This is the completion of the CCID3 fixes begun with the last patch set, plus work on DCCP passive-close. I apologize for the mass of output. On the other hand, it is probably not that much, since I have been anxious to split each logical change into a separate patch. I have not always managed to eliminate overriding by a later one, but tried to avoid this as good as possible. The patches divide into the following groups, the bulk of work is in II/III. All patches have been uploaded to http://www.erg.abdn.ac.uk/users/gerrit/dccp/patch-backlog/ , on top of the basic CCID3 patches. Due to the mass of outstanding patches, I offer to split them into separate directories and create separate blocks of descriptions -- please do let me know if you would like to make use of this. ========================================================================================================= I. Basic CCID3 Patches ====================== Patch #1: Implements circular modulo-16 arithmetic for working with window counters. Patch #2: Bug-fix: CCID3 receiver (hc_rx) does not set CCVal window counter. Patch #3: Remove the redundancy of keeping both p and p_inv = 1/p on the CCID3 socket. The socket size shrinks a bit further and the code becomes simpler. Patch #4: Avoids redundant history access when copying received-packet information over to send CCID3 feedback. Simplification and less prone to problems with concurrent access to data structures. Patch #5: Provide dedicated "tfrc_module.c" source file to mark structure of tfrc_lib; and provide the missing "tfrc_pr_debug()" macro for debugging. ========================================================================================================= II. New CCID3 RX History Structure | ==================================== The RX history structure has been much revised, in particular it has been reduced to a fixed-size rinbuffer of size 4. This is the minimally required size to perform loss detection. In addition, a more robust form of receiver-based RTT sampling is also performed on top of this structure. As a result, more functionality becomes available with simpler code. Patch #6: New RX history infrastructure - data structures. Patch #7: "Glue patch" to allow the transition, using mainly #if 0 ... to allow a transition. This patch is entirely overwritten by later patches, but is necessary if people want to test just the first 6 patches on their own. Patch #8: RX history constructor / destructor. NB: I left the old dccp_rx_hist_new() / dccp_rx_hist_delete() in case someone wants to build some other RX history in the future; but in fact none of the code uses them now. Patch #9: Takes care of always-tracking the highest-received sequence number in the new RX structure. Patch #10: Fallback RTT value from RFC 4340, 4.3. Seems to be an oversight, we need a fallback RTT in quite a few cases; this patch adds it for DCCP-wide use; plus update on RTT sanity-check. Patch #11: Receiver RTT sampling algorithm - I fine-tuned the algorithm from RFC 4342 for robustness. Patch #12: Integrate new RX history structure with CCID3 packet reception (hc_rx). Now has all the features and checks required by RFC 3448/4342 on receiving a new CCID3 packet. Patch #13: Add loss detection, complete with handling reordering/late arrival and filling of seqno holes. Patch #14: Loss detection continued - recycling of loss records under bursts of losses. Patch #15: Removes the old RX history functions which have been replaced. ========================================================================================================= II. New CCID3 Loss Intervals Database Structure | ================================================= This implements an entirely new loss intervals database which interfaces with the RX history structure. It is also part of the tfrc_lib module and so has the added benefit that (i) loss detection, (ii) detection whether a loss is indeed new and (iii) adding a new loss interval record to the LI database can all be done in one step - as a service provided by the tfrc_lib module. A nice feature is on-demand allocation, i.e. loss-free connections have very little memory consumption. Patch #16: Basic ringbuffer structure (just declarations and some inline test/initialisation functions). Patch #17: Access macros to encapsulate access to the internals of the structure. Patch #18: Slab cache on-demand allocation and de-allocation. There is one central slab cache which works as `array factory' for all clients. No memory is allocated when the connection is loss-free. Patch #19: Routines for computing and updating I_mean as required by RFC 3448, 6. Since this is a TFRC mechanism, it is also provided and exported as a service of the tfrc_lib module. Patch #20: Insert and Update routines for the LI database, as well as support for CCVal-based decision whether a loss is new (using the test from RFC 4342, 10.2). Patch #21: Makes calculation of first loss interval local to the client of the tfrc_lib module. This is better since the caller knows better its own variables. Patch #22: Integrate the Loss Interval database with RX history so that they can talk to each other. Patch #23: Fix header #includes so that services exported by tfrc_lib become accessible. Patch #24: Hook up the new LI database with CCID3. ========================================================================================================= III. Miscellaneous CCID3 patches | ================================== Patch #25: Update copyrights - this is Ian's patch. I have added University of Aberdeen; they paid for this work. Patch #26: Simplifies an often-repeated test for NULL by doing it centrally in the conversion/cast function. Patch #27: Bug fix - handle CCID3 "Idle and Application-Limited Periods" correctly. This wasn't done correctly in the current code. The actual interface is open to debate/modularisation. For now, I have used the definition of idleness from rfc3448bis -- found this a useful one. Patch #28: Bug fix - X_recv was not computed over the last R_m seconds as required by RFC 3448, 6. The solution used by this patch also makes the feedback handling more transparent, by using explicit states. Patch #29: Axe a few chars from the long variable name hcrx->ccid3hcrx_tstamp_last_feedback. ========================================================================================================= IV. Work on DCCP States | ========================= This contains updates to the internal naming and handling of states. It clarifies the interrelation with TCP statenames, and adds the two new states PASSIVE_1/PASSIVE_2 which are necessary to handle passive-close correctly (blurb is on http://www.erg.abdn.ac.uk/users/gerrit/dccp/docs/closing_states/ ) During the past days I have been leaning over backwards to try and make this break. I could not find any way to create a crashing or sick termination, it terminates cleanly in many settings now. But it could still be that I missed something, I'd be grateful for any comments and suggestions. Patch #30: Insert a divider to the DCCP-states enum, to separate specific from general states. Patch #31: Decouples PARTOPEN from TCP-specific stream-states, as suggested by a FIXME-comment. Patch #32: Implement dedicated PASSIVE_1/2 states to support passive-close. Patch #33: Integrate the new states from #32 and add necessary state transitions. Patch #34: Perform both SHUT_RD and SHUT_WR on close (to respect DCCP's absence of half-close). Patch #35: When the node is in state CLOSING, no more data is accepted (in the same spirit as #34). Patch #36: Shift retransmit timer for active-close from proto.c into output.c. Patch #37: Use the maximum RTO from RFC spec (64 secs instead of TCP 120 sec); another Fixme. Patch #38: Support for server holding TIMEWAIT state - as in the RFC Patch #39: Protect against Reset/Sync floods. This will happen when client crashes during connection. Patch #40: Collapse repeated `len' statements into one. ========================================================================================================= IV. Work on DCCP abnormal termination | ======================================= Some missing bits are added to support active reset when a client crashes or is terminated with a SO_LINGER time of 0; the API now behaves as in TCP. Patch #41: Check for unread data on close (same handling now as in TCP). Patch #42: Implement TCP-like ABORT function. ========================================================================================================= V. Last | ========= Patch #43: Update API Documentation. - To unsubscribe from this list: send the line "unsubscribe dccp" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
