From: esr@snark.thyrsus.com (Eric S. Raymond) Newsgroups: comp.unix.sysv386,comp.unix.bsd,comp.os.mach,news.answers Subject: Known Bugs in the USL UNIX distribution Message-ID: <1jWcLX#1Qbqls9Wy5rR5kw7xb4zw81M=esr@snark.thyrsus.com> Date: 20 Nov 92 21:45:47 GMT Expires: 20 Feb 93 00:00:00 GMT Lines: 1165 Archive-name: usl-bugs Last-update: Fri Nov 20 16:16:32 1992 Version: 8.0 [Note: This issue is being released nearly two weeks ahead of schedule. This is because (a) there have been been big changes in the hardware market since 7.0 --- the 386 is dead, long live the 386, and (b) I received an unusually large number of additions and corrections during the last week, including material for a new major section.] What's new in this issue: * A table of contents and some reorganization by topic * Several new bugs * Update on problems fixed and remaining in Dell 2.2. (In the table below, bugs new this issue or old bugs for which information has been added are marked with a * at the left margin) 0. Table of Contents I. Introduction II. General Bugs 1. Dropout problems with tty devices 2. Suid programs dump core when signalled 3. DMAs on large ISA machines may fail 4. There is a cylinder limit on disk size 5. shmat(2) vs. vfork(2) 6. X performance problem 7. FIONREAD fails on regular files 8. A security hole in login 9. COFF problems with long filenames 10. Flakeouts in the Wangtek device driver * 11. A kernel declaration bug 12. fread(3) does the wrong thing on pipes and FIFOs * 13. Process accounting is broken 14. tar(1) foos up in the presence of symbolic links * 15. Symbolic links can interfere with shellscript execution * 16. Piping a csh builtin causes the shell to hang. 17. Quick port setup option in sysadm is broken 18. COFF binaries linked with curses(3) and shared libc hang 19. shl hangs, sxt devices bad 20. num-lock prevents mouse from working properly 21. adjtime() doesn't work 22. ttymon drops DTR 23. cron mail doesn't go through aliasing * 24. fragility in xterm 25. csh lossage due to bad optimization 26. Bug in cp(1) 27. tbl -me doesn't work 28. who -r fragility leads to boot-time problems 29. at(1) breaks here-documents in shell scripts III. Networking and File-Sharing Bugs * 1. NFS locking is unusably slow 2. UFS file system problems * 3. Byte-order problem with NFS when accessing Sun disks 4. Under weird circumstances, lseek on UFS may cause corruption 5. FTP problems 6. A bug in the WD80x3 support IV. SCSI Support Problems 1. sar is confused by SCSI 2. A configuration problem * 3. Synchronous SCSI hang problem * 4. ps chokes on commands that do SCSI I/O 5. Transfer speed problems with Adaptec 1542B on 486s V. Development Tools Problems 1. General UCB library brokenness * 2. USL emulation of BSD signals doesn't work 3. Possible string library problems * 4. USL's ndbm support is broken. 5. An include file is missing 6. sscanf(3) has a potential bug 7. Compiler problems VI. The FUBYTE Problem VII. Destiny and Dell I. Introduction This posting lists known bugs in System V Release 4 implementations, and known fixes applied by various porting houses (there's also random bits of information about SCO UNIX here and there). It was formerly part of the 386-buyers-faq issues 1.0 through 4.0, and is still best read in conjunction with the pc-unix/software FAQ descended from that posting. This document is maintained and periodically updated as a service to the net by Eric S. Raymond , who began it for the very best self-interested reason that he was in the market and didn't believe in plonking down several grand without doing his homework first (no, I don't get paid for this, though I have had a bunch of free software and hardware dumped on me as a result of it!). Corrections, updates, and all pertinent information are welcomed at that address. This posting is periodically broadcast to the USENET group comp.unix.sysv386 and to a list of vendor addresses. If you are a vendor representative, please check to make sure the information on your company is current and correct. If it is not, please email me a correction ASAP. If you are a knowledgeable user of any of these products, please send me a precis of your experiences for the improvement of future issues. The bug descriptions often include indications of fixes by the various porting houses to their current releases. These are: Consensys UNIX Version 1.3 abbreviated as "Cons" below Dell UNIX Issue 2.2 abbreviated as "Dell" below Esix Revision A abbreviated as "Esix" below Micro Station Technology SVr4 UNIX abbreviated as "MST" below Microport System V Release 4.0 version 4 abbreviated as "uPort" below UHC Version 3.6 abbreviated as "UHC" below SCO Open DeskTop 1.1 abbreviated as "SCO" below II. General Bugs 1. Dropout problems with tty devices The most serious problem anyone has reported is that the USL asy driver is flaky and occasionally drops characters at above 4800 baud. Microport, Dell, Esix, and UHC say that they believe they've fixed this. However, Dell, at least, was mistaken when they first made this claim; a more detailed description of the problem is given below. I have been assured that this is on the fix list for the next Dell release. Bela Lubkin at SCO comments "386 interrupt latency vs. unbuffered UARTs. This is a tough problem. Nobody's driver should drop characters with a turned-on 16550. It's not so easy with a 16450. Anyone with 16450s or lower should be able to solve their problems by dropping in a 16550." 2. Suid programs dump core when signalled Mark Snitily of SGCS says that under many SVr4s, signalling a process that is running suid root will cause it to core-dump. He says Dell and MST have fixed this, and SCO doesn't suffer from this. 3. DMAs on large ISA machines may fail On ISA machines with more that 16MB of RAM, SVr4 may try to do DMA from outside the bus's address space, causing serious problems. UNIX ought to do an in-memory copy to within the low 16MB but the USL base code doesn't. Dell says they've fixed this, and that's been confirmed by a user. UHC says they've fixed this; they add that the special buffer-allocation logic to handle the problem can be turned off with a tunable kernel parameter if you've got less than 16M. Microport says they've fixed this in their new 4.1 release, shipping early March. Esix offers a patch to correct this problem. SCO used to have a similar bug but fixed it long ago. John Sully writes: "This was due to a bug in pre version 4 dma code. The USL code has always at least attempted to do a copy from low memory to high memory on systems with more than 16Mb of RAM. By the way UHC is wrong; the buffer allocation code only comes into play if you have more than 16Mb of memory. You can turn it off if you have a machine (ie. an EISA bus) which will allow you to do DMA above 16Mb. You *must* have this tunable (MAXDMAPAGE) turned on if you are using *ISA* bus masters in a system with more than 16Mb of ram. Unfortunately doing this will affect all drivers which do dma as there is no good way to do this on a per-driver basis." 4. There is a cylinder limit on disk size Stock USL code is limited to 1,024 cylinders per Winchester, which might cause problems with some disk drives. Microport, Dell, Esix, MST, and UHC have fixed this. Bela Lubkin says "SCO's boot filesystem must lie below 1024 cylinder mark; anything else can be anywhere. This is more-or-less a limitation of the BIOS interface that the bootstrap loader must use. Could be circumvented by going directly to controller hardware in the bootstrap loader, but that would be horrendously complex with all the controllers & host adapters to be supported." This limit probably applies to all other UNIXes as well. 5. shmat(2) vs. vfork(2) The shmat(2) call is known to interact bady with vfork(2). Specifically, if you attach a shared-memory segment, vfork(), and then the child releases the segment, the parent loses it too! Workaround; use fork(2). UHC and Microport both suspect that they still have this bug and opine that anyone who uses vfork deserves to lose. Dell has no plans to fix it. John Sully writes: "This is not a bug. It is completely consistent with the semantics of a change to the address space of the child. Think about it: any change to the address space of a child process created by vfork(2) is reflected in the parent since the child is actually executing in the parent's address space. Therefore if the child changes the address space (in this case by releasing the shared memory segment) what should happen? Right, the parent should have the same change happen. And what does happen? The segment is released in the parent. One can argue about the braindead semantics of vfork(2) all day, but the fact remains that this is exactly what one would expect to happen. To quote from the manual page: [...] vfork differs from fork in that the child borrows the parent's *memory* and thread of control until a call to execve or an exit (either by a call to exit or abnormally.) [ emphasis added ] and later: It does not work, however, to return while running in the child's context from the procedure which called vfork since the eventual return from vfork would then return to a no longer existent stack frame. Please note that the entire address space of the parent is used by the child created by vfork(2). The manual page also points out several other caveats involved in doing anything to the parent's address space except successfully calling an exec family function or _exit (note it specifically says *not* to call exit(2)). I do not believe that having a shared memory segment disappear from the parent's address space is out of line after reading the man page for vfork(2). It is interesting to note that Sun after implementing its new VM system in SunOS 4.0 initially had no plans to support vfork, since they felt that the COW semantics of the new fork would provide the necessary efficiency gain. Indeed they found that most programs which used vfork worked just fine by doing -Dvfork=fork. All that is, except for a certain popular command interpreter [ed: can you say C shell?]. So we are stuck with the legacy of this braindead system call. BTW, Microport has no plans to fix this :-)." 6. X performance problem Stock X11R4 and R5 (at least prior to 1.2E) is said to hog the processor if you use the LOCALCONNECT option. Jan Brittenson posted the following workaround: I don't know what causes the standard X server to hog the CPU, but it can be avoided. Use the following program instead of xinit. Compile it with `$CC -O -o xserv xserv.c -lX11' where CC is either /usr/ccs/bin/cc or gcc. Set DISPLAY and XINITRC and run `xserv' from your home directory. This is just a q&d hack, and not really a substitute for xinit -- but it works. /* xserv.c -- start X server Start X server. Similar to xinit, but intended to circumvent the X386 CPU Hog Mode Jan Brittenson, June 2 1992 05:15 am */ #include #include #include #include #include #include #include #include #include extern int errno; /* Start X server. Fork-exec server, passing the DISPLAY environment variable. Wait for server to get up and running (at which point it passes back a SIGUSR1), at which point the user xinitrc file is run. */ #define DEFAULT_XPATH "/usr/X386/bin/X386" #define XINITRC ".xinitrc" #define DEFAULT_XCOMMAND "xterm -g +1+1 -n login -display :0" extern void *malloc (), free (); extern char *basename (), *getenv (), *strcpy (); /* X stuff */ Display *top_display; /* This is supposed to be in libgen.a... */ static char *basename (s0) char *s0; { register char *s1; for (s1 = s0 + strlen (s0) - 1; s1 > s0 && *s1 != '/'; s1--); if (*s1 == '/') return s1+1; return s1; } jmp_buf sigusr1_frame; static void caught_sigusr1 (int dummy) { longjmp (sigusr1_frame, !0); } static char *dispname (s0) char *s0; { register char *s1; for (s1 = s0 + strlen (s0) - 1; s1 > s0 && *s1 != ':'; s1--); return s1; } /* No arguments */ int main (argc, argv) int argc; char **argv; { char *xserver_file, *xinitrc_file, *home_path, *display, *display_X_arg; int xserver_pid, orgmask; /* Not that it really matters, just to avoid being used as a direct replacement for xinit. */ if (argc != 1) { fprintf (stderr, "usage: %s\n", basename (*argv)); exit (1); } /* Resolve xinitrc path. This is done before the server is started. */ if (!(home_path = getenv ("HOME"))) home_path = "/etc"; if (!(xinitrc_file = getenv ("XINITRC"))) { xinitrc_file = malloc (strlen (home_path) + 1 + strlen (XINITRC) + 1); sprintf (xinitrc_file, "%s/%s", home_path, xinitrc_file); } else xinitrc_file = strdup (xinitrc_file); /* Resolve display */ if (!(display = getenv ("DISPLAY"))) display = display_X_arg = ":0.0"; else display_X_arg = dispname (display); /* Tell server to notify us when up and running */ signal (SIGUSR1, SIG_IGN); orgmask = sigblock (sigmask (SIGUSR1)); /* Start server */ if (!(xserver_pid = vfork ())) { xserver_file = DEFAULT_XPATH; execl (xserver_file, xserver_file, display_X_arg, NULL); fprintf (stderr, "%s: can't exec %s (errno = %d) -- start-up aborted\n", basename (*argv), xserver_file, errno); exit (1); } if (xserver_pid < 0) { fprintf (stderr, "%s: can't fork (errno = %d) -- start-up aborted\n", basename (*argv), errno); exit (1); } /* Await signal from server */ #if 0 /* Why the #@$*! doesn't this work?! */ sigsetmask (orgmask); alarm (20); sigpause (sigmask (SIGUSR1) | sigmask (SIGALRM)); #else sleep (5); #endif /* Open display */ if (!(top_display = XOpenDisplay (display))) { fprintf (stderr, "%s: unable to open display '%s' -- start-up aborted\n", basename (*argv), display); exit (1); } /* Execute xinitrc file */ if (system (xinitrc_file) < 0) system (DEFAULT_XCOMMAND); /* Close display */ XCloseDisplay (top_display); /* Terminate server */ kill (xserver_pid, SIGTERM); /* Finished */ free (xinitrc_file); } 7. FIONREAD fails on regular files Christoph Badura reports that the FIONREAD ioctl() fails on regular (disk) files. He has sent USL a one-line kernel fix. 8. A security hole in login David Wexelblat reports: "There is a HUGE security hole in /bin/login in all USL derived SVR4s before 4.0.4. Refer to CERT advisory CA-91:08, dated 5/23/91. This is known to be present in AT&T SVR4 2.1, and Microport SVR4 3.1. ESIX claims to have fixed it, Microport reports that it is fixed in 4.1. I won't give any more details unless necessary. Suffice to say that this bug allows any non-privileged user on an SVR4 system to get read-write access to any file on the system." 9. COFF problems with long filenames A source at Dell urges: "Our SVR4v2 did some stuff that USL didn't get around to until SVR4v4. Try Dell UNIX 2.1 with a COFF program on a large UFS filesystem in a directory with long names. Runs on Dell UNIX. Breaks on others." I don't have more definite info yet. 10. Flakeouts in the Wangtek device driver Dell reports that USL's Wangtek device driver is seriously flaky. "How'd you like a multi volume backup where the second and subsequent volumes don't follow on from the previous volumes?" UHC confirms this and is actively working on the problem. An anonymous SCOer says "The QIC02 tape controller `standard' is seriously flaky. Our driver's in pretty good shape but nobody will ever have a truly solid driver that supports every QIC02 controller you can find." Gordon Ross reports: "Actually, the SCSI tape target driver `st01' has a similar problem at version 4.0.3 which I corrected while I worked on the SVR4 code. The correction was provided to the support group at USL. The actual problem was that the SCSI tape would return a `check status' completion code which was just trying to inform the driver of the arrival of the `logical end of media' indication but the driver was treating it as an error. The tape drive had in fact written the data, but the driver incorrectly assumed that the "check status" return meant that it failed. The result of this is that when you write into the end of the tape, you can read back one more "chunk" than yu wrote. Of course, cpio does not like this at all when doing multi-volume backups..." 11. A kernel declaration bug A botch in USL's /etc/conf/pack.d/kernel/space.c (which is present in Consensys 1.3, Dell 2.1, Esix 4.0.3A, Microport 4.0.3 and 4.0.4 and may also be present in other SVr4s) can step on the linesw[] table. The problem is that the domain name array initialization is wrong and too short; thus, when it's set, data past the end of the array can be stomped. To fix this, find the following near line 247: char srpc_domain[] = SRPC_DOMAIN; and change it to char srpc_domain[SYS_NMLN] = SRPC_DOMAIN; then rebuild the kernel. Microport officially knows about this bug and plans to fix it in a near-future update release. It has been fixed in Dell 2.2. 12. fread(3) does the wrong thing on pipes and FIFOs Ed Hall writes: "Unlike the raw read() system call, fread() is supposed to be able to make several partial reads to satisfy the data requested by its arguments. The exceptions are an EOF or an error on the stream. This characteristic is quite useful when moving data through pipes or over network connections, since partial reads are quite common in these cases. Well, the version of fread() in ESIX 4.0.3 (and likely other Sys5R4's) only does a single physical read, and if it only satifies part of the requested number of bytes, that's all you get. This can sting you even if you carefully check the value returned by fread(), since the value returned is rounded down to the number of complete "nitems" read, although your position in the stream can be up to size-1 bytes beyond that point. Neither ferror() nor feof() indicate anything is wrong when this happens." This bug (which is also present in 4.0.4) is serious and nasty and should be high on every porting house's list to fix. It appears to be peculiar to USL 4.0.3 and 4.0.4; 4.0.2 does *not* have it, nor does SCO. A USL source claims it has been fixed in 4.1. 13. Process accounting is broken In 4.0.3, process accounting doesn't work. From examining the accounting scripts, it appears that /usr/lib/acct/accton is supposed to set a return code depending on whether accounting was switched on already or not. However, it always returns the same result - accounting switched off. This means that the /usr/lib/acct/ckpacct script, which is run every hour to keep the proccess accounting log in check, instead turns off accounting the first time it is run after booting. The same happens with the nightly /usr/lib/acct/monacct program. I don't yet know whether this bug is present in 4.0.4. It is definitely un-fixed in Dell 2.1 and Consensys 1.3. In Dell 2.2 the return bug is fixed, but accounting isn't automatically enabled at boot time. 14. tar(1) foos up in the presence of symbolic links Tar can get the names of symbolic links wrong when creating an archive. This bug can be demonstrated by doing the following: mkdir t cd t touch a 1234567890 ln -s 1234567890 b ln -s a c tar vcf ../t.tar . The output generated by tar is: a ./ 0 tape blocks a ./a 0 tape blocks a ./1234567890 0 tape blocks a ./b symbolic link to 1234567890 a ./c symbolic link to a234567890 (Note the above commands should be done in the order shown and in a new directory) This bug is nasty. Recommended solution: use GNU tar. This is reported from Esix 4.0.3 and Consensys 1.3, but probably exists on other SVr4s as well. 15. Symbolic links can interfere with shellscript execution There is a problem running #! scripts when symbolic links are involved. Typing in the following from a command shell demonstrates the problem: mkdir a b ln -s a c cd a cat > script <, who reported this bug, says he believes it is the sxt devices which are broken. It definitely exists in Consensys 1.3. 20. num-lock prevents mouse from working properly When using the Motif window manager, if your num lock is on, your mouse clicks are not recognized by the window manager. The mouse still works in xterm(1). This is allegedly fixed in Destiny (4.2). 21. adjtime() doesn't work Hugh Stearns reports that in 4.0.3.6 adjtime() doesn't. Calling `date -a' works to adjust the time slowly. 22. ttymon drops DTR Hugh Stearns reports that in 4.0.3.6 the ttymon(1) utility for HDB uucp drops DTR every few weeks. The workaround is to disable and re-enable it. 23. cron mail doesn't go through aliasing Hugh Stearns reports that in 4.0.3.6 cron mail to adm doesn't get redirected by the aliases file. 24. fragility in xterm Hugh Stearns reports that in 4.0.3.6, doing ~! from a cu in xterm kills xterm. This has been fixed in Dell 2.2. 25. csh lossage due to bad optimization If a csh user sources a non-existent file in their .cshrc (eg, source .alias, where .alias doesn't exist), then the system will hang for a couple of minutes. Eventually the user get an "Out of memory" error and the console logs "NOTICE: out of swap space - Insufficient memory to allocate 2 pages - system call failed". This appears to be due to over-optimization of code surrounding a longjmp call. (There are numerous other reports of memory leak bugs in csh). 26. Bug in cp(1) If ``copy'' encounters a directory before a file, it dumps core ... --- cut --- cd /tmp mkdir copybug jnk cd jnk mkdir directory >file cp -r * /tmp/copbug --- cut --- This was reported from Consensys 4.0.3 but is probably a generic SVr4 bug. 27. tbl -me doesn't work Wolfgang Denk reports that trying to use "tbl -me" for any input file causes tbl to quit. The problem is that newer tbl versions don't accept [nt]roff contol lines (".rm @W") after .TS. 28. who -r fragility leads to boot-time problems It coredumps if the name of the timezone is longer than three characters. This can be a real problem for European sites... and is potentially more hazardous than immediately apparent as _a lot_ of the initialization scripts (rc1.d, rc2.d) use ``who -r'' to see if the machine is in single- or multi-user mode. And when ``who'' bombs out, the ``set'' command is iven an empty command-line and can't do much else than print the shell variables, $1-$9 remain empty ... meaning that more or less all the scripts fail in various ways and the system has an exceptionally hard time coming up. 29. at(1) breaks here-documents in shell scripts at adds gratuitous empty lines to the job submitted by the user. This prevents shell here-documents from working. III. Networking and File Sharing Bugs 1. NFS locking is unusably slow Randy Terbush has posted code which demonstrates a serious bug in the SVr4 NFS locking daemon. In his own words: "The symptoms are ~30% cpu usage by 'lockd' and severe slowing of the machines on the network. This program demonstates that it takes ~20 seconds to obtain locks from an ailing 'lockd'. We have verified that this bug does not exist in HPUX 8.0x." Randy's code is too large to be included here. He is, quite rightly, exercised at USL's exceedingly slow response to this problem. The comment in his makefile reads, in part: # USL has admitted to the existance of this bug in version 4.0, 4.1, # and 4.2 of their distributed and yet to be released sources. This is # a network crippling problem that they have refused to fix until # release 4.3, which will be OVER 1 YEAR from today. (29 Oct 1992) # If your version of 'lockd' exhibits this same problem, I would # strongly urge you to contact your vendor and ask them to put some # pressure on USL to fix this problem. SVR4 is virtually useless in a # network of shared resources while this problem exists. 2. UFS file system problems In stock USL 4.0.3, you can't use a UFS file system as the root; the system hangs if you try. Consensys, Dell, Esix, Microport, MST, UHC, and ESIX all appear to have fixed this. David Aitken, the UNIX product manager at UHC, writes "The ufs as root file system [problem] was not really a bug, just a little oversight on USL's part - we have fixed it completely by adding one line to the /stand/boot script: rootfstype=ufs!" He adds that they've been using ufs on their lab machines for over 10 months with no trouble, and the latest UHC release defaults to ufs if you have more than 120MB of disk. 3. Byte-order problem with NFS when accessing Sun disks Christoph Badura notes that the stock USL resolver library suffers from serious confusion about the byte order in the socketaddr_in structure. This bug is acknowledged by USL for the 4.0.4 release. A symptom of this bug is that Sun disks will not mount correctly over NFS. As a workaround, try removing the references to /usr/lib/resolv.so from /etc/netconfig and rebooting your system. Unfortunately, this will mean you can't use nameservers. Alan Batie writes: "Actually, you don't have to remove resolv.so, just put tcpip.so first and have a hosts file with the names of hosts you want to do NFS mounts from. This way you can use nameservers for most things." 4. Under weird circumstances, lseek on UFS may cause corruption Christoph Badura reports that a UFS lseek() to an offset which is a multiple of 4096 but not a multiple of 8192, followed by a write(), may corrupt the file being written. The bug shows up only, if the file has no pages in the page pool associated with it at the seek offset and at 4k before the seek offset. He has sent USL kernel fix for this, which was included in 4.0.4. 5. FTP problems The in.ftpd on SVR4.0.3 does not support all the commands listed in RFC 959. When recent SCO UNIX/ODT versions ftp to SVR4.0.3, the SVR4 side will refuse, drop the connection, and core dump after you authenticate. This is because the SCO end sends the 'SYST' command ala RFC 959, and the SVR4.0.3 end doesn't recognise it. Some ports have fixed this. Christoph Badura adds: "The bug is do to a longjmp(3) on a sigjmpbuf obtained by sigsetjmp(3). ARGH. Testing led to a bug in the original BSD sources, which is still present in the NET/2 ftpd. " 6. A bug in the WD80x3 support MST reports a serious bug in the SVr4 kernel support for this card. Here's how to reproduce it: server: init 3 and share (export) /usr for example. client: mount -F nfs server:/usr /mnt cd /mnt find . -print | cpio -ocBuv > /dev/null what happens: server and client will "hang" together. "cue": hit keys on server and/or client, hang will go away for 10-20 seconds temporarily. Yank BNC connectors do the same trick. They say they've heard from customers that this happens on Dell, UHC as well as USL 4.0.4. PCNFS/BWNFS network xcopy suffers this as well. Client can be a Sun Sparc for that matter. IV. SCSI Support Problems 1. sar is confused by SCSI Sar -d doesn't work on SCSI drives. Dell fixed this in 2.1 and it's reported to work OK in Esix 4.0.3A; no report of any other SVr4 having fixed this yet. SCO fixed it in 3.2.4. 2. A configuration problem Stock USL requires you to jumper your SCSI devices to fixed IDs during installation (it can be changed to any other ID after). Dell says they've fixed this. The requirement is definitely still present in Esix and Consensys 1.3. UHC thinks they've fixed this, but their 4.0.3.6 release still seems to demand ID 1 to install. 3. Synchronous SCSI hang problem David Wexelblat reports: "Stock SVR4.0.3 will hang the SCSI bus with a 1542 in synchronous mode. Dell fixed this, and this has been given to Microport [ed note: Microport 4.0.4 and Consensys 4.0.3 have fixed the problem; MST UNIX and Esix 4.0.3 still have this problem; I have not yet been able to determine if ESIX 4.0.4 does]. In the file /sbin/bcheckrc, change the line: echo MARK > /dev/rswap to echo MARK | dd of=/dev/rswap bs=512 conv=sync > /dev/null 2>&1 The magic is apparently the conv=sync, which forces a 512 byte block to be written. The original echo writes 4 bytes, which apparently causes synchronous SCSI to go out to lunch. Now, you ask, how can I fix this, since the system won't boot? There are a couple of methods. First, if possible, disable synchronous negotiation (1542 jumper J5-1 removed, plus whatever you may need to do to your drive). Then boot up, edit /sbin/bcheckrc, then shutdown, restrap for synchronous, then reboot. Everything should be OK. That's the easy way. Unfortunately, some hard drives will only work in synchronous mode. Well, you can still recover from this phenomenon. Here's how: 1) Install on your hard drive 2) Boot from the first boot floppy. When it tells you to, insert the second boot floppy. At the first prompt, hit to break out to a shell. 3) Mount your hard drive under /mnt with the following command (replace FS-TYPE with s5, s52, or ufs, whichever you used for for your root partition): /etc/fs/FS-TYPE/mount /dev/dsk/c0t0d0s1 /mnt 4) Now edit /mnt/sbin/bcheckrc: ed /mnt/sbin/bcheckrc You may want the 'ed' man page handy (I barely remember how to to use 'ed' :->). For simplicity, you can delete/comment out the offending line, then replace it with the correct line later. 5) Unmount the hard drive: umount /mnt 6) Reboot from the hard drive. Everything should come up OK. and you can finish editing /sbin/bcheckrc, if necessary. Note that you perform these actions at your own risk. The first version was performed by me on Microport SVR4, and the second was performed by someone else (on my suggestion) on ESIX SVR4." This problem appears to be fixed on Consensys 1.3 and Dell 2.1. 4. ps chokes on commands that do SCSI I/O Hugh Stearns reports that in 4.0.3.6, ps doesn't work when a SCSI command in progress. It stops printing at the process executing the scsi command. This is still broken in Dell 2.2. 5. Transfer speed problems with Adaptec 1542B on 486s If a system mount or install fails, try setting the DMA speed to 5MB/s, rather than the default 5.7MB/s. This is accomplished by removing the jumper shorting the 12th pin pair of jumper block 5. V. Development Tools Problems 1. General UCB library brokenness The BSD compatibility libraries were badly broken in USL code. A Dell source adds "That meant that almost all the apps derived from them were broken too. Most stuff like automount will die when you send a SIGHUP, instead of rereading the map file. You can get a system into very strange states when that happens." John Sully of Microport opines: "This is a bug in automount itself rather than BSD compatibility, since the automount which comes with SVR4 is not compiled with the BSD libraries. (isn't this comforting?? :-()." Esix and UHC's BSD libraries are USL stock. I don't yet know the status of other ports. Microport has run into things they think may be symptoms of this but have no fix yet. John Sully of Microport counters with: "One common thread I find on reading of these problems is that the BSD compatibility libraries are *misused*. [...] The problem is that BSD and SYSV have similarly named .h files which sometimes contain different definitions for objects with the same name. This has been known to cause all sorts of problems because the SYSV headers are picked up and then the calls are satisfied from the BSD library rather than the shared object library. I have found that if you use /usr/ucb/cc that the BSD compatibility is much less broken than it would seem at first because it ensures that the correct headers are picked up." However, note that there is at least one *real* bug known --- as of 4.0.4 the signal emulation cannot explicitly set a handler to SIG_DFL or SIG_IGN. Ron Guilmette writes "[Library lossage] may be easily demonstrated by attempting to build and link the GNU C compiler with `-L/usr/ucblib -lucb'. The resulting compiler will most certainly crash and die." John Sully thinks this is because the /usr/ucb/cc compiler should have been used, but wasn't. 2. USL emulation of BSD signals doesn't work A different source reports that the the USL implementatation of BSD signals is broken in both 4.0.3 and 4.0.4; in particular, the sigvec() family doesn't work properly. It is possible to make minor tweaks to source to make such apps work properly with the native USL signals implementation. Here's more on the signals problem, thanks to Richard : ------------------------------------------------------------------------------ The problem is to do with the signal() function that is within the BSD compatability libc. To reproduce the problem do the following: #include #include #include #include main() { signal(SIGPIPE,SIG_IGN); pause(); } and compile it with cc xx.c -o xx /usr/ucblib/libucb.a (John Sully observes that this is definitely wrong; /usr/ucb/cc should have been used rather than "cc ... -L/usr/ucblib -lucb" or the equivalent "cc ... /usr/ucblib/libucb.a".) If you run the program and then signal it with a SIGPIPE, the program will die, even though you've told it to ignore SIGPIPE. The fix is difficult unless you've got source because there's a missing 'else' clause from the signal() code. This is the only signal fault I've found in the BSD signal functions, details of the rumoured sigvec problem would be useful? If you're trying to compile an application you could change the application code to do the following, this does work.. void catch(s) int s; { /* DO NOTHING */ ; } main() { signal(SIGPIPE,catch); pause(); } SUMMARY You can only change a signal handler to a function handler, any number of times. Any attempt to set the handler to SIG_DFL, or SIG_IGN will fail. This bug has given some people working with X11R5 aggro, causing the X server to die when you close a client. Christoph Badura confirms this bug He has sent USL a source fix. It appears already to have been fixed in Dell 2.2. ------------------------------------------------------------------------------ 3. Possible string library problems There are also persistent rumors of problems in the BSD-emulation string libraries. I have not been able to pin down specifics on this. 4. USL's ndbm support is broken. Christoph Badura reports "The ndbm functions in the ucb library are broken [apparently due to a compiler of optimizer bug in cc -- ed.]. Try makeing the whatis data base for /usr/share/man with Tom Christiansen's perl rewrite of man. The easiest way to fix this is to compile ndbm.c with gcc -fpcc-struct-return -traditional (gcc1.40 or 2.2 will do nicely). This option is only available to the SVR4 vendors." 5. An include file is missing Both 4.0.3 and 4.0.4 USL versions are missing the documented dial.h file from their /usr/include directory. Dell 2.1 has it. 6. sscanf(3) has a potential bug Anthony Shipman reports: " I found the following bug in SCO Unix 3.2.* and I think it may be common to many AT&T derived Unixes. sscanf() calls _doscan() to read from a pretend file. The file uses the string as a buffer and a fake file descriptor of 60 (=_NFILE). Since _NFILE (for SCO UNIX) is 60 it assumes that fd 60 can never be open. Then when fscanf() hits the end of the string it calls _filbuf() to read into the buffer (which is the string) from fd 60. This should fail with an errno=9 and then _filbuf() sets EOF and it all terminates. However in SCO Unix you can reconfigure the kernel to increase the number of files per process to a recommended maximum of 150. If you do this then your program might have fd 60 open one day. Then sscanf() will read from this file overwriting your string. The byte count to the read() in _filbuf() is some undefined but large value so a lot of memory will be overwritten. In my case the string was on the stack so my stack was wiped. In short if you configure your kernel to have NOFILES > _NFILE ie more than the default then sscanf() is a time bomb in your code." 7. Compiler problems Ronald Guilmette also reports the following: ------------------------------------------------------------------------------ /* Here is a bug in the original SVR4 C compiler (aka C Issue 5) which effectively prevents you from making good use of the `const' and `volatile' qualifiers defined by ANSI C in conjunction with pointer types and typedef statements. Compile this code and you will get: "qualifiers.c", line 23: left operand must be modifiable lvalue: op "=" ...if your copy of the svr4 C compiler still has the bug. Note that given these declarations, the ANSI C standard say that the thing pointed to by the variable `pci' should be considered to be constant... not the variable `pci' itself. (The GCC compiler, either version 1.x or version 2.x, correctly compiles this example without complaint.) */ typedef const int *ptr_to_const_int; ptr_to_const_int pci; int i; void main () { pci = &i; } ------------------------------------------------------------------------------ /* Here is a subtle bug in the original SVR4 C compiler (aka C Issue 5) which prevents you from first declaring a tagged type (i.e. a struct type or a union type) in a parameter list, and then defining that tagged type later on within the same scope. (Note that according to the ANSI C standard, the scope in which parameters get declared and the outermost block of a function body are one and the same scope. Thus, this really is legal ANSI C code!) Try compiling this with your C compiler on SVR4. If your compiler still has the bug, you will get: "tagged_type.c", line 24: warning: dubious tag declaration: struct S "tagged_type.c", line 28: warning: improper member use: i "tagged_type.c", line 28: warning: improper member use: i "tagged_type.c", line 31: warning: dubious tag declaration: struct S "tagged_type.c", line 35: warning: improper member use: i "tagged_type.c", line 35: warning: improper member use: i (The GCC compiler also had this bug in version 1.x, but it has been fixed in version 2.x.) */ void foobar1 (arg) /* use old-style without prototypes */ struct S *arg; { struct S { int i; }; /* define the type `struct S' */ arg->i = arg->i; /* legal according to ANSI C rules! */ } void foobar2 (struct S *arg) /* use new-style with prototypes */ { struct S { int i; }; /* define the type `struct S' */ arg->i = arg->i; /* legal according to ANSI C rules! */ } ------------------------------------------------------------------------------ /* Here is a serious bug in the original SVR4 `dump' program which dumps out parts of object files in either plain hex form or symbolically. To see the `dump' program get a segfault and die, save this code under the name `dump-bug.c' and then do: cc -g -c dump-bug.c dump -v -D dump-bug.o The bug arises whenever `dump' tries to read Dwarf debugging information for an array of pointers to any "user defined" type (e.g. `struct S' in this example). Past that point, `dump' is totally confused, so further Dwarf debugging information finally causes it to go belly-up. */ struct S { int i; }; struct S *array[10]; int j; ------------------------------------------------------------------------------ It appears that the svr4 C compiler (for x86 machines) doesn't conform real well to either the letter or the spirit of the IEEE 754 floating-point standard. In particular, "unordered comparisons" and other operations on NaNs don't always produce the result that that the IEEE 754 standard calls for. An AT&T source comments: "This is documented in the SVID as a future direction. We do not support NaNs in -Xa and -Xt modes, only in -Xc. Try isnan(sqrt(-1.0)) to determine which modes support it." ------------------------------------------------------------------------------ The compiler fails to issue diagnostics for cases where a floating point literal is given which exceeds the range of its type (either float or double). Actually this one could be argued either way, since IEEE FP format includes "infinities" and the compiler probably just changes any FP value which is out of range for its type into either positive infinity or negative infinity (as appropriate). The compiler fails to issue diagnostics in cases where a typedef name is reused to declare a formal parameter, as in: ----------------------------------------------------------------------- typedef int FOO; void bar (FOO) int FOO; { } ----------------------------------------------------------------------- The compiler crashes on the following invalid input: ----------------------------------------------------------------------- int i; volatile void *pvv; void pvv_test () { (i ? *pvv : *pvv); /* ERROR */ } ----------------------------------------------------------------------- The compiler fails to issue diagnostics for cases where an attempt is made to "forward declare" an enum type (without also defining it), as in: ----------------------------------------------------------------------- enum enum0 *ep; /* ERROR */ ----------------------------------------------------------------------- The compiler rejects the following code with an error, although there seems to be no good reason why it should (because no object is being declared). ----------------------------------------------------------------------- #include typedef char array_type[ULONG_MAX]; ----------------------------------------------------------------------- VI. The FUBYTE Problem (Thanks to Christoph Badura for this info) The kernel function fubyte() is documented to return a positive value when given a valid user space address and -1 otherwise. In the latter case u.u_error is set to EFAULT. USL SysV R4.0.3 has a sign extension bug in the implementation of fubyte() for local file descriptors (i.e. not opened via RFS), which causes fubyte() to return negative values if the byte fetched has its high bit set. This bug doesn't affect STREAMS drivers, as they don't call (and in fact are normally unable to call) fubyte(). Thus writing a byte with the high bit set to certain character device drivers returns with -1 and errno set to EFAULT. The bug may affect any character device driver that calls fubyte(). It's not limited to serial card drivers. The bug is noticed most often with serial card drivers, since uucp uses byte values > 127 very early during g-protocol setup and drivers for serial cards tend to use fubyte() quite often. Note also that the bug's effect is different if the driver checks for a -1 return value of fubyte() or just a negative one. In the former case it is possible to pass bytes with the 8 bit set through fubyte(), except for 0xff which is -1 in two's complement. That makes the bug more obscure. The fix is easy. First, make a backup copy of the kernel object file /etc/conf/pack.d/kernel/vm.o! A disassembly of vm.o(lfubyte) should reveal *exactly* one mov[s]bl (move byte to long w/sign extend). That one needs to be patched into a movzbl (zero extend). The difference is one bit in the second byte of the opcode. The movsbl has the bit pattern 00001111 1011111w mod/rm-byte. The movzbl has the bit pattern 00001111 1011011w mod/rm-byte. The 'w' bit is 0 for the instruction in question. So the opcodes are 0f be and 0f b6. Here is the diff -c from dis -F lfubyte showing the patch applied to the Dell 2.1 kernel: *** vm.o Mon Mar 9 00:31:38 1992 --- vm.o.org Mon Mar 9 00:32:40 1992 *************** *** 22,28 **** 11c90: 85 c0 testl %eax,%eax 11c92: 75 09 jne 0x9 <11c9d> 11c94: 8b 45 08 movl 8(%ebp),%eax ! 11c97: 0f b6 00 movzbl (%eax),%eax 11c9a: 89 45 fc movl %eax,-4(%ebp) 11c9d: c7 05 d8 13 00 00 00 00 00 00 movl $0x0,0x13d8 11ca7: 83 3d dc 13 00 00 00 cmpl $0x0,0x13dc --- 22,28 ---- 11c90: 85 c0 testl %eax,%eax 11c92: 75 09 jne 0x9 <11c9d> 11c94: 8b 45 08 movl 8(%ebp),%eax ! 11c97: 0f be 00 movsbl (%eax),%eax 11c9a: 89 45 fc movl %eax,-4(%ebp) 11c9d: c7 05 d8 13 00 00 00 00 00 00 movl $0x0,0x13d8 11ca7: 83 3d dc 13 00 00 00 cmpl $0x0,0x13dc Of course there is a workaround at the driver level. Canonically, one would do this by checking for fubyte() returning -1 *and* u.u_error being set to EFAULT (u.u_error is cleared upon entering a system call). However, in R4.0.3 fubyte() does NOT set u.u_error. It *does* set u.u_fault_catch.fc_errno. Cristoph reports that Dell V.4 can be object-patched successfully to fix this. I'm told that the offending 11c97 is at exactly the same address in the Consensys 1.3 kernel. I do not know the status of the other ports. Another poster (Marc Boucher ) adds: On ESIX SVR4.0.3 Rev. A, the instruction movsbl in question can be changed to movzbl (as described above) with a binary-editor on file /etc/conf/pack.d/kernel/vm.o. At offset 0x11eb0, change 0xbe to 0xb6. Before patching, verify that your /etc/conf/pack.d/kernel/vm.o is the same as mine! On my system, the /bin/sum generated checksum of vm.o was "4440 222". The problem results from a sign-extension bug. The function lfubyte(), which is called by fubyte(), is declared as int lfubyte(char *addr); /* actually caddr_t */ The byte is fetched with val = *addr; which triggers sign extension. Casting addr to a unsigned char * or declaring it as such solves the problem. This bug is still present in stock USL 4.0.4. However, it has been fixed in Dell 2.2. VI. Destiny and Dell A source at at UNIX System Labs Europe claims that `Destiny' (the new Release 4.2) incorporates all of Dell UNIX's fixes to 4.0.3; thus, any bug for which a Dell fix is indicated above should be gone in Destiny. -- Send your feedback to: Eric Raymond = esr@snark.thyrsus.com