On Thu, Mar 24, 2022 at 07:00:04PM +0530, Manivannan Sadhasivam wrote:
> On Thu, Mar 24, 2022 at 04:48:12AM +0300, Serge Semin wrote:
> > The dma_slave_config.direction field usage in the DW eDMA driver has been
> > introduced in the commit bd96f1b2f43a ("dmaengine: dw-edma: support local
> > dma device transfer semantics"). Mainly the change introduced there was
> > correct (indeed DEV_TO_MEM means using RD-channel and MEM_TO_DEV -
> > WR-channel for the case of having eDMA accessed locally from
> > CPU/Application side), but providing an additional
> > MEM_TO_MEM/DEV_TO_DEV-based semantics was quite redundant if not to say
> > potentially harmful (when it comes to removing the denoted field). First
> > of all since the dma_slave_config.direction field has been marked as
> > obsolete (see [1] and the structure dc [2]) and will be discarded in
> > future, using it especially in a non-standard way is discouraged. Secondly
> > in accordance with the commit denoted above the default
> > dw_edma_device_transfer() semantics has been changed despite what it's
> > message said. So claiming that the method was left backward compatible was
> > wrong.
> >
> > Anyway let's fix the problems denoted above and simplify the
> > dw_edma_device_transfer() method by dropping the parsing of the
> > DMA-channel direction field. Instead of having that implicit
> > dma_slave_config.direction field semantic we can use the recently added
> > DW_EDMA_CHIP_LOCAL flag to distinguish between the local and remote DW
> > eDMA setups thus preserving both cases support. In addition to that an
> > ASCII-figure has been added to clarify the complication out.
> >
> > [1] Documentation/driver-api/dmaengine/provider.rst
> > [2] include/linux/dmaengine.h: dma_slave_config.direction
> >
> > Co-developed-by: Manivannan Sadhasivam <manivannan.sadhasivam@xxxxxxxxxx>
> > Signed-off-by: Manivannan Sadhasivam <manivannan.sadhasivam@xxxxxxxxxx>
> > Signed-off-by: Serge Semin <Sergey.Semin@xxxxxxxxxxxxxxxxxxxx>
> >
> > ---
> >
> > In accordance with agreement with Frank and Manivannan this patch is
> > supposed to be moved to the series:
> > Link: https://lore.kernel.org/dmaengine/20220310192457.3090-1-Frank.Li@xxxxxxx/
> > in place of the patch:
> > [PATCH v5 6/9] dmaengine: dw-edma: Don't rely on the deprecated "direction" member
> > Link: https://lore.kernel.org/dmaengine/20220310192457.3090-7-Frank.Li@xxxxxxx/
> > ---
> > drivers/dma/dw-edma/dw-edma-core.c | 49 +++++++++++++++++++++---------
> > 1 file changed, 34 insertions(+), 15 deletions(-)
> >
> > diff --git a/drivers/dma/dw-edma/dw-edma-core.c b/drivers/dma/dw-edma/dw-edma-core.c
> > index 5be8a5944714..e9e32ed74aa9 100644
> > --- a/drivers/dma/dw-edma/dw-edma-core.c
> > +++ b/drivers/dma/dw-edma/dw-edma-core.c
> > @@ -339,21 +339,40 @@ dw_edma_device_transfer(struct dw_edma_transfer *xfer)
> > if (!chan->configured)
> > return NULL;
> >
> > - switch (chan->config.direction) {
> > - case DMA_DEV_TO_MEM: /* local DMA */
> > - if (dir == DMA_DEV_TO_MEM && chan->dir == EDMA_DIR_READ)
> > - break;
> > - return NULL;
> > - case DMA_MEM_TO_DEV: /* local DMA */
> > - if (dir == DMA_MEM_TO_DEV && chan->dir == EDMA_DIR_WRITE)
> > - break;
> > - return NULL;
> > - default: /* remote DMA */
> > - if (dir == DMA_MEM_TO_DEV && chan->dir == EDMA_DIR_READ)
> > - break;
> > - if (dir == DMA_DEV_TO_MEM && chan->dir == EDMA_DIR_WRITE)
> > - break;
> > - return NULL;
> > + /*
> > + * Local Root Port/End-point Remote End-point
> > + * +-----------------------+ PCIe bus +----------------------+
> > + * | | +-+ | |
> > + * | DEV_TO_MEM Rx Ch <----+ +---+ Tx Ch DEV_TO_MEM |
> > + * | | | | | |
> > + * | MEM_TO_DEV Tx Ch +----+ +---> Rx Ch MEM_TO_DEV |
> > + * | | +-+ | |
> > + * +-----------------------+ +----------------------+
> > + *
> > + * 1. Normal logic:
> > + * If eDMA is embedded into the DW PCIe RP/EP and controlled from the
> > + * CPU/Application side, the Rx channel (EDMA_DIR_READ) will be used
> > + * for the device read operations (DEV_TO_MEM) and the Tx channel
> > + * (EDMA_DIR_WRITE) - for the write operations (MEM_TO_DEV).
> > + *
> > + * 2. Inverted logic:
> > + * If eDMA is embedded into a Remote PCIe EP and is controlled by the
> > + * MWr/MRd TLPs sent from the CPU's PCIe host controller, the Tx
> > + * channel (EDMA_DIR_WRITE) will be used for the device read operations
> > + * (DEV_TO_MEM) and the Rx channel (EDMA_DIR_READ) - for the write
> > + * operations (MEM_TO_DEV).
> > + *
> > + * It is the client driver responsibility to choose a proper channel
> > + * for the DMA transfers.
> > + */
>
> I think it'd be good to document this using some form in "enum dw_edma_dir"
> declaration.
Well, I'd rather leave the comment here. The dw_edma_dir enumeration
only defines the actual device channel direction, while the actual
semantics is determined by the PCIe device residence/configuration
(local + host/end-point, remote + end-point) and is reflected in the
data transfer function - dw_edma_device_transfer(). So as I see it the
comment is more suitable to be left here in the place where the
semantic is actually implemented.
-Sergey
>
> Thanks,
> Mani
>
> > + if (chan->dw->chip->flags & DW_EDMA_CHIP_LOCAL) {
> > + if ((chan->dir == EDMA_DIR_READ && dir != DMA_DEV_TO_MEM) ||
> > + (chan->dir == EDMA_DIR_WRITE && dir != DMA_MEM_TO_DEV))
> > + return NULL;
> > + } else {
> > + if ((chan->dir == EDMA_DIR_WRITE && dir != DMA_DEV_TO_MEM) ||
> > + (chan->dir == EDMA_DIR_READ && dir != DMA_MEM_TO_DEV))
> > + return NULL;
> > }
> >
> > if (xfer->type == EDMA_XFER_CYCLIC) {
> > --
> > 2.35.1
> >
