Marc Kleine-Budde <mkl@xxxxxxxxxxxxxx> 於 2024年10月28日 週一 下午10:06寫道： > > On 28.10.2024 16:31:25, Ming Yu wrote: > > > > > > > > > The Linux USB stack can receive bulk messages longer than the max packet size. > > > > > > > > > > > > > > > > [Ming] Since NCT6694's bulk pipe endpoint size is 128 bytes for this MFD device. > > > > > > > > The core will divide packet 256 bytes for high speed USB device, but > > > > > > > > it is exceeds > > > > > > > > the hardware limitation, so I am dividing it manually. > > > > > > > > > > > > > > You say the endpoint descriptor is correctly reporting it's max packet > > > > > > > size of 128, but the Linux USB will send packets of 256 bytes? > > > > > > > > > > > > [Ming] The endpoint descriptor is correctly reporting it's max packet > > > > > > size of 256, but the Linux USB may send more than 256 (max is 512) > > > > > > https://elixir.bootlin.com/linux/v6.11.5/source/drivers/usb/host/xhci-mem.c#L1446 > > > > > > > > > > AFAIK according to the USB-2.0 spec the maximum packet size for > > > > > high-speed bulk transfers is fixed set to 512 bytes. Does this mean that > > > > > your device is a non-compliant USB device? > > > > > > > > We will reduce the endpoint size of other interfaces to ensure that MFD device > > > > meets the USB2.0 spec. In other words, I will remove the code for manual > > > > unpacking in the next patch. > > > > > > I was not talking about the driver, but your USB device. According to > > > the USB2.0 spec, the packet size is fixed to 512 for high-speed bulk > > > transfers. So your device must be able to handle 512 byte transfers or > > > it's a non-compliant USB device. > > > > I understand. Therefore, the USB device's firmware will be modified to support > > bulk pipe size of 512 bytes to comply with the USB 2.0 spec. > > Then you don't need manual segmentation of bulk transfers anymore! Understood, thank you very much. > > > > > > > > > > > + for (i = 0, len = length; len > 0; i++, len -= packet_len) { > > > > > > > > > > + if (len > nct6694->maxp) > > > > > > > > > > + packet_len = nct6694->maxp; > > > > > > > > > > + else > > > > > > > > > > + packet_len = len; > > > > > > > > > > + > > > > > > > > > > + ret = usb_bulk_msg(udev, usb_rcvbulkpipe(udev, BULK_IN_ENDPOINT), > > > > > > > > > > + nct6694->rx_buffer + nct6694->maxp * i, > > > > > > > > > > + packet_len, &rx_len, nct6694->timeout); > > > > > > > > > > + if (ret) > > > > > > > > > > + goto err; > > > > > > > > > > + } > > > > > > > > > > + > > > > > > > > > > + for (i = 0; i < rd_len; i++) > > > > > > > > > > + buf[i] = nct6694->rx_buffer[i + rd_idx]; > > > > > > > > > > > > > > > > > > memcpy()? > > > > > > > > > > > > > > > > > > Or why don't you directly receive data into the provided buffer? Copying > > > > > > > > > of the data doesn't make it faster. > > > > > > > > > > > > > > > > > > On the other hand, receiving directly into the target buffer means the > > > > > > > > > target buffer must not live on the stack. > > > > > > > > > > > > > > > > [Ming] Okay! I'll change it to memcpy(). > > > > > > > > > > > > > > fine! > > > > > > > > > > > > > > > This is my perspective: the data is uniformly received by the rx_bffer held > > > > > > > > by the MFD device. does it need to be changed? > > > > > > > > > > > > > > My question is: Why do you first receive into the nct6694->rx_buffer and > > > > > > > then memcpy() to the buffer provided by the caller, why don't you > > > > > > > directly receive into the memory provided by the caller? > > > > > > > > > > > > [Ming] Due to the bulk pipe maximum packet size limitation, I think consistently > > > > > > using the MFD'd dynamically allocated buffer to submit URBs will better > > > > > > manage USB-related operations > > > > > > > > > > The non-compliant max packet size limitation is unrelated to the > > > > > question which RX or TX buffer to use. > > > > > > > > I think these two USB functions can be easily called using the buffer > > > > dynamically > > > > allocated by the MFD. However, if they transfer data directly to the > > > > target buffer, > > > > they must ensure that it is not located on the stack. > > > > > > You have a high coupling between the MFD driver and the individual > > > drivers anyways, so why not directly use the dynamically allocated > > > buffer provided by the caller and get rid of the memcpy()? > > > > Okay! I will provide a function to request and free buffer for child devices, > > and update the caller's variables to use these two functions in the next patch. > > I don't see a need to provide dedicated function to allocate and free > the buffers. The caller can allocate them as part of their private data, > or allocate them during probe(). Okay, so each child device may allocate a buffer like this during probe(): priv->xmit_buf = devm_kcalloc(dev, MAX_PACKET_SZ, sizeof(unsigned char), GFP_KERNEL), right? > > regards, > Marc > > -- > Pengutronix e.K. | Marc Kleine-Budde | > Embedded Linux | https://www.pengutronix.de | > Vertretung Nürnberg | Phone: +49-5121-206917-129 | > Amtsgericht Hildesheim, HRA 2686 | Fax: +49-5121-206917-9 | Thanks, Ming
