Re: [EXT] Re: [PATCH] USB: gadget: Add ID numbers to configfs-gadget driver names

From: Chanh Nguyen
Date: Thu Dec 15 2022 - 04:08:29 EST



On 14/12/2022 11:20, Frank Li wrote:

Caution: EXT Email

On 13/12/2022 14:22, Christophe JAILLET wrote:
Le 13/12/2022 à 05:12, Chanh Nguyen a écrit :
It is unable to use configfs to attach more than one gadget. When
attaching the second gadget, it always fails and the kernel message
prints out:

Error: Driver 'configfs-gadget' is already registered, aborting...
UDC core: g1: driver registration failed: -16

This commit fixes the problem by a ".N" suffix added to each
configfs_gadget's driver name (where N is a unique ID number),
thus making the names distinct.

Fixes: fc274c1e9973 ("USB: gadget: Add a new bus for gadgets")
Signed-off-by: Chanh Nguyen
<chanh-shex6MNQR2J/SfDzf78azzKzEDxYleXD@xxxxxxxxxxxxxxxx>
---
drivers/usb/gadget/configfs.c | 42
+++++++++++++++++++++++++++++++++++
1 file changed, 42 insertions(+)

diff --git a/drivers/usb/gadget/configfs.c
b/drivers/usb/gadget/configfs.c
index 96121d1c8df4..d8c5156ad777 100644
--- a/drivers/usb/gadget/configfs.c
+++ b/drivers/usb/gadget/configfs.c
@@ -3,6 +3,7 @@
#include <linux/module.h>
#include <linux/slab.h>
#include <linux/device.h>
+#include <linux/idr.h>
#include <linux/kstrtox.h>
#include <linux/nls.h>
#include <linux/usb/composite.h>
@@ -11,6 +12,16 @@
#include "u_f.h"
#include "u_os_desc.h"
+static DEFINE_IDA(driver_id_numbers);
+
+/*
+ * Driver name has the form of "configfs-gadget.%d", where %d
+ * is id allocated by ida_alloc(). The max value returns by
+ * ida_alloc() is INT_MAX, in 64-bit system, it is about nine
+ * quintillion: 19 digits in decimal. Set the max length to 35.
+ */
+#define DRIVER_NAME_LENGTH_MAX 35

Hi,

if paranoiac, the final \0 seems to be missing in the max length
computation, but see below.

Thanks CJ! Indeed, I have missed that.


+
int check_user_usb_string(const char *name,
struct usb_gadget_strings *stringtab_dev)
{
@@ -52,6 +63,9 @@ struct gadget_info {
char qw_sign[OS_STRING_QW_SIGN_LEN];
spinlock_t spinlock;
bool unbind;
+
+ /* Make driver names unique */
+ int driver_id_number;
};
static inline struct gadget_info *to_gadget_info(struct config_item
*item)
@@ -1582,6 +1596,8 @@ static struct config_group *gadgets_make(
const char *name)
{
struct gadget_info *gi;
+ char *driver_name;
+ int ret;
gi = kzalloc(sizeof(*gi), GFP_KERNEL);
if (!gi)
@@ -1623,6 +1639,21 @@ static struct config_group *gadgets_make(
gi->composite.gadget_driver = configfs_driver_template;
+ ret = ida_alloc(&driver_id_numbers, GFP_KERNEL);
+ if (ret < 0)
+ goto err;
+ gi->driver_id_number = ret;
+
+ driver_name = kmalloc(DRIVER_NAME_LENGTH_MAX, GFP_KERNEL);
+ if (!driver_name)
+ goto out_free_driver_id_number;
+
+ ret = scnprintf(driver_name, DRIVER_NAME_LENGTH_MAX,
+ "configfs-gadget.%d", gi->driver_id_number);


using kasprintf() looks simpler here.
No need to kmalloc()+scnprintf(), and no need for
DRIVER_NAME_LENGTH_MAX.

Just my 2c,

CJ

Thanks CJ for the review!

I've made some changes as below (in gadgets_make() to remove
unnecessary
variables) and now trying to test it as much as possible. Will re-post
it as v2 if looks good soon.

static inline struct gadget_info *to_gadget_info(struct config_item *item)
@@ -1623,13 +1629,25 @@ static struct config_group *gadgets_make(

gi->composite.gadget_driver = configfs_driver_template;

+ gi->driver_id_number = ida_alloc(&driver_id_numbers, GFP_KERNEL);
+ if (gi->driver_id_number < 0)
+ goto err;
+
+ gi->composite.gadget_driver.driver.name =
+ kasprintf(GFP_KERNEL, "configfs-gadget.%d",
+ gi->driver_id_number);
+ if (!gi->composite.gadget_driver.driver.name)
+ goto out_free_driver_id_number;
+
gi->composite.gadget_driver.function = kstrdup(name, GFP_KERNEL);
gi->composite.name = gi->composite.gadget_driver.function;

if (!gi->composite.gadget_driver.function)
- goto err;
+ goto out_free_driver_id_number;

return &gi->group;
+
+out_free_driver_id_number:
+ ida_free(&driver_id_numbers, gi->driver_id_number);
err:
kfree(gi);
return ERR_PTR(-ENOMEM);



+ if (ret < 0)
+ goto out_free_driver_name;
+
+ gi->composite.gadget_driver.driver.name = driver_name;
gi->composite.gadget_driver.function = kstrdup(name, GFP_KERNEL);
gi->composite.name = gi->composite.gadget_driver.function;
@@ -1630,6 +1661,11 @@ static struct config_group *gadgets_make(
goto err;
return &gi->group;
+
+out_free_driver_name:
+ kfree(driver_name);
+out_free_driver_id_number:
+ ida_free(&driver_id_numbers, gi->driver_id_number);
err:
kfree(gi);
return ERR_PTR(-ENOMEM);
@@ -1637,6 +1673,12 @@ static struct config_group *gadgets_make(
static void gadgets_drop(struct config_group *group, struct
config_item *item)
{
+ struct gadget_info *gi = to_gadget_info(item);
+
+ mutex_lock(&gi->lock);
+ kfree(gi->composite.gadget_driver.driver.name);
+ ida_free(&driver_id_numbers, gi->driver_id_number);
+ mutex_unlock(&gi->lock);

Move all free into gadget_info_attr_release(), just before kfree(gi)
Driver.name and gi create at the same place,
Free should be the same place also.


Thanks a lot for the quick review comment.

As per my observation through the test, on the first mount, the virtual-media the gadgets_make() is called, then later, when unmount, the gadgets_drop() is called and followed by gadget_info_attr_release().

The gadget_info_attr_release() is registered as .release() of gadget_root_type for the gi->group via the call "config_group_init_type_name(&gi->group, name, &gadget_root_type);"

In general, the .release() will be called only for the group. There is nothing to guarantee that the group will always be registered, ie: incase without the call to "config_group_init_type_name(&gi->group, name, &gadget_root_type);"

In this patch, what is added is an ida number to be used to make up the composite driver name. This is done in gadgets_make() so we'd like to add the cleanup code to gadgets_drop() as they are registered together in the same place and would be a little easier to read than adding them to _release() as the code below:

    static struct configfs_group_operations gadgets_ops = {
        .make_group = &gadgets_make,
        .drop_item = &gadgets_drop,
    };

Anyway, we still doubt that there might be something that we have missed so please let me know the reason why putting cleanup codes to _release() would be a better solution.

Thank you and best regards,
- Chanh


config_item_put(item);
}