Re: 'native files', 'object fingerprints' [was: sendpath()]

[dean gaudet]

On Tue, 16 Jan 2001, Ingo Molnar wrote:

> But even user-space code could use 'native files', via the following, safe
> mechanizm:

so here's an alternative to ingo's proposal which i think solves some of
the other objections raised.  it's something i've proposed in the past
under the name "extended file handles".

struct extended_file_permission {
	int refcount;
	some form of mutex to protect refcount;
        some list structure head;
};

struct extended_file {
	struct file *file;
	struct extended_file_permission *perm;
        whatever list foo is needed to link with extended_file_perm above;
};

if you allocate a few huge arrays of struct extended_file, then you can
verify if a pointer passed from user space fits into one of those arrays
pretty quickly.

struct task has a struct extended_file_permission * added to it to
indicate which perm struct that task is associated with.

so you just compare the f->perm to current->extended_file_perm and you
know if the task is allowed to use it or not.

clone() allows you to create tasks sharing the same
extended_file_permissions.

fork()/exec() would create new extended_file_perms -- which implicitly
causes all those files to be closed.  this gives you pretty light cgi
fork()/exec() off a main "process" which is handling thousands of sockets.

i also proposed various methods of doing O_foo flag inheritance... but the
above is more interesting.

-dean