• If a 'mount /dev/cdrom' command is issued without a CD in the drive, or if an audio CD is present
• At various random times when Netscape or other memory-intensive program is running (see below).

On one occasion, the problem was associated with swapping. During this time, the system load was rapidly and continuously increasing, and the disk light was on continuously, as Linux went into page swap death:

PID USER PRI NI SIZE RSS SHARE STAT CPU MEM TIME COMMAND 3669 tjnelson 25 0 0 0 0 RW 71.2 0.0 5:01 bash 4 root 23 0 0 0 0 RW 20.6 0.0 2:18 kswapd 493 root 15 0 656 600 560 S 6.3 0.1 0:00 cron

This situation recovers after about 5 minutes if you exit X and logout. Other times, the freeze was not accompanied by swapping, and it was not possible to exit X.

Symptoms

1. During the page swap crisis, /proc/meminfo shows that the problem was related to low memory:

   NORMAL DURING FREEZE ------------------------- ------------------------- MemTotal: 516280 kB MemTotal: 516280 kB MemFree: 439636 kB MemFree: 2996 kB MemShared: 0 kB MemShared: 0 kB Buffers: 11684 kB Buffers: 7372 kB Cached: 30536 kB Cached: 17536 kB SwapCached: 1688 kB SwapCached: 56700 kB Active: 26908 kB Active: 34716 kB Inactive: 32992 kB Inactive: 461008 kB HighTotal: 0 kB HighTotal: 0 kB HighFree: 0 kB HighFree: 0 kB LowTotal: 516280 kB LowTotal: 516280 kB LowFree: 439636 kB LowFree: 2996 kB SwapTotal: 2106776 kB SwapTotal: 3723248 kB SwapFree: 2097136 kB SwapFree: 2097136 kB

   About 2GB of disk space is available to the swap partition. Note also that the kernel gets confused about the total swap space.

   Vmstat 1 normally gives:

   procs memory swap io system cpu r b w swpd free buff cache si so bi bo in cs us sy id 0 0 0 0 414192 19024 41508 0 0 0 0 101 306 1 0 99

   During a hang, vmstat 1 gives:

   procs memory swap io system cpu r b w swpd free buff cache si so bi bo in cs us sy id 2 4 3 15020 4716 378332 40712 0 0 0 0 101 326 0 0 100

   This shows that nothing unusual, such as high system load, is occurring. The only thing unusual is that idle memory is almost depleted, which would signify that virtual memory is about to be used.

The following all had no effect:

1. Changing IRQs of the sound card and serial port
2. Disabling USB
3. Swapping out the keyboard
4. Disabling APIC and APM by adding

   append = "disableapic apm=off "

   to /etc/lilo.conf, rerunning /sbin/lilo, and rebooting,
5. Setting hard limits for all user processes by modifying /etc/security/limits.conf, in case the problem was due to resource depletion. (Note that only the 2.5 kernels appear to respect the limits.conf file; in earlier versions, the resource limits must be set by the shell in /etc/profile. This means users can change their own limits. This is a good reason to upgrade to 2.5, since it means anyone with an account can crash the system.)
6. Switching to other browsers instead of Netscape
7. Running fdisk to lower the size of the swap partition from 2GB to 1GB. This stopped the inconsistencies in SwapTotal reported by the kernel. It appears that, contrary to the documentation, Linux has a maximum of 1GB swap partition size on x86 Pentium 4 machines.
8. Installing more recent kernels. A 2.5.7 and 2.5.10 kernel both gave the error message "null pointer dereference" at boot-up followed by a kernel panic. A 2.5.12 kernel not only failed to boot, but obliterated the root partition (Reiserfs), forcing a reinstallation. Apparently, all kernels in the neighborhood of 2.5.12 and 2.5.13 have this problem and should be avoided like the plague.

   Of the 2.5 kernels, 2.5.8 seems to be the most stable. Although this kernel required editing of main.c before it would compile, the new kernel seems to have much better memory management than previous versions, but it still had the same problem of hanging at random times on Pentium 4 computers.

   To compile 2.5.8:

   1. in init/main.c, line 263 add:

      static inline void setup_per_cpu_areas(void) { }

      (Note: This function should be empty unless you are using SMP).
   2. In drivers/ide/Config.in, line 52: Put double quotes (") around $CONFIG_BLK_DEV_IDE_TLQ_DEFAULT.
   Unfortunately, this new kernel couldn't control the Creative Labs SB Live Ensoniq EMU10000 sound card.

Solution

1. Changed processor type from Pentium 4 to MK6. It was a bit odd that this helped, considering the prominent "Intel Inside" stickers on the outside of the computer and the fact that /proc/cpuinfo correctly identifies the CPU as a "Intel(R) Pentium(R) 4 CPU 1.80GHz". The L1 cache shift setting was automatically changed from 7 to 5.
2. Activated multi-mode disk access
3. Deactivated loopback device (/dev/loop)
4. Deactivated PPP compression

     # CONFIG_MPENTIUM4 is not set
     CONFIG_MK6=y
     CONFIG_X86_L1_CACHE_SHIFT=5
     CONFIG_X86_ALIGNMENT_16=y
     # CONFIG_BLK_DEV_LOOP is not set
     CONFIG_IDEDISK_MULTI_MODE=y
     CONFIG_IDEDISK_STROKE=y
     # CONFIG_PPP_DEFLATE is not set
     # CONFIG_PPP_BSDCOMP is not set
     # CONFIG_ZLIB_INFLATE is not set
     # CONFIG_ZLIB_DEFLATE is not set    

Working 2.5.5 kernel configuration for a Sony VAIO
Working 2.5.8 kernel configuration for a Dell Precision 330

Back