Created attachment 42147 [details]
Little program to test the bug.

Hello,
I have a bug on many ATI computers with a debian squeeze with a homemade
kernel 2.6.37. (I join you the "lspci" of three of them.)


The bug appears whether I use fglrx or radeon, with or without KMS.
Bug does not appear if I use "vesa" driver or on Intel computers.

I think my bug is similar to this one:
https://bugs.freedesktop.org/show_bug.cgi?id=32830

I can reproduce this bug three different ways:
- Move the mouse and launch a setxkbmap (you don't need an argument).
- Move the mouse, launch Openoffice with a bad argument (Ex: "ooffice -lag")
   and wait (while always moving the mouse) until OpenOffice is loaded.
- Launch Openoffice, move the mouse and call a programme which using a program
   using libxtst. (See attachment "xtestfakekey.c".)



The bug can be described : your mouse will slow down, everything X-related on
your computer will freeze (Ex: videos will stop). But, when you stop the mouse,
everything will start again.

If we move the mouse long enough, X will show the following errors:
	"
Errors from xkbcomp are not fatal to the X server
  [mi] EQ overflowing. The server is probably stuck in an infinite loop.
Backtrace:
0: /usr/bin/X (xorg_backtrace+0x3b) [0x80e7aab]
1: /usr/bin/X (mieqEnqueue+0x1ab) [0x80e73ab]
2: /usr/bin/X (xf86PostMotionEventP+0xd2) [0x80c1882]
3: /usr/bin/X (xf86PostMotionEvent+0x68) [0x80c1a08]
4: /usr/lib/xorg/modules/input/synaptics_drv.so (0xb732c000+0x3508)[0xb732f508]
5: /usr/lib/xorg/modules/input/synaptics_drv.so (0xb732c000+0x5aed)[0xb7331aed]
6: /usr/bin/X (0x8048000+0x6c29f) [0x80b429f]
7: /usr/bin/X (0x8048000+0x11e3c4) [0x81663c4]
8: (vdso) (__kernel_sigreturn+0x0) [0xffffe400]
9: /lib/libpthread.so.0 (fork+0x14) [0xb76f9094]
10: /usr/bin/X (Popen+0x9b) [0x80a9f2b]
11: /usr/bin/X (XkbDDXLoadKeymapByNames+0x1c7) [0x81bb897]
12: /usr/bin/X (0x8048000+0xf1d0c) [0x8139d0c]
13: /usr/bin/X (0x8048000+0x2b027) [0x8073027]
14: /usr/bin/X (0x8048000+0x1e95a) [0x806695a]
15: /lib/libc.so.6 (__libc_start_main+0xe6) [0xb7467c76]
16: /usr/bin/X (0x8048000+0x1e541) [0x8066541]
"

And multiple times the following message:
"
Warning:          Multiple interpretations of "NoSymbol+AnyOf(all)"
                    Using last definition for duplicate fields
"

I tried changing distribution (debian squeeze, then gentoo), and changing my
window manager (gnome then E17), but nothing changed.

I try to resolve the bug with new or older version of the kernel.
The bug is a regression created between Kernel 2.6.32.27 an 2.6.33 and is
not since resolved.
I tried a git bisect between 2.6.32 and 2.6.33, but there's a bug between the
two releases preventing the kernel to boot (sometimes with KMS,
sometimes without, sometimes both)

During my git bisect, I notice that the bug appear sooner without KMS.
In Kernel 2.6.37, the bug appear in the two situations.

I try tu use strace in setxkbmap, the lag will appear in the following line:
"poll([{fd=3, events=POLLIN}], 1, -1)    = 1 ([{fd=3, revents=POLLIN}])"
The fd of this poll is defined:
"socket(PF_FILE, SOCK_STREAM|SOCK_CLOEXEC, 0) = 3"

I try a latrace in setxkbmap, and the lag will appear in the following line:
"
20689           xcb_wait_for_reply [/usr/lib/libxcb.so.1]
20689             pthread_mutex_lock [/lib/libc.so.6]
20689             pthread_mutex_unlock [/lib/libc.so.6]
20689             poll [/lib/libc.so.6]                *lag here*
20689             pthread_mutex_lock [/lib/libc.so.6]
20689             read [/lib/libc.so.6]
20689             malloc [/lib/libc.so.6]
20689             memcpy [/lib/libc.so.6]
20689             memmove [/lib/libc.so.6]
20689             read [/lib/libc.so.6]
20689             read [/lib/libc.so.6]
20689             __errno_location [/lib/libc.so.6]
20689             poll [/lib/libc.so.6]
20689             read [/lib/libc.so.6]
20689             malloc [/lib/libc.so.6]
20689             pthread_cond_signal [/lib/libc.so.6]
20689             free [/lib/libc.so.6]
20689             pthread_cond_destroy [/lib/libc.so.6]
20689             pthread_cond_signal [/lib/libc.so.6]
20689             pthread_mutex_unlock [/lib/libc.so.6]
"


After all these test, I always cannot find what I can do to stop this bug.
I also join report from ftrace and perf, in order to give you more information.
Ftrace cannot parse hrtimer functions (like "hrtimer_start").
And perf shows that the "strcmp" keeps all the processor.



Do you have any idea ?
Regards,