Re: [PATCH v10 07/26] mailbox: Add Gunyah message queue mailbox

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 





On 23/02/2023 23:15, Elliot Berman wrote:


On 2/23/2023 2:25 AM, Srinivas Kandagatla wrote:


On 23/02/2023 00:15, Elliot Berman wrote:


On 2/20/2023 5:59 AM, Srinivas Kandagatla wrote:


On 14/02/2023 21:23, Elliot Berman wrote:
Gunyah message queues are a unidirectional inter-VM pipe for messages up to 1024 bytes. This driver supports pairing a receiver message queue and
a transmitter message queue to expose a single mailbox channel.

Signed-off-by: Elliot Berman <quic_eberman@xxxxxxxxxxx>
---
  Documentation/virt/gunyah/message-queue.rst |   8 +
  drivers/mailbox/Makefile                   |   2 +
  drivers/mailbox/gunyah-msgq.c               | 214 ++++++++++++++++++++
  include/linux/gunyah.h                      |  56 +++++
  4 files changed, 280 insertions(+)
  create mode 100644 drivers/mailbox/gunyah-msgq.c

diff --git a/Documentation/virt/gunyah/message-queue.rst b/Documentation/virt/gunyah/message-queue.rst
index 0667b3eb1ff9..082085e981e0 100644
--- a/Documentation/virt/gunyah/message-queue.rst
+++ b/Documentation/virt/gunyah/message-queue.rst
@@ -59,3 +59,11 @@ vIRQ: two TX message queues will have two vIRQs (and two capability IDs).         |               |         |                 | |               |         |               |         |                 | |               |         +---------------+         +-----------------+ +---------------+
+
+Gunyah message queues are exposed as mailboxes. To create the mailbox, create +a mbox_client and call `gh_msgq_init`. On receipt of the RX_READY interrupt, +all messages in the RX message queue are read and pushed via the `rx_callback`
+of the registered mbox_client.
+
+.. kernel-doc:: drivers/mailbox/gunyah-msgq.c
+   :identifiers: gh_msgq_init
diff --git a/drivers/mailbox/Makefile b/drivers/mailbox/Makefile
index fc9376117111..5f929bb55e9a 100644
--- a/drivers/mailbox/Makefile
+++ b/drivers/mailbox/Makefile
@@ -55,6 +55,8 @@ obj-$(CONFIG_MTK_CMDQ_MBOX)    += mtk-cmdq-mailbox.o
  obj-$(CONFIG_ZYNQMP_IPI_MBOX)    += zynqmp-ipi-mailbox.o
+obj-$(CONFIG_GUNYAH)        += gunyah-msgq.o

Why are we reusing CONFIG_GUNYAH Kconfig symbol for mailbox, why not CONFIG_GUNYAH_MBOX?


There was some previous discussion about this:

https://lore.kernel.org/all/2a7bb5f2-1286-b661-659a-a5037150eae8@xxxxxxxxxxx/

+
  obj-$(CONFIG_SUN6I_MSGBOX)    += sun6i-msgbox.o
  obj-$(CONFIG_SPRD_MBOX)       += sprd-mailbox.o
diff --git a/drivers/mailbox/gunyah-msgq.c b/drivers/mailbox/gunyah-msgq.c
new file mode 100644
index 000000000000..03ffaa30ce9b
--- /dev/null
+++ b/drivers/mailbox/gunyah-msgq.c
@@ -0,0 +1,214 @@
+// SPDX-License-Identifier: GPL-2.0-only
+/*
+ * Copyright (c) 2022-2023 Qualcomm Innovation Center, Inc. All rights reserved.
+ */
+
+#include <linux/mailbox_controller.h>
+#include <linux/module.h>
+#include <linux/interrupt.h>
+#include <linux/gunyah.h>
+#include <linux/printk.h>
+#include <linux/init.h>
+#include <linux/slab.h>
+#include <linux/wait.h>

...

+/* Fired when message queue transitions from "full" to "space available" to send messages */
+static irqreturn_t gh_msgq_tx_irq_handler(int irq, void *data)
+{
+    struct gh_msgq *msgq = data;
+
+    mbox_chan_txdone(gh_msgq_chan(msgq), 0);
+
+    return IRQ_HANDLED;
+}
+
+/* Fired after sending message and hypercall told us there was more space available. */
+static void gh_msgq_txdone_tasklet(struct tasklet_struct *tasklet)

Tasklets have been long deprecated, consider using workqueues in this particular case.


Workqueues have higher latency and tasklets came as recommendation from Jassi. drivers/mailbox/imx-mailbox.c uses tasklets in the same way.

I did some quick unscientific measurements of ~1000x samples. The median latency for resource manager went from 25.5 us (tasklet) to 26 us (workqueue) (2% slower). The mean went from 28.7 us to 32.5 us (13% slower). Obviously, the outliers for workqueues were much more extreme.

TBH, this is expected because we are only testing resource manager, Note   the advantage that you will see shifting from tasket to workqueues is on overall system latencies and some drivers performance that need to react to events.

please take some time to read this nice article about this https://lwn.net/Articles/830964/


Hmm, this article is from 2020 and there was another effort in 2007. Neither seems to have succeeded. I'd like to stick to same mechanisms as other mailbox controllers.

I don't want to block this series because of this. We will have more opportunity to improve this once some system wide profiling is done.

AFAIU, In this system we will have atleast 2 tasklets between VM and RM and 2 per inter-vm, so if the number of tasklets increase in the system will be potentially spending more time in soft irq handling it.

At somepoint in time its good to get some profiling done using bcc/softirqs to see how much time is spent on softirqs.


--srini


Jassi, do you have any preferences?

Thanks,
Elliot





[Index of Archives]     [Linux ARM Kernel]     [Linux ARM]     [Linux Omap]     [Fedora ARM]     [Linux for Sparc]     [IETF Annouce]     [Security]     [Bugtraq]     [Linux MIPS]     [ECOS]     [Asterisk Internet PBX]     [Linux API]

  Powered by Linux