2.6.21-rc5-mm4: ia64: scheduling while atomic - utrace?
Running a 'usex -e' load [http://people.redhat.com/~anderson/usex/] on 2.6.21-rc5-mm4 on ia64, I see the following:
BUG: scheduling while atomic: strace/0x40000001/20162
Call Trace:
[<a000000100014ec0>] show_stack+0x80/0xa0
sp=e000076042dc7610 bsp=e000076042dc1260
[<a000000100014f10>] dump_stack+0x30/0x60
sp=e000076042dc77e0 bsp=e000076042dc1248
[<a0000001006f76e0>] schedule+0x1d00/0x22a0
sp=e000076042dc77e0 bsp=e000076042dc1108
[<a000000100099750>] __cond_resched+0x50/0xa0
sp=e000076042dc7800 bsp=e000076042dc10e8
[<a0000001006f8e30>] cond_resched+0xb0/0xe0
sp=e000076042dc7800 bsp=e000076042dc10d0
[<a0000001001561d0>] get_user_pages+0x1b0/0x7c0
sp=e000076042dc7800 bsp=e000076042dc1028
[<a0000001001568a0>] access_process_vm+0xc0/0x440
sp=e000076042dc7820 bsp=e000076042dc0f78
[<a00000010002fcc0>] ia64_sync_user_rbs+0x80/0x100
sp=e000076042dc7830 bsp=e000076042dc0f38
[<a00000010002fdf0>] do_gpregs_writeback+0xb0/0xe0
sp=e000076042dc7840 bsp=e000076042dc0f10
[<a00000010000cad0>] unw_init_running+0x70/0xa0
sp=e000076042dc7850 bsp=e000076042dc0ee8
[<a00000010002ed70>] do_regset_call+0x110/0x140
sp=e000076042dc7c30 bsp=e000076042dc0e88
[<a00000010002eea0>] gpregs_writeback+0x40/0x60
sp=e000076042dc7e30 bsp=e000076042dc0e60
[<a000000100123900>] ptrace_report+0xe0/0x1e0
sp=e000076042dc7e30 bsp=e000076042dc0e28
[<a000000100123aa0>] ptrace_report_syscall+0xa0/0xe0
sp=e000076042dc7e30 bsp=e000076042dc0e00
[<a000000100123b10>] ptrace_report_syscall_exit+0x30/0x60
sp=e000076042dc7e30 bsp=e000076042dc0dc8
[<a000000100122cb0>] utrace_report_syscall+0xf0/0x540
sp=e000076042dc7e30 bsp=e000076042dc0d48
[<a000000100031800>] syscall_trace_leave+0x60/0xc0
sp=e000076042dc7e30 bsp=e000076042dc0cf0
[<a00000010000c1c0>] ia64_trace_syscall+0x100/0x110
sp=e000076042dc7e30 bsp=e000076042dc0cf0Looks like get_ptrace_state(), called from ptrace_report_syscall calls rcu_read_lock() which disables preemption. Corresponding rcu_read_unlock() will be from put_ptrace_state() from ptrace_report() at end of report. However, ia64 needs to sync register backing store, and this requires access to process vm. get_user_pages' use of cond_sched() is tripping the "scheduling while atomic" bug.
May be related to:
http://marc.info/?a=102883379600003&r=1&w=4
Lee
Thanks for the report. I introduced this bug recently when I changed around some of the locking but forgot about the writeback issue. I don't think this is directly related to any other crash you might have seen.
I've moved the call out of the lock-holding region, where it didn't need to be. I'm updating my patch series now; I've appended the incremental patch.
Thanks,
Roland
--- kernel/ptrace.c | 20 ++++++++++---------- 1 files changed, 10 insertions(+), 10 deletions(-)
kernel/ptrace.c +10 -10
@@ -1473,16 +1473,6 @@ ptrace_report(struct utrace_attached_eng*/ utrace_set_flags (tsk, engine, engine->flags |UTRACE_ACTION_QUIESCE ); -/* -* If regset 0 has a writeback call, do it now. On register window -* machines, this makes sure the user memory backing the register -* data is up to date by the time wait_task_inactive returns to -* ptrace_start in our tracer doing a PTRACE_PEEKDATA or the like. -*/ - regset =utrace_regset (tsk, engine,utrace_native_view (tsk),0 ); -if (regset->writeback) - (*regset->writeback)(tsk, regset,0 ); -BUG_ON (code ==0 ); tsk->exit_code = code;do_notify (tsk, state->parent,CLD_TRAPPED );
@@ -1494,6 +1484,16 @@ ptrace_report(struct utrace_attached_engNO_LOCKS ; +/* +* If regset 0 has a writeback call, do it now. On register window +* machines, this makes sure the user memory backing the register +* data is up to date by the time wait_task_inactive returns to +* ptrace_start in our tracer doing a PTRACE_PEEKDATA or the like. +*/ + regset =utrace_regset (tsk, engine,utrace_native_view (tsk),0 ); +if (regset->writeback) + (*regset->writeback)(tsk, regset,0 ); +return UTRACE_ACTION_RESUME ; }