PERFORCE change 188035 for review

Fri Jan 21 16:53:54 UTC 2011

http://p4web.freebsd.org/@@188035?ac=10

Change 188035 by trasz at trasz_victim on 2011/01/21 16:53:15

	Update TODO, expand manual page.

Affected files ...

.. //depot/projects/soc2009/trasz_limits/TODO#40 edit
.. //depot/projects/soc2009/trasz_limits/usr.bin/rctl/rctl.8#3 edit

Differences ...

==== //depot/projects/soc2009/trasz_limits/TODO#40 (text+ko) ====

@@ -1,44 +1,29 @@
-Limits done:
+TODO:
 
- - CPU time (RUSAGE_CPU), in microseconds
- - data size (RUSAGE_DATA), in megabytes
- - maximum file size (RUSAGE_FSIZE), in megabytes
- - number of processes (RUSAGE_NPROC)
- - virtual memory usage (address space limit) (RUSAGE_VMEM), in megabytes
- - maximum core size (RUSAGE_CORE), in megabytes (core, as in ELF program state dump)
- - locked memory usage (RUSAGE_MEMLOCK), in megabytes
- - resident set size (physical memory usage) (RUSAGE_RSS), in megabytes
- - stack size (RUSAGE_STACK), in megabytes,
- - number of file descriptors (RUSAGE_NOFILE)
- - swap usage (RUSAGE_SWAP), in megabytes
- - amount of memory consumed by socket buffers (RUSAGE_SBSIZE), in megabytes
- - number of kernel-visible threads (RUSAGE_NTHR)
- - wallclock time (RUSAGE_WALLCLOCK)
- - %cpu time (RUSAGE_PCTCPU)
- - number of SysV shared memory segments (RUSAGE_NSHM)
- - SysV shared memory size, in megabytes (RUSAGE_SHMSIZE)
- - number of SysV semaphores modified in a single semop(2) call (RUSAGE_NSEMOP)
- - number of SysV semaphores (RUSAGE_NSEM)
- - number of SysV queues (RUSAGE_NMSGQ)
- - number of queued SysV messages (RUSAGE_MSGQQUEUED)
- - SysV message queue size, in megabytes (RUSAGE_MSGQSIZE)
+ - Fix %CPU limits for shortly living processes.
 
-Milestone 3:
+ - Consider replacing proc pointer with thread pointer in rusage_add(9) et al.
+   In most cases caller uses 'td->td_proc' anyway, and passing thread would
+   allow the HRL code to send a signal to the offending thread instead of the
+   offending process.  It would also mean that these the functions would
+   no longer need to be called with proc lock held, because we could use
+   td->td_ucred, which is stable across the syscall.
 
- - Fix %CPU limits for shortly living processes.
-
  - Get rid of container_lock.  Atomic instructions would be nice, but we really
    need 64 bits (per-process counters could be 32 bit, I guess, but the higher
    level containers could overflow), and atomic(9) doesn't support 64 bit values
    on 32 bit platforms.
 
- - Consider replacing proc pointer with thread pointer in rusage_add(9) et al.
-   In most cases caller uses 'td->td_proc' anyway, and passing thread would
-   allow the HRL code to send a signal to the offending thread instead of the
-   offending process.
+ - Don't account resources for system processes.  It's not as easy as checking
+   for P_SYSTEM flag, as the flag will be set for AIO kernel processes.
+
+   Also, try to figure out what's going on with 'intr' p_flag - checking for P_SYSTEM
+   didn't really work for that process.
 
 Issues:
 
+ - Setting RSS limit too low can make the system thrash to death.
+
  - We enforce limits when a process allocates a resource, and when it forks.
    We don't enforce limits when process changes its credentials, though.  This
    might be either a bug or feature, depending on point of view.
@@ -66,24 +51,6 @@
    container mechanism.  This is a part of HRL, and it's outside the scope
    of the containers project.  This means the code is full of "#ifdef CONTAINERS".
 
- - Don't account resources for system processes.  It's not as easy as checking
-   for P_SYSTEM flag, as the flag will be set for AIO kernel processes.
-
-   Also, try to figure out what's going on with 'intr' p_flag - checking for P_SYSTEM
-   didn't really work for that process.
-
- - Right now, the whole containers stuff is under a single mutex.  This is internal
-   to containers, i.e. the API consumers don't need to care, thus it's easy to change.
-
-   I'd need to run benchmarks first, but two strategies come to mind:
-
-   1. Replace container_lock with rmlock, protecting the container hierarchy.  The lock
-      would be acquired for write in operations changing the hierarchy, and for read
-      for all other operations.  When locked for read, the counters would be accessed
-      using atomic instructions.
-
-   2. Replace single container_lock with individual per-container mutexes.
-
  - RUSAGE_NOFILE accounts for size of file descriptor table, rather than the number
    of file descriptors.  This shouldn't be a problem, but might be worth keeping
    in mind.

==== //depot/projects/soc2009/trasz_limits/usr.bin/rctl/rctl.8#3 (text+ko) ====

@@ -116,6 +116,34 @@
 or, in short, ":".  A filter that matches all the login classes would be
 "loginclass:".  A filter that matches all defined limits for maxprocesses
 resource would be "::maxprocesses".
+.Pp
+.Sh RESOURCES
+.Bl -column -offset 3n "msgqqueued"
+.It cpu		CPU time, in microseconds
+.It fsize	maximum file size, in megabytes
+.It data	data size, in megabytes
+.It stack	stack size, in megabytes
+.It core	core dump size, in megabytes
+.It rss		resident set size, in megabytes
+.It memlock	locked memory, in megabytes
+.It nproc	number of processes
+.It nofile	file descriptor table size
+.It sbsize	memory consumed by socket buffers, in megabytes
+.It vmem	address space limit, in megabytes
+.It npts	number of PTYs
+.It swap	swap usage, in megabytes
+.It nthr	number of threads
+.It msgqqueued	number of queued SysV messages
+.It msgqsize	SysV message queue size, in megabytes
+.It nmsgq	number of SysV message queues
+.It nsem	number of SysV semaphores
+.It nsemop	number of SysV semaphores modified in a single semop(2) call
+.It nshm	number of SysV shared memory segments
+.It shmsize	SysV shared memory size, in megabytes
+.It wallclock	wallclock time
+.It pctcpu	%cpu time
+.El
+.Pp
 .Sh EXIT STATUS
 .Ex -std
 .Sh SEE ALSO
