Re: [PATCH] nvme: utilize two queue maps, one for reads and one for writes

[Guenter Roeck <linux@roeck-us.net>]

On Thu, Nov 15, 2018 at 03:12:48PM -0700, Jens Axboe wrote:
> On 11/15/18 3:06 PM, Guenter Roeck wrote:
> > On Thu, Nov 15, 2018 at 12:43:40PM -0700, Jens Axboe wrote:
> >> On 11/15/18 12:40 PM, Jens Axboe wrote:
> >>> On 11/15/18 12:38 PM, Guenter Roeck wrote:
> >>>> On Thu, Nov 15, 2018 at 12:29:04PM -0700, Jens Axboe wrote:
> >>>>> On 11/15/18 12:11 PM, Guenter Roeck wrote:
> >>>>>> On Wed, Nov 14, 2018 at 10:12:44AM -0700, Jens Axboe wrote:
> >>>>>>>
> >>>>>>> I think the below patch should fix it.
> >>>>>>>
> >>>>>>
> >>>>>> I spoke too early. sparc64, next-20181115:
> >>>>>>
> >>>>>> [   14.204370] nvme nvme0: pci function 0000:02:00.0
> >>>>>> [   14.249956] nvme nvme0: Removing after probe failure status: -5
> >>>>>> [   14.263496] ------------[ cut here ]------------
> >>>>>> [   14.263913] WARNING: CPU: 0 PID: 15 at kernel/irq/manage.c:1597 __free_irq+0xa4/0x320
> >>>>>> [   14.264265] Trying to free already-free IRQ 9
> >>>>>> [   14.264519] Modules linked in:
> >>>>>> [   14.264961] CPU: 0 PID: 15 Comm: kworker/u2:1 Not tainted 4.20.0-rc2-next-20181115 #1
> >>>>>> [   14.265555] Workqueue: nvme-reset-wq nvme_reset_work
> >>>>>> [   14.265899] Call Trace:
> >>>>>> [   14.266118]  [000000000046944c] __warn+0xcc/0x100
> >>>>>> [   14.266375]  [00000000004694b0] warn_slowpath_fmt+0x30/0x40
> >>>>>> [   14.266635]  [00000000004d4ce4] __free_irq+0xa4/0x320
> >>>>>> [   14.266867]  [00000000004d4ff8] free_irq+0x38/0x80
> >>>>>> [   14.267092]  [00000000007b1874] pci_free_irq+0x14/0x40
> >>>>>> [   14.267327]  [00000000008a5444] nvme_dev_disable+0xe4/0x520
> >>>>>> [   14.267576]  [00000000008a69b8] nvme_reset_work+0x138/0x1c60
> >>>>>> [   14.267827]  [0000000000488dd0] process_one_work+0x230/0x6e0
> >>>>>> [   14.268079]  [00000000004894f4] worker_thread+0x274/0x520
> >>>>>> [   14.268321]  [0000000000490624] kthread+0xe4/0x120
> >>>>>> [   14.268544]  [00000000004060c4] ret_from_fork+0x1c/0x2c
> >>>>>> [   14.268825]  [0000000000000000]           (null)
> >>>>>> [   14.269089] irq event stamp: 32796
> >>>>>> [   14.269350] hardirqs last  enabled at (32795): [<0000000000b624a4>] _raw_spin_unlock_irqrestore+0x24/0x80
> >>>>>> [   14.269757] hardirqs last disabled at (32796): [<0000000000b622f4>] _raw_spin_lock_irqsave+0x14/0x60
> >>>>>> [   14.270566] softirqs last  enabled at (32780): [<0000000000b64c18>] __do_softirq+0x238/0x520
> >>>>>> [   14.271206] softirqs last disabled at (32729): [<000000000042ceec>] do_softirq_own_stack+0x2c/0x40
> >>>>>> [   14.272288] ---[ end trace cb79ccd2a0a03f3c ]---
> >>>>>>
> >>>>>> Looks like an error during probe followed by an error cleanup problem.
> >>>>>
> >>>>> Did it previous probe fine? Or is the new thing just the fact that
> >>>>> we spew a warning on trying to free a non-existing vector?
> >>>>>
> >>>> This works fine in mainline, if that is your question.
> >>>
> >>> Yeah, as soon as I sent the other email I realized that. Let me send
> >>> you a quick patch.
> >>
> >> How's this?
> >>
> >>
> >> diff --git a/drivers/nvme/host/pci.c b/drivers/nvme/host/pci.c
> >> index ffbab5b01df4..fd73bfd2d1be 100644
> >> --- a/drivers/nvme/host/pci.c
> >> +++ b/drivers/nvme/host/pci.c
> >> @@ -2088,15 +2088,11 @@ static int nvme_setup_irqs(struct nvme_dev *dev, int nr_io_queues)
> >>  			affd.nr_sets = 1;
> >>  
> >>  		/*
> >> -		 * Need IRQs for read+write queues, and one for the admin queue.
> >> -		 * If we can't get more than one vector, we have to share the
> >> -		 * admin queue and IO queue vector. For that case, don't add
> >> -		 * an extra vector for the admin queue, or we'll continue
> >> -		 * asking for 2 and get -ENOSPC in return.
> >> +		 * If we got a failure and we're down to asking for just
> >> +		 * 1 + 1 queues, just ask for a single vector. We'll share
> >> +		 * that between the single IO queue and the admin queue.
> >>  		 */
> >> -		if (result == -ENOSPC && nr_io_queues == 1)
> >> -			nr_io_queues = 1;
> >> -		else
> >> +		if (!(result < 0 && nr_io_queues == 1))
> >>  			nr_io_queues = irq_sets[0] + irq_sets[1] + 1;
> >>  
> > 
> > Unfortunately, the code doesn't even get here because the call of
> > pci_alloc_irq_vectors_affinity in the first iteration fails with
> > -EINVAL, which results in an immediate return with -EIO.
> 
> Oh yeah... How about this then?
> 
Yes, this one works (at least on sparc64). Do I need to test
on other architectures as well ?

Thanks,
Guenter

> 
> diff --git a/drivers/nvme/host/pci.c b/drivers/nvme/host/pci.c
> index ffbab5b01df4..4d161daa9c3a 100644
> --- a/drivers/nvme/host/pci.c
> +++ b/drivers/nvme/host/pci.c
> @@ -2088,15 +2088,11 @@ static int nvme_setup_irqs(struct nvme_dev *dev, int nr_io_queues)
>  			affd.nr_sets = 1;
>  
>  		/*
> -		 * Need IRQs for read+write queues, and one for the admin queue.
> -		 * If we can't get more than one vector, we have to share the
> -		 * admin queue and IO queue vector. For that case, don't add
> -		 * an extra vector for the admin queue, or we'll continue
> -		 * asking for 2 and get -ENOSPC in return.
> +		 * If we got a failure and we're down to asking for just
> +		 * 1 + 1 queues, just ask for a single vector. We'll share
> +		 * that between the single IO queue and the admin queue.
>  		 */
> -		if (result == -ENOSPC && nr_io_queues == 1)
> -			nr_io_queues = 1;
> -		else
> +		if (!(result < 0 && nr_io_queues == 1))
>  			nr_io_queues = irq_sets[0] + irq_sets[1] + 1;
>  
>  		result = pci_alloc_irq_vectors_affinity(pdev, nr_io_queues,
> @@ -2111,6 +2107,9 @@ static int nvme_setup_irqs(struct nvme_dev *dev, int nr_io_queues)
>  			if (!nr_io_queues)
>  				return result;
>  			continue;
> +		} else if (result == -EINVAL) {

Add an explanation, maybe ?

> +			nr_io_queues = 1;
> +			continue;
>  		} else if (result <= 0)
>  			return -EIO;
>  		break;
> 
> -- 
> Jens Axboe
>