Re: [PATCH] 2.2.17pre7 VM enhancement Re: I/O performance on 2.4.0-test2

[Stephen C. Tweedie]

Hi,

On Wed, Jun 28, 2000 at 07:56:01PM +0200, Andrea Arcangeli wrote:
> 
> While you are oom, or while you're running mmap002, right? If so that's
> intentional and I'm very aware of the MAP_SHARED problem but what we did
> so far to fix that MAP_SHARED issue is band-aid (that could hurt the other
> cases that works right).

Right, and there is one trivially obvious bogosity in the VM which I
suspect is at the root of this particular behaviour.

try_to_free_pages() returns once it makes any progress.  However,
"progress" does not mean freeing a page.  Simply finding a page which
can be unmapped from a process's page tables is considered "progress".  
Unmapping those pages doesn't actually free any memory; we still need
to (a) wait for any writeback IO to complete; and (b) reenter
shrink_mmap afterwards to remove the page from the page cache before
this "progress" results in genuine free pages.

So, shared mappings inevitably result in try_to_free_pages() returning
without freeing any physical memory.  You can only do that so often
before processes start getting killed due to OOM.

--Stephen