lkml.org 
[lkml]   [2022]   [May]   [17]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
SubjectRe: [PATCH v6 5/5] x86/tdx: Add Quote generation support
From


On 5/17/22 4:06 PM, Kai Huang wrote:
> On Tue, 2022-05-17 at 13:08 -0700, Sathyanarayanan Kuppuswamy wrote:
>>
>> On 5/16/22 7:58 PM, Kai Huang wrote:
>>> On Fri, 2022-05-13 at 12:29 -0700, Sathyanarayanan Kuppuswamy wrote:
>>>>>
>>>>>
>>>>>> +
>>>>>> + /* Wait for attestation completion */
>>>>>> + ret = wait_for_completion_interruptible(&entry->compl);
>>>>>> + if (ret < 0) {
>>>>>> + /*
>>>>>> + * For premature termination, since VMM still owns the
>>>>>> + * shared buffer, mark the request invalid to let
>>>>>> + * quote_callback_handler() handle the memory cleanup
>>>>>> + * function.
>>>>>> + */
>>>>>> + invalidate_quote_request(entry);
>>>>>
>>>>> Interrupt can arrive after signal interrupt.  So invalidate_quote_request()
>>>>> should check if the request is already processed, and return 0 or -EINTR.
>>>>> Probably check the state always and del_list under single lock/unlock pair.
>>>>
>>>> Agree. But I think we should return -EINTR for the interrupted case
>>>> irrespective of the processed status (so no return 0).
>>>>
>>>> I will hold the lock and handle the cleanup for the processed
>>>> status.
>>>
>>> Even if we check the buffer status in invalidate_quote_request(), there's no
>>> guarantee the VMM won't change the buffer status _after_ we do the check, so
>>> looks such check isn't necessary.
>>>
>>
>> Consider the case where we get a callback interrupt, and before we
>> complete the processing for it, user terminates the request. In this
>> scenario, quote_callback_handler() will consider the request is
>> still valid and no do the memory cleanup. To handle this case,
>> we need to check the status in invalidate_quote_request() and do
>> the cleanup if required.
>>
>> /* Handles early termination of GetQuote requests */
>> void invalidate_quote_request(struct quote_entry *entry)
>> {
>> struct tdx_quote_hdr *quote_hdr;
>>
>> /*
>> * For early termination, if the request is not yet
>> * processed by VMM (GET_QUOTE_IN_FLIGHT), the VMM
>> * still owns the shared buffer, so mark the request
>> * invalid to let quote_callback_handler() handle the
>> * memory cleanup function. If the request is already
>> * processed, then do the cleanup and return.
>> */
>>
>> mutex_lock(&quote_lock);
>> quote_hdr = (struct tdx_quote_hdr *)entry->buf->vmaddr;
>> if (quote_hdr->status == GET_QUOTE_IN_FLIGHT) {
>
> What prevents VMM from updating quote_hdr->status from IN_FLIGHT to DONE _after_
> this check?

If IN_FLIGHT -> DONE happens at this point, you will get a separate
callback interrupt for it after this change. Since we hold quote_lock
here, processing of that callback work will wait till we make the
request invalid. For invalid requests, quote_callback_handler() will
handle the cleanup.

This logic is mainly for the case where quote_callback_handler() in
process and user terminates the request. In this case,
quote_callback_handler() will not do the cleanup, and we should add
logic here to handle it.

>
> If you want to add such check, you should check against GET_QUOTE_DONE, but not
> IN_FLIGHT. Only after status is DONE, VMM will not update the buffer.  Perhaps
> something like below:
>
> mutex_lock(&quote_lock);
> /* Skip invalidate the buffer if VMM has done with the buffer */
> if (quote_hdr->status == GET_QUOTE_DONE) {
> mutex_unlock(&quote_lock);
> return 0;
> }

There is no DONE status. Status can be either one of the following values.

GET_QUOTE_ERROR
GET_QUOTE_SUCCESS
GET_QUOTE_SERVICE_UNAVAILABLE
GET_QUOTE_IN_FLIGHT

>
> And in the IOCTL, you can perhaps to choose to return 0, instead of -EINTR in
> this case, as the Quote has been finished already.

I think your logic above is to return 0 for the completed request even
if it is interrupted. But IMO, since the user already interrupted the
request, then I think it is more appropriate to just return -EINTR for
this case.

>
> But I am not sure whether this is necessary. The worst case is one finished
> Quote is wasted I guess.

>
>> entry->valid = false;
>> mutex_unlock(&quote_lock);
>> return;
>> }
>> _del_quote_entry(entry);
>> mutex_unlock(&quote_lock);
>> }
>>
>>
>

--
Sathyanarayanan Kuppuswamy
Linux Kernel Developer

\
 
 \ /
  Last update: 2022-05-18 01:32    [W:3.081 / U:0.032 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site