* [PATCH rdma-next v1 0/5] MR cache cleanup
@ 2022-07-26 7:19 Michael Guralnik
2022-07-26 7:19 ` [PATCH rdma-next v1 1/5] RDMA/mlx5: Replace ent->lock with xa_lock Michael Guralnik
` (5 more replies)
0 siblings, 6 replies; 7+ messages in thread
From: Michael Guralnik @ 2022-07-26 7:19 UTC (permalink / raw)
To: jgg
Cc: leonro, maorg, linux-kernel, linux-rdma, netdev, saeedm,
Michael Guralnik
Hi,
In this series, Aharon continues to clean mlx5 MR cache logic.
Thanks
v1: Change push_mkey to eliminate locking on mail flow
v0: http://lore.kernel.org/all/cover.1654601897.git.leonro@nvidia.com/
Aharon Landau (5):
RDMA/mlx5: Replace ent->lock with xa_lock
RDMA/mlx5: Replace cache list with Xarray
RDMA/mlx5: Store the number of in_use cache mkeys instead of total_mrs
RDMA/mlx5: Store in the cache mkeys instead of mrs
RDMA/mlx5: Rename the mkey cache variables and functions
drivers/infiniband/hw/mlx5/main.c | 4 +-
drivers/infiniband/hw/mlx5/mlx5_ib.h | 61 ++--
drivers/infiniband/hw/mlx5/mr.c | 514 +++++++++++++++------------
drivers/infiniband/hw/mlx5/odp.c | 2 +-
include/linux/mlx5/driver.h | 6 +-
5 files changed, 307 insertions(+), 280 deletions(-)
--
2.17.2
^ permalink raw reply [flat|nested] 7+ messages in thread
* [PATCH rdma-next v1 1/5] RDMA/mlx5: Replace ent->lock with xa_lock
2022-07-26 7:19 [PATCH rdma-next v1 0/5] MR cache cleanup Michael Guralnik
@ 2022-07-26 7:19 ` Michael Guralnik
2022-07-26 7:19 ` [PATCH rdma-next v1 2/5] RDMA/mlx5: Replace cache list with Xarray Michael Guralnik
` (4 subsequent siblings)
5 siblings, 0 replies; 7+ messages in thread
From: Michael Guralnik @ 2022-07-26 7:19 UTC (permalink / raw)
To: jgg
Cc: leonro, maorg, linux-kernel, linux-rdma, netdev, saeedm, Aharon Landau
From: Aharon Landau <aharonl@nvidia.com>
In the next patch, ent->list will be replaced with an xarray. The xarray
uses an internal lock to protect the indexes. Use it to protect all the
entry fields, and get rid of ent->lock.
Signed-off-by: Aharon Landau <aharonl@nvidia.com>
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
---
drivers/infiniband/hw/mlx5/mlx5_ib.h | 5 +-
drivers/infiniband/hw/mlx5/mr.c | 92 ++++++++++++++--------------
2 files changed, 47 insertions(+), 50 deletions(-)
diff --git a/drivers/infiniband/hw/mlx5/mlx5_ib.h b/drivers/infiniband/hw/mlx5/mlx5_ib.h
index 688ee7c05a8f..42bc58967b1f 100644
--- a/drivers/infiniband/hw/mlx5/mlx5_ib.h
+++ b/drivers/infiniband/hw/mlx5/mlx5_ib.h
@@ -743,11 +743,8 @@ struct umr_common {
};
struct mlx5_cache_ent {
+ struct xarray mkeys;
struct list_head head;
- /* sync access to the cahce entry
- */
- spinlock_t lock;
-
char name[4];
u32 order;
diff --git a/drivers/infiniband/hw/mlx5/mr.c b/drivers/infiniband/hw/mlx5/mr.c
index aedfd7ff4846..d56e7ff74b98 100644
--- a/drivers/infiniband/hw/mlx5/mr.c
+++ b/drivers/infiniband/hw/mlx5/mr.c
@@ -153,10 +153,10 @@ static void create_mkey_callback(int status, struct mlx5_async_work *context)
if (status) {
create_mkey_warn(dev, status, mr->out);
kfree(mr);
- spin_lock_irqsave(&ent->lock, flags);
+ xa_lock_irqsave(&ent->mkeys, flags);
ent->pending--;
WRITE_ONCE(dev->fill_delay, 1);
- spin_unlock_irqrestore(&ent->lock, flags);
+ xa_unlock_irqrestore(&ent->mkeys, flags);
mod_timer(&dev->delay_timer, jiffies + HZ);
return;
}
@@ -168,14 +168,14 @@ static void create_mkey_callback(int status, struct mlx5_async_work *context)
WRITE_ONCE(dev->cache.last_add, jiffies);
- spin_lock_irqsave(&ent->lock, flags);
+ xa_lock_irqsave(&ent->mkeys, flags);
list_add_tail(&mr->list, &ent->head);
ent->available_mrs++;
ent->total_mrs++;
/* If we are doing fill_to_high_water then keep going. */
queue_adjust_cache_locked(ent);
ent->pending--;
- spin_unlock_irqrestore(&ent->lock, flags);
+ xa_unlock_irqrestore(&ent->mkeys, flags);
}
static int get_mkc_octo_size(unsigned int access_mode, unsigned int ndescs)
@@ -239,23 +239,23 @@ static int add_keys(struct mlx5_cache_ent *ent, unsigned int num)
err = -ENOMEM;
break;
}
- spin_lock_irq(&ent->lock);
+ xa_lock_irq(&ent->mkeys);
if (ent->pending >= MAX_PENDING_REG_MR) {
err = -EAGAIN;
- spin_unlock_irq(&ent->lock);
+ xa_unlock_irq(&ent->mkeys);
kfree(mr);
break;
}
ent->pending++;
- spin_unlock_irq(&ent->lock);
+ xa_unlock_irq(&ent->mkeys);
err = mlx5_ib_create_mkey_cb(ent->dev, &mr->mmkey,
&ent->dev->async_ctx, in, inlen,
mr->out, sizeof(mr->out),
&mr->cb_work);
if (err) {
- spin_lock_irq(&ent->lock);
+ xa_lock_irq(&ent->mkeys);
ent->pending--;
- spin_unlock_irq(&ent->lock);
+ xa_unlock_irq(&ent->mkeys);
mlx5_ib_warn(ent->dev, "create mkey failed %d\n", err);
kfree(mr);
break;
@@ -293,9 +293,9 @@ static struct mlx5_ib_mr *create_cache_mr(struct mlx5_cache_ent *ent)
init_waitqueue_head(&mr->mmkey.wait);
mr->mmkey.type = MLX5_MKEY_MR;
WRITE_ONCE(ent->dev->cache.last_add, jiffies);
- spin_lock_irq(&ent->lock);
+ xa_lock_irq(&ent->mkeys);
ent->total_mrs++;
- spin_unlock_irq(&ent->lock);
+ xa_unlock_irq(&ent->mkeys);
kfree(in);
return mr;
free_mr:
@@ -309,17 +309,17 @@ static void remove_cache_mr_locked(struct mlx5_cache_ent *ent)
{
struct mlx5_ib_mr *mr;
- lockdep_assert_held(&ent->lock);
+ lockdep_assert_held(&ent->mkeys.xa_lock);
if (list_empty(&ent->head))
return;
mr = list_first_entry(&ent->head, struct mlx5_ib_mr, list);
list_del(&mr->list);
ent->available_mrs--;
ent->total_mrs--;
- spin_unlock_irq(&ent->lock);
+ xa_unlock_irq(&ent->mkeys);
mlx5_core_destroy_mkey(ent->dev->mdev, mr->mmkey.key);
kfree(mr);
- spin_lock_irq(&ent->lock);
+ xa_lock_irq(&ent->mkeys);
}
static int resize_available_mrs(struct mlx5_cache_ent *ent, unsigned int target,
@@ -327,7 +327,7 @@ static int resize_available_mrs(struct mlx5_cache_ent *ent, unsigned int target,
{
int err;
- lockdep_assert_held(&ent->lock);
+ lockdep_assert_held(&ent->mkeys.xa_lock);
while (true) {
if (limit_fill)
@@ -337,11 +337,11 @@ static int resize_available_mrs(struct mlx5_cache_ent *ent, unsigned int target,
if (target > ent->available_mrs + ent->pending) {
u32 todo = target - (ent->available_mrs + ent->pending);
- spin_unlock_irq(&ent->lock);
+ xa_unlock_irq(&ent->mkeys);
err = add_keys(ent, todo);
if (err == -EAGAIN)
usleep_range(3000, 5000);
- spin_lock_irq(&ent->lock);
+ xa_lock_irq(&ent->mkeys);
if (err) {
if (err != -EAGAIN)
return err;
@@ -369,7 +369,7 @@ static ssize_t size_write(struct file *filp, const char __user *buf,
* cannot free MRs that are in use. Compute the target value for
* available_mrs.
*/
- spin_lock_irq(&ent->lock);
+ xa_lock_irq(&ent->mkeys);
if (target < ent->total_mrs - ent->available_mrs) {
err = -EINVAL;
goto err_unlock;
@@ -382,12 +382,12 @@ static ssize_t size_write(struct file *filp, const char __user *buf,
err = resize_available_mrs(ent, target, false);
if (err)
goto err_unlock;
- spin_unlock_irq(&ent->lock);
+ xa_unlock_irq(&ent->mkeys);
return count;
err_unlock:
- spin_unlock_irq(&ent->lock);
+ xa_unlock_irq(&ent->mkeys);
return err;
}
@@ -427,10 +427,10 @@ static ssize_t limit_write(struct file *filp, const char __user *buf,
* Upon set we immediately fill the cache to high water mark implied by
* the limit.
*/
- spin_lock_irq(&ent->lock);
+ xa_lock_irq(&ent->mkeys);
ent->limit = var;
err = resize_available_mrs(ent, 0, true);
- spin_unlock_irq(&ent->lock);
+ xa_unlock_irq(&ent->mkeys);
if (err)
return err;
return count;
@@ -465,9 +465,9 @@ static bool someone_adding(struct mlx5_mr_cache *cache)
struct mlx5_cache_ent *ent = &cache->ent[i];
bool ret;
- spin_lock_irq(&ent->lock);
+ xa_lock_irq(&ent->mkeys);
ret = ent->available_mrs < ent->limit;
- spin_unlock_irq(&ent->lock);
+ xa_unlock_irq(&ent->mkeys);
if (ret)
return true;
}
@@ -481,7 +481,7 @@ static bool someone_adding(struct mlx5_mr_cache *cache)
*/
static void queue_adjust_cache_locked(struct mlx5_cache_ent *ent)
{
- lockdep_assert_held(&ent->lock);
+ lockdep_assert_held(&ent->mkeys.xa_lock);
if (ent->disabled || READ_ONCE(ent->dev->fill_delay))
return;
@@ -514,16 +514,16 @@ static void __cache_work_func(struct mlx5_cache_ent *ent)
struct mlx5_mr_cache *cache = &dev->cache;
int err;
- spin_lock_irq(&ent->lock);
+ xa_lock_irq(&ent->mkeys);
if (ent->disabled)
goto out;
if (ent->fill_to_high_water &&
ent->available_mrs + ent->pending < 2 * ent->limit &&
!READ_ONCE(dev->fill_delay)) {
- spin_unlock_irq(&ent->lock);
+ xa_unlock_irq(&ent->mkeys);
err = add_keys(ent, 1);
- spin_lock_irq(&ent->lock);
+ xa_lock_irq(&ent->mkeys);
if (ent->disabled)
goto out;
if (err) {
@@ -556,11 +556,11 @@ static void __cache_work_func(struct mlx5_cache_ent *ent)
* the garbage collection work to try to run in next cycle, in
* order to free CPU resources to other tasks.
*/
- spin_unlock_irq(&ent->lock);
+ xa_unlock_irq(&ent->mkeys);
need_delay = need_resched() || someone_adding(cache) ||
!time_after(jiffies,
READ_ONCE(cache->last_add) + 300 * HZ);
- spin_lock_irq(&ent->lock);
+ xa_lock_irq(&ent->mkeys);
if (ent->disabled)
goto out;
if (need_delay) {
@@ -571,7 +571,7 @@ static void __cache_work_func(struct mlx5_cache_ent *ent)
queue_adjust_cache_locked(ent);
}
out:
- spin_unlock_irq(&ent->lock);
+ xa_unlock_irq(&ent->mkeys);
}
static void delayed_cache_work_func(struct work_struct *work)
@@ -592,11 +592,11 @@ struct mlx5_ib_mr *mlx5_mr_cache_alloc(struct mlx5_ib_dev *dev,
if (!mlx5r_umr_can_reconfig(dev, 0, access_flags))
return ERR_PTR(-EOPNOTSUPP);
- spin_lock_irq(&ent->lock);
+ xa_lock_irq(&ent->mkeys);
if (list_empty(&ent->head)) {
queue_adjust_cache_locked(ent);
ent->miss++;
- spin_unlock_irq(&ent->lock);
+ xa_unlock_irq(&ent->mkeys);
mr = create_cache_mr(ent);
if (IS_ERR(mr))
return mr;
@@ -605,7 +605,7 @@ struct mlx5_ib_mr *mlx5_mr_cache_alloc(struct mlx5_ib_dev *dev,
list_del(&mr->list);
ent->available_mrs--;
queue_adjust_cache_locked(ent);
- spin_unlock_irq(&ent->lock);
+ xa_unlock_irq(&ent->mkeys);
mlx5_clear_mr(mr);
}
@@ -617,11 +617,11 @@ static void mlx5_mr_cache_free(struct mlx5_ib_dev *dev, struct mlx5_ib_mr *mr)
struct mlx5_cache_ent *ent = mr->cache_ent;
WRITE_ONCE(dev->cache.last_add, jiffies);
- spin_lock_irq(&ent->lock);
+ xa_lock_irq(&ent->mkeys);
list_add_tail(&mr->list, &ent->head);
ent->available_mrs++;
queue_adjust_cache_locked(ent);
- spin_unlock_irq(&ent->lock);
+ xa_unlock_irq(&ent->mkeys);
}
static void clean_keys(struct mlx5_ib_dev *dev, int c)
@@ -634,16 +634,16 @@ static void clean_keys(struct mlx5_ib_dev *dev, int c)
cancel_delayed_work(&ent->dwork);
while (1) {
- spin_lock_irq(&ent->lock);
+ xa_lock_irq(&ent->mkeys);
if (list_empty(&ent->head)) {
- spin_unlock_irq(&ent->lock);
+ xa_unlock_irq(&ent->mkeys);
break;
}
mr = list_first_entry(&ent->head, struct mlx5_ib_mr, list);
list_move(&mr->list, &del_list);
ent->available_mrs--;
ent->total_mrs--;
- spin_unlock_irq(&ent->lock);
+ xa_unlock_irq(&ent->mkeys);
mlx5_core_destroy_mkey(dev->mdev, mr->mmkey.key);
}
@@ -710,7 +710,7 @@ int mlx5_mr_cache_init(struct mlx5_ib_dev *dev)
for (i = 0; i < MAX_MR_CACHE_ENTRIES; i++) {
ent = &cache->ent[i];
INIT_LIST_HEAD(&ent->head);
- spin_lock_init(&ent->lock);
+ xa_init_flags(&ent->mkeys, XA_FLAGS_LOCK_IRQ);
ent->order = i + 2;
ent->dev = dev;
ent->limit = 0;
@@ -734,9 +734,9 @@ int mlx5_mr_cache_init(struct mlx5_ib_dev *dev)
ent->limit = dev->mdev->profile.mr_cache[i].limit;
else
ent->limit = 0;
- spin_lock_irq(&ent->lock);
+ xa_lock_irq(&ent->mkeys);
queue_adjust_cache_locked(ent);
- spin_unlock_irq(&ent->lock);
+ xa_unlock_irq(&ent->mkeys);
}
mlx5_mr_cache_debugfs_init(dev);
@@ -754,9 +754,9 @@ int mlx5_mr_cache_cleanup(struct mlx5_ib_dev *dev)
for (i = 0; i < MAX_MR_CACHE_ENTRIES; i++) {
struct mlx5_cache_ent *ent = &dev->cache.ent[i];
- spin_lock_irq(&ent->lock);
+ xa_lock_irq(&ent->mkeys);
ent->disabled = true;
- spin_unlock_irq(&ent->lock);
+ xa_unlock_irq(&ent->mkeys);
cancel_delayed_work_sync(&ent->dwork);
}
@@ -1572,9 +1572,9 @@ int mlx5_ib_dereg_mr(struct ib_mr *ibmr, struct ib_udata *udata)
/* Stop DMA */
if (mr->cache_ent) {
if (mlx5r_umr_revoke_mr(mr)) {
- spin_lock_irq(&mr->cache_ent->lock);
+ xa_lock_irq(&mr->cache_ent->mkeys);
mr->cache_ent->total_mrs--;
- spin_unlock_irq(&mr->cache_ent->lock);
+ xa_unlock_irq(&mr->cache_ent->mkeys);
mr->cache_ent = NULL;
}
}
--
2.17.2
^ permalink raw reply related [flat|nested] 7+ messages in thread
* [PATCH rdma-next v1 2/5] RDMA/mlx5: Replace cache list with Xarray
2022-07-26 7:19 [PATCH rdma-next v1 0/5] MR cache cleanup Michael Guralnik
2022-07-26 7:19 ` [PATCH rdma-next v1 1/5] RDMA/mlx5: Replace ent->lock with xa_lock Michael Guralnik
@ 2022-07-26 7:19 ` Michael Guralnik
2022-07-26 7:19 ` [PATCH rdma-next v1 3/5] RDMA/mlx5: Store the number of in_use cache mkeys instead of total_mrs Michael Guralnik
` (3 subsequent siblings)
5 siblings, 0 replies; 7+ messages in thread
From: Michael Guralnik @ 2022-07-26 7:19 UTC (permalink / raw)
To: jgg
Cc: leonro, maorg, linux-kernel, linux-rdma, netdev, saeedm, Aharon Landau
From: Aharon Landau <aharonl@nvidia.com>
The Xarray allows us to store the cached mkeys in memory efficient way.
Entries are reserved in the Xarray using xa_cmpxchg before calling to
the upcoming callbacks to avoid allocations in interrupt context.
The xa_cmpxchg can sleep when using GFP_KERNEL, so we call it in
a loop to ensure one reserved entry for each process trying to reserve.
Signed-off-by: Aharon Landau <aharonl@nvidia.com>
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
---
drivers/infiniband/hw/mlx5/mlx5_ib.h | 14 +-
drivers/infiniband/hw/mlx5/mr.c | 226 ++++++++++++++++++---------
2 files changed, 152 insertions(+), 88 deletions(-)
diff --git a/drivers/infiniband/hw/mlx5/mlx5_ib.h b/drivers/infiniband/hw/mlx5/mlx5_ib.h
index 42bc58967b1f..e0eb666aefa1 100644
--- a/drivers/infiniband/hw/mlx5/mlx5_ib.h
+++ b/drivers/infiniband/hw/mlx5/mlx5_ib.h
@@ -651,8 +651,6 @@ struct mlx5_ib_mr {
struct {
u32 out[MLX5_ST_SZ_DW(create_mkey_out)];
struct mlx5_async_work cb_work;
- /* Cache list element */
- struct list_head list;
};
/* Used only by kernel MRs (umem == NULL) */
@@ -744,7 +742,8 @@ struct umr_common {
struct mlx5_cache_ent {
struct xarray mkeys;
- struct list_head head;
+ unsigned long stored;
+ unsigned long reserved;
char name[4];
u32 order;
@@ -756,18 +755,13 @@ struct mlx5_cache_ent {
u8 fill_to_high_water:1;
/*
- * - available_mrs is the length of list head, ie the number of MRs
- * available for immediate allocation.
- * - total_mrs is available_mrs plus all in use MRs that could be
+ * - total_mrs is stored mkeys plus all in use MRs that could be
* returned to the cache.
- * - limit is the low water mark for available_mrs, 2* limit is the
+ * - limit is the low water mark for stored mkeys, 2* limit is the
* upper water mark.
- * - pending is the number of MRs currently being created
*/
u32 total_mrs;
- u32 available_mrs;
u32 limit;
- u32 pending;
/* Statistics */
u32 miss;
diff --git a/drivers/infiniband/hw/mlx5/mr.c b/drivers/infiniband/hw/mlx5/mr.c
index d56e7ff74b98..cbb8882c7787 100644
--- a/drivers/infiniband/hw/mlx5/mr.c
+++ b/drivers/infiniband/hw/mlx5/mr.c
@@ -142,6 +142,104 @@ static void create_mkey_warn(struct mlx5_ib_dev *dev, int status, void *out)
mlx5_cmd_out_err(dev->mdev, MLX5_CMD_OP_CREATE_MKEY, 0, out);
}
+
+static int push_mkey(struct mlx5_cache_ent *ent, bool limit_pendings,
+ void *to_store)
+{
+ XA_STATE(xas, &ent->mkeys, 0);
+ void *curr;
+
+ xa_lock_irq(&ent->mkeys);
+ if (limit_pendings &&
+ (ent->reserved - ent->stored) > MAX_PENDING_REG_MR) {
+ xa_unlock_irq(&ent->mkeys);
+ return -EAGAIN;
+ }
+ while (1) {
+ /*
+ * This is cmpxchg (NULL, XA_ZERO_ENTRY) however this version
+ * doesn't transparently unlock. Instead we set the xas index to
+ * the current value of reserved every iteration.
+ */
+ xas_set(&xas, ent->reserved);
+ curr = xas_load(&xas);
+ if (!curr) {
+ if (to_store && ent->stored == ent->reserved)
+ xas_store(&xas, to_store);
+ else
+ xas_store(&xas, XA_ZERO_ENTRY);
+ if (xas_valid(&xas)) {
+ ent->reserved++;
+ if (to_store) {
+ if (ent->stored != ent->reserved)
+ __xa_store(&ent->mkeys,
+ ent->stored,
+ to_store,
+ GFP_KERNEL);
+ ent->stored++;
+ queue_adjust_cache_locked(ent);
+ WRITE_ONCE(ent->dev->cache.last_add,
+ jiffies);
+ }
+ }
+ }
+ xa_unlock_irq(&ent->mkeys);
+
+ /*
+ * Notice xas_nomem() must always be called as it cleans
+ * up any cached allocation.
+ */
+ if (!xas_nomem(&xas, GFP_KERNEL))
+ break;
+ xa_lock_irq(&ent->mkeys);
+ }
+ if (xas_error(&xas))
+ return xas_error(&xas);
+ if (WARN_ON(curr))
+ return -EINVAL;
+ return 0;
+}
+
+static void undo_push_reserve_mkey(struct mlx5_cache_ent *ent)
+{
+ void *old;
+
+ ent->reserved--;
+ old = __xa_erase(&ent->mkeys, ent->reserved);
+ WARN_ON(old);
+}
+
+static void push_to_reserved(struct mlx5_cache_ent *ent, struct mlx5_ib_mr *mr)
+{
+ void *old;
+
+ old = __xa_store(&ent->mkeys, ent->stored, mr, 0);
+ WARN_ON(old);
+ ent->stored++;
+}
+
+static struct mlx5_ib_mr *pop_stored_mkey(struct mlx5_cache_ent *ent)
+{
+ struct mlx5_ib_mr *mr;
+ void *old;
+
+ ent->stored--;
+ ent->reserved--;
+
+ if (ent->stored == ent->reserved) {
+ mr = __xa_erase(&ent->mkeys, ent->stored);
+ WARN_ON(!mr);
+ return mr;
+ }
+
+ mr = __xa_store(&ent->mkeys, ent->stored, XA_ZERO_ENTRY,
+ GFP_KERNEL);
+ WARN_ON(!mr || xa_is_err(mr));
+ old = __xa_erase(&ent->mkeys, ent->reserved);
+ WARN_ON(old);
+ return mr;
+}
+
static void create_mkey_callback(int status, struct mlx5_async_work *context)
{
struct mlx5_ib_mr *mr =
@@ -154,7 +252,7 @@ static void create_mkey_callback(int status, struct mlx5_async_work *context)
create_mkey_warn(dev, status, mr->out);
kfree(mr);
xa_lock_irqsave(&ent->mkeys, flags);
- ent->pending--;
+ undo_push_reserve_mkey(ent);
WRITE_ONCE(dev->fill_delay, 1);
xa_unlock_irqrestore(&ent->mkeys, flags);
mod_timer(&dev->delay_timer, jiffies + HZ);
@@ -169,12 +267,10 @@ static void create_mkey_callback(int status, struct mlx5_async_work *context)
WRITE_ONCE(dev->cache.last_add, jiffies);
xa_lock_irqsave(&ent->mkeys, flags);
- list_add_tail(&mr->list, &ent->head);
- ent->available_mrs++;
+ push_to_reserved(ent, mr);
ent->total_mrs++;
/* If we are doing fill_to_high_water then keep going. */
queue_adjust_cache_locked(ent);
- ent->pending--;
xa_unlock_irqrestore(&ent->mkeys, flags);
}
@@ -237,31 +333,33 @@ static int add_keys(struct mlx5_cache_ent *ent, unsigned int num)
mr = alloc_cache_mr(ent, mkc);
if (!mr) {
err = -ENOMEM;
- break;
+ goto free_in;
}
- xa_lock_irq(&ent->mkeys);
- if (ent->pending >= MAX_PENDING_REG_MR) {
- err = -EAGAIN;
- xa_unlock_irq(&ent->mkeys);
- kfree(mr);
- break;
- }
- ent->pending++;
- xa_unlock_irq(&ent->mkeys);
+
+ err = push_mkey(ent, true, NULL);
+ if (err)
+ goto free_mr;
+
err = mlx5_ib_create_mkey_cb(ent->dev, &mr->mmkey,
&ent->dev->async_ctx, in, inlen,
mr->out, sizeof(mr->out),
&mr->cb_work);
if (err) {
- xa_lock_irq(&ent->mkeys);
- ent->pending--;
- xa_unlock_irq(&ent->mkeys);
mlx5_ib_warn(ent->dev, "create mkey failed %d\n", err);
- kfree(mr);
- break;
+ goto err_undo_reserve;
}
}
+ kfree(in);
+ return 0;
+
+err_undo_reserve:
+ xa_lock_irq(&ent->mkeys);
+ undo_push_reserve_mkey(ent);
+ xa_unlock_irq(&ent->mkeys);
+free_mr:
+ kfree(mr);
+free_in:
kfree(in);
return err;
}
@@ -310,11 +408,9 @@ static void remove_cache_mr_locked(struct mlx5_cache_ent *ent)
struct mlx5_ib_mr *mr;
lockdep_assert_held(&ent->mkeys.xa_lock);
- if (list_empty(&ent->head))
+ if (!ent->stored)
return;
- mr = list_first_entry(&ent->head, struct mlx5_ib_mr, list);
- list_del(&mr->list);
- ent->available_mrs--;
+ mr = pop_stored_mkey(ent);
ent->total_mrs--;
xa_unlock_irq(&ent->mkeys);
mlx5_core_destroy_mkey(ent->dev->mdev, mr->mmkey.key);
@@ -324,6 +420,7 @@ static void remove_cache_mr_locked(struct mlx5_cache_ent *ent)
static int resize_available_mrs(struct mlx5_cache_ent *ent, unsigned int target,
bool limit_fill)
+ __acquires(&ent->mkeys) __releases(&ent->mkeys)
{
int err;
@@ -332,10 +429,10 @@ static int resize_available_mrs(struct mlx5_cache_ent *ent, unsigned int target,
while (true) {
if (limit_fill)
target = ent->limit * 2;
- if (target == ent->available_mrs + ent->pending)
+ if (target == ent->reserved)
return 0;
- if (target > ent->available_mrs + ent->pending) {
- u32 todo = target - (ent->available_mrs + ent->pending);
+ if (target > ent->reserved) {
+ u32 todo = target - ent->reserved;
xa_unlock_irq(&ent->mkeys);
err = add_keys(ent, todo);
@@ -366,15 +463,15 @@ static ssize_t size_write(struct file *filp, const char __user *buf,
/*
* Target is the new value of total_mrs the user requests, however we
- * cannot free MRs that are in use. Compute the target value for
- * available_mrs.
+ * cannot free MRs that are in use. Compute the target value for stored
+ * mkeys.
*/
xa_lock_irq(&ent->mkeys);
- if (target < ent->total_mrs - ent->available_mrs) {
+ if (target < ent->total_mrs - ent->stored) {
err = -EINVAL;
goto err_unlock;
}
- target = target - (ent->total_mrs - ent->available_mrs);
+ target = target - (ent->total_mrs - ent->stored);
if (target < ent->limit || target > ent->limit*2) {
err = -EINVAL;
goto err_unlock;
@@ -466,7 +563,7 @@ static bool someone_adding(struct mlx5_mr_cache *cache)
bool ret;
xa_lock_irq(&ent->mkeys);
- ret = ent->available_mrs < ent->limit;
+ ret = ent->stored < ent->limit;
xa_unlock_irq(&ent->mkeys);
if (ret)
return true;
@@ -485,22 +582,22 @@ static void queue_adjust_cache_locked(struct mlx5_cache_ent *ent)
if (ent->disabled || READ_ONCE(ent->dev->fill_delay))
return;
- if (ent->available_mrs < ent->limit) {
+ if (ent->stored < ent->limit) {
ent->fill_to_high_water = true;
mod_delayed_work(ent->dev->cache.wq, &ent->dwork, 0);
} else if (ent->fill_to_high_water &&
- ent->available_mrs + ent->pending < 2 * ent->limit) {
+ ent->reserved < 2 * ent->limit) {
/*
* Once we start populating due to hitting a low water mark
* continue until we pass the high water mark.
*/
mod_delayed_work(ent->dev->cache.wq, &ent->dwork, 0);
- } else if (ent->available_mrs == 2 * ent->limit) {
+ } else if (ent->stored == 2 * ent->limit) {
ent->fill_to_high_water = false;
- } else if (ent->available_mrs > 2 * ent->limit) {
+ } else if (ent->stored > 2 * ent->limit) {
/* Queue deletion of excess entries */
ent->fill_to_high_water = false;
- if (ent->pending)
+ if (ent->stored != ent->reserved)
queue_delayed_work(ent->dev->cache.wq, &ent->dwork,
msecs_to_jiffies(1000));
else
@@ -518,8 +615,7 @@ static void __cache_work_func(struct mlx5_cache_ent *ent)
if (ent->disabled)
goto out;
- if (ent->fill_to_high_water &&
- ent->available_mrs + ent->pending < 2 * ent->limit &&
+ if (ent->fill_to_high_water && ent->reserved < 2 * ent->limit &&
!READ_ONCE(dev->fill_delay)) {
xa_unlock_irq(&ent->mkeys);
err = add_keys(ent, 1);
@@ -528,8 +624,8 @@ static void __cache_work_func(struct mlx5_cache_ent *ent)
goto out;
if (err) {
/*
- * EAGAIN only happens if pending is positive, so we
- * will be rescheduled from reg_mr_callback(). The only
+ * EAGAIN only happens if there are pending MRs, so we
+ * will be rescheduled when storing them. The only
* failure path here is ENOMEM.
*/
if (err != -EAGAIN) {
@@ -541,7 +637,7 @@ static void __cache_work_func(struct mlx5_cache_ent *ent)
msecs_to_jiffies(1000));
}
}
- } else if (ent->available_mrs > 2 * ent->limit) {
+ } else if (ent->stored > 2 * ent->limit) {
bool need_delay;
/*
@@ -593,7 +689,7 @@ struct mlx5_ib_mr *mlx5_mr_cache_alloc(struct mlx5_ib_dev *dev,
return ERR_PTR(-EOPNOTSUPP);
xa_lock_irq(&ent->mkeys);
- if (list_empty(&ent->head)) {
+ if (!ent->stored) {
queue_adjust_cache_locked(ent);
ent->miss++;
xa_unlock_irq(&ent->mkeys);
@@ -601,9 +697,7 @@ struct mlx5_ib_mr *mlx5_mr_cache_alloc(struct mlx5_ib_dev *dev,
if (IS_ERR(mr))
return mr;
} else {
- mr = list_first_entry(&ent->head, struct mlx5_ib_mr, list);
- list_del(&mr->list);
- ent->available_mrs--;
+ mr = pop_stored_mkey(ent);
queue_adjust_cache_locked(ent);
xa_unlock_irq(&ent->mkeys);
@@ -612,45 +706,23 @@ struct mlx5_ib_mr *mlx5_mr_cache_alloc(struct mlx5_ib_dev *dev,
return mr;
}
-static void mlx5_mr_cache_free(struct mlx5_ib_dev *dev, struct mlx5_ib_mr *mr)
-{
- struct mlx5_cache_ent *ent = mr->cache_ent;
-
- WRITE_ONCE(dev->cache.last_add, jiffies);
- xa_lock_irq(&ent->mkeys);
- list_add_tail(&mr->list, &ent->head);
- ent->available_mrs++;
- queue_adjust_cache_locked(ent);
- xa_unlock_irq(&ent->mkeys);
-}
-
static void clean_keys(struct mlx5_ib_dev *dev, int c)
{
struct mlx5_mr_cache *cache = &dev->cache;
struct mlx5_cache_ent *ent = &cache->ent[c];
- struct mlx5_ib_mr *tmp_mr;
struct mlx5_ib_mr *mr;
- LIST_HEAD(del_list);
cancel_delayed_work(&ent->dwork);
- while (1) {
- xa_lock_irq(&ent->mkeys);
- if (list_empty(&ent->head)) {
- xa_unlock_irq(&ent->mkeys);
- break;
- }
- mr = list_first_entry(&ent->head, struct mlx5_ib_mr, list);
- list_move(&mr->list, &del_list);
- ent->available_mrs--;
+ xa_lock_irq(&ent->mkeys);
+ while (ent->stored) {
+ mr = pop_stored_mkey(ent);
ent->total_mrs--;
xa_unlock_irq(&ent->mkeys);
mlx5_core_destroy_mkey(dev->mdev, mr->mmkey.key);
- }
-
- list_for_each_entry_safe(mr, tmp_mr, &del_list, list) {
- list_del(&mr->list);
kfree(mr);
+ xa_lock_irq(&ent->mkeys);
}
+ xa_unlock_irq(&ent->mkeys);
}
static void mlx5_mr_cache_debugfs_cleanup(struct mlx5_ib_dev *dev)
@@ -680,7 +752,7 @@ static void mlx5_mr_cache_debugfs_init(struct mlx5_ib_dev *dev)
dir = debugfs_create_dir(ent->name, cache->root);
debugfs_create_file("size", 0600, dir, ent, &size_fops);
debugfs_create_file("limit", 0600, dir, ent, &limit_fops);
- debugfs_create_u32("cur", 0400, dir, &ent->available_mrs);
+ debugfs_create_ulong("cur", 0400, dir, &ent->stored);
debugfs_create_u32("miss", 0600, dir, &ent->miss);
}
}
@@ -709,7 +781,6 @@ int mlx5_mr_cache_init(struct mlx5_ib_dev *dev)
timer_setup(&dev->delay_timer, delay_time_func, 0);
for (i = 0; i < MAX_MR_CACHE_ENTRIES; i++) {
ent = &cache->ent[i];
- INIT_LIST_HEAD(&ent->head);
xa_init_flags(&ent->mkeys, XA_FLAGS_LOCK_IRQ);
ent->order = i + 2;
ent->dev = dev;
@@ -1571,7 +1642,8 @@ int mlx5_ib_dereg_mr(struct ib_mr *ibmr, struct ib_udata *udata)
/* Stop DMA */
if (mr->cache_ent) {
- if (mlx5r_umr_revoke_mr(mr)) {
+ if (mlx5r_umr_revoke_mr(mr) ||
+ push_mkey(mr->cache_ent, false, mr)) {
xa_lock_irq(&mr->cache_ent->mkeys);
mr->cache_ent->total_mrs--;
xa_unlock_irq(&mr->cache_ent->mkeys);
@@ -1595,9 +1667,7 @@ int mlx5_ib_dereg_mr(struct ib_mr *ibmr, struct ib_udata *udata)
mlx5_ib_free_odp_mr(mr);
}
- if (mr->cache_ent) {
- mlx5_mr_cache_free(dev, mr);
- } else {
+ if (!mr->cache_ent) {
mlx5_free_priv_descs(mr);
kfree(mr);
}
--
2.17.2
^ permalink raw reply related [flat|nested] 7+ messages in thread
* [PATCH rdma-next v1 3/5] RDMA/mlx5: Store the number of in_use cache mkeys instead of total_mrs
2022-07-26 7:19 [PATCH rdma-next v1 0/5] MR cache cleanup Michael Guralnik
2022-07-26 7:19 ` [PATCH rdma-next v1 1/5] RDMA/mlx5: Replace ent->lock with xa_lock Michael Guralnik
2022-07-26 7:19 ` [PATCH rdma-next v1 2/5] RDMA/mlx5: Replace cache list with Xarray Michael Guralnik
@ 2022-07-26 7:19 ` Michael Guralnik
2022-07-26 7:19 ` [PATCH rdma-next v1 4/5] RDMA/mlx5: Store in the cache mkeys instead of mrs Michael Guralnik
` (2 subsequent siblings)
5 siblings, 0 replies; 7+ messages in thread
From: Michael Guralnik @ 2022-07-26 7:19 UTC (permalink / raw)
To: jgg
Cc: leonro, maorg, linux-kernel, linux-rdma, netdev, saeedm, Aharon Landau
From: Aharon Landau <aharonl@nvidia.com>
total_mrs is used only to calculate the number of mkeys currently in
use. To simplify things, replace it with a new member called "in_use"
and directly store the number of mkeys currently in use.
Signed-off-by: Aharon Landau <aharonl@nvidia.com>
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
---
drivers/infiniband/hw/mlx5/mlx5_ib.h | 4 +---
drivers/infiniband/hw/mlx5/mr.c | 30 ++++++++++++++--------------
2 files changed, 16 insertions(+), 18 deletions(-)
diff --git a/drivers/infiniband/hw/mlx5/mlx5_ib.h b/drivers/infiniband/hw/mlx5/mlx5_ib.h
index e0eb666aefa1..da9202f4b5f3 100644
--- a/drivers/infiniband/hw/mlx5/mlx5_ib.h
+++ b/drivers/infiniband/hw/mlx5/mlx5_ib.h
@@ -755,12 +755,10 @@ struct mlx5_cache_ent {
u8 fill_to_high_water:1;
/*
- * - total_mrs is stored mkeys plus all in use MRs that could be
- * returned to the cache.
* - limit is the low water mark for stored mkeys, 2* limit is the
* upper water mark.
*/
- u32 total_mrs;
+ u32 in_use;
u32 limit;
/* Statistics */
diff --git a/drivers/infiniband/hw/mlx5/mr.c b/drivers/infiniband/hw/mlx5/mr.c
index cbb8882c7787..26bfdbba24b4 100644
--- a/drivers/infiniband/hw/mlx5/mr.c
+++ b/drivers/infiniband/hw/mlx5/mr.c
@@ -268,7 +268,6 @@ static void create_mkey_callback(int status, struct mlx5_async_work *context)
xa_lock_irqsave(&ent->mkeys, flags);
push_to_reserved(ent, mr);
- ent->total_mrs++;
/* If we are doing fill_to_high_water then keep going. */
queue_adjust_cache_locked(ent);
xa_unlock_irqrestore(&ent->mkeys, flags);
@@ -391,9 +390,6 @@ static struct mlx5_ib_mr *create_cache_mr(struct mlx5_cache_ent *ent)
init_waitqueue_head(&mr->mmkey.wait);
mr->mmkey.type = MLX5_MKEY_MR;
WRITE_ONCE(ent->dev->cache.last_add, jiffies);
- xa_lock_irq(&ent->mkeys);
- ent->total_mrs++;
- xa_unlock_irq(&ent->mkeys);
kfree(in);
return mr;
free_mr:
@@ -411,7 +407,6 @@ static void remove_cache_mr_locked(struct mlx5_cache_ent *ent)
if (!ent->stored)
return;
mr = pop_stored_mkey(ent);
- ent->total_mrs--;
xa_unlock_irq(&ent->mkeys);
mlx5_core_destroy_mkey(ent->dev->mdev, mr->mmkey.key);
kfree(mr);
@@ -467,11 +462,11 @@ static ssize_t size_write(struct file *filp, const char __user *buf,
* mkeys.
*/
xa_lock_irq(&ent->mkeys);
- if (target < ent->total_mrs - ent->stored) {
+ if (target < ent->in_use) {
err = -EINVAL;
goto err_unlock;
}
- target = target - (ent->total_mrs - ent->stored);
+ target = target - ent->in_use;
if (target < ent->limit || target > ent->limit*2) {
err = -EINVAL;
goto err_unlock;
@@ -495,7 +490,7 @@ static ssize_t size_read(struct file *filp, char __user *buf, size_t count,
char lbuf[20];
int err;
- err = snprintf(lbuf, sizeof(lbuf), "%d\n", ent->total_mrs);
+ err = snprintf(lbuf, sizeof(lbuf), "%ld\n", ent->stored + ent->in_use);
if (err < 0)
return err;
@@ -689,13 +684,19 @@ struct mlx5_ib_mr *mlx5_mr_cache_alloc(struct mlx5_ib_dev *dev,
return ERR_PTR(-EOPNOTSUPP);
xa_lock_irq(&ent->mkeys);
+ ent->in_use++;
+
if (!ent->stored) {
queue_adjust_cache_locked(ent);
ent->miss++;
xa_unlock_irq(&ent->mkeys);
mr = create_cache_mr(ent);
- if (IS_ERR(mr))
+ if (IS_ERR(mr)) {
+ xa_lock_irq(&ent->mkeys);
+ ent->in_use--;
+ xa_unlock_irq(&ent->mkeys);
return mr;
+ }
} else {
mr = pop_stored_mkey(ent);
queue_adjust_cache_locked(ent);
@@ -716,7 +717,6 @@ static void clean_keys(struct mlx5_ib_dev *dev, int c)
xa_lock_irq(&ent->mkeys);
while (ent->stored) {
mr = pop_stored_mkey(ent);
- ent->total_mrs--;
xa_unlock_irq(&ent->mkeys);
mlx5_core_destroy_mkey(dev->mdev, mr->mmkey.key);
kfree(mr);
@@ -1642,13 +1642,13 @@ int mlx5_ib_dereg_mr(struct ib_mr *ibmr, struct ib_udata *udata)
/* Stop DMA */
if (mr->cache_ent) {
+ xa_lock_irq(&mr->cache_ent->mkeys);
+ mr->cache_ent->in_use--;
+ xa_unlock_irq(&mr->cache_ent->mkeys);
+
if (mlx5r_umr_revoke_mr(mr) ||
- push_mkey(mr->cache_ent, false, mr)) {
- xa_lock_irq(&mr->cache_ent->mkeys);
- mr->cache_ent->total_mrs--;
- xa_unlock_irq(&mr->cache_ent->mkeys);
+ push_mkey(mr->cache_ent, false, mr))
mr->cache_ent = NULL;
- }
}
if (!mr->cache_ent) {
rc = destroy_mkey(to_mdev(mr->ibmr.device), mr);
--
2.17.2
^ permalink raw reply related [flat|nested] 7+ messages in thread
* [PATCH rdma-next v1 4/5] RDMA/mlx5: Store in the cache mkeys instead of mrs
2022-07-26 7:19 [PATCH rdma-next v1 0/5] MR cache cleanup Michael Guralnik
` (2 preceding siblings ...)
2022-07-26 7:19 ` [PATCH rdma-next v1 3/5] RDMA/mlx5: Store the number of in_use cache mkeys instead of total_mrs Michael Guralnik
@ 2022-07-26 7:19 ` Michael Guralnik
2022-07-26 7:19 ` [PATCH rdma-next v1 5/5] RDMA/mlx5: Rename the mkey cache variables and functions Michael Guralnik
2022-07-27 18:17 ` [PATCH rdma-next v1 0/5] MR cache cleanup Jason Gunthorpe
5 siblings, 0 replies; 7+ messages in thread
From: Michael Guralnik @ 2022-07-26 7:19 UTC (permalink / raw)
To: jgg
Cc: leonro, maorg, linux-kernel, linux-rdma, netdev, saeedm, Aharon Landau
From: Aharon Landau <aharonl@nvidia.com>
Currently, the driver stores mlx5_ib_mr struct in the cache entries,
although the only use of the cached MR is the mkey. Store only the mkey
in the cache.
Signed-off-by: Aharon Landau <aharonl@nvidia.com>
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
---
drivers/infiniband/hw/mlx5/mlx5_ib.h | 26 ++--
drivers/infiniband/hw/mlx5/mr.c | 200 ++++++++++++---------------
2 files changed, 97 insertions(+), 129 deletions(-)
diff --git a/drivers/infiniband/hw/mlx5/mlx5_ib.h b/drivers/infiniband/hw/mlx5/mlx5_ib.h
index da9202f4b5f3..91f985cd7d90 100644
--- a/drivers/infiniband/hw/mlx5/mlx5_ib.h
+++ b/drivers/infiniband/hw/mlx5/mlx5_ib.h
@@ -619,6 +619,7 @@ struct mlx5_ib_mkey {
unsigned int ndescs;
struct wait_queue_head wait;
refcount_t usecount;
+ struct mlx5_cache_ent *cache_ent;
};
#define MLX5_IB_MTT_PRESENT (MLX5_IB_MTT_READ | MLX5_IB_MTT_WRITE)
@@ -641,18 +642,9 @@ struct mlx5_ib_mr {
struct ib_mr ibmr;
struct mlx5_ib_mkey mmkey;
- /* User MR data */
- struct mlx5_cache_ent *cache_ent;
- /* Everything after cache_ent is zero'd when MR allocated */
struct ib_umem *umem;
union {
- /* Used only while the MR is in the cache */
- struct {
- u32 out[MLX5_ST_SZ_DW(create_mkey_out)];
- struct mlx5_async_work cb_work;
- };
-
/* Used only by kernel MRs (umem == NULL) */
struct {
void *descs;
@@ -692,12 +684,6 @@ struct mlx5_ib_mr {
};
};
-/* Zero the fields in the mr that are variant depending on usage */
-static inline void mlx5_clear_mr(struct mlx5_ib_mr *mr)
-{
- memset_after(mr, 0, cache_ent);
-}
-
static inline bool is_odp_mr(struct mlx5_ib_mr *mr)
{
return IS_ENABLED(CONFIG_INFINIBAND_ON_DEMAND_PAGING) && mr->umem &&
@@ -768,6 +754,16 @@ struct mlx5_cache_ent {
struct delayed_work dwork;
};
+struct mlx5r_async_create_mkey {
+ union {
+ u32 in[MLX5_ST_SZ_BYTES(create_mkey_in)];
+ u32 out[MLX5_ST_SZ_DW(create_mkey_out)];
+ };
+ struct mlx5_async_work cb_work;
+ struct mlx5_cache_ent *ent;
+ u32 mkey;
+};
+
struct mlx5_mr_cache {
struct workqueue_struct *wq;
struct mlx5_cache_ent ent[MAX_MR_CACHE_ENTRIES];
diff --git a/drivers/infiniband/hw/mlx5/mr.c b/drivers/infiniband/hw/mlx5/mr.c
index 26bfdbba24b4..edbc2990d151 100644
--- a/drivers/infiniband/hw/mlx5/mr.c
+++ b/drivers/infiniband/hw/mlx5/mr.c
@@ -82,15 +82,14 @@ static void set_mkc_access_pd_addr_fields(void *mkc, int acc, u64 start_addr,
MLX5_SET64(mkc, mkc, start_addr, start_addr);
}
-static void assign_mkey_variant(struct mlx5_ib_dev *dev,
- struct mlx5_ib_mkey *mkey, u32 *in)
+static void assign_mkey_variant(struct mlx5_ib_dev *dev, u32 *mkey, u32 *in)
{
u8 key = atomic_inc_return(&dev->mkey_var);
void *mkc;
mkc = MLX5_ADDR_OF(create_mkey_in, in, memory_key_mkey_entry);
MLX5_SET(mkc, mkc, mkey_7_0, key);
- mkey->key = key;
+ *mkey = key;
}
static int mlx5_ib_create_mkey(struct mlx5_ib_dev *dev,
@@ -98,7 +97,7 @@ static int mlx5_ib_create_mkey(struct mlx5_ib_dev *dev,
{
int ret;
- assign_mkey_variant(dev, mkey, in);
+ assign_mkey_variant(dev, &mkey->key, in);
ret = mlx5_core_create_mkey(dev->mdev, &mkey->key, in, inlen);
if (!ret)
init_waitqueue_head(&mkey->wait);
@@ -106,17 +105,18 @@ static int mlx5_ib_create_mkey(struct mlx5_ib_dev *dev,
return ret;
}
-static int
-mlx5_ib_create_mkey_cb(struct mlx5_ib_dev *dev,
- struct mlx5_ib_mkey *mkey,
- struct mlx5_async_ctx *async_ctx,
- u32 *in, int inlen, u32 *out, int outlen,
- struct mlx5_async_work *context)
+static int mlx5_ib_create_mkey_cb(struct mlx5r_async_create_mkey *async_create)
{
- MLX5_SET(create_mkey_in, in, opcode, MLX5_CMD_OP_CREATE_MKEY);
- assign_mkey_variant(dev, mkey, in);
- return mlx5_cmd_exec_cb(async_ctx, in, inlen, out, outlen,
- create_mkey_callback, context);
+ struct mlx5_ib_dev *dev = async_create->ent->dev;
+ size_t inlen = MLX5_ST_SZ_BYTES(create_mkey_in);
+ size_t outlen = MLX5_ST_SZ_BYTES(create_mkey_out);
+
+ MLX5_SET(create_mkey_in, async_create->in, opcode,
+ MLX5_CMD_OP_CREATE_MKEY);
+ assign_mkey_variant(dev, &async_create->mkey, async_create->in);
+ return mlx5_cmd_exec_cb(&dev->async_ctx, async_create->in, inlen,
+ async_create->out, outlen, create_mkey_callback,
+ &async_create->cb_work);
}
static int mr_cache_max_order(struct mlx5_ib_dev *dev);
@@ -209,48 +209,47 @@ static void undo_push_reserve_mkey(struct mlx5_cache_ent *ent)
WARN_ON(old);
}
-static void push_to_reserved(struct mlx5_cache_ent *ent, struct mlx5_ib_mr *mr)
+static void push_to_reserved(struct mlx5_cache_ent *ent, u32 mkey)
{
void *old;
- old = __xa_store(&ent->mkeys, ent->stored, mr, 0);
+ old = __xa_store(&ent->mkeys, ent->stored, xa_mk_value(mkey), 0);
WARN_ON(old);
ent->stored++;
}
-static struct mlx5_ib_mr *pop_stored_mkey(struct mlx5_cache_ent *ent)
+static u32 pop_stored_mkey(struct mlx5_cache_ent *ent)
{
- struct mlx5_ib_mr *mr;
- void *old;
+ void *old, *xa_mkey;
ent->stored--;
ent->reserved--;
if (ent->stored == ent->reserved) {
- mr = __xa_erase(&ent->mkeys, ent->stored);
- WARN_ON(!mr);
- return mr;
+ xa_mkey = __xa_erase(&ent->mkeys, ent->stored);
+ WARN_ON(!xa_mkey);
+ return (u32)xa_to_value(xa_mkey);
}
- mr = __xa_store(&ent->mkeys, ent->stored, XA_ZERO_ENTRY,
- GFP_KERNEL);
- WARN_ON(!mr || xa_is_err(mr));
+ xa_mkey = __xa_store(&ent->mkeys, ent->stored, XA_ZERO_ENTRY,
+ GFP_KERNEL);
+ WARN_ON(!xa_mkey || xa_is_err(xa_mkey));
old = __xa_erase(&ent->mkeys, ent->reserved);
WARN_ON(old);
- return mr;
+ return (u32)xa_to_value(xa_mkey);
}
static void create_mkey_callback(int status, struct mlx5_async_work *context)
{
- struct mlx5_ib_mr *mr =
- container_of(context, struct mlx5_ib_mr, cb_work);
- struct mlx5_cache_ent *ent = mr->cache_ent;
+ struct mlx5r_async_create_mkey *mkey_out =
+ container_of(context, struct mlx5r_async_create_mkey, cb_work);
+ struct mlx5_cache_ent *ent = mkey_out->ent;
struct mlx5_ib_dev *dev = ent->dev;
unsigned long flags;
if (status) {
- create_mkey_warn(dev, status, mr->out);
- kfree(mr);
+ create_mkey_warn(dev, status, mkey_out->out);
+ kfree(mkey_out);
xa_lock_irqsave(&ent->mkeys, flags);
undo_push_reserve_mkey(ent);
WRITE_ONCE(dev->fill_delay, 1);
@@ -259,18 +258,16 @@ static void create_mkey_callback(int status, struct mlx5_async_work *context)
return;
}
- mr->mmkey.type = MLX5_MKEY_MR;
- mr->mmkey.key |= mlx5_idx_to_mkey(
- MLX5_GET(create_mkey_out, mr->out, mkey_index));
- init_waitqueue_head(&mr->mmkey.wait);
-
+ mkey_out->mkey |= mlx5_idx_to_mkey(
+ MLX5_GET(create_mkey_out, mkey_out->out, mkey_index));
WRITE_ONCE(dev->cache.last_add, jiffies);
xa_lock_irqsave(&ent->mkeys, flags);
- push_to_reserved(ent, mr);
+ push_to_reserved(ent, mkey_out->mkey);
/* If we are doing fill_to_high_water then keep going. */
queue_adjust_cache_locked(ent);
xa_unlock_irqrestore(&ent->mkeys, flags);
+ kfree(mkey_out);
}
static int get_mkc_octo_size(unsigned int access_mode, unsigned int ndescs)
@@ -292,15 +289,8 @@ static int get_mkc_octo_size(unsigned int access_mode, unsigned int ndescs)
return ret;
}
-static struct mlx5_ib_mr *alloc_cache_mr(struct mlx5_cache_ent *ent, void *mkc)
+static void set_cache_mkc(struct mlx5_cache_ent *ent, void *mkc)
{
- struct mlx5_ib_mr *mr;
-
- mr = kzalloc(sizeof(*mr), GFP_KERNEL);
- if (!mr)
- return NULL;
- mr->cache_ent = ent;
-
set_mkc_access_pd_addr_fields(mkc, 0, 0, ent->dev->umrc.pd);
MLX5_SET(mkc, mkc, free, 1);
MLX5_SET(mkc, mkc, umr_en, 1);
@@ -310,106 +300,82 @@ static struct mlx5_ib_mr *alloc_cache_mr(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);
- return mr;
}
/* Asynchronously schedule new MRs to be populated in the cache. */
static int add_keys(struct mlx5_cache_ent *ent, unsigned int num)
{
- size_t inlen = MLX5_ST_SZ_BYTES(create_mkey_in);
- struct mlx5_ib_mr *mr;
+ struct mlx5r_async_create_mkey *async_create;
void *mkc;
- u32 *in;
int err = 0;
int i;
- in = kzalloc(inlen, GFP_KERNEL);
- if (!in)
- return -ENOMEM;
-
- mkc = MLX5_ADDR_OF(create_mkey_in, in, memory_key_mkey_entry);
for (i = 0; i < num; i++) {
- mr = alloc_cache_mr(ent, mkc);
- if (!mr) {
- err = -ENOMEM;
- goto free_in;
- }
+ async_create = kzalloc(sizeof(struct mlx5r_async_create_mkey),
+ GFP_KERNEL);
+ if (!async_create)
+ return -ENOMEM;
+ mkc = MLX5_ADDR_OF(create_mkey_in, async_create->in,
+ memory_key_mkey_entry);
+ set_cache_mkc(ent, mkc);
+ async_create->ent = ent;
err = push_mkey(ent, true, NULL);
if (err)
- goto free_mr;
+ goto free_async_create;
- err = mlx5_ib_create_mkey_cb(ent->dev, &mr->mmkey,
- &ent->dev->async_ctx, in, inlen,
- mr->out, sizeof(mr->out),
- &mr->cb_work);
+ err = mlx5_ib_create_mkey_cb(async_create);
if (err) {
mlx5_ib_warn(ent->dev, "create mkey failed %d\n", err);
goto err_undo_reserve;
}
}
- kfree(in);
return 0;
err_undo_reserve:
xa_lock_irq(&ent->mkeys);
undo_push_reserve_mkey(ent);
xa_unlock_irq(&ent->mkeys);
-free_mr:
- kfree(mr);
-free_in:
- kfree(in);
+free_async_create:
+ kfree(async_create);
return err;
}
/* Synchronously create a MR in the cache */
-static struct mlx5_ib_mr *create_cache_mr(struct mlx5_cache_ent *ent)
+static int create_cache_mkey(struct mlx5_cache_ent *ent, u32 *mkey)
{
size_t inlen = MLX5_ST_SZ_BYTES(create_mkey_in);
- struct mlx5_ib_mr *mr;
void *mkc;
u32 *in;
int err;
in = kzalloc(inlen, GFP_KERNEL);
if (!in)
- return ERR_PTR(-ENOMEM);
+ return -ENOMEM;
mkc = MLX5_ADDR_OF(create_mkey_in, in, memory_key_mkey_entry);
+ set_cache_mkc(ent, mkc);
- mr = alloc_cache_mr(ent, mkc);
- if (!mr) {
- err = -ENOMEM;
- goto free_in;
- }
-
- err = mlx5_core_create_mkey(ent->dev->mdev, &mr->mmkey.key, in, inlen);
+ err = mlx5_core_create_mkey(ent->dev->mdev, mkey, in, inlen);
if (err)
- goto free_mr;
+ goto free_in;
- init_waitqueue_head(&mr->mmkey.wait);
- mr->mmkey.type = MLX5_MKEY_MR;
WRITE_ONCE(ent->dev->cache.last_add, jiffies);
- kfree(in);
- return mr;
-free_mr:
- kfree(mr);
free_in:
kfree(in);
- return ERR_PTR(err);
+ return err;
}
static void remove_cache_mr_locked(struct mlx5_cache_ent *ent)
{
- struct mlx5_ib_mr *mr;
+ u32 mkey;
lockdep_assert_held(&ent->mkeys.xa_lock);
if (!ent->stored)
return;
- mr = pop_stored_mkey(ent);
+ mkey = pop_stored_mkey(ent);
xa_unlock_irq(&ent->mkeys);
- mlx5_core_destroy_mkey(ent->dev->mdev, mr->mmkey.key);
- kfree(mr);
+ mlx5_core_destroy_mkey(ent->dev->mdev, mkey);
xa_lock_irq(&ent->mkeys);
}
@@ -678,11 +644,15 @@ struct mlx5_ib_mr *mlx5_mr_cache_alloc(struct mlx5_ib_dev *dev,
int access_flags)
{
struct mlx5_ib_mr *mr;
+ int err;
- /* Matches access in alloc_cache_mr() */
if (!mlx5r_umr_can_reconfig(dev, 0, access_flags))
return ERR_PTR(-EOPNOTSUPP);
+ mr = kzalloc(sizeof(*mr), GFP_KERNEL);
+ if (!mr)
+ return ERR_PTR(-ENOMEM);
+
xa_lock_irq(&ent->mkeys);
ent->in_use++;
@@ -690,20 +660,22 @@ struct mlx5_ib_mr *mlx5_mr_cache_alloc(struct mlx5_ib_dev *dev,
queue_adjust_cache_locked(ent);
ent->miss++;
xa_unlock_irq(&ent->mkeys);
- mr = create_cache_mr(ent);
- if (IS_ERR(mr)) {
+ err = create_cache_mkey(ent, &mr->mmkey.key);
+ if (err) {
xa_lock_irq(&ent->mkeys);
ent->in_use--;
xa_unlock_irq(&ent->mkeys);
- return mr;
+ kfree(mr);
+ return ERR_PTR(err);
}
} else {
- mr = pop_stored_mkey(ent);
+ mr->mmkey.key = pop_stored_mkey(ent);
queue_adjust_cache_locked(ent);
xa_unlock_irq(&ent->mkeys);
-
- mlx5_clear_mr(mr);
}
+ mr->mmkey.cache_ent = ent;
+ mr->mmkey.type = MLX5_MKEY_MR;
+ init_waitqueue_head(&mr->mmkey.wait);
return mr;
}
@@ -711,15 +683,14 @@ static void clean_keys(struct mlx5_ib_dev *dev, int c)
{
struct mlx5_mr_cache *cache = &dev->cache;
struct mlx5_cache_ent *ent = &cache->ent[c];
- struct mlx5_ib_mr *mr;
+ u32 mkey;
cancel_delayed_work(&ent->dwork);
xa_lock_irq(&ent->mkeys);
while (ent->stored) {
- mr = pop_stored_mkey(ent);
+ mkey = pop_stored_mkey(ent);
xa_unlock_irq(&ent->mkeys);
- mlx5_core_destroy_mkey(dev->mdev, mr->mmkey.key);
- kfree(mr);
+ mlx5_core_destroy_mkey(dev->mdev, mkey);
xa_lock_irq(&ent->mkeys);
}
xa_unlock_irq(&ent->mkeys);
@@ -1391,7 +1362,7 @@ static bool can_use_umr_rereg_pas(struct mlx5_ib_mr *mr,
struct mlx5_ib_dev *dev = to_mdev(mr->ibmr.device);
/* We only track the allocated sizes of MRs from the cache */
- if (!mr->cache_ent)
+ if (!mr->mmkey.cache_ent)
return false;
if (!mlx5r_umr_can_load_pas(dev, new_umem->length))
return false;
@@ -1400,7 +1371,7 @@ static bool can_use_umr_rereg_pas(struct mlx5_ib_mr *mr,
mlx5_umem_find_best_pgsz(new_umem, mkc, log_page_size, 0, iova);
if (WARN_ON(!*page_size))
return false;
- return (1ULL << mr->cache_ent->order) >=
+ return (1ULL << mr->mmkey.cache_ent->order) >=
ib_umem_num_dma_blocks(new_umem, *page_size);
}
@@ -1641,16 +1612,17 @@ int mlx5_ib_dereg_mr(struct ib_mr *ibmr, struct ib_udata *udata)
}
/* Stop DMA */
- if (mr->cache_ent) {
- xa_lock_irq(&mr->cache_ent->mkeys);
- mr->cache_ent->in_use--;
- xa_unlock_irq(&mr->cache_ent->mkeys);
+ 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 (mlx5r_umr_revoke_mr(mr) ||
- push_mkey(mr->cache_ent, false, mr))
- mr->cache_ent = NULL;
+ push_mkey(mr->mmkey.cache_ent, false,
+ xa_mk_value(mr->mmkey.key)))
+ mr->mmkey.cache_ent = NULL;
}
- if (!mr->cache_ent) {
+ if (!mr->mmkey.cache_ent) {
rc = destroy_mkey(to_mdev(mr->ibmr.device), mr);
if (rc)
return rc;
@@ -1667,10 +1639,10 @@ int mlx5_ib_dereg_mr(struct ib_mr *ibmr, struct ib_udata *udata)
mlx5_ib_free_odp_mr(mr);
}
- if (!mr->cache_ent) {
+ if (!mr->mmkey.cache_ent)
mlx5_free_priv_descs(mr);
- kfree(mr);
- }
+
+ kfree(mr);
return 0;
}
--
2.17.2
^ permalink raw reply related [flat|nested] 7+ messages in thread
* [PATCH rdma-next v1 5/5] RDMA/mlx5: Rename the mkey cache variables and functions
2022-07-26 7:19 [PATCH rdma-next v1 0/5] MR cache cleanup Michael Guralnik
` (3 preceding siblings ...)
2022-07-26 7:19 ` [PATCH rdma-next v1 4/5] RDMA/mlx5: Store in the cache mkeys instead of mrs Michael Guralnik
@ 2022-07-26 7:19 ` Michael Guralnik
2022-07-27 18:17 ` [PATCH rdma-next v1 0/5] MR cache cleanup Jason Gunthorpe
5 siblings, 0 replies; 7+ messages in thread
From: Michael Guralnik @ 2022-07-26 7:19 UTC (permalink / raw)
To: jgg
Cc: leonro, maorg, linux-kernel, linux-rdma, netdev, saeedm, Aharon Landau
From: Aharon Landau <aharonl@nvidia.com>
After replacing the MR cache with an Mkey cache, rename the variables
and functions to fit the new meaning.
Signed-off-by: Aharon Landau <aharonl@nvidia.com>
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
---
drivers/infiniband/hw/mlx5/main.c | 4 +--
drivers/infiniband/hw/mlx5/mlx5_ib.h | 14 ++++----
drivers/infiniband/hw/mlx5/mr.c | 54 ++++++++++++++--------------
drivers/infiniband/hw/mlx5/odp.c | 2 +-
include/linux/mlx5/driver.h | 6 ++--
5 files changed, 40 insertions(+), 40 deletions(-)
diff --git a/drivers/infiniband/hw/mlx5/main.c b/drivers/infiniband/hw/mlx5/main.c
index b68fddeac0f1..a174a0eee8dc 100644
--- a/drivers/infiniband/hw/mlx5/main.c
+++ b/drivers/infiniband/hw/mlx5/main.c
@@ -4002,7 +4002,7 @@ static void mlx5_ib_stage_pre_ib_reg_umr_cleanup(struct mlx5_ib_dev *dev)
{
int err;
- err = mlx5_mr_cache_cleanup(dev);
+ err = mlx5_mkey_cache_cleanup(dev);
if (err)
mlx5_ib_warn(dev, "mr cache cleanup failed\n");
@@ -4022,7 +4022,7 @@ static int mlx5_ib_stage_post_ib_reg_umr_init(struct mlx5_ib_dev *dev)
if (ret)
return ret;
- ret = mlx5_mr_cache_init(dev);
+ ret = mlx5_mkey_cache_init(dev);
if (ret) {
mlx5_ib_warn(dev, "mr cache init failed %d\n", ret);
mlx5r_umr_resource_cleanup(dev);
diff --git a/drivers/infiniband/hw/mlx5/mlx5_ib.h b/drivers/infiniband/hw/mlx5/mlx5_ib.h
index 91f985cd7d90..2e2ad3918385 100644
--- a/drivers/infiniband/hw/mlx5/mlx5_ib.h
+++ b/drivers/infiniband/hw/mlx5/mlx5_ib.h
@@ -764,9 +764,9 @@ struct mlx5r_async_create_mkey {
u32 mkey;
};
-struct mlx5_mr_cache {
+struct mlx5_mkey_cache {
struct workqueue_struct *wq;
- struct mlx5_cache_ent ent[MAX_MR_CACHE_ENTRIES];
+ struct mlx5_cache_ent ent[MAX_MKEY_CACHE_ENTRIES];
struct dentry *root;
unsigned long last_add;
};
@@ -1065,7 +1065,7 @@ struct mlx5_ib_dev {
struct mlx5_ib_resources devr;
atomic_t mkey_var;
- struct mlx5_mr_cache cache;
+ struct mlx5_mkey_cache cache;
struct timer_list delay_timer;
/* Prevents soft lock on massive reg MRs */
struct mutex slow_path_mutex;
@@ -1310,8 +1310,8 @@ void mlx5_ib_populate_pas(struct ib_umem *umem, size_t page_size, __be64 *pas,
u64 access_flags);
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_mr_cache_init(struct mlx5_ib_dev *dev);
-int mlx5_mr_cache_cleanup(struct mlx5_ib_dev *dev);
+int mlx5_mkey_cache_init(struct mlx5_ib_dev *dev);
+int mlx5_mkey_cache_cleanup(struct mlx5_ib_dev *dev);
struct mlx5_ib_mr *mlx5_mr_cache_alloc(struct mlx5_ib_dev *dev,
struct mlx5_cache_ent *ent,
@@ -1339,7 +1339,7 @@ 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_mr_cache_entry(struct mlx5_cache_ent *ent);
+void mlx5_odp_init_mkey_cache_entry(struct mlx5_cache_ent *ent);
void mlx5_odp_populate_xlt(void *xlt, size_t idx, size_t nentries,
struct mlx5_ib_mr *mr, int flags);
@@ -1358,7 +1358,7 @@ 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_mr_cache_entry(struct mlx5_cache_ent *ent) {}
+static inline void mlx5_odp_init_mkey_cache_entry(struct mlx5_cache_ent *ent) {}
static inline void mlx5_odp_populate_xlt(void *xlt, size_t idx, size_t nentries,
struct mlx5_ib_mr *mr, int flags) {}
diff --git a/drivers/infiniband/hw/mlx5/mr.c b/drivers/infiniband/hw/mlx5/mr.c
index edbc2990d151..129d531bd01b 100644
--- a/drivers/infiniband/hw/mlx5/mr.c
+++ b/drivers/infiniband/hw/mlx5/mr.c
@@ -119,7 +119,7 @@ static int mlx5_ib_create_mkey_cb(struct mlx5r_async_create_mkey *async_create)
&async_create->cb_work);
}
-static int mr_cache_max_order(struct mlx5_ib_dev *dev);
+static int mkey_cache_max_order(struct mlx5_ib_dev *dev);
static void queue_adjust_cache_locked(struct mlx5_cache_ent *ent);
static int destroy_mkey(struct mlx5_ib_dev *dev, struct mlx5_ib_mr *mr)
@@ -515,11 +515,11 @@ static const struct file_operations limit_fops = {
.read = limit_read,
};
-static bool someone_adding(struct mlx5_mr_cache *cache)
+static bool someone_adding(struct mlx5_mkey_cache *cache)
{
unsigned int i;
- for (i = 0; i < MAX_MR_CACHE_ENTRIES; i++) {
+ for (i = 0; i < MAX_MKEY_CACHE_ENTRIES; i++) {
struct mlx5_cache_ent *ent = &cache->ent[i];
bool ret;
@@ -569,7 +569,7 @@ static void queue_adjust_cache_locked(struct mlx5_cache_ent *ent)
static void __cache_work_func(struct mlx5_cache_ent *ent)
{
struct mlx5_ib_dev *dev = ent->dev;
- struct mlx5_mr_cache *cache = &dev->cache;
+ struct mlx5_mkey_cache *cache = &dev->cache;
int err;
xa_lock_irq(&ent->mkeys);
@@ -681,7 +681,7 @@ struct mlx5_ib_mr *mlx5_mr_cache_alloc(struct mlx5_ib_dev *dev,
static void clean_keys(struct mlx5_ib_dev *dev, int c)
{
- struct mlx5_mr_cache *cache = &dev->cache;
+ struct mlx5_mkey_cache *cache = &dev->cache;
struct mlx5_cache_ent *ent = &cache->ent[c];
u32 mkey;
@@ -696,7 +696,7 @@ static void clean_keys(struct mlx5_ib_dev *dev, int c)
xa_unlock_irq(&ent->mkeys);
}
-static void mlx5_mr_cache_debugfs_cleanup(struct mlx5_ib_dev *dev)
+static void mlx5_mkey_cache_debugfs_cleanup(struct mlx5_ib_dev *dev)
{
if (!mlx5_debugfs_root || dev->is_rep)
return;
@@ -705,9 +705,9 @@ static void mlx5_mr_cache_debugfs_cleanup(struct mlx5_ib_dev *dev)
dev->cache.root = NULL;
}
-static void mlx5_mr_cache_debugfs_init(struct mlx5_ib_dev *dev)
+static void mlx5_mkey_cache_debugfs_init(struct mlx5_ib_dev *dev)
{
- struct mlx5_mr_cache *cache = &dev->cache;
+ struct mlx5_mkey_cache *cache = &dev->cache;
struct mlx5_cache_ent *ent;
struct dentry *dir;
int i;
@@ -717,7 +717,7 @@ static void mlx5_mr_cache_debugfs_init(struct mlx5_ib_dev *dev)
cache->root = debugfs_create_dir("mr_cache", mlx5_debugfs_get_dev_root(dev->mdev));
- for (i = 0; i < MAX_MR_CACHE_ENTRIES; i++) {
+ for (i = 0; i < MAX_MKEY_CACHE_ENTRIES; i++) {
ent = &cache->ent[i];
sprintf(ent->name, "%d", ent->order);
dir = debugfs_create_dir(ent->name, cache->root);
@@ -735,9 +735,9 @@ static void delay_time_func(struct timer_list *t)
WRITE_ONCE(dev->fill_delay, 0);
}
-int mlx5_mr_cache_init(struct mlx5_ib_dev *dev)
+int mlx5_mkey_cache_init(struct mlx5_ib_dev *dev)
{
- struct mlx5_mr_cache *cache = &dev->cache;
+ struct mlx5_mkey_cache *cache = &dev->cache;
struct mlx5_cache_ent *ent;
int i;
@@ -750,7 +750,7 @@ int mlx5_mr_cache_init(struct mlx5_ib_dev *dev)
mlx5_cmd_init_async_ctx(dev->mdev, &dev->async_ctx);
timer_setup(&dev->delay_timer, delay_time_func, 0);
- for (i = 0; i < MAX_MR_CACHE_ENTRIES; i++) {
+ for (i = 0; i < MAX_MKEY_CACHE_ENTRIES; i++) {
ent = &cache->ent[i];
xa_init_flags(&ent->mkeys, XA_FLAGS_LOCK_IRQ);
ent->order = i + 2;
@@ -759,12 +759,12 @@ int mlx5_mr_cache_init(struct mlx5_ib_dev *dev)
INIT_DELAYED_WORK(&ent->dwork, delayed_cache_work_func);
- if (i > MR_CACHE_LAST_STD_ENTRY) {
- mlx5_odp_init_mr_cache_entry(ent);
+ if (i > MKEY_CACHE_LAST_STD_ENTRY) {
+ mlx5_odp_init_mkey_cache_entry(ent);
continue;
}
- if (ent->order > mr_cache_max_order(dev))
+ if (ent->order > mkey_cache_max_order(dev))
continue;
ent->page = PAGE_SHIFT;
@@ -781,19 +781,19 @@ int mlx5_mr_cache_init(struct mlx5_ib_dev *dev)
xa_unlock_irq(&ent->mkeys);
}
- mlx5_mr_cache_debugfs_init(dev);
+ mlx5_mkey_cache_debugfs_init(dev);
return 0;
}
-int mlx5_mr_cache_cleanup(struct mlx5_ib_dev *dev)
+int mlx5_mkey_cache_cleanup(struct mlx5_ib_dev *dev)
{
unsigned int i;
if (!dev->cache.wq)
return 0;
- for (i = 0; i < MAX_MR_CACHE_ENTRIES; i++) {
+ for (i = 0; i < MAX_MKEY_CACHE_ENTRIES; i++) {
struct mlx5_cache_ent *ent = &dev->cache.ent[i];
xa_lock_irq(&ent->mkeys);
@@ -802,10 +802,10 @@ int mlx5_mr_cache_cleanup(struct mlx5_ib_dev *dev)
cancel_delayed_work_sync(&ent->dwork);
}
- mlx5_mr_cache_debugfs_cleanup(dev);
+ mlx5_mkey_cache_debugfs_cleanup(dev);
mlx5_cmd_cleanup_async_ctx(&dev->async_ctx);
- for (i = 0; i < MAX_MR_CACHE_ENTRIES; i++)
+ for (i = 0; i < MAX_MKEY_CACHE_ENTRIES; i++)
clean_keys(dev, i);
destroy_workqueue(dev->cache.wq);
@@ -872,22 +872,22 @@ static int get_octo_len(u64 addr, u64 len, int page_shift)
return (npages + 1) / 2;
}
-static int mr_cache_max_order(struct mlx5_ib_dev *dev)
+static int mkey_cache_max_order(struct mlx5_ib_dev *dev)
{
if (MLX5_CAP_GEN(dev->mdev, umr_extended_translation_offset))
- return MR_CACHE_LAST_STD_ENTRY + 2;
+ return MKEY_CACHE_LAST_STD_ENTRY + 2;
return MLX5_MAX_UMR_SHIFT;
}
-static struct mlx5_cache_ent *mr_cache_ent_from_order(struct mlx5_ib_dev *dev,
- unsigned int order)
+static struct mlx5_cache_ent *mkey_cache_ent_from_order(struct mlx5_ib_dev *dev,
+ unsigned int order)
{
- struct mlx5_mr_cache *cache = &dev->cache;
+ struct mlx5_mkey_cache *cache = &dev->cache;
if (order < cache->ent[0].order)
return &cache->ent[0];
order = order - cache->ent[0].order;
- if (order > MR_CACHE_LAST_STD_ENTRY)
+ if (order > MKEY_CACHE_LAST_STD_ENTRY)
return NULL;
return &cache->ent[order];
}
@@ -930,7 +930,7 @@ static struct mlx5_ib_mr *alloc_cacheable_mr(struct ib_pd *pd,
0, iova);
if (WARN_ON(!page_size))
return ERR_PTR(-EINVAL);
- ent = mr_cache_ent_from_order(
+ ent = mkey_cache_ent_from_order(
dev, order_base_2(ib_umem_num_dma_blocks(umem, page_size)));
/*
* Matches access in alloc_cache_mr(). If the MR can't come from the
diff --git a/drivers/infiniband/hw/mlx5/odp.c b/drivers/infiniband/hw/mlx5/odp.c
index 84da5674e1ab..e305bf1dc6c2 100644
--- a/drivers/infiniband/hw/mlx5/odp.c
+++ b/drivers/infiniband/hw/mlx5/odp.c
@@ -1588,7 +1588,7 @@ mlx5_ib_odp_destroy_eq(struct mlx5_ib_dev *dev, struct mlx5_ib_pf_eq *eq)
return err;
}
-void mlx5_odp_init_mr_cache_entry(struct mlx5_cache_ent *ent)
+void mlx5_odp_init_mkey_cache_entry(struct mlx5_cache_ent *ent)
{
if (!(ent->dev->odp_caps.general_caps & IB_ODP_SUPPORT_IMPLICIT))
return;
diff --git a/include/linux/mlx5/driver.h b/include/linux/mlx5/driver.h
index 76d7661e3e63..220597c2f436 100644
--- a/include/linux/mlx5/driver.h
+++ b/include/linux/mlx5/driver.h
@@ -728,10 +728,10 @@ enum {
};
enum {
- MR_CACHE_LAST_STD_ENTRY = 20,
+ MKEY_CACHE_LAST_STD_ENTRY = 20,
MLX5_IMR_MTT_CACHE_ENTRY,
MLX5_IMR_KSM_CACHE_ENTRY,
- MAX_MR_CACHE_ENTRIES
+ MAX_MKEY_CACHE_ENTRIES
};
struct mlx5_profile {
@@ -740,7 +740,7 @@ struct mlx5_profile {
struct {
int size;
int limit;
- } mr_cache[MAX_MR_CACHE_ENTRIES];
+ } mr_cache[MAX_MKEY_CACHE_ENTRIES];
};
struct mlx5_hca_cap {
--
2.17.2
^ permalink raw reply related [flat|nested] 7+ messages in thread
* Re: [PATCH rdma-next v1 0/5] MR cache cleanup
2022-07-26 7:19 [PATCH rdma-next v1 0/5] MR cache cleanup Michael Guralnik
` (4 preceding siblings ...)
2022-07-26 7:19 ` [PATCH rdma-next v1 5/5] RDMA/mlx5: Rename the mkey cache variables and functions Michael Guralnik
@ 2022-07-27 18:17 ` Jason Gunthorpe
5 siblings, 0 replies; 7+ messages in thread
From: Jason Gunthorpe @ 2022-07-27 18:17 UTC (permalink / raw)
To: Michael Guralnik; +Cc: leonro, maorg, linux-kernel, linux-rdma, netdev, saeedm
On Tue, Jul 26, 2022 at 10:19:06AM +0300, Michael Guralnik wrote:
> Hi,
>
> In this series, Aharon continues to clean mlx5 MR cache logic.
>
> Thanks
>
> v1: Change push_mkey to eliminate locking on mail flow
> v0: http://lore.kernel.org/all/cover.1654601897.git.leonro@nvidia.com/
>
> Aharon Landau (5):
> RDMA/mlx5: Replace ent->lock with xa_lock
> RDMA/mlx5: Replace cache list with Xarray
> RDMA/mlx5: Store the number of in_use cache mkeys instead of total_mrs
> RDMA/mlx5: Store in the cache mkeys instead of mrs
> RDMA/mlx5: Rename the mkey cache variables and functions
Applied to for-next, thanks
Jason
^ permalink raw reply [flat|nested] 7+ messages in thread
end of thread, other threads:[~2022-07-28 16:53 UTC | newest]
Thread overview: 7+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2022-07-26 7:19 [PATCH rdma-next v1 0/5] MR cache cleanup Michael Guralnik
2022-07-26 7:19 ` [PATCH rdma-next v1 1/5] RDMA/mlx5: Replace ent->lock with xa_lock Michael Guralnik
2022-07-26 7:19 ` [PATCH rdma-next v1 2/5] RDMA/mlx5: Replace cache list with Xarray Michael Guralnik
2022-07-26 7:19 ` [PATCH rdma-next v1 3/5] RDMA/mlx5: Store the number of in_use cache mkeys instead of total_mrs Michael Guralnik
2022-07-26 7:19 ` [PATCH rdma-next v1 4/5] RDMA/mlx5: Store in the cache mkeys instead of mrs Michael Guralnik
2022-07-26 7:19 ` [PATCH rdma-next v1 5/5] RDMA/mlx5: Rename the mkey cache variables and functions Michael Guralnik
2022-07-27 18:17 ` [PATCH rdma-next v1 0/5] MR cache cleanup Jason Gunthorpe
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).