2012/1/4 Hugh Dickins [off-list ref]:

On Tue, 3 Jan 2012, KOSAKI Motohiro wrote:

quoted

(1/3/12 8:51 PM), Hugh Dickins wrote:

quoted

In testing my fix for that, I find that there has been no attempt to
keep the Unevictable count accurate on SysVShm: SHM_LOCK pages get
marked unevictable lazily later as memory pressure discovers them -
which perhaps mirrors the way in which SHM_LOCK makes no attempt to
instantiate pages, unlike mlock.

Ugh, you are right. I'm recovering my remember gradually. Lee implemented
immediate lru off logic at first and I killed it
to close a race. I completely forgot. So, yes, now SHM_LOCK has no attempt to
instantiate pages. I'm ashamed.

Why ashamed?  The shmctl man-page documents "The caller must fault in any
pages that are required to be present after locking is enabled."  That's
just how it behaves.

hehe, I have big bad reputation about for bad remember capabilities from
my friends. I should have remembered what i implemented. ;-)

quoted

quoted

(But in writing this, realize I still don't quite understand why
the Unevictable count takes a second or two to get back to 0 after
SHM_UNLOCK: perhaps I've more to discover.)

Interesting. I'm looking at this too.

In case you got distracted before you found it, mm/vmstat.c's

static DEFINE_PER_CPU(struct delayed_work, vmstat_work);
int sysctl_stat_interval __read_mostly = HZ;

static void vmstat_update(struct work_struct *w)
{
       refresh_cpu_vm_stats(smp_processor_id());
       schedule_delayed_work(&__get_cpu_var(vmstat_work),
               round_jiffies_relative(sysctl_stat_interval));
}

would be why, I think.  And that implies to me that your
lru_add_drain_all_async() is not necessary, you'd get just as good
an effect, more cheaply, by doing a local lru_add_drain() before the
refresh in vmstat_update().

When, I implement lru_add_drain_all_async(), I thought this idea. I don't
dislike both. But if we take vmstat_update() one, I think we need more tricks.
pcp draining in refresh_cpu_vm_stats() delays up to 3 seconds. Why?
round_jiffies_relative() don't silly round to HZ boundary. Instead of, it adds
a few unique offset per each cpus. thus, 3 seconds mean max 3000cpus
don't make zone_{lru_}lock contention. pagevec draining also need same
trick for rescue SGI UV. It might be too pessimistic concern. but
vmstat_update() shouldn't make obsevable lock contention.

But it would still require your changes to ____pagevec_lru_add_fn(),
if those turn out to help more than they hurt.

I agree.

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>