RDMA/mlx5: Switch MR cache to use RB-tree
Subsystems (3)
v5 cover RDMA/mlx5: Switch MR cache to use RB-tree
This series moves the MR cache to use RB tree to store the entries of the cache. By doing so, enabling more flexibility when managing the cache entries.
The MR cache will now cache mkeys returned by the user even if they are not from one of the predefined pools, by that allowing restarting applications to reuse their released mkey and improve restart times.
v4->v5:
- Commit message fix: 'Remove implicit ODP' instead of 'explicit'
- Fix return value of init function in case of no ODP in configuration
v3->v4:
- remove 'change-id' and 'issue' git trailers
v2->v3:
- Refactor MR cache init flow
- Move rb_key decleration to rome unnecessary change in following patches
v1->v2:
- Rearrange patch order to first introduce the RB-tree and only then introduce the caching of previously non-cachable mkeys
v0->v1:
- Fix rb tree search from memcmp to dedicated cmp function
- Rewording of some commit messages
Aharon Landau (2):
RDMA/mlx5: Don't keep umrable 'page_shift' in cache entries RDMA/mlx5: Remove explicit ODP cache entry
Michael Guralnik (4):
RDMA/mlx5: Change the cache structure to an RB-tree RDMA/mlx5: Introduce mlx5r_cache_rb_key
RDMA/mlx5: Cache all user cacheable mkeys on dereg MR flow RDMA/mlx5: Add work to remove temporary entries from the cache
drivers/infiniband/hw/mlx5/mlx5_ib.h | 38 ++- drivers/infiniband/hw/mlx5/mr.c | 478 +++++++++++++++++++++------ drivers/infiniband/hw/mlx5/odp.c | 40 +-- include/linux/mlx5/driver.h | 1 - 4 files changed, 417 insertions(+), 140 deletions(-)
Signature · 1 line
2.17.2
PATCH v5 1/6 RDMA/mlx5: Don't keep umrable 'page_shift' in cache entries
From: Aharon Landau <[email protected]>
mkc.log_page_size can be changed using UMR. Therefore, don't treat it as a cache entry property.
Removing it from struct mlx5_cache_ent.
All cache mkeys will be created with default PAGE_SHIFT, and updated with the needed page_shift using UMR when passing them to a user.
--- drivers/infiniband/hw/mlx5/mlx5_ib.h | 1 - drivers/infiniband/hw/mlx5/mr.c | 3 +-- drivers/infiniband/hw/mlx5/odp.c | 2 -- 3 files changed, 1 insertion(+), 5 deletions(-)
drivers/infiniband/hw/mlx5/mlx5_ib.h +0 -1
@@ -739,7 +739,6 @@ struct mlx5_cache_ent {char name[4 ];u32 order;u32 access_mode; -u32 page;unsigned int ndescs; u8 disabled:1;
drivers/infiniband/hw/mlx5/mr.c +1 -2
@@ -297,7 +297,7 @@ static void set_cache_mkc(struct mlx5_cache_ent *ent, void *mkc)MLX5_SET (mkc, mkc, translations_octword_size,get_mkc_octo_size (ent->access_mode, ent->ndescs)); -MLX5_SET (mkc, mkc, log_page_size, ent->page); +MLX5_SET (mkc, mkc, log_page_size,PAGE_SHIFT ); }/* Asynchronously schedule new MRs to be populated in the cache. */
@@ -765,7 +765,6 @@ int mlx5_mkey_cache_init(struct mlx5_ib_dev *dev)if (ent->order >mkey_cache_max_order (dev))continue ; - ent->page =PAGE_SHIFT ; ent->ndescs =1 << ent->order; ent->access_mode =MLX5_MKC_ACCESS_MODE_MTT ; if ((dev->mdev->profile.mask & MLX5_PROF_MASK_MR_CACHE) &&
drivers/infiniband/hw/mlx5/odp.c +0 -2
@@ -1594,14 +1594,12 @@ void mlx5_odp_init_mkey_cache_entry(struct mlx5_cache_ent *ent)switch (ent->order -2 ) {case MLX5_IMR_MTT_CACHE_ENTRY : - ent->page =PAGE_SHIFT ; ent->ndescs =MLX5_IMR_MTT_ENTRIES ; ent->access_mode =MLX5_MKC_ACCESS_MODE_MTT ; ent->limit =0 ;break ;case MLX5_IMR_KSM_CACHE_ENTRY : - ent->page =MLX5_KSM_PAGE_SHIFT ; ent->ndescs = mlx5_imr_ksm_entries; ent->access_mode =MLX5_MKC_ACCESS_MODE_KSM ; ent->limit =0 ;
Signature · 1 line
2.17.2
PATCH v5 2/6 RDMA/mlx5: Remove implicit ODP cache entry
In reply to Michael GuralnikFrom: Aharon Landau <[email protected]>
Implicit ODP mkey doesn't have unique properties. It shares the same properties as the order 18 cache entry. There is no need to devote a special entry for that.
--- drivers/infiniband/hw/mlx5/odp.c | 20 +++++--------------- include/linux/mlx5/driver.h | 1 - 2 files changed, 5 insertions(+), 16 deletions(-)
drivers/infiniband/hw/mlx5/odp.c +5 -15
@@ -405,6 +405,7 @@ static void mlx5_ib_page_fault_resume(struct mlx5_ib_dev *dev,static struct mlx5_ib_mr *implicit_get_child_mr (struct mlx5_ib_mr *imr,unsigned long idx) { +int order =order_base_2 (MLX5_IMR_MTT_ENTRIES );struct mlx5_ib_dev *dev =mr_to_mdev (imr);struct ib_umem_odp *odp;struct mlx5_ib_mr *mr;
@@ -417,7 +418,8 @@ static struct mlx5_ib_mr *implicit_get_child_mr(struct mlx5_ib_mr *imr,if (IS_ERR (odp))return ERR_CAST (odp); - mr =mlx5_mr_cache_alloc (dev, &dev->cache.ent[MLX5_IMR_MTT_CACHE_ENTRY ], +BUILD_BUG_ON (order >MKEY_CACHE_LAST_STD_ENTRY ); + mr =mlx5_mr_cache_alloc (dev, &dev->cache.ent[order], imr->access_flags);if (IS_ERR (mr)) {ib_umem_odp_release (odp);
@@ -1591,20 +1593,8 @@ void mlx5_odp_init_mkey_cache_entry(struct mlx5_cache_ent *ent) {if (!(ent->dev->odp_caps.general_caps &IB_ODP_SUPPORT_IMPLICIT ))return ; - -switch (ent->order -2 ) { -case MLX5_IMR_MTT_CACHE_ENTRY : - ent->ndescs =MLX5_IMR_MTT_ENTRIES ; - ent->access_mode =MLX5_MKC_ACCESS_MODE_MTT ; - ent->limit =0 ; -break ; - -case MLX5_IMR_KSM_CACHE_ENTRY : - ent->ndescs = mlx5_imr_ksm_entries; - ent->access_mode =MLX5_MKC_ACCESS_MODE_KSM ; - ent->limit =0 ; -break ; - } + ent->ndescs = mlx5_imr_ksm_entries; + ent->access_mode =MLX5_MKC_ACCESS_MODE_KSM ; }static const struct ib_device_ops mlx5_ib_dev_odp_ops = {
include/linux/mlx5/driver.h +0 -1
@@ -734,7 +734,6 @@ enum {enum {MKEY_CACHE_LAST_STD_ENTRY =20 , -MLX5_IMR_MTT_CACHE_ENTRY ,MLX5_IMR_KSM_CACHE_ENTRY ,MAX_MKEY_CACHE_ENTRIES };
Signature · 1 line
2.17.2
PATCH v5 3/6 RDMA/mlx5: Change the cache structure to an RB-tree
In reply to Michael GuralnikCurrently, the cache structure is a static linear array. Therefore, his size is limited to the number of entries in it and is not expandable. The entries are dedicated to mkeys of size 2^x and no access_flags. Mkeys with different properties are not cacheable.
In this patch, we change the cache structure to an RB-tree. This will allow to extend the cache to support more entries with different mkey properties.
--- drivers/infiniband/hw/mlx5/mlx5_ib.h | 11 +- drivers/infiniband/hw/mlx5/mr.c | 160 ++++++++++++++++++++------- drivers/infiniband/hw/mlx5/odp.c | 8 +- 3 files changed, 132 insertions(+), 47 deletions(-)
drivers/infiniband/hw/mlx5/mlx5_ib.h +9 -2
@@ -741,6 +741,8 @@ struct mlx5_cache_ent {u32 access_mode;unsigned int ndescs; +struct rb_node node; + u8 disabled:1; u8 fill_to_high_water:1;
@@ -770,8 +772,9 @@ struct mlx5r_async_create_mkey {struct mlx5_mkey_cache {struct workqueue_struct *wq; -struct mlx5_cache_ent ent[MAX_MKEY_CACHE_ENTRIES ]; -struct dentry *root; +struct rb_root rb_root; +struct mutex rb_lock; +struct dentry *fs_root;unsigned long last_add; };
@@ -1316,11 +1319,15 @@ void mlx5_ib_copy_pas(u64 *old, u64 *new, int step, int num);int mlx5_ib_get_cqe_size (struct ib_cq *ibcq);int mlx5_mkey_cache_init (struct mlx5_ib_dev *dev);int mlx5_mkey_cache_cleanup (struct mlx5_ib_dev *dev); +struct mlx5_cache_ent *mlx5r_cache_create_ent (struct mlx5_ib_dev *dev, +int order);struct mlx5_ib_mr *mlx5_mr_cache_alloc (struct mlx5_ib_dev *dev,struct mlx5_cache_ent *ent,int access_flags); +struct mlx5_ib_mr *mlx5_mr_cache_alloc_order (struct mlx5_ib_dev *dev,u32 order, +int access_flags);int mlx5_ib_check_mr_status (struct ib_mr *ibmr,u32 check_mask,struct ib_mr_status *mr_status); struct ib_wq *mlx5_ib_create_wq(struct ib_pd *pd,
drivers/infiniband/hw/mlx5/mr.c +120 -40
240 lines not on this page · raw message
drivers/infiniband/hw/mlx5/odp.c +3 -5
@@ -419,8 +419,7 @@ static struct mlx5_ib_mr *implicit_get_child_mr(struct mlx5_ib_mr *imr,return ERR_CAST (odp);BUILD_BUG_ON (order >MKEY_CACHE_LAST_STD_ENTRY ); - mr =mlx5_mr_cache_alloc (dev, &dev->cache.ent[order], - imr->access_flags); + mr =mlx5_mr_cache_alloc_order (dev, order, imr->access_flags);if (IS_ERR (mr)) {ib_umem_odp_release (odp);return mr;
@@ -494,9 +493,8 @@ struct mlx5_ib_mr *mlx5_ib_alloc_implicit_mr(struct mlx5_ib_pd *pd,if (IS_ERR (umem_odp))return ERR_CAST (umem_odp); - imr =mlx5_mr_cache_alloc (dev, - &dev->cache.ent[MLX5_IMR_KSM_CACHE_ENTRY ], - access_flags); + imr =mlx5_mr_cache_alloc_order (dev,MLX5_IMR_KSM_CACHE_ENTRY , + access_flags);if (IS_ERR (imr)) {ib_umem_odp_release (umem_odp);return imr;
Signature · 1 line
2.17.2
PATCH v5 4/6 RDMA/mlx5: Introduce mlx5r_cache_rb_key
In reply to Michael GuralnikSwitch from using the mkey order to using the new struct as the key to the RB tree of cache entries.
The key is all the mkey properties that UMR operations can't modify. Using this key to define the cache entries and to search and create cache mkeys.
--- drivers/infiniband/hw/mlx5/mlx5_ib.h | 27 ++-- drivers/infiniband/hw/mlx5/mr.c | 228 +++++++++++++++++++-------- drivers/infiniband/hw/mlx5/odp.c | 30 ++-- 3 files changed, 201 insertions(+), 84 deletions(-)
drivers/infiniband/hw/mlx5/mlx5_ib.h +17 -10
@@ -617,6 +617,13 @@ enum mlx5_mkey_type { MLX5_MKEY_INDIRECT_DEVX, }; +struct mlx5r_cache_rb_key { +u8 ats:1 ; +unsigned int access_mode; +unsigned int access_flags; +unsigned int ndescs; +}; +struct mlx5_ib_mkey {u32 key;enum mlx5_mkey_type type;
@@ -737,11 +744,9 @@ struct mlx5_cache_ent {unsigned long reserved;char name[4 ]; -u32 order; -u32 access_mode; -unsigned int ndescs;struct rb_node node; +struct mlx5r_cache_rb_key rb_key; u8 disabled:1; u8 fill_to_high_water:1;
@@ -1320,14 +1325,13 @@ int mlx5_ib_get_cqe_size(struct ib_cq *ibcq);int mlx5_mkey_cache_init (struct mlx5_ib_dev *dev);int mlx5_mkey_cache_cleanup (struct mlx5_ib_dev *dev);struct mlx5_cache_ent *mlx5r_cache_create_ent (struct mlx5_ib_dev *dev, -int order); +struct mlx5r_cache_rb_key rb_key, +bool persistent_entry);struct mlx5_ib_mr *mlx5_mr_cache_alloc (struct mlx5_ib_dev *dev, -struct mlx5_cache_ent *ent, -int access_flags); +int access_flags,int access_mode, +int ndescs); -struct mlx5_ib_mr *mlx5_mr_cache_alloc_order (struct mlx5_ib_dev *dev,u32 order, -int access_flags);int mlx5_ib_check_mr_status (struct ib_mr *ibmr,u32 check_mask,struct ib_mr_status *mr_status); struct ib_wq *mlx5_ib_create_wq(struct ib_pd *pd,
@@ -1350,7 +1354,7 @@ int mlx5r_odp_create_eq(struct mlx5_ib_dev *dev, struct mlx5_ib_pf_eq *eq);void mlx5_ib_odp_cleanup_one (struct mlx5_ib_dev *ibdev);int __init mlx5_ib_odp_init (void );void mlx5_ib_odp_cleanup (void ); -void mlx5_odp_init_mkey_cache_entry (struct mlx5_cache_ent *ent); +int mlx5_odp_init_mkey_cache (struct mlx5_ib_dev *dev);void mlx5_odp_populate_xlt (void *xlt,size_t idx,size_t nentries,struct mlx5_ib_mr *mr,int flags);
@@ -1369,7 +1373,10 @@ static inline int mlx5r_odp_create_eq(struct mlx5_ib_dev *dev,static inline void mlx5_ib_odp_cleanup_one (struct mlx5_ib_dev *ibdev) {}static inline int mlx5_ib_odp_init (void ) {return 0 ; }static inline void mlx5_ib_odp_cleanup (void ) {} -static inline void mlx5_odp_init_mkey_cache_entry (struct mlx5_cache_ent *ent) {} +static inline int mlx5_odp_init_mkey_cache (struct mlx5_ib_dev *dev) +{ +return 0 ; +}static inline void mlx5_odp_populate_xlt (void *xlt,size_t idx,size_t nentries,struct mlx5_ib_mr *mr,int flags) {}
drivers/infiniband/hw/mlx5/mr.c +164 -64
382 lines not on this page · raw message
drivers/infiniband/hw/mlx5/odp.c +20 -10
@@ -405,7 +405,6 @@ static void mlx5_ib_page_fault_resume(struct mlx5_ib_dev *dev,static struct mlx5_ib_mr *implicit_get_child_mr (struct mlx5_ib_mr *imr,unsigned long idx) { -int order =order_base_2 (MLX5_IMR_MTT_ENTRIES );struct mlx5_ib_dev *dev =mr_to_mdev (imr);struct ib_umem_odp *odp;struct mlx5_ib_mr *mr;
@@ -418,8 +417,9 @@ static struct mlx5_ib_mr *implicit_get_child_mr(struct mlx5_ib_mr *imr,if (IS_ERR (odp))return ERR_CAST (odp); -BUILD_BUG_ON (order >MKEY_CACHE_LAST_STD_ENTRY ); - mr =mlx5_mr_cache_alloc_order (dev, order, imr->access_flags); + mr =mlx5_mr_cache_alloc (dev, imr->access_flags, +MLX5_MKC_ACCESS_MODE_MTT , +MLX5_IMR_MTT_ENTRIES );if (IS_ERR (mr)) {ib_umem_odp_release (odp);return mr;
@@ -493,8 +493,8 @@ struct mlx5_ib_mr *mlx5_ib_alloc_implicit_mr(struct mlx5_ib_pd *pd,if (IS_ERR (umem_odp))return ERR_CAST (umem_odp); - imr =mlx5_mr_cache_alloc_order (dev,MLX5_IMR_KSM_CACHE_ENTRY , - access_flags); + imr =mlx5_mr_cache_alloc (dev, access_flags,MLX5_MKC_ACCESS_MODE_KSM , + mlx5_imr_ksm_entries);if (IS_ERR (imr)) {ib_umem_odp_release (umem_odp);return imr;
@@ -1587,12 +1587,22 @@ mlx5_ib_odp_destroy_eq(struct mlx5_ib_dev *dev, struct mlx5_ib_pf_eq *eq)return err; } -void mlx5_odp_init_mkey_cache_entry (struct mlx5_cache_ent *ent) +int mlx5_odp_init_mkey_cache (struct mlx5_ib_dev *dev) { -if (!(ent->dev->odp_caps.general_caps &IB_ODP_SUPPORT_IMPLICIT )) -return ; - ent->ndescs = mlx5_imr_ksm_entries; - ent->access_mode =MLX5_MKC_ACCESS_MODE_KSM ; +struct mlx5r_cache_rb_key rb_key = { + .access_mode =MLX5_MKC_ACCESS_MODE_KSM , + .ndescs = mlx5_imr_ksm_entries, + }; +struct mlx5_cache_ent *ent; + +if (!(dev->odp_caps.general_caps &IB_ODP_SUPPORT_IMPLICIT )) +return 0 ; + + ent =mlx5r_cache_create_ent (dev, rb_key, true); +if (IS_ERR (ent)) +return PTR_ERR (ent); + +return 0 ; }static const struct ib_device_ops mlx5_ib_dev_odp_ops = {
Signature · 1 line
2.17.2
PATCH v5 5/6 RDMA/mlx5: Cache all user cacheable mkeys on dereg MR flow
In reply to Michael GuralnikCurrently, 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.
--- drivers/infiniband/hw/mlx5/mlx5_ib.h | 2 + drivers/infiniband/hw/mlx5/mr.c | 55 +++++++++++++++++++++------- 2 files changed, 44 insertions(+), 13 deletions(-)
drivers/infiniband/hw/mlx5/mlx5_ib.h +2 -0
@@ -630,6 +630,8 @@ struct mlx5_ib_mkey {unsigned int ndescs;struct wait_queue_head wait;refcount_t usecount; +/* User Mkey must hold either a rb_key or a cache_ent. */ +struct mlx5r_cache_rb_key rb_key;struct mlx5_cache_ent *cache_ent; };
drivers/infiniband/hw/mlx5/mr.c +42 -13
85 lines not on this page · raw message
Signature · 1 line
2.17.2
PATCH v5 6/6 RDMA/mlx5: Add work to remove temporary entries from the cache
In reply to Michael GuralnikThe non-cache mkeys are stored in the cache only to shorten restarting application time. Don't store them longer than needed.
Configure cache entries that store non-cache MRs as temporary entries. If 30 seconds have passed and no user reclaimed the temporarily cached mkeys, an asynchronous work will destroy the mkeys entries.
--- drivers/infiniband/hw/mlx5/mlx5_ib.h | 9 ++- drivers/infiniband/hw/mlx5/mr.c | 94 ++++++++++++++++++++++------ drivers/infiniband/hw/mlx5/odp.c | 2 +- 3 files changed, 82 insertions(+), 23 deletions(-)
drivers/infiniband/hw/mlx5/mlx5_ib.h +6 -3
@@ -750,6 +750,7 @@ struct mlx5_cache_ent {struct rb_node node;struct mlx5r_cache_rb_key rb_key; + u8 is_tmp:1; u8 disabled:1; u8 fill_to_high_water:1;
@@ -783,6 +784,7 @@ struct mlx5_mkey_cache {struct mutex rb_lock;struct dentry *fs_root;unsigned long last_add; +struct delayed_work remove_ent_dwork; };struct mlx5_ib_port_resources {
@@ -1326,9 +1328,10 @@ void mlx5_ib_copy_pas(u64 *old, u64 *new, int step, int num);int mlx5_ib_get_cqe_size (struct ib_cq *ibcq);int mlx5_mkey_cache_init (struct mlx5_ib_dev *dev);int mlx5_mkey_cache_cleanup (struct mlx5_ib_dev *dev); -struct mlx5_cache_ent *mlx5r_cache_create_ent (struct mlx5_ib_dev *dev, -struct mlx5r_cache_rb_key rb_key, -bool persistent_entry); +struct mlx5_cache_ent * +mlx5r_cache_create_ent_locked (struct mlx5_ib_dev *dev, +struct mlx5r_cache_rb_key rb_key, +bool persistent_entry); struct mlx5_ib_mr *mlx5_mr_cache_alloc(struct mlx5_ib_dev *dev, int access_flags, int access_mode,
drivers/infiniband/hw/mlx5/mr.c +75 -19
218 lines not on this page · raw message
drivers/infiniband/hw/mlx5/odp.c +1 -1
@@ -1598,7 +1598,7 @@ int mlx5_odp_init_mkey_cache(struct mlx5_ib_dev *dev)if (!(dev->odp_caps.general_caps &IB_ODP_SUPPORT_IMPLICIT ))return 0 ; - ent =mlx5r_cache_create_ent (dev, rb_key, true); + ent =mlx5r_cache_create_ent_locked (dev, rb_key, true);if (IS_ERR (ent))return PTR_ERR (ent);
Signature · 1 line
2.17.2
Quoted 32 lines · Michael Guralnik
This series moves the MR cache to use RB tree to store the entries of the cache. By doing so, enabling more flexibility when managing the cache entries.
The MR cache will now cache mkeys returned by the user even if they are not from one of the predefined pools, by that allowing restarting applications to reuse their released mkey and improve restart times.
v4->v5:
- Commit message fix: 'Remove implicit ODP' instead of 'explicit'
- Fix return value of init function in case of no ODP in configuration
v3->v4:
- remove 'change-id' and 'issue' git trailers
v2->v3:
- Refactor MR cache init flow
- Moverb_keydecleration to rome unnecessary change in following patches
v1->v2:
- Rearrange patch order to first introduce the RB-tree and only then introduce the caching of previously non-cachable mkeys
v0->v1:
- Fix rb tree search from memcmp to dedicated cmp function
- Rewording of some commit messagesAharon Landau (2):
RDMA/mlx5: Don't keep umrable 'page_shift' in cache entries RDMA/mlx5: Remove explicit ODP cache entry
Quoted 5 lines
Michael Guralnik (4):
RDMA/mlx5: Change the cache structure to an RB-tree RDMA/mlx5: Introducemlx5r_cache_rb_key
RDMA/mlx5: Cache all user cacheable mkeys on dereg MR flow RDMA/mlx5: Add work to remove temporary entries from the cache
Applied to for-next
Thanks,
Jason