2006-11-27 18:56:55

by linas

[permalink] [raw]
Subject: [RFC/PATCH] Uncap max number of evdev devices [was: Re: need more events]

On Fri, Nov 24, 2006 at 05:09:52PM -0800, [email protected] wrote:
>
> The object is to poll 20 usb keyboards in an elementary school
> classroom, each of which generates 2
> events (one keyboard and one mouse). The stock kernel maxes out at event
> 31, leaving me 4
> keyboards short.
> I thought to fix this by changing the definition of EVDEV_MINORS
> (line 12, evdev.c) from 32 to 64. It almost worked. The extra
> events showed up in /dev/input/* and /proc/bus/input/devices.
> However, attempting to access the new events in the application
> program only yields a segmentation fault. Obviously I've got to change
> something else.

The problem is in input_register_handler() in drivers/input/input.c
which does things like
if (input_table[handler->minor >> 5])
and
input_table[handler->minor >> 5] = handler;

which implicitly makes 32 the max.

The kernel path below removes this limitation. Does it fix your problem?

--linas


The evdev code is limited to 32 input devices max, because
the generic input layer puts a 32-device cap on the number
of minor device numbers that a device driver can register.

This patch hacks around the 32-device limitation for
the number of evdev drivers. Not clear to me if this
this is the best solution.

Not-even-compile-tested.

Signed-off-by: Linas Vepstas <[email protected]>

----
drivers/input/input.c | 16 ++++++++++------
1 file changed, 10 insertions(+), 6 deletions(-)

Index: linux-2.6.19-rc4-git3/drivers/input/input.c
===================================================================
--- linux-2.6.19-rc4-git3.orig/drivers/input/input.c 2006-11-01 16:15:19.000000000 -0600
+++ linux-2.6.19-rc4-git3/drivers/input/input.c 2006-11-27 12:50:12.000000000 -0600
@@ -28,12 +28,16 @@ MODULE_AUTHOR("Vojtech Pavlik <vojtech@s
MODULE_DESCRIPTION("Input core");
MODULE_LICENSE("GPL");

-#define INPUT_DEVICES 256
+#define INPUT_MINOR_BITS 8
+#define INPUT_DEVICES (1<<INPUT_MINOR_BITS)
+#define MINOR_RANGE_BITS 6
+#define MINOR_RANGES (1<<(INPUT_MINOR_BITS-MINOR_RANGE_BITS))
+#define TABLE_OFFSET(minor) ((minor)>>MINOR_RANGE_BITS)

static LIST_HEAD(input_dev_list);
static LIST_HEAD(input_handler_list);

-static struct input_handler *input_table[8];
+static struct input_handler *input_table[MINOR_RANGES];

/**
* input_event() - report new input event
@@ -1046,10 +1050,10 @@ int input_register_handler(struct input_
INIT_LIST_HEAD(&handler->h_list);

if (handler->fops != NULL) {
- if (input_table[handler->minor >> 5])
+ if (input_table[TABLE_OFFSET(handler->minor)])
return -EBUSY;

- input_table[handler->minor >> 5] = handler;
+ input_table[TABLE_OFFSET(handler->minor)] = handler;
}

list_add_tail(&handler->node, &input_handler_list);
@@ -1082,7 +1086,7 @@ void input_unregister_handler(struct inp
list_del_init(&handler->node);

if (handler->fops != NULL)
- input_table[handler->minor >> 5] = NULL;
+ input_table[TABLE_OFFSET(handler->minor)] = NULL;

input_wakeup_procfs_readers();
}
@@ -1090,7 +1094,7 @@ EXPORT_SYMBOL(input_unregister_handler);

static int input_open_file(struct inode *inode, struct file *file)
{
- struct input_handler *handler = input_table[iminor(inode) >> 5];
+ struct input_handler *handler = input_table[TABLE_OFFSET(iminor(inode))];
const struct file_operations *old_fops, *new_fops = NULL;
int err;


2006-11-28 03:45:19

by Dmitry Torokhov

[permalink] [raw]
Subject: Re: [RFC/PATCH] Uncap max number of evdev devices [was: Re: need more events]

On Monday 27 November 2006 13:56, Linas Vepstas wrote:
> On Fri, Nov 24, 2006 at 05:09:52PM -0800, [email protected] wrote:
> >
> > The object is to poll 20 usb keyboards in an elementary school
> > classroom, each of which generates 2
> > events (one keyboard and one mouse). The stock kernel maxes out at event
> > 31, leaving me 4
> > keyboards short.
> > I thought to fix this by changing the definition of EVDEV_MINORS
> > (line 12, evdev.c) from 32 to 64. It almost worked. The extra
> > events showed up in /dev/input/* and /proc/bus/input/devices.
> > However, attempting to access the new events in the application
> > program only yields a segmentation fault. Obviously I've got to change
> > something else.
>
> The problem is in input_register_handler() in drivers/input/input.c
> which does things like
> if (input_table[handler->minor >> 5])
> and
> input_table[handler->minor >> 5] = handler;
>
> which implicitly makes 32 the max.
>
> The kernel path below removes this limitation. Does it fix your problem?
>

This makes /dev/input/js32-63 take over /dev/input/mouse0-31 which is
contrary to what Documentation/devices.txt says and therefore not a good
idea for the mainline.

I am planning to convert input core to cdevs and that table will go away
completely and then it can be discussed how to extend range of input
devices (probably by going beyond 256 minors?).

--
Dmitry