Commit d5c435df4a890 ("intel_th: msu: Use the real device in case of IOMMU domain allocation") changes dma buffer allocation to use the actual underlying device, but forgets to change the deallocation path, which leads to (if you've got CAP_SYS_RAWIO):

# echo 0,0 > /sys/bus/intel_th/devices/0-msc0/nr_pages ------------[ cut here ]------------ kernel BUG at ../linux/drivers/iommu/intel-iommu.c:3670! CPU: 3 PID: 231 Comm: sh Not tainted 4.17.0-rc1+ #2729 RIP: 0010:intel_unmap+0x11e/0x130

...

Call Trace: intel_free_coherent+0x3e/0x60 msc_buffer_win_free+0x100/0x160 [intel_th_msu]

This patch fixes the buffer deallocation code to use the correct device.

Signed-off-by: Alexander Shishkin alexander.shishkin@linux.intel.com Fixes: d5c435df4a890 ("intel_th: msu: Use the real device in case of IOMMU domain allocation") Reported-by: Baofeng Tian baofeng.tian@intel.com CC: stable@vger.kernel.org # v4.14+ --- drivers/hwtracing/intel_th/msu.c | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/drivers/hwtracing/intel_th/msu.c b/drivers/hwtracing/intel_th/msu.c index ede388309376..634f58042c77 100644 --- a/drivers/hwtracing/intel_th/msu.c +++ b/drivers/hwtracing/intel_th/msu.c @@ -733,8 +733,8 @@ static int msc_buffer_win_alloc(struct msc *msc, unsigned int nr_blocks) /* Reset the page to write-back before releasing */ set_memory_wb((unsigned long)win->block[i].bdesc, 1); #endif - dma_free_coherent(msc_dev(msc), size, win->block[i].bdesc, - win->block[i].addr); + dma_free_coherent(msc_dev(msc)->parent->parent, size, + win->block[i].bdesc, win->block[i].addr); } kfree(win);

@@ -769,7 +769,7 @@ static void msc_buffer_win_free(struct msc *msc, struct msc_window *win) /* Reset the page to write-back before releasing */ set_memory_wb((unsigned long)win->block[i].bdesc, 1); #endif - dma_free_coherent(msc_dev(win->msc), PAGE_SIZE, + dma_free_coherent(msc_dev(win->msc)->parent->parent, PAGE_SIZE, win->block[i].bdesc, win->block[i].addr); }