kdeinit causes scheduling while atomic
A scheduling while atomic just started showing up in the kernel log on my machine, which unfortunately is running an odd combination of non-vanilla patches. Maybe someone who knows at least some of the code can give me some hints tracking this down? Or maybe the grsecurity folks have an idea?
Patches applied on top of 2.6.17.6:
grsecurity-2.1.9-2.6.17.4-200607120947 suspend2-2.2.7-for-2.6.17 (without 9920-linus-basic-trace - that plus grsecurity gives rejects on vmlinux.lds.S)
System is running Gentoo, the compiler is gcc 3.4.6 with pie & ssp although those features are turned off when compiling a kernel by the specs file. Userland is compiled with -fstack-protector, and I am using SEGMEXEC for userland NX support.
I have two different traces and some PaX output, here they all are:
---cut here---
BUG: scheduling while atomic: kdeinit/0x00000002/31816 <c0345c9c> schedule+0x53e/0x674 <c0346657> schedule_timeout+0x4d/0x9a <c0152a12> process_timeout+0x0/0x5 <c0145954> sched_fork+0x51/0x86 <c01483aa> copy_process+0x614/0xdea <c0148c66> do_fork+0x75/0x1b2 <c019864d> vfs_write+0xec/0x16e <c012dbf5> sys_clone+0x32/0x36 <c012f057> syscall_call+0x7/0xb <c012007b> init_script_binfmt+0x6/0xa
---end cut----
---cut here---
BUG: scheduling while atomic: kdeinit/0x00000002/31816 <c0345c9c> schedule+0x53e/0x674 <c0346657> schedule_timeout+0x4d/0x9a <c0152a12> process_timeout+0x0/0x5 <c0145954> sched_fork+0x51/0x86 <c01483aa> copy_process+0x614/0xdea <c0148c66> do_fork+0x75/0x1b2 <c012dbf5> sys_clone+0x32/0x36 <c012f057> syscall_call+0x7/0xb <c013007b> irq_entries_start+0xecb/0xef0
---end cut----
---cut here---
PAX: execution attempt in: <NULL>, 00000000-00000000 00000000
PAX: terminating task: /usr/kde/3.5/bin/konqueror(konqueror):21177, uid/euid: 1000/1000, PC: 00000010, SP: 5953cf80
PAX: bytes at PC: ?? ?? ?? ?? ?? ?? ?? ?? ?? ?? ?? ?? ?? ?? ?? ?? ?? ?? ?? ??
PAX: bytes at SP-4: 00000010 00000003 5953cfb0 00000000 0000000a 0000001c 40488ef8 0000001d 5618e6e0 40083d30 00000000 00000003 55efdb83 40189c00 56108d92 56b09377 5618e6e0 5953d000 00000006 404eadb8 55f2bef8 ---end cut----
Call of a null function pointer?
Sometimes konqueror is killed by PaX, sometimes it dies on its own (I think with a segfault). I can reproduce this 100% of the time by firing up Konqueror and going to www.wikipedia.org, just before the front page loads the window dissapears and leave those traces behind. Other sites seem to work fine. The process either segfaults or is killed by PaX immediately after causing the scheduling while atomic.
In the terminal where I start konqueror I see this (when PaX does the killing):
---cut here---
Try to load libthai dynamically...
Error, can't load libthai...
zsh: killed konqueror
---end cut----
... which seems to be an old unsolved bug in qt.
Do I or anybody else have any hope of tracking this down? It almost sounds like an interaction of multiple bugs.
A message that is not in the archive (sent off-list or not yet received): [email protected]
Quoted 7 lines · On 14:08 Sun 23 Jul , [email protected] wrote:
Quoted 5 lines · Will Simoneau
A scheduling while atomic just started showing up in the kernel log on my machine, which unfortunately is running an odd combination of non-vanilla patches. Maybe someone who knows at least some of the code can give me some hints tracking this down? Or maybe the grsecurity folks have an idea?
Quoted 5 lines
given that probably neither side is familiar with the other's patches, you should try to reproduce this with only one patch first. in the meantime, it'd help to find out what gets called at
copy_process+0x614/0xdea, there's probably a 'call' insn around copy_process+60f, you can use objdump to find out who the target is.
objdump -d:
c014838a: be 00 f0 ff ff mov $0xfffff000,%esi c014838f: 8b bb a0 00 00 00 mov 0xa0(%ebx),%edi c0148395: 21 e6 and %esp,%esi c0148397: 8b 06 mov (%esi),%eax c0148399: 85 ff test %edi,%edi c014839b: 0f b7 40 2c movzwl 0x2c(%eax),%eax c014839f: 66 89 43 2c mov %ax,0x2c(%ebx) c01483a3: 74 51 je c01483f6 <copy_process+0x660> -> c01483a5: e8 a7 ab 0f 00 call c0242f51 <gr_handle_brute_check> c01483aa: 8b 83 ac 00 00 00 mov 0xac(%ebx),%eax c01483b0: 05 b0 00 00 00 add $0xb0,%eax c01483b5: 8b 4c 24 0c mov 0xc(%esp),%ecx
Quoted 17 lines
Quoted 5 lines
Patches applied on top of 2.6.17.6:
grsecurity-2.1.9-2.6.17.4-200607120947 suspend2-2.2.7-for-2.6.17 (without 9920-linus-basic-trace - that plus grsecurity gives rejects onvmlinux.lds.S)from what i see, it's a trivial reject, you can apply it by hand after RODATA.
Quoted 8 lines
---cut here---
PAX: execution attempt in: <NULL>, 00000000-00000000 00000000
PAX: terminating task: /usr/kde/3.5/bin/konqueror(konqueror):21177, uid/euid: 1000/1000, PC: 00000010, SP:5953cf80
PAX: bytes at PC: ?? ?? ?? ?? ?? ?? ?? ?? ?? ?? ?? ?? ?? ?? ?? ?? ?? ?? ?? ??
PAX: bytes at SP-4: 00000010 000000035953cfb0000000000000000a0000001c40488ef80000001d5618e6e040083d3000000000 0000000355efdb8340189c0056108d9256b093775618e6e05953d00000000006404eadb855f2bef8---end cut----Call of a null function pointer?
Quoted 9 lines
it's interesting that at esp-4 you have
0x10, which happens to be the faulting eip value as well. this can occur if the fault occured not due to a 'call' but rather a 'retn' insn, that is, a function was trying to return to its caller, but the return address got overwritten on the stack. now whether that happened due to some programming error or a gcc/ssp bug is a good question. i'd first try to recompile it (and all related libraries!) w/o ssp, it's known to have code generation bugs for c++ code. if that cures the problem, you should enter it into the gentoo bugzilla.
It's caused by qt, with -fstack-protector(-all?) enabled. Version is x11-libs/qt-3.3.6-r1 from portage. Without ssp, going to wikipedia in konqueror works fine. Will file at bugs.gentoo.org when I get a chance.
Quoted 8 lines
Quoted 7 lines
Sometimes konqueror is killed by PaX, sometimes it dies on its own (I think with a segfault). I can reproduce this 100% of the time by firing up Konqueror and going to www.wikipedia.org, just before the front page loads the window dissapears and leave those traces behind. Other sites seem to work fine. The process either segfaults or is killed by PaX immediately after causing the scheduling while atomic.
Quoted 5 lines
you mean, the first mentioned schedule BUG triggered in kdeinit is related to this crash in konqueror? or are you getting a schedule BUG in konqueror as well? in any case, eliminating one variable at a time (like ssp) should help you nail the bug(s) down.
I'm not sure what's happening with the scheduling while atomic now.