[PATCH 16/28] [FLAT 1/6] Changes in data structures for flat model

[Eric W. Biederman]

Pavel Emelianov <xemul at openvz.org> writes:

> sukadev at us.ibm.com wrote:
>> Pavel Emelianov [xemul at openvz.org] wrote:
>> | This patch opens the flat model patches.
>> | 
>> | The flat model idea is that struct pid has two numbers. The first one 
>> | (pid->nr) is a global one and is unique in the system. The second one 
>> | (pid->vnr) is a virtual pid. It is used on the kernel user boundary only.
>> 
>> This approach duplicates 5 integers and 2 pointers per process for every
>> process in the system. While this may not be expensive for processes that
>> actually use multiple namespaces, doesn't it waste memory if majority of
>> processes exist only in one namespace ?
>
> task_struct alignment allows for it. so does the alignment of signal structure.
> and please note that this comes with appropriate ifdefs around. the only problem
> is with struct pid, but we're virtualizing it after all!
>
> moreover - two integers and a pointer to the namespace is the minimal set of
> fields for pid that is visible from two namespaces...

>> | --- ./include/linux/sched.h.flatdatast 2007-06-15 15:14:33.000000000 +0400
>> | +++ ./include/linux/sched.h	2007-06-15 15:19:14.000000000 +0400
>> | @@ -482,7 +482,10 @@ struct signal_struct {
>> |  		pid_t session __deprecated;
>> |  		pid_t __session;
>> |  	};
>> | -
>> | +#ifdef CONFIG_PID_NS_FLAT
>> | +	pid_t vpgrp;
>> | +	pid_t vsession;
>> | +#endif
>> |  	/* boolean value for session group leader */
>> |  	int leader;
>> | 
>> | @@ -944,6 +947,11 @@ struct task_struct {
>> |  	unsigned did_exec:1;
>> |  	pid_t pid;
>> |  	pid_t tgid;
>> | +#ifdef CONFIG_PID_NS_FLAT
>> | +	/* hash the virtual ids as well */
>> | +	pid_t vpid;
>> | +	pid_t vtgid;
>> | +#endif

Adding vpgrp, vsession, vpid, and vtgid is wrong.

A case can probably be made for caching the common case (users view),
but we already have fields for that.

For a global view we must use struct pid *, otherwise we are just asking
for trouble.

Eric