Re: memleaks, acpi + ext4 + tty

["Luis R. Rodriguez" <mcgrof@gmail.com>]

On Fri, Aug 28, 2009 at 3:09 PM, Luis R. Rodriguez<mcgrof@gmail.com> wrote:
> On Fri, Aug 28, 2009 at 2:50 PM, Luis R. Rodriguez<mcgrof@gmail.com> wrote:
>> On Fri, Aug 28, 2009 at 9:52 AM, Luis R. Rodriguez<mcgrof@gmail.com> wrote:
>>> On Fri, Aug 28, 2009 at 9:32 AM, Catalin Marinas<catalin.marinas@arm.com> wrote:
>>>> "Luis R. Rodriguez" <mcgrof@gmail.com> wrote:
>>>>> I have an assorted collection of kmemleak reports for acpi, ext4 and
>>>>> tty, not sure how to read these yet to fix so figure I'd at least post
>>>>> them. To reproduce I can just dd=/dev/zero to some big file and played
>>>>> some video.
>>>>
>>>> If you do a few echo scan > /sys/kernel/debug/kmemleak, do they
>>>> disappear (i.e. transient false positives)?
>>>
>>> Sure, I will once on rc8.
>>>
>>>> Which kernel version is this?
>>>
>>> v2.6.31-rc7-33172-gf4a9f9a
>>>
>>> This is from wireless-testing, which has wireless patches on top of
>>> rc7. John just rebased to rc8 so will give that a shot at work.
>>>
>>>>> unreferenced object 0xffff88003e0015c0 (size 64):
>>>>>   comm "swapper", pid 1, jiffies 4294892352
>>>>>   backtrace:
>>>>>     [<ffffffff81121fad>] create_object+0x13d/0x2d0
>>>>>     [<ffffffff81122265>] kmemleak_alloc+0x25/0x60
>>>>>     [<ffffffff81118a03>] kmem_cache_alloc_node+0x193/0x200
>>>>>     [<ffffffff8152509e>] process_zones+0x70/0x1cd
>>>>>     [<ffffffff81525230>] pageset_cpuup_callback+0x35/0x92
>>>>>     [<ffffffff8152c9b7>] notifier_call_chain+0x47/0x90
>>>>>     [<ffffffff81078549>] __raw_notifier_call_chain+0x9/0x10
>>>>>     [<ffffffff81523f25>] _cpu_up+0x75/0x130
>>>>>     [<ffffffff8152403a>] cpu_up+0x5a/0x6a
>>>>>     [<ffffffff8181969e>] kernel_init+0xcc/0x1ba
>>>>>     [<ffffffff810130ca>] child_rip+0xa/0x20
>>>>>     [<ffffffffffffffff>] 0xffffffffffffffff
>>>>
>>>> Can't really tell. Maybe a false positive caused by kmemleak not
>>>> scanning the pgdata node_zones. Can you post your .config file?
>>>
>>> Sure, attached.
>>>
>>>>> unreferenced object 0xffff88003cb5f700 (size 64):
>>>>>   comm "swapper", pid 1, jiffies 4294892459
>>>>>   backtrace:
>>>>>     [<ffffffff81121fad>] create_object+0x13d/0x2d0
>>>>>     [<ffffffff81122265>] kmemleak_alloc+0x25/0x60
>>>>>     [<ffffffff81119f3b>] __kmalloc+0x16b/0x250
>>>>>     [<ffffffff812bb549>] kzalloc+0xf/0x11
>>>>>     [<ffffffff812bbb53>] acpi_add_single_object+0x58e/0xd3c
>>>>>     [<ffffffff812bc51c>] acpi_bus_scan+0x125/0x1af
>>>>>     [<ffffffff81842361>] acpi_scan_init+0xc8/0xe9
>>>>>     [<ffffffff8184211c>] acpi_init+0x21f/0x265
>>>>>     [<ffffffff8100a05b>] do_one_initcall+0x4b/0x1b0
>>>>>     [<ffffffff81819736>] kernel_init+0x164/0x1ba
>>>>>     [<ffffffff810130ca>] child_rip+0xa/0x20
>>>>>     [<ffffffffffffffff>] 0xffffffffffffffff
>>>>
>>>> I get ACPI reports as well and they may be real leaks. However, I
>>>> didn't have time to analyse the code (pretty complicated reference
>>>> counting).
>>>
>>> Heh OK thanks for reviewing them though.
>>>
>>>>> unreferenced object 0xffff880039571800 (size 1024):
>>>>>   comm "exe", pid 1168, jiffies 4294893410
>>>>>   backtrace:
>>>>>     [<ffffffff81121fad>] create_object+0x13d/0x2d0
>>>>>     [<ffffffff81122265>] kmemleak_alloc+0x25/0x60
>>>>>     [<ffffffff81119f3b>] __kmalloc+0x16b/0x250
>>>>>     [<ffffffff811e1d71>] ext4_mb_init+0x1a1/0x590
>>>>>     [<ffffffff811d2da3>] ext4_fill_super+0x1df3/0x26c0
>>>>>     [<ffffffff8112774f>] get_sb_bdev+0x16f/0x1b0
>>>>>     [<ffffffff811c8fd3>] ext4_get_sb+0x13/0x20
>>>>>     [<ffffffff81127216>] vfs_kern_mount+0x76/0x180
>>>>>     [<ffffffff8112738d>] do_kern_mount+0x4d/0x130
>>>>>     [<ffffffff8113fc57>] do_mount+0x307/0x8b0
>>>>>     [<ffffffff8114028f>] sys_mount+0x8f/0xe0
>>>>>     [<ffffffff81011f02>] system_call_fastpath+0x16/0x1b
>>>>>     [<ffffffffffffffff>] 0xffffffffffffffff
>>>>
>>>> The ext4 reports are real leaks and patch was posted here -
>>>> http://lkml.org/lkml/2009/7/15/62. However, it hasn't been merged into
>>>> mainline yet (I cc'ed Aneesh).
>>>>
>>>> The patch is merged in my "kmemleak-fixes" branch on
>>>> git://linux-arm.org/linux-2.6.git.
>>>
>>> Will try to suck them out and try them.
>>
>> OK -- tested rc8 + a pull of your tree into mine. The bootup was
>> really slow and something was just not going right. After a while
>> memleak complained it had 8 kmemleak logs but I was not able to get my
>> system usable enough to cat the file.
>>
>> In cases like these I wish I would hookup my ctrl-alt-del to kexec() a
>> safe kernel.
>>
>> After a long period of time it seems X wished it would start, it tried
>> and then flashed back to the tty. This kept repeating in a loop.
>>
>> I am not sure if the culprit was rc8 or the kmemleak branch merge --
>> I'll find out after I boot into rc8 in a few.
>
> rc8 busted my bootup, the issues are present with just
> wireless-testing. I highly doubt the issues are wireless-testing
> related so I will not bisect there. Since I am unable to get anything
> useful from the kernel to determine what may have gone sour, any
> suggestions on a path to bisect, or should I just do the whole tree?

I tried 2.6.31-rc8 from hpa's linux-2.6-allstable.git tree instead of
Linus [1] as I already had that tree, git describe says:

v2.6.31-rc8-15-gadda766

Testing this would be the same as testing Linus' blessed rc8 --
correct me I'm wrong. Contrary to what I expected this tree with the
same config works well!

I have compiled a fresh checkout of wireless-testing origin/master to
double check the issue and it is indeed only present on
wireless-testing. A diff stat between John's merge of 2.6.31-rc8 and
current master branch on wireless-testing [2] doesn't reveal much
other than wireless specific stuff, as expected, so it seems this may
after all be introduced in a recent patches in wireless-testing. I
still find this a bit odd given I see no others reporting major
issues. My boot doesn't go very far, it stalls for a while after input
devices are being detected, then it spits out a kmemleak warning about
13 kmemleaks. Here's a picture [3]. I didn't bother waiting as I did
last time for X to try to come up, something is really wrong. I'll
bisect wireless-testing in the morning, starting with a good marker at
merge-2009-08-28 as that is when John pulled 2.6.31-rc8 (and I confirm
a diff stat between that and v2.6.31-rc8 yields nothing as it should)
and current master as the bad marker. I have 9 steps to go, will leave
first step compiling overnight.

[1] git://git.kernel.org/pub/scm/linux/kernel/git/hpa/linux-2.6-allstable.git
[2] git diff --stat merge-2009-08-28..HEAD
[3] http://bombadil.infradead.org/~mcgrof/images/2009/08/lag-wl-2009-08-31.jpg
[4] git diff --stat merge-2009-08-28..v2.6.31-rc8

  Luis