[PATCH v2 5/6] mailbox: Add generic mechanism for testing Mailbox Controllers

[Jassi Brar]

On Thu, Aug 13, 2015 at 5:10 PM, Lee Jones <lee.jones at linaro.org> wrote:
> On Thu, 13 Aug 2015, Jassi Brar wrote:
>
>> On Thu, Aug 13, 2015 at 4:30 PM, Lee Jones <lee.jones at linaro.org> wrote:
>> > On Thu, 13 Aug 2015, Jassi Brar wrote:
>> >
>> >> On Thu, Aug 13, 2015 at 3:53 PM, Lee Jones <lee.jones at linaro.org> wrote:
>> >> > On Thu, 13 Aug 2015, Jassi Brar wrote:
>> >> >
>> >> >> On Thu, Aug 13, 2015 at 2:49 PM, Lee Jones <lee.jones at linaro.org> wrote:
>> >> >> > On Thu, 13 Aug 2015, Jassi Brar wrote:
>> >> >>
>> >> >> >> >> > +
>> >> >> >> >> > +static void mbox_test_prepare_message(struct mbox_client *client, void *message)
>> >> >> >> >> > +{
>> >> >> >> >> > +       struct mbox_test_device *tdev = dev_get_drvdata(client->dev);
>> >> >> >> >> > +
>> >> >> >> >> > +       if (tdev->mmio)
>> >> >> >> >> > +               memcpy(tdev->mmio, message, MBOX_MAX_MSG_LEN);
>> >> >> >> >> >
>> >> >> >> >> This is unlikely to work. Those platforms that need to send a 2 part
>> >> >> >> >> message, they do :
>> >> >> >> >> (a) Signal/Command/Target code via some controller register (via
>> >> >> >> >> mbox_send_message).
>> >> >> >> >> (b) Setup the payload in Shared-Memory (via tx_prepare).
>> >> >> >> >>
>> >> >> >> >> This test driver assumes both are the same. I think you have to
>> >> >> >> >> declare something like
>> >> >> >> >
>> >> >> >> > This driver assumes that the framework will call client->tx_prepare()
>> >> >> >> > first, which satisfies (b).  It then assumes controller->send_data()
>> >> >> >> > will be invoked, which will send the platform specific
>> >> >> >> > signal/command/target code, which then satisfies (a).
>> >> >> >> >
>> >> >> >> Yeah, but what would be that code? Who provides that?
>> >> >> >>
>> >> >> >> > In what way does it assume they are the same?
>> >> >> >> >
>> >> >> >> notice the 'message' pointer in
>> >> >> >> mbox_send_message(tdev->tx_channel, message);
>> >> >> >>     and
>> >> >> >> memcpy(tdev->mmio, message, MBOX_MAX_MSG_LEN);
>> >> >> >>
>> >> >> >> >> struct mbox_test_message { /* same for TX and RX */
>> >> >> >> >>           unsigned sig_len;
>> >> >> >> >>           void *signal;               /* rx/tx via mailbox api */
>> >> >> >> >>           unsigned pl_len;
>> >> >> >> >>           void *payload;           /* rx/tx via shared-memory */
>> >> >> >> >> };
>> >> >> >> >
>> >> >> >> > How do you think this should be set and use these?
>> >> >> >> >
>> >> >> >> The userspace would want to specify the command code (32bits or not)
>> >> >> >> that would be passed via the fifo/register of the controller. So we
>> >> >> >> need signal[]
>> >> >> >>
>> >> >> >> The data to be passed via share-memory could be provided by userspace
>> >> >> >> via the payload[] array.
>> >> >> >
>> >> >> > Okay, so would the solution be two userspace files 'signal' and
>> >> >> > 'message', so in the case of a client specified signal we can write it
>> >> >> > into there first.
>> >> >> >
>> >> >> > echo 255  > signal
>> >> >> > echo test > message
>> >> >> >
>> >> >> > ... or in the case where no signal is required, or the controller
>> >> >> > driver taking care of it, we just don't write anything to signal?
>> >> >> >
>> >> >> file_operations.write() should parse signal and message, coming from
>> >> >> userspace, into a local structure before routing them via
>> >> >> mbox_send_message and tx_prepare respectively.
>> >> >
>> >> > Okay.  So before I code this up we should agree on syntax.
>> >> >
>> >> > How would you like to represent the break between signal and message?
>> >> > Obviously ' ' would be a bad idea, as some clients may want to send
>> >> > messages with white space contained.  How about '\t' or '\n'?
>> >> >
>> >> Yeah, I am not a fan of markers and flags either.
>> >>
>> >> Maybe we should share with userspace
>> >>   struct mbox_test_message { /* same for TX and RX */
>> >>            unsigned sig_len;
>> >>            void __user *signal;        /* rx/tx via mailbox api */
>> >>            unsigned pl_len;
>> >>            void __user *payload;    /* rx/tx via shared-memory */
>> >>   };
>> >>
>> >> First copy_from_user the structure of length sizof(struct
>> >> mbox_test_message) and then copy_from_user length sig_len and pl_len
>> >> from signal[] and payload[] respectively (usually ioctls would carry
>> >> such data).
>> >
>> > Simplicity and ease of use should be the goals here.  Testers should
>> > not have to write applications to use this driver.  Can we come up
>> > with a simple/effective method that uses SYSFS/DEBUGFS please?
>> >
>> > The easiest way I can think of which avoids markers/separators AND the
>> > requirement for users to have to write applications is to have two
>> > files, 'signal' and 'message' as mentioned before.  Once both are
>> > populated I can get this driver to draft the message appropriately and
>> > fire it off.
>> >
>> And then write to more files for RX data? ... which should also be in
>> the form of 'signal' and 'message'.
>>
>> BTW like for spi there is a stock application in
>> Documentation/spi/spidev_test.c we could do the same?
>
> Coming from personal experience, testing drivers with (even stock)
> applications is much more painful that simply writing/reading
> (cat/echo) to a file in SYSFS/DEBUGFS.  Particularly if people are
> using initramfs or thelike.  If it is possible to use SYSFS/DEBUGFS,
> which it is in this case, then I believe that's the route we could go
> down.
>
Well, where could sysfs/debugfs not be used? :)  Anyways I am ok if
prefer debugfs.

> In answer to your question; we only need those two files.  The reply
> can be placed back into 'message' and can be read from there.
>
Testing shouldn't be restricted to 'send command and receive reply'.
What if Linux only listens to broadcasts from the remote? Who knows
someone might want to (ab)use this test client to implement userspace
handler of remote commands?
So please see RX to be independent of TX.

Thanks.