lkml.org 
[lkml]   [2011]   [Mar]   [6]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
SubjectRe: BUG at mm/mmap.c:2309 when cx18.ko and cx18-alsa.ko loaded
From
On Sat, Mar 5, 2011 at 6:03 PM, Andy Walls <awalls@md.metrocast.net> wrote:
> On Sat, 2011-03-05 at 16:59 -0500, Andy Walls wrote:
>> On Thu, 2011-03-03 at 21:06 -0500, Andy Walls wrote:
>> > Hi,
>> >
>> > I got a BUG when loading the cx18.ko module (which in turn requests the
>> > cx18-alsa.ko module) on a kernel built from this repository
>> >
>> >     http://git.linuxtv.org/media_tree.git staging/for_v2.6.39
>> >
>> > which I beleive is based on 2.6.38-rc2.
>>
>> [snip]
>>
>> > So here is my transcription of a fuzzy digital photo of the screen:
>> >
>> > kernel BUG at /home/andy/cx18dev/git/media_tree/mm/mmap.c:2309!
>> > invalid opcode: 0000 [#1] SMP
>> > last sysfs file: /sys/module/snd_pcm/initstate
>> > Modules linked in: tda9887 tda8290 mxl5005s s5h1409 tuner_simple ...
>> > ...
>> > Pid: 2580, comm: udevd Not tainted 2.6.38-rc2-cx18-vb2-proto+
>> > RIP: 0010:[<ffffffff810eb50b>]  [<ffffffff810eb50b>] exit_mmap+0x10f/0x11e
>> > RAX: 0000000000000000 RBX: 0000000000000000 RCX: 0020000000000000
>> > RDX: 0000000000160011 RSI: ffffea____c42___ RDI: 0000000000000202
>> > RBP: ffff____18c1f_58 R08: ffff____________ R09: 0000000000000004
>> > R10: ffff_______bb_38 R11: 0000000000000000 R12: ffff____344a6680
>> > R13: 00007fff22______ R14: ffff____________ R15: 0000000000000001
>> > ...
>> > CR2: 0000000000000000 ...
>> > ....
>> > Process udevd (pid: 25__, threadinfo ffff________, ...
>> > Stack:
>> >  000000000000015e ffff00003bc0e1d0 0000000000000246 ....
>> > .....
>> > Call Trace:
>> > ... mmput+0x63/0xcf
>> > ... exit_mm+0x132/0x13f
>> > ... do_exit+0x238/0x749
>> > ... ? __dequeue_signal+0xfa/0x12f
>> > ... do_group_exit+0x7d/0xa5
>> > ... get_signal_to_deliver+0x371/0x395
>> > ... do_signal+0x72/0x692
>> > ... ? do_page_fault+0x24a/0x391
>> > ... ? printk+0x41/0x47
>> > ... ? sigprocmask+0xa3/0xcd
>> > ... do_notify_resume+0x2c/0x64
>> > ... retint_signal+0x48/0x8c
>> >
>> > Code: ff ff 48 8b 7d d8 4c 89 ea 31 f6 e8 3e fe ff ff 48 89 df e8 78 fe
>> > ff ff 48 85 c0 48 89 c3 75 f0 49 83 bc 24 e0 00 00 00 00 74 04 <0f> 0b
>> > eb fe 48 83 c4 18 5b 41 5c 41 5d c9 c3 55 48 89 e5 41 57
>> > RIP  [<ffffffff810eb50b>] exit_mmap+0x10f/0x11e
>> >  RSP <ffff880018c1fc28>
>> > general protection fault: 0000 [#2] SMP
>> > last sysfs file: /sys/devices/virtual/sound/card2/uevent
>> > CPU 1
>> > Modules linked in: cx18-alsa tda9887 tda8290 mxl5005s s5h1409
>> > tuner_simple tuner_types cs5345 tuner cx18 dvb_core cx2341x v4l2_common
>> > videodev v4l2_compat_ioctl32
>>
>>
>> I'm dumping all my previous assumtpions about this BUG.  After a bit of
>> reading, all I can say is that it's a page table deallocation problem at
>> process exit.  After all the page table deallocations on exit,
>> mm->nr_ptes is still > 0, and that's a bad thing.
>>
>> It apparently happened in a child udevd exiting shortly after cx18.ko
>> loaded.  The cx18 driver allocating large amounts kernel memory for DMA
>> buffers upon load may be related to triggering the problem, but I doubt
>> it is a root cause of the BUG.
>>
>>
>> This monsterous thread from 5 years ago is somewhat enlightening:
>>
>>       http://lkml.indiana.edu/hypermail/linux/kernel/0503.2/1680.html
>>       http://lkml.indiana.edu/hypermail/linux/kernel/0503.2/1787.html
>>
>> so it gives me a place to start looking for the problem.
>>
>> Any advice on what data to collect is appreciated.
>
> When attemtping to reproduce this BUG, I got another bug related to
> memory management:
>
> (Details handtyped from a photo):
> BUG: unable to handle kernel NULL pointer dereference at           (null)
> IP: [<ffffffff010f22fa>] remove_vm_area+0x42/0x77
> PGD 37cdd067 PUD 336c__67 PMD 0
> Oops: 0000 [#1] SMP
> last sysfs file: /sys/devices/pci0000:00/0000:00:14.4/0000:03:00.0/firmware/0000:03:00.0/loading
> CPU 0
> Modules linked in: tda9887 tda8290 mxl5005s s5h1409 tuner_simple tuner_types cx5345 tuner cx18(+) dvb_core cx2341x ...
> Pid: 2470, comm: work_for_cpu Tainted: G        W 2.6.28-rc2-cx18-vb2-proto+
> RIP: 0010:[<ffffffff010f22fa>]  [<ffffffff010f22fa>] remove_vm_area+0x42/0x77
> ...
> RAX: 0000000000000000 RBX: ffff____35e7c540 RCX: 0000000000001000
> RDX: 0000000000000000 ....
> ...
> CR2: 0000000000000000 ....
> Stack:
>  ffff__0011485968 000000000000001 ffff____1147dc9_ ffffffff_1_f23__
> ....
> Call Trace:
> ... __vunmap+0x3e/0xbd
> ... vfree+0x2e/0x30
> ... dvb_dmx_init+0x7e/0x253 [dvb_core]
> ... cx18_dvb_register+0xd2/0x75c [cx18]
> ... cx18_streams_resgister+0x6a/0x26a [cx18]
> ... cx18_streams_setup+0x3cc/0x486 [cx18]
> ... cx18_probe+0x11cc/0x12fb [cx18]
> ......
>
> The code appears to be failing here:
>
> /home/andy/cx18dev/git/media_tree/mm/vmalloc.c:1352
>    161d:       eb 06                   jmp    1625 <remove_vm_area+0x45>
>    161f:       48 89 c2                mov    %rax,%rdx
>    1622:       48 8b 00                mov    (%rax),%rax    <--- Oops  p = &tmp->next)  (tmp = *p)
>    1625:       48 39 d8                cmp    %rbx,%rax                (tmp = *p) != vm;
>    1628:       75 f5                   jne    161f <remove_vm_area+0x3f>
> /home/andy/cx18dev/git/media_tree/mm/vmalloc.c:1354
>
> Corresponding to this code in mm/vmalloc.c:
>
> struct vm_struct *remove_vm_area(const void *addr)
> {
>        struct vmap_area *va;
>
>        va = find_vmap_area((unsigned long)addr);
>        if (va && va->flags & VM_VM_AREA) {
>                struct vm_struct *vm = va->private;
>                struct vm_struct *tmp, **p;
>                /*
>                 * remove from list and disallow access to this vm_struct
>                 * before unmap. (address range confliction is maintained by
>                 * vmap.)
>                 */
>                write_lock(&vmlist_lock);
>                for (p = &vmlist; (tmp = *p) != vm; p = &tmp->next)  <--- Ooops
>                        ;
> [...]
>
> That for() loop appears to assume the vm_struct will be on the vmlist
> somewhere.  If it isn't, then I suppose the for() loop could end up
> doing a NULL dereference.
>
> This BUG happened in the final stages of the cx18 driver setting up a
> CX23418 card instance.  I have 2 cards in this machine, so a number of
> buffers had certainly been allocated using kmalloc().  The code in the
> dvb_core that is failing got BUG'ed in this case was this:
>
> int dvb_dmx_init(struct dvb_demux *dvbdemux)
> {
>        int i;
>        struct dmx_demux *dmx = &dvbdemux->dmx;
>
>        dvbdemux->cnt_storage = NULL;
>        dvbdemux->users = 0;
>        dvbdemux->filter = vmalloc(dvbdemux->filternum * sizeof(struct dvb_demux_filter));
>
>        if (!dvbdemux->filter)
>                return -ENOMEM;
>
>        dvbdemux->feed = vmalloc(dvbdemux->feednum * sizeof(struct dvb_demux_feed));
>        if (!dvbdemux->feed) {
>                vfree(dvbdemux->filter);     <------- BUG/Oops happened in this call
>                dvbdemux->filter = NULL;
>                return -ENOMEM;
>        }
> ...
>
> Which is kind of interesting:
> 1. The first vmalloc() succeeded.
> 2. The second vmalloc() failed.
> 3. The vfree() of the pointer from the first vmalloc() caused an
> Oops/BUG.
>
> I'm not sure where to go from here.

Thanks for all the effort you are putting into investigating this: you
deserve a better response than I can give you.

mm/vmalloc.c's vmap_area handling is entirely separate from
mm/mmap.c's vm_area_struct handling, yet both misbehaviors would be
explained if a next pointer has been corrupted to NULL.

Probably just coincidence that they both manifest that way, though the
underlying problem may turn out to be one.

If you have not already, it would be well worth turning on
CONFIG_DEBUG_LIST and CONFIG_DEBUG_SLAB or CONFIG_SLUB_DEBUG with
CONFIG_SLUB_DEBUG_ON.

If that BUG_ON(mm->nr_ptes ...) in exit_mmap() is preventing you from
getting on with your work, or slowing down reproduction of the
testcase, you should be able to replace it by a WARN_ON. You will
probably leak at least one page (the page table) and perhaps many
pages (those that that page table points to) each time it hits, but it
shouldn't actually be unsafe to continue - it's really a development
BUG_ON, to check that new architectures added are freeing all the
page tables they have allocated.

I do expect the underlying problem to be somewhere down the driver
end, given that nobody else has been reporting these issues. I'm
hoping that once the cx18 guys have time to try to reproduce it,
they'll be better able to track it down. But you are having trouble
reproducing it yourself? hitting this vmalloc one before you could
reproduce the exit_mmap one? No chance to bisect it to a particular
commit if you cannot reliably reproduce it.

There was a horrid list corruption bug in early 2.6.38-rc, fixed in
-rc6; but although I guess it could cause all kinds of havoc, its
particular signature was not like this, so I don't really believe that
one was to blame here.

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

\
 
 \ /
  Last update: 2011-03-06 19:41    [W:0.113 / U:0.180 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site