Re: [REPORT] Request for reviewing crypto code wrt wait_for_completion()

[Byungchul Park]

On Fri, Aug 06, 2021 at 07:40:58PM +0800, Herbert Xu wrote:
> On Fri, Aug 06, 2021 at 05:03:44PM +0900, Byungchul Park wrote:
> > Hello crypto folks,
> > 
> > I developed a tool for tracking waiters and reporting if any of the
> > events that the waiters are waiting for would never happen, say, a
> > deadlock. Yes, it would look like Lockdep but more inclusive.
> > 
> > While I ran the tool(Dept: Dependency Tracker) on v5.4.96, I got some
> > reports from the tool. One of them is related to crypto subsystem.
> > Because I'm not that familiar with the code, I'd like to ask you guys to
> > review the related code.
> > 
> > If I understand correctly, it doesn't actually cause deadlock but looks
> > like a problematic code. I know you are not used to the format of the
> > report from Dept so.. let me summerize the result.
> > 
> > The simplified call trace looks like when the problem araised :
> > 
> > THREAD A
> > --------
> > A1 crypto_alg_mod_lookup()
> > A2    crypto_probing_notify(CRYPTO_MSG_ALG_REQUEST)
> > A3       cryptomgr_schedule_probe()
> > A4          kthread_run(cyptomgr_probe) ---> Start THREAD B
> > 
> > A5    crypto_larval_wait()
> > A6       wait_for_completion_killable_timeout(c) /* waiting for B10 */
> 
> This larval would be an instantiation larval, and it can only be
> woken up by thread B, not C.

Yes. This is what I understood based on the code.

> > THREAD B
> > --------
> > B1 cryptomgr_probe()
> > B2    pkcslpad_create()
> > B3       crypto_wait_for_test()
> > B4          crypto_probing_notify(CRYPTO_MSG_ALG_REGISTER)
> > B5             cryptomgr_schedule_test()
> > B6                kthread_run(cyptomgr_test) ---> Start THREAD C
> > 
> > B7    tmpl->alloc()
> > B8    crupto_register_instance()
> > B9          wait_for_completion_killable(c) /* waiting for C3 */
> > B10   complete_all(c)
> 
> I presume you're talking about about the wait_for_completion from

Right. Sorry for confusing you.

> crypto_wait_for_test, in which case it can only be woken by thread
> C.  After which thread B will return to cryptomgr_probe and wake up
> thread A.

Yes. This is what I understood based on the code too.

> 
> > THREAD C
> > --------
> > C1 cryptomgr_test()
> > C2    crypto_alg_tested()
> > C3       complete_all(c)
> > 
> > ---
> > 
> > For example, in this situation, I think C3 could wake up both A6 and B9
> > before THREAD B reaches B10 which is not desired by A6. Say, is it okay
> > to wake up A6 with B7 ~ B9 having yet to complete?
> 
> AFAICS thread C only wakes up test larvals, not instantiation larvals.
> Please let me know if you have any further issues.

The both cases looks like to get the larvals from the same list,
crypto_alg_list, one from crypto_larval_lookup() and the other from
__crypto_register_alg(). So I thought a single larval can be used at the
same time both at crypto_wait_for_test() and crypto_alg_mod_lookup() by
any chance. It would be great if the code ensures it never happens :-)

The problematic scenario I wanted to ask you looks like - I was
wondering if it's okay to nest requesting CRYPTO_MSG_ALG_REQUEST and
CRYPTO_MSG_ALG_REGISTER in a single stack, in other words, if it's okay
to try CRYPTO_MSG_ALG_REGISTER before completing CRYPTO_MSG_ALG_REQUEST.

A1 crypto_alg_mod_lookup()
A2    crypto_probing_notify(CRYPTO_MSG_ALG_REQUEST)
A3       cryptomgr_schedule_probe()
A4          kthread_run(cyptomgr_probe) ---> Start THREAD B

B1 cryptomgr_probe()
B2    pkcslpad_create()
B3       crypto_wait_for_test()
B4          crypto_probing_notify(CRYPTO_MSG_ALG_REGISTER)
B5             cryptomgr_schedule_test()
B6                kthread_run(cyptomgr_test) ---> Start THREAD C

C1 cryptomgr_test()
C2    crypto_alg_tested()
C3       complete_all(c) <- *the point* that I'd like to ask you.

A5    crypto_larval_wait()
A6       wait_for_completion_killable_timeout(c) /* waiting for B10 */
         (wake up and go)

Bx          wait_for_completion_killable(c) /* waiting for C3 */
            (wake up and go)
Bx    tmpl->alloc()
Bx    crupto_register_instance()
B10   complete_all(c)

I think I've shown you all the detail about the problematic flow. If
it still looks okay to you, then it'd be great!

Thank you,
Byungchul

> Thanks,
> -- 
> Email: Herbert Xu <herbert@xxxxxxxxxxxxxxxxxxx>
> Home Page: http://gondor.apana.org.au/~herbert/
> PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt