On Wed, Apr 06, 2022, Peter Gonda wrote: > On Wed, Apr 6, 2022 at 12:26 PM Sean Christopherson <seanjc@xxxxxxxxxx> wrote: > > > > On Wed, Apr 06, 2022, Mingwei Zhang wrote: > > > Hi Sean, > > > > > > > > > diff --git a/arch/x86/kvm/svm/sev.c b/arch/x86/kvm/svm/sev.c > > > > > > index 75fa6dd268f0..c2fe89ecdb2d 100644 > > > > > > --- a/arch/x86/kvm/svm/sev.c > > > > > > +++ b/arch/x86/kvm/svm/sev.c > > > > > > @@ -465,6 +465,7 @@ static void sev_clflush_pages(struct page *pages[], unsigned long npages) > > > > > > page_virtual = kmap_atomic(pages[i]); > > > > > > clflush_cache_range(page_virtual, PAGE_SIZE); > > > > > > kunmap_atomic(page_virtual); > > > > > > + cond_resched(); > > > > > > > > > > If you add cond_resched() here, the frequency (once per 4K) might be > > > > > too high. You may want to do it once per X pages, where X could be > > > > > something like 1G/4K? > > > > > > > > No, every iteration is perfectly ok. The "cond"itional part means that this will > > > > reschedule if and only if it actually needs to be rescheduled, e.g. if the task's > > > > timeslice as expired. The check for a needed reschedule is cheap, using > > > > cond_resched() in tight-ish loops is ok and intended, e.g. KVM does a reched > > > > check prior to enterring the guest. > > > > > > Double check on the code again. I think the point is not about flag > > > checking. Obviously branch prediction could really help. The point I > > > think is the 'call' to cond_resched(). Depending on the kernel > > > configuration, cond_resched() may not always be inlined, at least this > > > is my understanding so far? So if that is true, then it still might > > > not always be the best to call cond_resched() that often. > > > > Eh, compared to the cost of 64 back-to-back CLFLUSHOPTs, the cost of __cond_resched() > > is peanuts. Even accounting for the rcu_all_qs() work, it's still dwarfed by the > > cost of flushing data from the cache. E.g. based on Agner Fog's wonderful uop > > latencies[*], the actual flush time for a single page is going to be upwards of > > 10k cycles, whereas __cond_resched() is going to well under 100 cycles in the happy > > case of no work. Even if those throughput numbers are off by an order of magnitude, > > e.g. CLFLUSHOPT can complete in 15 cycles, that's still ~1k cycles. > > > > Peter, don't we also theoretically need cond_resched() in the loops in > > sev_launch_update_data()? AFAICT, there's no articifical restriction on the size > > of the payload, i.e. the kernel is effectively relying on userspace to not update > > large swaths of memory. > > Yea we probably do want to cond_resched() in the for loop inside of > sev_launch_update_data(). Ithink in sev_dbg_crypt() userspace could > request a large number of pages to be decrypted/encrypted for > debugging but se have a call to sev_pin_memory() in the loop so that > will have a cond_resded() inside of __get_users_pages(). Or should we > have a cond_resded() inside of the loop in sev_dbg_crypt() too? I believe sev_dbg_crypt() needs a cond_resched() of its own, sev_pin_memory() isn't guaranteed to get into the slow path of internal_get_user_pages_fast().
