| Message ID | 878rrrh32q.fsf_-_@email.froward.int.ebiederm.org (mailing list archive) |
|---|---|
| Headers | show |
| Series | ptrace: cleaning up ptrace_stop | expand |
On Tue, Apr 26, 2022 at 05:50:21PM -0500, Eric W. Biederman wrote: > .... Peter Zijlstra has > been rewriting the classic freezer and in earlier parts of this > discussion so I presume it is also a problem for PREEMPT_RT. Ah, the freezer thing is in fact a sched/arm64 issue, the common issue between these two issues is ptrace though. Specifically, on recent arm64 chips only a subset of CPUs can execute arm32 code and 32bit processes are restricted to that subset. If by some mishap you try and execute a 32bit task on a non-capable CPU it gets terminated without prejudice. Now, the current freezer has this problem that tasks can spuriously thaw too soon (where too soon is before SMP is restored) which leads to these 32bit tasks being killed dead. That, and it was a good excuse to fix up the current freezer :-)
While looking at how ptrace is broken on PREEMPT_RT I realized that ptrace_stop would be much simpler and more maintainable if tsk->ptrace, tsk->parent, and tsk->real_parent were protected by siglock. Most of the changes are general cleanups in support of this locking change. While making the necessary changes to protect tsk->ptrace with siglock I discovered we have two architectures xtensa and um that were using tsk->ptrace for what most other architectures use TIF_SIGPENDING for and not protecting tsk->ptrace with any lock. By the end of this series ptrace should work on PREEMPT_RT with the CONFIG_FREEZER and CONFIG_CGROUPS disabled, by the simple fact that the ptrace_stop code becomes less special. The function cgroup_enter_frozen because it takes a lock which is a sleeping lock on PREEMPT_RT with preemption disabled definitely remains a problem. Peter Zijlstra has been rewriting the classic freezer and in earlier parts of this discussion so I presume it is also a problem for PREEMPT_RT. Peter's series rewriting the freezer[1] should work on top of this series with minimal changes and patch 2/5 removed. Eric W. Biederman (9): signal: Rename send_signal send_signal_locked signal: Replace __group_send_sig_info with send_signal_locked ptrace/um: Replace PT_DTRACE with TIF_SINGLESTEP ptrace/xtensa: Replace PT_SINGLESTEP with TIF_SINGLESTEP signal: Protect parent child relationships by childs siglock signal: Always call do_notify_parent_cldstop with siglock held ptrace: Simplify the wait_task_inactive call in ptrace_check_attach ptrace: Use siglock instead of tasklist_lock in ptrace_check_attach ptrace: Don't change __state arch/um/include/asm/thread_info.h | 2 + arch/um/kernel/exec.c | 2 +- arch/um/kernel/process.c | 2 +- arch/um/kernel/ptrace.c | 8 +- arch/um/kernel/signal.c | 4 +- arch/xtensa/kernel/ptrace.c | 4 +- arch/xtensa/kernel/signal.c | 4 +- drivers/tty/tty_jobctrl.c | 4 +- include/linux/ptrace.h | 7 -- include/linux/sched/jobctl.h | 2 + include/linux/sched/signal.h | 3 +- include/linux/signal.h | 3 +- kernel/exit.c | 4 + kernel/fork.c | 12 +-- kernel/ptrace.c | 61 ++++++------- kernel/signal.c | 187 ++++++++++++++------------------------ kernel/time/posix-cpu-timers.c | 6 +- 17 files changed, 131 insertions(+), 184 deletions(-) [1] https://lkml.kernel.org/r/20220421150248.667412396@infradead.org Eric