[PATCH 15/17] nvmet-tcp: enable TLS handshake upcall

Hannes Reinecke hare at suse.de
Mon Aug 14 23:29:34 PDT 2023


On 8/14/23 21:12, Sagi Grimberg wrote:
> 
>>>> @@ -1864,6 +1877,14 @@ static struct config_group 
>>>> *nvmet_ports_make(struct config_group *group,
>>>>           return ERR_PTR(-ENOMEM);
>>>>       }
>>>> +    if (nvme_keyring_id()) {
>>>> +        port->keyring = key_lookup(nvme_keyring_id());
>>>> +        if (IS_ERR(port->keyring)) {
>>>> +            pr_warn("NVMe keyring not available, disabling TLS\n");
>>>> +            port->keyring = NULL;
>>>
>>> why setting this to NULL?
>>>
>> It's check when changing TSAS; we can only enable TLS if the nvme 
>> keyring is available.
> 
> ok
> 
>>
>>>> +        }
>>>> +    }
>>>> +
>>>>       for (i = 1; i <= NVMET_MAX_ANAGRPS; i++) {
>>>>           if (i == NVMET_DEFAULT_ANA_GRPID)
>>>>               port->ana_state[1] = NVME_ANA_OPTIMIZED;
>>>> diff --git a/drivers/nvme/target/nvmet.h b/drivers/nvme/target/nvmet.h
>>>> index 8cfd60f3b564..7f9ae53c1df5 100644
>>>> --- a/drivers/nvme/target/nvmet.h
>>>> +++ b/drivers/nvme/target/nvmet.h
>>>> @@ -158,6 +158,7 @@ struct nvmet_port {
>>>>       struct config_group        ana_groups_group;
>>>>       struct nvmet_ana_group        ana_default_group;
>>>>       enum nvme_ana_state        *ana_state;
>>>> +    struct key            *keyring;
>>>>       void                *priv;
>>>>       bool                enabled;
>>>>       int                inline_data_size;
>>>> diff --git a/drivers/nvme/target/tcp.c b/drivers/nvme/target/tcp.c
>>>> index f19ea9d923fd..77fa339008e1 100644
>>>> --- a/drivers/nvme/target/tcp.c
>>>> +++ b/drivers/nvme/target/tcp.c
>>>> @@ -8,9 +8,13 @@
>>>>   #include <linux/init.h>
>>>>   #include <linux/slab.h>
>>>>   #include <linux/err.h>
>>>> +#include <linux/key.h>
>>>>   #include <linux/nvme-tcp.h>
>>>> +#include <linux/nvme-keyring.h>
>>>>   #include <net/sock.h>
>>>>   #include <net/tcp.h>
>>>> +#include <net/tls.h>
>>>> +#include <net/handshake.h>
>>>>   #include <linux/inet.h>
>>>>   #include <linux/llist.h>
>>>>   #include <crypto/hash.h>
>>>> @@ -66,6 +70,16 @@ device_param_cb(idle_poll_period_usecs, 
>>>> &set_param_ops,
>>>>   MODULE_PARM_DESC(idle_poll_period_usecs,
>>>>           "nvmet tcp io_work poll till idle time period in usecs: 
>>>> Default 0");
>>>> +#ifdef CONFIG_NVME_TARGET_TCP_TLS
>>>> +/*
>>>> + * TLS handshake timeout
>>>> + */
>>>> +static int tls_handshake_timeout = 10;
>>>> +module_param(tls_handshake_timeout, int, 0644);
>>>> +MODULE_PARM_DESC(tls_handshake_timeout,
>>>> +         "nvme TLS handshake timeout in seconds (default 10)");
>>>> +#endif
>>>> +
>>>>   #define NVMET_TCP_RECV_BUDGET        8
>>>>   #define NVMET_TCP_SEND_BUDGET        8
>>>>   #define NVMET_TCP_IO_WORK_BUDGET    64
>>>> @@ -122,11 +136,13 @@ struct nvmet_tcp_cmd {
>>>>   enum nvmet_tcp_queue_state {
>>>>       NVMET_TCP_Q_CONNECTING,
>>>> +    NVMET_TCP_Q_TLS_HANDSHAKE,
>>>>       NVMET_TCP_Q_LIVE,
>>>>       NVMET_TCP_Q_DISCONNECTING,
>>>>   };
>>>>   struct nvmet_tcp_queue {
>>>> +    struct kref        kref;
>>>
>>> Why is kref the first member of the struct?
>>>
>> Habit.
>> I don't mind where it'll end up.
> 
> Move it to the back together with the tls section.
> 
>>
>>>>       struct socket        *sock;
>>>>       struct nvmet_tcp_port    *port;
>>>>       struct work_struct    io_work;
>>>> @@ -155,6 +171,10 @@ struct nvmet_tcp_queue {
>>>>       struct ahash_request    *snd_hash;
>>>>       struct ahash_request    *rcv_hash;
>>>> +    /* TLS state */
>>>> +    key_serial_t        tls_pskid;
>>>> +    struct delayed_work    tls_handshake_work;
>>>> +
>>>>       unsigned long           poll_end;
>>>>       spinlock_t        state_lock;
>>>> @@ -1283,12 +1303,21 @@ static int nvmet_tcp_try_recv(struct 
>>>> nvmet_tcp_queue *queue,
>>>>       return ret;
>>>>   }
>>>> +static void nvmet_tcp_release_queue(struct kref *kref)
>>>> +{
>>>> +    struct nvmet_tcp_queue *queue =
>>>> +        container_of(kref, struct nvmet_tcp_queue, kref);
>>>> +
>>>> +    WARN_ON(queue->state != NVMET_TCP_Q_DISCONNECTING);
>>>> +    queue_work(nvmet_wq, &queue->release_work);
>>>> +}
>>>> +
>>>>   static void nvmet_tcp_schedule_release_queue(struct 
>>>> nvmet_tcp_queue *queue)
>>>>   {
>>>>       spin_lock_bh(&queue->state_lock);
>>>>       if (queue->state != NVMET_TCP_Q_DISCONNECTING) {
>>>>           queue->state = NVMET_TCP_Q_DISCONNECTING;
>>>> -        queue_work(nvmet_wq, &queue->release_work);
>>>> +        kref_put(&queue->kref, nvmet_tcp_release_queue);
>>>>       }
>>>>       spin_unlock_bh(&queue->state_lock);
>>>>   }
>>>> @@ -1485,6 +1514,8 @@ static void 
>>>> nvmet_tcp_release_queue_work(struct work_struct *w)
>>>>       mutex_unlock(&nvmet_tcp_queue_mutex);
>>>>       nvmet_tcp_restore_socket_callbacks(queue);
>>>> +    tls_handshake_cancel(queue->sock->sk);
>>>> +    cancel_delayed_work_sync(&queue->tls_handshake_work);
>>>
>>> We should call it tls_handshake_tmo_work or something to make it
>>> clear it is a timeout work.
>>>
>> Okay.
>>
>>>>       cancel_work_sync(&queue->io_work);
>>>>       /* stop accepting incoming data */
>>>>       queue->rcv_state = NVMET_TCP_RECV_ERR;
>>>> @@ -1512,8 +1543,13 @@ static void nvmet_tcp_data_ready(struct sock 
>>>> *sk)
>>>>       read_lock_bh(&sk->sk_callback_lock);
>>>>       queue = sk->sk_user_data;
>>>> -    if (likely(queue))
>>>> -        queue_work_on(queue_cpu(queue), nvmet_tcp_wq, 
>>>> &queue->io_work);
>>>> +    if (likely(queue)) {
>>>> +        if (queue->data_ready)
>>>> +            queue->data_ready(sk);
>>>> +        if (queue->state != NVMET_TCP_Q_TLS_HANDSHAKE)
>>>> +            queue_work_on(queue_cpu(queue), nvmet_tcp_wq,
>>>> +                      &queue->io_work);
>>>> +    }
>>>>       read_unlock_bh(&sk->sk_callback_lock);
>>>>   }
>>>> @@ -1621,6 +1657,83 @@ static int nvmet_tcp_set_queue_sock(struct 
>>>> nvmet_tcp_queue *queue)
>>>>       return ret;
>>>>   }
>>>> +#ifdef CONFIG_NVME_TARGET_TCP_TLS
>>>> +static void nvmet_tcp_tls_handshake_done(void *data, int status,
>>>> +                     key_serial_t peerid)
>>>> +{
>>>> +    struct nvmet_tcp_queue *queue = data;
>>>> +
>>>> +    pr_debug("queue %d: TLS handshake done, key %x, status %d\n",
>>>> +         queue->idx, peerid, status);
>>>> +    spin_lock_bh(&queue->state_lock);
>>>> +    if (queue->state != NVMET_TCP_Q_TLS_HANDSHAKE) {
>>>
>>> Is this even possible?
>>>
>> I guess it can happen when the socket closes during handshake; the 
>> daemon might still be sending a 'done' event but 
>> nvmet_tcp_schedule_release_queue() has been called.
> 
> Umm, if the socket closes during the handshake then the state
> is NVMET_TCP_Q_TLS_HANDSHAKE.
> 
But there's a race window between setting it to 
NVMET_TCP_Q_DISCONNECTING and calling tls_handshake_cancel().

> p.s. you call handshake cancel in the release flow so you should be
> fenced properly no?
Not really. But I'll check if I can fix it up.

Cheers,

Hannes
-- 
Dr. Hannes Reinecke                Kernel Storage Architect
hare at suse.de                              +49 911 74053 688
SUSE Software Solutions GmbH, Maxfeldstr. 5, 90409 Nürnberg
HRB 36809 (AG Nürnberg), Geschäftsführer: Ivo Totev, Andrew
Myers, Andrew McDonald, Martje Boudien Moerman




More information about the Linux-nvme mailing list