Re: RPi4B 3 GiByte pcie limitation: The following change survived my huge-file duplicate-and-diff tests so far

I have experimented with the segment size. I cannot have tested every scenario, but I did not find it made any difference, even if I went down to segments of only a page.

I understand the original Linux driver used a bounce buffer where each DMA transaction was limited to a few MB. However that design was dropped when the driver was merged into the mainline kernel.

— RHC.

‐‐‐‐‐‐‐ Original Message ‐‐‐‐‐‐‐
On Saturday, 3 October 2020 03:35, Mark Millard <[email protected]> wrote:

> [Operator error note and another note.]
>
> On 2020-Oct-2, at 13:56, Mark Millard <marklmi at yahoo.com> wrote:
>
> > It appears to me that 3 GiByte works for lowaddr and
> > maxsize but that the maxsegsz needs to be limited to
> > 1 GiByte [or so, maybe (1 Gi - 1) Bytes].
> > I did the following based on all those notes that I
> > sent out, that included the ?_TXFR_LEN being limited
> > 30 bits for (normal) DMA engines (0-6) and the DMA4
> > engines (11-14) [but not DMA LITE (7-10)]:
>
> Ignore: I had the wrong kernel build installed.
> In fact the below fails when the intended kernel
> is tested.
>
> > svnlite diff /usr/src/sys/arm/broadcom/bcm2835/bcm2838_pci.c /usr/src/sys/arm/broadcom/bcm2835/bcm2838_xhci.c
> >
> > ==============================================================================================================
> >
> > Index: /usr/src/sys/arm/broadcom/bcm2835/bcm2838_pci.c
> >
> > =======================================================
> >
> > --- /usr/src/sys/arm/broadcom/bcm2835/bcm2838_pci.c (revision 365932)
> > +++ /usr/src/sys/arm/broadcom/bcm2835/bcm2838_pci.c (working copy)
> > @@ -105,7 +105,8 @@
> > *
> >
> > -   Whatever the true maximum address, 960 MiB works.
> >     */
> >     -#define DMA_HIGH_LIMIT 0x3c000000
> >     +#define DMA_SEG_HIGH_LIMIT 0x3c000000
> >     +#define DMA_TOTAL_HIGH_LIMIT 0xc0000000
> >     #define MAX_MEMORY_LOG2 0x21
> >     #define REG_VALUE_DMA_WINDOW_LOW (MAX_MEMORY_LOG2 - 0xf)
> >     #define REG_VALUE_DMA_WINDOW_HIGH 0x0
> >     @@ -642,12 +643,12 @@
> >     /
> >     error = bus_dma_tag_create(bus_get_dma_tag(dev), / parent /
> >     1, 0, / alignment, bounds */
> >
> > -       DMA_HIGH_LIMIT,			/* lowaddr */
> >
> >
> >
> > -       DMA_TOTAL_HIGH_LIMIT,		/* lowaddr */
> >         BUS_SPACE_MAXADDR,			/* highaddr */
> >         NULL, NULL,				/* filter, filterarg */
> >
> >
> >
> > -       DMA_HIGH_LIMIT,			/* maxsize */
> >
> >
> >
> > -       DMA_TOTAL_HIGH_LIMIT,		/* maxsize */
> >         BUS_SPACE_UNRESTRICTED,		/* nsegments */
> >
> >
> >
> > -       DMA_HIGH_LIMIT,			/* maxsegsize */
> >
> >
> >
> > -       DMA_SEG_HIGH_LIMIT,			/* maxsegsize */
> >         0, 					/* flags */
> >         NULL, NULL,				/* lockfunc, lockarg */
> >         &sc->dmat);
> >
> >
> >
> > Then, with such a kernel installed, I used:
> > -rw-r--r-- 1 root wheel 11570948096 Jul 18 18:32:37 2020 clang-armv7-on-aarch64.tar
> > (so much larger than the 8 GiByte RAM) and did the following with
> > the file:
> >
> > cp -aRx clang-armv7-on-aarch64.tar clang-armv7-on-aarch64.alt_tar
> >
> > ==================================================================
> >
> > diff clang-armv7-on-aarch64.tar clang-armv7-on-aarch64.alt_tar
> >
> > ===============================================================
> >
> > ==
> >
> > Such tests have been passing.
> > I do not know what your test procedures were. But if the above
> > is suggestive, you might want to try testing something like the
> > above via your test procedures.
>
> Going in a different direction for potential maxsegsz figures
> (and the like): xhci instead of bcm2711 material . . .
>
> extensible-host-controler-interface-usb-xhci.pdf reports that
> normal TRB Transfer Length fields have bits 16:0 and "Valid
> values are 0 to 64K". (Transfer Event TRB Transfer Lengths
> have 23:0 but above 0x10000 as a value is only for Event Data
> flag being 1 or for condition code is stopped - short packet.)
>
> I've not experimented with this.
>
> ============================================================================================================================================================================================================================================================================================================================================================================================================================================================================
>
> Mark Millard
> marklmi at yahoo.com
> ( dsl-only.net went
> away in early 2018-Mar)


_______________________________________________
[email protected] mailing list
https://lists.freebsd.org/mailman/listinfo/freebsd-arm
To unsubscribe, send any mail to "[email protected]"