Re: [BUG] pwm: sysfs: It is not possible to export more than one channel

From: Michal VokÃÄ
Date: Tue Sep 25 2018 - 07:52:59 EST


On 25.9.2018 12:21, Thierry Reding wrote:
On Tue, Sep 25, 2018 at 11:11:32AM +0200, Michal VokÃÄ wrote:
Hi,

Since commit 7e5d1fd75c3d ("pwm: Set class for exported channels in
sysfs") it is not possible to export more than one PWM channel.

This is because for each exported channel a directory named pwmN is
created in /sys/class/pwm/. As channels for all PWM chips are numbered
from 0 it is not possible to export pwmN channel from one chip and
pwmN channel from another chip at the same time.

In theory if your SoC has N PWM chips and each chip has N channels you
should be able to export N channels. But only one channel from each chip.

I found the issue on i.MX6dl SoC with two PWM chips where each PWM chip
has just one channel.

This is how it can be reproduced:

root@hydraco:/sys/class/pwm# ls
pwmchip0 -> ../../devices/soc0/soc/2000000.aips-bus/2080000.pwm/pwm/pwmchip0
pwmchip1 -> ../../devices/soc0/soc/2000000.aips-bus/208c000.pwm/pwm/pwmchip1

root@hydraco:/sys/class/pwm# cat pwmchip0/npwm
1
root@hydraco:/sys/class/pwm# cat pwmchip1/npwm
1

root@hydraco:/sys/class/pwm# echo 0 > pwmchip0/export
root@hydraco:/sys/class/pwm# ls -l
pwm0 -> ../../devices/soc0/soc/2000000.aips-bus/2080000.pwm/pwm/pwmchip0/pwm0
pwmchip0 -> ../../devices/soc0/soc/2000000.aips-bus/2080000.pwm/pwm/pwmchip0
pwmchip1 -> ../../devices/soc0/soc/2000000.aips-bus/208c000.pwm/pwm/pwmchip1

root@hydraco:/sys/class/pwm# echo 0 > pwmchip1/export
[ 543.110824] sysfs: cannot create duplicate filename '/class/pwm/pwm0'
[ 543.117314] CPU: 1 PID: 351 Comm: sh Not tainted 4.19.0-rc5-00025-g249f3ed-dirty #8
[ 543.124990] Hardware name: Freescale i.MX6 Quad/DualLite (Device Tree)
[ 543.131587] [<80112d20>] (unwind_backtrace) from [<8010ddb4>] (show_stack+0x20/0x24)
[ 543.139383] [<8010ddb4>] (show_stack) from [<80ba1044>] (dump_stack+0x80/0x94)
[ 543.146652] [<80ba1044>] (dump_stack) from [<8031e2b4>] (sysfs_warn_dup+0x6c/0x78)
[ 543.154257] [<8031e2b4>] (sysfs_warn_dup) from [<8031e61c>] (sysfs_do_create_link_sd+0xd8/0xdc)
[ 543.162988] [<8031e61c>] (sysfs_do_create_link_sd) from [<8031e678>] (sysfs_create_link+0x38/0x44)
[ 543.171986] [<8031e678>] (sysfs_create_link) from [<806026b0>] (device_add+0x2c8/0x630)
[ 543.180025] [<806026b0>] (device_add) from [<80602a3c>] (device_register+0x24/0x28)
[ 543.187724] [<80602a3c>] (device_register) from [<80501acc>] (export_store+0x118/0x190)
[ 543.195762] [<80501acc>] (export_store) from [<805ffc10>] (dev_attr_store+0x28/0x34)
[ 543.203539] [<805ffc10>] (dev_attr_store) from [<8031d7dc>] (sysfs_kf_write+0x48/0x54)
[ 543.211485] [<8031d7dc>] (sysfs_kf_write) from [<8031cde0>] (kernfs_fop_write+0xf8/0x1e0)
[ 543.219696] [<8031cde0>] (kernfs_fop_write) from [<80294d78>] (__vfs_write+0x48/0x170)
[ 543.227645] [<80294d78>] (__vfs_write) from [<80295064>] (vfs_write+0xb4/0x1c0)
[ 543.234981] [<80295064>] (vfs_write) from [<802952e8>] (ksys_write+0x5c/0xbc)
[ 543.242144] [<802952e8>] (ksys_write) from [<80295360>] (sys_write+0x18/0x1c)
[ 543.249313] [<80295360>] (sys_write) from [<80101000>] (ret_fast_syscall+0x0/0x54)
[ 543.256901] Exception stack(0xecfbffa8 to 0xecfbfff0)
[ 543.261980] ffa0: 00000002 76fc8000 00000001 76fc8000 00000002 00000000
[ 543.270184] ffc0: 00000002 76fc8000 76f5cd60 00000004 00000002 000bd134 00000001 000ba730
[ 543.278380] ffe0: 00000000 7ebff954 76e8b988 76ee3cd0
-sh: echo: write error: File exists

I am not sure what will be the right solution. The problem is that
numbering of the PWM channels does not work the same way as in the GPIO
subsystem where each GPIO has its unique number.

Should we just revert the change? It was not documented in the PWM sysfs
interface. What do you think?

Fabrice has reported the same thing. We've been discussing possible
solutions here:

http://patchwork.ozlabs.org/patch/973224/

Thank you for the reference Thierry. I was convinced there must be someone
else who encountered the same problem. My search did not produce anything
useful though.

It looks to me like we'd need to at least revert the patch to restore
the sysfs ABI. At the same time I think we'll want to fix the issue that
the offending commit was trying to address.

Agree, ideally we want to solve both problems. My knowledge of the overall
architecture is still fairly limited though. I have no idea if/how we can
deal with the issue the offending commit was addressing if we revert it.

Would be great if we could somehow choose a different name just for the
symlink but leave the device name as is to not break the ABI as Fabrice
suggested.

Michal