On 01/15/2014 05:06 PM, Daniel Borkmann wrote:
> [keeping netdev in loop as well]
>
> On 01/15/2014 03:27 PM, Vlastimil Babka wrote:
>> On 01/13/2014 12:39 PM, Daniel Borkmann wrote:
>>> On 01/13/2014 11:16 AM, Vlastimil Babka wrote:
>>>> On 01/11/2014 02:32 PM, Daniel Borkmann wrote:
>>>>> On 01/11/2014 07:22 AM, Andrew Morton wrote:
>>>>>> On Fri, 10 Jan 2014 19:23:26 +0100 Daniel Borkmann <borkmann@xxxxxxxxxxxxx> wrote:
>>>>>>
>>>>>>> This is being reliably triggered for each mmaped() packet(7)
>>>>>>> socket from user space, basically during unmapping resp.
>>>>>>> closing the TX socket.
>>>>>>>
>>>>>>> I believe due to some change in transparent hugepages code ?
>>>>>>>
>>>>>>> When I disable transparent hugepages, everything works fine,
>>>>>>> no BUG triggered.
>>>>>>>
>>>>>>> I'd be happy to test patches.
>>>>>>
>>>>>> Did the inclusion of c424be1cbbf852e46acc8 ("mm: munlock: fix a bug
>>>>>> where THP tail page is encountered") in current mainline fix this?
>>>>>
>>>>> Thanks for your answer Andrew!
>>>>>
>>>>> Hm, I just cherry-picked that onto current net-next as I have some work
>>>>> there, and this time I got ...
>>>>>
>>>>> (User space uses packet mmap() and mlockall(MCL_CURRENT | MCL_FUTURE)
>>>>> and on shutdown munlockall() ...)
>>>>>
>>>>> [ 63.863672] ------------[ cut here ]------------
>>>>> [ 63.863702] kernel BUG at mm/mlock.c:507!
>>>>> [ 63.863721] invalid opcode: 0000 [#1] SMP
>>>>> [ 63.863743] Modules linked in: fuse ebtable_nat xt_CHECKSUM nf_conntrack_netbios_ns nf_conntrack_broadcast ipt_MASQUERADE ip6table_nat nf_nat_ipv6 ip6table_mangle ip6t_REJECT nf_conntrack_ipv6 nf_defrag_ipv6 iptable_nat nf_nat_ipv4 nf_nat iptable_mangle nf_conntrack_ipv4 nf_defrag_ipv4 xt_conntrack nf_conntrack bridge ebtable_filter ebtables stp llc ip6table_filter ip6_tables rfcomm bnep snd_hda_codec_hdmi snd_hda_codec_realtek snd_hda_intel snd_hda_codec iwlwifi cfg80211 snd_hwdep btusb snd_seq bluetooth sdhci_pci snd_seq_device e1000e tpm_tis snd_pcm thinkpad_acpi sdhci ptp tpm uvcvideo pps_core snd_page_alloc snd_timer snd rfkill mmc_core iTCO_wdt iTCO_vendor_support lpc_ich mfd_core soundcore joydev wmi videobuf2_vmalloc videobuf2_memops videobuf2_core i2c_i801 pcspkr videodev media uinput i915
>>>>> [ 63.864152] i2c_algo_bit drm_kms_helper drm i2c_core video
>>>>> [ 63.864181] CPU: 1 PID: 1617 Comm: trafgen Not tainted 3.13.0-rc6+ #15
>>>>> [ 63.864209] Hardware name: LENOVO 2429BP3/2429BP3, BIOS G4ET37WW (1.12 ) 05/29/2012
>>>>> [ 63.864242] task: ffff8801ee060000 ti: ffff8800b5954000 task.ti: ffff8800b5954000
>>>>> [ 63.864274] RIP: 0010:[<ffffffff8116fa9a>] [<ffffffff8116fa9a>] munlock_vma_pages_range+0x2ea/0x2f0
>>>>> [ 63.864318] RSP: 0018:ffff8800b5955e08 EFLAGS: 00010202
>>>>> [ 63.864341] RAX: 00000000000001ff RBX: ffff8800b58f7508 RCX: 0000000000000034
>>>>> [ 63.864372] RDX: 00000007f0708992 RSI: ffffea0002c3e700 RDI: ffffea0002c3e700
>>>>> [ 63.864402] RBP: ffff8800b5955ee0 R08: 3800000000000000 R09: a8000b0f9c000000
>>>>> [ 63.864432] R10: 57ffdef066c3e700 R11: ffffff5cfb00c14a R12: ffffea0002c3e700
>>>>> [ 63.864462] R13: ffff8800b5955f48 R14: 00007f0708992000 R15: 00007f0708992000
>>>>> [ 63.864492] FS: 00007f0708b92740(0000) GS:ffff88021e240000(0000) knlGS:0000000000000000
>>>>> [ 63.864526] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
>>>>> [ 63.864551] CR2: 00007f33bb373000 CR3: 00000000b2a2c000 CR4: 00000000001407e0
>>>>> [ 63.864581] Stack:
>>>>> [ 63.864593] ffff8800b5955ed0 00007f0708b91fff 00007f0708b92000 ffff8800b5955e48
>>>>> [ 63.864632] 000001ff810c864b ffff8801ee060000 0000000000000000 0000000000000000
>>>>> [ 63.864669] ffff8800b5955e58 ffff8801ee060000 0000000700000086 ffff8801ee060000
>>>>> [ 63.864708] Call Trace:
>>>>> [ 63.864724] [<ffffffff816956bc>] ? _raw_spin_unlock_irq+0x2c/0x30
>>>>> [ 63.864754] [<ffffffff81171b52>] ? vma_merge+0xc2/0x330
>>>>> [ 63.864786] [<ffffffff8116fb9c>] mlock_fixup+0xfc/0x190
>>>>> [ 63.864812] [<ffffffff8116fde7>] do_mlockall+0x87/0xc0
>>>>> [ 63.864836] [<ffffffff811702df>] sys_munlockall+0x2f/0x50
>>>>> [ 63.864873] [<ffffffff8169e192>] system_call_fastpath+0x16/0x1b
>>>>> [ 63.864898] Code: d7 48 89 95 28 ff ff ff e8 a4 04 fe ff 84 c0 48 8b 95 28 ff ff ff 0f 85 5a ff ff ff e9 46 ff ff ff e8 3f ac 51 00 e8 34 ac 51 00 <0f> 0b 0f 1f 40 00 0f 1f 44 00 00 55 48 89 e5 41 57 41 56 41 55
>>>>> [ 63.865114] RIP [<ffffffff8116fa9a>] munlock_vma_pages_range+0x2ea/0x2f0
>>>>> [ 63.865148] RSP <ffff8800b5955e08>
>>>>> [ 63.874968] ------------[ cut here ]------------
>>>>>
>>>>> ... when I find some time, I'll try with normal torvalds' tree, maybe some
>>>>> other patches are missing as well, not sure right now.
>>>>
>>>> Uh so the triggered assertion is the one added by this very patch, and there are no more changes wrt this in mainline.
>>>>
>>>> If you can still try debug patches, please try this. Thanks.
>>>
>>> Yes, thanks, I'll come back to you some time by today.
>>
>> Daniel sent me (off-list) instructions to reproduce:
>>
>>> Then in the kernel source tree, you'll find:
>>>
>>> tools/testing/selftests/net/
>>>
>>> There, just do a 'make' and run ./psock_tpacket
>>
>> It reproduces deterministically in mainline since 3.12, i.e. my munlock
>> performance series. Based on the initial debug output, I've expanded the
>> debug patch below a bit:
>>
>>>> From: Vlastimil Babka <vbabka@xxxxxxx>
>>>> Date: Mon, 13 Jan 2014 11:13:53 +0100
>>>> Subject: [PATCH] debug munlock_vma_pages_range
>>>>
>>>> ---
>>>> mm/mlock.c | 22 ++++++++++++++++++++--
>>>> 1 file changed, 20 insertions(+), 2 deletions(-)
>>>>
>>>> diff --git a/mm/mlock.c b/mm/mlock.c
>>>> index c59c420..7d0e29a 100644
>>>> --- a/mm/mlock.c
>>>> +++ b/mm/mlock.c
>>>> @@ -448,12 +448,14 @@ static unsigned long __munlock_pagevec_fill(struct pagevec *pvec,
>>>> void munlock_vma_pages_range(struct vm_area_struct *vma,
>>>> unsigned long start, unsigned long end)
>>>> {
>>>> + unsigned long orig_start = start;
>>>> + unsigned long page_increm = 0;
>>>> +
>>>> vma->vm_flags &= ~VM_LOCKED;
>>>>
>>>> while (start < end) {
>>>> struct page *page = NULL;
>>>> unsigned int page_mask;
>>>> - unsigned long page_increm;
>>>> struct pagevec pvec;
>>>> struct zone *zone;
>>>> int zoneid;
>>>> @@ -504,7 +506,23 @@ void munlock_vma_pages_range(struct vm_area_struct *vma,
>>>> }
>>>> }
>>>> /* It's a bug to munlock in the middle of a THP page */
>>>> - VM_BUG_ON((start >> PAGE_SHIFT) & page_mask);
>>>> + if ((start >> PAGE_SHIFT) & page_mask) {
>>>> + dump_page(page);
>>>> + printk("start=%lu pfn=%lu orig_start=%lu "
>>>> + "prev_page_increm=%lu page_mask=%u "
>>>> + "vm_start=%lu vm_end=%lu vm_flags=%lu\n",
>>>> + start, page_to_pfn(page), orig_start,
>>>> + page_increm, page_mask,
>>>> + vma->vm_start, vma->vm_end,
>>>> + vma->vm_flags);
>> + printk("vm_ops=%pF, open=%pF, fault=%pF, remap_pages=%pF\n", vma->vm_ops,
>> + vma->vm_ops->open, vma->vm_ops->fault, vma->vm_ops->remap_pages);
>> + if (PageCompound(page)) {
>> + printk("page is compound with order=%d\n", compound_order(page));
>> + }
>>>> + if (PageTail(page)) {
>>>> + struct page *first_page = page->first_page;
>>>> + printk("first_page pfn=%lu\n",
>>>> + page_to_pfn(first_page));
>>>> + dump_page(first_page);
>>>> + }
>>>> + VM_BUG_ON(true);
>>>> + }
>>>> page_increm = 1 + page_mask;
>>>> start += page_increm * PAGE_SIZE;
>>>> next:
>>>>
>>
>> And got output like this:
>>
>> page:ffffea0002474a40 count:5 mapcount:1 mapping: (null) index:0x0
>> page flags: 0x100000000004004(referenced|head)
>> start=140242647736320 pfn=682616 orig_start=140242647736320 prev_page_increm=0 page_mask=511 vm_start=140242647736320 vm_end=140242651930624 vm_flags=268435707
>> vm_ops=packet_mmap_ops+0x0/0xfffffffffffff8e0 [af_packet], open=packet_mm_open+0x0/0x30 [af_packet], fault= (null), remap_pages= (null)
>> page is compound with order=2
>>
>> Observations:
>> - address 140242647736320 is where the vma starts, and is not aligned to 512 pages
>> (so it cannot be a THP head which the munlock expects). Yet there is a head page
>> that triggers the PageTransHuge() and consequently hpage_nr_pages() in munlock_vma_page()
>> That's why page_mask is determined to be 511 and the code thinks it's in the
>> middle of a THP page.
>> - in fact, the page is a compound page with order=2
>> - the VM flags (except (may)read/write) are VM_SHARED and VM_MIXEDMAP
>> - the vma was mmapped by packet_mmap() (net/packet/af_packet.c) which uses
>> vm_insert_page(), which adds the VM_MIXEDMAP flag
>> - the buffers that are mapped were allocated by alloc_one_pg_vec_page()
>> where flags indeed include __GFP_COMP
>>
>> So clearly there is a way to have mlock/munlock operate on a vma that contains
>> compound pages and confuse the checks for PageTransHuge().
>>
>> The checks for THP in munlock came with commit ff6a6da60b89 ("mm: accelerate munlock()
>> treatment of THP pages"), i.e. since 3.9, but did not trigger a bug. It however
>> makes munlock_vma_pages_range() skip pages until the next 512-pages-aligned page,
>> when it encounters a head page. If the head page is of smaller order and is followed
>> by normal LRU pages (theoretically, I'm not sure if that's possible, or done anywhere),
>> they wouldn't get munlocked.
>>
>> My commit 7225522bb429 ("mm: munlock: batch non-THP page isolation and
>> munlock+putback using pagevec") (since 3.12) has added a new PageTransHuge() check
>> that can trigger on tail pages of the compound page here. Commit c424be1cbbf852e46acc8
>> ("mm: munlock: fix a bug where THP tail page is encountered") in current rc's removes
>> one class of bugs here, but still non-THP compound pages are not expected in mlock/munlock,
>> which leads to this assertion failing.
>>
>> The question is what is the correct fix, and I'm not that familiar with VM_MIXEDMAP
>> to decide.
>>
>> Option 1: mlocking VM_MIXEDMAP vma's has no sense. They should be treated like VM_PFNMAP
>> and added to VM_SPECIAL, which makes m(un)lock skip them completely.
>>
>> Option 2: if indeed VM_MIXEDMAP can contain PageLRU pages for which mlocking is useful,
>> VM_NO_THP should be checked in munlock before attempting PageTransHuge() and
>> friends. VM_NO_THP already contains VM_MIXEDMAP, so knowing that there can be
>> no THP means we don't try optimize for it and no unexpected head pages trip us.
>>
>> Thoughts?
OK, here's a RFC patch to hopefully help get us somewhere. I went for
Option1, as I didn't see anyone using VM_MIXEDMAP also for LRU pages,
and Option2 was ugly to implement and also seemed quite arbitrary. I'm
not sure if making VM_MIXEDMAP also non-mergeable this way is an issue
though.
Actually a nice third option would be that the vma in question would
include also another flag (other than VM_MIXEDMAP) that would declare it
can contain compound pages. But I didn't find any such rule documented,
perhaps VM_IO would be a candidate?
I added people that touched VM_MIXEDMAP in the past to CC.
-------8<-------
From: Vlastimil Babka <vbabka@xxxxxxx>
Date: Fri, 31 Jan 2014 11:50:21 +0100
Subject: [PATCH] mm: include VM_MIXEDMAP flag in the VM_SPECIAL list to avoid
m(un)locking
Daniel Borkmann reported a bug with VM_BUG_ON assertions failing where
munlock_vma_pages_range() thinks it's unexpectedly in the middle of a THP page.
This can be reproduced in tools/testing/selftests/net/ by running make and
then ./psock_tpacket.
The problem is that an order=2 compound page (allocated by
alloc_one_pg_vec_page() is part of the munlocked VM_MIXEDMAP vma (mapped by
packet_mmap()) and mistaken for a THP page and assumed to be order=9.
The checks for THP in munlock came with commit ff6a6da60b89 ("mm: accelerate
munlock() treatment of THP pages"), i.e. since 3.9, but did not trigger a bug.
It just makes munlock_vma_pages_range() skip such compound pages until the next
512-pages-aligned page, when it encounters a head page. This is however not a
problem for vma's where mlocking has no effect anyway, but it can distort the
accounting.
Since commit 7225522bb ("mm: munlock: batch non-THP page isolation and
munlock+putback using pagevec") this can trigger a VM_BUG_ON in PageTransHuge()
check.
This patch fixes the issue by adding VM_MIXEDMAP flag to VM_SPECIAL - a list of
flags that make vma's non-mlockable and non-mergeable. The reasoning is that
VM_MIXEDMAP vma's are similar to VM_PFNMAP, which is already on the VM_SPECIAL
list, and both are intended for non-LRU pages where mlocking makes no sense
anyway.
Signed-off-by: Vlastimil Babka <vbabka@xxxxxxx>
Reported-by: Daniel Borkmann <borkmann@xxxxxxxxxxxxx>
Cc: Thomas Hellstrom <thellstrom@xxxxxxxxxx>
Cc: John David Anglin <dave.anglin@xxxxxxxx>
Cc: HATAYAMA Daisuke <d.hatayama@xxxxxxxxxxxxxx>
Cc: Konstantin Khlebnikov <khlebnikov@xxxxxxxxxx>
Cc: Carsten Otte <cotte@xxxxxxxxxx>
Cc: Peter Zijlstra <peterz@xxxxxxxxxxxxx>
Cc: Jared Hulbert <jaredeh@xxxxxxxxx>
Signed-off-by: Vlastimil Babka <vbabka@xxxxxxx>
---
include/linux/mm.h | 2 +-
mm/huge_memory.c | 2 +-
2 files changed, 2 insertions(+), 2 deletions(-)
diff --git a/include/linux/mm.h b/include/linux/mm.h
index f28f46e..f9b04ac 100644
--- a/include/linux/mm.h
+++ b/include/linux/mm.h
@@ -175,7 +175,7 @@ extern unsigned int kobjsize(const void *objp);
* Special vmas that are non-mergable, non-mlock()able.
* Note: mm/huge_memory.c VM_NO_THP depends on this definition.
*/
-#define VM_SPECIAL (VM_IO | VM_DONTEXPAND | VM_PFNMAP)
+#define VM_SPECIAL (VM_IO | VM_DONTEXPAND | VM_PFNMAP | VM_MIXEDMAP)
/*
* mapping from the currently active vm_flags protection bits (the
diff --git a/mm/huge_memory.c b/mm/huge_memory.c
index 82166bf..f32fffb 100644
--- a/mm/huge_memory.c
+++ b/mm/huge_memory.c
@@ -1963,7 +1963,7 @@ out:
return ret;
}
-#define VM_NO_THP (VM_SPECIAL|VM_MIXEDMAP|VM_HUGETLB|VM_SHARED|VM_MAYSHARE)
+#define VM_NO_THP (VM_SPECIAL|VM_HUGETLB|VM_SHARED|VM_MAYSHARE)
int hugepage_madvise(struct vm_area_struct *vma,
unsigned long *vm_flags, int advice)
--
1.8.4.5
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@xxxxxxxxx. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>