xfrm_policy_lookup_bytype loops on seqcount mutex xfrm_policy_hash_generation
within an RCU read side critical section. Although ill advised, this is fine if
the loop is bounded.
xfrm_policy_hash_generation wraps mutex hash_resize_mutex, which is used to
serialize writers (xfrm_hash_resize, xfrm_hash_rebuild). This is fine too.
On PREEMPT_RT=y, the read_seqcount_begin call within xfrm_policy_lookup_bytype
emits a mutex lock/unlock for hash_resize_mutex. Mutex locking is fine, since
RCU read side critical sections are allowed to sleep with PREEMPT_RT.
xfrm_hash_resize can, however, block on synchronize_rcu while holding
hash_resize_mutex.
This leads to the following situation on PREEMPT_RT, where the writer is
blocked on RCU grace period expiry, while the reader is blocked on a lock held
by the writer:
Thead 1 (xfrm_hash_resize) Thread 2 (xfrm_policy_lookup_bytype)
rcu_read_lock();
mutex_lock(&hash_resize_mutex);
read_seqcount_begin(&xfrm_policy_hash_generation);
mutex_lock(&hash_resize_mutex); // block
xfrm_bydst_resize();
synchronize_rcu(); // block
<RCU stalls in xfrm_policy_lookup_bytype>
Move the read_seqcount_begin call outside of the RCU read side critical section,
and do an rcu_read_unlock/retry if we got stale data within the critical section.
On non-PREEMPT_RT, this shortens the time spent within RCU read side critical
section in case the seqcount needs a retry, and avoids unbounded looping.
Fixes: a7c44247f70 ("xfrm: policy: make xfrm_policy_lookup_bytype lockless")
Signed-off-by: Varad Gautam <[email protected]>
Cc: linux-rt-users <[email protected]>
Cc: [email protected]
Cc: [email protected] # v4.9
Cc: Steffen Klassert <[email protected]>
Cc: Herbert Xu <[email protected]>
Cc: "David S. Miller" <[email protected]>
Cc: Jakub Kicinski <[email protected]>
Cc: Florian Westphal <[email protected]>
---
net/xfrm/xfrm_policy.c | 21 ++++++++++++++-------
1 file changed, 14 insertions(+), 7 deletions(-)
diff --git a/net/xfrm/xfrm_policy.c b/net/xfrm/xfrm_policy.c
index ce500f847b99..e9d0df2a2ab1 100644
--- a/net/xfrm/xfrm_policy.c
+++ b/net/xfrm/xfrm_policy.c
@@ -2092,12 +2092,15 @@ static struct xfrm_policy *xfrm_policy_lookup_bytype(struct net *net, u8 type,
if (unlikely(!daddr || !saddr))
return NULL;
- rcu_read_lock();
retry:
- do {
- sequence = read_seqcount_begin(&xfrm_policy_hash_generation);
- chain = policy_hash_direct(net, daddr, saddr, family, dir);
- } while (read_seqcount_retry(&xfrm_policy_hash_generation, sequence));
+ sequence = read_seqcount_begin(&xfrm_policy_hash_generation);
+ rcu_read_lock();
+
+ chain = policy_hash_direct(net, daddr, saddr, family, dir);
+ if (read_seqcount_retry(&xfrm_policy_hash_generation, sequence)) {
+ rcu_read_unlock();
+ goto retry;
+ }
ret = NULL;
hlist_for_each_entry_rcu(pol, chain, bydst) {
@@ -2128,11 +2131,15 @@ static struct xfrm_policy *xfrm_policy_lookup_bytype(struct net *net, u8 type,
}
skip_inexact:
- if (read_seqcount_retry(&xfrm_policy_hash_generation, sequence))
+ if (read_seqcount_retry(&xfrm_policy_hash_generation, sequence)) {
+ rcu_read_unlock();
goto retry;
+ }
- if (ret && !xfrm_pol_hold_rcu(ret))
+ if (ret && !xfrm_pol_hold_rcu(ret)) {
+ rcu_read_unlock();
goto retry;
+ }
fail:
rcu_read_unlock();
--
2.26.2
On Fri, May 28, 2021, Varad Gautam wrote:
...
>
> Thead 1 (xfrm_hash_resize) Thread 2 (xfrm_policy_lookup_bytype)
>
> rcu_read_lock();
> mutex_lock(&hash_resize_mutex);
> read_seqcount_begin(&xfrm_policy_hash_generation);
> mutex_lock(&hash_resize_mutex); // block
> xfrm_bydst_resize();
> synchronize_rcu(); // block
> <RCU stalls in xfrm_policy_lookup_bytype>
>
...
> Fixes: a7c44247f70 ("xfrm: policy: make xfrm_policy_lookup_bytype lockless")
Minor note: the 'Fixes' commit should be 77cc278f7b20 ("xfrm: policy:
Use sequence counters with associated lock") instead.
The reason read_seqcount_begin() is emitting a mutex_lock() on
PREEMPT_RT is because of the s/seqcount_t/seqcount_mutex_t/ change.
Kind regards,
--
Ahmed S. Darwish
Linutronix GmbH
On 5/28/21 5:11 PM, Ahmed S. Darwish wrote:
> On Fri, May 28, 2021, Varad Gautam wrote:
> ...
>>
>> Thead 1 (xfrm_hash_resize) Thread 2 (xfrm_policy_lookup_bytype)
>>
>> rcu_read_lock();
>> mutex_lock(&hash_resize_mutex);
>> read_seqcount_begin(&xfrm_policy_hash_generation);
>> mutex_lock(&hash_resize_mutex); // block
>> xfrm_bydst_resize();
>> synchronize_rcu(); // block
>> <RCU stalls in xfrm_policy_lookup_bytype>
>>
> ...
>> Fixes: a7c44247f70 ("xfrm: policy: make xfrm_policy_lookup_bytype lockless")
>
> Minor note: the 'Fixes' commit should be 77cc278f7b20 ("xfrm: policy:
> Use sequence counters with associated lock") instead.
>
> The reason read_seqcount_begin() is emitting a mutex_lock() on
> PREEMPT_RT is because of the s/seqcount_t/seqcount_mutex_t/ change.
>
Thanks, corrected it in v2.
> Kind regards,
>
> --
> Ahmed S. Darwish
> Linutronix GmbH
>
--
SUSE Software Solutions Germany GmbH
Maxfeldstr. 5
90409 Nürnberg
Germany
HRB 36809, AG Nürnberg
Geschäftsführer: Felix Imendörffer
xfrm_policy_lookup_bytype loops on seqcount mutex xfrm_policy_hash_generation
within an RCU read side critical section. Although ill advised, this is fine if
the loop is bounded.
xfrm_policy_hash_generation wraps mutex hash_resize_mutex, which is used to
serialize writers (xfrm_hash_resize, xfrm_hash_rebuild). This is fine too.
On PREEMPT_RT=y, the read_seqcount_begin call within xfrm_policy_lookup_bytype
emits a mutex lock/unlock for hash_resize_mutex. Mutex locking is fine, since
RCU read side critical sections are allowed to sleep with PREEMPT_RT.
xfrm_hash_resize can, however, block on synchronize_rcu while holding
hash_resize_mutex.
This leads to the following situation on PREEMPT_RT, where the writer is
blocked on RCU grace period expiry, while the reader is blocked on a lock held
by the writer:
Thead 1 (xfrm_hash_resize) Thread 2 (xfrm_policy_lookup_bytype)
rcu_read_lock();
mutex_lock(&hash_resize_mutex);
read_seqcount_begin(&xfrm_policy_hash_generation);
mutex_lock(&hash_resize_mutex); // block
xfrm_bydst_resize();
synchronize_rcu(); // block
<RCU stalls in xfrm_policy_lookup_bytype>
Move the read_seqcount_begin call outside of the RCU read side critical section,
and do an rcu_read_unlock/retry if we got stale data within the critical section.
On non-PREEMPT_RT, this shortens the time spent within RCU read side critical
section in case the seqcount needs a retry, and avoids unbounded looping.
Fixes: 77cc278f7b20 ("xfrm: policy: Use sequence counters with associated lock")
Signed-off-by: Varad Gautam <[email protected]>
Cc: linux-rt-users <[email protected]>
Cc: [email protected]
Cc: [email protected] # v4.9
Cc: Steffen Klassert <[email protected]>
Cc: Herbert Xu <[email protected]>
Cc: "David S. Miller" <[email protected]>
Cc: Jakub Kicinski <[email protected]>
Cc: Florian Westphal <[email protected]>
Cc: "Ahmed S. Darwish" <[email protected]>
---
v2: Correct 'Fixes:' to the right commit.
net/xfrm/xfrm_policy.c | 21 ++++++++++++++-------
1 file changed, 14 insertions(+), 7 deletions(-)
diff --git a/net/xfrm/xfrm_policy.c b/net/xfrm/xfrm_policy.c
index ce500f847b99..e9d0df2a2ab1 100644
--- a/net/xfrm/xfrm_policy.c
+++ b/net/xfrm/xfrm_policy.c
@@ -2092,12 +2092,15 @@ static struct xfrm_policy *xfrm_policy_lookup_bytype(struct net *net, u8 type,
if (unlikely(!daddr || !saddr))
return NULL;
- rcu_read_lock();
retry:
- do {
- sequence = read_seqcount_begin(&xfrm_policy_hash_generation);
- chain = policy_hash_direct(net, daddr, saddr, family, dir);
- } while (read_seqcount_retry(&xfrm_policy_hash_generation, sequence));
+ sequence = read_seqcount_begin(&xfrm_policy_hash_generation);
+ rcu_read_lock();
+
+ chain = policy_hash_direct(net, daddr, saddr, family, dir);
+ if (read_seqcount_retry(&xfrm_policy_hash_generation, sequence)) {
+ rcu_read_unlock();
+ goto retry;
+ }
ret = NULL;
hlist_for_each_entry_rcu(pol, chain, bydst) {
@@ -2128,11 +2131,15 @@ static struct xfrm_policy *xfrm_policy_lookup_bytype(struct net *net, u8 type,
}
skip_inexact:
- if (read_seqcount_retry(&xfrm_policy_hash_generation, sequence))
+ if (read_seqcount_retry(&xfrm_policy_hash_generation, sequence)) {
+ rcu_read_unlock();
goto retry;
+ }
- if (ret && !xfrm_pol_hold_rcu(ret))
+ if (ret && !xfrm_pol_hold_rcu(ret)) {
+ rcu_read_unlock();
goto retry;
+ }
fail:
rcu_read_unlock();
--
2.26.2