[RESEND] nfs: prevent truncate on active swapfile

Commit Message

Omar Sandoval Jan. 8, 2015, 9:18 a.m. UTC

Most filesystems prevent truncation of an active swapfile by way of
inode_newsize_ok, called from inode_change_ok. NFS doesn't call either
from nfs_setattr, presumably because most of these checks are expected
to be done server-side. However, the IS_SWAPFILE check can only be done
client-side, and truncating a swapfile can't possibly be good.

Signed-off-by: Omar Sandoval <[e-mail]>
---
Hi, Trond,

Now that the holidays are over, could you take a look at this? It was
generated against v3.19-rc3.

Thanks!

 fs/nfs/inode.c | 7 ++++++-
 1 file changed, 6 insertions(+), 1 deletion(-)

Comments

Trond Myklebust Jan. 10, 2015, 10:08 p.m. UTC | #1

Hi Omar,

On Thu, Jan 8, 2015 at 4:18 AM, Omar Sandoval <[e-mail]> wrote:
> Most filesystems prevent truncation of an active swapfile by way of
> inode_newsize_ok, called from inode_change_ok. NFS doesn't call either
> from nfs_setattr, presumably because most of these checks are expected
> to be done server-side. However, the IS_SWAPFILE check can only be done
> client-side, and truncating a swapfile can't possibly be good.
>
> Signed-off-by: Omar Sandoval <[e-mail]>
> ---
> Hi, Trond,
>
> Now that the holidays are over, could you take a look at this? It was
> generated against v3.19-rc3.
>
> Thanks!
>
>  fs/nfs/inode.c | 7 ++++++-
>  1 file changed, 6 insertions(+), 1 deletion(-)
>
> diff --git a/fs/nfs/inode.c b/fs/nfs/inode.c
> index 4bffe63..9205513 100644
> --- a/fs/nfs/inode.c
> +++ b/fs/nfs/inode.c
> @@ -506,10 +506,15 @@ nfs_setattr(struct dentry *dentry, struct iattr *attr)
>                 attr->ia_valid &= ~ATTR_MODE;
>
>         if (attr->ia_valid & ATTR_SIZE) {
> +               loff_t i_size;
> +
>                 BUG_ON(!S_ISREG(inode->i_mode));
>
> -               if (attr->ia_size == i_size_read(inode))
> +               i_size = i_size_read(inode);
> +               if (attr->ia_size == i_size)
>                         attr->ia_valid &= ~ATTR_SIZE;
> +               else if (attr->ia_size < i_size && IS_SWAPFILE(inode))
> +                       return -ETXTBSY;
>         }
>
>         /* Optimization: if the end result is no change, don't RPC */
> --
> 2.2.1
>

I agree that truncating a swap file is bad, however as you point out,
this really only addresses the case on the client that knows about
this being a swap file.
I'll take the patch, but I'm wondering if we couldn't do better in the
case where we're using NFSv4 by using share deny modes (which are
enforced by the server). The problem is that there appears to be
nothing in swapon() that tells the filesystem this is an open of a
swap file...

Cheers
  Trond

Omar Sandoval Jan. 12, 2015, 7:43 p.m. UTC | #2

On Sat, Jan 10, 2015 at 05:08:03PM -0500, Trond Myklebust wrote:
> Hi Omar,
> 
> On Thu, Jan 8, 2015 at 4:18 AM, Omar Sandoval <[e-mail]> wrote:
> > Most filesystems prevent truncation of an active swapfile by way of
> > inode_newsize_ok, called from inode_change_ok. NFS doesn't call either
> > from nfs_setattr, presumably because most of these checks are expected
> > to be done server-side. However, the IS_SWAPFILE check can only be done
> > client-side, and truncating a swapfile can't possibly be good.
> >
> > Signed-off-by: Omar Sandoval <[e-mail]>
> > ---
> > Hi, Trond,
> >
> > Now that the holidays are over, could you take a look at this? It was
> > generated against v3.19-rc3.
> >
> > Thanks!
> >
> >  fs/nfs/inode.c | 7 ++++++-
> >  1 file changed, 6 insertions(+), 1 deletion(-)
> >
> > diff --git a/fs/nfs/inode.c b/fs/nfs/inode.c
> > index 4bffe63..9205513 100644
> > --- a/fs/nfs/inode.c
> > +++ b/fs/nfs/inode.c
> > @@ -506,10 +506,15 @@ nfs_setattr(struct dentry *dentry, struct iattr *attr)
> >                 attr->ia_valid &= ~ATTR_MODE;
> >
> >         if (attr->ia_valid & ATTR_SIZE) {
> > +               loff_t i_size;
> > +
> >                 BUG_ON(!S_ISREG(inode->i_mode));
> >
> > -               if (attr->ia_size == i_size_read(inode))
> > +               i_size = i_size_read(inode);
> > +               if (attr->ia_size == i_size)
> >                         attr->ia_valid &= ~ATTR_SIZE;
> > +               else if (attr->ia_size < i_size && IS_SWAPFILE(inode))
> > +                       return -ETXTBSY;
> >         }
> >
> >         /* Optimization: if the end result is no change, don't RPC */
> > --
> > 2.2.1
> >
> 
> I agree that truncating a swap file is bad, however as you point out,
> this really only addresses the case on the client that knows about
> this being a swap file.
> I'll take the patch,

Thanks, I appreciate it.

> but I'm wondering if we couldn't do better in the
> case where we're using NFSv4 by using share deny modes (which are
> enforced by the server). The problem is that there appears to be
> nothing in swapon() that tells the filesystem this is an open of a
> swap file...

Yeah, it would be nice for completeness to prevent one client from
truncating another client's swapfile. However, I'd hope that anyone
using swap-over-NFS on a shared NFS mount would take the necessary
precautions in terms of permissions, etc. to prevent someone from doing
that. Also, since the failure mode of truncating an NFS swapfile is a
corrupt swapfile rather than a corrupt filesystem (like on a local
filesystem), it's probably okay to just deal with the low-hanging fruit
for now.

Thanks!

> 
> Cheers
>   Trond
> -- 
> Trond Myklebust
> Linux NFS client maintainer, PrimaryData
> [e-mail]

Omar Sandoval Jan. 23, 2015, 4:07 a.m. UTC | #3

On Mon, Jan 12, 2015 at 11:43:02AM -0800, Omar Sandoval wrote:
> On Sat, Jan 10, 2015 at 05:08:03PM -0500, Trond Myklebust wrote:
> > Hi Omar,
> > 
> > On Thu, Jan 8, 2015 at 4:18 AM, Omar Sandoval <[e-mail]> wrote:
> > > Most filesystems prevent truncation of an active swapfile by way of
> > > inode_newsize_ok, called from inode_change_ok. NFS doesn't call either
> > > from nfs_setattr, presumably because most of these checks are expected
> > > to be done server-side. However, the IS_SWAPFILE check can only be done
> > > client-side, and truncating a swapfile can't possibly be good.
> > >
> > > Signed-off-by: Omar Sandoval <[e-mail]>
> > > ---
> > > Hi, Trond,
> > >
> > > Now that the holidays are over, could you take a look at this? It was
> > > generated against v3.19-rc3.
> > >
> > > Thanks!
> > >
> > >  fs/nfs/inode.c | 7 ++++++-
> > >  1 file changed, 6 insertions(+), 1 deletion(-)
> > >
> > > diff --git a/fs/nfs/inode.c b/fs/nfs/inode.c
> > > index 4bffe63..9205513 100644
> > > --- a/fs/nfs/inode.c
> > > +++ b/fs/nfs/inode.c
> > > @@ -506,10 +506,15 @@ nfs_setattr(struct dentry *dentry, struct iattr *attr)
> > >                 attr->ia_valid &= ~ATTR_MODE;
> > >
> > >         if (attr->ia_valid & ATTR_SIZE) {
> > > +               loff_t i_size;
> > > +
> > >                 BUG_ON(!S_ISREG(inode->i_mode));
> > >
> > > -               if (attr->ia_size == i_size_read(inode))
> > > +               i_size = i_size_read(inode);
> > > +               if (attr->ia_size == i_size)
> > >                         attr->ia_valid &= ~ATTR_SIZE;
> > > +               else if (attr->ia_size < i_size && IS_SWAPFILE(inode))
> > > +                       return -ETXTBSY;
> > >         }
> > >
> > >         /* Optimization: if the end result is no change, don't RPC */
> > > --
> > > 2.2.1
> > >
> > 
> > I agree that truncating a swap file is bad, however as you point out,
> > this really only addresses the case on the client that knows about
> > this being a swap file.
> > I'll take the patch,
> 
> Thanks, I appreciate it.
> 
> > but I'm wondering if we couldn't do better in the
> > case where we're using NFSv4 by using share deny modes (which are
> > enforced by the server). The problem is that there appears to be
> > nothing in swapon() that tells the filesystem this is an open of a
> > swap file...
> 
> Yeah, it would be nice for completeness to prevent one client from
> truncating another client's swapfile. However, I'd hope that anyone
> using swap-over-NFS on a shared NFS mount would take the necessary
> precautions in terms of permissions, etc. to prevent someone from doing
> that. Also, since the failure mode of truncating an NFS swapfile is a
> corrupt swapfile rather than a corrupt filesystem (like on a local
> filesystem), it's probably okay to just deal with the low-hanging fruit
> for now.
> 
> Thanks!
> 
> > 
> > Cheers
> >   Trond
> > -- 
> > Trond Myklebust
> > Linux NFS client maintainer, PrimaryData
> > [e-mail]
> 
> -- 
> Omar

Hi, Trond,

Are you still planning on taking this patch? I didn't see it in your
last pull request to Linus.

Thanks,

Trond Myklebust Jan. 23, 2015, 1:07 p.m. UTC | #4

On Thu, Jan 22, 2015 at 11:07 PM, Omar Sandoval <[e-mail]> wrote:
> On Mon, Jan 12, 2015 at 11:43:02AM -0800, Omar Sandoval wrote:
>> On Sat, Jan 10, 2015 at 05:08:03PM -0500, Trond Myklebust wrote:
>> > Hi Omar,
>> >
>> > On Thu, Jan 8, 2015 at 4:18 AM, Omar Sandoval <[e-mail]> wrote:
>> > > Most filesystems prevent truncation of an active swapfile by way of
>> > > inode_newsize_ok, called from inode_change_ok. NFS doesn't call either
>> > > from nfs_setattr, presumably because most of these checks are expected
>> > > to be done server-side. However, the IS_SWAPFILE check can only be done
>> > > client-side, and truncating a swapfile can't possibly be good.
>> > >
>> > > Signed-off-by: Omar Sandoval <[e-mail]>
>> > > ---
>> > > Hi, Trond,
>> > >
>> > > Now that the holidays are over, could you take a look at this? It was
>> > > generated against v3.19-rc3.
>> > >
>> > > Thanks!
>> > >
>> > >  fs/nfs/inode.c | 7 ++++++-
>> > >  1 file changed, 6 insertions(+), 1 deletion(-)
>> > >
>> > > diff --git a/fs/nfs/inode.c b/fs/nfs/inode.c
>> > > index 4bffe63..9205513 100644
>> > > --- a/fs/nfs/inode.c
>> > > +++ b/fs/nfs/inode.c
>> > > @@ -506,10 +506,15 @@ nfs_setattr(struct dentry *dentry, struct iattr *attr)
>> > >                 attr->ia_valid &= ~ATTR_MODE;
>> > >
>> > >         if (attr->ia_valid & ATTR_SIZE) {
>> > > +               loff_t i_size;
>> > > +
>> > >                 BUG_ON(!S_ISREG(inode->i_mode));
>> > >
>> > > -               if (attr->ia_size == i_size_read(inode))
>> > > +               i_size = i_size_read(inode);
>> > > +               if (attr->ia_size == i_size)
>> > >                         attr->ia_valid &= ~ATTR_SIZE;
>> > > +               else if (attr->ia_size < i_size && IS_SWAPFILE(inode))
>> > > +                       return -ETXTBSY;
>> > >         }
>> > >
>> > >         /* Optimization: if the end result is no change, don't RPC */
>> > > --
>> > > 2.2.1
>> > >
>> >
>> > I agree that truncating a swap file is bad, however as you point out,
>> > this really only addresses the case on the client that knows about
>> > this being a swap file.
>> > I'll take the patch,
>>
>> Thanks, I appreciate it.
>>
>> > but I'm wondering if we couldn't do better in the
>> > case where we're using NFSv4 by using share deny modes (which are
>> > enforced by the server). The problem is that there appears to be
>> > nothing in swapon() that tells the filesystem this is an open of a
>> > swap file...
>>
>> Yeah, it would be nice for completeness to prevent one client from
>> truncating another client's swapfile. However, I'd hope that anyone
>> using swap-over-NFS on a shared NFS mount would take the necessary
>> precautions in terms of permissions, etc. to prevent someone from doing
>> that. Also, since the failure mode of truncating an NFS swapfile is a
>> corrupt swapfile rather than a corrupt filesystem (like on a local
>> filesystem), it's probably okay to just deal with the low-hanging fruit
>> for now.
>>
>> Thanks!
>>
>> >
>> > Cheers
>> >   Trond
>> > --
>> > Trond Myklebust
>> > Linux NFS client maintainer, PrimaryData
>> > [e-mail]
>>
>> --
>> Omar
>
> Hi, Trond,
>
> Are you still planning on taking this patch? I didn't see it in your
> last pull request to Linus.

I was planning on pushing it in the 3.20 merge window.

Is there any reason to fasttrack it earlier as an important bugfix? To
me it seems more like an "assist user to not shoot self in foot" type
of thing.

Cheers
  Trond

Omar Sandoval Jan. 23, 2015, 6:24 p.m. UTC | #5

On Fri, Jan 23, 2015 at 08:07:41AM -0500, Trond Myklebust wrote:
> On Thu, Jan 22, 2015 at 11:07 PM, Omar Sandoval <[e-mail]> wrote:
> > On Mon, Jan 12, 2015 at 11:43:02AM -0800, Omar Sandoval wrote:
> >> On Sat, Jan 10, 2015 at 05:08:03PM -0500, Trond Myklebust wrote:
> >> > Hi Omar,
> >> >
> >> > On Thu, Jan 8, 2015 at 4:18 AM, Omar Sandoval <[e-mail]> wrote:
> >> > > Most filesystems prevent truncation of an active swapfile by way of
> >> > > inode_newsize_ok, called from inode_change_ok. NFS doesn't call either
> >> > > from nfs_setattr, presumably because most of these checks are expected
> >> > > to be done server-side. However, the IS_SWAPFILE check can only be done
> >> > > client-side, and truncating a swapfile can't possibly be good.
> >> > >
> >> > > Signed-off-by: Omar Sandoval <[e-mail]>
> >> > > ---
> >> > > Hi, Trond,
> >> > >
> >> > > Now that the holidays are over, could you take a look at this? It was
> >> > > generated against v3.19-rc3.
> >> > >
> >> > > Thanks!
> >> > >
> >> > >  fs/nfs/inode.c | 7 ++++++-
> >> > >  1 file changed, 6 insertions(+), 1 deletion(-)
> >> > >
> >> > > diff --git a/fs/nfs/inode.c b/fs/nfs/inode.c
> >> > > index 4bffe63..9205513 100644
> >> > > --- a/fs/nfs/inode.c
> >> > > +++ b/fs/nfs/inode.c
> >> > > @@ -506,10 +506,15 @@ nfs_setattr(struct dentry *dentry, struct iattr *attr)
> >> > >                 attr->ia_valid &= ~ATTR_MODE;
> >> > >
> >> > >         if (attr->ia_valid & ATTR_SIZE) {
> >> > > +               loff_t i_size;
> >> > > +
> >> > >                 BUG_ON(!S_ISREG(inode->i_mode));
> >> > >
> >> > > -               if (attr->ia_size == i_size_read(inode))
> >> > > +               i_size = i_size_read(inode);
> >> > > +               if (attr->ia_size == i_size)
> >> > >                         attr->ia_valid &= ~ATTR_SIZE;
> >> > > +               else if (attr->ia_size < i_size && IS_SWAPFILE(inode))
> >> > > +                       return -ETXTBSY;
> >> > >         }
> >> > >
> >> > >         /* Optimization: if the end result is no change, don't RPC */
> >> > > --
> >> > > 2.2.1
> >> > >
> >> >
> >> > I agree that truncating a swap file is bad, however as you point out,
> >> > this really only addresses the case on the client that knows about
> >> > this being a swap file.
> >> > I'll take the patch,
> >>
> >> Thanks, I appreciate it.
> >>
> >> > but I'm wondering if we couldn't do better in the
> >> > case where we're using NFSv4 by using share deny modes (which are
> >> > enforced by the server). The problem is that there appears to be
> >> > nothing in swapon() that tells the filesystem this is an open of a
> >> > swap file...
> >>
> >> Yeah, it would be nice for completeness to prevent one client from
> >> truncating another client's swapfile. However, I'd hope that anyone
> >> using swap-over-NFS on a shared NFS mount would take the necessary
> >> precautions in terms of permissions, etc. to prevent someone from doing
> >> that. Also, since the failure mode of truncating an NFS swapfile is a
> >> corrupt swapfile rather than a corrupt filesystem (like on a local
> >> filesystem), it's probably okay to just deal with the low-hanging fruit
> >> for now.
> >>
> >> Thanks!
> >>
> >> >
> >> > Cheers
> >> >   Trond
> >> > --
> >> > Trond Myklebust
> >> > Linux NFS client maintainer, PrimaryData
> >> > [e-mail]
> >>
> >> --
> >> Omar
> >
> > Hi, Trond,
> >
> > Are you still planning on taking this patch? I didn't see it in your
> > last pull request to Linus.
> 
> I was planning on pushing it in the 3.20 merge window.
> 
> Is there any reason to fasttrack it earlier as an important bugfix? To
> me it seems more like an "assist user to not shoot self in foot" type
> of thing.
> 
> Cheers
>   Trond
> 
> -- 
> Trond Myklebust
> Linux NFS client maintainer, PrimaryData
> [e-mail]

Ah, okay, 3.20 should be fine.

Thanks,

Patch

diff --git a/fs/nfs/inode.c b/fs/nfs/inode.c
index 4bffe63..9205513 100644
--- a/fs/nfs/inode.c
+++ b/fs/nfs/inode.c
@@ -506,10 +506,15 @@  nfs_setattr(struct dentry *dentry, struct iattr *attr)
 		attr->ia_valid &= ~ATTR_MODE;
 
 	if (attr->ia_valid & ATTR_SIZE) {
+		loff_t i_size;
+
 		BUG_ON(!S_ISREG(inode->i_mode));
 
-		if (attr->ia_size == i_size_read(inode))
+		i_size = i_size_read(inode);
+		if (attr->ia_size == i_size)
 			attr->ia_valid &= ~ATTR_SIZE;
+		else if (attr->ia_size < i_size && IS_SWAPFILE(inode))
+			return -ETXTBSY;
 	}
 
 	/* Optimization: if the end result is no change, don't RPC */