| Message ID | 157669169826.21991.16708899415880562587.stgit@warthog.procyon.org.uk (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | [1/2] rxrpc: Unlock new call in rxrpc_new_incoming_call() rather than the caller | expand |
On Wed, Dec 18, 2019 at 05:54:58PM +0000, David Howells wrote: > Standard kernel mutexes cannot be used in any way from interrupt or softirq > context, so the user_mutex which manages access to a call cannot be a mutex > since on a new call the mutex must start off locked and be unlocked within > the softirq handler to prevent userspace interfering with a call we're > setting up. > > Commit a0855d24fc22d49cdc25664fb224caee16998683 ("locking/mutex: Complain > upon mutex API misuse in IRQ contexts") causes big warnings to be splashed > in dmesg for each a new call that comes in from the server. Whilst it > *seems* like it should be okay, since the accept path uses trylock, there > are issues with PI boosting and marking the wrong task as the owner. > > Fix this by not taking the mutex in the softirq path at all. It's not > obvious that there should be any need for it as the state is set before the > first notification is generated for the new call. > > There's also no particular reason why the link-assessing ping should be > triggered inside the mutex. It's not actually transmitted there anyway, > but rather it has to be deferred to a workqueue. > > Further, I don't think that there's any particular reason that the socket > notification needs to be done from within rx->incoming_lock, so the amount > of time that lock is held can be shortened too and the ping prepared before > the new call notification is sent. > Assuming this works, this is the best solution possible! Excellent work. (I was about to suggest something based on wait_var_event() inside each mutex_lock(), but this is _much_ nicer) Thanks!
diff --git a/net/rxrpc/call_accept.c b/net/rxrpc/call_accept.c index 3685b1732f65..44fa22b020ef 100644 --- a/net/rxrpc/call_accept.c +++ b/net/rxrpc/call_accept.c @@ -381,18 +381,6 @@ struct rxrpc_call *rxrpc_new_incoming_call(struct rxrpc_local *local, trace_rxrpc_receive(call, rxrpc_receive_incoming, sp->hdr.serial, sp->hdr.seq); - /* Lock the call to prevent rxrpc_kernel_send/recv_data() and - * sendmsg()/recvmsg() inconveniently stealing the mutex once the - * notification is generated. - * - * The BUG should never happen because the kernel should be well - * behaved enough not to access the call before the first notification - * event and userspace is prevented from doing so until the state is - * appropriate. - */ - if (!mutex_trylock(&call->user_mutex)) - BUG(); - /* Make the call live. */ rxrpc_incoming_call(rx, call, skb); conn = call->conn; @@ -433,6 +421,9 @@ struct rxrpc_call *rxrpc_new_incoming_call(struct rxrpc_local *local, BUG(); } spin_unlock(&conn->state_lock); + spin_unlock(&rx->incoming_lock); + + rxrpc_send_ping(call, skb); if (call->state == RXRPC_CALL_SERVER_ACCEPTING) rxrpc_notify_socket(call); @@ -444,11 +435,6 @@ struct rxrpc_call *rxrpc_new_incoming_call(struct rxrpc_local *local, */ rxrpc_put_call(call, rxrpc_call_put); - spin_unlock(&rx->incoming_lock); - - rxrpc_send_ping(call, skb); - mutex_unlock(&call->user_mutex); - _leave(" = %p{%d}", call, call->debug_id); return call;
Standard kernel mutexes cannot be used in any way from interrupt or softirq context, so the user_mutex which manages access to a call cannot be a mutex since on a new call the mutex must start off locked and be unlocked within the softirq handler to prevent userspace interfering with a call we're setting up. Commit a0855d24fc22d49cdc25664fb224caee16998683 ("locking/mutex: Complain upon mutex API misuse in IRQ contexts") causes big warnings to be splashed in dmesg for each a new call that comes in from the server. Whilst it *seems* like it should be okay, since the accept path uses trylock, there are issues with PI boosting and marking the wrong task as the owner. Fix this by not taking the mutex in the softirq path at all. It's not obvious that there should be any need for it as the state is set before the first notification is generated for the new call. There's also no particular reason why the link-assessing ping should be triggered inside the mutex. It's not actually transmitted there anyway, but rather it has to be deferred to a workqueue. Further, I don't think that there's any particular reason that the socket notification needs to be done from within rx->incoming_lock, so the amount of time that lock is held can be shortened too and the ping prepared before the new call notification is sent. Fixes: 540b1c48c37a ("rxrpc: Fix deadlock between call creation and sendmsg/recvmsg") Signed-off-by: David Howells <dhowells@redhat.com> cc: Peter Zijlstra <peterz@infradead.org> cc: Ingo Molnar <mingo@redhat.com> cc: Will Deacon <will@kernel.org> cc: Davidlohr Bueso <dave@stgolabs.net> --- net/rxrpc/call_accept.c | 20 +++----------------- 1 file changed, 3 insertions(+), 17 deletions(-)