On Mon, Feb 18, 2019 at 04:22:13PM +0100, Michal Hocko wrote:
> On Mon 18-02-19 16:20:50, Michal Hocko wrote:
> > On Mon 18-02-19 16:05:15, Mike Rapoport wrote:
> > > On Mon, Feb 18, 2019 at 11:30:13AM +0100, Michal Hocko wrote:
> > > > On Mon 18-02-19 18:01:39, Rong Chen wrote:
> > > > >
> > > > > On 2/18/19 4:55 PM, Michal Hocko wrote:
> > > > > > [Sorry for an excessive quoting in the previous email]
> > > > > > [Cc Pavel - the full report is
> > > > > > http://lkml.kernel.org/r/20190218052823.GH29177@shao2-debian[]
> > > > > >
> > > > > > On Mon 18-02-19 08:08:44, Michal Hocko wrote:
> > > > > > > On Mon 18-02-19 13:28:23, kernel test robot wrote:
> > > > > > [...]
> > > > > > > > [ 40.305212] PGD 0 P4D 0
> > > > > > > > [ 40.308255] Oops: 0000 [#1] PREEMPT SMP PTI
> > > > > > > > [ 40.313055] CPU: 1 PID: 239 Comm: udevd Not tainted
> > > > > > > > 5.0.0-rc4-00149-gefad4e4 #1
> > > > > > > > [ 40.321348] Hardware name: QEMU Standard PC (i440FX + PIIX,
> > > > > > > > 1996), BIOS 1.10.2-1 04/01/2014
> > > > > > > > [ 40.330813] RIP: 0010:page_mapping+0x12/0x80
> > > > > > > > [ 40.335709] Code: 5d c3 48 89 df e8 0e ad 02 00 85 c0 75 da
> > > > > > > > 89 e8 5b 5d c3 0f 1f 44 00 00 53 48 89 fb 48 8b 43 08 48 8d 50
> > > > > > > > ff a8 01 48 0f 45 da <48> 8b 53 08 48 8d 42 ff 83 e2 01 48 0f
> > > > > > > > 44 c3 48 83 38 ff 74 2f 48
> > > > > > > > [ 40.356704] RSP: 0018:ffff88801fa87cd8 EFLAGS: 00010202
> > > > > > > > [ 40.362714] RAX: ffffffffffffffff RBX: fffffffffffffffe RCX:
> > > > > > > > 000000000000000a
> > > > > > > > [ 40.370798] RDX: fffffffffffffffe RSI: ffffffff820b9a20 RDI:
> > > > > > > > ffff88801e5c0000
> > > > > > > > [ 40.378830] RBP: 6db6db6db6db6db7 R08: ffff88801e8bb000 R09:
> > > > > > > > 0000000001b64d13
> > > > > > > > [ 40.386902] R10: ffff88801fa87cf8 R11: 0000000000000001 R12:
> > > > > > > > ffff88801e640000
> > > > > > > > [ 40.395033] R13: ffffffff820b9a20 R14: ffff88801f145258 R15:
> > > > > > > > 0000000000000001
> > > > > > > > [ 40.403138] FS: 00007fb2079817c0(0000)
> > > > > > > > GS:ffff88801dd00000(0000) knlGS:0000000000000000
> > > > > > > > [ 40.412243] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
> > > > > > > > [ 40.418846] CR2: 0000000000000006 CR3: 000000001fa82000 CR4:
> > > > > > > > 00000000000006a0
> > > > > > > > [ 40.426951] Call Trace:
> > > > > > > > [ 40.429843] __dump_page+0x14/0x2c0
> > > > > > > > [ 40.433947] is_mem_section_removable+0x24c/0x2c0
> > > > > > > This looks like we are stumbling over an unitialized struct page
> > > > > > > again.
> > > > > > > Something this patch should prevent from. Could you try to apply
> > > > > > > [1]
> > > > > > > which will make __dump_page more robust so that we do not blow up
> > > > > > > there
> > > > > > > and give some more details in return.
> > > > > > >
> > > > > > > Btw. is this reproducible all the time?
> > > > > > And forgot to ask whether this is reproducible with pending mmotm
> > > > > > patches in linux-next.
> > > > >
> > > > >
> > > > > Do you mean the below patch? I can reproduce the problem too.
> > > >
> > > > Yes, thanks for the swift response. The patch has just added a debugging
> > > > output
> > > > [ 0.013697] Early memory node ranges
> > > > [ 0.013701] node 0: [mem 0x0000000000001000-0x000000000009efff]
> > > > [ 0.013706] node 0: [mem 0x0000000000100000-0x000000001ffdffff]
> > > > [ 0.013711] zeroying 0-1
> > > >
> > > > This is the first pfn.
> > > >
> > > > [ 0.013715] zeroying 9f-100
> > > >
> > > > this is [mem 0x9f000, 0xfffff] so it fills up the whole hole between the
> > > > above two ranges. This is definitely good.
> > > >
> > > > [ 0.013722] zeroying 1ffe0-1ffe0
> > > >
> > > > this is a single page at 0x1ffe0000 right after the zone end.
> > > >
> > > > [ 0.013727] Zeroed struct page in unavailable ranges: 98 pages
> > > >
> > > > Hmm, so this is getting really interesting. The whole zone range should
> > > > be covered. So this is either some off-by-one or I something that I am
> > > > missing right now. Could you apply the following on top please? We
> > > > definitely need to see what pfn this is.
> > > >
> > > >
> > > > diff --git a/mm/memory_hotplug.c b/mm/memory_hotplug.c
> > > > index 124e794867c5..59bcfd934e37 100644
> > > > --- a/mm/memory_hotplug.c
> > > > +++ b/mm/memory_hotplug.c
> > > > @@ -1232,12 +1232,14 @@ static bool
> > > > is_pageblock_removable_nolock(struct page *page)
> > > > /* Checks if this range of memory is likely to be hot-removable. */
> > > > bool is_mem_section_removable(unsigned long start_pfn, unsigned long
> > > > nr_pages)
> > > > {
> > > > - struct page *page = pfn_to_page(start_pfn);
> > > > + struct page *page = pfn_to_page(start_pfn), *first_page;
> > > > unsigned long end_pfn = min(start_pfn + nr_pages,
> > > > zone_end_pfn(page_zone(page)));
> > > > struct page *end_page = pfn_to_page(end_pfn);
> > > >
> > > > /* Check the starting page of each pageblock within the range */
> > > > - for (; page < end_page; page = next_active_pageblock(page)) {
> > > > + for (first_page = page; page < end_page; page =
> > > > next_active_pageblock(page)) {
> > > > + if (PagePoisoned(page))
> > > > + pr_info("Unexpected poisoned page %px
> > > > pfn:%lx\n", page, start_pfn + page-first_page);
> > > > if (!is_pageblock_removable_nolock(page))
> > > > return false;
> > > > cond_resched();
> > >
> > > I've added more prints and somehow end_page gets too big (in brackets is
> > > the pfn):
> > >
> > > [ 11.183835] ===> start: ffff88801e240000(0), end:
> > > ffff88801e400000(8000)
> > > [ 11.188457] ===> start: ffff88801e400000(8000), end:
> > > ffff88801e640000(10000)
> > > [ 11.193266] ===> start: ffff88801e640000(10000), end:
> > > ffff88801e060000(18000)
> > >
> > > should be
> > > ffff88801e5c0000
> > >
> > > [ 11.197363] ===> start: ffff88801e060000(18000), end:
> > > ffff88801e21f900(1ffe0)
> > > [ 11.207547] Unexpected poisoned page ffff88801e5c0000 pfn:10000
> > >
> > >
> > > With the patch below the problem seem to disappear, although I have no
> > > idea
> > > why...
> > >
> > > diff --git a/mm/memory_hotplug.c b/mm/memory_hotplug.c
> > > index 91e6fef..53d15ff 100644
> > > --- a/mm/memory_hotplug.c
> > > +++ b/mm/memory_hotplug.c
> > > @@ -1234,7 +1234,7 @@ bool is_mem_section_removable(unsigned long
> > > start_pfn, unsigned long nr_pages)
> > > {
> > > struct page *page = pfn_to_page(start_pfn);
> > > unsigned long end_pfn = min(start_pfn + nr_pages,
> > > zone_end_pfn(page_zone(page)));
> > > - struct page *end_page = pfn_to_page(end_pfn);
> > > + struct page *end_page = page + (end_pfn - start_pfn);
> > >
> > > /* Check the starting page of each pageblock within the range */
> > > for (; page < end_page; page = next_active_pageblock(page)) {
> >
> > This is really interesting, because it would mean that the end_pfn is
> > out of the section and so the page pointer arithmetic doesn't really
> > work. But I am wondering how that could happen as nr_pages is
> > PAGES_PER_SECTION. Another option is that pfn_to_page doesn't work
> > properly here. It is CONFIG_SPARSEMEM. Could you print section_nr of
> > both start_pfn and end_pfn please?
[ 11.118745] ===> start: ffff88801e240000(0), end: ffff88801e400000(8000)
[ 11.118745] ===> s_sec: 0, e_sec: 1
[ 11.123876] ===> start: ffff88801e640000(10000), end: ffff88801e060000(18000)
[ 11.123876] ===> s_sec: 2, e_sec: 3
[ 11.126835] ===> start: ffff88801e400000(8000), end: ffff88801e640000(10000)
[ 11.126835] ===> s_sec: 1, e_sec: 2
[ 11.130546] ===> start: ffff88801e060000(18000), end: ffff88801e21f900(1ffe0)
[ 11.130546] ===> s_sec: 3, e_sec: 3
[ 11.149693] Unexpected poisoned page ffff88801e5c0000 pfn:10000
The sections for start and end are different.
> Thinking about it some more, is it possible that we are overflowing by 1
> here?
Looks like that, the end_pfn is actually the first pfn in the next section.
> diff --git a/mm/memory_hotplug.c b/mm/memory_hotplug.c
> index 124e794867c5..6618b9d3e53a 100644
> --- a/mm/memory_hotplug.c
> +++ b/mm/memory_hotplug.c
> @@ -1234,10 +1234,10 @@ bool is_mem_section_removable(unsigned long
> start_pfn, unsigned long nr_pages)
> {
> struct page *page = pfn_to_page(start_pfn);
> unsigned long end_pfn = min(start_pfn + nr_pages,
> zone_end_pfn(page_zone(page)));
> - struct page *end_page = pfn_to_page(end_pfn);
> + struct page *end_page = pfn_to_page(end_pfn - 1);
>
> /* Check the starting page of each pageblock within the range */
> - for (; page < end_page; page = next_active_pageblock(page)) {
> + for (; page <= end_page; page = next_active_pageblock(page)) {
> if (!is_pageblock_removable_nolock(page))
> return false;
> cond_resched();
Works with your fix, but I think mine is more intuitive ;-)
> --
> Michal Hocko
> SUSE Labs
>
--
Sincerely yours,
Mike.
