On Mon, Dec 07, 2020 at 04:54:21PM -0800, Dan Williams wrote:

[ add perf maintainers ]

On Sun, Nov 8, 2020 at 1:16 PM Vaibhav Jain [off-list ref] wrote:

quoted

Implement support for exposing generic nvdimm statistics via newly
introduced dimm-command ND_CMD_GET_STAT that can be handled by nvdimm
command handler function and provide values for these statistics back
to libnvdimm. Following generic nvdimm statistics are defined as an
enumeration in 'uapi/ndctl.h':

* "media_reads" : Number of media reads that have occurred since reboot.
* "media_writes" : Number of media writes that have occurred since reboot.
* "read_requests" : Number of read requests that have occurred since reboot.
* "write_requests" : Number of write requests that have occurred since reboot.

Perhaps document these as "since device reset"? As I can imagine some
devices might have a mechanism to reset the count outside of "reboot"
which is a bit ambiguous.

quoted

* "total_media_reads" : Total number of media reads that have occurred.
* "total_media_writes" : Total number of media writes that have occurred.
* "total_read_requests" : Total number of read requests that have occurred.
* "total_write_requests" : Total number of write requests that have occurred.

Apart from ND_CMD_GET_STAT ioctl these nvdimm statistics are also
exposed via sysfs '<nvdimm-device>/stats' directory for easy user-space
access like below:

/sys/class/nd/ndctl0/device/nmem0/stats # tail -n +1 *
==> media_reads <==
252197707602
==> media_writes <==
20684685172
==> read_requests <==
658810924962
==> write_requests <==
404464081574

Hmm, I haven't looked but how hard would it be to plumb these to be
perf counter-events. So someone could combine these with other perf
counters?

quoted

In case a specific nvdimm-statistic is not supported than nvdimm
command handler function can simply return an error (e.g -ENOENT) for
request to read that nvdimm-statistic.

Makes sense, but I expect the perf route also has a way to enumerate
which statistics / counters are supported. I'm not opposed to also
having them in sysfs, but I think perf support should be a first class
citizen.

arch/x86/events/msr.c might be a good starting point for a software pmu
delivering pure counters.