Re: [PATCH v3 04/38] fs: add mount_setattr()

[Christoph Hellwig <hch@lst.de>]

Lots of crazy long lines in the patch.  Remember that you should only
go past 80 lines if it clearly improves readability, and I don't
think it does anywhere in here.

> index a7cd0f64faa4..a5a6c470dc07 100644
> --- a/fs/internal.h
> +++ b/fs/internal.h
> @@ -82,6 +82,14 @@ int may_linkat(struct path *link);
>  /*
>   * namespace.c
>   */
> +struct mount_kattr {
> +	unsigned int attr_set;
> +	unsigned int attr_clr;
> +	unsigned int propagation;
> +	unsigned int lookup_flags;
> +	bool recurse;
> +};

Even with the whole series applied this structure is only used in
namespace.c, so it might be worth moving there.

> +static inline int mnt_hold_writers(struct mount *mnt)
>  {
> -	int ret = 0;
> -
>  	mnt->mnt.mnt_flags |= MNT_WRITE_HOLD;
>  	/*
>  	 * After storing MNT_WRITE_HOLD, we'll read the counters. This store
> @@ -497,15 +495,29 @@ static int mnt_make_readonly(struct mount *mnt)
>  	 * we're counting up here.
>  	 */
>  	if (mnt_get_writers(mnt) > 0)
> -		ret = -EBUSY;
> -	else
> -		mnt->mnt.mnt_flags |= MNT_READONLY;
> +		return -EBUSY;
> +
> +	return 0;
> +}
> +
> +static inline void mnt_unhold_writers(struct mount *mnt)
> +{
>  	/*
>  	 * MNT_READONLY must become visible before ~MNT_WRITE_HOLD, so writers
>  	 * that become unheld will see MNT_READONLY.
>  	 */
>  	smp_wmb();
>  	mnt->mnt.mnt_flags &= ~MNT_WRITE_HOLD;
> +}
> +
> +static int mnt_make_readonly(struct mount *mnt)
> +{
> +	int ret;
> +
> +	ret = mnt_hold_writers(mnt);
> +	if (!ret)
> +		mnt->mnt.mnt_flags |= MNT_READONLY;
> +	mnt_unhold_writers(mnt);
>  	return ret;
>  }
>  
> @@ -3438,6 +3450,33 @@ SYSCALL_DEFINE5(mount, char __user *, dev_name, char __user *, dir_name,
>  	return ret;
>  }

This refactoring seems worth a little prep patch.

>  
> +static int build_attr_flags(unsigned int attr_flags, unsigned int *flags)
> +{
> +	unsigned int aflags = 0;
> +
> +	if (attr_flags & ~(MOUNT_ATTR_RDONLY |
> +			   MOUNT_ATTR_NOSUID |
> +			   MOUNT_ATTR_NODEV |
> +			   MOUNT_ATTR_NOEXEC |
> +			   MOUNT_ATTR__ATIME |
> +			   MOUNT_ATTR_NODIRATIME))
> +		return -EINVAL;
> +
> +	if (attr_flags & MOUNT_ATTR_RDONLY)
> +		aflags |= MNT_READONLY;
> +	if (attr_flags & MOUNT_ATTR_NOSUID)
> +		aflags |= MNT_NOSUID;
> +	if (attr_flags & MOUNT_ATTR_NODEV)
> +		aflags |= MNT_NODEV;
> +	if (attr_flags & MOUNT_ATTR_NOEXEC)
> +		aflags |= MNT_NOEXEC;
> +	if (attr_flags & MOUNT_ATTR_NODIRATIME)
> +		aflags |= MNT_NODIRATIME;
> +
> +	*flags = aflags;
> +	return 0;
> +}

Same for adding this helper.

> +	*kattr = (struct mount_kattr){

Missing whitespace before the {.

> +	switch (attr->propagation) {
> +	case MAKE_PROPAGATION_UNCHANGED:
> +		kattr->propagation = 0;
> +		break;
> +	case MAKE_PROPAGATION_UNBINDABLE:
> +		kattr->propagation = MS_UNBINDABLE;
> +		break;
> +	case MAKE_PROPAGATION_PRIVATE:
> +		kattr->propagation = MS_PRIVATE;
> +		break;
> +	case MAKE_PROPAGATION_DEPENDENT:
> +		kattr->propagation = MS_SLAVE;
> +		break;
> +	case MAKE_PROPAGATION_SHARED:
> +		kattr->propagation = MS_SHARED;
> +		break;
> +	default:

Any reason to not just reuse the MS_* flags in the new API?  Yes, your
new names are more descriptive, but having different names for the same
thing is also rather confusing.

> +	if (upper_32_bits(attr->attr_set))
> +		return -EINVAL;
> +	if (build_attr_flags(lower_32_bits(attr->attr_set), &kattr->attr_set))
> +		return -EINVAL;
> +
> +	if (upper_32_bits(attr->attr_clr))
> +		return -EINVAL;
> +	if (build_attr_flags(lower_32_bits(attr->attr_clr), &kattr->attr_clr))
> +		return -EINVAL;

What is so magic about the upper and lower 32 bits?

> +		return -EINVAL;
> +	else if ((attr->attr_clr & MOUNT_ATTR__ATIME) &&
> +		 ((attr->attr_clr & MOUNT_ATTR__ATIME) != MOUNT_ATTR__ATIME))
> +		return -EINVAL;

No need for the else here.

That being said I'd reword the thing to be a little more obvious:

	if (attr->attr_clr & MOUNT_ATTR__ATIME) {
		if ((attr->attr_clr & MOUNT_ATTR__ATIME) != MOUNT_ATTR__ATIME)
			return -EINVAL;

		... code doing the update of the atime flags here
	} else {
		if (attr->attr_set & MOUNT_ATTR__ATIME)
			return -EINVAL;
	}


> +/* Change propagation through mount_setattr(). */
> +enum propagation_type {
> +	MAKE_PROPAGATION_UNCHANGED	= 0, /* Don't change mount propagation (default). */
> +	MAKE_PROPAGATION_UNBINDABLE	= 1, /* Make unbindable. */
> +	MAKE_PROPAGATION_PRIVATE	= 2, /* Do not receive or send mount events. */
> +	MAKE_PROPAGATION_DEPENDENT	= 3, /* Only receive mount events. */
> +	MAKE_PROPAGATION_SHARED		= 4, /* Send and receive mount events. */
> +};

FYI, in uapis using defines instead of enums is usually the better
choice, as that allows userspace to probe for later added defines.

But if we use MS_* here that would be void anyway.

> +/* List of all mount_attr versions. */
> +#define MOUNT_ATTR_SIZE_VER0	24 /* sizeof first published struct */
> +#define MOUNT_ATTR_SIZE_LATEST	MOUNT_ATTR_SIZE_VER0

The _LATEST things is pretty dangerous as there basically is no safe
and correct way for userspace to use it.