Date:   Tue, 29 Aug 2017 09:26:12 -0700
From:   Dan Williams <dan.j.williams@...el.com>
To:     Yasunori Goto <y-goto@...fujitsu.com>
Cc:     Jerry Hoemann <Jerry.Hoemann@....com>,
        "linux-nvdimm@...ts.01.org" <linux-nvdimm@...ts.01.org>,
        "linux-kernel@...r.kernel.org" <linux-kernel@...r.kernel.org>
Subject: Re: [PATCH] libnvdimm: clean up command definitions

On Tue, Aug 29, 2017 at 3:09 AM, Yasunori Goto <y-goto@...fujitsu.com> wrote:
>> On Mon, Aug 28, 2017 at 6:03 PM, Yasunori Goto <y-goto@...fujitsu.com> wrote:
>> >> On Mon, Aug 28, 2017 at 1:50 PM, Jerry Hoemann <jerry.hoemann@....com> wrote:
>> >> >
>> >> > On Mon, Aug 28, 2017 at 08:45:32AM -0700, Dan Williams wrote:
>> >> >> Remove the command payloads that do not have an associated libnvdimm
>> >> >> ioctl. I.e. remove the payloads that would only ever be carried in the
>> >> >> ND_CMD_CALL envelope. This prevents userspace from growing unnecessary
>> >> >> dependencies on this kernel header when userspace already has everything
>> >> >> it needs to craft and send these commands.
>> >> >
>> >> > Userspace needs to include linux/ndctl.h to make the call as
>> >> > that is where nd_cmd_pkg is defined.
>> >> >
>> >> > So you want to have some structures defined in ndctl.h and other
>> >> > defined in the to be created libndctl-nfit.h?  Plus a third header
>> >> > file for the HPE non-root calls?
>> >>
>> >> Yes. ndctl.h exports the ioctl command payloads, everything that goes
>> >> inside of ND_CMD_CALL is defined by userspace headers. The
>> >> libndctl-nfit.h header is proposed as a place to land vendor agnostic
>> >> NFIT-defined payloads, and any vendor specific definitions would
>> >> remain internal to libndctl as they are today.
>> >>
>> >> > Will libndctl-nfit.h be generally available and installed?
>> >>
>> >> Yes, that's the plan.
>> >>
>> >> > Will it be clean so that other applications can use it to get these
>> >> > definitions?  Or will it be loaded w/ a bunch of stuff only useful
>> >> > to your ndctl command?
>> >>
>> >> Yes, that's the plan. It's a bug if libndctl-nfit.h is not generically
>> >> clean for issuing the NFIT root device commands via some ND_CMD_CALL
>> >> helpers from the base libndctl library.
>> >>
>> >> In other words libndctl-nfit.h defines the payload and libndctl
>> >> defines some general helpers for issuing commands.
>> >
>> > Maybe I don't understand your idea yet, let me confirm it.
>> >
>> > Certainly, current acpi driver does not need these definitions.
>> > But, I think nfit_test.ko will need them to emulate these features.
>> >
>> > Do you intend that libndctl-nfit.h should be defined at "include/uapi/linux/"
>> > directory?
>> > Otherwise, it should be defined at "tools/testing/nvdimm/" or
>> > "tools/testing/nvdimm/test" ?
>>
>> nfit_test will need its own internal / private copy of these payloads
>> in tools/testing/nvdimm/test so it can emulate how the bios behaves.
>> The include/uapi/linux directory is for user to kernel interface
>> definitions and these command payloads are purely an interface to bios
>> / firmware.
>>
>
> Hmm,
>
> I'm trying to make libndctl-nfit.h under tools/testing/nvdimm/nfit/,
> and trying to move NFIT_CMD_TRANSLATE_SPA from drivers/acpi/nfit/core.c to it.
>
> But I noticed that drivers/acpi/nfit/core.c still uses the definition
> in acpi_nfit_init_dsms().
>
> ----
> static void acpi_nfit_init_dsms(struct acpi_nfit_desc *acpi_desc)
> {
>         struct nvdimm_bus_descriptor *nd_desc = &acpi_desc->nd_desc;
>         const guid_t *guid = to_nfit_uuid(NFIT_DEV_BUS);
>
>         :
>
>         dsm_mask =
>                 (1 << ND_CMD_ARS_CAP) |
>                 (1 << ND_CMD_ARS_START) |
>                 (1 << ND_CMD_ARS_STATUS) |
>                 (1 << ND_CMD_CLEAR_ERROR) |
>                 (1 << NFIT_CMD_TRANSLATE_SPA) |
>                 (1 << NFIT_CMD_ARS_INJECT_SET) |
>                 (1 << NFIT_CMD_ARS_INJECT_CLEAR) |
>                 (1 << NFIT_CMD_ARS_INJECT_GET);
>         for_each_set_bit(i, &dsm_mask, BITS_PER_LONG)
>                 if (acpi_check_dsm(adev->handle, guid, 1, 1ULL << i))
>                         set_bit(i, &nd_desc->bus_dsm_mask);
> ---
>
> Though it seems to be necessary yet,
> do you have any idea to remove it?

Here's the end state I'm looking for...

1/ drivers/acpi/nfit/core.c keeps its current definitions of the
NFIT_CMD_ numbers since it needs to validate the DSM command number
mask

2/ tools/testing/nvdimm/test/nfit.h grows private definitions of only
the command payloads that it emulates

3/ ndctl in userspace grows a new ndctl/libndctl-nfit.h with all the
payloads defined in ACPI. Ideally we can deprecate all the payload
definitions in ndctl.h and only use the ND_CMD_CALL ioctl plus the
ndctl/libndctl-nfit.h definitions. Although, since label reading has
moved away from DSMs we'll always need the separate
ND_CMD_{GET,SET}_CONFIG_DATA ioctls. The libndctl-nfit.h header should
be listed in pkginclude_HEADERS.

Powered by blists - more mailing lists