On Wed, Nov 2, 2022 at 12:14 AM Michal Hocko <mhocko@xxxxxxxx> wrote:
>
> On Tue 01-11-22 10:12:49, Yang Shi wrote:
> > On Tue, Nov 1, 2022 at 12:54 AM Michal Hocko <mhocko@xxxxxxxx> wrote:
> > >
> > > On Mon 31-10-22 17:05:06, Zach O'Keefe wrote:
> > > > On Mon, Oct 31, 2022 at 3:08 PM Michal Hocko <mhocko@xxxxxxxx> wrote:
> > > > >
> > > > > On Mon 31-10-22 11:31:22, Yang Shi wrote:
> > > > > > Syzbot reported the below splat:
> > > > > >
> > > > > > WARNING: CPU: 1 PID: 3646 at include/linux/gfp.h:221 __alloc_pages_node include/linux/gfp.h:221 [inline]
> > > > > > WARNING: CPU: 1 PID: 3646 at include/linux/gfp.h:221 hpage_collapse_alloc_page mm/khugepaged.c:807 [inline]
> > > > > > WARNING: CPU: 1 PID: 3646 at include/linux/gfp.h:221 alloc_charge_hpage+0x802/0xaa0 mm/khugepaged.c:963
> > > > > > Modules linked in:
> > > > > > CPU: 1 PID: 3646 Comm: syz-executor210 Not tainted 6.1.0-rc1-syzkaller-00454-ga70385240892 #0
> > > > > > Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS Google 10/11/2022
> > > > > > RIP: 0010:__alloc_pages_node include/linux/gfp.h:221 [inline]
> > > > > > RIP: 0010:hpage_collapse_alloc_page mm/khugepaged.c:807 [inline]
> > > > > > RIP: 0010:alloc_charge_hpage+0x802/0xaa0 mm/khugepaged.c:963
> > > > > > Code: e5 01 4c 89 ee e8 6e f9 ae ff 4d 85 ed 0f 84 28 fc ff ff e8 70 fc ae ff 48 8d 6b ff 4c 8d 63 07 e9 16 fc ff ff e8 5e fc ae ff <0f> 0b e9 96 fa ff ff 41 bc 1a 00 00 00 e9 86 fd ff ff e8 47 fc ae
> > > > > > RSP: 0018:ffffc90003fdf7d8 EFLAGS: 00010293
> > > > > > RAX: 0000000000000000 RBX: 0000000000000000 RCX: 0000000000000000
> > > > > > RDX: ffff888077f457c0 RSI: ffffffff81cd8f42 RDI: 0000000000000001
> > > > > > RBP: ffff888079388c0c R08: 0000000000000001 R09: 0000000000000000
> > > > > > R10: 0000000000000000 R11: 0000000000000000 R12: 0000000000000000
> > > > > > R13: dffffc0000000000 R14: 0000000000000000 R15: 0000000000000000
> > > > > > FS: 00007f6b48ccf700(0000) GS:ffff8880b9b00000(0000) knlGS:0000000000000000
> > > > > > CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
> > > > > > CR2: 00007f6b48a819f0 CR3: 00000000171e7000 CR4: 00000000003506e0
> > > > > > DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
> > > > > > DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
> > > > > > Call Trace:
> > > > > > <TASK>
> > > > > > collapse_file+0x1ca/0x5780 mm/khugepaged.c:1715
> > > > >
> > > > > This is quite weird, isn't it? alloc_charge_hpage is selecting the most
> > > > > busy node (as per collapse_control). How come this can be an offline
> > > > > node? Is a parallel memory hotplug happening?
> > > >
> > > > TBH -- I did not look closely at the syzbot reproducer (let alone
> > > > attempt to run it) and assumed this was the case. Taking a quick look,
> > > > at least memory hot remove is enabled:
> > > >
> > > > CONFIG_ARCH_ENABLE_MEMORY_HOTPLUG=y
> > > > CONFIG_ARCH_ENABLE_MEMORY_HOTREMOVE=y
> > > > CONFIG_MEMORY_HOTPLUG=y
> > > > CONFIG_MEMORY_HOTPLUG_DEFAULT_ONLINE=y
> > > > CONFIG_MEMORY_HOTREMOVE=y
> > > >
> > > > But looking at the C reproducer, I don't immediately see anywhere
> > > > where we offline nodes. I'll try to run this tomorrow to make sure I'm
> > > > not missing something real here.
> > >
> > > Looking slightly closer at hpage_collapse_scan_file I think that it is
> > > possible that xas_for_each simply doesn't find any entries in the page
> > > cache and with khugepaged_max_ptes_none == HPAGE_PMD_NR we can fall back
> > > to collapse_file even without any real entries.
> >
> > The khugepaged_max_ptes_none can't be HPAGE_PMD_NR, it must be <=
> > (HPAGE_PMD_NR - 1), but MADV_COLLAPSE does ignore it.
>
> OK, I see.
>
> > But a closer look at the code about how to pick up the preferred node,
> > there seems to be a corner case for MADV_COLLAPSE.
> >
> > The code tried to do some balance if several nodes have the same hit
> > record. Basically it does conceptually:
> > * If the target_node <= last_target_node, then iterate from
> > last_target_node + 1 to MAX_NUMNODES (1024 on default config)
> > * If the max_value == node_load[nid], then target_node = nid
>
> Correct
>
> > So assuming the system has 2 nodes, the target_node is 0 and the
> > last_target_node is 1, if MADV_COLLAPSE path is hit, then it may
> > return 2 for target_node, but it is actually not existing (offline),
> > so the warn is triggered.
>
> How can node_load[2] > 0 (IIUC max_value > 0 here) if the node is
> offline (other than a race with hotplug)?
The max_value may be 0 if there is no entry in page cache for that range IIUC.
>
> > The below patch should be able to fix it:
> >
> > diff --git a/mm/khugepaged.c b/mm/khugepaged.c
> > index ea0d186bc9d4..d24405e6736b 100644
> > --- a/mm/khugepaged.c
> > +++ b/mm/khugepaged.c
> > @@ -787,7 +787,8 @@ static int hpage_collapse_find_target_node(struct
> > collapse_control *cc)
> > if (target_node <= cc->last_target_node)
> > for (nid = cc->last_target_node + 1; nid < MAX_NUMNODES;
> > nid++)
> > - if (max_value == cc->node_load[nid]) {
> > + if (node_online(nid) &&
> > + max_value == cc->node_load[nid]) {
> > target_node = nid;
> > break;
> > }
>
> Node, this is equally racy. node_online might become false right after
> the check.
Yes, my point is this could filter out the non-existing node case, for
example, the system just has 2 nodes physically, but MAX_NUMNODES > 2.
It doesn't prevent offlining, it is still racy, just like other
__GFP_THISNODE call sites.
> --
> Michal Hocko
> SUSE Labs
