Hi,

On Wed, 20. Aug 13:00, Melissa Wen wrote:

On 19/08/2025 15:46, Fedor Pchelkin wrote:

...

amdgpu_dm_connector_ddc_get_modes() reinitializes a connector's probed modes list without cleaning it up. First time it is called during the driver's initialization phase, then via drm_mode_getconnector() ioctl. The leaks observed with Kmemleak are as following:

unreferenced object 0xffff88812f91b200 (size 128): comm "(udev-worker)", pid 388, jiffies 4294695475 hex dump (first 32 bytes): ac dd 07 00 80 02 70 0b 90 0b e0 0b 00 00 e0 01 ......p......... 0b 07 10 07 5c 07 00 00 0a 00 00 00 00 00 00 00 ............... backtrace (crc 89db554f): __kmalloc_cache_noprof+0x3a3/0x490 drm_mode_duplicate+0x8e/0x2b0 amdgpu_dm_create_common_mode+0x40/0x150 [amdgpu] amdgpu_dm_connector_add_common_modes+0x336/0x488 [amdgpu] amdgpu_dm_connector_get_modes+0x428/0x8a0 [amdgpu] amdgpu_dm_initialize_drm_device+0x1389/0x17b4 [amdgpu] amdgpu_dm_init.cold+0x157b/0x1a1e [amdgpu] dm_hw_init+0x3f/0x110 [amdgpu] amdgpu_device_ip_init+0xcf4/0x1180 [amdgpu] amdgpu_device_init.cold+0xb84/0x1863 [amdgpu] amdgpu_driver_load_kms+0x15/0x90 [amdgpu] amdgpu_pci_probe+0x391/0xce0 [amdgpu] local_pci_probe+0xd9/0x190 pci_call_probe+0x183/0x540 pci_device_probe+0x171/0x2c0 really_probe+0x1e1/0x890

Found by Linux Verification Center (linuxtesting.org).

Fixes: acc96ae0d127 ("drm/amd/display: set panel orientation before drm_dev_register") Cc: stable@vger.kernel.org Signed-off-by: Fedor Pchelkin pchelkin@ispras.ru

---

drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c | 3 +++ 1 file changed, 3 insertions(+)

diff --git a/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c b/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c index cd0e2976e268..7ec1f9afc081 100644 --- a/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c +++ b/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c @@ -8227,9 +8227,12 @@ static void amdgpu_dm_connector_ddc_get_modes(struct drm_connector *connector, { struct amdgpu_dm_connector *amdgpu_dm_connector = to_amdgpu_dm_connector(connector);

• struct drm_display_mode *mode, *t; if (drm_edid) { /* empty probed_modes */

• list_for_each_entry_safe(mode, t, &connector->probed_modes, head)
  

• 	drm_mode_remove(connector, mode);
  

  INIT_LIST_HEAD(&connector->probed_modes); amdgpu_dm_connector->num_modes = drm_edid_connector_add_modes(connector);

What if you update the connector with the drm_edid data and skip the INIT_LIST_HEAD instead?

Yep, getting rid of INIT_LIST_HEAD eliminates the leak, too. drm_edid_connector_add_modes() comments do also strongly recommend calling drm_edid_connector_update() before the function.

One thing remaining strange is that there'd be several different objects in the probed_modes list describing the same things I guess.

Something like:

if (drm_edid) {

At this point we already have the modes in the list added with the previous call to amdgpu_dm_connector_get_modes() from amdgpu_set_panel_orientation() - during the driver initialization phase.

drm_edid_connector_update(connector, drm_edid);    amdgpu_drm_connector->num_modes = drm_edid_connector_add_modes(connector);

Here we add them again (as new objects) to the list. By the way it leads to amdgpu_drm_connector->num_modes be less than the actual number of elements present in probed_modes list.

As far as I understand, *_get_modes() are supposed to be called only via drm_mode_get_connector ioctl, and not all things go as expected if they're firstly called in another path, as e.g. in amdgpu case through amdgpu_set_panel_orientation().

But it seems commit acc96ae0d127 ("drm/amd/display: set panel orientation before drm_dev_register") added that call deliberately.

I think we may update the connector with the drm_edid data and skip the INIT_LIST_HEAD part as you've suggested, but also need to flush the list - it might contain something left from the first amdgpu_dm_connector_get_modes() call.

If no objections, I'll send it out as v3 soon.

[...] }

Isn't it enough?