Message ID | 20180102163020.32473-1-nstange@suse.de |
---|---|
State | Changes Requested, archived |
Delegated to: | David Miller |
Headers | show |
Series | net: ipv4: emulate READ_ONCE() on ->hdrincl bit-field in raw_sendmsg() | expand |
Hi, On Tue, 2 Jan 2018 17:30:20 +0100 Nicolai Stange <nstange@suse.de> wrote: > [...] > > diff --git a/net/ipv4/raw.c b/net/ipv4/raw.c > index 5b9bd5c33d9d..e84290c28c0c 100644 > --- a/net/ipv4/raw.c > +++ b/net/ipv4/raw.c > @@ -513,16 +513,18 @@ static int raw_sendmsg(struct sock *sk, struct msghdr *msg, size_t len) > int err; > struct ip_options_data opt_copy; > struct raw_frag_vec rfv; > - int hdrincl; > + int hdrincl, __hdrincl; > > err = -EMSGSIZE; > if (len > 0xFFFF) > goto out; > > /* hdrincl should be READ_ONCE(inet->hdrincl) > - * but READ_ONCE() doesn't work with bit fields > + * but READ_ONCE() doesn't work with bit fields. > + * Emulate it by doing the READ_ONCE() from an intermediate int. > */ > - hdrincl = inet->hdrincl; > + __hdrincl = inet->hdrincl; > + hdrincl = READ_ONCE(__hdrincl); I guess you don't actually need to use a third variable. What about doing READ_ONCE() on hdrincl itself after the first assignment? Perhaps something like the patch below -- applies to net.git, yields same binary output as your version with gcc 6, looks IMHO more straightforward: diff --git a/net/ipv4/raw.c b/net/ipv4/raw.c index 125c1eab3eaa..8c2f783a95fc 100644 --- a/net/ipv4/raw.c +++ b/net/ipv4/raw.c @@ -519,10 +519,12 @@ static int raw_sendmsg(struct sock *sk, struct msghdr *msg, size_t len) if (len > 0xFFFF) goto out; - /* hdrincl should be READ_ONCE(inet->hdrincl) - * but READ_ONCE() doesn't work with bit fields + /* hdrincl should be READ_ONCE(inet->hdrincl) but READ_ONCE() doesn't + * work with bit fields. Emulate it by adding a further sequence point. */ hdrincl = inet->hdrincl; + hdrincl = READ_ONCE(hdrincl); + /* * Check the flags. */
Hi Stefano, Stefano Brivio <sbrivio@redhat.com> writes: > On Tue, 2 Jan 2018 17:30:20 +0100 > Nicolai Stange <nstange@suse.de> wrote: > >> [...] >> >> diff --git a/net/ipv4/raw.c b/net/ipv4/raw.c >> index 5b9bd5c33d9d..e84290c28c0c 100644 >> --- a/net/ipv4/raw.c >> +++ b/net/ipv4/raw.c >> @@ -513,16 +513,18 @@ static int raw_sendmsg(struct sock *sk, struct msghdr *msg, size_t len) >> int err; >> struct ip_options_data opt_copy; >> struct raw_frag_vec rfv; >> - int hdrincl; >> + int hdrincl, __hdrincl; >> >> err = -EMSGSIZE; >> if (len > 0xFFFF) >> goto out; >> >> /* hdrincl should be READ_ONCE(inet->hdrincl) >> - * but READ_ONCE() doesn't work with bit fields >> + * but READ_ONCE() doesn't work with bit fields. >> + * Emulate it by doing the READ_ONCE() from an intermediate int. >> */ >> - hdrincl = inet->hdrincl; >> + __hdrincl = inet->hdrincl; >> + hdrincl = READ_ONCE(__hdrincl); > > I guess you don't actually need to use a third variable. What about > doing READ_ONCE() on hdrincl itself after the first assignment? > > Perhaps something like the patch below -- applies to net.git, yields > same binary output as your version with gcc 6, looks IMHO more > straightforward: > > diff --git a/net/ipv4/raw.c b/net/ipv4/raw.c > index 125c1eab3eaa..8c2f783a95fc 100644 > --- a/net/ipv4/raw.c > +++ b/net/ipv4/raw.c > @@ -519,10 +519,12 @@ static int raw_sendmsg(struct sock *sk, struct msghdr *msg, size_t len) > if (len > 0xFFFF) > goto out; > > - /* hdrincl should be READ_ONCE(inet->hdrincl) > - * but READ_ONCE() doesn't work with bit fields > + /* hdrincl should be READ_ONCE(inet->hdrincl) but READ_ONCE() doesn't > + * work with bit fields. Emulate it by adding a further sequence point. > */ > hdrincl = inet->hdrincl; > + hdrincl = READ_ONCE(hdrincl); > + Yes, this does also work. In fact, after having been lowered into SSA form, it should be equivalent to what I posted. So, it's a matter of preference/style and I'd leave the decision on this to the maintainers -- for me, either way is fine. I don't like the "sequence point" wording in the comment above though: AFAICS, if taken in the meaning of C99, it's not any sequence point but the volatile access in READ_ONCE() which ensures that there won't be any reloads from ->hdrincl. If you don't mind, I'll adjust that comment if asked to resend with your solution. Thanks, Nicolai
Hi Nicolai, On Wed, 03 Jan 2018 10:28:20 +0100 Nicolai Stange <nstange@suse.de> wrote: > Hi Stefano, > > Stefano Brivio <sbrivio@redhat.com> writes: > > > On Tue, 2 Jan 2018 17:30:20 +0100 > > Nicolai Stange <nstange@suse.de> wrote: > > > >> [...] > >> > >> diff --git a/net/ipv4/raw.c b/net/ipv4/raw.c > >> index 5b9bd5c33d9d..e84290c28c0c 100644 > >> --- a/net/ipv4/raw.c > >> +++ b/net/ipv4/raw.c > >> @@ -513,16 +513,18 @@ static int raw_sendmsg(struct sock *sk, struct msghdr *msg, size_t len) > >> int err; > >> struct ip_options_data opt_copy; > >> struct raw_frag_vec rfv; > >> - int hdrincl; > >> + int hdrincl, __hdrincl; > >> > >> err = -EMSGSIZE; > >> if (len > 0xFFFF) > >> goto out; > >> > >> /* hdrincl should be READ_ONCE(inet->hdrincl) > >> - * but READ_ONCE() doesn't work with bit fields > >> + * but READ_ONCE() doesn't work with bit fields. > >> + * Emulate it by doing the READ_ONCE() from an intermediate int. > >> */ > >> - hdrincl = inet->hdrincl; > >> + __hdrincl = inet->hdrincl; > >> + hdrincl = READ_ONCE(__hdrincl); > > > > I guess you don't actually need to use a third variable. What about > > doing READ_ONCE() on hdrincl itself after the first assignment? > > > > Perhaps something like the patch below -- applies to net.git, yields > > same binary output as your version with gcc 6, looks IMHO more > > straightforward: > > > > diff --git a/net/ipv4/raw.c b/net/ipv4/raw.c > > index 125c1eab3eaa..8c2f783a95fc 100644 > > --- a/net/ipv4/raw.c > > +++ b/net/ipv4/raw.c > > @@ -519,10 +519,12 @@ static int raw_sendmsg(struct sock *sk, struct msghdr *msg, size_t len) > > if (len > 0xFFFF) > > goto out; > > > > - /* hdrincl should be READ_ONCE(inet->hdrincl) > > - * but READ_ONCE() doesn't work with bit fields > > + /* hdrincl should be READ_ONCE(inet->hdrincl) but READ_ONCE() doesn't > > + * work with bit fields. Emulate it by adding a further sequence point. > > */ > > hdrincl = inet->hdrincl; > > + hdrincl = READ_ONCE(hdrincl); > > + > > Yes, this does also work. In fact, after having been lowered into SSA > form, it should be equivalent to what I posted. > > So, it's a matter of preference/style and I'd leave the decision on > this to the maintainers -- for me, either way is fine. > > I don't like the "sequence point" wording in the comment above though: > AFAICS, if taken in the meaning of C99, it's not any sequence point but > the volatile access in READ_ONCE() which ensures that there won't be any > reloads from ->hdrincl. If you don't mind, I'll adjust that comment if > asked to resend with your solution. Well, by "by adding a further sequence point" I refer to what we have to do to emulate READ_ONCE(), not to the reason why we need READ_ONCE(). However, this is a likely sign that my comment isn't that clear either. So unless you have better ideas, I would go with: + /* hdrincl should be READ_ONCE(inet->hdrincl) but READ_ONCE() doesn't + * work with bit fields. Doing this indirectly yields the same result. but I really hope you have a better idea. :)
diff --git a/net/ipv4/raw.c b/net/ipv4/raw.c index 5b9bd5c33d9d..e84290c28c0c 100644 --- a/net/ipv4/raw.c +++ b/net/ipv4/raw.c @@ -513,16 +513,18 @@ static int raw_sendmsg(struct sock *sk, struct msghdr *msg, size_t len) int err; struct ip_options_data opt_copy; struct raw_frag_vec rfv; - int hdrincl; + int hdrincl, __hdrincl; err = -EMSGSIZE; if (len > 0xFFFF) goto out; /* hdrincl should be READ_ONCE(inet->hdrincl) - * but READ_ONCE() doesn't work with bit fields + * but READ_ONCE() doesn't work with bit fields. + * Emulate it by doing the READ_ONCE() from an intermediate int. */ - hdrincl = inet->hdrincl; + __hdrincl = inet->hdrincl; + hdrincl = READ_ONCE(__hdrincl); /* * Check the flags. */
Commit 8f659a03a0ba ("net: ipv4: fix for a race condition in raw_sendmsg") fixed the issue of possibly inconsistent ->hdrincl handling due to concurrent updates by reading this bit-field member into a local variable and using the thus stabilized value in subsequent tests. However, aforementioned commit also adds the (correct) comment that /* hdrincl should be READ_ONCE(inet->hdrincl) * but READ_ONCE() doesn't work with bit fields */ because as it stands, the compiler is free to shortcut or even eliminate the local variable at its will. Note that I have not seen anything like this happening in reality and thus, the concern is a theoretical one. However, in order to be on the safe side, emulate a READ_ONCE() on the bit-field by introducing an intermediate local variable and doing a READ_ONCE() from it: int __hdrincl = inet->hdrincl; int hdrincl = READ_ONCE(__hdrincl); This breaks the chain in the sense that the compiler is not allowed to replace subsequent reads from hdrincl with reloads from inet->hdrincl. Fixes: 8f659a03a0ba ("net: ipv4: fix for a race condition in raw_sendmsg") Signed-off-by: Nicolai Stange <nstange@suse.de> --- Compile-tested only (with inspection of compiler output on x86_64). Applicable to linux-next-20180102. net/ipv4/raw.c | 8 +++++--- 1 file changed, 5 insertions(+), 3 deletions(-)