Re: [Bug 206569] BUG at mm/filemap.c:2529: VM_BUG_ON_PAGE(page_to_pgoff(page) != offset)

[Song Liu <songliubraving@fb.com>]

On Feb 21, 2020, at 12:57 AM, Denis Lisov <dennis.lissov@gmail.com> wrote:
> 
> The reproducer I've been using was to start mass parallel rebuild of
> binary packages with
> 
> /usr/bin/emerge -1 --buildpkgonly --ask=n kde-plasma/bluedevil
> kde-plasma/breeze kde-plasma/breeze-gtk kde-plasma/drkonqi
> kde-plasma/kactivitymanagerd kde-plasma/kde-cli-tools
> kde-plasma/kde-gtk-config kde-plasma/kdecoration
> kde-plasma/kdeplasma-addons kde-plasma/kgamma kde-plasma/khotkeys
> kde-plasma/kinfocenter kde-plasma/kmenuedit kde-plasma/kscreen
> kde-plasma/kscreenlocker kde-plasma/ksshaskpass kde-plasma/ksysguard
> kde-plasma/kwallet-pam kde-plasma/kwayland-integration kde-plasma/kwin
> kde-plasma/kwrited kde-plasma/libkscreen kde-plasma/libksysguard
> kde-plasma/libkworkspace kde-plasma/milou kde-plasma/oxygen
> kde-plasma/plasma-browser-integration kde-plasma/plasma-desktop
> kde-plasma/plasma-integration kde-plasma/plasma-meta
> kde-plasma/plasma-vault kde-plasma/plasma-workspace
> kde-plasma/plasma-workspace-wallpapers kde-plasma/polkit-kde-agent
> kde-plasma/powerdevil kde-plasma/sddm-kcm kde-plasma/systemsettings
> kde-plasma/user-manager kde-plasma/xdg-desktop-portal-kde
> kde-plasma/xembed-sni-proxy
> 
> and the bug usually reproduces in 5-10 min after starting this. I've
> also seen it reproduce during a plain kernel build with `make clean &&
> make -j8`, but less reliably (the build could accidentally pass).
> 
> I'll see if I can make a self-contained reproducer in some way.
> 
> Denis Lisov.

Some updates:

On my VM, I saw khugepaged put .text of python binary in 2MB pages. 
This is caused by khugepaged_always() == true in hugepage_vma_check(). 
I guess this means the bug exists w/o CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS, 
but CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS makes it easier to trigger. 

However, I haven't got luck to repro this issue. 

Denis, what is the version of gcc are you using to repro this issue?
Also, could you please try the repro with he following patch:

diff --git a/mm/khugepaged.c b/mm/khugepaged.c
index b679908743cb..5b78590d4d6d 100644
--- a/mm/khugepaged.c
+++ b/mm/khugepaged.c
@@ -1718,6 +1718,10 @@ static void collapse_file(struct mm_struct *mm,
        else {
                __inc_node_page_state(new_page, NR_FILE_THPS);
                filemap_nr_thps_inc(mapping);
+               pr_info("%s ino %lu, d_iname %s comm %s\n", __func__,
+                       mapping->host->i_ino,
+                       file->f_path.dentry->d_iname,
+                       mm->owner->comm);
        }

        if (nr_none) {

Thanks,
Song