On 04/18/2018 10:40 PM, Fengguang Wu wrote: > [ 716.494065] PASS concurrent_autogo_5ghz_ht40 4.803608 2018-03-23 09:57:21.586794 > [ 716.494069] > [ 716.496923] passed all 1 test case(s) > [ 716.496926] > [ 716.511702] swap_info_get: Bad swap file entry 04000000 > [ 716.512731] BUG: Bad page map in process python2 pte:100_0000_0000 pmd:17e8be067 > [ 716.513844] addr:00000000860ba23b vm_flags:00000070 anon_vma: (null) mapping:000000004c76fece index:1e2 > [ 716.515160] file:libpcre.so.3.13.3 fault:filemap_fault mmap:generic_file_mmap readpage:simple_readpage > [ 716.516418] CPU: 2 PID: 8907 Comm: python2 Not tainted 4.16.0-rc5 #1 > [ 716.517533] Hardware name: /DH67GD, BIOS BLH6710H.86A.0132.2011.1007.1505 10/07/2011 Did you say that you have a few more examples of this? I would be really interested if it's always python or always the same shared library, or always file-backed memory, always the same bit, etc... From the vm_flags, I'd guess that this is the "rw-p" part of the file mapping. The bit that gets set is really weird. It's bit 40. I could definitely see scenarios where we might set the dirty bit, or even NX for that matter, or some *bit* that we mess with in software. It's not even close to the boundary where it could represent a swapoffset=1 or swapfile=1. It's also unlikely to be _PAGE_PSE having gone missing from the PMD since it's in the middle of a file-backed mapping and the PMD is obviously pointing to a 4k page. If I had to put money on it, I'd guess it's a hardware bit flip, or less likely, a rogue software bit flip. But, more examples will hopefully shed some more light.
