Re: [PATCH v3 1/5] PCI: dwc: Refactor dw_pcie_edma_find_chip() API

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

On Tue, Feb 27, 2024 at 12:00:41AM +0300, Serge Semin wrote:
> On Mon, Feb 26, 2024 at 08:57:57PM +0530, Manivannan Sadhasivam wrote:
> > On Mon, Feb 26, 2024 at 03:45:16PM +0300, Serge Semin wrote:
> > > Hi Manivannan
> > > 
> > > On Mon, Feb 26, 2024 at 05:07:26PM +0530, Manivannan Sadhasivam wrote:
> > > > In order to add support for Hyper DMA (HDMA), let's refactor the existing
> > > > dw_pcie_edma_find_chip() API by moving the common code to separate
> > > > functions.
> > > > 
> > > > No functional change.
> > > > 
> > > > Suggested-by: Serge Semin <fancer.lancer@gmail.com>
> > > > Signed-off-by: Manivannan Sadhasivam <manivannan.sadhasivam@linaro.org>
> > > > ---
> > > >  drivers/pci/controller/dwc/pcie-designware.c | 52 +++++++++++++++++++++-------
> > > >  1 file changed, 39 insertions(+), 13 deletions(-)
> > > > 
> > > > diff --git a/drivers/pci/controller/dwc/pcie-designware.c b/drivers/pci/controller/dwc/pcie-designware.c
> > > > index 250cf7f40b85..193fcd86cf93 100644
> > > > --- a/drivers/pci/controller/dwc/pcie-designware.c
> > > > +++ b/drivers/pci/controller/dwc/pcie-designware.c
> > > > @@ -880,7 +880,17 @@ static struct dw_edma_plat_ops dw_pcie_edma_ops = {
> > > >  	.irq_vector = dw_pcie_edma_irq_vector,
> > > >  };
> > > >  
> > > > -static int dw_pcie_edma_find_chip(struct dw_pcie *pci)
> > > > +static void dw_pcie_edma_init_data(struct dw_pcie *pci)
> > > > +{
> > > > +	pci->edma.dev = pci->dev;
> > > > +
> > > > +	if (!pci->edma.ops)
> > > > +		pci->edma.ops = &dw_pcie_edma_ops;
> > > > +
> > > > +	pci->edma.flags |= DW_EDMA_CHIP_LOCAL;
> > > > +}
> > > > +
> > > > +static int dw_pcie_edma_find_mf(struct dw_pcie *pci)
> > > >  {
> > > >  	u32 val;
> > > >  
> > > > @@ -900,24 +910,27 @@ static int dw_pcie_edma_find_chip(struct dw_pcie *pci)
> > > >  	else
> > > >  		val = dw_pcie_readl_dbi(pci, PCIE_DMA_VIEWPORT_BASE + PCIE_DMA_CTRL);
> > > > 
> > > 
> > > > -	if (val == 0xFFFFFFFF && pci->edma.reg_base) {
> > > > -		pci->edma.mf = EDMA_MF_EDMA_UNROLL;
> > > > -
> > > > -		val = dw_pcie_readl_dma(pci, PCIE_DMA_CTRL);
> > > > -	} else if (val != 0xFFFFFFFF) {
> > > > -		pci->edma.mf = EDMA_MF_EDMA_LEGACY;
> > > > +	/* Set default mapping format here and update it below if needed */
> > > > +	pci->edma.mf = EDMA_MF_EDMA_LEGACY;
> > > >  
> > > > +	if (val == 0xFFFFFFFF && pci->edma.reg_base)
> > > > +		pci->edma.mf = EDMA_MF_EDMA_UNROLL;
> > > > +	else if (val != 0xFFFFFFFF)
> > > >  		pci->edma.reg_base = pci->dbi_base + PCIE_DMA_VIEWPORT_BASE;
> > > > -	} else {
> > > > +	else
> > > >  		return -ENODEV;
> > > > -	}
> > > 
> > > Sorry for not posting my opinion about this earlier, but IMO v2 code
> > > was more correct than this one. This version makes the code being not
> > > linear as it was in v2, thus harder to comprehend:
> > > 
> > > 1. Setting up a default value and then overriding it or not makes the
> > > reader to keep in mind the initialized value which is harder than to
> > > just read what is done in the respective branch.
> > > 
> > 
> > No, I disagree. Whether we set the default value or not, EDMA_MF_EDMA_LEGACY is
> > indeed the default mapping format (this is one of the reasons why the enums
> > should start from 1 instead of 0). So initializing it to legacy is not changing
> > anything, rather making it explicit.
> > 
> > > 2. Splitting up the case clause with respective inits and the mapping
> > > format setting up also makes it harder to comprehend what's going on.
> > > In the legacy case the reg-base address and the mapping format init are
> > > split up while they should have been done simultaneously only if (val
> > > != 0xFFFFFFFF).
> > > 
> > 
> > Well again, this doesn't matter since the default mapping format is legacy. But
> > somewhat agree that the two clauses are setting different fields, but even if
> > the legacy mapping format is set inside the second clause, it still differs from
> > the first one since we are not setting reg_base.
> > 
> > > 3. The most of the current devices has the unrolled mapping (available
> > > since v4.9 IP-core), thus having the mf field pre-initialized produces
> > > a redundant store operation for the most of the modern devices.
> > > 
> > 
> > Ok, this one I agree. We could avoid the extra assignment.
> > 
> > > 4. Getting rid from the curly braces isn't something what should be
> > > avoided at any cost and doesn't give any optimization really. It
> > > doesn't cause having less C-lines of the source code and doesn't
> > > improve the code readability.
> > > 
> > 
> > Yeah, there is no benefit other than a simple view of the code. But for point
> > (3), I agree to roll back to v2 version.
> > 
> > > So to speak, I'd suggest to get back the v2 implementation here.
> > > 
> > > >  
> > > > -	pci->edma.dev = pci->dev;
> > > > +	return 0;
> > > > +}
> > > >  
> > > > -	if (!pci->edma.ops)
> > > > -		pci->edma.ops = &dw_pcie_edma_ops;
> > > > +static int dw_pcie_edma_find_channels(struct dw_pcie *pci)
> > > > +{
> > > > +	u32 val;
> > > >  
> > > > -	pci->edma.flags |= DW_EDMA_CHIP_LOCAL;
> > > 
> > > > +	if (pci->edma.mf == EDMA_MF_EDMA_LEGACY)
> > > > +		val = dw_pcie_readl_dbi(pci, PCIE_DMA_VIEWPORT_BASE + PCIE_DMA_CTRL);
> > > > +	else
> > > > +		val = dw_pcie_readl_dma(pci, PCIE_DMA_CTRL);
> > > 
> > > Just dw_pcie_readl_dma(pci, PCIE_DMA_CTRL)
> > > 
> > 
> > 'val' is uninitialized. Why should the assignment be skipped?
> 
> The entire
> 
> +	if (pci->edma.mf == EDMA_MF_EDMA_LEGACY)
> +		val = dw_pcie_readl_dbi(pci, PCIE_DMA_VIEWPORT_BASE + PCIE_DMA_CTRL);
> +	else
> +		val = dw_pcie_readl_dma(pci, PCIE_DMA_CTRL);
> 
> can be replaced with a single line
> 
> +	val = dw_pcie_readl_dma(pci, PCIE_DMA_CTRL);
> 
> since in the legacy case (reg_base = PCIE_DMA_VIEWPORT_BASE) and the
> reg_base has been initialized by now.
> 

Ah okay, got it!

- Mani

-- 
மணிவண்ணன் சதாசிவம்