Re: [PATCH v2 6/8] IB/hns: Replace counting semaphore event_sem with wait_event

From: Binoy Jayan
Date: Tue Oct 25 2016 - 09:00:37 EST


On 25 October 2016 at 17:58, Arnd Bergmann <arnd@xxxxxxxx> wrote:
> On Tuesday, October 25, 2016 5:31:57 PM CEST Binoy Jayan wrote:
>> static int __hns_roce_cmd_mbox_wait(struct hns_roce_dev *hr_dev, u64 in_param,
>> u64 out_param, unsigned long in_modifier,
>> @@ -198,11 +218,12 @@ static int __hns_roce_cmd_mbox_wait(struct hns_roce_dev *hr_dev, u64 in_param,
>> struct hns_roce_cmdq *cmd = &hr_dev->cmd;
>> struct device *dev = &hr_dev->pdev->dev;
>> struct hns_roce_cmd_context *context;
>> - int ret = 0;
>> + int orig_free_head, ret = 0;
>> +
>> + wait_event(cmd->wq, (orig_free_head = atomic_free_node(cmd, -1)) != -1);
>>
>> spin_lock(&cmd->context_lock);
>> - WARN_ON(cmd->free_head < 0);
>> - context = &cmd->context[cmd->free_head];
>> + context = &cmd->context[orig_free_head];
>> context->token += cmd->token_mask + 1;
>> cmd->free_head = context->next;
>> spin_unlock(&cmd->context_lock);
>>
>
> You get the lock in atomic_free_node() and then again right after that.
> Why not combine the two and only take the lock inside of that
> function that returns a context?


Hi Arnd,

I couldn't figure out a way to wait for a node to be free followed by
acquiring a lock
in an atomic fashion. If the lock is acquired after the wait_event,
there could be race
between the wait_event and acquiring the lock. If the lock is acquired
before the
wait_event, the process may goto sleep with the lock held which is not desired.
Could you suggest me of some way to circumvent this?

-Binoy