Sorry, I thought I solved the problem. Using `kzalloc` doesn't work all the time, I need to add `sleep(1)` in the test case after each syscall, like perf_event_open(...); sleep(1); ioctl(...); sleep(1); ioctl(...); sleep(1); read(...); I have tried these: 1) with `sleep(1)`, both kprobe and kretprobe are enabled 2) without `sleep(1)`, both kprobe and kretprobe are enabled 3) without `sleep(1)`, disable pre_handler of kprobe 4) same as 1), after the first run, comment the `sleep(1)` lines, and run the test again and 1) 3) 4) look fine, On 12/21/2016 07:35 PM, zerons wrote:
Hi everyone.
I wrote a kernel module to test something. The module uses kprobe and kretprobe, here is a bug I met today.
The pre_handler of kprobe, calls `do_something`. The probed instructions are in the middle of a function. The entry_handler of kretprobe, also calls `do_something`. `do_something` calls `kmalloc`+`memset`.
Back to userspace, when I have all the functions probed, then the test program cause a high CPU usage, and the keyboard doesn't work. The system does not panic when I set softlockup_panic=1.
If `do_something` is called by entry_handler of kretprobe, the module works fine. The bug happens when `do_something` called by the pre_handler of kprobe.
So I use "#if 0" to locate the bug. It turns out to be `kmalloc`+`memset`. When I change that to `kzalloc`, problem solved.
Then I get confused. `kzalloc` just calls `kmalloc` with a `__GFP_ZERO`. Why the bug only happens when pre_handler of kprobe gets called?
Is it necessary to post the source code here? Thanks.