On 2022/11/10 9:26, Kent Gibson wrote:
On Wed, Nov 09, 2022 at 04:47:08PM +0200, Andy Shevchenko wrote:
On Wed, Nov 09, 2022 at 05:31:20PM +0800, Zeng Heng wrote:
gcdev_register & gcdev_unregister call device_add & device_del to
request/release source. But in device_add, the dev->p allocated by
device_private_init is not released by device_del.
First of all, we refer to the functions like func().
Thanks, it would be updated in next version.
Further to this, the description of the problem could be clearer -
it would be helpful to indicate the code path that triggers the problem
- it is gpiochip_sysfs_register() returning an error?
So when calling gcdev_unregister to release gdev, it needs put_device
to release dev in the following.
Otherwise, kmemleak would report memory leak such as below:
unreferenced object 0xffff88810b406400 (size 512):
comm "python3", pid 1682, jiffies 4295346908 (age 24.090s)
hex dump (first 32 bytes):
00 00 00 00 ad 4e ad de ff ff ff ff 00 00 00 00 .....N..........
ff ff ff ff ff ff ff ff a0 5e 23 90 ff ff ff ff .........^#.....
backtrace:
Second, read Submitting Patches on how to provide your backtraces in the
message body.
Thanks, it would be updated in next version.
[<00000000a58ee5fe>] kmalloc_trace+0x22/0x110
[<0000000045fe2058>] device_add+0xb34/0x1130
[<00000000d778b45f>] cdev_device_add+0x83/0xe0
[<0000000089f948ed>] gpiolib_cdev_register+0x73/0xa0
[<00000000a3a8a316>] gpiochip_setup_dev+0x1c/0x70
[<00000000787227b4>] gpiochip_add_data_with_key+0x10f6/0x1bf0
[<000000009ac5742c>] devm_gpiochip_add_data_with_key+0x2e/0x80
[<00000000bf2b23d9>] xra1403_probe+0x192/0x1b0 [gpio_xra1403]
[<000000005b5ef2d4>] spi_probe+0xe1/0x140
[<000000002b26f6f1>] really_probe+0x17c/0x3f0
[<00000000dd2dad9c>] __driver_probe_device+0xe3/0x170
[<000000005ca60d2a>] device_driver_attach+0x34/0x80
[<00000000e9db90db>] bind_store+0x10b/0x1a0
[<00000000e2650f8a>] drv_attr_store+0x49/0x70
[<0000000080a80b2b>] sysfs_kf_write+0x8c/0xb0
[<00000000a28b45b9>] kernfs_fop_write_iter+0x216/0x2e0
unreferenced object 0xffff888100de9800 (size 512):
comm "python3", pid 264, jiffies 4294737615 (age 33.514s)
hex dump (first 32 bytes):
00 00 00 00 ad 4e ad de ff ff ff ff 00 00 00 00 .....N..........
ff ff ff ff ff ff ff ff a0 5e 63 a1 ff ff ff ff .........^c.....
backtrace:
[<00000000bcc571d0>] kmalloc_trace+0x22/0x110
[<00000000eeb06124>] device_add+0xb34/0x1130
[<000000007e5cd2fd>] cdev_device_add+0x83/0xe0
[<000000008f6bcd3a>] gpiolib_cdev_register+0x73/0xa0
[<0000000012c93b24>] gpiochip_setup_dev+0x1c/0x70
[<00000000a24b646a>] gpiochip_add_data_with_key+0x10f6/0x1bf0
[<000000000c225212>] tpic2810_probe+0x16e/0x196 [gpio_tpic2810]
[<00000000b52d04ff>] i2c_device_probe+0x651/0x680
[<0000000058d3ff6b>] really_probe+0x17c/0x3f0
[<00000000586f43d3>] __driver_probe_device+0xe3/0x170
[<000000003f428602>] device_driver_attach+0x34/0x80
[<0000000040e91a1b>] bind_store+0x10b/0x1a0
[<00000000c1d990b9>] drv_attr_store+0x49/0x70
[<00000000a23bfc22>] sysfs_kf_write+0x8c/0xb0
[<00000000064e6572>] kernfs_fop_write_iter+0x216/0x2e0
[<00000000026ce093>] vfs_write+0x658/0x810
Because at the point of gpiochip_setup_dev here, where dev.release
does not set yet, calling put_device would cause the warning of
no release function and double-free in the following fault handler
route (when kfree dev_name). So directly calling kfree to release
dev->p here in case of memory leak.
Again, this could be clearer. The dev->p is normally freed by
device_release() - why is that not happening in this case?
(as put_device() is never called in this path)
The double free you see if you do call put_device() appears to be due to
different expectations as to the cleanup that gpiochip_setup_dev() will
perform on error, depending on where it is called. gpiochip_setup_devs()
assumes any cleanup is performed by gpiochip_setup_dev(), while
gpiochip_add_data_with_key() assumes that it hasn't performed any cleanup.
Having gpiochip_setup_dev() perform its own cleanup makes the most sense
to me, so gpiochip_add_data_with_key() should be changed to allow for
that.
Right, the cleanup route of gpiochip_add_data_with_key() &
gpiochip_setup_dev()
has to be considered comprehensively after any possible cases of fault
injections.
...
@@ -539,6 +539,7 @@ static int gpiochip_setup_dev(struct gpio_device *gdev)
err_remove_device:
gcdev_unregister(gdev);
+ kfree(gdev->dev.p);
return ret;
Third, I do not believe it's a correct fix.
Have you read comments around device_del() / etc.?
Yes, not only the comments I read, but also the device_del() implement code.
Releasing the dev->p pointer is not the business with device_del(), but
it's relied on
put_device() calling release function.
Turning back here, the release function is not set yet at this point,
there is a gap
between device_add() and set release function pointer.
That's the reason why choose to free dev->p explicitly as the mail
mentioned above.
I agree - this is not the correct fix. The correct fix is to trigger the
normal cleanup mechanism, so put_device().
The fact that that triggers a warning:
"Device '%s' does not have a release() function, it is broken and must be
fixed. See Documentation/core-api/kobject.rst.\n"
is an indicator that dev.release should be set earlier.
If gpiodevice_release() is not appropriate, or cannot be modified to deal
with the device state at that point, then an appropriate interim release
function should be set.
And, as mentioned above, gpiochip_add_data_with_key() needs to be modified
to allow for gpiochip_setup_dev() cleaning up its own mess.
That is my take, but that is just from perusing the code so I may be
totally off base. Either way, an ACK/NACK on this from a maintainer or
other gpiolib expert would be helpful to expiditing a solution.
Cheers,
Kent.
Yes, exactly.
Thanks to all,
Zeng Heng
