Re: [PATCH v2 2/2] riscv: Enable context tracking

[Palmer Dabbelt <palmer@dabbelt.com>]

On Wed, 24 Jun 2020 02:03:16 PDT (-0700), greentime.hu@sifive.com wrote:
> This patch implements and enables context tracking for riscv (which is a
> prerequisite for CONFIG_NO_HZ_FULL support)
>
> It adds checking for previous state in the entry that all excepttions and
> interrupts goes to and calls context_tracking_user_exit() if it comes from
> user space. It also calls context_tracking_user_enter() if it will return
> to user space before restore_all.
>
> This patch is tested with the dynticks-testing testcase in
> qemu-system-riscv64 virt machine and Unleashed board.
> git://git.kernel.org/pub/scm/linux/kernel/git/frederic/dynticks-testing.git
>
> We can see the log here. The tick got mostly stopped during the execution
> of the user loop.
>
>                         _-----=> irqs-off
>                        / _----=> need-resched
>                       | / _---=> hardirq/softirq
>                       || / _--=> preempt-depth
>                       ||| /     delay
>      TASK-PID   CPU#  ||||    TIMESTAMP  FUNCTION
>         | |       |   ||||       |         |
>    <idle>-0     [001] d..2   604.183512: sched_switch: prev_comm=swapper/1 prev_pid=0 prev_prio=120 prev_state=R ==> next_comm=taskset next_pid=273 next_prio=120
> user_loop-273   [001] d.h1   604.184788: hrtimer_expire_entry: hrtimer=000000002eda5fab function=tick_sched_timer now=604176096300
> user_loop-273   [001] d.s2   604.184897: workqueue_queue_work: work struct=00000000383402c2 function=vmstat_update workqueue=00000000f36d35d4 req_cpu=1 cpu=1
> user_loop-273   [001] dns2   604.185039: tick_stop: success=0 dependency=SCHED
> user_loop-273   [001] dn.1   604.185103: tick_stop: success=0 dependency=SCHED
> user_loop-273   [001] d..2   604.185154: sched_switch: prev_comm=taskset prev_pid=273 prev_prio=120 prev_state=R+ ==> next_comm=kworker/1:1 next_pid=46 next_prio=120
>     <...>-46    [001] ....   604.185194: workqueue_execute_start: work struct 00000000383402c2: function vmstat_update
>     <...>-46    [001] d..2   604.185266: sched_switch: prev_comm=kworker/1:1 prev_pid=46 prev_prio=120 prev_state=I ==> next_comm=taskset next_pid=273 next_prio=120
> user_loop-273   [001] d.h1   604.188812: hrtimer_expire_entry: hrtimer=000000002eda5fab function=tick_sched_timer now=604180133400
> user_loop-273   [001] d..1   604.189050: tick_stop: success=1 dependency=NONE
> user_loop-273   [001] d..2   614.251386: sched_switch: prev_comm=user_loop prev_pid=273 prev_prio=120 prev_state=X ==> next_comm=swapper/1 next_pid=0 next_prio=120
>    <idle>-0     [001] d..2   614.315391: sched_switch: prev_comm=swapper/1 prev_pid=0 prev_prio=120 prev_state=R ==> next_comm=taskset next_pid=276 next_prio=120
>
> Signed-off-by: Greentime Hu <greentime.hu@sifive.com>
> ---
>  arch/riscv/Kconfig        |  1 +
>  arch/riscv/kernel/entry.S | 23 +++++++++++++++++++++++
>  2 files changed, 24 insertions(+)
>
> diff --git a/arch/riscv/Kconfig b/arch/riscv/Kconfig
> index 128192e14ff2..17520e11815b 100644
> --- a/arch/riscv/Kconfig
> +++ b/arch/riscv/Kconfig
> @@ -52,6 +52,7 @@ config RISCV
>  	select HAVE_ARCH_SECCOMP_FILTER
>  	select HAVE_ARCH_TRACEHOOK
>  	select HAVE_ASM_MODVERSIONS
> +	select HAVE_CONTEXT_TRACKING
>  	select HAVE_COPY_THREAD_TLS
>  	select HAVE_DMA_CONTIGUOUS if MMU
>  	select HAVE_EBPF_JIT if MMU
> diff --git a/arch/riscv/kernel/entry.S b/arch/riscv/kernel/entry.S
> index cae7e6d4c7ef..6ed579fc1073 100644
> --- a/arch/riscv/kernel/entry.S
> +++ b/arch/riscv/kernel/entry.S
> @@ -97,6 +97,14 @@ _save_context:
>  	la gp, __global_pointer$
>  .option pop
>
> +#ifdef CONFIG_CONTEXT_TRACKING
> +	/* If previous state is in user mode, call context_tracking_user_exit. */
> +	andi a0, s1, SR_SPP

I've changed that to SR_PP, as I don't see any reason why this should depend on
MMU.

I think this is correct: we're using scratch==0 elsewhere to detect recursive
traps, but we've blown that away by this point so it's not an option.  I don't
know of any reason why PP wouldn't be accurate.

> +	bnez a0, skip_context_tracking
> +	call context_tracking_user_exit
> +
> +skip_context_tracking:
> +#endif
>  	la ra, ret_from_exception
>  	/*
>  	 * MSB of cause differentiates between
> @@ -137,6 +145,17 @@ _save_context:
>  	tail do_trap_unknown
>
>  handle_syscall:
> +#ifdef CONFIG_CONTEXT_TRACKING
> +	/* Recover a0 - a7 for system calls */
> +	REG_L x10, PT_A0(sp)
> +	REG_L x11, PT_A1(sp)
> +	REG_L x12, PT_A2(sp)
> +	REG_L x13, PT_A3(sp)
> +	REG_L x14, PT_A4(sp)
> +	REG_L x15, PT_A5(sp)
> +	REG_L x16, PT_A6(sp)
> +	REG_L x17, PT_A7(sp)
> +#endif
>  	 /* save the initial A0 value (needed in signal handlers) */
>  	REG_S a0, PT_ORIG_A0(sp)
>  	/*
> @@ -205,6 +224,10 @@ resume_userspace:
>  	andi s1, s0, _TIF_WORK_MASK
>  	bnez s1, work_pending
>
> +#ifdef CONFIG_CONTEXT_TRACKING
> +	call context_tracking_user_enter
> +#endif
> +
>  	/* Save unwound kernel stack pointer in thread_info */
>  	addi s0, sp, PT_SIZE_ON_STACK
>  	REG_S s0, TASK_TI_KERNEL_SP(tp)