Re: [PATCH 0/6] RISC-V: Add eMMC support for TH1520 boards

[Geert Uytterhoeven]

Hi Jisheng,

On Wed, Oct 4, 2023 at 4:18 PM Jisheng Zhang <jszhang@xxxxxxxxxx> wrote:
> On Wed, Oct 04, 2023 at 02:49:56PM +0100, Robin Murphy wrote:
> > On 04/10/2023 2:02 pm, Lad, Prabhakar wrote:
> > > + CC linux-mm and Robin Murphy
> > >
> > > On Wed, Oct 4, 2023 at 12:42 PM Jisheng Zhang <jszhang@xxxxxxxxxx> wrote:
 > > >
> > > > On Mon, Oct 02, 2023 at 09:37:44PM -0700, Drew Fustini wrote:
> > > > > On Fri, Sep 22, 2023 at 05:48:21PM -0500, Robert Nelson wrote:
> > > > > > On Fri, Sep 22, 2023 at 2:08 PM Robert Nelson <robertcnelson@xxxxxxxxx> wrote:
> > > > > > >
> > > > > > > On Thu, Sep 21, 2023 at 8:51 PM Drew Fustini <dfustini@xxxxxxxxxxxx> wrote:
> > > > > > > >
> > > > > > > > This series adds support for the eMMC on the BeagleV Ahead and the
> > > > > > > > Sipeed LicheePi 4A. This allows the kernel to boot with the rootfs on
> > > > > > > > eMMC.
> > > > > > > >
> > > > > > > > I tested on top of v6.6-rc2 with this config [1]. I was able to boot
> > > > > > > > both the Ahead [2] and LPi4a [3] from eMMC. The following prerequisites
> > > > > > > > are required:
> > > > > > > >
> > > > > > > >    [PATCH v2] riscv: dts: thead: set dma-noncoherent to soc bus [4]
> > > > > > > >
> > > > > > > > I pushed a branch [5] with this patch series and the above patch for
> > > > > > > > those that find a git branch easier to test.
> > > > > > > >
> > > > > > > > Please note that only the MMC controller connected to the eMMC device
> > > > > > > > is enabled in the device trees for these two boards. I did not yet
> > > > > > > > attempt to configure and use the microSD card slot. My preference is to
> > > > > > > > address that in a future patch series.
> > > > > > > >
> > > > > > > > References:
> > > > > > > > [1] https://gist.github.com/pdp7/5fbdcf2a65eb1abdd3a29d519c19cdd2
> > > > > > > > [2] https://gist.github.com/pdp7/91a801a5f8d1070c53509eda9800ad78
> > > > > > > > [3] https://gist.github.com/pdp7/1445c3c991e88fd69c60165cef65726a
> > > > > > > > [4] https://lore.kernel.org/linux-riscv/20230912072232.2455-1-jszhang@xxxxxxxxxx/
> > > > > > > > [5] https://github.com/pdp7/linux/tree/b4/th1520-mmc
> > > > > > >
> > > > > > > This patchset came out very nice!
> > > > > > >
> > > > > > > v6.6-rc2 with Last RFC v2:
> > > > > > >
> > > > > > > [    4.066630] mmc0: SDHCI controller on ffe7080000.mmc
> > > > > > > [ffe7080000.mmc] using PIO
> > > > > > >
> > > > > > > debian@BeagleV:~$ sudo hdparm -tT /dev/mmcblk0
> > > > > > >
> > > > > > > /dev/mmcblk0:
> > > > > > >   Timing cached reads:   1516 MB in  2.00 seconds = 758.09 MB/sec
> > > > > > >   Timing buffered disk reads:  84 MB in  3.01 seconds =  27.94 MB/sec
> > > > > > >
> > > > > > > vs v6.6-rc2 with this patchset:
> > > > > > >
> > > > > > >   [    4.096837] mmc0: SDHCI controller on ffe7080000.mmc
> > > > > > > [ffe7080000.mmc] using DMA
> > > > > > >
> > > > > > > debian@BeagleV:~$ sudo hdparm -tT /dev/mmcblk0
> > > > > > >
> > > > > > > /dev/mmcblk0:
> > > > > > >   Timing cached reads:   1580 MB in  2.00 seconds = 790.97 MB/sec
> > > > > > >   Timing buffered disk reads: 418 MB in  3.00 seconds = 139.11 MB/sec
> > > > > >
> > > > > > Drew pointed out on Slack, this was not quite right.. After more
> > > > > > digging by Drew, CONFIG_DMA_GLOBAL_POOL is causing a DMA limitation
> > > > > > with the multiplatform defconfig. so with,
> > > > > >
> > > > > > ./scripts/config --disable CONFIG_ARCH_R9A07G043
> > > > > >
> > > > > > (to remove CONFIG_DMA_GLOBAL_POOL)... another 2x in buffered reads..
> > > > > >
> > > > > > [    4.059242] mmc0: SDHCI controller on ffe7080000.mmc
> > > > > > [ffe7080000.mmc] using ADMA 64-bit
> > > > > >
> > > > > > debian@BeagleV:~$ sudo hdparm -tT /dev/mmcblk0
> > > > > >
> > > > > > /dev/mmcblk0:
> > > > > >   Timing cached reads:   1600 MB in  2.00 seconds = 800.93 MB/sec
> > > > > >   Timing buffered disk reads: 892 MB in  3.00 seconds = 297.06 MB/sec
> > > > >
> > > > > It seems CONFIG_DMA_GLOBAL_POOL=y causes ADMA buffer alloc to fail [1]:
> > > > >
> > > > >    mmc0: Unable to allocate ADMA buffers - falling back to standard DMA
> > > > >
> > > > > Prabhakar's AX45MP non-coherent DMA support [2] series introduced the
> > > > > selection of DMA_GLOBAL_POOL for ARCH_R9A07G043 and the riscv defconfig
> > > > > selects ARCH_R9A07G043.
> > > > >
> > > > > Patch 5 in the series [3] states that:
> > > > >
> > > > >    With DMA_GLOBAL_POOL enabled all DMA allocations happen from this
> > > > >    region and synchronization callbacks are implemented to synchronize
> > > > >    when doing DMA transactions.
> > > > >
> > > > > This example of a "shared-dma-pool" node was given:
> > > > >
> > > > >          pma_resv0@58000000 {
> > > > >              compatible = "shared-dma-pool";
> > > > >              reg = <0x0 0x58000000 0x0 0x08000000>;
> > > > >              no-map;
> > > > >              linux,dma-default;
> > > > >          };
> > > > >
> > > > > I've copied that to th1520-beaglev-ahead.dts. The address of 0x58000000
> > > > > has no significance on th1520, but the existence of shared-dma-pool
> > > > > seems to fix the problem. ADMA mode [4] is now working even though
> > > > > CONFIG_DMA_GLOBAL_POOL=y.
> > > >
> > > > + Christoph, Lad
> > > >
> > > > IMHO, this is not TH1520 specific but a generic issue.
> > > >
> > > > I believe commit 484861e09f3e ("soc: renesas: Kconfig: Select the
> > > > required configs for RZ/Five SoC") can cause regression on all
> > > > non-dma-coherent riscv platforms with generic defconfig. This is
> > > > a common issue. The logic here is: generic riscv defconfig selects
> > > > ARCH_R9A07G043 which selects DMA_GLOBAL_POOL, which assumes all
> > > > non-dma-coherent riscv platforms have a dma global pool, this assumption
> > > > seems not correct. And I believe DMA_GLOBAL_POOL should not be
> > > > selected by ARCH_SOCFAMILIY, instead, only ARCH under some specific
> > > > conditions can select it globaly, for example NOMMU ARM and so on.
> > > >
> > > > Since this is a regression, what's proper fix? any suggestion is
> > > > appreciated.
> >
> > I think the answer is to not select DMA_GLOBAL_POOL, since that is only
> > designed for nommu cases where non-cacheable memory lives in a fixed place
> > in the physical address map, and regular kernel pages can't be remapped. As
> > far as I'm aware, RISCV_DMA_NONCOHERENT is the thing you want, such that
> > DMA_DIRECT_REMAP can dynamically provide non-cacheable coherent buffers for
> > non-hardware-coherent devices.
>
> Thank Robin!
> AFAIK, ARCH_R9A07G043 needs the dma global pool to handle its CMO. So
> it looks like ARCH_R9A07G043 can't be enabled in riscv generic
> defconfig. And we also need a special solution to prevent random config
> from selecting ARCH_R9A07G043 by chance for other platforms

There will be a similar issue with e.g. Starlight, as ERRATA_STARFIVE_JH7100
(not yet upstream, but in esmil/visionfive) also selects DMA_GLOBAL_POOL.

Gr{oetje,eeting}s,

                        Geert

-- 
Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- geert@xxxxxxxxxxxxxx

In personal conversations with technical people, I call myself a hacker. But
when I'm talking to journalists I just say "programmer" or something like that.
                                -- Linus Torvalds