Re: [Intel-wired-lan] Increased memory usage on NUMA nodes with ICE driver after upgrade to 6.13.y (regression in commit 492a044508ad)

[Maciej Fijalkowski <maciej.fijalkowski@intel.com>]

On Thu, Jul 03, 2025 at 09:16:35AM -0700, Jacob Keller wrote:
> 
> 
> On 7/2/2025 11:46 PM, Jaroslav Pulchart wrote:
> >>
> >> On 7/2/2025 2:48 AM, Jaroslav Pulchart wrote:
> >>>>
> >>>> On 6/30/2025 11:48 PM, Jaroslav Pulchart wrote:
> >>>>>> On 6/30/2025 2:56 PM, Jacob Keller wrote:
> >>>>>>> Unfortunately it looks like the fix I mentioned has landed in 6.14, so
> >>>>>>> its not a fix for your issue (since you mentioned 6.14 has failed
> >>>>>>> testing in your system)
> >>>>>>>
> >>>>>>> $ git describe --first-parent --contains --match=v* --exclude=*rc*
> >>>>>>> 743bbd93cf29f653fae0e1416a31f03231689911
> >>>>>>> v6.14~251^2~15^2~2
> >>>>>>>
> >>>>>>> I don't see any other relevant changes since v6.14. I can try to see if
> >>>>>>> I see similar issues with CONFIG_MEM_ALLOC_PROFILING on some test
> >>>>>>> systems here.
> >>>>>>
> >>>>>> On my system I see this at boot after loading the ice module from
> >>>>>>
> >>>>>> $ grep -F "/ice/" /proc/allocinfo | sort -g | tail | numfmt --to=iec>
> >>>>>>       26K      230 drivers/net/ethernet/intel/ice/ice_irq.c:84 [ice]
> >>>>>> func:ice_get_irq_res
> >>>>>>>          48K        2 drivers/net/ethernet/intel/ice/ice_arfs.c:565 [ice] func:ice_init_arfs
> >>>>>>>          57K      226 drivers/net/ethernet/intel/ice/ice_lib.c:397 [ice] func:ice_vsi_alloc_ring_stats
> >>>>>>>          57K      226 drivers/net/ethernet/intel/ice/ice_lib.c:416 [ice] func:ice_vsi_alloc_ring_stats
> >>>>>>>          85K      226 drivers/net/ethernet/intel/ice/ice_lib.c:1398 [ice] func:ice_vsi_alloc_rings
> >>>>>>>         339K      226 drivers/net/ethernet/intel/ice/ice_lib.c:1422 [ice] func:ice_vsi_alloc_rings
> >>>>>>>         678K      226 drivers/net/ethernet/intel/ice/ice_base.c:109 [ice] func:ice_vsi_alloc_q_vector
> >>>>>>>         1.1M      257 drivers/net/ethernet/intel/ice/ice_fwlog.c:40 [ice] func:ice_fwlog_alloc_ring_buffs
> >>>>>>>         7.2M      114 drivers/net/ethernet/intel/ice/ice_txrx.c:493 [ice] func:ice_setup_rx_ring
> >>>>>>>         896M   229264 drivers/net/ethernet/intel/ice/ice_txrx.c:680 [ice] func:ice_alloc_mapped_page
> >>>>>>
> >>>>>> Its about 1GB for the mapped pages. I don't see any increase moment to
> >>>>>> moment. I've started an iperf session to simulate some traffic, and I'll
> >>>>>> leave this running to see if anything changes overnight.
> >>>>>>
> >>>>>> Is there anything else that you can share about the traffic setup or
> >>>>>> otherwise that I could look into?  Your system seems to use ~2.5 x the
> >>>>>> buffer size as mine, but that might just be a smaller number of CPUs.
> >>>>>>
> >>>>>> Hopefully I'll get some more results overnight.
> >>>>>
> >>>>> The traffic is random production workloads from VMs, using standard
> >>>>> Linux or OVS bridges. There is no specific pattern to it. I haven’t
> >>>>> had any luck reproducing (or was not patient enough) this with iperf3
> >>>>> myself. The two active (UP) interfaces are in an LACP bonding setup.
> >>>>> Here are our ethtool settings for the two member ports (em1 and p3p1)
> >>>>>
> >>>>
> >>>> I had iperf3 running overnight and the memory usage for
> >>>> ice_alloc_mapped_pages is constant here. Mine was direct connections
> >>>> without bridge or bonding. From your description I assume there's no XDP
> >>>> happening either.
> >>>
> >>> Yes, no XDP in use.
> >>>
> >>> BTW the allocinfo after 6days uptime:
> >>> # uptime ; sort -g /proc/allocinfo| tail -n 15
> >>>  11:46:44 up 6 days,  2:18,  1 user,  load average: 9.24, 11.33, 15.07
> >>>    102489024   533797 fs/dcache.c:1681 func:__d_alloc
> >>>    106229760    25935 mm/shmem.c:1854 func:shmem_alloc_folio
> >>>    117118192   103097 fs/ext4/super.c:1388 [ext4] func:ext4_alloc_inode
> >>>    134479872    32832 kernel/events/ring_buffer.c:811 func:perf_mmap_alloc_page
> >>>    162783232     7656 mm/slub.c:2452 func:alloc_slab_page
> >>>    189906944    46364 mm/memory.c:1056 func:folio_prealloc
> >>>    499384320   121920 mm/percpu-vm.c:95 func:pcpu_alloc_pages
> >>>    530579456   129536 mm/page_ext.c:271 func:alloc_page_ext
> >>>    625876992    54186 mm/slub.c:2450 func:alloc_slab_page
> >>>    838860800      400 mm/huge_memory.c:1165 func:vma_alloc_anon_folio_pmd
> >>>   1014710272   247732 mm/filemap.c:1978 func:__filemap_get_folio
> >>>   1056710656   257986 mm/memory.c:1054 func:folio_prealloc
> >>>   1279262720      610 mm/khugepaged.c:1084 func:alloc_charge_folio
> >>>   1334530048   325763 mm/readahead.c:186 func:ractl_alloc_folio
> >>>   3341238272   412215 drivers/net/ethernet/intel/ice/ice_txrx.c:681
> >>> [ice] func:ice_alloc_mapped_page
> >>>
> >> I have a suspicion that the issue is related to the updating of
> >> page_count in ice_get_rx_pgcnt(). The i40e driver has a very similar
> >> logic for page reuse but doesn't do this. It also has a counter to track
> >> failure to re-use the Rx pages.
> >>
> >> Commit 11c4aa074d54 ("ice: gather page_count()'s of each frag right
> >> before XDP prog call") changed the logic to update page_count of the Rx
> >> page just prior to the XDP call instead of at the point where we get the
> >> page from ice_get_rx_buf(). I think this change was originally
> >> introduced while we were trying out an experimental refactor of the
> >> hotpath to handle fragments differently, which no longer happens since
> >> 743bbd93cf29 ("ice: put Rx buffers after being done with current
> >> frame"), which ironically was part of this very same series..
> >>
> >> I think this updating of page count is accidentally causing us to
> >> miscount when we could perform page-reuse, and ultimately causes us to
> >> leak the page somehow. I'm still investigating, but I think this might
> >> trigger if somehow the page pgcnt - pagecnt_bias becomes >1, we don't
> >> reuse the page.
> >>
> >> The i40e driver stores the page count in i40e_get_rx_buffer, and I think
> >> our updating it later can somehow get things out-of-sync.
> >>
> >> Do you know if your traffic pattern happens to send fragmented frames? I
> > 
> > Hmm, I check the
> > * node_netstat_Ip_Frag* metrics and they are empty(do-not-exists),
> > * shortly run "tcpdump -n -i any 'ip[6:2] & 0x3fff != 0'" and nothing was found
> > looks to me like there is no fragmentation.
> > 
> 
> Good to rule it out at least.
> 
> >> think iperf doesn't do that, which might be part of whats causing this
> >> issue. I'm going to try to see if I can generate such fragmentation to
> >> confirm. Is your MTU kept at the default ethernet size?
> > 
> > Our MTU size is set to 9000 everywhere.
> > 
> 
> Ok. I am re-trying with MTU 9000 and using some traffic generated by wrk
> now. I do see much larger memory use (~2GB) when using MTU 9000, so that
> tracks with what your system shows. Currently its fluctuating between
> 1.9 and 2G. I'll leave this going for a couple of days while on vacation
> and see if anything pops up.

I was thinking if order-1 pages might do the mess there for some reason
since for 9k mtu we pull them and split into half.

Maybe it would be worth trying out if legacy-rx (which will work on
order-0 pages) doesn't have this issue? but that would require 8k mtu.

> 
> Thanks,
> Jake