On Wed, 10 Nov 2021 22:03:07 +0100, Kai Vehmanen wrote:
Fix a corner case between PCI device driver remove callback and runtime PM idle callback.
Following sequence of events can happen:
- at azx_create, context is allocated with devm_kzalloc() and stored as pci_set_drvdata()
- user-space requests to unbind audio driver
- dd.c:__device_release_driver() calls PCI remove
- pci-driver.c:pci_device_remove() calls the audio driver azx_remove() callback and this is completed
- pci-driver.c:pm_runtime_put_sync() leads to a call to rpm_idle() which again calls azx_runtime_idle()
- the azx context object, as returned by dev_get_drvdata(), is no longer valid
-> access fault in azx_runtime_idle when executing struct snd_card *card = dev_get_drvdata(dev); chip = card->private_data; if (chip->disabled || hda->init_failed)
This was discovered by i915_module_load test with 5.15.0 based linux-next tree.
Example log caught by i915_module_load test with linux-next https://intel-gfx-ci.01.org/tree/linux-next/
<4> [264.038232] general protection fault, probably for non-canonical address 0x6b6b6b6b6b6b73f0: 0000 [#1] PREEMPT SMP NOPTI <4> [264.038248] CPU: 0 PID: 5374 Comm: i915_module_loa Not tainted 5.15.0-next-20211109-gc8109c2ba35e-next-20211109 #1 [...] <4> [264.038267] RIP: 0010:azx_runtime_idle+0x12/0x60 [snd_hda_intel] [...] <4> [264.038355] Call Trace: <4> [264.038359] <TASK> <4> [264.038362] __rpm_callback+0x3d/0x110 <4> [264.038371] rpm_idle+0x27f/0x380 <4> [264.038376] __pm_runtime_idle+0x3b/0x100 <4> [264.038382] pci_device_remove+0x6d/0xa0 <4> [264.038388] device_release_driver_internal+0xef/0x1e0 <4> [264.038395] unbind_store+0xeb/0x120 <4> [264.038400] kernfs_fop_write_iter+0x11a/0x1c0
Fix the issue by setting drvdata to NULL at end of azx_remove().
Signed-off-by: Kai Vehmanen kai.vehmanen@linux.intel.com
sound/pci/hda/hda_intel.c | 1 + 1 file changed, 1 insertion(+)
Some non-persistent direct links showing the bug trigger on different platforms with linux-next 20211109:
- https://intel-gfx-ci.01.org/tree/linux-next/next-20211109/fi-tgl-1115g4/igt@...
- https://intel-gfx-ci.01.org/tree/linux-next/next-20211109/fi-jsl-1/igt@i915_...
Notably with 20211110 linux-next, the bug does not trigger:
Is this the case with CONFIG_DEBUG_KOBJECT_RELEASE? This would be the only logical explanation I can think of for now.
In anyway, the code change itself looks good, so I took the fix now.
thanks,
Takashi