----------------------------------------------------------------------
This file is intended to help track internal changes to the library.
Latest changes should be added to the top of the file.
----------------------------------------------------------------------

2004/03/29 - zf
  * read_proc_file() is using close() to close the proc file
    Because it is called from the scheduler, it should use the
    SYS_close syscall directly.

2003/08/11 - zf
  * Fastpath improvement for buffer-cache hitting disk requests
  
2003/08/09 - zf
  * New eepoll patch for 2.6.0-test2-bk8

2003/08/07 - zf
  * Reformatted micro-benchmark results
  * Save errno in thread stack to avoid conflict with NPTL
  * Reduced I/O polling rate
  * Added group submit for nio. Default batch factor is 32.
    Tunable by IOQ_DEPTH_MAX in nio.c

2003/08/01 - zf
  * Added pipe scalability benchmark at src/bench/pipetest_thread.

2003/07/30 - jrvb
  * Added util/config.[ch], to clean up runtime options
  * Made several compile-time options to runtime options 

2003/07/28 - zf
  * Added new concurrent I/O benchmark p_bench_concurrent_io.c

2003/07/25 - jrvb
  * Changed init code in a number of places to use gcc's
    __attribute__((constructor)).  This is much more reliable than
    trying to do things by hand.  NOTE: in a few cases, things are
    still done by hand, eg. where order matters.
  * added diskio_immediate.c, to do immediate, blocking disk IO.  

2003/07/24 - zf
  * Pipe benchmark in bench/p_bench_pipe.c

2003/07/23 - jrvb
  * added some crude support for dynamic linking of libpthread.so to aio/Makefile  

2003/07/22 - jrvb
  * tweak Makefile in bench to kill orphaned capriccio kernel threads
  * fixed program name in bench/test_readdir.c

2003/07/21 - zf
  * automate benchmarking process of bench/ 
  * cosmetic changes to SYS_SET_SYSCALL (gcc 3.3+ complaining about 
    assigning const var to non-const)
  * nptl benchmarking
  * nptl build notes in doc/

2003/07/19 - zf
  * numerous signal fixes
  * signal test cases in bench/
  * apache now runs with no patching at all!
    ("-D DEBUG" command line argument still needed)

2003/07/15 - zf
  * (untested) full signal implementation in signal.c
  * fixed bug that stops pthread_exit() from exiting the main thread
  * now we pass most of test cases in bench/.  several we cannot do
    are either because we are cooperative or we do not have support some
    optional functionalities.

2003/07/11 - zf
  * __errno_location() now returns a thread-local error number instead
    of a global one
    
  * (untested) new select() wrapper in nio.c (not in default build)
  
  * Apache now works with nio (with Rob's configure command line and
    a slightly modified apache patch(in cvs))

2003/07/09 - zf
  * Extended epoll kernel patch in kernel/patches/eepoll-2.5.70.diff
    It's applicable to kernel 2.5.70 (or up hopefully)

2003/07/08 - zf
  * checked in i/o wrapping code using the extended epoll kernel interface
    Code is in nio/ and not used by default.  Add 'USE_IO_LIB=nio' to
    Make.opts to use it instead of the aio/ code.
    ( This is still preliminary and overlaps with the aio/ code.  
      It should be merged with aio/ code later.)

  * added CAP_NODEBUG env. var to disable debug messages at runtime.

  * added tests and benchmarks in bench, mainly taken from MIT pthread

2003/06/18 - jrvb
  * checked in initial (not tested) support for send/recv and friends
    to blocking_io.c

2003/04/24 - jrvb
  * switched util/occ_list.c to use spinlocks, instead of true
    optomistic concurrency control.
  * fixed kthread disk IO
  * made diskio_kthread the default
  * added more reporting to knot

2003/04/19 - jrvb
  * added line counts to makefiles

2003/03/24 - jrvb
  * have sockio_poll check IO first - faster w/ knot 

2003/03/23 - zf
  * fixes to assert()'s with side-effects.  the side-effects will not be
    executed when asserts are disable, causing potential bugs

2003/03/23 - jrvb
  * bugfixes all over the place
  * change timers, counters, etc to show diff b/w process time and
    real time when perfctr is available
  * insert yields in various places before IO calls, for better
    blocking graph and to avoid long edges
  * add tracking of which syscall we are in
  * ugly hacks to avoid init races w/ perfctr and IO libs
  * signal changes for better core dumps
  * bugfixes to diskio_kthread
  * many fixes to edge length tracking & stats gathering
  * fixes & improvements to scheduling, including basic admission control

2003/03/21 - jrvb
  * cleanup of blocking graph code
    - prepare for final integration w/ cil blocking graph
    - reduce overhead through sampling
  * fixes to cpu timing code & resource counting

2003/03/21 - zf
  * perfctr-based timing in timing.c, set USE_PERFCTR = 1 in Make.opts to
    use.  Then kill -sigusr1 <pid> will print out timing info

2003/03/19 - jrvb
  * optomistic concurrency control list structures are working w/
    diskio_kthread and linux clone()
  * make IO regime runtime selectable via environment vars:
       CAPRICCIO_SOCKIO and CAPRICCIO_DISKIO
  * change current_usec() to current_usecs() - for some reason the
    java programmer in me always wants the 's' from currentTimeMillis()
  * modify polling behavior in scheduler
    - if there are no active threads, allow sockio to block
    - max block time == min sleep time for sleeping threads
  * change src/aio/Makefile behavior - looks now for make variables
    HAVE_AIO and HAVE_SYS_EPOLL.  These should be defined in src/Make.opts
  * make debug selectable from w/in Make.opts.  See Make.opts.sample
    for info.

2003/03/18 - zf
  * timing routines using rdtsc in util/timing.h
    these can be used to measure our overhead of polling/scheduling

2003/03/15 - zf
  * updated apache patch - more signal ignoring
  * fix poll() race bug - we may return wrong result and corrupt the poll list
  * use backtrace_symbols() anyway because it gives us some function names
    NOTE: -all-static is preventing backtrace_symbols from working, while -static
    is ok.

2003/03/15
  * dump blocking graph w/ SIGUSR1
  * added dummy signal functions
    - use alt sig stack, so we don't hose ourselves (need to do this
      for each process on an MP system)
    - set default signal handlers
  * fixed thread_usleep() to take long long to avoid overflow
  * Makefile tweaks & reorg to better support compiling external progs
    - added include & lib directories
    - added dummy header files, to avoid getting libc-specific stuff
      in the way
  * makefile changes to support cil  

2003/03/13
  * added lifo scheduler
  * added simple http client
  * laughed about changelog timeskew....  
  * misc. fixes & tweaks to knot
  * add code to turn off nagle alg. in tcp


2003/03/13 - zf
  * mutex/cond variable timeout fixes
  * fix one year time skew of change log file  ;-)

2003/03/11 - jrvb
  * fixed sec -> usec conversions (should be times 10^6, not 10^3)
  * added nanosleep() hack to sockio_poll() (do similar for epoll?)
  * simpler version of priority scheduling


2003/03/10 - zf
  * misc sleep queue fixes


2003/03/10 - jrvb
  * thread changes
    - change some names: current -> current_thread, scheduler -> scheduler_thread
  * aio changes
    - make sockio_poll() try IO first before suspending
    - make all sockets non-blocking on creation (in accept, connect & open)
  * misc
    - changed order of includes, so we don't get multiple pthreads libs.  


2003/03/09 - jrvb
  * aio changes
    - removed ref to num_runnable_threads from blocking_io.c - should
      be managed only to threadlib.c
    - removed POLLOUT, POLLERR, POLLNVAL from requested events, as
      they are ignored there, and always added to returned events
      anyway.
    - changed sockio_poll.c to do IO in polling thread, to better
      handle lists of requests.  This is mainly useful for accept(),
      so we get returns from as many accept calls as possible.  (still
      need corresponding change in sockio_epoll.c)
    - per-fdsruct locking of request lists
  * thread changes
    - moved savederrno out of thread_t, and into stack frame of thread_yield_internal()
    - removed THREAD_ATTR_DEFAULT, which wasn't really used
    - unified new_thread(), thread_spawn(), and thread_spawn_with_attr()
    - fixed bug in timeout usage in mutex.c
    - streamlined thread_yield_internal
    - fixed stack dump to not call our IO
    - moved stack dump IO out of signal handler
    - changed stack dump signal to SIGUSR1 instead of SIGQUIT
  * util changes
    - added our own assert(), to avoid infinite loop b/w write() and assert()
  * general
    - streamline structures w/ bit fields


2003/03/09 - zf
  * complete the thread stack dump facility.  now it shows the state of each
    thread and their current state
  * add new API for coro to support thread stack dump, co_backtrace()
  * for the thread stack dump thing, add a linked list of all threads (global 
    variable threads) in threadlib.c

  * add thread id (tid) for each thread to enhance readability of thread dump
  * add generic "sleep queue" data structure to the scheduler
  * sleep(), usleep(), pthread_cond_timedwait() implemented on sleep queue

  * thread_suspend_self() now takes a timeout argument and returns FALSE if
    timeout actually expires before anyone else resumes the thread.  this will
    be used to implement all timeout related functionality

  * add poll()/select() wrappers

2003/03/08 - zf
  * added rudimentary thread stack dump functionality for debugging:
      kill -quit <pid>
    will generate a thread stack dump to stderr
  * dummy readv()/writev() mapping
  * bypass our mapping when user wants O_NONBLOCK
  * mutex fixes

2003/03/07 - jrvb
  * misc fixes to scheduling stuff
  * more tests in apps/schedtest

2003/03/06 - zf
  * set correct return value for lseek()
  * rewrite dup(), dup2() handling.  all dup'ed fds should have a single file
    state (offset etc).  so now they are connected together by a linked list
    and keep a single copy of state in one of them (root_fd).  
    p.s.  I know, this is ugly.  But the "cleaner" way to keep a separate data 
    structure for per-file state outside fdstruct_t requires too much change to
    the data structure and how we handle fds.  So I reject it.

2003/03/04 - jrvb
  * misc bugfixes to new scheduling code 
  * added graph-aware scheduler code
  * added env-var option & signal handler to dump blocking graph on exit
  * added basic test of batching scheduler

2003/03/03 - jrvb
  * clean up & abstract blocking graph functions, in preparation for
    integration w/ CIL
  * misc cleanup of thread lib, flags, etc.

2003/03/02 - jrvb
  * abstract scheduling functions to allow multiple schedulers

2003/03/02 - zf
  * src/aio/Makefile now outputs libcapriccio.a, which contains all needed 
    static libraries in one archive.
  * priliminary apache2 patch: src/apps/apache-2.0.44-capriccio.diff.
    it compiles and runs, but everything but serving the default page is not 
    working now.

2003/02/26 - jrvb
  * threadlib.c - don't deallocate main_thread in thread_exit(), since
    it may be used again by exit functions (ie, for flushing files in STDIO)
  * blocking_io.c - added set_nonblocking() function, and cleaned up accept()
  * added stats about yield points in threadlib.c and node_stats.c
  * removed last vestiges of threadlib.c interface from knot, in favor
    of pthreads interface
  * added Makefile targets for different thread libs to knot

2003/02/24 - zf
  * add a PRINT_STAT switch to node_stat.c, determining whether thread
    yield stats are printed
  * shared library makefile cleanup - now exports all symbols
  * pthreads tests in aio/tests

2003/02/20 - jrvb
  * fixed sockio_epoll to work w/ list of outstanding requests
  * added socketcall() args to iorequest_t

2003/02/19 - jrvb
  * added stats for all yield points, to examine graph nodes

2003/02/18 - jrvb
  * added latch macros to threadlib (for use in threadlib and aio)
    - can't be held over blocking points
    - not recursive
    - w/ one CPU, an error if they are already locked
  * used latches in mutex code, to fix multi-CPU races
  * switched aio routines to latches, rather than mutexes
  * cleaned up aiotest.c

2003/02/18 - zf
  * added thread_join()
  * aiotest working again with async disk i/o

2003/02/15 - zf
  * basic pthread layer in place (needs debugging).  this results in the
    following changes to threadlib API.  missing pthread API is tracked in 
    TODO file
  * added pthreadtest.c, 
      make testpthread
  * added pth-like thread attribute support
    - new API thread_attr_of(), ... to manage attributes
    - new API thread_spawn_with_attr() to create a thread with attribute
  * added mutex implementation
  * added rwlock
  * added condition variable
  * added key-based thread-specific storage

2003/02/13 - jrvb
  * traced aio/threads problems to stdio setbuf()
  * updated debugging statements many places
  * misc fixes to knot / aio

2003/02/12 - jrvb
  * change makefiles back to static linking
  * update knot to new interfaces
  * incorporate some runtime stats code

2003/02/11 - jrvb
  * added epoll support
  * misc fixes & minor reorg of sockio_poll

2003/02/08 - jrvb
  * bugfixes in aio: make everything use syscall() that should 
  * added aio/check_syscall to check for direct system call useage 
  * added many new syscall wrappers to aio, including pread/pwrite

2003/02/07 - jrvb
  * made blocking_io.c:debug() printf-like, and moved to util/
  * removed init_fdstruct(), since init is being done in get_fdstruct()
  * fixed bugs in lseek() and get_fdstruct()
  * moved thread_suspend_self() inside IO modules, to allow stack
    allocation of module-specific request info
  * changed diskio_aio.c to stack allocate diskio_cb structs
  * appended "_t" to diskio_aio.c:diskio_cb struct name
  * partial fix for fds not allocated locally 
    - breaks for system(). 
    - ultimately best to track all places that fds are created