Re: [PATCH 2/5] cxl: Rename CXL_DECODER_HOSTONLYMEM/DEVMEM

From: Huang, Ying
Date: Mon Oct 21 2024 - 00:44:38 EST


Dan Williams <dan.j.williams@xxxxxxxxx> writes:

> Huang, Ying wrote:
>> Hi, Dan,
>>
>> Dan Williams <dan.j.williams@xxxxxxxxx> writes:
>>
>> > Huang Ying wrote:
>> >> Previously, CXL type3 devices (memory expanders) use host only
>> >> coherence (HDM-H), while CXL type2 devices (accelerators) use dev
>> >> coherence (HDM-D). So the name of the target device type of a cxl
>> >> decoder is CXL_DECODER_HOSTONLYMEM for type3 devices and
>> >> CXL_DECODER_DEVMEM for type2 devices. However, this isn't true
>> >> anymore. CXL type3 devices can use dev coherence + back
>> >> invalidation (HDM-DB) too.
>> >>
>> >> To avoid confusion between the device type and coherence, the patch
>> >> renames CXL_DECODER_HOSTONLYMEM/DEVMEM to CXL_DECODER_EXPANDER/ACCEL.
>> >
>> > This does not look like an improvement to me. Type-3 devices that
>> > support back-invalidate are DEVMEM devices. The device plays a role in
>> > the coherence.
>> >
>> > Your explanation is the reverse of this commit:
>> >
>> > 5aa39a9165cf cxl/port: Rename CXL_DECODER_{EXPANDER, ACCELERATOR} => {HOSTONLYMEM, DEVMEM}
>> >
>> > ...so I am confused what motivated this rename?
>>
>> Sorry, I am confused about the target_type and coherence and forgot to
>> check the history. In some places, current kernel still hints
>> target_type (CXL_DECODER_HOSTONLYMEM/DEVMEM) as expander/accelerator.
>> Should we change them to avoid confusion in the future?
>>
>> $ grep expander -r drivers/cxl/
>> drivers/cxl/cxl.h:346: * @target_type: accelerator vs expander (type2 vs type3) selector
>> drivers/cxl/core/region.c:2450: * @type: select whether this is an expander or accelerator (type-2 or type-3)
>> drivers/cxl/core/port.c:141: return sysfs_emit(buf, "expander\n");
>>
>> The last one is
>>
>> static ssize_t target_type_show(struct device *dev,
>> struct device_attribute *attr, char *buf)
>> {
>> struct cxl_decoder *cxld = to_cxl_decoder(dev);
>>
>> switch (cxld->target_type) {
>> case CXL_DECODER_DEVMEM:
>> return sysfs_emit(buf, "accelerator\n");
>> case CXL_DECODER_HOSTONLYMEM:
>> return sysfs_emit(buf, "expander\n");
>> }
>> return -ENXIO;
>> }
>> static DEVICE_ATTR_RO(target_type);
>>
>> for decoder device. This is a testing ABI documented in,
>>
>> Documentation/ABI/testing/sysfs-bus-cxl
>>
>> Is it OK to change this?
>
> No, why does it need to change?

For example, if the target_type is CXL_DECODER_DEVMEM, while the device
is a memory expander with HDM-DB protocol. The sysfs will show it as
"accelerator". This may make users or developers confusing. If we can
show "hostonlymem"/"devmem", that may be better. But apparently, we
cannot change ABI.

> It is unfortunate, but ABI's are forever. The place to clarify that this
> decoder is participating in HDM-D[B] vs HDM-H protocol rather than being
> an "accelerator" or "expander" device would be in user tooling like
> cxl-cli. sysfs is just a transport, not a UI.

Although it's not perfect, this is a solution. Another way to solve
this is to separate device coherence (HOSTONLY vs. DEV) and device type
(ACCELERATOR vs. EXPANDER). In this way, if the "target_type" in sysfs
designates device type, it could show "expander" for memory expander
even if HDM-DB protocol is used.

Another possibility, can we just remove this sysfs file?

--
Best Regards,
Huang, Ying