8K stacks, I don't enable the CONFIG_4KSTACKS.
In addition, I also tried to enable CONFIG_4KSTACKS, the system still returns me lots "panic" during system bootup. :(

BRs,
Joseph Chan 

-----Original Message-----
From: Bruno Prémont [mailto:bonbons@linux-vserver.org] 
Sent: Friday, November 14, 2008 5:02 PM
To: Joseph Chan
Cc: akpm@linux-foundation.org; arjan@infradead.org; linux-fbdev-devel@lists.sourceforge.net; linux-kernel@vger.kernel.org; Scott Fang
Subject: Re: [PATCH] Fix crash in viafb due to 4k stack overflow

Joseph,

Are you testing with 4k stacks or with 8k stacks (CONFIG_4KSTACKS)?

Note that your failure looks pretty much like the ones I mentionned in
  http://lkml.org/lkml/2008/11/9/204
It looks like some kernel text/data got corrupted though I have no idea what caused it. For me this kind of NULL pointer dereference did happen a few times then stopped showing up (just a reboot, not kernel
change/recompile)

Bruno


On Fri, 14 Nov 2008 [off-list ref] wrote:

After testing the patch with 2.6.28-rc4 kernel, I found something 
strange in both viafb built-in and module modes. See the result on my 
EPIA-EX (CX700) below. Based on the result, it looks like something 
wrong with the new patch.

Kernel 2.6.28-rc4 + viafb-fix-crash-due-to-4k-stack-overflow.patch

1. w/o patching the viafb-fix-crash-due-to-4k-stack-overflow.patch
=> System works properly without "kernel panic" issue.

2. w/ patching the viafb-fix-crash-due-to-4k-stack-overflow.patch
=> System hangs with "kernel panic" randomly after loading the viafb 
module. Such as:
BUG: unable to handle kernel NULL pointer dereference at 000000c7
IP: [<c03dc6ec>] rt_worker_func+0xc7/0x15d Last sysfs file: 
/sys/class/graphics/fb0/dev Modules linked in: viafb i2c_algo_bit 
via_rhine sg pata_via sd_mod ehci_hcd

Pid: 6, comm: events/0 Not tainted (2.6.28-rc4b #1) pn test
EIP: 0060:[<c03dc6ec>] EFLAGS: 00010286 CPU:0 EIP is at 
rt_worker_func+0xc7/0x15d ......
....
.....


BRs,
Joseph

-----Original Message-----
From: Joseph Chan
Sent: Thursday, November 13, 2008 8:58 AM
To: 'Andrew Morton'; Bruno Prémont
Cc: arjan@infradead.org; linux-fbdev-devel@lists.sourceforge.net;
linux-kernel@vger.kernel.org Subject: RE: [PATCH] Fix crash in viafb 
due to 4k stack overflow

I will try this new patch later. :)

BRs,
Joseph Chan

-----Original Message-----
From: Andrew Morton [mailto:akpm@linux-foundation.org]
Sent: Thursday, November 13, 2008 7:01 AM
To: Bruno Prémont
Cc: arjan@infradead.org; Joseph Chan;
linux-fbdev-devel@lists.sourceforge.net; linux-kernel@vger.kernel.org
Subject: Re: [PATCH] Fix crash in viafb due to 4k stack overflow

On Mon, 10 Nov 2008 22:00:46 +0100
Bruno Pr__mont [off-list ref] wrote:

quoted

During conversion of viafb_ioctl() I noticed the following:

Those two cases just copy_from_user and do nothing with copied data, 
incomplete implementation?:
        case VIAFB_SET_PANEL_POSITION:
                if (copy_from_user(&u.panel_pos_size_para, argp,
                                   sizeof(u.panel_pos_size_para)))
                        return -EFAULT;
                break;
        case VIAFB_SET_PANEL_SIZE:
                if (copy_from_user(&u.panel_pos_size_para, argp,
                                   sizeof(u.panel_pos_size_para)))
                        return -EFAULT;
                break;

Handling of GET/SET GAMMA looks buggy:
In each case 256*4 bytes are allocated but only 4 bytes (size of
pointer) are copied to/from userspace though
viafb_(get|set)_gamma_table() operates on the full 256 elements...
        case VIAFB_SET_GAMMA_LUT:
                viafb_gamma_table = kmalloc(256 * sizeof(u32), 
GFP_KERNEL); if (!viafb_gamma_table)
                        return -ENOMEM;
                if (copy_from_user(viafb_gamma_table, argp,
                                sizeof(viafb_gamma_table))) {
                        kfree(viafb_gamma_table);
                        return -EFAULT;
                }
                viafb_set_gamma_table(viafb_bpp, viafb_gamma_table);
                kfree(viafb_gamma_table);
                break;

        case VIAFB_GET_GAMMA_LUT:
                viafb_gamma_table = kmalloc(256 * sizeof(u32), 
GFP_KERNEL); if (!viafb_gamma_table)
                        return -ENOMEM;
                viafb_get_gamma_table(viafb_gamma_table);
                if (copy_to_user(argp, viafb_gamma_table,
                        sizeof(viafb_gamma_table))) {
                        kfree(viafb_gamma_table);
                        return -EFAULT;
                }
                kfree(viafb_gamma_table);
                break;

I don't know if there is a userspace app that uses these VIA 
IOCTLs... so the ioctl part is just compile-tested.
After checking, fbset just uses some generic framebuffer IOCTLs 
outside of viafb's scope, thus not passing through viafb_ioctl().



---

--- linux-2.6.28-rc3/drivers/video/via/viafbdev.c.orig

2008-11-09 19:36:47.000000000 +0100 +++
linux-2.6.28-rc3/drivers/video/via/viafbdev.c	2008-11-10
20:50:32.000000000 +0100

hm, OK, I dropped the old patch and merged this one.

It'd be nice to have it runtime tested and reviewed by Joseph, but we 
haven't heard from him yet.  viafb might be 8k-stacks-only in 2.6.27, 
which would be bad.