Re: [PATCHv12 26/37] mm: rework mapcount accounting to enable 4k mapping of THPs

[Kirill A. Shutemov]

On Tue, Oct 27, 2015 at 11:24:22PM +0000, Naoya Horiguchi wrote:
> On Tue, Oct 27, 2015 at 11:30:55AM +0200, Kirill A. Shutemov wrote:
> > On Tue, Oct 27, 2015 at 06:18:01AM +0000, Naoya Horiguchi wrote:
> > > On Tue, Oct 06, 2015 at 06:23:53PM +0300, Kirill A. Shutemov wrote:
> > > > We're going to allow mapping of individual 4k pages of THP compound.
> > > > It means we need to track mapcount on per small page basis.
> > > >
> > > > Straight-forward approach is to use ->_mapcount in all subpages to track
> > > > how many time this subpage is mapped with PMDs or PTEs combined. But
> > > > this is rather expensive: mapping or unmapping of a THP page with PMD
> > > > would require HPAGE_PMD_NR atomic operations instead of single we have
> > > > now.
> > > >
> > > > The idea is to store separately how many times the page was mapped as
> > > > whole -- compound_mapcount. This frees up ->_mapcount in subpages to
> > > > track PTE mapcount.
> > > >
> > > > We use the same approach as with compound page destructor and compound
> > > > order to store compound_mapcount: use space in first tail page,
> > > > ->mapping this time.
> > > >
> > > > Any time we map/unmap whole compound page (THP or hugetlb) -- we
> > > > increment/decrement compound_mapcount. When we map part of compound page
> > > > with PTE we operate on ->_mapcount of the subpage.
> > > >
> > > > page_mapcount() counts both: PTE and PMD mappings of the page.
> > > >
> > > > Basically, we have mapcount for a subpage spread over two counters.
> > > > It makes tricky to detect when last mapcount for a page goes away.
> > > >
> > > > We introduced PageDoubleMap() for this. When we split THP PMD for the
> > > > first time and there's other PMD mapping left we offset up ->_mapcount
> > > > in all subpages by one and set PG_double_map on the compound page.
> > > > These additional references go away with last compound_mapcount.
> > > >
> > > > This approach provides a way to detect when last mapcount goes away on
> > > > per small page basis without introducing new overhead for most common
> > > > cases.
> > > >
> > > > Signed-off-by: Kirill A. Shutemov <kirill.shutemov@xxxxxxxxxxxxxxx>
> > > > Tested-by: Aneesh Kumar K.V <aneesh.kumar@xxxxxxxxxxxxxxxxxx>
> > > > Acked-by: Jerome Marchand <jmarchan@xxxxxxxxxx>
> > > 
> > > I found that recent mmotm hit the following BUG_ON() by reading
> > > /proc/kpageflags over pfn backed by a thp.
> > > 
> > >   [  268.024519] page:ffffea00033e0000 count:0 mapcount:0 mapping:          (null) index:0x700000200
> > >   [  268.026076] flags: 0x4000000000000000()
> > >   [  268.026778] page dumped because: VM_BUG_ON_PAGE(!PageHead(page))
> > >   [  268.027816] page->mem_cgroup:ffff88021588cc00
> > >   [  268.028638] ------------[ cut here ]------------
> > >   [  268.029932] kernel BUG at /src/linux-dev/include/linux/page-flags.h:552!
> > >   [  268.031092] invalid opcode: 0000 [#1] SMP DEBUG_PAGEALLOC
> > >   [  268.032125] Modules linked in: cfg80211 rfkill crc32c_intel virtio_balloon serio_raw i2c_piix4 virtio_blk virtio_net ata_generic pata_acpi
> > >   [  268.032598] CPU: 0 PID: 1183 Comm: page-types Not tainted 4.2.0-mmotm-2015-10-21-14-41-151027-1418-00014-41+ #179
> > >   [  268.032598] Hardware name: Bochs Bochs, BIOS Bochs 01/01/2011
> > >   [  268.032598] task: ffff880214a08bc0 ti: ffff880213e2c000 task.ti: ffff880213e2c000
> > >   [  268.032598] RIP: 0010:[<ffffffff812434b6>]  [<ffffffff812434b6>] stable_page_flags+0x336/0x340
> > >   [  268.032598] RSP: 0018:ffff880213e2fda8  EFLAGS: 00010292
> > >   [  268.032598] RAX: 0000000000000021 RBX: ffff8802150a39c0 RCX: 0000000000000000
> > >   [  268.032598] RDX: ffff88021ec0ff38 RSI: ffff88021ec0d658 RDI: ffff88021ec0d658
> > >   [  268.032598] RBP: ffff880213e2fdc8 R08: 000000000000000a R09: 000000000000132f
> > >   [  268.032598] R10: 0000000000000000 R11: 000000000000132f R12: 4000000000000000
> > >   [  268.032598] R13: ffffea00033e6340 R14: 00007fff8449e430 R15: ffffea00033e6340
> > >   [  268.032598] FS:  00007ff7f9525700(0000) GS:ffff88021ec00000(0000) knlGS:0000000000000000
> > >   [  268.032598] CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
> > >   [  268.032598] CR2: 000000000063b800 CR3: 00000000d9e71000 CR4: 00000000000006f0
> > >   [  268.032598] Stack:
> > >   [  268.032598]  ffff8800db82df80 ffff8802150a39c0 0000000000000008 00000000000cf98d
> > >   [  268.032598]  ffff880213e2fe18 ffffffff81243588 00007fff8449e430 ffff880213e2ff20
> > >   [  268.032598]  000000000063b800 ffff8802150a39c0 fffffffffffffffb ffff880213e2ff20
> > >   [  268.032598] Call Trace:
> > >   [  268.032598]  [<ffffffff81243588>] kpageflags_read+0xc8/0x130
> > >   [  268.032598]  [<ffffffff81235848>] proc_reg_read+0x48/0x70
> > >   [  268.032598]  [<ffffffff811d6b08>] __vfs_read+0x28/0xd0
> > >   [  268.032598]  [<ffffffff812ee43e>] ? security_file_permission+0xae/0xc0
> > >   [  268.032598]  [<ffffffff811d6f53>] ? rw_verify_area+0x53/0xf0
> > >   [  268.032598]  [<ffffffff811d707a>] vfs_read+0x8a/0x130
> > >   [  268.032598]  [<ffffffff811d7bf7>] SyS_pread64+0x77/0x90
> > >   [  268.032598]  [<ffffffff81648117>] entry_SYSCALL_64_fastpath+0x12/0x6a
> > >   [  268.032598] Code: ca 00 00 40 01 48 39 c1 48 0f 44 da e9 a2 fd ff ff 48 c7 c6 50 a6 a1 8 1 e8 58 ab f4 ff 0f 0b 48 c7 c6 90 a2 a1 81 e8 4a ab f4 ff <0f> 0b 0f 1f 84 00 00 00 00 00 66 66 66 66 90 55 48 89 e5 41 57
> > >   [  268.032598] RIP  [<ffffffff812434b6>] stable_page_flags+0x336/0x340
> > >   [  268.032598]  RSP <ffff880213e2fda8>
> > >   [  268.070504] ---[ end trace e5d18553088c026a ]---
> > > 
> > > page_mapcount() could be called for a tail page, so VM_BUG_ON_PAGE(!PageHead())
> > > in PageDoubleMap() introduced by this patch seems too strong restriction.
> > 
> > Hm. page_mapcount() calls PageDoubleMap() only for head pages.
> > 
> > > Could you handle this?
> > 
> > Do you have a reproducer?
> 
> Yes, first you run a simple thp allocation like below in the background:
> 
>   #include <sys/mman.h>
>   #include <string.h>
> 
>   #define BASEADDR 0x700000000000
> 
>   int main(int argc, char *argv[])
>   {
>           char *addr;
>           int length = 0x200000;
> 
>           while (1) {
>                   addr = mmap((void *)BASEADDR, length, PROT_READ|PROT_WRITE,
>                               MAP_PRIVATE|MAP_ANONYMOUS, -1, 0);
>                   memset(addr, 0, length);
>                   munmap(addr, length);
>           }
>           return 0;
>   }
> 
> , and then you call tools/vm/page-types.c tool with -p pid option.
> After a few calls, the above bug should reproduce.

Okay, the problem is that the page was freed under stable_page_flags().

Is the code performance sensitive? Can we get reference to the page before
touching it? If not, we can rewrite the helper like this:

static inline int PageDoubleMap(struct page *page)
{
	return PageHead(page) && test_bit(PG_double_map, &page[1].flags);                         
}

Just dropping the check would be wrong, I think, as we access the next
page.

-- 
 Kirill A. Shutemov
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/