On 8 Dec 2023, at 13:16, Trond Myklebust wrote:
> On Fri, 2023-12-08 at 11:33 -0500, Benjamin Coddington wrote:
>> After commit 59464b262ff5 ("SUNRPC: SOFTCONN tasks should time out
>> when on
>> the sending list"), any 4.1 backchannel tasks placed on the sending
>> queue
>> would immediately return with -ETIMEDOUT since their req timers are
>> zero.
>> We can fix this by keeping a copy of the rpc_clnt's timeout params on
>> the
>> transport and using them to properly setup the timeouts on the v4.1
>> backchannel tasks' req.
>>
>> Fixes: 59464b262ff5 ("SUNRPC: SOFTCONN tasks should time out when on
>> the sending list")
>> Signed-off-by: Benjamin Coddington <bcodding@xxxxxxxxxx>
>> ---
>> include/linux/sunrpc/xprt.h | 1 +
>> net/sunrpc/clnt.c | 3 +++
>> net/sunrpc/xprt.c | 15 +++++++++++++--
>> 3 files changed, 17 insertions(+), 2 deletions(-)
>>
>> diff --git a/include/linux/sunrpc/xprt.h
>> b/include/linux/sunrpc/xprt.h
>> index f85d3a0daca2..7565902053f3 100644
>> --- a/include/linux/sunrpc/xprt.h
>> +++ b/include/linux/sunrpc/xprt.h
>> @@ -285,6 +285,7 @@ struct rpc_xprt {
>> * items */
>> struct list_head bc_pa_list; /* List of
>> preallocated
>> * backchannel
>> rpc_rqst's */
>> + struct rpc_timeout bc_timeout; /* backchannel
>> timeout params */
>> #endif /* CONFIG_SUNRPC_BACKCHANNEL */
>>
>> struct rb_root recv_queue; /* Receive queue */
>> diff --git a/net/sunrpc/clnt.c b/net/sunrpc/clnt.c
>> index d6805c1268a7..5891757c88b1 100644
>> --- a/net/sunrpc/clnt.c
>> +++ b/net/sunrpc/clnt.c
>> @@ -279,6 +279,9 @@ static struct rpc_xprt
>> *rpc_clnt_set_transport(struct rpc_clnt *clnt,
>> clnt->cl_autobind = 1;
>>
>> clnt->cl_timeout = timeout;
>> +#if defined(CONFIG_SUNRPC_BACKCHANNEL)
>> + memcpy(&xprt->bc_timeout, timeout, sizeof(struct
>> rpc_timeout));
>> +#endif
>> rcu_assign_pointer(clnt->cl_xprt, xprt);
>> spin_unlock(&clnt->cl_lock);
>>
>> diff --git a/net/sunrpc/xprt.c b/net/sunrpc/xprt.c
>> index 92301e32cda4..d9cbe0814fd8 100644
>> --- a/net/sunrpc/xprt.c
>> +++ b/net/sunrpc/xprt.c
>> @@ -655,9 +655,14 @@ static unsigned long
>> xprt_abs_ktime_to_jiffies(ktime_t abstime)
>>
>> static unsigned long xprt_calc_majortimeo(struct rpc_rqst *req)
>> {
>> - const struct rpc_timeout *to = req->rq_task->tk_client-
>>> cl_timeout;
>> + const struct rpc_timeout *to;
>> unsigned long majortimeo = req->rq_timeout;
>>
>> + if (req->rq_task->tk_client)
>> + to = req->rq_task->tk_client->cl_timeout;
>> + else
>> + to = &req->rq_xprt->bc_timeout;
>>
>
> If you're going to convert this function for generic use, then please
> pass the timeout 'to' as a function parameter rather than making
> assumptions here.
No problem, I'll send it, but we'll end needing to make the same assumption
calling xprt_reset_majortimeo() from xprt_adjust_timeout().
.. actually it looks like backchannel tasks never currently call
rpc_check_timeout(), so we could just send the rpc_client's rpc_timeout
there, but that looks like a potential future problem. I'll send a v2 that
way and kick off my testing again.
I always thought that NULL tk_client was definitively backchannel. Is there
a case you're worried about?
We can fix this another way, probably. Looks like this fix won't actually
end up implementing normal timeout processing without adding
rpc_check_timeout() calls to the backchannel's tk_actions.
Ben
