Re: [PATCH v4] usb: gadget: u_ether: Replace netif_stop_queue with netif_device_detach
From: Ferry Toth
Date: Sun May 26 2024 - 16:53:08 EST
Hi
Op 15-05-2024 om 20:38 schreef Ferry Toth:
Hi,
Sorry to have kept you waiting, I was away for a short break.
Patch tested below.
Op 10-05-2024 om 11:45 schreef Hardik Gajjar:
On Thu, May 02, 2024 at 10:32:16PM +0200, Ferry Toth wrote:
Oops, sorry, wrong file attached . Now correct one.
Op 02-05-2024 om 22:13 schreef Ferry Toth:
Op 02-05-2024 om 17:29 schreef Hardik Gajjar:
On Tue, Apr 30, 2024 at 11:12:17PM +0200, Ferry Toth wrote:
Hi,
Op 30-04-2024 om 21:40 schreef Ferry Toth:
Hi,
Op 30-04-2024 om 17:32 schreef Hardik Gajjar:
On Sun, Apr 28, 2024 at 11:07:36PM +0200, Ferry Toth wrote:
Hi,
Op 25-04-2024 om 23:27 schreef Ferry Toth:
Hi,
Op 17-04-2024 om 17:13 schreef Hardik Gajjar:
On Tue, Apr 16, 2024 at 04:48:32PM +0300, Andy Shevchenko wrote:
On Thu, Apr 11, 2024 at 10:52:36PM +0200, Ferry Toth wrote:
Op 11-04-2024 om 18:39 schreef Andy Shevchenko:
On Thu, Apr 11, 2024 at 04:26:37PM +0200, Hardik Gajjar
wrote:
On Wed, Apr 10, 2024 at
08:37:42PM +0300, Andy
Shevchenko wrote:
On Sun, Apr 07, 2024 at 10:51:51PM +0200, Ferry Toth wrote:
Op 05-04-2024 om 13:38 schreef Hardik Gajjar:
...
Exactly. And this didn't happen before the 2 patches.
To be precise: /sys/class/net/usb0 is not
removed and it is a link, the link
target
/sys/devices/pci0000:00/0000:00:11.0/dwc3.0.auto/gadget.0/net/usb0
no
longer exists
So, it means that the /sys/class/net/usb0 is
present, but the symlink is
broken. In that case, the dwc3 driver should
recreate the device, and the
symlink should become active again
Yes, on first enabling gadget (when device mode is activated):
root@yuna:~# ls
/sys/devices/pci0000:00/0000:00:11.0/dwc3.0.auto/gadget.0/
driver net power sound subsystem suspended uevent
Then switching to host mode:
root@yuna:~# ls
/sys/devices/pci0000:00/0000:00:11.0/dwc3.0.auto/gadget.0/
ls: cannot access
'/sys/devices/pci0000:00/0000:00:11.0/dwc3.0.auto/gadget.0/':
No such file
or directory
Then back to device mode:
root@yuna:~# ls
/sys/devices/pci0000:00/0000:00:11.0/dwc3.0.auto/gadget.0/
driver power sound subsystem suspended uevent
net is missing. But, network functions:
root@yuna:~# ping 10.42.0.1
PING 10.42.0.1 (10.42.0.1): 56 data bytes
Mass storage device is created and removed each time as
expected.
So, what's the conclusion? Shall we move towards revert of
those
two changes?
As promised, I have the tested the this patch with the dwc3
gadget.
I could not reproduce
the issue.
I can see the usb0 exist all the time and accessible
regardless of
the role switching of the USB mode (peripheral <-> host)
Following are the logs:
//Host to device
console:/sys/bus/platform/devices/a800000.ssusb # echo
"peripheral"
mode
console:/sys/bus/platform/devices/a800000.ssusb # ls
a800000.dwc3/gadget/net/
usb0
//device to host
console:/sys/bus/platform/devices/a800000.ssusb
# echo "host" > mode
console:/sys/bus/platform/devices/a800000.ssusb # ls
a800000.dwc3/gadget/net/
usb0
That is weird. When I switch to host mode (using
the physical switch),
the whole gadget directory is removed (now testing 6.9.0-rc5)
Switching back to device mode, that gadget
directory is recreated. And
gadget/sound as well, but not gadget/net.
s a800000.dwc3/gadget/net/usb0
<
addr_assign_type duplex phys_port_name
addr_len flags phys_switch_id
address gro_flush_timeout power
broadcast ifalias proto_down
carrier ifindex queues
carrier_changes iflink speed
carrier_down_count link_mode statistics
carrier_up_count mtu subsystem
dev_id name_assign_type tx_queue_len
dev_port netdev_group type
device operstate uevent
dormant phys_port_id waiting_for_supplier
console:/sys/bus/platform/devices/a800000.ssusb # ifconfig -a
usb0
usb0 Link encap:Ethernet HWaddr 3a:8b:63:97:1a:9a
BROADCAST MULTICAST MTU:1500 Metric:1
RX packets:0 errors:0 dropped:0 overruns:0 frame:0
TX packets:0 errors:0 dropped:0 overruns:0
carrier:0
collisions:0 txqueuelen:1000
RX bytes:0 TX bytes:0
console:/sys/bus/platform/devices/a800000.ssusb #
I strongly advise against reverting the patch solely based on
the
observed issue of removing the /sys/class/net/usb0 directory
while
the usb0 interface remains available.
There's more to it. I also mentioned that switching the role or
unplugging the cable leaves the usb0 connection.
I have while in host mode:
root@yuna:~# ifconfig -a usb0
usb0: flags=-28605<UP,BROADCAST,RUNNING,MULTICAST,DYNAMIC>
mtu 1500
inet 10.42.0.221 netmask 255.255.255.0 broadcast
10.42.0.255
inet6 fe80::a8bb:ccff:fedd:eef1 prefixlen 64
scopeid 0x20<link>
You don't see that because you didn't create a connection at all.
Instead, I recommend enabling FTRACE to
trace the functions involved
and identify which faulty call is responsible for removing usb0.
Switching from device -> host -> device:
root@yuna:~# trace-cmd record -p function_graph -l *gether_*
plugin 'function_graph'
Hit Ctrl^C to stop recording
^CCPU0 data recorded at offset=0x1c8000
188 bytes in size (4096 uncompressed)
CPU1 data recorded at offset=0x1c9000
0 bytes in size (0 uncompressed)
root@yuna:~# trace-cmd report
cpus=2
irq/68-dwc3-725 [000] 514.575337:
funcgraph_entry: #
2079.480 us | gether_disconnect();
irq/68-dwc3-946 [000] 524.263731:
funcgraph_entry: +
11.640 us | gether_disconnect();
irq/68-dwc3-946 [000] 524.263743:
funcgraph_entry: !
116.520 us | gether_connect();
irq/68-dwc3-946 [000] 524.268029:
funcgraph_entry: #
2057.260 us | gether_disconnect();
irq/68-dwc3-946 [000] 524.270089:
funcgraph_entry: !
109.000 us | gether_connect();
I tried to get a more useful trace:
root@yuna:/sys/kernel/tracing# echo 'gether_*' > set_ftrace_filter
root@yuna:/sys/kernel/tracing# echo 'eem_*' >> set_ftrace_filter
root@yuna:/sys/kernel/tracing# echo function > current_tracer
root@yuna:/sys/kernel/tracing# echo 'reset_config'
set_ftrace_filter
-> switch to host mode then back to device
root@yuna:/sys/kernel/tracing# cat trace
# tracer: function
#
# entries-in-buffer/entries-written: 53/53 #P:2
#
# _-----=> irqs-off/BH-disabled
# / _----=> need-resched
# | / _---=> hardirq/softirq
# || / _--=> preempt-depth
# ||| / _-=> migrate-disable
# |||| / delay
# TASK-PID CPU# ||||| TIMESTAMP FUNCTION
# | | | ||||| | |
irq/68-dwc3-523 [000] D..3. 133.990254: reset_config
<-__composite_disconnect
irq/68-dwc3-523 [000] D..3. 133.992274: eem_disable
<-reset_config
irq/68-dwc3-523 [000] D..3. 133.992276:
gether_disconnect
<-reset_config
kworker/1:3-443 [001] ...1. 134.022453: eem_unbind
<-purge_configs_funcs
-> to device mode
kworker/1:3-443 [001] ...1. 148.630773: eem_bind
<-usb_add_function
irq/68-dwc3-734 [000] D..3. 149.155209: eem_set_alt
<-composite_setup
irq/68-dwc3-734 [000] D..3. 149.155215:
gether_disconnect
<-eem_set_alt
irq/68-dwc3-734 [000] D..3. 149.155220:
gether_connect
<-eem_set_alt
irq/68-dwc3-734 [000] D..3. 149.157287: eem_set_alt
<-composite_setup
irq/68-dwc3-734 [000] D..3. 149.157292:
gether_disconnect
<-eem_set_alt
irq/68-dwc3-734 [000] D..3. 149.159338:
gether_connect
<-eem_set_alt
irq/68-dwc3-734 [000] D..2. 149.239625: eem_unwrap
<-rx_complete
...
I don't know where to look exactly. Any hints?
do you see anything related to gether_cleanup() after
eem_unbind() ?
Nope. It's a pitty that the trace formatting got messed up
above. But as
you can see I traced gether_* and eem_*. After eem_unbind no traced
function is called, until I flip the switch to device mode.
The ... at the end is where I cut uninteresting eem_unwrap
<-rx_complete
and eem_wrap <-eth_start_xmit lines.
If not then, you may try to enable tracing of TCP/IP stack and
network side to check who deleting the sysfs entry
Yes, that's a vast amount of functions to trace. And I don't see how
that would be related to the patch we're discussing here. I was
hoping
for a little more targeted hint.
Now filtering 'gether_*', 'eem_*', '*configfs_*', 'composite_*',
'usb_fun*',
'reset_config' and 'device_remove_file' leads me to:
TIMESTAMP FUNCTION
| |
49.952477: eem_wrap <-eth_start_xmit
55.072455: eem_wrap <-eth_start_xmit
55.072621: eem_unwrap <-rx_complete
59.011540: configfs_composite_reset <-usb_gadget_udc_reset
59.011545: composite_reset <-configfs_composite_reset
59.011548: reset_config <-__composite_disconnect
59.013565: eem_disable <-reset_config
59.013567: gether_disconnect <-reset_config
59.049560: device_remove_file <-device_remove
59.051185: configfs_composite_disconnect <-usb_gadget_disco
59.051189: composite_disconnect <-configfs_composite_discon
59.051195: configfs_composite_unbind <-gadget_unbind_driver
59.052519: eem_unbind <-purge_configs_funcs
59.052529: composite_dev_cleanup <-configfs_composite_unbin
59.052537: device_remove_file <-composite_dev_cleanup
device_remove_file gets called twice, once by device_remove after
gether_disconnect (that the one). The 2nd time by
composite_dev_cleanup
(removing the gadget)
I believe that the device_remove_file function is only removing
suspend-specific attributes, not the complete gadget.
Typically, when you perform the role switch, the Gadget start/stop
function in your UDC driver is called. These functions should not
delete the gadget
To investigate further, could you please enable the DWC3 functions
in ftrace and check who is removing the gadget?
I can also enable this on my system and compare the logs with yours,
but I will be in PI planning for 1.5 weeks and may not be able to
provide immediate support.
Yes, but of course adding dwc3_* (and usb_*) also traces host mode, so
trace is 600kb. I cut uninteresting stuff before
configfs_composite_reset <-usb_gadget_udc_reset and after
__dwc3_set_mode,
<https://urldefense.proofpoint.com/v2/url?u=http-3A__300linesremain.Seeattachedtar.gzforyouto&d=DwIDaQ&c=euGZstcaTDllvimEN8b7jXrwqOf-v5A_CdpgnVfiiMM&r=SAhjP5GOmrADp1v_EE5jWoSuMlYCIt9gKduw-DCBPLs&m=zdiBhk-2V5AXxu707QAhbCgWR4qNVRARBmxN17nVB69gVOm-QPqrJeKpo4_trszw&s=ixagWKgLs6wQDJfwh4vIDQNDiy8GZnK9KnUELIfiJz0&e=>
compare.
Could you please try with the following patch and see if your issue
resolves ?
diff --git a/drivers/usb/gadget/function/f_eem.c
b/drivers/usb/gadget/function/f_eem.c
index 3b445bd88498..c2a904475083 100644
--- a/drivers/usb/gadget/function/f_eem.c
+++ b/drivers/usb/gadget/function/f_eem.c
@@ -247,7 +247,7 @@ static int eem_bind(struct usb_configuration *c,
struct usb_function *f)
struct usb_composite_dev *cdev = c->cdev;
struct f_eem *eem = func_to_eem(f);
struct usb_string *us;
- int status;
+ int status = 0;
struct usb_ep *ep;
struct f_eem_opts *eem_opts;
@@ -260,16 +260,20 @@ static int eem_bind(struct usb_configuration *c,
struct usb_function *f)
* with list_for_each_entry, so we assume no race condition
* with regard to eem_opts->bound access
*/
+ mutex_lock(&eem_opts->lock);
+ gether_set_gadget(eem_opts->net, cdev->gadget);
+
if (!eem_opts->bound) {
- mutex_lock(&eem_opts->lock);
- gether_set_gadget(eem_opts->net, cdev->gadget);
status = gether_register_netdev(eem_opts->net);
- mutex_unlock(&eem_opts->lock);
- if (status)
- return status;
- eem_opts->bound = true;
}
+ mutex_unlock(&eem_opts->lock);
+
+ if (status)
+ goto fail;
+
+ eem_opts->bound = true;
+
us = usb_gstrings_attach(cdev, eem_strings,
ARRAY_SIZE(eem_string_defs));
After switching from host -> device -> host the usb0 device as seen by
ifconfig remains "RUNNING" and in the routing table.
FYI This is the regression.
For us, I don't see any advantages in having this patch and only the
downside. Do you have any further ideas / patches that we could test.
Otherwise, it might be best to go with the original suggestion and
revert until it gets sorted out?
Also /sys/devices/pci0000:00/0000:00:11.0/dwc3.0.auto/gadget.0/net is
still missing.
After deeper diving into v6.1.0, I found the latter a longer existing
bug, not caused by your patch.
More over, this doesn't seem to do any harm.
The first issue does, as we try to route traffic over a device that no
longer exists.
Additionally, please check if you have any customized DWC patches
that may be causing this problem.
You may recall the whole issue did not occur before this
patch got applied.
Hardik
According to current kernel architecture of u_ether driver, only
gether_cleanup should remove the usb0 interface along with its
kobject and sysfs interface.
I suggest sharing the analysis here to understand why this
practice
is not followed in your use case or driver ?
Yes, I'll try to trace where that happens.
Nevertheless, the disappearance of the net/usb0 directory seems
harmless? But the usb: net device remaining after disconnect
or role
switch is not good, as the route remains.
May be they are 2 separate problems. Could you try to
reproduce what
happens if you make eem connection and then unplug?
I am curious why the driver was developed without adhering to
the
kernel's gadget architecture.
I don't know what you mean here. Which driver do you mean?
I have the dwc3 IP base usb controller, Let me check
with this patch and
share result here. May be we need some fix in dwc3
Would have been nice if someone could test on other
controller as well. But
another instance of dwc3 is also very welcome.
It's quite possible, please test on your side.
We are happy to test any fixes if you come up with.
--
With Best Regards,
Andy Shevchenko