Re: [PATCH v4 5/8] dmaengine: dw-edma: Fix programming the source & dest addresses for ep

[Manivannan Sadhasivam <manivannan.sadhasivam@linaro.org>]

On Thu, Mar 10, 2022 at 10:37:59PM +0300, Serge Semin wrote:
> On Thu, Mar 10, 2022 at 10:50:14AM -0600, Zhi Li wrote:
> > On Thu, Mar 10, 2022 at 10:32 AM Serge Semin <fancer.lancer@gmail.com> wrote:
> > >
> > > On Wed, Mar 09, 2022 at 03:12:01PM -0600, Frank Li wrote:
> > > > From: Manivannan Sadhasivam <manivannan.sadhasivam@linaro.org>
> > > >
> > > > When eDMA is controlled by the Endpoint (EP), the current logic incorrectly
> > > > programs the source and destination addresses for read and write. Since the
> > > > Root complex and Endpoint uses the opposite channels for read/write, fix the
> > > > issue by finding out the read operation first and program the eDMA accordingly.
> > > >
> > > > Cc: stable@vger.kernel.org
> > > > Fixes: bd96f1b2f43a ("dmaengine: dw-edma: support local dma device transfer semantics")
> > > > Fixes: e63d79d1ffcd ("dmaengine: Add Synopsys eDMA IP core driver")
> > > > Signed-off-by: Manivannan Sadhasivam <manivannan.sadhasivam@linaro.org>
> > > > Signed-off-by: Frank Li <Frank.Li@nxp.com>
> > > > ---
> > > > No change between v1 to v4
> > > >
> > > >  drivers/dma/dw-edma/dw-edma-core.c | 32 +++++++++++++++++++++++++++++-
> > > >  1 file changed, 31 insertions(+), 1 deletion(-)
> > > >
> > > > diff --git a/drivers/dma/dw-edma/dw-edma-core.c b/drivers/dma/dw-edma/dw-edma-core.c
> > > > index 66dc650577919..507f08db1aad3 100644
> > > > --- a/drivers/dma/dw-edma/dw-edma-core.c
> > > > +++ b/drivers/dma/dw-edma/dw-edma-core.c
> > > > @@ -334,6 +334,7 @@ dw_edma_device_transfer(struct dw_edma_transfer *xfer)
> > > >       struct dw_edma_chunk *chunk;
> > > >       struct dw_edma_burst *burst;
> > > >       struct dw_edma_desc *desc;
> > > > +     bool read = false;
> > > >       u32 cnt = 0;
> > > >       int i;
> > > >
> > > > @@ -424,7 +425,36 @@ dw_edma_device_transfer(struct dw_edma_transfer *xfer)
> > > >               chunk->ll_region.sz += burst->sz;
> > > >               desc->alloc_sz += burst->sz;
> > > >
> > > > -             if (chan->dir == EDMA_DIR_WRITE) {
> > > > +             /****************************************************************
> > > > +              *
> > >
> > > > +              *        Root Complex                           Endpoint
> > > > +              * +-----------------------+             +----------------------+
> > > > +              * |                       |    TX CH    |                      |
> > > > +              * |                       |             |                      |
> > > > +              * |      DEV_TO_MEM       <-------------+     MEM_TO_DEV       |
> > > > +              * |                       |             |                      |
> > > > +              * |                       |             |                      |
> > > > +              * |      MEM_TO_DEV       +------------->     DEV_TO_MEM       |
> > > > +              * |                       |             |                      |
> > > > +              * |                       |    RX CH    |                      |
> > > > +              * +-----------------------+             +----------------------+
> > > > +              *
> > > > +              * If eDMA is controlled by the Root complex, TX channel
> > > > +              * (EDMA_DIR_WRITE) is used for memory read (DEV_TO_MEM) and RX
> > > > +              * channel (EDMA_DIR_READ) is used for memory write (MEM_TO_DEV).
> > > > +              *
> > > > +              * If eDMA is controlled by the endpoint, RX channel
> > > > +              * (EDMA_DIR_READ) is used for memory read (DEV_TO_MEM) and TX
> > > > +              * channel (EDMA_DIR_WRITE) is used for memory write (MEM_TO_DEV).
> > >
> > > Either I have some wrong notion about this issue, or something wrong
> > > with the explanation above and with this fix below.
> > >
> > > From my understanding of the possible DW eDMA IP-core setups the
> > > scatch above and the text below it are incorrect. Here is the way the
> > > DW eDMA can be used:
> > > 1) Embedded into the DW PCIe Host/EP controller. In this case
> > > CPU/Application Memory is the memory of the CPU attached to the
> > > host/EP controller, while the remote (link partner) memory is the PCIe
> > > bus memory. In this case MEM_TO_DEV operation is supposed to be
> > > performed by the Tx/Write channels, while the DEV_TO_MEM operation -
> > > by the Rx/Read channels.
> > >
> > > Note it's applicable for both Host and End-point case, when Linux is
> > > running on the CPU-side of the eDMA controller. So if it's DW PCIe
> > > end-point, then MEM_TO_DEV means copying data from the local CPU
> > > memory into the remote memory. In general the remote memory can be
> > > either some PCIe device on the bus or the Root Complex' CPU memory,
> > > each of which is some remote device anyway from the Local CPU
> > > perspective.
> > >
> > > 2) Embedded into the PCIe EP. This case is implemented in the
> > > drivers/dma/dw-edma/dw-edma-pcie.c driver. AFAICS from the commits log
> > > and from the driver code, that device is a Synopsys PCIe EndPoint IP
> > > prototype kit. It is a normal PCIe peripheral device with eDMA
> > > embedded, which CPU/Application interface is connected to some
> > > embedded SRAM while remote (link partner) interface is directed
> > > towards the PCIe bus. At the same time the device is setup and handled
> > > by the code running on a CPU connected to the PCIe Host controller.  I
> > > think that in order to preserve the normal DMA operations semantics we
> > > still need to consider the MEM_TO_DEV/DEV_TO_MEM operations from the
> > > host CPU perspective, since that's the side the DMA controller is
> > > supposed to be setup from.  In this MEM_TO_DEV is supposed to be used
> > > to copy data from the host CPU memory into the remote device memory.
> > > It means to allocate Rx/Read channel on the eDMA controller, so one
> > > would be read data from the Local CPU memory and copied it to the PCIe
> > > device SRAM. The logic of the DEV_TO_MEM direction would be just
> > > flipped. The eDMA PCIe device shall use Tx/Write channel to copy data
> > > from it's SRAM into the Host CPU memory.
> > >
> > > Please note as I understand the case 2) describes the Synopsys PCIe
> > > EndPoint IP prototype kit, which is based on some FPGA code. It's just
> > > a test setup with no real application, while the case 1) is a real setup
> > > available on our SoC and I guess on yours.
> > 
> 
> > I think yes. But Remote EP also is a one kind of usage module. Just no one
> > writes an EP functional driver for it yet.  Even pci-epf-test was just
> > a test function.
> > I previously sent vNTB patches to implement a virtual network between
> > RC and EP,
> > you can look if you have interest.
> 
> AFAIU the remote EP case is the same as 1) anyway. The remote EP is
> handled by its own CPU, which sets up the DW PCIe EP controller
> together with eDMA synthesized into the CPU' SoC. Am I right? While
> the case 2) doesn't have any CPU attached on the PCIe EP. It's just an
> FPGA with PCIe interface and eDMA IP-core installed. In that case all
> the setups are performed by the PCIe Host CPU. That's the root problem
> that causes having all the DEV_TO_MEM/MEM_TO_DEV complications.
> 
> So to speak I would suggest for at least to have the scatch fixed in
> accordance with the logic explained in my message.
> 
> > 
> > >
> > > So what I suggest in the framework of this patch is just to implement
> > > the case 1) only. While the case 2) as it's an artificial one can be
> > > manually handled by the DMA client drivers. BTW There aren't ones available
> > > in the kernel anyway. The only exception is an old-time attempt to get
> > > an eDMA IP test-driver mainlined into the kernel:
> > > https://patchwork.kernel.org/project/linux-pci/patch/cc195ac53839b318764c8f6502002cd6d933a923.1547230339.git.gustavo.pimentel@synopsys.com/
> > > But it was long time ago. So it's unlikely to be accepted at all.
> > >
> > > What do you think?
> > >
> > > -Sergey
> > >
> > > > +              *
> > > > +              ****************************************************************/
> > > > +
> > >
> > > > +             if ((dir == DMA_DEV_TO_MEM && chan->dir == EDMA_DIR_READ) ||
> > > > +                 (dir == DMA_DEV_TO_MEM && chan->dir == EDMA_DIR_WRITE))
> > > > +                     read = true;
> > >
> 
> > > Seeing the driver support only two directions DMA_DEV_TO_MEM/DMA_DEV_TO_MEM
> > > and EDMA_DIR_READ/EDMA_DIR_WRITE, this conditional statement seems
> > > redundant.
> 
> Am I getting a response on this comment? In accordance with that
> conditional statement having dir == DMA_DEV_TO_MEM means performing
> read operation. If dir equals DMA_MEM_TO_DEV then a write operation
> will be performed. The code path doesn't depend on the chan->dir
> value.
> 

As per the eDMA test application [1] submitted a while ago, my logic is correct:

+	if (thread->direction == DMA_DEV_TO_MEM) {
+		/* DMA_DEV_TO_MEM - WRITE - DMA_FROM_DEVICE */
+		dev_dbg(dev, "%s: DMA_DEV_TO_MEM - WRITE - DMA_FROM_DEVICE\n",
+			dma_chan_name(chan));
+		err = dma_map_sg(dev, sgt->sgl, sgt->nents, DMA_FROM_DEVICE);
+		if (!err)
+			goto err_descs;
+
+		sgt->nents = err;
+		/* Endpoint memory */
+		sconf.src_addr = dt_region->paddr;
+		/* CPU memory */
+		sconf.dst_addr = descs[0].paddr;
+	} else {
+		/* DMA_MEM_TO_DEV - READ - DMA_TO_DEVICE */
+		dev_dbg(dev, "%s: DMA_MEM_TO_DEV - READ - DMA_TO_DEVICE\n",
+			dma_chan_name(chan));
+		err = dma_map_sg(dev, sgt->sgl, sgt->nents, DMA_TO_DEVICE);
+		if (!err)
+			goto err_descs;
+
+		sgt->nents = err;
+		/* CPU memory */
+		sconf.src_addr = descs[0].paddr;
+		/* Endpoint memory */
+		sconf.dst_addr = dt_region->paddr;
+	}

[1] https://patchwork.kernel.org/project/linux-pci/patch/cc195ac53839b318764c8f6502002cd6d933a923.1547230339.git.gustavo.pimentel@synopsys.com/

Thanks,
Mani

> -Sergey
> 
> > >
> > > > +
> > > > +             /* Program the source and destination addresses for DMA read/write */
> > > > +             if (read) {
> > > >                       burst->sar = src_addr;
> > > >                       if (xfer->type == EDMA_XFER_CYCLIC) {
> > > >                               burst->dar = xfer->xfer.cyclic.paddr;
> > > > --
> > > > 2.24.0.rc1
> > > >