[0/11] more compounding

Message

Ronnie Sahlberg Aug. 23, 2018, 11:31 p.m. UTC

Steve, all

Here is an updated version to test with.
It fixes an issue with RMDIR and smb3 encryption.
It also contain a temporary hack in the last patch to prevent
dequeue_mid() from crashing with a corrupted mid list.
That patch is only here until we understand that issue better and just
so we can play with the rest of the compounding improvements.

Comments

Aurélien Aptel Aug. 24, 2018, 9:50 a.m. UTC | #1

Ronnie Sahlberg <[e-mail]> writes:
> Here is an updated version to test with.
> It fixes an issue with RMDIR and smb3 encryption.
> It also contain a temporary hack in the last patch to prevent
> dequeue_mid() from crashing with a corrupted mid list.
> That patch is only here until we understand that issue better and just
> so we can play with the rest of the compounding improvements.

Ran xfstests again. It goes further along but still does a null ptr
deref while running generic/339:

| ==================================================================
| BUG: KASAN: null-ptr-deref in _raw_spin_lock_irqsave+0x17/0x40
| Write of size 4 at addr 0000000000000000 by task cifsd/19622
| 
| CPU: 0 PID: 19622 Comm: cifsd Not tainted 4.18.0+ #18
| Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.0.0-prebuilt.qemu-project.org 04/01/2014
| Call Trace:
|  dump_stack+0x5b/0x8b
|  kasan_report+0x253/0x2a0
|  ? _raw_spin_lock_irqsave+0x17/0x40
|  _raw_spin_lock_irqsave+0x17/0x40
|  remove_wait_queue+0x12/0x50
|  sk_wait_data+0xf6/0x110
|  ? autoremove_wake_function+0x30/0x30
|  tcp_recvmsg+0x434/0xb00
|  ? __update_load_avg_cfs_rq+0xfa/0x1a0
|  inet_recvmsg+0xa5/0xd0
|  cifs_readv_from_socket+0xfe/0x1e0
|  cifs_read_from_socket+0x3d/0x50
|  ? try_to_wake_up+0x413/0x430
|  ? allocate_buffers+0x85/0xf0
|  cifs_demultiplex_thread+0xe9/0xb40
|  kthread+0x126/0x130
|  ? cifs_handle_standard+0x180/0x180
|  ? kthread_destroy_worker+0x40/0x40
|  ret_from_fork+0x35/0x40

Full log can be found here [1]
It looks like a similar code path, when we read a packet. Also during
this test I see a lot of binary garbage in the file names on the log but
I think this is on purpose.

generic/339 calls:

    src/dirhash_collide -d -n 10000 $testdir >>$seqres.full 2>&1
    
implemented in src/dirhash_collide.c and its purpose is to "Generate
files or directories with hash collisions on a XFS filesystem". I'm not
sure the hashing part is relevant to cifs (is it using the same alg?)
but the -d flag makes the tool create directory. A lot of them
(10000).

1: http://zbeul.ist/tmp/cifs-compounding-crash-2018-08-24.txt.bz2

Cheers,

Pavel Shilovsky Aug. 24, 2018, 5:43 p.m. UTC | #2

пт, 24 авг. 2018 г. в 2:51, Aurélien Aptel <[e-mail]>:
>
> Ronnie Sahlberg <[e-mail]> writes:
> > Here is an updated version to test with.
> > It fixes an issue with RMDIR and smb3 encryption.
> > It also contain a temporary hack in the last patch to prevent
> > dequeue_mid() from crashing with a corrupted mid list.
> > That patch is only here until we understand that issue better and just
> > so we can play with the rest of the compounding improvements.
>
> Ran xfstests again. It goes further along but still does a null ptr
> deref while running generic/339:
>

Hi Aurelien, Ronnie,

Did you have see any crashes on the current master/for-next branches?

--
Best regards,
Pavel Shilovsky

Steve French Aug. 27, 2018, 1:42 a.m. UTC | #3

tentatively pushed to for-next to allow more experimentation and
analysis of the series - but needs some testing (and addressing the
problem that Pavle pointed out)
On Thu, Aug 23, 2018 at 6:31 PM Ronnie Sahlberg <[e-mail]> wrote:
>
> Steve, all
>
> Here is an updated version to test with.
> It fixes an issue with RMDIR and smb3 encryption.
> It also contain a temporary hack in the last patch to prevent
> dequeue_mid() from crashing with a corrupted mid list.
> That patch is only here until we understand that issue better and just
> so we can play with the rest of the compounding improvements.
>
>
>