Re: [PATCH] iommu: ipmmu-vmsa: Add DMA controllers to devices_allowlist

[Koichiro Den]

On Mon, Nov 03, 2025 at 05:56:12PM +0900, Ryo Takakura wrote:
> Hi, Geert!
> 
> I'm helping out Den-san a bit on the issue.
> 
> On Wed, 29 Oct 2025 17:08:07 +0100, Geert Uytterhoeven wrote:
> >On Mon, 27 Oct 2025 at 02:26, Koichiro Den <den@xxxxxxxxxxxxx> wrote:
> >> Add Renesas DMA controller devices to the devices_allowlist to enable
> >> their use with the IPMMU. This allows DMA channels to operate through
> >> the IOMMU when enabled.
> >>
> >> Signed-off-by: Koichiro Den <den@xxxxxxxxxxxxx>
> >
> >Thanks for your patch!
> >
> >> --- a/drivers/iommu/ipmmu-vmsa.c
> >> +++ b/drivers/iommu/ipmmu-vmsa.c
> >> @@ -741,7 +741,9 @@ static const char * const devices_allowlist[] = {
> >>         "ee100000.mmc",
> >>         "ee120000.mmc",
> >>         "ee140000.mmc",
> >> -       "ee160000.mmc"
> >> +       "ee160000.mmc",
> >> +       "e7350000.dma-controller",
> >> +       "e7351000.dma-controller"
> >
> >While your change looks correct to me, it causes DMA mapping failures on
> >Gray Hawk Single and Sparrow Hawk when IPMMU support is enabled
> >(on renesas-drivers-2025-10-28-v6.18-rc3 with renesas_defconfig
> >+ CONFIG_IPMMU_VMSA=y):
> >
> >    rcar-dmac e7351000.dma-controller: chan1: failed to map 1@0x00000000e656000c
> >    sh-sci e6560000.serial: Failed preparing Tx DMA descriptor
> >
> >0xe656000c = HSCIF2 Transmit FIFO data register.
> 
> Thank you for sharing.
> I was able to reproduce the issue on Gray Hawk Single(r8a779h2).
> 
> >With "#define DEBUG" added to drivers/dma/sh/rcar-dmac.c:
> >
> >    rcar-dmac e7351000.dma-controller: chan1: failed to map 1@0x00000000e656000c
> >    sh-sci e6560000.serial: Failed preparing Tx DMA descriptor
> >    rcar-dmac e7351000.dma-controller: chan2: map 1@0x00000000e6560014
> >to 0x00000000fff82014 dir: DMA_TO_DEVICE
> >    rcar-dmac e7351000.dma-controller: chan2: chunk
> >00000000e5110c20/000000005e0ede90 sgl 0@000000000d8c5440, 256/256
> >0x00000000fff82014 -> 0x00000000ffed6000
> >    rcar-dmac e7351000.dma-controller: chan2: submit #2@000000005e0ede90
> >    rcar-dmac e7351000.dma-controller: chan2: chunk
> >0000000025f2f66c/00000000e5f0dd15 sgl 0@0000000064f1067f, 256/256
> >0x00000000fff82014 -> 0x00000000ffed6100
> >    rcar-dmac e7351000.dma-controller: chan2: submit #3@00000000e5f0dd15
> >    rcar-dmac e7351000.dma-controller: chan2: queue chunk
> >00000000e5110c20: 256@0x00000000fff82014 -> 0x00000000ffed6000
> >
> >0xe6560014 = HSCIF2 Receive FIFO data register
> 
> Does IPMMU support handling FIFO registers?
> (Forgive me the question... I'm still trying to learn how it works.)
> 
> If not, with now dma-controller backed with IPMMU, maybe what
> we need is to map FIFO register directly like it used to do?
> 
> >Comparing to the CONFIG_IPMMU_VMSA=n case:
> >
> >    rcar-dmac e7351000.dma-controller: chan1: map 1@0x00000000e656000c
> >to 0x00000000e656000c dir: DMA_FROM_DEVICE
> >    rcar-dmac e7351000.dma-controller: chan1: chunk
> >0000000096396eb4/00000000f35357b6 sgl 0@0000000013546bf6, 74/74
> >0x0000000489ab5000 -> 0x00000000e656000c
> >    rcar-dmac e7351000.dma-controller: chan1: submit #2@00000000f35357b6
> >    rcar-dmac e7351000.dma-controller: chan1: queue chunk
> >0000000096396eb4: 74@0x0000000489ab5000 -> 0x00000000e656000c
> >
> >This confuses me even more: why no DMA_TO_DEVICE mapping in the
> >latter case?
> 
> I believe the DEVICE being reffered in DMA_FROM_DEVICE is
> DMAC rather than FIFO TX register.
> So from DMAC's point of view, the dma of DMAC->FIFO TX register
> should be think of it as DMA_FROM_DEVICE. 
> 
> >>  };
> >>
> >>  static bool ipmmu_device_is_allowed(struct device *dev)
> >
> >Gr{oetje,eeting}s,
> >
> >                        Geert
> 
> Sincerely,
> Ryo Takakura

Hi Geert and Takakura-san,

My sincere apologies for the very late follow-up email. Unfortunately, I
don't have access to Gray Hawk Single / Sparrow Hawk boards myself, so I
have not been able to reproduce the issue locally. Takakura-san does have
access to them and was able to reproduce the problem, thanks, Takakura-san.

As an update, at least for me, this patch will no longer be necessary due
to another ongoing work, "NTB transport backed by remote DW eDMA":
https://lore.kernel.org/all/20251129160405.2568284-1-den@xxxxxxxxxxxxx/

Given that, I'm rather inclined to keep this on hold for now, unless you
have any objections.

Thank you both for your time and support.

Koichiro