Re: Re: [PATCH] net: bonding: avoid use-after-free with tx_hashtbl/rx_hashtbl

From: sujing
Date: Tue Apr 11 2023 - 23:43:16 EST


Thank you so much for your reply!

 I see there are mainly 5 questions about my patch.
 Before answering your questions, let me provide more detailed
information about this issue.

In bonding mode 6, I encounter 3 crashes on ft2000+/64(arm64) server
with kernel 4.19.90 (the issue still exists in the latest version).
Crashes only occur when closing process preempts the lock before TX/RX,
the call traces are as follow:

Call trace 1 : TX vs. close
[21406.614627] tlb_choose_channel+0x5c/0x188 [bonding]
[21406.620009] bond_alb_xmit+0x1fc/0x5b8 [bonding]
[21406.625043] bond_start_xmit+0xf8/0x4c0 [bonding]
[21406.630162] dev_hard_start_xmit+0xac/0x258
[21406.634759] __dev_queue_xmit+0x754/0x940
[21406.639182] dev_queue_xmit+0x24/0x30
[21406.643261] ip_finish_output2+0x23c/0x3e8
[21406.647771] ip_finish_output+0x1c8/0x2a8
[21406.652194] ip_output+0x9c/0x100
[21406.655924] ip_local_out+0x58/0x68
[21406.659829] __ip_queue_xmit+0x12c/0x368
[21406.664167] ip_queue_xmit+0x10/0x18
[21406.668158] __tcp_transmit_skb+0x4f8/0xab8
[21406.672754] tcp_write_xmit+0x23c/0xf78

Timeline 1:
tlb_deinitialize ------------------ bond_start_xmit
spin_lock_bh ---------------------- bond_alb_xmit
tx_hashtbl = NULL ----------------- tlb_choose_channel
spin_unlock_bh -------------------- //wait
----------------------------------- spin_lock_bh
----------------------------------- __tlb_choose_channel
----------------------------------- tx_hashtbl[hash_index].tx_slave
----------------------------------- spin_unlock_bh

 Call trace 2 : RX vs. close
[ 76.809874] Call trace:
[ 76.810317] rlb_arp_recv+0x218/0x2f0 [bonding]
[ 76.811070] bond_handle_frame+0x58/0x270 [bonding]
[ 76.811905] __netif_receive_skb_core+0x2d4/0xd90
[ 76.812683] __netif_receive_skb_one_core+0x38/0x68
[ 76.813477] __netif_receive_skb+0x28/0x80
[ 76.814156] netif_receive_skb_internal+0x3c/0xa8
[ 76.814929] napi_gro_receive+0xf8/0x170
[ 76.815594] receive_buf+0xec/0xa08 [virtio_net]
[ 76.816353] virtnet_poll+0x144/0x310 [virtio_net]
[ 76.817134] net_rx_action+0x158/0x3a0

Timeline 2:
rlb_deinitialize ------------------ bond_handle_frame
spin_lock_bh ---------------------- rlb_arp_recv
rx_hashtbl = NULL ----------------- rlb_update_entry_from_arp
spin_unlock_bh -------------------- //wait
----------------------------------- spin_lock_bh
----------------------------------- rx_hashtbl[hash_index].assigned
----------------------------------- spin_unlock_bh

 Call trace 3 : TX vs. close
[ 144.146818] Call trace:
[ 144.147256] rlb_choose_channel+0x84/0x2e8 [bonding]
[ 144.148067] bond_alb_xmit+0x2d0/0x5b8 [bonding]
[ 144.148820] bond_start_xmit+0x4e8/0x4f0 [bonding]
[ 144.149649] dev_hard_start_xmit+0xac/0x258
[ 144.150335] __dev_queue_xmit+0x754/0x940
[ 144.150994] dev_queue_xmit+0x24/0x30
[ 144.151616] arp_xmit+0x24/0x90
[ 144.152148] arp_send_dst.part.3+0xb4/0xf0
[ 144.152820] arp_solicit+0x1ac/0x260
[ 144.153416] neigh_probe+0x64/0x88
[ 144.153986] __neigh_event_send+0x124/0x340
[ 144.154671] neigh_resolve_output+0x124/0x200
[ 144.155383] ip_finish_output2+0x138/0x3e8
[ 144.156055] ip_finish_output+0x1c8/0x2a8
[ 144.156720] ip_output+0x9c/0x100
[ 144.157276] ip_local_out+0x58/0x68
[ 144.157860] ip_send_skb+0x2c/0x80

Timeline 3:
tlb_deinitialize ------------------ bond_start_xmit
spin_lock_bh ---------------------- bond_alb_xmit
tx_hashtbl = NULL ----------------- rlb_arp_xmit
spin_unlock_bh -------------------- rlb_choose_channel
----------------------------------- spin_lock_bh
----------------------------------- rx_hashtbl[hash_index].assigned
----------------------------------- spin_unlock_bh

To solve the crashes above, the initial idea is to check whether
tx_hashtbl/rx_hashtbl is NULL before operating it ,
and then add some exception handling process.

In bonding driver, there is only one tx_hashtbl operation (in
'__tlb_choose_channel()'), TLB is easy to solve.
But there are so many rx_hashtbl operations that also aren't
checked whether the pointer is NULL.
It may cause other use-after-free while closing.

I am worried that the initial idea can't solve crash issues under other
uncertain race conditions, and it maybe affecting the normal process.

Besides, I think it is unnecessary to free and reallocate
TLB / RLB hash table every-time the bond device is down/up.
Therefore it should be allocated and freed once in
'bond_init()/bond_uninit()'.

Since we can't control whether upper level sends packets or not,
I think the best idea is to make sure tx_hashtbl/rx_hashtbl pointer will
not be set to NULL during the closing process.

Therefore, move 'bond_alb_deinitialize()' from 'bond_close()' to
'bond_uninit()', move 'bond_alb_initialize()' from 'bond_open()'
to 'bond_init()'.

So it becomes as below:
                                    enter(insmod)
bond_init------------------bond_init------------
bond_open------------------bond_alb_initialize--
bond_alb_initialize---==>--tlb_initialize-------
tlb_initialize-------------rlb_initialize-------
rlb_initialize-------------bond_open------------
                                    exit(rmmod)
bond_close-----------------bond_close-----------
bond_alb_deinitialize------bond_uninit----------
tlb_deinitialize------==>--bond_alb_deinitialize
rlb_deinitialize-----------tlb_deinitialize-----
bond_uninit----------------rlb_deinitialize-----

The original initialization is:
if (bond_is_lb(bond))
     bond_alb_initialize(bond, (BOND_MODE(bond) == BOND_MODE_ALB));
        -> tlb_initialize(bond);
        -> if (rlb_enabled)
                  rlb_initialize(bond);
bond_is_lb()
     -> return BOND_MODE(bond) == BOND_MODE_TLB ||
                       BOND_MODE(bond) == BOND_MODE_ALB;
rlb_enabled == (BOND_MODE(bond) == BOND_MODE_ALB);

The original de-initialization is:
if (bond_is_lb(bond))
     bond_alb_deinitialize(bond);
        -> tlb_deinitialize(bond);
        -> if (bond_info->rlb_enabled)
                   rlb_deinitialize(bond);


However, 'BOND_MODE(bond)' is set only after 'bond_init()', meaning that
there's no way to judge the bond device's mode in 'bond_init()'.
I haven't come up with any better idea, so I decide to set up / tear down
the TLB / RLB hash table even when they won't be used.

So the initialization changes to:
bond_alb_initialize(bond);
           -> tlb_initialize(bond);
           -> rlb_initialize(bond);

As well as the de initialization:
bond_alb_deinitialize(bond);
           -> tlb_deinitialize(bond);
           -> rlb_deinitialize(bond);

 Since the main points is the movement of applying and releasing
hash-table memory, the rest in 'bond_close() bond_open()' are keep in
original places.

 In order to make sure each time the status is clean, 'bond_open()'
should clear the TLB / RLB hash table before used.

The answers to your questions in the previous email are below.


On 2023/3/28 12:38, Jay Vosburgh wrote:
sujing <sujing@xxxxxxxxxx> wrote:

In bonding mode 6 (Balance-alb),
there are some potential race conditions between the 'bond_close' process
and the tx/rx processes that use tx_hashtbl/rx_hashtbl,
which may lead to use-after-free.

For instance, when the bond6 device is in the 'bond_close' process
while some backlogged packets from upper level are transmitted
to 'bond_start_xmit', there is a spinlock contention between
'tlb_deinitialize' and 'tlb_choose_channel'.

If 'tlb_deinitialize' preempts the lock before 'tlb_choose_channel',
a NULL pointer kernel panic will be triggered.

Here's the timeline:

bond_close ------------------ bond_start_xmit
bond_alb_deinitialize ------- __bond_start_xmit
tlb_deinitialize ------------ bond_alb_xmit
spin_lock_bh ---------------- bond_xmit_alb_slave_get
tx_hashtbl = NULL ----------- tlb_choose_channel
spin_unlock_bh -------------- //wait for spin_lock_bh
------------------------------ spin_lock_bh
------------------------------ __tlb_choose_channel
causing kernel panic ========> tx_hashtbl[hash_index].tx_slave
------------------------------ spin_unlock_bh
I'm still thinking on the race here, but have some questions
below about the implementation in the meantime.

Signed-off-by: sujing <sujing@xxxxxxxxxx>
---
drivers/net/bonding/bond_alb.c | 32 +++++++++------------------
drivers/net/bonding/bond_main.c | 39 +++++++++++++++++++++++++++------
include/net/bond_alb.h | 5 ++++-
3 files changed, 46 insertions(+), 30 deletions(-)

diff --git a/drivers/net/bonding/bond_alb.c b/drivers/net/bonding/bond_alb.c
index b9dbad3a8af8..f6ff5ea835c4 100644
--- a/drivers/net/bonding/bond_alb.c
+++ b/drivers/net/bonding/bond_alb.c
@@ -71,7 +71,7 @@ static inline u8 _simple_hash(const u8 *hash_start, int hash_size)

/*********************** tlb specific functions ***************************/

-static inline void tlb_init_table_entry(struct tlb_client_info *entry, int save_load)
+void tlb_init_table_entry(struct tlb_client_info *entry, int save_load)
{
if (save_load) {
entry->load_history = 1 + entry->tx_bytes /
@@ -269,8 +269,8 @@ static void rlb_update_entry_from_arp(struct bonding *bond, struct arp_pkt *arp)
spin_unlock_bh(&bond->mode_lock);
}

-static int rlb_arp_recv(const struct sk_buff *skb, struct bonding *bond,
- struct slave *slave)
+int rlb_arp_recv(const struct sk_buff *skb, struct bonding *bond,
+ struct slave *slave)
{
struct arp_pkt *arp, _arp;

@@ -756,7 +756,7 @@ static void rlb_init_table_entry_src(struct rlb_client_info *entry)
entry->src_next = RLB_NULL_INDEX;
}

-static void rlb_init_table_entry(struct rlb_client_info *entry)
+void rlb_init_table_entry(struct rlb_client_info *entry)
{
memset(entry, 0, sizeof(struct rlb_client_info));
rlb_init_table_entry_dst(entry);
@@ -874,9 +874,6 @@ static int rlb_initialize(struct bonding *bond)

spin_unlock_bh(&bond->mode_lock);

- /* register to receive ARPs */
- bond->recv_probe = rlb_arp_recv;
-
return 0;
}

@@ -888,7 +885,6 @@ static void rlb_deinitialize(struct bonding *bond)

kfree(bond_info->rx_hashtbl);
bond_info->rx_hashtbl = NULL;
- bond_info->rx_hashtbl_used_head = RLB_NULL_INDEX;
Why remove this line?

The idea is moving 'rlb_deinitialize()' from 'bond_close()' process to
'bond_uninit()' process, so there's no need to reset rx_hashtbl_used_head's
value while exiting the driver.

spin_unlock_bh(&bond->mode_lock);
}
@@ -1303,7 +1299,7 @@ static bool alb_determine_nd(struct sk_buff *skb, struct bonding *bond)

/************************ exported alb functions ************************/

-int bond_alb_initialize(struct bonding *bond, int rlb_enabled)
+int bond_alb_initialize(struct bonding *bond)
{
int res;

@@ -1311,15 +1307,10 @@ int bond_alb_initialize(struct bonding *bond, int rlb_enabled)
if (res)
return res;

- if (rlb_enabled) {
- res = rlb_initialize(bond);
- if (res) {
- tlb_deinitialize(bond);
- return res;
- }
- bond->alb_info.rlb_enabled = 1;
- } else {
- bond->alb_info.rlb_enabled = 0;
+ res = rlb_initialize(bond);
+ if (res) {
+ tlb_deinitialize(bond);
+ return res;
}

return 0;
@@ -1327,12 +1318,9 @@ int bond_alb_initialize(struct bonding *bond, int rlb_enabled)

void bond_alb_deinitialize(struct bonding *bond)
{
- struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
-
tlb_deinitialize(bond);

- if (bond_info->rlb_enabled)
- rlb_deinitialize(bond);
+ rlb_deinitialize(bond);
Why is rlb_deinitialize() now unconditionally called here and in
bond_alb_initialize()? if rlb_enabled is false, why set up / tear down
the RLB hash table that won't be used?

There may be some misunderstanding. 'rlb_deinitialize()' and
'tlb_deinitialize()'  are now called in 'bond_alb_deinitialize()'.

As for the next question, I explain it in the beginning, there's no way to
judge the bond device's mode in 'bond_init()'.
I haven't come up with a better way.

}

static netdev_tx_t bond_do_alb_xmit(struct sk_buff *skb, struct bonding *bond,
diff --git a/drivers/net/bonding/bond_main.c b/drivers/net/bonding/bond_main.c
index 236e5219c811..8fcb5d3ac0a2 100644
--- a/drivers/net/bonding/bond_main.c
+++ b/drivers/net/bonding/bond_main.c
@@ -4217,6 +4217,7 @@ static int bond_open(struct net_device *bond_dev)
struct bonding *bond = netdev_priv(bond_dev);
struct list_head *iter;
struct slave *slave;
+ int i;

if (BOND_MODE(bond) == BOND_MODE_ROUNDROBIN && !bond->rr_tx_counter) {
bond->rr_tx_counter = alloc_percpu(u32);
@@ -4239,11 +4240,29 @@ static int bond_open(struct net_device *bond_dev)
}

if (bond_is_lb(bond)) {
- /* bond_alb_initialize must be called before the timer
- * is started.
- */
- if (bond_alb_initialize(bond, (BOND_MODE(bond) == BOND_MODE_ALB)))
- return -ENOMEM;
+ struct alb_bond_info *bond_info = &(BOND_ALB_INFO(bond));
+
+ spin_lock_bh(&bond->mode_lock);
+
+ for (i = 0; i < TLB_HASH_TABLE_SIZE; i++)
+ tlb_init_table_entry(&bond_info->tx_hashtbl[i], 0);
+
+ spin_unlock_bh(&bond->mode_lock);
+
+ if (BOND_MODE(bond) == BOND_MODE_ALB) {
+ bond->alb_info.rlb_enabled = 1;
+ spin_lock_bh(&bond->mode_lock);
+
+ bond_info->rx_hashtbl_used_head = RLB_NULL_INDEX;
+ for (i = 0; i < RLB_HASH_TABLE_SIZE; i++)
+ rlb_init_table_entry(bond_info->rx_hashtbl + i);
+
+ spin_unlock_bh(&bond->mode_lock);
+ bond->recv_probe = rlb_arp_recv;
+ } else {
+ bond->alb_info.rlb_enabled = 0;
+ }
+
Why is all of the above done directly in bond_open() and not in
bond_alb.c somewhere? That would reduce some churn (changing some
functions away from static).

Also, I see that bond_alb_initialize() is now called from
bond_init() instead of bond_open(), and it only calls rlb_initialize().
However, this now duplicates most of the functionality of
rlb_initialize() and tlb_initialize() here. Why?

In general, the described race is TX vs. close processing, so
why is there so much change to the open processing?

-J

For the record, 'bond_alb_initialize()' now calls both 'rlb_initialize()'
and 'tlb_initialize()'.

You're right. There are two points that need to be improved.

point 1: create 2 new functions 'rlb_hash_initialize()' and
'tlb_hash_initialize()' in bond_alb.c, make them do all of the above.
The plan is as follow:
***************************************
bond_open() in bond_main.c:
if (bond_is_lb(bond)) {
     tlb_hash_initialize(bond);
     if (BOND_MODE(bond) == BOND_MODE_ALB) {
          bond->alb_info.rlb_enabled = 1;
          rlb_hash_initialize(bond);
          bond->recv_probe = rlb_arp_recv;
     } else
          bond->alb_info.rlb_enabled = 0;
}

tlb_hash_initialize() in bond_alb.c:
spin_lock_bh();
for (i)
     tlb_init_table_entry();
spin_unlock_bh();

rlb_hash_initialize() in bond_alb.c:
spin_lock_bh();
rx_hashtbl_used_head = RLB_NULL_INDEX;
for (i)
     rlb_init_table_entry();
spin_unlock_bh();

***************************************

point 2: delete the 'rlb_init_table_entry()' in 'rlb_initialize()' and
'tlb_init_table_entry()' in 'tlb_initialize()', there shouldn't double
the operations.

The key-point is avoiding TLB/RLB hash table is set to NULL while closing.
As for the last question, I think this issue due to timeline between
close processing and TLB / RLB operating.

If all these resetting operations are in 'bond_close()' instead of
'bond_open()', there still exist the same race conditions.

For example:
bond_close------------bond_start_xmit
spin_lock_bh----------bond_alb_xmit
tlb_init_table_entry--tlb_choose_channel
spin_unlock_bh--------//wait for lock
----------------------spin_lock_bh
----------------------tx_hashtbl[hash_index].tx_slave
----------------------spin_unlock_bh

If 'bond_close()' preempts the lock before TX, it won't be able to choose
the correct tx_slave in 'tlb_choose_channel()'.

However, if all these are in 'bond_open()', when the same race conditions
happen, TX can still work normally.


 Looking forward to your opinion on my reply!

-sujing

if (bond->params.tlb_dynamic_lb || BOND_MODE(bond) == BOND_MODE_ALB)
queue_delayed_work(bond->wq, &bond->alb_work, 0);
}
@@ -4279,8 +4298,6 @@ static int bond_close(struct net_device *bond_dev)

bond_work_cancel_all(bond);
bond->send_peer_notif = 0;
- if (bond_is_lb(bond))
- bond_alb_deinitialize(bond);
bond->recv_probe = NULL;

if (bond_uses_primary(bond)) {
@@ -5854,6 +5871,8 @@ static void bond_uninit(struct net_device *bond_dev)
struct list_head *iter;
struct slave *slave;

+ bond_alb_deinitialize(bond);
+
bond_netpoll_cleanup(bond_dev);

/* Release the bonded slaves */
@@ -6295,6 +6314,12 @@ static int bond_init(struct net_device *bond_dev)
bond_dev->addr_assign_type == NET_ADDR_PERM)
eth_hw_addr_random(bond_dev);

+ /* bond_alb_initialize must be called before the timer
+ * is started.
+ */
+ if (bond_alb_initialize(bond))
+ return -ENOMEM;
+
return 0;
}

diff --git a/include/net/bond_alb.h b/include/net/bond_alb.h
index 9dc082b2d543..9fd16e20ef82 100644
--- a/include/net/bond_alb.h
+++ b/include/net/bond_alb.h
@@ -150,7 +150,7 @@ struct alb_bond_info {
*/
};

-int bond_alb_initialize(struct bonding *bond, int rlb_enabled);
+int bond_alb_initialize(struct bonding *bond);
void bond_alb_deinitialize(struct bonding *bond);
int bond_alb_init_slave(struct bonding *bond, struct slave *slave);
void bond_alb_deinit_slave(struct bonding *bond, struct slave *slave);
@@ -165,5 +165,8 @@ struct slave *bond_xmit_tlb_slave_get(struct bonding *bond,
void bond_alb_monitor(struct work_struct *);
int bond_alb_set_mac_address(struct net_device *bond_dev, void *addr);
void bond_alb_clear_vlan(struct bonding *bond, unsigned short vlan_id);
+int rlb_arp_recv(const struct sk_buff *skb, struct bonding *bond, struct slave *slave);
+void tlb_init_table_entry(struct tlb_client_info *entry, int save_load);
+void rlb_init_table_entry(struct rlb_client_info *entry);
#endif /* _NET_BOND_ALB_H */

--
2.27.0

---
-Jay Vosburgh, jay.vosburgh@xxxxxxxxxxxxx
---
                    -sujing    <sujing@xxxxxxxxx>