ARM: print cma-reserved pages from show_mem

Commit Message

Gregory Fong April 13, 2015, 1:21 a.m. UTC

On Sun, Apr 12, 2015 at 06:09:13PM -0700, Gregory Fong wrote:
> On Fri, Apr 10, 2015 at 12:24:31PM +0100, Russell King - ARM Linux wrote:
> > On Fri, Apr 10, 2015 at 01:18:04PM +0800, Wang, Yalin wrote:
> > > > [   12.212102] active_anon:734 inactive_anon:1189 isolated_anon:0
> > > > [   12.212102]  active_file:0 inactive_file:0 isolated_file:0
> > > > [   12.212102]  unevictable:0 dirty:0 writeback:0 unstable:0
> > > > [   12.212102]  free:254104 slab_reclaimable:82 slab_unreclaimable:843
> > 
> > Here, we have 82 pages reclaimable, which is 328kB, and 843 unreclaimable
> > which is 3372kB, which is a total of 925 pages.
> > 
> > > > [   12.212102]  mapped:429 shmem:1815 pagetables:13 bounce:0
> > > > [   12.212102]  free_cma:4032
> > > > [   12.243172] DMA free:754080kB min:3472kB low:4340kB high:5208kB
> > > > active_anon:180kB inactive_anon:0kB active_file:0kB inactive_file:0kB
> > > > unevictable:0kB isolated(anon):0kB isolated(file):0kB present:778240kB
> > > > managed:759252kB mlocked:0kB dirty:0kB writeback:0kB mapped:0kB
> > > > shmem:0kB slab_reclaimable:328kB slab_unreclaimable:3372kB
> > 
> > Which agrees here.
> > 
> > > > [   12.401609] 834 slab pages
> > 
> > but not here... this is an interesting difference, because in the ARM
> > version of show_mem(), we count the actual number of pages where
> > PageSlab() returns true.  Can slab pages also be reserved pages or
> > swap cache pages?  I thought they were exclusive of those.  So, the
> > question comes... why the difference in accounting, and which one is
> > correct.
> > 
> > Maybe there's a bug in the accounting somewhere...
> 
> Yes, the ARM show_mem wasn't updated after the various allocators
> (SLUB, SLAB, SLOB) were updated to use compound pages.  Fixing it, you
> get
> 
> [    7.081826] sysrq: SysRq : Show Memory
> [    7.085610] Mem-info:
> [    7.087890] DMA per-cpu:
> [    7.090431] CPU    0: hi:  186, btch:  31 usd:  98
> [    7.095230] HighMem per-cpu:
> [    7.098116] CPU    0: hi:   90, btch:  15 usd:  29
> [    7.102923] active_anon:724 inactive_anon:1189 isolated_anon:0
> [    7.102923]  active_file:0 inactive_file:0 isolated_file:0
> [    7.102923]  unevictable:0 dirty:0 writeback:0 unstable:0
> [    7.102923]  free:253980 slab_reclaimable:83 slab_unreclaimable:846
> [    7.102923]  mapped:429 shmem:1815 pagetables:15 bounce:0
> [    7.102923]  free_cma:4032
> [    7.133995] DMA free:753344kB min:3472kB low:4340kB high:5208kB active_anon:292kB inactive_anon:0kB active_file:0kB inactive_file:0kB unevictable:0kB isolated(anon):0kB isolated(file):0kB present:778240kB managed:759188kB mlocked:0kB dirty:0kB writeback:0kB mapped:0kB shmem:0kB slab_reclaimable:332kB slab_unreclaimable:3384kB kernel_stack:256kB pagetables:36kB unstable:0kB bounce:0kB free_cma:0kB writeback_tmp:0kB pages_scanned:0 all_unreclaimable? yes
> [    7.174080] lowmem_reserve[]: 0 0 264 264
> [    7.178175] HighMem free:262576kB min:264kB low:572kB high:884kB active_anon:2604kB inactive_anon:4756kB active_file:0kB inactive_file:0kB unevictable:0kB isolated(anon):0kB isolated(file):0kB present:270336kB managed:270336kB mlocked:0kB dirty:0kB writeback:0kB mapped:1716kB shmem:7260kB slab_reclaimable:0kB slab_unreclaimable:0kB kernel_stack:0kB pagetables:24kB unstable:0kB bounce:0kB free_cma:16128kB writeback_tmp:0kB pages_scanned:0 all_unreclaimable? yes
> [    7.218955] lowmem_reserve[]: 0 0 0 0
> [    7.222687] DMA: 8*4kB (UEM) 8*8kB (UM) 4*16kB (UEM) 5*32kB (UM) 2*64kB (M) 4*128kB (UEM) 3*256kB (M) 2*512kB (EM) 5*1024kB (UEM) 6*2048kB (UEM) 3*4096kB (EM) 88*8192kB (MR) = 753344kB
> [    7.239477] HighMem: 0*4kB 0*8kB 1*16kB (M) 1*32kB (M) 2*64kB (UM) 2*128kB (UM) 2*256kB (UC) 3*512kB (UMC) 2*1024kB (UC) 2*2048kB (UC) 2*4096kB (UC) 30*8192kB (MRC) = 262576kB
> [    7.255455] 1815 total pagecache pages
> [    7.259211] 0 pages in swap cache
> [    7.262533] Swap cache stats: add 0, delete 0, find 0/0
> [    7.267764] Free swap  = 0kB
> [    7.270647] Total swap = 0kB
> [    7.282890] 262144 pages of RAM
> [    7.286041] 254274 free pages
> [    7.289013] 4763 reserved pages
> [    7.292157] 929 slab pages
> [    7.294868] 1063 pages shared
> [    7.297839] 0 pages swap cached
> 
> And now we see 83 slab_reclaimable + 846 slab_unreclaimable adds up
> correctly to the total of 929.
> 
> The patch below will end up with the correct count.
> 

Sorry, messed up the patch formatting.  Here it is fixed:

8<===

Comments

Russell King - ARM Linux April 13, 2015, 9:56 a.m. UTC | #1

On Sun, Apr 12, 2015 at 06:21:15PM -0700, Gregory Fong wrote:
> On Sun, Apr 12, 2015 at 06:09:13PM -0700, Gregory Fong wrote:
> > And now we see 83 slab_reclaimable + 846 slab_unreclaimable adds up
> > correctly to the total of 929.
> > 
> > The patch below will end up with the correct count.
> > 
> 
> Sorry, messed up the patch formatting.  Here it is fixed:

So now the question is: do we fix this, or do we use the generic version?
Given that the total number of slab pages can be easily deduced from the
generic statistics, do we need to modify the generic version to print an
additional line with this?  It seems wasteful to do so, and just adds
more noise to the kernel's debug output.

Mel Gorman April 13, 2015, 10:04 a.m. UTC | #2

On Mon, Apr 13, 2015 at 10:56:45AM +0100, Russell King - ARM Linux wrote:
> On Sun, Apr 12, 2015 at 06:21:15PM -0700, Gregory Fong wrote:
> > On Sun, Apr 12, 2015 at 06:09:13PM -0700, Gregory Fong wrote:
> > > And now we see 83 slab_reclaimable + 846 slab_unreclaimable adds up
> > > correctly to the total of 929.
> > > 
> > > The patch below will end up with the correct count.
> > > 
> > 
> > Sorry, messed up the patch formatting.  Here it is fixed:
> 
> So now the question is: do we fix this, or do we use the generic version?
> Given that the total number of slab pages can be easily deduced from the
> generic statistics, do we need to modify the generic version to print an
> additional line with this?

Whatever ARM decides, I do not think the generic version needs to do
a PFN walk to recaluate the SLAB statistics. The slab_reclaimable and
slab_unreclaimable stats based on the vmstat counters is sufficient.

Russell King - ARM Linux April 13, 2015, 10:05 a.m. UTC | #3

On Mon, Apr 13, 2015 at 11:04:26AM +0100, Mel Gorman wrote:
> On Mon, Apr 13, 2015 at 10:56:45AM +0100, Russell King - ARM Linux wrote:
> > On Sun, Apr 12, 2015 at 06:21:15PM -0700, Gregory Fong wrote:
> > > On Sun, Apr 12, 2015 at 06:09:13PM -0700, Gregory Fong wrote:
> > > > And now we see 83 slab_reclaimable + 846 slab_unreclaimable adds up
> > > > correctly to the total of 929.
> > > > 
> > > > The patch below will end up with the correct count.
> > > > 
> > > 
> > > Sorry, messed up the patch formatting.  Here it is fixed:
> > 
> > So now the question is: do we fix this, or do we use the generic version?
> > Given that the total number of slab pages can be easily deduced from the
> > generic statistics, do we need to modify the generic version to print an
> > additional line with this?
> 
> Whatever ARM decides, I do not think the generic version needs to do
> a PFN walk to recaluate the SLAB statistics. The slab_reclaimable and
> slab_unreclaimable stats based on the vmstat counters is sufficient. 

Yes, I agree.  My feeling is we just switch to the generic version and be
done with it.

Gregory Fong April 13, 2015, 12:29 p.m. UTC | #4

On Mon, Apr 13, 2015 at 3:05 AM, Russell King - ARM Linux
<[e-mail]> wrote:
> On Mon, Apr 13, 2015 at 11:04:26AM +0100, Mel Gorman wrote:
>> On Mon, Apr 13, 2015 at 10:56:45AM +0100, Russell King - ARM Linux wrote:
>> > On Sun, Apr 12, 2015 at 06:21:15PM -0700, Gregory Fong wrote:
>> > > On Sun, Apr 12, 2015 at 06:09:13PM -0700, Gregory Fong wrote:
>> > > > And now we see 83 slab_reclaimable + 846 slab_unreclaimable adds up
>> > > > correctly to the total of 929.
>> > > >
>> > > > The patch below will end up with the correct count.
>> > > >
>> > >
>> > > Sorry, messed up the patch formatting.  Here it is fixed:
>> >
>> > So now the question is: do we fix this, or do we use the generic version?
>> > Given that the total number of slab pages can be easily deduced from the
>> > generic statistics, do we need to modify the generic version to print an
>> > additional line with this?
>>
>> Whatever ARM decides, I do not think the generic version needs to do
>> a PFN walk to recaluate the SLAB statistics. The slab_reclaimable and
>> slab_unreclaimable stats based on the vmstat counters is sufficient.
>
> Yes, I agree.  My feeling is we just switch to the generic version and be
> done with it.

Agreed.  This is really what I was hoping for in the first place, but
didn't know before submitting the initial patch whether there was some
arcane reason for the arm-specific show_mem.

If someone like Yalin really wants the total slab pages, they can just
add slab_unreclaimable and slab_reclaimable (btw, Yalin, all emails
I've sent to you are bouncing, maybe you'll see this since it's going
to the list).

Thanks,
Gregory

Patch

diff --git a/arch/arm/mm/init.c b/arch/arm/mm/init.c
index 1609b02..8d606bb 100644
--- a/arch/arm/mm/init.c
+++ b/arch/arm/mm/init.c
@@ -117,7 +117,7 @@  void show_mem(unsigned int filter)
 			else if (PageSwapCache(page))
 				cached++;
 			else if (PageSlab(page))
-				slab++;
+				slab += 1 << compound_order(page);
 			else if (!page_count(page))
 				free++;
 			else