Skip to contentSkip to conversation

sys_sendfile oops in 2.6.13?

PATCH Grzegorz Nosek 4 participants 6 messages Oct 11, 2005 – Oct 12, 2005 VFS

lore
Share & tools
lore
#0 lore raw

Hi all

I found an (IMHO) silly bug in do_sendfile in 2.6.13.x kernels (at least in 2.6.13.3 and .4, didn't backtrack to find where it originated). Without the patch all I apparently get from sys_sendfile is an oops due to a call in sys_sendfile with ppos being NULL. With the patch it works OK. Noticed in vsftpd.

The patch may apply with some fuzz as my kernel is somehwat patched but the gist of the patch is the same anyway

Regards,
Grzegorz Nosek

fs/read_write.c 2005-10-05 19:14:04.000000000 +0200 renamed from fs/read_write.c~ 2005-10-06 21:35:03.000000000 +0200 +1 -1
@@ -719,7 +719,7 @@
       current->syscr++;
       current->syscw++;
 
-       if (*ppos > max)
+       if (ppos && *ppos > max)
               retval = -EOVERFLOW;
 
 fput_out:
#1 lore raw

Grzegorz Nosek napsal(a):

Quoted 17 lines

Hi all

I found an (IMHO) silly bug in do_sendfile in 2.6.13.x kernels (at least in 2.6.13.3 and .4, didn't backtrack to find where it originated). Without the patch all I apparently get from sys_sendfile is an oops due to a call in sys_sendfile with ppos being NULL. With the patch it works OK. Noticed in vsftpd.

The patch may apply with some fuzz as my kernel is somehwat patched but the gist of the patch is the same anyway

Regards,
Grzegorz Nosek

fs/read_write.c 2005-10-05 19:14:04.000000000 +0200 line 724

Quoted 7 lines
fs/read_write.c 2005-10-05 19:14:04.000000000 +0200 renamed from fs/read_write.c~ 2005-10-06 21:35:03.000000000 +0200 +1 -1
@@ -719,7 +719,7 @@
      current->syscr++;
      current->syscw++;
 
-       if (*ppos > max)
+       if (ppos && *ppos > max)
 
 

I don't know the code surrounding this, but shouldn't be this (!ppos || *ppos > max)?

Quoted 5 lines
fs/read_write.c 2005-10-05 19:14:04.000000000 +0200 renamed from fs/read_write.c~ 2005-10-06 21:35:03.000000000 +0200 +0 -0
              retval = -EOVERFLOW;

fput_out:

Signature · 1 line
js
#2 lore raw
In reply to Grzegorz Nosek

fs/read_write.c 2005-10-05 19:14:04.000000000 +0200 line 722

Quoted 12 lines · Grzegorz Nosek

I found an (IMHO) silly bug in do_sendfile in 2.6.13.x kernels (at least in 2.6.13.3 and .4, didn't backtrack to find where it originated). Without the patch all I apparently get from sys_sendfile is an oops due to a call in sys_sendfile with ppos being NULL. With the patch it works OK. Noticed in vsftpd.

fs/read_write.c 2005-10-05 19:14:04.000000000 +0200 renamed from fs/read_write.c~ 2005-10-06 21:35:03.000000000 +0200 +1 -1
@@ -719,7 +719,7 @@
       current->syscr++;
       current->syscw++;
 
-       if (*ppos > max)
+       if (ppos && *ppos > max)

That change can't fix a bug in 2.6.13, because ppos is forced to be non-null further up the file:

622 static ssize_t do_sendfile(int out_fd, int in_fd, loff_t *ppos,
...

    647         if (!ppos)
    648                 ppos = &in_file->f_pos;
...
    684         pos = *ppos;
...
    701         current->syscr++;
    702         current->syscw++;
    703 
    704         if (*ppos > max)
    705                 retval = -EOVERFLOW;

(line numbers from 2.6.13.)

So there must be something else at work. Perhaps your patches?

Quoted 2 lines · Jiri Slaby

I don't know the code surrounding this, but shouldn't be this (!ppos || *ppos > max)?

That would be wrong, too; if it were valid to call in with ppos==0, you wouldn't want to return EOVERFLOW; and if ppos==0 were not valid and you wanted to return an error, EOVERFLOW would be the wrong error to return.

-andy

#3 lore raw

2005/10/12, Andy Isaacson <[email protected]>:

Quoted 35 lines · Grzegorz Nosek

fs/read_write.c 2005-10-05 19:14:04.000000000 +0200 line 722

Quoted 12 lines

I found an (IMHO) silly bug in do_sendfile in 2.6.13.x kernels (at least in 2.6.13.3 and .4, didn't backtrack to find where it originated). Without the patch all I apparently get from sys_sendfile is an oops due to a call in sys_sendfile with ppos being NULL. With the patch it works OK. Noticed in vsftpd.

fs/read_write.c 2005-10-05 19:14:04.000000000 +0200 renamed from fs/read_write.c~ 2005-10-06 21:35:03.000000000 +0200 +1 -1
@@ -719,7 +719,7 @@
       current->syscr++;
       current->syscw++;
 
-       if (*ppos > max)
+       if (ppos && *ppos > max)

That change can't fix a bug in 2.6.13, because ppos is forced to be non-null further up the file:

622 static ssize_t do_sendfile(int out_fd, int in_fd, loff_t *ppos,
...

    647         if (!ppos)
    648                 ppos = &in_file->f_pos;
...
    684         pos = *ppos;
...
    701         current->syscr++;
    702         current->syscw++;
    703
    704         if (*ppos > max)
    705                 retval = -EOVERFLOW;

(line numbers from 2.6.13.)

So there must be something else at work. Perhaps your patches?

On Tue, Oct 11, 2005 at 04:53:47PM +0200, Jiri Slaby wrote:

Quoted 9 lines
Quoted 2 lines

I don't know the code surrounding this, but shouldn't be this (!ppos || *ppos > max)?

That would be wrong, too; if it were valid to call in with ppos==0, you wouldn't want to return EOVERFLOW; and if ppos==0 were not valid and you wanted to return an error, EOVERFLOW would be the wrong error to return.

-andy

OK so I must have a broken kernel tree then. The lines you mention in my version belong to vfs_sendfile function which indeed ensures ppos is a valid pointer but do_sendfile is called from sys_sendfile(64). I'll find the patch that did the change (some must have, obviously) and report it there (probably linux-vserver is to blame)

This section of the file in vanilla 2.6.13.4 looks nothing like in my file and the 2.6.13.3 and 2.6.13.4 patches have no changes there so at least that's cleared up.

Regards,
Greg

#4 lore raw
In reply to Grzegorz Nosek

fs/read_write.c 2005-10-05 19:14:04.000000000 +0200 line 723

Quoted 9 lines · * Grzegorz Nosek ([email protected]) wrote:
fs/read_write.c 2005-10-05 19:14:04.000000000 +0200 renamed from fs/read_write.c~ 2005-10-06 21:35:03.000000000 +0200 +1 -1
@@ -719,7 +719,7 @@
       current->syscr++;
       current->syscw++;
 
-       if (*ppos > max)
+       if (ppos && *ppos > max)
               retval = -EOVERFLOW;

This doesn't make sense. ppos must not be NULL.

Code looks like this:

if (!ppos)

	 ppos = &in_file->f_pos;

...
pos = *ppos;
...
if (*ppos > max)

So it can't be NULL, and if somehow it were, the oops would've already happened.

#5 lore raw

2005/10/12, Chris Wright <[email protected]>:

Quoted 16 lines · * Grzegorz Nosek ([email protected]) wrote:

fs/read_write.c 2005-10-05 19:14:04.000000000 +0200 line 723

Quoted 9 lines
fs/read_write.c 2005-10-05 19:14:04.000000000 +0200 renamed from fs/read_write.c~ 2005-10-06 21:35:03.000000000 +0200 +1 -1
@@ -719,7 +719,7 @@
       current->syscr++;
       current->syscw++;
 
-       if (*ppos > max)
+       if (ppos && *ppos > max)
               retval = -EOVERFLOW;

This doesn't make sense. ppos must not be NULL.

Code looks like this:

if (!ppos)

Quoted 9 lines
         ppos = &in_file->f_pos;

...
pos = *ppos;
...
if (*ppos > max)

So it can't be NULL, and if somehow it were, the oops would've already happened.

Yeah, I know. It turned out to be a 3rd-party patch (vserver probably). My read_write.c looks very different and somehow I overlooked that while checking with vanilla 2.6.13.3. Sorry for the confusion.

Regards,
Greg