[FreeVMS] pager question

[BERTRAND Joël]

Guido a écrit :
> Hi all,

	Hello,

> Looking at the pager algorithm I noticed that a page fault in user space could
> lead to allocation of a new page for the process that generated the page
> fault.

	Right.

> Maybe I'm missing something, but it seems to me that that is not
> desired behavior. Buggy or malicious software in user space should not be able
> to acquire resources simply by accessing invalid memory addresses. In stead
> this should lead to an access violation and termination of the originating
> process.

	I agree. For kernel itself, sys_pagefault always returns new 
ressources. For an unprivilegied process, pager has to save a list of 
allocated pages for this process. Pagefault handler only has to return 
new pages when process is loaded in memory, when stack is growing or 
when process uses something like mmap() or brk(). In all other cases, 
pagefault has to raise a new syscall to kill process.

> I'd suggest that page fault handling and mapping/granting memory are entirely
> separate functions. A page fault on a page that is not mapped is an access
> violation and should lead to termination of the originating process. If the
> page is mapped, the pager could (like OpenVMS does) distinguish between soft
> and hard page faults. A soft page fault means that the page still resides
> somewhere in cache, while a hard page fault means it is swapped out and has to
> be loaded from disk. This depends on the paging algorithm.

	I agree.

> Memory allocation and mapping/granting pages is something completely different.
> My suggestion would be that allocation requests by user tasks are handled by
> the task pager. If the task pager runs out of memory it can request one or
> more pages from the main pager through a page request system call. The main
> pager can map/grant a page from the free page pool it maintains or, if the
> pool is empty, it can request a new page from the kernel. If the kernel runs
> out of free pages, the pager should free up some pages by swapping out some
> pages to disk. Note that this could (should?) be optimised by initially
> mapping/granting all free pages to the main pager, so that the pager does not
> have to use system calls to request memory from the kernel once the system is
> up and running.
>
> I could be wrong, of course ...

	No, I don't think. In a first time, I propose that task pager doesn't 
use swapped memory as we need disk I/O drivers to handle swap and as 
these drivers require task pager.

	Kickstart loads L4/X2, sigma0 and several modules :
- vmskernel.sys (uses sigma0 pager) ;
- pager.sys (main pager that use sys_pagefault) that is launched 
_without_ swapped memory area as disk I/O are not available ;
- dq driver + partionning to add swapping capabilities.

	Regards,

	JKB