Skip to contentSkip to conversation

mux: Convert mux_control_ops to a flex array member in mux_chip

PATCH RESEND Superseded Thorsten Blum 3 participants 11 messages Mar 2, 2025 – Apr 29, 2025 Multiplexer

lore
Share & tools
lore
#0 lore raw

PATCH RESEND 1/1 mux: Convert mux_control_ops to a flex array member in mux_chip

Convert mux_control_ops to a flexible array member at the end of the mux_chip struct and add the __counted_by() compiler attribute to improve access bounds-checking via CONFIG_UBSAN_BOUNDS and CONFIG_FORTIFY_SOURCE.

Use struct_size() to calculate the number of bytes to allocate for a new mux chip and to remove the following Coccinelle/coccicheck warning:

WARNING: Use struct_size

Use size_add() to safely add any extra bytes.

Compile-tested only.

---
 drivers/mux/core.c         | 7 +++----
 include/linux/mux/driver.h | 4 ++--
 2 files changed, 5 insertions(+), 6 deletions(-)
drivers/mux/core.c +3 -4
index 02be4ba37257..a3840fe0995f 100644
@@ -98,13 +98,12 @@ struct mux_chip *mux_chip_alloc(struct device *dev,
 	if (WARN_ON(!dev || !controllers))
 		return ERR_PTR(-EINVAL);
 
-	mux_chip = kzalloc(sizeof(*mux_chip) +
-			   controllers * sizeof(*mux_chip->mux) +
-			   sizeof_priv, GFP_KERNEL);
+	mux_chip = kzalloc(size_add(struct_size(mux_chip, mux, controllers),
+				    sizeof_priv),
+			   GFP_KERNEL);
 	if (!mux_chip)
 		return ERR_PTR(-ENOMEM);
 
-	mux_chip->mux = (struct mux_control *)(mux_chip + 1);
 	mux_chip->dev.class = &mux_class;
 	mux_chip->dev.type = &mux_type;
 	mux_chip->dev.parent = dev;
include/linux/mux/driver.h +2 -2
index 18824064f8c0..e58e59354e23 100644
@@ -56,18 +56,18 @@ struct mux_control {
 /**
  * struct mux_chip -	Represents a chip holding mux controllers.
  * @controllers:	Number of mux controllers handled by the chip.
- * @mux:		Array of mux controllers that are handled.
  * @dev:		Device structure.
  * @id:			Used to identify the device internally.
  * @ops:		Mux controller operations.
+ * @mux:		Array of mux controllers that are handled.
  */
 struct mux_chip {
 	unsigned int controllers;
-	struct mux_control *mux;
 	struct device dev;
 	int id;
 
 	const struct mux_control_ops *ops;
+	struct mux_control mux[] __counted_by(controllers);
 };
 
 #define to_mux_chip(x) container_of((x), struct mux_chip, dev)
Signature · 1 line
2.48.1
#1 lore raw
Quoted 5 lines · Thorsten Blum

Convert mux_control_ops to a flexible array member at the end of the mux_chip struct and add the __counted_by() compiler attribute to improve access bounds-checking via CONFIG_UBSAN_BOUNDS and CONFIG_FORTIFY_SOURCE.

Quoted 8 lines

Use struct_size() to calculate the number of bytes to allocate for a new mux chip and to remove the following Coccinelle/coccicheck warning:

WARNING: Use struct_size

Use size_add() to safely add any extra bytes.

Compile-tested only.

I believe this will fail at runtime. Note that sizeof_priv follows the allocation, so at the very least, you'd need to update:

static inline void *mux_chip_priv(struct mux_chip *mux_chip)

{
        return &mux_chip->mux[mux_chip->controllers];
}

to not use the mux array itself as a location reference because it will be seen as out of bounds.

To deal with this, the location will need to be calculated using mux_chip as the base, not mux_chip->mux as the base. For example, see commit 838ae9f45c4e ("nouveau/gsp: Avoid addressing beyond end of rpc->entries")

-Kees

Quoted 58 lines
---
 drivers/mux/core.c         | 7 +++----
 include/linux/mux/driver.h | 4 ++--
 2 files changed, 5 insertions(+), 6 deletions(-)
drivers/mux/core.c +3 -4
index 02be4ba37257..a3840fe0995f 100644
@@ -98,13 +98,12 @@ struct mux_chip *mux_chip_alloc(struct device *dev,
 	if (WARN_ON(!dev || !controllers))
 		return ERR_PTR(-EINVAL);
 
-	mux_chip = kzalloc(sizeof(*mux_chip) +
-			   controllers * sizeof(*mux_chip->mux) +
-			   sizeof_priv, GFP_KERNEL);
+	mux_chip = kzalloc(size_add(struct_size(mux_chip, mux, controllers),
+				    sizeof_priv),
+			   GFP_KERNEL);
 	if (!mux_chip)
 		return ERR_PTR(-ENOMEM);
 
-	mux_chip->mux = (struct mux_control *)(mux_chip + 1);
 	mux_chip->dev.class = &mux_class;
 	mux_chip->dev.type = &mux_type;
 	mux_chip->dev.parent = dev;
include/linux/mux/driver.h +2 -2
index 18824064f8c0..e58e59354e23 100644
@@ -56,18 +56,18 @@ struct mux_control {
 /**
  * struct mux_chip -	Represents a chip holding mux controllers.
  * @controllers:	Number of mux controllers handled by the chip.
- * @mux:		Array of mux controllers that are handled.
  * @dev:		Device structure.
  * @id:			Used to identify the device internally.
  * @ops:		Mux controller operations.
+ * @mux:		Array of mux controllers that are handled.
  */
 struct mux_chip {
 	unsigned int controllers;
-	struct mux_control *mux;
 	struct device dev;
 	int id;
 
 	const struct mux_control_ops *ops;
+	struct mux_control mux[] __counted_by(controllers);
 };
 
 #define to_mux_chip(x) container_of((x), struct mux_chip, dev)
Signature · 1 line
2.48.1
Signature · 1 line
Kees Cook
#2 lore raw
Quoted 18 lines · Kees Cook
Quoted 5 lines · Thorsten Blum

Convert mux_control_ops to a flexible array member at the end of the mux_chip struct and add the __counted_by() compiler attribute to improve access bounds-checking via CONFIG_UBSAN_BOUNDS and CONFIG_FORTIFY_SOURCE.

Quoted 8 lines

Use struct_size() to calculate the number of bytes to allocate for a new mux chip and to remove the following Coccinelle/coccicheck warning:

WARNING: Use struct_size

Use size_add() to safely add any extra bytes.

Compile-tested only.

I believe this will fail at runtime. Note that sizeof_priv follows the allocation, so at the very least, you'd need to update:

Quoted 7 lines

static inline void *mux_chip_priv(struct mux_chip *mux_chip)

{
      return &mux_chip->mux[mux_chip->controllers];
}

to not use the mux array itself as a location reference because it will be seen as out of bounds.

Getting the address doesn't fail at runtime, does it? For this example it works, but maybe I'm missing some compiler flag?

https://godbolt.org/z/qTEdqn9WW

Thanks,
Thorsten

#3 lore raw
Quoted 24 lines · Thorsten Blum
Quoted 23 lines · Kees Cook
Quoted 5 lines · Thorsten Blum

Convert mux_control_ops to a flexible array member at the end of the mux_chip struct and add the __counted_by() compiler attribute to improve access bounds-checking via CONFIG_UBSAN_BOUNDS and CONFIG_FORTIFY_SOURCE.

Quoted 8 lines

Use struct_size() to calculate the number of bytes to allocate for a new mux chip and to remove the following Coccinelle/coccicheck warning:

WARNING: Use struct_size

Use size_add() to safely add any extra bytes.

Compile-tested only.

I believe this will fail at runtime. Note that sizeof_priv follows the allocation, so at the very least, you'd need to update:

static inline void *mux_chip_priv(struct mux_chip *mux_chip)

{
      return &mux_chip->mux[mux_chip->controllers];
}
Quoted 7 lines
Quoted 2 lines

to not use the mux array itself as a location reference because it will be seen as out of bounds.

Getting the address doesn't fail at runtime, does it? For this example it works, but maybe I'm missing some compiler flag?

https://godbolt.org/z/qTEdqn9WW

Uhn. I can't explain that. :( We've seen this calculation get tripped in the real world, though:

https://git.kernel.org/linus/a26a5107bc52

But yeah, when I build local test cases, grabbing an integral trips it, but taking an address does not, as your godbolt shows. This makes no sense to me at all.

Here's my version, doing a direct comparison of int to *(int *) ...
https://godbolt.org/z/e1bKGz739

#include <stdlib.h>
#include <stdio.h>

struct foo {
    int count;
    int array[] __attribute__((__counted_by__(count)));
};
int main(int argc, char *argv[]) {
    int num_elems = 2 + argc;

    struct foo *p = malloc(sizeof(*p) + num_elems * sizeof(*p->array) + sizeof(int));
    p->count = num_elems;

    // this correctly trips sanitizer:
    int val = p->array[num_elems];
    printf("%d\n", val);

    // this does not?!
    int *valp = &p->array[num_elems];
    printf("%p %d\n", valp, *valp);

    return 0;
}

Qing and Bill, are you able to explain this? If I set p->count = 0, 1, or 2, this trips. Is this somehow an off-by-one error in both GCC and Clang?

Signature · 1 line
Kees Cook
#4 lore raw
Quoted 28 lines · Kees Cook

#include <stdlib.h>
#include <stdio.h>

struct foo {
   int count;
   int array[] __attribute__((__counted_by__(count)));
};
int main(int argc, char *argv[]) {
   int num_elems = 2 + argc;

   struct foo *p = malloc(sizeof(*p) + num_elems * sizeof(*p->array) + sizeof(int));
   p->count = num_elems;

   // this correctly trips sanitizer:
   int val = p->array[num_elems];
   printf("%d\n", val);

   // this does not?!
   int *valp = &p->array[num_elems];
   printf("%p %d\n", valp, *valp);

   return 0;
}
#5 lore raw
In reply to Kees Cook
Quoted 74 lines · Kees Cook
Quoted 24 lines · Thorsten Blum
Quoted 23 lines · Kees Cook
Quoted 5 lines · Thorsten Blum

Convert mux_control_ops to a flexible array member at the end of the mux_chip struct and add the __counted_by() compiler attribute to improve access bounds-checking via CONFIG_UBSAN_BOUNDS and CONFIG_FORTIFY_SOURCE.

Quoted 8 lines

Use struct_size() to calculate the number of bytes to allocate for a new mux chip and to remove the following Coccinelle/coccicheck warning:

WARNING: Use struct_size

Use size_add() to safely add any extra bytes.

Compile-tested only.

I believe this will fail at runtime. Note that sizeof_priv follows the allocation, so at the very least, you'd need to update:

static inline void *mux_chip_priv(struct mux_chip *mux_chip)

{
     return &mux_chip->mux[mux_chip->controllers];
}
Quoted 7 lines
Quoted 2 lines

to not use the mux array itself as a location reference because it will be seen as out of bounds.

Getting the address doesn't fail at runtime, does it? For this example it works, but maybe I'm missing some compiler flag?

https://godbolt.org/z/qTEdqn9WW

Uhn. I can't explain that. :( We've seen this calculation get tripped in the real world, though:

https://git.kernel.org/linus/a26a5107bc52

But yeah, when I build local test cases, grabbing an integral trips it, but taking an address does not, as your godbolt shows. This makes no sense to me at all.

Here's my version, doing a direct comparison of int to *(int *) ...
https://godbolt.org/z/e1bKGz739

#include <stdlib.h>
#include <stdio.h>

struct foo {
   int count;
   int array[] __attribute__((__counted_by__(count)));
};
int main(int argc, char *argv[]) {
   int num_elems = 2 + argc;

   struct foo *p = malloc(sizeof(*p) + num_elems * sizeof(*p->array) + sizeof(int));
   p->count = num_elems;

   // this correctly trips sanitizer:
   int val = p->array[num_elems];
   printf("%d\n", val);

   // this does not?!
   int *valp = &p->array[num_elems];
   printf("%p %d\n", valp, *valp);

   return 0;
}
Quoted 2 lines

Qing and Bill, are you able to explain this? If I set p->count = 0, 1, or 2, this trips. Is this somehow an off-by-one error in both GCC and Clang?

This does look like a bug in the compiler, could you please file a bug against GCC on this?

thanks.

Qing

Quoted 3 lines
Signature · 2 lines
-- 
Kees Cook
#7 lore raw
In reply to Kees Cook
Quoted 23 lines · Kees Cook
Quoted 5 lines · Thorsten Blum

Convert mux_control_ops to a flexible array member at the end of the mux_chip struct and add the __counted_by() compiler attribute to improve access bounds-checking via CONFIG_UBSAN_BOUNDS and CONFIG_FORTIFY_SOURCE.

Quoted 8 lines

Use struct_size() to calculate the number of bytes to allocate for a new mux chip and to remove the following Coccinelle/coccicheck warning:

WARNING: Use struct_size

Use size_add() to safely add any extra bytes.

Compile-tested only.

I believe this will fail at runtime. Note that sizeof_priv follows the allocation, so at the very least, you'd need to update:

static inline void *mux_chip_priv(struct mux_chip *mux_chip)

{
       return &mux_chip->mux[mux_chip->controllers];
}
Quoted 6 lines

to not use the mux array itself as a location reference because it will be seen as out of bounds.

To deal with this, the location will need to be calculated using mux_chip as the base, not mux_chip->mux as the base. For example, see commit 838ae9f45c4e ("nouveau/gsp: Avoid addressing beyond end of rpc->entries")

Since this should work and is well-defined C code according to [1][2], could you give this patch another look or should I still change it and submit a v2?

Thanks,
Thorsten

[1] https://gcc.gnu.org/bugzilla/show_bug.cgi?id=119132
[2] https://github.com/llvm/llvm-project/issues/129951

#8 lore raw
Quoted 27 lines · Thorsten Blum
Quoted 26 lines · Kees Cook
Quoted 5 lines · Thorsten Blum

Convert mux_control_ops to a flexible array member at the end of the mux_chip struct and add the __counted_by() compiler attribute to improve access bounds-checking via CONFIG_UBSAN_BOUNDS and CONFIG_FORTIFY_SOURCE.

Quoted 8 lines

Use struct_size() to calculate the number of bytes to allocate for a new mux chip and to remove the following Coccinelle/coccicheck warning:

WARNING: Use struct_size

Use size_add() to safely add any extra bytes.

Compile-tested only.

I believe this will fail at runtime. Note that sizeof_priv follows the allocation, so at the very least, you'd need to update:

static inline void *mux_chip_priv(struct mux_chip *mux_chip)

{
       return &mux_chip->mux[mux_chip->controllers];
}

to not use the mux array itself as a location reference because it will be seen as out of bounds.

Quoted 7 lines
Quoted 3 lines

To deal with this, the location will need to be calculated using mux_chip as the base, not mux_chip->mux as the base. For example, see commit 838ae9f45c4e ("nouveau/gsp: Avoid addressing beyond end of rpc->entries")

Since this should work and is well-defined C code according to [1][2], could you give this patch another look or should I still change it and submit a v2?

I think C is wrong here, but it seems it will continue to accidentally work. I personally would like a v3 that fixes this, but I leave it to Peter who is the MUX maintainer...

-Kees

Signature · 1 line
Kees Cook
#9 lore raw

Hi Peter,

Quoted 32 lines · Kees Cook
Quoted 31 lines · Thorsten Blum
Quoted 29 lines · Kees Cook
Quoted 5 lines · Thorsten Blum

Convert mux_control_ops to a flexible array member at the end of the mux_chip struct and add the __counted_by() compiler attribute to improve access bounds-checking via CONFIG_UBSAN_BOUNDS and CONFIG_FORTIFY_SOURCE.

Quoted 8 lines

Use struct_size() to calculate the number of bytes to allocate for a new mux chip and to remove the following Coccinelle/coccicheck warning:

WARNING: Use struct_size

Use size_add() to safely add any extra bytes.

Compile-tested only.

I believe this will fail at runtime. Note that sizeof_priv follows the allocation, so at the very least, you'd need to update:

static inline void *mux_chip_priv(struct mux_chip *mux_chip)

{
      return &mux_chip->mux[mux_chip->controllers];
}

to not use the mux array itself as a location reference because it will be seen as out of bounds.

To deal with this, the location will need to be calculated using mux_chip as the base, not mux_chip->mux as the base. For example, see commit 838ae9f45c4e ("nouveau/gsp: Avoid addressing beyond end of rpc->entries")

Quoted 7 lines
Quoted 3 lines

Since this should work and is well-defined C code according to [1][2], could you give this patch another look or should I still change it and submit a v2?

I think C is wrong here, but it seems it will continue to accidentally work. I personally would like a v3 that fixes this, but I leave it to Peter who is the MUX maintainer...

What's your take on this?

Thanks,
Thorsten

#10 lore raw

Peter?

Quoted 42 lines · Thorsten Blum
Quoted 32 lines · Kees Cook
Quoted 31 lines · Thorsten Blum
Quoted 29 lines · Kees Cook
Quoted 5 lines · Thorsten Blum

Convert mux_control_ops to a flexible array member at the end of the mux_chip struct and add the __counted_by() compiler attribute to improve access bounds-checking via CONFIG_UBSAN_BOUNDS and CONFIG_FORTIFY_SOURCE.

Quoted 8 lines

Use struct_size() to calculate the number of bytes to allocate for a new mux chip and to remove the following Coccinelle/coccicheck warning:

WARNING: Use struct_size

Use size_add() to safely add any extra bytes.

Compile-tested only.

I believe this will fail at runtime. Note that sizeof_priv follows the allocation, so at the very least, you'd need to update:

static inline void *mux_chip_priv(struct mux_chip *mux_chip)

{
     return &mux_chip->mux[mux_chip->controllers];
}

to not use the mux array itself as a location reference because it will be seen as out of bounds.

To deal with this, the location will need to be calculated using mux_chip as the base, not mux_chip->mux as the base. For example, see commit 838ae9f45c4e ("nouveau/gsp: Avoid addressing beyond end of rpc->entries")

Quoted 7 lines
Quoted 3 lines

Since this should work and is well-defined C code according to [1][2], could you give this patch another look or should I still change it and submit a v2?

I think C is wrong here, but it seems it will continue to accidentally work. I personally would like a v3 that fixes this, but I leave it to Peter who is the MUX maintainer...

What's your take on this?