Message-ID: <ZgW6KB73Wh1X6911@matsya>
Date: Fri, 29 Mar 2024 00:12:48 +0530
From: Vinod Koul <vkoul@...nel.org>
To: Niklas Cassel <cassel@...nel.org>
Cc: Manivannan Sadhasivam <manivannan.sadhasivam@...aro.org>,
	Lorenzo Pieralisi <lpieralisi@...nel.org>,
	Krzysztof Wilczyński <kw@...ux.com>,
	Rob Herring <robh@...nel.org>, Bjorn Helgaas <bhelgaas@...gle.com>,
	Kishon Vijay Abraham I <kishon@...nel.org>,
	Thierry Reding <thierry.reding@...il.com>,
	Jonathan Hunter <jonathanh@...dia.com>,
	Jingoo Han <jingoohan1@...il.com>,
	Gustavo Pimentel <gustavo.pimentel@...opsys.com>,
	linux-pci@...r.kernel.org, linux-arm-msm@...r.kernel.org,
	linux-kernel@...r.kernel.org, mhi@...ts.linux.dev,
	linux-tegra@...r.kernel.org
Subject: Re: [PATCH 05/11] PCI: epf-{mhi/test}: Move DMA initialization to
 EPC init callback

On 27-03-24, 12:39, Niklas Cassel wrote:
> +CC Vinod
> 
> Hello Vinod,
> 
> I didn't know the answer, so I chose the "call a friend option" ;)
> I hope that you can help me out :)

Anytime :-)

> 
> 
> If you take a look at drivers/pci/endpoint/functions/pci-epf-test.c
> https://github.com/torvalds/linux/blob/v6.9-rc1/drivers/pci/endpoint/functions/pci-epf-test.c#L448-L471
> 
> You can see that the driver always does pci_epc_map_addr(),
> then it will either use:
> DMA API, e.g. dma_map_single() etc.
> or
> memcpy_fromio()/memcpy_toio()
> 
> based on flag FLAG_USE_DMA.
> 
> This flag is set via ioctl, so if we run:
> /usr/bin/pcitest -d
> the flag will be set, without the -d parameter the flag won't be set.
> 
> 
> If you look at how the DMA channel is requested:
> https://github.com/torvalds/linux/blob/v6.9-rc1/drivers/pci/endpoint/functions/pci-epf-test.c#L224-L258
> 
> If will try to get a private DMA channel, if that fails,
> it will use the "dummy memcpy" DMA channel.
> 
> If the FLAG_USE_DMA is set, the transfers itself will use:
> https://github.com/torvalds/linux/blob/v6.9-rc1/drivers/pci/endpoint/functions/pci-epf-test.c#L139-L155
> either dmaengine_prep_slave_single() or dmaengine_prep_dma_memcpy(),
> depending on if we are using "dummy memcpy" or not.
> 
> 
> 
> If you take e.g. the DWC PCIe EP controller, it can have an embedded DMA
> controller on the PCIe controller, and we will try to detect it when
> initializing the PCIe EP controller using dw_pcie_edma_detect():
> https://github.com/torvalds/linux/blob/v6.9-rc1/drivers/pci/controller/dwc/pcie-designware-ep.c#L759
> 
> For the PCIe EP controller that I am using, which have eDMA built-in,
> I noticed that if I do not enable the eDMA driver (# CONFIG_DW_EDMA is not
> set), I noticed that I can still run:
> /usr/bin/pcitest -d
> 
> Which will use the "dummy memcpy" DMA channel.
> Yes, the performance is poor, but it still works, so it appears that the
> fallback code is working properly.
> 
> 
> If I enable the eDMA driver (CONFIG_DW_EDMA=y),
> I can run:
> /usr/bin/pcitest -d
> 
> And the performance is good.
> 
> 
> So my question is:
> Is the "dummy memcpy" DMA channel always available?

That depends on the system, you may or maynot have such a system where
you have a generic memcpy dma controller which can provide you with
these channels

> 
> Because if it is, I think we could drop the path in the pci-epf-test.c
> driver which uses memcpy_fromio()/memcpy_toio() instead of DMA API.
> (Since just having a single path to do I/O in the driver would simplify
> the driver IMO.)
> 
> I assume that the "dummy memcpy" DMA channel just uses memcpy_fromio() and
> memcpy_toio() under the hood, so I assume that using the memcpy_fromio()/
> memcpy_toio/() is equivalent to using DMA API + dmaengine_prep_dma_memcpy().
> 
> Although it would be nice if we didn't need to have the two separate paths
> in pci_epf_test_data_transfer() (dmaengine_prep_slave_single() vs
> dmaengine_prep_dma_memcpy()) to support the "dummy memcpy" channel.
> But I guess that is not possible...

Based on my reading you might have this mechanism:
- eDMA provides dmaengine_prep_slave_single() which transfers data from
  mem to pci ep device, so fasted
- dmaengine_prep_dma_memcpy: This will copy the data but treat it as
  memory. I dont pci internals to figure out how both can work... So
  cant really make out why it is slowed
- memcpy_xxx that is IO mem functions, so ofc they will be slowest

I think the code is decent from fallback pov... chooses fastest path if
available on a system

> 
> 
> I hope that you can bring some clarity Vinod.
> (Please read my replies to Mani below before you compose your email,
> as it does provide more insight to this mess.)
> 
> Mani, I tried to reply to you inline below, with my limited understanding
> of how dmaengine works.
> 
> 
> On Wed, Mar 27, 2024 at 11:48:19AM +0530, Manivannan Sadhasivam wrote:
> > > So we still want to test:
> > > -DMA API using the eDMA
> > > -DMA API using the "dummy" memcpy dma-channel.
> > > 
> > 
> > IMO, the test driver should just test one form of data transfer. Either CPU
> > memcpy (using iATU or something similar) or DMA. But I think the motive behind
> > using DMA memcpy is that to support platforms that do not pass DMA slave
> > channels in devicetree.
> > 
> > It is applicable to test driver but not to MHI driver since all DMA supported
> > MHI platforms will pass the DMA slave channels in devicetree.
> 
> I don't understand how device tree is relevant here, e.g. qcom-ep.c
> specifies pcie_ep->pci.edma.nr_irqs = 1;
> https://github.com/torvalds/linux/blob/v6.9-rc1/drivers/pci/controller/dwc/pcie-qcom-ep.c#L818
> which is sufficient for you to be able to probe/detect eDMA successfully,
> no need for anything in device tree at all.
> 
> 
> > 
> > > However, it seems like both pci-epf-mhi.c and pci-epf-test.c
> > > do either:
> > > -Use DMA API
> > > or
> > > -Use memcpy_fromio()/memcpy_toio() instead of DMA API
> > > 
> > > 
> > > To me, it seems like we should always be able to use
> > > DMA API (using either a eDMA or "dummy" memcpy).
> > > 
> > 
> > No, there are platforms that don't support DMA at all. Like Qcom SDX55, so we
> > still need to do CPU memcpy.
> 
> I assume that you mean the the PCIe controller used in SDX55 does not
> have the eDMA on the PCIe controller, so dw_pcie_edma_detect() will
> fail to detect any eDMA. That is fine no?
> 
> I assume that this SoC will still able to use the "dummy" memcpy dma-channel?
> 
> 
> > 
> > > I don't really see the need to have the path that does:
> > > memcpy_fromio()/memcpy_toio().
> > > 
> > > I know that for DWC, when using memcpy (and this also
> > > memcpy via DMA API), we need to map the address using
> > > iATU first.
> > > 
> > > But that could probably be done using another flag,
> > > perhaps rename that flag FLAG_USE_DMA to NEEDS_MAP or
> > > something.
> > > (Such that we can change these drivers to only have a
> > > code path that uses DMA API.)
> > > (...and making sure that inheriting the DMA mask does
> > > not affect the DMA mask for DMA_MEMCPY.)
> 
> I was wrong here, pci-epf-test always calls pci_epc_map_addr()
> regardless if FLAG_USE_DMA is set or not.
> 
> (Even though this should be unnecessary when using the eDMA.)
> 
> However, if we look at pci-epf-mhi.c we can see that it does
> NOT call pci_epc_map_addr() when using DMA API + dmaengine.
> 
> Is it really safe to avoid pci_epc_map_addr() in all EPC controllers?
> I assume that it should be safe for all "real" DMA channels.
> We can see that it is not safe when using DMA API + "dummy" memcpy
> dma-channel. (That is why I was asking if we need a NEEDS_MAP, or
> MAP_NOT_NEEDED flag.)
> 
> 
> > > 
> > > But perhaps I am missing something... and DMA_MEMCPY is
> > > not always available?
> 
> Right now pci-epf-test driver has three ways:
> -DMA API + dmaengine dmaengine_prep_slave_single()
> -DMA API + dmaengine dmaengine_prep_dma_memcpy()
> -memcpy_toio()/memcpy_fromio().
> 
> pci-epf-mhi.c driver has two ways:
> -DMA API + dmaengine dmaengine_prep_slave_single()
> -memcpy_toio()/memcpy_fromio().
> 
> 
> pci-epf-test.c:
> -Always calls pci_epc_map_addr() when using DMA API.
> 
> pci-epf-mhi.c:
> -Never calls pci_epc_map_addr() when using DMA API.
> 
> 
> I honestly don't see any point of having three paths
> for pci-epf-test. Ideally I would want one, max two.
> 
> If you think that:
> -DMA API + dmaengine dmaengine_prep_slave_single()
> +
> -memcpy_toio()/memcpy_fromio().
> 
> is more logical than:
> -DMA API + dmaengine dmaengine_prep_slave_single()
> +
> -DMA API + dmaengine dmaengine_prep_dma_memcpy()
> 
> Then I think we should rip out the:
> -DMA API + dmaengine dmaengine_prep_dma_memcpy()
> it serves no purpose... if you don't have a "real" DMA channel,
> just run without the -d flag.
> 
> Or, if you argue that the dmaengine_prep_dma_memcpy() is there
> to test the DMA API code (which I can't say that it does, since
> it doesn't use the exact same code path as a "real" DMA channel, see:
> https://github.com/torvalds/linux/blob/v6.9-rc1/drivers/pci/endpoint/functions/pci-epf-test.c#L139-L155
> so this argument is questionable).
> 
> Put it under a --use_dummy_dma, and return failure by default
> if no "real" DMA channel is found.
> 
> 
> But even so, that would not address the pci-epf-test and
> pci-mhi-test inconsistency WRT pci_epc_map_addr().
> 
> I think if we rip out:
> -DMA API + dmaengine dmaengine_prep_dma_memcpy()
> we could also move the pci_epc_map_addr() so that it is
> only used for the memcpy_toio()/memcpy_fromio() path.
> 
> (Or if we add a --use_dummy_dma, we can move the pci_epc_map_addr() to
> that path, and remove it from the dmaengine_prep_slave_single() path.)
> 
> 
> Kind regards,
> Niklas

-- 
~Vinod

Powered by blists - more mailing lists