[patch 2.6.12-rc2] revert fs/char_dev.c CONFIG_BASE_FULL change

[patch 2.6.12-rc2] revert fs/char_dev.c CONFIG_BASE_FULL change

[David Brownell]

[-- Attachment #1: Type: text/plain, Size: 349 bytes --]

I tracked down a regression in PCMCIA (and other software) to a
new bogus register_chrdev() behavior that got merged last month;
a patch from Matt Mackall that misbehaves.

This patch just reverts Matt's, restoring the previous behavior
but at the cost of about a Kbyte of static memory on 32bit CPUs.
Someday a Real Fix(tm) would be good.

- Dave


[-- Attachment #2: chrdev.patch --]
[-- Type: text/x-diff, Size: 952 bytes --]

This reverts a fs/char_dev.c patch that was merged into BK on March 3.

The problem is that it breaks things ... __register_chrdev_region() has
a block of code, commented "temporary" for over two years now, which
fails rudely during PCMCIA initialization or other register_chrdev()
calls, because it doesn't "degrade to linked list".  This keeps whole
subsystems from working.

A real fix to that "temporary" code should be possible, using some better
scheme to allocate major numbers, but it's not something I want to spend
time on just now.

Signed-off-by: David Brownell <dbrownell@users.sourceforge.net>

--- 1.38/fs/char_dev.c	2005-03-09 09:03:28 -08:00
+++ edited/fs/char_dev.c	2005-04-17 08:45:19 -07:00
@@ -26,8 +26,7 @@
 
 static struct kobj_map *cdev_map;
 
-/* degrade to linked list for small systems */
-#define MAX_PROBE_HASH (CONFIG_BASE_SMALL ? 1 : 255)
+#define MAX_PROBE_HASH 255	/* random */
 
 static DECLARE_MUTEX(chrdevs_lock);
 

Re: [patch 2.6.12-rc2] revert fs/char_dev.c CONFIG_BASE_FULL change

[Matt Mackall]

On Sun, Apr 17, 2005 at 10:06:53AM -0700, David Brownell wrote:
> I tracked down a regression in PCMCIA (and other software) to a
> new bogus register_chrdev() behavior that got merged last month;
> a patch from Matt Mackall that misbehaves.

Thanks and sorry about that. I actually asked Linus to revert it right
after Andrew pushed it to him, but I hadn't actually seen a problem
report so I didn't pursue it.
 
> This patch just reverts Matt's, restoring the previous behavior
> but at the cost of about a Kbyte of static memory on 32bit CPUs.
> Someday a Real Fix(tm) would be good.

I've got a Real Fix (a refactor of block and chardev name
registration), will need to wait for .12.

> Signed-off-by: David Brownell <dbrownell@users.sourceforge.net>
Acked-by: Matt Mackall <mpm@selenic.com>

-- 
Mathematics is the supreme nostalgia of our time.