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

From: Guenter Roeck
Date: Thu Nov 15 2018 - 17:46:40 EST


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
>