Re: [RFC PATCH 6/9] cxl/mem: Trace Memory Module Event Record

[Jonathan Cameron <Jonathan.Cameron@huawei.com>]

On Wed, 14 Sep 2022 14:17:14 -0700
Ira Weiny <ira.weiny@intel.com> wrote:

> On Thu, Aug 25, 2022 at 11:58:42AM +0100, Jonathan Cameron wrote:
> > On Fri, 12 Aug 2022 22:32:40 -0700
> > ira.weiny@intel.com wrote:
> >   
> > > From: Ira Weiny <ira.weiny@intel.com>
> > > 
> > > CXL v3.0 section 8.2.9.2.1.3 defines the Memory Module Event Record.
> > > 
> > > Determine if the event read is memory module record and if so trace the
> > > record.
> > > 
> > > Signed-off-by: Ira Weiny <ira.weiny@intel.com>  
> > Similar comments to on previous patches around using
> > get_unaligned_le*()  
> 
> Yep...
> 
> [snip]
> 
> > >  
> > > +/*
> > > + * Get Health Info Record
> > > + * CXL v3.0 section 8.2.9.8.3.1; Table 8-100
> > > + */
> > > +struct cxl_get_health_info {
> > > +	u8 health_status;
> > > +	u8 media_status;
> > > +	u8 add_status;
> > > +	u8 life_used;
> > > +	u16 device_temp;  
> > 
> > As previous - even though they aren't aligned, I'd have thought
> > __le16 etc will still work.  The unaligned accessors are fine
> > taking __le16 * for example.  
> 
> Ok my bad on using u16 here and I will change it.  I 100% agree that these
> should be __le16/__le32.  That said there is no need to use the unaligned
> accessors for the 16/32 bit fields.
> 
> The unaligned accessors cast the pointer to a __le16/__le32 type and no
> architecture redefines those.  So using le{16,32}_to_cpu() should work just
> fine on all archs.

If they are unaligned, make sure to use the unaligned accessors.

Key is that it's not a simple cast, but rather a cast to a packed
structure.  The C spec guarantees that those will be handled correctly
even on platforms that don't do unaligned accesses - it will have to
use multiple instructions to construct the unaligned access from
a set of small aligned ones.
The C Spec doesn't guarantee the same for a simple cast to an __le16.

There are some hints on this in:
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/include/asm-generic/unaligned.h?id=778aaefb8e864fc61f850539ea479554dd4caea1

I recall a full explanation of why this worked, but no idea where
to find that now - might be the thread referred to in that patch from
Arnd.

Jonathan


> 
> [snip]
> 
> > > +
> > > +	TP_fast_assign(
> > > +		/* Common */
> > > +		__assign_str(dev_name, dev_name);
> > > +		memcpy(__entry->id, &rec->hdr.id, UUID_SIZE);
> > > +		__entry->log = log;
> > > +		__entry->flags = le32_to_cpu(rec->hdr.flags_length) >> 8;
> > > +		__entry->handle = le16_to_cpu(rec->hdr.handle);
> > > +		__entry->related_handle = le16_to_cpu(rec->hdr.related_handle);
> > > +		__entry->timestamp = le64_to_cpu(rec->hdr.timestamp);
> > > +
> > > +		/* Memory Module Event */
> > > +		__entry->event_type = rec->event_type;
> > > +
> > > +		/* Device Health Info */
> > > +		__entry->health_status = rec->info.health_status;
> > > +		__entry->media_status = rec->info.media_status;
> > > +		__entry->life_used = rec->info.life_used;
> > > +		__entry->dirty_shutdown_cnt = le32_to_cpu(rec->info.dirty_shutdown_cnt);
> > > +		__entry->cor_vol_err_cnt = le32_to_cpu(rec->info.cor_vol_err_cnt);  
> > 
> > I've lost track, but my guess is some / all of these need the unaligned_get_le32()
> > etc rather than aligned form.  Maybe just be lazy and use the unaligned versions
> > even when things happen to be aligned - then we don't have to think about it
> > when reviewing :)  
> 
> See above.  I think the 16/32 bit fields work as intended except for my lack of
> using the correct type.
> 
> Ira