Re: [PATCH v2] kernel/ucounts: expose count of inotify watches in use

[Andrew Morton]

On Fri,  1 Feb 2019 21:39:59 +0100 Albert Vaca Cintora <albertvaka@xxxxxxxxx> wrote:

> Adds a readonly 'current_inotify_watches' entry to the user sysctl table.
> The handler for this entry is a custom function that ends up calling
> proc_dointvec. Said sysctl table already contains 'max_inotify_watches'
> and it gets mounted under /proc/sys/user/.
> 
> Inotify watches are a finite resource, in a similar way to available file
> descriptors. The motivation for this patch is to be able to set up
> monitoring and alerting before an application starts failing because
> it runs out of inotify watches.

Matthias said "Albert found this problem while working on montitoring
software, so it fixes a real problem out there", so please include full
details of the problem which you encountered so that we are better able
to understand the value of the patch.

>
> ...
>
>  kernel/ucount.c | 29 +++++++++++++++++++++++++++++

Documentation, please.  Documentation/filesystems/inotify.txt and/or
Documentation/filesystems/proc.txt.

Also, max_inotify_instances (at least) also appears to be undocumented,
so it would be good to address this as well while you're in there.

> 
> diff --git a/kernel/ucount.c b/kernel/ucount.c
> index f48d1b6376a4..d8b11e53f098 100644
> --- a/kernel/ucount.c
> +++ b/kernel/ucount.c
> @@ -57,6 +57,11 @@ static struct ctl_table_root set_root = {
>  	.permissions = set_permissions,
>  };
>  
> +#ifdef CONFIG_INOTIFY_USER
> +int proc_read_inotify_watches(struct ctl_table *table, int write,
> +		     void __user *buffer, size_t *lenp, loff_t *ppos);
> +#endif

The ifdefs aren't really needed.  And this should be in a header file
if it is indeed to be non-static.

But it should be static, in which case the ifdef will be needed to
prevent a warning.  It's kinda irksome and perhaps it would be better
to move proc_read_inotify_watches() to be ahead of user_table[].


>  static int zero = 0;
>  static int int_max = INT_MAX;
>  #define UCOUNT_ENTRY(name)				\
> @@ -79,6 +84,12 @@ static struct ctl_table user_table[] = {
>  #ifdef CONFIG_INOTIFY_USER
>  	UCOUNT_ENTRY("max_inotify_instances"),
>  	UCOUNT_ENTRY("max_inotify_watches"),
> +	{
> +		.procname	= "current_inotify_watches",
> +		.maxlen		= sizeof(int),
> +		.mode		= 0444,
> +		.proc_handler	= proc_read_inotify_watches,
> +	},
>  #endif
>  	{ }
>  };
> @@ -226,6 +237,24 @@ void dec_ucount(struct ucounts *ucounts, enum ucount_type type)
>  	put_ucounts(ucounts);
>  }
>  
> +#ifdef CONFIG_INOTIFY_USER
> +int proc_read_inotify_watches(struct ctl_table *table, int write,
> +		     void __user *buffer, size_t *lenp, loff_t *ppos)
> +{
> +	struct ucounts *ucounts;
> +	struct ctl_table fake_table;
> +	int count;
> +
> +	ucounts = get_ucounts(current_user_ns(), current_euid());

get_ucounts() can return NULL.  The kernel will crash.

> +	count = atomic_read(&ucounts->ucount[UCOUNT_INOTIFY_WATCHES]);
> +	put_ucounts(ucounts);
> +
> +	fake_table.data = &count;
> +	fake_table.maxlen = sizeof(count);
> +	return proc_dointvec(&fake_table, write, buffer, lenp, ppos);
> +}
> +#endif