Re: [Qemu-devel] [PATCH v5 12/18] qht: QEMU's fast, resizable and scalable Hash Table

[Emilio G. Cota]

On Fri, May 20, 2016 at 22:48:11 -0400, Emilio G. Cota wrote:
> On Sat, May 21, 2016 at 01:13:20 +0300, Sergey Fedorov wrote:
> > > +static inline
> > > +void *qht_do_lookup(struct qht_bucket *head, qht_lookup_func_t func,
> > > +                    const void *userp, uint32_t hash)
> > > +{
> > > +    struct qht_bucket *b = head;
> > > +    int i;
> > > +
> > > +    do {
> > > +        for (i = 0; i < QHT_BUCKET_ENTRIES; i++) {
> > > +            if (atomic_read(&b->hashes[i]) == hash) {
> > > +                void *p = atomic_read(&b->pointers[i]);
> > 
> > Why do we need this atomic_read() and other (looking a bit inconsistent)
> > atomic operations on 'b->pointers' and 'b->hash'? if we always have to
> > access them protected properly by a seqlock together with a spinlock?
> 
> [ There should be consistency: read accesses use the atomic ops to read,
>   while write accesses have acquired the bucket lock so don't need them.
>   Well, they need care when they write, since there may be concurrent
>   readers. ]
> 
> I'm using atomic_read but what I really want is ACCESS_ONCE. That is:
> (1) Make sure that the accesses are done in a single instruction (even
>     though gcc doesn't explicitly guarantee it even to aligned addresses
>     anymore[1])
> (2) Make sure the pointer value is only read once, and never refetched.
>     This is what comes right after the pointer is read:
> > +                if (likely(p) && likely(func(p, userp))) {
> > +                    return p;
> > +                }
>     Refetching the pointer value might result in us passing something
>     a NULL p value to the comparison function (since there may be
>     concurrent updaters!), with an immediate segfault. See [2] for a
>     discussion on this (essentially the compiler assumes that there's
>     only a single thread).
> 
> Given that even reading a garbled hash is OK (we don't really need (1),
> since the seqlock will make us retry anyway), I've changed the code to:
> 
>          for (i = 0; i < QHT_BUCKET_ENTRIES; i++) {
> -            if (atomic_read(&b->hashes[i]) == hash) {
> +            if (b->hashes[i] == hash) {
> +                /* make sure the pointer is read only once */
>                  void *p = atomic_read(&b->pointers[i]);
> 
>                  if (likely(p) && likely(func(p, userp))) {
> 
> Performance-wise this is the impact after 10 tries for:
>       $ taskset -c 0 tests/qht-bench \
>         -d 5 -n 1 -u 0 -k 4096 -K 4096 -l 4096 -r 4096 -s 4096
> on my Haswell machine I get, in Mops/s:
>       atomic_read() for all           40.389 +- 0.20888327415622
>       atomic_read(p) only             40.759 +- 0.212835356294224
>       no atomic_read(p) (unsafe)      40.559 +- 0.121422128680622
> 
> Note that the unsafe version is slightly slower; I guess the CPU is trying
> to speculate too much and is gaining little from it.
> 
> [1] "Linux-Kernel Memory Model" by Paul McKenney
>     http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2015/n4374.html
> [2] https://lwn.net/Articles/508991/

A small update: I just got rid of all the atomic_read/set's that
apply to the hashes, since retries will take care of possible races.

The atomic_read/set's remain only for b->pointers[], for the
above reasons.

                E.