Re: [PATCH RFC 1/2] dmaengine: Introduce scheduled DMA framework

[Maxime Ripard <maxime.ripard@free-electrons.com>]

[-- Attachment #1: Type: text/plain, Size: 10791 bytes --]

Hi Ludo,

On Mon, Mar 23, 2015 at 02:25:13PM +0100, Ludovic Desroches wrote:
> On Sat, Mar 21, 2015 at 12:42:06PM -0700, Maxime Ripard wrote:
> > This framework aims at easing the development of dmaengine drivers by providing
> > generic implementations of the functions usually required by dmaengine, while
> > abstracting away most of the logic required.
> 
> For sure it will ease writing new dmaengine drivers! Moreover, adopting
> this framework will convert the driver to use virtual channels isn't it?

Yes, it relies on virt-dma

> > It is very relevant for controllers that have more requests than channels,
> > where you need to have some scheduling that is usually very bug prone, and
> > shouldn't be implemented in each and every driver.
> > 
> > This introduces a new set of callbacks for drivers to implement the device
> > specific behaviour. These new sets of callbacks aims at providing both how to
> > handle channel related operations (start the transfer of a new descriptor,
> > pause, resume, etc.) and the LLI related operations (iterator and various
> > accessors).
> > 
> > So far, the only transfer types supported are memcpy and slave transfers, but
> > eventually should support new transfer types as new drivers get converted.
> > 
> > Signed-off-by: Maxime Ripard <maxime.ripard@free-electrons.com>
> > ---
> >  drivers/dma/Kconfig         |   4 +
> >  drivers/dma/Makefile        |   1 +
> >  drivers/dma/scheduled-dma.c | 571 ++++++++++++++++++++++++++++++++++++++++++++
> >  drivers/dma/scheduled-dma.h | 140 +++++++++++
> >  4 files changed, 716 insertions(+)
> >  create mode 100644 drivers/dma/scheduled-dma.c
> >  create mode 100644 drivers/dma/scheduled-dma.h
> > 
> > diff --git a/drivers/dma/Kconfig b/drivers/dma/Kconfig
> > index a874b6ec6650..032bf5fcd58b 100644
> > --- a/drivers/dma/Kconfig
> > +++ b/drivers/dma/Kconfig
> > @@ -406,6 +406,7 @@ config DMA_SUN6I
> >  	depends on RESET_CONTROLLER
> >  	select DMA_ENGINE
> >  	select DMA_VIRTUAL_CHANNELS
> > +	select DMA_SCHEDULED
> >  	help
> >  	  Support for the DMA engine first found in Allwinner A31 SoCs.
> >  
> > @@ -431,6 +432,9 @@ config DMA_ENGINE
> >  config DMA_VIRTUAL_CHANNELS
> >  	tristate
> >  
> > +config DMA_SCHEDULED
> > +	bool
> > +
> 
> I think, it should select DMA_VIRTUAL_CHANNELS

Yep, indeed.

> >  config DMA_ACPI
> >  	def_bool y
> >  	depends on ACPI
> > diff --git a/drivers/dma/Makefile b/drivers/dma/Makefile
> > index f915f61ec574..1db31814c749 100644
> > --- a/drivers/dma/Makefile
> > +++ b/drivers/dma/Makefile
> > @@ -5,6 +5,7 @@ obj-$(CONFIG_DMA_ENGINE) += dmaengine.o
> >  obj-$(CONFIG_DMA_VIRTUAL_CHANNELS) += virt-dma.o
> >  obj-$(CONFIG_DMA_ACPI) += acpi-dma.o
> >  obj-$(CONFIG_DMA_OF) += of-dma.o
> > +obj-$(CONFIG_DMA_SCHEDULED) += scheduled-dma.o
> >  
> >  obj-$(CONFIG_INTEL_MID_DMAC) += intel_mid_dma.o
> >  obj-$(CONFIG_DMATEST) += dmatest.o
> > diff --git a/drivers/dma/scheduled-dma.c b/drivers/dma/scheduled-dma.c
> > new file mode 100644
> > index 000000000000..482d04f2ccbc
> > --- /dev/null
> > +++ b/drivers/dma/scheduled-dma.c
> > @@ -0,0 +1,571 @@
> > +/*
> > + * Copyright (C) 2015 Maxime Ripard
> > + * Maxime Ripard <maxime.ripard@free-electrons.com>
> > + *
> > + * This program is free software; you can redistribute it and/or modify
> > + * it under the terms of the GNU General Public License as published by
> > + * the Free Software Foundation; either version 2 of the License, or
> > + * (at your option) any later version.
> > + */
> > +
> > +#include <linux/dmaengine.h>
> > +#include <linux/dmapool.h>
> > +#include <linux/list.h>
> > +#include <linux/slab.h>
> > +
> > +#include "scheduled-dma.h"
> > +#include "virt-dma.h"
> > +
> > +static struct sdma_request *sdma_pop_queued_transfer(struct sdma *sdma,
> > +						     struct sdma_channel *schan)
> > +{
> > +	struct sdma_request *sreq = NULL;
> > +	unsigned long flags;
> > +
> > +	spin_lock_irqsave(&sdma->lock, flags);
> > +
> > +	/* No requests are awaiting an available channel */
> > +	if (list_empty(&sdma->pend_reqs))
> > +		goto out;
> > +
> > +	/*
> > +	 * If we don't have any validate_request callback, any request
> > +	 * can be dispatched to any channel.
> > +	 *
> > +	 * Remove the first entry and return it.
> > +	 */
> > +	if (!sdma->ops->validate_request) {
> > +		sreq = list_first_entry(&sdma->pend_reqs,
> > +					struct sdma_request, node);
> > +		list_del_init(&sreq->node);
> > +		goto out;
> > +	}
> > +
> > +	list_for_each_entry(sreq, &sdma->pend_reqs, node) {
> > +		/*
> > +		 * Ask the driver to validate that the request can
> > +		 * happen on the channel.
> > +		 */
> > +		if (sdma->ops->validate_request(schan, sreq)) {
> > +			list_del_init(&sreq->node);
> > +			goto out;
> > +		}
> > +
> > +		/* Otherwise, just keep looping */
> > +	}
> > +
> > +out:
> > +	spin_unlock_irqrestore(&sdma->lock, flags);
> > +
> > +	return sreq;
> > +}
> > +
> > +/*
> > + * Besoin d'une fonction pour pusher un descriptor vers un pchan
> > + *
> > + * Flow normal:
> > + *   - Election d'un pchan (Framework)
> > + *   - Push d'un descripteur vers le pchan (Driver)
> > + *   - idle....
> > + *   - interrupt (driver)
> > + *   - Transfert terminé, notification vers le framework (driver)
> > + *   - Nouveau transfert dans la queue?
> > + *     + Si oui, on est reparti
> > + *     + Si non, on sort de l'interrupt, le pchan est libéré
> > + */
> > +
> > +struct sdma_desc *sdma_report(struct sdma *sdma,
> > +			      struct sdma_channel *schan,
> > +			      enum sdma_report_status status)
> > +{
> > +	struct sdma_desc *sdesc = NULL;
> > +	struct virt_dma_desc *vdesc;
> > +	struct sdma_request *sreq;
> > +
> > +	switch (status) {
> > +	case SDMA_REPORT_TRANSFER:
> > +		/*
> > +		 * We're done with the current transfer that was in this
> > +		 * physical channel.
> > +		 */
> > +		vchan_cookie_complete(&schan->desc->vdesc);
> > +
> > +		/*
> > +		 * Now, try to see if there's any queued transfer
> > +		 * awaiting an available channel.
> > +		 *
> > +		 * If not, just bail out, and mark the pchan as
> > +		 * available.
> > +		 *
> > +		 * If there's such a transfer, validate that the
> > +		 * driver can handle it, and ask it to do the
> > +		 * transfer.
> > +		 */
> > +		sreq = sdma_pop_queued_transfer(sdma, schan);
> > +		if (!sreq) {
> > +			list_add_tail(&schan->node, &sdma->avail_chans);
> > +			return NULL;
> > +		}
> > +
> > +		/* Mark the request as assigned to a particular channel */
> > +		sreq->chan = schan;
> > +
> > +		/* Retrieve the next transfer descriptor */
> > +		vdesc = vchan_next_desc(&sreq->vchan);
> > +		schan->desc = sdesc = to_sdma_desc(&vdesc->tx);
> > +
> > +		break;
> > +
> > +	default:
> > +		break;
> > +	}
> > +
> > +	return sdesc;
> > +}
> > +EXPORT_SYMBOL_GPL(sdma_report);
> > +
> > +static enum dma_status sdma_tx_status(struct dma_chan *chan,
> > +				      dma_cookie_t cookie,
> > +				      struct dma_tx_state *state)
> > +{
> > +	struct sdma_request *sreq = to_sdma_request(chan);
> > +	struct sdma *sdma = to_sdma(chan->device);
> > +	struct sdma_channel *schan = sreq->chan;
> > +	struct virt_dma_desc *vd;
> > +	struct sdma_desc *desc;
> > +	enum dma_status ret;
> > +	unsigned long flags;
> > +	size_t bytes = 0;
> > +	void *lli;
> > +
> > +	spin_lock_irqsave(&sreq->vchan.lock, flags);
> > +
> > +	ret = dma_cookie_status(chan, cookie, state);
> > +	if (ret == DMA_COMPLETE)
> > +		goto out;
> > +
> > +	vd = vchan_find_desc(&sreq->vchan, cookie);
> > +	desc = to_sdma_desc(&vd->tx);
> > +
> > +	if (vd) {
> > +		lli = desc->v_lli;
> > +		while (true) {
> > +			bytes += sdma->ops->lli_size(lli);
> > +
> > +			if (!sdma->ops->lli_has_next(lli))
> > +				break;
> > +
> > +			lli = sdma->ops->lli_next(lli);
> > +		}
> > +	} else if (chan) {
> > +		bytes = sdma->ops->channel_residue(schan);
> > +	}
> > +
> > +	dma_set_residue(state, bytes);
> > +
> > +out:
> > +	spin_unlock_irqrestore(&sreq->vchan.lock, flags);
> > +
> > +	return ret;
> > +};
> > +
> > +static int sdma_config(struct dma_chan *chan,
> > +		       struct dma_slave_config *config)
> > +{
> > +	struct sdma_request *sreq = to_sdma_request(chan);
> > +	unsigned long flags;
> > +
> > +	spin_lock_irqsave(&sreq->vchan.lock, flags);
> > +	memcpy(&sreq->cfg, config, sizeof(*config));
> > +	spin_unlock_irqrestore(&sreq->vchan.lock, flags);
> > +
> > +	return 0;
> > +}
> > +
> > +static int sdma_pause(struct dma_chan *chan)
> > +{
> > +	struct sdma_request *sreq = to_sdma_request(chan);
> > +	struct sdma_channel *schan = sreq->chan;
> > +	struct sdma *sdma = to_sdma(chan->device);
> > +	unsigned long flags;
> > +
> > +	spin_lock_irqsave(&sreq->vchan.lock, flags);
> > +
> > +	/*
> > +	 * If the request is currently scheduled on a channel, just
> > +	 * pause the channel.
> > +	 *
> > +	 * If not, remove the request from the pending list.
> > +	 */
> > +	if (schan) {
> > +		sdma->ops->channel_pause(schan);
> > +	} else {
> > +		spin_lock(&sdma->lock);
> > +		list_del_init(&sreq->node);
> > +		spin_unlock(&sdma->lock);
> > +	}
> > +
> > +	spin_unlock_irqrestore(&sreq->vchan.lock, flags);
> > +
> > +	return 0;
> > +}
> > +
> > +static int sdma_resume(struct dma_chan *chan)
> > +{
> > +	struct sdma_request *sreq = to_sdma_request(chan);
> > +	struct sdma_channel *schan = sreq->chan;
> > +	struct sdma *sdma = to_sdma(chan->device);
> > +	unsigned long flags;
> > +
> > +	spin_lock_irqsave(&sreq->vchan.lock, flags);
> > +
> > +	/*
> > +	 * If the request is currently scheduled on a channel, just
> > +	 * resume the channel.
> > +	 *
> > +	 * If not, add the request from the pending list.
> > +	 */
> 
> Is such a case possible?

Yes, it's possible, if you have more requests than channels, and all
channels are currently busy with other transfers, we can end up in a
case where we have a requests where issue_pending has been called, but
is not currently scheduled on any channel, for an unknown amount of
time.

In such a case, you might very well end up with clients calling
pause/resume on these requests.

> The request should be already in the pending list, isn't it?

Not really, it would have been removed by sdma_pause.

> By the way, I may have missed something but where do you add a
> request to the pending list? I have seen it only in the resume case.

It should be in issue_pending and pause. If not, it's a bug :)

Thanks!
Maxime

-- 
Maxime Ripard, Free Electrons
Embedded Linux, Kernel and Android engineering
http://free-electrons.com

[-- Attachment #2: Digital signature --]
[-- Type: application/pgp-signature, Size: 819 bytes --]