On Wed, Nov 28, 2018 at 12:48 PM Russell King - ARM Linux
[off-list ref] wrote:

On Wed, Nov 28, 2018 at 11:27:17AM -0800, David Miller wrote:

quoted

From: Linus Torvalds <torvalds@linux-foundation.org>
Date: Wed, 28 Nov 2018 10:00:06 -0800

quoted

Not all memory is accessible even to the kernel. If you have memory
that shows up in the last page of phys_addr_t, you just mark it
reserved at boot-time.

It's not the physical memory at the end that needs to be reserved.

It's the IOMMU mapping arena.

True, if and only if you have an IOMMU.

Where there isn't an IOMMU, then we'd have to reserve every page that
that translates to a bus address in the top 4K of dma_addr_t on any
bus in the system - that means knowing early in the kernel
initialisation about all buses in the system so we can detect and
reserve these pages.

The arch and platform differences/confusion as to "what is DMA error
value" is the reason why dma_mapping_error is part dma ops. I
understand that iy would make sense to reduce the overhead of an
additional call, however, I am not sure if would be trivial change to
address this.

I was down this path of trying to address the missing mapping error
checks a few years ago and introduced dma_mapping_error checks in the
DMA DEBUG API. As you might already know that there is no common
definition for the mapping error.

Quick look at the defines shows:

#define CALGARY_MAPPING_ERROR 0
#define S390_MAPPING_ERROR  (~(dma_addr_t) 0x0)
#define SPARC_MAPPING_ERROR (~(dma_addr_t)0x0)

This is the reason why there is a arch or in some cases bus specific
mapping_error ops is needed.
We could unify this and fix all these first. I haven't looked at the
patch set closely, maybe you are already doing this.

thanks,
-- Shuah