Re: [PATCH 1/5] tcp/md5: Don't BUG_ON() failed kmemdup()

[Dmitry Safonov <0x7f454c46@gmail.com>]

On 11/5/21 02:55, Eric Dumazet wrote:
> 
> 
> On 11/4/21 6:49 PM, Dmitry Safonov wrote:
>> static_branch_unlikely(&tcp_md5_needed) is enabled by
>> tcp_alloc_md5sig_pool(), so as long as the code doesn't change
>> tcp_md5sig_pool has been already populated if this code is being
>> executed.
>>
>> In case tcptw->tw_md5_key allocaion failed - no reason to crash kernel:
>> tcp_{v4,v6}_send_ack() will send unsigned segment, the connection won't be
>> established, which is bad enough, but in OOM situation totally
>> acceptable and better than kernel crash.
>>
>> Introduce tcp_md5sig_pool_ready() helper.
>> tcp_alloc_md5sig_pool() usage is intentionally avoided here as it's
>> fast-path here and it's check for sanity rather than point of actual
>> pool allocation. That will allow to have generic slow-path allocator
>> for tcp crypto pool.
>>
>> Signed-off-by: Dmitry Safonov <dima@arista.com>
>> ---
>>  include/net/tcp.h        | 1 +
>>  net/ipv4/tcp.c           | 5 +++++
>>  net/ipv4/tcp_minisocks.c | 5 +++--
>>  3 files changed, 9 insertions(+), 2 deletions(-)
>>
>> diff --git a/include/net/tcp.h b/include/net/tcp.h
>> index 4da22b41bde6..3e5423a10a74 100644
>> --- a/include/net/tcp.h
>> +++ b/include/net/tcp.h
>> @@ -1672,6 +1672,7 @@ tcp_md5_do_lookup(const struct sock *sk, int l3index,
>>  #endif
>>  
>>  bool tcp_alloc_md5sig_pool(void);
>> +bool tcp_md5sig_pool_ready(void);
>>  
>>  struct tcp_md5sig_pool *tcp_get_md5sig_pool(void);
>>  static inline void tcp_put_md5sig_pool(void)
>> diff --git a/net/ipv4/tcp.c b/net/ipv4/tcp.c
>> index b7796b4cf0a0..c0856a6af9f5 100644
>> --- a/net/ipv4/tcp.c
>> +++ b/net/ipv4/tcp.c
>> @@ -4314,6 +4314,11 @@ bool tcp_alloc_md5sig_pool(void)
>>  }
>>  EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
>>  
>> +bool tcp_md5sig_pool_ready(void)
>> +{
>> +	return tcp_md5sig_pool_populated;
>> +}
>> +EXPORT_SYMBOL(tcp_md5sig_pool_ready);
>>  
>>  /**
>>   *	tcp_get_md5sig_pool - get md5sig_pool for this user
>> diff --git a/net/ipv4/tcp_minisocks.c b/net/ipv4/tcp_minisocks.c
>> index cf913a66df17..c99cdb529902 100644
>> --- a/net/ipv4/tcp_minisocks.c
>> +++ b/net/ipv4/tcp_minisocks.c
>> @@ -293,11 +293,12 @@ void tcp_time_wait(struct sock *sk, int state, int timeo)
>>  			tcptw->tw_md5_key = NULL;
>>  			if (static_branch_unlikely(&tcp_md5_needed)) {
>>  				struct tcp_md5sig_key *key;
>> +				bool err = WARN_ON(!tcp_md5sig_pool_ready());
>>  
>>  				key = tp->af_specific->md5_lookup(sk, sk);
>> -				if (key) {
>> +				if (key && !err) {
>>  					tcptw->tw_md5_key = kmemdup(key, sizeof(*key), GFP_ATOMIC);
>> -					BUG_ON(tcptw->tw_md5_key && !tcp_alloc_md5sig_pool());
>> +					WARN_ON_ONCE(tcptw->tw_md5_key == NULL);
>>  				}
>>  			}
>>  		} while (0);
>>
> 
> Hmmm.... how this BUG_ON() could trigger exactly ?
> 
> tcp_md5_needed can only be enabled after __tcp_alloc_md5sig_pool has succeeded.

Yeah, I've misread this part as
: BUG_ON(!tcptw->tw_md5_key || !tcp_alloc_md5sig_pool());

Still, there is an issue with checking tcp_alloc_md5sig_pool():
currently the condition is never true, but if it ever becomes true, the
tcp_alloc_md5sig_pool() call may cause tcp_time_wait() to sleep with bh
disabled (i.e. __tcp_close()). So, if this condition ever becomes true,
it will cause an issue checking it here.

I'll squash this with patch 3 and send when the merge window closes.

Thanks,
          Dmitry