Re: [PATCH v2 5/6] platform/x86: intel_tdx_attest: Add TDX Guest attestation interface driver

[Sathyanarayanan Kuppuswamy <sathyanarayanan.kuppuswamy@linux.intel.com>]

Hi Hans,

On 4/4/22 3:07 AM, Hans de Goede wrote:
>> +static int __init tdx_attest_init(void)
>> +{
>> +	dma_addr_t handle;
>> +	long ret = 0;
>> +
>> +	mutex_lock(&attestation_lock);
>> +
>> +	ret = misc_register(&tdx_attest_device);
>> +	if (ret) {
>> +		pr_err("misc device registration failed\n");
>> +		mutex_unlock(&attestation_lock);
>> +		return ret;
>> +	}
> Why not do this as the last thing of the probe?

We need misc device reference in dma_alloc_coherent() and
dma_set_coherent_mask() calls. This is the reason for keeping
misc_register() at the beginining of the init function.

> 
> That will avoid the need to unregister this again in all
> the error-exit paths and also fixes a possible deadlock.
> 

Agree. But, unless we create another device locally, I don't
think we can avoid this. Do you prefer this approach?

> Right now you possibly have:
> 
> 1. probe() locks attestation_lock
> 2. probe() registers misc-device
> 3. userspace calls tdx_attest_ioctl
> 4. tdx_attest_ioctl blocks waiting for attestastion_lock
> 5. Something goes wrong in probe, probe calls
>     misc_deregister()
> 6. misc_deregister waits for the ioctl to finish
> 7. deadlock
> 
> I'm not sure about 6, but if 6 does not happen then
> instead we now have tdx_attest_ioctl running
> after the misc_deregister, with tdquote_data and
> tdreport_data as NULL, or pointing to free-ed memory
> leading to various crash scenarios.

Makes sense. But as I have mentioned above, we have reason
for keeping the misc_register() at the begining of the
init function.

One way to avoid this deadlock is to use global initalization
check.

--- a/drivers/platform/x86/intel/tdx/intel_tdx_attest.c
+++ b/drivers/platform/x86/intel/tdx/intel_tdx_attest.c
@@ -48,6 +48,8 @@ static void *tdreport_data;
  /* DMA handle used to allocate and free tdquote DMA buffer */
  dma_addr_t tdquote_dma_handle;

+static bool device_initialized;
+
  static void attestation_callback_handler(void)
  {
         complete(&attestation_done);
@@ -60,6 +62,9 @@ static long tdx_attest_ioctl(struct file *file, 
unsigned int cmd,
         struct tdx_gen_quote tdquote_req;
         long ret = 0;

+       if (!device_initialized)
+               return -ENODEV;
+
         mutex_lock(&attestation_lock);

         switch (cmd) {
@@ -191,6 +196,8 @@ static int __init tdx_attest_init(void)

         mutex_unlock(&attestation_lock);

+       device_initialized = true;
+
         pr_debug("module initialization success\n");

         return 0;

Please let me know your comment on above solution.

> 
> TL;DR: you must always delay registering any
> interfaces for userspace until your code is
> ready to deal with userspace calls.
> 
> Regards,
> 
> Hans
> 
> p.s.
> 
> As I mentioned with v1:
> 
> 

-- 
Sathyanarayanan Kuppuswamy
Linux Kernel Developer