Re: [PATCH net-next 3/3] net: wwan: Allow WWAN drivers to provide blocking tx and poll function

On Wed, 16 Jun 2021 at 00:06, Stephan Gerhold [off-list ref] wrote:

On Tue, Jun 15, 2021 at 11:24:41PM +0200, Loic Poulain wrote:

quoted

On Tue, 15 Jun 2021 at 15:34, Stephan Gerhold [off-list ref] wrote:

quoted

At the moment, the WWAN core provides wwan_port_txon/off() to implement
blocking writes. The tx() port operation should not block, instead
wwan_port_txon/off() should be called when the TX queue is full or has
free space again.

However, in some cases it is not straightforward to make use of that
functionality. For example, the RPMSG API used by rpmsg_wwan_ctrl.c
does not provide any way to be notified when the TX queue has space
again. Instead, it only provides the following operations:

  - rpmsg_send(): blocking write (wait until there is space)
  - rpmsg_trysend(): non-blocking write (return error if no space)
  - rpmsg_poll(): set poll flags depending on TX queue state

Generally that's totally sufficient for implementing a char device,
but it does not fit well to the currently provided WWAN port ops.

Most of the time, using the non-blocking rpmsg_trysend() in the
WWAN tx() port operation works just fine. However, with high-frequent
writes to the char device it is possible to trigger a situation
where this causes issues. For example, consider the following
(somewhat unrealistic) example:

 # dd if=/dev/zero bs=1000 of=/dev/wwan0p2QMI
 dd: error writing '/dev/wwan0p2QMI': Resource temporarily unavailable
 1+0 records out

This fails immediately after writing the first record. It's likely
only a matter of time until this triggers issues for some real application
(e.g. ModemManager sending a lot of large QMI packets).

The rpmsg_char device does not have this problem, because it uses
rpmsg_trysend() and rpmsg_poll() to support non-blocking operations.
Make it possible to use the same in the RPMSG WWAN driver by extending
the tx() operation with a "nonblock" parameter and adding an optional
poll() callback. This integrates nicely with the RPMSG API and does
not break other WWAN drivers.

With these changes, the dd example above blocks instead of exiting
with an error.

Cc: Loic Poulain <redacted>
Signed-off-by: Stephan Gerhold <stephan@gerhold.net>
---
Note that rpmsg_poll() is an optional callback currently only implemented
by the qcom_smd RPMSG provider. However, it should be easy to implement
this for other RPMSG providers when needed.

Another potential solution suggested by Loic Poulain in [1] is to always
use the blocking rpmsg_send() from a workqueue/kthread and disable TX
until it is done. I think this could also work (perhaps a bit more
difficult to implement) but the main disadvantage is that I don't see
a way to return any kind of error to the client with this approach.
I assume we return immediately from the write() to the char device
after scheduling the rpmsg_send(), so we already reported success
when rpmsg_send() returns.

At the end all that matters to me is that it works properly, so I'm
open for any other suggestions. :)

[1]: https://lore.kernel.org/linux-arm-msm/CAMZdPi_-Qa=JnThHs_h-144dAfSAjF5s+QdBawdXZ3kk8Mx8ng@mail.gmail.com/ (local)
---
 drivers/net/wwan/iosm/iosm_ipc_port.c |  3 ++-
 drivers/net/wwan/mhi_wwan_ctrl.c      |  3 ++-
 drivers/net/wwan/rpmsg_wwan_ctrl.c    | 17 +++++++++++++++--
 drivers/net/wwan/wwan_core.c          |  9 ++++++---
 drivers/net/wwan/wwan_hwsim.c         |  3 ++-
 include/linux/wwan.h                  | 13 +++++++++----
 6 files changed, 36 insertions(+), 12 deletions(-)

diff --git a/drivers/net/wwan/iosm/iosm_ipc_port.c b/drivers/net/wwan/iosm/iosm_ipc_port.c
index beb944847398..2f874e41ceff 100644
--- a/drivers/net/wwan/iosm/iosm_ipc_port.c
+++ b/drivers/net/wwan/iosm/iosm_ipc_port.c

@@ -31,7 +31,8 @@ static void ipc_port_ctrl_stop(struct wwan_port *port)
 }

 /* transfer control data to modem */
-static int ipc_port_ctrl_tx(struct wwan_port *port, struct sk_buff *skb)
+static int ipc_port_ctrl_tx(struct wwan_port *port, struct sk_buff *skb,
+                           bool nonblock)
 {
        struct iosm_cdev *ipc_port = wwan_port_get_drvdata(port);

diff --git a/drivers/net/wwan/mhi_wwan_ctrl.c b/drivers/net/wwan/mhi_wwan_ctrl.c
index 1bc6b69aa530..9754f014d348 100644
--- a/drivers/net/wwan/mhi_wwan_ctrl.c
+++ b/drivers/net/wwan/mhi_wwan_ctrl.c

@@ -139,7 +139,8 @@ static void mhi_wwan_ctrl_stop(struct wwan_port *port)
        mhi_unprepare_from_transfer(mhiwwan->mhi_dev);
 }

-static int mhi_wwan_ctrl_tx(struct wwan_port *port, struct sk_buff *skb)
+static int mhi_wwan_ctrl_tx(struct wwan_port *port, struct sk_buff *skb,
+                           bool nonblock)
 {
        struct mhi_wwan_dev *mhiwwan = wwan_port_get_drvdata(port);
        int ret;

diff --git a/drivers/net/wwan/rpmsg_wwan_ctrl.c b/drivers/net/wwan/rpmsg_wwan_ctrl.c
index de226cdb69fd..63f431eada39 100644
--- a/drivers/net/wwan/rpmsg_wwan_ctrl.c
+++ b/drivers/net/wwan/rpmsg_wwan_ctrl.c

@@ -54,12 +54,16 @@ static void rpmsg_wwan_ctrl_stop(struct wwan_port *port)
        rpwwan->ept = NULL;
 }

-static int rpmsg_wwan_ctrl_tx(struct wwan_port *port, struct sk_buff *skb)
+static int rpmsg_wwan_ctrl_tx(struct wwan_port *port, struct sk_buff *skb,
+                             bool nonblock)
 {
        struct rpmsg_wwan_dev *rpwwan = wwan_port_get_drvdata(port);
        int ret;

-       ret = rpmsg_trysend(rpwwan->ept, skb->data, skb->len);
+       if (nonblock)
+               ret = rpmsg_trysend(rpwwan->ept, skb->data, skb->len);
+       else
+               ret = rpmsg_send(rpwwan->ept, skb->data, skb->len);
        if (ret)
                return ret;

@@ -67,10 +71,19 @@ static int rpmsg_wwan_ctrl_tx(struct wwan_port *port, struct sk_buff *skb)
        return 0;
 }

+static __poll_t rpmsg_wwan_ctrl_poll(struct wwan_port *port, struct file *filp,
+                                    poll_table *wait)
+{
+       struct rpmsg_wwan_dev *rpwwan = wwan_port_get_drvdata(port);
+
+       return rpmsg_poll(rpwwan->ept, filp, wait);
+}
+
 static const struct wwan_port_ops rpmsg_wwan_pops = {
        .start = rpmsg_wwan_ctrl_start,
        .stop = rpmsg_wwan_ctrl_stop,
        .tx = rpmsg_wwan_ctrl_tx,
+       .poll = rpmsg_wwan_ctrl_poll,
 };

 static struct device *rpmsg_wwan_find_parent(struct device *dev)

diff --git a/drivers/net/wwan/wwan_core.c b/drivers/net/wwan/wwan_core.c
index 7e728042fc41..c7fd0b897f87 100644
--- a/drivers/net/wwan/wwan_core.c
+++ b/drivers/net/wwan/wwan_core.c

@@ -500,7 +500,8 @@ static void wwan_port_op_stop(struct wwan_port *port)
        mutex_unlock(&port->ops_lock);
 }

-static int wwan_port_op_tx(struct wwan_port *port, struct sk_buff *skb)
+static int wwan_port_op_tx(struct wwan_port *port, struct sk_buff *skb,
+                          bool nonblock)
 {
        int ret;

@@ -510,7 +511,7 @@ static int wwan_port_op_tx(struct wwan_port *port, struct sk_buff *skb)
                goto out_unlock;
        }

-       ret = port->ops->tx(port, skb);
+       ret = port->ops->tx(port, skb, nonblock);

 out_unlock:
        mutex_unlock(&port->ops_lock);

@@ -637,7 +638,7 @@ static ssize_t wwan_port_fops_write(struct file *filp, const char __user *buf,
                return -EFAULT;
        }

-       ret = wwan_port_op_tx(port, skb);
+       ret = wwan_port_op_tx(port, skb, !!(filp->f_flags & O_NONBLOCK));
        if (ret) {
                kfree_skb(skb);
                return ret;

@@ -659,6 +660,8 @@ static __poll_t wwan_port_fops_poll(struct file *filp, poll_table *wait)
                mask |= EPOLLIN | EPOLLRDNORM;
        if (!port->ops)
                mask |= EPOLLHUP | EPOLLERR;
+       else if (port->ops->poll)
+               mask |= port->ops->poll(port, filp, wait);

I'm not sure it useful here because EPOLLOUT flag is already set above, right?

Oops, you're right - sorry! I thought the flags are inverted (only set
if (is_write_blocked())), then it would have worked fine. :)

I think this should be easy to fix though, I can just make the

        if (!is_write_blocked(port))
                mask |= EPOLLOUT | EPOLLWRNORM;

if statement conditional to (port->ops->poll == NULL). It only makes
sense to supply the poll() op if the built-in write-blocking cannot be
used easily (like in my case).

Yes, so maybe in that case poll ops should be renamed to something like tx_poll?

Regards,
Loic