Re: [PATCH 04/24] thunderbolt: Add MSI-X support

[Mika Westerberg]

On Sun, May 21, 2017 at 07:51:09PM +0200, Andreas Noever wrote:
> On Thu, May 18, 2017 at 4:38 PM, Mika Westerberg
> <mika.westerberg@xxxxxxxxxxxxxxx> wrote:
> > Intel Thunderbolt controllers support up to 16 MSI-X vectors. Using
> Is that true for all generations? If so can we remove the legacy path?

Yes, I think it has been the case from the beginning. I have here one
Mac Mini with Light Ridge and even it has working MSI-X.

Yehezkel, Michael, Amir, do you know if that's the case for all other
controller as well?

> > MSI-X is preferred over MSI or legacy interrupt and may bring additional
> > performance because there is no need to check the status registers which
> > interrupt was triggered.
> >
> > While there we convert comments in structs tb_ring and tb_nhi to follow
> > kernel-doc format more closely.
> >
> > This code is based on the work done by Amir Levy and Michael Jamet.
> >
> > Signed-off-by: Michael Jamet <michael.jamet@xxxxxxxxx>
> > Signed-off-by: Mika Westerberg <mika.westerberg@xxxxxxxxxxxxxxx>
> > Reviewed-by: Yehezkel Bernat <yehezkel.bernat@xxxxxxxxx>
> > ---
> >  drivers/thunderbolt/ctl.c      |   4 +-
> >  drivers/thunderbolt/nhi.c      | 159 ++++++++++++++++++++++++++++++++++++-----
> >  drivers/thunderbolt/nhi.h      |  56 +++++++++++----
> >  drivers/thunderbolt/nhi_regs.h |   9 +++
> >  4 files changed, 196 insertions(+), 32 deletions(-)
> >
> > diff --git a/drivers/thunderbolt/ctl.c b/drivers/thunderbolt/ctl.c
> > index 1031d97407a8..889a32dd21e7 100644
> > --- a/drivers/thunderbolt/ctl.c
> > +++ b/drivers/thunderbolt/ctl.c
> > @@ -488,11 +488,11 @@ struct tb_ctl *tb_ctl_alloc(struct tb_nhi *nhi, hotplug_cb cb, void *cb_data)
> >         if (!ctl->frame_pool)
> >                 goto err;
> >
> > -       ctl->tx = ring_alloc_tx(nhi, 0, 10);
> > +       ctl->tx = ring_alloc_tx(nhi, 0, 10, RING_FLAG_NO_SUSPEND);
> >         if (!ctl->tx)
> >                 goto err;
> >
> > -       ctl->rx = ring_alloc_rx(nhi, 0, 10);
> > +       ctl->rx = ring_alloc_rx(nhi, 0, 10, RING_FLAG_NO_SUSPEND);
> >         if (!ctl->rx)
> >                 goto err;
> >
> > diff --git a/drivers/thunderbolt/nhi.c b/drivers/thunderbolt/nhi.c
> > index a8c20413dbda..17f3b1bdb7da 100644
> > --- a/drivers/thunderbolt/nhi.c
> > +++ b/drivers/thunderbolt/nhi.c
> > @@ -21,6 +21,12 @@
> >
> >  #define RING_TYPE(ring) ((ring)->is_tx ? "TX ring" : "RX ring")
> >
> > +/*
> > + * Minimal number of vectors when we use MSI-X. Two for control channel
> > + * Rx/Tx and the rest four are for cross domain DMA paths.
> > + */
> > +#define MSIX_MIN_VECS          6
> > +#define MSIX_MAX_VECS          16
> >
> >  static int ring_interrupt_index(struct tb_ring *ring)
> >  {
> > @@ -239,8 +245,82 @@ int __ring_enqueue(struct tb_ring *ring, struct ring_frame *frame)
> >         return ret;
> >  }
> >
> > +static irqreturn_t ring_msix(int irq, void *data)
> > +{
> > +       struct tb_ring *ring = data;
> > +
> > +       schedule_work(&ring->work);
> > +       return IRQ_HANDLED;
> > +}
> > +
> > +static void ring_map_unmap_msix(struct tb_ring *ring, bool map)
> > +{
> This function does the same as ring_interrupt_active just for msix,
> right? Can we rename one (or both) to express that?
> For example rename ring_interrput_active -> ring_map_unmap_msi. Or
> just roll this one into ring_interrupt_active and do the right thing
> internally.

Yes, I think we can do that.

> > +       u32 step, shift, ivr, misc;
> > +       int index;
> > +
> > +       if (ring->irq <= 0)
> > +               return;
> > +
> > +       if (ring->is_tx)
> > +               index = ring->hop;
> > +       else
> > +               index = ring->hop + ring->nhi->hop_count;
> > +
> > +       /*
> > +        * Ask the hardware to clear interrupt status bits automatically
> > +        * since we already know which interrupt was triggered.
> > +        */
> > +       misc = ioread32(ring->nhi->iobase + REG_DMA_MISC);
> > +       if (!(misc & REG_DMA_MISC_INT_AUTO_CLEAR)) {
> > +               misc |= REG_DMA_MISC_INT_AUTO_CLEAR;
> > +               iowrite32(misc, ring->nhi->iobase + REG_DMA_MISC);
> > +       }
> > +
> > +       step = index / REG_INT_VEC_ALLOC_REGS * REG_INT_VEC_ALLOC_BITS;
> > +       shift = index % REG_INT_VEC_ALLOC_REGS * REG_INT_VEC_ALLOC_BITS;
> > +       ivr = ioread32(ring->nhi->iobase + REG_INT_VEC_ALLOC_BASE + step);
> > +       ivr &= ~(REG_INT_VEC_ALLOC_MASK << shift);
> > +       if (map)
> > +               ivr |= ring->vector << shift;
> > +       iowrite32(ivr, ring->nhi->iobase + REG_INT_VEC_ALLOC_BASE + step);
> > +}
> > +
> > +static int ring_request_msix(struct tb_ring *ring, bool no_suspend)
> > +{
> > +       struct tb_nhi *nhi = ring->nhi;
> > +       unsigned long irqflags;
> > +       int ret;
> > +
> > +       if (!nhi->pdev->msix_enabled)
> > +               return 0;
> > +
> > +       ret = ida_simple_get(&nhi->msix_ida, 0, MSIX_MAX_VECS, GFP_KERNEL);
> > +       if (ret < 0)
> > +               return ret;
> > +
> > +       ring->vector = ret;
> > +
> > +       ring->irq = pci_irq_vector(ring->nhi->pdev, ring->vector);
> > +       if (ring->irq < 0)
> > +               return ring->irq;
> > +
> > +       irqflags = no_suspend ? IRQF_NO_SUSPEND : 0;
> > +       return request_irq(ring->irq, ring_msix, irqflags, "thunderbolt", ring);
> > +}
> > +
> > +static void ring_release_msix(struct tb_ring *ring)
> > +{
> > +       if (ring->irq <= 0)
> > +               return;
> > +
> > +       free_irq(ring->irq, ring);
> > +       ida_simple_remove(&ring->nhi->msix_ida, ring->vector);
> > +       ring->vector = 0;
> > +       ring->irq = 0;
> > +}
> > +
> >  static struct tb_ring *ring_alloc(struct tb_nhi *nhi, u32 hop, int size,
> > -                                 bool transmit)
> > +                                 bool transmit, unsigned int flags)
> >  {
> >         struct tb_ring *ring = NULL;
> >         dev_info(&nhi->pdev->dev, "allocating %s ring %d of size %d\n",
> > @@ -271,9 +351,14 @@ static struct tb_ring *ring_alloc(struct tb_nhi *nhi, u32 hop, int size,
> >         ring->hop = hop;
> >         ring->is_tx = transmit;
> >         ring->size = size;
> > +       ring->flags = flags;
> >         ring->head = 0;
> >         ring->tail = 0;
> >         ring->running = false;
> > +
> > +       if (ring_request_msix(ring, flags & RING_FLAG_NO_SUSPEND))
> > +               goto err;
> > +
> >         ring->descriptors = dma_alloc_coherent(&ring->nhi->pdev->dev,
> >                         size * sizeof(*ring->descriptors),
> >                         &ring->descriptors_dma, GFP_KERNEL | __GFP_ZERO);
> > @@ -295,14 +380,16 @@ static struct tb_ring *ring_alloc(struct tb_nhi *nhi, u32 hop, int size,
> >         return NULL;
> >  }
> >
> > -struct tb_ring *ring_alloc_tx(struct tb_nhi *nhi, int hop, int size)
> > +struct tb_ring *ring_alloc_tx(struct tb_nhi *nhi, int hop, int size,
> > +                             unsigned int flags)
> >  {
> > -       return ring_alloc(nhi, hop, size, true);
> > +       return ring_alloc(nhi, hop, size, true, flags);
> >  }
> >
> > -struct tb_ring *ring_alloc_rx(struct tb_nhi *nhi, int hop, int size)
> > +struct tb_ring *ring_alloc_rx(struct tb_nhi *nhi, int hop, int size,
> > +                             unsigned int flags)
> >  {
> > -       return ring_alloc(nhi, hop, size, false);
> > +       return ring_alloc(nhi, hop, size, false, flags);
> >  }
> >
> >  /**
> > @@ -334,6 +421,7 @@ void ring_start(struct tb_ring *ring)
> >                 ring_iowrite32options(ring,
> >                                       RING_FLAG_ENABLE | RING_FLAG_RAW, 0);
> >         }
> > +       ring_map_unmap_msix(ring, true);
> >         ring_interrupt_active(ring, true);
> >         ring->running = true;
> >  err:
> > @@ -366,6 +454,7 @@ void ring_stop(struct tb_ring *ring)
> >                 goto err;
> >         }
> >         ring_interrupt_active(ring, false);
> > +       ring_map_unmap_msix(ring, false);
> >
> >         ring_iowrite32options(ring, 0, 0);
> >         ring_iowrite64desc(ring, 0, 0);
> > @@ -413,6 +502,8 @@ void ring_free(struct tb_ring *ring)
> >                          RING_TYPE(ring), ring->hop);
> >         }
> >
> > +       ring_release_msix(ring);
> > +
> >         dma_free_coherent(&ring->nhi->pdev->dev,
> >                           ring->size * sizeof(*ring->descriptors),
> >                           ring->descriptors, ring->descriptors_dma);
> There is a comment a couple lines below that states that ring->work is
> only scheduled by nhi_interrupt_work and ring_stop. Please add
> ring_msix to the list.

OK.

> > @@ -528,9 +619,51 @@ static void nhi_shutdown(struct tb_nhi *nhi)
> >          * We have to release the irq before calling flush_work. Otherwise an
> >          * already executing IRQ handler could call schedule_work again.
> >          */
> > -       devm_free_irq(&nhi->pdev->dev, nhi->pdev->irq, nhi);
> > +       if (!nhi->pdev->msix_enabled)
> > +               devm_free_irq(&nhi->pdev->dev, nhi->pdev->irq, nhi);
> >         flush_work(&nhi->interrupt_work);
> >         mutex_destroy(&nhi->lock);
> > +       ida_destroy(&nhi->msix_ida);
> > +}
> > +
> > +static int nhi_init_msi(struct tb_nhi *nhi)
> > +{
> > +       struct pci_dev *pdev = nhi->pdev;
> > +       int res, irq, nvec;
> > +
> > +       /* In case someone left them on. */
> > +       nhi_disable_interrupts(nhi);
> > +
> > +       ida_init(&nhi->msix_ida);
> > +
> > +       /*
> > +        * The NHI has 16 MSI-X vectors or a single MSI. We first try to
> > +        * get all MSI-X vectors and if we succeed, each ring will have
> > +        * one MSI-X. If for some reason that does not work out, we
> > +        * fallback to a single MSI.
> > +        */
> > +       nvec = pci_alloc_irq_vectors(pdev, MSIX_MIN_VECS, MSIX_MAX_VECS,
> > +                                    PCI_IRQ_MSIX);
> > +       if (nvec < 0) {
> > +               nvec = pci_alloc_irq_vectors(pdev, 1, 1, PCI_IRQ_MSI);
> > +               if (nvec < 0)
> > +                       return nvec;
> > +
> > +               INIT_WORK(&nhi->interrupt_work, nhi_interrupt_work);
> As far as I can see INIT_WORK is only called here in the fallback
> case. But flush_work is still called unconditionally in nhi_shutdown.

Good point. I should make it conditional as well (or INIT_WORK()
unconditional).