From: Johannes Helander To: lites Subject: Snapshot, Changes, TODO list. --text follows this line-- There is a new snapshot in leia.cs.hut.fi:foggy/lites/LITES.941116.tar.gz It doesn't differ from the previous intermediate snapshot that much but has a few fixes. Here are some changes from a period of about two weeks. - mapped gettimeofday is now enabled on all i386 configurations by default. The kernel crash is avoided by not inheriting the kernel page but instead remapping after fork. Somebody ought to fix the kernel in the future but the problem is no longer acute. Linux gettimeofday is now 500 times faster (in the worst vs. best case). The char devices /dev/time (25,0) and /dev/timezone (26,0) are needed (use mknod). - The server now uses Mach error numbers everywhere. The emulator translates the errors to a suitable errno depending on the running binary. sys/errno.h hides the differences. For source compatibilty reasons the server uses the traditional defines. The numeric values are just different. mach_error_string may be used to get a string corresponding to the internal error code. NOTE: mach_error_string has a bug that makes it return NULL sometimes. The explanation is simply that the library code assumes it is doing unsigned coparisons. Unfortunately kern_return_t (and transitively mach_error_t) is signed. Either libmach or kern_return.h needs to be fixed. - There is a preliminary mips port. The server works, binaries are loaded, system calls are performed. No signals, no Ultrix support yet. I might continue on this if I have time (I doubt it). Two CMU files in the mips port are not included in the list of files I have signed from CMU. They must be either replaced or cleared in a future additional list. - The server and emulator now share a new exec code. The exec is divided in more stages: - find file and fetch header - look at header and guess what binary it is - parse header according to guess and turn it into internal representation. Currently AOUT, ECOFF, and SOM are supported. The internal form is an array of sections and a load info struct. - walk through the sections and call a mapping method for each.. - Setup the execution environment according to the load info. BSS dirty page clearing used to be a pain from the server. The server now sets two registers in the user thread (it calls thread_set_state anyways) and ecrt0.c does the zeroing. In the emulator exec it is simpler: just bzero the area. ecrt0.c is a machine dependent part of the emulator. The register reading is easy to do with gcc asm variables. - The signal code is patched to work a little more. It is still a mess and needs a total redesign. - Job control works with Linux programs. Signal numbers are translated in more places such as wait4. - Lite system calls are on by default. The e_lite_sysent.c could be made MI soon. - There is a new alternative build environment that uses gmake, autoconf, and a couple of scripts. ODE still works but will probably deteriorate over time. conf/* has new configuration files. server/conf/* is ODE. For ODE to survive a bit longer it should be chnaged to use the configuration scripts in conf/* instead of the config program and related scripts. The gmake environment is most notably missing shadowing which shouldn't be impossible to add. Cross compilation works with both environments. The emulator building should be more similar to the server building. The emulator make files automatically picks up all files it can find. I don't think this is a particularly good idea. A simple list of files in the makefile makes it possible to be more selective. - SysV shared memory for making Doom faster. The totally emulator implementation makes files in /tmp/shm/ and pages to them. /etc/rc needs to create the directory and empty it on boot. The implementation is still untested as my X server binary did not support shared memory extensions. An X server with this feature is needed. - STD+WS now works without +DEBUG also on i386 This was due to mapin_user not initializing the address given to vm_map. vm_map's anywhere option is TRUE but it seems the kernel only scans from the given address forward and gives up when it hits the end. I think this behavior is ok even if not really expected. The simple fix is to zero the address hint always. TODO ---- - A library for code common to the server and emulator and perhaps later other programs. At least exec_file.c, bcopy etc., ntoh, and so on could be moved there. Requires changes to the build system. - Figure out how to avoid deadlocking risks arising from printf calling functions taking locks. Printf is called in a large variety of places where all kinds of locks may be held. Printing to ttys, selwakeup, etc. in this situation creates a big potential for deadlock. I added a comment next to printf noting this problem. - Fix and test build environment so that building on Linux and BSD really works. - Enhance the sysv support to make it real. - 64 bit cleanup. - Add support for Linux and or BSD X servers. This is mostly related to virtual consoles I think. - Write /sbin/cored -- the core dumping daemon. - Experiment with +ether_as_syscall and get it to work enough to make it default. This is about running protocol stacks on the network input thread instead of going through netisrs. It is important for simplifying splx and the whole spl mechanism. Also network latency ought to decrease. The first select in inetd didn't wake up last time I tried. After this I've fixed some bug related to select but there are still some uncertainty related to sockets so this might be a difficult thing to debug. - Fix /proc and other Lite stuff that is in conf/MASTER.h listed in the UNTESTED, ALMOST, and NOTYET categories. More later. Johannes -------------- - Add syncing of memory object cleaning to all fsyncs. - Wait for memory object cleaning to complete when should wait is set in fsync.