Skip to contentSkip to conversation

usb: core: Warn if an URB's transfer_buffer is on stack

PATCH Superseded Florian Fainelli 3 participants 6 messages Apr 22, 2017 – Apr 24, 2017 USB

v1 of 2 Superseded Latest is v2 →

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 usb: core: Warn if an URB's transfer_buffer is on stack

We see a large number of fixes to several drivers to remove the usage of on-stack buffers feeding into USB transfer functions. Make it easier to spot the offenders by adding a warning in usb_start_wait_urb() for urb->transfer_buffer to be located on the stack.

---
 drivers/usb/core/message.c | 3 +++
 1 file changed, 3 insertions(+)
drivers/usb/core/message.c +3 -0
index 2184ef40a82a..abefddbe9243 100644
@@ -8,6 +8,7 @@
 #include <linux/pci.h>	/* for scatterlist macros */
 #include <linux/usb.h>
 #include <linux/module.h>
+#include <linux/sched/task_stack.h> /* for object_is_on_stack */
 #include <linux/slab.h>
 #include <linux/mm.h>
 #include <linux/timer.h>
@@ -50,6 +51,8 @@ static int usb_start_wait_urb(struct urb *urb, int timeout, int *actual_length)
 	unsigned long expire;
 	int retval;
 
+	WARN_ON(object_is_on_stack(urb->transfer_buffer));
+
 	init_completion(&ctx.done);
 	urb->context = &ctx;
 	urb->actual_length = 0;
Signature · 1 line
2.11.0
#1 lore raw
Quoted 23 lines · Florian Fainelli

We see a large number of fixes to several drivers to remove the usage of on-stack buffers feeding into USB transfer functions. Make it easier to spot the offenders by adding a warning in usb_start_wait_urb() for urb->transfer_buffer to be located on the stack.

---
 drivers/usb/core/message.c | 3 +++
 1 file changed, 3 insertions(+)
drivers/usb/core/message.c +1 -0
index 2184ef40a82a..abefddbe9243 100644
@@ -8,6 +8,7 @@
 #include <linux/pci.h>	/* for scatterlist macros */
 #include <linux/usb.h>
 #include <linux/module.h>
+#include <linux/sched/task_stack.h> /* for object_is_on_stack */
 #include <linux/slab.h>
 #include <linux/mm.h>
 #include <linux/timer.h>

drivers/usb/core/message.c static int usb_start_wait_urb(struct urb *urb, int timeout, int *actual_length) · line 58

Quoted 8 lines
drivers/usb/core/message.c +2 -0
@@ -50,6 +51,8 @@ static int usb_start_wait_urb(struct urb *urb, int timeout, int *actual_length)
 	unsigned long expire;
 	int retval;
 
+	WARN_ON(object_is_on_stack(urb->transfer_buffer));
+
 	init_completion(&ctx.done);
 	urb->context = &ctx;
 	urb->actual_length = 0;

Does this really help? I mean, don't we already get warnings when the host controller drivers try to map on-stack buffers for DMA?

The only difference is that one wouldn't have to read as far back into the stack trace. But that hardly seems like an important consideration.

Alan Stern

#2 lore raw
Quoted 37 lines · Alan Stern
Quoted 23 lines · Florian Fainelli

We see a large number of fixes to several drivers to remove the usage of on-stack buffers feeding into USB transfer functions. Make it easier to spot the offenders by adding a warning in usb_start_wait_urb() for urb->transfer_buffer to be located on the stack.

---
 drivers/usb/core/message.c | 3 +++
 1 file changed, 3 insertions(+)
drivers/usb/core/message.c +1 -0
index 2184ef40a82a..abefddbe9243 100644
@@ -8,6 +8,7 @@
 #include <linux/pci.h>	/* for scatterlist macros */
 #include <linux/usb.h>
 #include <linux/module.h>
+#include <linux/sched/task_stack.h> /* for object_is_on_stack */
 #include <linux/slab.h>
 #include <linux/mm.h>
 #include <linux/timer.h>

drivers/usb/core/message.c static int usb_start_wait_urb(struct urb *urb, int timeout, int *actual_length) · line 58

Quoted 8 lines
drivers/usb/core/message.c +2 -0
@@ -50,6 +51,8 @@ static int usb_start_wait_urb(struct urb *urb, int timeout, int *actual_length)
 	unsigned long expire;
 	int retval;
 
+	WARN_ON(object_is_on_stack(urb->transfer_buffer));
+
 	init_completion(&ctx.done);
 	urb->context = &ctx;
 	urb->actual_length = 0;

Does this really help? I mean, don't we already get warnings when the host controller drivers try to map on-stack buffers for DMA?

Quoted 3 lines

The only difference is that one wouldn't have to read as far back into the stack trace. But that hardly seems like an important consideration.

I don't think this will show up if you don't have the VMALLOC_STACKS option enabled (or whatever it is), so this warning is good to have. I didn't know we had that macro, as the USB stack has always required this due to some platforms needing it, just not the "mainstream" ones...

I'll take this for the next kernel (after 4.12-rc1) and see what shakes out in linux-next...

thanks,

greg k-h

#3 lore raw
Quoted 38 lines · Greg Kroah-Hartman
Quoted 37 lines · Alan Stern
Quoted 23 lines · Florian Fainelli

We see a large number of fixes to several drivers to remove the usage of on-stack buffers feeding into USB transfer functions. Make it easier to spot the offenders by adding a warning in usb_start_wait_urb() for urb->transfer_buffer to be located on the stack.

---
 drivers/usb/core/message.c | 3 +++
 1 file changed, 3 insertions(+)
drivers/usb/core/message.c +1 -0
index 2184ef40a82a..abefddbe9243 100644
@@ -8,6 +8,7 @@
 #include <linux/pci.h>	/* for scatterlist macros */
 #include <linux/usb.h>
 #include <linux/module.h>
+#include <linux/sched/task_stack.h> /* for object_is_on_stack */
 #include <linux/slab.h>
 #include <linux/mm.h>
 #include <linux/timer.h>

drivers/usb/core/message.c static int usb_start_wait_urb(struct urb *urb, int timeout, int *actual_length) · line 58

Quoted 8 lines
drivers/usb/core/message.c +2 -0
@@ -50,6 +51,8 @@ static int usb_start_wait_urb(struct urb *urb, int timeout, int *actual_length)
 	unsigned long expire;
 	int retval;
 
+	WARN_ON(object_is_on_stack(urb->transfer_buffer));
+
 	init_completion(&ctx.done);
 	urb->context = &ctx;
 	urb->actual_length = 0;

Does this really help? I mean, don't we already get warnings when the host controller drivers try to map on-stack buffers for DMA?

Quoted 8 lines
Quoted 3 lines

The only difference is that one wouldn't have to read as far back into the stack trace. But that hardly seems like an important consideration.

I don't think this will show up if you don't have the VMALLOC_STACKS option enabled (or whatever it is), so this warning is good to have. I didn't know we had that macro, as the USB stack has always required this due to some platforms needing it, just not the "mainstream" ones...

In that case, it would be better to move the warning to a central place where it will always get triggered, such as map_urb_for_dma(). As it is, the patch will only issue a warning for callers of usb_bulk_msg(), usb_interrupt_msg(), or usb_control_msg().

Alan Stern

Quoted 6 lines

I'll take this for the next kernel (after 4.12-rc1) and see what shakes out in linux-next...

thanks,

greg k-h

#4 lore raw
Quoted 49 lines · Alan Stern
Quoted 38 lines · Greg Kroah-Hartman
Quoted 37 lines · Alan Stern
Quoted 23 lines · Florian Fainelli

We see a large number of fixes to several drivers to remove the usage of on-stack buffers feeding into USB transfer functions. Make it easier to spot the offenders by adding a warning in usb_start_wait_urb() for urb->transfer_buffer to be located on the stack.

---
 drivers/usb/core/message.c | 3 +++
 1 file changed, 3 insertions(+)
drivers/usb/core/message.c +1 -0
index 2184ef40a82a..abefddbe9243 100644
@@ -8,6 +8,7 @@
 #include <linux/pci.h>	/* for scatterlist macros */
 #include <linux/usb.h>
 #include <linux/module.h>
+#include <linux/sched/task_stack.h> /* for object_is_on_stack */
 #include <linux/slab.h>
 #include <linux/mm.h>
 #include <linux/timer.h>

drivers/usb/core/message.c static int usb_start_wait_urb(struct urb *urb, int timeout, int *actual_length) · line 58

Quoted 8 lines
drivers/usb/core/message.c +2 -0
@@ -50,6 +51,8 @@ static int usb_start_wait_urb(struct urb *urb, int timeout, int *actual_length)
 	unsigned long expire;
 	int retval;
 
+	WARN_ON(object_is_on_stack(urb->transfer_buffer));
+
 	init_completion(&ctx.done);
 	urb->context = &ctx;
 	urb->actual_length = 0;

Does this really help? I mean, don't we already get warnings when the host controller drivers try to map on-stack buffers for DMA?

Quoted 8 lines
Quoted 3 lines

The only difference is that one wouldn't have to read as far back into the stack trace. But that hardly seems like an important consideration.

I don't think this will show up if you don't have the VMALLOC_STACKS option enabled (or whatever it is), so this warning is good to have. I didn't know we had that macro, as the USB stack has always required this due to some platforms needing it, just not the "mainstream" ones...

Quoted 4 lines

In that case, it would be better to move the warning to a central place where it will always get triggered, such as map_urb_for_dma(). As it is, the patch will only issue a warning for callers of usb_bulk_msg(), usb_interrupt_msg(), or usb_control_msg().

map_urb_for_dma() seems like a better location indeed, let me submit a v2 with that then.

Most HCDs implementation that define a map_urb_for_dma function pointer also call usb_hcd_map_urb_for_dma() (except for max3421-hcd.c) so would it be better to move the is_object_on_stack() check in usb_hcd_map_urb_for_dma() for consistency with the existing is_vmalloc_addr() check?

Thanks!

NB: initially, I was looking for a compile-time warning, and first there does not seem to be a warning available in GCC that could tell you if a variable is on the stack (why would there be?) and second, we'd have to annotate functions that require some of their arguments to come from the heap anyway. The amount of work needed to annotate and fix these build-time warnings versus looking at reports and fixing warnings end-ups being pretty much the same anyway.

Signature · 1 line
Florian
#5 lore raw
Quoted 5 lines · Florian Fainelli
Quoted 4 lines

In that case, it would be better to move the warning to a central place where it will always get triggered, such as map_urb_for_dma(). As it is, the patch will only issue a warning for callers of usb_bulk_msg(), usb_interrupt_msg(), or usb_control_msg().

Quoted 8 lines

map_urb_for_dma() seems like a better location indeed, let me submit a v2 with that then.

Most HCDs implementation that define a map_urb_for_dma function pointer also call usb_hcd_map_urb_for_dma() (except for max3421-hcd.c) so would it be better to move the is_object_on_stack() check in usb_hcd_map_urb_for_dma() for consistency with the existing is_vmalloc_addr() check?

Or maybe move the is_vmalloc_addr() check to map_urb_for_dma().

The only issue is that some weird HCD might sometimes not need to do the DMA mapping (it might use PIO for some types of transfer). In those cases we don't want to issue the warning. So perhaps usb_hcd_map_urb_for_dma() would be better for both checks.

Alan Stern