Messages in this thread | | | Date | Wed, 11 Oct 2023 07:11:00 +0200 | From | Christoph Hellwig <> | Subject | Re: [PATCH v2] dma-mapping: fix dma_addressing_limited if dma_range_map can't cover all system RAM |
| |
On Tue, Oct 10, 2023 at 10:40:44AM +0100, Robin Murphy wrote: > On 2023-10-10 08:48, Christoph Hellwig wrote: >> dma_addressing_limited is called for every dma (direct) map, and this >> new code is way to heavy for that. We'll need to cache the information >> somehow. > > But is it? That was my instinctive reaction as well, but AFAICS it is only > actually used by drivers in setup paths, either directly or via > dma_max_mapping_size(). The dma_capable() check which we *do* use for every > actual mapping still just checks whether the given physical address exists > in the range map. I believe the underlying problem here is when > dma_capable() say the mapping needs bouncing, but then it's too big for > that to succeed, since dma_max_mapping_size() gave the wrong answer to > start with.
Ah, indeed, I somehw misremembered calling it in the mapping code.
Justing, can you still respin this a bit, add a prep patch to move dma_addressing_limited so that it is exported instead of the new low-level helper, and fix up coding style issues like the overly long lines of possible? If it's not perfect I'll fix up the rest later.
| |