Hi Hillf,
Please find my comments inline -
On 5/8/20 4:03 AM, Hillf Danton wrote:
> On Thu, 7 May 2020 12:36:29 Mark Bloch wrote:
>> On 5/7/2020 11:34, Divya Indi wrote:
>>> This patch fixes commit -
>>> commit 3ebd2fd0d011 ("IB/sa: Put netlink request into the request list before sending")'
>>>
>>> Above commit adds the query to the request list before ib_nl_snd_msg.
>>>
>>> However, if there is a delay in sending out the request (For
>>> eg: Delay due to low memory situation) the timer to handle request timeout
>>> might kick in before the request is sent out to ibacm via netlink.
>>> ib_nl_request_timeout may release the query if it fails to send it to SA
>>> as well causing a use after free situation.
>>>
>>> Call Trace for the above race:
>>>
>>> [<ffffffffa02f43cb>] ? ib_pack+0x17b/0x240 [ib_core]
>>> [<ffffffffa032aef1>] ib_sa_path_rec_get+0x181/0x200 [ib_sa]
>>> [<ffffffffa0379db0>] rdma_resolve_route+0x3c0/0x8d0 [rdma_cm]
>>> [<ffffffffa0374450>] ? cma_bind_port+0xa0/0xa0 [rdma_cm]
>>> [<ffffffffa040f850>] ? rds_rdma_cm_event_handler_cmn+0x850/0x850
>>> [rds_rdma]
>>> [<ffffffffa040f22c>] rds_rdma_cm_event_handler_cmn+0x22c/0x850
>>> [rds_rdma]
>>> [<ffffffffa040f860>] rds_rdma_cm_event_handler+0x10/0x20 [rds_rdma]
>>> [<ffffffffa037778e>] addr_handler+0x9e/0x140 [rdma_cm]
>>> [<ffffffffa026cdb4>] process_req+0x134/0x190 [ib_addr]
>>> [<ffffffff810a02f9>] process_one_work+0x169/0x4a0
>>> [<ffffffff810a0b2b>] worker_thread+0x5b/0x560
>>> [<ffffffff810a0ad0>] ? flush_delayed_work+0x50/0x50
>>> [<ffffffff810a68fb>] kthread+0xcb/0xf0
>>> [<ffffffff816ec49a>] ? __schedule+0x24a/0x810
>>> [<ffffffff816ec49a>] ? __schedule+0x24a/0x810
>>> [<ffffffff810a6830>] ? kthread_create_on_node+0x180/0x180
>>> [<ffffffff816f25a7>] ret_from_fork+0x47/0x90
>>> [<ffffffff810a6830>] ? kthread_create_on_node+0x180/0x180
>>> ....
>>> RIP [<ffffffffa03296cd>] send_mad+0x33d/0x5d0 [ib_sa]
>>>
>>> To resolve this issue, we introduce a new flag IB_SA_NL_QUERY_SENT.
>>> This flag Indicates if the request has been sent out to ibacm yet.
>>>
>>> If this flag is not set for a query and the query times out, we add it
>>> back to the list with the original delay.
>>>
>>> To handle the case where a response is received before we could set this
>>> flag, the response handler waits for the flag to be
>>> set before proceeding with the query.
>>>
>>> Signed-off-by: Divya Indi <divya.indi@xxxxxxxxxx>
>>> ---
>>> drivers/infiniband/core/sa_query.c | 45 ++++++++++++++++++++++++++++++++++++++
>>> 1 file changed, 45 insertions(+)
>>>
>>> diff --git a/drivers/infiniband/core/sa_query.c b/drivers/infiniband/core/sa_query.c
>>> index 30d4c12..ffbae2f 100644
>>> --- a/drivers/infiniband/core/sa_query.c
>>> +++ b/drivers/infiniband/core/sa_query.c
>>> @@ -59,6 +59,9 @@
>>> #define IB_SA_LOCAL_SVC_TIMEOUT_MAX 200000
>>> #define IB_SA_CPI_MAX_RETRY_CNT 3
>>> #define IB_SA_CPI_RETRY_WAIT 1000 /*msecs */
>>> +
>>> +DECLARE_WAIT_QUEUE_HEAD(wait_queue);
>>> +
>>> static int sa_local_svc_timeout_ms = IB_SA_LOCAL_SVC_TIMEOUT_DEFAULT;
>>>
>>> struct ib_sa_sm_ah {
>>> @@ -122,6 +125,7 @@ struct ib_sa_query {
>>> #define IB_SA_ENABLE_LOCAL_SERVICE 0x00000001
>>> #define IB_SA_CANCEL 0x00000002
>>> #define IB_SA_QUERY_OPA 0x00000004
>>> +#define IB_SA_NL_QUERY_SENT 0x00000008
>>>
>>> struct ib_sa_service_query {
>>> void (*callback)(int, struct ib_sa_service_rec *, void *);
>>> @@ -746,6 +750,11 @@ static inline int ib_sa_query_cancelled(struct ib_sa_query *query)
>>> return (query->flags & IB_SA_CANCEL);
>>> }
>>>
>>> +static inline int ib_sa_nl_query_sent(struct ib_sa_query *query)
>>> +{
>>> + return (query->flags & IB_SA_NL_QUERY_SENT);
>>> +}
>>> +
>>> static void ib_nl_set_path_rec_attrs(struct sk_buff *skb,
>>> struct ib_sa_query *query)
>>> {
>>> @@ -889,6 +898,15 @@ static int ib_nl_make_request(struct ib_sa_query *query, gfp_t gfp_mask)
>>> spin_lock_irqsave(&ib_nl_request_lock, flags);
>>> list_del(&query->list);
>>> spin_unlock_irqrestore(&ib_nl_request_lock, flags);
> Here it also needs to do wakeup as sleeper might come while sending, no
> matter the failure to send.
We only sleep in the response handler. If we could not send a query, then
response handler is not triggered and hence no one would be waiting on this query.
Thanks,
Divya
>>> + } else {
>>> + query->flags |= IB_SA_NL_QUERY_SENT;
>>> +
>>> + /*
>>> + * If response is received before this flag was set
>>> + * someone is waiting to process the response and release the
>>> + * query.
>>> + */
>>> + wake_up(&wait_queue);
>>> }
>>>
>>> return ret;
> With minor changes added on top of your work it now looks like the diff
> below (only including the wakeup part).
>
> --- a/drivers/infiniband/core/sa_query.c
> +++ b/drivers/infiniband/core/sa_query.c
> @@ -863,6 +863,11 @@ static int ib_nl_send_msg(struct ib_sa_q
> return rdma_nl_multicast(&init_net, skb, RDMA_NL_GROUP_LS, gfp_mask);
> }
>
> +static inline bool ib_sa_nl_query_sending(struct ib_sa_query *query)
> +{
> + return query->flags & IB_SA_NL_QUERY_SENDING;
> +}
> +
> static int ib_nl_make_request(struct ib_sa_query *query, gfp_t gfp_mask)
> {
> unsigned long flags;
> @@ -874,6 +879,8 @@ static int ib_nl_make_request(struct ib_
>
> /* Put the request on the list first.*/
> spin_lock_irqsave(&ib_nl_request_lock, flags);
> + /* info others we're having work to do */
> + query->flags |= IB_SA_NL_QUERY_SENDING;
> delay = msecs_to_jiffies(sa_local_svc_timeout_ms);
> query->timeout = delay + jiffies;
> list_add_tail(&query->list, &ib_nl_request_list);
> @@ -883,13 +890,15 @@ static int ib_nl_make_request(struct ib_
> spin_unlock_irqrestore(&ib_nl_request_lock, flags);
>
> ret = ib_nl_send_msg(query, gfp_mask);
> + /* safe to delete it with IB_SA_NL_QUERY_SENDING set */
> + spin_lock_irqsave(&ib_nl_request_lock, flags);
> if (ret) {
> ret = -EIO;
> - /* Remove the request */
> - spin_lock_irqsave(&ib_nl_request_lock, flags);
> list_del(&query->list);
> - spin_unlock_irqrestore(&ib_nl_request_lock, flags);
> }
> + query->flags &= ~IB_SA_NL_QUERY_SENDING;
> + wake_up(&wait_queue);
> + spin_unlock_irqrestore(&ib_nl_request_lock, flags);
>
> return ret;
> }
>
>
>>> @@ -994,6 +1012,21 @@ static void ib_nl_request_timeout(struct work_struct *work)
>>> }
>>>
>>> list_del(&query->list);
>>> +
>>> + /*
>>> + * If IB_SA_NL_QUERY_SENT is not set, this query has not been
>>> + * sent to ibacm yet. Reset the timer.
>>> + */
>>> + if (!ib_sa_nl_query_sent(query)) {
>>> + delay = msecs_to_jiffies(sa_local_svc_timeout_ms);
>>> + query->timeout = delay + jiffies;
>>> + list_add_tail(&query->list, &ib_nl_request_list);
>>> + /* Start the timeout if this is the only request */
>>> + if (ib_nl_request_list.next == &query->list)
>>> + queue_delayed_work(ib_nl_wq, &ib_nl_timed_work,
>>> + delay);
>>> + break;
>>> + }
>>> ib_sa_disable_local_svc(query);
>>> /* Hold the lock to protect against query cancellation */
>>> if (ib_sa_query_cancelled(query))
>>> @@ -1123,6 +1156,18 @@ int ib_nl_handle_resolve_resp(struct sk_buff *skb,
>>>
>>> send_buf = query->mad_buf;
>>>
>>> + /*
>>> + * Make sure the IB_SA_NL_QUERY_SENT flag is set before
>>> + * processing this query. If flag is not set, query can be accessed in
>>> + * another context while setting the flag and processing the query will
>>> + * eventually release it causing a possible use-after-free.
>>> + */
>>> + if (unlikely(!ib_sa_nl_query_sent(query))) {
>> Can't there be a race here where you check the flag (it isn't set)
>> and before you call wait_event() the flag is set and wake_up() is called
>> which means you will wait here forever? or worse, a timeout will happen
>> the query will be freed and them some other query will call wake_up()
>> and we have again a use-after-free.
>>
>>> + spin_unlock_irqrestore(&ib_nl_request_lock, flags);
>>> + wait_event(wait_queue, ib_sa_nl_query_sent(query));
>> What if there are two queries sent to userspace, shouldn't you check and make sure
>> you got woken up by the right one setting the flag?
>>
>> Other than that, the entire solution makes it very complicated to reason with (flags set/checked without locking etc)
>> maybe we should just revert and fix it the other way?
>>
>> Mark
>>
>>> + spin_lock_irqsave(&ib_nl_request_lock, flags);
>>> + }
>>> +
>>> if (!ib_nl_is_good_resolve_resp(nlh)) {
>>> /* if the result is a failure, send out the packet via IB */
>>> ib_sa_disable_local_svc(query);
>>>
