Re: [RFC PATCH] net: devlink: devlink_nl_info_fill: populate default information

Wed, Nov 23, 2022 at 05:08:10PM CET, mailhol.vincent@wanadoo.fr wrote:

On Wed. 23 Nov. 2022 at 21:10, Jiri Pirko [off-list ref] wrote:

quoted

Wed, Nov 23, 2022 at 12:00:44PM CET, mailhol.vincent@wanadoo.fr wrote:

quoted

On Wed. 23 nov. 2022 à 18:46, Jiri Pirko [off-list ref] wrote:

quoted

Tue, Nov 22, 2022 at 04:49:34PM CET, mailhol.vincent@wanadoo.fr wrote:

quoted

Some piece of information are common to the vast majority of the
devices. Examples are:

 * the driver name.
 * the serial number of a USB device.

Modify devlink_nl_info_fill() to retrieve those information so that
the drivers do not have to. Rationale: factorize code.

Signed-off-by: Vincent Mailhol <redacted>
---
I am sending this as an RFC because I just started to study devlink.

I can see a parallel with ethtool for which the core will fill
whatever it can. c.f.:
commit f20a0a0519f3 ("ethtool: doc: clarify what drivers can implement in their get_drvinfo()")
Link: https://git.kernel.org/netdev/net-next/c/f20a0a0519f3

I think that devlink should do the same.

Right now, I identified two fields. If this RFC receive positive
feedback, I will iron it up and try to see if there is more that can
be filled by default.

Thank you for your comments.
---
net/core/devlink.c | 36 ++++++++++++++++++++++++++++++++++++
1 file changed, 36 insertions(+)

diff --git a/net/core/devlink.c b/net/core/devlink.c
index 7f789bbcbbd7..1908b360caf7 100644
--- a/net/core/devlink.c
+++ b/net/core/devlink.c

@@ -18,6 +18,7 @@

#include <linux/netdevice.h>
#include <linux/spinlock.h>
#include <linux/refcount.h>
+#include <linux/usb.h>
#include <linux/workqueue.h>
#include <linux/u64_stats_sync.h>
#include <linux/timekeeping.h>

@@ -6685,12 +6686,37 @@ int devlink_info_version_running_put_ext(struct devlink_info_req *req,

}
EXPORT_SYMBOL_GPL(devlink_info_version_running_put_ext);

+static int devlink_nl_driver_info_get(struct device_driver *drv,
+                                    struct devlink_info_req *req)
+{
+      if (!drv)
+              return 0;
+
+      if (drv->name[0])
+              return devlink_info_driver_name_put(req, drv->name);

Make sure that this provides the same value for all existing drivers
using devlink.

There are 21 drivers so far which reports the driver name through devlink. c.f.:
 $ git grep "devlink_info_driver_name_put(" drivers | wc -l

Out of those 21, there is only one: the mlxsw which seems to report
something different than device_driver::name. Instead it reports some
bus_info:
 https://elixir.bootlin.com/linux/v6.1-rc1/source/drivers/net/ethernet/mellanox/mlxsw/core.c#L1462
 https://elixir.bootlin.com/linux/v6.1-rc1/source/drivers/net/ethernet/mellanox/mlxsw/core.h#L504

I am not sure what the bus_info is here, but it looks like a misuse of
the field here.

When you are not sure, look into the code to find out :) I see no misue.
What exactly do you mean by that?

I mean that device_kind, it does not sound like a field that would
hold the driver name.

Looking deeper in the code, I got the confirmation.
bus_info::device_kind is initialized here (among other):
https://elixir.bootlin.com/linux/v6.1-rc1/source/drivers/net/ethernet/mellanox/mlxsw/i2c.c#L714

and it uses ic2_client::name which indicate the type of the device
(e.g. chip name):
https://elixir.bootlin.com/linux/v6.1-rc1/source/include/linux/i2c.h#L317

So I confirm that this is a misuse. This driver does not report the
driver's name.

Okay, I think that is a bug of mlxsw_i2c implementation. You can fix it
along the way.

quoted

quoted

quoted

quoted

+
+      return 0;
+}
+
+static int devlink_nl_usb_info_get(struct usb_device *udev,
+                                 struct devlink_info_req *req)
+{
+      if (!udev)
+              return 0;
+
+      if (udev->serial[0])
+              return devlink_info_serial_number_put(req, udev->serial);
+
+      return 0;
+}
+
static int
devlink_nl_info_fill(struct sk_buff *msg, struct devlink *devlink,
                   enum devlink_command cmd, u32 portid,
                   u32 seq, int flags, struct netlink_ext_ack *extack)
{
      struct devlink_info_req req = {};
+      struct device *dev = devlink_to_dev(devlink);
      void *hdr;
      int err;

@@ -6707,6 +6733,16 @@ devlink_nl_info_fill(struct sk_buff *msg, struct devlink *devlink,
      if (err)
              goto err_cancel_msg;

+      err = devlink_nl_driver_info_get(dev->driver, &req);
+      if (err)
+              goto err_cancel_msg;
+
+      if (!strcmp(dev->parent->type->name, "usb_device")) {

Comparing to string does not seem correct here.

There is a is_usb_device() which does the check:
 https://elixir.bootlin.com/linux/v6.1-rc1/source/drivers/usb/core/usb.h#L152

but this macro is not exposed outside of the usb core. The string
comparison was the only solution I found.

Find a different one. String check here is wrong.

quoted

Do you have any other ideas? If not and if this goes further than the
RFC stage, I will ask the USB folks if there is a better way.

quoted

quoted

+              err = devlink_nl_usb_info_get(to_usb_device(dev->parent), &req);

As Jakub pointed out, you have to make sure that driver does not put the
same attrs again. You have to introduce this functionality with removing
the fill-ups in drivers atomically, in a single patch.

Either this, either track if the attribute is already set. I would
prefer to remove all drivers fill-ups but this is not feasible for the
serial number. c.f. my reply to Jacub in this thread:
 https://lore.kernel.org/netdev/CAMZ6RqJ8_=h1SS7WmBeEB=75wsvVUZrb-8ELCDtpZb0gSs=2+A@mail.gmail.com/ (local)

Sure, but for the driver name it is. Let's start there.

I will do a first patch only for the driver name and think again of
the USB serial later on.

Yep, thanks!

Yours sincerely,
Vincent Mailhol