Re: [RFC PATCH 5/9] cxl/mem: Trace DRAM Event Record

[Ira Weiny]

On Thu, Aug 25, 2022 at 11:46:32AM +0100, Jonathan Cameron wrote:
> On Fri, 12 Aug 2022 22:32:39 -0700
> ira.weiny@xxxxxxxxx wrote:
> 
> > From: Ira Weiny <ira.weiny@xxxxxxxxx>
> > 
> > CXL v3.0 section 8.2.9.2.1.2 defines the DRAM Event Record.
> > 
> > Determine if the event read is a DRAM event record and if so trace the
> > record.
> > 
> > Signed-off-by: Ira Weiny <ira.weiny@xxxxxxxxx>
> > 
> > ---
> > This record has a very odd byte layout with 2 - 16 bit fields
> > (validity_flags and column) aligned on an odd byte boundary.  In
> > addition nibble_mask and row are oddly aligned.
> > 
> > I've made my best guess as to how the endianess of these fields should
> > be resolved.  But I'm happy to hear from other folks if what I have is
> > wrong.
> My assumption is same as you.  We should sanity check of course by
> poking relevant people.  
> 
> Similar comments in here to previous.  Use the get_unaligned_le24()
> accessors + consider not printing invalid fields.

Yea I've already converted the 3 byte fields to get_unaligned_le24()

> > 
> > struct cxl_evt_dram_rec {
> > 	struct cxl_event_record_hdr hdr;
> > 	__le64 phys_addr;
> > 	u8 descriptor;
> > 	u8 type;
> > 	u8 transaction_type;
> > 	u16 validity_flags;
> > 	u8 channel;
> > 	u8 rank;
> > 	u8 nibble_mask[CXL_EVT_DER_NIBBLE_MASK_SIZE];
> > 	u8 bank_group;
> > 	u8 bank;
> > 	u8 row[CXL_EVT_DER_ROW_SIZE];
> > 	u16 column;
> > 	u8 correction_mask[CXL_EVT_DER_CORRECTION_MASK_SIZE];
> > } __packed;
> > ---
> >  drivers/cxl/core/mbox.c           |  16 +++++
> >  drivers/cxl/cxlmem.h              |  24 +++++++
> >  include/trace/events/cxl-events.h | 114 ++++++++++++++++++++++++++++++
> >  3 files changed, 154 insertions(+)
> > 
> > diff --git a/drivers/cxl/core/mbox.c b/drivers/cxl/core/mbox.c
> > index 0e433f072163..6414588a3c7b 100644
> > --- a/drivers/cxl/core/mbox.c
> > +++ b/drivers/cxl/core/mbox.c
> > @@ -717,6 +717,14 @@ static const uuid_t gen_media_event_uuid =
> >  	UUID_INIT(0xfbcd0a77, 0xc260, 0x417f,
> >  		  0x85, 0xa9, 0x08, 0x8b, 0x16, 0x21, 0xeb, 0xa6);
> >  
> > +/*
> > + * DRAM Event Record
> > + * CXL v3.0 section 8.2.9.2.1.2; Table 8-44
> rev3.0, r3.0 or just 3.0  

Already done.

> 
> > + */
> > +static const uuid_t dram_event_uuid =
> > +	UUID_INIT(0x601dcbb3, 0x9c06, 0x4eab,
> > +		  0xb8, 0xaf, 0x4e, 0x9b, 0xfb, 0x5c, 0x96, 0x24);
> > +
> >  static void cxl_trace_event_record(const char *dev_name,
> >  				   enum cxl_event_log_type type,
> >  				   struct cxl_get_event_payload *payload)
> > @@ -731,6 +739,14 @@ static void cxl_trace_event_record(const char *dev_name,
> >  		return;
> >  	}
> >  
> > +	if (uuid_equal(id, &dram_event_uuid)) {
> Why not else if?  Should be obvious to compiler that multiple uuid_equal
> conditions can't match, but even better to not make it try hard perhaps?

Sure else if can work.

> 
> > +		struct cxl_evt_dram_rec *rec =
> > +				(struct cxl_evt_dram_rec *)&payload->record;
> > +
> > +		trace_cxl_dram_event(dev_name, type, rec);
> > +		return;
> > +	}
> > +
> >  	/* For unknown record types print just the header */
> >  	trace_cxl_event(dev_name, type, &payload->record);
> >  }
> > diff --git a/drivers/cxl/cxlmem.h b/drivers/cxl/cxlmem.h
> > index 33669459ae4b..50536c0a7850 100644
> > --- a/drivers/cxl/cxlmem.h
> > +++ b/drivers/cxl/cxlmem.h
> > @@ -421,6 +421,30 @@ struct cxl_evt_gen_media {
> >  	u8 component_id[CXL_EVT_GEN_MED_COMP_ID_SIZE];
> >  } __packed;
> >  
> > +/*
> > + * DRAM Event Record - DER
> > + * CXL v3.0 section 8.2.9.2.1.2; Table 3-44
> > + */
> > +#define CXL_EVT_DER_NIBBLE_MASK_SIZE		3
> > +#define CXL_EVT_DER_ROW_SIZE			3
> > +#define CXL_EVT_DER_CORRECTION_MASK_SIZE	0x20
> > +struct cxl_evt_dram_rec {
> > +	struct cxl_event_record_hdr hdr;
> > +	__le64 phys_addr;
> > +	u8 descriptor;
> > +	u8 type;
> > +	u8 transaction_type;
> > +	u16 validity_flags;
> I've not tried it, but can we just mark these as __le16 and use
> the unaligned accessors?  get_unaligned_le16 etc

get_unaligned_le16() requires a byte array...

So I think this needs to be:

	u8 validity_flags[2];

Now that I know about those calls I think this does make a lot more sense.  The
test code works but I knew that it would be sketchy with real devices.

I'll adjust this.

> Also there is get_unaligned_le24() for the 3 byte ones.

Yea done.

[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);
> > +
> > +		/* DRAM */
> > +		__entry->phys_addr = le64_to_cpu(rec->phys_addr);
> > +		__entry->descriptor = rec->descriptor;
> > +		__entry->type = rec->type;
> > +		__entry->transaction_type = rec->transaction_type;
> > +		__entry->validity_flags = le16_to_cpu(rec->validity_flags);
> > +		__entry->channel = rec->channel;
> > +		__entry->rank = rec->rank;
> > +		__entry->nibble_mask = rec->nibble_mask[0] << 24 |
> > +				       rec->nibble_mask[1] << 16 |
> > +				       rec->nibble_mask[2] << 8; /* 3 byte LE ? */
> 
> Use get_unalinged_le24() ? I'd definitely expect these to be le24.
> 
> 
> > +		__entry->nibble_mask = le32_to_cpu(__entry->nibble_mask);
> 
> That doesn't look right.  You will have unwound the endianness using
> the shifts above. Don't convert it again (noop on le systems, so you
> probably won't see a problem when testing).

I thought I did it right with 2 shifts.  But regardless using
get_unalinged_le24() is better and I've already changed it.

> 
> > +		__entry->bank_group = rec->bank_group;
> > +		__entry->bank = rec->bank;
> > +		__entry->row = rec->row[0] << 24 |
> > +			       rec->row[1] << 16 |
> > +			       rec->row[2] << 8; /* 3 byte LE ? */
> 
> get_unaligned_le24()

... and this one.

> 
> > +		__entry->row = le32_to_cpu(__entry->row);
> 
> > +		__entry->column = le16_to_cpu(rec->column);
> > +		memcpy(__entry->cor_mask, &rec->correction_mask,
> > +			CXL_EVT_DER_CORRECTION_MASK_SIZE);
> > +	),
> > +
> > +	TP_printk("%s: %s time=%llu id=%pUl handle=%x related_handle=%x hdr_flags='%s': " \
> > +		  "phys_addr=%llx volatile=%s desc='%s' type='%s' trans_type='%s' channel=%u " \
> > +		  "rank=%u nibble_mask=%x bank_group=%u bank=%u row=%u column=%u " \
> > +		  "cor_mask=%s valid_flags='%s'",
> > +		__get_str(dev_name), show_log_type(__entry->log),
> > +		__entry->timestamp, __entry->id, __entry->handle,
> > +		__entry->related_handle, show_hdr_flags(__entry->flags),
> > +		__entry->phys_addr & ~CXL_GMER_PHYS_ADDR_MASK,
> > +		(__entry->phys_addr & CXL_GMER_PHYS_ADDR_VOLATILE) ? "TRUE" : "FALSE",
> > +		show_event_desc_flags(__entry->descriptor),
> As before can we not print the invalid ones based on the validity flags?
> 
> Few years ago now, but I did something along those lines for the CCIX equivalent of
> this stuff.  (honestly can't remember much about it now though!)
> Was a bit fiddly but lead to nicer prints in my opinion.
> 
> https://lore.kernel.org/all/20191114133919.32290-2-Jonathan.Cameron@xxxxxxxxxx/

I'm still not seeing anything which alters the actual print in this patch or
ras_event.h

Perhaps I'm missing what you mean by selecting the valid fields.

Something will have to change the TP_printk() format itself from what I can see
and I don't see a way to do that within the trace infrastructure.

We _could_ do that within the C code where trace_dram() is called.  But I'd
like to keep all the info together and let user space decode more than what the
kernel may know.

Ira