Skip to contentSkip to conversation

serial: qcom-geni: Fix port_ida handling for console and probe errors

PATCH In review Aniket Randive 4 participants 6 messages Aug 31 – Sep 1

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH v2 1/1 serial: qcom-geni: Fix port_ida handling for console and probe errors

The console port uses a fixed line number and is not allocated from port_ida. However, qcom_geni_serial_remove() unconditionally frees the line number for all ports, including the console port.

Skip ida_free() for console ports so that only IDs allocated from port_ida are returned.

Also release allocated IDs from the common probe error path. Currently, IDs are freed only on the wake-IRQ failure path, causing leaks when probe fails after a successful allocation. Move the ida_free() call to the common error path and guard it for console ports.

Route resources_init() failures through the common error path so the allocated line number is released correctly.

Allocate the port structure before calling ida_alloc_range() in get_port_from_line(). This prevents IDs from being left allocated when memory allocation fails.

---
Changes in v2:
- Route port->dev_data->resources_init() failures through the common error path so allocated line numbers are not leaked.
- Allocate the port structure before ida_alloc_range() in get_port_from_line() to prevent ID leaks on memory-allocation
failures.
---
Link to v1: https://patch.msgid.link/[email protected]

---
 drivers/tty/serial/qcom_geni_serial.c | 16 +++++++++-------
 1 file changed, 9 insertions(+), 7 deletions(-)
drivers/tty/serial/qcom_geni_serial.c +9 -7
index 3633723acef8..ceaf560fd8dd 100644
@@ -289,6 +289,10 @@ static struct qcom_geni_serial_port *get_port_from_line(int line, bool console,
 	} else {
 		int max_alias_num = of_alias_get_highest_id("serial");
 
+		port = devm_kzalloc(dev, sizeof(*port), GFP_KERNEL);
+		if (!port)
+			return ERR_PTR(-ENOMEM);
+
 		if (line < 0 || line >= nr_ports)
 			line = ida_alloc_range(&port_ida, max_alias_num + 1,
 					       nr_ports - 1, GFP_KERNEL);
@@ -299,10 +303,6 @@ static struct qcom_geni_serial_port *get_port_from_line(int line, bool console,
 		if (line < 0)
 			return ERR_PTR(-ENXIO);
 
-		port = devm_kzalloc(dev, sizeof(*port), GFP_KERNEL);
-		if (!port)
-			return ERR_PTR(-ENOMEM);
-
 		port->uport.iotype = UPIO_MEM;
 		port->uport.ops = &qcom_geni_uart_pops;
 		port->uport.flags = UPF_BOOT_AUTOCONF;
@@ -1897,7 +1897,7 @@ static int qcom_geni_serial_probe(struct platform_device *pdev)
 
 	ret = port->dev_data->resources_init(&port->se);
 	if (ret)
-		return ret;
+		goto error;
 
 	res = platform_get_resource(pdev, IORESOURCE_MEM, 0);
 	if (!res) {
@@ -1979,7 +1979,6 @@ static int qcom_geni_serial_probe(struct platform_device *pdev)
 						port->wakeup_irq);
 		if (ret) {
 			device_init_wakeup(&pdev->dev, false);
-			ida_free(&port_ida, uport->line);
 			goto error;
 		}
 	}
@@ -2003,6 +2002,8 @@ static int qcom_geni_serial_probe(struct platform_device *pdev)
 	return 0;
 
 error:
+	if (!data->console)
+		ida_free(&port_ida, uport->line);
 	if (port->rx_dma_addr) {
 		dma_unmap_single(pdev->dev.parent, port->rx_dma_addr,
 				 DMA_RX_BUF_SIZE, DMA_FROM_DEVICE);
@@ -2024,7 +2025,8 @@ static void qcom_geni_serial_remove(struct platform_device *pdev)
 	irq_work_sync(&port->tx_kick);
 	dev_pm_clear_wake_irq(&pdev->dev);
 	device_init_wakeup(&pdev->dev, false);
-	ida_free(&port_ida, uport->line);
+	if (!port->dev_data->console)
+		ida_free(&port_ida, uport->line);
 	uart_remove_one_port(drv, &port->uport);
 
 	if (port->rx_dma_addr) {

---
base-commit: a8406e6c0b793ce0788019683837c40855b55995
change-id: 20260828-master-3e806ba3492d

Best regards,

Signature · 1 line
Aniket Randive <[email protected]>
#1 lore raw
Thank you for your contribution! Sashiko AI review found 4 potential issue(s) to consider:

Thank you for your contribution! Sashiko AI review found 4 potential issue(s) to consider:

Pre-existing issues:
- [Critical] Manual detachment of the devres-managed PM domain list during probe error handling and device removal causes a double-free.
- [High] The driver incorrectly calls dma_unmap_single() with a DMA mapping error address in the probe error path.
- [High] The driver frees the port IDA before unregistering the UART port, exposing a race condition.
- [High] Unconditional probe failure if no serial aliases exist in the Device Tree.
--

commit db25e732789f738d8e62f2cfae1a09afe0942d97
Author: Aniket Randive <[email protected]>

serial: qcom-geni: Fix port_ida handling for console and probe errors

This commit ensures the console port line number is not allocated from or freed back to port_ida. It also fixes leaks by moving the ida_free() call to the common probe error path and allocating the port structure before allocating an ID.

Quoted 5 lines
drivers/tty/serial/qcom_geni_serial.c +0 -0
index 3633723acef8d..ceaf560fd8dda 100644

drivers/tty/serial/qcom_geni_serial.c static struct qcom_geni_serial_port *get_port_from_line(int line, bool console, · line 298

Quoted 10 lines
drivers/tty/serial/qcom_geni_serial.c +4 -0
@@ -289,6 +289,10 @@ static struct qcom_geni_serial_port *get_port_from_line(int line, bool console,
 	} else {
 		int max_alias_num = of_alias_get_highest_id("serial");
 
+		port = devm_kzalloc(dev, sizeof(*port), GFP_KERNEL);
+		if (!port)
+			return ERR_PTR(-ENOMEM);
+
 		if (line < 0 || line >= nr_ports)
 			line = ida_alloc_range(&port_ida, max_alias_num + 1,
 					       nr_ports - 1, GFP_KERNEL);

[Severity: High]
This is a pre-existing issue, but does this unconditionally fail the probe if there are no serial aliases in the Device Tree?

If of_alias_get_highest_id(serial) finds no aliases and returns -ENODEV (-19), the minimum value passed to ida_alloc_range() becomes -18. Since ida_alloc_range() performs a signedness bounds check ((int)min < 0), it immediately returns -ENOSPC, causing the probe to fail on systems without DT aliases.

[ ... ]

drivers/tty/serial/qcom_geni_serial.c static int qcom_geni_serial_probe(struct platform_device *pdev) · line 2009

Quoted 9 lines
drivers/tty/serial/qcom_geni_serial.c +2 -0
@@ -2003,6 +2002,8 @@ static int qcom_geni_serial_probe(struct platform_device *pdev)
 	return 0;
 
 error:
+	if (!data->console)
+		ida_free(&port_ida, uport->line);
 	if (port->rx_dma_addr) {
 		dma_unmap_single(pdev->dev.parent, port->rx_dma_addr,
 				 DMA_RX_BUF_SIZE, DMA_FROM_DEVICE);

[Severity: High]
This is a pre-existing issue, but does this erroneously unmap an invalid DMA address if dma_map_single() fails?

If dma_map_single() fails during probe, it returns DMA_MAPPING_ERROR (usually ~0ULL) which is non-zero. The failure path directly jumps to the error label without clearing port->rx_dma_addr, so this check considers it valid and calls dma_unmap_single() on the error address, potentially causing IOMMU faults or memory corruption.

[Severity: Critical]
This is a pre-existing issue, but does the manual detachment of the devres-managed PM domain list here in the probe error path cause a double-free?

Just below in the error path, dev_pm_domain_detach_list(port->se.pd_list) is called. Since pd_list was allocated and attached via devm_pm_domain_attach_list(), the devres subsystem will automatically free and detach it on probe failure. Manually calling dev_pm_domain_detach_list() results in kfree being called twice on the same pointer, leading to heap corruption and a kernel panic.

drivers/tty/serial/qcom_geni_serial.c static void qcom_geni_serial_remove(struct platform_device *pdev) · line 2030

Quoted 8 lines
drivers/tty/serial/qcom_geni_serial.c +2 -1
@@ -2024,7 +2025,8 @@ static void qcom_geni_serial_remove(struct platform_device *pdev)
 	irq_work_sync(&port->tx_kick);
 	dev_pm_clear_wake_irq(&pdev->dev);
 	device_init_wakeup(&pdev->dev, false);
-	ida_free(&port_ida, uport->line);
+	if (!port->dev_data->console)
+		ida_free(&port_ida, uport->line);
 	uart_remove_one_port(drv, &port->uport);

[Severity: High]
This is a pre-existing issue, but does freeing the port IDA before unregistering the UART port expose a TOCTOU race for the line ID?

If qcom_geni_serial_remove() frees the line IDA while the UART port is still registered in the TTY layer, a concurrent probe of another UART device could allocate the exact same line ID and call uart_add_one_port(), resulting in two active driver instances sharing the same TTY port line in the serial core.

[Severity: Critical]
This is a pre-existing issue, but does the manual detachment of the devres-managed PM domain list in this remove function cause a double-free?

Further down in qcom_geni_serial_remove(), the driver calls dev_pm_domain_detach_list(port->se.pd_list). Since pd_list is devm-managed, the devres subsystem will automatically handle its cleanup when the device is removed. Manually detaching it here causes a double-free of the pd_list memory.

Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=1
#2 lore raw
In reply to Aniket Randive
Quoted 19 lines · Aniket Randive

The console port uses a fixed line number and is not allocated from port_ida. However, qcom_geni_serial_remove() unconditionally frees the line number for all ports, including the console port.

Skip ida_free() for console ports so that only IDs allocated from port_ida are returned.

Also release allocated IDs from the common probe error path. Currently, IDs are freed only on the wake-IRQ failure path, causing leaks when probe fails after a successful allocation. Move the ida_free() call to the common error path and guard it for console ports.

Route resources_init() failures through the common error path so the allocated line number is released correctly.

Allocate the port structure before calling ida_alloc_range() in get_port_from_line(). This prevents IDs from being left allocated when memory allocation fails.

Quoted 5 lines

---
Changes in v2:
- Route port->dev_data->resources_init() failures through the common error path so allocated line numbers are not leaked.

I asked GPT what it thinks again and this time it found that this will cause a double calling of dev_pm_domain_detach_list(), which will mess up the refcounting..

I think the fix here is to drop the manual detaches since the common helper uses a devres-managed version of attach.

Konrad

#3 lore raw
Quoted 26 lines · Konrad Dybcio
Quoted 19 lines · Aniket Randive

The console port uses a fixed line number and is not allocated from port_ida. However, qcom_geni_serial_remove() unconditionally frees the line number for all ports, including the console port.

Skip ida_free() for console ports so that only IDs allocated from port_ida are returned.

Also release allocated IDs from the common probe error path. Currently, IDs are freed only on the wake-IRQ failure path, causing leaks when probe fails after a successful allocation. Move the ida_free() call to the common error path and guard it for console ports.

Route resources_init() failures through the common error path so the allocated line number is released correctly.

Allocate the port structure before calling ida_alloc_range() in get_port_from_line(). This prevents IDs from being left allocated when memory allocation fails.

Quoted 5 lines

---
Changes in v2:
- Route port->dev_data->resources_init() failures through the common error path so allocated line numbers are not leaked.

Quoted 7 lines

I asked GPT what it thinks again and this time it found that this will cause a double calling of dev_pm_domain_detach_list(), which will mess up the refcounting..

I think the fix here is to drop the manual detaches since the common helper uses a devres-managed version of attach.

Thanks for catching this.

Looking closer, this appears to be a separate issue from the port_ida fixes in this patch. The dev_pm_domain_detach_list() problem already exists in the current error path and is not introduced by this change.

Since this involves power-domain cleanup rather than port_ida handling, I'd prefer to address it in a separate follow-up patch to keep this fix focused and easier to review.

Is it OK to keep this patch limited to the port_ida fixes and handle the power-domain cleanup separately?

Thanks,
Aniket

Quoted 1 line

Konrad

Reviewed-by
#4 lore raw

Hi Aniket,

Quoted 37 lines · Aniket RANDIVE
Quoted 26 lines · Konrad Dybcio
Quoted 19 lines · Aniket Randive

The console port uses a fixed line number and is not allocated from port_ida. However, qcom_geni_serial_remove() unconditionally frees the line number for all ports, including the console port.

Skip ida_free() for console ports so that only IDs allocated from port_ida are returned.

Also release allocated IDs from the common probe error path. Currently, IDs are freed only on the wake-IRQ failure path, causing leaks when probe fails after a successful allocation. Move the ida_free() call to the common error path and guard it for console ports.

Route resources_init() failures through the common error path so the allocated line number is released correctly.

Allocate the port structure before calling ida_alloc_range() in get_port_from_line(). This prevents IDs from being left allocated when memory allocation fails.

Quoted 5 lines

---
Changes in v2:
- Route port->dev_data->resources_init() failures through the common error path so allocated line numbers are not leaked.

Quoted 7 lines

I asked GPT what it thinks again and this time it found that this will cause a double calling of dev_pm_domain_detach_list(), which will mess up the refcounting..

I think the fix here is to drop the manual detaches since the common helper uses a devres-managed version of attach.

Quoted 7 lines

Thanks for catching this.

Looking closer, this appears to be a separate issue from the port_ida fixes in this patch. The dev_pm_domain_detach_list() problem already exists in the current error path and is not introduced by this change.

Since this involves power-domain cleanup rather than port_ida handling, I'd prefer to address it in a separate follow-up patch to keep this fix focused and easier to review.

Is it OK to keep this patch limited to the port_ida fixes and handle the power-domain cleanup separately?

That should be separate patch.

For this patch.

Thanks,

Praveen Talari

Quoted 7 lines

Thanks,
Aniket

Quoted 1 line

Konrad

Reviewed-by
#5 lore raw
In reply to Aniket RANDIVE
Quoted 7 lines · Aniket RANDIVE
Quoted 4 lines · Konrad Dybcio
Quoted 3 lines · Aniket Randive

The console port uses a fixed line number and is not allocated from port_ida. However, qcom_geni_serial_remove() unconditionally frees the line number for all ports, including the console port.

[...]

Quoted 7 lines

Thanks for catching this.

Looking closer, this appears to be a separate issue from the port_ida fixes in this patch. The dev_pm_domain_detach_list() problem already exists in the current error path and is not introduced by this change.

Since this involves power-domain cleanup rather than port_ida handling, I'd prefer to address it in a separate follow-up patch to keep this fix focused and easier to review.

Is it OK to keep this patch limited to the port_ida fixes and handle the power-domain cleanup separately?

Yes

Konrad