Return-Path: From: "Peter Krystad" To: "'Marcel Holtmann'" Cc: "'Emeltchenko Andrei'" , References: <1322661272-32027-1-git-send-email-Andrei.Emeltchenko.news@gmail.com> <1322663484.26198.16.camel@aeonflux> <20111130151820.GB5158@aemeltch-MOBL1> <000001ccaff1$1ee6c0c0$5cb44240$@org> <1322738467.26198.26.camel@aeonflux> In-Reply-To: <1322738467.26198.26.camel@aeonflux> Subject: RE: [RFCv0 0/3] AMP HCI interface from A2MP Date: Thu, 1 Dec 2011 22:39:03 -0800 Message-ID: <000001ccb0bd$15c7e0b0$4157a210$@org> MIME-Version: 1.0 Content-Type: text/plain; charset="utf-8" List-ID: > -----Original Message----- > From: Marcel Holtmann [mailto:marcel@holtmann.org] > > Hi Peter, > > > > > > The code adds AMP HCI commands from A2MP protocol. HCI events are handled > > > > > similar way as mgmt interface. amp_pending is a kind of copy of mgmt_pending. > > > > > > > > this is all kernel internal code with no interface to user space. I do > > > > not see the need for such a complex infrastructure. Can not just have > > > > proper callbacks or event callback table like with L2CAP. Or just > > > > something similar. > > > > > > I see this as a simple callback. amp_pending is just keeping context of HCI > > > command we need to handle. I also included reference counting since we had > > > bad experience with l2cap and rfcomm. > > > > There does have to be some way to carry the A2MP message context while performing > > local HCI operations to service the message. A2MP Get Remote Assoc requires multiple > > HCI commands with data accumulated between the commands before the response can be > > sent. > > I agree, but this amp_pending seems to be wrong. I talked about changing > the init handling and at the same time we might can apply this to the > A2MP handling. I don't think this amp_pending is the right approach either. > Why not just set a bit in hdev->dev_flags or maybe even hcon->dev_flags > or something and based on that we do a nice async state machine that > sends the commands and results in calling back into A2MP core when its > finished. I agree we can build a nice async state machine. But it needs to be carry more data than a bit in the hci_dev->flags. 1) For every A2MP request received we have to hang onto the request identifier byte to be returned in the response. 2) For processing the A2MP Get AMP Assoc the partial assoc data has to be accumulated over multiple HCI calls before it can be returned in the response. 3) In the create physical link sequence the remote assoc has to be stored while it is fed to HCI in multiple commands, and after the channel select event the local assoc has to be accumulated before it can be sent back to the remote. > That seems a bit simpler to me and more aligned in a way I wanna redo > the whole HCI command/event handling anyway. In addition we can just > expose these flags via debugfs as text and have a nice way of knowing > current states if something goes wrong. In the CAF implementation there is a separate set of data structures that track all this context. I (and I think Andrei) think that that isn't the way to go. Storing the context in the hci_conn seems the best without adding more complexity. It needs to be on hci_conn not hci_dev because it possible for more than one remote AMP to send an A2MP command at one time. There will still need to be a flag in hci_dev->flags that serializes create physical link attempts, as PALs may require these to be serialized. There is a union in the structure amp_ctx in the CAF code that summarizes the data that needs to kept around. I've included it at the end of this e-mail. I'm not suggesting use this as is, just that it shows what state has to be kept while processing the A2MP requests. Regards, Peter. > > > > As far as I see it, we get an A2MP command over L2CAP fixed channel, we > > > > have to issue a HCI command or do some other task based on this and then > > > > respond to it. We only have one user of this first of all. And second of > > > > all, I think we can not really have pending A2MP commands anyway. This > > > > is pretty much one command at a time (per ACL link). > > > > A2MP commands are serialized by the sender, so A2MP message context could be > > associated with the hci_conn for BR-EDR. > > That is what I thought. Thanks for confirming. So using hcon->dev_flags > seems like a possible way to make this a lot simpler. > > And if the sender misbehaves, we just reject that commands that way. > Testing a flag is always cheaper then iterating a list. > > Regards > > Marcel > /* Get AMP Assoc sequence */ struct amp_gaa_state { __u8 req_ident; __u16 len_so_far; __u8 *assoc; }; /* Create Physical Link sequence */ struct amp_cpl_state { __u8 remote_id; __u16 max_len; __u8 *remote_assoc; __u8 *local_assoc; __u16 len_so_far; __u16 rem_len; __u8 phy_handle; }; /* Accept Physical Link sequence */ struct amp_apl_state { __u8 remote_id; __u8 req_ident; __u8 *remote_assoc; __u16 len_so_far; __u16 rem_len; __u8 phy_handle; }; struct amp_ctx { struct list_head list; struct amp_mgr *mgr; struct hci_dev *hdev; __u8 type; __u8 state; union { struct amp_gaa_state gaa; struct amp_cpl_state cpl; struct amp_apl_state apl; } d; __u8 evt_type; __u8 evt_code; __u16 opcode; __u8 id; __u8 rsp_ident; struct sock *sk; struct amp_ctx *deferred; struct timer_list timer; }; --Peter Krystad Employee of Qualcomm Innovation Center, Inc. Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum