Hi Kevin, Il giorno dom, 03/02/2008 alle 18.10 +0100, Kevin D. Kissell ha scritto: > Giuseppe Sacco wrote: [...] > > Thanks for your reply. I will try to understand how to use gdb on this > > context. (Any URI would be really appreciated.) > > Anyway I now understood that a dbe is a data bus error, so probably this > > is an error on the physical address, i.e. a kernel problem related to > > the mapping between vertical and physical addresses. Is this correct? > > > That's correct. You didn't say what processor you were running on, so > it's hard to be more specific - there are some which have a bus error > input pin that can be asserted by the system for other reasons - but > in general it means that there's a data reference at 0x2ac2bffc whose > valid translation goes to a bad address. Generally, that address > range is where shared libraries are mapped, so to find the instruction > you want to run the program that caused the crash under gdb, set a > breakpoint very early (e.g. main), run to the breakpoint, and > disassemble the virtual address. I find it interesting that the > register value reported for register $10 is a reasonable data address > shifted up by 32 bits. It's possible that code would have a real > reason to do that, but I can't help wonder if that isn't part of the > problem. We may be looking at a 2-level bug here: User(?) code > screwing up a base register used for a load or store, and the OS > failing to handle the upper reaches of the 64-bit address space > correctly. The complete bug report is available at http://bugs.debian.org/463808. The cpu is an "R5000 V2.1 FPU V1.0". The system is Debian stable, running mainly with courier-imap-ssl and exim4 (often in TLS mode). I cannot find a single program to debug, but I know for sure that if I leave the machine with those two daemons, it will hung in about 30 minutes. If I run a kernel build (using gcc-4.2 from dDebian testing), then the machine hungs in a few minutes. One time out of three gcc get a segmentation fault, other two times the machine stop. Thanks for your help, Giuseppe
