Hi Doug, On 7/28/20 5:48 PM, Doug Anderson wrote: > Hi, > > On Sun, Jul 26, 2020 at 2:44 AM Stanimir Varbanov > <stanimir.varbanov@xxxxxxxxxx> wrote: >> >> Hi Stephen, >> >> On 7/25/20 12:17 AM, Stephen Boyd wrote: >>> From: Stephen Boyd <swboyd@xxxxxxxxxxxx> >>> >>> The busy loop in rpmh_rsc_send_data() is written with the assumption >>> that the udelay will be preempted by the tcs_tx_done() irq handler when >>> the TCS slots are all full. This doesn't hold true when the calling >>> thread is an irqthread and the tcs_tx_done() irq is also an irqthread. >>> That's because kernel irqthreads are SCHED_FIFO and thus need to >>> voluntarily give up priority by calling into the scheduler so that other >>> threads can run. >>> >>> I see RCU stalls when I boot with irqthreads on the kernel commandline >>> because the modem remoteproc driver is trying to send an rpmh async >>> message from an irqthread that needs to give up the CPU for the rpmh >>> irqthread to run and clear out tcs slots. >>> >>> rcu: INFO: rcu_preempt self-detected stall on CPU >>> rcu: 0-....: (1 GPs behind) idle=402/1/0x4000000000000002 softirq=2108/2109 fqs=4920 >>> (t=21016 jiffies g=2933 q=590) >>> Task dump for CPU 0: >>> irq/11-smp2p R running task 0 148 2 0x00000028 >>> Call trace: >>> dump_backtrace+0x0/0x154 >>> show_stack+0x20/0x2c >>> sched_show_task+0xfc/0x108 >>> dump_cpu_task+0x44/0x50 >>> rcu_dump_cpu_stacks+0xa4/0xf8 >>> rcu_sched_clock_irq+0x7dc/0xaa8 >>> update_process_times+0x30/0x54 >>> tick_sched_handle+0x50/0x64 >>> tick_sched_timer+0x4c/0x8c >>> __hrtimer_run_queues+0x21c/0x36c >>> hrtimer_interrupt+0xf0/0x22c >>> arch_timer_handler_phys+0x40/0x50 >>> handle_percpu_devid_irq+0x114/0x25c >>> __handle_domain_irq+0x84/0xc4 >>> gic_handle_irq+0xd0/0x178 >>> el1_irq+0xbc/0x180 >>> save_return_addr+0x18/0x28 >>> return_address+0x54/0x88 >>> preempt_count_sub+0x40/0x88 >>> _raw_spin_unlock_irqrestore+0x4c/0x6c >>> ___ratelimit+0xd0/0x128 >>> rpmh_rsc_send_data+0x24c/0x378 >>> __rpmh_write+0x1b0/0x208 >>> rpmh_write_async+0x90/0xbc >>> rpmhpd_send_corner+0x60/0x8c >>> rpmhpd_aggregate_corner+0x8c/0x124 >>> rpmhpd_set_performance_state+0x8c/0xbc >>> _genpd_set_performance_state+0xdc/0x1b8 >>> dev_pm_genpd_set_performance_state+0xb8/0xf8 >>> q6v5_pds_disable+0x34/0x60 [qcom_q6v5_mss] >>> qcom_msa_handover+0x38/0x44 [qcom_q6v5_mss] >>> q6v5_handover_interrupt+0x24/0x3c [qcom_q6v5] >>> handle_nested_irq+0xd0/0x138 >>> qcom_smp2p_intr+0x188/0x200 >>> irq_thread_fn+0x2c/0x70 >>> irq_thread+0xfc/0x14c >>> kthread+0x11c/0x12c >>> ret_from_fork+0x10/0x18 >>> >>> This busy loop naturally lends itself to using a wait queue so that each >>> thread that tries to send a message will sleep waiting on the waitqueue >>> and only be woken up when a free slot is available. This should make >>> things more predictable too because the scheduler will be able to sleep >>> tasks that are waiting on a free tcs instead of the busy loop we >>> currently have today. >>> >>> Cc: Douglas Anderson <dianders@xxxxxxxxxxxx> >>> Cc: Maulik Shah <mkshah@xxxxxxxxxxxxxx> >>> Cc: Lina Iyer <ilina@xxxxxxxxxxxxxx> >>> Signed-off-by: Stephen Boyd <swboyd@xxxxxxxxxxxx> >>> --- >>> >>> Changes in v2: >>> * Document tcs_wait >>> * Move wake_up() outside of the spinlock >>> * Document claim_tcs_for_req() >>> >>> drivers/soc/qcom/rpmh-internal.h | 4 ++ >>> drivers/soc/qcom/rpmh-rsc.c | 115 +++++++++++++++---------------- >>> 2 files changed, 58 insertions(+), 61 deletions(-) >> >> This also fixes an issue related to TCS busy, seen with Venus driver >> with these [1] patches applied. >> >> Tested-by: Stanimir Varbanov <stanimir.varbanov@xxxxxxxxxx> >> >> -- >> regards, >> Stan >> >> [1] https://lkml.org/lkml/2020/7/23/394 > > It worries me that you say that this fixes any issues for you. > Specifically I don't see how this could fix anything except: > > 1. Fix the problem with irqthreads, which is how Stephen originally found this. > > 2. Fix things to be a little more efficient. > > 3. Avoid the small handful of messages that show up during normal > usage that look like: > > TCS Busy, retrying RPMH message send: addr=... > > I'm guessing you're referring to #3. Is that correct? Yes, correct. After Stephen's comment on the Venus DVFS series that he dropped the message I realized why I don't see the TCS Busy anymore. > > If so, you might want to double-check to confirm that you aren't > totally spamming the RPMh bus with your patch series. I found that > when I was seeing a lot of "TCS Busy, retrying RPMH message send" > spammed to the console that it was a sign that the code was being > really inefficient. The runtime_suspend in venus driver will not happen too often, but dev_pm_opp_set_rate could be called on every decoded/encoded frame (we have so called dynamic clock scaling depending on the input bitstream). In any case, you gave me more material to think of. > > Specifically the code to add interconnect bandwidth and OPP to the SPI > drivers would run at "runtime_suspend" and "runtime_resume". For our > SPI drivers this meant that they were running after every single > transfer, and they were quite slow. If your code is doing similar > then you probably have a problem. > > The problem was fixed by adding an autosuspend delay. > > References (from linuxnext): > > 8592eb959ad4 spi: spi-qcom-qspi: Set an autosuspend delay of 250 ms > cfdab2cd85ec spi: spi-geni-qcom: Set an autosuspend delay of 250 ms > > > Also if your code is adjusting its clock rate too frequently it can > now trigger similar problems. References: > > 2124331411a1 spi: spi-qcom-qspi: Avoid clock setting if not needed > 5f219524ec6e spi: spi-geni-qcom: Set the clock properly at runtime resume > 68890e20942b spi: spi-geni-qcom: Avoid clock setting if not needed > > > -Doug > -- regards, Stan
