[Devel] [PATCH RHEL7 COMMIT] ub: account shmem in cached in meminfo

[Konstantin Khorenko]

The commit is pushed to "branch-rh7-3.10.0-123.1.2-ovz" and will appear at https://src.openvz.org/scm/ovz/vzkernel.git
after rh7-3.10.0-123.1.2.vz7.5.9
------>
commit 1bd1fb1435323a78d7f5c647c62bdf7de08fa79b
Author: Vladimir Davydov <vdavydov at parallels.com>
Date:   Fri Jun 5 13:45:34 2015 +0400

    ub: account shmem in cached in meminfo
    
    This patch ports:
    
      diff-ub-account-shmem-to-meminfo-cached
      diff-ub-meminfo-cached-make-sure-meminfo-memFree-plus-Cached-plus-Slab-less-than-MemTotal
    
    Author: Vladimir Davydov
    Email: vdavydov at parallels.com
    Subject: ub: account Shmem to meminfo:Cached
    Date: Wed, 3 Dec 2014 16:30:22 +0300
    
    Upstream, Shmem is accounted as file pages so it must contribute to
    meminfo:Cached, otherwise output of certain versions of free(1) might
    get screwed (negative 'cached').
    
    https://jira.sw.ru/browse/PSBM-28682
    
    Signed-off-by: Vladimir Davydov <vdavydov at parallels.com>
    
    =============================================================================
    
    Author: Vladimir Davydov
    Email: vdavydov at parallels.com
    Subject: ub/cached: make sure meminfo:MemFree+Cached+Slab<=MemTotal
    Date: Tue, 30 Dec 2014 18:23:45 +0300
    
    Since commit diff-ub-account-shmem-to-meminfo-cached, container's
    meminfo:Cached includes meminfo:Shmem to conform to the behavior of
    vanilla kernels. However, the value of meminfo:Shmem reported inside
    containers is not quite the same as it would be in a non-virtualized
    environment: it includes swapped out memory, while on vanilla kernels it
    only shows the resident shared memory size. As a result,
    meminfo:MemFree+Cached can be greater than meminfo:MemTotal, in which
    case it's trimmed to MemTotal, so that meminfo:MemFree+Cached+Slab is
    likely to be greater than MemTotal. This, in turn, confuses our
    userspace tools (PACI autoscaling).
    
    The right fix would be to account only resident shared memory under
    meminfo:Shmem, but since it's tricky to implement, because shmem may be
    shadowed, let's simply trim meminfo:MemFree+Cached to MemTotal-Slab.
    
    https://jira.sw.ru/browse/PSBM-31006
    
    Signed-off-by: Vladimir Davydov <vdavydov at parallels.com>
    
    =============================================================================
    
    Related to https://jira.sw.ru/browse/PSBM-33637
    
    Signed-off-by: Vladimir Davydov <vdavydov at parallels.com>
---
 kernel/bc/vm_pages.c | 9 ++++++---
 1 file changed, 6 insertions(+), 3 deletions(-)

diff --git a/kernel/bc/vm_pages.c b/kernel/bc/vm_pages.c
index a6f311e..c52d34f 100644
--- a/kernel/bc/vm_pages.c
+++ b/kernel/bc/vm_pages.c
@@ -188,9 +188,6 @@ static int bc_fill_meminfo(struct user_beancounter *ub,
 	ub_sync_memcg(ub);
 	ub_page_stat(ub, &node_online_map, mi->pages);
 
-	mi->cached = min(mi->si->totalram - mi->si->freeram,
-			mi->pages[LRU_INACTIVE_FILE] +
-			mi->pages[LRU_ACTIVE_FILE]);
 	mi->locked = ub->ub_parms[UB_LOCKEDPAGES].held;
 	mi->shmem = ub->ub_parms[UB_SHMPAGES].held;
 	dcache = ub->ub_parms[UB_DCACHESIZE].held;
@@ -211,6 +208,12 @@ static int bc_fill_meminfo(struct user_beancounter *ub,
 	mi->slab_unreclaimable =
 		DIV_ROUND_UP(max(0L, (long)ub->ub_parms[UB_KMEMSIZE].held -
 							dcache), PAGE_SIZE);
+
+	mi->cached = min(mi->si->totalram - mi->si->freeram -
+			mi->slab_reclaimable - mi->slab_unreclaimable,
+			mi->pages[LRU_INACTIVE_FILE] +
+			mi->pages[LRU_ACTIVE_FILE] +
+			ub->ub_parms[UB_SHMPAGES].held);
 out:
 	return ret;
 }