Interesting... my guess with perl then is that the Linux kernel is supposed to be initializing some value in the thread-local storage that we're not initializing. Unfortunately the only way to track that down is usually to go reading the kernel source... though if you find a spot where they define a base TLS struct then that should give it to you. Anyone else out there on the list have any experience with this?
As far as specrand it's impossible to say what the problem is without going backward further in the trace to see where r20 is coming from. If r20 also comes from reading something out of the TLS area then it could well be the same bug. Steve On 9/9/07, Elliott Cooper-Balis <[EMAIL PROTECTED]> wrote: > > hey steve, > i tried both of your suggestions, and the latter of which i think might > give a good clue as the memory address which causes the fault is not > referenced at any other point in the program. > > here is the result of grep'ing for the address in the execution trace : > > >grep 12022e50 exec.out > 5278458500: system.cpu0 T0 : @__printf_fp+128 : addq r0,r1,r0 > : IntAlu : D=0x000000012022e508 > 5278459000: system.cpu0 T0 : @__printf_fp+132 : ldq r1,0(r0) > : MemRead : D=0x0000000000000000 A=0x12022e508 > > which are the 2 instructions right before the fault and the only 2 > instances of it being referenced. > > i tried digging around a little more to see if this address in particular > was causing the problems. unfortunately, that doesn't appear to be the > case. the benchmark we have been discussing is the Perl benchmark in > SPEC06. i ran the random number generator benchmark as well (999.specrand) > and here is the execution output just before its page fault : > > [EMAIL PROTECTED]:~/Development/M5/m5-2.0b3/build/ALPHA_SE$ ./m5.debug > --trace-flags=Exec,Syscall,SyscallVerbose --trace-start=2000000 > ../../configs/example/se.py -c > benchmarks/999.specrand/exe/specrand_base.amd64-m64-gcc41-nn -o "4 3943" > > .... > > 2183000: system.cpu0 T0 : @____strtoll_l_internal+52 : bis > r31,r18,r10 : IntAlu : D=0x000000000000000a > 2183500: system.cpu0 T0 : @____strtoll_l_internal+56 : bis > r31,r20,r11 : IntAlu : D=0x0000000000000000 > 2184000: system.cpu0 T0 : @____strtoll_l_internal+60 : ldq > r3,8(r20) : MemRead : A=0x8 > panic: Page table fault when accessing virtual address 0x8 > @ cycle 2184000 > [invoke:build/ALPHA_SE/sim/faults.cc, line 65] > Program aborted at cycle 2184000 > Aborted (core dumped) > > unfortunately, there doesn't appear to be (at least to me) any > similarities between the two benchmark's output. > > > elliott > > *Steve Reinhardt <[EMAIL PROTECTED]>* wrote: > > It's not obvious, but it does give some clues... > > The null pointer is being read from memory address 0x12022e508, so either > that's a bogus address or the memory location doesn't have the right value > (not getting initialized or getting clobbered at some point). > > The pointer address is computed by adding the uniq register (put into R0 > by "call_pal rduniq") and some value (0x28) read from -29160(r29)... I think > that's the global constant pool. The uniq reg is used as a pointer to > thread-local storage. So basically it's reading the null value out of > thread-local storage. It could be that that's a value that the OS is > supposed to provide but we're not initializing it properly. > > I'd do two more things to try and get some more clues: > > - run with just --trace-flags=Syscall (and no --trace-start) to get a > complete syscall trace, then look at whatever the last few syscalls are, and > see what they are and how closely they precede the crash > - run with just --trace-flags=Exec (and no --trace-start) and then pipe > the trace through "egrep -i '12022e50[0-7]' " to look at all the other > references to that memory location... is it ever written, if it's read > before is it always zero, etc. This will take a while... > > Steve > > On 9/7/07, Elliott Cooper-Balis <[EMAIL PROTECTED]> wrote: > > > > here is the output. is there anything obvious that might be broken? > > > > _______________________________________________ > m5-users mailing list > m5-users@m5sim.org > http://m5sim.org/cgi-bin/mailman/listinfo/m5-users > > > ------------------------------ > Yahoo! oneSearch: Finally, mobile search that gives > answers<http://us.rd.yahoo.com/evt=48252/*http://mobile.yahoo.com/mobileweb/onesearch?refer=1ONXIC>, > not web links. > > > _______________________________________________ > m5-users mailing list > m5-users@m5sim.org > http://m5sim.org/cgi-bin/mailman/listinfo/m5-users >
_______________________________________________ m5-users mailing list m5-users@m5sim.org http://m5sim.org/cgi-bin/mailman/listinfo/m5-users
