| Message ID | 20221207085752.82458-6-michaelgur@nvidia.com (mailing list archive) |
|---|---|
| State | Superseded |
| Headers | show |
| Series | RDMA/mlx5: Switch MR cache to use RB-tree | expand |
On Wed, Dec 07, 2022 at 10:57:51AM +0200, Michael Guralnik wrote: > Currently, when dereging an MR, if the mkey doesn't belong to a cache > entry, it will be destroyed. > As a result, the restart of applications with many non-cached mkeys is > not efficient since all the mkeys are destroyed and then recreated. > This process takes a long time (for 100,000 MRs, it is ~20 seconds for > dereg and ~28 seconds for re-reg). > > To shorten the restart runtime, insert all cacheable mkeys to the cache. > If there is no fitting entry to the mkey properties, create a temporary > entry that fits it. > > After a predetermined timeout, the cache entries will shrink to the > initial high limit. > > The mkeys will still be in the cache when consuming them again after an > application restart. Therefore, the registration will be much faster > (for 100,000 MRs, it is ~4 seconds for dereg and ~5 seconds for re-reg). > > The temporary cache entries created to store the non-cache mkeys are not > exposed through sysfs like the default cache entries. > > Signed-off-by: Michael Guralnik <michaelgur@nvidia.com> > --- > drivers/infiniband/hw/mlx5/mlx5_ib.h | 24 ++++++------ > drivers/infiniband/hw/mlx5/mr.c | 55 +++++++++++++++++++++------- > 2 files changed, 55 insertions(+), 24 deletions(-) > > diff --git a/drivers/infiniband/hw/mlx5/mlx5_ib.h b/drivers/infiniband/hw/mlx5/mlx5_ib.h > index be6d9ec5b127..8f0faa6bc9b5 100644 > --- a/drivers/infiniband/hw/mlx5/mlx5_ib.h > +++ b/drivers/infiniband/hw/mlx5/mlx5_ib.h > @@ -617,12 +617,25 @@ enum mlx5_mkey_type { > MLX5_MKEY_INDIRECT_DEVX, > }; > > +struct mlx5r_cache_rb_key { > + u8 ats:1; > + unsigned int access_mode; > + unsigned int access_flags; > + /* > + * keep ndescs as the last member so entries with about the same ndescs > + * will be close in the tree > + */ > + unsigned int ndescs; > +}; > + > struct mlx5_ib_mkey { > u32 key; > enum mlx5_mkey_type type; > unsigned int ndescs; > struct wait_queue_head wait; > refcount_t usecount; > + /* User Mkey must hold either a cache_key or a cache_ent. */ > + struct mlx5r_cache_rb_key rb_key; What is a cache_key? Why do we now have ndecs and rb_key.ndescs in the same struct? > struct mlx5_cache_ent *cache_ent; > }; > > @@ -731,17 +744,6 @@ struct umr_common { > unsigned int state; > }; > > -struct mlx5r_cache_rb_key { > - u8 ats:1; > - unsigned int access_mode; > - unsigned int access_flags; > - /* > - * keep ndescs as the last member so entries with about the same ndescs > - * will be close in the tree > - */ > - unsigned int ndescs; > -}; Don't move this, put it where it needs to be in the earlier patch > diff --git a/drivers/infiniband/hw/mlx5/mr.c b/drivers/infiniband/hw/mlx5/mr.c > index 6531e38ef4ec..2e984d436ad5 100644 > --- a/drivers/infiniband/hw/mlx5/mr.c > +++ b/drivers/infiniband/hw/mlx5/mr.c > @@ -1096,15 +1096,14 @@ static struct mlx5_ib_mr *alloc_cacheable_mr(struct ib_pd *pd, > rb_key.access_flags = get_unchangeable_access_flags(dev, access_flags); > ent = mkey_cache_ent_from_rb_key(dev, rb_key); > /* > - * Matches access in alloc_cache_mr(). If the MR can't come from the > - * cache then synchronously create an uncached one. > + * If the MR can't come from the cache then synchronously create an uncached > + * one. > */ > - if (!ent || ent->limit == 0 || > - !mlx5r_umr_can_reconfig(dev, 0, access_flags) || > - mlx5_umem_needs_ats(dev, umem, access_flags)) { > + if (!ent) { > mutex_lock(&dev->slow_path_mutex); > mr = reg_create(pd, umem, iova, access_flags, page_size, false); > mutex_unlock(&dev->slow_path_mutex); > + mr->mmkey.rb_key = rb_key; > return mr; > } Does this belong in this patch? Maybe these cleanups need their own patch Jason
diff --git a/drivers/infiniband/hw/mlx5/mlx5_ib.h b/drivers/infiniband/hw/mlx5/mlx5_ib.h index be6d9ec5b127..8f0faa6bc9b5 100644 --- a/drivers/infiniband/hw/mlx5/mlx5_ib.h +++ b/drivers/infiniband/hw/mlx5/mlx5_ib.h @@ -617,12 +617,25 @@ enum mlx5_mkey_type { MLX5_MKEY_INDIRECT_DEVX, }; +struct mlx5r_cache_rb_key { + u8 ats:1; + unsigned int access_mode; + unsigned int access_flags; + /* + * keep ndescs as the last member so entries with about the same ndescs + * will be close in the tree + */ + unsigned int ndescs; +}; + struct mlx5_ib_mkey { u32 key; enum mlx5_mkey_type type; unsigned int ndescs; struct wait_queue_head wait; refcount_t usecount; + /* User Mkey must hold either a cache_key or a cache_ent. */ + struct mlx5r_cache_rb_key rb_key; struct mlx5_cache_ent *cache_ent; }; @@ -731,17 +744,6 @@ struct umr_common { unsigned int state; }; -struct mlx5r_cache_rb_key { - u8 ats:1; - unsigned int access_mode; - unsigned int access_flags; - /* - * keep ndescs as the last member so entries with about the same ndescs - * will be close in the tree - */ - unsigned int ndescs; -}; - struct mlx5_cache_ent { struct xarray mkeys; unsigned long stored; diff --git a/drivers/infiniband/hw/mlx5/mr.c b/drivers/infiniband/hw/mlx5/mr.c index 6531e38ef4ec..2e984d436ad5 100644 --- a/drivers/infiniband/hw/mlx5/mr.c +++ b/drivers/infiniband/hw/mlx5/mr.c @@ -1096,15 +1096,14 @@ static struct mlx5_ib_mr *alloc_cacheable_mr(struct ib_pd *pd, rb_key.access_flags = get_unchangeable_access_flags(dev, access_flags); ent = mkey_cache_ent_from_rb_key(dev, rb_key); /* - * Matches access in alloc_cache_mr(). If the MR can't come from the - * cache then synchronously create an uncached one. + * If the MR can't come from the cache then synchronously create an uncached + * one. */ - if (!ent || ent->limit == 0 || - !mlx5r_umr_can_reconfig(dev, 0, access_flags) || - mlx5_umem_needs_ats(dev, umem, access_flags)) { + if (!ent) { mutex_lock(&dev->slow_path_mutex); mr = reg_create(pd, umem, iova, access_flags, page_size, false); mutex_unlock(&dev->slow_path_mutex); + mr->mmkey.rb_key = rb_key; return mr; } @@ -1195,6 +1194,7 @@ static struct mlx5_ib_mr *reg_create(struct ib_pd *pd, struct ib_umem *umem, goto err_2; } mr->mmkey.type = MLX5_MKEY_MR; + mr->mmkey.ndescs = get_octo_len(iova, umem->length, mr->page_shift); mr->umem = umem; set_mr_fields(dev, mr, umem->length, access_flags, iova); kvfree(in); @@ -1732,6 +1732,40 @@ mlx5_free_priv_descs(struct mlx5_ib_mr *mr) } } +static int cache_ent_find_and_store(struct mlx5_ib_dev *dev, + struct mlx5_ib_mr *mr) +{ + struct mlx5_mkey_cache *cache = &dev->cache; + struct mlx5_cache_ent *ent; + + if (mr->mmkey.cache_ent) { + xa_lock_irq(&mr->mmkey.cache_ent->mkeys); + mr->mmkey.cache_ent->in_use--; + xa_unlock_irq(&mr->mmkey.cache_ent->mkeys); + goto end; + } + + mutex_lock(&cache->rb_lock); + ent = mkey_cache_ent_from_rb_key(dev, mr->mmkey.rb_key); + mutex_unlock(&cache->rb_lock); + if (ent) { + if (ent->rb_key.ndescs == mr->mmkey.rb_key.ndescs) { + mr->mmkey.cache_ent = ent; + goto end; + } + } + + ent = mlx5r_cache_create_ent(dev, mr->mmkey.rb_key, false); + if (IS_ERR(ent)) + return PTR_ERR(ent); + + mr->mmkey.cache_ent = ent; + +end: + return push_mkey(mr->mmkey.cache_ent, false, + xa_mk_value(mr->mmkey.key)); +} + int mlx5_ib_dereg_mr(struct ib_mr *ibmr, struct ib_udata *udata) { struct mlx5_ib_mr *mr = to_mmr(ibmr); @@ -1777,16 +1811,11 @@ int mlx5_ib_dereg_mr(struct ib_mr *ibmr, struct ib_udata *udata) } /* Stop DMA */ - if (mr->mmkey.cache_ent) { - xa_lock_irq(&mr->mmkey.cache_ent->mkeys); - mr->mmkey.cache_ent->in_use--; - xa_unlock_irq(&mr->mmkey.cache_ent->mkeys); - + if (mr->umem && mlx5r_umr_can_load_pas(dev, mr->umem->length)) if (mlx5r_umr_revoke_mr(mr) || - push_mkey(mr->mmkey.cache_ent, false, - xa_mk_value(mr->mmkey.key))) + cache_ent_find_and_store(dev, mr)) mr->mmkey.cache_ent = NULL; - } + if (!mr->mmkey.cache_ent) { rc = destroy_mkey(to_mdev(mr->ibmr.device), mr); if (rc)
Currently, when dereging an MR, if the mkey doesn't belong to a cache entry, it will be destroyed. As a result, the restart of applications with many non-cached mkeys is not efficient since all the mkeys are destroyed and then recreated. This process takes a long time (for 100,000 MRs, it is ~20 seconds for dereg and ~28 seconds for re-reg). To shorten the restart runtime, insert all cacheable mkeys to the cache. If there is no fitting entry to the mkey properties, create a temporary entry that fits it. After a predetermined timeout, the cache entries will shrink to the initial high limit. The mkeys will still be in the cache when consuming them again after an application restart. Therefore, the registration will be much faster (for 100,000 MRs, it is ~4 seconds for dereg and ~5 seconds for re-reg). The temporary cache entries created to store the non-cache mkeys are not exposed through sysfs like the default cache entries. Signed-off-by: Michael Guralnik <michaelgur@nvidia.com> --- drivers/infiniband/hw/mlx5/mlx5_ib.h | 24 ++++++------ drivers/infiniband/hw/mlx5/mr.c | 55 +++++++++++++++++++++------- 2 files changed, 55 insertions(+), 24 deletions(-)