Messages in this thread | | | Date | Thu, 5 Jan 2012 03:06:03 +0100 | From | Frederic Weisbecker <> | Subject | Re: [PATCH] rcu: Improve detection of illegal synchronize_rcu() call from RCU read side |
| |
On Wed, Jan 04, 2012 at 06:01:08PM -0800, Paul E. McKenney wrote: > On Thu, Jan 05, 2012 at 02:45:20AM +0100, Frederic Weisbecker wrote: > > On Wed, Jan 04, 2012 at 01:30:35PM -0800, Paul E. McKenney wrote: > > > On Wed, Jan 04, 2012 at 08:03:39PM +0100, Frederic Weisbecker wrote: > > > > Actually for the case of RCU, the wait_for_completion() called by synchronize_rcu() > > > > has a might_sleep() call that triggers a warning in this case. > > > > > > > > But in the case of SMP with 1 online CPU, the rcu_blocking_is_gp() > > > > checks returns right away on rcutree. So probably we need this? > > > > > > I modified this to push the might_sleep() down into the > > > rcu_blocking_is_gp() function, queued the result, and retained your > > > Signed-off-by. (Please let me know if there is any problem with this.) > > > > > > This does work for TREE_PREEMPT_RCU and for synchronize_rcu_bh() in > > > TREE_RCU, but not for synchronize_sched() in TREE_RCU. This is because > > > rcu_read_lock() and rcu_read_unlock() are no-ops in the TREE_RCU case. > > > > Not sure about that. This calls preempt_disable() which, in any case with > > CONFIG_DEBUG_ATOMIC_SLEEP, handles the preempt count. And that even if > > !CONFIG_PREEMPT. > > Ah, of course! I keep forgetting that CONFIG_DEBUG_ATOMIC_SLEEP selects > CONFIG_PREEMPT_COUNT. > > > > So I queued up a separate patch using rcu_lockdep_assert() to check for > > > illegal RCU grace period within the same-type RCU read-side critical > > > section, including for SRCU. This is also a partial solution, as it > > > does not handle things like this: > > > > > > void foo(void) > > > { > > > mutex_lock(&my_mutex); > > > . . . > > > synchronize_srcu(&my_srcu); > > > . . . > > > mutex_unlock(&my_mutex); > > > } > > > > > > void bar(void) > > > { > > > int idx; > > > > > > idx = rcu_read_lock(&m_srcu); > > > . . . > > > mutex_lock(&my_mutex); > > > . . . > > > mutex_unlock(&my_mutex); > > > . . . > > > srcu_read_unlock(&m_srcu, idx); > > > } > > > > > > This can be extended into a chain of locks and a chain of SRCU instances. > > > For an example of the latter, consider an SRCU-A read-side critical > > > section containing an SRCU-B grace period, an SRCU-B read-side critical > > > section containing an SRCU-C grace period, and so on, with the SRCU-Z > > > read-side critical section containing an RCU-A grace period. > > > > Heh! Indeed... > > > > > But it > > > is OK to hold a mutex across one SRCU read-side critical section while > > > acquiring that same mutex within another same-flavor SRCU read-side > > > critical section. So the analogy with reader-writer locking only goes > > > so far. > > > > > > At the moment, a full solution seems to require some surgery on lockdep > > > itself, but perhaps there is a better way. > > > > Ok. > > > > > > > > > rcutiny seems to be fine with the cond_resched() call, but srcu needs > > > > a special treatment. > > > > > > For the moment, I just applied rcu_lockdep_assert() everywhere -- zero > > > cost on non-lockdep kernels, and fully handles all of the RCU simple > > > self-deadlock cases. > > > > So, for RCU I'm not sure this is useful given the might_sleep() things. > > But for srcu it is. > > One nice thing about the lockdep approach is that it tracks where the > conflicting RCU read-side critical section started. But I am planning > for these to be 3.4 material, so we do have some time to refine them.
Yeah sure. And in any case it's still good to keep might_sleep() early to spot other sources of illegal atomic sections (irqs disabled and co)
| |