SUNRPC: fix a memleak in gss_import_v2_context

  The ctx->mech_used.data allocated by kmemdup is not freed in neither
gss_import_v2_context nor it only caller radeon_driver_open_kms.
Thus, this patch reform the last call of gss_import_v2_context to the
gss_krb5_import_ctx_v2, preventing the memleak while keepping the return
formation.

Fixes: 47d848077629 ("gss_krb5: handle new context format from gssd")
Signed-off-by: Zhipeng Lu <alexious@zju.edu.cn>
---
 net/sunrpc/auth_gss/gss_krb5_mech.c | 9 ++++++++-
 1 file changed, 8 insertions(+), 1 deletion(-)
  

On Sun, Dec 24, 2023 at 04:20:33PM +0800, Zhipeng Lu wrote:
> The ctx->mech_used.data allocated by kmemdup is not freed in neither
> gss_import_v2_context nor it only caller radeon_driver_open_kms.
> Thus, this patch reform the last call of gss_import_v2_context to the
> gss_krb5_import_ctx_v2, preventing the memleak while keepping the return
> formation.
> 
> Fixes: 47d848077629 ("gss_krb5: handle new context format from gssd")
> Signed-off-by: Zhipeng Lu <alexious@zju.edu.cn>
> ---
>  net/sunrpc/auth_gss/gss_krb5_mech.c | 9 ++++++++-
>  1 file changed, 8 insertions(+), 1 deletion(-)
> 
> diff --git a/net/sunrpc/auth_gss/gss_krb5_mech.c b/net/sunrpc/auth_gss/gss_krb5_mech.c
> index e31cfdf7eadc..1e54bd63e3f0 100644
> --- a/net/sunrpc/auth_gss/gss_krb5_mech.c
> +++ b/net/sunrpc/auth_gss/gss_krb5_mech.c
> @@ -398,6 +398,7 @@ gss_import_v2_context(const void *p, const void *end, struct krb5_ctx *ctx,
>  	u64 seq_send64;
>  	int keylen;
>  	u32 time32;
> +	int ret;
>  
>  	p = simple_get_bytes(p, end, &ctx->flags, sizeof(ctx->flags));
>  	if (IS_ERR(p))
> @@ -450,8 +451,14 @@ gss_import_v2_context(const void *p, const void *end, struct krb5_ctx *ctx,
>  	}
>  	ctx->mech_used.len = gss_kerberos_mech.gm_oid.len;
>  
> -	return gss_krb5_import_ctx_v2(ctx, gfp_mask);
> +	ret = gss_krb5_import_ctx_v2(ctx, gfp_mask);
> +	if (ret) {
> +		p = ERR_PTR(ret);
> +		goto out_free;
> +	};
>  
> +out_free:
> +	kfree(ctx->mech_used.data);

If the caller's error flow does not invoke
gss_krb5_delete_sec_context(), then I would expect more than just
mech_used.data would be leaked. What if, instead, you changed
gss_krb5_import_sec_context() like this (untested):

471         ret = gss_import_v2_context(p, end, ctx, gfp_mask);
472         memzero_explicit(&ctx->Ksess, sizeof(ctx->Ksess));
473         if (ret) {      
   -                kfree(ctx);                      
   +                gss_krb5_delete_sec_context(ctx);
475                 return ret;
476         }    

Obviously you would need to add a forward declaration of
gss_krb5_import_sec_context() to make this compile. The question
is whether gss_krb5_delete_sec_context() will deal with a partially-
initialized @ctx.

How did you find this leak, and what kind of testing was done to
confirm the fix is safe?


>  out_err:
>  	return PTR_ERR(p);
>  }
> -- 
> 2.34.1
>
  

On Sun, Dec 24, 2023 at 04:20:33PM +0800, Zhipeng Lu wrote:
> The ctx->mech_used.data allocated by kmemdup is not freed in neither
> gss_import_v2_context nor it only caller radeon_driver_open_kms.
> Thus, this patch reform the last call of gss_import_v2_context to the
> gss_krb5_import_ctx_v2, preventing the memleak while keepping the return
> formation.
> 
> Fixes: 47d848077629 ("gss_krb5: handle new context format from gssd")
> Signed-off-by: Zhipeng Lu <alexious@zju.edu.cn>
> ---
>  net/sunrpc/auth_gss/gss_krb5_mech.c | 9 ++++++++-
>  1 file changed, 8 insertions(+), 1 deletion(-)
> 
> diff --git a/net/sunrpc/auth_gss/gss_krb5_mech.c b/net/sunrpc/auth_gss/gss_krb5_mech.c
> index e31cfdf7eadc..1e54bd63e3f0 100644
> --- a/net/sunrpc/auth_gss/gss_krb5_mech.c
> +++ b/net/sunrpc/auth_gss/gss_krb5_mech.c
> @@ -398,6 +398,7 @@ gss_import_v2_context(const void *p, const void *end, struct krb5_ctx *ctx,
>  	u64 seq_send64;
>  	int keylen;
>  	u32 time32;
> +	int ret;
>  
>  	p = simple_get_bytes(p, end, &ctx->flags, sizeof(ctx->flags));
>  	if (IS_ERR(p))
> @@ -450,8 +451,14 @@ gss_import_v2_context(const void *p, const void *end, struct krb5_ctx *ctx,
>  	}
>  	ctx->mech_used.len = gss_kerberos_mech.gm_oid.len;
>  
> -	return gss_krb5_import_ctx_v2(ctx, gfp_mask);
> +	ret = gss_krb5_import_ctx_v2(ctx, gfp_mask);
> +	if (ret) {
> +		p = ERR_PTR(ret);
> +		goto out_free;
> +	};

Hi Zhipeng Lu,

I think you need to handle the non-error case here:

	return 0;

>  
> +out_free:
> +	kfree(ctx->mech_used.data);
>  out_err:
>  	return PTR_ERR(p);
>  }
> -- 
> 2.34.1
>
  

> On Sun, Dec 24, 2023 at 04:20:33PM +0800, Zhipeng Lu wrote:
> > The ctx->mech_used.data allocated by kmemdup is not freed in neither
> > gss_import_v2_context nor it only caller radeon_driver_open_kms.
> > Thus, this patch reform the last call of gss_import_v2_context to the
> > gss_krb5_import_ctx_v2, preventing the memleak while keepping the return
> > formation.
> > 
> > Fixes: 47d848077629 ("gss_krb5: handle new context format from gssd")
> > Signed-off-by: Zhipeng Lu <alexious@zju.edu.cn>
> > ---
> >  net/sunrpc/auth_gss/gss_krb5_mech.c | 9 ++++++++-
> >  1 file changed, 8 insertions(+), 1 deletion(-)
> > 
> > diff --git a/net/sunrpc/auth_gss/gss_krb5_mech.c b/net/sunrpc/auth_gss/gss_krb5_mech.c
> > index e31cfdf7eadc..1e54bd63e3f0 100644
> > --- a/net/sunrpc/auth_gss/gss_krb5_mech.c
> > +++ b/net/sunrpc/auth_gss/gss_krb5_mech.c
> > @@ -398,6 +398,7 @@ gss_import_v2_context(const void *p, const void *end, struct krb5_ctx *ctx,
> >  	u64 seq_send64;
> >  	int keylen;
> >  	u32 time32;
> > +	int ret;
> >  
> >  	p = simple_get_bytes(p, end, &ctx->flags, sizeof(ctx->flags));
> >  	if (IS_ERR(p))
> > @@ -450,8 +451,14 @@ gss_import_v2_context(const void *p, const void *end, struct krb5_ctx *ctx,
> >  	}
> >  	ctx->mech_used.len = gss_kerberos_mech.gm_oid.len;
> >  
> > -	return gss_krb5_import_ctx_v2(ctx, gfp_mask);
> > +	ret = gss_krb5_import_ctx_v2(ctx, gfp_mask);
> > +	if (ret) {
> > +		p = ERR_PTR(ret);
> > +		goto out_free;
> > +	};
> >  
> > +out_free:
> > +	kfree(ctx->mech_used.data);
> 
> If the caller's error flow does not invoke
> gss_krb5_delete_sec_context(), then I would expect more than just
> mech_used.data would be leaked. What if, instead, you changed
> gss_krb5_import_sec_context() like this (untested):
> 
> 471         ret = gss_import_v2_context(p, end, ctx, gfp_mask);
> 472         memzero_explicit(&ctx->Ksess, sizeof(ctx->Ksess));
> 473         if (ret) {      
>    -                kfree(ctx);                      
>    +                gss_krb5_delete_sec_context(ctx);
> 475                 return ret;
> 476         }    
> 
> Obviously you would need to add a forward declaration of
> gss_krb5_import_sec_context() to make this compile. The question
> is whether gss_krb5_delete_sec_context() will deal with a partially-
> initialized @ctx.

Since the ctx is allocated just in gss_krb5_import_sec_context, 
together with that all of gss_krb5_import_sec_context, gss_import_v2_context(with this patch)
and gss_krb5_import_ctx_v2 are allocation-free balanced. It seems that we don't need to 
release anything else by invoking gss_krb5_delete_sec_context.

If I miss something leaked, please let me know.


> 
> How did you find this leak, and what kind of testing was done to
> confirm the fix is safe?

I found this memleak by static analysis. 
The safety issue can't be solved by automatic tools as far as I know.
So I check patches manuelly before sending patches.

> >  out_err:
> >  	return PTR_ERR(p);
> >  }
> > -- 
> > 2.34.1
> > 
> 
> -- 
> Chuck Lever
  

On Tue, Dec 26, 2023 at 05:01:05PM +0800, alexious@zju.edu.cn wrote:
> > On Sun, Dec 24, 2023 at 04:20:33PM +0800, Zhipeng Lu wrote:
> > > The ctx->mech_used.data allocated by kmemdup is not freed in neither
> > > gss_import_v2_context nor it only caller radeon_driver_open_kms.
> > > Thus, this patch reform the last call of gss_import_v2_context to the
> > > gss_krb5_import_ctx_v2, preventing the memleak while keepping the return
> > > formation.
> > > 
> > > Fixes: 47d848077629 ("gss_krb5: handle new context format from gssd")
> > > Signed-off-by: Zhipeng Lu <alexious@zju.edu.cn>
> > > ---
> > >  net/sunrpc/auth_gss/gss_krb5_mech.c | 9 ++++++++-
> > >  1 file changed, 8 insertions(+), 1 deletion(-)
> > > 
> > > diff --git a/net/sunrpc/auth_gss/gss_krb5_mech.c b/net/sunrpc/auth_gss/gss_krb5_mech.c
> > > index e31cfdf7eadc..1e54bd63e3f0 100644
> > > --- a/net/sunrpc/auth_gss/gss_krb5_mech.c
> > > +++ b/net/sunrpc/auth_gss/gss_krb5_mech.c
> > > @@ -398,6 +398,7 @@ gss_import_v2_context(const void *p, const void *end, struct krb5_ctx *ctx,
> > >  	u64 seq_send64;
> > >  	int keylen;
> > >  	u32 time32;
> > > +	int ret;
> > >  
> > >  	p = simple_get_bytes(p, end, &ctx->flags, sizeof(ctx->flags));
> > >  	if (IS_ERR(p))
> > > @@ -450,8 +451,14 @@ gss_import_v2_context(const void *p, const void *end, struct krb5_ctx *ctx,
> > >  	}
> > >  	ctx->mech_used.len = gss_kerberos_mech.gm_oid.len;
> > >  
> > > -	return gss_krb5_import_ctx_v2(ctx, gfp_mask);
> > > +	ret = gss_krb5_import_ctx_v2(ctx, gfp_mask);
> > > +	if (ret) {
> > > +		p = ERR_PTR(ret);
> > > +		goto out_free;
> > > +	};
> > >  
> > > +out_free:
> > > +	kfree(ctx->mech_used.data);
> > 
> > If the caller's error flow does not invoke
> > gss_krb5_delete_sec_context(), then I would expect more than just
> > mech_used.data would be leaked. What if, instead, you changed
> > gss_krb5_import_sec_context() like this (untested):
> > 
> > 471         ret = gss_import_v2_context(p, end, ctx, gfp_mask);
> > 472         memzero_explicit(&ctx->Ksess, sizeof(ctx->Ksess));
> > 473         if (ret) {      
> >    -                kfree(ctx);                      
> >    +                gss_krb5_delete_sec_context(ctx);
> > 475                 return ret;
> > 476         }    
> > 
> > Obviously you would need to add a forward declaration of
> > gss_krb5_import_sec_context() to make this compile. The question
> > is whether gss_krb5_delete_sec_context() will deal with a partially-
> > initialized @ctx.
> 
> Since the ctx is allocated just in gss_krb5_import_sec_context, 
> together with that all of gss_krb5_import_sec_context, gss_import_v2_context(with this patch)
> and gss_krb5_import_ctx_v2 are allocation-free balanced. It seems that we don't need to 
> release anything else by invoking gss_krb5_delete_sec_context.
> 
> If I miss something leaked, please let me know.

I see, if gss_krb5_import_ctx_v2() fails, it releases the ciphers
and hashes via out_free. So no leak there.

A nicer approach would be to handle that clean up in
gss_krb5_import_sec_context(): less code duplication.

But you're right, it's not broken today.


> > How did you find this leak, and what kind of testing was done to
> > confirm the fix is safe?
> 
> I found this memleak by static analysis. 
> The safety issue can't be solved by automatic tools as far as I know.
> So I check patches manuelly before sending patches.

Can you give some details about how you check the patches?