Re: patch for errno-issue (with soundcore)

Re: patch for errno-issue (with soundcore)

[Petr Vandrovec]

On 13 Jan 03 at 15:57, Thomas Schlichter wrote:
> On Mon, 13. Jan. 2003 16:13, Alan Cox wrote:
> > This actually shows a bug that has always been lurking. What if we load two
> > modules firmware at the same time. errno needs to be task private or we
> > perhaps need an errno_sem ?
> 
> OK, I think I see the problem now!
> But is soundcore the only place where 'errno' is used? Does this problem not 
> occur if any task modifies the errno value and an other one depends on its 
> previous value? I think this could happen even if no modules are used...

There is no problem currently, because of nobody uses errno value at
all (in the firmware loader), it is just that inline functions generated 
by syscallX() store error codes into errno...

Real problem is that firmware loader should use 
filp_open/vfs_read/filp_close (or sys_open/sys_llseek/sys_read/sys_close if 
you want to use fd interface, but filp_{open,close} and vfs_read are already 
exported for modules while sys_open/sys_llseek/sys_read are not).

As an alternative, do_mod_firmware_load should be standalone userspace
program executed through call_usermodehelper or something like that... 
Unfortunately we do not have an interface to distribute userspace binaries 
together with kernel (except initrd) yet, so it would require either
adding do_mod_firmware_load into module-init-tools, or some simillar
package required by 2.[56].x kernels.

Also adding "#define errno (current()->exit_code)" at the beginning of
sound_firmware.c (just below #define __KERNEL_SYSCALLS__) should do
the trick, but I do not recommend taking this path.
                                            Best regards,
                                                Petr Vandrovec
                                                vandrove@vc.cvut.cz
                                                

Re: patch for errno-issue (with soundcore)

[Thomas Schlichter]

On Mon, 13. Jan. 2003 17:33, Petr Vandrovec wrote:
> There is no problem currently, because of nobody uses errno value at
> all (in the firmware loader), it is just that inline functions generated
> by syscallX() store error codes into errno...
>
> Real problem is that firmware loader should use
> filp_open/vfs_read/filp_close (or sys_open/sys_llseek/sys_read/sys_close if
> you want to use fd interface, but filp_{open,close} and vfs_read are
> already exported for modules while sys_open/sys_llseek/sys_read are not).
>
> As an alternative, do_mod_firmware_load should be standalone userspace
> program executed through call_usermodehelper or something like that...
> Unfortunately we do not have an interface to distribute userspace binaries
> together with kernel (except initrd) yet, so it would require either
> adding do_mod_firmware_load into module-init-tools, or some simillar
> package required by 2.[56].x kernels.
>
> Also adding "#define errno (current()->exit_code)" at the beginning of
> sound_firmware.c (just below #define __KERNEL_SYSCALLS__) should do
> the trick, but I do not recommend taking this path.
>                                             Best regards,
>                                                 Petr Vandrovec
>                                                 vandrove@vc.cvut.cz

First of all a big THANKS!

Now, at least, I understood the real problem...
'errno' is only needed because open(), close(), llseek() and read() are used, 
which do a syscall and the return code is written to a variable called 
'errno'. This is NOT the 'errno' variable defined in lib/errno.c... OK.

And so the real problem is using these functions and as we ARE in the kernel 
mode we do not really need the kernel traps...

So after realizing the real problem I think I am not the right person to fix 
it. But if you want I could try...

Sincerely yours
   Thomas Schlichter

[PATCH] Re: patch for errno-issue (with soundcore)

[Petr Vandrovec]

> 
> Now, at least, I understood the real problem...
> 'errno' is only needed because open(), close(), llseek() and read() are used, 
> which do a syscall and the return code is written to a variable called 
> 'errno'. This is NOT the 'errno' variable defined in lib/errno.c... OK.
> 
> And so the real problem is using these functions and as we ARE in the kernel 
> mode we do not really need the kernel traps...

Hi, 
  can someone with sound card requiring firmware download (OSS's trix,
sb_common, msnd_pinnacle, sscape, maui or pss) verify that code below
works? I know that it compiles without warnings, and looks obviously
correct (to me), but...

  I think that I do not need any locking around reading i_size, as
nothing bad should happen if size changes, you'll get damaged firmware
upload at worst, and if you are updating firmware file while kernel 
reads it, you deserve it (besides that vfs_read is one who will do final 
check). And sys_stat (getattr) does not take any lock too, unless I 
missed something.

  Patch is for 2.5.57.

  BTW, should I convert it to goto? It was really annoying that I had to
fix 3 error paths instead of one ;-)
					Best regards,
						Petr Vandrovec
						vandrove@vc.cvut.cz	

diff -urdN linux/sound/sound_firmware.c linux/sound/sound_firmware.c
--- linux/sound/sound_firmware.c	2003-01-13 18:24:36.000000000 +0000
+++ linux/sound/sound_firmware.c	2003-01-13 18:45:12.000000000 +0000
@@ -9,39 +9,40 @@
 
 static int do_mod_firmware_load(const char *fn, char **fp)
 {
-	int fd;
+	struct file* filp;
 	long l;
 	char *dp;
+	loff_t pos;
 
-	fd = open(fn, 0, 0);
-	if (fd == -1)
+	filp = filp_open(fn, 0, 0);
+	if (IS_ERR(filp))
 	{
 		printk(KERN_INFO "Unable to load '%s'.\n", fn);
 		return 0;
 	}
-	l = lseek(fd, 0L, 2);
+	l = filp->f_dentry->d_inode->i_size;
 	if (l <= 0 || l > 131072)
 	{
 		printk(KERN_INFO "Invalid firmware '%s'\n", fn);
-		sys_close(fd);
+		filp_close(filp, current->files);
 		return 0;
 	}
-	lseek(fd, 0L, 0);
 	dp = vmalloc(l);
 	if (dp == NULL)
 	{
 		printk(KERN_INFO "Out of memory loading '%s'.\n", fn);
-		sys_close(fd);
+		filp_close(filp, current->files);
 		return 0;
 	}
-	if (read(fd, dp, l) != l)
+	pos = 0;
+	if (vfs_read(filp, dp, l, &pos) != l)
 	{
 		printk(KERN_INFO "Failed to read '%s'.\n", fn);
 		vfree(dp);
-		sys_close(fd);
+		filp_close(filp, current->files);
 		return 0;
 	}
-	close(fd);
+	filp_close(filp, current->files);
 	*fp = dp;
 	return (int) l;
 }