Re: [PATCH V2 02/10] mailbox: tegra-hsp: Add HSP(Hardware Synchronization Primitives) driver

From: Joseph Lo
Date: Thu Jul 07 2016 - 02:37:45 EST


On 07/06/2016 08:23 PM, Alexandre Courbot wrote:
On Wed, Jul 6, 2016 at 6:06 PM, Joseph Lo <josephl@xxxxxxxxxx> wrote:
On 07/06/2016 03:05 PM, Alexandre Courbot wrote:

On Tue, Jul 5, 2016 at 6:04 PM, Joseph Lo <josephl@xxxxxxxxxx> wrote:

The Tegra HSP mailbox driver implements the signaling doorbell-based
interprocessor communication (IPC) for remote processors currently. The
HSP HW modules support some different features for that, which are
shared mailboxes, shared semaphores, arbitrated semaphores, and
doorbells. And there are multiple HSP HW instances on the chip. So the
driver is extendable to support more features for different IPC
requirement.

The driver of remote processor can use it as a mailbox client and deal
with the IPC protocol to synchronize the data communications.

Signed-off-by: Joseph Lo <josephl@xxxxxxxxxx>
---
Changes in V2:
- Update the driver to support the binding changes in V2
- it's extendable to support multiple HSP sub-modules on the same HSP HW
block
now.
---
drivers/mailbox/Kconfig | 9 +
drivers/mailbox/Makefile | 2 +
drivers/mailbox/tegra-hsp.c | 418
++++++++++++++++++++++++++++++++++++++++++++
3 files changed, 429 insertions(+)
create mode 100644 drivers/mailbox/tegra-hsp.c

diff --git a/drivers/mailbox/Kconfig b/drivers/mailbox/Kconfig
index 5305923752d2..fe584cb54720 100644
--- a/drivers/mailbox/Kconfig
+++ b/drivers/mailbox/Kconfig
@@ -114,6 +114,15 @@ config MAILBOX_TEST
Test client to help with testing new Controller driver
implementations.

+config TEGRA_HSP_MBOX
+ bool "Tegra HSP(Hardware Synchronization Primitives) Driver"


Space missing before the opening parenthesis (same in the patch title
btw).

Okay.


+ depends on ARCH_TEGRA_186_SOC
+ help
+ The Tegra HSP driver is used for the interprocessor
communication
+ between different remote processors and host processors on
Tegra186
+ and later SoCs. Say Y here if you want to have this support.
+ If unsure say N.


Since this option is selected automatically by ARCH_TEGRA_186_SOC, you
should probably drop the last 2 sentences.

Okay.


+
config XGENE_SLIMPRO_MBOX
tristate "APM SoC X-Gene SLIMpro Mailbox Controller"
depends on ARCH_XGENE
diff --git a/drivers/mailbox/Makefile b/drivers/mailbox/Makefile
index 0be3e742bb7d..26d8f91c7fea 100644
--- a/drivers/mailbox/Makefile
+++ b/drivers/mailbox/Makefile
@@ -25,3 +25,5 @@ obj-$(CONFIG_TI_MESSAGE_MANAGER) += ti-msgmgr.o
obj-$(CONFIG_XGENE_SLIMPRO_MBOX) += mailbox-xgene-slimpro.o

obj-$(CONFIG_HI6220_MBOX) += hi6220-mailbox.o
+
+obj-${CONFIG_TEGRA_HSP_MBOX} += tegra-hsp.o
diff --git a/drivers/mailbox/tegra-hsp.c b/drivers/mailbox/tegra-hsp.c
new file mode 100644
index 000000000000..93c3ef58f29f
--- /dev/null
+++ b/drivers/mailbox/tegra-hsp.c
@@ -0,0 +1,418 @@
+/*
+ * Copyright (c) 2016, NVIDIA CORPORATION. All rights reserved.
+ *
+ * This program is free software; you can redistribute it and/or modify
it
+ * under the terms and conditions of the GNU General Public License,
+ * version 2, as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope it will be useful, but
WITHOUT
+ * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
+ * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
for
+ * more details.
+ */
+
+#include <linux/interrupt.h>
+#include <linux/io.h>
+#include <linux/mailbox_controller.h>
+#include <linux/of.h>
+#include <linux/of_device.h>
+#include <linux/platform_device.h>
+#include <dt-bindings/mailbox/tegra186-hsp.h>
+
+#define HSP_INT_DIMENSIONING 0x380
+#define HSP_nSM_OFFSET 0
+#define HSP_nSS_OFFSET 4
+#define HSP_nAS_OFFSET 8
+#define HSP_nDB_OFFSET 12
+#define HSP_nSI_OFFSET 16


Would be nice to have comments to understand what SM, SS, AS, etc.
stand for (Shared Mailboxes, Shared Semaphores, Arbitrated Semaphores
but you need to look at the patch description to understand that). A
top-of-file comment explaning the necessary concepts to read this code
would do the trick.

Yes, will fix that.


+#define HSP_nINT_MASK 0xf
+
+#define HSP_DB_REG_TRIGGER 0x0
+#define HSP_DB_REG_ENABLE 0x4
+#define HSP_DB_REG_RAW 0x8
+#define HSP_DB_REG_PENDING 0xc
+
+#define HSP_DB_CCPLEX 1
+#define HSP_DB_BPMP 3


Maybe turn this into enum and use that type for
tegra_hsp_db_chan::db_id? Also have MAX_NUM_HSP_DB here, since it is
related to these values?

Okay.


+
+#define MAX_NUM_HSP_CHAN 32
+#define MAX_NUM_HSP_DB 7
+
+#define hsp_db_offset(i, d) \
+ (d->base + ((1 + (d->nr_sm >> 1) + d->nr_ss + d->nr_as) << 16) +
\
+ (i) * 0x100)
+
+struct tegra_hsp_db_chan {
+ int master_id;
+ int db_id;
+};
+
+struct tegra_hsp_mbox_chan {
+ int type;
+ union {
+ struct tegra_hsp_db_chan db_chan;
+ };
+};
+
+struct tegra_hsp_mbox {
+ struct mbox_controller *mbox;
+ void __iomem *base;
+ void __iomem *db_base[MAX_NUM_HSP_DB];
+ int db_irq;
+ int nr_sm;
+ int nr_as;
+ int nr_ss;
+ int nr_db;
+ int nr_si;
+ spinlock_t lock;
+};
+
+static inline u32 hsp_readl(void __iomem *base, int reg)
+{
+ return readl(base + reg);
+}
+
+static inline void hsp_writel(void __iomem *base, int reg, u32 val)
+{
+ writel(val, base + reg);
+ readl(base + reg);
+}
+
+static int hsp_db_can_ring(void __iomem *db_base)
+{
+ u32 reg;
+
+ reg = hsp_readl(db_base, HSP_DB_REG_ENABLE);
+
+ return !!(reg & BIT(HSP_DB_MASTER_CCPLEX));
+}
+
+static irqreturn_t hsp_db_irq(int irq, void *p)
+{
+ struct tegra_hsp_mbox *hsp_mbox = p;
+ ulong val;
+ int master_id;
+
+ val = (ulong)hsp_readl(hsp_mbox->db_base[HSP_DB_CCPLEX],
+ HSP_DB_REG_PENDING);
+ hsp_writel(hsp_mbox->db_base[HSP_DB_CCPLEX], HSP_DB_REG_PENDING,
val);
+
+ spin_lock(&hsp_mbox->lock);
+ for_each_set_bit(master_id, &val, MAX_NUM_HSP_CHAN) {
+ struct mbox_chan *chan;
+ struct tegra_hsp_mbox_chan *mchan;
+ int i;
+
+ for (i = 0; i < MAX_NUM_HSP_CHAN; i++) {


I wonder if this could not be optimized. You are doing a double loop
on MAX_NUM_HSP_CHAN to look for an identical master_id. Since it seems
like the same master_id cannot be used twice (considering that the
inner loop only processes the first match), couldn't you just select
the free channel in of_hsp_mbox_xlate() by doing
&mbox->chans[master_id] (and returning an error if it is already
used), then simply getting chan as &hsp_mbox->mbox->chans[master_id]
instead of having the inner loop below? That would remove the need for
the second loop.


That was exactly what I did in the V1, which only supported one HSP
sub-module per HSP HW block. So we can just use the master_id as the mbox
channel ID.

Meanwhile, the V2 is purposed to support multiple HSP sub-modules to be
running on the same HSP HW block. The "ID" between different modules could
be conflict. So I dropped the mechanism that used the master_id as the mbox
channel ID.

Instead, the channel is allocated at the time, when the client is bound to
one of the HSP sub-modules. And we store the "ID" information into the
private mbox channel data, which can help us to figure out which mbox
channel should response to the interrupt.

In the doorbell case, because all the DB clients are shared the same DB IRQ
at the CPU side. So in the ISR, we need to figure out the IRQ source, which
is the master_id that the IRQ came from. This is the outer loop. The inner
loop, we figure out which channel should response to by checking the type
and ID.

And I think it should be pretty quick, because we only check the set bit
from the pending register. And finding the matching channel.

Yeah, I am not worried about the CPU time (although in interrupt
context, we always should), but rather about whether the code could be
simplified.

Ah, I think I get it. You want to be able to receive interrupts from
the same master, but not necessarily for the doorbell function.
Because of this you cannot use master_id as the index for the channel.
Am I understanding correctly?

Yes, the DB clients trigger the IRQ through the same master (HSP_DB_CCPLEX) with it's master_id. We (CPU) can check the ID to know who is requesting the HSP mbox service. Each ID is unique under the DB module.

But the ID could be conflict when the HSP mbox driver are working with multiple HSP sub-function under the same HSP HW block. So we can't just match the ID to the HSP mbox channel ID.




If having two channels use the same master_id is a valid scenario,
then all matches on master_id should probably be processed, not just
the first one.

Each DB channel should have different master_id.



+ chan = &hsp_mbox->mbox->chans[i];
+
+ if (!chan->con_priv)
+ continue;
+
+ mchan = chan->con_priv;
+ if (mchan->type == HSP_MBOX_TYPE_DB &&
+ mchan->db_chan.master_id == master_id)
+ break;
+ chan = NULL;
+ }
+
+ if (chan)
+ mbox_chan_received_data(chan, NULL);
+ }
+ spin_unlock(&hsp_mbox->lock);
+
+ return IRQ_HANDLED;
+}
+
+static int hsp_db_send_data(struct mbox_chan *chan, void *data)
+{
+ struct tegra_hsp_mbox_chan *mchan = chan->con_priv;
+ struct tegra_hsp_db_chan *db_chan = &mchan->db_chan;
+ struct tegra_hsp_mbox *hsp_mbox =
dev_get_drvdata(chan->mbox->dev);
+
+ hsp_writel(hsp_mbox->db_base[db_chan->db_id], HSP_DB_REG_TRIGGER,
1);
+
+ return 0;
+}
+
+static int hsp_db_startup(struct mbox_chan *chan)
+{
+ struct tegra_hsp_mbox_chan *mchan = chan->con_priv;
+ struct tegra_hsp_db_chan *db_chan = &mchan->db_chan;
+ struct tegra_hsp_mbox *hsp_mbox =
dev_get_drvdata(chan->mbox->dev);
+ u32 val;
+ unsigned long flag;
+
+ if (db_chan->master_id >= MAX_NUM_HSP_CHAN) {
+ dev_err(chan->mbox->dev, "invalid HSP chan: master ID:
%d\n",
+ db_chan->master_id);
+ return -EINVAL;
+ }
+
+ spin_lock_irqsave(&hsp_mbox->lock, flag);
+ val = hsp_readl(hsp_mbox->db_base[HSP_DB_CCPLEX],
HSP_DB_REG_ENABLE);
+ val |= BIT(db_chan->master_id);
+ hsp_writel(hsp_mbox->db_base[HSP_DB_CCPLEX], HSP_DB_REG_ENABLE,
val);
+ spin_unlock_irqrestore(&hsp_mbox->lock, flag);
+
+ if (!hsp_db_can_ring(hsp_mbox->db_base[db_chan->db_id]))
+ return -ENODEV;
+
+ return 0;
+}
+
+static void hsp_db_shutdown(struct mbox_chan *chan)
+{
+ struct tegra_hsp_mbox_chan *mchan = chan->con_priv;
+ struct tegra_hsp_db_chan *db_chan = &mchan->db_chan;
+ struct tegra_hsp_mbox *hsp_mbox =
dev_get_drvdata(chan->mbox->dev);
+ u32 val;
+ unsigned long flag;
+
+ spin_lock_irqsave(&hsp_mbox->lock, flag);
+ val = hsp_readl(hsp_mbox->db_base[HSP_DB_CCPLEX],
HSP_DB_REG_ENABLE);
+ val &= ~BIT(db_chan->master_id);
+ hsp_writel(hsp_mbox->db_base[HSP_DB_CCPLEX], HSP_DB_REG_ENABLE,
val);
+ spin_unlock_irqrestore(&hsp_mbox->lock, flag);
+}
+
+static bool hsp_db_last_tx_done(struct mbox_chan *chan)
+{
+ return true;
+}
+
+static int tegra_hsp_db_init(struct tegra_hsp_mbox *hsp_mbox,
+ struct mbox_chan *mchan, int master_id)
+{
+ struct platform_device *pdev =
to_platform_device(hsp_mbox->mbox->dev);
+ struct tegra_hsp_mbox_chan *hsp_mbox_chan;
+ int ret;
+
+ if (!hsp_mbox->db_irq) {
+ int i;
+
+ hsp_mbox->db_irq = platform_get_irq_byname(pdev,
"doorbell");


Getting the IRQ sounds more like a job for probe() - I don't see the
benefit of lazy-doing it?


We only need the IRQ when the client is requesting the DB service. For other
HSP sub-modules, they are using different IRQ. So I didn't do that at probe
time.

Ok, but probe() is where resources should be acquired... and at the
very least DT properties be looked up. In this case there is no hard
requirement for doing it elsewhere.

Is this interrupt absolutely required? Or can we tolerate to not use
the doorbell service? In the first case, the driver should fail during
probe(), not sometime later. In the second case, you should still get
all the interrupts in probe(), then disable them if they are not
needed, and check in this function whether db_irq is a valid interrupt
number to decide whether or not we can use doorbell.

Ah, I understand your concern now. It should be ok to move to probe(). Will fix that.




+ ret = devm_request_irq(&pdev->dev, hsp_mbox->db_irq,
+ hsp_db_irq, IRQF_NO_SUSPEND,
+ dev_name(&pdev->dev), hsp_mbox);
+ if (ret)
+ return ret;
+
+ for (i = 0; i < MAX_NUM_HSP_DB; i++)
+ hsp_mbox->db_base[i] = hsp_db_offset(i,
hsp_mbox);


Same here, cannot this be moved into probe()?

Same as above, only needed when the client requests it.

But you don't waste any resources by doing it preemptively in probe().
So let's keep related code in the same place.

Okay.

Thanks,
-Joseph