procps/TODO

177 lines
6.4 KiB
Plaintext
Raw Normal View History

2002-02-02 04:17:29 +05:30
-------------------------- general ------------------------
Implement /usr/proc/bin tools like Solaris has.
The main Makefile is broken. Try 'touch top.h' to see.
The Makefile is too clever, so it's still not fixed.
Don't these really belong in the procps package?
killall pstree fuser lsof who
(they are maintained elsewhere, which causes version problems)
Cache results of dev_to_tty.
---------------------- kernel -------------------------
Add an "adopted child" flag to mark processes that are not
natural children of init. This can make --forest work better.
Add a thread group ID, to be shared by all tasks that are related by
the clone() system call. This ID might be made unique from boot to
shutdown, perhaps being a 16-bit CPU number and 48-bit per-CPU
serial number.
Make the kernel group /proc listing output by thread group.
Without this, a thread-aware ps must always sort processes.
Supply the task ID (the "PID"/"TID") of the thread group leader.
I define "leader" as the first process of a thread group.
Don't reuse the task ID of a thread group leader until all threads
are dead. Better yet, don't let the leader exit.
Supply better data for top's CPU state display. Currently top has
to subtract old numbers from new numbers and divide that result by
the number of processors. The kernel won't even supply the number
of processors in a portable way.
Mark threads, and supply a list of other threads.
Supply data for the ADDR and JOBC fields.
Support & supply data for SL and RE.
Add a /proc/*/tty symlink to eliminate guessing when /proc/*/fd is
not accessable.
Put unique ID at the top of System.map and in /proc, to make sure
there is never a mismatch.
Maybe just put the symbols in the kernel like some other systems do.
Raul Miller writes: You might want to implement this as user data which just
happens to come from a wierd source -- instead of just freeing the memory
(like happens with init code), you want it to wind up in the buffer cache.
Add /proc/*/.bindata files to avoid string parsing. It should be an array
of 64-bit values on all machines. New entries go on the end and obsolete
ones get filled in with something logical -- entries must never be deleted!
Add all the stuff Solaris has. This would also replace ptrace.
The kernel should report these task_struct members:
has_cpu
processor
---------------------- w --------------------------
The LOGIN@ column sometimes has a space in it. This makes correct
scripting difficult.
Verify that DNS control does not give a user the power to specify
arbitrary data for the FROM column. (could set root's VGA color map!)
---------------------- vmstat --------------------------
Extract /proc/stat parsing from vmstat into libproc somewhere.
--------------------- libproc ----------------------
Stop storing fields with duplicate info (often different
units: kB and pages, seconds and jiffies) in the proc_t struct.
---------------------- top -------------------------
Share more stuff with ps.
---------------- ps for now, maybe move to libproc ------------------
With forest output and a tty named /dev/this_is_my_tty, the position
of the command name gets messed up. (we print too many spaces) (fixed?)
Fix missing stuff for these formats: FB_j FB_l FB_v HP_f HP_*_ HP_fl JFMT OL_m
(jobc,cpu,sl,re,cpu,prmgrp,cp,m_size,m_swap,m_share,vm_lib,m_dt)
Note that "cpu" has two meanings.
Add Beowulf support. This is ugly, since the current patches use a
daemon to collect info and add a HOST field after the PID field.
Query optimizer, put cheap/required process selection first.
Avoid reading both /proc/*/status and /proc/*/stat. Actually, avoid
reading lots of stuff! Avoidance got disabled. (at least don't read the
environment!!!)
Maybe ps should put a 'C' in front of fields when they are affected
by cumulative mode. Debian ps did that. (move flag to common.h...)
Support printing the client hostname (the FROM that w(1) uses) in place
of a pty. Maybe do this when PS_PERSONALITY=linux.
Disgusting color support, for completeness. (Debian ps had it)
Disambiguate narrow tty info. (/dev/tty7 == /dev/pts/7 now)
1------8 1--4
ttyS2 S2
ttyI31 I31
pts/7 7 Short form could be /999.
pts/9999 9999 Short form could just be trunctuated to /999.
tty7 7 Short form could be vc-7.
tty63 63 Short form could be vc63.
Internationalization, as specified by XPG3, Volume 1, Commands and Utilities.
(and suggested by Unix98) LC_TIME affects date format.
----------------------- ps -----------------------
Solve the "ps reports wrong date" bug. (recent processes on an SMP system
are reported as being 5 days old, while boot processes look normal)
(reported for procps-2.0.2 in May 1999)
Show real-time priorities. (type & number)
Add an option to select all processes that a user can kill.
(related to RUID, EUID, tty, etc. -- but maybe ignore root power)
Add a nice display option for killing things.
ruser,euser,ppid,pid,pmem,stime,args isn't too bad
Make the column alignment algorithm support this:
FOO BAR
8 44444
453 45
45 2989
63666 0
34 333
(useful for the UNIX tty and time values, since the time might look
like 100-10:40:32 for old processes and the tty might have extra room)
Improve long sort/format specifiers documentation and fill in the missing
code as much as the kernel can support. Make sure that memory amounts are in
pages when they should be and in kB when they should be, not backwards.
output encoding: UTF8 --nul --null
Make BSD formats use non-standard BSD time format, at least when it
doesn't violate the "no whitespace" rule.
Try to make -jl fit in 80 columns. Do we need more than 1000 pty devices,
9 flag bits, etc.? (hmmm, Linux supports 2048 pty devices now, and we
might also want to steal whitespace there when the time column overflows)
When not in strict Unix98 mode, let foo=8 specify that foo is 8
characters wide. Debian did that. Then foo=8=bar and foo=bar=8
could change both header and width.
Better unmangling of '?' as a tty. The shell destroys '?' when there
is a filename that matches. If the argument seems like garbage,
check for a file that might have screwed up the '?'.
If the 'O' option is given something already implied by 'O',
assume the user wanted a sorting option.
Conflict:
Digital THREAD is user,pcpu,pri,scnt,wchan,usertime,systime
AIX THREAD is uname,pid,ppid,tid,S,C,PRI,scount,WCHAN,F,tty,bnd,comm
AIX looks like this:
USER PID PPID TID S C PRI SC WCHAN FLAG TTY BND CMD