Re: BUG: aio/direct-io data corruption in 4.7

Re: BUG: aio/direct-io data corruption in 4.7

[Gregory Shapiro]

Hello, my name is Gregory Shapiro and I am a newbie on this list.
I recently encountered data corruption as I got a kernel to
acknowledge write ("io_getevents" system call with a correct number of
bytes) but undergoing write to disk failed.
After investigating the problem I found it is identical to issue found
in direct-io.c mentioned the bellow thread.
https://lore.kernel.org/lkml/20160921141539.GA17898@infradead.org/
Is there a reason proposed patch didn't apply to the kernel?
When can I expect it to be applied?
Thanks,
 Gregory

Re: BUG: aio/direct-io data corruption in 4.7

[Jack Wang]

Gregory Shapiro <shapiro.gregory@gmail.com> 于2018年11月5日周一 下午4:19写道：
>
> Hello, my name is Gregory Shapiro and I am a newbie on this list.
> I recently encountered data corruption as I got a kernel to
> acknowledge write ("io_getevents" system call with a correct number of
> bytes) but undergoing write to disk failed.
> After investigating the problem I found it is identical to issue found
> in direct-io.c mentioned the bellow thread.
> https://lore.kernel.org/lkml/20160921141539.GA17898@infradead.org/
> Is there a reason proposed patch didn't apply to the kernel?
> When can I expect it to be applied?
> Thanks,
>  Gregory

Hi Gregory,

Thanks for your info.
Have you tried with latest kernel other than 4.7, is the problem still there?

Could you share your test case?

Regards,
Jack Wang

Re: BUG: aio/direct-io data corruption in 4.7

[Gregory Shapiro]

Hi Jack,
I tested it in 4.9.102 and I checked the latest code from elixir
(versions 4.19 and 4.20) and the error in code is still present there.
More on the scenario and the bug:
I experienced data corruption in my application (nvme based storage).
The issue was caused because of faulty hardware, but the real problem
is I got a correct number of bytes in io_getevents thus couldn't
recognize it correctly the error.
Looking at the /var/log/messages and  I saw the following errors in
time of coruption:

Oct 11 14:55:15 block01-node05 kernel: [19272.951015]
blk_update_request: I/O error, dev nvme2n3, sector 117359360
Oct 11 14:55:15 block01-node05 kernel: [19272.952786]
blk_update_request: I/O error, dev nvme2n3, sector 117359872
Oct 11 14:55:16 block01-node05 kernel: [19273.544374]
blk_update_request: I/O error, dev nvme2n3, sector 117360384
...
So the block level does receive information about the error, but I
don't see it in the application.
running ftrace and doing code reading I find out that dio error status
is overridden.
In dio_complete it is propagated in (dio->io_error and if
dio->io_error is not zero in we are in async write
the status is overridden by transferred.

static ssize_t dio_complete(struct dio *dio, ssize_t ret, bool is_async)
{
...
        if (ret == 0)
                ret = dio->page_errors;
        if (ret == 0)
                ret = dio->io_error;
        if (ret == 0)
                ret = transferred;
...
        if (is_async) {
                /*
                 * generic_write_sync expects ki_pos to have been updated
                 * already, but the submission path only does this for
                 * synchronous I/O.
                 */
                dio->iocb->ki_pos += transferred;

                if (dio->op == REQ_OP_WRITE)
                        ret = generic_write_sync(dio->iocb,  transferred);
                dio->iocb->ki_complete(dio->iocb, ret, 0);



For your convenience I am attaching ftrace log to for easier tracking
the flow in the code:


 26)               |                nvme_complete_rq [nvme_core]() {
 26)               |                  blk_mq_end_request() {
 26)               |                    blk_update_request() { <----
log is from here
 26)   0.563 us    |                      blk_account_io_completion();
 26)   0.263 us    |                      bio_advance();
 26)               |                      bio_endio() {
 26)               |                        dio_bio_end_aio() {
 26)               |                          dio_bio_complete() {
 26)               |                            bio_check_pages_dirty() {
 26)               |                              bio_put() {
 26)               |                                bio_free() {
 26)               |                                  __bio_free() {
 26)   0.045 us    |                                    bio_disassociate_task();
 26)   0.497 us    |                                  }
 26)   0.042 us    |                                  bvec_free();
 26)               |                                  mempool_free() {
 26)               |                                    mempool_free_slab() {
 26)   0.264 us    |                                      kmem_cache_free();
 26)   0.606 us    |                                    }
 26)   1.125 us    |                                  }
 26)   2.588 us    |                                }
 26)   2.920 us    |                              }
 26)   3.979 us    |                            }
 26)   4.712 us    |                          }
 26)   0.040 us    |                          _raw_spin_lock_irqsave();
 26)   0.048 us    |                          _raw_spin_unlock_irqrestore();
 26)               |                          dio_complete() {
dio_complete(dio, 0, true);
 26)               |                            aio_complete() {
dio->iocb->ki_complete(dio->iocb, ret, 0); <<aio_complete(struct kiocb
*kiocb, long res, long res2)>>
 26)   0.073 us    |                              _raw_spin_lock_irqsave();
 26)   0.114 us    |                              refill_reqs_available();
 26)   0.048 us    |                              _raw_spin_unlock_irqrestore();
 26)               |                              kiocb_free() {
 26)   0.171 us    |                                fput();
 26)   0.102 us    |                                kmem_cache_free();
 26)   0.902 us    |                              }


        }

On Tue, Nov 6, 2018 at 9:29 AM Jack Wang <jack.wang.usish@gmail.com> wrote:
>
> Gregory Shapiro <shapiro.gregory@gmail.com> 于2018年11月5日周一 下午4:19写道：
> >
> > Hello, my name is Gregory Shapiro and I am a newbie on this list.
> > I recently encountered data corruption as I got a kernel to
> > acknowledge write ("io_getevents" system call with a correct number of
> > bytes) but undergoing write to disk failed.
> > After investigating the problem I found it is identical to issue found
> > in direct-io.c mentioned the bellow thread.
> > https://lore.kernel.org/lkml/20160921141539.GA17898@infradead.org/
> > Is there a reason proposed patch didn't apply to the kernel?
> > When can I expect it to be applied?
> > Thanks,
> >  Gregory
>
> Hi Gregory,
>
> Thanks for your info.
> Have you tried with latest kernel other than 4.7, is the problem still there?
>
> Could you share your test case?
>
> Regards,
> Jack Wang

Re: BUG: aio/direct-io data corruption in 4.7

[Jack Wang]

Gregory Shapiro <shapiro.gregory@gmail.com> 于2018年11月6日周二 下午12:31写道：
>
> Hi Jack,
> I tested it in 4.9.102 and I checked the latest code from elixir
> (versions 4.19 and 4.20) and the error in code is still present there.
> More on the scenario and the bug:
> I experienced data corruption in my application (nvme based storage).
> The issue was caused because of faulty hardware, but the real problem
> is I got a correct number of bytes in io_getevents thus couldn't
> recognize it correctly the error.
> Looking at the /var/log/messages and  I saw the following errors in
> time of coruption:
Thanks for the info, Gregory.

I noticed guys from Amazon pushing the fix to upstream:
https://lore.kernel.org/patchwork/patch/1008443/
I hope it will be in upstream soon.


Regards,
Jack Wang

BUG: aio/direct-io data corruption in 4.7

[Jonathan Nicklin]

In 4.7.2, the kernel is acknowledging block writes that have not
completed to disk. To reproduce: create an MD array, run FIO (direct +
libaio), and pull all drives. FIO will continue to run without
receiving I/O errors. I have also reproduced the bug using physical
drives. In this case, only a limited number of I/Os are incorrectly
acknowledged; FIO eventually receives an I/O error after the device
reference is removed.

The root cause of the problem is that dio_complete() does not
correctly propagate I/O errors in the is_async case. Specifically,
generic_write_sync() appears to be overwriting the return status
destined for ki_complete().

This bug appears to have been introduced by the following commit:

Description: "fs: simplify the generic_write_sync prototype"
Committed: Apr 7, 2016
Hash: e259221763a40403d5bb232209998e8c45804ab8
Affects: 4.7-rc1 - master

I have confirmed a fix for the AIO/Direct-IO failure condition but
have not reviewed the rest of the changes associated with that commit.
If you would like a small patch for direct-io.c, let me know.

Regards,
-Jonathan

Re: BUG: aio/direct-io data corruption in 4.7

[Christoph Hellwig]

Hi Jonathan,

please keep linux-fsdevel on the Cc list for something like, and if
you already track down a commit the author of that commit.

> Description: "fs: simplify the generic_write_sync prototype"
> Committed: Apr 7, 2016
> Hash: e259221763a40403d5bb232209998e8c45804ab8
> Affects: 4.7-rc1 - master
> 
> I have confirmed a fix for the AIO/Direct-IO failure condition but
> have not reviewed the rest of the changes associated with that commit.
> If you would like a small patch for direct-io.c, let me know.

On travel right now, but I suspect you want something like this fix?

diff --git a/fs/direct-io.c b/fs/direct-io.c
index 7c3ce73..891f71f 100644
--- a/fs/direct-io.c
+++ b/fs/direct-io.c
@@ -276,7 +276,7 @@ static ssize_t dio_complete(struct dio *dio, ssize_t ret, bool is_async)
 		dio->iocb->ki_pos += transferred;
 
 		if (dio->op == REQ_OP_WRITE)
-			ret = generic_write_sync(dio->iocb,  transferred);
+			ret = generic_write_sync(dio->iocb, ret);
 		dio->iocb->ki_complete(dio->iocb, ret, 0);
 	}