Re: [Patch v2 1/3] iommu: match the original algorithm

[Cong Wang <xiyou.wangcong@gmail.com>]

On Fri, Nov 29, 2019 at 6:43 AM John Garry <john.garry@huawei.com> wrote:
>
> On 29/11/2019 00:48, Cong Wang wrote:
> > The IOVA cache algorithm implemented in IOMMU code does not
> > exactly match the original algorithm described in the paper.
> >
>
> which paper?

It's in drivers/iommu/iova.c, from line 769:

 769 /*
 770  * Magazine caches for IOVA ranges.  For an introduction to magazines,
 771  * see the USENIX 2001 paper "Magazines and Vmem: Extending the Slab
 772  * Allocator to Many CPUs and Arbitrary Resources" by Bonwick and Adams.
 773  * For simplicity, we use a static magazine size and don't implement the
 774  * dynamic size tuning described in the paper.
 775  */


>
> > Particularly, it doesn't need to free the loaded empty magazine
> > when trying to put it back to global depot. To make it work, we
> > have to pre-allocate magazines in the depot and only recycle them
> > when all of them are full.
> >
> > Before this patch, rcache->depot[] contains either full or
> > freed entries, after this patch, it contains either full or
> > empty (but allocated) entries.
>
> I *quickly* tested this patch and got a small performance gain.

Thanks for testing! It requires a different workload to see bigger gain,
in our case, 24 memcache.parallel servers with 120 clients.


>
> >
> > Cc: Joerg Roedel <joro@8bytes.org>
> > Signed-off-by: Cong Wang <xiyou.wangcong@gmail.com>
> > ---
> >   drivers/iommu/iova.c | 45 +++++++++++++++++++++++++++-----------------
> >   1 file changed, 28 insertions(+), 17 deletions(-)
> >
> > diff --git a/drivers/iommu/iova.c b/drivers/iommu/iova.c
> > index 41c605b0058f..cb473ddce4cf 100644
> > --- a/drivers/iommu/iova.c
> > +++ b/drivers/iommu/iova.c
> > @@ -862,12 +862,16 @@ static void init_iova_rcaches(struct iova_domain *iovad)
> >       struct iova_cpu_rcache *cpu_rcache;
> >       struct iova_rcache *rcache;
> >       unsigned int cpu;
> > -     int i;
> > +     int i, j;
> >
> >       for (i = 0; i < IOVA_RANGE_CACHE_MAX_SIZE; ++i) {
> >               rcache = &iovad->rcaches[i];
> >               spin_lock_init(&rcache->lock);
> >               rcache->depot_size = 0;
> > +             for (j = 0; j < MAX_GLOBAL_MAGS; ++j) {
> > +                     rcache->depot[j] = iova_magazine_alloc(GFP_KERNEL);
> > +                     WARN_ON(!rcache->depot[j]);
> > +             }
> >               rcache->cpu_rcaches = __alloc_percpu(sizeof(*cpu_rcache), cache_line_size());
> >               if (WARN_ON(!rcache->cpu_rcaches))
> >                       continue;
> > @@ -900,24 +904,30 @@ static bool __iova_rcache_insert(struct iova_domain *iovad,
> >
> >       if (!iova_magazine_full(cpu_rcache->loaded)) {
> >               can_insert = true;
> > -     } else if (!iova_magazine_full(cpu_rcache->prev)) {
> > +     } else if (iova_magazine_empty(cpu_rcache->prev)) {
>
> is this change strictly necessary?

Yes, because it is what described in the paper. But it should be
functionally same because cpu_rcache->prev is either full or empty.



>
> >               swap(cpu_rcache->prev, cpu_rcache->loaded);
> >               can_insert = true;
> >       } else {
> > -             struct iova_magazine *new_mag = iova_magazine_alloc(GFP_ATOMIC);
> > +             spin_lock(&rcache->lock);
> > +             if (rcache->depot_size < MAX_GLOBAL_MAGS) {
> > +                     swap(rcache->depot[rcache->depot_size], cpu_rcache->prev);
> > +                     swap(cpu_rcache->prev, cpu_rcache->loaded);
> > +                     rcache->depot_size++;
> > +                     can_insert = true;
> > +             } else {
> > +                     mag_to_free = cpu_rcache->loaded;
> > +             }
> > +             spin_unlock(&rcache->lock);
> > +
> > +             if (mag_to_free) {
> > +                     struct iova_magazine *new_mag = iova_magazine_alloc(GFP_ATOMIC);
> >
> > -             if (new_mag) {
> > -                     spin_lock(&rcache->lock);
> > -                     if (rcache->depot_size < MAX_GLOBAL_MAGS) {
> > -                             rcache->depot[rcache->depot_size++] =
> > -                                             cpu_rcache->loaded;
> > +                     if (new_mag) {
> > +                             cpu_rcache->loaded = new_mag;
> > +                             can_insert = true;
> >                       } else {
> > -                             mag_to_free = cpu_rcache->loaded;
> > +                             mag_to_free = NULL;
> >                       }
> > -                     spin_unlock(&rcache->lock);
> > -
> > -                     cpu_rcache->loaded = new_mag;
> > -                     can_insert = true;
> >               }
> >       }
> >
> > @@ -963,14 +973,15 @@ static unsigned long __iova_rcache_get(struct iova_rcache *rcache,
> >
> >       if (!iova_magazine_empty(cpu_rcache->loaded)) {
> >               has_pfn = true;
> > -     } else if (!iova_magazine_empty(cpu_rcache->prev)) {
> > +     } else if (iova_magazine_full(cpu_rcache->prev)) {
> >               swap(cpu_rcache->prev, cpu_rcache->loaded);
> >               has_pfn = true;
> >       } else {
> >               spin_lock(&rcache->lock);
> >               if (rcache->depot_size > 0) {
> > -                     iova_magazine_free(cpu_rcache->loaded);
>
> it is good to remove this from under the lock, apart from this change
>
> > -                     cpu_rcache->loaded = rcache->depot[--rcache->depot_size];
> > +                     swap(rcache->depot[rcache->depot_size - 1], cpu_rcache->prev);
> > +                     swap(cpu_rcache->prev, cpu_rcache->loaded);
> > +                     rcache->depot_size--;
>
> I'm not sure how appropriate the name "depot_size" is any longer.

I think it is still okay, because empty ones don't count. However if you
have better names, I am open to your suggestion.

Thanks.