Re: flood of "dm-X: error: dax access failed" due to 5.9 commit 231609785cbfb

[Coly Li <colyli@suse.de>]

On 2020/9/3 13:20, Coly Li wrote:
> On 2020/9/3 00:51, Mike Snitzer wrote:
>> On Wed, Sep 02 2020 at 12:46pm -0400,
>> Coly Li <colyli@suse.de> wrote:
>>
>>> On 2020/9/3 00:44, Mike Snitzer wrote:
>>>> On Wed, Sep 02 2020 at 12:40pm -0400,
>>>> Coly Li <colyli@suse.de> wrote:
>>>>
>>>>> On 2020/9/3 00:04, Mike Snitzer wrote:
>>>>>> 5.9 commit 231609785cbfb ("dax: print error message by pr_info() in
>>>>>> __generic_fsdax_supported()") switched from pr_debug() to pr_info().
>>>>>>
>>>>>> The justification in the commit header is really inadequate.  If there
>>>>>> is a problem that you need to drill in on, repeat the testing after
>>>>>> enabling the dynamic debugging.
>>>>>>
>>>>>> Otherwise, now all DM devices that aren't layered on DAX capable devices
>>>>>> spew really confusing noise to users when they simply activate their
>>>>>> non-DAX DM devices:
>>>>>>
>>>>>> [66567.129798] dm-6: error: dax access failed (-5)
>>>>>> [66567.134400] dm-6: error: dax access failed (-5)
>>>>>> [66567.139152] dm-6: error: dax access failed (-5)
>>>>>> [66567.314546] dm-2: error: dax access failed (-95)
>>>>>> [66567.319380] dm-2: error: dax access failed (-95)
>>>>>> [66567.324254] dm-2: error: dax access failed (-95)
>>>>>> [66567.479025] dm-2: error: dax access failed (-95)
>>>>>> [66567.483713] dm-2: error: dax access failed (-95)
>>>>>> [66567.488722] dm-2: error: dax access failed (-95)
>>>>>> [66567.494061] dm-2: error: dax access failed (-95)
>>>>>> [66567.498823] dm-2: error: dax access failed (-95)
>>>>>> [66567.503693] dm-2: error: dax access failed (-95)
>>>>>>
>>>>>> commit 231609785cbfb must be reverted.
>>>>>>
>>>>>> Please advise, thanks.
>>>>>
>>>>> Adrian Huang from Lenovo posted a patch, which titled: dax: do not print
>>>>> error message for non-persistent memory block device
>>>>>
>>>>> It fixes the issue, but no response for now. Maybe we should take this fix.
>>>>
>>>> OK, yes sounds like it.  It was merged and is commit c2affe920b0e066
>>>> ("dax: do not print error message for non-persistent memory block
>>>> device")
>>>
>>> Thanks for informing me this patch is merged, I am going to update my
>>> local one :-)
>>
>> So the thing is I'm running v5.9-rc3 (which includes this commit) but
>> I'm still seeing all these warnings when I run the lvm2 testsuite.  The
>> reason _seems_ to be because the lvm2 testsuite uses brd devices for
>> test devices.  So there is something about the brd device that shows
>> commit c2affe920b0e066 isn't enough :(
> 
> [Resend and CC Adrian Huang]
> 
> Hi Mike,
> 
> Could you please apply and test this attached patch based on v5.9-rc3 ?
> 
> It seems the pointer dax_dev of __generic_fsdax_supported() parameter is
> not initialized (IMHO this is not a dm bug), therefore the && should be
> || to check the dax support state.
> 
> Also I add two pr_info() to print the variables value, let's see whether
> my guess makes sense.

Also I suggest some kind of change like this in drivers/md/dm.c,

diff --git a/drivers/md/dm.c b/drivers/md/dm.c
index fb0255d25e4b..566d8208df47 100644
--- a/drivers/md/dm.c
+++ b/drivers/md/dm.c
@@ -818,6 +818,8 @@ int dm_get_table_device(struct mapped_device *md,
dev_t dev, fmode_t mode,
                        return -ENOMEM;
                }

+               memset(td, 0, sizeof(struct table_device));
+
                td->dm_dev.mode = mode;
                td->dm_dev.bdev = NULL;


The above change may make sure *dax_dev sent into
__generic_fsdax_supported() is always NULL if the target does not
support DAX. But IMHO this is not 100% necessary, it just make
__generic_fsdax_supported() return false faster by the following change
in previous attached patch,

 -       if (!dax_dev && !bdev_dax_supported(bdev, blocksize)) {
 +       if (!dax_dev || !bdev_dax_supported(bdev, blocksize)) {

I am not very familiar with dm code, CMIIW, just for your information.

Coly Li
_______________________________________________
Linux-nvdimm mailing list -- linux-nvdimm@lists.01.org
To unsubscribe send an email to linux-nvdimm-leave@lists.01.org