Signal 11 while compiling the kernel

This FAQ describes what the possible causes are for an effect that bothers lots of people lately. Namely that a linux(*)-kernel (or any other large package for that matter) compile crashes with a "signal 11". The cause can be software or (most likely) hardware. Read on to find out more.
(*) Of course nothing is Linux specific. If your hardware is flaky, Linux, Windows 3.1, FreeBSD, Windows NT and NextStep will all crash.
If you are not reading this at http://www.BitWizard.nl/sig11/, that's where you can find the most recent version.
For those of you who prefer reading this in French, the French translation can be found at http://www.linux-france.org/article/sig11-fr/.
For those of you who prefer reading japanese, the Japanese translation can be found at http://www.linux.or.jp/JF/JFdocs/GCC-SIG11-FAQ/.
mailto:r.e.wolff@BitWizard.nlif you find any spelling errors, worthwhile additions or with an "it also happened to me" story. (Note that I reject some suggested additions on my belief that it is technical nonsense). I would appreciate it if you put "sig11" or something like that in the subject. You can also Email me about other subjects.

The Sig11 FAQ

QUESTION

Signal 11, what does that mean?

ANSWER

Signal 11, or officially know as "segmentation fault", means that the program accessed a memory location that was not assigned. That's usually a bug in the program. So if you're writing your own program, that's the most likely cause. However, this FAQ will concentrate on the possibilities besides that.

QUESTION

My (kernel) compile crashes with

      gcc: Internal compiler error: program cc1 got fatal signal 11

What is wrong with the compiler? Which version of the compiler do I need? Is there something wrong with the kernel?

ANSWER

Most likely there is nothing wrong with your installation, your compiler or kernel. It very likely has something to do with your hardware. There are a variety of subsystems that can be wrong, and there are a variety of ways to fix it. Read on, and you'll find out more. There are two exceptions to this "rule". You could be running low on virtual memory, or you could be installing Red Hat 5.x, 6.x or 7.x. There is more about this near the end.

QUESTION

Ok it may not be the software, How do I know for sure?

ANSWER

First lets make sure it is the hardware that is causing your trouble. When the "make" stops, simply type "make" again. If it compiles a few more files before stopping, it must be hardware that is causing you troubles. If it immediately stops again (i.e. scans a few directories with "nothing to be done for xxxx" before bombing at exactly the same place), try

        dd if=/dev/HARD_DISK of=/dev/null bs=1024k count=MEGS

Change HARD_DISK to "hda" to the name of your harddisk (e.g. hda or sda. Or use "df ."). Change the MEGS to the number of megabytes of main memory that you have. This will cause the first several megabytes of your harddisk to be read from disk, forcing the C source files and the gcc binary to be reread from disk the next time you run it. Now type make again. If it still stops in the same place I'm starting to wonder if you're reading the right FAQ, as it is starting to look like a software problem after all.... Take a peek at the "what are the other possibilities" question..... If without this "dd" command the compiler keeps on stopping at the same place, but moves to another place after you use the "dd" you definitely have a disk->ram transfer problem.

QUESTION

What does it really mean? Are you sure it's a hardware problem?

ANSWER

Well, the compiler accessed memory outside its memory range. If this happens on working hardware it's a programming error inside the compiler. That's why it says "internal compiler error". However when the hardware occasionally flips a bit, gcc uses so many pointers, that it is likely to end up accessing something outside of its addressing range. (random addresses are mostly outside your addressing range, as not very many people have a significant part of 4G as main memory... :-) It seems that nowadays, everybody with "signal 11" problems gets directed to this page. If you're developing your own software or have software that hasn't been debugged quite enough, "signal 11" (or segmentation fault) is still a very strong hint that there is something wrong with the program. Only when a program like "gcc" that works for almost everybody else to crash on a dataset (e.g. the Linux-kernel) that has also been well-tested, then it becomes a hint that there is something wrong with your hardware. If some software component like a hardware driver in your system is broken, it could cause symptoms that are VERY close to those of a hardware failure. However, when a driver is faulty it is more likely to cause serious trouble inside the kernel, than just causing the compiler to crash.

QUESTION

Ok. I may have a hardware problem what is it?

ANSWER

If it happens to be the hardware it can be:

Main memory. Your main memory might be getting an occasional bit wrong. If this happens on the "writes", you won't see any parity errors. There are several ways to fix it:
- The memory speed might be too slow. Increase the number of wait states in the BIOS.
  This could be caused by the AMIBIOSs autoconfig option: it may only know about 486s running upto 80 MHz, whereas you currently buy 100 MHz versions. -- Pat V.
- The memory speed might be too slow. Get faster DRAM SIMMs. For example current ASUS motherboards require 60 ns DRAM if you have a 100, or 133 MHz processor (Take a look in your motherboard's manual). I've heard reports that 70 ns also works, reliability problems like random sig11's belong to the possibilities.... (I wouldn't take the risk) -- Andrew Eskilsson (mpt95aes@pt.hk-r.se)
- You might think that you can run your 100MHz SDRAMs at 100MHz. Wrong! read http://www.bitwizard.nl/sig11/sdram.html why I think this is the case. You need at least one speed grade faster than the speed they are rated for.
- There is a bad chip on one of the SIMMs. If you own more than 1 bank of memory you might be able to pull SIMMs and see if the problem goes away. Be careful for STATIC!!!
- We handled a hard one here the last week. It turned out that ALL 4 16Mb SIMMs were broken in that they dropped a bit around once per hour. This was sufficient to crash the machine in about a day, or crash a kernel compile in about an hour. A new set of SIMMs works perfectly. It took a long while to diagnose this one, because all 4 of the SIMMs were affected equally, so leaving half of the memory out didn't change things.
  Mark Kettner (kettner@cat.et.tudelft.nl) reports that his system was capable of running my memory test for 2300 times faultlessly, but then detected around 10 errors. It then continued detecting no faults for a few hundred runs again..... In his case running kernel compiles was a much more efficient way of detecting the health of the system (in the most stable configuration the system could compile around 14 kernels before going bzurk). His solution was to "trade in" the old memory for a so called "memory upgrade". The shopkeeper then "tests" in their memory tester, which OKs the memory. He then got a good discount on the new memory :-).
- It seems that some 30-72 pin converters can cause memory errors. (See how old this entry is? Who remembers 30pin SIMMs? However all these things hold perfectly for SIMM <-> DIMM converters, or socket370 <-> slot 1 converters) (It hasn't been proven whether the 4 SIMMS in the converter had gone bad, or if the SIMM converter was at fault. The SIMMS had been functioning perfectly for years before they were moved into the converter....) -- Naresh Sharma (n.sharma@is.twi.tudelft.nl). Paul Gortmaker (paul.gortmaker@anu.edu.au) adds that the SIMM converters should have at least 4 bypass capacitors to keep the power supply of the SIMMs clean.
- If the refresh of the DRAM isn't functioning properly, the DRAMs will slowly lose their information. Some (486) motherboards stop refreshing correctly when you turn on "hidden refresh". There seems to be a program called "dram" around that can also mess up your refresh to cause sig11 problems. -- Hank Barta (hank@pswin.chi.il.us), Ron Tapia (tapia@nmia.com)
- The number of wait states could be too low. Increase the number of wait states in the BIOS for a fix. The Intel Endeavour board doesn't allow you to increase the memory wait states. This can supposedly be fixed by flashing a MR BIOS into the motherboard. -- David Halls (david.halls@cl.cam.ac.uk)
- Some memory modules simply don't like to work together with others. Separately they both work, together they don't. This is most likely to happen if you mix different brands and sizes. Officially if you stick to the specs for all the modules, it always works. Unofficially you sometimes run into trouble.
Cache memory. Your cache memory might be getting an occasional bit wrong. Caches are usually not equipped with parity. You can diagnose that this is the case by turning off the cache in the BIOS. If the problem goes away it is probably the cache. There are several ways to fix it:
- The cache memory speed might be too slow. Increase the number of wait states in the BIOS.
- The cache memory speed might be too slow. Get faster SRAM chips.
- There is a bad chip in your cache. It is unlikely that you can swap chips as easily as with SIMMs. Be careful for STATIC!!! -- Joseph Barone (barone@mntr02.psf.ge.com)
- The cache might be set to "write back" while there is a bug in the write back implementation of your chipset. The motherboard where this happened was a "MV020 486VL3H" (with 20M RAM) -- Scott Brumbaugh (scottb@borris.beachnet.com) (Mail address doesn't work. Scott: Get back at me with a valid return address)
- The motherboard may require a jumper to switch between Cache On A Stick and the old-fashioned dip chip cache. (JP16 on Rev 2.4 ASUS P/I-P55TP4XE motherboards)
Disk transfers. A block coming from disk might incur an occasional bit error.
- If you have this problem, you are most likely to have to do the "dd" command to "move" the problem from one place to the next....
- Some IDE harddisks cannot handle the "irq_unmasking" option. This may only show under load. And it could show as a sig11.
- Some setups cannot handle DMA in some configurations. Mario Moder reports that his system finally started working correctly after enabling 32-bit-io for both his HD and his CD drive. -- Mario Moder (clay-man@freenet.de)
- Do you have a kalok 31xx? Throw it in the garbage. (or sell it to a DOS user. Update: Haven't heard about kalok for years. They're probably bust. The drives also don't work with W95 by the way.)
- SCSI? Termination? A short bus might still work (unreliably that is) with bad termination. A long bus might get errors anyway. Can you turn on parity on the host and the DISK?
The CPU itself. Some batches of processors have a much higher percentage of them that happen to be "bad". Some years ago: original Intel-Pentium-120's. A few years ago AMD K6/2-300's (1998, produced in weeks 34 through 39!). And recently AMD K6/2-450's. Some people may decide that say 400MHz is acceptable to them, however if this turns out to be the problem, you're entitled to a new processor. Go and exchange it where you bought it. (Forget about those P120's, it's not worth the trouble... ;-) -- Guillaume Cottenceau (gcottenc@ens.insa-rennes.fr) & Mark Keegan (keegan@mx.qc.ca)
The CPU itself. Some batches of K6 processors simply have a design bug. Read http://www.multimania.com/poulot/k6bug.html and then make sure you get your K6 exchanged. -- Rongen (rongen@istar.ca).
Overclocking. Cyrix P-166 processors run at 133MHz, not at 166. This must be logical to the guys at Cyrix, but nobody else. You're overclocking them if you run them at 166Mhz..... (Note: Some of this FAQ is pretty old. Now AMD has started doing the same thing: XP1800's run at 1533MHz.)
Overclocking. Some vendors (or private people) think it is possible to overclock some CPUs. Some of them may work others don't. You might want to try turning off turbo (note that most pentium motherboards no longer support a non-turbo mode) and see if the problem goes away. Check the speed of your CPU compared (printed on it, carefully remove the fan if necessary) with what the motherboard jumpers or BIOS settings say.... It seems that even Intel may make mistakes in this area. I now have several reliable reports that official pentium would sig11 at their rated speed, but not at a lower speed. As for some speeds the motherboard is only stressed HARDER for a slower processor speed, (120 MHz-> motherboard runs at 60MHz, 100MHz-> motherboard runs at 66MHz), I think it is unlikely that this has anything to do with the motherboard. Moreover a new 120MHz processor is now functioning correctly. -- Samuel Ramac (sramac@vnet.ibm.com). This is not unique to Intel or any of its competitors.
Overclocking. Currently, processor speeds, power dissipation etc etc. are all so close to "the edge" that every now and then dependable corporations like Intel have to resort to the tricks that Overclockers use to improve performance. The results can also be comparable: Random lockups, sig11's etc. etc.
CPU temperature. A high speed processor might overheat without the correct heat sink. This can also be caused by a failing fan. (My personal '486 has a fan that takes a few minutes to get up to speed. It probably will never really FAIL because it's now decommissioned :-). The CPU can become erratic if "pushed" by compiling a kernel. This problem becomes worse if you disable "HALT" on the LILO command line. Linux tries to power-down the CPU by executing the "halt" instruction when the system is idle. This preserves power, and therefore the CPU temperature drops when the system is idle. You therefore might not notice this problem when simply editing, and it might only surface after hours of CPU intensive jobs when the ambient temp is high. If you have a Pentium with Fdiv bug, it is advisable to trade it in at Intel. They will send you a new one that pre-configured with an official Intel-approved FAN. Also note that most normal glues are very bad thermal conductors. There is special thermal glue available that should be used when a fan needs to be glued to a CPU. -- Arno Griffioen (arno@ixe.net), -- W. Paul Mills (wpmills@midusa.net) -- Alan Wind (wind@imada.ou.dk)
Intel says that the allowable temperature ranges for the outside of your CPU is:
0 to +85 C: Intel486 SX, Intel486 DX, IntelDX2, IntelDX4 processor
0 to +95 C: IntelDX2, IntelDX4 OverDrive® processors
0 to +80 C: 60 MHz Pentium® processor
0 to +70 C: 66 to 166 MHz Pentium processor
For information on how to measure this and some confirmation of what I say here, see: http://pentium.intel.com/procs/support/faqs/iarcfaq.htm (Especially questions Q5, Q6 and Q12. The document is getting slightly outdated, but it is still very accurate. It seems the questions move around a bit every now and then as well.) (Intel has now moved the file. I couldn't find the new one. Is there someone who can point me to the new location???)
CPU voltage. Some motherboards allow you to select the CPU voltage. Some motherboards badly document the jumper settings that manage this. It seems that a 5V processor might still work most of the time at 3.3 volts..... -- Karl Heyes (krheyes@comp.brad.ac.uk)
RAM voltage. It seems that vendors are preparing for 3.3V RAM now. Most memory is now 3.3V. (but be careful if you have a board capable of setting the RAM voltage: 3.3v RAM will break at 5V.....) (Having heard little about this, I think the switch must be automatic.)
Local bus overloading. At 25 MHz you're allowed to have 3 VesaLocalBus (VLB) cards, At 33MHz only two, at 40MHz only one and guess what at 50MHz NONE! (i.e. you are allowed to run your system with a 50MHz local bus, but then you're not allowed to use any VLB cards). Some systems start acting flaky when you overload the VLB. Even when your VLB isn't overloaded (over the limits stated above), the system may lose a few nanoseconds of margin by adding an extra VLB card, so you might need to add a cache wait state or something after you've added a new VLB card.... -- Richard Postgate (postgate@cafe.net)
Power management. Some laptops (and nowadays also "green" pc's) have power management features. These might interfere with Linux. One feature might save a memory image to HD and restore the RAM when you press a key. This sounds like fun, but Linux device drivers don't expect that the hardware has been turned off between two accesses. Some may recover, but others not. Try turning it off, or enabling "APM support" in your kernel. -- Elizabeth Ayer (eca23@cam.ac.uk)
Dust buildup. Some dust might conduct a bit and create a weak short. It might increase capacitances somewhere, and degrade timing characteristics. It might impede thermal flow, and lead to overheating components. It might even short a jumper connection! I recommend that every year or so, it is a good idea to open up your computer, and vacuum the inside. Tip: Those cotton-on-a-stick thingies help prodding the dust out of inaccessible spots... -- Craig Graham (c_graham@hinge.mistral.co.uk). Someone else told me: if you can, use pressurized air to blow things out instead of the vacuum cleaner. Preferably do this outside.
The CPU itself. Several people are reporting that they have found nothing to blame except the CPU. This could also have been an incompatibility between the CPU and the motherboard. A wave of reports concerning Intel CPUs has passed (Feb '97). A new wave of reports is coming in that are blaming Cyrix/IBM 6x86 CPUs. Although it could indeed be the CPU, it could also be that your motherboard is incompatible with your CPU. At least I've seen a motherboard manual mention that it isn't compatible with older 6x86's. My own experience is that these devices aren't bad at all, and on a kernel compile I benchmarked a P166+ to be equivalent with a P155 (1.3 times faster than a P120).

The Memory hole. Many modern motherboards allow you to use old ISA video cards with one or two megabytes of linear frame buffer. To achieve this, they have to map out the memory just below 16Mb. Nobody actually ever used this feature, but if you turn the memory hole (or LFB support in some BIOSes) on, your machine will certainly be flaky..... -- Paul Connolly (pconnolly@macdux.com.au)

X and AMD incompatibility. There is a problem with a bunch of modern AMD chips that don't handle some operations quite as good as they should. If you have an AMD, and X11 often quits with "signal 11 caught", then you might be the victim of this issue. Try booting with "mem=nopentium". -- Matthew Beale (mixonic@synitech.com).

The Microcode. Especially on SMP systems, the CPUS may need an upgrade. Since the Pentium division disaster, Intel have their CPUs field upgradable! The CPU can be bumped a few versions by a special instruction from the BIOS. These upgrades usually come with your BIOS, so make sure you're running the latest BIOS, especially if you have an SMP system. -- Jeffrey Friedl (Email withheld).

QUESTION

RAM timing problems? I fiddled with the bios settings more than a month ago. I've compiled numerous kernels in the mean time and nothing went wrong. It can't be the RAM timing. Right?

ANSWER

Wrong. Do you think that the RAM manufacturers have a machine that makes 60ns RAMs and another one that makes 70ns RAMs? Of course not! They make a bunch, and then test them. Some meet the specs for 60 ns, others don't. Those might be 61 ns if the manufacturer would have to put a number to it. In that case it is quite likely that it works in your computer when for example the temperature is below 40 degrees centigrade (chips become slower when the temp rises. That's why some supercomputers need so much cooling).

However "the coming of summer" or a long compile job may push the temperature inside your computer over the "limit". -- Philippe Troin (ptroin@compass-da.com)

QUESTION

I got suckered into not buying ECC memory because it was slightly cheaper. I feel like a fool. I should have bought the more expensive ECC memory. Right?

ANSWER

Buying the more expensive ECC memory and motherboards protects you against a certain type of errors: Those that occur randomly by passing alpha particles.
Because most people can reproduce "signal 11" problems within half an hour using "gcc" but cannot reproduce them by memory testing for hours in a row, that proves to me that it is not simply a random alpha particle flipping a bit. That would get noticed by the memory test too. This means that something else is going on. I have the impression that most sig11 problems are caused by timing errors on the CPU <-> cache <-> memory path. ECC on your main memory doesn't help you in that case. When should you buy ECC? a) When you feel you need it. b) When you have LOTS of RAM. (Why not a cut-off number? Because the cut-off changes with time, just like "LOTS".) Some people feel very strong about everybody using ECC memory. I refer them to reason "a)".

QUESTION

Memory problems? My BIOS tests my memory and tells me its ok. I have this fancy DOS program that tells me my memory is OK. Can't be memory right?

ANSWER

Wrong. The memory test in the BIOS is utterly useless. It may even occasionally OK more memory than really is available, let alone test whether it is good or not.
A friend of mine used to have a 640k PC (yeah, this was a long time ago) which had a single 64kbit chip instead of a 256kbit chip in the second 256k bank. This means that he effectively had 320k working memory. Sometimes the BIOS would test 384k as "OK". Anyway, only certain applications would fail. It was very hard to diagnose the actual problem....
Most memory problems only occur under special circumstances. Those circumstances are hardly ever known. gcc Seems to exercise them. Some memory tests, especially BIOS memory tests, don't. I'm no longer working on creating a floppy with a linux kernel and a good memory tester on it. Forget about bugging me about it......
The reason is that a memory test causes the CPU to execute just a few instructions, and the memory access patterns tend to be very regular. Under these circumstances only a very small subset of the memories breaks down. If you're studying Electrical Engineering and are interested in memory testing, a masters thesis could be to figure out what's going on. There are computer manufacturers that would want to sponsor such a project with some hardware that clients claim to be unreliable, but doesn't fail the production tests......

QUESTION

Does it only happen when I compile a kernel?

ANSWER

Nope. There is no way your hardware can know that you are compiling a kernel. It just so happens that a kernel compile is very tough on your hardware, so it just happens a lot when you are compiling a kernel. Compiling other large packages like gcc or glibc also often trigger the sig11.

People have seen "random" crashes for example while installing using the slackware installation script.... -- dhn@pluto.njcc.com
Others get "general protection errors" from the kernel (with the crashdump). These are usually in /var/adm/messages. -- fox@graphics.cs.nyu.edu
Some see bzip2crash with "signal 11" or with "internal assertion failure (#1007)." Bzip2 is pretty well-tested, so if it crashes, it's likely not a bug in bzip2. -- Julian Seward (jseward@acm.org)

QUESTION

Nothing crashes on NT, Windows 95, 98, Milennium or XP. It must be something Linux specific.

ANSWER

First of all, Linux stresses your hardware more than all of the above. Some OSes like the Microsoft ones named above crash in unpredictable ways anyway. Nobody is going to call Microsoft and say "hey, my windows box crashed today". If you do anyway, they will tell you that you, the user, made an error (see the interview with Bill Gates in a German magazine....) and that since it works now, you should shut up.
Those OSes are also somewhat more "predictable" than Linux. This means that Excel might always be loaded in the exact same memory area. Therefore when the bit-error occurs, it is always excel that gets it. Excel will crash. Or excel will crash another application. Anyway, it will seem to be a single application that fails, and not related to memory.
What I am sure of is that a cleanly installed Linux system should be able to compile the kernel without any errors. Certainly no sig-11 ones. (** Exception: Red Hat 5.0 with a Cyrix processor. See elsewhere. **)
Really Linux and gcc stress your hardware more than other OSes. If you need a non-linux thingy that stresses your hardware to the point of crashing, you can try winstone. -- Jonathan Bright (bright@informix.com)

QUESTION

Is it always signal 11?

ANSWER

Nope. Other signals like four, six and seven also occur occasionally. Signal 11 is most common though.

As long as memory is getting corrupted, anything can happen. I'd expect bad binaries to occur much more often than they really do. Anyway, it seems that the odds are heavily biased towards gcc getting a signal 11. Also seen:

free_one_pmd: bad directory entry 00000008
EXT2-fs warning (device 08:14): ext_2_free_blocks bit already cleared for block 127916
Internal error: bad swap device
Trying to free nonexistent swap-page
kfree of non-kmalloced memory ...
scsi0: REQ before WAIT DISCONNECT IID
Unable to handle kernel NULL pointer dereference at virtual address c0000004
put_page: page already exists 00000046
invalid operand: 0000
Whee.. inode changed from under us. Tell Linus
crc error -- System halted (During the uncompress of the Linux kernel)
Segmentation fault
"unable to resolve symbol"
make [1]: *** [sub_dirs] Error 139
make: *** [linuxsubdirs] Error 1
The X Window system can terminate with a "caught signal xx"

The first few ones are cases where the kernel "suspects" a kernel-programming-error that is actually caused by the bad memory. The last few point to application programs that end up with the trouble.

-- S.G.de Marinis (trance@interseg.it)
-- Dirk Nachtmann (nachtman@kogs.informatik.uni-hamburg.de)

QUESTION

What do I do?

ANSWER

Here are some things to try when you want to find out what is wrong... note: Some of these will significantly slow your computer down. These things are intended to get your computer to function properly and allow you to narrow down what's wrong with it. With this information you can for example try to get the faulty component replaced by your vendor.

Jumper the motherboard for lower CPU and bus speed.
Go into the BIOS and tell it "Load BIOS defaults". Make sure you write the disk drive settings down beforehand.
Disable the cache (BIOS) (or pull it out if it's on a "stick").
boot kernel with "linux mem=4M" (disables memory above 4Mb).
Try taking out half the memory. Try both halves in turn.
Fiddle with settings of the refresh (BIOS)
Try borrowing memory from someone else. Preferably this should be memory that runs Linux flawlessly in the other machine... (Silicon graphics Indy machines are also nice targets to borrow memory from)
If you want to verify if a solution really works try the following script:
```
   #!/bin/sh
   #set -x
   t=1
   while [ -f log.$t ] 
     do
     t=`expr $t + 1`
   done

   while true
     do
     make clean
     make -k bzImage > log.cur 2>&1
     mv log.cur log.$t
     t=`expr $t + 1`
   done
```
All the resulting logfiles should be the same (i.e. the same size, and the same contents). Every kernel build takes around 4 minutes on a 1GHz Athlon with 512Mb of memory. (and about 3 months on a 386 with 4Mb :-).
Another way to test if your current setup is stable might be to run "md5sum" on files of different sizes (dd if=/dev/random of=testfile bs=1024k count=). If you use a file twice the size of your RAM, you'll be exercising your disk. If you use a file 4 to 10 Mb smaller than your RAM, you'll exercise your RAM/CPU.
Whether this method catches all possible problems, however, is uncertain. Gcc executes lots of different instructions in different orders, and md5sum might simply not hit the right sequence of instructions that gcc does. But if md5sum leads to errors, it might do so quicker than a kernel compile. -- Rob Ludwick (rob@no-spam)

The hardest part is that most people will be able to do all of the above except borrowing memory from someone else, and it doesn't make a difference. This makes it likely that it really is the RAM. RAM used to be one of the priciest parts of a PC, so you would rather not arrive at this conclusion, but, I'm sorry, I get lots of reactions that in the end turn out to be the RAM. However don't despair just yet: your RAM may not be completely wasted: you can always try to trade it in for different or more RAM.

QUESTION

I had my RAMs tested in a RAM-tester device, and they are OK. Can't be the RAM right?

ANSWER

Wrong. It seems that the errors that are currently occurring in RAMS are not detectable by RAM-testers. It might be that your motherboard is accessing the RAMs in dubious ways or otherwise messing up the RAM while it is in YOUR computer. The advantage is that you can sell your RAM to someone who still has confidence in his RAM-tester......

QUESTION

What other hardware could be the problem?

ANSWER

Well, any hardware problem inside your computer. But things that are easy to check should be checked first. So, for example, all your cards should be correctly inserted into the mother board.

QUESTION

Why is the Red Hat install bombing on me?

ANSWER

The Red Hat 5.x, 6.x and 7.x install has problems on some machines. Try running the install with only 32M. This can usually be done with mem=32m as a boot parameter.

It could be that there is a read-error on the CD. The installer handles this less-than-perfect..... Make sure that your CD is flawless! It seems that the installer will bomb on marginal CDs!

People report, and I've seen with my own eyes, that Red Hat installs can go wrong (crash with signal 7 or signal 11) on machines that are perfectly in order. My machine was and still is 100% reliable (actually the machine I tested this on, is by now reliably dead). People are getting into trouble by wiping the old "working just fine" distribution, and then wanting to install a more recent Red Hat distribution. Going back is then no longer an option, because going back to 5.x also results in the same "crashes while installing".

Patrick Haley (haleyp@austin.rr.com) reports that he tried all memory configurations up to 96Mb (32 & 64) and found that only when he had 96Mb installed, the install would work. This is also consistent with my own experience (of Red Hat installs failing): I tried the install on a 32M machine.

NEW: It seems that this may be due to a kernel problem. The kernel may (temporarliy) run low on memory and kill the current process. The fix by Hubert Mantel (mantel@suse.de) is at: http://juanjox.linuxhq.com/patch/20-p0459.html.

If this is actually the case, try switching to the second virtual console (ctrl-alt-F2) and type "sync" there every few seconds. This reduces the amount of memory taken by harddisk-buffers... I would really appreciate hearing from you if you've seen the Red Hat install crash two or more times in a row, and then were able to finish the install using this trick!!!

What do you do to get around this problem?...

Use SuSE. It's better: It doesn't crash during the installation. (Moreover, it actually is better. ;-)
Maybe you're running into a bad-block on your CD. This can be drive-dependent. If that's the case, try making a copy of the CD in another drive. Try borrowing someone elses copy of Red Hat.
Try configuring a GIGABYTE of swap. I have two independent reports that report that they got through with a gig of swap. Please report to me if it helps!
Modify the "settings" for the harddisk. Changing the setting from "LBA" to "NORMAL" in the bios has helped for at least one person. If you try this, I'd really appreciate it if you'd EMail me: I would like to hear from you if it helps or not. (and what you exactly changed to get it to work)
I got my machine to install by installing a minimal base system, and then adding packages to the installed system.
Someone suggested that the machine might be out-of-memory when this happens. Try having a swap partition ready. Also, the install may be "prepared" to handle low mem situations, but misjudging the situation. For example, it may load a RAMDISK, leaving just 1M of free RAM, and then trying to load a 2M application. So if you have 16M of RAM, booting with mem=14M may actually help, as the "load RAMDISK" stage would then fail and the install would then know to run off the CD instead of off the RAMDISK. (installs used to work for >8M machines. Is that still true?)
Try, in one session to clear the disk of all the partitions that are going to be used by Linux. Reboot. Then try the install. Either by partitioning manually, or by letting the install program figure it out. (I take it that Red Hat has that possibility too, SuSE has it...) If this works for you, I'd appreciate it if you'd tell me.
A corrupted download can also cause this. Duh.
Someone reports that installs on 8Mb machines no longer work, and that the install ungracefully exits with a sig7. -- Chris Rocco (crocco@earthlink.net)
One person reports that disabling "BIOS shadow" (system & VIDEO), helped for him. As Linux doesn't use the BIOS, shadowing it doesn't help. Some computers may even give you 384k of extra RAM if you disable the shadowing. Just disable it, and see what happens. -- Philippe d'Offay (pdoffay@pmdsoft.com).

QUESTION

What are other possibilities?

ANSWER

Others have noted the following possibilities:

The compiler and libc included in Red Hat 5.0 have an odd interaction with the Cyrix processor. It crashes the compiler, This is VERY odd. I would think that the only way that this can be the case is when the Cyrix has a bug that has gone undetected all this time, and reliably gets triggered when THAT gcc compiles the Linux kernel. Anyway, if you just want compile a kernel, you should get a new compiler and/or libc from the Red Hat website. (start at the homepage, and click errata).
Compiling a 2.0.x kernel with a 2.8.x gcc or any egcs doesn't work. There are a few bugs in the kernel that don't show up because gcc 2.7.x does a lousy job optimizing it. gcc 2.8.x and egcs just dump some of the code because we didn't tell it not to. Anyway, you usually get a kernel that seems to work but has funny bugs. For example X may crash with a signal 11. Oh, and before you ask, no it's not going to be fixed. Don't bother Alan or Linus about this OK? -- Hans Peter Verne (h.p.verne@kjemi.uio.no)
The pentium-optimizing-gcc (the one with the version number ending in "p") fails with the default options on certain source files like floppy.c in the kernel. The "triggers" are in the kernel, libc and in gcc itself. This is easily diagnosed as "not a hardware problem" because it always happens in the same place. You can either disable some optimizations (try -fno-unroll-loops first) or use another gcc. -- Evan Cheng (evan@top.cis.syr.edu) (In other words: gcc 2.7.2p crashes with sig11 on floppy.c . Workaround-1: Use plain gcc. Workaround-2: Manually compile floppy.c with "-O" instead of "-O2". )
A bad connection between a disk and the system. For example IDE cables are only allowed to be 40cm (16") long. Many systems come with longer cables. Also a removable IDE rack may add enough trouble to crash a system.
A badly misconfigured gcc -- some parts from one version, some from another. After a few weeks I ended up re-installing from scratch to get everything right. -- Richard H. Derr III (rhd@Mars.mcs.com).
Gcc or the resulting application may terminate with sig11 when a program is linked against the SCO libraries (which come with iBCS). This occurs on some applications that have -L/lib in their LDFLAGS....
When compiling a kernel with an ELF compiler, but configured for a.out (or the other way around, I forgot) you will get a signal 11 on the first call to "ld". This is easily identified as a software problem, as it always occurs on the FIRST call to "ld" during the build. -- REW
An Ethernet card together with a badly configured PCI BIOS. If your (ISA) Ethernet card has an aperture on the ISA bus, you might need to configure it somewhere in the BIOS setup screens. Otherwise the hardware would look on the PCI bus for the shared memory area. As the ISA card can't react to the requests on the PCI bus, you are reading empty "air". This can result in segmentation faults and kernel crashes. -- REW
Corrupted swap partition. Tony Nugent (T.Nugent@sct.gu.edu.au) reports he used to have this problem and solved it by an mkswap on his swap partition. (Don't forget to type "sync" before doing anything else after an mkswap. -- Louis J. LaBash Jr. (lou@minuet.siue.edu))
NE2000 card. Some cheap Ne2000 cards might mess up the system. -- Danny ter Haar (dth@cistron.nl) I personally might have had similar problems, as my mail server crashed hard every now and then (once a day). It now seems that 1.2.13 and lots of the 1.3.x kernels have this bug. I haven't seen it in 1.3.48. Probably got fixed somewhere in the meantime.... -- REW
Power supply? No I don't think so. A modern heavy system with two or three harddisk, both SCSI and IDE will not exceed 120 Watts or so. If you have loads of old harddisks and old expansion cards the power requirements will be higher, but still it is very hard to reach the limits of the power supply. Of course some people manage to find loads of old full-size harddisks and install them into their big-tower. You can indeed overload a powersupply that way. -- Greg Nicholson (greg@job.cba.ua.edu) A faulty power supply CAN of course deliver marginal power, which causes all of the malfunctioning that you read about in this file.... -- Thorsten Kuehnemann (thorsten@actis.de)
An inconsistent ext2fs. Some circumstances can cause the kernel code of the ext2 file system to result in Signal 11 for Gcc. -- Morten Welinder (terra@diku.dk)
CMOS battery. Even if you set the BIOS as you want it, it could be changing back to "bad" settings under your nose if the CMOS battery is bad. -- Heonmin Lim (coco@me.umn.edu)
No or too little swap space. Gcc doesn't gracefully handle the "out of memory" condition. -- Paul Brannan (brannanp@musc.edu)
Incompatible libraries. When you have a symlink from "libc.so.5" pointing to "libc.so.6", some applications will bomb with sig11. -- Piete Brooks (piete.brooks@cl.cam.ac.uk).
Broken mouse. Somehow, a mouse seems to be able to break in a way that it causes some (mouse related) programs to crash with Sig11. I've seen it happen on an X server that would crash if you moved the mouse quickly. Matthew might not even have been moving his mouse. -- REW & Matthew Duggan (stauff@guarana.org).
Badly seated RAM. Make sure your RAM is correctly seated into the socket.... -- Carroll Kong (me@carrollkong.com).

QUESTION

I found that running ..... detects errors much quicker than just compiling kernels. Please mention this on your site.

ANSWER

Many people email me with notes like this. However, what many don't realize is that they encountered ONE case of problematic hardware. The person recommending "unzip -t" happened to have a certain broken DRAM stick. And unzip happened to "find" that much quicker than a kernel compile.

However, I'm sure that for many other problems, the kernel compile WOULD find it, while other tests don't. I think that the kernel compile is good because it stresses lots of different parts of the computer. Many other tests just exercise just one area. If that area happens to be broken in your case, it will show a problem much quicker than "kernel compile" will. But if your computer is OK on that area and broken in another, the "faster" test may just tell you your computer is OK, while the kernel compile test would have told you something was wrong.

In any case, I might just as well list what people think are good tests, which they are, but not as general as the "try and compile a kernel" test....

Run unzip while compiling kernels. Use a zipfile about as large as RAM.
use "memtest86" found at: http://www.memtest86.com/.
do dd if=/dev/hda of=/dev/null while compiling kernels.
run md5sum on large trees.

Note that whatever fast method you may find to tell you that your computer is broken, it won't guarantee your computer is fine if such a test suddenly doesn't fail anymore. I always recommend that after fiddling with things to make it work, you should run a 24-hour kernel-compile test.

QUESTION

Why isn't "memtest86" the first to try if I suspect memory problems?

ANSWER

Feel free to do so. Some of this is black magic. However, when "memtest86" tells you that your RAM is ok, you might be tempted to believe it. It's telling you that it couldn't find any problems. It's not telling you that your RAM is flawless.

In my experience, RAM related problems are sometimes not found using a memory tester. The patterns are all nice and regular. Some problematic RAM simply works well under that kind of stress, but fails under the more erratic stress patterns caused by "gcc" or "zip".

So, I still recommend that you try verifying your system using kernel compiles, and not trusting a memory tester....

QUESTION

I don't believe this. To whom has this happened?

ANSWER

Well for one it happened to me personally. But you don't have to believe me. It also happened to:

Johnny Stephens (icjps@asuvm.inre.asu.edu)
Dejan Ilic (d92dejil@und.ida.liu.se)
Rick Tessner (rick@myra.com)
David Fox (fox@graphics.cs.nyu.edu)
Darren White (dwhite@baker.cnw.com) (L2 cache)
Patrick J. Volkerding (volkerdi@mhd1.moorhead.msus.edu)
Jeff Coy Jr. (jcoy@gray.cscwc.pima.edu) (Temp problems)
Michael Blandford (mikey@azalea.lanl.gov) (Temp problems: CPU fan failed)
Alex Butcher (Alex.Butcher@bristol.ac.uk) (Memory waitstates)
Richard Postgate (postgate@cafe.net) (VLB loading)
Bert Meijs (L.Meijs@et.tudelft.nl) (bad SIMMs)
J. Van Stonecypher (scypher@cs.fsu.edu)
Mark Kettner (kettner@cat.et.tudelft.nl) (bad SIMMs)
Naresh Sharma (n.sharma@is.twi.tudelft.nl) (30->72 converter)
Rick Lim (ricklim@freenet.vancouver.bc.ca) (Bad cache)
Scott Brumbaugh (scottb@borris.beachnet.com)
Paul Gortmaker (paul.gortmaker@anu.edu.au)
Mike Tayter (tayter@ncats.newaygo.mi.us) (Something with the cache)
Benni ??? (benni@informatik.uni-frankfurt.de) (VLB Overloading)
Oliver Schoett (os@sdm.de) (Cache jumper)
Morten Welinder (terra@diku.dk)
Warwick Harvey (warwick@cs.mu.oz.au) (bit error in cache)
Hank Barta (hank@pswin.chi.il.us)
Jeffrey J. Radice (jjr@zilker.net) (Ram voltage)
Samuel Ramac (sramac@vnet.ibm.com) (CPU tops out)
Andrew Eskilsson (mpt95aes@pt.hk-r.se) (DRAM speed)
W. Paul Mills (wpmills@midusa.net) (CPU fan disconnected from CPU)
Joseph Barone (barone@mntr02.psf.ge.com) (Bad cache)
Philippe Troin (ptroin@compass-da.com) (delayed RAM timing trouble)
Koen D'Hondt (koen@dutlhs1.lr.tudelft.nl) (more kernel error messages)
Bill Faust (faust@pobox.com) (cache problem)
Tim Middlekoop (mtim@lab.housing.fsu.edu) (CPU temp: fan installed)
Andrew R. Cook (andy@anchtk.chm.anl.gov) (bad cache)
Allan Wind (wind@imada.ou.dk) (P66 overheating)
Michael Tuschik (mt2@irz.inf.tu-dresden.de) (gcc2.7.2p victim)
R.C.H. Li (chli@en.polyu.edu.hk) (Overclocking: ok for months...)
Florin (florin@monet.telebyte.nl) (Overclocked CPU by vendor)
Dale J March (dmarch@pcocd2.intel.com) (CPU overheating on laptop)
Markus Schulte (markus@dom.de) (Bad RAM)
Mark Davis (mark_d_davis@usa.pipeline.com) (Bad P120?)
Josep Lladonosa i Capell (jllado@arrakis.es) (PCI options overoptimization)
Emilio Federici (mc9995@mclink.it) (P120 overheating)
Conor McCarthy (conormc@cclana.ucd.ie) (Bad SIMM)
Matthias Petofalvi (mpetofal@ulb.ac.be) ("Simmverter" problem)
Jonathan Christopher Mckinney (jono@tamu.edu) (gcc2.7.2p victim)
Greg Nicholson (greg@job.cba.ua.edu) (many old disks)
Ismo Peltonen (iap@bigbang.hut.fi) (irq_unmasking)
Daniel Pancamo (pancamo@infocom.net) (70ns instead of 60 ns RAM)
David Halls (david.halls@cl.cam.ac.uk)
Mark Zusman (marklz@pointer.israel.net) (Bad motherboard)
Elizabeth Ayer (eca23@cam.ac.uk) (Power management features)
Thorsten Kuehnemann (thorsten@actis.de)
(Email me with your story, you might get to be mentioned here... :-) ---- Update: I like to hear what happened to you. This will allow me to guess what happens most, and keep this file as accurate as possible. However I now have around 500 different Email addresses of people who've had sig-11 problems. I don't think that it is useful to keep on adding "random" people's names on this list. What do YOU think?

I'm interested in new stories. If you have a problem and are unsure about what it is, it may help to mailto:R.E.Wolff@BitWizard.nl. My curiosity will usually drive me to answering your questions until you find what the problem is..... (on the other hand, I do get pissed when your problem is clearly described above :-)

This page is hosted by http://www.bitwizard.nl/

This text is (C) 1994-2003 R.E.Wolff@BitWizard.nl. You are permitted to copy and distribute this text, provided that you leave the attributions, this copyright message and the reference to the original (and up-to-date) location intact. If you're in the kingdom of the Netherlands, you're permitted to print or copy this document, provided you pay the standard fee of EUR 0.045 per page through the standard channels.