| Message ID | 1485532996-19719-1-git-send-email-geert+renesas@glider.be (mailing list archive) |
|---|---|
| State | Superseded |
| Delegated to: | Geert Uytterhoeven |
| Headers | show |
Hi Geert, On Sat, Jan 28, 2017 at 1:03 AM, Geert Uytterhoeven <geert+renesas@glider.be> wrote: > Some IOMMUs (e.g. Renesas IPMMU/VMSA) support only page sizes of 4 KiB, > 2 MiB, and 1 GiB. > > With the default setting of CONFIG_CMA_ALIGNMENT = 8, allocations larger > than 1 MiB are aligned to a 1 MiB boundary only. Hence a 2 MiB > allocation may not be aligned, leading to a mapping of 512 4 KiB pages. > > Increase CONFIG_CMA_ALIGNMENT to allow mapping a 2 MiB buffer using a > single PTE, decreasing memory usage and TLB pressure. > > Signed-off-by: Geert Uytterhoeven <geert+renesas@glider.be> > --- > Is this useful? > > Should there instead be different defaults in Kconfig, depending on > enabled platform support? I think there is a dependency on the kernel page size configuration as well. In case of 16 KiB or 64 KiB page size configuration other large page sizes may be required. Thanks, / magnus
On 01/02/17 13:45, Magnus Damm wrote: > Hi Geert, > > On Sat, Jan 28, 2017 at 1:03 AM, Geert Uytterhoeven > <geert+renesas@glider.be> wrote: >> Some IOMMUs (e.g. Renesas IPMMU/VMSA) support only page sizes of 4 KiB, >> 2 MiB, and 1 GiB. >> >> With the default setting of CONFIG_CMA_ALIGNMENT = 8, allocations larger >> than 1 MiB are aligned to a 1 MiB boundary only. Hence a 2 MiB >> allocation may not be aligned, leading to a mapping of 512 4 KiB pages. >> >> Increase CONFIG_CMA_ALIGNMENT to allow mapping a 2 MiB buffer using a >> single PTE, decreasing memory usage and TLB pressure. >> >> Signed-off-by: Geert Uytterhoeven <geert+renesas@glider.be> >> --- >> Is this useful? >> >> Should there instead be different defaults in Kconfig, depending on >> enabled platform support? > > I think there is a dependency on the kernel page size configuration as > well. In case of 16 KiB or 64 KiB page size configuration other large > page sizes may be required. I don't see the original patch (linux-arm-kernel doesn't look to have it archived either), but I think bumping the default up to 2MB to match our normal section size sounds generally reasonable - users can still override it manually in their config, right? For 16KB and 64KB granules, it is at least the size covered by contiguous-hinted pages; the actual section sizes there are far too big to be practical for alignment purposes anyway (32MB and 512MB respectively). Robin. > > Thanks, > > / magnus > > _______________________________________________ > linux-arm-kernel mailing list > linux-arm-kernel@lists.infradead.org > http://lists.infradead.org/mailman/listinfo/linux-arm-kernel >
On Wed, Feb 01, 2017 at 02:06:45PM +0000, Robin Murphy wrote: > On 01/02/17 13:45, Magnus Damm wrote: > > Hi Geert, > > > > On Sat, Jan 28, 2017 at 1:03 AM, Geert Uytterhoeven > > <geert+renesas@glider.be> wrote: > >> Some IOMMUs (e.g. Renesas IPMMU/VMSA) support only page sizes of 4 KiB, > >> 2 MiB, and 1 GiB. > >> > >> With the default setting of CONFIG_CMA_ALIGNMENT = 8, allocations larger > >> than 1 MiB are aligned to a 1 MiB boundary only. Hence a 2 MiB > >> allocation may not be aligned, leading to a mapping of 512 4 KiB pages. > >> > >> Increase CONFIG_CMA_ALIGNMENT to allow mapping a 2 MiB buffer using a > >> single PTE, decreasing memory usage and TLB pressure. > >> > >> Signed-off-by: Geert Uytterhoeven <geert+renesas@glider.be> > >> --- > >> Is this useful? > >> > >> Should there instead be different defaults in Kconfig, depending on > >> enabled platform support? > > > > I think there is a dependency on the kernel page size configuration as > > well. In case of 16 KiB or 64 KiB page size configuration other large > > page sizes may be required. > > I don't see the original patch (linux-arm-kernel doesn't look to have it > archived either), but I think bumping the default up to 2MB to match our > normal section size sounds generally reasonable - users can still > override it manually in their config, right? For 16KB and 64KB granules, > it is at least the size covered by contiguous-hinted pages; the actual > section sizes there are far too big to be practical for alignment > purposes anyway (32MB and 512MB respectively). I am happy to take this patch on the strength of the above but an Ack or similar tag from Robin would be nice. Geert, perhaps you could repost the patch with Robin CCed?
Hi Magnus, On Wed, Feb 1, 2017 at 2:45 PM, Magnus Damm <magnus.damm@gmail.com> wrote: > On Sat, Jan 28, 2017 at 1:03 AM, Geert Uytterhoeven > <geert+renesas@glider.be> wrote: >> Some IOMMUs (e.g. Renesas IPMMU/VMSA) support only page sizes of 4 KiB, >> 2 MiB, and 1 GiB. >> >> With the default setting of CONFIG_CMA_ALIGNMENT = 8, allocations larger >> than 1 MiB are aligned to a 1 MiB boundary only. Hence a 2 MiB >> allocation may not be aligned, leading to a mapping of 512 4 KiB pages. >> >> Increase CONFIG_CMA_ALIGNMENT to allow mapping a 2 MiB buffer using a >> single PTE, decreasing memory usage and TLB pressure. >> >> Signed-off-by: Geert Uytterhoeven <geert+renesas@glider.be> >> --- >> Is this useful? >> >> Should there instead be different defaults in Kconfig, depending on >> enabled platform support? > > I think there is a dependency on the kernel page size configuration as > well. In case of 16 KiB or 64 KiB page size configuration other large > page sizes may be required. This is about IOMMU pages, not (CPU)MMU pages? Still, there is indeed some dependency, as CONFIG_CMA_ALIGNMENT is a shift relative to PAGE_SHIFT (i.e. CONFIG_ARM64_PAGE_SHIFT = 12 in arm64_defconfig). Gr{oetje,eeting}s, Geert -- Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- geert@linux-m68k.org 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
Hi Robin, On Wed, Feb 1, 2017 at 3:06 PM, Robin Murphy <robin.murphy@arm.com> wrote: > On 01/02/17 13:45, Magnus Damm wrote: >> On Sat, Jan 28, 2017 at 1:03 AM, Geert Uytterhoeven >> <geert+renesas@glider.be> wrote: >>> Some IOMMUs (e.g. Renesas IPMMU/VMSA) support only page sizes of 4 KiB, >>> 2 MiB, and 1 GiB. >>> >>> With the default setting of CONFIG_CMA_ALIGNMENT = 8, allocations larger >>> than 1 MiB are aligned to a 1 MiB boundary only. Hence a 2 MiB >>> allocation may not be aligned, leading to a mapping of 512 4 KiB pages. >>> >>> Increase CONFIG_CMA_ALIGNMENT to allow mapping a 2 MiB buffer using a >>> single PTE, decreasing memory usage and TLB pressure. >>> >>> Signed-off-by: Geert Uytterhoeven <geert+renesas@glider.be> >>> --- >>> Is this useful? >>> >>> Should there instead be different defaults in Kconfig, depending on >>> enabled platform support? >> >> I think there is a dependency on the kernel page size configuration as >> well. In case of 16 KiB or 64 KiB page size configuration other large >> page sizes may be required. > > I don't see the original patch (linux-arm-kernel doesn't look to have it > archived either), but I think bumping the default up to 2MB to match our These days linux-arm-kernel (infradead) tends to reject my patches :-( linux-renesas-soc patchwork does have it: https://patchwork.kernel.org/patch/9542173/ > normal section size sounds generally reasonable - users can still > override it manually in their config, right? For 16KB and 64KB granules, > it is at least the size covered by contiguous-hinted pages; the actual > section sizes there are far too big to be practical for alignment > purposes anyway (32MB and 512MB respectively). Yes, it can be overridden by the user. One slight annoyance is the dependency on PAGE_SHIFT: if you increase ARM64_PAGE_SHIFT, you want to decrease CONFIG_CMA_ALIGNMENT. Gr{oetje,eeting}s, Geert -- Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- geert@linux-m68k.org 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
diff --git a/arch/arm64/configs/defconfig b/arch/arm64/configs/defconfig index cea17afae855f589..1acc422052050770 100644 --- a/arch/arm64/configs/defconfig +++ b/arch/arm64/configs/defconfig @@ -143,6 +143,7 @@ CONFIG_UEVENT_HELPER_PATH="/sbin/hotplug" CONFIG_DEVTMPFS=y CONFIG_DEVTMPFS_MOUNT=y CONFIG_DMA_CMA=y +CONFIG_CMA_ALIGNMENT=9 CONFIG_MTD=y CONFIG_MTD_M25P80=y CONFIG_MTD_SPI_NOR=y
Some IOMMUs (e.g. Renesas IPMMU/VMSA) support only page sizes of 4 KiB, 2 MiB, and 1 GiB. With the default setting of CONFIG_CMA_ALIGNMENT = 8, allocations larger than 1 MiB are aligned to a 1 MiB boundary only. Hence a 2 MiB allocation may not be aligned, leading to a mapping of 512 4 KiB pages. Increase CONFIG_CMA_ALIGNMENT to allow mapping a 2 MiB buffer using a single PTE, decreasing memory usage and TLB pressure. Signed-off-by: Geert Uytterhoeven <geert+renesas@glider.be> --- Is this useful? Should there instead be different defaults in Kconfig, depending on enabled platform support? --- arch/arm64/configs/defconfig | 1 + 1 file changed, 1 insertion(+)