From mboxrd@z Thu Jan 1 00:00:00 1970 References: <87wnqnz2my.fsf@xenomai.org> <10a0-60d0b180-85-29fb8780@204393696> From: Philippe Gerum Subject: Re: Large gpio interrupt latency In-reply-to: <10a0-60d0b180-85-29fb8780@204393696> Date: Mon, 21 Jun 2021 18:38:07 +0200 Message-ID: <87sg1byxzk.fsf@xenomai.org> MIME-Version: 1.0 Content-Type: text/plain; charset=utf-8 Content-Transfer-Encoding: quoted-printable List-Id: Discussions about the Xenomai project List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , To: =?utf-8?Q?Fran=C3=A7ois?= Legal Cc: Jan Kiszka , "Chen, Hongzhan" , "xenomai@xenomai.org" Fran=C3=A7ois Legal writes: > Le Lundi, Juin 21, 2021 16:57 CEST, Philippe Gerum a = =C3=A9crit:=20 >=20=20 >>=20 >> Jan Kiszka writes: >>=20 >> > On 21.06.21 16:28, Philippe Gerum wrote: >> >>=20 >> >> Jan Kiszka writes: >> >>=20 >> >>> On 21.06.21 15:54, Philippe Gerum wrote: >> >>>> >> >>>> Jan Kiszka writes: >> >>>> >> >>>>> On 21.06.21 11:39, Philippe Gerum wrote: >> >>>>>> >> >>>>>> Jan Kiszka writes: >> >>>>>> >> >>>>>>> On 18.06.21 20:41, Fran=C3=A7ois Legal wrote: >> >>>>>>>> Le Mercredi, Juin 16, 2021 17:10 CEST, Jan Kiszka a =C3=A9crit:=20 >> >>>>>>>>=20=20 >> >>>>>>>>> On 16.06.21 15:51, Fran=C3=A7ois Legal wrote: >> >>>>>>>>>> Le Mercredi, Juin 16, 2021 15:38 CEST, Jan Kiszka a =C3=A9crit:=20 >> >>>>>>>>>>=20=20 >> >>>>>>>>>>> On 16.06.21 15:29, Fran=C3=A7ois Legal wrote: >> >>>>>>>>>>>> Le Mercredi, Juin 16, 2021 11:40 CEST, Jan Kiszka a =C3=A9crit:=20 >> >>>>>>>>>>>>=20=20 >> >>>>>>>>>>>>> On 16.06.21 11:12, Fran=C3=A7ois Legal via Xenomai wrote: >> >>>>>>>>>>>>>> Le Mercredi, Juin 16, 2021 11:05 CEST, "Chen, Hongzhan" <= hongzhan.chen@intel.com> a =C3=A9crit:=20 >> >>>>>>>>>>>>>>=20=20 >> >>>>>>>>>>>>>>> >> >>>>>>>>>>>>>>> >> >>>>>>>>>>>>>>>> -----Original Message----- >> >>>>>>>>>>>>>>>> From: Fran=C3=A7ois Legal =20 >> >>>>>>>>>>>>>>>> Sent: Wednesday, June 16, 2021 4:19 PM >> >>>>>>>>>>>>>>>> To: Chen, Hongzhan >> >>>>>>>>>>>>>>>> Cc: xenomai@xenomai.org >> >>>>>>>>>>>>>>>> Subject: RE: Large gpio interrupt latency >> >>>>>>>>>>>>>>>> >> >>>>>>>>>>>>>>>> Le Mercredi, Juin 16, 2021 10:10 CEST, "Chen, Hongzhan"= a =C3=A9crit:=20 >> >>>>>>>>>>>>>>>> >> >>>>>>>>>>>>>>>>> >> >>>>>>>>>>>>>>>>>> >> >>>>>>>>>>>>>>>>>> -----Original Message----- >> >>>>>>>>>>>>>>>>>> From: Xenomai On Behalf= Of Fran=C3=A7ois Legal via Xenomai >> >>>>>>>>>>>>>>>>>> Sent: Wednesday, June 16, 2021 3:16 PM >> >>>>>>>>>>>>>>>>>> To: xenomai@xenomai.org >> >>>>>>>>>>>>>>>>>> Subject: Large gpio interrupt latency >> >>>>>>>>>>>>>>>>>> >> >>>>>>>>>>>>>>>>>> Hello, >> >>>>>>>>>>>>>>>>>> >> >>>>>>>>>>>>>>>>>> working on a realtime data recorder (for which I subm= itted a patch to add timestamp retrieval for net packets), I experience a s= trange latency problem with taking GPIO interrupts. >> >>>>>>>>>>>>>>>>>> >> >>>>>>>>>>>>>>>>>> So basically my app monitors network packets (which a= s received by the DUT would trigger toggling of a GPIO on the DUT) and one = GPIO. >> >>>>>>>>>>>>>>>>>> We know (from some ohter reference instrument) that t= he DUT would toggle the GPIO 1-2ms after receiving a specific network packe= t. >> >>>>>>>>>>>>>>>>>> >> >>>>>>>>>>>>>>>>>> My recorder app relies on the timestamping of events = done in interrupts service routines for GPIOs and network interface. By che= cking the timestamps returned by the application, I get an about 50ms delay= between the network packet and the GPIO (so something between 51 and 52ms = delay). >> >>>>>>>>>>>>>>>>>> >> >>>>>>>>>>>>>>>>> >> >>>>>>>>>>>>>>>>> Are you toggling gpio and access gpio device through = your rtdm device like on path /dev/rtdm/your_gpiocontroller/gpio*? >> >>>>>>>>>>>>>>>>> >> >>>>>>>>>>>>>>>> >> >>>>>>>>>>>>>>>> Maybe my setup was not very clear. >> >>>>>>>>>>>>>>>> I've got a DUT that receives network data, and toggles = one of its GPIOs depending one the network data. >> >>>>>>>>>>>>>>>> >> >>>>>>>>>>>>>>>> I've got another device running my realtime recording a= pp, that receives the same network data as the DUT, and that has one of its= GPIO connected to the DUT GPIO. >> >>>>>>>>>>>>>>>> >> >>>>>>>>>>>>>>>> On the recording app, I use the RTDM device to open, io= ctl (enable interrupt + timestamp), then select & read the GPIO value. >> >>>>>>>>>>>>>>> >> >>>>>>>>>>>>>>> When issue happen, the recording app side have got same = number of network data packages and gpio interrupts with that DUT have been= toggling?=20=20 >> >>>>>>>>>>>>>>> I am asking this because I met gpio hardware issue that = would cause gpio interrupt missing or invalid gpio interrupts when gpio is = connecting on two different boards. >> >>>>>>>>>>>>>>> >> >>>>>>>>>>>>>> >> >>>>>>>>>>>>>> AFAICT, I'm not missing GPIO edges/interrupts in the app.= Regarding network, the DUT and the recoding device are connected to the sa= me switch with the same port config. >> >>>>>>>>>>>>>> >> >>>>>>>>>>>>>> Moreover, I tried changing the interrupt polarity of the = recording device (switched from rising edge to falling edge), and the resul= t is 1-2ms + ~30ms (the width of the pulse of the DUT) + 50ms latency >> >>>>>>>>>>>>>> >> >>>>>>>>>>>>> >> >>>>>>>>>>>>> Can you (or did you already) break down the latencies on t= he DUT via >> >>>>>>>>>>>>> tracing? Is it really the GPIO output path? What is happen= ing in it, >> >>>>>>>>>>>>> starting with RT/Linux task switches, mode transitions etc= .? >> >>>>>>>>>>>>> >> >>>>>>>>>>>> >> >>>>>>>>>>>> I just did it. >> >>>>>>>>>>>> >> >>>>>>>>>>>> I did put gpio_pin_interrupt as trigger, then ran my applic= ation, triggered the GPIO. >> >>>>>>>>>>>> This is what I get in frozen, but I'm not quite sure how to= read it and what conclusion I could derive : >> >>>>>>>>>>>> cat frozen >> >>>>>>>>>>>> I-pipe frozen back-tracing service on 4.4.227+/ipipe releas= e #10 >> >>>>>>>>>>>> ------------------------------------------------------------ >> >>>>>>>>>>>> CPU: 0, Freeze: 218343820846 cycles, Trace Points: 100 (+10) >> >>>>>>>>>>>> Calibrated minimum trace-point overhead: 0.461 us >> >>>>>>>>>>>> >> >>>>>>>>>>>> +----- Hard IRQs ('|': locked) >> >>>>>>>>>>>> |+-- Xenomai >> >>>>>>>>>>>> ||+- Linux ('*': domain stalled, '+': current, '#': curren= t+stalled) >> >>>>>>>>>>>> ||| +---------- Delay flag ('+': > 1 = us, '!': > 10 us) >> >>>>>>>>>>>> ||| | +- NMI noise ('N') >> >>>>>>>>>>>> ||| | | >> >>>>>>>>>>>> Type User Val. Time Delay Function (Pare= nt) >> >>>>>>>>>>>> : +func -6087+ 1.302 load_balance+0x14 (r= un_rebalance_domains+0x7e8) >> >>>>>>>>>>>> : +func -6085 0.826 idle_cpu+0x10 (loa= d_balance+0x180) >> >>>>>>>>>>>> : +func -6084 0.892 find_busiest_group= +0x14 (load_balance+0x1a4) >> >>>>>>>>>>>> : +func -6084 0.757 update_group_capac= ity+0x14 (find_busiest_group+0x128) >> >>>>>>>>>>>> : +func -6083+ 1.452 __msecs_to_jiffies+0= x10 (update_group_capacity+0x30) >> >>>>>>>>>>>> : +func -6081+ 1.535 idle_cpu+0x10 (find_= busiest_group+0x1e4) >> >>>>>>>>>>>> : +func -6080+ 1.410 idle_cpu+0x10 (find_= busiest_group+0x1e4) >> >>>>>>>>>>>> : +func -6078 0.967 __msecs_to_jiffies= +0x10 (run_rebalance_domains+0x810) >> >>>>>>>>>>>> : +func -6077 0.886 __rcu_read_unlock+= 0x10 (run_rebalance_domains+0x648) >> >>>>>>>>>>>> : +func -6077 0.820 rcu_bh_qs+0x10 (__= do_softirq+0x1b0) >> >>>>>>>>>>>> : +func -6076 0.742 ipipe_stall_root+0= x10 (__do_softirq+0x1b4) >> >>>>>>>>>>>> : +func -6075 0.766 ipipe_root_only+0x= 10 (ipipe_stall_root+0x18) >> >>>>>>>>>>>> :| +begin 0x80000001 -6074 0.934 ipipe_trace_begin+= 0x24 (ipipe_root_only+0xb8) >> >>>>>>>>>>>> :| +end 0x80000001 -6073 0.811 ipipe_trace_end+0x= 24 (ipipe_root_only+0xfc) >> >>>>>>>>>>>> :| +begin 0x80000001 -6072 0.895 ipipe_trace_begin+= 0x24 (ipipe_stall_root+0x78) >> >>>>>>>>>>>> :| #end 0x80000001 -6072 0.847 ipipe_trace_end+0x= 24 (ipipe_stall_root+0xb8) >> >>>>>>>>>>>> : #func -6071 0.814 __local_bh_enable+= 0x10 (__do_softirq+0x214) >> >>>>>>>>>>>> : #func -6070 0.760 ipipe_test_root+0x= 10 (__local_bh_enable+0x1c) >> >>>>>>>>>>>> :| #begin 0x80000001 -6069 0.907 ipipe_trace_begin+= 0x24 (ipipe_test_root+0x74) >> >>>>>>>>>>>> :| #end 0x80000001 -6068 0.898 ipipe_trace_end+0x= 24 (ipipe_test_root+0xb8) >> >>>>>>>>>>>> : #func -6067 0.811 rcu_irq_exit+0x10 = (irq_exit+0x84) >> >>>>>>>>>>>> : #func -6067 0.781 ipipe_test_and_sta= ll_root+0x10 (rcu_irq_exit+0x18) >> >>>>>>>>>>>> : #func -6066 0.799 ipipe_root_only+0x= 10 (ipipe_test_and_stall_root+0x18) >> >>>>>>>>>>>> :| #begin 0x80000001 -6065+ 1.041 ipipe_trace_begin+0x= 24 (ipipe_root_only+0xb8) >> >>>>>>>>>>>> :| #end 0x80000001 -6064 0.769 ipipe_trace_end+0x= 24 (ipipe_root_only+0xfc) >> >>>>>>>>>>>> :| #begin 0x80000001 -6063 0.895 ipipe_trace_begin+= 0x24 (ipipe_test_and_stall_root+0x80) >> >>>>>>>>>>>> :| #end 0x80000001 -6062 0.841 ipipe_trace_end+0x= 24 (ipipe_test_and_stall_root+0xc4) >> >>>>>>>>>>>> : #func -6061+ 1.197 rcu_eqs_enter_common= .constprop.21+0x10 (rcu_irq_exit+0x80) >> >>>>>>>>>>>> :| #begin 0x80000000 -6060+ 1.413 ipipe_trace_begin+0x= 24 (__ipipe_do_sync_stage+0x2b8) >> >>>>>>>>>>>> :| +end 0x00000012 -6059+ 1.044 ipipe_trace_end+0x24= (__ipipe_grab_irq+0x84) >> >>>>>>>>>>>> :| +func -6058 0.988 __ipipe_check_root= _interruptible+0x10 (__irq_svc+0x70) >> >>>>>>>>>>>> :| +func -6057 0.976 ipipe_test_root+0x= 10 (__ipipe_check_root_interruptible+0x68) >> >>>>>>>>>>>> :| +func -6056 0.829 __ipipe_bugon_irqs= _enabled+0x10 (__ipipe_fast_svc_irq_exit+0x4) >> >>>>>>>>>>>> :| +end 0x90000000 -6055 0.913 __ipipe_fast_svc_i= rq_exit+0x20 (ipipe_unstall_root+0x88) >> >>>>>>>>>>>> : +func -6054 0.781 ipipe_test_root+0x= 10 (cpu_startup_entry+0x12c) >> >>>>>>>>>>>> :| +begin 0x80000001 -6053 0.868 ipipe_trace_begin+= 0x24 (ipipe_test_root+0x74) >> >>>>>>>>>>>> :| +end 0x80000001 -6052 0.781 ipipe_trace_end+0x= 24 (ipipe_test_root+0xb8) >> >>>>>>>>>>>> : +func -6052 0.748 rcu_idle_exit+0x10= (cpu_startup_entry+0x138) >> >>>>>>>>>>>> : +func -6051 0.739 ipipe_test_and_sta= ll_root+0x10 (rcu_idle_exit+0x18) >> >>>>>>>>>>>> : +func -6050 0.775 ipipe_root_only+0x= 10 (ipipe_test_and_stall_root+0x18) >> >>>>>>>>>>>> :| +begin 0x80000001 -6049+ 1.011 ipipe_trace_begin+0x= 24 (ipipe_root_only+0xb8) >> >>>>>>>>>>>> :| +end 0x80000001 -6048 0.742 ipipe_trace_end+0x= 24 (ipipe_root_only+0xfc) >> >>>>>>>>>>>> :| +begin 0x80000001 -6048 0.940 ipipe_trace_begin+= 0x24 (ipipe_test_and_stall_root+0x80) >> >>>>>>>>>>>> :| #end 0x80000001 -6047 0.790 ipipe_trace_end+0x= 24 (ipipe_test_and_stall_root+0xc4) >> >>>>>>>>>>>> : #func -6046 0.859 rcu_eqs_exit_commo= n.constprop.19+0x10 (rcu_idle_exit+0x8c) >> >>>>>>>>>>>> : #func -6045 0.772 ipipe_unstall_root= +0x10 (rcu_idle_exit+0x78) >> >>>>>>>>>>>> :| #begin 0x80000000 -6044 0.814 ipipe_trace_begin+= 0x24 (ipipe_unstall_root+0x98) >> >>>>>>>>>>>> :| #func -6043+ 1.077 ipipe_root_only+0x10= (ipipe_unstall_root+0x24) >> >>>>>>>>>>>> :| +end 0x80000000 -6042 0.835 ipipe_trace_end+0x= 24 (ipipe_unstall_root+0x84) >> >>>>>>>>>>>> : +func -6042 0.922 arch_cpu_idle_exit= +0x10 (cpu_startup_entry+0xfc) >> >>>>>>>>>>>> : +func -6041 0.793 ipipe_stall_root+0= x10 (cpu_startup_entry+0xc4) >> >>>>>>>>>>>> : +func -6040 0.724 ipipe_root_only+0x= 10 (ipipe_stall_root+0x18) >> >>>>>>>>>>>> :| +begin 0x80000001 -6039+ 1.098 ipipe_trace_begin+0x= 24 (ipipe_root_only+0xb8) >> >>>>>>>>>>>> :| +end 0x80000001 -6038 0.772 ipipe_trace_end+0x= 24 (ipipe_root_only+0xfc) >> >>>>>>>>>>>> :| +begin 0x80000001 -6037 0.841 ipipe_trace_begin+= 0x24 (ipipe_stall_root+0x78) >> >>>>>>>>>>>> :| #end 0x80000001 -6036 0.763 ipipe_trace_end+0x= 24 (ipipe_stall_root+0xb8) >> >>>>>>>>>>>> : #func -6036 0.838 arch_cpu_idle_ente= r+0x10 (cpu_startup_entry+0xc8) >> >>>>>>>>>>>> : #func -6035 0.745 arm_heavy_mb+0x10 = (arch_cpu_idle_enter+0x1c) >> >>>>>>>>>>>> : #func -6034 0.916 l2c210_sync+0x10 (= arm_heavy_mb+0x2c) >> >>>>>>>>>>>> : #func -6033+ 1.062 tick_check_broadcast= _expired+0x10 (cpu_startup_entry+0xd8) >> >>>>>>>>>>>> : #func -6032 0.787 rcu_idle_enter+0x1= 0 (cpu_startup_entry+0x124) >> >>>>>>>>>>>> : #func -6031 0.745 ipipe_test_and_sta= ll_root+0x10 (rcu_idle_enter+0x18) >> >>>>>>>>>>>> : #func -6031 0.751 ipipe_root_only+0x= 10 (ipipe_test_and_stall_root+0x18) >> >>>>>>>>>>>> :| #begin 0x80000001 -6030 0.991 ipipe_trace_begin+= 0x24 (ipipe_root_only+0xb8) >> >>>>>>>>>>>> :| #end 0x80000001 -6029 0.772 ipipe_trace_end+0x= 24 (ipipe_root_only+0xfc) >> >>>>>>>>>>>> :| #begin 0x80000001 -6028 0.892 ipipe_trace_begin+= 0x24 (ipipe_test_and_stall_root+0x80) >> >>>>>>>>>>>> :| #end 0x80000001 -6027 0.847 ipipe_trace_end+0x= 24 (ipipe_test_and_stall_root+0xc4) >> >>>>>>>>>>>> : #func -6026 0.922 rcu_eqs_enter_comm= on.constprop.21+0x10 (rcu_idle_enter+0x90) >> >>>>>>>>>>>> : #func -6025 0.862 default_idle_call+= 0x10 (cpu_startup_entry+0x128) >> >>>>>>>>>>>> : #func -6024 0.877 arch_cpu_idle+0x10= (default_idle_call+0x38) >> >>>>>>>>>>>> :| #begin 0x80000000 -6024! 5992.167 ipipe_trace_begin+0= x24 (arch_cpu_idle+0xb8) >> >>>>>>>>>>> >> >>>>>>>>>>> Here your system (or this core) went idle, waiting for the n= ext event. >> >>>>>>>>>>> >> >>>>>>>>>>>> :| +func -31 0.760 ipipe_unstall_root= +0x10 (arch_cpu_idle+0x30) >> >>>>>>>>>>> >> >>>>>>>>>>> Comming back from idle. >> >>>>>>>>>>> >> >>>>>>>>>>>> :| +func -31+ 1.116 ipipe_root_only+0x10= (ipipe_unstall_root+0x24) >> >>>>>>>>>>>> :| +end 0x80000000 -30 0.931 ipipe_trace_end+0x= 24 (ipipe_unstall_root+0x84) >> >>>>>>>>>>>> :| +begin 0x90000000 -29 0.844 __irq_svc+0x58 (ip= ipe_unstall_root+0x88) >> >>>>>>>>>>>> :| +func -28 0.925 gic_handle_irq+0x1= 0 (__irq_svc+0x6c) >> >>>>>>>>>>>> :| +func -27 0.904 irq_find_mapping+0= x10 (gic_handle_irq+0x50) >> >>>>>>>>>>>> :| +func -26 0.940 __ipipe_grab_irq+0= x10 (gic_handle_irq+0x58) >> >>>>>>>>>>>> :| +begin 0x000000c9 -25 0.826 ipipe_trace_begin+= 0x24 (__ipipe_grab_irq+0x58) >> >>>>>>>>>>>> :| +func -24 0.814 __ipipe_dispatch_i= rq+0x10 (__ipipe_grab_irq+0x7c) >> >>>>>>>>>>>> :| +func -23+ 1.275 irq_to_desc+0x10 (__= ipipe_dispatch_irq+0x184) >> >>>>>>>>>>>> :| +func -22+ 1.679 irq_to_desc+0x10 (__= ipipe_dispatch_irq+0x198) >> >>>>>>>>>>>> :| +func -20+ 2.092 ucc_gpio_irqhandler+= 0x14 (__ipipe_dispatch_irq+0x1fc) >> >>>>>>>>>>>> :| +func -18+ 1.413 irq_find_mapping+0x1= 0 (ucc_gpio_irqhandler+0x84) >> >>>>>>>>>>>> :| +begin 0x000000e3 -17 0.757 ipipe_trace_begin+= 0x24 (ucc_gpio_irqhandler+0x8c) >> >>>>>>>>>>>> :| +func -16 0.778 __ipipe_dispatch_i= rq+0x10 (ucc_gpio_irqhandler+0x98) >> >>>>>>>>>>>> :| +func -15+ 1.023 irq_to_desc+0x10 (__= ipipe_dispatch_irq+0x184) >> >>>>>>>>>>>> :| +func -14+ 1.494 irq_to_desc+0x10 (__= ipipe_dispatch_irq+0x198) >> >>>>>>>>>>>> :| +func -13+ 1.014 __ipipe_ack_level_ir= q+0x10 (__ipipe_dispatch_irq+0x1fc) >> >>>>>>>>>>>> :| +func -12 0.763 ucc_gpio_irq_mask+= 0x10 (__ipipe_ack_level_irq+0x54) >> >>>>>>>>>>>> :| +func -11+ 1.248 __ipipe_spin_lock_ir= qsave+0x10 (ucc_gpio_irq_mask+0x2c) >> >>>>>>>>>>>> :| #func -10+ 1.619 __ipipe_spin_unlock_= irqrestore+0x10 (ucc_gpio_irq_mask+0x4c) >> >>>>>>>>>>>> :| +func -8+ 1.239 __ipipe_set_irq_pend= ing+0x10 (__ipipe_dispatch_irq+0x3bc) >> >>>>>>>>>>>> :| +end 0x000000e3 -7 0.994 ipipe_trace_end+0x= 24 (ucc_gpio_irqhandler+0xa0) >> >>>>>>>>>>>> :| +func -6+ 1.029 gic_eoi_irq+0x10 (uc= c_gpio_irqhandler+0xd4) >> >>>>>>>>>>>> :| +func -5+ 1.353 __ipipe_do_sync_pipe= line+0x14 (__ipipe_dispatch_irq+0x17c) >> >>>>>>>>>>>> :|+ func -4+ 1.449 __ipipe_do_sync_stag= e+0x14 (__ipipe_do_sync_pipeline+0xf0) >> >>>>>>>>>>>> :|# func -2+ 1.191 xnintr_irq_handler+0= x14 (__ipipe_do_sync_stage+0x200) >> >>>>>>>>>>>> :|# func -1+ 1.455 ___xnlock_get+0x10 (= xnintr_irq_handler+0xc0) >> >>>>>>>>>>>> <|# func 0 1.107 gpio_pin_interrupt= +0x10 (xnintr_irq_handler+0xf4) >> >>>>>>>>>>> >> >>>>>>>>>>> And here we start to process that GPIO interrupt in the RTDM= handler, >> >>>>>>>>>>> roughly after 30 =C2=B5s (which are also due to tracing over= head). So far >> >>>>>>>>>>> nothing suspiciuos. >> >>>>>>>>>>> >> >>>>>>>>>>>> |# func 1 1.236 xnclock_core_read_= monotonic+0x10 (gpio_pin_interrupt+0x1c) >> >>>>>>>>>>>> |# func 2 1.224 rtdm_event_signal+= 0x10 (gpio_pin_interrupt+0x2c) >> >>>>>>>>>>> >> >>>>>>>>>>> Here some likely waiting RT task is signalled. Does that one= notice a >> >>>>>>>>>>> too high latency? >> >>>>>>>>>> >> >>>>>>>>>> I may have been wrong speaking of "latency". The problem I'm = trying to fix, is understanding why my reference instrument indicates a del= ay from network packet to GPIO on the DUT of about 1 to 2ms, whereas my xen= omai powered realtime recorder application, given the same network stream a= nd gpio access gives me ~50 to 52 ms. >> >>>>>>>>>> >> >>>>>>>>>> I first though there was something wrong/delayed on the GPIO = interrupt side (the network packet could not have been received before it i= s sent right). >> >>>>>>>>>> The trace seem to demonstrate ~30=C2=B5s of interrupt latency= (a number I expected for interrupt latency on that system) for the GPIO. S= o something is wrong on my system, but I don't know what ! >> >>>>>>>>>> >> >>>>>>>>> >> >>>>>>>>> Try to trace events, not functions, using regular ftrace ("tra= ce-cmd >> >>>>>>>>> record -e cobalt* -e sched* -e irq* -e signal*" e.g.). Check w= hen the >> >>>>>>>>> NIC gets the interrupt and compare that to when the GPIO event= is >> >>>>>>>>> triggered (or whatever is trigger and reaction). Function trac= ing is for >> >>>>>>>>> zooming in when you know where to zoom. >> >>>>>>>>> >> >>>>>>>> >> >>>>>>>> So I think I found the answer to my problem. >> >>>>>>>> Is there any reason why NET events (in NIC irq functions) are d= ated with rtdm_clock_read, whereas GPIO event are dated with rtdm_clock_rea= d_monotonic ? >> >>>>>>>> >> >>>>>>> >> >>>>>>> Inconsistency of the GPIO drivers, UART drivers use rtdm_clock_r= ead for >> >>>>>>> user-exposed timestamping as well. Maybe Philippe can comment on >> >>>>>>> thoughts behind this deviation. >> >>>>>>> >> >>>>>> >> >>>>>> When running over the I-pipe, rtdm_clock_read() is based on Xenom= ai's >> >>>>>> idea of real time, which is the Cobalt monotonic clock plus an ar= bitrary >> >>>>>> offset. For this reason, rtdm_clock_read() is not SMP-consistent >> >>>>>> (different CPUs might read different timestamps at the exact same= time), >> >>>>>> is not in sync with linux's wall clock either. For these reasons,= I >> >>>>>> don't see any practical way to synchronize multiple systems on th= e clock >> >>>>>> underlying rtdm_clock_read(). >> >>>>>> >> >>>>>> Therefore, there is no upside in using rtdm_clock_read() for >> >>>>>> timestamping in this context, only adding the potential for even = more >> >>>>>> surprising results due to the mono->real-time offset changing und= er our >> >>>>>> feet, since the epoch of the Xenomai real-time clock can be (re)s= et >> >>>>>> during runtime. >> >>>>>> >> >>>>>> I believe the UART driver is wrong here, it should return timesta= mps >> >>>>>> based on the monotonic source, which best fits the common need: g= etting >> >>>>>> timestamps from the local CPU for measuring delays between events >> >>>>>> received by drivers and the actions taken by the applications whi= ch >> >>>>>> consume them, immune from updates to the underlying clock >> >>>>>> epoch. Granted, there might be a catch when a timestamp is taken = from >> >>>>>> IRQ context, which is then consumed from a thread living on a dif= ferent >> >>>>>> CPU, if per-CPU clocks are not/badly synchronized. But that would= happen >> >>>>>> the exact same way with rtdm_clock_read() anyway. >> >>>>>> >> >>>>>> The situation improves when running on top of Dovetail, since Xen= omai >> >>>>>> now refers to the common linux clocks (mono / real) instead of pr= oviding >> >>>>>> its own idea of time, but the UART driver code predates the Dovet= ail >> >>>>>> port. >> >>>>>>=20=20 >> >>>>> >> >>>>> Well, it's not just UART. All hardware drivers - except for GPIO -= use >> >>>>> rtdm_clock_read. That was no problem in practice for their use cas= es so >> >>>>> far. One problem is that GPIO timestamps are now not comparable to= others. >> >>>>> >> >>>>> But IIRC, most Xenomai APIs using absolute timestamps are based on >> >>>>> Xenomai's real-time clock. Therefore, providing timestamps for that >> >>>> >> >>>> /* >> >>>> * The Copperplate clock shall be monotonic unless the threading >> >>>> * library has restrictions to support this over Mercury. >> >>>> * >> >>>> * In the normal case, this means that ongoing delays and timeouts >> >>>> * won't be affected when the host system date is changed. In the >> >>>> * restricted case by contrast, ongoing delays and timeouts may be >> >>>> * impacted by changes to the host system date. >> >>>> * >> >>>> * The implementation maintains a per-clock epoch value, so that >> >>>> * different emulators can have different (virtual) system dates. >> >>>> */ >> >>>> #ifdef CONFIG_XENO_COPPERPLATE_CLOCK_RESTRICTED >> >>>> #define CLOCK_COPPERPLATE CLOCK_REALTIME >> >>>> #else >> >>>> #define CLOCK_COPPERPLATE CLOCK_MONOTONIC >> >>>> #endif >> >>>> >> >>>> So no, only the POSIX API is using what the standard mandates, whic= h is >> >>>> CLOCK_REALTIME. All other APIs are based on copperplate, and they a= re >> >>>> using a monotonic source as documented above. >> >>> >> >>> OK, but that changed in 3.x. At the time that RTDM API was originally >> >>> added and then promoted, it was the other not this way. We became >> >>> inconsistent then. >> >>> >> >>>> >> >>>>> particular clock was the original idea of rtdm_clock_read (which >> >>>>> predates rtdm_clock_read_monotonic). GPIO breaks that and should be >> >>>>> fixed - unless I'm wrong with that assumption. >> >>>>> >> >>>> >> >>>> We cannot assume the epoch is going to remain stable with >> >>>> rtdm_clock_read() the way it is implemented, which is quite of a >> >>>> problem wrt the common use case. For timestamping, a majority of >> >>>> mainline drivers is using ktime_get() or a variant thereof which is >> >>>> based on the monotonic clock source, not the _real form. Why would = the >> >>>> real-time I/O drivers be different? >> >>> >> >>> We have two cases here: >> >>> >> >>> - I-pipe-based version where the realtime clock is under full >> >>> application control -> no problem to use rtdm_clock_read >> >>=20 >> >> Well, there is still the issue that rtdm_clock_read() is not immune to >> >> some part of userland changing the CLOCK_REALTIME epoch Xenomai-wise >> >> e.g. via a call to clock_settime(), which is the same problem than Li= nux >> >> changing the epoch of CLOCK_REALTIME over Dovetail. This would break = the >> >> application. >> > >> > Yes, but that's about the application(s) breaking themselves. Nothing >> > new, not going to change when we only avoid clock_realtime for stamps >> > but still use POSIX services basing timers on that clock. The key point >> > is that I-pipe gave that into RT application hands (with all related >> > downsides), with Dovetail it's in system hands. >> > >> >>=20 >> >>> - Dovetail where we share the realtime clock - with all its tuning - >> >>> with Linux -> here we have a problem with rtdm_clock_read and sho= uld >> >>> reconsider its usage (and promotion!) >> >>> >> >>> For stable 3.1, the proper fix is with GPIO going to rtdm_clock_read. >> >>=20 >> >> Wait, you have downstream users already depending on GPIO returning >> >> monotonic timestamps, and this is a _stable_ release. So why not fixi= ng >> >> other drivers based on the fact that timestamping with rtdm_clock_rea= d() >> >> is wrong instead? Same issue, right? >> > >> > GPIO was broken, but you are right that we may have users relying on >> > that breakage now. Obviously, we can change the other drivers for the >> > very same reasons: They are working like they work for more than 10 >> > years now. >> > >>=20 >> Sorry, but I my views, rtdm_read_clock() was broken since day one. GPIO >> had to work around the breakage.. :) >>=20 >> >>=20 >> >> So the best course of action to sort this out for 3.1.x may be to ext= end >> >> GPIO_RTIOC_TS with say, GPIO_RTIOC_TS_REAL, which would log and return >> >> timestamps based on the Xenomai wallclock. Applications which do want= to >> >> align on that clock would simply have to issue GPIO_RTIOC_TS_REAL >> >> instead of GPIO_RTIOC_TS. This would break backward ABI compat only f= or >> >> users of GPIO_RTIOC_TS_REAL, but that would be much better than >> >> introducing a sneaky change in behavior for the GPIO driver. >> > >> > Yeah, likely the way to go. >> > >> >>=20 >> >>> For 3.2, I'm not sure yet what to do with rtdm_clock_read. >> >>> >> >>=20 >> >> The decision looks pretty simple for the common use case: when >> >> timestamps are needed for performance/delay measurements, we want to = get >> >> them from a clock source which won't play funny games, warping back to >> >> the future. >> > >> > You only think of measurements. The other drivers used the stamping al= so >> > for real applications, means to calculate absolute clock-realtime >> > timeouts and wait for them to arrive. We will likely need to enhance >> > also the other driver APIs to select the desired clocksource, just like >> > for GPIO. >>=20 >> I'm referring to what is in the GPIO code, which is the problem at hand: >> that timestamping was designed since day #1 to provide a way to measure >> the scheduling latency. >>=20 >> I agree that the only way out is to enable all timestamp consumers to >> pick their base clock of choice (mono or wallclock). >>=20 >> --=20 >> Philippe. >=20=20 > Hi there, > > as we speak of breaking ABI, why not introducing some specific IOCTL to s= elect which timesource to use for each driver using it, then using a functi= on pointer to call the right rtdm_time_get_whatever function ? > That would mean adding a generic ioctl, and many changes all over the place (we would need no function pointer with distinct ioctl codes). Doable for sure, but likely too much for 3.1.x though. --=20 Philippe.