TODO

o (bug from yaz) On the kernel with your page-fault patch applied,
valgrind (2.2.0) can reliably cause the kernel to crash in
page_fault_wake_function.  Try, say, "valgrind --tool=addrcheck
/bin/ls" right after boot.

o the io_submit_try() thing could be done cleaner


CHANGELOG

10/3/2004:
  * interface change: replace io_modify system call with 
    IO_CMD_SET_EPOLLFD command to io_submit()

10/2/2004:
  * fixes from Yasushi. When adding fds with EPOLLOUT, the kernel
    sometimes segfaults.
  * fixes from Yasushi. Fixes cleanups in case of error in
    io_deliver_epoll_event()

9/21/2004:
  * experimental code for lockless event reaping (using cmpxchg),
    epoll_over_aio_lockless.patch, based on Kenneth Chen's patch
  * libaio code to reap events from the mapped event ring, also
    using cmpxchg

9/20/2004:
  * moved to libaio 0.3.99

9/19/2004:
  * cleaned up pipetest.c
  * use obj==-1 to indicate a canceled event in the event ring
  * changed sys_io_getevents() to deal with canceled events
  * added epio_test.c, epoll-over-aio test cases
  * changed the pagefault notification patch to also use io_modify().
    the test program does not work yet.

9/17/2004:
  * fixes from Yasushi
  * fixed the epoll event update race.  does proper locking when updating
    the event
  * removed "flex" from code, replaced with "aio" or "io"
  
9/10/2004:
  * fixed the pagefault bug.  we're missing a reference to the page
    when we finish loading the page.
  * found out that kgdb is incompatible with gcc 2.95.4 on vivace.
    we'll stick with 3.3.4
  * (Yosushi Saito) missing unlock in epoll
  * (Yosushi Saito) flex_deliver_epoll_event sometimes calls
  aio_ring_event w/o matching put_aio_ring_event

9/9/2004:
  * tried to fix the pagefault handling problem. the problem is when
    the process terminates, sys_exit() eventually calls free_hot_code_page()
    to free memory pages, when page->mapping should already be NULL.
    Now it's not, so the function complains.  the reason for this is
    still unknown.  the main memory-freeing function is zap_pte_range()

9/8/2004:
  * ported epoll code over to 2.6.8.1. pipetest is working

  * debugged io_submit_try().  for some reason, io_submit_one() is now
calling __aio_run_iocbs(), instead of aio_run_iocb().  but we rely on
io_submit_one() passing the user-level event object to aio_run_iocb()
to be filled.  For now I changed io_submit_one() to call
aio_run_iocb() again.  But I don't know whether this breaks anything.

  * trysubmit.c test is working


9/7/2004:
  * start using quilt+cvs to manage kernel patches. this seems to be 
    much more easier than using bitkeeper when it comes to tracking
    main-line kernel tree

  * port the latest patch (flex-031210.patch) to Linux 2.6.8.1
    (patches/flex.patch).

  * rewrote page_fault_wake_function because the pagetable
    maitainence API has changed. (no pte_chain_alloc() any more...)

  * the epoll code has changed much, back out our epoll changes for
    now

  * not tested yet

	
flex-031207.patch
	fixed page fault stale data bugs.
	page fault handling is fully working now

	status:
		pagefault - working
		event queue - working but possible races