Re: [PATCH 2/2] dmaengine: milbeaut-hdmac: Add HDMAC driver for Milbeaut platforms

From: Vinod Koul
Date: Fri Aug 16 2019 - 00:44:24 EST


On 15-08-19, 21:25, Jassi Brar wrote:
> On Mon, Jun 24, 2019 at 1:47 AM Vinod Koul <vkoul@xxxxxxxxxx> wrote:
> >
> > On 12-06-19, 19:52, jassisinghbrar@xxxxxxxxx wrote:
> >
> > > +#include <linux/bits.h>
> > > +#include <linux/clk.h>
> > > +#include <linux/dma-mapping.h>
> > > +#include <linux/dmaengine.h>
> > > +#include <linux/interrupt.h>
> > > +#include <linux/iopoll.h>
> > > +#include <linux/list.h>
> > > +#include <linux/module.h>
> > > +#include <linux/of.h>
> > > +#include <linux/of_dma.h>
> >
> > Do we need both, IIRC of_dma.h does include of.h!
> >
> OK
>
> > > +/* mc->vc.lock must be held by caller */
> > > +static void milbeaut_chan_start(struct milbeaut_hdmac_chan *mc,
> > > + struct milbeaut_hdmac_desc *md)
> > > +{
> > > + struct scatterlist *sg;
> > > + u32 cb, ca, src_addr, dest_addr, len;
> > ^^
> > double space
> >
> OK
>
> > > +static irqreturn_t milbeaut_hdmac_interrupt(int irq, void *dev_id)
> > > +{
> > > + struct milbeaut_hdmac_chan *mc = dev_id;
> > > + struct milbeaut_hdmac_desc *md;
> > > + irqreturn_t ret = IRQ_HANDLED;
> > > + u32 val;
> > > +
> > > + spin_lock(&mc->vc.lock);
> > > +
> > > + /* Ack and Disable irqs */
> > > + val = readl_relaxed(mc->reg_ch_base + MLB_HDMAC_DMACB);
> > > + val &= ~(FIELD_PREP(MLB_HDMAC_SS, 0x7));
> > ^^^^
> > Magic ..?
> >
> OK, will define a macro for 7
>
> > > +static int milbeaut_hdmac_chan_pause(struct dma_chan *chan)
> > > +{
> > > + struct virt_dma_chan *vc = to_virt_chan(chan);
> > > + struct milbeaut_hdmac_chan *mc = to_milbeaut_hdmac_chan(vc);
> > > + u32 val;
> > > +
> > > + spin_lock(&mc->vc.lock);
> > > + val = readl_relaxed(mc->reg_ch_base + MLB_HDMAC_DMACA);
> > > + val |= MLB_HDMAC_PB;
> > > + writel_relaxed(val, mc->reg_ch_base + MLB_HDMAC_DMACA);
> >
> > We really should have an updatel() and friends in kernel, feel free to
> > add in your driver though!
> >
> I'll pass on that for now.
>
> > > +static int milbeaut_hdmac_chan_init(struct platform_device *pdev,
> > > + struct milbeaut_hdmac_device *mdev,
> > > + int chan_id)
> > > +{
> > > + struct device *dev = &pdev->dev;
> > > + struct milbeaut_hdmac_chan *mc = &mdev->channels[chan_id];
> > > + char *irq_name;
> > > + int irq, ret;
> > > +
> > > + irq = platform_get_irq(pdev, chan_id);
> > > + if (irq < 0) {
> > > + dev_err(&pdev->dev, "failed to get IRQ number for ch%d\n",
> > > + chan_id);
> > > + return irq;
> > > + }
> > > +
> > > + irq_name = devm_kasprintf(dev, GFP_KERNEL, "milbeaut-hdmac-%d",
> > > + chan_id);
> > > + if (!irq_name)
> > > + return -ENOMEM;
> > > +
> > > + ret = devm_request_irq(dev, irq, milbeaut_hdmac_interrupt,
> > > + IRQF_SHARED, irq_name, mc);
> >
> > I tend to dislike using devm_request_irq(), we have no control over when
> > the irq is freed and what is a spirious irq is running while we are
> > unrolling, so IMHO it make sense to free up and ensure all tasklets are
> > quiesced when remove returns
> >
> If the code is written clean and tight we need not be so paranoid.
>
> > > + if (ret)
> > > + return ret;
> > > +
> > > + mc->mdev = mdev;
> > > + mc->reg_ch_base = mdev->reg_base + MLB_HDMAC_CH_STRIDE * (chan_id + 1);
> > > + mc->vc.desc_free = milbeaut_hdmac_desc_free;
> > > + vchan_init(&mc->vc, &mdev->ddev);
> >
> > who kills the vc->task?
> >
> vchan_synchronize() called from milbeaut_hdmac_synchronize()

But that can be skipped by called, from driver pov we need to ensure
that it is killed

> > > +static int milbeaut_hdmac_remove(struct platform_device *pdev)
> > > +{
> > > + struct milbeaut_hdmac_device *mdev = platform_get_drvdata(pdev);
> > > + struct dma_chan *chan;
> > > + int ret;
> > > +
> > > + /*
> > > + * Before reaching here, almost all descriptors have been freed by the
> > > + * ->device_free_chan_resources() hook. However, each channel might
> > > + * be still holding one descriptor that was on-flight at that moment.
> > > + * Terminate it to make sure this hardware is no longer running. Then,
> > > + * free the channel resources once again to avoid memory leak.
> > > + */
> > > + list_for_each_entry(chan, &mdev->ddev.channels, device_node) {
> > > + ret = dmaengine_terminate_sync(chan);
> > > + if (ret)
> > > + return ret;
> > > + milbeaut_hdmac_free_chan_resources(chan);
> > > + }
> > > +
> > > + of_dma_controller_free(pdev->dev.of_node);
> > > + dma_async_device_unregister(&mdev->ddev);
> > > + clk_disable_unprepare(mdev->clk);
> >
> > And as suspected we have active tasklets and irq at this time :(
> >
> Not sure how is that....

Since you use devm variants, and driver tasklet is not killed, irq can
be fired (spurious as well) and can run irq and schedule a tasklet. How
do you prevent that?

Thanks
--
~Vinod