Re: [PATCH RFC 1/1] mm, page_alloc: fix incorrect zone_statistics data

[hejianet]

On 20/12/2016 8:31 PM, Mel Gorman wrote:
> On Mon, Dec 12, 2016 at 01:59:07PM +0800, Jia He wrote:
> > In commit b9f00e147f27 ("mm, page_alloc: reduce branches in
> > zone_statistics"), it reconstructed codes to reduce the branch miss rate.
> > Compared with the original logic, it assumed if !(flag & __GFP_OTHER_NODE)
> >   z->node would not be equal to preferred_zone->node. That seems to be
> > incorrect.
> >
> > Fixes: commit b9f00e147f27 ("mm, page_alloc: reduce branches in
> > zone_statistics")
> >
> > Signed-off-by: Jia He <hejianet@xxxxxxxxx>
> This is slightly curious. It appear it would only occur if a process was
> running on a node that was outside the memory policy. Can you confirm
> that is the case?
Yes, here is what I caught:

dumpstack() is triggered when z->node(5) == preferred_zone->node(5) and z->node != numa_node_id(4)
without flag GET_OTHER_NODE

It is not a rare case. I saw hundreds of such cases when kernel boots up
[c000000cdcaef440] [c0000000002e88cc] cache_grow_begin+0xcc/0x500
[c000000cdcaef6f0] [c0000000002ecb44] do_tune_cpucache+0x64/0x100
[c000000cdcaef750] [c0000000002ecc7c] enable_cpucache+0x9c/0x180
[c000000cdcaef7d0] [c0000000002ed01c] __kmem_cache_create+0x1ec/00x2c0
[c000000cdcaef820] [c000000000291c98] create_cache+0xb8/0x240
[c000000cdcaef890] [c000000000291fa8] kmem_cache_create+0x188/0x2290
[c000000cdcaef950] [d000000011dc5c70] ext4_mb_init+0x3c0/0x5e0 [eext4]
[c000000cdcaef9f0] [d000000011daaedc] ext4_fill_super+0x266c/0x33390 [ext4]
[c000000cdcaefb30] [c000000000328b8c] mount_bdev+0x22c/0x260
[c000000cdcaefbd0] [d000000011da1fa8] ext4_mount+0x48/0x60 [ext4]
[c000000cdcaefc10] [c00000000032a11c] mount_fs+0x8c/0x230
[c000000cdcaefcb0] [c000000000351f98] vfs_kern_mount+0x78/0x180
[c000000cdcaefd00] [c000000000356d68] do_mount+0x258/0xea0
[c000000cdcaefde0] [c000000000357da0] SyS_mount+0xa0/0x110
[c000000cdcaefe30] [c00000000000bd84] system_call+0x38/0xe0


> If so, your patch is a a semantic curiousity because it's actually
> impossible for a NUMA allocation to be local and the definition of "HIT"
> is fuzzy enough to be useless.
>
> I won't object to the patch but it makes me trust "hit" even less than I
> already do for any analysis.
>
> Note that after this mail that I'll be unavailable by mail until early
> new years.