IRQ/nvme_pci_complete_rq: NULL pointer dereference yet again

[mr.nuke.me@gmail.com (Alex G.)]

On 04/09/2018 02:11 PM, Keith Busch wrote:
> On Mon, Apr 09, 2018@01:23:54PM -0500, Alex G. wrote:
>> On 04/06/2018 05:00 PM, Keith Busch wrote:
>> (snip)
>>> ---
>>> diff --git a/drivers/pci/pcie/aer/aerdrv_core.c b/drivers/pci/pcie/aer/aerdrv_core.c
>>> index a4bfea52e7d4..16ecbcd76373 100644
>>> --- a/drivers/pci/pcie/aer/aerdrv_core.c
>>> +++ b/drivers/pci/pcie/aer/aerdrv_core.c
>>> @@ -805,8 +805,10 @@ void aer_isr(struct work_struct *work)
>>>  	struct pcie_device *p_device = rpc->rpd;
>>>  	struct aer_err_source uninitialized_var(e_src);
>>>  
>>> +	pci_lock_rescan_remove();
>>>  	mutex_lock(&rpc->rpc_mutex);
>>>  	while (get_e_source(rpc, &e_src))
>>>  		aer_isr_one_error(p_device, &e_src);
>>>  	mutex_unlock(&rpc->rpc_mutex);
>>> +	pci_unlock_rescan_remove();
>>>  }
>>> --
>>
>> With this patch, I'm not seeing issues without LVM mirrors, but as soon
>> as I enable the mirror, we get the
>>  * use-after-free in swiotlb_unmap_sg_attrs
>>  * double-free or invalid-free in nvme_pci_complete_rq
>>
>> Alex
> 
> Awesome, thank you for the update. We'll do some more work on the
> above fixing the use-after-free in AER handling to make it ready for
> consideration.

No problem.

> Still not sure on the nvme double/incorrect completion, but have not
> given up yet.

Although I remember a time when I could run these sort of tests without
panic(), I don't have a solid base from which I can bisect things.
Especially since we've added a few fixes lately that won't be in any base.

I'll try a couple of branches, but don't hold your breath.

Alex