I found this when I was doing a bit of testing of my TTM patches. The issue is that during unload/unbind:
echo 0000:01:05.0 > /sys/bus/pci/drivers/radeon/unbind
the radeon driver would unload, but if compiled with the CONFIG_DMA_API_DEBUG=y it would throw out this warning:
[ 22.113420] drm: unregistered panic notifier^M [ 22.126500] [drm] radeon: finishing device.^M [ 22.130868] [drm] radeon: cp finalized^M [ 22.135384] [TTM] Finalizing pool allocator.^M [ 22.139669] [TTM] Zone kernel: Used memory at exit: 0 kiB.^M [ 22.145229] [drm] radeon: ttm finalized^M [ 22.149055] vga_switcheroo: disabled^M [ 22.153393] ------------[ cut here ]------------^M [ 22.158021] WARNING: at /home/konrad/ssd/linux/lib/dma-debug.c:689 dma_debug_device_change+0x12e/0x161()^M [ 22.167475] Hardware name: PowerEdge T105 ^M [ 22.171562] pci 0000:01:08.0: DMA-API: device driver has pending DMA allocations while released from device [count=1]^M [ 22.182139] Modules linked in: nouveau video sr_mod sg sd_mod cdrom ata_generic qla2xxx radeon sata_nv scsi_transport_fc scsi_tgt libata fbcon tileblit font bitblit softcursor ttm e1000e drm_kms_helper scsi_mod^M [ 22.201101] Pid: 2550, comm: primary_fb Not tainted 2.6.38-rc6-linux-next #1^M [ 22.208129] Call Trace:^M [ 22.210575] [<ffffffff81049580>] ? warn_slowpath_common+0x80/0x98^M [ 22.216737] [<ffffffff8104962c>] ? warn_slowpath_fmt+0x41/0x43^M [ 22.222643] [<ffffffff81208fe8>] ? dma_debug_device_change+0x12e/0x161^M [ 22.229240] [<ffffffff8144e467>] ? notifier_call_chain+0x32/0x5e^M [ 22.235319] [<ffffffff81067d5f>] ? __blocking_notifier_call_chain+0x4b/0x62^M [ 22.242350] [<ffffffff81067d85>] ? blocking_notifier_call_chain+0xf/0x11^M [ 22.249121] [<ffffffff812b54fd>] ? __device_release_driver+0xab/0xb0^M [ 22.255543] [<ffffffff812b55d0>] ? device_release_driver+0x1e/0x2b^M [ 22.261794] [<ffffffff812b4c86>] ? driver_unbind+0x57/0x8c^M [ 22.267352] [<ffffffff812b440d>] ? drv_attr_store+0x27/0x29^M [ 22.272998] [<ffffffff8114a64d>] ? sysfs_write_file+0xff/0x13b^M [ 22.278903] [<ffffffff810f5b88>] ? vfs_write+0xa9/0x105^M [ 22.284202] [<ffffffff810f5c9d>] ? sys_write+0x45/0x6c^M [ 22.289415] [<ffffffff8100a992>] ? system_call_fastpath+0x16/0x1b^M [ 22.295578] ---[ end trace f6e27fa61e0cc3ec ]---^M ERROR: Module radeon is in use^M
The radeon_gart_fini was missing an pci_unmap for the dummy page. Interestingly enough the r600.c has a call to radeon_dummy_page_fini() in its unwind path so there are definitly some chipsets on which this is done without any hinderance.
This above test is easily manifested with a stock 2.6.38-rc6 on a machine with a ATI ES1000 chipset (Dell T105, IBM x366, SuperMicro X7DB8).