[PATCH 0/1] ns: introduce proc_get_ns_by_fd()

Konstantin Khlebnikov khlebnikov at yandex-team.ru
Mon Sep 28 08:21:07 UTC 2015


On 25.09.2015 20:56, Oleg Nesterov wrote:
> On 09/25, Konstantin Khlebnikov wrote:
>>
>> +struct ns_common *proc_ns_fdget(int fd, int nstype, struct fd *fd_ref)
>>   {
>> -	struct file *file;
>> +	struct ns_common *ns;
>> +	struct fd f;
>>
>> -	file = fget(fd);
>> -	if (!file)
>> +	f = fdget(fd);
>> +	if (!f.file)
>>   		return ERR_PTR(-EBADF);
>>
>> -	if (file->f_op != &ns_file_operations)
>> +	if (f.file->f_op != &ns_file_operations)
>> +		goto out_invalid;
>> +
>> +	ns = get_proc_ns(file_inode(f.file));
>> +	if (nstype && (ns->ops->type != nstype))
>>   		goto out_invalid;
>>
>> -	return file;
>> +	*fd_ref = f;
>> +	return ns;
>>
>>   out_invalid:
>> -	fput(file);
>> +	fdput(f);
>>   	return ERR_PTR(-EINVAL);
>>   }
>
> Well yes, fdget() makes sense but this is minor.
>
> Honestly, I do not really like the new helper... I understand this
> is subjective, so I won't insist. But how about 1/1? We do not need
> fd/file at all. With this patch your sys_getvpid() can just use
> proc_get_ns_by_fd(fd, CLONE_NEWPID) and put_pid_ns().

Hmm. My version has 0 or 2 atomic ops per get-put sequence.
Your version: 2 or 4 atomic ops. Plus even in worst case pinning by
struct fd theoretically scales better because it touches refcount at
struct file: there're might be many of them for one namespace.

>
> Eric, what do you think?
>
> See also "TODO" in the changelog.
>
> Oleg.
>


-- 
Konstantin


More information about the Containers mailing list