[PATCH] seq_file: make seq_lseek accept SEEK_END

[PATCH] seq_file: make seq_lseek accept SEEK_END

[Alexey Dobriyan]

Apologies for delay, such simple thing should have been sent long ago.
Joakim, please, confirm.
-------------------------------------------
[PATCH] seq_file: make seq_lseek accept SEEK_END

and pretend seq_files have zero length. This should be enough
to fix busybox start-stop-daemon:
http://marc.info/?t=120836691600002&r=1&w=2

It does xlseek(fd, 0, SEEK_END) to estimate amount of memory to malloc
but satisfied with 0. Sudden -EINVAL from lseek(2) breaks it.

X-Introduced-By: f16278c679aa72e28288435b313ba2d4494d6be5
Signed-off-by: Alexey Dobriyan <adobriyan@parallels.com>
---

 fs/seq_file.c |    2 ++
 1 file changed, 2 insertions(+)

--- a/fs/seq_file.c
+++ b/fs/seq_file.c
@@ -254,6 +254,8 @@ loff_t seq_lseek(struct file *file, loff_t offset, int origin)
 	switch (origin) {
 		case 1:
 			offset += file->f_pos;
+		case 2:
+			/* pretend it's zero length */
 		case 0:
 			if (offset < 0)
 				break;

Re: [PATCH] seq_file: make seq_lseek accept SEEK_END

[Joakim Tjernlund]

busybox has been fixed, at least in trunk, not to do 
xlseek(fd, 0, SEEK_END) anymore to address this problem.
Perhaps there are other apps out there that needs this too, but I don't
know of any. Older busyboxes needs this change though.
I have moved on and don't have a setup handy where I can test this.
Perhaps someone at the bb list has, CC:ing bb list.

 Jocke

On Mon, 2008-06-09 at 15:01 +0400, Alexey Dobriyan wrote:
> Apologies for delay, such simple thing should have been sent long ago.
> Joakim, please, confirm.
> -------------------------------------------
> [PATCH] seq_file: make seq_lseek accept SEEK_END
> 
> and pretend seq_files have zero length. This should be enough
> to fix busybox start-stop-daemon:
> http://marc.info/?t=120836691600002&r=1&w=2
> 
> It does xlseek(fd, 0, SEEK_END) to estimate amount of memory to malloc
> but satisfied with 0. Sudden -EINVAL from lseek(2) breaks it.
> 
> X-Introduced-By: f16278c679aa72e28288435b313ba2d4494d6be5
> Signed-off-by: Alexey Dobriyan <adobriyan@parallels.com>
> ---
> 
>  fs/seq_file.c |    2 ++
>  1 file changed, 2 insertions(+)
> 
> --- a/fs/seq_file.c
> +++ b/fs/seq_file.c
> @@ -254,6 +254,8 @@ loff_t seq_lseek(struct file *file, loff_t offset, int origin)
>  	switch (origin) {
>  		case 1:
>  			offset += file->f_pos;
> +		case 2:
> +			/* pretend it's zero length */
>  		case 0:
>  			if (offset < 0)
>  				break;
> 
> 

Re: [PATCH] seq_file: make seq_lseek accept SEEK_END

[Andrew Morton]

On Mon, 09 Jun 2008 13:52:55 +0200
Joakim Tjernlund <joakim.tjernlund@transmode.se> wrote:

> busybox has been fixed, at least in trunk, not to do 
> xlseek(fd, 0, SEEK_END) anymore to address this problem.
> Perhaps there are other apps out there that needs this too, but I don't
> know of any. Older busyboxes needs this change though.
> I have moved on and don't have a setup handy where I can test this.
> Perhaps someone at the bb list has, CC:ing bb list.
> 

Nevertheless it'd be nice to fix old versions.

To some extent that depends on when we broke it.  See below.

> 
> On Mon, 2008-06-09 at 15:01 +0400, Alexey Dobriyan wrote:
> > Apologies for delay, such simple thing should have been sent long ago.
> > Joakim, please, confirm.
> > -------------------------------------------
> > [PATCH] seq_file: make seq_lseek accept SEEK_END
> > 
> > and pretend seq_files have zero length. This should be enough
> > to fix busybox start-stop-daemon:
> > http://marc.info/?t=120836691600002&r=1&w=2
> > 
> > It does xlseek(fd, 0, SEEK_END) to estimate amount of memory to malloc
> > but satisfied with 0. Sudden -EINVAL from lseek(2) breaks it.
> > 
> > X-Introduced-By: f16278c679aa72e28288435b313ba2d4494d6be5
> > Signed-off-by: Alexey Dobriyan <adobriyan@parallels.com>
> > ---
> > 
> >  fs/seq_file.c |    2 ++
> >  1 file changed, 2 insertions(+)
> > 
> > --- a/fs/seq_file.c
> > +++ b/fs/seq_file.c
> > @@ -254,6 +254,8 @@ loff_t seq_lseek(struct file *file, loff_t offset, int origin)
> >  	switch (origin) {
> >  		case 1:
> >  			offset += file->f_pos;
> > +		case 2:
> > +			/* pretend it's zero length */
> >  		case 0:
> >  			if (offset < 0)
> >  				break;
> > 
> > 

Is returning zero logical, given what we do with SEEK_SET?

afait we _could_ just do

	offset = <largenum>
	/* fall through */

and let the SEEK_SET code do its thing.  But I might have misread it. 
But still, it looks like it'd be possible to return a true(ish) value
from SEEK_END?


> > X-Introduced-By: f16278c679aa72e28288435b313ba2d4494d6be5

Please don't do that.  Instead, quote the commit ID _and_ the first few
lines of the commit itself, like:

commit f16278c679aa72e28288435b313ba2d4494d6be5
Author: Hans Rosenfeld <hans.rosenfeld@amd.com>
Date:   Fri Mar 21 18:46:59 2008 -0500

    Change pagemap output format to allow for future reporting of huge pages

And here we see one reason: that commit cannot be the one which
introduced this problem. 

Re: [PATCH] seq_file: make seq_lseek accept SEEK_END

[Alexey Dobriyan]

On Wed, Jun 11, 2008 at 02:29:09PM -0700, Andrew Morton wrote:
> On Mon, 09 Jun 2008 13:52:55 +0200
> Joakim Tjernlund <joakim.tjernlund@transmode.se> wrote:
> 
> > busybox has been fixed, at least in trunk, not to do 
> > xlseek(fd, 0, SEEK_END) anymore to address this problem.
> > Perhaps there are other apps out there that needs this too, but I don't
> > know of any. Older busyboxes needs this change though.
> > I have moved on and don't have a setup handy where I can test this.
> > Perhaps someone at the bb list has, CC:ing bb list.
> > 
> 
> Nevertheless it'd be nice to fix old versions.
> 
> To some extent that depends on when we broke it.  See below.
> 
> > 
> > On Mon, 2008-06-09 at 15:01 +0400, Alexey Dobriyan wrote:
> > > Apologies for delay, such simple thing should have been sent long ago.
> > > Joakim, please, confirm.
> > > -------------------------------------------
> > > [PATCH] seq_file: make seq_lseek accept SEEK_END
> > > 
> > > and pretend seq_files have zero length. This should be enough
> > > to fix busybox start-stop-daemon:
> > > http://marc.info/?t=120836691600002&r=1&w=2
> > > 
> > > It does xlseek(fd, 0, SEEK_END) to estimate amount of memory to malloc
> > > but satisfied with 0. Sudden -EINVAL from lseek(2) breaks it.
> > > 
> > > X-Introduced-By: f16278c679aa72e28288435b313ba2d4494d6be5
> > > Signed-off-by: Alexey Dobriyan <adobriyan@parallels.com>
> > > ---
> > > 
> > >  fs/seq_file.c |    2 ++
> > >  1 file changed, 2 insertions(+)
> > > 
> > > --- a/fs/seq_file.c
> > > +++ b/fs/seq_file.c
> > > @@ -254,6 +254,8 @@ loff_t seq_lseek(struct file *file, loff_t offset, int origin)
> > >  	switch (origin) {
> > >  		case 1:
> > >  			offset += file->f_pos;
> > > +		case 2:
> > > +			/* pretend it's zero length */
> > >  		case 0:
> > >  			if (offset < 0)
> > >  				break;
> > > 
> > > 
> 
> Is returning zero logical, given what we do with SEEK_SET?

Definitely illogical. :-)

> afait we _could_ just do
> 
> 	offset = <largenum>
> 	/* fall through */
> 
> and let the SEEK_SET code do its thing.  But I might have misread it. 

If it's largenum, not enough output will be generated during each
iteration until kmalloc() will refuse to give more memory.

> But still, it looks like it'd be possible to return a true(ish) value
> from SEEK_END?

I think this is doable.

On the other side is
a) ugliness (remember how much output each item generates, remember ->index,
   allocate memory to do that),
b) SEEK_END reported and reports zero-length for other proc files
   (and sysctls), and nobody really cared.