Skip to contentSkip to conversation

rds: synchronize info callbacks with module unload

PATCH Superseded Chengfeng Ye 2 participants 3 messages Jul 31 – Aug 1

Subsystems (4)

v3 of 4 Superseded Latest is v4 →

lore
Share & tools
lore
#0 lore raw

PATCH v3 1/1 rds: synchronize info callbacks with module unload

rds_info_getsockopt() reads a callback from rds_info_funcs and invokes it without protecting the callback's lifetime. Transport modules register functions stored in this array. For example, rds_tcp.ko registers rds_tcp_tc_info() for RDS_INFO_TCP_SOCKETS.

This permits the following interleaving:

CPU0 CPU1
rds_info_getsockopt()
func = rds_tcp_tc_info
rmmod rds_tcp
rds_tcp_exit()
rds_info_deregister_func()
rds_info_funcs[offset] = NULL
free rds_tcp module text
func()

The reader can therefore branch to an address in unloaded module text.

Protect callback invocation with SRCU. Enter the SRCU read-side critical section before loading the callback and leave it only after the callback returns. Clear the callback with release semantics and call synchronize_srcu() before deregistration returns, preventing module unload from freeing its text while an old reader is still executing it. SRCU is required because callbacks such as RDS_INFO_COUNTERS can sleep.

Keep the callback array unannotated and use acquire and release operations for publication so sparse does not have to apply __rcu through the function-pointer typedef. Replace the two callback-slot BUG_ON() checks with WARN_ON_ONCE() and return without changing the slot on mismatch.

---

Changes in v3:
- Use ordinary loads for callback-slot validation under rds_info_lock; retain acquire semantics only for the lockless reader.
- Drop the Fixes tag because this patch targets net-next.
- Add Allison Henderson's Reviewed-by tag.

Changes in v2:
- Split the info callback/module-unload race from the independent inc->i_conn lifetime bug.
- Keep the callback array unannotated and use smp_load_acquire() and smp_store_release() to resolve the sparse errors.
- Replace the two touched callback-slot BUG_ON() checks with WARN_ON_ONCE() error paths.
- Rewrite the commit message for the callback race and omit the unrelated inc->i_conn KASAN report.

 net/rds/info.c | 26 +++++++++++++++++++++-----
 1 file changed, 21 insertions(+), 5 deletions(-)
net/rds/info.c +21 -5
index 21b32eb16559..385fefbfa07f 100644
@@ -32,6 +32,7 @@
  */
 #include <linux/percpu.h>
 #include <linux/seq_file.h>
+#include <linux/srcu.h>
 #include <linux/slab.h>
 #include <linux/proc_fs.h>
 #include <linux/export.h>
@@ -68,6 +69,7 @@ struct rds_info_iterator {
 	unsigned long offset;
 };
 
+DEFINE_STATIC_SRCU(rds_info_srcu);
 static DEFINE_SPINLOCK(rds_info_lock);
 static rds_info_func rds_info_funcs[RDS_INFO_LAST - RDS_INFO_FIRST + 1];
 
@@ -78,8 +80,12 @@ void rds_info_register_func(int optname, rds_info_func func)
 	BUG_ON(optname < RDS_INFO_FIRST || optname > RDS_INFO_LAST);
 
 	spin_lock(&rds_info_lock);
-	BUG_ON(rds_info_funcs[offset]);
-	rds_info_funcs[offset] = func;
+	if (WARN_ON_ONCE(rds_info_funcs[offset])) {
+		spin_unlock(&rds_info_lock);
+		return;
+	}
+	/* Pair with lockless callback lookup. */
+	smp_store_release(&rds_info_funcs[offset], func);
 	spin_unlock(&rds_info_lock);
 }
 EXPORT_SYMBOL_GPL(rds_info_register_func);
@@ -91,9 +97,14 @@ void rds_info_deregister_func(int optname, rds_info_func func)
 	BUG_ON(optname < RDS_INFO_FIRST || optname > RDS_INFO_LAST);
 
 	spin_lock(&rds_info_lock);
-	BUG_ON(rds_info_funcs[offset] != func);
-	rds_info_funcs[offset] = NULL;
+	if (WARN_ON_ONCE(rds_info_funcs[offset] != func)) {
+		spin_unlock(&rds_info_lock);
+		return;
+	}
+	/* Hide the callback before waiting for old readers. */
+	smp_store_release(&rds_info_funcs[offset], NULL);
 	spin_unlock(&rds_info_lock);
+	synchronize_srcu(&rds_info_srcu);
 }
 EXPORT_SYMBOL_GPL(rds_info_deregister_func);
 
@@ -165,6 +176,7 @@ int rds_info_getsockopt(struct socket *sock, int optname, sockopt_t *opt)
 	int npages = 0;
 	int ret;
 	int len;
+	int srcu_idx;
 	int total;
 
 	len = opt->optlen;
@@ -214,8 +226,11 @@ int rds_info_getsockopt(struct socket *sock, int optname, sockopt_t *opt)
 	rdsdebug("len %d nr_pages %lu\n", len, nr_pages);
 
 call_func:
-	func = rds_info_funcs[optname - RDS_INFO_FIRST];
+	srcu_idx = srcu_read_lock(&rds_info_srcu);
+	/* Pair with callback slot publication and removal. */
+	func = smp_load_acquire(&rds_info_funcs[optname - RDS_INFO_FIRST]);
 	if (!func) {
+		srcu_read_unlock(&rds_info_srcu, srcu_idx);
 		ret = -ENOPROTOOPT;
 		goto out;
 	}
@@ -225,6 +240,7 @@ int rds_info_getsockopt(struct socket *sock, int optname, sockopt_t *opt)
 	iter.offset = offset0;
 
 	func(sock, len, &iter, &lens);
+	srcu_read_unlock(&rds_info_srcu, srcu_idx);
 	BUG_ON(lens.each == 0);
 
 	total = lens.nr * lens.each;
Signature · 1 line
2.43.0
#1 lore raw

From: Chengfeng Ye <[email protected]>
Date: Sat, 1 Aug 2026 00:09:42 +0800

Quoted 82 lines

rds_info_getsockopt() reads a callback from rds_info_funcs and invokes it without protecting the callback's lifetime. Transport modules register functions stored in this array. For example, rds_tcp.ko registers rds_tcp_tc_info() for RDS_INFO_TCP_SOCKETS.

This permits the following interleaving:

CPU0 CPU1
rds_info_getsockopt()
func = rds_tcp_tc_info
rmmod rds_tcp
rds_tcp_exit()
rds_info_deregister_func()
rds_info_funcs[offset] = NULL
free rds_tcp module text
func()

The reader can therefore branch to an address in unloaded module text.

Protect callback invocation with SRCU. Enter the SRCU read-side critical section before loading the callback and leave it only after the callback returns. Clear the callback with release semantics and call synchronize_srcu() before deregistration returns, preventing module unload from freeing its text while an old reader is still executing it. SRCU is required because callbacks such as RDS_INFO_COUNTERS can sleep.

Keep the callback array unannotated and use acquire and release operations for publication so sparse does not have to apply __rcu through the function-pointer typedef. Replace the two callback-slot BUG_ON() checks with WARN_ON_ONCE() and return without changing the slot on mismatch.

---

Changes in v3:
- Use ordinary loads for callback-slot validation under rds_info_lock; retain acquire semantics only for the lockless reader.
- Drop the Fixes tag because this patch targets net-next.
- Add Allison Henderson's Reviewed-by tag.

Changes in v2:
- Split the info callback/module-unload race from the independent inc->i_conn lifetime bug.
- Keep the callback array unannotated and use smp_load_acquire() and smp_store_release() to resolve the sparse errors.
- Replace the two touched callback-slot BUG_ON() checks with WARN_ON_ONCE() error paths.
- Rewrite the commit message for the callback race and omit the unrelated inc->i_conn KASAN report.

 net/rds/info.c | 26 +++++++++++++++++++++-----
 1 file changed, 21 insertions(+), 5 deletions(-)
net/rds/info.c +2 -0
index 21b32eb16559..385fefbfa07f 100644
@@ -32,6 +32,7 @@
  */
 #include <linux/percpu.h>
 #include <linux/seq_file.h>
+#include <linux/srcu.h>
 #include <linux/slab.h>
 #include <linux/proc_fs.h>
 #include <linux/export.h>
@@ -68,6 +69,7 @@ struct rds_info_iterator {
 	unsigned long offset;
 };
 
+DEFINE_STATIC_SRCU(rds_info_srcu);
 static DEFINE_SPINLOCK(rds_info_lock);
 static rds_info_func rds_info_funcs[RDS_INFO_LAST - RDS_INFO_FIRST + 1];
 
@@ -78,8 +80,12 @@ void rds_info_register_func(int optname, rds_info_func func)
 	BUG_ON(optname < RDS_INFO_FIRST || optname > RDS_INFO_LAST);

net/rds/info.c void rds_info_register_func(int optname, rds_info_func func) · line 88

Quoted 10 lines
net/rds/info.c +6 -2
 
 	spin_lock(&rds_info_lock);
-	BUG_ON(rds_info_funcs[offset]);
-	rds_info_funcs[offset] = func;
+	if (WARN_ON_ONCE(rds_info_funcs[offset])) {
+		spin_unlock(&rds_info_lock);
+		return;
+	}
+	/* Pair with lockless callback lookup. */
+	smp_store_release(&rds_info_funcs[offset], func);

Why is smp_store_release() necessary here ? and same for smp_load_acquire() on the reader.

I think just WRITE_ONCE() / READ_ONCE() should be enough.

It's just a function pointer, not struct, and the reader does not require any partial ordering, no ?

Quoted 21 lines
net/rds/info.c +7 -2
 	spin_unlock(&rds_info_lock);
 }
 EXPORT_SYMBOL_GPL(rds_info_register_func);
@@ -91,9 +97,14 @@ void rds_info_deregister_func(int optname, rds_info_func func)
 	BUG_ON(optname < RDS_INFO_FIRST || optname > RDS_INFO_LAST);
 
 	spin_lock(&rds_info_lock);
-	BUG_ON(rds_info_funcs[offset] != func);
-	rds_info_funcs[offset] = NULL;
+	if (WARN_ON_ONCE(rds_info_funcs[offset] != func)) {
+		spin_unlock(&rds_info_lock);
+		return;
+	}
+	/* Hide the callback before waiting for old readers. */
+	smp_store_release(&rds_info_funcs[offset], NULL);
 	spin_unlock(&rds_info_lock);
+	synchronize_srcu(&rds_info_srcu);
 }
 EXPORT_SYMBOL_GPL(rds_info_deregister_func);
 

net/rds/info.c int rds_info_getsockopt(struct socket *sock, int optname, sockopt_t *opt) · line 179

Quoted 4 lines
net/rds/info.c +1 -0
@@ -165,6 +176,7 @@ int rds_info_getsockopt(struct socket *sock, int optname, sockopt_t *opt)
 	int npages = 0;
 	int ret;
 	int len;
+	int srcu_idx;

nit: please keep the reverse xmas tree order.

Quoted 27 lines
net/rds/info.c +5 -1
 	int total;
 
 	len = opt->optlen;
@@ -214,8 +226,11 @@ int rds_info_getsockopt(struct socket *sock, int optname, sockopt_t *opt)
 	rdsdebug("len %d nr_pages %lu\n", len, nr_pages);
 
 call_func:
-	func = rds_info_funcs[optname - RDS_INFO_FIRST];
+	srcu_idx = srcu_read_lock(&rds_info_srcu);
+	/* Pair with callback slot publication and removal. */
+	func = smp_load_acquire(&rds_info_funcs[optname - RDS_INFO_FIRST]);
 	if (!func) {
+		srcu_read_unlock(&rds_info_srcu, srcu_idx);
 		ret = -ENOPROTOOPT;
 		goto out;
 	}
@@ -225,6 +240,7 @@ int rds_info_getsockopt(struct socket *sock, int optname, sockopt_t *opt)
 	iter.offset = offset0;
 
 	func(sock, len, &iter, &lens);
+	srcu_read_unlock(&rds_info_srcu, srcu_idx);
 	BUG_ON(lens.each == 0);
 
 	total = lens.nr * lens.each;
Signature · 1 line
2.43.0
#2 lore raw
Quoted 125 lines · Kuniyuki Iwashima

From: Chengfeng Ye <[email protected]>
Date: Sat, 1 Aug 2026 00:09:42 +0800

Quoted 82 lines

rds_info_getsockopt() reads a callback from rds_info_funcs and invokes it without protecting the callback's lifetime. Transport modules register functions stored in this array. For example, rds_tcp.ko registers rds_tcp_tc_info() for RDS_INFO_TCP_SOCKETS.

This permits the following interleaving:

CPU0 CPU1
rds_info_getsockopt()
func = rds_tcp_tc_info
rmmod rds_tcp
rds_tcp_exit()
rds_info_deregister_func()
rds_info_funcs[offset] = NULL
free rds_tcp module text
func()

The reader can therefore branch to an address in unloaded module text.

Protect callback invocation with SRCU. Enter the SRCU read-side critical section before loading the callback and leave it only after the callback returns. Clear the callback with release semantics and call synchronize_srcu() before deregistration returns, preventing module unload from freeing its text while an old reader is still executing it. SRCU is required because callbacks such as RDS_INFO_COUNTERS can sleep.

Keep the callback array unannotated and use acquire and release operations for publication so sparse does not have to apply __rcu through the function-pointer typedef. Replace the two callback-slot BUG_ON() checks with WARN_ON_ONCE() and return without changing the slot on mismatch.

---

Changes in v3:
- Use ordinary loads for callback-slot validation under rds_info_lock; retain acquire semantics only for the lockless reader.
- Drop the Fixes tag because this patch targets net-next.
- Add Allison Henderson's Reviewed-by tag.

Changes in v2:
- Split the info callback/module-unload race from the independent inc->i_conn lifetime bug.
- Keep the callback array unannotated and use smp_load_acquire() and smp_store_release() to resolve the sparse errors.
- Replace the two touched callback-slot BUG_ON() checks with WARN_ON_ONCE() error paths.
- Rewrite the commit message for the callback race and omit the unrelated inc->i_conn KASAN report.

 net/rds/info.c | 26 +++++++++++++++++++++-----
 1 file changed, 21 insertions(+), 5 deletions(-)
net/rds/info.c +2 -0
index 21b32eb16559..385fefbfa07f 100644
@@ -32,6 +32,7 @@
  */
 #include <linux/percpu.h>
 #include <linux/seq_file.h>
+#include <linux/srcu.h>
 #include <linux/slab.h>
 #include <linux/proc_fs.h>
 #include <linux/export.h>
@@ -68,6 +69,7 @@ struct rds_info_iterator {
      unsigned long offset;
 };
 
+DEFINE_STATIC_SRCU(rds_info_srcu);
 static DEFINE_SPINLOCK(rds_info_lock);
 static rds_info_func rds_info_funcs[RDS_INFO_LAST - RDS_INFO_FIRST + 1];
 
@@ -78,8 +80,12 @@ void rds_info_register_func(int optname, rds_info_func func)
      BUG_ON(optname < RDS_INFO_FIRST || optname > RDS_INFO_LAST);

net/rds/info.c void rds_info_register_func(int optname, rds_info_func func) · line 88

Quoted 10 lines
net/rds/info.c +6 -2
 
      spin_lock(&rds_info_lock);
-     BUG_ON(rds_info_funcs[offset]);
-     rds_info_funcs[offset] = func;
+     if (WARN_ON_ONCE(rds_info_funcs[offset])) {
+             spin_unlock(&rds_info_lock);
+             return;
+     }
+     /* Pair with lockless callback lookup. */
+     smp_store_release(&rds_info_funcs[offset], func);

Why is smp_store_release() necessary here ? and same for smp_load_acquire() on the reader.

I think just WRITE_ONCE() / READ_ONCE() should be enough.

It's just a function pointer, not struct, and the reader does not require any partial ordering, no ?

Quoted 20 lines
net/rds/info.c +7 -2
      spin_unlock(&rds_info_lock);
 }
 EXPORT_SYMBOL_GPL(rds_info_register_func);
@@ -91,9 +97,14 @@ void rds_info_deregister_func(int optname, rds_info_func func)
      BUG_ON(optname < RDS_INFO_FIRST || optname > RDS_INFO_LAST);
 
      spin_lock(&rds_info_lock);
-     BUG_ON(rds_info_funcs[offset] != func);
-     rds_info_funcs[offset] = NULL;
+     if (WARN_ON_ONCE(rds_info_funcs[offset] != func)) {
+             spin_unlock(&rds_info_lock);
+             return;
+     }
+     /* Hide the callback before waiting for old readers. */
+     smp_store_release(&rds_info_funcs[offset], NULL);
      spin_unlock(&rds_info_lock);
+     synchronize_srcu(&rds_info_srcu);
 }
 EXPORT_SYMBOL_GPL(rds_info_deregister_func);
 
Quoted 7 lines

net/rds/info.c int rds_info_getsockopt(struct socket *sock, int optname, sockopt_t *opt) · line 179

Quoted 5 lines
net/rds/info.c +1 -0
@@ -165,6 +176,7 @@ int rds_info_getsockopt(struct socket *sock, int optname, sockopt_t *opt)
      int npages = 0;
      int ret;
      int len;
+     int srcu_idx;

nit: please keep the reverse xmas tree order.

Hi,

Indeed WRITE/READ_ONCE should suffice for this code, maybe we don't need that smp store/load despite it looks more secure. I will make a v4 patch to address it as well as the reverse xmas tree order problem.

Best Regards,
Chengfeng