Re: [CHECKER] more potential deadlocks

[James Morris]

On Tue, 4 Mar 2003, Dawson Engler wrote:

> --------------------------------------------------
> BUG?  but might be that the two different head pointers cannot point to
>       the same object.  not clear alias analysis will actually help here
>       since things are so hairy.
> 
> ERROR: 1 thread deadlock.
>    <&tw_death_lock>-><struct tcp_bind_hashbucket.lock (<local>:0)> occurred 1 times
>    <struct tcp_bind_hashbucket.lock (<local>:0)>-><&tw_death_lock> occurred 1 times

Not a bug.  The call chain is invalid:

> 	   ->/u2/engler/mc/oses/linux/linux-2.5.62/net/ipv4/tcp_ipv4.c:tcp_v4_hash_connect:683
> 	   ->/u2/engler/mc/oses/linux/linux-2.5.62/net/ipv4/tcp_ipv4.c:tcp_v4_hash_connect:694
> 	   ->/u2/engler/mc/oses/linux/linux-2.5.62/net/ipv4/tcp_ipv4.c:__tcp_v4_check_established:561
> 	   ->/u2/engler/mc/oses/linux/linux-2.5.62/net/ipv4/tcp_ipv4.c:__tcp_v4_check_established:622
> 	   ->/u2/engler/mc/oses/linux/linux-2.5.62/net/ipv4/tcp_ipv4.c:__tcp_v4_check_established:629
> 	   ->end=/u2/engler/mc/oses/linux/linux-2.5.62/net/ipv4/tcp_minisocks.c:tcp_tw_deschedule:481

The call to __tcp_v4_check_established() at line 694 takes a &tw paramater
which cannot be null.  This means that the call to tcp_tw_deschedule() at
line 629 will not happen and tw_death_lock is thus not taken on this path
while also holding a tcp_bind_hashbucket.lock.

If the caller provides a &tw param, a time-wait bucket (if found) is
returned via it to be destroyed outside the hash bucket lock.

> BUG: seems like it, if they can point to the same thing.  ERROR: 1 thread deadlock.
>    <struct in_device.lock (<local>:0)>-><struct ip_mc_list.lock (<local>:0)> occurred 5 times
>    <struct ip_mc_list.lock (<local>:0)>-><struct in_device.lock (<local>:0)> occurred 5 times

See below.

> BUG? very hard to follow, but interesting if a real bug.  unfortunately,
> could also be a false positive because of 
> 	1. infeasible callchain path or 
> 
> 	2. the various in_dev and im pointers never actually point to
> 	   the same object.
> 
> requires three threads: 
> 	thread 1: acquires im->lock then tries to get inetdev_lock
> 	thread 2: acquires inetdev_lock and tries to get in_dev->lock.
> 	thread 3: acquires in_dev->lock and tries to get im->lock.
> 
> ERROR: 2 thread deadlock.
>    <struct ip_mc_list.lock (<local>:0)>-><&inetdev_lock> occurred 5 times
>    <&inetdev_lock>-><struct ip_mc_list.lock (<local>:0)> occurred 4 times

These are indeed potential deadlock cases, caused by holding im->lock for
too long, now fixed by Alexey (in 2.5 bk at least).

- James
-- 
James Morris
<[e-mail]>