On 07/29/2015 12:14 AM, Paolo Bonzini wrote: > > > On 28/07/2015 15:29, Wen Congyang wrote: >>>> >>> >>> If you call it just once, you have the same problem as before. In fact, >>> it's worse because instead of having an overflow every 2^31 periods, you >>> have one every 2 periods. Instead, by checking that rcu_reader went >>> through 1 _and_ 3 (or that it was at least once 0, i.e. the thread was >>> quiescent), you are sure that the thread went through _at least one_ >>> grace period. >> >> The overflow is acceptable. We only compare if rcu_reader.ctr is equal than >> rcu_gp_ctr. If not, we should wait that thread to call rcu_read_unlock(). >> We don't care which is bigger. If no threads calls sync_rcu(), all threads >> rcu_read.ctr is 0 or rcu_gp_ctr. If one thread calls sync_rcu(), all >> threads rcu_read.ctr is 0, old_rcu_gp_ctr, or new_rcu_gp_ctr. We only wait >> the >> thread that's rcu_read.ctr is old_rcu_gp_ctr. > > Suppose you have > > reader writer > --------------------------------------------------------------- > rcu_read_lock() > ctr = atomic_read(&rcu_gp_ctr); > ctr = 1 > synchronize_rcu() > rcu_gp_ctr = 3 > atomic_xchg(&p_rcu_reader->ctr, ctr); > rcu_reader.ctr = 1 > > We're in the critical section. > > synchronize_rcu() > rcu_gp_ctr = 1 > > rcu_gp_ongoing(&p_rcu_reader->ctr) returns false, so synchronize_rcu() > exits. But we're still in the critical section. > > p = atomic_rcu_read(&foo); > g_free(p); > > Boom. :) For more information see > http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.220.8810
Thanks for your explanation. I understand it now. If we use rcu_gp_ctr + RCU_GP_CTR for 32-bit host, the problem only exists when we call synchronize_rcu() 2^31 times while another thread calls rcu_read_lock(). Thanks Wen Congyang > > Paolo >
