24.01.2019 13:24, Mikko Perttunen пишет:
On 23.1.2019 21.42, Dmitry Osipenko wrote:
23.01.2019 18:55, Dmitry Osipenko пишет:
23.01.2019 17:04, Thierry Reding пишет:
On Wed, Jan 23, 2019 at 04:41:44PM +0300, Dmitry Osipenko wrote:
23.01.2019 12:39, Thierry Reding пишет:
From: Thierry Reding treding@nvidia.com
On Tegra186 and later, the ARM SMMU provides an input address space that is 48 bits wide. However, memory clients can only address up to 40 bits. If the geometry is used as-is, allocations of IOVA space can end up in a region that cannot be addressed by the memory clients.
To fix this, restrict the IOVA space to the DMA mask of the host1x device. Note that, technically, the IOVA space needs to be restricted to the intersection of the DMA masks for all clients that are attached to the IOMMU domain. In practice using the DMA mask of the host1x device is sufficient because all host1x clients share the same DMA mask.
Signed-off-by: Thierry Reding treding@nvidia.com
drivers/gpu/drm/tegra/drm.c | 5 +++-- 1 file changed, 3 insertions(+), 2 deletions(-)
diff --git a/drivers/gpu/drm/tegra/drm.c b/drivers/gpu/drm/tegra/drm.c index 271c7a5fc954..0c5f1e6a0446 100644 --- a/drivers/gpu/drm/tegra/drm.c +++ b/drivers/gpu/drm/tegra/drm.c @@ -136,11 +136,12 @@ static int tegra_drm_load(struct drm_device *drm, unsigned long flags) if (tegra->domain) { u64 carveout_start, carveout_end, gem_start, gem_end; + u64 dma_mask = dma_get_mask(&device->dev); dma_addr_t start, end; unsigned long order; - start = tegra->domain->geometry.aperture_start; - end = tegra->domain->geometry.aperture_end; + start = tegra->domain->geometry.aperture_start & dma_mask; + end = tegra->domain->geometry.aperture_end & dma_mask; gem_start = start; gem_end = end - CARVEOUT_SZ;
Wow, so IOVA could address >32bits on later Tegra's. AFAIK, currently there is no support for a proper programming of the 64bit addresses in the drivers code, hence.. won't it make sense to force IOVA mask to 32bit for now and hope that the second halve of address registers happen to be 0x00000000 in HW?
I think this restriction only applies to display at this point. In practice you'd be hard put to trigger that case because IOVA memory is allocated from the bottom, so you'd actually need to use up to 4 GiB of IOVA space before hitting that.
That said, I vaguely remember typing up the patch to support writing the WINBUF_START_ADDR_HI register and friends, but it looks as if that was never merged.
I'll try to dig out that patch (or rewrite it, shouldn't be very difficult) and make it part of this series. I'd rather fix that issue than arbitrarily restrict the IOVA space, because that's likely to come back and bite us at some point.
Looking at falcon.c.. it writes 32bit address only. Something need to be done about it as well, seems there is FALCON_DMATRFMOFFS register to facilitate >32bits addressing.
Although scratch about FALCON_DMATRFMOFFS, it should be about something else. Mikko, could you please clarify whether falcon could load firmware only from a 32bit AS?
The DMA base address is set using DMATRFBASE, which requires 256B alignment, meaning 40 bits are available for the address. The DMATRFFBOFFS I believe is then used as a 32-bit offset to that value.
TRM (up to T196) suggests that DMATRFMOFFS is a 16bit offset. Is it a kind of TRM bug or I'm missing something?
I suppose it should be fine to just reserve carveout from the bottom of IOVA space and done with it.