IB/mlx5: Multithreaded create MR

Use asynchronous commands to execute up to eight concurrent create MR
commands. This is to fill memory caches faster so we keep consuming
from there.  Also, increase timeout for shrinking caches to five
minutes.

Signed-off-by: Eli Cohen <eli@mellanox.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>

8 files changed, 255 insertions, 84 deletions

diff --git a/drivers/infiniband/hw/mlx5/main.c b/drivers/infiniband/hw/mlx5/main.c
index b1a6cb3a2809..306534109627 100644
--- a/drivers/infiniband/hw/mlx5/main.c
+++ b/drivers/infiniband/hw/mlx5/main.c
@@ -745,7 +745,8 @@ static int alloc_pa_mkey(struct mlx5_ib_dev *dev, u32 *key, u32 pdn)
 	seg->qpn_mkey7_0 = cpu_to_be32(0xffffff << 8);
 	seg->start_addr = 0;
 
-	err = mlx5_core_create_mkey(&dev->mdev, &mr, in, sizeof(*in));
+	err = mlx5_core_create_mkey(&dev->mdev, &mr, in, sizeof(*in),
+				    NULL, NULL, NULL);
 	if (err) {
 		mlx5_ib_warn(dev, "failed to create mkey, %d\n", err);
 		goto err_in;
diff --git a/drivers/infiniband/hw/mlx5/mlx5_ib.h b/drivers/infiniband/hw/mlx5/mlx5_ib.h
index 836be9157242..4c134d93d4fc 100644
--- a/drivers/infiniband/hw/mlx5/mlx5_ib.h
+++ b/drivers/infiniband/hw/mlx5/mlx5_ib.h
@@ -262,6 +262,9 @@ struct mlx5_ib_mr {
 	int			npages;
 	struct completion	done;
 	enum ib_wc_status	status;
+	struct mlx5_ib_dev     *dev;
+	struct mlx5_create_mkey_mbox_out out;
+	unsigned long		start;
 };
 
 struct mlx5_ib_fast_reg_page_list {
@@ -323,6 +326,7 @@ struct mlx5_cache_ent {
 	struct mlx5_ib_dev     *dev;
 	struct work_struct	work;
 	struct delayed_work	dwork;
+	int			pending;
 };
 
 struct mlx5_mr_cache {
@@ -358,6 +362,8 @@ struct mlx5_ib_dev {
 	spinlock_t			mr_lock;
 	struct mlx5_ib_resources	devr;
 	struct mlx5_mr_cache		cache;
+	struct timer_list		delay_timer;
+	int				fill_delay;
 };
 
 static inline struct mlx5_ib_cq *to_mibcq(struct mlx5_core_cq *mcq)
diff --git a/drivers/infiniband/hw/mlx5/mr.c b/drivers/infiniband/hw/mlx5/mr.c
index 3453580b1eb2..d7f202290747 100644
--- a/drivers/infiniband/hw/mlx5/mr.c
+++ b/drivers/infiniband/hw/mlx5/mr.c
@@ -35,11 +35,13 @@
 #include <linux/random.h>
 #include <linux/debugfs.h>
 #include <linux/export.h>
+#include <linux/delay.h>
 #include <rdma/ib_umem.h>
 #include "mlx5_ib.h"
 
 enum {
 	DEF_CACHE_SIZE	= 10,
+	MAX_PENDING_REG_MR = 8,
 };
 
 enum {
@@ -63,6 +65,57 @@ static int order2idx(struct mlx5_ib_dev *dev, int order)
 		return order - cache->ent[0].order;
 }
 
+static void reg_mr_callback(int status, void *context)
+{
+	struct mlx5_ib_mr *mr = context;
+	struct mlx5_ib_dev *dev = mr->dev;
+	struct mlx5_mr_cache *cache = &dev->cache;
+	int c = order2idx(dev, mr->order);
+	struct mlx5_cache_ent *ent = &cache->ent[c];
+	u8 key;
+	unsigned long delta = jiffies - mr->start;
+	unsigned long index;
+	unsigned long flags;
+
+	index = find_last_bit(&delta, 8 * sizeof(delta));
+	if (index == 64)
+		index = 0;
+
+	spin_lock_irqsave(&ent->lock, flags);
+	ent->pending--;
+	spin_unlock_irqrestore(&ent->lock, flags);
+	if (status) {
+		mlx5_ib_warn(dev, "async reg mr failed. status %d\n", status);
+		kfree(mr);
+		dev->fill_delay = 1;
+		mod_timer(&dev->delay_timer, jiffies + HZ);
+		return;
+	}
+
+	if (mr->out.hdr.status) {
+		mlx5_ib_warn(dev, "failed - status %d, syndorme 0x%x\n",
+			     mr->out.hdr.status,
+			     be32_to_cpu(mr->out.hdr.syndrome));
+		kfree(mr);
+		dev->fill_delay = 1;
+		mod_timer(&dev->delay_timer, jiffies + HZ);
+		return;
+	}
+
+	spin_lock_irqsave(&dev->mdev.priv.mkey_lock, flags);
+	key = dev->mdev.priv.mkey_key++;
+	spin_unlock_irqrestore(&dev->mdev.priv.mkey_lock, flags);
+	mr->mmr.key = mlx5_idx_to_mkey(be32_to_cpu(mr->out.mkey) & 0xffffff) | key;
+
+	cache->last_add = jiffies;
+
+	spin_lock_irqsave(&ent->lock, flags);
+	list_add_tail(&mr->list, &ent->head);
+	ent->cur++;
+	ent->size++;
+	spin_unlock_irqrestore(&ent->lock, flags);
+}
+
 static int add_keys(struct mlx5_ib_dev *dev, int c, int num)
 {
 	struct mlx5_mr_cache *cache = &dev->cache;
@@ -78,36 +131,39 @@ static int add_keys(struct mlx5_ib_dev *dev, int c, int num)
 		return -ENOMEM;
 
 	for (i = 0; i < num; i++) {
+		if (ent->pending >= MAX_PENDING_REG_MR) {
+			err = -EAGAIN;
+			break;
+		}
+
 		mr = kzalloc(sizeof(*mr), GFP_KERNEL);
 		if (!mr) {
 			err = -ENOMEM;
-			goto out;
+			break;
 		}
 		mr->order = ent->order;
 		mr->umred = 1;
+		mr->dev = dev;
 		in->seg.status = 1 << 6;
 		in->seg.xlt_oct_size = cpu_to_be32((npages + 1) / 2);
 		in->seg.qpn_mkey7_0 = cpu_to_be32(0xffffff << 8);
 		in->seg.flags = MLX5_ACCESS_MODE_MTT | MLX5_PERM_UMR_EN;
 		in->seg.log2_page_size = 12;
 
+		spin_lock_irq(&ent->lock);
+		ent->pending++;
+		spin_unlock_irq(&ent->lock);
+		mr->start = jiffies;
 		err = mlx5_core_create_mkey(&dev->mdev, &mr->mmr, in,
-					    sizeof(*in));
+					    sizeof(*in), reg_mr_callback,
+					    mr, &mr->out);
 		if (err) {
 			mlx5_ib_warn(dev, "create mkey failed %d\n", err);
 			kfree(mr);
-			goto out;
+			break;
 		}
-		cache->last_add = jiffies;
-
-		spin_lock(&ent->lock);
-		list_add_tail(&mr->list, &ent->head);
-		ent->cur++;
-		ent->size++;
-		spin_unlock(&ent->lock);
 	}
 
-out:
 	kfree(in);
 	return err;
 }
@@ -121,16 +177,16 @@ static void remove_keys(struct mlx5_ib_dev *dev, int c, int num)
 	int i;
 
 	for (i = 0; i < num; i++) {
-		spin_lock(&ent->lock);
+		spin_lock_irq(&ent->lock);
 		if (list_empty(&ent->head)) {
-			spin_unlock(&ent->lock);
+			spin_unlock_irq(&ent->lock);
 			return;
 		}
 		mr = list_first_entry(&ent->head, struct mlx5_ib_mr, list);
 		list_del(&mr->list);
 		ent->cur--;
 		ent->size--;
-		spin_unlock(&ent->lock);
+		spin_unlock_irq(&ent->lock);
 		err = mlx5_core_destroy_mkey(&dev->mdev, &mr->mmr);
 		if (err)
 			mlx5_ib_warn(dev, "failed destroy mkey\n");
@@ -162,9 +218,13 @@ static ssize_t size_write(struct file *filp, const char __user *buf,
 		return -EINVAL;
 
 	if (var > ent->size) {
-		err = add_keys(dev, c, var - ent->size);
-		if (err)
-			return err;
+		do {
+			err = add_keys(dev, c, var - ent->size);
+			if (err && err != -EAGAIN)
+				return err;
+
+			usleep_range(3000, 5000);
+		} while (err);
 	} else if (var < ent->size) {
 		remove_keys(dev, c, ent->size - var);
 	}
@@ -280,23 +340,37 @@ static void __cache_work_func(struct mlx5_cache_ent *ent)
 	struct mlx5_ib_dev *dev = ent->dev;
 	struct mlx5_mr_cache *cache = &dev->cache;
 	int i = order2idx(dev, ent->order);
+	int err;
 
 	if (cache->stopped)
 		return;
 
 	ent = &dev->cache.ent[i];
-	if (ent->cur < 2 * ent->limit) {
-		add_keys(dev, i, 1);
-		if (ent->cur < 2 * ent->limit)
-			queue_work(cache->wq, &ent->work);
+	if (ent->cur < 2 * ent->limit && !dev->fill_delay) {
+		err = add_keys(dev, i, 1);
+		if (ent->cur < 2 * ent->limit) {
+			if (err == -EAGAIN) {
+				mlx5_ib_dbg(dev, "returned eagain, order %d\n",
+					    i + 2);
+				queue_delayed_work(cache->wq, &ent->dwork,
+						   msecs_to_jiffies(3));
+			} else if (err) {
+				mlx5_ib_warn(dev, "command failed order %d, err %d\n",
+					     i + 2, err);
+				queue_delayed_work(cache->wq, &ent->dwork,
+						   msecs_to_jiffies(1000));
+			} else {
+				queue_work(cache->wq, &ent->work);
+			}
+		}
 	} else if (ent->cur > 2 * ent->limit) {
 		if (!someone_adding(cache) &&
-		    time_after(jiffies, cache->last_add + 60 * HZ)) {
+		    time_after(jiffies, cache->last_add + 300 * HZ)) {
 			remove_keys(dev, i, 1);
 			if (ent->cur > ent->limit)
 				queue_work(cache->wq, &ent->work);
 		} else {
-			queue_delayed_work(cache->wq, &ent->dwork, 60 * HZ);
+			queue_delayed_work(cache->wq, &ent->dwork, 300 * HZ);
 		}
 	}
 }
@@ -336,18 +410,18 @@ static struct mlx5_ib_mr *alloc_cached_mr(struct mlx5_ib_dev *dev, int order)
 
 		mlx5_ib_dbg(dev, "order %d, cache index %d\n", ent->order, i);
 
-		spin_lock(&ent->lock);
+		spin_lock_irq(&ent->lock);
 		if (!list_empty(&ent->head)) {
 			mr = list_first_entry(&ent->head, struct mlx5_ib_mr,
 					      list);
 			list_del(&mr->list);
 			ent->cur--;
-			spin_unlock(&ent->lock);
+			spin_unlock_irq(&ent->lock);
 			if (ent->cur < ent->limit)
 				queue_work(cache->wq, &ent->work);
 			break;
 		}
-		spin_unlock(&ent->lock);
+		spin_unlock_irq(&ent->lock);
 
 		queue_work(cache->wq, &ent->work);
 
@@ -374,12 +448,12 @@ static void free_cached_mr(struct mlx5_ib_dev *dev, struct mlx5_ib_mr *mr)
 		return;
 	}
 	ent = &cache->ent[c];
-	spin_lock(&ent->lock);
+	spin_lock_irq(&ent->lock);
 	list_add_tail(&mr->list, &ent->head);
 	ent->cur++;
 	if (ent->cur > 2 * ent->limit)
 		shrink = 1;
-	spin_unlock(&ent->lock);
+	spin_unlock_irq(&ent->lock);
 
 	if (shrink)
 		queue_work(cache->wq, &ent->work);
@@ -394,16 +468,16 @@ static void clean_keys(struct mlx5_ib_dev *dev, int c)
 
 	cancel_delayed_work(&ent->dwork);
 	while (1) {
-		spin_lock(&ent->lock);
+		spin_lock_irq(&ent->lock);
 		if (list_empty(&ent->head)) {
-			spin_unlock(&ent->lock);
+			spin_unlock_irq(&ent->lock);
 			return;
 		}
 		mr = list_first_entry(&ent->head, struct mlx5_ib_mr, list);
 		list_del(&mr->list);
 		ent->cur--;
 		ent->size--;
-		spin_unlock(&ent->lock);
+		spin_unlock_irq(&ent->lock);
 		err = mlx5_core_destroy_mkey(&dev->mdev, &mr->mmr);
 		if (err)
 			mlx5_ib_warn(dev, "failed destroy mkey\n");
@@ -464,6 +538,13 @@ static void mlx5_mr_cache_debugfs_cleanup(struct mlx5_ib_dev *dev)
 	debugfs_remove_recursive(dev->cache.root);
 }
 
+static void delay_time_func(unsigned long ctx)
+{
+	struct mlx5_ib_dev *dev = (struct mlx5_ib_dev *)ctx;
+
+	dev->fill_delay = 0;
+}
+
 int mlx5_mr_cache_init(struct mlx5_ib_dev *dev)
 {
 	struct mlx5_mr_cache *cache = &dev->cache;
@@ -479,6 +560,7 @@ int mlx5_mr_cache_init(struct mlx5_ib_dev *dev)
 		return -ENOMEM;
 	}
 
+	setup_timer(&dev->delay_timer, delay_time_func, (unsigned long)dev);
 	for (i = 0; i < MAX_MR_CACHE_ENTRIES; i++) {
 		INIT_LIST_HEAD(&cache->ent[i].head);
 		spin_lock_init(&cache->ent[i].lock);
@@ -522,6 +604,7 @@ int mlx5_mr_cache_cleanup(struct mlx5_ib_dev *dev)
 		clean_keys(dev, i);
 
 	destroy_workqueue(dev->cache.wq);
+	del_timer_sync(&dev->delay_timer);
 
 	return 0;
 }
@@ -551,7 +634,8 @@ struct ib_mr *mlx5_ib_get_dma_mr(struct ib_pd *pd, int acc)
 	seg->qpn_mkey7_0 = cpu_to_be32(0xffffff << 8);
 	seg->start_addr = 0;
 
-	err = mlx5_core_create_mkey(mdev, &mr->mmr, in, sizeof(*in));
+	err = mlx5_core_create_mkey(mdev, &mr->mmr, in, sizeof(*in), NULL, NULL,
+				    NULL);
 	if (err)
 		goto err_in;
 
@@ -660,14 +744,14 @@ static struct mlx5_ib_mr *reg_umr(struct ib_pd *pd, struct ib_umem *umem,
 	int err;
 	int i;
 
-	for (i = 0; i < 10; i++) {
+	for (i = 0; i < 1; i++) {
 		mr = alloc_cached_mr(dev, order);
 		if (mr)
 			break;
 
 		err = add_keys(dev, order2idx(dev, order), 1);
-		if (err) {
-			mlx5_ib_warn(dev, "add_keys failed\n");
+		if (err && err != -EAGAIN) {
+			mlx5_ib_warn(dev, "add_keys failed, err %d\n", err);
 			break;
 		}
 	}
@@ -759,8 +843,10 @@ static struct mlx5_ib_mr *reg_create(struct ib_pd *pd, u64 virt_addr,
 	in->seg.xlt_oct_size = cpu_to_be32(get_octo_len(virt_addr, length, 1 << page_shift));
 	in->seg.log2_page_size = page_shift;
 	in->seg.qpn_mkey7_0 = cpu_to_be32(0xffffff << 8);
-	in->xlat_oct_act_size = cpu_to_be32(get_octo_len(virt_addr, length, 1 << page_shift));
-	err = mlx5_core_create_mkey(&dev->mdev, &mr->mmr, in, inlen);
+	in->xlat_oct_act_size = cpu_to_be32(get_octo_len(virt_addr, length,
+							 1 << page_shift));
+	err = mlx5_core_create_mkey(&dev->mdev, &mr->mmr, in, inlen, NULL,
+				    NULL, NULL);
 	if (err) {
 		mlx5_ib_warn(dev, "create mkey failed\n");
 		goto err_2;
@@ -944,7 +1030,8 @@ struct ib_mr *mlx5_ib_alloc_fast_reg_mr(struct ib_pd *pd,
 	 * TBD not needed - issue 197292 */
 	in->seg.log2_page_size = PAGE_SHIFT;
 
-	err = mlx5_core_create_mkey(&dev->mdev, &mr->mmr, in, sizeof(*in));
+	err = mlx5_core_create_mkey(&dev->mdev, &mr->mmr, in, sizeof(*in), NULL,
+				    NULL, NULL);
 	kfree(in);
 	if (err)
 		goto err_free;
diff --git a/drivers/infiniband/hw/mlx5/qp.c b/drivers/infiniband/hw/mlx5/qp.c
index 5659ea880741..e3881433f5d7 100644
--- a/drivers/infiniband/hw/mlx5/qp.c
+++ b/drivers/infiniband/hw/mlx5/qp.c
@@ -1744,6 +1744,7 @@ static void set_reg_umr_segment(struct mlx5_wqe_umr_ctrl_seg *umr,
 			MLX5_MKEY_MASK_PD		|
 			MLX5_MKEY_MASK_LR		|
 			MLX5_MKEY_MASK_LW		|
+			MLX5_MKEY_MASK_KEY		|
 			MLX5_MKEY_MASK_RR		|
 			MLX5_MKEY_MASK_RW		|
 			MLX5_MKEY_MASK_A		|
@@ -1800,7 +1801,8 @@ static void set_reg_mkey_segment(struct mlx5_mkey_seg *seg, struct ib_send_wr *w
 	seg->start_addr = cpu_to_be64(wr->wr.fast_reg.iova_start);
 	seg->len = cpu_to_be64(wr->wr.fast_reg.length);
 	seg->log2_page_size = wr->wr.fast_reg.page_shift;
-	seg->qpn_mkey7_0 = cpu_to_be32(0xffffff << 8);
+	seg->qpn_mkey7_0 = cpu_to_be32(0xffffff00 |
+				       mlx5_mkey_variant(wr->wr.fast_reg.rkey));
 }
 
 static void set_frwr_pages(struct mlx5_wqe_data_seg *dseg,
diff --git a/drivers/net/ethernet/mellanox/mlx5/core/cmd.c b/drivers/net/ethernet/mellanox/mlx5/core/cmd.c
index 6ca30739625f..8675d26a678b 100644
--- a/drivers/net/ethernet/mellanox/mlx5/core/cmd.c
+++ b/drivers/net/ethernet/mellanox/mlx5/core/cmd.c
@@ -98,6 +98,7 @@ enum {
 static struct mlx5_cmd_work_ent *alloc_cmd(struct mlx5_cmd *cmd,
 					   struct mlx5_cmd_msg *in,
 					   struct mlx5_cmd_msg *out,
+					   void *uout, int uout_size,
 					   mlx5_cmd_cbk_t cbk,
 					   void *context, int page_queue)
 {
@@ -110,6 +111,8 @@ static struct mlx5_cmd_work_ent *alloc_cmd(struct mlx5_cmd *cmd,
 
 	ent->in		= in;
 	ent->out	= out;
+	ent->uout	= uout;
+	ent->uout_size	= uout_size;
 	ent->callback	= cbk;
 	ent->context	= context;
 	ent->cmd	= cmd;
@@ -534,6 +537,7 @@ static void cmd_work_handler(struct work_struct *work)
 	ent->lay = lay;
 	memset(lay, 0, sizeof(*lay));
 	memcpy(lay->in, ent->in->first.data, sizeof(lay->in));
+	ent->op = be32_to_cpu(lay->in[0]) >> 16;
 	if (ent->in->next)
 		lay->in_ptr = cpu_to_be64(ent->in->next->dma);
 	lay->inlen = cpu_to_be32(ent->in->len);
@@ -628,7 +632,8 @@ static int wait_func(struct mlx5_core_dev *dev, struct mlx5_cmd_work_ent *ent)
  *    2. page queue commands do not support asynchrous completion
  */
 static int mlx5_cmd_invoke(struct mlx5_core_dev *dev, struct mlx5_cmd_msg *in,
-			   struct mlx5_cmd_msg *out, mlx5_cmd_cbk_t callback,
+			   struct mlx5_cmd_msg *out, void *uout, int uout_size,
+			   mlx5_cmd_cbk_t callback,
 			   void *context, int page_queue, u8 *status)
 {
 	struct mlx5_cmd *cmd = &dev->cmd;
@@ -642,7 +647,8 @@ static int mlx5_cmd_invoke(struct mlx5_core_dev *dev, struct mlx5_cmd_msg *in,
 	if (callback && page_queue)
 		return -EINVAL;
 
-	ent = alloc_cmd(cmd, in, out, callback, context, page_queue);
+	ent = alloc_cmd(cmd, in, out, uout, uout_size, callback, context,
+			page_queue);
 	if (IS_ERR(ent))
 		return PTR_ERR(ent);
 
@@ -670,10 +676,10 @@ static int mlx5_cmd_invoke(struct mlx5_core_dev *dev, struct mlx5_cmd_msg *in,
 		op = be16_to_cpu(((struct mlx5_inbox_hdr *)in->first.data)->opcode);
 		if (op < ARRAY_SIZE(cmd->stats)) {
 			stats = &cmd->stats[op];
-			spin_lock(&stats->lock);
+			spin_lock_irq(&stats->lock);
 			stats->sum += ds;
 			++stats->n;
-			spin_unlock(&stats->lock);
+			spin_unlock_irq(&stats->lock);
 		}
 		mlx5_core_dbg_mask(dev, 1 << MLX5_CMD_TIME,
 				   "fw exec time for %s is %lld nsec\n",
@@ -826,7 +832,7 @@ static struct mlx5_cmd_msg *mlx5_alloc_cmd_msg(struct mlx5_core_dev *dev,
 	int n;
 	int i;
 
-	msg = kzalloc(sizeof(*msg), GFP_KERNEL);
+	msg = kzalloc(sizeof(*msg), flags);
 	if (!msg)
 		return ERR_PTR(-ENOMEM);
 
@@ -1109,6 +1115,19 @@ void mlx5_cmd_use_polling(struct mlx5_core_dev *dev)
 		up(&cmd->sem);
 }
 
+static void free_msg(struct mlx5_core_dev *dev, struct mlx5_cmd_msg *msg)
+{
+	unsigned long flags;
+
+	if (msg->cache) {
+		spin_lock_irqsave(&msg->cache->lock, flags);
+		list_add_tail(&msg->list, &msg->cache->head);
+		spin_unlock_irqrestore(&msg->cache->lock, flags);
+	} else {
+		mlx5_free_cmd_msg(dev, msg);
+	}
+}
+
 void mlx5_cmd_comp_handler(struct mlx5_core_dev *dev, unsigned long vector)
 {
 	struct mlx5_cmd *cmd = &dev->cmd;
@@ -1117,6 +1136,10 @@ void mlx5_cmd_comp_handler(struct mlx5_core_dev *dev, unsigned long vector)
 	void *context;
 	int err;
 	int i;
+	ktime_t t1, t2, delta;
+	s64 ds;
+	struct mlx5_cmd_stats *stats;
+	unsigned long flags;
 
 	for (i = 0; i < (1 << cmd->log_sz); i++) {
 		if (test_bit(i, &vector)) {
@@ -1141,9 +1164,29 @@ void mlx5_cmd_comp_handler(struct mlx5_core_dev *dev, unsigned long vector)
 			}
 			free_ent(cmd, ent->idx);
 			if (ent->callback) {
+				t1 = timespec_to_ktime(ent->ts1);
+				t2 = timespec_to_ktime(ent->ts2);
+				delta = ktime_sub(t2, t1);
+				ds = ktime_to_ns(delta);
+				if (ent->op < ARRAY_SIZE(cmd->stats)) {
+					stats = &cmd->stats[ent->op];
+					spin_lock_irqsave(&stats->lock, flags);
+					stats->sum += ds;
+					++stats->n;
+					spin_unlock_irqrestore(&stats->lock, flags);
+				}
+
 				callback = ent->callback;
 				context = ent->context;
 				err = ent->ret;
+				if (!err)
+					err = mlx5_copy_from_msg(ent->uout,
+								 ent->out,
+								 ent->uout_size);
+
+				mlx5_free_cmd_msg(dev, ent->out);
+				free_msg(dev, ent->in);
+
 				free_cmd(ent);
 				callback(err, context);
 			} else {
@@ -1160,7 +1203,8 @@ static int status_to_err(u8 status)
 	return status ? -1 : 0; /* TBD more meaningful codes */
 }
 
-static struct mlx5_cmd_msg *alloc_msg(struct mlx5_core_dev *dev, int in_size)
+static struct mlx5_cmd_msg *alloc_msg(struct mlx5_core_dev *dev, int in_size,
+				      gfp_t gfp)
 {
 	struct mlx5_cmd_msg *msg = ERR_PTR(-ENOMEM);
 	struct mlx5_cmd *cmd = &dev->cmd;
@@ -1172,7 +1216,7 @@ static struct mlx5_cmd_msg *alloc_msg(struct mlx5_core_dev *dev, int in_size)
 		ent = &cmd->cache.med;
 
 	if (ent) {
-		spin_lock(&ent->lock);
+		spin_lock_irq(&ent->lock);
 		if (!list_empty(&ent->head)) {
 			msg = list_entry(ent->head.next, typeof(*msg), list);
 			/* For cached lists, we must explicitly state what is
@@ -1181,43 +1225,34 @@ static struct mlx5_cmd_msg *alloc_msg(struct mlx5_core_dev *dev, int in_size)
 			msg->len = in_size;
 			list_del(&msg->list);
 		}
-		spin_unlock(&ent->lock);
+		spin_unlock_irq(&ent->lock);
 	}
 
 	if (IS_ERR(msg))
-		msg = mlx5_alloc_cmd_msg(dev, GFP_KERNEL, in_size);
+		msg = mlx5_alloc_cmd_msg(dev, gfp, in_size);
 
 	return msg;
 }
 
-static void free_msg(struct mlx5_core_dev *dev, struct mlx5_cmd_msg *msg)
-{
-	if (msg->cache) {
-		spin_lock(&msg->cache->lock);
-		list_add_tail(&msg->list, &msg->cache->head);
-		spin_unlock(&msg->cache->lock);
-	} else {
-		mlx5_free_cmd_msg(dev, msg);
-	}
-}
-
 static int is_manage_pages(struct mlx5_inbox_hdr *in)
 {
 	return be16_to_cpu(in->opcode) == MLX5_CMD_OP_MANAGE_PAGES;
 }
 
-int mlx5_cmd_exec(struct mlx5_core_dev *dev, void *in, int in_size, void *out,
-		  int out_size)
+static int cmd_exec(struct mlx5_core_dev *dev, void *in, int in_size, void *out,
+		    int out_size, mlx5_cmd_cbk_t callback, void *context)
 {
 	struct mlx5_cmd_msg *inb;
 	struct mlx5_cmd_msg *outb;
 	int pages_queue;
+	gfp_t gfp;
 	int err;
 	u8 status = 0;
 
 	pages_queue = is_manage_pages(in);
+	gfp = callback ? GFP_ATOMIC : GFP_KERNEL;
 
-	inb = alloc_msg(dev, in_size);
+	inb = alloc_msg(dev, in_size, gfp);
 	if (IS_ERR(inb)) {
 		err = PTR_ERR(inb);
 		return err;
@@ -1229,13 +1264,14 @@ int mlx5_cmd_exec(struct mlx5_core_dev *dev, void *in, int in_size, void *out,
 		goto out_in;
 	}
 
-	outb = mlx5_alloc_cmd_msg(dev, GFP_KERNEL, out_size);
+	outb = mlx5_alloc_cmd_msg(dev, gfp, out_size);
 	if (IS_ERR(outb)) {
 		err = PTR_ERR(outb);
 		goto out_in;
 	}
 
-	err = mlx5_cmd_invoke(dev, inb, outb, NULL, NULL, pages_queue, &status);
+	err = mlx5_cmd_invoke(dev, inb, outb, out, out_size, callback, context,
+			      pages_queue, &status);
 	if (err)
 		goto out_out;
 
@@ -1248,14 +1284,30 @@ int mlx5_cmd_exec(struct mlx5_core_dev *dev, void *in, int in_size, void *out,
 	err = mlx5_copy_from_msg(out, outb, out_size);
 
 out_out:
-	mlx5_free_cmd_msg(dev, outb);
+	if (!callback)
+		mlx5_free_cmd_msg(dev, outb);
 
 out_in:
-	free_msg(dev, inb);
+	if (!callback)
+		free_msg(dev, inb);
 	return err;
 }
+
+int mlx5_cmd_exec(struct mlx5_core_dev *dev, void *in, int in_size, void *out,
+		  int out_size)
+{
+	return cmd_exec(dev, in, in_size, out, out_size, NULL, NULL);
+}
 EXPORT_SYMBOL(mlx5_cmd_exec);
 
+int mlx5_cmd_exec_cb(struct mlx5_core_dev *dev, void *in, int in_size,
+		     void *out, int out_size, mlx5_cmd_cbk_t callback,
+		     void *context)
+{
+	return cmd_exec(dev, in, in_size, out, out_size, callback, context);
+}
+EXPORT_SYMBOL(mlx5_cmd_exec_cb);
+
 static void destroy_msg_cache(struct mlx5_core_dev *dev)
 {
 	struct mlx5_cmd *cmd = &dev->cmd;
diff --git a/drivers/net/ethernet/mellanox/mlx5/core/debugfs.c b/drivers/net/ethernet/mellanox/mlx5/core/debugfs.c
index 9c7194b26ee2..80f6d127257a 100644
--- a/drivers/net/ethernet/mellanox/mlx5/core/debugfs.c
+++ b/drivers/net/ethernet/mellanox/mlx5/core/debugfs.c
@@ -154,10 +154,10 @@ static ssize_t average_read(struct file *filp, char __user *buf, size_t count,
 		return 0;
 
 	stats = filp->private_data;
-	spin_lock(&stats->lock);
+	spin_lock_irq(&stats->lock);
 	if (stats->n)
 		field = div64_u64(stats->sum, stats->n);
-	spin_unlock(&stats->lock);
+	spin_unlock_irq(&stats->lock);
 	ret = snprintf(tbuf, sizeof(tbuf), "%llu\n", field);
 	if (ret > 0) {
 		if (copy_to_user(buf, tbuf, ret))
@@ -175,10 +175,10 @@ static ssize_t average_write(struct file *filp, const char __user *buf,
 	struct mlx5_cmd_stats *stats;
 
 	stats = filp->private_data;
-	spin_lock(&stats->lock);
+	spin_lock_irq(&stats->lock);
 	stats->sum = 0;
 	stats->n = 0;
-	spin_unlock(&stats->lock);
+	spin_unlock_irq(&stats->lock);
 
 	*pos += count;
 
diff --git a/drivers/net/ethernet/mellanox/mlx5/core/mr.c b/drivers/net/ethernet/mellanox/mlx5/core/mr.c
index 5b44e2e46daf..35e514dc7b7d 100644
--- a/drivers/net/ethernet/mellanox/mlx5/core/mr.c
+++ b/drivers/net/ethernet/mellanox/mlx5/core/mr.c
@@ -37,31 +37,41 @@
 #include "mlx5_core.h"
 
 int mlx5_core_create_mkey(struct mlx5_core_dev *dev, struct mlx5_core_mr *mr,
-			  struct mlx5_create_mkey_mbox_in *in, int inlen)
+			  struct mlx5_create_mkey_mbox_in *in, int inlen,
+			  mlx5_cmd_cbk_t callback, void *context,
+			  struct mlx5_create_mkey_mbox_out *out)
 {
-	struct mlx5_create_mkey_mbox_out out;
+	struct mlx5_create_mkey_mbox_out lout;
 	int err;
 	u8 key;
 
-	memset(&out, 0, sizeof(out));
-	spin_lock(&dev->priv.mkey_lock);
+	memset(&lout, 0, sizeof(lout));
+	spin_lock_irq(&dev->priv.mkey_lock);
 	key = dev->priv.mkey_key++;
-	spin_unlock(&dev->priv.mkey_lock);
+	spin_unlock_irq(&dev->priv.mkey_lock);
 	in->seg.qpn_mkey7_0 |= cpu_to_be32(key);
 	in->hdr.opcode = cpu_to_be16(MLX5_CMD_OP_CREATE_MKEY);
-	err = mlx5_cmd_exec(dev, in, inlen, &out, sizeof(out));
+	if (callback) {
+		err = mlx5_cmd_exec_cb(dev, in, inlen, out, sizeof(*out),
+				       callback, context);
+		return err;
+	} else {
+		err = mlx5_cmd_exec(dev, in, inlen, &lout, sizeof(lout));
+	}
+
 	if (err) {
 		mlx5_core_dbg(dev, "cmd exec faile %d\n", err);
 		return err;
 	}
 
-	if (out.hdr.status) {
-		mlx5_core_dbg(dev, "status %d\n", out.hdr.status);
-		return mlx5_cmd_status_to_err(&out.hdr);
+	if (lout.hdr.status) {
+		mlx5_core_dbg(dev, "status %d\n", lout.hdr.status);
+		return mlx5_cmd_status_to_err(&lout.hdr);
 	}
 
-	mr->key = mlx5_idx_to_mkey(be32_to_cpu(out.mkey) & 0xffffff) | key;
-	mlx5_core_dbg(dev, "out 0x%x, key 0x%x, mkey 0x%x\n", be32_to_cpu(out.mkey), key, mr->key);
+	mr->key = mlx5_idx_to_mkey(be32_to_cpu(lout.mkey) & 0xffffff) | key;
+	mlx5_core_dbg(dev, "out 0x%x, key 0x%x, mkey 0x%x\n",
+		      be32_to_cpu(lout.mkey), key, mr->key);
 
 	return err;
 }
diff --git a/include/linux/mlx5/driver.h b/include/linux/mlx5/driver.h
index 6b8c496572c8..513619a75695 100644
--- a/include/linux/mlx5/driver.h
+++ b/include/linux/mlx5/driver.h
@@ -557,9 +557,11 @@ typedef void (*mlx5_cmd_cbk_t)(int status, void *context);
 struct mlx5_cmd_work_ent {
 	struct mlx5_cmd_msg    *in;
 	struct mlx5_cmd_msg    *out;
+	void		       *uout;
+	int			uout_size;
 	mlx5_cmd_cbk_t		callback;
 	void		       *context;
-	int idx;
+	int			idx;
 	struct completion	done;
 	struct mlx5_cmd        *cmd;
 	struct work_struct	work;
@@ -570,6 +572,7 @@ struct mlx5_cmd_work_ent {
 	u8			token;
 	struct timespec		ts1;
 	struct timespec		ts2;
+	u16			op;
 };
 
 struct mlx5_pas {
@@ -653,6 +656,9 @@ void mlx5_cmd_use_polling(struct mlx5_core_dev *dev);
 int mlx5_cmd_status_to_err(struct mlx5_outbox_hdr *hdr);
 int mlx5_cmd_exec(struct mlx5_core_dev *dev, void *in, int in_size, void *out,
 		  int out_size);
+int mlx5_cmd_exec_cb(struct mlx5_core_dev *dev, void *in, int in_size,
+		     void *out, int out_size, mlx5_cmd_cbk_t callback,
+		     void *context);
 int mlx5_cmd_alloc_uar(struct mlx5_core_dev *dev, u32 *uarn);
 int mlx5_cmd_free_uar(struct mlx5_core_dev *dev, u32 uarn);
 int mlx5_alloc_uuars(struct mlx5_core_dev *dev, struct mlx5_uuar_info *uuari);
@@ -676,7 +682,9 @@ int mlx5_core_query_srq(struct mlx5_core_dev *dev, struct mlx5_core_srq *srq,
 int mlx5_core_arm_srq(struct mlx5_core_dev *dev, struct mlx5_core_srq *srq,
 		      u16 lwm, int is_srq);
 int mlx5_core_create_mkey(struct mlx5_core_dev *dev, struct mlx5_core_mr *mr,
-			  struct mlx5_create_mkey_mbox_in *in, int inlen);
+			  struct mlx5_create_mkey_mbox_in *in, int inlen,
+			  mlx5_cmd_cbk_t callback, void *context,
+			  struct mlx5_create_mkey_mbox_out *out);
 int mlx5_core_destroy_mkey(struct mlx5_core_dev *dev, struct mlx5_core_mr *mr);
 int mlx5_core_query_mkey(struct mlx5_core_dev *dev, struct mlx5_core_mr *mr,
 			 struct mlx5_query_mkey_mbox_out *out, int outlen);
@@ -745,6 +753,11 @@ static inline u32 mlx5_idx_to_mkey(u32 mkey_idx)
 	return mkey_idx << 8;
 }
 
+static inline u8 mlx5_mkey_variant(u32 mkey)
+{
+	return mkey & 0xff;
+}
+
 enum {
 	MLX5_PROF_MASK_QP_SIZE		= (u64)1 << 0,
 	MLX5_PROF_MASK_MR_CACHE		= (u64)1 << 1,