On 26/06/2018 01:23, Junaid Shahid wrote: >> >> 2) /* >> * Read sp->unsync before the spinlock. sp->unsync is written in >> * kvm_mmu->page_fault with mmu_lock held; this barrier pairs with >> * spin_lock's acquire barrier, which guarantees that the spinlock >> * is written before sp->unsync. >> */ >> smp_rmb(); >> >> Feel free to make it less verbose. :) >> > Hmm. I am not really convinced of the necessity of an smp_rmb() > between the sp->unsync test and the spin_lock (I assume that is where > you mean to place the smp_rmb, correct?), though a READ_ONCE of > sp->unsync may be warranted. To illustrate, consider that even if the > read of sp->unsync were to happen after the spin_lock, would it > really be any different than what used to happen before this change? I don't know, the loop in spin_lock probably saves it, but it's nevertheless the tricky control-dependency case. Since it's basically zero cost, I would feel safer adding a smp_rmb (or smp_load_acquire). Paolo
