Hi Leon,
Please find my comments inline -
On 6/13/20 11:41 PM, Leon Romanovsky wrote:
> On Tue, Jun 09, 2020 at 07:45:21AM -0700, Divya Indi wrote:
>> Hi Leon,
>>
>> Thanks for taking the time to review.
>>
>> Please find my comments inline -
>>
>> On 6/9/20 12:00 AM, Leon Romanovsky wrote:
>>> On Mon, Jun 08, 2020 at 07:46:16AM -0700, Divya Indi wrote:
>>>> Commit 3ebd2fd0d011 ("IB/sa: Put netlink request into the request list before sending")'
>>>> -
>>>> 1. Adds the query to the request list before ib_nl_snd_msg.
>>>> 2. Removes ib_nl_send_msg from within the spinlock which also makes it
>>>> possible to allocate memory with GFP_KERNEL.
>>>>
>>>> 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 causing a use after free situation
>>>> while accessing the query in ib_nl_send_msg.
>>>>
>>>> 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 the above issue -
>>>> 1. Add the req to the request list only after the request has been sent out.
>>>> 2. To handle the race where response comes in before adding request to
>>>> the request list, send(rdma_nl_multicast) and add to list while holding the
>>>> spinlock - request_lock.
>>>> 3. Use GFP_NOWAIT for rdma_nl_multicast since it is called while holding
>>>> a spinlock. In case of memory allocation failure, request will go out to SA.
>>>>
>>>> Signed-off-by: Divya Indi <divya.indi@xxxxxxxxxx>
>>>> Fixes: 3ebd2fd0d011 ("IB/sa: Put netlink request into the request list
>>>> before sending")
>>> Author SOB should be after "Fixes" line.
>> My bad. Noted.
>>
>>>> ---
>>>> drivers/infiniband/core/sa_query.c | 34 +++++++++++++++++-----------------
>>>> 1 file changed, 17 insertions(+), 17 deletions(-)
>>>>
>>>> diff --git a/drivers/infiniband/core/sa_query.c b/drivers/infiniband/core/sa_query.c
>>>> index 74e0058..042c99b 100644
>>>> --- a/drivers/infiniband/core/sa_query.c
>>>> +++ b/drivers/infiniband/core/sa_query.c
>>>> @@ -836,6 +836,9 @@ static int ib_nl_send_msg(struct ib_sa_query *query, gfp_t gfp_mask)
>>>> void *data;
>>>> struct ib_sa_mad *mad;
>>>> int len;
>>>> + unsigned long flags;
>>>> + unsigned long delay;
>>>> + int ret;
>>>>
>>>> mad = query->mad_buf->mad;
>>>> len = ib_nl_get_path_rec_attrs_len(mad->sa_hdr.comp_mask);
>>>> @@ -860,35 +863,32 @@ static int ib_nl_send_msg(struct ib_sa_query *query, gfp_t gfp_mask)
>>>> /* Repair the nlmsg header length */
>>>> nlmsg_end(skb, nlh);
>>>>
>>>> - return rdma_nl_multicast(&init_net, skb, RDMA_NL_GROUP_LS, gfp_mask);
>>>> + spin_lock_irqsave(&ib_nl_request_lock, flags);
>>>> + ret = rdma_nl_multicast(&init_net, skb, RDMA_NL_GROUP_LS, GFP_NOWAIT);
>>> It is hard to be convinced that this is correct solution. The mix of
>>> gfp_flags and GFP_NOWAIT at the same time and usage of
>>> ib_nl_request_lock to protect lists and suddenly rdma_nl_multicast() too
>>> makes this code unreadable/non-maintainable.
>> Prior to 3ebd2fd0d011 ("IB/sa: Put netlink request into the request list
>> before sending"), we had ib_nl_send_msg under the spinlock ib_nl_request_lock.
>>
>> ie we had -
>>
>> 1. Get spinlock - ib_nl_request_lock
>> 2. ib_nl_send_msg
>> 2.a) rdma_nl_multicast
>> 3. Add request to the req list
>> 4. Arm the timer if needed.
>> 5. Release spinlock
>>
>> However, ib_nl_send_msg involved a memory allocation using GFP_KERNEL.
>> hence, was moved out of the spinlock. In addition, req was now being
>> added prior to ib_nl_send_msg [To handle the race where response can
>> come in before we get a chance to add the request back to the list].
>>
>> This introduced another race resulting in use-after-free.[Described in the commit.]
>>
>> To resolve this, sending out the request and adding it to list need to
>> happen while holding the request_lock.
>> To ensure minimum allocations while holding the lock, instead of having
>> the entire ib_nl_send_msg under the lock, we only have rdma_nl_multicast
>> under this spinlock.
>>
>> However, do you think it would be a good idea to split ib_nl_send_msg
>> into 2 functions -
>> 1. Prepare the req/query [Outside the spinlock]
>> 2. Sending the req - rdma_nl_multicast [while holding spinlock]
>>
>> Would this be more intuitive?
> While it is always good idea to minimize the locked period. It still
> doesn't answer concern about mixing gfp_flags and direct GFP_NOWAIT.
> For example if user provides GFP_ATOMIC, the GFP_NOWAIT allocation will
> cause a trouble because latter is more lax than first one.
Makes sense, and we do have callers passing GFP_ATOMIC with gfp_mask.
However, in this case when we fail to send the request to ibacm,
we then fallback to sending it to the SA with gfp_mask. So, the
request will eventually go out with GFP_ATOMIC to SA. From the
caller perspective the request will not fail due to memory pressure.
-------
send_mad(...gfp_mask)
- send to ibacm with GFP_NOWAIT
- If fails, send to SA with gfp_mask
-------
So, using GFP_NOWAIT may not cause trouble here.
The other option might be to use GFP_NOWAIT conditionally ie
(only use GFP_NOWAIT when GFP_ATOMIC is not specified in gfp_mask else
use GFP_ATOMIC). Eventual goal being to not have a blocking memory allocation.
Your thoughts?
Really appreciate your feedback. Thanks!
Regards,
Divya
>
> Thanks
>
>>>> + if (!ret) {
>>> Please use kernel coding style.
>>>
>>> if (ret) {
>>> spin_unlock_irqrestore(&ib_nl_request_lock, flags);
>>> return ret;
>>> }
>>>
>>> ....
>> Noted. Will make this change.
>>
>>>> + /* Put the request on the list.*/
>>>> + 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);
>>>> + }
>>>> + spin_unlock_irqrestore(&ib_nl_request_lock, flags);
>>>> +
>>>> + return ret;
>>>> }
>>>>
>>>> static int ib_nl_make_request(struct ib_sa_query *query, gfp_t gfp_mask)
>>>> {
>>>> - unsigned long flags;
>>>> - unsigned long delay;
>>>> int ret;
>>>>
>>>> INIT_LIST_HEAD(&query->list);
>>>> query->seq = (u32)atomic_inc_return(&ib_nl_sa_request_seq);
>>>>
>>>> - /* Put the request on the list first.*/
>>>> - spin_lock_irqsave(&ib_nl_request_lock, flags);
>>>> - 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);
>>>> - spin_unlock_irqrestore(&ib_nl_request_lock, flags);
>>>> -
>>>> ret = ib_nl_send_msg(query, gfp_mask);
>>>> 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);
>>>> }
>>> Brackets should be removed too.
>> Noted.
>>>> return ret;
>>>> --
>>>> 1.8.3.1
>>>>
