Re: [PATCH] scsi: sd: unregister device if device_add_disk() failed in sd_probe()

[Yu Kuai]

Hi,

在 2024/01/30 1:46, Luis Chamberlain 写道:
> On Fri, Dec 22, 2023 at 04:27:16PM +0800, Yu Kuai wrote:
> > Hi,
> >
> > 在 2023/12/22 14:49, Luis Chamberlain 写道:
> > > On Fri, Dec 08, 2023 at 04:23:35PM +0800, linan666@xxxxxxxxxxxxxxx wrote:
> > > > From: Li Nan <linan122@xxxxxxxxxx>
> > > >
> > > > "if device_add() succeeds, you should call device_del() when you want to
> > > > get rid of it."
> > > >
> > > > In sd_probe(), device_add_disk() fails when device_add() has already
> > > > succeeded, so change put_device() to device_unregister() to ensure device
> > > > resources are released.
> > > >
> > > > Fixes: 2a7a891f4c40 ("scsi: sd: Add error handling support for add_disk()")
> > > > Signed-off-by: Li Nan <linan122@xxxxxxxxxx>
> > >
> > > Nacked-by: Luis Chamberlain <mcgrof@xxxxxxxxxx>
> > >
> > > > ---
> > > >    drivers/scsi/sd.c | 2 +-
> > > >    1 file changed, 1 insertion(+), 1 deletion(-)
> > > >
> > > > diff --git a/drivers/scsi/sd.c b/drivers/scsi/sd.c
> > > > index 542a4bbb21bc..d81cbeee06eb 100644
> > > > --- a/drivers/scsi/sd.c
> > > > +++ b/drivers/scsi/sd.c
> > > > @@ -3736,7 +3736,7 @@ static int sd_probe(struct device *dev)
> > > >    	error = device_add_disk(dev, gd, NULL);
> > > >    	if (error) {
> > > > -		put_device(&sdkp->disk_dev);
> > > > +		device_unregister(&sdkp->disk_dev);
> > > >    		put_disk(gd);
> > > >    		goto out;
> > > >    	}
> > >
> > > This is incorrect, device_unregister() calls:
> > >
> > > void device_unregister(struct device *dev)
> > > {
> > > 	pr_debug("device: '%s': %s\n", dev_name(dev), __func__);
> > > 	device_del(dev);
> > > 	put_device(dev);
> > > }
> > >
> > > So you're adding what you believe to be a correct missing device_del().
> > > But what you missed is that if device_add_disk() fails then device_add()
> > > did not succeed because the new code we have in the kernel *today* unwinds
> > > this for us now.
> >
> > I'm confused here, there are two device here, one is 'sdkp->disk_dev',
> > one is gendisk->part0->bd_device, and the order in which they
> > initialize:
> >
> > sd_probe
> > device_add(&sdkp->disk_dev) -> succeed
> > device_add_disk -> failed, and device_add(bd_device) did not succeed
> > put_device(&sdkp->disk_dev) -> device_del is missed
> >
> > I don't see that if device_add_disk() fail, device_del() for
> > 'sdkp->disk_dev'is called from anywhere. Do I missing anything?
>
> Ah then the fix is still incorrect and the commit log should
> describe that this is for another device.
>
> How about this instead?
>
> > From c3f6e03f4a82aa253b6c487a293dcd576393b606 Mon Sep 17 00:00:00 2001
> From: Luis Chamberlain <mcgrof@xxxxxxxxxx>
> Date: Mon, 29 Jan 2024 09:25:18 -0800
> Subject: [PATCH] sd: remove extra put_device() for extra scsi device
>
> The sd driver first device_add() its own device, and later use
> device_add_disk() with another device. When we added error handling
> for device_add_disk() we now call put_disk() and that will trigger
> disk_release() when the refcount is 0. That will end up calling
> the block driver's disk->fops->free_disk() if one is defined. The

This is incorrect. GD_ADDED will only set when device_add_disk()
succeed, and free_disk() will only be called from disk_release() if
GD_ADDED is set. I think Li Nan's patch is correct.

> sd driver has scsi_disk_free_disk() as its free_disk() and that
> does the proper put_device(&sdkp->disk_dev) for us so we should not
> need to call it, however we are left still missing the device_del()
> for it.
>
> While at it, unwind with scsi_autopm_put_device(sdp) *prior* to
> putting to device as we do in sd_remove().
>
> Reported-by: Li Nan <linan122@xxxxxxxxxx>
> Reported-by: Yu Kuai <yukuai1@xxxxxxxxxxxxxxx>
> Fixes: 2a7a891f4c40 ("scsi: sd: Add error handling support for add_disk()")
> Signed-off-by: Luis Chamberlain <mcgrof@xxxxxxxxxx>
> ---
>   drivers/scsi/sd.c | 8 +++++---
>   1 file changed, 5 insertions(+), 3 deletions(-)
>
> diff --git a/drivers/scsi/sd.c b/drivers/scsi/sd.c
> index 7f949adbadfd..6475a3c947f8 100644
> --- a/drivers/scsi/sd.c
> +++ b/drivers/scsi/sd.c
> @@ -3693,8 +3693,9 @@ static int sd_probe(struct device *dev)
>   
>   	error = device_add(&sdkp->disk_dev);
>   	if (error) {
> +		scsi_autopm_put_device(sdp);
>   		put_device(&sdkp->disk_dev);
> -		goto out;
> +		return error;

I don't see why this is necessary, the tag 'out' is still there. If
you think is a problem, I think you need a separate patch to call
scsi_autopm_put_device() before putting the device.

Thanks,
Kuai

>   	}
>   
>   	dev_set_drvdata(dev, sdkp);
> @@ -3734,9 +3735,10 @@ static int sd_probe(struct device *dev)
>   
>   	error = device_add_disk(dev, gd, NULL);
>   	if (error) {
> -		put_device(&sdkp->disk_dev);
> +		scsi_autopm_put_device(sdp);
> +		device_del(&sdkp->disk_dev);
>   		put_disk(gd);
> -		goto out;
> +		return error;
>   	}
>   
>   	if (sdkp->security) {