Re: [tpmdd-devel] [PATCH v2 2/7] tpm: two-phase chip management functions

[Jason Gunthorpe <jgunthorpe@obsidianresearch.com>]

On Tue, Oct 07, 2014 at 08:01:12PM +0300, Jarkko Sakkinen wrote:
> Added tpm_chip_alloc() and tpm_chip_register() where tpm_chip_alloc()
> reserves memory resources and tpm_chip_register() initializes the
> device driver. This way it is possible to alter struct tpm_chip
> attributes before passing it to tpm_chip_register().

This looks broadly reasonable to me

Please add a note to the commit that this is known to still have
problems with resource reference counting, but they are less severe
than what existed before, and this is only an interm step.

> +/**
> + * tpm_chip_alloc() - allocate a new struct tpm_chip instance

This is using devm so it should be called 'tpmm_chip_alloc()' for
clarity


I know that was there before, but it sure is racy:

> +	chip->dev_num = find_first_zero_bit(dev_mask, TPM_NUM_DEVICES);
[..]
> +	set_bit(chip->dev_num, dev_mask);

Someday it should use IDR.


> @@ -896,18 +872,7 @@ void tpm_remove_hardware(struct device *dev)
>  		return;
>  	}
>  
> -	spin_lock(&driver_lock);
> -	list_del_rcu(&chip->list);
> -	spin_unlock(&driver_lock);
> -	synchronize_rcu();
> -
> -	tpm_dev_del_device(chip);
> -	tpm_sysfs_del_device(chip);
> -	tpm_remove_ppi(&dev->kobj);
> -	tpm_bios_log_teardown(chip->bios_dir);
> -
> -	/* write it this way to be explicit (chip->dev == dev) */
> -	put_device(chip->dev);
> +	tpm_chip_unregister(chip);
>  }
>  EXPORT_SYMBOL_GPL(tpm_remove_hardware);

This can move to tpm-chip too, same with tpm_register_hardware

> @@ -714,15 +709,10 @@ static int tpm_tis_i2c_remove(struct i2c_client *client)
>  	struct tpm_chip *chip = tpm_dev.chip;
>  	release_locality(chip, chip->vendor.locality, 1);
>  
> -	/* close file handles */
> -	tpm_dev_vendor_release(chip);
> -
>  	/* remove hardware */
>  	tpm_remove_hardware(chip->dev);

Wrong ordering here, tpm_remove_hardware should always be first -
drivers should not tear down internal state before calling it, so
release_locality should be second.

Noting that since we use devm the kfree will not happen until
remove returns, so the chip pointer is still valid.

>  	/* reset these pointers, otherwise we oops */
> -	chip->dev->release = NULL;
> -	chip->release = NULL;
>  	tpm_dev.client = NULL;

The comment can go too

Note: tpm_dev should be driver private data, but that is not your
problem..

Did you test compile all the drivers? One of my git commits on github
has some hackery to make that possible on x86.

Jason