Building the ip_vs_sync code with CONFIG_OPTIMIZE_INLINING on x86
confuses the compiler to the point where it produces a rather
dubious warning message:
net/netfilter/ipvs/ip_vs_sync.c:1073:33: error: ‘opt.init_seq’ may be used uninitialized in this function [-Werror=maybe-uninitialized]
struct ip_vs_sync_conn_options opt;
^~~
net/netfilter/ipvs/ip_vs_sync.c:1073:33: error: ‘opt.delta’ may be used uninitialized in this function [-Werror=maybe-uninitialized]
net/netfilter/ipvs/ip_vs_sync.c:1073:33: error: ‘opt.previous_delta’ may be used uninitialized in this function [-Werror=maybe-uninitialized]
net/netfilter/ipvs/ip_vs_sync.c:1073:33: error: ‘*((void *)&opt+12).init_seq’ may be used uninitialized in this function [-Werror=maybe-uninitialized]
net/netfilter/ipvs/ip_vs_sync.c:1073:33: error: ‘*((void *)&opt+12).delta’ may be used uninitialized in this function [-Werror=maybe-uninitialized]
net/netfilter/ipvs/ip_vs_sync.c:1073:33: error: ‘*((void *)&opt+12).previous_delta’ may be used uninitialized in this function [-Werror=maybe-uninitialized]
The problem appears to be a combination of a number of factors, including
the __builtin_bswap32 compiler builtin being slightly odd, having a large
amount of code inlined into a single function, and the way that some
functions only get partially inlined here.
I've spent way too much time trying to work out a way to improve the
code, but the best I've come up with is to add an explicit memset
right before the ip_vs_seq structure is first initialized here. When
the compiler works correctly, this has absolutely no effect, but in the
case that produces the warning, the warning disappears.
In the process of analysing this warning, I also noticed that
we use memcpy to copy the larger ip_vs_sync_conn_options structure
over two members of the ip_vs_conn structure. This works because
the layout is identical, but seems error-prone, so I'm changing
this in the process to directly copy the two members. This change
seemed to have no effect on the object code or the warning, but
it deals with the same data, so I kept the two changes together.
Signed-off-by: Arnd Bergmann <[email protected]>
---
net/netfilter/ipvs/ip_vs_sync.c | 7 +++++--
1 file changed, 5 insertions(+), 2 deletions(-)
diff --git a/net/netfilter/ipvs/ip_vs_sync.c b/net/netfilter/ipvs/ip_vs_sync.c
index 1b07578bedf3..9350530c16c1 100644
--- a/net/netfilter/ipvs/ip_vs_sync.c
+++ b/net/netfilter/ipvs/ip_vs_sync.c
@@ -283,6 +283,7 @@ struct ip_vs_sync_buff {
*/
static void ntoh_seq(struct ip_vs_seq *no, struct ip_vs_seq *ho)
{
+ memset(ho, 0, sizeof(*ho));
ho->init_seq = get_unaligned_be32(&no->init_seq);
ho->delta = get_unaligned_be32(&no->delta);
ho->previous_delta = get_unaligned_be32(&no->previous_delta);
@@ -917,8 +918,10 @@ static void ip_vs_proc_conn(struct netns_ipvs *ipvs, struct ip_vs_conn_param *pa
kfree(param->pe_data);
}
- if (opt)
- memcpy(&cp->in_seq, opt, sizeof(*opt));
+ if (opt) {
+ cp->in_seq = opt->in_seq;
+ cp->out_seq = opt->out_seq;
+ }
atomic_set(&cp->in_pkts, sysctl_sync_threshold(ipvs));
cp->state = state;
cp->old_state = cp->state;
--
2.9.0
Hello,
On Mon, 24 Oct 2016, Arnd Bergmann wrote:
> Building the ip_vs_sync code with CONFIG_OPTIMIZE_INLINING on x86
> confuses the compiler to the point where it produces a rather
> dubious warning message:
>
> net/netfilter/ipvs/ip_vs_sync.c:1073:33: error: ‘opt.init_seq’ may be used uninitialized in this function [-Werror=maybe-uninitialized]
> struct ip_vs_sync_conn_options opt;
> ^~~
> net/netfilter/ipvs/ip_vs_sync.c:1073:33: error: ‘opt.delta’ may be used uninitialized in this function [-Werror=maybe-uninitialized]
> net/netfilter/ipvs/ip_vs_sync.c:1073:33: error: ‘opt.previous_delta’ may be used uninitialized in this function [-Werror=maybe-uninitialized]
> net/netfilter/ipvs/ip_vs_sync.c:1073:33: error: ‘*((void *)&opt+12).init_seq’ may be used uninitialized in this function [-Werror=maybe-uninitialized]
> net/netfilter/ipvs/ip_vs_sync.c:1073:33: error: ‘*((void *)&opt+12).delta’ may be used uninitialized in this function [-Werror=maybe-uninitialized]
> net/netfilter/ipvs/ip_vs_sync.c:1073:33: error: ‘*((void *)&opt+12).previous_delta’ may be used uninitialized in this function [-Werror=maybe-uninitialized]
>
> The problem appears to be a combination of a number of factors, including
> the __builtin_bswap32 compiler builtin being slightly odd, having a large
> amount of code inlined into a single function, and the way that some
> functions only get partially inlined here.
>
> I've spent way too much time trying to work out a way to improve the
> code, but the best I've come up with is to add an explicit memset
> right before the ip_vs_seq structure is first initialized here. When
> the compiler works correctly, this has absolutely no effect, but in the
> case that produces the warning, the warning disappears.
>
> In the process of analysing this warning, I also noticed that
> we use memcpy to copy the larger ip_vs_sync_conn_options structure
> over two members of the ip_vs_conn structure. This works because
> the layout is identical, but seems error-prone, so I'm changing
> this in the process to directly copy the two members. This change
> seemed to have no effect on the object code or the warning, but
> it deals with the same data, so I kept the two changes together.
>
> Signed-off-by: Arnd Bergmann <[email protected]>
OK,
Acked-by: Julian Anastasov <[email protected]>
I guess, Simon will take the patch for ipvs-next.
> ---
> net/netfilter/ipvs/ip_vs_sync.c | 7 +++++--
> 1 file changed, 5 insertions(+), 2 deletions(-)
>
> diff --git a/net/netfilter/ipvs/ip_vs_sync.c b/net/netfilter/ipvs/ip_vs_sync.c
> index 1b07578bedf3..9350530c16c1 100644
> --- a/net/netfilter/ipvs/ip_vs_sync.c
> +++ b/net/netfilter/ipvs/ip_vs_sync.c
> @@ -283,6 +283,7 @@ struct ip_vs_sync_buff {
> */
> static void ntoh_seq(struct ip_vs_seq *no, struct ip_vs_seq *ho)
> {
> + memset(ho, 0, sizeof(*ho));
> ho->init_seq = get_unaligned_be32(&no->init_seq);
> ho->delta = get_unaligned_be32(&no->delta);
> ho->previous_delta = get_unaligned_be32(&no->previous_delta);
So, now there is a double write here?
What about such constructs?:
*ho = (struct ip_vs_seq) {
.init_seq = get_unaligned_be32(&no->init_seq),
...
};
Any difference in the compiled code or warnings?
> @@ -917,8 +918,10 @@ static void ip_vs_proc_conn(struct netns_ipvs *ipvs, struct ip_vs_conn_param *pa
> kfree(param->pe_data);
> }
>
> - if (opt)
> - memcpy(&cp->in_seq, opt, sizeof(*opt));
> + if (opt) {
> + cp->in_seq = opt->in_seq;
> + cp->out_seq = opt->out_seq;
This fix is fine.
> + }
> atomic_set(&cp->in_pkts, sysctl_sync_threshold(ipvs));
> cp->state = state;
> cp->old_state = cp->state;
> --
> 2.9.0
Regards
--
Julian Anastasov <[email protected]>
On Monday, October 24, 2016 10:47:54 PM CEST Julian Anastasov wrote:
> > diff --git a/net/netfilter/ipvs/ip_vs_sync.c b/net/netfilter/ipvs/ip_vs_sync.c
> > index 1b07578bedf3..9350530c16c1 100644
> > --- a/net/netfilter/ipvs/ip_vs_sync.c
> > +++ b/net/netfilter/ipvs/ip_vs_sync.c
> > @@ -283,6 +283,7 @@ struct ip_vs_sync_buff {
> > */
> > static void ntoh_seq(struct ip_vs_seq *no, struct ip_vs_seq *ho)
> > {
> > + memset(ho, 0, sizeof(*ho));
> > ho->init_seq = get_unaligned_be32(&no->init_seq);
> > ho->delta = get_unaligned_be32(&no->delta);
> > ho->previous_delta = get_unaligned_be32(&no->previous_delta);
>
> So, now there is a double write here?
Correct. I would hope that a sane version of gcc would just not
perform the first write. What happens instead is that the version
that produces the warning here moves the initialization to the
top of the calling function.
> What about such constructs?:
>
> *ho = (struct ip_vs_seq) {
> .init_seq = get_unaligned_be32(&no->init_seq),
> ...
> };
>
> Any difference in the compiled code or warnings?
Yes, it's one of many things I tried. What happens here is that
the warning remains as long as all fields are initialized together,
e.g. these two produces the same warning:
a)
ho->init_seq = get_unaligned_be32(&no->init_seq);
ho->delta = get_unaligned_be32(&no->delta);
ho->previous_delta = get_unaligned_be32(&no->previous_delta);
b)
*ho = (struct ip_vs_seq) {
.init_seq = get_unaligned_be32(&no->init_seq);
.delta = get_unaligned_be32(&no->delta);
.previous_delta = get_unaligned_be32(&no->previous_delta);
};
but this one does not:
c)
*ho = (struct ip_vs_seq) {
.delta = get_unaligned_be32(&no->delta);
.previous_delta = get_unaligned_be32(&no->previous_delta);
};
ho->init_seq = get_unaligned_be32(&no->init_seq);
I have absolutely no idea what is going on inside of gcc here.
Arnd
From: Arnd Bergmann
> Sent: 24 October 2016 21:22
> On Monday, October 24, 2016 10:47:54 PM CEST Julian Anastasov wrote:
> > > diff --git a/net/netfilter/ipvs/ip_vs_sync.c b/net/netfilter/ipvs/ip_vs_sync.c
> > > index 1b07578bedf3..9350530c16c1 100644
> > > --- a/net/netfilter/ipvs/ip_vs_sync.c
> > > +++ b/net/netfilter/ipvs/ip_vs_sync.c
> > > @@ -283,6 +283,7 @@ struct ip_vs_sync_buff {
> > > */
> > > static void ntoh_seq(struct ip_vs_seq *no, struct ip_vs_seq *ho)
> > > {
> > > + memset(ho, 0, sizeof(*ho));
> > > ho->init_seq = get_unaligned_be32(&no->init_seq);
> > > ho->delta = get_unaligned_be32(&no->delta);
> > > ho->previous_delta = get_unaligned_be32(&no->previous_delta);
> >
> > So, now there is a double write here?
>
> Correct. I would hope that a sane version of gcc would just not
> perform the first write. What happens instead is that the version
> that produces the warning here moves the initialization to the
> top of the calling function.
Maybe doing the 3 get_unaligned_be32() before the memset will stop
the double-writes.
The problem is that the compiler doesn't know that the two structures
don't alias each other.
David
On Mon, Oct 24, 2016 at 10:47:54PM +0300, Julian Anastasov wrote:
>
> Hello,
>
> On Mon, 24 Oct 2016, Arnd Bergmann wrote:
>
> > Building the ip_vs_sync code with CONFIG_OPTIMIZE_INLINING on x86
> > confuses the compiler to the point where it produces a rather
> > dubious warning message:
> >
> > net/netfilter/ipvs/ip_vs_sync.c:1073:33: error: ‘opt.init_seq’ may be used uninitialized in this function [-Werror=maybe-uninitialized]
> > struct ip_vs_sync_conn_options opt;
> > ^~~
> > net/netfilter/ipvs/ip_vs_sync.c:1073:33: error: ‘opt.delta’ may be used uninitialized in this function [-Werror=maybe-uninitialized]
> > net/netfilter/ipvs/ip_vs_sync.c:1073:33: error: ‘opt.previous_delta’ may be used uninitialized in this function [-Werror=maybe-uninitialized]
> > net/netfilter/ipvs/ip_vs_sync.c:1073:33: error: ‘*((void *)&opt+12).init_seq’ may be used uninitialized in this function [-Werror=maybe-uninitialized]
> > net/netfilter/ipvs/ip_vs_sync.c:1073:33: error: ‘*((void *)&opt+12).delta’ may be used uninitialized in this function [-Werror=maybe-uninitialized]
> > net/netfilter/ipvs/ip_vs_sync.c:1073:33: error: ‘*((void *)&opt+12).previous_delta’ may be used uninitialized in this function [-Werror=maybe-uninitialized]
> >
> > The problem appears to be a combination of a number of factors, including
> > the __builtin_bswap32 compiler builtin being slightly odd, having a large
> > amount of code inlined into a single function, and the way that some
> > functions only get partially inlined here.
> >
> > I've spent way too much time trying to work out a way to improve the
> > code, but the best I've come up with is to add an explicit memset
> > right before the ip_vs_seq structure is first initialized here. When
> > the compiler works correctly, this has absolutely no effect, but in the
> > case that produces the warning, the warning disappears.
> >
> > In the process of analysing this warning, I also noticed that
> > we use memcpy to copy the larger ip_vs_sync_conn_options structure
> > over two members of the ip_vs_conn structure. This works because
> > the layout is identical, but seems error-prone, so I'm changing
> > this in the process to directly copy the two members. This change
> > seemed to have no effect on the object code or the warning, but
> > it deals with the same data, so I kept the two changes together.
> >
> > Signed-off-by: Arnd Bergmann <[email protected]>
>
> OK,
>
> Acked-by: Julian Anastasov <[email protected]>
>
> I guess, Simon will take the patch for ipvs-next.
@Simon: If you have no more pending updates, I can save you one pull
request for this small fix by placing this.
Thanks!
On Fri, Oct 28, 2016 at 11:34:22AM +0200, Pablo Neira Ayuso wrote:
> On Mon, Oct 24, 2016 at 10:47:54PM +0300, Julian Anastasov wrote:
> >
> > Hello,
> >
> > On Mon, 24 Oct 2016, Arnd Bergmann wrote:
> >
> > > Building the ip_vs_sync code with CONFIG_OPTIMIZE_INLINING on x86
> > > confuses the compiler to the point where it produces a rather
> > > dubious warning message:
> > >
> > > net/netfilter/ipvs/ip_vs_sync.c:1073:33: error: ‘opt.init_seq’ may be used uninitialized in this function [-Werror=maybe-uninitialized]
> > > struct ip_vs_sync_conn_options opt;
> > > ^~~
> > > net/netfilter/ipvs/ip_vs_sync.c:1073:33: error: ‘opt.delta’ may be used uninitialized in this function [-Werror=maybe-uninitialized]
> > > net/netfilter/ipvs/ip_vs_sync.c:1073:33: error: ‘opt.previous_delta’ may be used uninitialized in this function [-Werror=maybe-uninitialized]
> > > net/netfilter/ipvs/ip_vs_sync.c:1073:33: error: ‘*((void *)&opt+12).init_seq’ may be used uninitialized in this function [-Werror=maybe-uninitialized]
> > > net/netfilter/ipvs/ip_vs_sync.c:1073:33: error: ‘*((void *)&opt+12).delta’ may be used uninitialized in this function [-Werror=maybe-uninitialized]
> > > net/netfilter/ipvs/ip_vs_sync.c:1073:33: error: ‘*((void *)&opt+12).previous_delta’ may be used uninitialized in this function [-Werror=maybe-uninitialized]
> > >
> > > The problem appears to be a combination of a number of factors, including
> > > the __builtin_bswap32 compiler builtin being slightly odd, having a large
> > > amount of code inlined into a single function, and the way that some
> > > functions only get partially inlined here.
> > >
> > > I've spent way too much time trying to work out a way to improve the
> > > code, but the best I've come up with is to add an explicit memset
> > > right before the ip_vs_seq structure is first initialized here. When
> > > the compiler works correctly, this has absolutely no effect, but in the
> > > case that produces the warning, the warning disappears.
> > >
> > > In the process of analysing this warning, I also noticed that
> > > we use memcpy to copy the larger ip_vs_sync_conn_options structure
> > > over two members of the ip_vs_conn structure. This works because
> > > the layout is identical, but seems error-prone, so I'm changing
> > > this in the process to directly copy the two members. This change
> > > seemed to have no effect on the object code or the warning, but
> > > it deals with the same data, so I kept the two changes together.
> > >
> > > Signed-off-by: Arnd Bergmann <[email protected]>
> >
> > OK,
> >
> > Acked-by: Julian Anastasov <[email protected]>
> >
> > I guess, Simon will take the patch for ipvs-next.
>
> @Simon: If you have no more pending updates, I can save you one pull
> request for this small fix by placing this.
Thanks Pablo, please do.
Signed-off-by: Simon Horman <[email protected]>
On Fri, Oct 28, 2016 at 01:40:23PM +0200, Simon Horman wrote:
> On Fri, Oct 28, 2016 at 11:34:22AM +0200, Pablo Neira Ayuso wrote:
> > On Mon, Oct 24, 2016 at 10:47:54PM +0300, Julian Anastasov wrote:
> > >
> > > Hello,
> > >
> > > On Mon, 24 Oct 2016, Arnd Bergmann wrote:
> > >
> > > > Building the ip_vs_sync code with CONFIG_OPTIMIZE_INLINING on x86
> > > > confuses the compiler to the point where it produces a rather
> > > > dubious warning message:
> > > >
> > > > net/netfilter/ipvs/ip_vs_sync.c:1073:33: error: ‘opt.init_seq’ may be used uninitialized in this function [-Werror=maybe-uninitialized]
> > > > struct ip_vs_sync_conn_options opt;
> > > > ^~~
> > > > net/netfilter/ipvs/ip_vs_sync.c:1073:33: error: ‘opt.delta’ may be used uninitialized in this function [-Werror=maybe-uninitialized]
> > > > net/netfilter/ipvs/ip_vs_sync.c:1073:33: error: ‘opt.previous_delta’ may be used uninitialized in this function [-Werror=maybe-uninitialized]
> > > > net/netfilter/ipvs/ip_vs_sync.c:1073:33: error: ‘*((void *)&opt+12).init_seq’ may be used uninitialized in this function [-Werror=maybe-uninitialized]
> > > > net/netfilter/ipvs/ip_vs_sync.c:1073:33: error: ‘*((void *)&opt+12).delta’ may be used uninitialized in this function [-Werror=maybe-uninitialized]
> > > > net/netfilter/ipvs/ip_vs_sync.c:1073:33: error: ‘*((void *)&opt+12).previous_delta’ may be used uninitialized in this function [-Werror=maybe-uninitialized]
> > > >
> > > > The problem appears to be a combination of a number of factors, including
> > > > the __builtin_bswap32 compiler builtin being slightly odd, having a large
> > > > amount of code inlined into a single function, and the way that some
> > > > functions only get partially inlined here.
> > > >
> > > > I've spent way too much time trying to work out a way to improve the
> > > > code, but the best I've come up with is to add an explicit memset
> > > > right before the ip_vs_seq structure is first initialized here. When
> > > > the compiler works correctly, this has absolutely no effect, but in the
> > > > case that produces the warning, the warning disappears.
> > > >
> > > > In the process of analysing this warning, I also noticed that
> > > > we use memcpy to copy the larger ip_vs_sync_conn_options structure
> > > > over two members of the ip_vs_conn structure. This works because
> > > > the layout is identical, but seems error-prone, so I'm changing
> > > > this in the process to directly copy the two members. This change
> > > > seemed to have no effect on the object code or the warning, but
> > > > it deals with the same data, so I kept the two changes together.
> > > >
> > > > Signed-off-by: Arnd Bergmann <[email protected]>
> > >
> > > OK,
> > >
> > > Acked-by: Julian Anastasov <[email protected]>
> > >
> > > I guess, Simon will take the patch for ipvs-next.
> >
> > @Simon: If you have no more pending updates, I can save you one pull
> > request for this small fix by placing this.
>
> Thanks Pablo, please do.
Thanks Simon, feel free to exercise this path anytime.
> Signed-off-by: Simon Horman <[email protected]>
Applied to nf, thanks!