qdl: add QUD backend so on Windows flashing via the official QDLoader 9008 driver works

auto.c

@@ -0,0 +1,165 @@
+// SPDX-License-Identifier: BSD-3-Clause
+/*
+ * Copyright (c) Qualcomm Technologies, Inc. and/or its subsidiaries.
+ *
+ * QDL_DEVICE_AUTO: meta-backend that defers transport selection to the
+ * wait loop. Each 250 ms tick it runs both the libusb open attempt and
+ * (on Windows) a QUD SetupAPI probe; whichever first reaches an EDL
+ * device wins, and its concrete qdl_device is bound as the inner. All
+ * subsequent qdl_read/write/close calls on the outer forward to the
+ * inner.
+ */
+
+#include <errno.h>
+#include <stdbool.h>
+#include <stdlib.h>
+#include <string.h>
+#include <unistd.h>
+
+#include "qdl.h"
+
+struct qdl_device_auto {
+	struct qdl_device base;
+	struct qdl_device *inner;
+	long pending_chunk_size;
+	bool chunk_size_set;
+};
+
+static struct qdl_device_auto *to_auto(struct qdl_device *qdl)
+{
+	return container_of(qdl, struct qdl_device_auto, base);
+}
+
+static void auto_bind_inner(struct qdl_device_auto *wrap, struct qdl_device *inner)
+{
+	wrap->inner = inner;
+	wrap->base.max_payload_size = inner->max_payload_size;
+	if (wrap->chunk_size_set)
+		inner->set_out_chunk_size(inner, wrap->pending_chunk_size);
+}
+
+static int auto_open(struct qdl_device *qdl, const char *serial)
+{
+	struct qdl_device_auto *wrap = to_auto(qdl);
+	struct qdl_device *usb_dev;
+#ifdef _WIN32
+	struct qdl_device *qud_dev;
+	int qud_count;
+#endif
+	int visible_prev = -1;
+	int visible;
+	int ret;
+
+	usb_dev = usb_init();
+	if (!usb_dev)
+		return -1;
+
+#ifdef _WIN32
+	qud_dev = qud_init();
+	if (!qud_dev) {
+		qdl_deinit(usb_dev);
+		return -1;
+	}
+#endif
+
+	for (;;) {
+		ret = try_usb_open(usb_dev, serial, &visible);
+		if (ret == 0) {
+#ifdef _WIN32
+			qdl_deinit(qud_dev);
+#endif
+			auto_bind_inner(wrap, usb_dev);
+			return 0;
+		}
+		if (ret == -EIO)
+			goto fail;
+
+#ifdef _WIN32
+		qud_count = qud_probe_present();
+		if (qud_count > 0 || ret == -EBUSY) {
+			if (qud_dev->open(qud_dev, serial) == 0) {
+				qdl_deinit(usb_dev);
+				auto_bind_inner(wrap, qud_dev);
+				return 0;
+			}
+		}
+		visible += qud_count;
+#endif
+
+		if (visible != visible_prev) {
+			if (visible == 0)
+				ux_info("Waiting for EDL device\n");
+			else if (serial)
+				ux_info("%d EDL device(s) visible, none match serial \"%s\"\n",
+					visible, serial);
+			else
+				ux_info("%d EDL device(s) visible, none could be opened\n",
+					visible);
+			visible_prev = visible;
+		}
+
+		usleep(250000);
+	}
+
+fail:
+	qdl_deinit(usb_dev);
+#ifdef _WIN32
+	qdl_deinit(qud_dev);
+#endif
+	return -1;
+}
+
+static int auto_read(struct qdl_device *qdl, void *buf, size_t len, unsigned int timeout)
+{
+	struct qdl_device *inner = to_auto(qdl)->inner;
+
+	return inner->read(inner, buf, len, timeout);
+}
+
+static int auto_write(struct qdl_device *qdl, const void *buf, size_t len, unsigned int timeout)
+{
+	struct qdl_device *inner = to_auto(qdl)->inner;
+
+	return inner->write(inner, buf, len, timeout);
+}
+
+static void auto_close(struct qdl_device *qdl)
+{
+	struct qdl_device_auto *wrap = to_auto(qdl);
+
+	if (!wrap->inner)
+		return;
+	wrap->inner->close(wrap->inner);
+	qdl_deinit(wrap->inner);
+	wrap->inner = NULL;
+}
+
+static void auto_set_out_chunk_size(struct qdl_device *qdl, long size)
+{
+	struct qdl_device_auto *wrap = to_auto(qdl);
+
+	if (wrap->inner) {
+		wrap->inner->set_out_chunk_size(wrap->inner, size);
+		return;
+	}
+	wrap->pending_chunk_size = size;
+	wrap->chunk_size_set = true;
+}
+
+struct qdl_device *auto_init(void)
+{
+	struct qdl_device_auto *wrap = calloc(1, sizeof(*wrap));
+
+	if (!wrap)
+		return NULL;
+
+	wrap->base.dev_type = QDL_DEVICE_AUTO;
+	wrap->base.open = auto_open;
+	wrap->base.read = auto_read;
+	wrap->base.write = auto_write;
+	wrap->base.close = auto_close;
+	wrap->base.set_out_chunk_size = auto_set_out_chunk_size;
+	wrap->base.max_payload_size = 1048576;
+
+	return &wrap->base;
+}

firehose.c

@@ -179,15 +179,75 @@ static int firehose_read(struct qdl_device *qdl, int timeout_ms,
 
 		ux_debug("FIREHOSE READ: %s\n", buf);
 
-		node = firehose_response_parse(buf, n, &error);
-		if (!node)
-			return error;
+		/*
+		 * On stream-oriented transports (Windows COM port via the
+		 * QDLoader driver, virtio-console, ...) a single read can
+		 * deliver multiple back-to-back Firehose responses
+		 * concatenated, since the driver doesn't preserve USB bulk-
+		 * transfer boundaries. Walk the buffer using the "<?xml" ...
+		 * "</data>" envelope to bound each message; the closing tag
+		 * is what really delimits the document so that any rawmode
+		 * binary payload that arrives spliced onto the same read
+		 * doesn't end up fed into libxml2 as if it were XML.
+		 *
+		 * libusb preserves transfer boundaries, so on that path each
+		 * read still contains exactly one document and the loop runs
+		 * once.
+		 */
+		char *cursor = buf;
+		char *bufend = buf + n;
 
-		ret = response_parser(node, data, &rawmode);
-		xmlFreeDoc(node->doc);
+		while (cursor < bufend) {
+			char *start = strstr(cursor, "<?xml");
+			char *xml_end;
+			size_t chunk;
 
-		if (ret >= 0)
-			resp = ret;
+			if (!start)
+				break;
+
+			/*
+			 * Bound the XML on the closing </data> tag. If it's
+			 * missing the message was either truncated or doesn't
+			 * fit the schema we know how to parse; hand the rest
+			 * of the buffer to libxml2 and let it error out
+			 * gracefully.
+			 */
+			xml_end = strstr(start, "</data>");
+			if (xml_end) {
+				xml_end += sizeof("</data>") - 1;
+				chunk = (size_t)(xml_end - start);
+			} else {
+				chunk = (size_t)(bufend - start);
+			}
+
+			node = firehose_response_parse(start, chunk, &error);
+			if (!node)
+				return error;
+
+			ret = response_parser(node, data, &rawmode);
+			xmlFreeDoc(node->doc);
+
+			if (ret >= 0)
+				resp = ret;
+
+			cursor = start + chunk;
+
+			/*
+			 * The response we just parsed told the host to switch
+			 * to raw mode (e.g. the ACK that precedes the binary
+			 * sectors of a <read>). On a stream transport the
+			 * first chunk of that binary payload can have arrived
+			 * tacked onto this same read. Push it back so the
+			 * next qdl_read() picks it up before the transport
+			 * is touched again.
+			 */
+			if (rawmode) {
+				if (cursor < bufend)
+					qdl_push_back(qdl, cursor,
+						      (size_t)(bufend - cursor));
+				break;
+			}
+		}
 
 		if (rawmode)
 			break;
@@ -699,31 +759,47 @@ static int firehose_issue_read(struct qdl_device *qdl, struct firehose_op *read_
 
 	left = read_op->num_sectors;
 	while (left > 0) {
-		chunk_size = MIN(qdl->max_payload_size / sector_size, left);
+		size_t want;
+		size_t got;
 
-		n = qdl_read(qdl, buf, chunk_size * sector_size, 30000);
-		if (n < 0) {
-			ux_err("failed to read sector data\n");
-			ret = -1;
-			goto out;
-		}
+		chunk_size = MIN(qdl->max_payload_size / sector_size, left);
+		want = chunk_size * sector_size;
 
-		if ((size_t)n != chunk_size * sector_size) {
-			ux_err("failed to read full sector\n");
-			ret = -1;
-			goto out;
+		/*
+		 * Accumulate the chunk across qdl_read() calls. libusb usually
+		 * delivers an entire bulk transfer in one shot, but stream
+		 * transports (QUD's Windows COM port, virtio-console, ...) can
+		 * fragment it - including the rawmode tail that firehose_read()
+		 * pushed back from the same buffer as the ACK response.
+		 */
+		got = 0;
+		while (got < want) {
+			n = qdl_read(qdl, (char *)buf + got, want - got, 30000);
+			if (n < 0) {
+				ux_err("failed to read sector data\n");
+				ret = -1;
+				goto out;
+			}
+			if (n == 0) {
+				ux_err("unexpected EOF while reading sector data\n");
+				ret = -1;
+				goto out;
+			}
+			got += (size_t)n;
 		}
 
 		if (out_buf) {
-			if ((size_t)n > out_len - out_offset)
-				n = out_len - out_offset;
+			size_t copy = want;
+
+			if (copy > out_len - out_offset)
+				copy = out_len - out_offset;
 
-			memcpy(out_buf + out_offset, buf, n);
-			out_offset += n;
+			memcpy((char *)out_buf + out_offset, buf, copy);
+			out_offset += copy;
 		} else {
-			n = write(fd, buf, n);
+			n = write(fd, buf, want);
 
-			if (n < 0 || (size_t)n != chunk_size * sector_size) {
+			if (n < 0 || (size_t)n != want) {
 				ux_err("failed to write sector data\n");
 				ret = -1;
 				goto out;