lkml.org 
[lkml]   [2012]   [Nov]   [20]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [RFT PATCH v1 1/5] mm: introduce new field "managed_pages" to struct zone
On Tue, 20 Nov 2012 22:56:11 +0800
Jiang Liu <liuj97@gmail.com> wrote:

> On 11/20/2012 07:38 AM, Andrew Morton wrote:
> > On Mon, 19 Nov 2012 00:07:26 +0800
> > Jiang Liu <liuj97@gmail.com> wrote:
>
> ...
>
> > Also, the existing comment tells us that spanned_pages and
> > present_pages are protected by span_seqlock but has not been updated to
> > describe the locking (if any) for managed_pages.
> How about this?
>
> ...
>

Looks nice.

> >> + for (z = pgdat->node_zones; z < pgdat->node_zones + MAX_NR_ZONES; z++)
> >> + if (!is_highmem(z))
> >
> > Needs a comment explaining why we skip the highmem zone, please.
> How about this?
>
> ...
>

Ditto.

> >> @@ -106,6 +106,7 @@ static void get_page_bootmem(unsigned long info, struct page *page,
> >> void __ref put_page_bootmem(struct page *page)
> >> {
> >> unsigned long type;
> >> + static DEFINE_MUTEX(ppb_lock);
> >>
> >> type = (unsigned long) page->lru.next;
> >> BUG_ON(type < MEMORY_HOTPLUG_MIN_BOOTMEM_TYPE ||
> >> @@ -115,7 +116,9 @@ void __ref put_page_bootmem(struct page *page)
> >> ClearPagePrivate(page);
> >> set_page_private(page, 0);
> >> INIT_LIST_HEAD(&page->lru);
> >> + mutex_lock(&ppb_lock);
> >> __free_pages_bootmem(page, 0);
> >> + mutex_unlock(&ppb_lock);
> >
> > The mutex is odd. Nothing in the changelog, no code comment.
> > __free_pages_bootmem() is called from a lot of places but only this one
> > has locking. I'm madly guessing that the lock is here to handle two or
> > more concurrent memory hotpluggings, but I shouldn't need to guess!!
> Actually I'm a little hesitate whether we should add a lock here.
>
> All callers of __free_pages_bootmem() other than put_page_bootmem() should
> only be used at startup time. And currently the only caller of put_page_bootmem()
> has already been protected by pgdat_resize_lock(pgdat, &flags). So there's
> no real need for lock, just defensive.
>
> I'm not sure which is the best solution here.
> 1) add a comments into __free_pages_bootmem() to state that the caller should
> serialize themselves.
> 2) Use a dedicated lock to serialize updates to zone->managed_pages, this need
> modifications to page_alloc.c and memory_hotplug.c.
> 3) The above solution to serialize in put_page_bootmem().
> What's your suggestions here?

Firstly, let's be clear about what *data* we're protecting here. I
think it's only ->managed_pages?

I agree that no locking is needed during the init-time code.

So afaict we only need be concerned about concurrent updates to
->managed_pages via memory hotplug, and lock_memory_hotplug() is
sufficient there. We don't need to be concerned about readers of
managed_pages because it is an unsigned long (a u64 on 32-bit machines
would be a problem).

All correct? If so, the code is OK as-is and this can all be
described/formalised in code comments. If one wants to be really
confident, we could do something along the lines of

void mod_zone_managed_pages(struct zone *zone, signed long delta)
{
WARN_ON(system_state != SYSTEM_BOOTING &&
!is_locked_memory_hotplug());
zone->managed_pages += delta;
}

And yes, is_locked_memory_hotplug() is a dopey name.
[un]lock_memory_hotplug() should have been called
memory_hotplug_[un]lock()!




\
 
 \ /
  Last update: 2012-11-20 21:21    [W:0.075 / U:0.444 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site