Re: kernel BUG at mm/khugepaged.c:533 on 4.15.3

[Kirill A. Shutemov]

On Mon, Feb 19, 2018 at 10:51:01AM -0800, Laura Abbott wrote:
> Hi,
> 
> Fedora got a bug report of a BUG with 4.15.3:
> (https://bugzilla.redhat.com/show_bug.cgi?id=1546709)

Is it new to v4.15 kernel?
I don't see any recent change that could cause it.


> page:fffffac1800a0000 count:513 mapcount:1 mapping:ffff95657ef359a1 index:0x7f95d3400 compound_mapcount: 0
> flags: 0xffffe00048268(uptodate|lru|active|owner_priv_1|head|swapbacked)
> raw: 000ffffe00048268 ffff95657ef359a1 00000007f95d3400 0000020100000000
> raw: fffffac18edea9a0 fffffac18e7085a0 00000000000db400 ffff9567a3269800
> page dumped because: VM_BUG_ON_PAGE(PageCompound(page))
> page->mem_cgroup:ffff9567a3269800
> ------------[ cut here ]------------
> kernel BUG at mm/khugepaged.c:533!
> invalid opcode: 0000 [#1] SMP PTI
> Modules linked in: vhost_net vhost tap fuse xt_CHECKSUM ipt_MASQUERADE nf_nat_masquerade_ipv4 tun ip6t_rpfilter ip6t_REJECT nf_reject_ipv6 xt_conntrack ip_set nfnetlink ebtable_nat ebtable_broute bridge stp llc ip6table_nat nf_conntrack_ipv6 nf_defrag_ipv6 nf_nat_ipv6 ip6table_mangle ip6table_raw ip6table_security iptable_nat nf_conntrack_ipv4 nf_defrag_ipv4 nf_nat_ipv4 nf_nat nf_conntrack libcrc32c iptable_mangle iptable_raw iptable_security ebtable_filter ebtables ip6table_filter ip6_tables sunrpc vfat fat rmi_smbus rmi_core arc4 intel_rapl x86_pkg_temp_thermal intel_powerclamp coretemp snd_hda_codec_hdmi iwlmvm kvm_intel iTCO_wdt mac80211 snd_hda_codec_realtek iTCO_vendor_support mei_wdt kvm snd_hda_codec_generic irqbypass intel_cstate snd_hda_intel intel_uncore iwlwifi uvcvideo intel_rapl_perf
> snd_hda_codec videobuf2_vmalloc videobuf2_memops videobuf2_v4l2 videobuf2_core snd_hda_core cfg80211 videodev snd_hwdep snd_seq snd_seq_device snd_pcm media mei_me snd_timer thinkpad_acpi wmi_bmof rtsx_pci_ms joydev tpm_tis memstick i2c_i801 mei tpm_tis_core snd soundcore intel_pch_thermal tpm shpchp rfkill dm_crypt hid_logitech_hidpp hid_logitech_dj mmc_block nouveau i915 rtsx_pci_sdmmc mmc_core mxm_wmi ttm e1000e i2c_algo_bit drm_kms_helper crct10dif_pclmul crc32_pclmul crc32c_intel ptp drm ghash_clmulni_intel serio_raw rtsx_pci pps_core wmi video
> CPU: 2 PID: 66 Comm: khugepaged Not tainted 4.15.3-300.fc27.x86_64 #1
> Hardware name: LENOVO 20FXS0BB14/20FXS0BB14, BIOS R07ET63W (2.03 ) 03/15/2016
> RIP: 0010:khugepaged+0x1af6/0x2130
> RSP: 0018:ffffacacc1b4bdc0 EFLAGS: 00010282
> RAX: 0000000000000021 RBX: fffffac1800a0000 RCX: 0000000000000006
> RDX: 0000000000000000 RSI: 0000000000000086 RDI: ffff9567c14968f0
> RBP: fffffac18e3a5b40 R08: 00000000000004a8 R09: 0000000000000004
> R10: ffffacacc1b4bd70 R11: ffffffffb995b1ed R12: 00007f95f7e00000
> R13: ffff95661113eaf0 R14: ffff9567a9ea0000 R15: 8000000002800825
> FS:  0000000000000000(0000) GS:ffff9567c1480000(0000) knlGS:0000000000000000
> CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
> CR2: 0000000000481046 CR3: 00000002ee20a005 CR4: 00000000003626e0
> Call Trace:
> ? finish_wait+0x80/0x80
> ? collapse_shmem+0xdd0/0xdd0
> kthread+0x113/0x130
> ? kthread_create_worker_on_cpu+0x70/0x70
> ret_from_fork+0x35/0x40
> Code: ff e9 e7 fd ff ff bb 07 00 00 00 49 89 c7 e9 20 fb ff ff 48 83 ea 01 e9 66 fc ff ff 48 c7 c6 d8 3f 0a b9 48 89 df e8 0a 82 fa ff <0f> 0b 31 c9 4c 89 fa 48 89 de 4c 89 f7 e8 58 f1 fd ff e9 2e fa
> RIP: khugepaged+0x1af6/0x2130 RSP: ffffacacc1b4bdc0
> ---[ end trace a734c2f4d682e3bd ]---
> 
> Reporter said it happened several times. Config is attached.
> Any ideas?

Looks like somebody managed to insert THP into the range in split it back
between khugepaged_scan_pmd() and __collapse_huge_page_isolate().

That's rather unlikely chain of events, but I don't see other option.

Could you check if this works:

diff --git a/mm/khugepaged.c b/mm/khugepaged.c
index b7e2268dfc9a..c15da1ea7e63 100644
--- a/mm/khugepaged.c
+++ b/mm/khugepaged.c
@@ -530,7 +530,12 @@ static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
 			goto out;
 		}
 
-		VM_BUG_ON_PAGE(PageCompound(page), page);
+		/* TODO: teach khugepaged to collapse THP mapped with pte */
+		if (PageCompound(page)) {
+			result = SCAN_PAGE_COMPOUND;
+			goto out;
+		}
+
 		VM_BUG_ON_PAGE(!PageAnon(page), page);
 
 		/*
-- 
 Kirill A. Shutemov