From ea66b8de458f1b59f727948c0c2fd034f243adcd Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Tomasz=20Mo=C5=84?= Date: Wed, 30 Oct 2024 06:33:09 +0100 Subject: [PATCH] drivers: udc_nrf: handle overwritten Set Address commands MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit USBD peripheral automatically handles Set Address command which can lead to state mismatch between USB stack and the host. Keep track of device address and issue fake Set Address commands on mismatch. This fixes default vs addressed state mismatch that can occur due to sufficently high SETUP handling latency. The state mismatch was most commonly seen as SET CONFIGURATION failure when the enumeration happened during periods with increased latency. Signed-off-by: Tomasz Moń --- drivers/usb/udc/udc_nrf.c | 119 +++++++++++++++++++++++++++++++++++--- 1 file changed, 110 insertions(+), 9 deletions(-) diff --git a/drivers/usb/udc/udc_nrf.c b/drivers/usb/udc/udc_nrf.c index c024d2a6a184702..feb8ff40c83539d 100644 --- a/drivers/usb/udc/udc_nrf.c +++ b/drivers/usb/udc/udc_nrf.c @@ -70,7 +70,8 @@ static struct k_thread drv_stack_data; static struct udc_ep_config ep_cfg_out[CFG_EPOUT_CNT + CFG_EP_ISOOUT_CNT + 1]; static struct udc_ep_config ep_cfg_in[CFG_EPIN_CNT + CFG_EP_ISOIN_CNT + 1]; -static bool udc_nrf_setup_rcvd; +static bool udc_nrf_setup_rcvd, udc_nrf_setup_set_addr, udc_nrf_fake_setup; +static uint8_t udc_nrf_address; const static struct device *udc_nrf_dev; struct udc_nrf_config { @@ -143,7 +144,9 @@ static void udc_event_xfer_ctrl_in(const struct device *dev, /* Update to next stage of control transfer */ udc_ctrl_update_stage(dev, buf); - nrf_usbd_common_setup_clear(); + if (!udc_nrf_setup_set_addr) { + nrf_usbd_common_setup_clear(); + } } static void udc_event_fake_status_in(const struct device *dev) @@ -317,6 +320,7 @@ static int usbd_ctrl_feed_dout(const struct device *dev, static int udc_event_xfer_setup(const struct device *dev) { + nrf_usbd_common_setup_t *setup; struct net_buf *buf; int err; @@ -328,7 +332,77 @@ static int udc_event_xfer_setup(const struct device *dev) } udc_ep_buf_set_setup(buf); - nrf_usbd_common_setup_get((nrf_usbd_common_setup_t *)buf->data); + setup = (nrf_usbd_common_setup_t *)buf->data; + nrf_usbd_common_setup_get(setup); + + /* USBD peripheral automatically handles Set Address in slightly + * different manner than the USB stack. + * + * USBD peripheral doesn't care about wLength, but the peripheral + * switches to new address only after status stage. The device won't + * automatically accept Data Stage packets. + * + * However, in the case the host: + * * sends SETUP Set Address with non-zero wLength + * * does not set corresponding OUT DATA packets (to match wLength) + * or sends the packets but disregards NAK + * or sends the packets that device ACKs + * * sends IN token (either incorrectly proceeds to status stage, or + * manages to send IN before SW sets STALL) + * then the USBD peripheral will accept the address and USB stack won't. + * This will lead to state mismatch between the stack and peripheral. + * + * In cases where the SW stack would like to STALL the request there is + * a race condition between host issuing Set Address status stage (IN + * token) and SW setting STALL bit. If host wins the race, the device + * ACKs status stage and use new address. If device wins the race, the + * device STALLs status stage and address remains unchanged. + */ + udc_nrf_setup_set_addr = + setup->bmRequestType == 0 && + setup->bRequest == USB_SREQ_SET_ADDRESS; + if (udc_nrf_setup_set_addr) { + if (setup->wLength) { + /* Currently USB stack only STALLs OUT Data Stage when + * buffer allocation fails. To prevent the device from + * ACKing the Data Stage, simply ignore the request + * completely. + * + * If host incorrectly proceeds to status stage there + * will be address mismatch (unless the new address is + * equal to current device address). If host does not + * issue IN token then the mismatch will be avoided. + */ + net_buf_unref(buf); + return 0; + } + + /* nRF52/nRF53 USBD doesn't care about wValue bits 8..15 and + * wIndex value but USB device stack does. + * + * Just clear the bits so stack will handle the request in the + * same way as USBD peripheral does, avoiding the mismatch. + */ + setup->wValue &= 0x7F; + setup->wIndex = 0; + } + + if (!udc_nrf_setup_set_addr && udc_nrf_address != NRF_USBD->USBADDR) { + /* Address mismatch detected. Fake Set Address handling to + * correct the situation, then repeat handling. + */ + udc_nrf_fake_setup = true; + udc_nrf_setup_set_addr = true; + + setup->bmRequestType = 0; + setup->bRequest = USB_SREQ_SET_ADDRESS; + setup->wValue = NRF_USBD->USBADDR; + setup->wIndex = 0; + setup->wLength = 0; + } else { + udc_nrf_fake_setup = false; + } + net_buf_add(buf, sizeof(nrf_usbd_common_setup_t)); udc_nrf_setup_rcvd = true; @@ -494,7 +568,8 @@ static void udc_nrf_fake_status_in(const struct device *dev) .ep = USB_CONTROL_EP_IN, }; - if (nrf_usbd_common_last_setup_dir_get() == USB_CONTROL_EP_OUT) { + if (nrf_usbd_common_last_setup_dir_get() == USB_CONTROL_EP_OUT || + udc_nrf_fake_setup) { /* Let controller perform status IN stage */ k_msgq_put(&drv_msgq, &evt, K_NO_WAIT); } @@ -608,12 +683,38 @@ static int udc_nrf_ep_clear_halt(const struct device *dev, static int udc_nrf_set_address(const struct device *dev, const uint8_t addr) { - /** - * Nothing to do here. The USBD HW already takes care of initiating - * STATUS stage. Just double check the address for sanity. + /* + * If the status stage already finished (which depends entirely on when + * the host sends IN token) then NRF_USBD->USBADDR will have the same + * address, otherwise it won't (unless new address is unchanged). + * + * Store the address so the driver can detect address mismatches + * between USB stack and USBD peripheral. The mismatches can occur if: + * * SW has high enough latency in SETUP handling, or + * * Host did not issue Status Stage after Set Address request + * + * The SETUP handling latency is a problem because the Set Address is + * automatically handled by device. Because whole Set Address handling + * can finish in less than 21 us, the latency required (with perfect + * timing) to hit the issue is relatively short (2 ms Set Address + * recovery interval + negligible Set Address handling time). If host + * sends new SETUP before SW had a chance to read the Set Address one, + * the Set Address one will be overwritten without a trace. */ - if (addr != (uint8_t)NRF_USBD->USBADDR) { - LOG_WRN("USB Address incorrect 0x%02x", addr); + udc_nrf_address = addr; + + if (udc_nrf_fake_setup) { + struct udc_nrf_evt evt = { + .type = UDC_NRF_EVT_HAL, + .hal_evt = { + .type = NRF_USBD_COMMON_EVT_SETUP, + }, + }; + + /* Finished handling lost Set Address, now handle the pending + * SETUP transfer. + */ + k_msgq_put(&drv_msgq, &evt, K_NO_WAIT); } return 0;