AIX fixes - #700
Conversation
When running the install script on AIX, the build pipes most of the output through dist/isum.awk. This AWK script does not display any output on AIX using the default awk located in /usr/bin/awk. On the other platforms, this script displays the current item being compiled on a single line and updates that line as the build progresses. Using GNU AWK, which is available from the IBM AIX Toolbox for Open Source Software, allows the script to run correctly. This makes for a prettier build at the cost of requiring gawk on AIX. During the run of INSTALL, /opt/freeware/bin must also be in the PATH environment variable.
The INSTALL script detects if the fontsrv prerequsites are installed by compiling a test program. However, this does not work on AIX even with freetype installed from the IBM Toolbox for Open Source Software. This does not work for two reasons. First, the IBM- supplied package is installed into /opt/freeware, and therefore we need to include the following directories: /opt/freeware/include /opt/freeware/include/freetype2 Secondly, the INSTALL script uses cc to compile the script, and this does not set up enough PLAN9 environment for the compiler to find the right headers on AIX. Therefore, we use 9c which handles all of the AIX compiler envionment setup, and should also work for the other platforms.
| BUGGERED='CVS|faces|factotum|lp|ip|mailfs|upas|vncv|mnihongo|mpm|index|u9fs|secstore|smugfs|snarfer' | ||
| DIRS=lex `ls -l |sed -n 's/^d.* //p' |egrep -v "^($BUGGERED)$"|egrep -v '^lex$'` $FONTSRV | ||
| UNAME=`uname` | ||
| AIXBUGGERED=`if [ "$UNAME" == "AIX" ]; then echo rio; fi` |
There was a problem hiding this comment.
why not append to BUGGERED if on AIX?
There was a problem hiding this comment.
That is the effect here, but indirectly. I suppose I could add AIXBUGGERED to BUGGERED if you would prefer to see it done that way.
There was a problem hiding this comment.
Yes, I think adding to BUGGERED would be better.
| } | ||
| close(fd); | ||
| w = waitfor(pid); | ||
| sleep(10); |
There was a problem hiding this comment.
I know that you know this, but this is clearly the wrong fix. I recognize that this probably masks the symptoms of the underlying bug enough that things work on AIX with this in place, but I'm very hesitant to commit a bandaid like this into the repo.
There was a problem hiding this comment.
Indeed. This is one of those things that is going to be monstrously difficult to track down. Single-stepping through the code under the debugger adds enough of a delay that it works correctly while debugging it, so figuring out what is going wrong is pretty tough. Furthermore, if I compile the unpatched code on AIX 7.1, it works there but then when I move the binaries over to AIX 7.3, we also hit this issue. I'd welcome suggestions for how to proceed...
There was a problem hiding this comment.
Maybe, maybe not.
Looking at what's going on, it appears that, without the sleep, waitfor is returning a bogus, nil, value. Drilling down into waitfor, we see that it delegates to _wait and the first thing that does is see if n is <= 0 and return nil if it is. The other cases where it returns nil are if it can't tokenize whatever is in buf, and if malloc fails. Let's assume malloc isn't failing; then if tokenize failed, we should see the "couldn't parse wait message" error somewhere and we don't, so it stands to reason that n <= 0 is probably true. n in this case is the return value from awaitfor.
awaitfor delegates to _await, which in turn calls wait3 or wait4, depending on whether pid4 is -1. There's a comment there that is curiously incomplete; "On Linux, pid==-1 means anyone; on SunOS it's pid==0." That's nice, but there's no special case for SunOS; perhaps the comment is simply a vestige of something that's changed over time. But it's probably irrelevant here; we know that the pid we get here is the PID of a child process created by fork() in post9pservice, and the error case there was already dealt with. Therefore, we know that _await is calling wait4 in this instance. The the conclusion here is that this code is failing because _await is returning 0 or -1.
How that can that happen? There are three returns from this function, in three separate cases. Two of those are strlen, which can't return -1, and can only return 0 if passed an empty string. But in this case, I don't see how that can happen: the string in question is the result of a copying a formatted string into a buffer, and we can see statically that that buffer isn't empty. This suggests that _await is returning -1 since due to the check against the return value of wait3 or wait4. Since we know that we're calling wait4 here, we can deduce that this function is returning -1 because wait4 is returning 0 or something negative; I'm not a betting man, but if I were, my would be on -1.
I'll bet the, "Restart the system call" message mentioned in #699 is from actually from wait4; some random Google searches suggest that that's the strerror string for "ERESTART" on AIX. I don't have access to AIX (and haven't for decades); does the man page for wait4 mention anything about ERESTART?
Something you may try, instead of single-stepping in the debugger, is seeing if you can put a breakpoint on the return -1 on line 99 of await.c and see if you hit it without the sleep(10) and without single stepping.
|
|
||
| echo "* Compiler version:" | ||
| 9c -v 2>&1 | grep -v 'Configured with:' | grep -i version | sed 's/^/ /' | ||
| 9c -v -qversion 2>&1 | grep -v 'Configured with:' | grep -i version | sed 's/^/ /' |
There was a problem hiding this comment.
-qversion is presumably specific to xlc; with e.g. clang, the output here makes sense, but isn't what I'd normally expect to see:
* Compiler version:
clang: error: unknown argument '-qversion'; did you mean '--version'?
OpenBSD clang version 16.0.6
(Note the error message that sneaks in.)
Perhaps a way around this is add an extra variable in the large case OS statement above that adds this specifically for AIX. Something like:
CCVERSFLAG=''
case `uname` in
# ...
AIX)
# ...
CCVERSFLAG=-qversion
;;
esac
9c -v $CCVERSFLAG 2>&1 | grep ...
The INSTALL script typically prints the version of the compiler in use during a build, and does so by running "9c -v". Ultimately the -v argument gets passed to the actual compiler. However, the AIX compiler (XL C) does not use -v to display the compiler version, but instead uses the argument -qversion. Therefore, we add -qversion to the 9c invocation on this platform only.
The fonsrv program can build on AIX so long as the prerequsites are installed. These include freetype2 and fontconfig, both of which are available from the IBM Toolbox for Open Source Software. However, these get installed into /opt/freeware. Update the freetyperules.sh script to look for the headers and libraries in /opt/freeware on AIX.
When attempting to build rio on AIX, the following compiler error occurs: 9c -DDEBUG -DSHAPE -DDEBUG_EV -DDEBUG event.c event.c:9.10: 1506-296 (S) #include file <X11/extensions/shape.h> not found. event.c:32.59: 1506-046 (S) Syntax error. event.c:32.60: 1506-068 (S) Operation between types "void" and "union _XEvent" is not allowed. event.c:473.25: 1506-277 (S) Syntax error: possible missing ')' or ','? event.c:478.23: 1506-045 (S) Undeclared identifier e. mk: 9c -DDEBUG -DSHAPE ... : exit status=exit(1) mk: for i in ... : exit status=exit(1) This is because IBM does not supply X11/extensions/shape.h in any LPP. Unfortunately, this means that rio is permanently unusable on AIX, without significant modification. The easiest path forward is to add rio to the BUGGERED list in src/cmd/mkfile. However, rio works on most of the other platforms. Therefore, add an extra test and add rio to the BUGGERED list only if we are performing the build on AIX.
Suppress additional compiler and linker warnings on AIX.
The _await function did not retry wait3/wait4 when interrupted by a signal. On AIX 7.3, this causes post9pservice to fail because wait4 was interrupted during the window between forking 9pserve and the daemonize passer process exiting. Handle EINTR (POSIX) and ERESTART (non-POSIX, primarily AIX) by retrying the wait call. Fixes 9fans#699
2 participants
These commits address several issues on AIX:
Unfortunately, this introduces a requirement to install gawk on AIX, however doing so is fairly straightforward.