www.nico.schottelius.org/software/cinit/browse_source/cinit-0.3pre7/doc/ancient/thoughts.closed

53 lines
2.1 KiB
Text
Raw Normal View History

--------------------------------------------------------------------------------
Closed thoughts,
Nico Schottelius, 2005-05-XX (Last Modified: 2005-06-14)
--------------------------------------------------------------------------------
1. using SIDs (service IDs) to communicated with external processes
This was a very bad idea: The external program could exploit us by
specifying an arbitary big SID (as the SID is simply the index
of our service array).
2. using function pointers to handle messages
Seems like it works fine. We have handlers for each
message (do_*), which are called by client and server.
The function pointer is simply to read or write, dependending
if it is the client or the server. This way we don't need
to rewrite communications parts.
3. Using different storage
First all services were saved in a service array of the size
MAX_SVC. This has been replaced by a double-linked list.
Have a look at serv/list.c.
4. Using sockets for IPC (between cinit forks)
Works very fine, though we have to mount a temporarily fs before.
5. Removing the maximum number of direct dependencies
Is not planned, as it looks like the current number (32, see
conf/max_deps) seems to be much more than needed. And if
one really needs more, simply increase conf/max_deps.
6. Using a directory params with 1,2,3,4 for argv
This would make substituting a single argument easier, but only
if you know which one you have to change. This would add
an additional dirent(), which would not replace the current read(), but
add more open() and close()s. As this does not seem to make live
easier for system administrators, this is not implemented.
If you really like it, hack client/exec_svc.c
7. Using TCP/IP sockets
This would be a very small change in the code, but would allow
to control cinit over network. Since there is no authentication,
this would be highly insecure. On the other hand, cinit
could control the parallel start of many hosts, if they
should become 'one' computer at the end. As this is not needed
currently, it's not implemented.