On Fri, 21 Mar 2025 at 12:18, Ekansh Gupta <ekansh.gupta@xxxxxxxxxxxxxxxx> wrote: > > > > On 3/20/2025 9:27 PM, Ekansh Gupta wrote: > > > > On 3/20/2025 7:45 PM, Dmitry Baryshkov wrote: > >> On Thu, Mar 20, 2025 at 07:19:31PM +0530, Ekansh Gupta wrote: > >>> On 1/29/2025 4:10 PM, Dmitry Baryshkov wrote: > >>>> On Wed, Jan 29, 2025 at 11:12:16AM +0530, Ekansh Gupta wrote: > >>>>> On 1/29/2025 4:59 AM, Dmitry Baryshkov wrote: > >>>>>> On Mon, Jan 27, 2025 at 10:12:38AM +0530, Ekansh Gupta wrote: > >>>>>>> For any remote call to DSP, after sending an invocation message, > >>>>>>> fastRPC driver waits for glink response and during this time the > >>>>>>> CPU can go into low power modes. Adding a polling mode support > >>>>>>> with which fastRPC driver will poll continuously on a memory > >>>>>>> after sending a message to remote subsystem which will eliminate > >>>>>>> CPU wakeup and scheduling latencies and reduce fastRPC overhead. > >>>>>>> With this change, DSP always sends a glink response which will > >>>>>>> get ignored if polling mode didn't time out. > >>>>>> Is there a chance to implement actual async I/O protocol with the help > >>>>>> of the poll() call instead of hiding the polling / wait inside the > >>>>>> invoke2? > >>>>> This design is based on the implementation on DSP firmware as of today: > >>>>> Call flow: https://github.com/quic-ekangupt/fastrpc/blob/invokev2/Docs/invoke_v2.md#5-polling-mode > >>>>> > >>>>> Can you please give some reference to the async I/O protocol that you've > >>>>> suggested? I can check if it can be implemented here. > >>>> As with the typical poll() call implementation: > >>>> - write some data using ioctl > >>>> - call poll() / select() to wait for the data to be processed > >>>> - read data using another ioctl > >>>> > >>>> Getting back to your patch. from you commit message it is not clear, > >>>> which SoCs support this feature. Reminding you that we are supporting > >>>> all kinds of platforms, including the ones that are EoLed by Qualcomm. > >>>> > >>>> Next, you wrote that in-driver polling eliminates CPU wakeup and > >>>> scheduling. However this should also increase power consumption. Is > >>>> there any measurable difference in the latencies, granted that you > >>>> already use ioctl() syscall, as such there will be two context switches. > >>>> What is the actual impact? > >>> Hi Dmitry, > >>> > >>> Thank you for your feedback. > >>> > >>> I'm currently reworking this change and adding testing details. Regarding the SoC > >>> support, I'll add all the necessary information. > >> Please make sure that both the kernel and the userspace can handle the > >> 'non-supported' case properly. > > Yes, I will include changes to handle in both userspace and kernel. > > I am seeking additional suggestions on handling "non-supported" cases before making the > changes. > > Userspace: To enable DSP side polling, a remote call is made as defined in the DSP image. > If this call fails, polling mode will not be enabled from userspace. No. Instead userspace should check with the kernel, which capabilities are supported. Don't perform API calls which knowingly can fail. > > Kernel: Since this is a DSP-specific feature, I plan to add a devicetree property, such > as "qcom,polling-supported," under the fastrpc node if the DSP supports polling mode. This doesn't sound like a logical solution. The kernel already knows the hardware that it is running on. As such, there should be no need to further describe the hardware in DT. If the DSP firmware can report its capabilities, use that. If not, extend the schema to add an SoC-specific compatibility string. As a last resort we can use of_machine_is_compatible(). > > Does this approach seem appropriate, or is there a better way to handle this? > > Thanks, > Ekansh > > > > >>> For now, with in-driver > >>> polling, we are seeing significant performance improvements for calls > >>> with different sized buffers. On polling supporting platform, I've observed an > >>> ~80us improvement in latency. You can find more details in the test > >>> results here: > >>> https://github.com/quic/fastrpc/pull/134/files#diff-7dbc6537cd3ade7fea5766229cf585db585704e02730efd72e7afc9b148e28ed > >> Does the improvement come from the CPU not goint to idle or from the > >> glink response processing? > > Although both are contributing to performance improvement, the major > > improvement is coming from CPU not going to idle state. > > > > Thanks, > > Ekansh > > > >>> Regarding your concerns about power consumption, while in-driver polling > >>> eliminates CPU wakeup and scheduling, it does increase power consumption. > >>> However, the performance gains seem to outweigh this increase. > >>> > >>> Do you think the poll implementation that you suggested above could provide similar > >>> improvements? > >> No, I agree here. I was more concentrated on userspace polling rather > >> than hw polling. > >> > -- With best wishes Dmitry
