there as well, later in the + day + <tschwinge> Ack. But please try to finish your report before the meeting, + as discussed. + <NlightNFotis> oh, forgive me for that. I thought it was ok to write my + report a day or so later. Sorry. + <tschwinge> NlightNFotis: Please write your report as soon as possible -- + otherwise there's no useful way for me to know what your status is. + <NlightNFotis> I will. This week I have been mostly going through the + various sources (the Hurd, Mach and libpthread, especially the last two) + in my attempt to get a better understanding for how libpthread + works. Since yesterday I have attempted some small changes on my + libpthread repo that I plan on testing and reporting on them. That's why + I still have not written my report. + <tschwinge> NlightNFotis: Things don't need to be finished before you + report about them. It's often more useful to discuss issues *before* you + spend time on implementing them. + #hurd + <braunr> NlightNFotis: what kind of changes do you want to add to + libpthread ? + <tschwinge> Have a look at the asseriton failure, I would hope. :-) + <braunr> well no + <braunr> again, i did that + <braunr> and it's not easy to fix + <NlightNFotis> braunr: I was looking into ways that I could create the + thread pool you suggested into libpthread + <braunr> no, don't + <braunr> create it in your application + <braunr> not in libpthread + <braunr> well, this may not be an acceptable solution either .. + <tschwinge> Before doing that we have to understand what exactly the Go + runtime is doing. It may just be a weird itneraction with the setcontext + et al. functions that I failed to think about when implementing these? + <NlightNFotis> the other possibility is the go runtime libraries. But I + thought that libpthread might be a better idea, since you told me that + creation *and* destruction are buggy + <hacklu> braunr: you are right, the signal thread is always exist. I have + got a wrong understand before. + <NlightNFotis> tschwinge: I can look into that, now. I will also include + that in my report. + <braunr> NlightNFotis: i don't see how this is a relevant argument .. + <braunr> tschwinge: i'd suggest he first try with a custom pool in the go + runtime, so we exclude what you're suspecting + <braunr> if this pool actually works around the issues NlightNFotis is + having, it will confirm the offending problem comes from libpthread + <tschwinge> So, as a very first step make any thread + distruction/deallocation a no-op. + <braunr> yes + <NlightNFotis> braunr: I originally understood that a thread pool might + skip the thread's destruction, so that we escape the buggy part with the + thread's destruction. Since that was a problem with libpthread, it sure + affects other threads (instead of go's ) too. So I assumed that building + the thread pool into libpthread might help eliminate bugs that may affect + other code too. + <braunr> no, it's not a proper fix + <braunr> it's a work around + <braunr> and i'm working on a proper fix in parallel + <braunr> (when i have the time, that is :/) + <NlightNFotis> oh, I see. So for the time, I had better not touch + libpthread, and take a look at the go run time aye? + <tschwinge> NlightNFotis: Remember: one thing after the other. First + identify what is wrong exactly. Then think and discuss how to solve the + very specific issue. Then implement it. + <braunr> as tschwinge said, make thread destruction a nop in go + <braunr> see if that helps + <tschwinge> NlightNFotis: For example, you surely have noticed (per your + last report), that basically all Go language test pass (aside from the + handful of those testing select, etc.) -- but all those of the libgo + runtime library fail, literally all of them. + <tschwinge> You noticed they basically all fail with the same assertion + failure. But why do all the Go language ones work fine? + <tschwinge> Don't they execute the program they built, for example? + <tschwinge> (I haven't looked.) + <NlightNFotis> they do execute the program. the language ones that fail + too, fail due to the assertion failure + <tschwinge> Or, what else is different for them? How are they built, which + flags, how are they invoked. + <braunr> how many goroutines ? + <braunr> :p + <tschwinge> Do you also get the assertion failure when you built a small Go + program yourself and run that one. + <tschwinge> Don't get the assertion failure? Then add some more complex + stuff that are likely to invole adding/re-using new threads, such as + goroutines. + <NlightNFotis> I didn't get the assertion failure on a small test program, + but now that you suggest it it might be a good idea to build a custom + test suite + <tschwinge> Etc. That way you'll eventually get an understanding what + triggers the assertion failure. + <tschwinge> And that exeactly is the kind of analysis I'd like to read in + your weekly report. + <tschwinge> A list of things what you have done, which assuptions you've + made, how that directed your further analysis, what results that gave, + etc. + <NlightNFotis> I will do it. I will try to rush to finish it today before + you leave, so that you can inspect it. God I feel like all that time I + spent this week studying the particular source code (libpthread, and the + Mach) were in vain... + <NlightNFotis> on second thoughts, it was not in vain. I got a pretty good + understanding of how these pieces of software work, but now I will have + to do something completely different. + <tschwinge> Studying code is never in vain. + <tschwinge> Exactly. + <tschwinge> You must have had some motivation to study the code, so that + was surely a valid thing to do. + <tschwinge> But we'd link to understand your reasoning, so that we can + support you and direct you accordingly. + <braunr> but it's better to focus on your goals and determine an + appropriate course of actions, usually starting with good analysis + <tschwinge> Yes. + <pinotree> s/link/like/? + <tschwinge> pinotree: Indeed, thanks. + <braunr> makes me remember when i implemented radix trees to replace splay + trees, only to realize splay trees were barely used .. + <tschwinge> braunr: Yes. It has happened to all of us. ;-P + <tschwinge> NlightNFotis: So, don't worry -- but learn from such things. + :-) + <NlightNFotis> anyway, I will start right away with the courses of action + you suggested, and will try to have finished them by noon. Thanks for + your help, it really means a lot. + <tschwinge> In software generally, it is never a good idea to let you be + distracted, and don't follow your focus goal, because there are always so + many different things that could be improved/learned/fixed/etc. + <NlightNFotis> tschwinge, I am only nervous about one thing: the fact that + I have not submitted yet any patch or some piece of code in general. Then + again, the summer of code for me so far has been 70-80% reading about + stuff I didn't know about and 30-20% doing the stuff I should know + about... + <tschwinge> NlightNFotis: That's why we're here, to teach you something. + Which we're happy to do, but we all need to cooperate for that (and I'm + well aware that this is difficult if one is not in the same rooms, and + I'm also aware that my time is pretty limited). + <tschwinge> NlightNFotis: We're also very aware that the Hurd system, as + any operating system project (if you're not just doing "superficial" + things) is difficult, and takes lots of time to learn, and have concepts + and things sink into your brain. + <braunr> i wouldn't worry too much + <tschwinge> We're also still learning every day. + <braunr> go doesn't require a lot from the underlying system, but what is + required is critical + <braunr> once you identify it, coding will be quick + <NlightNFotis> tschwinge: braunr: thanks. I shall begin working following + the directions you gave to me. + <tschwinge> NlightNFotis: So yes, because Google wants us to grade you + based on that, you'll eventually have to write some code, but for + example, a patch to disable thread distruction/deallocation in libgo + would definitely count as such code. And that seems like one of your + next steps. + <NlightNFotis> tschwinge: i need to deliver that instantly, right? seeing + as the evaluation is today. + <tschwinge> NlightNFotis: No. Deliver it when you have something to + deliver. :-) + <NlightNFotis> tschwinge: I am nervous about the evaluation today. I have + not submitted a single piece of code, only some reports. How negatively + does this influence my performance report? + <tschwinge> NlightNFotis: If I can say so, in the evaluation today, Google + basically asks us mentors whether we want to fail our students right now. + Which I don'T plan to do, knowing about the complexity of the Hurd + system, and the learning required before you can do useful code changes. + <NlightNFotis> tschwinge: that really means a lot to me, and it got a + weight of my chest. + <braunr> uh ok, i have to be the rude guy again + <braunr> NlightNFotis: the gsoc is also a way for the student to prepare + for working in software development communities + <braunr> whether free software/open source and/or in companies + <braunr> people involved care a lot less about pathos than actual results + <pinotree> (or to prepare students to be hired by google, but that's + another story) + <braunr> NlightNFotis: in other words, stop apologizing that much, stop + focusing so much on that, and just work as you can + + +# IRC, freenode, #hurd, 2013-07-31 + + <nlightnfotis> teythoon: both samuel and thomas would be missing for the + week right? + <teythoon> nlightnfotis: they do, why? + <teythoon> nlightnfotis: err, they do?? why? + + +# IRC, freenode, #hurd, 2013-08-01 + + <nlightnfotis> braunr: I checked out what you (and Thomas) suggested and + did some research on go on the Hurd. I have found out that go works, + until you need to use anything that has to do with a goroutine. I am now + playing with the go runtime and checking to see if turning thread + destruction to noop will have any difference. + + +# IRC, freenode, #hurd, 2013-08-05 + + <nlightnfotis> youpi: whenever you have time, I would like to report my + progress as well. + <youpi> nlightnfotis: sure, go ahead + <youpi> but again, you should report before the meeting + <youpi> so we can read it before coming to the discussion + <nlightnfotis> I have written my report + <youpi> ah + <hacklu> nlightnfotis: I have read your report, these days you have make a + great progress. + <youpi> where is it? + <nlightnfotis> it was available since yesterday + <nlightnfotis> + http://www.fotiskoutoulakis.com/blog/2013/08/05/gsoc-partial-week-7-report/ + <nlightnfotis> thanks hacklu. The particular piece of code I was studying + was very very interesting :) + <hacklu> nlightnfotis: I think you should show your link in here or email + next time. I have spend a bit more time to find that :) + <nlightnfotis> youpi: for a tldr, at the last time I was told to check + gccgo's runtime for clues regarding the go routine failures. + <nlightnfotis> hacklu: will keep that in mind, thanks. + <nlightnfotis> youpi: thing is, gccgo operates on two different thread + types: G's (the goroutines, lightweight threads that are managed by the + runtime) and M's (the "real" kernel threads") + <nlightnfotis> none of which are really "destroyed" + <youpi> ok, makes sense + <nlightnfotis> G's are put in a pool of available goroutines when their + status is changed to "Gdead" so that they can be reused + <nlightnfotis> M's also don't seem to go away. There is always at least one + M (the bootstrap one) and all other M's that get created are also stashed + in a pool of available working threads. + <youpi> you could put some debugging printfs in libpthread, to make sure + whether threads do die or not + <nlightnfotis> I am studying this further as we speak, but they both don't + seem to get "destroyed", so that we can be sure that bugs are triggered + by thread destruction + <nlightnfotis> I was beginning to believe that maybe I was looking in the + wrong direction + <nlightnfotis> but then I looked at my past findings, and I noticed + something else + <nlightnfotis> if you take a look at the first failed go routine, it failed + at the time.sleep function, which puts a goroutine to sleep for ns + nanoseconds. That made me think if it was something that had to do with + the context functions and not the goroutines' creation. + <youpi> nlightnfotis: that's possible + <youpi> nlightnfotis: I'd say you can focus on this very simple example: a + mere sleep + <youpi> that's one of the simplest things a thread scheduler has to do, but + it has to do it right + <youpi> fixing that should fix a lot of other issues + <nlightnfotis> if I have understood correctly, there is at least one G + (Goroutine) and at least one M (kernel thread) running. Sleep does put + that goroutine at a hold, and restarting it might be an issue + <braunr> talking about thread scheduling ? :) + <youpi> nlightnfotis: go's runtime doesn't actually destroy kernel threads, + apparently + <nlightnfotis> youpi: yeah, that's what I have understood so far. And it + neither does destroy goroutines. If there was an issue with thread + creation, then I guess it should be triggered in the beginning of the + program too (seeing as both M's and G's are created there) + <nlightnfotis> the fact that it is triggered when a goroutine goes to sleep + makes me suspect the context functions + <youpi> yes + <nlightnfotis> again I am studying it the last days, in search of + clues. Will keep you all updated. + <nlightnfotis> braunr: I have written my report and it is available here + http://www.fotiskoutoulakis.com/blog/2013/08/05/gsoc-partial-week-7-report/ + If you could read it and tell me if you notice something weird tell me + so. + <braunr> nlightnfotis: ok + <braunr> nlightnfotis: quite busy here so don't worry if i suddenly + disappear + <braunr> nlightnfotis: hum, does go implement its own threads ?? + <nlightnfotis> braunr: yeah. It has 2 threads. Runtime managed (the + goroutines) and "real" (kernel managed) ones. + <braunr> i mean, does it still use libpthread ? + <nlightnfotis> thing is none of them "disappear" so as to explain the bug + with "thread creation **and** destruction) + <nlightnfotis> it must use libpthread for kernel threads as far as creation + goes. + <braunr> ok, good + <braunr> then, it schedules its own threads inside one pthread, right ? + <braunr> using the pthread as a virtual cpu + <nlightnfotis> yes. It matches kernel threads and runtime threads and runs + the kernel threads in reality + <nlightnfotis> the scheduler decides which goroutine will run on each + kernel thread. + <braunr> ew + <braunr> this is pretty much non portable + <braunr> and you're right to suspect context switching functions + <nlightnfotis> yeah my thought for it was the following: thread creation, + if it was buggy, should be triggered as soon as a program starts, seeing + as at least one kernel thread and at least one go routine starts. My + sleep experiment crashes when the goroutine is put on hold + <braunr> did you find the code putting on hold ? + <nlightnfotis> I will give you the exact link, wait a moment + <nlightnfotis> braunr: + https://github.com/NlightNFotis/gcc/blob/master/libgo/runtime/time.goc?source=c#L59 + <nlightnfotis> that is the exact location is line 26, which calls the one I + pointed you at + <braunr> ahah, tsleep + <braunr> old ghost from the past + <braunr> nlightnfotis: the real location is probably runtime_park + <nlightnfotis> I will check this out. + + <nlightnfotis> may I ask something non-technical but relevant to summer of + code? + <braunr> sure + <nlightnfotis> would it be okay if I took the day off tomorrow? + <braunr> nlightnfotis: ask tschwinge but i guess it's ok + + <braunr> have you found runtime_park ? + <braunr> i'm downloading your repository from github but it's slow :/ + <nlightnfotis> braunr: not yet. Grepping through the files didn't produce + any meaningful results and github's search is not working + <nlightnfotis> braunr: there is that strange thing with th gccgo sources, + where I can find a function's declaration but not it's definition. Funny + thing is those functions are not really extern, so I am playing a hide + and seek game, in which I am not always successful. + <nlightnfotis> runtime_park is declared in runtime.h. I have looked nearly + everywhere for it. There is only one last place I have not looked at. + <nlightnfotis> braunr: I found runtime_park. It's here: + https://github.com/NlightNFotis/gcc/blob/master/libgo/runtime/proc.c?source=c#L1372 + + <tschwinge> nlightnfotis: Taking the day off is fine. Have fun! + <nlightnfotis> tschwinge: I am still here; Thanks for that tschwinge. I + will be for the next half hour or something if you would like to ask me + anything + <tschwinge> nlightnfotis: I have no immediate questions (first have to read + your report and discussion in here) -- so feel free to log out and enjoy + the sun outside. :-) + + <teythoon> nlightnfotis, tschwinge: btw, have you seen + http://morsmachine.dk/go-scheduler ? + <nlightnfotis> teythoon: thanks for the link. It's really interesting. + + +# IRC, freenode, #hurd, 2013-08-12 + + <nlightnfotis> teythoon did you manage to build the Hurd successfuly? + <teythoon> ah yes, the Hurd is relatively easy + <teythoon> the libc is hard + <nlightnfotis> debian glibc or hurd upstream libc? + <teythoon> but my build on darnassus was successful + <nlightnfotis> *debian eglibc + <teythoon> well, I rebuilt the debian package with two tweaks + <nlightnfotis> do you build on linux and rsync on hurd or ...? + <teythoon> I built it on Hurd, though I thought about setting up a cross + compiler + <nlightnfotis> I see. The process was build Mach, build Hurd, and then + build glibc and it's ready or it needed more? + <teythoon> no, I never built Mach + <teythoon> I must admit I'm not sure about the "proper" procedure + <teythoon> if I change one of Hurds RPC definitions, I think the proper way + is to rebuild the libc against the new definitions and then the Hurd + <teythoon> but I found no way to do that, so everyone seems to build the + Hurd, install it, build the libc and then rebuild the Hurd again + <nlightnfotis> I see. Thanks for that :) + + <nlightnfotis> tschwinge, I have also written my report! It's available + here + http://www.fotiskoutoulakis.com/blog/2013/08/12/gsoc-week-8-partial-report/ + <nlightnfotis> I can sum it up if you want me to. + <tschwinge> nlightnfotis: I already read it! :-D + <tschwinge> Oh, I didn't. I read the week 7 one. Let me read week 8. ;-) + <nlightnfotis> ok. I am currently going through the assembly generated for + the sample program I have embedded my report. + <nlightnfotis> the weird thing is that the assembly generated is pretty + much the same for the program with 1 and 2 goroutine functions (with the + obvious difference that the one with 2 goroutine functions has 1 more + goroutine in it's assembly code) + <nlightnfotis> I can not understand why it is that when I have 1 goroutine, + an exception is triggered, but when I am having two (which are 99% + identical) it seems to be executed. + <nlightnfotis> and I do not understand why the exception is triggered when + I manually use a goroutine. + <nlightnfotis> To my understanding so far, there is at least 1 (kernel) + thread created at program startup to run main. The same thread gets + created to run a new goroutine (goroutines get associated with kernel + threads) + <nlightnfotis> and it's obvious from the assembly generated. + <nlightnfotis> go_init_main (the main function for go programs) starts with + a .cfi_startproc + <nlightnfotis> the same piece of code (.cfi_startproc) starts a new kernel + thread (on which a goroutine runs) + <tschwinge> nlightnfotis: Re your two-goroutines example: in that case I + assume, you're directly returning from the main function and the program + terminates normally. ;-) + <tschwinge> nlightnfotis: Studying the assembly code for this will be too + verbose, too low-level. What we need is a trace of steps that happen + until the error. + <nlightnfotis> tschwinge, that must be it, but it should trigger the bug, + since it still has at least one goroutine (and one is known to trigger + the bug) + <tschwinge> nlightnfotis: I guess the program exits before the first + gorouting would be scheduled for execution. + <nlightnfotis> the assembly for the goroutines is identical. You can't tell + one from the other. The only change is that it has 2 of these sections + instead of one + <nlightnfotis> actually it's the same for the first one + <tschwinge> nlightnfotis: I very much assume that the issue is not due to + the code generated by the Go compiler (which you're seeing in the + assembly code), but rather due to the runtime code in the libgo library. + <nlightnfotis> I didn't think of it this way. + <tschwinge> ... that improperly interacts with our libpthread. + <nlightnfotis> so my research should focus on the runtime from now on? + <tschwinge> Improperly may well imply that our libpthread is at fault, of + course, as we discussed. + <tschwinge> Back to the one-gouroutine case (that shows the assertion + failure). Simple case: one goroutine, plus the "main" thread. + <tschwinge> We need to get an understanding of the steps that happen until + the error happens. + <tschwinge> As this is a parallel problem, and it is involving "advanced" + things (such as setcontext), I would not trust GDB too much when used on + this code. + <nlightnfotis> I will have to manually step through the source myself, + right? + <tschwinge> What I would do, is add printf's (or similar) into the code at + critical points, to get an udnerstanding of what's going on. + <tschwinge> Such critical points are: pthread_create, setcontext, + swapcontext. + <nlightnfotis> It sounds like a good idea. Anything else to note? + <tschwinge> That way, you can isolate the steps required to trigger the + assertion failure. + <tschwinge> For example, it could be something like: makecontext, + swapcontext, pthread_creat, boom. + <nlightnfotis> pthread_create_internal is failing at an assertion. I wonder + what would happen if I remove that assertion. + <tschwinge> Not without understanding what the error is, and why it is + happening (which steps lead to it). We don't usually do »voodoo + computing and programming by coincidence«. + <nlightnfotis> tschwinge, I also figured out something. If it is a + libpthread issue, it should also get triggered when a simple C program + creates a thread (assuming _pthread_create is causing the issue) + <nlightnfotis> so maybe I should write a C program to test that + functionality and see if it provides any further clues? + <tschwinge> nlightnfotis: That's precile what the goal of »isolate the + steps required to trigger the assertion failure« is about: reduce the big + libgo code to a few function calls required to reproduce the problem. + <tschwinge> nlightnfotis: I simple C program just doing pthread_create + evidently does not fail. + <tschwinge> nlightnfotis: I assume you have a Go program dynamically linked + to the libgo you build? + <nlightnfotis> yes. To the latest go build from the source (4.9) + <nlightnfotis> *gccgo build from source + <braunr> removing an assertion is usually extremely bad practice + <tschwinge> Then you can just do something like make target-libgo (IIRC) + (or instead: cd i686-pc-gnu/libgo/ && make) to rebuild your changed + libgo, and then re-run the Go program. + <braunr> the thought of randomly removing assertions shouldn't even reach + your mind ! + <nlightnfotis> braunr: even if it is not permanent, but an experiment? + <braunr> yes + <nlightnfotis> can you explain to me why? + <tschwinge> nlightnfotis: <tschwinge> Not without understanding what the + error is, and why it is happening (which steps lead to it). We don't + usually do »voodoo computing and programming by coincidence«. + <braunr> an assertion exists to make sure something that should *never* + happen never happens + <braunr> removing it allows such events to silently occur + <teythoon> braunr: that's the theory, yes, to check invariants + <braunr> i dont' know what you mean by using assertions for "an experiment" + <teythoon> unfortunately some people use assert for error handling :/ + <braunr> that's wrong + <braunr> and i dont't remember it to be the case in libpthread + <braunr> nlightnfotis: can you point the faulting assertion again there + please ? + <nlightnfotis> braunr: sure: Assertion `({ mach_port_t ktid = + __mach_thread_self (); int ok = thread->kernel_thread == ktid; + <nlightnfotis> __mach_port_deallocate ((__mach_task_self + 0), ktid); ok; + })' failed. + <braunr> so basically, thread->kernel_thread != __mach_thread_self() + <braunr> this code is run only for num_threads == 1 + <braunr> but has there been any thread destruction before ? + <nlightnfotis> no. To my understanding kernel threads in the go runtime + never get destroyed (comments seem to support that) + <braunr> IOW: is it certain the only thread left *is* the main thread ? + <braunr> hm + <braunr> intuitively, i'd say this is wrong + <braunr> i'd say go doesn't destroy threads in most cases, but something in + the go runtime must have done it already + <braunr> i'm not even sure the main thread still exists + <braunr> check that + <braunr> where is the go code you're working on ? + <nlightnfotis> there are 3 files of interest + <braunr> i'd like the whole sources please + <nlightnfotis> I will find it in a moment + <tschwinge> braunr: GCC Git clone, tschwinge/t/hurd/go branch. + <nlightnfotis> it is <gcc_root>/libgo/runtime/runtime.h + <nlightnfotis> it is <gcc_root>/libgo/runtime/proc.c + <braunr> tschwinge: thanks + <tschwinge> braunr: git://gcc.gnu.org/git/gcc.git + <nlightnfotis> I will provide links on github + <braunr> nlightnfotis: i sayd the whole sources, why do you insist on + giving me separate files ? + <nlightnfotis> for checking it out quickly + <nlightnfotis> oh I misunderstood that sorry + <nlightnfotis> thought you wanted to check out thread creation and + destruction and that you were interested only in those specific files + <braunr> tschwinge: is it completely contained there or are there external + libraries ? + <tschwinge> braunr: You mean libgo? + <braunr> tschwinge: possibly + <nlightnfotis> tschwinge, I just made sure that yeah programs are + dynamically linked against the compiler's libgo + <nlightnfotis> libgo.so.3 + <braunr> does libgo come from gcc sources ? + <nlightnfotis> yeah + <braunr> ok + <nlightnfotis> go files on gcc sources are split under two directories: go, + which contains the frontend go, and libgo which contains the libraries + and the runtime code + <tschwinge> braunr: darnassus:~tschwinge/tmp/gcc/go.build/ is a recent + build, with sources in $PWD/../go/. + <tschwinge> braunr: libgo is in i686-unknown-gnu0.3/libgo/.libs/ + <nlightnfotis> so tschwinge to roundup for this week I should print debug + around the "hotspots" and see if I can extract more information about + where the specific problem is triggered right? + <tschwinge> nlightnfotis: Yes, for a start. + <braunr> nlightnfotis: identify the main thread, make sure it doesn't exit + <nlightnfotis> noted. + <nlightnfotis> braunr: do you have an idea about the issue I described + earlier? The one with the 1 goroutine triggering the bug, but the 2 + exiting successfully but with no output? + <braunr> nlightnfotis: i didn't read + <nlightnfotis> do you have 2 mins to read my report? I describe the issue + <braunr> something messed up in the context i suppose + <tschwinge> nlightnfotis: Uhm, I already explained that issue? + <braunr> you did ? + <nlightnfotis> tschwinge, I know, don't worry. I am trying to get all the + insight I can get. + <nlightnfotis> you mentioned that the scheduler might have an issue and + that the main thread returns before the goroutines execu + <nlightnfotis> *execute + <nlightnfotis> right? + <tschwinge> It is the normal thing for a process to terminate normally when + the main function returns. I would expect Go to behave the same way. + <braunr> "Now, if we change one of the say functions inside main to a + goroutine, this happens" + <braunr> how do you change it ? + <tschwinge> Or am I confused? + <braunr> tschwinge: i don't remember exactly + <nlightnfotis> braunr: from say("world") to go say("world") + <nlightnfotis> tschwinge, yeah I get that. What I still have not understood + is what is it specifically about the 2 goroutines that doesn't trigger + the issu when 1 goroutine does. + <nlightnfotis> You said that it might have something to do with the + scheduler; it does seem like a good explanation to me + <tschwinge> nlightnfotis: My understanding still is that the goroutinges + don't get executed before the main thread exits. + <braunr> which scheduler ? + <nlightnfotis> braunr: the runtime (go) scheduler. + <nlightnfotis> tschwinge, Yeah, they don't. But still, with 1 goroutine: + you get into main, attempt to execute it, and bam! With two, it should be + the same, but strangely it seems to exit main without an issue + <nlightnfotis> (attempt to execute the goroutine) + <braunr> why should it be the same ? + <nlightnfotis> braunr: seeing as one goroutine has problems, I can't see + why two wouldn't. At least one of the two should result in an exception. + <braunr> nlightnfotis: why ? + <braunr> nlightnfotis: they do have the problem + <braunr> they don't run + <braunr> they just don't run into that assertion, probably because there is + more than one thread + <nlightnfotis> wait a minute. You imply that they fail silently? But still + end up in the same situation + <braunr> yes + <braunr> in which case it does look like a go scheduler problem + <nlightnfotis> if I understood it correctly, that assertion fails when it + is only 1 thread? + <braunr> yes + <braunr> and since the main thread is always correct, i expect the main + thread has exited + <braunr> which this happens because the one thread left is *not* the main + thread + <braunr> (which is a libpthread bug) + <braunr> but it's a bug we've not seen because we don't have applications + creating threads while exiting + <nlightnfotis> I think I got it now. + <braunr> try to put something like getchar() in your go program + <braunr> something that introduces a break + <braunr> so that the main thread doesn't exit + <nlightnfotis> oh right. Thanks for that. And sorry tschwinge I reread what + you said, it seems I had misinterpreted what you suggested. + <tschwinge> braunr: If you're interested: for a Go program triggering the + asserition, I don't see any thread exiting (see + darnassus:~tschwinge/tmp/gcc/a.go, run: cd ~tschwinge/tmp/gcc/go.build/ + && ./a.out) -- but perhaps I've been looking for the wrong things in l_. + File l is without a goroutine. Have to leave now, sorry. + <tschwinge> braunr: If you want to rebuild: gcc/gccgo -B gcc -B + i686-unknown-gnu0.3/libgo ../a.go -Li686-unknown-gnu0.3/libgo/.libs + -Wl,-rpath,i686-unknown-gnu0.3/libgo/.libs + <braunr> tschwinge: no i won't touch anything + <braunr> but thanks + + +# IRC, freenode, #hurd, 2013-08-19 + + <youpi> nlightnfotis: how are you going with gcc go? + <nlightnfotis> I was print debugging all the week. + <nlightnfotis> I can tell you I haven't noticed anything weird so far. + <nlightnfotis> But I feel I am close to the solution + <nlightnfotis> I have not written my report yet. + <nlightnfotis> I will write it maximum until wednesday + <nlightnfotis> I hope I will have figured it all out until then + <pinotree> a report is not for writing solutions, but for the progress + <youpi> yes + <youpi> it's completely fine to be saying "I've been debugging, not found + anything yet" + <pinotree> results or not, always write your reports on time, so your + mentor(s) know what you are doing + <nlightnfotis> I see. Would you like me to write it right now, or is it + okay to write it a day or two later? + <hacklu__> nlightnfotis: FYI. this week my report is not finished. just + state some problem I face now. + <youpi> nlightnfotis: I'd say better write it now + <nlightnfotis> youpi: Ok I will write it and tell you when I am done with + it. + <nlightnfotis> youpi: here is my partial report describing what my course + of action looked like this + week. http://www.fotiskoutoulakis.com/blog/2013/08/19/gsoc-week-9-partial-report/ + <nlightnfotis> of course, I will write in a day or two (hopefully having + figured out the whole situation) an exhaustive report describing + everything I did in detail + <nlightnfotis> youpi: I have written my (partial) report describing how I + went about this week + http://www.fotiskoutoulakis.com/blog/2013/08/19/gsoc-week-9-partial-report/ + <youpi> nlightnfotis: good, thanks! + <nlightnfotis> youpi: please note that this is not an exhaustive link of my + findings or course of action, it merely acts as an example to demonstrate + the way I think and how I go about every day. + <nlightnfotis> I will write an exhaustive report of everything I did so + far, when I figure out what the issue is, and I feel I am close. + <youpi> well, you don't need to explain all bits in details + <youpi> this is fine to show an example of how you went + <youpi> but please also provide a summary of your other findings + <nlightnfotis> oh okay, I will keep this in mind. :) + + +# IRC, freenode, #hurd, 2013-08-22 + + < nlightnfotis> if I want to rebuild libpthread, I have to embed it into + eglibc's source, then build? + < pinotree> or pick the debian sources, patch libpthread there and rebuild + < nlightnfotis> that's most likely what I am going to do. Thanks pinotree. + < pinotree> yw + < braunr> nlightnfotis: i usually add my patches on top of the debian glibc + ones, yes + < braunr> it requires some tweaking + < braunr> but it's probably the easiest way + < nlightnfotis> braunr: I was studying my issues with gcc, and everyday I + was getting more and more confident it must be a libpthread issue + < nlightnfotis> and I figured out, that I might wanna play with libpthread + this time + < braunr> it probably is but + < braunr> i'm not so sure you should dive there + < nlightnfotis> why not? + < braunr> because it can be worked around in go + < braunr> i had a test for you last time + < braunr> do you remember what it was ? + < nlightnfotis> nope :/ care to remind it? + < braunr> iirc, it was running the go test you did but with an additional + instruction in the main function, that pauses + < braunr> something like getchar() in c + < braunr> to make sure main doesn't exit while the goroutines are still + running + < braunr> i'm almost positive that the bug you're seeing is main returning + and libpthread beleiving it's acting on the main thread because there is + only one left + < nlightnfotis> oh that's easy, I can do it now. But it's probably what + thomas had suggested: go routines may not be running at all. + < braunr> they probably aren't + < braunr> and that's a context bug + < braunr> not a libpthread bug + < braunr> and that's what you should focus on + < braunr> the libpthread bug is minor + < nlightnfotis> which is strange, because I had studied the assembly code + and it the code for the goroutine was there + < nlightnfotis> anyway I will proceed with what you suggested + < braunr> yes please + < braunr> that's becoming important + < nlightnfotis> would you mind me dumping some of my findings for you to + evaluate/ post on opinion on? + < braunr> no + < braunr> please do so + < nlightnfotis> I have found that the go runtime starts with a total number + of threads == 1 + < braunr> nlightnfotis: as all processes + < nlightnfotis> I would guess that's because of using fork () + < nlightnfotis> oh so it's ok + < braunr> there always is a main thread + < braunr> even for non-threaded applications + < nlightnfotis> yeah, that I know. The runtime proceeds to create + immediately one more. + < braunr> then it's 2 + < nlightnfotis> and that's ok, it doesn't have an issue with that + < nlightnfotis> yep + < nlightnfotis> the issue begins when it tries to create the 3rd one + < braunr> hum + < braunr> from what i remember + < nlightnfotis> it happily goes through the go runtime's kernel thread + allocation function (runtime_newm()) + < braunr> you also had an issue with the first goroutine + < nlightnfotis> that's with 1 go routine + < braunr> ok + < braunr> so 1 goroutine == 3 threads + < nlightnfotis> it seems so yes. + < braunr> depending on how the go scheduler is able to assign goroutines to + kernel threads i suppose + < nlightnfotis> mind you, (disclaimer: I am not so sure about that) that go + must be using one extra thread for the runtime scheduler and garbage + collector + < braunr> that's ok + < nlightnfotis> so that's where the two come from + < braunr> and expected from a modern runtime + < nlightnfotis> the third must be the go routime + < nlightnfotis> routine + < braunr> hum have to go + < braunr> brb in a few minutes + < braunr> keep posting + < nlightnfotis> it's ok take your time + < nlightnfotis> I will be here + < braunr> but i may not ;p + < braunr> in fact i will not + < braunr> i have like 15 mins ;) + < braunr> nlightnfotis: ^ + < nlightnfotis> I am trying what you told me to do with go + < nlightnfotis> it's ok if you have to go, I will continue investigating + and be back tomorrow + < braunr> ok + < nlightnfotis> braunr: I tried what you asked me to do, both we waiting to + read a string from stdin and with waiting to read an int from stdin + < nlightnfotis> it never waits, it still aborts with the assertion failure + < nlightnfotis> both with one and two go routines + < nlightnfotis> dumping it here just for the log, running the same code + without waiting for input results in two threads created (1 for main and + 1 for runtime, most likely) and "normal" execution. + < nlightnfotis> normal as in no assertion failure, + < nlightnfotis> it seems to skip the goroutines altogether + + +# IRC, freenode, #hurd, 2013-08-23 + + < braunr> nlightnfotis: can i see your last go test code please ? the one + with the read at the end of main + < nlightnfotis> braunr sure + < nlightnfotis> sorry I had gone to the toilet, now I am back + < nlightnfotis> I will send it right now + < nlightnfotis> braunr: http://pastebin.com/DVg3FipE + < nlightnfotis> it crashes when it attempts to create the 3rd thread (the + 1st goroutine), with the assertion fail + < nlightnfotis> if you remove the Scanf it will not fail, return 0, but + only create 2 threads (skip the goroutines alltogether) + < braunr> can you add a print right before main exits please ? + < braunr> so we know when it does + < nlightnfotis> doing it now + < nlightnfotis> braunr: If I enter a print statement right before main + exits, the assertion failure is triggered. If I remove it, it still runs + and creates only 2 threads. + < braunr> i don't understand + < braunr> 14:42 < nlightnfotis> it crashes when it attempts to create the + 3rd thread (the 1st goroutine), with the assertion fail + < braunr> why don't you get that ? + < nlightnfotis> This seems like having to do with the runtime. I mean, I + have seen the emitted assembly from the compiler, and the goroutines are + there. Something in the runtime must be skipping them + < braunr> context switching seems buggy + < nlightnfotis> if it's only goroutines in main + < nlightnfotis> if there's also something else in main, the assertion + failure is triggered. + < braunr> i want you to add a printf right before main exits, from the code + you pasted + < nlightnfotis> I did. It acts the same as before. + < braunr> do you see that last printf ? + < nlightnfotis> no. It aborts before that + < nlightnfotis> :q + < braunr> find a way to make sure the output buffer is flushed + < braunr> i don't know how it's done in go + < nlightnfotis> mistype the :q, was supposed to do it vim + < nlightnfotis> braunr will do right away + < nlightnfotis> there is one thing I still can not understand: Why is it + that two threads are ok, but when the next is going to get created, the + assertion is triggered. + < braunr> nlightnfotis: the assertion is triggered because a thread is + being created while there is only one thread left, and this thread isn't + the main thread + < braunr> so basically, the main thread has exited, and another (the last + one) is trying to create one + < nlightnfotis> the other one might be the runtime I guess. Let me check + out quickly what you suggested + < braunr> the main thread shouldn't exit at all + < braunr> so something with context switching is wrong + < nlightnfotis> the thing is: it doesn't seem to exit when this happens. My + debug statements (in the runtime) suggest that there are at least 2 + threads active, kernel threads don't get destroyed in gccgo + < braunr> 14:52 < braunr> so something with context switching is wrong + < braunr> how well have the context switching functions been tested ? + < nlightnfotis> to be honest I have not tested them; up until this point I + trusted they worked. Should I also take a look at them? + < braunr> how can you trust them ? + < braunr> they've never been used .. + < braunr> thomas added them recently if i'm right + < braunr> nothing has been using them except go + < braunr> piece of advice: don't trust anything + < nlightnfotis> I think they were in before, and thomas recently patched + them! + < braunr> they were in, but didn't work + < braunr> (if i'm right) + < braunr> nlightnfotis: you could patch libpthread to monitor the number of + threads + < braunr> or the go runtime, idk + < nlightnfotis> I have done so on the go runtime + < nlightnfotis> that's where I am getting the number of threads I + report. That's straight out from the scheduler's count. + < braunr> threads can exit by calling pthread_exit() or returning from the + thread routine + < braunr> make sure you catch both + < braunr> also check for pthread_cancel(), although i don't expect any in + go + < nlightnfotis> braunr: Should I really do that? I mean, from what I can + see in gccgo's comments, Kernel threads (m) never go away. They are added + to a pool of m's waiting for work if there is no goroutine running on + them + < nlightnfotis> I mean, I am not so sure they exit at all + < braunr> be sure + < braunr> point me the code please + < nlightnfotis> + https://github.com/NlightNFotis/gcc/blob/master/libgo/runtime/proc.c#L224 + < nlightnfotis> this is where it get's stated that m's never go away + < nlightnfotis> and at line 257 you can see the pool + < nlightnfotis> and wait for me to find the code that actually releases an + and places into the pool + < nlightnfotis> yep found it + < nlightnfotis> line 817 mput + < nlightnfotis> puts a kernel thread given as parameter to the pool + < nlightnfotis> another proof of the theory is at line 1177. It states: + "This point is never reached, because scheduler does not release os + threads at the moment." + < braunr> fetching git repository, bit busy, i'll have a look in 5-10 mins + < nlightnfotis> oh it's ok, I had pointed you to the file directly on + github to check it out instantly, but never mind, the file is + <gccroot>/libgo/runtime/proc.c + < braunr> damn github is so slow .. + < braunr> nlightnfotis: i much prefer my own text interface :) + < nlightnfotis> braunr: just out of curiosity what's your setup? I use vim + mainly (not that I am a vim expert or anything, I only know the basics, + but I love it) + < braunr> same + < braunr> nlightnfotis: add a trace at that comment to make SURE threads do + not exit + < braunr> you *cannot* get the libpthread assertion with more than 1 thread + < braunr> grep for pthread_exit() too + < nlightnfotis> will do it now. It will take about an hour to compile + though. + < braunr> i don't understand the stack trick at the start of runtime_mstart + < braunr> ah splitstack .. + < nlightnfotis> I think I should try cross compiling gcc, and then move + files on the hurd. It would be so much faster I believe. + < braunr> than what ? + < nlightnfotis> building gcc on the hurd + < nlightnfotis> I remember it taking about 10minutes with make -j4 on the + host + < nlightnfotis> it takes 45-50 minutes on the vm (kvm enabled) + < braunr> but you can merely rebuild the files you've changed + < nlightnfotis> I feel stupid now... + < braunr> nlightnfotis: have you tried setting GOMAXPROCS to 1 ? + < nlightnfotis> not really, but from what I know GOMAXPROCS defaults to 1 + if not set + < braunr> again, check that + < braunr> take the habit of checking things + < nlightnfotis> braunr: yeah sorry for that. I have checked these things + out before they don't come out of my head I just don't remember exactly + where I had seen this + < braunr> what you can also do is use gdb to catch the assertion and check + the number of threads at that time, as well as the number of threads as + seen by libpthread + < nlightnfotis> braunr: line 492 file proc.c: runtime_gomaxprocs = 1; + < braunr> also see runtime.LockOSThread + < braunr> to make sure the main thread is locked to its own pthread + < nlightnfotis> I can see in line 529 of the same file that the first + thread is getting locked + < nlightnfotis> the new threads that get initialised are non main threads + < braunr> if(!runtime_sched.lockmain) runtime_UnlockOSThread(); + < braunr> i'm suggesting you set runtime_sched.lockmain + < braunr> so it remains true for the whole execution + < braunr> this code looks like a revamp of plan9 lol + < nlightnfotis> it is + < nlightnfotis> in the paper from Ian Lance Taylor describing gccgo he + states somewhere that the original go compilers (the 3gs) are a modified + version of plan9's C compiler, and that gccgo tries to follow them + < nlightnfotis> they differ in a lot of ways though + < nlightnfotis> the 3gs generate a lot of code during link time + < nlightnfotis> gccgo follows the standard gcc procedures + < braunr> eh :D + < nlightnfotis> go -> gogo -> generic -> gimple -> rtl -> object + < nlightnfotis> that's how it flows as far as I recall + < nlightnfotis> gogo is an internal representation of go's structure inside + the gccgo frontend + < nlightnfotis> that's why you see many functions with gogo in their name + < nlightnfotis> I just revisited the paper: gogo is there to make it easy + to implement whatever analysis might seem desirable. It mirrors however + the Go source code read from the input files + < braunr> nlightnfotis: what are you trying now ? + < nlightnfotis> I am basically studying the runtime's source code while + waiting for gccgo to compile on the Hurd + < nlightnfotis> yes I did the stupid whole recompilation again. :/ + < braunr> nlightnfotis: compile for what ? + < braunr> what test ? + < nlightnfotis> to check out to see if M's really are added to the pool + instead of getting deleted + < braunr> nlightnfotis: but how ? + < nlightnfotis> braunr: I have added a statement in mput if we get there + first, and secondly the number of threads that the runtime scheduler + knows that are waiting (are in the pool of m's waiting for work) + < braunr> ok + < braunr> when you can, i'd really like you to do this test : + < braunr> 15:55 < braunr> what you can also do is use gdb to catch the + assertion and check the number of threads at that time, as well as the + number of threads as seen by libpthread + < nlightnfotis> the number of threads required by libpthread is gonna need + me to recompile the whole eglibc right? + < braunr> no + < braunr> just print it with gdb + < nlightnfotis> oh, ok + < braunr> it's __pthread_num_threads + < nlightnfotis> is gdb reliable? I remember thomas telling me that I can't + trust gdb at this point in time + < braunr> and also __pthread_total + < braunr> really ? + < braunr> i don't see why not :/ + < braunr> youpi: any idea about what nlightnfotis is speaking of ? + < nlightnfotis> I may have misunderstood it; don't take it by heart + < nlightnfotis> I don't wanna put words in other people's mouths because I + misunderstood something + < braunr> sure + < braunr> that's my habit to check things + < youpi> braunr: nope + < braunr> youpi: and am i right when i say we don't use context functions + on the hurd, and they're likely to be incomplete, even with the recent + changes from thomas ? + < braunr> (mcontext, ucontext) + < nlightnfotis> braunr: this is what had been said: 08:46:30< tschwinge> As + this is a parallel problem, and it is involving "advanced" things (such + as setcontext), I would not trust GDB too much when used on this code. + < pinotree> if thomas' changes were complete and polished, i guess he would + have sent them upstream already + < braunr> i see but + < braunr> you can normally trust gdb for global variables + < nlightnfotis> Didn't post it as an objection; I posted it because I felt + bad putting the wrong words on other people's mouths, as I said + before. So I posted his original comment which was more authoritative + than my interpretation of it + < braunr> i wonder if there is a tunable to strictly map one thread to one + goroutine + < braunr> nlightnfotis: more focus on the work, less on the rest please + < nlightnfotis> Did I do something wrong? + < braunr> you waste too much time apologizing + < braunr> for no reason + < braunr> nlightnfotis: i suppose you don't use splitstack, right ? + < nlightnfotis> no I didn't + < nlightnfotis> and here's something interesting: The code I just added, in + mput, to see if threads are added in the pool. It's not there, no matter + what I run + < nlightnfotis> So it seems that we the runtime is not reaching mput. + < nlightnfotis> Could this be normal behavior? I mean, on process + termination just release the resources so mput is skipped? + < braunr> i don't know the code well enough to answer that + < braunr> check closer to the lower interface + + +# IRC, freenode, #hurd, 2013-08-25 + + < nlightnfotis> braunr: what is initcontext supposed to be doing? + < braunr> nlightnfotis: didn't look + < braunr> i'll take a look later + < nlightnfotis> braunr: I am buffled by it. It seems to be doing nothing on + the Hurd branch and nothing in the Linux branch either. Why call a + function that does nothing? (it doesn't only seem to do nothing, I have + confirmed it) + < nlightnfotis> youpi: I was wondering if you could explain me + something. What is the initcontext function supposed to be doing? + < youpi> you mean initcontext ? + < nlightnfotis> yes + < youpi> ergl + < youpi> you mean makecontext? + < nlightnfotis> no initcontext. I am faced with this in the goruntime. It's + called in it, but it is doing nothing. Neither in the Hurd tree, nor in + the Linux one + < youpi> I don't know what initcontext is + < youpi> where do you read it? + < nlightnfotis> youpi: let me show you + < nlightnfotis> + https://github.com/NlightNFotis/gcc/blob/fotisk/goruntime_hurd/libgo/runtime/proc.c#L80 + < nlightnfotis> and it is called in quite a few places + < youpi> it's not doing nothing, see other implementations + < pinotree> if SETCONTEXT_CLOBBERS_TLS is not defined, initcontext and + fixcontext do nothing + < pinotree> otherwise (presuming if setcontext clobbers tls) there are two + implementations for solaris/x86_64 and netbsd + < youpi> I don't think we have the tls clobber bug + < youpi> so these functions being empty is completely fine + < nlightnfotis> pinotree: oh, you mean it's used as a workaround for these + two systems only? + < youpi> yes + < pinotree> yes + < nlightnfotis> That makes sense. Thanks both of you for the help :) + < nlightnfotis> youpi: if this counts as some progress, I have traced the + exact bootstrapping sequence of a new go process. I know a good deal of + what is done from it's spawn to it's end. There are some things I wanna + sort out, and later tonight I will write my report for it to be ready for + tomorrow. + < youpi> good + + +# IRC, freenode, #hurd, 2013-08-26 + + < nlightnfotis> Hi everyone, my report is here + http://www.fotiskoutoulakis.com/blog/2013/08/26/gsoc-week-10-report/ + < youpi> nlightnfotis: you should clearly put printfs inside libpthread + < youpi> to check what is happening with the ktids + < nlightnfotis> youpi: yep, that's my next course of action. I just want to + spend some more time in the go runtime to make sure that I understand the + flow perfectly, and to make sure that it is not the runtime's fault + < braunr> nlightnfotis: did you try gdb to print the number of threads ? + < youpi> nlightnfotis: to build it, the easiest way is to start building + eglibc, and when you see it compiling C files (i.e. run i486-gnu-gcc-4.7 + etc.) + < youpi> stop it + < youpi> and go into build/hurd-i386-libc, and run "make others" from there + < nlightnfotis> braunr: that was my plan for today or tomorrow :) + < braunr> start building *debian* glibc + < youpi> there's perhaps some way to only build libpthread, but I don't + remember + < braunr> nlightnfotis: ok + < braunr> youpi: i suggested he tried gdb first + < youpi> why not + < braunr> if you need quick glibc builds, you can use darnassus + < nlightnfotis> braunr: how much time on average should I expect it to + take? + < youpi> it highly depends on the machine + < youpi> it can be hours + < youpi> or a few minutes + < youpi> depending you already have a built tree, a fast disk, etc. + < braunr> make lib others on darnassus takes around 30 minutes + < braunr> a complete dpkg-buildpackage from fresh sources takes 5-6 hours + < braunr> make others from a built tree is very quick + < braunr> a few minutes at most + < braunr> nlightnfotis: i don't see any trace of thread exiting in your + report, is that normal ? + < nlightnfotis> yeah, I guess, since they don't exit prematurely, they are + released along with other resources at the process' exit + < braunr> i'll rephrase + < braunr> you said last time that you saw a function never got called + < braunr> i assumed it was because a thread exited prematurely + < nlightnfotis> oh I sorted it out with the help of youpi and pinotree + yesterday + < braunr> that's different + < braunr> i'm not talking about the function that does nothing + < braunr> i'm talking about the one never called + < nlightnfotis> oh, go on then, + < braunr> i don't remember its name + < braunr> anyway + < nlightnfotis> abort()? + < braunr> i hope abort doesn't get called :) + < nlightnfotis> it doesn't + < braunr> i thought it was the one right before + < braunr> what i mean is + < nlightnfotis> oh runtime_mstart, it does get called + < braunr> add traces at thread exit points + < nlightnfotis> I sorted it out too + < braunr> make *sure* threads don't exit + < nlightnfotis> it get's called to start the kernel thread created at + process spawn at the runtime_schedinit + < braunr> if they really don't, it's probably a context/tls issue + < nlightnfotis> I will do this right now. + < nlightnfotis> braunr: if it's a context/tls issue it's libpthread's + problem? + + +# IRC, freenode, #hurd, 2013-09-02 + + <nlightnfotis> Hello! My report for this week is online: + http://www.fotiskoutoulakis.com/blog/2013/09/02/gsoc-week-11-report/ + <braunr> nlightnfotis: there always is a signal thread in every hurd + program + <braunr> nlightnfotis: i also pointed out that there are two variables + involved in counting threads in libpthread, the other one being + __pthread_num_threads + <braunr> again, more attention to work and details, less showmanship + <braunr> i'm tired of repeating it + <youpi> nlightnfotis: doesn't backtrace work in gdb to tell you what + 0x01da48ec is? + <youpi> also, do you have libc0.3-dbg installed? + <nlightnfotis> braunr: __pthread_num_threads reports is 4. + <braunr> then why isn't it in your report ? + <braunr> it's acceptable that you overlook it + <nlightnfotis> and youpi: yeah I have got the backtrace, but 0x01da48ec is + ?? () from /lib/i386-gnu/libc.so.3 + <braunr> it's NOT when someone else has previously mentioned it to you + <youpi> nlightnfotis: only that line, no other line? + <nlightnfotis> it has 8 more youpi, the one after ?? is mach_msg () + form/lib/gni386-gnu/libc.so.0.3 + <braunr> yes mach_msg + <braunr> almost everything ends up in mach_msg + <youpi> you should probably pastebin somewhere the output of thread apply + all bt + <braunr> what's before that ? + <nlightnfotis> braunr: I don't know how I even missed it. I skimmed through + the code and only found __pthread_total and assumed that it was the total + number of threads + <braunr> nlightnfotis: i don't know either + <braunr> take notes + <nlightnfotis> before mach_msg ins __pthread_timedblock () from + /lib/i386-gnu/libpthread.so.0.3 + <nlightnfotis> I will add it to pastebin in a second + <braunr> i find it very disappointing that after several weeks blocking on + this, despite all the pointers you've been given, you still haven't made + enough progress to reach the context switching functions + <braunr> last week, most progress was made when we talked together + <braunr> then nothing + <braunr> it seems that you disappear, apparently searching on your own + <braunr> but for far too long + <nlightnfotis> braunr: I do search on my own, yes, + <braunr> almost like exploiting being blocked not to make progress on + purpose ... + <braunr> but too much + <nlightnfotis> braunr: I am not doing this on purpose, I believe you are + unfair to me. I am trying to make as much progress as I can alone, and + reach out only when I can't do much more alone + <braunr> then why is it only now that we get replies to questions such as + "how much is __pthread_num_threads" ? + <braunr> why do you stop discussions for almost a week, just to find + yourself blocked again ? + <nlightnfotis> I was working on gcc, going through the runtime making sure + about assumptions and going through various other goroutine or not + programs through gdb + <braunr> that doesn't take a week + <braunr> clearly not + <braunr> last time we talked was + <braunr> 10:40 < nlightnfotis> braunr: if it's a context/tls issue it's + libpthread's problem? + <nlightnfotis> it did for me... honestly, what is it you believe I am doing + wrong? I too am frustrated by my lack of progress, but I am doing my best + <braunr> august 26 + <nlightnfotis> yeah, I wanted to make sure about certain assumptions on the + gcc side. I don't want to start hacking on libpthread only to see that it + might have been something I msissed on the gcc side + <braunr> i told you + <braunr> it's probably not a libpthread issue + <braunr> the assertion is + <braunr> but it's minor + <braunr> it's not the realy problem, only a side effect + <braunr> i told you about __pthread_num_threads, why didn't you look at it + ? + <braunr> i told you about context switching functions, why nothing about it + ? + <braunr> doing a few printfs to check numbers and using gdb to check them + at break points should be quick + <braunr> when we talk,ed we had the results in a few minutes + <nlightnfotis> yeah, because I was guided, and that helped me target my + research. On my own things are quite different. I find out something + about gcc's behavior, then find out I need tons more information, and I + have a lot of things that I need to research to confirm any assumptions + from my side + <braunr> how did you miss the signal thread ? + <braunr> we even talked about it right here with hacklu + <braunr> i'll say it again + <braunr> if blocked more than one day, ask for help + <braunr> 2 days minimum each time is just too long + <nlightnfotis> I'm sorry. I will be online every day from now on and report + every 10 minutes, on my course of actions. + <nlightnfotis> I recognise that time is off the essence at this point in + time + <braunr> it's also NO + <braunr> NO + <braunr> *SIGH* + <hacklu> nlightnfotis: calm down. braunr just want to help you solve + problem quickly. + <braunr> 10 minutes is the other extreme + <hacklu> nlightnfotis: in my experiecence, if something block me, I will + keep asking him until I solve the problem. + <braunr> it's also very frustrating to see you answer questions quickly + when you're here, then wait days for unanswered questions that could have + taken little time if you kept being here + <braunr> this just gives the impression that you're doing something else in + parallel that keeps you busy + <braunr> and comfort me in believing you're not being serious enough + aboutit + <nlightnfotis> yeah, I understand that it gives that impression. The only + thing I can tell you now, is that I am *not* doing something else in + parallel. I am only trying to demonstrate some progress alone, and when + working alone things for me take quite some more time than when I am + guided + <braunr> hacklu: i'm actually the nervous one here + <nlightnfotis> braunr: ok, I understand I have dissapointed you. What would + you suggest me to do from now on? + <hacklu> braunr: :) + <braunr> manage your time correctly or you'll fail + <braunr> i'm not the main mentor of this project so it's not for me to + decide + <braunr> but if i were, and if i had to wait again for several days before + any notice of progress or blocking, i wouldn't even wait for the end of + the gsoc + <braunr> you're confronted with difficult issues + <braunr> tls, context switching, thread + <braunr> ing + <braunr> they're all complicated + <braunr> unless you're very experienced and/or gifted, don't assume you can + solve it on your own + <braunr> and the biggest concern for me is that it's not even the main + focus of your project + <braunr> you should be working on go + <braunr> on porting + <braunr> any side issues should be solved as quickly as possible + <braunr> and we're now in september ... + <nlightnfotis> go is working quite alright. It's goroutines that have + issues. + <braunr> nlightnfotis: same thing + <braunr> goroutines are part of go as far as i'm concerned + <braunr> and they're working too, something in the hurd isn't + <braunr> so it's a side issue + <braunr> you're very much entitled to ask as much help as you need for side + issues + <braunr> and i strongly feel you didn't + <nlightnfotis> yeah, you're right. I failed on that aspect, mainly because + of the way I work. I wanted to show some progress on my own, and not be + here and spam all day. I felt that spamming questions all day would + demonstrate incompetence from my side + <nlightnfotis> and I wanted to show that I am capable of solving my + problems on my own. + <braunr> well, in a sense it does, but that's not the skills we were + expecting from you so it's perfectly ok + <braunr> nlightnfotis: no development group, even in companies, in their + right mind, would expect you to grasp the low level dark details of an + operating system implementation in a few weeks ... + <nlightnfotis> braunr: ok, may I ask what you suggest to me that my next + course of action is? + <braunr> let me see + <braunr> nlightnfotis: your report mentions runtime_malg + <nlightnfotis> yes, I runtime malg always returns a new goroutine + <braunr> nlightnfotis: what's the problem ? + <nlightnfotis> a new m created is assigned a new goroutine via runtime_malg + <nlightnfotis> what happens to that goroutine? Is it destroyed? Because it + seems to be a bogus goroutine. Why isn't the kernel thread instantly + picking the one goroutine available at the global goroutine pool? + <braunr> let's see if it's that hard to figure out + <nlightnfotis> seeing as m's and g's have a 1:1 (in gccgo) relationship, + and a new kernel thread is created everytime there is a new goroutine + there to run. + <braunr> are you sure about that 1:1 relationship ? + <braunr> i hardly doubt it + <braunr> highly* + <nlightnfotis> yeah, that's what I thought too, but then again, my research + so far shows that when a new goroutine is created, a new kernel thread + creation follows suit + <nlightnfotis> what I have mentioned of course, happens in runtime_newm + <braunr> nlightnfotis: that's when you create a new m, not a new g + <nlightnfotis> yes, a new m is created when you create a new g. My issue is + that during m's creation, a new (bogus) g is created and assigned to the + m. I am looking into what happens to that. + <braunr> nlightnfotis: "a new m is created when you create a new g", can + you point me to the code ? + <nlightnfotis> braunr: matchmg line 1280 or close to that. Creates new m's + to run new g's up to (mcpumax) + <braunr> "Kick off new m's as needed (up to mcpumax)." + <braunr> so basically you have at most mcpumax m + <nlightnfotis> yeah. but for a small number of goroutines (as for example + in my experiments), a new m is created in order to run a new g. + <braunr> runtime_newm is called only if mget(gp)) == nil + <braunr> be rigorous please + <braunr> when i ask + <braunr> 11:01 < braunr> are you sure about that 1:1 relationship ? + <braunr> this conclusively proves it's *false* + <braunr> so don't answer yes to that + <braunr> it's true for a small number of goroutines, ok + <braunr> and at startup + <braunr> because then, mget returns an existing m + <braunr> nlightnfotis: this g0 goroutine is described in the struct as + <braunr> G runtime_g0; // idle goroutine for m0 + <braunr> runtime_malg builds it with just a stack + <braunr> apparently, that's the goroutine an m runs when there are no g + left + <braunr> so yes, the idle one + <braunr> it's not bogus + <nlightnfotis> I thought m0 and g0 where the bootstrap m and g for the + scheduler. + <nlightnfotis> *correction: runtime_m0 and runtime_g0 + <braunr> hm i got a bit fast + <braunr> G* g0; // goroutine with scheduling stack + <nlightnfotis> braunr: scheduling stack with stacksize = -1? + <nlightnfotis> unless it's not used as a parameter + <nlightnfotis> let me investigate that + <nlightnfotis> yeah now that I am seeing it, it might make sense, if it + using a default stack size, #defined as StackMin + <braunr> g0 looks like a placeholder + <braunr> i think it's used to reuse switching code when there is only one + goroutine involved + <braunr> e.g. when starting + <braunr> anyway i don't think we should waste too much time with it + <braunr> nlightnfotis: try to make a real 1:1 mapping + <braunr> that's something else i suggested last time + <nlightnfotis> braunr: ok. Where do you suspect the problem lies? + <braunr> context switching + <nlightnfotis> inside the goruntime? + <braunr> in glibc + <braunr> try to use runtime.LockOSThread + <braunr> http://code.google.com/p/go-wiki/wiki/LockOSThread + <braunr> nlightnfotis: http://golang.org/pkg/runtime/ is probably better + <nlightnfotis> what exactly do you mean by `use runtime.LockOSThread`? + LockOSThread locks the very first m and goroutine as the main threads + during process initialisation + <nlightnfotis> in proc.c line 565 or something + <braunr> i'm not sure it will help, because the problem is likely to occur + before even switching to the goroutine that locks its m, but worth trying + <braunr> 11:28 < braunr> nlightnfotis: http://golang.org/pkg/runtime/ is + probably better + <braunr> the first example is specific to GUIs that have requirements on + the main thread + <braunr> whereas i want every goroutine to run in its own thread + <nlightnfotis> I have also noticed that some context switching happens in + the goruntime even with a low number of goroutines and kernel threads + <braunr> that's expected + <braunr> goroutines must be viewed as works, and ms as worker threads + <braunr> everytime a goroutine sleeps, its m should be switching to useful + work + <braunr> nlightnfotis: i'd make prints (probably using mach_print) of + contexts when saved and restored + <braunr> and try to see if it makes any sense + <braunr> that's not simple to setup but not overly complicated either + <braunr> don't hesitate to ask for help + <nlightnfotis> from inside glibc, right? + <braunr> yes + <braunr> well + <braunr> no from go + <braunr> don't touch glibc from now + <braunr> put these prints near calls to makecontext/swapcontext + <braunr> and setcontext/getcontext + <braunr> wel + <braunr> you'll be using getcontext i think + <nlightnfotis> noted it all. I also have the gdb output you asked me for + http://pastebin.com/LdnMQDh1 + <braunr> i don't see main + <nlightnfotis> some notes first: The main thread is the one with id 4, and + the output on the top is its backtrace. + <braunr> and main.main is run in thread 6 + <nlightnfotis> Remember that main when it comes to go is in the file + go-main.c + <braunr> so main becomes runtime_MHeap_Scavenger + <nlightnfotis> yeah, main.main is the code of the program, (the one the + user wrote, not the runtime) + <nlightnfotis> yeah, it becomes a gc thread + <nlightnfotis> seeing as runtime_starttheworld reports that there is + already one gc thread + <braunr> and how much are __pthread_total and __pthread_num_threads for + that trace ? + <nlightnfotis> they were: __pthread_total = 2, and __pthread_num_threads = + 4 + <braunr> can you paste the assertion again please, just to make sure + <nlightnfotis> a.out: ./pthread/pt-create.c:167: __pthread_create_internal: + Assertion `({ mach_port_t ktid = __mach_thread_self (); int ok = + thread->kernel_thread == ktid; + <nlightnfotis> __mach_port_deallocate ((__mach_task_self + 0), ktid); ok; + })' failed. + <braunr> btw, install the -dbg packages too + <nlightnfotis> dbg for which one? gccgo? + <braunr> libc0.3 + <braunr> pthread/pt-create.c:167 is __pthread_sigstate (_pthread_self (), + 0, 0, &sigset, 0); here :/ + <braunr> that assertion should be in __pthread_thread_start + <braunr> let's just say gdb is confused + <pinotree> braunr: apt-get source eglibc ; cd eglibc-* ; debian/rules patch + <braunr> pinotree: i have + <braunr> and that assertion can only trigger if __pthread_total is 1 + <braunr> so let's say it just got to 2 + <nlightnfotis> it does from very early on in process initialisation + <nlightnfotis> let me check this out again + <braunr> hm + <braunr> actually, both __pthread_total and __pthread_num_threads must be 1 + <braunr> the context functions might be fine actually + <nlightnfotis> braunr: __pthread_num_threads = 2 right from the start of + the program + <nlightnfotis> 0x01da48ec is in mach_msg_trap + <braunr> something happened with libpthreads recently .. + <braunr> i can't even start iceweasel + <pinotree> braunr: what's the error? + <braunr> iceweasel: ./pthread/../sysdeps/generic/pt-mutex-timedlock.c:70: + __pthread_mutex_timedlock_internal: Assertion `__pthread_threads' failed. + +But not the [[open_issues/libpthread_dlopen]] issue? + + <braunr> considering __pthread_threads is a global variable, this is tough + <braunr> i wonder if that's the issue with nlightnfotis's work + <braunr> wrong symbol resolution, leading libpthread to consider there is + only one thread running + <pinotree> try with LD_PRELOAD=/lib/i386-gnu/libpthread.so.0 iceweasel + <braunr> same + <braunr> maybe the switch to glibc 2.17 + <braunr> this assertion is triggered by __pthread_self, assert + (__pthread_threads); + <braunr> __pthread_threads being the array of thread pointers + <braunr> so either corrupted (but we hardly changed anything ...) or wrong + resolution + <braunr> __pthread_num_threads includes the signal thread, __pthread_total + doesn't + <nlightnfotis> braunr: I recompiled with the libc debugging symbols and I + have new information + <nlightnfotis> the threads block at mach_msg_trap + <braunr> again, almost everything blocks there + <braunr> mach_msg is mach ipc, the way hurd system calls are implemented + <nlightnfotis> and the next calls (if it didn't block, from what I can see + from eip) are mach_reply_port and mach_thread_self + <braunr> please paste it + <nlightnfotis> yes give me 2 mins plz, brb + <braunr> pinotree: looks different for firefox + <braunr> it seems it calls pthread_key_create before pthread_create + <braunr> something our libpthread doesn't handle correctly + <nlightnfotis> braunr: http://pastebin.com/yNbT7nLn + <pinotree> braunr: what do you mean? + <braunr> pinotree: i mean libpthread needs to be fixed so thread-specific + data can be set even without a call to pthread_create + <braunr> nlightnfotis: hum, we already knew it was blocking in a semaphore + <braunr> nlightnfotis: ok forget the other things i told you to test + <braunr> nlightnfotis: track __pthread_total and __pthread_num_threads + <braunr> add prints (again, with mach_print) to see when (and why) they + change and go back to 1 + <pinotree> braunr: i see that pthread_key_create uses a mutex which in + turns needs _pthread_self(), but shouldn't at least one pthread_create be + done (directly by libc for the main thread)? + <braunr> pinotree: no :) + <braunr> well + <braunr> it should have been for the signal thread indeed + <braunr> and the signal thread exists + <pinotree> and the main thread? + <braunr> not the main, no + <pinotree> how so? + <braunr> a simple test program shows it does indeed work .. + <braunr> so this is again another problem in firefox too + <nlightnfotis> braunr: I don't think I understand this. I mean how can + pthread_total and __pthread_num_thread turn to 1, when , right before and + right after the crash they have numbers between 2, 3, and 4? + <braunr> how did you get their values "right before" the crash ? + <nlightnfotis> I have set a breakpoint to a printing function right before + the go statement + <nlightnfotis> (right before in this context, in the application code, not + the runtime code, but then again, I don't really think they are too far + each other) + <braunr> well, that's the mystery + <nlightnfotis> I am not challenging what you said, I will of course do, + just asking to understand some things + <braunr> they may either turn to 1, or there is some mess with symbol + resolution leading threads to see a value of 1 + <nlightnfotis> *do it + <braunr> there* + <nlightnfotis> braunr: ping + <teythoon> just ask ;) + <nlightnfotis> teythoon: have you used mach_print? + <teythoon> no + <nlightnfotis> I have some questions about it + <teythoon> ask them + <nlightnfotis> I was told to use them inside go's runtime, to print the + values of __pthread_total and __pthread_num_threads. The thing is, these + values (I believe) are unknown to the runtime, they are only known to the + executable (linking time and later) + <teythoon> so? if the requested information is bound to a symbol that is + resolved at link time, you can print it from within the runtime + <teythoon> the same way any function from the libc is not known to the + executable until linking against it, but you can still "use" it in your + executable + <nlightnfotis> yeah, ok I understand that, but these are references that + are resolved at link time. The values I want to print are totally unknown + to the runtime (0 references to them) + <teythoon> if the value you are interested in is bound to the symbol + __pthread_total at link time, then you've got a reference you can use + <teythoon> doesn't printing __pthread_total work? did you try that? + <nlightnfotis> no, whenever I printed these values I did it from gdb. I am + trying to do what you suggested atm + <braunr> nlightnfotis: im here + <braunr> printing those values from libgo will tell us what value libgo + actually sees + <nlightnfotis> I am trying to use mach_print. Could you give me some + pointers on its usage (inside the goruntime?) (I have already read your + document here + http://www.gnu.org/software/hurd/microkernel/mach/gnumach/interface/syscall/mach_print.html + and the example code) + <braunr> and symbol resolution may depend on where it's done from + <braunr> nlightnfotis: first, it only work with -dbg kernels + <braunr> so make sure you're running one + <braunr> actually, i'll write you a patch + <braunr> including a mach_printf function with argument parsing + <nlightnfotis> isn't it on by default? I read that on the document you are + discussing mach_printf + <nlightnfotis> ahh ok + <braunr> it's on by default on -dbg kernels + <braunr> i'll make a repository on darnassus too + <braunr> better store it there + <braunr> nlightnfotis: + http://darnassus.sceen.net/gitweb/rbraun/mach_print.git/ + <braunr> nlightnfotis: i suggest you implement mach_print with inline asm + statement in a C file, so that you don't need to alter the build system + configuration + <braunr> i'll make an example of that too + <nlightnfotis> braunr: that wasn't a problem. My only real problem atm is + that __atomic_t isn't recognised as a type, and I can not find the header + file for it on Hurd + <nlightnfotis> it was pt-internal.h in libpthread + <braunr> ah + <braunr> nlightnfotis: just in case, i updated the repository with an + inline assembly version + <braunr> let's see about __atomic_t + <braunr> sysdeps/i386/bits/pt-atomic.h:typedef __volatile int __atomic_t; + <braunr> nlightnfotis: just redeclare it as this locally + <braunr> nlightnfotis: ok ? + <nlightnfotis> I am working on it, because I still haven't found what + __atomic_t is typedefed from. Thinking of typedefing an int to it and see + how it goes + <nlightnfotis> braunr: found it just now: __volatile int + <braunr> "just now" ? + <braunr> 14:19 < braunr> sysdeps/i386/bits/pt-atomic.h:typedef __volatile + int __atomic_t; + <nlightnfotis> I was using cscope all this time + <braunr> why use cscope at all when i tell you where it is ? + <nlightnfotis> because I didn't notice it: your discussion was between + pino's and srs' and I wasn't tagged and thought it had something to do + with their discussion + <pinotree> (sorry) + <nlightnfotis> no it was my bad + <braunr> ok + <braunr> pinotree: there is indeed a special call to + __pthread_create_internal for the main thread + <pinotree> yeah + <pinotree> braunr: if there wouldn't be that libc→pthread bridge, things + like pthread_self() or so wouldn't work for the main thread + <braunr> pinotree: right + <pinotree> braunr: weird thing is that the error you got is usually a sign + that pthread is not linked in explicitly + <braunr> pinotree: yes + <braunr> pinotree: with firefox, gdb can't locate pthread symbols before a + call to a pthread function + <braunr> so yes, libpthread is loaded after main is called + <braunr> nlightnfotis: can you give me a quick procedure to build gcc with + go support from your repository, and then test a go program please ? + <braunr> to i can have a better look at it myself + <braunr> so* + <nlightnfotis> braunr: sure you want access to my go repo? If you already + have gcc repo add my github repo as a remote and checkout + fotisk/goruntime_hurd + <braunr> i have your github repo + <nlightnfotis> git checkout fotisk/goruntime_hurd (You may need to revert a + commit or two, because of my latest endeavour with mach_print + <nlightnfotis> braunr: check it out now, I reverted some messy commits for + you to rebuild + <braunr> nlightnfotis: i won't work on it right now, i'm building glibc to + check some things in libpthread + <braunr> since it seems to be the source of your problems and many others + <nlightnfotis> oh ok then. btw, it compiles ok, but when I try to compile + another program with gccgo collect2 cries about undefined references to + __pthread_num_threads and __pthread_total + <braunr> Oo + <braunr> another program ? + <nlightnfotis> braunr: will I get the same result if I slowly go through it + with gdb + <nlightnfotis> yep + <braunr> i don't understand + <braunr> what compiles ok, what fails ? + <nlightnfotis> gccgo compiles without errors (which is strange) but when I + use it to compile goroutine.go it fails with the errors I reported + <pinotree> (missing linking to pthread?) + <braunr> since when ? + <nlightnfotis> pinotree: perhaps braunr: since I made the changes with + mach_print + <nlightnfotis> pinotree: but what could be missing the link? GCC compiled + programs are getting linked automatically to the shared objects of the + headers they include right? + <nlightnfotis> (assuming it's not a huge program, only a tiny 10 liner for + instance) + <braunr> uh + <braunr> did you declare them as extern + <braunr> ? + <nlightnfotis> yes + <braunr> do you see -lpthread on the link line ? + <nlightnfotis> during gcc's compilation? I will have to rerun it again and + see. + <braunr> log the compilation output somewhere once + <braunr> nlightnfotis: why did you remove volatile from the definition of + __atomic_t ?? + <nlightnfotis> just for testing purposes, because I thought that the GNU + version is volatile with no __ in front of it and that might cause some + issues. + <braunr> i don't understand + <nlightnfotis> it was just an experiment gone wrong + <braunr> nlightnfotis: keep volatile there + <nlightnfotis> just did + <nlightnfotis> braunr: there is -lpthread on some lines. For instance when + libtool is invoked. + <youpi> braunr: the pthread assertion usually happens when libpthread gets + loaded from a plugin, I guess mozilla got rid of libpthread in the main + application recently, simply + <pinotree> youpi: he said that the LD_PRELOAD trick (which used to + workaround the issue in older iceweasel) does not work, though + <youpi> ah? it does work for me + <pinotree> dunno then... + <braunr> youpi: aouch, ok + <braunr> nlightnfotis: what about the specific gcc invocation that fails ? + <braunr> pinotree: /lib/i386-gnu/libpthread.so.0: ERROR: cannot open + `/lib/i386-gnu/libpthread.so.0' (No such file or directory) + <braunr> trying with a working path this time + <braunr> better + <pinotree> sorry, i typed it by hand :p + <braunr> Segmentation fault + <braunr> but no assertion + <nlightnfotis> braunr: gccgo hello.go + <braunr> nlightnfotis: ? + <pinotree> <braunr> nlightnfotis: what about the specific gcc invocation + that fails ? + <braunr> nlightnfotis: i'm asking if -lpthread is present when you have + these undefined reference errors + <nlightnfotis> it is. it seems so + <nlightnfotis> I wrote above that it is present when libtool is called + <nlightnfotis> I don't know what libtool is doing sadly + <braunr> you said some lines + <nlightnfotis> but I from what I've seen I believe it does some kind of + linking + <braunr> paste it somewhere please + <nlightnfotis> yeah it doesn't fail though + <braunr> that's far too vague ... + <braunr> it doesn't fail ? + <nlightnfotis> give me a second + <braunr> i thought it did + <nlightnfotis> no it doesn't + <braunr> 14:53 < nlightnfotis> gccgo compiles without errors (which is + strange) but when I use it to compile goroutine.go it fails with the + errors I reported + <nlightnfotis> yeah gccgo compiles. + <nlightnfotis> when I use the compiler, it fails + <braunr> so it fails running + <braunr> is gccgo built with -lpthread itself ? + <nlightnfotis> http://pastebin.com/1TkFrDcG + <nlightnfotis> check it out + <nlightnfotis> I think it does, but I would take an extra opinion + <nlightnfotis> line 782 + <nlightnfotis> and 784 + <braunr> (are you building as root ?) + <nlightnfotis> yes. for now + <pinotree> baaad :p + <nlightnfotis> I never had any particular problems...except that one time + that I rm -rf the source tree :P + <nlightnfotis> I know it's bad d/w + <nlightnfotis> braunr: I found something interesting (I don't know if it's + expected or not; probably not): If I set GOMAXPROCS to 2, and run the + goroutine program, it seems to be running for a while (with the + goroutines!) and then it segfaults. Will look more into it + <braunr> it's interesting, yes + <braunr> nlightnfotis: have you tried the preload trick too ? + <nlightnfotis> ldpreload? no. Could you tell me how to do it? export + LDPRELOAD and a path to libpthread? + <braunr> nlightnfotis: LD_PRELOAD=/lib/i386-gnu/libpthread.so.0.3 ... + <nlightnfotis> braunr: it also produces a very different backtrace. This + one heavily involves mig functions + <tschwinge> braunr, nlightnfotis: Thanks for working together, and sorry + for my lack of time. + <braunr> nlightnfotis: paste please + <nlightnfotis> tschwinge, Hello. It's ok, I am sorry for not showing good + amounts of progress from my part. + <nlightnfotis> braunr: http://pastebin.com/J4q2NN9p + <braunr> nlightnfotis: thread apply all bt full please + <nlightnfotis> braunr: http://pastebin.com/tbRkNzjw + <braunr> looks like an infinite loop of + __mach_port_mod_refs/__mig_dealloc_reply_port + <braunr> ... + <nlightnfotis> yes that's what I got from it too. Keep in mind these + results are with GOMAXPROCS=2 and they result in segmentation fault + <nlightnfotis> and I also can not understand the corrupted stack at the + beginning of the backtrace + <braunr> no please + <nlightnfotis> ? + <braunr> test LD_PRELOAD=/lib/i386-gnu/libpthread.so.0.3 without + GOMAXPROCS=2 + <nlightnfotis> braunr: LD_PRELOAD without GOMAXPROCS results in the usual + assertion failure and abortion of execution after it + <braunr> nlightnfotis: ok + <braunr> nlightnfotis: im sorry, i thought you couldn't launch a test since + you added mach_print + <nlightnfotis> I am not using mach_print, I couldn't fix the issue with the + references and thought I was losing time, so I went back to debugging + with gdb until I can't get anything more out of it + <nlightnfotis> braunr: should I focuse on mach_print? Will it produce very + different results than gdb? + <nlightnfotis> *focus + <nlightnfotis> (btw I didn't delete mach print or anything, it's still + there, in another branch) + <nlightnfotis> braunr: Now I stepped through the program in gdb, and got + something really really weird. Some close to a full execution + <nlightnfotis> Number of gorountines and machine threads according to + runtime was 3, __pthread_num_threads was 4 + <nlightnfotis> it did get SIGILL (illegal instruction some times though) + <nlightnfotis> and it exited with code 02 + <braunr> uh + <braunr> nlightnfotis: try with mach_print yes, it will show the values + from the real execution context, and be as close as what we can get + <braunr> i'm not sure about how gdb finds the values + <nlightnfotis> braunr: ok, will spend the rest of the day to find a way to + make mach_print and the other values work. Did you see my last messages, + with the goroutines that worked under gdb? + <braunr> yes + <nlightnfotis> it seemed to run. Didn't get the expected output, but also + didn't get any errors other than illegal instruction either + <nlightnfotis> braunr: I still have not found an easy way to do what you + asked me to from go's runtime. Would it be ok if I do it from inside + libpthread? + <braunr> nlightnfotis: do what ? + <nlightnfotis> print the values of __pthread_total and + __pthread_num_threads with mach_print. + <braunr> how ? + <braunr> oh wait + <braunr> well yes ofc, they're not exported :/ + <braunr> nlightnfotis: have you been able to use mach_print ? + <nlightnfotis> braunr: not really because of the problems I shared + earlier. I can try to use with in-gcc structures if you want me to, it's + nothing hard to do + <nlightnfotis> actually I will. Hang on + <braunr> proceed with debugging inside libpthread instead + <braunr> using mach_print to avoid deadlocks this time + <braunr> (mach_print was purposely built for debugging such low level code + parts) + <nlightnfotis> ok, I will patch this, but can I build it tomorrow? + <braunr> yes + <braunr> just keep us informed + <nlightnfotis> ok, thanks, and sorry for everything I have done. I want you + to know that I really appreciate that you are helping me. + <braunr> remember: the goal here is to understand why __pthread_total and + __pthread_num_threads have inconsistent values + <nlightnfotis> braunr: whenever you see it, mach_print works as expected + inside gcc. + + +# IRC, freenode, #hurd, 2013-09-03 + + <nlightnfotis> braunr: I have made the changes I want to glibc. After I + build it, how do I install it? make install or is it more involved? + <braunr> nlightnfotis: use LD_LIBRARY_PATH + <braunr> never install an experimental glibc unless you have backups or are + certain of what you're doing + <braunr> nlightnfotis: i didn't understand what you meant about mach_print + yesterday + <nlightnfotis> it works in gcc. + <braunr> what do you mean "in gcc" ? + <braunr> why would you put mach_print in gcc ? + <braunr> we want it in go programs .. + <nlightnfotis> yes, I understand it. gcc was the fastest way to test it's + usage at that moment (for me) and I just wanted to confirm it works. I + only had to change its signature to const char * because gcc wouldn't + accept it otherwise + <braunr> doesn't my example include const ? + <braunr> nlightnfotis: why did you rebuild glibc ? + <nlightnfotis> braunr: I have not started yet, will do now, to apply the + changes to libpthread + <braunr> you mean add the print calls there ? + <nlightnfotis> yes + <braunr> ok + <braunr> use debian/rules build, interrupt when you see gcc invocations + <braunr> then switch to the build directory (hurd-libc-i386 iirc), and make + others + <braunr> nlightnfotis: did you send me the instructions to build and test + your work ? + <braunr> so i can reproduce these weird threading problems at my side + <nlightnfotis> braunr: sorry, I was in the toilet, where would you like me + to send the instructions? + <braunr> nlightnfotis: i should be fine i guess, let's check here + <braunr> nlightnfotis: i simply used configure + --enable-languages=c,c++,go,lto + <braunr> and i'll see how it goes + <nlightnfotis> I configure with --enable-languages=go (it automatically + builds c and c++ for that as go depends on them), --disable-bootstrap, + and use a custom prefix to install at a custom location + <braunr> yes + <braunr> ok + <braunr> nlightnfotis: how long does it take you ? + <nlightnfotis> complete non-bootstrap build about 45 minutes. With a build + tree ready and only simple changes, about 2-3 minutes + <nlightnfotis> braunr: In an hour I will go offline for 2-3 hours, I am + gonna move back to my other home in the other city. It won't take long, + the whole process will be about 4 hours, and I will compensate for the + time lost by staying up late up until 3 o clock in the morning + <braunr> i'd prefer you didn't "compensate" + <nlightnfotis> ? + <braunr> work if you want to + <braunr> noone if forcing you to work late at night for gsoc, unless you + want to + <nlightnfotis> no, I do it because I want to. I **really** really want to + succeed, and time is off the essence for me at this point + <braunr> then ok + <braunr> nlok i have a gccgo compiler + <pinotree> nlok? + <braunr> nl being nlightnfotis but he's gone + <pinotree> oh + * pinotree was trying to parse that as "now" or "look" or the like + <nlightnfotis> braunr: 08:19:56< braunr> use debian/rules build, interrupt + when you see gcc invocations: Are gcc invocations related to + i486-gnu-gcc-4.7? + <nlightnfotis> nvm I'm good now :) + <gnu_srs> of course not, that's only for compiling applications using the + newly built libc + <nlightnfotis> gnu_srs: I didn't exactly understand what you said? Care to + elaborate? which one is for compiling applications using the newly build + libc? -486-gnu-gcc-4.7? + <gnu_srs> when you see gcc ... -llibc.so you know libc.so is built, and + that is sufficient to use it. + <gnu_srs> with LD_PRELOAD or LD_LIBRARY_PATH (after cding and building + others) + <nlightnfotis> gnu_srs: thanks for the tip :) + <gnu_srs> :-D + <nlightnfotis> is anyone else getting glibc build problems? (from apt-get + source glibc, at cxa-finalize.c)? + <gnu_srs> apt-get source eglibc; apt-get build-dep eglibc (as root); + dpkg-buildpackage -b ... + <braunr> nlightnfotis: just debian/rules build + <braunr> to start the glibc build + <nlightnfotis> braunr: oh I have now, it's building without issues so far + <braunr> when you see gcc processes, it means the build process has + switched from configuring to making + <braunr> then interrupt (ctrl-c) + <braunr> cd build-tree/hurd-i386-libc + <braunr> make others + <braunr> or make lib others + <braunr> lib is glibc, others is some addons which include our libpthread + <nlightnfotis> thanks for the tip braunr. + <nlightnfotis> braunr: I have managed to get a working version of glibc and + libpthread with mach_print working. I have also run 2 test programs and + it works as expected. Will continue researching tomorrow if that's ok + with you, I am too tired to keep on now. + <nlightnfotis> for the record compilation of glibc right from the start was + about 1 hour and 20 - 30 minutes + + +# IRC, freenode, #hurd, 2013-09-04 + + <braunr> i've taken a deeper look at this assertion failure + <braunr> and ... + <braunr> it has nothing to do with pthread_create + <braunr> i assumed it was the one in sysdeps/mach/pt-thread-start.c + <nlightnfotis> pthread_self ()? + <braunr> but it's actually from sysdeps/mach/hurd/pt-sysdep.h, in + _pthread_self() + <braunr> and looking there : + <braunr> thread = *(struct __pthread **)__hurd_threadvar_location + (_HURD_THREADVAR_THREAD); + <braunr> so simply put, context switching doesn't fix up thread specific + data ... + <braunr> it's that simple + <nlightnfotis> wow + <nlightnfotis> today I was running programs all day long with mach_print on + to print __pthread_total and __pthread_num_threads to see when both + become 1 and couldn't find anything + <nlightnfotis> I was nearly desperate. You just made my day! :) + <braunr> now the problem is + <braunr> thread specific data is highly dependent on the stack + <braunr> it's illegal to make a thread switch stack and expect it to keep + working on the hurd + <nlightnfotis> unless split stack is activated? + <nlightnfotis> no wait + <braunr> split stack is completely unsupported on the hurd + <teythoon> uh, why would that be? + <braunr> teythoon: about split stack ? + <teythoon> yes + <braunr> i'm not sure + <nlightnfotis> at least now we do know what the problem is and I can start + working on a solution. + <nlightnfotis> braunr: we should tell tschwinge and youpi about it. + <braunr> nlightnfotis: sure but + <braunr> nlightnfotis: you can also start looking at a workaround + <braunr> nlightnfotis: also, let's makre sure that's the reason first + <braunr> nlightnfotis: use mach_print to display the stack pointer when + switching + <braunr> nlightnfotis: + http://stackoverflow.com/questions/1880262/go-forcing-goroutines-into-the-same-thread + <braunr> " I believe runtime.LockOSThread() is necessary if you are + creating a library binding from C code which uses thread-local storage" + <braunr> oh, a paper about the go runtime scheduler + <braunr> let's have a look .. + <teythoon> braunr: have you seen the high level overview presented in that + blog post I once posted here? + <braunr> no + <nlightnfotis> braunr, just came back, and read the log. Which paper are + you reading? The one from columbia university? + <braunr> but i need to know about details here, specifically, if threads do + change stack + <braunr> nlightnfotis: yes + <teythoon> braunr: ok + <braunr> this could be caused either by true stack switching, or by "stack + segmentation" as implemented by go + <braunr> it is interesting that there are stack related members per + goroutine + <braunr> nlightnfotis: in particular, pthread_attr_setstacksize() doesn't + work on the hurd + <nlightnfotis> <braunr> it is interesting that there are stack related + members per goroutine -> I think that's go's policy. All goroutines run + on a shared address space (that is the kernel thread's address space) + <braunr> nlightnfotis: that's obvious + <braunr> and not the problem + <braunr> and yes, it's "stack segmentation" + <braunr> and on linux, and probably other archs, switching stack may be + perfectly legit + <braunr> on the hurd, we still have threadvars + <braunr> which are the hurd specific thread local storage mechanism + <braunr> it means 1/ all stacks in a process must have the same size + <braunr> 2/ stack size must be a power of two + <braunr> 3/ threads can't switch stack + <braunr> this hardly prevents goroutines from being run by just any thread + <braunr> i see there already hard hurd specific changes about stack + handling + <nlightnfotis> so we should only make changes to the specific gccgo + scheduler as a workaround under the Hurd right? + <braunr> i don't know + <braunr> this might also push the switch to tls + <nlightnfotis> this sounds better as a long term fix + <nlightnfotis> but it must also involve a great amount of work, right? + <braunr> most of it has already been done + <braunr> by youpi and tschwinge + <nlightnfotis> with the changes to tls early in the summer? + <braunr> maybe + <braunr> 14:36 < braunr> nlightnfotis: also, let's makre sure that's the + reason first + <braunr> 14:36 < braunr> nlightnfotis: use mach_print to display the stack + pointer when switching + <braunr> check what goes wrong with the stack + <braunr> then we'll see + <braunr> as a very simple workaround, i expect locking g's on m's to be a + good first step + <nlightnfotis> braunr: noted everything. that's my work for tonight. I + expect myself to stay up late like yesterday and have this all figured + out by tomorrow. + <braunr> nlightnfotis: why not now ? + <nlightnfotis> I am starting from now, but I expect myself to stop about 6 + o clock here (2 hours) because I have an appointment with a doctor. + <nlightnfotis> and keep on when I come back home + <braunr> well adding a few printfs to track the stack should be doable + before 2 hours + <nlightnfotis> braunr: I am doing it now. Will report as soon as I have + results :) + <nlightnfotis> braunr: have I messed up with the way I read esp's value? + https://github.com/NlightNFotis/glibc/commit/fdab1f5d45a43db5c5c288c4579b3d8251ee0f64#L1R67 + <braunr> nlightnfotis: +unsigned + <braunr> nlightnfotis: using gdb : + <braunr> (gdb) info registers + <braunr> esp 0x203ff7c0 0x203ff7c0 + <braunr> (gdb) print thread->stackaddr + <braunr> $2 = (void *) 0x2000000 + <nlightnfotis> oh yes, I know about gdb, I thought you wanted me to use + mach_print + <braunr> nlightnfotis: yes + <braunr> this is just my own attempt + <braunr> and it does show the stack pointer is completely outside the + thread stack + <braunr> nlightnfotis: in your code, i suggest using + __builtin_frame_address() + <braunr> well __builtin_frame_address(0) + <braunr> see + http://gcc.gnu.org/onlinedocs/gcc-4.7.3/gcc/Return-Address.html#Return-Address + <braunr> it's not exactly the stack pointer but close enough, unless of + course the stack is changed in the middle of the function + <nlightnfotis> I see. I am gonna try one more time with esp the way I + worked it and if it fails to work, I am gonna use return address + <braunr> nlightnfotis: be very careful about signed/unsigned and type + widths + <braunr> not return address, frame address + <braunr> return address is code, frame address is data (stack) + <nlightnfotis> ah, I see, thanks for the correction. + <braunr> youpi: not sure you catched it earlier, the problem fotis has been + having with goroutines is about threadvars + <braunr> simply put, threads use setcontext functions to save/restore + goroutines state, which make them switch stack, rendering the location of + threadvars invalid, and making _pthread_self() choke + + +# IRC, freenode, #hurd, 2013-09-05 + + <nlightnfotis> I am having very weird behavior with my code, something that + I can not explain and seems likely to be a bug, could someone else take a + look? + <nlightnfotis> pinotree are you available at the moment to take a look at + something? + <pinotree> nlightnfotis: dont ask to ask, just ask + <nlightnfotis> I have made some modifications to pthread_self as also + suggested by braunr to see if the stack pointer is within the bounds of + the frame address after context switching. I can get the values of both + esp and frame_address to be shown before the context switch, but I can + only get the value of esp to be shown after the context switch, and it + always results to the program getting killed + <nlightnfotis> + https://github.com/NlightNFotis/glibc/blob/7e72da09a42b1518865f6f4882d68689e681f25b/libpthread/sysdeps/mach/hurd/pt-sysdep.h#L97 + <nlightnfotis> thing is a dummy print value I have right after the code + that was supposed to print the frame_address after the context switching + is executing without any issues. + <pinotree> oh assembler... cannot help, sorry :/ + <nlightnfotis> oh no, I am not asking for assembler help, that part works + quite alright. I am asking why from the 4 identical pieces of code that + print debugging values the last one doesn't work. I am on it all day, and + still have not found an answer + <braunr> nlightnfotis: i can + <nlightnfotis> hello braunr, + <braunr> nlightnfotis: do you have a backtrace ? + <braunr> uh + <nlightnfotis> nope, it crashes right after I execute something. Let me + compile glibc once again and see if a fix I attempted works + <braunr> malloc and free use locks + <braunr> so they probably use _pthread_self + <braunr> don't use them + <braunr> for debugging, a simple statically allocated buffer on the stack + will do + <braunr> nlightnfotis: so ? + <nlightnfotis> Ι got past my original problem, but now I am trying to get + past the sigkills that kill the program at the beginning + <nlightnfotis> i remember not having this problem, so I am compiling my + master branch to see if it is reproducible. If it is, it means something + is very wrong. If it's not, it means I screwed up somewhere + <braunr> i don't understand, how do you know if you get past the problem if + you still have trouble reaching that code ? + <nlightnfotis> braunr: I fixed all my problems now. I can see that both esp + and the frame_address are the same after context switching though? + <braunr> always ? + <braunr> for all goroutines ? + <nlightnfotis> for all kernel threads, not go routines. We are in + libpthread + <braunr> if they're the same after a context switch, it usually means the + scheduler didn't switch + <braunr> well obviously + <braunr> but what i asked you was to trace calls to setcontext functions + <nlightnfotis> I will run some tests again. May I show you my code to see + if there is anything wrong with it? + <braunr> what address do you have ? + <braunr> not yet + <braunr> i'm not sure you understand what i want to check + <braunr> do you see how threadvars work basically ? + <nlightnfotis> I think so yes, they keep in the stack the local variables + of a thread right? + <nlightnfotis> and the globals + <nlightnfotis> or + <nlightnfotis> wait a minute... + <braunr> yes but do you see how the thread specific data are fetched ? + <nlightnfotis> with __hurd_threadvar_location_from_sp? + <braunr> yes but "basically", what does it do ? + <nlightnfotis> it get's a stack pointer as a parameter, and returns the + location of that specific data based on that stack pointer, right? + <braunr> and how ? + <nlightnfotis> I believe it must compare the base value of the stack and + the value of the end of the stack, and if the results are consistent, it + returns a pointer to the data? + <braunr> and how does it determine the start and end of the stack ? + <nlightnfotis> stack_pointer must be pointing at the base of the + stack. That + stack_size must be the stack limit I guess. + <braunr> so you're saying the caller of __hurd_threadvar_location_from_sp + knows the stack base ? + <nlightnfotis> I am not so sure I understand this question. + <braunr> i want to know if you understand how threadvars work + <braunr> apparently you don't + <braunr> the caller only has its current stack pointer + <braunr> which does *not* point to the stack base + <braunr> threadvars work by assuming a *fixed* stack size, power of two, + aligned (obviously) + <braunr> in our case, 2MiB (except in hurd servers where a kludge reduces + that to 64k) + <braunr> this is why stack size can't be changed + <braunr> this is also why the stack pointer can't ever point outside the + initial stack + <braunr> i want you to make sure go violates this last assumption + <braunr> so 1/ show the initial stack boundaries of your threads, then show + that, after loading a goroutine, the stack pointer is outside + <braunr> which is what, if i'm right, triggers the assertion + <braunr> ask if there is anything confusing + <braunr> this is important, it should already have been done + <nlightnfotis> ok, I noted it all, I am starting to work on it right now. I + only have one question. My results, the ones with the stack pointer and + the frame address, are expected or unexpected? + <braunr> i don't know + <braunr> show me the code again please + <braunr> and explain your intent + <nlightnfotis> + https://github.com/NlightNFotis/glibc/blob/7fe202317db4c3947f8ae1d1a4e52f7f0642e9ed/libpthread/sysdeps/mach/hurd/pt-sysdep.h + <nlightnfotis> At first I print the value of esp and the frame_address + before the context switching and after the context switching. + <nlightnfotis> The different variables were introduced as part of a test to + see if my results were consistent, + <braunr> what context switch ? + <nlightnfotis> in hurd_threadvar_location + <braunr> what makes you think this is a context switch ? + <nlightnfotis> in threadvar.h, it calls __hurd_threadvar_location_from_sp. + <nlightnfotis> the full path for it is glibc/hurd/hurd/threadvar.h + <braunr> i don't see how giving me the path will explain why it's a context + switch + <braunr> and i can tell you right away it's not + <braunr> hurd_threadvar_location is basically a lookup returning the + address of the thread specific data + <nlightnfotis> wait a minute...does this mean that + hurd_threadvar_location_from_sp is also a lookup function for the same + reason + <nlightnfotis> ? + <braunr> yes + <braunr> isn't the name meaningful enough ? + <braunr> "location of the threadvars from stack pointer" + <nlightnfotis> I guess I made wrong deductions from when you originally + shared your findings... + <nlightnfotis> <braunr> thread = *(struct __pthread + **)__hurd_threadvar_location (_HURD_THREADVAR_THREAD); + <nlightnfotis> <braunr> so simply put, context switching doesn't fix up + thread specific data ... + <nlightnfotis> I thought that hurd_threadvar_location was doing the context + switching + <braunr> nlightnfotis: by context switching, i mean setcontext functions + <nlightnfotis> braunr: You mean the one in sysdeps/mach/hurd/i386? + <braunr> yes + <braunr> but + <braunr> do you understand what i want you to check now ? + <nlightnfotis> I think I got this time: Let me explain it: + <nlightnfotis> You suggested that stack sizes are fixed. That is the main + reason that the stack pointer should not be able to point outside of it. + <braunr> no + <braunr> locating threadvars is done by applying a mask, computed from the + stack size, on the stack pointer, to determine its base + <nlightnfotis> yeah, what __hurd_threadvar_location_from_sp is doing + <braunr> if size is a power of two, size - 1 is a mask that, if + complemented, aligns the address + <braunr> yes + <braunr> so, threadvars expect the stack pointer to always point to the + initial stack + <nlightnfotis> and we wanna prove that go violates this rule right? That + the stack pointer is not pointing at the initial stack + <braunr> yes |