kernel space getcwd()? (using current() to find out cwd)

kernel space getcwd()? (using current() to find out cwd)

[Michael L. Welles]

This is probably a stupid question, and probably directed to the wrong
list.  Apologies in advance, but I'm stumped

I've been working on a kernel module to report on "changed files". It
works just fine -- I wrap the orignal system calls with my
replacements which queue the filenames being modified, and when
another proccess reads from the device or proc entry, they get a nice
snapshot of what's going on in the system -- except that all the paths
are relative to the calling process.

So, a little ignorance being a dangerous thing, I thought I'd be clever
and manually reconstruct the full path by walking up
current->fs->pwd->d_parent and padding d_name to the filename until it
hits root.

Unfortunatly, this approach causes kernel panics.  e.g., the attached
code snippet will inevitably bring down the machine if I call it
during in my replacement open, mkdir, rmdir, unlink routines -- and
tehy all work fine without itq. 

What am I not getting? I do see, before I go down, that there's a few
occasions where current() is NULL... 

Apologies in advance for a wordy, probably stupid question, but I'm
stumped.   

If this is not the right approach for what I'm trying to do (e.g. a
kernel space getcwd()), can someone point me to something else I can try?

Thanks in advance, 

Mike Welles


----------------------------------------

(this is the greatly reduced version which does nothing but try and
reference current->fs->pwd)

void fill_full_path(char *name)
{ 
    if (current==NULL) 
    { 
#ifdef DEBUG
	  printk("ERROR! current == NULL\n"); 
#endif
	  return; 
    }

  if (current->fs==NULL) 
    { 
#ifdef DEBUG
	  printk("ERROR! current-> == NULL\n"); 
#endif
	  return; 
    }

  if (current->fs->pwd==NULL) 
    { 
#ifdef DEBUG
	  printk("ERROR! current->fs->pwd == NULL\n"); 
#endif
	  return; 
    }

    return; 
}








This is probably a stupid question: I've been working on a kernel
module to report on "changed files". 

It works just fine -- I wrap the orignal system calls with my
replacements which queue the filenames being modified, and when
another proccess reads from the device or proc entry, they get a nice
snapshot of what's going on in the system -- except that all the paths
are relative to the calling process. 

So, a little ignorance being a dangerous thing, I thought I'd be clever
and manually reconstruct the full path by walking up
current->fs->pwd->d_parent and padding d_name to the filename until it
hits root.

Unfortunatly, this approach causes kernel panics.  e.g., the attached
code snippet will inevitably bring down the machine if I call it
during in my replacement open, mkdir, rmdir, unlink routines -- and
tehy all work fine without itq. 

What am I not getting? I do see, before I go down, that there's a few
occasions where current() is NULL... 

Apologies in advance for a wordy, probably stupid question, but I'm
stumped.   

If this is not the right approach for what I'm trying to do (e.g. a
kernel space getcwd()), can someone point me to where I should look?

Thanks in advance, 

Mike Welles


----------------------------------------

(this is the greatly reduced version which does nothing but try and
reference current->fs->pwd)

void fill_full_path(char *name)
{ 
    if (current==NULL) 
    { 
#ifdef DEBUG
	  printk("ERROR! current == NULL\n"); 
#endif
	  return; 
    }

  if (current->fs==NULL) 
    { 
#ifdef DEBUG
	  printk("ERROR! current-> == NULL\n"); 
#endif
	  return; 
    }

  if (current->fs->pwd==NULL) 
    { 
#ifdef DEBUG
	  printk("ERROR! current->fs->pwd == NULL\n"); 
#endif
	  return; 
    }

    return; 
}

Re: kernel space getcwd()? (using current() to find out cwd)

[Christoph Hellwig]

In article <15067.30060.436790.458922@bangstate.com> you wrote:
> So, a little ignorance being a dangerous thing, I thought I'd be clever
> and manually reconstruct the full path by walking up
> current->fs->pwd->d_parent and padding d_name to the filename until it
> hits root.
>
> Unfortunatly, this approach causes kernel panics.  e.g., the attached
> code snippet will inevitably bring down the machine if I call it
> during in my replacement open, mkdir, rmdir, unlink routines -- and
> tehy all work fine without itq. 


Use d_path.  NOTE:  the buffer in which the pathname is returned is
the return value of the function and _not_ the buffer you gave to it.

	Christoph

-- 
Of course it doesn't work. We've performed a software upgrade.

Re: kernel space getcwd()? (using current() to find out cwd)

[Brian J. Watson]

> This is probably a stupid question, and probably directed to the wrong
> list. Apologies in advance, but I'm stumped
> 
> I've been working on a kernel module to report on "changed files". It
> works just fine -- I wrap the orignal system calls with my
> [...]


At least in the 2.4 kernels, there's already a __d_path() routine (fs/dcache.c)
that builds the pathname using the mechanism you discussed.

Here's one way you could use it:

char *
kgetcwd()
{
	char *path = (char *) __get_free_page(GFP_USER);
        struct vfsmnt *pwdmnt;
        struct dentry *pwd;

        if (!path)
                return ERR_PTR(-ENOMEM);

        read_lock(&current->fs->lock);
        pwdmnt = mntget(current->fs->pwdmnt);
        pwd = dget(current->fs->pwd);
        read_unlock(&current->fs->lock);

        spin_lock(&dcache_lock);
        path = __d_path(pwd, pwdmnt, NULL, NULL, path, PAGE_SIZE);
        spin_unlock(&dcache_lock);

        mntput(pwdmnt);
        dput(pwd);

        return path;
}


If you only want the pathname back to the process root, use d_path() instead
(and don't grab the dcache_lock).

When you're done with path, free it with free_page() and not kfree().

BTW, I'm not subscribed to the kernel mailing list (I just read it on the web),
so please copy me on any response.


--
Brian Watson
Compaq Computer

Re: kernel space getcwd()? (using current() to find out cwd)

[Brian J. Watson]

"Brian J. Watson" wrote:
>         path = __d_path(pwd, pwdmnt, NULL, NULL, path, PAGE_SIZE);


Oops! That's no good. Here's the new and improved version:

char *
kgetcwd(char **bufp)
{
        char *path, *buf = (char *) __get_free_page(GFP_USER);
        struct vfsmnt *pwdmnt;
        struct dentry *pwd;

        *bufp = NULL;
        if (!buf)
                return ERR_PTR(-ENOMEM);

        read_lock(&current->fs->lock);
        pwdmnt = mntget(current->fs->pwdmnt);
        pwd = dget(current->fs->pwd);
        read_unlock(&current->fs->lock);

        spin_lock(&dcache_lock);
        path = __d_path(pwd, pwdmnt, NULL, NULL, buf, PAGE_SIZE);
        spin_unlock(&dcache_lock);

        mntput(pwdmnt);
        dput(pwd);

        *bufp = buf;
        return path;
}


The returned pointer is for the beginning of the path name. The pointer filled
into bufp is for the beginning of the allocated space. To deallocate, call
free_page() on the value in bufp.

The reason for the distinction is that __d_path builds the pathname from the end
of the buffer, working its way back toward the beginning. Rarely will the string
begin at the same address as the allocated buffer.


--
Brian Watson
Compaq Computer