Re: [PATCH v2 1/4] trace, cxl: Introduce a TRACE_EVENT for CXL poison records

[Alison Schofield <alison.schofield@intel.com>]

On Wed, Oct 12, 2022 at 05:46:20PM -0400, Steven Rostedt wrote:
> On Wed, 12 Oct 2022 14:28:17 -0700
> alison.schofield@intel.com wrote:
> 
> > +TRACE_EVENT(cxl_poison,
> > +
> > +	    TP_PROTO(pid_t pid, const char *region, const char *memdev,
> > +		     const char *pcidev, u64 hpa, u64 dpa, u32 length,
> > +		     u8 source, u8 flags, u64 overflow_t),
> > +
> > +	    TP_ARGS(pid, region, memdev, pcidev, hpa, dpa,
> > +		    length, source, flags, overflow_t),
> > +
> > +	    TP_STRUCT__entry(
> > +		__field(pid_t, pid)
> > +		__string(region, region ? region : "")
> > +		__string(memdev, memdev)
> > +		__string(pcidev, pcidev)
> > +		__field(u64, hpa)
> > +		__field(u64, dpa)
> > +		__field(u32, length)
> > +		__field(u8, source)
> > +		__field(u8, flags)
> > +		__field(u64, overflow_t)
> 
> The above looks nice and compact.
> 
> > +	    ),
> > +
> > +	    TP_fast_assign(
> > +		__entry->pid = pid;
> > +		__assign_str(region, region ? region : "");
> > +		__assign_str(memdev, memdev);
> > +		__assign_str(pcidev, pcidev);
> > +		__entry->hpa = hpa;
> > +		__entry->dpa = dpa;
> > +		__entry->length = length;
> > +		__entry->source = source;
> > +		__entry->flags = flags;
> > +		__entry->overflow_t = overflow_t;
> 
> But I wonder if it would be better to move the computation of patch 2
> here?
> 
> That is, this:
> 
> > +		for (i = 0; i < le16_to_cpu(po->count); i++) {
> > +			u32 len = le32_to_cpu(po->record[i].length) *
> > +					      CXL_POISON_LEN_MULT;
> > +			u64 addr = le64_to_cpu(po->record[i].address);
> > +			u8 source = addr & CXL_POISON_SOURCE_MASK;
> > +			u64 dpa = addr & CXL_POISON_START_MASK;
> > +			u64 hpa = 0;
> > +
> > +			trace_cxl_poison(current->pid, region_name,
> > +					 dev_name(&cxlmd->dev),
> > +					 dev_name(cxlds->dev), hpa, dpa, len,
> > +					 source, po->flags, overflow_t);
> > +		}
> 
> As:
> 
> // The trace_*_enabled() is a static branch which is true when the
> // tracepoint is enabled, and a nop when not (skipping the for block)

Thanks for pointing it out Steve. I will take advantage of that in next
version.

In this case, we'd expect, but not enforce, that the tracepoint would
be enabled, because otherwise we're just reading the data from the
device and throwing it away. 

I appreciate your help here!
Alison

> 
> 		for (i = 0; trace_cxl_poison_enabled() && i < le16_to_cpu(po->count); i++)
> 			trace_cxl_poison(region_name, cxlmd, cxlds, &po->record[i]);
> 
> And then have:
> 
> 	    TP_fast_assign(
> 		u32 len = le32_to_cpu(record->length) *
> 				      CXL_POISON_LEN_MULT;
> 		u64 addr = le64_to_cpu(record->address);
> 		u8 source = addr & CXL_POISON_SOURCE_MASK;
> 		u64 dpa = addr & CXL_POISON_START_MASK;
> 		u64 hpa = 0;
> 
> 		__entry->pid = current->pid;
> 
> { the above isn't needed as the trace event will have common_pid = current->pid }

got it, thanks.

> 
> 		__assign_str(region, region ? region : "");
> 		__assign_str(memdev, dev_name(&cxlmd->dev));
> 		__assign_str(pcidev, dev_name(cxlds->dev));
> 		__entry->hpa = hpa;
> 		__entry->dpa = dpa;
> 		__entry->length = length;
> 		__entry->source = source;
> 		__entry->flags = flags;
> 		__entry->overflow_t = overflow_t;
> 
> 
> Or something similar. This will keep the work out of the code path.
> 
> -- Steve
> 
> 
> > +	    ),
> > +
> > +	    TP_printk("pid:%d region:%s memdev:%s pcidev:%s hpa:0x%llx dpa:0x%llx length:0x%x source:%s flags:%s overflow_time:%llu",
> > +		__entry->pid,
> > +		__get_str(region),
> > +		__get_str(memdev),
> > +		__get_str(pcidev),
> > +		__entry->hpa,
> > +		__entry->dpa,
> > +		__entry->length,
> > +		show_poison_source(__entry->source),
> > +		show_poison_flags(__entry->flags),
> > +		__entry->overflow_t)
> > +);
> > +#endif /* _CXL_TRACE_H */