diff mbox series

[for-next,1/3] RDMA/hns: Enable all CMDQ context

Message ID 1617262341-37571-2-git-send-email-liweihang@huawei.com (mailing list archive)
State Accepted
Delegated to: Jason Gunthorpe
Headers show
Series RDMA/hns: Updates of CMDQ | expand

Commit Message

Weihang Li April 1, 2021, 7:32 a.m. UTC
From: Lang Cheng <chenglang@huawei.com>

Fix error of cmd's context number calculation algorithm to enable all of
32 cmd entries and support 32 concurrent accesses.

Signed-off-by: Lang Cheng <chenglang@huawei.com>
Signed-off-by: Weihang Li <liweihang@huawei.com>
---
 drivers/infiniband/hw/hns/hns_roce_cmd.c    | 62 ++++++++++++-----------------
 drivers/infiniband/hw/hns/hns_roce_device.h |  6 +--
 2 files changed, 27 insertions(+), 41 deletions(-)

Comments

Leon Romanovsky April 1, 2021, 1:08 p.m. UTC | #1
On Thu, Apr 01, 2021 at 03:32:19PM +0800, Weihang Li wrote:
> From: Lang Cheng <chenglang@huawei.com>
> 
> Fix error of cmd's context number calculation algorithm to enable all of
> 32 cmd entries and support 32 concurrent accesses.
> 
> Signed-off-by: Lang Cheng <chenglang@huawei.com>
> Signed-off-by: Weihang Li <liweihang@huawei.com>
> ---
>  drivers/infiniband/hw/hns/hns_roce_cmd.c    | 62 ++++++++++++-----------------
>  drivers/infiniband/hw/hns/hns_roce_device.h |  6 +--
>  2 files changed, 27 insertions(+), 41 deletions(-)

<...>

> -	WARN_ON(cmd->free_head < 0);
> -	context = &cmd->context[cmd->free_head];
> -	context->token += cmd->token_mask + 1;
> -	cmd->free_head = context->next;
> +
> +	do {
> +		context = &cmd->context[cmd->free_head];
> +		cmd->free_head = context->next;
> +	} while (context->busy);
> +
> +	context->busy = 1;

This "busy" flag after do-while together with release in __hns_roce_cmd_mbox_wait()
is interesting thing. Are you sure that it won't loop forever here?

Thanks
Lang Cheng April 2, 2021, 1:26 a.m. UTC | #2
On 2021/4/1 21:08, Leon Romanovsky wrote:
> On Thu, Apr 01, 2021 at 03:32:19PM +0800, Weihang Li wrote:
>> From: Lang Cheng <chenglang@huawei.com>
>>
>> Fix error of cmd's context number calculation algorithm to enable all of
>> 32 cmd entries and support 32 concurrent accesses.
>>
>> Signed-off-by: Lang Cheng <chenglang@huawei.com>
>> Signed-off-by: Weihang Li <liweihang@huawei.com>
>> ---
>>   drivers/infiniband/hw/hns/hns_roce_cmd.c    | 62 ++++++++++++-----------------
>>   drivers/infiniband/hw/hns/hns_roce_device.h |  6 +--
>>   2 files changed, 27 insertions(+), 41 deletions(-)
> 
> <...>
> 
>> -	WARN_ON(cmd->free_head < 0);
>> -	context = &cmd->context[cmd->free_head];
>> -	context->token += cmd->token_mask + 1;
>> -	cmd->free_head = context->next;
>> +
>> +	do {
>> +		context = &cmd->context[cmd->free_head];
>> +		cmd->free_head = context->next;
>> +	} while (context->busy);
>> +
>> +	context->busy = 1;
> 
> This "busy" flag after do-while together with release in __hns_roce_cmd_mbox_wait()
> is interesting thing. Are you sure that it won't loop forever here?
> 

When initializing resources in hns_roce_cmd_use_events(), ensure that 
the number of semaphores is consistent with the depth of context[].

int hns_roce_cmd_use_events( )
{
	hr_cmd->context = kcalloc(hr_cmd->max_cmds, ...);
	sema_init(&hr_cmd->event_sem, hr_cmd->max_cmds);
}

Then, when someone gets the event_sem in hns_roce_cmd_mbox_wait(), it 
means that there must be a not busy context.

Thanks.

> Thanks
> .
>
Leon Romanovsky April 4, 2021, 10:14 a.m. UTC | #3
On Fri, Apr 02, 2021 at 09:26:38AM +0800, Lang Cheng wrote:
> 
> 
> On 2021/4/1 21:08, Leon Romanovsky wrote:
> > On Thu, Apr 01, 2021 at 03:32:19PM +0800, Weihang Li wrote:
> > > From: Lang Cheng <chenglang@huawei.com>
> > > 
> > > Fix error of cmd's context number calculation algorithm to enable all of
> > > 32 cmd entries and support 32 concurrent accesses.
> > > 
> > > Signed-off-by: Lang Cheng <chenglang@huawei.com>
> > > Signed-off-by: Weihang Li <liweihang@huawei.com>
> > > ---
> > >   drivers/infiniband/hw/hns/hns_roce_cmd.c    | 62 ++++++++++++-----------------
> > >   drivers/infiniband/hw/hns/hns_roce_device.h |  6 +--
> > >   2 files changed, 27 insertions(+), 41 deletions(-)
> > 
> > <...>
> > 
> > > -	WARN_ON(cmd->free_head < 0);
> > > -	context = &cmd->context[cmd->free_head];
> > > -	context->token += cmd->token_mask + 1;
> > > -	cmd->free_head = context->next;
> > > +
> > > +	do {
> > > +		context = &cmd->context[cmd->free_head];
> > > +		cmd->free_head = context->next;
> > > +	} while (context->busy);
> > > +
> > > +	context->busy = 1;
> > 
> > This "busy" flag after do-while together with release in __hns_roce_cmd_mbox_wait()
> > is interesting thing. Are you sure that it won't loop forever here?
> > 
> 
> When initializing resources in hns_roce_cmd_use_events(), ensure that the
> number of semaphores is consistent with the depth of context[].
> 
> int hns_roce_cmd_use_events( )
> {
> 	hr_cmd->context = kcalloc(hr_cmd->max_cmds, ...);
> 	sema_init(&hr_cmd->event_sem, hr_cmd->max_cmds);
> }
> 
> Then, when someone gets the event_sem in hns_roce_cmd_mbox_wait(), it means
> that there must be a not busy context.

OK, thanks

> 
> Thanks.
> 
> > Thanks
> > .
> > 

> null
diff mbox series

Patch

diff --git a/drivers/infiniband/hw/hns/hns_roce_cmd.c b/drivers/infiniband/hw/hns/hns_roce_cmd.c
index 339e3fd..46900d0 100644
--- a/drivers/infiniband/hw/hns/hns_roce_cmd.c
+++ b/drivers/infiniband/hw/hns/hns_roce_cmd.c
@@ -38,22 +38,14 @@ 
 
 #define CMD_POLL_TOKEN 0xffff
 #define CMD_MAX_NUM 32
-#define CMD_TOKEN_MASK 0x1f
 
 static int hns_roce_cmd_mbox_post_hw(struct hns_roce_dev *hr_dev, u64 in_param,
 				     u64 out_param, u32 in_modifier,
 				     u8 op_modifier, u16 op, u16 token,
 				     int event)
 {
-	struct hns_roce_cmdq *cmd = &hr_dev->cmd;
-	int ret;
-
-	mutex_lock(&cmd->hcr_mutex);
-	ret = hr_dev->hw->post_mbox(hr_dev, in_param, out_param, in_modifier,
-				    op_modifier, op, token, event);
-	mutex_unlock(&cmd->hcr_mutex);
-
-	return ret;
+	return hr_dev->hw->post_mbox(hr_dev, in_param, out_param, in_modifier,
+				     op_modifier, op, token, event);
 }
 
 /* this should be called with "poll_sem" */
@@ -96,15 +88,18 @@  void hns_roce_cmd_event(struct hns_roce_dev *hr_dev, u16 token, u8 status,
 	struct hns_roce_cmd_context *context =
 		&hr_dev->cmd.context[token % hr_dev->cmd.max_cmds];
 
-	if (token != context->token)
+	if (unlikely(token != context->token)) {
+		dev_err_ratelimited(hr_dev->dev,
+				    "[cmd] invalid ae token %x,context token is %x!\n",
+				    token, context->token);
 		return;
+	}
 
 	context->result = (status == HNS_ROCE_CMD_SUCCESS) ? 0 : (-EIO);
 	context->out_param = out_param;
 	complete(&context->done);
 }
 
-/* this should be called with "use_events" */
 static int __hns_roce_cmd_mbox_wait(struct hns_roce_dev *hr_dev, u64 in_param,
 				    u64 out_param, unsigned long in_modifier,
 				    u8 op_modifier, u16 op,
@@ -116,13 +111,18 @@  static int __hns_roce_cmd_mbox_wait(struct hns_roce_dev *hr_dev, u64 in_param,
 	int ret;
 
 	spin_lock(&cmd->context_lock);
-	WARN_ON(cmd->free_head < 0);
-	context = &cmd->context[cmd->free_head];
-	context->token += cmd->token_mask + 1;
-	cmd->free_head = context->next;
+
+	do {
+		context = &cmd->context[cmd->free_head];
+		cmd->free_head = context->next;
+	} while (context->busy);
+
+	context->busy = 1;
+	context->token += cmd->max_cmds;
+
 	spin_unlock(&cmd->context_lock);
 
-	init_completion(&context->done);
+	reinit_completion(&context->done);
 
 	ret = hns_roce_cmd_mbox_post_hw(hr_dev, in_param, out_param,
 					in_modifier, op_modifier, op,
@@ -130,30 +130,19 @@  static int __hns_roce_cmd_mbox_wait(struct hns_roce_dev *hr_dev, u64 in_param,
 	if (ret)
 		goto out;
 
-	/*
-	 * It is timeout when wait_for_completion_timeout return 0
-	 * The return value is the time limit set in advance
-	 * how many seconds showing
-	 */
 	if (!wait_for_completion_timeout(&context->done,
 					 msecs_to_jiffies(timeout))) {
-		dev_err(dev, "[cmd]wait_for_completion_timeout timeout\n");
+		dev_err(dev, "[cmd] token %x timeout, drop.\n", context->token);
 		ret = -EBUSY;
 		goto out;
 	}
 
 	ret = context->result;
-	if (ret) {
-		dev_err(dev, "[cmd]event mod cmd process error!err=%d\n", ret);
-		goto out;
-	}
+	if (ret)
+		dev_err(dev, "[cmd] token %x error %d\n", context->token, ret);
 
 out:
-	spin_lock(&cmd->context_lock);
-	context->next = cmd->free_head;
-	cmd->free_head = context - cmd->context;
-	spin_unlock(&cmd->context_lock);
-
+	context->busy = 0;
 	return ret;
 }
 
@@ -208,7 +197,6 @@  int hns_roce_cmd_init(struct hns_roce_dev *hr_dev)
 {
 	struct device *dev = hr_dev->dev;
 
-	mutex_init(&hr_dev->cmd.hcr_mutex);
 	sema_init(&hr_dev->cmd.poll_sem, 1);
 	hr_dev->cmd.use_events = 0;
 	hr_dev->cmd.max_cmds = CMD_MAX_NUM;
@@ -239,16 +227,16 @@  int hns_roce_cmd_use_events(struct hns_roce_dev *hr_dev)
 	for (i = 0; i < hr_cmd->max_cmds; ++i) {
 		hr_cmd->context[i].token = i;
 		hr_cmd->context[i].next = i + 1;
+		init_completion(&hr_cmd->context[i].done);
 	}
-
-	hr_cmd->context[hr_cmd->max_cmds - 1].next = -1;
+	hr_cmd->context[hr_cmd->max_cmds - 1].next = 0;
 	hr_cmd->free_head = 0;
 
 	sema_init(&hr_cmd->event_sem, hr_cmd->max_cmds);
 	spin_lock_init(&hr_cmd->context_lock);
 
-	hr_cmd->token_mask = CMD_TOKEN_MASK;
 	hr_cmd->use_events = 1;
+	down(&hr_cmd->poll_sem);
 
 	return 0;
 }
@@ -259,6 +247,8 @@  void hns_roce_cmd_use_polling(struct hns_roce_dev *hr_dev)
 
 	kfree(hr_cmd->context);
 	hr_cmd->use_events = 0;
+
+	up(&hr_cmd->poll_sem);
 }
 
 struct hns_roce_cmd_mailbox *
diff --git a/drivers/infiniband/hw/hns/hns_roce_device.h b/drivers/infiniband/hw/hns/hns_roce_device.h
index eb2ccb8..9c20c3a 100644
--- a/drivers/infiniband/hw/hns/hns_roce_device.h
+++ b/drivers/infiniband/hw/hns/hns_roce_device.h
@@ -556,6 +556,7 @@  struct hns_roce_cmd_context {
 	int			next;
 	u64			out_param;
 	u16			token;
+	u16			busy;
 };
 
 struct hns_roce_cmdq {
@@ -572,11 +573,6 @@  struct hns_roce_cmdq {
 	int			free_head;
 	struct hns_roce_cmd_context *context;
 	/*
-	 * Result of get integer part
-	 * which max_comds compute according a power of 2
-	 */
-	u16			token_mask;
-	/*
 	 * Process whether use event mode, init default non-zero
 	 * After the event queue of cmd event ready,
 	 * can switch into event mode