Message-ID: <1b8bda3b-c514-7e46-08bf-3ea50ea68096@redhat.com>
Date:   Wed, 27 Sep 2017 08:21:13 -0400
From:   Doug Ledford <dledford@...hat.com>
To:     "Wei Hu (Xavier)" <xavier.huwei@....com>,
        Leon Romanovsky <leon@...nel.org>
Cc:     "Wei Hu (Xavier)" <xavier.huwei@...wei.com>,
        linux-rdma@...r.kernel.org, lijun_nudt@....com, oulijun@...wei.com,
        charles.chenxin@...wei.com, liuyixian@...wei.com,
        xushaobo2@...wei.com, zhangxiping3@...wei.com, linuxarm@...wei.com,
        linux-kernel@...r.kernel.org
Subject: Re: [PATCH for-next 05/20] RDMA/hns: Add command queue support for
 hip08 RoCE driver

On 9/26/2017 10:46 PM, Wei Hu (Xavier) wrote:
> 
> 
> On 2017/9/27 0:18, Doug Ledford wrote:
>> On 9/26/2017 9:13 AM, Wei Hu (Xavier) wrote:
>>>
>>> On 2017/9/26 1:36, Doug Ledford wrote:
>>>> On Mon, 2017-09-25 at 20:18 +0300, Leon Romanovsky wrote:
>>>>> On Mon, Sep 25, 2017 at 01:06:53PM -0400, Doug Ledford wrote:
>>>>>> On Wed, 2017-08-30 at 17:23 +0800, Wei Hu (Xavier) wrote:
>>>>>>
>>>>>>> +    /*
>>>>>>> +     * If the command is sync, wait for the firmware to
>>>>>>> write
>>>>>>> back,
>>>>>>> +     * if multi descriptors to be sent, use the first one to
>>>>>>> check
>>>>>>> +     */
>>>>>>> +    if ((desc->flag) & HNS_ROCE_CMD_FLAG_NO_INTR) {
>>>>>>> +        do {
>>>>>>> +            if (hns_roce_cmq_csq_done(hr_dev))
>>>>>>> +                break;
>>>>>>> +            usleep_range(1000, 2000);
>>>>>>> +            timeout++;
>>>>>>> +        } while (timeout < priv->cmq.tx_timeout);
>>>>>>> +    }
>>>>>> then we spin here for a maximum amount of time between 200 and
>>>>>> 400ms,
>>>>>> so 1/4 to 1/2 a second.  All the time we are holding the bh lock on
>>>>>> this CPU.  That seems excessive to me.  If we are going to spin
>>>>>> that
>>>>>> long, can you find a way to allocate/reserve your resources, send
>>>>>> the
>>>>>> command, then drop the bh lock while you spin, and retake it before
>>>>>> you
>>>>>> complete once the spinning is done?
>>>>> They don't allocate anything in this loop, but checking the pointers
>>>>> are
>>>>> the same, see hns_roce_cmq_csq_done.
>>>> I'm not sure I understand your intended implication of your comment.  I
>>>> wasn't concerned about them allocating anything, only that if the
>>>> hardware is hung, then this loop will hang out for 1/4 to 1/2 a second
>>>> and hold up all bottom half processing on this CPU in the meantime.
>>>> That's the sort of things that provides poor overall system behavior.
>>>>
>>>> Now, since they are really only checking to see if the hardware has
>>>> gotten around to their particular command, and their command is part of
>>>> a ring structure, it's possible to record the original head command,
>>>> and our new head command, and then release the spin_lock_bh around the
>>>> entire do{ }while construct, and in hns_roce_cmd_csq_done() you could
>>>> check that head is not in the range old_head:new_head.  That would
>>>> protect you in case something in the bottom half processing queued up
>>>> some more commands and from one sleep to the next the head jumped from
>>>> something other than the new_head to something past new_head, so that
>>>> head == priv->cmq.csq.next_to_use ends up being perpetually false.
>>>> But, that's just from a quick read of the code, I could easily be
>>>> missing something here...
>>> Hi, Doug
>>>      Driver issues the cmds in cmq, and firmware gets and processes
>>> them.
>>>      The firmware process only one cmd at the same time, and it will
>>> take
>>>      about serveral to 200 us in one cmd currently, so the driver need
>>>      not to use stream mode to issue cmd.
>> I'm not sure I understand your response here.
>>
>> I get that the driver issues cmds in the cmq, and that the firmware gets
>> them and processes them.
>>
>> I get that the firmware will only work on one command at a time and only
>> move to the next one once the current one is complete.
>>
>> I get that commands take anywhere from a few usec to a couple hundred
>> usec.
>>
>> I also get that because you are sleeping for somewhere in between 1000
>> and 2000 usecs, that the driver could easily finish a whole slew of
>> commands.  It could do 10 slow commands, or 100 or more fast commands.
>> What this tells me is that the only reason your current implementation
>> of hns_roce_cmq_csq_done() works at all is because you keep the device
>> locked out from any other commands being put on the queue.  As far as I
>> can tell, that's the only way you can guarantee that at some point you
>> will wake up and the head pointer will be exactly at csq->next_to_use.
>> Otherwise, if you didn't block them out, then you could sleep with the
>> head pointer before csq->next_to_use and wake up the next time with it
>> already well past csq->next_to_use.  Am I right about that?  While you
>> are waiting on this command queue, any other commands are blocked from
>> being placed on the command queue?
> Hi, Doug,
> you are right.
> And one "hns_x" ib device only has one command queue in hip08,
> other commands will be blocked when waiting on the command queue.
>>
>> I don't understand what you mean by "so the driver need not to use
>> stream mode to issue cmd".
> Sorry, my expression error.
> stream -> pipeline
> 
> And if you argee, after this patchset has been accepted we will send a
> following up patch :
>     In hns_roce_cmq_send function, replace
>         usleep_range(1000, 2000);
>     with the following statement:
>          udelay(1);
>     And if so, we can avoid using usleep_range function in spin_lock_bh
> spin region,
>     because it probally cause calltrace.

Ok, I'm fine with that.  I'll pull these in.



-- 
Doug Ledford <dledford@...hat.com>
    GPG Key ID: B826A3330E572FDD
    Key fingerprint = AE6B 1BDA 122B 23B4 265B  1274 B826 A333 0E57 2FDD



Download attachment "signature.asc" of type "application/pgp-signature" (885 bytes)

Powered by blists - more mailing lists