Re: [PATCH 2/4] soc: apple: rtkit: Implement OSLog buffers properly

["Sven Peter" <sven@svenpeter.dev>]

On Wed, Feb 26, 2025, at 18:05, Alyssa Rosenzweig wrote:
>> >> +	if (ep == APPLE_RTKIT_EP_OSLOG) {
>> >> +		buffer->size = FIELD_GET(APPLE_RTKIT_OSLOG_SIZE, msg);
>> >> +		buffer->iova = FIELD_GET(APPLE_RTKIT_OSLOG_IOVA, msg) << 12;
>> >> +	} else {
>> >> +		buffer->size = FIELD_GET(APPLE_RTKIT_BUFFER_REQUEST_SIZE, msg) << 12;
>> >> +		buffer->iova = FIELD_GET(APPLE_RTKIT_BUFFER_REQUEST_IOVA, msg);
>> >> +	}
>> >
>> > The shifts are suspiciously asymmetric. Are we really sure this is
>> > correct? My guess is that both size & iova for both oslog & buffer need
>> > to be page-aligned, so all 4 lines should be shifted, and the bit
>> > offsets should be adjusted in turn, and the lower 12-bits in oslog_size
>> > and buffer_iova are reserved. But that's just a guess.
>> >
>> > Anyway if this logic is really what we want it deserves a comment
>> > because it looks like a typo.
>> 
>> That guess can't be true for syslog since there's no change in behavior here
>> and the syslog endpoint has been working fine so far. This common code is
>> also used for other endpoints that request buffers and there haven't been
>> any issues there either. The size is just passed in "number of 4k chunks"
>> and the IOVA needs no additional fixups.
>> 
>> 
>> The entire reason for this commit is because this common logic just didn't
>> work for oslog. Our u-boot fork uses the same logic as used here [1]. We're stealing
>> a chunk of MTP's SRAM to make hand-off to Linux easier there. If either size or
>> IOVA was off by a factor 0x4000 this would've never worked in the first
>> place.
>
> I'm not suggesting things are off by a factor of 4k. Rather I'm
> questioning what the behaviour is when we're not 4k aligned. (I.e.
> the syslog or oslog buffer does not both start and end at 4k
> boundaries.)
>
> If we're aligned, all our bottom bits are 0, and hypothetically we're
> putting 0 into "reserved-must be zero" bits.
>
> I guess it's inconsequential if everything is 4k aligned in practice.
> But .. is it? I don't know.


For the common buffers at least we sometimes _get_ the address from the
co-processor when it e.g. points to its internal SRAM. We could access any
buffer aligned to a 4 byte boundary in that case because it's just MMIO
access then and I'm not even sure how strongly the buffer passed to us will be aligned.
I'd rather not zero out bits there because we just cannot be sure those really
are reserved.
If the co-processor only asks for a buffer we'll grab it using dma_alloc_coherent
and it's going to be aligned anyway then.


For oslog there aren't even any "reserved lower bits" in the message. It really
expects the down-shifted address on the wire starting at bit 0.



Sven