spi_rtio: fix transactions for default handler

This patch fixes transaction op items not performed within a single
SPI transfer. This is common for Write + Read commands, that depend on
the CS kept asserted until the end, otherwise the context will be lost.
A similar fix was applied to i2c_rtio_default on #79890.

Signed-off-by: Luis Ubieda <[email protected]>
(cherry picked from commit b5ca5b7)

Author: ubieda

drivers/spi/Kconfig

@@ -37,6 +37,24 @@ config SPI_RTIO
 	  This option enables the RTIO API calls. RTIO support is
 	  experimental as the API itself is unstable.
 
+if SPI_RTIO
+
+config SPI_RTIO_FALLBACK_MSGS
+	int "Number of available spi_buf structs for the default handler to use"
+	default 4
+	help
+		When RTIO is used with a driver that does not yet implement the submit API
+		natively the submissions are converted back to struct spi_buf values that
+		are given to spi_transfer. This requires some number of msgs be available to convert
+		the submissions into on the stack. MISRA rules dictate we must know this in
+		advance.
+
+		In all likelihood 4 is going to work for everyone, but in case you do end up with
+		an issue where you are using RTIO, your driver does not implement submit natively,
+		and get an error relating to not enough spi msgs this is the Kconfig to manipulate.
+
+endif # SPI_RTIO
+
 config SPI_SLAVE
 	bool "Slave support [EXPERIMENTAL]"
 	select EXPERIMENTAL

drivers/spi/spi_rtio.c

@@ -22,6 +22,7 @@ static void spi_rtio_iodev_default_submit_sync(struct rtio_iodev_sqe *iodev_sqe)
 {
 	struct spi_dt_spec *dt_spec = iodev_sqe->sqe.iodev->data;
 	const struct device *dev = dt_spec->bus;
+	uint8_t num_msgs = 0;
 	int err = 0;
 
 	LOG_DBG("Sync RTIO work item for: %p", (void *)dev);
@@ -33,67 +34,103 @@ static void spi_rtio_iodev_default_submit_sync(struct rtio_iodev_sqe *iodev_sqe)
 	struct rtio_iodev_sqe *txn_head = iodev_sqe;
 	struct rtio_iodev_sqe *txn_curr = iodev_sqe;
 
+	/* We allocate the spi_buf's on the stack, to do so
+	 * the count of messages needs to be determined to
+	 * ensure we don't go over the statically sized array.
+	 */
 	do {
-		struct rtio_sqe *sqe = &txn_curr->sqe;
-		struct spi_buf tx_buf = {0};
-		struct spi_buf_set tx_buf_set = {
-			.buffers = &tx_buf,
-		};
+		switch (txn_curr->sqe.op) {
+		case RTIO_OP_RX:
+		case RTIO_OP_TX:
+		case RTIO_OP_TINY_TX:
+		case RTIO_OP_TXRX:
+			num_msgs++;
+			break;
+		default:
+			LOG_ERR("Invalid op code %d for submission %p", txn_curr->sqe.op,
+				(void *)&txn_curr->sqe);
+			err = -EIO;
+			break;
+		}
+		txn_curr = rtio_txn_next(txn_curr);
+	} while (err == 0 && txn_curr != NULL);
 
-		struct spi_buf rx_buf = {0};
-		struct spi_buf_set rx_buf_set = {
-			.buffers = &rx_buf,
-		};
+	if (err != 0) {
+		rtio_iodev_sqe_err(txn_head, err);
+		return;
+	}
+
+	/* Allocate msgs on the stack, MISRA doesn't like VLAs so we need a statically
+	 * sized array here. It's pretty unlikely we have more than 4 spi messages
+	 * in a transaction as we typically would only have 2, one to write a
+	 * register address, and another to read/write the register into an array
+	 */
+	if (num_msgs > CONFIG_SPI_RTIO_FALLBACK_MSGS) {
+		LOG_ERR("At most CONFIG_SPI_RTIO_FALLBACK_MSGS"
+			" submissions in a transaction are"
+			" allowed in the default handler");
+		rtio_iodev_sqe_err(txn_head, -ENOMEM);
+		return;
+	}
+
+	struct spi_buf tx_bufs[CONFIG_SPI_RTIO_FALLBACK_MSGS];
+	struct spi_buf rx_bufs[CONFIG_SPI_RTIO_FALLBACK_MSGS];
+	struct spi_buf_set tx_buf_set = {
+		.buffers = tx_bufs,
+		.count = num_msgs,
+	};
+	struct spi_buf_set rx_buf_set = {
+		.buffers = rx_bufs,
+		.count = num_msgs,
+	};
 
-		LOG_DBG("Preparing transfer: %p", txn_curr);
+	txn_curr = txn_head;
+
+	for (size_t i = 0 ; i < num_msgs ; i++) {
+		struct rtio_sqe *sqe = &txn_curr->sqe;
 
 		switch (sqe->op) {
 		case RTIO_OP_RX:
-			rx_buf.buf = sqe->rx.buf;
-			rx_buf.len = sqe->rx.buf_len;
-			rx_buf_set.count = 1;
+			rx_bufs[i].buf = sqe->rx.buf;
+			rx_bufs[i].len = sqe->rx.buf_len;
+			tx_bufs[i].buf = NULL;
+			tx_bufs[i].len = sqe->rx.buf_len;
 			break;
 		case RTIO_OP_TX:
-			tx_buf.buf = (uint8_t *)sqe->tx.buf;
-			tx_buf.len = sqe->tx.buf_len;
-			tx_buf_set.count = 1;
+			rx_bufs[i].buf = NULL;
+			rx_bufs[i].len = sqe->tx.buf_len;
+			tx_bufs[i].buf = (uint8_t *)sqe->tx.buf;
+			tx_bufs[i].len = sqe->tx.buf_len;
 			break;
 		case RTIO_OP_TINY_TX:
-			tx_buf.buf = (uint8_t *)sqe->tiny_tx.buf;
-			tx_buf.len = sqe->tiny_tx.buf_len;
-			tx_buf_set.count = 1;
+			rx_bufs[i].buf = NULL;
+			rx_bufs[i].len = sqe->tiny_tx.buf_len;
+			tx_bufs[i].buf = (uint8_t *)sqe->tiny_tx.buf;
+			tx_bufs[i].len = sqe->tiny_tx.buf_len;
 			break;
 		case RTIO_OP_TXRX:
-			rx_buf.buf = sqe->txrx.rx_buf;
-			rx_buf.len = sqe->txrx.buf_len;
-			tx_buf.buf = (uint8_t *)sqe->txrx.tx_buf;
-			tx_buf.len = sqe->txrx.buf_len;
-			rx_buf_set.count = 1;
-			tx_buf_set.count = 1;
+			rx_bufs[i].buf = sqe->txrx.rx_buf;
+			rx_bufs[i].len = sqe->txrx.buf_len;
+			tx_bufs[i].buf = (uint8_t *)sqe->txrx.tx_buf;
+			tx_bufs[i].len = sqe->txrx.buf_len;
 			break;
 		default:
-			LOG_ERR("Invalid op code %d for submission %p\n", sqe->op, (void *)sqe);
 			err = -EIO;
 			break;
 		}
 
-		if (!err) {
-			struct spi_buf_set *tx_buf_ptr = tx_buf_set.count > 0 ? &tx_buf_set : NULL;
-			struct spi_buf_set *rx_buf_ptr = rx_buf_set.count > 0 ? &rx_buf_set : NULL;
-
-			err = spi_transceive_dt(dt_spec, tx_buf_ptr, rx_buf_ptr);
+		txn_curr = rtio_txn_next(txn_curr);
+	}
 
-			/* NULL if this submission is not a transaction */
-			txn_curr = rtio_txn_next(txn_curr);
-		}
-	} while (err >= 0 && txn_curr != NULL);
+	if (err == 0) {
+		__ASSERT_NO_MSG(num_msgs > 0);
+		err = spi_transceive_dt(dt_spec, &tx_buf_set, &rx_buf_set);
+	}
 
-	if (err < 0) {
-		LOG_ERR("Transfer failed: %d", err);
+	if (err != 0) {
 		rtio_iodev_sqe_err(txn_head, err);
 	} else {
-		LOG_DBG("Transfer OK: %d", err);
-		rtio_iodev_sqe_ok(txn_head, err);
+		rtio_iodev_sqe_ok(txn_head, 0);
 	}
 }