On Mon, Feb 1, 2021 at 11:13 AM Ben Widawsky <ben.widawsky(a)intel.com> wrote:
On 21-02-01 12:54:00, Konrad Rzeszutek Wilk wrote:
> > +#define cxl_doorbell_busy(cxlm)
\
> > + (cxl_read_mbox_reg32(cxlm, CXLDEV_MB_CTRL_OFFSET) &
\
> > + CXLDEV_MB_CTRL_DOORBELL)
> > +
> > +#define CXL_MAILBOX_TIMEOUT_US 2000
>
> You been using the spec for the values. Is that number also from it ?
>
Yes it is. I'll add a comment with the spec reference.
> > +
> > +enum opcode {
> > + CXL_MBOX_OP_IDENTIFY = 0x4000,
> > + CXL_MBOX_OP_MAX = 0x10000
> > +};
> > +
> > +/**
> > + * struct mbox_cmd - A command to be submitted to hardware.
> > + * @opcode: (input) The command set and command submitted to hardware.
> > + * @payload_in: (input) Pointer to the input payload.
> > + * @payload_out: (output) Pointer to the output payload. Must be allocated by
> > + * the caller.
> > + * @size_in: (input) Number of bytes to load from @payload.
> > + * @size_out: (output) Number of bytes loaded into @payload.
> > + * @return_code: (output) Error code returned from hardware.
> > + *
> > + * This is the primary mechanism used to send commands to the hardware.
> > + * All the fields except @payload_* correspond exactly to the fields described
in
> > + * Command Register section of the CXL 2.0 spec (8.2.8.4.5). @payload_in and
> > + * @payload_out are written to, and read from the Command Payload Registers
> > + * defined in (8.2.8.4.8).
> > + */
> > +struct mbox_cmd {
> > + u16 opcode;
> > + void *payload_in;
> > + void *payload_out;
>
> On a 32-bit OS (not that we use those that more, but lets assume
> someone really wants to), the void is 4-bytes, while on 64-bit it is
> 8-bytes.
>
> `pahole` is your friend as I think there is a gap between opcode and
> payload_in in the structure.
>
> > + size_t size_in;
> > + size_t size_out;
>
> And those can also change depending on 32-bit/64-bit.
>
> > + u16 return_code;
> > +#define CXL_MBOX_SUCCESS 0
> > +};
>
> Do you want to use __packed to match with the spec?
>
> Ah, reading later you don't care about it.
>
> In that case may I recommend you move 'return_code' (or perhaps just
> call it rc?) to be right after opcode? Less of gaps in that structure.
>
I guess I hadn't realized we're supposed to try to fully pack structs by
default.
This is just the internal parsed context of a command, I can't imagine
packing is relevant here. pahole optimization feels premature as well.
> > +
> > +static int cxl_mem_wait_for_doorbell(struct cxl_mem *cxlm)
> > +{
> > + const int timeout = msecs_to_jiffies(CXL_MAILBOX_TIMEOUT_US);
> > + const unsigned long start = jiffies;
> > + unsigned long end = start;
> > +
> > + while (cxl_doorbell_busy(cxlm)) {
> > + end = jiffies;
> > +
> > + if (time_after(end, start + timeout)) {
> > + /* Check again in case preempted before timeout test */
> > + if (!cxl_doorbell_busy(cxlm))
> > + break;
> > + return -ETIMEDOUT;
> > + }
> > + cpu_relax();
> > + }
>
> Hm, that is not very scheduler friendly. I mean we are sitting here for
> 2000us (2 ms) - that is quite the amount of time spinning.
>
> Should this perhaps be put in a workqueue?
So let me first point you to the friendlier version which was shot down:
https://lore.kernel.org/linux-cxl/20201111054356.793390-8-ben.widawsky@in...
I'm not opposed to this being moved to a workqueue at some point, but I think
that's unnecessary complexity currently. The reality is that it's expected that
commands will finish way sooner than this or be implemented as background
commands. I've heard a person who makes a lot of the spec decisions say, "if
it's 2 seconds, nobody will use these things".
That said, asynchronous probe needs to be enabled for the next driver update.