svn commit: r309017 - in head/sys: cddl/compat/opensolaris/sys compat/linprocfs fs/tmpfs sys vm

[Alexey Dokuchaev]

On Tue, Nov 22, 2016 at 06:13:46PM +0000, Alan Cox wrote:
> New Revision: 309017
> URL: https://svnweb.freebsd.org/changeset/base/309017
> 
> Log:
>   Remove PG_CACHED-related fields from struct vmmeter, because they are no
>   longer used.  More precisely, they are always zero because the code that
>   decremented and incremented them no longer exists.
>   
>   Bump __FreeBSD_version to mark this change.
>   
>   Reviewed by:	kib, markj
>   Sponsored by:	Dell EMC Isilon
>   Differential Revision:	https://reviews.freebsd.org/D8583
> 
> Modified: head/sys/cddl/compat/opensolaris/sys/kmem.h
> ...
> -#define	freemem				(vm_cnt.v_free_count + vm_cnt.v_cache_count)
> +#define	freemem				vm_cnt.v_free_count

This looks correct now.

> Modified: head/sys/compat/linprocfs/linprocfs.c
> ...
> @@ -176,7 +176,7 @@ linprocfs_domeminfo(PFS_FILL_ARGS)
>  	 * like unstaticizing it just for linprocfs's sake.
>  	 */
>  	buffers = 0;
> -	cached = vm_cnt.v_cache_count * PAGE_SIZE;
> +	cached = vm_cnt.v_inactive_count * PAGE_SIZE;

Some applications that make calculations based on the readings from the
/usr/compat/linux/proc/meminfo are broken after this change, e.g. those
that do things like:

    return (totalMem - (freeMem + buffers + cache)) / totalMem;

because now free memory includes cached one, and above formula gives
bogus negative result.

In `sys/compat/linprocfs/linprocfs.c', memfree is calculated as follows:

    memtotal = physmem * PAGE_SIZE;
    /*
     * The correct thing here would be:
     *
    memfree = vm_cnt.v_free_count * PAGE_SIZE;
    memused = memtotal - memfree;
     *
     * but it might mislead linux binaries into thinking there
     * is very little memory left, so we cheat and tell them that
     * all memory that isn't wired down is free.
     */
    memused = vm_cnt.v_wire_count * PAGE_SIZE;
    memfree = memtotal - memused;

So, when vm.stats.vm.v_cache_count was yielding some (typically small)
result, using the aforementioned trick made sense (because even when
vm_cnt.v_free_count added to vm.stats.vm.v_cache_count it would still
be rather small).  This logic no longer applies after this change.

On my 12-CURRENT i386 system, /usr/compat/linux/proc/meminfo reports:

    % cat /usr/compat/linux/proc/meminfo | head -4
    MemTotal:   3132504 kB
    MemFree:    2925936 kB			<<< that's a lot
    Buffers:          0 kB
    Cached:     2212196 kB			<<< that's also a lot

On an older 4.11-STABLE box (also i386), it reports:

    MemTotal:    521828 kB
    MemFree:     383976 kB
    MemShared:     7480 kB
    Buffers:          0 kB
    Cached:       22732 kB (can be added to MemFree and still < MemTotal)

So either the "cheat" have to go away and vm_cnt.v_free_count used as it
should, or cached memory should not be read from vm_cnt.v_inactive_count
but estimated somehow else.  Which approach would be more correct?

./danfe