Re: [v2 PATCH 2/2] booke/kprobe: remove unnecessary preempt_enable_no_resched

Ananth N Mavinakayanahalli wrote:

On Mon, Jul 11, 2011 at 10:39:35AM +0800, Tiejun Chen wrote:

quoted

When enable CONFIG_PREEMPT we will trigger the following call trace:

BUG: scheduling while atomic: swapper/1/0x10000000
...

krpobe always goes through the following path:

program_check_exception()
        |
        + notify_die(DIE_BPT, "breakpoint",...)
                |
                + kprobe_handler()
                        |
                        + preempt_disable();
                        + break_handler() <- preempt_enable_no_resched()
                        + emulate_step()
                        + preempt_enable_no_resched()
                        ...
        exit

We should remove unnecessary preempt_enable_no_resched() inside of break_handler()
since looks longjmp_break_handler() always go the above path.

The current code is correct. Reasoning follows...

setjmp_pre_handler() and longjmp_break_handler() are used only for
jprobes. In the case of a jprobe, the code flow would be:

bp hit -> kprobe_handler() -> preempt_disable() -> setjmp_pre_handler()
(not that since this routine returns 1, we skip sstep here) -> jp->entry()
-> jprobe_return() -> bp hit -> kprobe_handler() -> preempt_disable() again
-> longjmp_break_handler() -> preempt_enable() -> sstep -> preempt_enable()
(for the second kprobe_handler() entry).

You could verify this with a preempt_count() printk with a
CONFIG_PREEMPT=y kernel.

quoted

Signed-off-by: Tiejun Chen <redacted>

Nack, sorry :-)

You're right.

When use EXC_LEVEL_EXCEPTION_PROLOG for Critical/Machine check, if the exception
came from kernel mode, we copy thread_info flags, *preempt*, and task pointer
from the process thread_info. So here I steal EXC_LEVEL_EXCEPTION_PROLOG for
Program Exception, preempt count would be corrupted incorrectly.

Thanks
Tiejun

Ananth