On 08/07/2016 21:55, Sergey Fedorov wrote: > On 08/07/16 17:07, Paolo Bonzini wrote: >> >> On 08/07/2016 14:32, Sergey Fedorov wrote: >>>>>>> I think we can do even better. One option is using a separate tiny lock >>>>>>> to protect direct jump set/reset instead of tb_lock. >>>>> If you have to use a separate tiny lock, you don't gain anything compared >>>>> to the two critical sections, do you? >>> If we have a separate lock for direct jump set/reset then we can do fast >>> TB lookup + direct jump patching without taking tb_lock at all. How much >>> this would reduce lock contention largely depends on the workload we use. >> Yeah, it probably would be easy enough that it's hard to object to it >> (unlike the other idea below, which I'm not very comfortable with, at >> least without seeing patches). >> >> The main advantage would be that this tiny lock could be a spinlock >> rather than a mutex. > > Well, the problem is more subtle than we thought: tb_find_fast() can > race with tb_phys_invalidate(). The first tb_find_phys() out of the lock > can return a TB which is being invalidated. Then a direct jump can be > set up to this TB. It can happen after concurrent tb_phys_invalidate() > resets all the direct jumps to the TB. Thus we can end up with a direct > jump to an invalidated TB. Even extending tb_lock critical section > wouldn't help if at least one tb_find_phys() is performed out of the lock.
Ahem, isn't this exactly why tb_find_phys was invalidating the PC in my patches, as the very first step?... (The smp_wmb after invalidating the PC paired with an atomic_rcu_read in tb_find_fast; now we could do it after computing the hash and before calling qht_remove). It turned out that invalidating the PC wasn't as easy as writing -1 to the pc, but it's possible to do one of these: 1) set cs_base to an invalid value (all-ones works for everything except x86---instead anything nonzero is enough except on x86 and SPARC) 2) set the flags to an invalid combination (x86 can use all ones or rename the useless HF_SOFTMMU_MASK to HF_INVALID_MASK). Paolo