lwip/CHANGELOG
2006-03-01 16:57:43 +00:00

587 lines
20 KiB
Plaintext

FUTURE
* TODO: The lwIP source code makes some invalid assumptions on processor
word-length, storage sizes and alignment. See the mailing lists for
problems with exoteric (/DSP) architectures showing these problems.
We still have to fix some of these issues neatly.
* TODO: the ARP layer is not protected against concurrent access. If
you run from a multitasking OS, serialize access to ARP (called from
your network device driver and from a timeout thread.)
* TODO: the PPP code is broken in a few ways. There are namespace
collisions on BSD systems and many assumptions on word-length
(sizeof(int)). In ppp.c an assumption is made on the availability of
a thread subsystem. Either PPP needs to be moved to contrib/ports/???
or rearranged to be more generic.
HISTORY
2006-01-01 Leon Woestenberg <leon.woestenberg@gmx.net>
* tcp_{in,out}.c, tcp_out.c: Removed 'even sndbuf' fix in TCP, which is
now handled by the checksum routine properly.
2006-02-27 Leon Woestenberg <leon.woestenberg@gmx.net>
* pbuf.c: Fix alignment; pbuf_init() would not work unless
pbuf_pool_memory[] was properly aligned. (Patch by Curt McDowell.)
2005-12-20 Leon Woestenberg <leon.woestenberg@gmx.net>
* tcp.c: Remove PCBs which stay in LAST_ACK state too long. Patch
submitted by Mitrani Hiroshi.
2005-12-15 Christiaan Simons
* inet.c: Disabled the added summing routine to preserve code space.
2005-12-14 Leon Woestenberg <leon.woestenberg@gmx.net>
* tcp_in.c: Duplicate FIN ACK race condition fix by Kelvin Lawson.
Added Curt McDowell's optimized checksumming routine for future
inclusion. Need to create test case for unaliged, aligned, odd,
even length combination of cases on various endianess machines.
2005-12-09 Christiaan Simons
* inet.c: Rewrote standard checksum routine in proper portable C.
2005-11-25 Christiaan Simons
* udp.c tcp.c: Removed SO_REUSE hack. Should reside in socket code only.
* *.c: introduced cc.h LWIP_DEBUG formatters matching the u16_t, s16_t,
u32_t, s32_t typedefs. This solves most debug word-length assumes.
2005-07-17 Leon Woestenberg <leon.woestenberg@gmx.net>
* inet.c: Fixed unaligned 16-bit access in the standard checksum
routine by Peter Jolasson.
* slipif.c: Fixed implementation assumption of single-pbuf datagrams.
2005-02-04 Leon Woestenberg <leon.woestenberg@gmx.net>
* tcp_out.c: Fixed uninitialized 'queue' referenced in memerr branch.
* tcp_{out|in}.c: Applied patch fixing unaligned access.
2005-01-04 Leon Woestenberg <leon.woestenberg@gmx.net>
* pbuf.c: Fixed missing semicolon after LWIP_DEBUG statement.
2005-01-03 Leon Woestenberg <leon.woestenberg@gmx.net>
* udp.c: UDP pcb->recv() was called even when it was NULL.
(STABLE-1_1_0)
2004-12-28 Leon Woestenberg <leon.woestenberg@gmx.net>
* etharp.*: Disabled multiple packets on the ARP queue.
This clashes with TCP queueing.
2004-11-28 Leon Woestenberg <leon.woestenberg@gmx.net>
* etharp.*: Fixed race condition from ARP request to ARP timeout.
Halved the ARP period, doubled the period counts.
ETHARP_MAX_PENDING now should be at least 2. This prevents
the counter from reaching 0 right away (which would allow
too little time for ARP responses to be received).
2004-11-25 Leon Woestenberg <leon.woestenberg@gmx.net>
* dhcp.c: Decline messages were not multicast but unicast.
* etharp.c: ETHARP_CREATE is renamed to ETHARP_TRY_HARD.
Do not try hard to insert arbitrary packet's source address,
etharp_ip_input() now calls etharp_update() without ETHARP_TRY_HARD.
etharp_query() now always DOES call ETHARP_TRY_HARD so that users
querying an address will see it appear in the cache (DHCP could
suffer from this when a server invalidly gave an in-use address.)
* ipv4/ip_addr.h: Renamed ip_addr_maskcmp() to _netcmp() as we are
comparing network addresses (identifiers), not the network masks
themselves.
* ipv4/ip_addr.c: ip_addr_isbroadcast() now checks that the given
IP address actually belongs to the network of the given interface.
2004-11-24 Kieran Mansley <kjm25@cam.ac.uk>
* tcp.c: Increment pcb->snd_buf when ACK is received in SYN_SENT state.
(STABLE-1_1_0-RC1)
2004-10-16 Kieran Mansley <kjm25@cam.ac.uk>
* tcp.c: Add code to tcp_recved() to send an ACK (window update) immediately,
even if one is already pending, if the rcv_wnd is above a threshold
(currently TCP_WND/2). This avoids waiting for a timer to expire to send a
delayed ACK in order to open the window if the stack is only receiving data.
2004-09-12 Kieran Mansley <kjm25@cam.ac.uk>
* tcp*.*: Retransmit time-out handling improvement by Sam Jansen.
2004-08-20 Tony Mountifield <tony@softins.co.uk>
* etharp.c: Make sure the first pbuf queued on an ARP entry
is properly ref counted.
2004-07-27 Tony Mountifield <tony@softins.co.uk>
* debug.h: Added (int) cast in LWIP_DEBUGF() to avoid compiler
warnings about comparison.
* pbuf.c: Stopped compiler complaining of empty if statement
when LWIP_DEBUGF() empty. Closed an unclosed comment.
* tcp.c: Stopped compiler complaining of empty if statement
when LWIP_DEBUGF() empty.
* ip.h Corrected IPH_TOS() macro: returns a byte, so doesn't need htons().
* inet.c: Added a couple of casts to quiet the compiler.
No need to test isascii(c) before isdigit(c) or isxdigit(c).
2004-07-22 Tony Mountifield <tony@softins.co.uk>
* inet.c: Made data types consistent in inet_ntoa().
Added casts for return values of checksum routines, to pacify compiler.
* ip_frag.c, tcp_out.c, sockets.c, pbuf.c
Small corrections to some debugging statements, to pacify compiler.
2004-07-21 Tony Mountifield <tony@softins.co.uk>
* etharp.c: Removed spurious semicolon and added missing end-of-comment.
* ethernetif.c Updated low_level_output() to match prototype for
netif->linkoutput and changed low_level_input() similarly for consistency.
* api_msg.c: Changed recv_raw() from int to u8_t, to match prototype
of raw_recv() in raw.h and so avoid compiler error.
* sockets.c: Added trivial (int) cast to keep compiler happier.
* ip.c, netif.c Changed debug statements to use the tidier ip4_addrN() macros.
(STABLE-1_0_0)
++ Changes:
2004-07-05 Leon Woestenberg <leon.woestenberg@gmx.net>
* sockets.*: Restructured LWIP_PRIVATE_TIMEVAL. Make sure
your cc.h file defines this either 1 or 0. If non-defined,
defaults to 1.
* .c: Added <string.h> and <errno.h> includes where used.
* etharp.c: Made some array indices unsigned.
2004-06-27 Leon Woestenberg <leon.woestenberg@gmx.net>
* netif.*: Added netif_set_up()/down().
* dhcp.c: Changes to restart program flow.
2004-05-07 Leon Woestenberg <leon.woestenberg@gmx.net>
* etharp.c: In find_entry(), instead of a list traversal per candidate, do a
single-pass lookup for different candidates. Should exploit locality.
2004-04-29 Leon Woestenberg <leon.woestenberg@gmx.net>
* tcp*.c: Cleaned up source comment documentation for Doxygen processing.
* opt.h: ETHARP_ALWAYS_INSERT option removed to comply with ARP RFC.
* etharp.c: update_arp_entry() only adds new ARP entries when adviced to by
the caller. This deprecates the ETHARP_ALWAYS_INSERT overrule option.
++ Bug fixes:
2004-04-27 Leon Woestenberg <leon.woestenberg@gmx.net>
* etharp.c: Applied patch of bug #8708 by Toni Mountifield with a solution
suggested by Timmy Brolin. Fix for 32-bit processors that cannot access
non-aligned 32-bit words, such as soms 32-bit TCP/IP header fields. Fix
is to prefix the 14-bit Ethernet headers with two padding bytes.
2004-04-23 Leon Woestenberg <leon.woestenberg@gmx.net>
* ip_addr.c: Fix in the ip_addr_isbroadcast() check.
* etharp.c: Fixed the case where the packet that initiates the ARP request
is not queued, and gets lost. Fixed the case where the packets destination
address is already known; we now always queue the packet and perform an ARP
request.
(STABLE-0_7_0)
++ Bug fixes:
* Fixed TCP bug for SYN_SENT to ESTABLISHED state transition.
* Fixed TCP bug in dequeueing of FIN from out of order segment queue.
* Fixed two possible NULL references in rare cases.
(STABLE-0_6_6)
++ Bug fixes:
* Fixed DHCP which did not include the IP address in DECLINE messages.
++ Changes:
* etharp.c has been hauled over a bit.
(STABLE-0_6_5)
++ Bug fixes:
* Fixed TCP bug induced by bad window resizing with unidirectional TCP traffic.
* Packets sent from ARP queue had invalid source hardware address.
++ Changes:
* Pass-by ARP requests do now update the cache.
++ New features:
* No longer dependent on ctype.h.
* New socket options.
* Raw IP pcb support.
(STABLE-0_6_4)
++ Bug fixes:
* Some debug formatters and casts fixed.
* Numereous fixes in PPP.
++ Changes:
* DEBUGF now is LWIP_DEBUGF
* pbuf_dechain() has been re-enabled.
* Mentioned the changed use of CVS branches in README.
(STABLE-0_6_3)
++ Bug fixes:
* Fixed pool pbuf memory leak in pbuf_alloc().
Occured if not enough PBUF_POOL pbufs for a packet pbuf chain.
Reported by Savin Zlobec.
* PBUF_POOL chains had their tot_len field not set for non-first
pbufs. Fixed in pbuf_alloc().
++ New features:
* Added PPP stack contributed by Marc Boucher
++ Changes:
* Now drops short packets for ICMP/UDP/TCP protocols. More robust.
* ARP queueuing now queues the latest packet instead of the first.
This is the RFC recommended behaviour, but can be overridden in
lwipopts.h.
(0.6.2)
++ Bugfixes:
* TCP has been fixed to deal with the new use of the pbuf->ref
counter.
* DHCP dhcp_inform() crash bug fixed.
++ Changes:
* Removed pbuf_pool_free_cache and pbuf_pool_alloc_cache. Also removed
pbuf_refresh(). This has sped up pbuf pool operations considerably.
Implemented by David Haas.
(0.6.1)
++ New features:
* The packet buffer implementation has been enhanced to support
zero-copy and copy-on-demand for packet buffers which have their
payloads in application-managed memory.
Implemented by David Haas.
Use PBUF_REF to make a pbuf refer to RAM. lwIP will use zero-copy
if an outgoing packet can be directly sent on the link, or perform
a copy-on-demand when necessary.
The application can safely assume the packet is sent, and the RAM
is available to the application directly after calling udp_send()
or similar function.
++ Bugfixes:
* ARP_QUEUEING should now correctly work for all cases, including
PBUF_REF.
Implemented by Leon Woestenberg.
++ Changes:
* IP_ADDR_ANY is no longer a NULL pointer. Instead, it is a pointer
to a '0.0.0.0' IP address.
* The packet buffer implementation is changed. The pbuf->ref counter
meaning has changed, and several pbuf functions have been
adapted accordingly.
* netif drivers have to be changed to set the hardware address length field
that must be initialized correctly by the driver (hint: 6 for Ethernet MAC).
See the contrib/ports/c16x cs8900 driver as a driver example.
* netif's have a dhcp field that must be initialized to NULL by the driver.
See the contrib/ports/c16x cs8900 driver as a driver example.
(0.5.x) This file has been unmaintained up to 0.6.1. All changes are
logged in CVS but have not been explained here.
(0.5.3) Changes since version 0.5.2
++ Bugfixes:
* memp_malloc(MEMP_API_MSG) could fail with multiple application
threads because it wasn't protected by semaphores.
++ Other changes:
* struct ip_addr now packed.
* The name of the time variable in arp.c has been changed to ctime
to avoid conflicts with the time() function.
(0.5.2) Changes since version 0.5.1
++ New features:
* A new TCP function, tcp_tmr(), now handles both TCP timers.
++ Bugfixes:
* A bug in tcp_parseopt() could cause the stack to hang because of a
malformed TCP option.
* The address of new connections in the accept() function in the BSD
socket library was not handled correctly.
* pbuf_dechain() did not update the ->tot_len field of the tail.
* Aborted TCP connections were not handled correctly in all
situations.
++ Other changes:
* All protocol header structs are now packed.
* The ->len field in the tcp_seg structure now counts the actual
amount of data, and does not add one for SYN and FIN segments.
(0.5.1) Changes since version 0.5.0
++ New features:
* Possible to run as a user process under Linux.
* Preliminary support for cross platform packed structs.
* ARP timer now implemented.
++ Bugfixes:
* TCP output queue length was badly initialized when opening
connections.
* TCP delayed ACKs were not sent correctly.
* Explicit initialization of BSS segment variables.
* read() in BSD socket library could drop data.
* Problems with memory alignment.
* Situations when all TCP buffers were used could lead to
starvation.
* TCP MSS option wasn't parsed correctly.
* Problems with UDP checksum calculation.
* IP multicast address tests had endianess problems.
* ARP requests had wrong destination hardware address.
++ Other changes:
* struct eth_addr changed from u16_t[3] array to u8_t[6].
* A ->linkoutput() member was added to struct netif.
* TCP and UDP ->dest_* struct members where changed to ->remote_*.
* ntoh* macros are now null definitions for big endian CPUs.
(0.5.0) Changes since version 0.4.2
++ New features:
* Redesigned operating system emulation layer to make porting easier.
* Better control over TCP output buffers.
* Documenation added.
++ Bugfixes:
* Locking issues in buffer management.
* Bugfixes in the sequential API.
* IP forwarding could cause memory leakage. This has been fixed.
++ Other changes:
* Directory structure somewhat changed; the core/ tree has been
collapsed.
(0.4.2) Changes since version 0.4.1
++ New features:
* Experimental ARP implementation added.
* Skeleton Ethernet driver added.
* Experimental BSD socket API library added.
++ Bugfixes:
* In very intense situations, memory leakage could occur. This has
been fixed.
++ Other changes:
* Variables named "data" and "code" have been renamed in order to
avoid name conflicts in certain compilers.
* Variable++ have in appliciable cases been translated to ++variable
since some compilers generate better code in the latter case.
(0.4.1) Changes since version 0.4
++ New features:
* TCP: Connection attempts time out earlier than data
transmissions. Nagle algorithm implemented. Push flag set on the
last segment in a burst.
* UDP: experimental support for UDP-Lite extensions.
++ Bugfixes:
* TCP: out of order segments were in some cases handled incorrectly,
and this has now been fixed. Delayed acknowledgements was broken
in 0.4, has now been fixed. Binding to an address that is in use
now results in an error. Reset connections sometimes hung an
application; this has been fixed.
* Checksum calculation sometimes failed for chained pbufs with odd
lengths. This has been fixed.
* API: a lot of bug fixes in the API. The UDP API has been improved
and tested. Error reporting and handling has been
improved. Logical flaws and race conditions for incoming TCP
connections has been found and removed.
* Memory manager: alignment issues. Reallocating memory sometimes
failed, this has been fixed.
* Generic library: bcopy was flawed and has been fixed.
++ Other changes:
* API: all datatypes has been changed from generic ones such as
ints, to specified ones such as u16_t. Functions that return
errors now have the correct type (err_t).
* General: A lot of code cleaned up and debugging code removed. Many
portability issues have been fixed.
* The license was changed; the advertising clause was removed.
* C64 port added.
* Thanks: Huge thanks go to Dagan Galarneau, Horst Garnetzke, Petri
Kosunen, Mikael Caleres, and Frits Wilmink for reporting and
fixing bugs!
(0.4) Changes since version 0.3.1
* Memory management has been radically changed; instead of
allocating memory from a shared heap, memory for objects that are
rapidly allocated and deallocated is now kept in pools. Allocation
and deallocation from those memory pools is very fast. The shared
heap is still present but is used less frequently.
* The memory, memory pool, and packet buffer subsystems now support
4-, 2-, or 1-byte alignment.
* "Out of memory" situations are handled in a more robust way.
* Stack usage has been reduced.
* Easier configuration of lwIP parameters such as memory usage,
TTLs, statistics gathering, etc. All configuration parameters are
now kept in a single header file "lwipopts.h".
* The directory structure has been changed slightly so that all
architecture specific files are kept under the src/arch
hierarchy.
* Error propagation has been improved, both in the protocol modules
and in the API.
* The code for the RTXC architecture has been implemented, tested
and put to use.
* Bugs have been found and corrected in the TCP, UDP, IP, API, and
the Internet checksum modules.
* Bugs related to porting between a 32-bit and a 16-bit architecture
have been found and corrected.
* The license has been changed slightly to conform more with the
original BSD license, including the advertisement clause.
(0.3.1) Changes since version 0.3
* Fix of a fatal bug in the buffer management. Pbufs with allocated
RAM never returned the RAM when the pbuf was deallocated.
* TCP congestion control, window updates and retransmissions did not
work correctly. This has now been fixed.
* Bugfixes in the API.
(0.3) Changes since version 0.2
* New and improved directory structure. All include files are now
kept in a dedicated include/ directory.
* The API now has proper error handling. A new function,
netconn_err(), now returns an error code for the connection in
case of errors.
* Improvements in the memory management subsystem. The system now
keeps a pointer to the lowest free memory block. A new function,
mem_malloc2() tries to allocate memory once, and if it fails tries
to free some memory and retry the allocation.
* Much testing has been done with limited memory
configurations. lwIP now does a better job when overloaded.
* Some bugfixes and improvements to the buffer (pbuf) subsystem.
* Many bugfixes in the TCP code:
- Fixed a bug in tcp_close().
- The TCP receive window was incorrectly closed when out of
sequence segments was received. This has been fixed.
- Connections are now timed-out of the FIN-WAIT-2 state.
- The initial congestion window could in some cases be too
large. This has been fixed.
- The retransmission queue could in some cases be screwed up. This
has been fixed.
- TCP RST flag now handled correctly.
- Out of sequence data was in some cases never delivered to the
application. This has been fixed.
- Retransmitted segments now contain the correct acknowledgment
number and advertised window.
- TCP retransmission timeout backoffs are not correctly computed
(ala BSD). After a number of retransmissions, TCP now gives up
the connection.
* TCP connections now are kept on three lists, one for active
connections, one for listening connections, and one for
connections that are in TIME-WAIT. This greatly speeds up the fast
timeout processing for sending delayed ACKs.
* TCP now provides proper feedback to the application when a
connection has been successfully set up.
* More comments have been added to the code. The code has also been
somewhat cleaned up.
(0.2) Initial public release.