Re: [RFC][PATCH] proc: export more page flags in /proc/kpageflags

[Wu Fengguang <fengguang.wu@intel.com>]

On Thu, Apr 16, 2009 at 11:54:43AM +0800, Andi Kleen wrote:
> On Thu, Apr 16, 2009 at 10:41:33AM +0800, Wu Fengguang wrote:
> > On Wed, Apr 15, 2009 at 09:57:49PM +0800, Andi Kleen wrote:
> > > > That's pretty good separations. I guess it would be convenient to make the
> > > > extra kernel flags available under CONFIG_DEBUG_KERNEL?
> > > 
> > > Yes.
> > > 
> > > BTW an alternative would be just someone implementing a suitable
> > > command/macro in crash(1) and tell the kernel hackers to run that on
> > > /proc/kcore. That would have the advantage to not require code.
> > 
> > Hmm, that would be horrible to code/maintain. i
> 
> Actually the bits are enums and crash is able to read C type 
> information.

Great! That dismissed my main concern with crash.

> > One major purpose of
> > /proc/kpageflags is to export the unstable kernel page flag bits as
> > stable ones to user space. 
> 
> That's the first case ("administrator"), but not the second one
> ("kernel hacker")
> 
> BTW not saying that crash is the best solution for this, but
> it's certainly an serious alternative for the kernel hacker
> case. 

OK.

> > Note that the exact internal flag bits can
> > not only change slowly with kernel versions, but more likely with
> > different kconfig combinations.
> 
> Really? The numbers should be the same, at least for a given
> architecture with 32bit/64bit.

For example, the presence of CONFIG_PAGEFLAGS_EXTENDED will shift
all the following flag bits by 1.

        #ifdef CONFIG_PAGEFLAGS_EXTENDED
                PG_head,                /* A head page */
                PG_tail,                /* A tail page */
        #else  
                PG_compound,            /* A compound page */
        #endif 
                PG_swapcache,           /* Swap page: swp_entry_t in private */
                PG_mappedtodisk,        /* Has blocks allocated on-disk */
                PG_reclaim,             /* To be reclaimed asap */
                PG_buddy,               /* Page is free, on buddy lists */
                PG_swapbacked,          /* Page is backed by RAM/swap */
        #ifdef CONFIG_UNEVICTABLE_LRU
                PG_unevictable,         /* Page is "unevictable"  */
        #endif
        #ifdef CONFIG_HAVE_MLOCKED_PAGE_BIT
                PG_mlocked,             /* Page is vma mlocked */
        #endif
        #ifdef CONFIG_IA64_UNCACHED_ALLOCATOR
                PG_uncached,            /* Page has been mapped as uncached */
        #endif  
                __NR_PAGEFLAGS,


> > Followed are their detailed locations. Did we found a bug? ;-)
> 
> I think all pages > 0 in a larger page are tails.  But I don't
> claim to understand all the finer details of compound pages.

Right. Tail pages will outnumber head pages. But I found that the
tail page _ranges_ greatly outnumber head pages. There should be
exactly one tail page range for one head page. 

Thanks,
Fengguang