Re: [RFC] ACPI: PCC: Support shared interrupt for multiple subspaces

From: lihuisong (C)
Date: Mon Nov 21 2022 - 22:42:19 EST



在 2022/11/22 0:59, Robbie King 写道:
On 11/19/2022 2:32 AM, lihuisong (C) wrote:
在 2022/11/18 2:09, Robbie King 写道:
On 11/7/2022 1:24 AM, lihuisong (C) wrote:
在 2022/11/4 23:39, Robbie King 写道:
On 11/4/2022 11:15 AM, Sudeep Holla wrote:
On Fri, Nov 04, 2022 at 11:04:22AM -0400, Robbie King wrote:
Hello Huisong, your raising of the shared interrupt issue is very timely, I
am working to implement "Extended PCC subspaces (types 3 and 4)" using PCC
on the ARM RDN2 reference platform as a proof of concept, and encountered
this issue as well.  FWIW, I am currently testing using Sudeep's patch with
the "chan_in_use" flag removed, and so far have not encountered any issues.

Interesting, do you mean the patch I post in this thread but without the
whole chan_in_use flag ?
That's right, diff I'm running with is attached to end of message.
Hello Robbie, In multiple subspaces scenario, there is a problem
that OS doesn't know which channel should respond to the interrupt
if no this chan_in_use flag. If you have not not encountered any
issues in this case, it may be related to your register settings.

Hi Huisong, apologies, I see your point now concerning multiple subspaces.

I have started stress testing where I continuously generate both requests
and notifications as quickly as possible, and unfortunately found an issue
even with the original chan_in_use patch.  I first had to modify the patch
to get the type 4 channel notifications to function at all, essentially
ignoring the chan_in_use flag for that channel.  With that change, I still
hit my original stress issue, where the pcc_mbox_irq function did not
correctly ignore an interrupt for the type 3 channel.

The issue occurs when a request from AP to SCP over the type 3 channel is
outstanding, and simultaneously the SCP initiates a notification over the
type 4 channel.  Since the two channels share an interrupt, both handlers
are invoked.

I've tried to draw out the state of the channel status "free" bits along
with the AP and SCP function calls involved.

type 3
------

 (1)pcc.c:pcc_send_data()
       |                         (5) mailbox.c:mbox_chan_receive_data()
_______v                      (4)pcc.c:pcc_mbox_irq()
free   \_________________________________________

                              ^
type 4                        ^
------                        ^
_____________________
free                 \_____________________________
                     ^        ^
                     |        |
(2)mod_smt.c:smt_transmit()   |
                              |
(3)mod_mhu2.c:raise_interrupt()

The sequence of events are:

1) OS initiates request to SCP by clearing FREE in status and ringing SCP doorbell
2) SCP initiates notification by filling shared memory and clearing FREE in status
3) SCP notifies OS by ringing OS doorbell
4) OS first invokes interrupt handler for type 3 channel

   At this step, the issue is that "val" from reading status (i.e. CommandCompleteCheck)
   is zero (SCP has not responded yet) so the code below falls through and continues
   to processes the interrupt as if the request has been acknowledged by the SCP.

    if (val) { /* Ensure GAS exists and value is non-zero */
        val &= pchan->cmd_complete.status_mask;
        if (!val)
            return IRQ_NONE;
    }

   The chan_in_use flag does not address this because the channel is indeed in use.

5) ACPI:PCC client kernel module is incorrectly notified that response data is
   available
Indeed, chan_in_use flag is invalid for type4.
Thinking about this some more, I believe there is a need for the chan_in_use flag
for the type 4 channels. If there are multiple SCP to AP channels sharing an
interrupt, and the PCC client code chooses to acknowledge the transfer from
process level (i.e. call mbox_send outside of the mbox_chan_received_data callback),
then I believe a window exists where the callback could be invoked twice for the
same SCP to AP channel. I've attached a diff.
I don't understand your concern. type4 need to set command complete bit and
ring doorbell after processing mbox_chan_received_data callback. I think it
is ok without chan_in_use flag.

Here's what I think:
For type4, set the command complete bit to 1 by default in platform.
Clear the command complete when do platform notification.
If a given channel detects the command complete bit is 0, it should respond
the interrupt, otherwise there is nothing to do.

I put all points we discussed into the RFC V2 patch. Let's go to V2 to discuss.

I added the following fix (applied on top of Sudeep's original patch for clarity)
for the issue above which solved the stress test issue.  I've changed the interrupt
handler to explicitly verify that the status value matches the mask for type 3
interrupts before acknowledging them.  Conversely, a type 4 channel verifies that
the status value does *not* match the mask, since in this case we are functioning
as the recipient, not the initiator.

One concern is that since this fundamentally changes handling of the channel status,
that existing platforms could be impacted.
[snip]
+    /*
+     * When we control data flow on the channel, we expect
+     * to see the mask bit(s) set by the remote to indicate
+     * the presence of a valid response.  When we do not
+     * control the flow (i.e. type 4) the opposite is true.
+     */
+    if (pchan->is_controller)
+        cmp = pchan->cmd_complete.status_mask;
+    else
+        cmp = 0;
+
+    val &= pchan->cmd_complete.status_mask;
+    if (cmp != val)
+        return IRQ_NONE;

We don't have to use the pchan->cmd_complete.status_mask as above.

For the communication from AP to SCP, if this channel is in use, command
complete bit is 1 indicates that the command being executed has been
completed.
For the communication from SCP to AP, if this channel is in use, command
complete bit is 0 indicates that the bit has been cleared and OS should
response the interrupt.

So you concern should be gone if we do as following approach:
"
val &= pchan->cmd_complete.status_mask;
need_rsp_irq = pchan->is_controller ? val != 0 : val == 0;
if (!need_rsp_irq)
    return IRQ_NONE
"

This may depend on the default value of the command complete register
for each channel(must be 1, means that the channel is free for use).
It is ok for type3 because of chan_in_use flag, while something needs
to do in platform or OS for type4.
ret = pcc_chan_reg_read(&pchan->error, &val);
     if (ret)
@@ -704,6 +717,9 @@ static int pcc_mbox_probe(struct platform_device *pdev)
         pcc_mbox_channels[i].con_priv = pchan;
         pchan->chan.mchan = &pcc_mbox_channels[i];

+        pchan->is_controller =
+            (pcct_entry->type != ACPI_PCCT_TYPE_EXT_PCC_SLAVE_SUBSPACE);
+
This definition does not apply to all types because type1 and type2
support bidirectional communication.

if (pcct_entry->type == ACPI_PCCT_TYPE_EXT_PCC_SLAVE_SUBSPACE &&
             !pcc_mbox_ctrl->txdone_irq) {
             pr_err("Plaform Interrupt flag must be set to 1");

I put all points we discussed into the following modifcation.
Robbie, can you try it again for type 4 and see what else needs to be
done?

Regards,
Huisong

Thanks Huisong, I ran my current stress test scenario against your diff
with no issues (I did have to manually merge due to a tabs to spaces issue
which may be totally on my end, still investigating).

Here is the proposed change to support chan_in_use for type 4 (which I've
also successfully tested with). I think I have solved the tabs to spaces
issue for my sent messages, apologies if that's not the case.

diff --git a/drivers/mailbox/pcc.c b/drivers/mailbox/pcc.c
index 057e00ee83c8..d4fcc913a9a8 100644
--- a/drivers/mailbox/pcc.c
+++ b/drivers/mailbox/pcc.c
@@ -292,7 +292,7 @@ static irqreturn_t pcc_mbox_irq(int irq, void *p)
int ret;

pchan = chan->con_priv;
- if (pchan->mesg_dir == PCC_ONLY_AP_TO_SCP && !pchan->chan_in_use)
+ if (!pchan->chan_in_use)
return IRQ_NONE;

ret = pcc_chan_reg_read(&pchan->cmd_complete, &val);
@@ -320,8 +320,16 @@ static irqreturn_t pcc_mbox_irq(int irq, void *p)
goto out;
}

+ /*
+ * Clearing in_use before RX callback allows calls to mbox_send
+ * (which sets in_use) within the callback so SCP to AP channels
+ * can acknowledge transfer within IRQ context
+ */
+ if (pchan->cmd_complete.gas)
+ pchan->chan_in_use = false;
+
mbox_chan_received_data(chan, NULL);
- rc = IRQ_HANDLED;
+ return IRQ_HANDLED;

out:
if (pchan->cmd_complete.gas)
@@ -772,6 +780,8 @@ static int pcc_mbox_probe(struct platform_device *pdev)
goto err;
}
pcc_set_chan_mesg_dir(pchan, pcct_entry->type);
+ if (pchan->mesg_dir == PCC_ONLY_SCP_TO_AP)
+ pchan->chan_in_use = true;

if (pcc_mbox_ctrl->txdone_irq) {
rc = pcc_parse_subspace_irq(pchan, pcct_entry);

.